]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Copyright (C) 2009 Red Hat, Inc. | |
3 | * | |
4 | * This work is licensed under the terms of the GNU GPL, version 2. See | |
5 | * the COPYING file in the top-level directory. | |
6 | */ | |
7 | ||
8 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt | |
9 | ||
10 | #include <linux/mm.h> | |
11 | #include <linux/sched.h> | |
12 | #include <linux/sched/coredump.h> | |
13 | #include <linux/sched/numa_balancing.h> | |
14 | #include <linux/highmem.h> | |
15 | #include <linux/hugetlb.h> | |
16 | #include <linux/mmu_notifier.h> | |
17 | #include <linux/rmap.h> | |
18 | #include <linux/swap.h> | |
19 | #include <linux/shrinker.h> | |
20 | #include <linux/mm_inline.h> | |
21 | #include <linux/swapops.h> | |
22 | #include <linux/dax.h> | |
23 | #include <linux/khugepaged.h> | |
24 | #include <linux/freezer.h> | |
25 | #include <linux/pfn_t.h> | |
26 | #include <linux/mman.h> | |
27 | #include <linux/memremap.h> | |
28 | #include <linux/pagemap.h> | |
29 | #include <linux/debugfs.h> | |
30 | #include <linux/migrate.h> | |
31 | #include <linux/hashtable.h> | |
32 | #include <linux/userfaultfd_k.h> | |
33 | #include <linux/page_idle.h> | |
34 | #include <linux/shmem_fs.h> | |
35 | #include <linux/oom.h> | |
36 | ||
37 | #include <asm/tlb.h> | |
38 | #include <asm/pgalloc.h> | |
39 | #include "internal.h" | |
40 | ||
41 | /* | |
42 | * By default transparent hugepage support is disabled in order that avoid | |
43 | * to risk increase the memory footprint of applications without a guaranteed | |
44 | * benefit. When transparent hugepage support is enabled, is for all mappings, | |
45 | * and khugepaged scans all mappings. | |
46 | * Defrag is invoked by khugepaged hugepage allocations and by page faults | |
47 | * for all hugepage allocations. | |
48 | */ | |
49 | unsigned long transparent_hugepage_flags __read_mostly = | |
50 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS | |
51 | (1<<TRANSPARENT_HUGEPAGE_FLAG)| | |
52 | #endif | |
53 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE | |
54 | (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)| | |
55 | #endif | |
56 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)| | |
57 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)| | |
58 | (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
59 | ||
60 | static struct shrinker deferred_split_shrinker; | |
61 | ||
62 | static atomic_t huge_zero_refcount; | |
63 | struct page *huge_zero_page __read_mostly; | |
64 | ||
65 | static struct page *get_huge_zero_page(void) | |
66 | { | |
67 | struct page *zero_page; | |
68 | retry: | |
69 | if (likely(atomic_inc_not_zero(&huge_zero_refcount))) | |
70 | return READ_ONCE(huge_zero_page); | |
71 | ||
72 | zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE, | |
73 | HPAGE_PMD_ORDER); | |
74 | if (!zero_page) { | |
75 | count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED); | |
76 | return NULL; | |
77 | } | |
78 | count_vm_event(THP_ZERO_PAGE_ALLOC); | |
79 | preempt_disable(); | |
80 | if (cmpxchg(&huge_zero_page, NULL, zero_page)) { | |
81 | preempt_enable(); | |
82 | __free_pages(zero_page, compound_order(zero_page)); | |
83 | goto retry; | |
84 | } | |
85 | ||
86 | /* We take additional reference here. It will be put back by shrinker */ | |
87 | atomic_set(&huge_zero_refcount, 2); | |
88 | preempt_enable(); | |
89 | return READ_ONCE(huge_zero_page); | |
90 | } | |
91 | ||
92 | static void put_huge_zero_page(void) | |
93 | { | |
94 | /* | |
95 | * Counter should never go to zero here. Only shrinker can put | |
96 | * last reference. | |
97 | */ | |
98 | BUG_ON(atomic_dec_and_test(&huge_zero_refcount)); | |
99 | } | |
100 | ||
101 | struct page *mm_get_huge_zero_page(struct mm_struct *mm) | |
102 | { | |
103 | if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) | |
104 | return READ_ONCE(huge_zero_page); | |
105 | ||
106 | if (!get_huge_zero_page()) | |
107 | return NULL; | |
108 | ||
109 | if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) | |
110 | put_huge_zero_page(); | |
111 | ||
112 | return READ_ONCE(huge_zero_page); | |
113 | } | |
114 | ||
115 | void mm_put_huge_zero_page(struct mm_struct *mm) | |
116 | { | |
117 | if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags)) | |
118 | put_huge_zero_page(); | |
119 | } | |
120 | ||
121 | static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink, | |
122 | struct shrink_control *sc) | |
123 | { | |
124 | /* we can free zero page only if last reference remains */ | |
125 | return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0; | |
126 | } | |
127 | ||
128 | static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink, | |
129 | struct shrink_control *sc) | |
130 | { | |
131 | if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) { | |
132 | struct page *zero_page = xchg(&huge_zero_page, NULL); | |
133 | BUG_ON(zero_page == NULL); | |
134 | __free_pages(zero_page, compound_order(zero_page)); | |
135 | return HPAGE_PMD_NR; | |
136 | } | |
137 | ||
138 | return 0; | |
139 | } | |
140 | ||
141 | static struct shrinker huge_zero_page_shrinker = { | |
142 | .count_objects = shrink_huge_zero_page_count, | |
143 | .scan_objects = shrink_huge_zero_page_scan, | |
144 | .seeks = DEFAULT_SEEKS, | |
145 | }; | |
146 | ||
147 | #ifdef CONFIG_SYSFS | |
148 | static ssize_t enabled_show(struct kobject *kobj, | |
149 | struct kobj_attribute *attr, char *buf) | |
150 | { | |
151 | if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags)) | |
152 | return sprintf(buf, "[always] madvise never\n"); | |
153 | else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags)) | |
154 | return sprintf(buf, "always [madvise] never\n"); | |
155 | else | |
156 | return sprintf(buf, "always madvise [never]\n"); | |
157 | } | |
158 | ||
159 | static ssize_t enabled_store(struct kobject *kobj, | |
160 | struct kobj_attribute *attr, | |
161 | const char *buf, size_t count) | |
162 | { | |
163 | ssize_t ret = count; | |
164 | ||
165 | if (!memcmp("always", buf, | |
166 | min(sizeof("always")-1, count))) { | |
167 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
168 | set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); | |
169 | } else if (!memcmp("madvise", buf, | |
170 | min(sizeof("madvise")-1, count))) { | |
171 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); | |
172 | set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
173 | } else if (!memcmp("never", buf, | |
174 | min(sizeof("never")-1, count))) { | |
175 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags); | |
176 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
177 | } else | |
178 | ret = -EINVAL; | |
179 | ||
180 | if (ret > 0) { | |
181 | int err = start_stop_khugepaged(); | |
182 | if (err) | |
183 | ret = err; | |
184 | } | |
185 | return ret; | |
186 | } | |
187 | static struct kobj_attribute enabled_attr = | |
188 | __ATTR(enabled, 0644, enabled_show, enabled_store); | |
189 | ||
190 | ssize_t single_hugepage_flag_show(struct kobject *kobj, | |
191 | struct kobj_attribute *attr, char *buf, | |
192 | enum transparent_hugepage_flag flag) | |
193 | { | |
194 | return sprintf(buf, "%d\n", | |
195 | !!test_bit(flag, &transparent_hugepage_flags)); | |
196 | } | |
197 | ||
198 | ssize_t single_hugepage_flag_store(struct kobject *kobj, | |
199 | struct kobj_attribute *attr, | |
200 | const char *buf, size_t count, | |
201 | enum transparent_hugepage_flag flag) | |
202 | { | |
203 | unsigned long value; | |
204 | int ret; | |
205 | ||
206 | ret = kstrtoul(buf, 10, &value); | |
207 | if (ret < 0) | |
208 | return ret; | |
209 | if (value > 1) | |
210 | return -EINVAL; | |
211 | ||
212 | if (value) | |
213 | set_bit(flag, &transparent_hugepage_flags); | |
214 | else | |
215 | clear_bit(flag, &transparent_hugepage_flags); | |
216 | ||
217 | return count; | |
218 | } | |
219 | ||
220 | static ssize_t defrag_show(struct kobject *kobj, | |
221 | struct kobj_attribute *attr, char *buf) | |
222 | { | |
223 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) | |
224 | return sprintf(buf, "[always] defer defer+madvise madvise never\n"); | |
225 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) | |
226 | return sprintf(buf, "always [defer] defer+madvise madvise never\n"); | |
227 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) | |
228 | return sprintf(buf, "always defer [defer+madvise] madvise never\n"); | |
229 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) | |
