]>
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> | |
f7ccbae4 | 12 | #include <linux/sched/coredump.h> |
6a3827d7 | 13 | #include <linux/sched/numa_balancing.h> |
71e3aac0 AA |
14 | #include <linux/highmem.h> |
15 | #include <linux/hugetlb.h> | |
16 | #include <linux/mmu_notifier.h> | |
17 | #include <linux/rmap.h> | |
18 | #include <linux/swap.h> | |
97ae1749 | 19 | #include <linux/shrinker.h> |
ba76149f | 20 | #include <linux/mm_inline.h> |
e9b61f19 | 21 | #include <linux/swapops.h> |
4897c765 | 22 | #include <linux/dax.h> |
ba76149f | 23 | #include <linux/khugepaged.h> |
878aee7d | 24 | #include <linux/freezer.h> |
f25748e3 | 25 | #include <linux/pfn_t.h> |
a664b2d8 | 26 | #include <linux/mman.h> |
3565fce3 | 27 | #include <linux/memremap.h> |
325adeb5 | 28 | #include <linux/pagemap.h> |
49071d43 | 29 | #include <linux/debugfs.h> |
4daae3b4 | 30 | #include <linux/migrate.h> |
43b5fbbd | 31 | #include <linux/hashtable.h> |
6b251fc9 | 32 | #include <linux/userfaultfd_k.h> |
33c3fc71 | 33 | #include <linux/page_idle.h> |
baa355fd | 34 | #include <linux/shmem_fs.h> |
6b31d595 | 35 | #include <linux/oom.h> |
be765d99 | 36 | #include <linux/page_owner.h> |
97ae1749 | 37 | |
71e3aac0 AA |
38 | #include <asm/tlb.h> |
39 | #include <asm/pgalloc.h> | |
40 | #include "internal.h" | |
41 | ||
ba76149f | 42 | /* |
b14d595a MD |
43 | * By default, transparent hugepage support is disabled in order to avoid |
44 | * risking an increased memory footprint for applications that are not | |
45 | * guaranteed to benefit from it. When transparent hugepage support is | |
46 | * enabled, it is for all mappings, and khugepaged scans all mappings. | |
8bfa3f9a JW |
47 | * Defrag is invoked by khugepaged hugepage allocations and by page faults |
48 | * for all hugepage allocations. | |
ba76149f | 49 | */ |
71e3aac0 | 50 | unsigned long transparent_hugepage_flags __read_mostly = |
13ece886 | 51 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS |
ba76149f | 52 | (1<<TRANSPARENT_HUGEPAGE_FLAG)| |
13ece886 AA |
53 | #endif |
54 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE | |
55 | (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| | |
56 | #endif | |
444eb2a4 | 57 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)| |
79da5407 KS |
58 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| |
59 | (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
ba76149f | 60 | |
9a982250 | 61 | static struct shrinker deferred_split_shrinker; |
f000565a | 62 | |
97ae1749 | 63 | static atomic_t huge_zero_refcount; |
56873f43 | 64 | struct page *huge_zero_page __read_mostly; |
4a6c1297 | 65 | |
6fcb52a5 | 66 | static struct page *get_huge_zero_page(void) |
97ae1749 KS |
67 | { |
68 | struct page *zero_page; | |
69 | retry: | |
70 | if (likely(atomic_inc_not_zero(&huge_zero_refcount))) | |
4db0c3c2 | 71 | return READ_ONCE(huge_zero_page); |
97ae1749 KS |
72 | |
73 | zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, | |
4a6c1297 | 74 | HPAGE_PMD_ORDER); |
d8a8e1f0 KS |
75 | if (!zero_page) { |
76 | count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); | |
5918d10a | 77 | return NULL; |
d8a8e1f0 KS |
78 | } |
79 | count_vm_event(THP_ZERO_PAGE_ALLOC); | |
97ae1749 | 80 | preempt_disable(); |
5918d10a | 81 | if (cmpxchg(&huge_zero_page, NULL, zero_page)) { |
97ae1749 | 82 | preempt_enable(); |
5ddacbe9 | 83 | __free_pages(zero_page, compound_order(zero_page)); |
97ae1749 KS |
84 | goto retry; |
85 | } | |
86 | ||
87 | /* We take additional reference here. It will be put back by shrinker */ | |
88 | atomic_set(&huge_zero_refcount, 2); | |
89 | preempt_enable(); | |
4db0c3c2 | 90 | return READ_ONCE(huge_zero_page); |
4a6c1297 KS |
91 | } |
92 | ||
6fcb52a5 | 93 | static void put_huge_zero_page(void) |
4a6c1297 | 94 | { |
97ae1749 KS |
95 | /* |
96 | * Counter should never go to zero here. Only shrinker can put | |
97 | * last reference. | |
98 | */ | |
99 | BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); | |
4a6c1297 KS |
100 | } |
101 | ||
6fcb52a5 AL |
102 | struct page *mm_get_huge_zero_page(struct mm_struct *mm) |
103 | { | |
104 | if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) | |
105 | return READ_ONCE(huge_zero_page); | |
106 | ||
107 | if (!get_huge_zero_page()) | |
108 | return NULL; | |
109 | ||
110 | if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) | |
111 | put_huge_zero_page(); | |
112 | ||
113 | return READ_ONCE(huge_zero_page); | |
114 | } | |
115 | ||
116 | void mm_put_huge_zero_page(struct mm_struct *mm) | |
117 | { | |
118 | if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) | |
119 | put_huge_zero_page(); | |
120 | } | |
121 | ||
48896466 GC |
122 | static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, |
123 | struct shrink_control *sc) | |
4a6c1297 | 124 | { |
48896466 GC |
125 | /* we can free zero page only if last reference remains */ |
126 | return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; | |
127 | } | |
97ae1749 | 128 | |
48896466 GC |
129 | static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, |
130 | struct shrink_control *sc) | |
131 | { | |
97ae1749 | 132 | if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { |
5918d10a KS |
133 | struct page *zero_page = xchg(&huge_zero_page, NULL); |
134 | BUG_ON(zero_page == NULL); | |
5ddacbe9 | 135 | __free_pages(zero_page, compound_order(zero_page)); |
48896466 | 136 | return HPAGE_PMD_NR; |
97ae1749 KS |
137 | } |
138 | ||
139 | return 0; | |
4a6c1297 KS |
140 | } |
141 | ||
97ae1749 | 142 | static struct shrinker huge_zero_page_shrinker = { |
48896466 GC |
143 | .count_objects = shrink_huge_zero_page_count, |
144 | .scan_objects = shrink_huge_zero_page_scan, | |
97ae1749 KS |
145 | .seeks = DEFAULT_SEEKS, |
146 | }; | |
147 | ||
71e3aac0 | 148 | #ifdef CONFIG_SYSFS |
71e3aac0 AA |
149 | static ssize_t enabled_show(struct kobject *kobj, |
150 | struct kobj_attribute *attr, char *buf) | |
151 | { | |
444eb2a4 MG |
152 | if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags)) |
153 | return sprintf(buf, "[always] madvise never\n"); | |
154 | else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags)) | |
155 | return sprintf(buf, "always [madvise] never\n"); | |
156 | else | |
157 | return sprintf(buf, "always madvise [never]\n"); | |
71e3aac0 | 158 | } |
444eb2a4 | 159 | |
71e3aac0 AA |
160 | static ssize_t enabled_store(struct kobject *kobj, |
161 | struct kobj_attribute *attr, | |
162 | const char *buf, size_t count) | |
163 | { | |
21440d7e | 164 | ssize_t ret = count; |
ba76149f | 165 | |
21440d7e DR |
166 | if (!memcmp("always", buf, |
167 | min(sizeof("always")-1, count))) { | |
168 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
169 | set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); | |
170 | } else if (!memcmp("madvise", buf, | |
171 | min(sizeof("madvise")-1, count))) { | |
172 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); | |
173 | set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
174 | } else if (!memcmp("never", buf, | |
175 | min(sizeof("never")-1, count))) { | |
176 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); | |
177 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
178 | } else | |
179 | ret = -EINVAL; | |
ba76149f AA |
180 | |
181 | if (ret > 0) { | |
b46e756f | 182 | int err = start_stop_khugepaged(); |
ba76149f AA |
183 | if (err) |
184 | ret = err; | |
185 | } | |
ba76149f | 186 | return ret; |
71e3aac0 AA |
187 | } |
188 | static struct kobj_attribute enabled_attr = | |
189 | __ATTR(enabled, 0644, enabled_show, enabled_store); | |
190 | ||
b46e756f | 191 | ssize_t single_hugepage_flag_show(struct kobject *kobj, |
71e3aac0 AA |
192 | struct kobj_attribute *attr, char *buf, |
193 | enum transparent_hugepage_flag flag) | |
194 | { | |
e27e6151 BH |
195 | return sprintf(buf, "%d\n", |
196 | !!test_bit(flag, &transparent_hugepage_flags)); | |
71e3aac0 | 197 | } |
e27e6151 | 198 | |
b46e756f | 199 | ssize_t single_hugepage_flag_store(struct kobject *kobj, |
71e3aac0 AA |
200 | struct kobj_attribute *attr, |
201 | const char *buf, size_t count, | |
202 | enum transparent_hugepage_flag flag) | |
203 | { | |
e27e6151 BH |
204 | unsigned long value; |
205 | int ret; | |
206 | ||
207 | ret = kstrtoul(buf, 10, &value); | |
208 | if (ret < 0) | |
209 | return ret; | |
210 | if (value > 1) | |
211 | return -EINVAL; | |
212 | ||
213 | if (value) | |
71e3aac0 | 214 | set_bit(flag, &transparent_hugepage_flags); |
e27e6151 | 215 | else |
71e3aac0 | 216 | clear_bit(flag, &transparent_hugepage_flags); |
71e3aac0 AA |
217 | |
218 | return count; | |
219 | } | |
220 | ||
71e3aac0 AA |
221 | static ssize_t defrag_show(struct kobject *kobj, |
222 | struct kobj_attribute *attr, char *buf) | |
223 | { | |
444eb2a4 | 224 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) |
21440d7e | 225 | return sprintf(buf, "[always] defer defer+madvise madvise never\n"); |
444eb2a4 | 226 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) |
21440d7e DR |
227 | return sprintf(buf, "always [defer] defer+madvise madvise never\n"); |
228 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) | |
229 | return sprintf(buf, "always defer [defer+madvise] madvise never\n"); | |
230 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) | |
231 | return sprintf(buf, "always defer defer+madvise [madvise] never\n"); | |
232 | return sprintf(buf, "always defer defer+madvise madvise [never]\n"); | |
71e3aac0 | 233 | } |
21440d7e | 234 | |
71e3aac0 AA |
235 | static ssize_t defrag_store(struct kobject *kobj, |
236 | struct kobj_attribute *attr, | |
237 | const char *buf, size_t count) | |
238 | { | |
21440d7e DR |
239 | if (!memcmp("always", buf, |
240 | min(sizeof("always")-1, count))) { | |
241 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
242 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
243 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
244 | set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); | |
21440d7e DR |
245 | } else if (!memcmp("defer+madvise", buf, |
246 | min(sizeof("defer+madvise")-1, count))) { | |
247 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); | |
248 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
249 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
250 | set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
4fad7fb6 DR |
251 | } else if (!memcmp("defer", buf, |
252 | min(sizeof("defer")-1, count))) { | |
253 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); | |
254 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
255 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
256 | set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
21440d7e DR |
257 | } else if (!memcmp("madvise", buf, |
258 | min(sizeof("madvise")-1, count))) { | |
259 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); | |
260 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
261 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
262 | set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
263 | } else if (!memcmp("never", buf, | |
264 | min(sizeof("never")-1, count))) { | |
265 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); | |
266 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
267 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
268 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
269 | } else | |
270 | return -EINVAL; | |
271 | ||
272 | return count; | |
71e3aac0 AA |
273 | } |
274 | static struct kobj_attribute defrag_attr = | |
275 | __ATTR(defrag, 0644, defrag_show, defrag_store); | |
276 | ||
79da5407 KS |
277 | static ssize_t use_zero_page_show(struct kobject *kobj, |
278 | struct kobj_attribute *attr, char *buf) | |
279 | { | |
b46e756f | 280 | return single_hugepage_flag_show(kobj, attr, buf, |
79da5407 KS |
281 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); |
282 | } | |
283 | static ssize_t use_zero_page_store(struct kobject *kobj, | |
284 | struct kobj_attribute *attr, const char *buf, size_t count) | |
285 | { | |
b46e756f | 286 | return single_hugepage_flag_store(kobj, attr, buf, count, |
79da5407 KS |
287 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); |
288 | } | |
289 | static struct kobj_attribute use_zero_page_attr = | |
290 | __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); | |
49920d28 HD |
291 | |
292 | static ssize_t hpage_pmd_size_show(struct kobject *kobj, | |
293 | struct kobj_attribute *attr, char *buf) | |
294 | { | |
295 | return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE); | |
296 | } | |
297 | static struct kobj_attribute hpage_pmd_size_attr = | |
298 | __ATTR_RO(hpage_pmd_size); | |
299 | ||
71e3aac0 AA |
300 | #ifdef CONFIG_DEBUG_VM |
301 | static ssize_t debug_cow_show(struct kobject *kobj, | |
302 | struct kobj_attribute *attr, char *buf) | |
303 | { | |
b46e756f | 304 | return single_hugepage_flag_show(kobj, attr, buf, |
71e3aac0 AA |
305 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); |
306 | } | |
307 | static ssize_t debug_cow_store(struct kobject *kobj, | |
308 | struct kobj_attribute *attr, | |
309 | const char *buf, size_t count) | |
310 | { | |
b46e756f | 311 | return single_hugepage_flag_store(kobj, attr, buf, count, |
71e3aac0 AA |
312 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); |
313 | } | |
314 | static struct kobj_attribute debug_cow_attr = | |
315 | __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); | |
316 | #endif /* CONFIG_DEBUG_VM */ | |
317 | ||
318 | static struct attribute *hugepage_attr[] = { | |
319 | &enabled_attr.attr, | |
320 | &defrag_attr.attr, | |
79da5407 | 321 | &use_zero_page_attr.attr, |
49920d28 | 322 | &hpage_pmd_size_attr.attr, |
e496cf3d | 323 | #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE) |
5a6e75f8 KS |
324 | &shmem_enabled_attr.attr, |
325 | #endif | |
71e3aac0 AA |
326 | #ifdef CONFIG_DEBUG_VM |
327 | &debug_cow_attr.attr, | |
328 | #endif | |
329 | NULL, | |
330 | }; | |
331 | ||
8aa95a21 | 332 | static const struct attribute_group hugepage_attr_group = { |
71e3aac0 | 333 | .attrs = hugepage_attr, |
ba76149f AA |
334 | }; |
335 | ||
569e5590 | 336 | static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) |
71e3aac0 | 337 | { |
71e3aac0 AA |
338 | int err; |
339 | ||
569e5590 SL |
340 | *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); |
341 | if (unlikely(!*hugepage_kobj)) { | |
ae3a8c1c | 342 | pr_err("failed to create transparent hugepage kobject\n"); |
569e5590 | 343 | return -ENOMEM; |
ba76149f AA |
344 | } |
345 | ||
569e5590 | 346 | err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); |
ba76149f | 347 | if (err) { |
ae3a8c1c | 348 | pr_err("failed to register transparent hugepage group\n"); |
569e5590 | 349 | goto delete_obj; |
ba76149f AA |
350 | } |
351 | ||
569e5590 | 352 | err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); |
ba76149f | 353 | if (err) { |
ae3a8c1c | 354 | pr_err("failed to register transparent hugepage group\n"); |
569e5590 | 355 | goto remove_hp_group; |
ba76149f | 356 | } |
569e5590 SL |
357 | |
358 | return 0; | |
359 | ||
360 | remove_hp_group: | |
361 | sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); | |
362 | delete_obj: | |
