]>
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 | { | |
bfb0ffeb JP |
166 | const char *output; |
167 | ||
444eb2a4 | 168 | if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags)) |
bfb0ffeb JP |
169 | output = "[always] madvise never"; |
170 | else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
171 | &transparent_hugepage_flags)) | |
172 | output = "always [madvise] never"; | |
444eb2a4 | 173 | else |
bfb0ffeb JP |
174 | output = "always madvise [never]"; |
175 | ||
176 | return sysfs_emit(buf, "%s\n", output); | |
71e3aac0 | 177 | } |
444eb2a4 | 178 | |
71e3aac0 AA |
179 | static ssize_t enabled_store(struct kobject *kobj, |
180 | struct kobj_attribute *attr, | |
181 | const char *buf, size_t count) | |
182 | { | |
21440d7e | 183 | ssize_t ret = count; |
ba76149f | 184 | |
f42f2552 | 185 | if (sysfs_streq(buf, "always")) { |
21440d7e DR |
186 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); |
187 | set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); | |
f42f2552 | 188 | } else if (sysfs_streq(buf, "madvise")) { |
21440d7e DR |
189 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); |
190 | set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
f42f2552 | 191 | } else if (sysfs_streq(buf, "never")) { |
21440d7e DR |
192 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); |
193 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
194 | } else | |
195 | ret = -EINVAL; | |
ba76149f AA |
196 | |
197 | if (ret > 0) { | |
b46e756f | 198 | int err = start_stop_khugepaged(); |
ba76149f AA |
199 | if (err) |
200 | ret = err; | |
201 | } | |
ba76149f | 202 | return ret; |
71e3aac0 AA |
203 | } |
204 | static struct kobj_attribute enabled_attr = | |
205 | __ATTR(enabled, 0644, enabled_show, enabled_store); | |
206 | ||
b46e756f | 207 | ssize_t single_hugepage_flag_show(struct kobject *kobj, |
bfb0ffeb JP |
208 | struct kobj_attribute *attr, char *buf, |
209 | enum transparent_hugepage_flag flag) | |
71e3aac0 | 210 | { |
bfb0ffeb JP |
211 | return sysfs_emit(buf, "%d\n", |
212 | !!test_bit(flag, &transparent_hugepage_flags)); | |
71e3aac0 | 213 | } |
e27e6151 | 214 | |
b46e756f | 215 | ssize_t single_hugepage_flag_store(struct kobject *kobj, |
71e3aac0 AA |
216 | struct kobj_attribute *attr, |
217 | const char *buf, size_t count, | |
218 | enum transparent_hugepage_flag flag) | |
219 | { | |
e27e6151 BH |
220 | unsigned long value; |
221 | int ret; | |
222 | ||
223 | ret = kstrtoul(buf, 10, &value); | |
224 | if (ret < 0) | |
225 | return ret; | |
226 | if (value > 1) | |
227 | return -EINVAL; | |
228 | ||
229 | if (value) | |
71e3aac0 | 230 | set_bit(flag, &transparent_hugepage_flags); |
e27e6151 | 231 | else |
71e3aac0 | 232 | clear_bit(flag, &transparent_hugepage_flags); |
71e3aac0 AA |
233 | |
234 | return count; | |
235 | } | |
236 | ||
71e3aac0 AA |
237 | static ssize_t defrag_show(struct kobject *kobj, |
238 | struct kobj_attribute *attr, char *buf) | |
239 | { | |
bfb0ffeb JP |
240 | const char *output; |
241 | ||
242 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, | |
243 | &transparent_hugepage_flags)) | |
244 | output = "[always] defer defer+madvise madvise never"; | |
245 | else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, | |
246 | &transparent_hugepage_flags)) | |
247 | output = "always [defer] defer+madvise madvise never"; | |
248 | else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, | |
249 | &transparent_hugepage_flags)) | |
250 | output = "always defer [defer+madvise] madvise never"; | |
251 | else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, | |
252 | &transparent_hugepage_flags)) | |
253 | output = "always defer defer+madvise [madvise] never"; | |
254 | else | |
255 | output = "always defer defer+madvise madvise [never]"; | |
256 | ||
257 | return sysfs_emit(buf, "%s\n", output); | |
71e3aac0 | 258 | } |
21440d7e | 259 | |
71e3aac0 AA |
260 | static ssize_t defrag_store(struct kobject *kobj, |
261 | struct kobj_attribute *attr, | |
262 | const char *buf, size_t count) | |
263 | { | |
f42f2552 | 264 | if (sysfs_streq(buf, "always")) { |
21440d7e DR |
265 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); |
266 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
267 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
268 | set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); | |
f42f2552 | 269 | } else if (sysfs_streq(buf, "defer+madvise")) { |
21440d7e DR |
270 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); |
271 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
272 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
273 | set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
f42f2552 | 274 | } else if (sysfs_streq(buf, "defer")) { |
4fad7fb6 DR |
275 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); |
276 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
277 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
278 | set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
f42f2552 | 279 | } else if (sysfs_streq(buf, "madvise")) { |
21440d7e DR |
280 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); |
281 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
282 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
283 | set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
f42f2552 | 284 | } else if (sysfs_streq(buf, "never")) { |
21440d7e DR |
285 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); |
286 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
287 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
288 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
289 | } else | |
290 | return -EINVAL; | |
291 | ||
292 | return count; | |
71e3aac0 AA |
293 | } |
294 | static struct kobj_attribute defrag_attr = | |
295 | __ATTR(defrag, 0644, defrag_show, defrag_store); | |
296 | ||
79da5407 | 297 | static ssize_t use_zero_page_show(struct kobject *kobj, |
ae7a927d | 298 | struct kobj_attribute *attr, char *buf) |
79da5407 | 299 | { |
b46e756f | 300 | return single_hugepage_flag_show(kobj, attr, buf, |
ae7a927d | 301 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); |
79da5407 KS |
302 | } |
303 | static ssize_t use_zero_page_store(struct kobject *kobj, | |
304 | struct kobj_attribute *attr, const char *buf, size_t count) | |
305 | { | |
b46e756f | 306 | return single_hugepage_flag_store(kobj, attr, buf, count, |
79da5407 KS |
307 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); |
308 | } | |
309 | static struct kobj_attribute use_zero_page_attr = | |
310 | __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); | |
49920d28 HD |
311 | |
312 | static ssize_t hpage_pmd_size_show(struct kobject *kobj, | |
ae7a927d | 313 | struct kobj_attribute *attr, char *buf) |
49920d28 | 314 | { |
ae7a927d | 315 | return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE); |
49920d28 HD |
316 | } |
317 | static struct kobj_attribute hpage_pmd_size_attr = | |
318 | __ATTR_RO(hpage_pmd_size); | |
319 | ||
71e3aac0 AA |
320 | static struct attribute *hugepage_attr[] = { |
321 | &enabled_attr.attr, | |
322 | &defrag_attr.attr, | |
79da5407 | 323 | &use_zero_page_attr.attr, |
49920d28 | 324 | &hpage_pmd_size_attr.attr, |
396bcc52 | 325 | #ifdef CONFIG_SHMEM |
5a6e75f8 | 326 | &shmem_enabled_attr.attr, |
71e3aac0 AA |
327 | #endif |
328 | NULL, | |
329 | }; | |
330 | ||
8aa95a21 | 331 | static const struct attribute_group hugepage_attr_group = { |
71e3aac0 | 332 | .attrs = hugepage_attr, |
ba76149f AA |
333 | }; |
334 | ||
569e5590 | 335 | static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) |
71e3aac0 | 336 | { |
71e3aac0 AA |
337 | int err; |
338 | ||
569e5590 SL |
339 | *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); |
340 | if (unlikely(!*hugepage_kobj)) { | |
ae3a8c1c | 341 | pr_err("failed to create transparent hugepage kobject\n"); |
569e5590 | 342 | return -ENOMEM; |
ba76149f AA |
343 | } |
344 | ||
569e5590 | 345 | err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); |
ba76149f | 346 | if (err) { |
ae3a8c1c | 347 | pr_err("failed to register transparent hugepage group\n"); |
569e5590 | 348 | goto delete_obj; |
ba76149f AA |
349 | } |
350 | ||
569e5590 | 351 | err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); |
ba76149f | 352 | if (err) { |
ae3a8c1c | 353 | pr_err("failed to register transparent hugepage group\n"); |
569e5590 | 354 | goto remove_hp_group; |
ba76149f | 355 | } |
569e5590 SL |
356 | |
357 | return 0; | |
358 | ||
359 | remove_hp_group: | |
360 | sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); | |
361 | delete_obj: | |
362 | kobject_put(*hugepage_kobj); | |
363 | return err; | |
364 | } | |
365 | ||
366 | static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
367 | { | |
368 | sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); | |
369 | sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); | |
370 | kobject_put(hugepage_kobj); | |
371 | } | |
372 | #else | |
373 | static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) | |
374 | { | |
375 | return 0; | |
376 | } | |
377 | ||
378 | static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
379 | { | |
380 | } | |
381 | #endif /* CONFIG_SYSFS */ | |
382 | ||
383 | static int __init hugepage_init(void) | |
384 | { | |
385 | int err; | |
386 | struct kobject *hugepage_kobj; | |
387 | ||
388 | if (!has_transparent_hugepage()) { | |
389 | transparent_hugepage_flags = 0; | |
390 | return -EINVAL; | |
391 | } | |
392 | ||
ff20c2e0 KS |
393 | /* |
394 | * hugepages can't be allocated by the buddy allocator | |
395 | */ | |
396 | MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER); | |
397 | /* | |
398 | * we use page->mapping and page->index in second tail page | |
399 | * as list_head: assuming THP order >= 2 | |
400 | */ | |
401 | MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2); | |
402 | ||
569e5590 SL |
403 | err = hugepage_init_sysfs(&hugepage_kobj); |
404 | if (err) | |
65ebb64f | 405 | goto err_sysfs; |
ba76149f | 406 | |
b46e756f | 407 | err = khugepaged_init(); |
ba76149f | 408 | if (err) |
65ebb64f | 409 | goto err_slab; |
ba76149f | 410 | |
65ebb64f KS |
411 | err = register_shrinker(&huge_zero_page_shrinker); |
412 | if (err) | |
413 | goto err_hzp_shrinker; | |
9a982250 KS |
414 | err = register_shrinker(&deferred_split_shrinker); |
415 | if (err) | |
416 | goto err_split_shrinker; | |
97ae1749 | 417 | |
97562cd2 RR |
418 | /* |
419 | * By default disable transparent hugepages on smaller systems, | |
420 | * where the extra memory used could hurt more than TLB overhead | |
421 | * is likely to save. The admin can still enable it through /sys. | |
422 | */ | |
ca79b0c2 | 423 | if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) { |
97562cd2 | 424 | transparent_hugepage_flags = 0; |
79553da2 KS |
425 | return 0; |
426 | } | |
97562cd2 | 427 | |
79553da2 | 428 | err = start_stop_khugepaged(); |
65ebb64f KS |
429 | if (err) |
430 | goto err_khugepaged; | |
ba76149f | 431 | |
569e5590 | 432 | return 0; |
65ebb64f | 433 | err_khugepaged: |
9a982250 KS |
434 | unregister_shrinker(&deferred_split_shrinker); |
435 | err_split_shrinker: | |
65ebb64f KS |
436 | unregister_shrinker(&huge_zero_page_shrinker); |
437 | err_hzp_shrinker: | |
b46e756f | 438 | khugepaged_destroy(); |
65ebb64f | 439 | err_slab: |
569e5590 | 440 | hugepage_exit_sysfs(hugepage_kobj); |
65ebb64f | 441 | err_sysfs: |
ba76149f | 442 | return err; |
71e3aac0 | 443 | } |
a64fb3cd | 444 | subsys_initcall(hugepage_init); |
71e3aac0 AA |
445 | |
446 | static int __init setup_transparent_hugepage(char *str) | |
447 | { | |
448 | int ret = 0; | |
449 | if (!str) | |
450 | goto out; | |
451 | if (!strcmp(str, "always")) { | |
452 | set_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
453 | &transparent_hugepage_flags); | |
454 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
455 | &transparent_hugepage_flags); | |
456 | ret = 1; | |
457 | } else if (!strcmp(str, "madvise")) { | |
458 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
459 | &transparent_hugepage_flags); | |
460 | set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
461 | &transparent_hugepage_flags); | |
462 | ret = 1; | |
463 | } else if (!strcmp(str, "never")) { | |
464 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
465 | &transparent_hugepage_flags); | |
466 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
467 | &transparent_hugepage_flags); | |
468 | ret = 1; | |
469 | } | |
470 | out: | |
471 | if (!ret) | |
ae3a8c1c | 472 | pr_warn("transparent_hugepage= cannot parse, ignored\n"); |
71e3aac0 AA |
473 | return ret; |
474 | } | |
475 | __setup("transparent_hugepage=", setup_transparent_hugepage); | |
476 | ||
f55e1014 | 477 | pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) |
71e3aac0 | 478 | { |
f55e1014 | 479 | if (likely(vma->vm_flags & VM_WRITE)) |
71e3aac0 AA |
480 | pmd = pmd_mkwrite(pmd); |
481 | return pmd; | |
482 | } | |
483 | ||
87eaceb3 YS |
484 | #ifdef CONFIG_MEMCG |
485 | static inline struct deferred_split *get_deferred_split_queue(struct page *page) | |
9a982250 | 486 | { |
bcfe06bf | 487 | struct mem_cgroup *memcg = page_memcg(compound_head(page)); |
87eaceb3 YS |
488 | struct pglist_data *pgdat = NODE_DATA(page_to_nid(page)); |
489 | ||
490 | if (memcg) | |
491 | return &memcg->deferred_split_queue; | |
492 | else | |
493 | return &pgdat->deferred_split_queue; | |
9a982250 | 494 | } |
87eaceb3 YS |
495 | #else |
496 | static inline struct deferred_split *get_deferred_split_queue(struct page *page) | |
497 | { | |
498 | struct pglist_data *pgdat = NODE_DATA(page_to_nid(page)); | |
499 | ||
500 | return &pgdat->deferred_split_queue; | |
501 | } | |
502 | #endif | |
9a982250 KS |
503 | |
504 | void prep_transhuge_page(struct page *page) | |
505 | { | |
506 | /* | |
507 | * we use page->mapping and page->indexlru in second tail page | |
508 | * as list_head: assuming THP order >= 2 | |
509 | */ | |
9a982250 KS |
510 | |
511 | INIT_LIST_HEAD(page_deferred_list(page)); | |
512 | set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR); | |
513 | } | |
514 | ||
005ba37c SC |
515 | bool is_transparent_hugepage(struct page *page) |
516 | { | |
517 | if (!PageCompound(page)) | |
fa1f68cc | 518 | return false; |
005ba37c SC |
519 | |
520 | page = compound_head(page); | |
521 | return is_huge_zero_page(page) || | |
522 | page[1].compound_dtor == TRANSHUGE_PAGE_DTOR; | |
523 | } | |
524 | EXPORT_SYMBOL_GPL(is_transparent_hugepage); | |
525 | ||
97d3d0f9 KS |
526 | static unsigned long __thp_get_unmapped_area(struct file *filp, |
527 | unsigned long addr, unsigned long len, | |
74d2fad1 TK |
528 | loff_t off, unsigned long flags, unsigned long size) |
529 | { | |
74d2fad1 TK |
530 | loff_t off_end = off + len; |
