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, | |
48 | SCAN_ANY_PROCESS, | |
49 | SCAN_VMA_NULL, | |
50 | SCAN_VMA_CHECK, | |
51 | SCAN_ADDRESS_RANGE, | |
52 | SCAN_SWAP_CACHE_PAGE, | |
53 | SCAN_DEL_PAGE_LRU, | |
54 | SCAN_ALLOC_HUGE_PAGE_FAIL, | |
55 | SCAN_CGROUP_CHARGE_FAIL | |
56 | }; | |
57 | ||
58 | #define CREATE_TRACE_POINTS | |
59 | #include <trace/events/huge_memory.h> | |
60 | ||
ba76149f | 61 | /* |
8bfa3f9a JW |
62 | * By default transparent hugepage support is disabled in order that avoid |
63 | * to risk increase the memory footprint of applications without a guaranteed | |
64 | * benefit. When transparent hugepage support is enabled, is for all mappings, | |
65 | * and khugepaged scans all mappings. | |
66 | * Defrag is invoked by khugepaged hugepage allocations and by page faults | |
67 | * for all hugepage allocations. | |
ba76149f | 68 | */ |
71e3aac0 | 69 | unsigned long transparent_hugepage_flags __read_mostly = |
13ece886 | 70 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS |
ba76149f | 71 | (1<<TRANSPARENT_HUGEPAGE_FLAG)| |
13ece886 AA |
72 | #endif |
73 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE | |
74 | (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| | |
75 | #endif | |
d39d33c3 | 76 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)| |
79da5407 KS |
77 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| |
78 | (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
ba76149f AA |
79 | |
80 | /* default scan 8*512 pte (or vmas) every 30 second */ | |
81 | static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8; | |
82 | static unsigned int khugepaged_pages_collapsed; | |
83 | static unsigned int khugepaged_full_scans; | |
84 | static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; | |
85 | /* during fragmentation poll the hugepage allocator once every minute */ | |
86 | static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; | |
87 | static struct task_struct *khugepaged_thread __read_mostly; | |
88 | static DEFINE_MUTEX(khugepaged_mutex); | |
89 | static DEFINE_SPINLOCK(khugepaged_mm_lock); | |
90 | static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); | |
91 | /* | |
92 | * default collapse hugepages if there is at least one pte mapped like | |
93 | * it would have happened if the vma was large enough during page | |
94 | * fault. | |
95 | */ | |
96 | static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1; | |
97 | ||
98 | static int khugepaged(void *none); | |
ba76149f | 99 | static int khugepaged_slab_init(void); |
65ebb64f | 100 | static void khugepaged_slab_exit(void); |
ba76149f | 101 | |
43b5fbbd SL |
102 | #define MM_SLOTS_HASH_BITS 10 |
103 | static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); | |
104 | ||
ba76149f AA |
105 | static struct kmem_cache *mm_slot_cache __read_mostly; |
106 | ||
107 | /** | |
108 | * struct mm_slot - hash lookup from mm to mm_slot | |
109 | * @hash: hash collision list | |
110 | * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head | |
111 | * @mm: the mm that this information is valid for | |
112 | */ | |
113 | struct mm_slot { | |
114 | struct hlist_node hash; | |
115 | struct list_head mm_node; | |
116 | struct mm_struct *mm; | |
117 | }; | |
118 | ||
119 | /** | |
120 | * struct khugepaged_scan - cursor for scanning | |
121 | * @mm_head: the head of the mm list to scan | |
122 | * @mm_slot: the current mm_slot we are scanning | |
123 | * @address: the next address inside that to be scanned | |
124 | * | |
125 | * There is only the one khugepaged_scan instance of this cursor structure. | |
126 | */ | |
127 | struct khugepaged_scan { | |
128 | struct list_head mm_head; | |
129 | struct mm_slot *mm_slot; | |
130 | unsigned long address; | |
2f1da642 HS |
131 | }; |
132 | static struct khugepaged_scan khugepaged_scan = { | |
ba76149f AA |
133 | .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), |
134 | }; | |
135 | ||
f000565a | 136 | |
2c0b80d4 | 137 | static void set_recommended_min_free_kbytes(void) |
f000565a AA |
138 | { |
139 | struct zone *zone; | |
140 | int nr_zones = 0; | |
141 | unsigned long recommended_min; | |
f000565a | 142 | |
f000565a AA |
143 | for_each_populated_zone(zone) |
144 | nr_zones++; | |
145 | ||
974a786e | 146 | /* Ensure 2 pageblocks are free to assist fragmentation avoidance */ |
f000565a AA |
147 | recommended_min = pageblock_nr_pages * nr_zones * 2; |
148 | ||
149 | /* | |
150 | * Make sure that on average at least two pageblocks are almost free | |
151 | * of another type, one for a migratetype to fall back to and a | |
152 | * second to avoid subsequent fallbacks of other types There are 3 | |
153 | * MIGRATE_TYPES we care about. | |
154 | */ | |
155 | recommended_min += pageblock_nr_pages * nr_zones * | |
156 | MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; | |
157 | ||
158 | /* don't ever allow to reserve more than 5% of the lowmem */ | |
159 | recommended_min = min(recommended_min, | |
160 | (unsigned long) nr_free_buffer_pages() / 20); | |
161 | recommended_min <<= (PAGE_SHIFT-10); | |
162 | ||
42aa83cb HP |
163 | if (recommended_min > min_free_kbytes) { |
164 | if (user_min_free_kbytes >= 0) | |
165 | pr_info("raising min_free_kbytes from %d to %lu " | |
166 | "to help transparent hugepage allocations\n", | |
167 | min_free_kbytes, recommended_min); | |
168 | ||
f000565a | 169 | min_free_kbytes = recommended_min; |
42aa83cb | 170 | } |
f000565a | 171 | setup_per_zone_wmarks(); |
f000565a | 172 | } |
f000565a | 173 | |
79553da2 | 174 | static int start_stop_khugepaged(void) |
ba76149f AA |
175 | { |
176 | int err = 0; | |
177 | if (khugepaged_enabled()) { | |
ba76149f AA |
178 | if (!khugepaged_thread) |
179 | khugepaged_thread = kthread_run(khugepaged, NULL, | |
180 | "khugepaged"); | |
18e8e5c7 | 181 | if (IS_ERR(khugepaged_thread)) { |
ae3a8c1c | 182 | pr_err("khugepaged: kthread_run(khugepaged) failed\n"); |
ba76149f AA |
183 | err = PTR_ERR(khugepaged_thread); |
184 | khugepaged_thread = NULL; | |
79553da2 | 185 | goto fail; |
ba76149f | 186 | } |
911891af XG |
187 | |
188 | if (!list_empty(&khugepaged_scan.mm_head)) | |
ba76149f | 189 | wake_up_interruptible(&khugepaged_wait); |
f000565a AA |
190 | |
191 | set_recommended_min_free_kbytes(); | |
911891af | 192 | } else if (khugepaged_thread) { |
911891af XG |
193 | kthread_stop(khugepaged_thread); |
194 | khugepaged_thread = NULL; | |
195 | } | |
79553da2 | 196 | fail: |
ba76149f AA |
197 | return err; |
198 | } | |
71e3aac0 | 199 | |
97ae1749 | 200 | static atomic_t huge_zero_refcount; |
56873f43 | 201 | struct page *huge_zero_page __read_mostly; |
4a6c1297 | 202 | |
fc437044 | 203 | struct page *get_huge_zero_page(void) |
97ae1749 KS |
204 | { |
205 | struct page *zero_page; | |
206 | retry: | |
207 | if (likely(atomic_inc_not_zero(&huge_zero_refcount))) | |
4db0c3c2 | 208 | return READ_ONCE(huge_zero_page); |
97ae1749 KS |
209 | |
210 | zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, | |
4a6c1297 | 211 | HPAGE_PMD_ORDER); |
d8a8e1f0 KS |
212 | if (!zero_page) { |
213 | count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); | |
5918d10a | 214 | return NULL; |
d8a8e1f0 KS |
215 | } |
216 | count_vm_event(THP_ZERO_PAGE_ALLOC); | |
97ae1749 | 217 | preempt_disable(); |
5918d10a | 218 | if (cmpxchg(&huge_zero_page, NULL, zero_page)) { |
97ae1749 | 219 | preempt_enable(); |
5ddacbe9 | 220 | __free_pages(zero_page, compound_order(zero_page)); |
97ae1749 KS |
221 | goto retry; |
222 | } | |
223 | ||
224 | /* We take additional reference here. It will be put back by shrinker */ | |
225 | atomic_set(&huge_zero_refcount, 2); | |
226 | preempt_enable(); | |
4db0c3c2 | 227 | return READ_ONCE(huge_zero_page); |
4a6c1297 KS |
228 | } |
229 | ||
97ae1749 | 230 | static void put_huge_zero_page(void) |
4a6c1297 | 231 | { |
97ae1749 KS |
232 | /* |
233 | * Counter should never go to zero here. Only shrinker can put | |
234 | * last reference. | |
235 | */ | |
236 | BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); | |
4a6c1297 KS |
237 | } |
238 | ||
48896466 GC |
239 | static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, |
240 | struct shrink_control *sc) | |
4a6c1297 | 241 | { |
48896466 GC |
242 | /* we can free zero page only if last reference remains */ |
243 | return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; | |
244 | } | |
97ae1749 | 245 | |
48896466 GC |
246 | static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, |
247 | struct shrink_control *sc) | |
248 | { | |
97ae1749 | 249 | if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { |
5918d10a KS |
250 | struct page *zero_page = xchg(&huge_zero_page, NULL); |
251 | BUG_ON(zero_page == NULL); | |
5ddacbe9 | 252 | __free_pages(zero_page, compound_order(zero_page)); |
48896466 | 253 | return HPAGE_PMD_NR; |
97ae1749 KS |
254 | } |
255 | ||
256 | return 0; | |
4a6c1297 KS |
257 | } |
258 | ||
97ae1749 | 259 | static struct shrinker huge_zero_page_shrinker = { |
48896466 GC |
260 | .count_objects = shrink_huge_zero_page_count, |
261 | .scan_objects = shrink_huge_zero_page_scan, | |
97ae1749 KS |
262 | .seeks = DEFAULT_SEEKS, |
263 | }; | |
264 | ||
71e3aac0 | 265 | #ifdef CONFIG_SYSFS |
ba76149f | 266 | |
71e3aac0 AA |
267 | static ssize_t double_flag_show(struct kobject *kobj, |
268 | struct kobj_attribute *attr, char *buf, | |
269 | enum transparent_hugepage_flag enabled, | |
270 | enum transparent_hugepage_flag req_madv) | |
271 | { | |
272 | if (test_bit(enabled, &transparent_hugepage_flags)) { | |
273 | VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags)); | |
274 | return sprintf(buf, "[always] madvise never\n"); | |
275 | } else if (test_bit(req_madv, &transparent_hugepage_flags)) | |
276 | return sprintf(buf, "always [madvise] never\n"); | |
277 | else | |
278 | return sprintf(buf, "always madvise [never]\n"); | |
279 | } | |
280 | static ssize_t double_flag_store(struct kobject *kobj, | |
281 | struct kobj_attribute *attr, | |
282 | const char *buf, size_t count, | |
283 | enum transparent_hugepage_flag enabled, | |
284 | enum transparent_hugepage_flag req_madv) | |
285 | { | |
286 | if (!memcmp("always", buf, | |
287 | min(sizeof("always")-1, count))) { | |
288 | set_bit(enabled, &transparent_hugepage_flags); | |
289 | clear_bit(req_madv, &transparent_hugepage_flags); | |
290 | } else if (!memcmp("madvise", buf, | |
291 | min(sizeof("madvise")-1, count))) { | |
292 | clear_bit(enabled, &transparent_hugepage_flags); | |
293 | set_bit(req_madv, &transparent_hugepage_flags); | |
294 | } else if (!memcmp("never", buf, | |
295 | min(sizeof("never")-1, count))) { | |
296 | clear_bit(enabled, &transparent_hugepage_flags); | |
297 | clear_bit(req_madv, &transparent_hugepage_flags); | |
298 | } else | |
299 | return -EINVAL; | |
300 | ||
301 | return count; | |
302 | } | |
303 | ||
304 | static ssize_t enabled_show(struct kobject *kobj, | |
305 | struct kobj_attribute *attr, char *buf) | |
306 | { | |
307 | return double_flag_show(kobj, attr, buf, | |
308 | TRANSPARENT_HUGEPAGE_FLAG, | |
309 | TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); | |
310 | } | |
311 | static ssize_t enabled_store(struct kobject *kobj, | |
312 | struct kobj_attribute *attr, | |
313 | const char *buf, size_t count) | |
314 | { | |
ba76149f AA |
315 | ssize_t ret; |
316 | ||
317 | ret = double_flag_store(kobj, attr, buf, count, | |
318 | TRANSPARENT_HUGEPAGE_FLAG, | |
319 | TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); | |
320 | ||
321 | if (ret > 0) { | |
911891af XG |
322 | int err; |
323 | ||
324 | mutex_lock(&khugepaged_mutex); | |
79553da2 | 325 | err = start_stop_khugepaged(); |
911891af XG |
326 | mutex_unlock(&khugepaged_mutex); |
327 | ||
ba76149f AA |
328 | if (err) |
329 | ret = err; | |
330 | } | |
331 | ||
332 | return ret; | |
71e3aac0 AA |
333 | } |
334 | static struct kobj_attribute enabled_attr = | |
335 | __ATTR(enabled, 0644, enabled_show, enabled_store); | |
336 | ||
337 | static ssize_t single_flag_show(struct kobject *kobj, | |
338 | struct kobj_attribute *attr, char *buf, | |
339 | enum transparent_hugepage_flag flag) | |
340 | { | |
e27e6151 BH |
341 | return sprintf(buf, "%d\n", |
342 | !!test_bit(flag, &transparent_hugepage_flags)); | |
71e3aac0 | 343 | } |
e27e6151 | 344 | |
71e3aac0 AA |
345 | static ssize_t single_flag_store(struct kobject *kobj, |
346 | struct kobj_attribute *attr, | |
347 | const char *buf, size_t count, | |
348 | enum transparent_hugepage_flag flag) | |
349 | { | |
e27e6151 BH |
350 | unsigned long value; |
351 | int ret; | |
352 | ||
353 | ret = kstrtoul(buf, 10, &value); | |
354 | if (ret < 0) | |
355 | return ret; | |
356 | if (value > 1) | |
357 | return -EINVAL; | |
358 | ||
359 | if (value) | |
71e3aac0 | 360 | set_bit(flag, &transparent_hugepage_flags); |
e27e6151 | 361 | else |
71e3aac0 | 362 | clear_bit(flag, &transparent_hugepage_flags); |
71e3aac0 AA |
363 | |
364 | return count; | |
365 | } | |
366 | ||
367 | /* | |
368 | * Currently defrag only disables __GFP_NOWAIT for allocation. A blind | |
369 | * __GFP_REPEAT is too aggressive, it's never worth swapping tons of | |
370 | * memory just to allocate one more hugepage. | |
371 | */ | |
372 | static ssize_t defrag_show(struct kobject *kobj, | |
373 | struct kobj_attribute *attr, char *buf) | |
374 | { | |
375 | return double_flag_show(kobj, attr, buf, | |
376 | TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, | |
377 | TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); | |
378 | } | |
379 | static ssize_t defrag_store(struct kobject *kobj, | |
380 | struct kobj_attribute *attr, | |
381 | const char *buf, size_t count) | |
382 | { | |
383 | return double_flag_store(kobj, attr, buf, count, | |
384 | TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, | |
385 | TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); | |
386 | } | |
387 | static struct kobj_attribute defrag_attr = | |
388 | __ATTR(defrag, 0644, defrag_show, defrag_store); | |
389 | ||
79da5407 KS |
390 | static ssize_t use_zero_page_show(struct kobject *kobj, |
391 | struct kobj_attribute *attr, char *buf) | |
392 | { | |
393 | return single_flag_show(kobj, attr, buf, | |
394 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
395 | } | |
396 | static ssize_t use_zero_page_store(struct kobject *kobj, | |
397 | struct kobj_attribute *attr, const char *buf, size_t count) | |
398 | { | |
399 | return single_flag_store(kobj, attr, buf, count, | |
400 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
401 | } | |
402 | static struct kobj_attribute use_zero_page_attr = | |
403 | __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); | |
71e3aac0 AA |
404 | #ifdef CONFIG_DEBUG_VM |
405 | static ssize_t debug_cow_show(struct kobject *kobj, | |
406 | struct kobj_attribute *attr, char *buf) | |
407 | { | |
408 | return single_flag_show(kobj, attr, buf, | |
409 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); | |
410 | } | |
411 | static ssize_t debug_cow_store(struct kobject *kobj, | |
412 | struct kobj_attribute *attr, | |
413 | const char *buf, size_t count) | |
414 | { | |
415 | return single_flag_store(kobj, attr, buf, count, | |
416 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); | |
417 | } | |
418 | static struct kobj_attribute debug_cow_attr = | |
419 | __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); | |
420 | #endif /* CONFIG_DEBUG_VM */ | |
421 | ||
422 | static struct attribute *hugepage_attr[] = { | |
423 | &enabled_attr.attr, | |
424 | &defrag_attr.attr, | |