363 | kobject_put(*hugepage_kobj); | |
364 | return err; | |
365 | } | |
366 | ||
367 | static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
368 | { | |
369 | sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); | |
370 | sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); | |
371 | kobject_put(hugepage_kobj); | |
372 | } | |
373 | #else | |
374 | static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) | |
375 | { | |
376 | return 0; | |
377 | } | |
378 | ||
379 | static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
380 | { | |
381 | } | |
382 | #endif /* CONFIG_SYSFS */ | |
383 | ||
384 | static int __init hugepage_init(void) | |
385 | { | |
386 | int err; | |
387 | struct kobject *hugepage_kobj; | |
388 | ||
389 | if (!has_transparent_hugepage()) { | |
390 | transparent_hugepage_flags = 0; | |
391 | return -EINVAL; | |
392 | } | |
393 | ||
ff20c2e0 KS |
394 | /* |
395 | * hugepages can't be allocated by the buddy allocator | |
396 | */ | |
397 | MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER); | |
398 | /* | |
399 | * we use page->mapping and page->index in second tail page | |
400 | * as list_head: assuming THP order >= 2 | |
401 | */ | |
402 | MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2); | |
403 | ||
569e5590 SL |
404 | err = hugepage_init_sysfs(&hugepage_kobj); |
405 | if (err) | |
65ebb64f | 406 | goto err_sysfs; |
ba76149f | 407 | |
b46e756f | 408 | err = khugepaged_init(); |
ba76149f | 409 | if (err) |
65ebb64f | 410 | goto err_slab; |
ba76149f | 411 | |
65ebb64f KS |
412 | err = register_shrinker(&huge_zero_page_shrinker); |
413 | if (err) | |
414 | goto err_hzp_shrinker; | |
9a982250 KS |
415 | err = register_shrinker(&deferred_split_shrinker); |
416 | if (err) | |
417 | goto err_split_shrinker; | |
97ae1749 | 418 | |
97562cd2 RR |
419 | /* |
420 | * By default disable transparent hugepages on smaller systems, | |
421 | * where the extra memory used could hurt more than TLB overhead | |
422 | * is likely to save. The admin can still enable it through /sys. | |
423 | */ | |
79553da2 | 424 | if (totalram_pages < (512 << (20 - PAGE_SHIFT))) { |
97562cd2 | 425 | transparent_hugepage_flags = 0; |
79553da2 KS |
426 | return 0; |
427 | } | |
97562cd2 | 428 | |
79553da2 | 429 | err = start_stop_khugepaged(); |
65ebb64f KS |
430 | if (err) |
431 | goto err_khugepaged; | |
ba76149f | 432 | |
569e5590 | 433 | return 0; |
65ebb64f | 434 | err_khugepaged: |
9a982250 KS |
435 | unregister_shrinker(&deferred_split_shrinker); |
436 | err_split_shrinker: | |
65ebb64f KS |
437 | unregister_shrinker(&huge_zero_page_shrinker); |
438 | err_hzp_shrinker: | |
b46e756f | 439 | khugepaged_destroy(); |
65ebb64f | 440 | err_slab: |
569e5590 | 441 | hugepage_exit_sysfs(hugepage_kobj); |
65ebb64f | 442 | err_sysfs: |
ba76149f | 443 | return err; |
71e3aac0 | 444 | } |
a64fb3cd | 445 | subsys_initcall(hugepage_init); |
71e3aac0 AA |
446 | |
447 | static int __init setup_transparent_hugepage(char *str) | |
448 | { | |
449 | int ret = 0; | |
450 | if (!str) | |
451 | goto out; | |
452 | if (!strcmp(str, "always")) { | |
453 | set_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
454 | &transparent_hugepage_flags); | |
455 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
456 | &transparent_hugepage_flags); | |
457 | ret = 1; | |
458 | } else if (!strcmp(str, "madvise")) { | |
459 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
460 | &transparent_hugepage_flags); | |
461 | set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
462 | &transparent_hugepage_flags); | |
463 | ret = 1; | |
464 | } else if (!strcmp(str, "never")) { | |
465 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
466 | &transparent_hugepage_flags); | |
467 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
468 | &transparent_hugepage_flags); | |
469 | ret = 1; | |
470 | } | |
471 | out: | |
472 | if (!ret) | |
ae3a8c1c | 473 | pr_warn("transparent_hugepage= cannot parse, ignored\n"); |
71e3aac0 AA |
474 | return ret; |
475 | } | |
476 | __setup("transparent_hugepage=", setup_transparent_hugepage); | |
477 | ||
f55e1014 | 478 | pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) |
71e3aac0 | 479 | { |
f55e1014 | 480 | if (likely(vma->vm_flags & VM_WRITE)) |
71e3aac0 AA |
481 | pmd = pmd_mkwrite(pmd); |
482 | return pmd; | |
483 | } | |
484 | ||
9a982250 KS |
485 | static inline struct list_head *page_deferred_list(struct page *page) |
486 | { | |
487 | /* | |
488 | * ->lru in the tail pages is occupied by compound_head. | |
489 | * Let's use ->mapping + ->index in the second tail page as list_head. | |
490 | */ | |
491 | return (struct list_head *)&page[2].mapping; | |
492 | } | |
493 | ||
494 | void prep_transhuge_page(struct page *page) | |
495 | { | |
496 | /* | |
497 | * we use page->mapping and page->indexlru in second tail page | |
498 | * as list_head: assuming THP order >= 2 | |
499 | */ | |
9a982250 KS |
500 | |
501 | INIT_LIST_HEAD(page_deferred_list(page)); | |
502 | set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR); | |
503 | } | |
504 | ||
1e0f9f66 KS |
505 | static unsigned long __thp_get_unmapped_area(struct file *filp, |
506 | unsigned long addr, unsigned long len, | |
74d2fad1 TK |
507 | loff_t off, unsigned long flags, unsigned long size) |
508 | { | |
74d2fad1 TK |
509 | loff_t off_end = off + len; |
510 | loff_t off_align = round_up(off, size); | |
1e0f9f66 | 511 | unsigned long len_pad, ret; |
74d2fad1 TK |
512 | |
513 | if (off_end <= off_align || (off_end - off_align) < size) | |
514 | return 0; | |
515 | ||
516 | len_pad = len + size; | |
517 | if (len_pad < len || (off + len_pad) < off) | |
518 | return 0; | |
519 | ||
1e0f9f66 | 520 | ret = current->mm->get_unmapped_area(filp, addr, len_pad, |
74d2fad1 | 521 | off >> PAGE_SHIFT, flags); |
1e0f9f66 KS |
522 | |
523 | /* | |
524 | * The failure might be due to length padding. The caller will retry | |
525 | * without the padding. | |
526 | */ | |
527 | if (IS_ERR_VALUE(ret)) | |
74d2fad1 TK |
528 | return 0; |
529 | ||
1e0f9f66 KS |
530 | /* |
531 | * Do not try to align to THP boundary if allocation at the address | |
532 | * hint succeeds. | |
533 | */ | |
534 | if (ret == addr) | |
535 | return addr; | |
536 | ||
537 | ret += (off - ret) & (size - 1); | |
538 | return ret; | |
74d2fad1 TK |
539 | } |
540 | ||
541 | unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr, | |
542 | unsigned long len, unsigned long pgoff, unsigned long flags) | |
543 | { | |
1e0f9f66 | 544 | unsigned long ret; |
74d2fad1 TK |
545 | loff_t off = (loff_t)pgoff << PAGE_SHIFT; |
546 | ||
74d2fad1 TK |
547 | if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD)) |
548 | goto out; | |
549 | ||
1e0f9f66 KS |
550 | ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE); |
551 | if (ret) | |
552 | return ret; | |
553 | out: | |
74d2fad1 TK |
554 | return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags); |
555 | } | |
556 | EXPORT_SYMBOL_GPL(thp_get_unmapped_area); | |
557 | ||
82b0f8c3 | 558 | static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page, |
bae473a4 | 559 | gfp_t gfp) |
71e3aac0 | 560 | { |
82b0f8c3 | 561 | struct vm_area_struct *vma = vmf->vma; |
00501b53 | 562 | struct mem_cgroup *memcg; |
71e3aac0 | 563 | pgtable_t pgtable; |
82b0f8c3 | 564 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; |
6b31d595 | 565 | int ret = 0; |
71e3aac0 | 566 | |
309381fe | 567 | VM_BUG_ON_PAGE(!PageCompound(page), page); |
00501b53 | 568 | |
46fe940f DR |
569 | if (mem_cgroup_try_charge(page, vma->vm_mm, gfp | __GFP_NORETRY, &memcg, |
570 | true)) { | |
6b251fc9 AA |
571 | put_page(page); |
572 | count_vm_event(THP_FAULT_FALLBACK); | |
573 | return VM_FAULT_FALLBACK; | |
574 | } | |
00501b53 | 575 | |
bae473a4 | 576 | pgtable = pte_alloc_one(vma->vm_mm, haddr); |
00501b53 | 577 | if (unlikely(!pgtable)) { |
6b31d595 MH |
578 | ret = VM_FAULT_OOM; |
579 | goto release; | |
00501b53 | 580 | } |
71e3aac0 | 581 | |
c79b57e4 | 582 | clear_huge_page(page, vmf->address, HPAGE_PMD_NR); |
52f37629 MK |
583 | /* |
584 | * The memory barrier inside __SetPageUptodate makes sure that | |
585 | * clear_huge_page writes become visible before the set_pmd_at() | |
586 | * write. | |
587 | */ | |
71e3aac0 AA |
588 | __SetPageUptodate(page); |
589 | ||
82b0f8c3 JK |
590 | vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); |
591 | if (unlikely(!pmd_none(*vmf->pmd))) { | |
6b31d595 | 592 | goto unlock_release; |
71e3aac0 AA |
593 | } else { |
594 | pmd_t entry; | |
6b251fc9 | 595 | |
6b31d595 MH |
596 | ret = check_stable_address_space(vma->vm_mm); |
597 | if (ret) | |
598 | goto unlock_release; | |
599 | ||
6b251fc9 AA |
600 | /* Deliver the page fault to userland */ |
601 | if (userfaultfd_missing(vma)) { | |
602 | int ret; | |
603 | ||
82b0f8c3 | 604 | spin_unlock(vmf->ptl); |
f627c2f5 | 605 | mem_cgroup_cancel_charge(page, memcg, true); |
6b251fc9 | 606 | put_page(page); |
bae473a4 | 607 | pte_free(vma->vm_mm, pgtable); |
82b0f8c3 | 608 | ret = handle_userfault(vmf, VM_UFFD_MISSING); |
6b251fc9 AA |
609 | VM_BUG_ON(ret & VM_FAULT_FALLBACK); |
610 | return ret; | |
611 | } | |
612 | ||
3122359a | 613 | entry = mk_huge_pmd(page, vma->vm_page_prot); |
f55e1014 | 614 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
d281ee61 | 615 | page_add_new_anon_rmap(page, vma, haddr, true); |
f627c2f5 | 616 | mem_cgroup_commit_charge(page, memcg, false, true); |
00501b53 | 617 | lru_cache_add_active_or_unevictable(page, vma); |
82b0f8c3 JK |
618 | pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable); |
619 | set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); | |
bae473a4 | 620 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); |
c4812909 | 621 | mm_inc_nr_ptes(vma->vm_mm); |
82b0f8c3 | 622 | spin_unlock(vmf->ptl); |
6b251fc9 | 623 | count_vm_event(THP_FAULT_ALLOC); |
71e3aac0 AA |
624 | } |
625 | ||
aa2e878e | 626 | return 0; |
6b31d595 MH |
627 | unlock_release: |
628 | spin_unlock(vmf->ptl); | |
629 | release: | |
630 | if (pgtable) | |
631 | pte_free(vma->vm_mm, pgtable); | |
632 | mem_cgroup_cancel_charge(page, memcg, true); | |
633 | put_page(page); | |
634 | return ret; | |
635 | ||
71e3aac0 AA |
636 | } |
637 | ||
444eb2a4 | 638 | /* |
21440d7e DR |
639 | * always: directly stall for all thp allocations |
640 | * defer: wake kswapd and fail if not immediately available | |
641 | * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise | |
642 | * fail if not immediately available | |
643 | * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately | |
644 | * available | |
645 | * never: never stall for any thp allocation | |
444eb2a4 MG |
646 | */ |
647 | static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma) | |
648 | { | |
21440d7e | 649 | const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE); |
25160354 | 650 | |
21440d7e | 651 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) |
25160354 | 652 | return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY); |
21440d7e DR |
653 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) |
654 | return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM; | |
655 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) | |
656 | return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM : | |
657 | __GFP_KSWAPD_RECLAIM); | |
658 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) | |
659 | return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM : | |
660 | 0); | |
25160354 | 661 | return GFP_TRANSHUGE_LIGHT; |
444eb2a4 MG |
662 | } |
663 | ||
c4088ebd | 664 | /* Caller must hold page table lock. */ |
d295e341 | 665 | static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, |
97ae1749 | 666 | struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, |
5918d10a | 667 | struct page *zero_page) |
fc9fe822 KS |
668 | { |
669 | pmd_t entry; | |
7c414164 AM |
670 | if (!pmd_none(*pmd)) |
671 | return false; | |
5918d10a | 672 | entry = mk_pmd(zero_page, vma->vm_page_prot); |
fc9fe822 | 673 | entry = pmd_mkhuge(entry); |
12c9d70b MW |
674 | if (pgtable) |
675 | pgtable_trans_huge_deposit(mm, pmd, pgtable); | |
fc9fe822 | 676 | set_pmd_at(mm, haddr, pmd, entry); |
c4812909 | 677 | mm_inc_nr_ptes(mm); |
7c414164 | 678 | return true; |
fc9fe822 KS |
679 | } |
680 | ||
82b0f8c3 | 681 | int do_huge_pmd_anonymous_page(struct vm_fault *vmf) |
71e3aac0 | 682 | { |
82b0f8c3 | 683 | struct vm_area_struct *vma = vmf->vma; |
077fcf11 | 684 | gfp_t gfp; |
71e3aac0 | 685 | struct page *page; |
82b0f8c3 | 686 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; |
71e3aac0 | 687 | |
128ec037 | 688 | if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end) |
c0292554 | 689 | return VM_FAULT_FALLBACK; |
128ec037 KS |
690 | if (unlikely(anon_vma_prepare(vma))) |
691 | return VM_FAULT_OOM; | |
6d50e60c | 692 | if (unlikely(khugepaged_enter(vma, vma->vm_flags))) |
128ec037 | 693 | return VM_FAULT_OOM; |
82b0f8c3 | 694 | if (!(vmf->flags & FAULT_FLAG_WRITE) && |
bae473a4 | 695 | !mm_forbids_zeropage(vma->vm_mm) && |
128ec037 KS |
696 | transparent_hugepage_use_zero_page()) { |
697 | pgtable_t pgtable; | |
698 | struct page *zero_page; | |
699 | bool set; | |
6b251fc9 | 700 | int ret; |
bae473a4 | 701 | pgtable = pte_alloc_one(vma->vm_mm, haddr); |
128ec037 | 702 | if (unlikely(!pgtable)) |
ba76149f | 703 | return VM_FAULT_OOM; |
6fcb52a5 | 704 | zero_page = mm_get_huge_zero_page(vma->vm_mm); |
128ec037 | 705 | if (unlikely(!zero_page)) { |
bae473a4 | 706 | pte_free(vma->vm_mm, pgtable); |
81ab4201 | 707 | count_vm_event(THP_FAULT_FALLBACK); |
c0292554 | 708 | return VM_FAULT_FALLBACK; |
b9bbfbe3 | 709 | } |
82b0f8c3 | 710 | vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); |
6b251fc9 AA |
711 | ret = 0; |
712 | set = false; | |
82b0f8c3 | 713 | if (pmd_none(*vmf->pmd)) { |
6b31d595 MH |
714 | ret = check_stable_address_space(vma->vm_mm); |
715 | if (ret) { | |
716 | spin_unlock(vmf->ptl); | |
717 | } else if (userfaultfd_missing(vma)) { | |
82b0f8c3 JK |
718 | spin_unlock(vmf->ptl); |
719 | ret = handle_userfault(vmf, VM_UFFD_MISSING); | |
6b251fc9 AA |
720 | VM_BUG_ON(ret & VM_FAULT_FALLBACK); |
721 | } else { | |
bae473a4 | 722 | set_huge_zero_page(pgtable, vma->vm_mm, vma, |
82b0f8c3 JK |
723 | haddr, vmf->pmd, zero_page); |
724 | spin_unlock(vmf->ptl); | |
6b251fc9 AA |
725 | set = true; |
726 | } | |
727 | } else | |
82b0f8c3 | 728 | spin_unlock(vmf->ptl); |
6fcb52a5 | 729 | if (!set) |
bae473a4 | 730 | pte_free(vma->vm_mm, pgtable); |
6b251fc9 | 731 | return ret; |
71e3aac0 | 732 | } |
444eb2a4 | 733 | gfp = alloc_hugepage_direct_gfpmask(vma); |
077fcf11 | 734 | page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); |
128ec037 KS |
735 | if (unlikely(!page)) { |
736 | count_vm_event(THP_FAULT_FALLBACK); | |
c0292554 | 737 | return VM_FAULT_FALLBACK; |
128ec037 | 738 | } |
9a982250 | 739 | prep_transhuge_page(page); |
82b0f8c3 | 740 | return __do_huge_pmd_anonymous_page(vmf, page, gfp); |
71e3aac0 AA |
741 | } |
742 | ||
ae18d6dc | 743 | static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, |
3b6521f5 OH |
744 | pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write, |
745 | pgtable_t pgtable) | |
5cad465d MW |
746 | { |