531 | loff_t off_align = round_up(off, size); | |
97d3d0f9 | 532 | unsigned long len_pad, ret; |
74d2fad1 TK |
533 | |
534 | if (off_end <= off_align || (off_end - off_align) < size) | |
535 | return 0; | |
536 | ||
537 | len_pad = len + size; | |
538 | if (len_pad < len || (off + len_pad) < off) | |
539 | return 0; | |
540 | ||
97d3d0f9 | 541 | ret = current->mm->get_unmapped_area(filp, addr, len_pad, |
74d2fad1 | 542 | off >> PAGE_SHIFT, flags); |
97d3d0f9 KS |
543 | |
544 | /* | |
545 | * The failure might be due to length padding. The caller will retry | |
546 | * without the padding. | |
547 | */ | |
548 | if (IS_ERR_VALUE(ret)) | |
74d2fad1 TK |
549 | return 0; |
550 | ||
97d3d0f9 KS |
551 | /* |
552 | * Do not try to align to THP boundary if allocation at the address | |
553 | * hint succeeds. | |
554 | */ | |
555 | if (ret == addr) | |
556 | return addr; | |
557 | ||
558 | ret += (off - ret) & (size - 1); | |
559 | return ret; | |
74d2fad1 TK |
560 | } |
561 | ||
562 | unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr, | |
563 | unsigned long len, unsigned long pgoff, unsigned long flags) | |
564 | { | |
97d3d0f9 | 565 | unsigned long ret; |
74d2fad1 TK |
566 | loff_t off = (loff_t)pgoff << PAGE_SHIFT; |
567 | ||
74d2fad1 TK |
568 | if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD)) |
569 | goto out; | |
570 | ||
97d3d0f9 KS |
571 | ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE); |
572 | if (ret) | |
573 | return ret; | |
574 | out: | |
74d2fad1 TK |
575 | return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags); |
576 | } | |
577 | EXPORT_SYMBOL_GPL(thp_get_unmapped_area); | |
578 | ||
2b740303 SJ |
579 | static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf, |
580 | struct page *page, gfp_t gfp) | |
71e3aac0 | 581 | { |
82b0f8c3 | 582 | struct vm_area_struct *vma = vmf->vma; |
71e3aac0 | 583 | pgtable_t pgtable; |
82b0f8c3 | 584 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; |
2b740303 | 585 | vm_fault_t ret = 0; |
71e3aac0 | 586 | |
309381fe | 587 | VM_BUG_ON_PAGE(!PageCompound(page), page); |
00501b53 | 588 | |
d9eb1ea2 | 589 | if (mem_cgroup_charge(page, vma->vm_mm, gfp)) { |
6b251fc9 AA |
590 | put_page(page); |
591 | count_vm_event(THP_FAULT_FALLBACK); | |
85b9f46e | 592 | count_vm_event(THP_FAULT_FALLBACK_CHARGE); |
6b251fc9 AA |
593 | return VM_FAULT_FALLBACK; |
594 | } | |
9d82c694 | 595 | cgroup_throttle_swaprate(page, gfp); |
00501b53 | 596 | |
4cf58924 | 597 | pgtable = pte_alloc_one(vma->vm_mm); |
00501b53 | 598 | if (unlikely(!pgtable)) { |
6b31d595 MH |
599 | ret = VM_FAULT_OOM; |
600 | goto release; | |
00501b53 | 601 | } |
71e3aac0 | 602 | |
c79b57e4 | 603 | clear_huge_page(page, vmf->address, HPAGE_PMD_NR); |
52f37629 MK |
604 | /* |
605 | * The memory barrier inside __SetPageUptodate makes sure that | |
606 | * clear_huge_page writes become visible before the set_pmd_at() | |
607 | * write. | |
608 | */ | |
71e3aac0 AA |
609 | __SetPageUptodate(page); |
610 | ||
82b0f8c3 JK |
611 | vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); |
612 | if (unlikely(!pmd_none(*vmf->pmd))) { | |
6b31d595 | 613 | goto unlock_release; |
71e3aac0 AA |
614 | } else { |
615 | pmd_t entry; | |
6b251fc9 | 616 | |
6b31d595 MH |
617 | ret = check_stable_address_space(vma->vm_mm); |
618 | if (ret) | |
619 | goto unlock_release; | |
620 | ||
6b251fc9 AA |
621 | /* Deliver the page fault to userland */ |
622 | if (userfaultfd_missing(vma)) { | |
2b740303 | 623 | vm_fault_t ret2; |
6b251fc9 | 624 | |
82b0f8c3 | 625 | spin_unlock(vmf->ptl); |
6b251fc9 | 626 | put_page(page); |
bae473a4 | 627 | pte_free(vma->vm_mm, pgtable); |
2b740303 SJ |
628 | ret2 = handle_userfault(vmf, VM_UFFD_MISSING); |
629 | VM_BUG_ON(ret2 & VM_FAULT_FALLBACK); | |
630 | return ret2; | |
6b251fc9 AA |
631 | } |
632 | ||
3122359a | 633 | entry = mk_huge_pmd(page, vma->vm_page_prot); |
f55e1014 | 634 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
d281ee61 | 635 | page_add_new_anon_rmap(page, vma, haddr, true); |
b518154e | 636 | lru_cache_add_inactive_or_unevictable(page, vma); |
82b0f8c3 JK |
637 | pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable); |
638 | set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); | |
fca40573 | 639 | update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); |
bae473a4 | 640 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); |
c4812909 | 641 | mm_inc_nr_ptes(vma->vm_mm); |
82b0f8c3 | 642 | spin_unlock(vmf->ptl); |
6b251fc9 | 643 | count_vm_event(THP_FAULT_ALLOC); |
9d82c694 | 644 | count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC); |
71e3aac0 AA |
645 | } |
646 | ||
aa2e878e | 647 | return 0; |
6b31d595 MH |
648 | unlock_release: |
649 | spin_unlock(vmf->ptl); | |
650 | release: | |
651 | if (pgtable) | |
652 | pte_free(vma->vm_mm, pgtable); | |
6b31d595 MH |
653 | put_page(page); |
654 | return ret; | |
655 | ||
71e3aac0 AA |
656 | } |
657 | ||
444eb2a4 | 658 | /* |
21440d7e DR |
659 | * always: directly stall for all thp allocations |
660 | * defer: wake kswapd and fail if not immediately available | |
661 | * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise | |
662 | * fail if not immediately available | |
663 | * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately | |
664 | * available | |
665 | * never: never stall for any thp allocation | |
444eb2a4 | 666 | */ |
19deb769 | 667 | static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma) |
444eb2a4 | 668 | { |
21440d7e | 669 | const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE); |
2f0799a0 | 670 | |
ac79f78d | 671 | /* Always do synchronous compaction */ |
a8282608 AA |
672 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) |
673 | return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY); | |
ac79f78d DR |
674 | |
675 | /* Kick kcompactd and fail quickly */ | |
21440d7e | 676 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) |
19deb769 | 677 | return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM; |
ac79f78d DR |
678 | |
679 | /* Synchronous compaction if madvised, otherwise kick kcompactd */ | |
21440d7e | 680 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) |
19deb769 DR |
681 | return GFP_TRANSHUGE_LIGHT | |
682 | (vma_madvised ? __GFP_DIRECT_RECLAIM : | |
683 | __GFP_KSWAPD_RECLAIM); | |
ac79f78d DR |
684 | |
685 | /* Only do synchronous compaction if madvised */ | |
21440d7e | 686 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) |
19deb769 DR |
687 | return GFP_TRANSHUGE_LIGHT | |
688 | (vma_madvised ? __GFP_DIRECT_RECLAIM : 0); | |
ac79f78d | 689 | |
19deb769 | 690 | return GFP_TRANSHUGE_LIGHT; |
444eb2a4 MG |
691 | } |
692 | ||
c4088ebd | 693 | /* Caller must hold page table lock. */ |
2efeb8da | 694 | static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, |
97ae1749 | 695 | struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, |
5918d10a | 696 | struct page *zero_page) |
fc9fe822 KS |
697 | { |
698 | pmd_t entry; | |
7c414164 | 699 | if (!pmd_none(*pmd)) |
2efeb8da | 700 | return; |
5918d10a | 701 | entry = mk_pmd(zero_page, vma->vm_page_prot); |
fc9fe822 | 702 | entry = pmd_mkhuge(entry); |
12c9d70b MW |
703 | if (pgtable) |
704 | pgtable_trans_huge_deposit(mm, pmd, pgtable); | |
fc9fe822 | 705 | set_pmd_at(mm, haddr, pmd, entry); |
c4812909 | 706 | mm_inc_nr_ptes(mm); |
fc9fe822 KS |
707 | } |
708 | ||
2b740303 | 709 | vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf) |
71e3aac0 | 710 | { |
82b0f8c3 | 711 | struct vm_area_struct *vma = vmf->vma; |
077fcf11 | 712 | gfp_t gfp; |
71e3aac0 | 713 | struct page *page; |
82b0f8c3 | 714 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; |
71e3aac0 | 715 | |
43675e6f | 716 | if (!transhuge_vma_suitable(vma, haddr)) |
c0292554 | 717 | return VM_FAULT_FALLBACK; |
128ec037 KS |
718 | if (unlikely(anon_vma_prepare(vma))) |
719 | return VM_FAULT_OOM; | |
6d50e60c | 720 | if (unlikely(khugepaged_enter(vma, vma->vm_flags))) |
128ec037 | 721 | return VM_FAULT_OOM; |
82b0f8c3 | 722 | if (!(vmf->flags & FAULT_FLAG_WRITE) && |
bae473a4 | 723 | !mm_forbids_zeropage(vma->vm_mm) && |
128ec037 KS |
724 | transparent_hugepage_use_zero_page()) { |
725 | pgtable_t pgtable; | |
726 | struct page *zero_page; | |
2b740303 | 727 | vm_fault_t ret; |
4cf58924 | 728 | pgtable = pte_alloc_one(vma->vm_mm); |
128ec037 | 729 | if (unlikely(!pgtable)) |
ba76149f | 730 | return VM_FAULT_OOM; |
6fcb52a5 | 731 | zero_page = mm_get_huge_zero_page(vma->vm_mm); |
128ec037 | 732 | if (unlikely(!zero_page)) { |
bae473a4 | 733 | pte_free(vma->vm_mm, pgtable); |
81ab4201 | 734 | count_vm_event(THP_FAULT_FALLBACK); |
c0292554 | 735 | return VM_FAULT_FALLBACK; |
b9bbfbe3 | 736 | } |
82b0f8c3 | 737 | vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); |
6b251fc9 | 738 | ret = 0; |
82b0f8c3 | 739 | if (pmd_none(*vmf->pmd)) { |
6b31d595 MH |
740 | ret = check_stable_address_space(vma->vm_mm); |
741 | if (ret) { | |
742 | spin_unlock(vmf->ptl); | |
bfe8cc1d | 743 | pte_free(vma->vm_mm, pgtable); |
6b31d595 | 744 | } else if (userfaultfd_missing(vma)) { |
82b0f8c3 | 745 | spin_unlock(vmf->ptl); |
bfe8cc1d | 746 | pte_free(vma->vm_mm, pgtable); |
82b0f8c3 | 747 | ret = handle_userfault(vmf, VM_UFFD_MISSING); |
6b251fc9 AA |
748 | VM_BUG_ON(ret & VM_FAULT_FALLBACK); |
749 | } else { | |
bae473a4 | 750 | set_huge_zero_page(pgtable, vma->vm_mm, vma, |
82b0f8c3 | 751 | haddr, vmf->pmd, zero_page); |
fca40573 | 752 | update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); |
82b0f8c3 | 753 | spin_unlock(vmf->ptl); |
6b251fc9 | 754 | } |
bfe8cc1d | 755 | } else { |
82b0f8c3 | 756 | spin_unlock(vmf->ptl); |
bae473a4 | 757 | pte_free(vma->vm_mm, pgtable); |
bfe8cc1d | 758 | } |
6b251fc9 | 759 | return ret; |
71e3aac0 | 760 | } |
19deb769 DR |
761 | gfp = alloc_hugepage_direct_gfpmask(vma); |
762 | page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); | |
128ec037 KS |
763 | if (unlikely(!page)) { |
764 | count_vm_event(THP_FAULT_FALLBACK); | |
c0292554 | 765 | return VM_FAULT_FALLBACK; |
128ec037 | 766 | } |
9a982250 | 767 | prep_transhuge_page(page); |
82b0f8c3 | 768 | return __do_huge_pmd_anonymous_page(vmf, page, gfp); |
71e3aac0 AA |
769 | } |
770 | ||
ae18d6dc | 771 | static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, |
3b6521f5 OH |
772 | pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write, |
773 | pgtable_t pgtable) | |
5cad465d MW |
774 | { |
775 | struct mm_struct *mm = vma->vm_mm; | |
776 | pmd_t entry; | |
777 | spinlock_t *ptl; | |
778 | ||
779 | ptl = pmd_lock(mm, pmd); | |
c6f3c5ee AK |
780 | if (!pmd_none(*pmd)) { |
781 | if (write) { | |
782 | if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) { | |
783 | WARN_ON_ONCE(!is_huge_zero_pmd(*pmd)); | |
784 | goto out_unlock; | |
785 | } | |
786 | entry = pmd_mkyoung(*pmd); | |
787 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
788 | if (pmdp_set_access_flags(vma, addr, pmd, entry, 1)) | |
789 | update_mmu_cache_pmd(vma, addr, pmd); | |
790 | } | |
791 | ||
792 | goto out_unlock; | |
793 | } | |
794 | ||
f25748e3 DW |
795 | entry = pmd_mkhuge(pfn_t_pmd(pfn, prot)); |
796 | if (pfn_t_devmap(pfn)) | |
797 | entry = pmd_mkdevmap(entry); | |
01871e59 | 798 | if (write) { |
f55e1014 LT |
799 | entry = pmd_mkyoung(pmd_mkdirty(entry)); |
800 | entry = maybe_pmd_mkwrite(entry, vma); | |
5cad465d | 801 | } |
3b6521f5 OH |
802 | |
803 | if (pgtable) { | |
804 | pgtable_trans_huge_deposit(mm, pmd, pgtable); | |
c4812909 | 805 | mm_inc_nr_ptes(mm); |
c6f3c5ee | 806 | pgtable = NULL; |
3b6521f5 OH |
807 | } |
808 | ||
01871e59 RZ |
809 | set_pmd_at(mm, addr, pmd, entry); |
810 | update_mmu_cache_pmd(vma, addr, pmd); | |
c6f3c5ee AK |
811 | |
812 | out_unlock: | |
5cad465d | 813 | spin_unlock(ptl); |
c6f3c5ee AK |
814 | if (pgtable) |
815 | pte_free(mm, pgtable); | |
5cad465d MW |
816 | } |
817 | ||
9a9731b1 THV |
818 | /** |
819 | * vmf_insert_pfn_pmd_prot - insert a pmd size pfn | |
820 | * @vmf: Structure describing the fault | |
821 | * @pfn: pfn to insert | |
822 | * @pgprot: page protection to use | |
823 | * @write: whether it's a write fault | |
824 | * | |
825 | * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and | |
826 | * also consult the vmf_insert_mixed_prot() documentation when | |
827 | * @pgprot != @vmf->vma->vm_page_prot. | |
828 | * | |
829 | * Return: vm_fault_t value. | |
830 | */ | |
831 | vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn, | |
832 | pgprot_t pgprot, bool write) | |
5cad465d | 833 | { |
fce86ff5 DW |
834 | unsigned long addr = vmf->address & PMD_MASK; |
835 | struct vm_area_struct *vma = vmf->vma; | |
3b6521f5 | 836 | pgtable_t pgtable = NULL; |
fce86ff5 | 837 | |
5cad465d MW |
838 | /* |
839 | * If we had pmd_special, we could avoid all these restrictions, | |
840 | * but we need to be consistent with PTEs and architectures that | |
841 | * can't support a 'special' bit. | |
842 | */ | |
e1fb4a08 DJ |
843 | BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && |
844 | !pfn_t_devmap(pfn)); | |
5cad465d MW |
845 | BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == |
846 | (VM_PFNMAP|VM_MIXEDMAP)); | |
847 | BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); | |
5cad465d MW |
848 | |
849 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
850 | return VM_FAULT_SIGBUS; | |
308a047c | 851 | |
3b6521f5 | 852 | if (arch_needs_pgtable_deposit()) { |
4cf58924 | 853 | pgtable = pte_alloc_one(vma->vm_mm); |
3b6521f5 OH |
854 | if (!pgtable) |
855 | return VM_FAULT_OOM; | |
856 | } | |
857 | ||
308a047c BP |
858 | track_pfn_insert(vma, &pgprot, pfn); |
859 | ||
fce86ff5 | 860 | insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable); |
ae18d6dc | 861 | return VM_FAULT_NOPAGE; |
5cad465d | 862 | } |
9a9731b1 | 863 | EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot); |
5cad465d | 864 | |
a00cc7d9 | 865 | #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD |
f55e1014 | 866 | static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma) |
a00cc7d9 | 867 | { |
f55e1014 | 868 | if (likely(vma->vm_flags & VM_WRITE)) |
a00cc7d9 MW |
869 | pud = pud_mkwrite(pud); |
870 | return pud; | |
871 | } | |
872 | ||
873 | static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, | |
874 | pud_t *pud, pfn_t pfn, pgprot_t prot, bool write) | |
875 | { | |
876 | struct mm_struct *mm = vma->vm_mm; | |
877 | pud_t entry; | |
878 | spinlock_t *ptl; | |
879 | ||
880 | ptl = pud_lock(mm, pud); | |
c6f3c5ee AK |
881 | if (!pud_none(*pud)) { |
882 | if (write) { | |
883 | if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) { | |
884 | WARN_ON_ONCE(!is_huge_zero_pud(*pud)); | |
885 | goto out_unlock; | |
886 | } | |
887 | entry = pud_mkyoung(*pud); | |
