Commit | Line | Data |
---|---|---|
20c8ccb1 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
71e3aac0 AA |
2 | /* |
3 | * Copyright (C) 2009 Red Hat, Inc. | |
71e3aac0 AA |
4 | */ |
5 | ||
ae3a8c1c AM |
6 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
7 | ||
71e3aac0 AA |
8 | #include <linux/mm.h> |
9 | #include <linux/sched.h> | |
f7ccbae4 | 10 | #include <linux/sched/coredump.h> |
6a3827d7 | 11 | #include <linux/sched/numa_balancing.h> |
71e3aac0 AA |
12 | #include <linux/highmem.h> |
13 | #include <linux/hugetlb.h> | |
14 | #include <linux/mmu_notifier.h> | |
15 | #include <linux/rmap.h> | |
16 | #include <linux/swap.h> | |
97ae1749 | 17 | #include <linux/shrinker.h> |
ba76149f | 18 | #include <linux/mm_inline.h> |
e9b61f19 | 19 | #include <linux/swapops.h> |
4897c765 | 20 | #include <linux/dax.h> |
ba76149f | 21 | #include <linux/khugepaged.h> |
878aee7d | 22 | #include <linux/freezer.h> |
f25748e3 | 23 | #include <linux/pfn_t.h> |
a664b2d8 | 24 | #include <linux/mman.h> |
3565fce3 | 25 | #include <linux/memremap.h> |
325adeb5 | 26 | #include <linux/pagemap.h> |
49071d43 | 27 | #include <linux/debugfs.h> |
4daae3b4 | 28 | #include <linux/migrate.h> |
43b5fbbd | 29 | #include <linux/hashtable.h> |
6b251fc9 | 30 | #include <linux/userfaultfd_k.h> |
33c3fc71 | 31 | #include <linux/page_idle.h> |
baa355fd | 32 | #include <linux/shmem_fs.h> |
6b31d595 | 33 | #include <linux/oom.h> |
98fa15f3 | 34 | #include <linux/numa.h> |
f7da677b | 35 | #include <linux/page_owner.h> |
97ae1749 | 36 | |
71e3aac0 AA |
37 | #include <asm/tlb.h> |
38 | #include <asm/pgalloc.h> | |
39 | #include "internal.h" | |
40 | ||
ba76149f | 41 | /* |
b14d595a MD |
42 | * By default, transparent hugepage support is disabled in order to avoid |
43 | * risking an increased memory footprint for applications that are not | |
44 | * guaranteed to benefit from it. When transparent hugepage support is | |
45 | * enabled, it is for all mappings, and khugepaged scans all mappings. | |
8bfa3f9a JW |
46 | * Defrag is invoked by khugepaged hugepage allocations and by page faults |
47 | * for all hugepage allocations. | |
ba76149f | 48 | */ |
71e3aac0 | 49 | unsigned long transparent_hugepage_flags __read_mostly = |
13ece886 | 50 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS |
ba76149f | 51 | (1<<TRANSPARENT_HUGEPAGE_FLAG)| |
13ece886 AA |
52 | #endif |
53 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE | |
54 | (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| | |
55 | #endif | |
444eb2a4 | 56 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)| |
79da5407 KS |
57 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| |
58 | (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
ba76149f | 59 | |
9a982250 | 60 | static struct shrinker deferred_split_shrinker; |
f000565a | 61 | |
97ae1749 | 62 | static atomic_t huge_zero_refcount; |
56873f43 | 63 | struct page *huge_zero_page __read_mostly; |
4a6c1297 | 64 | |
7635d9cb MH |
65 | bool transparent_hugepage_enabled(struct vm_area_struct *vma) |
66 | { | |
c0630669 YS |
67 | /* The addr is used to check if the vma size fits */ |
68 | unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE; | |
69 | ||
70 | if (!transhuge_vma_suitable(vma, addr)) | |
71 | return false; | |
7635d9cb MH |
72 | if (vma_is_anonymous(vma)) |
73 | return __transparent_hugepage_enabled(vma); | |
c0630669 YS |
74 | if (vma_is_shmem(vma)) |
75 | return shmem_huge_enabled(vma); | |
7635d9cb MH |
76 | |
77 | return false; | |
78 | } | |
79 | ||
6fcb52a5 | 80 | static struct page *get_huge_zero_page(void) |
97ae1749 KS |
81 | { |
82 | struct page *zero_page; | |
83 | retry: | |
84 | if (likely(atomic_inc_not_zero(&huge_zero_refcount))) | |
4db0c3c2 | 85 | return READ_ONCE(huge_zero_page); |
97ae1749 KS |
86 | |
87 | zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, | |
4a6c1297 | 88 | HPAGE_PMD_ORDER); |
d8a8e1f0 KS |
89 | if (!zero_page) { |
90 | count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); | |
5918d10a | 91 | return NULL; |
d8a8e1f0 KS |
92 | } |
93 | count_vm_event(THP_ZERO_PAGE_ALLOC); | |
97ae1749 | 94 | preempt_disable(); |
5918d10a | 95 | if (cmpxchg(&huge_zero_page, NULL, zero_page)) { |
97ae1749 | 96 | preempt_enable(); |
5ddacbe9 | 97 | __free_pages(zero_page, compound_order(zero_page)); |
97ae1749 KS |
98 | goto retry; |
99 | } | |
100 | ||
101 | /* We take additional reference here. It will be put back by shrinker */ | |
102 | atomic_set(&huge_zero_refcount, 2); | |
103 | preempt_enable(); | |
4db0c3c2 | 104 | return READ_ONCE(huge_zero_page); |
4a6c1297 KS |
105 | } |
106 | ||
6fcb52a5 | 107 | static void put_huge_zero_page(void) |
4a6c1297 | 108 | { |
97ae1749 KS |
109 | /* |
110 | * Counter should never go to zero here. Only shrinker can put | |
111 | * last reference. | |
112 | */ | |
113 | BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); | |
4a6c1297 KS |
114 | } |
115 | ||
6fcb52a5 AL |
116 | struct page *mm_get_huge_zero_page(struct mm_struct *mm) |
117 | { | |
118 | if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) | |
119 | return READ_ONCE(huge_zero_page); | |
120 | ||
121 | if (!get_huge_zero_page()) | |
122 | return NULL; | |
123 | ||
124 | if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) | |
125 | put_huge_zero_page(); | |
126 | ||
127 | return READ_ONCE(huge_zero_page); | |
128 | } | |
129 | ||
130 | void mm_put_huge_zero_page(struct mm_struct *mm) | |
131 | { | |
132 | if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) | |
133 | put_huge_zero_page(); | |
134 | } | |
135 | ||
48896466 GC |
136 | static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, |
137 | struct shrink_control *sc) | |
4a6c1297 | 138 | { |
48896466 GC |
139 | /* we can free zero page only if last reference remains */ |
140 | return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; | |
141 | } | |
97ae1749 | 142 | |
48896466 GC |
143 | static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, |
144 | struct shrink_control *sc) | |
145 | { | |
97ae1749 | 146 | if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { |
5918d10a KS |
147 | struct page *zero_page = xchg(&huge_zero_page, NULL); |
148 | BUG_ON(zero_page == NULL); | |
5ddacbe9 | 149 | __free_pages(zero_page, compound_order(zero_page)); |
48896466 | 150 | return HPAGE_PMD_NR; |
97ae1749 KS |
151 | } |
152 | ||
153 | return 0; | |
4a6c1297 KS |
154 | } |
155 | ||
97ae1749 | 156 | static struct shrinker huge_zero_page_shrinker = { |
48896466 GC |
157 | .count_objects = shrink_huge_zero_page_count, |
158 | .scan_objects = shrink_huge_zero_page_scan, | |
97ae1749 KS |
159 | .seeks = DEFAULT_SEEKS, |
160 | }; | |
161 | ||
71e3aac0 | 162 | #ifdef CONFIG_SYSFS |
71e3aac0 AA |
163 | static ssize_t enabled_show(struct kobject *kobj, |
164 | struct kobj_attribute *attr, char *buf) | |
165 | { | |
444eb2a4 MG |
166 | if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags)) |
167 | return sprintf(buf, "[always] madvise never\n"); | |
168 | else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags)) | |
169 | return sprintf(buf, "always [madvise] never\n"); | |
170 | else | |
171 | return sprintf(buf, "always madvise [never]\n"); | |
71e3aac0 | 172 | } |
444eb2a4 | 173 | |
71e3aac0 AA |
174 | static ssize_t enabled_store(struct kobject *kobj, |
175 | struct kobj_attribute *attr, | |
176 | const char *buf, size_t count) | |
177 | { | |
21440d7e | 178 | ssize_t ret = count; |
ba76149f | 179 | |
f42f2552 | 180 | if (sysfs_streq(buf, "always")) { |
21440d7e DR |
181 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); |
182 | set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); | |
f42f2552 | 183 | } else if (sysfs_streq(buf, "madvise")) { |
21440d7e DR |
184 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); |
185 | set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
f42f2552 | 186 | } else if (sysfs_streq(buf, "never")) { |
21440d7e DR |
187 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); |
188 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
189 | } else | |
190 | ret = -EINVAL; | |
ba76149f AA |
191 | |
192 | if (ret > 0) { | |
b46e756f | 193 | int err = start_stop_khugepaged(); |
ba76149f AA |
194 | if (err) |
195 | ret = err; | |
196 | } | |
ba76149f | 197 | return ret; |
71e3aac0 AA |
198 | } |
199 | static struct kobj_attribute enabled_attr = | |
200 | __ATTR(enabled, 0644, enabled_show, enabled_store); | |
201 | ||
b46e756f | 202 | ssize_t single_hugepage_flag_show(struct kobject *kobj, |
71e3aac0 AA |
203 | struct kobj_attribute *attr, char *buf, |
204 | enum transparent_hugepage_flag flag) | |
205 | { | |
e27e6151 BH |
206 | return sprintf(buf, "%d\n", |
207 | !!test_bit(flag, &transparent_hugepage_flags)); | |
71e3aac0 | 208 | } |
e27e6151 | 209 | |
b46e756f | 210 | ssize_t single_hugepage_flag_store(struct kobject *kobj, |
71e3aac0 AA |
211 | struct kobj_attribute *attr, |
212 | const char *buf, size_t count, | |
213 | enum transparent_hugepage_flag flag) | |
214 | { | |
e27e6151 BH |
215 | unsigned long value; |
216 | int ret; | |
217 | ||
218 | ret = kstrtoul(buf, 10, &value); | |
219 | if (ret < 0) | |
220 | return ret; | |
221 | if (value > 1) | |
222 | return -EINVAL; | |
223 | ||
224 | if (value) | |
71e3aac0 | 225 | set_bit(flag, &transparent_hugepage_flags); |
e27e6151 | 226 | else |
71e3aac0 | 227 | clear_bit(flag, &transparent_hugepage_flags); |
71e3aac0 AA |
228 | |
229 | return count; | |
230 | } | |
231 | ||
71e3aac0 AA |
232 | static ssize_t defrag_show(struct kobject *kobj, |
233 | struct kobj_attribute *attr, char *buf) | |
234 | { | |
444eb2a4 | 235 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) |
21440d7e | 236 | return sprintf(buf, "[always] defer defer+madvise madvise never\n"); |
444eb2a4 | 237 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) |
21440d7e DR |
238 | return sprintf(buf, "always [defer] defer+madvise madvise never\n"); |
239 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) | |
240 | return sprintf(buf, "always defer [defer+madvise] madvise never\n"); | |
241 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) | |
242 | return sprintf(buf, "always defer defer+madvise [madvise] never\n"); | |
243 | return sprintf(buf, "always defer defer+madvise madvise [never]\n"); | |
71e3aac0 | 244 | } |
21440d7e | 245 | |
71e3aac0 AA |
246 | static ssize_t defrag_store(struct kobject *kobj, |
247 | struct kobj_attribute *attr, | |
248 | const char *buf, size_t count) | |
249 | { | |
f42f2552 | 250 | if (sysfs_streq(buf, "always")) { |
21440d7e DR |
251 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); |
252 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
253 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
254 | set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); | |
f42f2552 | 255 | } else if (sysfs_streq(buf, "defer+madvise")) { |
21440d7e DR |
256 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); |
257 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
258 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
259 | set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
f42f2552 | 260 | } else if (sysfs_streq(buf, "defer")) { |
4fad7fb6 DR |
261 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); |
262 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
263 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
264 | set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
f42f2552 | 265 | } else if (sysfs_streq(buf, "madvise")) { |
21440d7e DR |
266 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); |
267 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
268 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
269 | set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
f42f2552 | 270 | } else if (sysfs_streq(buf, "never")) { |
21440d7e DR |
271 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); |
272 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
273 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
274 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
275 | } else | |
276 | return -EINVAL; | |
277 | ||
278 | return count; | |
71e3aac0 AA |
279 | } |
280 | static struct kobj_attribute defrag_attr = | |
281 | __ATTR(defrag, 0644, defrag_show, defrag_store); | |
282 | ||
79da5407 KS |
283 | static ssize_t use_zero_page_show(struct kobject *kobj, |
284 | struct kobj_attribute *attr, char *buf) | |
285 | { | |
b46e756f | 286 | return single_hugepage_flag_show(kobj, attr, buf, |
79da5407 KS |
287 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); |
288 | } | |
289 | static ssize_t use_zero_page_store(struct kobject *kobj, | |
290 | struct kobj_attribute *attr, const char *buf, size_t count) | |
291 | { | |
b46e756f | 292 | return single_hugepage_flag_store(kobj, attr, buf, count, |
79da5407 KS |
293 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); |
294 | } | |
295 | static struct kobj_attribute use_zero_page_attr = | |
296 | __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); | |
49920d28 HD |
297 | |
298 | static ssize_t hpage_pmd_size_show(struct kobject *kobj, | |
299 | struct kobj_attribute *attr, char *buf) | |
300 | { | |
301 | return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE); | |
302 | } | |
303 | static struct kobj_attribute hpage_pmd_size_attr = | |
304 | __ATTR_RO(hpage_pmd_size); | |
305 | ||
71e3aac0 AA |
306 | static struct attribute *hugepage_attr[] = { |
307 | &enabled_attr.attr, | |
308 | &defrag_attr.attr, | |
79da5407 | 309 | &use_zero_page_attr.attr, |
49920d28 | 310 | &hpage_pmd_size_attr.attr, |
396bcc52 | 311 | #ifdef CONFIG_SHMEM |
5a6e75f8 | 312 | &shmem_enabled_attr.attr, |
71e3aac0 AA |
313 | #endif |
314 | NULL, | |
315 | }; | |
316 | ||
8aa95a21 | 317 | static const struct attribute_group hugepage_attr_group = { |
71e3aac0 | 318 | .attrs = hugepage_attr, |
ba76149f AA |
319 | }; |
320 | ||
569e5590 | 321 | static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) |
71e3aac0 | 322 | { |
71e3aac0 AA |
323 | int err; |
324 | ||
569e5590 SL |
325 | *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); |
326 | if (unlikely(!*hugepage_kobj)) { | |
ae3a8c1c | 327 | pr_err("failed to create transparent hugepage kobject\n"); |
569e5590 | 328 | return -ENOMEM; |
ba76149f AA |
329 | } |
330 | ||
569e5590 | 331 | err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); |
ba76149f | 332 | if (err) { |
ae3a8c1c | 333 | pr_err("failed to register transparent hugepage group\n"); |
569e5590 | 334 | goto delete_obj; |
ba76149f AA |
335 | } |
336 | ||
569e5590 | 337 | err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); |
ba76149f | 338 | if (err) { |
ae3a8c1c | 339 | pr_err("failed to register transparent hugepage group\n"); |
569e5590 | 340 | goto remove_hp_group; |
ba76149f | 341 | } |
569e5590 SL |
342 | |
343 | return 0; | |
344 | ||
345 | remove_hp_group: | |
346 | sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); | |
347 | delete_obj: | |
348 | kobject_put(*hugepage_kobj); | |
349 | return err; | |
350 | } | |
351 | ||
352 | static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
353 | { | |