230 | return sprintf(buf, "always defer defer+madvise [madvise] never\n"); | |
231 | return sprintf(buf, "always defer defer+madvise madvise [never]\n"); | |
232 | } | |
233 | ||
234 | static ssize_t defrag_store(struct kobject *kobj, | |
235 | struct kobj_attribute *attr, | |
236 | const char *buf, size_t count) | |
237 | { | |
238 | if (!memcmp("always", buf, | |
239 | min(sizeof("always")-1, count))) { | |
240 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
241 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
242 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
243 | set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); | |
244 | } else if (!memcmp("defer+madvise", buf, | |
245 | min(sizeof("defer+madvise")-1, count))) { | |
246 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); | |
247 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
248 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
249 | set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
250 | } else if (!memcmp("defer", buf, | |
251 | min(sizeof("defer")-1, count))) { | |
252 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); | |
253 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
254 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
255 | set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
256 | } else if (!memcmp("madvise", buf, | |
257 | min(sizeof("madvise")-1, count))) { | |
258 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); | |
259 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
260 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
261 | set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
262 | } else if (!memcmp("never", buf, | |
263 | min(sizeof("never")-1, count))) { | |
264 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags); | |
265 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags); | |
266 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags); | |
267 | clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags); | |
268 | } else | |
269 | return -EINVAL; | |
270 | ||
271 | return count; | |
272 | } | |
273 | static struct kobj_attribute defrag_attr = | |
274 | __ATTR(defrag, 0644, defrag_show, defrag_store); | |
275 | ||
276 | static ssize_t use_zero_page_show(struct kobject *kobj, | |
277 | struct kobj_attribute *attr, char *buf) | |
278 | { | |
279 | return single_hugepage_flag_show(kobj, attr, buf, | |
280 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
281 | } | |
282 | static ssize_t use_zero_page_store(struct kobject *kobj, | |
283 | struct kobj_attribute *attr, const char *buf, size_t count) | |
284 | { | |
285 | return single_hugepage_flag_store(kobj, attr, buf, count, | |
286 | TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG); | |
287 | } | |
288 | static struct kobj_attribute use_zero_page_attr = | |
289 | __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store); | |
290 | ||
291 | static ssize_t hpage_pmd_size_show(struct kobject *kobj, | |
292 | struct kobj_attribute *attr, char *buf) | |
293 | { | |
294 | return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE); | |
295 | } | |
296 | static struct kobj_attribute hpage_pmd_size_attr = | |
297 | __ATTR_RO(hpage_pmd_size); | |
298 | ||
299 | #ifdef CONFIG_DEBUG_VM | |
300 | static ssize_t debug_cow_show(struct kobject *kobj, | |
301 | struct kobj_attribute *attr, char *buf) | |
302 | { | |
303 | return single_hugepage_flag_show(kobj, attr, buf, | |
304 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); | |
305 | } | |
306 | static ssize_t debug_cow_store(struct kobject *kobj, | |
307 | struct kobj_attribute *attr, | |
308 | const char *buf, size_t count) | |
309 | { | |
310 | return single_hugepage_flag_store(kobj, attr, buf, count, | |
311 | TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG); | |
312 | } | |
313 | static struct kobj_attribute debug_cow_attr = | |
314 | __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store); | |
315 | #endif /* CONFIG_DEBUG_VM */ | |
316 | ||
317 | static struct attribute *hugepage_attr[] = { | |
318 | &enabled_attr.attr, | |
319 | &defrag_attr.attr, | |
320 | &use_zero_page_attr.attr, | |
321 | &hpage_pmd_size_attr.attr, | |
322 | #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE) | |
323 | &shmem_enabled_attr.attr, | |
324 | #endif | |
325 | #ifdef CONFIG_DEBUG_VM | |
326 | &debug_cow_attr.attr, | |
327 | #endif | |
328 | NULL, | |
329 | }; | |
330 | ||
331 | static struct attribute_group hugepage_attr_group = { | |
332 | .attrs = hugepage_attr, | |
333 | }; | |
334 | ||
335 | static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj) | |
336 | { | |
337 | int err; | |
338 | ||
339 | *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj); | |
340 | if (unlikely(!*hugepage_kobj)) { | |
341 | pr_err("failed to create transparent hugepage kobject\n"); | |
342 | return -ENOMEM; | |
343 | } | |
344 | ||
345 | err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group); | |
346 | if (err) { | |
347 | pr_err("failed to register transparent hugepage group\n"); | |
348 | goto delete_obj; | |
349 | } | |
350 | ||
351 | err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group); | |
352 | if (err) { | |
353 | pr_err("failed to register transparent hugepage group\n"); | |
354 | goto remove_hp_group; | |
355 | } | |
356 | ||
357 | return 0; | |
358 | ||
359 | remove_hp_group: | |
360 | sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group); | |
361 | delete_obj: | |
362 | kobject_put(*hugepage_kobj); | |
363 | return err; | |
364 | } | |
365 | ||
366 | static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
367 | { | |
368 | sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group); | |
369 | sysfs_remove_group(hugepage_kobj, &hugepage_attr_group); | |
370 | kobject_put(hugepage_kobj); | |
371 | } | |
372 | #else | |
373 | static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj) | |
374 | { | |
375 | return 0; | |
376 | } | |
377 | ||
378 | static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj) | |
379 | { | |
380 | } | |
381 | #endif /* CONFIG_SYSFS */ | |
382 | ||
383 | static int __init hugepage_init(void) | |
384 | { | |
385 | int err; | |
386 | struct kobject *hugepage_kobj; | |
387 | ||
388 | if (!has_transparent_hugepage()) { | |
389 | transparent_hugepage_flags = 0; | |
390 | return -EINVAL; | |
391 | } | |
392 | ||
393 | /* | |
394 | * hugepages can't be allocated by the buddy allocator | |
395 | */ | |
396 | MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER); | |
397 | /* | |
398 | * we use page->mapping and page->index in second tail page | |
399 | * as list_head: assuming THP order >= 2 | |
400 | */ | |
401 | MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2); | |
402 | ||
403 | err = hugepage_init_sysfs(&hugepage_kobj); | |
404 | if (err) | |
405 | goto err_sysfs; | |
406 | ||
407 | err = khugepaged_init(); | |
408 | if (err) | |
409 | goto err_slab; | |
410 | ||
411 | err = register_shrinker(&huge_zero_page_shrinker); | |
412 | if (err) | |
413 | goto err_hzp_shrinker; | |
414 | err = register_shrinker(&deferred_split_shrinker); | |
415 | if (err) | |
416 | goto err_split_shrinker; | |
417 | ||
418 | /* | |
419 | * By default disable transparent hugepages on smaller systems, | |
420 | * where the extra memory used could hurt more than TLB overhead | |
421 | * is likely to save. The admin can still enable it through /sys. | |
422 | */ | |
423 | if (totalram_pages < (512 << (20 - PAGE_SHIFT))) { | |
424 | transparent_hugepage_flags = 0; | |
425 | return 0; | |
426 | } | |
427 | ||
428 | err = start_stop_khugepaged(); | |
429 | if (err) | |
430 | goto err_khugepaged; | |
431 | ||
432 | return 0; | |
433 | err_khugepaged: | |
434 | unregister_shrinker(&deferred_split_shrinker); | |
435 | err_split_shrinker: | |
436 | unregister_shrinker(&huge_zero_page_shrinker); | |
437 | err_hzp_shrinker: | |
438 | khugepaged_destroy(); | |
439 | err_slab: | |
440 | hugepage_exit_sysfs(hugepage_kobj); | |
441 | err_sysfs: | |
442 | return err; | |
443 | } | |
444 | subsys_initcall(hugepage_init); | |
445 | ||
446 | static int __init setup_transparent_hugepage(char *str) | |
447 | { | |
448 | int ret = 0; | |
449 | if (!str) | |
450 | goto out; | |
451 | if (!strcmp(str, "always")) { | |
452 | set_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
453 | &transparent_hugepage_flags); | |
454 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
