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 | 18 | #include <linux/mm_inline.h> |
4897c765 | 19 | #include <linux/dax.h> |
ba76149f AA |
20 | #include <linux/kthread.h> |
21 | #include <linux/khugepaged.h> | |
878aee7d | 22 | #include <linux/freezer.h> |
a664b2d8 | 23 | #include <linux/mman.h> |
325adeb5 | 24 | #include <linux/pagemap.h> |
4daae3b4 | 25 | #include <linux/migrate.h> |
43b5fbbd | 26 | #include <linux/hashtable.h> |
6b251fc9 | 27 | #include <linux/userfaultfd_k.h> |
33c3fc71 | 28 | #include <linux/page_idle.h> |
97ae1749 | 29 | |
71e3aac0 AA |
30 | #include <asm/tlb.h> |
31 | #include <asm/pgalloc.h> | |
32 | #include "internal.h" | |
33 | ||
7d2eba05 EA |
34 | enum scan_result { |
35 | SCAN_FAIL, | |
36 | SCAN_SUCCEED, | |
37 | SCAN_PMD_NULL, | |
38 | SCAN_EXCEED_NONE_PTE, | |
39 | SCAN_PTE_NON_PRESENT, | |
40 | SCAN_PAGE_RO, | |
41 | SCAN_NO_REFERENCED_PAGE, | |
42 | SCAN_PAGE_NULL, | |
43 | SCAN_SCAN_ABORT, | |
44 | SCAN_PAGE_COUNT, | |
45 | SCAN_PAGE_LRU, | |
46 | SCAN_PAGE_LOCK, | |
47 | SCAN_PAGE_ANON, | |
b1caa957 | 48 | SCAN_PAGE_COMPOUND, |
7d2eba05 EA |
49 | SCAN_ANY_PROCESS, |
50 | SCAN_VMA_NULL, | |
51 | SCAN_VMA_CHECK, | |
52 | SCAN_ADDRESS_RANGE, | |
53 | SCAN_SWAP_CACHE_PAGE, | |
54 | SCAN_DEL_PAGE_LRU, | |
55 | SCAN_ALLOC_HUGE_PAGE_FAIL, | |
56 | SCAN_CGROUP_CHARGE_FAIL | |
57 | }; | |
58 | ||
59 | #define CREATE_TRACE_POINTS | |
60 | #include <trace/events/huge_memory.h> | |
61 | ||
ba76149f | 62 | /* |
8bfa3f9a JW |
63 | * By default transparent hugepage support is disabled in order that avoid |
64 | * to risk increase the memory footprint of applications without a guaranteed | |
65 | * benefit. When transparent hugepage support is enabled, is for all mappings, | |
66 | * and khugepaged scans all mappings. | |
67 | * Defrag is invoked by khugepaged hugepage allocations and by page faults | |
68 | * for all hugepage allocations. | |
ba76149f | 69 | */ |
71e3aac0 | 70 | unsigned long transparent_hugepage_flags __read_mostly = |
13ece886 | 71 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS |
ba76149f | 72 | (1<<TRANSPARENT_HUGEPAGE_FLAG)| |
13ece886 AA |
73 | #endif |
74 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE | |
75 | (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| | |
76 | #endif | |
d39d33c3 | 77 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)| |
79da5407 KS |
78 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| |
79 | (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
ba76149f AA |
80 | |
81 | /* default scan 8*512 pte (or vmas) every 30 second */ | |
82 | static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8; | |
83 | static unsigned int khugepaged_pages_collapsed; | |
84 | static unsigned int khugepaged_full_scans; | |
85 | static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; | |
86 | /* during fragmentation poll the hugepage allocator once every minute */ | |
87 | static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; | |
88 | static struct task_struct *khugepaged_thread __read_mostly; | |
89 | static DEFINE_MUTEX(khugepaged_mutex); | |
90 | static DEFINE_SPINLOCK(khugepaged_mm_lock); | |
91 | static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); | |
92 | /* | |
93 | * default collapse hugepages if there is at least one pte mapped like | |
94 | * it would have happened if the vma was large enough during page | |
95 | * fault. | |
96 | */ | |
97 | static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1; | |
98 | ||
99 | static int khugepaged(void *none); | |
ba76149f | 100 | static int khugepaged_slab_init(void); |
65ebb64f | 101 | static void khugepaged_slab_exit(void); |
ba76149f | 102 | |
43b5fbbd SL |
103 | #define MM_SLOTS_HASH_BITS 10 |
104 | static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); | |
105 | ||
ba76149f AA |
106 | static struct kmem_cache *mm_slot_cache __read_mostly; |
107 | ||
108 | /** | |
109 | * struct mm_slot - hash lookup from mm to mm_slot | |
110 | * @hash: hash collision list | |
111 | * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head | |
112 | * @mm: the mm that this information is valid for | |
113 | */ | |
114 | struct mm_slot { | |
115 | struct hlist_node hash; | |
116 | struct list_head mm_node; | |
117 | struct mm_struct *mm; | |
118 | }; | |
119 | ||
120 | /** | |
121 | * struct khugepaged_scan - cursor for scanning | |
122 | * @mm_head: the head of the mm list to scan | |
123 | * @mm_slot: the current mm_slot we are scanning | |
124 | * @address: the next address inside that to be scanned | |
125 | * | |
126 | * There is only the one khugepaged_scan instance of this cursor structure. | |
127 | */ | |
128 | struct khugepaged_scan { | |
129 | struct list_head mm_head; | |
130 | struct mm_slot *mm_slot; | |
131 | unsigned long address; | |
2f1da642 HS |
132 | }; |
133 | static struct khugepaged_scan khugepaged_scan = { | |
ba76149f AA |
134 | .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), |
135 | }; | |
136 | ||
f000565a | 137 | |
2c0b80d4 | 138 | static void set_recommended_min_free_kbytes(void) |
f000565a AA |
139 | { |
140 | struct zone *zone; | |
141 | int nr_zones = 0; | |
142 | unsigned long recommended_min; | |
f000565a | 143 | |
f000565a AA |
144 | for_each_populated_zone(zone) |
145 | nr_zones++; | |
146 | ||
974a786e | 147 | /* Ensure 2 pageblocks are free to assist fragmentation avoidance */ |
f000565a AA |
148 | recommended_min = pageblock_nr_pages * nr_zones * 2; |
149 | ||
150 | /* | |
151 | * Make sure that on average at least two pageblocks are almost free | |
152 | * of another type, one for a migratetype to fall back to and a | |
153 | * second to avoid subsequent fallbacks of other types There are 3 | |
154 | * MIGRATE_TYPES we care about. | |
155 | */ | |
156 | recommended_min += pageblock_nr_pages * nr_zones * | |
157 | MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; | |
158 | ||
159 | /* don't ever allow to reserve more than 5% of the lowmem */ | |
160 | recommended_min = min(recommended_min, | |
161 | (unsigned long) nr_free_buffer_pages() / 20); | |
162 | recommended_min <<= (PAGE_SHIFT-10); | |
163 | ||
42aa83cb HP |
164 | if (recommended_min > min_free_kbytes) { |
165 | if (user_min_free_kbytes >= 0) | |
166 | pr_info("raising min_free_kbytes from %d to %lu " | |
167 | "to help transparent hugepage allocations\n", | |
168 | min_free_kbytes, recommended_min); | |
169 | ||
f000565a | 170 | min_free_kbytes = recommended_min; |
42aa83cb | 171 | } |
f000565a | 172 | setup_per_zone_wmarks(); |
f000565a | 173 | } |
f000565a | 174 | |
79553da2 | 175 | static int start_stop_khugepaged(void) |
ba76149f AA |
176 | { |
177 | int err = 0; | |
178 | if (khugepaged_enabled()) { | |
ba76149f AA |
179 | if (!khugepaged_thread) |
180 | khugepaged_thread = kthread_run(khugepaged, NULL, | |
181 | "khugepaged"); | |
18e8e5c7 | 182 | if (IS_ERR(khugepaged_thread)) { |
ae3a8c1c | 183 | pr_err("khugepaged: kthread_run(khugepaged) failed\n"); |
ba76149f AA |
184 | err = PTR_ERR(khugepaged_thread); |
185 | khugepaged_thread = NULL; | |
79553da2 | 186 | goto fail; |
ba76149f | 187 | } |
911891af XG |
188 | |
189 | if (!list_empty(&khugepaged_scan.mm_head)) | |
ba76149f | 190 | wake_up_interruptible(&khugepaged_wait); |
f000565a AA |
191 | |
192 | set_recommended_min_free_kbytes(); | |
911891af | 193 | } else if (khugepaged_thread) { |
911891af XG |
194 | kthread_stop(khugepaged_thread); |
195 | khugepaged_thread = NULL; | |
196 | } | |
79553da2 | 197 | fail: |
ba76149f AA |
198 | return err; |
199 | } | |
71e3aac0 | 200 | |
97ae1749 | 201 | static atomic_t huge_zero_refcount; |
56873f43 | 202 | struct page *huge_zero_page __read_mostly; |
4a6c1297 | 203 | |
fc437044 | 204 | struct page *get_huge_zero_page(void) |
97ae1749 KS |
205 | { |
206 | struct page *zero_page; | |
207 | retry: | |
208 | if (likely(atomic_inc_not_zero(&huge_zero_refcount))) | |
4db0c3c2 | 209 | return READ_ONCE(huge_zero_page); |
97ae1749 KS |
210 | |
211 | zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, | |
4a6c1297 | 212 | HPAGE_PMD_ORDER); |
d8a8e1f0 KS |
213 | if (!zero_page) { |
214 | count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); | |
5918d10a | 215 | return NULL; |
d8a8e1f0 KS |
216 | } |
217 | count_vm_event(THP_ZERO_PAGE_ALLOC); | |
97ae1749 | 218 | preempt_disable(); |
5918d10a | 219 | if (cmpxchg(&huge_zero_page, NULL, zero_page)) { |
97ae1749 | 220 | preempt_enable(); |
5ddacbe9 | 221 | __free_pages(zero_page, compound_order(zero_page)); |
97ae1749 KS |
222 | goto retry; |
223 | } | |
224 | ||
225 | /* We take additional reference here. It will be put back by shrinker */ | |
226 | atomic_set(&huge_zero_refcount, 2); | |
227 | preempt_enable(); | |
4db0c3c2 | 228 | return READ_ONCE(huge_zero_page); |
4a6c1297 KS |
229 | } |
230 | ||
97ae1749 | 231 | static void put_huge_zero_page(void) |
4a6c1297 | 232 | { |
97ae1749 KS |
233 | /* |
234 | * Counter should never go to zero here. Only shrinker can put | |
235 | * last reference. | |
236 | */ | |
237 | BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); | |
4a6c1297 KS |
238 | } |
239 | ||
48896466 GC |
240 | static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, |
241 | struct shrink_control *sc) | |
4a6c1297 | 242 | { |
48896466 GC |
243 | /* we can free zero page only if last reference remains */ |
244 | return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; | |
245 | } | |
97ae1749 | 246 | |
48896466 GC |
247 | static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, |
248 | struct shrink_control *sc) | |
249 | { | |
97ae1749 | 250 | if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { |
5918d10a KS |
251 | struct page *zero_page = xchg(&huge_zero_page, NULL); |
252 | BUG_ON(zero_page == NULL); | |
5ddacbe9 | 253 | __free_pages(zero_page, compound_order(zero_page)); |
48896466 | 254 | return HPAGE_PMD_NR; |
97ae1749 KS |
255 | } |
256 | ||
257 | return 0; | |
4a6c1297 KS |
258 | } |
259 | ||
97ae1749 | 260 | static struct shrinker huge_zero_page_shrinker = { |
48896466 GC |
261 | .count_objects = shrink_huge_zero_page_count, |
262 | .scan_objects = shrink_huge_zero_page_scan, | |
97ae1749 KS |
263 | .seeks = DEFAULT_SEEKS, |
264 | }; | |
265 | ||
71e3aac0 | 266 | #ifdef CONFIG_SYSFS |
ba76149f | 267 | |
71e3aac0 AA |
268 | static ssize_t double_flag_show(struct kobject *kobj, |
269 | struct kobj_attribute *attr, char *buf, | |
270 | enum transparent_hugepage_flag enabled, | |
271 | enum transparent_hugepage_flag req_madv) | |
272 | { | |
273 | if (test_bit(enabled, &transparent_hugepage_flags)) { | |
274 | VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags)); | |
275 | return sprintf(buf, "[always] madvise never\n"); | |
276 | } else if (test_bit(req_madv, &transparent_hugepage_flags)) | |
277 | return sprintf(buf, "always [madvise] never\n"); | |
278 | else | |
279 | return sprintf(buf, "always madvise [never]\n"); | |
280 | } | |
281 | static ssize_t double_flag_store(struct kobject *kobj, | |
282 | struct kobj_attribute *attr, | |
283 | const char *buf, size_t count, | |
284 | enum transparent_hugepage_flag enabled, | |
285 | enum transparent_hugepage_flag req_madv) | |
286 | { | |
287 | if (!memcmp("always", buf, | |
288 | min(sizeof("always")-1, count))) { | |
289 | set_bit(enabled, &transparent_hugepage_flags); | |
290 | clear_bit(req_madv, &transparent_hugepage_flags); | |
291 | } else if (!memcmp("madvise", buf, | |
292 | min(sizeof("madvise")-1, count))) { | |
293 | clear_bit(enabled, &transparent_hugepage_flags); | |
294 | set_bit(req_madv, &transparent_hugepage_flags); | |
295 | } else if (!memcmp("never", buf, | |
296 | min(sizeof("never")-1, count))) { | |
297 | clear_bit(enabled, &transparent_hugepage_flags); | |
298 | clear_bit(req_madv, &transparent_hugepage_flags); | |
299 | } else | |
300 | return -EINVAL; | |
301 | ||
302 | return count; | |
303 | } | |
304 | ||
305 | static ssize_t enabled_show(struct kobject *kobj, | |
306 | struct kobj_attribute *attr, char *buf) | |
307 | { | |
308 | return double_flag_show(kobj, attr, buf, | |
309 | TRANSPARENT_HUGEPAGE_FLAG, | |
310 | TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); | |
311 | } | |
312 | static ssize_t enabled_store(struct kobject *kobj, | |
313 | struct kobj_attribute *attr, | |
314 | const char *buf, size_t count) | |
315 | { | |
ba76149f AA |
316 | ssize_t ret; |
317 | ||
318 | ret = double_flag_store(kobj, attr, buf, count, | |
319 | TRANSPARENT_HUGEPAGE_FLAG, | |
320 | TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); | |
321 | ||
322 | if (ret > 0) { | |
911891af XG |
323 | int err; |
324 | ||
325 | mutex_lock(&khugepaged_mutex); | |
79553da2 | 326 | err = start_stop_khugepaged(); |
911891af XG |
327 | mutex_unlock(&khugepaged_mutex); |
328 | ||
ba76149f AA |
329 | if (err) |
330 | ret = err; | |
331 | } | |
332 | ||
333 | return ret; | |
71e3aac0 AA |
334 | } |
335 | static struct kobj_attribute enabled_attr = | |
336 | __ATTR(enabled, 0644, enabled_show, enabled_store); | |
337 | ||
338 | static ssize_t single_flag_show(struct kobject *kobj, | |
339 | struct kobj_attribute *attr, char *buf, | |
340 | enum transparent_hugepage_flag flag) | |
341 | { | |
e27e6151 BH |
342 | return sprintf(buf, "%d\n", |
343 | !!test_bit(flag, &transparent_hugepage_flags)); | |
71e3aac0 | 344 | } |
e27e6151 | 345 | |
71e3aac0 AA |
346 | static ssize_t single_flag_store(struct kobject *kobj, |
347 | struct kobj_attribute *attr, | |
348 | const char *buf, size_t count, | |
349 | enum transparent_hugepage_flag flag) | |
350 | { | |
e27e6151 BH |
351 | unsigned long value; |
352 | int ret; | |
353 | ||
354 | ret = kstrtoul(buf, 10, &value); | |
355 | if (ret < 0) | |
356 | return ret; | |
357 | if (value > 1) | |
358 | return -EINVAL; | |
359 | ||
360 | if (value) | |
71e3aac0 | 361 | set_bit(flag, &transparent_hugepage_flags); |
e27e6151 | 362 | else |
71e3aac0 | 363 | clear_bit(flag, &transparent_hugepage_flags); |
71e3aac0 AA |
364 | |
365 | return count; | |
366 | } | |
367 | ||
368 | /* | |
369 | * Currently defrag only disables __GFP_NOWAIT for allocation. A blind | |
370 | * __GFP_REPEAT is too aggressive, it's never worth swapping tons of | |
371 | * memory just to allocate one more hugepage. | |
372 | */ | |
373 | static ssize_t defrag_show(struct kobject *kobj, | |
374 | struct kobj_attribute *attr, char *buf) | |
375 | { | |
376 | return double_flag_show(kobj, attr, buf, | |
377 | TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, | |
378 | TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); | |
379 | } | |
380 | static ssize_t defrag_store(struct kobject *kobj, | |
381 | struct kobj_attribute *attr, | |
382 | const char *buf, size_t count) | |
383 | { | |
384 | return double_flag_store(kobj, attr, buf, count, | |
385 | TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, | |
386 | TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); | |
387 | } | |
388 | static struct kobj_attribute defrag_attr = | |
389 | __ATTR(defrag, 0644, defrag_show, defrag_store); | |
390 | ||
79da5407 KS |
391 | static ssize_t use_zero_page_show(struct kobject *kobj, |
392 | struct kobj_attribute *attr, char *buf) | |
393 | { | |
394 | return single_flag_show(kobj, attr, buf, | |
395 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
396 | } | |
397 | static ssize_t use_zero_page_store(struct kobject *kobj, | |
398 | struct kobj_attribute *attr, const char *buf, size_t count) | |
399 | { | |
400 | return single_flag_store(kobj, attr, buf, count, | |
401 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
402 | } | |
403 | static struct kobj_attribute use_zero_page_attr = | |
404 | __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); | |
