]>
Commit | Line | Data |
---|---|---|
71e3aac0 AA |
1 | /* |
2 | * Copyright (C) 2009 Red Hat, Inc. | |
3 | * | |
4 | * This work is licensed under the terms of the GNU GPL, version 2. See | |
5 | * the COPYING file in the top-level directory. | |
6 | */ | |
7 | ||
ae3a8c1c AM |
8 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
9 | ||
71e3aac0 AA |
10 | #include <linux/mm.h> |
11 | #include <linux/sched.h> | |
12 | #include <linux/highmem.h> | |
13 | #include <linux/hugetlb.h> | |
14 | #include <linux/mmu_notifier.h> | |
15 | #include <linux/rmap.h> | |
16 | #include <linux/swap.h> | |
97ae1749 | 17 | #include <linux/shrinker.h> |
ba76149f | 18 | #include <linux/mm_inline.h> |
e9b61f19 | 19 | #include <linux/swapops.h> |
4897c765 | 20 | #include <linux/dax.h> |
ba76149f | 21 | #include <linux/khugepaged.h> |
878aee7d | 22 | #include <linux/freezer.h> |
f25748e3 | 23 | #include <linux/pfn_t.h> |
a664b2d8 | 24 | #include <linux/mman.h> |
3565fce3 | 25 | #include <linux/memremap.h> |
325adeb5 | 26 | #include <linux/pagemap.h> |
49071d43 | 27 | #include <linux/debugfs.h> |
4daae3b4 | 28 | #include <linux/migrate.h> |
43b5fbbd | 29 | #include <linux/hashtable.h> |
6b251fc9 | 30 | #include <linux/userfaultfd_k.h> |
33c3fc71 | 31 | #include <linux/page_idle.h> |
baa355fd | 32 | #include <linux/shmem_fs.h> |
97ae1749 | 33 | |
71e3aac0 AA |
34 | #include <asm/tlb.h> |
35 | #include <asm/pgalloc.h> | |
36 | #include "internal.h" | |
37 | ||
ba76149f | 38 | /* |
8bfa3f9a JW |
39 | * By default transparent hugepage support is disabled in order that avoid |
40 | * to risk increase the memory footprint of applications without a guaranteed | |
41 | * benefit. When transparent hugepage support is enabled, is for all mappings, | |
42 | * and khugepaged scans all mappings. | |
43 | * Defrag is invoked by khugepaged hugepage allocations and by page faults | |
44 | * for all hugepage allocations. | |
ba76149f | 45 | */ |
71e3aac0 | 46 | unsigned long transparent_hugepage_flags __read_mostly = |
13ece886 | 47 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS |
ba76149f | 48 | (1<<TRANSPARENT_HUGEPAGE_FLAG)| |
13ece886 AA |
49 | #endif |
50 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE | |
51 | (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| | |
52 | #endif | |
444eb2a4 | 53 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)| |
79da5407 KS |
54 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| |
55 | (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
ba76149f | 56 | |
9a982250 | 57 | static struct shrinker deferred_split_shrinker; |
f000565a | 58 | |
97ae1749 | 59 | static atomic_t huge_zero_refcount; |
56873f43 | 60 | struct page *huge_zero_page __read_mostly; |
4a6c1297 | 61 | |
fc437044 | 62 | struct page *get_huge_zero_page(void) |
97ae1749 KS |
63 | { |
64 | struct page *zero_page; | |
65 | retry: | |
66 | if (likely(atomic_inc_not_zero(&huge_zero_refcount))) | |
4db0c3c2 | 67 | return READ_ONCE(huge_zero_page); |
97ae1749 KS |
68 | |
69 | zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, | |
4a6c1297 | 70 | HPAGE_PMD_ORDER); |
d8a8e1f0 KS |
71 | if (!zero_page) { |
72 | count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); | |
5918d10a | 73 | return NULL; |
d8a8e1f0 KS |
74 | } |
75 | count_vm_event(THP_ZERO_PAGE_ALLOC); | |
97ae1749 | 76 | preempt_disable(); |
5918d10a | 77 | if (cmpxchg(&huge_zero_page, NULL, zero_page)) { |
97ae1749 | 78 | preempt_enable(); |
5ddacbe9 | 79 | __free_pages(zero_page, compound_order(zero_page)); |
97ae1749 KS |
80 | goto retry; |
81 | } | |
82 | ||
83 | /* We take additional reference here. It will be put back by shrinker */ | |
84 | atomic_set(&huge_zero_refcount, 2); | |
85 | preempt_enable(); | |
4db0c3c2 | 86 | return READ_ONCE(huge_zero_page); |
4a6c1297 KS |
87 | } |
88 | ||
aa88b68c | 89 | void put_huge_zero_page(void) |
4a6c1297 | 90 | { |
97ae1749 KS |
91 | /* |
92 | * Counter should never go to zero here. Only shrinker can put | |
93 | * last reference. | |
94 | */ | |
95 | BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); | |
4a6c1297 KS |
96 | } |
97 | ||
48896466 GC |
98 | static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, |
99 | struct shrink_control *sc) | |
4a6c1297 | 100 | { |
48896466 GC |
101 | /* we can free zero page only if last reference remains */ |
102 | return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; | |
103 | } | |
97ae1749 | 104 | |
48896466 GC |
105 | static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, |
106 | struct shrink_control *sc) | |
107 | { | |
97ae1749 | 108 | if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { |
5918d10a KS |
109 | struct page *zero_page = xchg(&huge_zero_page, NULL); |
110 | BUG_ON(zero_page == NULL); | |
5ddacbe9 | 111 | __free_pages(zero_page, compound_order(zero_page)); |
48896466 | 112 | return HPAGE_PMD_NR; |
97ae1749 KS |
113 | } |
114 | ||
115 | return 0; | |
4a6c1297 KS |
116 | } |
117 | ||
97ae1749 | 118 | static struct shrinker huge_zero_page_shrinker = { |
48896466 GC |
119 | .count_objects = shrink_huge_zero_page_count, |
120 | .scan_objects = shrink_huge_zero_page_scan, | |
97ae1749 KS |
121 | .seeks = DEFAULT_SEEKS, |
122 | }; | |
123 | ||
71e3aac0 | 124 | #ifdef CONFIG_SYSFS |
ba76149f | 125 | |
444eb2a4 | 126 | static ssize_t triple_flag_store(struct kobject *kobj, |
71e3aac0 AA |
127 | struct kobj_attribute *attr, |
128 | const char *buf, size_t count, | |
129 | enum transparent_hugepage_flag enabled, | |
444eb2a4 | 130 | enum transparent_hugepage_flag deferred, |
71e3aac0 AA |
131 | enum transparent_hugepage_flag req_madv) |
132 | { | |
444eb2a4 MG |
133 | if (!memcmp("defer", buf, |
134 | min(sizeof("defer")-1, count))) { | |
135 | if (enabled == deferred) | |
136 | return -EINVAL; | |
137 | clear_bit(enabled, &transparent_hugepage_flags); | |
138 | clear_bit(req_madv, &transparent_hugepage_flags); | |
139 | set_bit(deferred, &transparent_hugepage_flags); | |
140 | } else if (!memcmp("always", buf, | |
71e3aac0 | 141 | min(sizeof("always")-1, count))) { |
444eb2a4 | 142 | clear_bit(deferred, &transparent_hugepage_flags); |
71e3aac0 | 143 | clear_bit(req_madv, &transparent_hugepage_flags); |
444eb2a4 | 144 | set_bit(enabled, &transparent_hugepage_flags); |
71e3aac0 AA |
145 | } else if (!memcmp("madvise", buf, |
146 | min(sizeof("madvise")-1, count))) { | |
147 | clear_bit(enabled, &transparent_hugepage_flags); | |
444eb2a4 | 148 | clear_bit(deferred, &transparent_hugepage_flags); |
71e3aac0 AA |
149 | set_bit(req_madv, &transparent_hugepage_flags); |
150 | } else if (!memcmp("never", buf, | |
151 | min(sizeof("never")-1, count))) { | |
152 | clear_bit(enabled, &transparent_hugepage_flags); | |
153 | clear_bit(req_madv, &transparent_hugepage_flags); | |
444eb2a4 | 154 | clear_bit(deferred, &transparent_hugepage_flags); |
71e3aac0 AA |
155 | } else |
156 | return -EINVAL; | |
157 | ||
158 | return count; | |
159 | } | |
160 | ||
161 | static ssize_t enabled_show(struct kobject *kobj, | |
162 | struct kobj_attribute *attr, char *buf) | |
163 | { | |
444eb2a4 MG |
164 | if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags)) |
165 | return sprintf(buf, "[always] madvise never\n"); | |
166 | else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags)) | |
167 | return sprintf(buf, "always [madvise] never\n"); | |
168 | else | |
169 | return sprintf(buf, "always madvise [never]\n"); | |
71e3aac0 | 170 | } |
444eb2a4 | 171 | |
71e3aac0 AA |
172 | static ssize_t enabled_store(struct kobject *kobj, |
173 | struct kobj_attribute *attr, | |
174 | const char *buf, size_t count) | |
175 | { | |
ba76149f AA |
176 | ssize_t ret; |
177 | ||
444eb2a4 MG |
178 | ret = triple_flag_store(kobj, attr, buf, count, |
179 | TRANSPARENT_HUGEPAGE_FLAG, | |
ba76149f AA |
180 | TRANSPARENT_HUGEPAGE_FLAG, |
181 | TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG); | |
182 | ||
183 | if (ret > 0) { | |
b46e756f | 184 | int err = start_stop_khugepaged(); |
ba76149f AA |
185 | if (err) |
186 | ret = err; | |
187 | } | |
188 | ||
189 | return ret; | |
71e3aac0 AA |
190 | } |
191 | static struct kobj_attribute enabled_attr = | |
192 | __ATTR(enabled, 0644, enabled_show, enabled_store); | |
193 | ||
b46e756f | 194 | ssize_t single_hugepage_flag_show(struct kobject *kobj, |
71e3aac0 AA |
195 | struct kobj_attribute *attr, char *buf, |
196 | enum transparent_hugepage_flag flag) | |
197 | { | |
e27e6151 BH |
198 | return sprintf(buf, "%d\n", |
199 | !!test_bit(flag, &transparent_hugepage_flags)); | |
71e3aac0 | 200 | } |
e27e6151 | 201 | |
b46e756f | 202 | ssize_t single_hugepage_flag_store(struct kobject *kobj, |
71e3aac0 AA |
203 | struct kobj_attribute *attr, |
204 | const char *buf, size_t count, | |
205 | enum transparent_hugepage_flag flag) | |
206 | { | |
e27e6151 BH |
207 | unsigned long value; |
208 | int ret; | |
209 | ||
210 | ret = kstrtoul(buf, 10, &value); | |
211 | if (ret < 0) | |
212 | return ret; | |
213 | if (value > 1) | |
214 | return -EINVAL; | |
215 | ||
216 | if (value) | |
71e3aac0 | 217 | set_bit(flag, &transparent_hugepage_flags); |
e27e6151 | 218 | else |
71e3aac0 | 219 | clear_bit(flag, &transparent_hugepage_flags); |
71e3aac0 AA |
220 | |
221 | return count; | |
222 | } | |
223 | ||
224 | /* | |
225 | * Currently defrag only disables __GFP_NOWAIT for allocation. A blind | |
226 | * __GFP_REPEAT is too aggressive, it's never worth swapping tons of | |
227 | * memory just to allocate one more hugepage. | |
228 | */ | |
229 | static ssize_t defrag_show(struct kobject *kobj, | |
230 | struct kobj_attribute *attr, char *buf) | |
231 | { | |
444eb2a4 MG |
232 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) |
233 | return sprintf(buf, "[always] defer madvise never\n"); | |
234 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) | |
235 | return sprintf(buf, "always [defer] madvise never\n"); | |
236 | else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) | |
237 | return sprintf(buf, "always defer [madvise] never\n"); | |
238 | else | |
239 | return sprintf(buf, "always defer madvise [never]\n"); | |
240 | ||
71e3aac0 AA |
241 | } |
242 | static ssize_t defrag_store(struct kobject *kobj, | |
243 | struct kobj_attribute *attr, | |
244 | const char *buf, size_t count) | |
245 | { | |
444eb2a4 MG |
246 | return triple_flag_store(kobj, attr, buf, count, |
247 | TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, | |
248 | TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, | |
71e3aac0 AA |
249 | TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG); |
250 | } | |
251 | static struct kobj_attribute defrag_attr = | |
252 | __ATTR(defrag, 0644, defrag_show, defrag_store); | |
253 | ||
79da5407 KS |
254 | static ssize_t use_zero_page_show(struct kobject *kobj, |
255 | struct kobj_attribute *attr, char *buf) | |
256 | { | |
b46e756f | 257 | return single_hugepage_flag_show(kobj, attr, buf, |
79da5407 KS |
258 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); |
259 | } | |
260 | static ssize_t use_zero_page_store(struct kobject *kobj, | |
261 | struct kobj_attribute *attr, const char *buf, size_t count) | |
262 | { | |
b46e756f | 263 | return single_hugepage_flag_store(kobj, attr, buf, count, |
79da5407 KS |
264 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); |
265 | } | |
266 | static struct kobj_attribute use_zero_page_attr = | |
267 | __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); | |
71e3aac0 AA |
268 | #ifdef CONFIG_DEBUG_VM |
269 | static ssize_t debug_cow_show(struct kobject *kobj, | |
