]>
Commit | Line | Data |
---|---|---|
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 | ||
8 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt | |
9 | ||
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> | |
17 | #include <linux/shrinker.h> | |
18 | #include <linux/mm_inline.h> | |
19 | #include <linux/swapops.h> | |
20 | #include <linux/dax.h> | |
21 | #include <linux/kthread.h> | |
22 | #include <linux/khugepaged.h> | |
23 | #include <linux/freezer.h> | |
24 | #include <linux/pfn_t.h> | |
25 | #include <linux/mman.h> | |
26 | #include <linux/memremap.h> | |
27 | #include <linux/pagemap.h> | |
28 | #include <linux/debugfs.h> | |
29 | #include <linux/migrate.h> | |
30 | #include <linux/hashtable.h> | |
31 | #include <linux/userfaultfd_k.h> | |
32 | #include <linux/page_idle.h> | |
33 | ||
34 | #include <asm/tlb.h> | |
35 | #include <asm/pgalloc.h> | |
36 | #include "internal.h" | |
37 | ||
38 | enum scan_result { | |
39 | SCAN_FAIL, | |
40 | SCAN_SUCCEED, | |
41 | SCAN_PMD_NULL, | |
42 | SCAN_EXCEED_NONE_PTE, | |
43 | SCAN_PTE_NON_PRESENT, | |
44 | SCAN_PAGE_RO, | |
45 | SCAN_NO_REFERENCED_PAGE, | |
46 | SCAN_PAGE_NULL, | |
47 | SCAN_SCAN_ABORT, | |
48 | SCAN_PAGE_COUNT, | |
49 | SCAN_PAGE_LRU, | |
50 | SCAN_PAGE_LOCK, | |
51 | SCAN_PAGE_ANON, | |
52 | SCAN_PAGE_COMPOUND, | |
53 | SCAN_ANY_PROCESS, | |
54 | SCAN_VMA_NULL, | |
55 | SCAN_VMA_CHECK, | |
56 | SCAN_ADDRESS_RANGE, | |
57 | SCAN_SWAP_CACHE_PAGE, | |
58 | SCAN_DEL_PAGE_LRU, | |
59 | SCAN_ALLOC_HUGE_PAGE_FAIL, | |
60 | SCAN_CGROUP_CHARGE_FAIL, | |
61 | SCAN_EXCEED_SWAP_PTE | |
62 | }; | |
63 | ||
64 | #define CREATE_TRACE_POINTS | |
65 | #include <trace/events/huge_memory.h> | |
66 | ||
67 | /* | |
68 | * By default transparent hugepage support is disabled in order that avoid | |
69 | * to risk increase the memory footprint of applications without a guaranteed | |
70 | * benefit. When transparent hugepage support is enabled, is for all mappings, | |
71 | * and khugepaged scans all mappings. | |
72 | * Defrag is invoked by khugepaged hugepage allocations and by page faults | |
73 | * for all hugepage allocations. | |
74 | */ | |
75 | unsigned long transparent_hugepage_flags __read_mostly = | |
76 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS | |
77 | (1<<TRANSPARENT_HUGEPAGE_FLAG)| | |
78 | #endif | |
79 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE | |
80 | (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| | |
81 | #endif | |
82 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)| | |
83 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| | |
84 | (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
85 | ||
86 | /* default scan 8*512 pte (or vmas) every 30 second */ | |
87 | static unsigned int khugepaged_pages_to_scan __read_mostly; | |
88 | static unsigned int khugepaged_pages_collapsed; | |
89 | static unsigned int khugepaged_full_scans; | |
90 | static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; | |
91 | /* during fragmentation poll the hugepage allocator once every minute */ | |
92 | static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; | |
93 | static unsigned long khugepaged_sleep_expire; | |
94 | static struct task_struct *khugepaged_thread __read_mostly; | |
95 | static DEFINE_MUTEX(khugepaged_mutex); | |
96 | static DEFINE_SPINLOCK(khugepaged_mm_lock); | |
97 | static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); | |
98 | /* | |
99 | * default collapse hugepages if there is at least one pte mapped like | |
100 | * it would have happened if the vma was large enough during page | |
101 | * fault. | |
102 | */ | |
103 | static unsigned int khugepaged_max_ptes_none __read_mostly; | |
104 | static unsigned int khugepaged_max_ptes_swap __read_mostly; | |
105 | ||
106 | static int khugepaged(void *none); | |
107 | static int khugepaged_slab_init(void); | |
108 | static void khugepaged_slab_exit(void); | |
109 | ||
110 | #define MM_SLOTS_HASH_BITS 10 | |
111 | static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); | |
112 | ||
113 | static struct kmem_cache *mm_slot_cache __read_mostly; | |
114 | ||
115 | /** | |
116 | * struct mm_slot - hash lookup from mm to mm_slot | |
117 | * @hash: hash collision list | |
118 | * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head | |
119 | * @mm: the mm that this information is valid for | |
120 | */ | |
121 | struct mm_slot { | |
122 | struct hlist_node hash; | |
123 | struct list_head mm_node; | |
124 | struct mm_struct *mm; | |
125 | }; | |
126 | ||
127 | /** | |
128 | * struct khugepaged_scan - cursor for scanning | |
129 | * @mm_head: the head of the mm list to scan | |
130 | * @mm_slot: the current mm_slot we are scanning | |
131 | * @address: the next address inside that to be scanned | |
132 | * | |
133 | * There is only the one khugepaged_scan instance of this cursor structure. | |
134 | */ | |
135 | struct khugepaged_scan { | |
136 | struct list_head mm_head; | |
137 | struct mm_slot *mm_slot; | |
138 | unsigned long address; | |
139 | }; | |
140 | static struct khugepaged_scan khugepaged_scan = { | |
141 | .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), | |
142 | }; | |
143 | ||
144 | static struct shrinker deferred_split_shrinker; | |
145 | ||
146 | static void set_recommended_min_free_kbytes(void) | |
147 | { | |
148 | struct zone *zone; | |
149 | int nr_zones = 0; | |
150 | unsigned long recommended_min; | |
151 | ||
152 | for_each_populated_zone(zone) | |
153 | nr_zones++; | |
154 | ||
155 | /* Ensure 2 pageblocks are free to assist fragmentation avoidance */ | |
156 | recommended_min = pageblock_nr_pages * nr_zones * 2; | |
157 | ||
158 | /* | |
159 | * Make sure that on average at least two pageblocks are almost free | |
160 | * of another type, one for a migratetype to fall back to and a | |
161 | * second to avoid subsequent fallbacks of other types There are 3 | |
162 | * MIGRATE_TYPES we care about. | |
163 | */ | |
164 | recommended_min += pageblock_nr_pages * nr_zones * | |
165 | MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; | |
166 | ||
167 | /* don't ever allow to reserve more than 5% of the lowmem */ | |
168 | recommended_min = min(recommended_min, | |
169 | (unsigned long) nr_free_buffer_pages() / 20); | |
170 | recommended_min <<= (PAGE_SHIFT-10); | |
171 | ||
172 | if (recommended_min > min_free_kbytes) { | |
173 | if (user_min_free_kbytes >= 0) | |
174 | pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n", | |
175 | min_free_kbytes, recommended_min); | |
176 | ||
177 | min_free_kbytes = recommended_min; | |
178 | } | |
179 | setup_per_zone_wmarks(); | |
180 | } | |
181 | ||
182 | static int start_stop_khugepaged(void) | |
183 | { | |
184 | int err = 0; | |
185 | if (khugepaged_enabled()) { | |
186 | if (!khugepaged_thread) | |
187 | khugepaged_thread = kthread_run(khugepaged, NULL, | |
188 | "khugepaged"); | |
189 | if (IS_ERR(khugepaged_thread)) { | |
190 | pr_err("khugepaged: kthread_run(khugepaged) failed\n"); | |
191 | err = PTR_ERR(khugepaged_thread); | |
192 | khugepaged_thread = NULL; | |
193 | goto fail; | |
194 | } | |
195 | ||
196 | if (!list_empty(&khugepaged_scan.mm_head)) | |
197 | wake_up_interruptible(&khugepaged_wait); | |
198 | ||
199 | set_recommended_min_free_kbytes(); | |
200 | } else if (khugepaged_thread) { | |
201 | kthread_stop(khugepaged_thread); | |
202 | khugepaged_thread = NULL; | |
203 | } | |
204 | fail: | |
205 | return err; | |
206 | } | |
207 | ||
208 | static atomic_t huge_zero_refcount; | |
209 | struct page *huge_zero_page __read_mostly; | |
210 | ||
211 | struct page *get_huge_zero_page(void) | |
212 | { | |
213 | struct page *zero_page; | |
214 | retry: | |
215 | if (likely(atomic_inc_not_zero(&huge_zero_refcount))) | |
216 | return READ_ONCE(huge_zero_page); | |
217 | ||
218 | zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, | |
219 | HPAGE_PMD_ORDER); | |
220 | if (!zero_page) { | |
221 | count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); | |
222 | return NULL; | |
223 | } | |
224 | count_vm_event(THP_ZERO_PAGE_ALLOC); | |
225 | preempt_disable(); | |
226 | if (cmpxchg(&huge_zero_page, NULL, zero_page)) { | |
227 | preempt_enable(); | |
228 | __free_pages(zero_page, compound_order(zero_page)); | |
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(); | |
235 | return READ_ONCE(huge_zero_page); | |
236 | } | |
237 | ||
238 | void put_huge_zero_page(void) | |
239 | { | |
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)); | |
245 | } | |
246 | ||
247 | static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, | |
248 | struct shrink_control *sc) | |
249 | { | |
250 | /* we can free zero page only if last reference remains */ | |
251 | return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; | |
252 | } | |
253 | ||
254 | static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, | |
255 | struct shrink_control *sc) | |
256 | { | |
257 | if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { | |
258 | struct page *zero_page = xchg(&huge_zero_page, NULL); | |
259 | BUG_ON(zero_page == NULL); | |
260 | __free_pages(zero_page, compound_order(zero_page)); | |
261 | return HPAGE_PMD_NR; | |
262 | } | |
263 | ||
264 | return 0; | |
265 | } | |
266 | ||
267 | static struct shrinker huge_zero_page_shrinker = { | |
268 | .count_objects = shrink_huge_zero_page_count, | |
269 | .scan_objects = shrink_huge_zero_page_scan, | |
270 | .seeks = DEFAULT_SEEKS, | |
271 | }; | |
272 | ||
273 | #ifdef CONFIG_SYSFS | |
274 | ||
275 | static ssize_t triple_flag_store(struct kobject *kobj, | |
276 | struct kobj_attribute *attr, | |
277 | const char *buf, size_t count, | |
278 | enum transparent_hugepage_flag enabled, | |
279 | enum transparent_hugepage_flag deferred, | |
280 | enum transparent_hugepage_flag req_madv) | |
281 | { | |
282 | if (!memcmp("defer", buf, | |
283 | min(sizeof("defer")-1, count))) { | |
284 | if (enabled == deferred) | |
285 | return -EINVAL; | |
286 | clear_bit(enabled, &transparent_hugepage_flags); | |
287 | clear_bit(req_madv, &transparent_hugepage_flags); | |
288 | set_bit(deferred, &transparent_hugepage_flags); | |
289 | } else if (!memcmp("always", buf, | |
290 | min(sizeof("always")-1, count))) { | |
291 | clear_bit(deferred, &transparent_hugepage_flags); | |
292 | clear_bit(req_madv, &transparent_hugepage_flags); | |
293 | set_bit(enabled, &transparent_hugepage_flags); | |
294 | } else if (!memcmp("madvise", buf, | |
295 | min(sizeof("madvise")-1, count))) { | |
296 | clear_bit(enabled, &transparent_hugepage_flags); | |
297 | clear_bit(deferred, &transparent_hugepage_flags); | |
298 | set_bit(req_madv, &transparent_hugepage_flags); | |
299 | } else if (!memcmp("never", buf, | |
300 | min(sizeof("never")-1, count))) { | |
301 | clear_bit(enabled, &transparent_hugepage_flags); | |
302 | clear_bit(req_madv, &transparent_hugepage_flags); | |
303 | clear_bit(deferred, &transparent_hugepage_flags); | |
304 | } else | |
305 | return -EINVAL; | |
306 | ||
307 | return count; | |
308 | } | |
309 | ||
310 | static ssize_t enabled_show(struct kobject *kobj, | |
311 | struct kobj_attribute *attr, char *buf) | |
312 | { | |
313 | if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags)) | |
314 | return sprintf(buf, "[always] madvise never\n"); | |
315 | else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags)) | |
316 | return sprintf(buf, "always [madvise] never\n"); | |
317 | else | |
318 | return sprintf(buf, "always madvise [never]\n"); | |
319 | } | |
320 | ||
321 | static ssize_t enabled_store(struct kobject *kobj, | |
322 | struct kobj_attribute *attr, | |
323 | const char *buf, size_t count) | |
324 | { | |
325 | ssize_t ret; | |
326 | ||
327 | ret = triple_flag_store(kobj, attr, buf, count, | |
328 | TRANSPARENT_HUGEPAGE_FLAG, | |
329 | TRANSPARENT_HUGEPAGE_FLAG, | |
330 | TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); | |
331 | ||
332 | if (ret > 0) { | |
333 | int err; | |
334 | ||
335 | mutex_lock(&khugepaged_mutex); | |
336 | err = start_stop_khugepaged(); | |
337 | mutex_unlock(&khugepaged_mutex); | |
338 | ||
339 | if (err) | |
340 | ret = err; | |
341 | } | |
342 | ||
343 | return ret; | |
344 | } | |
345 | static struct kobj_attribute enabled_attr = | |
346 | __ATTR(enabled, 0644, enabled_show, enabled_store); | |
347 | ||
348 | static ssize_t single_flag_show(struct kobject *kobj, | |
349 | struct kobj_attribute *attr, char *buf, | |
350 | enum transparent_hugepage_flag flag) | |
351 | { | |
352 | return sprintf(buf, "%d\n", | |
353 | !!test_bit(flag, &transparent_hugepage_flags)); | |
354 | } | |
355 | ||
356 | static ssize_t single_flag_store(struct kobject *kobj, | |
357 | struct kobj_attribute *attr, | |
358 | const char *buf, size_t count, | |
359 | enum transparent_hugepage_flag flag) | |
360 | { | |
361 | unsigned long value; | |
362 | int ret; | |
363 | ||
364 | ret = kstrtoul(buf, 10, &value); | |
365 | if (ret < 0) | |
366 | return ret; | |
367 | if (value > 1) | |
368 | return -EINVAL; | |
369 | ||
370 | if (value) | |
371 | set_bit(flag, &transparent_hugepage_flags); | |
372 | else | |
373 | clear_bit(flag, &transparent_hugepage_flags); | |
374 | ||
375 | return count; | |
376 | } | |
377 | ||
378 | /* | |
379 | * Currently defrag only disables __GFP_NOWAIT for allocation. A blind | |
380 | * __GFP_REPEAT is too aggressive, it's never worth swapping tons of | |
381 | * memory just to allocate one more hugepage. | |
382 | */ | |
383 | static ssize_t defrag_show(struct kobject *kobj, | |
384 | struct kobj_attribute *attr, char *buf) | |
385 | { | |
386 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) | |
387 | return sprintf(buf, "[always] defer madvise never\n"); | |
388 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) | |
389 | return sprintf(buf, "always [defer] madvise never\n"); | |
390 | else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) | |
391 | return sprintf(buf, "always defer [madvise] never\n"); | |
392 | else | |
393 | return sprintf(buf, "always defer madvise [never]\n"); | |
394 | ||
395 | } | |
396 | static ssize_t defrag_store(struct kobject *kobj, | |
397 | struct kobj_attribute *attr, | |
398 | const char *buf, size_t count) | |
399 | { | |
400 | return triple_flag_store(kobj, attr, buf, count, | |
401 | TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, | |
402 | TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, | |
403 | TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); | |
404 | } | |
405 | static struct kobj_attribute defrag_attr = | |
406 | __ATTR(defrag, 0644, defrag_show, defrag_store); | |
407 | ||
408 | static ssize_t use_zero_page_show(struct kobject *kobj, | |
409 | struct kobj_attribute *attr, char *buf) | |
410 | { | |
411 | return single_flag_show(kobj, attr, buf, | |
412 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
413 | } | |
414 | static ssize_t use_zero_page_store(struct kobject *kobj, | |
415 | struct kobj_attribute *attr, const char *buf, size_t count) | |
416 | { | |
417 | return single_flag_store(kobj, attr, buf, count, | |
418 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
419 | } | |
420 | static struct kobj_attribute use_zero_page_attr = | |
421 | __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); | |
422 | #ifdef CONFIG_DEBUG_VM | |
423 | static ssize_t debug_cow_show(struct kobject *kobj, | |
424 | struct kobj_attribute *attr, char *buf) | |
425 | { | |
426 | return single_flag_show(kobj, attr, buf, | |
427 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); | |
428 | } | |
429 | static ssize_t debug_cow_store(struct kobject *kobj, | |
430 | struct kobj_attribute *attr, | |
431 | const char *buf, size_t count) | |
432 | { | |
433 | return single_flag_store(kobj, attr, buf, count, | |
434 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); | |
435 | } | |
436 | static struct kobj_attribute debug_cow_attr = | |
437 | __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); | |
438 | #endif /* CONFIG_DEBUG_VM */ | |
439 | ||
440 | static struct attribute *hugepage_attr[] = { | |
441 | &enabled_attr.attr, | |
442 | &defrag_attr.attr, | |
443 | &use_zero_page_attr.attr, | |
444 | #ifdef CONFIG_DEBUG_VM | |
445 | &debug_cow_attr.attr, | |
446 | #endif | |
