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> | |
80110bbf | 19 | #include <linux/page_table_check.h> |
b46e756f | 20 | #include <linux/swapops.h> |
f3f0e1d2 | 21 | #include <linux/shmem_fs.h> |
b46e756f KS |
22 | |
23 | #include <asm/tlb.h> | |
24 | #include <asm/pgalloc.h> | |
25 | #include "internal.h" | |
26 | ||
27 | enum scan_result { | |
28 | SCAN_FAIL, | |
29 | SCAN_SUCCEED, | |
30 | SCAN_PMD_NULL, | |
31 | SCAN_EXCEED_NONE_PTE, | |
71a2c112 KS |
32 | SCAN_EXCEED_SWAP_PTE, |
33 | SCAN_EXCEED_SHARED_PTE, | |
b46e756f | 34 | SCAN_PTE_NON_PRESENT, |
e1e267c7 | 35 | SCAN_PTE_UFFD_WP, |
b46e756f | 36 | SCAN_PAGE_RO, |
0db501f7 | 37 | SCAN_LACK_REFERENCED_PAGE, |
b46e756f KS |
38 | SCAN_PAGE_NULL, |
39 | SCAN_SCAN_ABORT, | |
40 | SCAN_PAGE_COUNT, | |
41 | SCAN_PAGE_LRU, | |
42 | SCAN_PAGE_LOCK, | |
43 | SCAN_PAGE_ANON, | |
44 | SCAN_PAGE_COMPOUND, | |
45 | SCAN_ANY_PROCESS, | |
46 | SCAN_VMA_NULL, | |
47 | SCAN_VMA_CHECK, | |
48 | SCAN_ADDRESS_RANGE, | |
b46e756f KS |
49 | SCAN_DEL_PAGE_LRU, |
50 | SCAN_ALLOC_HUGE_PAGE_FAIL, | |
51 | SCAN_CGROUP_CHARGE_FAIL, | |
f3f0e1d2 | 52 | SCAN_TRUNCATED, |
99cb0dbd | 53 | SCAN_PAGE_HAS_PRIVATE, |
b46e756f KS |
54 | }; |
55 | ||
56 | #define CREATE_TRACE_POINTS | |
57 | #include <trace/events/huge_memory.h> | |
58 | ||
4aab2be0 VB |
59 | static struct task_struct *khugepaged_thread __read_mostly; |
60 | static DEFINE_MUTEX(khugepaged_mutex); | |
61 | ||
b46e756f KS |
62 | /* default scan 8*512 pte (or vmas) every 30 second */ |
63 | static unsigned int khugepaged_pages_to_scan __read_mostly; | |
64 | static unsigned int khugepaged_pages_collapsed; | |
65 | static unsigned int khugepaged_full_scans; | |
66 | static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; | |
67 | /* during fragmentation poll the hugepage allocator once every minute */ | |
68 | static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; | |
69 | static unsigned long khugepaged_sleep_expire; | |
70 | static DEFINE_SPINLOCK(khugepaged_mm_lock); | |
71 | static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); | |
72 | /* | |
73 | * default collapse hugepages if there is at least one pte mapped like | |
74 | * it would have happened if the vma was large enough during page | |
75 | * fault. | |
76 | */ | |
77 | static unsigned int khugepaged_max_ptes_none __read_mostly; | |
78 | static unsigned int khugepaged_max_ptes_swap __read_mostly; | |
71a2c112 | 79 | static unsigned int khugepaged_max_ptes_shared __read_mostly; |
b46e756f KS |
80 | |
81 | #define MM_SLOTS_HASH_BITS 10 | |
82 | static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); | |
83 | ||
84 | static struct kmem_cache *mm_slot_cache __read_mostly; | |
85 | ||
27e1f827 SL |
86 | #define MAX_PTE_MAPPED_THP 8 |
87 | ||
b46e756f KS |
88 | /** |
89 | * struct mm_slot - hash lookup from mm to mm_slot | |
90 | * @hash: hash collision list | |
91 | * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head | |
92 | * @mm: the mm that this information is valid for | |
336e6b53 AS |
93 | * @nr_pte_mapped_thp: number of pte mapped THP |
94 | * @pte_mapped_thp: address array corresponding pte mapped THP | |
b46e756f KS |
95 | */ |
96 | struct mm_slot { | |
97 | struct hlist_node hash; | |
98 | struct list_head mm_node; | |
99 | struct mm_struct *mm; | |
27e1f827 SL |
100 | |
101 | /* pte-mapped THP in this mm */ | |
102 | int nr_pte_mapped_thp; | |
103 | unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP]; | |
b46e756f KS |
104 | }; |
105 | ||
106 | /** | |
107 | * struct khugepaged_scan - cursor for scanning | |
108 | * @mm_head: the head of the mm list to scan | |
109 | * @mm_slot: the current mm_slot we are scanning | |
110 | * @address: the next address inside that to be scanned | |
111 | * | |
112 | * There is only the one khugepaged_scan instance of this cursor structure. | |
113 | */ | |
114 | struct khugepaged_scan { | |
115 | struct list_head mm_head; | |
116 | struct mm_slot *mm_slot; | |
117 | unsigned long address; | |
118 | }; | |
119 | ||
120 | static struct khugepaged_scan khugepaged_scan = { | |
121 | .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), | |
122 | }; | |
123 | ||
e1465d12 | 124 | #ifdef CONFIG_SYSFS |
b46e756f KS |
125 | static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, |
126 | struct kobj_attribute *attr, | |
127 | char *buf) | |
128 | { | |
ae7a927d | 129 | return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs); |
b46e756f KS |
130 | } |
131 | ||
132 | static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, | |
133 | struct kobj_attribute *attr, | |
134 | const char *buf, size_t count) | |
135 | { | |
dfefd226 | 136 | unsigned int msecs; |
b46e756f KS |
137 | int err; |
138 | ||
dfefd226 AD |
139 | err = kstrtouint(buf, 10, &msecs); |
140 | if (err) | |
b46e756f KS |
141 | return -EINVAL; |
142 | ||
143 | khugepaged_scan_sleep_millisecs = msecs; | |
144 | khugepaged_sleep_expire = 0; | |
145 | wake_up_interruptible(&khugepaged_wait); | |
146 | ||
147 | return count; | |
148 | } | |
149 | static struct kobj_attribute scan_sleep_millisecs_attr = | |
150 | __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show, | |
151 | scan_sleep_millisecs_store); | |
152 | ||
153 | static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, | |
154 | struct kobj_attribute *attr, | |
155 | char *buf) | |
156 | { | |
ae7a927d | 157 | return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs); |
b46e756f KS |
158 | } |
159 | ||
160 | static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, | |
161 | struct kobj_attribute *attr, | |
162 | const char *buf, size_t count) | |
163 | { | |
dfefd226 | 164 | unsigned int msecs; |
b46e756f KS |
165 | int err; |
166 | ||
dfefd226 AD |
167 | err = kstrtouint(buf, 10, &msecs); |
168 | if (err) | |
b46e756f KS |
169 | return -EINVAL; |
170 | ||
171 | khugepaged_alloc_sleep_millisecs = msecs; | |
172 | khugepaged_sleep_expire = 0; | |
173 | wake_up_interruptible(&khugepaged_wait); | |
174 | ||
175 | return count; | |
176 | } | |
177 | static struct kobj_attribute alloc_sleep_millisecs_attr = | |
178 | __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show, | |
179 | alloc_sleep_millisecs_store); | |
180 | ||
181 | static ssize_t pages_to_scan_show(struct kobject *kobj, | |
182 | struct kobj_attribute *attr, | |
183 | char *buf) | |
184 | { | |
ae7a927d | 185 | return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan); |
b46e756f KS |
186 | } |
187 | static ssize_t pages_to_scan_store(struct kobject *kobj, | |
188 | struct kobj_attribute *attr, | |
189 | const char *buf, size_t count) | |
190 | { | |
dfefd226 | 191 | unsigned int pages; |
b46e756f | 192 | int err; |
b46e756f | 193 | |
dfefd226 AD |
194 | err = kstrtouint(buf, 10, &pages); |
195 | if (err || !pages) | |
b46e756f KS |
196 | return -EINVAL; |
197 | ||
198 | khugepaged_pages_to_scan = pages; | |
199 | ||
200 | return count; | |
201 | } | |
202 | static struct kobj_attribute pages_to_scan_attr = | |
203 | __ATTR(pages_to_scan, 0644, pages_to_scan_show, | |
204 | pages_to_scan_store); | |
205 | ||
206 | static ssize_t pages_collapsed_show(struct kobject *kobj, | |
207 | struct kobj_attribute *attr, | |
208 | char *buf) | |
209 | { | |
ae7a927d | 210 | return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed); |
b46e756f KS |
211 | } |
212 | static struct kobj_attribute pages_collapsed_attr = | |
213 | __ATTR_RO(pages_collapsed); | |
214 | ||
215 | static ssize_t full_scans_show(struct kobject *kobj, | |
216 | struct kobj_attribute *attr, | |
217 | char *buf) | |
218 | { | |
ae7a927d | 219 | return sysfs_emit(buf, "%u\n", khugepaged_full_scans); |
b46e756f KS |
220 | } |
221 | static struct kobj_attribute full_scans_attr = | |
222 | __ATTR_RO(full_scans); | |
223 | ||
224 | static ssize_t khugepaged_defrag_show(struct kobject *kobj, | |
225 | struct kobj_attribute *attr, char *buf) | |
226 | { | |
227 | return single_hugepage_flag_show(kobj, attr, buf, | |
ae7a927d | 228 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); |
b46e756f KS |
229 | } |
230 | static ssize_t khugepaged_defrag_store(struct kobject *kobj, | |
231 | struct kobj_attribute *attr, | |
232 | const char *buf, size_t count) | |
233 | { | |
234 | return single_hugepage_flag_store(kobj, attr, buf, count, | |
235 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); | |
236 | } | |
237 | static struct kobj_attribute khugepaged_defrag_attr = | |
238 | __ATTR(defrag, 0644, khugepaged_defrag_show, | |
239 | khugepaged_defrag_store); | |
240 | ||
241 | /* | |
242 | * max_ptes_none controls if khugepaged should collapse hugepages over | |
243 | * any unmapped ptes in turn potentially increasing the memory | |
244 | * footprint of the vmas. When max_ptes_none is 0 khugepaged will not | |
245 | * reduce the available free memory in the system as it | |
246 | * runs. Increasing max_ptes_none will instead potentially reduce the | |
247 | * free memory in the system during the khugepaged scan. | |
248 | */ | |
249 | static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj, | |
250 | struct kobj_attribute *attr, | |
251 | char *buf) | |
252 | { | |
ae7a927d | 253 | return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none); |
b46e756f KS |
254 | } |
255 | static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj, | |
256 | struct kobj_attribute *attr, | |
257 | const char *buf, size_t count) | |
258 | { | |
259 | int err; | |
260 | unsigned long max_ptes_none; | |
261 | ||
262 | err = kstrtoul(buf, 10, &max_ptes_none); | |
36ee2c78 | 263 | if (err || max_ptes_none > HPAGE_PMD_NR - 1) |
b46e756f KS |
264 | return -EINVAL; |
265 | ||
266 | khugepaged_max_ptes_none = max_ptes_none; | |
267 | ||
268 | return count; | |
269 | } | |
270 | static struct kobj_attribute khugepaged_max_ptes_none_attr = | |
271 | __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show, | |
272 | khugepaged_max_ptes_none_store); | |
273 | ||
274 | static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj, | |
275 | struct kobj_attribute *attr, | |
276 | char *buf) | |
277 | { | |
ae7a927d | 278 | return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap); |
b46e756f KS |
279 | } |
280 | ||
281 | static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj, | |
282 | struct kobj_attribute *attr, | |
283 | const char *buf, size_t count) | |
284 | { | |
285 | int err; | |
286 | unsigned long max_ptes_swap; | |
287 | ||
288 | err = kstrtoul(buf, 10, &max_ptes_swap); | |
36ee2c78 | 289 | if (err || max_ptes_swap > HPAGE_PMD_NR - 1) |
b46e756f KS |
290 | return -EINVAL; |
291 | ||
292 | khugepaged_max_ptes_swap = max_ptes_swap; | |
293 | ||
294 | return count; | |
295 | } | |
296 | ||
297 | static struct kobj_attribute khugepaged_max_ptes_swap_attr = | |
298 | __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show, | |
299 | khugepaged_max_ptes_swap_store); | |
300 | ||
71a2c112 | 301 | static ssize_t khugepaged_max_ptes_shared_show(struct kobject *kobj, |
ae7a927d JP |
302 | struct kobj_attribute *attr, |