270 | struct kobj_attribute *attr, char *buf) | |
271 | { | |
b46e756f | 272 | return single_hugepage_flag_show(kobj, attr, buf, |
71e3aac0 AA |
273 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); |
274 | } | |
275 | static ssize_t debug_cow_store(struct kobject *kobj, | |
276 | struct kobj_attribute *attr, | |
277 | const char *buf, size_t count) | |
278 | { | |
b46e756f | 279 | return single_hugepage_flag_store(kobj, attr, buf, count, |
71e3aac0 AA |
280 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); |
281 | } | |
282 | static struct kobj_attribute debug_cow_attr = | |
283 | __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); | |
284 | #endif /* CONFIG_DEBUG_VM */ | |
285 | ||
286 | static struct attribute *hugepage_attr[] = { | |
287 | &enabled_attr.attr, | |
288 | &defrag_attr.attr, | |
79da5407 | 289 | &use_zero_page_attr.attr, |
e496cf3d | 290 | #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE) |
5a6e75f8 KS |
291 | &shmem_enabled_attr.attr, |
292 | #endif | |
71e3aac0 AA |
293 | #ifdef CONFIG_DEBUG_VM |
294 | &debug_cow_attr.attr, | |
295 | #endif | |
296 | NULL, | |
297 | }; | |
298 | ||
299 | static struct attribute_group hugepage_attr_group = { | |
300 | .attrs = hugepage_attr, | |
ba76149f AA |
301 | }; |
302 | ||
569e5590 | 303 | static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) |
71e3aac0 | 304 | { |
71e3aac0 AA |
305 | int err; |
306 | ||
569e5590 SL |
307 | *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); |
308 | if (unlikely(!*hugepage_kobj)) { | |
ae3a8c1c | 309 | pr_err("failed to create transparent hugepage kobject\n"); |
569e5590 | 310 | return -ENOMEM; |
ba76149f AA |
311 | } |
312 | ||
569e5590 | 313 | err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); |
ba76149f | 314 | if (err) { |
ae3a8c1c | 315 | pr_err("failed to register transparent hugepage group\n"); |
569e5590 | 316 | goto delete_obj; |
ba76149f AA |
317 | } |
318 | ||
569e5590 | 319 | err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); |
ba76149f | 320 | if (err) { |
ae3a8c1c | 321 | pr_err("failed to register transparent hugepage group\n"); |
569e5590 | 322 | goto remove_hp_group; |
ba76149f | 323 | } |
569e5590 SL |
324 | |
325 | return 0; | |
326 | ||
327 | remove_hp_group: | |
328 | sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); | |
329 | delete_obj: | |
330 | kobject_put(*hugepage_kobj); | |
331 | return err; | |
332 | } | |
333 | ||
334 | static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
335 | { | |
336 | sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); | |
337 | sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); | |
338 | kobject_put(hugepage_kobj); | |
339 | } | |
340 | #else | |
341 | static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) | |
342 | { | |
343 | return 0; | |
344 | } | |
345 | ||
346 | static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
347 | { | |
348 | } | |
349 | #endif /* CONFIG_SYSFS */ | |
350 | ||
351 | static int __init hugepage_init(void) | |
352 | { | |
353 | int err; | |
354 | struct kobject *hugepage_kobj; | |
355 | ||
356 | if (!has_transparent_hugepage()) { | |
357 | transparent_hugepage_flags = 0; | |
358 | return -EINVAL; | |
359 | } | |
360 | ||
ff20c2e0 KS |
361 | /* |
362 | * hugepages can't be allocated by the buddy allocator | |
363 | */ | |
364 | MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER); | |
365 | /* | |
366 | * we use page->mapping and page->index in second tail page | |
367 | * as list_head: assuming THP order >= 2 | |
368 | */ | |
369 | MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2); | |
370 | ||
569e5590 SL |
371 | err = hugepage_init_sysfs(&hugepage_kobj); |
372 | if (err) | |
65ebb64f | 373 | goto err_sysfs; |
ba76149f | 374 | |
b46e756f | 375 | err = khugepaged_init(); |
ba76149f | 376 | if (err) |
65ebb64f | 377 | goto err_slab; |
ba76149f | 378 | |
65ebb64f KS |
379 | err = register_shrinker(&huge_zero_page_shrinker); |
380 | if (err) | |
381 | goto err_hzp_shrinker; | |
9a982250 KS |
382 | err = register_shrinker(&deferred_split_shrinker); |
383 | if (err) | |
384 | goto err_split_shrinker; | |
97ae1749 | 385 | |
97562cd2 RR |
386 | /* |
387 | * By default disable transparent hugepages on smaller systems, | |
388 | * where the extra memory used could hurt more than TLB overhead | |
389 | * is likely to save. The admin can still enable it through /sys. | |
390 | */ | |
79553da2 | 391 | if (totalram_pages < (512 << (20 - PAGE_SHIFT))) { |
97562cd2 | 392 | transparent_hugepage_flags = 0; |
79553da2 KS |
393 | return 0; |
394 | } | |
97562cd2 | 395 | |
79553da2 | 396 | err = start_stop_khugepaged(); |
65ebb64f KS |
397 | if (err) |
398 | goto err_khugepaged; | |
ba76149f | 399 | |
569e5590 | 400 | return 0; |
65ebb64f | 401 | err_khugepaged: |
9a982250 KS |
402 | unregister_shrinker(&deferred_split_shrinker); |
403 | err_split_shrinker: | |
65ebb64f KS |
404 | unregister_shrinker(&huge_zero_page_shrinker); |
405 | err_hzp_shrinker: | |
b46e756f | 406 | khugepaged_destroy(); |
65ebb64f | 407 | err_slab: |
569e5590 | 408 | hugepage_exit_sysfs(hugepage_kobj); |
65ebb64f | 409 | err_sysfs: |
ba76149f | 410 | return err; |
71e3aac0 | 411 | } |
a64fb3cd | 412 | subsys_initcall(hugepage_init); |
71e3aac0 AA |
413 | |
414 | static int __init setup_transparent_hugepage(char *str) | |
415 | { | |
416 | int ret = 0; | |
417 | if (!str) | |
418 | goto out; | |
419 | if (!strcmp(str, "always")) { | |
420 | set_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
421 | &transparent_hugepage_flags); | |
422 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
423 | &transparent_hugepage_flags); | |
424 | ret = 1; | |
425 | } else if (!strcmp(str, "madvise")) { | |
426 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
427 | &transparent_hugepage_flags); | |
428 | set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
429 | &transparent_hugepage_flags); | |
430 | ret = 1; | |
431 | } else if (!strcmp(str, "never")) { | |
432 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
433 | &transparent_hugepage_flags); | |
434 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
435 | &transparent_hugepage_flags); | |
436 | ret = 1; | |
437 | } | |
438 | out: | |
439 | if (!ret) | |
ae3a8c1c | 440 | pr_warn("transparent_hugepage= cannot parse, ignored\n"); |
71e3aac0 AA |
441 | return ret; |
442 | } | |
443 | __setup("transparent_hugepage=", setup_transparent_hugepage); | |
444 | ||
b32967ff | 445 | pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) |
71e3aac0 AA |
446 | { |
447 | if (likely(vma->vm_flags & VM_WRITE)) | |
448 | pmd = pmd_mkwrite(pmd); | |
449 | return pmd; | |
450 | } | |
451 | ||
9a982250 KS |
452 | static inline struct list_head *page_deferred_list(struct page *page) |
453 | { | |
454 | /* | |
455 | * ->lru in the tail pages is occupied by compound_head. | |
456 | * Let's use ->mapping + ->index in the second tail page as list_head. | |
457 | */ | |
458 | return (struct list_head *)&page[2].mapping; | |
459 | } | |
460 | ||
461 | void prep_transhuge_page(struct page *page) | |
462 | { | |
463 | /* | |
464 | * we use page->mapping and page->indexlru in second tail page | |
465 | * as list_head: assuming THP order >= 2 | |
466 | */ | |
9a982250 KS |
467 | |
468 | INIT_LIST_HEAD(page_deferred_list(page)); | |
469 | set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR); | |
470 | } | |
471 | ||
bae473a4 KS |
472 | static int __do_huge_pmd_anonymous_page(struct fault_env *fe, struct page *page, |
473 | gfp_t gfp) | |
71e3aac0 | 474 | { |
bae473a4 | 475 | struct vm_area_struct *vma = fe->vma; |
00501b53 | 476 | struct mem_cgroup *memcg; |
71e3aac0 | 477 | pgtable_t pgtable; |
bae473a4 | 478 | unsigned long haddr = fe->address & HPAGE_PMD_MASK; |
71e3aac0 | 479 | |
309381fe | 480 | VM_BUG_ON_PAGE(!PageCompound(page), page); |
00501b53 | 481 | |
bae473a4 | 482 | if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) { |
6b251fc9 AA |
483 | put_page(page); |
484 | count_vm_event(THP_FAULT_FALLBACK); | |
485 | return VM_FAULT_FALLBACK; | |
486 | } | |
00501b53 | 487 | |
bae473a4 | 488 | pgtable = pte_alloc_one(vma->vm_mm, haddr); |
00501b53 | 489 | if (unlikely(!pgtable)) { |
f627c2f5 | 490 | mem_cgroup_cancel_charge(page, memcg, true); |
6b251fc9 | 491 | put_page(page); |
71e3aac0 | 492 | return VM_FAULT_OOM; |
00501b53 | 493 | } |
71e3aac0 AA |
494 | |
495 | clear_huge_page(page, haddr, HPAGE_PMD_NR); | |
52f37629 MK |
496 | /* |
497 | * The memory barrier inside __SetPageUptodate makes sure that | |
498 | * clear_huge_page writes become visible before the set_pmd_at() | |
499 | * write. | |
500 | */ | |
71e3aac0 AA |
501 | __SetPageUptodate(page); |
502 | ||
bae473a4 KS |
503 | fe->ptl = pmd_lock(vma->vm_mm, fe->pmd); |
504 | if (unlikely(!pmd_none(*fe->pmd))) { | |
505 | spin_unlock(fe->ptl); | |
f627c2f5 | 506 | mem_cgroup_cancel_charge(page, memcg, true); |
71e3aac0 | 507 | put_page(page); |
bae473a4 | 508 | pte_free(vma->vm_mm, pgtable); |
71e3aac0 AA |
509 | } else { |
510 | pmd_t entry; | |
6b251fc9 AA |
511 | |
512 | /* Deliver the page fault to userland */ | |
513 | if (userfaultfd_missing(vma)) { | |
514 | int ret; | |
515 | ||
bae473a4 | 516 | spin_unlock(fe->ptl); |
f627c2f5 | 517 | mem_cgroup_cancel_charge(page, memcg, true); |
6b251fc9 | 518 | put_page(page); |
bae473a4 KS |
519 | pte_free(vma->vm_mm, pgtable); |
520 | ret = handle_userfault(fe, VM_UFFD_MISSING); | |
6b251fc9 AA |
521 | VM_BUG_ON(ret & VM_FAULT_FALLBACK); |
522 | return ret; | |
523 | } | |
524 | ||
3122359a KS |
525 | entry = mk_huge_pmd(page, vma->vm_page_prot); |
526 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
d281ee61 | 527 | page_add_new_anon_rmap(page, vma, haddr, true); |
f627c2f5 | 528 | mem_cgroup_commit_charge(page, memcg, false, true); |
00501b53 | 529 | lru_cache_add_active_or_unevictable(page, vma); |
bae473a4 KS |
530 | pgtable_trans_huge_deposit(vma->vm_mm, fe->pmd, pgtable); |
531 | set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry); | |
532 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); | |
533 | atomic_long_inc(&vma->vm_mm->nr_ptes); | |
534 | spin_unlock(fe->ptl); | |
6b251fc9 | 535 | count_vm_event(THP_FAULT_ALLOC); |
71e3aac0 AA |
536 | } |
537 | ||
aa2e878e | 538 | return 0; |
71e3aac0 AA |
539 | } |
540 | ||
444eb2a4 | 541 | /* |
25160354 VB |
542 | * If THP defrag is set to always then directly reclaim/compact as necessary |
543 | * If set to defer then do only background reclaim/compact and defer to khugepaged | |
444eb2a4 | 544 | * If set to madvise and the VMA is flagged then directly reclaim/compact |
25160354 | 545 | * When direct reclaim/compact is allowed, don't retry except for flagged VMA's |
444eb2a4 MG |
546 | */ |
547 | static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma) | |
548 | { | |
25160354 VB |
549 | bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE); |
550 | ||
551 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, | |
552 | &transparent_hugepage_flags) && vma_madvised) | |
553 | return GFP_TRANSHUGE; | |
554 | else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, | |
555 | &transparent_hugepage_flags)) | |
556 | return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM; | |
557 | else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, | |
558 | &transparent_hugepage_flags)) | |
559 | return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY); | |
560 | ||
561 | return GFP_TRANSHUGE_LIGHT; | |
444eb2a4 MG |
562 | } |
563 | ||
