Commit | Line | Data |
---|---|---|
71e3aac0 AA |
1 | /* |
2 | * Copyright (C) 2009 Red Hat, Inc. | |
3 | * | |
4 | * This work is licensed under the terms of the GNU GPL, version 2. See | |
5 | * the COPYING file in the top-level directory. | |
6 | */ | |
7 | ||
8 | #include <linux/mm.h> | |
9 | #include <linux/sched.h> | |
10 | #include <linux/highmem.h> | |
11 | #include <linux/hugetlb.h> | |
12 | #include <linux/mmu_notifier.h> | |
13 | #include <linux/rmap.h> | |
14 | #include <linux/swap.h> | |
ba76149f AA |
15 | #include <linux/mm_inline.h> |
16 | #include <linux/kthread.h> | |
17 | #include <linux/khugepaged.h> | |
71e3aac0 AA |
18 | #include <asm/tlb.h> |
19 | #include <asm/pgalloc.h> | |
20 | #include "internal.h" | |
21 | ||
ba76149f AA |
22 | /* |
23 | * By default transparent hugepage support is enabled for all mappings | |
24 | * and khugepaged scans all mappings. Defrag is only invoked by | |
25 | * khugepaged hugepage allocations and by page faults inside | |
26 | * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived | |
27 | * allocations. | |
28 | */ | |
71e3aac0 | 29 | unsigned long transparent_hugepage_flags __read_mostly = |
13ece886 | 30 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS |
ba76149f | 31 | (1<<TRANSPARENT_HUGEPAGE_FLAG)| |
13ece886 AA |
32 | #endif |
33 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE | |
34 | (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| | |
35 | #endif | |
d39d33c3 | 36 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)| |
ba76149f AA |
37 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); |
38 | ||
39 | /* default scan 8*512 pte (or vmas) every 30 second */ | |
40 | static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8; | |
41 | static unsigned int khugepaged_pages_collapsed; | |
42 | static unsigned int khugepaged_full_scans; | |
43 | static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; | |
44 | /* during fragmentation poll the hugepage allocator once every minute */ | |
45 | static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; | |
46 | static struct task_struct *khugepaged_thread __read_mostly; | |
47 | static DEFINE_MUTEX(khugepaged_mutex); | |
48 | static DEFINE_SPINLOCK(khugepaged_mm_lock); | |
49 | static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); | |
50 | /* | |
51 | * default collapse hugepages if there is at least one pte mapped like | |
52 | * it would have happened if the vma was large enough during page | |
53 | * fault. | |
54 | */ | |
55 | static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1; | |
56 | ||
57 | static int khugepaged(void *none); | |
58 | static int mm_slots_hash_init(void); | |
59 | static int khugepaged_slab_init(void); | |
60 | static void khugepaged_slab_free(void); | |
61 | ||
62 | #define MM_SLOTS_HASH_HEADS 1024 | |
63 | static struct hlist_head *mm_slots_hash __read_mostly; | |
64 | static struct kmem_cache *mm_slot_cache __read_mostly; | |
65 | ||
66 | /** | |
67 | * struct mm_slot - hash lookup from mm to mm_slot | |
68 | * @hash: hash collision list | |
69 | * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head | |
70 | * @mm: the mm that this information is valid for | |
71 | */ | |
72 | struct mm_slot { | |
73 | struct hlist_node hash; | |
74 | struct list_head mm_node; | |
75 | struct mm_struct *mm; | |
76 | }; | |
77 | ||
78 | /** | |
79 | * struct khugepaged_scan - cursor for scanning | |
80 | * @mm_head: the head of the mm list to scan | |
81 | * @mm_slot: the current mm_slot we are scanning | |
82 | * @address: the next address inside that to be scanned | |
83 | * | |
84 | * There is only the one khugepaged_scan instance of this cursor structure. | |
85 | */ | |
86 | struct khugepaged_scan { | |
87 | struct list_head mm_head; | |
88 | struct mm_slot *mm_slot; | |
89 | unsigned long address; | |
90 | } khugepaged_scan = { | |
91 | .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), | |
92 | }; | |
93 | ||
f000565a AA |
94 | |
95 | static int set_recommended_min_free_kbytes(void) | |
96 | { | |
97 | struct zone *zone; | |
98 | int nr_zones = 0; | |
99 | unsigned long recommended_min; | |
100 | extern int min_free_kbytes; | |
101 | ||
102 | if (!test_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
103 | &transparent_hugepage_flags) && | |
104 | !test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
105 | &transparent_hugepage_flags)) | |
106 | return 0; | |
107 | ||
108 | for_each_populated_zone(zone) | |
109 | nr_zones++; | |
110 | ||
111 | /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */ | |
112 | recommended_min = pageblock_nr_pages * nr_zones * 2; | |
113 | ||
114 | /* | |
115 | * Make sure that on average at least two pageblocks are almost free | |
116 | * of another type, one for a migratetype to fall back to and a | |
117 | * second to avoid subsequent fallbacks of other types There are 3 | |
118 | * MIGRATE_TYPES we care about. | |
119 | */ | |
120 | recommended_min += pageblock_nr_pages * nr_zones * | |
121 | MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; | |
122 | ||
123 | /* don't ever allow to reserve more than 5% of the lowmem */ | |
124 | recommended_min = min(recommended_min, | |
125 | (unsigned long) nr_free_buffer_pages() / 20); | |
126 | recommended_min <<= (PAGE_SHIFT-10); | |
127 | ||
128 | if (recommended_min > min_free_kbytes) | |
129 | min_free_kbytes = recommended_min; | |
130 | setup_per_zone_wmarks(); | |
131 | return 0; | |
132 | } | |
133 | late_initcall(set_recommended_min_free_kbytes); | |
134 | ||
ba76149f AA |
135 | static int start_khugepaged(void) |
136 | { | |
137 | int err = 0; | |
138 | if (khugepaged_enabled()) { | |
139 | int wakeup; | |
140 | if (unlikely(!mm_slot_cache || !mm_slots_hash)) { | |
141 | err = -ENOMEM; | |
142 | goto out; | |
143 | } | |
144 | mutex_lock(&khugepaged_mutex); | |
145 | if (!khugepaged_thread) | |
146 | khugepaged_thread = kthread_run(khugepaged, NULL, | |
147 | "khugepaged"); | |
148 | if (unlikely(IS_ERR(khugepaged_thread))) { | |
149 | printk(KERN_ERR | |
150 | "khugepaged: kthread_run(khugepaged) failed\n"); | |
151 | err = PTR_ERR(khugepaged_thread); | |
152 | khugepaged_thread = NULL; | |
153 | } | |
154 | wakeup = !list_empty(&khugepaged_scan.mm_head); | |
155 | mutex_unlock(&khugepaged_mutex); | |
156 | if (wakeup) | |
157 | wake_up_interruptible(&khugepaged_wait); | |
f000565a AA |
158 | |
159 | set_recommended_min_free_kbytes(); | |
ba76149f AA |
160 | } else |
161 | /* wakeup to exit */ | |
162 | wake_up_interruptible(&khugepaged_wait); | |
163 | out: | |
164 | return err; | |
165 | } | |
71e3aac0 AA |
166 | |
167 | #ifdef CONFIG_SYSFS | |
ba76149f | 168 | |
71e3aac0 AA |
169 | static ssize_t double_flag_show(struct kobject *kobj, |
170 | struct kobj_attribute *attr, char *buf, | |
171 | enum transparent_hugepage_flag enabled, | |
172 | enum transparent_hugepage_flag req_madv) | |
173 | { | |
174 | if (test_bit(enabled, &transparent_hugepage_flags)) { | |
175 | VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags)); | |
176 | return sprintf(buf, "[always] madvise never\n"); | |
177 | } else if (test_bit(req_madv, &transparent_hugepage_flags)) | |
178 | return sprintf(buf, "always [madvise] never\n"); | |
179 | else | |
180 | return sprintf(buf, "always madvise [never]\n"); | |
181 | } | |
182 | static ssize_t double_flag_store(struct kobject *kobj, | |
183 | struct kobj_attribute *attr, | |
184 | const char *buf, size_t count, | |
185 | enum transparent_hugepage_flag enabled, | |
186 | enum transparent_hugepage_flag req_madv) | |
187 | { | |
188 | if (!memcmp("always", buf, | |
189 | min(sizeof("always")-1, count))) { | |
190 | set_bit(enabled, &transparent_hugepage_flags); | |
191 | clear_bit(req_madv, &transparent_hugepage_flags); | |
192 | } else if (!memcmp("madvise", buf, | |
193 | min(sizeof("madvise")-1, count))) { | |
194 | clear_bit(enabled, &transparent_hugepage_flags); | |
195 | set_bit(req_madv, &transparent_hugepage_flags); | |
196 | } else if (!memcmp("never", buf, | |
197 | min(sizeof("never")-1, count))) { | |
198 | clear_bit(enabled, &transparent_hugepage_flags); | |
199 | clear_bit(req_madv, &transparent_hugepage_flags); | |
200 | } else | |
201 | return -EINVAL; | |
202 | ||
203 | return count; | |
204 | } | |
205 | ||
206 | static ssize_t enabled_show(struct kobject *kobj, | |
207 | struct kobj_attribute *attr, char *buf) | |
208 | { | |
209 | return double_flag_show(kobj, attr, buf, | |
210 | TRANSPARENT_HUGEPAGE_FLAG, | |
211 | TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); | |
212 | } | |
213 | static ssize_t enabled_store(struct kobject *kobj, | |
214 | struct kobj_attribute *attr, | |
215 | const char *buf, size_t count) | |
216 | { | |
ba76149f AA |
217 | ssize_t ret; |
218 | ||
219 | ret = double_flag_store(kobj, attr, buf, count, | |
220 | TRANSPARENT_HUGEPAGE_FLAG, | |
221 | TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); | |
222 | ||
223 | if (ret > 0) { | |
224 | int err = start_khugepaged(); | |
225 | if (err) | |
226 | ret = err; | |
227 | } | |
228 | ||
f000565a AA |
229 | if (ret > 0 && |
230 | (test_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
231 | &transparent_hugepage_flags) || | |
232 | test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
233 | &transparent_hugepage_flags))) | |
234 | set_recommended_min_free_kbytes(); | |
235 | ||
ba76149f | 236 | return ret; |
71e3aac0 AA |
237 | } |
238 | static struct kobj_attribute enabled_attr = | |
239 | __ATTR(enabled, 0644, enabled_show, enabled_store); | |
240 | ||
241 | static ssize_t single_flag_show(struct kobject *kobj, | |
242 | struct kobj_attribute *attr, char *buf, | |
243 | enum transparent_hugepage_flag flag) | |
244 | { | |
245 | if (test_bit(flag, &transparent_hugepage_flags)) | |
