]>
Commit | Line | Data |
---|---|---|
71e3aac0 AA |
1 | /* |
2 | * Copyright (C) 2009 Red Hat, Inc. | |
3 | * | |
4 | * This work is licensed under the terms of the GNU GPL, version 2. See | |
5 | * the COPYING file in the top-level directory. | |
6 | */ | |
7 | ||
ae3a8c1c AM |
8 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
9 | ||
71e3aac0 AA |
10 | #include <linux/mm.h> |
11 | #include <linux/sched.h> | |
12 | #include <linux/highmem.h> | |
13 | #include <linux/hugetlb.h> | |
14 | #include <linux/mmu_notifier.h> | |
15 | #include <linux/rmap.h> | |
16 | #include <linux/swap.h> | |
97ae1749 | 17 | #include <linux/shrinker.h> |
ba76149f AA |
18 | #include <linux/mm_inline.h> |
19 | #include <linux/kthread.h> | |
20 | #include <linux/khugepaged.h> | |
878aee7d | 21 | #include <linux/freezer.h> |
a664b2d8 | 22 | #include <linux/mman.h> |
325adeb5 | 23 | #include <linux/pagemap.h> |
4daae3b4 | 24 | #include <linux/migrate.h> |
43b5fbbd | 25 | #include <linux/hashtable.h> |
97ae1749 | 26 | |
71e3aac0 AA |
27 | #include <asm/tlb.h> |
28 | #include <asm/pgalloc.h> | |
29 | #include "internal.h" | |
30 | ||
ba76149f | 31 | /* |
8bfa3f9a JW |
32 | * By default transparent hugepage support is disabled in order that avoid |
33 | * to risk increase the memory footprint of applications without a guaranteed | |
34 | * benefit. When transparent hugepage support is enabled, is for all mappings, | |
35 | * and khugepaged scans all mappings. | |
36 | * Defrag is invoked by khugepaged hugepage allocations and by page faults | |
37 | * for all hugepage allocations. | |
ba76149f | 38 | */ |
71e3aac0 | 39 | unsigned long transparent_hugepage_flags __read_mostly = |
13ece886 | 40 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS |
ba76149f | 41 | (1<<TRANSPARENT_HUGEPAGE_FLAG)| |
13ece886 AA |
42 | #endif |
43 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE | |
44 | (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| | |
45 | #endif | |
d39d33c3 | 46 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)| |
79da5407 KS |
47 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| |
48 | (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
ba76149f AA |
49 | |
50 | /* default scan 8*512 pte (or vmas) every 30 second */ | |
51 | static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8; | |
52 | static unsigned int khugepaged_pages_collapsed; | |
53 | static unsigned int khugepaged_full_scans; | |
54 | static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; | |
55 | /* during fragmentation poll the hugepage allocator once every minute */ | |
56 | static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; | |
57 | static struct task_struct *khugepaged_thread __read_mostly; | |
58 | static DEFINE_MUTEX(khugepaged_mutex); | |
59 | static DEFINE_SPINLOCK(khugepaged_mm_lock); | |
60 | static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); | |
61 | /* | |
62 | * default collapse hugepages if there is at least one pte mapped like | |
63 | * it would have happened if the vma was large enough during page | |
64 | * fault. | |
65 | */ | |
66 | static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1; | |
67 | ||
68 | static int khugepaged(void *none); | |
ba76149f | 69 | static int khugepaged_slab_init(void); |
ba76149f | 70 | |
43b5fbbd SL |
71 | #define MM_SLOTS_HASH_BITS 10 |
72 | static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); | |
73 | ||
ba76149f AA |
74 | static struct kmem_cache *mm_slot_cache __read_mostly; |
75 | ||
76 | /** | |
77 | * struct mm_slot - hash lookup from mm to mm_slot | |
78 | * @hash: hash collision list | |
79 | * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head | |
80 | * @mm: the mm that this information is valid for | |
81 | */ | |
82 | struct mm_slot { | |
83 | struct hlist_node hash; | |
84 | struct list_head mm_node; | |
85 | struct mm_struct *mm; | |
86 | }; | |
87 | ||
88 | /** | |
89 | * struct khugepaged_scan - cursor for scanning | |
90 | * @mm_head: the head of the mm list to scan | |
91 | * @mm_slot: the current mm_slot we are scanning | |
92 | * @address: the next address inside that to be scanned | |
93 | * | |
94 | * There is only the one khugepaged_scan instance of this cursor structure. | |
95 | */ | |
96 | struct khugepaged_scan { | |
97 | struct list_head mm_head; | |
98 | struct mm_slot *mm_slot; | |
99 | unsigned long address; | |
2f1da642 HS |
100 | }; |
101 | static struct khugepaged_scan khugepaged_scan = { | |
ba76149f AA |
102 | .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), |
103 | }; | |
104 | ||
f000565a AA |
105 | |
106 | static int set_recommended_min_free_kbytes(void) | |
107 | { | |
108 | struct zone *zone; | |
109 | int nr_zones = 0; | |
110 | unsigned long recommended_min; | |
f000565a | 111 | |
17c230af | 112 | if (!khugepaged_enabled()) |
f000565a AA |
113 | return 0; |
114 | ||
115 | for_each_populated_zone(zone) | |
116 | nr_zones++; | |
117 | ||
118 | /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */ | |
119 | recommended_min = pageblock_nr_pages * nr_zones * 2; | |
120 | ||
121 | /* | |
122 | * Make sure that on average at least two pageblocks are almost free | |
123 | * of another type, one for a migratetype to fall back to and a | |
124 | * second to avoid subsequent fallbacks of other types There are 3 | |
125 | * MIGRATE_TYPES we care about. | |
126 | */ | |
127 | recommended_min += pageblock_nr_pages * nr_zones * | |
128 | MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; | |
129 | ||
130 | /* don't ever allow to reserve more than 5% of the lowmem */ | |
131 | recommended_min = min(recommended_min, | |
132 | (unsigned long) nr_free_buffer_pages() / 20); | |
133 | recommended_min <<= (PAGE_SHIFT-10); | |
134 | ||
42aa83cb HP |
135 | if (recommended_min > min_free_kbytes) { |
136 | if (user_min_free_kbytes >= 0) | |
137 | pr_info("raising min_free_kbytes from %d to %lu " | |
138 | "to help transparent hugepage allocations\n", | |
139 | min_free_kbytes, recommended_min); | |
140 | ||
f000565a | 141 | min_free_kbytes = recommended_min; |
42aa83cb | 142 | } |
f000565a AA |
143 | setup_per_zone_wmarks(); |
144 | return 0; | |
145 | } | |
146 | late_initcall(set_recommended_min_free_kbytes); | |
147 | ||
ba76149f AA |
148 | static int start_khugepaged(void) |
149 | { | |
150 | int err = 0; | |
151 | if (khugepaged_enabled()) { | |
ba76149f AA |
152 | if (!khugepaged_thread) |
153 | khugepaged_thread = kthread_run(khugepaged, NULL, | |
154 | "khugepaged"); | |
155 | if (unlikely(IS_ERR(khugepaged_thread))) { | |
ae3a8c1c | 156 | pr_err("khugepaged: kthread_run(khugepaged) failed\n"); |
ba76149f AA |
157 | err = PTR_ERR(khugepaged_thread); |
158 | khugepaged_thread = NULL; | |
159 | } | |
911891af XG |
160 | |
161 | if (!list_empty(&khugepaged_scan.mm_head)) | |
ba76149f | 162 | wake_up_interruptible(&khugepaged_wait); |
f000565a AA |
163 | |
164 | set_recommended_min_free_kbytes(); | |
911891af | 165 | } else if (khugepaged_thread) { |
911891af XG |
166 | kthread_stop(khugepaged_thread); |
167 | khugepaged_thread = NULL; | |
168 | } | |
637e3a27 | 169 | |
ba76149f AA |
170 | return err; |
171 | } | |
71e3aac0 | 172 | |
97ae1749 | 173 | static atomic_t huge_zero_refcount; |
56873f43 | 174 | struct page *huge_zero_page __read_mostly; |
4a6c1297 | 175 | |
97ae1749 KS |
176 | static inline bool is_huge_zero_pmd(pmd_t pmd) |
177 | { | |
5918d10a | 178 | return is_huge_zero_page(pmd_page(pmd)); |
97ae1749 KS |
179 | } |
180 | ||
5918d10a | 181 | static struct page *get_huge_zero_page(void) |
97ae1749 KS |
182 | { |
183 | struct page *zero_page; | |
184 | retry: | |
185 | if (likely(atomic_inc_not_zero(&huge_zero_refcount))) | |
5918d10a | 186 | return ACCESS_ONCE(huge_zero_page); |
97ae1749 KS |
187 | |
188 | zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, | |
4a6c1297 | 189 | HPAGE_PMD_ORDER); |
d8a8e1f0 KS |
190 | if (!zero_page) { |
191 | count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); | |
5918d10a | 192 | return NULL; |
d8a8e1f0 KS |
193 | } |
194 | count_vm_event(THP_ZERO_PAGE_ALLOC); | |
97ae1749 | 195 | preempt_disable(); |
5918d10a | 196 | if (cmpxchg(&huge_zero_page, NULL, zero_page)) { |
97ae1749 | 197 | preempt_enable(); |
5ddacbe9 | 198 | __free_pages(zero_page, compound_order(zero_page)); |
97ae1749 KS |
199 | goto retry; |
200 | } | |
201 | ||
202 | /* We take additional reference here. It will be put back by shrinker */ | |
203 | atomic_set(&huge_zero_refcount, 2); | |
204 | preempt_enable(); | |
5918d10a | 205 | return ACCESS_ONCE(huge_zero_page); |
4a6c1297 KS |
206 | } |
207 | ||
97ae1749 | 208 | static void put_huge_zero_page(void) |
4a6c1297 | 209 | { |
97ae1749 KS |
210 | /* |
211 | * Counter should never go to zero here. Only shrinker can put | |
212 | * last reference. | |
213 | */ | |
214 | BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); | |
4a6c1297 KS |
215 | } |
216 | ||
48896466 GC |
217 | static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, |
218 | struct shrink_control *sc) | |
4a6c1297 | 219 | { |
48896466 GC |
220 | /* we can free zero page only if last reference remains */ |
221 | return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; | |
222 | } | |
97ae1749 | 223 | |
48896466 GC |
224 | static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, |
225 | struct shrink_control *sc) | |
226 | { | |
97ae1749 | 227 | if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { |
5918d10a KS |
228 | struct page *zero_page = xchg(&huge_zero_page, NULL); |
229 | BUG_ON(zero_page == NULL); | |
5ddacbe9 | 230 | __free_pages(zero_page, compound_order(zero_page)); |
48896466 | 231 | return HPAGE_PMD_NR; |
97ae1749 KS |
232 | } |
233 | ||
234 | return 0; | |
4a6c1297 KS |
235 | } |
236 | ||
97ae1749 | 237 | static struct shrinker huge_zero_page_shrinker = { |
48896466 GC |
238 | .count_objects = shrink_huge_zero_page_count, |
239 | .scan_objects = shrink_huge_zero_page_scan, | |
97ae1749 KS |
240 | .seeks = DEFAULT_SEEKS, |
241 | }; | |
242 | ||
71e3aac0 | 243 | #ifdef CONFIG_SYSFS |
ba76149f | 244 | |
71e3aac0 AA |
245 | static ssize_t double_flag_show(struct kobject *kobj, |
246 | struct kobj_attribute *attr, char *buf, | |
247 | enum transparent_hugepage_flag enabled, | |
248 | enum transparent_hugepage_flag req_madv) | |
249 | { | |
250 | if (test_bit(enabled, &transparent_hugepage_flags)) { | |
251 | VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags)); | |
252 | return sprintf(buf, "[always] madvise never\n"); | |
253 | } else if (test_bit(req_madv, &transparent_hugepage_flags)) | |
254 | return sprintf(buf, "always [madvise] never\n"); | |
255 | else | |
256 | return sprintf(buf, "always madvise [never]\n"); | |
257 | } | |
258 | static ssize_t double_flag_store(struct kobject *kobj, | |
259 | struct kobj_attribute *attr, | |
260 | const char *buf, size_t count, | |
261 | enum transparent_hugepage_flag enabled, | |
262 | enum transparent_hugepage_flag req_madv) | |
263 | { | |
264 | if (!memcmp("always", buf, | |
265 | min(sizeof("always")-1, count))) { | |
266 | set_bit(enabled, &transparent_hugepage_flags); | |
267 | clear_bit(req_madv, &transparent_hugepage_flags); | |
268 | } else if (!memcmp("madvise", buf, | |
269 | min(sizeof("madvise")-1, count))) { | |
270 | clear_bit(enabled, &transparent_hugepage_flags); | |
271 | set_bit(req_madv, &transparent_hugepage_flags); | |
272 | } else if (!memcmp("never", buf, | |
273 | min(sizeof("never")-1, count))) { | |
274 | clear_bit(enabled, &transparent_hugepage_flags); | |
275 | clear_bit(req_madv, &transparent_hugepage_flags); | |
276 | } else | |
277 | return -EINVAL; | |
278 | ||
279 | return count; | |
280 | } | |
281 | ||
282 | static ssize_t enabled_show(struct kobject *kobj, | |
283 | struct kobj_attribute *attr, char *buf) | |
284 | { | |
285 | return double_flag_show(kobj, attr, buf, | |
286 | TRANSPARENT_HUGEPAGE_FLAG, | |
287 | TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); | |
288 | } | |
289 | static ssize_t enabled_store(struct kobject *kobj, | |
290 | struct kobj_attribute *attr, | |
291 | const char *buf, size_t count) | |
292 | { | |
ba76149f AA |
293 | ssize_t ret; |
294 | ||
295 | ret = double_flag_store(kobj, attr, buf, count, | |
296 | TRANSPARENT_HUGEPAGE_FLAG, | |
297 | TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); | |
298 | ||
299 | if (ret > 0) { | |
911891af XG |
300 | int err; |
301 | ||
302 | mutex_lock(&khugepaged_mutex); | |
303 | err = start_khugepaged(); | |
304 | mutex_unlock(&khugepaged_mutex); | |
305 | ||
ba76149f AA |
306 | if (err) |
307 | ret = err; | |
308 | } | |
309 | ||
310 | return ret; | |
71e3aac0 AA |
311 | } |
312 | static struct kobj_attribute enabled_attr = | |
313 | __ATTR(enabled, 0644, enabled_show, enabled_store); | |
314 | ||
315 | static ssize_t single_flag_show(struct kobject *kobj, | |
316 | struct kobj_attribute *attr, char *buf, | |
317 | enum transparent_hugepage_flag flag) | |
318 | { | |
e27e6151 BH |
319 | return sprintf(buf, "%d\n", |
320 | !!test_bit(flag, &transparent_hugepage_flags)); | |
71e3aac0 | 321 | } |
e27e6151 | 322 | |
71e3aac0 AA |
323 | static ssize_t single_flag_store(struct kobject *kobj, |
324 | struct kobj_attribute *attr, | |
325 | const char *buf, size_t count, | |
326 | enum transparent_hugepage_flag flag) | |
327 | { | |
e27e6151 BH |
328 | unsigned long value; |
329 | int ret; | |
330 | ||
331 | ret = kstrtoul(buf, 10, &value); | |
332 | if (ret < 0) | |
333 | return ret; | |
334 | if (value > 1) | |
335 | return -EINVAL; | |
336 | ||
337 | if (value) | |
71e3aac0 | 338 | set_bit(flag, &transparent_hugepage_flags); |
e27e6151 | 339 | else |
71e3aac0 | 340 | clear_bit(flag, &transparent_hugepage_flags); |
71e3aac0 AA |
341 | |
342 | return count; | |
343 | } | |
344 | ||
345 | /* | |
346 | * Currently defrag only disables __GFP_NOWAIT for allocation. A blind | |
347 | * __GFP_REPEAT is too aggressive, it's never worth swapping tons of | |
348 | * memory just to allocate one more hugepage. | |
349 | */ | |
350 | static ssize_t defrag_show(struct kobject *kobj, | |
351 | struct kobj_attribute *attr, char *buf) | |
352 | { | |
353 | return double_flag_show(kobj, attr, buf, | |
354 | TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, | |
355 | TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); | |
356 | } | |
357 | static ssize_t defrag_store(struct kobject *kobj, | |
358 | struct kobj_attribute *attr, | |
359 | const char *buf, size_t count) | |
360 | { | |
361 | return double_flag_store(kobj, attr, buf, count, | |
