]>
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> |
f25748e3 | 24 | #include <linux/pfn_t.h> |
a664b2d8 | 25 | #include <linux/mman.h> |
325adeb5 | 26 | #include <linux/pagemap.h> |
49071d43 | 27 | #include <linux/debugfs.h> |
4daae3b4 | 28 | #include <linux/migrate.h> |
43b5fbbd | 29 | #include <linux/hashtable.h> |
6b251fc9 | 30 | #include <linux/userfaultfd_k.h> |
33c3fc71 | 31 | #include <linux/page_idle.h> |
97ae1749 | 32 | |
71e3aac0 AA |
33 | #include <asm/tlb.h> |
34 | #include <asm/pgalloc.h> | |
35 | #include "internal.h" | |
36 | ||
7d2eba05 EA |
37 | enum scan_result { |
38 | SCAN_FAIL, | |
39 | SCAN_SUCCEED, | |
40 | SCAN_PMD_NULL, | |
41 | SCAN_EXCEED_NONE_PTE, | |
42 | SCAN_PTE_NON_PRESENT, | |
43 | SCAN_PAGE_RO, | |
44 | SCAN_NO_REFERENCED_PAGE, | |
45 | SCAN_PAGE_NULL, | |
46 | SCAN_SCAN_ABORT, | |
47 | SCAN_PAGE_COUNT, | |
48 | SCAN_PAGE_LRU, | |
49 | SCAN_PAGE_LOCK, | |
50 | SCAN_PAGE_ANON, | |
b1caa957 | 51 | SCAN_PAGE_COMPOUND, |
7d2eba05 EA |
52 | SCAN_ANY_PROCESS, |
53 | SCAN_VMA_NULL, | |
54 | SCAN_VMA_CHECK, | |
55 | SCAN_ADDRESS_RANGE, | |
56 | SCAN_SWAP_CACHE_PAGE, | |
57 | SCAN_DEL_PAGE_LRU, | |
58 | SCAN_ALLOC_HUGE_PAGE_FAIL, | |
59 | SCAN_CGROUP_CHARGE_FAIL | |
60 | }; | |
61 | ||
62 | #define CREATE_TRACE_POINTS | |
63 | #include <trace/events/huge_memory.h> | |
64 | ||
ba76149f | 65 | /* |
8bfa3f9a JW |
66 | * By default transparent hugepage support is disabled in order that avoid |
67 | * to risk increase the memory footprint of applications without a guaranteed | |
68 | * benefit. When transparent hugepage support is enabled, is for all mappings, | |
69 | * and khugepaged scans all mappings. | |
70 | * Defrag is invoked by khugepaged hugepage allocations and by page faults | |
71 | * for all hugepage allocations. | |
ba76149f | 72 | */ |
71e3aac0 | 73 | unsigned long transparent_hugepage_flags __read_mostly = |
13ece886 | 74 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS |
ba76149f | 75 | (1<<TRANSPARENT_HUGEPAGE_FLAG)| |
13ece886 AA |
76 | #endif |
77 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE | |
78 | (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| | |
79 | #endif | |
d39d33c3 | 80 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)| |
79da5407 KS |
81 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| |
82 | (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
ba76149f AA |
83 | |
84 | /* default scan 8*512 pte (or vmas) every 30 second */ | |
85 | static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8; | |
86 | static unsigned int khugepaged_pages_collapsed; | |
87 | static unsigned int khugepaged_full_scans; | |
88 | static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; | |
89 | /* during fragmentation poll the hugepage allocator once every minute */ | |
90 | static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; | |
91 | static struct task_struct *khugepaged_thread __read_mostly; | |
92 | static DEFINE_MUTEX(khugepaged_mutex); | |
93 | static DEFINE_SPINLOCK(khugepaged_mm_lock); | |
94 | static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); | |
95 | /* | |
96 | * default collapse hugepages if there is at least one pte mapped like | |
97 | * it would have happened if the vma was large enough during page | |
98 | * fault. | |
99 | */ | |
100 | static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1; | |
101 | ||
102 | static int khugepaged(void *none); | |
ba76149f | 103 | static int khugepaged_slab_init(void); |
65ebb64f | 104 | static void khugepaged_slab_exit(void); |
ba76149f | 105 | |
43b5fbbd SL |
106 | #define MM_SLOTS_HASH_BITS 10 |
107 | static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); | |
108 | ||
ba76149f AA |
109 | static struct kmem_cache *mm_slot_cache __read_mostly; |
110 | ||
111 | /** | |
112 | * struct mm_slot - hash lookup from mm to mm_slot | |
113 | * @hash: hash collision list | |
114 | * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head | |
115 | * @mm: the mm that this information is valid for | |
116 | */ | |
117 | struct mm_slot { | |
118 | struct hlist_node hash; | |
119 | struct list_head mm_node; | |
120 | struct mm_struct *mm; | |
121 | }; | |
122 | ||
123 | /** | |
124 | * struct khugepaged_scan - cursor for scanning | |
125 | * @mm_head: the head of the mm list to scan | |
126 | * @mm_slot: the current mm_slot we are scanning | |
127 | * @address: the next address inside that to be scanned | |
128 | * | |
129 | * There is only the one khugepaged_scan instance of this cursor structure. | |
130 | */ | |
131 | struct khugepaged_scan { | |
132 | struct list_head mm_head; | |
133 | struct mm_slot *mm_slot; | |
134 | unsigned long address; | |
2f1da642 HS |
135 | }; |
136 | static struct khugepaged_scan khugepaged_scan = { | |
ba76149f AA |
137 | .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), |
138 | }; | |
139 | ||
9a982250 KS |
140 | static DEFINE_SPINLOCK(split_queue_lock); |
141 | static LIST_HEAD(split_queue); | |
142 | static unsigned long split_queue_len; | |
143 | static struct shrinker deferred_split_shrinker; | |
f000565a | 144 | |
2c0b80d4 | 145 | static void set_recommended_min_free_kbytes(void) |
f000565a AA |
146 | { |
147 | struct zone *zone; | |
148 | int nr_zones = 0; | |
149 | unsigned long recommended_min; | |
f000565a | 150 | |
f000565a AA |
151 | for_each_populated_zone(zone) |
152 | nr_zones++; | |
153 | ||
974a786e | 154 | /* Ensure 2 pageblocks are free to assist fragmentation avoidance */ |
f000565a AA |
155 | recommended_min = pageblock_nr_pages * nr_zones * 2; |
156 | ||
157 | /* | |
158 | * Make sure that on average at least two pageblocks are almost free | |
159 | * of another type, one for a migratetype to fall back to and a | |
160 | * second to avoid subsequent fallbacks of other types There are 3 | |
161 | * MIGRATE_TYPES we care about. | |
162 | */ | |
163 | recommended_min += pageblock_nr_pages * nr_zones * | |
164 | MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; | |
165 | ||
166 | /* don't ever allow to reserve more than 5% of the lowmem */ | |
167 | recommended_min = min(recommended_min, | |
168 | (unsigned long) nr_free_buffer_pages() / 20); | |
169 | recommended_min <<= (PAGE_SHIFT-10); | |
170 | ||
42aa83cb HP |
171 | if (recommended_min > min_free_kbytes) { |
172 | if (user_min_free_kbytes >= 0) | |
173 | pr_info("raising min_free_kbytes from %d to %lu " | |
174 | "to help transparent hugepage allocations\n", | |
175 | min_free_kbytes, recommended_min); | |
176 | ||
f000565a | 177 | min_free_kbytes = recommended_min; |
42aa83cb | 178 | } |
f000565a | 179 | setup_per_zone_wmarks(); |
f000565a | 180 | } |
f000565a | 181 | |
79553da2 | 182 | static int start_stop_khugepaged(void) |
ba76149f AA |
183 | { |
184 | int err = 0; | |
185 | if (khugepaged_enabled()) { | |
ba76149f AA |
186 | if (!khugepaged_thread) |
187 | khugepaged_thread = kthread_run(khugepaged, NULL, | |
188 | "khugepaged"); | |
18e8e5c7 | 189 | if (IS_ERR(khugepaged_thread)) { |
ae3a8c1c | 190 | pr_err("khugepaged: kthread_run(khugepaged) failed\n"); |
ba76149f AA |
191 | err = PTR_ERR(khugepaged_thread); |
192 | khugepaged_thread = NULL; | |
79553da2 | 193 | goto fail; |
ba76149f | 194 | } |
911891af XG |
195 | |
196 | if (!list_empty(&khugepaged_scan.mm_head)) | |
ba76149f | 197 | wake_up_interruptible(&khugepaged_wait); |
f000565a AA |
198 | |
199 | set_recommended_min_free_kbytes(); | |
911891af | 200 | } else if (khugepaged_thread) { |
911891af XG |
201 | kthread_stop(khugepaged_thread); |
202 | khugepaged_thread = NULL; | |
203 | } | |
79553da2 | 204 | fail: |
ba76149f AA |
205 | return err; |
206 | } | |
71e3aac0 | 207 | |
97ae1749 | 208 | static atomic_t huge_zero_refcount; |
56873f43 | 209 | struct page *huge_zero_page __read_mostly; |
4a6c1297 | 210 | |
fc437044 | 211 | struct page *get_huge_zero_page(void) |
97ae1749 KS |
212 | { |
213 | struct page *zero_page; | |
214 | retry: | |
215 | if (likely(atomic_inc_not_zero(&huge_zero_refcount))) | |
4db0c3c2 | 216 | return READ_ONCE(huge_zero_page); |
97ae1749 KS |
217 | |
218 | zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, | |
4a6c1297 | 219 | HPAGE_PMD_ORDER); |
d8a8e1f0 KS |
220 | if (!zero_page) { |
221 | count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); | |
5918d10a | 222 | return NULL; |
d8a8e1f0 KS |
223 | } |
224 | count_vm_event(THP_ZERO_PAGE_ALLOC); | |
97ae1749 | 225 | preempt_disable(); |
5918d10a | 226 | if (cmpxchg(&huge_zero_page, NULL, zero_page)) { |
97ae1749 | 227 | preempt_enable(); |
5ddacbe9 | 228 | __free_pages(zero_page, compound_order(zero_page)); |
97ae1749 KS |
229 | goto retry; |
230 | } | |
231 | ||
232 | /* We take additional reference here. It will be put back by shrinker */ | |
233 | atomic_set(&huge_zero_refcount, 2); | |
234 | preempt_enable(); | |
4db0c3c2 | 235 | return READ_ONCE(huge_zero_page); |
4a6c1297 KS |
236 | } |
237 | ||
97ae1749 | 238 | static void put_huge_zero_page(void) |
4a6c1297 | 239 | { |
97ae1749 KS |
240 | /* |
241 | * Counter should never go to zero here. Only shrinker can put | |
242 | * last reference. | |
243 | */ | |
244 | BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); | |
4a6c1297 KS |
245 | } |
246 | ||
48896466 GC |
247 | static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, |
248 | struct shrink_control *sc) | |
4a6c1297 | 249 | { |
48896466 GC |
250 | /* we can free zero page only if last reference remains */ |
251 | return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; | |
252 | } | |
97ae1749 | 253 | |
48896466 GC |
254 | static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, |
255 | struct shrink_control *sc) | |
256 | { | |
97ae1749 | 257 | if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { |
5918d10a KS |
258 | struct page *zero_page = xchg(&huge_zero_page, NULL); |
259 | BUG_ON(zero_page == NULL); | |
5ddacbe9 | 260 | __free_pages(zero_page, compound_order(zero_page)); |
48896466 | 261 | return HPAGE_PMD_NR; |
97ae1749 KS |
262 | } |
263 | ||
264 | return 0; | |
4a6c1297 KS |
265 | } |
266 | ||
97ae1749 | 267 | static struct shrinker huge_zero_page_shrinker = { |
48896466 GC |
268 | .count_objects = shrink_huge_zero_page_count, |
269 | .scan_objects = shrink_huge_zero_page_scan, | |
97ae1749 KS |
270 | .seeks = DEFAULT_SEEKS, |
271 | }; | |
272 | ||
71e3aac0 | 273 | #ifdef CONFIG_SYSFS |
ba76149f | 274 | |
71e3aac0 AA |
275 | static ssize_t double_flag_show(struct kobject *kobj, |
276 | struct kobj_attribute *attr, char *buf, | |
277 | enum transparent_hugepage_flag enabled, | |
278 | enum transparent_hugepage_flag req_madv) | |
279 | { | |
280 | if (test_bit(enabled, &transparent_hugepage_flags)) { | |
281 | VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags)); | |
282 | return sprintf(buf, "[always] madvise never\n"); | |
283 | } else if (test_bit(req_madv, &transparent_hugepage_flags)) | |
284 | return sprintf(buf, "always [madvise] never\n"); | |
285 | else | |
286 | return sprintf(buf, "always madvise [never]\n"); | |
287 | } | |
288 | static ssize_t double_flag_store(struct kobject *kobj, | |
289 | struct kobj_attribute *attr, | |
290 | const char *buf, size_t count, | |
291 | enum transparent_hugepage_flag enabled, | |
292 | enum transparent_hugepage_flag req_madv) | |
293 | { | |
294 | if (!memcmp("always", buf, | |
295 | min(sizeof("always")-1, count))) { | |
296 | set_bit(enabled, &transparent_hugepage_flags); | |
297 | clear_bit(req_madv, &transparent_hugepage_flags); | |
298 | } else if (!memcmp("madvise", buf, | |
299 | min(sizeof("madvise")-1, count))) { | |
300 | clear_bit(enabled, &transparent_hugepage_flags); | |
301 | set_bit(req_madv, &transparent_hugepage_flags); | |
302 | } else if (!memcmp("never", buf, | |
303 | min(sizeof("never")-1, count))) { | |
304 | clear_bit(enabled, &transparent_hugepage_flags); | |
305 | clear_bit(req_madv, &transparent_hugepage_flags); | |
306 | } else | |
307 | return -EINVAL; | |
308 | ||
309 | return count; | |
310 | } | |
311 | ||
312 | static ssize_t enabled_show(struct kobject *kobj, | |
313 | struct kobj_attribute *attr, char *buf) | |
314 | { | |
315 | return double_flag_show(kobj, attr, buf, | |
316 | TRANSPARENT_HUGEPAGE_FLAG, | |
317 | TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); | |
318 | } | |
319 | static ssize_t enabled_store(struct kobject *kobj, | |
320 | struct kobj_attribute *attr, | |
321 | const char *buf, size_t count) | |
322 | { | |
ba76149f AA |
323 | ssize_t ret; |
324 | ||
325 | ret = double_flag_store(kobj, attr, buf, count, | |
326 | TRANSPARENT_HUGEPAGE_FLAG, | |
327 | TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); | |
328 | ||
329 | if (ret > 0) { | |
911891af XG |
330 | int err; |
331 | ||
332 | mutex_lock(&khugepaged_mutex); | |
79553da2 | 333 | err = start_stop_khugepaged(); |
911891af XG |
334 | mutex_unlock(&khugepaged_mutex); |
335 | ||
ba76149f AA |
336 | if (err) |
337 | ret = err; | |
338 | } | |
339 | ||
340 | return ret; | |
71e3aac0 AA |
341 | } |
342 | static struct kobj_attribute enabled_attr = | |
343 | __ATTR(enabled, 0644, enabled_show, enabled_store); | |
344 | ||
345 | static ssize_t single_flag_show(struct kobject *kobj, | |
346 | struct kobj_attribute *attr, char *buf, | |
347 | enum transparent_hugepage_flag flag) | |
348 | { | |
e27e6151 BH |
349 | return sprintf(buf, "%d\n", |
350 | !!test_bit(flag, &transparent_hugepage_flags)); | |
71e3aac0 | 351 | } |
e27e6151 | 352 | |
71e3aac0 AA |
353 | static ssize_t single_flag_store(struct kobject *kobj, |
354 | struct kobj_attribute *attr, | |
355 | const char *buf, size_t count, | |
356 | enum transparent_hugepage_flag flag) | |
357 | { | |
e27e6151 BH |
358 | unsigned long value; |
359 | int ret; | |
360 | ||
361 | ret = kstrtoul(buf, 10, &value); | |
362 | if (ret < 0) | |
363 | return ret; | |
364 | if (value > 1) | |
365 | return -EINVAL; | |
366 | ||
367 | if (value) | |
71e3aac0 | 368 | set_bit(flag, &transparent_hugepage_flags); |
e27e6151 | 369 | else |
71e3aac0 | 370 | clear_bit(flag, &transparent_hugepage_flags); |
71e3aac0 AA |
371 | |
372 | return count; | |
373 | } | |
374 | ||
375 | /* | |
376 | * Currently defrag only disables __GFP_NOWAIT for allocation. A blind | |
377 | * __GFP_REPEAT is too aggressive, it's never worth swapping tons of | |
378 | * memory just to allocate one more hugepage. | |
379 | */ | |
380 | static ssize_t defrag_show(struct kobject *kobj, | |
381 | struct kobj_attribute *attr, char *buf) | |
382 | { | |
383 | return double_flag_show(kobj, attr, buf, | |
384 | TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, | |
385 | TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); | |
386 | } | |
387 | static ssize_t defrag_store(struct kobject *kobj, | |
388 | struct kobj_attribute *attr, | |
389 | const char *buf, size_t count) | |
390 | { | |
391 | return double_flag_store(kobj, attr, buf, count, | |
392 | TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, | |
393 | TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); | |
394 | } | |
395 | static struct kobj_attribute defrag_attr = | |
396 | __ATTR(defrag, 0644, defrag_show, defrag_store); | |
397 | ||
79da5407 KS |
398 | static ssize_t use_zero_page_show(struct kobject *kobj, |
399 | struct kobj_attribute *attr, char *buf) | |
400 | { | |
401 | return single_flag_show(kobj, attr, buf, | |
402 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
403 | } | |
404 | static ssize_t use_zero_page_store(struct kobject *kobj, | |
405 | struct kobj_attribute *attr, const char *buf, size_t count) | |
406 | { | |
407 | return single_flag_store(kobj, attr, buf, count, | |
408 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
409 | } | |
410 | static struct kobj_attribute use_zero_page_attr = | |
411 | __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); | |
71e3aac0 AA |
412 | #ifdef CONFIG_DEBUG_VM |
413 | static ssize_t debug_cow_show(struct kobject *kobj, | |
414 | struct kobj_attribute *attr, char *buf) | |
415 | { | |
416 | return single_flag_show(kobj, attr, buf, | |
417 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); | |
418 | } | |
419 | static ssize_t debug_cow_store(struct kobject *kobj, | |
420 | struct kobj_attribute *attr, | |
421 | const char *buf, size_t count) | |
422 | { | |
423 | return single_flag_store(kobj, attr, buf, count, | |
424 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); | |
425 | } | |
426 | static struct kobj_attribute debug_cow_attr = | |
427 | __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); | |
428 | #endif /* CONFIG_DEBUG_VM */ | |
429 | ||
430 | static struct attribute *hugepage_attr[] = { | |
431 | &enabled_attr.attr, | |
432 | &defrag_attr.attr, | |
79da5407 | 433 | &use_zero_page_attr.attr, |
71e3aac0 AA |
434 | #ifdef CONFIG_DEBUG_VM |
435 | &debug_cow_attr.attr, | |
436 | #endif | |
437 | NULL, | |
438 | }; | |
439 | ||
440 | static struct attribute_group hugepage_attr_group = { | |
441 | .attrs = hugepage_attr, | |