747 | struct mm_struct *mm = vma->vm_mm; | |
748 | pmd_t entry; | |
749 | spinlock_t *ptl; | |
750 | ||
751 | ptl = pmd_lock(mm, pmd); | |
0da17885 AK |
752 | if (!pmd_none(*pmd)) { |
753 | if (write) { | |
754 | if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) { | |
755 | WARN_ON_ONCE(!is_huge_zero_pmd(*pmd)); | |
756 | goto out_unlock; | |
757 | } | |
758 | entry = pmd_mkyoung(*pmd); | |
759 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
760 | if (pmdp_set_access_flags(vma, addr, pmd, entry, 1)) | |
761 | update_mmu_cache_pmd(vma, addr, pmd); | |
762 | } | |
763 | ||
764 | goto out_unlock; | |
765 | } | |
766 | ||
f25748e3 DW |
767 | entry = pmd_mkhuge(pfn_t_pmd(pfn, prot)); |
768 | if (pfn_t_devmap(pfn)) | |
769 | entry = pmd_mkdevmap(entry); | |
01871e59 | 770 | if (write) { |
f55e1014 LT |
771 | entry = pmd_mkyoung(pmd_mkdirty(entry)); |
772 | entry = maybe_pmd_mkwrite(entry, vma); | |
5cad465d | 773 | } |
3b6521f5 OH |
774 | |
775 | if (pgtable) { | |
776 | pgtable_trans_huge_deposit(mm, pmd, pgtable); | |
c4812909 | 777 | mm_inc_nr_ptes(mm); |
0da17885 | 778 | pgtable = NULL; |
3b6521f5 OH |
779 | } |
780 | ||
01871e59 RZ |
781 | set_pmd_at(mm, addr, pmd, entry); |
782 | update_mmu_cache_pmd(vma, addr, pmd); | |
0da17885 AK |
783 | |
784 | out_unlock: | |
5cad465d | 785 | spin_unlock(ptl); |
0da17885 AK |
786 | if (pgtable) |
787 | pte_free(mm, pgtable); | |
5cad465d MW |
788 | } |
789 | ||
218fa1c8 | 790 | int vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write) |
5cad465d | 791 | { |
218fa1c8 DW |
792 | unsigned long addr = vmf->address & PMD_MASK; |
793 | struct vm_area_struct *vma = vmf->vma; | |
5cad465d | 794 | pgprot_t pgprot = vma->vm_page_prot; |
3b6521f5 | 795 | pgtable_t pgtable = NULL; |
218fa1c8 | 796 | |
5cad465d MW |
797 | /* |
798 | * If we had pmd_special, we could avoid all these restrictions, | |
799 | * but we need to be consistent with PTEs and architectures that | |
800 | * can't support a 'special' bit. | |
801 | */ | |
802 | BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); | |
803 | BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == | |
804 | (VM_PFNMAP|VM_MIXEDMAP)); | |
805 | BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); | |
f25748e3 | 806 | BUG_ON(!pfn_t_devmap(pfn)); |
5cad465d MW |
807 | |
808 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
809 | return VM_FAULT_SIGBUS; | |
308a047c | 810 | |
3b6521f5 OH |
811 | if (arch_needs_pgtable_deposit()) { |
812 | pgtable = pte_alloc_one(vma->vm_mm, addr); | |
813 | if (!pgtable) | |
814 | return VM_FAULT_OOM; | |
815 | } | |
816 | ||
308a047c BP |
817 | track_pfn_insert(vma, &pgprot, pfn); |
818 | ||
218fa1c8 | 819 | insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable); |
ae18d6dc | 820 | return VM_FAULT_NOPAGE; |
5cad465d | 821 | } |
dee41079 | 822 | EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd); |
5cad465d | 823 | |
a00cc7d9 | 824 | #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD |
f55e1014 | 825 | static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma) |
a00cc7d9 | 826 | { |
f55e1014 | 827 | if (likely(vma->vm_flags & VM_WRITE)) |
a00cc7d9 MW |
828 | pud = pud_mkwrite(pud); |
829 | return pud; | |
830 | } | |
831 | ||
832 | static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, | |
833 | pud_t *pud, pfn_t pfn, pgprot_t prot, bool write) | |
834 | { | |
835 | struct mm_struct *mm = vma->vm_mm; | |
836 | pud_t entry; | |
837 | spinlock_t *ptl; | |
838 | ||
839 | ptl = pud_lock(mm, pud); | |
0da17885 AK |
840 | if (!pud_none(*pud)) { |
841 | if (write) { | |
842 | if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) { | |
843 | WARN_ON_ONCE(!is_huge_zero_pud(*pud)); | |
844 | goto out_unlock; | |
845 | } | |
846 | entry = pud_mkyoung(*pud); | |
847 | entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma); | |
848 | if (pudp_set_access_flags(vma, addr, pud, entry, 1)) | |
849 | update_mmu_cache_pud(vma, addr, pud); | |
850 | } | |
851 | goto out_unlock; | |
852 | } | |
853 | ||
a00cc7d9 MW |
854 | entry = pud_mkhuge(pfn_t_pud(pfn, prot)); |
855 | if (pfn_t_devmap(pfn)) | |
856 | entry = pud_mkdevmap(entry); | |
857 | if (write) { | |
f55e1014 LT |
858 | entry = pud_mkyoung(pud_mkdirty(entry)); |
859 | entry = maybe_pud_mkwrite(entry, vma); | |
a00cc7d9 MW |
860 | } |
861 | set_pud_at(mm, addr, pud, entry); | |
862 | update_mmu_cache_pud(vma, addr, pud); | |
0da17885 AK |
863 | |
864 | out_unlock: | |
a00cc7d9 MW |
865 | spin_unlock(ptl); |
866 | } | |
867 | ||
218fa1c8 | 868 | int vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write) |
a00cc7d9 | 869 | { |
218fa1c8 DW |
870 | unsigned long addr = vmf->address & PUD_MASK; |
871 | struct vm_area_struct *vma = vmf->vma; | |
a00cc7d9 | 872 | pgprot_t pgprot = vma->vm_page_prot; |
218fa1c8 | 873 | |
a00cc7d9 MW |
874 | /* |
875 | * If we had pud_special, we could avoid all these restrictions, | |
876 | * but we need to be consistent with PTEs and architectures that | |
877 | * can't support a 'special' bit. | |
878 | */ | |
879 | BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); | |
880 | BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == | |
881 | (VM_PFNMAP|VM_MIXEDMAP)); | |
882 | BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); | |
883 | BUG_ON(!pfn_t_devmap(pfn)); | |
884 | ||
885 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
886 | return VM_FAULT_SIGBUS; | |
887 | ||
888 | track_pfn_insert(vma, &pgprot, pfn); | |
889 | ||
218fa1c8 | 890 | insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write); |
a00cc7d9 MW |
891 | return VM_FAULT_NOPAGE; |
892 | } | |
893 | EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud); | |
894 | #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ | |
895 | ||
3565fce3 | 896 | static void touch_pmd(struct vm_area_struct *vma, unsigned long addr, |
a8f97366 | 897 | pmd_t *pmd, int flags) |
3565fce3 DW |
898 | { |
899 | pmd_t _pmd; | |
900 | ||
a8f97366 KS |
901 | _pmd = pmd_mkyoung(*pmd); |
902 | if (flags & FOLL_WRITE) | |
903 | _pmd = pmd_mkdirty(_pmd); | |
3565fce3 | 904 | if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, |
a8f97366 | 905 | pmd, _pmd, flags & FOLL_WRITE)) |
3565fce3 DW |
906 | update_mmu_cache_pmd(vma, addr, pmd); |
907 | } | |
908 | ||
909 | struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr, | |
910 | pmd_t *pmd, int flags) | |
911 | { | |
912 | unsigned long pfn = pmd_pfn(*pmd); | |
913 | struct mm_struct *mm = vma->vm_mm; | |
914 | struct dev_pagemap *pgmap; | |
915 | struct page *page; | |
916 | ||
917 | assert_spin_locked(pmd_lockptr(mm, pmd)); | |
918 | ||
8310d48b KF |
919 | /* |
920 | * When we COW a devmap PMD entry, we split it into PTEs, so we should | |
921 | * not be in this function with `flags & FOLL_COW` set. | |
922 | */ | |
923 | WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set"); | |
924 | ||
f6f37321 | 925 | if (flags & FOLL_WRITE && !pmd_write(*pmd)) |
3565fce3 DW |
926 | return NULL; |
927 | ||
928 | if (pmd_present(*pmd) && pmd_devmap(*pmd)) | |
929 | /* pass */; | |
930 | else | |
931 | return NULL; | |
932 | ||
933 | if (flags & FOLL_TOUCH) | |
a8f97366 | 934 | touch_pmd(vma, addr, pmd, flags); |
3565fce3 DW |
935 | |
936 | /* | |
937 | * device mapped pages can only be returned if the | |
938 | * caller will manage the page reference count. | |
939 | */ | |
940 | if (!(flags & FOLL_GET)) | |
941 | return ERR_PTR(-EEXIST); | |
942 | ||
943 | pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT; | |
944 | pgmap = get_dev_pagemap(pfn, NULL); | |
945 | if (!pgmap) | |
946 | return ERR_PTR(-EFAULT); | |
947 | page = pfn_to_page(pfn); | |
948 | get_page(page); | |
949 | put_dev_pagemap(pgmap); | |
950 | ||
951 | return page; | |
952 | } | |
953 | ||
71e3aac0 AA |
954 | int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, |
955 | pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, | |
956 | struct vm_area_struct *vma) | |
957 | { | |
c4088ebd | 958 | spinlock_t *dst_ptl, *src_ptl; |
71e3aac0 AA |
959 | struct page *src_page; |
960 | pmd_t pmd; | |
12c9d70b | 961 | pgtable_t pgtable = NULL; |
628d47ce | 962 | int ret = -ENOMEM; |
71e3aac0 | 963 | |
628d47ce KS |
964 | /* Skip if can be re-fill on fault */ |
965 | if (!vma_is_anonymous(vma)) | |
966 | return 0; | |
967 | ||
968 | pgtable = pte_alloc_one(dst_mm, addr); | |
969 | if (unlikely(!pgtable)) | |
970 | goto out; | |
71e3aac0 | 971 | |
c4088ebd KS |
972 | dst_ptl = pmd_lock(dst_mm, dst_pmd); |
973 | src_ptl = pmd_lockptr(src_mm, src_pmd); | |
974 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
71e3aac0 AA |
975 | |
976 | ret = -EAGAIN; | |
977 | pmd = *src_pmd; | |
84c3fc4e ZY |
978 | |
979 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION | |
980 | if (unlikely(is_swap_pmd(pmd))) { | |
981 | swp_entry_t entry = pmd_to_swp_entry(pmd); | |
982 | ||
983 | VM_BUG_ON(!is_pmd_migration_entry(pmd)); | |
984 | if (is_write_migration_entry(entry)) { | |
985 | make_migration_entry_read(&entry); | |
986 | pmd = swp_entry_to_pmd(entry); | |
ab6e3d09 NH |
987 | if (pmd_swp_soft_dirty(*src_pmd)) |
988 | pmd = pmd_swp_mksoft_dirty(pmd); | |
84c3fc4e ZY |
989 | set_pmd_at(src_mm, addr, src_pmd, pmd); |
990 | } | |
dd8a67f9 | 991 | add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); |
af5b0f6a | 992 | mm_inc_nr_ptes(dst_mm); |
dd8a67f9 | 993 | pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); |
84c3fc4e ZY |
994 | set_pmd_at(dst_mm, addr, dst_pmd, pmd); |
995 | ret = 0; | |
996 | goto out_unlock; | |
997 | } | |
998 | #endif | |
999 | ||
628d47ce | 1000 | if (unlikely(!pmd_trans_huge(pmd))) { |
71e3aac0 AA |
1001 | pte_free(dst_mm, pgtable); |
1002 | goto out_unlock; | |
1003 | } | |
fc9fe822 | 1004 | /* |
c4088ebd | 1005 | * When page table lock is held, the huge zero pmd should not be |
fc9fe822 KS |
1006 | * under splitting since we don't split the page itself, only pmd to |
1007 | * a page table. | |
1008 | */ | |
1009 | if (is_huge_zero_pmd(pmd)) { | |
5918d10a | 1010 | struct page *zero_page; |
97ae1749 KS |
1011 | /* |
1012 | * get_huge_zero_page() will never allocate a new page here, | |
1013 | * since we already have a zero page to copy. It just takes a | |
1014 | * reference. | |
1015 | */ | |
6fcb52a5 | 1016 | zero_page = mm_get_huge_zero_page(dst_mm); |
6b251fc9 | 1017 | set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, |
5918d10a | 1018 | zero_page); |
fc9fe822 KS |
1019 | ret = 0; |
1020 | goto out_unlock; | |
1021 | } | |
de466bd6 | 1022 | |
628d47ce KS |
1023 | src_page = pmd_page(pmd); |
1024 | VM_BUG_ON_PAGE(!PageHead(src_page), src_page); | |
1025 | get_page(src_page); | |
1026 | page_dup_rmap(src_page, true); | |
1027 | add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); | |
c4812909 | 1028 | mm_inc_nr_ptes(dst_mm); |
628d47ce | 1029 | pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); |
71e3aac0 AA |
1030 | |
1031 | pmdp_set_wrprotect(src_mm, addr, src_pmd); | |
1032 | pmd = pmd_mkold(pmd_wrprotect(pmd)); | |
1033 | set_pmd_at(dst_mm, addr, dst_pmd, pmd); | |
71e3aac0 AA |
1034 | |
1035 | ret = 0; | |
1036 | out_unlock: | |
c4088ebd KS |
1037 | spin_unlock(src_ptl); |
1038 | spin_unlock(dst_ptl); | |
71e3aac0 AA |
1039 | out: |
1040 | return ret; | |
1041 | } | |
1042 | ||
a00cc7d9 MW |
1043 | #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD |
1044 | static void touch_pud(struct vm_area_struct *vma, unsigned long addr, | |
a8f97366 | 1045 | pud_t *pud, int flags) |
a00cc7d9 MW |
1046 | { |
1047 | pud_t _pud; | |
1048 | ||
a8f97366 KS |
1049 | _pud = pud_mkyoung(*pud); |
1050 | if (flags & FOLL_WRITE) | |
1051 | _pud = pud_mkdirty(_pud); | |
a00cc7d9 | 1052 | if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK, |
a8f97366 | 1053 | pud, _pud, flags & FOLL_WRITE)) |
a00cc7d9 MW |
1054 | update_mmu_cache_pud(vma, addr, pud); |
1055 | } | |
1056 | ||
1057 | struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr, | |
1058 | pud_t *pud, int flags) | |
1059 | { | |
1060 | unsigned long pfn = pud_pfn(*pud); | |
1061 | struct mm_struct *mm = vma->vm_mm; | |
1062 | struct dev_pagemap *pgmap; | |
1063 | struct page *page; | |
1064 | ||
1065 | assert_spin_locked(pud_lockptr(mm, pud)); | |
1066 | ||
f6f37321 | 1067 | if (flags & FOLL_WRITE && !pud_write(*pud)) |
a00cc7d9 MW |
1068 | return NULL; |
1069 | ||
1070 | if (pud_present(*pud) && pud_devmap(*pud)) | |
1071 | /* pass */; | |
1072 | else | |
1073 | return NULL; | |
1074 | ||
1075 | if (flags & FOLL_TOUCH) | |
a8f97366 | 1076 | touch_pud(vma, addr, pud, flags); |
a00cc7d9 MW |
1077 | |
1078 | /* | |
1079 | * device mapped pages can only be returned if the | |
1080 | * caller will manage the page reference count. | |
1081 | */ | |
1082 | if (!(flags & FOLL_GET)) | |
1083 | return ERR_PTR(-EEXIST); | |
1084 | ||
1085 | pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT; | |
1086 | pgmap = get_dev_pagemap(pfn, NULL); | |
1087 | if (!pgmap) | |
1088 | return ERR_PTR(-EFAULT); | |
1089 | page = pfn_to_page(pfn); | |
1090 | get_page(page); | |
1091 | put_dev_pagemap(pgmap); | |
1092 | ||
1093 | return page; | |
1094 | } | |
1095 | ||
1096 | int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm, | |
1097 | pud_t *dst_pud, pud_t *src_pud, unsigned long addr, | |
1098 | struct vm_area_struct *vma) | |
1099 | { | |
1100 | spinlock_t *dst_ptl, *src_ptl; | |
1101 | pud_t pud; | |
1102 | int ret; | |
1103 | ||
1104 | dst_ptl = pud_lock(dst_mm, dst_pud); | |
1105 | src_ptl = pud_lockptr(src_mm, src_pud); | |
1106 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
1107 | ||
1108 | ret = -EAGAIN; | |
1109 | pud = *src_pud; | |
1110 | if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud))) | |
1111 | goto out_unlock; | |
1112 | ||
1113 | /* | |
1114 | * When page table lock is held, the huge zero pud should not be | |
1115 | * under splitting since we don't split the page itself, only pud to | |
1116 | * a page table. | |
1117 | */ | |
1118 | if (is_huge_zero_pud(pud)) { | |
1119 | /* No huge zero pud yet */ | |
1120 | } | |
1121 | ||
1122 | pudp_set_wrprotect(src_mm, addr, src_pud); | |
1123 | pud = pud_mkold(pud_wrprotect(pud)); | |
1124 | set_pud_at(dst_mm, addr, dst_pud, pud); | |
1125 | ||
1126 | ret = 0; | |
1127 | out_unlock: | |
1128 | spin_unlock(src_ptl); | |
1129 | spin_unlock(dst_ptl); | |
1130 | return ret; | |
1131 | } | |
1132 | ||
1133 | void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud) | |
1134 | { | |
1135 | pud_t entry; | |
1136 | unsigned long haddr; | |
1137 | bool write = vmf->flags & FAULT_FLAG_WRITE; | |
1138 | ||
1139 | vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud); | |
1140 | if (unlikely(!pud_same(*vmf->pud, orig_pud))) | |
1141 | goto unlock; | |
1142 | ||
1143 | entry = pud_mkyoung(orig_pud); | |
1144 | if (write) | |
1145 | entry = pud_mkdirty(entry); | |
1146 | haddr = vmf->address & HPAGE_PUD_MASK; | |
1147 | if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write)) | |
1148 | update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud); | |
1149 | ||
1150 | unlock: | |
1151 | spin_unlock(vmf->ptl); | |
1152 | } | |
1153 | #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ | |
1154 | ||