362 | TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, | |
363 | TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); | |
364 | } | |
365 | static struct kobj_attribute defrag_attr = | |
366 | __ATTR(defrag, 0644, defrag_show, defrag_store); | |
367 | ||
79da5407 KS |
368 | static ssize_t use_zero_page_show(struct kobject *kobj, |
369 | struct kobj_attribute *attr, char *buf) | |
370 | { | |
371 | return single_flag_show(kobj, attr, buf, | |
372 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
373 | } | |
374 | static ssize_t use_zero_page_store(struct kobject *kobj, | |
375 | struct kobj_attribute *attr, const char *buf, size_t count) | |
376 | { | |
377 | return single_flag_store(kobj, attr, buf, count, | |
378 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
379 | } | |
380 | static struct kobj_attribute use_zero_page_attr = | |
381 | __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); | |
71e3aac0 AA |
382 | #ifdef CONFIG_DEBUG_VM |
383 | static ssize_t debug_cow_show(struct kobject *kobj, | |
384 | struct kobj_attribute *attr, char *buf) | |
385 | { | |
386 | return single_flag_show(kobj, attr, buf, | |
387 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); | |
388 | } | |
389 | static ssize_t debug_cow_store(struct kobject *kobj, | |
390 | struct kobj_attribute *attr, | |
391 | const char *buf, size_t count) | |
392 | { | |
393 | return single_flag_store(kobj, attr, buf, count, | |
394 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); | |
395 | } | |
396 | static struct kobj_attribute debug_cow_attr = | |
397 | __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); | |
398 | #endif /* CONFIG_DEBUG_VM */ | |
399 | ||
400 | static struct attribute *hugepage_attr[] = { | |
401 | &enabled_attr.attr, | |
402 | &defrag_attr.attr, | |
79da5407 | 403 | &use_zero_page_attr.attr, |
71e3aac0 AA |
404 | #ifdef CONFIG_DEBUG_VM |
405 | &debug_cow_attr.attr, | |
406 | #endif | |
407 | NULL, | |
408 | }; | |
409 | ||
410 | static struct attribute_group hugepage_attr_group = { | |
411 | .attrs = hugepage_attr, | |
ba76149f AA |
412 | }; |
413 | ||
414 | static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, | |
415 | struct kobj_attribute *attr, | |
416 | char *buf) | |
417 | { | |
418 | return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs); | |
419 | } | |
420 | ||
421 | static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, | |
422 | struct kobj_attribute *attr, | |
423 | const char *buf, size_t count) | |
424 | { | |
425 | unsigned long msecs; | |
426 | int err; | |
427 | ||
3dbb95f7 | 428 | err = kstrtoul(buf, 10, &msecs); |
ba76149f AA |
429 | if (err || msecs > UINT_MAX) |
430 | return -EINVAL; | |
431 | ||
432 | khugepaged_scan_sleep_millisecs = msecs; | |
433 | wake_up_interruptible(&khugepaged_wait); | |
434 | ||
435 | return count; | |
436 | } | |
437 | static struct kobj_attribute scan_sleep_millisecs_attr = | |
438 | __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show, | |
439 | scan_sleep_millisecs_store); | |
440 | ||
441 | static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, | |
442 | struct kobj_attribute *attr, | |
443 | char *buf) | |
444 | { | |
445 | return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs); | |
446 | } | |
447 | ||
448 | static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, | |
449 | struct kobj_attribute *attr, | |
450 | const char *buf, size_t count) | |
451 | { | |
452 | unsigned long msecs; | |
453 | int err; | |
454 | ||
3dbb95f7 | 455 | err = kstrtoul(buf, 10, &msecs); |
ba76149f AA |
456 | if (err || msecs > UINT_MAX) |
457 | return -EINVAL; | |
458 | ||
459 | khugepaged_alloc_sleep_millisecs = msecs; | |
460 | wake_up_interruptible(&khugepaged_wait); | |
461 | ||
462 | return count; | |
463 | } | |
464 | static struct kobj_attribute alloc_sleep_millisecs_attr = | |
465 | __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show, | |
466 | alloc_sleep_millisecs_store); | |
467 | ||
468 | static ssize_t pages_to_scan_show(struct kobject *kobj, | |
469 | struct kobj_attribute *attr, | |
470 | char *buf) | |
471 | { | |
472 | return sprintf(buf, "%u\n", khugepaged_pages_to_scan); | |
473 | } | |
474 | static ssize_t pages_to_scan_store(struct kobject *kobj, | |
475 | struct kobj_attribute *attr, | |
476 | const char *buf, size_t count) | |
477 | { | |
478 | int err; | |
479 | unsigned long pages; | |
480 | ||
3dbb95f7 | 481 | err = kstrtoul(buf, 10, &pages); |
ba76149f AA |
482 | if (err || !pages || pages > UINT_MAX) |
483 | return -EINVAL; | |
484 | ||
485 | khugepaged_pages_to_scan = pages; | |
486 | ||
487 | return count; | |
488 | } | |
489 | static struct kobj_attribute pages_to_scan_attr = | |
490 | __ATTR(pages_to_scan, 0644, pages_to_scan_show, | |
491 | pages_to_scan_store); | |
492 | ||
493 | static ssize_t pages_collapsed_show(struct kobject *kobj, | |
494 | struct kobj_attribute *attr, | |
495 | char *buf) | |
496 | { | |
497 | return sprintf(buf, "%u\n", khugepaged_pages_collapsed); | |
498 | } | |
499 | static struct kobj_attribute pages_collapsed_attr = | |
500 | __ATTR_RO(pages_collapsed); | |
501 | ||
502 | static ssize_t full_scans_show(struct kobject *kobj, | |
503 | struct kobj_attribute *attr, | |
504 | char *buf) | |
505 | { | |
506 | return sprintf(buf, "%u\n", khugepaged_full_scans); | |
507 | } | |
508 | static struct kobj_attribute full_scans_attr = | |
509 | __ATTR_RO(full_scans); | |
510 | ||
511 | static ssize_t khugepaged_defrag_show(struct kobject *kobj, | |
512 | struct kobj_attribute *attr, char *buf) | |
513 | { | |
514 | return single_flag_show(kobj, attr, buf, | |
515 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); | |
516 | } | |
517 | static ssize_t khugepaged_defrag_store(struct kobject *kobj, | |
518 | struct kobj_attribute *attr, | |
519 | const char *buf, size_t count) | |
520 | { | |
521 | return single_flag_store(kobj, attr, buf, count, | |
522 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); | |
523 | } | |
524 | static struct kobj_attribute khugepaged_defrag_attr = | |
525 | __ATTR(defrag, 0644, khugepaged_defrag_show, | |
526 | khugepaged_defrag_store); | |
527 | ||
528 | /* | |
529 | * max_ptes_none controls if khugepaged should collapse hugepages over | |
530 | * any unmapped ptes in turn potentially increasing the memory | |
531 | * footprint of the vmas. When max_ptes_none is 0 khugepaged will not | |
532 | * reduce the available free memory in the system as it | |
533 | * runs. Increasing max_ptes_none will instead potentially reduce the | |
534 | * free memory in the system during the khugepaged scan. | |
535 | */ | |
536 | static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj, | |
537 | struct kobj_attribute *attr, | |
538 | char *buf) | |
539 | { | |
540 | return sprintf(buf, "%u\n", khugepaged_max_ptes_none); | |
541 | } | |
542 | static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj, | |
543 | struct kobj_attribute *attr, | |
544 | const char *buf, size_t count) | |
545 | { | |
546 | int err; | |
547 | unsigned long max_ptes_none; | |
548 | ||
3dbb95f7 | 549 | err = kstrtoul(buf, 10, &max_ptes_none); |
ba76149f AA |
550 | if (err || max_ptes_none > HPAGE_PMD_NR-1) |
551 | return -EINVAL; | |
552 | ||
553 | khugepaged_max_ptes_none = max_ptes_none; | |
554 | ||
555 | return count; | |
556 | } | |
557 | static struct kobj_attribute khugepaged_max_ptes_none_attr = | |
558 | __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show, | |
559 | khugepaged_max_ptes_none_store); | |
560 | ||
561 | static struct attribute *khugepaged_attr[] = { | |
562 | &khugepaged_defrag_attr.attr, | |
563 | &khugepaged_max_ptes_none_attr.attr, | |
564 | &pages_to_scan_attr.attr, | |
565 | &pages_collapsed_attr.attr, | |
566 | &full_scans_attr.attr, | |
567 | &scan_sleep_millisecs_attr.attr, | |
568 | &alloc_sleep_millisecs_attr.attr, | |
569 | NULL, | |
570 | }; | |
571 | ||
572 | static struct attribute_group khugepaged_attr_group = { | |
573 | .attrs = khugepaged_attr, | |
574 | .name = "khugepaged", | |
71e3aac0 | 575 | }; |
71e3aac0 | 576 | |
569e5590 | 577 | static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) |
71e3aac0 | 578 | { |
71e3aac0 AA |
579 | int err; |
580 | ||
569e5590 SL |
581 | *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); |
582 | if (unlikely(!*hugepage_kobj)) { | |
ae3a8c1c | 583 | pr_err("failed to create transparent hugepage kobject\n"); |
569e5590 | 584 | return -ENOMEM; |
ba76149f AA |
585 | } |
586 | ||
569e5590 | 587 | err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); |
ba76149f | 588 | if (err) { |
ae3a8c1c | 589 | pr_err("failed to register transparent hugepage group\n"); |
569e5590 | 590 | goto delete_obj; |
ba76149f AA |
591 | } |
592 | ||
569e5590 | 593 | err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); |
ba76149f | 594 | if (err) { |
ae3a8c1c | 595 | pr_err("failed to register transparent hugepage group\n"); |
569e5590 | 596 | goto remove_hp_group; |
ba76149f | 597 | } |
569e5590 SL |
598 | |
599 | return 0; | |
600 | ||
601 | remove_hp_group: | |
602 | sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); | |
603 | delete_obj: | |
604 | kobject_put(*hugepage_kobj); | |
605 | return err; | |
606 | } | |
607 | ||
608 | static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
609 | { | |
610 | sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); | |
611 | sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); | |
612 | kobject_put(hugepage_kobj); | |
613 | } | |
614 | #else | |
615 | static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) | |
616 | { | |
617 | return 0; | |
618 | } | |
619 | ||
620 | static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
621 | { | |
622 | } | |
623 | #endif /* CONFIG_SYSFS */ | |
624 | ||
625 | static int __init hugepage_init(void) | |
626 | { | |
627 | int err; | |
628 | struct kobject *hugepage_kobj; | |
629 | ||
630 | if (!has_transparent_hugepage()) { | |
631 | transparent_hugepage_flags = 0; | |
632 | return -EINVAL; | |
633 | } | |
634 | ||
635 | err = hugepage_init_sysfs(&hugepage_kobj); | |
636 | if (err) | |
637 | return err; | |
ba76149f AA |
638 | |
639 | err = khugepaged_slab_init(); | |
640 | if (err) | |
641 | goto out; | |
642 | ||
97ae1749 KS |
643 | register_shrinker(&huge_zero_page_shrinker); |
644 | ||
97562cd2 RR |
645 | /* |
646 | * By default disable transparent hugepages on smaller systems, | |
647 | * where the extra memory used could hurt more than TLB overhead | |
648 | * is likely to save. The admin can still enable it through /sys. | |
649 | */ | |
650 | if (totalram_pages < (512 << (20 - PAGE_SHIFT))) | |
651 | transparent_hugepage_flags = 0; | |
652 | ||
ba76149f AA |
653 | start_khugepaged(); |
654 | ||
569e5590 | 655 | return 0; |
ba76149f | 656 | out: |
569e5590 | 657 | hugepage_exit_sysfs(hugepage_kobj); |
ba76149f | 658 | return err; |
71e3aac0 | 659 | } |
a64fb3cd | 660 | subsys_initcall(hugepage_init); |
71e3aac0 AA |
661 | |
662 | static int __init setup_transparent_hugepage(char *str) | |
663 | { | |
664 | int ret = 0; | |
665 | if (!str) | |
666 | goto out; | |
667 | if (!strcmp(str, "always")) { | |
668 | set_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
669 | &transparent_hugepage_flags); | |
670 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
671 | &transparent_hugepage_flags); | |
672 | ret = 1; | |
673 | } else if (!strcmp(str, "madvise")) { | |
674 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
675 | &transparent_hugepage_flags); | |
676 | set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
677 | &transparent_hugepage_flags); | |
678 | ret = 1; | |
679 | } else if (!strcmp(str, "never")) { | |
680 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
681 | &transparent_hugepage_flags); | |
682 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
683 | &transparent_hugepage_flags); | |
684 | ret = 1; | |
685 | } | |
686 | out: | |
687 | if (!ret) | |
ae3a8c1c | 688 | pr_warn("transparent_hugepage= cannot parse, ignored\n"); |
71e3aac0 AA |
689 | return ret; |
690 | } | |
691 | __setup("transparent_hugepage=", setup_transparent_hugepage); | |
692 | ||
b32967ff | 693 | pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) |
71e3aac0 AA |
694 | { |
695 | if (likely(vma->vm_flags & VM_WRITE)) | |
696 | pmd = pmd_mkwrite(pmd); | |
697 | return pmd; | |
698 | } | |
699 | ||
3122359a | 700 | static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot) |
b3092b3b BL |
701 | { |
702 | pmd_t entry; | |
3122359a | 703 | entry = mk_pmd(page, prot); |
b3092b3b BL |
704 | entry = pmd_mkhuge(entry); |
705 | return entry; | |
706 | } | |
707 | ||
71e3aac0 AA |
708 | static int __do_huge_pmd_anonymous_page(struct mm_struct *mm, |
709 | struct vm_area_struct *vma, | |
710 | unsigned long haddr, pmd_t *pmd, | |
711 | struct page *page) | |
712 | { | |
00501b53 | 713 | struct mem_cgroup *memcg; |
71e3aac0 | 714 | pgtable_t pgtable; |
c4088ebd | 715 | spinlock_t *ptl; |
71e3aac0 | 716 | |
309381fe | 717 | VM_BUG_ON_PAGE(!PageCompound(page), page); |
00501b53 JW |
718 | |
719 | if (mem_cgroup_try_charge(page, mm, GFP_TRANSHUGE, &memcg)) | |
720 | return VM_FAULT_OOM; | |
721 | ||
71e3aac0 | 722 | pgtable = pte_alloc_one(mm, haddr); |
00501b53 JW |
723 | if (unlikely(!pgtable)) { |
724 | mem_cgroup_cancel_charge(page, memcg); | |
71e3aac0 | 725 | return VM_FAULT_OOM; |
00501b53 | 726 | } |
71e3aac0 AA |
727 | |
728 | clear_huge_page(page, haddr, HPAGE_PMD_NR); | |
52f37629 MK |
729 | /* |
730 | * The memory barrier inside __SetPageUptodate makes sure that | |
731 | * clear_huge_page writes become visible before the set_pmd_at() | |
732 | * write. | |
733 | */ | |
71e3aac0 AA |
734 | __SetPageUptodate(page); |
735 | ||
c4088ebd | 736 | ptl = pmd_lock(mm, pmd); |
71e3aac0 | 737 | if (unlikely(!pmd_none(*pmd))) { |
c4088ebd | 738 | spin_unlock(ptl); |
00501b53 | 739 | mem_cgroup_cancel_charge(page, memcg); |
71e3aac0 AA |
740 | put_page(page); |
741 | pte_free(mm, pgtable); | |
742 | } else { | |
743 | pmd_t entry; | |
3122359a KS |
744 | entry = mk_huge_pmd(page, vma->vm_page_prot); |
745 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
71e3aac0 | 746 | page_add_new_anon_rmap(page, vma, haddr); |
00501b53 JW |
747 | mem_cgroup_commit_charge(page, memcg, false); |
748 | lru_cache_add_active_or_unevictable(page, vma); | |
6b0b50b0 | 749 | pgtable_trans_huge_deposit(mm, pmd, pgtable); |
71e3aac0 | 750 | set_pmd_at(mm, haddr, pmd, entry); |
71e3aac0 | 751 | add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); |