79da5407 | 425 | &use_zero_page_attr.attr, |
71e3aac0 AA |
426 | #ifdef CONFIG_DEBUG_VM |
427 | &debug_cow_attr.attr, | |
428 | #endif | |
429 | NULL, | |
430 | }; | |
431 | ||
432 | static struct attribute_group hugepage_attr_group = { | |
433 | .attrs = hugepage_attr, | |
ba76149f AA |
434 | }; |
435 | ||
436 | static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, | |
437 | struct kobj_attribute *attr, | |
438 | char *buf) | |
439 | { | |
440 | return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs); | |
441 | } | |
442 | ||
443 | static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, | |
444 | struct kobj_attribute *attr, | |
445 | const char *buf, size_t count) | |
446 | { | |
447 | unsigned long msecs; | |
448 | int err; | |
449 | ||
3dbb95f7 | 450 | err = kstrtoul(buf, 10, &msecs); |
ba76149f AA |
451 | if (err || msecs > UINT_MAX) |
452 | return -EINVAL; | |
453 | ||
454 | khugepaged_scan_sleep_millisecs = msecs; | |
455 | wake_up_interruptible(&khugepaged_wait); | |
456 | ||
457 | return count; | |
458 | } | |
459 | static struct kobj_attribute scan_sleep_millisecs_attr = | |
460 | __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show, | |
461 | scan_sleep_millisecs_store); | |
462 | ||
463 | static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, | |
464 | struct kobj_attribute *attr, | |
465 | char *buf) | |
466 | { | |
467 | return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs); | |
468 | } | |
469 | ||
470 | static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, | |
471 | struct kobj_attribute *attr, | |
472 | const char *buf, size_t count) | |
473 | { | |
474 | unsigned long msecs; | |
475 | int err; | |
476 | ||
3dbb95f7 | 477 | err = kstrtoul(buf, 10, &msecs); |
ba76149f AA |
478 | if (err || msecs > UINT_MAX) |
479 | return -EINVAL; | |
480 | ||
481 | khugepaged_alloc_sleep_millisecs = msecs; | |
482 | wake_up_interruptible(&khugepaged_wait); | |
483 | ||
484 | return count; | |
485 | } | |
486 | static struct kobj_attribute alloc_sleep_millisecs_attr = | |
487 | __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show, | |
488 | alloc_sleep_millisecs_store); | |
489 | ||
490 | static ssize_t pages_to_scan_show(struct kobject *kobj, | |
491 | struct kobj_attribute *attr, | |
492 | char *buf) | |
493 | { | |
494 | return sprintf(buf, "%u\n", khugepaged_pages_to_scan); | |
495 | } | |
496 | static ssize_t pages_to_scan_store(struct kobject *kobj, | |
497 | struct kobj_attribute *attr, | |
498 | const char *buf, size_t count) | |
499 | { | |
500 | int err; | |
501 | unsigned long pages; | |
502 | ||
3dbb95f7 | 503 | err = kstrtoul(buf, 10, &pages); |
ba76149f AA |
504 | if (err || !pages || pages > UINT_MAX) |
505 | return -EINVAL; | |
506 | ||
507 | khugepaged_pages_to_scan = pages; | |
508 | ||
509 | return count; | |
510 | } | |
511 | static struct kobj_attribute pages_to_scan_attr = | |
512 | __ATTR(pages_to_scan, 0644, pages_to_scan_show, | |
513 | pages_to_scan_store); | |
514 | ||
515 | static ssize_t pages_collapsed_show(struct kobject *kobj, | |
516 | struct kobj_attribute *attr, | |
517 | char *buf) | |
518 | { | |
519 | return sprintf(buf, "%u\n", khugepaged_pages_collapsed); | |
520 | } | |
521 | static struct kobj_attribute pages_collapsed_attr = | |
522 | __ATTR_RO(pages_collapsed); | |
523 | ||
524 | static ssize_t full_scans_show(struct kobject *kobj, | |
525 | struct kobj_attribute *attr, | |
526 | char *buf) | |
527 | { | |
528 | return sprintf(buf, "%u\n", khugepaged_full_scans); | |
529 | } | |
530 | static struct kobj_attribute full_scans_attr = | |
531 | __ATTR_RO(full_scans); | |
532 | ||
533 | static ssize_t khugepaged_defrag_show(struct kobject *kobj, | |
534 | struct kobj_attribute *attr, char *buf) | |
535 | { | |
536 | return single_flag_show(kobj, attr, buf, | |
537 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); | |
538 | } | |
539 | static ssize_t khugepaged_defrag_store(struct kobject *kobj, | |
540 | struct kobj_attribute *attr, | |
541 | const char *buf, size_t count) | |
542 | { | |
543 | return single_flag_store(kobj, attr, buf, count, | |
544 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); | |
545 | } | |
546 | static struct kobj_attribute khugepaged_defrag_attr = | |
547 | __ATTR(defrag, 0644, khugepaged_defrag_show, | |
548 | khugepaged_defrag_store); | |
549 | ||
550 | /* | |
551 | * max_ptes_none controls if khugepaged should collapse hugepages over | |
552 | * any unmapped ptes in turn potentially increasing the memory | |
553 | * footprint of the vmas. When max_ptes_none is 0 khugepaged will not | |
554 | * reduce the available free memory in the system as it | |
555 | * runs. Increasing max_ptes_none will instead potentially reduce the | |
556 | * free memory in the system during the khugepaged scan. | |
557 | */ | |
558 | static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj, | |
559 | struct kobj_attribute *attr, | |
560 | char *buf) | |
561 | { | |
562 | return sprintf(buf, "%u\n", khugepaged_max_ptes_none); | |
563 | } | |
564 | static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj, | |
565 | struct kobj_attribute *attr, | |
566 | const char *buf, size_t count) | |
567 | { | |
568 | int err; | |
569 | unsigned long max_ptes_none; | |
570 | ||
3dbb95f7 | 571 | err = kstrtoul(buf, 10, &max_ptes_none); |
ba76149f AA |
572 | if (err || max_ptes_none > HPAGE_PMD_NR-1) |
573 | return -EINVAL; | |
574 | ||
575 | khugepaged_max_ptes_none = max_ptes_none; | |
576 | ||
577 | return count; | |
578 | } | |
579 | static struct kobj_attribute khugepaged_max_ptes_none_attr = | |
580 | __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show, | |
581 | khugepaged_max_ptes_none_store); | |
582 | ||
583 | static struct attribute *khugepaged_attr[] = { | |
584 | &khugepaged_defrag_attr.attr, | |
585 | &khugepaged_max_ptes_none_attr.attr, | |
586 | &pages_to_scan_attr.attr, | |
587 | &pages_collapsed_attr.attr, | |
588 | &full_scans_attr.attr, | |
589 | &scan_sleep_millisecs_attr.attr, | |
590 | &alloc_sleep_millisecs_attr.attr, | |
591 | NULL, | |
592 | }; | |
593 | ||
594 | static struct attribute_group khugepaged_attr_group = { | |
595 | .attrs = khugepaged_attr, | |
596 | .name = "khugepaged", | |
71e3aac0 | 597 | }; |
71e3aac0 | 598 | |
569e5590 | 599 | static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) |
71e3aac0 | 600 | { |
71e3aac0 AA |
601 | int err; |
602 | ||
569e5590 SL |
603 | *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); |
604 | if (unlikely(!*hugepage_kobj)) { | |
ae3a8c1c | 605 | pr_err("failed to create transparent hugepage kobject\n"); |
569e5590 | 606 | return -ENOMEM; |
ba76149f AA |
607 | } |
608 | ||
569e5590 | 609 | err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); |
ba76149f | 610 | if (err) { |
ae3a8c1c | 611 | pr_err("failed to register transparent hugepage group\n"); |
569e5590 | 612 | goto delete_obj; |
ba76149f AA |
613 | } |
614 | ||
569e5590 | 615 | err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); |
ba76149f | 616 | if (err) { |
ae3a8c1c | 617 | pr_err("failed to register transparent hugepage group\n"); |
569e5590 | 618 | goto remove_hp_group; |
ba76149f | 619 | } |
569e5590 SL |
620 | |
621 | return 0; | |
622 | ||
623 | remove_hp_group: | |
624 | sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); | |
625 | delete_obj: | |
626 | kobject_put(*hugepage_kobj); | |
627 | return err; | |
628 | } | |
629 | ||
630 | static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
631 | { | |
632 | sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); | |
633 | sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); | |
634 | kobject_put(hugepage_kobj); | |
635 | } | |
636 | #else | |
637 | static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) | |
638 | { | |
639 | return 0; | |
640 | } | |
641 | ||
642 | static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
643 | { | |
644 | } | |
645 | #endif /* CONFIG_SYSFS */ | |
646 | ||
647 | static int __init hugepage_init(void) | |
648 | { | |
649 | int err; | |
650 | struct kobject *hugepage_kobj; | |
651 | ||
652 | if (!has_transparent_hugepage()) { | |
653 | transparent_hugepage_flags = 0; | |
654 | return -EINVAL; | |
655 | } | |
656 | ||
657 | err = hugepage_init_sysfs(&hugepage_kobj); | |
658 | if (err) | |
65ebb64f | 659 | goto err_sysfs; |
ba76149f AA |
660 | |
661 | err = khugepaged_slab_init(); | |
662 | if (err) | |
65ebb64f | 663 | goto err_slab; |
ba76149f | 664 | |
65ebb64f KS |
665 | err = register_shrinker(&huge_zero_page_shrinker); |
666 | if (err) | |
667 | goto err_hzp_shrinker; | |
97ae1749 | 668 | |
97562cd2 RR |
669 | /* |
670 | * By default disable transparent hugepages on smaller systems, | |
671 | * where the extra memory used could hurt more than TLB overhead | |
672 | * is likely to save. The admin can still enable it through /sys. | |
673 | */ | |
79553da2 | 674 | if (totalram_pages < (512 << (20 - PAGE_SHIFT))) { |
97562cd2 | 675 | transparent_hugepage_flags = 0; |
79553da2 KS |
676 | return 0; |
677 | } | |
97562cd2 | 678 | |
79553da2 | 679 | err = start_stop_khugepaged(); |
65ebb64f KS |
680 | if (err) |
681 | goto err_khugepaged; | |
ba76149f | 682 | |
569e5590 | 683 | return 0; |
65ebb64f KS |
684 | err_khugepaged: |
685 | unregister_shrinker(&huge_zero_page_shrinker); | |
686 | err_hzp_shrinker: | |
687 | khugepaged_slab_exit(); | |
688 | err_slab: | |
569e5590 | 689 | hugepage_exit_sysfs(hugepage_kobj); |
65ebb64f | 690 | err_sysfs: |
ba76149f | 691 | return err; |
71e3aac0 | 692 | } |
a64fb3cd | 693 | subsys_initcall(hugepage_init); |
71e3aac0 AA |
694 | |
695 | static int __init setup_transparent_hugepage(char *str) | |
696 | { | |
697 | int ret = 0; | |
698 | if (!str) | |
699 | goto out; | |
700 | if (!strcmp(str, "always")) { | |
701 | set_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
702 | &transparent_hugepage_flags); | |
703 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
704 | &transparent_hugepage_flags); | |
705 | ret = 1; | |
706 | } else if (!strcmp(str, "madvise")) { | |
707 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
708 | &transparent_hugepage_flags); | |
709 | set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
710 | &transparent_hugepage_flags); | |
711 | ret = 1; | |
712 | } else if (!strcmp(str, "never")) { | |
713 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
714 | &transparent_hugepage_flags); | |
715 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
716 | &transparent_hugepage_flags); | |
717 | ret = 1; | |
718 | } | |
719 | out: | |
720 | if (!ret) | |
ae3a8c1c | 721 | pr_warn("transparent_hugepage= cannot parse, ignored\n"); |
71e3aac0 AA |
722 | return ret; |
723 | } | |
724 | __setup("transparent_hugepage=", setup_transparent_hugepage); | |
725 | ||
b32967ff | 726 | pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) |
71e3aac0 AA |
727 | { |
728 | if (likely(vma->vm_flags & VM_WRITE)) | |
729 | pmd = pmd_mkwrite(pmd); | |
730 | return pmd; | |
731 | } | |
732 | ||
3122359a | 733 | static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot) |
b3092b3b BL |
734 | { |
735 | pmd_t entry; | |
3122359a | 736 | entry = mk_pmd(page, prot); |
b3092b3b BL |
737 | entry = pmd_mkhuge(entry); |
738 | return entry; | |
739 | } | |
740 | ||
71e3aac0 AA |
741 | static int __do_huge_pmd_anonymous_page(struct mm_struct *mm, |
742 | struct vm_area_struct *vma, | |
230c92a8 | 743 | unsigned long address, pmd_t *pmd, |
6b251fc9 AA |
744 | struct page *page, gfp_t gfp, |
745 | unsigned int flags) | |
71e3aac0 | 746 | { |
00501b53 | 747 | struct mem_cgroup *memcg; |
71e3aac0 | 748 | pgtable_t pgtable; |
c4088ebd | 749 | spinlock_t *ptl; |
230c92a8 | 750 | unsigned long haddr = address & HPAGE_PMD_MASK; |
71e3aac0 | 751 | |
309381fe | 752 | VM_BUG_ON_PAGE(!PageCompound(page), page); |
00501b53 | 753 | |
f627c2f5 | 754 | if (mem_cgroup_try_charge(page, mm, gfp, &memcg, true)) { |
6b251fc9 AA |
755 | put_page(page); |
756 | count_vm_event(THP_FAULT_FALLBACK); | |
757 | return VM_FAULT_FALLBACK; | |
758 | } | |
00501b53 | 759 | |
71e3aac0 | 760 | pgtable = pte_alloc_one(mm, haddr); |
00501b53 | 761 | if (unlikely(!pgtable)) { |
f627c2f5 | 762 | mem_cgroup_cancel_charge(page, memcg, true); |
6b251fc9 | 763 | put_page(page); |
71e3aac0 | 764 | return VM_FAULT_OOM; |
00501b53 | 765 | } |
71e3aac0 AA |
766 | |
767 | clear_huge_page(page, haddr, HPAGE_PMD_NR); | |
52f37629 MK |
768 | /* |
769 | * The memory barrier inside __SetPageUptodate makes sure that | |
770 | * clear_huge_page writes become visible before the set_pmd_at() | |
771 | * write. | |
772 | */ | |
71e3aac0 AA |
773 | __SetPageUptodate(page); |
774 | ||
c4088ebd | 775 | ptl = pmd_lock(mm, pmd); |
71e3aac0 | 776 | if (unlikely(!pmd_none(*pmd))) { |
c4088ebd | 777 | spin_unlock(ptl); |
f627c2f5 | 778 | mem_cgroup_cancel_charge(page, memcg, true); |
71e3aac0 AA |
779 | put_page(page); |
780 | pte_free(mm, pgtable); | |
781 | } else { | |
782 | pmd_t entry; | |
6b251fc9 AA |
783 | |
784 | /* Deliver the page fault to userland */ | |
785 | if (userfaultfd_missing(vma)) { | |
786 | int ret; | |
787 | ||
788 | spin_unlock(ptl); | |
f627c2f5 | 789 | mem_cgroup_cancel_charge(page, memcg, true); |
6b251fc9 AA |
790 | put_page(page); |
791 | pte_free(mm, pgtable); | |
230c92a8 | 792 | ret = handle_userfault(vma, address, flags, |
6b251fc9 AA |
793 | VM_UFFD_MISSING); |
794 | VM_BUG_ON(ret & VM_FAULT_FALLBACK); | |
795 | return ret; | |
796 | } | |
797 | ||
3122359a KS |
798 | entry = mk_huge_pmd(page, vma->vm_page_prot); |
799 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
d281ee61 | 800 | page_add_new_anon_rmap(page, vma, haddr, true); |
f627c2f5 | 801 | mem_cgroup_commit_charge(page, memcg, false, true); |
00501b53 | 802 | lru_cache_add_active_or_unevictable(page, vma); |
6b0b50b0 | 803 | pgtable_trans_huge_deposit(mm, pmd, pgtable); |
71e3aac0 | 804 | set_pmd_at(mm, haddr, pmd, entry); |
71e3aac0 | 805 | add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); |
e1f56c89 | 806 | atomic_long_inc(&mm->nr_ptes); |
c4088ebd | 807 | spin_unlock(ptl); |
6b251fc9 | 808 | count_vm_event(THP_FAULT_ALLOC); |
71e3aac0 AA |
809 | } |
810 | ||
aa2e878e | 811 | return 0; |
71e3aac0 AA |
812 | } |
813 | ||
cc5d462f | 814 | static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp) |
0bbbc0b3 | 815 | { |
71baba4b | 816 | return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_RECLAIM)) | extra_gfp; |
0bbbc0b3 AA |
817 | } |
818 | ||
c4088ebd | 819 | /* Caller must hold page table lock. */ |
d295e341 | 820 | static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, |
97ae1749 | 821 | struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, |
5918d10a | 822 | struct page *zero_page) |
fc9fe822 KS |
823 | { |
824 | pmd_t entry; | |
7c414164 AM |
825 | if (!pmd_none(*pmd)) |
826 | return false; | |
5918d10a | 827 | entry = mk_pmd(zero_page, vma->vm_page_prot); |
fc9fe822 | 828 | entry = pmd_mkhuge(entry); |
6b0b50b0 | 829 | pgtable_trans_huge_deposit(mm, pmd, pgtable); |
fc9fe822 | 830 | set_pmd_at(mm, haddr, pmd, entry); |
e1f56c89 | 831 | atomic_long_inc(&mm->nr_ptes); |
7c414164 | 832 | return true; |
fc9fe822 KS |
833 | } |
834 | ||
71e3aac0 AA |
835 | int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma, |
836 | unsigned long address, pmd_t *pmd, | |
837 | unsigned int flags) | |
838 | { | |
077fcf11 | 839 | gfp_t gfp; |
71e3aac0 AA |
840 | struct page *page; |
841 | unsigned long haddr = address & HPAGE_PMD_MASK; | |
71e3aac0 | 842 | |