455 | &transparent_hugepage_flags); | |
456 | ret = 1; | |
457 | } else if (!strcmp(str, "madvise")) { | |
458 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
459 | &transparent_hugepage_flags); | |
460 | set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
461 | &transparent_hugepage_flags); | |
462 | ret = 1; | |
463 | } else if (!strcmp(str, "never")) { | |
464 | clear_bit(TRANSPARENT_HUGEPAGE_FLAG, | |
465 | &transparent_hugepage_flags); | |
466 | clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, | |
467 | &transparent_hugepage_flags); | |
468 | ret = 1; | |
469 | } | |
470 | out: | |
471 | if (!ret) | |
472 | pr_warn("transparent_hugepage= cannot parse, ignored\n"); | |
473 | return ret; | |
474 | } | |
475 | __setup("transparent_hugepage=", setup_transparent_hugepage); | |
476 | ||
477 | pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma) | |
478 | { | |
479 | if (likely(vma->vm_flags & VM_WRITE)) | |
480 | pmd = pmd_mkwrite(pmd); | |
481 | return pmd; | |
482 | } | |
483 | ||
484 | static inline struct list_head *page_deferred_list(struct page *page) | |
485 | { | |
486 | /* | |
487 | * ->lru in the tail pages is occupied by compound_head. | |
488 | * Let's use ->mapping + ->index in the second tail page as list_head. | |
489 | */ | |
490 | return (struct list_head *)&page[2].mapping; | |
491 | } | |
492 | ||
493 | void prep_transhuge_page(struct page *page) | |
494 | { | |
495 | /* | |
496 | * we use page->mapping and page->indexlru in second tail page | |
497 | * as list_head: assuming THP order >= 2 | |
498 | */ | |
499 | ||
500 | INIT_LIST_HEAD(page_deferred_list(page)); | |
501 | set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR); | |
502 | } | |
503 | ||
504 | unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len, | |
505 | loff_t off, unsigned long flags, unsigned long size) | |
506 | { | |
507 | unsigned long addr; | |
508 | loff_t off_end = off + len; | |
509 | loff_t off_align = round_up(off, size); | |
510 | unsigned long len_pad; | |
511 | ||
512 | if (off_end <= off_align || (off_end - off_align) < size) | |
513 | return 0; | |
514 | ||
515 | len_pad = len + size; | |
516 | if (len_pad < len || (off + len_pad) < off) | |
517 | return 0; | |
518 | ||
519 | addr = current->mm->get_unmapped_area(filp, 0, len_pad, | |
520 | off >> PAGE_SHIFT, flags); | |
521 | if (IS_ERR_VALUE(addr)) | |
522 | return 0; | |
523 | ||
524 | addr += (off - addr) & (size - 1); | |
525 | return addr; | |
526 | } | |
527 | ||
528 | unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr, | |
529 | unsigned long len, unsigned long pgoff, unsigned long flags) | |
530 | { | |
531 | loff_t off = (loff_t)pgoff << PAGE_SHIFT; | |
532 | ||
533 | if (addr) | |
534 | goto out; | |
535 | if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD)) | |
536 | goto out; | |
537 | ||
538 | addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE); | |
539 | if (addr) | |
540 | return addr; | |
541 | ||
542 | out: | |
543 | return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags); | |
544 | } | |
545 | EXPORT_SYMBOL_GPL(thp_get_unmapped_area); | |
546 | ||
547 | static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page, | |
548 | gfp_t gfp) | |
549 | { | |
550 | struct vm_area_struct *vma = vmf->vma; | |
551 | struct mem_cgroup *memcg; | |
552 | pgtable_t pgtable; | |
553 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; | |
554 | int ret = 0; | |
555 | ||
556 | VM_BUG_ON_PAGE(!PageCompound(page), page); | |
557 | ||
558 | if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) { | |
559 | put_page(page); | |
560 | count_vm_event(THP_FAULT_FALLBACK); | |
561 | return VM_FAULT_FALLBACK; | |
562 | } | |
563 | ||
564 | pgtable = pte_alloc_one(vma->vm_mm, haddr); | |
565 | if (unlikely(!pgtable)) { | |
566 | ret = VM_FAULT_OOM; | |
567 | goto release; | |
568 | } | |
569 | ||
570 | clear_huge_page(page, haddr, HPAGE_PMD_NR); | |
571 | /* | |
572 | * The memory barrier inside __SetPageUptodate makes sure that | |
573 | * clear_huge_page writes become visible before the set_pmd_at() | |
574 | * write. | |
575 | */ | |
576 | __SetPageUptodate(page); | |
577 | ||
578 | vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); | |
579 | if (unlikely(!pmd_none(*vmf->pmd))) { | |
580 | goto unlock_release; | |
581 | } else { | |
582 | pmd_t entry; | |
583 | ||
584 | ret = check_stable_address_space(vma->vm_mm); | |
585 | if (ret) | |
586 | goto unlock_release; | |
587 | ||
588 | /* Deliver the page fault to userland */ | |
589 | if (userfaultfd_missing(vma)) { | |
590 | int ret; | |
591 | ||
592 | spin_unlock(vmf->ptl); | |
593 | mem_cgroup_cancel_charge(page, memcg, true); | |
594 | put_page(page); | |
595 | pte_free(vma->vm_mm, pgtable); | |
596 | ret = handle_userfault(vmf, VM_UFFD_MISSING); | |
597 | VM_BUG_ON(ret & VM_FAULT_FALLBACK); | |
598 | return ret; | |
599 | } | |
600 | ||
601 | entry = mk_huge_pmd(page, vma->vm_page_prot); | |
602 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
603 | page_add_new_anon_rmap(page, vma, haddr, true); | |
604 | mem_cgroup_commit_charge(page, memcg, false, true); | |
605 | lru_cache_add_active_or_unevictable(page, vma); | |
606 | pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable); | |
607 | set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); | |
608 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); | |
609 | atomic_long_inc(&vma->vm_mm->nr_ptes); | |
610 | spin_unlock(vmf->ptl); | |
611 | count_vm_event(THP_FAULT_ALLOC); | |
612 | } | |
613 | ||
614 | return 0; | |
615 | unlock_release: | |
616 | spin_unlock(vmf->ptl); | |
617 | release: | |
618 | if (pgtable) | |
619 | pte_free(vma->vm_mm, pgtable); | |
620 | mem_cgroup_cancel_charge(page, memcg, true); | |
621 | put_page(page); | |
622 | return ret; | |
623 | ||
624 | } | |
625 | ||
626 | /* | |
627 | * always: directly stall for all thp allocations | |
628 | * defer: wake kswapd and fail if not immediately available | |
629 | * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise | |
630 | * fail if not immediately available | |
631 | * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately | |
632 | * available | |
633 | * never: never stall for any thp allocation | |
634 | */ | |
635 | static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma) | |
636 | { | |
637 | const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE); | |
638 | ||
639 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags)) | |
640 | return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY); | |
641 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags)) | |
642 | return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM; | |
643 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags)) | |
644 | return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM : | |
645 | __GFP_KSWAPD_RECLAIM); | |
646 | if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags)) | |
647 | return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM : | |
648 | 0); | |
649 | return GFP_TRANSHUGE_LIGHT; | |
650 | } | |
651 | ||
652 | /* Caller must hold page table lock. */ | |
653 | static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm, | |
654 | struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd, | |
655 | struct page *zero_page) | |
656 | { | |
657 | pmd_t entry; | |
658 | if (!pmd_none(*pmd)) | |
659 | return false; | |
660 | entry = mk_pmd(zero_page, vma->vm_page_prot); | |
661 | entry = pmd_mkhuge(entry); | |
662 | if (pgtable) | |
663 | pgtable_trans_huge_deposit(mm, pmd, pgtable); | |
664 | set_pmd_at(mm, haddr, pmd, entry); | |
665 | atomic_long_inc(&mm->nr_ptes); | |
666 | return true; | |
667 | } | |
668 | ||
669 | int do_huge_pmd_anonymous_page(struct vm_fault *vmf) | |
670 | { | |
671 | struct vm_area_struct *vma = vmf->vma; | |
672 | gfp_t gfp; | |
673 | struct page *page; | |
674 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; | |
675 | ||
676 | if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end) | |
677 | return VM_FAULT_FALLBACK; | |
678 | if (unlikely(anon_vma_prepare(vma))) | |
679 | return VM_FAULT_OOM; | |
680 | if (unlikely(khugepaged_enter(vma, vma->vm_flags))) | |
681 | return VM_FAULT_OOM; | |
682 | if (!(vmf->flags & FAULT_FLAG_WRITE) && | |
683 | !mm_forbids_zeropage(vma->vm_mm) && | |
684 | transparent_hugepage_use_zero_page()) { | |
685 | pgtable_t pgtable; | |
686 | struct page *zero_page; | |
687 | bool set; | |
688 | int ret; | |