354 | sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); | |
355 | sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); | |
356 | kobject_put(hugepage_kobj); | |
357 | } | |
358 | #else | |
359 | static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) | |
360 | { | |
361 | return 0; | |
362 | } | |
363 | ||
364 | static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
365 | { | |
366 | } | |
367 | #endif /* CONFIG_SYSFS */ | |
368 | ||
369 | static int __init hugepage_init(void) | |
370 | { | |
371 | int err; | |
372 | struct kobject *hugepage_kobj; | |
373 | ||
374 | if (!has_transparent_hugepage()) { | |
375 | transparent_hugepage_flags = 0; | |
376 | return -EINVAL; | |
377 | } | |
378 | ||
ff20c2e0 KS |
379 | /* |
380 | * hugepages can't be allocated by the buddy allocator | |
381 | */ | |
382 | MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER); | |
383 | /* | |
384 | * we use page->mapping and page->index in second tail page | |
385 | * as list_head: assuming THP order >= 2 | |
386 | */ | |
387 | MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2); | |
388 | ||
569e5590 SL |
389 | err = hugepage_init_sysfs(&hugepage_kobj); |
390 | if (err) | |
65ebb64f | 391 | goto err_sysfs; |
ba76149f | 392 | |
b46e756f | 393 | err = khugepaged_init(); |
ba76149f | 394 | if (err) |
65ebb64f | 395 | goto err_slab; |
ba76149f | 396 | |
65ebb64f KS |
397 | err = register_shrinker(&huge_zero_page_shrinker); |
398 | if (err) | |
399 | goto err_hzp_shrinker; | |
9a982250 KS |
400 | err = register_shrinker(&deferred_split_shrinker); |
401 | if (err) | |
402 | goto err_split_shrinker; | |
97ae1749 | 403 | |
97562cd2 RR |
404 | /* |
405 | * By default disable transparent hugepages on smaller systems, | |
406 | * where the extra memory used could hurt more than TLB overhead | |
407 | * is likely to save. The admin can still enable it through /sys. | |
408 | */ | |
ca79b0c2 | 409 | if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) { |
97562cd2 | 410 | transparent_hugepage_flags = 0; |
79553da2 KS |
411 | return 0; |
412 | } | |
97562cd2 | 413 | |
79553da2 | 414 | err = start_stop_khugepaged(); |
65ebb64f KS |
415 | if (err) |
416 | goto err_khugepaged; | |
ba76149f | 417 | |
569e5590 | 418 | return 0; |
65ebb64f | 419 | err_khugepaged: |
9a982250 KS |
420 | unregister_shrinker(&deferred_split_shrinker); |
421 | err_split_shrinker: | |
65ebb64f KS |
422 | unregister_shrinker(&huge_zero_page_shrinker); |
423 | err_hzp_shrinker: | |
b46e756f | 424 | khugepaged_destroy(); |
65ebb64f | 425 | err_slab: |
569e5590 | 426 | hugepage_exit_sysfs(hugepage_kobj); |
65ebb64f | 427 | err_sysfs: |
ba76149f | 428 | return err; |
71e3aac0 | 429 | } |
a64fb3cd | 430 | subsys_initcall(hugepage_init); |
71e3aac0 AA |
431 | |
432 | static int __init setup_transparent_hugepage(char *str) | |
433 | { | |
434 | int ret = 0; | |
435 | if (!str) | |
436 | goto out; | |
437 | if (!strcmp(str, "always")) { | |
438 | set_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
439 | &transparent_hugepage_flags); | |
440 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
441 | &transparent_hugepage_flags); | |
442 | ret = 1; | |
443 | } else if (!strcmp(str, "madvise")) { | |
444 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
445 | &transparent_hugepage_flags); | |
446 | set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
447 | &transparent_hugepage_flags); | |
448 | ret = 1; | |
449 | } else if (!strcmp(str, "never")) { | |
450 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
451 | &transparent_hugepage_flags); | |
452 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
453 | &transparent_hugepage_flags); | |
454 | ret = 1; | |
455 | } | |
456 | out: | |
457 | if (!ret) | |
ae3a8c1c | 458 | pr_warn("transparent_hugepage= cannot parse, ignored\n"); |
71e3aac0 AA |
459 | return ret; |
460 | } | |
461 | __setup("transparent_hugepage=", setup_transparent_hugepage); | |
462 | ||
f55e1014 | 463 | pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) |
71e3aac0 | 464 | { |
f55e1014 | 465 | if (likely(vma->vm_flags & VM_WRITE)) |
71e3aac0 AA |
466 | pmd = pmd_mkwrite(pmd); |
467 | return pmd; | |
468 | } | |
469 | ||
87eaceb3 YS |
470 | #ifdef CONFIG_MEMCG |
471 | static inline struct deferred_split *get_deferred_split_queue(struct page *page) | |
9a982250 | 472 | { |
87eaceb3 YS |
473 | struct mem_cgroup *memcg = compound_head(page)->mem_cgroup; |
474 | struct pglist_data *pgdat = NODE_DATA(page_to_nid(page)); | |
475 | ||
476 | if (memcg) | |
477 | return &memcg->deferred_split_queue; | |
478 | else | |
479 | return &pgdat->deferred_split_queue; | |
9a982250 | 480 | } |
87eaceb3 YS |
481 | #else |
482 | static inline struct deferred_split *get_deferred_split_queue(struct page *page) | |
483 | { | |
484 | struct pglist_data *pgdat = NODE_DATA(page_to_nid(page)); | |
485 | ||
486 | return &pgdat->deferred_split_queue; | |
487 | } | |
488 | #endif | |
9a982250 KS |
489 | |
490 | void prep_transhuge_page(struct page *page) | |
491 | { | |
492 | /* | |
493 | * we use page->mapping and page->indexlru in second tail page | |
494 | * as list_head: assuming THP order >= 2 | |
495 | */ | |
9a982250 KS |
496 | |
497 | INIT_LIST_HEAD(page_deferred_list(page)); | |
498 | set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR); | |
499 | } | |
500 | ||
005ba37c SC |
501 | bool is_transparent_hugepage(struct page *page) |
502 | { | |
503 | if (!PageCompound(page)) | |
fa1f68cc | 504 | return false; |
005ba37c SC |
505 | |
506 | page = compound_head(page); | |
507 | return is_huge_zero_page(page) || | |
508 | page[1].compound_dtor == TRANSHUGE_PAGE_DTOR; | |
509 | } | |
510 | EXPORT_SYMBOL_GPL(is_transparent_hugepage); | |
511 | ||
97d3d0f9 KS |
512 | static unsigned long __thp_get_unmapped_area(struct file *filp, |
513 | unsigned long addr, unsigned long len, | |
74d2fad1 TK |
514 | loff_t off, unsigned long flags, unsigned long size) |
515 | { | |
74d2fad1 TK |
516 | loff_t off_end = off + len; |
517 | loff_t off_align = round_up(off, size); | |
97d3d0f9 | 518 | unsigned long len_pad, ret; |
74d2fad1 TK |
519 | |
520 | if (off_end <= off_align || (off_end - off_align) < size) | |
521 | return 0; | |
522 | ||
523 | len_pad = len + size; | |
524 | if (len_pad < len || (off + len_pad) < off) | |
525 | return 0; | |
526 | ||
97d3d0f9 | 527 | ret = current->mm->get_unmapped_area(filp, addr, len_pad, |
74d2fad1 | 528 | off >> PAGE_SHIFT, flags); |
97d3d0f9 KS |
529 | |
530 | /* | |
531 | * The failure might be due to length padding. The caller will retry | |
532 | * without the padding. | |
533 | */ | |
534 | if (IS_ERR_VALUE(ret)) | |
74d2fad1 TK |
535 | return 0; |
536 | ||
97d3d0f9 KS |
537 | /* |
538 | * Do not try to align to THP boundary if allocation at the address | |
539 | * hint succeeds. | |
540 | */ | |
541 | if (ret == addr) | |
542 | return addr; | |
543 | ||
544 | ret += (off - ret) & (size - 1); | |
545 | return ret; | |
74d2fad1 TK |
546 | } |
547 | ||
548 | unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr, | |
549 | unsigned long len, unsigned long pgoff, unsigned long flags) | |
550 | { | |
97d3d0f9 | 551 | unsigned long ret; |
74d2fad1 TK |
552 | loff_t off = (loff_t)pgoff << PAGE_SHIFT; |
553 | ||
74d2fad1 TK |
554 | if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD)) |
555 | goto out; | |
556 | ||
97d3d0f9 KS |
557 | ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE); |
558 | if (ret) | |
559 | return ret; | |
560 | out: | |
74d2fad1 TK |
561 | return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags); |
562 | } | |
563 | EXPORT_SYMBOL_GPL(thp_get_unmapped_area); | |
564 | ||
2b740303 SJ |
565 | static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf, |
566 | struct page *page, gfp_t gfp) | |
71e3aac0 | 567 | { |
82b0f8c3 | 568 | struct vm_area_struct *vma = vmf->vma; |
71e3aac0 | 569 | pgtable_t pgtable; |
82b0f8c3 | 570 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; |
2b740303 | 571 | vm_fault_t ret = 0; |
71e3aac0 | 572 | |
309381fe | 573 | VM_BUG_ON_PAGE(!PageCompound(page), page); |
00501b53 | 574 | |
d9eb1ea2 | 575 | if (mem_cgroup_charge(page, vma->vm_mm, gfp)) { |
6b251fc9 AA |
576 | put_page(page); |
577 | count_vm_event(THP_FAULT_FALLBACK); | |
85b9f46e | 578 | count_vm_event(THP_FAULT_FALLBACK_CHARGE); |
6b251fc9 AA |
579 | return VM_FAULT_FALLBACK; |
580 | } | |
9d82c694 | 581 | cgroup_throttle_swaprate(page, gfp); |
00501b53 | 582 | |
4cf58924 | 583 | pgtable = pte_alloc_one(vma->vm_mm); |
00501b53 | 584 | if (unlikely(!pgtable)) { |
6b31d595 MH |
585 | ret = VM_FAULT_OOM; |
586 | goto release; | |
00501b53 | 587 | } |
71e3aac0 | 588 | |
c79b57e4 | 589 | clear_huge_page(page, vmf->address, HPAGE_PMD_NR); |
52f37629 MK |
590 | /* |
591 | * The memory barrier inside __SetPageUptodate makes sure that | |
592 | * clear_huge_page writes become visible before the set_pmd_at() | |
593 | * write. | |
594 | */ | |
71e3aac0 AA |
595 | __SetPageUptodate(page); |
596 | ||
82b0f8c3 JK |
597 | vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); |
598 | if (unlikely(!pmd_none(*vmf->pmd))) { | |
6b31d595 | 599 | goto unlock_release; |
71e3aac0 AA |
600 | } else { |
601 | pmd_t entry; | |
6b251fc9 | 602 | |
6b31d595 MH |
603 | ret = check_stable_address_space(vma->vm_mm); |
604 | if (ret) | |
605 | goto unlock_release; | |
606 | ||
6b251fc9 AA |
607 | /* Deliver the page fault to userland */ |
608 | if (userfaultfd_missing(vma)) { | |
2b740303 | 609 | vm_fault_t ret2; |
6b251fc9 | 610 | |
82b0f8c3 | 611 | spin_unlock(vmf->ptl); |
6b251fc9 | 612 | put_page(page); |
bae473a4 | 613 | pte_free(vma->vm_mm, pgtable); |
2b740303 SJ |
614 | ret2 = handle_userfault(vmf, VM_UFFD_MISSING); |
615 | VM_BUG_ON(ret2 & VM_FAULT_FALLBACK); | |
616 | return ret2; | |
6b251fc9 AA |
617 | } |
618 | ||
3122359a | 619 | entry = mk_huge_pmd(page, vma->vm_page_prot); |
f55e1014 | 620 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
d281ee61 | 621 | page_add_new_anon_rmap(page, vma, haddr, true); |
b518154e | 622 | lru_cache_add_inactive_or_unevictable(page, vma); |
82b0f8c3 JK |
623 | pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable); |
624 | set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); | |
bae473a4 | 625 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); |
c4812909 | 626 | mm_inc_nr_ptes(vma->vm_mm); |
82b0f8c3 | 627 | spin_unlock(vmf->ptl); |
6b251fc9 | 628 | count_vm_event(THP_FAULT_ALLOC); |
9d82c694 | 629 | count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC); |
71e3aac0 AA |
630 | } |
631 | ||
aa2e878e | 632 | return 0; |
6b31d595 MH |
633 | unlock_release: |
634 | spin_unlock(vmf->ptl); | |
635 | release: | |
636 | if (pgtable) | |
637 | pte_free(vma->vm_mm, pgtable); | |
6b31d595 MH |
638 | put_page(page); |
639 | return ret; | |
640 | ||
71e3aac0 AA |
641 | } |
642 | ||
444eb2a4 | 643 | /* |
21440d7e DR |
644 | * always: directly stall for all thp allocations |
645 | * defer: wake kswapd and fail if not immediately available | |
646 | * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise | |
647 | * fail if not immediately available | |
648 | * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately | |
649 | * available | |
650 | * never: never stall for any thp allocation | |
444eb2a4 | 651 | */ |
19deb769 | 652 | static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma) |
444eb2a4 | 653 | { |
21440d7e | 654 | const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE); |
2f0799a0 | 655 | |
ac79f78d | 656 | /* Always do synchronous compaction */ |
a8282608 AA |
657 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) |
658 | return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY); | |
ac79f78d DR |
659 | |
660 | /* Kick kcompactd and fail quickly */ | |
21440d7e | 661 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) |
19deb769 | 662 | return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM; |
ac79f78d DR |
663 | |
664 | /* Synchronous compaction if madvised, otherwise kick kcompactd */ | |
21440d7e | 665 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) |
19deb769 DR |
666 | return GFP_TRANSHUGE_LIGHT | |
667 | (vma_madvised ? __GFP_DIRECT_RECLAIM : | |
668 | __GFP_KSWAPD_RECLAIM); | |
ac79f78d DR |
669 | |
670 | /* Only do synchronous compaction if madvised */ | |
21440d7e | 671 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) |
19deb769 DR |
672 | return GFP_TRANSHUGE_LIGHT | |
673 | (vma_madvised ? __GFP_DIRECT_RECLAIM : 0); | |
ac79f78d | 674 | |
19deb769 | 675 | return GFP_TRANSHUGE_LIGHT; |
444eb2a4 MG |
676 | } |
677 | ||
c4088ebd | 678 | /* Caller must hold page table lock. */ |
d295e341 | 679 | static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, |
97ae1749 | 680 | struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, |
5918d10a | 681 | struct page *zero_page) |
fc9fe822 KS |
682 | { |
683 | pmd_t entry; | |
7c414164 AM |
684 | if (!pmd_none(*pmd)) |
685 | return false; | |
5918d10a | 686 | entry = mk_pmd(zero_page, vma->vm_page_prot); |
fc9fe822 | 687 | entry = pmd_mkhuge(entry); |
12c9d70b MW |
688 | if (pgtable) |
689 | pgtable_trans_huge_deposit(mm, pmd, pgtable); | |
fc9fe822 | 690 | set_pmd_at(mm, haddr, pmd, entry); |
c4812909 | 691 | mm_inc_nr_ptes(mm); |
7c414164 | 692 | return true; |
fc9fe822 KS |
693 | } |
694 | ||
2b740303 | 695 | vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf) |
71e3aac0 | 696 | { |
82b0f8c3 | 697 | struct vm_area_struct *vma = vmf->vma; |
077fcf11 | 698 | gfp_t gfp; |
71e3aac0 | 699 | struct page *page; |
82b0f8c3 | 700 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; |
71e3aac0 | 701 | |
43675e6f | 702 | if (!transhuge_vma_suitable(vma, haddr)) |
c0292554 | 703 | return VM_FAULT_FALLBACK; |
128ec037 KS |
704 | if (unlikely(anon_vma_prepare(vma))) |
705 | return VM_FAULT_OOM; | |
6d50e60c | 706 | if (unlikely(khugepaged_enter(vma, vma->vm_flags))) |
128ec037 | 707 | return VM_FAULT_OOM; |
82b0f8c3 | 708 | if (!(vmf->flags & FAULT_FLAG_WRITE) && |
bae473a4 | 709 | !mm_forbids_zeropage(vma->vm_mm) && |
128ec037 KS |
710 | transparent_hugepage_use_zero_page()) { |
711 | pgtable_t pgtable; | |
712 | struct page *zero_page; | |
713 | bool set; | |
2b740303 | 714 | vm_fault_t ret; |
4cf58924 | 715 | pgtable = pte_alloc_one(vma->vm_mm); |
128ec037 | 716 | if (unlikely(!pgtable)) |
ba76149f | 717 | return VM_FAULT_OOM; |
6fcb52a5 | 718 | zero_page = mm_get_huge_zero_page(vma->vm_mm); |
128ec037 | 719 | if (unlikely(!zero_page)) { |
bae473a4 | 720 | pte_free(vma->vm_mm, pgtable); |
81ab4201 | 721 | count_vm_event(THP_FAULT_FALLBACK); |
c0292554 | 722 | return VM_FAULT_FALLBACK; |
b9bbfbe3 | 723 | } |
82b0f8c3 | 724 | vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); |
6b251fc9 AA |
725 | ret = 0; |
726 | set = false; | |
82b0f8c3 | 727 | if (pmd_none(*vmf->pmd)) { |
6b31d595 MH |
728 | ret = check_stable_address_space(vma->vm_mm); |
729 | if (ret) { | |
730 | spin_unlock(vmf->ptl); | |
731 | } else if (userfaultfd_missing(vma)) { | |
82b0f8c3 JK |
732 | spin_unlock(vmf->ptl); |
733 | ret = handle_userfault(vmf, VM_UFFD_MISSING); | |
6b251fc9 AA |
734 | VM_BUG_ON(ret & VM_FAULT_FALLBACK); |