303 | char *buf) | |
71a2c112 | 304 | { |
ae7a927d | 305 | return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared); |
71a2c112 KS |
306 | } |
307 | ||
308 | static ssize_t khugepaged_max_ptes_shared_store(struct kobject *kobj, | |
309 | struct kobj_attribute *attr, | |
310 | const char *buf, size_t count) | |
311 | { | |
312 | int err; | |
313 | unsigned long max_ptes_shared; | |
314 | ||
315 | err = kstrtoul(buf, 10, &max_ptes_shared); | |
36ee2c78 | 316 | if (err || max_ptes_shared > HPAGE_PMD_NR - 1) |
71a2c112 KS |
317 | return -EINVAL; |
318 | ||
319 | khugepaged_max_ptes_shared = max_ptes_shared; | |
320 | ||
321 | return count; | |
322 | } | |
323 | ||
324 | static struct kobj_attribute khugepaged_max_ptes_shared_attr = | |
325 | __ATTR(max_ptes_shared, 0644, khugepaged_max_ptes_shared_show, | |
326 | khugepaged_max_ptes_shared_store); | |
327 | ||
b46e756f KS |
328 | static struct attribute *khugepaged_attr[] = { |
329 | &khugepaged_defrag_attr.attr, | |
330 | &khugepaged_max_ptes_none_attr.attr, | |
71a2c112 KS |
331 | &khugepaged_max_ptes_swap_attr.attr, |
332 | &khugepaged_max_ptes_shared_attr.attr, | |
b46e756f KS |
333 | &pages_to_scan_attr.attr, |
334 | &pages_collapsed_attr.attr, | |
335 | &full_scans_attr.attr, | |
336 | &scan_sleep_millisecs_attr.attr, | |
337 | &alloc_sleep_millisecs_attr.attr, | |
b46e756f KS |
338 | NULL, |
339 | }; | |
340 | ||
341 | struct attribute_group khugepaged_attr_group = { | |
342 | .attrs = khugepaged_attr, | |
343 | .name = "khugepaged", | |
344 | }; | |
e1465d12 | 345 | #endif /* CONFIG_SYSFS */ |
b46e756f | 346 | |
b46e756f KS |
347 | int hugepage_madvise(struct vm_area_struct *vma, |
348 | unsigned long *vm_flags, int advice) | |
349 | { | |
350 | switch (advice) { | |
351 | case MADV_HUGEPAGE: | |
352 | #ifdef CONFIG_S390 | |
353 | /* | |
354 | * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390 | |
355 | * can't handle this properly after s390_enable_sie, so we simply | |
356 | * ignore the madvise to prevent qemu from causing a SIGSEGV. | |
357 | */ | |
358 | if (mm_has_pgste(vma->vm_mm)) | |
359 | return 0; | |
360 | #endif | |
361 | *vm_flags &= ~VM_NOHUGEPAGE; | |
362 | *vm_flags |= VM_HUGEPAGE; | |
363 | /* | |
364 | * If the vma become good for khugepaged to scan, | |
365 | * register it here without waiting a page fault that | |
366 | * may not happen any time soon. | |
367 | */ | |
c791576c | 368 | khugepaged_enter_vma(vma, *vm_flags); |
b46e756f KS |
369 | break; |
370 | case MADV_NOHUGEPAGE: | |
371 | *vm_flags &= ~VM_HUGEPAGE; | |
372 | *vm_flags |= VM_NOHUGEPAGE; | |
373 | /* | |
374 | * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning | |
375 | * this vma even if we leave the mm registered in khugepaged if | |
376 | * it got registered before VM_NOHUGEPAGE was set. | |
377 | */ | |
378 | break; | |
379 | } | |
380 | ||
381 | return 0; | |
382 | } | |
383 | ||
384 | int __init khugepaged_init(void) | |
385 | { | |
386 | mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", | |
387 | sizeof(struct mm_slot), | |
388 | __alignof__(struct mm_slot), 0, NULL); | |
389 | if (!mm_slot_cache) | |
390 | return -ENOMEM; | |
391 | ||
392 | khugepaged_pages_to_scan = HPAGE_PMD_NR * 8; | |
393 | khugepaged_max_ptes_none = HPAGE_PMD_NR - 1; | |
394 | khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8; | |
71a2c112 | 395 | khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2; |
b46e756f KS |
396 | |
397 | return 0; | |
398 | } | |
399 | ||
400 | void __init khugepaged_destroy(void) | |
401 | { | |
402 | kmem_cache_destroy(mm_slot_cache); | |
403 | } | |
404 | ||
405 | static inline struct mm_slot *alloc_mm_slot(void) | |
406 | { | |
407 | if (!mm_slot_cache) /* initialization failed */ | |
408 | return NULL; | |
409 | return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); | |
410 | } | |
411 | ||
412 | static inline void free_mm_slot(struct mm_slot *mm_slot) | |
413 | { | |
414 | kmem_cache_free(mm_slot_cache, mm_slot); | |
415 | } | |
416 | ||
417 | static struct mm_slot *get_mm_slot(struct mm_struct *mm) | |
418 | { | |
419 | struct mm_slot *mm_slot; | |
420 | ||
421 | hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm) | |
422 | if (mm == mm_slot->mm) | |
423 | return mm_slot; | |
424 | ||
425 | return NULL; | |
426 | } | |
427 | ||
428 | static void insert_to_mm_slots_hash(struct mm_struct *mm, | |
429 | struct mm_slot *mm_slot) | |
430 | { | |
431 | mm_slot->mm = mm; | |
432 | hash_add(mm_slots_hash, &mm_slot->hash, (long)mm); | |
433 | } | |
434 | ||
435 | static inline int khugepaged_test_exit(struct mm_struct *mm) | |
436 | { | |
4d45e75a | 437 | return atomic_read(&mm->mm_users) == 0; |
b46e756f KS |
438 | } |
439 | ||
2647d11b YS |
440 | bool hugepage_vma_check(struct vm_area_struct *vma, |
441 | unsigned long vm_flags) | |
c2231020 | 442 | { |
e6be37b2 | 443 | if (!transhuge_vma_enabled(vma, vm_flags)) |
c2231020 | 444 | return false; |
99cb0dbd | 445 | |
cb648754 YS |
446 | if (vm_flags & VM_NO_KHUGEPAGED) |
447 | return false; | |
448 | ||
52b52bf1 YS |
449 | /* Don't run khugepaged against DAX vma */ |
450 | if (vma_is_dax(vma)) | |
451 | return false; | |
452 | ||
a4aeaa06 YS |
453 | if (vma->vm_file && !IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - |
454 | vma->vm_pgoff, HPAGE_PMD_NR)) | |
455 | return false; | |
456 | ||
cd89fb06 | 457 | /* Enabled via shmem mount options or sysfs settings. */ |
a4aeaa06 YS |
458 | if (shmem_file(vma->vm_file)) |
459 | return shmem_huge_enabled(vma); | |
cd89fb06 RR |
460 | |
461 | /* THP settings require madvise. */ | |
462 | if (!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) | |
463 | return false; | |
464 | ||
a4aeaa06 | 465 | /* Only regular file is valid */ |
78d12c19 YS |
466 | if (file_thp_enabled(vma)) |
467 | return true; | |
cd89fb06 | 468 | |
25fa414a | 469 | if (!vma->anon_vma || !vma_is_anonymous(vma)) |
c2231020 | 470 | return false; |
222100ee | 471 | if (vma_is_temporary_stack(vma)) |
c2231020 | 472 | return false; |
cb648754 YS |
473 | |
474 | return true; | |
c2231020 YS |
475 | } |
476 | ||
d2081b2b | 477 | void __khugepaged_enter(struct mm_struct *mm) |
b46e756f KS |
478 | { |
479 | struct mm_slot *mm_slot; | |
480 | int wakeup; | |
481 | ||
482 | mm_slot = alloc_mm_slot(); | |
483 | if (!mm_slot) | |
d2081b2b | 484 | return; |
b46e756f KS |
485 | |
486 | /* __khugepaged_exit() must not run from under us */ | |
28ff0a3c | 487 | VM_BUG_ON_MM(khugepaged_test_exit(mm), mm); |
b46e756f KS |
488 | if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { |
489 | free_mm_slot(mm_slot); | |
d2081b2b | 490 | return; |
b46e756f KS |
491 | } |
492 | ||
493 | spin_lock(&khugepaged_mm_lock); | |
494 | insert_to_mm_slots_hash(mm, mm_slot); | |
495 | /* | |
496 | * Insert just behind the scanning cursor, to let the area settle | |
497 | * down a little. | |
498 | */ | |
499 | wakeup = list_empty(&khugepaged_scan.mm_head); | |
500 | list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head); | |
501 | spin_unlock(&khugepaged_mm_lock); | |
502 | ||
f1f10076 | 503 | mmgrab(mm); |
b46e756f KS |
504 | if (wakeup) |
505 | wake_up_interruptible(&khugepaged_wait); | |
b46e756f KS |
506 | } |
507 | ||
c791576c YS |
508 | void khugepaged_enter_vma(struct vm_area_struct *vma, |
509 | unsigned long vm_flags) | |
b46e756f | 510 | { |
2647d11b YS |
511 | if (!test_bit(MMF_VM_HUGEPAGE, &vma->vm_mm->flags) && |
512 | khugepaged_enabled() && | |
513 | (((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) < | |
514 | (vma->vm_end & HPAGE_PMD_MASK))) { | |
515 | if (hugepage_vma_check(vma, vm_flags)) | |
516 | __khugepaged_enter(vma->vm_mm); | |
517 | } | |
b46e756f KS |
518 | } |
519 | ||
520 | void __khugepaged_exit(struct mm_struct *mm) | |
521 | { | |
522 | struct mm_slot *mm_slot; | |
523 | int free = 0; | |
524 | ||
525 | spin_lock(&khugepaged_mm_lock); | |
526 | mm_slot = get_mm_slot(mm); | |
527 | if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { | |
528 | hash_del(&mm_slot->hash); | |
529 | list_del(&mm_slot->mm_node); | |
530 | free = 1; | |
531 | } | |
532 | spin_unlock(&khugepaged_mm_lock); | |
533 | ||
534 | if (free) { | |
535 | clear_bit(MMF_VM_HUGEPAGE, &mm->flags); | |
536 | free_mm_slot(mm_slot); | |
537 | mmdrop(mm); | |
538 | } else if (mm_slot) { | |
539 | /* | |
540 | * This is required to serialize against | |
541 | * khugepaged_test_exit() (which is guaranteed to run | |
542 | * under mmap sem read mode). Stop here (after we | |
543 | * return all pagetables will be destroyed) until | |
544 | * khugepaged has finished working on the pagetables | |
c1e8d7c6 | 545 | * under the mmap_lock. |
b46e756f | 546 | */ |
d8ed45c5 ML |
547 | mmap_write_lock(mm); |
548 | mmap_write_unlock(mm); | |
b46e756f KS |
549 | } |
550 | } | |
551 | ||
552 | static void release_pte_page(struct page *page) | |
553 | { | |
5503fbf2 KS |
554 | mod_node_page_state(page_pgdat(page), |
555 | NR_ISOLATED_ANON + page_is_file_lru(page), | |
556 | -compound_nr(page)); | |
b46e756f KS |
557 | unlock_page(page); |
558 | putback_lru_page(page); | |
559 | } | |
560 | ||
5503fbf2 KS |
561 | static void release_pte_pages(pte_t *pte, pte_t *_pte, |
562 | struct list_head *compound_pagelist) | |
b46e756f | 563 | { |
5503fbf2 KS |
564 | struct page *page, *tmp; |
565 | ||
b46e756f KS |
566 | while (--_pte >= pte) { |
567 | pte_t pteval = *_pte; | |
5503fbf2 KS |
568 | |
569 | page = pte_page(pteval); | |
570 | if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) && | |
571 | !PageCompound(page)) | |
572 | release_pte_page(page); | |
573 | } | |
574 | ||
575 | list_for_each_entry_safe(page, tmp, compound_pagelist, lru) { | |
576 | list_del(&page->lru); | |
577 | release_pte_page(page); | |
b46e756f KS |
578 | } |
579 | } | |
580 | ||
9445689f KS |
581 | static bool is_refcount_suitable(struct page *page) |
582 | { | |
583 | int expected_refcount; | |
584 | ||
585 | expected_refcount = total_mapcount(page); | |
586 | if (PageSwapCache(page)) | |
587 | expected_refcount += compound_nr(page); | |
588 | ||
589 | return page_count(page) == expected_refcount; | |
590 | } | |
591 | ||
b46e756f KS |
592 | static int __collapse_huge_page_isolate(struct vm_area_struct *vma, |
593 | unsigned long address, | |
5503fbf2 KS |
594 | pte_t *pte, |
595 | struct list_head *compound_pagelist) | |
b46e756f KS |
596 | { |
597 | struct page *page = NULL; | |
598 | pte_t *_pte; | |
71a2c112 | 599 | int none_or_zero = 0, shared = 0, result = 0, referenced = 0; |
0db501f7 | 600 | bool writable = false; |
b46e756f | 601 | |
36ee2c78 | 602 | for (_pte = pte; _pte < pte + HPAGE_PMD_NR; |
b46e756f KS |
603 | _pte++, address += PAGE_SIZE) { |
604 | pte_t pteval = *_pte; | |
605 | if (pte_none(pteval) || (pte_present(pteval) && | |
606 | is_zero_pfn(pte_pfn(pteval)))) { | |
607 | if (!userfaultfd_armed(vma) && | |
608 | ++none_or_zero <= khugepaged_max_ptes_none) { | |
609 | continue; | |
610 | } else { | |
611 | result = SCAN_EXCEED_NONE_PTE; | |