689 | pgtable = pte_alloc_one(vma->vm_mm, haddr); | |
690 | if (unlikely(!pgtable)) | |
691 | return VM_FAULT_OOM; | |
692 | zero_page = mm_get_huge_zero_page(vma->vm_mm); | |
693 | if (unlikely(!zero_page)) { | |
694 | pte_free(vma->vm_mm, pgtable); | |
695 | count_vm_event(THP_FAULT_FALLBACK); | |
696 | return VM_FAULT_FALLBACK; | |
697 | } | |
698 | vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); | |
699 | ret = 0; | |
700 | set = false; | |
701 | if (pmd_none(*vmf->pmd)) { | |
702 | ret = check_stable_address_space(vma->vm_mm); | |
703 | if (ret) { | |
704 | spin_unlock(vmf->ptl); | |
705 | } else if (userfaultfd_missing(vma)) { | |
706 | spin_unlock(vmf->ptl); | |
707 | ret = handle_userfault(vmf, VM_UFFD_MISSING); | |
708 | VM_BUG_ON(ret & VM_FAULT_FALLBACK); | |
709 | } else { | |
710 | set_huge_zero_page(pgtable, vma->vm_mm, vma, | |
711 | haddr, vmf->pmd, zero_page); | |
712 | spin_unlock(vmf->ptl); | |
713 | set = true; | |
714 | } | |
715 | } else | |
716 | spin_unlock(vmf->ptl); | |
717 | if (!set) | |
718 | pte_free(vma->vm_mm, pgtable); | |
719 | return ret; | |
720 | } | |
721 | gfp = alloc_hugepage_direct_gfpmask(vma); | |
722 | page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER); | |
723 | if (unlikely(!page)) { | |
724 | count_vm_event(THP_FAULT_FALLBACK); | |
725 | return VM_FAULT_FALLBACK; | |
726 | } | |
727 | prep_transhuge_page(page); | |
728 | return __do_huge_pmd_anonymous_page(vmf, page, gfp); | |
729 | } | |
730 | ||
731 | static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, | |
732 | pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write, | |
733 | pgtable_t pgtable) | |
734 | { | |
735 | struct mm_struct *mm = vma->vm_mm; | |
736 | pmd_t entry; | |
737 | spinlock_t *ptl; | |
738 | ||
739 | ptl = pmd_lock(mm, pmd); | |
740 | entry = pmd_mkhuge(pfn_t_pmd(pfn, prot)); | |
741 | if (pfn_t_devmap(pfn)) | |
742 | entry = pmd_mkdevmap(entry); | |
743 | if (write) { | |
744 | entry = pmd_mkyoung(pmd_mkdirty(entry)); | |
745 | entry = maybe_pmd_mkwrite(entry, vma); | |
746 | } | |
747 | ||
748 | if (pgtable) { | |
749 | pgtable_trans_huge_deposit(mm, pmd, pgtable); | |
750 | atomic_long_inc(&mm->nr_ptes); | |
751 | } | |
752 | ||
753 | set_pmd_at(mm, addr, pmd, entry); | |
754 | update_mmu_cache_pmd(vma, addr, pmd); | |
755 | spin_unlock(ptl); | |
756 | } | |
757 | ||
758 | int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr, | |
759 | pmd_t *pmd, pfn_t pfn, bool write) | |
760 | { | |
761 | pgprot_t pgprot = vma->vm_page_prot; | |
762 | pgtable_t pgtable = NULL; | |
763 | /* | |
764 | * If we had pmd_special, we could avoid all these restrictions, | |
765 | * but we need to be consistent with PTEs and architectures that | |
766 | * can't support a 'special' bit. | |
767 | */ | |
768 | BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); | |
769 | BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == | |
770 | (VM_PFNMAP|VM_MIXEDMAP)); | |
771 | BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); | |
772 | BUG_ON(!pfn_t_devmap(pfn)); | |
773 | ||
774 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
775 | return VM_FAULT_SIGBUS; | |
776 | ||
777 | if (arch_needs_pgtable_deposit()) { | |
778 | pgtable = pte_alloc_one(vma->vm_mm, addr); | |
779 | if (!pgtable) | |
780 | return VM_FAULT_OOM; | |
781 | } | |
782 | ||
783 | track_pfn_insert(vma, &pgprot, pfn); | |
784 | ||
785 | insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable); | |
786 | return VM_FAULT_NOPAGE; | |
787 | } | |
788 | EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd); | |
789 | ||
790 | #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD | |
791 | static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma) | |
792 | { | |
793 | if (likely(vma->vm_flags & VM_WRITE)) | |
794 | pud = pud_mkwrite(pud); | |
795 | return pud; | |
796 | } | |
797 | ||
798 | static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, | |
799 | pud_t *pud, pfn_t pfn, pgprot_t prot, bool write) | |
800 | { | |
801 | struct mm_struct *mm = vma->vm_mm; | |
802 | pud_t entry; | |
803 | spinlock_t *ptl; | |
804 | ||
805 | ptl = pud_lock(mm, pud); | |
806 | entry = pud_mkhuge(pfn_t_pud(pfn, prot)); | |
807 | if (pfn_t_devmap(pfn)) | |
808 | entry = pud_mkdevmap(entry); | |
809 | if (write) { | |
810 | entry = pud_mkyoung(pud_mkdirty(entry)); | |
811 | entry = maybe_pud_mkwrite(entry, vma); | |
812 | } | |
813 | set_pud_at(mm, addr, pud, entry); | |
814 | update_mmu_cache_pud(vma, addr, pud); | |
815 | spin_unlock(ptl); | |
816 | } | |
817 | ||
818 | int vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr, | |
819 | pud_t *pud, pfn_t pfn, bool write) | |
820 | { | |
821 | pgprot_t pgprot = vma->vm_page_prot; | |
822 | /* | |
823 | * If we had pud_special, we could avoid all these restrictions, | |
824 | * but we need to be consistent with PTEs and architectures that | |
825 | * can't support a 'special' bit. | |
826 | */ | |
827 | BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); | |
828 | BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == | |
829 | (VM_PFNMAP|VM_MIXEDMAP)); | |
830 | BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); | |
831 | BUG_ON(!pfn_t_devmap(pfn)); | |
832 | ||
833 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
834 | return VM_FAULT_SIGBUS; | |
835 | ||
836 | track_pfn_insert(vma, &pgprot, pfn); | |
837 | ||
838 | insert_pfn_pud(vma, addr, pud, pfn, pgprot, write); | |
839 | return VM_FAULT_NOPAGE; | |
840 | } | |
841 | EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud); | |
842 | #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ | |
843 | ||
844 | static void touch_pmd(struct vm_area_struct *vma, unsigned long addr, | |
845 | pmd_t *pmd, int flags) | |
846 | { | |
847 | pmd_t _pmd; | |
848 | ||
849 | _pmd = pmd_mkyoung(*pmd); | |
850 | if (flags & FOLL_WRITE) | |
851 | _pmd = pmd_mkdirty(_pmd); | |
852 | if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK, | |
853 | pmd, _pmd, flags & FOLL_WRITE)) | |
854 | update_mmu_cache_pmd(vma, addr, pmd); | |
855 | } | |
856 | ||
857 | struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr, | |
858 | pmd_t *pmd, int flags) | |
859 | { | |
860 | unsigned long pfn = pmd_pfn(*pmd); | |
861 | struct mm_struct *mm = vma->vm_mm; | |
862 | struct dev_pagemap *pgmap; | |
863 | struct page *page; | |
864 | ||
865 | assert_spin_locked(pmd_lockptr(mm, pmd)); | |
866 | ||
867 | /* | |
868 | * When we COW a devmap PMD entry, we split it into PTEs, so we should | |
869 | * not be in this function with `flags & FOLL_COW` set. | |
870 | */ | |
871 | WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set"); | |
872 | ||
873 | if (flags & FOLL_WRITE && !pmd_write(*pmd)) | |
874 | return NULL; | |
875 | ||
876 | if (pmd_present(*pmd) && pmd_devmap(*pmd)) | |
877 | /* pass */; | |
878 | else | |
879 | return NULL; | |
880 | ||
881 | if (flags & FOLL_TOUCH) | |
882 | touch_pmd(vma, addr, pmd, flags); | |
883 | ||
884 | /* | |
885 | * device mapped pages can only be returned if the | |
886 | * caller will manage the page reference count. | |
887 | */ | |
888 | if (!(flags & FOLL_GET)) | |
889 | return ERR_PTR(-EEXIST); | |
890 | ||
891 | pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT; | |
892 | pgmap = get_dev_pagemap(pfn, NULL); | |
893 | if (!pgmap) | |
894 | return ERR_PTR(-EFAULT); | |
895 | page = pfn_to_page(pfn); | |
896 | get_page(page); | |
897 | put_dev_pagemap(pgmap); | |
898 | ||
899 | return page; | |
900 | } | |
901 | ||
902 | int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm, | |
903 | pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr, | |
904 | struct vm_area_struct *vma) | |
905 | { | |
906 | spinlock_t *dst_ptl, *src_ptl; | |
907 | struct page *src_page; | |
908 | pmd_t pmd; | |
909 | pgtable_t pgtable = NULL; | |
910 | int ret = -ENOMEM; | |
911 | ||
912 | /* Skip if can be re-fill on fault */ | |
913 | if (!vma_is_anonymous(vma)) | |
914 | return 0; | |
915 | ||
916 | pgtable = pte_alloc_one(dst_mm, addr); | |
917 | if (unlikely(!pgtable)) | |
918 | goto out; | |
919 | ||
920 | dst_ptl = pmd_lock(dst_mm, dst_pmd); | |
921 | src_ptl = pmd_lockptr(src_mm, src_pmd); | |
922 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
923 | ||
924 | ret = -EAGAIN; | |
925 | pmd = *src_pmd; | |
926 | if (unlikely(!pmd_trans_huge(pmd))) { | |
927 | pte_free(dst_mm, pgtable); | |
928 | goto out_unlock; | |
929 | } | |
930 | /* | |
931 | * When page table lock is held, the huge zero pmd should not be | |