888 | entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma); | |
889 | if (pudp_set_access_flags(vma, addr, pud, entry, 1)) | |
890 | update_mmu_cache_pud(vma, addr, pud); | |
891 | } | |
892 | goto out_unlock; | |
893 | } | |
894 | ||
a00cc7d9 MW |
895 | entry = pud_mkhuge(pfn_t_pud(pfn, prot)); |
896 | if (pfn_t_devmap(pfn)) | |
897 | entry = pud_mkdevmap(entry); | |
898 | if (write) { | |
f55e1014 LT |
899 | entry = pud_mkyoung(pud_mkdirty(entry)); |
900 | entry = maybe_pud_mkwrite(entry, vma); | |
a00cc7d9 MW |
901 | } |
902 | set_pud_at(mm, addr, pud, entry); | |
903 | update_mmu_cache_pud(vma, addr, pud); | |
c6f3c5ee AK |
904 | |
905 | out_unlock: | |
a00cc7d9 MW |
906 | spin_unlock(ptl); |
907 | } | |
908 | ||
9a9731b1 THV |
909 | /** |
910 | * vmf_insert_pfn_pud_prot - insert a pud size pfn | |
911 | * @vmf: Structure describing the fault | |
912 | * @pfn: pfn to insert | |
913 | * @pgprot: page protection to use | |
914 | * @write: whether it's a write fault | |
915 | * | |
916 | * Insert a pud size pfn. See vmf_insert_pfn() for additional info and | |
917 | * also consult the vmf_insert_mixed_prot() documentation when | |
918 | * @pgprot != @vmf->vma->vm_page_prot. | |
919 | * | |
920 | * Return: vm_fault_t value. | |
921 | */ | |
922 | vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn, | |
923 | pgprot_t pgprot, bool write) | |
a00cc7d9 | 924 | { |
fce86ff5 DW |
925 | unsigned long addr = vmf->address & PUD_MASK; |
926 | struct vm_area_struct *vma = vmf->vma; | |
fce86ff5 | 927 | |
a00cc7d9 MW |
928 | /* |
929 | * If we had pud_special, we could avoid all these restrictions, | |
930 | * but we need to be consistent with PTEs and architectures that | |
931 | * can't support a 'special' bit. | |
932 | */ | |
62ec0d8c DJ |
933 | BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && |
934 | !pfn_t_devmap(pfn)); | |
a00cc7d9 MW |
935 | BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == |
936 | (VM_PFNMAP|VM_MIXEDMAP)); | |
937 | BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); | |
a00cc7d9 MW |
938 | |
939 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
940 | return VM_FAULT_SIGBUS; | |
941 | ||
942 | track_pfn_insert(vma, &pgprot, pfn); | |
943 | ||
fce86ff5 | 944 | insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write); |
a00cc7d9 MW |
945 | return VM_FAULT_NOPAGE; |
946 | } | |
9a9731b1 | 947 | EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot); |
a00cc7d9 MW |
948 | #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ |
949 | ||
3565fce3 | 950 | static void touch_pmd(struct vm_area_struct *vma, unsigned long addr, |
a8f97366 | 951 | pmd_t *pmd, int flags) |
3565fce3 DW |
952 | { |
953 | pmd_t _pmd; | |
954 | ||
a8f97366 KS |
955 | _pmd = pmd_mkyoung(*pmd); |
956 | if (flags & FOLL_WRITE) | |
957 | _pmd = pmd_mkdirty(_pmd); | |
3565fce3 | 958 | if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, |
a8f97366 | 959 | pmd, _pmd, flags & FOLL_WRITE)) |
3565fce3 DW |
960 | update_mmu_cache_pmd(vma, addr, pmd); |
961 | } | |
962 | ||
963 | struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr, | |
df06b37f | 964 | pmd_t *pmd, int flags, struct dev_pagemap **pgmap) |
3565fce3 DW |
965 | { |
966 | unsigned long pfn = pmd_pfn(*pmd); | |
967 | struct mm_struct *mm = vma->vm_mm; | |
3565fce3 DW |
968 | struct page *page; |
969 | ||
970 | assert_spin_locked(pmd_lockptr(mm, pmd)); | |
971 | ||
8310d48b KF |
972 | /* |
973 | * When we COW a devmap PMD entry, we split it into PTEs, so we should | |
974 | * not be in this function with `flags & FOLL_COW` set. | |
975 | */ | |
976 | WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set"); | |
977 | ||
3faa52c0 JH |
978 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ |
979 | if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == | |
980 | (FOLL_PIN | FOLL_GET))) | |
981 | return NULL; | |
982 | ||
f6f37321 | 983 | if (flags & FOLL_WRITE && !pmd_write(*pmd)) |
3565fce3 DW |
984 | return NULL; |
985 | ||
986 | if (pmd_present(*pmd) && pmd_devmap(*pmd)) | |
987 | /* pass */; | |
988 | else | |
989 | return NULL; | |
990 | ||
991 | if (flags & FOLL_TOUCH) | |
a8f97366 | 992 | touch_pmd(vma, addr, pmd, flags); |
3565fce3 DW |
993 | |
994 | /* | |
995 | * device mapped pages can only be returned if the | |
996 | * caller will manage the page reference count. | |
997 | */ | |
3faa52c0 | 998 | if (!(flags & (FOLL_GET | FOLL_PIN))) |
3565fce3 DW |
999 | return ERR_PTR(-EEXIST); |
1000 | ||
1001 | pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT; | |
df06b37f KB |
1002 | *pgmap = get_dev_pagemap(pfn, *pgmap); |
1003 | if (!*pgmap) | |
3565fce3 DW |
1004 | return ERR_PTR(-EFAULT); |
1005 | page = pfn_to_page(pfn); | |
3faa52c0 JH |
1006 | if (!try_grab_page(page, flags)) |
1007 | page = ERR_PTR(-ENOMEM); | |
3565fce3 DW |
1008 | |
1009 | return page; | |
1010 | } | |
1011 | ||
71e3aac0 AA |
1012 | int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, |
1013 | pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, | |
1014 | struct vm_area_struct *vma) | |
1015 | { | |
c4088ebd | 1016 | spinlock_t *dst_ptl, *src_ptl; |
71e3aac0 AA |
1017 | struct page *src_page; |
1018 | pmd_t pmd; | |
12c9d70b | 1019 | pgtable_t pgtable = NULL; |
628d47ce | 1020 | int ret = -ENOMEM; |
71e3aac0 | 1021 | |
628d47ce KS |
1022 | /* Skip if can be re-fill on fault */ |
1023 | if (!vma_is_anonymous(vma)) | |
1024 | return 0; | |
1025 | ||
4cf58924 | 1026 | pgtable = pte_alloc_one(dst_mm); |
628d47ce KS |
1027 | if (unlikely(!pgtable)) |
1028 | goto out; | |
71e3aac0 | 1029 | |
c4088ebd KS |
1030 | dst_ptl = pmd_lock(dst_mm, dst_pmd); |
1031 | src_ptl = pmd_lockptr(src_mm, src_pmd); | |
1032 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
71e3aac0 AA |
1033 | |
1034 | ret = -EAGAIN; | |
1035 | pmd = *src_pmd; | |
84c3fc4e | 1036 | |
b569a176 PX |
1037 | /* |
1038 | * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA | |
1039 | * does not have the VM_UFFD_WP, which means that the uffd | |
1040 | * fork event is not enabled. | |
1041 | */ | |
1042 | if (!(vma->vm_flags & VM_UFFD_WP)) | |
1043 | pmd = pmd_clear_uffd_wp(pmd); | |
1044 | ||
84c3fc4e ZY |
1045 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
1046 | if (unlikely(is_swap_pmd(pmd))) { | |
1047 | swp_entry_t entry = pmd_to_swp_entry(pmd); | |
1048 | ||
1049 | VM_BUG_ON(!is_pmd_migration_entry(pmd)); | |
1050 | if (is_write_migration_entry(entry)) { | |
1051 | make_migration_entry_read(&entry); | |
1052 | pmd = swp_entry_to_pmd(entry); | |
ab6e3d09 NH |
1053 | if (pmd_swp_soft_dirty(*src_pmd)) |
1054 | pmd = pmd_swp_mksoft_dirty(pmd); | |
84c3fc4e ZY |
1055 | set_pmd_at(src_mm, addr, src_pmd, pmd); |
1056 | } | |
dd8a67f9 | 1057 | add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); |
af5b0f6a | 1058 | mm_inc_nr_ptes(dst_mm); |
dd8a67f9 | 1059 | pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); |
84c3fc4e ZY |
1060 | set_pmd_at(dst_mm, addr, dst_pmd, pmd); |
1061 | ret = 0; | |
1062 | goto out_unlock; | |
1063 | } | |
1064 | #endif | |
1065 | ||
628d47ce | 1066 | if (unlikely(!pmd_trans_huge(pmd))) { |
71e3aac0 AA |
1067 | pte_free(dst_mm, pgtable); |
1068 | goto out_unlock; | |
1069 | } | |
fc9fe822 | 1070 | /* |
c4088ebd | 1071 | * When page table lock is held, the huge zero pmd should not be |
fc9fe822 KS |
1072 | * under splitting since we don't split the page itself, only pmd to |
1073 | * a page table. | |
1074 | */ | |
1075 | if (is_huge_zero_pmd(pmd)) { | |
5918d10a | 1076 | struct page *zero_page; |
97ae1749 KS |
1077 | /* |
1078 | * get_huge_zero_page() will never allocate a new page here, | |
1079 | * since we already have a zero page to copy. It just takes a | |
1080 | * reference. | |
1081 | */ | |
6fcb52a5 | 1082 | zero_page = mm_get_huge_zero_page(dst_mm); |
6b251fc9 | 1083 | set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, |
5918d10a | 1084 | zero_page); |
fc9fe822 KS |
1085 | ret = 0; |
1086 | goto out_unlock; | |
1087 | } | |
de466bd6 | 1088 | |
628d47ce KS |
1089 | src_page = pmd_page(pmd); |
1090 | VM_BUG_ON_PAGE(!PageHead(src_page), src_page); | |
d042035e PX |
1091 | |
1092 | /* | |
1093 | * If this page is a potentially pinned page, split and retry the fault | |
1094 | * with smaller page size. Normally this should not happen because the | |
1095 | * userspace should use MADV_DONTFORK upon pinned regions. This is a | |
1096 | * best effort that the pinned pages won't be replaced by another | |
1097 | * random page during the coming copy-on-write. | |
1098 | */ | |
1099 | if (unlikely(is_cow_mapping(vma->vm_flags) && | |
1100 | atomic_read(&src_mm->has_pinned) && | |
1101 | page_maybe_dma_pinned(src_page))) { | |
1102 | pte_free(dst_mm, pgtable); | |
1103 | spin_unlock(src_ptl); | |
1104 | spin_unlock(dst_ptl); | |
1105 | __split_huge_pmd(vma, src_pmd, addr, false, NULL); | |
1106 | return -EAGAIN; | |
1107 | } | |
1108 | ||
628d47ce KS |
1109 | get_page(src_page); |
1110 | page_dup_rmap(src_page, true); | |
1111 | add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); | |
c4812909 | 1112 | mm_inc_nr_ptes(dst_mm); |
628d47ce | 1113 | pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); |
71e3aac0 AA |
1114 | |
1115 | pmdp_set_wrprotect(src_mm, addr, src_pmd); | |
1116 | pmd = pmd_mkold(pmd_wrprotect(pmd)); | |
1117 | set_pmd_at(dst_mm, addr, dst_pmd, pmd); | |
71e3aac0 AA |
1118 | |
1119 | ret = 0; | |
1120 | out_unlock: | |
c4088ebd KS |
1121 | spin_unlock(src_ptl); |
1122 | spin_unlock(dst_ptl); | |
71e3aac0 AA |
1123 | out: |
1124 | return ret; | |
1125 | } | |
1126 | ||
a00cc7d9 MW |
1127 | #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD |
1128 | static void touch_pud(struct vm_area_struct *vma, unsigned long addr, | |
a8f97366 | 1129 | pud_t *pud, int flags) |
a00cc7d9 MW |
1130 | { |
1131 | pud_t _pud; | |
1132 | ||
a8f97366 KS |
1133 | _pud = pud_mkyoung(*pud); |
1134 | if (flags & FOLL_WRITE) | |
1135 | _pud = pud_mkdirty(_pud); | |
a00cc7d9 | 1136 | if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK, |
a8f97366 | 1137 | pud, _pud, flags & FOLL_WRITE)) |
a00cc7d9 MW |
1138 | update_mmu_cache_pud(vma, addr, pud); |
1139 | } | |
1140 | ||
1141 | struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr, | |
df06b37f | 1142 | pud_t *pud, int flags, struct dev_pagemap **pgmap) |
a00cc7d9 MW |
1143 | { |
1144 | unsigned long pfn = pud_pfn(*pud); | |
1145 | struct mm_struct *mm = vma->vm_mm; | |
a00cc7d9 MW |
1146 | struct page *page; |
1147 | ||
1148 | assert_spin_locked(pud_lockptr(mm, pud)); | |
1149 | ||
f6f37321 | 1150 | if (flags & FOLL_WRITE && !pud_write(*pud)) |
a00cc7d9 MW |
1151 | return NULL; |
1152 | ||
3faa52c0 JH |
1153 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ |
1154 | if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == | |
1155 | (FOLL_PIN | FOLL_GET))) | |
1156 | return NULL; | |
1157 | ||
a00cc7d9 MW |
1158 | if (pud_present(*pud) && pud_devmap(*pud)) |
1159 | /* pass */; | |
1160 | else | |
1161 | return NULL; | |
1162 | ||
1163 | if (flags & FOLL_TOUCH) | |
a8f97366 | 1164 | touch_pud(vma, addr, pud, flags); |
a00cc7d9 MW |
1165 | |
1166 | /* | |
1167 | * device mapped pages can only be returned if the | |
1168 | * caller will manage the page reference count. | |
3faa52c0 JH |
1169 | * |
1170 | * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here: | |
a00cc7d9 | 1171 | */ |
3faa52c0 | 1172 | if (!(flags & (FOLL_GET | FOLL_PIN))) |
a00cc7d9 MW |
1173 | return ERR_PTR(-EEXIST); |
1174 | ||
1175 | pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT; | |
df06b37f KB |
1176 | *pgmap = get_dev_pagemap(pfn, *pgmap); |
1177 | if (!*pgmap) | |
a00cc7d9 MW |
1178 | return ERR_PTR(-EFAULT); |
1179 | page = pfn_to_page(pfn); | |
3faa52c0 JH |
1180 | if (!try_grab_page(page, flags)) |
1181 | page = ERR_PTR(-ENOMEM); | |
a00cc7d9 MW |
1182 | |
1183 | return page; | |
1184 | } | |
1185 | ||
1186 | int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm, | |
1187 | pud_t *dst_pud, pud_t *src_pud, unsigned long addr, | |
1188 | struct vm_area_struct *vma) | |
1189 | { | |
1190 | spinlock_t *dst_ptl, *src_ptl; | |
1191 | pud_t pud; | |
1192 | int ret; | |
1193 | ||
1194 | dst_ptl = pud_lock(dst_mm, dst_pud); | |
1195 | src_ptl = pud_lockptr(src_mm, src_pud); | |
1196 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
1197 | ||
1198 | ret = -EAGAIN; | |
1199 | pud = *src_pud; | |
1200 | if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud))) | |
1201 | goto out_unlock; | |
1202 | ||
1203 | /* | |
1204 | * When page table lock is held, the huge zero pud should not be | |
1205 | * under splitting since we don't split the page itself, only pud to | |
1206 | * a page table. | |
1207 | */ | |
1208 | if (is_huge_zero_pud(pud)) { | |
1209 | /* No huge zero pud yet */ | |
1210 | } | |
1211 | ||
d042035e PX |
1212 | /* Please refer to comments in copy_huge_pmd() */ |
1213 | if (unlikely(is_cow_mapping(vma->vm_flags) && | |
1214 | atomic_read(&src_mm->has_pinned) && | |
1215 | page_maybe_dma_pinned(pud_page(pud)))) { | |
1216 | spin_unlock(src_ptl); | |
1217 | spin_unlock(dst_ptl); | |
1218 | __split_huge_pud(vma, src_pud, addr); | |
1219 | return -EAGAIN; | |
1220 | } | |
1221 | ||
a00cc7d9 MW |
1222 | pudp_set_wrprotect(src_mm, addr, src_pud); |
1223 | pud = pud_mkold(pud_wrprotect(pud)); | |
1224 | set_pud_at(dst_mm, addr, dst_pud, pud); | |
1225 | ||
1226 | ret = 0; | |
1227 | out_unlock: | |
1228 | spin_unlock(src_ptl); | |
1229 | spin_unlock(dst_ptl); | |
1230 | return ret; | |
1231 | } | |
1232 | ||
1233 | void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud) | |
1234 | { | |
1235 | pud_t entry; | |
1236 | unsigned long haddr; | |
1237 | bool write = vmf->flags & FAULT_FLAG_WRITE; | |
1238 | ||
1239 | vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud); | |
1240 | if (unlikely(!pud_same(*vmf->pud, orig_pud))) | |
1241 | goto unlock; | |
1242 | ||
1243 | entry = pud_mkyoung(orig_pud); | |
1244 | if (write) | |
1245 | entry = pud_mkdirty(entry); | |
1246 | haddr = vmf->address & HPAGE_PUD_MASK; | |
1247 | if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write)) | |
1248 | update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud); | |
1249 | ||
1250 | unlock: | |
1251 | spin_unlock(vmf->ptl); | |
1252 | } | |
1253 | #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ | |
1254 | ||
82b0f8c3 | 1255 | void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd) |
a1dd450b WD |
1256 | { |
1257 | pmd_t entry; | |
1258 | unsigned long haddr; | |
20f664aa | 1259 | bool write = vmf->flags & FAULT_FLAG_WRITE; |
a1dd450b | 1260 | |
82b0f8c3 JK |
1261 | vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); |
1262 | if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) | |
a1dd450b WD |
1263 | goto unlock; |