735 | } else { | |
bae473a4 | 736 | set_huge_zero_page(pgtable, vma->vm_mm, vma, |
82b0f8c3 JK |
737 | haddr, vmf->pmd, zero_page); |
738 | spin_unlock(vmf->ptl); | |
6b251fc9 AA |
739 | set = true; |
740 | } | |
741 | } else | |
82b0f8c3 | 742 | spin_unlock(vmf->ptl); |
6fcb52a5 | 743 | if (!set) |
bae473a4 | 744 | pte_free(vma->vm_mm, pgtable); |
6b251fc9 | 745 | return ret; |
71e3aac0 | 746 | } |
19deb769 DR |
747 | gfp = alloc_hugepage_direct_gfpmask(vma); |
748 | page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); | |
128ec037 KS |
749 | if (unlikely(!page)) { |
750 | count_vm_event(THP_FAULT_FALLBACK); | |
c0292554 | 751 | return VM_FAULT_FALLBACK; |
128ec037 | 752 | } |
9a982250 | 753 | prep_transhuge_page(page); |
82b0f8c3 | 754 | return __do_huge_pmd_anonymous_page(vmf, page, gfp); |
71e3aac0 AA |
755 | } |
756 | ||
ae18d6dc | 757 | static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, |
3b6521f5 OH |
758 | pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write, |
759 | pgtable_t pgtable) | |
5cad465d MW |
760 | { |
761 | struct mm_struct *mm = vma->vm_mm; | |
762 | pmd_t entry; | |
763 | spinlock_t *ptl; | |
764 | ||
765 | ptl = pmd_lock(mm, pmd); | |
c6f3c5ee AK |
766 | if (!pmd_none(*pmd)) { |
767 | if (write) { | |
768 | if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) { | |
769 | WARN_ON_ONCE(!is_huge_zero_pmd(*pmd)); | |
770 | goto out_unlock; | |
771 | } | |
772 | entry = pmd_mkyoung(*pmd); | |
773 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
774 | if (pmdp_set_access_flags(vma, addr, pmd, entry, 1)) | |
775 | update_mmu_cache_pmd(vma, addr, pmd); | |
776 | } | |
777 | ||
778 | goto out_unlock; | |
779 | } | |
780 | ||
f25748e3 DW |
781 | entry = pmd_mkhuge(pfn_t_pmd(pfn, prot)); |
782 | if (pfn_t_devmap(pfn)) | |
783 | entry = pmd_mkdevmap(entry); | |
01871e59 | 784 | if (write) { |
f55e1014 LT |
785 | entry = pmd_mkyoung(pmd_mkdirty(entry)); |
786 | entry = maybe_pmd_mkwrite(entry, vma); | |
5cad465d | 787 | } |
3b6521f5 OH |
788 | |
789 | if (pgtable) { | |
790 | pgtable_trans_huge_deposit(mm, pmd, pgtable); | |
c4812909 | 791 | mm_inc_nr_ptes(mm); |
c6f3c5ee | 792 | pgtable = NULL; |
3b6521f5 OH |
793 | } |
794 | ||
01871e59 RZ |
795 | set_pmd_at(mm, addr, pmd, entry); |
796 | update_mmu_cache_pmd(vma, addr, pmd); | |
c6f3c5ee AK |
797 | |
798 | out_unlock: | |
5cad465d | 799 | spin_unlock(ptl); |
c6f3c5ee AK |
800 | if (pgtable) |
801 | pte_free(mm, pgtable); | |
5cad465d MW |
802 | } |
803 | ||
9a9731b1 THV |
804 | /** |
805 | * vmf_insert_pfn_pmd_prot - insert a pmd size pfn | |
806 | * @vmf: Structure describing the fault | |
807 | * @pfn: pfn to insert | |
808 | * @pgprot: page protection to use | |
809 | * @write: whether it's a write fault | |
810 | * | |
811 | * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and | |
812 | * also consult the vmf_insert_mixed_prot() documentation when | |
813 | * @pgprot != @vmf->vma->vm_page_prot. | |
814 | * | |
815 | * Return: vm_fault_t value. | |
816 | */ | |
817 | vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn, | |
818 | pgprot_t pgprot, bool write) | |
5cad465d | 819 | { |
fce86ff5 DW |
820 | unsigned long addr = vmf->address & PMD_MASK; |
821 | struct vm_area_struct *vma = vmf->vma; | |
3b6521f5 | 822 | pgtable_t pgtable = NULL; |
fce86ff5 | 823 | |
5cad465d MW |
824 | /* |
825 | * If we had pmd_special, we could avoid all these restrictions, | |
826 | * but we need to be consistent with PTEs and architectures that | |
827 | * can't support a 'special' bit. | |
828 | */ | |
e1fb4a08 DJ |
829 | BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && |
830 | !pfn_t_devmap(pfn)); | |
5cad465d MW |
831 | BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == |
832 | (VM_PFNMAP|VM_MIXEDMAP)); | |
833 | BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); | |
5cad465d MW |
834 | |
835 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
836 | return VM_FAULT_SIGBUS; | |
308a047c | 837 | |
3b6521f5 | 838 | if (arch_needs_pgtable_deposit()) { |
4cf58924 | 839 | pgtable = pte_alloc_one(vma->vm_mm); |
3b6521f5 OH |
840 | if (!pgtable) |
841 | return VM_FAULT_OOM; | |
842 | } | |
843 | ||
308a047c BP |
844 | track_pfn_insert(vma, &pgprot, pfn); |
845 | ||
fce86ff5 | 846 | insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable); |
ae18d6dc | 847 | return VM_FAULT_NOPAGE; |
5cad465d | 848 | } |
9a9731b1 | 849 | EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot); |
5cad465d | 850 | |
a00cc7d9 | 851 | #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD |
f55e1014 | 852 | static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma) |
a00cc7d9 | 853 | { |
f55e1014 | 854 | if (likely(vma->vm_flags & VM_WRITE)) |
a00cc7d9 MW |
855 | pud = pud_mkwrite(pud); |
856 | return pud; | |
857 | } | |
858 | ||
859 | static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, | |
860 | pud_t *pud, pfn_t pfn, pgprot_t prot, bool write) | |
861 | { | |
862 | struct mm_struct *mm = vma->vm_mm; | |
863 | pud_t entry; | |
864 | spinlock_t *ptl; | |
865 | ||
866 | ptl = pud_lock(mm, pud); | |
c6f3c5ee AK |
867 | if (!pud_none(*pud)) { |
868 | if (write) { | |
869 | if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) { | |
870 | WARN_ON_ONCE(!is_huge_zero_pud(*pud)); | |
871 | goto out_unlock; | |
872 | } | |
873 | entry = pud_mkyoung(*pud); | |
874 | entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma); | |
875 | if (pudp_set_access_flags(vma, addr, pud, entry, 1)) | |
876 | update_mmu_cache_pud(vma, addr, pud); | |
877 | } | |
878 | goto out_unlock; | |
879 | } | |
880 | ||
a00cc7d9 MW |
881 | entry = pud_mkhuge(pfn_t_pud(pfn, prot)); |
882 | if (pfn_t_devmap(pfn)) | |
883 | entry = pud_mkdevmap(entry); | |
884 | if (write) { | |
f55e1014 LT |
885 | entry = pud_mkyoung(pud_mkdirty(entry)); |
886 | entry = maybe_pud_mkwrite(entry, vma); | |
a00cc7d9 MW |
887 | } |
888 | set_pud_at(mm, addr, pud, entry); | |
889 | update_mmu_cache_pud(vma, addr, pud); | |
c6f3c5ee AK |
890 | |
891 | out_unlock: | |
a00cc7d9 MW |
892 | spin_unlock(ptl); |
893 | } | |
894 | ||
9a9731b1 THV |
895 | /** |
896 | * vmf_insert_pfn_pud_prot - insert a pud size pfn | |
897 | * @vmf: Structure describing the fault | |
898 | * @pfn: pfn to insert | |
899 | * @pgprot: page protection to use | |
900 | * @write: whether it's a write fault | |
901 | * | |
902 | * Insert a pud size pfn. See vmf_insert_pfn() for additional info and | |
903 | * also consult the vmf_insert_mixed_prot() documentation when | |
904 | * @pgprot != @vmf->vma->vm_page_prot. | |
905 | * | |
906 | * Return: vm_fault_t value. | |
907 | */ | |
908 | vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn, | |
909 | pgprot_t pgprot, bool write) | |
a00cc7d9 | 910 | { |
fce86ff5 DW |
911 | unsigned long addr = vmf->address & PUD_MASK; |
912 | struct vm_area_struct *vma = vmf->vma; | |
fce86ff5 | 913 | |
a00cc7d9 MW |
914 | /* |
915 | * If we had pud_special, we could avoid all these restrictions, | |
916 | * but we need to be consistent with PTEs and architectures that | |
917 | * can't support a 'special' bit. | |
918 | */ | |
62ec0d8c DJ |
919 | BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && |
920 | !pfn_t_devmap(pfn)); | |
a00cc7d9 MW |
921 | BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == |
922 | (VM_PFNMAP|VM_MIXEDMAP)); | |
923 | BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); | |
a00cc7d9 MW |
924 | |
925 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
926 | return VM_FAULT_SIGBUS; | |
927 | ||
928 | track_pfn_insert(vma, &pgprot, pfn); | |
929 | ||
fce86ff5 | 930 | insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write); |
a00cc7d9 MW |
931 | return VM_FAULT_NOPAGE; |
932 | } | |
9a9731b1 | 933 | EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot); |
a00cc7d9 MW |
934 | #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ |
935 | ||
3565fce3 | 936 | static void touch_pmd(struct vm_area_struct *vma, unsigned long addr, |
a8f97366 | 937 | pmd_t *pmd, int flags) |
3565fce3 DW |
938 | { |
939 | pmd_t _pmd; | |
940 | ||
a8f97366 KS |
941 | _pmd = pmd_mkyoung(*pmd); |
942 | if (flags & FOLL_WRITE) | |
943 | _pmd = pmd_mkdirty(_pmd); | |
3565fce3 | 944 | if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, |
a8f97366 | 945 | pmd, _pmd, flags & FOLL_WRITE)) |
3565fce3 DW |
946 | update_mmu_cache_pmd(vma, addr, pmd); |
947 | } | |
948 | ||
949 | struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr, | |
df06b37f | 950 | pmd_t *pmd, int flags, struct dev_pagemap **pgmap) |
3565fce3 DW |
951 | { |
952 | unsigned long pfn = pmd_pfn(*pmd); | |
953 | struct mm_struct *mm = vma->vm_mm; | |
3565fce3 DW |
954 | struct page *page; |
955 | ||
956 | assert_spin_locked(pmd_lockptr(mm, pmd)); | |
957 | ||
8310d48b KF |
958 | /* |
959 | * When we COW a devmap PMD entry, we split it into PTEs, so we should | |
960 | * not be in this function with `flags & FOLL_COW` set. | |
961 | */ | |
962 | WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set"); | |
963 | ||
3faa52c0 JH |
964 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ |
965 | if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == | |
966 | (FOLL_PIN | FOLL_GET))) | |
967 | return NULL; | |
968 | ||
f6f37321 | 969 | if (flags & FOLL_WRITE && !pmd_write(*pmd)) |
3565fce3 DW |
970 | return NULL; |
971 | ||
972 | if (pmd_present(*pmd) && pmd_devmap(*pmd)) | |
973 | /* pass */; | |
974 | else | |
975 | return NULL; | |
976 | ||
977 | if (flags & FOLL_TOUCH) | |
a8f97366 | 978 | touch_pmd(vma, addr, pmd, flags); |
3565fce3 DW |
979 | |
980 | /* | |
981 | * device mapped pages can only be returned if the | |
982 | * caller will manage the page reference count. | |
983 | */ | |
3faa52c0 | 984 | if (!(flags & (FOLL_GET | FOLL_PIN))) |
3565fce3 DW |
985 | return ERR_PTR(-EEXIST); |
986 | ||
987 | pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT; | |
df06b37f KB |
988 | *pgmap = get_dev_pagemap(pfn, *pgmap); |
989 | if (!*pgmap) | |
3565fce3 DW |
990 | return ERR_PTR(-EFAULT); |
991 | page = pfn_to_page(pfn); | |
3faa52c0 JH |
992 | if (!try_grab_page(page, flags)) |
993 | page = ERR_PTR(-ENOMEM); | |
3565fce3 DW |
994 | |
995 | return page; | |
996 | } | |
997 | ||
71e3aac0 AA |
998 | int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, |
999 | pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, | |
1000 | struct vm_area_struct *vma) | |
1001 | { | |
c4088ebd | 1002 | spinlock_t *dst_ptl, *src_ptl; |
71e3aac0 AA |
1003 | struct page *src_page; |
1004 | pmd_t pmd; | |
12c9d70b | 1005 | pgtable_t pgtable = NULL; |
628d47ce | 1006 | int ret = -ENOMEM; |
71e3aac0 | 1007 | |
628d47ce KS |
1008 | /* Skip if can be re-fill on fault */ |
1009 | if (!vma_is_anonymous(vma)) | |
1010 | return 0; | |
1011 | ||
4cf58924 | 1012 | pgtable = pte_alloc_one(dst_mm); |
628d47ce KS |
1013 | if (unlikely(!pgtable)) |
1014 | goto out; | |
71e3aac0 | 1015 | |
c4088ebd KS |
1016 | dst_ptl = pmd_lock(dst_mm, dst_pmd); |
1017 | src_ptl = pmd_lockptr(src_mm, src_pmd); | |
1018 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
71e3aac0 AA |
1019 | |
1020 | ret = -EAGAIN; | |
1021 | pmd = *src_pmd; | |
84c3fc4e | 1022 | |
b569a176 PX |
1023 | /* |
1024 | * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA | |
1025 | * does not have the VM_UFFD_WP, which means that the uffd | |
1026 | * fork event is not enabled. | |
1027 | */ | |
1028 | if (!(vma->vm_flags & VM_UFFD_WP)) | |
1029 | pmd = pmd_clear_uffd_wp(pmd); | |
1030 | ||
84c3fc4e ZY |
1031 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
1032 | if (unlikely(is_swap_pmd(pmd))) { | |
1033 | swp_entry_t entry = pmd_to_swp_entry(pmd); | |
1034 | ||
1035 | VM_BUG_ON(!is_pmd_migration_entry(pmd)); | |
1036 | if (is_write_migration_entry(entry)) { | |
1037 | make_migration_entry_read(&entry); | |
1038 | pmd = swp_entry_to_pmd(entry); | |
ab6e3d09 NH |
1039 | if (pmd_swp_soft_dirty(*src_pmd)) |
1040 | pmd = pmd_swp_mksoft_dirty(pmd); | |
84c3fc4e ZY |
1041 | set_pmd_at(src_mm, addr, src_pmd, pmd); |
1042 | } | |
dd8a67f9 | 1043 | add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); |
af5b0f6a | 1044 | mm_inc_nr_ptes(dst_mm); |
dd8a67f9 | 1045 | pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); |
84c3fc4e ZY |
1046 | set_pmd_at(dst_mm, addr, dst_pmd, pmd); |
1047 | ret = 0; | |
1048 | goto out_unlock; | |
1049 | } | |
1050 | #endif | |
1051 | ||
628d47ce | 1052 | if (unlikely(!pmd_trans_huge(pmd))) { |
71e3aac0 AA |
1053 | pte_free(dst_mm, pgtable); |
1054 | goto out_unlock; | |
1055 | } | |
fc9fe822 | 1056 | /* |
c4088ebd | 1057 | * When page table lock is held, the huge zero pmd should not be |
fc9fe822 KS |
1058 | * under splitting since we don't split the page itself, only pmd to |
1059 | * a page table. | |
1060 | */ | |
1061 | if (is_huge_zero_pmd(pmd)) { | |
5918d10a | 1062 | struct page *zero_page; |
97ae1749 KS |
1063 | /* |
1064 | * get_huge_zero_page() will never allocate a new page here, | |
1065 | * since we already have a zero page to copy. It just takes a | |
1066 | * reference. | |
1067 | */ | |
6fcb52a5 | 1068 | zero_page = mm_get_huge_zero_page(dst_mm); |
6b251fc9 | 1069 | set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, |
5918d10a | 1070 | zero_page); |
fc9fe822 KS |
1071 | ret = 0; |
1072 | goto out_unlock; | |
1073 | } | |
de466bd6 | 1074 | |
628d47ce KS |
1075 | src_page = pmd_page(pmd); |
1076 | VM_BUG_ON_PAGE(!PageHead(src_page), src_page); | |
1077 | get_page(src_page); | |
1078 | page_dup_rmap(src_page, true); | |
1079 | add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); | |
c4812909 | 1080 | mm_inc_nr_ptes(dst_mm); |
628d47ce | 1081 | pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); |
71e3aac0 AA |
1082 | |
1083 | pmdp_set_wrprotect(src_mm, addr, src_pmd); | |
1084 | pmd = pmd_mkold(pmd_wrprotect(pmd)); | |
1085 | set_pmd_at(dst_mm, addr, dst_pmd, pmd); | |
71e3aac0 AA |
1086 | |
1087 | ret = 0; | |
1088 | out_unlock: | |
c4088ebd KS |
1089 | spin_unlock(src_ptl); |
1090 | spin_unlock(dst_ptl); | |
71e3aac0 AA |
1091 | out: |
1092 | return ret; | |
1093 | } | |
1094 | ||
a00cc7d9 MW |
1095 | #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD |
1096 | static void touch_pud(struct vm_area_struct *vma, unsigned long addr, | |
a8f97366 | 1097 | pud_t *pud, int flags) |
a00cc7d9 MW |
1098 | { |
1099 | pud_t _pud; | |
1100 | ||
a8f97366 KS |
1101 | _pud = pud_mkyoung(*pud); |
1102 | if (flags & FOLL_WRITE) | |
1103 | _pud = pud_mkdirty(_pud); | |
a00cc7d9 | 1104 | if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK, |
a8f97366 | 1105 | pud, _pud, flags & FOLL_WRITE)) |
a00cc7d9 MW |
1106 | update_mmu_cache_pud(vma, addr, pud); |
1107 | } | |
1108 | ||
1109 | struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr, | |
df06b37f | 1110 | pud_t *pud, int flags, struct dev_pagemap **pgmap) |
a00cc7d9 MW |
1111 | { |
1112 | unsigned long pfn = pud_pfn(*pud); | |
1113 | struct mm_struct *mm = vma->vm_mm; | |
a00cc7d9 MW |
1114 | struct page *page; |
1115 | ||
1116 | assert_spin_locked(pud_lockptr(mm, pud)); | |
1117 | ||
f6f37321 | 1118 | if (flags & FOLL_WRITE && !pud_write(*pud)) |