71e3aac0 AA |
405 | #ifdef CONFIG_DEBUG_VM |
406 | static ssize_t debug_cow_show(struct kobject *kobj, | |
407 | struct kobj_attribute *attr, char *buf) | |
408 | { | |
409 | return single_flag_show(kobj, attr, buf, | |
410 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); | |
411 | } | |
412 | static ssize_t debug_cow_store(struct kobject *kobj, | |
413 | struct kobj_attribute *attr, | |
414 | const char *buf, size_t count) | |
415 | { | |
416 | return single_flag_store(kobj, attr, buf, count, | |
417 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); | |
418 | } | |
419 | static struct kobj_attribute debug_cow_attr = | |
420 | __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); | |
421 | #endif /* CONFIG_DEBUG_VM */ | |
422 | ||
423 | static struct attribute *hugepage_attr[] = { | |
424 | &enabled_attr.attr, | |
425 | &defrag_attr.attr, | |
79da5407 | 426 | &use_zero_page_attr.attr, |
71e3aac0 AA |
427 | #ifdef CONFIG_DEBUG_VM |
428 | &debug_cow_attr.attr, | |
429 | #endif | |
430 | NULL, | |
431 | }; | |
432 | ||
433 | static struct attribute_group hugepage_attr_group = { | |
434 | .attrs = hugepage_attr, | |
ba76149f AA |
435 | }; |
436 | ||
437 | static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, | |
438 | struct kobj_attribute *attr, | |
439 | char *buf) | |
440 | { | |
441 | return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs); | |
442 | } | |
443 | ||
444 | static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, | |
445 | struct kobj_attribute *attr, | |
446 | const char *buf, size_t count) | |
447 | { | |
448 | unsigned long msecs; | |
449 | int err; | |
450 | ||
3dbb95f7 | 451 | err = kstrtoul(buf, 10, &msecs); |
ba76149f AA |
452 | if (err || msecs > UINT_MAX) |
453 | return -EINVAL; | |
454 | ||
455 | khugepaged_scan_sleep_millisecs = msecs; | |
456 | wake_up_interruptible(&khugepaged_wait); | |
457 | ||
458 | return count; | |
459 | } | |
460 | static struct kobj_attribute scan_sleep_millisecs_attr = | |
461 | __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show, | |
462 | scan_sleep_millisecs_store); | |
463 | ||
464 | static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, | |
465 | struct kobj_attribute *attr, | |
466 | char *buf) | |
467 | { | |
468 | return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs); | |
469 | } | |
470 | ||
471 | static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, | |
472 | struct kobj_attribute *attr, | |
473 | const char *buf, size_t count) | |
474 | { | |
475 | unsigned long msecs; | |
476 | int err; | |
477 | ||
3dbb95f7 | 478 | err = kstrtoul(buf, 10, &msecs); |
ba76149f AA |
479 | if (err || msecs > UINT_MAX) |
480 | return -EINVAL; | |
481 | ||
482 | khugepaged_alloc_sleep_millisecs = msecs; | |
483 | wake_up_interruptible(&khugepaged_wait); | |
484 | ||
485 | return count; | |
486 | } | |
487 | static struct kobj_attribute alloc_sleep_millisecs_attr = | |
488 | __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show, | |
489 | alloc_sleep_millisecs_store); | |
490 | ||
491 | static ssize_t pages_to_scan_show(struct kobject *kobj, | |
492 | struct kobj_attribute *attr, | |
493 | char *buf) | |
494 | { | |
495 | return sprintf(buf, "%u\n", khugepaged_pages_to_scan); | |
496 | } | |
497 | static ssize_t pages_to_scan_store(struct kobject *kobj, | |
498 | struct kobj_attribute *attr, | |
499 | const char *buf, size_t count) | |
500 | { | |
501 | int err; | |
502 | unsigned long pages; | |
503 | ||
3dbb95f7 | 504 | err = kstrtoul(buf, 10, &pages); |
ba76149f AA |
505 | if (err || !pages || pages > UINT_MAX) |
506 | return -EINVAL; | |
507 | ||
508 | khugepaged_pages_to_scan = pages; | |
509 | ||
510 | return count; | |
511 | } | |
512 | static struct kobj_attribute pages_to_scan_attr = | |
513 | __ATTR(pages_to_scan, 0644, pages_to_scan_show, | |
514 | pages_to_scan_store); | |
515 | ||
516 | static ssize_t pages_collapsed_show(struct kobject *kobj, | |
517 | struct kobj_attribute *attr, | |
518 | char *buf) | |
519 | { | |
520 | return sprintf(buf, "%u\n", khugepaged_pages_collapsed); | |
521 | } | |
522 | static struct kobj_attribute pages_collapsed_attr = | |
523 | __ATTR_RO(pages_collapsed); | |
524 | ||
525 | static ssize_t full_scans_show(struct kobject *kobj, | |
526 | struct kobj_attribute *attr, | |
527 | char *buf) | |
528 | { | |
529 | return sprintf(buf, "%u\n", khugepaged_full_scans); | |
530 | } | |
531 | static struct kobj_attribute full_scans_attr = | |
532 | __ATTR_RO(full_scans); | |
533 | ||
534 | static ssize_t khugepaged_defrag_show(struct kobject *kobj, | |
535 | struct kobj_attribute *attr, char *buf) | |
536 | { | |
537 | return single_flag_show(kobj, attr, buf, | |
538 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); | |
539 | } | |
540 | static ssize_t khugepaged_defrag_store(struct kobject *kobj, | |
541 | struct kobj_attribute *attr, | |
542 | const char *buf, size_t count) | |
543 | { | |
544 | return single_flag_store(kobj, attr, buf, count, | |
545 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); | |
546 | } | |
547 | static struct kobj_attribute khugepaged_defrag_attr = | |
548 | __ATTR(defrag, 0644, khugepaged_defrag_show, | |
549 | khugepaged_defrag_store); | |
550 | ||
551 | /* | |
552 | * max_ptes_none controls if khugepaged should collapse hugepages over | |
553 | * any unmapped ptes in turn potentially increasing the memory | |
554 | * footprint of the vmas. When max_ptes_none is 0 khugepaged will not | |
555 | * reduce the available free memory in the system as it | |
556 | * runs. Increasing max_ptes_none will instead potentially reduce the | |
557 | * free memory in the system during the khugepaged scan. | |
558 | */ | |
559 | static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj, | |
560 | struct kobj_attribute *attr, | |
561 | char *buf) | |
562 | { | |
563 | return sprintf(buf, "%u\n", khugepaged_max_ptes_none); | |
564 | } | |
565 | static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj, | |
566 | struct kobj_attribute *attr, | |
567 | const char *buf, size_t count) | |
568 | { | |
569 | int err; | |
570 | unsigned long max_ptes_none; | |
571 | ||
3dbb95f7 | 572 | err = kstrtoul(buf, 10, &max_ptes_none); |
ba76149f AA |
573 | if (err || max_ptes_none > HPAGE_PMD_NR-1) |
574 | return -EINVAL; | |
575 | ||
576 | khugepaged_max_ptes_none = max_ptes_none; | |
577 | ||
578 | return count; | |
579 | } | |
580 | static struct kobj_attribute khugepaged_max_ptes_none_attr = | |
581 | __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show, | |
582 | khugepaged_max_ptes_none_store); | |
583 | ||
584 | static struct attribute *khugepaged_attr[] = { | |
585 | &khugepaged_defrag_attr.attr, | |
586 | &khugepaged_max_ptes_none_attr.attr, | |
587 | &pages_to_scan_attr.attr, | |
588 | &pages_collapsed_attr.attr, | |
589 | &full_scans_attr.attr, | |
590 | &scan_sleep_millisecs_attr.attr, | |
591 | &alloc_sleep_millisecs_attr.attr, | |
592 | NULL, | |
593 | }; | |
594 | ||
595 | static struct attribute_group khugepaged_attr_group = { | |
596 | .attrs = khugepaged_attr, | |
597 | .name = "khugepaged", | |
71e3aac0 | 598 | }; |
71e3aac0 | 599 | |
569e5590 | 600 | static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) |
71e3aac0 | 601 | { |
71e3aac0 AA |
602 | int err; |
603 | ||
569e5590 SL |
604 | *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); |
605 | if (unlikely(!*hugepage_kobj)) { | |
ae3a8c1c | 606 | pr_err("failed to create transparent hugepage kobject\n"); |
569e5590 | 607 | return -ENOMEM; |
ba76149f AA |
608 | } |
609 | ||
569e5590 | 610 | err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); |
ba76149f | 611 | if (err) { |
ae3a8c1c | 612 | pr_err("failed to register transparent hugepage group\n"); |
569e5590 | 613 | goto delete_obj; |
ba76149f AA |
614 | } |
615 | ||
569e5590 | 616 | err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); |
ba76149f | 617 | if (err) { |
ae3a8c1c | 618 | pr_err("failed to register transparent hugepage group\n"); |
569e5590 | 619 | goto remove_hp_group; |
ba76149f | 620 | } |
569e5590 SL |
621 | |
622 | return 0; | |
623 | ||
624 | remove_hp_group: | |
625 | sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); | |
626 | delete_obj: | |
627 | kobject_put(*hugepage_kobj); | |
628 | return err; | |
629 | } | |
630 | ||
631 | static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
632 | { | |
633 | sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); | |
634 | sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); | |
635 | kobject_put(hugepage_kobj); | |
636 | } | |
637 | #else | |
638 | static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) | |
639 | { | |
640 | return 0; | |
641 | } | |
642 | ||
643 | static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
644 | { | |
645 | } | |
646 | #endif /* CONFIG_SYSFS */ | |
647 | ||
648 | static int __init hugepage_init(void) | |
649 | { | |
650 | int err; | |
651 | struct kobject *hugepage_kobj; | |
652 | ||
653 | if (!has_transparent_hugepage()) { | |
654 | transparent_hugepage_flags = 0; | |
655 | return -EINVAL; | |
656 | } | |
657 | ||
658 | err = hugepage_init_sysfs(&hugepage_kobj); | |
659 | if (err) | |
65ebb64f | 660 | goto err_sysfs; |
ba76149f AA |
661 | |
662 | err = khugepaged_slab_init(); | |
663 | if (err) | |
65ebb64f | 664 | goto err_slab; |
ba76149f | 665 | |
65ebb64f KS |
666 | err = register_shrinker(&huge_zero_page_shrinker); |
667 | if (err) | |
668 | goto err_hzp_shrinker; | |
97ae1749 | 669 | |
97562cd2 RR |
670 | /* |
671 | * By default disable transparent hugepages on smaller systems, | |
672 | * where the extra memory used could hurt more than TLB overhead | |
673 | * is likely to save. The admin can still enable it through /sys. | |
674 | */ | |
79553da2 | 675 | if (totalram_pages < (512 << (20 - PAGE_SHIFT))) { |
97562cd2 | 676 | transparent_hugepage_flags = 0; |
79553da2 KS |
677 | return 0; |
678 | } | |
97562cd2 | 679 | |
79553da2 | 680 | err = start_stop_khugepaged(); |
65ebb64f KS |
681 | if (err) |
682 | goto err_khugepaged; | |
ba76149f | 683 | |
569e5590 | 684 | return 0; |
65ebb64f KS |
685 | err_khugepaged: |
686 | unregister_shrinker(&huge_zero_page_shrinker); | |
687 | err_hzp_shrinker: | |
688 | khugepaged_slab_exit(); | |
689 | err_slab: | |
569e5590 | 690 | hugepage_exit_sysfs(hugepage_kobj); |
65ebb64f | 691 | err_sysfs: |
ba76149f | 692 | return err; |
71e3aac0 | 693 | } |
a64fb3cd | 694 | subsys_initcall(hugepage_init); |
71e3aac0 AA |
695 | |
696 | static int __init setup_transparent_hugepage(char *str) | |
697 | { | |
698 | int ret = 0; | |
699 | if (!str) | |
700 | goto out; | |
701 | if (!strcmp(str, "always")) { | |
702 | set_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
703 | &transparent_hugepage_flags); | |
704 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
705 | &transparent_hugepage_flags); | |
706 | ret = 1; | |
707 | } else if (!strcmp(str, "madvise")) { | |
708 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
709 | &transparent_hugepage_flags); | |
710 | set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
711 | &transparent_hugepage_flags); | |
712 | ret = 1; | |
713 | } else if (!strcmp(str, "never")) { | |
714 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
715 | &transparent_hugepage_flags); | |
716 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
717 | &transparent_hugepage_flags); | |
718 | ret = 1; | |
719 | } | |
720 | out: | |
721 | if (!ret) | |
ae3a8c1c | 722 | pr_warn("transparent_hugepage= cannot parse, ignored\n"); |
71e3aac0 AA |
723 | return ret; |
724 | } | |
725 | __setup("transparent_hugepage=", setup_transparent_hugepage); | |
726 | ||
b32967ff | 727 | pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) |
71e3aac0 AA |
728 | { |
729 | if (likely(vma->vm_flags & VM_WRITE)) | |
730 | pmd = pmd_mkwrite(pmd); | |
731 | return pmd; | |
732 | } | |
733 | ||
3122359a | 734 | static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot) |
b3092b3b BL |
735 | { |
736 | pmd_t entry; | |
3122359a | 737 | entry = mk_pmd(page, prot); |
b3092b3b BL |
738 | entry = pmd_mkhuge(entry); |
739 | return entry; | |
740 | } | |
741 | ||
71e3aac0 AA |
742 | static int __do_huge_pmd_anonymous_page(struct mm_struct *mm, |
743 | struct vm_area_struct *vma, | |
230c92a8 | 744 | unsigned long address, pmd_t *pmd, |
6b251fc9 AA |
745 | struct page *page, gfp_t gfp, |
746 | unsigned int flags) | |
71e3aac0 | 747 | { |
00501b53 | 748 | struct mem_cgroup *memcg; |
71e3aac0 | 749 | pgtable_t pgtable; |
c4088ebd | 750 | spinlock_t *ptl; |
230c92a8 | 751 | unsigned long haddr = address & HPAGE_PMD_MASK; |
71e3aac0 | 752 | |
309381fe | 753 | VM_BUG_ON_PAGE(!PageCompound(page), page); |
00501b53 | 754 | |
f627c2f5 | 755 | if (mem_cgroup_try_charge(page, mm, gfp, &memcg, true)) { |
6b251fc9 AA |
756 | put_page(page); |
757 | count_vm_event(THP_FAULT_FALLBACK); | |
758 | return VM_FAULT_FALLBACK; | |
759 | } | |
00501b53 | 760 | |
71e3aac0 | 761 | pgtable = pte_alloc_one(mm, haddr); |
00501b53 | 762 | if (unlikely(!pgtable)) { |
f627c2f5 | 763 | mem_cgroup_cancel_charge(page, memcg, true); |
6b251fc9 | 764 | put_page(page); |
71e3aac0 | 765 | return VM_FAULT_OOM; |
00501b53 | 766 | } |
71e3aac0 AA |
767 | |
768 | clear_huge_page(page, haddr, HPAGE_PMD_NR); | |
52f37629 MK |
769 | /* |
770 | * The memory barrier inside __SetPageUptodate makes sure that | |
771 | * clear_huge_page writes become visible before the set_pmd_at() | |
772 | * write. | |
773 | */ | |
71e3aac0 AA |
774 | __SetPageUptodate(page); |
775 | ||
c4088ebd | 776 | ptl = pmd_lock(mm, pmd); |
71e3aac0 | 777 | if (unlikely(!pmd_none(*pmd))) { |
c4088ebd | 778 | spin_unlock(ptl); |
f627c2f5 | 779 | mem_cgroup_cancel_charge(page, memcg, true); |
71e3aac0 AA |
780 | put_page(page); |
781 | pte_free(mm, pgtable); | |
782 | } else { | |
783 | pmd_t entry; | |
6b251fc9 AA |
784 | |
785 | /* Deliver the page fault to userland */ | |
786 | if (userfaultfd_missing(vma)) { | |
787 | int ret; | |
788 | ||
789 | spin_unlock(ptl); | |
f627c2f5 | 790 | mem_cgroup_cancel_charge(page, memcg, true); |
6b251fc9 AA |
791 | put_page(page); |
792 | pte_free(mm, pgtable); | |
230c92a8 | 793 | ret = handle_userfault(vma, address, flags, |
6b251fc9 AA |
794 | VM_UFFD_MISSING); |
795 | VM_BUG_ON(ret & VM_FAULT_FALLBACK); | |
796 | return ret; | |
797 | } | |
798 | ||
3122359a KS |
799 | entry = mk_huge_pmd(page, vma->vm_page_prot); |
800 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
d281ee61 | 801 | page_add_new_anon_rmap(page, vma, haddr, true); |
f627c2f5 | 802 | mem_cgroup_commit_charge(page, memcg, false, true); |
00501b53 | 803 | lru_cache_add_active_or_unevictable(page, vma); |
6b0b50b0 | 804 | pgtable_trans_huge_deposit(mm, pmd, pgtable); |
71e3aac0 | 805 | set_pmd_at(mm, haddr, pmd, entry); |
71e3aac0 | 806 | add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); |
e1f56c89 | 807 | atomic_long_inc(&mm->nr_ptes); |
c4088ebd | 808 | spin_unlock(ptl); |
6b251fc9 | 809 | count_vm_event(THP_FAULT_ALLOC); |
71e3aac0 AA |
810 | } |
811 | ||
aa2e878e | 812 | return 0; |
71e3aac0 AA |
813 | } |
814 | ||
cc5d462f | 815 | static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp) |
0bbbc0b3 | 816 | { |
71baba4b | 817 | return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_RECLAIM)) | extra_gfp; |
0bbbc0b3 AA |
818 | } |
819 | ||
c4088ebd | 820 | /* Caller must hold page table lock. */ |
d295e341 | 821 | static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, |
97ae1749 | 822 | struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, |
5918d10a | 823 | struct page *zero_page) |
fc9fe822 KS |
824 | { |
