]>
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> | |
fa6c0231 | 10 | #include <linux/sched/mm.h> |
f7ccbae4 | 11 | #include <linux/sched/coredump.h> |
6a3827d7 | 12 | #include <linux/sched/numa_balancing.h> |
71e3aac0 AA |
13 | #include <linux/highmem.h> |
14 | #include <linux/hugetlb.h> | |
15 | #include <linux/mmu_notifier.h> | |
16 | #include <linux/rmap.h> | |
17 | #include <linux/swap.h> | |
97ae1749 | 18 | #include <linux/shrinker.h> |
ba76149f | 19 | #include <linux/mm_inline.h> |
e9b61f19 | 20 | #include <linux/swapops.h> |
4897c765 | 21 | #include <linux/dax.h> |
ba76149f | 22 | #include <linux/khugepaged.h> |
878aee7d | 23 | #include <linux/freezer.h> |
f25748e3 | 24 | #include <linux/pfn_t.h> |
a664b2d8 | 25 | #include <linux/mman.h> |
3565fce3 | 26 | #include <linux/memremap.h> |
325adeb5 | 27 | #include <linux/pagemap.h> |
49071d43 | 28 | #include <linux/debugfs.h> |
4daae3b4 | 29 | #include <linux/migrate.h> |
43b5fbbd | 30 | #include <linux/hashtable.h> |
6b251fc9 | 31 | #include <linux/userfaultfd_k.h> |
33c3fc71 | 32 | #include <linux/page_idle.h> |
baa355fd | 33 | #include <linux/shmem_fs.h> |
6b31d595 | 34 | #include <linux/oom.h> |
98fa15f3 | 35 | #include <linux/numa.h> |
f7da677b | 36 | #include <linux/page_owner.h> |
97ae1749 | 37 | |
71e3aac0 AA |
38 | #include <asm/tlb.h> |
39 | #include <asm/pgalloc.h> | |
40 | #include "internal.h" | |
41 | ||
ba76149f | 42 | /* |
b14d595a MD |
43 | * By default, transparent hugepage support is disabled in order to avoid |
44 | * risking an increased memory footprint for applications that are not | |
45 | * guaranteed to benefit from it. When transparent hugepage support is | |
46 | * enabled, it is for all mappings, and khugepaged scans all mappings. | |
8bfa3f9a JW |
47 | * Defrag is invoked by khugepaged hugepage allocations and by page faults |
48 | * for all hugepage allocations. | |
ba76149f | 49 | */ |
71e3aac0 | 50 | unsigned long transparent_hugepage_flags __read_mostly = |
13ece886 | 51 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS |
ba76149f | 52 | (1<<TRANSPARENT_HUGEPAGE_FLAG)| |
13ece886 AA |
53 | #endif |
54 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE | |
55 | (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| | |
56 | #endif | |
444eb2a4 | 57 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)| |
79da5407 KS |
58 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| |
59 | (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
ba76149f | 60 | |
9a982250 | 61 | static struct shrinker deferred_split_shrinker; |
f000565a | 62 | |
97ae1749 | 63 | static atomic_t huge_zero_refcount; |
56873f43 | 64 | struct page *huge_zero_page __read_mostly; |
3b77e8c8 | 65 | unsigned long huge_zero_pfn __read_mostly = ~0UL; |
4a6c1297 | 66 | |
e6be37b2 ML |
67 | static inline bool file_thp_enabled(struct vm_area_struct *vma) |
68 | { | |
69 | return transhuge_vma_enabled(vma, vma->vm_flags) && vma->vm_file && | |
70 | !inode_is_open_for_write(vma->vm_file->f_inode) && | |
71 | (vma->vm_flags & VM_EXEC); | |
72 | } | |
73 | ||
74 | bool transparent_hugepage_active(struct vm_area_struct *vma) | |
7635d9cb | 75 | { |
c0630669 YS |
76 | /* The addr is used to check if the vma size fits */ |
77 | unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE; | |
78 | ||
79 | if (!transhuge_vma_suitable(vma, addr)) | |
80 | return false; | |
7635d9cb MH |
81 | if (vma_is_anonymous(vma)) |
82 | return __transparent_hugepage_enabled(vma); | |
c0630669 YS |
83 | if (vma_is_shmem(vma)) |
84 | return shmem_huge_enabled(vma); | |
e6be37b2 ML |
85 | if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS)) |
86 | return file_thp_enabled(vma); | |
7635d9cb MH |
87 | |
88 | return false; | |
89 | } | |
90 | ||
aaa9705b | 91 | static bool get_huge_zero_page(void) |
97ae1749 KS |
92 | { |
93 | struct page *zero_page; | |
94 | retry: | |
95 | if (likely(atomic_inc_not_zero(&huge_zero_refcount))) | |
aaa9705b | 96 | return true; |
97ae1749 KS |
97 | |
98 | zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, | |
4a6c1297 | 99 | HPAGE_PMD_ORDER); |
d8a8e1f0 KS |
100 | if (!zero_page) { |
101 | count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); | |
aaa9705b | 102 | return false; |
d8a8e1f0 KS |
103 | } |
104 | count_vm_event(THP_ZERO_PAGE_ALLOC); | |
97ae1749 | 105 | preempt_disable(); |
5918d10a | 106 | if (cmpxchg(&huge_zero_page, NULL, zero_page)) { |
97ae1749 | 107 | preempt_enable(); |
5ddacbe9 | 108 | __free_pages(zero_page, compound_order(zero_page)); |
97ae1749 KS |
109 | goto retry; |
110 | } | |
3b77e8c8 | 111 | WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page)); |
97ae1749 KS |
112 | |
113 | /* We take additional reference here. It will be put back by shrinker */ | |
114 | atomic_set(&huge_zero_refcount, 2); | |
115 | preempt_enable(); | |
aaa9705b | 116 | return true; |
4a6c1297 KS |
117 | } |
118 | ||
6fcb52a5 | 119 | static void put_huge_zero_page(void) |
4a6c1297 | 120 | { |
97ae1749 KS |
121 | /* |
122 | * Counter should never go to zero here. Only shrinker can put | |
123 | * last reference. | |
124 | */ | |
125 | BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); | |
4a6c1297 KS |
126 | } |
127 | ||
6fcb52a5 AL |
128 | struct page *mm_get_huge_zero_page(struct mm_struct *mm) |
129 | { | |
130 | if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) | |
131 | return READ_ONCE(huge_zero_page); | |
132 | ||
133 | if (!get_huge_zero_page()) | |
134 | return NULL; | |
135 | ||
136 | if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) | |
137 | put_huge_zero_page(); | |
138 | ||
139 | return READ_ONCE(huge_zero_page); | |
140 | } | |
141 | ||
142 | void mm_put_huge_zero_page(struct mm_struct *mm) | |
143 | { | |
144 | if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) | |
145 | put_huge_zero_page(); | |
146 | } | |
147 | ||
48896466 GC |
148 | static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, |
149 | struct shrink_control *sc) | |
4a6c1297 | 150 | { |
48896466 GC |
151 | /* we can free zero page only if last reference remains */ |
152 | return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; | |
153 | } | |
97ae1749 | 154 | |
48896466 GC |
155 | static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, |
156 | struct shrink_control *sc) | |
157 | { | |
97ae1749 | 158 | if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { |
5918d10a KS |
159 | struct page *zero_page = xchg(&huge_zero_page, NULL); |
160 | BUG_ON(zero_page == NULL); | |
3b77e8c8 | 161 | WRITE_ONCE(huge_zero_pfn, ~0UL); |
5ddacbe9 | 162 | __free_pages(zero_page, compound_order(zero_page)); |
48896466 | 163 | return HPAGE_PMD_NR; |
97ae1749 KS |
164 | } |
165 | ||
166 | return 0; | |
4a6c1297 KS |
167 | } |
168 | ||
97ae1749 | 169 | static struct shrinker huge_zero_page_shrinker = { |
48896466 GC |
170 | .count_objects = shrink_huge_zero_page_count, |
171 | .scan_objects = shrink_huge_zero_page_scan, | |
97ae1749 KS |
172 | .seeks = DEFAULT_SEEKS, |
173 | }; | |
174 | ||
71e3aac0 | 175 | #ifdef CONFIG_SYSFS |
71e3aac0 AA |
176 | static ssize_t enabled_show(struct kobject *kobj, |
177 | struct kobj_attribute *attr, char *buf) | |
178 | { | |
bfb0ffeb JP |
179 | const char *output; |
180 | ||
444eb2a4 | 181 | if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags)) |
bfb0ffeb JP |
182 | output = "[always] madvise never"; |
183 | else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
184 | &transparent_hugepage_flags)) | |
185 | output = "always [madvise] never"; | |
444eb2a4 | 186 | else |
bfb0ffeb JP |
187 | output = "always madvise [never]"; |
188 | ||
189 | return sysfs_emit(buf, "%s\n", output); | |
71e3aac0 | 190 | } |
444eb2a4 | 191 | |
71e3aac0 AA |
192 | static ssize_t enabled_store(struct kobject *kobj, |
193 | struct kobj_attribute *attr, | |
194 | const char *buf, size_t count) | |
195 | { | |
21440d7e | 196 | ssize_t ret = count; |
ba76149f | 197 | |
f42f2552 | 198 | if (sysfs_streq(buf, "always")) { |
21440d7e DR |
199 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); |
200 | set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); | |
f42f2552 | 201 | } else if (sysfs_streq(buf, "madvise")) { |
21440d7e DR |
202 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); |
203 | set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
f42f2552 | 204 | } else if (sysfs_streq(buf, "never")) { |
21440d7e DR |
205 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); |
206 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
207 | } else | |
208 | ret = -EINVAL; | |
ba76149f AA |
209 | |
210 | if (ret > 0) { | |
b46e756f | 211 | int err = start_stop_khugepaged(); |
ba76149f AA |
212 | if (err) |
213 | ret = err; | |
214 | } | |
ba76149f | 215 | return ret; |
71e3aac0 AA |
216 | } |
217 | static struct kobj_attribute enabled_attr = | |
218 | __ATTR(enabled, 0644, enabled_show, enabled_store); | |
219 | ||
b46e756f | 220 | ssize_t single_hugepage_flag_show(struct kobject *kobj, |
bfb0ffeb JP |
221 | struct kobj_attribute *attr, char *buf, |
222 | enum transparent_hugepage_flag flag) | |
71e3aac0 | 223 | { |
bfb0ffeb JP |
224 | return sysfs_emit(buf, "%d\n", |
225 | !!test_bit(flag, &transparent_hugepage_flags)); | |
71e3aac0 | 226 | } |
e27e6151 | 227 | |
b46e756f | 228 | ssize_t single_hugepage_flag_store(struct kobject *kobj, |
71e3aac0 AA |
229 | struct kobj_attribute *attr, |
230 | const char *buf, size_t count, | |
231 | enum transparent_hugepage_flag flag) | |
232 | { | |
e27e6151 BH |
233 | unsigned long value; |
234 | int ret; | |
235 | ||
236 | ret = kstrtoul(buf, 10, &value); | |
237 | if (ret < 0) | |
238 | return ret; | |
239 | if (value > 1) | |
240 | return -EINVAL; | |
241 | ||
242 | if (value) | |
71e3aac0 | 243 | set_bit(flag, &transparent_hugepage_flags); |
e27e6151 | 244 | else |
71e3aac0 | 245 | clear_bit(flag, &transparent_hugepage_flags); |
71e3aac0 AA |
246 | |
247 | return count; | |
248 | } | |
249 | ||
71e3aac0 AA |
250 | static ssize_t defrag_show(struct kobject *kobj, |
251 | struct kobj_attribute *attr, char *buf) | |
252 | { | |
bfb0ffeb JP |
253 | const char *output; |
254 | ||
255 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, | |
256 | &transparent_hugepage_flags)) | |
257 | output = "[always] defer defer+madvise madvise never"; | |
258 | else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, | |
259 | &transparent_hugepage_flags)) | |
260 | output = "always [defer] defer+madvise madvise never"; | |
261 | else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, | |
262 | &transparent_hugepage_flags)) | |
263 | output = "always defer [defer+madvise] madvise never"; | |
264 | else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, | |
265 | &transparent_hugepage_flags)) | |
266 | output = "always defer defer+madvise [madvise] never"; | |
267 | else | |
268 | output = "always defer defer+madvise madvise [never]"; | |
269 | ||
270 | return sysfs_emit(buf, "%s\n", output); | |
71e3aac0 | 271 | } |
21440d7e | 272 | |
71e3aac0 AA |
273 | static ssize_t defrag_store(struct kobject *kobj, |
274 | struct kobj_attribute *attr, | |
275 | const char *buf, size_t count) | |
276 | { | |
f42f2552 | 277 | if (sysfs_streq(buf, "always")) { |
21440d7e DR |
278 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); |
279 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
280 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
281 | set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); | |
f42f2552 | 282 | } else if (sysfs_streq(buf, "defer+madvise")) { |
21440d7e DR |
283 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); |
284 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
285 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
286 | set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
f42f2552 | 287 | } else if (sysfs_streq(buf, "defer")) { |
4fad7fb6 DR |
288 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); |
289 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
290 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
291 | set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
f42f2552 | 292 | } else if (sysfs_streq(buf, "madvise")) { |
21440d7e DR |
293 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); |
294 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
295 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
296 | set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
f42f2552 | 297 | } else if (sysfs_streq(buf, "never")) { |
21440d7e DR |
298 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); |
299 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
300 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
301 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
302 | } else | |
303 | return -EINVAL; | |
304 | ||
305 | return count; | |
71e3aac0 AA |
306 | } |
307 | static struct kobj_attribute defrag_attr = | |
308 | __ATTR(defrag, 0644, defrag_show, defrag_store); | |
309 | ||
79da5407 | 310 | static ssize_t use_zero_page_show(struct kobject *kobj, |
ae7a927d | 311 | struct kobj_attribute *attr, char *buf) |
79da5407 | 312 | { |
b46e756f | 313 | return single_hugepage_flag_show(kobj, attr, buf, |
ae7a927d | 314 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); |
79da5407 KS |
315 | } |
316 | static ssize_t use_zero_page_store(struct kobject *kobj, | |
317 | struct kobj_attribute *attr, const char *buf, size_t count) | |
318 | { | |
b46e756f | 319 | return single_hugepage_flag_store(kobj, attr, buf, count, |
79da5407 KS |
320 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); |
321 | } | |
322 | static struct kobj_attribute use_zero_page_attr = | |
323 | __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); | |
49920d28 HD |
324 | |
325 | static ssize_t hpage_pmd_size_show(struct kobject *kobj, | |
ae7a927d | 326 | struct kobj_attribute *attr, char *buf) |
49920d28 | 327 | { |
ae7a927d | 328 | return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE); |
49920d28 HD |
329 | } |
330 | static struct kobj_attribute hpage_pmd_size_attr = | |
331 | __ATTR_RO(hpage_pmd_size); | |
332 | ||
71e3aac0 AA |
333 | static struct attribute *hugepage_attr[] = { |
334 | &enabled_attr.attr, | |
335 | &defrag_attr.attr, | |
79da5407 | 336 | &use_zero_page_attr.attr, |
49920d28 | 337 | &hpage_pmd_size_attr.attr, |
396bcc52 | 338 | #ifdef CONFIG_SHMEM |
5a6e75f8 | 339 | &shmem_enabled_attr.attr, |
71e3aac0 AA |
340 | #endif |
341 | NULL, | |
342 | }; | |
343 | ||
8aa95a21 | 344 | static const struct attribute_group hugepage_attr_group = { |
71e3aac0 | 345 | .attrs = hugepage_attr, |
ba76149f AA |
346 | }; |
347 | ||
569e5590 | 348 | static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) |
71e3aac0 | 349 | { |
71e3aac0 AA |
350 | int err; |
351 | ||
569e5590 SL |
352 | *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); |
353 | if (unlikely(!*hugepage_kobj)) { | |
ae3a8c1c | 354 | pr_err("failed to create transparent hugepage kobject\n"); |
569e5590 | 355 | return -ENOMEM; |
ba76149f AA |
356 | } |
357 | ||
569e5590 | 358 | err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); |
ba76149f | 359 | if (err) { |
ae3a8c1c | 360 | pr_err("failed to register transparent hugepage group\n"); |
569e5590 | 361 | goto delete_obj; |
ba76149f AA |
362 | } |
363 | ||
569e5590 | 364 | err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); |
ba76149f | 365 | if (err) { |
ae3a8c1c | 366 | pr_err("failed to register transparent hugepage group\n"); |
569e5590 | 367 | goto remove_hp_group; |
ba76149f | 368 | } |
569e5590 SL |
369 | |
370 | return 0; | |
371 | ||
372 | remove_hp_group: | |
373 | sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); | |
374 | delete_obj: | |
375 | kobject_put(*hugepage_kobj); | |
376 | return err; | |
377 | } | |
378 | ||
379 | static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
380 | { | |
381 | sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); | |
382 | sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); | |
383 | kobject_put(hugepage_kobj); | |
384 | } | |
385 | #else | |
386 | static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) | |
387 | { | |
388 | return 0; | |
389 | } | |
390 | ||
391 | static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
392 | { | |
393 | } | |
394 | #endif /* CONFIG_SYSFS */ | |
395 | ||
396 | static int __init hugepage_init(void) | |
397 | { | |