246 | return sprintf(buf, "[yes] no\n"); | |
247 | else | |
248 | return sprintf(buf, "yes [no]\n"); | |
249 | } | |
250 | static ssize_t single_flag_store(struct kobject *kobj, | |
251 | struct kobj_attribute *attr, | |
252 | const char *buf, size_t count, | |
253 | enum transparent_hugepage_flag flag) | |
254 | { | |
255 | if (!memcmp("yes", buf, | |
256 | min(sizeof("yes")-1, count))) { | |
257 | set_bit(flag, &transparent_hugepage_flags); | |
258 | } else if (!memcmp("no", buf, | |
259 | min(sizeof("no")-1, count))) { | |
260 | clear_bit(flag, &transparent_hugepage_flags); | |
261 | } else | |
262 | return -EINVAL; | |
263 | ||
264 | return count; | |
265 | } | |
266 | ||
267 | /* | |
268 | * Currently defrag only disables __GFP_NOWAIT for allocation. A blind | |
269 | * __GFP_REPEAT is too aggressive, it's never worth swapping tons of | |
270 | * memory just to allocate one more hugepage. | |
271 | */ | |
272 | static ssize_t defrag_show(struct kobject *kobj, | |
273 | struct kobj_attribute *attr, char *buf) | |
274 | { | |
275 | return double_flag_show(kobj, attr, buf, | |
276 | TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, | |
277 | TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); | |
278 | } | |
279 | static ssize_t defrag_store(struct kobject *kobj, | |
280 | struct kobj_attribute *attr, | |
281 | const char *buf, size_t count) | |
282 | { | |
283 | return double_flag_store(kobj, attr, buf, count, | |
284 | TRANSPARENT_HUGEPAGE_DEFRAG_FLAG, | |
285 | TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); | |
286 | } | |
287 | static struct kobj_attribute defrag_attr = | |
288 | __ATTR(defrag, 0644, defrag_show, defrag_store); | |
289 | ||
290 | #ifdef CONFIG_DEBUG_VM | |
291 | static ssize_t debug_cow_show(struct kobject *kobj, | |
292 | struct kobj_attribute *attr, char *buf) | |
293 | { | |
294 | return single_flag_show(kobj, attr, buf, | |
295 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); | |
296 | } | |
297 | static ssize_t debug_cow_store(struct kobject *kobj, | |
298 | struct kobj_attribute *attr, | |
299 | const char *buf, size_t count) | |
300 | { | |
301 | return single_flag_store(kobj, attr, buf, count, | |
302 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); | |
303 | } | |
304 | static struct kobj_attribute debug_cow_attr = | |
305 | __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); | |
306 | #endif /* CONFIG_DEBUG_VM */ | |
307 | ||
308 | static struct attribute *hugepage_attr[] = { | |
309 | &enabled_attr.attr, | |
310 | &defrag_attr.attr, | |
311 | #ifdef CONFIG_DEBUG_VM | |
312 | &debug_cow_attr.attr, | |
313 | #endif | |
314 | NULL, | |
315 | }; | |
316 | ||
317 | static struct attribute_group hugepage_attr_group = { | |
318 | .attrs = hugepage_attr, | |
ba76149f AA |
319 | }; |
320 | ||
321 | static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, | |
322 | struct kobj_attribute *attr, | |
323 | char *buf) | |
324 | { | |
325 | return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs); | |
326 | } | |
327 | ||
328 | static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, | |
329 | struct kobj_attribute *attr, | |
330 | const char *buf, size_t count) | |
331 | { | |
332 | unsigned long msecs; | |
333 | int err; | |
334 | ||
335 | err = strict_strtoul(buf, 10, &msecs); | |
336 | if (err || msecs > UINT_MAX) | |
337 | return -EINVAL; | |
338 | ||
339 | khugepaged_scan_sleep_millisecs = msecs; | |
340 | wake_up_interruptible(&khugepaged_wait); | |
341 | ||
342 | return count; | |
343 | } | |
344 | static struct kobj_attribute scan_sleep_millisecs_attr = | |
345 | __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show, | |
346 | scan_sleep_millisecs_store); | |
347 | ||
348 | static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, | |
349 | struct kobj_attribute *attr, | |
350 | char *buf) | |
351 | { | |
352 | return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs); | |
353 | } | |
354 | ||
355 | static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, | |
356 | struct kobj_attribute *attr, | |
357 | const char *buf, size_t count) | |
358 | { | |
359 | unsigned long msecs; | |
360 | int err; | |
361 | ||
362 | err = strict_strtoul(buf, 10, &msecs); | |
363 | if (err || msecs > UINT_MAX) | |
364 | return -EINVAL; | |
365 | ||
366 | khugepaged_alloc_sleep_millisecs = msecs; | |
367 | wake_up_interruptible(&khugepaged_wait); | |
368 | ||
369 | return count; | |
370 | } | |
371 | static struct kobj_attribute alloc_sleep_millisecs_attr = | |
372 | __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show, | |
373 | alloc_sleep_millisecs_store); | |
374 | ||
375 | static ssize_t pages_to_scan_show(struct kobject *kobj, | |
376 | struct kobj_attribute *attr, | |
377 | char *buf) | |
378 | { | |
379 | return sprintf(buf, "%u\n", khugepaged_pages_to_scan); | |
380 | } | |
381 | static ssize_t pages_to_scan_store(struct kobject *kobj, | |
382 | struct kobj_attribute *attr, | |
383 | const char *buf, size_t count) | |
384 | { | |
385 | int err; | |
386 | unsigned long pages; | |
387 | ||
388 | err = strict_strtoul(buf, 10, &pages); | |
389 | if (err || !pages || pages > UINT_MAX) | |
390 | return -EINVAL; | |
391 | ||
392 | khugepaged_pages_to_scan = pages; | |
393 | ||
394 | return count; | |
395 | } | |
396 | static struct kobj_attribute pages_to_scan_attr = | |
397 | __ATTR(pages_to_scan, 0644, pages_to_scan_show, | |
398 | pages_to_scan_store); | |
399 | ||
400 | static ssize_t pages_collapsed_show(struct kobject *kobj, | |
401 | struct kobj_attribute *attr, | |
402 | char *buf) | |
403 | { | |
404 | return sprintf(buf, "%u\n", khugepaged_pages_collapsed); | |
405 | } | |
406 | static struct kobj_attribute pages_collapsed_attr = | |
407 | __ATTR_RO(pages_collapsed); | |
408 | ||
409 | static ssize_t full_scans_show(struct kobject *kobj, | |
410 | struct kobj_attribute *attr, | |
411 | char *buf) | |
412 | { | |
413 | return sprintf(buf, "%u\n", khugepaged_full_scans); | |
414 | } | |
415 | static struct kobj_attribute full_scans_attr = | |
416 | __ATTR_RO(full_scans); | |
417 | ||
418 | static ssize_t khugepaged_defrag_show(struct kobject *kobj, | |
419 | struct kobj_attribute *attr, char *buf) | |
420 | { | |
421 | return single_flag_show(kobj, attr, buf, | |
422 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); | |
423 | } | |
424 | static ssize_t khugepaged_defrag_store(struct kobject *kobj, | |
425 | struct kobj_attribute *attr, | |
426 | const char *buf, size_t count) | |
427 | { | |
428 | return single_flag_store(kobj, attr, buf, count, | |
429 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); | |
430 | } | |
431 | static struct kobj_attribute khugepaged_defrag_attr = | |
432 | __ATTR(defrag, 0644, khugepaged_defrag_show, | |
433 | khugepaged_defrag_store); | |
434 | ||
435 | /* | |
436 | * max_ptes_none controls if khugepaged should collapse hugepages over | |
437 | * any unmapped ptes in turn potentially increasing the memory | |
438 | * footprint of the vmas. When max_ptes_none is 0 khugepaged will not | |
439 | * reduce the available free memory in the system as it | |
440 | * runs. Increasing max_ptes_none will instead potentially reduce the | |
441 | * free memory in the system during the khugepaged scan. | |
442 | */ | |
443 | static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj, | |
444 | struct kobj_attribute *attr, | |
445 | char *buf) | |
446 | { | |
447 | return sprintf(buf, "%u\n", khugepaged_max_ptes_none); | |
448 | } | |
449 | static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj, | |
450 | struct kobj_attribute *attr, | |
451 | const char *buf, size_t count) | |
452 | { | |
453 | int err; | |
454 | unsigned long max_ptes_none; | |
455 | ||
456 | err = strict_strtoul(buf, 10, &max_ptes_none); | |
457 | if (err || max_ptes_none > HPAGE_PMD_NR-1) | |
458 | return -EINVAL; | |
459 | ||
460 | khugepaged_max_ptes_none = max_ptes_none; | |
461 | ||
462 | return count; | |
463 | } | |
464 | static struct kobj_attribute khugepaged_max_ptes_none_attr = | |
465 | __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show, | |
466 | khugepaged_max_ptes_none_store); | |
467 | ||
468 | static struct attribute *khugepaged_attr[] = { | |
469 | &khugepaged_defrag_attr.attr, | |
470 | &khugepaged_max_ptes_none_attr.attr, | |
471 | &pages_to_scan_attr.attr, | |
472 | &pages_collapsed_attr.attr, | |
473 | &full_scans_attr.attr, | |
474 | &scan_sleep_millisecs_attr.attr, | |
475 | &alloc_sleep_millisecs_attr.attr, | |
476 | NULL, | |
477 | }; | |
478 | ||
479 | static struct attribute_group khugepaged_attr_group = { | |
480 | .attrs = khugepaged_attr, | |
481 | .name = "khugepaged", | |
71e3aac0 AA |
482 | }; |
483 | #endif /* CONFIG_SYSFS */ | |
484 | ||
485 | static int __init hugepage_init(void) | |
486 | { | |
71e3aac0 | 487 | int err; |
ba76149f AA |
488 | #ifdef CONFIG_SYSFS |
489 | static struct kobject *hugepage_kobj; | |
71e3aac0 | 490 | |
ba76149f AA |
491 | err = -ENOMEM; |
492 | hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); | |
493 | if (unlikely(!hugepage_kobj)) { | |
494 | printk(KERN_ERR "hugepage: failed kobject create\n"); | |
495 | goto out; | |
496 | } | |
497 | ||
498 | err = sysfs_create_group(hugepage_kobj, &hugepage_attr_group); | |
499 | if (err) { | |
500 | printk(KERN_ERR "hugepage: failed register hugeage group\n"); | |
501 | goto out; | |
502 | } | |
503 | ||
504 | err = sysfs_create_group(hugepage_kobj, &khugepaged_attr_group); | |
505 | if (err) { | |
506 | printk(KERN_ERR "hugepage: failed register hugeage group\n"); | |
507 | goto out; | |
508 | } | |
71e3aac0 | 509 | #endif |
ba76149f AA |
510 | |
511 | err = khugepaged_slab_init(); | |
512 | if (err) | |
513 | goto out; | |
514 | ||
515 | err = mm_slots_hash_init(); | |