447 | NULL, | |
448 | }; | |
449 | ||
450 | static struct attribute_group hugepage_attr_group = { | |
451 | .attrs = hugepage_attr, | |
452 | }; | |
453 | ||
454 | static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, | |
455 | struct kobj_attribute *attr, | |
456 | char *buf) | |
457 | { | |
458 | return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs); | |
459 | } | |
460 | ||
461 | static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, | |
462 | struct kobj_attribute *attr, | |
463 | const char *buf, size_t count) | |
464 | { | |
465 | unsigned long msecs; | |
466 | int err; | |
467 | ||
468 | err = kstrtoul(buf, 10, &msecs); | |
469 | if (err || msecs > UINT_MAX) | |
470 | return -EINVAL; | |
471 | ||
472 | khugepaged_scan_sleep_millisecs = msecs; | |
473 | khugepaged_sleep_expire = 0; | |
474 | wake_up_interruptible(&khugepaged_wait); | |
475 | ||
476 | return count; | |
477 | } | |
478 | static struct kobj_attribute scan_sleep_millisecs_attr = | |
479 | __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show, | |
480 | scan_sleep_millisecs_store); | |
481 | ||
482 | static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, | |
483 | struct kobj_attribute *attr, | |
484 | char *buf) | |
485 | { | |
486 | return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs); | |
487 | } | |
488 | ||
489 | static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, | |
490 | struct kobj_attribute *attr, | |
491 | const char *buf, size_t count) | |
492 | { | |
493 | unsigned long msecs; | |
494 | int err; | |
495 | ||
496 | err = kstrtoul(buf, 10, &msecs); | |
497 | if (err || msecs > UINT_MAX) | |
498 | return -EINVAL; | |
499 | ||
500 | khugepaged_alloc_sleep_millisecs = msecs; | |
501 | khugepaged_sleep_expire = 0; | |
502 | wake_up_interruptible(&khugepaged_wait); | |
503 | ||
504 | return count; | |
505 | } | |
506 | static struct kobj_attribute alloc_sleep_millisecs_attr = | |
507 | __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show, | |
508 | alloc_sleep_millisecs_store); | |
509 | ||
510 | static ssize_t pages_to_scan_show(struct kobject *kobj, | |
511 | struct kobj_attribute *attr, | |
512 | char *buf) | |
513 | { | |
514 | return sprintf(buf, "%u\n", khugepaged_pages_to_scan); | |
515 | } | |
516 | static ssize_t pages_to_scan_store(struct kobject *kobj, | |
517 | struct kobj_attribute *attr, | |
518 | const char *buf, size_t count) | |
519 | { | |
520 | int err; | |
521 | unsigned long pages; | |
522 | ||
523 | err = kstrtoul(buf, 10, &pages); | |
524 | if (err || !pages || pages > UINT_MAX) | |
525 | return -EINVAL; | |
526 | ||
527 | khugepaged_pages_to_scan = pages; | |
528 | ||
529 | return count; | |
530 | } | |
531 | static struct kobj_attribute pages_to_scan_attr = | |
532 | __ATTR(pages_to_scan, 0644, pages_to_scan_show, | |
533 | pages_to_scan_store); | |
534 | ||
535 | static ssize_t pages_collapsed_show(struct kobject *kobj, | |
536 | struct kobj_attribute *attr, | |
537 | char *buf) | |
538 | { | |
539 | return sprintf(buf, "%u\n", khugepaged_pages_collapsed); | |
540 | } | |
541 | static struct kobj_attribute pages_collapsed_attr = | |
542 | __ATTR_RO(pages_collapsed); | |
543 | ||
544 | static ssize_t full_scans_show(struct kobject *kobj, | |
545 | struct kobj_attribute *attr, | |
546 | char *buf) | |
547 | { | |
548 | return sprintf(buf, "%u\n", khugepaged_full_scans); | |
549 | } | |
550 | static struct kobj_attribute full_scans_attr = | |
551 | __ATTR_RO(full_scans); | |
552 | ||
553 | static ssize_t khugepaged_defrag_show(struct kobject *kobj, | |
554 | struct kobj_attribute *attr, char *buf) | |
555 | { | |
556 | return single_flag_show(kobj, attr, buf, | |
557 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); | |
558 | } | |
559 | static ssize_t khugepaged_defrag_store(struct kobject *kobj, | |
560 | struct kobj_attribute *attr, | |
561 | const char *buf, size_t count) | |
562 | { | |
563 | return single_flag_store(kobj, attr, buf, count, | |
564 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); | |
565 | } | |
566 | static struct kobj_attribute khugepaged_defrag_attr = | |
567 | __ATTR(defrag, 0644, khugepaged_defrag_show, | |
568 | khugepaged_defrag_store); | |
569 | ||
570 | /* | |
571 | * max_ptes_none controls if khugepaged should collapse hugepages over | |
572 | * any unmapped ptes in turn potentially increasing the memory | |
573 | * footprint of the vmas. When max_ptes_none is 0 khugepaged will not | |
574 | * reduce the available free memory in the system as it | |
575 | * runs. Increasing max_ptes_none will instead potentially reduce the | |
576 | * free memory in the system during the khugepaged scan. | |
577 | */ | |
578 | static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj, | |
579 | struct kobj_attribute *attr, | |
580 | char *buf) | |
581 | { | |
582 | return sprintf(buf, "%u\n", khugepaged_max_ptes_none); | |
583 | } | |
584 | static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj, | |
585 | struct kobj_attribute *attr, | |
586 | const char *buf, size_t count) | |
587 | { | |
588 | int err; | |
589 | unsigned long max_ptes_none; | |
590 | ||
591 | err = kstrtoul(buf, 10, &max_ptes_none); | |
592 | if (err || max_ptes_none > HPAGE_PMD_NR-1) | |
593 | return -EINVAL; | |
594 | ||
595 | khugepaged_max_ptes_none = max_ptes_none; | |
596 | ||
597 | return count; | |
598 | } | |
599 | static struct kobj_attribute khugepaged_max_ptes_none_attr = | |
600 | __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show, | |
601 | khugepaged_max_ptes_none_store); | |
602 | ||
603 | static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj, | |
604 | struct kobj_attribute *attr, | |
605 | char *buf) | |
606 | { | |
607 | return sprintf(buf, "%u\n", khugepaged_max_ptes_swap); | |
608 | } | |
609 | ||
610 | static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj, | |
611 | struct kobj_attribute *attr, | |
612 | const char *buf, size_t count) | |
613 | { | |
614 | int err; | |
615 | unsigned long max_ptes_swap; | |
616 | ||
617 | err = kstrtoul(buf, 10, &max_ptes_swap); | |
618 | if (err || max_ptes_swap > HPAGE_PMD_NR-1) | |
619 | return -EINVAL; | |
620 | ||
621 | khugepaged_max_ptes_swap = max_ptes_swap; | |
622 | ||
623 | return count; | |
624 | } | |
625 | ||
626 | static struct kobj_attribute khugepaged_max_ptes_swap_attr = | |
627 | __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show, | |
628 | khugepaged_max_ptes_swap_store); | |
629 | ||
630 | static struct attribute *khugepaged_attr[] = { | |
631 | &khugepaged_defrag_attr.attr, | |
632 | &khugepaged_max_ptes_none_attr.attr, | |
633 | &pages_to_scan_attr.attr, | |
634 | &pages_collapsed_attr.attr, | |
635 | &full_scans_attr.attr, | |
636 | &scan_sleep_millisecs_attr.attr, | |
637 | &alloc_sleep_millisecs_attr.attr, | |
638 | &khugepaged_max_ptes_swap_attr.attr, | |
639 | NULL, | |
640 | }; | |
641 | ||
642 | static struct attribute_group khugepaged_attr_group = { | |
643 | .attrs = khugepaged_attr, | |
644 | .name = "khugepaged", | |
645 | }; | |
646 | ||
647 | static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) | |
648 | { | |
649 | int err; | |
650 | ||
651 | *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); | |
652 | if (unlikely(!*hugepage_kobj)) { | |
653 | pr_err("failed to create transparent hugepage kobject\n"); | |
654 | return -ENOMEM; | |
655 | } | |
656 | ||
657 | err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); | |
658 | if (err) { | |
659 | pr_err("failed to register transparent hugepage group\n"); | |
660 | goto delete_obj; | |
661 | } | |
662 | ||
663 | err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); | |
664 | if (err) { | |
665 | pr_err("failed to register transparent hugepage group\n"); | |
666 | goto remove_hp_group; | |
667 | } | |
668 | ||
669 | return 0; | |
670 | ||
671 | remove_hp_group: | |
672 | sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); | |
673 | delete_obj: | |
674 | kobject_put(*hugepage_kobj); | |
675 | return err; | |
676 | } | |
677 | ||
678 | static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
679 | { | |
680 | sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); | |
681 | sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); | |
682 | kobject_put(hugepage_kobj); | |
683 | } | |
684 | #else | |
685 | static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) | |
686 | { | |
687 | return 0; | |
688 | } | |
689 | ||
690 | static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
691 | { | |
692 | } | |
693 | #endif /* CONFIG_SYSFS */ | |
694 | ||
695 | static int __init hugepage_init(void) | |
696 | { | |
697 | int err; | |
698 | struct kobject *hugepage_kobj; | |
699 | ||
700 | if (!has_transparent_hugepage()) { | |
701 | transparent_hugepage_flags = 0; | |
702 | return -EINVAL; | |
703 | } | |
704 | ||
705 | khugepaged_pages_to_scan = HPAGE_PMD_NR * 8; | |
706 | khugepaged_max_ptes_none = HPAGE_PMD_NR - 1; | |
707 | khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8; | |
708 | /* | |
709 | * hugepages can't be allocated by the buddy allocator | |
710 | */ | |
711 | MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER); | |
712 | /* | |
713 | * we use page->mapping and page->index in second tail page | |
714 | * as list_head: assuming THP order >= 2 | |
715 | */ | |
716 | MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2); | |
717 | ||
718 | err = hugepage_init_sysfs(&hugepage_kobj); | |
719 | if (err) | |
720 | goto err_sysfs; | |
721 | ||
722 | err = khugepaged_slab_init(); | |
723 | if (err) | |
724 | goto err_slab; | |
725 | ||
726 | err = register_shrinker(&huge_zero_page_shrinker); | |
727 | if (err) | |
728 | goto err_hzp_shrinker; | |
729 | err = register_shrinker(&deferred_split_shrinker); | |
730 | if (err) | |
731 | goto err_split_shrinker; | |
732 | ||
733 | /* | |
734 | * By default disable transparent hugepages on smaller systems, | |
735 | * where the extra memory used could hurt more than TLB overhead | |
736 | * is likely to save. The admin can still enable it through /sys. | |
737 | */ | |
738 | if (totalram_pages < (512 << (20 - PAGE_SHIFT))) { | |
739 | transparent_hugepage_flags = 0; | |
740 | return 0; | |
741 | } | |
742 | ||
743 | err = start_stop_khugepaged(); | |
744 | if (err) | |
745 | goto err_khugepaged; | |
746 | ||
747 | return 0; | |
748 | err_khugepaged: | |
749 | unregister_shrinker(&deferred_split_shrinker); | |
750 | err_split_shrinker: | |
751 | unregister_shrinker(&huge_zero_page_shrinker); | |
752 | err_hzp_shrinker: | |
753 | khugepaged_slab_exit(); | |
754 | err_slab: | |
755 | hugepage_exit_sysfs(hugepage_kobj); | |
756 | err_sysfs: | |
757 | return err; | |
758 | } | |
759 | subsys_initcall(hugepage_init); | |
760 | ||
761 | static int __init setup_transparent_hugepage(char *str) | |
762 | { | |
763 | int ret = 0; | |
764 | if (!str) | |
765 | goto out; | |
766 | if (!strcmp(str, "always")) { | |
767 | set_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
768 | &transparent_hugepage_flags); | |
769 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
770 | &transparent_hugepage_flags); | |
771 | ret = 1; | |
772 | } else if (!strcmp(str, "madvise")) { | |
773 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
774 | &transparent_hugepage_flags); | |
775 | set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
776 | &transparent_hugepage_flags); | |
777 | ret = 1; | |
778 | } else if (!strcmp(str, "never")) { | |
779 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
780 | &transparent_hugepage_flags); | |
781 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
782 | &transparent_hugepage_flags); | |
783 | ret = 1; | |
784 | } | |
785 | out: | |
786 | if (!ret) | |
787 | pr_warn("transparent_hugepage= cannot parse, ignored\n"); | |
788 | return ret; | |
789 | } | |
790 | __setup("transparent_hugepage=", setup_transparent_hugepage); | |
791 | ||
792 | pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) | |
793 | { | |
794 | if (likely(vma->vm_flags & VM_WRITE)) | |
795 | pmd = pmd_mkwrite(pmd); | |
796 | return pmd; | |
797 | } | |
798 | ||
799 | static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot) | |
800 | { | |
801 | return pmd_mkhuge(mk_pmd(page, prot)); | |
802 | } | |
803 | ||
804 | static inline struct list_head *page_deferred_list(struct page *page) | |
805 | { | |
806 | /* | |
807 | * ->lru in the tail pages is occupied by compound_head. | |
808 | * Let's use ->mapping + ->index in the second tail page as list_head. | |
809 | */ | |
810 | return (struct list_head *)&page[2].mapping; | |
811 | } | |
812 | ||
813 | void prep_transhuge_page(struct page *page) | |
814 | { | |
815 | /* | |
816 | * we use page->mapping and page->indexlru in second tail page | |
817 | * as list_head: assuming THP order >= 2 | |
818 | */ | |
819 | ||
820 | INIT_LIST_HEAD(page_deferred_list(page)); | |
821 | set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR); | |
822 | } | |
823 | ||
824 | static int __do_huge_pmd_anonymous_page(struct fault_env *fe, struct page *page, | |
825 | gfp_t gfp) | |
826 | { | |
827 | struct vm_area_struct *vma = fe->vma; | |
828 | struct mem_cgroup *memcg; | |
829 | pgtable_t pgtable; | |
830 | unsigned long haddr = fe->address & HPAGE_PMD_MASK; | |
831 | ||
832 | VM_BUG_ON_PAGE(!PageCompound(page), page); | |
833 | ||
834 | if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) { | |
835 | put_page(page); | |
836 | count_vm_event(THP_FAULT_FALLBACK); | |
837 | return VM_FAULT_FALLBACK; | |
838 | } | |
839 | ||
840 | pgtable = pte_alloc_one(vma->vm_mm, haddr); | |
841 | if (unlikely(!pgtable)) { | |
842 | mem_cgroup_cancel_charge(page, memcg, true); | |
843 | put_page(page); | |
844 | return VM_FAULT_OOM; | |
845 | } | |
846 | ||
847 | clear_huge_page(page, haddr, HPAGE_PMD_NR); | |
848 | /* | |
849 | * The memory barrier inside __SetPageUptodate makes sure that | |
850 | * clear_huge_page writes become visible before the set_pmd_at() | |
851 | * write. | |
852 | */ | |
853 | __SetPageUptodate(page); | |
854 | ||
855 | fe->ptl = pmd_lock(vma->vm_mm, fe->pmd); | |
856 | if (unlikely(!pmd_none(*fe->pmd))) { | |
857 | spin_unlock(fe->ptl); | |
858 | mem_cgroup_cancel_charge(page, memcg, true); | |
859 | put_page(page); | |
860 | pte_free(vma->vm_mm, pgtable); | |
861 | } else { | |
862 | pmd_t entry; | |
863 | ||
864 | /* Deliver the page fault to userland */ | |
865 | if (userfaultfd_missing(vma)) { | |
866 | int ret; | |
867 | ||
868 | spin_unlock(fe->ptl); | |
869 | mem_cgroup_cancel_charge(page, memcg, true); | |
870 | put_page(page); | |
871 | pte_free(vma->vm_mm, pgtable); | |
872 | ret = handle_userfault(fe, VM_UFFD_MISSING); | |
873 | VM_BUG_ON(ret & VM_FAULT_FALLBACK); | |
874 | return ret; | |
875 | } | |
876 | ||
877 | entry = mk_huge_pmd(page, vma->vm_page_prot); | |
878 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
879 | page_add_new_anon_rmap(page, vma, haddr, true); | |
880 | mem_cgroup_commit_charge(page, memcg, false, true); | |
881 | lru_cache_add_active_or_unevictable(page, vma); | |
882 | pgtable_trans_huge_deposit(vma->vm_mm, fe->pmd, pgtable); | |
883 | set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry); | |
884 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); | |
885 | atomic_long_inc(&vma->vm_mm->nr_ptes); | |
886 | spin_unlock(fe->ptl); | |
887 | count_vm_event(THP_FAULT_ALLOC); | |
888 | } | |
889 | ||
890 | return 0; | |
891 | } | |
892 | ||
893 | /* | |
894 | * If THP is set to always then directly reclaim/compact as necessary | |
895 | * If set to defer then do no reclaim and defer to khugepaged | |
896 | * If set to madvise and the VMA is flagged then directly reclaim/compact | |
897 | */ | |
898 | static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma) | |
899 | { | |
900 | gfp_t reclaim_flags = 0; | |
901 | ||
902 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags) && | |
903 | (vma->vm_flags & VM_HUGEPAGE)) | |
904 | reclaim_flags = __GFP_DIRECT_RECLAIM; | |
905 | else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) | |
906 | reclaim_flags = __GFP_KSWAPD_RECLAIM; | |
907 | else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) | |
908 | reclaim_flags = __GFP_DIRECT_RECLAIM; | |