e1f56c89 | 752 | atomic_long_inc(&mm->nr_ptes); |
c4088ebd | 753 | spin_unlock(ptl); |
71e3aac0 AA |
754 | } |
755 | ||
aa2e878e | 756 | return 0; |
71e3aac0 AA |
757 | } |
758 | ||
cc5d462f | 759 | static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp) |
0bbbc0b3 | 760 | { |
cc5d462f | 761 | return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp; |
0bbbc0b3 AA |
762 | } |
763 | ||
c4088ebd | 764 | /* Caller must hold page table lock. */ |
3ea41e62 | 765 | static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, |
97ae1749 | 766 | struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, |
5918d10a | 767 | struct page *zero_page) |
fc9fe822 KS |
768 | { |
769 | pmd_t entry; | |
3ea41e62 KS |
770 | if (!pmd_none(*pmd)) |
771 | return false; | |
5918d10a | 772 | entry = mk_pmd(zero_page, vma->vm_page_prot); |
fc9fe822 | 773 | entry = pmd_mkhuge(entry); |
6b0b50b0 | 774 | pgtable_trans_huge_deposit(mm, pmd, pgtable); |
fc9fe822 | 775 | set_pmd_at(mm, haddr, pmd, entry); |
e1f56c89 | 776 | atomic_long_inc(&mm->nr_ptes); |
3ea41e62 | 777 | return true; |
fc9fe822 KS |
778 | } |
779 | ||
71e3aac0 AA |
780 | int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma, |
781 | unsigned long address, pmd_t *pmd, | |
782 | unsigned int flags) | |
783 | { | |
077fcf11 | 784 | gfp_t gfp; |
71e3aac0 AA |
785 | struct page *page; |
786 | unsigned long haddr = address & HPAGE_PMD_MASK; | |
71e3aac0 | 787 | |
128ec037 | 788 | if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end) |
c0292554 | 789 | return VM_FAULT_FALLBACK; |
128ec037 KS |
790 | if (unlikely(anon_vma_prepare(vma))) |
791 | return VM_FAULT_OOM; | |
6d50e60c | 792 | if (unlikely(khugepaged_enter(vma, vma->vm_flags))) |
128ec037 | 793 | return VM_FAULT_OOM; |
593befa6 | 794 | if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) && |
128ec037 | 795 | transparent_hugepage_use_zero_page()) { |
c4088ebd | 796 | spinlock_t *ptl; |
128ec037 KS |
797 | pgtable_t pgtable; |
798 | struct page *zero_page; | |
799 | bool set; | |
800 | pgtable = pte_alloc_one(mm, haddr); | |
801 | if (unlikely(!pgtable)) | |
ba76149f | 802 | return VM_FAULT_OOM; |
128ec037 KS |
803 | zero_page = get_huge_zero_page(); |
804 | if (unlikely(!zero_page)) { | |
805 | pte_free(mm, pgtable); | |
81ab4201 | 806 | count_vm_event(THP_FAULT_FALLBACK); |
c0292554 | 807 | return VM_FAULT_FALLBACK; |
b9bbfbe3 | 808 | } |
c4088ebd | 809 | ptl = pmd_lock(mm, pmd); |
128ec037 KS |
810 | set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd, |
811 | zero_page); | |
c4088ebd | 812 | spin_unlock(ptl); |
128ec037 KS |
813 | if (!set) { |
814 | pte_free(mm, pgtable); | |
815 | put_huge_zero_page(); | |
edad9d2c | 816 | } |
edad9d2c | 817 | return 0; |
71e3aac0 | 818 | } |
077fcf11 AK |
819 | gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0); |
820 | page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); | |
128ec037 KS |
821 | if (unlikely(!page)) { |
822 | count_vm_event(THP_FAULT_FALLBACK); | |
c0292554 | 823 | return VM_FAULT_FALLBACK; |
128ec037 | 824 | } |
128ec037 | 825 | if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) { |
128ec037 | 826 | put_page(page); |
17766dde | 827 | count_vm_event(THP_FAULT_FALLBACK); |
c0292554 | 828 | return VM_FAULT_FALLBACK; |
128ec037 KS |
829 | } |
830 | ||
17766dde | 831 | count_vm_event(THP_FAULT_ALLOC); |
128ec037 | 832 | return 0; |
71e3aac0 AA |
833 | } |
834 | ||
835 | int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, | |
836 | pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, | |
837 | struct vm_area_struct *vma) | |
838 | { | |
c4088ebd | 839 | spinlock_t *dst_ptl, *src_ptl; |
71e3aac0 AA |
840 | struct page *src_page; |
841 | pmd_t pmd; | |
842 | pgtable_t pgtable; | |
843 | int ret; | |
844 | ||
845 | ret = -ENOMEM; | |
846 | pgtable = pte_alloc_one(dst_mm, addr); | |
847 | if (unlikely(!pgtable)) | |
848 | goto out; | |
849 | ||
c4088ebd KS |
850 | dst_ptl = pmd_lock(dst_mm, dst_pmd); |
851 | src_ptl = pmd_lockptr(src_mm, src_pmd); | |
852 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
71e3aac0 AA |
853 | |
854 | ret = -EAGAIN; | |
855 | pmd = *src_pmd; | |
856 | if (unlikely(!pmd_trans_huge(pmd))) { | |
857 | pte_free(dst_mm, pgtable); | |
858 | goto out_unlock; | |
859 | } | |
fc9fe822 | 860 | /* |
c4088ebd | 861 | * When page table lock is held, the huge zero pmd should not be |
fc9fe822 KS |
862 | * under splitting since we don't split the page itself, only pmd to |
863 | * a page table. | |
864 | */ | |
865 | if (is_huge_zero_pmd(pmd)) { | |
5918d10a | 866 | struct page *zero_page; |
3ea41e62 | 867 | bool set; |
97ae1749 KS |
868 | /* |
869 | * get_huge_zero_page() will never allocate a new page here, | |
870 | * since we already have a zero page to copy. It just takes a | |
871 | * reference. | |
872 | */ | |
5918d10a | 873 | zero_page = get_huge_zero_page(); |
3ea41e62 | 874 | set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, |
5918d10a | 875 | zero_page); |
3ea41e62 | 876 | BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */ |
fc9fe822 KS |
877 | ret = 0; |
878 | goto out_unlock; | |
879 | } | |
de466bd6 | 880 | |
71e3aac0 AA |
881 | if (unlikely(pmd_trans_splitting(pmd))) { |
882 | /* split huge page running from under us */ | |
c4088ebd KS |
883 | spin_unlock(src_ptl); |
884 | spin_unlock(dst_ptl); | |
71e3aac0 AA |
885 | pte_free(dst_mm, pgtable); |
886 | ||
887 | wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */ | |
888 | goto out; | |
889 | } | |
890 | src_page = pmd_page(pmd); | |
309381fe | 891 | VM_BUG_ON_PAGE(!PageHead(src_page), src_page); |
71e3aac0 AA |
892 | get_page(src_page); |
893 | page_dup_rmap(src_page); | |
894 | add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); | |
895 | ||
896 | pmdp_set_wrprotect(src_mm, addr, src_pmd); | |
897 | pmd = pmd_mkold(pmd_wrprotect(pmd)); | |
6b0b50b0 | 898 | pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); |
71e3aac0 | 899 | set_pmd_at(dst_mm, addr, dst_pmd, pmd); |
e1f56c89 | 900 | atomic_long_inc(&dst_mm->nr_ptes); |
71e3aac0 AA |
901 | |
902 | ret = 0; | |
903 | out_unlock: | |
c4088ebd KS |
904 | spin_unlock(src_ptl); |
905 | spin_unlock(dst_ptl); | |
71e3aac0 AA |
906 | out: |
907 | return ret; | |
908 | } | |
909 | ||
a1dd450b WD |
910 | void huge_pmd_set_accessed(struct mm_struct *mm, |
911 | struct vm_area_struct *vma, | |
912 | unsigned long address, | |
913 | pmd_t *pmd, pmd_t orig_pmd, | |
914 | int dirty) | |
915 | { | |
c4088ebd | 916 | spinlock_t *ptl; |
a1dd450b WD |
917 | pmd_t entry; |
918 | unsigned long haddr; | |
919 | ||
c4088ebd | 920 | ptl = pmd_lock(mm, pmd); |
a1dd450b WD |
921 | if (unlikely(!pmd_same(*pmd, orig_pmd))) |
922 | goto unlock; | |
923 | ||
924 | entry = pmd_mkyoung(orig_pmd); | |
925 | haddr = address & HPAGE_PMD_MASK; | |
926 | if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty)) | |
927 | update_mmu_cache_pmd(vma, address, pmd); | |
928 | ||
929 | unlock: | |
c4088ebd | 930 | spin_unlock(ptl); |
a1dd450b WD |
931 | } |
932 | ||
5338a937 HD |
933 | /* |
934 | * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages | |
935 | * during copy_user_huge_page()'s copy_page_rep(): in the case when | |
936 | * the source page gets split and a tail freed before copy completes. | |
937 | * Called under pmd_lock of checked pmd, so safe from splitting itself. | |
938 | */ | |
939 | static void get_user_huge_page(struct page *page) | |
940 | { | |
941 | if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) { | |
942 | struct page *endpage = page + HPAGE_PMD_NR; | |
943 | ||
944 | atomic_add(HPAGE_PMD_NR, &page->_count); | |
945 | while (++page < endpage) | |
946 | get_huge_page_tail(page); | |
947 | } else { | |
948 | get_page(page); | |
949 | } | |
950 | } | |
951 | ||
952 | static void put_user_huge_page(struct page *page) | |
953 | { | |
954 | if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) { | |
955 | struct page *endpage = page + HPAGE_PMD_NR; | |
956 | ||
957 | while (page < endpage) | |
958 | put_page(page++); | |
959 | } else { | |
960 | put_page(page); | |
961 | } | |
962 | } | |
963 | ||
71e3aac0 AA |
964 | static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm, |
965 | struct vm_area_struct *vma, | |
966 | unsigned long address, | |
967 | pmd_t *pmd, pmd_t orig_pmd, | |
968 | struct page *page, | |
969 | unsigned long haddr) | |
970 | { | |
00501b53 | 971 | struct mem_cgroup *memcg; |
c4088ebd | 972 | spinlock_t *ptl; |
71e3aac0 AA |
973 | pgtable_t pgtable; |
974 | pmd_t _pmd; | |
975 | int ret = 0, i; | |
976 | struct page **pages; | |
2ec74c3e SG |
977 | unsigned long mmun_start; /* For mmu_notifiers */ |
978 | unsigned long mmun_end; /* For mmu_notifiers */ | |
71e3aac0 AA |
979 | |
980 | pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR, | |
981 | GFP_KERNEL); | |
982 | if (unlikely(!pages)) { | |
983 | ret |= VM_FAULT_OOM; | |
984 | goto out; | |
985 | } | |
986 | ||
987 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
cc5d462f AK |
988 | pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE | |
989 | __GFP_OTHER_NODE, | |
19ee151e | 990 | vma, address, page_to_nid(page)); |
b9bbfbe3 | 991 | if (unlikely(!pages[i] || |
00501b53 JW |
992 | mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL, |
993 | &memcg))) { | |
b9bbfbe3 | 994 | if (pages[i]) |
71e3aac0 | 995 | put_page(pages[i]); |
b9bbfbe3 | 996 | while (--i >= 0) { |
00501b53 JW |
997 | memcg = (void *)page_private(pages[i]); |
998 | set_page_private(pages[i], 0); | |
999 | mem_cgroup_cancel_charge(pages[i], memcg); | |
b9bbfbe3 AA |
1000 | put_page(pages[i]); |
1001 | } | |
71e3aac0 AA |
1002 | kfree(pages); |
1003 | ret |= VM_FAULT_OOM; | |
1004 | goto out; | |
1005 | } | |
00501b53 | 1006 | set_page_private(pages[i], (unsigned long)memcg); |
71e3aac0 AA |
1007 | } |
1008 | ||
1009 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
1010 | copy_user_highpage(pages[i], page + i, | |
0089e485 | 1011 | haddr + PAGE_SIZE * i, vma); |
71e3aac0 AA |
1012 | __SetPageUptodate(pages[i]); |
1013 | cond_resched(); | |
1014 | } | |
1015 | ||
2ec74c3e SG |
1016 | mmun_start = haddr; |
1017 | mmun_end = haddr + HPAGE_PMD_SIZE; | |
1018 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
1019 | ||
c4088ebd | 1020 | ptl = pmd_lock(mm, pmd); |
71e3aac0 AA |
1021 | if (unlikely(!pmd_same(*pmd, orig_pmd))) |
1022 | goto out_free_pages; | |
309381fe | 1023 | VM_BUG_ON_PAGE(!PageHead(page), page); |
71e3aac0 | 1024 | |
34ee645e | 1025 | pmdp_clear_flush_notify(vma, haddr, pmd); |
71e3aac0 AA |
1026 | /* leave pmd empty until pte is filled */ |
1027 | ||
6b0b50b0 | 1028 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); |
71e3aac0 AA |
1029 | pmd_populate(mm, &_pmd, pgtable); |
1030 | ||
1031 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
1032 | pte_t *pte, entry; | |
1033 | entry = mk_pte(pages[i], vma->vm_page_prot); | |
1034 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); | |
00501b53 JW |
1035 | memcg = (void *)page_private(pages[i]); |
1036 | set_page_private(pages[i], 0); | |
71e3aac0 | 1037 | page_add_new_anon_rmap(pages[i], vma, haddr); |
00501b53 JW |
1038 | mem_cgroup_commit_charge(pages[i], memcg, false); |
1039 | lru_cache_add_active_or_unevictable(pages[i], vma); | |
71e3aac0 AA |
1040 | pte = pte_offset_map(&_pmd, haddr); |
1041 | VM_BUG_ON(!pte_none(*pte)); | |
1042 | set_pte_at(mm, haddr, pte, entry); | |
1043 | pte_unmap(pte); | |
1044 | } | |
1045 | kfree(pages); | |
1046 | ||
71e3aac0 AA |
1047 | smp_wmb(); /* make pte visible before pmd */ |
1048 | pmd_populate(mm, pmd, pgtable); | |
1049 | page_remove_rmap(page); | |
c4088ebd | 1050 | spin_unlock(ptl); |
71e3aac0 | 1051 | |
2ec74c3e SG |
1052 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
1053 | ||
71e3aac0 AA |
1054 | ret |= VM_FAULT_WRITE; |
1055 | put_page(page); | |
1056 | ||
1057 | out: | |
1058 | return ret; | |
1059 | ||
1060 | out_free_pages: | |
c4088ebd | 1061 | spin_unlock(ptl); |
2ec74c3e | 1062 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
b9bbfbe3 | 1063 | for (i = 0; i < HPAGE_PMD_NR; i++) { |
00501b53 JW |
1064 | memcg = (void *)page_private(pages[i]); |
1065 | set_page_private(pages[i], 0); | |
1066 | mem_cgroup_cancel_charge(pages[i], memcg); | |
71e3aac0 | 1067 | put_page(pages[i]); |
b9bbfbe3 | 1068 | } |
71e3aac0 AA |
1069 | kfree(pages); |
1070 | goto out; | |
1071 | } | |
1072 | ||
1073 | int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma, | |
1074 | unsigned long address, pmd_t *pmd, pmd_t orig_pmd) | |
1075 | { | |
c4088ebd | 1076 | spinlock_t *ptl; |
71e3aac0 | 1077 | int ret = 0; |
93b4796d | 1078 | struct page *page = NULL, *new_page; |
00501b53 | 1079 | struct mem_cgroup *memcg; |
71e3aac0 | 1080 | unsigned long haddr; |
2ec74c3e SG |
1081 | unsigned long mmun_start; /* For mmu_notifiers */ |
1082 | unsigned long mmun_end; /* For mmu_notifiers */ | |
71e3aac0 | 1083 | |
c4088ebd | 1084 | ptl = pmd_lockptr(mm, pmd); |
81d1b09c | 1085 | VM_BUG_ON_VMA(!vma->anon_vma, vma); |
93b4796d KS |
1086 | haddr = address & HPAGE_PMD_MASK; |
1087 | if (is_huge_zero_pmd(orig_pmd)) | |
1088 | goto alloc; | |
c4088ebd | 1089 | spin_lock(ptl); |
71e3aac0 AA |
1090 | if (unlikely(!pmd_same(*pmd, orig_pmd))) |
1091 | goto out_unlock; | |
1092 | ||
1093 | page = pmd_page(orig_pmd); | |
309381fe | 1094 | VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); |
71e3aac0 AA |
1095 | if (page_mapcount(page) == 1) { |
1096 | pmd_t entry; | |
1097 | entry = pmd_mkyoung(orig_pmd); | |
1098 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
1099 | if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1)) | |
b113da65 | 1100 | update_mmu_cache_pmd(vma, address, pmd); |
71e3aac0 AA |
1101 | ret |= VM_FAULT_WRITE; |
1102 | goto out_unlock; | |
1103 | } | |
5338a937 | 1104 | get_user_huge_page(page); |
c4088ebd | 1105 | spin_unlock(ptl); |
93b4796d | 1106 | alloc: |
71e3aac0 | 1107 | if (transparent_hugepage_enabled(vma) && |
077fcf11 AK |
1108 | !transparent_hugepage_debug_cow()) { |
1109 | gfp_t gfp; | |
1110 | ||
1111 | gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0); | |
1112 | new_page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); | |
1113 | } else | |
71e3aac0 AA |
1114 | new_page = NULL; |
1115 | ||
1116 | if (unlikely(!new_page)) { | |
eecc1e42 | 1117 | if (!page) { |
e9b71ca9 KS |
1118 | split_huge_page_pmd(vma, address, pmd); |
1119 | ret |= VM_FAULT_FALLBACK; | |
93b4796d KS |
1120 | } else { |
1121 | ret = do_huge_pmd_wp_page_fallback(mm, vma, address, | |