398 | int err; | |
399 | struct kobject *hugepage_kobj; | |
400 | ||
401 | if (!has_transparent_hugepage()) { | |
bae84953 AK |
402 | /* |
403 | * Hardware doesn't support hugepages, hence disable | |
404 | * DAX PMD support. | |
405 | */ | |
406 | transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX; | |
569e5590 SL |
407 | return -EINVAL; |
408 | } | |
409 | ||
ff20c2e0 KS |
410 | /* |
411 | * hugepages can't be allocated by the buddy allocator | |
412 | */ | |
413 | MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER); | |
414 | /* | |
415 | * we use page->mapping and page->index in second tail page | |
416 | * as list_head: assuming THP order >= 2 | |
417 | */ | |
418 | MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2); | |
419 | ||
569e5590 SL |
420 | err = hugepage_init_sysfs(&hugepage_kobj); |
421 | if (err) | |
65ebb64f | 422 | goto err_sysfs; |
ba76149f | 423 | |
b46e756f | 424 | err = khugepaged_init(); |
ba76149f | 425 | if (err) |
65ebb64f | 426 | goto err_slab; |
ba76149f | 427 | |
65ebb64f KS |
428 | err = register_shrinker(&huge_zero_page_shrinker); |
429 | if (err) | |
430 | goto err_hzp_shrinker; | |
9a982250 KS |
431 | err = register_shrinker(&deferred_split_shrinker); |
432 | if (err) | |
433 | goto err_split_shrinker; | |
97ae1749 | 434 | |
97562cd2 RR |
435 | /* |
436 | * By default disable transparent hugepages on smaller systems, | |
437 | * where the extra memory used could hurt more than TLB overhead | |
438 | * is likely to save. The admin can still enable it through /sys. | |
439 | */ | |
ca79b0c2 | 440 | if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) { |
97562cd2 | 441 | transparent_hugepage_flags = 0; |
79553da2 KS |
442 | return 0; |
443 | } | |
97562cd2 | 444 | |
79553da2 | 445 | err = start_stop_khugepaged(); |
65ebb64f KS |
446 | if (err) |
447 | goto err_khugepaged; | |
ba76149f | 448 | |
569e5590 | 449 | return 0; |
65ebb64f | 450 | err_khugepaged: |
9a982250 KS |
451 | unregister_shrinker(&deferred_split_shrinker); |
452 | err_split_shrinker: | |
65ebb64f KS |
453 | unregister_shrinker(&huge_zero_page_shrinker); |
454 | err_hzp_shrinker: | |
b46e756f | 455 | khugepaged_destroy(); |
65ebb64f | 456 | err_slab: |
569e5590 | 457 | hugepage_exit_sysfs(hugepage_kobj); |
65ebb64f | 458 | err_sysfs: |
ba76149f | 459 | return err; |
71e3aac0 | 460 | } |
a64fb3cd | 461 | subsys_initcall(hugepage_init); |
71e3aac0 AA |
462 | |
463 | static int __init setup_transparent_hugepage(char *str) | |
464 | { | |
465 | int ret = 0; | |
466 | if (!str) | |
467 | goto out; | |
468 | if (!strcmp(str, "always")) { | |
469 | set_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
470 | &transparent_hugepage_flags); | |
471 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
472 | &transparent_hugepage_flags); | |
473 | ret = 1; | |
474 | } else if (!strcmp(str, "madvise")) { | |
475 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
476 | &transparent_hugepage_flags); | |
477 | set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
478 | &transparent_hugepage_flags); | |
479 | ret = 1; | |
480 | } else if (!strcmp(str, "never")) { | |
481 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
482 | &transparent_hugepage_flags); | |
483 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
484 | &transparent_hugepage_flags); | |
485 | ret = 1; | |
486 | } | |
487 | out: | |
488 | if (!ret) | |
ae3a8c1c | 489 | pr_warn("transparent_hugepage= cannot parse, ignored\n"); |
71e3aac0 AA |
490 | return ret; |
491 | } | |
492 | __setup("transparent_hugepage=", setup_transparent_hugepage); | |
493 | ||
f55e1014 | 494 | pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) |
71e3aac0 | 495 | { |
f55e1014 | 496 | if (likely(vma->vm_flags & VM_WRITE)) |
71e3aac0 AA |
497 | pmd = pmd_mkwrite(pmd); |
498 | return pmd; | |
499 | } | |
500 | ||
87eaceb3 YS |
501 | #ifdef CONFIG_MEMCG |
502 | static inline struct deferred_split *get_deferred_split_queue(struct page *page) | |
9a982250 | 503 | { |
bcfe06bf | 504 | struct mem_cgroup *memcg = page_memcg(compound_head(page)); |
87eaceb3 YS |
505 | struct pglist_data *pgdat = NODE_DATA(page_to_nid(page)); |
506 | ||
507 | if (memcg) | |
508 | return &memcg->deferred_split_queue; | |
509 | else | |
510 | return &pgdat->deferred_split_queue; | |
9a982250 | 511 | } |
87eaceb3 YS |
512 | #else |
513 | static inline struct deferred_split *get_deferred_split_queue(struct page *page) | |
514 | { | |
515 | struct pglist_data *pgdat = NODE_DATA(page_to_nid(page)); | |
516 | ||
517 | return &pgdat->deferred_split_queue; | |
518 | } | |
519 | #endif | |
9a982250 KS |
520 | |
521 | void prep_transhuge_page(struct page *page) | |
522 | { | |
523 | /* | |
524 | * we use page->mapping and page->indexlru in second tail page | |
525 | * as list_head: assuming THP order >= 2 | |
526 | */ | |
9a982250 KS |
527 | |
528 | INIT_LIST_HEAD(page_deferred_list(page)); | |
529 | set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR); | |
530 | } | |
531 | ||
005ba37c SC |
532 | bool is_transparent_hugepage(struct page *page) |
533 | { | |
534 | if (!PageCompound(page)) | |
fa1f68cc | 535 | return false; |
005ba37c SC |
536 | |
537 | page = compound_head(page); | |
538 | return is_huge_zero_page(page) || | |
539 | page[1].compound_dtor == TRANSHUGE_PAGE_DTOR; | |
540 | } | |
541 | EXPORT_SYMBOL_GPL(is_transparent_hugepage); | |
542 | ||
97d3d0f9 KS |
543 | static unsigned long __thp_get_unmapped_area(struct file *filp, |
544 | unsigned long addr, unsigned long len, | |
74d2fad1 TK |
545 | loff_t off, unsigned long flags, unsigned long size) |
546 | { | |
74d2fad1 TK |
547 | loff_t off_end = off + len; |
548 | loff_t off_align = round_up(off, size); | |
97d3d0f9 | 549 | unsigned long len_pad, ret; |
74d2fad1 TK |
550 | |
551 | if (off_end <= off_align || (off_end - off_align) < size) | |
552 | return 0; | |
553 | ||
554 | len_pad = len + size; | |
555 | if (len_pad < len || (off + len_pad) < off) | |
556 | return 0; | |
557 | ||
97d3d0f9 | 558 | ret = current->mm->get_unmapped_area(filp, addr, len_pad, |
74d2fad1 | 559 | off >> PAGE_SHIFT, flags); |
97d3d0f9 KS |
560 | |
561 | /* | |
562 | * The failure might be due to length padding. The caller will retry | |
563 | * without the padding. | |
564 | */ | |
565 | if (IS_ERR_VALUE(ret)) | |
74d2fad1 TK |
566 | return 0; |
567 | ||
97d3d0f9 KS |
568 | /* |
569 | * Do not try to align to THP boundary if allocation at the address | |
570 | * hint succeeds. | |
571 | */ | |
572 | if (ret == addr) | |
573 | return addr; | |
574 | ||
575 | ret += (off - ret) & (size - 1); | |
576 | return ret; | |
74d2fad1 TK |
577 | } |
578 | ||
579 | unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr, | |
580 | unsigned long len, unsigned long pgoff, unsigned long flags) | |
581 | { | |
97d3d0f9 | 582 | unsigned long ret; |
74d2fad1 TK |
583 | loff_t off = (loff_t)pgoff << PAGE_SHIFT; |
584 | ||
74d2fad1 TK |
585 | if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD)) |
586 | goto out; | |
587 | ||
97d3d0f9 KS |
588 | ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE); |
589 | if (ret) | |
590 | return ret; | |
591 | out: | |
74d2fad1 TK |
592 | return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags); |
593 | } | |
594 | EXPORT_SYMBOL_GPL(thp_get_unmapped_area); | |
595 | ||
2b740303 SJ |
596 | static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf, |
597 | struct page *page, gfp_t gfp) | |
71e3aac0 | 598 | { |
82b0f8c3 | 599 | struct vm_area_struct *vma = vmf->vma; |
71e3aac0 | 600 | pgtable_t pgtable; |
82b0f8c3 | 601 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; |
2b740303 | 602 | vm_fault_t ret = 0; |
71e3aac0 | 603 | |
309381fe | 604 | VM_BUG_ON_PAGE(!PageCompound(page), page); |
00501b53 | 605 | |
d9eb1ea2 | 606 | if (mem_cgroup_charge(page, vma->vm_mm, gfp)) { |
6b251fc9 AA |
607 | put_page(page); |
608 | count_vm_event(THP_FAULT_FALLBACK); | |
85b9f46e | 609 | count_vm_event(THP_FAULT_FALLBACK_CHARGE); |
6b251fc9 AA |
610 | return VM_FAULT_FALLBACK; |
611 | } | |
9d82c694 | 612 | cgroup_throttle_swaprate(page, gfp); |
00501b53 | 613 | |
4cf58924 | 614 | pgtable = pte_alloc_one(vma->vm_mm); |
00501b53 | 615 | if (unlikely(!pgtable)) { |
6b31d595 MH |
616 | ret = VM_FAULT_OOM; |
617 | goto release; | |
00501b53 | 618 | } |
71e3aac0 | 619 | |
c79b57e4 | 620 | clear_huge_page(page, vmf->address, HPAGE_PMD_NR); |
52f37629 MK |
621 | /* |
622 | * The memory barrier inside __SetPageUptodate makes sure that | |
623 | * clear_huge_page writes become visible before the set_pmd_at() | |
624 | * write. | |
625 | */ | |
71e3aac0 AA |
626 | __SetPageUptodate(page); |
627 | ||
82b0f8c3 JK |
628 | vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); |
629 | if (unlikely(!pmd_none(*vmf->pmd))) { | |
6b31d595 | 630 | goto unlock_release; |
71e3aac0 AA |
631 | } else { |
632 | pmd_t entry; | |
6b251fc9 | 633 | |
6b31d595 MH |
634 | ret = check_stable_address_space(vma->vm_mm); |
635 | if (ret) | |
636 | goto unlock_release; | |
637 | ||
6b251fc9 AA |
638 | /* Deliver the page fault to userland */ |
639 | if (userfaultfd_missing(vma)) { | |
82b0f8c3 | 640 | spin_unlock(vmf->ptl); |
6b251fc9 | 641 | put_page(page); |
bae473a4 | 642 | pte_free(vma->vm_mm, pgtable); |
8fd5eda4 ML |
643 | ret = handle_userfault(vmf, VM_UFFD_MISSING); |
644 | VM_BUG_ON(ret & VM_FAULT_FALLBACK); | |
645 | return ret; | |
6b251fc9 AA |
646 | } |
647 | ||
3122359a | 648 | entry = mk_huge_pmd(page, vma->vm_page_prot); |
f55e1014 | 649 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
d281ee61 | 650 | page_add_new_anon_rmap(page, vma, haddr, true); |
b518154e | 651 | lru_cache_add_inactive_or_unevictable(page, vma); |
82b0f8c3 JK |
652 | pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable); |
653 | set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); | |
fca40573 | 654 | update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); |
bae473a4 | 655 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); |
c4812909 | 656 | mm_inc_nr_ptes(vma->vm_mm); |
82b0f8c3 | 657 | spin_unlock(vmf->ptl); |
6b251fc9 | 658 | count_vm_event(THP_FAULT_ALLOC); |
9d82c694 | 659 | count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC); |
71e3aac0 AA |
660 | } |
661 | ||
aa2e878e | 662 | return 0; |
6b31d595 MH |
663 | unlock_release: |
664 | spin_unlock(vmf->ptl); | |
665 | release: | |
666 | if (pgtable) | |
667 | pte_free(vma->vm_mm, pgtable); | |
6b31d595 MH |
668 | put_page(page); |
669 | return ret; | |
670 | ||
71e3aac0 AA |
671 | } |
672 | ||
444eb2a4 | 673 | /* |
21440d7e DR |
674 | * always: directly stall for all thp allocations |
675 | * defer: wake kswapd and fail if not immediately available | |
676 | * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise | |
677 | * fail if not immediately available | |
678 | * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately | |
679 | * available | |
680 | * never: never stall for any thp allocation | |
444eb2a4 | 681 | */ |
164cc4fe | 682 | gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma) |
444eb2a4 | 683 | { |
164cc4fe | 684 | const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE); |
2f0799a0 | 685 | |
ac79f78d | 686 | /* Always do synchronous compaction */ |
a8282608 AA |
687 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) |
688 | return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY); | |
ac79f78d DR |
689 | |
690 | /* Kick kcompactd and fail quickly */ | |
21440d7e | 691 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) |
19deb769 | 692 | return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM; |
ac79f78d DR |
693 | |
694 | /* Synchronous compaction if madvised, otherwise kick kcompactd */ | |
21440d7e | 695 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) |
19deb769 DR |
696 | return GFP_TRANSHUGE_LIGHT | |
697 | (vma_madvised ? __GFP_DIRECT_RECLAIM : | |
698 | __GFP_KSWAPD_RECLAIM); | |
ac79f78d DR |
699 | |
700 | /* Only do synchronous compaction if madvised */ | |
21440d7e | 701 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) |
19deb769 DR |
702 | return GFP_TRANSHUGE_LIGHT | |
703 | (vma_madvised ? __GFP_DIRECT_RECLAIM : 0); | |
ac79f78d | 704 | |
19deb769 | 705 | return GFP_TRANSHUGE_LIGHT; |
444eb2a4 MG |
706 | } |
707 | ||
c4088ebd | 708 | /* Caller must hold page table lock. */ |
2efeb8da | 709 | static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, |
97ae1749 | 710 | struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, |
5918d10a | 711 | struct page *zero_page) |
fc9fe822 KS |
712 | { |
713 | pmd_t entry; | |
7c414164 | 714 | if (!pmd_none(*pmd)) |
2efeb8da | 715 | return; |
5918d10a | 716 | entry = mk_pmd(zero_page, vma->vm_page_prot); |
fc9fe822 | 717 | entry = pmd_mkhuge(entry); |
12c9d70b MW |
718 | if (pgtable) |
719 | pgtable_trans_huge_deposit(mm, pmd, pgtable); | |
fc9fe822 | 720 | set_pmd_at(mm, haddr, pmd, entry); |
c4812909 | 721 | mm_inc_nr_ptes(mm); |
fc9fe822 KS |
722 | } |
723 | ||
2b740303 | 724 | vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf) |
71e3aac0 | 725 | { |
82b0f8c3 | 726 | struct vm_area_struct *vma = vmf->vma; |
077fcf11 | 727 | gfp_t gfp; |
71e3aac0 | 728 | struct page *page; |
82b0f8c3 | 729 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; |
71e3aac0 | 730 | |
43675e6f | 731 | if (!transhuge_vma_suitable(vma, haddr)) |
c0292554 | 732 | return VM_FAULT_FALLBACK; |
128ec037 KS |
733 | if (unlikely(anon_vma_prepare(vma))) |
734 | return VM_FAULT_OOM; | |
6d50e60c | 735 | if (unlikely(khugepaged_enter(vma, vma->vm_flags))) |
128ec037 | 736 | return VM_FAULT_OOM; |
82b0f8c3 | 737 | if (!(vmf->flags & FAULT_FLAG_WRITE) && |
bae473a4 | 738 | !mm_forbids_zeropage(vma->vm_mm) && |
128ec037 KS |
739 | transparent_hugepage_use_zero_page()) { |
740 | pgtable_t pgtable; | |
741 | struct page *zero_page; | |
2b740303 | 742 | vm_fault_t ret; |
4cf58924 | 743 | pgtable = pte_alloc_one(vma->vm_mm); |
128ec037 | 744 | if (unlikely(!pgtable)) |
ba76149f | 745 | return VM_FAULT_OOM; |
6fcb52a5 | 746 | zero_page = mm_get_huge_zero_page(vma->vm_mm); |
128ec037 | 747 | if (unlikely(!zero_page)) { |
bae473a4 | 748 | pte_free(vma->vm_mm, pgtable); |
81ab4201 | 749 | count_vm_event(THP_FAULT_FALLBACK); |
c0292554 | 750 | return VM_FAULT_FALLBACK; |
b9bbfbe3 | 751 | } |
82b0f8c3 | 752 | vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); |
6b251fc9 | 753 | ret = 0; |
82b0f8c3 | 754 | if (pmd_none(*vmf->pmd)) { |
6b31d595 MH |
755 | ret = check_stable_address_space(vma->vm_mm); |
756 | if (ret) { | |
757 | spin_unlock(vmf->ptl); | |
bfe8cc1d | 758 | pte_free(vma->vm_mm, pgtable); |
6b31d595 | 759 | } else if (userfaultfd_missing(vma)) { |
82b0f8c3 | 760 | spin_unlock(vmf->ptl); |
bfe8cc1d | 761 | pte_free(vma->vm_mm, pgtable); |
82b0f8c3 | 762 | ret = handle_userfault(vmf, VM_UFFD_MISSING); |
6b251fc9 AA |
763 | VM_BUG_ON(ret & VM_FAULT_FALLBACK); |
764 | } else { | |
bae473a4 | 765 | set_huge_zero_page(pgtable, vma->vm_mm, vma, |
82b0f8c3 | 766 | haddr, vmf->pmd, zero_page); |
fca40573 | 767 | update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); |
82b0f8c3 | 768 | spin_unlock(vmf->ptl); |
6b251fc9 | 769 | } |
bfe8cc1d | 770 | } else { |
82b0f8c3 | 771 | spin_unlock(vmf->ptl); |
bae473a4 | 772 | pte_free(vma->vm_mm, pgtable); |
bfe8cc1d | 773 | } |
6b251fc9 | 774 | return ret; |
71e3aac0 | 775 | } |
164cc4fe | 776 | gfp = vma_thp_gfp_mask(vma); |
19deb769 | 777 | page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); |
128ec037 KS |
778 | if (unlikely(!page)) { |
779 | count_vm_event(THP_FAULT_FALLBACK); | |
c0292554 | 780 | return VM_FAULT_FALLBACK; |
128ec037 | 781 | } |
9a982250 | 782 | prep_transhuge_page(page); |
82b0f8c3 | 783 | return __do_huge_pmd_anonymous_page(vmf, page, gfp); |
71e3aac0 AA |
784 | } |
785 | ||
ae18d6dc | 786 | static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, |
3b6521f5 OH |
787 | pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write, |
788 | pgtable_t pgtable) | |
5cad465d MW |
789 | { |
790 | struct mm_struct *mm = vma->vm_mm; | |
791 | pmd_t entry; | |
792 | spinlock_t *ptl; | |
793 | ||
794 | ptl = pmd_lock(mm, pmd); | |
c6f3c5ee AK |
795 | if (!pmd_none(*pmd)) { |
796 | if (write) { | |
797 | if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) { | |
798 | WARN_ON_ONCE(!is_huge_zero_pmd(*pmd)); | |
799 | goto out_unlock; | |
800 | } | |
801 | entry = pmd_mkyoung(*pmd); | |