c4088ebd | 564 | /* Caller must hold page table lock. */ |
d295e341 | 565 | static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, |
97ae1749 | 566 | struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, |
5918d10a | 567 | struct page *zero_page) |
fc9fe822 KS |
568 | { |
569 | pmd_t entry; | |
7c414164 AM |
570 | if (!pmd_none(*pmd)) |
571 | return false; | |
5918d10a | 572 | entry = mk_pmd(zero_page, vma->vm_page_prot); |
fc9fe822 | 573 | entry = pmd_mkhuge(entry); |
12c9d70b MW |
574 | if (pgtable) |
575 | pgtable_trans_huge_deposit(mm, pmd, pgtable); | |
fc9fe822 | 576 | set_pmd_at(mm, haddr, pmd, entry); |
e1f56c89 | 577 | atomic_long_inc(&mm->nr_ptes); |
7c414164 | 578 | return true; |
fc9fe822 KS |
579 | } |
580 | ||
bae473a4 | 581 | int do_huge_pmd_anonymous_page(struct fault_env *fe) |
71e3aac0 | 582 | { |
bae473a4 | 583 | struct vm_area_struct *vma = fe->vma; |
077fcf11 | 584 | gfp_t gfp; |
71e3aac0 | 585 | struct page *page; |
bae473a4 | 586 | unsigned long haddr = fe->address & HPAGE_PMD_MASK; |
71e3aac0 | 587 | |
128ec037 | 588 | if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end) |
c0292554 | 589 | return VM_FAULT_FALLBACK; |
128ec037 KS |
590 | if (unlikely(anon_vma_prepare(vma))) |
591 | return VM_FAULT_OOM; | |
6d50e60c | 592 | if (unlikely(khugepaged_enter(vma, vma->vm_flags))) |
128ec037 | 593 | return VM_FAULT_OOM; |
bae473a4 KS |
594 | if (!(fe->flags & FAULT_FLAG_WRITE) && |
595 | !mm_forbids_zeropage(vma->vm_mm) && | |
128ec037 KS |
596 | transparent_hugepage_use_zero_page()) { |
597 | pgtable_t pgtable; | |
598 | struct page *zero_page; | |
599 | bool set; | |
6b251fc9 | 600 | int ret; |
bae473a4 | 601 | pgtable = pte_alloc_one(vma->vm_mm, haddr); |
128ec037 | 602 | if (unlikely(!pgtable)) |
ba76149f | 603 | return VM_FAULT_OOM; |
128ec037 KS |
604 | zero_page = get_huge_zero_page(); |
605 | if (unlikely(!zero_page)) { | |
bae473a4 | 606 | pte_free(vma->vm_mm, pgtable); |
81ab4201 | 607 | count_vm_event(THP_FAULT_FALLBACK); |
c0292554 | 608 | return VM_FAULT_FALLBACK; |
b9bbfbe3 | 609 | } |
bae473a4 | 610 | fe->ptl = pmd_lock(vma->vm_mm, fe->pmd); |
6b251fc9 AA |
611 | ret = 0; |
612 | set = false; | |
bae473a4 | 613 | if (pmd_none(*fe->pmd)) { |
6b251fc9 | 614 | if (userfaultfd_missing(vma)) { |
bae473a4 KS |
615 | spin_unlock(fe->ptl); |
616 | ret = handle_userfault(fe, VM_UFFD_MISSING); | |
6b251fc9 AA |
617 | VM_BUG_ON(ret & VM_FAULT_FALLBACK); |
618 | } else { | |
bae473a4 KS |
619 | set_huge_zero_page(pgtable, vma->vm_mm, vma, |
620 | haddr, fe->pmd, zero_page); | |
621 | spin_unlock(fe->ptl); | |
6b251fc9 AA |
622 | set = true; |
623 | } | |
624 | } else | |
bae473a4 | 625 | spin_unlock(fe->ptl); |
128ec037 | 626 | if (!set) { |
bae473a4 | 627 | pte_free(vma->vm_mm, pgtable); |
128ec037 | 628 | put_huge_zero_page(); |
edad9d2c | 629 | } |
6b251fc9 | 630 | return ret; |
71e3aac0 | 631 | } |
444eb2a4 | 632 | gfp = alloc_hugepage_direct_gfpmask(vma); |
077fcf11 | 633 | page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); |
128ec037 KS |
634 | if (unlikely(!page)) { |
635 | count_vm_event(THP_FAULT_FALLBACK); | |
c0292554 | 636 | return VM_FAULT_FALLBACK; |
128ec037 | 637 | } |
9a982250 | 638 | prep_transhuge_page(page); |
bae473a4 | 639 | return __do_huge_pmd_anonymous_page(fe, page, gfp); |
71e3aac0 AA |
640 | } |
641 | ||
ae18d6dc | 642 | static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, |
f25748e3 | 643 | pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write) |
5cad465d MW |
644 | { |
645 | struct mm_struct *mm = vma->vm_mm; | |
646 | pmd_t entry; | |
647 | spinlock_t *ptl; | |
648 | ||
649 | ptl = pmd_lock(mm, pmd); | |
f25748e3 DW |
650 | entry = pmd_mkhuge(pfn_t_pmd(pfn, prot)); |
651 | if (pfn_t_devmap(pfn)) | |
652 | entry = pmd_mkdevmap(entry); | |
01871e59 RZ |
653 | if (write) { |
654 | entry = pmd_mkyoung(pmd_mkdirty(entry)); | |
655 | entry = maybe_pmd_mkwrite(entry, vma); | |
5cad465d | 656 | } |
01871e59 RZ |
657 | set_pmd_at(mm, addr, pmd, entry); |
658 | update_mmu_cache_pmd(vma, addr, pmd); | |
5cad465d | 659 | spin_unlock(ptl); |
5cad465d MW |
660 | } |
661 | ||
662 | int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, | |
f25748e3 | 663 | pmd_t *pmd, pfn_t pfn, bool write) |
5cad465d MW |
664 | { |
665 | pgprot_t pgprot = vma->vm_page_prot; | |
666 | /* | |
667 | * If we had pmd_special, we could avoid all these restrictions, | |
668 | * but we need to be consistent with PTEs and architectures that | |
669 | * can't support a 'special' bit. | |
670 | */ | |
671 | BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); | |
672 | BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == | |
673 | (VM_PFNMAP|VM_MIXEDMAP)); | |
674 | BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); | |
f25748e3 | 675 | BUG_ON(!pfn_t_devmap(pfn)); |
5cad465d MW |
676 | |
677 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
678 | return VM_FAULT_SIGBUS; | |
679 | if (track_pfn_insert(vma, &pgprot, pfn)) | |
680 | return VM_FAULT_SIGBUS; | |
ae18d6dc MW |
681 | insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write); |
682 | return VM_FAULT_NOPAGE; | |
5cad465d | 683 | } |
dee41079 | 684 | EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd); |
5cad465d | 685 | |
3565fce3 DW |
686 | static void touch_pmd(struct vm_area_struct *vma, unsigned long addr, |
687 | pmd_t *pmd) | |
688 | { | |
689 | pmd_t _pmd; | |
690 | ||
691 | /* | |
692 | * We should set the dirty bit only for FOLL_WRITE but for now | |
693 | * the dirty bit in the pmd is meaningless. And if the dirty | |
694 | * bit will become meaningful and we'll only set it with | |
695 | * FOLL_WRITE, an atomic set_bit will be required on the pmd to | |
696 | * set the young bit, instead of the current set_pmd_at. | |
697 | */ | |
698 | _pmd = pmd_mkyoung(pmd_mkdirty(*pmd)); | |
699 | if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, | |
700 | pmd, _pmd, 1)) | |
701 | update_mmu_cache_pmd(vma, addr, pmd); | |
702 | } | |
703 | ||
704 | struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr, | |
705 | pmd_t *pmd, int flags) | |
706 | { | |
707 | unsigned long pfn = pmd_pfn(*pmd); | |
708 | struct mm_struct *mm = vma->vm_mm; | |
709 | struct dev_pagemap *pgmap; | |
710 | struct page *page; | |
711 | ||
712 | assert_spin_locked(pmd_lockptr(mm, pmd)); | |
713 | ||
714 | if (flags & FOLL_WRITE && !pmd_write(*pmd)) | |
715 | return NULL; | |
716 | ||
717 | if (pmd_present(*pmd) && pmd_devmap(*pmd)) | |
718 | /* pass */; | |
719 | else | |
720 | return NULL; | |
721 | ||
722 | if (flags & FOLL_TOUCH) | |
723 | touch_pmd(vma, addr, pmd); | |
724 | ||
725 | /* | |
726 | * device mapped pages can only be returned if the | |
727 | * caller will manage the page reference count. | |
728 | */ | |
729 | if (!(flags & FOLL_GET)) | |
730 | return ERR_PTR(-EEXIST); | |
731 | ||
732 | pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT; | |
733 | pgmap = get_dev_pagemap(pfn, NULL); | |
734 | if (!pgmap) | |
735 | return ERR_PTR(-EFAULT); | |
736 | page = pfn_to_page(pfn); | |
737 | get_page(page); | |
738 | put_dev_pagemap(pgmap); | |
739 | ||
740 | return page; | |
741 | } | |
742 | ||
71e3aac0 AA |
743 | int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, |
744 | pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, | |
745 | struct vm_area_struct *vma) | |
746 | { | |
c4088ebd | 747 | spinlock_t *dst_ptl, *src_ptl; |
71e3aac0 AA |
748 | struct page *src_page; |
749 | pmd_t pmd; | |
12c9d70b | 750 | pgtable_t pgtable = NULL; |
628d47ce | 751 | int ret = -ENOMEM; |
71e3aac0 | 752 | |
628d47ce KS |
753 | /* Skip if can be re-fill on fault */ |
754 | if (!vma_is_anonymous(vma)) | |
755 | return 0; | |
756 | ||
757 | pgtable = pte_alloc_one(dst_mm, addr); | |
758 | if (unlikely(!pgtable)) | |
759 | goto out; | |
71e3aac0 | 760 | |
c4088ebd KS |
761 | dst_ptl = pmd_lock(dst_mm, dst_pmd); |
762 | src_ptl = pmd_lockptr(src_mm, src_pmd); | |
763 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
71e3aac0 AA |
764 | |
765 | ret = -EAGAIN; | |
766 | pmd = *src_pmd; | |
628d47ce | 767 | if (unlikely(!pmd_trans_huge(pmd))) { |
71e3aac0 AA |
768 | pte_free(dst_mm, pgtable); |
769 | goto out_unlock; | |
770 | } | |
fc9fe822 | 771 | /* |
c4088ebd | 772 | * When page table lock is held, the huge zero pmd should not be |
fc9fe822 KS |
773 | * under splitting since we don't split the page itself, only pmd to |
774 | * a page table. | |
775 | */ | |
776 | if (is_huge_zero_pmd(pmd)) { | |
5918d10a | 777 | struct page *zero_page; |
97ae1749 KS |
778 | /* |
779 | * get_huge_zero_page() will never allocate a new page here, | |
780 | * since we already have a zero page to copy. It just takes a | |
781 | * reference. | |
782 | */ | |
5918d10a | 783 | zero_page = get_huge_zero_page(); |
6b251fc9 | 784 | set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, |
5918d10a | 785 | zero_page); |
fc9fe822 KS |
786 | ret = 0; |
787 | goto out_unlock; | |
788 | } | |
de466bd6 | 789 | |
628d47ce KS |
790 | src_page = pmd_page(pmd); |
791 | VM_BUG_ON_PAGE(!PageHead(src_page), src_page); | |
792 | get_page(src_page); | |
793 | page_dup_rmap(src_page, true); | |
794 | add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); | |
795 | atomic_long_inc(&dst_mm->nr_ptes); | |
796 | pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); | |
71e3aac0 AA |
797 | |
798 | pmdp_set_wrprotect(src_mm, addr, src_pmd); | |
799 | pmd = pmd_mkold(pmd_wrprotect(pmd)); | |
800 | set_pmd_at(dst_mm, addr, dst_pmd, pmd); | |
71e3aac0 AA |
801 | |
802 | ret = 0; | |
803 | out_unlock: | |
c4088ebd KS |
804 | spin_unlock(src_ptl); |
805 | spin_unlock(dst_ptl); | |
71e3aac0 AA |
806 | out: |
807 | return ret; | |
808 | } | |
809 | ||
bae473a4 | 810 | void huge_pmd_set_accessed(struct fault_env *fe, pmd_t orig_pmd) |
a1dd450b WD |
811 | { |
812 | pmd_t entry; | |
813 | unsigned long haddr; | |
814 | ||
bae473a4 KS |
815 | fe->ptl = pmd_lock(fe->vma->vm_mm, fe->pmd); |
816 | if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) | |
a1dd450b WD |
817 | goto unlock; |
818 | ||
819 | entry = pmd_mkyoung(orig_pmd); | |
bae473a4 KS |
820 | haddr = fe->address & HPAGE_PMD_MASK; |
821 | if (pmdp_set_access_flags(fe->vma, haddr, fe->pmd, entry, | |
822 | fe->flags & FAULT_FLAG_WRITE)) | |
823 | update_mmu_cache_pmd(fe->vma, fe->address, fe->pmd); | |
a1dd450b WD |
824 | |
825 | unlock: | |
bae473a4 | 826 | spin_unlock(fe->ptl); |
a1dd450b WD |
827 | } |
828 | ||
bae473a4 KS |
829 | static int do_huge_pmd_wp_page_fallback(struct fault_env *fe, pmd_t orig_pmd, |
830 | struct page *page) | |
71e3aac0 | 831 | { |
bae473a4 KS |
832 | struct vm_area_struct *vma = fe->vma; |
833 | unsigned long haddr = fe->address & HPAGE_PMD_MASK; | |
00501b53 | 834 | struct mem_cgroup *memcg; |
71e3aac0 AA |
835 | pgtable_t pgtable; |
836 | pmd_t _pmd; | |
837 | int ret = 0, i; | |
838 | struct page **pages; | |
2ec74c3e SG |
839 | unsigned long mmun_start; /* For mmu_notifiers */ |
840 | unsigned long mmun_end; /* For mmu_notifiers */ | |
71e3aac0 AA |
841 | |
842 | pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR, | |
843 | GFP_KERNEL); | |
844 | if (unlikely(!pages)) { | |
845 | ret |= VM_FAULT_OOM; | |
846 | goto out; | |
847 | } | |
848 | ||
849 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
cc5d462f | 850 | pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE | |
bae473a4 KS |
851 | __GFP_OTHER_NODE, vma, |
852 | fe->address, page_to_nid(page)); | |
b9bbfbe3 | 853 | if (unlikely(!pages[i] || |
bae473a4 KS |