1264 | ||
1265 | entry = pmd_mkyoung(orig_pmd); | |
20f664aa MK |
1266 | if (write) |
1267 | entry = pmd_mkdirty(entry); | |
82b0f8c3 | 1268 | haddr = vmf->address & HPAGE_PMD_MASK; |
20f664aa | 1269 | if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write)) |
82b0f8c3 | 1270 | update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd); |
a1dd450b WD |
1271 | |
1272 | unlock: | |
82b0f8c3 | 1273 | spin_unlock(vmf->ptl); |
a1dd450b WD |
1274 | } |
1275 | ||
2b740303 | 1276 | vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd) |
71e3aac0 | 1277 | { |
82b0f8c3 | 1278 | struct vm_area_struct *vma = vmf->vma; |
3917c802 | 1279 | struct page *page; |
82b0f8c3 | 1280 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; |
71e3aac0 | 1281 | |
82b0f8c3 | 1282 | vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd); |
81d1b09c | 1283 | VM_BUG_ON_VMA(!vma->anon_vma, vma); |
3917c802 | 1284 | |
93b4796d | 1285 | if (is_huge_zero_pmd(orig_pmd)) |
3917c802 KS |
1286 | goto fallback; |
1287 | ||
82b0f8c3 | 1288 | spin_lock(vmf->ptl); |
3917c802 KS |
1289 | |
1290 | if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { | |
1291 | spin_unlock(vmf->ptl); | |
1292 | return 0; | |
1293 | } | |
71e3aac0 AA |
1294 | |
1295 | page = pmd_page(orig_pmd); | |
309381fe | 1296 | VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); |
3917c802 KS |
1297 | |
1298 | /* Lock page for reuse_swap_page() */ | |
ba3c4ce6 HY |
1299 | if (!trylock_page(page)) { |
1300 | get_page(page); | |
1301 | spin_unlock(vmf->ptl); | |
1302 | lock_page(page); | |
1303 | spin_lock(vmf->ptl); | |
1304 | if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { | |
3917c802 | 1305 | spin_unlock(vmf->ptl); |
ba3c4ce6 HY |
1306 | unlock_page(page); |
1307 | put_page(page); | |
3917c802 | 1308 | return 0; |
ba3c4ce6 HY |
1309 | } |
1310 | put_page(page); | |
1311 | } | |
3917c802 KS |
1312 | |
1313 | /* | |
1314 | * We can only reuse the page if nobody else maps the huge page or it's | |
1315 | * part. | |
1316 | */ | |
ba3c4ce6 | 1317 | if (reuse_swap_page(page, NULL)) { |
71e3aac0 AA |
1318 | pmd_t entry; |
1319 | entry = pmd_mkyoung(orig_pmd); | |
f55e1014 | 1320 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
3917c802 | 1321 | if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1)) |
82b0f8c3 | 1322 | update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); |
ba3c4ce6 | 1323 | unlock_page(page); |
82b0f8c3 | 1324 | spin_unlock(vmf->ptl); |
3917c802 | 1325 | return VM_FAULT_WRITE; |
71e3aac0 | 1326 | } |
3917c802 KS |
1327 | |
1328 | unlock_page(page); | |
82b0f8c3 | 1329 | spin_unlock(vmf->ptl); |
3917c802 KS |
1330 | fallback: |
1331 | __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL); | |
1332 | return VM_FAULT_FALLBACK; | |
71e3aac0 AA |
1333 | } |
1334 | ||
8310d48b | 1335 | /* |
a308c71b PX |
1336 | * FOLL_FORCE can write to even unwritable pmd's, but only |
1337 | * after we've gone through a COW cycle and they are dirty. | |
8310d48b KF |
1338 | */ |
1339 | static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags) | |
1340 | { | |
a308c71b PX |
1341 | return pmd_write(pmd) || |
1342 | ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd)); | |
8310d48b KF |
1343 | } |
1344 | ||
b676b293 | 1345 | struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, |
71e3aac0 AA |
1346 | unsigned long addr, |
1347 | pmd_t *pmd, | |
1348 | unsigned int flags) | |
1349 | { | |
b676b293 | 1350 | struct mm_struct *mm = vma->vm_mm; |
71e3aac0 AA |
1351 | struct page *page = NULL; |
1352 | ||
c4088ebd | 1353 | assert_spin_locked(pmd_lockptr(mm, pmd)); |
71e3aac0 | 1354 | |
8310d48b | 1355 | if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags)) |
71e3aac0 AA |
1356 | goto out; |
1357 | ||
85facf25 KS |
1358 | /* Avoid dumping huge zero page */ |
1359 | if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) | |
1360 | return ERR_PTR(-EFAULT); | |
1361 | ||
2b4847e7 | 1362 | /* Full NUMA hinting faults to serialise migration in fault paths */ |
8a0516ed | 1363 | if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) |
2b4847e7 MG |
1364 | goto out; |
1365 | ||
71e3aac0 | 1366 | page = pmd_page(*pmd); |
ca120cf6 | 1367 | VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page); |
3faa52c0 JH |
1368 | |
1369 | if (!try_grab_page(page, flags)) | |
1370 | return ERR_PTR(-ENOMEM); | |
1371 | ||
3565fce3 | 1372 | if (flags & FOLL_TOUCH) |
a8f97366 | 1373 | touch_pmd(vma, addr, pmd, flags); |
3faa52c0 | 1374 | |
de60f5f1 | 1375 | if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { |
e90309c9 KS |
1376 | /* |
1377 | * We don't mlock() pte-mapped THPs. This way we can avoid | |
1378 | * leaking mlocked pages into non-VM_LOCKED VMAs. | |
1379 | * | |
9a73f61b KS |
1380 | * For anon THP: |
1381 | * | |
e90309c9 KS |
1382 | * In most cases the pmd is the only mapping of the page as we |
1383 | * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for | |
1384 | * writable private mappings in populate_vma_page_range(). | |
1385 | * | |
1386 | * The only scenario when we have the page shared here is if we | |
1387 | * mlocking read-only mapping shared over fork(). We skip | |
1388 | * mlocking such pages. | |
9a73f61b KS |
1389 | * |
1390 | * For file THP: | |
1391 | * | |
1392 | * We can expect PageDoubleMap() to be stable under page lock: | |
1393 | * for file pages we set it in page_add_file_rmap(), which | |
1394 | * requires page to be locked. | |
e90309c9 | 1395 | */ |
9a73f61b KS |
1396 | |
1397 | if (PageAnon(page) && compound_mapcount(page) != 1) | |
1398 | goto skip_mlock; | |
1399 | if (PageDoubleMap(page) || !page->mapping) | |
1400 | goto skip_mlock; | |
1401 | if (!trylock_page(page)) | |
1402 | goto skip_mlock; | |
9a73f61b KS |
1403 | if (page->mapping && !PageDoubleMap(page)) |
1404 | mlock_vma_page(page); | |
1405 | unlock_page(page); | |
b676b293 | 1406 | } |
9a73f61b | 1407 | skip_mlock: |
71e3aac0 | 1408 | page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; |
ca120cf6 | 1409 | VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page); |
71e3aac0 AA |
1410 | |
1411 | out: | |
1412 | return page; | |
1413 | } | |
1414 | ||
d10e63f2 | 1415 | /* NUMA hinting page fault entry point for trans huge pmds */ |
2b740303 | 1416 | vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd) |
d10e63f2 | 1417 | { |
82b0f8c3 | 1418 | struct vm_area_struct *vma = vmf->vma; |
b8916634 | 1419 | struct anon_vma *anon_vma = NULL; |
b32967ff | 1420 | struct page *page; |
82b0f8c3 | 1421 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; |
98fa15f3 | 1422 | int page_nid = NUMA_NO_NODE, this_nid = numa_node_id(); |
90572890 | 1423 | int target_nid, last_cpupid = -1; |
8191acbd MG |
1424 | bool page_locked; |
1425 | bool migrated = false; | |
b191f9b1 | 1426 | bool was_writable; |
6688cc05 | 1427 | int flags = 0; |
d10e63f2 | 1428 | |
82b0f8c3 JK |
1429 | vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); |
1430 | if (unlikely(!pmd_same(pmd, *vmf->pmd))) | |
d10e63f2 MG |
1431 | goto out_unlock; |
1432 | ||
de466bd6 MG |
1433 | /* |
1434 | * If there are potential migrations, wait for completion and retry | |
1435 | * without disrupting NUMA hinting information. Do not relock and | |
1436 | * check_same as the page may no longer be mapped. | |
1437 | */ | |
82b0f8c3 JK |
1438 | if (unlikely(pmd_trans_migrating(*vmf->pmd))) { |
1439 | page = pmd_page(*vmf->pmd); | |
3c226c63 MR |
1440 | if (!get_page_unless_zero(page)) |
1441 | goto out_unlock; | |
82b0f8c3 | 1442 | spin_unlock(vmf->ptl); |
48054625 | 1443 | put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE); |
de466bd6 MG |
1444 | goto out; |
1445 | } | |
1446 | ||
d10e63f2 | 1447 | page = pmd_page(pmd); |
a1a46184 | 1448 | BUG_ON(is_huge_zero_page(page)); |
8191acbd | 1449 | page_nid = page_to_nid(page); |
90572890 | 1450 | last_cpupid = page_cpupid_last(page); |
03c5a6e1 | 1451 | count_vm_numa_event(NUMA_HINT_FAULTS); |
04bb2f94 | 1452 | if (page_nid == this_nid) { |
03c5a6e1 | 1453 | count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); |
04bb2f94 RR |
1454 | flags |= TNF_FAULT_LOCAL; |
1455 | } | |
4daae3b4 | 1456 | |
bea66fbd | 1457 | /* See similar comment in do_numa_page for explanation */ |
288bc549 | 1458 | if (!pmd_savedwrite(pmd)) |
6688cc05 PZ |
1459 | flags |= TNF_NO_GROUP; |
1460 | ||
ff9042b1 MG |
1461 | /* |
1462 | * Acquire the page lock to serialise THP migrations but avoid dropping | |
1463 | * page_table_lock if at all possible | |
1464 | */ | |
b8916634 MG |
1465 | page_locked = trylock_page(page); |
1466 | target_nid = mpol_misplaced(page, vma, haddr); | |
98fa15f3 | 1467 | if (target_nid == NUMA_NO_NODE) { |
b8916634 | 1468 | /* If the page was locked, there are no parallel migrations */ |
a54a407f | 1469 | if (page_locked) |
b8916634 | 1470 | goto clear_pmdnuma; |
2b4847e7 | 1471 | } |
4daae3b4 | 1472 | |
de466bd6 | 1473 | /* Migration could have started since the pmd_trans_migrating check */ |
2b4847e7 | 1474 | if (!page_locked) { |
98fa15f3 | 1475 | page_nid = NUMA_NO_NODE; |
3c226c63 MR |
1476 | if (!get_page_unless_zero(page)) |
1477 | goto out_unlock; | |
82b0f8c3 | 1478 | spin_unlock(vmf->ptl); |
48054625 | 1479 | put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE); |
b8916634 MG |
1480 | goto out; |
1481 | } | |
1482 | ||
2b4847e7 MG |
1483 | /* |
1484 | * Page is misplaced. Page lock serialises migrations. Acquire anon_vma | |
1485 | * to serialises splits | |
1486 | */ | |
b8916634 | 1487 | get_page(page); |
82b0f8c3 | 1488 | spin_unlock(vmf->ptl); |
b8916634 | 1489 | anon_vma = page_lock_anon_vma_read(page); |
4daae3b4 | 1490 | |
c69307d5 | 1491 | /* Confirm the PMD did not change while page_table_lock was released */ |
82b0f8c3 JK |
1492 | spin_lock(vmf->ptl); |
1493 | if (unlikely(!pmd_same(pmd, *vmf->pmd))) { | |
b32967ff MG |
1494 | unlock_page(page); |
1495 | put_page(page); | |
98fa15f3 | 1496 | page_nid = NUMA_NO_NODE; |
4daae3b4 | 1497 | goto out_unlock; |
b32967ff | 1498 | } |
ff9042b1 | 1499 | |
c3a489ca MG |
1500 | /* Bail if we fail to protect against THP splits for any reason */ |
1501 | if (unlikely(!anon_vma)) { | |
1502 | put_page(page); | |
98fa15f3 | 1503 | page_nid = NUMA_NO_NODE; |
c3a489ca MG |
1504 | goto clear_pmdnuma; |
1505 | } | |
1506 | ||
8b1b436d PZ |
1507 | /* |
1508 | * Since we took the NUMA fault, we must have observed the !accessible | |
1509 | * bit. Make sure all other CPUs agree with that, to avoid them | |
1510 | * modifying the page we're about to migrate. | |
1511 | * | |
1512 | * Must be done under PTL such that we'll observe the relevant | |
ccde85ba PZ |
1513 | * inc_tlb_flush_pending(). |
1514 | * | |
1515 | * We are not sure a pending tlb flush here is for a huge page | |
1516 | * mapping or not. Hence use the tlb range variant | |
8b1b436d | 1517 | */ |
7066f0f9 | 1518 | if (mm_tlb_flush_pending(vma->vm_mm)) { |
ccde85ba | 1519 | flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE); |
7066f0f9 AA |
1520 | /* |
1521 | * change_huge_pmd() released the pmd lock before | |
1522 | * invalidating the secondary MMUs sharing the primary | |
1523 | * MMU pagetables (with ->invalidate_range()). The | |
1524 | * mmu_notifier_invalidate_range_end() (which | |
1525 | * internally calls ->invalidate_range()) in | |
1526 | * change_pmd_range() will run after us, so we can't | |
1527 | * rely on it here and we need an explicit invalidate. | |
1528 | */ | |
1529 | mmu_notifier_invalidate_range(vma->vm_mm, haddr, | |
1530 | haddr + HPAGE_PMD_SIZE); | |
1531 | } | |
8b1b436d | 1532 | |
a54a407f MG |
1533 | /* |
1534 | * Migrate the THP to the requested node, returns with page unlocked | |
8a0516ed | 1535 | * and access rights restored. |
a54a407f | 1536 | */ |
82b0f8c3 | 1537 | spin_unlock(vmf->ptl); |
8b1b436d | 1538 | |
bae473a4 | 1539 | migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma, |
82b0f8c3 | 1540 | vmf->pmd, pmd, vmf->address, page, target_nid); |
6688cc05 PZ |
1541 | if (migrated) { |
1542 | flags |= TNF_MIGRATED; | |
8191acbd | 1543 | page_nid = target_nid; |
074c2381 MG |
1544 | } else |
1545 | flags |= TNF_MIGRATE_FAIL; | |
b32967ff | 1546 | |
8191acbd | 1547 | goto out; |
b32967ff | 1548 | clear_pmdnuma: |
a54a407f | 1549 | BUG_ON(!PageLocked(page)); |
288bc549 | 1550 | was_writable = pmd_savedwrite(pmd); |
4d942466 | 1551 | pmd = pmd_modify(pmd, vma->vm_page_prot); |
b7b04004 | 1552 | pmd = pmd_mkyoung(pmd); |
b191f9b1 MG |
1553 | if (was_writable) |
1554 | pmd = pmd_mkwrite(pmd); | |
82b0f8c3 JK |
1555 | set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd); |
1556 | update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); | |
a54a407f | 1557 | unlock_page(page); |
d10e63f2 | 1558 | out_unlock: |
82b0f8c3 | 1559 | spin_unlock(vmf->ptl); |
b8916634 MG |
1560 | |
1561 | out: | |
1562 | if (anon_vma) | |
1563 | page_unlock_anon_vma_read(anon_vma); | |
1564 | ||
98fa15f3 | 1565 | if (page_nid != NUMA_NO_NODE) |
82b0f8c3 | 1566 | task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, |
9a8b300f | 1567 | flags); |
8191acbd | 1568 | |
d10e63f2 MG |
1569 | return 0; |
1570 | } | |
1571 | ||
319904ad HY |
1572 | /* |
1573 | * Return true if we do MADV_FREE successfully on entire pmd page. | |
1574 | * Otherwise, return false. | |
1575 | */ | |
1576 | bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, | |
b8d3c4c3 | 1577 | pmd_t *pmd, unsigned long addr, unsigned long next) |
b8d3c4c3 MK |
1578 | { |
1579 | spinlock_t *ptl; | |
1580 | pmd_t orig_pmd; | |
1581 | struct page *page; | |
1582 | struct mm_struct *mm = tlb->mm; | |
319904ad | 1583 | bool ret = false; |
b8d3c4c3 | 1584 | |
ed6a7935 | 1585 | tlb_change_page_size(tlb, HPAGE_PMD_SIZE); |
07e32661 | 1586 | |
b6ec57f4 KS |
1587 | ptl = pmd_trans_huge_lock(pmd, vma); |
1588 | if (!ptl) | |
25eedabe | 1589 | goto out_unlocked; |
b8d3c4c3 MK |
1590 | |
1591 | orig_pmd = *pmd; | |
319904ad | 1592 | if (is_huge_zero_pmd(orig_pmd)) |
b8d3c4c3 | 1593 | goto out; |
b8d3c4c3 | 1594 | |
84c3fc4e ZY |
1595 | if (unlikely(!pmd_present(orig_pmd))) { |
1596 | VM_BUG_ON(thp_migration_supported() && | |
1597 | !is_pmd_migration_entry(orig_pmd)); | |
1598 | goto out; | |
1599 | } | |
1600 | ||
b8d3c4c3 MK |
1601 | page = pmd_page(orig_pmd); |
1602 | /* | |
1603 | * If other processes are mapping this page, we couldn't discard | |
1604 | * the page unless they all do MADV_FREE so let's skip the page. | |
1605 | */ | |
1606 | if (page_mapcount(page) != 1) | |
1607 | goto out; | |
1608 | ||
1609 | if (!trylock_page(page)) | |
1610 | goto out; | |
1611 | ||
1612 | /* | |
1613 | * If user want to discard part-pages of THP, split it so MADV_FREE | |
1614 | * will deactivate only them. | |
1615 | */ | |
1616 | if (next - addr != HPAGE_PMD_SIZE) { | |