802 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
803 | if (pmdp_set_access_flags(vma, addr, pmd, entry, 1)) | |
804 | update_mmu_cache_pmd(vma, addr, pmd); | |
805 | } | |
806 | ||
807 | goto out_unlock; | |
808 | } | |
809 | ||
f25748e3 DW |
810 | entry = pmd_mkhuge(pfn_t_pmd(pfn, prot)); |
811 | if (pfn_t_devmap(pfn)) | |
812 | entry = pmd_mkdevmap(entry); | |
01871e59 | 813 | if (write) { |
f55e1014 LT |
814 | entry = pmd_mkyoung(pmd_mkdirty(entry)); |
815 | entry = maybe_pmd_mkwrite(entry, vma); | |
5cad465d | 816 | } |
3b6521f5 OH |
817 | |
818 | if (pgtable) { | |
819 | pgtable_trans_huge_deposit(mm, pmd, pgtable); | |
c4812909 | 820 | mm_inc_nr_ptes(mm); |
c6f3c5ee | 821 | pgtable = NULL; |
3b6521f5 OH |
822 | } |
823 | ||
01871e59 RZ |
824 | set_pmd_at(mm, addr, pmd, entry); |
825 | update_mmu_cache_pmd(vma, addr, pmd); | |
c6f3c5ee AK |
826 | |
827 | out_unlock: | |
5cad465d | 828 | spin_unlock(ptl); |
c6f3c5ee AK |
829 | if (pgtable) |
830 | pte_free(mm, pgtable); | |
5cad465d MW |
831 | } |
832 | ||
9a9731b1 THV |
833 | /** |
834 | * vmf_insert_pfn_pmd_prot - insert a pmd size pfn | |
835 | * @vmf: Structure describing the fault | |
836 | * @pfn: pfn to insert | |
837 | * @pgprot: page protection to use | |
838 | * @write: whether it's a write fault | |
839 | * | |
840 | * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and | |
841 | * also consult the vmf_insert_mixed_prot() documentation when | |
842 | * @pgprot != @vmf->vma->vm_page_prot. | |
843 | * | |
844 | * Return: vm_fault_t value. | |
845 | */ | |
846 | vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn, | |
847 | pgprot_t pgprot, bool write) | |
5cad465d | 848 | { |
fce86ff5 DW |
849 | unsigned long addr = vmf->address & PMD_MASK; |
850 | struct vm_area_struct *vma = vmf->vma; | |
3b6521f5 | 851 | pgtable_t pgtable = NULL; |
fce86ff5 | 852 | |
5cad465d MW |
853 | /* |
854 | * If we had pmd_special, we could avoid all these restrictions, | |
855 | * but we need to be consistent with PTEs and architectures that | |
856 | * can't support a 'special' bit. | |
857 | */ | |
e1fb4a08 DJ |
858 | BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && |
859 | !pfn_t_devmap(pfn)); | |
5cad465d MW |
860 | BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == |
861 | (VM_PFNMAP|VM_MIXEDMAP)); | |
862 | BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); | |
5cad465d MW |
863 | |
864 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
865 | return VM_FAULT_SIGBUS; | |
308a047c | 866 | |
3b6521f5 | 867 | if (arch_needs_pgtable_deposit()) { |
4cf58924 | 868 | pgtable = pte_alloc_one(vma->vm_mm); |
3b6521f5 OH |
869 | if (!pgtable) |
870 | return VM_FAULT_OOM; | |
871 | } | |
872 | ||
308a047c BP |
873 | track_pfn_insert(vma, &pgprot, pfn); |
874 | ||
fce86ff5 | 875 | insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable); |
ae18d6dc | 876 | return VM_FAULT_NOPAGE; |
5cad465d | 877 | } |
9a9731b1 | 878 | EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot); |
5cad465d | 879 | |
a00cc7d9 | 880 | #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD |
f55e1014 | 881 | static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma) |
a00cc7d9 | 882 | { |
f55e1014 | 883 | if (likely(vma->vm_flags & VM_WRITE)) |
a00cc7d9 MW |
884 | pud = pud_mkwrite(pud); |
885 | return pud; | |
886 | } | |
887 | ||
888 | static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, | |
889 | pud_t *pud, pfn_t pfn, pgprot_t prot, bool write) | |
890 | { | |
891 | struct mm_struct *mm = vma->vm_mm; | |
892 | pud_t entry; | |
893 | spinlock_t *ptl; | |
894 | ||
895 | ptl = pud_lock(mm, pud); | |
c6f3c5ee AK |
896 | if (!pud_none(*pud)) { |
897 | if (write) { | |
898 | if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) { | |
899 | WARN_ON_ONCE(!is_huge_zero_pud(*pud)); | |
900 | goto out_unlock; | |
901 | } | |
902 | entry = pud_mkyoung(*pud); | |
903 | entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma); | |
904 | if (pudp_set_access_flags(vma, addr, pud, entry, 1)) | |
905 | update_mmu_cache_pud(vma, addr, pud); | |
906 | } | |
907 | goto out_unlock; | |
908 | } | |
909 | ||
a00cc7d9 MW |
910 | entry = pud_mkhuge(pfn_t_pud(pfn, prot)); |
911 | if (pfn_t_devmap(pfn)) | |
912 | entry = pud_mkdevmap(entry); | |
913 | if (write) { | |
f55e1014 LT |
914 | entry = pud_mkyoung(pud_mkdirty(entry)); |
915 | entry = maybe_pud_mkwrite(entry, vma); | |
a00cc7d9 MW |
916 | } |
917 | set_pud_at(mm, addr, pud, entry); | |
918 | update_mmu_cache_pud(vma, addr, pud); | |
c6f3c5ee AK |
919 | |
920 | out_unlock: | |
a00cc7d9 MW |
921 | spin_unlock(ptl); |
922 | } | |
923 | ||
9a9731b1 THV |
924 | /** |
925 | * vmf_insert_pfn_pud_prot - insert a pud size pfn | |
926 | * @vmf: Structure describing the fault | |
927 | * @pfn: pfn to insert | |
928 | * @pgprot: page protection to use | |
929 | * @write: whether it's a write fault | |
930 | * | |
931 | * Insert a pud size pfn. See vmf_insert_pfn() for additional info and | |
932 | * also consult the vmf_insert_mixed_prot() documentation when | |
933 | * @pgprot != @vmf->vma->vm_page_prot. | |
934 | * | |
935 | * Return: vm_fault_t value. | |
936 | */ | |
937 | vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn, | |
938 | pgprot_t pgprot, bool write) | |
a00cc7d9 | 939 | { |
fce86ff5 DW |
940 | unsigned long addr = vmf->address & PUD_MASK; |
941 | struct vm_area_struct *vma = vmf->vma; | |
fce86ff5 | 942 | |
a00cc7d9 MW |
943 | /* |
944 | * If we had pud_special, we could avoid all these restrictions, | |
945 | * but we need to be consistent with PTEs and architectures that | |
946 | * can't support a 'special' bit. | |
947 | */ | |
62ec0d8c DJ |
948 | BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) && |
949 | !pfn_t_devmap(pfn)); | |
a00cc7d9 MW |
950 | BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == |
951 | (VM_PFNMAP|VM_MIXEDMAP)); | |
952 | BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); | |
a00cc7d9 MW |
953 | |
954 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
955 | return VM_FAULT_SIGBUS; | |
956 | ||
957 | track_pfn_insert(vma, &pgprot, pfn); | |
958 | ||
fce86ff5 | 959 | insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write); |
a00cc7d9 MW |
960 | return VM_FAULT_NOPAGE; |
961 | } | |
9a9731b1 | 962 | EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot); |
a00cc7d9 MW |
963 | #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ |
964 | ||
3565fce3 | 965 | static void touch_pmd(struct vm_area_struct *vma, unsigned long addr, |
a8f97366 | 966 | pmd_t *pmd, int flags) |
3565fce3 DW |
967 | { |
968 | pmd_t _pmd; | |
969 | ||
a8f97366 KS |
970 | _pmd = pmd_mkyoung(*pmd); |
971 | if (flags & FOLL_WRITE) | |
972 | _pmd = pmd_mkdirty(_pmd); | |
3565fce3 | 973 | if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, |
a8f97366 | 974 | pmd, _pmd, flags & FOLL_WRITE)) |
3565fce3 DW |
975 | update_mmu_cache_pmd(vma, addr, pmd); |
976 | } | |
977 | ||
978 | struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr, | |
df06b37f | 979 | pmd_t *pmd, int flags, struct dev_pagemap **pgmap) |
3565fce3 DW |
980 | { |
981 | unsigned long pfn = pmd_pfn(*pmd); | |
982 | struct mm_struct *mm = vma->vm_mm; | |
3565fce3 DW |
983 | struct page *page; |
984 | ||
985 | assert_spin_locked(pmd_lockptr(mm, pmd)); | |
986 | ||
8310d48b KF |
987 | /* |
988 | * When we COW a devmap PMD entry, we split it into PTEs, so we should | |
989 | * not be in this function with `flags & FOLL_COW` set. | |
990 | */ | |
991 | WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set"); | |
992 | ||
3faa52c0 JH |
993 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ |
994 | if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == | |
995 | (FOLL_PIN | FOLL_GET))) | |
996 | return NULL; | |
997 | ||
f6f37321 | 998 | if (flags & FOLL_WRITE && !pmd_write(*pmd)) |
3565fce3 DW |
999 | return NULL; |
1000 | ||
1001 | if (pmd_present(*pmd) && pmd_devmap(*pmd)) | |
1002 | /* pass */; | |
1003 | else | |
1004 | return NULL; | |
1005 | ||
1006 | if (flags & FOLL_TOUCH) | |
a8f97366 | 1007 | touch_pmd(vma, addr, pmd, flags); |
3565fce3 DW |
1008 | |
1009 | /* | |
1010 | * device mapped pages can only be returned if the | |
1011 | * caller will manage the page reference count. | |
1012 | */ | |
3faa52c0 | 1013 | if (!(flags & (FOLL_GET | FOLL_PIN))) |
3565fce3 DW |
1014 | return ERR_PTR(-EEXIST); |
1015 | ||
1016 | pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT; | |
df06b37f KB |
1017 | *pgmap = get_dev_pagemap(pfn, *pgmap); |
1018 | if (!*pgmap) | |
3565fce3 DW |
1019 | return ERR_PTR(-EFAULT); |
1020 | page = pfn_to_page(pfn); | |
3faa52c0 JH |
1021 | if (!try_grab_page(page, flags)) |
1022 | page = ERR_PTR(-ENOMEM); | |
3565fce3 DW |
1023 | |
1024 | return page; | |
1025 | } | |
1026 | ||
71e3aac0 AA |
1027 | int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, |
1028 | pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, | |
1029 | struct vm_area_struct *vma) | |
1030 | { | |
c4088ebd | 1031 | spinlock_t *dst_ptl, *src_ptl; |
71e3aac0 AA |
1032 | struct page *src_page; |
1033 | pmd_t pmd; | |
12c9d70b | 1034 | pgtable_t pgtable = NULL; |
628d47ce | 1035 | int ret = -ENOMEM; |
71e3aac0 | 1036 | |
628d47ce KS |
1037 | /* Skip if can be re-fill on fault */ |
1038 | if (!vma_is_anonymous(vma)) | |
1039 | return 0; | |
1040 | ||
4cf58924 | 1041 | pgtable = pte_alloc_one(dst_mm); |
628d47ce KS |
1042 | if (unlikely(!pgtable)) |
1043 | goto out; | |
71e3aac0 | 1044 | |
c4088ebd KS |
1045 | dst_ptl = pmd_lock(dst_mm, dst_pmd); |
1046 | src_ptl = pmd_lockptr(src_mm, src_pmd); | |
1047 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
71e3aac0 AA |
1048 | |
1049 | ret = -EAGAIN; | |
1050 | pmd = *src_pmd; | |
84c3fc4e | 1051 | |
b569a176 PX |
1052 | /* |
1053 | * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA | |
1054 | * does not have the VM_UFFD_WP, which means that the uffd | |
1055 | * fork event is not enabled. | |
1056 | */ | |
1057 | if (!(vma->vm_flags & VM_UFFD_WP)) | |
1058 | pmd = pmd_clear_uffd_wp(pmd); | |
1059 | ||
84c3fc4e ZY |
1060 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
1061 | if (unlikely(is_swap_pmd(pmd))) { | |
1062 | swp_entry_t entry = pmd_to_swp_entry(pmd); | |
1063 | ||
1064 | VM_BUG_ON(!is_pmd_migration_entry(pmd)); | |
1065 | if (is_write_migration_entry(entry)) { | |
1066 | make_migration_entry_read(&entry); | |
1067 | pmd = swp_entry_to_pmd(entry); | |
ab6e3d09 NH |
1068 | if (pmd_swp_soft_dirty(*src_pmd)) |
1069 | pmd = pmd_swp_mksoft_dirty(pmd); | |
84c3fc4e ZY |
1070 | set_pmd_at(src_mm, addr, src_pmd, pmd); |
1071 | } | |
dd8a67f9 | 1072 | add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); |
af5b0f6a | 1073 | mm_inc_nr_ptes(dst_mm); |
dd8a67f9 | 1074 | pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); |
84c3fc4e ZY |
1075 | set_pmd_at(dst_mm, addr, dst_pmd, pmd); |
1076 | ret = 0; | |
1077 | goto out_unlock; | |
1078 | } | |
1079 | #endif | |
1080 | ||
628d47ce | 1081 | if (unlikely(!pmd_trans_huge(pmd))) { |
71e3aac0 AA |
1082 | pte_free(dst_mm, pgtable); |
1083 | goto out_unlock; | |
1084 | } | |
fc9fe822 | 1085 | /* |
c4088ebd | 1086 | * When page table lock is held, the huge zero pmd should not be |
fc9fe822 KS |
1087 | * under splitting since we don't split the page itself, only pmd to |
1088 | * a page table. | |
1089 | */ | |
1090 | if (is_huge_zero_pmd(pmd)) { | |
97ae1749 KS |
1091 | /* |
1092 | * get_huge_zero_page() will never allocate a new page here, | |
1093 | * since we already have a zero page to copy. It just takes a | |
1094 | * reference. | |
1095 | */ | |
5fc7a5f6 PX |
1096 | mm_get_huge_zero_page(dst_mm); |
1097 | goto out_zero_page; | |
fc9fe822 | 1098 | } |
de466bd6 | 1099 | |
628d47ce KS |
1100 | src_page = pmd_page(pmd); |
1101 | VM_BUG_ON_PAGE(!PageHead(src_page), src_page); | |
d042035e PX |
1102 | |
1103 | /* | |
1104 | * If this page is a potentially pinned page, split and retry the fault | |
1105 | * with smaller page size. Normally this should not happen because the | |
1106 | * userspace should use MADV_DONTFORK upon pinned regions. This is a | |
1107 | * best effort that the pinned pages won't be replaced by another | |
1108 | * random page during the coming copy-on-write. | |
1109 | */ | |
97a7e473 | 1110 | if (unlikely(page_needs_cow_for_dma(vma, src_page))) { |
d042035e PX |
1111 | pte_free(dst_mm, pgtable); |
1112 | spin_unlock(src_ptl); | |
1113 | spin_unlock(dst_ptl); | |
1114 | __split_huge_pmd(vma, src_pmd, addr, false, NULL); | |
1115 | return -EAGAIN; | |
1116 | } | |
1117 | ||
628d47ce KS |
1118 | get_page(src_page); |
1119 | page_dup_rmap(src_page, true); | |
1120 | add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); | |
5fc7a5f6 | 1121 | out_zero_page: |
c4812909 | 1122 | mm_inc_nr_ptes(dst_mm); |
628d47ce | 1123 | pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); |
71e3aac0 AA |
1124 | |
1125 | pmdp_set_wrprotect(src_mm, addr, src_pmd); | |
1126 | pmd = pmd_mkold(pmd_wrprotect(pmd)); | |
1127 | set_pmd_at(dst_mm, addr, dst_pmd, pmd); | |
71e3aac0 AA |
1128 | |
1129 | ret = 0; | |
1130 | out_unlock: | |
c4088ebd KS |
1131 | spin_unlock(src_ptl); |
1132 | spin_unlock(dst_ptl); | |
71e3aac0 AA |
1133 | out: |
1134 | return ret; | |
1135 | } | |
1136 | ||
a00cc7d9 MW |
1137 | #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD |
1138 | static void touch_pud(struct vm_area_struct *vma, unsigned long addr, | |
a8f97366 | 1139 | pud_t *pud, int flags) |
a00cc7d9 MW |
1140 | { |
1141 | pud_t _pud; | |
1142 | ||
a8f97366 KS |
1143 | _pud = pud_mkyoung(*pud); |
1144 | if (flags & FOLL_WRITE) | |
1145 | _pud = pud_mkdirty(_pud); | |
a00cc7d9 | 1146 | if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK, |
a8f97366 | 1147 | pud, _pud, flags & FOLL_WRITE)) |
a00cc7d9 MW |
1148 | update_mmu_cache_pud(vma, addr, pud); |
1149 | } | |
1150 | ||
1151 | struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr, | |
df06b37f | 1152 | pud_t *pud, int flags, struct dev_pagemap **pgmap) |
a00cc7d9 MW |
1153 | { |
1154 | unsigned long pfn = pud_pfn(*pud); | |
1155 | struct mm_struct *mm = vma->vm_mm; | |
a00cc7d9 MW |
1156 | struct page *page; |
1157 | ||
1158 | assert_spin_locked(pud_lockptr(mm, pud)); | |
1159 | ||
f6f37321 | 1160 | if (flags & FOLL_WRITE && !pud_write(*pud)) |
a00cc7d9 MW |
1161 | return NULL; |
1162 | ||
3faa52c0 JH |
1163 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ |
1164 | if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == | |
1165 | (FOLL_PIN | FOLL_GET))) | |
1166 | return NULL; | |
1167 | ||
a00cc7d9 MW |
1168 | if (pud_present(*pud) && pud_devmap(*pud)) |
1169 | /* pass */; | |
1170 | else | |
1171 | return NULL; | |
1172 | ||
1173 | if (flags & FOLL_TOUCH) | |
a8f97366 | 1174 | touch_pud(vma, addr, pud, flags); |
a00cc7d9 MW |
1175 | |
1176 | /* | |
1177 | * device mapped pages can only be returned if the | |
1178 | * caller will manage the page reference count. | |
3faa52c0 JH |
1179 | * |
1180 | * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here: | |
a00cc7d9 | 1181 | */ |
3faa52c0 | 1182 | if (!(flags & (FOLL_GET | FOLL_PIN))) |
a00cc7d9 MW |
1183 | return ERR_PTR(-EEXIST); |
1184 | ||
1185 | pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT; | |
df06b37f KB |
1186 | *pgmap = get_dev_pagemap(pfn, *pgmap); |
1187 | if (!*pgmap) | |
a00cc7d9 MW |
1188 | return ERR_PTR(-EFAULT); |
1189 | page = pfn_to_page(pfn); | |
3faa52c0 JH |
1190 | if (!try_grab_page(page, flags)) |
1191 | page = ERR_PTR(-ENOMEM); | |
a00cc7d9 MW |
1192 | |
1193 | return page; | |
1194 | } | |
1195 | ||
1196 | int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm, | |
1197 | pud_t *dst_pud, pud_t *src_pud, unsigned long addr, | |
1198 | struct vm_area_struct *vma) | |
1199 | { | |
1200 | spinlock_t *dst_ptl, *src_ptl; | |
1201 | pud_t pud; | |
1202 | int ret; | |
1203 | ||
1204 | dst_ptl = pud_lock(dst_mm, dst_pud); | |
1205 | src_ptl = pud_lockptr(src_mm, src_pud); | |
1206 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
1207 | ||
1208 | ret = -EAGAIN; | |
1209 | pud = *src_pud; | |
1210 | if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud))) | |
1211 | goto out_unlock; | |
1212 | ||
1213 | /* | |
1214 | * When page table lock is held, the huge zero pud should not be | |
1215 | * under splitting since we don't split the page itself, only pud to | |
1216 | * a page table. | |