854 | mem_cgroup_try_charge(pages[i], vma->vm_mm, |
855 | GFP_KERNEL, &memcg, false))) { | |
b9bbfbe3 | 856 | if (pages[i]) |
71e3aac0 | 857 | put_page(pages[i]); |
b9bbfbe3 | 858 | while (--i >= 0) { |
00501b53 JW |
859 | memcg = (void *)page_private(pages[i]); |
860 | set_page_private(pages[i], 0); | |
f627c2f5 KS |
861 | mem_cgroup_cancel_charge(pages[i], memcg, |
862 | false); | |
b9bbfbe3 AA |
863 | put_page(pages[i]); |
864 | } | |
71e3aac0 AA |
865 | kfree(pages); |
866 | ret |= VM_FAULT_OOM; | |
867 | goto out; | |
868 | } | |
00501b53 | 869 | set_page_private(pages[i], (unsigned long)memcg); |
71e3aac0 AA |
870 | } |
871 | ||
872 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
873 | copy_user_highpage(pages[i], page + i, | |
0089e485 | 874 | haddr + PAGE_SIZE * i, vma); |
71e3aac0 AA |
875 | __SetPageUptodate(pages[i]); |
876 | cond_resched(); | |
877 | } | |
878 | ||
2ec74c3e SG |
879 | mmun_start = haddr; |
880 | mmun_end = haddr + HPAGE_PMD_SIZE; | |
bae473a4 | 881 | mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end); |
2ec74c3e | 882 | |
bae473a4 KS |
883 | fe->ptl = pmd_lock(vma->vm_mm, fe->pmd); |
884 | if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) | |
71e3aac0 | 885 | goto out_free_pages; |
309381fe | 886 | VM_BUG_ON_PAGE(!PageHead(page), page); |
71e3aac0 | 887 | |
bae473a4 | 888 | pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd); |
71e3aac0 AA |
889 | /* leave pmd empty until pte is filled */ |
890 | ||
bae473a4 KS |
891 | pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, fe->pmd); |
892 | pmd_populate(vma->vm_mm, &_pmd, pgtable); | |
71e3aac0 AA |
893 | |
894 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
bae473a4 | 895 | pte_t entry; |
71e3aac0 AA |
896 | entry = mk_pte(pages[i], vma->vm_page_prot); |
897 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); | |
00501b53 JW |
898 | memcg = (void *)page_private(pages[i]); |
899 | set_page_private(pages[i], 0); | |
bae473a4 | 900 | page_add_new_anon_rmap(pages[i], fe->vma, haddr, false); |
f627c2f5 | 901 | mem_cgroup_commit_charge(pages[i], memcg, false, false); |
00501b53 | 902 | lru_cache_add_active_or_unevictable(pages[i], vma); |
bae473a4 KS |
903 | fe->pte = pte_offset_map(&_pmd, haddr); |
904 | VM_BUG_ON(!pte_none(*fe->pte)); | |
905 | set_pte_at(vma->vm_mm, haddr, fe->pte, entry); | |
906 | pte_unmap(fe->pte); | |
71e3aac0 AA |
907 | } |
908 | kfree(pages); | |
909 | ||
71e3aac0 | 910 | smp_wmb(); /* make pte visible before pmd */ |
bae473a4 | 911 | pmd_populate(vma->vm_mm, fe->pmd, pgtable); |
d281ee61 | 912 | page_remove_rmap(page, true); |
bae473a4 | 913 | spin_unlock(fe->ptl); |
71e3aac0 | 914 | |
bae473a4 | 915 | mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); |
2ec74c3e | 916 | |
71e3aac0 AA |
917 | ret |= VM_FAULT_WRITE; |
918 | put_page(page); | |
919 | ||
920 | out: | |
921 | return ret; | |
922 | ||
923 | out_free_pages: | |
bae473a4 KS |
924 | spin_unlock(fe->ptl); |
925 | mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); | |
b9bbfbe3 | 926 | for (i = 0; i < HPAGE_PMD_NR; i++) { |
00501b53 JW |
927 | memcg = (void *)page_private(pages[i]); |
928 | set_page_private(pages[i], 0); | |
f627c2f5 | 929 | mem_cgroup_cancel_charge(pages[i], memcg, false); |
71e3aac0 | 930 | put_page(pages[i]); |
b9bbfbe3 | 931 | } |
71e3aac0 AA |
932 | kfree(pages); |
933 | goto out; | |
934 | } | |
935 | ||
bae473a4 | 936 | int do_huge_pmd_wp_page(struct fault_env *fe, pmd_t orig_pmd) |
71e3aac0 | 937 | { |
bae473a4 | 938 | struct vm_area_struct *vma = fe->vma; |
93b4796d | 939 | struct page *page = NULL, *new_page; |
00501b53 | 940 | struct mem_cgroup *memcg; |
bae473a4 | 941 | unsigned long haddr = fe->address & HPAGE_PMD_MASK; |
2ec74c3e SG |
942 | unsigned long mmun_start; /* For mmu_notifiers */ |
943 | unsigned long mmun_end; /* For mmu_notifiers */ | |
3b363692 | 944 | gfp_t huge_gfp; /* for allocation and charge */ |
bae473a4 | 945 | int ret = 0; |
71e3aac0 | 946 | |
bae473a4 | 947 | fe->ptl = pmd_lockptr(vma->vm_mm, fe->pmd); |
81d1b09c | 948 | VM_BUG_ON_VMA(!vma->anon_vma, vma); |
93b4796d KS |
949 | if (is_huge_zero_pmd(orig_pmd)) |
950 | goto alloc; | |
bae473a4 KS |
951 | spin_lock(fe->ptl); |
952 | if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) | |
71e3aac0 AA |
953 | goto out_unlock; |
954 | ||
955 | page = pmd_page(orig_pmd); | |
309381fe | 956 | VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); |
1f25fe20 KS |
957 | /* |
958 | * We can only reuse the page if nobody else maps the huge page or it's | |
6d0a07ed | 959 | * part. |
1f25fe20 | 960 | */ |
6d0a07ed | 961 | if (page_trans_huge_mapcount(page, NULL) == 1) { |
71e3aac0 AA |
962 | pmd_t entry; |
963 | entry = pmd_mkyoung(orig_pmd); | |
964 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
bae473a4 KS |
965 | if (pmdp_set_access_flags(vma, haddr, fe->pmd, entry, 1)) |
966 | update_mmu_cache_pmd(vma, fe->address, fe->pmd); | |
71e3aac0 AA |
967 | ret |= VM_FAULT_WRITE; |
968 | goto out_unlock; | |
969 | } | |
ddc58f27 | 970 | get_page(page); |
bae473a4 | 971 | spin_unlock(fe->ptl); |
93b4796d | 972 | alloc: |
71e3aac0 | 973 | if (transparent_hugepage_enabled(vma) && |
077fcf11 | 974 | !transparent_hugepage_debug_cow()) { |
444eb2a4 | 975 | huge_gfp = alloc_hugepage_direct_gfpmask(vma); |
3b363692 | 976 | new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER); |
077fcf11 | 977 | } else |
71e3aac0 AA |
978 | new_page = NULL; |
979 | ||
9a982250 KS |
980 | if (likely(new_page)) { |
981 | prep_transhuge_page(new_page); | |
982 | } else { | |
eecc1e42 | 983 | if (!page) { |
bae473a4 | 984 | split_huge_pmd(vma, fe->pmd, fe->address); |
e9b71ca9 | 985 | ret |= VM_FAULT_FALLBACK; |
93b4796d | 986 | } else { |
bae473a4 | 987 | ret = do_huge_pmd_wp_page_fallback(fe, orig_pmd, page); |
9845cbbd | 988 | if (ret & VM_FAULT_OOM) { |
bae473a4 | 989 | split_huge_pmd(vma, fe->pmd, fe->address); |
9845cbbd KS |
990 | ret |= VM_FAULT_FALLBACK; |
991 | } | |
ddc58f27 | 992 | put_page(page); |
93b4796d | 993 | } |
17766dde | 994 | count_vm_event(THP_FAULT_FALLBACK); |
71e3aac0 AA |
995 | goto out; |
996 | } | |
997 | ||
bae473a4 KS |
998 | if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm, |
999 | huge_gfp, &memcg, true))) { | |
b9bbfbe3 | 1000 | put_page(new_page); |
bae473a4 KS |
1001 | split_huge_pmd(vma, fe->pmd, fe->address); |
1002 | if (page) | |
ddc58f27 | 1003 | put_page(page); |
9845cbbd | 1004 | ret |= VM_FAULT_FALLBACK; |
17766dde | 1005 | count_vm_event(THP_FAULT_FALLBACK); |
b9bbfbe3 AA |
1006 | goto out; |
1007 | } | |
1008 | ||
17766dde DR |
1009 | count_vm_event(THP_FAULT_ALLOC); |
1010 | ||
eecc1e42 | 1011 | if (!page) |
93b4796d KS |
1012 | clear_huge_page(new_page, haddr, HPAGE_PMD_NR); |
1013 | else | |
1014 | copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR); | |
71e3aac0 AA |
1015 | __SetPageUptodate(new_page); |
1016 | ||
2ec74c3e SG |
1017 | mmun_start = haddr; |
1018 | mmun_end = haddr + HPAGE_PMD_SIZE; | |
bae473a4 | 1019 | mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end); |
2ec74c3e | 1020 | |
bae473a4 | 1021 | spin_lock(fe->ptl); |
93b4796d | 1022 | if (page) |
ddc58f27 | 1023 | put_page(page); |
bae473a4 KS |
1024 | if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) { |
1025 | spin_unlock(fe->ptl); | |
f627c2f5 | 1026 | mem_cgroup_cancel_charge(new_page, memcg, true); |
71e3aac0 | 1027 | put_page(new_page); |
2ec74c3e | 1028 | goto out_mn; |
b9bbfbe3 | 1029 | } else { |
71e3aac0 | 1030 | pmd_t entry; |
3122359a KS |
1031 | entry = mk_huge_pmd(new_page, vma->vm_page_prot); |
1032 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
bae473a4 | 1033 | pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd); |
d281ee61 | 1034 | page_add_new_anon_rmap(new_page, vma, haddr, true); |
f627c2f5 | 1035 | mem_cgroup_commit_charge(new_page, memcg, false, true); |
00501b53 | 1036 | lru_cache_add_active_or_unevictable(new_page, vma); |
bae473a4 KS |
1037 | set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry); |
1038 | update_mmu_cache_pmd(vma, fe->address, fe->pmd); | |
eecc1e42 | 1039 | if (!page) { |
bae473a4 | 1040 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); |
97ae1749 KS |
1041 | put_huge_zero_page(); |
1042 | } else { | |
309381fe | 1043 | VM_BUG_ON_PAGE(!PageHead(page), page); |
d281ee61 | 1044 | page_remove_rmap(page, true); |
93b4796d KS |
1045 | put_page(page); |
1046 | } | |
71e3aac0 AA |
1047 | ret |= VM_FAULT_WRITE; |
1048 | } | |
bae473a4 | 1049 | spin_unlock(fe->ptl); |
2ec74c3e | 1050 | out_mn: |
bae473a4 | 1051 | mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); |
71e3aac0 AA |
1052 | out: |
1053 | return ret; | |
2ec74c3e | 1054 | out_unlock: |
bae473a4 | 1055 | spin_unlock(fe->ptl); |
2ec74c3e | 1056 | return ret; |
71e3aac0 AA |
1057 | } |
1058 | ||
b676b293 | 1059 | struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, |
71e3aac0 AA |
1060 | unsigned long addr, |
1061 | pmd_t *pmd, | |
1062 | unsigned int flags) | |
1063 | { | |
b676b293 | 1064 | struct mm_struct *mm = vma->vm_mm; |
71e3aac0 AA |
1065 | struct page *page = NULL; |
1066 | ||
c4088ebd | 1067 | assert_spin_locked(pmd_lockptr(mm, pmd)); |
71e3aac0 AA |
1068 | |
1069 | if (flags & FOLL_WRITE && !pmd_write(*pmd)) | |
1070 | goto out; | |
1071 | ||
85facf25 KS |
1072 | /* Avoid dumping huge zero page */ |
1073 | if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) | |
1074 | return ERR_PTR(-EFAULT); | |
1075 | ||
2b4847e7 | 1076 | /* Full NUMA hinting faults to serialise migration in fault paths */ |
8a0516ed | 1077 | if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) |
2b4847e7 MG |
1078 | goto out; |
1079 | ||
71e3aac0 | 1080 | page = pmd_page(*pmd); |
309381fe | 1081 | VM_BUG_ON_PAGE(!PageHead(page), page); |
3565fce3 DW |
1082 | if (flags & FOLL_TOUCH) |
1083 | touch_pmd(vma, addr, pmd); | |
de60f5f1 | 1084 | if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { |
e90309c9 KS |
1085 | /* |
1086 | * We don't mlock() pte-mapped THPs. This way we can avoid | |
1087 | * leaking mlocked pages into non-VM_LOCKED VMAs. | |
1088 | * | |
9a73f61b KS |
1089 | * For anon THP: |
1090 | * | |
e90309c9 KS |
1091 | * In most cases the pmd is the only mapping of the page as we |
1092 | * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for | |
1093 | * writable private mappings in populate_vma_page_range(). | |
1094 | * | |
1095 | * The only scenario when we have the page shared here is if we | |
1096 | * mlocking read-only mapping shared over fork(). We skip | |
1097 | * mlocking such pages. | |
9a73f61b KS |
1098 | * |
1099 | * For file THP: | |
1100 | * | |
1101 | * We can expect PageDoubleMap() to be stable under page lock: | |
1102 | * for file pages we set it in page_add_file_rmap(), which | |
1103 | * requires page to be locked. | |
e90309c9 | 1104 | */ |
9a73f61b KS |
1105 | |
1106 | if (PageAnon(page) && compound_mapcount(page) != 1) | |
1107 | goto skip_mlock; | |
1108 | if (PageDoubleMap(page) || !page->mapping) | |
1109 | goto skip_mlock; | |
1110 | if (!trylock_page(page)) | |
1111 | goto skip_mlock; | |
1112 | lru_add_drain(); | |
1113 | if (page->mapping && !PageDoubleMap(page)) | |
1114 | mlock_vma_page(page); | |
1115 | unlock_page(page); | |
b676b293 | 1116 | } |
9a73f61b | 1117 | skip_mlock: |
71e3aac0 | 1118 | page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; |
309381fe | 1119 | VM_BUG_ON_PAGE(!PageCompound(page), page); |
71e3aac0 | 1120 | if (flags & FOLL_GET) |
ddc58f27 | 1121 | get_page(page); |
71e3aac0 AA |