516 | if (err) { | |
517 | khugepaged_slab_free(); | |
518 | goto out; | |
519 | } | |
520 | ||
521 | start_khugepaged(); | |
522 | ||
f000565a AA |
523 | set_recommended_min_free_kbytes(); |
524 | ||
ba76149f AA |
525 | out: |
526 | return err; | |
71e3aac0 AA |
527 | } |
528 | module_init(hugepage_init) | |
529 | ||
530 | static int __init setup_transparent_hugepage(char *str) | |
531 | { | |
532 | int ret = 0; | |
533 | if (!str) | |
534 | goto out; | |
535 | if (!strcmp(str, "always")) { | |
536 | set_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
537 | &transparent_hugepage_flags); | |
538 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
539 | &transparent_hugepage_flags); | |
540 | ret = 1; | |
541 | } else if (!strcmp(str, "madvise")) { | |
542 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
543 | &transparent_hugepage_flags); | |
544 | set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
545 | &transparent_hugepage_flags); | |
546 | ret = 1; | |
547 | } else if (!strcmp(str, "never")) { | |
548 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
549 | &transparent_hugepage_flags); | |
550 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
551 | &transparent_hugepage_flags); | |
552 | ret = 1; | |
553 | } | |
554 | out: | |
555 | if (!ret) | |
556 | printk(KERN_WARNING | |
557 | "transparent_hugepage= cannot parse, ignored\n"); | |
558 | return ret; | |
559 | } | |
560 | __setup("transparent_hugepage=", setup_transparent_hugepage); | |
561 | ||
562 | static void prepare_pmd_huge_pte(pgtable_t pgtable, | |
563 | struct mm_struct *mm) | |
564 | { | |
565 | assert_spin_locked(&mm->page_table_lock); | |
566 | ||
567 | /* FIFO */ | |
568 | if (!mm->pmd_huge_pte) | |
569 | INIT_LIST_HEAD(&pgtable->lru); | |
570 | else | |
571 | list_add(&pgtable->lru, &mm->pmd_huge_pte->lru); | |
572 | mm->pmd_huge_pte = pgtable; | |
573 | } | |
574 | ||
575 | static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) | |
576 | { | |
577 | if (likely(vma->vm_flags & VM_WRITE)) | |
578 | pmd = pmd_mkwrite(pmd); | |
579 | return pmd; | |
580 | } | |
581 | ||
582 | static int __do_huge_pmd_anonymous_page(struct mm_struct *mm, | |
583 | struct vm_area_struct *vma, | |
584 | unsigned long haddr, pmd_t *pmd, | |
585 | struct page *page) | |
586 | { | |
587 | int ret = 0; | |
588 | pgtable_t pgtable; | |
589 | ||
590 | VM_BUG_ON(!PageCompound(page)); | |
591 | pgtable = pte_alloc_one(mm, haddr); | |
592 | if (unlikely(!pgtable)) { | |
b9bbfbe3 | 593 | mem_cgroup_uncharge_page(page); |
71e3aac0 AA |
594 | put_page(page); |
595 | return VM_FAULT_OOM; | |
596 | } | |
597 | ||
598 | clear_huge_page(page, haddr, HPAGE_PMD_NR); | |
599 | __SetPageUptodate(page); | |
600 | ||
601 | spin_lock(&mm->page_table_lock); | |
602 | if (unlikely(!pmd_none(*pmd))) { | |
603 | spin_unlock(&mm->page_table_lock); | |
b9bbfbe3 | 604 | mem_cgroup_uncharge_page(page); |
71e3aac0 AA |
605 | put_page(page); |
606 | pte_free(mm, pgtable); | |
607 | } else { | |
608 | pmd_t entry; | |
609 | entry = mk_pmd(page, vma->vm_page_prot); | |
610 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
611 | entry = pmd_mkhuge(entry); | |
612 | /* | |
613 | * The spinlocking to take the lru_lock inside | |
614 | * page_add_new_anon_rmap() acts as a full memory | |
615 | * barrier to be sure clear_huge_page writes become | |
616 | * visible after the set_pmd_at() write. | |
617 | */ | |
618 | page_add_new_anon_rmap(page, vma, haddr); | |
619 | set_pmd_at(mm, haddr, pmd, entry); | |
620 | prepare_pmd_huge_pte(pgtable, mm); | |
621 | add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR); | |
622 | spin_unlock(&mm->page_table_lock); | |
623 | } | |
624 | ||
625 | return ret; | |
626 | } | |
627 | ||
0bbbc0b3 AA |
628 | static inline gfp_t alloc_hugepage_gfpmask(int defrag) |
629 | { | |
630 | return GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT); | |
631 | } | |
632 | ||
633 | static inline struct page *alloc_hugepage_vma(int defrag, | |
634 | struct vm_area_struct *vma, | |
635 | unsigned long haddr) | |
636 | { | |
637 | return alloc_pages_vma(alloc_hugepage_gfpmask(defrag), | |
638 | HPAGE_PMD_ORDER, vma, haddr); | |
639 | } | |
640 | ||
641 | #ifndef CONFIG_NUMA | |
71e3aac0 AA |
642 | static inline struct page *alloc_hugepage(int defrag) |
643 | { | |
0bbbc0b3 | 644 | return alloc_pages(alloc_hugepage_gfpmask(defrag), |
71e3aac0 AA |
645 | HPAGE_PMD_ORDER); |
646 | } | |
0bbbc0b3 | 647 | #endif |
71e3aac0 AA |
648 | |
649 | int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma, | |
650 | unsigned long address, pmd_t *pmd, | |
651 | unsigned int flags) | |
652 | { | |
653 | struct page *page; | |
654 | unsigned long haddr = address & HPAGE_PMD_MASK; | |
655 | pte_t *pte; | |
656 | ||
657 | if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) { | |
658 | if (unlikely(anon_vma_prepare(vma))) | |
659 | return VM_FAULT_OOM; | |
ba76149f AA |
660 | if (unlikely(khugepaged_enter(vma))) |
661 | return VM_FAULT_OOM; | |
0bbbc0b3 AA |
662 | page = alloc_hugepage_vma(transparent_hugepage_defrag(vma), |
663 | vma, haddr); | |
71e3aac0 AA |
664 | if (unlikely(!page)) |
665 | goto out; | |
b9bbfbe3 AA |
666 | if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) { |
667 | put_page(page); | |
668 | goto out; | |
669 | } | |
71e3aac0 AA |
670 | |
671 | return __do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page); | |
672 | } | |
673 | out: | |
674 | /* | |
675 | * Use __pte_alloc instead of pte_alloc_map, because we can't | |
676 | * run pte_offset_map on the pmd, if an huge pmd could | |
677 | * materialize from under us from a different thread. | |
678 | */ | |
679 | if (unlikely(__pte_alloc(mm, vma, pmd, address))) | |
680 | return VM_FAULT_OOM; | |
681 | /* if an huge pmd materialized from under us just retry later */ | |
682 | if (unlikely(pmd_trans_huge(*pmd))) | |
683 | return 0; | |
684 | /* | |
685 | * A regular pmd is established and it can't morph into a huge pmd | |
686 | * from under us anymore at this point because we hold the mmap_sem | |
687 | * read mode and khugepaged takes it in write mode. So now it's | |
688 | * safe to run pte_offset_map(). | |
689 | */ | |
690 | pte = pte_offset_map(pmd, address); | |
691 | return handle_pte_fault(mm, vma, address, pte, pmd, flags); | |
692 | } | |
693 | ||
694 | int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, | |
695 | pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, | |
696 | struct vm_area_struct *vma) | |
697 | { | |
698 | struct page *src_page; | |
699 | pmd_t pmd; | |
700 | pgtable_t pgtable; | |
701 | int ret; | |
702 | ||
703 | ret = -ENOMEM; | |
704 | pgtable = pte_alloc_one(dst_mm, addr); | |
705 | if (unlikely(!pgtable)) | |
706 | goto out; | |
707 | ||
708 | spin_lock(&dst_mm->page_table_lock); | |
709 | spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING); | |
710 | ||
711 | ret = -EAGAIN; | |
712 | pmd = *src_pmd; | |
713 | if (unlikely(!pmd_trans_huge(pmd))) { | |
714 | pte_free(dst_mm, pgtable); | |
715 | goto out_unlock; | |
716 | } | |
717 | if (unlikely(pmd_trans_splitting(pmd))) { | |
718 | /* split huge page running from under us */ | |
719 | spin_unlock(&src_mm->page_table_lock); | |
720 | spin_unlock(&dst_mm->page_table_lock); | |
721 | pte_free(dst_mm, pgtable); | |
722 | ||
723 | wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */ | |
724 | goto out; | |
725 | } | |
726 | src_page = pmd_page(pmd); | |
727 | VM_BUG_ON(!PageHead(src_page)); | |
728 | get_page(src_page); | |
729 | page_dup_rmap(src_page); | |
730 | add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); | |
731 | ||
732 | pmdp_set_wrprotect(src_mm, addr, src_pmd); | |
733 | pmd = pmd_mkold(pmd_wrprotect(pmd)); | |
734 | set_pmd_at(dst_mm, addr, dst_pmd, pmd); | |
735 | prepare_pmd_huge_pte(pgtable, dst_mm); | |
736 | ||
737 | ret = 0; | |
738 | out_unlock: | |
739 | spin_unlock(&src_mm->page_table_lock); | |
740 | spin_unlock(&dst_mm->page_table_lock); | |
741 | out: | |
742 | return ret; | |
743 | } | |
744 | ||
745 | /* no "address" argument so destroys page coloring of some arch */ | |
746 | pgtable_t get_pmd_huge_pte(struct mm_struct *mm) | |
747 | { | |
748 | pgtable_t pgtable; | |
749 | ||
750 | assert_spin_locked(&mm->page_table_lock); | |
751 | ||
752 | /* FIFO */ | |
753 | pgtable = mm->pmd_huge_pte; | |
754 | if (list_empty(&pgtable->lru)) | |
755 | mm->pmd_huge_pte = NULL; | |
756 | else { | |
757 | mm->pmd_huge_pte = list_entry(pgtable->lru.next, | |
758 | struct page, lru); | |
759 | list_del(&pgtable->lru); | |
760 | } | |
761 | return pgtable; | |
762 | } | |
763 | ||
764 | static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm, | |
765 | struct vm_area_struct *vma, | |
766 | unsigned long address, | |
767 | pmd_t *pmd, pmd_t orig_pmd, | |
768 | struct page *page, | |
769 | unsigned long haddr) | |
770 | { | |
771 | pgtable_t pgtable; | |
772 | pmd_t _pmd; | |
773 | int ret = 0, i; | |
774 | struct page **pages; | |
775 | ||
776 | pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR, | |
777 | GFP_KERNEL); | |
778 | if (unlikely(!pages)) { | |
779 | ret |= VM_FAULT_OOM; | |
780 | goto out; | |
781 | } | |
782 | ||
783 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
784 | pages[i] = alloc_page_vma(GFP_HIGHUSER_MOVABLE, | |
785 | vma, address); | |
b9bbfbe3 AA |
786 | if (unlikely(!pages[i] || |
787 | mem_cgroup_newpage_charge(pages[i], mm, | |
788 | GFP_KERNEL))) { | |
789 | if (pages[i]) | |
71e3aac0 | 790 | put_page(pages[i]); |
b9bbfbe3 AA |
791 | mem_cgroup_uncharge_start(); |
792 | while (--i >= 0) { | |
793 | mem_cgroup_uncharge_page(pages[i]); | |
794 | put_page(pages[i]); | |
795 | } | |
796 | mem_cgroup_uncharge_end(); | |
71e3aac0 AA |
797 | kfree(pages); |
798 | ret |= VM_FAULT_OOM; | |