1617 | get_page(page); | |
1618 | spin_unlock(ptl); | |
9818b8cd | 1619 | split_huge_page(page); |
b8d3c4c3 | 1620 | unlock_page(page); |
bbf29ffc | 1621 | put_page(page); |
b8d3c4c3 MK |
1622 | goto out_unlocked; |
1623 | } | |
1624 | ||
1625 | if (PageDirty(page)) | |
1626 | ClearPageDirty(page); | |
1627 | unlock_page(page); | |
1628 | ||
b8d3c4c3 | 1629 | if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { |
58ceeb6b | 1630 | pmdp_invalidate(vma, addr, pmd); |
b8d3c4c3 MK |
1631 | orig_pmd = pmd_mkold(orig_pmd); |
1632 | orig_pmd = pmd_mkclean(orig_pmd); | |
1633 | ||
1634 | set_pmd_at(mm, addr, pmd, orig_pmd); | |
1635 | tlb_remove_pmd_tlb_entry(tlb, pmd, addr); | |
1636 | } | |
802a3a92 SL |
1637 | |
1638 | mark_page_lazyfree(page); | |
319904ad | 1639 | ret = true; |
b8d3c4c3 MK |
1640 | out: |
1641 | spin_unlock(ptl); | |
1642 | out_unlocked: | |
1643 | return ret; | |
1644 | } | |
1645 | ||
953c66c2 AK |
1646 | static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd) |
1647 | { | |
1648 | pgtable_t pgtable; | |
1649 | ||
1650 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); | |
1651 | pte_free(mm, pgtable); | |
c4812909 | 1652 | mm_dec_nr_ptes(mm); |
953c66c2 AK |
1653 | } |
1654 | ||
71e3aac0 | 1655 | int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, |
f21760b1 | 1656 | pmd_t *pmd, unsigned long addr) |
71e3aac0 | 1657 | { |
da146769 | 1658 | pmd_t orig_pmd; |
bf929152 | 1659 | spinlock_t *ptl; |
71e3aac0 | 1660 | |
ed6a7935 | 1661 | tlb_change_page_size(tlb, HPAGE_PMD_SIZE); |
07e32661 | 1662 | |
b6ec57f4 KS |
1663 | ptl = __pmd_trans_huge_lock(pmd, vma); |
1664 | if (!ptl) | |
da146769 KS |
1665 | return 0; |
1666 | /* | |
1667 | * For architectures like ppc64 we look at deposited pgtable | |
1668 | * when calling pmdp_huge_get_and_clear. So do the | |
1669 | * pgtable_trans_huge_withdraw after finishing pmdp related | |
1670 | * operations. | |
1671 | */ | |
93a98695 AK |
1672 | orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd, |
1673 | tlb->fullmm); | |
da146769 | 1674 | tlb_remove_pmd_tlb_entry(tlb, pmd, addr); |
2484ca9b | 1675 | if (vma_is_special_huge(vma)) { |
3b6521f5 OH |
1676 | if (arch_needs_pgtable_deposit()) |
1677 | zap_deposited_table(tlb->mm, pmd); | |
da146769 KS |
1678 | spin_unlock(ptl); |
1679 | if (is_huge_zero_pmd(orig_pmd)) | |
c0f2e176 | 1680 | tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); |
da146769 | 1681 | } else if (is_huge_zero_pmd(orig_pmd)) { |
c14a6eb4 | 1682 | zap_deposited_table(tlb->mm, pmd); |
da146769 | 1683 | spin_unlock(ptl); |
c0f2e176 | 1684 | tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); |
da146769 | 1685 | } else { |
616b8371 ZY |
1686 | struct page *page = NULL; |
1687 | int flush_needed = 1; | |
1688 | ||
1689 | if (pmd_present(orig_pmd)) { | |
1690 | page = pmd_page(orig_pmd); | |
1691 | page_remove_rmap(page, true); | |
1692 | VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); | |
1693 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
1694 | } else if (thp_migration_supported()) { | |
1695 | swp_entry_t entry; | |
1696 | ||
1697 | VM_BUG_ON(!is_pmd_migration_entry(orig_pmd)); | |
1698 | entry = pmd_to_swp_entry(orig_pmd); | |
1699 | page = pfn_to_page(swp_offset(entry)); | |
1700 | flush_needed = 0; | |
1701 | } else | |
1702 | WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!"); | |
1703 | ||
b5072380 | 1704 | if (PageAnon(page)) { |
c14a6eb4 | 1705 | zap_deposited_table(tlb->mm, pmd); |
b5072380 KS |
1706 | add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); |
1707 | } else { | |
953c66c2 AK |
1708 | if (arch_needs_pgtable_deposit()) |
1709 | zap_deposited_table(tlb->mm, pmd); | |
fadae295 | 1710 | add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR); |
b5072380 | 1711 | } |
616b8371 | 1712 | |
da146769 | 1713 | spin_unlock(ptl); |
616b8371 ZY |
1714 | if (flush_needed) |
1715 | tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE); | |
025c5b24 | 1716 | } |
da146769 | 1717 | return 1; |
71e3aac0 AA |
1718 | } |
1719 | ||
1dd38b6c AK |
1720 | #ifndef pmd_move_must_withdraw |
1721 | static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl, | |
1722 | spinlock_t *old_pmd_ptl, | |
1723 | struct vm_area_struct *vma) | |
1724 | { | |
1725 | /* | |
1726 | * With split pmd lock we also need to move preallocated | |
1727 | * PTE page table if new_pmd is on different PMD page table. | |
1728 | * | |
1729 | * We also don't deposit and withdraw tables for file pages. | |
1730 | */ | |
1731 | return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma); | |
1732 | } | |
1733 | #endif | |
1734 | ||
ab6e3d09 NH |
1735 | static pmd_t move_soft_dirty_pmd(pmd_t pmd) |
1736 | { | |
1737 | #ifdef CONFIG_MEM_SOFT_DIRTY | |
1738 | if (unlikely(is_pmd_migration_entry(pmd))) | |
1739 | pmd = pmd_swp_mksoft_dirty(pmd); | |
1740 | else if (pmd_present(pmd)) | |
1741 | pmd = pmd_mksoft_dirty(pmd); | |
1742 | #endif | |
1743 | return pmd; | |
1744 | } | |
1745 | ||
bf8616d5 | 1746 | bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, |
b8aa9d9d | 1747 | unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd) |
37a1c49a | 1748 | { |
bf929152 | 1749 | spinlock_t *old_ptl, *new_ptl; |
37a1c49a | 1750 | pmd_t pmd; |
37a1c49a | 1751 | struct mm_struct *mm = vma->vm_mm; |
5d190420 | 1752 | bool force_flush = false; |
37a1c49a | 1753 | |
37a1c49a AA |
1754 | /* |
1755 | * The destination pmd shouldn't be established, free_pgtables() | |
1756 | * should have release it. | |
1757 | */ | |
1758 | if (WARN_ON(!pmd_none(*new_pmd))) { | |
1759 | VM_BUG_ON(pmd_trans_huge(*new_pmd)); | |
4b471e88 | 1760 | return false; |
37a1c49a AA |
1761 | } |
1762 | ||
bf929152 KS |
1763 | /* |
1764 | * We don't have to worry about the ordering of src and dst | |
c1e8d7c6 | 1765 | * ptlocks because exclusive mmap_lock prevents deadlock. |
bf929152 | 1766 | */ |
b6ec57f4 KS |
1767 | old_ptl = __pmd_trans_huge_lock(old_pmd, vma); |
1768 | if (old_ptl) { | |
bf929152 KS |
1769 | new_ptl = pmd_lockptr(mm, new_pmd); |
1770 | if (new_ptl != old_ptl) | |
1771 | spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); | |
8809aa2d | 1772 | pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); |
eb66ae03 | 1773 | if (pmd_present(pmd)) |
a2ce2666 | 1774 | force_flush = true; |
025c5b24 | 1775 | VM_BUG_ON(!pmd_none(*new_pmd)); |
3592806c | 1776 | |
1dd38b6c | 1777 | if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) { |
b3084f4d | 1778 | pgtable_t pgtable; |
3592806c KS |
1779 | pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); |
1780 | pgtable_trans_huge_deposit(mm, new_pmd, pgtable); | |
3592806c | 1781 | } |
ab6e3d09 NH |
1782 | pmd = move_soft_dirty_pmd(pmd); |
1783 | set_pmd_at(mm, new_addr, new_pmd, pmd); | |
5d190420 AL |
1784 | if (force_flush) |
1785 | flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE); | |
eb66ae03 LT |
1786 | if (new_ptl != old_ptl) |
1787 | spin_unlock(new_ptl); | |
bf929152 | 1788 | spin_unlock(old_ptl); |
4b471e88 | 1789 | return true; |
37a1c49a | 1790 | } |
4b471e88 | 1791 | return false; |
37a1c49a AA |
1792 | } |
1793 | ||
f123d74a MG |
1794 | /* |
1795 | * Returns | |
1796 | * - 0 if PMD could not be locked | |
1797 | * - 1 if PMD was locked but protections unchange and TLB flush unnecessary | |
1798 | * - HPAGE_PMD_NR is protections changed and TLB flush necessary | |
1799 | */ | |
cd7548ab | 1800 | int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, |
58705444 | 1801 | unsigned long addr, pgprot_t newprot, unsigned long cp_flags) |
cd7548ab JW |
1802 | { |
1803 | struct mm_struct *mm = vma->vm_mm; | |
bf929152 | 1804 | spinlock_t *ptl; |
0a85e51d KS |
1805 | pmd_t entry; |
1806 | bool preserve_write; | |
1807 | int ret; | |
58705444 | 1808 | bool prot_numa = cp_flags & MM_CP_PROT_NUMA; |
292924b2 PX |
1809 | bool uffd_wp = cp_flags & MM_CP_UFFD_WP; |
1810 | bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; | |
cd7548ab | 1811 | |
b6ec57f4 | 1812 | ptl = __pmd_trans_huge_lock(pmd, vma); |
0a85e51d KS |
1813 | if (!ptl) |
1814 | return 0; | |
e944fd67 | 1815 | |
0a85e51d KS |
1816 | preserve_write = prot_numa && pmd_write(*pmd); |
1817 | ret = 1; | |
e944fd67 | 1818 | |
84c3fc4e ZY |
1819 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
1820 | if (is_swap_pmd(*pmd)) { | |
1821 | swp_entry_t entry = pmd_to_swp_entry(*pmd); | |
1822 | ||
1823 | VM_BUG_ON(!is_pmd_migration_entry(*pmd)); | |
1824 | if (is_write_migration_entry(entry)) { | |
1825 | pmd_t newpmd; | |
1826 | /* | |
1827 | * A protection check is difficult so | |
1828 | * just be safe and disable write | |
1829 | */ | |
1830 | make_migration_entry_read(&entry); | |
1831 | newpmd = swp_entry_to_pmd(entry); | |
ab6e3d09 NH |
1832 | if (pmd_swp_soft_dirty(*pmd)) |
1833 | newpmd = pmd_swp_mksoft_dirty(newpmd); | |
84c3fc4e ZY |
1834 | set_pmd_at(mm, addr, pmd, newpmd); |
1835 | } | |
1836 | goto unlock; | |
1837 | } | |
1838 | #endif | |
1839 | ||
0a85e51d KS |
1840 | /* |
1841 | * Avoid trapping faults against the zero page. The read-only | |
1842 | * data is likely to be read-cached on the local CPU and | |
1843 | * local/remote hits to the zero page are not interesting. | |
1844 | */ | |
1845 | if (prot_numa && is_huge_zero_pmd(*pmd)) | |
1846 | goto unlock; | |
025c5b24 | 1847 | |
0a85e51d KS |
1848 | if (prot_numa && pmd_protnone(*pmd)) |
1849 | goto unlock; | |
1850 | ||
ced10803 | 1851 | /* |
3e4e28c5 | 1852 | * In case prot_numa, we are under mmap_read_lock(mm). It's critical |
ced10803 | 1853 | * to not clear pmd intermittently to avoid race with MADV_DONTNEED |
3e4e28c5 | 1854 | * which is also under mmap_read_lock(mm): |
ced10803 KS |
1855 | * |
1856 | * CPU0: CPU1: | |
1857 | * change_huge_pmd(prot_numa=1) | |
1858 | * pmdp_huge_get_and_clear_notify() | |
1859 | * madvise_dontneed() | |
1860 | * zap_pmd_range() | |
1861 | * pmd_trans_huge(*pmd) == 0 (without ptl) | |
1862 | * // skip the pmd | |
1863 | * set_pmd_at(); | |
1864 | * // pmd is re-established | |
1865 | * | |
1866 | * The race makes MADV_DONTNEED miss the huge pmd and don't clear it | |
1867 | * which may break userspace. | |
1868 | * | |
1869 | * pmdp_invalidate() is required to make sure we don't miss | |
1870 | * dirty/young flags set by hardware. | |
1871 | */ | |
a3cf988f | 1872 | entry = pmdp_invalidate(vma, addr, pmd); |
ced10803 | 1873 | |
0a85e51d KS |
1874 | entry = pmd_modify(entry, newprot); |
1875 | if (preserve_write) | |
1876 | entry = pmd_mk_savedwrite(entry); | |
292924b2 PX |
1877 | if (uffd_wp) { |
1878 | entry = pmd_wrprotect(entry); | |
1879 | entry = pmd_mkuffd_wp(entry); | |
1880 | } else if (uffd_wp_resolve) { | |
1881 | /* | |
1882 | * Leave the write bit to be handled by PF interrupt | |
1883 | * handler, then things like COW could be properly | |
1884 | * handled. | |
1885 | */ | |
1886 | entry = pmd_clear_uffd_wp(entry); | |
1887 | } | |
0a85e51d KS |
1888 | ret = HPAGE_PMD_NR; |
1889 | set_pmd_at(mm, addr, pmd, entry); | |
1890 | BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry)); | |
1891 | unlock: | |
1892 | spin_unlock(ptl); | |
025c5b24 NH |
1893 | return ret; |
1894 | } | |
1895 | ||
1896 | /* | |
8f19b0c0 | 1897 | * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise. |
025c5b24 | 1898 | * |
8f19b0c0 HY |
1899 | * Note that if it returns page table lock pointer, this routine returns without |
1900 | * unlocking page table lock. So callers must unlock it. | |
025c5b24 | 1901 | */ |
b6ec57f4 | 1902 | spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) |
025c5b24 | 1903 | { |
b6ec57f4 KS |
1904 | spinlock_t *ptl; |
1905 | ptl = pmd_lock(vma->vm_mm, pmd); | |
84c3fc4e ZY |
1906 | if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || |
1907 | pmd_devmap(*pmd))) | |
b6ec57f4 KS |
1908 | return ptl; |
1909 | spin_unlock(ptl); | |
1910 | return NULL; | |
cd7548ab JW |
1911 | } |
1912 | ||
a00cc7d9 MW |
1913 | /* |
1914 | * Returns true if a given pud maps a thp, false otherwise. | |
1915 | * | |
1916 | * Note that if it returns true, this routine returns without unlocking page | |
1917 | * table lock. So callers must unlock it. | |
1918 | */ | |
1919 | spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma) | |
1920 | { | |
1921 | spinlock_t *ptl; | |
1922 | ||
1923 | ptl = pud_lock(vma->vm_mm, pud); | |
1924 | if (likely(pud_trans_huge(*pud) || pud_devmap(*pud))) | |
1925 | return ptl; | |
1926 | spin_unlock(ptl); | |
1927 | return NULL; | |
1928 | } | |
1929 | ||
1930 | #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD | |
1931 | int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, | |
1932 | pud_t *pud, unsigned long addr) | |
1933 | { | |
a00cc7d9 MW |
1934 | spinlock_t *ptl; |
1935 | ||
1936 | ptl = __pud_trans_huge_lock(pud, vma); | |
1937 | if (!ptl) | |
1938 | return 0; | |
1939 | /* | |
1940 | * For architectures like ppc64 we look at deposited pgtable | |
1941 | * when calling pudp_huge_get_and_clear. So do the | |
1942 | * pgtable_trans_huge_withdraw after finishing pudp related | |
1943 | * operations. | |
1944 | */ | |
70516b93 | 1945 | pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm); |
a00cc7d9 | 1946 | tlb_remove_pud_tlb_entry(tlb, pud, addr); |
2484ca9b | 1947 | if (vma_is_special_huge(vma)) { |
a00cc7d9 MW |
1948 | spin_unlock(ptl); |
1949 | /* No zero page support yet */ | |
1950 | } else { | |
1951 | /* No support for anonymous PUD pages yet */ | |
1952 | BUG(); | |
1953 | } | |
1954 | return 1; | |
1955 | } | |
1956 | ||
1957 | static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud, | |
1958 | unsigned long haddr) | |
1959 | { | |
1960 | VM_BUG_ON(haddr & ~HPAGE_PUD_MASK); | |
1961 | VM_BUG_ON_VMA(vma->vm_start > haddr, vma); | |
1962 | VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma); | |
1963 | VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud)); | |
1964 | ||
ce9311cf | 1965 | count_vm_event(THP_SPLIT_PUD); |
a00cc7d9 MW |
1966 | |
1967 | pudp_huge_clear_flush_notify(vma, haddr, pud); | |
1968 | } | |
1969 | ||
1970 | void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud, | |
1971 | unsigned long address) | |
1972 | { | |
1973 | spinlock_t *ptl; | |
ac46d4f3 | 1974 | struct mmu_notifier_range range; |
a00cc7d9 | 1975 | |
7269f999 | 1976 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, |
6f4f13e8 | 1977 | address & HPAGE_PUD_MASK, |
ac46d4f3 JG |
1978 | (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE); |
1979 | mmu_notifier_invalidate_range_start(&range); | |
1980 | ptl = pud_lock(vma->vm_mm, pud); | |
a00cc7d9 MW |
1981 | if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud))) |
1982 | goto out; | |
ac46d4f3 | 1983 | __split_huge_pud_locked(vma, pud, range.start); |
a00cc7d9 MW |
1984 | |
1985 | out: | |
1986 | spin_unlock(ptl); | |