1122 | |
1123 | out: | |
1124 | return page; | |
1125 | } | |
1126 | ||
d10e63f2 | 1127 | /* NUMA hinting page fault entry point for trans huge pmds */ |
bae473a4 | 1128 | int do_huge_pmd_numa_page(struct fault_env *fe, pmd_t pmd) |
d10e63f2 | 1129 | { |
bae473a4 | 1130 | struct vm_area_struct *vma = fe->vma; |
b8916634 | 1131 | struct anon_vma *anon_vma = NULL; |
b32967ff | 1132 | struct page *page; |
bae473a4 | 1133 | unsigned long haddr = fe->address & HPAGE_PMD_MASK; |
8191acbd | 1134 | int page_nid = -1, this_nid = numa_node_id(); |
90572890 | 1135 | int target_nid, last_cpupid = -1; |
8191acbd MG |
1136 | bool page_locked; |
1137 | bool migrated = false; | |
b191f9b1 | 1138 | bool was_writable; |
6688cc05 | 1139 | int flags = 0; |
d10e63f2 | 1140 | |
c0e7cad9 MG |
1141 | /* A PROT_NONE fault should not end up here */ |
1142 | BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))); | |
1143 | ||
bae473a4 KS |
1144 | fe->ptl = pmd_lock(vma->vm_mm, fe->pmd); |
1145 | if (unlikely(!pmd_same(pmd, *fe->pmd))) | |
d10e63f2 MG |
1146 | goto out_unlock; |
1147 | ||
de466bd6 MG |
1148 | /* |
1149 | * If there are potential migrations, wait for completion and retry | |
1150 | * without disrupting NUMA hinting information. Do not relock and | |
1151 | * check_same as the page may no longer be mapped. | |
1152 | */ | |
bae473a4 KS |
1153 | if (unlikely(pmd_trans_migrating(*fe->pmd))) { |
1154 | page = pmd_page(*fe->pmd); | |
1155 | spin_unlock(fe->ptl); | |
5d833062 | 1156 | wait_on_page_locked(page); |
de466bd6 MG |
1157 | goto out; |
1158 | } | |
1159 | ||
d10e63f2 | 1160 | page = pmd_page(pmd); |
a1a46184 | 1161 | BUG_ON(is_huge_zero_page(page)); |
8191acbd | 1162 | page_nid = page_to_nid(page); |
90572890 | 1163 | last_cpupid = page_cpupid_last(page); |
03c5a6e1 | 1164 | count_vm_numa_event(NUMA_HINT_FAULTS); |
04bb2f94 | 1165 | if (page_nid == this_nid) { |
03c5a6e1 | 1166 | count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); |
04bb2f94 RR |
1167 | flags |= TNF_FAULT_LOCAL; |
1168 | } | |
4daae3b4 | 1169 | |
bea66fbd MG |
1170 | /* See similar comment in do_numa_page for explanation */ |
1171 | if (!(vma->vm_flags & VM_WRITE)) | |
6688cc05 PZ |
1172 | flags |= TNF_NO_GROUP; |
1173 | ||
ff9042b1 MG |
1174 | /* |
1175 | * Acquire the page lock to serialise THP migrations but avoid dropping | |
1176 | * page_table_lock if at all possible | |
1177 | */ | |
b8916634 MG |
1178 | page_locked = trylock_page(page); |
1179 | target_nid = mpol_misplaced(page, vma, haddr); | |
1180 | if (target_nid == -1) { | |
1181 | /* If the page was locked, there are no parallel migrations */ | |
a54a407f | 1182 | if (page_locked) |
b8916634 | 1183 | goto clear_pmdnuma; |
2b4847e7 | 1184 | } |
4daae3b4 | 1185 | |
de466bd6 | 1186 | /* Migration could have started since the pmd_trans_migrating check */ |
2b4847e7 | 1187 | if (!page_locked) { |
bae473a4 | 1188 | spin_unlock(fe->ptl); |
b8916634 | 1189 | wait_on_page_locked(page); |
a54a407f | 1190 | page_nid = -1; |
b8916634 MG |
1191 | goto out; |
1192 | } | |
1193 | ||
2b4847e7 MG |
1194 | /* |
1195 | * Page is misplaced. Page lock serialises migrations. Acquire anon_vma | |
1196 | * to serialises splits | |
1197 | */ | |
b8916634 | 1198 | get_page(page); |
bae473a4 | 1199 | spin_unlock(fe->ptl); |
b8916634 | 1200 | anon_vma = page_lock_anon_vma_read(page); |
4daae3b4 | 1201 | |
c69307d5 | 1202 | /* Confirm the PMD did not change while page_table_lock was released */ |
bae473a4 KS |
1203 | spin_lock(fe->ptl); |
1204 | if (unlikely(!pmd_same(pmd, *fe->pmd))) { | |
b32967ff MG |
1205 | unlock_page(page); |
1206 | put_page(page); | |
a54a407f | 1207 | page_nid = -1; |
4daae3b4 | 1208 | goto out_unlock; |
b32967ff | 1209 | } |
ff9042b1 | 1210 | |
c3a489ca MG |
1211 | /* Bail if we fail to protect against THP splits for any reason */ |
1212 | if (unlikely(!anon_vma)) { | |
1213 | put_page(page); | |
1214 | page_nid = -1; | |
1215 | goto clear_pmdnuma; | |
1216 | } | |
1217 | ||
a54a407f MG |
1218 | /* |
1219 | * Migrate the THP to the requested node, returns with page unlocked | |
8a0516ed | 1220 | * and access rights restored. |
a54a407f | 1221 | */ |
bae473a4 KS |
1222 | spin_unlock(fe->ptl); |
1223 | migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma, | |
1224 | fe->pmd, pmd, fe->address, page, target_nid); | |
6688cc05 PZ |
1225 | if (migrated) { |
1226 | flags |= TNF_MIGRATED; | |
8191acbd | 1227 | page_nid = target_nid; |
074c2381 MG |
1228 | } else |
1229 | flags |= TNF_MIGRATE_FAIL; | |
b32967ff | 1230 | |
8191acbd | 1231 | goto out; |
b32967ff | 1232 | clear_pmdnuma: |
a54a407f | 1233 | BUG_ON(!PageLocked(page)); |
b191f9b1 | 1234 | was_writable = pmd_write(pmd); |
4d942466 | 1235 | pmd = pmd_modify(pmd, vma->vm_page_prot); |
b7b04004 | 1236 | pmd = pmd_mkyoung(pmd); |
b191f9b1 MG |
1237 | if (was_writable) |
1238 | pmd = pmd_mkwrite(pmd); | |
bae473a4 KS |
1239 | set_pmd_at(vma->vm_mm, haddr, fe->pmd, pmd); |
1240 | update_mmu_cache_pmd(vma, fe->address, fe->pmd); | |
a54a407f | 1241 | unlock_page(page); |
d10e63f2 | 1242 | out_unlock: |
bae473a4 | 1243 | spin_unlock(fe->ptl); |
b8916634 MG |
1244 | |
1245 | out: | |
1246 | if (anon_vma) | |
1247 | page_unlock_anon_vma_read(anon_vma); | |
1248 | ||
8191acbd | 1249 | if (page_nid != -1) |
bae473a4 | 1250 | task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, fe->flags); |
8191acbd | 1251 | |
d10e63f2 MG |
1252 | return 0; |
1253 | } | |
1254 | ||
319904ad HY |
1255 | /* |
1256 | * Return true if we do MADV_FREE successfully on entire pmd page. | |
1257 | * Otherwise, return false. | |
1258 | */ | |
1259 | bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, | |
b8d3c4c3 | 1260 | pmd_t *pmd, unsigned long addr, unsigned long next) |
b8d3c4c3 MK |
1261 | { |
1262 | spinlock_t *ptl; | |
1263 | pmd_t orig_pmd; | |
1264 | struct page *page; | |
1265 | struct mm_struct *mm = tlb->mm; | |
319904ad | 1266 | bool ret = false; |
b8d3c4c3 | 1267 | |
b6ec57f4 KS |
1268 | ptl = pmd_trans_huge_lock(pmd, vma); |
1269 | if (!ptl) | |
25eedabe | 1270 | goto out_unlocked; |
b8d3c4c3 MK |
1271 | |
1272 | orig_pmd = *pmd; | |
319904ad | 1273 | if (is_huge_zero_pmd(orig_pmd)) |
b8d3c4c3 | 1274 | goto out; |
b8d3c4c3 MK |
1275 | |
1276 | page = pmd_page(orig_pmd); | |
1277 | /* | |
1278 | * If other processes are mapping this page, we couldn't discard | |
1279 | * the page unless they all do MADV_FREE so let's skip the page. | |
1280 | */ | |
1281 | if (page_mapcount(page) != 1) | |
1282 | goto out; | |
1283 | ||
1284 | if (!trylock_page(page)) | |
1285 | goto out; | |
1286 | ||
1287 | /* | |
1288 | * If user want to discard part-pages of THP, split it so MADV_FREE | |
1289 | * will deactivate only them. | |
1290 | */ | |
1291 | if (next - addr != HPAGE_PMD_SIZE) { | |
1292 | get_page(page); | |
1293 | spin_unlock(ptl); | |
9818b8cd | 1294 | split_huge_page(page); |
b8d3c4c3 MK |
1295 | put_page(page); |
1296 | unlock_page(page); | |
b8d3c4c3 MK |
1297 | goto out_unlocked; |
1298 | } | |
1299 | ||
1300 | if (PageDirty(page)) | |
1301 | ClearPageDirty(page); | |
1302 | unlock_page(page); | |
1303 | ||
1304 | if (PageActive(page)) | |
1305 | deactivate_page(page); | |
1306 | ||
1307 | if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { | |
1308 | orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd, | |
1309 | tlb->fullmm); | |
1310 | orig_pmd = pmd_mkold(orig_pmd); | |
1311 | orig_pmd = pmd_mkclean(orig_pmd); | |
1312 | ||
1313 | set_pmd_at(mm, addr, pmd, orig_pmd); | |
1314 | tlb_remove_pmd_tlb_entry(tlb, pmd, addr); | |
1315 | } | |
319904ad | 1316 | ret = true; |
b8d3c4c3 MK |
1317 | out: |
1318 | spin_unlock(ptl); | |
1319 | out_unlocked: | |
1320 | return ret; | |
1321 | } | |
1322 | ||
71e3aac0 | 1323 | int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, |
f21760b1 | 1324 | pmd_t *pmd, unsigned long addr) |
71e3aac0 | 1325 | { |
da146769 | 1326 | pmd_t orig_pmd; |
bf929152 | 1327 | spinlock_t *ptl; |
71e3aac0 | 1328 | |
b6ec57f4 KS |
1329 | ptl = __pmd_trans_huge_lock(pmd, vma); |
1330 | if (!ptl) | |
da146769 KS |
1331 | return 0; |
1332 | /* | |
1333 | * For architectures like ppc64 we look at deposited pgtable | |
1334 | * when calling pmdp_huge_get_and_clear. So do the | |
1335 | * pgtable_trans_huge_withdraw after finishing pmdp related | |
1336 | * operations. | |
1337 | */ | |
1338 | orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd, | |
1339 | tlb->fullmm); | |
1340 | tlb_remove_pmd_tlb_entry(tlb, pmd, addr); | |
1341 | if (vma_is_dax(vma)) { | |
1342 | spin_unlock(ptl); | |
1343 | if (is_huge_zero_pmd(orig_pmd)) | |
aa88b68c | 1344 | tlb_remove_page(tlb, pmd_page(orig_pmd)); |
da146769 KS |
1345 | } else if (is_huge_zero_pmd(orig_pmd)) { |
1346 | pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd)); | |
1347 | atomic_long_dec(&tlb->mm->nr_ptes); | |
1348 | spin_unlock(ptl); | |
aa88b68c | 1349 | tlb_remove_page(tlb, pmd_page(orig_pmd)); |
da146769 KS |
1350 | } else { |
1351 | struct page *page = pmd_page(orig_pmd); | |
d281ee61 | 1352 | page_remove_rmap(page, true); |
da146769 | 1353 | VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); |
da146769 | 1354 | VM_BUG_ON_PAGE(!PageHead(page), page); |
b5072380 KS |
1355 | if (PageAnon(page)) { |
1356 | pgtable_t pgtable; | |
1357 | pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd); | |
1358 | pte_free(tlb->mm, pgtable); | |
1359 | atomic_long_dec(&tlb->mm->nr_ptes); | |
1360 | add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); | |
1361 | } else { | |
1362 | add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR); | |
1363 | } | |
da146769 | 1364 | spin_unlock(ptl); |
e77b0852 | 1365 | tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE); |
025c5b24 | 1366 | } |
da146769 | 1367 | return 1; |
71e3aac0 AA |
1368 | } |
1369 | ||
bf8616d5 | 1370 | bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, |
37a1c49a AA |
1371 | unsigned long new_addr, unsigned long old_end, |
1372 | pmd_t *old_pmd, pmd_t *new_pmd) | |
1373 | { | |
bf929152 | 1374 | spinlock_t *old_ptl, *new_ptl; |
37a1c49a | 1375 | pmd_t pmd; |
37a1c49a AA |
1376 | struct mm_struct *mm = vma->vm_mm; |
1377 | ||
1378 | if ((old_addr & ~HPAGE_PMD_MASK) || | |
1379 | (new_addr & ~HPAGE_PMD_MASK) || | |
bf8616d5 | 1380 | old_end - old_addr < HPAGE_PMD_SIZE) |
4b471e88 | 1381 | return false; |
37a1c49a AA |
1382 | |
1383 | /* | |
1384 | * The destination pmd shouldn't be established, free_pgtables() | |
1385 | * should have release it. | |
1386 | */ | |
1387 | if (WARN_ON(!pmd_none(*new_pmd))) { | |
1388 | VM_BUG_ON(pmd_trans_huge(*new_pmd)); | |
4b471e88 | 1389 | return false; |
37a1c49a AA |
1390 | } |
1391 | ||
bf929152 KS |
1392 | /* |
1393 | * We don't have to worry about the ordering of src and dst | |
1394 | * ptlocks because exclusive mmap_sem prevents deadlock. | |
1395 | */ | |
b6ec57f4 KS |
1396 | old_ptl = __pmd_trans_huge_lock(old_pmd, vma); |
1397 | if (old_ptl) { | |
bf929152 KS |
1398 | new_ptl = pmd_lockptr(mm, new_pmd); |
1399 | if (new_ptl != old_ptl) | |
1400 | spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); | |
8809aa2d | 1401 | pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); |
025c5b24 | 1402 | VM_BUG_ON(!pmd_none(*new_pmd)); |
3592806c | 1403 | |
69a8ec2d KS |
1404 | if (pmd_move_must_withdraw(new_ptl, old_ptl) && |
1405 | vma_is_anonymous(vma)) { | |
b3084f4d | 1406 | pgtable_t pgtable; |
3592806c KS |
1407 | pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); |
1408 | pgtable_trans_huge_deposit(mm, new_pmd, pgtable); | |