825 | pmd_t entry; | |
7c414164 AM |
826 | if (!pmd_none(*pmd)) |
827 | return false; | |
5918d10a | 828 | entry = mk_pmd(zero_page, vma->vm_page_prot); |
fc9fe822 | 829 | entry = pmd_mkhuge(entry); |
6b0b50b0 | 830 | pgtable_trans_huge_deposit(mm, pmd, pgtable); |
fc9fe822 | 831 | set_pmd_at(mm, haddr, pmd, entry); |
e1f56c89 | 832 | atomic_long_inc(&mm->nr_ptes); |
7c414164 | 833 | return true; |
fc9fe822 KS |
834 | } |
835 | ||
71e3aac0 AA |
836 | int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma, |
837 | unsigned long address, pmd_t *pmd, | |
838 | unsigned int flags) | |
839 | { | |
077fcf11 | 840 | gfp_t gfp; |
71e3aac0 AA |
841 | struct page *page; |
842 | unsigned long haddr = address & HPAGE_PMD_MASK; | |
71e3aac0 | 843 | |
128ec037 | 844 | if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end) |
c0292554 | 845 | return VM_FAULT_FALLBACK; |
7479df6d KS |
846 | if (vma->vm_flags & VM_LOCKED) |
847 | return VM_FAULT_FALLBACK; | |
128ec037 KS |
848 | if (unlikely(anon_vma_prepare(vma))) |
849 | return VM_FAULT_OOM; | |
6d50e60c | 850 | if (unlikely(khugepaged_enter(vma, vma->vm_flags))) |
128ec037 | 851 | return VM_FAULT_OOM; |
593befa6 | 852 | if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) && |
128ec037 | 853 | transparent_hugepage_use_zero_page()) { |
c4088ebd | 854 | spinlock_t *ptl; |
128ec037 KS |
855 | pgtable_t pgtable; |
856 | struct page *zero_page; | |
857 | bool set; | |
6b251fc9 | 858 | int ret; |
128ec037 KS |
859 | pgtable = pte_alloc_one(mm, haddr); |
860 | if (unlikely(!pgtable)) | |
ba76149f | 861 | return VM_FAULT_OOM; |
128ec037 KS |
862 | zero_page = get_huge_zero_page(); |
863 | if (unlikely(!zero_page)) { | |
864 | pte_free(mm, pgtable); | |
81ab4201 | 865 | count_vm_event(THP_FAULT_FALLBACK); |
c0292554 | 866 | return VM_FAULT_FALLBACK; |
b9bbfbe3 | 867 | } |
c4088ebd | 868 | ptl = pmd_lock(mm, pmd); |
6b251fc9 AA |
869 | ret = 0; |
870 | set = false; | |
871 | if (pmd_none(*pmd)) { | |
872 | if (userfaultfd_missing(vma)) { | |
873 | spin_unlock(ptl); | |
230c92a8 | 874 | ret = handle_userfault(vma, address, flags, |
6b251fc9 AA |
875 | VM_UFFD_MISSING); |
876 | VM_BUG_ON(ret & VM_FAULT_FALLBACK); | |
877 | } else { | |
878 | set_huge_zero_page(pgtable, mm, vma, | |
879 | haddr, pmd, | |
880 | zero_page); | |
881 | spin_unlock(ptl); | |
882 | set = true; | |
883 | } | |
884 | } else | |
885 | spin_unlock(ptl); | |
128ec037 KS |
886 | if (!set) { |
887 | pte_free(mm, pgtable); | |
888 | put_huge_zero_page(); | |
edad9d2c | 889 | } |
6b251fc9 | 890 | return ret; |
71e3aac0 | 891 | } |
077fcf11 AK |
892 | gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0); |
893 | page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); | |
128ec037 KS |
894 | if (unlikely(!page)) { |
895 | count_vm_event(THP_FAULT_FALLBACK); | |
c0292554 | 896 | return VM_FAULT_FALLBACK; |
128ec037 | 897 | } |
230c92a8 AA |
898 | return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp, |
899 | flags); | |
71e3aac0 AA |
900 | } |
901 | ||
ae18d6dc | 902 | static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, |
5cad465d MW |
903 | pmd_t *pmd, unsigned long pfn, pgprot_t prot, bool write) |
904 | { | |
905 | struct mm_struct *mm = vma->vm_mm; | |
906 | pmd_t entry; | |
907 | spinlock_t *ptl; | |
908 | ||
909 | ptl = pmd_lock(mm, pmd); | |
910 | if (pmd_none(*pmd)) { | |
911 | entry = pmd_mkhuge(pfn_pmd(pfn, prot)); | |
912 | if (write) { | |
913 | entry = pmd_mkyoung(pmd_mkdirty(entry)); | |
914 | entry = maybe_pmd_mkwrite(entry, vma); | |
915 | } | |
916 | set_pmd_at(mm, addr, pmd, entry); | |
917 | update_mmu_cache_pmd(vma, addr, pmd); | |
918 | } | |
919 | spin_unlock(ptl); | |
5cad465d MW |
920 | } |
921 | ||
922 | int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, | |
923 | pmd_t *pmd, unsigned long pfn, bool write) | |
924 | { | |
925 | pgprot_t pgprot = vma->vm_page_prot; | |
926 | /* | |
927 | * If we had pmd_special, we could avoid all these restrictions, | |
928 | * but we need to be consistent with PTEs and architectures that | |
929 | * can't support a 'special' bit. | |
930 | */ | |
931 | BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); | |
932 | BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == | |
933 | (VM_PFNMAP|VM_MIXEDMAP)); | |
934 | BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); | |
935 | BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn)); | |
936 | ||
937 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
938 | return VM_FAULT_SIGBUS; | |
939 | if (track_pfn_insert(vma, &pgprot, pfn)) | |
940 | return VM_FAULT_SIGBUS; | |
ae18d6dc MW |
941 | insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write); |
942 | return VM_FAULT_NOPAGE; | |
5cad465d MW |
943 | } |
944 | ||
71e3aac0 AA |
945 | int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, |
946 | pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, | |
947 | struct vm_area_struct *vma) | |
948 | { | |
c4088ebd | 949 | spinlock_t *dst_ptl, *src_ptl; |
71e3aac0 AA |
950 | struct page *src_page; |
951 | pmd_t pmd; | |
952 | pgtable_t pgtable; | |
953 | int ret; | |
954 | ||
955 | ret = -ENOMEM; | |
956 | pgtable = pte_alloc_one(dst_mm, addr); | |
957 | if (unlikely(!pgtable)) | |
958 | goto out; | |
959 | ||
c4088ebd KS |
960 | dst_ptl = pmd_lock(dst_mm, dst_pmd); |
961 | src_ptl = pmd_lockptr(src_mm, src_pmd); | |
962 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
71e3aac0 AA |
963 | |
964 | ret = -EAGAIN; | |
965 | pmd = *src_pmd; | |
966 | if (unlikely(!pmd_trans_huge(pmd))) { | |
967 | pte_free(dst_mm, pgtable); | |
968 | goto out_unlock; | |
969 | } | |
fc9fe822 | 970 | /* |
c4088ebd | 971 | * When page table lock is held, the huge zero pmd should not be |
fc9fe822 KS |
972 | * under splitting since we don't split the page itself, only pmd to |
973 | * a page table. | |
974 | */ | |
975 | if (is_huge_zero_pmd(pmd)) { | |
5918d10a | 976 | struct page *zero_page; |
97ae1749 KS |
977 | /* |
978 | * get_huge_zero_page() will never allocate a new page here, | |
979 | * since we already have a zero page to copy. It just takes a | |
980 | * reference. | |
981 | */ | |
5918d10a | 982 | zero_page = get_huge_zero_page(); |
6b251fc9 | 983 | set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, |
5918d10a | 984 | zero_page); |
fc9fe822 KS |
985 | ret = 0; |
986 | goto out_unlock; | |
987 | } | |
de466bd6 | 988 | |
71e3aac0 AA |
989 | if (unlikely(pmd_trans_splitting(pmd))) { |
990 | /* split huge page running from under us */ | |
c4088ebd KS |
991 | spin_unlock(src_ptl); |
992 | spin_unlock(dst_ptl); | |
71e3aac0 AA |
993 | pte_free(dst_mm, pgtable); |
994 | ||
995 | wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */ | |
996 | goto out; | |
997 | } | |
998 | src_page = pmd_page(pmd); | |
309381fe | 999 | VM_BUG_ON_PAGE(!PageHead(src_page), src_page); |
71e3aac0 AA |
1000 | get_page(src_page); |
1001 | page_dup_rmap(src_page); | |
1002 | add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); | |
1003 | ||
1004 | pmdp_set_wrprotect(src_mm, addr, src_pmd); | |
1005 | pmd = pmd_mkold(pmd_wrprotect(pmd)); | |
6b0b50b0 | 1006 | pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); |
71e3aac0 | 1007 | set_pmd_at(dst_mm, addr, dst_pmd, pmd); |
e1f56c89 | 1008 | atomic_long_inc(&dst_mm->nr_ptes); |
71e3aac0 AA |
1009 | |
1010 | ret = 0; | |
1011 | out_unlock: | |
c4088ebd KS |
1012 | spin_unlock(src_ptl); |
1013 | spin_unlock(dst_ptl); | |
71e3aac0 AA |
1014 | out: |
1015 | return ret; | |
1016 | } | |
1017 | ||
a1dd450b WD |
1018 | void huge_pmd_set_accessed(struct mm_struct *mm, |
1019 | struct vm_area_struct *vma, | |
1020 | unsigned long address, | |
1021 | pmd_t *pmd, pmd_t orig_pmd, | |
1022 | int dirty) | |
1023 | { | |
c4088ebd | 1024 | spinlock_t *ptl; |
a1dd450b WD |
1025 | pmd_t entry; |
1026 | unsigned long haddr; | |
1027 | ||
c4088ebd | 1028 | ptl = pmd_lock(mm, pmd); |
a1dd450b WD |
1029 | if (unlikely(!pmd_same(*pmd, orig_pmd))) |
1030 | goto unlock; | |
1031 | ||
1032 | entry = pmd_mkyoung(orig_pmd); | |
1033 | haddr = address & HPAGE_PMD_MASK; | |
1034 | if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty)) | |
1035 | update_mmu_cache_pmd(vma, address, pmd); | |
1036 | ||
1037 | unlock: | |
c4088ebd | 1038 | spin_unlock(ptl); |
a1dd450b WD |
1039 | } |
1040 | ||
5338a937 HD |
1041 | /* |
1042 | * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages | |
1043 | * during copy_user_huge_page()'s copy_page_rep(): in the case when | |
1044 | * the source page gets split and a tail freed before copy completes. | |
1045 | * Called under pmd_lock of checked pmd, so safe from splitting itself. | |
1046 | */ | |
1047 | static void get_user_huge_page(struct page *page) | |
1048 | { | |
1049 | if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) { | |
1050 | struct page *endpage = page + HPAGE_PMD_NR; | |
1051 | ||
1052 | atomic_add(HPAGE_PMD_NR, &page->_count); | |
1053 | while (++page < endpage) | |
1054 | get_huge_page_tail(page); | |
1055 | } else { | |
1056 | get_page(page); | |
1057 | } | |
1058 | } | |
1059 | ||
1060 | static void put_user_huge_page(struct page *page) | |
1061 | { | |
1062 | if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) { | |
1063 | struct page *endpage = page + HPAGE_PMD_NR; | |
1064 | ||
1065 | while (page < endpage) | |
1066 | put_page(page++); | |
1067 | } else { | |
1068 | put_page(page); | |
1069 | } | |
1070 | } | |
1071 | ||
71e3aac0 AA |
1072 | static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm, |
1073 | struct vm_area_struct *vma, | |
1074 | unsigned long address, | |
1075 | pmd_t *pmd, pmd_t orig_pmd, | |
1076 | struct page *page, | |
1077 | unsigned long haddr) | |
1078 | { | |
00501b53 | 1079 | struct mem_cgroup *memcg; |
c4088ebd | 1080 | spinlock_t *ptl; |
71e3aac0 AA |
1081 | pgtable_t pgtable; |
1082 | pmd_t _pmd; | |
1083 | int ret = 0, i; | |
1084 | struct page **pages; | |
2ec74c3e SG |
1085 | unsigned long mmun_start; /* For mmu_notifiers */ |
1086 | unsigned long mmun_end; /* For mmu_notifiers */ | |
71e3aac0 AA |
1087 | |
1088 | pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR, | |
1089 | GFP_KERNEL); | |
1090 | if (unlikely(!pages)) { | |
1091 | ret |= VM_FAULT_OOM; | |
1092 | goto out; | |
1093 | } | |
1094 | ||
1095 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
cc5d462f AK |
1096 | pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE | |
1097 | __GFP_OTHER_NODE, | |
19ee151e | 1098 | vma, address, page_to_nid(page)); |
b9bbfbe3 | 1099 | if (unlikely(!pages[i] || |
00501b53 | 1100 | mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL, |
f627c2f5 | 1101 | &memcg, false))) { |
b9bbfbe3 | 1102 | if (pages[i]) |
71e3aac0 | 1103 | put_page(pages[i]); |
b9bbfbe3 | 1104 | while (--i >= 0) { |
00501b53 JW |
1105 | memcg = (void *)page_private(pages[i]); |
1106 | set_page_private(pages[i], 0); | |
f627c2f5 KS |
1107 | mem_cgroup_cancel_charge(pages[i], memcg, |
1108 | false); | |
b9bbfbe3 AA |
1109 | put_page(pages[i]); |
1110 | } | |
71e3aac0 AA |
1111 | kfree(pages); |
1112 | ret |= VM_FAULT_OOM; | |
1113 | goto out; | |
1114 | } | |
00501b53 | 1115 | set_page_private(pages[i], (unsigned long)memcg); |
71e3aac0 AA |
1116 | } |
1117 | ||
1118 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
1119 | copy_user_highpage(pages[i], page + i, | |
0089e485 | 1120 | haddr + PAGE_SIZE * i, vma); |
71e3aac0 AA |
1121 | __SetPageUptodate(pages[i]); |
1122 | cond_resched(); | |
1123 | } | |
1124 | ||
2ec74c3e SG |
1125 | mmun_start = haddr; |
1126 | mmun_end = haddr + HPAGE_PMD_SIZE; | |
1127 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
1128 | ||
c4088ebd | 1129 | ptl = pmd_lock(mm, pmd); |
71e3aac0 AA |
1130 | if (unlikely(!pmd_same(*pmd, orig_pmd))) |
1131 | goto out_free_pages; | |
309381fe | 1132 | VM_BUG_ON_PAGE(!PageHead(page), page); |
71e3aac0 | 1133 | |
8809aa2d | 1134 | pmdp_huge_clear_flush_notify(vma, haddr, pmd); |
71e3aac0 AA |
1135 | /* leave pmd empty until pte is filled */ |
1136 | ||
6b0b50b0 | 1137 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); |
71e3aac0 AA |
1138 | pmd_populate(mm, &_pmd, pgtable); |
1139 | ||
1140 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
1141 | pte_t *pte, entry; | |
1142 | entry = mk_pte(pages[i], vma->vm_page_prot); | |
1143 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); | |
00501b53 JW |
1144 | memcg = (void *)page_private(pages[i]); |
1145 | set_page_private(pages[i], 0); | |
d281ee61 | 1146 | page_add_new_anon_rmap(pages[i], vma, haddr, false); |
f627c2f5 | 1147 | mem_cgroup_commit_charge(pages[i], memcg, false, false); |
00501b53 | 1148 | lru_cache_add_active_or_unevictable(pages[i], vma); |
71e3aac0 AA |
1149 | pte = pte_offset_map(&_pmd, haddr); |
1150 | VM_BUG_ON(!pte_none(*pte)); | |
1151 | set_pte_at(mm, haddr, pte, entry); | |
1152 | pte_unmap(pte); | |
1153 | } | |
1154 | kfree(pages); | |
1155 | ||
71e3aac0 AA |
1156 | smp_wmb(); /* make pte visible before pmd */ |
1157 | pmd_populate(mm, pmd, pgtable); | |
d281ee61 | 1158 | page_remove_rmap(page, true); |
c4088ebd | 1159 | spin_unlock(ptl); |
71e3aac0 | 1160 | |
2ec74c3e SG |
1161 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
1162 | ||
71e3aac0 AA |
1163 | ret |= VM_FAULT_WRITE; |
1164 | put_page(page); | |
1165 | ||
1166 | out: | |
1167 | return ret; | |
1168 | ||
1169 | out_free_pages: | |
c4088ebd | 1170 | spin_unlock(ptl); |
2ec74c3e | 1171 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
b9bbfbe3 | 1172 | for (i = 0; i < HPAGE_PMD_NR; i++) { |
00501b53 JW |
1173 | memcg = (void *)page_private(pages[i]); |
1174 | set_page_private(pages[i], 0); | |
f627c2f5 | 1175 | mem_cgroup_cancel_charge(pages[i], memcg, false); |
71e3aac0 | 1176 | put_page(pages[i]); |
b9bbfbe3 | 1177 | } |
71e3aac0 AA |
1178 | kfree(pages); |
1179 | goto out; | |
1180 | } | |
1181 | ||
1182 | int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma, | |
1183 | unsigned long address, pmd_t *pmd, pmd_t orig_pmd) | |
1184 | { | |
c4088ebd | 1185 | spinlock_t *ptl; |
71e3aac0 | 1186 | int ret = 0; |
93b4796d | 1187 | struct page *page = NULL, *new_page; |
00501b53 | 1188 | struct mem_cgroup *memcg; |
71e3aac0 | 1189 | unsigned long haddr; |
2ec74c3e SG |
1190 | unsigned long mmun_start; /* For mmu_notifiers */ |
1191 | unsigned long mmun_end; /* For mmu_notifiers */ | |
3b363692 | 1192 | gfp_t huge_gfp; /* for allocation and charge */ |
71e3aac0 | 1193 | |
c4088ebd | 1194 | ptl = pmd_lockptr(mm, pmd); |
81d1b09c | 1195 | VM_BUG_ON_VMA(!vma->anon_vma, vma); |
93b4796d KS |
1196 | haddr = address & HPAGE_PMD_MASK; |
1197 | if (is_huge_zero_pmd(orig_pmd)) | |
1198 | goto alloc; | |
c4088ebd | 1199 | spin_lock(ptl); |
71e3aac0 AA |
1200 | if (unlikely(!pmd_same(*pmd, orig_pmd))) |
1201 | goto out_unlock; | |
1202 | ||
1203 | page = pmd_page(orig_pmd); | |
309381fe | 1204 | VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); |
1f25fe20 KS |
1205 | /* |
1206 | * We can only reuse the page if nobody else maps the huge page or it's | |
1207 | * part. We can do it by checking page_mapcount() on each sub-page, but | |
1208 | * it's expensive. | |
1209 | * The cheaper way is to check page_count() to be equal 1: every | |
1210 | * mapcount takes page reference reference, so this way we can | |
1211 | * guarantee, that the PMD is the only mapping. | |
1212 | * This can give false negative if somebody pinned the page, but that's | |
1213 | * fine. | |
1214 | */ | |
1215 | if (page_mapcount(page) == 1 && page_count(page) == 1) { | |
71e3aac0 AA |
1216 | pmd_t entry; |
1217 | entry = pmd_mkyoung(orig_pmd); | |
1218 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