799 | goto out; | |
800 | } | |
801 | } | |
802 | ||
803 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
804 | copy_user_highpage(pages[i], page + i, | |
805 | haddr + PAGE_SHIFT*i, vma); | |
806 | __SetPageUptodate(pages[i]); | |
807 | cond_resched(); | |
808 | } | |
809 | ||
810 | spin_lock(&mm->page_table_lock); | |
811 | if (unlikely(!pmd_same(*pmd, orig_pmd))) | |
812 | goto out_free_pages; | |
813 | VM_BUG_ON(!PageHead(page)); | |
814 | ||
815 | pmdp_clear_flush_notify(vma, haddr, pmd); | |
816 | /* leave pmd empty until pte is filled */ | |
817 | ||
818 | pgtable = get_pmd_huge_pte(mm); | |
819 | pmd_populate(mm, &_pmd, pgtable); | |
820 | ||
821 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
822 | pte_t *pte, entry; | |
823 | entry = mk_pte(pages[i], vma->vm_page_prot); | |
824 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); | |
825 | page_add_new_anon_rmap(pages[i], vma, haddr); | |
826 | pte = pte_offset_map(&_pmd, haddr); | |
827 | VM_BUG_ON(!pte_none(*pte)); | |
828 | set_pte_at(mm, haddr, pte, entry); | |
829 | pte_unmap(pte); | |
830 | } | |
831 | kfree(pages); | |
832 | ||
833 | mm->nr_ptes++; | |
834 | smp_wmb(); /* make pte visible before pmd */ | |
835 | pmd_populate(mm, pmd, pgtable); | |
836 | page_remove_rmap(page); | |
837 | spin_unlock(&mm->page_table_lock); | |
838 | ||
839 | ret |= VM_FAULT_WRITE; | |
840 | put_page(page); | |
841 | ||
842 | out: | |
843 | return ret; | |
844 | ||
845 | out_free_pages: | |
846 | spin_unlock(&mm->page_table_lock); | |
b9bbfbe3 AA |
847 | mem_cgroup_uncharge_start(); |
848 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
849 | mem_cgroup_uncharge_page(pages[i]); | |
71e3aac0 | 850 | put_page(pages[i]); |
b9bbfbe3 AA |
851 | } |
852 | mem_cgroup_uncharge_end(); | |
71e3aac0 AA |
853 | kfree(pages); |
854 | goto out; | |
855 | } | |
856 | ||
857 | int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma, | |
858 | unsigned long address, pmd_t *pmd, pmd_t orig_pmd) | |
859 | { | |
860 | int ret = 0; | |
861 | struct page *page, *new_page; | |
862 | unsigned long haddr; | |
863 | ||
864 | VM_BUG_ON(!vma->anon_vma); | |
865 | spin_lock(&mm->page_table_lock); | |
866 | if (unlikely(!pmd_same(*pmd, orig_pmd))) | |
867 | goto out_unlock; | |
868 | ||
869 | page = pmd_page(orig_pmd); | |
870 | VM_BUG_ON(!PageCompound(page) || !PageHead(page)); | |
871 | haddr = address & HPAGE_PMD_MASK; | |
872 | if (page_mapcount(page) == 1) { | |
873 | pmd_t entry; | |
874 | entry = pmd_mkyoung(orig_pmd); | |
875 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
876 | if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1)) | |
877 | update_mmu_cache(vma, address, entry); | |
878 | ret |= VM_FAULT_WRITE; | |
879 | goto out_unlock; | |
880 | } | |
881 | get_page(page); | |
882 | spin_unlock(&mm->page_table_lock); | |
883 | ||
884 | if (transparent_hugepage_enabled(vma) && | |
885 | !transparent_hugepage_debug_cow()) | |
0bbbc0b3 AA |
886 | new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma), |
887 | vma, haddr); | |
71e3aac0 AA |
888 | else |
889 | new_page = NULL; | |
890 | ||
891 | if (unlikely(!new_page)) { | |
892 | ret = do_huge_pmd_wp_page_fallback(mm, vma, address, | |
893 | pmd, orig_pmd, page, haddr); | |
894 | put_page(page); | |
895 | goto out; | |
896 | } | |
897 | ||
b9bbfbe3 AA |
898 | if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) { |
899 | put_page(new_page); | |
900 | put_page(page); | |
901 | ret |= VM_FAULT_OOM; | |
902 | goto out; | |
903 | } | |
904 | ||
71e3aac0 AA |
905 | copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR); |
906 | __SetPageUptodate(new_page); | |
907 | ||
908 | spin_lock(&mm->page_table_lock); | |
909 | put_page(page); | |
b9bbfbe3 AA |
910 | if (unlikely(!pmd_same(*pmd, orig_pmd))) { |
911 | mem_cgroup_uncharge_page(new_page); | |
71e3aac0 | 912 | put_page(new_page); |
b9bbfbe3 | 913 | } else { |
71e3aac0 AA |
914 | pmd_t entry; |
915 | VM_BUG_ON(!PageHead(page)); | |
916 | entry = mk_pmd(new_page, vma->vm_page_prot); | |
917 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
918 | entry = pmd_mkhuge(entry); | |
919 | pmdp_clear_flush_notify(vma, haddr, pmd); | |
920 | page_add_new_anon_rmap(new_page, vma, haddr); | |
921 | set_pmd_at(mm, haddr, pmd, entry); | |
922 | update_mmu_cache(vma, address, entry); | |
923 | page_remove_rmap(page); | |
924 | put_page(page); | |
925 | ret |= VM_FAULT_WRITE; | |
926 | } | |
927 | out_unlock: | |
928 | spin_unlock(&mm->page_table_lock); | |
929 | out: | |
930 | return ret; | |
931 | } | |
932 | ||
933 | struct page *follow_trans_huge_pmd(struct mm_struct *mm, | |
934 | unsigned long addr, | |
935 | pmd_t *pmd, | |
936 | unsigned int flags) | |
937 | { | |
938 | struct page *page = NULL; | |
939 | ||
940 | assert_spin_locked(&mm->page_table_lock); | |
941 | ||
942 | if (flags & FOLL_WRITE && !pmd_write(*pmd)) | |
943 | goto out; | |
944 | ||
945 | page = pmd_page(*pmd); | |
946 | VM_BUG_ON(!PageHead(page)); | |
947 | if (flags & FOLL_TOUCH) { | |
948 | pmd_t _pmd; | |
949 | /* | |
950 | * We should set the dirty bit only for FOLL_WRITE but | |
951 | * for now the dirty bit in the pmd is meaningless. | |
952 | * And if the dirty bit will become meaningful and | |
953 | * we'll only set it with FOLL_WRITE, an atomic | |
954 | * set_bit will be required on the pmd to set the | |
955 | * young bit, instead of the current set_pmd_at. | |
956 | */ | |
957 | _pmd = pmd_mkyoung(pmd_mkdirty(*pmd)); | |
958 | set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd); | |
959 | } | |
960 | page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; | |
961 | VM_BUG_ON(!PageCompound(page)); | |
962 | if (flags & FOLL_GET) | |
963 | get_page(page); | |
964 | ||
965 | out: | |
966 | return page; | |
967 | } | |
968 | ||
969 | int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, | |
970 | pmd_t *pmd) | |
971 | { | |
972 | int ret = 0; | |
973 | ||
974 | spin_lock(&tlb->mm->page_table_lock); | |
975 | if (likely(pmd_trans_huge(*pmd))) { | |
976 | if (unlikely(pmd_trans_splitting(*pmd))) { | |
977 | spin_unlock(&tlb->mm->page_table_lock); | |
978 | wait_split_huge_page(vma->anon_vma, | |
979 | pmd); | |
980 | } else { | |
981 | struct page *page; | |
982 | pgtable_t pgtable; | |
983 | pgtable = get_pmd_huge_pte(tlb->mm); | |
984 | page = pmd_page(*pmd); | |
985 | pmd_clear(pmd); | |
986 | page_remove_rmap(page); | |
987 | VM_BUG_ON(page_mapcount(page) < 0); | |
988 | add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); | |
989 | VM_BUG_ON(!PageHead(page)); | |
990 | spin_unlock(&tlb->mm->page_table_lock); | |
991 | tlb_remove_page(tlb, page); | |
992 | pte_free(tlb->mm, pgtable); | |
993 | ret = 1; | |
994 | } | |
995 | } else | |
996 | spin_unlock(&tlb->mm->page_table_lock); | |
997 | ||
998 | return ret; | |
999 | } | |
1000 | ||
0ca1634d JW |
1001 | int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, |
1002 | unsigned long addr, unsigned long end, | |
1003 | unsigned char *vec) | |
1004 | { | |
1005 | int ret = 0; | |
1006 | ||
1007 | spin_lock(&vma->vm_mm->page_table_lock); | |
1008 | if (likely(pmd_trans_huge(*pmd))) { | |
1009 | ret = !pmd_trans_splitting(*pmd); | |
1010 | spin_unlock(&vma->vm_mm->page_table_lock); | |
1011 | if (unlikely(!ret)) | |
1012 | wait_split_huge_page(vma->anon_vma, pmd); | |
1013 | else { | |
1014 | /* | |
1015 | * All logical pages in the range are present | |
1016 | * if backed by a huge page. | |
1017 | */ | |
1018 | memset(vec, 1, (end - addr) >> PAGE_SHIFT); | |
1019 | } | |
1020 | } else | |
1021 | spin_unlock(&vma->vm_mm->page_table_lock); | |
1022 | ||
1023 | return ret; | |
1024 | } | |
1025 | ||
cd7548ab JW |
1026 | int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, |
1027 | unsigned long addr, pgprot_t newprot) | |
1028 | { | |
1029 | struct mm_struct *mm = vma->vm_mm; | |
1030 | int ret = 0; | |
1031 | ||
1032 | spin_lock(&mm->page_table_lock); | |
1033 | if (likely(pmd_trans_huge(*pmd))) { | |
1034 | if (unlikely(pmd_trans_splitting(*pmd))) { | |
1035 | spin_unlock(&mm->page_table_lock); | |
1036 | wait_split_huge_page(vma->anon_vma, pmd); | |
1037 | } else { | |
1038 | pmd_t entry; | |
1039 | ||
1040 | entry = pmdp_get_and_clear(mm, addr, pmd); | |
1041 | entry = pmd_modify(entry, newprot); | |
1042 | set_pmd_at(mm, addr, pmd, entry); | |
1043 | spin_unlock(&vma->vm_mm->page_table_lock); | |
1044 | flush_tlb_range(vma, addr, addr + HPAGE_PMD_SIZE); | |
1045 | ret = 1; | |
1046 | } | |
1047 | } else | |
1048 | spin_unlock(&vma->vm_mm->page_table_lock); | |
1049 | ||
1050 | return ret; | |
1051 | } | |
1052 | ||
71e3aac0 AA |
1053 | pmd_t *page_check_address_pmd(struct page *page, |
1054 | struct mm_struct *mm, | |
1055 | unsigned long address, | |
1056 | enum page_check_address_pmd_flag flag) | |
1057 | { | |
1058 | pgd_t *pgd; | |
1059 | pud_t *pud; | |
1060 | pmd_t *pmd, *ret = NULL; | |
1061 | ||
1062 | if (address & ~HPAGE_PMD_MASK) | |
1063 | goto out; | |
1064 | ||
1065 | pgd = pgd_offset(mm, address); | |
1066 | if (!pgd_present(*pgd)) | |
1067 | goto out; | |
1068 | ||
1069 | pud = pud_offset(pgd, address); | |
1070 | if (!pud_present(*pud)) | |
1071 | goto out; | |
1072 | ||
1073 | pmd = pmd_offset(pud, address); | |
1074 | if (pmd_none(*pmd)) | |
1075 | goto out; | |
1076 | if (pmd_page(*pmd) != page) | |
1077 | goto out; | |
1078 | VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG && | |
1079 | pmd_trans_splitting(*pmd)); | |
1080 | if (pmd_trans_huge(*pmd)) { | |
1081 | VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG && | |
1082 | !pmd_trans_splitting(*pmd)); | |
1083 | ret = pmd; | |
1084 | } | |
1085 | out: | |
1086 | return ret; | |
1087 | } | |
1088 | ||
1089 | static int __split_huge_page_splitting(struct page *page, | |