a00cc7d9 MW |
1119 | return NULL; |
1120 | ||
3faa52c0 JH |
1121 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ |
1122 | if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == | |
1123 | (FOLL_PIN | FOLL_GET))) | |
1124 | return NULL; | |
1125 | ||
a00cc7d9 MW |
1126 | if (pud_present(*pud) && pud_devmap(*pud)) |
1127 | /* pass */; | |
1128 | else | |
1129 | return NULL; | |
1130 | ||
1131 | if (flags & FOLL_TOUCH) | |
a8f97366 | 1132 | touch_pud(vma, addr, pud, flags); |
a00cc7d9 MW |
1133 | |
1134 | /* | |
1135 | * device mapped pages can only be returned if the | |
1136 | * caller will manage the page reference count. | |
3faa52c0 JH |
1137 | * |
1138 | * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here: | |
a00cc7d9 | 1139 | */ |
3faa52c0 | 1140 | if (!(flags & (FOLL_GET | FOLL_PIN))) |
a00cc7d9 MW |
1141 | return ERR_PTR(-EEXIST); |
1142 | ||
1143 | pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT; | |
df06b37f KB |
1144 | *pgmap = get_dev_pagemap(pfn, *pgmap); |
1145 | if (!*pgmap) | |
a00cc7d9 MW |
1146 | return ERR_PTR(-EFAULT); |
1147 | page = pfn_to_page(pfn); | |
3faa52c0 JH |
1148 | if (!try_grab_page(page, flags)) |
1149 | page = ERR_PTR(-ENOMEM); | |
a00cc7d9 MW |
1150 | |
1151 | return page; | |
1152 | } | |
1153 | ||
1154 | int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm, | |
1155 | pud_t *dst_pud, pud_t *src_pud, unsigned long addr, | |
1156 | struct vm_area_struct *vma) | |
1157 | { | |
1158 | spinlock_t *dst_ptl, *src_ptl; | |
1159 | pud_t pud; | |
1160 | int ret; | |
1161 | ||
1162 | dst_ptl = pud_lock(dst_mm, dst_pud); | |
1163 | src_ptl = pud_lockptr(src_mm, src_pud); | |
1164 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
1165 | ||
1166 | ret = -EAGAIN; | |
1167 | pud = *src_pud; | |
1168 | if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud))) | |
1169 | goto out_unlock; | |
1170 | ||
1171 | /* | |
1172 | * When page table lock is held, the huge zero pud should not be | |
1173 | * under splitting since we don't split the page itself, only pud to | |
1174 | * a page table. | |
1175 | */ | |
1176 | if (is_huge_zero_pud(pud)) { | |
1177 | /* No huge zero pud yet */ | |
1178 | } | |
1179 | ||
1180 | pudp_set_wrprotect(src_mm, addr, src_pud); | |
1181 | pud = pud_mkold(pud_wrprotect(pud)); | |
1182 | set_pud_at(dst_mm, addr, dst_pud, pud); | |
1183 | ||
1184 | ret = 0; | |
1185 | out_unlock: | |
1186 | spin_unlock(src_ptl); | |
1187 | spin_unlock(dst_ptl); | |
1188 | return ret; | |
1189 | } | |
1190 | ||
1191 | void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud) | |
1192 | { | |
1193 | pud_t entry; | |
1194 | unsigned long haddr; | |
1195 | bool write = vmf->flags & FAULT_FLAG_WRITE; | |
1196 | ||
1197 | vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud); | |
1198 | if (unlikely(!pud_same(*vmf->pud, orig_pud))) | |
1199 | goto unlock; | |
1200 | ||
1201 | entry = pud_mkyoung(orig_pud); | |
1202 | if (write) | |
1203 | entry = pud_mkdirty(entry); | |
1204 | haddr = vmf->address & HPAGE_PUD_MASK; | |
1205 | if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write)) | |
1206 | update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud); | |
1207 | ||
1208 | unlock: | |
1209 | spin_unlock(vmf->ptl); | |
1210 | } | |
1211 | #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ | |
1212 | ||
82b0f8c3 | 1213 | void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd) |
a1dd450b WD |
1214 | { |
1215 | pmd_t entry; | |
1216 | unsigned long haddr; | |
20f664aa | 1217 | bool write = vmf->flags & FAULT_FLAG_WRITE; |
a1dd450b | 1218 | |
82b0f8c3 JK |
1219 | vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); |
1220 | if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) | |
a1dd450b WD |
1221 | goto unlock; |
1222 | ||
1223 | entry = pmd_mkyoung(orig_pmd); | |
20f664aa MK |
1224 | if (write) |
1225 | entry = pmd_mkdirty(entry); | |
82b0f8c3 | 1226 | haddr = vmf->address & HPAGE_PMD_MASK; |
20f664aa | 1227 | if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write)) |
82b0f8c3 | 1228 | update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd); |
a1dd450b WD |
1229 | |
1230 | unlock: | |
82b0f8c3 | 1231 | spin_unlock(vmf->ptl); |
a1dd450b WD |
1232 | } |
1233 | ||
2b740303 | 1234 | vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd) |
71e3aac0 | 1235 | { |
82b0f8c3 | 1236 | struct vm_area_struct *vma = vmf->vma; |
3917c802 | 1237 | struct page *page; |
82b0f8c3 | 1238 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; |
71e3aac0 | 1239 | |
82b0f8c3 | 1240 | vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd); |
81d1b09c | 1241 | VM_BUG_ON_VMA(!vma->anon_vma, vma); |
3917c802 | 1242 | |
93b4796d | 1243 | if (is_huge_zero_pmd(orig_pmd)) |
3917c802 KS |
1244 | goto fallback; |
1245 | ||
82b0f8c3 | 1246 | spin_lock(vmf->ptl); |
3917c802 KS |
1247 | |
1248 | if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { | |
1249 | spin_unlock(vmf->ptl); | |
1250 | return 0; | |
1251 | } | |
71e3aac0 AA |
1252 | |
1253 | page = pmd_page(orig_pmd); | |
309381fe | 1254 | VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); |
3917c802 KS |
1255 | |
1256 | /* Lock page for reuse_swap_page() */ | |
ba3c4ce6 HY |
1257 | if (!trylock_page(page)) { |
1258 | get_page(page); | |
1259 | spin_unlock(vmf->ptl); | |
1260 | lock_page(page); | |
1261 | spin_lock(vmf->ptl); | |
1262 | if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { | |
3917c802 | 1263 | spin_unlock(vmf->ptl); |
ba3c4ce6 HY |
1264 | unlock_page(page); |
1265 | put_page(page); | |
3917c802 | 1266 | return 0; |
ba3c4ce6 HY |
1267 | } |
1268 | put_page(page); | |
1269 | } | |
3917c802 KS |
1270 | |
1271 | /* | |
1272 | * We can only reuse the page if nobody else maps the huge page or it's | |
1273 | * part. | |
1274 | */ | |
ba3c4ce6 | 1275 | if (reuse_swap_page(page, NULL)) { |
71e3aac0 AA |
1276 | pmd_t entry; |
1277 | entry = pmd_mkyoung(orig_pmd); | |
f55e1014 | 1278 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
3917c802 | 1279 | if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1)) |
82b0f8c3 | 1280 | update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); |
ba3c4ce6 | 1281 | unlock_page(page); |
82b0f8c3 | 1282 | spin_unlock(vmf->ptl); |
3917c802 | 1283 | return VM_FAULT_WRITE; |
71e3aac0 | 1284 | } |
3917c802 KS |
1285 | |
1286 | unlock_page(page); | |
82b0f8c3 | 1287 | spin_unlock(vmf->ptl); |
3917c802 KS |
1288 | fallback: |
1289 | __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL); | |
1290 | return VM_FAULT_FALLBACK; | |
71e3aac0 AA |
1291 | } |
1292 | ||
8310d48b | 1293 | /* |
a308c71b PX |
1294 | * FOLL_FORCE can write to even unwritable pmd's, but only |
1295 | * after we've gone through a COW cycle and they are dirty. | |
8310d48b KF |
1296 | */ |
1297 | static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags) | |
1298 | { | |
a308c71b PX |
1299 | return pmd_write(pmd) || |
1300 | ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd)); | |
8310d48b KF |
1301 | } |
1302 | ||
b676b293 | 1303 | struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, |
71e3aac0 AA |
1304 | unsigned long addr, |
1305 | pmd_t *pmd, | |
1306 | unsigned int flags) | |
1307 | { | |
b676b293 | 1308 | struct mm_struct *mm = vma->vm_mm; |
71e3aac0 AA |
1309 | struct page *page = NULL; |
1310 | ||
c4088ebd | 1311 | assert_spin_locked(pmd_lockptr(mm, pmd)); |
71e3aac0 | 1312 | |
8310d48b | 1313 | if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags)) |
71e3aac0 AA |
1314 | goto out; |
1315 | ||
85facf25 KS |
1316 | /* Avoid dumping huge zero page */ |
1317 | if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) | |
1318 | return ERR_PTR(-EFAULT); | |
1319 | ||
2b4847e7 | 1320 | /* Full NUMA hinting faults to serialise migration in fault paths */ |
8a0516ed | 1321 | if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) |
2b4847e7 MG |
1322 | goto out; |
1323 | ||
71e3aac0 | 1324 | page = pmd_page(*pmd); |
ca120cf6 | 1325 | VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page); |
3faa52c0 JH |
1326 | |
1327 | if (!try_grab_page(page, flags)) | |
1328 | return ERR_PTR(-ENOMEM); | |
1329 | ||
3565fce3 | 1330 | if (flags & FOLL_TOUCH) |
a8f97366 | 1331 | touch_pmd(vma, addr, pmd, flags); |
3faa52c0 | 1332 | |
de60f5f1 | 1333 | if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { |
e90309c9 KS |
1334 | /* |
1335 | * We don't mlock() pte-mapped THPs. This way we can avoid | |
1336 | * leaking mlocked pages into non-VM_LOCKED VMAs. | |
1337 | * | |
9a73f61b KS |
1338 | * For anon THP: |
1339 | * | |
e90309c9 KS |
1340 | * In most cases the pmd is the only mapping of the page as we |
1341 | * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for | |
1342 | * writable private mappings in populate_vma_page_range(). | |
1343 | * | |
1344 | * The only scenario when we have the page shared here is if we | |
1345 | * mlocking read-only mapping shared over fork(). We skip | |
1346 | * mlocking such pages. | |
9a73f61b KS |
1347 | * |
1348 | * For file THP: | |
1349 | * | |
1350 | * We can expect PageDoubleMap() to be stable under page lock: | |
1351 | * for file pages we set it in page_add_file_rmap(), which | |
1352 | * requires page to be locked. | |
e90309c9 | 1353 | */ |
9a73f61b KS |
1354 | |
1355 | if (PageAnon(page) && compound_mapcount(page) != 1) | |
1356 | goto skip_mlock; | |
1357 | if (PageDoubleMap(page) || !page->mapping) | |
1358 | goto skip_mlock; | |
1359 | if (!trylock_page(page)) | |
1360 | goto skip_mlock; | |
9a73f61b KS |
1361 | if (page->mapping && !PageDoubleMap(page)) |
1362 | mlock_vma_page(page); | |
1363 | unlock_page(page); | |
b676b293 | 1364 | } |
9a73f61b | 1365 | skip_mlock: |
71e3aac0 | 1366 | page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; |
ca120cf6 | 1367 | VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page); |
71e3aac0 AA |
1368 | |
1369 | out: | |
1370 | return page; | |
1371 | } | |
1372 | ||
d10e63f2 | 1373 | /* NUMA hinting page fault entry point for trans huge pmds */ |
2b740303 | 1374 | vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd) |
d10e63f2 | 1375 | { |
82b0f8c3 | 1376 | struct vm_area_struct *vma = vmf->vma; |
b8916634 | 1377 | struct anon_vma *anon_vma = NULL; |
b32967ff | 1378 | struct page *page; |
82b0f8c3 | 1379 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; |
98fa15f3 | 1380 | int page_nid = NUMA_NO_NODE, this_nid = numa_node_id(); |
90572890 | 1381 | int target_nid, last_cpupid = -1; |
8191acbd MG |
1382 | bool page_locked; |
1383 | bool migrated = false; | |
b191f9b1 | 1384 | bool was_writable; |
6688cc05 | 1385 | int flags = 0; |
d10e63f2 | 1386 | |
82b0f8c3 JK |
1387 | vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); |
1388 | if (unlikely(!pmd_same(pmd, *vmf->pmd))) | |
d10e63f2 MG |
1389 | goto out_unlock; |
1390 | ||
de466bd6 MG |
1391 | /* |
1392 | * If there are potential migrations, wait for completion and retry | |
1393 | * without disrupting NUMA hinting information. Do not relock and | |
1394 | * check_same as the page may no longer be mapped. | |
1395 | */ | |
82b0f8c3 JK |
1396 | if (unlikely(pmd_trans_migrating(*vmf->pmd))) { |
1397 | page = pmd_page(*vmf->pmd); | |
3c226c63 MR |
1398 | if (!get_page_unless_zero(page)) |
1399 | goto out_unlock; | |
82b0f8c3 | 1400 | spin_unlock(vmf->ptl); |
9a1ea439 | 1401 | put_and_wait_on_page_locked(page); |
de466bd6 MG |
1402 | goto out; |
1403 | } | |
1404 | ||
d10e63f2 | 1405 | page = pmd_page(pmd); |
a1a46184 | 1406 | BUG_ON(is_huge_zero_page(page)); |
8191acbd | 1407 | page_nid = page_to_nid(page); |
90572890 | 1408 | last_cpupid = page_cpupid_last(page); |
03c5a6e1 | 1409 | count_vm_numa_event(NUMA_HINT_FAULTS); |
04bb2f94 | 1410 | if (page_nid == this_nid) { |
03c5a6e1 | 1411 | count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); |
04bb2f94 RR |
1412 | flags |= TNF_FAULT_LOCAL; |
1413 | } | |
4daae3b4 | 1414 | |
bea66fbd | 1415 | /* See similar comment in do_numa_page for explanation */ |
288bc549 | 1416 | if (!pmd_savedwrite(pmd)) |
6688cc05 PZ |
1417 | flags |= TNF_NO_GROUP; |
1418 | ||
ff9042b1 MG |
1419 | /* |
1420 | * Acquire the page lock to serialise THP migrations but avoid dropping | |
1421 | * page_table_lock if at all possible | |
1422 | */ | |
b8916634 MG |
1423 | page_locked = trylock_page(page); |
1424 | target_nid = mpol_misplaced(page, vma, haddr); | |
98fa15f3 | 1425 | if (target_nid == NUMA_NO_NODE) { |
b8916634 | 1426 | /* If the page was locked, there are no parallel migrations */ |
a54a407f | 1427 | if (page_locked) |
b8916634 | 1428 | goto clear_pmdnuma; |
2b4847e7 | 1429 | } |
4daae3b4 | 1430 | |
de466bd6 | 1431 | /* Migration could have started since the pmd_trans_migrating check */ |
2b4847e7 | 1432 | if (!page_locked) { |
98fa15f3 | 1433 | page_nid = NUMA_NO_NODE; |
3c226c63 MR |
1434 | if (!get_page_unless_zero(page)) |
1435 | goto out_unlock; | |
82b0f8c3 | 1436 | spin_unlock(vmf->ptl); |
9a1ea439 | 1437 | put_and_wait_on_page_locked(page); |
b8916634 MG |
1438 | goto out; |
1439 | } | |
1440 | ||
2b4847e7 MG |
1441 | /* |
1442 | * Page is misplaced. Page lock serialises migrations. Acquire anon_vma | |
1443 | * to serialises splits | |
1444 | */ | |
b8916634 | 1445 | get_page(page); |
82b0f8c3 | 1446 | spin_unlock(vmf->ptl); |
b8916634 | 1447 | anon_vma = page_lock_anon_vma_read(page); |
4daae3b4 | 1448 | |
c69307d5 | 1449 | /* Confirm the PMD did not change while page_table_lock was released */ |
82b0f8c3 JK |
1450 | spin_lock(vmf->ptl); |
1451 | if (unlikely(!pmd_same(pmd, *vmf->pmd))) { | |
b32967ff MG |
1452 | unlock_page(page); |
1453 | put_page(page); | |
98fa15f3 | 1454 | page_nid = NUMA_NO_NODE; |
4daae3b4 | 1455 | goto out_unlock; |
b32967ff | 1456 | } |
ff9042b1 | 1457 | |
c3a489ca MG |
1458 | /* Bail if we fail to protect against THP splits for any reason */ |
1459 | if (unlikely(!anon_vma)) { | |
1460 | put_page(page); | |
98fa15f3 | 1461 | page_nid = NUMA_NO_NODE; |
c3a489ca MG |
1462 | goto clear_pmdnuma; |
1463 | } | |
1464 | ||
8b1b436d PZ |
1465 | /* |
1466 | * Since we took the NUMA fault, we must have observed the !accessible | |
1467 | * bit. Make sure all other CPUs agree with that, to avoid them | |
1468 | * modifying the page we're about to migrate. | |
1469 | * | |
1470 | * Must be done under PTL such that we'll observe the relevant | |
ccde85ba PZ |
1471 | * inc_tlb_flush_pending(). |
1472 | * | |
1473 | * We are not sure a pending tlb flush here is for a huge page | |
1474 | * mapping or not. Hence use the tlb range variant | |
8b1b436d | 1475 | */ |
7066f0f9 | 1476 | if (mm_tlb_flush_pending(vma->vm_mm)) { |
ccde85ba | 1477 | flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE); |
7066f0f9 AA |
1478 | /* |
1479 | * change_huge_pmd() released the pmd lock before | |
1480 | * invalidating the secondary MMUs sharing the primary | |
1481 | * MMU pagetables (with ->invalidate_range()). The | |
1482 | * mmu_notifier_invalidate_range_end() (which | |
1483 | * internally calls ->invalidate_range()) in | |
1484 | * change_pmd_range() will run after us, so we can't | |
1485 | * rely on it here and we need an explicit invalidate. | |