ba76149f AA |
442 | }; |
443 | ||
444 | static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, | |
445 | struct kobj_attribute *attr, | |
446 | char *buf) | |
447 | { | |
448 | return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs); | |
449 | } | |
450 | ||
451 | static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, | |
452 | struct kobj_attribute *attr, | |
453 | const char *buf, size_t count) | |
454 | { | |
455 | unsigned long msecs; | |
456 | int err; | |
457 | ||
3dbb95f7 | 458 | err = kstrtoul(buf, 10, &msecs); |
ba76149f AA |
459 | if (err || msecs > UINT_MAX) |
460 | return -EINVAL; | |
461 | ||
462 | khugepaged_scan_sleep_millisecs = msecs; | |
463 | wake_up_interruptible(&khugepaged_wait); | |
464 | ||
465 | return count; | |
466 | } | |
467 | static struct kobj_attribute scan_sleep_millisecs_attr = | |
468 | __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show, | |
469 | scan_sleep_millisecs_store); | |
470 | ||
471 | static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, | |
472 | struct kobj_attribute *attr, | |
473 | char *buf) | |
474 | { | |
475 | return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs); | |
476 | } | |
477 | ||
478 | static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, | |
479 | struct kobj_attribute *attr, | |
480 | const char *buf, size_t count) | |
481 | { | |
482 | unsigned long msecs; | |
483 | int err; | |
484 | ||
3dbb95f7 | 485 | err = kstrtoul(buf, 10, &msecs); |
ba76149f AA |
486 | if (err || msecs > UINT_MAX) |
487 | return -EINVAL; | |
488 | ||
489 | khugepaged_alloc_sleep_millisecs = msecs; | |
490 | wake_up_interruptible(&khugepaged_wait); | |
491 | ||
492 | return count; | |
493 | } | |
494 | static struct kobj_attribute alloc_sleep_millisecs_attr = | |
495 | __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show, | |
496 | alloc_sleep_millisecs_store); | |
497 | ||
498 | static ssize_t pages_to_scan_show(struct kobject *kobj, | |
499 | struct kobj_attribute *attr, | |
500 | char *buf) | |
501 | { | |
502 | return sprintf(buf, "%u\n", khugepaged_pages_to_scan); | |
503 | } | |
504 | static ssize_t pages_to_scan_store(struct kobject *kobj, | |
505 | struct kobj_attribute *attr, | |
506 | const char *buf, size_t count) | |
507 | { | |
508 | int err; | |
509 | unsigned long pages; | |
510 | ||
3dbb95f7 | 511 | err = kstrtoul(buf, 10, &pages); |
ba76149f AA |
512 | if (err || !pages || pages > UINT_MAX) |
513 | return -EINVAL; | |
514 | ||
515 | khugepaged_pages_to_scan = pages; | |
516 | ||
517 | return count; | |
518 | } | |
519 | static struct kobj_attribute pages_to_scan_attr = | |
520 | __ATTR(pages_to_scan, 0644, pages_to_scan_show, | |
521 | pages_to_scan_store); | |
522 | ||
523 | static ssize_t pages_collapsed_show(struct kobject *kobj, | |
524 | struct kobj_attribute *attr, | |
525 | char *buf) | |
526 | { | |
527 | return sprintf(buf, "%u\n", khugepaged_pages_collapsed); | |
528 | } | |
529 | static struct kobj_attribute pages_collapsed_attr = | |
530 | __ATTR_RO(pages_collapsed); | |
531 | ||
532 | static ssize_t full_scans_show(struct kobject *kobj, | |
533 | struct kobj_attribute *attr, | |
534 | char *buf) | |
535 | { | |
536 | return sprintf(buf, "%u\n", khugepaged_full_scans); | |
537 | } | |
538 | static struct kobj_attribute full_scans_attr = | |
539 | __ATTR_RO(full_scans); | |
540 | ||
541 | static ssize_t khugepaged_defrag_show(struct kobject *kobj, | |
542 | struct kobj_attribute *attr, char *buf) | |
543 | { | |
544 | return single_flag_show(kobj, attr, buf, | |
545 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); | |
546 | } | |
547 | static ssize_t khugepaged_defrag_store(struct kobject *kobj, | |
548 | struct kobj_attribute *attr, | |
549 | const char *buf, size_t count) | |
550 | { | |
551 | return single_flag_store(kobj, attr, buf, count, | |
552 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); | |
553 | } | |
554 | static struct kobj_attribute khugepaged_defrag_attr = | |
555 | __ATTR(defrag, 0644, khugepaged_defrag_show, | |
556 | khugepaged_defrag_store); | |
557 | ||
558 | /* | |
559 | * max_ptes_none controls if khugepaged should collapse hugepages over | |
560 | * any unmapped ptes in turn potentially increasing the memory | |
561 | * footprint of the vmas. When max_ptes_none is 0 khugepaged will not | |
562 | * reduce the available free memory in the system as it | |
563 | * runs. Increasing max_ptes_none will instead potentially reduce the | |
564 | * free memory in the system during the khugepaged scan. | |
565 | */ | |
566 | static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj, | |
567 | struct kobj_attribute *attr, | |
568 | char *buf) | |
569 | { | |
570 | return sprintf(buf, "%u\n", khugepaged_max_ptes_none); | |
571 | } | |
572 | static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj, | |
573 | struct kobj_attribute *attr, | |
574 | const char *buf, size_t count) | |
575 | { | |
576 | int err; | |
577 | unsigned long max_ptes_none; | |
578 | ||
3dbb95f7 | 579 | err = kstrtoul(buf, 10, &max_ptes_none); |
ba76149f AA |
580 | if (err || max_ptes_none > HPAGE_PMD_NR-1) |
581 | return -EINVAL; | |
582 | ||
583 | khugepaged_max_ptes_none = max_ptes_none; | |
584 | ||
585 | return count; | |
586 | } | |
587 | static struct kobj_attribute khugepaged_max_ptes_none_attr = | |
588 | __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show, | |
589 | khugepaged_max_ptes_none_store); | |
590 | ||
591 | static struct attribute *khugepaged_attr[] = { | |
592 | &khugepaged_defrag_attr.attr, | |
593 | &khugepaged_max_ptes_none_attr.attr, | |
594 | &pages_to_scan_attr.attr, | |
595 | &pages_collapsed_attr.attr, | |
596 | &full_scans_attr.attr, | |
597 | &scan_sleep_millisecs_attr.attr, | |
598 | &alloc_sleep_millisecs_attr.attr, | |
599 | NULL, | |
600 | }; | |
601 | ||
602 | static struct attribute_group khugepaged_attr_group = { | |
603 | .attrs = khugepaged_attr, | |
604 | .name = "khugepaged", | |
71e3aac0 | 605 | }; |
71e3aac0 | 606 | |
569e5590 | 607 | static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) |
71e3aac0 | 608 | { |
71e3aac0 AA |
609 | int err; |
610 | ||
569e5590 SL |
611 | *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); |
612 | if (unlikely(!*hugepage_kobj)) { | |
ae3a8c1c | 613 | pr_err("failed to create transparent hugepage kobject\n"); |
569e5590 | 614 | return -ENOMEM; |
ba76149f AA |
615 | } |
616 | ||
569e5590 | 617 | err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); |
ba76149f | 618 | if (err) { |
ae3a8c1c | 619 | pr_err("failed to register transparent hugepage group\n"); |
569e5590 | 620 | goto delete_obj; |
ba76149f AA |
621 | } |
622 | ||
569e5590 | 623 | err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); |
ba76149f | 624 | if (err) { |
ae3a8c1c | 625 | pr_err("failed to register transparent hugepage group\n"); |
569e5590 | 626 | goto remove_hp_group; |
ba76149f | 627 | } |
569e5590 SL |
628 | |
629 | return 0; | |
630 | ||
631 | remove_hp_group: | |
632 | sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); | |
633 | delete_obj: | |
634 | kobject_put(*hugepage_kobj); | |
635 | return err; | |
636 | } | |
637 | ||
638 | static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
639 | { | |
640 | sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); | |
641 | sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); | |
642 | kobject_put(hugepage_kobj); | |
643 | } | |
644 | #else | |
645 | static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) | |
646 | { | |
647 | return 0; | |
648 | } | |
649 | ||
650 | static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
651 | { | |
652 | } | |
653 | #endif /* CONFIG_SYSFS */ | |
654 | ||
655 | static int __init hugepage_init(void) | |
656 | { | |
657 | int err; | |
658 | struct kobject *hugepage_kobj; | |
659 | ||
660 | if (!has_transparent_hugepage()) { | |
661 | transparent_hugepage_flags = 0; | |
662 | return -EINVAL; | |
663 | } | |
664 | ||
665 | err = hugepage_init_sysfs(&hugepage_kobj); | |
666 | if (err) | |
65ebb64f | 667 | goto err_sysfs; |
ba76149f AA |
668 | |
669 | err = khugepaged_slab_init(); | |
670 | if (err) | |
65ebb64f | 671 | goto err_slab; |
ba76149f | 672 | |
65ebb64f KS |
673 | err = register_shrinker(&huge_zero_page_shrinker); |
674 | if (err) | |
675 | goto err_hzp_shrinker; | |
9a982250 KS |
676 | err = register_shrinker(&deferred_split_shrinker); |
677 | if (err) | |
678 | goto err_split_shrinker; | |
97ae1749 | 679 | |
97562cd2 RR |
680 | /* |
681 | * By default disable transparent hugepages on smaller systems, | |
682 | * where the extra memory used could hurt more than TLB overhead | |
683 | * is likely to save. The admin can still enable it through /sys. | |
684 | */ | |
79553da2 | 685 | if (totalram_pages < (512 << (20 - PAGE_SHIFT))) { |
97562cd2 | 686 | transparent_hugepage_flags = 0; |
79553da2 KS |
687 | return 0; |
688 | } | |
97562cd2 | 689 | |
79553da2 | 690 | err = start_stop_khugepaged(); |
65ebb64f KS |
691 | if (err) |
692 | goto err_khugepaged; | |
ba76149f | 693 | |
569e5590 | 694 | return 0; |
65ebb64f | 695 | err_khugepaged: |
9a982250 KS |
696 | unregister_shrinker(&deferred_split_shrinker); |
697 | err_split_shrinker: | |
65ebb64f KS |
698 | unregister_shrinker(&huge_zero_page_shrinker); |
699 | err_hzp_shrinker: | |
700 | khugepaged_slab_exit(); | |
701 | err_slab: | |
569e5590 | 702 | hugepage_exit_sysfs(hugepage_kobj); |
65ebb64f | 703 | err_sysfs: |
ba76149f | 704 | return err; |
71e3aac0 | 705 | } |
a64fb3cd | 706 | subsys_initcall(hugepage_init); |
71e3aac0 AA |
707 | |
708 | static int __init setup_transparent_hugepage(char *str) | |
709 | { | |
710 | int ret = 0; | |
711 | if (!str) | |
712 | goto out; | |
713 | if (!strcmp(str, "always")) { | |
714 | set_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
715 | &transparent_hugepage_flags); | |
716 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
717 | &transparent_hugepage_flags); | |
718 | ret = 1; | |
719 | } else if (!strcmp(str, "madvise")) { | |
720 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
721 | &transparent_hugepage_flags); | |
722 | set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
723 | &transparent_hugepage_flags); | |
724 | ret = 1; | |
725 | } else if (!strcmp(str, "never")) { | |
726 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
727 | &transparent_hugepage_flags); | |
728 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
729 | &transparent_hugepage_flags); | |
730 | ret = 1; | |
731 | } | |
732 | out: | |
733 | if (!ret) | |
ae3a8c1c | 734 | pr_warn("transparent_hugepage= cannot parse, ignored\n"); |
71e3aac0 AA |
735 | return ret; |
736 | } | |
737 | __setup("transparent_hugepage=", setup_transparent_hugepage); | |
738 | ||
b32967ff | 739 | pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) |
71e3aac0 AA |
740 | { |
741 | if (likely(vma->vm_flags & VM_WRITE)) | |
742 | pmd = pmd_mkwrite(pmd); | |
743 | return pmd; | |
744 | } | |
745 | ||
3122359a | 746 | static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot) |
b3092b3b BL |
747 | { |
748 | pmd_t entry; | |
3122359a | 749 | entry = mk_pmd(page, prot); |
b3092b3b BL |
750 | entry = pmd_mkhuge(entry); |
751 | return entry; | |
752 | } | |
753 | ||
9a982250 KS |
754 | static inline struct list_head *page_deferred_list(struct page *page) |
755 | { | |
756 | /* | |
757 | * ->lru in the tail pages is occupied by compound_head. | |
758 | * Let's use ->mapping + ->index in the second tail page as list_head. | |
759 | */ | |
760 | return (struct list_head *)&page[2].mapping; | |
761 | } | |
762 | ||
763 | void prep_transhuge_page(struct page *page) | |
764 | { | |
765 | /* | |
766 | * we use page->mapping and page->indexlru in second tail page | |
767 | * as list_head: assuming THP order >= 2 | |
768 | */ | |
769 | BUILD_BUG_ON(HPAGE_PMD_ORDER < 2); | |
770 | ||
771 | INIT_LIST_HEAD(page_deferred_list(page)); | |
772 | set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR); | |
773 | } | |
774 | ||
71e3aac0 AA |
775 | static int __do_huge_pmd_anonymous_page(struct mm_struct *mm, |
776 | struct vm_area_struct *vma, | |
230c92a8 | 777 | unsigned long address, pmd_t *pmd, |
6b251fc9 AA |
778 | struct page *page, gfp_t gfp, |
779 | unsigned int flags) | |
71e3aac0 | 780 | { |
00501b53 | 781 | struct mem_cgroup *memcg; |
71e3aac0 | 782 | pgtable_t pgtable; |
c4088ebd | 783 | spinlock_t *ptl; |
230c92a8 | 784 | unsigned long haddr = address & HPAGE_PMD_MASK; |
71e3aac0 | 785 | |
309381fe | 786 | VM_BUG_ON_PAGE(!PageCompound(page), page); |
00501b53 | 787 | |
f627c2f5 | 788 | if (mem_cgroup_try_charge(page, mm, gfp, &memcg, true)) { |
6b251fc9 AA |
789 | put_page(page); |
790 | count_vm_event(THP_FAULT_FALLBACK); | |
791 | return VM_FAULT_FALLBACK; | |
792 | } | |
00501b53 | 793 | |
71e3aac0 | 794 | pgtable = pte_alloc_one(mm, haddr); |
00501b53 | 795 | if (unlikely(!pgtable)) { |
f627c2f5 | 796 | mem_cgroup_cancel_charge(page, memcg, true); |
6b251fc9 | 797 | put_page(page); |
71e3aac0 | 798 | return VM_FAULT_OOM; |
00501b53 | 799 | } |
71e3aac0 AA |
800 | |
801 | clear_huge_page(page, haddr, HPAGE_PMD_NR); | |
52f37629 MK |
802 | /* |
803 | * The memory barrier inside __SetPageUptodate makes sure that | |
804 | * clear_huge_page writes become visible before the set_pmd_at() | |
805 | * write. | |
806 | */ | |
71e3aac0 AA |
807 | __SetPageUptodate(page); |
808 | ||
c4088ebd | 809 | ptl = pmd_lock(mm, pmd); |
71e3aac0 | 810 | if (unlikely(!pmd_none(*pmd))) { |
c4088ebd | 811 | spin_unlock(ptl); |
f627c2f5 | 812 | mem_cgroup_cancel_charge(page, memcg, true); |
71e3aac0 AA |
813 | put_page(page); |
814 | pte_free(mm, pgtable); | |
815 | } else { | |
816 | pmd_t entry; | |
6b251fc9 AA |
817 | |
818 | /* Deliver the page fault to userland */ | |
819 | if (userfaultfd_missing(vma)) { | |
820 | int ret; | |
821 | ||
822 | spin_unlock(ptl); | |
f627c2f5 | 823 | mem_cgroup_cancel_charge(page, memcg, true); |
6b251fc9 AA |
824 | put_page(page); |
825 | pte_free(mm, pgtable); | |
230c92a8 | 826 | ret = handle_userfault(vma, address, flags, |
6b251fc9 AA |
827 | VM_UFFD_MISSING); |
828 | VM_BUG_ON(ret & VM_FAULT_FALLBACK); | |
829 | return ret; | |
830 | } | |
831 | ||
3122359a KS |
832 | entry = mk_huge_pmd(page, vma->vm_page_prot); |
833 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
d281ee61 | 834 | page_add_new_anon_rmap(page, vma, haddr, true); |
f627c2f5 | 835 | mem_cgroup_commit_charge(page, memcg, false, true); |
00501b53 | 836 | lru_cache_add_active_or_unevictable(page, vma); |
6b0b50b0 | 837 | pgtable_trans_huge_deposit(mm, pmd, pgtable); |
71e3aac0 | 838 | set_pmd_at(mm, haddr, pmd, entry); |
71e3aac0 | 839 | add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); |
e1f56c89 | 840 | atomic_long_inc(&mm->nr_ptes); |
c4088ebd | 841 | spin_unlock(ptl); |
6b251fc9 | 842 | count_vm_event(THP_FAULT_ALLOC); |
71e3aac0 AA |
843 | } |
844 | ||
aa2e878e | 845 | return 0; |
71e3aac0 AA |
846 | } |
847 | ||
cc5d462f | 848 | static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp) |
0bbbc0b3 | 849 | { |
71baba4b | 850 | return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_RECLAIM)) | extra_gfp; |
0bbbc0b3 AA |
851 | } |
852 | ||
c4088ebd | 853 | /* Caller must hold page table lock. */ |
d295e341 | 854 | static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, |
97ae1749 | 855 | struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, |
5918d10a | 856 | struct page *zero_page) |
fc9fe822 KS |
857 | { |
858 | pmd_t entry; | |
7c414164 AM |
859 | if (!pmd_none(*pmd)) |
860 | return false; | |
5918d10a | 861 | entry = mk_pmd(zero_page, vma->vm_page_prot); |
fc9fe822 | 862 | entry = pmd_mkhuge(entry); |
6b0b50b0 | 863 | pgtable_trans_huge_deposit(mm, pmd, pgtable); |
fc9fe822 | 864 | set_pmd_at(mm, haddr, pmd, entry); |
e1f56c89 | 865 | atomic_long_inc(&mm->nr_ptes); |
7c414164 | 866 | return true; |
fc9fe822 KS |
867 | } |
868 | ||
71e3aac0 AA |
869 | int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma, |
870 | unsigned long address, pmd_t *pmd, | |
871 | unsigned int flags) | |
872 | { | |
077fcf11 | 873 | gfp_t gfp; |
71e3aac0 AA |
874 | struct page *page; |
875 | unsigned long haddr = address & HPAGE_PMD_MASK; | |
71e3aac0 | 876 | |
128ec037 | 877 | if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end) |
c0292554 | 878 | return VM_FAULT_FALLBACK; |
128ec037 KS |
879 | if (unlikely(anon_vma_prepare(vma))) |
880 | return VM_FAULT_OOM; | |
6d50e60c | 881 | if (unlikely(khugepaged_enter(vma, vma->vm_flags))) |
128ec037 | 882 | return VM_FAULT_OOM; |
593befa6 | 883 | if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) && |
128ec037 | 884 | transparent_hugepage_use_zero_page()) { |
c4088ebd | 885 | spinlock_t *ptl; |
128ec037 KS |
886 | pgtable_t pgtable; |
887 | struct page *zero_page; | |
888 | bool set; | |
6b251fc9 | 889 | int ret; |
128ec037 KS |
890 | pgtable = pte_alloc_one(mm, haddr); |