1217 | */ | |
1218 | if (is_huge_zero_pud(pud)) { | |
1219 | /* No huge zero pud yet */ | |
1220 | } | |
1221 | ||
d042035e | 1222 | /* Please refer to comments in copy_huge_pmd() */ |
97a7e473 | 1223 | if (unlikely(page_needs_cow_for_dma(vma, pud_page(pud)))) { |
d042035e PX |
1224 | spin_unlock(src_ptl); |
1225 | spin_unlock(dst_ptl); | |
1226 | __split_huge_pud(vma, src_pud, addr); | |
1227 | return -EAGAIN; | |
1228 | } | |
1229 | ||
a00cc7d9 MW |
1230 | pudp_set_wrprotect(src_mm, addr, src_pud); |
1231 | pud = pud_mkold(pud_wrprotect(pud)); | |
1232 | set_pud_at(dst_mm, addr, dst_pud, pud); | |
1233 | ||
1234 | ret = 0; | |
1235 | out_unlock: | |
1236 | spin_unlock(src_ptl); | |
1237 | spin_unlock(dst_ptl); | |
1238 | return ret; | |
1239 | } | |
1240 | ||
1241 | void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud) | |
1242 | { | |
1243 | pud_t entry; | |
1244 | unsigned long haddr; | |
1245 | bool write = vmf->flags & FAULT_FLAG_WRITE; | |
1246 | ||
1247 | vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud); | |
1248 | if (unlikely(!pud_same(*vmf->pud, orig_pud))) | |
1249 | goto unlock; | |
1250 | ||
1251 | entry = pud_mkyoung(orig_pud); | |
1252 | if (write) | |
1253 | entry = pud_mkdirty(entry); | |
1254 | haddr = vmf->address & HPAGE_PUD_MASK; | |
1255 | if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write)) | |
1256 | update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud); | |
1257 | ||
1258 | unlock: | |
1259 | spin_unlock(vmf->ptl); | |
1260 | } | |
1261 | #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ | |
1262 | ||
82b0f8c3 | 1263 | void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd) |
a1dd450b WD |
1264 | { |
1265 | pmd_t entry; | |
1266 | unsigned long haddr; | |
20f664aa | 1267 | bool write = vmf->flags & FAULT_FLAG_WRITE; |
a1dd450b | 1268 | |
82b0f8c3 JK |
1269 | vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); |
1270 | if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) | |
a1dd450b WD |
1271 | goto unlock; |
1272 | ||
1273 | entry = pmd_mkyoung(orig_pmd); | |
20f664aa MK |
1274 | if (write) |
1275 | entry = pmd_mkdirty(entry); | |
82b0f8c3 | 1276 | haddr = vmf->address & HPAGE_PMD_MASK; |
20f664aa | 1277 | if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write)) |
82b0f8c3 | 1278 | update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd); |
a1dd450b WD |
1279 | |
1280 | unlock: | |
82b0f8c3 | 1281 | spin_unlock(vmf->ptl); |
a1dd450b WD |
1282 | } |
1283 | ||
2b740303 | 1284 | vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd) |
71e3aac0 | 1285 | { |
82b0f8c3 | 1286 | struct vm_area_struct *vma = vmf->vma; |
3917c802 | 1287 | struct page *page; |
82b0f8c3 | 1288 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; |
71e3aac0 | 1289 | |
82b0f8c3 | 1290 | vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd); |
81d1b09c | 1291 | VM_BUG_ON_VMA(!vma->anon_vma, vma); |
3917c802 | 1292 | |
93b4796d | 1293 | if (is_huge_zero_pmd(orig_pmd)) |
3917c802 KS |
1294 | goto fallback; |
1295 | ||
82b0f8c3 | 1296 | spin_lock(vmf->ptl); |
3917c802 KS |
1297 | |
1298 | if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { | |
1299 | spin_unlock(vmf->ptl); | |
1300 | return 0; | |
1301 | } | |
71e3aac0 AA |
1302 | |
1303 | page = pmd_page(orig_pmd); | |
f6004e73 | 1304 | VM_BUG_ON_PAGE(!PageHead(page), page); |
3917c802 KS |
1305 | |
1306 | /* Lock page for reuse_swap_page() */ | |
ba3c4ce6 HY |
1307 | if (!trylock_page(page)) { |
1308 | get_page(page); | |
1309 | spin_unlock(vmf->ptl); | |
1310 | lock_page(page); | |
1311 | spin_lock(vmf->ptl); | |
1312 | if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { | |
3917c802 | 1313 | spin_unlock(vmf->ptl); |
ba3c4ce6 HY |
1314 | unlock_page(page); |
1315 | put_page(page); | |
3917c802 | 1316 | return 0; |
ba3c4ce6 HY |
1317 | } |
1318 | put_page(page); | |
1319 | } | |
3917c802 KS |
1320 | |
1321 | /* | |
1322 | * We can only reuse the page if nobody else maps the huge page or it's | |
1323 | * part. | |
1324 | */ | |
ba3c4ce6 | 1325 | if (reuse_swap_page(page, NULL)) { |
71e3aac0 AA |
1326 | pmd_t entry; |
1327 | entry = pmd_mkyoung(orig_pmd); | |
f55e1014 | 1328 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); |
3917c802 | 1329 | if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1)) |
82b0f8c3 | 1330 | update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); |
ba3c4ce6 | 1331 | unlock_page(page); |
82b0f8c3 | 1332 | spin_unlock(vmf->ptl); |
3917c802 | 1333 | return VM_FAULT_WRITE; |
71e3aac0 | 1334 | } |
3917c802 KS |
1335 | |
1336 | unlock_page(page); | |
82b0f8c3 | 1337 | spin_unlock(vmf->ptl); |
3917c802 KS |
1338 | fallback: |
1339 | __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL); | |
1340 | return VM_FAULT_FALLBACK; | |
71e3aac0 AA |
1341 | } |
1342 | ||
8310d48b | 1343 | /* |
a308c71b PX |
1344 | * FOLL_FORCE can write to even unwritable pmd's, but only |
1345 | * after we've gone through a COW cycle and they are dirty. | |
8310d48b KF |
1346 | */ |
1347 | static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags) | |
1348 | { | |
a308c71b PX |
1349 | return pmd_write(pmd) || |
1350 | ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd)); | |
8310d48b KF |
1351 | } |
1352 | ||
b676b293 | 1353 | struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, |
71e3aac0 AA |
1354 | unsigned long addr, |
1355 | pmd_t *pmd, | |
1356 | unsigned int flags) | |
1357 | { | |
b676b293 | 1358 | struct mm_struct *mm = vma->vm_mm; |
71e3aac0 AA |
1359 | struct page *page = NULL; |
1360 | ||
c4088ebd | 1361 | assert_spin_locked(pmd_lockptr(mm, pmd)); |
71e3aac0 | 1362 | |
8310d48b | 1363 | if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags)) |
71e3aac0 AA |
1364 | goto out; |
1365 | ||
85facf25 KS |
1366 | /* Avoid dumping huge zero page */ |
1367 | if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) | |
1368 | return ERR_PTR(-EFAULT); | |
1369 | ||
2b4847e7 | 1370 | /* Full NUMA hinting faults to serialise migration in fault paths */ |
8a0516ed | 1371 | if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) |
2b4847e7 MG |
1372 | goto out; |
1373 | ||
71e3aac0 | 1374 | page = pmd_page(*pmd); |
ca120cf6 | 1375 | VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page); |
3faa52c0 JH |
1376 | |
1377 | if (!try_grab_page(page, flags)) | |
1378 | return ERR_PTR(-ENOMEM); | |
1379 | ||
3565fce3 | 1380 | if (flags & FOLL_TOUCH) |
a8f97366 | 1381 | touch_pmd(vma, addr, pmd, flags); |
3faa52c0 | 1382 | |
de60f5f1 | 1383 | if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { |
e90309c9 KS |
1384 | /* |
1385 | * We don't mlock() pte-mapped THPs. This way we can avoid | |
1386 | * leaking mlocked pages into non-VM_LOCKED VMAs. | |
1387 | * | |
9a73f61b KS |
1388 | * For anon THP: |
1389 | * | |
e90309c9 KS |
1390 | * In most cases the pmd is the only mapping of the page as we |
1391 | * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for | |
1392 | * writable private mappings in populate_vma_page_range(). | |
1393 | * | |
1394 | * The only scenario when we have the page shared here is if we | |
1395 | * mlocking read-only mapping shared over fork(). We skip | |
1396 | * mlocking such pages. | |
9a73f61b KS |
1397 | * |
1398 | * For file THP: | |
1399 | * | |
1400 | * We can expect PageDoubleMap() to be stable under page lock: | |
1401 | * for file pages we set it in page_add_file_rmap(), which | |
1402 | * requires page to be locked. | |
e90309c9 | 1403 | */ |
9a73f61b KS |
1404 | |
1405 | if (PageAnon(page) && compound_mapcount(page) != 1) | |
1406 | goto skip_mlock; | |
1407 | if (PageDoubleMap(page) || !page->mapping) | |
1408 | goto skip_mlock; | |
1409 | if (!trylock_page(page)) | |
1410 | goto skip_mlock; | |
9a73f61b KS |
1411 | if (page->mapping && !PageDoubleMap(page)) |
1412 | mlock_vma_page(page); | |
1413 | unlock_page(page); | |
b676b293 | 1414 | } |
9a73f61b | 1415 | skip_mlock: |
71e3aac0 | 1416 | page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; |
ca120cf6 | 1417 | VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page); |
71e3aac0 AA |
1418 | |
1419 | out: | |
1420 | return page; | |
1421 | } | |
1422 | ||
d10e63f2 | 1423 | /* NUMA hinting page fault entry point for trans huge pmds */ |
2b740303 | 1424 | vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd) |
d10e63f2 | 1425 | { |
82b0f8c3 | 1426 | struct vm_area_struct *vma = vmf->vma; |
b8916634 | 1427 | struct anon_vma *anon_vma = NULL; |
b32967ff | 1428 | struct page *page; |
82b0f8c3 | 1429 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; |
98fa15f3 | 1430 | int page_nid = NUMA_NO_NODE, this_nid = numa_node_id(); |
90572890 | 1431 | int target_nid, last_cpupid = -1; |
8191acbd MG |
1432 | bool page_locked; |
1433 | bool migrated = false; | |
b191f9b1 | 1434 | bool was_writable; |
6688cc05 | 1435 | int flags = 0; |
d10e63f2 | 1436 | |
82b0f8c3 JK |
1437 | vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); |
1438 | if (unlikely(!pmd_same(pmd, *vmf->pmd))) | |
d10e63f2 MG |
1439 | goto out_unlock; |
1440 | ||
de466bd6 MG |
1441 | /* |
1442 | * If there are potential migrations, wait for completion and retry | |
1443 | * without disrupting NUMA hinting information. Do not relock and | |
1444 | * check_same as the page may no longer be mapped. | |
1445 | */ | |
82b0f8c3 JK |
1446 | if (unlikely(pmd_trans_migrating(*vmf->pmd))) { |
1447 | page = pmd_page(*vmf->pmd); | |
3c226c63 MR |
1448 | if (!get_page_unless_zero(page)) |
1449 | goto out_unlock; | |
82b0f8c3 | 1450 | spin_unlock(vmf->ptl); |
48054625 | 1451 | put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE); |
de466bd6 MG |
1452 | goto out; |
1453 | } | |
1454 | ||
d10e63f2 | 1455 | page = pmd_page(pmd); |
a1a46184 | 1456 | BUG_ON(is_huge_zero_page(page)); |
8191acbd | 1457 | page_nid = page_to_nid(page); |
90572890 | 1458 | last_cpupid = page_cpupid_last(page); |
03c5a6e1 | 1459 | count_vm_numa_event(NUMA_HINT_FAULTS); |
04bb2f94 | 1460 | if (page_nid == this_nid) { |
03c5a6e1 | 1461 | count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); |
04bb2f94 RR |
1462 | flags |= TNF_FAULT_LOCAL; |
1463 | } | |
4daae3b4 | 1464 | |
bea66fbd | 1465 | /* See similar comment in do_numa_page for explanation */ |
288bc549 | 1466 | if (!pmd_savedwrite(pmd)) |
6688cc05 PZ |
1467 | flags |= TNF_NO_GROUP; |
1468 | ||
ff9042b1 MG |
1469 | /* |
1470 | * Acquire the page lock to serialise THP migrations but avoid dropping | |
1471 | * page_table_lock if at all possible | |
1472 | */ | |
b8916634 MG |
1473 | page_locked = trylock_page(page); |
1474 | target_nid = mpol_misplaced(page, vma, haddr); | |
de466bd6 | 1475 | /* Migration could have started since the pmd_trans_migrating check */ |
2b4847e7 | 1476 | if (!page_locked) { |
98fa15f3 | 1477 | page_nid = NUMA_NO_NODE; |
3c226c63 MR |
1478 | if (!get_page_unless_zero(page)) |
1479 | goto out_unlock; | |
82b0f8c3 | 1480 | spin_unlock(vmf->ptl); |
48054625 | 1481 | put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE); |
b8916634 | 1482 | goto out; |
6beb5e8b ML |
1483 | } else if (target_nid == NUMA_NO_NODE) { |
1484 | /* There are no parallel migrations and page is in the right | |
1485 | * node. Clear the numa hinting info in this pmd. | |
1486 | */ | |
1487 | goto clear_pmdnuma; | |
b8916634 MG |
1488 | } |
1489 | ||
2b4847e7 MG |
1490 | /* |
1491 | * Page is misplaced. Page lock serialises migrations. Acquire anon_vma | |
1492 | * to serialises splits | |
1493 | */ | |
b8916634 | 1494 | get_page(page); |
82b0f8c3 | 1495 | spin_unlock(vmf->ptl); |
b8916634 | 1496 | anon_vma = page_lock_anon_vma_read(page); |
4daae3b4 | 1497 | |
c69307d5 | 1498 | /* Confirm the PMD did not change while page_table_lock was released */ |
82b0f8c3 JK |
1499 | spin_lock(vmf->ptl); |
1500 | if (unlikely(!pmd_same(pmd, *vmf->pmd))) { | |
b32967ff MG |
1501 | unlock_page(page); |
1502 | put_page(page); | |
98fa15f3 | 1503 | page_nid = NUMA_NO_NODE; |
4daae3b4 | 1504 | goto out_unlock; |
b32967ff | 1505 | } |
ff9042b1 | 1506 | |
c3a489ca MG |
1507 | /* Bail if we fail to protect against THP splits for any reason */ |
1508 | if (unlikely(!anon_vma)) { | |
1509 | put_page(page); | |
98fa15f3 | 1510 | page_nid = NUMA_NO_NODE; |
c3a489ca MG |
1511 | goto clear_pmdnuma; |
1512 | } | |
1513 | ||
8b1b436d PZ |
1514 | /* |
1515 | * Since we took the NUMA fault, we must have observed the !accessible | |
1516 | * bit. Make sure all other CPUs agree with that, to avoid them | |
1517 | * modifying the page we're about to migrate. | |
1518 | * | |
1519 | * Must be done under PTL such that we'll observe the relevant | |
ccde85ba PZ |
1520 | * inc_tlb_flush_pending(). |
1521 | * | |
1522 | * We are not sure a pending tlb flush here is for a huge page | |
1523 | * mapping or not. Hence use the tlb range variant | |
8b1b436d | 1524 | */ |
7066f0f9 | 1525 | if (mm_tlb_flush_pending(vma->vm_mm)) { |
ccde85ba | 1526 | flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE); |
7066f0f9 AA |
1527 | /* |
1528 | * change_huge_pmd() released the pmd lock before | |
1529 | * invalidating the secondary MMUs sharing the primary | |
1530 | * MMU pagetables (with ->invalidate_range()). The | |
1531 | * mmu_notifier_invalidate_range_end() (which | |
1532 | * internally calls ->invalidate_range()) in | |
1533 | * change_pmd_range() will run after us, so we can't | |
1534 | * rely on it here and we need an explicit invalidate. | |
1535 | */ | |
1536 | mmu_notifier_invalidate_range(vma->vm_mm, haddr, | |
1537 | haddr + HPAGE_PMD_SIZE); | |
1538 | } | |
8b1b436d | 1539 | |
a54a407f MG |
1540 | /* |
1541 | * Migrate the THP to the requested node, returns with page unlocked | |
8a0516ed | 1542 | * and access rights restored. |
a54a407f | 1543 | */ |
82b0f8c3 | 1544 | spin_unlock(vmf->ptl); |
8b1b436d | 1545 | |
bae473a4 | 1546 | migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma, |
82b0f8c3 | 1547 | vmf->pmd, pmd, vmf->address, page, target_nid); |
6688cc05 PZ |
1548 | if (migrated) { |
1549 | flags |= TNF_MIGRATED; | |
8191acbd | 1550 | page_nid = target_nid; |
074c2381 MG |
1551 | } else |
1552 | flags |= TNF_MIGRATE_FAIL; | |
b32967ff | 1553 | |
8191acbd | 1554 | goto out; |
b32967ff | 1555 | clear_pmdnuma: |
a54a407f | 1556 | BUG_ON(!PageLocked(page)); |
288bc549 | 1557 | was_writable = pmd_savedwrite(pmd); |
4d942466 | 1558 | pmd = pmd_modify(pmd, vma->vm_page_prot); |
b7b04004 | 1559 | pmd = pmd_mkyoung(pmd); |
b191f9b1 MG |
1560 | if (was_writable) |
1561 | pmd = pmd_mkwrite(pmd); | |
82b0f8c3 JK |
1562 | set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd); |
1563 | update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); | |
a54a407f | 1564 | unlock_page(page); |
d10e63f2 | 1565 | out_unlock: |
82b0f8c3 | 1566 | spin_unlock(vmf->ptl); |
b8916634 MG |
1567 | |
1568 | out: | |
1569 | if (anon_vma) | |
1570 | page_unlock_anon_vma_read(anon_vma); | |
1571 | ||
98fa15f3 | 1572 | if (page_nid != NUMA_NO_NODE) |
82b0f8c3 | 1573 | task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, |
9a8b300f | 1574 | flags); |
8191acbd | 1575 | |
d10e63f2 MG |
1576 | return 0; |
1577 | } | |
1578 | ||
319904ad HY |
1579 | /* |
1580 | * Return true if we do MADV_FREE successfully on entire pmd page. | |
1581 | * Otherwise, return false. | |
1582 | */ | |
1583 | bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, | |
b8d3c4c3 | 1584 | pmd_t *pmd, unsigned long addr, unsigned long next) |
b8d3c4c3 MK |
1585 | { |
1586 | spinlock_t *ptl; | |
1587 | pmd_t orig_pmd; | |
1588 | struct page *page; | |
1589 | struct mm_struct *mm = tlb->mm; | |
319904ad | 1590 | bool ret = false; |
b8d3c4c3 | 1591 | |
ed6a7935 | 1592 | tlb_change_page_size(tlb, HPAGE_PMD_SIZE); |
07e32661 | 1593 | |
b6ec57f4 KS |
1594 | ptl = pmd_trans_huge_lock(pmd, vma); |
1595 | if (!ptl) | |
25eedabe | 1596 | goto out_unlocked; |
b8d3c4c3 MK |
1597 | |
1598 | orig_pmd = *pmd; | |
319904ad | 1599 | if (is_huge_zero_pmd(orig_pmd)) |
b8d3c4c3 | 1600 | goto out; |
b8d3c4c3 | 1601 | |
84c3fc4e ZY |
1602 | if (unlikely(!pmd_present(orig_pmd))) { |
1603 | VM_BUG_ON(thp_migration_supported() && | |
1604 | !is_pmd_migration_entry(orig_pmd)); | |
1605 | goto out; | |
1606 | } | |
1607 | ||
b8d3c4c3 MK |
1608 | page = pmd_page(orig_pmd); |
1609 | /* | |
1610 | * If other processes are mapping this page, we couldn't discard | |
1611 | * the page unless they all do MADV_FREE so let's skip the page. | |