e9ea874a | 612 | count_vm_event(THP_SCAN_EXCEED_NONE_PTE); |
b46e756f KS |
613 | goto out; |
614 | } | |
615 | } | |
616 | if (!pte_present(pteval)) { | |
617 | result = SCAN_PTE_NON_PRESENT; | |
618 | goto out; | |
619 | } | |
620 | page = vm_normal_page(vma, address, pteval); | |
621 | if (unlikely(!page)) { | |
622 | result = SCAN_PAGE_NULL; | |
623 | goto out; | |
624 | } | |
625 | ||
5503fbf2 KS |
626 | VM_BUG_ON_PAGE(!PageAnon(page), page); |
627 | ||
71a2c112 KS |
628 | if (page_mapcount(page) > 1 && |
629 | ++shared > khugepaged_max_ptes_shared) { | |
630 | result = SCAN_EXCEED_SHARED_PTE; | |
e9ea874a | 631 | count_vm_event(THP_SCAN_EXCEED_SHARED_PTE); |
71a2c112 KS |
632 | goto out; |
633 | } | |
634 | ||
fece2029 | 635 | if (PageCompound(page)) { |
5503fbf2 KS |
636 | struct page *p; |
637 | page = compound_head(page); | |
fece2029 | 638 | |
5503fbf2 KS |
639 | /* |
640 | * Check if we have dealt with the compound page | |
641 | * already | |
642 | */ | |
643 | list_for_each_entry(p, compound_pagelist, lru) { | |
644 | if (page == p) | |
645 | goto next; | |
646 | } | |
647 | } | |
b46e756f KS |
648 | |
649 | /* | |
650 | * We can do it before isolate_lru_page because the | |
651 | * page can't be freed from under us. NOTE: PG_lock | |
652 | * is needed to serialize against split_huge_page | |
653 | * when invoked from the VM. | |
654 | */ | |
655 | if (!trylock_page(page)) { | |
656 | result = SCAN_PAGE_LOCK; | |
657 | goto out; | |
658 | } | |
659 | ||
660 | /* | |
9445689f KS |
661 | * Check if the page has any GUP (or other external) pins. |
662 | * | |
663 | * The page table that maps the page has been already unlinked | |
664 | * from the page table tree and this process cannot get | |
f0953a1b | 665 | * an additional pin on the page. |
9445689f KS |
666 | * |
667 | * New pins can come later if the page is shared across fork, | |
668 | * but not from this process. The other process cannot write to | |
669 | * the page, only trigger CoW. | |
b46e756f | 670 | */ |
9445689f | 671 | if (!is_refcount_suitable(page)) { |
b46e756f KS |
672 | unlock_page(page); |
673 | result = SCAN_PAGE_COUNT; | |
674 | goto out; | |
675 | } | |
b46e756f KS |
676 | |
677 | /* | |
678 | * Isolate the page to avoid collapsing an hugepage | |
679 | * currently in use by the VM. | |
680 | */ | |
681 | if (isolate_lru_page(page)) { | |
682 | unlock_page(page); | |
683 | result = SCAN_DEL_PAGE_LRU; | |
684 | goto out; | |
685 | } | |
5503fbf2 KS |
686 | mod_node_page_state(page_pgdat(page), |
687 | NR_ISOLATED_ANON + page_is_file_lru(page), | |
688 | compound_nr(page)); | |
b46e756f KS |
689 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
690 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
691 | ||
5503fbf2 KS |
692 | if (PageCompound(page)) |
693 | list_add_tail(&page->lru, compound_pagelist); | |
694 | next: | |
0db501f7 | 695 | /* There should be enough young pte to collapse the page */ |
b46e756f KS |
696 | if (pte_young(pteval) || |
697 | page_is_young(page) || PageReferenced(page) || | |
698 | mmu_notifier_test_young(vma->vm_mm, address)) | |
0db501f7 | 699 | referenced++; |
5503fbf2 KS |
700 | |
701 | if (pte_write(pteval)) | |
702 | writable = true; | |
b46e756f | 703 | } |
74e579bf ML |
704 | |
705 | if (unlikely(!writable)) { | |
b46e756f | 706 | result = SCAN_PAGE_RO; |
74e579bf ML |
707 | } else if (unlikely(!referenced)) { |
708 | result = SCAN_LACK_REFERENCED_PAGE; | |
709 | } else { | |
710 | result = SCAN_SUCCEED; | |
711 | trace_mm_collapse_huge_page_isolate(page, none_or_zero, | |
712 | referenced, writable, result); | |
713 | return 1; | |
b46e756f | 714 | } |
b46e756f | 715 | out: |
5503fbf2 | 716 | release_pte_pages(pte, _pte, compound_pagelist); |
b46e756f KS |
717 | trace_mm_collapse_huge_page_isolate(page, none_or_zero, |
718 | referenced, writable, result); | |
719 | return 0; | |
720 | } | |
721 | ||
722 | static void __collapse_huge_page_copy(pte_t *pte, struct page *page, | |
723 | struct vm_area_struct *vma, | |
724 | unsigned long address, | |
5503fbf2 KS |
725 | spinlock_t *ptl, |
726 | struct list_head *compound_pagelist) | |
b46e756f | 727 | { |
5503fbf2 | 728 | struct page *src_page, *tmp; |
b46e756f | 729 | pte_t *_pte; |
338a16ba DR |
730 | for (_pte = pte; _pte < pte + HPAGE_PMD_NR; |
731 | _pte++, page++, address += PAGE_SIZE) { | |
b46e756f | 732 | pte_t pteval = *_pte; |
b46e756f KS |
733 | |
734 | if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { | |
735 | clear_user_highpage(page, address); | |
736 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); | |
737 | if (is_zero_pfn(pte_pfn(pteval))) { | |
738 | /* | |
739 | * ptl mostly unnecessary. | |
740 | */ | |
741 | spin_lock(ptl); | |
08d5b29e | 742 | ptep_clear(vma->vm_mm, address, _pte); |
b46e756f KS |
743 | spin_unlock(ptl); |
744 | } | |
745 | } else { | |
746 | src_page = pte_page(pteval); | |
747 | copy_user_highpage(page, src_page, address, vma); | |
5503fbf2 KS |
748 | if (!PageCompound(src_page)) |
749 | release_pte_page(src_page); | |
b46e756f KS |
750 | /* |
751 | * ptl mostly unnecessary, but preempt has to | |
752 | * be disabled to update the per-cpu stats | |
753 | * inside page_remove_rmap(). | |
754 | */ | |
755 | spin_lock(ptl); | |
08d5b29e | 756 | ptep_clear(vma->vm_mm, address, _pte); |
cea86fe2 | 757 | page_remove_rmap(src_page, vma, false); |
b46e756f KS |
758 | spin_unlock(ptl); |
759 | free_page_and_swap_cache(src_page); | |
760 | } | |
b46e756f | 761 | } |
5503fbf2 KS |
762 | |
763 | list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) { | |
764 | list_del(&src_page->lru); | |
765 | release_pte_page(src_page); | |
766 | } | |
b46e756f KS |
767 | } |
768 | ||
769 | static void khugepaged_alloc_sleep(void) | |
770 | { | |
771 | DEFINE_WAIT(wait); | |
772 | ||
773 | add_wait_queue(&khugepaged_wait, &wait); | |
774 | freezable_schedule_timeout_interruptible( | |
775 | msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); | |
776 | remove_wait_queue(&khugepaged_wait, &wait); | |
777 | } | |
778 | ||
779 | static int khugepaged_node_load[MAX_NUMNODES]; | |
780 | ||
781 | static bool khugepaged_scan_abort(int nid) | |
782 | { | |
783 | int i; | |
784 | ||
785 | /* | |
a5f5f91d | 786 | * If node_reclaim_mode is disabled, then no extra effort is made to |
b46e756f KS |
787 | * allocate memory locally. |
788 | */ | |
202e35db | 789 | if (!node_reclaim_enabled()) |
b46e756f KS |
790 | return false; |
791 | ||
792 | /* If there is a count for this node already, it must be acceptable */ | |
793 | if (khugepaged_node_load[nid]) | |
794 | return false; | |
795 | ||
796 | for (i = 0; i < MAX_NUMNODES; i++) { | |
797 | if (!khugepaged_node_load[i]) | |
798 | continue; | |
a55c7454 | 799 | if (node_distance(nid, i) > node_reclaim_distance) |
b46e756f KS |
800 | return true; |
801 | } | |
802 | return false; | |
803 | } | |
804 | ||
805 | /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */ | |
806 | static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void) | |
807 | { | |
25160354 | 808 | return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT; |
b46e756f KS |
809 | } |
810 | ||
811 | #ifdef CONFIG_NUMA | |
812 | static int khugepaged_find_target_node(void) | |
813 | { | |
814 | static int last_khugepaged_target_node = NUMA_NO_NODE; | |
815 | int nid, target_node = 0, max_value = 0; | |
816 | ||
817 | /* find first node with max normal pages hit */ | |
818 | for (nid = 0; nid < MAX_NUMNODES; nid++) | |
819 | if (khugepaged_node_load[nid] > max_value) { | |
820 | max_value = khugepaged_node_load[nid]; | |
821 | target_node = nid; | |
822 | } | |
823 | ||
824 | /* do some balance if several nodes have the same hit record */ | |
825 | if (target_node <= last_khugepaged_target_node) | |
826 | for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES; | |
827 | nid++) | |
828 | if (max_value == khugepaged_node_load[nid]) { | |
829 | target_node = nid; | |
830 | break; | |
831 | } | |
832 | ||
833 | last_khugepaged_target_node = target_node; | |
834 | return target_node; | |
835 | } | |
836 | ||
837 | static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) | |
838 | { | |
839 | if (IS_ERR(*hpage)) { | |
840 | if (!*wait) | |
841 | return false; | |
842 | ||
843 | *wait = false; | |
844 | *hpage = NULL; | |
845 | khugepaged_alloc_sleep(); | |
846 | } else if (*hpage) { | |
847 | put_page(*hpage); | |
848 | *hpage = NULL; | |
849 | } | |
850 | ||
851 | return true; | |
852 | } | |
853 | ||
854 | static struct page * | |
988ddb71 | 855 | khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node) |
b46e756f KS |
856 | { |
857 | VM_BUG_ON_PAGE(*hpage, *hpage); | |
858 | ||
b46e756f KS |
859 | *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER); |
860 | if (unlikely(!*hpage)) { | |
861 | count_vm_event(THP_COLLAPSE_ALLOC_FAILED); | |
862 | *hpage = ERR_PTR(-ENOMEM); | |
863 | return NULL; | |
864 | } | |
865 | ||
866 | prep_transhuge_page(*hpage); | |
867 | count_vm_event(THP_COLLAPSE_ALLOC); | |
868 | return *hpage; | |
869 | } | |
870 | #else | |
871 | static int khugepaged_find_target_node(void) | |
872 | { | |
873 | return 0; | |
874 | } | |
875 | ||
876 | static inline struct page *alloc_khugepaged_hugepage(void) | |
877 | { | |
878 | struct page *page; | |
879 | ||
880 | page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(), | |
881 | HPAGE_PMD_ORDER); | |
882 | if (page) | |
883 | prep_transhuge_page(page); | |
884 | return page; | |
885 | } | |
886 | ||
887 | static struct page *khugepaged_alloc_hugepage(bool *wait) | |
888 | { | |
889 | struct page *hpage; | |
890 | ||
891 | do { | |
892 | hpage = alloc_khugepaged_hugepage(); | |
893 | if (!hpage) { | |
894 | count_vm_event(THP_COLLAPSE_ALLOC_FAILED); | |
895 | if (!*wait) | |
896 | return NULL; | |
897 | ||
898 | *wait = false; | |
899 | khugepaged_alloc_sleep(); | |
900 | } else | |
901 | count_vm_event(THP_COLLAPSE_ALLOC); | |
902 | } while (unlikely(!hpage) && likely(khugepaged_enabled())); | |
903 | ||
904 | return hpage; | |
905 | } | |
906 | ||
907 | static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) | |
908 | { | |
033b5d77 HD |
909 | /* |
910 | * If the hpage allocated earlier was briefly exposed in page cache | |
911 | * before collapse_file() failed, it is possible that racing lookups | |
912 | * have not yet completed, and would then be unpleasantly surprised by | |
913 | * finding the hpage reused for the same mapping at a different offset. | |
914 | * Just release the previous allocation if there is any danger of that. | |
915 | */ | |
916 | if (*hpage && page_count(*hpage) > 1) { | |
917 | put_page(*hpage); | |
918 | *hpage = NULL; | |
919 | } | |
920 | ||
b46e756f KS |
921 | if (!*hpage) |
922 | *hpage = khugepaged_alloc_hugepage(wait); | |
923 | ||
924 | if (unlikely(!*hpage)) | |
925 | return false; | |
926 | ||
927 | return true; | |
928 | } | |
929 | ||
930 | static struct page * | |
988ddb71 | 931 | khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node) |
b46e756f | 932 | { |
b46e756f KS |