1122 | pmd, orig_pmd, page, haddr); | |
9845cbbd | 1123 | if (ret & VM_FAULT_OOM) { |
93b4796d | 1124 | split_huge_page(page); |
9845cbbd KS |
1125 | ret |= VM_FAULT_FALLBACK; |
1126 | } | |
5338a937 | 1127 | put_user_huge_page(page); |
93b4796d | 1128 | } |
17766dde | 1129 | count_vm_event(THP_FAULT_FALLBACK); |
71e3aac0 AA |
1130 | goto out; |
1131 | } | |
1132 | ||
00501b53 JW |
1133 | if (unlikely(mem_cgroup_try_charge(new_page, mm, |
1134 | GFP_TRANSHUGE, &memcg))) { | |
b9bbfbe3 | 1135 | put_page(new_page); |
93b4796d KS |
1136 | if (page) { |
1137 | split_huge_page(page); | |
5338a937 | 1138 | put_user_huge_page(page); |
9845cbbd KS |
1139 | } else |
1140 | split_huge_page_pmd(vma, address, pmd); | |
1141 | ret |= VM_FAULT_FALLBACK; | |
17766dde | 1142 | count_vm_event(THP_FAULT_FALLBACK); |
b9bbfbe3 AA |
1143 | goto out; |
1144 | } | |
1145 | ||
17766dde DR |
1146 | count_vm_event(THP_FAULT_ALLOC); |
1147 | ||
eecc1e42 | 1148 | if (!page) |
93b4796d KS |
1149 | clear_huge_page(new_page, haddr, HPAGE_PMD_NR); |
1150 | else | |
1151 | copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR); | |
71e3aac0 AA |
1152 | __SetPageUptodate(new_page); |
1153 | ||
2ec74c3e SG |
1154 | mmun_start = haddr; |
1155 | mmun_end = haddr + HPAGE_PMD_SIZE; | |
1156 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
1157 | ||
c4088ebd | 1158 | spin_lock(ptl); |
93b4796d | 1159 | if (page) |
5338a937 | 1160 | put_user_huge_page(page); |
b9bbfbe3 | 1161 | if (unlikely(!pmd_same(*pmd, orig_pmd))) { |
c4088ebd | 1162 | spin_unlock(ptl); |
00501b53 | 1163 | mem_cgroup_cancel_charge(new_page, memcg); |
71e3aac0 | 1164 | put_page(new_page); |
2ec74c3e | 1165 | goto out_mn; |
b9bbfbe3 | 1166 | } else { |
71e3aac0 | 1167 | pmd_t entry; |
3122359a KS |
1168 | entry = mk_huge_pmd(new_page, vma->vm_page_prot); |
1169 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
34ee645e | 1170 | pmdp_clear_flush_notify(vma, haddr, pmd); |
71e3aac0 | 1171 | page_add_new_anon_rmap(new_page, vma, haddr); |
00501b53 JW |
1172 | mem_cgroup_commit_charge(new_page, memcg, false); |
1173 | lru_cache_add_active_or_unevictable(new_page, vma); | |
71e3aac0 | 1174 | set_pmd_at(mm, haddr, pmd, entry); |
b113da65 | 1175 | update_mmu_cache_pmd(vma, address, pmd); |
eecc1e42 | 1176 | if (!page) { |
93b4796d | 1177 | add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); |
97ae1749 KS |
1178 | put_huge_zero_page(); |
1179 | } else { | |
309381fe | 1180 | VM_BUG_ON_PAGE(!PageHead(page), page); |
93b4796d KS |
1181 | page_remove_rmap(page); |
1182 | put_page(page); | |
1183 | } | |
71e3aac0 AA |
1184 | ret |= VM_FAULT_WRITE; |
1185 | } | |
c4088ebd | 1186 | spin_unlock(ptl); |
2ec74c3e SG |
1187 | out_mn: |
1188 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); | |
71e3aac0 AA |
1189 | out: |
1190 | return ret; | |
2ec74c3e | 1191 | out_unlock: |
c4088ebd | 1192 | spin_unlock(ptl); |
2ec74c3e | 1193 | return ret; |
71e3aac0 AA |
1194 | } |
1195 | ||
b676b293 | 1196 | struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, |
71e3aac0 AA |
1197 | unsigned long addr, |
1198 | pmd_t *pmd, | |
1199 | unsigned int flags) | |
1200 | { | |
b676b293 | 1201 | struct mm_struct *mm = vma->vm_mm; |
71e3aac0 AA |
1202 | struct page *page = NULL; |
1203 | ||
c4088ebd | 1204 | assert_spin_locked(pmd_lockptr(mm, pmd)); |
71e3aac0 AA |
1205 | |
1206 | if (flags & FOLL_WRITE && !pmd_write(*pmd)) | |
1207 | goto out; | |
1208 | ||
85facf25 KS |
1209 | /* Avoid dumping huge zero page */ |
1210 | if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) | |
1211 | return ERR_PTR(-EFAULT); | |
1212 | ||
2b4847e7 | 1213 | /* Full NUMA hinting faults to serialise migration in fault paths */ |
8a0516ed | 1214 | if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) |
2b4847e7 MG |
1215 | goto out; |
1216 | ||
71e3aac0 | 1217 | page = pmd_page(*pmd); |
309381fe | 1218 | VM_BUG_ON_PAGE(!PageHead(page), page); |
71e3aac0 AA |
1219 | if (flags & FOLL_TOUCH) { |
1220 | pmd_t _pmd; | |
1221 | /* | |
1222 | * We should set the dirty bit only for FOLL_WRITE but | |
1223 | * for now the dirty bit in the pmd is meaningless. | |
1224 | * And if the dirty bit will become meaningful and | |
1225 | * we'll only set it with FOLL_WRITE, an atomic | |
1226 | * set_bit will be required on the pmd to set the | |
1227 | * young bit, instead of the current set_pmd_at. | |
1228 | */ | |
1229 | _pmd = pmd_mkyoung(pmd_mkdirty(*pmd)); | |
8663890a AK |
1230 | if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, |
1231 | pmd, _pmd, 1)) | |
1232 | update_mmu_cache_pmd(vma, addr, pmd); | |
71e3aac0 | 1233 | } |
b676b293 DR |
1234 | if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { |
1235 | if (page->mapping && trylock_page(page)) { | |
1236 | lru_add_drain(); | |
1237 | if (page->mapping) | |
1238 | mlock_vma_page(page); | |
1239 | unlock_page(page); | |
1240 | } | |
1241 | } | |
71e3aac0 | 1242 | page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; |
309381fe | 1243 | VM_BUG_ON_PAGE(!PageCompound(page), page); |
71e3aac0 | 1244 | if (flags & FOLL_GET) |
70b50f94 | 1245 | get_page_foll(page); |
71e3aac0 AA |
1246 | |
1247 | out: | |
1248 | return page; | |
1249 | } | |
1250 | ||
d10e63f2 | 1251 | /* NUMA hinting page fault entry point for trans huge pmds */ |
4daae3b4 MG |
1252 | int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma, |
1253 | unsigned long addr, pmd_t pmd, pmd_t *pmdp) | |
d10e63f2 | 1254 | { |
c4088ebd | 1255 | spinlock_t *ptl; |
b8916634 | 1256 | struct anon_vma *anon_vma = NULL; |
b32967ff | 1257 | struct page *page; |
d10e63f2 | 1258 | unsigned long haddr = addr & HPAGE_PMD_MASK; |
8191acbd | 1259 | int page_nid = -1, this_nid = numa_node_id(); |
90572890 | 1260 | int target_nid, last_cpupid = -1; |
8191acbd MG |
1261 | bool page_locked; |
1262 | bool migrated = false; | |
6688cc05 | 1263 | int flags = 0; |
d10e63f2 | 1264 | |
c0e7cad9 MG |
1265 | /* A PROT_NONE fault should not end up here */ |
1266 | BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))); | |
1267 | ||
c4088ebd | 1268 | ptl = pmd_lock(mm, pmdp); |
d10e63f2 MG |
1269 | if (unlikely(!pmd_same(pmd, *pmdp))) |
1270 | goto out_unlock; | |
1271 | ||
de466bd6 MG |
1272 | /* |
1273 | * If there are potential migrations, wait for completion and retry | |
1274 | * without disrupting NUMA hinting information. Do not relock and | |
1275 | * check_same as the page may no longer be mapped. | |
1276 | */ | |
1277 | if (unlikely(pmd_trans_migrating(*pmdp))) { | |
5d833062 | 1278 | page = pmd_page(*pmdp); |
de466bd6 | 1279 | spin_unlock(ptl); |
5d833062 | 1280 | wait_on_page_locked(page); |
de466bd6 MG |
1281 | goto out; |
1282 | } | |
1283 | ||
d10e63f2 | 1284 | page = pmd_page(pmd); |
a1a46184 | 1285 | BUG_ON(is_huge_zero_page(page)); |
8191acbd | 1286 | page_nid = page_to_nid(page); |
90572890 | 1287 | last_cpupid = page_cpupid_last(page); |
03c5a6e1 | 1288 | count_vm_numa_event(NUMA_HINT_FAULTS); |
04bb2f94 | 1289 | if (page_nid == this_nid) { |
03c5a6e1 | 1290 | count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); |
04bb2f94 RR |
1291 | flags |= TNF_FAULT_LOCAL; |
1292 | } | |
4daae3b4 | 1293 | |
bea66fbd MG |
1294 | /* See similar comment in do_numa_page for explanation */ |
1295 | if (!(vma->vm_flags & VM_WRITE)) | |
6688cc05 PZ |
1296 | flags |= TNF_NO_GROUP; |
1297 | ||
ff9042b1 MG |
1298 | /* |
1299 | * Acquire the page lock to serialise THP migrations but avoid dropping | |
1300 | * page_table_lock if at all possible | |
1301 | */ | |
b8916634 MG |
1302 | page_locked = trylock_page(page); |
1303 | target_nid = mpol_misplaced(page, vma, haddr); | |
1304 | if (target_nid == -1) { | |
1305 | /* If the page was locked, there are no parallel migrations */ | |
a54a407f | 1306 | if (page_locked) |
b8916634 | 1307 | goto clear_pmdnuma; |
2b4847e7 | 1308 | } |
4daae3b4 | 1309 | |
de466bd6 | 1310 | /* Migration could have started since the pmd_trans_migrating check */ |
2b4847e7 | 1311 | if (!page_locked) { |
c4088ebd | 1312 | spin_unlock(ptl); |
b8916634 | 1313 | wait_on_page_locked(page); |
a54a407f | 1314 | page_nid = -1; |
b8916634 MG |
1315 | goto out; |
1316 | } | |
1317 | ||
2b4847e7 MG |
1318 | /* |
1319 | * Page is misplaced. Page lock serialises migrations. Acquire anon_vma | |
1320 | * to serialises splits | |
1321 | */ | |
b8916634 | 1322 | get_page(page); |
c4088ebd | 1323 | spin_unlock(ptl); |
b8916634 | 1324 | anon_vma = page_lock_anon_vma_read(page); |
4daae3b4 | 1325 | |
c69307d5 | 1326 | /* Confirm the PMD did not change while page_table_lock was released */ |
c4088ebd | 1327 | spin_lock(ptl); |
b32967ff MG |
1328 | if (unlikely(!pmd_same(pmd, *pmdp))) { |
1329 | unlock_page(page); | |
1330 | put_page(page); | |
a54a407f | 1331 | page_nid = -1; |
4daae3b4 | 1332 | goto out_unlock; |
b32967ff | 1333 | } |
ff9042b1 | 1334 | |
c3a489ca MG |
1335 | /* Bail if we fail to protect against THP splits for any reason */ |
1336 | if (unlikely(!anon_vma)) { | |
1337 | put_page(page); | |
1338 | page_nid = -1; | |
1339 | goto clear_pmdnuma; | |
1340 | } | |
1341 | ||
a54a407f MG |
1342 | /* |
1343 | * Migrate the THP to the requested node, returns with page unlocked | |
8a0516ed | 1344 | * and access rights restored. |
a54a407f | 1345 | */ |
c4088ebd | 1346 | spin_unlock(ptl); |
b32967ff | 1347 | migrated = migrate_misplaced_transhuge_page(mm, vma, |
340ef390 | 1348 | pmdp, pmd, addr, page, target_nid); |
6688cc05 PZ |
1349 | if (migrated) { |
1350 | flags |= TNF_MIGRATED; | |
8191acbd | 1351 | page_nid = target_nid; |
6688cc05 | 1352 | } |
b32967ff | 1353 | |
8191acbd | 1354 | goto out; |
b32967ff | 1355 | clear_pmdnuma: |
a54a407f | 1356 | BUG_ON(!PageLocked(page)); |
4d942466 | 1357 | pmd = pmd_modify(pmd, vma->vm_page_prot); |
d10e63f2 | 1358 | set_pmd_at(mm, haddr, pmdp, pmd); |
d10e63f2 | 1359 | update_mmu_cache_pmd(vma, addr, pmdp); |
a54a407f | 1360 | unlock_page(page); |
d10e63f2 | 1361 | out_unlock: |
c4088ebd | 1362 | spin_unlock(ptl); |
b8916634 MG |
1363 | |
1364 | out: | |
1365 | if (anon_vma) | |
1366 | page_unlock_anon_vma_read(anon_vma); | |
1367 | ||
8191acbd | 1368 | if (page_nid != -1) |
6688cc05 | 1369 | task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags); |
8191acbd | 1370 | |
d10e63f2 MG |
1371 | return 0; |
1372 | } | |
1373 | ||
71e3aac0 | 1374 | int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, |
f21760b1 | 1375 | pmd_t *pmd, unsigned long addr) |
71e3aac0 | 1376 | { |
bf929152 | 1377 | spinlock_t *ptl; |
71e3aac0 AA |
1378 | int ret = 0; |
1379 | ||
bf929152 | 1380 | if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) { |
025c5b24 NH |
1381 | struct page *page; |
1382 | pgtable_t pgtable; | |
f5c8ad47 | 1383 | pmd_t orig_pmd; |
a6bf2bb0 AK |
1384 | /* |
1385 | * For architectures like ppc64 we look at deposited pgtable | |
1386 | * when calling pmdp_get_and_clear. So do the | |
1387 | * pgtable_trans_huge_withdraw after finishing pmdp related | |
1388 | * operations. | |
1389 | */ | |
fcbe08d6 MS |
1390 | orig_pmd = pmdp_get_and_clear_full(tlb->mm, addr, pmd, |
1391 | tlb->fullmm); | |
025c5b24 | 1392 | tlb_remove_pmd_tlb_entry(tlb, pmd, addr); |
a6bf2bb0 | 1393 | pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd); |
479f0abb | 1394 | if (is_huge_zero_pmd(orig_pmd)) { |
e1f56c89 | 1395 | atomic_long_dec(&tlb->mm->nr_ptes); |
bf929152 | 1396 | spin_unlock(ptl); |
97ae1749 | 1397 | put_huge_zero_page(); |
479f0abb KS |
1398 | } else { |
1399 | page = pmd_page(orig_pmd); | |
1400 | page_remove_rmap(page); | |
309381fe | 1401 | VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); |
479f0abb | 1402 | add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); |
309381fe | 1403 | VM_BUG_ON_PAGE(!PageHead(page), page); |
e1f56c89 | 1404 | atomic_long_dec(&tlb->mm->nr_ptes); |
bf929152 | 1405 | spin_unlock(ptl); |
479f0abb KS |
1406 | tlb_remove_page(tlb, page); |
1407 | } | |
025c5b24 NH |
1408 | pte_free(tlb->mm, pgtable); |
1409 | ret = 1; | |
1410 | } | |
71e3aac0 AA |
1411 | return ret; |
1412 | } | |
1413 | ||
37a1c49a AA |
1414 | int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma, |
1415 | unsigned long old_addr, | |
1416 | unsigned long new_addr, unsigned long old_end, | |
1417 | pmd_t *old_pmd, pmd_t *new_pmd) | |
1418 | { | |
bf929152 | 1419 | spinlock_t *old_ptl, *new_ptl; |
37a1c49a AA |
1420 | int ret = 0; |
1421 | pmd_t pmd; | |
1422 | ||
1423 | struct mm_struct *mm = vma->vm_mm; | |
1424 | ||
1425 | if ((old_addr & ~HPAGE_PMD_MASK) || | |
1426 | (new_addr & ~HPAGE_PMD_MASK) || | |
1427 | old_end - old_addr < HPAGE_PMD_SIZE || | |
1428 | (new_vma->vm_flags & VM_NOHUGEPAGE)) | |
1429 | goto out; | |
1430 | ||
1431 | /* | |
1432 | * The destination pmd shouldn't be established, free_pgtables() | |
1433 | * should have release it. | |
1434 | */ | |
1435 | if (WARN_ON(!pmd_none(*new_pmd))) { | |
1436 | VM_BUG_ON(pmd_trans_huge(*new_pmd)); | |
1437 | goto out; | |
1438 | } | |
1439 | ||
bf929152 KS |
1440 | /* |
1441 | * We don't have to worry about the ordering of src and dst | |
1442 | * ptlocks because exclusive mmap_sem prevents deadlock. | |
1443 | */ | |
1444 | ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl); | |
025c5b24 | 1445 | if (ret == 1) { |
bf929152 KS |
1446 | new_ptl = pmd_lockptr(mm, new_pmd); |
1447 | if (new_ptl != old_ptl) | |
1448 | spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); | |
025c5b24 NH |
1449 | pmd = pmdp_get_and_clear(mm, old_addr, old_pmd); |
1450 | VM_BUG_ON(!pmd_none(*new_pmd)); | |
3592806c | 1451 | |
b3084f4d AK |
1452 | if (pmd_move_must_withdraw(new_ptl, old_ptl)) { |
1453 | pgtable_t pgtable; | |
3592806c KS |
1454 | pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); |
1455 | pgtable_trans_huge_deposit(mm, new_pmd, pgtable); | |
3592806c | 1456 | } |
b3084f4d AK |
1457 | set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd)); |
1458 | if (new_ptl != old_ptl) | |
1459 | spin_unlock(new_ptl); | |
bf929152 | 1460 | spin_unlock(old_ptl); |
37a1c49a AA |
1461 | } |
1462 | out: | |
1463 | return ret; | |
1464 | } | |
1465 | ||
f123d74a MG |
1466 | /* |
1467 | * Returns | |
1468 | * - 0 if PMD could not be locked | |
1469 | * - 1 if PMD was locked but protections unchange and TLB flush unnecessary | |
1470 | * - HPAGE_PMD_NR is protections changed and TLB flush necessary | |
1471 | */ | |
cd7548ab | 1472 | int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, |
e944fd67 | 1473 | unsigned long addr, pgprot_t newprot, int prot_numa) |
cd7548ab JW |
1474 | { |
1475 | struct mm_struct *mm = vma->vm_mm; | |