891 | if (unlikely(!pgtable)) | |
ba76149f | 892 | return VM_FAULT_OOM; |
128ec037 KS |
893 | zero_page = get_huge_zero_page(); |
894 | if (unlikely(!zero_page)) { | |
895 | pte_free(mm, pgtable); | |
81ab4201 | 896 | count_vm_event(THP_FAULT_FALLBACK); |
c0292554 | 897 | return VM_FAULT_FALLBACK; |
b9bbfbe3 | 898 | } |
c4088ebd | 899 | ptl = pmd_lock(mm, pmd); |
6b251fc9 AA |
900 | ret = 0; |
901 | set = false; | |
902 | if (pmd_none(*pmd)) { | |
903 | if (userfaultfd_missing(vma)) { | |
904 | spin_unlock(ptl); | |
230c92a8 | 905 | ret = handle_userfault(vma, address, flags, |
6b251fc9 AA |
906 | VM_UFFD_MISSING); |
907 | VM_BUG_ON(ret & VM_FAULT_FALLBACK); | |
908 | } else { | |
909 | set_huge_zero_page(pgtable, mm, vma, | |
910 | haddr, pmd, | |
911 | zero_page); | |
912 | spin_unlock(ptl); | |
913 | set = true; | |
914 | } | |
915 | } else | |
916 | spin_unlock(ptl); | |
128ec037 KS |
917 | if (!set) { |
918 | pte_free(mm, pgtable); | |
919 | put_huge_zero_page(); | |
edad9d2c | 920 | } |
6b251fc9 | 921 | return ret; |
71e3aac0 | 922 | } |
077fcf11 AK |
923 | gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0); |
924 | page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); | |
128ec037 KS |
925 | if (unlikely(!page)) { |
926 | count_vm_event(THP_FAULT_FALLBACK); | |
c0292554 | 927 | return VM_FAULT_FALLBACK; |
128ec037 | 928 | } |
9a982250 | 929 | prep_transhuge_page(page); |
230c92a8 AA |
930 | return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp, |
931 | flags); | |
71e3aac0 AA |
932 | } |
933 | ||
ae18d6dc | 934 | static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, |
f25748e3 | 935 | pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write) |
5cad465d MW |
936 | { |
937 | struct mm_struct *mm = vma->vm_mm; | |
938 | pmd_t entry; | |
939 | spinlock_t *ptl; | |
940 | ||
941 | ptl = pmd_lock(mm, pmd); | |
f25748e3 DW |
942 | entry = pmd_mkhuge(pfn_t_pmd(pfn, prot)); |
943 | if (pfn_t_devmap(pfn)) | |
944 | entry = pmd_mkdevmap(entry); | |
01871e59 RZ |
945 | if (write) { |
946 | entry = pmd_mkyoung(pmd_mkdirty(entry)); | |
947 | entry = maybe_pmd_mkwrite(entry, vma); | |
5cad465d | 948 | } |
01871e59 RZ |
949 | set_pmd_at(mm, addr, pmd, entry); |
950 | update_mmu_cache_pmd(vma, addr, pmd); | |
5cad465d | 951 | spin_unlock(ptl); |
5cad465d MW |
952 | } |
953 | ||
954 | int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, | |
f25748e3 | 955 | pmd_t *pmd, pfn_t pfn, bool write) |
5cad465d MW |
956 | { |
957 | pgprot_t pgprot = vma->vm_page_prot; | |
958 | /* | |
959 | * If we had pmd_special, we could avoid all these restrictions, | |
960 | * but we need to be consistent with PTEs and architectures that | |
961 | * can't support a 'special' bit. | |
962 | */ | |
963 | BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); | |
964 | BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == | |
965 | (VM_PFNMAP|VM_MIXEDMAP)); | |
966 | BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); | |
f25748e3 | 967 | BUG_ON(!pfn_t_devmap(pfn)); |
5cad465d MW |
968 | |
969 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
970 | return VM_FAULT_SIGBUS; | |
971 | if (track_pfn_insert(vma, &pgprot, pfn)) | |
972 | return VM_FAULT_SIGBUS; | |
ae18d6dc MW |
973 | insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write); |
974 | return VM_FAULT_NOPAGE; | |
5cad465d MW |
975 | } |
976 | ||
71e3aac0 AA |
977 | int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, |
978 | pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, | |
979 | struct vm_area_struct *vma) | |
980 | { | |
c4088ebd | 981 | spinlock_t *dst_ptl, *src_ptl; |
71e3aac0 AA |
982 | struct page *src_page; |
983 | pmd_t pmd; | |
984 | pgtable_t pgtable; | |
985 | int ret; | |
986 | ||
987 | ret = -ENOMEM; | |
988 | pgtable = pte_alloc_one(dst_mm, addr); | |
989 | if (unlikely(!pgtable)) | |
990 | goto out; | |
991 | ||
c4088ebd KS |
992 | dst_ptl = pmd_lock(dst_mm, dst_pmd); |
993 | src_ptl = pmd_lockptr(src_mm, src_pmd); | |
994 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
71e3aac0 AA |
995 | |
996 | ret = -EAGAIN; | |
997 | pmd = *src_pmd; | |
5c7fb56e | 998 | if (unlikely(!pmd_trans_huge(pmd) && !pmd_devmap(pmd))) { |
71e3aac0 AA |
999 | pte_free(dst_mm, pgtable); |
1000 | goto out_unlock; | |
1001 | } | |
fc9fe822 | 1002 | /* |
c4088ebd | 1003 | * When page table lock is held, the huge zero pmd should not be |
fc9fe822 KS |
1004 | * under splitting since we don't split the page itself, only pmd to |
1005 | * a page table. | |
1006 | */ | |
1007 | if (is_huge_zero_pmd(pmd)) { | |
5918d10a | 1008 | struct page *zero_page; |
97ae1749 KS |
1009 | /* |
1010 | * get_huge_zero_page() will never allocate a new page here, | |
1011 | * since we already have a zero page to copy. It just takes a | |
1012 | * reference. | |
1013 | */ | |
5918d10a | 1014 | zero_page = get_huge_zero_page(); |
6b251fc9 | 1015 | set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, |
5918d10a | 1016 | zero_page); |
fc9fe822 KS |
1017 | ret = 0; |
1018 | goto out_unlock; | |
1019 | } | |
de466bd6 | 1020 | |
5c7fb56e DW |
1021 | if (pmd_trans_huge(pmd)) { |
1022 | /* thp accounting separate from pmd_devmap accounting */ | |
1023 | src_page = pmd_page(pmd); | |
1024 | VM_BUG_ON_PAGE(!PageHead(src_page), src_page); | |
1025 | get_page(src_page); | |
1026 | page_dup_rmap(src_page, true); | |
1027 | add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); | |
1028 | atomic_long_inc(&dst_mm->nr_ptes); | |
1029 | pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); | |
1030 | } | |
71e3aac0 AA |
1031 | |
1032 | pmdp_set_wrprotect(src_mm, addr, src_pmd); | |
1033 | pmd = pmd_mkold(pmd_wrprotect(pmd)); | |
1034 | set_pmd_at(dst_mm, addr, dst_pmd, pmd); | |
71e3aac0 AA |
1035 | |
1036 | ret = 0; | |
1037 | out_unlock: | |
c4088ebd KS |
1038 | spin_unlock(src_ptl); |
1039 | spin_unlock(dst_ptl); | |
71e3aac0 AA |
1040 | out: |
1041 | return ret; | |
1042 | } | |
1043 | ||
a1dd450b WD |
1044 | void huge_pmd_set_accessed(struct mm_struct *mm, |
1045 | struct vm_area_struct *vma, | |
1046 | unsigned long address, | |
1047 | pmd_t *pmd, pmd_t orig_pmd, | |
1048 | int dirty) | |
1049 | { | |
c4088ebd | 1050 | spinlock_t *ptl; |
a1dd450b WD |
1051 | pmd_t entry; |
1052 | unsigned long haddr; | |
1053 | ||
c4088ebd | 1054 | ptl = pmd_lock(mm, pmd); |
a1dd450b WD |
1055 | if (unlikely(!pmd_same(*pmd, orig_pmd))) |
1056 | goto unlock; | |
1057 | ||
1058 | entry = pmd_mkyoung(orig_pmd); | |
1059 | haddr = address & HPAGE_PMD_MASK; | |
1060 | if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty)) | |
1061 | update_mmu_cache_pmd(vma, address, pmd); | |
1062 | ||
1063 | unlock: | |
c4088ebd | 1064 | spin_unlock(ptl); |
a1dd450b WD |
1065 | } |
1066 | ||
71e3aac0 AA |
1067 | static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm, |
1068 | struct vm_area_struct *vma, | |
1069 | unsigned long address, | |
1070 | pmd_t *pmd, pmd_t orig_pmd, | |
1071 | struct page *page, | |
1072 | unsigned long haddr) | |
1073 | { | |
00501b53 | 1074 | struct mem_cgroup *memcg; |
c4088ebd | 1075 | spinlock_t *ptl; |
71e3aac0 AA |
1076 | pgtable_t pgtable; |
1077 | pmd_t _pmd; | |
1078 | int ret = 0, i; | |
1079 | struct page **pages; | |
2ec74c3e SG |
1080 | unsigned long mmun_start; /* For mmu_notifiers */ |
1081 | unsigned long mmun_end; /* For mmu_notifiers */ | |
71e3aac0 AA |
1082 | |
1083 | pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR, | |
1084 | GFP_KERNEL); | |
1085 | if (unlikely(!pages)) { | |
1086 | ret |= VM_FAULT_OOM; | |
1087 | goto out; | |
1088 | } | |
1089 | ||
1090 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
cc5d462f AK |
1091 | pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE | |
1092 | __GFP_OTHER_NODE, | |
19ee151e | 1093 | vma, address, page_to_nid(page)); |
b9bbfbe3 | 1094 | if (unlikely(!pages[i] || |
00501b53 | 1095 | mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL, |
f627c2f5 | 1096 | &memcg, false))) { |
b9bbfbe3 | 1097 | if (pages[i]) |
71e3aac0 | 1098 | put_page(pages[i]); |
b9bbfbe3 | 1099 | while (--i >= 0) { |
00501b53 JW |
1100 | memcg = (void *)page_private(pages[i]); |
1101 | set_page_private(pages[i], 0); | |
f627c2f5 KS |
1102 | mem_cgroup_cancel_charge(pages[i], memcg, |
1103 | false); | |
b9bbfbe3 AA |
1104 | put_page(pages[i]); |
1105 | } | |
71e3aac0 AA |
1106 | kfree(pages); |
1107 | ret |= VM_FAULT_OOM; | |
1108 | goto out; | |
1109 | } | |
00501b53 | 1110 | set_page_private(pages[i], (unsigned long)memcg); |
71e3aac0 AA |
1111 | } |
1112 | ||
1113 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
1114 | copy_user_highpage(pages[i], page + i, | |
0089e485 | 1115 | haddr + PAGE_SIZE * i, vma); |
71e3aac0 AA |
1116 | __SetPageUptodate(pages[i]); |
1117 | cond_resched(); | |
1118 | } | |
1119 | ||
2ec74c3e SG |
1120 | mmun_start = haddr; |
1121 | mmun_end = haddr + HPAGE_PMD_SIZE; | |
1122 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
1123 | ||
c4088ebd | 1124 | ptl = pmd_lock(mm, pmd); |
71e3aac0 AA |
1125 | if (unlikely(!pmd_same(*pmd, orig_pmd))) |
1126 | goto out_free_pages; | |
309381fe | 1127 | VM_BUG_ON_PAGE(!PageHead(page), page); |
71e3aac0 | 1128 | |
8809aa2d | 1129 | pmdp_huge_clear_flush_notify(vma, haddr, pmd); |
71e3aac0 AA |
1130 | /* leave pmd empty until pte is filled */ |
1131 | ||
6b0b50b0 | 1132 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); |
71e3aac0 AA |
1133 | pmd_populate(mm, &_pmd, pgtable); |
1134 | ||
1135 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
1136 | pte_t *pte, entry; | |
1137 | entry = mk_pte(pages[i], vma->vm_page_prot); | |
1138 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); | |
00501b53 JW |
1139 | memcg = (void *)page_private(pages[i]); |
1140 | set_page_private(pages[i], 0); | |
d281ee61 | 1141 | page_add_new_anon_rmap(pages[i], vma, haddr, false); |
f627c2f5 | 1142 | mem_cgroup_commit_charge(pages[i], memcg, false, false); |
00501b53 | 1143 | lru_cache_add_active_or_unevictable(pages[i], vma); |
71e3aac0 AA |
1144 | pte = pte_offset_map(&_pmd, haddr); |
1145 | VM_BUG_ON(!pte_none(*pte)); | |
1146 | set_pte_at(mm, haddr, pte, entry); | |
1147 | pte_unmap(pte); | |
1148 | } | |
1149 | kfree(pages); | |
1150 | ||
71e3aac0 AA |
1151 | smp_wmb(); /* make pte visible before pmd */ |
1152 | pmd_populate(mm, pmd, pgtable); | |
d281ee61 | 1153 | page_remove_rmap(page, true); |
c4088ebd | 1154 | spin_unlock(ptl); |
71e3aac0 | 1155 | |
2ec74c3e SG |
1156 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
1157 | ||
71e3aac0 AA |
1158 | ret |= VM_FAULT_WRITE; |
1159 | put_page(page); | |
1160 | ||
1161 | out: | |
1162 | return ret; | |
1163 | ||
1164 | out_free_pages: | |
c4088ebd | 1165 | spin_unlock(ptl); |
2ec74c3e | 1166 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
b9bbfbe3 | 1167 | for (i = 0; i < HPAGE_PMD_NR; i++) { |
00501b53 JW |
1168 | memcg = (void *)page_private(pages[i]); |
1169 | set_page_private(pages[i], 0); | |
f627c2f5 | 1170 | mem_cgroup_cancel_charge(pages[i], memcg, false); |
71e3aac0 | 1171 | put_page(pages[i]); |
b9bbfbe3 | 1172 | } |
71e3aac0 AA |
1173 | kfree(pages); |
1174 | goto out; | |
1175 | } | |
1176 | ||
1177 | int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma, | |
1178 | unsigned long address, pmd_t *pmd, pmd_t orig_pmd) | |
1179 | { | |
c4088ebd | 1180 | spinlock_t *ptl; |
71e3aac0 | 1181 | int ret = 0; |
93b4796d | 1182 | struct page *page = NULL, *new_page; |
00501b53 | 1183 | struct mem_cgroup *memcg; |
71e3aac0 | 1184 | unsigned long haddr; |
2ec74c3e SG |
1185 | unsigned long mmun_start; /* For mmu_notifiers */ |
1186 | unsigned long mmun_end; /* For mmu_notifiers */ | |
3b363692 | 1187 | gfp_t huge_gfp; /* for allocation and charge */ |
71e3aac0 | 1188 | |
c4088ebd | 1189 | ptl = pmd_lockptr(mm, pmd); |
81d1b09c | 1190 | VM_BUG_ON_VMA(!vma->anon_vma, vma); |
93b4796d KS |
1191 | haddr = address & HPAGE_PMD_MASK; |
1192 | if (is_huge_zero_pmd(orig_pmd)) | |
1193 | goto alloc; | |
c4088ebd | 1194 | spin_lock(ptl); |
71e3aac0 AA |
1195 | if (unlikely(!pmd_same(*pmd, orig_pmd))) |
1196 | goto out_unlock; | |
1197 | ||
1198 | page = pmd_page(orig_pmd); | |
309381fe | 1199 | VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); |
1f25fe20 KS |
1200 | /* |
1201 | * We can only reuse the page if nobody else maps the huge page or it's | |
1202 | * part. We can do it by checking page_mapcount() on each sub-page, but | |
1203 | * it's expensive. | |
1204 | * The cheaper way is to check page_count() to be equal 1: every | |
1205 | * mapcount takes page reference reference, so this way we can | |
1206 | * guarantee, that the PMD is the only mapping. | |
1207 | * This can give false negative if somebody pinned the page, but that's | |
1208 | * fine. | |
1209 | */ | |
1210 | if (page_mapcount(page) == 1 && page_count(page) == 1) { | |
71e3aac0 AA |
1211 | pmd_t entry; |
1212 | entry = pmd_mkyoung(orig_pmd); | |
1213 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
1214 | if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1)) | |
b113da65 | 1215 | update_mmu_cache_pmd(vma, address, pmd); |
71e3aac0 AA |
1216 | ret |= VM_FAULT_WRITE; |
1217 | goto out_unlock; | |
1218 | } | |
ddc58f27 | 1219 | get_page(page); |
c4088ebd | 1220 | spin_unlock(ptl); |
93b4796d | 1221 | alloc: |
71e3aac0 | 1222 | if (transparent_hugepage_enabled(vma) && |
077fcf11 | 1223 | !transparent_hugepage_debug_cow()) { |
3b363692 MH |
1224 | huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0); |
1225 | new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER); | |
077fcf11 | 1226 | } else |
71e3aac0 AA |
1227 | new_page = NULL; |
1228 | ||
9a982250 KS |
1229 | if (likely(new_page)) { |
1230 | prep_transhuge_page(new_page); | |
1231 | } else { | |
eecc1e42 | 1232 | if (!page) { |
78ddc534 | 1233 | split_huge_pmd(vma, pmd, address); |
e9b71ca9 | 1234 | ret |= VM_FAULT_FALLBACK; |
93b4796d KS |
1235 | } else { |
1236 | ret = do_huge_pmd_wp_page_fallback(mm, vma, address, | |
1237 | pmd, orig_pmd, page, haddr); | |
9845cbbd | 1238 | if (ret & VM_FAULT_OOM) { |
78ddc534 | 1239 | split_huge_pmd(vma, pmd, address); |
9845cbbd KS |
1240 | ret |= VM_FAULT_FALLBACK; |
1241 | } | |
ddc58f27 | 1242 | put_page(page); |
93b4796d | 1243 | } |
17766dde | 1244 | count_vm_event(THP_FAULT_FALLBACK); |
71e3aac0 AA |
1245 | goto out; |
1246 | } | |
1247 | ||
f627c2f5 KS |
1248 | if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg, |
1249 | true))) { | |
b9bbfbe3 | 1250 | put_page(new_page); |
93b4796d | 1251 | if (page) { |
78ddc534 | 1252 | split_huge_pmd(vma, pmd, address); |
ddc58f27 | 1253 | put_page(page); |
9845cbbd | 1254 | } else |
78ddc534 | 1255 | split_huge_pmd(vma, pmd, address); |
9845cbbd | 1256 | ret |= VM_FAULT_FALLBACK; |
17766dde | 1257 | count_vm_event(THP_FAULT_FALLBACK); |
b9bbfbe3 AA |
1258 | goto out; |
1259 | } | |
1260 | ||
17766dde DR |
1261 | count_vm_event(THP_FAULT_ALLOC); |
1262 | ||
eecc1e42 | 1263 | if (!page) |
93b4796d KS |
1264 | clear_huge_page(new_page, haddr, HPAGE_PMD_NR); |
1265 | else | |
1266 | copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR); | |
71e3aac0 AA |
1267 | __SetPageUptodate(new_page); |
1268 | ||
2ec74c3e SG |
1269 | mmun_start = haddr; |
1270 | mmun_end = haddr + HPAGE_PMD_SIZE; | |
1271 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
1272 | ||
c4088ebd | 1273 | spin_lock(ptl); |
93b4796d | 1274 | if (page) |
ddc58f27 | 1275 | put_page(page); |
b9bbfbe3 | 1276 | if (unlikely(!pmd_same(*pmd, orig_pmd))) { |
c4088ebd | 1277 | spin_unlock(ptl); |
f627c2f5 | 1278 | mem_cgroup_cancel_charge(new_page, memcg, true); |
71e3aac0 | 1279 | put_page(new_page); |
2ec74c3e | 1280 | goto out_mn; |
b9bbfbe3 | 1281 | } else { |
71e3aac0 | 1282 | pmd_t entry; |
3122359a KS |
1283 | entry = mk_huge_pmd(new_page, vma->vm_page_prot); |
1284 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
8809aa2d | 1285 | pmdp_huge_clear_flush_notify(vma, haddr, pmd); |
d281ee61 | 1286 | page_add_new_anon_rmap(new_page, vma, haddr, true); |
f627c2f5 | 1287 | mem_cgroup_commit_charge(new_page, memcg, false, true); |
00501b53 | 1288 | lru_cache_add_active_or_unevictable(new_page, vma); |
71e3aac0 | 1289 | set_pmd_at(mm, haddr, pmd, entry); |
b113da65 | 1290 | update_mmu_cache_pmd(vma, address, pmd); |
eecc1e42 | 1291 | if (!page) { |
93b4796d | 1292 | add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); |
97ae1749 KS |
1293 | put_huge_zero_page(); |
1294 | } else { | |
309381fe | 1295 | VM_BUG_ON_PAGE(!PageHead(page), page); |
d281ee61 | 1296 | page_remove_rmap(page, true); |
93b4796d KS |
1297 | put_page(page); |
1298 | } | |
71e3aac0 AA |
1299 | ret |= VM_FAULT_WRITE; |
1300 | } | |
c4088ebd | 1301 | spin_unlock(ptl); |
2ec74c3e SG |
1302 | out_mn: |
1303 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); | |
71e3aac0 AA |
1304 | out: |
1305 | return ret; | |
2ec74c3e | 1306 | out_unlock: |
c4088ebd | 1307 | spin_unlock(ptl); |