932 | * under splitting since we don't split the page itself, only pmd to | |
933 | * a page table. | |
934 | */ | |
935 | if (is_huge_zero_pmd(pmd)) { | |
936 | struct page *zero_page; | |
937 | /* | |
938 | * get_huge_zero_page() will never allocate a new page here, | |
939 | * since we already have a zero page to copy. It just takes a | |
940 | * reference. | |
941 | */ | |
942 | zero_page = mm_get_huge_zero_page(dst_mm); | |
943 | set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd, | |
944 | zero_page); | |
945 | ret = 0; | |
946 | goto out_unlock; | |
947 | } | |
948 | ||
949 | src_page = pmd_page(pmd); | |
950 | VM_BUG_ON_PAGE(!PageHead(src_page), src_page); | |
951 | get_page(src_page); | |
952 | page_dup_rmap(src_page, true); | |
953 | add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR); | |
954 | atomic_long_inc(&dst_mm->nr_ptes); | |
955 | pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable); | |
956 | ||
957 | pmdp_set_wrprotect(src_mm, addr, src_pmd); | |
958 | pmd = pmd_mkold(pmd_wrprotect(pmd)); | |
959 | set_pmd_at(dst_mm, addr, dst_pmd, pmd); | |
960 | ||
961 | ret = 0; | |
962 | out_unlock: | |
963 | spin_unlock(src_ptl); | |
964 | spin_unlock(dst_ptl); | |
965 | out: | |
966 | return ret; | |
967 | } | |
968 | ||
969 | #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD | |
970 | static void touch_pud(struct vm_area_struct *vma, unsigned long addr, | |
971 | pud_t *pud, int flags) | |
972 | { | |
973 | pud_t _pud; | |
974 | ||
975 | _pud = pud_mkyoung(*pud); | |
976 | if (flags & FOLL_WRITE) | |
977 | _pud = pud_mkdirty(_pud); | |
978 | if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK, | |
979 | pud, _pud, flags & FOLL_WRITE)) | |
980 | update_mmu_cache_pud(vma, addr, pud); | |
981 | } | |
982 | ||
983 | struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr, | |
984 | pud_t *pud, int flags) | |
985 | { | |
986 | unsigned long pfn = pud_pfn(*pud); | |
987 | struct mm_struct *mm = vma->vm_mm; | |
988 | struct dev_pagemap *pgmap; | |
989 | struct page *page; | |
990 | ||
991 | assert_spin_locked(pud_lockptr(mm, pud)); | |
992 | ||
993 | if (flags & FOLL_WRITE && !pud_write(*pud)) | |
994 | return NULL; | |
995 | ||
996 | if (pud_present(*pud) && pud_devmap(*pud)) | |
997 | /* pass */; | |
998 | else | |
999 | return NULL; | |
1000 | ||
1001 | if (flags & FOLL_TOUCH) | |
1002 | touch_pud(vma, addr, pud, flags); | |
1003 | ||
1004 | /* | |
1005 | * device mapped pages can only be returned if the | |
1006 | * caller will manage the page reference count. | |
1007 | */ | |
1008 | if (!(flags & FOLL_GET)) | |
1009 | return ERR_PTR(-EEXIST); | |
1010 | ||
1011 | pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT; | |
1012 | pgmap = get_dev_pagemap(pfn, NULL); | |
1013 | if (!pgmap) | |
1014 | return ERR_PTR(-EFAULT); | |
1015 | page = pfn_to_page(pfn); | |
1016 | get_page(page); | |
1017 | put_dev_pagemap(pgmap); | |
1018 | ||
1019 | return page; | |
1020 | } | |
1021 | ||
1022 | int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm, | |
1023 | pud_t *dst_pud, pud_t *src_pud, unsigned long addr, | |
1024 | struct vm_area_struct *vma) | |
1025 | { | |
1026 | spinlock_t *dst_ptl, *src_ptl; | |
1027 | pud_t pud; | |
1028 | int ret; | |
1029 | ||
1030 | dst_ptl = pud_lock(dst_mm, dst_pud); | |
1031 | src_ptl = pud_lockptr(src_mm, src_pud); | |
1032 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
1033 | ||
1034 | ret = -EAGAIN; | |
1035 | pud = *src_pud; | |
1036 | if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud))) | |
1037 | goto out_unlock; | |
1038 | ||
1039 | /* | |
1040 | * When page table lock is held, the huge zero pud should not be | |
1041 | * under splitting since we don't split the page itself, only pud to | |
1042 | * a page table. | |
1043 | */ | |
1044 | if (is_huge_zero_pud(pud)) { | |
1045 | /* No huge zero pud yet */ | |
1046 | } | |
1047 | ||
1048 | pudp_set_wrprotect(src_mm, addr, src_pud); | |
1049 | pud = pud_mkold(pud_wrprotect(pud)); | |
1050 | set_pud_at(dst_mm, addr, dst_pud, pud); | |
1051 | ||
1052 | ret = 0; | |
1053 | out_unlock: | |
1054 | spin_unlock(src_ptl); | |
1055 | spin_unlock(dst_ptl); | |
1056 | return ret; | |
1057 | } | |
1058 | ||
1059 | void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud) | |
1060 | { | |
1061 | pud_t entry; | |
1062 | unsigned long haddr; | |
1063 | bool write = vmf->flags & FAULT_FLAG_WRITE; | |
1064 | ||
1065 | vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud); | |
1066 | if (unlikely(!pud_same(*vmf->pud, orig_pud))) | |
1067 | goto unlock; | |
1068 | ||
1069 | entry = pud_mkyoung(orig_pud); | |
1070 | if (write) | |
1071 | entry = pud_mkdirty(entry); | |
1072 | haddr = vmf->address & HPAGE_PUD_MASK; | |
1073 | if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write)) | |
1074 | update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud); | |
1075 | ||
1076 | unlock: | |
1077 | spin_unlock(vmf->ptl); | |
1078 | } | |
1079 | #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ | |
1080 | ||
1081 | void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd) | |
1082 | { | |
1083 | pmd_t entry; | |
1084 | unsigned long haddr; | |
1085 | bool write = vmf->flags & FAULT_FLAG_WRITE; | |
1086 | ||
1087 | vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); | |
1088 | if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) | |
1089 | goto unlock; | |
1090 | ||
1091 | entry = pmd_mkyoung(orig_pmd); | |
1092 | if (write) | |
1093 | entry = pmd_mkdirty(entry); | |
1094 | haddr = vmf->address & HPAGE_PMD_MASK; | |
1095 | if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write)) | |
1096 | update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd); | |
1097 | ||
1098 | unlock: | |
1099 | spin_unlock(vmf->ptl); | |
1100 | } | |
1101 | ||
1102 | static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd, | |
1103 | struct page *page) | |
1104 | { | |
1105 | struct vm_area_struct *vma = vmf->vma; | |
1106 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; | |
1107 | struct mem_cgroup *memcg; | |
1108 | pgtable_t pgtable; | |
1109 | pmd_t _pmd; | |
1110 | int ret = 0, i; | |
1111 | struct page **pages; | |
1112 | unsigned long mmun_start; /* For mmu_notifiers */ | |
1113 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1114 | ||
1115 | pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR, | |
1116 | GFP_KERNEL); | |
1117 | if (unlikely(!pages)) { | |
1118 | ret |= VM_FAULT_OOM; | |
1119 | goto out; | |
1120 | } | |
1121 | ||
1122 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
1123 | pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma, | |
1124 | vmf->address, page_to_nid(page)); | |
1125 | if (unlikely(!pages[i] || | |
1126 | mem_cgroup_try_charge(pages[i], vma->vm_mm, | |
1127 | GFP_KERNEL, &memcg, false))) { | |
1128 | if (pages[i]) | |
1129 | put_page(pages[i]); | |
1130 | while (--i >= 0) { | |
1131 | memcg = (void *)page_private(pages[i]); | |
1132 | set_page_private(pages[i], 0); | |
1133 | mem_cgroup_cancel_charge(pages[i], memcg, | |
1134 | false); | |
1135 | put_page(pages[i]); | |
1136 | } | |
1137 | kfree(pages); | |
1138 | ret |= VM_FAULT_OOM; | |
1139 | goto out; | |
1140 | } | |
1141 | set_page_private(pages[i], (unsigned long)memcg); | |
1142 | } | |
1143 | ||
1144 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
1145 | copy_user_highpage(pages[i], page + i, | |
1146 | haddr + PAGE_SIZE * i, vma); | |
1147 | __SetPageUptodate(pages[i]); | |
1148 | cond_resched(); | |
1149 | } | |
1150 | ||
1151 | mmun_start = haddr; | |
1152 | mmun_end = haddr + HPAGE_PMD_SIZE; | |
1153 | mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end); | |
1154 | ||
1155 | vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); | |
1156 | if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) | |
1157 | goto out_free_pages; | |
1158 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
1159 | ||
1160 | pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd); | |
1161 | /* leave pmd empty until pte is filled */ | |
1162 | ||
1163 | pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd); | |
1164 | pmd_populate(vma->vm_mm, &_pmd, pgtable); | |
1165 | ||
1166 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
1167 | pte_t entry; | |
1168 | entry = mk_pte(pages[i], vma->vm_page_prot); | |
1169 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); | |
1170 | memcg = (void *)page_private(pages[i]); | |
1171 | set_page_private(pages[i], 0); | |