1486 | */ | |
1487 | mmu_notifier_invalidate_range(vma->vm_mm, haddr, | |
1488 | haddr + HPAGE_PMD_SIZE); | |
1489 | } | |
8b1b436d | 1490 | |
a54a407f MG |
1491 | /* |
1492 | * Migrate the THP to the requested node, returns with page unlocked | |
8a0516ed | 1493 | * and access rights restored. |
a54a407f | 1494 | */ |
82b0f8c3 | 1495 | spin_unlock(vmf->ptl); |
8b1b436d | 1496 | |
bae473a4 | 1497 | migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma, |
82b0f8c3 | 1498 | vmf->pmd, pmd, vmf->address, page, target_nid); |
6688cc05 PZ |
1499 | if (migrated) { |
1500 | flags |= TNF_MIGRATED; | |
8191acbd | 1501 | page_nid = target_nid; |
074c2381 MG |
1502 | } else |
1503 | flags |= TNF_MIGRATE_FAIL; | |
b32967ff | 1504 | |
8191acbd | 1505 | goto out; |
b32967ff | 1506 | clear_pmdnuma: |
a54a407f | 1507 | BUG_ON(!PageLocked(page)); |
288bc549 | 1508 | was_writable = pmd_savedwrite(pmd); |
4d942466 | 1509 | pmd = pmd_modify(pmd, vma->vm_page_prot); |
b7b04004 | 1510 | pmd = pmd_mkyoung(pmd); |
b191f9b1 MG |
1511 | if (was_writable) |
1512 | pmd = pmd_mkwrite(pmd); | |
82b0f8c3 JK |
1513 | set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd); |
1514 | update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); | |
a54a407f | 1515 | unlock_page(page); |
d10e63f2 | 1516 | out_unlock: |
82b0f8c3 | 1517 | spin_unlock(vmf->ptl); |
b8916634 MG |
1518 | |
1519 | out: | |
1520 | if (anon_vma) | |
1521 | page_unlock_anon_vma_read(anon_vma); | |
1522 | ||
98fa15f3 | 1523 | if (page_nid != NUMA_NO_NODE) |
82b0f8c3 | 1524 | task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, |
9a8b300f | 1525 | flags); |
8191acbd | 1526 | |
d10e63f2 MG |
1527 | return 0; |
1528 | } | |
1529 | ||
319904ad HY |
1530 | /* |
1531 | * Return true if we do MADV_FREE successfully on entire pmd page. | |
1532 | * Otherwise, return false. | |
1533 | */ | |
1534 | bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, | |
b8d3c4c3 | 1535 | pmd_t *pmd, unsigned long addr, unsigned long next) |
b8d3c4c3 MK |
1536 | { |
1537 | spinlock_t *ptl; | |
1538 | pmd_t orig_pmd; | |
1539 | struct page *page; | |
1540 | struct mm_struct *mm = tlb->mm; | |
319904ad | 1541 | bool ret = false; |
b8d3c4c3 | 1542 | |
ed6a7935 | 1543 | tlb_change_page_size(tlb, HPAGE_PMD_SIZE); |
07e32661 | 1544 | |
b6ec57f4 KS |
1545 | ptl = pmd_trans_huge_lock(pmd, vma); |
1546 | if (!ptl) | |
25eedabe | 1547 | goto out_unlocked; |
b8d3c4c3 MK |
1548 | |
1549 | orig_pmd = *pmd; | |
319904ad | 1550 | if (is_huge_zero_pmd(orig_pmd)) |
b8d3c4c3 | 1551 | goto out; |
b8d3c4c3 | 1552 | |
84c3fc4e ZY |
1553 | if (unlikely(!pmd_present(orig_pmd))) { |
1554 | VM_BUG_ON(thp_migration_supported() && | |
1555 | !is_pmd_migration_entry(orig_pmd)); | |
1556 | goto out; | |
1557 | } | |
1558 | ||
b8d3c4c3 MK |
1559 | page = pmd_page(orig_pmd); |
1560 | /* | |
1561 | * If other processes are mapping this page, we couldn't discard | |
1562 | * the page unless they all do MADV_FREE so let's skip the page. | |
1563 | */ | |
1564 | if (page_mapcount(page) != 1) | |
1565 | goto out; | |
1566 | ||
1567 | if (!trylock_page(page)) | |
1568 | goto out; | |
1569 | ||
1570 | /* | |
1571 | * If user want to discard part-pages of THP, split it so MADV_FREE | |
1572 | * will deactivate only them. | |
1573 | */ | |
1574 | if (next - addr != HPAGE_PMD_SIZE) { | |
1575 | get_page(page); | |
1576 | spin_unlock(ptl); | |
9818b8cd | 1577 | split_huge_page(page); |
b8d3c4c3 | 1578 | unlock_page(page); |
bbf29ffc | 1579 | put_page(page); |
b8d3c4c3 MK |
1580 | goto out_unlocked; |
1581 | } | |
1582 | ||
1583 | if (PageDirty(page)) | |
1584 | ClearPageDirty(page); | |
1585 | unlock_page(page); | |
1586 | ||
b8d3c4c3 | 1587 | if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { |
58ceeb6b | 1588 | pmdp_invalidate(vma, addr, pmd); |
b8d3c4c3 MK |
1589 | orig_pmd = pmd_mkold(orig_pmd); |
1590 | orig_pmd = pmd_mkclean(orig_pmd); | |
1591 | ||
1592 | set_pmd_at(mm, addr, pmd, orig_pmd); | |
1593 | tlb_remove_pmd_tlb_entry(tlb, pmd, addr); | |
1594 | } | |
802a3a92 SL |
1595 | |
1596 | mark_page_lazyfree(page); | |
319904ad | 1597 | ret = true; |
b8d3c4c3 MK |
1598 | out: |
1599 | spin_unlock(ptl); | |
1600 | out_unlocked: | |
1601 | return ret; | |
1602 | } | |
1603 | ||
953c66c2 AK |
1604 | static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd) |
1605 | { | |
1606 | pgtable_t pgtable; | |
1607 | ||
1608 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); | |
1609 | pte_free(mm, pgtable); | |
c4812909 | 1610 | mm_dec_nr_ptes(mm); |
953c66c2 AK |
1611 | } |
1612 | ||
71e3aac0 | 1613 | int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, |
f21760b1 | 1614 | pmd_t *pmd, unsigned long addr) |
71e3aac0 | 1615 | { |
da146769 | 1616 | pmd_t orig_pmd; |
bf929152 | 1617 | spinlock_t *ptl; |
71e3aac0 | 1618 | |
ed6a7935 | 1619 | tlb_change_page_size(tlb, HPAGE_PMD_SIZE); |
07e32661 | 1620 | |
b6ec57f4 KS |
1621 | ptl = __pmd_trans_huge_lock(pmd, vma); |
1622 | if (!ptl) | |
da146769 KS |
1623 | return 0; |
1624 | /* | |
1625 | * For architectures like ppc64 we look at deposited pgtable | |
1626 | * when calling pmdp_huge_get_and_clear. So do the | |
1627 | * pgtable_trans_huge_withdraw after finishing pmdp related | |
1628 | * operations. | |
1629 | */ | |
93a98695 AK |
1630 | orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd, |
1631 | tlb->fullmm); | |
da146769 | 1632 | tlb_remove_pmd_tlb_entry(tlb, pmd, addr); |
2484ca9b | 1633 | if (vma_is_special_huge(vma)) { |
3b6521f5 OH |
1634 | if (arch_needs_pgtable_deposit()) |
1635 | zap_deposited_table(tlb->mm, pmd); | |
da146769 KS |
1636 | spin_unlock(ptl); |
1637 | if (is_huge_zero_pmd(orig_pmd)) | |
c0f2e176 | 1638 | tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); |
da146769 | 1639 | } else if (is_huge_zero_pmd(orig_pmd)) { |
c14a6eb4 | 1640 | zap_deposited_table(tlb->mm, pmd); |
da146769 | 1641 | spin_unlock(ptl); |
c0f2e176 | 1642 | tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); |
da146769 | 1643 | } else { |
616b8371 ZY |
1644 | struct page *page = NULL; |
1645 | int flush_needed = 1; | |
1646 | ||
1647 | if (pmd_present(orig_pmd)) { | |
1648 | page = pmd_page(orig_pmd); | |
1649 | page_remove_rmap(page, true); | |
1650 | VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); | |
1651 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
1652 | } else if (thp_migration_supported()) { | |
1653 | swp_entry_t entry; | |
1654 | ||
1655 | VM_BUG_ON(!is_pmd_migration_entry(orig_pmd)); | |
1656 | entry = pmd_to_swp_entry(orig_pmd); | |
1657 | page = pfn_to_page(swp_offset(entry)); | |
1658 | flush_needed = 0; | |
1659 | } else | |
1660 | WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!"); | |
1661 | ||
b5072380 | 1662 | if (PageAnon(page)) { |
c14a6eb4 | 1663 | zap_deposited_table(tlb->mm, pmd); |
b5072380 KS |
1664 | add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); |
1665 | } else { | |
953c66c2 AK |
1666 | if (arch_needs_pgtable_deposit()) |
1667 | zap_deposited_table(tlb->mm, pmd); | |
fadae295 | 1668 | add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR); |
b5072380 | 1669 | } |
616b8371 | 1670 | |
da146769 | 1671 | spin_unlock(ptl); |
616b8371 ZY |
1672 | if (flush_needed) |
1673 | tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE); | |
025c5b24 | 1674 | } |
da146769 | 1675 | return 1; |
71e3aac0 AA |
1676 | } |
1677 | ||
1dd38b6c AK |
1678 | #ifndef pmd_move_must_withdraw |
1679 | static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl, | |
1680 | spinlock_t *old_pmd_ptl, | |
1681 | struct vm_area_struct *vma) | |
1682 | { | |
1683 | /* | |
1684 | * With split pmd lock we also need to move preallocated | |
1685 | * PTE page table if new_pmd is on different PMD page table. | |
1686 | * | |
1687 | * We also don't deposit and withdraw tables for file pages. | |
1688 | */ | |
1689 | return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma); | |
1690 | } | |
1691 | #endif | |
1692 | ||
ab6e3d09 NH |
1693 | static pmd_t move_soft_dirty_pmd(pmd_t pmd) |
1694 | { | |
1695 | #ifdef CONFIG_MEM_SOFT_DIRTY | |
1696 | if (unlikely(is_pmd_migration_entry(pmd))) | |
1697 | pmd = pmd_swp_mksoft_dirty(pmd); | |
1698 | else if (pmd_present(pmd)) | |
1699 | pmd = pmd_mksoft_dirty(pmd); | |
1700 | #endif | |
1701 | return pmd; | |
1702 | } | |
1703 | ||
bf8616d5 | 1704 | bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, |
b8aa9d9d | 1705 | unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd) |
37a1c49a | 1706 | { |
bf929152 | 1707 | spinlock_t *old_ptl, *new_ptl; |
37a1c49a | 1708 | pmd_t pmd; |
37a1c49a | 1709 | struct mm_struct *mm = vma->vm_mm; |
5d190420 | 1710 | bool force_flush = false; |
37a1c49a | 1711 | |
37a1c49a AA |
1712 | /* |
1713 | * The destination pmd shouldn't be established, free_pgtables() | |
1714 | * should have release it. | |
1715 | */ | |
1716 | if (WARN_ON(!pmd_none(*new_pmd))) { | |
1717 | VM_BUG_ON(pmd_trans_huge(*new_pmd)); | |
4b471e88 | 1718 | return false; |
37a1c49a AA |
1719 | } |
1720 | ||
bf929152 KS |
1721 | /* |
1722 | * We don't have to worry about the ordering of src and dst | |
c1e8d7c6 | 1723 | * ptlocks because exclusive mmap_lock prevents deadlock. |
bf929152 | 1724 | */ |
b6ec57f4 KS |
1725 | old_ptl = __pmd_trans_huge_lock(old_pmd, vma); |
1726 | if (old_ptl) { | |
bf929152 KS |
1727 | new_ptl = pmd_lockptr(mm, new_pmd); |
1728 | if (new_ptl != old_ptl) | |
1729 | spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); | |
8809aa2d | 1730 | pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); |
eb66ae03 | 1731 | if (pmd_present(pmd)) |
a2ce2666 | 1732 | force_flush = true; |
025c5b24 | 1733 | VM_BUG_ON(!pmd_none(*new_pmd)); |
3592806c | 1734 | |
1dd38b6c | 1735 | if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) { |
b3084f4d | 1736 | pgtable_t pgtable; |
3592806c KS |
1737 | pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); |
1738 | pgtable_trans_huge_deposit(mm, new_pmd, pgtable); | |
3592806c | 1739 | } |
ab6e3d09 NH |
1740 | pmd = move_soft_dirty_pmd(pmd); |
1741 | set_pmd_at(mm, new_addr, new_pmd, pmd); | |
5d190420 AL |
1742 | if (force_flush) |
1743 | flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE); | |
eb66ae03 LT |
1744 | if (new_ptl != old_ptl) |
1745 | spin_unlock(new_ptl); | |
bf929152 | 1746 | spin_unlock(old_ptl); |
4b471e88 | 1747 | return true; |
37a1c49a | 1748 | } |
4b471e88 | 1749 | return false; |
37a1c49a AA |
1750 | } |
1751 | ||
f123d74a MG |
1752 | /* |
1753 | * Returns | |
1754 | * - 0 if PMD could not be locked | |
1755 | * - 1 if PMD was locked but protections unchange and TLB flush unnecessary | |
1756 | * - HPAGE_PMD_NR is protections changed and TLB flush necessary | |
1757 | */ | |
cd7548ab | 1758 | int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, |
58705444 | 1759 | unsigned long addr, pgprot_t newprot, unsigned long cp_flags) |
cd7548ab JW |
1760 | { |
1761 | struct mm_struct *mm = vma->vm_mm; | |
bf929152 | 1762 | spinlock_t *ptl; |
0a85e51d KS |
1763 | pmd_t entry; |
1764 | bool preserve_write; | |
1765 | int ret; | |
58705444 | 1766 | bool prot_numa = cp_flags & MM_CP_PROT_NUMA; |
292924b2 PX |
1767 | bool uffd_wp = cp_flags & MM_CP_UFFD_WP; |
1768 | bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; | |
cd7548ab | 1769 | |
b6ec57f4 | 1770 | ptl = __pmd_trans_huge_lock(pmd, vma); |
0a85e51d KS |
1771 | if (!ptl) |
1772 | return 0; | |
e944fd67 | 1773 | |
0a85e51d KS |
1774 | preserve_write = prot_numa && pmd_write(*pmd); |
1775 | ret = 1; | |
e944fd67 | 1776 | |
84c3fc4e ZY |
1777 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
1778 | if (is_swap_pmd(*pmd)) { | |
1779 | swp_entry_t entry = pmd_to_swp_entry(*pmd); | |
1780 | ||
1781 | VM_BUG_ON(!is_pmd_migration_entry(*pmd)); | |
1782 | if (is_write_migration_entry(entry)) { | |
1783 | pmd_t newpmd; | |
1784 | /* | |
1785 | * A protection check is difficult so | |
1786 | * just be safe and disable write | |
1787 | */ | |
1788 | make_migration_entry_read(&entry); | |
1789 | newpmd = swp_entry_to_pmd(entry); | |
ab6e3d09 NH |
1790 | if (pmd_swp_soft_dirty(*pmd)) |
1791 | newpmd = pmd_swp_mksoft_dirty(newpmd); | |
84c3fc4e ZY |
1792 | set_pmd_at(mm, addr, pmd, newpmd); |
1793 | } | |
1794 | goto unlock; | |
1795 | } | |
1796 | #endif | |
1797 | ||
0a85e51d KS |
1798 | /* |
1799 | * Avoid trapping faults against the zero page. The read-only | |
1800 | * data is likely to be read-cached on the local CPU and | |
1801 | * local/remote hits to the zero page are not interesting. | |
1802 | */ | |
1803 | if (prot_numa && is_huge_zero_pmd(*pmd)) | |
1804 | goto unlock; | |
025c5b24 | 1805 | |
0a85e51d KS |
1806 | if (prot_numa && pmd_protnone(*pmd)) |
1807 | goto unlock; | |
1808 | ||
ced10803 | 1809 | /* |
3e4e28c5 | 1810 | * In case prot_numa, we are under mmap_read_lock(mm). It's critical |
ced10803 | 1811 | * to not clear pmd intermittently to avoid race with MADV_DONTNEED |
3e4e28c5 | 1812 | * which is also under mmap_read_lock(mm): |
ced10803 KS |
1813 | * |
1814 | * CPU0: CPU1: | |
1815 | * change_huge_pmd(prot_numa=1) | |
1816 | * pmdp_huge_get_and_clear_notify() | |
1817 | * madvise_dontneed() | |
1818 | * zap_pmd_range() | |
1819 | * pmd_trans_huge(*pmd) == 0 (without ptl) | |
1820 | * // skip the pmd | |
1821 | * set_pmd_at(); | |
1822 | * // pmd is re-established | |
1823 | * | |
1824 | * The race makes MADV_DONTNEED miss the huge pmd and don't clear it | |
1825 | * which may break userspace. | |
1826 | * | |
1827 | * pmdp_invalidate() is required to make sure we don't miss | |
1828 | * dirty/young flags set by hardware. | |
1829 | */ | |
a3cf988f | 1830 | entry = pmdp_invalidate(vma, addr, pmd); |
ced10803 | 1831 | |
0a85e51d KS |
1832 | entry = pmd_modify(entry, newprot); |
1833 | if (preserve_write) | |
1834 | entry = pmd_mk_savedwrite(entry); | |
292924b2 PX |
1835 | if (uffd_wp) { |
1836 | entry = pmd_wrprotect(entry); | |
1837 | entry = pmd_mkuffd_wp(entry); | |
1838 | } else if (uffd_wp_resolve) { | |
1839 | /* | |
1840 | * Leave the write bit to be handled by PF interrupt | |
1841 | * handler, then things like COW could be properly | |
1842 | * handled. | |
1843 | */ | |
1844 | entry = pmd_clear_uffd_wp(entry); | |
1845 | } | |
0a85e51d KS |
1846 | ret = HPAGE_PMD_NR; |
1847 | set_pmd_at(mm, addr, pmd, entry); | |
1848 | BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry)); | |
1849 | unlock: | |
1850 | spin_unlock(ptl); | |
025c5b24 NH |
1851 | return ret; |
1852 | } | |
1853 | ||
1854 | /* | |
8f19b0c0 | 1855 | * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise. |
025c5b24 | 1856 | * |
8f19b0c0 HY |
1857 | * Note that if it returns page table lock pointer, this routine returns without |
1858 | * unlocking page table lock. So callers must unlock it. | |
025c5b24 | 1859 | */ |
b6ec57f4 | 1860 | spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) |
025c5b24 | 1861 | { |
b6ec57f4 KS |
1862 | spinlock_t *ptl; |
1863 | ptl = pmd_lock(vma->vm_mm, pmd); | |
84c3fc4e ZY |
1864 | if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || |
1865 | pmd_devmap(*pmd))) | |
b6ec57f4 KS |
1866 | return ptl; |
