]>
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> |
e9b61f19 | 19 | #include <linux/swapops.h> |
4897c765 | 20 | #include <linux/dax.h> |
ba76149f AA |
21 | #include <linux/kthread.h> |
22 | #include <linux/khugepaged.h> | |
878aee7d | 23 | #include <linux/freezer.h> |
a664b2d8 | 24 | #include <linux/mman.h> |
325adeb5 | 25 | #include <linux/pagemap.h> |
4daae3b4 | 26 | #include <linux/migrate.h> |
43b5fbbd | 27 | #include <linux/hashtable.h> |
6b251fc9 | 28 | #include <linux/userfaultfd_k.h> |
33c3fc71 | 29 | #include <linux/page_idle.h> |
97ae1749 | 30 | |
71e3aac0 AA |
31 | #include <asm/tlb.h> |
32 | #include <asm/pgalloc.h> | |
33 | #include "internal.h" | |
34 | ||
7d2eba05 EA |
35 | enum scan_result { |
36 | SCAN_FAIL, | |
37 | SCAN_SUCCEED, | |
38 | SCAN_PMD_NULL, | |
39 | SCAN_EXCEED_NONE_PTE, | |
40 | SCAN_PTE_NON_PRESENT, | |
41 | SCAN_PAGE_RO, | |
42 | SCAN_NO_REFERENCED_PAGE, | |
43 | SCAN_PAGE_NULL, | |
44 | SCAN_SCAN_ABORT, | |
45 | SCAN_PAGE_COUNT, | |
46 | SCAN_PAGE_LRU, | |
47 | SCAN_PAGE_LOCK, | |
48 | SCAN_PAGE_ANON, | |
b1caa957 | 49 | SCAN_PAGE_COMPOUND, |
7d2eba05 EA |
50 | SCAN_ANY_PROCESS, |
51 | SCAN_VMA_NULL, | |
52 | SCAN_VMA_CHECK, | |
53 | SCAN_ADDRESS_RANGE, | |
54 | SCAN_SWAP_CACHE_PAGE, | |
55 | SCAN_DEL_PAGE_LRU, | |
56 | SCAN_ALLOC_HUGE_PAGE_FAIL, | |
57 | SCAN_CGROUP_CHARGE_FAIL | |
58 | }; | |
59 | ||
60 | #define CREATE_TRACE_POINTS | |
61 | #include <trace/events/huge_memory.h> | |
62 | ||
ba76149f | 63 | /* |
8bfa3f9a JW |
64 | * By default transparent hugepage support is disabled in order that avoid |
65 | * to risk increase the memory footprint of applications without a guaranteed | |
66 | * benefit. When transparent hugepage support is enabled, is for all mappings, | |
67 | * and khugepaged scans all mappings. | |
68 | * Defrag is invoked by khugepaged hugepage allocations and by page faults | |
69 | * for all hugepage allocations. | |
ba76149f | 70 | */ |
71e3aac0 | 71 | unsigned long transparent_hugepage_flags __read_mostly = |
13ece886 | 72 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS |
ba76149f | 73 | (1<<TRANSPARENT_HUGEPAGE_FLAG)| |
13ece886 AA |
74 | #endif |
75 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE | |
76 | (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| | |
77 | #endif | |
d39d33c3 | 78 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)| |
79da5407 KS |
79 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| |
80 | (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
ba76149f AA |
81 | |
82 | /* default scan 8*512 pte (or vmas) every 30 second */ | |
83 | static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8; | |
84 | static unsigned int khugepaged_pages_collapsed; | |
85 | static unsigned int khugepaged_full_scans; | |
86 | static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; | |
87 | /* during fragmentation poll the hugepage allocator once every minute */ | |
88 | static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; | |
89 | static struct task_struct *khugepaged_thread __read_mostly; | |
90 | static DEFINE_MUTEX(khugepaged_mutex); | |
91 | static DEFINE_SPINLOCK(khugepaged_mm_lock); | |
92 | static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); | |
93 | /* | |
94 | * default collapse hugepages if there is at least one pte mapped like | |
95 | * it would have happened if the vma was large enough during page | |
96 | * fault. | |
97 | */ | |
98 | static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1; | |
99 | ||
100 | static int khugepaged(void *none); | |
ba76149f | 101 | static int khugepaged_slab_init(void); |
65ebb64f | 102 | static void khugepaged_slab_exit(void); |
ba76149f | 103 | |
43b5fbbd SL |
104 | #define MM_SLOTS_HASH_BITS 10 |
105 | static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); | |
106 | ||
ba76149f AA |
107 | static struct kmem_cache *mm_slot_cache __read_mostly; |
108 | ||
109 | /** | |
110 | * struct mm_slot - hash lookup from mm to mm_slot | |
111 | * @hash: hash collision list | |
112 | * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head | |
113 | * @mm: the mm that this information is valid for | |
114 | */ | |
115 | struct mm_slot { | |
116 | struct hlist_node hash; | |
117 | struct list_head mm_node; | |
118 | struct mm_struct *mm; | |
119 | }; | |
120 | ||
121 | /** | |
122 | * struct khugepaged_scan - cursor for scanning | |
123 | * @mm_head: the head of the mm list to scan | |
124 | * @mm_slot: the current mm_slot we are scanning | |
125 | * @address: the next address inside that to be scanned | |
126 | * | |
127 | * There is only the one khugepaged_scan instance of this cursor structure. | |
128 | */ | |
129 | struct khugepaged_scan { | |
130 | struct list_head mm_head; | |
131 | struct mm_slot *mm_slot; | |
132 | unsigned long address; | |
2f1da642 HS |
133 | }; |
134 | static struct khugepaged_scan khugepaged_scan = { | |
ba76149f AA |
135 | .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), |
136 | }; | |
137 | ||
f000565a | 138 | |
2c0b80d4 | 139 | static void set_recommended_min_free_kbytes(void) |
f000565a AA |
140 | { |
141 | struct zone *zone; | |
142 | int nr_zones = 0; | |
143 | unsigned long recommended_min; | |
f000565a | 144 | |
f000565a AA |
145 | for_each_populated_zone(zone) |
146 | nr_zones++; | |
147 | ||
974a786e | 148 | /* Ensure 2 pageblocks are free to assist fragmentation avoidance */ |
f000565a AA |
149 | recommended_min = pageblock_nr_pages * nr_zones * 2; |
150 | ||
151 | /* | |
152 | * Make sure that on average at least two pageblocks are almost free | |
153 | * of another type, one for a migratetype to fall back to and a | |
154 | * second to avoid subsequent fallbacks of other types There are 3 | |
155 | * MIGRATE_TYPES we care about. | |
156 | */ | |
157 | recommended_min += pageblock_nr_pages * nr_zones * | |
158 | MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; | |
159 | ||
160 | /* don't ever allow to reserve more than 5% of the lowmem */ | |
161 | recommended_min = min(recommended_min, | |
162 | (unsigned long) nr_free_buffer_pages() / 20); | |
163 | recommended_min <<= (PAGE_SHIFT-10); | |
164 | ||
42aa83cb HP |
165 | if (recommended_min > min_free_kbytes) { |
166 | if (user_min_free_kbytes >= 0) | |
167 | pr_info("raising min_free_kbytes from %d to %lu " | |
168 | "to help transparent hugepage allocations\n", | |
169 | min_free_kbytes, recommended_min); | |
170 | ||
f000565a | 171 | min_free_kbytes = recommended_min; |
42aa83cb | 172 | } |
f000565a | 173 | setup_per_zone_wmarks(); |
f000565a | 174 | } |
f000565a | 175 | |
79553da2 | 176 | static int start_stop_khugepaged(void) |
ba76149f AA |
177 | { |
178 | int err = 0; | |
179 | if (khugepaged_enabled()) { | |
ba76149f AA |
180 | if (!khugepaged_thread) |
181 | khugepaged_thread = kthread_run(khugepaged, NULL, | |
182 | "khugepaged"); | |
18e8e5c7 | 183 | if (IS_ERR(khugepaged_thread)) { |
ae3a8c1c | 184 | pr_err("khugepaged: kthread_run(khugepaged) failed\n"); |
ba76149f AA |
185 | err = PTR_ERR(khugepaged_thread); |
186 | khugepaged_thread = NULL; | |
79553da2 | 187 | goto fail; |
ba76149f | 188 | } |
911891af XG |
189 | |
190 | if (!list_empty(&khugepaged_scan.mm_head)) | |
ba76149f | 191 | wake_up_interruptible(&khugepaged_wait); |
f000565a AA |
192 | |
193 | set_recommended_min_free_kbytes(); | |
911891af | 194 | } else if (khugepaged_thread) { |
911891af XG |
195 | kthread_stop(khugepaged_thread); |
196 | khugepaged_thread = NULL; | |
197 | } | |
79553da2 | 198 | fail: |
ba76149f AA |
199 | return err; |
200 | } | |
71e3aac0 | 201 | |
97ae1749 | 202 | static atomic_t huge_zero_refcount; |
56873f43 | 203 | struct page *huge_zero_page __read_mostly; |
4a6c1297 | 204 | |
fc437044 | 205 | struct page *get_huge_zero_page(void) |
97ae1749 KS |
206 | { |
207 | struct page *zero_page; | |
208 | retry: | |
209 | if (likely(atomic_inc_not_zero(&huge_zero_refcount))) | |
4db0c3c2 | 210 | return READ_ONCE(huge_zero_page); |
97ae1749 KS |
211 | |
212 | zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, | |
4a6c1297 | 213 | HPAGE_PMD_ORDER); |
d8a8e1f0 KS |
214 | if (!zero_page) { |
215 | count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); | |
5918d10a | 216 | return NULL; |
d8a8e1f0 KS |
217 | } |
218 | count_vm_event(THP_ZERO_PAGE_ALLOC); | |
97ae1749 | 219 | preempt_disable(); |
5918d10a | 220 | if (cmpxchg(&huge_zero_page, NULL, zero_page)) { |
97ae1749 | 221 | preempt_enable(); |
5ddacbe9 | 222 | __free_pages(zero_page, compound_order(zero_page)); |
97ae1749 KS |
223 | goto retry; |
224 | } | |
225 | ||
226 | /* We take additional reference here. It will be put back by shrinker */ | |
227 | atomic_set(&huge_zero_refcount, 2); | |
228 | preempt_enable(); | |
4db0c3c2 | 229 | return READ_ONCE(huge_zero_page); |
4a6c1297 KS |
230 | } |
231 | ||
97ae1749 | 232 | static void put_huge_zero_page(void) |
4a6c1297 | 233 | { |
97ae1749 KS |
234 | /* |
235 | * Counter should never go to zero here. Only shrinker can put | |
236 | * last reference. | |
237 | */ | |
238 | BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); | |
4a6c1297 KS |
239 | } |
240 | ||
48896466 GC |
241 | static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, |
242 | struct shrink_control *sc) | |
4a6c1297 | 243 | { |
48896466 GC |
244 | /* we can free zero page only if last reference remains */ |
245 | return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; | |
246 | } | |
97ae1749 | 247 | |
48896466 GC |
248 | static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, |
249 | struct shrink_control *sc) | |
250 | { | |
97ae1749 | 251 | if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { |
5918d10a KS |
252 | struct page *zero_page = xchg(&huge_zero_page, NULL); |
253 | BUG_ON(zero_page == NULL); | |
5ddacbe9 | 254 | __free_pages(zero_page, compound_order(zero_page)); |
48896466 | 255 | return HPAGE_PMD_NR; |
97ae1749 KS |
256 | } |
257 | ||
258 | return 0; | |
4a6c1297 KS |
259 | } |
260 | ||
97ae1749 | 261 | static struct shrinker huge_zero_page_shrinker = { |
48896466 GC |
262 | .count_objects = shrink_huge_zero_page_count, |
263 | .scan_objects = shrink_huge_zero_page_scan, | |
97ae1749 KS |
264 | .seeks = DEFAULT_SEEKS, |
265 | }; | |
266 | ||
71e3aac0 | 267 | #ifdef CONFIG_SYSFS |
ba76149f | 268 | |
71e3aac0 AA |
269 | static ssize_t double_flag_show(struct kobject *kobj, |
270 | struct kobj_attribute *attr, char *buf, | |
271 | enum transparent_hugepage_flag enabled, | |
272 | enum transparent_hugepage_flag req_madv) | |
273 | { | |
274 | if (test_bit(enabled, &transparent_hugepage_flags)) { | |
275 | VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags)); | |
276 | return sprintf(buf, "[always] madvise never\n"); | |
277 | } else if (test_bit(req_madv, &transparent_hugepage_flags)) | |
278 | return sprintf(buf, "always [madvise] never\n"); | |
279 | else | |
280 | return sprintf(buf, "always madvise [never]\n"); | |
281 | } | |
282 | static ssize_t double_flag_store(struct kobject *kobj, | |
283 | struct kobj_attribute *attr, | |
284 | const char *buf, size_t count, | |
285 | enum transparent_hugepage_flag enabled, | |
286 | enum transparent_hugepage_flag req_madv) | |
287 | { | |
288 | if (!memcmp("always", buf, | |
289 | min(sizeof("always")-1, count))) { | |
290 | set_bit(enabled, &transparent_hugepage_flags); | |
291 | clear_bit(req_madv, &transparent_hugepage_flags); | |
292 | } else if (!memcmp("madvise", buf, | |
293 | min(sizeof("madvise")-1, count))) { | |
294 | clear_bit(enabled, &transparent_hugepage_flags); | |
295 | set_bit(req_madv, &transparent_hugepage_flags); | |
296 | } else if (!memcmp("never", buf, | |
297 | min(sizeof("never")-1, count))) { | |
298 | clear_bit(enabled, &transparent_hugepage_flags); | |
299 | clear_bit(req_madv, &transparent_hugepage_flags); | |
300 | } else | |
301 | return -EINVAL; | |
302 | ||
303 | return count; | |
304 | } | |
305 | ||
306 | static ssize_t enabled_show(struct kobject *kobj, | |
307 | struct kobj_attribute *attr, char *buf) | |
308 | { | |
309 | return double_flag_show(kobj, attr, buf, | |
310 | TRANSPARENT_HUGEPAGE_FLAG, | |
311 | TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); | |
312 | } | |
313 | static ssize_t enabled_store(struct kobject *kobj, | |
314 | struct kobj_attribute *attr, | |
315 | const char *buf, size_t count) | |
316 | { | |
ba76149f AA |
317 | ssize_t ret; |
318 | ||
319 | ret = double_flag_store(kobj, attr, buf, count, | |
320 | TRANSPARENT_HUGEPAGE_FLAG, | |
321 | TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); | |
322 | ||
323 | if (ret > 0) { | |
911891af XG |
324 | int err; |
325 | ||
326 | mutex_lock(&khugepaged_mutex); | |
79553da2 | 327 | err = start_stop_khugepaged(); |
911891af XG |
328 | mutex_unlock(&khugepaged_mutex); |
329 | ||
ba76149f AA |
330 | if (err) |
331 | ret = err; | |
332 | } | |
333 | ||
334 | return ret; | |
71e3aac0 AA |
335 | } |
336 | static struct kobj_attribute enabled_attr = | |
337 | __ATTR(enabled, 0644, enabled_show, enabled_store); | |
338 | ||
339 | static ssize_t single_flag_show(struct kobject *kobj, | |
340 | struct kobj_attribute *attr, char *buf, | |
341 | enum transparent_hugepage_flag flag) | |
342 | { | |
e27e6151 BH |
343 | return sprintf(buf, "%d\n", |
344 | !!test_bit(flag, &transparent_hugepage_flags)); | |
71e3aac0 | 345 | } |
e27e6151 | 346 | |
71e3aac0 AA |
347 | static ssize_t single_flag_store(struct kobject *kobj, |
348 | struct kobj_attribute *attr, | |
349 | const char *buf, size_t count, | |
350 | enum transparent_hugepage_flag flag) | |
351 | { | |
e27e6151 BH |
352 | unsigned long value; |
353 | int ret; | |
354 | ||
355 | ret = kstrtoul(buf, 10, &value); | |
356 | if (ret < 0) | |
357 | return ret; | |
358 | if (value > 1) | |
359 | return -EINVAL; | |
360 | ||
361 | if (value) | |
71e3aac0 | 362 | set_bit(flag, &transparent_hugepage_flags); |
e27e6151 | 363 | else |
71e3aac0 | 364 | clear_bit(flag, &transparent_hugepage_flags); |
71e3aac0 AA |
365 | |
366 | return count; | |
367 | } | |
368 | ||
369 | /* | |
370 | * Currently defrag only disables __GFP_NOWAIT for allocation. A blind | |
371 | * __GFP_REPEAT is too aggressive, it's never worth swapping tons of | |
372 | * memory just to allocate one more hugepage. | |
373 | */ | |
374 | static ssize_t defrag_show(struct kobject *kobj, | |
375 | struct kobj_attribute *attr, char *buf) | |
376 | { | |
377 | return double_flag_show(kobj, attr, buf, | |
378 | TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, | |
379 | TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); | |
380 | } | |
381 | static ssize_t defrag_store(struct kobject *kobj, | |
382 | struct kobj_attribute *attr, | |
383 | const char *buf, size_t count) | |
384 | { | |
385 | return double_flag_store(kobj, attr, buf, count, | |
386 | TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, | |
387 | TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); | |
388 | } | |
389 | static struct kobj_attribute defrag_attr = | |
390 | __ATTR(defrag, 0644, defrag_show, defrag_store); | |
391 | ||
79da5407 KS |
392 | static ssize_t use_zero_page_show(struct kobject *kobj, |
393 | struct kobj_attribute *attr, char *buf) | |
394 | { | |
395 | return single_flag_show(kobj, attr, buf, | |
396 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
397 | } | |
398 | static ssize_t use_zero_page_store(struct kobject *kobj, | |
399 | struct kobj_attribute *attr, const char *buf, size_t count) | |
400 | { | |
401 | return single_flag_store(kobj, attr, buf, count, | |
402 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