909 | ||
910 | return GFP_TRANSHUGE | reclaim_flags; | |
911 | } | |
912 | ||
913 | /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */ | |
914 | static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void) | |
915 | { | |
916 | return GFP_TRANSHUGE | (khugepaged_defrag() ? __GFP_DIRECT_RECLAIM : 0); | |
917 | } | |
918 | ||
919 | /* Caller must hold page table lock. */ | |
920 | static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, | |
921 | struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, | |
922 | struct page *zero_page) | |
923 | { | |
924 | pmd_t entry; | |
925 | if (!pmd_none(*pmd)) | |
926 | return false; | |
927 | entry = mk_pmd(zero_page, vma->vm_page_prot); | |
928 | entry = pmd_mkhuge(entry); | |
929 | if (pgtable) | |
930 | pgtable_trans_huge_deposit(mm, pmd, pgtable); | |
931 | set_pmd_at(mm, haddr, pmd, entry); | |
932 | atomic_long_inc(&mm->nr_ptes); | |
933 | return true; | |
934 | } | |
935 | ||
936 | int do_huge_pmd_anonymous_page(struct fault_env *fe) | |
937 | { | |
938 | struct vm_area_struct *vma = fe->vma; | |
939 | gfp_t gfp; | |
940 | struct page *page; | |
941 | unsigned long haddr = fe->address & HPAGE_PMD_MASK; | |
942 | ||
943 | if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end) | |
944 | return VM_FAULT_FALLBACK; | |
945 | if (unlikely(anon_vma_prepare(vma))) | |
946 | return VM_FAULT_OOM; | |
947 | if (unlikely(khugepaged_enter(vma, vma->vm_flags))) | |
948 | return VM_FAULT_OOM; | |
949 | if (!(fe->flags & FAULT_FLAG_WRITE) && | |
950 | !mm_forbids_zeropage(vma->vm_mm) && | |
951 | transparent_hugepage_use_zero_page()) { | |
952 | pgtable_t pgtable; | |
953 | struct page *zero_page; | |
954 | bool set; | |
955 | int ret; | |
956 | pgtable = pte_alloc_one(vma->vm_mm, haddr); | |
957 | if (unlikely(!pgtable)) | |
958 | return VM_FAULT_OOM; | |
959 | zero_page = get_huge_zero_page(); | |
960 | if (unlikely(!zero_page)) { | |
961 | pte_free(vma->vm_mm, pgtable); | |
962 | count_vm_event(THP_FAULT_FALLBACK); | |
963 | return VM_FAULT_FALLBACK; | |
964 | } | |
965 | fe->ptl = pmd_lock(vma->vm_mm, fe->pmd); | |
966 | ret = 0; | |
967 | set = false; | |
968 | if (pmd_none(*fe->pmd)) { | |
969 | if (userfaultfd_missing(vma)) { | |
970 | spin_unlock(fe->ptl); | |
971 | ret = handle_userfault(fe, VM_UFFD_MISSING); | |
972 | VM_BUG_ON(ret & VM_FAULT_FALLBACK); | |
973 | } else { | |
974 | set_huge_zero_page(pgtable, vma->vm_mm, vma, | |
975 | haddr, fe->pmd, zero_page); | |
976 | spin_unlock(fe->ptl); | |
977 | set = true; | |
978 | } | |
979 | } else | |
980 | spin_unlock(fe->ptl); | |
981 | if (!set) { | |
982 | pte_free(vma->vm_mm, pgtable); | |
983 | put_huge_zero_page(); | |
984 | } | |
985 | return ret; | |
986 | } | |
987 | gfp = alloc_hugepage_direct_gfpmask(vma); | |
988 | page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); | |
989 | if (unlikely(!page)) { | |
990 | count_vm_event(THP_FAULT_FALLBACK); | |
991 | return VM_FAULT_FALLBACK; | |
992 | } | |
993 | prep_transhuge_page(page); | |
994 | return __do_huge_pmd_anonymous_page(fe, page, gfp); | |
995 | } | |
996 | ||
997 | static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, | |
998 | pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write) | |
999 | { | |
1000 | struct mm_struct *mm = vma->vm_mm; | |
1001 | pmd_t entry; | |
1002 | spinlock_t *ptl; | |
1003 | ||
1004 | ptl = pmd_lock(mm, pmd); | |
1005 | entry = pmd_mkhuge(pfn_t_pmd(pfn, prot)); | |
1006 | if (pfn_t_devmap(pfn)) | |
1007 | entry = pmd_mkdevmap(entry); | |
1008 | if (write) { | |
1009 | entry = pmd_mkyoung(pmd_mkdirty(entry)); | |
1010 | entry = maybe_pmd_mkwrite(entry, vma); | |
1011 | } | |
1012 | set_pmd_at(mm, addr, pmd, entry); | |
1013 | update_mmu_cache_pmd(vma, addr, pmd); | |
1014 | spin_unlock(ptl); | |
1015 | } | |
1016 | ||
1017 | int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, | |
1018 | pmd_t *pmd, pfn_t pfn, bool write) | |
1019 | { | |
1020 | pgprot_t pgprot = vma->vm_page_prot; | |
1021 | /* | |
1022 | * If we had pmd_special, we could avoid all these restrictions, | |
1023 | * but we need to be consistent with PTEs and architectures that | |
1024 | * can't support a 'special' bit. | |
1025 | */ | |
1026 | BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); | |
1027 | BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == | |
1028 | (VM_PFNMAP|VM_MIXEDMAP)); | |
1029 | BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); | |
1030 | BUG_ON(!pfn_t_devmap(pfn)); | |
1031 | ||
1032 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
1033 | return VM_FAULT_SIGBUS; | |
1034 | if (track_pfn_insert(vma, &pgprot, pfn)) | |
1035 | return VM_FAULT_SIGBUS; | |
1036 | insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write); | |
1037 | return VM_FAULT_NOPAGE; | |
1038 | } | |
1039 | EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd); | |
1040 | ||
1041 | static void touch_pmd(struct vm_area_struct *vma, unsigned long addr, | |
1042 | pmd_t *pmd) | |
1043 | { | |
1044 | pmd_t _pmd; | |
1045 | ||
1046 | /* | |
1047 | * We should set the dirty bit only for FOLL_WRITE but for now | |
1048 | * the dirty bit in the pmd is meaningless. And if the dirty | |
1049 | * bit will become meaningful and we'll only set it with | |
1050 | * FOLL_WRITE, an atomic set_bit will be required on the pmd to | |
1051 | * set the young bit, instead of the current set_pmd_at. | |
1052 | */ | |
1053 | _pmd = pmd_mkyoung(pmd_mkdirty(*pmd)); | |
1054 | if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, | |
1055 | pmd, _pmd, 1)) | |
1056 | update_mmu_cache_pmd(vma, addr, pmd); | |
1057 | } | |
1058 | ||
1059 | struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr, | |
1060 | pmd_t *pmd, int flags) | |
1061 | { | |
1062 | unsigned long pfn = pmd_pfn(*pmd); | |
1063 | struct mm_struct *mm = vma->vm_mm; | |
1064 | struct dev_pagemap *pgmap; | |
1065 | struct page *page; | |
1066 | ||
1067 | assert_spin_locked(pmd_lockptr(mm, pmd)); | |
1068 | ||
1069 | if (flags & FOLL_WRITE && !pmd_write(*pmd)) | |
1070 | return NULL; | |
1071 | ||
1072 | if (pmd_present(*pmd) && pmd_devmap(*pmd)) | |
1073 | /* pass */; | |
1074 | else | |
1075 | return NULL; | |
1076 | ||
1077 | if (flags & FOLL_TOUCH) | |
1078 | touch_pmd(vma, addr, pmd); | |
1079 | ||
1080 | /* | |
1081 | * device mapped pages can only be returned if the | |
1082 | * caller will manage the page reference count. | |
1083 | */ | |
1084 | if (!(flags & FOLL_GET)) | |
1085 | return ERR_PTR(-EEXIST); | |
1086 | ||
1087 | pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT; | |
1088 | pgmap = get_dev_pagemap(pfn, NULL); | |
1089 | if (!pgmap) | |
1090 | return ERR_PTR(-EFAULT); | |
1091 | page = pfn_to_page(pfn); | |
1092 | get_page(page); | |
1093 | put_dev_pagemap(pgmap); | |
1094 | ||
1095 | return page; | |
1096 | } | |
1097 | ||
1098 | int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, | |
1099 | pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, | |
1100 | struct vm_area_struct *vma) | |
1101 | { | |
1102 | spinlock_t *dst_ptl, *src_ptl; | |
1103 | struct page *src_page; | |
1104 | pmd_t pmd; | |
1105 | pgtable_t pgtable = NULL; | |
1106 | int ret; | |
1107 | ||
1108 | if (!vma_is_dax(vma)) { | |
1109 | ret = -ENOMEM; | |
1110 | pgtable = pte_alloc_one(dst_mm, addr); | |
1111 | if (unlikely(!pgtable)) | |
1112 | goto out; | |
1113 | } | |
1114 | ||
1115 | dst_ptl = pmd_lock(dst_mm, dst_pmd); | |
1116 | src_ptl = pmd_lockptr(src_mm, src_pmd); | |
1117 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
1118 | ||
1119 | ret = -EAGAIN; | |
1120 | pmd = *src_pmd; | |
1121 | if (unlikely(!pmd_trans_huge(pmd) && !pmd_devmap(pmd))) { | |
1122 | pte_free(dst_mm, pgtable); | |
1123 | goto out_unlock; | |
1124 | } | |
1125 | /* | |
1126 | * When page table lock is held, the huge zero pmd should not be | |
1127 | * under splitting since we don't split the page itself, only pmd to | |
1128 | * a page table. | |
1129 | */ | |
1130 | if (is_huge_zero_pmd(pmd)) { | |
1131 | struct page *zero_page; | |
1132 | /* | |
1133 | * get_huge_zero_page() will never allocate a new page here, | |
1134 | * since we already have a zero page to copy. It just takes a | |
1135 | * reference. | |
1136 | */ | |
1137 | zero_page = get_huge_zero_page(); | |
1138 | set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, | |
1139 | zero_page); | |
1140 | ret = 0; | |
1141 | goto out_unlock; | |
1142 | } | |
1143 | ||
1144 | if (!vma_is_dax(vma)) { | |
1145 | /* thp accounting separate from pmd_devmap accounting */ | |
1146 | src_page = pmd_page(pmd); | |
1147 | VM_BUG_ON_PAGE(!PageHead(src_page), src_page); | |
1148 | get_page(src_page); | |
1149 | page_dup_rmap(src_page, true); | |
1150 | add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); | |
1151 | atomic_long_inc(&dst_mm->nr_ptes); | |
1152 | pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); | |
1153 | } | |
1154 | ||
1155 | pmdp_set_wrprotect(src_mm, addr, src_pmd); | |
1156 | pmd = pmd_mkold(pmd_wrprotect(pmd)); | |
1157 | set_pmd_at(dst_mm, addr, dst_pmd, pmd); | |
1158 | ||
1159 | ret = 0; | |
1160 | out_unlock: | |
1161 | spin_unlock(src_ptl); | |
1162 | spin_unlock(dst_ptl); | |
1163 | out: | |
1164 | return ret; | |
1165 | } | |
1166 | ||
1167 | void huge_pmd_set_accessed(struct fault_env *fe, pmd_t orig_pmd) | |
1168 | { | |
1169 | pmd_t entry; | |
1170 | unsigned long haddr; | |
1171 | ||
1172 | fe->ptl = pmd_lock(fe->vma->vm_mm, fe->pmd); | |
1173 | if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) | |
1174 | goto unlock; | |
1175 | ||
1176 | entry = pmd_mkyoung(orig_pmd); | |
1177 | haddr = fe->address & HPAGE_PMD_MASK; | |
1178 | if (pmdp_set_access_flags(fe->vma, haddr, fe->pmd, entry, | |
1179 | fe->flags & FAULT_FLAG_WRITE)) | |
1180 | update_mmu_cache_pmd(fe->vma, fe->address, fe->pmd); | |
1181 | ||
1182 | unlock: | |
1183 | spin_unlock(fe->ptl); | |
1184 | } | |
1185 | ||
1186 | static int do_huge_pmd_wp_page_fallback(struct fault_env *fe, pmd_t orig_pmd, | |
1187 | struct page *page) | |
1188 | { | |
1189 | struct vm_area_struct *vma = fe->vma; | |
1190 | unsigned long haddr = fe->address & HPAGE_PMD_MASK; | |
1191 | struct mem_cgroup *memcg; | |
1192 | pgtable_t pgtable; | |
1193 | pmd_t _pmd; | |
1194 | int ret = 0, i; | |
1195 | struct page **pages; | |
1196 | unsigned long mmun_start; /* For mmu_notifiers */ | |
1197 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1198 | ||
1199 | pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR, | |
1200 | GFP_KERNEL); | |
1201 | if (unlikely(!pages)) { | |
1202 | ret |= VM_FAULT_OOM; | |
1203 | goto out; | |
1204 | } | |
1205 | ||
1206 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
1207 | pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE | | |
1208 | __GFP_OTHER_NODE, vma, | |
1209 | fe->address, page_to_nid(page)); | |
1210 | if (unlikely(!pages[i] || | |
1211 | mem_cgroup_try_charge(pages[i], vma->vm_mm, | |
1212 | GFP_KERNEL, &memcg, false))) { | |
1213 | if (pages[i]) | |
1214 | put_page(pages[i]); | |
1215 | while (--i >= 0) { | |
1216 | memcg = (void *)page_private(pages[i]); | |
1217 | set_page_private(pages[i], 0); | |
1218 | mem_cgroup_cancel_charge(pages[i], memcg, | |
1219 | false); | |
1220 | put_page(pages[i]); | |
1221 | } | |
1222 | kfree(pages); | |
1223 | ret |= VM_FAULT_OOM; | |
1224 | goto out; | |
1225 | } | |
1226 | set_page_private(pages[i], (unsigned long)memcg); | |
1227 | } | |
1228 | ||
1229 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
1230 | copy_user_highpage(pages[i], page + i, | |
1231 | haddr + PAGE_SIZE * i, vma); | |
1232 | __SetPageUptodate(pages[i]); | |
1233 | cond_resched(); | |
1234 | } | |
1235 | ||
1236 | mmun_start = haddr; | |
1237 | mmun_end = haddr + HPAGE_PMD_SIZE; | |
1238 | mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end); | |
1239 | ||
1240 | fe->ptl = pmd_lock(vma->vm_mm, fe->pmd); | |
1241 | if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) | |
1242 | goto out_free_pages; | |
1243 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
1244 | ||
1245 | pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd); | |
1246 | /* leave pmd empty until pte is filled */ | |
1247 | ||
1248 | pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, fe->pmd); | |
1249 | pmd_populate(vma->vm_mm, &_pmd, pgtable); | |
1250 | ||
1251 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
1252 | pte_t entry; | |
1253 | entry = mk_pte(pages[i], vma->vm_page_prot); | |
1254 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); | |
1255 | memcg = (void *)page_private(pages[i]); | |
1256 | set_page_private(pages[i], 0); | |
1257 | page_add_new_anon_rmap(pages[i], fe->vma, haddr, false); | |
1258 | mem_cgroup_commit_charge(pages[i], memcg, false, false); | |
1259 | lru_cache_add_active_or_unevictable(pages[i], vma); | |
1260 | fe->pte = pte_offset_map(&_pmd, haddr); | |
1261 | VM_BUG_ON(!pte_none(*fe->pte)); | |
1262 | set_pte_at(vma->vm_mm, haddr, fe->pte, entry); | |
1263 | pte_unmap(fe->pte); | |
1264 | } | |
1265 | kfree(pages); | |
1266 | ||
1267 | smp_wmb(); /* make pte visible before pmd */ | |
1268 | pmd_populate(vma->vm_mm, fe->pmd, pgtable); | |
1269 | page_remove_rmap(page, true); | |
1270 | spin_unlock(fe->ptl); | |
1271 | ||
1272 | mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); | |
1273 | ||
1274 | ret |= VM_FAULT_WRITE; | |
1275 | put_page(page); | |
1276 | ||
1277 | out: | |
1278 | return ret; | |
1279 | ||
1280 | out_free_pages: | |
1281 | spin_unlock(fe->ptl); | |
1282 | mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); | |
1283 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
1284 | memcg = (void *)page_private(pages[i]); | |
1285 | set_page_private(pages[i], 0); | |
1286 | mem_cgroup_cancel_charge(pages[i], memcg, false); | |
1287 | put_page(pages[i]); | |
1288 | } | |
1289 | kfree(pages); | |
1290 | goto out; | |
1291 | } | |
1292 | ||
1293 | int do_huge_pmd_wp_page(struct fault_env *fe, pmd_t orig_pmd) | |
1294 | { | |
1295 | struct vm_area_struct *vma = fe->vma; | |
1296 | struct page *page = NULL, *new_page; | |
1297 | struct mem_cgroup *memcg; | |
1298 | unsigned long haddr = fe->address & HPAGE_PMD_MASK; | |
1299 | unsigned long mmun_start; /* For mmu_notifiers */ | |
1300 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1301 | gfp_t huge_gfp; /* for allocation and charge */ | |
1302 | int ret = 0; | |
1303 | ||
1304 | fe->ptl = pmd_lockptr(vma->vm_mm, fe->pmd); | |
1305 | VM_BUG_ON_VMA(!vma->anon_vma, vma); | |
1306 | if (is_huge_zero_pmd(orig_pmd)) | |
1307 | goto alloc; | |
1308 | spin_lock(fe->ptl); | |
1309 | if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) | |
1310 | goto out_unlock; | |
1311 | ||
1312 | page = pmd_page(orig_pmd); | |
1313 | VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); | |
1314 | /* | |
1315 | * We can only reuse the page if nobody else maps the huge page or it's | |
1316 | * part. | |
1317 | */ | |
1318 | if (page_trans_huge_mapcount(page, NULL) == 1) { | |
1319 | pmd_t entry; | |
1320 | entry = pmd_mkyoung(orig_pmd); | |
1321 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
1322 | if (pmdp_set_access_flags(vma, haddr, fe->pmd, entry, 1)) | |
1323 | update_mmu_cache_pmd(vma, fe->address, fe->pmd); | |
1324 | ret |= VM_FAULT_WRITE; | |
1325 | goto out_unlock; | |
1326 | } | |
1327 | get_page(page); | |
1328 | spin_unlock(fe->ptl); | |
1329 | alloc: | |
1330 | if (transparent_hugepage_enabled(vma) && | |
1331 | !transparent_hugepage_debug_cow()) { | |
1332 | huge_gfp = alloc_hugepage_direct_gfpmask(vma); | |
1333 | new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER); | |
1334 | } else | |
1335 | new_page = NULL; | |
1336 | ||
1337 | if (likely(new_page)) { | |
1338 | prep_transhuge_page(new_page); | |
1339 | } else { | |
1340 | if (!page) { | |
1341 | split_huge_pmd(vma, fe->pmd, fe->address); | |
1342 | ret |= VM_FAULT_FALLBACK; | |
1343 | } else { | |
1344 | ret = do_huge_pmd_wp_page_fallback(fe, orig_pmd, page); | |
1345 | if (ret & VM_FAULT_OOM) { | |
1346 | split_huge_pmd(vma, fe->pmd, fe->address); | |
1347 | ret |= VM_FAULT_FALLBACK; | |
1348 | } | |
1349 | put_page(page); | |
1350 | } | |
1351 | count_vm_event(THP_FAULT_FALLBACK); | |