1867 | spin_unlock(ptl); | |
1868 | return NULL; | |
cd7548ab JW |
1869 | } |
1870 | ||
a00cc7d9 MW |
1871 | /* |
1872 | * Returns true if a given pud maps a thp, false otherwise. | |
1873 | * | |
1874 | * Note that if it returns true, this routine returns without unlocking page | |
1875 | * table lock. So callers must unlock it. | |
1876 | */ | |
1877 | spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma) | |
1878 | { | |
1879 | spinlock_t *ptl; | |
1880 | ||
1881 | ptl = pud_lock(vma->vm_mm, pud); | |
1882 | if (likely(pud_trans_huge(*pud) || pud_devmap(*pud))) | |
1883 | return ptl; | |
1884 | spin_unlock(ptl); | |
1885 | return NULL; | |
1886 | } | |
1887 | ||
1888 | #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD | |
1889 | int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, | |
1890 | pud_t *pud, unsigned long addr) | |
1891 | { | |
a00cc7d9 MW |
1892 | spinlock_t *ptl; |
1893 | ||
1894 | ptl = __pud_trans_huge_lock(pud, vma); | |
1895 | if (!ptl) | |
1896 | return 0; | |
1897 | /* | |
1898 | * For architectures like ppc64 we look at deposited pgtable | |
1899 | * when calling pudp_huge_get_and_clear. So do the | |
1900 | * pgtable_trans_huge_withdraw after finishing pudp related | |
1901 | * operations. | |
1902 | */ | |
70516b93 | 1903 | pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm); |
a00cc7d9 | 1904 | tlb_remove_pud_tlb_entry(tlb, pud, addr); |
2484ca9b | 1905 | if (vma_is_special_huge(vma)) { |
a00cc7d9 MW |
1906 | spin_unlock(ptl); |
1907 | /* No zero page support yet */ | |
1908 | } else { | |
1909 | /* No support for anonymous PUD pages yet */ | |
1910 | BUG(); | |
1911 | } | |
1912 | return 1; | |
1913 | } | |
1914 | ||
1915 | static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud, | |
1916 | unsigned long haddr) | |
1917 | { | |
1918 | VM_BUG_ON(haddr & ~HPAGE_PUD_MASK); | |
1919 | VM_BUG_ON_VMA(vma->vm_start > haddr, vma); | |
1920 | VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma); | |
1921 | VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud)); | |
1922 | ||
ce9311cf | 1923 | count_vm_event(THP_SPLIT_PUD); |
a00cc7d9 MW |
1924 | |
1925 | pudp_huge_clear_flush_notify(vma, haddr, pud); | |
1926 | } | |
1927 | ||
1928 | void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud, | |
1929 | unsigned long address) | |
1930 | { | |
1931 | spinlock_t *ptl; | |
ac46d4f3 | 1932 | struct mmu_notifier_range range; |
a00cc7d9 | 1933 | |
7269f999 | 1934 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, |
6f4f13e8 | 1935 | address & HPAGE_PUD_MASK, |
ac46d4f3 JG |
1936 | (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE); |
1937 | mmu_notifier_invalidate_range_start(&range); | |
1938 | ptl = pud_lock(vma->vm_mm, pud); | |
a00cc7d9 MW |
1939 | if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud))) |
1940 | goto out; | |
ac46d4f3 | 1941 | __split_huge_pud_locked(vma, pud, range.start); |
a00cc7d9 MW |
1942 | |
1943 | out: | |
1944 | spin_unlock(ptl); | |
4645b9fe JG |
1945 | /* |
1946 | * No need to double call mmu_notifier->invalidate_range() callback as | |
1947 | * the above pudp_huge_clear_flush_notify() did already call it. | |
1948 | */ | |
ac46d4f3 | 1949 | mmu_notifier_invalidate_range_only_end(&range); |
a00cc7d9 MW |
1950 | } |
1951 | #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ | |
1952 | ||
eef1b3ba KS |
1953 | static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, |
1954 | unsigned long haddr, pmd_t *pmd) | |
1955 | { | |
1956 | struct mm_struct *mm = vma->vm_mm; | |
1957 | pgtable_t pgtable; | |
1958 | pmd_t _pmd; | |
1959 | int i; | |
1960 | ||
0f10851e JG |
1961 | /* |
1962 | * Leave pmd empty until pte is filled note that it is fine to delay | |
1963 | * notification until mmu_notifier_invalidate_range_end() as we are | |
1964 | * replacing a zero pmd write protected page with a zero pte write | |
1965 | * protected page. | |
1966 | * | |
ad56b738 | 1967 | * See Documentation/vm/mmu_notifier.rst |
0f10851e JG |
1968 | */ |
1969 | pmdp_huge_clear_flush(vma, haddr, pmd); | |
eef1b3ba KS |
1970 | |
1971 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); | |
1972 | pmd_populate(mm, &_pmd, pgtable); | |
1973 | ||
1974 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
1975 | pte_t *pte, entry; | |
1976 | entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); | |
1977 | entry = pte_mkspecial(entry); | |
1978 | pte = pte_offset_map(&_pmd, haddr); | |
1979 | VM_BUG_ON(!pte_none(*pte)); | |
1980 | set_pte_at(mm, haddr, pte, entry); | |
1981 | pte_unmap(pte); | |
1982 | } | |
1983 | smp_wmb(); /* make pte visible before pmd */ | |
1984 | pmd_populate(mm, pmd, pgtable); | |
eef1b3ba KS |
1985 | } |
1986 | ||
1987 | static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, | |
ba988280 | 1988 | unsigned long haddr, bool freeze) |
eef1b3ba KS |
1989 | { |
1990 | struct mm_struct *mm = vma->vm_mm; | |
1991 | struct page *page; | |
1992 | pgtable_t pgtable; | |
423ac9af | 1993 | pmd_t old_pmd, _pmd; |
292924b2 | 1994 | bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false; |
2ac015e2 | 1995 | unsigned long addr; |
eef1b3ba KS |
1996 | int i; |
1997 | ||
1998 | VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); | |
1999 | VM_BUG_ON_VMA(vma->vm_start > haddr, vma); | |
2000 | VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); | |
84c3fc4e ZY |
2001 | VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd) |
2002 | && !pmd_devmap(*pmd)); | |
eef1b3ba KS |
2003 | |
2004 | count_vm_event(THP_SPLIT_PMD); | |
2005 | ||
d21b9e57 KS |
2006 | if (!vma_is_anonymous(vma)) { |
2007 | _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); | |
953c66c2 AK |
2008 | /* |
2009 | * We are going to unmap this huge page. So | |
2010 | * just go ahead and zap it | |
2011 | */ | |
2012 | if (arch_needs_pgtable_deposit()) | |
2013 | zap_deposited_table(mm, pmd); | |
2484ca9b | 2014 | if (vma_is_special_huge(vma)) |
d21b9e57 KS |
2015 | return; |
2016 | page = pmd_page(_pmd); | |
e1f1b157 HD |
2017 | if (!PageDirty(page) && pmd_dirty(_pmd)) |
2018 | set_page_dirty(page); | |
d21b9e57 KS |
2019 | if (!PageReferenced(page) && pmd_young(_pmd)) |
2020 | SetPageReferenced(page); | |
2021 | page_remove_rmap(page, true); | |
2022 | put_page(page); | |
fadae295 | 2023 | add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR); |
eef1b3ba | 2024 | return; |
ec0abae6 | 2025 | } else if (pmd_trans_huge(*pmd) && is_huge_zero_pmd(*pmd)) { |
4645b9fe JG |
2026 | /* |
2027 | * FIXME: Do we want to invalidate secondary mmu by calling | |
2028 | * mmu_notifier_invalidate_range() see comments below inside | |
2029 | * __split_huge_pmd() ? | |
2030 | * | |
2031 | * We are going from a zero huge page write protected to zero | |
2032 | * small page also write protected so it does not seems useful | |
2033 | * to invalidate secondary mmu at this time. | |
2034 | */ | |
eef1b3ba KS |
2035 | return __split_huge_zero_page_pmd(vma, haddr, pmd); |
2036 | } | |
2037 | ||
423ac9af AK |
2038 | /* |
2039 | * Up to this point the pmd is present and huge and userland has the | |
2040 | * whole access to the hugepage during the split (which happens in | |
2041 | * place). If we overwrite the pmd with the not-huge version pointing | |
2042 | * to the pte here (which of course we could if all CPUs were bug | |
2043 | * free), userland could trigger a small page size TLB miss on the | |
2044 | * small sized TLB while the hugepage TLB entry is still established in | |
2045 | * the huge TLB. Some CPU doesn't like that. | |
42742d9b AK |
2046 | * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum |
2047 | * 383 on page 105. Intel should be safe but is also warns that it's | |
423ac9af AK |
2048 | * only safe if the permission and cache attributes of the two entries |
2049 | * loaded in the two TLB is identical (which should be the case here). | |
2050 | * But it is generally safer to never allow small and huge TLB entries | |
2051 | * for the same virtual address to be loaded simultaneously. So instead | |
2052 | * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the | |
2053 | * current pmd notpresent (atomically because here the pmd_trans_huge | |
2054 | * must remain set at all times on the pmd until the split is complete | |
2055 | * for this pmd), then we flush the SMP TLB and finally we write the | |
2056 | * non-huge version of the pmd entry with pmd_populate. | |
2057 | */ | |
2058 | old_pmd = pmdp_invalidate(vma, haddr, pmd); | |
2059 | ||
423ac9af | 2060 | pmd_migration = is_pmd_migration_entry(old_pmd); |
2e83ee1d | 2061 | if (unlikely(pmd_migration)) { |
84c3fc4e ZY |
2062 | swp_entry_t entry; |
2063 | ||
423ac9af | 2064 | entry = pmd_to_swp_entry(old_pmd); |
84c3fc4e | 2065 | page = pfn_to_page(swp_offset(entry)); |
2e83ee1d PX |
2066 | write = is_write_migration_entry(entry); |
2067 | young = false; | |
2068 | soft_dirty = pmd_swp_soft_dirty(old_pmd); | |
f45ec5ff | 2069 | uffd_wp = pmd_swp_uffd_wp(old_pmd); |
2e83ee1d | 2070 | } else { |
423ac9af | 2071 | page = pmd_page(old_pmd); |
2e83ee1d PX |
2072 | if (pmd_dirty(old_pmd)) |
2073 | SetPageDirty(page); | |
2074 | write = pmd_write(old_pmd); | |
2075 | young = pmd_young(old_pmd); | |
2076 | soft_dirty = pmd_soft_dirty(old_pmd); | |
292924b2 | 2077 | uffd_wp = pmd_uffd_wp(old_pmd); |
2e83ee1d | 2078 | } |
eef1b3ba | 2079 | VM_BUG_ON_PAGE(!page_count(page), page); |
fe896d18 | 2080 | page_ref_add(page, HPAGE_PMD_NR - 1); |
eef1b3ba | 2081 | |
423ac9af AK |
2082 | /* |
2083 | * Withdraw the table only after we mark the pmd entry invalid. | |
2084 | * This's critical for some architectures (Power). | |
2085 | */ | |
eef1b3ba KS |
2086 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); |
2087 | pmd_populate(mm, &_pmd, pgtable); | |
2088 | ||
2ac015e2 | 2089 | for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { |
eef1b3ba KS |
2090 | pte_t entry, *pte; |
2091 | /* | |
2092 | * Note that NUMA hinting access restrictions are not | |
2093 | * transferred to avoid any possibility of altering | |
2094 | * permissions across VMAs. | |
2095 | */ | |
84c3fc4e | 2096 | if (freeze || pmd_migration) { |
ba988280 KS |
2097 | swp_entry_t swp_entry; |
2098 | swp_entry = make_migration_entry(page + i, write); | |
2099 | entry = swp_entry_to_pte(swp_entry); | |
804dd150 AA |
2100 | if (soft_dirty) |
2101 | entry = pte_swp_mksoft_dirty(entry); | |
f45ec5ff PX |
2102 | if (uffd_wp) |
2103 | entry = pte_swp_mkuffd_wp(entry); | |
ba988280 | 2104 | } else { |
6d2329f8 | 2105 | entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot)); |
b8d3c4c3 | 2106 | entry = maybe_mkwrite(entry, vma); |
ba988280 KS |
2107 | if (!write) |
2108 | entry = pte_wrprotect(entry); | |
2109 | if (!young) | |
2110 | entry = pte_mkold(entry); | |
804dd150 AA |
2111 | if (soft_dirty) |
2112 | entry = pte_mksoft_dirty(entry); | |
292924b2 PX |
2113 | if (uffd_wp) |
2114 | entry = pte_mkuffd_wp(entry); | |
ba988280 | 2115 | } |
2ac015e2 | 2116 | pte = pte_offset_map(&_pmd, addr); |
eef1b3ba | 2117 | BUG_ON(!pte_none(*pte)); |
2ac015e2 | 2118 | set_pte_at(mm, addr, pte, entry); |
ec0abae6 | 2119 | if (!pmd_migration) |
eef1b3ba | 2120 | atomic_inc(&page[i]._mapcount); |
ec0abae6 | 2121 | pte_unmap(pte); |
eef1b3ba KS |
2122 | } |
2123 | ||
ec0abae6 RC |
2124 | if (!pmd_migration) { |
2125 | /* | |
2126 | * Set PG_double_map before dropping compound_mapcount to avoid | |
2127 | * false-negative page_mapped(). | |
2128 | */ | |
2129 | if (compound_mapcount(page) > 1 && | |
2130 | !TestSetPageDoubleMap(page)) { | |
eef1b3ba | 2131 | for (i = 0; i < HPAGE_PMD_NR; i++) |
ec0abae6 RC |
2132 | atomic_inc(&page[i]._mapcount); |
2133 | } | |
2134 | ||
2135 | lock_page_memcg(page); | |
2136 | if (atomic_add_negative(-1, compound_mapcount_ptr(page))) { | |
2137 | /* Last compound_mapcount is gone. */ | |
2138 | __dec_lruvec_page_state(page, NR_ANON_THPS); | |
2139 | if (TestClearPageDoubleMap(page)) { | |
2140 | /* No need in mapcount reference anymore */ | |
2141 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
2142 | atomic_dec(&page[i]._mapcount); | |
2143 | } | |
eef1b3ba | 2144 | } |
ec0abae6 | 2145 | unlock_page_memcg(page); |
eef1b3ba KS |
2146 | } |
2147 | ||
2148 | smp_wmb(); /* make pte visible before pmd */ | |
2149 | pmd_populate(mm, pmd, pgtable); | |
e9b61f19 KS |
2150 | |
2151 | if (freeze) { | |
2ac015e2 | 2152 | for (i = 0; i < HPAGE_PMD_NR; i++) { |
e9b61f19 KS |
2153 | page_remove_rmap(page + i, false); |
2154 | put_page(page + i); | |
2155 | } | |
2156 | } | |
eef1b3ba KS |
2157 | } |
2158 | ||
2159 | void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, | |
33f4751e | 2160 | unsigned long address, bool freeze, struct page *page) |
eef1b3ba KS |
2161 | { |
2162 | spinlock_t *ptl; | |
ac46d4f3 | 2163 | struct mmu_notifier_range range; |
c444eb56 AA |
2164 | bool was_locked = false; |
2165 | pmd_t _pmd; | |
eef1b3ba | 2166 | |
7269f999 | 2167 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, |
6f4f13e8 | 2168 | address & HPAGE_PMD_MASK, |
ac46d4f3 JG |
2169 | (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE); |
2170 | mmu_notifier_invalidate_range_start(&range); | |
2171 | ptl = pmd_lock(vma->vm_mm, pmd); | |
33f4751e NH |
2172 | |
2173 | /* | |
2174 | * If caller asks to setup a migration entries, we need a page to check | |
2175 | * pmd against. Otherwise we can end up replacing wrong page. | |
2176 | */ | |
2177 | VM_BUG_ON(freeze && !page); | |
c444eb56 AA |
2178 | if (page) { |
2179 | VM_WARN_ON_ONCE(!PageLocked(page)); | |
2180 | was_locked = true; | |
2181 | if (page != pmd_page(*pmd)) | |
2182 | goto out; | |
2183 | } | |
33f4751e | 2184 | |
c444eb56 | 2185 | repeat: |
5c7fb56e | 2186 | if (pmd_trans_huge(*pmd)) { |
c444eb56 AA |
2187 | if (!page) { |
2188 | page = pmd_page(*pmd); | |
2189 | if (unlikely(!trylock_page(page))) { | |
2190 | get_page(page); | |
2191 | _pmd = *pmd; | |
2192 | spin_unlock(ptl); | |
2193 | lock_page(page); | |
2194 | spin_lock(ptl); | |
2195 | if (unlikely(!pmd_same(*pmd, _pmd))) { | |
2196 | unlock_page(page); | |
2197 | put_page(page); | |
2198 | page = NULL; | |
2199 | goto repeat; | |
2200 | } | |
2201 | put_page(page); | |
2202 | } | |
2203 | } | |
5c7fb56e | 2204 | if (PageMlocked(page)) |
5f737714 | 2205 | clear_page_mlock(page); |
84c3fc4e | 2206 | } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd))) |
e90309c9 | 2207 | goto out; |
ac46d4f3 | 2208 | __split_huge_pmd_locked(vma, pmd, range.start, freeze); |
e90309c9 | 2209 | out: |
eef1b3ba | 2210 | spin_unlock(ptl); |
c444eb56 AA |
2211 | if (!was_locked && page) |
2212 | unlock_page(page); | |
4645b9fe JG |
2213 | /* |
2214 | * No need to double call mmu_notifier->invalidate_range() callback. | |
2215 | * They are 3 cases to consider inside __split_huge_pmd_locked(): | |
2216 | * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious | |
2217 | * 2) __split_huge_zero_page_pmd() read only zero page and any write | |
2218 | * fault will trigger a flush_notify before pointing to a new page | |
2219 | * (it is fine if the secondary mmu keeps pointing to the old zero | |
2220 | * page in the meantime) | |