4645b9fe JG |
1987 | /* |
1988 | * No need to double call mmu_notifier->invalidate_range() callback as | |
1989 | * the above pudp_huge_clear_flush_notify() did already call it. | |
1990 | */ | |
ac46d4f3 | 1991 | mmu_notifier_invalidate_range_only_end(&range); |
a00cc7d9 MW |
1992 | } |
1993 | #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ | |
1994 | ||
eef1b3ba KS |
1995 | static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, |
1996 | unsigned long haddr, pmd_t *pmd) | |
1997 | { | |
1998 | struct mm_struct *mm = vma->vm_mm; | |
1999 | pgtable_t pgtable; | |
2000 | pmd_t _pmd; | |
2001 | int i; | |
2002 | ||
0f10851e JG |
2003 | /* |
2004 | * Leave pmd empty until pte is filled note that it is fine to delay | |
2005 | * notification until mmu_notifier_invalidate_range_end() as we are | |
2006 | * replacing a zero pmd write protected page with a zero pte write | |
2007 | * protected page. | |
2008 | * | |
ad56b738 | 2009 | * See Documentation/vm/mmu_notifier.rst |
0f10851e JG |
2010 | */ |
2011 | pmdp_huge_clear_flush(vma, haddr, pmd); | |
eef1b3ba KS |
2012 | |
2013 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); | |
2014 | pmd_populate(mm, &_pmd, pgtable); | |
2015 | ||
2016 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
2017 | pte_t *pte, entry; | |
2018 | entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); | |
2019 | entry = pte_mkspecial(entry); | |
2020 | pte = pte_offset_map(&_pmd, haddr); | |
2021 | VM_BUG_ON(!pte_none(*pte)); | |
2022 | set_pte_at(mm, haddr, pte, entry); | |
2023 | pte_unmap(pte); | |
2024 | } | |
2025 | smp_wmb(); /* make pte visible before pmd */ | |
2026 | pmd_populate(mm, pmd, pgtable); | |
eef1b3ba KS |
2027 | } |
2028 | ||
2029 | static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, | |
ba988280 | 2030 | unsigned long haddr, bool freeze) |
eef1b3ba KS |
2031 | { |
2032 | struct mm_struct *mm = vma->vm_mm; | |
2033 | struct page *page; | |
2034 | pgtable_t pgtable; | |
423ac9af | 2035 | pmd_t old_pmd, _pmd; |
292924b2 | 2036 | bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false; |
2ac015e2 | 2037 | unsigned long addr; |
eef1b3ba KS |
2038 | int i; |
2039 | ||
2040 | VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); | |
2041 | VM_BUG_ON_VMA(vma->vm_start > haddr, vma); | |
2042 | VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); | |
84c3fc4e ZY |
2043 | VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd) |
2044 | && !pmd_devmap(*pmd)); | |
eef1b3ba KS |
2045 | |
2046 | count_vm_event(THP_SPLIT_PMD); | |
2047 | ||
d21b9e57 KS |
2048 | if (!vma_is_anonymous(vma)) { |
2049 | _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); | |
953c66c2 AK |
2050 | /* |
2051 | * We are going to unmap this huge page. So | |
2052 | * just go ahead and zap it | |
2053 | */ | |
2054 | if (arch_needs_pgtable_deposit()) | |
2055 | zap_deposited_table(mm, pmd); | |
2484ca9b | 2056 | if (vma_is_special_huge(vma)) |
d21b9e57 KS |
2057 | return; |
2058 | page = pmd_page(_pmd); | |
e1f1b157 HD |
2059 | if (!PageDirty(page) && pmd_dirty(_pmd)) |
2060 | set_page_dirty(page); | |
d21b9e57 KS |
2061 | if (!PageReferenced(page) && pmd_young(_pmd)) |
2062 | SetPageReferenced(page); | |
2063 | page_remove_rmap(page, true); | |
2064 | put_page(page); | |
fadae295 | 2065 | add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR); |
eef1b3ba | 2066 | return; |
ec0abae6 | 2067 | } else if (pmd_trans_huge(*pmd) && is_huge_zero_pmd(*pmd)) { |
4645b9fe JG |
2068 | /* |
2069 | * FIXME: Do we want to invalidate secondary mmu by calling | |
2070 | * mmu_notifier_invalidate_range() see comments below inside | |
2071 | * __split_huge_pmd() ? | |
2072 | * | |
2073 | * We are going from a zero huge page write protected to zero | |
2074 | * small page also write protected so it does not seems useful | |
2075 | * to invalidate secondary mmu at this time. | |
2076 | */ | |
eef1b3ba KS |
2077 | return __split_huge_zero_page_pmd(vma, haddr, pmd); |
2078 | } | |
2079 | ||
423ac9af AK |
2080 | /* |
2081 | * Up to this point the pmd is present and huge and userland has the | |
2082 | * whole access to the hugepage during the split (which happens in | |
2083 | * place). If we overwrite the pmd with the not-huge version pointing | |
2084 | * to the pte here (which of course we could if all CPUs were bug | |
2085 | * free), userland could trigger a small page size TLB miss on the | |
2086 | * small sized TLB while the hugepage TLB entry is still established in | |
2087 | * the huge TLB. Some CPU doesn't like that. | |
42742d9b AK |
2088 | * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum |
2089 | * 383 on page 105. Intel should be safe but is also warns that it's | |
423ac9af AK |
2090 | * only safe if the permission and cache attributes of the two entries |
2091 | * loaded in the two TLB is identical (which should be the case here). | |
2092 | * But it is generally safer to never allow small and huge TLB entries | |
2093 | * for the same virtual address to be loaded simultaneously. So instead | |
2094 | * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the | |
2095 | * current pmd notpresent (atomically because here the pmd_trans_huge | |
2096 | * must remain set at all times on the pmd until the split is complete | |
2097 | * for this pmd), then we flush the SMP TLB and finally we write the | |
2098 | * non-huge version of the pmd entry with pmd_populate. | |
2099 | */ | |
2100 | old_pmd = pmdp_invalidate(vma, haddr, pmd); | |
2101 | ||
423ac9af | 2102 | pmd_migration = is_pmd_migration_entry(old_pmd); |
2e83ee1d | 2103 | if (unlikely(pmd_migration)) { |
84c3fc4e ZY |
2104 | swp_entry_t entry; |
2105 | ||
423ac9af | 2106 | entry = pmd_to_swp_entry(old_pmd); |
84c3fc4e | 2107 | page = pfn_to_page(swp_offset(entry)); |
2e83ee1d PX |
2108 | write = is_write_migration_entry(entry); |
2109 | young = false; | |
2110 | soft_dirty = pmd_swp_soft_dirty(old_pmd); | |
f45ec5ff | 2111 | uffd_wp = pmd_swp_uffd_wp(old_pmd); |
2e83ee1d | 2112 | } else { |
423ac9af | 2113 | page = pmd_page(old_pmd); |
2e83ee1d PX |
2114 | if (pmd_dirty(old_pmd)) |
2115 | SetPageDirty(page); | |
2116 | write = pmd_write(old_pmd); | |
2117 | young = pmd_young(old_pmd); | |
2118 | soft_dirty = pmd_soft_dirty(old_pmd); | |
292924b2 | 2119 | uffd_wp = pmd_uffd_wp(old_pmd); |
2e83ee1d | 2120 | } |
eef1b3ba | 2121 | VM_BUG_ON_PAGE(!page_count(page), page); |
fe896d18 | 2122 | page_ref_add(page, HPAGE_PMD_NR - 1); |
eef1b3ba | 2123 | |
423ac9af AK |
2124 | /* |
2125 | * Withdraw the table only after we mark the pmd entry invalid. | |
2126 | * This's critical for some architectures (Power). | |
2127 | */ | |
eef1b3ba KS |
2128 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); |
2129 | pmd_populate(mm, &_pmd, pgtable); | |
2130 | ||
2ac015e2 | 2131 | for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { |
eef1b3ba KS |
2132 | pte_t entry, *pte; |
2133 | /* | |
2134 | * Note that NUMA hinting access restrictions are not | |
2135 | * transferred to avoid any possibility of altering | |
2136 | * permissions across VMAs. | |
2137 | */ | |
84c3fc4e | 2138 | if (freeze || pmd_migration) { |
ba988280 KS |
2139 | swp_entry_t swp_entry; |
2140 | swp_entry = make_migration_entry(page + i, write); | |
2141 | entry = swp_entry_to_pte(swp_entry); | |
804dd150 AA |
2142 | if (soft_dirty) |
2143 | entry = pte_swp_mksoft_dirty(entry); | |
f45ec5ff PX |
2144 | if (uffd_wp) |
2145 | entry = pte_swp_mkuffd_wp(entry); | |
ba988280 | 2146 | } else { |
6d2329f8 | 2147 | entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot)); |
b8d3c4c3 | 2148 | entry = maybe_mkwrite(entry, vma); |
ba988280 KS |
2149 | if (!write) |
2150 | entry = pte_wrprotect(entry); | |
2151 | if (!young) | |
2152 | entry = pte_mkold(entry); | |
804dd150 AA |
2153 | if (soft_dirty) |
2154 | entry = pte_mksoft_dirty(entry); | |
292924b2 PX |
2155 | if (uffd_wp) |
2156 | entry = pte_mkuffd_wp(entry); | |
ba988280 | 2157 | } |
2ac015e2 | 2158 | pte = pte_offset_map(&_pmd, addr); |
eef1b3ba | 2159 | BUG_ON(!pte_none(*pte)); |
2ac015e2 | 2160 | set_pte_at(mm, addr, pte, entry); |
ec0abae6 | 2161 | if (!pmd_migration) |
eef1b3ba | 2162 | atomic_inc(&page[i]._mapcount); |
ec0abae6 | 2163 | pte_unmap(pte); |
eef1b3ba KS |
2164 | } |
2165 | ||
ec0abae6 RC |
2166 | if (!pmd_migration) { |
2167 | /* | |
2168 | * Set PG_double_map before dropping compound_mapcount to avoid | |
2169 | * false-negative page_mapped(). | |
2170 | */ | |
2171 | if (compound_mapcount(page) > 1 && | |
2172 | !TestSetPageDoubleMap(page)) { | |
eef1b3ba | 2173 | for (i = 0; i < HPAGE_PMD_NR; i++) |
ec0abae6 RC |
2174 | atomic_inc(&page[i]._mapcount); |
2175 | } | |
2176 | ||
2177 | lock_page_memcg(page); | |
2178 | if (atomic_add_negative(-1, compound_mapcount_ptr(page))) { | |
2179 | /* Last compound_mapcount is gone. */ | |
69473e5d MS |
2180 | __mod_lruvec_page_state(page, NR_ANON_THPS, |
2181 | -HPAGE_PMD_NR); | |
ec0abae6 RC |
2182 | if (TestClearPageDoubleMap(page)) { |
2183 | /* No need in mapcount reference anymore */ | |
2184 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
2185 | atomic_dec(&page[i]._mapcount); | |
2186 | } | |
eef1b3ba | 2187 | } |
ec0abae6 | 2188 | unlock_page_memcg(page); |
eef1b3ba KS |
2189 | } |
2190 | ||
2191 | smp_wmb(); /* make pte visible before pmd */ | |
2192 | pmd_populate(mm, pmd, pgtable); | |
e9b61f19 KS |
2193 | |
2194 | if (freeze) { | |
2ac015e2 | 2195 | for (i = 0; i < HPAGE_PMD_NR; i++) { |
e9b61f19 KS |
2196 | page_remove_rmap(page + i, false); |
2197 | put_page(page + i); | |
2198 | } | |
2199 | } | |
eef1b3ba KS |
2200 | } |
2201 | ||
2202 | void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, | |
33f4751e | 2203 | unsigned long address, bool freeze, struct page *page) |
eef1b3ba KS |
2204 | { |
2205 | spinlock_t *ptl; | |
ac46d4f3 | 2206 | struct mmu_notifier_range range; |
1c2f6730 | 2207 | bool do_unlock_page = false; |
c444eb56 | 2208 | pmd_t _pmd; |
eef1b3ba | 2209 | |
7269f999 | 2210 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, |
6f4f13e8 | 2211 | address & HPAGE_PMD_MASK, |
ac46d4f3 JG |
2212 | (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE); |
2213 | mmu_notifier_invalidate_range_start(&range); | |
2214 | ptl = pmd_lock(vma->vm_mm, pmd); | |
33f4751e NH |
2215 | |
2216 | /* | |
2217 | * If caller asks to setup a migration entries, we need a page to check | |
2218 | * pmd against. Otherwise we can end up replacing wrong page. | |
2219 | */ | |
2220 | VM_BUG_ON(freeze && !page); | |
c444eb56 AA |
2221 | if (page) { |
2222 | VM_WARN_ON_ONCE(!PageLocked(page)); | |
c444eb56 AA |
2223 | if (page != pmd_page(*pmd)) |
2224 | goto out; | |
2225 | } | |
33f4751e | 2226 | |
c444eb56 | 2227 | repeat: |
5c7fb56e | 2228 | if (pmd_trans_huge(*pmd)) { |
c444eb56 AA |
2229 | if (!page) { |
2230 | page = pmd_page(*pmd); | |
1c2f6730 HD |
2231 | /* |
2232 | * An anonymous page must be locked, to ensure that a | |
2233 | * concurrent reuse_swap_page() sees stable mapcount; | |
2234 | * but reuse_swap_page() is not used on shmem or file, | |
2235 | * and page lock must not be taken when zap_pmd_range() | |
2236 | * calls __split_huge_pmd() while i_mmap_lock is held. | |
2237 | */ | |
2238 | if (PageAnon(page)) { | |
2239 | if (unlikely(!trylock_page(page))) { | |
2240 | get_page(page); | |
2241 | _pmd = *pmd; | |
2242 | spin_unlock(ptl); | |
2243 | lock_page(page); | |
2244 | spin_lock(ptl); | |
2245 | if (unlikely(!pmd_same(*pmd, _pmd))) { | |
2246 | unlock_page(page); | |
2247 | put_page(page); | |
2248 | page = NULL; | |
2249 | goto repeat; | |
2250 | } | |
c444eb56 | 2251 | put_page(page); |
c444eb56 | 2252 | } |
1c2f6730 | 2253 | do_unlock_page = true; |
c444eb56 AA |
2254 | } |
2255 | } | |
5c7fb56e | 2256 | if (PageMlocked(page)) |
5f737714 | 2257 | clear_page_mlock(page); |
84c3fc4e | 2258 | } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd))) |
e90309c9 | 2259 | goto out; |
ac46d4f3 | 2260 | __split_huge_pmd_locked(vma, pmd, range.start, freeze); |
e90309c9 | 2261 | out: |
eef1b3ba | 2262 | spin_unlock(ptl); |
1c2f6730 | 2263 | if (do_unlock_page) |
c444eb56 | 2264 | unlock_page(page); |
4645b9fe JG |
2265 | /* |
2266 | * No need to double call mmu_notifier->invalidate_range() callback. | |
2267 | * They are 3 cases to consider inside __split_huge_pmd_locked(): | |
2268 | * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious | |
2269 | * 2) __split_huge_zero_page_pmd() read only zero page and any write | |
2270 | * fault will trigger a flush_notify before pointing to a new page | |
2271 | * (it is fine if the secondary mmu keeps pointing to the old zero | |
2272 | * page in the meantime) | |
2273 | * 3) Split a huge pmd into pte pointing to the same page. No need | |
2274 | * to invalidate secondary tlb entry they are all still valid. | |
2275 | * any further changes to individual pte will notify. So no need | |
2276 | * to call mmu_notifier->invalidate_range() | |
2277 | */ | |
ac46d4f3 | 2278 | mmu_notifier_invalidate_range_only_end(&range); |
eef1b3ba KS |
2279 | } |
2280 | ||
fec89c10 KS |
2281 | void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address, |
2282 | bool freeze, struct page *page) | |
94fcc585 | 2283 | { |
f72e7dcd | 2284 | pgd_t *pgd; |
c2febafc | 2285 | p4d_t *p4d; |
f72e7dcd | 2286 | pud_t *pud; |
94fcc585 AA |
2287 | pmd_t *pmd; |
2288 | ||
78ddc534 | 2289 | pgd = pgd_offset(vma->vm_mm, address); |
f72e7dcd HD |
2290 | if (!pgd_present(*pgd)) |
2291 | return; | |
2292 | ||
c2febafc KS |
2293 | p4d = p4d_offset(pgd, address); |
2294 | if (!p4d_present(*p4d)) | |
2295 | return; | |
2296 | ||
2297 | pud = pud_offset(p4d, address); | |
f72e7dcd HD |
2298 | if (!pud_present(*pud)) |
2299 | return; | |
2300 | ||
2301 | pmd = pmd_offset(pud, address); | |
fec89c10 | 2302 | |
33f4751e | 2303 | __split_huge_pmd(vma, pmd, address, freeze, page); |
94fcc585 AA |
2304 | } |
2305 | ||
e1b9996b | 2306 | void vma_adjust_trans_huge(struct vm_area_struct *vma, |
94fcc585 AA |
2307 | unsigned long start, |
2308 | unsigned long end, | |
2309 | long adjust_next) | |
2310 | { | |
2311 | /* | |
2312 | * If the new start address isn't hpage aligned and it could | |
2313 | * previously contain an hugepage: check if we need to split | |
2314 | * an huge pmd. | |
2315 | */ | |
2316 | if (start & ~HPAGE_PMD_MASK && | |
2317 | (start & HPAGE_PMD_MASK) >= vma->vm_start && | |
2318 | (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
fec89c10 | 2319 | split_huge_pmd_address(vma, start, false, NULL); |
94fcc585 AA |
2320 | |
2321 | /* | |
2322 | * If the new end address isn't hpage aligned and it could | |
2323 | * previously contain an hugepage: check if we need to split | |
2324 | * an huge pmd. | |
2325 | */ | |