933 | VM_BUG_ON(!*hpage); |
934 | ||
935 | return *hpage; | |
936 | } | |
937 | #endif | |
938 | ||
b46e756f | 939 | /* |
c1e8d7c6 ML |
940 | * If mmap_lock temporarily dropped, revalidate vma |
941 | * before taking mmap_lock. | |
b46e756f KS |
942 | * Return 0 if succeeds, otherwise return none-zero |
943 | * value (scan code). | |
944 | */ | |
945 | ||
c131f751 KS |
946 | static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address, |
947 | struct vm_area_struct **vmap) | |
b46e756f KS |
948 | { |
949 | struct vm_area_struct *vma; | |
950 | unsigned long hstart, hend; | |
951 | ||
952 | if (unlikely(khugepaged_test_exit(mm))) | |
953 | return SCAN_ANY_PROCESS; | |
954 | ||
c131f751 | 955 | *vmap = vma = find_vma(mm, address); |
b46e756f KS |
956 | if (!vma) |
957 | return SCAN_VMA_NULL; | |
958 | ||
959 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; | |
960 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
961 | if (address < hstart || address + HPAGE_PMD_SIZE > hend) | |
962 | return SCAN_ADDRESS_RANGE; | |
50f8b92f | 963 | if (!hugepage_vma_check(vma, vma->vm_flags)) |
b46e756f | 964 | return SCAN_VMA_CHECK; |
594cced1 | 965 | /* Anon VMA expected */ |
25fa414a | 966 | if (!vma->anon_vma || !vma_is_anonymous(vma)) |
594cced1 | 967 | return SCAN_VMA_CHECK; |
b46e756f KS |
968 | return 0; |
969 | } | |
970 | ||
971 | /* | |
972 | * Bring missing pages in from swap, to complete THP collapse. | |
973 | * Only done if khugepaged_scan_pmd believes it is worthwhile. | |
974 | * | |
4d928e20 ML |
975 | * Called and returns without pte mapped or spinlocks held. |
976 | * Note that if false is returned, mmap_lock will be released. | |
b46e756f KS |
977 | */ |
978 | ||
979 | static bool __collapse_huge_page_swapin(struct mm_struct *mm, | |
980 | struct vm_area_struct *vma, | |
2b635dd3 | 981 | unsigned long haddr, pmd_t *pmd, |
0db501f7 | 982 | int referenced) |
b46e756f | 983 | { |
2b740303 SJ |
984 | int swapped_in = 0; |
985 | vm_fault_t ret = 0; | |
2b635dd3 WD |
986 | unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE); |
987 | ||
988 | for (address = haddr; address < end; address += PAGE_SIZE) { | |
989 | struct vm_fault vmf = { | |
990 | .vma = vma, | |
991 | .address = address, | |
992 | .pgoff = linear_page_index(vma, haddr), | |
993 | .flags = FAULT_FLAG_ALLOW_RETRY, | |
994 | .pmd = pmd, | |
995 | }; | |
996 | ||
997 | vmf.pte = pte_offset_map(pmd, address); | |
2994302b | 998 | vmf.orig_pte = *vmf.pte; |
2b635dd3 WD |
999 | if (!is_swap_pte(vmf.orig_pte)) { |
1000 | pte_unmap(vmf.pte); | |
b46e756f | 1001 | continue; |
2b635dd3 | 1002 | } |
2994302b | 1003 | ret = do_swap_page(&vmf); |
0db501f7 | 1004 | |
4d928e20 ML |
1005 | /* |
1006 | * do_swap_page returns VM_FAULT_RETRY with released mmap_lock. | |
1007 | * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because | |
1008 | * we do not retry here and swap entry will remain in pagetable | |
1009 | * resulting in later failure. | |
1010 | */ | |
b46e756f | 1011 | if (ret & VM_FAULT_RETRY) { |
4d928e20 ML |
1012 | trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); |
1013 | return false; | |
b46e756f KS |
1014 | } |
1015 | if (ret & VM_FAULT_ERROR) { | |
4d928e20 | 1016 | mmap_read_unlock(mm); |
0db501f7 | 1017 | trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); |
b46e756f KS |
1018 | return false; |
1019 | } | |
4d928e20 | 1020 | swapped_in++; |
b46e756f | 1021 | } |
ae2c5d80 KS |
1022 | |
1023 | /* Drain LRU add pagevec to remove extra pin on the swapped in pages */ | |
1024 | if (swapped_in) | |
1025 | lru_add_drain(); | |
1026 | ||
0db501f7 | 1027 | trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1); |
b46e756f KS |
1028 | return true; |
1029 | } | |
1030 | ||
1031 | static void collapse_huge_page(struct mm_struct *mm, | |
1032 | unsigned long address, | |
1033 | struct page **hpage, | |
ffe945e6 | 1034 | int node, int referenced, int unmapped) |
b46e756f | 1035 | { |
5503fbf2 | 1036 | LIST_HEAD(compound_pagelist); |
b46e756f KS |
1037 | pmd_t *pmd, _pmd; |
1038 | pte_t *pte; | |
1039 | pgtable_t pgtable; | |
1040 | struct page *new_page; | |
1041 | spinlock_t *pmd_ptl, *pte_ptl; | |
1042 | int isolated = 0, result = 0; | |
c131f751 | 1043 | struct vm_area_struct *vma; |
ac46d4f3 | 1044 | struct mmu_notifier_range range; |
b46e756f KS |
1045 | gfp_t gfp; |
1046 | ||
1047 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
1048 | ||
1049 | /* Only allocate from the target node */ | |
41b6167e | 1050 | gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE; |
b46e756f | 1051 | |
988ddb71 | 1052 | /* |
c1e8d7c6 | 1053 | * Before allocating the hugepage, release the mmap_lock read lock. |
988ddb71 | 1054 | * The allocation can take potentially a long time if it involves |
c1e8d7c6 | 1055 | * sync compaction, and we do not need to hold the mmap_lock during |
988ddb71 KS |
1056 | * that. We will recheck the vma after taking it again in write mode. |
1057 | */ | |
d8ed45c5 | 1058 | mmap_read_unlock(mm); |
988ddb71 | 1059 | new_page = khugepaged_alloc_page(hpage, gfp, node); |
b46e756f KS |
1060 | if (!new_page) { |
1061 | result = SCAN_ALLOC_HUGE_PAGE_FAIL; | |
1062 | goto out_nolock; | |
1063 | } | |
1064 | ||
8f425e4e | 1065 | if (unlikely(mem_cgroup_charge(page_folio(new_page), mm, gfp))) { |
b46e756f KS |
1066 | result = SCAN_CGROUP_CHARGE_FAIL; |
1067 | goto out_nolock; | |
1068 | } | |
9d82c694 | 1069 | count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC); |
b46e756f | 1070 | |
d8ed45c5 | 1071 | mmap_read_lock(mm); |
c131f751 | 1072 | result = hugepage_vma_revalidate(mm, address, &vma); |
b46e756f | 1073 | if (result) { |
d8ed45c5 | 1074 | mmap_read_unlock(mm); |
b46e756f KS |
1075 | goto out_nolock; |
1076 | } | |
1077 | ||
1078 | pmd = mm_find_pmd(mm, address); | |
1079 | if (!pmd) { | |
1080 | result = SCAN_PMD_NULL; | |
d8ed45c5 | 1081 | mmap_read_unlock(mm); |
b46e756f KS |
1082 | goto out_nolock; |
1083 | } | |
1084 | ||
1085 | /* | |
4d928e20 ML |
1086 | * __collapse_huge_page_swapin will return with mmap_lock released |
1087 | * when it fails. So we jump out_nolock directly in that case. | |
b46e756f KS |
1088 | * Continuing to collapse causes inconsistency. |
1089 | */ | |
ffe945e6 KS |
1090 | if (unmapped && !__collapse_huge_page_swapin(mm, vma, address, |
1091 | pmd, referenced)) { | |
b46e756f KS |
1092 | goto out_nolock; |
1093 | } | |
1094 | ||
d8ed45c5 | 1095 | mmap_read_unlock(mm); |
b46e756f KS |
1096 | /* |
1097 | * Prevent all access to pagetables with the exception of | |
1098 | * gup_fast later handled by the ptep_clear_flush and the VM | |
1099 | * handled by the anon_vma lock + PG_lock. | |
1100 | */ | |
d8ed45c5 | 1101 | mmap_write_lock(mm); |
c131f751 | 1102 | result = hugepage_vma_revalidate(mm, address, &vma); |
b46e756f | 1103 | if (result) |
18d24a7c | 1104 | goto out_up_write; |
b46e756f KS |
1105 | /* check if the pmd is still valid */ |
1106 | if (mm_find_pmd(mm, address) != pmd) | |
18d24a7c | 1107 | goto out_up_write; |
b46e756f KS |
1108 | |
1109 | anon_vma_lock_write(vma->anon_vma); | |
1110 | ||
7269f999 | 1111 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm, |
6f4f13e8 | 1112 | address, address + HPAGE_PMD_SIZE); |
ac46d4f3 | 1113 | mmu_notifier_invalidate_range_start(&range); |
ec649c9d VS |
1114 | |
1115 | pte = pte_offset_map(pmd, address); | |
1116 | pte_ptl = pte_lockptr(mm, pmd); | |
1117 | ||
b46e756f KS |
1118 | pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ |
1119 | /* | |
1120 | * After this gup_fast can't run anymore. This also removes | |
1121 | * any huge TLB entry from the CPU so we won't allow | |
1122 | * huge and small TLB entries for the same virtual address | |
1123 | * to avoid the risk of CPU bugs in that area. | |
1124 | */ | |
1125 | _pmd = pmdp_collapse_flush(vma, address, pmd); | |
1126 | spin_unlock(pmd_ptl); | |
ac46d4f3 | 1127 | mmu_notifier_invalidate_range_end(&range); |
b46e756f KS |
1128 | |
1129 | spin_lock(pte_ptl); | |
5503fbf2 KS |
1130 | isolated = __collapse_huge_page_isolate(vma, address, pte, |
1131 | &compound_pagelist); | |
b46e756f KS |
1132 | spin_unlock(pte_ptl); |
1133 | ||
1134 | if (unlikely(!isolated)) { | |
1135 | pte_unmap(pte); | |
1136 | spin_lock(pmd_ptl); | |
1137 | BUG_ON(!pmd_none(*pmd)); | |
1138 | /* | |
1139 | * We can only use set_pmd_at when establishing | |
1140 | * hugepmds and never for establishing regular pmds that | |
1141 | * points to regular pagetables. Use pmd_populate for that | |
1142 | */ | |
1143 | pmd_populate(mm, pmd, pmd_pgtable(_pmd)); | |
1144 | spin_unlock(pmd_ptl); | |
1145 | anon_vma_unlock_write(vma->anon_vma); | |
1146 | result = SCAN_FAIL; | |
18d24a7c | 1147 | goto out_up_write; |
b46e756f KS |
1148 | } |
1149 | ||
1150 | /* | |
1151 | * All pages are isolated and locked so anon_vma rmap | |
1152 | * can't run anymore. | |
1153 | */ | |
1154 | anon_vma_unlock_write(vma->anon_vma); | |
1155 | ||
5503fbf2 KS |
1156 | __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl, |
1157 | &compound_pagelist); | |
b46e756f | 1158 | pte_unmap(pte); |
588d01f9 ML |
1159 | /* |
1160 | * spin_lock() below is not the equivalent of smp_wmb(), but | |
1161 | * the smp_wmb() inside __SetPageUptodate() can be reused to | |
1162 | * avoid the copy_huge_page writes to become visible after | |
1163 | * the set_pmd_at() write. | |
1164 | */ | |
b46e756f KS |
1165 | __SetPageUptodate(new_page); |
1166 | pgtable = pmd_pgtable(_pmd); | |
1167 | ||
1168 | _pmd = mk_huge_pmd(new_page, vma->vm_page_prot); | |
f55e1014 | 1169 | _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); |
b46e756f | 1170 | |
b46e756f KS |
1171 | spin_lock(pmd_ptl); |
1172 | BUG_ON(!pmd_none(*pmd)); | |
40f2bbf7 | 1173 | page_add_new_anon_rmap(new_page, vma, address); |
b518154e | 1174 | lru_cache_add_inactive_or_unevictable(new_page, vma); |
b46e756f KS |
1175 | pgtable_trans_huge_deposit(mm, pmd, pgtable); |
1176 | set_pmd_at(mm, address, pmd, _pmd); | |
1177 | update_mmu_cache_pmd(vma, address, pmd); | |
1178 | spin_unlock(pmd_ptl); | |
1179 | ||
1180 | *hpage = NULL; | |
1181 | ||
1182 | khugepaged_pages_collapsed++; | |
1183 | result = SCAN_SUCCEED; | |
1184 | out_up_write: | |
d8ed45c5 | 1185 | mmap_write_unlock(mm); |
b46e756f | 1186 | out_nolock: |
9d82c694 | 1187 | if (!IS_ERR_OR_NULL(*hpage)) |
bbc6b703 | 1188 | mem_cgroup_uncharge(page_folio(*hpage)); |
b46e756f KS |
1189 | trace_mm_collapse_huge_page(mm, isolated, result); |
1190 | return; | |
b46e756f KS |
1191 | } |
1192 | ||
1193 | static int khugepaged_scan_pmd(struct mm_struct *mm, | |
1194 | struct vm_area_struct *vma, | |
1195 | unsigned long address, | |
1196 | struct page **hpage) | |
1197 | { | |
1198 | pmd_t *pmd; | |
1199 | pte_t *pte, *_pte; | |
71a2c112 KS |
1200 | int ret = 0, result = 0, referenced = 0; |
1201 | int none_or_zero = 0, shared = 0; | |
b46e756f KS |
1202 | struct page *page = NULL; |
1203 | unsigned long _address; | |
1204 | spinlock_t *ptl; | |
1205 | int node = NUMA_NO_NODE, unmapped = 0; | |