403 | } | |
404 | static struct kobj_attribute use_zero_page_attr = | |
405 | __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); | |
71e3aac0 AA |
406 | #ifdef CONFIG_DEBUG_VM |
407 | static ssize_t debug_cow_show(struct kobject *kobj, | |
408 | struct kobj_attribute *attr, char *buf) | |
409 | { | |
410 | return single_flag_show(kobj, attr, buf, | |
411 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); | |
412 | } | |
413 | static ssize_t debug_cow_store(struct kobject *kobj, | |
414 | struct kobj_attribute *attr, | |
415 | const char *buf, size_t count) | |
416 | { | |
417 | return single_flag_store(kobj, attr, buf, count, | |
418 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); | |
419 | } | |
420 | static struct kobj_attribute debug_cow_attr = | |
421 | __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); | |
422 | #endif /* CONFIG_DEBUG_VM */ | |
423 | ||
424 | static struct attribute *hugepage_attr[] = { | |
425 | &enabled_attr.attr, | |
426 | &defrag_attr.attr, | |
79da5407 | 427 | &use_zero_page_attr.attr, |
71e3aac0 AA |
428 | #ifdef CONFIG_DEBUG_VM |
429 | &debug_cow_attr.attr, | |
430 | #endif | |
431 | NULL, | |
432 | }; | |
433 | ||
434 | static struct attribute_group hugepage_attr_group = { | |
435 | .attrs = hugepage_attr, | |
ba76149f AA |
436 | }; |
437 | ||
438 | static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, | |
439 | struct kobj_attribute *attr, | |
440 | char *buf) | |
441 | { | |
442 | return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs); | |
443 | } | |
444 | ||
445 | static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, | |
446 | struct kobj_attribute *attr, | |
447 | const char *buf, size_t count) | |
448 | { | |
449 | unsigned long msecs; | |
450 | int err; | |
451 | ||
3dbb95f7 | 452 | err = kstrtoul(buf, 10, &msecs); |
ba76149f AA |
453 | if (err || msecs > UINT_MAX) |
454 | return -EINVAL; | |
455 | ||
456 | khugepaged_scan_sleep_millisecs = msecs; | |
457 | wake_up_interruptible(&khugepaged_wait); | |
458 | ||
459 | return count; | |
460 | } | |
461 | static struct kobj_attribute scan_sleep_millisecs_attr = | |
462 | __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show, | |
463 | scan_sleep_millisecs_store); | |
464 | ||
465 | static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, | |
466 | struct kobj_attribute *attr, | |
467 | char *buf) | |
468 | { | |
469 | return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs); | |
470 | } | |
471 | ||
472 | static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, | |
473 | struct kobj_attribute *attr, | |
474 | const char *buf, size_t count) | |
475 | { | |
476 | unsigned long msecs; | |
477 | int err; | |
478 | ||
3dbb95f7 | 479 | err = kstrtoul(buf, 10, &msecs); |
ba76149f AA |
480 | if (err || msecs > UINT_MAX) |
481 | return -EINVAL; | |
482 | ||
483 | khugepaged_alloc_sleep_millisecs = msecs; | |
484 | wake_up_interruptible(&khugepaged_wait); | |
485 | ||
486 | return count; | |
487 | } | |
488 | static struct kobj_attribute alloc_sleep_millisecs_attr = | |
489 | __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show, | |
490 | alloc_sleep_millisecs_store); | |
491 | ||
492 | static ssize_t pages_to_scan_show(struct kobject *kobj, | |
493 | struct kobj_attribute *attr, | |
494 | char *buf) | |
495 | { | |
496 | return sprintf(buf, "%u\n", khugepaged_pages_to_scan); | |
497 | } | |
498 | static ssize_t pages_to_scan_store(struct kobject *kobj, | |
499 | struct kobj_attribute *attr, | |
500 | const char *buf, size_t count) | |
501 | { | |
502 | int err; | |
503 | unsigned long pages; | |
504 | ||
3dbb95f7 | 505 | err = kstrtoul(buf, 10, &pages); |
ba76149f AA |
506 | if (err || !pages || pages > UINT_MAX) |
507 | return -EINVAL; | |
508 | ||
509 | khugepaged_pages_to_scan = pages; | |
510 | ||
511 | return count; | |
512 | } | |
513 | static struct kobj_attribute pages_to_scan_attr = | |
514 | __ATTR(pages_to_scan, 0644, pages_to_scan_show, | |
515 | pages_to_scan_store); | |
516 | ||
517 | static ssize_t pages_collapsed_show(struct kobject *kobj, | |
518 | struct kobj_attribute *attr, | |
519 | char *buf) | |
520 | { | |
521 | return sprintf(buf, "%u\n", khugepaged_pages_collapsed); | |
522 | } | |
523 | static struct kobj_attribute pages_collapsed_attr = | |
524 | __ATTR_RO(pages_collapsed); | |
525 | ||
526 | static ssize_t full_scans_show(struct kobject *kobj, | |
527 | struct kobj_attribute *attr, | |
528 | char *buf) | |
529 | { | |
530 | return sprintf(buf, "%u\n", khugepaged_full_scans); | |
531 | } | |
532 | static struct kobj_attribute full_scans_attr = | |
533 | __ATTR_RO(full_scans); | |
534 | ||
535 | static ssize_t khugepaged_defrag_show(struct kobject *kobj, | |
536 | struct kobj_attribute *attr, char *buf) | |
537 | { | |
538 | return single_flag_show(kobj, attr, buf, | |
539 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); | |
540 | } | |
541 | static ssize_t khugepaged_defrag_store(struct kobject *kobj, | |
542 | struct kobj_attribute *attr, | |
543 | const char *buf, size_t count) | |
544 | { | |
545 | return single_flag_store(kobj, attr, buf, count, | |
546 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); | |
547 | } | |
548 | static struct kobj_attribute khugepaged_defrag_attr = | |
549 | __ATTR(defrag, 0644, khugepaged_defrag_show, | |
550 | khugepaged_defrag_store); | |
551 | ||
552 | /* | |
553 | * max_ptes_none controls if khugepaged should collapse hugepages over | |
554 | * any unmapped ptes in turn potentially increasing the memory | |
555 | * footprint of the vmas. When max_ptes_none is 0 khugepaged will not | |
556 | * reduce the available free memory in the system as it | |
557 | * runs. Increasing max_ptes_none will instead potentially reduce the | |
558 | * free memory in the system during the khugepaged scan. | |
559 | */ | |
560 | static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj, | |
561 | struct kobj_attribute *attr, | |
562 | char *buf) | |
563 | { | |
564 | return sprintf(buf, "%u\n", khugepaged_max_ptes_none); | |
565 | } | |
566 | static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj, | |
567 | struct kobj_attribute *attr, | |
568 | const char *buf, size_t count) | |
569 | { | |
570 | int err; | |
571 | unsigned long max_ptes_none; | |
572 | ||
3dbb95f7 | 573 | err = kstrtoul(buf, 10, &max_ptes_none); |
ba76149f AA |
574 | if (err || max_ptes_none > HPAGE_PMD_NR-1) |
575 | return -EINVAL; | |
576 | ||
577 | khugepaged_max_ptes_none = max_ptes_none; | |
578 | ||
579 | return count; | |
580 | } | |
581 | static struct kobj_attribute khugepaged_max_ptes_none_attr = | |
582 | __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show, | |
583 | khugepaged_max_ptes_none_store); | |
584 | ||
585 | static struct attribute *khugepaged_attr[] = { | |
586 | &khugepaged_defrag_attr.attr, | |
587 | &khugepaged_max_ptes_none_attr.attr, | |
588 | &pages_to_scan_attr.attr, | |
589 | &pages_collapsed_attr.attr, | |
590 | &full_scans_attr.attr, | |
591 | &scan_sleep_millisecs_attr.attr, | |
592 | &alloc_sleep_millisecs_attr.attr, | |
593 | NULL, | |
594 | }; | |
595 | ||
596 | static struct attribute_group khugepaged_attr_group = { | |
597 | .attrs = khugepaged_attr, | |
598 | .name = "khugepaged", | |
71e3aac0 | 599 | }; |
71e3aac0 | 600 | |
569e5590 | 601 | static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) |
71e3aac0 | 602 | { |
71e3aac0 AA |
603 | int err; |
604 | ||
569e5590 SL |
605 | *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); |
606 | if (unlikely(!*hugepage_kobj)) { | |
ae3a8c1c | 607 | pr_err("failed to create transparent hugepage kobject\n"); |
569e5590 | 608 | return -ENOMEM; |
ba76149f AA |
609 | } |
610 | ||
569e5590 | 611 | err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); |
ba76149f | 612 | if (err) { |
ae3a8c1c | 613 | pr_err("failed to register transparent hugepage group\n"); |
569e5590 | 614 | goto delete_obj; |
ba76149f AA |
615 | } |
616 | ||
569e5590 | 617 | err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); |
ba76149f | 618 | if (err) { |
ae3a8c1c | 619 | pr_err("failed to register transparent hugepage group\n"); |
569e5590 | 620 | goto remove_hp_group; |
ba76149f | 621 | } |
569e5590 SL |
622 | |
623 | return 0; | |
624 | ||
625 | remove_hp_group: | |
626 | sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); | |
627 | delete_obj: | |
628 | kobject_put(*hugepage_kobj); | |
629 | return err; | |
630 | } | |
631 | ||
632 | static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
633 | { | |
634 | sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); | |
635 | sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); | |
636 | kobject_put(hugepage_kobj); | |
637 | } | |
638 | #else | |
639 | static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) | |
640 | { | |
641 | return 0; | |
642 | } | |
643 | ||
644 | static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
645 | { | |
646 | } | |
647 | #endif /* CONFIG_SYSFS */ | |
648 | ||
649 | static int __init hugepage_init(void) | |
650 | { | |
651 | int err; | |
652 | struct kobject *hugepage_kobj; | |
653 | ||
654 | if (!has_transparent_hugepage()) { | |
655 | transparent_hugepage_flags = 0; | |
656 | return -EINVAL; | |
657 | } | |
658 | ||
659 | err = hugepage_init_sysfs(&hugepage_kobj); | |
660 | if (err) | |
65ebb64f | 661 | goto err_sysfs; |
ba76149f AA |
662 | |
663 | err = khugepaged_slab_init(); | |
664 | if (err) | |
65ebb64f | 665 | goto err_slab; |
ba76149f | 666 | |
65ebb64f KS |
667 | err = register_shrinker(&huge_zero_page_shrinker); |
668 | if (err) | |
669 | goto err_hzp_shrinker; | |
97ae1749 | 670 | |
97562cd2 RR |
671 | /* |
672 | * By default disable transparent hugepages on smaller systems, | |
673 | * where the extra memory used could hurt more than TLB overhead | |
674 | * is likely to save. The admin can still enable it through /sys. | |
675 | */ | |
79553da2 | 676 | if (totalram_pages < (512 << (20 - PAGE_SHIFT))) { |
97562cd2 | 677 | transparent_hugepage_flags = 0; |
79553da2 KS |
678 | return 0; |
679 | } | |
97562cd2 | 680 | |
79553da2 | 681 | err = start_stop_khugepaged(); |
65ebb64f KS |
682 | if (err) |
683 | goto err_khugepaged; | |
ba76149f | 684 | |
569e5590 | 685 | return 0; |
65ebb64f KS |
686 | err_khugepaged: |
687 | unregister_shrinker(&huge_zero_page_shrinker); | |
688 | err_hzp_shrinker: | |
689 | khugepaged_slab_exit(); | |
690 | err_slab: | |
569e5590 | 691 | hugepage_exit_sysfs(hugepage_kobj); |
65ebb64f | 692 | err_sysfs: |
ba76149f | 693 | return err; |
71e3aac0 | 694 | } |
a64fb3cd | 695 | subsys_initcall(hugepage_init); |
71e3aac0 AA |
696 | |
697 | static int __init setup_transparent_hugepage(char *str) | |
698 | { | |
699 | int ret = 0; | |
700 | if (!str) | |
701 | goto out; | |
702 | if (!strcmp(str, "always")) { | |
703 | set_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
704 | &transparent_hugepage_flags); | |
705 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
706 | &transparent_hugepage_flags); | |
707 | ret = 1; | |
708 | } else if (!strcmp(str, "madvise")) { | |
709 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
710 | &transparent_hugepage_flags); | |
711 | set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
712 | &transparent_hugepage_flags); | |
713 | ret = 1; | |
714 | } else if (!strcmp(str, "never")) { | |
715 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
716 | &transparent_hugepage_flags); | |
717 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
718 | &transparent_hugepage_flags); | |
719 | ret = 1; | |
720 | } | |
721 | out: | |
722 | if (!ret) | |
ae3a8c1c | 723 | pr_warn("transparent_hugepage= cannot parse, ignored\n"); |
71e3aac0 AA |
724 | return ret; |
725 | } | |
726 | __setup("transparent_hugepage=", setup_transparent_hugepage); | |
727 | ||
b32967ff | 728 | pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) |
71e3aac0 AA |
729 | { |
730 | if (likely(vma->vm_flags & VM_WRITE)) | |
731 | pmd = pmd_mkwrite(pmd); | |
732 | return pmd; | |
733 | } | |
734 | ||
3122359a | 735 | static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot) |
b3092b3b BL |
736 | { |
737 | pmd_t entry; | |
3122359a | 738 | entry = mk_pmd(page, prot); |
b3092b3b BL |
739 | entry = pmd_mkhuge(entry); |
740 | return entry; | |
741 | } | |
742 | ||
71e3aac0 AA |
743 | static int __do_huge_pmd_anonymous_page(struct mm_struct *mm, |
744 | struct vm_area_struct *vma, | |
230c92a8 | 745 | unsigned long address, pmd_t *pmd, |
6b251fc9 AA |
746 | struct page *page, gfp_t gfp, |
747 | unsigned int flags) | |
71e3aac0 | 748 | { |
00501b53 | 749 | struct mem_cgroup *memcg; |
71e3aac0 | 750 | pgtable_t pgtable; |
c4088ebd | 751 | spinlock_t *ptl; |
230c92a8 | 752 | unsigned long haddr = address & HPAGE_PMD_MASK; |
71e3aac0 | 753 | |
309381fe | 754 | VM_BUG_ON_PAGE(!PageCompound(page), page); |
00501b53 | 755 | |
f627c2f5 | 756 | if (mem_cgroup_try_charge(page, mm, gfp, &memcg, true)) { |
6b251fc9 AA |
757 | put_page(page); |
758 | count_vm_event(THP_FAULT_FALLBACK); | |
759 | return VM_FAULT_FALLBACK; | |
760 | } | |
00501b53 | 761 | |
71e3aac0 | 762 | pgtable = pte_alloc_one(mm, haddr); |
00501b53 | 763 | if (unlikely(!pgtable)) { |
f627c2f5 | 764 | mem_cgroup_cancel_charge(page, memcg, true); |
6b251fc9 | 765 | put_page(page); |
71e3aac0 | 766 | return VM_FAULT_OOM; |
00501b53 | 767 | } |
71e3aac0 AA |
768 | |
769 | clear_huge_page(page, haddr, HPAGE_PMD_NR); | |
52f37629 MK |
770 | /* |
771 | * The memory barrier inside __SetPageUptodate makes sure that | |
772 | * clear_huge_page writes become visible before the set_pmd_at() | |
773 | * write. | |
774 | */ | |
71e3aac0 AA |
775 | __SetPageUptodate(page); |
776 | ||
c4088ebd | 777 | ptl = pmd_lock(mm, pmd); |
71e3aac0 | 778 | if (unlikely(!pmd_none(*pmd))) { |
c4088ebd | 779 | spin_unlock(ptl); |
f627c2f5 | 780 | mem_cgroup_cancel_charge(page, memcg, true); |
71e3aac0 AA |
781 | put_page(page); |
782 | pte_free(mm, pgtable); | |
783 | } else { | |
784 | pmd_t entry; | |
6b251fc9 AA |
785 | |
786 | /* Deliver the page fault to userland */ | |
787 | if (userfaultfd_missing(vma)) { | |
788 | int ret; | |
789 | ||
790 | spin_unlock(ptl); | |
f627c2f5 | 791 | mem_cgroup_cancel_charge(page, memcg, true); |
6b251fc9 AA |
792 | put_page(page); |
793 | pte_free(mm, pgtable); | |
230c92a8 | 794 | ret = handle_userfault(vma, address, flags, |
6b251fc9 AA |
795 | VM_UFFD_MISSING); |
796 | VM_BUG_ON(ret & VM_FAULT_FALLBACK); | |
797 | return ret; | |
798 | } | |
799 | ||
3122359a KS |
800 | entry = mk_huge_pmd(page, vma->vm_page_prot); |
801 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
d281ee61 | 802 | page_add_new_anon_rmap(page, vma, haddr, true); |
f627c2f5 | 803 | mem_cgroup_commit_charge(page, memcg, false, true); |
00501b53 | 804 | lru_cache_add_active_or_unevictable(page, vma); |
6b0b50b0 | 805 | pgtable_trans_huge_deposit(mm, pmd, pgtable); |
71e3aac0 | 806 | set_pmd_at(mm, haddr, pmd, entry); |
71e3aac0 | 807 | add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); |
e1f56c89 | 808 | atomic_long_inc(&mm->nr_ptes); |
c4088ebd | 809 | spin_unlock(ptl); |
6b251fc9 | 810 | count_vm_event(THP_FAULT_ALLOC); |
71e3aac0 AA |
811 | } |
812 | ||
aa2e878e | 813 | return 0; |
71e3aac0 AA |
814 | } |
815 | ||
cc5d462f | 816 | static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp) |
0bbbc0b3 | 817 | { |
71baba4b | 818 | return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_RECLAIM)) | extra_gfp; |
0bbbc0b3 AA |
819 | } |
820 | ||
c4088ebd | 821 | /* Caller must hold page table lock. */ |