128ec037 | 843 | if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end) |
c0292554 | 844 | return VM_FAULT_FALLBACK; |
7479df6d KS |
845 | if (vma->vm_flags & VM_LOCKED) |
846 | return VM_FAULT_FALLBACK; | |
128ec037 KS |
847 | if (unlikely(anon_vma_prepare(vma))) |
848 | return VM_FAULT_OOM; | |
6d50e60c | 849 | if (unlikely(khugepaged_enter(vma, vma->vm_flags))) |
128ec037 | 850 | return VM_FAULT_OOM; |
593befa6 | 851 | if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) && |
128ec037 | 852 | transparent_hugepage_use_zero_page()) { |
c4088ebd | 853 | spinlock_t *ptl; |
128ec037 KS |
854 | pgtable_t pgtable; |
855 | struct page *zero_page; | |
856 | bool set; | |
6b251fc9 | 857 | int ret; |
128ec037 KS |
858 | pgtable = pte_alloc_one(mm, haddr); |
859 | if (unlikely(!pgtable)) | |
ba76149f | 860 | return VM_FAULT_OOM; |
128ec037 KS |
861 | zero_page = get_huge_zero_page(); |
862 | if (unlikely(!zero_page)) { | |
863 | pte_free(mm, pgtable); | |
81ab4201 | 864 | count_vm_event(THP_FAULT_FALLBACK); |
c0292554 | 865 | return VM_FAULT_FALLBACK; |
b9bbfbe3 | 866 | } |
c4088ebd | 867 | ptl = pmd_lock(mm, pmd); |
6b251fc9 AA |
868 | ret = 0; |
869 | set = false; | |
870 | if (pmd_none(*pmd)) { | |
871 | if (userfaultfd_missing(vma)) { | |
872 | spin_unlock(ptl); | |
230c92a8 | 873 | ret = handle_userfault(vma, address, flags, |
6b251fc9 AA |
874 | VM_UFFD_MISSING); |
875 | VM_BUG_ON(ret & VM_FAULT_FALLBACK); | |
876 | } else { | |
877 | set_huge_zero_page(pgtable, mm, vma, | |
878 | haddr, pmd, | |
879 | zero_page); | |
880 | spin_unlock(ptl); | |
881 | set = true; | |
882 | } | |
883 | } else | |
884 | spin_unlock(ptl); | |
128ec037 KS |
885 | if (!set) { |
886 | pte_free(mm, pgtable); | |
887 | put_huge_zero_page(); | |
edad9d2c | 888 | } |
6b251fc9 | 889 | return ret; |
71e3aac0 | 890 | } |
077fcf11 AK |
891 | gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0); |
892 | page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); | |
128ec037 KS |
893 | if (unlikely(!page)) { |
894 | count_vm_event(THP_FAULT_FALLBACK); | |
c0292554 | 895 | return VM_FAULT_FALLBACK; |
128ec037 | 896 | } |
230c92a8 AA |
897 | return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp, |
898 | flags); | |
71e3aac0 AA |
899 | } |
900 | ||
ae18d6dc | 901 | static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, |
5cad465d MW |
902 | pmd_t *pmd, unsigned long pfn, pgprot_t prot, bool write) |
903 | { | |
904 | struct mm_struct *mm = vma->vm_mm; | |
905 | pmd_t entry; | |
906 | spinlock_t *ptl; | |
907 | ||
908 | ptl = pmd_lock(mm, pmd); | |
909 | if (pmd_none(*pmd)) { | |
910 | entry = pmd_mkhuge(pfn_pmd(pfn, prot)); | |
911 | if (write) { | |
912 | entry = pmd_mkyoung(pmd_mkdirty(entry)); | |
913 | entry = maybe_pmd_mkwrite(entry, vma); | |
914 | } | |
915 | set_pmd_at(mm, addr, pmd, entry); | |
916 | update_mmu_cache_pmd(vma, addr, pmd); | |
917 | } | |
918 | spin_unlock(ptl); | |
5cad465d MW |
919 | } |
920 | ||
921 | int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, | |
922 | pmd_t *pmd, unsigned long pfn, bool write) | |
923 | { | |
924 | pgprot_t pgprot = vma->vm_page_prot; | |
925 | /* | |
926 | * If we had pmd_special, we could avoid all these restrictions, | |
927 | * but we need to be consistent with PTEs and architectures that | |
928 | * can't support a 'special' bit. | |
929 | */ | |
930 | BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); | |
931 | BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == | |
932 | (VM_PFNMAP|VM_MIXEDMAP)); | |
933 | BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); | |
934 | BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn)); | |
935 | ||
936 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
937 | return VM_FAULT_SIGBUS; | |
938 | if (track_pfn_insert(vma, &pgprot, pfn)) | |
939 | return VM_FAULT_SIGBUS; | |
ae18d6dc MW |
940 | insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write); |
941 | return VM_FAULT_NOPAGE; | |
5cad465d MW |
942 | } |
943 | ||
71e3aac0 AA |
944 | int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, |
945 | pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, | |
946 | struct vm_area_struct *vma) | |
947 | { | |
c4088ebd | 948 | spinlock_t *dst_ptl, *src_ptl; |
71e3aac0 AA |
949 | struct page *src_page; |
950 | pmd_t pmd; | |
951 | pgtable_t pgtable; | |
952 | int ret; | |
953 | ||
954 | ret = -ENOMEM; | |
955 | pgtable = pte_alloc_one(dst_mm, addr); | |
956 | if (unlikely(!pgtable)) | |
957 | goto out; | |
958 | ||
c4088ebd KS |
959 | dst_ptl = pmd_lock(dst_mm, dst_pmd); |
960 | src_ptl = pmd_lockptr(src_mm, src_pmd); | |
961 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
71e3aac0 AA |
962 | |
963 | ret = -EAGAIN; | |
964 | pmd = *src_pmd; | |
965 | if (unlikely(!pmd_trans_huge(pmd))) { | |
966 | pte_free(dst_mm, pgtable); | |
967 | goto out_unlock; | |
968 | } | |
fc9fe822 | 969 | /* |
c4088ebd | 970 | * When page table lock is held, the huge zero pmd should not be |
fc9fe822 KS |
971 | * under splitting since we don't split the page itself, only pmd to |
972 | * a page table. | |
973 | */ | |
974 | if (is_huge_zero_pmd(pmd)) { | |
5918d10a | 975 | struct page *zero_page; |
97ae1749 KS |
976 | /* |
977 | * get_huge_zero_page() will never allocate a new page here, | |
978 | * since we already have a zero page to copy. It just takes a | |
979 | * reference. | |
980 | */ | |
5918d10a | 981 | zero_page = get_huge_zero_page(); |
6b251fc9 | 982 | set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, |
5918d10a | 983 | zero_page); |
fc9fe822 KS |
984 | ret = 0; |
985 | goto out_unlock; | |
986 | } | |
de466bd6 | 987 | |
71e3aac0 AA |
988 | if (unlikely(pmd_trans_splitting(pmd))) { |
989 | /* split huge page running from under us */ | |
c4088ebd KS |
990 | spin_unlock(src_ptl); |
991 | spin_unlock(dst_ptl); | |
71e3aac0 AA |
992 | pte_free(dst_mm, pgtable); |
993 | ||
994 | wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */ | |
995 | goto out; | |
996 | } | |
997 | src_page = pmd_page(pmd); | |
309381fe | 998 | VM_BUG_ON_PAGE(!PageHead(src_page), src_page); |
71e3aac0 AA |
999 | get_page(src_page); |
1000 | page_dup_rmap(src_page); | |
1001 | add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); | |
1002 | ||
1003 | pmdp_set_wrprotect(src_mm, addr, src_pmd); | |
1004 | pmd = pmd_mkold(pmd_wrprotect(pmd)); | |
6b0b50b0 | 1005 | pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); |
71e3aac0 | 1006 | set_pmd_at(dst_mm, addr, dst_pmd, pmd); |
e1f56c89 | 1007 | atomic_long_inc(&dst_mm->nr_ptes); |
71e3aac0 AA |
1008 | |
1009 | ret = 0; | |
1010 | out_unlock: | |
c4088ebd KS |
1011 | spin_unlock(src_ptl); |
1012 | spin_unlock(dst_ptl); | |
71e3aac0 AA |
1013 | out: |
1014 | return ret; | |
1015 | } | |
1016 | ||
a1dd450b WD |
1017 | void huge_pmd_set_accessed(struct mm_struct *mm, |
1018 | struct vm_area_struct *vma, | |
1019 | unsigned long address, | |
1020 | pmd_t *pmd, pmd_t orig_pmd, | |
1021 | int dirty) | |
1022 | { | |
c4088ebd | 1023 | spinlock_t *ptl; |
a1dd450b WD |
1024 | pmd_t entry; |
1025 | unsigned long haddr; | |
1026 | ||
c4088ebd | 1027 | ptl = pmd_lock(mm, pmd); |
a1dd450b WD |
1028 | if (unlikely(!pmd_same(*pmd, orig_pmd))) |
1029 | goto unlock; | |
1030 | ||
1031 | entry = pmd_mkyoung(orig_pmd); | |
1032 | haddr = address & HPAGE_PMD_MASK; | |
1033 | if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty)) | |
1034 | update_mmu_cache_pmd(vma, address, pmd); | |
1035 | ||
1036 | unlock: | |
c4088ebd | 1037 | spin_unlock(ptl); |
a1dd450b WD |
1038 | } |
1039 | ||
5338a937 HD |
1040 | /* |
1041 | * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages | |
1042 | * during copy_user_huge_page()'s copy_page_rep(): in the case when | |
1043 | * the source page gets split and a tail freed before copy completes. | |
1044 | * Called under pmd_lock of checked pmd, so safe from splitting itself. | |
1045 | */ | |
1046 | static void get_user_huge_page(struct page *page) | |
1047 | { | |
1048 | if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) { | |
1049 | struct page *endpage = page + HPAGE_PMD_NR; | |
1050 | ||
1051 | atomic_add(HPAGE_PMD_NR, &page->_count); | |
1052 | while (++page < endpage) | |
1053 | get_huge_page_tail(page); | |
1054 | } else { | |
1055 | get_page(page); | |
1056 | } | |
1057 | } | |
1058 | ||
1059 | static void put_user_huge_page(struct page *page) | |
1060 | { | |
1061 | if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) { | |
1062 | struct page *endpage = page + HPAGE_PMD_NR; | |
1063 | ||
1064 | while (page < endpage) | |
1065 | put_page(page++); | |
1066 | } else { | |
1067 | put_page(page); | |
1068 | } | |
1069 | } | |
1070 | ||
71e3aac0 AA |
1071 | static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm, |
1072 | struct vm_area_struct *vma, | |
1073 | unsigned long address, | |
1074 | pmd_t *pmd, pmd_t orig_pmd, | |
1075 | struct page *page, | |
1076 | unsigned long haddr) | |
1077 | { | |
00501b53 | 1078 | struct mem_cgroup *memcg; |
c4088ebd | 1079 | spinlock_t *ptl; |
71e3aac0 AA |
1080 | pgtable_t pgtable; |
1081 | pmd_t _pmd; | |
1082 | int ret = 0, i; | |
1083 | struct page **pages; | |
2ec74c3e SG |
1084 | unsigned long mmun_start; /* For mmu_notifiers */ |
1085 | unsigned long mmun_end; /* For mmu_notifiers */ | |
71e3aac0 AA |
1086 | |
1087 | pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR, | |
1088 | GFP_KERNEL); | |
1089 | if (unlikely(!pages)) { | |
1090 | ret |= VM_FAULT_OOM; | |
1091 | goto out; | |
1092 | } | |
1093 | ||
1094 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
cc5d462f AK |
1095 | pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE | |
1096 | __GFP_OTHER_NODE, | |
19ee151e | 1097 | vma, address, page_to_nid(page)); |
b9bbfbe3 | 1098 | if (unlikely(!pages[i] || |
00501b53 | 1099 | mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL, |
f627c2f5 | 1100 | &memcg, false))) { |
b9bbfbe3 | 1101 | if (pages[i]) |
71e3aac0 | 1102 | put_page(pages[i]); |
b9bbfbe3 | 1103 | while (--i >= 0) { |
00501b53 JW |
1104 | memcg = (void *)page_private(pages[i]); |
1105 | set_page_private(pages[i], 0); | |
f627c2f5 KS |
1106 | mem_cgroup_cancel_charge(pages[i], memcg, |
1107 | false); | |
b9bbfbe3 AA |
1108 | put_page(pages[i]); |
1109 | } | |
71e3aac0 AA |
1110 | kfree(pages); |
1111 | ret |= VM_FAULT_OOM; | |
1112 | goto out; | |
1113 | } | |
00501b53 | 1114 | set_page_private(pages[i], (unsigned long)memcg); |
71e3aac0 AA |
1115 | } |
1116 | ||
1117 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
1118 | copy_user_highpage(pages[i], page + i, | |
0089e485 | 1119 | haddr + PAGE_SIZE * i, vma); |
71e3aac0 AA |
1120 | __SetPageUptodate(pages[i]); |
1121 | cond_resched(); | |
1122 | } | |
1123 | ||
2ec74c3e SG |
1124 | mmun_start = haddr; |
1125 | mmun_end = haddr + HPAGE_PMD_SIZE; | |
1126 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
1127 | ||
c4088ebd | 1128 | ptl = pmd_lock(mm, pmd); |
71e3aac0 AA |
1129 | if (unlikely(!pmd_same(*pmd, orig_pmd))) |
1130 | goto out_free_pages; | |
309381fe | 1131 | VM_BUG_ON_PAGE(!PageHead(page), page); |
71e3aac0 | 1132 | |
8809aa2d | 1133 | pmdp_huge_clear_flush_notify(vma, haddr, pmd); |
71e3aac0 AA |
1134 | /* leave pmd empty until pte is filled */ |
1135 | ||
6b0b50b0 | 1136 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); |
71e3aac0 AA |
1137 | pmd_populate(mm, &_pmd, pgtable); |
1138 | ||
1139 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
1140 | pte_t *pte, entry; | |
1141 | entry = mk_pte(pages[i], vma->vm_page_prot); | |
1142 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); | |
00501b53 JW |
1143 | memcg = (void *)page_private(pages[i]); |
1144 | set_page_private(pages[i], 0); | |
d281ee61 | 1145 | page_add_new_anon_rmap(pages[i], vma, haddr, false); |
f627c2f5 | 1146 | mem_cgroup_commit_charge(pages[i], memcg, false, false); |
00501b53 | 1147 | lru_cache_add_active_or_unevictable(pages[i], vma); |
71e3aac0 AA |
1148 | pte = pte_offset_map(&_pmd, haddr); |
1149 | VM_BUG_ON(!pte_none(*pte)); | |
1150 | set_pte_at(mm, haddr, pte, entry); | |
1151 | pte_unmap(pte); | |
1152 | } | |
1153 | kfree(pages); | |
1154 | ||
71e3aac0 AA |
1155 | smp_wmb(); /* make pte visible before pmd */ |
1156 | pmd_populate(mm, pmd, pgtable); | |
d281ee61 | 1157 | page_remove_rmap(page, true); |
c4088ebd | 1158 | spin_unlock(ptl); |
71e3aac0 | 1159 | |
2ec74c3e SG |
1160 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
1161 | ||
71e3aac0 AA |
1162 | ret |= VM_FAULT_WRITE; |
1163 | put_page(page); | |
1164 | ||
1165 | out: | |
1166 | return ret; | |
1167 | ||
1168 | out_free_pages: | |
c4088ebd | 1169 | spin_unlock(ptl); |
2ec74c3e | 1170 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
b9bbfbe3 | 1171 | for (i = 0; i < HPAGE_PMD_NR; i++) { |
00501b53 JW |
1172 | memcg = (void *)page_private(pages[i]); |
1173 | set_page_private(pages[i], 0); | |
f627c2f5 | 1174 | mem_cgroup_cancel_charge(pages[i], memcg, false); |
71e3aac0 | 1175 | put_page(pages[i]); |
b9bbfbe3 | 1176 | } |
71e3aac0 AA |
1177 | kfree(pages); |
1178 | goto out; | |
1179 | } | |
1180 | ||
1181 | int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma, | |
1182 | unsigned long address, pmd_t *pmd, pmd_t orig_pmd) | |
1183 | { | |
c4088ebd | 1184 | spinlock_t *ptl; |
71e3aac0 | 1185 | int ret = 0; |
93b4796d | 1186 | struct page *page = NULL, *new_page; |
00501b53 | 1187 | struct mem_cgroup *memcg; |
71e3aac0 | 1188 | unsigned long haddr; |
2ec74c3e SG |
1189 | unsigned long mmun_start; /* For mmu_notifiers */ |
1190 | unsigned long mmun_end; /* For mmu_notifiers */ | |
3b363692 | 1191 | gfp_t huge_gfp; /* for allocation and charge */ |
71e3aac0 | 1192 | |
c4088ebd | 1193 | ptl = pmd_lockptr(mm, pmd); |
81d1b09c | 1194 | VM_BUG_ON_VMA(!vma->anon_vma, vma); |
93b4796d KS |
1195 | haddr = address & HPAGE_PMD_MASK; |
1196 | if (is_huge_zero_pmd(orig_pmd)) | |
1197 | goto alloc; | |
c4088ebd | 1198 | spin_lock(ptl); |
71e3aac0 AA |
1199 | if (unlikely(!pmd_same(*pmd, orig_pmd))) |
1200 | goto out_unlock; | |
1201 | ||
1202 | page = pmd_page(orig_pmd); | |
309381fe | 1203 | VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); |
1f25fe20 KS |
1204 | /* |
1205 | * We can only reuse the page if nobody else maps the huge page or it's | |
1206 | * part. We can do it by checking page_mapcount() on each sub-page, but | |
1207 | * it's expensive. | |
1208 | * The cheaper way is to check page_count() to be equal 1: every | |
1209 | * mapcount takes page reference reference, so this way we can | |
1210 | * guarantee, that the PMD is the only mapping. | |
1211 | * This can give false negative if somebody pinned the page, but that's | |
1212 | * fine. | |
1213 | */ | |
1214 | if (page_mapcount(page) == 1 && page_count(page) == 1) { | |
71e3aac0 AA |
1215 | pmd_t entry; |
1216 | entry = pmd_mkyoung(orig_pmd); | |
1217 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
1218 | if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1)) | |
b113da65 | 1219 | update_mmu_cache_pmd(vma, address, pmd); |
71e3aac0 AA |
1220 | ret |= VM_FAULT_WRITE; |
1221 | goto out_unlock; | |
1222 | } | |
5338a937 | 1223 | get_user_huge_page(page); |
c4088ebd | 1224 | spin_unlock(ptl); |
93b4796d | 1225 | alloc: |
71e3aac0 | 1226 | if (transparent_hugepage_enabled(vma) && |
077fcf11 | 1227 | !transparent_hugepage_debug_cow()) { |
3b363692 MH |
1228 | huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0); |
1229 | new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER); | |
077fcf11 | 1230 | } else |
71e3aac0 AA |
1231 | new_page = NULL; |
1232 | ||