bf929152 | 1476 | spinlock_t *ptl; |
cd7548ab JW |
1477 | int ret = 0; |
1478 | ||
bf929152 | 1479 | if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) { |
025c5b24 | 1480 | pmd_t entry; |
ba68bc01 | 1481 | ret = 1; |
e944fd67 MG |
1482 | |
1483 | /* | |
1484 | * Avoid trapping faults against the zero page. The read-only | |
1485 | * data is likely to be read-cached on the local CPU and | |
1486 | * local/remote hits to the zero page are not interesting. | |
1487 | */ | |
1488 | if (prot_numa && is_huge_zero_pmd(*pmd)) { | |
1489 | spin_unlock(ptl); | |
ba68bc01 | 1490 | return ret; |
e944fd67 MG |
1491 | } |
1492 | ||
10c1045f | 1493 | if (!prot_numa || !pmd_protnone(*pmd)) { |
10c1045f MG |
1494 | entry = pmdp_get_and_clear_notify(mm, addr, pmd); |
1495 | entry = pmd_modify(entry, newprot); | |
1496 | ret = HPAGE_PMD_NR; | |
1497 | set_pmd_at(mm, addr, pmd, entry); | |
1498 | BUG_ON(pmd_write(entry)); | |
1499 | } | |
bf929152 | 1500 | spin_unlock(ptl); |
025c5b24 NH |
1501 | } |
1502 | ||
1503 | return ret; | |
1504 | } | |
1505 | ||
1506 | /* | |
1507 | * Returns 1 if a given pmd maps a stable (not under splitting) thp. | |
1508 | * Returns -1 if it maps a thp under splitting. Returns 0 otherwise. | |
1509 | * | |
1510 | * Note that if it returns 1, this routine returns without unlocking page | |
1511 | * table locks. So callers must unlock them. | |
1512 | */ | |
bf929152 KS |
1513 | int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma, |
1514 | spinlock_t **ptl) | |
025c5b24 | 1515 | { |
bf929152 | 1516 | *ptl = pmd_lock(vma->vm_mm, pmd); |
cd7548ab JW |
1517 | if (likely(pmd_trans_huge(*pmd))) { |
1518 | if (unlikely(pmd_trans_splitting(*pmd))) { | |
bf929152 | 1519 | spin_unlock(*ptl); |
cd7548ab | 1520 | wait_split_huge_page(vma->anon_vma, pmd); |
025c5b24 | 1521 | return -1; |
cd7548ab | 1522 | } else { |
025c5b24 NH |
1523 | /* Thp mapped by 'pmd' is stable, so we can |
1524 | * handle it as it is. */ | |
1525 | return 1; | |
cd7548ab | 1526 | } |
025c5b24 | 1527 | } |
bf929152 | 1528 | spin_unlock(*ptl); |
025c5b24 | 1529 | return 0; |
cd7548ab JW |
1530 | } |
1531 | ||
117b0791 KS |
1532 | /* |
1533 | * This function returns whether a given @page is mapped onto the @address | |
1534 | * in the virtual space of @mm. | |
1535 | * | |
1536 | * When it's true, this function returns *pmd with holding the page table lock | |
1537 | * and passing it back to the caller via @ptl. | |
1538 | * If it's false, returns NULL without holding the page table lock. | |
1539 | */ | |
71e3aac0 AA |
1540 | pmd_t *page_check_address_pmd(struct page *page, |
1541 | struct mm_struct *mm, | |
1542 | unsigned long address, | |
117b0791 KS |
1543 | enum page_check_address_pmd_flag flag, |
1544 | spinlock_t **ptl) | |
71e3aac0 | 1545 | { |
b5a8cad3 KS |
1546 | pgd_t *pgd; |
1547 | pud_t *pud; | |
117b0791 | 1548 | pmd_t *pmd; |
71e3aac0 AA |
1549 | |
1550 | if (address & ~HPAGE_PMD_MASK) | |
117b0791 | 1551 | return NULL; |
71e3aac0 | 1552 | |
b5a8cad3 KS |
1553 | pgd = pgd_offset(mm, address); |
1554 | if (!pgd_present(*pgd)) | |
117b0791 | 1555 | return NULL; |
b5a8cad3 KS |
1556 | pud = pud_offset(pgd, address); |
1557 | if (!pud_present(*pud)) | |
1558 | return NULL; | |
1559 | pmd = pmd_offset(pud, address); | |
1560 | ||
117b0791 | 1561 | *ptl = pmd_lock(mm, pmd); |
b5a8cad3 | 1562 | if (!pmd_present(*pmd)) |
117b0791 | 1563 | goto unlock; |
71e3aac0 | 1564 | if (pmd_page(*pmd) != page) |
117b0791 | 1565 | goto unlock; |
94fcc585 AA |
1566 | /* |
1567 | * split_vma() may create temporary aliased mappings. There is | |
1568 | * no risk as long as all huge pmd are found and have their | |
1569 | * splitting bit set before __split_huge_page_refcount | |
1570 | * runs. Finding the same huge pmd more than once during the | |
1571 | * same rmap walk is not a problem. | |
1572 | */ | |
1573 | if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG && | |
1574 | pmd_trans_splitting(*pmd)) | |
117b0791 | 1575 | goto unlock; |
71e3aac0 AA |
1576 | if (pmd_trans_huge(*pmd)) { |
1577 | VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG && | |
1578 | !pmd_trans_splitting(*pmd)); | |
117b0791 | 1579 | return pmd; |
71e3aac0 | 1580 | } |
117b0791 KS |
1581 | unlock: |
1582 | spin_unlock(*ptl); | |
1583 | return NULL; | |
71e3aac0 AA |
1584 | } |
1585 | ||
1586 | static int __split_huge_page_splitting(struct page *page, | |
1587 | struct vm_area_struct *vma, | |
1588 | unsigned long address) | |
1589 | { | |
1590 | struct mm_struct *mm = vma->vm_mm; | |
117b0791 | 1591 | spinlock_t *ptl; |
71e3aac0 AA |
1592 | pmd_t *pmd; |
1593 | int ret = 0; | |
2ec74c3e SG |
1594 | /* For mmu_notifiers */ |
1595 | const unsigned long mmun_start = address; | |
1596 | const unsigned long mmun_end = address + HPAGE_PMD_SIZE; | |
71e3aac0 | 1597 | |
2ec74c3e | 1598 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
71e3aac0 | 1599 | pmd = page_check_address_pmd(page, mm, address, |
117b0791 | 1600 | PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl); |
71e3aac0 AA |
1601 | if (pmd) { |
1602 | /* | |
1603 | * We can't temporarily set the pmd to null in order | |
1604 | * to split it, the pmd must remain marked huge at all | |
1605 | * times or the VM won't take the pmd_trans_huge paths | |
5a505085 | 1606 | * and it won't wait on the anon_vma->root->rwsem to |
71e3aac0 AA |
1607 | * serialize against split_huge_page*. |
1608 | */ | |
2ec74c3e | 1609 | pmdp_splitting_flush(vma, address, pmd); |
34ee645e | 1610 | |
71e3aac0 | 1611 | ret = 1; |
117b0791 | 1612 | spin_unlock(ptl); |
71e3aac0 | 1613 | } |
2ec74c3e | 1614 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
71e3aac0 AA |
1615 | |
1616 | return ret; | |
1617 | } | |
1618 | ||
5bc7b8ac SL |
1619 | static void __split_huge_page_refcount(struct page *page, |
1620 | struct list_head *list) | |
71e3aac0 AA |
1621 | { |
1622 | int i; | |
71e3aac0 | 1623 | struct zone *zone = page_zone(page); |
fa9add64 | 1624 | struct lruvec *lruvec; |
70b50f94 | 1625 | int tail_count = 0; |
71e3aac0 AA |
1626 | |
1627 | /* prevent PageLRU to go away from under us, and freeze lru stats */ | |
1628 | spin_lock_irq(&zone->lru_lock); | |
fa9add64 HD |
1629 | lruvec = mem_cgroup_page_lruvec(page, zone); |
1630 | ||
71e3aac0 | 1631 | compound_lock(page); |
e94c8a9c KH |
1632 | /* complete memcg works before add pages to LRU */ |
1633 | mem_cgroup_split_huge_fixup(page); | |
71e3aac0 | 1634 | |
45676885 | 1635 | for (i = HPAGE_PMD_NR - 1; i >= 1; i--) { |
71e3aac0 AA |
1636 | struct page *page_tail = page + i; |
1637 | ||
70b50f94 AA |
1638 | /* tail_page->_mapcount cannot change */ |
1639 | BUG_ON(page_mapcount(page_tail) < 0); | |
1640 | tail_count += page_mapcount(page_tail); | |
1641 | /* check for overflow */ | |
1642 | BUG_ON(tail_count < 0); | |
1643 | BUG_ON(atomic_read(&page_tail->_count) != 0); | |
1644 | /* | |
1645 | * tail_page->_count is zero and not changing from | |
1646 | * under us. But get_page_unless_zero() may be running | |
1647 | * from under us on the tail_page. If we used | |
1648 | * atomic_set() below instead of atomic_add(), we | |
1649 | * would then run atomic_set() concurrently with | |
1650 | * get_page_unless_zero(), and atomic_set() is | |
1651 | * implemented in C not using locked ops. spin_unlock | |
1652 | * on x86 sometime uses locked ops because of PPro | |
1653 | * errata 66, 92, so unless somebody can guarantee | |
1654 | * atomic_set() here would be safe on all archs (and | |
1655 | * not only on x86), it's safer to use atomic_add(). | |
1656 | */ | |
1657 | atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1, | |
1658 | &page_tail->_count); | |
71e3aac0 AA |
1659 | |
1660 | /* after clearing PageTail the gup refcount can be released */ | |
3a79d52a | 1661 | smp_mb__after_atomic(); |
71e3aac0 | 1662 | |
a6d30ddd JD |
1663 | /* |
1664 | * retain hwpoison flag of the poisoned tail page: | |
1665 | * fix for the unsuitable process killed on Guest Machine(KVM) | |
1666 | * by the memory-failure. | |
1667 | */ | |
1668 | page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON; | |
71e3aac0 AA |
1669 | page_tail->flags |= (page->flags & |
1670 | ((1L << PG_referenced) | | |
1671 | (1L << PG_swapbacked) | | |
1672 | (1L << PG_mlocked) | | |
e180cf80 KS |
1673 | (1L << PG_uptodate) | |
1674 | (1L << PG_active) | | |
1675 | (1L << PG_unevictable))); | |
71e3aac0 AA |
1676 | page_tail->flags |= (1L << PG_dirty); |
1677 | ||
70b50f94 | 1678 | /* clear PageTail before overwriting first_page */ |
71e3aac0 AA |
1679 | smp_wmb(); |
1680 | ||
1681 | /* | |
1682 | * __split_huge_page_splitting() already set the | |
1683 | * splitting bit in all pmd that could map this | |
1684 | * hugepage, that will ensure no CPU can alter the | |
1685 | * mapcount on the head page. The mapcount is only | |
1686 | * accounted in the head page and it has to be | |
1687 | * transferred to all tail pages in the below code. So | |
1688 | * for this code to be safe, the split the mapcount | |
1689 | * can't change. But that doesn't mean userland can't | |
1690 | * keep changing and reading the page contents while | |
1691 | * we transfer the mapcount, so the pmd splitting | |
1692 | * status is achieved setting a reserved bit in the | |
1693 | * pmd, not by clearing the present bit. | |
1694 | */ | |
71e3aac0 AA |
1695 | page_tail->_mapcount = page->_mapcount; |
1696 | ||
1697 | BUG_ON(page_tail->mapping); | |
1698 | page_tail->mapping = page->mapping; | |
1699 | ||
45676885 | 1700 | page_tail->index = page->index + i; |
90572890 | 1701 | page_cpupid_xchg_last(page_tail, page_cpupid_last(page)); |
71e3aac0 AA |
1702 | |
1703 | BUG_ON(!PageAnon(page_tail)); | |
1704 | BUG_ON(!PageUptodate(page_tail)); | |
1705 | BUG_ON(!PageDirty(page_tail)); | |
1706 | BUG_ON(!PageSwapBacked(page_tail)); | |
1707 | ||
5bc7b8ac | 1708 | lru_add_page_tail(page, page_tail, lruvec, list); |
71e3aac0 | 1709 | } |
70b50f94 AA |
1710 | atomic_sub(tail_count, &page->_count); |
1711 | BUG_ON(atomic_read(&page->_count) <= 0); | |
71e3aac0 | 1712 | |
fa9add64 | 1713 | __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1); |
79134171 | 1714 | |
71e3aac0 AA |
1715 | ClearPageCompound(page); |
1716 | compound_unlock(page); | |
1717 | spin_unlock_irq(&zone->lru_lock); | |
1718 | ||
1719 | for (i = 1; i < HPAGE_PMD_NR; i++) { | |
1720 | struct page *page_tail = page + i; | |
1721 | BUG_ON(page_count(page_tail) <= 0); | |
1722 | /* | |
1723 | * Tail pages may be freed if there wasn't any mapping | |
1724 | * like if add_to_swap() is running on a lru page that | |
1725 | * had its mapping zapped. And freeing these pages | |
1726 | * requires taking the lru_lock so we do the put_page | |
1727 | * of the tail pages after the split is complete. | |
1728 | */ | |
1729 | put_page(page_tail); | |
1730 | } | |
1731 | ||
1732 | /* | |
1733 | * Only the head page (now become a regular page) is required | |
1734 | * to be pinned by the caller. | |
1735 | */ | |
1736 | BUG_ON(page_count(page) <= 0); | |
1737 | } | |
1738 | ||
1739 | static int __split_huge_page_map(struct page *page, | |
1740 | struct vm_area_struct *vma, | |
1741 | unsigned long address) | |
1742 | { | |
1743 | struct mm_struct *mm = vma->vm_mm; | |
117b0791 | 1744 | spinlock_t *ptl; |
71e3aac0 AA |
1745 | pmd_t *pmd, _pmd; |
1746 | int ret = 0, i; | |
1747 | pgtable_t pgtable; | |
1748 | unsigned long haddr; | |
1749 | ||
71e3aac0 | 1750 | pmd = page_check_address_pmd(page, mm, address, |
117b0791 | 1751 | PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl); |
71e3aac0 | 1752 | if (pmd) { |
6b0b50b0 | 1753 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); |
71e3aac0 | 1754 | pmd_populate(mm, &_pmd, pgtable); |
f8303c25 WL |
1755 | if (pmd_write(*pmd)) |
1756 | BUG_ON(page_mapcount(page) != 1); | |
71e3aac0 | 1757 | |
e3ebcf64 GS |
1758 | haddr = address; |
1759 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
71e3aac0 AA |
1760 | pte_t *pte, entry; |
1761 | BUG_ON(PageCompound(page+i)); | |
abc40bd2 | 1762 | /* |
8a0516ed MG |
1763 | * Note that NUMA hinting access restrictions are not |
1764 | * transferred to avoid any possibility of altering | |
1765 | * permissions across VMAs. | |
abc40bd2 | 1766 | */ |
71e3aac0 AA |
1767 | entry = mk_pte(page + i, vma->vm_page_prot); |
1768 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); | |
1769 | if (!pmd_write(*pmd)) | |
1770 | entry = pte_wrprotect(entry); | |
71e3aac0 AA |
1771 | if (!pmd_young(*pmd)) |
1772 | entry = pte_mkold(entry); | |
1773 | pte = pte_offset_map(&_pmd, haddr); | |
1774 | BUG_ON(!pte_none(*pte)); | |
1775 | set_pte_at(mm, haddr, pte, entry); | |
1776 | pte_unmap(pte); | |
1777 | } | |
1778 | ||
71e3aac0 AA |
1779 | smp_wmb(); /* make pte visible before pmd */ |
1780 | /* | |
1781 | * Up to this point the pmd is present and huge and | |
1782 | * userland has the whole access to the hugepage | |
1783 | * during the split (which happens in place). If we | |
1784 | * overwrite the pmd with the not-huge version | |
1785 | * pointing to the pte here (which of course we could | |
1786 | * if all CPUs were bug free), userland could trigger | |
1787 | * a small page size TLB miss on the small sized TLB | |
1788 | * while the hugepage TLB entry is still established | |
1789 | * in the huge TLB. Some CPU doesn't like that. See | |
1790 | * http://support.amd.com/us/Processor_TechDocs/41322.pdf, | |
1791 | * Erratum 383 on page 93. Intel should be safe but is | |
1792 | * also warns that it's only safe if the permission | |
1793 | * and cache attributes of the two entries loaded in | |
1794 | * the two TLB is identical (which should be the case | |
1795 | * here). But it is generally safer to never allow | |
1796 | * small and huge TLB entries for the same virtual | |
1797 | * address to be loaded simultaneously. So instead of | |
1798 | * doing "pmd_populate(); flush_tlb_range();" we first | |
1799 | * mark the current pmd notpresent (atomically because | |
1800 | * here the pmd_trans_huge and pmd_trans_splitting | |
1801 | * must remain set at all times on the pmd until the | |
1802 | * split is complete for this pmd), then we flush the | |
1803 | * SMP TLB and finally we write the non-huge version | |
1804 | * of the pmd entry with pmd_populate. | |
1805 | */ | |
46dcde73 | 1806 | pmdp_invalidate(vma, address, pmd); |
71e3aac0 AA |
1807 | pmd_populate(mm, pmd, pgtable); |
1808 | ret = 1; | |
117b0791 | 1809 | spin_unlock(ptl); |
71e3aac0 | 1810 | } |
71e3aac0 AA |
1811 | |
1812 | return ret; | |
1813 | } | |