2ec74c3e | 1308 | return ret; |
71e3aac0 AA |
1309 | } |
1310 | ||
b676b293 | 1311 | struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, |
71e3aac0 AA |
1312 | unsigned long addr, |
1313 | pmd_t *pmd, | |
1314 | unsigned int flags) | |
1315 | { | |
b676b293 | 1316 | struct mm_struct *mm = vma->vm_mm; |
71e3aac0 AA |
1317 | struct page *page = NULL; |
1318 | ||
c4088ebd | 1319 | assert_spin_locked(pmd_lockptr(mm, pmd)); |
71e3aac0 AA |
1320 | |
1321 | if (flags & FOLL_WRITE && !pmd_write(*pmd)) | |
1322 | goto out; | |
1323 | ||
85facf25 KS |
1324 | /* Avoid dumping huge zero page */ |
1325 | if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) | |
1326 | return ERR_PTR(-EFAULT); | |
1327 | ||
2b4847e7 | 1328 | /* Full NUMA hinting faults to serialise migration in fault paths */ |
8a0516ed | 1329 | if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) |
2b4847e7 MG |
1330 | goto out; |
1331 | ||
71e3aac0 | 1332 | page = pmd_page(*pmd); |
309381fe | 1333 | VM_BUG_ON_PAGE(!PageHead(page), page); |
71e3aac0 AA |
1334 | if (flags & FOLL_TOUCH) { |
1335 | pmd_t _pmd; | |
1336 | /* | |
1337 | * We should set the dirty bit only for FOLL_WRITE but | |
1338 | * for now the dirty bit in the pmd is meaningless. | |
1339 | * And if the dirty bit will become meaningful and | |
1340 | * we'll only set it with FOLL_WRITE, an atomic | |
1341 | * set_bit will be required on the pmd to set the | |
1342 | * young bit, instead of the current set_pmd_at. | |
1343 | */ | |
1344 | _pmd = pmd_mkyoung(pmd_mkdirty(*pmd)); | |
8663890a AK |
1345 | if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, |
1346 | pmd, _pmd, 1)) | |
1347 | update_mmu_cache_pmd(vma, addr, pmd); | |
71e3aac0 | 1348 | } |
de60f5f1 | 1349 | if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { |
e90309c9 KS |
1350 | /* |
1351 | * We don't mlock() pte-mapped THPs. This way we can avoid | |
1352 | * leaking mlocked pages into non-VM_LOCKED VMAs. | |
1353 | * | |
1354 | * In most cases the pmd is the only mapping of the page as we | |
1355 | * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for | |
1356 | * writable private mappings in populate_vma_page_range(). | |
1357 | * | |
1358 | * The only scenario when we have the page shared here is if we | |
1359 | * mlocking read-only mapping shared over fork(). We skip | |
1360 | * mlocking such pages. | |
1361 | */ | |
1362 | if (compound_mapcount(page) == 1 && !PageDoubleMap(page) && | |
1363 | page->mapping && trylock_page(page)) { | |
b676b293 DR |
1364 | lru_add_drain(); |
1365 | if (page->mapping) | |
1366 | mlock_vma_page(page); | |
1367 | unlock_page(page); | |
1368 | } | |
1369 | } | |
71e3aac0 | 1370 | page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; |
309381fe | 1371 | VM_BUG_ON_PAGE(!PageCompound(page), page); |
71e3aac0 | 1372 | if (flags & FOLL_GET) |
ddc58f27 | 1373 | get_page(page); |
71e3aac0 AA |
1374 | |
1375 | out: | |
1376 | return page; | |
1377 | } | |
1378 | ||
d10e63f2 | 1379 | /* NUMA hinting page fault entry point for trans huge pmds */ |
4daae3b4 MG |
1380 | int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma, |
1381 | unsigned long addr, pmd_t pmd, pmd_t *pmdp) | |
d10e63f2 | 1382 | { |
c4088ebd | 1383 | spinlock_t *ptl; |
b8916634 | 1384 | struct anon_vma *anon_vma = NULL; |
b32967ff | 1385 | struct page *page; |
d10e63f2 | 1386 | unsigned long haddr = addr & HPAGE_PMD_MASK; |
8191acbd | 1387 | int page_nid = -1, this_nid = numa_node_id(); |
90572890 | 1388 | int target_nid, last_cpupid = -1; |
8191acbd MG |
1389 | bool page_locked; |
1390 | bool migrated = false; | |
b191f9b1 | 1391 | bool was_writable; |
6688cc05 | 1392 | int flags = 0; |
d10e63f2 | 1393 | |
c0e7cad9 MG |
1394 | /* A PROT_NONE fault should not end up here */ |
1395 | BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))); | |
1396 | ||
c4088ebd | 1397 | ptl = pmd_lock(mm, pmdp); |
d10e63f2 MG |
1398 | if (unlikely(!pmd_same(pmd, *pmdp))) |
1399 | goto out_unlock; | |
1400 | ||
de466bd6 MG |
1401 | /* |
1402 | * If there are potential migrations, wait for completion and retry | |
1403 | * without disrupting NUMA hinting information. Do not relock and | |
1404 | * check_same as the page may no longer be mapped. | |
1405 | */ | |
1406 | if (unlikely(pmd_trans_migrating(*pmdp))) { | |
5d833062 | 1407 | page = pmd_page(*pmdp); |
de466bd6 | 1408 | spin_unlock(ptl); |
5d833062 | 1409 | wait_on_page_locked(page); |
de466bd6 MG |
1410 | goto out; |
1411 | } | |
1412 | ||
d10e63f2 | 1413 | page = pmd_page(pmd); |
a1a46184 | 1414 | BUG_ON(is_huge_zero_page(page)); |
8191acbd | 1415 | page_nid = page_to_nid(page); |
90572890 | 1416 | last_cpupid = page_cpupid_last(page); |
03c5a6e1 | 1417 | count_vm_numa_event(NUMA_HINT_FAULTS); |
04bb2f94 | 1418 | if (page_nid == this_nid) { |
03c5a6e1 | 1419 | count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); |
04bb2f94 RR |
1420 | flags |= TNF_FAULT_LOCAL; |
1421 | } | |
4daae3b4 | 1422 | |
bea66fbd MG |
1423 | /* See similar comment in do_numa_page for explanation */ |
1424 | if (!(vma->vm_flags & VM_WRITE)) | |
6688cc05 PZ |
1425 | flags |= TNF_NO_GROUP; |
1426 | ||
ff9042b1 MG |
1427 | /* |
1428 | * Acquire the page lock to serialise THP migrations but avoid dropping | |
1429 | * page_table_lock if at all possible | |
1430 | */ | |
b8916634 MG |
1431 | page_locked = trylock_page(page); |
1432 | target_nid = mpol_misplaced(page, vma, haddr); | |
1433 | if (target_nid == -1) { | |
1434 | /* If the page was locked, there are no parallel migrations */ | |
a54a407f | 1435 | if (page_locked) |
b8916634 | 1436 | goto clear_pmdnuma; |
2b4847e7 | 1437 | } |
4daae3b4 | 1438 | |
de466bd6 | 1439 | /* Migration could have started since the pmd_trans_migrating check */ |
2b4847e7 | 1440 | if (!page_locked) { |
c4088ebd | 1441 | spin_unlock(ptl); |
b8916634 | 1442 | wait_on_page_locked(page); |
a54a407f | 1443 | page_nid = -1; |
b8916634 MG |
1444 | goto out; |
1445 | } | |
1446 | ||
2b4847e7 MG |
1447 | /* |
1448 | * Page is misplaced. Page lock serialises migrations. Acquire anon_vma | |
1449 | * to serialises splits | |
1450 | */ | |
b8916634 | 1451 | get_page(page); |
c4088ebd | 1452 | spin_unlock(ptl); |
b8916634 | 1453 | anon_vma = page_lock_anon_vma_read(page); |
4daae3b4 | 1454 | |
c69307d5 | 1455 | /* Confirm the PMD did not change while page_table_lock was released */ |
c4088ebd | 1456 | spin_lock(ptl); |
b32967ff MG |
1457 | if (unlikely(!pmd_same(pmd, *pmdp))) { |
1458 | unlock_page(page); | |
1459 | put_page(page); | |
a54a407f | 1460 | page_nid = -1; |
4daae3b4 | 1461 | goto out_unlock; |
b32967ff | 1462 | } |
ff9042b1 | 1463 | |
c3a489ca MG |
1464 | /* Bail if we fail to protect against THP splits for any reason */ |
1465 | if (unlikely(!anon_vma)) { | |
1466 | put_page(page); | |
1467 | page_nid = -1; | |
1468 | goto clear_pmdnuma; | |
1469 | } | |
1470 | ||
a54a407f MG |
1471 | /* |
1472 | * Migrate the THP to the requested node, returns with page unlocked | |
8a0516ed | 1473 | * and access rights restored. |
a54a407f | 1474 | */ |
c4088ebd | 1475 | spin_unlock(ptl); |
b32967ff | 1476 | migrated = migrate_misplaced_transhuge_page(mm, vma, |
340ef390 | 1477 | pmdp, pmd, addr, page, target_nid); |
6688cc05 PZ |
1478 | if (migrated) { |
1479 | flags |= TNF_MIGRATED; | |
8191acbd | 1480 | page_nid = target_nid; |
074c2381 MG |
1481 | } else |
1482 | flags |= TNF_MIGRATE_FAIL; | |
b32967ff | 1483 | |
8191acbd | 1484 | goto out; |
b32967ff | 1485 | clear_pmdnuma: |
a54a407f | 1486 | BUG_ON(!PageLocked(page)); |
b191f9b1 | 1487 | was_writable = pmd_write(pmd); |
4d942466 | 1488 | pmd = pmd_modify(pmd, vma->vm_page_prot); |
b7b04004 | 1489 | pmd = pmd_mkyoung(pmd); |
b191f9b1 MG |
1490 | if (was_writable) |
1491 | pmd = pmd_mkwrite(pmd); | |
d10e63f2 | 1492 | set_pmd_at(mm, haddr, pmdp, pmd); |
d10e63f2 | 1493 | update_mmu_cache_pmd(vma, addr, pmdp); |
a54a407f | 1494 | unlock_page(page); |
d10e63f2 | 1495 | out_unlock: |
c4088ebd | 1496 | spin_unlock(ptl); |
b8916634 MG |
1497 | |
1498 | out: | |
1499 | if (anon_vma) | |
1500 | page_unlock_anon_vma_read(anon_vma); | |
1501 | ||
8191acbd | 1502 | if (page_nid != -1) |
6688cc05 | 1503 | task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags); |
8191acbd | 1504 | |
d10e63f2 MG |
1505 | return 0; |
1506 | } | |
1507 | ||
b8d3c4c3 MK |
1508 | int madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, |
1509 | pmd_t *pmd, unsigned long addr, unsigned long next) | |
1510 | ||
1511 | { | |
1512 | spinlock_t *ptl; | |
1513 | pmd_t orig_pmd; | |
1514 | struct page *page; | |
1515 | struct mm_struct *mm = tlb->mm; | |
1516 | int ret = 0; | |
1517 | ||
1518 | if (!pmd_trans_huge_lock(pmd, vma, &ptl)) | |
1519 | goto out; | |
1520 | ||
1521 | orig_pmd = *pmd; | |
1522 | if (is_huge_zero_pmd(orig_pmd)) { | |
1523 | ret = 1; | |
1524 | goto out; | |
1525 | } | |
1526 | ||
1527 | page = pmd_page(orig_pmd); | |
1528 | /* | |
1529 | * If other processes are mapping this page, we couldn't discard | |
1530 | * the page unless they all do MADV_FREE so let's skip the page. | |
1531 | */ | |
1532 | if (page_mapcount(page) != 1) | |
1533 | goto out; | |
1534 | ||
1535 | if (!trylock_page(page)) | |
1536 | goto out; | |
1537 | ||
1538 | /* | |
1539 | * If user want to discard part-pages of THP, split it so MADV_FREE | |
1540 | * will deactivate only them. | |
1541 | */ | |
1542 | if (next - addr != HPAGE_PMD_SIZE) { | |
1543 | get_page(page); | |
1544 | spin_unlock(ptl); | |
1545 | if (split_huge_page(page)) { | |
1546 | put_page(page); | |
1547 | unlock_page(page); | |
1548 | goto out_unlocked; | |
1549 | } | |
1550 | put_page(page); | |
1551 | unlock_page(page); | |
1552 | ret = 1; | |
1553 | goto out_unlocked; | |
1554 | } | |
1555 | ||
1556 | if (PageDirty(page)) | |
1557 | ClearPageDirty(page); | |
1558 | unlock_page(page); | |
1559 | ||
1560 | if (PageActive(page)) | |
1561 | deactivate_page(page); | |
1562 | ||
1563 | if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { | |
1564 | orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd, | |
1565 | tlb->fullmm); | |
1566 | orig_pmd = pmd_mkold(orig_pmd); | |
1567 | orig_pmd = pmd_mkclean(orig_pmd); | |
1568 | ||
1569 | set_pmd_at(mm, addr, pmd, orig_pmd); | |
1570 | tlb_remove_pmd_tlb_entry(tlb, pmd, addr); | |
1571 | } | |
1572 | ret = 1; | |
1573 | out: | |
1574 | spin_unlock(ptl); | |
1575 | out_unlocked: | |
1576 | return ret; | |
1577 | } | |
1578 | ||
71e3aac0 | 1579 | int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, |
f21760b1 | 1580 | pmd_t *pmd, unsigned long addr) |
71e3aac0 | 1581 | { |
da146769 | 1582 | pmd_t orig_pmd; |
bf929152 | 1583 | spinlock_t *ptl; |
71e3aac0 | 1584 | |
4b471e88 | 1585 | if (!__pmd_trans_huge_lock(pmd, vma, &ptl)) |
da146769 KS |
1586 | return 0; |
1587 | /* | |
1588 | * For architectures like ppc64 we look at deposited pgtable | |
1589 | * when calling pmdp_huge_get_and_clear. So do the | |
1590 | * pgtable_trans_huge_withdraw after finishing pmdp related | |
1591 | * operations. | |
1592 | */ | |
1593 | orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd, | |
1594 | tlb->fullmm); | |
1595 | tlb_remove_pmd_tlb_entry(tlb, pmd, addr); | |
1596 | if (vma_is_dax(vma)) { | |
1597 | spin_unlock(ptl); | |
1598 | if (is_huge_zero_pmd(orig_pmd)) | |
97ae1749 | 1599 | put_huge_zero_page(); |
da146769 KS |
1600 | } else if (is_huge_zero_pmd(orig_pmd)) { |
1601 | pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd)); | |
1602 | atomic_long_dec(&tlb->mm->nr_ptes); | |
1603 | spin_unlock(ptl); | |
1604 | put_huge_zero_page(); | |
1605 | } else { | |
1606 | struct page *page = pmd_page(orig_pmd); | |
d281ee61 | 1607 | page_remove_rmap(page, true); |
da146769 KS |
1608 | VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); |
1609 | add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); | |
1610 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
1611 | pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd)); | |
1612 | atomic_long_dec(&tlb->mm->nr_ptes); | |
1613 | spin_unlock(ptl); | |
1614 | tlb_remove_page(tlb, page); | |
025c5b24 | 1615 | } |
da146769 | 1616 | return 1; |
71e3aac0 AA |
1617 | } |
1618 | ||
4b471e88 | 1619 | bool move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma, |
37a1c49a AA |
1620 | unsigned long old_addr, |
1621 | unsigned long new_addr, unsigned long old_end, | |
1622 | pmd_t *old_pmd, pmd_t *new_pmd) | |
1623 | { | |
bf929152 | 1624 | spinlock_t *old_ptl, *new_ptl; |
37a1c49a AA |
1625 | pmd_t pmd; |
1626 | ||
1627 | struct mm_struct *mm = vma->vm_mm; | |
1628 | ||
1629 | if ((old_addr & ~HPAGE_PMD_MASK) || | |
1630 | (new_addr & ~HPAGE_PMD_MASK) || | |
1631 | old_end - old_addr < HPAGE_PMD_SIZE || | |
1632 | (new_vma->vm_flags & VM_NOHUGEPAGE)) | |
4b471e88 | 1633 | return false; |
37a1c49a AA |
1634 | |
1635 | /* | |
1636 | * The destination pmd shouldn't be established, free_pgtables() | |
1637 | * should have release it. | |
1638 | */ | |
1639 | if (WARN_ON(!pmd_none(*new_pmd))) { | |
1640 | VM_BUG_ON(pmd_trans_huge(*new_pmd)); | |
4b471e88 | 1641 | return false; |
37a1c49a AA |
1642 | } |
1643 | ||
bf929152 KS |
1644 | /* |
1645 | * We don't have to worry about the ordering of src and dst | |
1646 | * ptlocks because exclusive mmap_sem prevents deadlock. | |
1647 | */ | |
4b471e88 | 1648 | if (__pmd_trans_huge_lock(old_pmd, vma, &old_ptl)) { |
bf929152 KS |
1649 | new_ptl = pmd_lockptr(mm, new_pmd); |
1650 | if (new_ptl != old_ptl) | |
1651 | spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); | |
8809aa2d | 1652 | pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); |
025c5b24 | 1653 | VM_BUG_ON(!pmd_none(*new_pmd)); |
3592806c | 1654 | |
b3084f4d AK |
1655 | if (pmd_move_must_withdraw(new_ptl, old_ptl)) { |
1656 | pgtable_t pgtable; | |
3592806c KS |
1657 | pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); |
1658 | pgtable_trans_huge_deposit(mm, new_pmd, pgtable); | |
3592806c | 1659 | } |
b3084f4d AK |
1660 | set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd)); |
1661 | if (new_ptl != old_ptl) | |
1662 | spin_unlock(new_ptl); | |
bf929152 | 1663 | spin_unlock(old_ptl); |
4b471e88 | 1664 | return true; |
37a1c49a | 1665 | } |
4b471e88 | 1666 | return false; |
37a1c49a AA |
1667 | } |
1668 | ||
f123d74a MG |
1669 | /* |
1670 | * Returns | |
1671 | * - 0 if PMD could not be locked | |
1672 | * - 1 if PMD was locked but protections unchange and TLB flush unnecessary | |
1673 | * - HPAGE_PMD_NR is protections changed and TLB flush necessary | |
1674 | */ | |
cd7548ab | 1675 | int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, |
e944fd67 | 1676 | unsigned long addr, pgprot_t newprot, int prot_numa) |
cd7548ab JW |
1677 | { |
1678 | struct mm_struct *mm = vma->vm_mm; | |
bf929152 | 1679 | spinlock_t *ptl; |
cd7548ab JW |
1680 | int ret = 0; |
1681 | ||
4b471e88 | 1682 | if (__pmd_trans_huge_lock(pmd, vma, &ptl)) { |
025c5b24 | 1683 | pmd_t entry; |
b191f9b1 | 1684 | bool preserve_write = prot_numa && pmd_write(*pmd); |
ba68bc01 | 1685 | ret = 1; |
e944fd67 MG |
1686 | |
1687 | /* | |
1688 | * Avoid trapping faults against the zero page. The read-only | |
1689 | * data is likely to be read-cached on the local CPU and | |
1690 | * local/remote hits to the zero page are not interesting. | |
1691 | */ | |
1692 | if (prot_numa && is_huge_zero_pmd(*pmd)) { | |
1693 | spin_unlock(ptl); | |
ba68bc01 | 1694 | return ret; |
e944fd67 MG |
1695 | } |
1696 | ||
10c1045f | 1697 | if (!prot_numa || !pmd_protnone(*pmd)) { |
8809aa2d | 1698 | entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd); |
10c1045f | 1699 | entry = pmd_modify(entry, newprot); |
b191f9b1 MG |
1700 | if (preserve_write) |
1701 | entry = pmd_mkwrite(entry); | |
10c1045f MG |
1702 | ret = HPAGE_PMD_NR; |
1703 | set_pmd_at(mm, addr, pmd, entry); | |
b191f9b1 | 1704 | BUG_ON(!preserve_write && pmd_write(entry)); |
10c1045f | 1705 | } |
bf929152 | 1706 | spin_unlock(ptl); |
025c5b24 NH |
1707 | } |
1708 | ||
1709 | return ret; | |
1710 | } | |
1711 | ||
1712 | /* | |
4b471e88 | 1713 | * Returns true if a given pmd maps a thp, false otherwise. |
025c5b24 | 1714 | * |
4b471e88 KS |
1715 | * Note that if it returns true, this routine returns without unlocking page |
1716 | * table lock. So callers must unlock it. | |
025c5b24 | 1717 | */ |
4b471e88 | 1718 | bool __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma, |
bf929152 | 1719 | spinlock_t **ptl) |
025c5b24 | 1720 | { |
bf929152 | 1721 | *ptl = pmd_lock(vma->vm_mm, pmd); |
5c7fb56e | 1722 | if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd))) |
4b471e88 | 1723 | return true; |
bf929152 | 1724 | spin_unlock(*ptl); |
4b471e88 | 1725 | return false; |
cd7548ab JW |
1726 | } |
1727 | ||
9050d7eb | 1728 | #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE) |
78f11a25 | 1729 | |
60ab3244 AA |
1730 | int hugepage_madvise(struct vm_area_struct *vma, |
1731 | unsigned long *vm_flags, int advice) | |
0af4e98b | 1732 | { |