1172 | page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false); | |
1173 | mem_cgroup_commit_charge(pages[i], memcg, false, false); | |
1174 | lru_cache_add_active_or_unevictable(pages[i], vma); | |
1175 | vmf->pte = pte_offset_map(&_pmd, haddr); | |
1176 | VM_BUG_ON(!pte_none(*vmf->pte)); | |
1177 | set_pte_at(vma->vm_mm, haddr, vmf->pte, entry); | |
1178 | pte_unmap(vmf->pte); | |
1179 | } | |
1180 | kfree(pages); | |
1181 | ||
1182 | smp_wmb(); /* make pte visible before pmd */ | |
1183 | pmd_populate(vma->vm_mm, vmf->pmd, pgtable); | |
1184 | page_remove_rmap(page, true); | |
1185 | spin_unlock(vmf->ptl); | |
1186 | ||
1187 | mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); | |
1188 | ||
1189 | ret |= VM_FAULT_WRITE; | |
1190 | put_page(page); | |
1191 | ||
1192 | out: | |
1193 | return ret; | |
1194 | ||
1195 | out_free_pages: | |
1196 | spin_unlock(vmf->ptl); | |
1197 | mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); | |
1198 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
1199 | memcg = (void *)page_private(pages[i]); | |
1200 | set_page_private(pages[i], 0); | |
1201 | mem_cgroup_cancel_charge(pages[i], memcg, false); | |
1202 | put_page(pages[i]); | |
1203 | } | |
1204 | kfree(pages); | |
1205 | goto out; | |
1206 | } | |
1207 | ||
1208 | int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd) | |
1209 | { | |
1210 | struct vm_area_struct *vma = vmf->vma; | |
1211 | struct page *page = NULL, *new_page; | |
1212 | struct mem_cgroup *memcg; | |
1213 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; | |
1214 | unsigned long mmun_start; /* For mmu_notifiers */ | |
1215 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1216 | gfp_t huge_gfp; /* for allocation and charge */ | |
1217 | int ret = 0; | |
1218 | ||
1219 | vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd); | |
1220 | VM_BUG_ON_VMA(!vma->anon_vma, vma); | |
1221 | if (is_huge_zero_pmd(orig_pmd)) | |
1222 | goto alloc; | |
1223 | spin_lock(vmf->ptl); | |
1224 | if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) | |
1225 | goto out_unlock; | |
1226 | ||
1227 | page = pmd_page(orig_pmd); | |
1228 | VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page); | |
1229 | /* | |
1230 | * We can only reuse the page if nobody else maps the huge page or it's | |
1231 | * part. | |
1232 | */ | |
1233 | if (page_trans_huge_mapcount(page, NULL) == 1) { | |
1234 | pmd_t entry; | |
1235 | entry = pmd_mkyoung(orig_pmd); | |
1236 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
1237 | if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1)) | |
1238 | update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); | |
1239 | ret |= VM_FAULT_WRITE; | |
1240 | goto out_unlock; | |
1241 | } | |
1242 | get_page(page); | |
1243 | spin_unlock(vmf->ptl); | |
1244 | alloc: | |
1245 | if (transparent_hugepage_enabled(vma) && | |
1246 | !transparent_hugepage_debug_cow()) { | |
1247 | huge_gfp = alloc_hugepage_direct_gfpmask(vma); | |
1248 | new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER); | |
1249 | } else | |
1250 | new_page = NULL; | |
1251 | ||
1252 | if (likely(new_page)) { | |
1253 | prep_transhuge_page(new_page); | |
1254 | } else { | |
1255 | if (!page) { | |
1256 | split_huge_pmd(vma, vmf->pmd, vmf->address); | |
1257 | ret |= VM_FAULT_FALLBACK; | |
1258 | } else { | |
1259 | ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page); | |
1260 | if (ret & VM_FAULT_OOM) { | |
1261 | split_huge_pmd(vma, vmf->pmd, vmf->address); | |
1262 | ret |= VM_FAULT_FALLBACK; | |
1263 | } | |
1264 | put_page(page); | |
1265 | } | |
1266 | count_vm_event(THP_FAULT_FALLBACK); | |
1267 | goto out; | |
1268 | } | |
1269 | ||
1270 | if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm, | |
1271 | huge_gfp, &memcg, true))) { | |
1272 | put_page(new_page); | |
1273 | split_huge_pmd(vma, vmf->pmd, vmf->address); | |
1274 | if (page) | |
1275 | put_page(page); | |
1276 | ret |= VM_FAULT_FALLBACK; | |
1277 | count_vm_event(THP_FAULT_FALLBACK); | |
1278 | goto out; | |
1279 | } | |
1280 | ||
1281 | count_vm_event(THP_FAULT_ALLOC); | |
1282 | ||
1283 | if (!page) | |
1284 | clear_huge_page(new_page, haddr, HPAGE_PMD_NR); | |
1285 | else | |
1286 | copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR); | |
1287 | __SetPageUptodate(new_page); | |
1288 | ||
1289 | mmun_start = haddr; | |
1290 | mmun_end = haddr + HPAGE_PMD_SIZE; | |
1291 | mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end); | |
1292 | ||
1293 | spin_lock(vmf->ptl); | |
1294 | if (page) | |
1295 | put_page(page); | |
1296 | if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) { | |
1297 | spin_unlock(vmf->ptl); | |
1298 | mem_cgroup_cancel_charge(new_page, memcg, true); | |
1299 | put_page(new_page); | |
1300 | goto out_mn; | |
1301 | } else { | |
1302 | pmd_t entry; | |
1303 | entry = mk_huge_pmd(new_page, vma->vm_page_prot); | |
1304 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
1305 | pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd); | |
1306 | page_add_new_anon_rmap(new_page, vma, haddr, true); | |
1307 | mem_cgroup_commit_charge(new_page, memcg, false, true); | |
1308 | lru_cache_add_active_or_unevictable(new_page, vma); | |
1309 | set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry); | |
1310 | update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); | |
1311 | if (!page) { | |
1312 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR); | |
1313 | } else { | |
1314 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
1315 | page_remove_rmap(page, true); | |
1316 | put_page(page); | |
1317 | } | |
1318 | ret |= VM_FAULT_WRITE; | |
1319 | } | |
1320 | spin_unlock(vmf->ptl); | |
1321 | out_mn: | |
1322 | mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end); | |
1323 | out: | |
1324 | return ret; | |
1325 | out_unlock: | |
1326 | spin_unlock(vmf->ptl); | |
1327 | return ret; | |
1328 | } | |
1329 | ||
1330 | /* | |
1331 | * FOLL_FORCE can write to even unwritable pmd's, but only | |
1332 | * after we've gone through a COW cycle and they are dirty. | |
1333 | */ | |
1334 | static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags) | |
1335 | { | |
1336 | return pmd_write(pmd) || | |
1337 | ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd)); | |
1338 | } | |
1339 | ||
1340 | struct page *follow_trans_huge_pmd(struct vm_area_struct *vma, | |
1341 | unsigned long addr, | |
1342 | pmd_t *pmd, | |
1343 | unsigned int flags) | |
1344 | { | |
1345 | struct mm_struct *mm = vma->vm_mm; | |
1346 | struct page *page = NULL; | |
1347 | ||
1348 | assert_spin_locked(pmd_lockptr(mm, pmd)); | |
1349 | ||
1350 | if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags)) | |
1351 | goto out; | |
1352 | ||
1353 | /* Avoid dumping huge zero page */ | |
1354 | if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd)) | |
1355 | return ERR_PTR(-EFAULT); | |
1356 | ||
1357 | /* Full NUMA hinting faults to serialise migration in fault paths */ | |
1358 | if ((flags & FOLL_NUMA) && pmd_protnone(*pmd)) | |
1359 | goto out; | |
1360 | ||
1361 | page = pmd_page(*pmd); | |
1362 | VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page); | |
1363 | if (flags & FOLL_TOUCH) | |
1364 | touch_pmd(vma, addr, pmd, flags); | |
1365 | if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) { | |
1366 | /* | |
1367 | * We don't mlock() pte-mapped THPs. This way we can avoid | |
1368 | * leaking mlocked pages into non-VM_LOCKED VMAs. | |
1369 | * | |
1370 | * For anon THP: | |
1371 | * | |
1372 | * In most cases the pmd is the only mapping of the page as we | |
1373 | * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for | |
1374 | * writable private mappings in populate_vma_page_range(). | |
1375 | * | |
1376 | * The only scenario when we have the page shared here is if we | |
1377 | * mlocking read-only mapping shared over fork(). We skip | |
1378 | * mlocking such pages. | |
1379 | * | |
1380 | * For file THP: | |
1381 | * | |
1382 | * We can expect PageDoubleMap() to be stable under page lock: | |
1383 | * for file pages we set it in page_add_file_rmap(), which | |
1384 | * requires page to be locked. | |
1385 | */ | |
1386 | ||
1387 | if (PageAnon(page) && compound_mapcount(page) != 1) | |
1388 | goto skip_mlock; | |
1389 | if (PageDoubleMap(page) || !page->mapping) | |
1390 | goto skip_mlock; | |
1391 | if (!trylock_page(page)) | |
1392 | goto skip_mlock; | |
1393 | lru_add_drain(); | |