1814 | ||
5a505085 | 1815 | /* must be called with anon_vma->root->rwsem held */ |
71e3aac0 | 1816 | static void __split_huge_page(struct page *page, |
5bc7b8ac SL |
1817 | struct anon_vma *anon_vma, |
1818 | struct list_head *list) | |
71e3aac0 AA |
1819 | { |
1820 | int mapcount, mapcount2; | |
bf181b9f | 1821 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); |
71e3aac0 AA |
1822 | struct anon_vma_chain *avc; |
1823 | ||
1824 | BUG_ON(!PageHead(page)); | |
1825 | BUG_ON(PageTail(page)); | |
1826 | ||
1827 | mapcount = 0; | |
bf181b9f | 1828 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { |
71e3aac0 AA |
1829 | struct vm_area_struct *vma = avc->vma; |
1830 | unsigned long addr = vma_address(page, vma); | |
1831 | BUG_ON(is_vma_temporary_stack(vma)); | |
71e3aac0 AA |
1832 | mapcount += __split_huge_page_splitting(page, vma, addr); |
1833 | } | |
05759d38 AA |
1834 | /* |
1835 | * It is critical that new vmas are added to the tail of the | |
1836 | * anon_vma list. This guarantes that if copy_huge_pmd() runs | |
1837 | * and establishes a child pmd before | |
1838 | * __split_huge_page_splitting() freezes the parent pmd (so if | |
1839 | * we fail to prevent copy_huge_pmd() from running until the | |
1840 | * whole __split_huge_page() is complete), we will still see | |
1841 | * the newly established pmd of the child later during the | |
1842 | * walk, to be able to set it as pmd_trans_splitting too. | |
1843 | */ | |
ff9e43eb | 1844 | if (mapcount != page_mapcount(page)) { |
ae3a8c1c AM |
1845 | pr_err("mapcount %d page_mapcount %d\n", |
1846 | mapcount, page_mapcount(page)); | |
ff9e43eb KS |
1847 | BUG(); |
1848 | } | |
71e3aac0 | 1849 | |
5bc7b8ac | 1850 | __split_huge_page_refcount(page, list); |
71e3aac0 AA |
1851 | |
1852 | mapcount2 = 0; | |
bf181b9f | 1853 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { |
71e3aac0 AA |
1854 | struct vm_area_struct *vma = avc->vma; |
1855 | unsigned long addr = vma_address(page, vma); | |
1856 | BUG_ON(is_vma_temporary_stack(vma)); | |
71e3aac0 AA |
1857 | mapcount2 += __split_huge_page_map(page, vma, addr); |
1858 | } | |
ff9e43eb | 1859 | if (mapcount != mapcount2) { |
ae3a8c1c AM |
1860 | pr_err("mapcount %d mapcount2 %d page_mapcount %d\n", |
1861 | mapcount, mapcount2, page_mapcount(page)); | |
ff9e43eb KS |
1862 | BUG(); |
1863 | } | |
71e3aac0 AA |
1864 | } |
1865 | ||
5bc7b8ac SL |
1866 | /* |
1867 | * Split a hugepage into normal pages. This doesn't change the position of head | |
1868 | * page. If @list is null, tail pages will be added to LRU list, otherwise, to | |
1869 | * @list. Both head page and tail pages will inherit mapping, flags, and so on | |
1870 | * from the hugepage. | |
1871 | * Return 0 if the hugepage is split successfully otherwise return 1. | |
1872 | */ | |
1873 | int split_huge_page_to_list(struct page *page, struct list_head *list) | |
71e3aac0 AA |
1874 | { |
1875 | struct anon_vma *anon_vma; | |
1876 | int ret = 1; | |
1877 | ||
5918d10a | 1878 | BUG_ON(is_huge_zero_page(page)); |
71e3aac0 | 1879 | BUG_ON(!PageAnon(page)); |
062f1af2 MG |
1880 | |
1881 | /* | |
1882 | * The caller does not necessarily hold an mmap_sem that would prevent | |
1883 | * the anon_vma disappearing so we first we take a reference to it | |
1884 | * and then lock the anon_vma for write. This is similar to | |
1885 | * page_lock_anon_vma_read except the write lock is taken to serialise | |
1886 | * against parallel split or collapse operations. | |
1887 | */ | |
1888 | anon_vma = page_get_anon_vma(page); | |
71e3aac0 AA |
1889 | if (!anon_vma) |
1890 | goto out; | |
062f1af2 MG |
1891 | anon_vma_lock_write(anon_vma); |
1892 | ||
71e3aac0 AA |
1893 | ret = 0; |
1894 | if (!PageCompound(page)) | |
1895 | goto out_unlock; | |
1896 | ||
1897 | BUG_ON(!PageSwapBacked(page)); | |
5bc7b8ac | 1898 | __split_huge_page(page, anon_vma, list); |
81ab4201 | 1899 | count_vm_event(THP_SPLIT); |
71e3aac0 AA |
1900 | |
1901 | BUG_ON(PageCompound(page)); | |
1902 | out_unlock: | |
08b52706 | 1903 | anon_vma_unlock_write(anon_vma); |
062f1af2 | 1904 | put_anon_vma(anon_vma); |
71e3aac0 AA |
1905 | out: |
1906 | return ret; | |
1907 | } | |
1908 | ||
9050d7eb | 1909 | #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE) |
78f11a25 | 1910 | |
60ab3244 AA |
1911 | int hugepage_madvise(struct vm_area_struct *vma, |
1912 | unsigned long *vm_flags, int advice) | |
0af4e98b | 1913 | { |
a664b2d8 AA |
1914 | switch (advice) { |
1915 | case MADV_HUGEPAGE: | |
1e1836e8 AT |
1916 | #ifdef CONFIG_S390 |
1917 | /* | |
1918 | * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390 | |
1919 | * can't handle this properly after s390_enable_sie, so we simply | |
1920 | * ignore the madvise to prevent qemu from causing a SIGSEGV. | |
1921 | */ | |
1922 | if (mm_has_pgste(vma->vm_mm)) | |
1923 | return 0; | |
1924 | #endif | |
a664b2d8 AA |
1925 | /* |
1926 | * Be somewhat over-protective like KSM for now! | |
1927 | */ | |
78f11a25 | 1928 | if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP)) |
a664b2d8 AA |
1929 | return -EINVAL; |
1930 | *vm_flags &= ~VM_NOHUGEPAGE; | |
1931 | *vm_flags |= VM_HUGEPAGE; | |
60ab3244 AA |
1932 | /* |
1933 | * If the vma become good for khugepaged to scan, | |
1934 | * register it here without waiting a page fault that | |
1935 | * may not happen any time soon. | |
1936 | */ | |
6d50e60c | 1937 | if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags))) |
60ab3244 | 1938 | return -ENOMEM; |
a664b2d8 AA |
1939 | break; |
1940 | case MADV_NOHUGEPAGE: | |
1941 | /* | |
1942 | * Be somewhat over-protective like KSM for now! | |
1943 | */ | |
78f11a25 | 1944 | if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP)) |
a664b2d8 AA |
1945 | return -EINVAL; |
1946 | *vm_flags &= ~VM_HUGEPAGE; | |
1947 | *vm_flags |= VM_NOHUGEPAGE; | |
60ab3244 AA |
1948 | /* |
1949 | * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning | |
1950 | * this vma even if we leave the mm registered in khugepaged if | |
1951 | * it got registered before VM_NOHUGEPAGE was set. | |
1952 | */ | |
a664b2d8 AA |
1953 | break; |
1954 | } | |
0af4e98b AA |
1955 | |
1956 | return 0; | |
1957 | } | |
1958 | ||
ba76149f AA |
1959 | static int __init khugepaged_slab_init(void) |
1960 | { | |
1961 | mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", | |
1962 | sizeof(struct mm_slot), | |
1963 | __alignof__(struct mm_slot), 0, NULL); | |
1964 | if (!mm_slot_cache) | |
1965 | return -ENOMEM; | |
1966 | ||
1967 | return 0; | |
1968 | } | |
1969 | ||
ba76149f AA |
1970 | static inline struct mm_slot *alloc_mm_slot(void) |
1971 | { | |
1972 | if (!mm_slot_cache) /* initialization failed */ | |
1973 | return NULL; | |
1974 | return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); | |
1975 | } | |
1976 | ||
1977 | static inline void free_mm_slot(struct mm_slot *mm_slot) | |
1978 | { | |
1979 | kmem_cache_free(mm_slot_cache, mm_slot); | |
1980 | } | |
1981 | ||
ba76149f AA |
1982 | static struct mm_slot *get_mm_slot(struct mm_struct *mm) |
1983 | { | |
1984 | struct mm_slot *mm_slot; | |
ba76149f | 1985 | |
b67bfe0d | 1986 | hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm) |
ba76149f AA |
1987 | if (mm == mm_slot->mm) |
1988 | return mm_slot; | |
43b5fbbd | 1989 | |
ba76149f AA |
1990 | return NULL; |
1991 | } | |
1992 | ||
1993 | static void insert_to_mm_slots_hash(struct mm_struct *mm, | |
1994 | struct mm_slot *mm_slot) | |
1995 | { | |
ba76149f | 1996 | mm_slot->mm = mm; |
43b5fbbd | 1997 | hash_add(mm_slots_hash, &mm_slot->hash, (long)mm); |
ba76149f AA |
1998 | } |
1999 | ||
2000 | static inline int khugepaged_test_exit(struct mm_struct *mm) | |
2001 | { | |
2002 | return atomic_read(&mm->mm_users) == 0; | |
2003 | } | |
2004 | ||
2005 | int __khugepaged_enter(struct mm_struct *mm) | |
2006 | { | |
2007 | struct mm_slot *mm_slot; | |
2008 | int wakeup; | |
2009 | ||
2010 | mm_slot = alloc_mm_slot(); | |
2011 | if (!mm_slot) | |
2012 | return -ENOMEM; | |
2013 | ||
2014 | /* __khugepaged_exit() must not run from under us */ | |
96dad67f | 2015 | VM_BUG_ON_MM(khugepaged_test_exit(mm), mm); |
ba76149f AA |
2016 | if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { |
2017 | free_mm_slot(mm_slot); | |
2018 | return 0; | |
2019 | } | |
2020 | ||
2021 | spin_lock(&khugepaged_mm_lock); | |
2022 | insert_to_mm_slots_hash(mm, mm_slot); | |
2023 | /* | |
2024 | * Insert just behind the scanning cursor, to let the area settle | |
2025 | * down a little. | |
2026 | */ | |
2027 | wakeup = list_empty(&khugepaged_scan.mm_head); | |
2028 | list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head); | |
2029 | spin_unlock(&khugepaged_mm_lock); | |
2030 | ||
2031 | atomic_inc(&mm->mm_count); | |
2032 | if (wakeup) | |
2033 | wake_up_interruptible(&khugepaged_wait); | |
2034 | ||
2035 | return 0; | |
2036 | } | |
2037 | ||
6d50e60c DR |
2038 | int khugepaged_enter_vma_merge(struct vm_area_struct *vma, |
2039 | unsigned long vm_flags) | |
ba76149f AA |
2040 | { |
2041 | unsigned long hstart, hend; | |
2042 | if (!vma->anon_vma) | |
2043 | /* | |
2044 | * Not yet faulted in so we will register later in the | |
2045 | * page fault if needed. | |
2046 | */ | |
2047 | return 0; | |
78f11a25 | 2048 | if (vma->vm_ops) |
ba76149f AA |
2049 | /* khugepaged not yet working on file or special mappings */ |
2050 | return 0; | |
6d50e60c | 2051 | VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma); |
ba76149f AA |
2052 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; |
2053 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
2054 | if (hstart < hend) | |
6d50e60c | 2055 | return khugepaged_enter(vma, vm_flags); |
ba76149f AA |
2056 | return 0; |
2057 | } | |
2058 | ||
2059 | void __khugepaged_exit(struct mm_struct *mm) | |
2060 | { | |
2061 | struct mm_slot *mm_slot; | |
2062 | int free = 0; | |
2063 | ||
2064 | spin_lock(&khugepaged_mm_lock); | |
2065 | mm_slot = get_mm_slot(mm); | |
2066 | if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { | |
43b5fbbd | 2067 | hash_del(&mm_slot->hash); |
ba76149f AA |
2068 | list_del(&mm_slot->mm_node); |
2069 | free = 1; | |
2070 | } | |
d788e80a | 2071 | spin_unlock(&khugepaged_mm_lock); |
ba76149f AA |
2072 | |
2073 | if (free) { | |
ba76149f AA |
2074 | clear_bit(MMF_VM_HUGEPAGE, &mm->flags); |
2075 | free_mm_slot(mm_slot); | |
2076 | mmdrop(mm); | |
2077 | } else if (mm_slot) { | |
ba76149f AA |
2078 | /* |
2079 | * This is required to serialize against | |
2080 | * khugepaged_test_exit() (which is guaranteed to run | |
2081 | * under mmap sem read mode). Stop here (after we | |
2082 | * return all pagetables will be destroyed) until | |
2083 | * khugepaged has finished working on the pagetables | |
2084 | * under the mmap_sem. | |
2085 | */ | |
2086 | down_write(&mm->mmap_sem); | |
2087 | up_write(&mm->mmap_sem); | |
d788e80a | 2088 | } |
ba76149f AA |
2089 | } |
2090 | ||
2091 | static void release_pte_page(struct page *page) | |
2092 | { | |
2093 | /* 0 stands for page_is_file_cache(page) == false */ | |
2094 | dec_zone_page_state(page, NR_ISOLATED_ANON + 0); | |
2095 | unlock_page(page); | |
2096 | putback_lru_page(page); | |
2097 | } | |
2098 | ||
2099 | static void release_pte_pages(pte_t *pte, pte_t *_pte) | |
2100 | { | |
2101 | while (--_pte >= pte) { | |
2102 | pte_t pteval = *_pte; | |
2103 | if (!pte_none(pteval)) | |
2104 | release_pte_page(pte_page(pteval)); | |
2105 | } | |
2106 | } | |
2107 | ||
ba76149f AA |
2108 | static int __collapse_huge_page_isolate(struct vm_area_struct *vma, |
2109 | unsigned long address, | |
2110 | pte_t *pte) | |
2111 | { | |
2112 | struct page *page; | |
2113 | pte_t *_pte; | |
10359213 EA |
2114 | int none = 0; |
2115 | bool referenced = false, writable = false; | |
ba76149f AA |
2116 | for (_pte = pte; _pte < pte+HPAGE_PMD_NR; |
2117 | _pte++, address += PAGE_SIZE) { | |
2118 | pte_t pteval = *_pte; | |
2119 | if (pte_none(pteval)) { | |
2120 | if (++none <= khugepaged_max_ptes_none) | |
2121 | continue; | |
344aa35c | 2122 | else |
ba76149f | 2123 | goto out; |
ba76149f | 2124 | } |
10359213 | 2125 | if (!pte_present(pteval)) |
ba76149f | 2126 | goto out; |
ba76149f | 2127 | page = vm_normal_page(vma, address, pteval); |
344aa35c | 2128 | if (unlikely(!page)) |
ba76149f | 2129 | goto out; |
344aa35c | 2130 | |
309381fe SL |
2131 | VM_BUG_ON_PAGE(PageCompound(page), page); |
2132 | VM_BUG_ON_PAGE(!PageAnon(page), page); | |
2133 | VM_BUG_ON_PAGE(!PageSwapBacked(page), page); | |
ba76149f | 2134 | |
ba76149f AA |
2135 | /* |
2136 | * We can do it before isolate_lru_page because the | |
2137 | * page can't be freed from under us. NOTE: PG_lock | |
2138 | * is needed to serialize against split_huge_page | |
2139 | * when invoked from the VM. | |
2140 | */ | |
344aa35c | 2141 | if (!trylock_page(page)) |
ba76149f | 2142 | goto out; |
10359213 EA |
2143 | |
2144 | /* | |
2145 | * cannot use mapcount: can't collapse if there's a gup pin. | |
2146 | * The page must only be referenced by the scanned process | |
2147 | * and page swap cache. | |
2148 | */ | |
2149 | if (page_count(page) != 1 + !!PageSwapCache(page)) { | |
2150 | unlock_page(page); | |
2151 | goto out; | |
2152 | } | |
2153 | if (pte_write(pteval)) { | |
2154 | writable = true; | |
2155 | } else { | |
2156 | if (PageSwapCache(page) && !reuse_swap_page(page)) { | |
2157 | unlock_page(page); | |
2158 | goto out; | |
2159 | } | |
2160 | /* | |
2161 | * Page is not in the swap cache. It can be collapsed | |
2162 | * into a THP. | |
2163 | */ | |
2164 | } | |
2165 | ||
ba76149f AA |
2166 | /* |
2167 | * Isolate the page to avoid collapsing an hugepage | |
2168 | * currently in use by the VM. | |
2169 | */ | |
2170 | if (isolate_lru_page(page)) { | |
2171 | unlock_page(page); | |
ba76149f AA |
2172 | goto out; |
2173 | } | |
2174 | /* 0 stands for page_is_file_cache(page) == false */ | |
2175 | inc_zone_page_state(page, NR_ISOLATED_ANON + 0); | |
309381fe SL |
2176 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
2177 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
ba76149f AA |
2178 | |
2179 | /* If there is no mapped pte young don't collapse the page */ | |
8ee53820 AA |
2180 | if (pte_young(pteval) || PageReferenced(page) || |
2181 | mmu_notifier_test_young(vma->vm_mm, address)) | |
10359213 | 2182 | referenced = true; |
ba76149f | 2183 | } |
10359213 | 2184 | if (likely(referenced && writable)) |
344aa35c | 2185 | return 1; |
ba76149f | 2186 | out: |
344aa35c BL |
2187 | release_pte_pages(pte, _pte); |
2188 | return 0; | |
ba76149f AA |
2189 | } |