2326 | if (end & ~HPAGE_PMD_MASK && | |
2327 | (end & HPAGE_PMD_MASK) >= vma->vm_start && | |
2328 | (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
fec89c10 | 2329 | split_huge_pmd_address(vma, end, false, NULL); |
94fcc585 AA |
2330 | |
2331 | /* | |
2332 | * If we're also updating the vma->vm_next->vm_start, if the new | |
f9d86a60 | 2333 | * vm_next->vm_start isn't hpage aligned and it could previously |
94fcc585 AA |
2334 | * contain an hugepage: check if we need to split an huge pmd. |
2335 | */ | |
2336 | if (adjust_next > 0) { | |
2337 | struct vm_area_struct *next = vma->vm_next; | |
2338 | unsigned long nstart = next->vm_start; | |
f9d86a60 | 2339 | nstart += adjust_next; |
94fcc585 AA |
2340 | if (nstart & ~HPAGE_PMD_MASK && |
2341 | (nstart & HPAGE_PMD_MASK) >= next->vm_start && | |
2342 | (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) | |
fec89c10 | 2343 | split_huge_pmd_address(next, nstart, false, NULL); |
94fcc585 AA |
2344 | } |
2345 | } | |
e9b61f19 | 2346 | |
906f9cdf | 2347 | static void unmap_page(struct page *page) |
e9b61f19 | 2348 | { |
013339df | 2349 | enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | |
c7ab0d2f | 2350 | TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD; |
666e5a40 | 2351 | bool unmap_success; |
e9b61f19 KS |
2352 | |
2353 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
2354 | ||
baa355fd | 2355 | if (PageAnon(page)) |
b5ff8161 | 2356 | ttu_flags |= TTU_SPLIT_FREEZE; |
baa355fd | 2357 | |
666e5a40 MK |
2358 | unmap_success = try_to_unmap(page, ttu_flags); |
2359 | VM_BUG_ON_PAGE(!unmap_success, page); | |
e9b61f19 KS |
2360 | } |
2361 | ||
8cce5475 | 2362 | static void remap_page(struct page *page, unsigned int nr) |
e9b61f19 | 2363 | { |
fec89c10 | 2364 | int i; |
ace71a19 KS |
2365 | if (PageTransHuge(page)) { |
2366 | remove_migration_ptes(page, page, true); | |
2367 | } else { | |
8cce5475 | 2368 | for (i = 0; i < nr; i++) |
ace71a19 KS |
2369 | remove_migration_ptes(page + i, page + i, true); |
2370 | } | |
e9b61f19 KS |
2371 | } |
2372 | ||
94866635 | 2373 | static void lru_add_page_tail(struct page *head, struct page *tail, |
88dcb9a3 AS |
2374 | struct lruvec *lruvec, struct list_head *list) |
2375 | { | |
94866635 AS |
2376 | VM_BUG_ON_PAGE(!PageHead(head), head); |
2377 | VM_BUG_ON_PAGE(PageCompound(tail), head); | |
2378 | VM_BUG_ON_PAGE(PageLRU(tail), head); | |
6168d0da | 2379 | lockdep_assert_held(&lruvec->lru_lock); |
88dcb9a3 | 2380 | |
6dbb5741 | 2381 | if (list) { |
88dcb9a3 | 2382 | /* page reclaim is reclaiming a huge page */ |
6dbb5741 | 2383 | VM_WARN_ON(PageLRU(head)); |
94866635 AS |
2384 | get_page(tail); |
2385 | list_add_tail(&tail->lru, list); | |
88dcb9a3 | 2386 | } else { |
6dbb5741 AS |
2387 | /* head is still on lru (and we have it frozen) */ |
2388 | VM_WARN_ON(!PageLRU(head)); | |
2389 | SetPageLRU(tail); | |
2390 | list_add_tail(&tail->lru, &head->lru); | |
88dcb9a3 AS |
2391 | } |
2392 | } | |
2393 | ||
8df651c7 | 2394 | static void __split_huge_page_tail(struct page *head, int tail, |
e9b61f19 KS |
2395 | struct lruvec *lruvec, struct list_head *list) |
2396 | { | |
e9b61f19 KS |
2397 | struct page *page_tail = head + tail; |
2398 | ||
8df651c7 | 2399 | VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail); |
e9b61f19 KS |
2400 | |
2401 | /* | |
605ca5ed KK |
2402 | * Clone page flags before unfreezing refcount. |
2403 | * | |
2404 | * After successful get_page_unless_zero() might follow flags change, | |
8958b249 | 2405 | * for example lock_page() which set PG_waiters. |
e9b61f19 | 2406 | */ |
e9b61f19 KS |
2407 | page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; |
2408 | page_tail->flags |= (head->flags & | |
2409 | ((1L << PG_referenced) | | |
2410 | (1L << PG_swapbacked) | | |
38d8b4e6 | 2411 | (1L << PG_swapcache) | |
e9b61f19 KS |
2412 | (1L << PG_mlocked) | |
2413 | (1L << PG_uptodate) | | |
2414 | (1L << PG_active) | | |
1899ad18 | 2415 | (1L << PG_workingset) | |
e9b61f19 | 2416 | (1L << PG_locked) | |
b8d3c4c3 | 2417 | (1L << PG_unevictable) | |
72e6afa0 CM |
2418 | #ifdef CONFIG_64BIT |
2419 | (1L << PG_arch_2) | | |
2420 | #endif | |
b8d3c4c3 | 2421 | (1L << PG_dirty))); |
e9b61f19 | 2422 | |
173d9d9f HD |
2423 | /* ->mapping in first tail page is compound_mapcount */ |
2424 | VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING, | |
2425 | page_tail); | |
2426 | page_tail->mapping = head->mapping; | |
2427 | page_tail->index = head->index + tail; | |
2428 | ||
605ca5ed | 2429 | /* Page flags must be visible before we make the page non-compound. */ |
e9b61f19 KS |
2430 | smp_wmb(); |
2431 | ||
605ca5ed KK |
2432 | /* |
2433 | * Clear PageTail before unfreezing page refcount. | |
2434 | * | |
2435 | * After successful get_page_unless_zero() might follow put_page() | |
2436 | * which needs correct compound_head(). | |
2437 | */ | |
e9b61f19 KS |
2438 | clear_compound_head(page_tail); |
2439 | ||
605ca5ed KK |
2440 | /* Finally unfreeze refcount. Additional reference from page cache. */ |
2441 | page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) || | |
2442 | PageSwapCache(head))); | |
2443 | ||
e9b61f19 KS |
2444 | if (page_is_young(head)) |
2445 | set_page_young(page_tail); | |
2446 | if (page_is_idle(head)) | |
2447 | set_page_idle(page_tail); | |
2448 | ||
e9b61f19 | 2449 | page_cpupid_xchg_last(page_tail, page_cpupid_last(head)); |
94723aaf MH |
2450 | |
2451 | /* | |
2452 | * always add to the tail because some iterators expect new | |
2453 | * pages to show after the currently processed elements - e.g. | |
2454 | * migrate_pages | |
2455 | */ | |
e9b61f19 | 2456 | lru_add_page_tail(head, page_tail, lruvec, list); |
e9b61f19 KS |
2457 | } |
2458 | ||
baa355fd | 2459 | static void __split_huge_page(struct page *page, struct list_head *list, |
b6769834 | 2460 | pgoff_t end) |
e9b61f19 KS |
2461 | { |
2462 | struct page *head = compound_head(page); | |
e9b61f19 | 2463 | struct lruvec *lruvec; |
4101196b MWO |
2464 | struct address_space *swap_cache = NULL; |
2465 | unsigned long offset = 0; | |
8cce5475 | 2466 | unsigned int nr = thp_nr_pages(head); |
8df651c7 | 2467 | int i; |
e9b61f19 | 2468 | |
e9b61f19 KS |
2469 | /* complete memcg works before add pages to LRU */ |
2470 | mem_cgroup_split_huge_fixup(head); | |
2471 | ||
4101196b MWO |
2472 | if (PageAnon(head) && PageSwapCache(head)) { |
2473 | swp_entry_t entry = { .val = page_private(head) }; | |
2474 | ||
2475 | offset = swp_offset(entry); | |
2476 | swap_cache = swap_address_space(entry); | |
2477 | xa_lock(&swap_cache->i_pages); | |
2478 | } | |
2479 | ||
6168d0da AS |
2480 | /* lock lru list/PageCompound, ref freezed by page_ref_freeze */ |
2481 | lruvec = lock_page_lruvec(head); | |
b6769834 | 2482 | |
8cce5475 | 2483 | for (i = nr - 1; i >= 1; i--) { |
8df651c7 | 2484 | __split_huge_page_tail(head, i, lruvec, list); |
baa355fd KS |
2485 | /* Some pages can be beyond i_size: drop them from page cache */ |
2486 | if (head[i].index >= end) { | |
2d077d4b | 2487 | ClearPageDirty(head + i); |
baa355fd | 2488 | __delete_from_page_cache(head + i, NULL); |
800d8c63 KS |
2489 | if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head)) |
2490 | shmem_uncharge(head->mapping->host, 1); | |
baa355fd | 2491 | put_page(head + i); |
4101196b MWO |
2492 | } else if (!PageAnon(page)) { |
2493 | __xa_store(&head->mapping->i_pages, head[i].index, | |
2494 | head + i, 0); | |
2495 | } else if (swap_cache) { | |
2496 | __xa_store(&swap_cache->i_pages, offset + i, | |
2497 | head + i, 0); | |
baa355fd KS |
2498 | } |
2499 | } | |
e9b61f19 KS |
2500 | |
2501 | ClearPageCompound(head); | |
6168d0da | 2502 | unlock_page_lruvec(lruvec); |
b6769834 | 2503 | /* Caller disabled irqs, so they are still disabled here */ |
f7da677b | 2504 | |
8cce5475 | 2505 | split_page_owner(head, nr); |
f7da677b | 2506 | |
baa355fd KS |
2507 | /* See comment in __split_huge_page_tail() */ |
2508 | if (PageAnon(head)) { | |
aa5dc07f | 2509 | /* Additional pin to swap cache */ |
4101196b | 2510 | if (PageSwapCache(head)) { |
38d8b4e6 | 2511 | page_ref_add(head, 2); |
4101196b MWO |
2512 | xa_unlock(&swap_cache->i_pages); |
2513 | } else { | |
38d8b4e6 | 2514 | page_ref_inc(head); |
4101196b | 2515 | } |
baa355fd | 2516 | } else { |
aa5dc07f | 2517 | /* Additional pin to page cache */ |
baa355fd | 2518 | page_ref_add(head, 2); |
b93b0163 | 2519 | xa_unlock(&head->mapping->i_pages); |
baa355fd | 2520 | } |
b6769834 | 2521 | local_irq_enable(); |
e9b61f19 | 2522 | |
8cce5475 | 2523 | remap_page(head, nr); |
e9b61f19 | 2524 | |
c4f9c701 HY |
2525 | if (PageSwapCache(head)) { |
2526 | swp_entry_t entry = { .val = page_private(head) }; | |
2527 | ||
2528 | split_swap_cluster(entry); | |
2529 | } | |
2530 | ||
8cce5475 | 2531 | for (i = 0; i < nr; i++) { |
e9b61f19 KS |
2532 | struct page *subpage = head + i; |
2533 | if (subpage == page) | |
2534 | continue; | |
2535 | unlock_page(subpage); | |
2536 | ||
2537 | /* | |
2538 | * Subpages may be freed if there wasn't any mapping | |
2539 | * like if add_to_swap() is running on a lru page that | |
2540 | * had its mapping zapped. And freeing these pages | |
2541 | * requires taking the lru_lock so we do the put_page | |
2542 | * of the tail pages after the split is complete. | |
2543 | */ | |
2544 | put_page(subpage); | |
2545 | } | |
2546 | } | |
2547 | ||
b20ce5e0 KS |
2548 | int total_mapcount(struct page *page) |
2549 | { | |
86b562b6 | 2550 | int i, compound, nr, ret; |
b20ce5e0 KS |
2551 | |
2552 | VM_BUG_ON_PAGE(PageTail(page), page); | |
2553 | ||
2554 | if (likely(!PageCompound(page))) | |
2555 | return atomic_read(&page->_mapcount) + 1; | |
2556 | ||
dd78fedd | 2557 | compound = compound_mapcount(page); |
86b562b6 | 2558 | nr = compound_nr(page); |
b20ce5e0 | 2559 | if (PageHuge(page)) |
dd78fedd KS |
2560 | return compound; |
2561 | ret = compound; | |
86b562b6 | 2562 | for (i = 0; i < nr; i++) |
b20ce5e0 | 2563 | ret += atomic_read(&page[i]._mapcount) + 1; |
dd78fedd KS |
2564 | /* File pages has compound_mapcount included in _mapcount */ |
2565 | if (!PageAnon(page)) | |
86b562b6 | 2566 | return ret - compound * nr; |
b20ce5e0 | 2567 | if (PageDoubleMap(page)) |
86b562b6 | 2568 | ret -= nr; |
b20ce5e0 KS |
2569 | return ret; |
2570 | } | |
2571 | ||
6d0a07ed AA |
2572 | /* |
2573 | * This calculates accurately how many mappings a transparent hugepage | |
2574 | * has (unlike page_mapcount() which isn't fully accurate). This full | |
2575 | * accuracy is primarily needed to know if copy-on-write faults can | |
2576 | * reuse the page and change the mapping to read-write instead of | |
2577 | * copying them. At the same time this returns the total_mapcount too. | |
2578 | * | |
2579 | * The function returns the highest mapcount any one of the subpages | |
2580 | * has. If the return value is one, even if different processes are | |
2581 | * mapping different subpages of the transparent hugepage, they can | |
2582 | * all reuse it, because each process is reusing a different subpage. | |
2583 | * | |
2584 | * The total_mapcount is instead counting all virtual mappings of the | |
2585 | * subpages. If the total_mapcount is equal to "one", it tells the | |
2586 | * caller all mappings belong to the same "mm" and in turn the | |
2587 | * anon_vma of the transparent hugepage can become the vma->anon_vma | |
2588 | * local one as no other process may be mapping any of the subpages. | |
2589 | * | |
2590 | * It would be more accurate to replace page_mapcount() with | |
2591 | * page_trans_huge_mapcount(), however we only use | |
2592 | * page_trans_huge_mapcount() in the copy-on-write faults where we | |
2593 | * need full accuracy to avoid breaking page pinning, because | |
2594 | * page_trans_huge_mapcount() is slower than page_mapcount(). | |
2595 | */ | |
2596 | int page_trans_huge_mapcount(struct page *page, int *total_mapcount) | |
2597 | { | |
2598 | int i, ret, _total_mapcount, mapcount; | |
2599 | ||
2600 | /* hugetlbfs shouldn't call it */ | |
2601 | VM_BUG_ON_PAGE(PageHuge(page), page); | |
2602 | ||
2603 | if (likely(!PageTransCompound(page))) { | |
2604 | mapcount = atomic_read(&page->_mapcount) + 1; | |
2605 | if (total_mapcount) | |
2606 | *total_mapcount = mapcount; | |
2607 | return mapcount; | |
2608 | } | |
2609 | ||
2610 | page = compound_head(page); | |
2611 | ||
2612 | _total_mapcount = ret = 0; | |
65dfe3c3 | 2613 | for (i = 0; i < thp_nr_pages(page); i++) { |
6d0a07ed AA |
2614 | mapcount = atomic_read(&page[i]._mapcount) + 1; |
2615 | ret = max(ret, mapcount); | |
2616 | _total_mapcount += mapcount; | |
2617 | } | |
2618 | if (PageDoubleMap(page)) { | |
2619 | ret -= 1; | |
65dfe3c3 | 2620 | _total_mapcount -= thp_nr_pages(page); |
6d0a07ed AA |
2621 | } |
2622 | mapcount = compound_mapcount(page); | |
2623 | ret += mapcount; | |
2624 | _total_mapcount += mapcount; | |
2625 | if (total_mapcount) | |
2626 | *total_mapcount = _total_mapcount; | |
2627 | return ret; | |
2628 | } | |
2629 | ||
b8f593cd HY |
2630 | /* Racy check whether the huge page can be split */ |
2631 | bool can_split_huge_page(struct page *page, int *pextra_pins) | |
2632 | { | |
2633 | int extra_pins; | |
2634 | ||
aa5dc07f | 2635 | /* Additional pins from page cache */ |
b8f593cd | 2636 | if (PageAnon(page)) |
e2333dad | 2637 | extra_pins = PageSwapCache(page) ? thp_nr_pages(page) : 0; |
b8f593cd | 2638 | else |
e2333dad | 2639 | extra_pins = thp_nr_pages(page); |
b8f593cd HY |
2640 | if (pextra_pins) |
2641 | *pextra_pins = extra_pins; | |
2642 | return total_mapcount(page) == page_count(page) - extra_pins - 1; | |
2643 | } | |
2644 | ||
e9b61f19 KS |
2645 | /* |
2646 | * This function splits huge page into normal pages. @page can point to any | |
2647 | * subpage of huge page to split. Split doesn't change the position of @page. | |
2648 | * | |
2649 | * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. | |
2650 | * The huge page must be locked. | |
2651 | * | |
2652 | * If @list is null, tail pages will be added to LRU list, otherwise, to @list. | |
2653 | * | |
2654 | * Both head page and tail pages will inherit mapping, flags, and so on from | |
2655 | * the hugepage. | |
2656 | * | |
2657 | * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if | |
2658 | * they are not mapped. | |
2659 | * | |
2660 | * Returns 0 if the hugepage is split successfully. | |
2661 | * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under | |
2662 | * us. | |
2663 | */ | |
2664 | int split_huge_page_to_list(struct page *page, struct list_head *list) | |
2665 | { | |
2666 | struct page *head = compound_head(page); | |
a8803e6c | 2667 | struct deferred_split *ds_queue = get_deferred_split_queue(head); |
baa355fd KS |
2668 | struct anon_vma *anon_vma = NULL; |