1219 | if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1)) | |
b113da65 | 1220 | update_mmu_cache_pmd(vma, address, pmd); |
71e3aac0 AA |
1221 | ret |= VM_FAULT_WRITE; |
1222 | goto out_unlock; | |
1223 | } | |
5338a937 | 1224 | get_user_huge_page(page); |
c4088ebd | 1225 | spin_unlock(ptl); |
93b4796d | 1226 | alloc: |
71e3aac0 | 1227 | if (transparent_hugepage_enabled(vma) && |
077fcf11 | 1228 | !transparent_hugepage_debug_cow()) { |
3b363692 MH |
1229 | huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0); |
1230 | new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER); | |
077fcf11 | 1231 | } else |
71e3aac0 AA |
1232 | new_page = NULL; |
1233 | ||
1234 | if (unlikely(!new_page)) { | |
eecc1e42 | 1235 | if (!page) { |
78ddc534 | 1236 | split_huge_pmd(vma, pmd, address); |
e9b71ca9 | 1237 | ret |= VM_FAULT_FALLBACK; |
93b4796d KS |
1238 | } else { |
1239 | ret = do_huge_pmd_wp_page_fallback(mm, vma, address, | |
1240 | pmd, orig_pmd, page, haddr); | |
9845cbbd | 1241 | if (ret & VM_FAULT_OOM) { |
78ddc534 | 1242 | split_huge_pmd(vma, pmd, address); |
9845cbbd KS |
1243 | ret |= VM_FAULT_FALLBACK; |
1244 | } | |
5338a937 | 1245 | put_user_huge_page(page); |
93b4796d | 1246 | } |
17766dde | 1247 | count_vm_event(THP_FAULT_FALLBACK); |
71e3aac0 AA |
1248 | goto out; |
1249 | } | |
1250 | ||
f627c2f5 KS |
1251 | if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg, |
1252 | true))) { | |
b9bbfbe3 | 1253 | put_page(new_page); |
93b4796d | 1254 | if (page) { |
78ddc534 | 1255 | split_huge_pmd(vma, pmd, address); |
5338a937 | 1256 | put_user_huge_page(page); |
9845cbbd | 1257 | } else |
78ddc534 | 1258 | split_huge_pmd(vma, pmd, address); |
9845cbbd | 1259 | ret |= VM_FAULT_FALLBACK; |
17766dde | 1260 | count_vm_event(THP_FAULT_FALLBACK); |
b9bbfbe3 AA |
1261 | goto out; |
1262 | } | |
1263 | ||
17766dde DR |
1264 | count_vm_event(THP_FAULT_ALLOC); |
1265 | ||
eecc1e42 | 1266 | if (!page) |
93b4796d KS |
1267 | clear_huge_page(new_page, haddr, HPAGE_PMD_NR); |
1268 | else | |
1269 | copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR); | |
71e3aac0 AA |
1270 | __SetPageUptodate(new_page); |
1271 | ||
2ec74c3e SG |
1272 | mmun_start = haddr; |
1273 | mmun_end = haddr + HPAGE_PMD_SIZE; | |
1274 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
1275 | ||
c4088ebd | 1276 | spin_lock(ptl); |
93b4796d | 1277 | if (page) |
5338a937 | 1278 | put_user_huge_page(page); |
b9bbfbe3 | 1279 | if (unlikely(!pmd_same(*pmd, orig_pmd))) { |
c4088ebd | 1280 | spin_unlock(ptl); |
f627c2f5 | 1281 | mem_cgroup_cancel_charge(new_page, memcg, true); |
71e3aac0 | 1282 | put_page(new_page); |
2ec74c3e | 1283 | goto out_mn; |
b9bbfbe3 | 1284 | } else { |
71e3aac0 | 1285 | pmd_t entry; |
3122359a KS |
1286 | entry = mk_huge_pmd(new_page, vma->vm_page_prot); |
1287 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
8809aa2d | 1288 | pmdp_huge_clear_flush_notify(vma, haddr, pmd); |
d281ee61 | 1289 | page_add_new_anon_rmap(new_page, vma, haddr, true); |
f627c2f5 | 1290 | mem_cgroup_commit_charge(new_page, memcg, false, true); |
00501b53 | 1291 | lru_cache_add_active_or_unevictable(new_page, vma); |
71e3aac0 | 1292 | set_pmd_at(mm, haddr, pmd, entry); |
b113da65 | 1293 | update_mmu_cache_pmd(vma, address, pmd); |
eecc1e42 | 1294 | if (!page) { |
93b4796d | 1295 | add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); |
97ae1749 KS |
1296 | put_huge_zero_page(); |
1297 | } else { | |
309381fe | 1298 | VM_BUG_ON_PAGE(!PageHead(page), page); |
d281ee61 | 1299 | page_remove_rmap(page, true); |
93b4796d KS |
1300 | put_page(page); |
1301 | } | |
71e3aac0 AA |
1302 | ret |= VM_FAULT_WRITE; |
1303 | } | |
c4088ebd | 1304 | spin_unlock(ptl); |
2ec74c3e SG |
1305 | out_mn: |
1306 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); | |
71e3aac0 AA |
1307 | out: |
1308 | return ret; | |
2ec74c3e | 1309 | out_unlock: |
c4088ebd | 1310 | spin_unlock(ptl); |
2ec74c3e | 1311 | return ret; |
71e3aac0 AA |
1312 | } |
1313 | ||
b676b293 | 1314 | struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, |
71e3aac0 AA |
1315 | unsigned long addr, |
1316 | pmd_t *pmd, | |
1317 | unsigned int flags) | |
1318 | { | |
b676b293 | 1319 | struct mm_struct *mm = vma->vm_mm; |
71e3aac0 AA |
1320 | struct page *page = NULL; |
1321 | ||
c4088ebd | 1322 | assert_spin_locked(pmd_lockptr(mm, pmd)); |
71e3aac0 AA |
1323 | |
1324 | if (flags & FOLL_WRITE && !pmd_write(*pmd)) | |
1325 | goto out; | |
1326 | ||
85facf25 KS |
1327 | /* Avoid dumping huge zero page */ |
1328 | if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) | |
1329 | return ERR_PTR(-EFAULT); | |
1330 | ||
2b4847e7 | 1331 | /* Full NUMA hinting faults to serialise migration in fault paths */ |
8a0516ed | 1332 | if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) |
2b4847e7 MG |
1333 | goto out; |
1334 | ||
71e3aac0 | 1335 | page = pmd_page(*pmd); |
309381fe | 1336 | VM_BUG_ON_PAGE(!PageHead(page), page); |
71e3aac0 AA |
1337 | if (flags & FOLL_TOUCH) { |
1338 | pmd_t _pmd; | |
1339 | /* | |
1340 | * We should set the dirty bit only for FOLL_WRITE but | |
1341 | * for now the dirty bit in the pmd is meaningless. | |
1342 | * And if the dirty bit will become meaningful and | |
1343 | * we'll only set it with FOLL_WRITE, an atomic | |
1344 | * set_bit will be required on the pmd to set the | |
1345 | * young bit, instead of the current set_pmd_at. | |
1346 | */ | |
1347 | _pmd = pmd_mkyoung(pmd_mkdirty(*pmd)); | |
8663890a AK |
1348 | if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, |
1349 | pmd, _pmd, 1)) | |
1350 | update_mmu_cache_pmd(vma, addr, pmd); | |
71e3aac0 | 1351 | } |
de60f5f1 | 1352 | if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { |
b676b293 DR |
1353 | if (page->mapping && trylock_page(page)) { |
1354 | lru_add_drain(); | |
1355 | if (page->mapping) | |
1356 | mlock_vma_page(page); | |
1357 | unlock_page(page); | |
1358 | } | |
1359 | } | |
71e3aac0 | 1360 | page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; |
309381fe | 1361 | VM_BUG_ON_PAGE(!PageCompound(page), page); |
71e3aac0 | 1362 | if (flags & FOLL_GET) |
70b50f94 | 1363 | get_page_foll(page); |
71e3aac0 AA |
1364 | |
1365 | out: | |
1366 | return page; | |
1367 | } | |
1368 | ||
d10e63f2 | 1369 | /* NUMA hinting page fault entry point for trans huge pmds */ |
4daae3b4 MG |
1370 | int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma, |
1371 | unsigned long addr, pmd_t pmd, pmd_t *pmdp) | |
d10e63f2 | 1372 | { |
c4088ebd | 1373 | spinlock_t *ptl; |
b8916634 | 1374 | struct anon_vma *anon_vma = NULL; |
b32967ff | 1375 | struct page *page; |
d10e63f2 | 1376 | unsigned long haddr = addr & HPAGE_PMD_MASK; |
8191acbd | 1377 | int page_nid = -1, this_nid = numa_node_id(); |
90572890 | 1378 | int target_nid, last_cpupid = -1; |
8191acbd MG |
1379 | bool page_locked; |
1380 | bool migrated = false; | |
b191f9b1 | 1381 | bool was_writable; |
6688cc05 | 1382 | int flags = 0; |
d10e63f2 | 1383 | |
c0e7cad9 MG |
1384 | /* A PROT_NONE fault should not end up here */ |
1385 | BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))); | |
1386 | ||
c4088ebd | 1387 | ptl = pmd_lock(mm, pmdp); |
d10e63f2 MG |
1388 | if (unlikely(!pmd_same(pmd, *pmdp))) |
1389 | goto out_unlock; | |
1390 | ||
de466bd6 MG |
1391 | /* |
1392 | * If there are potential migrations, wait for completion and retry | |
1393 | * without disrupting NUMA hinting information. Do not relock and | |
1394 | * check_same as the page may no longer be mapped. | |
1395 | */ | |
1396 | if (unlikely(pmd_trans_migrating(*pmdp))) { | |
5d833062 | 1397 | page = pmd_page(*pmdp); |
de466bd6 | 1398 | spin_unlock(ptl); |
5d833062 | 1399 | wait_on_page_locked(page); |
de466bd6 MG |
1400 | goto out; |
1401 | } | |
1402 | ||
d10e63f2 | 1403 | page = pmd_page(pmd); |
a1a46184 | 1404 | BUG_ON(is_huge_zero_page(page)); |
8191acbd | 1405 | page_nid = page_to_nid(page); |
90572890 | 1406 | last_cpupid = page_cpupid_last(page); |
03c5a6e1 | 1407 | count_vm_numa_event(NUMA_HINT_FAULTS); |
04bb2f94 | 1408 | if (page_nid == this_nid) { |
03c5a6e1 | 1409 | count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); |
04bb2f94 RR |
1410 | flags |= TNF_FAULT_LOCAL; |
1411 | } | |
4daae3b4 | 1412 | |
bea66fbd MG |
1413 | /* See similar comment in do_numa_page for explanation */ |
1414 | if (!(vma->vm_flags & VM_WRITE)) | |
6688cc05 PZ |
1415 | flags |= TNF_NO_GROUP; |
1416 | ||
ff9042b1 MG |
1417 | /* |
1418 | * Acquire the page lock to serialise THP migrations but avoid dropping | |
1419 | * page_table_lock if at all possible | |
1420 | */ | |
b8916634 MG |
1421 | page_locked = trylock_page(page); |
1422 | target_nid = mpol_misplaced(page, vma, haddr); | |
1423 | if (target_nid == -1) { | |
1424 | /* If the page was locked, there are no parallel migrations */ | |
a54a407f | 1425 | if (page_locked) |
b8916634 | 1426 | goto clear_pmdnuma; |
2b4847e7 | 1427 | } |
4daae3b4 | 1428 | |
de466bd6 | 1429 | /* Migration could have started since the pmd_trans_migrating check */ |
2b4847e7 | 1430 | if (!page_locked) { |
c4088ebd | 1431 | spin_unlock(ptl); |
b8916634 | 1432 | wait_on_page_locked(page); |
a54a407f | 1433 | page_nid = -1; |
b8916634 MG |
1434 | goto out; |
1435 | } | |
1436 | ||
2b4847e7 MG |
1437 | /* |
1438 | * Page is misplaced. Page lock serialises migrations. Acquire anon_vma | |
1439 | * to serialises splits | |
1440 | */ | |
b8916634 | 1441 | get_page(page); |
c4088ebd | 1442 | spin_unlock(ptl); |
b8916634 | 1443 | anon_vma = page_lock_anon_vma_read(page); |
4daae3b4 | 1444 | |
c69307d5 | 1445 | /* Confirm the PMD did not change while page_table_lock was released */ |
c4088ebd | 1446 | spin_lock(ptl); |
b32967ff MG |
1447 | if (unlikely(!pmd_same(pmd, *pmdp))) { |
1448 | unlock_page(page); | |
1449 | put_page(page); | |
a54a407f | 1450 | page_nid = -1; |
4daae3b4 | 1451 | goto out_unlock; |
b32967ff | 1452 | } |
ff9042b1 | 1453 | |
c3a489ca MG |
1454 | /* Bail if we fail to protect against THP splits for any reason */ |
1455 | if (unlikely(!anon_vma)) { | |
1456 | put_page(page); | |
1457 | page_nid = -1; | |
1458 | goto clear_pmdnuma; | |
1459 | } | |
1460 | ||
a54a407f MG |
1461 | /* |
1462 | * Migrate the THP to the requested node, returns with page unlocked | |
8a0516ed | 1463 | * and access rights restored. |
a54a407f | 1464 | */ |
c4088ebd | 1465 | spin_unlock(ptl); |
b32967ff | 1466 | migrated = migrate_misplaced_transhuge_page(mm, vma, |
340ef390 | 1467 | pmdp, pmd, addr, page, target_nid); |
6688cc05 PZ |
1468 | if (migrated) { |
1469 | flags |= TNF_MIGRATED; | |
8191acbd | 1470 | page_nid = target_nid; |
074c2381 MG |
1471 | } else |
1472 | flags |= TNF_MIGRATE_FAIL; | |
b32967ff | 1473 | |
8191acbd | 1474 | goto out; |
b32967ff | 1475 | clear_pmdnuma: |
a54a407f | 1476 | BUG_ON(!PageLocked(page)); |
b191f9b1 | 1477 | was_writable = pmd_write(pmd); |
4d942466 | 1478 | pmd = pmd_modify(pmd, vma->vm_page_prot); |
b7b04004 | 1479 | pmd = pmd_mkyoung(pmd); |
b191f9b1 MG |
1480 | if (was_writable) |
1481 | pmd = pmd_mkwrite(pmd); | |
d10e63f2 | 1482 | set_pmd_at(mm, haddr, pmdp, pmd); |
d10e63f2 | 1483 | update_mmu_cache_pmd(vma, addr, pmdp); |
a54a407f | 1484 | unlock_page(page); |
d10e63f2 | 1485 | out_unlock: |
c4088ebd | 1486 | spin_unlock(ptl); |
b8916634 MG |
1487 | |
1488 | out: | |
1489 | if (anon_vma) | |
1490 | page_unlock_anon_vma_read(anon_vma); | |
1491 | ||
8191acbd | 1492 | if (page_nid != -1) |
6688cc05 | 1493 | task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags); |
8191acbd | 1494 | |
d10e63f2 MG |
1495 | return 0; |
1496 | } | |
1497 | ||
71e3aac0 | 1498 | int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, |
f21760b1 | 1499 | pmd_t *pmd, unsigned long addr) |
71e3aac0 | 1500 | { |
da146769 | 1501 | pmd_t orig_pmd; |
bf929152 | 1502 | spinlock_t *ptl; |
71e3aac0 | 1503 | |
da146769 KS |
1504 | if (__pmd_trans_huge_lock(pmd, vma, &ptl) != 1) |
1505 | return 0; | |
1506 | /* | |
1507 | * For architectures like ppc64 we look at deposited pgtable | |
1508 | * when calling pmdp_huge_get_and_clear. So do the | |
1509 | * pgtable_trans_huge_withdraw after finishing pmdp related | |
1510 | * operations. | |
1511 | */ | |
1512 | orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd, | |
1513 | tlb->fullmm); | |
1514 | tlb_remove_pmd_tlb_entry(tlb, pmd, addr); | |
1515 | if (vma_is_dax(vma)) { | |
1516 | spin_unlock(ptl); | |
1517 | if (is_huge_zero_pmd(orig_pmd)) | |
97ae1749 | 1518 | put_huge_zero_page(); |
da146769 KS |
1519 | } else if (is_huge_zero_pmd(orig_pmd)) { |
1520 | pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd)); | |
1521 | atomic_long_dec(&tlb->mm->nr_ptes); | |
1522 | spin_unlock(ptl); | |
1523 | put_huge_zero_page(); | |
1524 | } else { | |
1525 | struct page *page = pmd_page(orig_pmd); | |
d281ee61 | 1526 | page_remove_rmap(page, true); |
da146769 KS |
1527 | VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); |
1528 | add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); | |
1529 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
1530 | pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd)); | |
1531 | atomic_long_dec(&tlb->mm->nr_ptes); | |
1532 | spin_unlock(ptl); | |
1533 | tlb_remove_page(tlb, page); | |
025c5b24 | 1534 | } |
da146769 | 1535 | return 1; |
71e3aac0 AA |
1536 | } |
1537 | ||
37a1c49a AA |
1538 | int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma, |
1539 | unsigned long old_addr, | |
1540 | unsigned long new_addr, unsigned long old_end, | |
1541 | pmd_t *old_pmd, pmd_t *new_pmd) | |
1542 | { | |
bf929152 | 1543 | spinlock_t *old_ptl, *new_ptl; |
37a1c49a AA |
1544 | int ret = 0; |
1545 | pmd_t pmd; | |
1546 | ||
1547 | struct mm_struct *mm = vma->vm_mm; | |
1548 | ||
1549 | if ((old_addr & ~HPAGE_PMD_MASK) || | |
1550 | (new_addr & ~HPAGE_PMD_MASK) || | |
1551 | old_end - old_addr < HPAGE_PMD_SIZE || | |
1552 | (new_vma->vm_flags & VM_NOHUGEPAGE)) | |
1553 | goto out; | |
1554 | ||
1555 | /* | |
1556 | * The destination pmd shouldn't be established, free_pgtables() | |
1557 | * should have release it. | |
1558 | */ | |
1559 | if (WARN_ON(!pmd_none(*new_pmd))) { | |
1560 | VM_BUG_ON(pmd_trans_huge(*new_pmd)); | |
1561 | goto out; | |
1562 | } | |
1563 | ||
bf929152 KS |
1564 | /* |
1565 | * We don't have to worry about the ordering of src and dst | |
1566 | * ptlocks because exclusive mmap_sem prevents deadlock. | |
1567 | */ | |
1568 | ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl); | |
025c5b24 | 1569 | if (ret == 1) { |
bf929152 KS |
1570 | new_ptl = pmd_lockptr(mm, new_pmd); |
1571 | if (new_ptl != old_ptl) | |
1572 | spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); | |
8809aa2d | 1573 | pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); |
025c5b24 | 1574 | VM_BUG_ON(!pmd_none(*new_pmd)); |
3592806c | 1575 | |
b3084f4d AK |
1576 | if (pmd_move_must_withdraw(new_ptl, old_ptl)) { |
1577 | pgtable_t pgtable; | |
3592806c KS |
1578 | pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); |
1579 | pgtable_trans_huge_deposit(mm, new_pmd, pgtable); | |
3592806c | 1580 | } |
b3084f4d AK |
1581 | set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd)); |
1582 | if (new_ptl != old_ptl) | |
1583 | spin_unlock(new_ptl); | |
bf929152 | 1584 | spin_unlock(old_ptl); |
37a1c49a AA |
1585 | } |
1586 | out: | |
1587 | return ret; | |
1588 | } | |
1589 | ||
f123d74a MG |
1590 | /* |
1591 | * Returns | |
1592 | * - 0 if PMD could not be locked | |
1593 | * - 1 if PMD was locked but protections unchange and TLB flush unnecessary | |
1594 | * - HPAGE_PMD_NR is protections changed and TLB flush necessary | |
1595 | */ | |
cd7548ab | 1596 | int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, |