1233 | if (unlikely(!new_page)) { | |
eecc1e42 | 1234 | if (!page) { |
e9b71ca9 KS |
1235 | split_huge_page_pmd(vma, address, pmd); |
1236 | ret |= VM_FAULT_FALLBACK; | |
93b4796d KS |
1237 | } else { |
1238 | ret = do_huge_pmd_wp_page_fallback(mm, vma, address, | |
1239 | pmd, orig_pmd, page, haddr); | |
9845cbbd | 1240 | if (ret & VM_FAULT_OOM) { |
93b4796d | 1241 | split_huge_page(page); |
9845cbbd KS |
1242 | ret |= VM_FAULT_FALLBACK; |
1243 | } | |
5338a937 | 1244 | put_user_huge_page(page); |
93b4796d | 1245 | } |
17766dde | 1246 | count_vm_event(THP_FAULT_FALLBACK); |
71e3aac0 AA |
1247 | goto out; |
1248 | } | |
1249 | ||
f627c2f5 KS |
1250 | if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg, |
1251 | true))) { | |
b9bbfbe3 | 1252 | put_page(new_page); |
93b4796d KS |
1253 | if (page) { |
1254 | split_huge_page(page); | |
5338a937 | 1255 | put_user_huge_page(page); |
9845cbbd KS |
1256 | } else |
1257 | split_huge_page_pmd(vma, address, pmd); | |
1258 | ret |= VM_FAULT_FALLBACK; | |
17766dde | 1259 | count_vm_event(THP_FAULT_FALLBACK); |
b9bbfbe3 AA |
1260 | goto out; |
1261 | } | |
1262 | ||
17766dde DR |
1263 | count_vm_event(THP_FAULT_ALLOC); |
1264 | ||
eecc1e42 | 1265 | if (!page) |
93b4796d KS |
1266 | clear_huge_page(new_page, haddr, HPAGE_PMD_NR); |
1267 | else | |
1268 | copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR); | |
71e3aac0 AA |
1269 | __SetPageUptodate(new_page); |
1270 | ||
2ec74c3e SG |
1271 | mmun_start = haddr; |
1272 | mmun_end = haddr + HPAGE_PMD_SIZE; | |
1273 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
1274 | ||
c4088ebd | 1275 | spin_lock(ptl); |
93b4796d | 1276 | if (page) |
5338a937 | 1277 | put_user_huge_page(page); |
b9bbfbe3 | 1278 | if (unlikely(!pmd_same(*pmd, orig_pmd))) { |
c4088ebd | 1279 | spin_unlock(ptl); |
f627c2f5 | 1280 | mem_cgroup_cancel_charge(new_page, memcg, true); |
71e3aac0 | 1281 | put_page(new_page); |
2ec74c3e | 1282 | goto out_mn; |
b9bbfbe3 | 1283 | } else { |
71e3aac0 | 1284 | pmd_t entry; |
3122359a KS |
1285 | entry = mk_huge_pmd(new_page, vma->vm_page_prot); |
1286 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
8809aa2d | 1287 | pmdp_huge_clear_flush_notify(vma, haddr, pmd); |
d281ee61 | 1288 | page_add_new_anon_rmap(new_page, vma, haddr, true); |
f627c2f5 | 1289 | mem_cgroup_commit_charge(new_page, memcg, false, true); |
00501b53 | 1290 | lru_cache_add_active_or_unevictable(new_page, vma); |
71e3aac0 | 1291 | set_pmd_at(mm, haddr, pmd, entry); |
b113da65 | 1292 | update_mmu_cache_pmd(vma, address, pmd); |
eecc1e42 | 1293 | if (!page) { |
93b4796d | 1294 | add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); |
97ae1749 KS |
1295 | put_huge_zero_page(); |
1296 | } else { | |
309381fe | 1297 | VM_BUG_ON_PAGE(!PageHead(page), page); |
d281ee61 | 1298 | page_remove_rmap(page, true); |
93b4796d KS |
1299 | put_page(page); |
1300 | } | |
71e3aac0 AA |
1301 | ret |= VM_FAULT_WRITE; |
1302 | } | |
c4088ebd | 1303 | spin_unlock(ptl); |
2ec74c3e SG |
1304 | out_mn: |
1305 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); | |
71e3aac0 AA |
1306 | out: |
1307 | return ret; | |
2ec74c3e | 1308 | out_unlock: |
c4088ebd | 1309 | spin_unlock(ptl); |
2ec74c3e | 1310 | return ret; |
71e3aac0 AA |
1311 | } |
1312 | ||
b676b293 | 1313 | struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, |
71e3aac0 AA |
1314 | unsigned long addr, |
1315 | pmd_t *pmd, | |
1316 | unsigned int flags) | |
1317 | { | |
b676b293 | 1318 | struct mm_struct *mm = vma->vm_mm; |
71e3aac0 AA |
1319 | struct page *page = NULL; |
1320 | ||
c4088ebd | 1321 | assert_spin_locked(pmd_lockptr(mm, pmd)); |
71e3aac0 AA |
1322 | |
1323 | if (flags & FOLL_WRITE && !pmd_write(*pmd)) | |
1324 | goto out; | |
1325 | ||
85facf25 KS |
1326 | /* Avoid dumping huge zero page */ |
1327 | if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) | |
1328 | return ERR_PTR(-EFAULT); | |
1329 | ||
2b4847e7 | 1330 | /* Full NUMA hinting faults to serialise migration in fault paths */ |
8a0516ed | 1331 | if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) |
2b4847e7 MG |
1332 | goto out; |
1333 | ||
71e3aac0 | 1334 | page = pmd_page(*pmd); |
309381fe | 1335 | VM_BUG_ON_PAGE(!PageHead(page), page); |
71e3aac0 AA |
1336 | if (flags & FOLL_TOUCH) { |
1337 | pmd_t _pmd; | |
1338 | /* | |
1339 | * We should set the dirty bit only for FOLL_WRITE but | |
1340 | * for now the dirty bit in the pmd is meaningless. | |
1341 | * And if the dirty bit will become meaningful and | |
1342 | * we'll only set it with FOLL_WRITE, an atomic | |
1343 | * set_bit will be required on the pmd to set the | |
1344 | * young bit, instead of the current set_pmd_at. | |
1345 | */ | |
1346 | _pmd = pmd_mkyoung(pmd_mkdirty(*pmd)); | |
8663890a AK |
1347 | if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, |
1348 | pmd, _pmd, 1)) | |
1349 | update_mmu_cache_pmd(vma, addr, pmd); | |
71e3aac0 | 1350 | } |
de60f5f1 | 1351 | if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { |
b676b293 DR |
1352 | if (page->mapping && trylock_page(page)) { |
1353 | lru_add_drain(); | |
1354 | if (page->mapping) | |
1355 | mlock_vma_page(page); | |
1356 | unlock_page(page); | |
1357 | } | |
1358 | } | |
71e3aac0 | 1359 | page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; |
309381fe | 1360 | VM_BUG_ON_PAGE(!PageCompound(page), page); |
71e3aac0 | 1361 | if (flags & FOLL_GET) |
70b50f94 | 1362 | get_page_foll(page); |
71e3aac0 AA |
1363 | |
1364 | out: | |
1365 | return page; | |
1366 | } | |
1367 | ||
d10e63f2 | 1368 | /* NUMA hinting page fault entry point for trans huge pmds */ |
4daae3b4 MG |
1369 | int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma, |
1370 | unsigned long addr, pmd_t pmd, pmd_t *pmdp) | |
d10e63f2 | 1371 | { |
c4088ebd | 1372 | spinlock_t *ptl; |
b8916634 | 1373 | struct anon_vma *anon_vma = NULL; |
b32967ff | 1374 | struct page *page; |
d10e63f2 | 1375 | unsigned long haddr = addr & HPAGE_PMD_MASK; |
8191acbd | 1376 | int page_nid = -1, this_nid = numa_node_id(); |
90572890 | 1377 | int target_nid, last_cpupid = -1; |
8191acbd MG |
1378 | bool page_locked; |
1379 | bool migrated = false; | |
b191f9b1 | 1380 | bool was_writable; |
6688cc05 | 1381 | int flags = 0; |
d10e63f2 | 1382 | |
c0e7cad9 MG |
1383 | /* A PROT_NONE fault should not end up here */ |
1384 | BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))); | |
1385 | ||
c4088ebd | 1386 | ptl = pmd_lock(mm, pmdp); |
d10e63f2 MG |
1387 | if (unlikely(!pmd_same(pmd, *pmdp))) |
1388 | goto out_unlock; | |
1389 | ||
de466bd6 MG |
1390 | /* |
1391 | * If there are potential migrations, wait for completion and retry | |
1392 | * without disrupting NUMA hinting information. Do not relock and | |
1393 | * check_same as the page may no longer be mapped. | |
1394 | */ | |
1395 | if (unlikely(pmd_trans_migrating(*pmdp))) { | |
5d833062 | 1396 | page = pmd_page(*pmdp); |
de466bd6 | 1397 | spin_unlock(ptl); |
5d833062 | 1398 | wait_on_page_locked(page); |
de466bd6 MG |
1399 | goto out; |
1400 | } | |
1401 | ||
d10e63f2 | 1402 | page = pmd_page(pmd); |
a1a46184 | 1403 | BUG_ON(is_huge_zero_page(page)); |
8191acbd | 1404 | page_nid = page_to_nid(page); |
90572890 | 1405 | last_cpupid = page_cpupid_last(page); |
03c5a6e1 | 1406 | count_vm_numa_event(NUMA_HINT_FAULTS); |
04bb2f94 | 1407 | if (page_nid == this_nid) { |
03c5a6e1 | 1408 | count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); |
04bb2f94 RR |
1409 | flags |= TNF_FAULT_LOCAL; |
1410 | } | |
4daae3b4 | 1411 | |
bea66fbd MG |
1412 | /* See similar comment in do_numa_page for explanation */ |
1413 | if (!(vma->vm_flags & VM_WRITE)) | |
6688cc05 PZ |
1414 | flags |= TNF_NO_GROUP; |
1415 | ||
ff9042b1 MG |
1416 | /* |
1417 | * Acquire the page lock to serialise THP migrations but avoid dropping | |
1418 | * page_table_lock if at all possible | |
1419 | */ | |
b8916634 MG |
1420 | page_locked = trylock_page(page); |
1421 | target_nid = mpol_misplaced(page, vma, haddr); | |
1422 | if (target_nid == -1) { | |
1423 | /* If the page was locked, there are no parallel migrations */ | |
a54a407f | 1424 | if (page_locked) |
b8916634 | 1425 | goto clear_pmdnuma; |
2b4847e7 | 1426 | } |
4daae3b4 | 1427 | |
de466bd6 | 1428 | /* Migration could have started since the pmd_trans_migrating check */ |
2b4847e7 | 1429 | if (!page_locked) { |
c4088ebd | 1430 | spin_unlock(ptl); |
b8916634 | 1431 | wait_on_page_locked(page); |
a54a407f | 1432 | page_nid = -1; |
b8916634 MG |
1433 | goto out; |
1434 | } | |
1435 | ||
2b4847e7 MG |
1436 | /* |
1437 | * Page is misplaced. Page lock serialises migrations. Acquire anon_vma | |
1438 | * to serialises splits | |
1439 | */ | |
b8916634 | 1440 | get_page(page); |
c4088ebd | 1441 | spin_unlock(ptl); |
b8916634 | 1442 | anon_vma = page_lock_anon_vma_read(page); |
4daae3b4 | 1443 | |
c69307d5 | 1444 | /* Confirm the PMD did not change while page_table_lock was released */ |
c4088ebd | 1445 | spin_lock(ptl); |
b32967ff MG |
1446 | if (unlikely(!pmd_same(pmd, *pmdp))) { |
1447 | unlock_page(page); | |
1448 | put_page(page); | |
a54a407f | 1449 | page_nid = -1; |
4daae3b4 | 1450 | goto out_unlock; |
b32967ff | 1451 | } |
ff9042b1 | 1452 | |
c3a489ca MG |
1453 | /* Bail if we fail to protect against THP splits for any reason */ |
1454 | if (unlikely(!anon_vma)) { | |
1455 | put_page(page); | |
1456 | page_nid = -1; | |
1457 | goto clear_pmdnuma; | |
1458 | } | |
1459 | ||
a54a407f MG |
1460 | /* |
1461 | * Migrate the THP to the requested node, returns with page unlocked | |
8a0516ed | 1462 | * and access rights restored. |
a54a407f | 1463 | */ |
c4088ebd | 1464 | spin_unlock(ptl); |
b32967ff | 1465 | migrated = migrate_misplaced_transhuge_page(mm, vma, |
340ef390 | 1466 | pmdp, pmd, addr, page, target_nid); |
6688cc05 PZ |
1467 | if (migrated) { |
1468 | flags |= TNF_MIGRATED; | |
8191acbd | 1469 | page_nid = target_nid; |
074c2381 MG |
1470 | } else |
1471 | flags |= TNF_MIGRATE_FAIL; | |
b32967ff | 1472 | |
8191acbd | 1473 | goto out; |
b32967ff | 1474 | clear_pmdnuma: |
a54a407f | 1475 | BUG_ON(!PageLocked(page)); |
b191f9b1 | 1476 | was_writable = pmd_write(pmd); |
4d942466 | 1477 | pmd = pmd_modify(pmd, vma->vm_page_prot); |
b7b04004 | 1478 | pmd = pmd_mkyoung(pmd); |
b191f9b1 MG |
1479 | if (was_writable) |
1480 | pmd = pmd_mkwrite(pmd); | |
d10e63f2 | 1481 | set_pmd_at(mm, haddr, pmdp, pmd); |
d10e63f2 | 1482 | update_mmu_cache_pmd(vma, addr, pmdp); |
a54a407f | 1483 | unlock_page(page); |
d10e63f2 | 1484 | out_unlock: |
c4088ebd | 1485 | spin_unlock(ptl); |
b8916634 MG |
1486 | |
1487 | out: | |
1488 | if (anon_vma) | |
1489 | page_unlock_anon_vma_read(anon_vma); | |
1490 | ||
8191acbd | 1491 | if (page_nid != -1) |
6688cc05 | 1492 | task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags); |
8191acbd | 1493 | |
d10e63f2 MG |
1494 | return 0; |
1495 | } | |
1496 | ||
71e3aac0 | 1497 | int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, |
f21760b1 | 1498 | pmd_t *pmd, unsigned long addr) |
71e3aac0 | 1499 | { |
da146769 | 1500 | pmd_t orig_pmd; |
bf929152 | 1501 | spinlock_t *ptl; |
71e3aac0 | 1502 | |
da146769 KS |
1503 | if (__pmd_trans_huge_lock(pmd, vma, &ptl) != 1) |
1504 | return 0; | |
1505 | /* | |
1506 | * For architectures like ppc64 we look at deposited pgtable | |
1507 | * when calling pmdp_huge_get_and_clear. So do the | |
1508 | * pgtable_trans_huge_withdraw after finishing pmdp related | |
1509 | * operations. | |
1510 | */ | |
1511 | orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd, | |
1512 | tlb->fullmm); | |
1513 | tlb_remove_pmd_tlb_entry(tlb, pmd, addr); | |
1514 | if (vma_is_dax(vma)) { | |
1515 | spin_unlock(ptl); | |
1516 | if (is_huge_zero_pmd(orig_pmd)) | |
97ae1749 | 1517 | put_huge_zero_page(); |
da146769 KS |
1518 | } else if (is_huge_zero_pmd(orig_pmd)) { |
1519 | pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd)); | |
1520 | atomic_long_dec(&tlb->mm->nr_ptes); | |
1521 | spin_unlock(ptl); | |
1522 | put_huge_zero_page(); | |
1523 | } else { | |
1524 | struct page *page = pmd_page(orig_pmd); | |
d281ee61 | 1525 | page_remove_rmap(page, true); |
da146769 KS |
1526 | VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); |
1527 | add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); | |
1528 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
1529 | pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd)); | |
1530 | atomic_long_dec(&tlb->mm->nr_ptes); | |
1531 | spin_unlock(ptl); | |
1532 | tlb_remove_page(tlb, page); | |
025c5b24 | 1533 | } |
da146769 | 1534 | return 1; |
71e3aac0 AA |
1535 | } |
1536 | ||
37a1c49a AA |
1537 | int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma, |
1538 | unsigned long old_addr, | |
1539 | unsigned long new_addr, unsigned long old_end, | |
1540 | pmd_t *old_pmd, pmd_t *new_pmd) | |
1541 | { | |
bf929152 | 1542 | spinlock_t *old_ptl, *new_ptl; |
37a1c49a AA |
1543 | int ret = 0; |
1544 | pmd_t pmd; | |
1545 | ||
1546 | struct mm_struct *mm = vma->vm_mm; | |
1547 | ||
1548 | if ((old_addr & ~HPAGE_PMD_MASK) || | |
1549 | (new_addr & ~HPAGE_PMD_MASK) || | |
1550 | old_end - old_addr < HPAGE_PMD_SIZE || | |
1551 | (new_vma->vm_flags & VM_NOHUGEPAGE)) | |
1552 | goto out; | |
1553 | ||
1554 | /* | |
1555 | * The destination pmd shouldn't be established, free_pgtables() | |
1556 | * should have release it. | |
1557 | */ | |
1558 | if (WARN_ON(!pmd_none(*new_pmd))) { | |
1559 | VM_BUG_ON(pmd_trans_huge(*new_pmd)); | |
1560 | goto out; | |
1561 | } | |
1562 | ||
bf929152 KS |
1563 | /* |
1564 | * We don't have to worry about the ordering of src and dst | |
1565 | * ptlocks because exclusive mmap_sem prevents deadlock. | |
1566 | */ | |
1567 | ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl); | |
025c5b24 | 1568 | if (ret == 1) { |
bf929152 KS |
1569 | new_ptl = pmd_lockptr(mm, new_pmd); |
1570 | if (new_ptl != old_ptl) | |
1571 | spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); | |
8809aa2d | 1572 | pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); |
025c5b24 | 1573 | VM_BUG_ON(!pmd_none(*new_pmd)); |
3592806c | 1574 | |
b3084f4d AK |
1575 | if (pmd_move_must_withdraw(new_ptl, old_ptl)) { |
1576 | pgtable_t pgtable; | |
3592806c KS |
1577 | pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); |
1578 | pgtable_trans_huge_deposit(mm, new_pmd, pgtable); | |
3592806c | 1579 | } |
b3084f4d AK |
1580 | set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd)); |
1581 | if (new_ptl != old_ptl) | |
1582 | spin_unlock(new_ptl); | |
bf929152 | 1583 | spin_unlock(old_ptl); |
37a1c49a AA |
1584 | } |
1585 | out: | |
1586 | return ret; | |
1587 | } | |
1588 | ||
f123d74a MG |
1589 | /* |
1590 | * Returns | |
1591 | * - 0 if PMD could not be locked | |
1592 | * - 1 if PMD was locked but protections unchange and TLB flush unnecessary | |
1593 | * - HPAGE_PMD_NR is protections changed and TLB flush necessary | |
1594 | */ | |
cd7548ab | 1595 | int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, |
e944fd67 | 1596 | unsigned long addr, pgprot_t newprot, int prot_numa) |
cd7548ab JW |
1597 | { |
1598 | struct mm_struct *mm = vma->vm_mm; | |
bf929152 | 1599 | spinlock_t *ptl; |
cd7548ab JW |
1600 | int ret = 0; |
1601 | ||
bf929152 | 1602 | if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) { |
025c5b24 | 1603 | pmd_t entry; |
b191f9b1 | 1604 | bool preserve_write = prot_numa && pmd_write(*pmd); |
ba68bc01 | 1605 | ret = 1; |
e944fd67 MG |
1606 | |
1607 | /* | |
1608 | * Avoid trapping faults against the zero page. The read-only | |