82b0f8c3 | 1155 | void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd) |
a1dd450b WD |
1156 | { |
1157 | pmd_t entry; | |
1158 | unsigned long haddr; | |
20f664aa | 1159 | bool write = vmf->flags & FAULT_FLAG_WRITE; |
a1dd450b | 1160 | |
82b0f8c3 JK |
1161 | vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); |
1162 | if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) | |
a1dd450b WD |
1163 | goto unlock; |
1164 | ||
1165 | entry = pmd_mkyoung(orig_pmd); | |
20f664aa MK |
1166 | if (write) |
1167 | entry = pmd_mkdirty(entry); | |
82b0f8c3 | 1168 | haddr = vmf->address & HPAGE_PMD_MASK; |
20f664aa | 1169 | if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write)) |
82b0f8c3 | 1170 | update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd); |
a1dd450b WD |
1171 | |
1172 | unlock: | |
82b0f8c3 | 1173 | spin_unlock(vmf->ptl); |
a1dd450b WD |
1174 | } |
1175 | ||
82b0f8c3 | 1176 | static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd, |
bae473a4 | 1177 | struct page *page) |
71e3aac0 | 1178 | { |
82b0f8c3 JK |
1179 | struct vm_area_struct *vma = vmf->vma; |
1180 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; | |
00501b53 | 1181 | struct mem_cgroup *memcg; |
71e3aac0 AA |
1182 | pgtable_t pgtable; |
1183 | pmd_t _pmd; | |
1184 | int ret = 0, i; | |
1185 | struct page **pages; | |
2ec74c3e SG |
1186 | unsigned long mmun_start; /* For mmu_notifiers */ |
1187 | unsigned long mmun_end; /* For mmu_notifiers */ | |
71e3aac0 AA |
1188 | |
1189 | pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR, | |
1190 | GFP_KERNEL); | |
1191 | if (unlikely(!pages)) { | |
1192 | ret |= VM_FAULT_OOM; | |
1193 | goto out; | |
1194 | } | |
1195 | ||
1196 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
41b6167e | 1197 | pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma, |
82b0f8c3 | 1198 | vmf->address, page_to_nid(page)); |
b9bbfbe3 | 1199 | if (unlikely(!pages[i] || |
bae473a4 KS |
1200 | mem_cgroup_try_charge(pages[i], vma->vm_mm, |
1201 | GFP_KERNEL, &memcg, false))) { | |
b9bbfbe3 | 1202 | if (pages[i]) |
71e3aac0 | 1203 | put_page(pages[i]); |
b9bbfbe3 | 1204 | while (--i >= 0) { |
00501b53 JW |
1205 | memcg = (void *)page_private(pages[i]); |
1206 | set_page_private(pages[i], 0); | |
f627c2f5 KS |
1207 | mem_cgroup_cancel_charge(pages[i], memcg, |
1208 | false); | |
b9bbfbe3 AA |
1209 | put_page(pages[i]); |
1210 | } | |
71e3aac0 AA |
1211 | kfree(pages); |
1212 | ret |= VM_FAULT_OOM; | |
1213 | goto out; | |
1214 | } | |
00501b53 | 1215 | set_page_private(pages[i], (unsigned long)memcg); |
71e3aac0 AA |
1216 | } |
1217 | ||
1218 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
1219 | copy_user_highpage(pages[i], page + i, | |
0089e485 | 1220 | haddr + PAGE_SIZE * i, vma); |
71e3aac0 AA |
1221 | __SetPageUptodate(pages[i]); |
1222 | cond_resched(); | |
1223 | } | |
1224 | ||
2ec74c3e SG |
1225 | mmun_start = haddr; |
1226 | mmun_end = haddr + HPAGE_PMD_SIZE; | |
bae473a4 | 1227 | mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end); |
2ec74c3e | 1228 | |
82b0f8c3 JK |
1229 | vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); |
1230 | if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) | |
71e3aac0 | 1231 | goto out_free_pages; |
309381fe | 1232 | VM_BUG_ON_PAGE(!PageHead(page), page); |
71e3aac0 | 1233 | |
0f10851e JG |
1234 | /* |
1235 | * Leave pmd empty until pte is filled note we must notify here as | |
1236 | * concurrent CPU thread might write to new page before the call to | |
1237 | * mmu_notifier_invalidate_range_end() happens which can lead to a | |
1238 | * device seeing memory write in different order than CPU. | |
1239 | * | |
1240 | * See Documentation/vm/mmu_notifier.txt | |
1241 | */ | |
82b0f8c3 | 1242 | pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd); |
71e3aac0 | 1243 | |
82b0f8c3 | 1244 | pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd); |
bae473a4 | 1245 | pmd_populate(vma->vm_mm, &_pmd, pgtable); |
71e3aac0 AA |
1246 | |
1247 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
bae473a4 | 1248 | pte_t entry; |
71e3aac0 AA |
1249 | entry = mk_pte(pages[i], vma->vm_page_prot); |
1250 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); | |
00501b53 JW |
1251 | memcg = (void *)page_private(pages[i]); |
1252 | set_page_private(pages[i], 0); | |
82b0f8c3 | 1253 | page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false); |
f627c2f5 | 1254 | mem_cgroup_commit_charge(pages[i], memcg, false, false); |
00501b53 | 1255 | lru_cache_add_active_or_unevictable(pages[i], vma); |
82b0f8c3 JK |
1256 | vmf->pte = pte_offset_map(&_pmd, haddr); |
1257 | VM_BUG_ON(!pte_none(*vmf->pte)); | |
1258 | set_pte_at(vma->vm_mm, haddr, vmf->pte, entry); | |
1259 | pte_unmap(vmf->pte); | |
71e3aac0 AA |
1260 | } |
1261 | kfree(pages); | |
1262 | ||
71e3aac0 | 1263 | smp_wmb(); /* make pte visible before pmd */ |
82b0f8c3 | 1264 | pmd_populate(vma->vm_mm, vmf->pmd, pgtable); |
d281ee61 | 1265 | page_remove_rmap(page, true); |
82b0f8c3 | 1266 | spin_unlock(vmf->ptl); |
71e3aac0 | 1267 | |
4645b9fe JG |
1268 | /* |
1269 | * No need to double call mmu_notifier->invalidate_range() callback as | |
1270 | * the above pmdp_huge_clear_flush_notify() did already call it. | |
1271 | */ | |
1272 | mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start, | |
1273 | mmun_end); | |
2ec74c3e | 1274 | |
71e3aac0 AA |
1275 | ret |= VM_FAULT_WRITE; |
1276 | put_page(page); | |
1277 | ||
1278 | out: | |
1279 | return ret; | |
1280 | ||
1281 | out_free_pages: | |
82b0f8c3 | 1282 | spin_unlock(vmf->ptl); |
bae473a4 | 1283 | mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); |
b9bbfbe3 | 1284 | for (i = 0; i < HPAGE_PMD_NR; i++) { |
00501b53 JW |
1285 | memcg = (void *)page_private(pages[i]); |
1286 | set_page_private(pages[i], 0); | |
f627c2f5 | 1287 | mem_cgroup_cancel_charge(pages[i], memcg, false); |
71e3aac0 | 1288 | put_page(pages[i]); |
b9bbfbe3 | 1289 | } |
71e3aac0 AA |
1290 | kfree(pages); |
1291 | goto out; | |
1292 | } | |
1293 | ||
82b0f8c3 | 1294 | int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd) |
71e3aac0 | 1295 | { |
82b0f8c3 | 1296 | struct vm_area_struct *vma = vmf->vma; |
93b4796d | 1297 | struct page *page = NULL, *new_page; |
00501b53 | 1298 | struct mem_cgroup *memcg; |
82b0f8c3 | 1299 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; |
2ec74c3e SG |
1300 | unsigned long mmun_start; /* For mmu_notifiers */ |
1301 | unsigned long mmun_end; /* For mmu_notifiers */ | |
3b363692 | 1302 | gfp_t huge_gfp; /* for allocation and charge */ |
bae473a4 | 1303 | int ret = 0; |
71e3aac0 | 1304 | |
82b0f8c3 | 1305 | vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd); |
81d1b09c | 1306 | VM_BUG_ON_VMA(!vma->anon_vma, vma); |
93b4796d KS |
1307 | if (is_huge_zero_pmd(orig_pmd)) |
1308 | goto alloc; | |
82b0f8c3 JK |
1309 | spin_lock(vmf->ptl); |
1310 | if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) | |
71e3aac0 AA |
1311 | goto out_unlock; |
1312 | ||
1313 | page = pmd_page(orig_pmd); | |
309381fe | 1314 | VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); |
1f25fe20 KS |
1315 | /* |
1316 | * We can only reuse the page if nobody else maps the huge page or it's | |
6d0a07ed | 1317 | * part. |
1f25fe20 | 1318 | */ |
ba3c4ce6 HY |
1319 | if (!trylock_page(page)) { |
1320 | get_page(page); | |
1321 | spin_unlock(vmf->ptl); | |
1322 | lock_page(page); | |
1323 | spin_lock(vmf->ptl); | |
1324 | if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { | |
1325 | unlock_page(page); | |
1326 | put_page(page); | |
1327 | goto out_unlock; | |
1328 | } | |
1329 | put_page(page); | |
1330 | } | |
1331 | if (reuse_swap_page(page, NULL)) { | |
71e3aac0 AA |
1332 | pmd_t entry; |
1333 | entry = pmd_mkyoung(orig_pmd); | |
f55e1014 | 1334 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
82b0f8c3 JK |
1335 | if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1)) |
1336 | update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); | |
71e3aac0 | 1337 | ret |= VM_FAULT_WRITE; |
ba3c4ce6 | 1338 | unlock_page(page); |
71e3aac0 AA |
1339 | goto out_unlock; |
1340 | } | |
ba3c4ce6 | 1341 | unlock_page(page); |
ddc58f27 | 1342 | get_page(page); |
82b0f8c3 | 1343 | spin_unlock(vmf->ptl); |
93b4796d | 1344 | alloc: |
71e3aac0 | 1345 | if (transparent_hugepage_enabled(vma) && |
077fcf11 | 1346 | !transparent_hugepage_debug_cow()) { |
444eb2a4 | 1347 | huge_gfp = alloc_hugepage_direct_gfpmask(vma); |
3b363692 | 1348 | new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER); |
077fcf11 | 1349 | } else |
71e3aac0 AA |
1350 | new_page = NULL; |
1351 | ||
9a982250 KS |
1352 | if (likely(new_page)) { |
1353 | prep_transhuge_page(new_page); | |
1354 | } else { | |
eecc1e42 | 1355 | if (!page) { |
82b0f8c3 | 1356 | split_huge_pmd(vma, vmf->pmd, vmf->address); |
e9b71ca9 | 1357 | ret |= VM_FAULT_FALLBACK; |
93b4796d | 1358 | } else { |
82b0f8c3 | 1359 | ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page); |
9845cbbd | 1360 | if (ret & VM_FAULT_OOM) { |
82b0f8c3 | 1361 | split_huge_pmd(vma, vmf->pmd, vmf->address); |
9845cbbd KS |
1362 | ret |= VM_FAULT_FALLBACK; |
1363 | } | |
ddc58f27 | 1364 | put_page(page); |
93b4796d | 1365 | } |
17766dde | 1366 | count_vm_event(THP_FAULT_FALLBACK); |
71e3aac0 AA |
1367 | goto out; |
1368 | } | |
1369 | ||
bae473a4 | 1370 | if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm, |
46fe940f | 1371 | huge_gfp | __GFP_NORETRY, &memcg, true))) { |
b9bbfbe3 | 1372 | put_page(new_page); |
82b0f8c3 | 1373 | split_huge_pmd(vma, vmf->pmd, vmf->address); |
bae473a4 | 1374 | if (page) |
ddc58f27 | 1375 | put_page(page); |
9845cbbd | 1376 | ret |= VM_FAULT_FALLBACK; |
17766dde | 1377 | count_vm_event(THP_FAULT_FALLBACK); |
b9bbfbe3 AA |
1378 | goto out; |
1379 | } | |
1380 | ||
17766dde DR |
1381 | count_vm_event(THP_FAULT_ALLOC); |
1382 | ||
eecc1e42 | 1383 | if (!page) |
c79b57e4 | 1384 | clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR); |
93b4796d KS |
1385 | else |
1386 | copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR); | |
71e3aac0 AA |
1387 | __SetPageUptodate(new_page); |
1388 | ||
2ec74c3e SG |
1389 | mmun_start = haddr; |
1390 | mmun_end = haddr + HPAGE_PMD_SIZE; | |
bae473a4 | 1391 | mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end); |
2ec74c3e | 1392 | |
82b0f8c3 | 1393 | spin_lock(vmf->ptl); |
93b4796d | 1394 | if (page) |
ddc58f27 | 1395 | put_page(page); |
82b0f8c3 JK |
1396 | if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { |
1397 | spin_unlock(vmf->ptl); | |
f627c2f5 | 1398 | mem_cgroup_cancel_charge(new_page, memcg, true); |
71e3aac0 | 1399 | put_page(new_page); |
2ec74c3e | 1400 | goto out_mn; |
b9bbfbe3 | 1401 | } else { |
71e3aac0 | 1402 | pmd_t entry; |
3122359a | 1403 | entry = mk_huge_pmd(new_page, vma->vm_page_prot); |
f55e1014 | 1404 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
82b0f8c3 | 1405 | pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd); |
d281ee61 | 1406 | page_add_new_anon_rmap(new_page, vma, haddr, true); |
f627c2f5 | 1407 | mem_cgroup_commit_charge(new_page, memcg, false, true); |
00501b53 | 1408 | lru_cache_add_active_or_unevictable(new_page, vma); |
82b0f8c3 JK |
1409 | set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); |
1410 | update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); | |
eecc1e42 | 1411 | if (!page) { |
bae473a4 | 1412 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); |
97ae1749 | 1413 | } else { |
309381fe | 1414 | VM_BUG_ON_PAGE(!PageHead(page), page); |
d281ee61 | 1415 | page_remove_rmap(page, true); |
93b4796d KS |
1416 | put_page(page); |
1417 | } | |
71e3aac0 AA |
1418 | ret |= VM_FAULT_WRITE; |
1419 | } | |
82b0f8c3 | 1420 | spin_unlock(vmf->ptl); |
2ec74c3e | 1421 | out_mn: |
4645b9fe JG |
1422 | /* |
1423 | * No need to double call mmu_notifier->invalidate_range() callback as | |
1424 | * the above pmdp_huge_clear_flush_notify() did already call it. | |
1425 | */ | |
1426 | mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start, | |
1427 | mmun_end); | |
71e3aac0 AA |
1428 | out: |
1429 | return ret; | |
2ec74c3e | 1430 | out_unlock: |
82b0f8c3 | 1431 | spin_unlock(vmf->ptl); |
2ec74c3e | 1432 | return ret; |
71e3aac0 AA |
1433 | } |
1434 | ||
8310d48b KF |
1435 | /* |
1436 | * FOLL_FORCE can write to even unwritable pmd's, but only | |
1437 | * after we've gone through a COW cycle and they are dirty. | |
1438 | */ | |
1439 | static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags) | |
1440 | { | |
f6f37321 | 1441 | return pmd_write(pmd) || |
8310d48b KF |
1442 | ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd)); |
1443 | } | |
1444 | ||
b676b293 | 1445 | struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, |
71e3aac0 AA |
1446 | unsigned long addr, |
1447 | pmd_t *pmd, | |
1448 | unsigned int flags) | |
1449 | { | |
b676b293 | 1450 | struct mm_struct *mm = vma->vm_mm; |
71e3aac0 AA |
1451 | struct page *page = NULL; |
1452 | ||
c4088ebd | 1453 | assert_spin_locked(pmd_lockptr(mm, pmd)); |
71e3aac0 | 1454 | |
8310d48b | 1455 | if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags)) |
71e3aac0 AA |
1456 | goto out; |
1457 | ||
85facf25 KS |
1458 | /* Avoid dumping huge zero page */ |
1459 | if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) | |
1460 | return ERR_PTR(-EFAULT); | |
1461 | ||
2b4847e7 | 1462 | /* Full NUMA hinting faults to serialise migration in fault paths */ |
8a0516ed | 1463 | if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) |
2b4847e7 MG |
1464 | goto out; |
1465 | ||
71e3aac0 | 1466 | page = pmd_page(*pmd); |
ca120cf6 | 1467 | VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page); |
3565fce3 | 1468 | if (flags & FOLL_TOUCH) |
a8f97366 | 1469 | touch_pmd(vma, addr, pmd, flags); |
de60f5f1 | 1470 | if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { |
e90309c9 KS |
1471 | /* |
1472 | * We don't mlock() pte-mapped THPs. This way we can avoid | |
1473 | * leaking mlocked pages into non-VM_LOCKED VMAs. | |
1474 | * | |
9a73f61b KS |
1475 | * For anon THP: |
1476 | * | |
e90309c9 KS |
1477 | * In most cases the pmd is the only mapping of the page as we |
1478 | * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for | |
1479 | * writable private mappings in populate_vma_page_range(). | |
1480 | * | |
1481 | * The only scenario when we have the page shared here is if we | |
1482 | * mlocking read-only mapping shared over fork(). We skip | |
1483 | * mlocking such pages. | |
9a73f61b KS |
1484 | * |
1485 | * For file THP: | |
1486 | * | |
1487 | * We can expect PageDoubleMap() to be stable under page lock: | |
1488 | * for file pages we set it in page_add_file_rmap(), which | |
1489 | * requires page to be locked. | |
e90309c9 | 1490 | */ |
9a73f61b KS |
1491 | |
1492 | if (PageAnon(page) && compound_mapcount(page) != 1) | |
1493 | goto skip_mlock; | |
1494 | if (PageDoubleMap(page) || !page->mapping) | |
1495 | goto skip_mlock; | |
1496 | if (!trylock_page(page)) | |
1497 | goto skip_mlock; | |
1498 | lru_add_drain(); | |
1499 | if (page->mapping && !PageDoubleMap(page)) | |
1500 | mlock_vma_page(page); | |
1501 | unlock_page(page); | |
b676b293 | 1502 | } |
9a73f61b | 1503 | skip_mlock: |