1352 | goto out; | |
1353 | } | |
1354 | ||
1355 | if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm, | |
1356 | huge_gfp, &memcg, true))) { | |
1357 | put_page(new_page); | |
1358 | split_huge_pmd(vma, fe->pmd, fe->address); | |
1359 | if (page) | |
1360 | put_page(page); | |
1361 | ret |= VM_FAULT_FALLBACK; | |
1362 | count_vm_event(THP_FAULT_FALLBACK); | |
1363 | goto out; | |
1364 | } | |
1365 | ||
1366 | count_vm_event(THP_FAULT_ALLOC); | |
1367 | ||
1368 | if (!page) | |
1369 | clear_huge_page(new_page, haddr, HPAGE_PMD_NR); | |
1370 | else | |
1371 | copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR); | |
1372 | __SetPageUptodate(new_page); | |
1373 | ||
1374 | mmun_start = haddr; | |
1375 | mmun_end = haddr + HPAGE_PMD_SIZE; | |
1376 | mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end); | |
1377 | ||
1378 | spin_lock(fe->ptl); | |
1379 | if (page) | |
1380 | put_page(page); | |
1381 | if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) { | |
1382 | spin_unlock(fe->ptl); | |
1383 | mem_cgroup_cancel_charge(new_page, memcg, true); | |
1384 | put_page(new_page); | |
1385 | goto out_mn; | |
1386 | } else { | |
1387 | pmd_t entry; | |
1388 | entry = mk_huge_pmd(new_page, vma->vm_page_prot); | |
1389 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
1390 | pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd); | |
1391 | page_add_new_anon_rmap(new_page, vma, haddr, true); | |
1392 | mem_cgroup_commit_charge(new_page, memcg, false, true); | |
1393 | lru_cache_add_active_or_unevictable(new_page, vma); | |
1394 | set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry); | |
1395 | update_mmu_cache_pmd(vma, fe->address, fe->pmd); | |
1396 | if (!page) { | |
1397 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); | |
1398 | put_huge_zero_page(); | |
1399 | } else { | |
1400 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
1401 | page_remove_rmap(page, true); | |
1402 | put_page(page); | |
1403 | } | |
1404 | ret |= VM_FAULT_WRITE; | |
1405 | } | |
1406 | spin_unlock(fe->ptl); | |
1407 | out_mn: | |
1408 | mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); | |
1409 | out: | |
1410 | return ret; | |
1411 | out_unlock: | |
1412 | spin_unlock(fe->ptl); | |
1413 | return ret; | |
1414 | } | |
1415 | ||
1416 | struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, | |
1417 | unsigned long addr, | |
1418 | pmd_t *pmd, | |
1419 | unsigned int flags) | |
1420 | { | |
1421 | struct mm_struct *mm = vma->vm_mm; | |
1422 | struct page *page = NULL; | |
1423 | ||
1424 | assert_spin_locked(pmd_lockptr(mm, pmd)); | |
1425 | ||
1426 | if (flags & FOLL_WRITE && !pmd_write(*pmd)) | |
1427 | goto out; | |
1428 | ||
1429 | /* Avoid dumping huge zero page */ | |
1430 | if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) | |
1431 | return ERR_PTR(-EFAULT); | |
1432 | ||
1433 | /* Full NUMA hinting faults to serialise migration in fault paths */ | |
1434 | if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) | |
1435 | goto out; | |
1436 | ||
1437 | page = pmd_page(*pmd); | |
1438 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
1439 | if (flags & FOLL_TOUCH) | |
1440 | touch_pmd(vma, addr, pmd); | |
1441 | if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { | |
1442 | /* | |
1443 | * We don't mlock() pte-mapped THPs. This way we can avoid | |
1444 | * leaking mlocked pages into non-VM_LOCKED VMAs. | |
1445 | * | |
1446 | * In most cases the pmd is the only mapping of the page as we | |
1447 | * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for | |
1448 | * writable private mappings in populate_vma_page_range(). | |
1449 | * | |
1450 | * The only scenario when we have the page shared here is if we | |
1451 | * mlocking read-only mapping shared over fork(). We skip | |
1452 | * mlocking such pages. | |
1453 | */ | |
1454 | if (compound_mapcount(page) == 1 && !PageDoubleMap(page) && | |
1455 | page->mapping && trylock_page(page)) { | |
1456 | lru_add_drain(); | |
1457 | if (page->mapping) | |
1458 | mlock_vma_page(page); | |
1459 | unlock_page(page); | |
1460 | } | |
1461 | } | |
1462 | page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; | |
1463 | VM_BUG_ON_PAGE(!PageCompound(page), page); | |
1464 | if (flags & FOLL_GET) | |
1465 | get_page(page); | |
1466 | ||
1467 | out: | |
1468 | return page; | |
1469 | } | |
1470 | ||
1471 | /* NUMA hinting page fault entry point for trans huge pmds */ | |
1472 | int do_huge_pmd_numa_page(struct fault_env *fe, pmd_t pmd) | |
1473 | { | |
1474 | struct vm_area_struct *vma = fe->vma; | |
1475 | struct anon_vma *anon_vma = NULL; | |
1476 | struct page *page; | |
1477 | unsigned long haddr = fe->address & HPAGE_PMD_MASK; | |
1478 | int page_nid = -1, this_nid = numa_node_id(); | |
1479 | int target_nid, last_cpupid = -1; | |
1480 | bool page_locked; | |
1481 | bool migrated = false; | |
1482 | bool was_writable; | |
1483 | int flags = 0; | |
1484 | ||
1485 | /* A PROT_NONE fault should not end up here */ | |
1486 | BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))); | |
1487 | ||
1488 | fe->ptl = pmd_lock(vma->vm_mm, fe->pmd); | |
1489 | if (unlikely(!pmd_same(pmd, *fe->pmd))) | |
1490 | goto out_unlock; | |
1491 | ||
1492 | /* | |
1493 | * If there are potential migrations, wait for completion and retry | |
1494 | * without disrupting NUMA hinting information. Do not relock and | |
1495 | * check_same as the page may no longer be mapped. | |
1496 | */ | |
1497 | if (unlikely(pmd_trans_migrating(*fe->pmd))) { | |
1498 | page = pmd_page(*fe->pmd); | |
1499 | spin_unlock(fe->ptl); | |
1500 | wait_on_page_locked(page); | |
1501 | goto out; | |
1502 | } | |
1503 | ||
1504 | page = pmd_page(pmd); | |
1505 | BUG_ON(is_huge_zero_page(page)); | |
1506 | page_nid = page_to_nid(page); | |
1507 | last_cpupid = page_cpupid_last(page); | |
1508 | count_vm_numa_event(NUMA_HINT_FAULTS); | |
1509 | if (page_nid == this_nid) { | |
1510 | count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); | |
1511 | flags |= TNF_FAULT_LOCAL; | |
1512 | } | |
1513 | ||
1514 | /* See similar comment in do_numa_page for explanation */ | |
1515 | if (!(vma->vm_flags & VM_WRITE)) | |
1516 | flags |= TNF_NO_GROUP; | |
1517 | ||
1518 | /* | |
1519 | * Acquire the page lock to serialise THP migrations but avoid dropping | |
1520 | * page_table_lock if at all possible | |
1521 | */ | |
1522 | page_locked = trylock_page(page); | |
1523 | target_nid = mpol_misplaced(page, vma, haddr); | |
1524 | if (target_nid == -1) { | |
1525 | /* If the page was locked, there are no parallel migrations */ | |
1526 | if (page_locked) | |
1527 | goto clear_pmdnuma; | |
1528 | } | |
1529 | ||
1530 | /* Migration could have started since the pmd_trans_migrating check */ | |
1531 | if (!page_locked) { | |
1532 | spin_unlock(fe->ptl); | |
1533 | wait_on_page_locked(page); | |
1534 | page_nid = -1; | |
1535 | goto out; | |
1536 | } | |
1537 | ||
1538 | /* | |
1539 | * Page is misplaced. Page lock serialises migrations. Acquire anon_vma | |
1540 | * to serialises splits | |
1541 | */ | |
1542 | get_page(page); | |
1543 | spin_unlock(fe->ptl); | |
1544 | anon_vma = page_lock_anon_vma_read(page); | |
1545 | ||
1546 | /* Confirm the PMD did not change while page_table_lock was released */ | |
1547 | spin_lock(fe->ptl); | |
1548 | if (unlikely(!pmd_same(pmd, *fe->pmd))) { | |
1549 | unlock_page(page); | |
1550 | put_page(page); | |
1551 | page_nid = -1; | |
1552 | goto out_unlock; | |
1553 | } | |
1554 | ||
1555 | /* Bail if we fail to protect against THP splits for any reason */ | |
1556 | if (unlikely(!anon_vma)) { | |
1557 | put_page(page); | |
1558 | page_nid = -1; | |
1559 | goto clear_pmdnuma; | |
1560 | } | |
1561 | ||
1562 | /* | |
1563 | * Migrate the THP to the requested node, returns with page unlocked | |
1564 | * and access rights restored. | |
1565 | */ | |
1566 | spin_unlock(fe->ptl); | |
1567 | migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma, | |
1568 | fe->pmd, pmd, fe->address, page, target_nid); | |
1569 | if (migrated) { | |
1570 | flags |= TNF_MIGRATED; | |
1571 | page_nid = target_nid; | |
1572 | } else | |
1573 | flags |= TNF_MIGRATE_FAIL; | |
1574 | ||
1575 | goto out; | |
1576 | clear_pmdnuma: | |
1577 | BUG_ON(!PageLocked(page)); | |
1578 | was_writable = pmd_write(pmd); | |
1579 | pmd = pmd_modify(pmd, vma->vm_page_prot); | |
1580 | pmd = pmd_mkyoung(pmd); | |
1581 | if (was_writable) | |
1582 | pmd = pmd_mkwrite(pmd); | |
1583 | set_pmd_at(vma->vm_mm, haddr, fe->pmd, pmd); | |
1584 | update_mmu_cache_pmd(vma, fe->address, fe->pmd); | |
1585 | unlock_page(page); | |
1586 | out_unlock: | |
1587 | spin_unlock(fe->ptl); | |
1588 | ||
1589 | out: | |
1590 | if (anon_vma) | |
1591 | page_unlock_anon_vma_read(anon_vma); | |
1592 | ||
1593 | if (page_nid != -1) | |
1594 | task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, fe->flags); | |
1595 | ||
1596 | return 0; | |
1597 | } | |
1598 | ||
1599 | int madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, | |
1600 | pmd_t *pmd, unsigned long addr, unsigned long next) | |
1601 | ||
1602 | { | |
1603 | spinlock_t *ptl; | |
1604 | pmd_t orig_pmd; | |
1605 | struct page *page; | |
1606 | struct mm_struct *mm = tlb->mm; | |
1607 | int ret = 0; | |
1608 | ||
1609 | ptl = pmd_trans_huge_lock(pmd, vma); | |
1610 | if (!ptl) | |
1611 | goto out_unlocked; | |
1612 | ||
1613 | orig_pmd = *pmd; | |
1614 | if (is_huge_zero_pmd(orig_pmd)) { | |
1615 | ret = 1; | |
1616 | goto out; | |
1617 | } | |
1618 | ||
1619 | page = pmd_page(orig_pmd); | |
1620 | /* | |
1621 | * If other processes are mapping this page, we couldn't discard | |
1622 | * the page unless they all do MADV_FREE so let's skip the page. | |
1623 | */ | |
1624 | if (page_mapcount(page) != 1) | |
1625 | goto out; | |
1626 | ||
1627 | if (!trylock_page(page)) | |
1628 | goto out; | |
1629 | ||
1630 | /* | |
1631 | * If user want to discard part-pages of THP, split it so MADV_FREE | |
1632 | * will deactivate only them. | |
1633 | */ | |
1634 | if (next - addr != HPAGE_PMD_SIZE) { | |
1635 | get_page(page); | |
1636 | spin_unlock(ptl); | |
1637 | split_huge_page(page); | |
1638 | put_page(page); | |
1639 | unlock_page(page); | |
1640 | goto out_unlocked; | |
1641 | } | |
1642 | ||
1643 | if (PageDirty(page)) | |
1644 | ClearPageDirty(page); | |
1645 | unlock_page(page); | |
1646 | ||
1647 | if (PageActive(page)) | |
1648 | deactivate_page(page); | |
1649 | ||
1650 | if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { | |
1651 | orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd, | |
1652 | tlb->fullmm); | |
1653 | orig_pmd = pmd_mkold(orig_pmd); | |
1654 | orig_pmd = pmd_mkclean(orig_pmd); | |
1655 | ||
1656 | set_pmd_at(mm, addr, pmd, orig_pmd); | |
1657 | tlb_remove_pmd_tlb_entry(tlb, pmd, addr); | |
1658 | } | |
1659 | ret = 1; | |
1660 | out: | |
1661 | spin_unlock(ptl); | |
1662 | out_unlocked: | |
1663 | return ret; | |
1664 | } | |
1665 | ||
1666 | int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, | |
1667 | pmd_t *pmd, unsigned long addr) | |
1668 | { | |
1669 | pmd_t orig_pmd; | |
1670 | spinlock_t *ptl; | |
1671 | ||
1672 | ptl = __pmd_trans_huge_lock(pmd, vma); | |
1673 | if (!ptl) | |
1674 | return 0; | |
1675 | /* | |
1676 | * For architectures like ppc64 we look at deposited pgtable | |
1677 | * when calling pmdp_huge_get_and_clear. So do the | |
1678 | * pgtable_trans_huge_withdraw after finishing pmdp related | |
1679 | * operations. | |
1680 | */ | |
1681 | orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd, | |
1682 | tlb->fullmm); | |
1683 | tlb_remove_pmd_tlb_entry(tlb, pmd, addr); | |
1684 | if (vma_is_dax(vma)) { | |
1685 | spin_unlock(ptl); | |
1686 | if (is_huge_zero_pmd(orig_pmd)) | |
1687 | tlb_remove_page(tlb, pmd_page(orig_pmd)); | |
1688 | } else if (is_huge_zero_pmd(orig_pmd)) { | |
1689 | pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd)); | |
1690 | atomic_long_dec(&tlb->mm->nr_ptes); | |
1691 | spin_unlock(ptl); | |
1692 | tlb_remove_page(tlb, pmd_page(orig_pmd)); | |
1693 | } else { | |
1694 | struct page *page = pmd_page(orig_pmd); | |
1695 | page_remove_rmap(page, true); | |
1696 | VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); | |
1697 | add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); | |
1698 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
1699 | pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd)); | |
1700 | atomic_long_dec(&tlb->mm->nr_ptes); | |
1701 | spin_unlock(ptl); | |
1702 | tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE); | |
1703 | } | |
1704 | return 1; | |
1705 | } | |
1706 | ||
1707 | bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, | |
1708 | unsigned long new_addr, unsigned long old_end, | |
1709 | pmd_t *old_pmd, pmd_t *new_pmd) | |
1710 | { | |
1711 | spinlock_t *old_ptl, *new_ptl; | |
1712 | pmd_t pmd; | |
1713 | struct mm_struct *mm = vma->vm_mm; | |
1714 | ||
1715 | if ((old_addr & ~HPAGE_PMD_MASK) || | |
1716 | (new_addr & ~HPAGE_PMD_MASK) || | |
1717 | old_end - old_addr < HPAGE_PMD_SIZE) | |
1718 | return false; | |
1719 | ||
1720 | /* | |
1721 | * The destination pmd shouldn't be established, free_pgtables() | |
1722 | * should have release it. | |
1723 | */ | |
1724 | if (WARN_ON(!pmd_none(*new_pmd))) { | |
1725 | VM_BUG_ON(pmd_trans_huge(*new_pmd)); | |
1726 | return false; | |
1727 | } | |
1728 | ||
1729 | /* | |
1730 | * We don't have to worry about the ordering of src and dst | |
1731 | * ptlocks because exclusive mmap_sem prevents deadlock. | |
1732 | */ | |
1733 | old_ptl = __pmd_trans_huge_lock(old_pmd, vma); | |
1734 | if (old_ptl) { | |
1735 | new_ptl = pmd_lockptr(mm, new_pmd); | |
1736 | if (new_ptl != old_ptl) | |
1737 | spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); | |
1738 | pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); | |
1739 | VM_BUG_ON(!pmd_none(*new_pmd)); | |
1740 | ||
1741 | if (pmd_move_must_withdraw(new_ptl, old_ptl) && | |
1742 | vma_is_anonymous(vma)) { | |
1743 | pgtable_t pgtable; | |
1744 | pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); | |
1745 | pgtable_trans_huge_deposit(mm, new_pmd, pgtable); | |
1746 | } | |
1747 | set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd)); | |
1748 | if (new_ptl != old_ptl) | |
1749 | spin_unlock(new_ptl); | |
1750 | spin_unlock(old_ptl); | |
1751 | return true; | |
1752 | } | |
1753 | return false; | |
1754 | } | |
1755 | ||
1756 | /* | |
1757 | * Returns | |
1758 | * - 0 if PMD could not be locked | |
1759 | * - 1 if PMD was locked but protections unchange and TLB flush unnecessary | |
1760 | * - HPAGE_PMD_NR is protections changed and TLB flush necessary | |
1761 | */ | |
1762 | int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, | |
1763 | unsigned long addr, pgprot_t newprot, int prot_numa) | |
1764 | { | |
1765 | struct mm_struct *mm = vma->vm_mm; | |
1766 | spinlock_t *ptl; | |
1767 | int ret = 0; | |
1768 | ||
1769 | ptl = __pmd_trans_huge_lock(pmd, vma); | |
1770 | if (ptl) { | |
1771 | pmd_t entry; | |
1772 | bool preserve_write = prot_numa && pmd_write(*pmd); | |
1773 | ret = 1; | |
1774 | ||
1775 | /* | |
1776 | * Avoid trapping faults against the zero page. The read-only | |
1777 | * data is likely to be read-cached on the local CPU and | |
1778 | * local/remote hits to the zero page are not interesting. | |
1779 | */ | |
1780 | if (prot_numa && is_huge_zero_pmd(*pmd)) { | |
1781 | spin_unlock(ptl); | |
1782 | return ret; | |
1783 | } | |
1784 | ||
1785 | if (!prot_numa || !pmd_protnone(*pmd)) { | |
1786 | entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd); | |
1787 | entry = pmd_modify(entry, newprot); | |
1788 | if (preserve_write) | |
1789 | entry = pmd_mkwrite(entry); | |
1790 | ret = HPAGE_PMD_NR; | |
1791 | set_pmd_at(mm, addr, pmd, entry); | |
1792 | BUG_ON(!preserve_write && pmd_write(entry)); | |
1793 | } | |
1794 | spin_unlock(ptl); | |
1795 | } | |
1796 | ||
1797 | return ret; | |
1798 | } | |
1799 | ||
1800 | /* | |
1801 | * Returns true if a given pmd maps a thp, false otherwise. | |
1802 | * | |
1803 | * Note that if it returns true, this routine returns without unlocking page | |
1804 | * table lock. So callers must unlock it. | |