1612 | */ | |
babbbdd0 | 1613 | if (total_mapcount(page) != 1) |
b8d3c4c3 MK |
1614 | goto out; |
1615 | ||
1616 | if (!trylock_page(page)) | |
1617 | goto out; | |
1618 | ||
1619 | /* | |
1620 | * If user want to discard part-pages of THP, split it so MADV_FREE | |
1621 | * will deactivate only them. | |
1622 | */ | |
1623 | if (next - addr != HPAGE_PMD_SIZE) { | |
1624 | get_page(page); | |
1625 | spin_unlock(ptl); | |
9818b8cd | 1626 | split_huge_page(page); |
b8d3c4c3 | 1627 | unlock_page(page); |
bbf29ffc | 1628 | put_page(page); |
b8d3c4c3 MK |
1629 | goto out_unlocked; |
1630 | } | |
1631 | ||
1632 | if (PageDirty(page)) | |
1633 | ClearPageDirty(page); | |
1634 | unlock_page(page); | |
1635 | ||
b8d3c4c3 | 1636 | if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { |
58ceeb6b | 1637 | pmdp_invalidate(vma, addr, pmd); |
b8d3c4c3 MK |
1638 | orig_pmd = pmd_mkold(orig_pmd); |
1639 | orig_pmd = pmd_mkclean(orig_pmd); | |
1640 | ||
1641 | set_pmd_at(mm, addr, pmd, orig_pmd); | |
1642 | tlb_remove_pmd_tlb_entry(tlb, pmd, addr); | |
1643 | } | |
802a3a92 SL |
1644 | |
1645 | mark_page_lazyfree(page); | |
319904ad | 1646 | ret = true; |
b8d3c4c3 MK |
1647 | out: |
1648 | spin_unlock(ptl); | |
1649 | out_unlocked: | |
1650 | return ret; | |
1651 | } | |
1652 | ||
953c66c2 AK |
1653 | static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd) |
1654 | { | |
1655 | pgtable_t pgtable; | |
1656 | ||
1657 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); | |
1658 | pte_free(mm, pgtable); | |
c4812909 | 1659 | mm_dec_nr_ptes(mm); |
953c66c2 AK |
1660 | } |
1661 | ||
71e3aac0 | 1662 | int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, |
f21760b1 | 1663 | pmd_t *pmd, unsigned long addr) |
71e3aac0 | 1664 | { |
da146769 | 1665 | pmd_t orig_pmd; |
bf929152 | 1666 | spinlock_t *ptl; |
71e3aac0 | 1667 | |
ed6a7935 | 1668 | tlb_change_page_size(tlb, HPAGE_PMD_SIZE); |
07e32661 | 1669 | |
b6ec57f4 KS |
1670 | ptl = __pmd_trans_huge_lock(pmd, vma); |
1671 | if (!ptl) | |
da146769 KS |
1672 | return 0; |
1673 | /* | |
1674 | * For architectures like ppc64 we look at deposited pgtable | |
1675 | * when calling pmdp_huge_get_and_clear. So do the | |
1676 | * pgtable_trans_huge_withdraw after finishing pmdp related | |
1677 | * operations. | |
1678 | */ | |
93a98695 AK |
1679 | orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd, |
1680 | tlb->fullmm); | |
da146769 | 1681 | tlb_remove_pmd_tlb_entry(tlb, pmd, addr); |
2484ca9b | 1682 | if (vma_is_special_huge(vma)) { |
3b6521f5 OH |
1683 | if (arch_needs_pgtable_deposit()) |
1684 | zap_deposited_table(tlb->mm, pmd); | |
da146769 | 1685 | spin_unlock(ptl); |
da146769 | 1686 | } else if (is_huge_zero_pmd(orig_pmd)) { |
c14a6eb4 | 1687 | zap_deposited_table(tlb->mm, pmd); |
da146769 | 1688 | spin_unlock(ptl); |
da146769 | 1689 | } else { |
616b8371 ZY |
1690 | struct page *page = NULL; |
1691 | int flush_needed = 1; | |
1692 | ||
1693 | if (pmd_present(orig_pmd)) { | |
1694 | page = pmd_page(orig_pmd); | |
1695 | page_remove_rmap(page, true); | |
1696 | VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); | |
1697 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
1698 | } else if (thp_migration_supported()) { | |
1699 | swp_entry_t entry; | |
1700 | ||
1701 | VM_BUG_ON(!is_pmd_migration_entry(orig_pmd)); | |
1702 | entry = pmd_to_swp_entry(orig_pmd); | |
a44f89dc | 1703 | page = migration_entry_to_page(entry); |
616b8371 ZY |
1704 | flush_needed = 0; |
1705 | } else | |
1706 | WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!"); | |
1707 | ||
b5072380 | 1708 | if (PageAnon(page)) { |
c14a6eb4 | 1709 | zap_deposited_table(tlb->mm, pmd); |
b5072380 KS |
1710 | add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); |
1711 | } else { | |
953c66c2 AK |
1712 | if (arch_needs_pgtable_deposit()) |
1713 | zap_deposited_table(tlb->mm, pmd); | |
fadae295 | 1714 | add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR); |
b5072380 | 1715 | } |
616b8371 | 1716 | |
da146769 | 1717 | spin_unlock(ptl); |
616b8371 ZY |
1718 | if (flush_needed) |
1719 | tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE); | |
025c5b24 | 1720 | } |
da146769 | 1721 | return 1; |
71e3aac0 AA |
1722 | } |
1723 | ||
1dd38b6c AK |
1724 | #ifndef pmd_move_must_withdraw |
1725 | static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl, | |
1726 | spinlock_t *old_pmd_ptl, | |
1727 | struct vm_area_struct *vma) | |
1728 | { | |
1729 | /* | |
1730 | * With split pmd lock we also need to move preallocated | |
1731 | * PTE page table if new_pmd is on different PMD page table. | |
1732 | * | |
1733 | * We also don't deposit and withdraw tables for file pages. | |
1734 | */ | |
1735 | return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma); | |
1736 | } | |
1737 | #endif | |
1738 | ||
ab6e3d09 NH |
1739 | static pmd_t move_soft_dirty_pmd(pmd_t pmd) |
1740 | { | |
1741 | #ifdef CONFIG_MEM_SOFT_DIRTY | |
1742 | if (unlikely(is_pmd_migration_entry(pmd))) | |
1743 | pmd = pmd_swp_mksoft_dirty(pmd); | |
1744 | else if (pmd_present(pmd)) | |
1745 | pmd = pmd_mksoft_dirty(pmd); | |
1746 | #endif | |
1747 | return pmd; | |
1748 | } | |
1749 | ||
bf8616d5 | 1750 | bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, |
b8aa9d9d | 1751 | unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd) |
37a1c49a | 1752 | { |
bf929152 | 1753 | spinlock_t *old_ptl, *new_ptl; |
37a1c49a | 1754 | pmd_t pmd; |
37a1c49a | 1755 | struct mm_struct *mm = vma->vm_mm; |
5d190420 | 1756 | bool force_flush = false; |
37a1c49a | 1757 | |
37a1c49a AA |
1758 | /* |
1759 | * The destination pmd shouldn't be established, free_pgtables() | |
1760 | * should have release it. | |
1761 | */ | |
1762 | if (WARN_ON(!pmd_none(*new_pmd))) { | |
1763 | VM_BUG_ON(pmd_trans_huge(*new_pmd)); | |
4b471e88 | 1764 | return false; |
37a1c49a AA |
1765 | } |
1766 | ||
bf929152 KS |
1767 | /* |
1768 | * We don't have to worry about the ordering of src and dst | |
c1e8d7c6 | 1769 | * ptlocks because exclusive mmap_lock prevents deadlock. |
bf929152 | 1770 | */ |
b6ec57f4 KS |
1771 | old_ptl = __pmd_trans_huge_lock(old_pmd, vma); |
1772 | if (old_ptl) { | |
bf929152 KS |
1773 | new_ptl = pmd_lockptr(mm, new_pmd); |
1774 | if (new_ptl != old_ptl) | |
1775 | spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); | |
8809aa2d | 1776 | pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); |
eb66ae03 | 1777 | if (pmd_present(pmd)) |
a2ce2666 | 1778 | force_flush = true; |
025c5b24 | 1779 | VM_BUG_ON(!pmd_none(*new_pmd)); |
3592806c | 1780 | |
1dd38b6c | 1781 | if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) { |
b3084f4d | 1782 | pgtable_t pgtable; |
3592806c KS |
1783 | pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); |
1784 | pgtable_trans_huge_deposit(mm, new_pmd, pgtable); | |
3592806c | 1785 | } |
ab6e3d09 NH |
1786 | pmd = move_soft_dirty_pmd(pmd); |
1787 | set_pmd_at(mm, new_addr, new_pmd, pmd); | |
5d190420 AL |
1788 | if (force_flush) |
1789 | flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE); | |
eb66ae03 LT |
1790 | if (new_ptl != old_ptl) |
1791 | spin_unlock(new_ptl); | |
bf929152 | 1792 | spin_unlock(old_ptl); |
4b471e88 | 1793 | return true; |
37a1c49a | 1794 | } |
4b471e88 | 1795 | return false; |
37a1c49a AA |
1796 | } |
1797 | ||
f123d74a MG |
1798 | /* |
1799 | * Returns | |
1800 | * - 0 if PMD could not be locked | |
f0953a1b IM |
1801 | * - 1 if PMD was locked but protections unchanged and TLB flush unnecessary |
1802 | * - HPAGE_PMD_NR if protections changed and TLB flush necessary | |
f123d74a | 1803 | */ |
cd7548ab | 1804 | int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, |
58705444 | 1805 | unsigned long addr, pgprot_t newprot, unsigned long cp_flags) |
cd7548ab JW |
1806 | { |
1807 | struct mm_struct *mm = vma->vm_mm; | |
bf929152 | 1808 | spinlock_t *ptl; |
0a85e51d KS |
1809 | pmd_t entry; |
1810 | bool preserve_write; | |
1811 | int ret; | |
58705444 | 1812 | bool prot_numa = cp_flags & MM_CP_PROT_NUMA; |
292924b2 PX |
1813 | bool uffd_wp = cp_flags & MM_CP_UFFD_WP; |
1814 | bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; | |
cd7548ab | 1815 | |
b6ec57f4 | 1816 | ptl = __pmd_trans_huge_lock(pmd, vma); |
0a85e51d KS |
1817 | if (!ptl) |
1818 | return 0; | |
e944fd67 | 1819 | |
0a85e51d KS |
1820 | preserve_write = prot_numa && pmd_write(*pmd); |
1821 | ret = 1; | |
e944fd67 | 1822 | |
84c3fc4e ZY |
1823 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION |
1824 | if (is_swap_pmd(*pmd)) { | |
1825 | swp_entry_t entry = pmd_to_swp_entry(*pmd); | |
1826 | ||
1827 | VM_BUG_ON(!is_pmd_migration_entry(*pmd)); | |
1828 | if (is_write_migration_entry(entry)) { | |
1829 | pmd_t newpmd; | |
1830 | /* | |
1831 | * A protection check is difficult so | |
1832 | * just be safe and disable write | |
1833 | */ | |
1834 | make_migration_entry_read(&entry); | |
1835 | newpmd = swp_entry_to_pmd(entry); | |
ab6e3d09 NH |
1836 | if (pmd_swp_soft_dirty(*pmd)) |
1837 | newpmd = pmd_swp_mksoft_dirty(newpmd); | |
84c3fc4e ZY |
1838 | set_pmd_at(mm, addr, pmd, newpmd); |
1839 | } | |
1840 | goto unlock; | |
1841 | } | |
1842 | #endif | |
1843 | ||
0a85e51d KS |
1844 | /* |
1845 | * Avoid trapping faults against the zero page. The read-only | |
1846 | * data is likely to be read-cached on the local CPU and | |
1847 | * local/remote hits to the zero page are not interesting. | |
1848 | */ | |
1849 | if (prot_numa && is_huge_zero_pmd(*pmd)) | |
1850 | goto unlock; | |
025c5b24 | 1851 | |
0a85e51d KS |
1852 | if (prot_numa && pmd_protnone(*pmd)) |
1853 | goto unlock; | |
1854 | ||
ced10803 | 1855 | /* |
3e4e28c5 | 1856 | * In case prot_numa, we are under mmap_read_lock(mm). It's critical |
ced10803 | 1857 | * to not clear pmd intermittently to avoid race with MADV_DONTNEED |
3e4e28c5 | 1858 | * which is also under mmap_read_lock(mm): |
ced10803 KS |
1859 | * |
1860 | * CPU0: CPU1: | |
1861 | * change_huge_pmd(prot_numa=1) | |
1862 | * pmdp_huge_get_and_clear_notify() | |
1863 | * madvise_dontneed() | |
1864 | * zap_pmd_range() | |
1865 | * pmd_trans_huge(*pmd) == 0 (without ptl) | |
1866 | * // skip the pmd | |
1867 | * set_pmd_at(); | |
1868 | * // pmd is re-established | |
1869 | * | |
1870 | * The race makes MADV_DONTNEED miss the huge pmd and don't clear it | |
1871 | * which may break userspace. | |
1872 | * | |
1873 | * pmdp_invalidate() is required to make sure we don't miss | |
1874 | * dirty/young flags set by hardware. | |
1875 | */ | |
a3cf988f | 1876 | entry = pmdp_invalidate(vma, addr, pmd); |
ced10803 | 1877 | |
0a85e51d KS |
1878 | entry = pmd_modify(entry, newprot); |
1879 | if (preserve_write) | |
1880 | entry = pmd_mk_savedwrite(entry); | |
292924b2 PX |
1881 | if (uffd_wp) { |
1882 | entry = pmd_wrprotect(entry); | |
1883 | entry = pmd_mkuffd_wp(entry); | |
1884 | } else if (uffd_wp_resolve) { | |
1885 | /* | |
1886 | * Leave the write bit to be handled by PF interrupt | |
1887 | * handler, then things like COW could be properly | |
1888 | * handled. | |
1889 | */ | |
1890 | entry = pmd_clear_uffd_wp(entry); | |
1891 | } | |
0a85e51d KS |
1892 | ret = HPAGE_PMD_NR; |
1893 | set_pmd_at(mm, addr, pmd, entry); | |
1894 | BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry)); | |
1895 | unlock: | |
1896 | spin_unlock(ptl); | |
025c5b24 NH |
1897 | return ret; |
1898 | } | |
1899 | ||
1900 | /* | |
8f19b0c0 | 1901 | * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise. |
025c5b24 | 1902 | * |
8f19b0c0 HY |
1903 | * Note that if it returns page table lock pointer, this routine returns without |
1904 | * unlocking page table lock. So callers must unlock it. | |
025c5b24 | 1905 | */ |
b6ec57f4 | 1906 | spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) |
025c5b24 | 1907 | { |
b6ec57f4 KS |
1908 | spinlock_t *ptl; |
1909 | ptl = pmd_lock(vma->vm_mm, pmd); | |
84c3fc4e ZY |
1910 | if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || |
1911 | pmd_devmap(*pmd))) | |
b6ec57f4 KS |
1912 | return ptl; |
1913 | spin_unlock(ptl); | |
1914 | return NULL; | |
cd7548ab JW |
1915 | } |
1916 | ||
a00cc7d9 MW |
1917 | /* |
1918 | * Returns true if a given pud maps a thp, false otherwise. | |
1919 | * | |
1920 | * Note that if it returns true, this routine returns without unlocking page | |
1921 | * table lock. So callers must unlock it. | |
1922 | */ | |
1923 | spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma) | |
1924 | { | |
1925 | spinlock_t *ptl; | |
1926 | ||
1927 | ptl = pud_lock(vma->vm_mm, pud); | |
1928 | if (likely(pud_trans_huge(*pud) || pud_devmap(*pud))) | |
1929 | return ptl; | |
1930 | spin_unlock(ptl); | |
1931 | return NULL; | |
1932 | } | |
1933 | ||
1934 | #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD | |
1935 | int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, | |
1936 | pud_t *pud, unsigned long addr) | |
1937 | { | |
a00cc7d9 MW |
1938 | spinlock_t *ptl; |
1939 | ||
1940 | ptl = __pud_trans_huge_lock(pud, vma); | |
1941 | if (!ptl) | |
1942 | return 0; | |
1943 | /* | |
1944 | * For architectures like ppc64 we look at deposited pgtable | |
1945 | * when calling pudp_huge_get_and_clear. So do the | |
1946 | * pgtable_trans_huge_withdraw after finishing pudp related | |
1947 | * operations. | |
1948 | */ | |
70516b93 | 1949 | pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm); |
a00cc7d9 | 1950 | tlb_remove_pud_tlb_entry(tlb, pud, addr); |
2484ca9b | 1951 | if (vma_is_special_huge(vma)) { |
a00cc7d9 MW |
1952 | spin_unlock(ptl); |
1953 | /* No zero page support yet */ | |
1954 | } else { | |
1955 | /* No support for anonymous PUD pages yet */ | |
1956 | BUG(); | |
1957 | } | |
1958 | return 1; | |
1959 | } | |
1960 | ||
1961 | static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud, | |
1962 | unsigned long haddr) | |
1963 | { | |
1964 | VM_BUG_ON(haddr & ~HPAGE_PUD_MASK); | |
1965 | VM_BUG_ON_VMA(vma->vm_start > haddr, vma); | |
1966 | VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma); | |
1967 | VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud)); | |
1968 | ||
ce9311cf | 1969 | count_vm_event(THP_SPLIT_PUD); |
a00cc7d9 MW |
1970 | |
1971 | pudp_huge_clear_flush_notify(vma, haddr, pud); | |
1972 | } | |
1973 | ||
1974 | void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud, | |
1975 | unsigned long address) | |
1976 | { | |
1977 | spinlock_t *ptl; | |
ac46d4f3 | 1978 | struct mmu_notifier_range range; |
a00cc7d9 | 1979 | |
7269f999 | 1980 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, |
6f4f13e8 | 1981 | address & HPAGE_PUD_MASK, |
ac46d4f3 JG |
1982 | (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE); |
1983 | mmu_notifier_invalidate_range_start(&range); | |
1984 | ptl = pud_lock(vma->vm_mm, pud); | |
a00cc7d9 MW |
1985 | if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud))) |
1986 | goto out; | |
ac46d4f3 | 1987 | __split_huge_pud_locked(vma, pud, range.start); |
a00cc7d9 MW |
1988 | |
1989 | out: | |
1990 | spin_unlock(ptl); | |
4645b9fe JG |
1991 | /* |
1992 | * No need to double call mmu_notifier->invalidate_range() callback as | |
1993 | * the above pudp_huge_clear_flush_notify() did already call it. | |
1994 | */ | |
ac46d4f3 | 1995 | mmu_notifier_invalidate_range_only_end(&range); |
a00cc7d9 MW |
1996 | } |
1997 | #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ | |
1998 | ||
eef1b3ba KS |
1999 | static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, |
2000 | unsigned long haddr, pmd_t *pmd) | |
2001 | { | |
2002 | struct mm_struct *mm = vma->vm_mm; | |
2003 | pgtable_t pgtable; | |
2004 | pmd_t _pmd; | |
2005 | int i; | |
2006 | ||
0f10851e JG |
2007 | /* |
2008 | * Leave pmd empty until pte is filled note that it is fine to delay | |
2009 | * notification until mmu_notifier_invalidate_range_end() as we are | |
2010 | * replacing a zero pmd write protected page with a zero pte write | |
2011 | * protected page. | |
2012 | * | |
ad56b738 | 2013 | * See Documentation/vm/mmu_notifier.rst |
0f10851e JG |
2014 | */ |
2015 | pmdp_huge_clear_flush(vma, haddr, pmd); | |
eef1b3ba KS |
2016 | |
2017 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); | |
2018 | pmd_populate(mm, &_pmd, pgtable); | |
2019 | ||
2020 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
2021 | pte_t *pte, entry; | |
2022 | entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); | |
2023 | entry = pte_mkspecial(entry); | |