d295e341 | 822 | static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, |
97ae1749 | 823 | struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, |
5918d10a | 824 | struct page *zero_page) |
fc9fe822 KS |
825 | { |
826 | pmd_t entry; | |
7c414164 AM |
827 | if (!pmd_none(*pmd)) |
828 | return false; | |
5918d10a | 829 | entry = mk_pmd(zero_page, vma->vm_page_prot); |
fc9fe822 | 830 | entry = pmd_mkhuge(entry); |
6b0b50b0 | 831 | pgtable_trans_huge_deposit(mm, pmd, pgtable); |
fc9fe822 | 832 | set_pmd_at(mm, haddr, pmd, entry); |
e1f56c89 | 833 | atomic_long_inc(&mm->nr_ptes); |
7c414164 | 834 | return true; |
fc9fe822 KS |
835 | } |
836 | ||
71e3aac0 AA |
837 | int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma, |
838 | unsigned long address, pmd_t *pmd, | |
839 | unsigned int flags) | |
840 | { | |
077fcf11 | 841 | gfp_t gfp; |
71e3aac0 AA |
842 | struct page *page; |
843 | unsigned long haddr = address & HPAGE_PMD_MASK; | |
71e3aac0 | 844 | |
128ec037 | 845 | if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end) |
c0292554 | 846 | return VM_FAULT_FALLBACK; |
7479df6d KS |
847 | if (vma->vm_flags & VM_LOCKED) |
848 | return VM_FAULT_FALLBACK; | |
128ec037 KS |
849 | if (unlikely(anon_vma_prepare(vma))) |
850 | return VM_FAULT_OOM; | |
6d50e60c | 851 | if (unlikely(khugepaged_enter(vma, vma->vm_flags))) |
128ec037 | 852 | return VM_FAULT_OOM; |
593befa6 | 853 | if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) && |
128ec037 | 854 | transparent_hugepage_use_zero_page()) { |
c4088ebd | 855 | spinlock_t *ptl; |
128ec037 KS |
856 | pgtable_t pgtable; |
857 | struct page *zero_page; | |
858 | bool set; | |
6b251fc9 | 859 | int ret; |
128ec037 KS |
860 | pgtable = pte_alloc_one(mm, haddr); |
861 | if (unlikely(!pgtable)) | |
ba76149f | 862 | return VM_FAULT_OOM; |
128ec037 KS |
863 | zero_page = get_huge_zero_page(); |
864 | if (unlikely(!zero_page)) { | |
865 | pte_free(mm, pgtable); | |
81ab4201 | 866 | count_vm_event(THP_FAULT_FALLBACK); |
c0292554 | 867 | return VM_FAULT_FALLBACK; |
b9bbfbe3 | 868 | } |
c4088ebd | 869 | ptl = pmd_lock(mm, pmd); |
6b251fc9 AA |
870 | ret = 0; |
871 | set = false; | |
872 | if (pmd_none(*pmd)) { | |
873 | if (userfaultfd_missing(vma)) { | |
874 | spin_unlock(ptl); | |
230c92a8 | 875 | ret = handle_userfault(vma, address, flags, |
6b251fc9 AA |
876 | VM_UFFD_MISSING); |
877 | VM_BUG_ON(ret & VM_FAULT_FALLBACK); | |
878 | } else { | |
879 | set_huge_zero_page(pgtable, mm, vma, | |
880 | haddr, pmd, | |
881 | zero_page); | |
882 | spin_unlock(ptl); | |
883 | set = true; | |
884 | } | |
885 | } else | |
886 | spin_unlock(ptl); | |
128ec037 KS |
887 | if (!set) { |
888 | pte_free(mm, pgtable); | |
889 | put_huge_zero_page(); | |
edad9d2c | 890 | } |
6b251fc9 | 891 | return ret; |
71e3aac0 | 892 | } |
077fcf11 AK |
893 | gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0); |
894 | page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); | |
128ec037 KS |
895 | if (unlikely(!page)) { |
896 | count_vm_event(THP_FAULT_FALLBACK); | |
c0292554 | 897 | return VM_FAULT_FALLBACK; |
128ec037 | 898 | } |
230c92a8 AA |
899 | return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp, |
900 | flags); | |
71e3aac0 AA |
901 | } |
902 | ||
ae18d6dc | 903 | static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, |
5cad465d MW |
904 | pmd_t *pmd, unsigned long pfn, pgprot_t prot, bool write) |
905 | { | |
906 | struct mm_struct *mm = vma->vm_mm; | |
907 | pmd_t entry; | |
908 | spinlock_t *ptl; | |
909 | ||
910 | ptl = pmd_lock(mm, pmd); | |
911 | if (pmd_none(*pmd)) { | |
912 | entry = pmd_mkhuge(pfn_pmd(pfn, prot)); | |
913 | if (write) { | |
914 | entry = pmd_mkyoung(pmd_mkdirty(entry)); | |
915 | entry = maybe_pmd_mkwrite(entry, vma); | |
916 | } | |
917 | set_pmd_at(mm, addr, pmd, entry); | |
918 | update_mmu_cache_pmd(vma, addr, pmd); | |
919 | } | |
920 | spin_unlock(ptl); | |
5cad465d MW |
921 | } |
922 | ||
923 | int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, | |
924 | pmd_t *pmd, unsigned long pfn, bool write) | |
925 | { | |
926 | pgprot_t pgprot = vma->vm_page_prot; | |
927 | /* | |
928 | * If we had pmd_special, we could avoid all these restrictions, | |
929 | * but we need to be consistent with PTEs and architectures that | |
930 | * can't support a 'special' bit. | |
931 | */ | |
932 | BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); | |
933 | BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == | |
934 | (VM_PFNMAP|VM_MIXEDMAP)); | |
935 | BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); | |
936 | BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn)); | |
937 | ||
938 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
939 | return VM_FAULT_SIGBUS; | |
940 | if (track_pfn_insert(vma, &pgprot, pfn)) | |
941 | return VM_FAULT_SIGBUS; | |
ae18d6dc MW |
942 | insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write); |
943 | return VM_FAULT_NOPAGE; | |
5cad465d MW |
944 | } |
945 | ||
71e3aac0 AA |
946 | int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, |
947 | pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, | |
948 | struct vm_area_struct *vma) | |
949 | { | |
c4088ebd | 950 | spinlock_t *dst_ptl, *src_ptl; |
71e3aac0 AA |
951 | struct page *src_page; |
952 | pmd_t pmd; | |
953 | pgtable_t pgtable; | |
954 | int ret; | |
955 | ||
956 | ret = -ENOMEM; | |
957 | pgtable = pte_alloc_one(dst_mm, addr); | |
958 | if (unlikely(!pgtable)) | |
959 | goto out; | |
960 | ||
c4088ebd KS |
961 | dst_ptl = pmd_lock(dst_mm, dst_pmd); |
962 | src_ptl = pmd_lockptr(src_mm, src_pmd); | |
963 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
71e3aac0 AA |
964 | |
965 | ret = -EAGAIN; | |
966 | pmd = *src_pmd; | |
967 | if (unlikely(!pmd_trans_huge(pmd))) { | |
968 | pte_free(dst_mm, pgtable); | |
969 | goto out_unlock; | |
970 | } | |
fc9fe822 | 971 | /* |
c4088ebd | 972 | * When page table lock is held, the huge zero pmd should not be |
fc9fe822 KS |
973 | * under splitting since we don't split the page itself, only pmd to |
974 | * a page table. | |
975 | */ | |
976 | if (is_huge_zero_pmd(pmd)) { | |
5918d10a | 977 | struct page *zero_page; |
97ae1749 KS |
978 | /* |
979 | * get_huge_zero_page() will never allocate a new page here, | |
980 | * since we already have a zero page to copy. It just takes a | |
981 | * reference. | |
982 | */ | |
5918d10a | 983 | zero_page = get_huge_zero_page(); |
6b251fc9 | 984 | set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, |
5918d10a | 985 | zero_page); |
fc9fe822 KS |
986 | ret = 0; |
987 | goto out_unlock; | |
988 | } | |
de466bd6 | 989 | |
71e3aac0 | 990 | src_page = pmd_page(pmd); |
309381fe | 991 | VM_BUG_ON_PAGE(!PageHead(src_page), src_page); |
71e3aac0 | 992 | get_page(src_page); |
53f9263b | 993 | page_dup_rmap(src_page, true); |
71e3aac0 AA |
994 | add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); |
995 | ||
996 | pmdp_set_wrprotect(src_mm, addr, src_pmd); | |
997 | pmd = pmd_mkold(pmd_wrprotect(pmd)); | |
6b0b50b0 | 998 | pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); |
71e3aac0 | 999 | set_pmd_at(dst_mm, addr, dst_pmd, pmd); |
e1f56c89 | 1000 | atomic_long_inc(&dst_mm->nr_ptes); |
71e3aac0 AA |
1001 | |
1002 | ret = 0; | |
1003 | out_unlock: | |
c4088ebd KS |
1004 | spin_unlock(src_ptl); |
1005 | spin_unlock(dst_ptl); | |
71e3aac0 AA |
1006 | out: |
1007 | return ret; | |
1008 | } | |
1009 | ||
a1dd450b WD |
1010 | void huge_pmd_set_accessed(struct mm_struct *mm, |
1011 | struct vm_area_struct *vma, | |
1012 | unsigned long address, | |
1013 | pmd_t *pmd, pmd_t orig_pmd, | |
1014 | int dirty) | |
1015 | { | |
c4088ebd | 1016 | spinlock_t *ptl; |
a1dd450b WD |
1017 | pmd_t entry; |
1018 | unsigned long haddr; | |
1019 | ||
c4088ebd | 1020 | ptl = pmd_lock(mm, pmd); |
a1dd450b WD |
1021 | if (unlikely(!pmd_same(*pmd, orig_pmd))) |
1022 | goto unlock; | |
1023 | ||
1024 | entry = pmd_mkyoung(orig_pmd); | |
1025 | haddr = address & HPAGE_PMD_MASK; | |
1026 | if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty)) | |
1027 | update_mmu_cache_pmd(vma, address, pmd); | |
1028 | ||
1029 | unlock: | |
c4088ebd | 1030 | spin_unlock(ptl); |
a1dd450b WD |
1031 | } |
1032 | ||
71e3aac0 AA |
1033 | static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm, |
1034 | struct vm_area_struct *vma, | |
1035 | unsigned long address, | |
1036 | pmd_t *pmd, pmd_t orig_pmd, | |
1037 | struct page *page, | |
1038 | unsigned long haddr) | |
1039 | { | |
00501b53 | 1040 | struct mem_cgroup *memcg; |
c4088ebd | 1041 | spinlock_t *ptl; |
71e3aac0 AA |
1042 | pgtable_t pgtable; |
1043 | pmd_t _pmd; | |
1044 | int ret = 0, i; | |
1045 | struct page **pages; | |
2ec74c3e SG |
1046 | unsigned long mmun_start; /* For mmu_notifiers */ |
1047 | unsigned long mmun_end; /* For mmu_notifiers */ | |
71e3aac0 AA |
1048 | |
1049 | pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR, | |
1050 | GFP_KERNEL); | |
1051 | if (unlikely(!pages)) { | |
1052 | ret |= VM_FAULT_OOM; | |
1053 | goto out; | |
1054 | } | |
1055 | ||
1056 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
cc5d462f AK |
1057 | pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE | |
1058 | __GFP_OTHER_NODE, | |
19ee151e | 1059 | vma, address, page_to_nid(page)); |
b9bbfbe3 | 1060 | if (unlikely(!pages[i] || |
00501b53 | 1061 | mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL, |
f627c2f5 | 1062 | &memcg, false))) { |
b9bbfbe3 | 1063 | if (pages[i]) |
71e3aac0 | 1064 | put_page(pages[i]); |
b9bbfbe3 | 1065 | while (--i >= 0) { |
00501b53 JW |
1066 | memcg = (void *)page_private(pages[i]); |
1067 | set_page_private(pages[i], 0); | |
f627c2f5 KS |
1068 | mem_cgroup_cancel_charge(pages[i], memcg, |
1069 | false); | |
b9bbfbe3 AA |
1070 | put_page(pages[i]); |
1071 | } | |
71e3aac0 AA |
1072 | kfree(pages); |
1073 | ret |= VM_FAULT_OOM; | |
1074 | goto out; | |
1075 | } | |
00501b53 | 1076 | set_page_private(pages[i], (unsigned long)memcg); |
71e3aac0 AA |
1077 | } |
1078 | ||
1079 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
1080 | copy_user_highpage(pages[i], page + i, | |
0089e485 | 1081 | haddr + PAGE_SIZE * i, vma); |
71e3aac0 AA |
1082 | __SetPageUptodate(pages[i]); |
1083 | cond_resched(); | |
1084 | } | |
1085 | ||
2ec74c3e SG |
1086 | mmun_start = haddr; |
1087 | mmun_end = haddr + HPAGE_PMD_SIZE; | |
1088 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
1089 | ||
c4088ebd | 1090 | ptl = pmd_lock(mm, pmd); |
71e3aac0 AA |
1091 | if (unlikely(!pmd_same(*pmd, orig_pmd))) |
1092 | goto out_free_pages; | |
309381fe | 1093 | VM_BUG_ON_PAGE(!PageHead(page), page); |
71e3aac0 | 1094 | |
8809aa2d | 1095 | pmdp_huge_clear_flush_notify(vma, haddr, pmd); |
71e3aac0 AA |
1096 | /* leave pmd empty until pte is filled */ |
1097 | ||
6b0b50b0 | 1098 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); |
71e3aac0 AA |
1099 | pmd_populate(mm, &_pmd, pgtable); |
1100 | ||
1101 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
1102 | pte_t *pte, entry; | |
1103 | entry = mk_pte(pages[i], vma->vm_page_prot); | |
1104 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); | |
00501b53 JW |
1105 | memcg = (void *)page_private(pages[i]); |
1106 | set_page_private(pages[i], 0); | |
d281ee61 | 1107 | page_add_new_anon_rmap(pages[i], vma, haddr, false); |
f627c2f5 | 1108 | mem_cgroup_commit_charge(pages[i], memcg, false, false); |
00501b53 | 1109 | lru_cache_add_active_or_unevictable(pages[i], vma); |
71e3aac0 AA |
1110 | pte = pte_offset_map(&_pmd, haddr); |
1111 | VM_BUG_ON(!pte_none(*pte)); | |
1112 | set_pte_at(mm, haddr, pte, entry); | |
1113 | pte_unmap(pte); | |
1114 | } | |
1115 | kfree(pages); | |
1116 | ||
71e3aac0 AA |
1117 | smp_wmb(); /* make pte visible before pmd */ |
1118 | pmd_populate(mm, pmd, pgtable); | |
d281ee61 | 1119 | page_remove_rmap(page, true); |
c4088ebd | 1120 | spin_unlock(ptl); |
71e3aac0 | 1121 | |
2ec74c3e SG |
1122 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
1123 | ||
71e3aac0 AA |
1124 | ret |= VM_FAULT_WRITE; |
1125 | put_page(page); | |
1126 | ||
1127 | out: | |
1128 | return ret; | |
1129 | ||
1130 | out_free_pages: | |
c4088ebd | 1131 | spin_unlock(ptl); |
2ec74c3e | 1132 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
b9bbfbe3 | 1133 | for (i = 0; i < HPAGE_PMD_NR; i++) { |
00501b53 JW |
1134 | memcg = (void *)page_private(pages[i]); |
1135 | set_page_private(pages[i], 0); | |
f627c2f5 | 1136 | mem_cgroup_cancel_charge(pages[i], memcg, false); |
71e3aac0 | 1137 | put_page(pages[i]); |
b9bbfbe3 | 1138 | } |
71e3aac0 AA |
1139 | kfree(pages); |
1140 | goto out; | |
1141 | } | |
1142 | ||
1143 | int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma, | |
1144 | unsigned long address, pmd_t *pmd, pmd_t orig_pmd) | |
1145 | { | |
c4088ebd | 1146 | spinlock_t *ptl; |
71e3aac0 | 1147 | int ret = 0; |
93b4796d | 1148 | struct page *page = NULL, *new_page; |
00501b53 | 1149 | struct mem_cgroup *memcg; |
71e3aac0 | 1150 | unsigned long haddr; |
2ec74c3e SG |
1151 | unsigned long mmun_start; /* For mmu_notifiers */ |
1152 | unsigned long mmun_end; /* For mmu_notifiers */ | |
3b363692 | 1153 | gfp_t huge_gfp; /* for allocation and charge */ |
71e3aac0 | 1154 | |
c4088ebd | 1155 | ptl = pmd_lockptr(mm, pmd); |
81d1b09c | 1156 | VM_BUG_ON_VMA(!vma->anon_vma, vma); |
93b4796d KS |
1157 | haddr = address & HPAGE_PMD_MASK; |
1158 | if (is_huge_zero_pmd(orig_pmd)) | |
1159 | goto alloc; | |
c4088ebd | 1160 | spin_lock(ptl); |
71e3aac0 AA |
1161 | if (unlikely(!pmd_same(*pmd, orig_pmd))) |
1162 | goto out_unlock; | |
1163 | ||
1164 | page = pmd_page(orig_pmd); | |
309381fe | 1165 | VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); |
1f25fe20 KS |
1166 | /* |
1167 | * We can only reuse the page if nobody else maps the huge page or it's | |
1168 | * part. We can do it by checking page_mapcount() on each sub-page, but | |
1169 | * it's expensive. | |
1170 | * The cheaper way is to check page_count() to be equal 1: every | |
1171 | * mapcount takes page reference reference, so this way we can | |
1172 | * guarantee, that the PMD is the only mapping. | |
1173 | * This can give false negative if somebody pinned the page, but that's | |
1174 | * fine. | |
1175 | */ | |
1176 | if (page_mapcount(page) == 1 && page_count(page) == 1) { | |
71e3aac0 AA |
1177 | pmd_t entry; |
1178 | entry = pmd_mkyoung(orig_pmd); | |
1179 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
1180 | if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1)) | |
b113da65 | 1181 | update_mmu_cache_pmd(vma, address, pmd); |
71e3aac0 AA |
1182 | ret |= VM_FAULT_WRITE; |
1183 | goto out_unlock; | |
1184 | } | |
ddc58f27 | 1185 | get_page(page); |
c4088ebd | 1186 | spin_unlock(ptl); |
93b4796d | 1187 | alloc: |
71e3aac0 | 1188 | if (transparent_hugepage_enabled(vma) && |
077fcf11 | 1189 | !transparent_hugepage_debug_cow()) { |
3b363692 MH |
1190 | huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0); |
1191 | new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER); | |
077fcf11 | 1192 | } else |
71e3aac0 AA |
1193 | new_page = NULL; |
1194 | ||
1195 | if (unlikely(!new_page)) { | |
eecc1e42 | 1196 | if (!page) { |
78ddc534 | 1197 | split_huge_pmd(vma, pmd, address); |
e9b71ca9 | 1198 | ret |= VM_FAULT_FALLBACK; |
93b4796d KS |
1199 | } else { |
1200 | ret = do_huge_pmd_wp_page_fallback(mm, vma, address, | |
1201 | pmd, orig_pmd, page, haddr); | |
9845cbbd | 1202 | if (ret & VM_FAULT_OOM) { |
78ddc534 | 1203 | split_huge_pmd(vma, pmd, address); |
9845cbbd KS |
1204 | ret |= VM_FAULT_FALLBACK; |
1205 | } | |
ddc58f27 | 1206 | put_page(page); |
93b4796d | 1207 | } |