1609 | * data is likely to be read-cached on the local CPU and | |
1610 | * local/remote hits to the zero page are not interesting. | |
1611 | */ | |
1612 | if (prot_numa && is_huge_zero_pmd(*pmd)) { | |
1613 | spin_unlock(ptl); | |
ba68bc01 | 1614 | return ret; |
e944fd67 MG |
1615 | } |
1616 | ||
10c1045f | 1617 | if (!prot_numa || !pmd_protnone(*pmd)) { |
8809aa2d | 1618 | entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd); |
10c1045f | 1619 | entry = pmd_modify(entry, newprot); |
b191f9b1 MG |
1620 | if (preserve_write) |
1621 | entry = pmd_mkwrite(entry); | |
10c1045f MG |
1622 | ret = HPAGE_PMD_NR; |
1623 | set_pmd_at(mm, addr, pmd, entry); | |
b191f9b1 | 1624 | BUG_ON(!preserve_write && pmd_write(entry)); |
10c1045f | 1625 | } |
bf929152 | 1626 | spin_unlock(ptl); |
025c5b24 NH |
1627 | } |
1628 | ||
1629 | return ret; | |
1630 | } | |
1631 | ||
1632 | /* | |
1633 | * Returns 1 if a given pmd maps a stable (not under splitting) thp. | |
1634 | * Returns -1 if it maps a thp under splitting. Returns 0 otherwise. | |
1635 | * | |
1636 | * Note that if it returns 1, this routine returns without unlocking page | |
1637 | * table locks. So callers must unlock them. | |
1638 | */ | |
bf929152 KS |
1639 | int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma, |
1640 | spinlock_t **ptl) | |
025c5b24 | 1641 | { |
bf929152 | 1642 | *ptl = pmd_lock(vma->vm_mm, pmd); |
cd7548ab JW |
1643 | if (likely(pmd_trans_huge(*pmd))) { |
1644 | if (unlikely(pmd_trans_splitting(*pmd))) { | |
bf929152 | 1645 | spin_unlock(*ptl); |
cd7548ab | 1646 | wait_split_huge_page(vma->anon_vma, pmd); |
025c5b24 | 1647 | return -1; |
cd7548ab | 1648 | } else { |
025c5b24 NH |
1649 | /* Thp mapped by 'pmd' is stable, so we can |
1650 | * handle it as it is. */ | |
1651 | return 1; | |
cd7548ab | 1652 | } |
025c5b24 | 1653 | } |
bf929152 | 1654 | spin_unlock(*ptl); |
025c5b24 | 1655 | return 0; |
cd7548ab JW |
1656 | } |
1657 | ||
117b0791 KS |
1658 | /* |
1659 | * This function returns whether a given @page is mapped onto the @address | |
1660 | * in the virtual space of @mm. | |
1661 | * | |
1662 | * When it's true, this function returns *pmd with holding the page table lock | |
1663 | * and passing it back to the caller via @ptl. | |
1664 | * If it's false, returns NULL without holding the page table lock. | |
1665 | */ | |
71e3aac0 AA |
1666 | pmd_t *page_check_address_pmd(struct page *page, |
1667 | struct mm_struct *mm, | |
1668 | unsigned long address, | |
117b0791 KS |
1669 | enum page_check_address_pmd_flag flag, |
1670 | spinlock_t **ptl) | |
71e3aac0 | 1671 | { |
b5a8cad3 KS |
1672 | pgd_t *pgd; |
1673 | pud_t *pud; | |
117b0791 | 1674 | pmd_t *pmd; |
71e3aac0 AA |
1675 | |
1676 | if (address & ~HPAGE_PMD_MASK) | |
117b0791 | 1677 | return NULL; |
71e3aac0 | 1678 | |
b5a8cad3 KS |
1679 | pgd = pgd_offset(mm, address); |
1680 | if (!pgd_present(*pgd)) | |
117b0791 | 1681 | return NULL; |
b5a8cad3 KS |
1682 | pud = pud_offset(pgd, address); |
1683 | if (!pud_present(*pud)) | |
1684 | return NULL; | |
1685 | pmd = pmd_offset(pud, address); | |
1686 | ||
117b0791 | 1687 | *ptl = pmd_lock(mm, pmd); |
b5a8cad3 | 1688 | if (!pmd_present(*pmd)) |
117b0791 | 1689 | goto unlock; |
71e3aac0 | 1690 | if (pmd_page(*pmd) != page) |
117b0791 | 1691 | goto unlock; |
94fcc585 AA |
1692 | /* |
1693 | * split_vma() may create temporary aliased mappings. There is | |
1694 | * no risk as long as all huge pmd are found and have their | |
1695 | * splitting bit set before __split_huge_page_refcount | |
1696 | * runs. Finding the same huge pmd more than once during the | |
1697 | * same rmap walk is not a problem. | |
1698 | */ | |
1699 | if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG && | |
1700 | pmd_trans_splitting(*pmd)) | |
117b0791 | 1701 | goto unlock; |
71e3aac0 AA |
1702 | if (pmd_trans_huge(*pmd)) { |
1703 | VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG && | |
1704 | !pmd_trans_splitting(*pmd)); | |
117b0791 | 1705 | return pmd; |
71e3aac0 | 1706 | } |
117b0791 KS |
1707 | unlock: |
1708 | spin_unlock(*ptl); | |
1709 | return NULL; | |
71e3aac0 AA |
1710 | } |
1711 | ||
1712 | static int __split_huge_page_splitting(struct page *page, | |
1713 | struct vm_area_struct *vma, | |
1714 | unsigned long address) | |
1715 | { | |
1716 | struct mm_struct *mm = vma->vm_mm; | |
117b0791 | 1717 | spinlock_t *ptl; |
71e3aac0 AA |
1718 | pmd_t *pmd; |
1719 | int ret = 0; | |
2ec74c3e SG |
1720 | /* For mmu_notifiers */ |
1721 | const unsigned long mmun_start = address; | |
1722 | const unsigned long mmun_end = address + HPAGE_PMD_SIZE; | |
71e3aac0 | 1723 | |
2ec74c3e | 1724 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
71e3aac0 | 1725 | pmd = page_check_address_pmd(page, mm, address, |
117b0791 | 1726 | PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl); |
71e3aac0 AA |
1727 | if (pmd) { |
1728 | /* | |
1729 | * We can't temporarily set the pmd to null in order | |
1730 | * to split it, the pmd must remain marked huge at all | |
1731 | * times or the VM won't take the pmd_trans_huge paths | |
5a505085 | 1732 | * and it won't wait on the anon_vma->root->rwsem to |
71e3aac0 AA |
1733 | * serialize against split_huge_page*. |
1734 | */ | |
2ec74c3e | 1735 | pmdp_splitting_flush(vma, address, pmd); |
34ee645e | 1736 | |
71e3aac0 | 1737 | ret = 1; |
117b0791 | 1738 | spin_unlock(ptl); |
71e3aac0 | 1739 | } |
2ec74c3e | 1740 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
71e3aac0 AA |
1741 | |
1742 | return ret; | |
1743 | } | |
1744 | ||
5bc7b8ac SL |
1745 | static void __split_huge_page_refcount(struct page *page, |
1746 | struct list_head *list) | |
71e3aac0 AA |
1747 | { |
1748 | int i; | |
71e3aac0 | 1749 | struct zone *zone = page_zone(page); |
fa9add64 | 1750 | struct lruvec *lruvec; |
70b50f94 | 1751 | int tail_count = 0; |
71e3aac0 AA |
1752 | |
1753 | /* prevent PageLRU to go away from under us, and freeze lru stats */ | |
1754 | spin_lock_irq(&zone->lru_lock); | |
fa9add64 HD |
1755 | lruvec = mem_cgroup_page_lruvec(page, zone); |
1756 | ||
71e3aac0 | 1757 | compound_lock(page); |
e94c8a9c KH |
1758 | /* complete memcg works before add pages to LRU */ |
1759 | mem_cgroup_split_huge_fixup(page); | |
71e3aac0 | 1760 | |
45676885 | 1761 | for (i = HPAGE_PMD_NR - 1; i >= 1; i--) { |
71e3aac0 AA |
1762 | struct page *page_tail = page + i; |
1763 | ||
70b50f94 AA |
1764 | /* tail_page->_mapcount cannot change */ |
1765 | BUG_ON(page_mapcount(page_tail) < 0); | |
1766 | tail_count += page_mapcount(page_tail); | |
1767 | /* check for overflow */ | |
1768 | BUG_ON(tail_count < 0); | |
1769 | BUG_ON(atomic_read(&page_tail->_count) != 0); | |
1770 | /* | |
1771 | * tail_page->_count is zero and not changing from | |
1772 | * under us. But get_page_unless_zero() may be running | |
1773 | * from under us on the tail_page. If we used | |
1774 | * atomic_set() below instead of atomic_add(), we | |
1775 | * would then run atomic_set() concurrently with | |
1776 | * get_page_unless_zero(), and atomic_set() is | |
1777 | * implemented in C not using locked ops. spin_unlock | |
1778 | * on x86 sometime uses locked ops because of PPro | |
1779 | * errata 66, 92, so unless somebody can guarantee | |
1780 | * atomic_set() here would be safe on all archs (and | |
1781 | * not only on x86), it's safer to use atomic_add(). | |
1782 | */ | |
1783 | atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1, | |
1784 | &page_tail->_count); | |
71e3aac0 AA |
1785 | |
1786 | /* after clearing PageTail the gup refcount can be released */ | |
3a79d52a | 1787 | smp_mb__after_atomic(); |
71e3aac0 | 1788 | |
f4c18e6f | 1789 | page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; |
71e3aac0 AA |
1790 | page_tail->flags |= (page->flags & |
1791 | ((1L << PG_referenced) | | |
1792 | (1L << PG_swapbacked) | | |
1793 | (1L << PG_mlocked) | | |
e180cf80 KS |
1794 | (1L << PG_uptodate) | |
1795 | (1L << PG_active) | | |
1796 | (1L << PG_unevictable))); | |
71e3aac0 AA |
1797 | page_tail->flags |= (1L << PG_dirty); |
1798 | ||
1d798ca3 | 1799 | clear_compound_head(page_tail); |
71e3aac0 | 1800 | |
33c3fc71 VD |
1801 | if (page_is_young(page)) |
1802 | set_page_young(page_tail); | |
1803 | if (page_is_idle(page)) | |
1804 | set_page_idle(page_tail); | |
1805 | ||
71e3aac0 AA |
1806 | /* |
1807 | * __split_huge_page_splitting() already set the | |
1808 | * splitting bit in all pmd that could map this | |
1809 | * hugepage, that will ensure no CPU can alter the | |
1810 | * mapcount on the head page. The mapcount is only | |
1811 | * accounted in the head page and it has to be | |
1812 | * transferred to all tail pages in the below code. So | |
1813 | * for this code to be safe, the split the mapcount | |
1814 | * can't change. But that doesn't mean userland can't | |
1815 | * keep changing and reading the page contents while | |
1816 | * we transfer the mapcount, so the pmd splitting | |
1817 | * status is achieved setting a reserved bit in the | |
1818 | * pmd, not by clearing the present bit. | |
1819 | */ | |
71e3aac0 AA |
1820 | page_tail->_mapcount = page->_mapcount; |
1821 | ||
1c290f64 | 1822 | BUG_ON(page_tail->mapping != TAIL_MAPPING); |
71e3aac0 AA |
1823 | page_tail->mapping = page->mapping; |
1824 | ||
45676885 | 1825 | page_tail->index = page->index + i; |
90572890 | 1826 | page_cpupid_xchg_last(page_tail, page_cpupid_last(page)); |
71e3aac0 AA |
1827 | |
1828 | BUG_ON(!PageAnon(page_tail)); | |
1829 | BUG_ON(!PageUptodate(page_tail)); | |
1830 | BUG_ON(!PageDirty(page_tail)); | |
1831 | BUG_ON(!PageSwapBacked(page_tail)); | |
1832 | ||
5bc7b8ac | 1833 | lru_add_page_tail(page, page_tail, lruvec, list); |
71e3aac0 | 1834 | } |
70b50f94 AA |
1835 | atomic_sub(tail_count, &page->_count); |
1836 | BUG_ON(atomic_read(&page->_count) <= 0); | |
71e3aac0 | 1837 | |
fa9add64 | 1838 | __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1); |
79134171 | 1839 | |
71e3aac0 AA |
1840 | ClearPageCompound(page); |
1841 | compound_unlock(page); | |
1842 | spin_unlock_irq(&zone->lru_lock); | |
1843 | ||
1844 | for (i = 1; i < HPAGE_PMD_NR; i++) { | |
1845 | struct page *page_tail = page + i; | |
1846 | BUG_ON(page_count(page_tail) <= 0); | |
1847 | /* | |
1848 | * Tail pages may be freed if there wasn't any mapping | |
1849 | * like if add_to_swap() is running on a lru page that | |
1850 | * had its mapping zapped. And freeing these pages | |
1851 | * requires taking the lru_lock so we do the put_page | |
1852 | * of the tail pages after the split is complete. | |
1853 | */ | |
1854 | put_page(page_tail); | |
1855 | } | |
1856 | ||
1857 | /* | |
1858 | * Only the head page (now become a regular page) is required | |
1859 | * to be pinned by the caller. | |
1860 | */ | |
1861 | BUG_ON(page_count(page) <= 0); | |
1862 | } | |
1863 | ||
1864 | static int __split_huge_page_map(struct page *page, | |
1865 | struct vm_area_struct *vma, | |
1866 | unsigned long address) | |
1867 | { | |
1868 | struct mm_struct *mm = vma->vm_mm; | |
117b0791 | 1869 | spinlock_t *ptl; |
71e3aac0 AA |
1870 | pmd_t *pmd, _pmd; |
1871 | int ret = 0, i; | |
1872 | pgtable_t pgtable; | |
1873 | unsigned long haddr; | |
1874 | ||
71e3aac0 | 1875 | pmd = page_check_address_pmd(page, mm, address, |
117b0791 | 1876 | PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl); |
71e3aac0 | 1877 | if (pmd) { |
6b0b50b0 | 1878 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); |
71e3aac0 | 1879 | pmd_populate(mm, &_pmd, pgtable); |
f8303c25 WL |
1880 | if (pmd_write(*pmd)) |
1881 | BUG_ON(page_mapcount(page) != 1); | |
71e3aac0 | 1882 | |
e3ebcf64 GS |
1883 | haddr = address; |
1884 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
71e3aac0 AA |
1885 | pte_t *pte, entry; |
1886 | BUG_ON(PageCompound(page+i)); | |
abc40bd2 | 1887 | /* |
8a0516ed MG |
1888 | * Note that NUMA hinting access restrictions are not |
1889 | * transferred to avoid any possibility of altering | |
1890 | * permissions across VMAs. | |
abc40bd2 | 1891 | */ |
71e3aac0 AA |
1892 | entry = mk_pte(page + i, vma->vm_page_prot); |
1893 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); | |
1894 | if (!pmd_write(*pmd)) | |
1895 | entry = pte_wrprotect(entry); | |
71e3aac0 AA |
1896 | if (!pmd_young(*pmd)) |
1897 | entry = pte_mkold(entry); | |
1898 | pte = pte_offset_map(&_pmd, haddr); | |
1899 | BUG_ON(!pte_none(*pte)); | |
1900 | set_pte_at(mm, haddr, pte, entry); | |
1901 | pte_unmap(pte); | |
1902 | } | |
1903 | ||
71e3aac0 AA |
1904 | smp_wmb(); /* make pte visible before pmd */ |
1905 | /* | |
1906 | * Up to this point the pmd is present and huge and | |
1907 | * userland has the whole access to the hugepage | |
1908 | * during the split (which happens in place). If we | |
1909 | * overwrite the pmd with the not-huge version | |
1910 | * pointing to the pte here (which of course we could | |
1911 | * if all CPUs were bug free), userland could trigger | |
1912 | * a small page size TLB miss on the small sized TLB | |
1913 | * while the hugepage TLB entry is still established | |
1914 | * in the huge TLB. Some CPU doesn't like that. See | |
1915 | * http://support.amd.com/us/Processor_TechDocs/41322.pdf, | |
1916 | * Erratum 383 on page 93. Intel should be safe but is | |
1917 | * also warns that it's only safe if the permission | |
1918 | * and cache attributes of the two entries loaded in | |
1919 | * the two TLB is identical (which should be the case | |
1920 | * here). But it is generally safer to never allow | |
1921 | * small and huge TLB entries for the same virtual | |
1922 | * address to be loaded simultaneously. So instead of | |
12ebc158 | 1923 | * doing "pmd_populate(); flush_pmd_tlb_range();" we first |
71e3aac0 AA |
1924 | * mark the current pmd notpresent (atomically because |
1925 | * here the pmd_trans_huge and pmd_trans_splitting | |
1926 | * must remain set at all times on the pmd until the | |
1927 | * split is complete for this pmd), then we flush the | |
1928 | * SMP TLB and finally we write the non-huge version | |
1929 | * of the pmd entry with pmd_populate. | |
1930 | */ | |
46dcde73 | 1931 | pmdp_invalidate(vma, address, pmd); |
71e3aac0 AA |
1932 | pmd_populate(mm, pmd, pgtable); |
1933 | ret = 1; | |
117b0791 | 1934 | spin_unlock(ptl); |
71e3aac0 | 1935 | } |
71e3aac0 AA |
1936 | |
1937 | return ret; | |
1938 | } | |
1939 | ||
5a505085 | 1940 | /* must be called with anon_vma->root->rwsem held */ |
71e3aac0 | 1941 | static void __split_huge_page(struct page *page, |
5bc7b8ac SL |
1942 | struct anon_vma *anon_vma, |
1943 | struct list_head *list) | |
71e3aac0 AA |
1944 | { |
1945 | int mapcount, mapcount2; | |
bf181b9f | 1946 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); |
71e3aac0 AA |
1947 | struct anon_vma_chain *avc; |
1948 | ||
1949 | BUG_ON(!PageHead(page)); | |
1950 | BUG_ON(PageTail(page)); | |
1951 | ||
1952 | mapcount = 0; | |
bf181b9f | 1953 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { |
71e3aac0 AA |
1954 | struct vm_area_struct *vma = avc->vma; |
1955 | unsigned long addr = vma_address(page, vma); | |
1956 | BUG_ON(is_vma_temporary_stack(vma)); | |
71e3aac0 AA |
1957 | mapcount += __split_huge_page_splitting(page, vma, addr); |
1958 | } | |
05759d38 AA |
1959 | /* |
1960 | * It is critical that new vmas are added to the tail of the | |
1961 | * anon_vma list. This guarantes that if copy_huge_pmd() runs | |
1962 | * and establishes a child pmd before | |
1963 | * __split_huge_page_splitting() freezes the parent pmd (so if | |
1964 | * we fail to prevent copy_huge_pmd() from running until the | |