1394 | if (page->mapping && !PageDoubleMap(page)) | |
1395 | mlock_vma_page(page); | |
1396 | unlock_page(page); | |
1397 | } | |
1398 | skip_mlock: | |
1399 | page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT; | |
1400 | VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page); | |
1401 | if (flags & FOLL_GET) | |
1402 | get_page(page); | |
1403 | ||
1404 | out: | |
1405 | return page; | |
1406 | } | |
1407 | ||
1408 | /* NUMA hinting page fault entry point for trans huge pmds */ | |
1409 | int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd) | |
1410 | { | |
1411 | struct vm_area_struct *vma = vmf->vma; | |
1412 | struct anon_vma *anon_vma = NULL; | |
1413 | struct page *page; | |
1414 | unsigned long haddr = vmf->address & HPAGE_PMD_MASK; | |
1415 | int page_nid = -1, this_nid = numa_node_id(); | |
1416 | int target_nid, last_cpupid = -1; | |
1417 | bool page_locked; | |
1418 | bool migrated = false; | |
1419 | bool was_writable; | |
1420 | int flags = 0; | |
1421 | ||
1422 | vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd); | |
1423 | if (unlikely(!pmd_same(pmd, *vmf->pmd))) | |
1424 | goto out_unlock; | |
1425 | ||
1426 | /* | |
1427 | * If there are potential migrations, wait for completion and retry | |
1428 | * without disrupting NUMA hinting information. Do not relock and | |
1429 | * check_same as the page may no longer be mapped. | |
1430 | */ | |
1431 | if (unlikely(pmd_trans_migrating(*vmf->pmd))) { | |
1432 | page = pmd_page(*vmf->pmd); | |
1433 | if (!get_page_unless_zero(page)) | |
1434 | goto out_unlock; | |
1435 | spin_unlock(vmf->ptl); | |
1436 | wait_on_page_locked(page); | |
1437 | put_page(page); | |
1438 | goto out; | |
1439 | } | |
1440 | ||
1441 | page = pmd_page(pmd); | |
1442 | BUG_ON(is_huge_zero_page(page)); | |
1443 | page_nid = page_to_nid(page); | |
1444 | last_cpupid = page_cpupid_last(page); | |
1445 | count_vm_numa_event(NUMA_HINT_FAULTS); | |
1446 | if (page_nid == this_nid) { | |
1447 | count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); | |
1448 | flags |= TNF_FAULT_LOCAL; | |
1449 | } | |
1450 | ||
1451 | /* See similar comment in do_numa_page for explanation */ | |
1452 | if (!pmd_savedwrite(pmd)) | |
1453 | flags |= TNF_NO_GROUP; | |
1454 | ||
1455 | /* | |
1456 | * Acquire the page lock to serialise THP migrations but avoid dropping | |
1457 | * page_table_lock if at all possible | |
1458 | */ | |
1459 | page_locked = trylock_page(page); | |
1460 | target_nid = mpol_misplaced(page, vma, haddr); | |
1461 | if (target_nid == -1) { | |
1462 | /* If the page was locked, there are no parallel migrations */ | |
1463 | if (page_locked) | |
1464 | goto clear_pmdnuma; | |
1465 | } | |
1466 | ||
1467 | /* Migration could have started since the pmd_trans_migrating check */ | |
1468 | if (!page_locked) { | |
1469 | page_nid = -1; | |
1470 | if (!get_page_unless_zero(page)) | |
1471 | goto out_unlock; | |
1472 | spin_unlock(vmf->ptl); | |
1473 | wait_on_page_locked(page); | |
1474 | put_page(page); | |
1475 | goto out; | |
1476 | } | |
1477 | ||
1478 | /* | |
1479 | * Page is misplaced. Page lock serialises migrations. Acquire anon_vma | |
1480 | * to serialises splits | |
1481 | */ | |
1482 | get_page(page); | |
1483 | spin_unlock(vmf->ptl); | |
1484 | anon_vma = page_lock_anon_vma_read(page); | |
1485 | ||
1486 | /* Confirm the PMD did not change while page_table_lock was released */ | |
1487 | spin_lock(vmf->ptl); | |
1488 | if (unlikely(!pmd_same(pmd, *vmf->pmd))) { | |
1489 | unlock_page(page); | |
1490 | put_page(page); | |
1491 | page_nid = -1; | |
1492 | goto out_unlock; | |
1493 | } | |
1494 | ||
1495 | /* Bail if we fail to protect against THP splits for any reason */ | |
1496 | if (unlikely(!anon_vma)) { | |
1497 | put_page(page); | |
1498 | page_nid = -1; | |
1499 | goto clear_pmdnuma; | |
1500 | } | |
1501 | ||
1502 | /* | |
1503 | * The page_table_lock above provides a memory barrier | |
1504 | * with change_protection_range. | |
1505 | */ | |
1506 | if (mm_tlb_flush_pending(vma->vm_mm)) | |
1507 | flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE); | |
1508 | ||
1509 | /* | |
1510 | * Migrate the THP to the requested node, returns with page unlocked | |
1511 | * and access rights restored. | |
1512 | */ | |
1513 | spin_unlock(vmf->ptl); | |
1514 | migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma, | |
1515 | vmf->pmd, pmd, vmf->address, page, target_nid); | |
1516 | if (migrated) { | |
1517 | flags |= TNF_MIGRATED; | |
1518 | page_nid = target_nid; | |
1519 | } else | |
1520 | flags |= TNF_MIGRATE_FAIL; | |
1521 | ||
1522 | goto out; | |
1523 | clear_pmdnuma: | |
1524 | BUG_ON(!PageLocked(page)); | |
1525 | was_writable = pmd_savedwrite(pmd); | |
1526 | pmd = pmd_modify(pmd, vma->vm_page_prot); | |
1527 | pmd = pmd_mkyoung(pmd); | |
1528 | if (was_writable) | |
1529 | pmd = pmd_mkwrite(pmd); | |
1530 | set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd); | |
1531 | update_mmu_cache_pmd(vma, vmf->address, vmf->pmd); | |
1532 | unlock_page(page); | |
1533 | out_unlock: | |
1534 | spin_unlock(vmf->ptl); | |
1535 | ||
1536 | out: | |
1537 | if (anon_vma) | |
1538 | page_unlock_anon_vma_read(anon_vma); | |
1539 | ||
1540 | if (page_nid != -1) | |
1541 | task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, | |
1542 | flags); | |
1543 | ||
1544 | return 0; | |
1545 | } | |
1546 | ||
1547 | /* | |
1548 | * Return true if we do MADV_FREE successfully on entire pmd page. | |
1549 | * Otherwise, return false. | |
1550 | */ | |
1551 | bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, | |
1552 | pmd_t *pmd, unsigned long addr, unsigned long next) | |
1553 | { | |
1554 | spinlock_t *ptl; | |
1555 | pmd_t orig_pmd; | |
1556 | struct page *page; | |
1557 | struct mm_struct *mm = tlb->mm; | |
1558 | bool ret = false; | |
1559 | ||
1560 | tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE); | |
1561 | ||
1562 | ptl = pmd_trans_huge_lock(pmd, vma); | |
1563 | if (!ptl) | |
1564 | goto out_unlocked; | |
1565 | ||
1566 | orig_pmd = *pmd; | |
1567 | if (is_huge_zero_pmd(orig_pmd)) | |
1568 | goto out; | |
1569 | ||
1570 | page = pmd_page(orig_pmd); | |
1571 | /* | |
1572 | * If other processes are mapping this page, we couldn't discard | |
1573 | * the page unless they all do MADV_FREE so let's skip the page. | |
1574 | */ | |
1575 | if (page_mapcount(page) != 1) | |
1576 | goto out; | |
1577 | ||
1578 | if (!trylock_page(page)) | |
1579 | goto out; | |
1580 | ||
1581 | /* | |
1582 | * If user want to discard part-pages of THP, split it so MADV_FREE | |
1583 | * will deactivate only them. | |
1584 | */ | |
1585 | if (next - addr != HPAGE_PMD_SIZE) { | |
1586 | get_page(page); | |
1587 | spin_unlock(ptl); | |
1588 | split_huge_page(page); | |
1589 | unlock_page(page); | |
1590 | put_page(page); | |
1591 | goto out_unlocked; | |
1592 | } | |
1593 | ||
1594 | if (PageDirty(page)) | |
1595 | ClearPageDirty(page); | |
1596 | unlock_page(page); | |
1597 | ||
1598 | if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) { | |
1599 | pmdp_invalidate(vma, addr, pmd); | |
1600 | orig_pmd = pmd_mkold(orig_pmd); | |
1601 | orig_pmd = pmd_mkclean(orig_pmd); | |
1602 | ||
1603 | set_pmd_at(mm, addr, pmd, orig_pmd); | |
1604 | tlb_remove_pmd_tlb_entry(tlb, pmd, addr); | |
1605 | } | |
1606 | ||
1607 | mark_page_lazyfree(page); | |
1608 | ret = true; | |
1609 | out: | |
1610 | spin_unlock(ptl); | |
1611 | out_unlocked: | |
1612 | return ret; | |
1613 | } | |
1614 | ||
1615 | static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd) | |
1616 | { | |
1617 | pgtable_t pgtable; | |
1618 | ||
1619 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); | |
1620 | pte_free(mm, pgtable); | |
1621 | atomic_long_dec(&mm->nr_ptes); | |
1622 | } | |
1623 | ||
1624 | int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, | |
1625 | pmd_t *pmd, unsigned long addr) | |
1626 | { | |
1627 | pmd_t orig_pmd; | |
1628 | spinlock_t *ptl; | |
1629 | ||
1630 | tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE); | |
1631 | ||
1632 | ptl = __pmd_trans_huge_lock(pmd, vma); | |
1633 | if (!ptl) | |
1634 | return 0; | |
1635 | /* | |
1636 | * For architectures like ppc64 we look at deposited pgtable | |
1637 | * when calling pmdp_huge_get_and_clear. So do the | |
1638 | * pgtable_trans_huge_withdraw after finishing pmdp related | |
1639 | * operations. | |
1640 | */ | |
1641 | orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd, | |
1642 | tlb->fullmm); | |