0db501f7 | 1206 | bool writable = false; |
b46e756f KS |
1207 | |
1208 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
1209 | ||
1210 | pmd = mm_find_pmd(mm, address); | |
1211 | if (!pmd) { | |
1212 | result = SCAN_PMD_NULL; | |
1213 | goto out; | |
1214 | } | |
1215 | ||
1216 | memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load)); | |
1217 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); | |
36ee2c78 | 1218 | for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR; |
b46e756f KS |
1219 | _pte++, _address += PAGE_SIZE) { |
1220 | pte_t pteval = *_pte; | |
1221 | if (is_swap_pte(pteval)) { | |
1222 | if (++unmapped <= khugepaged_max_ptes_swap) { | |
e1e267c7 PX |
1223 | /* |
1224 | * Always be strict with uffd-wp | |
1225 | * enabled swap entries. Please see | |
1226 | * comment below for pte_uffd_wp(). | |
1227 | */ | |
1228 | if (pte_swp_uffd_wp(pteval)) { | |
1229 | result = SCAN_PTE_UFFD_WP; | |
1230 | goto out_unmap; | |
1231 | } | |
b46e756f KS |
1232 | continue; |
1233 | } else { | |
1234 | result = SCAN_EXCEED_SWAP_PTE; | |
e9ea874a | 1235 | count_vm_event(THP_SCAN_EXCEED_SWAP_PTE); |
b46e756f KS |
1236 | goto out_unmap; |
1237 | } | |
1238 | } | |
1239 | if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { | |
1240 | if (!userfaultfd_armed(vma) && | |
1241 | ++none_or_zero <= khugepaged_max_ptes_none) { | |
1242 | continue; | |
1243 | } else { | |
1244 | result = SCAN_EXCEED_NONE_PTE; | |
e9ea874a | 1245 | count_vm_event(THP_SCAN_EXCEED_NONE_PTE); |
b46e756f KS |
1246 | goto out_unmap; |
1247 | } | |
1248 | } | |
e1e267c7 PX |
1249 | if (pte_uffd_wp(pteval)) { |
1250 | /* | |
1251 | * Don't collapse the page if any of the small | |
1252 | * PTEs are armed with uffd write protection. | |
1253 | * Here we can also mark the new huge pmd as | |
1254 | * write protected if any of the small ones is | |
8958b249 | 1255 | * marked but that could bring unknown |
e1e267c7 PX |
1256 | * userfault messages that falls outside of |
1257 | * the registered range. So, just be simple. | |
1258 | */ | |
1259 | result = SCAN_PTE_UFFD_WP; | |
1260 | goto out_unmap; | |
1261 | } | |
b46e756f KS |
1262 | if (pte_write(pteval)) |
1263 | writable = true; | |
1264 | ||
1265 | page = vm_normal_page(vma, _address, pteval); | |
1266 | if (unlikely(!page)) { | |
1267 | result = SCAN_PAGE_NULL; | |
1268 | goto out_unmap; | |
1269 | } | |
1270 | ||
71a2c112 KS |
1271 | if (page_mapcount(page) > 1 && |
1272 | ++shared > khugepaged_max_ptes_shared) { | |
1273 | result = SCAN_EXCEED_SHARED_PTE; | |
e9ea874a | 1274 | count_vm_event(THP_SCAN_EXCEED_SHARED_PTE); |
71a2c112 KS |
1275 | goto out_unmap; |
1276 | } | |
1277 | ||
5503fbf2 | 1278 | page = compound_head(page); |
b46e756f KS |
1279 | |
1280 | /* | |
1281 | * Record which node the original page is from and save this | |
1282 | * information to khugepaged_node_load[]. | |
0b8f0d87 | 1283 | * Khugepaged will allocate hugepage from the node has the max |
b46e756f KS |
1284 | * hit record. |
1285 | */ | |
1286 | node = page_to_nid(page); | |
1287 | if (khugepaged_scan_abort(node)) { | |
1288 | result = SCAN_SCAN_ABORT; | |
1289 | goto out_unmap; | |
1290 | } | |
1291 | khugepaged_node_load[node]++; | |
1292 | if (!PageLRU(page)) { | |
1293 | result = SCAN_PAGE_LRU; | |
1294 | goto out_unmap; | |
1295 | } | |
1296 | if (PageLocked(page)) { | |
1297 | result = SCAN_PAGE_LOCK; | |
1298 | goto out_unmap; | |
1299 | } | |
1300 | if (!PageAnon(page)) { | |
1301 | result = SCAN_PAGE_ANON; | |
1302 | goto out_unmap; | |
1303 | } | |
1304 | ||
1305 | /* | |
9445689f KS |
1306 | * Check if the page has any GUP (or other external) pins. |
1307 | * | |
36ee2c78 | 1308 | * Here the check is racy it may see total_mapcount > refcount |
9445689f KS |
1309 | * in some cases. |
1310 | * For example, one process with one forked child process. | |
1311 | * The parent has the PMD split due to MADV_DONTNEED, then | |
1312 | * the child is trying unmap the whole PMD, but khugepaged | |
1313 | * may be scanning the parent between the child has | |
1314 | * PageDoubleMap flag cleared and dec the mapcount. So | |
1315 | * khugepaged may see total_mapcount > refcount. | |
1316 | * | |
1317 | * But such case is ephemeral we could always retry collapse | |
1318 | * later. However it may report false positive if the page | |
1319 | * has excessive GUP pins (i.e. 512). Anyway the same check | |
1320 | * will be done again later the risk seems low. | |
b46e756f | 1321 | */ |
9445689f | 1322 | if (!is_refcount_suitable(page)) { |
b46e756f KS |
1323 | result = SCAN_PAGE_COUNT; |
1324 | goto out_unmap; | |
1325 | } | |
1326 | if (pte_young(pteval) || | |
1327 | page_is_young(page) || PageReferenced(page) || | |
1328 | mmu_notifier_test_young(vma->vm_mm, address)) | |
0db501f7 | 1329 | referenced++; |
b46e756f | 1330 | } |
ffe945e6 | 1331 | if (!writable) { |
b46e756f | 1332 | result = SCAN_PAGE_RO; |
ffe945e6 KS |
1333 | } else if (!referenced || (unmapped && referenced < HPAGE_PMD_NR/2)) { |
1334 | result = SCAN_LACK_REFERENCED_PAGE; | |
1335 | } else { | |
1336 | result = SCAN_SUCCEED; | |
1337 | ret = 1; | |
b46e756f KS |
1338 | } |
1339 | out_unmap: | |
1340 | pte_unmap_unlock(pte, ptl); | |
1341 | if (ret) { | |
1342 | node = khugepaged_find_target_node(); | |
c1e8d7c6 | 1343 | /* collapse_huge_page will return with the mmap_lock released */ |
ffe945e6 KS |
1344 | collapse_huge_page(mm, address, hpage, node, |
1345 | referenced, unmapped); | |
b46e756f KS |
1346 | } |
1347 | out: | |
1348 | trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced, | |
1349 | none_or_zero, result, unmapped); | |
1350 | return ret; | |
1351 | } | |
1352 | ||
1353 | static void collect_mm_slot(struct mm_slot *mm_slot) | |
1354 | { | |
1355 | struct mm_struct *mm = mm_slot->mm; | |
1356 | ||
35f3aa39 | 1357 | lockdep_assert_held(&khugepaged_mm_lock); |
b46e756f KS |
1358 | |
1359 | if (khugepaged_test_exit(mm)) { | |
1360 | /* free mm_slot */ | |
1361 | hash_del(&mm_slot->hash); | |
1362 | list_del(&mm_slot->mm_node); | |
1363 | ||
1364 | /* | |
1365 | * Not strictly needed because the mm exited already. | |
1366 | * | |
1367 | * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); | |
1368 | */ | |
1369 | ||
1370 | /* khugepaged_mm_lock actually not necessary for the below */ | |
1371 | free_mm_slot(mm_slot); | |
1372 | mmdrop(mm); | |
1373 | } | |
1374 | } | |
1375 | ||
396bcc52 | 1376 | #ifdef CONFIG_SHMEM |
27e1f827 SL |
1377 | /* |
1378 | * Notify khugepaged that given addr of the mm is pte-mapped THP. Then | |
1379 | * khugepaged should try to collapse the page table. | |
1380 | */ | |
1381 | static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm, | |
1382 | unsigned long addr) | |
1383 | { | |
1384 | struct mm_slot *mm_slot; | |
1385 | ||
1386 | VM_BUG_ON(addr & ~HPAGE_PMD_MASK); | |
1387 | ||
1388 | spin_lock(&khugepaged_mm_lock); | |
1389 | mm_slot = get_mm_slot(mm); | |
1390 | if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP)) | |
1391 | mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr; | |
1392 | spin_unlock(&khugepaged_mm_lock); | |
1393 | return 0; | |
1394 | } | |
1395 | ||
e59a47b8 PT |
1396 | static void collapse_and_free_pmd(struct mm_struct *mm, struct vm_area_struct *vma, |
1397 | unsigned long addr, pmd_t *pmdp) | |
1398 | { | |
1399 | spinlock_t *ptl; | |
1400 | pmd_t pmd; | |
1401 | ||
80110bbf | 1402 | mmap_assert_write_locked(mm); |
e59a47b8 PT |
1403 | ptl = pmd_lock(vma->vm_mm, pmdp); |
1404 | pmd = pmdp_collapse_flush(vma, addr, pmdp); | |
1405 | spin_unlock(ptl); | |
1406 | mm_dec_nr_ptes(mm); | |
80110bbf | 1407 | page_table_check_pte_clear_range(mm, addr, pmd); |
e59a47b8 PT |
1408 | pte_free(mm, pmd_pgtable(pmd)); |
1409 | } | |
1410 | ||
27e1f827 | 1411 | /** |
336e6b53 AS |
1412 | * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at |
1413 | * address haddr. | |
1414 | * | |
1415 | * @mm: process address space where collapse happens | |
1416 | * @addr: THP collapse address | |
27e1f827 SL |
1417 | * |
1418 | * This function checks whether all the PTEs in the PMD are pointing to the | |
1419 | * right THP. If so, retract the page table so the THP can refault in with | |
1420 | * as pmd-mapped. | |
1421 | */ | |
1422 | void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr) | |
1423 | { | |
1424 | unsigned long haddr = addr & HPAGE_PMD_MASK; | |
1425 | struct vm_area_struct *vma = find_vma(mm, haddr); | |
119a5fc1 | 1426 | struct page *hpage; |
27e1f827 | 1427 | pte_t *start_pte, *pte; |
e59a47b8 | 1428 | pmd_t *pmd; |
27e1f827 SL |
1429 | spinlock_t *ptl; |
1430 | int count = 0; | |
1431 | int i; | |
1432 | ||
1433 | if (!vma || !vma->vm_file || | |
fef792a4 | 1434 | !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE)) |
27e1f827 SL |
1435 | return; |
1436 | ||
1437 | /* | |
1438 | * This vm_flags may not have VM_HUGEPAGE if the page was not | |
1439 | * collapsed by this mm. But we can still collapse if the page is | |
1440 | * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check() | |
1441 | * will not fail the vma for missing VM_HUGEPAGE | |
1442 | */ | |
1443 | if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE)) | |
1444 | return; | |
1445 | ||
deb4c93a PX |
1446 | /* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */ |
1447 | if (userfaultfd_wp(vma)) | |
1448 | return; | |
1449 | ||
119a5fc1 HD |
1450 | hpage = find_lock_page(vma->vm_file->f_mapping, |
1451 | linear_page_index(vma, haddr)); | |
1452 | if (!hpage) | |
1453 | return; | |
1454 | ||
1455 | if (!PageHead(hpage)) | |
1456 | goto drop_hpage; | |
1457 | ||
27e1f827 SL |
1458 | pmd = mm_find_pmd(mm, haddr); |
1459 | if (!pmd) | |
119a5fc1 | 1460 | goto drop_hpage; |
27e1f827 SL |
1461 | |
1462 | start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl); | |
1463 | ||
1464 | /* step 1: check all mapped PTEs are to the right huge page */ | |
1465 | for (i = 0, addr = haddr, pte = start_pte; | |
1466 | i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { | |
1467 | struct page *page; | |
1468 | ||
1469 | /* empty pte, skip */ | |
1470 | if (pte_none(*pte)) | |
1471 | continue; | |
1472 | ||
1473 | /* page swapped out, abort */ | |
1474 | if (!pte_present(*pte)) | |
1475 | goto abort; | |
1476 | ||
1477 | page = vm_normal_page(vma, addr, *pte); | |
1478 | ||
27e1f827 | 1479 | /* |
119a5fc1 HD |
1480 | * Note that uprobe, debugger, or MAP_PRIVATE may change the |
1481 | * page table, but the new page will not be a subpage of hpage. | |
27e1f827 | 1482 | */ |
119a5fc1 | 1483 | if (hpage + i != page) |
27e1f827 SL |
1484 | goto abort; |
1485 | count++; | |
1486 | } | |
1487 | ||
1488 | /* step 2: adjust rmap */ | |
1489 | for (i = 0, addr = haddr, pte = start_pte; | |
1490 | i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { | |
1491 | struct page *page; | |
1492 | ||
1493 | if (pte_none(*pte)) | |
1494 | continue; | |
1495 | page = vm_normal_page(vma, addr, *pte); | |
cea86fe2 | 1496 | page_remove_rmap(page, vma, false); |
27e1f827 SL |
1497 | } |