1090 | struct vm_area_struct *vma, | |
1091 | unsigned long address) | |
1092 | { | |
1093 | struct mm_struct *mm = vma->vm_mm; | |
1094 | pmd_t *pmd; | |
1095 | int ret = 0; | |
1096 | ||
1097 | spin_lock(&mm->page_table_lock); | |
1098 | pmd = page_check_address_pmd(page, mm, address, | |
1099 | PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG); | |
1100 | if (pmd) { | |
1101 | /* | |
1102 | * We can't temporarily set the pmd to null in order | |
1103 | * to split it, the pmd must remain marked huge at all | |
1104 | * times or the VM won't take the pmd_trans_huge paths | |
1105 | * and it won't wait on the anon_vma->root->lock to | |
1106 | * serialize against split_huge_page*. | |
1107 | */ | |
1108 | pmdp_splitting_flush_notify(vma, address, pmd); | |
1109 | ret = 1; | |
1110 | } | |
1111 | spin_unlock(&mm->page_table_lock); | |
1112 | ||
1113 | return ret; | |
1114 | } | |
1115 | ||
1116 | static void __split_huge_page_refcount(struct page *page) | |
1117 | { | |
1118 | int i; | |
1119 | unsigned long head_index = page->index; | |
1120 | struct zone *zone = page_zone(page); | |
1121 | ||
1122 | /* prevent PageLRU to go away from under us, and freeze lru stats */ | |
1123 | spin_lock_irq(&zone->lru_lock); | |
1124 | compound_lock(page); | |
1125 | ||
1126 | for (i = 1; i < HPAGE_PMD_NR; i++) { | |
1127 | struct page *page_tail = page + i; | |
1128 | ||
1129 | /* tail_page->_count cannot change */ | |
1130 | atomic_sub(atomic_read(&page_tail->_count), &page->_count); | |
1131 | BUG_ON(page_count(page) <= 0); | |
1132 | atomic_add(page_mapcount(page) + 1, &page_tail->_count); | |
1133 | BUG_ON(atomic_read(&page_tail->_count) <= 0); | |
1134 | ||
1135 | /* after clearing PageTail the gup refcount can be released */ | |
1136 | smp_mb(); | |
1137 | ||
1138 | page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; | |
1139 | page_tail->flags |= (page->flags & | |
1140 | ((1L << PG_referenced) | | |
1141 | (1L << PG_swapbacked) | | |
1142 | (1L << PG_mlocked) | | |
1143 | (1L << PG_uptodate))); | |
1144 | page_tail->flags |= (1L << PG_dirty); | |
1145 | ||
1146 | /* | |
1147 | * 1) clear PageTail before overwriting first_page | |
1148 | * 2) clear PageTail before clearing PageHead for VM_BUG_ON | |
1149 | */ | |
1150 | smp_wmb(); | |
1151 | ||
1152 | /* | |
1153 | * __split_huge_page_splitting() already set the | |
1154 | * splitting bit in all pmd that could map this | |
1155 | * hugepage, that will ensure no CPU can alter the | |
1156 | * mapcount on the head page. The mapcount is only | |
1157 | * accounted in the head page and it has to be | |
1158 | * transferred to all tail pages in the below code. So | |
1159 | * for this code to be safe, the split the mapcount | |
1160 | * can't change. But that doesn't mean userland can't | |
1161 | * keep changing and reading the page contents while | |
1162 | * we transfer the mapcount, so the pmd splitting | |
1163 | * status is achieved setting a reserved bit in the | |
1164 | * pmd, not by clearing the present bit. | |
1165 | */ | |
1166 | BUG_ON(page_mapcount(page_tail)); | |
1167 | page_tail->_mapcount = page->_mapcount; | |
1168 | ||
1169 | BUG_ON(page_tail->mapping); | |
1170 | page_tail->mapping = page->mapping; | |
1171 | ||
1172 | page_tail->index = ++head_index; | |
1173 | ||
1174 | BUG_ON(!PageAnon(page_tail)); | |
1175 | BUG_ON(!PageUptodate(page_tail)); | |
1176 | BUG_ON(!PageDirty(page_tail)); | |
1177 | BUG_ON(!PageSwapBacked(page_tail)); | |
1178 | ||
1179 | lru_add_page_tail(zone, page, page_tail); | |
1180 | } | |
1181 | ||
79134171 AA |
1182 | __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES); |
1183 | __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR); | |
1184 | ||
71e3aac0 AA |
1185 | ClearPageCompound(page); |
1186 | compound_unlock(page); | |
1187 | spin_unlock_irq(&zone->lru_lock); | |
1188 | ||
1189 | for (i = 1; i < HPAGE_PMD_NR; i++) { | |
1190 | struct page *page_tail = page + i; | |
1191 | BUG_ON(page_count(page_tail) <= 0); | |
1192 | /* | |
1193 | * Tail pages may be freed if there wasn't any mapping | |
1194 | * like if add_to_swap() is running on a lru page that | |
1195 | * had its mapping zapped. And freeing these pages | |
1196 | * requires taking the lru_lock so we do the put_page | |
1197 | * of the tail pages after the split is complete. | |
1198 | */ | |
1199 | put_page(page_tail); | |
1200 | } | |
1201 | ||
1202 | /* | |
1203 | * Only the head page (now become a regular page) is required | |
1204 | * to be pinned by the caller. | |
1205 | */ | |
1206 | BUG_ON(page_count(page) <= 0); | |
1207 | } | |
1208 | ||
1209 | static int __split_huge_page_map(struct page *page, | |
1210 | struct vm_area_struct *vma, | |
1211 | unsigned long address) | |
1212 | { | |
1213 | struct mm_struct *mm = vma->vm_mm; | |
1214 | pmd_t *pmd, _pmd; | |
1215 | int ret = 0, i; | |
1216 | pgtable_t pgtable; | |
1217 | unsigned long haddr; | |
1218 | ||
1219 | spin_lock(&mm->page_table_lock); | |
1220 | pmd = page_check_address_pmd(page, mm, address, | |
1221 | PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG); | |
1222 | if (pmd) { | |
1223 | pgtable = get_pmd_huge_pte(mm); | |
1224 | pmd_populate(mm, &_pmd, pgtable); | |
1225 | ||
1226 | for (i = 0, haddr = address; i < HPAGE_PMD_NR; | |
1227 | i++, haddr += PAGE_SIZE) { | |
1228 | pte_t *pte, entry; | |
1229 | BUG_ON(PageCompound(page+i)); | |
1230 | entry = mk_pte(page + i, vma->vm_page_prot); | |
1231 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); | |
1232 | if (!pmd_write(*pmd)) | |
1233 | entry = pte_wrprotect(entry); | |
1234 | else | |
1235 | BUG_ON(page_mapcount(page) != 1); | |
1236 | if (!pmd_young(*pmd)) | |
1237 | entry = pte_mkold(entry); | |
1238 | pte = pte_offset_map(&_pmd, haddr); | |
1239 | BUG_ON(!pte_none(*pte)); | |
1240 | set_pte_at(mm, haddr, pte, entry); | |
1241 | pte_unmap(pte); | |
1242 | } | |
1243 | ||
1244 | mm->nr_ptes++; | |
1245 | smp_wmb(); /* make pte visible before pmd */ | |
1246 | /* | |
1247 | * Up to this point the pmd is present and huge and | |
1248 | * userland has the whole access to the hugepage | |
1249 | * during the split (which happens in place). If we | |
1250 | * overwrite the pmd with the not-huge version | |
1251 | * pointing to the pte here (which of course we could | |
1252 | * if all CPUs were bug free), userland could trigger | |
1253 | * a small page size TLB miss on the small sized TLB | |
1254 | * while the hugepage TLB entry is still established | |
1255 | * in the huge TLB. Some CPU doesn't like that. See | |
1256 | * http://support.amd.com/us/Processor_TechDocs/41322.pdf, | |
1257 | * Erratum 383 on page 93. Intel should be safe but is | |
1258 | * also warns that it's only safe if the permission | |
1259 | * and cache attributes of the two entries loaded in | |
1260 | * the two TLB is identical (which should be the case | |
1261 | * here). But it is generally safer to never allow | |
1262 | * small and huge TLB entries for the same virtual | |
1263 | * address to be loaded simultaneously. So instead of | |
1264 | * doing "pmd_populate(); flush_tlb_range();" we first | |
1265 | * mark the current pmd notpresent (atomically because | |
1266 | * here the pmd_trans_huge and pmd_trans_splitting | |
1267 | * must remain set at all times on the pmd until the | |
1268 | * split is complete for this pmd), then we flush the | |
1269 | * SMP TLB and finally we write the non-huge version | |
1270 | * of the pmd entry with pmd_populate. | |
1271 | */ | |
1272 | set_pmd_at(mm, address, pmd, pmd_mknotpresent(*pmd)); | |
1273 | flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE); | |
1274 | pmd_populate(mm, pmd, pgtable); | |
1275 | ret = 1; | |
1276 | } | |
1277 | spin_unlock(&mm->page_table_lock); | |
1278 | ||
1279 | return ret; | |
1280 | } | |
1281 | ||
1282 | /* must be called with anon_vma->root->lock hold */ | |
1283 | static void __split_huge_page(struct page *page, | |
1284 | struct anon_vma *anon_vma) | |
1285 | { | |
1286 | int mapcount, mapcount2; | |
1287 | struct anon_vma_chain *avc; | |
1288 | ||
1289 | BUG_ON(!PageHead(page)); | |
1290 | BUG_ON(PageTail(page)); | |
1291 | ||
1292 | mapcount = 0; | |
1293 | list_for_each_entry(avc, &anon_vma->head, same_anon_vma) { | |
1294 | struct vm_area_struct *vma = avc->vma; | |
1295 | unsigned long addr = vma_address(page, vma); | |
1296 | BUG_ON(is_vma_temporary_stack(vma)); | |
1297 | if (addr == -EFAULT) | |
1298 | continue; | |
1299 | mapcount += __split_huge_page_splitting(page, vma, addr); | |
1300 | } | |
05759d38 AA |
1301 | /* |
1302 | * It is critical that new vmas are added to the tail of the | |
1303 | * anon_vma list. This guarantes that if copy_huge_pmd() runs | |
1304 | * and establishes a child pmd before | |
1305 | * __split_huge_page_splitting() freezes the parent pmd (so if | |
1306 | * we fail to prevent copy_huge_pmd() from running until the | |
1307 | * whole __split_huge_page() is complete), we will still see | |
1308 | * the newly established pmd of the child later during the | |
1309 | * walk, to be able to set it as pmd_trans_splitting too. | |
1310 | */ | |
1311 | if (mapcount != page_mapcount(page)) | |
1312 | printk(KERN_ERR "mapcount %d page_mapcount %d\n", | |
1313 | mapcount, page_mapcount(page)); | |
71e3aac0 AA |
1314 | BUG_ON(mapcount != page_mapcount(page)); |
1315 | ||
1316 | __split_huge_page_refcount(page); | |
1317 | ||
1318 | mapcount2 = 0; | |
1319 | list_for_each_entry(avc, &anon_vma->head, same_anon_vma) { | |
1320 | struct vm_area_struct *vma = avc->vma; | |
1321 | unsigned long addr = vma_address(page, vma); | |
1322 | BUG_ON(is_vma_temporary_stack(vma)); | |
1323 | if (addr == -EFAULT) | |
1324 | continue; | |
1325 | mapcount2 += __split_huge_page_map(page, vma, addr); | |
1326 | } | |
05759d38 AA |