71e3aac0 | 1504 | page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; |
ca120cf6 | 1505 | VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page); |
71e3aac0 | 1506 | if (flags & FOLL_GET) |
ddc58f27 | 1507 | get_page(page); |
71e3aac0 AA |
1508 | |
1509 | out: | |
1510 | return page; | |
1511 | } | |
1512 | ||
d10e63f2 | 1513 | /* NUMA hinting page fault entry point for trans huge pmds */ |
82b0f8c3 | 1514 | int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd) |
d10e63f2 | 1515 | { |
82b0f8c3 | 1516 | struct vm_area_struct *vma = vmf->vma; |
b8916634 | 1517 | struct anon_vma *anon_vma = NULL; |
b32967ff | 1518 | struct page *page; |
82b0f8c3 | 1519 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; |
8191acbd | 1520 | int page_nid = -1, this_nid = numa_node_id(); |
90572890 | 1521 | int target_nid, last_cpupid = -1; |
8191acbd MG |
1522 | bool page_locked; |
1523 | bool migrated = false; | |
b191f9b1 | 1524 | bool was_writable; |
6688cc05 | 1525 | int flags = 0; |
d10e63f2 | 1526 | |
82b0f8c3 JK |
1527 | vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); |
1528 | if (unlikely(!pmd_same(pmd, *vmf->pmd))) | |
d10e63f2 MG |
1529 | goto out_unlock; |
1530 | ||
de466bd6 MG |
1531 | /* |
1532 | * If there are potential migrations, wait for completion and retry | |
1533 | * without disrupting NUMA hinting information. Do not relock and | |
1534 | * check_same as the page may no longer be mapped. | |
1535 | */ | |
82b0f8c3 JK |
1536 | if (unlikely(pmd_trans_migrating(*vmf->pmd))) { |
1537 | page = pmd_page(*vmf->pmd); | |
3c226c63 MR |
1538 | if (!get_page_unless_zero(page)) |
1539 | goto out_unlock; | |
82b0f8c3 | 1540 | spin_unlock(vmf->ptl); |
5d833062 | 1541 | wait_on_page_locked(page); |
3c226c63 | 1542 | put_page(page); |
de466bd6 MG |
1543 | goto out; |
1544 | } | |
1545 | ||
d10e63f2 | 1546 | page = pmd_page(pmd); |
a1a46184 | 1547 | BUG_ON(is_huge_zero_page(page)); |
8191acbd | 1548 | page_nid = page_to_nid(page); |
90572890 | 1549 | last_cpupid = page_cpupid_last(page); |
03c5a6e1 | 1550 | count_vm_numa_event(NUMA_HINT_FAULTS); |
04bb2f94 | 1551 | if (page_nid == this_nid) { |
03c5a6e1 | 1552 | count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); |
04bb2f94 RR |
1553 | flags |= TNF_FAULT_LOCAL; |
1554 | } | |
4daae3b4 | 1555 | |
bea66fbd | 1556 | /* See similar comment in do_numa_page for explanation */ |
288bc549 | 1557 | if (!pmd_savedwrite(pmd)) |
6688cc05 PZ |
1558 | flags |= TNF_NO_GROUP; |
1559 | ||
ff9042b1 MG |
1560 | /* |
1561 | * Acquire the page lock to serialise THP migrations but avoid dropping | |
1562 | * page_table_lock if at all possible | |
1563 | */ | |
b8916634 MG |
1564 | page_locked = trylock_page(page); |
1565 | target_nid = mpol_misplaced(page, vma, haddr); | |
1566 | if (target_nid == -1) { | |
1567 | /* If the page was locked, there are no parallel migrations */ | |
a54a407f | 1568 | if (page_locked) |
b8916634 | 1569 | goto clear_pmdnuma; |
2b4847e7 | 1570 | } |
4daae3b4 | 1571 | |
de466bd6 | 1572 | /* Migration could have started since the pmd_trans_migrating check */ |
2b4847e7 | 1573 | if (!page_locked) { |
3c226c63 MR |
1574 | page_nid = -1; |
1575 | if (!get_page_unless_zero(page)) | |
1576 | goto out_unlock; | |
82b0f8c3 | 1577 | spin_unlock(vmf->ptl); |
b8916634 | 1578 | wait_on_page_locked(page); |
3c226c63 | 1579 | put_page(page); |
b8916634 MG |
1580 | goto out; |
1581 | } | |
1582 | ||
2b4847e7 MG |
1583 | /* |
1584 | * Page is misplaced. Page lock serialises migrations. Acquire anon_vma | |
1585 | * to serialises splits | |
1586 | */ | |
b8916634 | 1587 | get_page(page); |
82b0f8c3 | 1588 | spin_unlock(vmf->ptl); |
b8916634 | 1589 | anon_vma = page_lock_anon_vma_read(page); |
4daae3b4 | 1590 | |
c69307d5 | 1591 | /* Confirm the PMD did not change while page_table_lock was released */ |
82b0f8c3 JK |
1592 | spin_lock(vmf->ptl); |
1593 | if (unlikely(!pmd_same(pmd, *vmf->pmd))) { | |
b32967ff MG |
1594 | unlock_page(page); |
1595 | put_page(page); | |
a54a407f | 1596 | page_nid = -1; |
4daae3b4 | 1597 | goto out_unlock; |
b32967ff | 1598 | } |
ff9042b1 | 1599 | |
c3a489ca MG |
1600 | /* Bail if we fail to protect against THP splits for any reason */ |
1601 | if (unlikely(!anon_vma)) { | |
1602 | put_page(page); | |
1603 | page_nid = -1; | |
1604 | goto clear_pmdnuma; | |
1605 | } | |
1606 | ||
8b1b436d PZ |
1607 | /* |
1608 | * Since we took the NUMA fault, we must have observed the !accessible | |
1609 | * bit. Make sure all other CPUs agree with that, to avoid them | |
1610 | * modifying the page we're about to migrate. | |
1611 | * | |
1612 | * Must be done under PTL such that we'll observe the relevant | |
ccde85ba PZ |
1613 | * inc_tlb_flush_pending(). |
1614 | * | |
1615 | * We are not sure a pending tlb flush here is for a huge page | |
1616 | * mapping or not. Hence use the tlb range variant | |
8b1b436d PZ |
1617 | */ |
1618 | if (mm_tlb_flush_pending(vma->vm_mm)) | |
ccde85ba | 1619 | flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE); |
8b1b436d | 1620 | |
a54a407f MG |
1621 | /* |
1622 | * Migrate the THP to the requested node, returns with page unlocked | |
8a0516ed | 1623 | * and access rights restored. |
a54a407f | 1624 | */ |
82b0f8c3 | 1625 | spin_unlock(vmf->ptl); |
8b1b436d | 1626 | |
bae473a4 | 1627 | migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma, |
82b0f8c3 | 1628 | vmf->pmd, pmd, vmf->address, page, target_nid); |
6688cc05 PZ |
1629 | if (migrated) { |
1630 | flags |= TNF_MIGRATED; | |
8191acbd | 1631 | page_nid = target_nid; |
074c2381 MG |
1632 | } else |
1633 | flags |= TNF_MIGRATE_FAIL; | |
b32967ff | 1634 | |
8191acbd | 1635 | goto out; |
b32967ff | 1636 | clear_pmdnuma: |
a54a407f | 1637 | BUG_ON(!PageLocked(page)); |
288bc549 | 1638 | was_writable = pmd_savedwrite(pmd); |
4d942466 | 1639 | pmd = pmd_modify(pmd, vma->vm_page_prot); |
b7b04004 | 1640 | pmd = pmd_mkyoung(pmd); |
b191f9b1 MG |
1641 | if (was_writable) |
1642 | pmd = pmd_mkwrite(pmd); | |
82b0f8c3 JK |
1643 | set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd); |
1644 | update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); | |
a54a407f | 1645 | unlock_page(page); |
d10e63f2 | 1646 | out_unlock: |
82b0f8c3 | 1647 | spin_unlock(vmf->ptl); |
b8916634 MG |
1648 | |
1649 | out: | |
1650 | if (anon_vma) | |
1651 | page_unlock_anon_vma_read(anon_vma); | |
1652 | ||
8191acbd | 1653 | if (page_nid != -1) |
82b0f8c3 | 1654 | task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, |
9a8b300f | 1655 | flags); |
8191acbd | 1656 | |
d10e63f2 MG |
1657 | return 0; |
1658 | } | |
1659 | ||
319904ad HY |
1660 | /* |
1661 | * Return true if we do MADV_FREE successfully on entire pmd page. | |
1662 | * Otherwise, return false. | |
1663 | */ | |
1664 | bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, | |
b8d3c4c3 | 1665 | pmd_t *pmd, unsigned long addr, unsigned long next) |
b8d3c4c3 MK |
1666 | { |
1667 | spinlock_t *ptl; | |
1668 | pmd_t orig_pmd; | |
1669 | struct page *page; | |
1670 | struct mm_struct *mm = tlb->mm; | |
319904ad | 1671 | bool ret = false; |
b8d3c4c3 | 1672 | |
07e32661 AK |
1673 | tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE); |
1674 | ||
b6ec57f4 KS |
1675 | ptl = pmd_trans_huge_lock(pmd, vma); |
1676 | if (!ptl) | |
25eedabe | 1677 | goto out_unlocked; |
b8d3c4c3 MK |
1678 | |
1679 | orig_pmd = *pmd; | |
319904ad | 1680 | if (is_huge_zero_pmd(orig_pmd)) |
b8d3c4c3 | 1681 | goto out; |
b8d3c4c3 | 1682 | |
84c3fc4e ZY |
1683 | if (unlikely(!pmd_present(orig_pmd))) { |
1684 | VM_BUG_ON(thp_migration_supported() && | |
1685 | !is_pmd_migration_entry(orig_pmd)); | |
1686 | goto out; | |
1687 | } | |
1688 | ||
b8d3c4c3 MK |
1689 | page = pmd_page(orig_pmd); |
1690 | /* | |
1691 | * If other processes are mapping this page, we couldn't discard | |
1692 | * the page unless they all do MADV_FREE so let's skip the page. | |
1693 | */ | |
1694 | if (page_mapcount(page) != 1) | |
1695 | goto out; | |
1696 | ||
1697 | if (!trylock_page(page)) | |
1698 | goto out; | |
1699 | ||
1700 | /* | |
1701 | * If user want to discard part-pages of THP, split it so MADV_FREE | |
1702 | * will deactivate only them. | |
1703 | */ | |
1704 | if (next - addr != HPAGE_PMD_SIZE) { | |
1705 | get_page(page); | |
1706 | spin_unlock(ptl); | |
9818b8cd | 1707 | split_huge_page(page); |
b8d3c4c3 | 1708 | unlock_page(page); |
bbf29ffc | 1709 | put_page(page); |
b8d3c4c3 MK |
1710 | goto out_unlocked; |
1711 | } | |
1712 | ||
1713 | if (PageDirty(page)) | |
1714 | ClearPageDirty(page); | |
1715 | unlock_page(page); | |
1716 | ||
b8d3c4c3 | 1717 | if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { |
58ceeb6b | 1718 | pmdp_invalidate(vma, addr, pmd); |
b8d3c4c3 MK |
1719 | orig_pmd = pmd_mkold(orig_pmd); |
1720 | orig_pmd = pmd_mkclean(orig_pmd); | |
1721 | ||
1722 | set_pmd_at(mm, addr, pmd, orig_pmd); | |
1723 | tlb_remove_pmd_tlb_entry(tlb, pmd, addr); | |
1724 | } | |
802a3a92 SL |
1725 | |
1726 | mark_page_lazyfree(page); | |
319904ad | 1727 | ret = true; |
b8d3c4c3 MK |
1728 | out: |
1729 | spin_unlock(ptl); | |
1730 | out_unlocked: | |
1731 | return ret; | |
1732 | } | |
1733 | ||
953c66c2 AK |
1734 | static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd) |
1735 | { | |
1736 | pgtable_t pgtable; | |
1737 | ||
1738 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); | |
1739 | pte_free(mm, pgtable); | |
c4812909 | 1740 | mm_dec_nr_ptes(mm); |
953c66c2 AK |
1741 | } |
1742 | ||
71e3aac0 | 1743 | int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, |
f21760b1 | 1744 | pmd_t *pmd, unsigned long addr) |
71e3aac0 | 1745 | { |
da146769 | 1746 | pmd_t orig_pmd; |
bf929152 | 1747 | spinlock_t *ptl; |
71e3aac0 | 1748 | |
07e32661 AK |
1749 | tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE); |
1750 | ||
b6ec57f4 KS |
1751 | ptl = __pmd_trans_huge_lock(pmd, vma); |
1752 | if (!ptl) | |
da146769 KS |
1753 | return 0; |
1754 | /* | |
1755 | * For architectures like ppc64 we look at deposited pgtable | |
1756 | * when calling pmdp_huge_get_and_clear. So do the | |
1757 | * pgtable_trans_huge_withdraw after finishing pmdp related | |
1758 | * operations. | |
1759 | */ | |
1760 | orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd, | |
1761 | tlb->fullmm); | |
1762 | tlb_remove_pmd_tlb_entry(tlb, pmd, addr); | |
1763 | if (vma_is_dax(vma)) { | |
3b6521f5 OH |
1764 | if (arch_needs_pgtable_deposit()) |
1765 | zap_deposited_table(tlb->mm, pmd); | |
da146769 KS |
1766 | spin_unlock(ptl); |
1767 | if (is_huge_zero_pmd(orig_pmd)) | |
c0f2e176 | 1768 | tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); |
da146769 | 1769 | } else if (is_huge_zero_pmd(orig_pmd)) { |
c14a6eb4 | 1770 | zap_deposited_table(tlb->mm, pmd); |
da146769 | 1771 | spin_unlock(ptl); |
c0f2e176 | 1772 | tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); |
da146769 | 1773 | } else { |
616b8371 ZY |
1774 | struct page *page = NULL; |
1775 | int flush_needed = 1; | |
1776 | ||
1777 | if (pmd_present(orig_pmd)) { | |
1778 | page = pmd_page(orig_pmd); | |
1779 | page_remove_rmap(page, true); | |
1780 | VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); | |
1781 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
1782 | } else if (thp_migration_supported()) { | |
1783 | swp_entry_t entry; | |
1784 | ||
1785 | VM_BUG_ON(!is_pmd_migration_entry(orig_pmd)); | |
1786 | entry = pmd_to_swp_entry(orig_pmd); | |
1787 | page = pfn_to_page(swp_offset(entry)); | |
1788 | flush_needed = 0; | |
1789 | } else | |
1790 | WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!"); | |
1791 | ||
b5072380 | 1792 | if (PageAnon(page)) { |
c14a6eb4 | 1793 | zap_deposited_table(tlb->mm, pmd); |
b5072380 KS |
1794 | add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); |
1795 | } else { | |
953c66c2 AK |
1796 | if (arch_needs_pgtable_deposit()) |
1797 | zap_deposited_table(tlb->mm, pmd); | |
b5072380 KS |
1798 | add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR); |
1799 | } | |
616b8371 | 1800 | |
da146769 | 1801 | spin_unlock(ptl); |
616b8371 ZY |
1802 | if (flush_needed) |
1803 | tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE); | |
025c5b24 | 1804 | } |
da146769 | 1805 | return 1; |
71e3aac0 AA |
1806 | } |
1807 | ||
1dd38b6c AK |
1808 | #ifndef pmd_move_must_withdraw |
1809 | static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl, | |
1810 | spinlock_t *old_pmd_ptl, | |
1811 | struct vm_area_struct *vma) | |
1812 | { | |
1813 | /* | |
1814 | * With split pmd lock we also need to move preallocated | |
1815 | * PTE page table if new_pmd is on different PMD page table. | |
1816 | * | |
1817 | * We also don't deposit and withdraw tables for file pages. | |
1818 | */ | |
1819 | return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma); | |
1820 | } | |
1821 | #endif | |
1822 | ||
ab6e3d09 NH |
1823 | static pmd_t move_soft_dirty_pmd(pmd_t pmd) |
1824 | { | |
1825 | #ifdef CONFIG_MEM_SOFT_DIRTY | |
1826 | if (unlikely(is_pmd_migration_entry(pmd))) | |
1827 | pmd = pmd_swp_mksoft_dirty(pmd); | |
1828 | else if (pmd_present(pmd)) | |
1829 | pmd = pmd_mksoft_dirty(pmd); | |
1830 | #endif | |
1831 | return pmd; | |
1832 | } | |
1833 | ||
bf8616d5 | 1834 | bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, |
37a1c49a | 1835 | unsigned long new_addr, unsigned long old_end, |
5ad4de30 | 1836 | pmd_t *old_pmd, pmd_t *new_pmd) |
37a1c49a | 1837 | { |
bf929152 | 1838 | spinlock_t *old_ptl, *new_ptl; |
37a1c49a | 1839 | pmd_t pmd; |
37a1c49a | 1840 | struct mm_struct *mm = vma->vm_mm; |
5d190420 | 1841 | bool force_flush = false; |
37a1c49a AA |
1842 | |
1843 | if ((old_addr & ~HPAGE_PMD_MASK) || | |
1844 | (new_addr & ~HPAGE_PMD_MASK) || | |
bf8616d5 | 1845 | old_end - old_addr < HPAGE_PMD_SIZE) |
4b471e88 | 1846 | return false; |
37a1c49a AA |
1847 | |
1848 | /* | |
1849 | * The destination pmd shouldn't be established, free_pgtables() | |
1850 | * should have release it. | |
1851 | */ | |
1852 | if (WARN_ON(!pmd_none(*new_pmd))) { | |
1853 | VM_BUG_ON(pmd_trans_huge(*new_pmd)); | |
4b471e88 | 1854 | return false; |
37a1c49a AA |
1855 | } |
1856 | ||
bf929152 KS |
1857 | /* |
1858 | * We don't have to worry about the ordering of src and dst | |
1859 | * ptlocks because exclusive mmap_sem prevents deadlock. | |
1860 | */ | |
b6ec57f4 KS |
1861 | old_ptl = __pmd_trans_huge_lock(old_pmd, vma); |
1862 | if (old_ptl) { | |
bf929152 KS |
1863 | new_ptl = pmd_lockptr(mm, new_pmd); |
1864 | if (new_ptl != old_ptl) | |
1865 | spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); | |
8809aa2d | 1866 | pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); |
5ad4de30 | 1867 | if (pmd_present(pmd)) |
a2ce2666 | 1868 | force_flush = true; |
025c5b24 | 1869 | VM_BUG_ON(!pmd_none(*new_pmd)); |
3592806c | 1870 | |
1dd38b6c | 1871 | if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) { |
b3084f4d | 1872 | pgtable_t pgtable; |
3592806c KS |
1873 | pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); |
1874 | pgtable_trans_huge_deposit(mm, new_pmd, pgtable); | |
3592806c | 1875 | } |
ab6e3d09 NH |
1876 | pmd = move_soft_dirty_pmd(pmd); |
1877 | set_pmd_at(mm, new_addr, new_pmd, pmd); | |
5d190420 AL |
1878 | if (force_flush) |