e944fd67 | 1597 | unsigned long addr, pgprot_t newprot, int prot_numa) |
cd7548ab JW |
1598 | { |
1599 | struct mm_struct *mm = vma->vm_mm; | |
bf929152 | 1600 | spinlock_t *ptl; |
cd7548ab JW |
1601 | int ret = 0; |
1602 | ||
bf929152 | 1603 | if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) { |
025c5b24 | 1604 | pmd_t entry; |
b191f9b1 | 1605 | bool preserve_write = prot_numa && pmd_write(*pmd); |
ba68bc01 | 1606 | ret = 1; |
e944fd67 MG |
1607 | |
1608 | /* | |
1609 | * Avoid trapping faults against the zero page. The read-only | |
1610 | * data is likely to be read-cached on the local CPU and | |
1611 | * local/remote hits to the zero page are not interesting. | |
1612 | */ | |
1613 | if (prot_numa && is_huge_zero_pmd(*pmd)) { | |
1614 | spin_unlock(ptl); | |
ba68bc01 | 1615 | return ret; |
e944fd67 MG |
1616 | } |
1617 | ||
10c1045f | 1618 | if (!prot_numa || !pmd_protnone(*pmd)) { |
8809aa2d | 1619 | entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd); |
10c1045f | 1620 | entry = pmd_modify(entry, newprot); |
b191f9b1 MG |
1621 | if (preserve_write) |
1622 | entry = pmd_mkwrite(entry); | |
10c1045f MG |
1623 | ret = HPAGE_PMD_NR; |
1624 | set_pmd_at(mm, addr, pmd, entry); | |
b191f9b1 | 1625 | BUG_ON(!preserve_write && pmd_write(entry)); |
10c1045f | 1626 | } |
bf929152 | 1627 | spin_unlock(ptl); |
025c5b24 NH |
1628 | } |
1629 | ||
1630 | return ret; | |
1631 | } | |
1632 | ||
1633 | /* | |
1634 | * Returns 1 if a given pmd maps a stable (not under splitting) thp. | |
1635 | * Returns -1 if it maps a thp under splitting. Returns 0 otherwise. | |
1636 | * | |
1637 | * Note that if it returns 1, this routine returns without unlocking page | |
1638 | * table locks. So callers must unlock them. | |
1639 | */ | |
bf929152 KS |
1640 | int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma, |
1641 | spinlock_t **ptl) | |
025c5b24 | 1642 | { |
bf929152 | 1643 | *ptl = pmd_lock(vma->vm_mm, pmd); |
cd7548ab JW |
1644 | if (likely(pmd_trans_huge(*pmd))) { |
1645 | if (unlikely(pmd_trans_splitting(*pmd))) { | |
bf929152 | 1646 | spin_unlock(*ptl); |
cd7548ab | 1647 | wait_split_huge_page(vma->anon_vma, pmd); |
025c5b24 | 1648 | return -1; |
cd7548ab | 1649 | } else { |
025c5b24 NH |
1650 | /* Thp mapped by 'pmd' is stable, so we can |
1651 | * handle it as it is. */ | |
1652 | return 1; | |
cd7548ab | 1653 | } |
025c5b24 | 1654 | } |
bf929152 | 1655 | spin_unlock(*ptl); |
025c5b24 | 1656 | return 0; |
cd7548ab JW |
1657 | } |
1658 | ||
117b0791 KS |
1659 | /* |
1660 | * This function returns whether a given @page is mapped onto the @address | |
1661 | * in the virtual space of @mm. | |
1662 | * | |
1663 | * When it's true, this function returns *pmd with holding the page table lock | |
1664 | * and passing it back to the caller via @ptl. | |
1665 | * If it's false, returns NULL without holding the page table lock. | |
1666 | */ | |
71e3aac0 AA |
1667 | pmd_t *page_check_address_pmd(struct page *page, |
1668 | struct mm_struct *mm, | |
1669 | unsigned long address, | |
117b0791 KS |
1670 | enum page_check_address_pmd_flag flag, |
1671 | spinlock_t **ptl) | |
71e3aac0 | 1672 | { |
b5a8cad3 KS |
1673 | pgd_t *pgd; |
1674 | pud_t *pud; | |
117b0791 | 1675 | pmd_t *pmd; |
71e3aac0 AA |
1676 | |
1677 | if (address & ~HPAGE_PMD_MASK) | |
117b0791 | 1678 | return NULL; |
71e3aac0 | 1679 | |
b5a8cad3 KS |
1680 | pgd = pgd_offset(mm, address); |
1681 | if (!pgd_present(*pgd)) | |
117b0791 | 1682 | return NULL; |
b5a8cad3 KS |
1683 | pud = pud_offset(pgd, address); |
1684 | if (!pud_present(*pud)) | |
1685 | return NULL; | |
1686 | pmd = pmd_offset(pud, address); | |
1687 | ||
117b0791 | 1688 | *ptl = pmd_lock(mm, pmd); |
b5a8cad3 | 1689 | if (!pmd_present(*pmd)) |
117b0791 | 1690 | goto unlock; |
71e3aac0 | 1691 | if (pmd_page(*pmd) != page) |
117b0791 | 1692 | goto unlock; |
94fcc585 AA |
1693 | /* |
1694 | * split_vma() may create temporary aliased mappings. There is | |
1695 | * no risk as long as all huge pmd are found and have their | |
1696 | * splitting bit set before __split_huge_page_refcount | |
1697 | * runs. Finding the same huge pmd more than once during the | |
1698 | * same rmap walk is not a problem. | |
1699 | */ | |
1700 | if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG && | |
1701 | pmd_trans_splitting(*pmd)) | |
117b0791 | 1702 | goto unlock; |
71e3aac0 AA |
1703 | if (pmd_trans_huge(*pmd)) { |
1704 | VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG && | |
1705 | !pmd_trans_splitting(*pmd)); | |
117b0791 | 1706 | return pmd; |
71e3aac0 | 1707 | } |
117b0791 KS |
1708 | unlock: |
1709 | spin_unlock(*ptl); | |
1710 | return NULL; | |
71e3aac0 AA |
1711 | } |
1712 | ||
1713 | static int __split_huge_page_splitting(struct page *page, | |
1714 | struct vm_area_struct *vma, | |
1715 | unsigned long address) | |
1716 | { | |
1717 | struct mm_struct *mm = vma->vm_mm; | |
117b0791 | 1718 | spinlock_t *ptl; |
71e3aac0 AA |
1719 | pmd_t *pmd; |
1720 | int ret = 0; | |
2ec74c3e SG |
1721 | /* For mmu_notifiers */ |
1722 | const unsigned long mmun_start = address; | |
1723 | const unsigned long mmun_end = address + HPAGE_PMD_SIZE; | |
71e3aac0 | 1724 | |
2ec74c3e | 1725 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
71e3aac0 | 1726 | pmd = page_check_address_pmd(page, mm, address, |
117b0791 | 1727 | PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl); |
71e3aac0 AA |
1728 | if (pmd) { |
1729 | /* | |
1730 | * We can't temporarily set the pmd to null in order | |
1731 | * to split it, the pmd must remain marked huge at all | |
1732 | * times or the VM won't take the pmd_trans_huge paths | |
5a505085 | 1733 | * and it won't wait on the anon_vma->root->rwsem to |
71e3aac0 AA |
1734 | * serialize against split_huge_page*. |
1735 | */ | |
2ec74c3e | 1736 | pmdp_splitting_flush(vma, address, pmd); |
34ee645e | 1737 | |
71e3aac0 | 1738 | ret = 1; |
117b0791 | 1739 | spin_unlock(ptl); |
71e3aac0 | 1740 | } |
2ec74c3e | 1741 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
71e3aac0 AA |
1742 | |
1743 | return ret; | |
1744 | } | |
1745 | ||
5bc7b8ac SL |
1746 | static void __split_huge_page_refcount(struct page *page, |
1747 | struct list_head *list) | |
71e3aac0 AA |
1748 | { |
1749 | int i; | |
71e3aac0 | 1750 | struct zone *zone = page_zone(page); |
fa9add64 | 1751 | struct lruvec *lruvec; |
70b50f94 | 1752 | int tail_count = 0; |
71e3aac0 AA |
1753 | |
1754 | /* prevent PageLRU to go away from under us, and freeze lru stats */ | |
1755 | spin_lock_irq(&zone->lru_lock); | |
fa9add64 HD |
1756 | lruvec = mem_cgroup_page_lruvec(page, zone); |
1757 | ||
71e3aac0 | 1758 | compound_lock(page); |
e94c8a9c KH |
1759 | /* complete memcg works before add pages to LRU */ |
1760 | mem_cgroup_split_huge_fixup(page); | |
71e3aac0 | 1761 | |
45676885 | 1762 | for (i = HPAGE_PMD_NR - 1; i >= 1; i--) { |
71e3aac0 AA |
1763 | struct page *page_tail = page + i; |
1764 | ||
70b50f94 AA |
1765 | /* tail_page->_mapcount cannot change */ |
1766 | BUG_ON(page_mapcount(page_tail) < 0); | |
1767 | tail_count += page_mapcount(page_tail); | |
1768 | /* check for overflow */ | |
1769 | BUG_ON(tail_count < 0); | |
1770 | BUG_ON(atomic_read(&page_tail->_count) != 0); | |
1771 | /* | |
1772 | * tail_page->_count is zero and not changing from | |
1773 | * under us. But get_page_unless_zero() may be running | |
1774 | * from under us on the tail_page. If we used | |
1775 | * atomic_set() below instead of atomic_add(), we | |
1776 | * would then run atomic_set() concurrently with | |
1777 | * get_page_unless_zero(), and atomic_set() is | |
1778 | * implemented in C not using locked ops. spin_unlock | |
1779 | * on x86 sometime uses locked ops because of PPro | |
1780 | * errata 66, 92, so unless somebody can guarantee | |
1781 | * atomic_set() here would be safe on all archs (and | |
1782 | * not only on x86), it's safer to use atomic_add(). | |
1783 | */ | |
1784 | atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1, | |
1785 | &page_tail->_count); | |
71e3aac0 AA |
1786 | |
1787 | /* after clearing PageTail the gup refcount can be released */ | |
3a79d52a | 1788 | smp_mb__after_atomic(); |
71e3aac0 | 1789 | |
f4c18e6f | 1790 | page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; |
71e3aac0 AA |
1791 | page_tail->flags |= (page->flags & |
1792 | ((1L << PG_referenced) | | |
1793 | (1L << PG_swapbacked) | | |
1794 | (1L << PG_mlocked) | | |
e180cf80 KS |
1795 | (1L << PG_uptodate) | |
1796 | (1L << PG_active) | | |
1797 | (1L << PG_unevictable))); | |
71e3aac0 AA |
1798 | page_tail->flags |= (1L << PG_dirty); |
1799 | ||
1d798ca3 | 1800 | clear_compound_head(page_tail); |
71e3aac0 | 1801 | |
33c3fc71 VD |
1802 | if (page_is_young(page)) |
1803 | set_page_young(page_tail); | |
1804 | if (page_is_idle(page)) | |
1805 | set_page_idle(page_tail); | |
1806 | ||
71e3aac0 AA |
1807 | /* |
1808 | * __split_huge_page_splitting() already set the | |
1809 | * splitting bit in all pmd that could map this | |
1810 | * hugepage, that will ensure no CPU can alter the | |
1811 | * mapcount on the head page. The mapcount is only | |
1812 | * accounted in the head page and it has to be | |
1813 | * transferred to all tail pages in the below code. So | |
1814 | * for this code to be safe, the split the mapcount | |
1815 | * can't change. But that doesn't mean userland can't | |
1816 | * keep changing and reading the page contents while | |
1817 | * we transfer the mapcount, so the pmd splitting | |
1818 | * status is achieved setting a reserved bit in the | |
1819 | * pmd, not by clearing the present bit. | |
1820 | */ | |
71e3aac0 AA |
1821 | page_tail->_mapcount = page->_mapcount; |
1822 | ||
1c290f64 | 1823 | BUG_ON(page_tail->mapping != TAIL_MAPPING); |
71e3aac0 AA |
1824 | page_tail->mapping = page->mapping; |
1825 | ||
45676885 | 1826 | page_tail->index = page->index + i; |
90572890 | 1827 | page_cpupid_xchg_last(page_tail, page_cpupid_last(page)); |
71e3aac0 AA |
1828 | |
1829 | BUG_ON(!PageAnon(page_tail)); | |
1830 | BUG_ON(!PageUptodate(page_tail)); | |
1831 | BUG_ON(!PageDirty(page_tail)); | |
1832 | BUG_ON(!PageSwapBacked(page_tail)); | |
1833 | ||
5bc7b8ac | 1834 | lru_add_page_tail(page, page_tail, lruvec, list); |
71e3aac0 | 1835 | } |
70b50f94 AA |
1836 | atomic_sub(tail_count, &page->_count); |
1837 | BUG_ON(atomic_read(&page->_count) <= 0); | |
71e3aac0 | 1838 | |
fa9add64 | 1839 | __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1); |
79134171 | 1840 | |
71e3aac0 AA |
1841 | ClearPageCompound(page); |
1842 | compound_unlock(page); | |
1843 | spin_unlock_irq(&zone->lru_lock); | |
1844 | ||
1845 | for (i = 1; i < HPAGE_PMD_NR; i++) { | |
1846 | struct page *page_tail = page + i; | |
1847 | BUG_ON(page_count(page_tail) <= 0); | |
1848 | /* | |
1849 | * Tail pages may be freed if there wasn't any mapping | |
1850 | * like if add_to_swap() is running on a lru page that | |
1851 | * had its mapping zapped. And freeing these pages | |
1852 | * requires taking the lru_lock so we do the put_page | |
1853 | * of the tail pages after the split is complete. | |
1854 | */ | |
1855 | put_page(page_tail); | |
1856 | } | |
1857 | ||
1858 | /* | |
1859 | * Only the head page (now become a regular page) is required | |
1860 | * to be pinned by the caller. | |
1861 | */ | |
1862 | BUG_ON(page_count(page) <= 0); | |
1863 | } | |
1864 | ||
1865 | static int __split_huge_page_map(struct page *page, | |
1866 | struct vm_area_struct *vma, | |
1867 | unsigned long address) | |
1868 | { | |
1869 | struct mm_struct *mm = vma->vm_mm; | |
117b0791 | 1870 | spinlock_t *ptl; |
71e3aac0 AA |
1871 | pmd_t *pmd, _pmd; |
1872 | int ret = 0, i; | |
1873 | pgtable_t pgtable; | |
1874 | unsigned long haddr; | |
1875 | ||
71e3aac0 | 1876 | pmd = page_check_address_pmd(page, mm, address, |
117b0791 | 1877 | PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl); |
71e3aac0 | 1878 | if (pmd) { |
6b0b50b0 | 1879 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); |
71e3aac0 | 1880 | pmd_populate(mm, &_pmd, pgtable); |
f8303c25 WL |
1881 | if (pmd_write(*pmd)) |
1882 | BUG_ON(page_mapcount(page) != 1); | |
71e3aac0 | 1883 | |
e3ebcf64 GS |
1884 | haddr = address; |
1885 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
71e3aac0 AA |
1886 | pte_t *pte, entry; |
1887 | BUG_ON(PageCompound(page+i)); | |
abc40bd2 | 1888 | /* |
8a0516ed MG |
1889 | * Note that NUMA hinting access restrictions are not |
1890 | * transferred to avoid any possibility of altering | |
1891 | * permissions across VMAs. | |
abc40bd2 | 1892 | */ |
71e3aac0 AA |
1893 | entry = mk_pte(page + i, vma->vm_page_prot); |
1894 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); | |
1895 | if (!pmd_write(*pmd)) | |
1896 | entry = pte_wrprotect(entry); | |
71e3aac0 AA |
1897 | if (!pmd_young(*pmd)) |
1898 | entry = pte_mkold(entry); | |
1899 | pte = pte_offset_map(&_pmd, haddr); | |
1900 | BUG_ON(!pte_none(*pte)); | |
1901 | set_pte_at(mm, haddr, pte, entry); | |
1902 | pte_unmap(pte); | |
1903 | } | |
1904 | ||
71e3aac0 AA |
1905 | smp_wmb(); /* make pte visible before pmd */ |
1906 | /* | |
1907 | * Up to this point the pmd is present and huge and | |
1908 | * userland has the whole access to the hugepage | |
1909 | * during the split (which happens in place). If we | |
1910 | * overwrite the pmd with the not-huge version | |
1911 | * pointing to the pte here (which of course we could | |
1912 | * if all CPUs were bug free), userland could trigger | |
1913 | * a small page size TLB miss on the small sized TLB | |
1914 | * while the hugepage TLB entry is still established | |
1915 | * in the huge TLB. Some CPU doesn't like that. See | |
1916 | * http://support.amd.com/us/Processor_TechDocs/41322.pdf, | |
1917 | * Erratum 383 on page 93. Intel should be safe but is | |
1918 | * also warns that it's only safe if the permission | |
1919 | * and cache attributes of the two entries loaded in | |
1920 | * the two TLB is identical (which should be the case | |
1921 | * here). But it is generally safer to never allow | |
1922 | * small and huge TLB entries for the same virtual | |
1923 | * address to be loaded simultaneously. So instead of | |
12ebc158 | 1924 | * doing "pmd_populate(); flush_pmd_tlb_range();" we first |
71e3aac0 AA |
1925 | * mark the current pmd notpresent (atomically because |
1926 | * here the pmd_trans_huge and pmd_trans_splitting | |
1927 | * must remain set at all times on the pmd until the | |
1928 | * split is complete for this pmd), then we flush the | |
1929 | * SMP TLB and finally we write the non-huge version | |
1930 | * of the pmd entry with pmd_populate. | |
1931 | */ | |
46dcde73 | 1932 | pmdp_invalidate(vma, address, pmd); |
71e3aac0 AA |
1933 | pmd_populate(mm, pmd, pgtable); |
1934 | ret = 1; | |
117b0791 | 1935 | spin_unlock(ptl); |
71e3aac0 | 1936 | } |
71e3aac0 AA |
1937 | |
1938 | return ret; | |
1939 | } | |
1940 | ||
5a505085 | 1941 | /* must be called with anon_vma->root->rwsem held */ |
71e3aac0 | 1942 | static void __split_huge_page(struct page *page, |
5bc7b8ac SL |
1943 | struct anon_vma *anon_vma, |
1944 | struct list_head *list) | |
71e3aac0 AA |
1945 | { |
1946 | int mapcount, mapcount2; | |
bf181b9f | 1947 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); |
71e3aac0 AA |
1948 | struct anon_vma_chain *avc; |
1949 | ||
1950 | BUG_ON(!PageHead(page)); | |
1951 | BUG_ON(PageTail(page)); | |
1952 | ||
1953 | mapcount = 0; | |
bf181b9f | 1954 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { |
71e3aac0 AA |
1955 | struct vm_area_struct *vma = avc->vma; |
1956 | unsigned long addr = vma_address(page, vma); | |
1957 | BUG_ON(is_vma_temporary_stack(vma)); | |
71e3aac0 AA |