2024 | pte = pte_offset_map(&_pmd, haddr); | |
2025 | VM_BUG_ON(!pte_none(*pte)); | |
2026 | set_pte_at(mm, haddr, pte, entry); | |
2027 | pte_unmap(pte); | |
2028 | } | |
2029 | smp_wmb(); /* make pte visible before pmd */ | |
2030 | pmd_populate(mm, pmd, pgtable); | |
eef1b3ba KS |
2031 | } |
2032 | ||
2033 | static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, | |
ba988280 | 2034 | unsigned long haddr, bool freeze) |
eef1b3ba KS |
2035 | { |
2036 | struct mm_struct *mm = vma->vm_mm; | |
2037 | struct page *page; | |
2038 | pgtable_t pgtable; | |
423ac9af | 2039 | pmd_t old_pmd, _pmd; |
292924b2 | 2040 | bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false; |
2ac015e2 | 2041 | unsigned long addr; |
eef1b3ba KS |
2042 | int i; |
2043 | ||
2044 | VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); | |
2045 | VM_BUG_ON_VMA(vma->vm_start > haddr, vma); | |
2046 | VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); | |
84c3fc4e ZY |
2047 | VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd) |
2048 | && !pmd_devmap(*pmd)); | |
eef1b3ba KS |
2049 | |
2050 | count_vm_event(THP_SPLIT_PMD); | |
2051 | ||
d21b9e57 | 2052 | if (!vma_is_anonymous(vma)) { |
99fa8a48 | 2053 | old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); |
953c66c2 AK |
2054 | /* |
2055 | * We are going to unmap this huge page. So | |
2056 | * just go ahead and zap it | |
2057 | */ | |
2058 | if (arch_needs_pgtable_deposit()) | |
2059 | zap_deposited_table(mm, pmd); | |
2484ca9b | 2060 | if (vma_is_special_huge(vma)) |
d21b9e57 | 2061 | return; |
99fa8a48 HD |
2062 | if (unlikely(is_pmd_migration_entry(old_pmd))) { |
2063 | swp_entry_t entry; | |
2064 | ||
2065 | entry = pmd_to_swp_entry(old_pmd); | |
2066 | page = migration_entry_to_page(entry); | |
2067 | } else { | |
2068 | page = pmd_page(old_pmd); | |
2069 | if (!PageDirty(page) && pmd_dirty(old_pmd)) | |
2070 | set_page_dirty(page); | |
2071 | if (!PageReferenced(page) && pmd_young(old_pmd)) | |
2072 | SetPageReferenced(page); | |
2073 | page_remove_rmap(page, true); | |
2074 | put_page(page); | |
2075 | } | |
fadae295 | 2076 | add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR); |
eef1b3ba | 2077 | return; |
99fa8a48 HD |
2078 | } |
2079 | ||
3b77e8c8 | 2080 | if (is_huge_zero_pmd(*pmd)) { |
4645b9fe JG |
2081 | /* |
2082 | * FIXME: Do we want to invalidate secondary mmu by calling | |
2083 | * mmu_notifier_invalidate_range() see comments below inside | |
2084 | * __split_huge_pmd() ? | |
2085 | * | |
2086 | * We are going from a zero huge page write protected to zero | |
2087 | * small page also write protected so it does not seems useful | |
2088 | * to invalidate secondary mmu at this time. | |
2089 | */ | |
eef1b3ba KS |
2090 | return __split_huge_zero_page_pmd(vma, haddr, pmd); |
2091 | } | |
2092 | ||
423ac9af AK |
2093 | /* |
2094 | * Up to this point the pmd is present and huge and userland has the | |
2095 | * whole access to the hugepage during the split (which happens in | |
2096 | * place). If we overwrite the pmd with the not-huge version pointing | |
2097 | * to the pte here (which of course we could if all CPUs were bug | |
2098 | * free), userland could trigger a small page size TLB miss on the | |
2099 | * small sized TLB while the hugepage TLB entry is still established in | |
2100 | * the huge TLB. Some CPU doesn't like that. | |
42742d9b AK |
2101 | * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum |
2102 | * 383 on page 105. Intel should be safe but is also warns that it's | |
423ac9af AK |
2103 | * only safe if the permission and cache attributes of the two entries |
2104 | * loaded in the two TLB is identical (which should be the case here). | |
2105 | * But it is generally safer to never allow small and huge TLB entries | |
2106 | * for the same virtual address to be loaded simultaneously. So instead | |
2107 | * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the | |
2108 | * current pmd notpresent (atomically because here the pmd_trans_huge | |
2109 | * must remain set at all times on the pmd until the split is complete | |
2110 | * for this pmd), then we flush the SMP TLB and finally we write the | |
2111 | * non-huge version of the pmd entry with pmd_populate. | |
2112 | */ | |
2113 | old_pmd = pmdp_invalidate(vma, haddr, pmd); | |
2114 | ||
423ac9af | 2115 | pmd_migration = is_pmd_migration_entry(old_pmd); |
2e83ee1d | 2116 | if (unlikely(pmd_migration)) { |
84c3fc4e ZY |
2117 | swp_entry_t entry; |
2118 | ||
423ac9af | 2119 | entry = pmd_to_swp_entry(old_pmd); |
a44f89dc | 2120 | page = migration_entry_to_page(entry); |
2e83ee1d PX |
2121 | write = is_write_migration_entry(entry); |
2122 | young = false; | |
2123 | soft_dirty = pmd_swp_soft_dirty(old_pmd); | |
f45ec5ff | 2124 | uffd_wp = pmd_swp_uffd_wp(old_pmd); |
2e83ee1d | 2125 | } else { |
423ac9af | 2126 | page = pmd_page(old_pmd); |
2e83ee1d PX |
2127 | if (pmd_dirty(old_pmd)) |
2128 | SetPageDirty(page); | |
2129 | write = pmd_write(old_pmd); | |
2130 | young = pmd_young(old_pmd); | |
2131 | soft_dirty = pmd_soft_dirty(old_pmd); | |
292924b2 | 2132 | uffd_wp = pmd_uffd_wp(old_pmd); |
2e83ee1d | 2133 | } |
eef1b3ba | 2134 | VM_BUG_ON_PAGE(!page_count(page), page); |
fe896d18 | 2135 | page_ref_add(page, HPAGE_PMD_NR - 1); |
eef1b3ba | 2136 | |
423ac9af AK |
2137 | /* |
2138 | * Withdraw the table only after we mark the pmd entry invalid. | |
2139 | * This's critical for some architectures (Power). | |
2140 | */ | |
eef1b3ba KS |
2141 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); |
2142 | pmd_populate(mm, &_pmd, pgtable); | |
2143 | ||
2ac015e2 | 2144 | for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { |
eef1b3ba KS |
2145 | pte_t entry, *pte; |
2146 | /* | |
2147 | * Note that NUMA hinting access restrictions are not | |
2148 | * transferred to avoid any possibility of altering | |
2149 | * permissions across VMAs. | |
2150 | */ | |
84c3fc4e | 2151 | if (freeze || pmd_migration) { |
ba988280 KS |
2152 | swp_entry_t swp_entry; |
2153 | swp_entry = make_migration_entry(page + i, write); | |
2154 | entry = swp_entry_to_pte(swp_entry); | |
804dd150 AA |
2155 | if (soft_dirty) |
2156 | entry = pte_swp_mksoft_dirty(entry); | |
f45ec5ff PX |
2157 | if (uffd_wp) |
2158 | entry = pte_swp_mkuffd_wp(entry); | |
ba988280 | 2159 | } else { |
6d2329f8 | 2160 | entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot)); |
b8d3c4c3 | 2161 | entry = maybe_mkwrite(entry, vma); |
ba988280 KS |
2162 | if (!write) |
2163 | entry = pte_wrprotect(entry); | |
2164 | if (!young) | |
2165 | entry = pte_mkold(entry); | |
804dd150 AA |
2166 | if (soft_dirty) |
2167 | entry = pte_mksoft_dirty(entry); | |
292924b2 PX |
2168 | if (uffd_wp) |
2169 | entry = pte_mkuffd_wp(entry); | |
ba988280 | 2170 | } |
2ac015e2 | 2171 | pte = pte_offset_map(&_pmd, addr); |
eef1b3ba | 2172 | BUG_ON(!pte_none(*pte)); |
2ac015e2 | 2173 | set_pte_at(mm, addr, pte, entry); |
ec0abae6 | 2174 | if (!pmd_migration) |
eef1b3ba | 2175 | atomic_inc(&page[i]._mapcount); |
ec0abae6 | 2176 | pte_unmap(pte); |
eef1b3ba KS |
2177 | } |
2178 | ||
ec0abae6 RC |
2179 | if (!pmd_migration) { |
2180 | /* | |
2181 | * Set PG_double_map before dropping compound_mapcount to avoid | |
2182 | * false-negative page_mapped(). | |
2183 | */ | |
2184 | if (compound_mapcount(page) > 1 && | |
2185 | !TestSetPageDoubleMap(page)) { | |
eef1b3ba | 2186 | for (i = 0; i < HPAGE_PMD_NR; i++) |
ec0abae6 RC |
2187 | atomic_inc(&page[i]._mapcount); |
2188 | } | |
2189 | ||
2190 | lock_page_memcg(page); | |
2191 | if (atomic_add_negative(-1, compound_mapcount_ptr(page))) { | |
2192 | /* Last compound_mapcount is gone. */ | |
69473e5d MS |
2193 | __mod_lruvec_page_state(page, NR_ANON_THPS, |
2194 | -HPAGE_PMD_NR); | |
ec0abae6 RC |
2195 | if (TestClearPageDoubleMap(page)) { |
2196 | /* No need in mapcount reference anymore */ | |
2197 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
2198 | atomic_dec(&page[i]._mapcount); | |
2199 | } | |
eef1b3ba | 2200 | } |
ec0abae6 | 2201 | unlock_page_memcg(page); |
eef1b3ba KS |
2202 | } |
2203 | ||
2204 | smp_wmb(); /* make pte visible before pmd */ | |
2205 | pmd_populate(mm, pmd, pgtable); | |
e9b61f19 KS |
2206 | |
2207 | if (freeze) { | |
2ac015e2 | 2208 | for (i = 0; i < HPAGE_PMD_NR; i++) { |
e9b61f19 KS |
2209 | page_remove_rmap(page + i, false); |
2210 | put_page(page + i); | |
2211 | } | |
2212 | } | |
eef1b3ba KS |
2213 | } |
2214 | ||
2215 | void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, | |
33f4751e | 2216 | unsigned long address, bool freeze, struct page *page) |
eef1b3ba KS |
2217 | { |
2218 | spinlock_t *ptl; | |
ac46d4f3 | 2219 | struct mmu_notifier_range range; |
1c2f6730 | 2220 | bool do_unlock_page = false; |
c444eb56 | 2221 | pmd_t _pmd; |
eef1b3ba | 2222 | |
7269f999 | 2223 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, |
6f4f13e8 | 2224 | address & HPAGE_PMD_MASK, |
ac46d4f3 JG |
2225 | (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE); |
2226 | mmu_notifier_invalidate_range_start(&range); | |
2227 | ptl = pmd_lock(vma->vm_mm, pmd); | |
33f4751e NH |
2228 | |
2229 | /* | |
2230 | * If caller asks to setup a migration entries, we need a page to check | |
2231 | * pmd against. Otherwise we can end up replacing wrong page. | |
2232 | */ | |
2233 | VM_BUG_ON(freeze && !page); | |
c444eb56 AA |
2234 | if (page) { |
2235 | VM_WARN_ON_ONCE(!PageLocked(page)); | |
c444eb56 AA |
2236 | if (page != pmd_page(*pmd)) |
2237 | goto out; | |
2238 | } | |
33f4751e | 2239 | |
c444eb56 | 2240 | repeat: |
5c7fb56e | 2241 | if (pmd_trans_huge(*pmd)) { |
c444eb56 AA |
2242 | if (!page) { |
2243 | page = pmd_page(*pmd); | |
1c2f6730 HD |
2244 | /* |
2245 | * An anonymous page must be locked, to ensure that a | |
2246 | * concurrent reuse_swap_page() sees stable mapcount; | |
2247 | * but reuse_swap_page() is not used on shmem or file, | |
2248 | * and page lock must not be taken when zap_pmd_range() | |
2249 | * calls __split_huge_pmd() while i_mmap_lock is held. | |
2250 | */ | |
2251 | if (PageAnon(page)) { | |
2252 | if (unlikely(!trylock_page(page))) { | |
2253 | get_page(page); | |
2254 | _pmd = *pmd; | |
2255 | spin_unlock(ptl); | |
2256 | lock_page(page); | |
2257 | spin_lock(ptl); | |
2258 | if (unlikely(!pmd_same(*pmd, _pmd))) { | |
2259 | unlock_page(page); | |
2260 | put_page(page); | |
2261 | page = NULL; | |
2262 | goto repeat; | |
2263 | } | |
c444eb56 | 2264 | put_page(page); |
c444eb56 | 2265 | } |
1c2f6730 | 2266 | do_unlock_page = true; |
c444eb56 AA |
2267 | } |
2268 | } | |
5c7fb56e | 2269 | if (PageMlocked(page)) |
5f737714 | 2270 | clear_page_mlock(page); |
84c3fc4e | 2271 | } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd))) |
e90309c9 | 2272 | goto out; |
ac46d4f3 | 2273 | __split_huge_pmd_locked(vma, pmd, range.start, freeze); |
e90309c9 | 2274 | out: |
eef1b3ba | 2275 | spin_unlock(ptl); |
1c2f6730 | 2276 | if (do_unlock_page) |
c444eb56 | 2277 | unlock_page(page); |
4645b9fe JG |
2278 | /* |
2279 | * No need to double call mmu_notifier->invalidate_range() callback. | |
2280 | * They are 3 cases to consider inside __split_huge_pmd_locked(): | |
2281 | * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious | |
2282 | * 2) __split_huge_zero_page_pmd() read only zero page and any write | |
2283 | * fault will trigger a flush_notify before pointing to a new page | |
2284 | * (it is fine if the secondary mmu keeps pointing to the old zero | |
2285 | * page in the meantime) | |
2286 | * 3) Split a huge pmd into pte pointing to the same page. No need | |
2287 | * to invalidate secondary tlb entry they are all still valid. | |
2288 | * any further changes to individual pte will notify. So no need | |
2289 | * to call mmu_notifier->invalidate_range() | |
2290 | */ | |
ac46d4f3 | 2291 | mmu_notifier_invalidate_range_only_end(&range); |
eef1b3ba KS |
2292 | } |
2293 | ||
fec89c10 KS |
2294 | void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address, |
2295 | bool freeze, struct page *page) | |
94fcc585 | 2296 | { |
f72e7dcd | 2297 | pgd_t *pgd; |
c2febafc | 2298 | p4d_t *p4d; |
f72e7dcd | 2299 | pud_t *pud; |
94fcc585 AA |
2300 | pmd_t *pmd; |
2301 | ||
78ddc534 | 2302 | pgd = pgd_offset(vma->vm_mm, address); |
f72e7dcd HD |
2303 | if (!pgd_present(*pgd)) |
2304 | return; | |
2305 | ||
c2febafc KS |
2306 | p4d = p4d_offset(pgd, address); |
2307 | if (!p4d_present(*p4d)) | |
2308 | return; | |
2309 | ||
2310 | pud = pud_offset(p4d, address); | |
f72e7dcd HD |
2311 | if (!pud_present(*pud)) |
2312 | return; | |
2313 | ||
2314 | pmd = pmd_offset(pud, address); | |
fec89c10 | 2315 | |
33f4751e | 2316 | __split_huge_pmd(vma, pmd, address, freeze, page); |
94fcc585 AA |
2317 | } |
2318 | ||
71f9e58e ML |
2319 | static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address) |
2320 | { | |
2321 | /* | |
2322 | * If the new address isn't hpage aligned and it could previously | |
2323 | * contain an hugepage: check if we need to split an huge pmd. | |
2324 | */ | |
2325 | if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) && | |
2326 | range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE), | |
2327 | ALIGN(address, HPAGE_PMD_SIZE))) | |
2328 | split_huge_pmd_address(vma, address, false, NULL); | |
2329 | } | |
2330 | ||
e1b9996b | 2331 | void vma_adjust_trans_huge(struct vm_area_struct *vma, |
94fcc585 AA |
2332 | unsigned long start, |
2333 | unsigned long end, | |
2334 | long adjust_next) | |
2335 | { | |
71f9e58e ML |
2336 | /* Check if we need to split start first. */ |
2337 | split_huge_pmd_if_needed(vma, start); | |
94fcc585 | 2338 | |
71f9e58e ML |
2339 | /* Check if we need to split end next. */ |
2340 | split_huge_pmd_if_needed(vma, end); | |
94fcc585 AA |
2341 | |
2342 | /* | |
71f9e58e ML |
2343 | * If we're also updating the vma->vm_next->vm_start, |
2344 | * check if we need to split it. | |
94fcc585 AA |
2345 | */ |
2346 | if (adjust_next > 0) { | |
2347 | struct vm_area_struct *next = vma->vm_next; | |
2348 | unsigned long nstart = next->vm_start; | |
f9d86a60 | 2349 | nstart += adjust_next; |
71f9e58e | 2350 | split_huge_pmd_if_needed(next, nstart); |
94fcc585 AA |
2351 | } |
2352 | } | |
e9b61f19 | 2353 | |
906f9cdf | 2354 | static void unmap_page(struct page *page) |
e9b61f19 | 2355 | { |
732ed558 | 2356 | enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_SYNC | |
c7ab0d2f | 2357 | TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD; |
e9b61f19 KS |
2358 | |
2359 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
2360 | ||
baa355fd | 2361 | if (PageAnon(page)) |
b5ff8161 | 2362 | ttu_flags |= TTU_SPLIT_FREEZE; |
baa355fd | 2363 | |
504e070d YS |
2364 | try_to_unmap(page, ttu_flags); |
2365 | ||
2366 | VM_WARN_ON_ONCE_PAGE(page_mapped(page), page); | |
e9b61f19 KS |
2367 | } |
2368 | ||
8cce5475 | 2369 | static void remap_page(struct page *page, unsigned int nr) |
e9b61f19 | 2370 | { |
fec89c10 | 2371 | int i; |
ace71a19 KS |
2372 | if (PageTransHuge(page)) { |
2373 | remove_migration_ptes(page, page, true); | |
2374 | } else { | |
8cce5475 | 2375 | for (i = 0; i < nr; i++) |
ace71a19 KS |
2376 | remove_migration_ptes(page + i, page + i, true); |
2377 | } | |
e9b61f19 KS |
2378 | } |
2379 | ||
94866635 | 2380 | static void lru_add_page_tail(struct page *head, struct page *tail, |
88dcb9a3 AS |
2381 | struct lruvec *lruvec, struct list_head *list) |
2382 | { | |
94866635 AS |
2383 | VM_BUG_ON_PAGE(!PageHead(head), head); |
2384 | VM_BUG_ON_PAGE(PageCompound(tail), head); | |
2385 | VM_BUG_ON_PAGE(PageLRU(tail), head); | |
6168d0da | 2386 | lockdep_assert_held(&lruvec->lru_lock); |
88dcb9a3 | 2387 | |
6dbb5741 | 2388 | if (list) { |
88dcb9a3 | 2389 | /* page reclaim is reclaiming a huge page */ |
6dbb5741 | 2390 | VM_WARN_ON(PageLRU(head)); |
94866635 AS |
2391 | get_page(tail); |
2392 | list_add_tail(&tail->lru, list); | |
88dcb9a3 | 2393 | } else { |
6dbb5741 AS |
2394 | /* head is still on lru (and we have it frozen) */ |
2395 | VM_WARN_ON(!PageLRU(head)); | |
2396 | SetPageLRU(tail); | |
2397 | list_add_tail(&tail->lru, &head->lru); | |
88dcb9a3 AS |
2398 | } |
2399 | } | |
2400 | ||
8df651c7 | 2401 | static void __split_huge_page_tail(struct page *head, int tail, |
e9b61f19 KS |
2402 | struct lruvec *lruvec, struct list_head *list) |
2403 | { | |
e9b61f19 KS |
2404 | struct page *page_tail = head + tail; |
2405 | ||