1805 | */ | |
1806 | spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) | |
1807 | { | |
1808 | spinlock_t *ptl; | |
1809 | ptl = pmd_lock(vma->vm_mm, pmd); | |
1810 | if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd))) | |
1811 | return ptl; | |
1812 | spin_unlock(ptl); | |
1813 | return NULL; | |
1814 | } | |
1815 | ||
1816 | #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE) | |
1817 | ||
1818 | int hugepage_madvise(struct vm_area_struct *vma, | |
1819 | unsigned long *vm_flags, int advice) | |
1820 | { | |
1821 | switch (advice) { | |
1822 | case MADV_HUGEPAGE: | |
1823 | #ifdef CONFIG_S390 | |
1824 | /* | |
1825 | * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390 | |
1826 | * can't handle this properly after s390_enable_sie, so we simply | |
1827 | * ignore the madvise to prevent qemu from causing a SIGSEGV. | |
1828 | */ | |
1829 | if (mm_has_pgste(vma->vm_mm)) | |
1830 | return 0; | |
1831 | #endif | |
1832 | /* | |
1833 | * Be somewhat over-protective like KSM for now! | |
1834 | */ | |
1835 | if (*vm_flags & VM_NO_THP) | |
1836 | return -EINVAL; | |
1837 | *vm_flags &= ~VM_NOHUGEPAGE; | |
1838 | *vm_flags |= VM_HUGEPAGE; | |
1839 | /* | |
1840 | * If the vma become good for khugepaged to scan, | |
1841 | * register it here without waiting a page fault that | |
1842 | * may not happen any time soon. | |
1843 | */ | |
1844 | if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags))) | |
1845 | return -ENOMEM; | |
1846 | break; | |
1847 | case MADV_NOHUGEPAGE: | |
1848 | /* | |
1849 | * Be somewhat over-protective like KSM for now! | |
1850 | */ | |
1851 | if (*vm_flags & VM_NO_THP) | |
1852 | return -EINVAL; | |
1853 | *vm_flags &= ~VM_HUGEPAGE; | |
1854 | *vm_flags |= VM_NOHUGEPAGE; | |
1855 | /* | |
1856 | * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning | |
1857 | * this vma even if we leave the mm registered in khugepaged if | |
1858 | * it got registered before VM_NOHUGEPAGE was set. | |
1859 | */ | |
1860 | break; | |
1861 | } | |
1862 | ||
1863 | return 0; | |
1864 | } | |
1865 | ||
1866 | static int __init khugepaged_slab_init(void) | |
1867 | { | |
1868 | mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", | |
1869 | sizeof(struct mm_slot), | |
1870 | __alignof__(struct mm_slot), 0, NULL); | |
1871 | if (!mm_slot_cache) | |
1872 | return -ENOMEM; | |
1873 | ||
1874 | return 0; | |
1875 | } | |
1876 | ||
1877 | static void __init khugepaged_slab_exit(void) | |
1878 | { | |
1879 | kmem_cache_destroy(mm_slot_cache); | |
1880 | } | |
1881 | ||
1882 | static inline struct mm_slot *alloc_mm_slot(void) | |
1883 | { | |
1884 | if (!mm_slot_cache) /* initialization failed */ | |
1885 | return NULL; | |
1886 | return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); | |
1887 | } | |
1888 | ||
1889 | static inline void free_mm_slot(struct mm_slot *mm_slot) | |
1890 | { | |
1891 | kmem_cache_free(mm_slot_cache, mm_slot); | |
1892 | } | |
1893 | ||
1894 | static struct mm_slot *get_mm_slot(struct mm_struct *mm) | |
1895 | { | |
1896 | struct mm_slot *mm_slot; | |
1897 | ||
1898 | hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm) | |
1899 | if (mm == mm_slot->mm) | |
1900 | return mm_slot; | |
1901 | ||
1902 | return NULL; | |
1903 | } | |
1904 | ||
1905 | static void insert_to_mm_slots_hash(struct mm_struct *mm, | |
1906 | struct mm_slot *mm_slot) | |
1907 | { | |
1908 | mm_slot->mm = mm; | |
1909 | hash_add(mm_slots_hash, &mm_slot->hash, (long)mm); | |
1910 | } | |
1911 | ||
1912 | static inline int khugepaged_test_exit(struct mm_struct *mm) | |
1913 | { | |
1914 | return atomic_read(&mm->mm_users) == 0; | |
1915 | } | |
1916 | ||
1917 | int __khugepaged_enter(struct mm_struct *mm) | |
1918 | { | |
1919 | struct mm_slot *mm_slot; | |
1920 | int wakeup; | |
1921 | ||
1922 | mm_slot = alloc_mm_slot(); | |
1923 | if (!mm_slot) | |
1924 | return -ENOMEM; | |
1925 | ||
1926 | /* __khugepaged_exit() must not run from under us */ | |
1927 | VM_BUG_ON_MM(khugepaged_test_exit(mm), mm); | |
1928 | if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { | |
1929 | free_mm_slot(mm_slot); | |
1930 | return 0; | |
1931 | } | |
1932 | ||
1933 | spin_lock(&khugepaged_mm_lock); | |
1934 | insert_to_mm_slots_hash(mm, mm_slot); | |
1935 | /* | |
1936 | * Insert just behind the scanning cursor, to let the area settle | |
1937 | * down a little. | |
1938 | */ | |
1939 | wakeup = list_empty(&khugepaged_scan.mm_head); | |
1940 | list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head); | |
1941 | spin_unlock(&khugepaged_mm_lock); | |
1942 | ||
1943 | atomic_inc(&mm->mm_count); | |
1944 | if (wakeup) | |
1945 | wake_up_interruptible(&khugepaged_wait); | |
1946 | ||
1947 | return 0; | |
1948 | } | |
1949 | ||
1950 | int khugepaged_enter_vma_merge(struct vm_area_struct *vma, | |
1951 | unsigned long vm_flags) | |
1952 | { | |
1953 | unsigned long hstart, hend; | |
1954 | if (!vma->anon_vma) | |
1955 | /* | |
1956 | * Not yet faulted in so we will register later in the | |
1957 | * page fault if needed. | |
1958 | */ | |
1959 | return 0; | |
1960 | if (vma->vm_ops || (vm_flags & VM_NO_THP)) | |
1961 | /* khugepaged not yet working on file or special mappings */ | |
1962 | return 0; | |
1963 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; | |
1964 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
1965 | if (hstart < hend) | |
1966 | return khugepaged_enter(vma, vm_flags); | |
1967 | return 0; | |
1968 | } | |
1969 | ||
1970 | void __khugepaged_exit(struct mm_struct *mm) | |
1971 | { | |
1972 | struct mm_slot *mm_slot; | |
1973 | int free = 0; | |
1974 | ||
1975 | spin_lock(&khugepaged_mm_lock); | |
1976 | mm_slot = get_mm_slot(mm); | |
1977 | if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { | |
1978 | hash_del(&mm_slot->hash); | |
1979 | list_del(&mm_slot->mm_node); | |
1980 | free = 1; | |
1981 | } | |
1982 | spin_unlock(&khugepaged_mm_lock); | |
1983 | ||
1984 | if (free) { | |
1985 | clear_bit(MMF_VM_HUGEPAGE, &mm->flags); | |
1986 | free_mm_slot(mm_slot); | |
1987 | mmdrop(mm); | |
1988 | } else if (mm_slot) { | |
1989 | /* | |
1990 | * This is required to serialize against | |
1991 | * khugepaged_test_exit() (which is guaranteed to run | |
1992 | * under mmap sem read mode). Stop here (after we | |
1993 | * return all pagetables will be destroyed) until | |
1994 | * khugepaged has finished working on the pagetables | |
1995 | * under the mmap_sem. | |
1996 | */ | |
1997 | down_write(&mm->mmap_sem); | |
1998 | up_write(&mm->mmap_sem); | |
1999 | } | |
2000 | } | |
2001 | ||
2002 | static void release_pte_page(struct page *page) | |
2003 | { | |
2004 | /* 0 stands for page_is_file_cache(page) == false */ | |
2005 | dec_zone_page_state(page, NR_ISOLATED_ANON + 0); | |
2006 | unlock_page(page); | |
2007 | putback_lru_page(page); | |
2008 | } | |
2009 | ||
2010 | static void release_pte_pages(pte_t *pte, pte_t *_pte) | |
2011 | { | |
2012 | while (--_pte >= pte) { | |
2013 | pte_t pteval = *_pte; | |
2014 | if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval))) | |
2015 | release_pte_page(pte_page(pteval)); | |
2016 | } | |
2017 | } | |
2018 | ||
2019 | static int __collapse_huge_page_isolate(struct vm_area_struct *vma, | |
2020 | unsigned long address, | |
2021 | pte_t *pte) | |
2022 | { | |
2023 | struct page *page = NULL; | |
2024 | pte_t *_pte; | |
2025 | int none_or_zero = 0, result = 0; | |
2026 | bool referenced = false, writable = false; | |
2027 | ||
2028 | for (_pte = pte; _pte < pte+HPAGE_PMD_NR; | |
2029 | _pte++, address += PAGE_SIZE) { | |
2030 | pte_t pteval = *_pte; | |
2031 | if (pte_none(pteval) || (pte_present(pteval) && | |
2032 | is_zero_pfn(pte_pfn(pteval)))) { | |
2033 | if (!userfaultfd_armed(vma) && | |
2034 | ++none_or_zero <= khugepaged_max_ptes_none) { | |
2035 | continue; | |
2036 | } else { | |
2037 | result = SCAN_EXCEED_NONE_PTE; | |
2038 | goto out; | |
2039 | } | |
2040 | } | |
2041 | if (!pte_present(pteval)) { | |
2042 | result = SCAN_PTE_NON_PRESENT; | |
2043 | goto out; | |
2044 | } | |
2045 | page = vm_normal_page(vma, address, pteval); | |
2046 | if (unlikely(!page)) { | |
2047 | result = SCAN_PAGE_NULL; | |
2048 | goto out; | |
2049 | } | |
2050 | ||
2051 | VM_BUG_ON_PAGE(PageCompound(page), page); | |
2052 | VM_BUG_ON_PAGE(!PageAnon(page), page); | |
2053 | VM_BUG_ON_PAGE(!PageSwapBacked(page), page); | |
2054 | ||
2055 | /* | |
2056 | * We can do it before isolate_lru_page because the | |
2057 | * page can't be freed from under us. NOTE: PG_lock | |
2058 | * is needed to serialize against split_huge_page | |
2059 | * when invoked from the VM. | |
2060 | */ | |
2061 | if (!trylock_page(page)) { | |
2062 | result = SCAN_PAGE_LOCK; | |
2063 | goto out; | |
2064 | } | |
2065 | ||
2066 | /* | |
2067 | * cannot use mapcount: can't collapse if there's a gup pin. | |
2068 | * The page must only be referenced by the scanned process | |
2069 | * and page swap cache. | |
2070 | */ | |
2071 | if (page_count(page) != 1 + !!PageSwapCache(page)) { | |
2072 | unlock_page(page); | |
2073 | result = SCAN_PAGE_COUNT; | |
2074 | goto out; | |
2075 | } | |
2076 | if (pte_write(pteval)) { | |
2077 | writable = true; | |
2078 | } else { | |
2079 | if (PageSwapCache(page) && | |
2080 | !reuse_swap_page(page, NULL)) { | |
2081 | unlock_page(page); | |
2082 | result = SCAN_SWAP_CACHE_PAGE; | |
2083 | goto out; | |
2084 | } | |
2085 | /* | |
2086 | * Page is not in the swap cache. It can be collapsed | |
2087 | * into a THP. | |
2088 | */ | |
2089 | } | |
2090 | ||
2091 | /* | |
2092 | * Isolate the page to avoid collapsing an hugepage | |
2093 | * currently in use by the VM. | |
2094 | */ | |
2095 | if (isolate_lru_page(page)) { | |
2096 | unlock_page(page); | |
2097 | result = SCAN_DEL_PAGE_LRU; | |
2098 | goto out; | |
2099 | } | |
2100 | /* 0 stands for page_is_file_cache(page) == false */ | |
2101 | inc_zone_page_state(page, NR_ISOLATED_ANON + 0); | |
2102 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
2103 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
2104 | ||
2105 | /* If there is no mapped pte young don't collapse the page */ | |
2106 | if (pte_young(pteval) || | |
2107 | page_is_young(page) || PageReferenced(page) || | |
2108 | mmu_notifier_test_young(vma->vm_mm, address)) | |
2109 | referenced = true; | |
2110 | } | |
2111 | if (likely(writable)) { | |
2112 | if (likely(referenced)) { | |
2113 | result = SCAN_SUCCEED; | |
2114 | trace_mm_collapse_huge_page_isolate(page, none_or_zero, | |
2115 | referenced, writable, result); | |
2116 | return 1; | |
2117 | } | |
2118 | } else { | |
2119 | result = SCAN_PAGE_RO; | |
2120 | } | |
2121 | ||
2122 | out: | |
2123 | release_pte_pages(pte, _pte); | |
2124 | trace_mm_collapse_huge_page_isolate(page, none_or_zero, | |
2125 | referenced, writable, result); | |
2126 | return 0; | |
2127 | } | |
2128 | ||
2129 | static void __collapse_huge_page_copy(pte_t *pte, struct page *page, | |
2130 | struct vm_area_struct *vma, | |
2131 | unsigned long address, | |
2132 | spinlock_t *ptl) | |
2133 | { | |
2134 | pte_t *_pte; | |
2135 | for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) { | |
2136 | pte_t pteval = *_pte; | |
2137 | struct page *src_page; | |
2138 | ||
2139 | if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { | |
2140 | clear_user_highpage(page, address); | |
2141 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); | |
2142 | if (is_zero_pfn(pte_pfn(pteval))) { | |
2143 | /* | |
2144 | * ptl mostly unnecessary. | |
2145 | */ | |
2146 | spin_lock(ptl); | |
2147 | /* | |
2148 | * paravirt calls inside pte_clear here are | |
2149 | * superfluous. | |
2150 | */ | |
2151 | pte_clear(vma->vm_mm, address, _pte); | |
2152 | spin_unlock(ptl); | |
2153 | } | |
2154 | } else { | |
2155 | src_page = pte_page(pteval); | |
2156 | copy_user_highpage(page, src_page, address, vma); | |
2157 | VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page); | |
2158 | release_pte_page(src_page); | |
2159 | /* | |
2160 | * ptl mostly unnecessary, but preempt has to | |
2161 | * be disabled to update the per-cpu stats | |
2162 | * inside page_remove_rmap(). | |
2163 | */ | |
2164 | spin_lock(ptl); | |
2165 | /* | |
2166 | * paravirt calls inside pte_clear here are | |
2167 | * superfluous. | |
2168 | */ | |
2169 | pte_clear(vma->vm_mm, address, _pte); | |
2170 | page_remove_rmap(src_page, false); | |
2171 | spin_unlock(ptl); | |
2172 | free_page_and_swap_cache(src_page); | |
2173 | } | |
2174 | ||
2175 | address += PAGE_SIZE; | |
2176 | page++; | |
2177 | } | |
2178 | } | |
2179 | ||
2180 | static void khugepaged_alloc_sleep(void) | |
2181 | { | |
2182 | DEFINE_WAIT(wait); | |
2183 | ||
2184 | add_wait_queue(&khugepaged_wait, &wait); | |
2185 | freezable_schedule_timeout_interruptible( | |
2186 | msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); | |
2187 | remove_wait_queue(&khugepaged_wait, &wait); | |
2188 | } | |
2189 | ||
2190 | static int khugepaged_node_load[MAX_NUMNODES]; | |
2191 | ||
2192 | static bool khugepaged_scan_abort(int nid) | |
2193 | { | |
2194 | int i; | |
2195 | ||
2196 | /* | |
2197 | * If zone_reclaim_mode is disabled, then no extra effort is made to | |
2198 | * allocate memory locally. | |
2199 | */ | |
2200 | if (!zone_reclaim_mode) | |
2201 | return false; | |
2202 | ||
2203 | /* If there is a count for this node already, it must be acceptable */ | |
2204 | if (khugepaged_node_load[nid]) | |
2205 | return false; | |
2206 | ||
2207 | for (i = 0; i < MAX_NUMNODES; i++) { | |
2208 | if (!khugepaged_node_load[i]) | |
2209 | continue; | |
2210 | if (node_distance(nid, i) > RECLAIM_DISTANCE) | |
2211 | return true; | |
2212 | } | |
2213 | return false; | |
2214 | } | |
2215 | ||
2216 | #ifdef CONFIG_NUMA | |
2217 | static int khugepaged_find_target_node(void) | |
2218 | { | |
2219 | static int last_khugepaged_target_node = NUMA_NO_NODE; | |
2220 | int nid, target_node = 0, max_value = 0; | |
2221 | ||
2222 | /* find first node with max normal pages hit */ | |
2223 | for (nid = 0; nid < MAX_NUMNODES; nid++) | |
2224 | if (khugepaged_node_load[nid] > max_value) { | |
2225 | max_value = khugepaged_node_load[nid]; | |
2226 | target_node = nid; | |
2227 | } | |
2228 | ||
2229 | /* do some balance if several nodes have the same hit record */ | |
2230 | if (target_node <= last_khugepaged_target_node) | |
2231 | for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES; | |
2232 | nid++) | |
2233 | if (max_value == khugepaged_node_load[nid]) { | |
2234 | target_node = nid; | |
2235 | break; | |
2236 | } | |
2237 | ||
2238 | last_khugepaged_target_node = target_node; | |
2239 | return target_node; | |
2240 | } | |
2241 | ||
2242 | static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) | |
2243 | { | |
2244 | if (IS_ERR(*hpage)) { | |
2245 | if (!*wait) | |
2246 | return false; | |
2247 | ||
2248 | *wait = false; | |
2249 | *hpage = NULL; | |
2250 | khugepaged_alloc_sleep(); | |
2251 | } else if (*hpage) { | |
2252 | put_page(*hpage); | |
2253 | *hpage = NULL; | |
2254 | } | |
2255 | ||
2256 | return true; | |
2257 | } | |
2258 | ||
2259 | static struct page * | |
2260 | khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm, | |
2261 | unsigned long address, int node) | |
2262 | { | |
2263 | VM_BUG_ON_PAGE(*hpage, *hpage); | |
2264 | ||
2265 | /* | |
2266 | * Before allocating the hugepage, release the mmap_sem read lock. | |
2267 | * The allocation can take potentially a long time if it involves | |
2268 | * sync compaction, and we do not need to hold the mmap_sem during | |
2269 | * that. We will recheck the vma after taking it again in write mode. | |
2270 | */ | |
2271 | up_read(&mm->mmap_sem); | |
2272 | ||
2273 | *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER); | |
2274 | if (unlikely(!*hpage)) { | |
2275 | count_vm_event(THP_COLLAPSE_ALLOC_FAILED); | |
2276 | *hpage = ERR_PTR(-ENOMEM); | |
2277 | return NULL; | |
2278 | } | |
2279 | ||
2280 | prep_transhuge_page(*hpage); | |
2281 | count_vm_event(THP_COLLAPSE_ALLOC); | |
2282 | return *hpage; | |
2283 | } | |
2284 | #else | |
2285 | static int khugepaged_find_target_node(void) | |
2286 | { | |
2287 | return 0; | |
2288 | } | |
2289 | ||