1958 | mapcount += __split_huge_page_splitting(page, vma, addr); |
1959 | } | |
05759d38 AA |
1960 | /* |
1961 | * It is critical that new vmas are added to the tail of the | |
1962 | * anon_vma list. This guarantes that if copy_huge_pmd() runs | |
1963 | * and establishes a child pmd before | |
1964 | * __split_huge_page_splitting() freezes the parent pmd (so if | |
1965 | * we fail to prevent copy_huge_pmd() from running until the | |
1966 | * whole __split_huge_page() is complete), we will still see | |
1967 | * the newly established pmd of the child later during the | |
1968 | * walk, to be able to set it as pmd_trans_splitting too. | |
1969 | */ | |
ff9e43eb | 1970 | if (mapcount != page_mapcount(page)) { |
ae3a8c1c AM |
1971 | pr_err("mapcount %d page_mapcount %d\n", |
1972 | mapcount, page_mapcount(page)); | |
ff9e43eb KS |
1973 | BUG(); |
1974 | } | |
71e3aac0 | 1975 | |
5bc7b8ac | 1976 | __split_huge_page_refcount(page, list); |
71e3aac0 AA |
1977 | |
1978 | mapcount2 = 0; | |
bf181b9f | 1979 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { |
71e3aac0 AA |
1980 | struct vm_area_struct *vma = avc->vma; |
1981 | unsigned long addr = vma_address(page, vma); | |
1982 | BUG_ON(is_vma_temporary_stack(vma)); | |
71e3aac0 AA |
1983 | mapcount2 += __split_huge_page_map(page, vma, addr); |
1984 | } | |
ff9e43eb | 1985 | if (mapcount != mapcount2) { |
ae3a8c1c AM |
1986 | pr_err("mapcount %d mapcount2 %d page_mapcount %d\n", |
1987 | mapcount, mapcount2, page_mapcount(page)); | |
ff9e43eb KS |
1988 | BUG(); |
1989 | } | |
71e3aac0 AA |
1990 | } |
1991 | ||
5bc7b8ac SL |
1992 | /* |
1993 | * Split a hugepage into normal pages. This doesn't change the position of head | |
1994 | * page. If @list is null, tail pages will be added to LRU list, otherwise, to | |
1995 | * @list. Both head page and tail pages will inherit mapping, flags, and so on | |
1996 | * from the hugepage. | |
1997 | * Return 0 if the hugepage is split successfully otherwise return 1. | |
1998 | */ | |
1999 | int split_huge_page_to_list(struct page *page, struct list_head *list) | |
71e3aac0 AA |
2000 | { |
2001 | struct anon_vma *anon_vma; | |
2002 | int ret = 1; | |
2003 | ||
5918d10a | 2004 | BUG_ON(is_huge_zero_page(page)); |
71e3aac0 | 2005 | BUG_ON(!PageAnon(page)); |
062f1af2 MG |
2006 | |
2007 | /* | |
2008 | * The caller does not necessarily hold an mmap_sem that would prevent | |
2009 | * the anon_vma disappearing so we first we take a reference to it | |
2010 | * and then lock the anon_vma for write. This is similar to | |
2011 | * page_lock_anon_vma_read except the write lock is taken to serialise | |
2012 | * against parallel split or collapse operations. | |
2013 | */ | |
2014 | anon_vma = page_get_anon_vma(page); | |
71e3aac0 AA |
2015 | if (!anon_vma) |
2016 | goto out; | |
062f1af2 MG |
2017 | anon_vma_lock_write(anon_vma); |
2018 | ||
71e3aac0 AA |
2019 | ret = 0; |
2020 | if (!PageCompound(page)) | |
2021 | goto out_unlock; | |
2022 | ||
2023 | BUG_ON(!PageSwapBacked(page)); | |
5bc7b8ac | 2024 | __split_huge_page(page, anon_vma, list); |
122afea9 | 2025 | count_vm_event(THP_SPLIT_PAGE); |
71e3aac0 AA |
2026 | |
2027 | BUG_ON(PageCompound(page)); | |
2028 | out_unlock: | |
08b52706 | 2029 | anon_vma_unlock_write(anon_vma); |
062f1af2 | 2030 | put_anon_vma(anon_vma); |
71e3aac0 AA |
2031 | out: |
2032 | return ret; | |
2033 | } | |
2034 | ||
9050d7eb | 2035 | #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE) |
78f11a25 | 2036 | |
60ab3244 AA |
2037 | int hugepage_madvise(struct vm_area_struct *vma, |
2038 | unsigned long *vm_flags, int advice) | |
0af4e98b | 2039 | { |
a664b2d8 AA |
2040 | switch (advice) { |
2041 | case MADV_HUGEPAGE: | |
1e1836e8 AT |
2042 | #ifdef CONFIG_S390 |
2043 | /* | |
2044 | * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390 | |
2045 | * can't handle this properly after s390_enable_sie, so we simply | |
2046 | * ignore the madvise to prevent qemu from causing a SIGSEGV. | |
2047 | */ | |
2048 | if (mm_has_pgste(vma->vm_mm)) | |
2049 | return 0; | |
2050 | #endif | |
a664b2d8 AA |
2051 | /* |
2052 | * Be somewhat over-protective like KSM for now! | |
2053 | */ | |
1a763615 | 2054 | if (*vm_flags & VM_NO_THP) |
a664b2d8 AA |
2055 | return -EINVAL; |
2056 | *vm_flags &= ~VM_NOHUGEPAGE; | |
2057 | *vm_flags |= VM_HUGEPAGE; | |
60ab3244 AA |
2058 | /* |
2059 | * If the vma become good for khugepaged to scan, | |
2060 | * register it here without waiting a page fault that | |
2061 | * may not happen any time soon. | |
2062 | */ | |
6d50e60c | 2063 | if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags))) |
60ab3244 | 2064 | return -ENOMEM; |
a664b2d8 AA |
2065 | break; |
2066 | case MADV_NOHUGEPAGE: | |
2067 | /* | |
2068 | * Be somewhat over-protective like KSM for now! | |
2069 | */ | |
1a763615 | 2070 | if (*vm_flags & VM_NO_THP) |
a664b2d8 AA |
2071 | return -EINVAL; |
2072 | *vm_flags &= ~VM_HUGEPAGE; | |
2073 | *vm_flags |= VM_NOHUGEPAGE; | |
60ab3244 AA |
2074 | /* |
2075 | * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning | |
2076 | * this vma even if we leave the mm registered in khugepaged if | |
2077 | * it got registered before VM_NOHUGEPAGE was set. | |
2078 | */ | |
a664b2d8 AA |
2079 | break; |
2080 | } | |
0af4e98b AA |
2081 | |
2082 | return 0; | |
2083 | } | |
2084 | ||
ba76149f AA |
2085 | static int __init khugepaged_slab_init(void) |
2086 | { | |
2087 | mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", | |
2088 | sizeof(struct mm_slot), | |
2089 | __alignof__(struct mm_slot), 0, NULL); | |
2090 | if (!mm_slot_cache) | |
2091 | return -ENOMEM; | |
2092 | ||
2093 | return 0; | |
2094 | } | |
2095 | ||
65ebb64f KS |
2096 | static void __init khugepaged_slab_exit(void) |
2097 | { | |
2098 | kmem_cache_destroy(mm_slot_cache); | |
2099 | } | |
2100 | ||
ba76149f AA |
2101 | static inline struct mm_slot *alloc_mm_slot(void) |
2102 | { | |
2103 | if (!mm_slot_cache) /* initialization failed */ | |
2104 | return NULL; | |
2105 | return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); | |
2106 | } | |
2107 | ||
2108 | static inline void free_mm_slot(struct mm_slot *mm_slot) | |
2109 | { | |
2110 | kmem_cache_free(mm_slot_cache, mm_slot); | |
2111 | } | |
2112 | ||
ba76149f AA |
2113 | static struct mm_slot *get_mm_slot(struct mm_struct *mm) |
2114 | { | |
2115 | struct mm_slot *mm_slot; | |
ba76149f | 2116 | |
b67bfe0d | 2117 | hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm) |
ba76149f AA |
2118 | if (mm == mm_slot->mm) |
2119 | return mm_slot; | |
43b5fbbd | 2120 | |
ba76149f AA |
2121 | return NULL; |
2122 | } | |
2123 | ||
2124 | static void insert_to_mm_slots_hash(struct mm_struct *mm, | |
2125 | struct mm_slot *mm_slot) | |
2126 | { | |
ba76149f | 2127 | mm_slot->mm = mm; |
43b5fbbd | 2128 | hash_add(mm_slots_hash, &mm_slot->hash, (long)mm); |
ba76149f AA |
2129 | } |
2130 | ||
2131 | static inline int khugepaged_test_exit(struct mm_struct *mm) | |
2132 | { | |
2133 | return atomic_read(&mm->mm_users) == 0; | |
2134 | } | |
2135 | ||
2136 | int __khugepaged_enter(struct mm_struct *mm) | |
2137 | { | |
2138 | struct mm_slot *mm_slot; | |
2139 | int wakeup; | |
2140 | ||
2141 | mm_slot = alloc_mm_slot(); | |
2142 | if (!mm_slot) | |
2143 | return -ENOMEM; | |
2144 | ||
2145 | /* __khugepaged_exit() must not run from under us */ | |
96dad67f | 2146 | VM_BUG_ON_MM(khugepaged_test_exit(mm), mm); |
ba76149f AA |
2147 | if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { |
2148 | free_mm_slot(mm_slot); | |
2149 | return 0; | |
2150 | } | |
2151 | ||
2152 | spin_lock(&khugepaged_mm_lock); | |
2153 | insert_to_mm_slots_hash(mm, mm_slot); | |
2154 | /* | |
2155 | * Insert just behind the scanning cursor, to let the area settle | |
2156 | * down a little. | |
2157 | */ | |
2158 | wakeup = list_empty(&khugepaged_scan.mm_head); | |
2159 | list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head); | |
2160 | spin_unlock(&khugepaged_mm_lock); | |
2161 | ||
2162 | atomic_inc(&mm->mm_count); | |
2163 | if (wakeup) | |
2164 | wake_up_interruptible(&khugepaged_wait); | |
2165 | ||
2166 | return 0; | |
2167 | } | |
2168 | ||
6d50e60c DR |
2169 | int khugepaged_enter_vma_merge(struct vm_area_struct *vma, |
2170 | unsigned long vm_flags) | |
ba76149f AA |
2171 | { |
2172 | unsigned long hstart, hend; | |
2173 | if (!vma->anon_vma) | |
2174 | /* | |
2175 | * Not yet faulted in so we will register later in the | |
2176 | * page fault if needed. | |
2177 | */ | |
2178 | return 0; | |
78f11a25 | 2179 | if (vma->vm_ops) |
ba76149f AA |
2180 | /* khugepaged not yet working on file or special mappings */ |
2181 | return 0; | |
6d50e60c | 2182 | VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma); |
ba76149f AA |
2183 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; |
2184 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
2185 | if (hstart < hend) | |
6d50e60c | 2186 | return khugepaged_enter(vma, vm_flags); |
ba76149f AA |
2187 | return 0; |
2188 | } | |
2189 | ||
2190 | void __khugepaged_exit(struct mm_struct *mm) | |
2191 | { | |
2192 | struct mm_slot *mm_slot; | |
2193 | int free = 0; | |
2194 | ||
2195 | spin_lock(&khugepaged_mm_lock); | |
2196 | mm_slot = get_mm_slot(mm); | |
2197 | if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { | |
43b5fbbd | 2198 | hash_del(&mm_slot->hash); |
ba76149f AA |
2199 | list_del(&mm_slot->mm_node); |
2200 | free = 1; | |
2201 | } | |
d788e80a | 2202 | spin_unlock(&khugepaged_mm_lock); |
ba76149f AA |
2203 | |
2204 | if (free) { | |
ba76149f AA |
2205 | clear_bit(MMF_VM_HUGEPAGE, &mm->flags); |
2206 | free_mm_slot(mm_slot); | |
2207 | mmdrop(mm); | |
2208 | } else if (mm_slot) { | |
ba76149f AA |
2209 | /* |
2210 | * This is required to serialize against | |
2211 | * khugepaged_test_exit() (which is guaranteed to run | |
2212 | * under mmap sem read mode). Stop here (after we | |
2213 | * return all pagetables will be destroyed) until | |
2214 | * khugepaged has finished working on the pagetables | |
2215 | * under the mmap_sem. | |
2216 | */ | |
2217 | down_write(&mm->mmap_sem); | |
2218 | up_write(&mm->mmap_sem); | |
d788e80a | 2219 | } |
ba76149f AA |
2220 | } |
2221 | ||
2222 | static void release_pte_page(struct page *page) | |
2223 | { | |
2224 | /* 0 stands for page_is_file_cache(page) == false */ | |
2225 | dec_zone_page_state(page, NR_ISOLATED_ANON + 0); | |
2226 | unlock_page(page); | |
2227 | putback_lru_page(page); | |
2228 | } | |
2229 | ||
2230 | static void release_pte_pages(pte_t *pte, pte_t *_pte) | |
2231 | { | |
2232 | while (--_pte >= pte) { | |
2233 | pte_t pteval = *_pte; | |
ca0984ca | 2234 | if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval))) |
ba76149f AA |
2235 | release_pte_page(pte_page(pteval)); |
2236 | } | |
2237 | } | |
2238 | ||
ba76149f AA |
2239 | static int __collapse_huge_page_isolate(struct vm_area_struct *vma, |
2240 | unsigned long address, | |
2241 | pte_t *pte) | |
2242 | { | |
7d2eba05 | 2243 | struct page *page = NULL; |
ba76149f | 2244 | pte_t *_pte; |
7d2eba05 | 2245 | int none_or_zero = 0, result = 0; |
10359213 | 2246 | bool referenced = false, writable = false; |
7d2eba05 | 2247 | |
ba76149f AA |
2248 | for (_pte = pte; _pte < pte+HPAGE_PMD_NR; |
2249 | _pte++, address += PAGE_SIZE) { | |
2250 | pte_t pteval = *_pte; | |
47aee4d8 MK |
2251 | if (pte_none(pteval) || (pte_present(pteval) && |
2252 | is_zero_pfn(pte_pfn(pteval)))) { | |
c1294d05 | 2253 | if (!userfaultfd_armed(vma) && |
7d2eba05 | 2254 | ++none_or_zero <= khugepaged_max_ptes_none) { |
ba76149f | 2255 | continue; |
7d2eba05 EA |
2256 | } else { |
2257 | result = SCAN_EXCEED_NONE_PTE; | |
ba76149f | 2258 | goto out; |
7d2eba05 | 2259 | } |
ba76149f | 2260 | } |
7d2eba05 EA |
2261 | if (!pte_present(pteval)) { |
2262 | result = SCAN_PTE_NON_PRESENT; | |
ba76149f | 2263 | goto out; |
7d2eba05 | 2264 | } |
ba76149f | 2265 | page = vm_normal_page(vma, address, pteval); |
7d2eba05 EA |
2266 | if (unlikely(!page)) { |
2267 | result = SCAN_PAGE_NULL; | |
ba76149f | 2268 | goto out; |
7d2eba05 | 2269 | } |
344aa35c | 2270 | |
309381fe SL |
2271 | VM_BUG_ON_PAGE(PageCompound(page), page); |
2272 | VM_BUG_ON_PAGE(!PageAnon(page), page); | |
2273 | VM_BUG_ON_PAGE(!PageSwapBacked(page), page); | |
ba76149f | 2274 | |
ba76149f AA |
2275 | /* |
2276 | * We can do it before isolate_lru_page because the | |
2277 | * page can't be freed from under us. NOTE: PG_lock | |
2278 | * is needed to serialize against split_huge_page | |
2279 | * when invoked from the VM. | |
2280 | */ | |
7d2eba05 EA |
2281 | if (!trylock_page(page)) { |
2282 | result = SCAN_PAGE_LOCK; | |
ba76149f | 2283 | goto out; |
7d2eba05 | 2284 | } |
10359213 EA |
2285 | |
2286 | /* | |
2287 | * cannot use mapcount: can't collapse if there's a gup pin. | |
2288 | * The page must only be referenced by the scanned process | |
2289 | * and page swap cache. | |
2290 | */ | |
2291 | if (page_count(page) != 1 + !!PageSwapCache(page)) { | |
2292 | unlock_page(page); | |
7d2eba05 | 2293 | result = SCAN_PAGE_COUNT; |
10359213 EA |
2294 | goto out; |
2295 | } | |
2296 | if (pte_write(pteval)) { | |
2297 | writable = true; | |
2298 | } else { | |
2299 | if (PageSwapCache(page) && !reuse_swap_page(page)) { | |
2300 | unlock_page(page); | |
7d2eba05 | 2301 | result = SCAN_SWAP_CACHE_PAGE; |
10359213 EA |
2302 | goto out; |
2303 | } | |
2304 | /* | |
2305 | * Page is not in the swap cache. It can be collapsed | |
2306 | * into a THP. | |
2307 | */ | |
2308 | } | |
2309 | ||
ba76149f AA |
2310 | /* |
2311 | * Isolate the page to avoid collapsing an hugepage | |
2312 | * currently in use by the VM. | |
2313 | */ | |
2314 | if (isolate_lru_page(page)) { | |
2315 | unlock_page(page); | |
7d2eba05 | 2316 | result = SCAN_DEL_PAGE_LRU; |
ba76149f AA |
2317 | goto out; |
2318 | } | |
2319 | /* 0 stands for page_is_file_cache(page) == false */ | |
2320 | inc_zone_page_state(page, NR_ISOLATED_ANON + 0); | |
309381fe SL |
2321 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
2322 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
ba76149f AA |
2323 | |
2324 | /* If there is no mapped pte young don't collapse the page */ | |
33c3fc71 VD |
2325 | if (pte_young(pteval) || |
2326 | page_is_young(page) || PageReferenced(page) || | |
8ee53820 | 2327 | mmu_notifier_test_young(vma->vm_mm, address)) |
10359213 | 2328 | referenced = true; |
ba76149f | 2329 | } |
7d2eba05 EA |
2330 | if (likely(writable)) { |
2331 | if (likely(referenced)) { | |
2332 | result = SCAN_SUCCEED; | |
2333 | trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero, | |
2334 | referenced, writable, result); | |
2335 | return 1; | |
2336 | } | |
2337 | } else { | |
2338 | result = SCAN_PAGE_RO; | |
2339 | } | |
2340 | ||
ba76149f | 2341 | out: |
344aa35c | 2342 | release_pte_pages(pte, _pte); |
7d2eba05 EA |
2343 | trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero, |
2344 | referenced, writable, result); | |
344aa35c | 2345 | return 0; |
ba76149f AA |
2346 | } |
2347 | ||
2348 | static void __collapse_huge_page_copy(pte_t *pte, struct page *page, | |
2349 | struct vm_area_struct *vma, | |
2350 | unsigned long address, | |
2351 | spinlock_t *ptl) | |
2352 | { | |
2353 | pte_t *_pte; | |
2354 | for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) { | |
2355 | pte_t pteval = *_pte; | |
2356 | struct page *src_page; | |
2357 | ||
ca0984ca | 2358 | if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { |
ba76149f AA |
2359 | clear_user_highpage(page, address); |
2360 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); | |
ca0984ca EA |
2361 | if (is_zero_pfn(pte_pfn(pteval))) { |
2362 | /* | |
2363 | * ptl mostly unnecessary. | |