3592806c | 1409 | } |
b3084f4d AK |
1410 | set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd)); |
1411 | if (new_ptl != old_ptl) | |
1412 | spin_unlock(new_ptl); | |
bf929152 | 1413 | spin_unlock(old_ptl); |
4b471e88 | 1414 | return true; |
37a1c49a | 1415 | } |
4b471e88 | 1416 | return false; |
37a1c49a AA |
1417 | } |
1418 | ||
f123d74a MG |
1419 | /* |
1420 | * Returns | |
1421 | * - 0 if PMD could not be locked | |
1422 | * - 1 if PMD was locked but protections unchange and TLB flush unnecessary | |
1423 | * - HPAGE_PMD_NR is protections changed and TLB flush necessary | |
1424 | */ | |
cd7548ab | 1425 | int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, |
e944fd67 | 1426 | unsigned long addr, pgprot_t newprot, int prot_numa) |
cd7548ab JW |
1427 | { |
1428 | struct mm_struct *mm = vma->vm_mm; | |
bf929152 | 1429 | spinlock_t *ptl; |
cd7548ab JW |
1430 | int ret = 0; |
1431 | ||
b6ec57f4 KS |
1432 | ptl = __pmd_trans_huge_lock(pmd, vma); |
1433 | if (ptl) { | |
025c5b24 | 1434 | pmd_t entry; |
b191f9b1 | 1435 | bool preserve_write = prot_numa && pmd_write(*pmd); |
ba68bc01 | 1436 | ret = 1; |
e944fd67 MG |
1437 | |
1438 | /* | |
1439 | * Avoid trapping faults against the zero page. The read-only | |
1440 | * data is likely to be read-cached on the local CPU and | |
1441 | * local/remote hits to the zero page are not interesting. | |
1442 | */ | |
1443 | if (prot_numa && is_huge_zero_pmd(*pmd)) { | |
1444 | spin_unlock(ptl); | |
ba68bc01 | 1445 | return ret; |
e944fd67 MG |
1446 | } |
1447 | ||
10c1045f | 1448 | if (!prot_numa || !pmd_protnone(*pmd)) { |
8809aa2d | 1449 | entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd); |
10c1045f | 1450 | entry = pmd_modify(entry, newprot); |
b191f9b1 MG |
1451 | if (preserve_write) |
1452 | entry = pmd_mkwrite(entry); | |
10c1045f MG |
1453 | ret = HPAGE_PMD_NR; |
1454 | set_pmd_at(mm, addr, pmd, entry); | |
b237aded KS |
1455 | BUG_ON(vma_is_anonymous(vma) && !preserve_write && |
1456 | pmd_write(entry)); | |
10c1045f | 1457 | } |
bf929152 | 1458 | spin_unlock(ptl); |
025c5b24 NH |
1459 | } |
1460 | ||
1461 | return ret; | |
1462 | } | |
1463 | ||
1464 | /* | |
8f19b0c0 | 1465 | * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise. |
025c5b24 | 1466 | * |
8f19b0c0 HY |
1467 | * Note that if it returns page table lock pointer, this routine returns without |
1468 | * unlocking page table lock. So callers must unlock it. | |
025c5b24 | 1469 | */ |
b6ec57f4 | 1470 | spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) |
025c5b24 | 1471 | { |
b6ec57f4 KS |
1472 | spinlock_t *ptl; |
1473 | ptl = pmd_lock(vma->vm_mm, pmd); | |
5c7fb56e | 1474 | if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd))) |
b6ec57f4 KS |
1475 | return ptl; |
1476 | spin_unlock(ptl); | |
1477 | return NULL; | |
cd7548ab JW |
1478 | } |
1479 | ||
eef1b3ba KS |
1480 | static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, |
1481 | unsigned long haddr, pmd_t *pmd) | |
1482 | { | |
1483 | struct mm_struct *mm = vma->vm_mm; | |
1484 | pgtable_t pgtable; | |
1485 | pmd_t _pmd; | |
1486 | int i; | |
1487 | ||
1488 | /* leave pmd empty until pte is filled */ | |
1489 | pmdp_huge_clear_flush_notify(vma, haddr, pmd); | |
1490 | ||
1491 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); | |
1492 | pmd_populate(mm, &_pmd, pgtable); | |
1493 | ||
1494 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
1495 | pte_t *pte, entry; | |
1496 | entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); | |
1497 | entry = pte_mkspecial(entry); | |
1498 | pte = pte_offset_map(&_pmd, haddr); | |
1499 | VM_BUG_ON(!pte_none(*pte)); | |
1500 | set_pte_at(mm, haddr, pte, entry); | |
1501 | pte_unmap(pte); | |
1502 | } | |
1503 | smp_wmb(); /* make pte visible before pmd */ | |
1504 | pmd_populate(mm, pmd, pgtable); | |
1505 | put_huge_zero_page(); | |
1506 | } | |
1507 | ||
1508 | static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, | |
ba988280 | 1509 | unsigned long haddr, bool freeze) |
eef1b3ba KS |
1510 | { |
1511 | struct mm_struct *mm = vma->vm_mm; | |
1512 | struct page *page; | |
1513 | pgtable_t pgtable; | |
1514 | pmd_t _pmd; | |
b8d3c4c3 | 1515 | bool young, write, dirty; |
2ac015e2 | 1516 | unsigned long addr; |
eef1b3ba KS |
1517 | int i; |
1518 | ||
1519 | VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); | |
1520 | VM_BUG_ON_VMA(vma->vm_start > haddr, vma); | |
1521 | VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); | |
5c7fb56e | 1522 | VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)); |
eef1b3ba KS |
1523 | |
1524 | count_vm_event(THP_SPLIT_PMD); | |
1525 | ||
d21b9e57 KS |
1526 | if (!vma_is_anonymous(vma)) { |
1527 | _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); | |
eef1b3ba KS |
1528 | if (is_huge_zero_pmd(_pmd)) |
1529 | put_huge_zero_page(); | |
d21b9e57 KS |
1530 | if (vma_is_dax(vma)) |
1531 | return; | |
1532 | page = pmd_page(_pmd); | |
1533 | if (!PageReferenced(page) && pmd_young(_pmd)) | |
1534 | SetPageReferenced(page); | |
1535 | page_remove_rmap(page, true); | |
1536 | put_page(page); | |
1537 | add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR); | |
eef1b3ba KS |
1538 | return; |
1539 | } else if (is_huge_zero_pmd(*pmd)) { | |
1540 | return __split_huge_zero_page_pmd(vma, haddr, pmd); | |
1541 | } | |
1542 | ||
1543 | page = pmd_page(*pmd); | |
1544 | VM_BUG_ON_PAGE(!page_count(page), page); | |
fe896d18 | 1545 | page_ref_add(page, HPAGE_PMD_NR - 1); |
eef1b3ba KS |
1546 | write = pmd_write(*pmd); |
1547 | young = pmd_young(*pmd); | |
b8d3c4c3 | 1548 | dirty = pmd_dirty(*pmd); |
eef1b3ba | 1549 | |
c777e2a8 | 1550 | pmdp_huge_split_prepare(vma, haddr, pmd); |
eef1b3ba KS |
1551 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); |
1552 | pmd_populate(mm, &_pmd, pgtable); | |
1553 | ||
2ac015e2 | 1554 | for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { |
eef1b3ba KS |
1555 | pte_t entry, *pte; |
1556 | /* | |
1557 | * Note that NUMA hinting access restrictions are not | |
1558 | * transferred to avoid any possibility of altering | |
1559 | * permissions across VMAs. | |
1560 | */ | |
ba988280 KS |
1561 | if (freeze) { |
1562 | swp_entry_t swp_entry; | |
1563 | swp_entry = make_migration_entry(page + i, write); | |
1564 | entry = swp_entry_to_pte(swp_entry); | |
1565 | } else { | |
1566 | entry = mk_pte(page + i, vma->vm_page_prot); | |
b8d3c4c3 | 1567 | entry = maybe_mkwrite(entry, vma); |
ba988280 KS |
1568 | if (!write) |
1569 | entry = pte_wrprotect(entry); | |
1570 | if (!young) | |
1571 | entry = pte_mkold(entry); | |
1572 | } | |
b8d3c4c3 MK |
1573 | if (dirty) |
1574 | SetPageDirty(page + i); | |
2ac015e2 | 1575 | pte = pte_offset_map(&_pmd, addr); |
eef1b3ba | 1576 | BUG_ON(!pte_none(*pte)); |
2ac015e2 | 1577 | set_pte_at(mm, addr, pte, entry); |
eef1b3ba KS |
1578 | atomic_inc(&page[i]._mapcount); |
1579 | pte_unmap(pte); | |
1580 | } | |
1581 | ||
1582 | /* | |
1583 | * Set PG_double_map before dropping compound_mapcount to avoid | |
1584 | * false-negative page_mapped(). | |
1585 | */ | |
1586 | if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) { | |
1587 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
1588 | atomic_inc(&page[i]._mapcount); | |
1589 | } | |
1590 | ||
1591 | if (atomic_add_negative(-1, compound_mapcount_ptr(page))) { | |
1592 | /* Last compound_mapcount is gone. */ | |
11fb9989 | 1593 | __dec_node_page_state(page, NR_ANON_THPS); |
eef1b3ba KS |
1594 | if (TestClearPageDoubleMap(page)) { |
1595 | /* No need in mapcount reference anymore */ | |
1596 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
1597 | atomic_dec(&page[i]._mapcount); | |
1598 | } | |
1599 | } | |
1600 | ||
1601 | smp_wmb(); /* make pte visible before pmd */ | |
e9b61f19 KS |
1602 | /* |
1603 | * Up to this point the pmd is present and huge and userland has the | |
1604 | * whole access to the hugepage during the split (which happens in | |
1605 | * place). If we overwrite the pmd with the not-huge version pointing | |
1606 | * to the pte here (which of course we could if all CPUs were bug | |
1607 | * free), userland could trigger a small page size TLB miss on the | |
1608 | * small sized TLB while the hugepage TLB entry is still established in | |
1609 | * the huge TLB. Some CPU doesn't like that. | |
1610 | * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum | |
1611 | * 383 on page 93. Intel should be safe but is also warns that it's | |
1612 | * only safe if the permission and cache attributes of the two entries | |
1613 | * loaded in the two TLB is identical (which should be the case here). | |
1614 | * But it is generally safer to never allow small and huge TLB entries | |
1615 | * for the same virtual address to be loaded simultaneously. So instead | |
1616 | * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the | |
1617 | * current pmd notpresent (atomically because here the pmd_trans_huge | |
1618 | * and pmd_trans_splitting must remain set at all times on the pmd | |
1619 | * until the split is complete for this pmd), then we flush the SMP TLB | |
1620 | * and finally we write the non-huge version of the pmd entry with | |
1621 | * pmd_populate. | |
1622 | */ | |
1623 | pmdp_invalidate(vma, haddr, pmd); | |
eef1b3ba | 1624 | pmd_populate(mm, pmd, pgtable); |
e9b61f19 KS |
1625 | |
1626 | if (freeze) { | |
2ac015e2 | 1627 | for (i = 0; i < HPAGE_PMD_NR; i++) { |
e9b61f19 KS |
1628 | page_remove_rmap(page + i, false); |
1629 | put_page(page + i); | |
1630 | } | |
1631 | } | |
eef1b3ba KS |
1632 | } |
1633 | ||
1634 | void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, | |
33f4751e | 1635 | unsigned long address, bool freeze, struct page *page) |
eef1b3ba KS |
1636 | { |
1637 | spinlock_t *ptl; | |
1638 | struct mm_struct *mm = vma->vm_mm; | |
1639 | unsigned long haddr = address & HPAGE_PMD_MASK; | |
1640 | ||
1641 | mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE); | |
1642 | ptl = pmd_lock(mm, pmd); | |
33f4751e NH |
1643 | |
1644 | /* | |
1645 | * If caller asks to setup a migration entries, we need a page to check | |
1646 | * pmd against. Otherwise we can end up replacing wrong page. | |
1647 | */ | |
1648 | VM_BUG_ON(freeze && !page); | |
1649 | if (page && page != pmd_page(*pmd)) | |
1650 | goto out; | |
1651 | ||
5c7fb56e | 1652 | if (pmd_trans_huge(*pmd)) { |
33f4751e | 1653 | page = pmd_page(*pmd); |
5c7fb56e | 1654 | if (PageMlocked(page)) |
5f737714 | 1655 | clear_page_mlock(page); |
5c7fb56e | 1656 | } else if (!pmd_devmap(*pmd)) |
e90309c9 | 1657 | goto out; |
fec89c10 | 1658 | __split_huge_pmd_locked(vma, pmd, haddr, freeze); |
e90309c9 | 1659 | out: |
eef1b3ba KS |
1660 | spin_unlock(ptl); |
1661 | mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE); | |
1662 | } | |
1663 | ||
fec89c10 KS |
1664 | void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address, |
1665 | bool freeze, struct page *page) | |
94fcc585 | 1666 | { |
f72e7dcd HD |
1667 | pgd_t *pgd; |
1668 | pud_t *pud; | |
94fcc585 AA |
1669 | pmd_t *pmd; |
1670 | ||
78ddc534 | 1671 | pgd = pgd_offset(vma->vm_mm, address); |
f72e7dcd HD |
1672 | if (!pgd_present(*pgd)) |
1673 | return; | |
1674 | ||
1675 | pud = pud_offset(pgd, address); | |
1676 | if (!pud_present(*pud)) | |
1677 | return; | |
1678 | ||
1679 | pmd = pmd_offset(pud, address); | |
fec89c10 | 1680 | |
33f4751e | 1681 | __split_huge_pmd(vma, pmd, address, freeze, page); |
94fcc585 AA |
1682 | } |
1683 | ||
e1b9996b | 1684 | void vma_adjust_trans_huge(struct vm_area_struct *vma, |