2290 | static inline struct page *alloc_khugepaged_hugepage(void) | |
2291 | { | |
2292 | struct page *page; | |
2293 | ||
2294 | page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(), | |
2295 | HPAGE_PMD_ORDER); | |
2296 | if (page) | |
2297 | prep_transhuge_page(page); | |
2298 | return page; | |
2299 | } | |
2300 | ||
2301 | static struct page *khugepaged_alloc_hugepage(bool *wait) | |
2302 | { | |
2303 | struct page *hpage; | |
2304 | ||
2305 | do { | |
2306 | hpage = alloc_khugepaged_hugepage(); | |
2307 | if (!hpage) { | |
2308 | count_vm_event(THP_COLLAPSE_ALLOC_FAILED); | |
2309 | if (!*wait) | |
2310 | return NULL; | |
2311 | ||
2312 | *wait = false; | |
2313 | khugepaged_alloc_sleep(); | |
2314 | } else | |
2315 | count_vm_event(THP_COLLAPSE_ALLOC); | |
2316 | } while (unlikely(!hpage) && likely(khugepaged_enabled())); | |
2317 | ||
2318 | return hpage; | |
2319 | } | |
2320 | ||
2321 | static bool khugepaged_prealloc_page(struct page **hpage, bool *wait) | |
2322 | { | |
2323 | if (!*hpage) | |
2324 | *hpage = khugepaged_alloc_hugepage(wait); | |
2325 | ||
2326 | if (unlikely(!*hpage)) | |
2327 | return false; | |
2328 | ||
2329 | return true; | |
2330 | } | |
2331 | ||
2332 | static struct page * | |
2333 | khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm, | |
2334 | unsigned long address, int node) | |
2335 | { | |
2336 | up_read(&mm->mmap_sem); | |
2337 | VM_BUG_ON(!*hpage); | |
2338 | ||
2339 | return *hpage; | |
2340 | } | |
2341 | #endif | |
2342 | ||
2343 | static bool hugepage_vma_check(struct vm_area_struct *vma) | |
2344 | { | |
2345 | if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) || | |
2346 | (vma->vm_flags & VM_NOHUGEPAGE)) | |
2347 | return false; | |
2348 | if (!vma->anon_vma || vma->vm_ops) | |
2349 | return false; | |
2350 | if (is_vma_temporary_stack(vma)) | |
2351 | return false; | |
2352 | return !(vma->vm_flags & VM_NO_THP); | |
2353 | } | |
2354 | ||
2355 | /* | |
2356 | * If mmap_sem temporarily dropped, revalidate vma | |
2357 | * before taking mmap_sem. | |
2358 | * Return 0 if succeeds, otherwise return none-zero | |
2359 | * value (scan code). | |
2360 | */ | |
2361 | ||
2362 | static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address) | |
2363 | { | |
2364 | struct vm_area_struct *vma; | |
2365 | unsigned long hstart, hend; | |
2366 | ||
2367 | if (unlikely(khugepaged_test_exit(mm))) | |
2368 | return SCAN_ANY_PROCESS; | |
2369 | ||
2370 | vma = find_vma(mm, address); | |
2371 | if (!vma) | |
2372 | return SCAN_VMA_NULL; | |
2373 | ||
2374 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; | |
2375 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
2376 | if (address < hstart || address + HPAGE_PMD_SIZE > hend) | |
2377 | return SCAN_ADDRESS_RANGE; | |
2378 | if (!hugepage_vma_check(vma)) | |
2379 | return SCAN_VMA_CHECK; | |
2380 | return 0; | |
2381 | } | |
2382 | ||
2383 | /* | |
2384 | * Bring missing pages in from swap, to complete THP collapse. | |
2385 | * Only done if khugepaged_scan_pmd believes it is worthwhile. | |
2386 | * | |
2387 | * Called and returns without pte mapped or spinlocks held, | |
2388 | * but with mmap_sem held to protect against vma changes. | |
2389 | */ | |
2390 | ||
2391 | static bool __collapse_huge_page_swapin(struct mm_struct *mm, | |
2392 | struct vm_area_struct *vma, | |
2393 | unsigned long address, pmd_t *pmd) | |
2394 | { | |
2395 | pte_t pteval; | |
2396 | int swapped_in = 0, ret = 0; | |
2397 | struct fault_env fe = { | |
2398 | .vma = vma, | |
2399 | .address = address, | |
2400 | .flags = FAULT_FLAG_ALLOW_RETRY, | |
2401 | .pmd = pmd, | |
2402 | }; | |
2403 | ||
2404 | fe.pte = pte_offset_map(pmd, address); | |
2405 | for (; fe.address < address + HPAGE_PMD_NR*PAGE_SIZE; | |
2406 | fe.pte++, fe.address += PAGE_SIZE) { | |
2407 | pteval = *fe.pte; | |
2408 | if (!is_swap_pte(pteval)) | |
2409 | continue; | |
2410 | swapped_in++; | |
2411 | ret = do_swap_page(&fe, pteval); | |
2412 | /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */ | |
2413 | if (ret & VM_FAULT_RETRY) { | |
2414 | down_read(&mm->mmap_sem); | |
2415 | /* vma is no longer available, don't continue to swapin */ | |
2416 | if (hugepage_vma_revalidate(mm, address)) | |
2417 | return false; | |
2418 | /* check if the pmd is still valid */ | |
2419 | if (mm_find_pmd(mm, address) != pmd) | |
2420 | return false; | |
2421 | } | |
2422 | if (ret & VM_FAULT_ERROR) { | |
2423 | trace_mm_collapse_huge_page_swapin(mm, swapped_in, 0); | |
2424 | return false; | |
2425 | } | |
2426 | /* pte is unmapped now, we need to map it */ | |
2427 | fe.pte = pte_offset_map(pmd, fe.address); | |
2428 | } | |
2429 | fe.pte--; | |
2430 | pte_unmap(fe.pte); | |
2431 | trace_mm_collapse_huge_page_swapin(mm, swapped_in, 1); | |
2432 | return true; | |
2433 | } | |
2434 | ||
2435 | static void collapse_huge_page(struct mm_struct *mm, | |
2436 | unsigned long address, | |
2437 | struct page **hpage, | |
2438 | struct vm_area_struct *vma, | |
2439 | int node) | |
2440 | { | |
2441 | pmd_t *pmd, _pmd; | |
2442 | pte_t *pte; | |
2443 | pgtable_t pgtable; | |
2444 | struct page *new_page; | |
2445 | spinlock_t *pmd_ptl, *pte_ptl; | |
2446 | int isolated = 0, result = 0; | |
2447 | struct mem_cgroup *memcg; | |
2448 | unsigned long mmun_start; /* For mmu_notifiers */ | |
2449 | unsigned long mmun_end; /* For mmu_notifiers */ | |
2450 | gfp_t gfp; | |
2451 | ||
2452 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
2453 | ||
2454 | /* Only allocate from the target node */ | |
2455 | gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_OTHER_NODE | __GFP_THISNODE; | |
2456 | ||
2457 | /* release the mmap_sem read lock. */ | |
2458 | new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node); | |
2459 | if (!new_page) { | |
2460 | result = SCAN_ALLOC_HUGE_PAGE_FAIL; | |
2461 | goto out_nolock; | |
2462 | } | |
2463 | ||
2464 | if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) { | |
2465 | result = SCAN_CGROUP_CHARGE_FAIL; | |
2466 | goto out_nolock; | |
2467 | } | |
2468 | ||
2469 | down_read(&mm->mmap_sem); | |
2470 | result = hugepage_vma_revalidate(mm, address); | |
2471 | if (result) { | |
2472 | mem_cgroup_cancel_charge(new_page, memcg, true); | |
2473 | up_read(&mm->mmap_sem); | |
2474 | goto out_nolock; | |
2475 | } | |
2476 | ||
2477 | pmd = mm_find_pmd(mm, address); | |
2478 | if (!pmd) { | |
2479 | result = SCAN_PMD_NULL; | |
2480 | mem_cgroup_cancel_charge(new_page, memcg, true); | |
2481 | up_read(&mm->mmap_sem); | |
2482 | goto out_nolock; | |
2483 | } | |
2484 | ||
2485 | /* | |
2486 | * __collapse_huge_page_swapin always returns with mmap_sem locked. | |
2487 | * If it fails, release mmap_sem and jump directly out. | |
2488 | * Continuing to collapse causes inconsistency. | |
2489 | */ | |
2490 | if (!__collapse_huge_page_swapin(mm, vma, address, pmd)) { | |
2491 | mem_cgroup_cancel_charge(new_page, memcg, true); | |
2492 | up_read(&mm->mmap_sem); | |
2493 | goto out_nolock; | |
2494 | } | |
2495 | ||
2496 | up_read(&mm->mmap_sem); | |
2497 | /* | |
2498 | * Prevent all access to pagetables with the exception of | |
2499 | * gup_fast later handled by the ptep_clear_flush and the VM | |
2500 | * handled by the anon_vma lock + PG_lock. | |
2501 | */ | |
2502 | down_write(&mm->mmap_sem); | |
2503 | result = hugepage_vma_revalidate(mm, address); | |
2504 | if (result) | |
2505 | goto out; | |
2506 | /* check if the pmd is still valid */ | |
2507 | if (mm_find_pmd(mm, address) != pmd) | |
2508 | goto out; | |
2509 | ||
2510 | anon_vma_lock_write(vma->anon_vma); | |
2511 | ||
2512 | pte = pte_offset_map(pmd, address); | |
2513 | pte_ptl = pte_lockptr(mm, pmd); | |
2514 | ||
2515 | mmun_start = address; | |
2516 | mmun_end = address + HPAGE_PMD_SIZE; | |
2517 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
2518 | pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ | |
2519 | /* | |
2520 | * After this gup_fast can't run anymore. This also removes | |
2521 | * any huge TLB entry from the CPU so we won't allow | |
2522 | * huge and small TLB entries for the same virtual address | |
2523 | * to avoid the risk of CPU bugs in that area. | |
2524 | */ | |
2525 | _pmd = pmdp_collapse_flush(vma, address, pmd); | |
2526 | spin_unlock(pmd_ptl); | |
2527 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); | |
2528 | ||
2529 | spin_lock(pte_ptl); | |
2530 | isolated = __collapse_huge_page_isolate(vma, address, pte); | |
2531 | spin_unlock(pte_ptl); | |
2532 | ||
2533 | if (unlikely(!isolated)) { | |
2534 | pte_unmap(pte); | |
2535 | spin_lock(pmd_ptl); | |
2536 | BUG_ON(!pmd_none(*pmd)); | |
2537 | /* | |
2538 | * We can only use set_pmd_at when establishing | |
2539 | * hugepmds and never for establishing regular pmds that | |
2540 | * points to regular pagetables. Use pmd_populate for that | |
2541 | */ | |
2542 | pmd_populate(mm, pmd, pmd_pgtable(_pmd)); | |
2543 | spin_unlock(pmd_ptl); | |
2544 | anon_vma_unlock_write(vma->anon_vma); | |
2545 | result = SCAN_FAIL; | |
2546 | goto out; | |
2547 | } | |
2548 | ||
2549 | /* | |
2550 | * All pages are isolated and locked so anon_vma rmap | |
2551 | * can't run anymore. | |
2552 | */ | |
2553 | anon_vma_unlock_write(vma->anon_vma); | |
2554 | ||
2555 | __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl); | |
2556 | pte_unmap(pte); | |
2557 | __SetPageUptodate(new_page); | |
2558 | pgtable = pmd_pgtable(_pmd); | |
2559 | ||
2560 | _pmd = mk_huge_pmd(new_page, vma->vm_page_prot); | |
2561 | _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); | |
2562 | ||
2563 | /* | |
2564 | * spin_lock() below is not the equivalent of smp_wmb(), so | |
2565 | * this is needed to avoid the copy_huge_page writes to become | |
2566 | * visible after the set_pmd_at() write. | |
2567 | */ | |
2568 | smp_wmb(); | |
2569 | ||
2570 | spin_lock(pmd_ptl); | |
2571 | BUG_ON(!pmd_none(*pmd)); | |
2572 | page_add_new_anon_rmap(new_page, vma, address, true); | |
2573 | mem_cgroup_commit_charge(new_page, memcg, false, true); | |
2574 | lru_cache_add_active_or_unevictable(new_page, vma); | |
2575 | pgtable_trans_huge_deposit(mm, pmd, pgtable); | |
2576 | set_pmd_at(mm, address, pmd, _pmd); | |
2577 | update_mmu_cache_pmd(vma, address, pmd); | |
2578 | spin_unlock(pmd_ptl); | |
2579 | ||
2580 | *hpage = NULL; | |
2581 | ||
2582 | khugepaged_pages_collapsed++; | |
2583 | result = SCAN_SUCCEED; | |
2584 | out_up_write: | |
2585 | up_write(&mm->mmap_sem); | |
2586 | out_nolock: | |
2587 | trace_mm_collapse_huge_page(mm, isolated, result); | |
2588 | return; | |
2589 | out: | |
2590 | mem_cgroup_cancel_charge(new_page, memcg, true); | |
2591 | goto out_up_write; | |
2592 | } | |
2593 | ||
2594 | static int khugepaged_scan_pmd(struct mm_struct *mm, | |
2595 | struct vm_area_struct *vma, | |
2596 | unsigned long address, | |
2597 | struct page **hpage) | |
2598 | { | |
2599 | pmd_t *pmd; | |
2600 | pte_t *pte, *_pte; | |
2601 | int ret = 0, none_or_zero = 0, result = 0; | |
2602 | struct page *page = NULL; | |
2603 | unsigned long _address; | |
2604 | spinlock_t *ptl; | |
2605 | int node = NUMA_NO_NODE, unmapped = 0; | |
2606 | bool writable = false, referenced = false; | |
2607 | ||
2608 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
2609 | ||
2610 | pmd = mm_find_pmd(mm, address); | |
2611 | if (!pmd) { | |
2612 | result = SCAN_PMD_NULL; | |
2613 | goto out; | |
2614 | } | |
2615 | ||
2616 | memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load)); | |
2617 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); | |
2618 | for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR; | |
2619 | _pte++, _address += PAGE_SIZE) { | |
2620 | pte_t pteval = *_pte; | |
2621 | if (is_swap_pte(pteval)) { | |
2622 | if (++unmapped <= khugepaged_max_ptes_swap) { | |
2623 | continue; | |
2624 | } else { | |
2625 | result = SCAN_EXCEED_SWAP_PTE; | |
2626 | goto out_unmap; | |
2627 | } | |
2628 | } | |
2629 | if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { | |
2630 | if (!userfaultfd_armed(vma) && | |
2631 | ++none_or_zero <= khugepaged_max_ptes_none) { | |
2632 | continue; | |
2633 | } else { | |
2634 | result = SCAN_EXCEED_NONE_PTE; | |
2635 | goto out_unmap; | |
2636 | } | |
2637 | } | |
2638 | if (!pte_present(pteval)) { | |
2639 | result = SCAN_PTE_NON_PRESENT; | |
2640 | goto out_unmap; | |
2641 | } | |
2642 | if (pte_write(pteval)) | |
2643 | writable = true; | |
2644 | ||
2645 | page = vm_normal_page(vma, _address, pteval); | |
2646 | if (unlikely(!page)) { | |
2647 | result = SCAN_PAGE_NULL; | |
2648 | goto out_unmap; | |
2649 | } | |
2650 | ||
2651 | /* TODO: teach khugepaged to collapse THP mapped with pte */ | |
2652 | if (PageCompound(page)) { | |
2653 | result = SCAN_PAGE_COMPOUND; | |
2654 | goto out_unmap; | |
2655 | } | |
2656 | ||
2657 | /* | |
2658 | * Record which node the original page is from and save this | |
2659 | * information to khugepaged_node_load[]. | |
2660 | * Khupaged will allocate hugepage from the node has the max | |
2661 | * hit record. | |
2662 | */ | |
2663 | node = page_to_nid(page); | |
2664 | if (khugepaged_scan_abort(node)) { | |
2665 | result = SCAN_SCAN_ABORT; | |
2666 | goto out_unmap; | |
2667 | } | |
2668 | khugepaged_node_load[node]++; | |
2669 | if (!PageLRU(page)) { | |
2670 | result = SCAN_PAGE_LRU; | |
2671 | goto out_unmap; | |
2672 | } | |
2673 | if (PageLocked(page)) { | |
2674 | result = SCAN_PAGE_LOCK; | |
2675 | goto out_unmap; | |
2676 | } | |
2677 | if (!PageAnon(page)) { | |
2678 | result = SCAN_PAGE_ANON; | |
2679 | goto out_unmap; | |
2680 | } | |
2681 | ||
2682 | /* | |
2683 | * cannot use mapcount: can't collapse if there's a gup pin. | |
2684 | * The page must only be referenced by the scanned process | |
2685 | * and page swap cache. | |
2686 | */ | |
2687 | if (page_count(page) != 1 + !!PageSwapCache(page)) { | |
2688 | result = SCAN_PAGE_COUNT; | |
2689 | goto out_unmap; | |
2690 | } | |
2691 | if (pte_young(pteval) || | |
2692 | page_is_young(page) || PageReferenced(page) || | |
2693 | mmu_notifier_test_young(vma->vm_mm, address)) | |
2694 | referenced = true; | |
2695 | } | |
2696 | if (writable) { | |
2697 | if (referenced) { | |
2698 | result = SCAN_SUCCEED; | |
2699 | ret = 1; | |
2700 | } else { | |
2701 | result = SCAN_NO_REFERENCED_PAGE; | |
2702 | } | |
2703 | } else { | |
2704 | result = SCAN_PAGE_RO; | |
2705 | } | |
2706 | out_unmap: | |
2707 | pte_unmap_unlock(pte, ptl); | |
2708 | if (ret) { | |
2709 | node = khugepaged_find_target_node(); | |
2710 | /* collapse_huge_page will return with the mmap_sem released */ | |
2711 | collapse_huge_page(mm, address, hpage, vma, node); | |
2712 | } | |
2713 | out: | |
2714 | trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced, | |
2715 | none_or_zero, result, unmapped); | |
2716 | return ret; | |
2717 | } | |
2718 | ||
2719 | static void collect_mm_slot(struct mm_slot *mm_slot) | |
2720 | { | |
2721 | struct mm_struct *mm = mm_slot->mm; | |
2722 | ||
2723 | VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); | |
2724 | ||
2725 | if (khugepaged_test_exit(mm)) { | |
2726 | /* free mm_slot */ | |
2727 | hash_del(&mm_slot->hash); | |
2728 | list_del(&mm_slot->mm_node); | |
2729 | ||
2730 | /* | |
2731 | * Not strictly needed because the mm exited already. | |
2732 | * | |
2733 | * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); | |
2734 | */ | |
2735 | ||
2736 | /* khugepaged_mm_lock actually not necessary for the below */ | |
2737 | free_mm_slot(mm_slot); | |
2738 | mmdrop(mm); | |
2739 | } | |
2740 | } | |
2741 | ||
2742 | static unsigned int khugepaged_scan_mm_slot(unsigned int pages, | |
2743 | struct page **hpage) | |
2744 | __releases(&khugepaged_mm_lock) | |
2745 | __acquires(&khugepaged_mm_lock) | |
2746 | { | |
2747 | struct mm_slot *mm_slot; | |
2748 | struct mm_struct *mm; | |
2749 | struct vm_area_struct *vma; | |
2750 | int progress = 0; | |
2751 | ||
2752 | VM_BUG_ON(!pages); | |
2753 | VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock)); | |
2754 | ||
2755 | if (khugepaged_scan.mm_slot) | |
2756 | mm_slot = khugepaged_scan.mm_slot; | |
2757 | else { | |
2758 | mm_slot = list_entry(khugepaged_scan.mm_head.next, | |
2759 | struct mm_slot, mm_node); | |
2760 | khugepaged_scan.address = 0; | |