1879 | flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE); | |
5ad4de30 LT |
1880 | if (new_ptl != old_ptl) |
1881 | spin_unlock(new_ptl); | |
bf929152 | 1882 | spin_unlock(old_ptl); |
4b471e88 | 1883 | return true; |
37a1c49a | 1884 | } |
4b471e88 | 1885 | return false; |
37a1c49a AA |
1886 | } |
1887 | ||
f123d74a MG |
1888 | /* |
1889 | * Returns | |
1890 | * - 0 if PMD could not be locked | |
1891 | * - 1 if PMD was locked but protections unchange and TLB flush unnecessary | |
1892 | * - HPAGE_PMD_NR is protections changed and TLB flush necessary | |
1893 | */ | |
cd7548ab | 1894 | int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, |
e944fd67 | 1895 | unsigned long addr, pgprot_t newprot, int prot_numa) |
cd7548ab JW |
1896 | { |
1897 | struct mm_struct *mm = vma->vm_mm; | |
bf929152 | 1898 | spinlock_t *ptl; |
0a85e51d KS |
1899 | pmd_t entry; |
1900 | bool preserve_write; | |
1901 | int ret; | |
cd7548ab | 1902 | |
b6ec57f4 | 1903 | ptl = __pmd_trans_huge_lock(pmd, vma); |
0a85e51d KS |
1904 | if (!ptl) |
1905 | return 0; | |
e944fd67 | 1906 | |
0a85e51d KS |
1907 | preserve_write = prot_numa && pmd_write(*pmd); |
1908 | ret = 1; | |
e944fd67 | 1909 | |
84c3fc4e ZY |
1910 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
1911 | if (is_swap_pmd(*pmd)) { | |
1912 | swp_entry_t entry = pmd_to_swp_entry(*pmd); | |
1913 | ||
1914 | VM_BUG_ON(!is_pmd_migration_entry(*pmd)); | |
1915 | if (is_write_migration_entry(entry)) { | |
1916 | pmd_t newpmd; | |
1917 | /* | |
1918 | * A protection check is difficult so | |
1919 | * just be safe and disable write | |
1920 | */ | |
1921 | make_migration_entry_read(&entry); | |
1922 | newpmd = swp_entry_to_pmd(entry); | |
ab6e3d09 NH |
1923 | if (pmd_swp_soft_dirty(*pmd)) |
1924 | newpmd = pmd_swp_mksoft_dirty(newpmd); | |
84c3fc4e ZY |
1925 | set_pmd_at(mm, addr, pmd, newpmd); |
1926 | } | |
1927 | goto unlock; | |
1928 | } | |
1929 | #endif | |
1930 | ||
0a85e51d KS |
1931 | /* |
1932 | * Avoid trapping faults against the zero page. The read-only | |
1933 | * data is likely to be read-cached on the local CPU and | |
1934 | * local/remote hits to the zero page are not interesting. | |
1935 | */ | |
1936 | if (prot_numa && is_huge_zero_pmd(*pmd)) | |
1937 | goto unlock; | |
025c5b24 | 1938 | |
0a85e51d KS |
1939 | if (prot_numa && pmd_protnone(*pmd)) |
1940 | goto unlock; | |
1941 | ||
ced10803 KS |
1942 | /* |
1943 | * In case prot_numa, we are under down_read(mmap_sem). It's critical | |
1944 | * to not clear pmd intermittently to avoid race with MADV_DONTNEED | |
1945 | * which is also under down_read(mmap_sem): | |
1946 | * | |
1947 | * CPU0: CPU1: | |
1948 | * change_huge_pmd(prot_numa=1) | |
1949 | * pmdp_huge_get_and_clear_notify() | |
1950 | * madvise_dontneed() | |
1951 | * zap_pmd_range() | |
1952 | * pmd_trans_huge(*pmd) == 0 (without ptl) | |
1953 | * // skip the pmd | |
1954 | * set_pmd_at(); | |
1955 | * // pmd is re-established | |
1956 | * | |
1957 | * The race makes MADV_DONTNEED miss the huge pmd and don't clear it | |
1958 | * which may break userspace. | |
1959 | * | |
1960 | * pmdp_invalidate() is required to make sure we don't miss | |
1961 | * dirty/young flags set by hardware. | |
1962 | */ | |
1963 | entry = *pmd; | |
1964 | pmdp_invalidate(vma, addr, pmd); | |
1965 | ||
1966 | /* | |
1967 | * Recover dirty/young flags. It relies on pmdp_invalidate to not | |
1968 | * corrupt them. | |
1969 | */ | |
1970 | if (pmd_dirty(*pmd)) | |
1971 | entry = pmd_mkdirty(entry); | |
1972 | if (pmd_young(*pmd)) | |
1973 | entry = pmd_mkyoung(entry); | |
1974 | ||
0a85e51d KS |
1975 | entry = pmd_modify(entry, newprot); |
1976 | if (preserve_write) | |
1977 | entry = pmd_mk_savedwrite(entry); | |
1978 | ret = HPAGE_PMD_NR; | |
1979 | set_pmd_at(mm, addr, pmd, entry); | |
1980 | BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry)); | |
1981 | unlock: | |
1982 | spin_unlock(ptl); | |
025c5b24 NH |
1983 | return ret; |
1984 | } | |
1985 | ||
1986 | /* | |
8f19b0c0 | 1987 | * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise. |
025c5b24 | 1988 | * |
8f19b0c0 HY |
1989 | * Note that if it returns page table lock pointer, this routine returns without |
1990 | * unlocking page table lock. So callers must unlock it. | |
025c5b24 | 1991 | */ |
b6ec57f4 | 1992 | spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) |
025c5b24 | 1993 | { |
b6ec57f4 KS |
1994 | spinlock_t *ptl; |
1995 | ptl = pmd_lock(vma->vm_mm, pmd); | |
84c3fc4e ZY |
1996 | if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || |
1997 | pmd_devmap(*pmd))) | |
b6ec57f4 KS |
1998 | return ptl; |
1999 | spin_unlock(ptl); | |
2000 | return NULL; | |
cd7548ab JW |
2001 | } |
2002 | ||
a00cc7d9 MW |
2003 | /* |
2004 | * Returns true if a given pud maps a thp, false otherwise. | |
2005 | * | |
2006 | * Note that if it returns true, this routine returns without unlocking page | |
2007 | * table lock. So callers must unlock it. | |
2008 | */ | |
2009 | spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma) | |
2010 | { | |
2011 | spinlock_t *ptl; | |
2012 | ||
2013 | ptl = pud_lock(vma->vm_mm, pud); | |
2014 | if (likely(pud_trans_huge(*pud) || pud_devmap(*pud))) | |
2015 | return ptl; | |
2016 | spin_unlock(ptl); | |
2017 | return NULL; | |
2018 | } | |
2019 | ||
2020 | #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD | |
2021 | int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, | |
2022 | pud_t *pud, unsigned long addr) | |
2023 | { | |
2024 | pud_t orig_pud; | |
2025 | spinlock_t *ptl; | |
2026 | ||
2027 | ptl = __pud_trans_huge_lock(pud, vma); | |
2028 | if (!ptl) | |
2029 | return 0; | |
2030 | /* | |
2031 | * For architectures like ppc64 we look at deposited pgtable | |
2032 | * when calling pudp_huge_get_and_clear. So do the | |
2033 | * pgtable_trans_huge_withdraw after finishing pudp related | |
2034 | * operations. | |
2035 | */ | |
2036 | orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud, | |
2037 | tlb->fullmm); | |
2038 | tlb_remove_pud_tlb_entry(tlb, pud, addr); | |
2039 | if (vma_is_dax(vma)) { | |
2040 | spin_unlock(ptl); | |
2041 | /* No zero page support yet */ | |
2042 | } else { | |
2043 | /* No support for anonymous PUD pages yet */ | |
2044 | BUG(); | |
2045 | } | |
2046 | return 1; | |
2047 | } | |
2048 | ||
2049 | static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud, | |
2050 | unsigned long haddr) | |
2051 | { | |
2052 | VM_BUG_ON(haddr & ~HPAGE_PUD_MASK); | |
2053 | VM_BUG_ON_VMA(vma->vm_start > haddr, vma); | |
2054 | VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma); | |
2055 | VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud)); | |
2056 | ||
ce9311cf | 2057 | count_vm_event(THP_SPLIT_PUD); |
a00cc7d9 MW |
2058 | |
2059 | pudp_huge_clear_flush_notify(vma, haddr, pud); | |
2060 | } | |
2061 | ||
2062 | void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud, | |
2063 | unsigned long address) | |
2064 | { | |
2065 | spinlock_t *ptl; | |
2066 | struct mm_struct *mm = vma->vm_mm; | |
2067 | unsigned long haddr = address & HPAGE_PUD_MASK; | |
2068 | ||
2069 | mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE); | |
2070 | ptl = pud_lock(mm, pud); | |
2071 | if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud))) | |
2072 | goto out; | |
2073 | __split_huge_pud_locked(vma, pud, haddr); | |
2074 | ||
2075 | out: | |
2076 | spin_unlock(ptl); | |
4645b9fe JG |
2077 | /* |
2078 | * No need to double call mmu_notifier->invalidate_range() callback as | |
2079 | * the above pudp_huge_clear_flush_notify() did already call it. | |
2080 | */ | |
2081 | mmu_notifier_invalidate_range_only_end(mm, haddr, haddr + | |
2082 | HPAGE_PUD_SIZE); | |
a00cc7d9 MW |
2083 | } |
2084 | #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ | |
2085 | ||
eef1b3ba KS |
2086 | static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, |
2087 | unsigned long haddr, pmd_t *pmd) | |
2088 | { | |
2089 | struct mm_struct *mm = vma->vm_mm; | |
2090 | pgtable_t pgtable; | |
2091 | pmd_t _pmd; | |
2092 | int i; | |
2093 | ||
0f10851e JG |
2094 | /* |
2095 | * Leave pmd empty until pte is filled note that it is fine to delay | |
2096 | * notification until mmu_notifier_invalidate_range_end() as we are | |
2097 | * replacing a zero pmd write protected page with a zero pte write | |
2098 | * protected page. | |
2099 | * | |
2100 | * See Documentation/vm/mmu_notifier.txt | |
2101 | */ | |
2102 | pmdp_huge_clear_flush(vma, haddr, pmd); | |
eef1b3ba KS |
2103 | |
2104 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); | |
2105 | pmd_populate(mm, &_pmd, pgtable); | |
2106 | ||
2107 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
2108 | pte_t *pte, entry; | |
2109 | entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); | |
2110 | entry = pte_mkspecial(entry); | |
2111 | pte = pte_offset_map(&_pmd, haddr); | |
2112 | VM_BUG_ON(!pte_none(*pte)); | |
2113 | set_pte_at(mm, haddr, pte, entry); | |
2114 | pte_unmap(pte); | |
2115 | } | |
2116 | smp_wmb(); /* make pte visible before pmd */ | |
2117 | pmd_populate(mm, pmd, pgtable); | |
eef1b3ba KS |
2118 | } |
2119 | ||
2120 | static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, | |
ba988280 | 2121 | unsigned long haddr, bool freeze) |
eef1b3ba KS |
2122 | { |
2123 | struct mm_struct *mm = vma->vm_mm; | |
2124 | struct page *page; | |
2125 | pgtable_t pgtable; | |
2126 | pmd_t _pmd; | |
84c3fc4e | 2127 | bool young, write, dirty, soft_dirty, pmd_migration = false; |
2ac015e2 | 2128 | unsigned long addr; |
eef1b3ba KS |
2129 | int i; |
2130 | ||
2131 | VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); | |
2132 | VM_BUG_ON_VMA(vma->vm_start > haddr, vma); | |
2133 | VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); | |
84c3fc4e ZY |
2134 | VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd) |
2135 | && !pmd_devmap(*pmd)); | |
eef1b3ba KS |
2136 | |
2137 | count_vm_event(THP_SPLIT_PMD); | |
2138 | ||
d21b9e57 KS |
2139 | if (!vma_is_anonymous(vma)) { |
2140 | _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); | |
953c66c2 AK |
2141 | /* |
2142 | * We are going to unmap this huge page. So | |
2143 | * just go ahead and zap it | |
2144 | */ | |
2145 | if (arch_needs_pgtable_deposit()) | |
2146 | zap_deposited_table(mm, pmd); | |
d21b9e57 KS |
2147 | if (vma_is_dax(vma)) |
2148 | return; | |
2149 | page = pmd_page(_pmd); | |
30820520 HD |
2150 | if (!PageDirty(page) && pmd_dirty(_pmd)) |
2151 | set_page_dirty(page); | |
d21b9e57 KS |
2152 | if (!PageReferenced(page) && pmd_young(_pmd)) |
2153 | SetPageReferenced(page); | |
2154 | page_remove_rmap(page, true); | |
2155 | put_page(page); | |
2156 | add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR); | |
eef1b3ba KS |
2157 | return; |
2158 | } else if (is_huge_zero_pmd(*pmd)) { | |
4645b9fe JG |
2159 | /* |
2160 | * FIXME: Do we want to invalidate secondary mmu by calling | |
2161 | * mmu_notifier_invalidate_range() see comments below inside | |
2162 | * __split_huge_pmd() ? | |
2163 | * | |
2164 | * We are going from a zero huge page write protected to zero | |
2165 | * small page also write protected so it does not seems useful | |
2166 | * to invalidate secondary mmu at this time. | |
2167 | */ | |
eef1b3ba KS |
2168 | return __split_huge_zero_page_pmd(vma, haddr, pmd); |
2169 | } | |
2170 | ||
84c3fc4e ZY |
2171 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
2172 | pmd_migration = is_pmd_migration_entry(*pmd); | |
2173 | if (pmd_migration) { | |
2174 | swp_entry_t entry; | |
2175 | ||
2176 | entry = pmd_to_swp_entry(*pmd); | |
2177 | page = pfn_to_page(swp_offset(entry)); | |
2178 | } else | |
2179 | #endif | |
2180 | page = pmd_page(*pmd); | |
eef1b3ba | 2181 | VM_BUG_ON_PAGE(!page_count(page), page); |
fe896d18 | 2182 | page_ref_add(page, HPAGE_PMD_NR - 1); |
eef1b3ba KS |
2183 | write = pmd_write(*pmd); |
2184 | young = pmd_young(*pmd); | |
b8d3c4c3 | 2185 | dirty = pmd_dirty(*pmd); |
804dd150 | 2186 | soft_dirty = pmd_soft_dirty(*pmd); |
eef1b3ba | 2187 | |
c777e2a8 | 2188 | pmdp_huge_split_prepare(vma, haddr, pmd); |
eef1b3ba KS |
2189 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); |
2190 | pmd_populate(mm, &_pmd, pgtable); | |
2191 | ||
2ac015e2 | 2192 | for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { |
eef1b3ba KS |
2193 | pte_t entry, *pte; |
2194 | /* | |
2195 | * Note that NUMA hinting access restrictions are not | |
2196 | * transferred to avoid any possibility of altering | |
2197 | * permissions across VMAs. | |
2198 | */ | |
84c3fc4e | 2199 | if (freeze || pmd_migration) { |
ba988280 KS |
2200 | swp_entry_t swp_entry; |
2201 | swp_entry = make_migration_entry(page + i, write); | |
2202 | entry = swp_entry_to_pte(swp_entry); | |
804dd150 AA |
2203 | if (soft_dirty) |
2204 | entry = pte_swp_mksoft_dirty(entry); | |
ba988280 | 2205 | } else { |
6d2329f8 | 2206 | entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot)); |
b8d3c4c3 | 2207 | entry = maybe_mkwrite(entry, vma); |
ba988280 KS |
2208 | if (!write) |
2209 | entry = pte_wrprotect(entry); | |
2210 | if (!young) | |
2211 | entry = pte_mkold(entry); | |
804dd150 AA |
2212 | if (soft_dirty) |
2213 | entry = pte_mksoft_dirty(entry); | |
ba988280 | 2214 | } |
b8d3c4c3 MK |
2215 | if (dirty) |
2216 | SetPageDirty(page + i); | |
2ac015e2 | 2217 | pte = pte_offset_map(&_pmd, addr); |
eef1b3ba | 2218 | BUG_ON(!pte_none(*pte)); |
2ac015e2 | 2219 | set_pte_at(mm, addr, pte, entry); |
eef1b3ba KS |
2220 | atomic_inc(&page[i]._mapcount); |
2221 | pte_unmap(pte); | |
2222 | } | |
2223 | ||
2224 | /* | |
2225 | * Set PG_double_map before dropping compound_mapcount to avoid | |
2226 | * false-negative page_mapped(). | |
2227 | */ | |
2228 | if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) { | |
2229 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
2230 | atomic_inc(&page[i]._mapcount); | |
2231 | } | |
2232 | ||
2233 | if (atomic_add_negative(-1, compound_mapcount_ptr(page))) { | |
2234 | /* Last compound_mapcount is gone. */ | |
11fb9989 | 2235 | __dec_node_page_state(page, NR_ANON_THPS); |
eef1b3ba KS |
2236 | if (TestClearPageDoubleMap(page)) { |
2237 | /* No need in mapcount reference anymore */ | |
2238 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
2239 | atomic_dec(&page[i]._mapcount); | |
2240 | } | |
2241 | } | |
2242 | ||
2243 | smp_wmb(); /* make pte visible before pmd */ | |
e9b61f19 KS |
2244 | /* |
2245 | * Up to this point the pmd is present and huge and userland has the | |
2246 | * whole access to the hugepage during the split (which happens in | |
2247 | * place). If we overwrite the pmd with the not-huge version pointing | |
2248 | * to the pte here (which of course we could if all CPUs were bug | |
2249 | * free), userland could trigger a small page size TLB miss on the | |
2250 | * small sized TLB while the hugepage TLB entry is still established in | |
2251 | * the huge TLB. Some CPU doesn't like that. | |
2252 | * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum | |
2253 | * 383 on page 93. Intel should be safe but is also warns that it's | |
2254 | * only safe if the permission and cache attributes of the two entries | |