17766dde | 1208 | count_vm_event(THP_FAULT_FALLBACK); |
71e3aac0 AA |
1209 | goto out; |
1210 | } | |
1211 | ||
f627c2f5 KS |
1212 | if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg, |
1213 | true))) { | |
b9bbfbe3 | 1214 | put_page(new_page); |
93b4796d | 1215 | if (page) { |
78ddc534 | 1216 | split_huge_pmd(vma, pmd, address); |
ddc58f27 | 1217 | put_page(page); |
9845cbbd | 1218 | } else |
78ddc534 | 1219 | split_huge_pmd(vma, pmd, address); |
9845cbbd | 1220 | ret |= VM_FAULT_FALLBACK; |
17766dde | 1221 | count_vm_event(THP_FAULT_FALLBACK); |
b9bbfbe3 AA |
1222 | goto out; |
1223 | } | |
1224 | ||
17766dde DR |
1225 | count_vm_event(THP_FAULT_ALLOC); |
1226 | ||
eecc1e42 | 1227 | if (!page) |
93b4796d KS |
1228 | clear_huge_page(new_page, haddr, HPAGE_PMD_NR); |
1229 | else | |
1230 | copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR); | |
71e3aac0 AA |
1231 | __SetPageUptodate(new_page); |
1232 | ||
2ec74c3e SG |
1233 | mmun_start = haddr; |
1234 | mmun_end = haddr + HPAGE_PMD_SIZE; | |
1235 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
1236 | ||
c4088ebd | 1237 | spin_lock(ptl); |
93b4796d | 1238 | if (page) |
ddc58f27 | 1239 | put_page(page); |
b9bbfbe3 | 1240 | if (unlikely(!pmd_same(*pmd, orig_pmd))) { |
c4088ebd | 1241 | spin_unlock(ptl); |
f627c2f5 | 1242 | mem_cgroup_cancel_charge(new_page, memcg, true); |
71e3aac0 | 1243 | put_page(new_page); |
2ec74c3e | 1244 | goto out_mn; |
b9bbfbe3 | 1245 | } else { |
71e3aac0 | 1246 | pmd_t entry; |
3122359a KS |
1247 | entry = mk_huge_pmd(new_page, vma->vm_page_prot); |
1248 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
8809aa2d | 1249 | pmdp_huge_clear_flush_notify(vma, haddr, pmd); |
d281ee61 | 1250 | page_add_new_anon_rmap(new_page, vma, haddr, true); |
f627c2f5 | 1251 | mem_cgroup_commit_charge(new_page, memcg, false, true); |
00501b53 | 1252 | lru_cache_add_active_or_unevictable(new_page, vma); |
71e3aac0 | 1253 | set_pmd_at(mm, haddr, pmd, entry); |
b113da65 | 1254 | update_mmu_cache_pmd(vma, address, pmd); |
eecc1e42 | 1255 | if (!page) { |
93b4796d | 1256 | add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); |
97ae1749 KS |
1257 | put_huge_zero_page(); |
1258 | } else { | |
309381fe | 1259 | VM_BUG_ON_PAGE(!PageHead(page), page); |
d281ee61 | 1260 | page_remove_rmap(page, true); |
93b4796d KS |
1261 | put_page(page); |
1262 | } | |
71e3aac0 AA |
1263 | ret |= VM_FAULT_WRITE; |
1264 | } | |
c4088ebd | 1265 | spin_unlock(ptl); |
2ec74c3e SG |
1266 | out_mn: |
1267 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); | |
71e3aac0 AA |
1268 | out: |
1269 | return ret; | |
2ec74c3e | 1270 | out_unlock: |
c4088ebd | 1271 | spin_unlock(ptl); |
2ec74c3e | 1272 | return ret; |
71e3aac0 AA |
1273 | } |
1274 | ||
b676b293 | 1275 | struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, |
71e3aac0 AA |
1276 | unsigned long addr, |
1277 | pmd_t *pmd, | |
1278 | unsigned int flags) | |
1279 | { | |
b676b293 | 1280 | struct mm_struct *mm = vma->vm_mm; |
71e3aac0 AA |
1281 | struct page *page = NULL; |
1282 | ||
c4088ebd | 1283 | assert_spin_locked(pmd_lockptr(mm, pmd)); |
71e3aac0 AA |
1284 | |
1285 | if (flags & FOLL_WRITE && !pmd_write(*pmd)) | |
1286 | goto out; | |
1287 | ||
85facf25 KS |
1288 | /* Avoid dumping huge zero page */ |
1289 | if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) | |
1290 | return ERR_PTR(-EFAULT); | |
1291 | ||
2b4847e7 | 1292 | /* Full NUMA hinting faults to serialise migration in fault paths */ |
8a0516ed | 1293 | if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) |
2b4847e7 MG |
1294 | goto out; |
1295 | ||
71e3aac0 | 1296 | page = pmd_page(*pmd); |
309381fe | 1297 | VM_BUG_ON_PAGE(!PageHead(page), page); |
71e3aac0 AA |
1298 | if (flags & FOLL_TOUCH) { |
1299 | pmd_t _pmd; | |
1300 | /* | |
1301 | * We should set the dirty bit only for FOLL_WRITE but | |
1302 | * for now the dirty bit in the pmd is meaningless. | |
1303 | * And if the dirty bit will become meaningful and | |
1304 | * we'll only set it with FOLL_WRITE, an atomic | |
1305 | * set_bit will be required on the pmd to set the | |
1306 | * young bit, instead of the current set_pmd_at. | |
1307 | */ | |
1308 | _pmd = pmd_mkyoung(pmd_mkdirty(*pmd)); | |
8663890a AK |
1309 | if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, |
1310 | pmd, _pmd, 1)) | |
1311 | update_mmu_cache_pmd(vma, addr, pmd); | |
71e3aac0 | 1312 | } |
de60f5f1 | 1313 | if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { |
b676b293 DR |
1314 | if (page->mapping && trylock_page(page)) { |
1315 | lru_add_drain(); | |
1316 | if (page->mapping) | |
1317 | mlock_vma_page(page); | |
1318 | unlock_page(page); | |
1319 | } | |
1320 | } | |
71e3aac0 | 1321 | page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; |
309381fe | 1322 | VM_BUG_ON_PAGE(!PageCompound(page), page); |
71e3aac0 | 1323 | if (flags & FOLL_GET) |
ddc58f27 | 1324 | get_page(page); |
71e3aac0 AA |
1325 | |
1326 | out: | |
1327 | return page; | |
1328 | } | |
1329 | ||
d10e63f2 | 1330 | /* NUMA hinting page fault entry point for trans huge pmds */ |
4daae3b4 MG |
1331 | int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma, |
1332 | unsigned long addr, pmd_t pmd, pmd_t *pmdp) | |
d10e63f2 | 1333 | { |
c4088ebd | 1334 | spinlock_t *ptl; |
b8916634 | 1335 | struct anon_vma *anon_vma = NULL; |
b32967ff | 1336 | struct page *page; |
d10e63f2 | 1337 | unsigned long haddr = addr & HPAGE_PMD_MASK; |
8191acbd | 1338 | int page_nid = -1, this_nid = numa_node_id(); |
90572890 | 1339 | int target_nid, last_cpupid = -1; |
8191acbd MG |
1340 | bool page_locked; |
1341 | bool migrated = false; | |
b191f9b1 | 1342 | bool was_writable; |
6688cc05 | 1343 | int flags = 0; |
d10e63f2 | 1344 | |
c0e7cad9 MG |
1345 | /* A PROT_NONE fault should not end up here */ |
1346 | BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))); | |
1347 | ||
c4088ebd | 1348 | ptl = pmd_lock(mm, pmdp); |
d10e63f2 MG |
1349 | if (unlikely(!pmd_same(pmd, *pmdp))) |
1350 | goto out_unlock; | |
1351 | ||
de466bd6 MG |
1352 | /* |
1353 | * If there are potential migrations, wait for completion and retry | |
1354 | * without disrupting NUMA hinting information. Do not relock and | |
1355 | * check_same as the page may no longer be mapped. | |
1356 | */ | |
1357 | if (unlikely(pmd_trans_migrating(*pmdp))) { | |
5d833062 | 1358 | page = pmd_page(*pmdp); |
de466bd6 | 1359 | spin_unlock(ptl); |
5d833062 | 1360 | wait_on_page_locked(page); |
de466bd6 MG |
1361 | goto out; |
1362 | } | |
1363 | ||
d10e63f2 | 1364 | page = pmd_page(pmd); |
a1a46184 | 1365 | BUG_ON(is_huge_zero_page(page)); |
8191acbd | 1366 | page_nid = page_to_nid(page); |
90572890 | 1367 | last_cpupid = page_cpupid_last(page); |
03c5a6e1 | 1368 | count_vm_numa_event(NUMA_HINT_FAULTS); |
04bb2f94 | 1369 | if (page_nid == this_nid) { |
03c5a6e1 | 1370 | count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); |
04bb2f94 RR |
1371 | flags |= TNF_FAULT_LOCAL; |
1372 | } | |
4daae3b4 | 1373 | |
bea66fbd MG |
1374 | /* See similar comment in do_numa_page for explanation */ |
1375 | if (!(vma->vm_flags & VM_WRITE)) | |
6688cc05 PZ |
1376 | flags |= TNF_NO_GROUP; |
1377 | ||
ff9042b1 MG |
1378 | /* |
1379 | * Acquire the page lock to serialise THP migrations but avoid dropping | |
1380 | * page_table_lock if at all possible | |
1381 | */ | |
b8916634 MG |
1382 | page_locked = trylock_page(page); |
1383 | target_nid = mpol_misplaced(page, vma, haddr); | |
1384 | if (target_nid == -1) { | |
1385 | /* If the page was locked, there are no parallel migrations */ | |
a54a407f | 1386 | if (page_locked) |
b8916634 | 1387 | goto clear_pmdnuma; |
2b4847e7 | 1388 | } |
4daae3b4 | 1389 | |
de466bd6 | 1390 | /* Migration could have started since the pmd_trans_migrating check */ |
2b4847e7 | 1391 | if (!page_locked) { |
c4088ebd | 1392 | spin_unlock(ptl); |
b8916634 | 1393 | wait_on_page_locked(page); |
a54a407f | 1394 | page_nid = -1; |
b8916634 MG |
1395 | goto out; |
1396 | } | |
1397 | ||
2b4847e7 MG |
1398 | /* |
1399 | * Page is misplaced. Page lock serialises migrations. Acquire anon_vma | |
1400 | * to serialises splits | |
1401 | */ | |
b8916634 | 1402 | get_page(page); |
c4088ebd | 1403 | spin_unlock(ptl); |
b8916634 | 1404 | anon_vma = page_lock_anon_vma_read(page); |
4daae3b4 | 1405 | |
c69307d5 | 1406 | /* Confirm the PMD did not change while page_table_lock was released */ |
c4088ebd | 1407 | spin_lock(ptl); |
b32967ff MG |
1408 | if (unlikely(!pmd_same(pmd, *pmdp))) { |
1409 | unlock_page(page); | |
1410 | put_page(page); | |
a54a407f | 1411 | page_nid = -1; |
4daae3b4 | 1412 | goto out_unlock; |
b32967ff | 1413 | } |
ff9042b1 | 1414 | |
c3a489ca MG |
1415 | /* Bail if we fail to protect against THP splits for any reason */ |
1416 | if (unlikely(!anon_vma)) { | |
1417 | put_page(page); | |
1418 | page_nid = -1; | |
1419 | goto clear_pmdnuma; | |
1420 | } | |
1421 | ||
a54a407f MG |
1422 | /* |
1423 | * Migrate the THP to the requested node, returns with page unlocked | |
8a0516ed | 1424 | * and access rights restored. |
a54a407f | 1425 | */ |
c4088ebd | 1426 | spin_unlock(ptl); |
b32967ff | 1427 | migrated = migrate_misplaced_transhuge_page(mm, vma, |
340ef390 | 1428 | pmdp, pmd, addr, page, target_nid); |
6688cc05 PZ |
1429 | if (migrated) { |
1430 | flags |= TNF_MIGRATED; | |
8191acbd | 1431 | page_nid = target_nid; |
074c2381 MG |
1432 | } else |
1433 | flags |= TNF_MIGRATE_FAIL; | |
b32967ff | 1434 | |
8191acbd | 1435 | goto out; |
b32967ff | 1436 | clear_pmdnuma: |
a54a407f | 1437 | BUG_ON(!PageLocked(page)); |
b191f9b1 | 1438 | was_writable = pmd_write(pmd); |
4d942466 | 1439 | pmd = pmd_modify(pmd, vma->vm_page_prot); |
b7b04004 | 1440 | pmd = pmd_mkyoung(pmd); |
b191f9b1 MG |
1441 | if (was_writable) |
1442 | pmd = pmd_mkwrite(pmd); | |
d10e63f2 | 1443 | set_pmd_at(mm, haddr, pmdp, pmd); |
d10e63f2 | 1444 | update_mmu_cache_pmd(vma, addr, pmdp); |
a54a407f | 1445 | unlock_page(page); |
d10e63f2 | 1446 | out_unlock: |
c4088ebd | 1447 | spin_unlock(ptl); |
b8916634 MG |
1448 | |
1449 | out: | |
1450 | if (anon_vma) | |
1451 | page_unlock_anon_vma_read(anon_vma); | |
1452 | ||
8191acbd | 1453 | if (page_nid != -1) |
6688cc05 | 1454 | task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags); |
8191acbd | 1455 | |
d10e63f2 MG |
1456 | return 0; |
1457 | } | |
1458 | ||
71e3aac0 | 1459 | int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, |
f21760b1 | 1460 | pmd_t *pmd, unsigned long addr) |
71e3aac0 | 1461 | { |
da146769 | 1462 | pmd_t orig_pmd; |
bf929152 | 1463 | spinlock_t *ptl; |
71e3aac0 | 1464 | |
4b471e88 | 1465 | if (!__pmd_trans_huge_lock(pmd, vma, &ptl)) |
da146769 KS |
1466 | return 0; |
1467 | /* | |
1468 | * For architectures like ppc64 we look at deposited pgtable | |
1469 | * when calling pmdp_huge_get_and_clear. So do the | |
1470 | * pgtable_trans_huge_withdraw after finishing pmdp related | |
1471 | * operations. | |
1472 | */ | |
1473 | orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd, | |
1474 | tlb->fullmm); | |
1475 | tlb_remove_pmd_tlb_entry(tlb, pmd, addr); | |
1476 | if (vma_is_dax(vma)) { | |
1477 | spin_unlock(ptl); | |
1478 | if (is_huge_zero_pmd(orig_pmd)) | |
97ae1749 | 1479 | put_huge_zero_page(); |
da146769 KS |
1480 | } else if (is_huge_zero_pmd(orig_pmd)) { |
1481 | pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd)); | |
1482 | atomic_long_dec(&tlb->mm->nr_ptes); | |
1483 | spin_unlock(ptl); | |
1484 | put_huge_zero_page(); | |
1485 | } else { | |
1486 | struct page *page = pmd_page(orig_pmd); | |
d281ee61 | 1487 | page_remove_rmap(page, true); |
da146769 KS |
1488 | VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); |
1489 | add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); | |
1490 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
1491 | pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd)); | |
1492 | atomic_long_dec(&tlb->mm->nr_ptes); | |
1493 | spin_unlock(ptl); | |
1494 | tlb_remove_page(tlb, page); | |
025c5b24 | 1495 | } |
da146769 | 1496 | return 1; |
71e3aac0 AA |
1497 | } |
1498 | ||
4b471e88 | 1499 | bool move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma, |
37a1c49a AA |
1500 | unsigned long old_addr, |
1501 | unsigned long new_addr, unsigned long old_end, | |
1502 | pmd_t *old_pmd, pmd_t *new_pmd) | |
1503 | { | |
bf929152 | 1504 | spinlock_t *old_ptl, *new_ptl; |
37a1c49a AA |
1505 | pmd_t pmd; |
1506 | ||
1507 | struct mm_struct *mm = vma->vm_mm; | |
1508 | ||
1509 | if ((old_addr & ~HPAGE_PMD_MASK) || | |
1510 | (new_addr & ~HPAGE_PMD_MASK) || | |
1511 | old_end - old_addr < HPAGE_PMD_SIZE || | |
1512 | (new_vma->vm_flags & VM_NOHUGEPAGE)) | |
4b471e88 | 1513 | return false; |
37a1c49a AA |
1514 | |
1515 | /* | |
1516 | * The destination pmd shouldn't be established, free_pgtables() | |
1517 | * should have release it. | |
1518 | */ | |
1519 | if (WARN_ON(!pmd_none(*new_pmd))) { | |
1520 | VM_BUG_ON(pmd_trans_huge(*new_pmd)); | |
4b471e88 | 1521 | return false; |
37a1c49a AA |
1522 | } |
1523 | ||
bf929152 KS |
1524 | /* |
1525 | * We don't have to worry about the ordering of src and dst | |
1526 | * ptlocks because exclusive mmap_sem prevents deadlock. | |
1527 | */ | |
4b471e88 | 1528 | if (__pmd_trans_huge_lock(old_pmd, vma, &old_ptl)) { |
bf929152 KS |
1529 | new_ptl = pmd_lockptr(mm, new_pmd); |
1530 | if (new_ptl != old_ptl) | |
1531 | spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); | |
8809aa2d | 1532 | pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); |
025c5b24 | 1533 | VM_BUG_ON(!pmd_none(*new_pmd)); |
3592806c | 1534 | |
b3084f4d AK |
1535 | if (pmd_move_must_withdraw(new_ptl, old_ptl)) { |
1536 | pgtable_t pgtable; | |
3592806c KS |
1537 | pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); |
1538 | pgtable_trans_huge_deposit(mm, new_pmd, pgtable); | |
3592806c | 1539 | } |
b3084f4d AK |
1540 | set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd)); |
1541 | if (new_ptl != old_ptl) | |
1542 | spin_unlock(new_ptl); | |
bf929152 | 1543 | spin_unlock(old_ptl); |
4b471e88 | 1544 | return true; |
37a1c49a | 1545 | } |
4b471e88 | 1546 | return false; |
37a1c49a AA |
1547 | } |
1548 | ||
f123d74a MG |
1549 | /* |
1550 | * Returns | |
1551 | * - 0 if PMD could not be locked | |
1552 | * - 1 if PMD was locked but protections unchange and TLB flush unnecessary | |
1553 | * - HPAGE_PMD_NR is protections changed and TLB flush necessary | |
1554 | */ | |
cd7548ab | 1555 | int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, |
e944fd67 | 1556 | unsigned long addr, pgprot_t newprot, int prot_numa) |
cd7548ab JW |
1557 | { |
1558 | struct mm_struct *mm = vma->vm_mm; | |
bf929152 | 1559 | spinlock_t *ptl; |
cd7548ab JW |
1560 | int ret = 0; |
1561 | ||
4b471e88 | 1562 | if (__pmd_trans_huge_lock(pmd, vma, &ptl)) { |
025c5b24 | 1563 | pmd_t entry; |
b191f9b1 | 1564 | bool preserve_write = prot_numa && pmd_write(*pmd); |
ba68bc01 | 1565 | ret = 1; |
e944fd67 MG |
1566 | |
1567 | /* | |
1568 | * Avoid trapping faults against the zero page. The read-only | |
1569 | * data is likely to be read-cached on the local CPU and | |
1570 | * local/remote hits to the zero page are not interesting. | |
1571 | */ | |
1572 | if (prot_numa && is_huge_zero_pmd(*pmd)) { | |
1573 | spin_unlock(ptl); | |
ba68bc01 | 1574 | return ret; |
e944fd67 MG |
1575 | } |
1576 | ||
10c1045f | 1577 | if (!prot_numa || !pmd_protnone(*pmd)) { |
8809aa2d | 1578 | entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd); |
10c1045f | 1579 | entry = pmd_modify(entry, newprot); |
b191f9b1 MG |
1580 | if (preserve_write) |