8df651c7 | 2406 | VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail); |
e9b61f19 KS |
2407 | |
2408 | /* | |
605ca5ed KK |
2409 | * Clone page flags before unfreezing refcount. |
2410 | * | |
2411 | * After successful get_page_unless_zero() might follow flags change, | |
8958b249 | 2412 | * for example lock_page() which set PG_waiters. |
e9b61f19 | 2413 | */ |
e9b61f19 KS |
2414 | page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; |
2415 | page_tail->flags |= (head->flags & | |
2416 | ((1L << PG_referenced) | | |
2417 | (1L << PG_swapbacked) | | |
38d8b4e6 | 2418 | (1L << PG_swapcache) | |
e9b61f19 KS |
2419 | (1L << PG_mlocked) | |
2420 | (1L << PG_uptodate) | | |
2421 | (1L << PG_active) | | |
1899ad18 | 2422 | (1L << PG_workingset) | |
e9b61f19 | 2423 | (1L << PG_locked) | |
b8d3c4c3 | 2424 | (1L << PG_unevictable) | |
72e6afa0 CM |
2425 | #ifdef CONFIG_64BIT |
2426 | (1L << PG_arch_2) | | |
2427 | #endif | |
b8d3c4c3 | 2428 | (1L << PG_dirty))); |
e9b61f19 | 2429 | |
173d9d9f HD |
2430 | /* ->mapping in first tail page is compound_mapcount */ |
2431 | VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING, | |
2432 | page_tail); | |
2433 | page_tail->mapping = head->mapping; | |
2434 | page_tail->index = head->index + tail; | |
2435 | ||
605ca5ed | 2436 | /* Page flags must be visible before we make the page non-compound. */ |
e9b61f19 KS |
2437 | smp_wmb(); |
2438 | ||
605ca5ed KK |
2439 | /* |
2440 | * Clear PageTail before unfreezing page refcount. | |
2441 | * | |
2442 | * After successful get_page_unless_zero() might follow put_page() | |
2443 | * which needs correct compound_head(). | |
2444 | */ | |
e9b61f19 KS |
2445 | clear_compound_head(page_tail); |
2446 | ||
605ca5ed KK |
2447 | /* Finally unfreeze refcount. Additional reference from page cache. */ |
2448 | page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) || | |
2449 | PageSwapCache(head))); | |
2450 | ||
e9b61f19 KS |
2451 | if (page_is_young(head)) |
2452 | set_page_young(page_tail); | |
2453 | if (page_is_idle(head)) | |
2454 | set_page_idle(page_tail); | |
2455 | ||
e9b61f19 | 2456 | page_cpupid_xchg_last(page_tail, page_cpupid_last(head)); |
94723aaf MH |
2457 | |
2458 | /* | |
2459 | * always add to the tail because some iterators expect new | |
2460 | * pages to show after the currently processed elements - e.g. | |
2461 | * migrate_pages | |
2462 | */ | |
e9b61f19 | 2463 | lru_add_page_tail(head, page_tail, lruvec, list); |
e9b61f19 KS |
2464 | } |
2465 | ||
baa355fd | 2466 | static void __split_huge_page(struct page *page, struct list_head *list, |
b6769834 | 2467 | pgoff_t end) |
e9b61f19 KS |
2468 | { |
2469 | struct page *head = compound_head(page); | |
e9b61f19 | 2470 | struct lruvec *lruvec; |
4101196b MWO |
2471 | struct address_space *swap_cache = NULL; |
2472 | unsigned long offset = 0; | |
8cce5475 | 2473 | unsigned int nr = thp_nr_pages(head); |
8df651c7 | 2474 | int i; |
e9b61f19 | 2475 | |
e9b61f19 | 2476 | /* complete memcg works before add pages to LRU */ |
be6c8982 | 2477 | split_page_memcg(head, nr); |
e9b61f19 | 2478 | |
4101196b MWO |
2479 | if (PageAnon(head) && PageSwapCache(head)) { |
2480 | swp_entry_t entry = { .val = page_private(head) }; | |
2481 | ||
2482 | offset = swp_offset(entry); | |
2483 | swap_cache = swap_address_space(entry); | |
2484 | xa_lock(&swap_cache->i_pages); | |
2485 | } | |
2486 | ||
f0953a1b | 2487 | /* lock lru list/PageCompound, ref frozen by page_ref_freeze */ |
6168d0da | 2488 | lruvec = lock_page_lruvec(head); |
b6769834 | 2489 | |
8cce5475 | 2490 | for (i = nr - 1; i >= 1; i--) { |
8df651c7 | 2491 | __split_huge_page_tail(head, i, lruvec, list); |
baa355fd KS |
2492 | /* Some pages can be beyond i_size: drop them from page cache */ |
2493 | if (head[i].index >= end) { | |
2d077d4b | 2494 | ClearPageDirty(head + i); |
baa355fd | 2495 | __delete_from_page_cache(head + i, NULL); |
800d8c63 KS |
2496 | if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head)) |
2497 | shmem_uncharge(head->mapping->host, 1); | |
baa355fd | 2498 | put_page(head + i); |
4101196b MWO |
2499 | } else if (!PageAnon(page)) { |
2500 | __xa_store(&head->mapping->i_pages, head[i].index, | |
2501 | head + i, 0); | |
2502 | } else if (swap_cache) { | |
2503 | __xa_store(&swap_cache->i_pages, offset + i, | |
2504 | head + i, 0); | |
baa355fd KS |
2505 | } |
2506 | } | |
e9b61f19 KS |
2507 | |
2508 | ClearPageCompound(head); | |
6168d0da | 2509 | unlock_page_lruvec(lruvec); |
b6769834 | 2510 | /* Caller disabled irqs, so they are still disabled here */ |
f7da677b | 2511 | |
8cce5475 | 2512 | split_page_owner(head, nr); |
f7da677b | 2513 | |
baa355fd KS |
2514 | /* See comment in __split_huge_page_tail() */ |
2515 | if (PageAnon(head)) { | |
aa5dc07f | 2516 | /* Additional pin to swap cache */ |
4101196b | 2517 | if (PageSwapCache(head)) { |
38d8b4e6 | 2518 | page_ref_add(head, 2); |
4101196b MWO |
2519 | xa_unlock(&swap_cache->i_pages); |
2520 | } else { | |
38d8b4e6 | 2521 | page_ref_inc(head); |
4101196b | 2522 | } |
baa355fd | 2523 | } else { |
aa5dc07f | 2524 | /* Additional pin to page cache */ |
baa355fd | 2525 | page_ref_add(head, 2); |
b93b0163 | 2526 | xa_unlock(&head->mapping->i_pages); |
baa355fd | 2527 | } |
b6769834 | 2528 | local_irq_enable(); |
e9b61f19 | 2529 | |
8cce5475 | 2530 | remap_page(head, nr); |
e9b61f19 | 2531 | |
c4f9c701 HY |
2532 | if (PageSwapCache(head)) { |
2533 | swp_entry_t entry = { .val = page_private(head) }; | |
2534 | ||
2535 | split_swap_cluster(entry); | |
2536 | } | |
2537 | ||
8cce5475 | 2538 | for (i = 0; i < nr; i++) { |
e9b61f19 KS |
2539 | struct page *subpage = head + i; |
2540 | if (subpage == page) | |
2541 | continue; | |
2542 | unlock_page(subpage); | |
2543 | ||
2544 | /* | |
2545 | * Subpages may be freed if there wasn't any mapping | |
2546 | * like if add_to_swap() is running on a lru page that | |
2547 | * had its mapping zapped. And freeing these pages | |
2548 | * requires taking the lru_lock so we do the put_page | |
2549 | * of the tail pages after the split is complete. | |
2550 | */ | |
2551 | put_page(subpage); | |
2552 | } | |
2553 | } | |
2554 | ||
b20ce5e0 KS |
2555 | int total_mapcount(struct page *page) |
2556 | { | |
86b562b6 | 2557 | int i, compound, nr, ret; |
b20ce5e0 KS |
2558 | |
2559 | VM_BUG_ON_PAGE(PageTail(page), page); | |
2560 | ||
2561 | if (likely(!PageCompound(page))) | |
2562 | return atomic_read(&page->_mapcount) + 1; | |
2563 | ||
dd78fedd | 2564 | compound = compound_mapcount(page); |
86b562b6 | 2565 | nr = compound_nr(page); |
b20ce5e0 | 2566 | if (PageHuge(page)) |
dd78fedd KS |
2567 | return compound; |
2568 | ret = compound; | |
86b562b6 | 2569 | for (i = 0; i < nr; i++) |
b20ce5e0 | 2570 | ret += atomic_read(&page[i]._mapcount) + 1; |
dd78fedd KS |
2571 | /* File pages has compound_mapcount included in _mapcount */ |
2572 | if (!PageAnon(page)) | |
86b562b6 | 2573 | return ret - compound * nr; |
b20ce5e0 | 2574 | if (PageDoubleMap(page)) |
86b562b6 | 2575 | ret -= nr; |
b20ce5e0 KS |
2576 | return ret; |
2577 | } | |
2578 | ||
6d0a07ed AA |
2579 | /* |
2580 | * This calculates accurately how many mappings a transparent hugepage | |
2581 | * has (unlike page_mapcount() which isn't fully accurate). This full | |
2582 | * accuracy is primarily needed to know if copy-on-write faults can | |
2583 | * reuse the page and change the mapping to read-write instead of | |
2584 | * copying them. At the same time this returns the total_mapcount too. | |
2585 | * | |
2586 | * The function returns the highest mapcount any one of the subpages | |
2587 | * has. If the return value is one, even if different processes are | |
2588 | * mapping different subpages of the transparent hugepage, they can | |
2589 | * all reuse it, because each process is reusing a different subpage. | |
2590 | * | |
2591 | * The total_mapcount is instead counting all virtual mappings of the | |
2592 | * subpages. If the total_mapcount is equal to "one", it tells the | |
2593 | * caller all mappings belong to the same "mm" and in turn the | |
2594 | * anon_vma of the transparent hugepage can become the vma->anon_vma | |
2595 | * local one as no other process may be mapping any of the subpages. | |
2596 | * | |
2597 | * It would be more accurate to replace page_mapcount() with | |
2598 | * page_trans_huge_mapcount(), however we only use | |
2599 | * page_trans_huge_mapcount() in the copy-on-write faults where we | |
2600 | * need full accuracy to avoid breaking page pinning, because | |
2601 | * page_trans_huge_mapcount() is slower than page_mapcount(). | |
2602 | */ | |
2603 | int page_trans_huge_mapcount(struct page *page, int *total_mapcount) | |
2604 | { | |
2605 | int i, ret, _total_mapcount, mapcount; | |
2606 | ||
2607 | /* hugetlbfs shouldn't call it */ | |
2608 | VM_BUG_ON_PAGE(PageHuge(page), page); | |
2609 | ||
2610 | if (likely(!PageTransCompound(page))) { | |
2611 | mapcount = atomic_read(&page->_mapcount) + 1; | |
2612 | if (total_mapcount) | |
2613 | *total_mapcount = mapcount; | |
2614 | return mapcount; | |
2615 | } | |
2616 | ||
2617 | page = compound_head(page); | |
2618 | ||
2619 | _total_mapcount = ret = 0; | |
65dfe3c3 | 2620 | for (i = 0; i < thp_nr_pages(page); i++) { |
6d0a07ed AA |
2621 | mapcount = atomic_read(&page[i]._mapcount) + 1; |
2622 | ret = max(ret, mapcount); | |
2623 | _total_mapcount += mapcount; | |
2624 | } | |
2625 | if (PageDoubleMap(page)) { | |
2626 | ret -= 1; | |
65dfe3c3 | 2627 | _total_mapcount -= thp_nr_pages(page); |
6d0a07ed AA |
2628 | } |
2629 | mapcount = compound_mapcount(page); | |
2630 | ret += mapcount; | |
2631 | _total_mapcount += mapcount; | |
2632 | if (total_mapcount) | |
2633 | *total_mapcount = _total_mapcount; | |
2634 | return ret; | |
2635 | } | |
2636 | ||
b8f593cd HY |
2637 | /* Racy check whether the huge page can be split */ |
2638 | bool can_split_huge_page(struct page *page, int *pextra_pins) | |
2639 | { | |
2640 | int extra_pins; | |
2641 | ||
aa5dc07f | 2642 | /* Additional pins from page cache */ |
b8f593cd | 2643 | if (PageAnon(page)) |
e2333dad | 2644 | extra_pins = PageSwapCache(page) ? thp_nr_pages(page) : 0; |
b8f593cd | 2645 | else |
e2333dad | 2646 | extra_pins = thp_nr_pages(page); |
b8f593cd HY |
2647 | if (pextra_pins) |
2648 | *pextra_pins = extra_pins; | |
2649 | return total_mapcount(page) == page_count(page) - extra_pins - 1; | |
2650 | } | |
2651 | ||
e9b61f19 KS |
2652 | /* |
2653 | * This function splits huge page into normal pages. @page can point to any | |
2654 | * subpage of huge page to split. Split doesn't change the position of @page. | |
2655 | * | |
2656 | * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. | |
2657 | * The huge page must be locked. | |
2658 | * | |
2659 | * If @list is null, tail pages will be added to LRU list, otherwise, to @list. | |
2660 | * | |
2661 | * Both head page and tail pages will inherit mapping, flags, and so on from | |
2662 | * the hugepage. | |
2663 | * | |
2664 | * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if | |
2665 | * they are not mapped. | |
2666 | * | |
2667 | * Returns 0 if the hugepage is split successfully. | |
2668 | * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under | |
2669 | * us. | |
2670 | */ | |
2671 | int split_huge_page_to_list(struct page *page, struct list_head *list) | |
2672 | { | |
2673 | struct page *head = compound_head(page); | |
a8803e6c | 2674 | struct deferred_split *ds_queue = get_deferred_split_queue(head); |
baa355fd KS |
2675 | struct anon_vma *anon_vma = NULL; |
2676 | struct address_space *mapping = NULL; | |
504e070d | 2677 | int extra_pins, ret; |
006d3ff2 | 2678 | pgoff_t end; |
e9b61f19 | 2679 | |
cb829624 | 2680 | VM_BUG_ON_PAGE(is_huge_zero_page(head), head); |
a8803e6c WY |
2681 | VM_BUG_ON_PAGE(!PageLocked(head), head); |
2682 | VM_BUG_ON_PAGE(!PageCompound(head), head); | |
e9b61f19 | 2683 | |
a8803e6c | 2684 | if (PageWriteback(head)) |
59807685 HY |
2685 | return -EBUSY; |
2686 | ||
baa355fd KS |
2687 | if (PageAnon(head)) { |
2688 | /* | |
c1e8d7c6 | 2689 | * The caller does not necessarily hold an mmap_lock that would |
baa355fd KS |
2690 | * prevent the anon_vma disappearing so we first we take a |
2691 | * reference to it and then lock the anon_vma for write. This | |
2692 | * is similar to page_lock_anon_vma_read except the write lock | |
2693 | * is taken to serialise against parallel split or collapse | |
2694 | * operations. | |
2695 | */ | |
2696 | anon_vma = page_get_anon_vma(head); | |
2697 | if (!anon_vma) { | |
2698 | ret = -EBUSY; | |
2699 | goto out; | |
2700 | } | |
006d3ff2 | 2701 | end = -1; |
baa355fd KS |
2702 | mapping = NULL; |
2703 | anon_vma_lock_write(anon_vma); | |
2704 | } else { | |
2705 | mapping = head->mapping; | |
2706 | ||
2707 | /* Truncated ? */ | |
2708 | if (!mapping) { | |
2709 | ret = -EBUSY; | |
2710 | goto out; | |
2711 | } | |
2712 | ||
baa355fd KS |
2713 | anon_vma = NULL; |
2714 | i_mmap_lock_read(mapping); | |
006d3ff2 HD |
2715 | |
2716 | /* | |
2717 | *__split_huge_page() may need to trim off pages beyond EOF: | |
2718 | * but on 32-bit, i_size_read() takes an irq-unsafe seqlock, | |
2719 | * which cannot be nested inside the page tree lock. So note | |
2720 | * end now: i_size itself may be changed at any moment, but | |
2721 | * head page lock is good enough to serialize the trimming. | |
2722 | */ | |
2723 | end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE); | |
e9b61f19 | 2724 | } |
e9b61f19 KS |
2725 | |
2726 | /* | |
906f9cdf | 2727 | * Racy check if we can split the page, before unmap_page() will |
e9b61f19 KS |
2728 | * split PMDs |
2729 | */ | |
b8f593cd | 2730 | if (!can_split_huge_page(head, &extra_pins)) { |
e9b61f19 KS |
2731 | ret = -EBUSY; |
2732 | goto out_unlock; | |
2733 | } | |
2734 | ||
906f9cdf | 2735 | unmap_page(head); |
e9b61f19 | 2736 | |
b6769834 AS |
2737 | /* block interrupt reentry in xa_lock and spinlock */ |
2738 | local_irq_disable(); | |
baa355fd | 2739 | if (mapping) { |
aa5dc07f | 2740 | XA_STATE(xas, &mapping->i_pages, page_index(head)); |
baa355fd | 2741 | |
baa355fd | 2742 | /* |
aa5dc07f | 2743 | * Check if the head page is present in page cache. |
baa355fd KS |
2744 | * We assume all tail are present too, if head is there. |
2745 | */ | |
aa5dc07f MW |
2746 | xa_lock(&mapping->i_pages); |
2747 | if (xas_load(&xas) != head) | |
baa355fd KS |
2748 | goto fail; |
2749 | } | |
2750 | ||
0139aa7b | 2751 | /* Prevent deferred_split_scan() touching ->_refcount */ |
364c1eeb | 2752 | spin_lock(&ds_queue->split_queue_lock); |
504e070d | 2753 | if (page_ref_freeze(head, 1 + extra_pins)) { |
9a982250 | 2754 | if (!list_empty(page_deferred_list(head))) { |
364c1eeb | 2755 | ds_queue->split_queue_len--; |
9a982250 KS |
2756 | list_del(page_deferred_list(head)); |
2757 | } | |
afb97172 | 2758 | spin_unlock(&ds_queue->split_queue_lock); |
06d3eff6 | 2759 | if (mapping) { |
bf9ecead MS |
2760 | int nr = thp_nr_pages(head); |
2761 | ||
a8803e6c | 2762 | if (PageSwapBacked(head)) |
57b2847d MS |
2763 | __mod_lruvec_page_state(head, NR_SHMEM_THPS, |
2764 | -nr); | |
06d3eff6 | 2765 | else |
bf9ecead MS |
2766 | __mod_lruvec_page_state(head, NR_FILE_THPS, |
2767 | -nr); | |
06d3eff6 KS |
2768 | } |
2769 | ||
b6769834 | 2770 | __split_huge_page(page, list, end); |
c4f9c701 | 2771 | ret = 0; |
e9b61f19 | 2772 | } else { |
364c1eeb | 2773 | spin_unlock(&ds_queue->split_queue_lock); |
504e070d YS |
2774 | fail: |
2775 | if (mapping) | |
b93b0163 | 2776 | xa_unlock(&mapping->i_pages); |
b6769834 | 2777 | local_irq_enable(); |
8cce5475 | 2778 | remap_page(head, thp_nr_pages(head)); |
e9b61f19 KS |
2779 | ret = -EBUSY; |
2780 | } | |
2781 | ||
2782 | out_unlock: | |
baa355fd KS |
2783 | if (anon_vma) { |
2784 | anon_vma_unlock_write(anon_vma); | |
2785 | put_anon_vma(anon_vma); | |
2786 | } | |
2787 | if (mapping) | |
2788 | i_mmap_unlock_read(mapping); | |
e9b61f19 KS |
2789 | out: |
2790 | count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); | |
2791 | return ret; | |
2792 | } | |
9a982250 KS |
2793 | |
2794 | void free_transhuge_page(struct page *page) | |
2795 | { | |
87eaceb3 | 2796 | struct deferred_split *ds_queue = get_deferred_split_queue(page); |
9a982250 KS |
2797 | unsigned long flags; |
2798 | ||
364c1eeb | 2799 | spin_lock_irqsave(&ds_queue->split_queue_lock, flags); |
9a982250 | 2800 | if (!list_empty(page_deferred_list(page))) { |