1498 | ||
1499 | pte_unmap_unlock(start_pte, ptl); | |
1500 | ||
1501 | /* step 3: set proper refcount and mm_counters. */ | |
119a5fc1 | 1502 | if (count) { |
27e1f827 SL |
1503 | page_ref_sub(hpage, count); |
1504 | add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count); | |
1505 | } | |
1506 | ||
1507 | /* step 4: collapse pmd */ | |
e59a47b8 | 1508 | collapse_and_free_pmd(mm, vma, haddr, pmd); |
119a5fc1 HD |
1509 | drop_hpage: |
1510 | unlock_page(hpage); | |
1511 | put_page(hpage); | |
27e1f827 SL |
1512 | return; |
1513 | ||
1514 | abort: | |
1515 | pte_unmap_unlock(start_pte, ptl); | |
119a5fc1 | 1516 | goto drop_hpage; |
27e1f827 SL |
1517 | } |
1518 | ||
0edf61e5 | 1519 | static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot) |
27e1f827 SL |
1520 | { |
1521 | struct mm_struct *mm = mm_slot->mm; | |
1522 | int i; | |
1523 | ||
1524 | if (likely(mm_slot->nr_pte_mapped_thp == 0)) | |
0edf61e5 | 1525 | return; |
27e1f827 | 1526 | |
d8ed45c5 | 1527 | if (!mmap_write_trylock(mm)) |
0edf61e5 | 1528 | return; |
27e1f827 SL |
1529 | |
1530 | if (unlikely(khugepaged_test_exit(mm))) | |
1531 | goto out; | |
1532 | ||
1533 | for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++) | |
1534 | collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]); | |
1535 | ||
1536 | out: | |
1537 | mm_slot->nr_pte_mapped_thp = 0; | |
d8ed45c5 | 1538 | mmap_write_unlock(mm); |
27e1f827 SL |
1539 | } |
1540 | ||
f3f0e1d2 KS |
1541 | static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff) |
1542 | { | |
1543 | struct vm_area_struct *vma; | |
18e77600 | 1544 | struct mm_struct *mm; |
f3f0e1d2 | 1545 | unsigned long addr; |
e59a47b8 | 1546 | pmd_t *pmd; |
f3f0e1d2 KS |
1547 | |
1548 | i_mmap_lock_write(mapping); | |
1549 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { | |
27e1f827 SL |
1550 | /* |
1551 | * Check vma->anon_vma to exclude MAP_PRIVATE mappings that | |
1552 | * got written to. These VMAs are likely not worth investing | |
3e4e28c5 | 1553 | * mmap_write_lock(mm) as PMD-mapping is likely to be split |
27e1f827 SL |
1554 | * later. |
1555 | * | |
36ee2c78 | 1556 | * Note that vma->anon_vma check is racy: it can be set up after |
c1e8d7c6 | 1557 | * the check but before we took mmap_lock by the fault path. |
27e1f827 SL |
1558 | * But page lock would prevent establishing any new ptes of the |
1559 | * page, so we are safe. | |
1560 | * | |
1561 | * An alternative would be drop the check, but check that page | |
1562 | * table is clear before calling pmdp_collapse_flush() under | |
1563 | * ptl. It has higher chance to recover THP for the VMA, but | |
1564 | * has higher cost too. | |
1565 | */ | |
f3f0e1d2 KS |
1566 | if (vma->anon_vma) |
1567 | continue; | |
1568 | addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); | |
1569 | if (addr & ~HPAGE_PMD_MASK) | |
1570 | continue; | |
1571 | if (vma->vm_end < addr + HPAGE_PMD_SIZE) | |
1572 | continue; | |
18e77600 HD |
1573 | mm = vma->vm_mm; |
1574 | pmd = mm_find_pmd(mm, addr); | |
f3f0e1d2 KS |
1575 | if (!pmd) |
1576 | continue; | |
1577 | /* | |
c1e8d7c6 | 1578 | * We need exclusive mmap_lock to retract page table. |
27e1f827 SL |
1579 | * |
1580 | * We use trylock due to lock inversion: we need to acquire | |
c1e8d7c6 | 1581 | * mmap_lock while holding page lock. Fault path does it in |
27e1f827 | 1582 | * reverse order. Trylock is a way to avoid deadlock. |
f3f0e1d2 | 1583 | */ |
18e77600 | 1584 | if (mmap_write_trylock(mm)) { |
deb4c93a PX |
1585 | /* |
1586 | * When a vma is registered with uffd-wp, we can't | |
1587 | * recycle the pmd pgtable because there can be pte | |
1588 | * markers installed. Skip it only, so the rest mm/vma | |
1589 | * can still have the same file mapped hugely, however | |
1590 | * it'll always mapped in small page size for uffd-wp | |
1591 | * registered ranges. | |
1592 | */ | |
1593 | if (!khugepaged_test_exit(mm) && !userfaultfd_wp(vma)) | |
e59a47b8 | 1594 | collapse_and_free_pmd(mm, vma, addr, pmd); |
18e77600 | 1595 | mmap_write_unlock(mm); |
27e1f827 SL |
1596 | } else { |
1597 | /* Try again later */ | |
18e77600 | 1598 | khugepaged_add_pte_mapped_thp(mm, addr); |
f3f0e1d2 KS |
1599 | } |
1600 | } | |
1601 | i_mmap_unlock_write(mapping); | |
1602 | } | |
1603 | ||
1604 | /** | |
99cb0dbd | 1605 | * collapse_file - collapse filemap/tmpfs/shmem pages into huge one. |
f3f0e1d2 | 1606 | * |
336e6b53 AS |
1607 | * @mm: process address space where collapse happens |
1608 | * @file: file that collapse on | |
1609 | * @start: collapse start address | |
1610 | * @hpage: new allocated huge page for collapse | |
1611 | * @node: appointed node the new huge page allocate from | |
1612 | * | |
f3f0e1d2 | 1613 | * Basic scheme is simple, details are more complex: |
87c460a0 | 1614 | * - allocate and lock a new huge page; |
77da9389 | 1615 | * - scan page cache replacing old pages with the new one |
99cb0dbd | 1616 | * + swap/gup in pages if necessary; |
f3f0e1d2 | 1617 | * + fill in gaps; |
77da9389 MW |
1618 | * + keep old pages around in case rollback is required; |
1619 | * - if replacing succeeds: | |
f3f0e1d2 KS |
1620 | * + copy data over; |
1621 | * + free old pages; | |
87c460a0 | 1622 | * + unlock huge page; |
f3f0e1d2 KS |
1623 | * - if replacing failed; |
1624 | * + put all pages back and unfreeze them; | |
77da9389 | 1625 | * + restore gaps in the page cache; |
87c460a0 | 1626 | * + unlock and free huge page; |
f3f0e1d2 | 1627 | */ |
579c571e SL |
1628 | static void collapse_file(struct mm_struct *mm, |
1629 | struct file *file, pgoff_t start, | |
f3f0e1d2 KS |
1630 | struct page **hpage, int node) |
1631 | { | |
579c571e | 1632 | struct address_space *mapping = file->f_mapping; |
f3f0e1d2 | 1633 | gfp_t gfp; |
77da9389 | 1634 | struct page *new_page; |
f3f0e1d2 KS |
1635 | pgoff_t index, end = start + HPAGE_PMD_NR; |
1636 | LIST_HEAD(pagelist); | |
77da9389 | 1637 | XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER); |
f3f0e1d2 | 1638 | int nr_none = 0, result = SCAN_SUCCEED; |
99cb0dbd | 1639 | bool is_shmem = shmem_file(file); |
bf9ecead | 1640 | int nr; |
f3f0e1d2 | 1641 | |
99cb0dbd | 1642 | VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem); |
f3f0e1d2 KS |
1643 | VM_BUG_ON(start & (HPAGE_PMD_NR - 1)); |
1644 | ||
1645 | /* Only allocate from the target node */ | |
41b6167e | 1646 | gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE; |
f3f0e1d2 KS |
1647 | |
1648 | new_page = khugepaged_alloc_page(hpage, gfp, node); | |
1649 | if (!new_page) { | |
1650 | result = SCAN_ALLOC_HUGE_PAGE_FAIL; | |
1651 | goto out; | |
1652 | } | |
1653 | ||
8f425e4e | 1654 | if (unlikely(mem_cgroup_charge(page_folio(new_page), mm, gfp))) { |
f3f0e1d2 KS |
1655 | result = SCAN_CGROUP_CHARGE_FAIL; |
1656 | goto out; | |
1657 | } | |
9d82c694 | 1658 | count_memcg_page_event(new_page, THP_COLLAPSE_ALLOC); |
f3f0e1d2 | 1659 | |
6b24ca4a MWO |
1660 | /* |
1661 | * Ensure we have slots for all the pages in the range. This is | |
1662 | * almost certainly a no-op because most of the pages must be present | |
1663 | */ | |
95feeabb HD |
1664 | do { |
1665 | xas_lock_irq(&xas); | |
1666 | xas_create_range(&xas); | |
1667 | if (!xas_error(&xas)) | |
1668 | break; | |
1669 | xas_unlock_irq(&xas); | |
1670 | if (!xas_nomem(&xas, GFP_KERNEL)) { | |
95feeabb HD |
1671 | result = SCAN_FAIL; |
1672 | goto out; | |
1673 | } | |
1674 | } while (1); | |
1675 | ||
042a3082 | 1676 | __SetPageLocked(new_page); |
99cb0dbd SL |
1677 | if (is_shmem) |
1678 | __SetPageSwapBacked(new_page); | |
f3f0e1d2 KS |
1679 | new_page->index = start; |
1680 | new_page->mapping = mapping; | |
f3f0e1d2 | 1681 | |
f3f0e1d2 | 1682 | /* |
87c460a0 HD |
1683 | * At this point the new_page is locked and not up-to-date. |
1684 | * It's safe to insert it into the page cache, because nobody would | |
1685 | * be able to map it or use it in another way until we unlock it. | |
f3f0e1d2 KS |
1686 | */ |
1687 | ||
77da9389 MW |
1688 | xas_set(&xas, start); |
1689 | for (index = start; index < end; index++) { | |
1690 | struct page *page = xas_next(&xas); | |
1691 | ||
1692 | VM_BUG_ON(index != xas.xa_index); | |
99cb0dbd SL |
1693 | if (is_shmem) { |
1694 | if (!page) { | |
1695 | /* | |
1696 | * Stop if extent has been truncated or | |
1697 | * hole-punched, and is now completely | |
1698 | * empty. | |
1699 | */ | |
1700 | if (index == start) { | |
1701 | if (!xas_next_entry(&xas, end - 1)) { | |
1702 | result = SCAN_TRUNCATED; | |
1703 | goto xa_locked; | |
1704 | } | |
1705 | xas_set(&xas, index); | |
1706 | } | |
1707 | if (!shmem_charge(mapping->host, 1)) { | |
1708 | result = SCAN_FAIL; | |
042a3082 | 1709 | goto xa_locked; |
701270fa | 1710 | } |
99cb0dbd SL |
1711 | xas_store(&xas, new_page); |
1712 | nr_none++; | |
1713 | continue; | |
701270fa | 1714 | } |
99cb0dbd SL |
1715 | |
1716 | if (xa_is_value(page) || !PageUptodate(page)) { | |
1717 | xas_unlock_irq(&xas); | |
1718 | /* swap in or instantiate fallocated page */ | |
1719 | if (shmem_getpage(mapping->host, index, &page, | |
acdd9f8e | 1720 | SGP_NOALLOC)) { |
99cb0dbd SL |
1721 | result = SCAN_FAIL; |
1722 | goto xa_unlocked; | |
1723 | } | |
1724 | } else if (trylock_page(page)) { | |
1725 | get_page(page); | |
1726 | xas_unlock_irq(&xas); | |
1727 | } else { | |
1728 | result = SCAN_PAGE_LOCK; | |
042a3082 | 1729 | goto xa_locked; |
77da9389 | 1730 | } |
99cb0dbd SL |
1731 | } else { /* !is_shmem */ |
1732 | if (!page || xa_is_value(page)) { | |
1733 | xas_unlock_irq(&xas); | |
1734 | page_cache_sync_readahead(mapping, &file->f_ra, | |
1735 | file, index, | |
e5a59d30 | 1736 | end - index); |
99cb0dbd SL |
1737 | /* drain pagevecs to help isolate_lru_page() */ |
1738 | lru_add_drain(); | |
1739 | page = find_lock_page(mapping, index); | |
1740 | if (unlikely(page == NULL)) { | |
1741 | result = SCAN_FAIL; | |
1742 | goto xa_unlocked; | |
1743 | } | |
75f36069 SL |
1744 | } else if (PageDirty(page)) { |
1745 | /* | |
1746 | * khugepaged only works on read-only fd, | |
1747 | * so this page is dirty because it hasn't | |
1748 | * been flushed since first write. There | |
1749 | * won't be new dirty pages. | |
1750 | * | |
1751 | * Trigger async flush here and hope the | |
1752 | * writeback is done when khugepaged | |
1753 | * revisits this page. | |
1754 | * | |
1755 | * This is a one-off situation. We are not | |
1756 | * forcing writeback in loop. | |
1757 | */ | |
1758 | xas_unlock_irq(&xas); | |
1759 | filemap_flush(mapping); | |
1760 | result = SCAN_FAIL; | |
1761 | goto xa_unlocked; | |
74c42e1b RW |
1762 | } else if (PageWriteback(page)) { |
1763 | xas_unlock_irq(&xas); | |
1764 | result = SCAN_FAIL; | |
1765 | goto xa_unlocked; | |
99cb0dbd SL |
1766 | } else if (trylock_page(page)) { |
1767 | get_page(page); | |
1768 | xas_unlock_irq(&xas); | |
1769 | } else { | |
1770 | result = SCAN_PAGE_LOCK; | |
1771 | goto xa_locked; | |
f3f0e1d2 | 1772 | } |
f3f0e1d2 KS |
1773 | } |
1774 | ||
1775 | /* | |
b93b0163 | 1776 | * The page must be locked, so we can drop the i_pages lock |