1581 | entry = pmd_mkwrite(entry); | |
10c1045f MG |
1582 | ret = HPAGE_PMD_NR; |
1583 | set_pmd_at(mm, addr, pmd, entry); | |
b191f9b1 | 1584 | BUG_ON(!preserve_write && pmd_write(entry)); |
10c1045f | 1585 | } |
bf929152 | 1586 | spin_unlock(ptl); |
025c5b24 NH |
1587 | } |
1588 | ||
1589 | return ret; | |
1590 | } | |
1591 | ||
1592 | /* | |
4b471e88 | 1593 | * Returns true if a given pmd maps a thp, false otherwise. |
025c5b24 | 1594 | * |
4b471e88 KS |
1595 | * Note that if it returns true, this routine returns without unlocking page |
1596 | * table lock. So callers must unlock it. | |
025c5b24 | 1597 | */ |
4b471e88 | 1598 | bool __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma, |
bf929152 | 1599 | spinlock_t **ptl) |
025c5b24 | 1600 | { |
bf929152 | 1601 | *ptl = pmd_lock(vma->vm_mm, pmd); |
4b471e88 KS |
1602 | if (likely(pmd_trans_huge(*pmd))) |
1603 | return true; | |
bf929152 | 1604 | spin_unlock(*ptl); |
4b471e88 | 1605 | return false; |
cd7548ab JW |
1606 | } |
1607 | ||
117b0791 KS |
1608 | /* |
1609 | * This function returns whether a given @page is mapped onto the @address | |
1610 | * in the virtual space of @mm. | |
1611 | * | |
1612 | * When it's true, this function returns *pmd with holding the page table lock | |
1613 | * and passing it back to the caller via @ptl. | |
1614 | * If it's false, returns NULL without holding the page table lock. | |
1615 | */ | |
71e3aac0 AA |
1616 | pmd_t *page_check_address_pmd(struct page *page, |
1617 | struct mm_struct *mm, | |
1618 | unsigned long address, | |
117b0791 | 1619 | spinlock_t **ptl) |
71e3aac0 | 1620 | { |
b5a8cad3 KS |
1621 | pgd_t *pgd; |
1622 | pud_t *pud; | |
117b0791 | 1623 | pmd_t *pmd; |
71e3aac0 AA |
1624 | |
1625 | if (address & ~HPAGE_PMD_MASK) | |
117b0791 | 1626 | return NULL; |
71e3aac0 | 1627 | |
b5a8cad3 KS |
1628 | pgd = pgd_offset(mm, address); |
1629 | if (!pgd_present(*pgd)) | |
117b0791 | 1630 | return NULL; |
b5a8cad3 KS |
1631 | pud = pud_offset(pgd, address); |
1632 | if (!pud_present(*pud)) | |
1633 | return NULL; | |
1634 | pmd = pmd_offset(pud, address); | |
1635 | ||
117b0791 | 1636 | *ptl = pmd_lock(mm, pmd); |
b5a8cad3 | 1637 | if (!pmd_present(*pmd)) |
117b0791 | 1638 | goto unlock; |
71e3aac0 | 1639 | if (pmd_page(*pmd) != page) |
117b0791 | 1640 | goto unlock; |
4b471e88 | 1641 | if (pmd_trans_huge(*pmd)) |
117b0791 | 1642 | return pmd; |
117b0791 KS |
1643 | unlock: |
1644 | spin_unlock(*ptl); | |
1645 | return NULL; | |
71e3aac0 AA |
1646 | } |
1647 | ||
9050d7eb | 1648 | #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE) |
78f11a25 | 1649 | |
60ab3244 AA |
1650 | int hugepage_madvise(struct vm_area_struct *vma, |
1651 | unsigned long *vm_flags, int advice) | |
0af4e98b | 1652 | { |
a664b2d8 AA |
1653 | switch (advice) { |
1654 | case MADV_HUGEPAGE: | |
1e1836e8 AT |
1655 | #ifdef CONFIG_S390 |
1656 | /* | |
1657 | * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390 | |
1658 | * can't handle this properly after s390_enable_sie, so we simply | |
1659 | * ignore the madvise to prevent qemu from causing a SIGSEGV. | |
1660 | */ | |
1661 | if (mm_has_pgste(vma->vm_mm)) | |
1662 | return 0; | |
1663 | #endif | |
a664b2d8 AA |
1664 | /* |
1665 | * Be somewhat over-protective like KSM for now! | |
1666 | */ | |
1a763615 | 1667 | if (*vm_flags & VM_NO_THP) |
a664b2d8 AA |
1668 | return -EINVAL; |
1669 | *vm_flags &= ~VM_NOHUGEPAGE; | |
1670 | *vm_flags |= VM_HUGEPAGE; | |
60ab3244 AA |
1671 | /* |
1672 | * If the vma become good for khugepaged to scan, | |
1673 | * register it here without waiting a page fault that | |
1674 | * may not happen any time soon. | |
1675 | */ | |
6d50e60c | 1676 | if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags))) |
60ab3244 | 1677 | return -ENOMEM; |
a664b2d8 AA |
1678 | break; |
1679 | case MADV_NOHUGEPAGE: | |
1680 | /* | |
1681 | * Be somewhat over-protective like KSM for now! | |
1682 | */ | |
1a763615 | 1683 | if (*vm_flags & VM_NO_THP) |
a664b2d8 AA |
1684 | return -EINVAL; |
1685 | *vm_flags &= ~VM_HUGEPAGE; | |
1686 | *vm_flags |= VM_NOHUGEPAGE; | |
60ab3244 AA |
1687 | /* |
1688 | * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning | |
1689 | * this vma even if we leave the mm registered in khugepaged if | |
1690 | * it got registered before VM_NOHUGEPAGE was set. | |
1691 | */ | |
a664b2d8 AA |
1692 | break; |
1693 | } | |
0af4e98b AA |
1694 | |
1695 | return 0; | |
1696 | } | |
1697 | ||
ba76149f AA |
1698 | static int __init khugepaged_slab_init(void) |
1699 | { | |
1700 | mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", | |
1701 | sizeof(struct mm_slot), | |
1702 | __alignof__(struct mm_slot), 0, NULL); | |
1703 | if (!mm_slot_cache) | |
1704 | return -ENOMEM; | |
1705 | ||
1706 | return 0; | |
1707 | } | |
1708 | ||
65ebb64f KS |
1709 | static void __init khugepaged_slab_exit(void) |
1710 | { | |
1711 | kmem_cache_destroy(mm_slot_cache); | |
1712 | } | |
1713 | ||
ba76149f AA |
1714 | static inline struct mm_slot *alloc_mm_slot(void) |
1715 | { | |
1716 | if (!mm_slot_cache) /* initialization failed */ | |
1717 | return NULL; | |
1718 | return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); | |
1719 | } | |
1720 | ||
1721 | static inline void free_mm_slot(struct mm_slot *mm_slot) | |
1722 | { | |
1723 | kmem_cache_free(mm_slot_cache, mm_slot); | |
1724 | } | |
1725 | ||
ba76149f AA |
1726 | static struct mm_slot *get_mm_slot(struct mm_struct *mm) |
1727 | { | |
1728 | struct mm_slot *mm_slot; | |
ba76149f | 1729 | |
b67bfe0d | 1730 | hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm) |
ba76149f AA |
1731 | if (mm == mm_slot->mm) |
1732 | return mm_slot; | |
43b5fbbd | 1733 | |
ba76149f AA |
1734 | return NULL; |
1735 | } | |
1736 | ||
1737 | static void insert_to_mm_slots_hash(struct mm_struct *mm, | |
1738 | struct mm_slot *mm_slot) | |
1739 | { | |
ba76149f | 1740 | mm_slot->mm = mm; |
43b5fbbd | 1741 | hash_add(mm_slots_hash, &mm_slot->hash, (long)mm); |
ba76149f AA |
1742 | } |
1743 | ||
1744 | static inline int khugepaged_test_exit(struct mm_struct *mm) | |
1745 | { | |
1746 | return atomic_read(&mm->mm_users) == 0; | |
1747 | } | |
1748 | ||
1749 | int __khugepaged_enter(struct mm_struct *mm) | |
1750 | { | |
1751 | struct mm_slot *mm_slot; | |
1752 | int wakeup; | |
1753 | ||
1754 | mm_slot = alloc_mm_slot(); | |
1755 | if (!mm_slot) | |
1756 | return -ENOMEM; | |
1757 | ||
1758 | /* __khugepaged_exit() must not run from under us */ | |
96dad67f | 1759 | VM_BUG_ON_MM(khugepaged_test_exit(mm), mm); |
ba76149f AA |
1760 | if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { |
1761 | free_mm_slot(mm_slot); | |
1762 | return 0; | |
1763 | } | |
1764 | ||
1765 | spin_lock(&khugepaged_mm_lock); | |
1766 | insert_to_mm_slots_hash(mm, mm_slot); | |
1767 | /* | |
1768 | * Insert just behind the scanning cursor, to let the area settle | |
1769 | * down a little. | |
1770 | */ | |
1771 | wakeup = list_empty(&khugepaged_scan.mm_head); | |
1772 | list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head); | |
1773 | spin_unlock(&khugepaged_mm_lock); | |
1774 | ||
1775 | atomic_inc(&mm->mm_count); | |
1776 | if (wakeup) | |
1777 | wake_up_interruptible(&khugepaged_wait); | |
1778 | ||
1779 | return 0; | |
1780 | } | |
1781 | ||
6d50e60c DR |
1782 | int khugepaged_enter_vma_merge(struct vm_area_struct *vma, |
1783 | unsigned long vm_flags) | |
ba76149f AA |
1784 | { |
1785 | unsigned long hstart, hend; | |
1786 | if (!vma->anon_vma) | |
1787 | /* | |
1788 | * Not yet faulted in so we will register later in the | |
1789 | * page fault if needed. | |
1790 | */ | |
1791 | return 0; | |
78f11a25 | 1792 | if (vma->vm_ops) |
ba76149f AA |
1793 | /* khugepaged not yet working on file or special mappings */ |
1794 | return 0; | |
6d50e60c | 1795 | VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma); |
ba76149f AA |
1796 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; |
1797 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
1798 | if (hstart < hend) | |
6d50e60c | 1799 | return khugepaged_enter(vma, vm_flags); |
ba76149f AA |
1800 | return 0; |
1801 | } | |
1802 | ||
1803 | void __khugepaged_exit(struct mm_struct *mm) | |
1804 | { | |
1805 | struct mm_slot *mm_slot; | |
1806 | int free = 0; | |
1807 | ||
1808 | spin_lock(&khugepaged_mm_lock); | |
1809 | mm_slot = get_mm_slot(mm); | |
1810 | if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { | |
43b5fbbd | 1811 | hash_del(&mm_slot->hash); |
ba76149f AA |
1812 | list_del(&mm_slot->mm_node); |
1813 | free = 1; | |
1814 | } | |
d788e80a | 1815 | spin_unlock(&khugepaged_mm_lock); |
ba76149f AA |
1816 | |
1817 | if (free) { | |
ba76149f AA |
1818 | clear_bit(MMF_VM_HUGEPAGE, &mm->flags); |
1819 | free_mm_slot(mm_slot); | |
1820 | mmdrop(mm); | |
1821 | } else if (mm_slot) { | |
ba76149f AA |
1822 | /* |
1823 | * This is required to serialize against | |
1824 | * khugepaged_test_exit() (which is guaranteed to run | |
1825 | * under mmap sem read mode). Stop here (after we | |
1826 | * return all pagetables will be destroyed) until | |
1827 | * khugepaged has finished working on the pagetables | |
1828 | * under the mmap_sem. | |
1829 | */ | |
1830 | down_write(&mm->mmap_sem); | |
1831 | up_write(&mm->mmap_sem); | |
d788e80a | 1832 | } |
ba76149f AA |
1833 | } |
1834 | ||
1835 | static void release_pte_page(struct page *page) | |
1836 | { | |
1837 | /* 0 stands for page_is_file_cache(page) == false */ | |
1838 | dec_zone_page_state(page, NR_ISOLATED_ANON + 0); | |
1839 | unlock_page(page); | |
1840 | putback_lru_page(page); | |
1841 | } | |
1842 | ||
1843 | static void release_pte_pages(pte_t *pte, pte_t *_pte) | |
1844 | { | |
1845 | while (--_pte >= pte) { | |
1846 | pte_t pteval = *_pte; | |
ca0984ca | 1847 | if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval))) |
ba76149f AA |
1848 | release_pte_page(pte_page(pteval)); |
1849 | } | |
1850 | } | |
1851 | ||
ba76149f AA |
1852 | static int __collapse_huge_page_isolate(struct vm_area_struct *vma, |
1853 | unsigned long address, | |
1854 | pte_t *pte) | |
1855 | { | |
7d2eba05 | 1856 | struct page *page = NULL; |
ba76149f | 1857 | pte_t *_pte; |
7d2eba05 | 1858 | int none_or_zero = 0, result = 0; |
10359213 | 1859 | bool referenced = false, writable = false; |
7d2eba05 | 1860 | |
ba76149f AA |
1861 | for (_pte = pte; _pte < pte+HPAGE_PMD_NR; |
1862 | _pte++, address += PAGE_SIZE) { | |
1863 | pte_t pteval = *_pte; | |
47aee4d8 MK |
1864 | if (pte_none(pteval) || (pte_present(pteval) && |
1865 | is_zero_pfn(pte_pfn(pteval)))) { | |
c1294d05 | 1866 | if (!userfaultfd_armed(vma) && |
7d2eba05 | 1867 | ++none_or_zero <= khugepaged_max_ptes_none) { |
ba76149f | 1868 | continue; |
7d2eba05 EA |
1869 | } else { |
1870 | result = SCAN_EXCEED_NONE_PTE; | |
ba76149f | 1871 | goto out; |
7d2eba05 | 1872 | } |
ba76149f | 1873 | } |
7d2eba05 EA |
1874 | if (!pte_present(pteval)) { |
1875 | result = SCAN_PTE_NON_PRESENT; | |
ba76149f | 1876 | goto out; |
7d2eba05 | 1877 | } |
ba76149f | 1878 | page = vm_normal_page(vma, address, pteval); |
7d2eba05 EA |
1879 | if (unlikely(!page)) { |
1880 | result = SCAN_PAGE_NULL; | |
ba76149f | 1881 | goto out; |
7d2eba05 | 1882 | } |
344aa35c | 1883 | |
309381fe SL |
1884 | VM_BUG_ON_PAGE(PageCompound(page), page); |
1885 | VM_BUG_ON_PAGE(!PageAnon(page), page); | |
1886 | VM_BUG_ON_PAGE(!PageSwapBacked(page), page); | |
ba76149f | 1887 | |
ba76149f AA |
1888 | /* |
1889 | * We can do it before isolate_lru_page because the | |
1890 | * page can't be freed from under us. NOTE: PG_lock | |
1891 | * is needed to serialize against split_huge_page | |
1892 | * when invoked from the VM. | |
1893 | */ | |
7d2eba05 EA |
1894 | if (!trylock_page(page)) { |
1895 | result = SCAN_PAGE_LOCK; | |
ba76149f | 1896 | goto out; |
7d2eba05 | 1897 | } |
10359213 EA |
1898 | |
1899 | /* | |
1900 | * cannot use mapcount: can't collapse if there's a gup pin. | |
1901 | * The page must only be referenced by the scanned process | |
1902 | * and page swap cache. | |
1903 | */ | |
1904 | if (page_count(page) != 1 + !!PageSwapCache(page)) { | |
1905 | unlock_page(page); | |
7d2eba05 | 1906 | result = SCAN_PAGE_COUNT; |
10359213 EA |
1907 | goto out; |
1908 | } | |
1909 | if (pte_write(pteval)) { | |
1910 | writable = true; | |
1911 | } else { | |
1912 | if (PageSwapCache(page) && !reuse_swap_page(page)) { | |
1913 | unlock_page(page); | |
7d2eba05 | 1914 | result = SCAN_SWAP_CACHE_PAGE; |
10359213 EA |
1915 | goto out; |
1916 | } | |
1917 | /* | |
1918 | * Page is not in the swap cache. It can be collapsed | |
1919 | * into a THP. | |
1920 | */ | |
1921 | } | |
1922 | ||
ba76149f AA |
1923 | /* |
1924 | * Isolate the page to avoid collapsing an hugepage | |
1925 | * currently in use by the VM. | |
1926 | */ | |
1927 | if (isolate_lru_page(page)) { | |
1928 | unlock_page(page); | |
7d2eba05 | 1929 | result = SCAN_DEL_PAGE_LRU; |
ba76149f AA |
1930 | goto out; |
1931 | } | |
1932 | /* 0 stands for page_is_file_cache(page) == false */ | |
1933 | inc_zone_page_state(page, NR_ISOLATED_ANON + 0); | |
309381fe SL |
1934 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
1935 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
ba76149f AA |
1936 | |
1937 | /* If there is no mapped pte young don't collapse the page */ | |
33c3fc71 VD |
1938 | if (pte_young(pteval) || |
1939 | page_is_young(page) || PageReferenced(page) || | |
8ee53820 | 1940 | mmu_notifier_test_young(vma->vm_mm, address)) |
10359213 | 1941 | referenced = true; |
ba76149f | 1942 | } |
7d2eba05 EA |
1943 | if (likely(writable)) { |
1944 | if (likely(referenced)) { | |
1945 | result = SCAN_SUCCEED; | |
1946 | trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero, | |
1947 | referenced, writable, result); | |
1948 | return 1; | |
1949 | } | |
1950 | } else { | |
1951 | result = SCAN_PAGE_RO; | |
1952 | } | |
1953 | ||
ba76149f | 1954 | out: |
344aa35c | 1955 | release_pte_pages(pte, _pte); |
7d2eba05 EA |
1956 | trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero, |
1957 | referenced, writable, result); | |
344aa35c | 1958 | return 0; |
ba76149f AA |
1959 | } |
1960 | ||
1961 | static void __collapse_huge_page_copy(pte_t *pte, struct page *page, | |
1962 | struct vm_area_struct *vma, | |
1963 | unsigned long address, | |
1964 | spinlock_t *ptl) | |
1965 | { | |
1966 | pte_t *_pte; | |
1967 | for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) { | |
1968 | pte_t pteval = *_pte; | |
1969 | struct page *src_page; | |
1970 | ||
ca0984ca | 1971 | if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { |
ba76149f AA |
1972 | clear_user_highpage(page, address); |
1973 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); | |
ca0984ca EA |
1974 | if (is_zero_pfn(pte_pfn(pteval))) { |
1975 | /* | |
1976 | * ptl mostly unnecessary. | |
1977 | */ | |
1978 | spin_lock(ptl); | |
1979 | /* | |
1980 | * paravirt calls inside pte_clear here are | |
1981 | * superfluous. | |
1982 | */ | |
1983 | pte_clear(vma->vm_mm, address, _pte); | |
1984 | spin_unlock(ptl); | |
1985 | } | |
ba76149f AA |
1986 | } else { |
1987 | src_page = pte_page(pteval); | |
1988 | copy_user_highpage(page, src_page, address, vma); | |
309381fe | 1989 | VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page); |
ba76149f AA |
1990 | release_pte_page(src_page); |
1991 | /* | |
1992 | * ptl mostly unnecessary, but preempt has to | |
1993 | * be disabled to update the per-cpu stats | |
1994 | * inside page_remove_rmap(). | |
1995 | */ | |
1996 | spin_lock(ptl); | |
1997 | /* | |
1998 | * paravirt calls inside pte_clear here are | |