1327 | if (mapcount != mapcount2) |
1328 | printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n", | |
1329 | mapcount, mapcount2, page_mapcount(page)); | |
71e3aac0 AA |
1330 | BUG_ON(mapcount != mapcount2); |
1331 | } | |
1332 | ||
1333 | int split_huge_page(struct page *page) | |
1334 | { | |
1335 | struct anon_vma *anon_vma; | |
1336 | int ret = 1; | |
1337 | ||
1338 | BUG_ON(!PageAnon(page)); | |
1339 | anon_vma = page_lock_anon_vma(page); | |
1340 | if (!anon_vma) | |
1341 | goto out; | |
1342 | ret = 0; | |
1343 | if (!PageCompound(page)) | |
1344 | goto out_unlock; | |
1345 | ||
1346 | BUG_ON(!PageSwapBacked(page)); | |
1347 | __split_huge_page(page, anon_vma); | |
1348 | ||
1349 | BUG_ON(PageCompound(page)); | |
1350 | out_unlock: | |
1351 | page_unlock_anon_vma(anon_vma); | |
1352 | out: | |
1353 | return ret; | |
1354 | } | |
1355 | ||
0af4e98b AA |
1356 | int hugepage_madvise(unsigned long *vm_flags) |
1357 | { | |
1358 | /* | |
1359 | * Be somewhat over-protective like KSM for now! | |
1360 | */ | |
1361 | if (*vm_flags & (VM_HUGEPAGE | VM_SHARED | VM_MAYSHARE | | |
1362 | VM_PFNMAP | VM_IO | VM_DONTEXPAND | | |
1363 | VM_RESERVED | VM_HUGETLB | VM_INSERTPAGE | | |
1364 | VM_MIXEDMAP | VM_SAO)) | |
1365 | return -EINVAL; | |
1366 | ||
1367 | *vm_flags |= VM_HUGEPAGE; | |
1368 | ||
1369 | return 0; | |
1370 | } | |
1371 | ||
ba76149f AA |
1372 | static int __init khugepaged_slab_init(void) |
1373 | { | |
1374 | mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", | |
1375 | sizeof(struct mm_slot), | |
1376 | __alignof__(struct mm_slot), 0, NULL); | |
1377 | if (!mm_slot_cache) | |
1378 | return -ENOMEM; | |
1379 | ||
1380 | return 0; | |
1381 | } | |
1382 | ||
1383 | static void __init khugepaged_slab_free(void) | |
1384 | { | |
1385 | kmem_cache_destroy(mm_slot_cache); | |
1386 | mm_slot_cache = NULL; | |
1387 | } | |
1388 | ||
1389 | static inline struct mm_slot *alloc_mm_slot(void) | |
1390 | { | |
1391 | if (!mm_slot_cache) /* initialization failed */ | |
1392 | return NULL; | |
1393 | return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL); | |
1394 | } | |
1395 | ||
1396 | static inline void free_mm_slot(struct mm_slot *mm_slot) | |
1397 | { | |
1398 | kmem_cache_free(mm_slot_cache, mm_slot); | |
1399 | } | |
1400 | ||
1401 | static int __init mm_slots_hash_init(void) | |
1402 | { | |
1403 | mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head), | |
1404 | GFP_KERNEL); | |
1405 | if (!mm_slots_hash) | |
1406 | return -ENOMEM; | |
1407 | return 0; | |
1408 | } | |
1409 | ||
1410 | #if 0 | |
1411 | static void __init mm_slots_hash_free(void) | |
1412 | { | |
1413 | kfree(mm_slots_hash); | |
1414 | mm_slots_hash = NULL; | |
1415 | } | |
1416 | #endif | |
1417 | ||
1418 | static struct mm_slot *get_mm_slot(struct mm_struct *mm) | |
1419 | { | |
1420 | struct mm_slot *mm_slot; | |
1421 | struct hlist_head *bucket; | |
1422 | struct hlist_node *node; | |
1423 | ||
1424 | bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct)) | |
1425 | % MM_SLOTS_HASH_HEADS]; | |
1426 | hlist_for_each_entry(mm_slot, node, bucket, hash) { | |
1427 | if (mm == mm_slot->mm) | |
1428 | return mm_slot; | |
1429 | } | |
1430 | return NULL; | |
1431 | } | |
1432 | ||
1433 | static void insert_to_mm_slots_hash(struct mm_struct *mm, | |
1434 | struct mm_slot *mm_slot) | |
1435 | { | |
1436 | struct hlist_head *bucket; | |
1437 | ||
1438 | bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct)) | |
1439 | % MM_SLOTS_HASH_HEADS]; | |
1440 | mm_slot->mm = mm; | |
1441 | hlist_add_head(&mm_slot->hash, bucket); | |
1442 | } | |
1443 | ||
1444 | static inline int khugepaged_test_exit(struct mm_struct *mm) | |
1445 | { | |
1446 | return atomic_read(&mm->mm_users) == 0; | |
1447 | } | |
1448 | ||
1449 | int __khugepaged_enter(struct mm_struct *mm) | |
1450 | { | |
1451 | struct mm_slot *mm_slot; | |
1452 | int wakeup; | |
1453 | ||
1454 | mm_slot = alloc_mm_slot(); | |
1455 | if (!mm_slot) | |
1456 | return -ENOMEM; | |
1457 | ||
1458 | /* __khugepaged_exit() must not run from under us */ | |
1459 | VM_BUG_ON(khugepaged_test_exit(mm)); | |
1460 | if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { | |
1461 | free_mm_slot(mm_slot); | |
1462 | return 0; | |
1463 | } | |
1464 | ||
1465 | spin_lock(&khugepaged_mm_lock); | |
1466 | insert_to_mm_slots_hash(mm, mm_slot); | |
1467 | /* | |
1468 | * Insert just behind the scanning cursor, to let the area settle | |
1469 | * down a little. | |
1470 | */ | |
1471 | wakeup = list_empty(&khugepaged_scan.mm_head); | |
1472 | list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head); | |
1473 | spin_unlock(&khugepaged_mm_lock); | |
1474 | ||
1475 | atomic_inc(&mm->mm_count); | |
1476 | if (wakeup) | |
1477 | wake_up_interruptible(&khugepaged_wait); | |
1478 | ||
1479 | return 0; | |
1480 | } | |
1481 | ||
1482 | int khugepaged_enter_vma_merge(struct vm_area_struct *vma) | |
1483 | { | |
1484 | unsigned long hstart, hend; | |
1485 | if (!vma->anon_vma) | |
1486 | /* | |
1487 | * Not yet faulted in so we will register later in the | |
1488 | * page fault if needed. | |
1489 | */ | |
1490 | return 0; | |
1491 | if (vma->vm_file || vma->vm_ops) | |
1492 | /* khugepaged not yet working on file or special mappings */ | |
1493 | return 0; | |
1494 | VM_BUG_ON(is_linear_pfn_mapping(vma) || is_pfn_mapping(vma)); | |
1495 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; | |
1496 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
1497 | if (hstart < hend) | |
1498 | return khugepaged_enter(vma); | |
1499 | return 0; | |
1500 | } | |
1501 | ||
1502 | void __khugepaged_exit(struct mm_struct *mm) | |
1503 | { | |
1504 | struct mm_slot *mm_slot; | |
1505 | int free = 0; | |
1506 | ||
1507 | spin_lock(&khugepaged_mm_lock); | |
1508 | mm_slot = get_mm_slot(mm); | |
1509 | if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { | |
1510 | hlist_del(&mm_slot->hash); | |
1511 | list_del(&mm_slot->mm_node); | |
1512 | free = 1; | |
1513 | } | |
1514 | ||
1515 | if (free) { | |
1516 | spin_unlock(&khugepaged_mm_lock); | |
1517 | clear_bit(MMF_VM_HUGEPAGE, &mm->flags); | |
1518 | free_mm_slot(mm_slot); | |
1519 | mmdrop(mm); | |
1520 | } else if (mm_slot) { | |
1521 | spin_unlock(&khugepaged_mm_lock); | |
1522 | /* | |
1523 | * This is required to serialize against | |
1524 | * khugepaged_test_exit() (which is guaranteed to run | |
1525 | * under mmap sem read mode). Stop here (after we | |
1526 | * return all pagetables will be destroyed) until | |
1527 | * khugepaged has finished working on the pagetables | |
1528 | * under the mmap_sem. | |
1529 | */ | |
1530 | down_write(&mm->mmap_sem); | |
1531 | up_write(&mm->mmap_sem); | |
1532 | } else | |
1533 | spin_unlock(&khugepaged_mm_lock); | |
1534 | } | |
1535 | ||
1536 | static void release_pte_page(struct page *page) | |
1537 | { | |
1538 | /* 0 stands for page_is_file_cache(page) == false */ | |
1539 | dec_zone_page_state(page, NR_ISOLATED_ANON + 0); | |
1540 | unlock_page(page); | |
1541 | putback_lru_page(page); | |
1542 | } | |
1543 | ||
1544 | static void release_pte_pages(pte_t *pte, pte_t *_pte) | |
1545 | { | |
1546 | while (--_pte >= pte) { | |
1547 | pte_t pteval = *_pte; | |
1548 | if (!pte_none(pteval)) | |
1549 | release_pte_page(pte_page(pteval)); | |
1550 | } | |
1551 | } | |
1552 | ||
1553 | static void release_all_pte_pages(pte_t *pte) | |
1554 | { | |
1555 | release_pte_pages(pte, pte + HPAGE_PMD_NR); | |
1556 | } | |
1557 | ||
1558 | static int __collapse_huge_page_isolate(struct vm_area_struct *vma, | |
1559 | unsigned long address, | |
1560 | pte_t *pte) | |
1561 | { | |
1562 | struct page *page; | |
1563 | pte_t *_pte; | |
1564 | int referenced = 0, isolated = 0, none = 0; | |
1565 | for (_pte = pte; _pte < pte+HPAGE_PMD_NR; | |
1566 | _pte++, address += PAGE_SIZE) { | |
1567 | pte_t pteval = *_pte; | |
1568 | if (pte_none(pteval)) { | |
1569 | if (++none <= khugepaged_max_ptes_none) | |
1570 | continue; | |
1571 | else { | |
1572 | release_pte_pages(pte, _pte); | |
1573 | goto out; | |
1574 | } | |
1575 | } | |
1576 | if (!pte_present(pteval) || !pte_write(pteval)) { | |
1577 | release_pte_pages(pte, _pte); | |
1578 | goto out; | |
1579 | } | |
1580 | page = vm_normal_page(vma, address, pteval); | |
1581 | if (unlikely(!page)) { | |
1582 | release_pte_pages(pte, _pte); | |
1583 | goto out; | |
1584 | } | |
1585 | VM_BUG_ON(PageCompound(page)); | |
1586 | BUG_ON(!PageAnon(page)); | |
1587 | VM_BUG_ON(!PageSwapBacked(page)); | |
1588 | ||
1589 | /* cannot use mapcount: can't collapse if there's a gup pin */ | |
1590 | if (page_count(page) != 1) { | |
1591 | release_pte_pages(pte, _pte); | |
1592 | goto out; | |
1593 | } | |
1594 | /* | |
1595 | * We can do it before isolate_lru_page because the | |
1596 | * page can't be freed from under us. NOTE: PG_lock | |
1597 | * is needed to serialize against split_huge_page | |
1598 | * when invoked from the VM. | |
1599 | */ | |
1600 | if (!trylock_page(page)) { | |
1601 | release_pte_pages(pte, _pte); | |
1602 | goto out; | |
1603 | } | |
1604 | /* | |
1605 | * Isolate the page to avoid collapsing an hugepage | |
1606 | * currently in use by the VM. | |
1607 | */ | |
1608 | if (isolate_lru_page(page)) { | |
1609 | unlock_page(page); | |
1610 | release_pte_pages(pte, _pte); | |
1611 | goto out; | |
1612 | } | |
1613 | /* 0 stands for page_is_file_cache(page) == false */ | |
1614 | inc_zone_page_state(page, NR_ISOLATED_ANON + 0); | |
1615 | VM_BUG_ON(!PageLocked(page)); | |
1616 | VM_BUG_ON(PageLRU(page)); | |
1617 | ||
1618 | /* If there is no mapped pte young don't collapse the page */ | |
1619 | if (pte_young(pteval)) | |
1620 | referenced = 1; | |
1621 | } | |
1622 | if (unlikely(!referenced)) | |
1623 | release_all_pte_pages(pte); | |