f3f0e1d2 KS |
1777 | * without racing with truncate. |
1778 | */ | |
1779 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
4655e5e5 SL |
1780 | |
1781 | /* make sure the page is up to date */ | |
1782 | if (unlikely(!PageUptodate(page))) { | |
1783 | result = SCAN_FAIL; | |
1784 | goto out_unlock; | |
1785 | } | |
06a5e126 HD |
1786 | |
1787 | /* | |
1788 | * If file was truncated then extended, or hole-punched, before | |
1789 | * we locked the first page, then a THP might be there already. | |
1790 | */ | |
1791 | if (PageTransCompound(page)) { | |
1792 | result = SCAN_PAGE_COMPOUND; | |
1793 | goto out_unlock; | |
1794 | } | |
f3f0e1d2 KS |
1795 | |
1796 | if (page_mapping(page) != mapping) { | |
1797 | result = SCAN_TRUNCATED; | |
1798 | goto out_unlock; | |
1799 | } | |
f3f0e1d2 | 1800 | |
74c42e1b RW |
1801 | if (!is_shmem && (PageDirty(page) || |
1802 | PageWriteback(page))) { | |
4655e5e5 SL |
1803 | /* |
1804 | * khugepaged only works on read-only fd, so this | |
1805 | * page is dirty because it hasn't been flushed | |
1806 | * since first write. | |
1807 | */ | |
1808 | result = SCAN_FAIL; | |
1809 | goto out_unlock; | |
1810 | } | |
1811 | ||
f3f0e1d2 KS |
1812 | if (isolate_lru_page(page)) { |
1813 | result = SCAN_DEL_PAGE_LRU; | |
042a3082 | 1814 | goto out_unlock; |
f3f0e1d2 KS |
1815 | } |
1816 | ||
99cb0dbd SL |
1817 | if (page_has_private(page) && |
1818 | !try_to_release_page(page, GFP_KERNEL)) { | |
1819 | result = SCAN_PAGE_HAS_PRIVATE; | |
2f33a706 | 1820 | putback_lru_page(page); |
99cb0dbd SL |
1821 | goto out_unlock; |
1822 | } | |
1823 | ||
f3f0e1d2 | 1824 | if (page_mapped(page)) |
869f7ee6 MWO |
1825 | try_to_unmap(page_folio(page), |
1826 | TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH); | |
f3f0e1d2 | 1827 | |
77da9389 MW |
1828 | xas_lock_irq(&xas); |
1829 | xas_set(&xas, index); | |
f3f0e1d2 | 1830 | |
77da9389 | 1831 | VM_BUG_ON_PAGE(page != xas_load(&xas), page); |
f3f0e1d2 KS |
1832 | |
1833 | /* | |
1834 | * The page is expected to have page_count() == 3: | |
1835 | * - we hold a pin on it; | |
77da9389 | 1836 | * - one reference from page cache; |
f3f0e1d2 KS |
1837 | * - one from isolate_lru_page; |
1838 | */ | |
1839 | if (!page_ref_freeze(page, 3)) { | |
1840 | result = SCAN_PAGE_COUNT; | |
042a3082 HD |
1841 | xas_unlock_irq(&xas); |
1842 | putback_lru_page(page); | |
1843 | goto out_unlock; | |
f3f0e1d2 KS |
1844 | } |
1845 | ||
1846 | /* | |
1847 | * Add the page to the list to be able to undo the collapse if | |
1848 | * something go wrong. | |
1849 | */ | |
1850 | list_add_tail(&page->lru, &pagelist); | |
1851 | ||
1852 | /* Finally, replace with the new page. */ | |
4101196b | 1853 | xas_store(&xas, new_page); |
f3f0e1d2 | 1854 | continue; |
f3f0e1d2 KS |
1855 | out_unlock: |
1856 | unlock_page(page); | |
1857 | put_page(page); | |
042a3082 | 1858 | goto xa_unlocked; |
f3f0e1d2 | 1859 | } |
bf9ecead | 1860 | nr = thp_nr_pages(new_page); |
f3f0e1d2 | 1861 | |
99cb0dbd | 1862 | if (is_shmem) |
57b2847d | 1863 | __mod_lruvec_page_state(new_page, NR_SHMEM_THPS, nr); |
09d91cda | 1864 | else { |
bf9ecead | 1865 | __mod_lruvec_page_state(new_page, NR_FILE_THPS, nr); |
09d91cda | 1866 | filemap_nr_thps_inc(mapping); |
eb6ecbed CF |
1867 | /* |
1868 | * Paired with smp_mb() in do_dentry_open() to ensure | |
1869 | * i_writecount is up to date and the update to nr_thps is | |
1870 | * visible. Ensures the page cache will be truncated if the | |
1871 | * file is opened writable. | |
1872 | */ | |
1873 | smp_mb(); | |
1874 | if (inode_is_open_for_write(mapping->host)) { | |
1875 | result = SCAN_FAIL; | |
1876 | __mod_lruvec_page_state(new_page, NR_FILE_THPS, -nr); | |
1877 | filemap_nr_thps_dec(mapping); | |
1878 | goto xa_locked; | |
1879 | } | |
09d91cda | 1880 | } |
99cb0dbd | 1881 | |
042a3082 | 1882 | if (nr_none) { |
9d82c694 | 1883 | __mod_lruvec_page_state(new_page, NR_FILE_PAGES, nr_none); |
2f55f070 ML |
1884 | /* nr_none is always 0 for non-shmem. */ |
1885 | __mod_lruvec_page_state(new_page, NR_SHMEM, nr_none); | |
042a3082 HD |
1886 | } |
1887 | ||
6b24ca4a MWO |
1888 | /* Join all the small entries into a single multi-index entry */ |
1889 | xas_set_order(&xas, start, HPAGE_PMD_ORDER); | |
1890 | xas_store(&xas, new_page); | |
042a3082 HD |
1891 | xa_locked: |
1892 | xas_unlock_irq(&xas); | |
77da9389 | 1893 | xa_unlocked: |
042a3082 | 1894 | |
6d9df8a5 HD |
1895 | /* |
1896 | * If collapse is successful, flush must be done now before copying. | |
1897 | * If collapse is unsuccessful, does flush actually need to be done? | |
1898 | * Do it anyway, to clear the state. | |
1899 | */ | |
1900 | try_to_unmap_flush(); | |
1901 | ||
f3f0e1d2 | 1902 | if (result == SCAN_SUCCEED) { |
77da9389 | 1903 | struct page *page, *tmp; |
f3f0e1d2 KS |
1904 | |
1905 | /* | |
77da9389 MW |
1906 | * Replacing old pages with new one has succeeded, now we |
1907 | * need to copy the content and free the old pages. | |
f3f0e1d2 | 1908 | */ |
2af8ff29 | 1909 | index = start; |
f3f0e1d2 | 1910 | list_for_each_entry_safe(page, tmp, &pagelist, lru) { |
2af8ff29 HD |
1911 | while (index < page->index) { |
1912 | clear_highpage(new_page + (index % HPAGE_PMD_NR)); | |
1913 | index++; | |
1914 | } | |
f3f0e1d2 KS |
1915 | copy_highpage(new_page + (page->index % HPAGE_PMD_NR), |
1916 | page); | |
1917 | list_del(&page->lru); | |
f3f0e1d2 | 1918 | page->mapping = NULL; |
042a3082 | 1919 | page_ref_unfreeze(page, 1); |
f3f0e1d2 KS |
1920 | ClearPageActive(page); |
1921 | ClearPageUnevictable(page); | |
042a3082 | 1922 | unlock_page(page); |
f3f0e1d2 | 1923 | put_page(page); |
2af8ff29 HD |
1924 | index++; |
1925 | } | |
1926 | while (index < end) { | |
1927 | clear_highpage(new_page + (index % HPAGE_PMD_NR)); | |
1928 | index++; | |
f3f0e1d2 KS |
1929 | } |
1930 | ||
f3f0e1d2 | 1931 | SetPageUptodate(new_page); |
87c460a0 | 1932 | page_ref_add(new_page, HPAGE_PMD_NR - 1); |
6058eaec | 1933 | if (is_shmem) |
99cb0dbd | 1934 | set_page_dirty(new_page); |
6058eaec | 1935 | lru_cache_add(new_page); |
f3f0e1d2 | 1936 | |
042a3082 HD |
1937 | /* |
1938 | * Remove pte page tables, so we can re-fault the page as huge. | |
1939 | */ | |
1940 | retract_page_tables(mapping, start); | |
f3f0e1d2 | 1941 | *hpage = NULL; |
87aa7529 YS |
1942 | |
1943 | khugepaged_pages_collapsed++; | |
f3f0e1d2 | 1944 | } else { |
77da9389 | 1945 | struct page *page; |
aaa52e34 | 1946 | |
77da9389 | 1947 | /* Something went wrong: roll back page cache changes */ |
77da9389 | 1948 | xas_lock_irq(&xas); |
2f55f070 ML |
1949 | if (nr_none) { |
1950 | mapping->nrpages -= nr_none; | |
99cb0dbd | 1951 | shmem_uncharge(mapping->host, nr_none); |
2f55f070 | 1952 | } |
aaa52e34 | 1953 | |
77da9389 MW |
1954 | xas_set(&xas, start); |
1955 | xas_for_each(&xas, page, end - 1) { | |
f3f0e1d2 KS |
1956 | page = list_first_entry_or_null(&pagelist, |
1957 | struct page, lru); | |
77da9389 | 1958 | if (!page || xas.xa_index < page->index) { |
f3f0e1d2 KS |
1959 | if (!nr_none) |
1960 | break; | |
f3f0e1d2 | 1961 | nr_none--; |
59749e6c | 1962 | /* Put holes back where they were */ |
77da9389 | 1963 | xas_store(&xas, NULL); |
f3f0e1d2 KS |
1964 | continue; |
1965 | } | |
1966 | ||
77da9389 | 1967 | VM_BUG_ON_PAGE(page->index != xas.xa_index, page); |
f3f0e1d2 KS |
1968 | |
1969 | /* Unfreeze the page. */ | |
1970 | list_del(&page->lru); | |
1971 | page_ref_unfreeze(page, 2); | |
77da9389 MW |
1972 | xas_store(&xas, page); |
1973 | xas_pause(&xas); | |
1974 | xas_unlock_irq(&xas); | |
f3f0e1d2 | 1975 | unlock_page(page); |
042a3082 | 1976 | putback_lru_page(page); |
77da9389 | 1977 | xas_lock_irq(&xas); |
f3f0e1d2 KS |
1978 | } |
1979 | VM_BUG_ON(nr_none); | |
77da9389 | 1980 | xas_unlock_irq(&xas); |
f3f0e1d2 | 1981 | |
f3f0e1d2 KS |
1982 | new_page->mapping = NULL; |
1983 | } | |
042a3082 HD |
1984 | |
1985 | unlock_page(new_page); | |
f3f0e1d2 KS |
1986 | out: |
1987 | VM_BUG_ON(!list_empty(&pagelist)); | |
9d82c694 | 1988 | if (!IS_ERR_OR_NULL(*hpage)) |
bbc6b703 | 1989 | mem_cgroup_uncharge(page_folio(*hpage)); |
f3f0e1d2 KS |
1990 | /* TODO: tracepoints */ |
1991 | } | |
1992 | ||
579c571e SL |
1993 | static void khugepaged_scan_file(struct mm_struct *mm, |
1994 | struct file *file, pgoff_t start, struct page **hpage) | |
f3f0e1d2 KS |
1995 | { |
1996 | struct page *page = NULL; | |
579c571e | 1997 | struct address_space *mapping = file->f_mapping; |
85b392db | 1998 | XA_STATE(xas, &mapping->i_pages, start); |
f3f0e1d2 KS |
1999 | int present, swap; |
2000 | int node = NUMA_NO_NODE; | |
2001 | int result = SCAN_SUCCEED; | |
2002 | ||
2003 | present = 0; | |
2004 | swap = 0; | |
2005 | memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load)); | |
2006 | rcu_read_lock(); | |
85b392db MW |
2007 | xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) { |
2008 | if (xas_retry(&xas, page)) | |
f3f0e1d2 | 2009 | continue; |
f3f0e1d2 | 2010 | |
85b392db | 2011 | if (xa_is_value(page)) { |
f3f0e1d2 KS |
2012 | if (++swap > khugepaged_max_ptes_swap) { |
2013 | result = SCAN_EXCEED_SWAP_PTE; | |
e9ea874a | 2014 | count_vm_event(THP_SCAN_EXCEED_SWAP_PTE); |
f3f0e1d2 KS |
2015 | break; |
2016 | } | |
2017 | continue; | |
2018 | } | |
2019 | ||
6b24ca4a MWO |
2020 | /* |
2021 | * XXX: khugepaged should compact smaller compound pages | |
2022 | * into a PMD sized page | |
2023 | */ | |
f3f0e1d2 KS |
2024 | if (PageTransCompound(page)) { |
2025 | result = SCAN_PAGE_COMPOUND; | |
2026 | break; | |
2027 | } | |
2028 | ||
2029 | node = page_to_nid(page); | |
2030 | if (khugepaged_scan_abort(node)) { | |
2031 | result = SCAN_SCAN_ABORT; | |
2032 | break; | |
2033 | } | |
2034 | khugepaged_node_load[node]++; | |
2035 | ||
2036 | if (!PageLRU(page)) { | |
2037 | result = SCAN_PAGE_LRU; | |
2038 | break; | |
2039 | } | |
2040 | ||
99cb0dbd SL |
2041 | if (page_count(page) != |
2042 | 1 + page_mapcount(page) + page_has_private(page)) { | |
f3f0e1d2 KS |
2043 | result = SCAN_PAGE_COUNT; |
2044 | break; | |
2045 | } | |
2046 | ||
2047 | /* | |
2048 | * We probably should check if the page is referenced here, but | |
2049 | * nobody would transfer pte_young() to PageReferenced() for us. | |
2050 | * And rmap walk here is just too costly... | |
2051 | */ | |
2052 | ||
2053 | present++; | |
2054 | ||
2055 | if (need_resched()) { | |
85b392db | 2056 | xas_pause(&xas); |
f3f0e1d2 | 2057 | cond_resched_rcu(); |
f3f0e1d2 KS |
2058 | } |
2059 | } | |
2060 | rcu_read_unlock(); | |
2061 | ||
2062 | if (result == SCAN_SUCCEED) { | |
2063 | if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) { | |
2064 | result = SCAN_EXCEED_NONE_PTE; | |
e9ea874a | 2065 | count_vm_event(THP_SCAN_EXCEED_NONE_PTE); |
f3f0e1d2 KS |
2066 | } else { |
2067 | node = khugepaged_find_target_node(); | |
579c571e | 2068 | collapse_file(mm, file, start, hpage, node); |
f3f0e1d2 KS |
2069 | } |
2070 | } | |
2071 | ||
2072 | /* TODO: tracepoints */ | |
2073 | } | |
2074 | #else | |
579c571e SL |
2075 | static void khugepaged_scan_file(struct mm_struct *mm, |