1999 | * superfluous. | |
2000 | */ | |
2001 | pte_clear(vma->vm_mm, address, _pte); | |
d281ee61 | 2002 | page_remove_rmap(src_page, false); |
ba76149f AA |
2003 | spin_unlock(ptl); |
2004 | free_page_and_swap_cache(src_page); | |
2005 | } | |
2006 | ||
2007 | address += PAGE_SIZE; | |
2008 | page++; | |
2009 | } | |
2010 | } | |
2011 | ||
26234f36 | 2012 | static void khugepaged_alloc_sleep(void) |
ba76149f | 2013 | { |
bde43c6c PM |
2014 | DEFINE_WAIT(wait); |
2015 | ||
2016 | add_wait_queue(&khugepaged_wait, &wait); | |
2017 | freezable_schedule_timeout_interruptible( | |
2018 | msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); | |
2019 | remove_wait_queue(&khugepaged_wait, &wait); | |
26234f36 | 2020 | } |
ba76149f | 2021 | |
9f1b868a BL |
2022 | static int khugepaged_node_load[MAX_NUMNODES]; |
2023 | ||
14a4e214 DR |
2024 | static bool khugepaged_scan_abort(int nid) |
2025 | { | |
2026 | int i; | |
2027 | ||
2028 | /* | |
2029 | * If zone_reclaim_mode is disabled, then no extra effort is made to | |
2030 | * allocate memory locally. | |
2031 | */ | |
2032 | if (!zone_reclaim_mode) | |
2033 | return false; | |
2034 | ||
2035 | /* If there is a count for this node already, it must be acceptable */ | |
2036 | if (khugepaged_node_load[nid]) | |
2037 | return false; | |
2038 | ||
2039 | for (i = 0; i < MAX_NUMNODES; i++) { | |
2040 | if (!khugepaged_node_load[i]) | |
2041 | continue; | |
2042 | if (node_distance(nid, i) > RECLAIM_DISTANCE) | |
2043 | return true; | |
2044 | } | |
2045 | return false; | |
2046 | } | |
2047 | ||
26234f36 | 2048 | #ifdef CONFIG_NUMA |
9f1b868a BL |
2049 | static int khugepaged_find_target_node(void) |
2050 | { | |
2051 | static int last_khugepaged_target_node = NUMA_NO_NODE; | |
2052 | int nid, target_node = 0, max_value = 0; | |
2053 | ||
2054 | /* find first node with max normal pages hit */ | |
2055 | for (nid = 0; nid < MAX_NUMNODES; nid++) | |
2056 | if (khugepaged_node_load[nid] > max_value) { | |
2057 | max_value = khugepaged_node_load[nid]; | |
2058 | target_node = nid; | |
2059 | } | |
2060 | ||
2061 | /* do some balance if several nodes have the same hit record */ | |
2062 | if (target_node <= last_khugepaged_target_node) | |
2063 | for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES; | |
2064 | nid++) | |
2065 | if (max_value == khugepaged_node_load[nid]) { | |
2066 | target_node = nid; | |
2067 | break; | |
2068 | } | |
2069 | ||
2070 | last_khugepaged_target_node = target_node; | |
2071 | return target_node; | |
2072 | } | |
2073 | ||
26234f36 XG |
2074 | static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) |
2075 | { | |
2076 | if (IS_ERR(*hpage)) { | |
2077 | if (!*wait) | |
2078 | return false; | |
2079 | ||
2080 | *wait = false; | |
e3b4126c | 2081 | *hpage = NULL; |
26234f36 XG |
2082 | khugepaged_alloc_sleep(); |
2083 | } else if (*hpage) { | |
2084 | put_page(*hpage); | |
2085 | *hpage = NULL; | |
2086 | } | |
2087 | ||
2088 | return true; | |
2089 | } | |
2090 | ||
3b363692 MH |
2091 | static struct page * |
2092 | khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm, | |
d6669d68 | 2093 | unsigned long address, int node) |
26234f36 | 2094 | { |
309381fe | 2095 | VM_BUG_ON_PAGE(*hpage, *hpage); |
8b164568 | 2096 | |
ce83d217 | 2097 | /* |
8b164568 VB |
2098 | * Before allocating the hugepage, release the mmap_sem read lock. |
2099 | * The allocation can take potentially a long time if it involves | |
2100 | * sync compaction, and we do not need to hold the mmap_sem during | |
2101 | * that. We will recheck the vma after taking it again in write mode. | |
ce83d217 | 2102 | */ |
8b164568 VB |
2103 | up_read(&mm->mmap_sem); |
2104 | ||
96db800f | 2105 | *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER); |
26234f36 | 2106 | if (unlikely(!*hpage)) { |
81ab4201 | 2107 | count_vm_event(THP_COLLAPSE_ALLOC_FAILED); |
ce83d217 | 2108 | *hpage = ERR_PTR(-ENOMEM); |
26234f36 | 2109 | return NULL; |
ce83d217 | 2110 | } |
26234f36 | 2111 | |
65b3c07b | 2112 | count_vm_event(THP_COLLAPSE_ALLOC); |
26234f36 XG |
2113 | return *hpage; |
2114 | } | |
2115 | #else | |
9f1b868a BL |
2116 | static int khugepaged_find_target_node(void) |
2117 | { | |
2118 | return 0; | |
2119 | } | |
2120 | ||
10dc4155 BL |
2121 | static inline struct page *alloc_hugepage(int defrag) |
2122 | { | |
2123 | return alloc_pages(alloc_hugepage_gfpmask(defrag, 0), | |
2124 | HPAGE_PMD_ORDER); | |
2125 | } | |
2126 | ||
26234f36 XG |
2127 | static struct page *khugepaged_alloc_hugepage(bool *wait) |
2128 | { | |
2129 | struct page *hpage; | |
2130 | ||
2131 | do { | |
2132 | hpage = alloc_hugepage(khugepaged_defrag()); | |
2133 | if (!hpage) { | |
2134 | count_vm_event(THP_COLLAPSE_ALLOC_FAILED); | |
2135 | if (!*wait) | |
2136 | return NULL; | |
2137 | ||
2138 | *wait = false; | |
2139 | khugepaged_alloc_sleep(); | |
2140 | } else | |
2141 | count_vm_event(THP_COLLAPSE_ALLOC); | |
2142 | } while (unlikely(!hpage) && likely(khugepaged_enabled())); | |
2143 | ||
2144 | return hpage; | |
2145 | } | |
2146 | ||
2147 | static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) | |
2148 | { | |
2149 | if (!*hpage) | |
2150 | *hpage = khugepaged_alloc_hugepage(wait); | |
2151 | ||
2152 | if (unlikely(!*hpage)) | |
2153 | return false; | |
2154 | ||
2155 | return true; | |
2156 | } | |
2157 | ||
3b363692 MH |
2158 | static struct page * |
2159 | khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm, | |
d6669d68 | 2160 | unsigned long address, int node) |
26234f36 XG |
2161 | { |
2162 | up_read(&mm->mmap_sem); | |
2163 | VM_BUG_ON(!*hpage); | |
3b363692 | 2164 | |
26234f36 XG |
2165 | return *hpage; |
2166 | } | |
692e0b35 AA |
2167 | #endif |
2168 | ||
fa475e51 BL |
2169 | static bool hugepage_vma_check(struct vm_area_struct *vma) |
2170 | { | |
2171 | if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) || | |
2172 | (vma->vm_flags & VM_NOHUGEPAGE)) | |
2173 | return false; | |
7479df6d KS |
2174 | if (vma->vm_flags & VM_LOCKED) |
2175 | return false; | |
fa475e51 BL |
2176 | if (!vma->anon_vma || vma->vm_ops) |
2177 | return false; | |
2178 | if (is_vma_temporary_stack(vma)) | |
2179 | return false; | |
81d1b09c | 2180 | VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma); |
fa475e51 BL |
2181 | return true; |
2182 | } | |
2183 | ||
26234f36 XG |
2184 | static void collapse_huge_page(struct mm_struct *mm, |
2185 | unsigned long address, | |
2186 | struct page **hpage, | |
2187 | struct vm_area_struct *vma, | |
2188 | int node) | |
2189 | { | |
26234f36 XG |
2190 | pmd_t *pmd, _pmd; |
2191 | pte_t *pte; | |
2192 | pgtable_t pgtable; | |
2193 | struct page *new_page; | |
c4088ebd | 2194 | spinlock_t *pmd_ptl, *pte_ptl; |
7d2eba05 | 2195 | int isolated, result = 0; |
26234f36 | 2196 | unsigned long hstart, hend; |
00501b53 | 2197 | struct mem_cgroup *memcg; |
2ec74c3e SG |
2198 | unsigned long mmun_start; /* For mmu_notifiers */ |
2199 | unsigned long mmun_end; /* For mmu_notifiers */ | |
3b363692 | 2200 | gfp_t gfp; |
26234f36 XG |
2201 | |
2202 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
2203 | ||
3b363692 MH |
2204 | /* Only allocate from the target node */ |
2205 | gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) | | |
2206 | __GFP_THISNODE; | |
2207 | ||
26234f36 | 2208 | /* release the mmap_sem read lock. */ |
d6669d68 | 2209 | new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node); |
7d2eba05 EA |
2210 | if (!new_page) { |
2211 | result = SCAN_ALLOC_HUGE_PAGE_FAIL; | |
2212 | goto out_nolock; | |
2213 | } | |
26234f36 | 2214 | |
f627c2f5 | 2215 | if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) { |
7d2eba05 EA |
2216 | result = SCAN_CGROUP_CHARGE_FAIL; |
2217 | goto out_nolock; | |
2218 | } | |
ba76149f AA |
2219 | |
2220 | /* | |
2221 | * Prevent all access to pagetables with the exception of | |
2222 | * gup_fast later hanlded by the ptep_clear_flush and the VM | |
2223 | * handled by the anon_vma lock + PG_lock. | |
2224 | */ | |
2225 | down_write(&mm->mmap_sem); | |
7d2eba05 EA |
2226 | if (unlikely(khugepaged_test_exit(mm))) { |
2227 | result = SCAN_ANY_PROCESS; | |
ba76149f | 2228 | goto out; |
7d2eba05 | 2229 | } |
ba76149f AA |
2230 | |
2231 | vma = find_vma(mm, address); | |
7d2eba05 EA |
2232 | if (!vma) { |
2233 | result = SCAN_VMA_NULL; | |
a8f531eb | 2234 | goto out; |
7d2eba05 | 2235 | } |
ba76149f AA |
2236 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; |
2237 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
7d2eba05 EA |
2238 | if (address < hstart || address + HPAGE_PMD_SIZE > hend) { |
2239 | result = SCAN_ADDRESS_RANGE; | |
ba76149f | 2240 | goto out; |
7d2eba05 EA |
2241 | } |
2242 | if (!hugepage_vma_check(vma)) { | |
2243 | result = SCAN_VMA_CHECK; | |
a7d6e4ec | 2244 | goto out; |
7d2eba05 | 2245 | } |
6219049a | 2246 | pmd = mm_find_pmd(mm, address); |
7d2eba05 EA |
2247 | if (!pmd) { |
2248 | result = SCAN_PMD_NULL; | |
ba76149f | 2249 | goto out; |
7d2eba05 | 2250 | } |
ba76149f | 2251 | |
4fc3f1d6 | 2252 | anon_vma_lock_write(vma->anon_vma); |
ba76149f AA |
2253 | |
2254 | pte = pte_offset_map(pmd, address); | |
c4088ebd | 2255 | pte_ptl = pte_lockptr(mm, pmd); |
ba76149f | 2256 | |
2ec74c3e SG |
2257 | mmun_start = address; |
2258 | mmun_end = address + HPAGE_PMD_SIZE; | |
2259 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
c4088ebd | 2260 | pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ |
ba76149f AA |
2261 | /* |
2262 | * After this gup_fast can't run anymore. This also removes | |
2263 | * any huge TLB entry from the CPU so we won't allow | |
2264 | * huge and small TLB entries for the same virtual address | |
2265 | * to avoid the risk of CPU bugs in that area. | |
2266 | */ | |
15a25b2e | 2267 | _pmd = pmdp_collapse_flush(vma, address, pmd); |
c4088ebd | 2268 | spin_unlock(pmd_ptl); |
2ec74c3e | 2269 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
ba76149f | 2270 | |
c4088ebd | 2271 | spin_lock(pte_ptl); |
ba76149f | 2272 | isolated = __collapse_huge_page_isolate(vma, address, pte); |
c4088ebd | 2273 | spin_unlock(pte_ptl); |
ba76149f AA |
2274 | |
2275 | if (unlikely(!isolated)) { | |
453c7192 | 2276 | pte_unmap(pte); |
c4088ebd | 2277 | spin_lock(pmd_ptl); |
ba76149f | 2278 | BUG_ON(!pmd_none(*pmd)); |
7c342512 AK |
2279 | /* |
2280 | * We can only use set_pmd_at when establishing | |
2281 | * hugepmds and never for establishing regular pmds that | |
2282 | * points to regular pagetables. Use pmd_populate for that | |
2283 | */ | |
2284 | pmd_populate(mm, pmd, pmd_pgtable(_pmd)); | |
c4088ebd | 2285 | spin_unlock(pmd_ptl); |
08b52706 | 2286 | anon_vma_unlock_write(vma->anon_vma); |
7d2eba05 | 2287 | result = SCAN_FAIL; |
ce83d217 | 2288 | goto out; |
ba76149f AA |
2289 | } |
2290 | ||
2291 | /* | |
2292 | * All pages are isolated and locked so anon_vma rmap | |
2293 | * can't run anymore. | |
2294 | */ | |
08b52706 | 2295 | anon_vma_unlock_write(vma->anon_vma); |
ba76149f | 2296 | |
c4088ebd | 2297 | __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl); |
453c7192 | 2298 | pte_unmap(pte); |
ba76149f AA |
2299 | __SetPageUptodate(new_page); |
2300 | pgtable = pmd_pgtable(_pmd); | |
ba76149f | 2301 | |
3122359a KS |
2302 | _pmd = mk_huge_pmd(new_page, vma->vm_page_prot); |
2303 | _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); | |
ba76149f AA |
2304 | |
2305 | /* | |
2306 | * spin_lock() below is not the equivalent of smp_wmb(), so | |
2307 | * this is needed to avoid the copy_huge_page writes to become | |
2308 | * visible after the set_pmd_at() write. | |
2309 | */ | |
2310 | smp_wmb(); | |
2311 | ||
c4088ebd | 2312 | spin_lock(pmd_ptl); |
ba76149f | 2313 | BUG_ON(!pmd_none(*pmd)); |
d281ee61 | 2314 | page_add_new_anon_rmap(new_page, vma, address, true); |
f627c2f5 | 2315 | mem_cgroup_commit_charge(new_page, memcg, false, true); |
00501b53 | 2316 | lru_cache_add_active_or_unevictable(new_page, vma); |
fce144b4 | 2317 | pgtable_trans_huge_deposit(mm, pmd, pgtable); |
ba76149f | 2318 | set_pmd_at(mm, address, pmd, _pmd); |
b113da65 | 2319 | update_mmu_cache_pmd(vma, address, pmd); |
c4088ebd | 2320 | spin_unlock(pmd_ptl); |
ba76149f AA |
2321 | |
2322 | *hpage = NULL; | |
420256ef | 2323 | |
ba76149f | 2324 | khugepaged_pages_collapsed++; |
7d2eba05 | 2325 | result = SCAN_SUCCEED; |
ce83d217 | 2326 | out_up_write: |
ba76149f | 2327 | up_write(&mm->mmap_sem); |
7d2eba05 | 2328 | trace_mm_collapse_huge_page(mm, isolated, result); |
0bbbc0b3 AA |
2329 | return; |
2330 | ||
7d2eba05 EA |
2331 | out_nolock: |
2332 | trace_mm_collapse_huge_page(mm, isolated, result); | |
2333 | return; | |
ce83d217 | 2334 | out: |
f627c2f5 | 2335 | mem_cgroup_cancel_charge(new_page, memcg, true); |
ce83d217 | 2336 | goto out_up_write; |
ba76149f AA |
2337 | } |
2338 | ||
2339 | static int khugepaged_scan_pmd(struct mm_struct *mm, | |
2340 | struct vm_area_struct *vma, | |
2341 | unsigned long address, | |
2342 | struct page **hpage) | |
2343 | { | |
ba76149f AA |
2344 | pmd_t *pmd; |
2345 | pte_t *pte, *_pte; | |
7d2eba05 EA |
2346 | int ret = 0, none_or_zero = 0, result = 0; |
2347 | struct page *page = NULL; | |
ba76149f AA |
2348 | unsigned long _address; |
2349 | spinlock_t *ptl; | |
00ef2d2f | 2350 | int node = NUMA_NO_NODE; |
10359213 | 2351 | bool writable = false, referenced = false; |
ba76149f AA |
2352 | |
2353 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
2354 | ||
6219049a | 2355 | pmd = mm_find_pmd(mm, address); |
7d2eba05 EA |
2356 | if (!pmd) { |
2357 | result = SCAN_PMD_NULL; | |
ba76149f | 2358 | goto out; |
7d2eba05 | 2359 | } |
ba76149f | 2360 | |
9f1b868a | 2361 | memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load)); |
ba76149f AA |
2362 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); |
2363 | for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR; | |
2364 | _pte++, _address += PAGE_SIZE) { | |
2365 | pte_t pteval = *_pte; | |
ca0984ca | 2366 | if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { |
c1294d05 | 2367 | if (!userfaultfd_armed(vma) && |
7d2eba05 | 2368 | ++none_or_zero <= khugepaged_max_ptes_none) { |
ba76149f | 2369 | continue; |
7d2eba05 EA |
2370 | } else { |
2371 | result = SCAN_EXCEED_NONE_PTE; | |
ba76149f | 2372 | goto out_unmap; |
7d2eba05 | 2373 | } |
ba76149f | 2374 | } |
7d2eba05 EA |
2375 | if (!pte_present(pteval)) { |
2376 | result = SCAN_PTE_NON_PRESENT; | |
ba76149f | 2377 | goto out_unmap; |
7d2eba05 | 2378 | } |
10359213 EA |
2379 | if (pte_write(pteval)) |
2380 | writable = true; | |
2381 | ||
ba76149f | 2382 | page = vm_normal_page(vma, _address, pteval); |
7d2eba05 EA |
2383 | if (unlikely(!page)) { |
2384 | result = SCAN_PAGE_NULL; | |
ba76149f | 2385 | goto out_unmap; |
7d2eba05 | 2386 | } |
b1caa957 KS |
2387 | |
2388 | /* TODO: teach khugepaged to collapse THP mapped with pte */ | |
2389 | if (PageCompound(page)) { | |
2390 | result = SCAN_PAGE_COMPOUND; | |
2391 | goto out_unmap; | |
2392 | } | |
2393 | ||
5c4b4be3 | 2394 | /* |
9f1b868a BL |
2395 | * Record which node the original page is from and save this |
2396 | * information to khugepaged_node_load[]. | |
2397 | * Khupaged will allocate hugepage from the node has the max | |
2398 | * hit record. | |
5c4b4be3 | 2399 | */ |
9f1b868a | 2400 | node = page_to_nid(page); |
7d2eba05 EA |
2401 | if (khugepaged_scan_abort(node)) { |
2402 | result = SCAN_SCAN_ABORT; | |
14a4e214 | 2403 | goto out_unmap; |
7d2eba05 | 2404 | } |
9f1b868a | 2405 | khugepaged_node_load[node]++; |
7d2eba05 EA |
2406 | if (!PageLRU(page)) { |
2407 | result = SCAN_SCAN_ABORT; | |
2408 | goto out_unmap; | |
2409 | } | |