1965 | * whole __split_huge_page() is complete), we will still see | |
1966 | * the newly established pmd of the child later during the | |
1967 | * walk, to be able to set it as pmd_trans_splitting too. | |
1968 | */ | |
ff9e43eb | 1969 | if (mapcount != page_mapcount(page)) { |
ae3a8c1c AM |
1970 | pr_err("mapcount %d page_mapcount %d\n", |
1971 | mapcount, page_mapcount(page)); | |
ff9e43eb KS |
1972 | BUG(); |
1973 | } | |
71e3aac0 | 1974 | |
5bc7b8ac | 1975 | __split_huge_page_refcount(page, list); |
71e3aac0 AA |
1976 | |
1977 | mapcount2 = 0; | |
bf181b9f | 1978 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { |
71e3aac0 AA |
1979 | struct vm_area_struct *vma = avc->vma; |
1980 | unsigned long addr = vma_address(page, vma); | |
1981 | BUG_ON(is_vma_temporary_stack(vma)); | |
71e3aac0 AA |
1982 | mapcount2 += __split_huge_page_map(page, vma, addr); |
1983 | } | |
ff9e43eb | 1984 | if (mapcount != mapcount2) { |
ae3a8c1c AM |
1985 | pr_err("mapcount %d mapcount2 %d page_mapcount %d\n", |
1986 | mapcount, mapcount2, page_mapcount(page)); | |
ff9e43eb KS |
1987 | BUG(); |
1988 | } | |
71e3aac0 AA |
1989 | } |
1990 | ||
5bc7b8ac SL |
1991 | /* |
1992 | * Split a hugepage into normal pages. This doesn't change the position of head | |
1993 | * page. If @list is null, tail pages will be added to LRU list, otherwise, to | |
1994 | * @list. Both head page and tail pages will inherit mapping, flags, and so on | |
1995 | * from the hugepage. | |
1996 | * Return 0 if the hugepage is split successfully otherwise return 1. | |
1997 | */ | |
1998 | int split_huge_page_to_list(struct page *page, struct list_head *list) | |
71e3aac0 AA |
1999 | { |
2000 | struct anon_vma *anon_vma; | |
2001 | int ret = 1; | |
2002 | ||
5918d10a | 2003 | BUG_ON(is_huge_zero_page(page)); |
71e3aac0 | 2004 | BUG_ON(!PageAnon(page)); |
062f1af2 MG |
2005 | |
2006 | /* | |
2007 | * The caller does not necessarily hold an mmap_sem that would prevent | |
2008 | * the anon_vma disappearing so we first we take a reference to it | |
2009 | * and then lock the anon_vma for write. This is similar to | |
2010 | * page_lock_anon_vma_read except the write lock is taken to serialise | |
2011 | * against parallel split or collapse operations. | |
2012 | */ | |
2013 | anon_vma = page_get_anon_vma(page); | |
71e3aac0 AA |
2014 | if (!anon_vma) |
2015 | goto out; | |
062f1af2 MG |
2016 | anon_vma_lock_write(anon_vma); |
2017 | ||
71e3aac0 AA |
2018 | ret = 0; |
2019 | if (!PageCompound(page)) | |
2020 | goto out_unlock; | |
2021 | ||
2022 | BUG_ON(!PageSwapBacked(page)); | |
5bc7b8ac | 2023 | __split_huge_page(page, anon_vma, list); |
81ab4201 | 2024 | count_vm_event(THP_SPLIT); |
71e3aac0 AA |
2025 | |
2026 | BUG_ON(PageCompound(page)); | |
2027 | out_unlock: | |
08b52706 | 2028 | anon_vma_unlock_write(anon_vma); |
062f1af2 | 2029 | put_anon_vma(anon_vma); |
71e3aac0 AA |
2030 | out: |
2031 | return ret; | |
2032 | } | |
2033 | ||
9050d7eb | 2034 | #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE) |
78f11a25 | 2035 | |
60ab3244 AA |
2036 | int hugepage_madvise(struct vm_area_struct *vma, |
2037 | unsigned long *vm_flags, int advice) | |
0af4e98b | 2038 | { |
a664b2d8 AA |
2039 | switch (advice) { |
2040 | case MADV_HUGEPAGE: | |
1e1836e8 AT |
2041 | #ifdef CONFIG_S390 |
2042 | /* | |
2043 | * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390 | |
2044 | * can't handle this properly after s390_enable_sie, so we simply | |
2045 | * ignore the madvise to prevent qemu from causing a SIGSEGV. | |
2046 | */ | |
2047 | if (mm_has_pgste(vma->vm_mm)) | |
2048 | return 0; | |
2049 | #endif | |
a664b2d8 AA |
2050 | /* |
2051 | * Be somewhat over-protective like KSM for now! | |
2052 | */ | |
1a763615 | 2053 | if (*vm_flags & VM_NO_THP) |
a664b2d8 AA |
2054 | return -EINVAL; |
2055 | *vm_flags &= ~VM_NOHUGEPAGE; | |
2056 | *vm_flags |= VM_HUGEPAGE; | |
60ab3244 AA |
2057 | /* |
2058 | * If the vma become good for khugepaged to scan, | |
2059 | * register it here without waiting a page fault that | |
2060 | * may not happen any time soon. | |
2061 | */ | |
6d50e60c | 2062 | if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags))) |
60ab3244 | 2063 | return -ENOMEM; |
a664b2d8 AA |
2064 | break; |
2065 | case MADV_NOHUGEPAGE: | |
2066 | /* | |
2067 | * Be somewhat over-protective like KSM for now! | |
2068 | */ | |
1a763615 | 2069 | if (*vm_flags & VM_NO_THP) |
a664b2d8 AA |
2070 | return -EINVAL; |
2071 | *vm_flags &= ~VM_HUGEPAGE; | |
2072 | *vm_flags |= VM_NOHUGEPAGE; | |
60ab3244 AA |
2073 | /* |
2074 | * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning | |
2075 | * this vma even if we leave the mm registered in khugepaged if | |
2076 | * it got registered before VM_NOHUGEPAGE was set. | |
2077 | */ | |
a664b2d8 AA |
2078 | break; |
2079 | } | |
0af4e98b AA |
2080 | |
2081 | return 0; | |
2082 | } | |
2083 | ||
ba76149f AA |
2084 | static int __init khugepaged_slab_init(void) |
2085 | { | |
2086 | mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", | |
2087 | sizeof(struct mm_slot), | |
2088 | __alignof__(struct mm_slot), 0, NULL); | |
2089 | if (!mm_slot_cache) | |
2090 | return -ENOMEM; | |
2091 | ||
2092 | return 0; | |
2093 | } | |
2094 | ||
65ebb64f KS |
2095 | static void __init khugepaged_slab_exit(void) |
2096 | { | |
2097 | kmem_cache_destroy(mm_slot_cache); | |
2098 | } | |
2099 | ||
ba76149f AA |
2100 | static inline struct mm_slot *alloc_mm_slot(void) |
2101 | { | |
2102 | if (!mm_slot_cache) /* initialization failed */ | |
2103 | return NULL; | |
2104 | return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); | |
2105 | } | |
2106 | ||
2107 | static inline void free_mm_slot(struct mm_slot *mm_slot) | |
2108 | { | |
2109 | kmem_cache_free(mm_slot_cache, mm_slot); | |
2110 | } | |
2111 | ||
ba76149f AA |
2112 | static struct mm_slot *get_mm_slot(struct mm_struct *mm) |
2113 | { | |
2114 | struct mm_slot *mm_slot; | |
ba76149f | 2115 | |
b67bfe0d | 2116 | hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm) |
ba76149f AA |
2117 | if (mm == mm_slot->mm) |
2118 | return mm_slot; | |
43b5fbbd | 2119 | |
ba76149f AA |
2120 | return NULL; |
2121 | } | |
2122 | ||
2123 | static void insert_to_mm_slots_hash(struct mm_struct *mm, | |
2124 | struct mm_slot *mm_slot) | |
2125 | { | |
ba76149f | 2126 | mm_slot->mm = mm; |
43b5fbbd | 2127 | hash_add(mm_slots_hash, &mm_slot->hash, (long)mm); |
ba76149f AA |
2128 | } |
2129 | ||
2130 | static inline int khugepaged_test_exit(struct mm_struct *mm) | |
2131 | { | |
2132 | return atomic_read(&mm->mm_users) == 0; | |
2133 | } | |
2134 | ||
2135 | int __khugepaged_enter(struct mm_struct *mm) | |
2136 | { | |
2137 | struct mm_slot *mm_slot; | |
2138 | int wakeup; | |
2139 | ||
2140 | mm_slot = alloc_mm_slot(); | |
2141 | if (!mm_slot) | |
2142 | return -ENOMEM; | |
2143 | ||
2144 | /* __khugepaged_exit() must not run from under us */ | |
96dad67f | 2145 | VM_BUG_ON_MM(khugepaged_test_exit(mm), mm); |
ba76149f AA |
2146 | if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { |
2147 | free_mm_slot(mm_slot); | |
2148 | return 0; | |
2149 | } | |
2150 | ||
2151 | spin_lock(&khugepaged_mm_lock); | |
2152 | insert_to_mm_slots_hash(mm, mm_slot); | |
2153 | /* | |
2154 | * Insert just behind the scanning cursor, to let the area settle | |
2155 | * down a little. | |
2156 | */ | |
2157 | wakeup = list_empty(&khugepaged_scan.mm_head); | |
2158 | list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head); | |
2159 | spin_unlock(&khugepaged_mm_lock); | |
2160 | ||
2161 | atomic_inc(&mm->mm_count); | |
2162 | if (wakeup) | |
2163 | wake_up_interruptible(&khugepaged_wait); | |
2164 | ||
2165 | return 0; | |
2166 | } | |
2167 | ||
6d50e60c DR |
2168 | int khugepaged_enter_vma_merge(struct vm_area_struct *vma, |
2169 | unsigned long vm_flags) | |
ba76149f AA |
2170 | { |
2171 | unsigned long hstart, hend; | |
2172 | if (!vma->anon_vma) | |
2173 | /* | |
2174 | * Not yet faulted in so we will register later in the | |
2175 | * page fault if needed. | |
2176 | */ | |
2177 | return 0; | |
78f11a25 | 2178 | if (vma->vm_ops) |
ba76149f AA |
2179 | /* khugepaged not yet working on file or special mappings */ |
2180 | return 0; | |
6d50e60c | 2181 | VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma); |
ba76149f AA |
2182 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; |
2183 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
2184 | if (hstart < hend) | |
6d50e60c | 2185 | return khugepaged_enter(vma, vm_flags); |
ba76149f AA |
2186 | return 0; |
2187 | } | |
2188 | ||
2189 | void __khugepaged_exit(struct mm_struct *mm) | |
2190 | { | |
2191 | struct mm_slot *mm_slot; | |
2192 | int free = 0; | |
2193 | ||
2194 | spin_lock(&khugepaged_mm_lock); | |
2195 | mm_slot = get_mm_slot(mm); | |
2196 | if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { | |
43b5fbbd | 2197 | hash_del(&mm_slot->hash); |
ba76149f AA |
2198 | list_del(&mm_slot->mm_node); |
2199 | free = 1; | |
2200 | } | |
d788e80a | 2201 | spin_unlock(&khugepaged_mm_lock); |
ba76149f AA |
2202 | |
2203 | if (free) { | |
ba76149f AA |
2204 | clear_bit(MMF_VM_HUGEPAGE, &mm->flags); |
2205 | free_mm_slot(mm_slot); | |
2206 | mmdrop(mm); | |
2207 | } else if (mm_slot) { | |
ba76149f AA |
2208 | /* |
2209 | * This is required to serialize against | |
2210 | * khugepaged_test_exit() (which is guaranteed to run | |
2211 | * under mmap sem read mode). Stop here (after we | |
2212 | * return all pagetables will be destroyed) until | |
2213 | * khugepaged has finished working on the pagetables | |
2214 | * under the mmap_sem. | |
2215 | */ | |
2216 | down_write(&mm->mmap_sem); | |
2217 | up_write(&mm->mmap_sem); | |
d788e80a | 2218 | } |
ba76149f AA |
2219 | } |
2220 | ||
2221 | static void release_pte_page(struct page *page) | |
2222 | { | |
2223 | /* 0 stands for page_is_file_cache(page) == false */ | |
2224 | dec_zone_page_state(page, NR_ISOLATED_ANON + 0); | |
2225 | unlock_page(page); | |
2226 | putback_lru_page(page); | |
2227 | } | |
2228 | ||
2229 | static void release_pte_pages(pte_t *pte, pte_t *_pte) | |
2230 | { | |
2231 | while (--_pte >= pte) { | |
2232 | pte_t pteval = *_pte; | |
ca0984ca | 2233 | if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval))) |
ba76149f AA |
2234 | release_pte_page(pte_page(pteval)); |
2235 | } | |
2236 | } | |
2237 | ||
ba76149f AA |
2238 | static int __collapse_huge_page_isolate(struct vm_area_struct *vma, |
2239 | unsigned long address, | |
2240 | pte_t *pte) | |
2241 | { | |
7d2eba05 | 2242 | struct page *page = NULL; |
ba76149f | 2243 | pte_t *_pte; |
7d2eba05 | 2244 | int none_or_zero = 0, result = 0; |
10359213 | 2245 | bool referenced = false, writable = false; |
7d2eba05 | 2246 | |
ba76149f AA |
2247 | for (_pte = pte; _pte < pte+HPAGE_PMD_NR; |
2248 | _pte++, address += PAGE_SIZE) { | |
2249 | pte_t pteval = *_pte; | |
47aee4d8 MK |
2250 | if (pte_none(pteval) || (pte_present(pteval) && |
2251 | is_zero_pfn(pte_pfn(pteval)))) { | |
c1294d05 | 2252 | if (!userfaultfd_armed(vma) && |
7d2eba05 | 2253 | ++none_or_zero <= khugepaged_max_ptes_none) { |
ba76149f | 2254 | continue; |
7d2eba05 EA |
2255 | } else { |
2256 | result = SCAN_EXCEED_NONE_PTE; | |
ba76149f | 2257 | goto out; |
7d2eba05 | 2258 | } |
ba76149f | 2259 | } |
7d2eba05 EA |
2260 | if (!pte_present(pteval)) { |
2261 | result = SCAN_PTE_NON_PRESENT; | |
ba76149f | 2262 | goto out; |
7d2eba05 | 2263 | } |
ba76149f | 2264 | page = vm_normal_page(vma, address, pteval); |
7d2eba05 EA |
2265 | if (unlikely(!page)) { |
2266 | result = SCAN_PAGE_NULL; | |
ba76149f | 2267 | goto out; |
7d2eba05 | 2268 | } |
344aa35c | 2269 | |
309381fe SL |
2270 | VM_BUG_ON_PAGE(PageCompound(page), page); |
2271 | VM_BUG_ON_PAGE(!PageAnon(page), page); | |
2272 | VM_BUG_ON_PAGE(!PageSwapBacked(page), page); | |
ba76149f | 2273 | |
ba76149f AA |
2274 | /* |
2275 | * We can do it before isolate_lru_page because the | |
2276 | * page can't be freed from under us. NOTE: PG_lock | |
2277 | * is needed to serialize against split_huge_page | |
2278 | * when invoked from the VM. | |
2279 | */ | |
7d2eba05 EA |
2280 | if (!trylock_page(page)) { |
2281 | result = SCAN_PAGE_LOCK; | |
ba76149f | 2282 | goto out; |
7d2eba05 | 2283 | } |
10359213 EA |
2284 | |
2285 | /* | |
2286 | * cannot use mapcount: can't collapse if there's a gup pin. | |
2287 | * The page must only be referenced by the scanned process | |
2288 | * and page swap cache. | |
2289 | */ | |
2290 | if (page_count(page) != 1 + !!PageSwapCache(page)) { | |
2291 | unlock_page(page); | |
7d2eba05 | 2292 | result = SCAN_PAGE_COUNT; |
10359213 EA |
2293 | goto out; |
2294 | } | |
2295 | if (pte_write(pteval)) { | |
2296 | writable = true; | |
2297 | } else { | |
2298 | if (PageSwapCache(page) && !reuse_swap_page(page)) { | |
2299 | unlock_page(page); | |
7d2eba05 | 2300 | result = SCAN_SWAP_CACHE_PAGE; |
10359213 EA |
2301 | goto out; |
2302 | } | |
2303 | /* | |
2304 | * Page is not in the swap cache. It can be collapsed | |
2305 | * into a THP. | |
2306 | */ | |
2307 | } | |
2308 | ||
ba76149f AA |
2309 | /* |
2310 | * Isolate the page to avoid collapsing an hugepage | |
2311 | * currently in use by the VM. | |
2312 | */ | |
2313 | if (isolate_lru_page(page)) { | |
2314 | unlock_page(page); | |
7d2eba05 | 2315 | result = SCAN_DEL_PAGE_LRU; |
ba76149f AA |
2316 | goto out; |
2317 | } | |
2318 | /* 0 stands for page_is_file_cache(page) == false */ | |
2319 | inc_zone_page_state(page, NR_ISOLATED_ANON + 0); | |
309381fe SL |
2320 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
2321 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
ba76149f AA |
2322 | |
2323 | /* If there is no mapped pte young don't collapse the page */ | |
33c3fc71 VD |
2324 | if (pte_young(pteval) || |
2325 | page_is_young(page) || PageReferenced(page) || | |
8ee53820 | 2326 | mmu_notifier_test_young(vma->vm_mm, address)) |
10359213 | 2327 | referenced = true; |
ba76149f | 2328 | } |
7d2eba05 EA |
2329 | if (likely(writable)) { |
2330 | if (likely(referenced)) { | |
2331 | result = SCAN_SUCCEED; | |
2332 | trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero, | |
2333 | referenced, writable, result); | |
2334 | return 1; | |
2335 | } | |
2336 | } else { | |
2337 | result = SCAN_PAGE_RO; | |
2338 | } | |
2339 | ||
ba76149f | 2340 | out: |
344aa35c | 2341 | release_pte_pages(pte, _pte); |
7d2eba05 EA |
2342 | trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero, |
2343 | referenced, writable, result); | |
344aa35c | 2344 | return 0; |
ba76149f AA |
2345 | } |
2346 | ||
2347 | static void __collapse_huge_page_copy(pte_t *pte, struct page *page, | |
2348 | struct vm_area_struct *vma, | |
2349 | unsigned long address, | |
2350 | spinlock_t *ptl) | |
2351 | { | |
2352 | pte_t *_pte; | |
2353 | for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) { | |
2354 | pte_t pteval = *_pte; | |
2355 | struct page *src_page; | |
2356 | ||
ca0984ca | 2357 | if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { |
ba76149f AA |
2358 | clear_user_highpage(page, address); |
2359 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); | |
ca0984ca EA |
2360 | if (is_zero_pfn(pte_pfn(pteval))) { |
2361 | /* | |
2362 | * ptl mostly unnecessary. | |
2363 | */ | |
2364 | spin_lock(ptl); | |
2365 | /* | |
2366 | * paravirt calls inside pte_clear here are | |
2367 | * superfluous. | |
2368 | */ | |
2369 | pte_clear(vma->vm_mm, address, _pte); | |
2370 | spin_unlock(ptl); | |
2371 | } | |
ba76149f AA |
2372 | } else { |
2373 | src_page = pte_page(pteval); | |