a664b2d8 AA |
1733 | switch (advice) { |
1734 | case MADV_HUGEPAGE: | |
1e1836e8 AT |
1735 | #ifdef CONFIG_S390 |
1736 | /* | |
1737 | * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390 | |
1738 | * can't handle this properly after s390_enable_sie, so we simply | |
1739 | * ignore the madvise to prevent qemu from causing a SIGSEGV. | |
1740 | */ | |
1741 | if (mm_has_pgste(vma->vm_mm)) | |
1742 | return 0; | |
1743 | #endif | |
a664b2d8 AA |
1744 | /* |
1745 | * Be somewhat over-protective like KSM for now! | |
1746 | */ | |
1a763615 | 1747 | if (*vm_flags & VM_NO_THP) |
a664b2d8 AA |
1748 | return -EINVAL; |
1749 | *vm_flags &= ~VM_NOHUGEPAGE; | |
1750 | *vm_flags |= VM_HUGEPAGE; | |
60ab3244 AA |
1751 | /* |
1752 | * If the vma become good for khugepaged to scan, | |
1753 | * register it here without waiting a page fault that | |
1754 | * may not happen any time soon. | |
1755 | */ | |
6d50e60c | 1756 | if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags))) |
60ab3244 | 1757 | return -ENOMEM; |
a664b2d8 AA |
1758 | break; |
1759 | case MADV_NOHUGEPAGE: | |
1760 | /* | |
1761 | * Be somewhat over-protective like KSM for now! | |
1762 | */ | |
1a763615 | 1763 | if (*vm_flags & VM_NO_THP) |
a664b2d8 AA |
1764 | return -EINVAL; |
1765 | *vm_flags &= ~VM_HUGEPAGE; | |
1766 | *vm_flags |= VM_NOHUGEPAGE; | |
60ab3244 AA |
1767 | /* |
1768 | * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning | |
1769 | * this vma even if we leave the mm registered in khugepaged if | |
1770 | * it got registered before VM_NOHUGEPAGE was set. | |
1771 | */ | |
a664b2d8 AA |
1772 | break; |
1773 | } | |
0af4e98b AA |
1774 | |
1775 | return 0; | |
1776 | } | |
1777 | ||
ba76149f AA |
1778 | static int __init khugepaged_slab_init(void) |
1779 | { | |
1780 | mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", | |
1781 | sizeof(struct mm_slot), | |
1782 | __alignof__(struct mm_slot), 0, NULL); | |
1783 | if (!mm_slot_cache) | |
1784 | return -ENOMEM; | |
1785 | ||
1786 | return 0; | |
1787 | } | |
1788 | ||
65ebb64f KS |
1789 | static void __init khugepaged_slab_exit(void) |
1790 | { | |
1791 | kmem_cache_destroy(mm_slot_cache); | |
1792 | } | |
1793 | ||
ba76149f AA |
1794 | static inline struct mm_slot *alloc_mm_slot(void) |
1795 | { | |
1796 | if (!mm_slot_cache) /* initialization failed */ | |
1797 | return NULL; | |
1798 | return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); | |
1799 | } | |
1800 | ||
1801 | static inline void free_mm_slot(struct mm_slot *mm_slot) | |
1802 | { | |
1803 | kmem_cache_free(mm_slot_cache, mm_slot); | |
1804 | } | |
1805 | ||
ba76149f AA |
1806 | static struct mm_slot *get_mm_slot(struct mm_struct *mm) |
1807 | { | |
1808 | struct mm_slot *mm_slot; | |
ba76149f | 1809 | |
b67bfe0d | 1810 | hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm) |
ba76149f AA |
1811 | if (mm == mm_slot->mm) |
1812 | return mm_slot; | |
43b5fbbd | 1813 | |
ba76149f AA |
1814 | return NULL; |
1815 | } | |
1816 | ||
1817 | static void insert_to_mm_slots_hash(struct mm_struct *mm, | |
1818 | struct mm_slot *mm_slot) | |
1819 | { | |
ba76149f | 1820 | mm_slot->mm = mm; |
43b5fbbd | 1821 | hash_add(mm_slots_hash, &mm_slot->hash, (long)mm); |
ba76149f AA |
1822 | } |
1823 | ||
1824 | static inline int khugepaged_test_exit(struct mm_struct *mm) | |
1825 | { | |
1826 | return atomic_read(&mm->mm_users) == 0; | |
1827 | } | |
1828 | ||
1829 | int __khugepaged_enter(struct mm_struct *mm) | |
1830 | { | |
1831 | struct mm_slot *mm_slot; | |
1832 | int wakeup; | |
1833 | ||
1834 | mm_slot = alloc_mm_slot(); | |
1835 | if (!mm_slot) | |
1836 | return -ENOMEM; | |
1837 | ||
1838 | /* __khugepaged_exit() must not run from under us */ | |
96dad67f | 1839 | VM_BUG_ON_MM(khugepaged_test_exit(mm), mm); |
ba76149f AA |
1840 | if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { |
1841 | free_mm_slot(mm_slot); | |
1842 | return 0; | |
1843 | } | |
1844 | ||
1845 | spin_lock(&khugepaged_mm_lock); | |
1846 | insert_to_mm_slots_hash(mm, mm_slot); | |
1847 | /* | |
1848 | * Insert just behind the scanning cursor, to let the area settle | |
1849 | * down a little. | |
1850 | */ | |
1851 | wakeup = list_empty(&khugepaged_scan.mm_head); | |
1852 | list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head); | |
1853 | spin_unlock(&khugepaged_mm_lock); | |
1854 | ||
1855 | atomic_inc(&mm->mm_count); | |
1856 | if (wakeup) | |
1857 | wake_up_interruptible(&khugepaged_wait); | |
1858 | ||
1859 | return 0; | |
1860 | } | |
1861 | ||
6d50e60c DR |
1862 | int khugepaged_enter_vma_merge(struct vm_area_struct *vma, |
1863 | unsigned long vm_flags) | |
ba76149f AA |
1864 | { |
1865 | unsigned long hstart, hend; | |
1866 | if (!vma->anon_vma) | |
1867 | /* | |
1868 | * Not yet faulted in so we will register later in the | |
1869 | * page fault if needed. | |
1870 | */ | |
1871 | return 0; | |
78f11a25 | 1872 | if (vma->vm_ops) |
ba76149f AA |
1873 | /* khugepaged not yet working on file or special mappings */ |
1874 | return 0; | |
6d50e60c | 1875 | VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma); |
ba76149f AA |
1876 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; |
1877 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
1878 | if (hstart < hend) | |
6d50e60c | 1879 | return khugepaged_enter(vma, vm_flags); |
ba76149f AA |
1880 | return 0; |
1881 | } | |
1882 | ||
1883 | void __khugepaged_exit(struct mm_struct *mm) | |
1884 | { | |
1885 | struct mm_slot *mm_slot; | |
1886 | int free = 0; | |
1887 | ||
1888 | spin_lock(&khugepaged_mm_lock); | |
1889 | mm_slot = get_mm_slot(mm); | |
1890 | if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { | |
43b5fbbd | 1891 | hash_del(&mm_slot->hash); |
ba76149f AA |
1892 | list_del(&mm_slot->mm_node); |
1893 | free = 1; | |
1894 | } | |
d788e80a | 1895 | spin_unlock(&khugepaged_mm_lock); |
ba76149f AA |
1896 | |
1897 | if (free) { | |
ba76149f AA |
1898 | clear_bit(MMF_VM_HUGEPAGE, &mm->flags); |
1899 | free_mm_slot(mm_slot); | |
1900 | mmdrop(mm); | |
1901 | } else if (mm_slot) { | |
ba76149f AA |
1902 | /* |
1903 | * This is required to serialize against | |
1904 | * khugepaged_test_exit() (which is guaranteed to run | |
1905 | * under mmap sem read mode). Stop here (after we | |
1906 | * return all pagetables will be destroyed) until | |
1907 | * khugepaged has finished working on the pagetables | |
1908 | * under the mmap_sem. | |
1909 | */ | |
1910 | down_write(&mm->mmap_sem); | |
1911 | up_write(&mm->mmap_sem); | |
d788e80a | 1912 | } |
ba76149f AA |
1913 | } |
1914 | ||
1915 | static void release_pte_page(struct page *page) | |
1916 | { | |
1917 | /* 0 stands for page_is_file_cache(page) == false */ | |
1918 | dec_zone_page_state(page, NR_ISOLATED_ANON + 0); | |
1919 | unlock_page(page); | |
1920 | putback_lru_page(page); | |
1921 | } | |
1922 | ||
1923 | static void release_pte_pages(pte_t *pte, pte_t *_pte) | |
1924 | { | |
1925 | while (--_pte >= pte) { | |
1926 | pte_t pteval = *_pte; | |
ca0984ca | 1927 | if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval))) |
ba76149f AA |
1928 | release_pte_page(pte_page(pteval)); |
1929 | } | |
1930 | } | |
1931 | ||
ba76149f AA |
1932 | static int __collapse_huge_page_isolate(struct vm_area_struct *vma, |
1933 | unsigned long address, | |
1934 | pte_t *pte) | |
1935 | { | |
7d2eba05 | 1936 | struct page *page = NULL; |
ba76149f | 1937 | pte_t *_pte; |
7d2eba05 | 1938 | int none_or_zero = 0, result = 0; |
10359213 | 1939 | bool referenced = false, writable = false; |
7d2eba05 | 1940 | |
ba76149f AA |
1941 | for (_pte = pte; _pte < pte+HPAGE_PMD_NR; |
1942 | _pte++, address += PAGE_SIZE) { | |
1943 | pte_t pteval = *_pte; | |
47aee4d8 MK |
1944 | if (pte_none(pteval) || (pte_present(pteval) && |
1945 | is_zero_pfn(pte_pfn(pteval)))) { | |
c1294d05 | 1946 | if (!userfaultfd_armed(vma) && |
7d2eba05 | 1947 | ++none_or_zero <= khugepaged_max_ptes_none) { |
ba76149f | 1948 | continue; |
7d2eba05 EA |
1949 | } else { |
1950 | result = SCAN_EXCEED_NONE_PTE; | |
ba76149f | 1951 | goto out; |
7d2eba05 | 1952 | } |
ba76149f | 1953 | } |
7d2eba05 EA |
1954 | if (!pte_present(pteval)) { |
1955 | result = SCAN_PTE_NON_PRESENT; | |
ba76149f | 1956 | goto out; |
7d2eba05 | 1957 | } |
ba76149f | 1958 | page = vm_normal_page(vma, address, pteval); |
7d2eba05 EA |
1959 | if (unlikely(!page)) { |
1960 | result = SCAN_PAGE_NULL; | |
ba76149f | 1961 | goto out; |
7d2eba05 | 1962 | } |
344aa35c | 1963 | |
309381fe SL |
1964 | VM_BUG_ON_PAGE(PageCompound(page), page); |
1965 | VM_BUG_ON_PAGE(!PageAnon(page), page); | |
1966 | VM_BUG_ON_PAGE(!PageSwapBacked(page), page); | |
ba76149f | 1967 | |
ba76149f AA |
1968 | /* |
1969 | * We can do it before isolate_lru_page because the | |
1970 | * page can't be freed from under us. NOTE: PG_lock | |
1971 | * is needed to serialize against split_huge_page | |
1972 | * when invoked from the VM. | |
1973 | */ | |
7d2eba05 EA |
1974 | if (!trylock_page(page)) { |
1975 | result = SCAN_PAGE_LOCK; | |
ba76149f | 1976 | goto out; |
7d2eba05 | 1977 | } |
10359213 EA |
1978 | |
1979 | /* | |
1980 | * cannot use mapcount: can't collapse if there's a gup pin. | |
1981 | * The page must only be referenced by the scanned process | |
1982 | * and page swap cache. | |
1983 | */ | |
1984 | if (page_count(page) != 1 + !!PageSwapCache(page)) { | |
1985 | unlock_page(page); | |
7d2eba05 | 1986 | result = SCAN_PAGE_COUNT; |
10359213 EA |
1987 | goto out; |
1988 | } | |
1989 | if (pte_write(pteval)) { | |
1990 | writable = true; | |
1991 | } else { | |
1992 | if (PageSwapCache(page) && !reuse_swap_page(page)) { | |
1993 | unlock_page(page); | |
7d2eba05 | 1994 | result = SCAN_SWAP_CACHE_PAGE; |
10359213 EA |
1995 | goto out; |
1996 | } | |
1997 | /* | |
1998 | * Page is not in the swap cache. It can be collapsed | |
1999 | * into a THP. | |
2000 | */ | |
2001 | } | |
2002 | ||
ba76149f AA |
2003 | /* |
2004 | * Isolate the page to avoid collapsing an hugepage | |
2005 | * currently in use by the VM. | |
2006 | */ | |
2007 | if (isolate_lru_page(page)) { | |
2008 | unlock_page(page); | |
7d2eba05 | 2009 | result = SCAN_DEL_PAGE_LRU; |
ba76149f AA |
2010 | goto out; |
2011 | } | |
2012 | /* 0 stands for page_is_file_cache(page) == false */ | |
2013 | inc_zone_page_state(page, NR_ISOLATED_ANON + 0); | |
309381fe SL |
2014 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
2015 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
ba76149f AA |
2016 | |
2017 | /* If there is no mapped pte young don't collapse the page */ | |
33c3fc71 VD |
2018 | if (pte_young(pteval) || |
2019 | page_is_young(page) || PageReferenced(page) || | |
8ee53820 | 2020 | mmu_notifier_test_young(vma->vm_mm, address)) |
10359213 | 2021 | referenced = true; |
ba76149f | 2022 | } |
7d2eba05 EA |
2023 | if (likely(writable)) { |
2024 | if (likely(referenced)) { | |
2025 | result = SCAN_SUCCEED; | |
2026 | trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero, | |
2027 | referenced, writable, result); | |
2028 | return 1; | |
2029 | } | |
2030 | } else { | |
2031 | result = SCAN_PAGE_RO; | |
2032 | } | |
2033 | ||
ba76149f | 2034 | out: |
344aa35c | 2035 | release_pte_pages(pte, _pte); |
7d2eba05 EA |
2036 | trace_mm_collapse_huge_page_isolate(page_to_pfn(page), none_or_zero, |
2037 | referenced, writable, result); | |
344aa35c | 2038 | return 0; |
ba76149f AA |
2039 | } |
2040 | ||
2041 | static void __collapse_huge_page_copy(pte_t *pte, struct page *page, | |
2042 | struct vm_area_struct *vma, | |
2043 | unsigned long address, | |
2044 | spinlock_t *ptl) | |
2045 | { | |
2046 | pte_t *_pte; | |
2047 | for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) { | |
2048 | pte_t pteval = *_pte; | |
2049 | struct page *src_page; | |
2050 | ||
ca0984ca | 2051 | if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { |
ba76149f AA |
2052 | clear_user_highpage(page, address); |
2053 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); | |
ca0984ca EA |
2054 | if (is_zero_pfn(pte_pfn(pteval))) { |
2055 | /* | |
2056 | * ptl mostly unnecessary. | |
2057 | */ | |
2058 | spin_lock(ptl); | |
2059 | /* | |
2060 | * paravirt calls inside pte_clear here are | |
2061 | * superfluous. | |
2062 | */ | |
2063 | pte_clear(vma->vm_mm, address, _pte); | |
2064 | spin_unlock(ptl); | |
2065 | } | |
ba76149f AA |
2066 | } else { |
2067 | src_page = pte_page(pteval); | |
2068 | copy_user_highpage(page, src_page, address, vma); | |
309381fe | 2069 | VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page); |
ba76149f AA |
2070 | release_pte_page(src_page); |
2071 | /* | |
2072 | * ptl mostly unnecessary, but preempt has to | |
2073 | * be disabled to update the per-cpu stats | |
2074 | * inside page_remove_rmap(). | |
2075 | */ | |
2076 | spin_lock(ptl); | |
2077 | /* | |
2078 | * paravirt calls inside pte_clear here are | |
2079 | * superfluous. | |
2080 | */ | |
2081 | pte_clear(vma->vm_mm, address, _pte); | |
d281ee61 | 2082 | page_remove_rmap(src_page, false); |
ba76149f AA |
2083 | spin_unlock(ptl); |
2084 | free_page_and_swap_cache(src_page); | |
2085 | } | |
2086 | ||
2087 | address += PAGE_SIZE; | |
2088 | page++; | |
2089 | } | |
2090 | } | |
2091 | ||
26234f36 | 2092 | static void khugepaged_alloc_sleep(void) |
ba76149f | 2093 | { |
bde43c6c PM |
2094 | DEFINE_WAIT(wait); |
2095 | ||
2096 | add_wait_queue(&khugepaged_wait, &wait); | |
2097 | freezable_schedule_timeout_interruptible( | |
2098 | msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); | |
2099 | remove_wait_queue(&khugepaged_wait, &wait); | |
26234f36 | 2100 | } |
ba76149f | 2101 | |
9f1b868a BL |
2102 | static int khugepaged_node_load[MAX_NUMNODES]; |
2103 | ||
14a4e214 DR |
2104 | static bool khugepaged_scan_abort(int nid) |
2105 | { | |
2106 | int i; | |
2107 | ||
2108 | /* | |
2109 | * If zone_reclaim_mode is disabled, then no extra effort is made to | |
2110 | * allocate memory locally. | |
2111 | */ | |
2112 | if (!zone_reclaim_mode) | |
2113 | return false; | |
2114 | ||
2115 | /* If there is a count for this node already, it must be acceptable */ | |
2116 | if (khugepaged_node_load[nid]) | |
2117 | return false; | |
2118 | ||
2119 | for (i = 0; i < MAX_NUMNODES; i++) { | |
2120 | if (!khugepaged_node_load[i]) | |
2121 | continue; | |
2122 | if (node_distance(nid, i) > RECLAIM_DISTANCE) | |
2123 | return true; | |
2124 | } | |
2125 | return false; | |
2126 | } | |
2127 | ||
26234f36 | 2128 | #ifdef CONFIG_NUMA |
9f1b868a BL |
2129 | static int khugepaged_find_target_node(void) |
2130 | { | |
2131 | static int last_khugepaged_target_node = NUMA_NO_NODE; | |
2132 | int nid, target_node = 0, max_value = 0; | |
2133 | ||
2134 | /* find first node with max normal pages hit */ | |
2135 | for (nid = 0; nid < MAX_NUMNODES; nid++) | |
2136 | if (khugepaged_node_load[nid] > max_value) { | |
2137 | max_value = khugepaged_node_load[nid]; | |
2138 | target_node = nid; | |
2139 | } | |
2140 | ||
2141 | /* do some balance if several nodes have the same hit record */ | |
2142 | if (target_node <= last_khugepaged_target_node) | |
2143 | for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES; | |
2144 | nid++) | |
2145 | if (max_value == khugepaged_node_load[nid]) { | |
2146 | target_node = nid; | |
2147 | break; | |
2148 | } | |
2149 | ||
2150 | last_khugepaged_target_node = target_node; | |
2151 | return target_node; | |
2152 | } | |
2153 | ||
26234f36 XG |
2154 | static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) |
2155 | { | |
2156 | if (IS_ERR(*hpage)) { | |
2157 | if (!*wait) | |
2158 | return false; | |
2159 | ||
2160 | *wait = false; | |
e3b4126c | 2161 | *hpage = NULL; |
26234f36 XG |
2162 | khugepaged_alloc_sleep(); |
2163 | } else if (*hpage) { | |
2164 | put_page(*hpage); | |
2165 | *hpage = NULL; | |
2166 | } | |
2167 | ||
2168 | return true; | |
2169 | } | |
2170 | ||
3b363692 MH |
2171 | static struct page * |
2172 | khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm, | |
d6669d68 | 2173 | unsigned long address, int node) |
26234f36 | 2174 | { |
309381fe | 2175 | VM_BUG_ON_PAGE(*hpage, *hpage); |
8b164568 | 2176 | |
ce83d217 | 2177 | /* |
8b164568 VB |
2178 | * Before allocating the hugepage, release the mmap_sem read lock. |
2179 | * The allocation can take potentially a long time if it involves | |
2180 | * sync compaction, and we do not need to hold the mmap_sem during | |
2181 | * that. We will recheck the vma after taking it again in write mode. | |
ce83d217 | 2182 | */ |
8b164568 VB |
2183 | up_read(&mm->mmap_sem); |
2184 | ||
96db800f | 2185 | *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER); |
26234f36 | 2186 | if (unlikely(!*hpage)) { |
81ab4201 | 2187 | count_vm_event(THP_COLLAPSE_ALLOC_FAILED); |
ce83d217 | 2188 | *hpage = ERR_PTR(-ENOMEM); |
26234f36 | 2189 | return NULL; |
ce83d217 | 2190 | } |
26234f36 | 2191 | |
9a982250 | 2192 | prep_transhuge_page(*hpage); |
65b3c07b | 2193 | count_vm_event(THP_COLLAPSE_ALLOC); |