2761 | khugepaged_scan.mm_slot = mm_slot; | |
2762 | } | |
2763 | spin_unlock(&khugepaged_mm_lock); | |
2764 | ||
2765 | mm = mm_slot->mm; | |
2766 | down_read(&mm->mmap_sem); | |
2767 | if (unlikely(khugepaged_test_exit(mm))) | |
2768 | vma = NULL; | |
2769 | else | |
2770 | vma = find_vma(mm, khugepaged_scan.address); | |
2771 | ||
2772 | progress++; | |
2773 | for (; vma; vma = vma->vm_next) { | |
2774 | unsigned long hstart, hend; | |
2775 | ||
2776 | cond_resched(); | |
2777 | if (unlikely(khugepaged_test_exit(mm))) { | |
2778 | progress++; | |
2779 | break; | |
2780 | } | |
2781 | if (!hugepage_vma_check(vma)) { | |
2782 | skip: | |
2783 | progress++; | |
2784 | continue; | |
2785 | } | |
2786 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; | |
2787 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
2788 | if (hstart >= hend) | |
2789 | goto skip; | |
2790 | if (khugepaged_scan.address > hend) | |
2791 | goto skip; | |
2792 | if (khugepaged_scan.address < hstart) | |
2793 | khugepaged_scan.address = hstart; | |
2794 | VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); | |
2795 | ||
2796 | while (khugepaged_scan.address < hend) { | |
2797 | int ret; | |
2798 | cond_resched(); | |
2799 | if (unlikely(khugepaged_test_exit(mm))) | |
2800 | goto breakouterloop; | |
2801 | ||
2802 | VM_BUG_ON(khugepaged_scan.address < hstart || | |
2803 | khugepaged_scan.address + HPAGE_PMD_SIZE > | |
2804 | hend); | |
2805 | ret = khugepaged_scan_pmd(mm, vma, | |
2806 | khugepaged_scan.address, | |
2807 | hpage); | |
2808 | /* move to next address */ | |
2809 | khugepaged_scan.address += HPAGE_PMD_SIZE; | |
2810 | progress += HPAGE_PMD_NR; | |
2811 | if (ret) | |
2812 | /* we released mmap_sem so break loop */ | |
2813 | goto breakouterloop_mmap_sem; | |
2814 | if (progress >= pages) | |
2815 | goto breakouterloop; | |
2816 | } | |
2817 | } | |
2818 | breakouterloop: | |
2819 | up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */ | |
2820 | breakouterloop_mmap_sem: | |
2821 | ||
2822 | spin_lock(&khugepaged_mm_lock); | |
2823 | VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); | |
2824 | /* | |
2825 | * Release the current mm_slot if this mm is about to die, or | |
2826 | * if we scanned all vmas of this mm. | |
2827 | */ | |
2828 | if (khugepaged_test_exit(mm) || !vma) { | |
2829 | /* | |
2830 | * Make sure that if mm_users is reaching zero while | |
2831 | * khugepaged runs here, khugepaged_exit will find | |
2832 | * mm_slot not pointing to the exiting mm. | |
2833 | */ | |
2834 | if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) { | |
2835 | khugepaged_scan.mm_slot = list_entry( | |
2836 | mm_slot->mm_node.next, | |
2837 | struct mm_slot, mm_node); | |
2838 | khugepaged_scan.address = 0; | |
2839 | } else { | |
2840 | khugepaged_scan.mm_slot = NULL; | |
2841 | khugepaged_full_scans++; | |
2842 | } | |
2843 | ||
2844 | collect_mm_slot(mm_slot); | |
2845 | } | |
2846 | ||
2847 | return progress; | |
2848 | } | |
2849 | ||
2850 | static int khugepaged_has_work(void) | |
2851 | { | |
2852 | return !list_empty(&khugepaged_scan.mm_head) && | |
2853 | khugepaged_enabled(); | |
2854 | } | |
2855 | ||
2856 | static int khugepaged_wait_event(void) | |
2857 | { | |
2858 | return !list_empty(&khugepaged_scan.mm_head) || | |
2859 | kthread_should_stop(); | |
2860 | } | |
2861 | ||
2862 | static void khugepaged_do_scan(void) | |
2863 | { | |
2864 | struct page *hpage = NULL; | |
2865 | unsigned int progress = 0, pass_through_head = 0; | |
2866 | unsigned int pages = khugepaged_pages_to_scan; | |
2867 | bool wait = true; | |
2868 | ||
2869 | barrier(); /* write khugepaged_pages_to_scan to local stack */ | |
2870 | ||
2871 | while (progress < pages) { | |
2872 | if (!khugepaged_prealloc_page(&hpage, &wait)) | |
2873 | break; | |
2874 | ||
2875 | cond_resched(); | |
2876 | ||
2877 | if (unlikely(kthread_should_stop() || try_to_freeze())) | |
2878 | break; | |
2879 | ||
2880 | spin_lock(&khugepaged_mm_lock); | |
2881 | if (!khugepaged_scan.mm_slot) | |
2882 | pass_through_head++; | |
2883 | if (khugepaged_has_work() && | |
2884 | pass_through_head < 2) | |
2885 | progress += khugepaged_scan_mm_slot(pages - progress, | |
2886 | &hpage); | |
2887 | else | |
2888 | progress = pages; | |
2889 | spin_unlock(&khugepaged_mm_lock); | |
2890 | } | |
2891 | ||
2892 | if (!IS_ERR_OR_NULL(hpage)) | |
2893 | put_page(hpage); | |
2894 | } | |
2895 | ||
2896 | static bool khugepaged_should_wakeup(void) | |
2897 | { | |
2898 | return kthread_should_stop() || | |
2899 | time_after_eq(jiffies, khugepaged_sleep_expire); | |
2900 | } | |
2901 | ||
2902 | static void khugepaged_wait_work(void) | |
2903 | { | |
2904 | if (khugepaged_has_work()) { | |
2905 | const unsigned long scan_sleep_jiffies = | |
2906 | msecs_to_jiffies(khugepaged_scan_sleep_millisecs); | |
2907 | ||
2908 | if (!scan_sleep_jiffies) | |
2909 | return; | |
2910 | ||
2911 | khugepaged_sleep_expire = jiffies + scan_sleep_jiffies; | |
2912 | wait_event_freezable_timeout(khugepaged_wait, | |
2913 | khugepaged_should_wakeup(), | |
2914 | scan_sleep_jiffies); | |
2915 | return; | |
2916 | } | |
2917 | ||
2918 | if (khugepaged_enabled()) | |
2919 | wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); | |
2920 | } | |
2921 | ||
2922 | static int khugepaged(void *none) | |
2923 | { | |
2924 | struct mm_slot *mm_slot; | |
2925 | ||
2926 | set_freezable(); | |
2927 | set_user_nice(current, MAX_NICE); | |
2928 | ||
2929 | while (!kthread_should_stop()) { | |
2930 | khugepaged_do_scan(); | |
2931 | khugepaged_wait_work(); | |
2932 | } | |
2933 | ||
2934 | spin_lock(&khugepaged_mm_lock); | |
2935 | mm_slot = khugepaged_scan.mm_slot; | |
2936 | khugepaged_scan.mm_slot = NULL; | |
2937 | if (mm_slot) | |
2938 | collect_mm_slot(mm_slot); | |
2939 | spin_unlock(&khugepaged_mm_lock); | |
2940 | return 0; | |
2941 | } | |
2942 | ||
2943 | static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, | |
2944 | unsigned long haddr, pmd_t *pmd) | |
2945 | { | |
2946 | struct mm_struct *mm = vma->vm_mm; | |
2947 | pgtable_t pgtable; | |
2948 | pmd_t _pmd; | |
2949 | int i; | |
2950 | ||
2951 | /* leave pmd empty until pte is filled */ | |
2952 | pmdp_huge_clear_flush_notify(vma, haddr, pmd); | |
2953 | ||
2954 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); | |
2955 | pmd_populate(mm, &_pmd, pgtable); | |
2956 | ||
2957 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
2958 | pte_t *pte, entry; | |
2959 | entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); | |
2960 | entry = pte_mkspecial(entry); | |
2961 | pte = pte_offset_map(&_pmd, haddr); | |
2962 | VM_BUG_ON(!pte_none(*pte)); | |
2963 | set_pte_at(mm, haddr, pte, entry); | |
2964 | pte_unmap(pte); | |
2965 | } | |
2966 | smp_wmb(); /* make pte visible before pmd */ | |
2967 | pmd_populate(mm, pmd, pgtable); | |
2968 | put_huge_zero_page(); | |
2969 | } | |
2970 | ||
2971 | static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, | |
2972 | unsigned long haddr, bool freeze) | |
2973 | { | |
2974 | struct mm_struct *mm = vma->vm_mm; | |
2975 | struct page *page; | |
2976 | pgtable_t pgtable; | |
2977 | pmd_t _pmd; | |
2978 | bool young, write, dirty; | |
2979 | unsigned long addr; | |
2980 | int i; | |
2981 | ||
2982 | VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); | |
2983 | VM_BUG_ON_VMA(vma->vm_start > haddr, vma); | |
2984 | VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); | |
2985 | VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)); | |
2986 | ||
2987 | count_vm_event(THP_SPLIT_PMD); | |
2988 | ||
2989 | if (vma_is_dax(vma)) { | |
2990 | pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); | |
2991 | if (is_huge_zero_pmd(_pmd)) | |
2992 | put_huge_zero_page(); | |
2993 | return; | |
2994 | } else if (is_huge_zero_pmd(*pmd)) { | |
2995 | return __split_huge_zero_page_pmd(vma, haddr, pmd); | |
2996 | } | |
2997 | ||
2998 | page = pmd_page(*pmd); | |
2999 | VM_BUG_ON_PAGE(!page_count(page), page); | |
3000 | page_ref_add(page, HPAGE_PMD_NR - 1); | |
3001 | write = pmd_write(*pmd); | |
3002 | young = pmd_young(*pmd); | |
3003 | dirty = pmd_dirty(*pmd); | |
3004 | ||
3005 | pmdp_huge_split_prepare(vma, haddr, pmd); | |
3006 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); | |
3007 | pmd_populate(mm, &_pmd, pgtable); | |
3008 | ||
3009 | for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { | |
3010 | pte_t entry, *pte; | |
3011 | /* | |
3012 | * Note that NUMA hinting access restrictions are not | |
3013 | * transferred to avoid any possibility of altering | |
3014 | * permissions across VMAs. | |
3015 | */ | |
3016 | if (freeze) { | |
3017 | swp_entry_t swp_entry; | |
3018 | swp_entry = make_migration_entry(page + i, write); | |
3019 | entry = swp_entry_to_pte(swp_entry); | |
3020 | } else { | |
3021 | entry = mk_pte(page + i, vma->vm_page_prot); | |
3022 | entry = maybe_mkwrite(entry, vma); | |
3023 | if (!write) | |
3024 | entry = pte_wrprotect(entry); | |
3025 | if (!young) | |
3026 | entry = pte_mkold(entry); | |
3027 | } | |
3028 | if (dirty) | |
3029 | SetPageDirty(page + i); | |
3030 | pte = pte_offset_map(&_pmd, addr); | |
3031 | BUG_ON(!pte_none(*pte)); | |
3032 | set_pte_at(mm, addr, pte, entry); | |
3033 | atomic_inc(&page[i]._mapcount); | |
3034 | pte_unmap(pte); | |
3035 | } | |
3036 | ||
3037 | /* | |
3038 | * Set PG_double_map before dropping compound_mapcount to avoid | |
3039 | * false-negative page_mapped(). | |
3040 | */ | |
3041 | if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) { | |
3042 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
3043 | atomic_inc(&page[i]._mapcount); | |
3044 | } | |
3045 | ||
3046 | if (atomic_add_negative(-1, compound_mapcount_ptr(page))) { | |
3047 | /* Last compound_mapcount is gone. */ | |
3048 | __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES); | |
3049 | if (TestClearPageDoubleMap(page)) { | |
3050 | /* No need in mapcount reference anymore */ | |
3051 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
3052 | atomic_dec(&page[i]._mapcount); | |
3053 | } | |
3054 | } | |
3055 | ||
3056 | smp_wmb(); /* make pte visible before pmd */ | |
3057 | /* | |
3058 | * Up to this point the pmd is present and huge and userland has the | |
3059 | * whole access to the hugepage during the split (which happens in | |
3060 | * place). If we overwrite the pmd with the not-huge version pointing | |
3061 | * to the pte here (which of course we could if all CPUs were bug | |
3062 | * free), userland could trigger a small page size TLB miss on the | |
3063 | * small sized TLB while the hugepage TLB entry is still established in | |
3064 | * the huge TLB. Some CPU doesn't like that. | |
3065 | * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum | |
3066 | * 383 on page 93. Intel should be safe but is also warns that it's | |
3067 | * only safe if the permission and cache attributes of the two entries | |
3068 | * loaded in the two TLB is identical (which should be the case here). | |
3069 | * But it is generally safer to never allow small and huge TLB entries | |
3070 | * for the same virtual address to be loaded simultaneously. So instead | |
3071 | * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the | |
3072 | * current pmd notpresent (atomically because here the pmd_trans_huge | |
3073 | * and pmd_trans_splitting must remain set at all times on the pmd | |
3074 | * until the split is complete for this pmd), then we flush the SMP TLB | |
3075 | * and finally we write the non-huge version of the pmd entry with | |
3076 | * pmd_populate. | |
3077 | */ | |
3078 | pmdp_invalidate(vma, haddr, pmd); | |
3079 | pmd_populate(mm, pmd, pgtable); | |
3080 | ||
3081 | if (freeze) { | |
3082 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
3083 | page_remove_rmap(page + i, false); | |
3084 | put_page(page + i); | |
3085 | } | |
3086 | } | |
3087 | } | |
3088 | ||
3089 | void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, | |
3090 | unsigned long address, bool freeze, struct page *page) | |
3091 | { | |
3092 | spinlock_t *ptl; | |
3093 | struct mm_struct *mm = vma->vm_mm; | |
3094 | unsigned long haddr = address & HPAGE_PMD_MASK; | |
3095 | ||
3096 | mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE); | |
3097 | ptl = pmd_lock(mm, pmd); | |
3098 | ||
3099 | /* | |
3100 | * If caller asks to setup a migration entries, we need a page to check | |
3101 | * pmd against. Otherwise we can end up replacing wrong page. | |
3102 | */ | |
3103 | VM_BUG_ON(freeze && !page); | |
3104 | if (page && page != pmd_page(*pmd)) | |
3105 | goto out; | |
3106 | ||
3107 | if (pmd_trans_huge(*pmd)) { | |
3108 | page = pmd_page(*pmd); | |
3109 | if (PageMlocked(page)) | |
3110 | clear_page_mlock(page); | |
3111 | } else if (!pmd_devmap(*pmd)) | |
3112 | goto out; | |
3113 | __split_huge_pmd_locked(vma, pmd, haddr, freeze); | |
3114 | out: | |
3115 | spin_unlock(ptl); | |
3116 | mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE); | |
3117 | } | |
3118 | ||
3119 | void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address, | |
3120 | bool freeze, struct page *page) | |
3121 | { | |
3122 | pgd_t *pgd; | |
3123 | pud_t *pud; | |
3124 | pmd_t *pmd; | |
3125 | ||
3126 | pgd = pgd_offset(vma->vm_mm, address); | |
3127 | if (!pgd_present(*pgd)) | |
3128 | return; | |
3129 | ||
3130 | pud = pud_offset(pgd, address); | |
3131 | if (!pud_present(*pud)) | |
3132 | return; | |
3133 | ||
3134 | pmd = pmd_offset(pud, address); | |
3135 | ||
3136 | __split_huge_pmd(vma, pmd, address, freeze, page); | |
3137 | } | |
3138 | ||
3139 | void vma_adjust_trans_huge(struct vm_area_struct *vma, | |
3140 | unsigned long start, | |
3141 | unsigned long end, | |
3142 | long adjust_next) | |
3143 | { | |
3144 | /* | |
3145 | * If the new start address isn't hpage aligned and it could | |
3146 | * previously contain an hugepage: check if we need to split | |
3147 | * an huge pmd. | |
3148 | */ | |
3149 | if (start & ~HPAGE_PMD_MASK && | |
3150 | (start & HPAGE_PMD_MASK) >= vma->vm_start && | |
3151 | (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
3152 | split_huge_pmd_address(vma, start, false, NULL); | |
3153 | ||
3154 | /* | |
3155 | * If the new end address isn't hpage aligned and it could | |
3156 | * previously contain an hugepage: check if we need to split | |
3157 | * an huge pmd. | |
3158 | */ | |
3159 | if (end & ~HPAGE_PMD_MASK && | |
3160 | (end & HPAGE_PMD_MASK) >= vma->vm_start && | |
3161 | (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
3162 | split_huge_pmd_address(vma, end, false, NULL); | |
3163 | ||
3164 | /* | |
3165 | * If we're also updating the vma->vm_next->vm_start, if the new | |
3166 | * vm_next->vm_start isn't page aligned and it could previously | |
3167 | * contain an hugepage: check if we need to split an huge pmd. | |
3168 | */ | |
3169 | if (adjust_next > 0) { | |
3170 | struct vm_area_struct *next = vma->vm_next; | |
3171 | unsigned long nstart = next->vm_start; | |
3172 | nstart += adjust_next << PAGE_SHIFT; | |
3173 | if (nstart & ~HPAGE_PMD_MASK && | |
3174 | (nstart & HPAGE_PMD_MASK) >= next->vm_start && | |
3175 | (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) | |
3176 | split_huge_pmd_address(next, nstart, false, NULL); | |
3177 | } | |
3178 | } | |
3179 | ||
3180 | static void freeze_page(struct page *page) | |
3181 | { | |
3182 | enum ttu_flags ttu_flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | | |
3183 | TTU_IGNORE_ACCESS | TTU_RMAP_LOCKED; | |
3184 | int i, ret; | |
3185 | ||
3186 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
3187 | ||
3188 | /* We only need TTU_SPLIT_HUGE_PMD once */ | |
3189 | ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD); | |
3190 | for (i = 1; !ret && i < HPAGE_PMD_NR; i++) { | |
3191 | /* Cut short if the page is unmapped */ | |
3192 | if (page_count(page) == 1) | |
3193 | return; | |
3194 | ||
3195 | ret = try_to_unmap(page + i, ttu_flags); | |
3196 | } | |
3197 | VM_BUG_ON(ret); | |
3198 | } | |
3199 | ||
3200 | static void unfreeze_page(struct page *page) | |
3201 | { | |
3202 | int i; | |
3203 | ||
3204 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
3205 | remove_migration_ptes(page + i, page + i, true); | |