2190 | ||
2191 | static void __collapse_huge_page_copy(pte_t *pte, struct page *page, | |
2192 | struct vm_area_struct *vma, | |
2193 | unsigned long address, | |
2194 | spinlock_t *ptl) | |
2195 | { | |
2196 | pte_t *_pte; | |
2197 | for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) { | |
2198 | pte_t pteval = *_pte; | |
2199 | struct page *src_page; | |
2200 | ||
2201 | if (pte_none(pteval)) { | |
2202 | clear_user_highpage(page, address); | |
2203 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); | |
2204 | } else { | |
2205 | src_page = pte_page(pteval); | |
2206 | copy_user_highpage(page, src_page, address, vma); | |
309381fe | 2207 | VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page); |
ba76149f AA |
2208 | release_pte_page(src_page); |
2209 | /* | |
2210 | * ptl mostly unnecessary, but preempt has to | |
2211 | * be disabled to update the per-cpu stats | |
2212 | * inside page_remove_rmap(). | |
2213 | */ | |
2214 | spin_lock(ptl); | |
2215 | /* | |
2216 | * paravirt calls inside pte_clear here are | |
2217 | * superfluous. | |
2218 | */ | |
2219 | pte_clear(vma->vm_mm, address, _pte); | |
2220 | page_remove_rmap(src_page); | |
2221 | spin_unlock(ptl); | |
2222 | free_page_and_swap_cache(src_page); | |
2223 | } | |
2224 | ||
2225 | address += PAGE_SIZE; | |
2226 | page++; | |
2227 | } | |
2228 | } | |
2229 | ||
26234f36 | 2230 | static void khugepaged_alloc_sleep(void) |
ba76149f | 2231 | { |
26234f36 XG |
2232 | wait_event_freezable_timeout(khugepaged_wait, false, |
2233 | msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); | |
2234 | } | |
ba76149f | 2235 | |
9f1b868a BL |
2236 | static int khugepaged_node_load[MAX_NUMNODES]; |
2237 | ||
14a4e214 DR |
2238 | static bool khugepaged_scan_abort(int nid) |
2239 | { | |
2240 | int i; | |
2241 | ||
2242 | /* | |
2243 | * If zone_reclaim_mode is disabled, then no extra effort is made to | |
2244 | * allocate memory locally. | |
2245 | */ | |
2246 | if (!zone_reclaim_mode) | |
2247 | return false; | |
2248 | ||
2249 | /* If there is a count for this node already, it must be acceptable */ | |
2250 | if (khugepaged_node_load[nid]) | |
2251 | return false; | |
2252 | ||
2253 | for (i = 0; i < MAX_NUMNODES; i++) { | |
2254 | if (!khugepaged_node_load[i]) | |
2255 | continue; | |
2256 | if (node_distance(nid, i) > RECLAIM_DISTANCE) | |
2257 | return true; | |
2258 | } | |
2259 | return false; | |
2260 | } | |
2261 | ||
26234f36 | 2262 | #ifdef CONFIG_NUMA |
9f1b868a BL |
2263 | static int khugepaged_find_target_node(void) |
2264 | { | |
2265 | static int last_khugepaged_target_node = NUMA_NO_NODE; | |
2266 | int nid, target_node = 0, max_value = 0; | |
2267 | ||
2268 | /* find first node with max normal pages hit */ | |
2269 | for (nid = 0; nid < MAX_NUMNODES; nid++) | |
2270 | if (khugepaged_node_load[nid] > max_value) { | |
2271 | max_value = khugepaged_node_load[nid]; | |
2272 | target_node = nid; | |
2273 | } | |
2274 | ||
2275 | /* do some balance if several nodes have the same hit record */ | |
2276 | if (target_node <= last_khugepaged_target_node) | |
2277 | for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES; | |
2278 | nid++) | |
2279 | if (max_value == khugepaged_node_load[nid]) { | |
2280 | target_node = nid; | |
2281 | break; | |
2282 | } | |
2283 | ||
2284 | last_khugepaged_target_node = target_node; | |
2285 | return target_node; | |
2286 | } | |
2287 | ||
26234f36 XG |
2288 | static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) |
2289 | { | |
2290 | if (IS_ERR(*hpage)) { | |
2291 | if (!*wait) | |
2292 | return false; | |
2293 | ||
2294 | *wait = false; | |
e3b4126c | 2295 | *hpage = NULL; |
26234f36 XG |
2296 | khugepaged_alloc_sleep(); |
2297 | } else if (*hpage) { | |
2298 | put_page(*hpage); | |
2299 | *hpage = NULL; | |
2300 | } | |
2301 | ||
2302 | return true; | |
2303 | } | |
2304 | ||
2305 | static struct page | |
2306 | *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm, | |
2307 | struct vm_area_struct *vma, unsigned long address, | |
2308 | int node) | |
2309 | { | |
309381fe | 2310 | VM_BUG_ON_PAGE(*hpage, *hpage); |
8b164568 | 2311 | |
ce83d217 | 2312 | /* |
8b164568 VB |
2313 | * Before allocating the hugepage, release the mmap_sem read lock. |
2314 | * The allocation can take potentially a long time if it involves | |
2315 | * sync compaction, and we do not need to hold the mmap_sem during | |
2316 | * that. We will recheck the vma after taking it again in write mode. | |
ce83d217 | 2317 | */ |
8b164568 VB |
2318 | up_read(&mm->mmap_sem); |
2319 | ||
9f1b868a BL |
2320 | *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask( |
2321 | khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER); | |
26234f36 | 2322 | if (unlikely(!*hpage)) { |
81ab4201 | 2323 | count_vm_event(THP_COLLAPSE_ALLOC_FAILED); |
ce83d217 | 2324 | *hpage = ERR_PTR(-ENOMEM); |
26234f36 | 2325 | return NULL; |
ce83d217 | 2326 | } |
26234f36 | 2327 | |
65b3c07b | 2328 | count_vm_event(THP_COLLAPSE_ALLOC); |
26234f36 XG |
2329 | return *hpage; |
2330 | } | |
2331 | #else | |
9f1b868a BL |
2332 | static int khugepaged_find_target_node(void) |
2333 | { | |
2334 | return 0; | |
2335 | } | |
2336 | ||
10dc4155 BL |
2337 | static inline struct page *alloc_hugepage(int defrag) |
2338 | { | |
2339 | return alloc_pages(alloc_hugepage_gfpmask(defrag, 0), | |
2340 | HPAGE_PMD_ORDER); | |
2341 | } | |
2342 | ||
26234f36 XG |
2343 | static struct page *khugepaged_alloc_hugepage(bool *wait) |
2344 | { | |
2345 | struct page *hpage; | |
2346 | ||
2347 | do { | |
2348 | hpage = alloc_hugepage(khugepaged_defrag()); | |
2349 | if (!hpage) { | |
2350 | count_vm_event(THP_COLLAPSE_ALLOC_FAILED); | |
2351 | if (!*wait) | |
2352 | return NULL; | |
2353 | ||
2354 | *wait = false; | |
2355 | khugepaged_alloc_sleep(); | |
2356 | } else | |
2357 | count_vm_event(THP_COLLAPSE_ALLOC); | |
2358 | } while (unlikely(!hpage) && likely(khugepaged_enabled())); | |
2359 | ||
2360 | return hpage; | |
2361 | } | |
2362 | ||
2363 | static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) | |
2364 | { | |
2365 | if (!*hpage) | |
2366 | *hpage = khugepaged_alloc_hugepage(wait); | |
2367 | ||
2368 | if (unlikely(!*hpage)) | |
2369 | return false; | |
2370 | ||
2371 | return true; | |
2372 | } | |
2373 | ||
2374 | static struct page | |
2375 | *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm, | |
2376 | struct vm_area_struct *vma, unsigned long address, | |
2377 | int node) | |
2378 | { | |
2379 | up_read(&mm->mmap_sem); | |
2380 | VM_BUG_ON(!*hpage); | |
2381 | return *hpage; | |
2382 | } | |
692e0b35 AA |
2383 | #endif |
2384 | ||
fa475e51 BL |
2385 | static bool hugepage_vma_check(struct vm_area_struct *vma) |
2386 | { | |
2387 | if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) || | |
2388 | (vma->vm_flags & VM_NOHUGEPAGE)) | |
2389 | return false; | |
2390 | ||
2391 | if (!vma->anon_vma || vma->vm_ops) | |
2392 | return false; | |
2393 | if (is_vma_temporary_stack(vma)) | |
2394 | return false; | |
81d1b09c | 2395 | VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma); |
fa475e51 BL |
2396 | return true; |
2397 | } | |
2398 | ||
26234f36 XG |
2399 | static void collapse_huge_page(struct mm_struct *mm, |
2400 | unsigned long address, | |
2401 | struct page **hpage, | |
2402 | struct vm_area_struct *vma, | |
2403 | int node) | |
2404 | { | |
26234f36 XG |
2405 | pmd_t *pmd, _pmd; |
2406 | pte_t *pte; | |
2407 | pgtable_t pgtable; | |
2408 | struct page *new_page; | |
c4088ebd | 2409 | spinlock_t *pmd_ptl, *pte_ptl; |
26234f36 XG |
2410 | int isolated; |
2411 | unsigned long hstart, hend; | |
00501b53 | 2412 | struct mem_cgroup *memcg; |
2ec74c3e SG |
2413 | unsigned long mmun_start; /* For mmu_notifiers */ |
2414 | unsigned long mmun_end; /* For mmu_notifiers */ | |
26234f36 XG |
2415 | |
2416 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
2417 | ||
2418 | /* release the mmap_sem read lock. */ | |
2419 | new_page = khugepaged_alloc_page(hpage, mm, vma, address, node); | |
2420 | if (!new_page) | |
2421 | return; | |
2422 | ||
00501b53 JW |
2423 | if (unlikely(mem_cgroup_try_charge(new_page, mm, |
2424 | GFP_TRANSHUGE, &memcg))) | |
ce83d217 | 2425 | return; |
ba76149f AA |
2426 | |
2427 | /* | |
2428 | * Prevent all access to pagetables with the exception of | |
2429 | * gup_fast later hanlded by the ptep_clear_flush and the VM | |
2430 | * handled by the anon_vma lock + PG_lock. | |
2431 | */ | |
2432 | down_write(&mm->mmap_sem); | |
2433 | if (unlikely(khugepaged_test_exit(mm))) | |
2434 | goto out; | |
2435 | ||
2436 | vma = find_vma(mm, address); | |
a8f531eb L |
2437 | if (!vma) |
2438 | goto out; | |
ba76149f AA |
2439 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; |
2440 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
2441 | if (address < hstart || address + HPAGE_PMD_SIZE > hend) | |
2442 | goto out; | |
fa475e51 | 2443 | if (!hugepage_vma_check(vma)) |
a7d6e4ec | 2444 | goto out; |
6219049a BL |
2445 | pmd = mm_find_pmd(mm, address); |
2446 | if (!pmd) | |
ba76149f | 2447 | goto out; |
ba76149f | 2448 | |
4fc3f1d6 | 2449 | anon_vma_lock_write(vma->anon_vma); |
ba76149f AA |
2450 | |
2451 | pte = pte_offset_map(pmd, address); | |
c4088ebd | 2452 | pte_ptl = pte_lockptr(mm, pmd); |
ba76149f | 2453 | |
2ec74c3e SG |
2454 | mmun_start = address; |
2455 | mmun_end = address + HPAGE_PMD_SIZE; | |
2456 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
c4088ebd | 2457 | pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ |
ba76149f AA |
2458 | /* |
2459 | * After this gup_fast can't run anymore. This also removes | |
2460 | * any huge TLB entry from the CPU so we won't allow | |
2461 | * huge and small TLB entries for the same virtual address | |
2462 | * to avoid the risk of CPU bugs in that area. | |
2463 | */ | |
2ec74c3e | 2464 | _pmd = pmdp_clear_flush(vma, address, pmd); |
c4088ebd | 2465 | spin_unlock(pmd_ptl); |
2ec74c3e | 2466 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
ba76149f | 2467 | |
c4088ebd | 2468 | spin_lock(pte_ptl); |
ba76149f | 2469 | isolated = __collapse_huge_page_isolate(vma, address, pte); |
c4088ebd | 2470 | spin_unlock(pte_ptl); |
ba76149f AA |
2471 | |
2472 | if (unlikely(!isolated)) { | |
453c7192 | 2473 | pte_unmap(pte); |
c4088ebd | 2474 | spin_lock(pmd_ptl); |
ba76149f | 2475 | BUG_ON(!pmd_none(*pmd)); |
7c342512 AK |
2476 | /* |
2477 | * We can only use set_pmd_at when establishing | |
2478 | * hugepmds and never for establishing regular pmds that | |
2479 | * points to regular pagetables. Use pmd_populate for that | |
2480 | */ | |
2481 | pmd_populate(mm, pmd, pmd_pgtable(_pmd)); | |
c4088ebd | 2482 | spin_unlock(pmd_ptl); |
08b52706 | 2483 | anon_vma_unlock_write(vma->anon_vma); |
ce83d217 | 2484 | goto out; |
ba76149f AA |
2485 | } |
2486 | ||
2487 | /* | |
2488 | * All pages are isolated and locked so anon_vma rmap | |
2489 | * can't run anymore. | |
2490 | */ | |
08b52706 | 2491 | anon_vma_unlock_write(vma->anon_vma); |
ba76149f | 2492 | |
c4088ebd | 2493 | __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl); |
453c7192 | 2494 | pte_unmap(pte); |
ba76149f AA |
2495 | __SetPageUptodate(new_page); |
2496 | pgtable = pmd_pgtable(_pmd); | |
ba76149f | 2497 | |
3122359a KS |
2498 | _pmd = mk_huge_pmd(new_page, vma->vm_page_prot); |
2499 | _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); | |
ba76149f AA |
2500 | |
2501 | /* | |
2502 | * spin_lock() below is not the equivalent of smp_wmb(), so | |
2503 | * this is needed to avoid the copy_huge_page writes to become | |
2504 | * visible after the set_pmd_at() write. | |
2505 | */ | |
2506 | smp_wmb(); | |
2507 | ||
c4088ebd | 2508 | spin_lock(pmd_ptl); |
ba76149f AA |
2509 | BUG_ON(!pmd_none(*pmd)); |
2510 | page_add_new_anon_rmap(new_page, vma, address); | |
00501b53 JW |
2511 | mem_cgroup_commit_charge(new_page, memcg, false); |
2512 | lru_cache_add_active_or_unevictable(new_page, vma); | |
fce144b4 | 2513 | pgtable_trans_huge_deposit(mm, pmd, pgtable); |
ba76149f | 2514 | set_pmd_at(mm, address, pmd, _pmd); |
b113da65 | 2515 | update_mmu_cache_pmd(vma, address, pmd); |
c4088ebd | 2516 | spin_unlock(pmd_ptl); |
ba76149f AA |
2517 | |
2518 | *hpage = NULL; | |
420256ef | 2519 | |
ba76149f | 2520 | khugepaged_pages_collapsed++; |
ce83d217 | 2521 | out_up_write: |
ba76149f | 2522 | up_write(&mm->mmap_sem); |
0bbbc0b3 AA |
2523 | return; |
2524 | ||
ce83d217 | 2525 | out: |
00501b53 | 2526 | mem_cgroup_cancel_charge(new_page, memcg); |
ce83d217 | 2527 | goto out_up_write; |
ba76149f AA |
2528 | } |
2529 | ||
2530 | static int khugepaged_scan_pmd(struct mm_struct *mm, | |
2531 | struct vm_area_struct *vma, | |
2532 | unsigned long address, | |
2533 | struct page **hpage) | |
2534 | { | |
ba76149f AA |
2535 | pmd_t *pmd; |
2536 | pte_t *pte, *_pte; | |
10359213 | 2537 | int ret = 0, none = 0; |
ba76149f AA |
2538 | struct page *page; |
2539 | unsigned long _address; | |
2540 | spinlock_t *ptl; | |
00ef2d2f | 2541 | int node = NUMA_NO_NODE; |
10359213 | 2542 | bool writable = false, referenced = false; |
ba76149f AA |
2543 | |
2544 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
2545 | ||
6219049a BL |
2546 | pmd = mm_find_pmd(mm, address); |
2547 | if (!pmd) | |
ba76149f | 2548 | goto out; |
ba76149f | 2549 | |
9f1b868a | 2550 | memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load)); |
ba76149f AA |
2551 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); |
2552 | for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR; | |
2553 | _pte++, _address += PAGE_SIZE) { | |
2554 | pte_t pteval = *_pte; | |
2555 | if (pte_none(pteval)) { | |
2556 | if (++none <= khugepaged_max_ptes_none) | |
2557 | continue; | |
2558 | else | |
2559 | goto out_unmap; | |
2560 | } | |
10359213 | 2561 | if (!pte_present(pteval)) |
ba76149f | 2562 | goto out_unmap; |
10359213 EA |
2563 | if (pte_write(pteval)) |
2564 | writable = true; | |
2565 | ||
ba76149f AA |
2566 | page = vm_normal_page(vma, _address, pteval); |
2567 | if (unlikely(!page)) | |
2568 | goto out_unmap; | |
5c4b4be3 | 2569 | /* |
9f1b868a BL |
2570 | * Record which node the original page is from and save this |
2571 | * information to khugepaged_node_load[]. | |
2572 | * Khupaged will allocate hugepage from the node has the max | |
2573 | * hit record. | |
5c4b4be3 | 2574 | */ |
9f1b868a | 2575 | node = page_to_nid(page); |
14a4e214 DR |
2576 | if (khugepaged_scan_abort(node)) |
2577 | goto out_unmap; | |
9f1b868a | 2578 | khugepaged_node_load[node]++; |
309381fe | 2579 | VM_BUG_ON_PAGE(PageCompound(page), page); |