2255 | * loaded in the two TLB is identical (which should be the case here). | |
2256 | * But it is generally safer to never allow small and huge TLB entries | |
2257 | * for the same virtual address to be loaded simultaneously. So instead | |
2258 | * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the | |
2259 | * current pmd notpresent (atomically because here the pmd_trans_huge | |
2260 | * and pmd_trans_splitting must remain set at all times on the pmd | |
2261 | * until the split is complete for this pmd), then we flush the SMP TLB | |
2262 | * and finally we write the non-huge version of the pmd entry with | |
2263 | * pmd_populate. | |
2264 | */ | |
2265 | pmdp_invalidate(vma, haddr, pmd); | |
eef1b3ba | 2266 | pmd_populate(mm, pmd, pgtable); |
e9b61f19 KS |
2267 | |
2268 | if (freeze) { | |
2ac015e2 | 2269 | for (i = 0; i < HPAGE_PMD_NR; i++) { |
e9b61f19 KS |
2270 | page_remove_rmap(page + i, false); |
2271 | put_page(page + i); | |
2272 | } | |
2273 | } | |
eef1b3ba KS |
2274 | } |
2275 | ||
2276 | void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, | |
33f4751e | 2277 | unsigned long address, bool freeze, struct page *page) |
eef1b3ba KS |
2278 | { |
2279 | spinlock_t *ptl; | |
2280 | struct mm_struct *mm = vma->vm_mm; | |
2281 | unsigned long haddr = address & HPAGE_PMD_MASK; | |
2282 | ||
2283 | mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE); | |
2284 | ptl = pmd_lock(mm, pmd); | |
33f4751e NH |
2285 | |
2286 | /* | |
2287 | * If caller asks to setup a migration entries, we need a page to check | |
2288 | * pmd against. Otherwise we can end up replacing wrong page. | |
2289 | */ | |
2290 | VM_BUG_ON(freeze && !page); | |
2291 | if (page && page != pmd_page(*pmd)) | |
2292 | goto out; | |
2293 | ||
5c7fb56e | 2294 | if (pmd_trans_huge(*pmd)) { |
33f4751e | 2295 | page = pmd_page(*pmd); |
5c7fb56e | 2296 | if (PageMlocked(page)) |
5f737714 | 2297 | clear_page_mlock(page); |
84c3fc4e | 2298 | } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd))) |
e90309c9 | 2299 | goto out; |
fec89c10 | 2300 | __split_huge_pmd_locked(vma, pmd, haddr, freeze); |
e90309c9 | 2301 | out: |
eef1b3ba | 2302 | spin_unlock(ptl); |
4645b9fe JG |
2303 | /* |
2304 | * No need to double call mmu_notifier->invalidate_range() callback. | |
2305 | * They are 3 cases to consider inside __split_huge_pmd_locked(): | |
2306 | * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious | |
2307 | * 2) __split_huge_zero_page_pmd() read only zero page and any write | |
2308 | * fault will trigger a flush_notify before pointing to a new page | |
2309 | * (it is fine if the secondary mmu keeps pointing to the old zero | |
2310 | * page in the meantime) | |
2311 | * 3) Split a huge pmd into pte pointing to the same page. No need | |
2312 | * to invalidate secondary tlb entry they are all still valid. | |
2313 | * any further changes to individual pte will notify. So no need | |
2314 | * to call mmu_notifier->invalidate_range() | |
2315 | */ | |
2316 | mmu_notifier_invalidate_range_only_end(mm, haddr, haddr + | |
2317 | HPAGE_PMD_SIZE); | |
eef1b3ba KS |
2318 | } |
2319 | ||
fec89c10 KS |
2320 | void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address, |
2321 | bool freeze, struct page *page) | |
94fcc585 | 2322 | { |
f72e7dcd | 2323 | pgd_t *pgd; |
c2febafc | 2324 | p4d_t *p4d; |
f72e7dcd | 2325 | pud_t *pud; |
94fcc585 AA |
2326 | pmd_t *pmd; |
2327 | ||
78ddc534 | 2328 | pgd = pgd_offset(vma->vm_mm, address); |
f72e7dcd HD |
2329 | if (!pgd_present(*pgd)) |
2330 | return; | |
2331 | ||
c2febafc KS |
2332 | p4d = p4d_offset(pgd, address); |
2333 | if (!p4d_present(*p4d)) | |
2334 | return; | |
2335 | ||
2336 | pud = pud_offset(p4d, address); | |
f72e7dcd HD |
2337 | if (!pud_present(*pud)) |
2338 | return; | |
2339 | ||
2340 | pmd = pmd_offset(pud, address); | |
fec89c10 | 2341 | |
33f4751e | 2342 | __split_huge_pmd(vma, pmd, address, freeze, page); |
94fcc585 AA |
2343 | } |
2344 | ||
e1b9996b | 2345 | void vma_adjust_trans_huge(struct vm_area_struct *vma, |
94fcc585 AA |
2346 | unsigned long start, |
2347 | unsigned long end, | |
2348 | long adjust_next) | |
2349 | { | |
2350 | /* | |
2351 | * If the new start address isn't hpage aligned and it could | |
2352 | * previously contain an hugepage: check if we need to split | |
2353 | * an huge pmd. | |
2354 | */ | |
2355 | if (start & ~HPAGE_PMD_MASK && | |
2356 | (start & HPAGE_PMD_MASK) >= vma->vm_start && | |
2357 | (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
fec89c10 | 2358 | split_huge_pmd_address(vma, start, false, NULL); |
94fcc585 AA |
2359 | |
2360 | /* | |
2361 | * If the new end address isn't hpage aligned and it could | |
2362 | * previously contain an hugepage: check if we need to split | |
2363 | * an huge pmd. | |
2364 | */ | |
2365 | if (end & ~HPAGE_PMD_MASK && | |
2366 | (end & HPAGE_PMD_MASK) >= vma->vm_start && | |
2367 | (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
fec89c10 | 2368 | split_huge_pmd_address(vma, end, false, NULL); |
94fcc585 AA |
2369 | |
2370 | /* | |
2371 | * If we're also updating the vma->vm_next->vm_start, if the new | |
2372 | * vm_next->vm_start isn't page aligned and it could previously | |
2373 | * contain an hugepage: check if we need to split an huge pmd. | |
2374 | */ | |
2375 | if (adjust_next > 0) { | |
2376 | struct vm_area_struct *next = vma->vm_next; | |
2377 | unsigned long nstart = next->vm_start; | |
2378 | nstart += adjust_next << PAGE_SHIFT; | |
2379 | if (nstart & ~HPAGE_PMD_MASK && | |
2380 | (nstart & HPAGE_PMD_MASK) >= next->vm_start && | |
2381 | (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) | |
fec89c10 | 2382 | split_huge_pmd_address(next, nstart, false, NULL); |
94fcc585 AA |
2383 | } |
2384 | } | |
e9b61f19 | 2385 | |
d7b66bd2 | 2386 | static void unmap_page(struct page *page) |
e9b61f19 | 2387 | { |
baa355fd | 2388 | enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS | |
c7ab0d2f | 2389 | TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD; |
666e5a40 | 2390 | bool unmap_success; |
e9b61f19 KS |
2391 | |
2392 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
2393 | ||
baa355fd | 2394 | if (PageAnon(page)) |
b5ff8161 | 2395 | ttu_flags |= TTU_SPLIT_FREEZE; |
baa355fd | 2396 | |
666e5a40 MK |
2397 | unmap_success = try_to_unmap(page, ttu_flags); |
2398 | VM_BUG_ON_PAGE(!unmap_success, page); | |
e9b61f19 KS |
2399 | } |
2400 | ||
d7b66bd2 | 2401 | static void remap_page(struct page *page) |
e9b61f19 | 2402 | { |
fec89c10 | 2403 | int i; |
ace71a19 KS |
2404 | if (PageTransHuge(page)) { |
2405 | remove_migration_ptes(page, page, true); | |
2406 | } else { | |
2407 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
2408 | remove_migration_ptes(page + i, page + i, true); | |
2409 | } | |
e9b61f19 KS |
2410 | } |
2411 | ||
8df651c7 | 2412 | static void __split_huge_page_tail(struct page *head, int tail, |
e9b61f19 KS |
2413 | struct lruvec *lruvec, struct list_head *list) |
2414 | { | |
e9b61f19 KS |
2415 | struct page *page_tail = head + tail; |
2416 | ||
8df651c7 | 2417 | VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail); |
e9b61f19 KS |
2418 | |
2419 | /* | |
0eb2c7d7 KK |
2420 | * Clone page flags before unfreezing refcount. |
2421 | * | |
2422 | * After successful get_page_unless_zero() might follow flags change, | |
2423 | * for exmaple lock_page() which set PG_waiters. | |
e9b61f19 | 2424 | */ |
e9b61f19 KS |
2425 | page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; |
2426 | page_tail->flags |= (head->flags & | |
2427 | ((1L << PG_referenced) | | |
2428 | (1L << PG_swapbacked) | | |
38d8b4e6 | 2429 | (1L << PG_swapcache) | |
e9b61f19 KS |
2430 | (1L << PG_mlocked) | |
2431 | (1L << PG_uptodate) | | |
2432 | (1L << PG_active) | | |
2433 | (1L << PG_locked) | | |
b8d3c4c3 MK |
2434 | (1L << PG_unevictable) | |
2435 | (1L << PG_dirty))); | |
e9b61f19 | 2436 | |
7c34f8cd HD |
2437 | /* ->mapping in first tail page is compound_mapcount */ |
2438 | VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING, | |
2439 | page_tail); | |
2440 | page_tail->mapping = head->mapping; | |
2441 | page_tail->index = head->index + tail; | |
2442 | ||
0eb2c7d7 | 2443 | /* Page flags must be visible before we make the page non-compound. */ |
e9b61f19 KS |
2444 | smp_wmb(); |
2445 | ||
0eb2c7d7 KK |
2446 | /* |
2447 | * Clear PageTail before unfreezing page refcount. | |
2448 | * | |
2449 | * After successful get_page_unless_zero() might follow put_page() | |
2450 | * which needs correct compound_head(). | |
2451 | */ | |
e9b61f19 KS |
2452 | clear_compound_head(page_tail); |
2453 | ||
0eb2c7d7 KK |
2454 | /* Finally unfreeze refcount. Additional reference from page cache. */ |
2455 | page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) || | |
2456 | PageSwapCache(head))); | |
2457 | ||
e9b61f19 KS |
2458 | if (page_is_young(head)) |
2459 | set_page_young(page_tail); | |
2460 | if (page_is_idle(head)) | |
2461 | set_page_idle(page_tail); | |
2462 | ||
e9b61f19 KS |
2463 | page_cpupid_xchg_last(page_tail, page_cpupid_last(head)); |
2464 | lru_add_page_tail(head, page_tail, lruvec, list); | |
e9b61f19 KS |
2465 | } |
2466 | ||
baa355fd | 2467 | static void __split_huge_page(struct page *page, struct list_head *list, |
5ba09cb7 | 2468 | pgoff_t end, unsigned long flags) |
e9b61f19 KS |
2469 | { |
2470 | struct page *head = compound_head(page); | |
2471 | struct zone *zone = page_zone(head); | |
2472 | struct lruvec *lruvec; | |
8df651c7 | 2473 | int i; |
e9b61f19 | 2474 | |
599d0c95 | 2475 | lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat); |
e9b61f19 KS |
2476 | |
2477 | /* complete memcg works before add pages to LRU */ | |
2478 | mem_cgroup_split_huge_fixup(head); | |
2479 | ||
baa355fd | 2480 | for (i = HPAGE_PMD_NR - 1; i >= 1; i--) { |
8df651c7 | 2481 | __split_huge_page_tail(head, i, lruvec, list); |
baa355fd KS |
2482 | /* Some pages can be beyond i_size: drop them from page cache */ |
2483 | if (head[i].index >= end) { | |
8759535b | 2484 | ClearPageDirty(head + i); |
baa355fd | 2485 | __delete_from_page_cache(head + i, NULL); |
800d8c63 KS |
2486 | if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head)) |
2487 | shmem_uncharge(head->mapping->host, 1); | |
baa355fd KS |
2488 | put_page(head + i); |
2489 | } | |
2490 | } | |
e9b61f19 KS |
2491 | |
2492 | ClearPageCompound(head); | |
be765d99 VB |
2493 | |
2494 | split_page_owner(head, HPAGE_PMD_ORDER); | |
2495 | ||
baa355fd KS |
2496 | /* See comment in __split_huge_page_tail() */ |
2497 | if (PageAnon(head)) { | |
38d8b4e6 HY |
2498 | /* Additional pin to radix tree of swap cache */ |
2499 | if (PageSwapCache(head)) | |
2500 | page_ref_add(head, 2); | |
2501 | else | |
2502 | page_ref_inc(head); | |
baa355fd KS |
2503 | } else { |
2504 | /* Additional pin to radix tree */ | |
2505 | page_ref_add(head, 2); | |
2506 | spin_unlock(&head->mapping->tree_lock); | |
2507 | } | |
2508 | ||
a52633d8 | 2509 | spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags); |
e9b61f19 | 2510 | |
d7b66bd2 | 2511 | remap_page(head); |
e9b61f19 KS |
2512 | |
2513 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
2514 | struct page *subpage = head + i; | |
2515 | if (subpage == page) | |
2516 | continue; | |
2517 | unlock_page(subpage); | |
2518 | ||
2519 | /* | |
2520 | * Subpages may be freed if there wasn't any mapping | |
2521 | * like if add_to_swap() is running on a lru page that | |
2522 | * had its mapping zapped. And freeing these pages | |
2523 | * requires taking the lru_lock so we do the put_page | |
2524 | * of the tail pages after the split is complete. | |
2525 | */ | |
2526 | put_page(subpage); | |
2527 | } | |
2528 | } | |
2529 | ||
b20ce5e0 KS |
2530 | int total_mapcount(struct page *page) |
2531 | { | |
dd78fedd | 2532 | int i, compound, ret; |
b20ce5e0 KS |
2533 | |
2534 | VM_BUG_ON_PAGE(PageTail(page), page); | |
2535 | ||
2536 | if (likely(!PageCompound(page))) | |
2537 | return atomic_read(&page->_mapcount) + 1; | |
2538 | ||
dd78fedd | 2539 | compound = compound_mapcount(page); |
b20ce5e0 | 2540 | if (PageHuge(page)) |
dd78fedd KS |
2541 | return compound; |
2542 | ret = compound; | |
b20ce5e0 KS |
2543 | for (i = 0; i < HPAGE_PMD_NR; i++) |
2544 | ret += atomic_read(&page[i]._mapcount) + 1; | |
dd78fedd KS |
2545 | /* File pages has compound_mapcount included in _mapcount */ |
2546 | if (!PageAnon(page)) | |
2547 | return ret - compound * HPAGE_PMD_NR; | |
b20ce5e0 KS |
2548 | if (PageDoubleMap(page)) |
2549 | ret -= HPAGE_PMD_NR; | |
2550 | return ret; | |
2551 | } | |
2552 | ||
6d0a07ed AA |
2553 | /* |
2554 | * This calculates accurately how many mappings a transparent hugepage | |
2555 | * has (unlike page_mapcount() which isn't fully accurate). This full | |
2556 | * accuracy is primarily needed to know if copy-on-write faults can | |
2557 | * reuse the page and change the mapping to read-write instead of | |
2558 | * copying them. At the same time this returns the total_mapcount too. | |
2559 | * | |
2560 | * The function returns the highest mapcount any one of the subpages | |
2561 | * has. If the return value is one, even if different processes are | |
2562 | * mapping different subpages of the transparent hugepage, they can | |
2563 | * all reuse it, because each process is reusing a different subpage. | |
2564 | * | |
2565 | * The total_mapcount is instead counting all virtual mappings of the | |
2566 | * subpages. If the total_mapcount is equal to "one", it tells the | |
2567 | * caller all mappings belong to the same "mm" and in turn the | |
2568 | * anon_vma of the transparent hugepage can become the vma->anon_vma | |
2569 | * local one as no other process may be mapping any of the subpages. | |
2570 | * | |
2571 | * It would be more accurate to replace page_mapcount() with | |
2572 | * page_trans_huge_mapcount(), however we only use | |
2573 | * page_trans_huge_mapcount() in the copy-on-write faults where we | |
2574 | * need full accuracy to avoid breaking page pinning, because | |
2575 | * page_trans_huge_mapcount() is slower than page_mapcount(). | |
2576 | */ | |
2577 | int page_trans_huge_mapcount(struct page *page, int *total_mapcount) | |
2578 | { | |
2579 | int i, ret, _total_mapcount, mapcount; | |
2580 | ||
2581 | /* hugetlbfs shouldn't call it */ | |
2582 | VM_BUG_ON_PAGE(PageHuge(page), page); | |
2583 | ||
2584 | if (likely(!PageTransCompound(page))) { | |
2585 | mapcount = atomic_read(&page->_mapcount) + 1; | |
2586 | if (total_mapcount) | |
2587 | *total_mapcount = mapcount; | |
2588 | return mapcount; | |
2589 | } | |
2590 | ||
2591 | page = compound_head(page); | |
2592 | ||
2593 | _total_mapcount = ret = 0; | |
2594 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
2595 | mapcount = atomic_read(&page[i]._mapcount) + 1; | |
2596 | ret = max(ret, mapcount); | |
2597 | _total_mapcount += mapcount; | |
2598 | } | |
2599 | if (PageDoubleMap(page)) { | |
2600 | ret -= 1; | |
2601 | _total_mapcount -= HPAGE_PMD_NR; | |
2602 | } | |
2603 | mapcount = compound_mapcount(page); | |
2604 | ret += mapcount; | |
2605 | _total_mapcount += mapcount; | |
2606 | if (total_mapcount) | |
2607 | *total_mapcount = _total_mapcount; | |
2608 | return ret; | |