94fcc585 AA |
1685 | unsigned long start, |
1686 | unsigned long end, | |
1687 | long adjust_next) | |
1688 | { | |
1689 | /* | |
1690 | * If the new start address isn't hpage aligned and it could | |
1691 | * previously contain an hugepage: check if we need to split | |
1692 | * an huge pmd. | |
1693 | */ | |
1694 | if (start & ~HPAGE_PMD_MASK && | |
1695 | (start & HPAGE_PMD_MASK) >= vma->vm_start && | |
1696 | (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
fec89c10 | 1697 | split_huge_pmd_address(vma, start, false, NULL); |
94fcc585 AA |
1698 | |
1699 | /* | |
1700 | * If the new end address isn't hpage aligned and it could | |
1701 | * previously contain an hugepage: check if we need to split | |
1702 | * an huge pmd. | |
1703 | */ | |
1704 | if (end & ~HPAGE_PMD_MASK && | |
1705 | (end & HPAGE_PMD_MASK) >= vma->vm_start && | |
1706 | (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
fec89c10 | 1707 | split_huge_pmd_address(vma, end, false, NULL); |
94fcc585 AA |
1708 | |
1709 | /* | |
1710 | * If we're also updating the vma->vm_next->vm_start, if the new | |
1711 | * vm_next->vm_start isn't page aligned and it could previously | |
1712 | * contain an hugepage: check if we need to split an huge pmd. | |
1713 | */ | |
1714 | if (adjust_next > 0) { | |
1715 | struct vm_area_struct *next = vma->vm_next; | |
1716 | unsigned long nstart = next->vm_start; | |
1717 | nstart += adjust_next << PAGE_SHIFT; | |
1718 | if (nstart & ~HPAGE_PMD_MASK && | |
1719 | (nstart & HPAGE_PMD_MASK) >= next->vm_start && | |
1720 | (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) | |
fec89c10 | 1721 | split_huge_pmd_address(next, nstart, false, NULL); |
94fcc585 AA |
1722 | } |
1723 | } | |
e9b61f19 | 1724 | |
fec89c10 | 1725 | static void freeze_page(struct page *page) |
e9b61f19 | 1726 | { |
baa355fd KS |
1727 | enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS | |
1728 | TTU_RMAP_LOCKED; | |
fec89c10 | 1729 | int i, ret; |
e9b61f19 KS |
1730 | |
1731 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
1732 | ||
baa355fd KS |
1733 | if (PageAnon(page)) |
1734 | ttu_flags |= TTU_MIGRATION; | |
1735 | ||
fec89c10 KS |
1736 | /* We only need TTU_SPLIT_HUGE_PMD once */ |
1737 | ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD); | |
1738 | for (i = 1; !ret && i < HPAGE_PMD_NR; i++) { | |
1739 | /* Cut short if the page is unmapped */ | |
1740 | if (page_count(page) == 1) | |
1741 | return; | |
e9b61f19 | 1742 | |
fec89c10 | 1743 | ret = try_to_unmap(page + i, ttu_flags); |
e9b61f19 | 1744 | } |
baa355fd | 1745 | VM_BUG_ON_PAGE(ret, page + i - 1); |
e9b61f19 KS |
1746 | } |
1747 | ||
fec89c10 | 1748 | static void unfreeze_page(struct page *page) |
e9b61f19 | 1749 | { |
fec89c10 | 1750 | int i; |
e9b61f19 | 1751 | |
fec89c10 KS |
1752 | for (i = 0; i < HPAGE_PMD_NR; i++) |
1753 | remove_migration_ptes(page + i, page + i, true); | |
e9b61f19 KS |
1754 | } |
1755 | ||
8df651c7 | 1756 | static void __split_huge_page_tail(struct page *head, int tail, |
e9b61f19 KS |
1757 | struct lruvec *lruvec, struct list_head *list) |
1758 | { | |
e9b61f19 KS |
1759 | struct page *page_tail = head + tail; |
1760 | ||
8df651c7 | 1761 | VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail); |
fe896d18 | 1762 | VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail); |
e9b61f19 KS |
1763 | |
1764 | /* | |
0139aa7b | 1765 | * tail_page->_refcount is zero and not changing from under us. But |
e9b61f19 | 1766 | * get_page_unless_zero() may be running from under us on the |
baa355fd KS |
1767 | * tail_page. If we used atomic_set() below instead of atomic_inc() or |
1768 | * atomic_add(), we would then run atomic_set() concurrently with | |
e9b61f19 KS |
1769 | * get_page_unless_zero(), and atomic_set() is implemented in C not |
1770 | * using locked ops. spin_unlock on x86 sometime uses locked ops | |
1771 | * because of PPro errata 66, 92, so unless somebody can guarantee | |
1772 | * atomic_set() here would be safe on all archs (and not only on x86), | |
baa355fd | 1773 | * it's safer to use atomic_inc()/atomic_add(). |
e9b61f19 | 1774 | */ |
baa355fd KS |
1775 | if (PageAnon(head)) { |
1776 | page_ref_inc(page_tail); | |
1777 | } else { | |
1778 | /* Additional pin to radix tree */ | |
1779 | page_ref_add(page_tail, 2); | |
1780 | } | |
e9b61f19 KS |
1781 | |
1782 | page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; | |
1783 | page_tail->flags |= (head->flags & | |
1784 | ((1L << PG_referenced) | | |
1785 | (1L << PG_swapbacked) | | |
1786 | (1L << PG_mlocked) | | |
1787 | (1L << PG_uptodate) | | |
1788 | (1L << PG_active) | | |
1789 | (1L << PG_locked) | | |
b8d3c4c3 MK |
1790 | (1L << PG_unevictable) | |
1791 | (1L << PG_dirty))); | |
e9b61f19 KS |
1792 | |
1793 | /* | |
1794 | * After clearing PageTail the gup refcount can be released. | |
1795 | * Page flags also must be visible before we make the page non-compound. | |
1796 | */ | |
1797 | smp_wmb(); | |
1798 | ||
1799 | clear_compound_head(page_tail); | |
1800 | ||
1801 | if (page_is_young(head)) | |
1802 | set_page_young(page_tail); | |
1803 | if (page_is_idle(head)) | |
1804 | set_page_idle(page_tail); | |
1805 | ||
1806 | /* ->mapping in first tail page is compound_mapcount */ | |
9a982250 | 1807 | VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING, |
e9b61f19 KS |
1808 | page_tail); |
1809 | page_tail->mapping = head->mapping; | |
1810 | ||
1811 | page_tail->index = head->index + tail; | |
1812 | page_cpupid_xchg_last(page_tail, page_cpupid_last(head)); | |
1813 | lru_add_page_tail(head, page_tail, lruvec, list); | |
e9b61f19 KS |
1814 | } |
1815 | ||
baa355fd KS |
1816 | static void __split_huge_page(struct page *page, struct list_head *list, |
1817 | unsigned long flags) | |
e9b61f19 KS |
1818 | { |
1819 | struct page *head = compound_head(page); | |
1820 | struct zone *zone = page_zone(head); | |
1821 | struct lruvec *lruvec; | |
baa355fd | 1822 | pgoff_t end = -1; |
8df651c7 | 1823 | int i; |
e9b61f19 | 1824 | |
599d0c95 | 1825 | lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat); |
e9b61f19 KS |
1826 | |
1827 | /* complete memcg works before add pages to LRU */ | |
1828 | mem_cgroup_split_huge_fixup(head); | |
1829 | ||
baa355fd KS |
1830 | if (!PageAnon(page)) |
1831 | end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE); | |
1832 | ||
1833 | for (i = HPAGE_PMD_NR - 1; i >= 1; i--) { | |
8df651c7 | 1834 | __split_huge_page_tail(head, i, lruvec, list); |
baa355fd KS |
1835 | /* Some pages can be beyond i_size: drop them from page cache */ |
1836 | if (head[i].index >= end) { | |
1837 | __ClearPageDirty(head + i); | |
1838 | __delete_from_page_cache(head + i, NULL); | |
800d8c63 KS |
1839 | if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head)) |
1840 | shmem_uncharge(head->mapping->host, 1); | |
baa355fd KS |
1841 | put_page(head + i); |
1842 | } | |
1843 | } | |
e9b61f19 KS |
1844 | |
1845 | ClearPageCompound(head); | |
baa355fd KS |
1846 | /* See comment in __split_huge_page_tail() */ |
1847 | if (PageAnon(head)) { | |
1848 | page_ref_inc(head); | |
1849 | } else { | |
1850 | /* Additional pin to radix tree */ | |
1851 | page_ref_add(head, 2); | |
1852 | spin_unlock(&head->mapping->tree_lock); | |
1853 | } | |
1854 | ||
a52633d8 | 1855 | spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags); |
e9b61f19 | 1856 | |
fec89c10 | 1857 | unfreeze_page(head); |
e9b61f19 KS |
1858 | |
1859 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
1860 | struct page *subpage = head + i; | |
1861 | if (subpage == page) | |
1862 | continue; | |
1863 | unlock_page(subpage); | |
1864 | ||
1865 | /* | |
1866 | * Subpages may be freed if there wasn't any mapping | |
1867 | * like if add_to_swap() is running on a lru page that | |
1868 | * had its mapping zapped. And freeing these pages | |
1869 | * requires taking the lru_lock so we do the put_page | |
1870 | * of the tail pages after the split is complete. | |
1871 | */ | |
1872 | put_page(subpage); | |
1873 | } | |
1874 | } | |
1875 | ||
b20ce5e0 KS |
1876 | int total_mapcount(struct page *page) |
1877 | { | |
dd78fedd | 1878 | int i, compound, ret; |
b20ce5e0 KS |
1879 | |
1880 | VM_BUG_ON_PAGE(PageTail(page), page); | |
1881 | ||
1882 | if (likely(!PageCompound(page))) | |
1883 | return atomic_read(&page->_mapcount) + 1; | |
1884 | ||
dd78fedd | 1885 | compound = compound_mapcount(page); |
b20ce5e0 | 1886 | if (PageHuge(page)) |
dd78fedd KS |
1887 | return compound; |
1888 | ret = compound; | |
b20ce5e0 KS |
1889 | for (i = 0; i < HPAGE_PMD_NR; i++) |
1890 | ret += atomic_read(&page[i]._mapcount) + 1; | |
dd78fedd KS |
1891 | /* File pages has compound_mapcount included in _mapcount */ |
1892 | if (!PageAnon(page)) | |
1893 | return ret - compound * HPAGE_PMD_NR; | |
b20ce5e0 KS |
1894 | if (PageDoubleMap(page)) |
1895 | ret -= HPAGE_PMD_NR; | |
1896 | return ret; | |
1897 | } | |
1898 | ||
6d0a07ed AA |
1899 | /* |
1900 | * This calculates accurately how many mappings a transparent hugepage | |
1901 | * has (unlike page_mapcount() which isn't fully accurate). This full | |
1902 | * accuracy is primarily needed to know if copy-on-write faults can | |
1903 | * reuse the page and change the mapping to read-write instead of | |
1904 | * copying them. At the same time this returns the total_mapcount too. | |
1905 | * | |
1906 | * The function returns the highest mapcount any one of the subpages | |
1907 | * has. If the return value is one, even if different processes are | |
1908 | * mapping different subpages of the transparent hugepage, they can | |
1909 | * all reuse it, because each process is reusing a different subpage. | |
1910 | * | |
1911 | * The total_mapcount is instead counting all virtual mappings of the | |
1912 | * subpages. If the total_mapcount is equal to "one", it tells the | |
1913 | * caller all mappings belong to the same "mm" and in turn the | |
1914 | * anon_vma of the transparent hugepage can become the vma->anon_vma | |
1915 | * local one as no other process may be mapping any of the subpages. | |
1916 | * | |
1917 | * It would be more accurate to replace page_mapcount() with | |
1918 | * page_trans_huge_mapcount(), however we only use | |
1919 | * page_trans_huge_mapcount() in the copy-on-write faults where we | |
1920 | * need full accuracy to avoid breaking page pinning, because | |
1921 | * page_trans_huge_mapcount() is slower than page_mapcount(). | |
1922 | */ | |
1923 | int page_trans_huge_mapcount(struct page *page, int *total_mapcount) | |
1924 | { | |
1925 | int i, ret, _total_mapcount, mapcount; | |
1926 | ||
1927 | /* hugetlbfs shouldn't call it */ | |
1928 | VM_BUG_ON_PAGE(PageHuge(page), page); | |
1929 | ||
1930 | if (likely(!PageTransCompound(page))) { | |
1931 | mapcount = atomic_read(&page->_mapcount) + 1; | |
1932 | if (total_mapcount) | |
1933 | *total_mapcount = mapcount; | |
1934 | return mapcount; | |
1935 | } | |
1936 | ||
1937 | page = compound_head(page); | |
1938 | ||
1939 | _total_mapcount = ret = 0; | |
1940 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
1941 | mapcount = atomic_read(&page[i]._mapcount) + 1; | |
1942 | ret = max(ret, mapcount); | |
1943 | _total_mapcount += mapcount; | |
1944 | } | |
1945 | if (PageDoubleMap(page)) { | |
1946 | ret -= 1; | |
1947 | _total_mapcount -= HPAGE_PMD_NR; | |
1948 | } | |
1949 | mapcount = compound_mapcount(page); | |
1950 | ret += mapcount; | |
1951 | _total_mapcount += mapcount; | |
1952 | if (total_mapcount) | |
1953 | *total_mapcount = _total_mapcount; | |