26234f36 XG |
2194 | return *hpage; |
2195 | } | |
2196 | #else | |
9f1b868a BL |
2197 | static int khugepaged_find_target_node(void) |
2198 | { | |
2199 | return 0; | |
2200 | } | |
2201 | ||
10dc4155 BL |
2202 | static inline struct page *alloc_hugepage(int defrag) |
2203 | { | |
9a982250 KS |
2204 | struct page *page; |
2205 | ||
2206 | page = alloc_pages(alloc_hugepage_gfpmask(defrag, 0), HPAGE_PMD_ORDER); | |
2207 | if (page) | |
2208 | prep_transhuge_page(page); | |
2209 | return page; | |
10dc4155 BL |
2210 | } |
2211 | ||
26234f36 XG |
2212 | static struct page *khugepaged_alloc_hugepage(bool *wait) |
2213 | { | |
2214 | struct page *hpage; | |
2215 | ||
2216 | do { | |
2217 | hpage = alloc_hugepage(khugepaged_defrag()); | |
2218 | if (!hpage) { | |
2219 | count_vm_event(THP_COLLAPSE_ALLOC_FAILED); | |
2220 | if (!*wait) | |
2221 | return NULL; | |
2222 | ||
2223 | *wait = false; | |
2224 | khugepaged_alloc_sleep(); | |
2225 | } else | |
2226 | count_vm_event(THP_COLLAPSE_ALLOC); | |
2227 | } while (unlikely(!hpage) && likely(khugepaged_enabled())); | |
2228 | ||
2229 | return hpage; | |
2230 | } | |
2231 | ||
2232 | static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) | |
2233 | { | |
2234 | if (!*hpage) | |
2235 | *hpage = khugepaged_alloc_hugepage(wait); | |
2236 | ||
2237 | if (unlikely(!*hpage)) | |
2238 | return false; | |
2239 | ||
2240 | return true; | |
2241 | } | |
2242 | ||
3b363692 MH |
2243 | static struct page * |
2244 | khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm, | |
d6669d68 | 2245 | unsigned long address, int node) |
26234f36 XG |
2246 | { |
2247 | up_read(&mm->mmap_sem); | |
2248 | VM_BUG_ON(!*hpage); | |
3b363692 | 2249 | |
26234f36 XG |
2250 | return *hpage; |
2251 | } | |
692e0b35 AA |
2252 | #endif |
2253 | ||
fa475e51 BL |
2254 | static bool hugepage_vma_check(struct vm_area_struct *vma) |
2255 | { | |
2256 | if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) || | |
2257 | (vma->vm_flags & VM_NOHUGEPAGE)) | |
2258 | return false; | |
fa475e51 BL |
2259 | if (!vma->anon_vma || vma->vm_ops) |
2260 | return false; | |
2261 | if (is_vma_temporary_stack(vma)) | |
2262 | return false; | |
81d1b09c | 2263 | VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma); |
fa475e51 BL |
2264 | return true; |
2265 | } | |
2266 | ||
26234f36 XG |
2267 | static void collapse_huge_page(struct mm_struct *mm, |
2268 | unsigned long address, | |
2269 | struct page **hpage, | |
2270 | struct vm_area_struct *vma, | |
2271 | int node) | |
2272 | { | |
26234f36 XG |
2273 | pmd_t *pmd, _pmd; |
2274 | pte_t *pte; | |
2275 | pgtable_t pgtable; | |
2276 | struct page *new_page; | |
c4088ebd | 2277 | spinlock_t *pmd_ptl, *pte_ptl; |
7d2eba05 | 2278 | int isolated, result = 0; |
26234f36 | 2279 | unsigned long hstart, hend; |
00501b53 | 2280 | struct mem_cgroup *memcg; |
2ec74c3e SG |
2281 | unsigned long mmun_start; /* For mmu_notifiers */ |
2282 | unsigned long mmun_end; /* For mmu_notifiers */ | |
3b363692 | 2283 | gfp_t gfp; |
26234f36 XG |
2284 | |
2285 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
2286 | ||
3b363692 MH |
2287 | /* Only allocate from the target node */ |
2288 | gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) | | |
2289 | __GFP_THISNODE; | |
2290 | ||
26234f36 | 2291 | /* release the mmap_sem read lock. */ |
d6669d68 | 2292 | new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node); |
7d2eba05 EA |
2293 | if (!new_page) { |
2294 | result = SCAN_ALLOC_HUGE_PAGE_FAIL; | |
2295 | goto out_nolock; | |
2296 | } | |
26234f36 | 2297 | |
f627c2f5 | 2298 | if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) { |
7d2eba05 EA |
2299 | result = SCAN_CGROUP_CHARGE_FAIL; |
2300 | goto out_nolock; | |
2301 | } | |
ba76149f AA |
2302 | |
2303 | /* | |
2304 | * Prevent all access to pagetables with the exception of | |
2305 | * gup_fast later hanlded by the ptep_clear_flush and the VM | |
2306 | * handled by the anon_vma lock + PG_lock. | |
2307 | */ | |
2308 | down_write(&mm->mmap_sem); | |
7d2eba05 EA |
2309 | if (unlikely(khugepaged_test_exit(mm))) { |
2310 | result = SCAN_ANY_PROCESS; | |
ba76149f | 2311 | goto out; |
7d2eba05 | 2312 | } |
ba76149f AA |
2313 | |
2314 | vma = find_vma(mm, address); | |
7d2eba05 EA |
2315 | if (!vma) { |
2316 | result = SCAN_VMA_NULL; | |
a8f531eb | 2317 | goto out; |
7d2eba05 | 2318 | } |
ba76149f AA |
2319 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; |
2320 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
7d2eba05 EA |
2321 | if (address < hstart || address + HPAGE_PMD_SIZE > hend) { |
2322 | result = SCAN_ADDRESS_RANGE; | |
ba76149f | 2323 | goto out; |
7d2eba05 EA |
2324 | } |
2325 | if (!hugepage_vma_check(vma)) { | |
2326 | result = SCAN_VMA_CHECK; | |
a7d6e4ec | 2327 | goto out; |
7d2eba05 | 2328 | } |
6219049a | 2329 | pmd = mm_find_pmd(mm, address); |
7d2eba05 EA |
2330 | if (!pmd) { |
2331 | result = SCAN_PMD_NULL; | |
ba76149f | 2332 | goto out; |
7d2eba05 | 2333 | } |
ba76149f | 2334 | |
4fc3f1d6 | 2335 | anon_vma_lock_write(vma->anon_vma); |
ba76149f AA |
2336 | |
2337 | pte = pte_offset_map(pmd, address); | |
c4088ebd | 2338 | pte_ptl = pte_lockptr(mm, pmd); |
ba76149f | 2339 | |
2ec74c3e SG |
2340 | mmun_start = address; |
2341 | mmun_end = address + HPAGE_PMD_SIZE; | |
2342 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
c4088ebd | 2343 | pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ |
ba76149f AA |
2344 | /* |
2345 | * After this gup_fast can't run anymore. This also removes | |
2346 | * any huge TLB entry from the CPU so we won't allow | |
2347 | * huge and small TLB entries for the same virtual address | |
2348 | * to avoid the risk of CPU bugs in that area. | |
2349 | */ | |
15a25b2e | 2350 | _pmd = pmdp_collapse_flush(vma, address, pmd); |
c4088ebd | 2351 | spin_unlock(pmd_ptl); |
2ec74c3e | 2352 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
ba76149f | 2353 | |
c4088ebd | 2354 | spin_lock(pte_ptl); |
ba76149f | 2355 | isolated = __collapse_huge_page_isolate(vma, address, pte); |
c4088ebd | 2356 | spin_unlock(pte_ptl); |
ba76149f AA |
2357 | |
2358 | if (unlikely(!isolated)) { | |
453c7192 | 2359 | pte_unmap(pte); |
c4088ebd | 2360 | spin_lock(pmd_ptl); |
ba76149f | 2361 | BUG_ON(!pmd_none(*pmd)); |
7c342512 AK |
2362 | /* |
2363 | * We can only use set_pmd_at when establishing | |
2364 | * hugepmds and never for establishing regular pmds that | |
2365 | * points to regular pagetables. Use pmd_populate for that | |
2366 | */ | |
2367 | pmd_populate(mm, pmd, pmd_pgtable(_pmd)); | |
c4088ebd | 2368 | spin_unlock(pmd_ptl); |
08b52706 | 2369 | anon_vma_unlock_write(vma->anon_vma); |
7d2eba05 | 2370 | result = SCAN_FAIL; |
ce83d217 | 2371 | goto out; |
ba76149f AA |
2372 | } |
2373 | ||
2374 | /* | |
2375 | * All pages are isolated and locked so anon_vma rmap | |
2376 | * can't run anymore. | |
2377 | */ | |
08b52706 | 2378 | anon_vma_unlock_write(vma->anon_vma); |
ba76149f | 2379 | |
c4088ebd | 2380 | __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl); |
453c7192 | 2381 | pte_unmap(pte); |
ba76149f AA |
2382 | __SetPageUptodate(new_page); |
2383 | pgtable = pmd_pgtable(_pmd); | |
ba76149f | 2384 | |
3122359a KS |
2385 | _pmd = mk_huge_pmd(new_page, vma->vm_page_prot); |
2386 | _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); | |
ba76149f AA |
2387 | |
2388 | /* | |
2389 | * spin_lock() below is not the equivalent of smp_wmb(), so | |
2390 | * this is needed to avoid the copy_huge_page writes to become | |
2391 | * visible after the set_pmd_at() write. | |
2392 | */ | |
2393 | smp_wmb(); | |
2394 | ||
c4088ebd | 2395 | spin_lock(pmd_ptl); |
ba76149f | 2396 | BUG_ON(!pmd_none(*pmd)); |
d281ee61 | 2397 | page_add_new_anon_rmap(new_page, vma, address, true); |
f627c2f5 | 2398 | mem_cgroup_commit_charge(new_page, memcg, false, true); |
00501b53 | 2399 | lru_cache_add_active_or_unevictable(new_page, vma); |
fce144b4 | 2400 | pgtable_trans_huge_deposit(mm, pmd, pgtable); |
ba76149f | 2401 | set_pmd_at(mm, address, pmd, _pmd); |
b113da65 | 2402 | update_mmu_cache_pmd(vma, address, pmd); |
c4088ebd | 2403 | spin_unlock(pmd_ptl); |
ba76149f AA |
2404 | |
2405 | *hpage = NULL; | |
420256ef | 2406 | |
ba76149f | 2407 | khugepaged_pages_collapsed++; |
7d2eba05 | 2408 | result = SCAN_SUCCEED; |
ce83d217 | 2409 | out_up_write: |
ba76149f | 2410 | up_write(&mm->mmap_sem); |
7d2eba05 | 2411 | trace_mm_collapse_huge_page(mm, isolated, result); |
0bbbc0b3 AA |
2412 | return; |
2413 | ||
7d2eba05 EA |
2414 | out_nolock: |
2415 | trace_mm_collapse_huge_page(mm, isolated, result); | |
2416 | return; | |
ce83d217 | 2417 | out: |
f627c2f5 | 2418 | mem_cgroup_cancel_charge(new_page, memcg, true); |
ce83d217 | 2419 | goto out_up_write; |
ba76149f AA |
2420 | } |
2421 | ||
2422 | static int khugepaged_scan_pmd(struct mm_struct *mm, | |
2423 | struct vm_area_struct *vma, | |
2424 | unsigned long address, | |
2425 | struct page **hpage) | |
2426 | { | |
ba76149f AA |
2427 | pmd_t *pmd; |
2428 | pte_t *pte, *_pte; | |
7d2eba05 EA |
2429 | int ret = 0, none_or_zero = 0, result = 0; |
2430 | struct page *page = NULL; | |
ba76149f AA |
2431 | unsigned long _address; |
2432 | spinlock_t *ptl; | |
00ef2d2f | 2433 | int node = NUMA_NO_NODE; |
10359213 | 2434 | bool writable = false, referenced = false; |
ba76149f AA |
2435 | |
2436 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
2437 | ||
6219049a | 2438 | pmd = mm_find_pmd(mm, address); |
7d2eba05 EA |
2439 | if (!pmd) { |
2440 | result = SCAN_PMD_NULL; | |
ba76149f | 2441 | goto out; |
7d2eba05 | 2442 | } |
ba76149f | 2443 | |
9f1b868a | 2444 | memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load)); |
ba76149f AA |
2445 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); |
2446 | for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR; | |
2447 | _pte++, _address += PAGE_SIZE) { | |
2448 | pte_t pteval = *_pte; | |
ca0984ca | 2449 | if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { |
c1294d05 | 2450 | if (!userfaultfd_armed(vma) && |
7d2eba05 | 2451 | ++none_or_zero <= khugepaged_max_ptes_none) { |
ba76149f | 2452 | continue; |
7d2eba05 EA |
2453 | } else { |
2454 | result = SCAN_EXCEED_NONE_PTE; | |
ba76149f | 2455 | goto out_unmap; |
7d2eba05 | 2456 | } |
ba76149f | 2457 | } |
7d2eba05 EA |
2458 | if (!pte_present(pteval)) { |
2459 | result = SCAN_PTE_NON_PRESENT; | |
ba76149f | 2460 | goto out_unmap; |
7d2eba05 | 2461 | } |
10359213 EA |
2462 | if (pte_write(pteval)) |
2463 | writable = true; | |
2464 | ||
ba76149f | 2465 | page = vm_normal_page(vma, _address, pteval); |
7d2eba05 EA |
2466 | if (unlikely(!page)) { |
2467 | result = SCAN_PAGE_NULL; | |
ba76149f | 2468 | goto out_unmap; |
7d2eba05 | 2469 | } |
b1caa957 KS |
2470 | |
2471 | /* TODO: teach khugepaged to collapse THP mapped with pte */ | |
2472 | if (PageCompound(page)) { | |
2473 | result = SCAN_PAGE_COMPOUND; | |
2474 | goto out_unmap; | |
2475 | } | |
2476 | ||
5c4b4be3 | 2477 | /* |
9f1b868a BL |
2478 | * Record which node the original page is from and save this |
2479 | * information to khugepaged_node_load[]. | |
2480 | * Khupaged will allocate hugepage from the node has the max | |
2481 | * hit record. | |
5c4b4be3 | 2482 | */ |
9f1b868a | 2483 | node = page_to_nid(page); |
7d2eba05 EA |
2484 | if (khugepaged_scan_abort(node)) { |
2485 | result = SCAN_SCAN_ABORT; | |
14a4e214 | 2486 | goto out_unmap; |
7d2eba05 | 2487 | } |
9f1b868a | 2488 | khugepaged_node_load[node]++; |
7d2eba05 EA |
2489 | if (!PageLRU(page)) { |
2490 | result = SCAN_SCAN_ABORT; | |
2491 | goto out_unmap; | |
2492 | } | |
2493 | if (PageLocked(page)) { | |
2494 | result = SCAN_PAGE_LOCK; | |
ba76149f | 2495 | goto out_unmap; |
7d2eba05 EA |
2496 | } |
2497 | if (!PageAnon(page)) { | |
2498 | result = SCAN_PAGE_ANON; | |
2499 | goto out_unmap; | |
2500 | } | |
2501 | ||
10359213 EA |
2502 | /* |
2503 | * cannot use mapcount: can't collapse if there's a gup pin. | |
2504 | * The page must only be referenced by the scanned process | |
2505 | * and page swap cache. | |
2506 | */ | |
7d2eba05 EA |
2507 | if (page_count(page) != 1 + !!PageSwapCache(page)) { |
2508 | result = SCAN_PAGE_COUNT; | |
ba76149f | 2509 | goto out_unmap; |
7d2eba05 | 2510 | } |
33c3fc71 VD |
2511 | if (pte_young(pteval) || |
2512 | page_is_young(page) || PageReferenced(page) || | |
8ee53820 | 2513 | mmu_notifier_test_young(vma->vm_mm, address)) |
10359213 | 2514 | referenced = true; |
ba76149f | 2515 | } |
7d2eba05 EA |
2516 | if (writable) { |
2517 | if (referenced) { | |
2518 | result = SCAN_SUCCEED; | |
2519 | ret = 1; | |
2520 | } else { | |
2521 | result = SCAN_NO_REFERENCED_PAGE; | |
2522 | } | |
2523 | } else { | |
2524 | result = SCAN_PAGE_RO; | |
2525 | } | |
ba76149f AA |
2526 | out_unmap: |
2527 | pte_unmap_unlock(pte, ptl); | |
9f1b868a BL |
2528 | if (ret) { |
2529 | node = khugepaged_find_target_node(); | |
ce83d217 | 2530 | /* collapse_huge_page will return with the mmap_sem released */ |
5c4b4be3 | 2531 | collapse_huge_page(mm, address, hpage, vma, node); |
9f1b868a | 2532 | } |
ba76149f | 2533 | out: |
7d2eba05 EA |
2534 | trace_mm_khugepaged_scan_pmd(mm, page_to_pfn(page), writable, referenced, |
2535 | none_or_zero, result); | |
ba76149f AA |
2536 | return ret; |
2537 | } | |
2538 | ||
2539 | static void collect_mm_slot(struct mm_slot *mm_slot) | |
2540 | { | |
2541 | struct mm_struct *mm = mm_slot->mm; | |
2542 | ||
b9980cdc | 2543 | VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); |
ba76149f AA |
2544 | |
2545 | if (khugepaged_test_exit(mm)) { | |
2546 | /* free mm_slot */ | |
43b5fbbd | 2547 | hash_del(&mm_slot->hash); |
ba76149f AA |
2548 | list_del(&mm_slot->mm_node); |
2549 | ||
2550 | /* | |
2551 | * Not strictly needed because the mm exited already. | |
2552 | * | |
2553 | * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); | |
2554 | */ | |
2555 | ||
2556 | /* khugepaged_mm_lock actually not necessary for the below */ | |
2557 | free_mm_slot(mm_slot); | |
2558 | mmdrop(mm); | |
2559 | } | |
2560 | } | |
2561 | ||
2562 | static unsigned int khugepaged_scan_mm_slot(unsigned int pages, | |
2563 | struct page **hpage) | |
2f1da642 HS |
2564 | __releases(&khugepaged_mm_lock) |
2565 | __acquires(&khugepaged_mm_lock) | |
ba76149f AA |
2566 | { |
2567 | struct mm_slot *mm_slot; | |
2568 | struct mm_struct *mm; | |
2569 | struct vm_area_struct *vma; | |
2570 | int progress = 0; | |
2571 | ||
2572 | VM_BUG_ON(!pages); | |
b9980cdc | 2573 | VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); |
ba76149f AA |
2574 | |
2575 | if (khugepaged_scan.mm_slot) | |
2576 | mm_slot = khugepaged_scan.mm_slot; | |
2577 | else { | |
2578 | mm_slot = list_entry(khugepaged_scan.mm_head.next, | |
2579 | struct mm_slot, mm_node); | |
2580 | khugepaged_scan.address = 0; | |
2581 | khugepaged_scan.mm_slot = mm_slot; | |
2582 | } | |
2583 | spin_unlock(&khugepaged_mm_lock); | |
2584 | ||
2585 | mm = mm_slot->mm; | |
2586 | down_read(&mm->mmap_sem); | |
2587 | if (unlikely(khugepaged_test_exit(mm))) | |
2588 | vma = NULL; | |
2589 | else | |
2590 | vma = find_vma(mm, khugepaged_scan.address); | |
2591 | ||
2592 | progress++; | |
2593 | for (; vma; vma = vma->vm_next) { | |
2594 | unsigned long hstart, hend; | |
2595 | ||
2596 | cond_resched(); | |
2597 | if (unlikely(khugepaged_test_exit(mm))) { | |
2598 | progress++; | |
2599 | break; | |
2600 | } | |
fa475e51 BL |
2601 | if (!hugepage_vma_check(vma)) { |
2602 | skip: | |
ba76149f AA |
2603 | progress++; |
2604 | continue; | |
2605 | } | |
ba76149f AA |
2606 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; |
2607 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
a7d6e4ec AA |
2608 | if (hstart >= hend) |
2609 | goto skip; | |
2610 | if (khugepaged_scan.address > hend) | |
2611 | goto skip; | |
ba76149f AA |
2612 | if (khugepaged_scan.address < hstart) |
2613 | khugepaged_scan.address = hstart; | |
a7d6e4ec | 2614 | VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); |
ba76149f AA |
2615 | |
2616 | while (khugepaged_scan.address < hend) { | |
2617 | int ret; | |
2618 | cond_resched(); | |
2619 | if (unlikely(khugepaged_test_exit(mm))) | |
2620 | goto breakouterloop; | |
2621 | ||
2622 | VM_BUG_ON(khugepaged_scan.address < hstart || | |
2623 | khugepaged_scan.address + HPAGE_PMD_SIZE > | |
2624 | hend); | |
2625 | ret = khugepaged_scan_pmd(mm, vma, | |
2626 | khugepaged_scan.address, | |
2627 | hpage); | |
2628 | /* move to next address */ | |
2629 | khugepaged_scan.address += HPAGE_PMD_SIZE; | |
2630 | progress += HPAGE_PMD_NR; | |
2631 | if (ret) | |
2632 | /* we released mmap_sem so break loop */ | |
2633 | goto breakouterloop_mmap_sem; | |
2634 | if (progress >= pages) | |
2635 | goto breakouterloop; | |
2636 | } | |
2637 | } | |
2638 | breakouterloop: | |
2639 | up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */ | |
2640 | breakouterloop_mmap_sem: | |
2641 | ||
2642 | spin_lock(&khugepaged_mm_lock); | |