2374 | copy_user_highpage(page, src_page, address, vma); | |
309381fe | 2375 | VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page); |
ba76149f AA |
2376 | release_pte_page(src_page); |
2377 | /* | |
2378 | * ptl mostly unnecessary, but preempt has to | |
2379 | * be disabled to update the per-cpu stats | |
2380 | * inside page_remove_rmap(). | |
2381 | */ | |
2382 | spin_lock(ptl); | |
2383 | /* | |
2384 | * paravirt calls inside pte_clear here are | |
2385 | * superfluous. | |
2386 | */ | |
2387 | pte_clear(vma->vm_mm, address, _pte); | |
d281ee61 | 2388 | page_remove_rmap(src_page, false); |
ba76149f AA |
2389 | spin_unlock(ptl); |
2390 | free_page_and_swap_cache(src_page); | |
2391 | } | |
2392 | ||
2393 | address += PAGE_SIZE; | |
2394 | page++; | |
2395 | } | |
2396 | } | |
2397 | ||
26234f36 | 2398 | static void khugepaged_alloc_sleep(void) |
ba76149f | 2399 | { |
bde43c6c PM |
2400 | DEFINE_WAIT(wait); |
2401 | ||
2402 | add_wait_queue(&khugepaged_wait, &wait); | |
2403 | freezable_schedule_timeout_interruptible( | |
2404 | msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); | |
2405 | remove_wait_queue(&khugepaged_wait, &wait); | |
26234f36 | 2406 | } |
ba76149f | 2407 | |
9f1b868a BL |
2408 | static int khugepaged_node_load[MAX_NUMNODES]; |
2409 | ||
14a4e214 DR |
2410 | static bool khugepaged_scan_abort(int nid) |
2411 | { | |
2412 | int i; | |
2413 | ||
2414 | /* | |
2415 | * If zone_reclaim_mode is disabled, then no extra effort is made to | |
2416 | * allocate memory locally. | |
2417 | */ | |
2418 | if (!zone_reclaim_mode) | |
2419 | return false; | |
2420 | ||
2421 | /* If there is a count for this node already, it must be acceptable */ | |
2422 | if (khugepaged_node_load[nid]) | |
2423 | return false; | |
2424 | ||
2425 | for (i = 0; i < MAX_NUMNODES; i++) { | |
2426 | if (!khugepaged_node_load[i]) | |
2427 | continue; | |
2428 | if (node_distance(nid, i) > RECLAIM_DISTANCE) | |
2429 | return true; | |
2430 | } | |
2431 | return false; | |
2432 | } | |
2433 | ||
26234f36 | 2434 | #ifdef CONFIG_NUMA |
9f1b868a BL |
2435 | static int khugepaged_find_target_node(void) |
2436 | { | |
2437 | static int last_khugepaged_target_node = NUMA_NO_NODE; | |
2438 | int nid, target_node = 0, max_value = 0; | |
2439 | ||
2440 | /* find first node with max normal pages hit */ | |
2441 | for (nid = 0; nid < MAX_NUMNODES; nid++) | |
2442 | if (khugepaged_node_load[nid] > max_value) { | |
2443 | max_value = khugepaged_node_load[nid]; | |
2444 | target_node = nid; | |
2445 | } | |
2446 | ||
2447 | /* do some balance if several nodes have the same hit record */ | |
2448 | if (target_node <= last_khugepaged_target_node) | |
2449 | for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES; | |
2450 | nid++) | |
2451 | if (max_value == khugepaged_node_load[nid]) { | |
2452 | target_node = nid; | |
2453 | break; | |
2454 | } | |
2455 | ||
2456 | last_khugepaged_target_node = target_node; | |
2457 | return target_node; | |
2458 | } | |
2459 | ||
26234f36 XG |
2460 | static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) |
2461 | { | |
2462 | if (IS_ERR(*hpage)) { | |
2463 | if (!*wait) | |
2464 | return false; | |
2465 | ||
2466 | *wait = false; | |
e3b4126c | 2467 | *hpage = NULL; |
26234f36 XG |
2468 | khugepaged_alloc_sleep(); |
2469 | } else if (*hpage) { | |
2470 | put_page(*hpage); | |
2471 | *hpage = NULL; | |
2472 | } | |
2473 | ||
2474 | return true; | |
2475 | } | |
2476 | ||
3b363692 MH |
2477 | static struct page * |
2478 | khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm, | |
d6669d68 | 2479 | unsigned long address, int node) |
26234f36 | 2480 | { |
309381fe | 2481 | VM_BUG_ON_PAGE(*hpage, *hpage); |
8b164568 | 2482 | |
ce83d217 | 2483 | /* |
8b164568 VB |
2484 | * Before allocating the hugepage, release the mmap_sem read lock. |
2485 | * The allocation can take potentially a long time if it involves | |
2486 | * sync compaction, and we do not need to hold the mmap_sem during | |
2487 | * that. We will recheck the vma after taking it again in write mode. | |
ce83d217 | 2488 | */ |
8b164568 VB |
2489 | up_read(&mm->mmap_sem); |
2490 | ||
96db800f | 2491 | *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER); |
26234f36 | 2492 | if (unlikely(!*hpage)) { |
81ab4201 | 2493 | count_vm_event(THP_COLLAPSE_ALLOC_FAILED); |
ce83d217 | 2494 | *hpage = ERR_PTR(-ENOMEM); |
26234f36 | 2495 | return NULL; |
ce83d217 | 2496 | } |
26234f36 | 2497 | |
65b3c07b | 2498 | count_vm_event(THP_COLLAPSE_ALLOC); |
26234f36 XG |
2499 | return *hpage; |
2500 | } | |
2501 | #else | |
9f1b868a BL |
2502 | static int khugepaged_find_target_node(void) |
2503 | { | |
2504 | return 0; | |
2505 | } | |
2506 | ||
10dc4155 BL |
2507 | static inline struct page *alloc_hugepage(int defrag) |
2508 | { | |
2509 | return alloc_pages(alloc_hugepage_gfpmask(defrag, 0), | |
2510 | HPAGE_PMD_ORDER); | |
2511 | } | |
2512 | ||
26234f36 XG |
2513 | static struct page *khugepaged_alloc_hugepage(bool *wait) |
2514 | { | |
2515 | struct page *hpage; | |
2516 | ||
2517 | do { | |
2518 | hpage = alloc_hugepage(khugepaged_defrag()); | |
2519 | if (!hpage) { | |
2520 | count_vm_event(THP_COLLAPSE_ALLOC_FAILED); | |
2521 | if (!*wait) | |
2522 | return NULL; | |
2523 | ||
2524 | *wait = false; | |
2525 | khugepaged_alloc_sleep(); | |
2526 | } else | |
2527 | count_vm_event(THP_COLLAPSE_ALLOC); | |
2528 | } while (unlikely(!hpage) && likely(khugepaged_enabled())); | |
2529 | ||
2530 | return hpage; | |
2531 | } | |
2532 | ||
2533 | static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) | |
2534 | { | |
2535 | if (!*hpage) | |
2536 | *hpage = khugepaged_alloc_hugepage(wait); | |
2537 | ||
2538 | if (unlikely(!*hpage)) | |
2539 | return false; | |
2540 | ||
2541 | return true; | |
2542 | } | |
2543 | ||
3b363692 MH |
2544 | static struct page * |
2545 | khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm, | |
d6669d68 | 2546 | unsigned long address, int node) |
26234f36 XG |
2547 | { |
2548 | up_read(&mm->mmap_sem); | |
2549 | VM_BUG_ON(!*hpage); | |
3b363692 | 2550 | |
26234f36 XG |
2551 | return *hpage; |
2552 | } | |
692e0b35 AA |
2553 | #endif |
2554 | ||
fa475e51 BL |
2555 | static bool hugepage_vma_check(struct vm_area_struct *vma) |
2556 | { | |
2557 | if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) || | |
2558 | (vma->vm_flags & VM_NOHUGEPAGE)) | |
2559 | return false; | |
7479df6d KS |
2560 | if (vma->vm_flags & VM_LOCKED) |
2561 | return false; | |
fa475e51 BL |
2562 | if (!vma->anon_vma || vma->vm_ops) |
2563 | return false; | |
2564 | if (is_vma_temporary_stack(vma)) | |
2565 | return false; | |
81d1b09c | 2566 | VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma); |
fa475e51 BL |
2567 | return true; |
2568 | } | |
2569 | ||
26234f36 XG |
2570 | static void collapse_huge_page(struct mm_struct *mm, |
2571 | unsigned long address, | |
2572 | struct page **hpage, | |
2573 | struct vm_area_struct *vma, | |
2574 | int node) | |
2575 | { | |
26234f36 XG |
2576 | pmd_t *pmd, _pmd; |
2577 | pte_t *pte; | |
2578 | pgtable_t pgtable; | |
2579 | struct page *new_page; | |
c4088ebd | 2580 | spinlock_t *pmd_ptl, *pte_ptl; |
7d2eba05 | 2581 | int isolated, result = 0; |
26234f36 | 2582 | unsigned long hstart, hend; |
00501b53 | 2583 | struct mem_cgroup *memcg; |
2ec74c3e SG |
2584 | unsigned long mmun_start; /* For mmu_notifiers */ |
2585 | unsigned long mmun_end; /* For mmu_notifiers */ | |
3b363692 | 2586 | gfp_t gfp; |
26234f36 XG |
2587 | |
2588 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
2589 | ||
3b363692 MH |
2590 | /* Only allocate from the target node */ |
2591 | gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) | | |
2592 | __GFP_THISNODE; | |
2593 | ||
26234f36 | 2594 | /* release the mmap_sem read lock. */ |
d6669d68 | 2595 | new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node); |
7d2eba05 EA |
2596 | if (!new_page) { |
2597 | result = SCAN_ALLOC_HUGE_PAGE_FAIL; | |
2598 | goto out_nolock; | |
2599 | } | |
26234f36 | 2600 | |
f627c2f5 | 2601 | if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) { |
7d2eba05 EA |
2602 | result = SCAN_CGROUP_CHARGE_FAIL; |
2603 | goto out_nolock; | |
2604 | } | |
ba76149f AA |
2605 | |
2606 | /* | |
2607 | * Prevent all access to pagetables with the exception of | |
2608 | * gup_fast later hanlded by the ptep_clear_flush and the VM | |
2609 | * handled by the anon_vma lock + PG_lock. | |
2610 | */ | |
2611 | down_write(&mm->mmap_sem); | |
7d2eba05 EA |
2612 | if (unlikely(khugepaged_test_exit(mm))) { |
2613 | result = SCAN_ANY_PROCESS; | |
ba76149f | 2614 | goto out; |
7d2eba05 | 2615 | } |
ba76149f AA |
2616 | |
2617 | vma = find_vma(mm, address); | |
7d2eba05 EA |
2618 | if (!vma) { |
2619 | result = SCAN_VMA_NULL; | |
a8f531eb | 2620 | goto out; |
7d2eba05 | 2621 | } |
ba76149f AA |
2622 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; |
2623 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
7d2eba05 EA |
2624 | if (address < hstart || address + HPAGE_PMD_SIZE > hend) { |
2625 | result = SCAN_ADDRESS_RANGE; | |
ba76149f | 2626 | goto out; |
7d2eba05 EA |
2627 | } |
2628 | if (!hugepage_vma_check(vma)) { | |
2629 | result = SCAN_VMA_CHECK; | |
a7d6e4ec | 2630 | goto out; |
7d2eba05 | 2631 | } |
6219049a | 2632 | pmd = mm_find_pmd(mm, address); |
7d2eba05 EA |
2633 | if (!pmd) { |
2634 | result = SCAN_PMD_NULL; | |
ba76149f | 2635 | goto out; |
7d2eba05 | 2636 | } |
ba76149f | 2637 | |
4fc3f1d6 | 2638 | anon_vma_lock_write(vma->anon_vma); |
ba76149f AA |
2639 | |
2640 | pte = pte_offset_map(pmd, address); | |
c4088ebd | 2641 | pte_ptl = pte_lockptr(mm, pmd); |
ba76149f | 2642 | |
2ec74c3e SG |
2643 | mmun_start = address; |
2644 | mmun_end = address + HPAGE_PMD_SIZE; | |
2645 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
c4088ebd | 2646 | pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ |
ba76149f AA |
2647 | /* |
2648 | * After this gup_fast can't run anymore. This also removes | |
2649 | * any huge TLB entry from the CPU so we won't allow | |
2650 | * huge and small TLB entries for the same virtual address | |
2651 | * to avoid the risk of CPU bugs in that area. | |
2652 | */ | |
15a25b2e | 2653 | _pmd = pmdp_collapse_flush(vma, address, pmd); |
c4088ebd | 2654 | spin_unlock(pmd_ptl); |
2ec74c3e | 2655 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
ba76149f | 2656 | |
c4088ebd | 2657 | spin_lock(pte_ptl); |
ba76149f | 2658 | isolated = __collapse_huge_page_isolate(vma, address, pte); |
c4088ebd | 2659 | spin_unlock(pte_ptl); |
ba76149f AA |
2660 | |
2661 | if (unlikely(!isolated)) { | |
453c7192 | 2662 | pte_unmap(pte); |
c4088ebd | 2663 | spin_lock(pmd_ptl); |
ba76149f | 2664 | BUG_ON(!pmd_none(*pmd)); |
7c342512 AK |
2665 | /* |
2666 | * We can only use set_pmd_at when establishing | |
2667 | * hugepmds and never for establishing regular pmds that | |
2668 | * points to regular pagetables. Use pmd_populate for that | |
2669 | */ | |
2670 | pmd_populate(mm, pmd, pmd_pgtable(_pmd)); | |
c4088ebd | 2671 | spin_unlock(pmd_ptl); |
08b52706 | 2672 | anon_vma_unlock_write(vma->anon_vma); |
7d2eba05 | 2673 | result = SCAN_FAIL; |
ce83d217 | 2674 | goto out; |
ba76149f AA |
2675 | } |
2676 | ||
2677 | /* | |
2678 | * All pages are isolated and locked so anon_vma rmap | |
2679 | * can't run anymore. | |
2680 | */ | |
08b52706 | 2681 | anon_vma_unlock_write(vma->anon_vma); |
ba76149f | 2682 | |
c4088ebd | 2683 | __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl); |
453c7192 | 2684 | pte_unmap(pte); |
ba76149f AA |
2685 | __SetPageUptodate(new_page); |
2686 | pgtable = pmd_pgtable(_pmd); | |
ba76149f | 2687 | |
3122359a KS |
2688 | _pmd = mk_huge_pmd(new_page, vma->vm_page_prot); |
2689 | _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); | |
ba76149f AA |
2690 | |
2691 | /* | |
2692 | * spin_lock() below is not the equivalent of smp_wmb(), so | |
2693 | * this is needed to avoid the copy_huge_page writes to become | |
2694 | * visible after the set_pmd_at() write. | |
2695 | */ | |
2696 | smp_wmb(); | |
2697 | ||
c4088ebd | 2698 | spin_lock(pmd_ptl); |
ba76149f | 2699 | BUG_ON(!pmd_none(*pmd)); |
d281ee61 | 2700 | page_add_new_anon_rmap(new_page, vma, address, true); |
f627c2f5 | 2701 | mem_cgroup_commit_charge(new_page, memcg, false, true); |
00501b53 | 2702 | lru_cache_add_active_or_unevictable(new_page, vma); |
fce144b4 | 2703 | pgtable_trans_huge_deposit(mm, pmd, pgtable); |
ba76149f | 2704 | set_pmd_at(mm, address, pmd, _pmd); |
b113da65 | 2705 | update_mmu_cache_pmd(vma, address, pmd); |
c4088ebd | 2706 | spin_unlock(pmd_ptl); |
ba76149f AA |
2707 | |
2708 | *hpage = NULL; | |
420256ef | 2709 | |
ba76149f | 2710 | khugepaged_pages_collapsed++; |
7d2eba05 | 2711 | result = SCAN_SUCCEED; |
ce83d217 | 2712 | out_up_write: |
ba76149f | 2713 | up_write(&mm->mmap_sem); |
7d2eba05 | 2714 | trace_mm_collapse_huge_page(mm, isolated, result); |
0bbbc0b3 AA |
2715 | return; |
2716 | ||
7d2eba05 EA |
2717 | out_nolock: |
2718 | trace_mm_collapse_huge_page(mm, isolated, result); | |
2719 | return; | |
ce83d217 | 2720 | out: |
f627c2f5 | 2721 | mem_cgroup_cancel_charge(new_page, memcg, true); |
ce83d217 | 2722 | goto out_up_write; |
ba76149f AA |
2723 | } |
2724 | ||
2725 | static int khugepaged_scan_pmd(struct mm_struct *mm, | |
2726 | struct vm_area_struct *vma, | |
2727 | unsigned long address, | |
2728 | struct page **hpage) | |
2729 | { | |
ba76149f AA |
2730 | pmd_t *pmd; |
2731 | pte_t *pte, *_pte; | |
7d2eba05 EA |
2732 | int ret = 0, none_or_zero = 0, result = 0; |
2733 | struct page *page = NULL; | |
ba76149f AA |
2734 | unsigned long _address; |
2735 | spinlock_t *ptl; | |
00ef2d2f | 2736 | int node = NUMA_NO_NODE; |
10359213 | 2737 | bool writable = false, referenced = false; |
ba76149f AA |
2738 | |
2739 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
2740 | ||
6219049a | 2741 | pmd = mm_find_pmd(mm, address); |
7d2eba05 EA |
2742 | if (!pmd) { |
2743 | result = SCAN_PMD_NULL; | |
ba76149f | 2744 | goto out; |
7d2eba05 | 2745 | } |
ba76149f | 2746 | |
9f1b868a | 2747 | memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load)); |
ba76149f AA |
2748 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); |
2749 | for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR; | |
2750 | _pte++, _address += PAGE_SIZE) { | |
2751 | pte_t pteval = *_pte; | |
ca0984ca | 2752 | if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { |
c1294d05 | 2753 | if (!userfaultfd_armed(vma) && |
7d2eba05 | 2754 | ++none_or_zero <= khugepaged_max_ptes_none) { |
ba76149f | 2755 | continue; |
7d2eba05 EA |
2756 | } else { |
2757 | result = SCAN_EXCEED_NONE_PTE; | |
ba76149f | 2758 | goto out_unmap; |
7d2eba05 | 2759 | } |
ba76149f | 2760 | } |
7d2eba05 EA |
2761 | if (!pte_present(pteval)) { |
2762 | result = SCAN_PTE_NON_PRESENT; | |
ba76149f | 2763 | goto out_unmap; |
7d2eba05 | 2764 | } |
10359213 EA |
2765 | if (pte_write(pteval)) |
2766 | writable = true; | |
2767 | ||
ba76149f | 2768 | page = vm_normal_page(vma, _address, pteval); |
7d2eba05 EA |
2769 | if (unlikely(!page)) { |
2770 | result = SCAN_PAGE_NULL; | |
ba76149f | 2771 | goto out_unmap; |
7d2eba05 | 2772 | } |
5c4b4be3 | 2773 | /* |
9f1b868a BL |
2774 | * Record which node the original page is from and save this |
2775 | * information to khugepaged_node_load[]. | |
2776 | * Khupaged will allocate hugepage from the node has the max | |
2777 | * hit record. | |
5c4b4be3 | 2778 | */ |
9f1b868a | 2779 | node = page_to_nid(page); |
7d2eba05 EA |
2780 | if (khugepaged_scan_abort(node)) { |