2076 | struct file *file, pgoff_t start, struct page **hpage) | |
f3f0e1d2 KS |
2077 | { |
2078 | BUILD_BUG(); | |
2079 | } | |
27e1f827 | 2080 | |
0edf61e5 | 2081 | static void khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot) |
27e1f827 | 2082 | { |
27e1f827 | 2083 | } |
f3f0e1d2 KS |
2084 | #endif |
2085 | ||
b46e756f KS |
2086 | static unsigned int khugepaged_scan_mm_slot(unsigned int pages, |
2087 | struct page **hpage) | |
2088 | __releases(&khugepaged_mm_lock) | |
2089 | __acquires(&khugepaged_mm_lock) | |
2090 | { | |
2091 | struct mm_slot *mm_slot; | |
2092 | struct mm_struct *mm; | |
2093 | struct vm_area_struct *vma; | |
2094 | int progress = 0; | |
2095 | ||
2096 | VM_BUG_ON(!pages); | |
35f3aa39 | 2097 | lockdep_assert_held(&khugepaged_mm_lock); |
b46e756f KS |
2098 | |
2099 | if (khugepaged_scan.mm_slot) | |
2100 | mm_slot = khugepaged_scan.mm_slot; | |
2101 | else { | |
2102 | mm_slot = list_entry(khugepaged_scan.mm_head.next, | |
2103 | struct mm_slot, mm_node); | |
2104 | khugepaged_scan.address = 0; | |
2105 | khugepaged_scan.mm_slot = mm_slot; | |
2106 | } | |
2107 | spin_unlock(&khugepaged_mm_lock); | |
27e1f827 | 2108 | khugepaged_collapse_pte_mapped_thps(mm_slot); |
b46e756f KS |
2109 | |
2110 | mm = mm_slot->mm; | |
3b454ad3 YS |
2111 | /* |
2112 | * Don't wait for semaphore (to avoid long wait times). Just move to | |
2113 | * the next mm on the list. | |
2114 | */ | |
2115 | vma = NULL; | |
d8ed45c5 | 2116 | if (unlikely(!mmap_read_trylock(mm))) |
c1e8d7c6 | 2117 | goto breakouterloop_mmap_lock; |
3b454ad3 | 2118 | if (likely(!khugepaged_test_exit(mm))) |
b46e756f KS |
2119 | vma = find_vma(mm, khugepaged_scan.address); |
2120 | ||
2121 | progress++; | |
2122 | for (; vma; vma = vma->vm_next) { | |
2123 | unsigned long hstart, hend; | |
2124 | ||
2125 | cond_resched(); | |
2126 | if (unlikely(khugepaged_test_exit(mm))) { | |
2127 | progress++; | |
2128 | break; | |
2129 | } | |
50f8b92f | 2130 | if (!hugepage_vma_check(vma, vma->vm_flags)) { |
b46e756f KS |
2131 | skip: |
2132 | progress++; | |
2133 | continue; | |
2134 | } | |
2135 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; | |
2136 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
2137 | if (hstart >= hend) | |
2138 | goto skip; | |
2139 | if (khugepaged_scan.address > hend) | |
2140 | goto skip; | |
2141 | if (khugepaged_scan.address < hstart) | |
2142 | khugepaged_scan.address = hstart; | |
2143 | VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); | |
2144 | ||
2145 | while (khugepaged_scan.address < hend) { | |
2146 | int ret; | |
2147 | cond_resched(); | |
2148 | if (unlikely(khugepaged_test_exit(mm))) | |
2149 | goto breakouterloop; | |
2150 | ||
2151 | VM_BUG_ON(khugepaged_scan.address < hstart || | |
2152 | khugepaged_scan.address + HPAGE_PMD_SIZE > | |
2153 | hend); | |
99cb0dbd | 2154 | if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) { |
396bcc52 | 2155 | struct file *file = get_file(vma->vm_file); |
f3f0e1d2 KS |
2156 | pgoff_t pgoff = linear_page_index(vma, |
2157 | khugepaged_scan.address); | |
99cb0dbd | 2158 | |
d8ed45c5 | 2159 | mmap_read_unlock(mm); |
f3f0e1d2 | 2160 | ret = 1; |
579c571e | 2161 | khugepaged_scan_file(mm, file, pgoff, hpage); |
f3f0e1d2 KS |
2162 | fput(file); |
2163 | } else { | |
2164 | ret = khugepaged_scan_pmd(mm, vma, | |
2165 | khugepaged_scan.address, | |
2166 | hpage); | |
2167 | } | |
b46e756f KS |
2168 | /* move to next address */ |
2169 | khugepaged_scan.address += HPAGE_PMD_SIZE; | |
2170 | progress += HPAGE_PMD_NR; | |
2171 | if (ret) | |
c1e8d7c6 ML |
2172 | /* we released mmap_lock so break loop */ |
2173 | goto breakouterloop_mmap_lock; | |
b46e756f KS |
2174 | if (progress >= pages) |
2175 | goto breakouterloop; | |
2176 | } | |
2177 | } | |
2178 | breakouterloop: | |
d8ed45c5 | 2179 | mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */ |
c1e8d7c6 | 2180 | breakouterloop_mmap_lock: |
b46e756f KS |
2181 | |
2182 | spin_lock(&khugepaged_mm_lock); | |
2183 | VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); | |
2184 | /* | |
2185 | * Release the current mm_slot if this mm is about to die, or | |
2186 | * if we scanned all vmas of this mm. | |
2187 | */ | |
2188 | if (khugepaged_test_exit(mm) || !vma) { | |
2189 | /* | |
2190 | * Make sure that if mm_users is reaching zero while | |
2191 | * khugepaged runs here, khugepaged_exit will find | |
2192 | * mm_slot not pointing to the exiting mm. | |
2193 | */ | |
2194 | if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) { | |
2195 | khugepaged_scan.mm_slot = list_entry( | |
2196 | mm_slot->mm_node.next, | |
2197 | struct mm_slot, mm_node); | |
2198 | khugepaged_scan.address = 0; | |
2199 | } else { | |
2200 | khugepaged_scan.mm_slot = NULL; | |
2201 | khugepaged_full_scans++; | |
2202 | } | |
2203 | ||
2204 | collect_mm_slot(mm_slot); | |
2205 | } | |
2206 | ||
2207 | return progress; | |
2208 | } | |
2209 | ||
2210 | static int khugepaged_has_work(void) | |
2211 | { | |
2212 | return !list_empty(&khugepaged_scan.mm_head) && | |
2213 | khugepaged_enabled(); | |
2214 | } | |
2215 | ||
2216 | static int khugepaged_wait_event(void) | |
2217 | { | |
2218 | return !list_empty(&khugepaged_scan.mm_head) || | |
2219 | kthread_should_stop(); | |
2220 | } | |
2221 | ||
2222 | static void khugepaged_do_scan(void) | |
2223 | { | |
2224 | struct page *hpage = NULL; | |
2225 | unsigned int progress = 0, pass_through_head = 0; | |
89dc6a96 | 2226 | unsigned int pages = READ_ONCE(khugepaged_pages_to_scan); |
b46e756f KS |
2227 | bool wait = true; |
2228 | ||
a980df33 KS |
2229 | lru_add_drain_all(); |
2230 | ||
b46e756f KS |
2231 | while (progress < pages) { |
2232 | if (!khugepaged_prealloc_page(&hpage, &wait)) | |
2233 | break; | |
2234 | ||
2235 | cond_resched(); | |
2236 | ||
2237 | if (unlikely(kthread_should_stop() || try_to_freeze())) | |
2238 | break; | |
2239 | ||
2240 | spin_lock(&khugepaged_mm_lock); | |
2241 | if (!khugepaged_scan.mm_slot) | |
2242 | pass_through_head++; | |
2243 | if (khugepaged_has_work() && | |
2244 | pass_through_head < 2) | |
2245 | progress += khugepaged_scan_mm_slot(pages - progress, | |
2246 | &hpage); | |
2247 | else | |
2248 | progress = pages; | |
2249 | spin_unlock(&khugepaged_mm_lock); | |
2250 | } | |
2251 | ||
2252 | if (!IS_ERR_OR_NULL(hpage)) | |
2253 | put_page(hpage); | |
2254 | } | |
2255 | ||
2256 | static bool khugepaged_should_wakeup(void) | |
2257 | { | |
2258 | return kthread_should_stop() || | |
2259 | time_after_eq(jiffies, khugepaged_sleep_expire); | |
2260 | } | |
2261 | ||
2262 | static void khugepaged_wait_work(void) | |
2263 | { | |
2264 | if (khugepaged_has_work()) { | |
2265 | const unsigned long scan_sleep_jiffies = | |
2266 | msecs_to_jiffies(khugepaged_scan_sleep_millisecs); | |
2267 | ||
2268 | if (!scan_sleep_jiffies) | |
2269 | return; | |
2270 | ||
2271 | khugepaged_sleep_expire = jiffies + scan_sleep_jiffies; | |
2272 | wait_event_freezable_timeout(khugepaged_wait, | |
2273 | khugepaged_should_wakeup(), | |
2274 | scan_sleep_jiffies); | |
2275 | return; | |
2276 | } | |
2277 | ||
2278 | if (khugepaged_enabled()) | |
2279 | wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); | |
2280 | } | |
2281 | ||
2282 | static int khugepaged(void *none) | |
2283 | { | |
2284 | struct mm_slot *mm_slot; | |
2285 | ||
2286 | set_freezable(); | |
2287 | set_user_nice(current, MAX_NICE); | |
2288 | ||
2289 | while (!kthread_should_stop()) { | |
2290 | khugepaged_do_scan(); | |
2291 | khugepaged_wait_work(); | |
2292 | } | |
2293 | ||
2294 | spin_lock(&khugepaged_mm_lock); | |
2295 | mm_slot = khugepaged_scan.mm_slot; | |
2296 | khugepaged_scan.mm_slot = NULL; | |
2297 | if (mm_slot) | |
2298 | collect_mm_slot(mm_slot); | |
2299 | spin_unlock(&khugepaged_mm_lock); | |
2300 | return 0; | |
2301 | } | |
2302 | ||
2303 | static void set_recommended_min_free_kbytes(void) | |
2304 | { | |
2305 | struct zone *zone; | |
2306 | int nr_zones = 0; | |
2307 | unsigned long recommended_min; | |
2308 | ||
bd3400ea LF |
2309 | if (!khugepaged_enabled()) { |
2310 | calculate_min_free_kbytes(); | |
2311 | goto update_wmarks; | |
2312 | } | |
2313 | ||
b7d349c7 JK |
2314 | for_each_populated_zone(zone) { |
2315 | /* | |
2316 | * We don't need to worry about fragmentation of | |
2317 | * ZONE_MOVABLE since it only has movable pages. | |
2318 | */ | |
2319 | if (zone_idx(zone) > gfp_zone(GFP_USER)) | |
2320 | continue; | |
2321 | ||
b46e756f | 2322 | nr_zones++; |
b7d349c7 | 2323 | } |
b46e756f KS |
2324 | |
2325 | /* Ensure 2 pageblocks are free to assist fragmentation avoidance */ | |
2326 | recommended_min = pageblock_nr_pages * nr_zones * 2; | |
2327 | ||
2328 | /* | |
2329 | * Make sure that on average at least two pageblocks are almost free | |
2330 | * of another type, one for a migratetype to fall back to and a | |
2331 | * second to avoid subsequent fallbacks of other types There are 3 | |
2332 | * MIGRATE_TYPES we care about. | |
2333 | */ | |
2334 | recommended_min += pageblock_nr_pages * nr_zones * | |
2335 | MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; | |
2336 | ||
2337 | /* don't ever allow to reserve more than 5% of the lowmem */ | |
2338 | recommended_min = min(recommended_min, | |
2339 | (unsigned long) nr_free_buffer_pages() / 20); | |
2340 | recommended_min <<= (PAGE_SHIFT-10); | |
2341 | ||
2342 | if (recommended_min > min_free_kbytes) { | |
2343 | if (user_min_free_kbytes >= 0) | |
2344 | pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n", | |
2345 | min_free_kbytes, recommended_min); | |
2346 | ||
2347 | min_free_kbytes = recommended_min; | |
2348 | } | |
bd3400ea LF |
2349 | |
2350 | update_wmarks: | |
b46e756f KS |
2351 | setup_per_zone_wmarks(); |
2352 | } | |
2353 | ||
2354 | int start_stop_khugepaged(void) | |
2355 | { | |
b46e756f KS |
2356 | int err = 0; |
2357 | ||
2358 | mutex_lock(&khugepaged_mutex); | |
2359 | if (khugepaged_enabled()) { | |
2360 | if (!khugepaged_thread) | |
2361 | khugepaged_thread = kthread_run(khugepaged, NULL, | |
2362 | "khugepaged"); | |
2363 | if (IS_ERR(khugepaged_thread)) { | |
2364 | pr_err("khugepaged: kthread_run(khugepaged) failed\n"); | |
2365 | err = PTR_ERR(khugepaged_thread); | |
2366 | khugepaged_thread = NULL; | |
2367 | goto fail; | |
2368 | } | |
2369 | ||
2370 | if (!list_empty(&khugepaged_scan.mm_head)) | |
2371 | wake_up_interruptible(&khugepaged_wait); | |
b46e756f KS |
2372 | } else if (khugepaged_thread) { |
2373 | kthread_stop(khugepaged_thread); | |
2374 | khugepaged_thread = NULL; | |
2375 | } | |
bd3400ea | 2376 | set_recommended_min_free_kbytes(); |
b46e756f KS |
2377 | fail: |
2378 | mutex_unlock(&khugepaged_mutex); | |
2379 | return err; | |
2380 | } | |
4aab2be0 VB |
2381 | |
2382 | void khugepaged_min_free_kbytes_update(void) | |
2383 | { | |
2384 | mutex_lock(&khugepaged_mutex); | |
2385 | if (khugepaged_enabled() && khugepaged_thread) | |
2386 | set_recommended_min_free_kbytes(); | |
2387 | mutex_unlock(&khugepaged_mutex); | |
2388 | } |