2364 | */ | |
2365 | spin_lock(ptl); | |
2366 | /* | |
2367 | * paravirt calls inside pte_clear here are | |
2368 | * superfluous. | |
2369 | */ | |
2370 | pte_clear(vma->vm_mm, address, _pte); | |
2371 | spin_unlock(ptl); | |
2372 | } | |
ba76149f AA |
2373 | } else { |
2374 | src_page = pte_page(pteval); | |
2375 | copy_user_highpage(page, src_page, address, vma); | |
309381fe | 2376 | VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page); |
ba76149f AA |
2377 | release_pte_page(src_page); |
2378 | /* | |
2379 | * ptl mostly unnecessary, but preempt has to | |
2380 | * be disabled to update the per-cpu stats | |
2381 | * inside page_remove_rmap(). | |
2382 | */ | |
2383 | spin_lock(ptl); | |
2384 | /* | |
2385 | * paravirt calls inside pte_clear here are | |
2386 | * superfluous. | |
2387 | */ | |
2388 | pte_clear(vma->vm_mm, address, _pte); | |
d281ee61 | 2389 | page_remove_rmap(src_page, false); |
ba76149f AA |
2390 | spin_unlock(ptl); |
2391 | free_page_and_swap_cache(src_page); | |
2392 | } | |
2393 | ||
2394 | address += PAGE_SIZE; | |
2395 | page++; | |
2396 | } | |
2397 | } | |
2398 | ||
26234f36 | 2399 | static void khugepaged_alloc_sleep(void) |
ba76149f | 2400 | { |
bde43c6c PM |
2401 | DEFINE_WAIT(wait); |
2402 | ||
2403 | add_wait_queue(&khugepaged_wait, &wait); | |
2404 | freezable_schedule_timeout_interruptible( | |
2405 | msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); | |
2406 | remove_wait_queue(&khugepaged_wait, &wait); | |
26234f36 | 2407 | } |
ba76149f | 2408 | |
9f1b868a BL |
2409 | static int khugepaged_node_load[MAX_NUMNODES]; |
2410 | ||
14a4e214 DR |
2411 | static bool khugepaged_scan_abort(int nid) |
2412 | { | |
2413 | int i; | |
2414 | ||
2415 | /* | |
2416 | * If zone_reclaim_mode is disabled, then no extra effort is made to | |
2417 | * allocate memory locally. | |
2418 | */ | |
2419 | if (!zone_reclaim_mode) | |
2420 | return false; | |
2421 | ||
2422 | /* If there is a count for this node already, it must be acceptable */ | |
2423 | if (khugepaged_node_load[nid]) | |
2424 | return false; | |
2425 | ||
2426 | for (i = 0; i < MAX_NUMNODES; i++) { | |
2427 | if (!khugepaged_node_load[i]) | |
2428 | continue; | |
2429 | if (node_distance(nid, i) > RECLAIM_DISTANCE) | |
2430 | return true; | |
2431 | } | |
2432 | return false; | |
2433 | } | |
2434 | ||
26234f36 | 2435 | #ifdef CONFIG_NUMA |
9f1b868a BL |
2436 | static int khugepaged_find_target_node(void) |
2437 | { | |
2438 | static int last_khugepaged_target_node = NUMA_NO_NODE; | |
2439 | int nid, target_node = 0, max_value = 0; | |
2440 | ||
2441 | /* find first node with max normal pages hit */ | |
2442 | for (nid = 0; nid < MAX_NUMNODES; nid++) | |
2443 | if (khugepaged_node_load[nid] > max_value) { | |
2444 | max_value = khugepaged_node_load[nid]; | |
2445 | target_node = nid; | |
2446 | } | |
2447 | ||
2448 | /* do some balance if several nodes have the same hit record */ | |
2449 | if (target_node <= last_khugepaged_target_node) | |
2450 | for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES; | |
2451 | nid++) | |
2452 | if (max_value == khugepaged_node_load[nid]) { | |
2453 | target_node = nid; | |
2454 | break; | |
2455 | } | |
2456 | ||
2457 | last_khugepaged_target_node = target_node; | |
2458 | return target_node; | |
2459 | } | |
2460 | ||
26234f36 XG |
2461 | static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) |
2462 | { | |
2463 | if (IS_ERR(*hpage)) { | |
2464 | if (!*wait) | |
2465 | return false; | |
2466 | ||
2467 | *wait = false; | |
e3b4126c | 2468 | *hpage = NULL; |
26234f36 XG |
2469 | khugepaged_alloc_sleep(); |
2470 | } else if (*hpage) { | |
2471 | put_page(*hpage); | |
2472 | *hpage = NULL; | |
2473 | } | |
2474 | ||
2475 | return true; | |
2476 | } | |
2477 | ||
3b363692 MH |
2478 | static struct page * |
2479 | khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm, | |
d6669d68 | 2480 | unsigned long address, int node) |
26234f36 | 2481 | { |
309381fe | 2482 | VM_BUG_ON_PAGE(*hpage, *hpage); |
8b164568 | 2483 | |
ce83d217 | 2484 | /* |
8b164568 VB |
2485 | * Before allocating the hugepage, release the mmap_sem read lock. |
2486 | * The allocation can take potentially a long time if it involves | |
2487 | * sync compaction, and we do not need to hold the mmap_sem during | |
2488 | * that. We will recheck the vma after taking it again in write mode. | |
ce83d217 | 2489 | */ |
8b164568 VB |
2490 | up_read(&mm->mmap_sem); |
2491 | ||
96db800f | 2492 | *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER); |
26234f36 | 2493 | if (unlikely(!*hpage)) { |
81ab4201 | 2494 | count_vm_event(THP_COLLAPSE_ALLOC_FAILED); |
ce83d217 | 2495 | *hpage = ERR_PTR(-ENOMEM); |
26234f36 | 2496 | return NULL; |
ce83d217 | 2497 | } |
26234f36 | 2498 | |
65b3c07b | 2499 | count_vm_event(THP_COLLAPSE_ALLOC); |
26234f36 XG |
2500 | return *hpage; |
2501 | } | |
2502 | #else | |
9f1b868a BL |
2503 | static int khugepaged_find_target_node(void) |
2504 | { | |
2505 | return 0; | |
2506 | } | |
2507 | ||
10dc4155 BL |
2508 | static inline struct page *alloc_hugepage(int defrag) |
2509 | { | |
2510 | return alloc_pages(alloc_hugepage_gfpmask(defrag, 0), | |
2511 | HPAGE_PMD_ORDER); | |
2512 | } | |
2513 | ||
26234f36 XG |
2514 | static struct page *khugepaged_alloc_hugepage(bool *wait) |
2515 | { | |
2516 | struct page *hpage; | |
2517 | ||
2518 | do { | |
2519 | hpage = alloc_hugepage(khugepaged_defrag()); | |
2520 | if (!hpage) { | |
2521 | count_vm_event(THP_COLLAPSE_ALLOC_FAILED); | |
2522 | if (!*wait) | |
2523 | return NULL; | |
2524 | ||
2525 | *wait = false; | |
2526 | khugepaged_alloc_sleep(); | |
2527 | } else | |
2528 | count_vm_event(THP_COLLAPSE_ALLOC); | |
2529 | } while (unlikely(!hpage) && likely(khugepaged_enabled())); | |
2530 | ||
2531 | return hpage; | |
2532 | } | |
2533 | ||
2534 | static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) | |
2535 | { | |
2536 | if (!*hpage) | |
2537 | *hpage = khugepaged_alloc_hugepage(wait); | |
2538 | ||
2539 | if (unlikely(!*hpage)) | |
2540 | return false; | |
2541 | ||
2542 | return true; | |
2543 | } | |
2544 | ||
3b363692 MH |
2545 | static struct page * |
2546 | khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm, | |
d6669d68 | 2547 | unsigned long address, int node) |
26234f36 XG |
2548 | { |
2549 | up_read(&mm->mmap_sem); | |
2550 | VM_BUG_ON(!*hpage); | |
3b363692 | 2551 | |
26234f36 XG |
2552 | return *hpage; |
2553 | } | |
692e0b35 AA |
2554 | #endif |
2555 | ||
fa475e51 BL |
2556 | static bool hugepage_vma_check(struct vm_area_struct *vma) |
2557 | { | |
2558 | if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) || | |
2559 | (vma->vm_flags & VM_NOHUGEPAGE)) | |
2560 | return false; | |
7479df6d KS |
2561 | if (vma->vm_flags & VM_LOCKED) |
2562 | return false; | |
fa475e51 BL |
2563 | if (!vma->anon_vma || vma->vm_ops) |
2564 | return false; | |
2565 | if (is_vma_temporary_stack(vma)) | |
2566 | return false; | |
81d1b09c | 2567 | VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma); |
fa475e51 BL |
2568 | return true; |
2569 | } | |
2570 | ||
26234f36 XG |
2571 | static void collapse_huge_page(struct mm_struct *mm, |
2572 | unsigned long address, | |
2573 | struct page **hpage, | |
2574 | struct vm_area_struct *vma, | |
2575 | int node) | |
2576 | { | |
26234f36 XG |
2577 | pmd_t *pmd, _pmd; |
2578 | pte_t *pte; | |
2579 | pgtable_t pgtable; | |
2580 | struct page *new_page; | |
c4088ebd | 2581 | spinlock_t *pmd_ptl, *pte_ptl; |
7d2eba05 | 2582 | int isolated, result = 0; |
26234f36 | 2583 | unsigned long hstart, hend; |
00501b53 | 2584 | struct mem_cgroup *memcg; |
2ec74c3e SG |
2585 | unsigned long mmun_start; /* For mmu_notifiers */ |
2586 | unsigned long mmun_end; /* For mmu_notifiers */ | |
3b363692 | 2587 | gfp_t gfp; |
26234f36 XG |
2588 | |
2589 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
2590 | ||
3b363692 MH |
2591 | /* Only allocate from the target node */ |
2592 | gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) | | |
2593 | __GFP_THISNODE; | |
2594 | ||
26234f36 | 2595 | /* release the mmap_sem read lock. */ |
d6669d68 | 2596 | new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node); |
7d2eba05 EA |
2597 | if (!new_page) { |
2598 | result = SCAN_ALLOC_HUGE_PAGE_FAIL; | |
2599 | goto out_nolock; | |
2600 | } | |
26234f36 | 2601 | |
f627c2f5 | 2602 | if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) { |
7d2eba05 EA |
2603 | result = SCAN_CGROUP_CHARGE_FAIL; |
2604 | goto out_nolock; | |
2605 | } | |
ba76149f AA |
2606 | |
2607 | /* | |
2608 | * Prevent all access to pagetables with the exception of | |
2609 | * gup_fast later hanlded by the ptep_clear_flush and the VM | |
2610 | * handled by the anon_vma lock + PG_lock. | |
2611 | */ | |
2612 | down_write(&mm->mmap_sem); | |
7d2eba05 EA |
2613 | if (unlikely(khugepaged_test_exit(mm))) { |
2614 | result = SCAN_ANY_PROCESS; | |
ba76149f | 2615 | goto out; |
7d2eba05 | 2616 | } |
ba76149f AA |
2617 | |
2618 | vma = find_vma(mm, address); | |
7d2eba05 EA |
2619 | if (!vma) { |
2620 | result = SCAN_VMA_NULL; | |
a8f531eb | 2621 | goto out; |
7d2eba05 | 2622 | } |
ba76149f AA |
2623 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; |
2624 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
7d2eba05 EA |
2625 | if (address < hstart || address + HPAGE_PMD_SIZE > hend) { |
2626 | result = SCAN_ADDRESS_RANGE; | |
ba76149f | 2627 | goto out; |
7d2eba05 EA |
2628 | } |
2629 | if (!hugepage_vma_check(vma)) { | |
2630 | result = SCAN_VMA_CHECK; | |
a7d6e4ec | 2631 | goto out; |
7d2eba05 | 2632 | } |
6219049a | 2633 | pmd = mm_find_pmd(mm, address); |
7d2eba05 EA |
2634 | if (!pmd) { |
2635 | result = SCAN_PMD_NULL; | |
ba76149f | 2636 | goto out; |
7d2eba05 | 2637 | } |
ba76149f | 2638 | |
4fc3f1d6 | 2639 | anon_vma_lock_write(vma->anon_vma); |
ba76149f AA |
2640 | |
2641 | pte = pte_offset_map(pmd, address); | |
c4088ebd | 2642 | pte_ptl = pte_lockptr(mm, pmd); |
ba76149f | 2643 | |
2ec74c3e SG |
2644 | mmun_start = address; |
2645 | mmun_end = address + HPAGE_PMD_SIZE; | |
2646 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
c4088ebd | 2647 | pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ |
ba76149f AA |
2648 | /* |
2649 | * After this gup_fast can't run anymore. This also removes | |
2650 | * any huge TLB entry from the CPU so we won't allow | |
2651 | * huge and small TLB entries for the same virtual address | |
2652 | * to avoid the risk of CPU bugs in that area. | |
2653 | */ | |
15a25b2e | 2654 | _pmd = pmdp_collapse_flush(vma, address, pmd); |
c4088ebd | 2655 | spin_unlock(pmd_ptl); |
2ec74c3e | 2656 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
ba76149f | 2657 | |
c4088ebd | 2658 | spin_lock(pte_ptl); |
ba76149f | 2659 | isolated = __collapse_huge_page_isolate(vma, address, pte); |
c4088ebd | 2660 | spin_unlock(pte_ptl); |
ba76149f AA |
2661 | |
2662 | if (unlikely(!isolated)) { | |
453c7192 | 2663 | pte_unmap(pte); |
c4088ebd | 2664 | spin_lock(pmd_ptl); |
ba76149f | 2665 | BUG_ON(!pmd_none(*pmd)); |
7c342512 AK |
2666 | /* |
2667 | * We can only use set_pmd_at when establishing | |
2668 | * hugepmds and never for establishing regular pmds that | |
2669 | * points to regular pagetables. Use pmd_populate for that | |
2670 | */ | |
2671 | pmd_populate(mm, pmd, pmd_pgtable(_pmd)); | |
c4088ebd | 2672 | spin_unlock(pmd_ptl); |
08b52706 | 2673 | anon_vma_unlock_write(vma->anon_vma); |
7d2eba05 | 2674 | result = SCAN_FAIL; |
ce83d217 | 2675 | goto out; |
ba76149f AA |
2676 | } |
2677 | ||
2678 | /* | |
2679 | * All pages are isolated and locked so anon_vma rmap | |
2680 | * can't run anymore. | |
2681 | */ | |
08b52706 | 2682 | anon_vma_unlock_write(vma->anon_vma); |
ba76149f | 2683 | |
c4088ebd | 2684 | __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl); |
453c7192 | 2685 | pte_unmap(pte); |
ba76149f AA |
2686 | __SetPageUptodate(new_page); |
2687 | pgtable = pmd_pgtable(_pmd); | |
ba76149f | 2688 | |
3122359a KS |
2689 | _pmd = mk_huge_pmd(new_page, vma->vm_page_prot); |
2690 | _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); | |
ba76149f AA |
2691 | |
2692 | /* | |
2693 | * spin_lock() below is not the equivalent of smp_wmb(), so | |
2694 | * this is needed to avoid the copy_huge_page writes to become | |
2695 | * visible after the set_pmd_at() write. | |
2696 | */ | |
2697 | smp_wmb(); | |
2698 | ||
c4088ebd | 2699 | spin_lock(pmd_ptl); |
ba76149f | 2700 | BUG_ON(!pmd_none(*pmd)); |
d281ee61 | 2701 | page_add_new_anon_rmap(new_page, vma, address, true); |
f627c2f5 | 2702 | mem_cgroup_commit_charge(new_page, memcg, false, true); |
00501b53 | 2703 | lru_cache_add_active_or_unevictable(new_page, vma); |
fce144b4 | 2704 | pgtable_trans_huge_deposit(mm, pmd, pgtable); |
ba76149f | 2705 | set_pmd_at(mm, address, pmd, _pmd); |
b113da65 | 2706 | update_mmu_cache_pmd(vma, address, pmd); |
c4088ebd | 2707 | spin_unlock(pmd_ptl); |
ba76149f AA |
2708 | |
2709 | *hpage = NULL; | |
420256ef | 2710 | |
ba76149f | 2711 | khugepaged_pages_collapsed++; |
7d2eba05 | 2712 | result = SCAN_SUCCEED; |
ce83d217 | 2713 | out_up_write: |
ba76149f | 2714 | up_write(&mm->mmap_sem); |
7d2eba05 | 2715 | trace_mm_collapse_huge_page(mm, isolated, result); |
0bbbc0b3 AA |
2716 | return; |
2717 | ||
7d2eba05 EA |
2718 | out_nolock: |
2719 | trace_mm_collapse_huge_page(mm, isolated, result); | |
2720 | return; | |
ce83d217 | 2721 | out: |
f627c2f5 | 2722 | mem_cgroup_cancel_charge(new_page, memcg, true); |
ce83d217 | 2723 | goto out_up_write; |
ba76149f AA |
2724 | } |
2725 | ||
2726 | static int khugepaged_scan_pmd(struct mm_struct *mm, | |
2727 | struct vm_area_struct *vma, | |
2728 | unsigned long address, | |
2729 | struct page **hpage) | |
2730 | { | |
ba76149f AA |
2731 | pmd_t *pmd; |
2732 | pte_t *pte, *_pte; | |
7d2eba05 EA |
2733 | int ret = 0, none_or_zero = 0, result = 0; |
2734 | struct page *page = NULL; | |
ba76149f AA |
2735 | unsigned long _address; |
2736 | spinlock_t *ptl; | |
00ef2d2f | 2737 | int node = NUMA_NO_NODE; |
10359213 | 2738 | bool writable = false, referenced = false; |
ba76149f AA |
2739 | |
2740 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
2741 | ||
6219049a | 2742 | pmd = mm_find_pmd(mm, address); |
7d2eba05 EA |
2743 | if (!pmd) { |
2744 | result = SCAN_PMD_NULL; | |
ba76149f | 2745 | goto out; |
7d2eba05 | 2746 | } |
ba76149f | 2747 | |
9f1b868a | 2748 | memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load)); |
ba76149f AA |
2749 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); |
2750 | for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR; | |
2751 | _pte++, _address += PAGE_SIZE) { | |
2752 | pte_t pteval = *_pte; | |
ca0984ca | 2753 | if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { |
c1294d05 | 2754 | if (!userfaultfd_armed(vma) && |
7d2eba05 | 2755 | ++none_or_zero <= khugepaged_max_ptes_none) { |
ba76149f | 2756 | continue; |
7d2eba05 EA |
2757 | } else { |
2758 | result = SCAN_EXCEED_NONE_PTE; | |
ba76149f | 2759 | goto out_unmap; |
7d2eba05 | 2760 | } |
ba76149f | 2761 | } |
7d2eba05 EA |
2762 | if (!pte_present(pteval)) { |
2763 | result = SCAN_PTE_NON_PRESENT; | |
ba76149f | 2764 | goto out_unmap; |
7d2eba05 | 2765 | } |
10359213 EA |
2766 | if (pte_write(pteval)) |
2767 | writable = true; | |
2768 | ||
ba76149f | 2769 | page = vm_normal_page(vma, _address, pteval); |
7d2eba05 EA |
2770 | if (unlikely(!page)) { |
2771 | result = SCAN_PAGE_NULL; | |
ba76149f | 2772 | goto out_unmap; |
7d2eba05 | 2773 | } |
b1caa957 KS |
2774 | |
2775 | /* TODO: teach khugepaged to collapse THP mapped with pte */ | |
2776 | if (PageCompound(page)) { | |
2777 | result = SCAN_PAGE_COMPOUND; | |
2778 | goto out_unmap; | |