1954 | return ret; | |
1955 | } | |
1956 | ||
e9b61f19 KS |
1957 | /* |
1958 | * This function splits huge page into normal pages. @page can point to any | |
1959 | * subpage of huge page to split. Split doesn't change the position of @page. | |
1960 | * | |
1961 | * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. | |
1962 | * The huge page must be locked. | |
1963 | * | |
1964 | * If @list is null, tail pages will be added to LRU list, otherwise, to @list. | |
1965 | * | |
1966 | * Both head page and tail pages will inherit mapping, flags, and so on from | |
1967 | * the hugepage. | |
1968 | * | |
1969 | * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if | |
1970 | * they are not mapped. | |
1971 | * | |
1972 | * Returns 0 if the hugepage is split successfully. | |
1973 | * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under | |
1974 | * us. | |
1975 | */ | |
1976 | int split_huge_page_to_list(struct page *page, struct list_head *list) | |
1977 | { | |
1978 | struct page *head = compound_head(page); | |
a3d0a918 | 1979 | struct pglist_data *pgdata = NODE_DATA(page_to_nid(head)); |
baa355fd KS |
1980 | struct anon_vma *anon_vma = NULL; |
1981 | struct address_space *mapping = NULL; | |
1982 | int count, mapcount, extra_pins, ret; | |
d9654322 | 1983 | bool mlocked; |
0b9b6fff | 1984 | unsigned long flags; |
e9b61f19 KS |
1985 | |
1986 | VM_BUG_ON_PAGE(is_huge_zero_page(page), page); | |
e9b61f19 KS |
1987 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
1988 | VM_BUG_ON_PAGE(!PageSwapBacked(page), page); | |
1989 | VM_BUG_ON_PAGE(!PageCompound(page), page); | |
1990 | ||
baa355fd KS |
1991 | if (PageAnon(head)) { |
1992 | /* | |
1993 | * The caller does not necessarily hold an mmap_sem that would | |
1994 | * prevent the anon_vma disappearing so we first we take a | |
1995 | * reference to it and then lock the anon_vma for write. This | |
1996 | * is similar to page_lock_anon_vma_read except the write lock | |
1997 | * is taken to serialise against parallel split or collapse | |
1998 | * operations. | |
1999 | */ | |
2000 | anon_vma = page_get_anon_vma(head); | |
2001 | if (!anon_vma) { | |
2002 | ret = -EBUSY; | |
2003 | goto out; | |
2004 | } | |
2005 | extra_pins = 0; | |
2006 | mapping = NULL; | |
2007 | anon_vma_lock_write(anon_vma); | |
2008 | } else { | |
2009 | mapping = head->mapping; | |
2010 | ||
2011 | /* Truncated ? */ | |
2012 | if (!mapping) { | |
2013 | ret = -EBUSY; | |
2014 | goto out; | |
2015 | } | |
2016 | ||
2017 | /* Addidional pins from radix tree */ | |
2018 | extra_pins = HPAGE_PMD_NR; | |
2019 | anon_vma = NULL; | |
2020 | i_mmap_lock_read(mapping); | |
e9b61f19 | 2021 | } |
e9b61f19 KS |
2022 | |
2023 | /* | |
2024 | * Racy check if we can split the page, before freeze_page() will | |
2025 | * split PMDs | |
2026 | */ | |
baa355fd | 2027 | if (total_mapcount(head) != page_count(head) - extra_pins - 1) { |
e9b61f19 KS |
2028 | ret = -EBUSY; |
2029 | goto out_unlock; | |
2030 | } | |
2031 | ||
d9654322 | 2032 | mlocked = PageMlocked(page); |
fec89c10 | 2033 | freeze_page(head); |
e9b61f19 KS |
2034 | VM_BUG_ON_PAGE(compound_mapcount(head), head); |
2035 | ||
d9654322 KS |
2036 | /* Make sure the page is not on per-CPU pagevec as it takes pin */ |
2037 | if (mlocked) | |
2038 | lru_add_drain(); | |
2039 | ||
baa355fd | 2040 | /* prevent PageLRU to go away from under us, and freeze lru stats */ |
a52633d8 | 2041 | spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags); |
baa355fd KS |
2042 | |
2043 | if (mapping) { | |
2044 | void **pslot; | |
2045 | ||
2046 | spin_lock(&mapping->tree_lock); | |
2047 | pslot = radix_tree_lookup_slot(&mapping->page_tree, | |
2048 | page_index(head)); | |
2049 | /* | |
2050 | * Check if the head page is present in radix tree. | |
2051 | * We assume all tail are present too, if head is there. | |
2052 | */ | |
2053 | if (radix_tree_deref_slot_protected(pslot, | |
2054 | &mapping->tree_lock) != head) | |
2055 | goto fail; | |
2056 | } | |
2057 | ||
0139aa7b | 2058 | /* Prevent deferred_split_scan() touching ->_refcount */ |
baa355fd | 2059 | spin_lock(&pgdata->split_queue_lock); |
e9b61f19 KS |
2060 | count = page_count(head); |
2061 | mapcount = total_mapcount(head); | |
baa355fd | 2062 | if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) { |
9a982250 | 2063 | if (!list_empty(page_deferred_list(head))) { |
a3d0a918 | 2064 | pgdata->split_queue_len--; |
9a982250 KS |
2065 | list_del(page_deferred_list(head)); |
2066 | } | |
65c45377 | 2067 | if (mapping) |
11fb9989 | 2068 | __dec_node_page_state(page, NR_SHMEM_THPS); |
baa355fd KS |
2069 | spin_unlock(&pgdata->split_queue_lock); |
2070 | __split_huge_page(page, list, flags); | |
e9b61f19 | 2071 | ret = 0; |
e9b61f19 | 2072 | } else { |
baa355fd KS |
2073 | if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) { |
2074 | pr_alert("total_mapcount: %u, page_count(): %u\n", | |
2075 | mapcount, count); | |
2076 | if (PageTail(page)) | |
2077 | dump_page(head, NULL); | |
2078 | dump_page(page, "total_mapcount(head) > 0"); | |
2079 | BUG(); | |
2080 | } | |
2081 | spin_unlock(&pgdata->split_queue_lock); | |
2082 | fail: if (mapping) | |
2083 | spin_unlock(&mapping->tree_lock); | |
a52633d8 | 2084 | spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags); |
fec89c10 | 2085 | unfreeze_page(head); |
e9b61f19 KS |
2086 | ret = -EBUSY; |
2087 | } | |
2088 | ||
2089 | out_unlock: | |
baa355fd KS |
2090 | if (anon_vma) { |
2091 | anon_vma_unlock_write(anon_vma); | |
2092 | put_anon_vma(anon_vma); | |
2093 | } | |
2094 | if (mapping) | |
2095 | i_mmap_unlock_read(mapping); | |
e9b61f19 KS |
2096 | out: |
2097 | count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); | |
2098 | return ret; | |
2099 | } | |
9a982250 KS |
2100 | |
2101 | void free_transhuge_page(struct page *page) | |
2102 | { | |
a3d0a918 | 2103 | struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); |
9a982250 KS |
2104 | unsigned long flags; |
2105 | ||
a3d0a918 | 2106 | spin_lock_irqsave(&pgdata->split_queue_lock, flags); |
9a982250 | 2107 | if (!list_empty(page_deferred_list(page))) { |
a3d0a918 | 2108 | pgdata->split_queue_len--; |
9a982250 KS |
2109 | list_del(page_deferred_list(page)); |
2110 | } | |
a3d0a918 | 2111 | spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); |
9a982250 KS |
2112 | free_compound_page(page); |
2113 | } | |
2114 | ||
2115 | void deferred_split_huge_page(struct page *page) | |
2116 | { | |
a3d0a918 | 2117 | struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); |
9a982250 KS |
2118 | unsigned long flags; |
2119 | ||
2120 | VM_BUG_ON_PAGE(!PageTransHuge(page), page); | |
2121 | ||
a3d0a918 | 2122 | spin_lock_irqsave(&pgdata->split_queue_lock, flags); |
9a982250 | 2123 | if (list_empty(page_deferred_list(page))) { |
f9719a03 | 2124 | count_vm_event(THP_DEFERRED_SPLIT_PAGE); |
a3d0a918 KS |
2125 | list_add_tail(page_deferred_list(page), &pgdata->split_queue); |
2126 | pgdata->split_queue_len++; | |
9a982250 | 2127 | } |
a3d0a918 | 2128 | spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); |
9a982250 KS |
2129 | } |
2130 | ||
2131 | static unsigned long deferred_split_count(struct shrinker *shrink, | |
2132 | struct shrink_control *sc) | |
2133 | { | |
a3d0a918 | 2134 | struct pglist_data *pgdata = NODE_DATA(sc->nid); |
cb8d68ec | 2135 | return ACCESS_ONCE(pgdata->split_queue_len); |
9a982250 KS |
2136 | } |
2137 | ||
2138 | static unsigned long deferred_split_scan(struct shrinker *shrink, | |
2139 | struct shrink_control *sc) | |
2140 | { | |
a3d0a918 | 2141 | struct pglist_data *pgdata = NODE_DATA(sc->nid); |
9a982250 KS |
2142 | unsigned long flags; |
2143 | LIST_HEAD(list), *pos, *next; | |
2144 | struct page *page; | |
2145 | int split = 0; | |
2146 | ||
a3d0a918 | 2147 | spin_lock_irqsave(&pgdata->split_queue_lock, flags); |
9a982250 | 2148 | /* Take pin on all head pages to avoid freeing them under us */ |
ae026204 | 2149 | list_for_each_safe(pos, next, &pgdata->split_queue) { |
9a982250 KS |
2150 | page = list_entry((void *)pos, struct page, mapping); |
2151 | page = compound_head(page); | |
e3ae1953 KS |
2152 | if (get_page_unless_zero(page)) { |
2153 | list_move(page_deferred_list(page), &list); | |
2154 | } else { | |
2155 | /* We lost race with put_compound_page() */ | |
9a982250 | 2156 | list_del_init(page_deferred_list(page)); |
a3d0a918 | 2157 | pgdata->split_queue_len--; |
9a982250 | 2158 | } |
e3ae1953 KS |
2159 | if (!--sc->nr_to_scan) |
2160 | break; | |
9a982250 | 2161 | } |
a3d0a918 | 2162 | spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); |
9a982250 KS |
2163 | |
2164 | list_for_each_safe(pos, next, &list) { | |
2165 | page = list_entry((void *)pos, struct page, mapping); | |
2166 | lock_page(page); | |
2167 | /* split_huge_page() removes page from list on success */ | |
2168 | if (!split_huge_page(page)) | |
2169 | split++; | |
2170 | unlock_page(page); | |
2171 | put_page(page); | |
2172 | } | |
2173 | ||
a3d0a918 KS |
2174 | spin_lock_irqsave(&pgdata->split_queue_lock, flags); |
2175 | list_splice_tail(&list, &pgdata->split_queue); | |
2176 | spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); | |
9a982250 | 2177 | |
cb8d68ec KS |
2178 | /* |
2179 | * Stop shrinker if we didn't split any page, but the queue is empty. | |
2180 | * This can happen if pages were freed under us. | |
2181 | */ | |
2182 | if (!split && list_empty(&pgdata->split_queue)) | |
2183 | return SHRINK_STOP; | |
2184 | return split; | |
9a982250 KS |
2185 | } |
2186 | ||
2187 | static struct shrinker deferred_split_shrinker = { | |
2188 | .count_objects = deferred_split_count, | |
2189 | .scan_objects = deferred_split_scan, | |
2190 | .seeks = DEFAULT_SEEKS, | |
a3d0a918 | 2191 | .flags = SHRINKER_NUMA_AWARE, |
9a982250 | 2192 | }; |
49071d43 KS |
2193 | |
2194 | #ifdef CONFIG_DEBUG_FS | |
2195 | static int split_huge_pages_set(void *data, u64 val) | |
2196 | { | |
2197 | struct zone *zone; | |
2198 | struct page *page; | |
2199 | unsigned long pfn, max_zone_pfn; | |
2200 | unsigned long total = 0, split = 0; | |
2201 | ||
2202 | if (val != 1) | |
2203 | return -EINVAL; | |
2204 | ||
2205 | for_each_populated_zone(zone) { | |
2206 | max_zone_pfn = zone_end_pfn(zone); | |
2207 | for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { | |
2208 | if (!pfn_valid(pfn)) | |
2209 | continue; | |
2210 | ||
2211 | page = pfn_to_page(pfn); | |
2212 | if (!get_page_unless_zero(page)) | |
2213 | continue; | |
2214 | ||
2215 | if (zone != page_zone(page)) | |
2216 | goto next; | |
2217 | ||
baa355fd | 2218 | if (!PageHead(page) || PageHuge(page) || !PageLRU(page)) |
49071d43 KS |
2219 | goto next; |
2220 | ||
2221 | total++; | |
2222 | lock_page(page); | |
2223 | if (!split_huge_page(page)) | |
2224 | split++; | |
2225 | unlock_page(page); | |
2226 | next: | |
2227 | put_page(page); | |
2228 | } | |
2229 | } | |
2230 | ||
145bdaa1 | 2231 | pr_info("%lu of %lu THP split\n", split, total); |
49071d43 KS |
2232 | |
2233 | return 0; | |
2234 | } | |
2235 | DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set, | |
2236 | "%llu\n"); | |
2237 | ||
2238 | static int __init split_huge_pages_debugfs(void) | |
2239 | { | |
2240 | void *ret; | |
2241 | ||
145bdaa1 | 2242 | ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL, |
49071d43 KS |
2243 | &split_huge_pages_fops); |
2244 | if (!ret) | |
2245 | pr_warn("Failed to create split_huge_pages in debugfs"); | |
2246 | return 0; | |
2247 | } | |
2248 | late_initcall(split_huge_pages_debugfs); | |
2249 | #endif |