364c1eeb | 2801 | ds_queue->split_queue_len--; |
9a982250 KS |
2802 | list_del(page_deferred_list(page)); |
2803 | } | |
364c1eeb | 2804 | spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); |
9a982250 KS |
2805 | free_compound_page(page); |
2806 | } | |
2807 | ||
2808 | void deferred_split_huge_page(struct page *page) | |
2809 | { | |
87eaceb3 YS |
2810 | struct deferred_split *ds_queue = get_deferred_split_queue(page); |
2811 | #ifdef CONFIG_MEMCG | |
bcfe06bf | 2812 | struct mem_cgroup *memcg = page_memcg(compound_head(page)); |
87eaceb3 | 2813 | #endif |
9a982250 KS |
2814 | unsigned long flags; |
2815 | ||
2816 | VM_BUG_ON_PAGE(!PageTransHuge(page), page); | |
2817 | ||
87eaceb3 YS |
2818 | /* |
2819 | * The try_to_unmap() in page reclaim path might reach here too, | |
2820 | * this may cause a race condition to corrupt deferred split queue. | |
2821 | * And, if page reclaim is already handling the same page, it is | |
2822 | * unnecessary to handle it again in shrinker. | |
2823 | * | |
2824 | * Check PageSwapCache to determine if the page is being | |
2825 | * handled by page reclaim since THP swap would add the page into | |
2826 | * swap cache before calling try_to_unmap(). | |
2827 | */ | |
2828 | if (PageSwapCache(page)) | |
2829 | return; | |
2830 | ||
364c1eeb | 2831 | spin_lock_irqsave(&ds_queue->split_queue_lock, flags); |
9a982250 | 2832 | if (list_empty(page_deferred_list(page))) { |
f9719a03 | 2833 | count_vm_event(THP_DEFERRED_SPLIT_PAGE); |
364c1eeb YS |
2834 | list_add_tail(page_deferred_list(page), &ds_queue->split_queue); |
2835 | ds_queue->split_queue_len++; | |
87eaceb3 YS |
2836 | #ifdef CONFIG_MEMCG |
2837 | if (memcg) | |
2bfd3637 YS |
2838 | set_shrinker_bit(memcg, page_to_nid(page), |
2839 | deferred_split_shrinker.id); | |
87eaceb3 | 2840 | #endif |
9a982250 | 2841 | } |
364c1eeb | 2842 | spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); |
9a982250 KS |
2843 | } |
2844 | ||
2845 | static unsigned long deferred_split_count(struct shrinker *shrink, | |
2846 | struct shrink_control *sc) | |
2847 | { | |
a3d0a918 | 2848 | struct pglist_data *pgdata = NODE_DATA(sc->nid); |
364c1eeb | 2849 | struct deferred_split *ds_queue = &pgdata->deferred_split_queue; |
87eaceb3 YS |
2850 | |
2851 | #ifdef CONFIG_MEMCG | |
2852 | if (sc->memcg) | |
2853 | ds_queue = &sc->memcg->deferred_split_queue; | |
2854 | #endif | |
364c1eeb | 2855 | return READ_ONCE(ds_queue->split_queue_len); |
9a982250 KS |
2856 | } |
2857 | ||
2858 | static unsigned long deferred_split_scan(struct shrinker *shrink, | |
2859 | struct shrink_control *sc) | |
2860 | { | |
a3d0a918 | 2861 | struct pglist_data *pgdata = NODE_DATA(sc->nid); |
364c1eeb | 2862 | struct deferred_split *ds_queue = &pgdata->deferred_split_queue; |
9a982250 KS |
2863 | unsigned long flags; |
2864 | LIST_HEAD(list), *pos, *next; | |
2865 | struct page *page; | |
2866 | int split = 0; | |
2867 | ||
87eaceb3 YS |
2868 | #ifdef CONFIG_MEMCG |
2869 | if (sc->memcg) | |
2870 | ds_queue = &sc->memcg->deferred_split_queue; | |
2871 | #endif | |
2872 | ||
364c1eeb | 2873 | spin_lock_irqsave(&ds_queue->split_queue_lock, flags); |
9a982250 | 2874 | /* Take pin on all head pages to avoid freeing them under us */ |
364c1eeb | 2875 | list_for_each_safe(pos, next, &ds_queue->split_queue) { |
dfe5c51c | 2876 | page = list_entry((void *)pos, struct page, deferred_list); |
9a982250 | 2877 | page = compound_head(page); |
e3ae1953 KS |
2878 | if (get_page_unless_zero(page)) { |
2879 | list_move(page_deferred_list(page), &list); | |
2880 | } else { | |
2881 | /* We lost race with put_compound_page() */ | |
9a982250 | 2882 | list_del_init(page_deferred_list(page)); |
364c1eeb | 2883 | ds_queue->split_queue_len--; |
9a982250 | 2884 | } |
e3ae1953 KS |
2885 | if (!--sc->nr_to_scan) |
2886 | break; | |
9a982250 | 2887 | } |
364c1eeb | 2888 | spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); |
9a982250 KS |
2889 | |
2890 | list_for_each_safe(pos, next, &list) { | |
dfe5c51c | 2891 | page = list_entry((void *)pos, struct page, deferred_list); |
fa41b900 KS |
2892 | if (!trylock_page(page)) |
2893 | goto next; | |
9a982250 KS |
2894 | /* split_huge_page() removes page from list on success */ |
2895 | if (!split_huge_page(page)) | |
2896 | split++; | |
2897 | unlock_page(page); | |
fa41b900 | 2898 | next: |
9a982250 KS |
2899 | put_page(page); |
2900 | } | |
2901 | ||
364c1eeb YS |
2902 | spin_lock_irqsave(&ds_queue->split_queue_lock, flags); |
2903 | list_splice_tail(&list, &ds_queue->split_queue); | |
2904 | spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags); | |
9a982250 | 2905 | |
cb8d68ec KS |
2906 | /* |
2907 | * Stop shrinker if we didn't split any page, but the queue is empty. | |
2908 | * This can happen if pages were freed under us. | |
2909 | */ | |
364c1eeb | 2910 | if (!split && list_empty(&ds_queue->split_queue)) |
cb8d68ec KS |
2911 | return SHRINK_STOP; |
2912 | return split; | |
9a982250 KS |
2913 | } |
2914 | ||
2915 | static struct shrinker deferred_split_shrinker = { | |
2916 | .count_objects = deferred_split_count, | |
2917 | .scan_objects = deferred_split_scan, | |
2918 | .seeks = DEFAULT_SEEKS, | |
87eaceb3 YS |
2919 | .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE | |
2920 | SHRINKER_NONSLAB, | |
9a982250 | 2921 | }; |
49071d43 KS |
2922 | |
2923 | #ifdef CONFIG_DEBUG_FS | |
fa6c0231 | 2924 | static void split_huge_pages_all(void) |
49071d43 KS |
2925 | { |
2926 | struct zone *zone; | |
2927 | struct page *page; | |
2928 | unsigned long pfn, max_zone_pfn; | |
2929 | unsigned long total = 0, split = 0; | |
2930 | ||
fa6c0231 | 2931 | pr_debug("Split all THPs\n"); |
49071d43 KS |
2932 | for_each_populated_zone(zone) { |
2933 | max_zone_pfn = zone_end_pfn(zone); | |
2934 | for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { | |
2935 | if (!pfn_valid(pfn)) | |
2936 | continue; | |
2937 | ||
2938 | page = pfn_to_page(pfn); | |
2939 | if (!get_page_unless_zero(page)) | |
2940 | continue; | |
2941 | ||
2942 | if (zone != page_zone(page)) | |
2943 | goto next; | |
2944 | ||
baa355fd | 2945 | if (!PageHead(page) || PageHuge(page) || !PageLRU(page)) |
49071d43 KS |
2946 | goto next; |
2947 | ||
2948 | total++; | |
2949 | lock_page(page); | |
2950 | if (!split_huge_page(page)) | |
2951 | split++; | |
2952 | unlock_page(page); | |
2953 | next: | |
2954 | put_page(page); | |
fa6c0231 | 2955 | cond_resched(); |
49071d43 KS |
2956 | } |
2957 | } | |
2958 | ||
fa6c0231 ZY |
2959 | pr_debug("%lu of %lu THP split\n", split, total); |
2960 | } | |
49071d43 | 2961 | |
fa6c0231 ZY |
2962 | static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma) |
2963 | { | |
2964 | return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) || | |
2965 | is_vm_hugetlb_page(vma); | |
2966 | } | |
2967 | ||
2968 | static int split_huge_pages_pid(int pid, unsigned long vaddr_start, | |
2969 | unsigned long vaddr_end) | |
2970 | { | |
2971 | int ret = 0; | |
2972 | struct task_struct *task; | |
2973 | struct mm_struct *mm; | |
2974 | unsigned long total = 0, split = 0; | |
2975 | unsigned long addr; | |
2976 | ||
2977 | vaddr_start &= PAGE_MASK; | |
2978 | vaddr_end &= PAGE_MASK; | |
2979 | ||
2980 | /* Find the task_struct from pid */ | |
2981 | rcu_read_lock(); | |
2982 | task = find_task_by_vpid(pid); | |
2983 | if (!task) { | |
2984 | rcu_read_unlock(); | |
2985 | ret = -ESRCH; | |
2986 | goto out; | |
2987 | } | |
2988 | get_task_struct(task); | |
2989 | rcu_read_unlock(); | |
2990 | ||
2991 | /* Find the mm_struct */ | |
2992 | mm = get_task_mm(task); | |
2993 | put_task_struct(task); | |
2994 | ||
2995 | if (!mm) { | |
2996 | ret = -EINVAL; | |
2997 | goto out; | |
2998 | } | |
2999 | ||
3000 | pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n", | |
3001 | pid, vaddr_start, vaddr_end); | |
3002 | ||
3003 | mmap_read_lock(mm); | |
3004 | /* | |
3005 | * always increase addr by PAGE_SIZE, since we could have a PTE page | |
3006 | * table filled with PTE-mapped THPs, each of which is distinct. | |
3007 | */ | |
3008 | for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) { | |
3009 | struct vm_area_struct *vma = find_vma(mm, addr); | |
3010 | unsigned int follflags; | |
3011 | struct page *page; | |
3012 | ||
3013 | if (!vma || addr < vma->vm_start) | |
3014 | break; | |
3015 | ||
3016 | /* skip special VMA and hugetlb VMA */ | |
3017 | if (vma_not_suitable_for_thp_split(vma)) { | |
3018 | addr = vma->vm_end; | |
3019 | continue; | |
3020 | } | |
3021 | ||
3022 | /* FOLL_DUMP to ignore special (like zero) pages */ | |
3023 | follflags = FOLL_GET | FOLL_DUMP; | |
3024 | page = follow_page(vma, addr, follflags); | |
3025 | ||
3026 | if (IS_ERR(page)) | |
3027 | continue; | |
3028 | if (!page) | |
3029 | continue; | |
3030 | ||
3031 | if (!is_transparent_hugepage(page)) | |
3032 | goto next; | |
3033 | ||
3034 | total++; | |
3035 | if (!can_split_huge_page(compound_head(page), NULL)) | |
3036 | goto next; | |
3037 | ||
3038 | if (!trylock_page(page)) | |
3039 | goto next; | |
3040 | ||
3041 | if (!split_huge_page(page)) | |
3042 | split++; | |
3043 | ||
3044 | unlock_page(page); | |
3045 | next: | |
3046 | put_page(page); | |
3047 | cond_resched(); | |
3048 | } | |
3049 | mmap_read_unlock(mm); | |
3050 | mmput(mm); | |
3051 | ||
3052 | pr_debug("%lu of %lu THP split\n", split, total); | |
3053 | ||
3054 | out: | |
3055 | return ret; | |
49071d43 | 3056 | } |
fa6c0231 | 3057 | |
fbe37501 ZY |
3058 | static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start, |
3059 | pgoff_t off_end) | |
3060 | { | |
3061 | struct filename *file; | |
3062 | struct file *candidate; | |
3063 | struct address_space *mapping; | |
3064 | int ret = -EINVAL; | |
3065 | pgoff_t index; | |
3066 | int nr_pages = 1; | |
3067 | unsigned long total = 0, split = 0; | |
3068 | ||
3069 | file = getname_kernel(file_path); | |
3070 | if (IS_ERR(file)) | |
3071 | return ret; | |
3072 | ||
3073 | candidate = file_open_name(file, O_RDONLY, 0); | |
3074 | if (IS_ERR(candidate)) | |
3075 | goto out; | |
3076 | ||
3077 | pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n", | |
3078 | file_path, off_start, off_end); | |
3079 | ||
3080 | mapping = candidate->f_mapping; | |
3081 | ||
3082 | for (index = off_start; index < off_end; index += nr_pages) { | |
3083 | struct page *fpage = pagecache_get_page(mapping, index, | |
3084 | FGP_ENTRY | FGP_HEAD, 0); | |
3085 | ||
3086 | nr_pages = 1; | |
3087 | if (xa_is_value(fpage) || !fpage) | |
3088 | continue; | |
3089 | ||
3090 | if (!is_transparent_hugepage(fpage)) | |
3091 | goto next; | |
3092 | ||
3093 | total++; | |
3094 | nr_pages = thp_nr_pages(fpage); | |
3095 | ||
3096 | if (!trylock_page(fpage)) | |
3097 | goto next; | |
3098 | ||
3099 | if (!split_huge_page(fpage)) | |
3100 | split++; | |
3101 | ||
3102 | unlock_page(fpage); | |
3103 | next: | |
3104 | put_page(fpage); | |
3105 | cond_resched(); | |
3106 | } | |
3107 | ||
3108 | filp_close(candidate, NULL); | |
3109 | ret = 0; | |
3110 | ||
3111 | pr_debug("%lu of %lu file-backed THP split\n", split, total); | |
3112 | out: | |
3113 | putname(file); | |
3114 | return ret; | |
3115 | } | |
3116 | ||
fa6c0231 ZY |
3117 | #define MAX_INPUT_BUF_SZ 255 |
3118 | ||
3119 | static ssize_t split_huge_pages_write(struct file *file, const char __user *buf, | |
3120 | size_t count, loff_t *ppops) | |
3121 | { | |
3122 | static DEFINE_MUTEX(split_debug_mutex); | |
3123 | ssize_t ret; | |
fbe37501 ZY |
3124 | /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */ |
3125 | char input_buf[MAX_INPUT_BUF_SZ]; | |
fa6c0231 ZY |
3126 | int pid; |
3127 | unsigned long vaddr_start, vaddr_end; | |
3128 | ||
3129 | ret = mutex_lock_interruptible(&split_debug_mutex); | |
3130 | if (ret) | |
3131 | return ret; | |
3132 | ||
3133 | ret = -EFAULT; | |
3134 | ||
3135 | memset(input_buf, 0, MAX_INPUT_BUF_SZ); | |
3136 | if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ))) | |
3137 | goto out; | |
3138 | ||
3139 | input_buf[MAX_INPUT_BUF_SZ - 1] = '\0'; | |
fbe37501 ZY |
3140 | |
3141 | if (input_buf[0] == '/') { | |
3142 | char *tok; | |
3143 | char *buf = input_buf; | |
3144 | char file_path[MAX_INPUT_BUF_SZ]; | |
3145 | pgoff_t off_start = 0, off_end = 0; | |
3146 | size_t input_len = strlen(input_buf); | |
3147 | ||
3148 | tok = strsep(&buf, ","); | |
3149 | if (tok) { | |
3150 | strncpy(file_path, tok, MAX_INPUT_BUF_SZ); | |
3151 | } else { | |
3152 | ret = -EINVAL; | |
3153 | goto out; | |
3154 | } | |
3155 | ||
3156 | ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end); | |
3157 | if (ret != 2) { | |
3158 | ret = -EINVAL; | |
3159 | goto out; | |
3160 | } | |
3161 | ret = split_huge_pages_in_file(file_path, off_start, off_end); | |
3162 | if (!ret) | |
3163 | ret = input_len; | |
3164 | ||
3165 | goto out; | |
3166 | } | |
3167 | ||
fa6c0231 ZY |
3168 | ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end); |
3169 | if (ret == 1 && pid == 1) { | |
3170 | split_huge_pages_all(); | |
3171 | ret = strlen(input_buf); | |
3172 | goto out; | |
3173 | } else if (ret != 3) { | |
3174 | ret = -EINVAL; | |
3175 | goto out; | |
3176 | } | |
3177 | ||
3178 | ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end); | |
3179 | if (!ret) | |
3180 | ret = strlen(input_buf); | |
3181 | out: | |
3182 | mutex_unlock(&split_debug_mutex); | |
3183 | return ret; | |
3184 | ||
3185 | } | |
3186 | ||
3187 | static const struct file_operations split_huge_pages_fops = { | |
3188 | .owner = THIS_MODULE, | |
3189 | .write = split_huge_pages_write, | |
3190 | .llseek = no_llseek, | |
3191 | }; | |
49071d43 KS |
3192 | |
3193 | static int __init split_huge_pages_debugfs(void) | |
3194 | { | |
d9f7979c GKH |
3195 | debugfs_create_file("split_huge_pages", 0200, NULL, NULL, |
3196 | &split_huge_pages_fops); | |
49071d43 KS |
3197 | return 0; |
3198 | } | |
3199 | late_initcall(split_huge_pages_debugfs); | |
3200 | #endif | |
616b8371 ZY |
3201 | |
3202 | #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION | |
3203 | void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw, | |
3204 | struct page *page) | |
3205 | { | |
3206 | struct vm_area_struct *vma = pvmw->vma; | |
3207 | struct mm_struct *mm = vma->vm_mm; | |
3208 | unsigned long address = pvmw->address; | |
3209 | pmd_t pmdval; | |
3210 | swp_entry_t entry; | |
ab6e3d09 | 3211 | pmd_t pmdswp; |
616b8371 ZY |
3212 | |
3213 | if (!(pvmw->pmd && !pvmw->pte)) | |
3214 | return; | |
3215 | ||
616b8371 | 3216 | flush_cache_range(vma, address, address + HPAGE_PMD_SIZE); |
8a8683ad | 3217 | pmdval = pmdp_invalidate(vma, address, pvmw->pmd); |
616b8371 ZY |
3218 | if (pmd_dirty(pmdval)) |
3219 | set_page_dirty(page); | |
3220 | entry = make_migration_entry(page, pmd_write(pmdval)); | |
ab6e3d09 NH |
3221 | pmdswp = swp_entry_to_pmd(entry); |
3222 | if (pmd_soft_dirty(pmdval)) | |
3223 | pmdswp = pmd_swp_mksoft_dirty(pmdswp); | |
3224 | set_pmd_at(mm, address, pvmw->pmd, pmdswp); | |
616b8371 ZY |
3225 | page_remove_rmap(page, true); |
3226 | put_page(page); | |
616b8371 ZY |
3227 | } |
3228 | ||
3229 | void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new) | |
3230 | { | |
3231 | struct vm_area_struct *vma = pvmw->vma; | |
3232 | struct mm_struct *mm = vma->vm_mm; | |
3233 | unsigned long address = pvmw->address; | |
3234 | unsigned long mmun_start = address & HPAGE_PMD_MASK; | |
3235 | pmd_t pmde; | |
3236 | swp_entry_t entry; | |
3237 | ||
3238 | if (!(pvmw->pmd && !pvmw->pte)) | |
3239 | return; | |
3240 | ||
3241 | entry = pmd_to_swp_entry(*pvmw->pmd); | |
3242 | get_page(new); | |
3243 | pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot)); | |
ab6e3d09 NH |
3244 | if (pmd_swp_soft_dirty(*pvmw->pmd)) |
3245 | pmde = pmd_mksoft_dirty(pmde); | |
616b8371 | 3246 | if (is_write_migration_entry(entry)) |
f55e1014 | 3247 | pmde = maybe_pmd_mkwrite(pmde, vma); |
616b8371 ZY |
3248 | |
3249 | flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE); | |
e71769ae NH |
3250 | if (PageAnon(new)) |
3251 | page_add_anon_rmap(new, vma, mmun_start, true); | |
3252 | else | |
3253 | page_add_file_rmap(new, true); | |
616b8371 | 3254 | set_pmd_at(mm, mmun_start, pvmw->pmd, pmde); |
e125fe40 | 3255 | if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new)) |
616b8371 ZY |
3256 | mlock_vma_page(new); |
3257 | update_mmu_cache_pmd(vma, address, pvmw->pmd); | |
3258 | } | |
3259 | #endif |