2410 | if (PageLocked(page)) { | |
2411 | result = SCAN_PAGE_LOCK; | |
ba76149f | 2412 | goto out_unmap; |
7d2eba05 EA |
2413 | } |
2414 | if (!PageAnon(page)) { | |
2415 | result = SCAN_PAGE_ANON; | |
2416 | goto out_unmap; | |
2417 | } | |
2418 | ||
10359213 EA |
2419 | /* |
2420 | * cannot use mapcount: can't collapse if there's a gup pin. | |
2421 | * The page must only be referenced by the scanned process | |
2422 | * and page swap cache. | |
2423 | */ | |
7d2eba05 EA |
2424 | if (page_count(page) != 1 + !!PageSwapCache(page)) { |
2425 | result = SCAN_PAGE_COUNT; | |
ba76149f | 2426 | goto out_unmap; |
7d2eba05 | 2427 | } |
33c3fc71 VD |
2428 | if (pte_young(pteval) || |
2429 | page_is_young(page) || PageReferenced(page) || | |
8ee53820 | 2430 | mmu_notifier_test_young(vma->vm_mm, address)) |
10359213 | 2431 | referenced = true; |
ba76149f | 2432 | } |
7d2eba05 EA |
2433 | if (writable) { |
2434 | if (referenced) { | |
2435 | result = SCAN_SUCCEED; | |
2436 | ret = 1; | |
2437 | } else { | |
2438 | result = SCAN_NO_REFERENCED_PAGE; | |
2439 | } | |
2440 | } else { | |
2441 | result = SCAN_PAGE_RO; | |
2442 | } | |
ba76149f AA |
2443 | out_unmap: |
2444 | pte_unmap_unlock(pte, ptl); | |
9f1b868a BL |
2445 | if (ret) { |
2446 | node = khugepaged_find_target_node(); | |
ce83d217 | 2447 | /* collapse_huge_page will return with the mmap_sem released */ |
5c4b4be3 | 2448 | collapse_huge_page(mm, address, hpage, vma, node); |
9f1b868a | 2449 | } |
ba76149f | 2450 | out: |
7d2eba05 EA |
2451 | trace_mm_khugepaged_scan_pmd(mm, page_to_pfn(page), writable, referenced, |
2452 | none_or_zero, result); | |
ba76149f AA |
2453 | return ret; |
2454 | } | |
2455 | ||
2456 | static void collect_mm_slot(struct mm_slot *mm_slot) | |
2457 | { | |
2458 | struct mm_struct *mm = mm_slot->mm; | |
2459 | ||
b9980cdc | 2460 | VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); |
ba76149f AA |
2461 | |
2462 | if (khugepaged_test_exit(mm)) { | |
2463 | /* free mm_slot */ | |
43b5fbbd | 2464 | hash_del(&mm_slot->hash); |
ba76149f AA |
2465 | list_del(&mm_slot->mm_node); |
2466 | ||
2467 | /* | |
2468 | * Not strictly needed because the mm exited already. | |
2469 | * | |
2470 | * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); | |
2471 | */ | |
2472 | ||
2473 | /* khugepaged_mm_lock actually not necessary for the below */ | |
2474 | free_mm_slot(mm_slot); | |
2475 | mmdrop(mm); | |
2476 | } | |
2477 | } | |
2478 | ||
2479 | static unsigned int khugepaged_scan_mm_slot(unsigned int pages, | |
2480 | struct page **hpage) | |
2f1da642 HS |
2481 | __releases(&khugepaged_mm_lock) |
2482 | __acquires(&khugepaged_mm_lock) | |
ba76149f AA |
2483 | { |
2484 | struct mm_slot *mm_slot; | |
2485 | struct mm_struct *mm; | |
2486 | struct vm_area_struct *vma; | |
2487 | int progress = 0; | |
2488 | ||
2489 | VM_BUG_ON(!pages); | |
b9980cdc | 2490 | VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); |
ba76149f AA |
2491 | |
2492 | if (khugepaged_scan.mm_slot) | |
2493 | mm_slot = khugepaged_scan.mm_slot; | |
2494 | else { | |
2495 | mm_slot = list_entry(khugepaged_scan.mm_head.next, | |
2496 | struct mm_slot, mm_node); | |
2497 | khugepaged_scan.address = 0; | |
2498 | khugepaged_scan.mm_slot = mm_slot; | |
2499 | } | |
2500 | spin_unlock(&khugepaged_mm_lock); | |
2501 | ||
2502 | mm = mm_slot->mm; | |
2503 | down_read(&mm->mmap_sem); | |
2504 | if (unlikely(khugepaged_test_exit(mm))) | |
2505 | vma = NULL; | |
2506 | else | |
2507 | vma = find_vma(mm, khugepaged_scan.address); | |
2508 | ||
2509 | progress++; | |
2510 | for (; vma; vma = vma->vm_next) { | |
2511 | unsigned long hstart, hend; | |
2512 | ||
2513 | cond_resched(); | |
2514 | if (unlikely(khugepaged_test_exit(mm))) { | |
2515 | progress++; | |
2516 | break; | |
2517 | } | |
fa475e51 BL |
2518 | if (!hugepage_vma_check(vma)) { |
2519 | skip: | |
ba76149f AA |
2520 | progress++; |
2521 | continue; | |
2522 | } | |
ba76149f AA |
2523 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; |
2524 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
a7d6e4ec AA |
2525 | if (hstart >= hend) |
2526 | goto skip; | |
2527 | if (khugepaged_scan.address > hend) | |
2528 | goto skip; | |
ba76149f AA |
2529 | if (khugepaged_scan.address < hstart) |
2530 | khugepaged_scan.address = hstart; | |
a7d6e4ec | 2531 | VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); |
ba76149f AA |
2532 | |
2533 | while (khugepaged_scan.address < hend) { | |
2534 | int ret; | |
2535 | cond_resched(); | |
2536 | if (unlikely(khugepaged_test_exit(mm))) | |
2537 | goto breakouterloop; | |
2538 | ||
2539 | VM_BUG_ON(khugepaged_scan.address < hstart || | |
2540 | khugepaged_scan.address + HPAGE_PMD_SIZE > | |
2541 | hend); | |
2542 | ret = khugepaged_scan_pmd(mm, vma, | |
2543 | khugepaged_scan.address, | |
2544 | hpage); | |
2545 | /* move to next address */ | |
2546 | khugepaged_scan.address += HPAGE_PMD_SIZE; | |
2547 | progress += HPAGE_PMD_NR; | |
2548 | if (ret) | |
2549 | /* we released mmap_sem so break loop */ | |
2550 | goto breakouterloop_mmap_sem; | |
2551 | if (progress >= pages) | |
2552 | goto breakouterloop; | |
2553 | } | |
2554 | } | |
2555 | breakouterloop: | |
2556 | up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */ | |
2557 | breakouterloop_mmap_sem: | |
2558 | ||
2559 | spin_lock(&khugepaged_mm_lock); | |
a7d6e4ec | 2560 | VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); |
ba76149f AA |
2561 | /* |
2562 | * Release the current mm_slot if this mm is about to die, or | |
2563 | * if we scanned all vmas of this mm. | |
2564 | */ | |
2565 | if (khugepaged_test_exit(mm) || !vma) { | |
2566 | /* | |
2567 | * Make sure that if mm_users is reaching zero while | |
2568 | * khugepaged runs here, khugepaged_exit will find | |
2569 | * mm_slot not pointing to the exiting mm. | |
2570 | */ | |
2571 | if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) { | |
2572 | khugepaged_scan.mm_slot = list_entry( | |
2573 | mm_slot->mm_node.next, | |
2574 | struct mm_slot, mm_node); | |
2575 | khugepaged_scan.address = 0; | |
2576 | } else { | |
2577 | khugepaged_scan.mm_slot = NULL; | |
2578 | khugepaged_full_scans++; | |
2579 | } | |
2580 | ||
2581 | collect_mm_slot(mm_slot); | |
2582 | } | |
2583 | ||
2584 | return progress; | |
2585 | } | |
2586 | ||
2587 | static int khugepaged_has_work(void) | |
2588 | { | |
2589 | return !list_empty(&khugepaged_scan.mm_head) && | |
2590 | khugepaged_enabled(); | |
2591 | } | |
2592 | ||
2593 | static int khugepaged_wait_event(void) | |
2594 | { | |
2595 | return !list_empty(&khugepaged_scan.mm_head) || | |
2017c0bf | 2596 | kthread_should_stop(); |
ba76149f AA |
2597 | } |
2598 | ||
d516904b | 2599 | static void khugepaged_do_scan(void) |
ba76149f | 2600 | { |
d516904b | 2601 | struct page *hpage = NULL; |
ba76149f AA |
2602 | unsigned int progress = 0, pass_through_head = 0; |
2603 | unsigned int pages = khugepaged_pages_to_scan; | |
d516904b | 2604 | bool wait = true; |
ba76149f AA |
2605 | |
2606 | barrier(); /* write khugepaged_pages_to_scan to local stack */ | |
2607 | ||
2608 | while (progress < pages) { | |
26234f36 | 2609 | if (!khugepaged_prealloc_page(&hpage, &wait)) |
d516904b | 2610 | break; |
26234f36 | 2611 | |
420256ef | 2612 | cond_resched(); |
ba76149f | 2613 | |
cd092411 | 2614 | if (unlikely(kthread_should_stop() || try_to_freeze())) |
878aee7d AA |
2615 | break; |
2616 | ||
ba76149f AA |
2617 | spin_lock(&khugepaged_mm_lock); |
2618 | if (!khugepaged_scan.mm_slot) | |
2619 | pass_through_head++; | |
2620 | if (khugepaged_has_work() && | |
2621 | pass_through_head < 2) | |
2622 | progress += khugepaged_scan_mm_slot(pages - progress, | |
d516904b | 2623 | &hpage); |
ba76149f AA |
2624 | else |
2625 | progress = pages; | |
2626 | spin_unlock(&khugepaged_mm_lock); | |
2627 | } | |
ba76149f | 2628 | |
d516904b XG |
2629 | if (!IS_ERR_OR_NULL(hpage)) |
2630 | put_page(hpage); | |
0bbbc0b3 AA |
2631 | } |
2632 | ||
2017c0bf XG |
2633 | static void khugepaged_wait_work(void) |
2634 | { | |
2017c0bf XG |
2635 | if (khugepaged_has_work()) { |
2636 | if (!khugepaged_scan_sleep_millisecs) | |
2637 | return; | |
2638 | ||
2639 | wait_event_freezable_timeout(khugepaged_wait, | |
2640 | kthread_should_stop(), | |
2641 | msecs_to_jiffies(khugepaged_scan_sleep_millisecs)); | |
2642 | return; | |
2643 | } | |
2644 | ||
2645 | if (khugepaged_enabled()) | |
2646 | wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); | |
2647 | } | |
2648 | ||
ba76149f AA |
2649 | static int khugepaged(void *none) |
2650 | { | |
2651 | struct mm_slot *mm_slot; | |
2652 | ||
878aee7d | 2653 | set_freezable(); |
8698a745 | 2654 | set_user_nice(current, MAX_NICE); |
ba76149f | 2655 | |
b7231789 XG |
2656 | while (!kthread_should_stop()) { |
2657 | khugepaged_do_scan(); | |
2658 | khugepaged_wait_work(); | |
2659 | } | |
ba76149f AA |
2660 | |
2661 | spin_lock(&khugepaged_mm_lock); | |
2662 | mm_slot = khugepaged_scan.mm_slot; | |
2663 | khugepaged_scan.mm_slot = NULL; | |
2664 | if (mm_slot) | |
2665 | collect_mm_slot(mm_slot); | |
2666 | spin_unlock(&khugepaged_mm_lock); | |
ba76149f AA |
2667 | return 0; |
2668 | } | |
2669 | ||
eef1b3ba KS |
2670 | static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, |
2671 | unsigned long haddr, pmd_t *pmd) | |
2672 | { | |
2673 | struct mm_struct *mm = vma->vm_mm; | |
2674 | pgtable_t pgtable; | |
2675 | pmd_t _pmd; | |
2676 | int i; | |
2677 | ||
2678 | /* leave pmd empty until pte is filled */ | |
2679 | pmdp_huge_clear_flush_notify(vma, haddr, pmd); | |
2680 | ||
2681 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); | |
2682 | pmd_populate(mm, &_pmd, pgtable); | |
2683 | ||
2684 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
2685 | pte_t *pte, entry; | |
2686 | entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); | |
2687 | entry = pte_mkspecial(entry); | |
2688 | pte = pte_offset_map(&_pmd, haddr); | |
2689 | VM_BUG_ON(!pte_none(*pte)); | |
2690 | set_pte_at(mm, haddr, pte, entry); | |
2691 | pte_unmap(pte); | |
2692 | } | |
2693 | smp_wmb(); /* make pte visible before pmd */ | |
2694 | pmd_populate(mm, pmd, pgtable); | |
2695 | put_huge_zero_page(); | |
2696 | } | |
2697 | ||
2698 | static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, | |
ba988280 | 2699 | unsigned long haddr, bool freeze) |
eef1b3ba KS |
2700 | { |
2701 | struct mm_struct *mm = vma->vm_mm; | |
2702 | struct page *page; | |
2703 | pgtable_t pgtable; | |
2704 | pmd_t _pmd; | |
2705 | bool young, write; | |
2706 | int i; | |
2707 | ||
2708 | VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); | |
2709 | VM_BUG_ON_VMA(vma->vm_start > haddr, vma); | |
2710 | VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); | |
2711 | VM_BUG_ON(!pmd_trans_huge(*pmd)); | |
2712 | ||
2713 | count_vm_event(THP_SPLIT_PMD); | |
2714 | ||
2715 | if (vma_is_dax(vma)) { | |
2716 | pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); | |
2717 | if (is_huge_zero_pmd(_pmd)) | |
2718 | put_huge_zero_page(); | |
2719 | return; | |
2720 | } else if (is_huge_zero_pmd(*pmd)) { | |
2721 | return __split_huge_zero_page_pmd(vma, haddr, pmd); | |
2722 | } | |
2723 | ||
2724 | page = pmd_page(*pmd); | |
2725 | VM_BUG_ON_PAGE(!page_count(page), page); | |
2726 | atomic_add(HPAGE_PMD_NR - 1, &page->_count); | |
2727 | write = pmd_write(*pmd); | |
2728 | young = pmd_young(*pmd); | |
2729 | ||
eef1b3ba KS |
2730 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); |
2731 | pmd_populate(mm, &_pmd, pgtable); | |
2732 | ||
2733 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
2734 | pte_t entry, *pte; | |
2735 | /* | |
2736 | * Note that NUMA hinting access restrictions are not | |
2737 | * transferred to avoid any possibility of altering | |
2738 | * permissions across VMAs. | |
2739 | */ | |
ba988280 KS |
2740 | if (freeze) { |
2741 | swp_entry_t swp_entry; | |
2742 | swp_entry = make_migration_entry(page + i, write); | |
2743 | entry = swp_entry_to_pte(swp_entry); | |
2744 | } else { | |
2745 | entry = mk_pte(page + i, vma->vm_page_prot); | |
2746 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); | |
2747 | if (!write) | |
2748 | entry = pte_wrprotect(entry); | |
2749 | if (!young) | |
2750 | entry = pte_mkold(entry); | |
2751 | } | |
eef1b3ba KS |
2752 | pte = pte_offset_map(&_pmd, haddr); |
2753 | BUG_ON(!pte_none(*pte)); | |
2754 | set_pte_at(mm, haddr, pte, entry); | |
2755 | atomic_inc(&page[i]._mapcount); | |
2756 | pte_unmap(pte); | |
2757 | } | |
2758 | ||
2759 | /* | |
2760 | * Set PG_double_map before dropping compound_mapcount to avoid | |
2761 | * false-negative page_mapped(). | |
2762 | */ | |
2763 | if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) { | |
2764 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
2765 | atomic_inc(&page[i]._mapcount); | |
2766 | } | |
2767 | ||
2768 | if (atomic_add_negative(-1, compound_mapcount_ptr(page))) { | |
2769 | /* Last compound_mapcount is gone. */ | |
2770 | __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES); | |
2771 | if (TestClearPageDoubleMap(page)) { | |
2772 | /* No need in mapcount reference anymore */ | |
2773 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
2774 | atomic_dec(&page[i]._mapcount); | |
2775 | } | |
2776 | } | |
2777 | ||
2778 | smp_wmb(); /* make pte visible before pmd */ | |
e9b61f19 KS |
2779 | /* |
2780 | * Up to this point the pmd is present and huge and userland has the | |
2781 | * whole access to the hugepage during the split (which happens in | |
2782 | * place). If we overwrite the pmd with the not-huge version pointing | |
2783 | * to the pte here (which of course we could if all CPUs were bug | |
2784 | * free), userland could trigger a small page size TLB miss on the | |
2785 | * small sized TLB while the hugepage TLB entry is still established in | |
2786 | * the huge TLB. Some CPU doesn't like that. | |
2787 | * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum | |
2788 | * 383 on page 93. Intel should be safe but is also warns that it's | |
2789 | * only safe if the permission and cache attributes of the two entries | |
2790 | * loaded in the two TLB is identical (which should be the case here). | |
2791 | * But it is generally safer to never allow small and huge TLB entries | |
2792 | * for the same virtual address to be loaded simultaneously. So instead | |
2793 | * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the | |
2794 | * current pmd notpresent (atomically because here the pmd_trans_huge | |
2795 | * and pmd_trans_splitting must remain set at all times on the pmd | |
2796 | * until the split is complete for this pmd), then we flush the SMP TLB | |
2797 | * and finally we write the non-huge version of the pmd entry with | |
2798 | * pmd_populate. | |
2799 | */ | |
2800 | pmdp_invalidate(vma, haddr, pmd); | |
eef1b3ba | 2801 | pmd_populate(mm, pmd, pgtable); |
e9b61f19 KS |
2802 | |
2803 | if (freeze) { | |
2804 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
2805 | page_remove_rmap(page + i, false); | |
2806 | put_page(page + i); | |
2807 | } | |
2808 | } | |
eef1b3ba KS |
2809 | } |
2810 | ||
2811 | void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, | |
2812 | unsigned long address) | |
2813 | { | |
2814 | spinlock_t *ptl; | |
2815 | struct mm_struct *mm = vma->vm_mm; | |
2816 | unsigned long haddr = address & HPAGE_PMD_MASK; | |
2817 | ||
2818 | mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE); | |
2819 | ptl = pmd_lock(mm, pmd); | |
2820 | if (likely(pmd_trans_huge(*pmd))) | |
ba988280 | 2821 | __split_huge_pmd_locked(vma, pmd, haddr, false); |
eef1b3ba KS |