a7d6e4ec | 2643 | VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); |
ba76149f AA |
2644 | /* |
2645 | * Release the current mm_slot if this mm is about to die, or | |
2646 | * if we scanned all vmas of this mm. | |
2647 | */ | |
2648 | if (khugepaged_test_exit(mm) || !vma) { | |
2649 | /* | |
2650 | * Make sure that if mm_users is reaching zero while | |
2651 | * khugepaged runs here, khugepaged_exit will find | |
2652 | * mm_slot not pointing to the exiting mm. | |
2653 | */ | |
2654 | if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) { | |
2655 | khugepaged_scan.mm_slot = list_entry( | |
2656 | mm_slot->mm_node.next, | |
2657 | struct mm_slot, mm_node); | |
2658 | khugepaged_scan.address = 0; | |
2659 | } else { | |
2660 | khugepaged_scan.mm_slot = NULL; | |
2661 | khugepaged_full_scans++; | |
2662 | } | |
2663 | ||
2664 | collect_mm_slot(mm_slot); | |
2665 | } | |
2666 | ||
2667 | return progress; | |
2668 | } | |
2669 | ||
2670 | static int khugepaged_has_work(void) | |
2671 | { | |
2672 | return !list_empty(&khugepaged_scan.mm_head) && | |
2673 | khugepaged_enabled(); | |
2674 | } | |
2675 | ||
2676 | static int khugepaged_wait_event(void) | |
2677 | { | |
2678 | return !list_empty(&khugepaged_scan.mm_head) || | |
2017c0bf | 2679 | kthread_should_stop(); |
ba76149f AA |
2680 | } |
2681 | ||
d516904b | 2682 | static void khugepaged_do_scan(void) |
ba76149f | 2683 | { |
d516904b | 2684 | struct page *hpage = NULL; |
ba76149f AA |
2685 | unsigned int progress = 0, pass_through_head = 0; |
2686 | unsigned int pages = khugepaged_pages_to_scan; | |
d516904b | 2687 | bool wait = true; |
ba76149f AA |
2688 | |
2689 | barrier(); /* write khugepaged_pages_to_scan to local stack */ | |
2690 | ||
2691 | while (progress < pages) { | |
26234f36 | 2692 | if (!khugepaged_prealloc_page(&hpage, &wait)) |
d516904b | 2693 | break; |
26234f36 | 2694 | |
420256ef | 2695 | cond_resched(); |
ba76149f | 2696 | |
cd092411 | 2697 | if (unlikely(kthread_should_stop() || try_to_freeze())) |
878aee7d AA |
2698 | break; |
2699 | ||
ba76149f AA |
2700 | spin_lock(&khugepaged_mm_lock); |
2701 | if (!khugepaged_scan.mm_slot) | |
2702 | pass_through_head++; | |
2703 | if (khugepaged_has_work() && | |
2704 | pass_through_head < 2) | |
2705 | progress += khugepaged_scan_mm_slot(pages - progress, | |
d516904b | 2706 | &hpage); |
ba76149f AA |
2707 | else |
2708 | progress = pages; | |
2709 | spin_unlock(&khugepaged_mm_lock); | |
2710 | } | |
ba76149f | 2711 | |
d516904b XG |
2712 | if (!IS_ERR_OR_NULL(hpage)) |
2713 | put_page(hpage); | |
0bbbc0b3 AA |
2714 | } |
2715 | ||
2017c0bf XG |
2716 | static void khugepaged_wait_work(void) |
2717 | { | |
2017c0bf XG |
2718 | if (khugepaged_has_work()) { |
2719 | if (!khugepaged_scan_sleep_millisecs) | |
2720 | return; | |
2721 | ||
2722 | wait_event_freezable_timeout(khugepaged_wait, | |
2723 | kthread_should_stop(), | |
2724 | msecs_to_jiffies(khugepaged_scan_sleep_millisecs)); | |
2725 | return; | |
2726 | } | |
2727 | ||
2728 | if (khugepaged_enabled()) | |
2729 | wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); | |
2730 | } | |
2731 | ||
ba76149f AA |
2732 | static int khugepaged(void *none) |
2733 | { | |
2734 | struct mm_slot *mm_slot; | |
2735 | ||
878aee7d | 2736 | set_freezable(); |
8698a745 | 2737 | set_user_nice(current, MAX_NICE); |
ba76149f | 2738 | |
b7231789 XG |
2739 | while (!kthread_should_stop()) { |
2740 | khugepaged_do_scan(); | |
2741 | khugepaged_wait_work(); | |
2742 | } | |
ba76149f AA |
2743 | |
2744 | spin_lock(&khugepaged_mm_lock); | |
2745 | mm_slot = khugepaged_scan.mm_slot; | |
2746 | khugepaged_scan.mm_slot = NULL; | |
2747 | if (mm_slot) | |
2748 | collect_mm_slot(mm_slot); | |
2749 | spin_unlock(&khugepaged_mm_lock); | |
ba76149f AA |
2750 | return 0; |
2751 | } | |
2752 | ||
eef1b3ba KS |
2753 | static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, |
2754 | unsigned long haddr, pmd_t *pmd) | |
2755 | { | |
2756 | struct mm_struct *mm = vma->vm_mm; | |
2757 | pgtable_t pgtable; | |
2758 | pmd_t _pmd; | |
2759 | int i; | |
2760 | ||
2761 | /* leave pmd empty until pte is filled */ | |
2762 | pmdp_huge_clear_flush_notify(vma, haddr, pmd); | |
2763 | ||
2764 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); | |
2765 | pmd_populate(mm, &_pmd, pgtable); | |
2766 | ||
2767 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
2768 | pte_t *pte, entry; | |
2769 | entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); | |
2770 | entry = pte_mkspecial(entry); | |
2771 | pte = pte_offset_map(&_pmd, haddr); | |
2772 | VM_BUG_ON(!pte_none(*pte)); | |
2773 | set_pte_at(mm, haddr, pte, entry); | |
2774 | pte_unmap(pte); | |
2775 | } | |
2776 | smp_wmb(); /* make pte visible before pmd */ | |
2777 | pmd_populate(mm, pmd, pgtable); | |
2778 | put_huge_zero_page(); | |
2779 | } | |
2780 | ||
2781 | static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, | |
ba988280 | 2782 | unsigned long haddr, bool freeze) |
eef1b3ba KS |
2783 | { |
2784 | struct mm_struct *mm = vma->vm_mm; | |
2785 | struct page *page; | |
2786 | pgtable_t pgtable; | |
2787 | pmd_t _pmd; | |
b8d3c4c3 | 2788 | bool young, write, dirty; |
eef1b3ba KS |
2789 | int i; |
2790 | ||
2791 | VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); | |
2792 | VM_BUG_ON_VMA(vma->vm_start > haddr, vma); | |
2793 | VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); | |
5c7fb56e | 2794 | VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)); |
eef1b3ba KS |
2795 | |
2796 | count_vm_event(THP_SPLIT_PMD); | |
2797 | ||
2798 | if (vma_is_dax(vma)) { | |
2799 | pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); | |
2800 | if (is_huge_zero_pmd(_pmd)) | |
2801 | put_huge_zero_page(); | |
2802 | return; | |
2803 | } else if (is_huge_zero_pmd(*pmd)) { | |
2804 | return __split_huge_zero_page_pmd(vma, haddr, pmd); | |
2805 | } | |
2806 | ||
2807 | page = pmd_page(*pmd); | |
2808 | VM_BUG_ON_PAGE(!page_count(page), page); | |
2809 | atomic_add(HPAGE_PMD_NR - 1, &page->_count); | |
2810 | write = pmd_write(*pmd); | |
2811 | young = pmd_young(*pmd); | |
b8d3c4c3 | 2812 | dirty = pmd_dirty(*pmd); |
eef1b3ba | 2813 | |
eef1b3ba KS |
2814 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); |
2815 | pmd_populate(mm, &_pmd, pgtable); | |
2816 | ||
2817 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
2818 | pte_t entry, *pte; | |
2819 | /* | |
2820 | * Note that NUMA hinting access restrictions are not | |
2821 | * transferred to avoid any possibility of altering | |
2822 | * permissions across VMAs. | |
2823 | */ | |
ba988280 KS |
2824 | if (freeze) { |
2825 | swp_entry_t swp_entry; | |
2826 | swp_entry = make_migration_entry(page + i, write); | |
2827 | entry = swp_entry_to_pte(swp_entry); | |
2828 | } else { | |
2829 | entry = mk_pte(page + i, vma->vm_page_prot); | |
b8d3c4c3 | 2830 | entry = maybe_mkwrite(entry, vma); |
ba988280 KS |
2831 | if (!write) |
2832 | entry = pte_wrprotect(entry); | |
2833 | if (!young) | |
2834 | entry = pte_mkold(entry); | |
2835 | } | |
b8d3c4c3 MK |
2836 | if (dirty) |
2837 | SetPageDirty(page + i); | |
eef1b3ba KS |
2838 | pte = pte_offset_map(&_pmd, haddr); |
2839 | BUG_ON(!pte_none(*pte)); | |
2840 | set_pte_at(mm, haddr, pte, entry); | |
2841 | atomic_inc(&page[i]._mapcount); | |
2842 | pte_unmap(pte); | |
2843 | } | |
2844 | ||
2845 | /* | |
2846 | * Set PG_double_map before dropping compound_mapcount to avoid | |
2847 | * false-negative page_mapped(). | |
2848 | */ | |
2849 | if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) { | |
2850 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
2851 | atomic_inc(&page[i]._mapcount); | |
2852 | } | |
2853 | ||
2854 | if (atomic_add_negative(-1, compound_mapcount_ptr(page))) { | |
2855 | /* Last compound_mapcount is gone. */ | |
2856 | __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES); | |
2857 | if (TestClearPageDoubleMap(page)) { | |
2858 | /* No need in mapcount reference anymore */ | |
2859 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
2860 | atomic_dec(&page[i]._mapcount); | |
2861 | } | |
2862 | } | |
2863 | ||
2864 | smp_wmb(); /* make pte visible before pmd */ | |
e9b61f19 KS |
2865 | /* |
2866 | * Up to this point the pmd is present and huge and userland has the | |
2867 | * whole access to the hugepage during the split (which happens in | |
2868 | * place). If we overwrite the pmd with the not-huge version pointing | |
2869 | * to the pte here (which of course we could if all CPUs were bug | |
2870 | * free), userland could trigger a small page size TLB miss on the | |
2871 | * small sized TLB while the hugepage TLB entry is still established in | |
2872 | * the huge TLB. Some CPU doesn't like that. | |
2873 | * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum | |
2874 | * 383 on page 93. Intel should be safe but is also warns that it's | |
2875 | * only safe if the permission and cache attributes of the two entries | |
2876 | * loaded in the two TLB is identical (which should be the case here). | |
2877 | * But it is generally safer to never allow small and huge TLB entries | |
2878 | * for the same virtual address to be loaded simultaneously. So instead | |
2879 | * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the | |
2880 | * current pmd notpresent (atomically because here the pmd_trans_huge | |
2881 | * and pmd_trans_splitting must remain set at all times on the pmd | |
2882 | * until the split is complete for this pmd), then we flush the SMP TLB | |
2883 | * and finally we write the non-huge version of the pmd entry with | |
2884 | * pmd_populate. | |
2885 | */ | |
2886 | pmdp_invalidate(vma, haddr, pmd); | |
eef1b3ba | 2887 | pmd_populate(mm, pmd, pgtable); |
e9b61f19 KS |
2888 | |
2889 | if (freeze) { | |
2890 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
2891 | page_remove_rmap(page + i, false); | |
2892 | put_page(page + i); | |
2893 | } | |
2894 | } | |
eef1b3ba KS |
2895 | } |
2896 | ||
2897 | void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, | |
2898 | unsigned long address) | |
2899 | { | |
2900 | spinlock_t *ptl; | |
2901 | struct mm_struct *mm = vma->vm_mm; | |
e90309c9 | 2902 | struct page *page = NULL; |
eef1b3ba KS |
2903 | unsigned long haddr = address & HPAGE_PMD_MASK; |
2904 | ||
2905 | mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE); | |
2906 | ptl = pmd_lock(mm, pmd); | |
5c7fb56e DW |
2907 | if (pmd_trans_huge(*pmd)) { |
2908 | page = pmd_page(*pmd); | |
2909 | if (PageMlocked(page)) | |
2910 | get_page(page); | |
2911 | else | |
2912 | page = NULL; | |
2913 | } else if (!pmd_devmap(*pmd)) | |
e90309c9 | 2914 | goto out; |
e90309c9 | 2915 | __split_huge_pmd_locked(vma, pmd, haddr, false); |
e90309c9 | 2916 | out: |
eef1b3ba KS |
2917 | spin_unlock(ptl); |
2918 | mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE); | |
e90309c9 KS |
2919 | if (page) { |
2920 | lock_page(page); | |
2921 | munlock_vma_page(page); | |
2922 | unlock_page(page); | |
2923 | put_page(page); | |
2924 | } | |
eef1b3ba KS |
2925 | } |
2926 | ||
78ddc534 | 2927 | static void split_huge_pmd_address(struct vm_area_struct *vma, |
94fcc585 AA |
2928 | unsigned long address) |
2929 | { | |
f72e7dcd HD |
2930 | pgd_t *pgd; |
2931 | pud_t *pud; | |
94fcc585 AA |
2932 | pmd_t *pmd; |
2933 | ||
2934 | VM_BUG_ON(!(address & ~HPAGE_PMD_MASK)); | |
2935 | ||
78ddc534 | 2936 | pgd = pgd_offset(vma->vm_mm, address); |
f72e7dcd HD |
2937 | if (!pgd_present(*pgd)) |
2938 | return; | |
2939 | ||
2940 | pud = pud_offset(pgd, address); | |
2941 | if (!pud_present(*pud)) | |
2942 | return; | |
2943 | ||
2944 | pmd = pmd_offset(pud, address); | |
5c7fb56e | 2945 | if (!pmd_present(*pmd) || (!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd))) |
94fcc585 AA |
2946 | return; |
2947 | /* | |
2948 | * Caller holds the mmap_sem write mode, so a huge pmd cannot | |
2949 | * materialize from under us. | |
2950 | */ | |
ad0bed24 | 2951 | split_huge_pmd(vma, pmd, address); |
94fcc585 AA |
2952 | } |
2953 | ||
e1b9996b | 2954 | void vma_adjust_trans_huge(struct vm_area_struct *vma, |
94fcc585 AA |
2955 | unsigned long start, |
2956 | unsigned long end, | |
2957 | long adjust_next) | |
2958 | { | |
2959 | /* | |
2960 | * If the new start address isn't hpage aligned and it could | |
2961 | * previously contain an hugepage: check if we need to split | |
2962 | * an huge pmd. | |
2963 | */ | |
2964 | if (start & ~HPAGE_PMD_MASK && | |
2965 | (start & HPAGE_PMD_MASK) >= vma->vm_start && | |
2966 | (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
78ddc534 | 2967 | split_huge_pmd_address(vma, start); |
94fcc585 AA |
2968 | |
2969 | /* | |
2970 | * If the new end address isn't hpage aligned and it could | |
2971 | * previously contain an hugepage: check if we need to split | |
2972 | * an huge pmd. | |
2973 | */ | |
2974 | if (end & ~HPAGE_PMD_MASK && | |
2975 | (end & HPAGE_PMD_MASK) >= vma->vm_start && | |
2976 | (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
78ddc534 | 2977 | split_huge_pmd_address(vma, end); |
94fcc585 AA |
2978 | |
2979 | /* | |
2980 | * If we're also updating the vma->vm_next->vm_start, if the new | |
2981 | * vm_next->vm_start isn't page aligned and it could previously | |
2982 | * contain an hugepage: check if we need to split an huge pmd. | |
2983 | */ | |
2984 | if (adjust_next > 0) { | |
2985 | struct vm_area_struct *next = vma->vm_next; | |
2986 | unsigned long nstart = next->vm_start; | |
2987 | nstart += adjust_next << PAGE_SHIFT; | |
2988 | if (nstart & ~HPAGE_PMD_MASK && | |
2989 | (nstart & HPAGE_PMD_MASK) >= next->vm_start && | |
2990 | (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) | |
78ddc534 | 2991 | split_huge_pmd_address(next, nstart); |
94fcc585 AA |
2992 | } |
2993 | } | |
e9b61f19 KS |
2994 | |
2995 | static void freeze_page_vma(struct vm_area_struct *vma, struct page *page, | |
2996 | unsigned long address) | |
2997 | { | |
bd56086f | 2998 | unsigned long haddr = address & HPAGE_PMD_MASK; |
e9b61f19 KS |
2999 | spinlock_t *ptl; |
3000 | pgd_t *pgd; | |
3001 | pud_t *pud; | |
3002 | pmd_t *pmd; | |
3003 | pte_t *pte; | |
3004 | int i, nr = HPAGE_PMD_NR; | |
3005 | ||
3006 | /* Skip pages which doesn't belong to the VMA */ | |
3007 | if (address < vma->vm_start) { | |
3008 | int off = (vma->vm_start - address) >> PAGE_SHIFT; | |
3009 | page += off; | |
3010 | nr -= off; | |
3011 | address = vma->vm_start; | |
3012 | } | |
3013 | ||
3014 | pgd = pgd_offset(vma->vm_mm, address); | |
3015 | if (!pgd_present(*pgd)) | |
3016 | return; | |
3017 | pud = pud_offset(pgd, address); | |
3018 | if (!pud_present(*pud)) | |
3019 | return; | |
3020 | pmd = pmd_offset(pud, address); | |
3021 | ptl = pmd_lock(vma->vm_mm, pmd); | |
3022 | if (!pmd_present(*pmd)) { | |
3023 | spin_unlock(ptl); | |
3024 | return; | |
3025 | } | |
3026 | if (pmd_trans_huge(*pmd)) { | |
3027 | if (page == pmd_page(*pmd)) | |
bd56086f | 3028 | __split_huge_pmd_locked(vma, pmd, haddr, true); |
e9b61f19 KS |
3029 | spin_unlock(ptl); |
3030 | return; | |
3031 | } | |
3032 | spin_unlock(ptl); | |
3033 | ||
3034 | pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl); | |
bd56086f | 3035 | for (i = 0; i < nr; i++, address += PAGE_SIZE, page++, pte++) { |
e9b61f19 KS |
3036 | pte_t entry, swp_pte; |
3037 | swp_entry_t swp_entry; | |
3038 | ||
bd56086f KS |
3039 | /* |
3040 | * We've just crossed page table boundary: need to map next one. | |
3041 | * It can happen if THP was mremaped to non PMD-aligned address. | |
3042 | */ | |
3043 | if (unlikely(address == haddr + HPAGE_PMD_SIZE)) { | |
3044 | pte_unmap_unlock(pte - 1, ptl); | |
3045 | pmd = mm_find_pmd(vma->vm_mm, address); | |
3046 | if (!pmd) | |
3047 | return; | |
3048 | pte = pte_offset_map_lock(vma->vm_mm, pmd, | |
3049 | address, &ptl); | |
3050 | } | |
3051 | ||
3052 | if (!pte_present(*pte)) | |
e9b61f19 | 3053 | continue; |
bd56086f | 3054 | if (page_to_pfn(page) != pte_pfn(*pte)) |
e9b61f19 KS |
3055 | continue; |
3056 | flush_cache_page(vma, address, page_to_pfn(page)); | |
bd56086f | 3057 | entry = ptep_clear_flush(vma, address, pte); |
b8d3c4c3 MK |
3058 | if (pte_dirty(entry)) |
3059 | SetPageDirty(page); | |
e9b61f19 KS |
3060 | swp_entry = make_migration_entry(page, pte_write(entry)); |
3061 | swp_pte = swp_entry_to_pte(swp_entry); | |
3062 | if (pte_soft_dirty(entry)) | |
3063 | swp_pte = pte_swp_mksoft_dirty(swp_pte); | |
bd56086f | 3064 | set_pte_at(vma->vm_mm, address, pte, swp_pte); |
e9b61f19 KS |
3065 | page_remove_rmap(page, false); |
3066 | put_page(page); | |
3067 | } | |
bd56086f | 3068 | pte_unmap_unlock(pte - 1, ptl); |
e9b61f19 KS |
3069 | } |
3070 | ||
3071 | static void freeze_page(struct anon_vma *anon_vma, struct page *page) | |
3072 | { | |
3073 | struct anon_vma_chain *avc; | |
3074 | pgoff_t pgoff = page_to_pgoff(page); | |
3075 | ||
3076 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
3077 | ||
3078 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, | |