1624 | else | |
1625 | isolated = 1; | |
1626 | out: | |
1627 | return isolated; | |
1628 | } | |
1629 | ||
1630 | static void __collapse_huge_page_copy(pte_t *pte, struct page *page, | |
1631 | struct vm_area_struct *vma, | |
1632 | unsigned long address, | |
1633 | spinlock_t *ptl) | |
1634 | { | |
1635 | pte_t *_pte; | |
1636 | for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) { | |
1637 | pte_t pteval = *_pte; | |
1638 | struct page *src_page; | |
1639 | ||
1640 | if (pte_none(pteval)) { | |
1641 | clear_user_highpage(page, address); | |
1642 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); | |
1643 | } else { | |
1644 | src_page = pte_page(pteval); | |
1645 | copy_user_highpage(page, src_page, address, vma); | |
1646 | VM_BUG_ON(page_mapcount(src_page) != 1); | |
1647 | VM_BUG_ON(page_count(src_page) != 2); | |
1648 | release_pte_page(src_page); | |
1649 | /* | |
1650 | * ptl mostly unnecessary, but preempt has to | |
1651 | * be disabled to update the per-cpu stats | |
1652 | * inside page_remove_rmap(). | |
1653 | */ | |
1654 | spin_lock(ptl); | |
1655 | /* | |
1656 | * paravirt calls inside pte_clear here are | |
1657 | * superfluous. | |
1658 | */ | |
1659 | pte_clear(vma->vm_mm, address, _pte); | |
1660 | page_remove_rmap(src_page); | |
1661 | spin_unlock(ptl); | |
1662 | free_page_and_swap_cache(src_page); | |
1663 | } | |
1664 | ||
1665 | address += PAGE_SIZE; | |
1666 | page++; | |
1667 | } | |
1668 | } | |
1669 | ||
1670 | static void collapse_huge_page(struct mm_struct *mm, | |
1671 | unsigned long address, | |
ce83d217 AA |
1672 | struct page **hpage, |
1673 | struct vm_area_struct *vma) | |
ba76149f | 1674 | { |
ba76149f AA |
1675 | pgd_t *pgd; |
1676 | pud_t *pud; | |
1677 | pmd_t *pmd, _pmd; | |
1678 | pte_t *pte; | |
1679 | pgtable_t pgtable; | |
1680 | struct page *new_page; | |
1681 | spinlock_t *ptl; | |
1682 | int isolated; | |
1683 | unsigned long hstart, hend; | |
1684 | ||
1685 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
0bbbc0b3 | 1686 | #ifndef CONFIG_NUMA |
ba76149f | 1687 | VM_BUG_ON(!*hpage); |
ce83d217 | 1688 | new_page = *hpage; |
0bbbc0b3 AA |
1689 | #else |
1690 | VM_BUG_ON(*hpage); | |
ce83d217 AA |
1691 | /* |
1692 | * Allocate the page while the vma is still valid and under | |
1693 | * the mmap_sem read mode so there is no memory allocation | |
1694 | * later when we take the mmap_sem in write mode. This is more | |
1695 | * friendly behavior (OTOH it may actually hide bugs) to | |
1696 | * filesystems in userland with daemons allocating memory in | |
1697 | * the userland I/O paths. Allocating memory with the | |
1698 | * mmap_sem in read mode is good idea also to allow greater | |
1699 | * scalability. | |
1700 | */ | |
1701 | new_page = alloc_hugepage_vma(khugepaged_defrag(), vma, address); | |
1702 | if (unlikely(!new_page)) { | |
1703 | up_read(&mm->mmap_sem); | |
1704 | *hpage = ERR_PTR(-ENOMEM); | |
1705 | return; | |
1706 | } | |
0bbbc0b3 | 1707 | #endif |
ce83d217 AA |
1708 | if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) { |
1709 | up_read(&mm->mmap_sem); | |
1710 | put_page(new_page); | |
1711 | return; | |
1712 | } | |
1713 | ||
1714 | /* after allocating the hugepage upgrade to mmap_sem write mode */ | |
1715 | up_read(&mm->mmap_sem); | |
ba76149f AA |
1716 | |
1717 | /* | |
1718 | * Prevent all access to pagetables with the exception of | |
1719 | * gup_fast later hanlded by the ptep_clear_flush and the VM | |
1720 | * handled by the anon_vma lock + PG_lock. | |
1721 | */ | |
1722 | down_write(&mm->mmap_sem); | |
1723 | if (unlikely(khugepaged_test_exit(mm))) | |
1724 | goto out; | |
1725 | ||
1726 | vma = find_vma(mm, address); | |
1727 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; | |
1728 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
1729 | if (address < hstart || address + HPAGE_PMD_SIZE > hend) | |
1730 | goto out; | |
1731 | ||
1732 | if (!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) | |
1733 | goto out; | |
1734 | ||
1735 | /* VM_PFNMAP vmas may have vm_ops null but vm_file set */ | |
1736 | if (!vma->anon_vma || vma->vm_ops || vma->vm_file) | |
1737 | goto out; | |
1738 | VM_BUG_ON(is_linear_pfn_mapping(vma) || is_pfn_mapping(vma)); | |
1739 | ||
1740 | pgd = pgd_offset(mm, address); | |
1741 | if (!pgd_present(*pgd)) | |
1742 | goto out; | |
1743 | ||
1744 | pud = pud_offset(pgd, address); | |
1745 | if (!pud_present(*pud)) | |
1746 | goto out; | |
1747 | ||
1748 | pmd = pmd_offset(pud, address); | |
1749 | /* pmd can't go away or become huge under us */ | |
1750 | if (!pmd_present(*pmd) || pmd_trans_huge(*pmd)) | |
1751 | goto out; | |
1752 | ||
ba76149f AA |
1753 | anon_vma_lock(vma->anon_vma); |
1754 | ||
1755 | pte = pte_offset_map(pmd, address); | |
1756 | ptl = pte_lockptr(mm, pmd); | |
1757 | ||
1758 | spin_lock(&mm->page_table_lock); /* probably unnecessary */ | |
1759 | /* | |
1760 | * After this gup_fast can't run anymore. This also removes | |
1761 | * any huge TLB entry from the CPU so we won't allow | |
1762 | * huge and small TLB entries for the same virtual address | |
1763 | * to avoid the risk of CPU bugs in that area. | |
1764 | */ | |
1765 | _pmd = pmdp_clear_flush_notify(vma, address, pmd); | |
1766 | spin_unlock(&mm->page_table_lock); | |
1767 | ||
1768 | spin_lock(ptl); | |
1769 | isolated = __collapse_huge_page_isolate(vma, address, pte); | |
1770 | spin_unlock(ptl); | |
1771 | pte_unmap(pte); | |
1772 | ||
1773 | if (unlikely(!isolated)) { | |
1774 | spin_lock(&mm->page_table_lock); | |
1775 | BUG_ON(!pmd_none(*pmd)); | |
1776 | set_pmd_at(mm, address, pmd, _pmd); | |
1777 | spin_unlock(&mm->page_table_lock); | |
1778 | anon_vma_unlock(vma->anon_vma); | |
1779 | mem_cgroup_uncharge_page(new_page); | |
ce83d217 | 1780 | goto out; |
ba76149f AA |
1781 | } |
1782 | ||
1783 | /* | |
1784 | * All pages are isolated and locked so anon_vma rmap | |
1785 | * can't run anymore. | |
1786 | */ | |
1787 | anon_vma_unlock(vma->anon_vma); | |
1788 | ||
1789 | __collapse_huge_page_copy(pte, new_page, vma, address, ptl); | |
1790 | __SetPageUptodate(new_page); | |
1791 | pgtable = pmd_pgtable(_pmd); | |
1792 | VM_BUG_ON(page_count(pgtable) != 1); | |
1793 | VM_BUG_ON(page_mapcount(pgtable) != 0); | |
1794 | ||
1795 | _pmd = mk_pmd(new_page, vma->vm_page_prot); | |
1796 | _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); | |
1797 | _pmd = pmd_mkhuge(_pmd); | |
1798 | ||
1799 | /* | |
1800 | * spin_lock() below is not the equivalent of smp_wmb(), so | |
1801 | * this is needed to avoid the copy_huge_page writes to become | |
1802 | * visible after the set_pmd_at() write. | |
1803 | */ | |
1804 | smp_wmb(); | |
1805 | ||
1806 | spin_lock(&mm->page_table_lock); | |
1807 | BUG_ON(!pmd_none(*pmd)); | |
1808 | page_add_new_anon_rmap(new_page, vma, address); | |
1809 | set_pmd_at(mm, address, pmd, _pmd); | |
1810 | update_mmu_cache(vma, address, entry); | |
1811 | prepare_pmd_huge_pte(pgtable, mm); | |
1812 | mm->nr_ptes--; | |
1813 | spin_unlock(&mm->page_table_lock); | |
1814 | ||
0bbbc0b3 | 1815 | #ifndef CONFIG_NUMA |
ba76149f | 1816 | *hpage = NULL; |
0bbbc0b3 | 1817 | #endif |
ba76149f | 1818 | khugepaged_pages_collapsed++; |
ce83d217 | 1819 | out_up_write: |
ba76149f | 1820 | up_write(&mm->mmap_sem); |
0bbbc0b3 AA |
1821 | return; |
1822 | ||
ce83d217 | 1823 | out: |
0bbbc0b3 AA |
1824 | #ifdef CONFIG_NUMA |
1825 | put_page(new_page); | |
1826 | #endif | |
ce83d217 | 1827 | goto out_up_write; |
ba76149f AA |
1828 | } |
1829 | ||
1830 | static int khugepaged_scan_pmd(struct mm_struct *mm, | |
1831 | struct vm_area_struct *vma, | |
1832 | unsigned long address, | |
1833 | struct page **hpage) | |
1834 | { | |
1835 | pgd_t *pgd; | |
1836 | pud_t *pud; | |
1837 | pmd_t *pmd; | |
1838 | pte_t *pte, *_pte; | |
1839 | int ret = 0, referenced = 0, none = 0; | |
1840 | struct page *page; | |
1841 | unsigned long _address; | |
1842 | spinlock_t *ptl; | |
1843 | ||
1844 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
1845 | ||
1846 | pgd = pgd_offset(mm, address); | |
1847 | if (!pgd_present(*pgd)) | |
1848 | goto out; | |
1849 | ||
1850 | pud = pud_offset(pgd, address); | |
1851 | if (!pud_present(*pud)) | |
1852 | goto out; | |
1853 | ||
1854 | pmd = pmd_offset(pud, address); | |
1855 | if (!pmd_present(*pmd) || pmd_trans_huge(*pmd)) | |
1856 | goto out; | |
1857 | ||
1858 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); | |
1859 | for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR; | |
1860 | _pte++, _address += PAGE_SIZE) { | |
1861 | pte_t pteval = *_pte; | |
1862 | if (pte_none(pteval)) { | |
1863 | if (++none <= khugepaged_max_ptes_none) | |
1864 | continue; | |
1865 | else | |
1866 | goto out_unmap; | |
1867 | } | |
1868 | if (!pte_present(pteval) || !pte_write(pteval)) | |
1869 | goto out_unmap; | |
1870 | page = vm_normal_page(vma, _address, pteval); | |
1871 | if (unlikely(!page)) | |
1872 | goto out_unmap; | |
1873 | VM_BUG_ON(PageCompound(page)); | |
1874 | if (!PageLRU(page) || PageLocked(page) || !PageAnon(page)) | |
1875 | goto out_unmap; | |
1876 | /* cannot use mapcount: can't collapse if there's a gup pin */ | |
1877 | if (page_count(page) != 1) | |
1878 | goto out_unmap; | |
1879 | if (pte_young(pteval)) | |
1880 | referenced = 1; | |
1881 | } | |
1882 | if (referenced) | |
1883 | ret = 1; | |
1884 | out_unmap: | |
1885 | pte_unmap_unlock(pte, ptl); | |
ce83d217 AA |
1886 | if (ret) |
1887 | /* collapse_huge_page will return with the mmap_sem released */ | |
1888 | collapse_huge_page(mm, address, hpage, vma); | |
ba76149f AA |
1889 | out: |
1890 | return ret; | |
1891 | } | |
1892 | ||
1893 | static void collect_mm_slot(struct mm_slot *mm_slot) | |
1894 | { | |
1895 | struct mm_struct *mm = mm_slot->mm; | |
1896 | ||