2822 | spin_unlock(ptl); |
2823 | mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE); | |
2824 | } | |
2825 | ||
78ddc534 | 2826 | static void split_huge_pmd_address(struct vm_area_struct *vma, |
94fcc585 AA |
2827 | unsigned long address) |
2828 | { | |
f72e7dcd HD |
2829 | pgd_t *pgd; |
2830 | pud_t *pud; | |
94fcc585 AA |
2831 | pmd_t *pmd; |
2832 | ||
2833 | VM_BUG_ON(!(address & ~HPAGE_PMD_MASK)); | |
2834 | ||
78ddc534 | 2835 | pgd = pgd_offset(vma->vm_mm, address); |
f72e7dcd HD |
2836 | if (!pgd_present(*pgd)) |
2837 | return; | |
2838 | ||
2839 | pud = pud_offset(pgd, address); | |
2840 | if (!pud_present(*pud)) | |
2841 | return; | |
2842 | ||
2843 | pmd = pmd_offset(pud, address); | |
78ddc534 | 2844 | if (!pmd_present(*pmd) || !pmd_trans_huge(*pmd)) |
94fcc585 AA |
2845 | return; |
2846 | /* | |
2847 | * Caller holds the mmap_sem write mode, so a huge pmd cannot | |
2848 | * materialize from under us. | |
2849 | */ | |
ad0bed24 | 2850 | split_huge_pmd(vma, pmd, address); |
94fcc585 AA |
2851 | } |
2852 | ||
e1b9996b | 2853 | void vma_adjust_trans_huge(struct vm_area_struct *vma, |
94fcc585 AA |
2854 | unsigned long start, |
2855 | unsigned long end, | |
2856 | long adjust_next) | |
2857 | { | |
2858 | /* | |
2859 | * If the new start address isn't hpage aligned and it could | |
2860 | * previously contain an hugepage: check if we need to split | |
2861 | * an huge pmd. | |
2862 | */ | |
2863 | if (start & ~HPAGE_PMD_MASK && | |
2864 | (start & HPAGE_PMD_MASK) >= vma->vm_start && | |
2865 | (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
78ddc534 | 2866 | split_huge_pmd_address(vma, start); |
94fcc585 AA |
2867 | |
2868 | /* | |
2869 | * If the new end address isn't hpage aligned and it could | |
2870 | * previously contain an hugepage: check if we need to split | |
2871 | * an huge pmd. | |
2872 | */ | |
2873 | if (end & ~HPAGE_PMD_MASK && | |
2874 | (end & HPAGE_PMD_MASK) >= vma->vm_start && | |
2875 | (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
78ddc534 | 2876 | split_huge_pmd_address(vma, end); |
94fcc585 AA |
2877 | |
2878 | /* | |
2879 | * If we're also updating the vma->vm_next->vm_start, if the new | |
2880 | * vm_next->vm_start isn't page aligned and it could previously | |
2881 | * contain an hugepage: check if we need to split an huge pmd. | |
2882 | */ | |
2883 | if (adjust_next > 0) { | |
2884 | struct vm_area_struct *next = vma->vm_next; | |
2885 | unsigned long nstart = next->vm_start; | |
2886 | nstart += adjust_next << PAGE_SHIFT; | |
2887 | if (nstart & ~HPAGE_PMD_MASK && | |
2888 | (nstart & HPAGE_PMD_MASK) >= next->vm_start && | |
2889 | (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) | |
78ddc534 | 2890 | split_huge_pmd_address(next, nstart); |
94fcc585 AA |
2891 | } |
2892 | } | |
e9b61f19 KS |
2893 | |
2894 | static void freeze_page_vma(struct vm_area_struct *vma, struct page *page, | |
2895 | unsigned long address) | |
2896 | { | |
2897 | spinlock_t *ptl; | |
2898 | pgd_t *pgd; | |
2899 | pud_t *pud; | |
2900 | pmd_t *pmd; | |
2901 | pte_t *pte; | |
2902 | int i, nr = HPAGE_PMD_NR; | |
2903 | ||
2904 | /* Skip pages which doesn't belong to the VMA */ | |
2905 | if (address < vma->vm_start) { | |
2906 | int off = (vma->vm_start - address) >> PAGE_SHIFT; | |
2907 | page += off; | |
2908 | nr -= off; | |
2909 | address = vma->vm_start; | |
2910 | } | |
2911 | ||
2912 | pgd = pgd_offset(vma->vm_mm, address); | |
2913 | if (!pgd_present(*pgd)) | |
2914 | return; | |
2915 | pud = pud_offset(pgd, address); | |
2916 | if (!pud_present(*pud)) | |
2917 | return; | |
2918 | pmd = pmd_offset(pud, address); | |
2919 | ptl = pmd_lock(vma->vm_mm, pmd); | |
2920 | if (!pmd_present(*pmd)) { | |
2921 | spin_unlock(ptl); | |
2922 | return; | |
2923 | } | |
2924 | if (pmd_trans_huge(*pmd)) { | |
2925 | if (page == pmd_page(*pmd)) | |
2926 | __split_huge_pmd_locked(vma, pmd, address, true); | |
2927 | spin_unlock(ptl); | |
2928 | return; | |
2929 | } | |
2930 | spin_unlock(ptl); | |
2931 | ||
2932 | pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl); | |
2933 | for (i = 0; i < nr; i++, address += PAGE_SIZE, page++) { | |
2934 | pte_t entry, swp_pte; | |
2935 | swp_entry_t swp_entry; | |
2936 | ||
2937 | if (!pte_present(pte[i])) | |
2938 | continue; | |
2939 | if (page_to_pfn(page) != pte_pfn(pte[i])) | |
2940 | continue; | |
2941 | flush_cache_page(vma, address, page_to_pfn(page)); | |
2942 | entry = ptep_clear_flush(vma, address, pte + i); | |
2943 | swp_entry = make_migration_entry(page, pte_write(entry)); | |
2944 | swp_pte = swp_entry_to_pte(swp_entry); | |
2945 | if (pte_soft_dirty(entry)) | |
2946 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
2947 | set_pte_at(vma->vm_mm, address, pte + i, swp_pte); | |
2948 | page_remove_rmap(page, false); | |
2949 | put_page(page); | |
2950 | } | |
2951 | pte_unmap_unlock(pte, ptl); | |
2952 | } | |
2953 | ||
2954 | static void freeze_page(struct anon_vma *anon_vma, struct page *page) | |
2955 | { | |
2956 | struct anon_vma_chain *avc; | |
2957 | pgoff_t pgoff = page_to_pgoff(page); | |
2958 | ||
2959 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
2960 | ||
2961 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, | |
2962 | pgoff + HPAGE_PMD_NR - 1) { | |
2963 | unsigned long haddr; | |
2964 | ||
2965 | haddr = __vma_address(page, avc->vma) & HPAGE_PMD_MASK; | |
2966 | mmu_notifier_invalidate_range_start(avc->vma->vm_mm, | |
2967 | haddr, haddr + HPAGE_PMD_SIZE); | |
2968 | freeze_page_vma(avc->vma, page, haddr); | |
2969 | mmu_notifier_invalidate_range_end(avc->vma->vm_mm, | |
2970 | haddr, haddr + HPAGE_PMD_SIZE); | |
2971 | } | |
2972 | } | |
2973 | ||
2974 | static void unfreeze_page_vma(struct vm_area_struct *vma, struct page *page, | |
2975 | unsigned long address) | |
2976 | { | |
2977 | spinlock_t *ptl; | |
2978 | pmd_t *pmd; | |
2979 | pte_t *pte, entry; | |
2980 | swp_entry_t swp_entry; | |
2981 | int i, nr = HPAGE_PMD_NR; | |
2982 | ||
2983 | /* Skip pages which doesn't belong to the VMA */ | |
2984 | if (address < vma->vm_start) { | |
2985 | int off = (vma->vm_start - address) >> PAGE_SHIFT; | |
2986 | page += off; | |
2987 | nr -= off; | |
2988 | address = vma->vm_start; | |
2989 | } | |
2990 | ||
2991 | pmd = mm_find_pmd(vma->vm_mm, address); | |
2992 | if (!pmd) | |
2993 | return; | |
2994 | pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl); | |
2995 | for (i = 0; i < nr; i++, address += PAGE_SIZE, page++) { | |
2996 | if (!is_swap_pte(pte[i])) | |
2997 | continue; | |
2998 | ||
2999 | swp_entry = pte_to_swp_entry(pte[i]); | |
3000 | if (!is_migration_entry(swp_entry)) | |
3001 | continue; | |
3002 | if (migration_entry_to_page(swp_entry) != page) | |
3003 | continue; | |
3004 | ||
3005 | get_page(page); | |
3006 | page_add_anon_rmap(page, vma, address, false); | |
3007 | ||
3008 | entry = pte_mkold(mk_pte(page, vma->vm_page_prot)); | |
3009 | entry = pte_mkdirty(entry); | |
3010 | if (is_write_migration_entry(swp_entry)) | |
3011 | entry = maybe_mkwrite(entry, vma); | |
3012 | ||
3013 | flush_dcache_page(page); | |
3014 | set_pte_at(vma->vm_mm, address, pte + i, entry); | |
3015 | ||
3016 | /* No need to invalidate - it was non-present before */ | |
3017 | update_mmu_cache(vma, address, pte + i); | |
3018 | } | |
3019 | pte_unmap_unlock(pte, ptl); | |
3020 | } | |
3021 | ||
3022 | static void unfreeze_page(struct anon_vma *anon_vma, struct page *page) | |
3023 | { | |
3024 | struct anon_vma_chain *avc; | |
3025 | pgoff_t pgoff = page_to_pgoff(page); | |
3026 | ||
3027 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, | |
3028 | pgoff, pgoff + HPAGE_PMD_NR - 1) { | |
3029 | unsigned long address = __vma_address(page, avc->vma); | |
3030 | ||
3031 | mmu_notifier_invalidate_range_start(avc->vma->vm_mm, | |
3032 | address, address + HPAGE_PMD_SIZE); | |
3033 | unfreeze_page_vma(avc->vma, page, address); | |
3034 | mmu_notifier_invalidate_range_end(avc->vma->vm_mm, | |
3035 | address, address + HPAGE_PMD_SIZE); | |
3036 | } | |
3037 | } | |
3038 | ||
3039 | static int total_mapcount(struct page *page) | |
3040 | { | |
3041 | int i, ret; | |
3042 | ||
3043 | ret = compound_mapcount(page); | |
3044 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
3045 | ret += atomic_read(&page[i]._mapcount) + 1; | |
3046 | ||
3047 | if (PageDoubleMap(page)) | |
3048 | ret -= HPAGE_PMD_NR; | |
3049 | ||
3050 | return ret; | |
3051 | } | |
3052 | ||
3053 | static int __split_huge_page_tail(struct page *head, int tail, | |
3054 | struct lruvec *lruvec, struct list_head *list) | |
3055 | { | |
3056 | int mapcount; | |
3057 | struct page *page_tail = head + tail; | |
3058 | ||
3059 | mapcount = atomic_read(&page_tail->_mapcount) + 1; | |
3060 | VM_BUG_ON_PAGE(atomic_read(&page_tail->_count) != 0, page_tail); | |
3061 | ||
3062 | /* | |
3063 | * tail_page->_count is zero and not changing from under us. But | |
3064 | * get_page_unless_zero() may be running from under us on the | |
3065 | * tail_page. If we used atomic_set() below instead of atomic_add(), we | |
3066 | * would then run atomic_set() concurrently with | |
3067 | * get_page_unless_zero(), and atomic_set() is implemented in C not | |
3068 | * using locked ops. spin_unlock on x86 sometime uses locked ops | |
3069 | * because of PPro errata 66, 92, so unless somebody can guarantee | |
3070 | * atomic_set() here would be safe on all archs (and not only on x86), | |
3071 | * it's safer to use atomic_add(). | |
3072 | */ | |
3073 | atomic_add(mapcount + 1, &page_tail->_count); | |
3074 | ||
3075 | ||
3076 | page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; | |
3077 | page_tail->flags |= (head->flags & | |
3078 | ((1L << PG_referenced) | | |
3079 | (1L << PG_swapbacked) | | |
3080 | (1L << PG_mlocked) | | |
3081 | (1L << PG_uptodate) | | |
3082 | (1L << PG_active) | | |
3083 | (1L << PG_locked) | | |
3084 | (1L << PG_unevictable))); | |
3085 | page_tail->flags |= (1L << PG_dirty); | |
3086 | ||
3087 | /* | |
3088 | * After clearing PageTail the gup refcount can be released. | |
3089 | * Page flags also must be visible before we make the page non-compound. | |
3090 | */ | |
3091 | smp_wmb(); | |
3092 | ||
3093 | clear_compound_head(page_tail); | |
3094 | ||
3095 | if (page_is_young(head)) | |
3096 | set_page_young(page_tail); | |
3097 | if (page_is_idle(head)) | |
3098 | set_page_idle(page_tail); | |
3099 | ||
3100 | /* ->mapping in first tail page is compound_mapcount */ | |
3101 | VM_BUG_ON_PAGE(tail != 1 && page_tail->mapping != TAIL_MAPPING, | |
3102 | page_tail); | |
3103 | page_tail->mapping = head->mapping; | |
3104 | ||
3105 | page_tail->index = head->index + tail; | |
3106 | page_cpupid_xchg_last(page_tail, page_cpupid_last(head)); | |
3107 | lru_add_page_tail(head, page_tail, lruvec, list); | |
3108 | ||
3109 | return mapcount; | |
3110 | } | |
3111 | ||
3112 | static void __split_huge_page(struct page *page, struct list_head *list) | |
3113 | { | |
3114 | struct page *head = compound_head(page); | |
3115 | struct zone *zone = page_zone(head); | |
3116 | struct lruvec *lruvec; | |
3117 | int i, tail_mapcount; | |
3118 | ||
3119 | /* prevent PageLRU to go away from under us, and freeze lru stats */ | |
3120 | spin_lock_irq(&zone->lru_lock); | |
3121 | lruvec = mem_cgroup_page_lruvec(head, zone); | |
3122 | ||
3123 | /* complete memcg works before add pages to LRU */ | |
3124 | mem_cgroup_split_huge_fixup(head); | |
3125 | ||
3126 | tail_mapcount = 0; | |
3127 | for (i = HPAGE_PMD_NR - 1; i >= 1; i--) | |
3128 | tail_mapcount += __split_huge_page_tail(head, i, lruvec, list); | |
3129 | atomic_sub(tail_mapcount, &head->_count); | |
3130 | ||
3131 | ClearPageCompound(head); | |
3132 | spin_unlock_irq(&zone->lru_lock); | |
3133 | ||
3134 | unfreeze_page(page_anon_vma(head), head); | |
3135 | ||
3136 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
3137 | struct page *subpage = head + i; | |
3138 | if (subpage == page) | |
3139 | continue; | |
3140 | unlock_page(subpage); | |
3141 | ||
3142 | /* | |
3143 | * Subpages may be freed if there wasn't any mapping | |
3144 | * like if add_to_swap() is running on a lru page that | |
3145 | * had its mapping zapped. And freeing these pages | |
3146 | * requires taking the lru_lock so we do the put_page | |
3147 | * of the tail pages after the split is complete. | |
3148 | */ | |
3149 | put_page(subpage); | |
3150 | } | |
3151 | } | |
3152 | ||
3153 | /* | |
3154 | * This function splits huge page into normal pages. @page can point to any | |
3155 | * subpage of huge page to split. Split doesn't change the position of @page. | |
3156 | * | |
3157 | * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. | |
3158 | * The huge page must be locked. | |
3159 | * | |
3160 | * If @list is null, tail pages will be added to LRU list, otherwise, to @list. | |
3161 | * | |
3162 | * Both head page and tail pages will inherit mapping, flags, and so on from | |
3163 | * the hugepage. | |
3164 | * | |
3165 | * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if | |
3166 | * they are not mapped. | |
3167 | * | |
3168 | * Returns 0 if the hugepage is split successfully. | |
3169 | * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under | |
3170 | * us. | |
3171 | */ | |
3172 | int split_huge_page_to_list(struct page *page, struct list_head *list) | |
3173 | { | |
3174 | struct page *head = compound_head(page); | |
3175 | struct anon_vma *anon_vma; | |
3176 | int count, mapcount, ret; | |
3177 | ||
3178 | VM_BUG_ON_PAGE(is_huge_zero_page(page), page); | |
3179 | VM_BUG_ON_PAGE(!PageAnon(page), page); | |
3180 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
3181 | VM_BUG_ON_PAGE(!PageSwapBacked(page), page); | |
3182 | VM_BUG_ON_PAGE(!PageCompound(page), page); | |
3183 | ||
3184 | /* | |
3185 | * The caller does not necessarily hold an mmap_sem that would prevent | |
3186 | * the anon_vma disappearing so we first we take a reference to it | |
3187 | * and then lock the anon_vma for write. This is similar to | |
3188 | * page_lock_anon_vma_read except the write lock is taken to serialise | |
3189 | * against parallel split or collapse operations. | |
3190 | */ | |
3191 | anon_vma = page_get_anon_vma(head); | |
3192 | if (!anon_vma) { | |
3193 | ret = -EBUSY; | |
3194 | goto out; | |
3195 | } | |
3196 | anon_vma_lock_write(anon_vma); | |
3197 | ||
3198 | /* | |
3199 | * Racy check if we can split the page, before freeze_page() will | |
3200 | * split PMDs | |
3201 | */ | |
3202 | if (total_mapcount(head) != page_count(head) - 1) { | |
3203 | ret = -EBUSY; | |
3204 | goto out_unlock; | |
3205 | } | |
3206 | ||
3207 | freeze_page(anon_vma, head); | |
3208 | VM_BUG_ON_PAGE(compound_mapcount(head), head); | |
3209 | ||
3210 | count = page_count(head); | |
3211 | mapcount = total_mapcount(head); | |
3212 | if (mapcount == count - 1) { | |
3213 | __split_huge_page(page, list); | |
3214 | ret = 0; | |
3215 | } else if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount > count - 1) { | |
3216 | pr_alert("total_mapcount: %u, page_count(): %u\n", | |
3217 | mapcount, count); | |
3218 | if (PageTail(page)) | |
3219 | dump_page(head, NULL); | |
3220 | dump_page(page, "total_mapcount(head) > page_count(head) - 1"); | |
3221 | BUG(); | |
3222 | } else { | |
3223 | unfreeze_page(anon_vma, head); | |
3224 | ret = -EBUSY; | |
3225 | } | |
3226 | ||
3227 | out_unlock: | |
3228 | anon_vma_unlock_write(anon_vma); | |
3229 | put_anon_vma(anon_vma); | |
3230 | out: | |
3231 | count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); | |
3232 | return ret; | |
3233 | } |