3079 | pgoff + HPAGE_PMD_NR - 1) { | |
bd56086f | 3080 | unsigned long address = __vma_address(page, avc->vma); |
e9b61f19 | 3081 | |
e9b61f19 | 3082 | mmu_notifier_invalidate_range_start(avc->vma->vm_mm, |
bd56086f KS |
3083 | address, address + HPAGE_PMD_SIZE); |
3084 | freeze_page_vma(avc->vma, page, address); | |
e9b61f19 | 3085 | mmu_notifier_invalidate_range_end(avc->vma->vm_mm, |
bd56086f | 3086 | address, address + HPAGE_PMD_SIZE); |
e9b61f19 KS |
3087 | } |
3088 | } | |
3089 | ||
3090 | static void unfreeze_page_vma(struct vm_area_struct *vma, struct page *page, | |
3091 | unsigned long address) | |
3092 | { | |
3093 | spinlock_t *ptl; | |
3094 | pmd_t *pmd; | |
3095 | pte_t *pte, entry; | |
3096 | swp_entry_t swp_entry; | |
bd56086f | 3097 | unsigned long haddr = address & HPAGE_PMD_MASK; |
e9b61f19 KS |
3098 | int i, nr = HPAGE_PMD_NR; |
3099 | ||
3100 | /* Skip pages which doesn't belong to the VMA */ | |
3101 | if (address < vma->vm_start) { | |
3102 | int off = (vma->vm_start - address) >> PAGE_SHIFT; | |
3103 | page += off; | |
3104 | nr -= off; | |
3105 | address = vma->vm_start; | |
3106 | } | |
3107 | ||
3108 | pmd = mm_find_pmd(vma->vm_mm, address); | |
3109 | if (!pmd) | |
3110 | return; | |
bd56086f | 3111 | |
e9b61f19 | 3112 | pte = pte_offset_map_lock(vma->vm_mm, pmd, address, &ptl); |
bd56086f KS |
3113 | for (i = 0; i < nr; i++, address += PAGE_SIZE, page++, pte++) { |
3114 | /* | |
3115 | * We've just crossed page table boundary: need to map next one. | |
3116 | * It can happen if THP was mremaped to non-PMD aligned address. | |
3117 | */ | |
3118 | if (unlikely(address == haddr + HPAGE_PMD_SIZE)) { | |
3119 | pte_unmap_unlock(pte - 1, ptl); | |
3120 | pmd = mm_find_pmd(vma->vm_mm, address); | |
3121 | if (!pmd) | |
3122 | return; | |
3123 | pte = pte_offset_map_lock(vma->vm_mm, pmd, | |
3124 | address, &ptl); | |
3125 | } | |
3126 | ||
3127 | if (!is_swap_pte(*pte)) | |
e9b61f19 KS |
3128 | continue; |
3129 | ||
bd56086f | 3130 | swp_entry = pte_to_swp_entry(*pte); |
e9b61f19 KS |
3131 | if (!is_migration_entry(swp_entry)) |
3132 | continue; | |
3133 | if (migration_entry_to_page(swp_entry) != page) | |
3134 | continue; | |
3135 | ||
3136 | get_page(page); | |
3137 | page_add_anon_rmap(page, vma, address, false); | |
3138 | ||
3139 | entry = pte_mkold(mk_pte(page, vma->vm_page_prot)); | |
b8d3c4c3 MK |
3140 | if (PageDirty(page)) |
3141 | entry = pte_mkdirty(entry); | |
e9b61f19 KS |
3142 | if (is_write_migration_entry(swp_entry)) |
3143 | entry = maybe_mkwrite(entry, vma); | |
3144 | ||
3145 | flush_dcache_page(page); | |
bd56086f | 3146 | set_pte_at(vma->vm_mm, address, pte, entry); |
e9b61f19 KS |
3147 | |
3148 | /* No need to invalidate - it was non-present before */ | |
bd56086f | 3149 | update_mmu_cache(vma, address, pte); |
e9b61f19 | 3150 | } |
bd56086f | 3151 | pte_unmap_unlock(pte - 1, ptl); |
e9b61f19 KS |
3152 | } |
3153 | ||
3154 | static void unfreeze_page(struct anon_vma *anon_vma, struct page *page) | |
3155 | { | |
3156 | struct anon_vma_chain *avc; | |
3157 | pgoff_t pgoff = page_to_pgoff(page); | |
3158 | ||
3159 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, | |
3160 | pgoff, pgoff + HPAGE_PMD_NR - 1) { | |
3161 | unsigned long address = __vma_address(page, avc->vma); | |
3162 | ||
3163 | mmu_notifier_invalidate_range_start(avc->vma->vm_mm, | |
3164 | address, address + HPAGE_PMD_SIZE); | |
3165 | unfreeze_page_vma(avc->vma, page, address); | |
3166 | mmu_notifier_invalidate_range_end(avc->vma->vm_mm, | |
3167 | address, address + HPAGE_PMD_SIZE); | |
3168 | } | |
3169 | } | |
3170 | ||
e9b61f19 KS |
3171 | static int __split_huge_page_tail(struct page *head, int tail, |
3172 | struct lruvec *lruvec, struct list_head *list) | |
3173 | { | |
3174 | int mapcount; | |
3175 | struct page *page_tail = head + tail; | |
3176 | ||
3177 | mapcount = atomic_read(&page_tail->_mapcount) + 1; | |
3178 | VM_BUG_ON_PAGE(atomic_read(&page_tail->_count) != 0, page_tail); | |
3179 | ||
3180 | /* | |
3181 | * tail_page->_count is zero and not changing from under us. But | |
3182 | * get_page_unless_zero() may be running from under us on the | |
3183 | * tail_page. If we used atomic_set() below instead of atomic_add(), we | |
3184 | * would then run atomic_set() concurrently with | |
3185 | * get_page_unless_zero(), and atomic_set() is implemented in C not | |
3186 | * using locked ops. spin_unlock on x86 sometime uses locked ops | |
3187 | * because of PPro errata 66, 92, so unless somebody can guarantee | |
3188 | * atomic_set() here would be safe on all archs (and not only on x86), | |
3189 | * it's safer to use atomic_add(). | |
3190 | */ | |
3191 | atomic_add(mapcount + 1, &page_tail->_count); | |
3192 | ||
3193 | ||
3194 | page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; | |
3195 | page_tail->flags |= (head->flags & | |
3196 | ((1L << PG_referenced) | | |
3197 | (1L << PG_swapbacked) | | |
3198 | (1L << PG_mlocked) | | |
3199 | (1L << PG_uptodate) | | |
3200 | (1L << PG_active) | | |
3201 | (1L << PG_locked) | | |
b8d3c4c3 MK |
3202 | (1L << PG_unevictable) | |
3203 | (1L << PG_dirty))); | |
e9b61f19 KS |
3204 | |
3205 | /* | |
3206 | * After clearing PageTail the gup refcount can be released. | |
3207 | * Page flags also must be visible before we make the page non-compound. | |
3208 | */ | |
3209 | smp_wmb(); | |
3210 | ||
3211 | clear_compound_head(page_tail); | |
3212 | ||
3213 | if (page_is_young(head)) | |
3214 | set_page_young(page_tail); | |
3215 | if (page_is_idle(head)) | |
3216 | set_page_idle(page_tail); | |
3217 | ||
3218 | /* ->mapping in first tail page is compound_mapcount */ | |
9a982250 | 3219 | VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING, |
e9b61f19 KS |
3220 | page_tail); |
3221 | page_tail->mapping = head->mapping; | |
3222 | ||
3223 | page_tail->index = head->index + tail; | |
3224 | page_cpupid_xchg_last(page_tail, page_cpupid_last(head)); | |
3225 | lru_add_page_tail(head, page_tail, lruvec, list); | |
3226 | ||
3227 | return mapcount; | |
3228 | } | |
3229 | ||
3230 | static void __split_huge_page(struct page *page, struct list_head *list) | |
3231 | { | |
3232 | struct page *head = compound_head(page); | |
3233 | struct zone *zone = page_zone(head); | |
3234 | struct lruvec *lruvec; | |
3235 | int i, tail_mapcount; | |
3236 | ||
3237 | /* prevent PageLRU to go away from under us, and freeze lru stats */ | |
3238 | spin_lock_irq(&zone->lru_lock); | |
3239 | lruvec = mem_cgroup_page_lruvec(head, zone); | |
3240 | ||
3241 | /* complete memcg works before add pages to LRU */ | |
3242 | mem_cgroup_split_huge_fixup(head); | |
3243 | ||
3244 | tail_mapcount = 0; | |
3245 | for (i = HPAGE_PMD_NR - 1; i >= 1; i--) | |
3246 | tail_mapcount += __split_huge_page_tail(head, i, lruvec, list); | |
3247 | atomic_sub(tail_mapcount, &head->_count); | |
3248 | ||
3249 | ClearPageCompound(head); | |
3250 | spin_unlock_irq(&zone->lru_lock); | |
3251 | ||
3252 | unfreeze_page(page_anon_vma(head), head); | |
3253 | ||
3254 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
3255 | struct page *subpage = head + i; | |
3256 | if (subpage == page) | |
3257 | continue; | |
3258 | unlock_page(subpage); | |
3259 | ||
3260 | /* | |
3261 | * Subpages may be freed if there wasn't any mapping | |
3262 | * like if add_to_swap() is running on a lru page that | |
3263 | * had its mapping zapped. And freeing these pages | |
3264 | * requires taking the lru_lock so we do the put_page | |
3265 | * of the tail pages after the split is complete. | |
3266 | */ | |
3267 | put_page(subpage); | |
3268 | } | |
3269 | } | |
3270 | ||
b20ce5e0 KS |
3271 | int total_mapcount(struct page *page) |
3272 | { | |
3273 | int i, ret; | |
3274 | ||
3275 | VM_BUG_ON_PAGE(PageTail(page), page); | |
3276 | ||
3277 | if (likely(!PageCompound(page))) | |
3278 | return atomic_read(&page->_mapcount) + 1; | |
3279 | ||
3280 | ret = compound_mapcount(page); | |
3281 | if (PageHuge(page)) | |
3282 | return ret; | |
3283 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
3284 | ret += atomic_read(&page[i]._mapcount) + 1; | |
3285 | if (PageDoubleMap(page)) | |
3286 | ret -= HPAGE_PMD_NR; | |
3287 | return ret; | |
3288 | } | |
3289 | ||
e9b61f19 KS |
3290 | /* |
3291 | * This function splits huge page into normal pages. @page can point to any | |
3292 | * subpage of huge page to split. Split doesn't change the position of @page. | |
3293 | * | |
3294 | * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. | |
3295 | * The huge page must be locked. | |
3296 | * | |
3297 | * If @list is null, tail pages will be added to LRU list, otherwise, to @list. | |
3298 | * | |
3299 | * Both head page and tail pages will inherit mapping, flags, and so on from | |
3300 | * the hugepage. | |
3301 | * | |
3302 | * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if | |
3303 | * they are not mapped. | |
3304 | * | |
3305 | * Returns 0 if the hugepage is split successfully. | |
3306 | * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under | |
3307 | * us. | |
3308 | */ | |
3309 | int split_huge_page_to_list(struct page *page, struct list_head *list) | |
3310 | { | |
3311 | struct page *head = compound_head(page); | |
3312 | struct anon_vma *anon_vma; | |
3313 | int count, mapcount, ret; | |
d9654322 | 3314 | bool mlocked; |
e9b61f19 KS |
3315 | |
3316 | VM_BUG_ON_PAGE(is_huge_zero_page(page), page); | |
3317 | VM_BUG_ON_PAGE(!PageAnon(page), page); | |
3318 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
3319 | VM_BUG_ON_PAGE(!PageSwapBacked(page), page); | |
3320 | VM_BUG_ON_PAGE(!PageCompound(page), page); | |
3321 | ||
3322 | /* | |
3323 | * The caller does not necessarily hold an mmap_sem that would prevent | |
3324 | * the anon_vma disappearing so we first we take a reference to it | |
3325 | * and then lock the anon_vma for write. This is similar to | |
3326 | * page_lock_anon_vma_read except the write lock is taken to serialise | |
3327 | * against parallel split or collapse operations. | |
3328 | */ | |
3329 | anon_vma = page_get_anon_vma(head); | |
3330 | if (!anon_vma) { | |
3331 | ret = -EBUSY; | |
3332 | goto out; | |
3333 | } | |
3334 | anon_vma_lock_write(anon_vma); | |
3335 | ||
3336 | /* | |
3337 | * Racy check if we can split the page, before freeze_page() will | |
3338 | * split PMDs | |
3339 | */ | |
3340 | if (total_mapcount(head) != page_count(head) - 1) { | |
3341 | ret = -EBUSY; | |
3342 | goto out_unlock; | |
3343 | } | |
3344 | ||
d9654322 | 3345 | mlocked = PageMlocked(page); |
e9b61f19 KS |
3346 | freeze_page(anon_vma, head); |
3347 | VM_BUG_ON_PAGE(compound_mapcount(head), head); | |
3348 | ||
d9654322 KS |
3349 | /* Make sure the page is not on per-CPU pagevec as it takes pin */ |
3350 | if (mlocked) | |
3351 | lru_add_drain(); | |
3352 | ||
9a982250 KS |
3353 | /* Prevent deferred_split_scan() touching ->_count */ |
3354 | spin_lock(&split_queue_lock); | |
e9b61f19 KS |
3355 | count = page_count(head); |
3356 | mapcount = total_mapcount(head); | |
bd56086f | 3357 | if (!mapcount && count == 1) { |
9a982250 KS |
3358 | if (!list_empty(page_deferred_list(head))) { |
3359 | split_queue_len--; | |
3360 | list_del(page_deferred_list(head)); | |
3361 | } | |
3362 | spin_unlock(&split_queue_lock); | |
e9b61f19 KS |
3363 | __split_huge_page(page, list); |
3364 | ret = 0; | |
bd56086f | 3365 | } else if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) { |
9a982250 | 3366 | spin_unlock(&split_queue_lock); |
e9b61f19 KS |
3367 | pr_alert("total_mapcount: %u, page_count(): %u\n", |
3368 | mapcount, count); | |
3369 | if (PageTail(page)) | |
3370 | dump_page(head, NULL); | |
bd56086f | 3371 | dump_page(page, "total_mapcount(head) > 0"); |
e9b61f19 KS |
3372 | BUG(); |
3373 | } else { | |
9a982250 | 3374 | spin_unlock(&split_queue_lock); |
e9b61f19 KS |
3375 | unfreeze_page(anon_vma, head); |
3376 | ret = -EBUSY; | |
3377 | } | |
3378 | ||
3379 | out_unlock: | |
3380 | anon_vma_unlock_write(anon_vma); | |
3381 | put_anon_vma(anon_vma); | |
3382 | out: | |
3383 | count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); | |
3384 | return ret; | |
3385 | } | |
9a982250 KS |
3386 | |
3387 | void free_transhuge_page(struct page *page) | |
3388 | { | |
3389 | unsigned long flags; | |
3390 | ||
3391 | spin_lock_irqsave(&split_queue_lock, flags); | |
3392 | if (!list_empty(page_deferred_list(page))) { | |
3393 | split_queue_len--; | |
3394 | list_del(page_deferred_list(page)); | |
3395 | } | |
3396 | spin_unlock_irqrestore(&split_queue_lock, flags); | |
3397 | free_compound_page(page); | |
3398 | } | |
3399 | ||
3400 | void deferred_split_huge_page(struct page *page) | |
3401 | { | |
3402 | unsigned long flags; | |
3403 | ||
3404 | VM_BUG_ON_PAGE(!PageTransHuge(page), page); | |
3405 | ||
3406 | spin_lock_irqsave(&split_queue_lock, flags); | |
3407 | if (list_empty(page_deferred_list(page))) { | |
3408 | list_add_tail(page_deferred_list(page), &split_queue); | |
3409 | split_queue_len++; | |
3410 | } | |
3411 | spin_unlock_irqrestore(&split_queue_lock, flags); | |
3412 | } | |
3413 | ||
3414 | static unsigned long deferred_split_count(struct shrinker *shrink, | |
3415 | struct shrink_control *sc) | |
3416 | { | |
3417 | /* | |
3418 | * Split a page from split_queue will free up at least one page, | |
3419 | * at most HPAGE_PMD_NR - 1. We don't track exact number. | |
3420 | * Let's use HPAGE_PMD_NR / 2 as ballpark. | |
3421 | */ | |
3422 | return ACCESS_ONCE(split_queue_len) * HPAGE_PMD_NR / 2; | |
3423 | } | |
3424 | ||
3425 | static unsigned long deferred_split_scan(struct shrinker *shrink, | |
3426 | struct shrink_control *sc) | |
3427 | { | |
3428 | unsigned long flags; | |
3429 | LIST_HEAD(list), *pos, *next; | |
3430 | struct page *page; | |
3431 | int split = 0; | |
3432 | ||
3433 | spin_lock_irqsave(&split_queue_lock, flags); | |
3434 | list_splice_init(&split_queue, &list); | |
3435 | ||
3436 | /* Take pin on all head pages to avoid freeing them under us */ | |
3437 | list_for_each_safe(pos, next, &list) { | |
3438 | page = list_entry((void *)pos, struct page, mapping); | |
3439 | page = compound_head(page); | |
3440 | /* race with put_compound_page() */ | |
3441 | if (!get_page_unless_zero(page)) { | |
3442 | list_del_init(page_deferred_list(page)); | |
3443 | split_queue_len--; | |
3444 | } | |
3445 | } | |
3446 | spin_unlock_irqrestore(&split_queue_lock, flags); | |
3447 | ||
3448 | list_for_each_safe(pos, next, &list) { | |
3449 | page = list_entry((void *)pos, struct page, mapping); | |
3450 | lock_page(page); | |
3451 | /* split_huge_page() removes page from list on success */ | |
3452 | if (!split_huge_page(page)) | |
3453 | split++; | |
3454 | unlock_page(page); | |
3455 | put_page(page); | |
3456 | } | |
3457 | ||
3458 | spin_lock_irqsave(&split_queue_lock, flags); | |
3459 | list_splice_tail(&list, &split_queue); | |
3460 | spin_unlock_irqrestore(&split_queue_lock, flags); | |
3461 | ||
3462 | return split * HPAGE_PMD_NR / 2; | |
3463 | } | |
3464 | ||
3465 | static struct shrinker deferred_split_shrinker = { | |
3466 | .count_objects = deferred_split_count, | |
3467 | .scan_objects = deferred_split_scan, | |
3468 | .seeks = DEFAULT_SEEKS, | |
3469 | }; | |
49071d43 KS |
3470 | |
3471 | #ifdef CONFIG_DEBUG_FS | |
3472 | static int split_huge_pages_set(void *data, u64 val) | |
3473 | { | |
3474 | struct zone *zone; | |
3475 | struct page *page; | |
3476 | unsigned long pfn, max_zone_pfn; | |
3477 | unsigned long total = 0, split = 0; | |
3478 | ||
3479 | if (val != 1) | |
3480 | return -EINVAL; | |
3481 | ||
3482 | for_each_populated_zone(zone) { | |
3483 | max_zone_pfn = zone_end_pfn(zone); | |
3484 | for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { | |
3485 | if (!pfn_valid(pfn)) | |
3486 | continue; | |
3487 | ||
3488 | page = pfn_to_page(pfn); | |
3489 | if (!get_page_unless_zero(page)) | |
3490 | continue; | |
3491 | ||
3492 | if (zone != page_zone(page)) | |
3493 | goto next; | |
3494 | ||
3495 | if (!PageHead(page) || !PageAnon(page) || | |
3496 | PageHuge(page)) | |
3497 | goto next; | |
3498 | ||
3499 | total++; | |
3500 | lock_page(page); | |
3501 | if (!split_huge_page(page)) | |
3502 | split++; | |
3503 | unlock_page(page); | |
3504 | next: | |
3505 | put_page(page); | |
3506 | } | |
3507 | } | |
3508 | ||
3509 | pr_info("%lu of %lu THP split", split, total); | |
3510 | ||
3511 | return 0; | |
3512 | } | |
3513 | DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set, | |
3514 | "%llu\n"); | |
3515 | ||
3516 | static int __init split_huge_pages_debugfs(void) | |
3517 | { | |
3518 | void *ret; | |
3519 | ||
3520 | ret = debugfs_create_file("split_huge_pages", 0644, NULL, NULL, | |
3521 | &split_huge_pages_fops); | |
3522 | if (!ret) | |
3523 | pr_warn("Failed to create split_huge_pages in debugfs"); | |
3524 | return 0; | |
3525 | } | |
3526 | late_initcall(split_huge_pages_debugfs); | |
3527 | #endif |