1897 | VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock)); | |
1898 | ||
1899 | if (khugepaged_test_exit(mm)) { | |
1900 | /* free mm_slot */ | |
1901 | hlist_del(&mm_slot->hash); | |
1902 | list_del(&mm_slot->mm_node); | |
1903 | ||
1904 | /* | |
1905 | * Not strictly needed because the mm exited already. | |
1906 | * | |
1907 | * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); | |
1908 | */ | |
1909 | ||
1910 | /* khugepaged_mm_lock actually not necessary for the below */ | |
1911 | free_mm_slot(mm_slot); | |
1912 | mmdrop(mm); | |
1913 | } | |
1914 | } | |
1915 | ||
1916 | static unsigned int khugepaged_scan_mm_slot(unsigned int pages, | |
1917 | struct page **hpage) | |
1918 | { | |
1919 | struct mm_slot *mm_slot; | |
1920 | struct mm_struct *mm; | |
1921 | struct vm_area_struct *vma; | |
1922 | int progress = 0; | |
1923 | ||
1924 | VM_BUG_ON(!pages); | |
1925 | VM_BUG_ON(!spin_is_locked(&khugepaged_mm_lock)); | |
1926 | ||
1927 | if (khugepaged_scan.mm_slot) | |
1928 | mm_slot = khugepaged_scan.mm_slot; | |
1929 | else { | |
1930 | mm_slot = list_entry(khugepaged_scan.mm_head.next, | |
1931 | struct mm_slot, mm_node); | |
1932 | khugepaged_scan.address = 0; | |
1933 | khugepaged_scan.mm_slot = mm_slot; | |
1934 | } | |
1935 | spin_unlock(&khugepaged_mm_lock); | |
1936 | ||
1937 | mm = mm_slot->mm; | |
1938 | down_read(&mm->mmap_sem); | |
1939 | if (unlikely(khugepaged_test_exit(mm))) | |
1940 | vma = NULL; | |
1941 | else | |
1942 | vma = find_vma(mm, khugepaged_scan.address); | |
1943 | ||
1944 | progress++; | |
1945 | for (; vma; vma = vma->vm_next) { | |
1946 | unsigned long hstart, hend; | |
1947 | ||
1948 | cond_resched(); | |
1949 | if (unlikely(khugepaged_test_exit(mm))) { | |
1950 | progress++; | |
1951 | break; | |
1952 | } | |
1953 | ||
1954 | if (!(vma->vm_flags & VM_HUGEPAGE) && | |
1955 | !khugepaged_always()) { | |
1956 | progress++; | |
1957 | continue; | |
1958 | } | |
1959 | ||
1960 | /* VM_PFNMAP vmas may have vm_ops null but vm_file set */ | |
1961 | if (!vma->anon_vma || vma->vm_ops || vma->vm_file) { | |
1962 | khugepaged_scan.address = vma->vm_end; | |
1963 | progress++; | |
1964 | continue; | |
1965 | } | |
1966 | VM_BUG_ON(is_linear_pfn_mapping(vma) || is_pfn_mapping(vma)); | |
1967 | ||
1968 | hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; | |
1969 | hend = vma->vm_end & HPAGE_PMD_MASK; | |
1970 | if (hstart >= hend) { | |
1971 | progress++; | |
1972 | continue; | |
1973 | } | |
1974 | if (khugepaged_scan.address < hstart) | |
1975 | khugepaged_scan.address = hstart; | |
1976 | if (khugepaged_scan.address > hend) { | |
1977 | khugepaged_scan.address = hend + HPAGE_PMD_SIZE; | |
1978 | progress++; | |
1979 | continue; | |
1980 | } | |
1981 | BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); | |
1982 | ||
1983 | while (khugepaged_scan.address < hend) { | |
1984 | int ret; | |
1985 | cond_resched(); | |
1986 | if (unlikely(khugepaged_test_exit(mm))) | |
1987 | goto breakouterloop; | |
1988 | ||
1989 | VM_BUG_ON(khugepaged_scan.address < hstart || | |
1990 | khugepaged_scan.address + HPAGE_PMD_SIZE > | |
1991 | hend); | |
1992 | ret = khugepaged_scan_pmd(mm, vma, | |
1993 | khugepaged_scan.address, | |
1994 | hpage); | |
1995 | /* move to next address */ | |
1996 | khugepaged_scan.address += HPAGE_PMD_SIZE; | |
1997 | progress += HPAGE_PMD_NR; | |
1998 | if (ret) | |
1999 | /* we released mmap_sem so break loop */ | |
2000 | goto breakouterloop_mmap_sem; | |
2001 | if (progress >= pages) | |
2002 | goto breakouterloop; | |
2003 | } | |
2004 | } | |
2005 | breakouterloop: | |
2006 | up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */ | |
2007 | breakouterloop_mmap_sem: | |
2008 | ||
2009 | spin_lock(&khugepaged_mm_lock); | |
2010 | BUG_ON(khugepaged_scan.mm_slot != mm_slot); | |
2011 | /* | |
2012 | * Release the current mm_slot if this mm is about to die, or | |
2013 | * if we scanned all vmas of this mm. | |
2014 | */ | |
2015 | if (khugepaged_test_exit(mm) || !vma) { | |
2016 | /* | |
2017 | * Make sure that if mm_users is reaching zero while | |
2018 | * khugepaged runs here, khugepaged_exit will find | |
2019 | * mm_slot not pointing to the exiting mm. | |
2020 | */ | |
2021 | if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) { | |
2022 | khugepaged_scan.mm_slot = list_entry( | |
2023 | mm_slot->mm_node.next, | |
2024 | struct mm_slot, mm_node); | |
2025 | khugepaged_scan.address = 0; | |
2026 | } else { | |
2027 | khugepaged_scan.mm_slot = NULL; | |
2028 | khugepaged_full_scans++; | |
2029 | } | |
2030 | ||
2031 | collect_mm_slot(mm_slot); | |
2032 | } | |
2033 | ||
2034 | return progress; | |
2035 | } | |
2036 | ||
2037 | static int khugepaged_has_work(void) | |
2038 | { | |
2039 | return !list_empty(&khugepaged_scan.mm_head) && | |
2040 | khugepaged_enabled(); | |
2041 | } | |
2042 | ||
2043 | static int khugepaged_wait_event(void) | |
2044 | { | |
2045 | return !list_empty(&khugepaged_scan.mm_head) || | |
2046 | !khugepaged_enabled(); | |
2047 | } | |
2048 | ||
2049 | static void khugepaged_do_scan(struct page **hpage) | |
2050 | { | |
2051 | unsigned int progress = 0, pass_through_head = 0; | |
2052 | unsigned int pages = khugepaged_pages_to_scan; | |
2053 | ||
2054 | barrier(); /* write khugepaged_pages_to_scan to local stack */ | |
2055 | ||
2056 | while (progress < pages) { | |
2057 | cond_resched(); | |
2058 | ||
0bbbc0b3 | 2059 | #ifndef CONFIG_NUMA |
ba76149f AA |
2060 | if (!*hpage) { |
2061 | *hpage = alloc_hugepage(khugepaged_defrag()); | |
2062 | if (unlikely(!*hpage)) | |
2063 | break; | |
2064 | } | |
0bbbc0b3 AA |
2065 | #else |
2066 | if (IS_ERR(*hpage)) | |
2067 | break; | |
2068 | #endif | |
ba76149f AA |
2069 | |
2070 | spin_lock(&khugepaged_mm_lock); | |
2071 | if (!khugepaged_scan.mm_slot) | |
2072 | pass_through_head++; | |
2073 | if (khugepaged_has_work() && | |
2074 | pass_through_head < 2) | |
2075 | progress += khugepaged_scan_mm_slot(pages - progress, | |
2076 | hpage); | |
2077 | else | |
2078 | progress = pages; | |
2079 | spin_unlock(&khugepaged_mm_lock); | |
2080 | } | |
2081 | } | |
2082 | ||
0bbbc0b3 AA |
2083 | static void khugepaged_alloc_sleep(void) |
2084 | { | |
2085 | DEFINE_WAIT(wait); | |
2086 | add_wait_queue(&khugepaged_wait, &wait); | |
2087 | schedule_timeout_interruptible( | |
2088 | msecs_to_jiffies( | |
2089 | khugepaged_alloc_sleep_millisecs)); | |
2090 | remove_wait_queue(&khugepaged_wait, &wait); | |
2091 | } | |
2092 | ||
2093 | #ifndef CONFIG_NUMA | |
ba76149f AA |
2094 | static struct page *khugepaged_alloc_hugepage(void) |
2095 | { | |
2096 | struct page *hpage; | |
2097 | ||
2098 | do { | |
2099 | hpage = alloc_hugepage(khugepaged_defrag()); | |
0bbbc0b3 AA |
2100 | if (!hpage) |
2101 | khugepaged_alloc_sleep(); | |
ba76149f AA |
2102 | } while (unlikely(!hpage) && |
2103 | likely(khugepaged_enabled())); | |
2104 | return hpage; | |
2105 | } | |
0bbbc0b3 | 2106 | #endif |
ba76149f AA |
2107 | |
2108 | static void khugepaged_loop(void) | |
2109 | { | |
2110 | struct page *hpage; | |
2111 | ||
0bbbc0b3 AA |
2112 | #ifdef CONFIG_NUMA |
2113 | hpage = NULL; | |
2114 | #endif | |
ba76149f | 2115 | while (likely(khugepaged_enabled())) { |
0bbbc0b3 | 2116 | #ifndef CONFIG_NUMA |
ba76149f AA |
2117 | hpage = khugepaged_alloc_hugepage(); |
2118 | if (unlikely(!hpage)) | |
2119 | break; | |
0bbbc0b3 AA |
2120 | #else |
2121 | if (IS_ERR(hpage)) { | |
2122 | khugepaged_alloc_sleep(); | |
2123 | hpage = NULL; | |
2124 | } | |
2125 | #endif | |
ba76149f AA |
2126 | |
2127 | khugepaged_do_scan(&hpage); | |
0bbbc0b3 | 2128 | #ifndef CONFIG_NUMA |
ba76149f AA |
2129 | if (hpage) |
2130 | put_page(hpage); | |
0bbbc0b3 | 2131 | #endif |
ba76149f AA |
2132 | if (khugepaged_has_work()) { |
2133 | DEFINE_WAIT(wait); | |
2134 | if (!khugepaged_scan_sleep_millisecs) | |
2135 | continue; | |
2136 | add_wait_queue(&khugepaged_wait, &wait); | |
2137 | schedule_timeout_interruptible( | |
2138 | msecs_to_jiffies( | |
2139 | khugepaged_scan_sleep_millisecs)); | |
2140 | remove_wait_queue(&khugepaged_wait, &wait); | |
2141 | } else if (khugepaged_enabled()) | |
2142 | wait_event_interruptible(khugepaged_wait, | |
2143 | khugepaged_wait_event()); | |
2144 | } | |
2145 | } | |
2146 | ||
2147 | static int khugepaged(void *none) | |
2148 | { | |
2149 | struct mm_slot *mm_slot; | |
2150 | ||
2151 | set_user_nice(current, 19); | |
2152 | ||
2153 | /* serialize with start_khugepaged() */ | |
2154 | mutex_lock(&khugepaged_mutex); | |
2155 | ||
2156 | for (;;) { | |
2157 | mutex_unlock(&khugepaged_mutex); | |
2158 | BUG_ON(khugepaged_thread != current); | |
2159 | khugepaged_loop(); | |
2160 | BUG_ON(khugepaged_thread != current); | |
2161 | ||
2162 | mutex_lock(&khugepaged_mutex); | |
2163 | if (!khugepaged_enabled()) | |
2164 | break; | |
2165 | } | |
2166 | ||
2167 | spin_lock(&khugepaged_mm_lock); | |
2168 | mm_slot = khugepaged_scan.mm_slot; | |
2169 | khugepaged_scan.mm_slot = NULL; | |
2170 | if (mm_slot) | |
2171 | collect_mm_slot(mm_slot); | |
2172 | spin_unlock(&khugepaged_mm_lock); | |
2173 | ||
2174 | khugepaged_thread = NULL; | |
2175 | mutex_unlock(&khugepaged_mutex); | |
2176 | ||
2177 | return 0; | |
2178 | } | |
2179 | ||
71e3aac0 AA |
2180 | void __split_huge_page_pmd(struct mm_struct *mm, pmd_t *pmd) |
2181 | { | |
2182 | struct page *page; | |
2183 | ||
2184 | spin_lock(&mm->page_table_lock); | |
2185 | if (unlikely(!pmd_trans_huge(*pmd))) { | |
2186 | spin_unlock(&mm->page_table_lock); | |
2187 | return; | |
2188 | } | |
2189 | page = pmd_page(*pmd); | |
2190 | VM_BUG_ON(!page_count(page)); | |
2191 | get_page(page); | |
2192 | spin_unlock(&mm->page_table_lock); | |
2193 | ||
2194 | split_huge_page(page); | |
2195 | ||
2196 | put_page(page); | |
2197 | BUG_ON(pmd_trans_huge(*pmd)); | |
2198 | } |