2221 | * 3) Split a huge pmd into pte pointing to the same page. No need | |
2222 | * to invalidate secondary tlb entry they are all still valid. | |
2223 | * any further changes to individual pte will notify. So no need | |
2224 | * to call mmu_notifier->invalidate_range() | |
2225 | */ | |
ac46d4f3 | 2226 | mmu_notifier_invalidate_range_only_end(&range); |
eef1b3ba KS |
2227 | } |
2228 | ||
fec89c10 KS |
2229 | void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address, |
2230 | bool freeze, struct page *page) | |
94fcc585 | 2231 | { |
f72e7dcd | 2232 | pgd_t *pgd; |
c2febafc | 2233 | p4d_t *p4d; |
f72e7dcd | 2234 | pud_t *pud; |
94fcc585 AA |
2235 | pmd_t *pmd; |
2236 | ||
78ddc534 | 2237 | pgd = pgd_offset(vma->vm_mm, address); |
f72e7dcd HD |
2238 | if (!pgd_present(*pgd)) |
2239 | return; | |
2240 | ||
c2febafc KS |
2241 | p4d = p4d_offset(pgd, address); |
2242 | if (!p4d_present(*p4d)) | |
2243 | return; | |
2244 | ||
2245 | pud = pud_offset(p4d, address); | |
f72e7dcd HD |
2246 | if (!pud_present(*pud)) |
2247 | return; | |
2248 | ||
2249 | pmd = pmd_offset(pud, address); | |
fec89c10 | 2250 | |
33f4751e | 2251 | __split_huge_pmd(vma, pmd, address, freeze, page); |
94fcc585 AA |
2252 | } |
2253 | ||
e1b9996b | 2254 | void vma_adjust_trans_huge(struct vm_area_struct *vma, |
94fcc585 AA |
2255 | unsigned long start, |
2256 | unsigned long end, | |
2257 | long adjust_next) | |
2258 | { | |
2259 | /* | |
2260 | * If the new start address isn't hpage aligned and it could | |
2261 | * previously contain an hugepage: check if we need to split | |
2262 | * an huge pmd. | |
2263 | */ | |
2264 | if (start & ~HPAGE_PMD_MASK && | |
2265 | (start & HPAGE_PMD_MASK) >= vma->vm_start && | |
2266 | (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
fec89c10 | 2267 | split_huge_pmd_address(vma, start, false, NULL); |
94fcc585 AA |
2268 | |
2269 | /* | |
2270 | * If the new end address isn't hpage aligned and it could | |
2271 | * previously contain an hugepage: check if we need to split | |
2272 | * an huge pmd. | |
2273 | */ | |
2274 | if (end & ~HPAGE_PMD_MASK && | |
2275 | (end & HPAGE_PMD_MASK) >= vma->vm_start && | |
2276 | (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
fec89c10 | 2277 | split_huge_pmd_address(vma, end, false, NULL); |
94fcc585 AA |
2278 | |
2279 | /* | |
2280 | * If we're also updating the vma->vm_next->vm_start, if the new | |
2281 | * vm_next->vm_start isn't page aligned and it could previously | |
2282 | * contain an hugepage: check if we need to split an huge pmd. | |
2283 | */ | |
2284 | if (adjust_next > 0) { | |
2285 | struct vm_area_struct *next = vma->vm_next; | |
2286 | unsigned long nstart = next->vm_start; | |
2287 | nstart += adjust_next << PAGE_SHIFT; | |
2288 | if (nstart & ~HPAGE_PMD_MASK && | |
2289 | (nstart & HPAGE_PMD_MASK) >= next->vm_start && | |
2290 | (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) | |
fec89c10 | 2291 | split_huge_pmd_address(next, nstart, false, NULL); |
94fcc585 AA |
2292 | } |
2293 | } | |
e9b61f19 | 2294 | |
906f9cdf | 2295 | static void unmap_page(struct page *page) |
e9b61f19 | 2296 | { |
baa355fd | 2297 | enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS | |
c7ab0d2f | 2298 | TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD; |
666e5a40 | 2299 | bool unmap_success; |
e9b61f19 KS |
2300 | |
2301 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
2302 | ||
baa355fd | 2303 | if (PageAnon(page)) |
b5ff8161 | 2304 | ttu_flags |= TTU_SPLIT_FREEZE; |
baa355fd | 2305 | |
666e5a40 MK |
2306 | unmap_success = try_to_unmap(page, ttu_flags); |
2307 | VM_BUG_ON_PAGE(!unmap_success, page); | |
e9b61f19 KS |
2308 | } |
2309 | ||
906f9cdf | 2310 | static void remap_page(struct page *page) |
e9b61f19 | 2311 | { |
fec89c10 | 2312 | int i; |
ace71a19 KS |
2313 | if (PageTransHuge(page)) { |
2314 | remove_migration_ptes(page, page, true); | |
2315 | } else { | |
2316 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
2317 | remove_migration_ptes(page + i, page + i, true); | |
2318 | } | |
e9b61f19 KS |
2319 | } |
2320 | ||
8df651c7 | 2321 | static void __split_huge_page_tail(struct page *head, int tail, |
e9b61f19 KS |
2322 | struct lruvec *lruvec, struct list_head *list) |
2323 | { | |
e9b61f19 KS |
2324 | struct page *page_tail = head + tail; |
2325 | ||
8df651c7 | 2326 | VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail); |
e9b61f19 KS |
2327 | |
2328 | /* | |
605ca5ed KK |
2329 | * Clone page flags before unfreezing refcount. |
2330 | * | |
2331 | * After successful get_page_unless_zero() might follow flags change, | |
2332 | * for exmaple lock_page() which set PG_waiters. | |
e9b61f19 | 2333 | */ |
e9b61f19 KS |
2334 | page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; |
2335 | page_tail->flags |= (head->flags & | |
2336 | ((1L << PG_referenced) | | |
2337 | (1L << PG_swapbacked) | | |
38d8b4e6 | 2338 | (1L << PG_swapcache) | |
e9b61f19 KS |
2339 | (1L << PG_mlocked) | |
2340 | (1L << PG_uptodate) | | |
2341 | (1L << PG_active) | | |
1899ad18 | 2342 | (1L << PG_workingset) | |
e9b61f19 | 2343 | (1L << PG_locked) | |
b8d3c4c3 MK |
2344 | (1L << PG_unevictable) | |
2345 | (1L << PG_dirty))); | |
e9b61f19 | 2346 | |
173d9d9f HD |
2347 | /* ->mapping in first tail page is compound_mapcount */ |
2348 | VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING, | |
2349 | page_tail); | |
2350 | page_tail->mapping = head->mapping; | |
2351 | page_tail->index = head->index + tail; | |
2352 | ||
605ca5ed | 2353 | /* Page flags must be visible before we make the page non-compound. */ |
e9b61f19 KS |
2354 | smp_wmb(); |
2355 | ||
605ca5ed KK |
2356 | /* |
2357 | * Clear PageTail before unfreezing page refcount. | |
2358 | * | |
2359 | * After successful get_page_unless_zero() might follow put_page() | |
2360 | * which needs correct compound_head(). | |
2361 | */ | |
e9b61f19 KS |
2362 | clear_compound_head(page_tail); |
2363 | ||
605ca5ed KK |
2364 | /* Finally unfreeze refcount. Additional reference from page cache. */ |
2365 | page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) || | |
2366 | PageSwapCache(head))); | |
2367 | ||
e9b61f19 KS |
2368 | if (page_is_young(head)) |
2369 | set_page_young(page_tail); | |
2370 | if (page_is_idle(head)) | |
2371 | set_page_idle(page_tail); | |
2372 | ||
e9b61f19 | 2373 | page_cpupid_xchg_last(page_tail, page_cpupid_last(head)); |
94723aaf MH |
2374 | |
2375 | /* | |
2376 | * always add to the tail because some iterators expect new | |
2377 | * pages to show after the currently processed elements - e.g. | |
2378 | * migrate_pages | |
2379 | */ | |
e9b61f19 | 2380 | lru_add_page_tail(head, page_tail, lruvec, list); |
e9b61f19 KS |
2381 | } |
2382 | ||
baa355fd | 2383 | static void __split_huge_page(struct page *page, struct list_head *list, |
006d3ff2 | 2384 | pgoff_t end, unsigned long flags) |
e9b61f19 KS |
2385 | { |
2386 | struct page *head = compound_head(page); | |
f4b7e272 | 2387 | pg_data_t *pgdat = page_pgdat(head); |
e9b61f19 | 2388 | struct lruvec *lruvec; |
4101196b MWO |
2389 | struct address_space *swap_cache = NULL; |
2390 | unsigned long offset = 0; | |
8df651c7 | 2391 | int i; |
e9b61f19 | 2392 | |
f4b7e272 | 2393 | lruvec = mem_cgroup_page_lruvec(head, pgdat); |
e9b61f19 KS |
2394 | |
2395 | /* complete memcg works before add pages to LRU */ | |
2396 | mem_cgroup_split_huge_fixup(head); | |
2397 | ||
4101196b MWO |
2398 | if (PageAnon(head) && PageSwapCache(head)) { |
2399 | swp_entry_t entry = { .val = page_private(head) }; | |
2400 | ||
2401 | offset = swp_offset(entry); | |
2402 | swap_cache = swap_address_space(entry); | |
2403 | xa_lock(&swap_cache->i_pages); | |
2404 | } | |
2405 | ||
baa355fd | 2406 | for (i = HPAGE_PMD_NR - 1; i >= 1; i--) { |
8df651c7 | 2407 | __split_huge_page_tail(head, i, lruvec, list); |
baa355fd KS |
2408 | /* Some pages can be beyond i_size: drop them from page cache */ |
2409 | if (head[i].index >= end) { | |
2d077d4b | 2410 | ClearPageDirty(head + i); |
baa355fd | 2411 | __delete_from_page_cache(head + i, NULL); |
800d8c63 KS |
2412 | if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head)) |
2413 | shmem_uncharge(head->mapping->host, 1); | |
baa355fd | 2414 | put_page(head + i); |
4101196b MWO |
2415 | } else if (!PageAnon(page)) { |
2416 | __xa_store(&head->mapping->i_pages, head[i].index, | |
2417 | head + i, 0); | |
2418 | } else if (swap_cache) { | |
2419 | __xa_store(&swap_cache->i_pages, offset + i, | |
2420 | head + i, 0); | |
baa355fd KS |
2421 | } |
2422 | } | |
e9b61f19 KS |
2423 | |
2424 | ClearPageCompound(head); | |
f7da677b VB |
2425 | |
2426 | split_page_owner(head, HPAGE_PMD_ORDER); | |
2427 | ||
baa355fd KS |
2428 | /* See comment in __split_huge_page_tail() */ |
2429 | if (PageAnon(head)) { | |
aa5dc07f | 2430 | /* Additional pin to swap cache */ |
4101196b | 2431 | if (PageSwapCache(head)) { |
38d8b4e6 | 2432 | page_ref_add(head, 2); |
4101196b MWO |
2433 | xa_unlock(&swap_cache->i_pages); |
2434 | } else { | |
38d8b4e6 | 2435 | page_ref_inc(head); |
4101196b | 2436 | } |
baa355fd | 2437 | } else { |
aa5dc07f | 2438 | /* Additional pin to page cache */ |
baa355fd | 2439 | page_ref_add(head, 2); |
b93b0163 | 2440 | xa_unlock(&head->mapping->i_pages); |
baa355fd KS |
2441 | } |
2442 | ||
f4b7e272 | 2443 | spin_unlock_irqrestore(&pgdat->lru_lock, flags); |
e9b61f19 | 2444 | |
906f9cdf | 2445 | remap_page(head); |
e9b61f19 KS |
2446 | |
2447 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
2448 | struct page *subpage = head + i; | |
2449 | if (subpage == page) | |
2450 | continue; | |
2451 | unlock_page(subpage); | |
2452 | ||
2453 | /* | |
2454 | * Subpages may be freed if there wasn't any mapping | |
2455 | * like if add_to_swap() is running on a lru page that | |
2456 | * had its mapping zapped. And freeing these pages | |
2457 | * requires taking the lru_lock so we do the put_page | |
2458 | * of the tail pages after the split is complete. | |
2459 | */ | |
2460 | put_page(subpage); | |
2461 | } | |
2462 | } | |
2463 | ||
b20ce5e0 KS |
2464 | int total_mapcount(struct page *page) |
2465 | { | |
dd78fedd | 2466 | int i, compound, ret; |
b20ce5e0 KS |
2467 | |
2468 | VM_BUG_ON_PAGE(PageTail(page), page); | |
2469 | ||
2470 | if (likely(!PageCompound(page))) | |
2471 | return atomic_read(&page->_mapcount) + 1; | |
2472 | ||
dd78fedd | 2473 | compound = compound_mapcount(page); |
b20ce5e0 | 2474 | if (PageHuge(page)) |
dd78fedd KS |
2475 | return compound; |
2476 | ret = compound; | |
b20ce5e0 KS |
2477 | for (i = 0; i < HPAGE_PMD_NR; i++) |
2478 | ret += atomic_read(&page[i]._mapcount) + 1; | |
dd78fedd KS |
2479 | /* File pages has compound_mapcount included in _mapcount */ |
2480 | if (!PageAnon(page)) | |
2481 | return ret - compound * HPAGE_PMD_NR; | |
b20ce5e0 KS |
2482 | if (PageDoubleMap(page)) |
2483 | ret -= HPAGE_PMD_NR; | |
2484 | return ret; | |
2485 | } | |
2486 | ||
6d0a07ed AA |
2487 | /* |
2488 | * This calculates accurately how many mappings a transparent hugepage | |
2489 | * has (unlike page_mapcount() which isn't fully accurate). This full | |
2490 | * accuracy is primarily needed to know if copy-on-write faults can | |
2491 | * reuse the page and change the mapping to read-write instead of | |
2492 | * copying them. At the same time this returns the total_mapcount too. | |
2493 | * | |
2494 | * The function returns the highest mapcount any one of the subpages | |
2495 | * has. If the return value is one, even if different processes are | |
2496 | * mapping different subpages of the transparent hugepage, they can | |
2497 | * all reuse it, because each process is reusing a different subpage. | |
2498 | * | |
2499 | * The total_mapcount is instead counting all virtual mappings of the | |
2500 | * subpages. If the total_mapcount is equal to "one", it tells the | |
2501 | * caller all mappings belong to the same "mm" and in turn the | |
2502 | * anon_vma of the transparent hugepage can become the vma->anon_vma | |
2503 | * local one as no other process may be mapping any of the subpages. | |
2504 | * | |
2505 | * It would be more accurate to replace page_mapcount() with | |
2506 | * page_trans_huge_mapcount(), however we only use | |
2507 | * page_trans_huge_mapcount() in the copy-on-write faults where we | |
2508 | * need full accuracy to avoid breaking page pinning, because | |
2509 | * page_trans_huge_mapcount() is slower than page_mapcount(). | |
2510 | */ | |
2511 | int page_trans_huge_mapcount(struct page *page, int *total_mapcount) | |
2512 | { | |
2513 | int i, ret, _total_mapcount, mapcount; | |
2514 | ||
2515 | /* hugetlbfs shouldn't call it */ | |
2516 | VM_BUG_ON_PAGE(PageHuge(page), page); | |
2517 | ||
2518 | if (likely(!PageTransCompound(page))) { | |
2519 | mapcount = atomic_read(&page->_mapcount) + 1; | |
2520 | if (total_mapcount) | |
2521 | *total_mapcount = mapcount; | |
2522 | return mapcount; | |
2523 | } | |
2524 | ||
2525 | page = compound_head(page); | |
2526 | ||
2527 | _total_mapcount = ret = 0; | |
2528 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
2529 | mapcount = atomic_read(&page[i]._mapcount) + 1; | |
2530 | ret = max(ret, mapcount); | |
2531 | _total_mapcount += mapcount; | |
2532 | } | |
2533 | if (PageDoubleMap(page)) { | |
2534 | ret -= 1; | |
2535 | _total_mapcount -= HPAGE_PMD_NR; | |
2536 | } | |
2537 | mapcount = compound_mapcount(page); | |
2538 | ret += mapcount; | |
2539 | _total_mapcount += mapcount; | |
2540 | if (total_mapcount) | |
2541 | *total_mapcount = _total_mapcount; | |
2542 | return ret; | |
2543 | } | |
2544 | ||
b8f593cd HY |
2545 | /* Racy check whether the huge page can be split */ |
2546 | bool can_split_huge_page(struct page *page, int *pextra_pins) | |
2547 | { | |
2548 | int extra_pins; | |
2549 | ||
aa5dc07f | 2550 | /* Additional pins from page cache */ |
b8f593cd HY |
2551 | if (PageAnon(page)) |
2552 | extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0; | |
2553 | else | |
2554 | extra_pins = HPAGE_PMD_NR; | |
2555 | if (pextra_pins) | |
2556 | *pextra_pins = extra_pins; | |
2557 | return total_mapcount(page) == page_count(page) - extra_pins - 1; | |
2558 | } | |
2559 | ||
e9b61f19 KS |
2560 | /* |
2561 | * This function splits huge page into normal pages. @page can point to any | |
2562 | * subpage of huge page to split. Split doesn't change the position of @page. | |
2563 | * | |
2564 | * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. | |
2565 | * The huge page must be locked. | |
2566 | * | |
2567 | * If @list is null, tail pages will be added to LRU list, otherwise, to @list. | |
2568 | * | |
2569 | * Both head page and tail pages will inherit mapping, flags, and so on from | |
2570 | * the hugepage. | |
2571 | * | |
2572 | * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if | |
2573 | * they are not mapped. | |
2574 | * | |
2575 | * Returns 0 if the hugepage is split successfully. | |
2576 | * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under | |
2577 | * us. | |
2578 | */ | |
2579 | int split_huge_page_to_list(struct page *page, struct list_head *list) | |
2580 | { | |
2581 | struct page *head = compound_head(page); | |