ba76149f AA |
2580 | if (!PageLRU(page) || PageLocked(page) || !PageAnon(page)) |
2581 | goto out_unmap; | |
10359213 EA |
2582 | /* |
2583 | * cannot use mapcount: can't collapse if there's a gup pin. | |
2584 | * The page must only be referenced by the scanned process | |
2585 | * and page swap cache. | |
2586 | */ | |
2587 | if (page_count(page) != 1 + !!PageSwapCache(page)) | |
ba76149f | 2588 | goto out_unmap; |
8ee53820 AA |
2589 | if (pte_young(pteval) || PageReferenced(page) || |
2590 | mmu_notifier_test_young(vma->vm_mm, address)) | |
10359213 | 2591 | referenced = true; |
ba76149f | 2592 | } |
10359213 | 2593 | if (referenced && writable) |
ba76149f AA |
2594 | ret = 1; |
2595 | out_unmap: | |
2596 | pte_unmap_unlock(pte, ptl); | |
9f1b868a BL |
2597 | if (ret) { |
2598 | node = khugepaged_find_target_node(); | |
ce83d217 | 2599 | /* collapse_huge_page will return with the mmap_sem released */ |
5c4b4be3 | 2600 | collapse_huge_page(mm, address, hpage, vma, node); |
9f1b868a | 2601 | } |
ba76149f AA |
2602 | out: |
2603 | return ret; | |
2604 | } | |
2605 | ||
2606 | static void collect_mm_slot(struct mm_slot *mm_slot) | |
2607 | { | |
2608 | struct mm_struct *mm = mm_slot->mm; | |
2609 | ||
b9980cdc | 2610 | VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); |
ba76149f AA |
2611 | |
2612 | if (khugepaged_test_exit(mm)) { | |
2613 | /* free mm_slot */ | |
43b5fbbd | 2614 | hash_del(&mm_slot->hash); |
ba76149f AA |
2615 | list_del(&mm_slot->mm_node); |
2616 | ||
2617 | /* | |
2618 | * Not strictly needed because the mm exited already. | |
2619 | * | |
2620 | * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); | |
2621 | */ | |
2622 | ||
2623 | /* khugepaged_mm_lock actually not necessary for the below */ | |
2624 | free_mm_slot(mm_slot); | |
2625 | mmdrop(mm); | |
2626 | } | |
2627 | } | |
2628 | ||
2629 | static unsigned int khugepaged_scan_mm_slot(unsigned int pages, | |
2630 | struct page **hpage) | |
2f1da642 HS |
2631 | __releases(&khugepaged_mm_lock) |
2632 | __acquires(&khugepaged_mm_lock) | |
ba76149f AA |
2633 | { |
2634 | struct mm_slot *mm_slot; | |
2635 | struct mm_struct *mm; | |
2636 | struct vm_area_struct *vma; | |
2637 | int progress = 0; | |
2638 | ||
2639 | VM_BUG_ON(!pages); | |
b9980cdc | 2640 | VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); |
ba76149f AA |
2641 | |
2642 | if (khugepaged_scan.mm_slot) | |
2643 | mm_slot = khugepaged_scan.mm_slot; | |
2644 | else { | |
2645 | mm_slot = list_entry(khugepaged_scan.mm_head.next, | |
2646 | struct mm_slot, mm_node); | |
2647 | khugepaged_scan.address = 0; | |
2648 | khugepaged_scan.mm_slot = mm_slot; | |
2649 | } | |
2650 | spin_unlock(&khugepaged_mm_lock); | |
2651 | ||
2652 | mm = mm_slot->mm; | |
2653 | down_read(&mm->mmap_sem); | |
2654 | if (unlikely(khugepaged_test_exit(mm))) | |
2655 | vma = NULL; | |
2656 | else | |
2657 | vma = find_vma(mm, khugepaged_scan.address); | |
2658 | ||
2659 | progress++; | |
2660 | for (; vma; vma = vma->vm_next) { | |
2661 | unsigned long hstart, hend; | |
2662 | ||
2663 | cond_resched(); | |
2664 | if (unlikely(khugepaged_test_exit(mm))) { | |
2665 | progress++; | |
2666 | break; | |
2667 | } | |
fa475e51 BL |
2668 | if (!hugepage_vma_check(vma)) { |
2669 | skip: | |
ba76149f AA |
2670 | progress++; |
2671 | continue; | |
2672 | } | |
ba76149f AA |
2673 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; |
2674 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
a7d6e4ec AA |
2675 | if (hstart >= hend) |
2676 | goto skip; | |
2677 | if (khugepaged_scan.address > hend) | |
2678 | goto skip; | |
ba76149f AA |
2679 | if (khugepaged_scan.address < hstart) |
2680 | khugepaged_scan.address = hstart; | |
a7d6e4ec | 2681 | VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); |
ba76149f AA |
2682 | |
2683 | while (khugepaged_scan.address < hend) { | |
2684 | int ret; | |
2685 | cond_resched(); | |
2686 | if (unlikely(khugepaged_test_exit(mm))) | |
2687 | goto breakouterloop; | |
2688 | ||
2689 | VM_BUG_ON(khugepaged_scan.address < hstart || | |
2690 | khugepaged_scan.address + HPAGE_PMD_SIZE > | |
2691 | hend); | |
2692 | ret = khugepaged_scan_pmd(mm, vma, | |
2693 | khugepaged_scan.address, | |
2694 | hpage); | |
2695 | /* move to next address */ | |
2696 | khugepaged_scan.address += HPAGE_PMD_SIZE; | |
2697 | progress += HPAGE_PMD_NR; | |
2698 | if (ret) | |
2699 | /* we released mmap_sem so break loop */ | |
2700 | goto breakouterloop_mmap_sem; | |
2701 | if (progress >= pages) | |
2702 | goto breakouterloop; | |
2703 | } | |
2704 | } | |
2705 | breakouterloop: | |
2706 | up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */ | |
2707 | breakouterloop_mmap_sem: | |
2708 | ||
2709 | spin_lock(&khugepaged_mm_lock); | |
a7d6e4ec | 2710 | VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); |
ba76149f AA |
2711 | /* |
2712 | * Release the current mm_slot if this mm is about to die, or | |
2713 | * if we scanned all vmas of this mm. | |
2714 | */ | |
2715 | if (khugepaged_test_exit(mm) || !vma) { | |
2716 | /* | |
2717 | * Make sure that if mm_users is reaching zero while | |
2718 | * khugepaged runs here, khugepaged_exit will find | |
2719 | * mm_slot not pointing to the exiting mm. | |
2720 | */ | |
2721 | if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) { | |
2722 | khugepaged_scan.mm_slot = list_entry( | |
2723 | mm_slot->mm_node.next, | |
2724 | struct mm_slot, mm_node); | |
2725 | khugepaged_scan.address = 0; | |
2726 | } else { | |
2727 | khugepaged_scan.mm_slot = NULL; | |
2728 | khugepaged_full_scans++; | |
2729 | } | |
2730 | ||
2731 | collect_mm_slot(mm_slot); | |
2732 | } | |
2733 | ||
2734 | return progress; | |
2735 | } | |
2736 | ||
2737 | static int khugepaged_has_work(void) | |
2738 | { | |
2739 | return !list_empty(&khugepaged_scan.mm_head) && | |
2740 | khugepaged_enabled(); | |
2741 | } | |
2742 | ||
2743 | static int khugepaged_wait_event(void) | |
2744 | { | |
2745 | return !list_empty(&khugepaged_scan.mm_head) || | |
2017c0bf | 2746 | kthread_should_stop(); |
ba76149f AA |
2747 | } |
2748 | ||
d516904b | 2749 | static void khugepaged_do_scan(void) |
ba76149f | 2750 | { |
d516904b | 2751 | struct page *hpage = NULL; |
ba76149f AA |
2752 | unsigned int progress = 0, pass_through_head = 0; |
2753 | unsigned int pages = khugepaged_pages_to_scan; | |
d516904b | 2754 | bool wait = true; |
ba76149f AA |
2755 | |
2756 | barrier(); /* write khugepaged_pages_to_scan to local stack */ | |
2757 | ||
2758 | while (progress < pages) { | |
26234f36 | 2759 | if (!khugepaged_prealloc_page(&hpage, &wait)) |
d516904b | 2760 | break; |
26234f36 | 2761 | |
420256ef | 2762 | cond_resched(); |
ba76149f | 2763 | |
878aee7d AA |
2764 | if (unlikely(kthread_should_stop() || freezing(current))) |
2765 | break; | |
2766 | ||
ba76149f AA |
2767 | spin_lock(&khugepaged_mm_lock); |
2768 | if (!khugepaged_scan.mm_slot) | |
2769 | pass_through_head++; | |
2770 | if (khugepaged_has_work() && | |
2771 | pass_through_head < 2) | |
2772 | progress += khugepaged_scan_mm_slot(pages - progress, | |
d516904b | 2773 | &hpage); |
ba76149f AA |
2774 | else |
2775 | progress = pages; | |
2776 | spin_unlock(&khugepaged_mm_lock); | |
2777 | } | |
ba76149f | 2778 | |
d516904b XG |
2779 | if (!IS_ERR_OR_NULL(hpage)) |
2780 | put_page(hpage); | |
0bbbc0b3 AA |
2781 | } |
2782 | ||
2017c0bf XG |
2783 | static void khugepaged_wait_work(void) |
2784 | { | |
2785 | try_to_freeze(); | |
2786 | ||
2787 | if (khugepaged_has_work()) { | |
2788 | if (!khugepaged_scan_sleep_millisecs) | |
2789 | return; | |
2790 | ||
2791 | wait_event_freezable_timeout(khugepaged_wait, | |
2792 | kthread_should_stop(), | |
2793 | msecs_to_jiffies(khugepaged_scan_sleep_millisecs)); | |
2794 | return; | |
2795 | } | |
2796 | ||
2797 | if (khugepaged_enabled()) | |
2798 | wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); | |
2799 | } | |
2800 | ||
ba76149f AA |
2801 | static int khugepaged(void *none) |
2802 | { | |
2803 | struct mm_slot *mm_slot; | |
2804 | ||
878aee7d | 2805 | set_freezable(); |
8698a745 | 2806 | set_user_nice(current, MAX_NICE); |
ba76149f | 2807 | |
b7231789 XG |
2808 | while (!kthread_should_stop()) { |
2809 | khugepaged_do_scan(); | |
2810 | khugepaged_wait_work(); | |
2811 | } | |
ba76149f AA |
2812 | |
2813 | spin_lock(&khugepaged_mm_lock); | |
2814 | mm_slot = khugepaged_scan.mm_slot; | |
2815 | khugepaged_scan.mm_slot = NULL; | |
2816 | if (mm_slot) | |
2817 | collect_mm_slot(mm_slot); | |
2818 | spin_unlock(&khugepaged_mm_lock); | |
ba76149f AA |
2819 | return 0; |
2820 | } | |
2821 | ||
c5a647d0 KS |
2822 | static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, |
2823 | unsigned long haddr, pmd_t *pmd) | |
2824 | { | |
2825 | struct mm_struct *mm = vma->vm_mm; | |
2826 | pgtable_t pgtable; | |
2827 | pmd_t _pmd; | |
2828 | int i; | |
2829 | ||
34ee645e | 2830 | pmdp_clear_flush_notify(vma, haddr, pmd); |
c5a647d0 KS |
2831 | /* leave pmd empty until pte is filled */ |
2832 | ||
6b0b50b0 | 2833 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); |
c5a647d0 KS |
2834 | pmd_populate(mm, &_pmd, pgtable); |
2835 | ||
2836 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
2837 | pte_t *pte, entry; | |
2838 | entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); | |
2839 | entry = pte_mkspecial(entry); | |
2840 | pte = pte_offset_map(&_pmd, haddr); | |
2841 | VM_BUG_ON(!pte_none(*pte)); | |
2842 | set_pte_at(mm, haddr, pte, entry); | |
2843 | pte_unmap(pte); | |
2844 | } | |
2845 | smp_wmb(); /* make pte visible before pmd */ | |
2846 | pmd_populate(mm, pmd, pgtable); | |
97ae1749 | 2847 | put_huge_zero_page(); |
c5a647d0 KS |
2848 | } |
2849 | ||
e180377f KS |
2850 | void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address, |
2851 | pmd_t *pmd) | |
71e3aac0 | 2852 | { |
c4088ebd | 2853 | spinlock_t *ptl; |
71e3aac0 | 2854 | struct page *page; |
e180377f | 2855 | struct mm_struct *mm = vma->vm_mm; |
c5a647d0 KS |
2856 | unsigned long haddr = address & HPAGE_PMD_MASK; |
2857 | unsigned long mmun_start; /* For mmu_notifiers */ | |
2858 | unsigned long mmun_end; /* For mmu_notifiers */ | |
e180377f KS |
2859 | |
2860 | BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE); | |
71e3aac0 | 2861 | |
c5a647d0 KS |
2862 | mmun_start = haddr; |
2863 | mmun_end = haddr + HPAGE_PMD_SIZE; | |
750e8165 | 2864 | again: |
c5a647d0 | 2865 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
c4088ebd | 2866 | ptl = pmd_lock(mm, pmd); |
71e3aac0 | 2867 | if (unlikely(!pmd_trans_huge(*pmd))) { |
c4088ebd | 2868 | spin_unlock(ptl); |
c5a647d0 KS |
2869 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
2870 | return; | |
2871 | } | |
2872 | if (is_huge_zero_pmd(*pmd)) { | |
2873 | __split_huge_zero_page_pmd(vma, haddr, pmd); | |
c4088ebd | 2874 | spin_unlock(ptl); |
c5a647d0 | 2875 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
71e3aac0 AA |
2876 | return; |
2877 | } | |
2878 | page = pmd_page(*pmd); | |
309381fe | 2879 | VM_BUG_ON_PAGE(!page_count(page), page); |
71e3aac0 | 2880 | get_page(page); |
c4088ebd | 2881 | spin_unlock(ptl); |
c5a647d0 | 2882 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
71e3aac0 AA |
2883 | |
2884 | split_huge_page(page); | |
2885 | ||
2886 | put_page(page); | |
750e8165 HD |
2887 | |
2888 | /* | |
2889 | * We don't always have down_write of mmap_sem here: a racing | |
2890 | * do_huge_pmd_wp_page() might have copied-on-write to another | |
2891 | * huge page before our split_huge_page() got the anon_vma lock. | |
2892 | */ | |
2893 | if (unlikely(pmd_trans_huge(*pmd))) | |
2894 | goto again; | |
71e3aac0 | 2895 | } |
94fcc585 | 2896 | |
e180377f KS |
2897 | void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address, |
2898 | pmd_t *pmd) | |
2899 | { | |
2900 | struct vm_area_struct *vma; | |
2901 | ||
2902 | vma = find_vma(mm, address); | |
2903 | BUG_ON(vma == NULL); | |
2904 | split_huge_page_pmd(vma, address, pmd); | |
2905 | } | |
2906 | ||
94fcc585 AA |
2907 | static void split_huge_page_address(struct mm_struct *mm, |
2908 | unsigned long address) | |
2909 | { | |
f72e7dcd HD |
2910 | pgd_t *pgd; |
2911 | pud_t *pud; | |
94fcc585 AA |
2912 | pmd_t *pmd; |
2913 | ||
2914 | VM_BUG_ON(!(address & ~HPAGE_PMD_MASK)); | |
2915 | ||
f72e7dcd HD |
2916 | pgd = pgd_offset(mm, address); |
2917 | if (!pgd_present(*pgd)) | |
2918 | return; | |
2919 | ||
2920 | pud = pud_offset(pgd, address); | |
2921 | if (!pud_present(*pud)) | |
2922 | return; | |
2923 | ||
2924 | pmd = pmd_offset(pud, address); | |
2925 | if (!pmd_present(*pmd)) | |
94fcc585 AA |
2926 | return; |
2927 | /* | |
2928 | * Caller holds the mmap_sem write mode, so a huge pmd cannot | |
2929 | * materialize from under us. | |
2930 | */ | |
e180377f | 2931 | split_huge_page_pmd_mm(mm, address, pmd); |
94fcc585 AA |
2932 | } |
2933 | ||
2934 | void __vma_adjust_trans_huge(struct vm_area_struct *vma, | |
2935 | unsigned long start, | |
2936 | unsigned long end, | |
2937 | long adjust_next) | |
2938 | { | |
2939 | /* | |
2940 | * If the new start address isn't hpage aligned and it could | |
2941 | * previously contain an hugepage: check if we need to split | |
2942 | * an huge pmd. | |
2943 | */ | |
2944 | if (start & ~HPAGE_PMD_MASK && | |
2945 | (start & HPAGE_PMD_MASK) >= vma->vm_start && | |
2946 | (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
2947 | split_huge_page_address(vma->vm_mm, start); | |
2948 | ||
2949 | /* | |
2950 | * If the new end address isn't hpage aligned and it could | |
2951 | * previously contain an hugepage: check if we need to split | |
2952 | * an huge pmd. | |
2953 | */ | |
2954 | if (end & ~HPAGE_PMD_MASK && | |
2955 | (end & HPAGE_PMD_MASK) >= vma->vm_start && | |
2956 | (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
2957 | split_huge_page_address(vma->vm_mm, end); | |
2958 | ||
2959 | /* | |
2960 | * If we're also updating the vma->vm_next->vm_start, if the new | |
2961 | * vm_next->vm_start isn't page aligned and it could previously | |
2962 | * contain an hugepage: check if we need to split an huge pmd. | |
2963 | */ | |
2964 | if (adjust_next > 0) { | |
2965 | struct vm_area_struct *next = vma->vm_next; | |
2966 | unsigned long nstart = next->vm_start; | |
2967 | nstart += adjust_next << PAGE_SHIFT; | |
2968 | if (nstart & ~HPAGE_PMD_MASK && | |
2969 | (nstart & HPAGE_PMD_MASK) >= next->vm_start && | |
2970 | (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) | |
2971 | split_huge_page_address(next->vm_mm, nstart); | |
2972 | } | |
2973 | } |