2781 | result = SCAN_SCAN_ABORT; | |
14a4e214 | 2782 | goto out_unmap; |
7d2eba05 | 2783 | } |
9f1b868a | 2784 | khugepaged_node_load[node]++; |
309381fe | 2785 | VM_BUG_ON_PAGE(PageCompound(page), page); |
7d2eba05 EA |
2786 | if (!PageLRU(page)) { |
2787 | result = SCAN_SCAN_ABORT; | |
2788 | goto out_unmap; | |
2789 | } | |
2790 | if (PageLocked(page)) { | |
2791 | result = SCAN_PAGE_LOCK; | |
ba76149f | 2792 | goto out_unmap; |
7d2eba05 EA |
2793 | } |
2794 | if (!PageAnon(page)) { | |
2795 | result = SCAN_PAGE_ANON; | |
2796 | goto out_unmap; | |
2797 | } | |
2798 | ||
10359213 EA |
2799 | /* |
2800 | * cannot use mapcount: can't collapse if there's a gup pin. | |
2801 | * The page must only be referenced by the scanned process | |
2802 | * and page swap cache. | |
2803 | */ | |
7d2eba05 EA |
2804 | if (page_count(page) != 1 + !!PageSwapCache(page)) { |
2805 | result = SCAN_PAGE_COUNT; | |
ba76149f | 2806 | goto out_unmap; |
7d2eba05 | 2807 | } |
33c3fc71 VD |
2808 | if (pte_young(pteval) || |
2809 | page_is_young(page) || PageReferenced(page) || | |
8ee53820 | 2810 | mmu_notifier_test_young(vma->vm_mm, address)) |
10359213 | 2811 | referenced = true; |
ba76149f | 2812 | } |
7d2eba05 EA |
2813 | if (writable) { |
2814 | if (referenced) { | |
2815 | result = SCAN_SUCCEED; | |
2816 | ret = 1; | |
2817 | } else { | |
2818 | result = SCAN_NO_REFERENCED_PAGE; | |
2819 | } | |
2820 | } else { | |
2821 | result = SCAN_PAGE_RO; | |
2822 | } | |
ba76149f AA |
2823 | out_unmap: |
2824 | pte_unmap_unlock(pte, ptl); | |
9f1b868a BL |
2825 | if (ret) { |
2826 | node = khugepaged_find_target_node(); | |
ce83d217 | 2827 | /* collapse_huge_page will return with the mmap_sem released */ |
5c4b4be3 | 2828 | collapse_huge_page(mm, address, hpage, vma, node); |
9f1b868a | 2829 | } |
ba76149f | 2830 | out: |
7d2eba05 EA |
2831 | trace_mm_khugepaged_scan_pmd(mm, page_to_pfn(page), writable, referenced, |
2832 | none_or_zero, result); | |
ba76149f AA |
2833 | return ret; |
2834 | } | |
2835 | ||
2836 | static void collect_mm_slot(struct mm_slot *mm_slot) | |
2837 | { | |
2838 | struct mm_struct *mm = mm_slot->mm; | |
2839 | ||
b9980cdc | 2840 | VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); |
ba76149f AA |
2841 | |
2842 | if (khugepaged_test_exit(mm)) { | |
2843 | /* free mm_slot */ | |
43b5fbbd | 2844 | hash_del(&mm_slot->hash); |
ba76149f AA |
2845 | list_del(&mm_slot->mm_node); |
2846 | ||
2847 | /* | |
2848 | * Not strictly needed because the mm exited already. | |
2849 | * | |
2850 | * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); | |
2851 | */ | |
2852 | ||
2853 | /* khugepaged_mm_lock actually not necessary for the below */ | |
2854 | free_mm_slot(mm_slot); | |
2855 | mmdrop(mm); | |
2856 | } | |
2857 | } | |
2858 | ||
2859 | static unsigned int khugepaged_scan_mm_slot(unsigned int pages, | |
2860 | struct page **hpage) | |
2f1da642 HS |
2861 | __releases(&khugepaged_mm_lock) |
2862 | __acquires(&khugepaged_mm_lock) | |
ba76149f AA |
2863 | { |
2864 | struct mm_slot *mm_slot; | |
2865 | struct mm_struct *mm; | |
2866 | struct vm_area_struct *vma; | |
2867 | int progress = 0; | |
2868 | ||
2869 | VM_BUG_ON(!pages); | |
b9980cdc | 2870 | VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); |
ba76149f AA |
2871 | |
2872 | if (khugepaged_scan.mm_slot) | |
2873 | mm_slot = khugepaged_scan.mm_slot; | |
2874 | else { | |
2875 | mm_slot = list_entry(khugepaged_scan.mm_head.next, | |
2876 | struct mm_slot, mm_node); | |
2877 | khugepaged_scan.address = 0; | |
2878 | khugepaged_scan.mm_slot = mm_slot; | |
2879 | } | |
2880 | spin_unlock(&khugepaged_mm_lock); | |
2881 | ||
2882 | mm = mm_slot->mm; | |
2883 | down_read(&mm->mmap_sem); | |
2884 | if (unlikely(khugepaged_test_exit(mm))) | |
2885 | vma = NULL; | |
2886 | else | |
2887 | vma = find_vma(mm, khugepaged_scan.address); | |
2888 | ||
2889 | progress++; | |
2890 | for (; vma; vma = vma->vm_next) { | |
2891 | unsigned long hstart, hend; | |
2892 | ||
2893 | cond_resched(); | |
2894 | if (unlikely(khugepaged_test_exit(mm))) { | |
2895 | progress++; | |
2896 | break; | |
2897 | } | |
fa475e51 BL |
2898 | if (!hugepage_vma_check(vma)) { |
2899 | skip: | |
ba76149f AA |
2900 | progress++; |
2901 | continue; | |
2902 | } | |
ba76149f AA |
2903 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; |
2904 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
a7d6e4ec AA |
2905 | if (hstart >= hend) |
2906 | goto skip; | |
2907 | if (khugepaged_scan.address > hend) | |
2908 | goto skip; | |
ba76149f AA |
2909 | if (khugepaged_scan.address < hstart) |
2910 | khugepaged_scan.address = hstart; | |
a7d6e4ec | 2911 | VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); |
ba76149f AA |
2912 | |
2913 | while (khugepaged_scan.address < hend) { | |
2914 | int ret; | |
2915 | cond_resched(); | |
2916 | if (unlikely(khugepaged_test_exit(mm))) | |
2917 | goto breakouterloop; | |
2918 | ||
2919 | VM_BUG_ON(khugepaged_scan.address < hstart || | |
2920 | khugepaged_scan.address + HPAGE_PMD_SIZE > | |
2921 | hend); | |
2922 | ret = khugepaged_scan_pmd(mm, vma, | |
2923 | khugepaged_scan.address, | |
2924 | hpage); | |
2925 | /* move to next address */ | |
2926 | khugepaged_scan.address += HPAGE_PMD_SIZE; | |
2927 | progress += HPAGE_PMD_NR; | |
2928 | if (ret) | |
2929 | /* we released mmap_sem so break loop */ | |
2930 | goto breakouterloop_mmap_sem; | |
2931 | if (progress >= pages) | |
2932 | goto breakouterloop; | |
2933 | } | |
2934 | } | |
2935 | breakouterloop: | |
2936 | up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */ | |
2937 | breakouterloop_mmap_sem: | |
2938 | ||
2939 | spin_lock(&khugepaged_mm_lock); | |
a7d6e4ec | 2940 | VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); |
ba76149f AA |
2941 | /* |
2942 | * Release the current mm_slot if this mm is about to die, or | |
2943 | * if we scanned all vmas of this mm. | |
2944 | */ | |
2945 | if (khugepaged_test_exit(mm) || !vma) { | |
2946 | /* | |
2947 | * Make sure that if mm_users is reaching zero while | |
2948 | * khugepaged runs here, khugepaged_exit will find | |
2949 | * mm_slot not pointing to the exiting mm. | |
2950 | */ | |
2951 | if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) { | |
2952 | khugepaged_scan.mm_slot = list_entry( | |
2953 | mm_slot->mm_node.next, | |
2954 | struct mm_slot, mm_node); | |
2955 | khugepaged_scan.address = 0; | |
2956 | } else { | |
2957 | khugepaged_scan.mm_slot = NULL; | |
2958 | khugepaged_full_scans++; | |
2959 | } | |
2960 | ||
2961 | collect_mm_slot(mm_slot); | |
2962 | } | |
2963 | ||
2964 | return progress; | |
2965 | } | |
2966 | ||
2967 | static int khugepaged_has_work(void) | |
2968 | { | |
2969 | return !list_empty(&khugepaged_scan.mm_head) && | |
2970 | khugepaged_enabled(); | |
2971 | } | |
2972 | ||
2973 | static int khugepaged_wait_event(void) | |
2974 | { | |
2975 | return !list_empty(&khugepaged_scan.mm_head) || | |
2017c0bf | 2976 | kthread_should_stop(); |
ba76149f AA |
2977 | } |
2978 | ||
d516904b | 2979 | static void khugepaged_do_scan(void) |
ba76149f | 2980 | { |
d516904b | 2981 | struct page *hpage = NULL; |
ba76149f AA |
2982 | unsigned int progress = 0, pass_through_head = 0; |
2983 | unsigned int pages = khugepaged_pages_to_scan; | |
d516904b | 2984 | bool wait = true; |
ba76149f AA |
2985 | |
2986 | barrier(); /* write khugepaged_pages_to_scan to local stack */ | |
2987 | ||
2988 | while (progress < pages) { | |
26234f36 | 2989 | if (!khugepaged_prealloc_page(&hpage, &wait)) |
d516904b | 2990 | break; |
26234f36 | 2991 | |
420256ef | 2992 | cond_resched(); |
ba76149f | 2993 | |
cd092411 | 2994 | if (unlikely(kthread_should_stop() || try_to_freeze())) |
878aee7d AA |
2995 | break; |
2996 | ||
ba76149f AA |
2997 | spin_lock(&khugepaged_mm_lock); |
2998 | if (!khugepaged_scan.mm_slot) | |
2999 | pass_through_head++; | |
3000 | if (khugepaged_has_work() && | |
3001 | pass_through_head < 2) | |
3002 | progress += khugepaged_scan_mm_slot(pages - progress, | |
d516904b | 3003 | &hpage); |
ba76149f AA |
3004 | else |
3005 | progress = pages; | |
3006 | spin_unlock(&khugepaged_mm_lock); | |
3007 | } | |
ba76149f | 3008 | |
d516904b XG |
3009 | if (!IS_ERR_OR_NULL(hpage)) |
3010 | put_page(hpage); | |
0bbbc0b3 AA |
3011 | } |
3012 | ||
2017c0bf XG |
3013 | static void khugepaged_wait_work(void) |
3014 | { | |
2017c0bf XG |
3015 | if (khugepaged_has_work()) { |
3016 | if (!khugepaged_scan_sleep_millisecs) | |
3017 | return; | |
3018 | ||
3019 | wait_event_freezable_timeout(khugepaged_wait, | |
3020 | kthread_should_stop(), | |
3021 | msecs_to_jiffies(khugepaged_scan_sleep_millisecs)); | |
3022 | return; | |
3023 | } | |
3024 | ||
3025 | if (khugepaged_enabled()) | |
3026 | wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); | |
3027 | } | |
3028 | ||
ba76149f AA |
3029 | static int khugepaged(void *none) |
3030 | { | |
3031 | struct mm_slot *mm_slot; | |
3032 | ||
878aee7d | 3033 | set_freezable(); |
8698a745 | 3034 | set_user_nice(current, MAX_NICE); |
ba76149f | 3035 | |
b7231789 XG |
3036 | while (!kthread_should_stop()) { |
3037 | khugepaged_do_scan(); | |
3038 | khugepaged_wait_work(); | |
3039 | } | |
ba76149f AA |
3040 | |
3041 | spin_lock(&khugepaged_mm_lock); | |
3042 | mm_slot = khugepaged_scan.mm_slot; | |
3043 | khugepaged_scan.mm_slot = NULL; | |
3044 | if (mm_slot) | |
3045 | collect_mm_slot(mm_slot); | |
3046 | spin_unlock(&khugepaged_mm_lock); | |
ba76149f AA |
3047 | return 0; |
3048 | } | |
3049 | ||
c5a647d0 KS |
3050 | static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, |
3051 | unsigned long haddr, pmd_t *pmd) | |
3052 | { | |
3053 | struct mm_struct *mm = vma->vm_mm; | |
3054 | pgtable_t pgtable; | |
3055 | pmd_t _pmd; | |
3056 | int i; | |
3057 | ||
8809aa2d | 3058 | pmdp_huge_clear_flush_notify(vma, haddr, pmd); |
c5a647d0 KS |
3059 | /* leave pmd empty until pte is filled */ |
3060 | ||
6b0b50b0 | 3061 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); |
c5a647d0 KS |
3062 | pmd_populate(mm, &_pmd, pgtable); |
3063 | ||
3064 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
3065 | pte_t *pte, entry; | |
3066 | entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); | |
3067 | entry = pte_mkspecial(entry); | |
3068 | pte = pte_offset_map(&_pmd, haddr); | |
3069 | VM_BUG_ON(!pte_none(*pte)); | |
3070 | set_pte_at(mm, haddr, pte, entry); | |
3071 | pte_unmap(pte); | |
3072 | } | |
3073 | smp_wmb(); /* make pte visible before pmd */ | |
3074 | pmd_populate(mm, pmd, pgtable); | |
97ae1749 | 3075 | put_huge_zero_page(); |
c5a647d0 KS |
3076 | } |
3077 | ||
e180377f KS |
3078 | void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address, |
3079 | pmd_t *pmd) | |
71e3aac0 | 3080 | { |
c4088ebd | 3081 | spinlock_t *ptl; |
4897c765 | 3082 | struct page *page = NULL; |
e180377f | 3083 | struct mm_struct *mm = vma->vm_mm; |
c5a647d0 KS |
3084 | unsigned long haddr = address & HPAGE_PMD_MASK; |
3085 | unsigned long mmun_start; /* For mmu_notifiers */ | |
3086 | unsigned long mmun_end; /* For mmu_notifiers */ | |
e180377f KS |
3087 | |
3088 | BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE); | |
71e3aac0 | 3089 | |
c5a647d0 KS |
3090 | mmun_start = haddr; |
3091 | mmun_end = haddr + HPAGE_PMD_SIZE; | |
750e8165 | 3092 | again: |
c5a647d0 | 3093 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
c4088ebd | 3094 | ptl = pmd_lock(mm, pmd); |
4897c765 MW |
3095 | if (unlikely(!pmd_trans_huge(*pmd))) |
3096 | goto unlock; | |
3097 | if (vma_is_dax(vma)) { | |
5b701b84 KS |
3098 | pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); |
3099 | if (is_huge_zero_pmd(_pmd)) | |
3100 | put_huge_zero_page(); | |
4897c765 | 3101 | } else if (is_huge_zero_pmd(*pmd)) { |
c5a647d0 | 3102 | __split_huge_zero_page_pmd(vma, haddr, pmd); |
4897c765 MW |
3103 | } else { |
3104 | page = pmd_page(*pmd); | |
3105 | VM_BUG_ON_PAGE(!page_count(page), page); | |
3106 | get_page(page); | |
71e3aac0 | 3107 | } |
4897c765 | 3108 | unlock: |
c4088ebd | 3109 | spin_unlock(ptl); |
c5a647d0 | 3110 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
71e3aac0 | 3111 | |
4897c765 MW |
3112 | if (!page) |
3113 | return; | |
71e3aac0 | 3114 | |
4897c765 | 3115 | split_huge_page(page); |
71e3aac0 | 3116 | put_page(page); |
750e8165 HD |
3117 | |
3118 | /* | |
3119 | * We don't always have down_write of mmap_sem here: a racing | |
3120 | * do_huge_pmd_wp_page() might have copied-on-write to another | |
3121 | * huge page before our split_huge_page() got the anon_vma lock. | |
3122 | */ | |
3123 | if (unlikely(pmd_trans_huge(*pmd))) | |
3124 | goto again; | |
71e3aac0 | 3125 | } |
94fcc585 | 3126 | |
e180377f KS |
3127 | void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address, |
3128 | pmd_t *pmd) | |
3129 | { | |
3130 | struct vm_area_struct *vma; | |
3131 | ||
3132 | vma = find_vma(mm, address); | |
3133 | BUG_ON(vma == NULL); | |
3134 | split_huge_page_pmd(vma, address, pmd); | |
3135 | } | |
3136 | ||
94fcc585 AA |
3137 | static void split_huge_page_address(struct mm_struct *mm, |
3138 | unsigned long address) | |
3139 | { | |
f72e7dcd HD |
3140 | pgd_t *pgd; |
3141 | pud_t *pud; | |
94fcc585 AA |
3142 | pmd_t *pmd; |
3143 | ||
3144 | VM_BUG_ON(!(address & ~HPAGE_PMD_MASK)); | |
3145 | ||
f72e7dcd HD |
3146 | pgd = pgd_offset(mm, address); |
3147 | if (!pgd_present(*pgd)) | |
3148 | return; | |
3149 | ||
3150 | pud = pud_offset(pgd, address); | |
3151 | if (!pud_present(*pud)) | |
3152 | return; | |
3153 | ||
3154 | pmd = pmd_offset(pud, address); | |
3155 | if (!pmd_present(*pmd)) | |
94fcc585 AA |
3156 | return; |
3157 | /* | |
3158 | * Caller holds the mmap_sem write mode, so a huge pmd cannot | |
3159 | * materialize from under us. | |
3160 | */ | |
e180377f | 3161 | split_huge_page_pmd_mm(mm, address, pmd); |
94fcc585 AA |
3162 | } |
3163 | ||
e1b9996b | 3164 | void vma_adjust_trans_huge(struct vm_area_struct *vma, |
94fcc585 AA |
3165 | unsigned long start, |
3166 | unsigned long end, | |
3167 | long adjust_next) | |
3168 | { | |
3169 | /* | |
3170 | * If the new start address isn't hpage aligned and it could | |
3171 | * previously contain an hugepage: check if we need to split | |
3172 | * an huge pmd. | |
3173 | */ | |
3174 | if (start & ~HPAGE_PMD_MASK && | |
3175 | (start & HPAGE_PMD_MASK) >= vma->vm_start && | |
3176 | (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
3177 | split_huge_page_address(vma->vm_mm, start); | |
3178 | ||
3179 | /* | |
3180 | * If the new end address isn't hpage aligned and it could | |
3181 | * previously contain an hugepage: check if we need to split | |
3182 | * an huge pmd. | |
3183 | */ | |
3184 | if (end & ~HPAGE_PMD_MASK && | |
3185 | (end & HPAGE_PMD_MASK) >= vma->vm_start && | |
3186 | (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
3187 | split_huge_page_address(vma->vm_mm, end); | |
3188 | ||
3189 | /* | |
3190 | * If we're also updating the vma->vm_next->vm_start, if the new | |
3191 | * vm_next->vm_start isn't page aligned and it could previously | |
3192 | * contain an hugepage: check if we need to split an huge pmd. | |
3193 | */ | |
3194 | if (adjust_next > 0) { | |
3195 | struct vm_area_struct *next = vma->vm_next; | |
3196 | unsigned long nstart = next->vm_start; | |
3197 | nstart += adjust_next << PAGE_SHIFT; | |
3198 | if (nstart & ~HPAGE_PMD_MASK && | |
3199 | (nstart & HPAGE_PMD_MASK) >= next->vm_start && | |
3200 | (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) | |
3201 | split_huge_page_address(next->vm_mm, nstart); | |
3202 | } | |
3203 | } |