2609 | } | |
2610 | ||
b8f593cd HY |
2611 | /* Racy check whether the huge page can be split */ |
2612 | bool can_split_huge_page(struct page *page, int *pextra_pins) | |
2613 | { | |
2614 | int extra_pins; | |
2615 | ||
2616 | /* Additional pins from radix tree */ | |
2617 | if (PageAnon(page)) | |
2618 | extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0; | |
2619 | else | |
2620 | extra_pins = HPAGE_PMD_NR; | |
2621 | if (pextra_pins) | |
2622 | *pextra_pins = extra_pins; | |
2623 | return total_mapcount(page) == page_count(page) - extra_pins - 1; | |
2624 | } | |
2625 | ||
e9b61f19 KS |
2626 | /* |
2627 | * This function splits huge page into normal pages. @page can point to any | |
2628 | * subpage of huge page to split. Split doesn't change the position of @page. | |
2629 | * | |
2630 | * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. | |
2631 | * The huge page must be locked. | |
2632 | * | |
2633 | * If @list is null, tail pages will be added to LRU list, otherwise, to @list. | |
2634 | * | |
2635 | * Both head page and tail pages will inherit mapping, flags, and so on from | |
2636 | * the hugepage. | |
2637 | * | |
2638 | * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if | |
2639 | * they are not mapped. | |
2640 | * | |
2641 | * Returns 0 if the hugepage is split successfully. | |
2642 | * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under | |
2643 | * us. | |
2644 | */ | |
2645 | int split_huge_page_to_list(struct page *page, struct list_head *list) | |
2646 | { | |
2647 | struct page *head = compound_head(page); | |
a3d0a918 | 2648 | struct pglist_data *pgdata = NODE_DATA(page_to_nid(head)); |
baa355fd KS |
2649 | struct anon_vma *anon_vma = NULL; |
2650 | struct address_space *mapping = NULL; | |
2651 | int count, mapcount, extra_pins, ret; | |
d9654322 | 2652 | bool mlocked; |
0b9b6fff | 2653 | unsigned long flags; |
5ba09cb7 | 2654 | pgoff_t end; |
e9b61f19 KS |
2655 | |
2656 | VM_BUG_ON_PAGE(is_huge_zero_page(page), page); | |
e9b61f19 | 2657 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
e9b61f19 KS |
2658 | VM_BUG_ON_PAGE(!PageCompound(page), page); |
2659 | ||
59807685 HY |
2660 | if (PageWriteback(page)) |
2661 | return -EBUSY; | |
2662 | ||
baa355fd KS |
2663 | if (PageAnon(head)) { |
2664 | /* | |
2665 | * The caller does not necessarily hold an mmap_sem that would | |
2666 | * prevent the anon_vma disappearing so we first we take a | |
2667 | * reference to it and then lock the anon_vma for write. This | |
2668 | * is similar to page_lock_anon_vma_read except the write lock | |
2669 | * is taken to serialise against parallel split or collapse | |
2670 | * operations. | |
2671 | */ | |
2672 | anon_vma = page_get_anon_vma(head); | |
2673 | if (!anon_vma) { | |
2674 | ret = -EBUSY; | |
2675 | goto out; | |
2676 | } | |
5ba09cb7 | 2677 | end = -1; |
baa355fd KS |
2678 | mapping = NULL; |
2679 | anon_vma_lock_write(anon_vma); | |
2680 | } else { | |
2681 | mapping = head->mapping; | |
2682 | ||
2683 | /* Truncated ? */ | |
2684 | if (!mapping) { | |
2685 | ret = -EBUSY; | |
2686 | goto out; | |
2687 | } | |
2688 | ||
baa355fd KS |
2689 | anon_vma = NULL; |
2690 | i_mmap_lock_read(mapping); | |
5ba09cb7 HD |
2691 | |
2692 | /* | |
2693 | *__split_huge_page() may need to trim off pages beyond EOF: | |
2694 | * but on 32-bit, i_size_read() takes an irq-unsafe seqlock, | |
2695 | * which cannot be nested inside the page tree lock. So note | |
2696 | * end now: i_size itself may be changed at any moment, but | |
2697 | * head page lock is good enough to serialize the trimming. | |
2698 | */ | |
2699 | end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE); | |
e9b61f19 | 2700 | } |
e9b61f19 KS |
2701 | |
2702 | /* | |
d7b66bd2 | 2703 | * Racy check if we can split the page, before unmap_page() will |
e9b61f19 KS |
2704 | * split PMDs |
2705 | */ | |
b8f593cd | 2706 | if (!can_split_huge_page(head, &extra_pins)) { |
e9b61f19 KS |
2707 | ret = -EBUSY; |
2708 | goto out_unlock; | |
2709 | } | |
2710 | ||
d9654322 | 2711 | mlocked = PageMlocked(page); |
d7b66bd2 | 2712 | unmap_page(head); |
e9b61f19 KS |
2713 | VM_BUG_ON_PAGE(compound_mapcount(head), head); |
2714 | ||
d9654322 KS |
2715 | /* Make sure the page is not on per-CPU pagevec as it takes pin */ |
2716 | if (mlocked) | |
2717 | lru_add_drain(); | |
2718 | ||
baa355fd | 2719 | /* prevent PageLRU to go away from under us, and freeze lru stats */ |
a52633d8 | 2720 | spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags); |
baa355fd KS |
2721 | |
2722 | if (mapping) { | |
2723 | void **pslot; | |
2724 | ||
2725 | spin_lock(&mapping->tree_lock); | |
2726 | pslot = radix_tree_lookup_slot(&mapping->page_tree, | |
2727 | page_index(head)); | |
2728 | /* | |
2729 | * Check if the head page is present in radix tree. | |
2730 | * We assume all tail are present too, if head is there. | |
2731 | */ | |
2732 | if (radix_tree_deref_slot_protected(pslot, | |
2733 | &mapping->tree_lock) != head) | |
2734 | goto fail; | |
2735 | } | |
2736 | ||
0139aa7b | 2737 | /* Prevent deferred_split_scan() touching ->_refcount */ |
baa355fd | 2738 | spin_lock(&pgdata->split_queue_lock); |
e9b61f19 KS |
2739 | count = page_count(head); |
2740 | mapcount = total_mapcount(head); | |
baa355fd | 2741 | if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) { |
9a982250 | 2742 | if (!list_empty(page_deferred_list(head))) { |
a3d0a918 | 2743 | pgdata->split_queue_len--; |
9a982250 KS |
2744 | list_del(page_deferred_list(head)); |
2745 | } | |
65c45377 | 2746 | if (mapping) |
11fb9989 | 2747 | __dec_node_page_state(page, NR_SHMEM_THPS); |
baa355fd | 2748 | spin_unlock(&pgdata->split_queue_lock); |
5ba09cb7 | 2749 | __split_huge_page(page, list, end, flags); |
59807685 HY |
2750 | if (PageSwapCache(head)) { |
2751 | swp_entry_t entry = { .val = page_private(head) }; | |
2752 | ||
2753 | ret = split_swap_cluster(entry); | |
2754 | } else | |
2755 | ret = 0; | |
e9b61f19 | 2756 | } else { |
baa355fd KS |
2757 | if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) { |
2758 | pr_alert("total_mapcount: %u, page_count(): %u\n", | |
2759 | mapcount, count); | |
2760 | if (PageTail(page)) | |
2761 | dump_page(head, NULL); | |
2762 | dump_page(page, "total_mapcount(head) > 0"); | |
2763 | BUG(); | |
2764 | } | |
2765 | spin_unlock(&pgdata->split_queue_lock); | |
2766 | fail: if (mapping) | |
2767 | spin_unlock(&mapping->tree_lock); | |
a52633d8 | 2768 | spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags); |
d7b66bd2 | 2769 | remap_page(head); |
e9b61f19 KS |
2770 | ret = -EBUSY; |
2771 | } | |
2772 | ||
2773 | out_unlock: | |
baa355fd KS |
2774 | if (anon_vma) { |
2775 | anon_vma_unlock_write(anon_vma); | |
2776 | put_anon_vma(anon_vma); | |
2777 | } | |
2778 | if (mapping) | |
2779 | i_mmap_unlock_read(mapping); | |
e9b61f19 KS |
2780 | out: |
2781 | count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); | |
2782 | return ret; | |
2783 | } | |
9a982250 KS |
2784 | |
2785 | void free_transhuge_page(struct page *page) | |
2786 | { | |
a3d0a918 | 2787 | struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); |
9a982250 KS |
2788 | unsigned long flags; |
2789 | ||
a3d0a918 | 2790 | spin_lock_irqsave(&pgdata->split_queue_lock, flags); |
9a982250 | 2791 | if (!list_empty(page_deferred_list(page))) { |
a3d0a918 | 2792 | pgdata->split_queue_len--; |
9a982250 KS |
2793 | list_del(page_deferred_list(page)); |
2794 | } | |
a3d0a918 | 2795 | spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); |
9a982250 KS |
2796 | free_compound_page(page); |
2797 | } | |
2798 | ||
2799 | void deferred_split_huge_page(struct page *page) | |
2800 | { | |
a3d0a918 | 2801 | struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); |
9a982250 KS |
2802 | unsigned long flags; |
2803 | ||
2804 | VM_BUG_ON_PAGE(!PageTransHuge(page), page); | |
2805 | ||
a3d0a918 | 2806 | spin_lock_irqsave(&pgdata->split_queue_lock, flags); |
9a982250 | 2807 | if (list_empty(page_deferred_list(page))) { |
f9719a03 | 2808 | count_vm_event(THP_DEFERRED_SPLIT_PAGE); |
a3d0a918 KS |
2809 | list_add_tail(page_deferred_list(page), &pgdata->split_queue); |
2810 | pgdata->split_queue_len++; | |
9a982250 | 2811 | } |
a3d0a918 | 2812 | spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); |
9a982250 KS |
2813 | } |
2814 | ||
2815 | static unsigned long deferred_split_count(struct shrinker *shrink, | |
2816 | struct shrink_control *sc) | |
2817 | { | |
a3d0a918 | 2818 | struct pglist_data *pgdata = NODE_DATA(sc->nid); |
6aa7de05 | 2819 | return READ_ONCE(pgdata->split_queue_len); |
9a982250 KS |
2820 | } |
2821 | ||
2822 | static unsigned long deferred_split_scan(struct shrinker *shrink, | |
2823 | struct shrink_control *sc) | |
2824 | { | |
a3d0a918 | 2825 | struct pglist_data *pgdata = NODE_DATA(sc->nid); |
9a982250 KS |
2826 | unsigned long flags; |
2827 | LIST_HEAD(list), *pos, *next; | |
2828 | struct page *page; | |
2829 | int split = 0; | |
2830 | ||
a3d0a918 | 2831 | spin_lock_irqsave(&pgdata->split_queue_lock, flags); |
9a982250 | 2832 | /* Take pin on all head pages to avoid freeing them under us */ |
ae026204 | 2833 | list_for_each_safe(pos, next, &pgdata->split_queue) { |
9a982250 KS |
2834 | page = list_entry((void *)pos, struct page, mapping); |
2835 | page = compound_head(page); | |
e3ae1953 KS |
2836 | if (get_page_unless_zero(page)) { |
2837 | list_move(page_deferred_list(page), &list); | |
2838 | } else { | |
2839 | /* We lost race with put_compound_page() */ | |
9a982250 | 2840 | list_del_init(page_deferred_list(page)); |
a3d0a918 | 2841 | pgdata->split_queue_len--; |
9a982250 | 2842 | } |
e3ae1953 KS |
2843 | if (!--sc->nr_to_scan) |
2844 | break; | |
9a982250 | 2845 | } |
a3d0a918 | 2846 | spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); |
9a982250 KS |
2847 | |
2848 | list_for_each_safe(pos, next, &list) { | |
2849 | page = list_entry((void *)pos, struct page, mapping); | |
002fe7f9 KS |
2850 | if (!trylock_page(page)) |
2851 | goto next; | |
9a982250 KS |
2852 | /* split_huge_page() removes page from list on success */ |
2853 | if (!split_huge_page(page)) | |
2854 | split++; | |
2855 | unlock_page(page); | |
002fe7f9 | 2856 | next: |
9a982250 KS |
2857 | put_page(page); |
2858 | } | |
2859 | ||
a3d0a918 KS |
2860 | spin_lock_irqsave(&pgdata->split_queue_lock, flags); |
2861 | list_splice_tail(&list, &pgdata->split_queue); | |
2862 | spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); | |
9a982250 | 2863 | |
cb8d68ec KS |
2864 | /* |
2865 | * Stop shrinker if we didn't split any page, but the queue is empty. | |
2866 | * This can happen if pages were freed under us. | |
2867 | */ | |
2868 | if (!split && list_empty(&pgdata->split_queue)) | |
2869 | return SHRINK_STOP; | |
2870 | return split; | |
9a982250 KS |
2871 | } |
2872 | ||
2873 | static struct shrinker deferred_split_shrinker = { | |
2874 | .count_objects = deferred_split_count, | |
2875 | .scan_objects = deferred_split_scan, | |
2876 | .seeks = DEFAULT_SEEKS, | |
a3d0a918 | 2877 | .flags = SHRINKER_NUMA_AWARE, |
9a982250 | 2878 | }; |
49071d43 KS |
2879 | |
2880 | #ifdef CONFIG_DEBUG_FS | |
2881 | static int split_huge_pages_set(void *data, u64 val) | |
2882 | { | |
2883 | struct zone *zone; | |
2884 | struct page *page; | |
2885 | unsigned long pfn, max_zone_pfn; | |
2886 | unsigned long total = 0, split = 0; | |
2887 | ||
2888 | if (val != 1) | |
2889 | return -EINVAL; | |
2890 | ||
2891 | for_each_populated_zone(zone) { | |
2892 | max_zone_pfn = zone_end_pfn(zone); | |
2893 | for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { | |
2894 | if (!pfn_valid(pfn)) | |
2895 | continue; | |
2896 | ||
2897 | page = pfn_to_page(pfn); | |
2898 | if (!get_page_unless_zero(page)) | |
2899 | continue; | |
2900 | ||
2901 | if (zone != page_zone(page)) | |
2902 | goto next; | |
2903 | ||
baa355fd | 2904 | if (!PageHead(page) || PageHuge(page) || !PageLRU(page)) |
49071d43 KS |
2905 | goto next; |
2906 | ||
2907 | total++; | |
2908 | lock_page(page); | |
2909 | if (!split_huge_page(page)) | |
2910 | split++; | |
2911 | unlock_page(page); | |
2912 | next: | |
2913 | put_page(page); | |
2914 | } | |
2915 | } | |
2916 | ||
145bdaa1 | 2917 | pr_info("%lu of %lu THP split\n", split, total); |
49071d43 KS |
2918 | |
2919 | return 0; | |
2920 | } | |
2921 | DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set, | |
2922 | "%llu\n"); | |
2923 | ||
2924 | static int __init split_huge_pages_debugfs(void) | |
2925 | { | |
2926 | void *ret; | |
2927 | ||
145bdaa1 | 2928 | ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL, |
49071d43 KS |
2929 | &split_huge_pages_fops); |
2930 | if (!ret) | |
2931 | pr_warn("Failed to create split_huge_pages in debugfs"); | |
2932 | return 0; | |
2933 | } | |
2934 | late_initcall(split_huge_pages_debugfs); | |
2935 | #endif | |
616b8371 ZY |
2936 | |
2937 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION | |
2938 | void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw, | |
2939 | struct page *page) | |
2940 | { | |
2941 | struct vm_area_struct *vma = pvmw->vma; | |
2942 | struct mm_struct *mm = vma->vm_mm; | |
2943 | unsigned long address = pvmw->address; | |
2944 | pmd_t pmdval; | |
2945 | swp_entry_t entry; | |
ab6e3d09 | 2946 | pmd_t pmdswp; |
616b8371 ZY |
2947 | |
2948 | if (!(pvmw->pmd && !pvmw->pte)) | |
2949 | return; | |
2950 | ||
616b8371 ZY |
2951 | flush_cache_range(vma, address, address + HPAGE_PMD_SIZE); |
2952 | pmdval = *pvmw->pmd; | |
2953 | pmdp_invalidate(vma, address, pvmw->pmd); | |
2954 | if (pmd_dirty(pmdval)) | |
2955 | set_page_dirty(page); | |
2956 | entry = make_migration_entry(page, pmd_write(pmdval)); | |
ab6e3d09 NH |
2957 | pmdswp = swp_entry_to_pmd(entry); |
2958 | if (pmd_soft_dirty(pmdval)) | |
2959 | pmdswp = pmd_swp_mksoft_dirty(pmdswp); | |
2960 | set_pmd_at(mm, address, pvmw->pmd, pmdswp); | |
616b8371 ZY |
2961 | page_remove_rmap(page, true); |
2962 | put_page(page); | |
616b8371 ZY |
2963 | } |
2964 | ||
2965 | void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new) | |
2966 | { | |
2967 | struct vm_area_struct *vma = pvmw->vma; | |
2968 | struct mm_struct *mm = vma->vm_mm; | |
2969 | unsigned long address = pvmw->address; | |
2970 | unsigned long mmun_start = address & HPAGE_PMD_MASK; | |
2971 | pmd_t pmde; | |
2972 | swp_entry_t entry; | |
2973 | ||
2974 | if (!(pvmw->pmd && !pvmw->pte)) | |
2975 | return; | |
2976 | ||
2977 | entry = pmd_to_swp_entry(*pvmw->pmd); | |
2978 | get_page(new); | |
2979 | pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot)); | |
ab6e3d09 NH |
2980 | if (pmd_swp_soft_dirty(*pvmw->pmd)) |
2981 | pmde = pmd_mksoft_dirty(pmde); | |
616b8371 | 2982 | if (is_write_migration_entry(entry)) |
f55e1014 | 2983 | pmde = maybe_pmd_mkwrite(pmde, vma); |
616b8371 ZY |
2984 | |
2985 | flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE); | |
2986 | page_add_anon_rmap(new, vma, mmun_start, true); | |
2987 | set_pmd_at(mm, mmun_start, pvmw->pmd, pmde); | |
0e1e3cac | 2988 | if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new)) |
616b8371 ZY |
2989 | mlock_vma_page(new); |
2990 | update_mmu_cache_pmd(vma, address, pvmw->pmd); | |
2991 | } | |
2992 | #endif |