2779 | } | |
2780 | ||
5c4b4be3 | 2781 | /* |
9f1b868a BL |
2782 | * Record which node the original page is from and save this |
2783 | * information to khugepaged_node_load[]. | |
2784 | * Khupaged will allocate hugepage from the node has the max | |
2785 | * hit record. | |
5c4b4be3 | 2786 | */ |
9f1b868a | 2787 | node = page_to_nid(page); |
7d2eba05 EA |
2788 | if (khugepaged_scan_abort(node)) { |
2789 | result = SCAN_SCAN_ABORT; | |
14a4e214 | 2790 | goto out_unmap; |
7d2eba05 | 2791 | } |
9f1b868a | 2792 | khugepaged_node_load[node]++; |
7d2eba05 EA |
2793 | if (!PageLRU(page)) { |
2794 | result = SCAN_SCAN_ABORT; | |
2795 | goto out_unmap; | |
2796 | } | |
2797 | if (PageLocked(page)) { | |
2798 | result = SCAN_PAGE_LOCK; | |
ba76149f | 2799 | goto out_unmap; |
7d2eba05 EA |
2800 | } |
2801 | if (!PageAnon(page)) { | |
2802 | result = SCAN_PAGE_ANON; | |
2803 | goto out_unmap; | |
2804 | } | |
2805 | ||
10359213 EA |
2806 | /* |
2807 | * cannot use mapcount: can't collapse if there's a gup pin. | |
2808 | * The page must only be referenced by the scanned process | |
2809 | * and page swap cache. | |
2810 | */ | |
7d2eba05 EA |
2811 | if (page_count(page) != 1 + !!PageSwapCache(page)) { |
2812 | result = SCAN_PAGE_COUNT; | |
ba76149f | 2813 | goto out_unmap; |
7d2eba05 | 2814 | } |
33c3fc71 VD |
2815 | if (pte_young(pteval) || |
2816 | page_is_young(page) || PageReferenced(page) || | |
8ee53820 | 2817 | mmu_notifier_test_young(vma->vm_mm, address)) |
10359213 | 2818 | referenced = true; |
ba76149f | 2819 | } |
7d2eba05 EA |
2820 | if (writable) { |
2821 | if (referenced) { | |
2822 | result = SCAN_SUCCEED; | |
2823 | ret = 1; | |
2824 | } else { | |
2825 | result = SCAN_NO_REFERENCED_PAGE; | |
2826 | } | |
2827 | } else { | |
2828 | result = SCAN_PAGE_RO; | |
2829 | } | |
ba76149f AA |
2830 | out_unmap: |
2831 | pte_unmap_unlock(pte, ptl); | |
9f1b868a BL |
2832 | if (ret) { |
2833 | node = khugepaged_find_target_node(); | |
ce83d217 | 2834 | /* collapse_huge_page will return with the mmap_sem released */ |
5c4b4be3 | 2835 | collapse_huge_page(mm, address, hpage, vma, node); |
9f1b868a | 2836 | } |
ba76149f | 2837 | out: |
7d2eba05 EA |
2838 | trace_mm_khugepaged_scan_pmd(mm, page_to_pfn(page), writable, referenced, |
2839 | none_or_zero, result); | |
ba76149f AA |
2840 | return ret; |
2841 | } | |
2842 | ||
2843 | static void collect_mm_slot(struct mm_slot *mm_slot) | |
2844 | { | |
2845 | struct mm_struct *mm = mm_slot->mm; | |
2846 | ||
b9980cdc | 2847 | VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); |
ba76149f AA |
2848 | |
2849 | if (khugepaged_test_exit(mm)) { | |
2850 | /* free mm_slot */ | |
43b5fbbd | 2851 | hash_del(&mm_slot->hash); |
ba76149f AA |
2852 | list_del(&mm_slot->mm_node); |
2853 | ||
2854 | /* | |
2855 | * Not strictly needed because the mm exited already. | |
2856 | * | |
2857 | * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); | |
2858 | */ | |
2859 | ||
2860 | /* khugepaged_mm_lock actually not necessary for the below */ | |
2861 | free_mm_slot(mm_slot); | |
2862 | mmdrop(mm); | |
2863 | } | |
2864 | } | |
2865 | ||
2866 | static unsigned int khugepaged_scan_mm_slot(unsigned int pages, | |
2867 | struct page **hpage) | |
2f1da642 HS |
2868 | __releases(&khugepaged_mm_lock) |
2869 | __acquires(&khugepaged_mm_lock) | |
ba76149f AA |
2870 | { |
2871 | struct mm_slot *mm_slot; | |
2872 | struct mm_struct *mm; | |
2873 | struct vm_area_struct *vma; | |
2874 | int progress = 0; | |
2875 | ||
2876 | VM_BUG_ON(!pages); | |
b9980cdc | 2877 | VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); |
ba76149f AA |
2878 | |
2879 | if (khugepaged_scan.mm_slot) | |
2880 | mm_slot = khugepaged_scan.mm_slot; | |
2881 | else { | |
2882 | mm_slot = list_entry(khugepaged_scan.mm_head.next, | |
2883 | struct mm_slot, mm_node); | |
2884 | khugepaged_scan.address = 0; | |
2885 | khugepaged_scan.mm_slot = mm_slot; | |
2886 | } | |
2887 | spin_unlock(&khugepaged_mm_lock); | |
2888 | ||
2889 | mm = mm_slot->mm; | |
2890 | down_read(&mm->mmap_sem); | |
2891 | if (unlikely(khugepaged_test_exit(mm))) | |
2892 | vma = NULL; | |
2893 | else | |
2894 | vma = find_vma(mm, khugepaged_scan.address); | |
2895 | ||
2896 | progress++; | |
2897 | for (; vma; vma = vma->vm_next) { | |
2898 | unsigned long hstart, hend; | |
2899 | ||
2900 | cond_resched(); | |
2901 | if (unlikely(khugepaged_test_exit(mm))) { | |
2902 | progress++; | |
2903 | break; | |
2904 | } | |
fa475e51 BL |
2905 | if (!hugepage_vma_check(vma)) { |
2906 | skip: | |
ba76149f AA |
2907 | progress++; |
2908 | continue; | |
2909 | } | |
ba76149f AA |
2910 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; |
2911 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
a7d6e4ec AA |
2912 | if (hstart >= hend) |
2913 | goto skip; | |
2914 | if (khugepaged_scan.address > hend) | |
2915 | goto skip; | |
ba76149f AA |
2916 | if (khugepaged_scan.address < hstart) |
2917 | khugepaged_scan.address = hstart; | |
a7d6e4ec | 2918 | VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); |
ba76149f AA |
2919 | |
2920 | while (khugepaged_scan.address < hend) { | |
2921 | int ret; | |
2922 | cond_resched(); | |
2923 | if (unlikely(khugepaged_test_exit(mm))) | |
2924 | goto breakouterloop; | |
2925 | ||
2926 | VM_BUG_ON(khugepaged_scan.address < hstart || | |
2927 | khugepaged_scan.address + HPAGE_PMD_SIZE > | |
2928 | hend); | |
2929 | ret = khugepaged_scan_pmd(mm, vma, | |
2930 | khugepaged_scan.address, | |
2931 | hpage); | |
2932 | /* move to next address */ | |
2933 | khugepaged_scan.address += HPAGE_PMD_SIZE; | |
2934 | progress += HPAGE_PMD_NR; | |
2935 | if (ret) | |
2936 | /* we released mmap_sem so break loop */ | |
2937 | goto breakouterloop_mmap_sem; | |
2938 | if (progress >= pages) | |
2939 | goto breakouterloop; | |
2940 | } | |
2941 | } | |
2942 | breakouterloop: | |
2943 | up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */ | |
2944 | breakouterloop_mmap_sem: | |
2945 | ||
2946 | spin_lock(&khugepaged_mm_lock); | |
a7d6e4ec | 2947 | VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); |
ba76149f AA |
2948 | /* |
2949 | * Release the current mm_slot if this mm is about to die, or | |
2950 | * if we scanned all vmas of this mm. | |
2951 | */ | |
2952 | if (khugepaged_test_exit(mm) || !vma) { | |
2953 | /* | |
2954 | * Make sure that if mm_users is reaching zero while | |
2955 | * khugepaged runs here, khugepaged_exit will find | |
2956 | * mm_slot not pointing to the exiting mm. | |
2957 | */ | |
2958 | if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) { | |
2959 | khugepaged_scan.mm_slot = list_entry( | |
2960 | mm_slot->mm_node.next, | |
2961 | struct mm_slot, mm_node); | |
2962 | khugepaged_scan.address = 0; | |
2963 | } else { | |
2964 | khugepaged_scan.mm_slot = NULL; | |
2965 | khugepaged_full_scans++; | |
2966 | } | |
2967 | ||
2968 | collect_mm_slot(mm_slot); | |
2969 | } | |
2970 | ||
2971 | return progress; | |
2972 | } | |
2973 | ||
2974 | static int khugepaged_has_work(void) | |
2975 | { | |
2976 | return !list_empty(&khugepaged_scan.mm_head) && | |
2977 | khugepaged_enabled(); | |
2978 | } | |
2979 | ||
2980 | static int khugepaged_wait_event(void) | |
2981 | { | |
2982 | return !list_empty(&khugepaged_scan.mm_head) || | |
2017c0bf | 2983 | kthread_should_stop(); |
ba76149f AA |
2984 | } |
2985 | ||
d516904b | 2986 | static void khugepaged_do_scan(void) |
ba76149f | 2987 | { |
d516904b | 2988 | struct page *hpage = NULL; |
ba76149f AA |
2989 | unsigned int progress = 0, pass_through_head = 0; |
2990 | unsigned int pages = khugepaged_pages_to_scan; | |
d516904b | 2991 | bool wait = true; |
ba76149f AA |
2992 | |
2993 | barrier(); /* write khugepaged_pages_to_scan to local stack */ | |
2994 | ||
2995 | while (progress < pages) { | |
26234f36 | 2996 | if (!khugepaged_prealloc_page(&hpage, &wait)) |
d516904b | 2997 | break; |
26234f36 | 2998 | |
420256ef | 2999 | cond_resched(); |
ba76149f | 3000 | |
cd092411 | 3001 | if (unlikely(kthread_should_stop() || try_to_freeze())) |
878aee7d AA |
3002 | break; |
3003 | ||
ba76149f AA |
3004 | spin_lock(&khugepaged_mm_lock); |
3005 | if (!khugepaged_scan.mm_slot) | |
3006 | pass_through_head++; | |
3007 | if (khugepaged_has_work() && | |
3008 | pass_through_head < 2) | |
3009 | progress += khugepaged_scan_mm_slot(pages - progress, | |
d516904b | 3010 | &hpage); |
ba76149f AA |
3011 | else |
3012 | progress = pages; | |
3013 | spin_unlock(&khugepaged_mm_lock); | |
3014 | } | |
ba76149f | 3015 | |
d516904b XG |
3016 | if (!IS_ERR_OR_NULL(hpage)) |
3017 | put_page(hpage); | |
0bbbc0b3 AA |
3018 | } |
3019 | ||
2017c0bf XG |
3020 | static void khugepaged_wait_work(void) |
3021 | { | |
2017c0bf XG |
3022 | if (khugepaged_has_work()) { |
3023 | if (!khugepaged_scan_sleep_millisecs) | |
3024 | return; | |
3025 | ||
3026 | wait_event_freezable_timeout(khugepaged_wait, | |
3027 | kthread_should_stop(), | |
3028 | msecs_to_jiffies(khugepaged_scan_sleep_millisecs)); | |
3029 | return; | |
3030 | } | |
3031 | ||
3032 | if (khugepaged_enabled()) | |
3033 | wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); | |
3034 | } | |
3035 | ||
ba76149f AA |
3036 | static int khugepaged(void *none) |
3037 | { | |
3038 | struct mm_slot *mm_slot; | |
3039 | ||
878aee7d | 3040 | set_freezable(); |
8698a745 | 3041 | set_user_nice(current, MAX_NICE); |
ba76149f | 3042 | |
b7231789 XG |
3043 | while (!kthread_should_stop()) { |
3044 | khugepaged_do_scan(); | |
3045 | khugepaged_wait_work(); | |
3046 | } | |
ba76149f AA |
3047 | |
3048 | spin_lock(&khugepaged_mm_lock); | |
3049 | mm_slot = khugepaged_scan.mm_slot; | |
3050 | khugepaged_scan.mm_slot = NULL; | |
3051 | if (mm_slot) | |
3052 | collect_mm_slot(mm_slot); | |
3053 | spin_unlock(&khugepaged_mm_lock); | |
ba76149f AA |
3054 | return 0; |
3055 | } | |
3056 | ||
c5a647d0 KS |
3057 | static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, |
3058 | unsigned long haddr, pmd_t *pmd) | |
3059 | { | |
3060 | struct mm_struct *mm = vma->vm_mm; | |
3061 | pgtable_t pgtable; | |
3062 | pmd_t _pmd; | |
3063 | int i; | |
3064 | ||
8809aa2d | 3065 | pmdp_huge_clear_flush_notify(vma, haddr, pmd); |
c5a647d0 KS |
3066 | /* leave pmd empty until pte is filled */ |
3067 | ||
6b0b50b0 | 3068 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); |
c5a647d0 KS |
3069 | pmd_populate(mm, &_pmd, pgtable); |
3070 | ||
3071 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
3072 | pte_t *pte, entry; | |
3073 | entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); | |
3074 | entry = pte_mkspecial(entry); | |
3075 | pte = pte_offset_map(&_pmd, haddr); | |
3076 | VM_BUG_ON(!pte_none(*pte)); | |
3077 | set_pte_at(mm, haddr, pte, entry); | |
3078 | pte_unmap(pte); | |
3079 | } | |
3080 | smp_wmb(); /* make pte visible before pmd */ | |
3081 | pmd_populate(mm, pmd, pgtable); | |
97ae1749 | 3082 | put_huge_zero_page(); |
c5a647d0 KS |
3083 | } |
3084 | ||
e180377f KS |
3085 | void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address, |
3086 | pmd_t *pmd) | |
71e3aac0 | 3087 | { |
c4088ebd | 3088 | spinlock_t *ptl; |
4897c765 | 3089 | struct page *page = NULL; |
e180377f | 3090 | struct mm_struct *mm = vma->vm_mm; |
c5a647d0 KS |
3091 | unsigned long haddr = address & HPAGE_PMD_MASK; |
3092 | unsigned long mmun_start; /* For mmu_notifiers */ | |
3093 | unsigned long mmun_end; /* For mmu_notifiers */ | |
e180377f KS |
3094 | |
3095 | BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE); | |
71e3aac0 | 3096 | |
c5a647d0 KS |
3097 | mmun_start = haddr; |
3098 | mmun_end = haddr + HPAGE_PMD_SIZE; | |
750e8165 | 3099 | again: |
c5a647d0 | 3100 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
c4088ebd | 3101 | ptl = pmd_lock(mm, pmd); |
4897c765 MW |
3102 | if (unlikely(!pmd_trans_huge(*pmd))) |
3103 | goto unlock; | |
3104 | if (vma_is_dax(vma)) { | |
5b701b84 KS |
3105 | pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); |
3106 | if (is_huge_zero_pmd(_pmd)) | |
3107 | put_huge_zero_page(); | |
4897c765 | 3108 | } else if (is_huge_zero_pmd(*pmd)) { |
c5a647d0 | 3109 | __split_huge_zero_page_pmd(vma, haddr, pmd); |
4897c765 MW |
3110 | } else { |
3111 | page = pmd_page(*pmd); | |
3112 | VM_BUG_ON_PAGE(!page_count(page), page); | |
3113 | get_page(page); | |
71e3aac0 | 3114 | } |
4897c765 | 3115 | unlock: |
c4088ebd | 3116 | spin_unlock(ptl); |
c5a647d0 | 3117 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
71e3aac0 | 3118 | |
4897c765 MW |
3119 | if (!page) |
3120 | return; | |
71e3aac0 | 3121 | |
4897c765 | 3122 | split_huge_page(page); |
71e3aac0 | 3123 | put_page(page); |
750e8165 HD |
3124 | |
3125 | /* | |
3126 | * We don't always have down_write of mmap_sem here: a racing | |
3127 | * do_huge_pmd_wp_page() might have copied-on-write to another | |
3128 | * huge page before our split_huge_page() got the anon_vma lock. | |
3129 | */ | |
3130 | if (unlikely(pmd_trans_huge(*pmd))) | |
3131 | goto again; | |
71e3aac0 | 3132 | } |
94fcc585 | 3133 | |
78ddc534 | 3134 | static void split_huge_pmd_address(struct vm_area_struct *vma, |
94fcc585 AA |
3135 | unsigned long address) |
3136 | { | |
f72e7dcd HD |
3137 | pgd_t *pgd; |
3138 | pud_t *pud; | |
94fcc585 AA |
3139 | pmd_t *pmd; |
3140 | ||
3141 | VM_BUG_ON(!(address & ~HPAGE_PMD_MASK)); | |
3142 | ||
78ddc534 | 3143 | pgd = pgd_offset(vma->vm_mm, address); |
f72e7dcd HD |
3144 | if (!pgd_present(*pgd)) |
3145 | return; | |
3146 | ||
3147 | pud = pud_offset(pgd, address); | |
3148 | if (!pud_present(*pud)) | |
3149 | return; | |
3150 | ||
3151 | pmd = pmd_offset(pud, address); | |
78ddc534 | 3152 | if (!pmd_present(*pmd) || !pmd_trans_huge(*pmd)) |
94fcc585 AA |
3153 | return; |
3154 | /* | |
3155 | * Caller holds the mmap_sem write mode, so a huge pmd cannot | |
3156 | * materialize from under us. | |
3157 | */ | |
78ddc534 | 3158 | __split_huge_page_pmd(vma, address, pmd); |
94fcc585 AA |
3159 | } |
3160 | ||
e1b9996b | 3161 | void vma_adjust_trans_huge(struct vm_area_struct *vma, |
94fcc585 AA |
3162 | unsigned long start, |
3163 | unsigned long end, | |
3164 | long adjust_next) | |
3165 | { | |
3166 | /* | |
3167 | * If the new start address isn't hpage aligned and it could | |
3168 | * previously contain an hugepage: check if we need to split | |
3169 | * an huge pmd. | |
3170 | */ | |
3171 | if (start & ~HPAGE_PMD_MASK && | |
3172 | (start & HPAGE_PMD_MASK) >= vma->vm_start && | |
3173 | (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
78ddc534 | 3174 | split_huge_pmd_address(vma, start); |
94fcc585 AA |
3175 | |
3176 | /* | |
3177 | * If the new end address isn't hpage aligned and it could | |
3178 | * previously contain an hugepage: check if we need to split | |
3179 | * an huge pmd. | |
3180 | */ | |
3181 | if (end & ~HPAGE_PMD_MASK && | |
3182 | (end & HPAGE_PMD_MASK) >= vma->vm_start && | |
3183 | (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
78ddc534 | 3184 | split_huge_pmd_address(vma, end); |
94fcc585 AA |
3185 | |
3186 | /* | |
3187 | * If we're also updating the vma->vm_next->vm_start, if the new | |
3188 | * vm_next->vm_start isn't page aligned and it could previously | |
3189 | * contain an hugepage: check if we need to split an huge pmd. | |
3190 | */ | |
3191 | if (adjust_next > 0) { | |
3192 | struct vm_area_struct *next = vma->vm_next; | |
3193 | unsigned long nstart = next->vm_start; | |
3194 | nstart += adjust_next << PAGE_SHIFT; | |
3195 | if (nstart & ~HPAGE_PMD_MASK && | |
3196 | (nstart & HPAGE_PMD_MASK) >= next->vm_start && | |
3197 | (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) | |
78ddc534 | 3198 | split_huge_pmd_address(next, nstart); |
94fcc585 AA |
3199 | } |
3200 | } |