3206 | } | |
3207 | ||
3208 | static void __split_huge_page_tail(struct page *head, int tail, | |
3209 | struct lruvec *lruvec, struct list_head *list) | |
3210 | { | |
3211 | struct page *page_tail = head + tail; | |
3212 | ||
3213 | VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail); | |
3214 | VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail); | |
3215 | ||
3216 | /* | |
3217 | * tail_page->_refcount is zero and not changing from under us. But | |
3218 | * get_page_unless_zero() may be running from under us on the | |
3219 | * tail_page. If we used atomic_set() below instead of atomic_inc(), we | |
3220 | * would then run atomic_set() concurrently with | |
3221 | * get_page_unless_zero(), and atomic_set() is implemented in C not | |
3222 | * using locked ops. spin_unlock on x86 sometime uses locked ops | |
3223 | * because of PPro errata 66, 92, so unless somebody can guarantee | |
3224 | * atomic_set() here would be safe on all archs (and not only on x86), | |
3225 | * it's safer to use atomic_inc(). | |
3226 | */ | |
3227 | page_ref_inc(page_tail); | |
3228 | ||
3229 | page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; | |
3230 | page_tail->flags |= (head->flags & | |
3231 | ((1L << PG_referenced) | | |
3232 | (1L << PG_swapbacked) | | |
3233 | (1L << PG_mlocked) | | |
3234 | (1L << PG_uptodate) | | |
3235 | (1L << PG_active) | | |
3236 | (1L << PG_locked) | | |
3237 | (1L << PG_unevictable) | | |
3238 | (1L << PG_dirty))); | |
3239 | ||
3240 | /* | |
3241 | * After clearing PageTail the gup refcount can be released. | |
3242 | * Page flags also must be visible before we make the page non-compound. | |
3243 | */ | |
3244 | smp_wmb(); | |
3245 | ||
3246 | clear_compound_head(page_tail); | |
3247 | ||
3248 | if (page_is_young(head)) | |
3249 | set_page_young(page_tail); | |
3250 | if (page_is_idle(head)) | |
3251 | set_page_idle(page_tail); | |
3252 | ||
3253 | /* ->mapping in first tail page is compound_mapcount */ | |
3254 | VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING, | |
3255 | page_tail); | |
3256 | page_tail->mapping = head->mapping; | |
3257 | ||
3258 | page_tail->index = head->index + tail; | |
3259 | page_cpupid_xchg_last(page_tail, page_cpupid_last(head)); | |
3260 | lru_add_page_tail(head, page_tail, lruvec, list); | |
3261 | } | |
3262 | ||
3263 | static void __split_huge_page(struct page *page, struct list_head *list) | |
3264 | { | |
3265 | struct page *head = compound_head(page); | |
3266 | struct zone *zone = page_zone(head); | |
3267 | struct lruvec *lruvec; | |
3268 | int i; | |
3269 | ||
3270 | /* prevent PageLRU to go away from under us, and freeze lru stats */ | |
3271 | spin_lock_irq(&zone->lru_lock); | |
3272 | lruvec = mem_cgroup_page_lruvec(head, zone); | |
3273 | ||
3274 | /* complete memcg works before add pages to LRU */ | |
3275 | mem_cgroup_split_huge_fixup(head); | |
3276 | ||
3277 | for (i = HPAGE_PMD_NR - 1; i >= 1; i--) | |
3278 | __split_huge_page_tail(head, i, lruvec, list); | |
3279 | ||
3280 | ClearPageCompound(head); | |
3281 | spin_unlock_irq(&zone->lru_lock); | |
3282 | ||
3283 | unfreeze_page(head); | |
3284 | ||
3285 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
3286 | struct page *subpage = head + i; | |
3287 | if (subpage == page) | |
3288 | continue; | |
3289 | unlock_page(subpage); | |
3290 | ||
3291 | /* | |
3292 | * Subpages may be freed if there wasn't any mapping | |
3293 | * like if add_to_swap() is running on a lru page that | |
3294 | * had its mapping zapped. And freeing these pages | |
3295 | * requires taking the lru_lock so we do the put_page | |
3296 | * of the tail pages after the split is complete. | |
3297 | */ | |
3298 | put_page(subpage); | |
3299 | } | |
3300 | } | |
3301 | ||
3302 | int total_mapcount(struct page *page) | |
3303 | { | |
3304 | int i, ret; | |
3305 | ||
3306 | VM_BUG_ON_PAGE(PageTail(page), page); | |
3307 | ||
3308 | if (likely(!PageCompound(page))) | |
3309 | return atomic_read(&page->_mapcount) + 1; | |
3310 | ||
3311 | ret = compound_mapcount(page); | |
3312 | if (PageHuge(page)) | |
3313 | return ret; | |
3314 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
3315 | ret += atomic_read(&page[i]._mapcount) + 1; | |
3316 | if (PageDoubleMap(page)) | |
3317 | ret -= HPAGE_PMD_NR; | |
3318 | return ret; | |
3319 | } | |
3320 | ||
3321 | /* | |
3322 | * This calculates accurately how many mappings a transparent hugepage | |
3323 | * has (unlike page_mapcount() which isn't fully accurate). This full | |
3324 | * accuracy is primarily needed to know if copy-on-write faults can | |
3325 | * reuse the page and change the mapping to read-write instead of | |
3326 | * copying them. At the same time this returns the total_mapcount too. | |
3327 | * | |
3328 | * The function returns the highest mapcount any one of the subpages | |
3329 | * has. If the return value is one, even if different processes are | |
3330 | * mapping different subpages of the transparent hugepage, they can | |
3331 | * all reuse it, because each process is reusing a different subpage. | |
3332 | * | |
3333 | * The total_mapcount is instead counting all virtual mappings of the | |
3334 | * subpages. If the total_mapcount is equal to "one", it tells the | |
3335 | * caller all mappings belong to the same "mm" and in turn the | |
3336 | * anon_vma of the transparent hugepage can become the vma->anon_vma | |
3337 | * local one as no other process may be mapping any of the subpages. | |
3338 | * | |
3339 | * It would be more accurate to replace page_mapcount() with | |
3340 | * page_trans_huge_mapcount(), however we only use | |
3341 | * page_trans_huge_mapcount() in the copy-on-write faults where we | |
3342 | * need full accuracy to avoid breaking page pinning, because | |
3343 | * page_trans_huge_mapcount() is slower than page_mapcount(). | |
3344 | */ | |
3345 | int page_trans_huge_mapcount(struct page *page, int *total_mapcount) | |
3346 | { | |
3347 | int i, ret, _total_mapcount, mapcount; | |
3348 | ||
3349 | /* hugetlbfs shouldn't call it */ | |
3350 | VM_BUG_ON_PAGE(PageHuge(page), page); | |
3351 | ||
3352 | if (likely(!PageTransCompound(page))) { | |
3353 | mapcount = atomic_read(&page->_mapcount) + 1; | |
3354 | if (total_mapcount) | |
3355 | *total_mapcount = mapcount; | |
3356 | return mapcount; | |
3357 | } | |
3358 | ||
3359 | page = compound_head(page); | |
3360 | ||
3361 | _total_mapcount = ret = 0; | |
3362 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
3363 | mapcount = atomic_read(&page[i]._mapcount) + 1; | |
3364 | ret = max(ret, mapcount); | |
3365 | _total_mapcount += mapcount; | |
3366 | } | |
3367 | if (PageDoubleMap(page)) { | |
3368 | ret -= 1; | |
3369 | _total_mapcount -= HPAGE_PMD_NR; | |
3370 | } | |
3371 | mapcount = compound_mapcount(page); | |
3372 | ret += mapcount; | |
3373 | _total_mapcount += mapcount; | |
3374 | if (total_mapcount) | |
3375 | *total_mapcount = _total_mapcount; | |
3376 | return ret; | |
3377 | } | |
3378 | ||
3379 | /* | |
3380 | * This function splits huge page into normal pages. @page can point to any | |
3381 | * subpage of huge page to split. Split doesn't change the position of @page. | |
3382 | * | |
3383 | * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. | |
3384 | * The huge page must be locked. | |
3385 | * | |
3386 | * If @list is null, tail pages will be added to LRU list, otherwise, to @list. | |
3387 | * | |
3388 | * Both head page and tail pages will inherit mapping, flags, and so on from | |
3389 | * the hugepage. | |
3390 | * | |
3391 | * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if | |
3392 | * they are not mapped. | |
3393 | * | |
3394 | * Returns 0 if the hugepage is split successfully. | |
3395 | * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under | |
3396 | * us. | |
3397 | */ | |
3398 | int split_huge_page_to_list(struct page *page, struct list_head *list) | |
3399 | { | |
3400 | struct page *head = compound_head(page); | |
3401 | struct pglist_data *pgdata = NODE_DATA(page_to_nid(head)); | |
3402 | struct anon_vma *anon_vma; | |
3403 | int count, mapcount, ret; | |
3404 | bool mlocked; | |
3405 | unsigned long flags; | |
3406 | ||
3407 | VM_BUG_ON_PAGE(is_huge_zero_page(page), page); | |
3408 | VM_BUG_ON_PAGE(!PageAnon(page), page); | |
3409 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
3410 | VM_BUG_ON_PAGE(!PageSwapBacked(page), page); | |
3411 | VM_BUG_ON_PAGE(!PageCompound(page), page); | |
3412 | ||
3413 | /* | |
3414 | * The caller does not necessarily hold an mmap_sem that would prevent | |
3415 | * the anon_vma disappearing so we first we take a reference to it | |
3416 | * and then lock the anon_vma for write. This is similar to | |
3417 | * page_lock_anon_vma_read except the write lock is taken to serialise | |
3418 | * against parallel split or collapse operations. | |
3419 | */ | |
3420 | anon_vma = page_get_anon_vma(head); | |
3421 | if (!anon_vma) { | |
3422 | ret = -EBUSY; | |
3423 | goto out; | |
3424 | } | |
3425 | anon_vma_lock_write(anon_vma); | |
3426 | ||
3427 | /* | |
3428 | * Racy check if we can split the page, before freeze_page() will | |
3429 | * split PMDs | |
3430 | */ | |
3431 | if (total_mapcount(head) != page_count(head) - 1) { | |
3432 | ret = -EBUSY; | |
3433 | goto out_unlock; | |
3434 | } | |
3435 | ||
3436 | mlocked = PageMlocked(page); | |
3437 | freeze_page(head); | |
3438 | VM_BUG_ON_PAGE(compound_mapcount(head), head); | |
3439 | ||
3440 | /* Make sure the page is not on per-CPU pagevec as it takes pin */ | |
3441 | if (mlocked) | |
3442 | lru_add_drain(); | |
3443 | ||
3444 | /* Prevent deferred_split_scan() touching ->_refcount */ | |
3445 | spin_lock_irqsave(&pgdata->split_queue_lock, flags); | |
3446 | count = page_count(head); | |
3447 | mapcount = total_mapcount(head); | |
3448 | if (!mapcount && count == 1) { | |
3449 | if (!list_empty(page_deferred_list(head))) { | |
3450 | pgdata->split_queue_len--; | |
3451 | list_del(page_deferred_list(head)); | |
3452 | } | |
3453 | spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); | |
3454 | __split_huge_page(page, list); | |
3455 | ret = 0; | |
3456 | } else if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) { | |
3457 | spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); | |
3458 | pr_alert("total_mapcount: %u, page_count(): %u\n", | |
3459 | mapcount, count); | |
3460 | if (PageTail(page)) | |
3461 | dump_page(head, NULL); | |
3462 | dump_page(page, "total_mapcount(head) > 0"); | |
3463 | BUG(); | |
3464 | } else { | |
3465 | spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); | |
3466 | unfreeze_page(head); | |
3467 | ret = -EBUSY; | |
3468 | } | |
3469 | ||
3470 | out_unlock: | |
3471 | anon_vma_unlock_write(anon_vma); | |
3472 | put_anon_vma(anon_vma); | |
3473 | out: | |
3474 | count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); | |
3475 | return ret; | |
3476 | } | |
3477 | ||
3478 | void free_transhuge_page(struct page *page) | |
3479 | { | |
3480 | struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); | |
3481 | unsigned long flags; | |
3482 | ||
3483 | spin_lock_irqsave(&pgdata->split_queue_lock, flags); | |
3484 | if (!list_empty(page_deferred_list(page))) { | |
3485 | pgdata->split_queue_len--; | |
3486 | list_del(page_deferred_list(page)); | |
3487 | } | |
3488 | spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); | |
3489 | free_compound_page(page); | |
3490 | } | |
3491 | ||
3492 | void deferred_split_huge_page(struct page *page) | |
3493 | { | |
3494 | struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); | |
3495 | unsigned long flags; | |
3496 | ||
3497 | VM_BUG_ON_PAGE(!PageTransHuge(page), page); | |
3498 | ||
3499 | spin_lock_irqsave(&pgdata->split_queue_lock, flags); | |
3500 | if (list_empty(page_deferred_list(page))) { | |
3501 | count_vm_event(THP_DEFERRED_SPLIT_PAGE); | |
3502 | list_add_tail(page_deferred_list(page), &pgdata->split_queue); | |
3503 | pgdata->split_queue_len++; | |
3504 | } | |
3505 | spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); | |
3506 | } | |
3507 | ||
3508 | static unsigned long deferred_split_count(struct shrinker *shrink, | |
3509 | struct shrink_control *sc) | |
3510 | { | |
3511 | struct pglist_data *pgdata = NODE_DATA(sc->nid); | |
3512 | return ACCESS_ONCE(pgdata->split_queue_len); | |
3513 | } | |
3514 | ||
3515 | static unsigned long deferred_split_scan(struct shrinker *shrink, | |
3516 | struct shrink_control *sc) | |
3517 | { | |
3518 | struct pglist_data *pgdata = NODE_DATA(sc->nid); | |
3519 | unsigned long flags; | |
3520 | LIST_HEAD(list), *pos, *next; | |
3521 | struct page *page; | |
3522 | int split = 0; | |
3523 | ||
3524 | spin_lock_irqsave(&pgdata->split_queue_lock, flags); | |
3525 | /* Take pin on all head pages to avoid freeing them under us */ | |
3526 | list_for_each_safe(pos, next, &pgdata->split_queue) { | |
3527 | page = list_entry((void *)pos, struct page, mapping); | |
3528 | page = compound_head(page); | |
3529 | if (get_page_unless_zero(page)) { | |
3530 | list_move(page_deferred_list(page), &list); | |
3531 | } else { | |
3532 | /* We lost race with put_compound_page() */ | |
3533 | list_del_init(page_deferred_list(page)); | |
3534 | pgdata->split_queue_len--; | |
3535 | } | |
3536 | if (!--sc->nr_to_scan) | |
3537 | break; | |
3538 | } | |
3539 | spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); | |
3540 | ||
3541 | list_for_each_safe(pos, next, &list) { | |
3542 | page = list_entry((void *)pos, struct page, mapping); | |
3543 | lock_page(page); | |
3544 | /* split_huge_page() removes page from list on success */ | |
3545 | if (!split_huge_page(page)) | |
3546 | split++; | |
3547 | unlock_page(page); | |
3548 | put_page(page); | |
3549 | } | |
3550 | ||
3551 | spin_lock_irqsave(&pgdata->split_queue_lock, flags); | |
3552 | list_splice_tail(&list, &pgdata->split_queue); | |
3553 | spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); | |
3554 | ||
3555 | /* | |
3556 | * Stop shrinker if we didn't split any page, but the queue is empty. | |
3557 | * This can happen if pages were freed under us. | |
3558 | */ | |
3559 | if (!split && list_empty(&pgdata->split_queue)) | |
3560 | return SHRINK_STOP; | |
3561 | return split; | |
3562 | } | |
3563 | ||
3564 | static struct shrinker deferred_split_shrinker = { | |
3565 | .count_objects = deferred_split_count, | |
3566 | .scan_objects = deferred_split_scan, | |
3567 | .seeks = DEFAULT_SEEKS, | |
3568 | .flags = SHRINKER_NUMA_AWARE, | |
3569 | }; | |
3570 | ||
3571 | #ifdef CONFIG_DEBUG_FS | |
3572 | static int split_huge_pages_set(void *data, u64 val) | |
3573 | { | |
3574 | struct zone *zone; | |
3575 | struct page *page; | |
3576 | unsigned long pfn, max_zone_pfn; | |
3577 | unsigned long total = 0, split = 0; | |
3578 | ||
3579 | if (val != 1) | |
3580 | return -EINVAL; | |
3581 | ||
3582 | for_each_populated_zone(zone) { | |
3583 | max_zone_pfn = zone_end_pfn(zone); | |
3584 | for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { | |
3585 | if (!pfn_valid(pfn)) | |
3586 | continue; | |
3587 | ||
3588 | page = pfn_to_page(pfn); | |
3589 | if (!get_page_unless_zero(page)) | |
3590 | continue; | |
3591 | ||
3592 | if (zone != page_zone(page)) | |
3593 | goto next; | |
3594 | ||
3595 | if (!PageHead(page) || !PageAnon(page) || | |
3596 | PageHuge(page)) | |
3597 | goto next; | |
3598 | ||
3599 | total++; | |
3600 | lock_page(page); | |
3601 | if (!split_huge_page(page)) | |
3602 | split++; | |
3603 | unlock_page(page); | |
3604 | next: | |
3605 | put_page(page); | |
3606 | } | |
3607 | } | |
3608 | ||
3609 | pr_info("%lu of %lu THP split\n", split, total); | |
3610 | ||
3611 | return 0; | |
3612 | } | |
3613 | DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set, | |
3614 | "%llu\n"); | |
3615 | ||
3616 | static int __init split_huge_pages_debugfs(void) | |
3617 | { | |
3618 | void *ret; | |
3619 | ||
3620 | ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL, | |
3621 | &split_huge_pages_fops); | |
3622 | if (!ret) | |
3623 | pr_warn("Failed to create split_huge_pages in debugfs"); | |
3624 | return 0; | |
3625 | } | |
3626 | late_initcall(split_huge_pages_debugfs); | |
3627 | #endif |