2669 | struct address_space *mapping = NULL; | |
2670 | int count, mapcount, extra_pins, ret; | |
006d3ff2 | 2671 | pgoff_t end; |
e9b61f19 | 2672 | |
cb829624 | 2673 | VM_BUG_ON_PAGE(is_huge_zero_page(head), head); |
a8803e6c WY |
2674 | VM_BUG_ON_PAGE(!PageLocked(head), head); |
2675 | VM_BUG_ON_PAGE(!PageCompound(head), head); | |
e9b61f19 | 2676 | |
a8803e6c | 2677 | if (PageWriteback(head)) |
59807685 HY |
2678 | return -EBUSY; |
2679 | ||
baa355fd KS |
2680 | if (PageAnon(head)) { |
2681 | /* | |
c1e8d7c6 | 2682 | * The caller does not necessarily hold an mmap_lock that would |
baa355fd KS |
2683 | * prevent the anon_vma disappearing so we first we take a |
2684 | * reference to it and then lock the anon_vma for write. This | |
2685 | * is similar to page_lock_anon_vma_read except the write lock | |
2686 | * is taken to serialise against parallel split or collapse | |
2687 | * operations. | |
2688 | */ | |
2689 | anon_vma = page_get_anon_vma(head); | |
2690 | if (!anon_vma) { | |
2691 | ret = -EBUSY; | |
2692 | goto out; | |
2693 | } | |
006d3ff2 | 2694 | end = -1; |
baa355fd KS |
2695 | mapping = NULL; |
2696 | anon_vma_lock_write(anon_vma); | |
2697 | } else { | |
2698 | mapping = head->mapping; | |
2699 | ||
2700 | /* Truncated ? */ | |
2701 | if (!mapping) { | |
2702 | ret = -EBUSY; | |
2703 | goto out; | |
2704 | } | |
2705 | ||
baa355fd KS |
2706 | anon_vma = NULL; |
2707 | i_mmap_lock_read(mapping); | |
006d3ff2 HD |
2708 | |
2709 | /* | |
2710 | *__split_huge_page() may need to trim off pages beyond EOF: | |
2711 | * but on 32-bit, i_size_read() takes an irq-unsafe seqlock, | |
2712 | * which cannot be nested inside the page tree lock. So note | |
2713 | * end now: i_size itself may be changed at any moment, but | |
2714 | * head page lock is good enough to serialize the trimming. | |
2715 | */ | |
2716 | end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE); | |
e9b61f19 | 2717 | } |
e9b61f19 KS |
2718 | |
2719 | /* | |
906f9cdf | 2720 | * Racy check if we can split the page, before unmap_page() will |
e9b61f19 KS |
2721 | * split PMDs |
2722 | */ | |
b8f593cd | 2723 | if (!can_split_huge_page(head, &extra_pins)) { |
e9b61f19 KS |
2724 | ret = -EBUSY; |
2725 | goto out_unlock; | |
2726 | } | |
2727 | ||
906f9cdf | 2728 | unmap_page(head); |
e9b61f19 KS |
2729 | VM_BUG_ON_PAGE(compound_mapcount(head), head); |
2730 | ||
b6769834 AS |
2731 | /* block interrupt reentry in xa_lock and spinlock */ |
2732 | local_irq_disable(); | |
baa355fd | 2733 | if (mapping) { |
aa5dc07f | 2734 | XA_STATE(xas, &mapping->i_pages, page_index(head)); |
baa355fd | 2735 | |
baa355fd | 2736 | /* |
aa5dc07f | 2737 | * Check if the head page is present in page cache. |
baa355fd KS |
2738 | * We assume all tail are present too, if head is there. |
2739 | */ | |
aa5dc07f MW |
2740 | xa_lock(&mapping->i_pages); |
2741 | if (xas_load(&xas) != head) | |
baa355fd KS |
2742 | goto fail; |
2743 | } | |
2744 | ||
0139aa7b | 2745 | /* Prevent deferred_split_scan() touching ->_refcount */ |
364c1eeb | 2746 | spin_lock(&ds_queue->split_queue_lock); |
e9b61f19 KS |
2747 | count = page_count(head); |
2748 | mapcount = total_mapcount(head); | |
baa355fd | 2749 | if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) { |
9a982250 | 2750 | if (!list_empty(page_deferred_list(head))) { |
364c1eeb | 2751 | ds_queue->split_queue_len--; |
9a982250 KS |
2752 | list_del(page_deferred_list(head)); |
2753 | } | |
afb97172 | 2754 | spin_unlock(&ds_queue->split_queue_lock); |
06d3eff6 | 2755 | if (mapping) { |
bf9ecead MS |
2756 | int nr = thp_nr_pages(head); |
2757 | ||
a8803e6c | 2758 | if (PageSwapBacked(head)) |
57b2847d MS |
2759 | __mod_lruvec_page_state(head, NR_SHMEM_THPS, |
2760 | -nr); | |
06d3eff6 | 2761 | else |
bf9ecead MS |
2762 | __mod_lruvec_page_state(head, NR_FILE_THPS, |
2763 | -nr); | |
06d3eff6 KS |
2764 | } |
2765 | ||
b6769834 | 2766 | __split_huge_page(page, list, end); |
c4f9c701 | 2767 | ret = 0; |
e9b61f19 | 2768 | } else { |
baa355fd KS |
2769 | if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) { |
2770 | pr_alert("total_mapcount: %u, page_count(): %u\n", | |
2771 | mapcount, count); | |
2772 | if (PageTail(page)) | |
2773 | dump_page(head, NULL); | |
2774 | dump_page(page, "total_mapcount(head) > 0"); | |
2775 | BUG(); | |
2776 | } | |
364c1eeb | 2777 | spin_unlock(&ds_queue->split_queue_lock); |
baa355fd | 2778 | fail: if (mapping) |
b93b0163 | 2779 | xa_unlock(&mapping->i_pages); |
b6769834 | 2780 | local_irq_enable(); |
8cce5475 | 2781 | remap_page(head, thp_nr_pages(head)); |
e9b61f19 KS |
2782 | ret = -EBUSY; |
2783 | } | |
2784 | ||
2785 | out_unlock: | |
baa355fd KS |
2786 | if (anon_vma) { |
2787 | anon_vma_unlock_write(anon_vma); | |
2788 | put_anon_vma(anon_vma); | |
2789 | } | |
2790 | if (mapping) | |
2791 | i_mmap_unlock_read(mapping); | |
e9b61f19 KS |
2792 | out: |
2793 | count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); | |
2794 | return ret; | |
2795 | } | |
9a982250 KS |
2796 | |
2797 | void free_transhuge_page(struct page *page) | |
2798 | { | |
87eaceb3 | 2799 | struct deferred_split *ds_queue = get_deferred_split_queue(page); |
9a982250 KS |
2800 | unsigned long flags; |
2801 | ||
364c1eeb | 2802 | spin_lock_irqsave(&ds_queue->split_queue_lock, flags); |
9a982250 | 2803 | if (!list_empty(page_deferred_list(page))) { |
364c1eeb | 2804 | ds_queue->split_queue_len--; |
9a982250 KS |
2805 | list_del(page_deferred_list(page)); |
2806 | } | |
364c1eeb | 2807 | spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); |
9a982250 KS |
2808 | free_compound_page(page); |
2809 | } | |
2810 | ||
2811 | void deferred_split_huge_page(struct page *page) | |
2812 | { | |
87eaceb3 YS |
2813 | struct deferred_split *ds_queue = get_deferred_split_queue(page); |
2814 | #ifdef CONFIG_MEMCG | |
bcfe06bf | 2815 | struct mem_cgroup *memcg = page_memcg(compound_head(page)); |
87eaceb3 | 2816 | #endif |
9a982250 KS |
2817 | unsigned long flags; |
2818 | ||
2819 | VM_BUG_ON_PAGE(!PageTransHuge(page), page); | |
2820 | ||
87eaceb3 YS |
2821 | /* |
2822 | * The try_to_unmap() in page reclaim path might reach here too, | |
2823 | * this may cause a race condition to corrupt deferred split queue. | |
2824 | * And, if page reclaim is already handling the same page, it is | |
2825 | * unnecessary to handle it again in shrinker. | |
2826 | * | |
2827 | * Check PageSwapCache to determine if the page is being | |
2828 | * handled by page reclaim since THP swap would add the page into | |
2829 | * swap cache before calling try_to_unmap(). | |
2830 | */ | |
2831 | if (PageSwapCache(page)) | |
2832 | return; | |
2833 | ||
364c1eeb | 2834 | spin_lock_irqsave(&ds_queue->split_queue_lock, flags); |
9a982250 | 2835 | if (list_empty(page_deferred_list(page))) { |
f9719a03 | 2836 | count_vm_event(THP_DEFERRED_SPLIT_PAGE); |
364c1eeb YS |
2837 | list_add_tail(page_deferred_list(page), &ds_queue->split_queue); |
2838 | ds_queue->split_queue_len++; | |
87eaceb3 YS |
2839 | #ifdef CONFIG_MEMCG |
2840 | if (memcg) | |
2841 | memcg_set_shrinker_bit(memcg, page_to_nid(page), | |
2842 | deferred_split_shrinker.id); | |
2843 | #endif | |
9a982250 | 2844 | } |
364c1eeb | 2845 | spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); |
9a982250 KS |
2846 | } |
2847 | ||
2848 | static unsigned long deferred_split_count(struct shrinker *shrink, | |
2849 | struct shrink_control *sc) | |
2850 | { | |
a3d0a918 | 2851 | struct pglist_data *pgdata = NODE_DATA(sc->nid); |
364c1eeb | 2852 | struct deferred_split *ds_queue = &pgdata->deferred_split_queue; |
87eaceb3 YS |
2853 | |
2854 | #ifdef CONFIG_MEMCG | |
2855 | if (sc->memcg) | |
2856 | ds_queue = &sc->memcg->deferred_split_queue; | |
2857 | #endif | |
364c1eeb | 2858 | return READ_ONCE(ds_queue->split_queue_len); |
9a982250 KS |
2859 | } |
2860 | ||
2861 | static unsigned long deferred_split_scan(struct shrinker *shrink, | |
2862 | struct shrink_control *sc) | |
2863 | { | |
a3d0a918 | 2864 | struct pglist_data *pgdata = NODE_DATA(sc->nid); |
364c1eeb | 2865 | struct deferred_split *ds_queue = &pgdata->deferred_split_queue; |
9a982250 KS |
2866 | unsigned long flags; |
2867 | LIST_HEAD(list), *pos, *next; | |
2868 | struct page *page; | |
2869 | int split = 0; | |
2870 | ||
87eaceb3 YS |
2871 | #ifdef CONFIG_MEMCG |
2872 | if (sc->memcg) | |
2873 | ds_queue = &sc->memcg->deferred_split_queue; | |
2874 | #endif | |
2875 | ||
364c1eeb | 2876 | spin_lock_irqsave(&ds_queue->split_queue_lock, flags); |
9a982250 | 2877 | /* Take pin on all head pages to avoid freeing them under us */ |
364c1eeb | 2878 | list_for_each_safe(pos, next, &ds_queue->split_queue) { |
9a982250 KS |
2879 | page = list_entry((void *)pos, struct page, mapping); |
2880 | page = compound_head(page); | |
e3ae1953 KS |
2881 | if (get_page_unless_zero(page)) { |
2882 | list_move(page_deferred_list(page), &list); | |
2883 | } else { | |
2884 | /* We lost race with put_compound_page() */ | |
9a982250 | 2885 | list_del_init(page_deferred_list(page)); |
364c1eeb | 2886 | ds_queue->split_queue_len--; |
9a982250 | 2887 | } |
e3ae1953 KS |
2888 | if (!--sc->nr_to_scan) |
2889 | break; | |
9a982250 | 2890 | } |
364c1eeb | 2891 | spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); |
9a982250 KS |
2892 | |
2893 | list_for_each_safe(pos, next, &list) { | |
2894 | page = list_entry((void *)pos, struct page, mapping); | |
fa41b900 KS |
2895 | if (!trylock_page(page)) |
2896 | goto next; | |
9a982250 KS |
2897 | /* split_huge_page() removes page from list on success */ |
2898 | if (!split_huge_page(page)) | |
2899 | split++; | |
2900 | unlock_page(page); | |
fa41b900 | 2901 | next: |
9a982250 KS |
2902 | put_page(page); |
2903 | } | |
2904 | ||
364c1eeb YS |
2905 | spin_lock_irqsave(&ds_queue->split_queue_lock, flags); |
2906 | list_splice_tail(&list, &ds_queue->split_queue); | |
2907 | spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); | |
9a982250 | 2908 | |
cb8d68ec KS |
2909 | /* |
2910 | * Stop shrinker if we didn't split any page, but the queue is empty. | |
2911 | * This can happen if pages were freed under us. | |
2912 | */ | |
364c1eeb | 2913 | if (!split && list_empty(&ds_queue->split_queue)) |
cb8d68ec KS |
2914 | return SHRINK_STOP; |
2915 | return split; | |
9a982250 KS |
2916 | } |
2917 | ||
2918 | static struct shrinker deferred_split_shrinker = { | |
2919 | .count_objects = deferred_split_count, | |
2920 | .scan_objects = deferred_split_scan, | |
2921 | .seeks = DEFAULT_SEEKS, | |
87eaceb3 YS |
2922 | .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE | |
2923 | SHRINKER_NONSLAB, | |
9a982250 | 2924 | }; |
49071d43 KS |
2925 | |
2926 | #ifdef CONFIG_DEBUG_FS | |
2927 | static int split_huge_pages_set(void *data, u64 val) | |
2928 | { | |
2929 | struct zone *zone; | |
2930 | struct page *page; | |
2931 | unsigned long pfn, max_zone_pfn; | |
2932 | unsigned long total = 0, split = 0; | |
2933 | ||
2934 | if (val != 1) | |
2935 | return -EINVAL; | |
2936 | ||
2937 | for_each_populated_zone(zone) { | |
2938 | max_zone_pfn = zone_end_pfn(zone); | |
2939 | for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { | |
2940 | if (!pfn_valid(pfn)) | |
2941 | continue; | |
2942 | ||
2943 | page = pfn_to_page(pfn); | |
2944 | if (!get_page_unless_zero(page)) | |
2945 | continue; | |
2946 | ||
2947 | if (zone != page_zone(page)) | |
2948 | goto next; | |
2949 | ||
baa355fd | 2950 | if (!PageHead(page) || PageHuge(page) || !PageLRU(page)) |
49071d43 KS |
2951 | goto next; |
2952 | ||
2953 | total++; | |
2954 | lock_page(page); | |
2955 | if (!split_huge_page(page)) | |
2956 | split++; | |
2957 | unlock_page(page); | |
2958 | next: | |
2959 | put_page(page); | |
2960 | } | |
2961 | } | |
2962 | ||
145bdaa1 | 2963 | pr_info("%lu of %lu THP split\n", split, total); |
49071d43 KS |
2964 | |
2965 | return 0; | |
2966 | } | |
f1287869 | 2967 | DEFINE_DEBUGFS_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set, |
49071d43 KS |
2968 | "%llu\n"); |
2969 | ||
2970 | static int __init split_huge_pages_debugfs(void) | |
2971 | { | |
d9f7979c GKH |
2972 | debugfs_create_file("split_huge_pages", 0200, NULL, NULL, |
2973 | &split_huge_pages_fops); | |
49071d43 KS |
2974 | return 0; |
2975 | } | |
2976 | late_initcall(split_huge_pages_debugfs); | |
2977 | #endif | |
616b8371 ZY |
2978 | |
2979 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION | |
2980 | void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw, | |
2981 | struct page *page) | |
2982 | { | |
2983 | struct vm_area_struct *vma = pvmw->vma; | |
2984 | struct mm_struct *mm = vma->vm_mm; | |
2985 | unsigned long address = pvmw->address; | |
2986 | pmd_t pmdval; | |
2987 | swp_entry_t entry; | |
ab6e3d09 | 2988 | pmd_t pmdswp; |
616b8371 ZY |
2989 | |
2990 | if (!(pvmw->pmd && !pvmw->pte)) | |
2991 | return; | |
2992 | ||
616b8371 | 2993 | flush_cache_range(vma, address, address + HPAGE_PMD_SIZE); |
8a8683ad | 2994 | pmdval = pmdp_invalidate(vma, address, pvmw->pmd); |
616b8371 ZY |
2995 | if (pmd_dirty(pmdval)) |
2996 | set_page_dirty(page); | |
2997 | entry = make_migration_entry(page, pmd_write(pmdval)); | |
ab6e3d09 NH |
2998 | pmdswp = swp_entry_to_pmd(entry); |
2999 | if (pmd_soft_dirty(pmdval)) | |
3000 | pmdswp = pmd_swp_mksoft_dirty(pmdswp); | |
3001 | set_pmd_at(mm, address, pvmw->pmd, pmdswp); | |
616b8371 ZY |
3002 | page_remove_rmap(page, true); |
3003 | put_page(page); | |
616b8371 ZY |
3004 | } |
3005 | ||
3006 | void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new) | |
3007 | { | |
3008 | struct vm_area_struct *vma = pvmw->vma; | |
3009 | struct mm_struct *mm = vma->vm_mm; | |
3010 | unsigned long address = pvmw->address; | |
3011 | unsigned long mmun_start = address & HPAGE_PMD_MASK; | |
3012 | pmd_t pmde; | |
3013 | swp_entry_t entry; | |
3014 | ||
3015 | if (!(pvmw->pmd && !pvmw->pte)) | |
3016 | return; | |
3017 | ||
3018 | entry = pmd_to_swp_entry(*pvmw->pmd); | |
3019 | get_page(new); | |
3020 | pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot)); | |
ab6e3d09 NH |
3021 | if (pmd_swp_soft_dirty(*pvmw->pmd)) |
3022 | pmde = pmd_mksoft_dirty(pmde); | |
616b8371 | 3023 | if (is_write_migration_entry(entry)) |
f55e1014 | 3024 | pmde = maybe_pmd_mkwrite(pmde, vma); |
616b8371 ZY |
3025 | |
3026 | flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE); | |
e71769ae NH |
3027 | if (PageAnon(new)) |
3028 | page_add_anon_rmap(new, vma, mmun_start, true); | |
3029 | else | |
3030 | page_add_file_rmap(new, true); | |
616b8371 | 3031 | set_pmd_at(mm, mmun_start, pvmw->pmd, pmde); |
e125fe40 | 3032 | if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new)) |
616b8371 ZY |
3033 | mlock_vma_page(new); |
3034 | update_mmu_cache_pmd(vma, address, pvmw->pmd); | |
3035 | } | |
3036 | #endif |