a3d0a918 | 2582 | struct pglist_data *pgdata = NODE_DATA(page_to_nid(head)); |
a8803e6c | 2583 | struct deferred_split *ds_queue = get_deferred_split_queue(head); |
baa355fd KS |
2584 | struct anon_vma *anon_vma = NULL; |
2585 | struct address_space *mapping = NULL; | |
2586 | int count, mapcount, extra_pins, ret; | |
0b9b6fff | 2587 | unsigned long flags; |
006d3ff2 | 2588 | pgoff_t end; |
e9b61f19 | 2589 | |
cb829624 | 2590 | VM_BUG_ON_PAGE(is_huge_zero_page(head), head); |
a8803e6c WY |
2591 | VM_BUG_ON_PAGE(!PageLocked(head), head); |
2592 | VM_BUG_ON_PAGE(!PageCompound(head), head); | |
e9b61f19 | 2593 | |
a8803e6c | 2594 | if (PageWriteback(head)) |
59807685 HY |
2595 | return -EBUSY; |
2596 | ||
baa355fd KS |
2597 | if (PageAnon(head)) { |
2598 | /* | |
c1e8d7c6 | 2599 | * The caller does not necessarily hold an mmap_lock that would |
baa355fd KS |
2600 | * prevent the anon_vma disappearing so we first we take a |
2601 | * reference to it and then lock the anon_vma for write. This | |
2602 | * is similar to page_lock_anon_vma_read except the write lock | |
2603 | * is taken to serialise against parallel split or collapse | |
2604 | * operations. | |
2605 | */ | |
2606 | anon_vma = page_get_anon_vma(head); | |
2607 | if (!anon_vma) { | |
2608 | ret = -EBUSY; | |
2609 | goto out; | |
2610 | } | |
006d3ff2 | 2611 | end = -1; |
baa355fd KS |
2612 | mapping = NULL; |
2613 | anon_vma_lock_write(anon_vma); | |
2614 | } else { | |
2615 | mapping = head->mapping; | |
2616 | ||
2617 | /* Truncated ? */ | |
2618 | if (!mapping) { | |
2619 | ret = -EBUSY; | |
2620 | goto out; | |
2621 | } | |
2622 | ||
baa355fd KS |
2623 | anon_vma = NULL; |
2624 | i_mmap_lock_read(mapping); | |
006d3ff2 HD |
2625 | |
2626 | /* | |
2627 | *__split_huge_page() may need to trim off pages beyond EOF: | |
2628 | * but on 32-bit, i_size_read() takes an irq-unsafe seqlock, | |
2629 | * which cannot be nested inside the page tree lock. So note | |
2630 | * end now: i_size itself may be changed at any moment, but | |
2631 | * head page lock is good enough to serialize the trimming. | |
2632 | */ | |
2633 | end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE); | |
e9b61f19 | 2634 | } |
e9b61f19 KS |
2635 | |
2636 | /* | |
906f9cdf | 2637 | * Racy check if we can split the page, before unmap_page() will |
e9b61f19 KS |
2638 | * split PMDs |
2639 | */ | |
b8f593cd | 2640 | if (!can_split_huge_page(head, &extra_pins)) { |
e9b61f19 KS |
2641 | ret = -EBUSY; |
2642 | goto out_unlock; | |
2643 | } | |
2644 | ||
906f9cdf | 2645 | unmap_page(head); |
e9b61f19 KS |
2646 | VM_BUG_ON_PAGE(compound_mapcount(head), head); |
2647 | ||
baa355fd | 2648 | /* prevent PageLRU to go away from under us, and freeze lru stats */ |
f4b7e272 | 2649 | spin_lock_irqsave(&pgdata->lru_lock, flags); |
baa355fd KS |
2650 | |
2651 | if (mapping) { | |
aa5dc07f | 2652 | XA_STATE(xas, &mapping->i_pages, page_index(head)); |
baa355fd | 2653 | |
baa355fd | 2654 | /* |
aa5dc07f | 2655 | * Check if the head page is present in page cache. |
baa355fd KS |
2656 | * We assume all tail are present too, if head is there. |
2657 | */ | |
aa5dc07f MW |
2658 | xa_lock(&mapping->i_pages); |
2659 | if (xas_load(&xas) != head) | |
baa355fd KS |
2660 | goto fail; |
2661 | } | |
2662 | ||
0139aa7b | 2663 | /* Prevent deferred_split_scan() touching ->_refcount */ |
364c1eeb | 2664 | spin_lock(&ds_queue->split_queue_lock); |
e9b61f19 KS |
2665 | count = page_count(head); |
2666 | mapcount = total_mapcount(head); | |
baa355fd | 2667 | if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) { |
9a982250 | 2668 | if (!list_empty(page_deferred_list(head))) { |
364c1eeb | 2669 | ds_queue->split_queue_len--; |
9a982250 KS |
2670 | list_del(page_deferred_list(head)); |
2671 | } | |
afb97172 | 2672 | spin_unlock(&ds_queue->split_queue_lock); |
06d3eff6 | 2673 | if (mapping) { |
a8803e6c WY |
2674 | if (PageSwapBacked(head)) |
2675 | __dec_node_page_state(head, NR_SHMEM_THPS); | |
06d3eff6 | 2676 | else |
a8803e6c | 2677 | __dec_node_page_state(head, NR_FILE_THPS); |
06d3eff6 KS |
2678 | } |
2679 | ||
006d3ff2 | 2680 | __split_huge_page(page, list, end, flags); |
59807685 HY |
2681 | if (PageSwapCache(head)) { |
2682 | swp_entry_t entry = { .val = page_private(head) }; | |
2683 | ||
2684 | ret = split_swap_cluster(entry); | |
2685 | } else | |
2686 | ret = 0; | |
e9b61f19 | 2687 | } else { |
baa355fd KS |
2688 | if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) { |
2689 | pr_alert("total_mapcount: %u, page_count(): %u\n", | |
2690 | mapcount, count); | |
2691 | if (PageTail(page)) | |
2692 | dump_page(head, NULL); | |
2693 | dump_page(page, "total_mapcount(head) > 0"); | |
2694 | BUG(); | |
2695 | } | |
364c1eeb | 2696 | spin_unlock(&ds_queue->split_queue_lock); |
baa355fd | 2697 | fail: if (mapping) |
b93b0163 | 2698 | xa_unlock(&mapping->i_pages); |
f4b7e272 | 2699 | spin_unlock_irqrestore(&pgdata->lru_lock, flags); |
906f9cdf | 2700 | remap_page(head); |
e9b61f19 KS |
2701 | ret = -EBUSY; |
2702 | } | |
2703 | ||
2704 | out_unlock: | |
baa355fd KS |
2705 | if (anon_vma) { |
2706 | anon_vma_unlock_write(anon_vma); | |
2707 | put_anon_vma(anon_vma); | |
2708 | } | |
2709 | if (mapping) | |
2710 | i_mmap_unlock_read(mapping); | |
e9b61f19 KS |
2711 | out: |
2712 | count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); | |
2713 | return ret; | |
2714 | } | |
9a982250 KS |
2715 | |
2716 | void free_transhuge_page(struct page *page) | |
2717 | { | |
87eaceb3 | 2718 | struct deferred_split *ds_queue = get_deferred_split_queue(page); |
9a982250 KS |
2719 | unsigned long flags; |
2720 | ||
364c1eeb | 2721 | spin_lock_irqsave(&ds_queue->split_queue_lock, flags); |
9a982250 | 2722 | if (!list_empty(page_deferred_list(page))) { |
364c1eeb | 2723 | ds_queue->split_queue_len--; |
9a982250 KS |
2724 | list_del(page_deferred_list(page)); |
2725 | } | |
364c1eeb | 2726 | spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); |
9a982250 KS |
2727 | free_compound_page(page); |
2728 | } | |
2729 | ||
2730 | void deferred_split_huge_page(struct page *page) | |
2731 | { | |
87eaceb3 YS |
2732 | struct deferred_split *ds_queue = get_deferred_split_queue(page); |
2733 | #ifdef CONFIG_MEMCG | |
2734 | struct mem_cgroup *memcg = compound_head(page)->mem_cgroup; | |
2735 | #endif | |
9a982250 KS |
2736 | unsigned long flags; |
2737 | ||
2738 | VM_BUG_ON_PAGE(!PageTransHuge(page), page); | |
2739 | ||
87eaceb3 YS |
2740 | /* |
2741 | * The try_to_unmap() in page reclaim path might reach here too, | |
2742 | * this may cause a race condition to corrupt deferred split queue. | |
2743 | * And, if page reclaim is already handling the same page, it is | |
2744 | * unnecessary to handle it again in shrinker. | |
2745 | * | |
2746 | * Check PageSwapCache to determine if the page is being | |
2747 | * handled by page reclaim since THP swap would add the page into | |
2748 | * swap cache before calling try_to_unmap(). | |
2749 | */ | |
2750 | if (PageSwapCache(page)) | |
2751 | return; | |
2752 | ||
364c1eeb | 2753 | spin_lock_irqsave(&ds_queue->split_queue_lock, flags); |
9a982250 | 2754 | if (list_empty(page_deferred_list(page))) { |
f9719a03 | 2755 | count_vm_event(THP_DEFERRED_SPLIT_PAGE); |
364c1eeb YS |
2756 | list_add_tail(page_deferred_list(page), &ds_queue->split_queue); |
2757 | ds_queue->split_queue_len++; | |
87eaceb3 YS |
2758 | #ifdef CONFIG_MEMCG |
2759 | if (memcg) | |
2760 | memcg_set_shrinker_bit(memcg, page_to_nid(page), | |
2761 | deferred_split_shrinker.id); | |
2762 | #endif | |
9a982250 | 2763 | } |
364c1eeb | 2764 | spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); |
9a982250 KS |
2765 | } |
2766 | ||
2767 | static unsigned long deferred_split_count(struct shrinker *shrink, | |
2768 | struct shrink_control *sc) | |
2769 | { | |
a3d0a918 | 2770 | struct pglist_data *pgdata = NODE_DATA(sc->nid); |
364c1eeb | 2771 | struct deferred_split *ds_queue = &pgdata->deferred_split_queue; |
87eaceb3 YS |
2772 | |
2773 | #ifdef CONFIG_MEMCG | |
2774 | if (sc->memcg) | |
2775 | ds_queue = &sc->memcg->deferred_split_queue; | |
2776 | #endif | |
364c1eeb | 2777 | return READ_ONCE(ds_queue->split_queue_len); |
9a982250 KS |
2778 | } |
2779 | ||
2780 | static unsigned long deferred_split_scan(struct shrinker *shrink, | |
2781 | struct shrink_control *sc) | |
2782 | { | |
a3d0a918 | 2783 | struct pglist_data *pgdata = NODE_DATA(sc->nid); |
364c1eeb | 2784 | struct deferred_split *ds_queue = &pgdata->deferred_split_queue; |
9a982250 KS |
2785 | unsigned long flags; |
2786 | LIST_HEAD(list), *pos, *next; | |
2787 | struct page *page; | |
2788 | int split = 0; | |
2789 | ||
87eaceb3 YS |
2790 | #ifdef CONFIG_MEMCG |
2791 | if (sc->memcg) | |
2792 | ds_queue = &sc->memcg->deferred_split_queue; | |
2793 | #endif | |
2794 | ||
364c1eeb | 2795 | spin_lock_irqsave(&ds_queue->split_queue_lock, flags); |
9a982250 | 2796 | /* Take pin on all head pages to avoid freeing them under us */ |
364c1eeb | 2797 | list_for_each_safe(pos, next, &ds_queue->split_queue) { |
9a982250 KS |
2798 | page = list_entry((void *)pos, struct page, mapping); |
2799 | page = compound_head(page); | |
e3ae1953 KS |
2800 | if (get_page_unless_zero(page)) { |
2801 | list_move(page_deferred_list(page), &list); | |
2802 | } else { | |
2803 | /* We lost race with put_compound_page() */ | |
9a982250 | 2804 | list_del_init(page_deferred_list(page)); |
364c1eeb | 2805 | ds_queue->split_queue_len--; |
9a982250 | 2806 | } |
e3ae1953 KS |
2807 | if (!--sc->nr_to_scan) |
2808 | break; | |
9a982250 | 2809 | } |
364c1eeb | 2810 | spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); |
9a982250 KS |
2811 | |
2812 | list_for_each_safe(pos, next, &list) { | |
2813 | page = list_entry((void *)pos, struct page, mapping); | |
fa41b900 KS |
2814 | if (!trylock_page(page)) |
2815 | goto next; | |
9a982250 KS |
2816 | /* split_huge_page() removes page from list on success */ |
2817 | if (!split_huge_page(page)) | |
2818 | split++; | |
2819 | unlock_page(page); | |
fa41b900 | 2820 | next: |
9a982250 KS |
2821 | put_page(page); |
2822 | } | |
2823 | ||
364c1eeb YS |
2824 | spin_lock_irqsave(&ds_queue->split_queue_lock, flags); |
2825 | list_splice_tail(&list, &ds_queue->split_queue); | |
2826 | spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); | |
9a982250 | 2827 | |
cb8d68ec KS |
2828 | /* |
2829 | * Stop shrinker if we didn't split any page, but the queue is empty. | |
2830 | * This can happen if pages were freed under us. | |
2831 | */ | |
364c1eeb | 2832 | if (!split && list_empty(&ds_queue->split_queue)) |
cb8d68ec KS |
2833 | return SHRINK_STOP; |
2834 | return split; | |
9a982250 KS |
2835 | } |
2836 | ||
2837 | static struct shrinker deferred_split_shrinker = { | |
2838 | .count_objects = deferred_split_count, | |
2839 | .scan_objects = deferred_split_scan, | |
2840 | .seeks = DEFAULT_SEEKS, | |
87eaceb3 YS |
2841 | .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE | |
2842 | SHRINKER_NONSLAB, | |
9a982250 | 2843 | }; |
49071d43 KS |
2844 | |
2845 | #ifdef CONFIG_DEBUG_FS | |
2846 | static int split_huge_pages_set(void *data, u64 val) | |
2847 | { | |
2848 | struct zone *zone; | |
2849 | struct page *page; | |
2850 | unsigned long pfn, max_zone_pfn; | |
2851 | unsigned long total = 0, split = 0; | |
2852 | ||
2853 | if (val != 1) | |
2854 | return -EINVAL; | |
2855 | ||
2856 | for_each_populated_zone(zone) { | |
2857 | max_zone_pfn = zone_end_pfn(zone); | |
2858 | for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { | |
2859 | if (!pfn_valid(pfn)) | |
2860 | continue; | |
2861 | ||
2862 | page = pfn_to_page(pfn); | |
2863 | if (!get_page_unless_zero(page)) | |
2864 | continue; | |
2865 | ||
2866 | if (zone != page_zone(page)) | |
2867 | goto next; | |
2868 | ||
baa355fd | 2869 | if (!PageHead(page) || PageHuge(page) || !PageLRU(page)) |
49071d43 KS |
2870 | goto next; |
2871 | ||
2872 | total++; | |
2873 | lock_page(page); | |
2874 | if (!split_huge_page(page)) | |
2875 | split++; | |
2876 | unlock_page(page); | |
2877 | next: | |
2878 | put_page(page); | |
2879 | } | |
2880 | } | |
2881 | ||
145bdaa1 | 2882 | pr_info("%lu of %lu THP split\n", split, total); |
49071d43 KS |
2883 | |
2884 | return 0; | |
2885 | } | |
f1287869 | 2886 | DEFINE_DEBUGFS_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set, |
49071d43 KS |
2887 | "%llu\n"); |
2888 | ||
2889 | static int __init split_huge_pages_debugfs(void) | |
2890 | { | |
d9f7979c GKH |
2891 | debugfs_create_file("split_huge_pages", 0200, NULL, NULL, |
2892 | &split_huge_pages_fops); | |
49071d43 KS |
2893 | return 0; |
2894 | } | |
2895 | late_initcall(split_huge_pages_debugfs); | |
2896 | #endif | |
616b8371 ZY |
2897 | |
2898 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION | |
2899 | void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw, | |
2900 | struct page *page) | |
2901 | { | |
2902 | struct vm_area_struct *vma = pvmw->vma; | |
2903 | struct mm_struct *mm = vma->vm_mm; | |
2904 | unsigned long address = pvmw->address; | |
2905 | pmd_t pmdval; | |
2906 | swp_entry_t entry; | |
ab6e3d09 | 2907 | pmd_t pmdswp; |
616b8371 ZY |
2908 | |
2909 | if (!(pvmw->pmd && !pvmw->pte)) | |
2910 | return; | |
2911 | ||
616b8371 | 2912 | flush_cache_range(vma, address, address + HPAGE_PMD_SIZE); |
8a8683ad | 2913 | pmdval = pmdp_invalidate(vma, address, pvmw->pmd); |
616b8371 ZY |
2914 | if (pmd_dirty(pmdval)) |
2915 | set_page_dirty(page); | |
2916 | entry = make_migration_entry(page, pmd_write(pmdval)); | |
ab6e3d09 NH |
2917 | pmdswp = swp_entry_to_pmd(entry); |
2918 | if (pmd_soft_dirty(pmdval)) | |
2919 | pmdswp = pmd_swp_mksoft_dirty(pmdswp); | |
2920 | set_pmd_at(mm, address, pvmw->pmd, pmdswp); | |
616b8371 ZY |
2921 | page_remove_rmap(page, true); |
2922 | put_page(page); | |
616b8371 ZY |
2923 | } |
2924 | ||
2925 | void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new) | |
2926 | { | |
2927 | struct vm_area_struct *vma = pvmw->vma; | |
2928 | struct mm_struct *mm = vma->vm_mm; | |
2929 | unsigned long address = pvmw->address; | |
2930 | unsigned long mmun_start = address & HPAGE_PMD_MASK; | |
2931 | pmd_t pmde; | |
2932 | swp_entry_t entry; | |
2933 | ||
2934 | if (!(pvmw->pmd && !pvmw->pte)) | |
2935 | return; | |
2936 | ||
2937 | entry = pmd_to_swp_entry(*pvmw->pmd); | |
2938 | get_page(new); | |
2939 | pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot)); | |
ab6e3d09 NH |
2940 | if (pmd_swp_soft_dirty(*pvmw->pmd)) |
2941 | pmde = pmd_mksoft_dirty(pmde); | |
616b8371 | 2942 | if (is_write_migration_entry(entry)) |
f55e1014 | 2943 | pmde = maybe_pmd_mkwrite(pmde, vma); |
616b8371 ZY |
2944 | |
2945 | flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE); | |
e71769ae NH |
2946 | if (PageAnon(new)) |
2947 | page_add_anon_rmap(new, vma, mmun_start, true); | |
2948 | else | |
2949 | page_add_file_rmap(new, true); | |
616b8371 | 2950 | set_pmd_at(mm, mmun_start, pvmw->pmd, pmde); |
e125fe40 | 2951 | if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new)) |
616b8371 ZY |
2952 | mlock_vma_page(new); |
2953 | update_mmu_cache_pmd(vma, address, pvmw->pmd); | |
2954 | } | |
2955 | #endif |