1643 | tlb_remove_pmd_tlb_entry(tlb, pmd, addr); | |
1644 | if (vma_is_dax(vma)) { | |
1645 | if (arch_needs_pgtable_deposit()) | |
1646 | zap_deposited_table(tlb->mm, pmd); | |
1647 | spin_unlock(ptl); | |
1648 | if (is_huge_zero_pmd(orig_pmd)) | |
1649 | tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); | |
1650 | } else if (is_huge_zero_pmd(orig_pmd)) { | |
1651 | zap_deposited_table(tlb->mm, pmd); | |
1652 | spin_unlock(ptl); | |
1653 | tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE); | |
1654 | } else { | |
1655 | struct page *page = pmd_page(orig_pmd); | |
1656 | page_remove_rmap(page, true); | |
1657 | VM_BUG_ON_PAGE(page_mapcount(page) < 0, page); | |
1658 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
1659 | if (PageAnon(page)) { | |
1660 | zap_deposited_table(tlb->mm, pmd); | |
1661 | add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR); | |
1662 | } else { | |
1663 | if (arch_needs_pgtable_deposit()) | |
1664 | zap_deposited_table(tlb->mm, pmd); | |
1665 | add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR); | |
1666 | } | |
1667 | spin_unlock(ptl); | |
1668 | tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE); | |
1669 | } | |
1670 | return 1; | |
1671 | } | |
1672 | ||
1673 | #ifndef pmd_move_must_withdraw | |
1674 | static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl, | |
1675 | spinlock_t *old_pmd_ptl, | |
1676 | struct vm_area_struct *vma) | |
1677 | { | |
1678 | /* | |
1679 | * With split pmd lock we also need to move preallocated | |
1680 | * PTE page table if new_pmd is on different PMD page table. | |
1681 | * | |
1682 | * We also don't deposit and withdraw tables for file pages. | |
1683 | */ | |
1684 | return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma); | |
1685 | } | |
1686 | #endif | |
1687 | ||
1688 | bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, | |
1689 | unsigned long new_addr, unsigned long old_end, | |
1690 | pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush) | |
1691 | { | |
1692 | spinlock_t *old_ptl, *new_ptl; | |
1693 | pmd_t pmd; | |
1694 | struct mm_struct *mm = vma->vm_mm; | |
1695 | bool force_flush = false; | |
1696 | ||
1697 | if ((old_addr & ~HPAGE_PMD_MASK) || | |
1698 | (new_addr & ~HPAGE_PMD_MASK) || | |
1699 | old_end - old_addr < HPAGE_PMD_SIZE) | |
1700 | return false; | |
1701 | ||
1702 | /* | |
1703 | * The destination pmd shouldn't be established, free_pgtables() | |
1704 | * should have release it. | |
1705 | */ | |
1706 | if (WARN_ON(!pmd_none(*new_pmd))) { | |
1707 | VM_BUG_ON(pmd_trans_huge(*new_pmd)); | |
1708 | return false; | |
1709 | } | |
1710 | ||
1711 | /* | |
1712 | * We don't have to worry about the ordering of src and dst | |
1713 | * ptlocks because exclusive mmap_sem prevents deadlock. | |
1714 | */ | |
1715 | old_ptl = __pmd_trans_huge_lock(old_pmd, vma); | |
1716 | if (old_ptl) { | |
1717 | new_ptl = pmd_lockptr(mm, new_pmd); | |
1718 | if (new_ptl != old_ptl) | |
1719 | spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING); | |
1720 | pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd); | |
1721 | if (pmd_present(pmd) && pmd_dirty(pmd)) | |
1722 | force_flush = true; | |
1723 | VM_BUG_ON(!pmd_none(*new_pmd)); | |
1724 | ||
1725 | if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) { | |
1726 | pgtable_t pgtable; | |
1727 | pgtable = pgtable_trans_huge_withdraw(mm, old_pmd); | |
1728 | pgtable_trans_huge_deposit(mm, new_pmd, pgtable); | |
1729 | } | |
1730 | set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd)); | |
1731 | if (new_ptl != old_ptl) | |
1732 | spin_unlock(new_ptl); | |
1733 | if (force_flush) | |
1734 | flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE); | |
1735 | else | |
1736 | *need_flush = true; | |
1737 | spin_unlock(old_ptl); | |
1738 | return true; | |
1739 | } | |
1740 | return false; | |
1741 | } | |
1742 | ||
1743 | /* | |
1744 | * Returns | |
1745 | * - 0 if PMD could not be locked | |
1746 | * - 1 if PMD was locked but protections unchange and TLB flush unnecessary | |
1747 | * - HPAGE_PMD_NR is protections changed and TLB flush necessary | |
1748 | */ | |
1749 | int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, | |
1750 | unsigned long addr, pgprot_t newprot, int prot_numa) | |
1751 | { | |
1752 | struct mm_struct *mm = vma->vm_mm; | |
1753 | spinlock_t *ptl; | |
1754 | pmd_t entry; | |
1755 | bool preserve_write; | |
1756 | int ret; | |
1757 | ||
1758 | ptl = __pmd_trans_huge_lock(pmd, vma); | |
1759 | if (!ptl) | |
1760 | return 0; | |
1761 | ||
1762 | preserve_write = prot_numa && pmd_write(*pmd); | |
1763 | ret = 1; | |
1764 | ||
1765 | /* | |
1766 | * Avoid trapping faults against the zero page. The read-only | |
1767 | * data is likely to be read-cached on the local CPU and | |
1768 | * local/remote hits to the zero page are not interesting. | |
1769 | */ | |
1770 | if (prot_numa && is_huge_zero_pmd(*pmd)) | |
1771 | goto unlock; | |
1772 | ||
1773 | if (prot_numa && pmd_protnone(*pmd)) | |
1774 | goto unlock; | |
1775 | ||
1776 | /* | |
1777 | * In case prot_numa, we are under down_read(mmap_sem). It's critical | |
1778 | * to not clear pmd intermittently to avoid race with MADV_DONTNEED | |
1779 | * which is also under down_read(mmap_sem): | |
1780 | * | |
1781 | * CPU0: CPU1: | |
1782 | * change_huge_pmd(prot_numa=1) | |
1783 | * pmdp_huge_get_and_clear_notify() | |
1784 | * madvise_dontneed() | |
1785 | * zap_pmd_range() | |
1786 | * pmd_trans_huge(*pmd) == 0 (without ptl) | |
1787 | * // skip the pmd | |
1788 | * set_pmd_at(); | |
1789 | * // pmd is re-established | |
1790 | * | |
1791 | * The race makes MADV_DONTNEED miss the huge pmd and don't clear it | |
1792 | * which may break userspace. | |
1793 | * | |
1794 | * pmdp_invalidate() is required to make sure we don't miss | |
1795 | * dirty/young flags set by hardware. | |
1796 | */ | |
1797 | entry = *pmd; | |
1798 | pmdp_invalidate(vma, addr, pmd); | |
1799 | ||
1800 | /* | |
1801 | * Recover dirty/young flags. It relies on pmdp_invalidate to not | |
1802 | * corrupt them. | |
1803 | */ | |
1804 | if (pmd_dirty(*pmd)) | |
1805 | entry = pmd_mkdirty(entry); | |
1806 | if (pmd_young(*pmd)) | |
1807 | entry = pmd_mkyoung(entry); | |
1808 | ||
1809 | entry = pmd_modify(entry, newprot); | |
1810 | if (preserve_write) | |
1811 | entry = pmd_mk_savedwrite(entry); | |
1812 | ret = HPAGE_PMD_NR; | |
1813 | set_pmd_at(mm, addr, pmd, entry); | |
1814 | BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry)); | |
1815 | unlock: | |
1816 | spin_unlock(ptl); | |
1817 | return ret; | |
1818 | } | |
1819 | ||
1820 | /* | |
1821 | * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise. | |
1822 | * | |
1823 | * Note that if it returns page table lock pointer, this routine returns without | |
1824 | * unlocking page table lock. So callers must unlock it. | |
1825 | */ | |
1826 | spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma) | |
1827 | { | |
1828 | spinlock_t *ptl; | |
1829 | ptl = pmd_lock(vma->vm_mm, pmd); | |
1830 | if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd))) | |
1831 | return ptl; | |
1832 | spin_unlock(ptl); | |
1833 | return NULL; | |
1834 | } | |
1835 | ||
1836 | /* | |
1837 | * Returns true if a given pud maps a thp, false otherwise. | |
1838 | * | |
1839 | * Note that if it returns true, this routine returns without unlocking page | |
1840 | * table lock. So callers must unlock it. | |
1841 | */ | |
1842 | spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma) | |
1843 | { | |
1844 | spinlock_t *ptl; | |
1845 | ||
1846 | ptl = pud_lock(vma->vm_mm, pud); | |
1847 | if (likely(pud_trans_huge(*pud) || pud_devmap(*pud))) | |
1848 | return ptl; | |
1849 | spin_unlock(ptl); | |
1850 | return NULL; | |
1851 | } | |
1852 | ||
1853 | #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD | |
1854 | int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, | |
1855 | pud_t *pud, unsigned long addr) | |
1856 | { | |
1857 | pud_t orig_pud; | |
1858 | spinlock_t *ptl; | |
1859 | ||
1860 | ptl = __pud_trans_huge_lock(pud, vma); | |
1861 | if (!ptl) | |
1862 | return 0; | |
1863 | /* | |
1864 | * For architectures like ppc64 we look at deposited pgtable | |
1865 | * when calling pudp_huge_get_and_clear. So do the | |
1866 | * pgtable_trans_huge_withdraw after finishing pudp related | |
1867 | * operations. | |
1868 | */ | |
1869 | orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud, | |
1870 | tlb->fullmm); | |
1871 | tlb_remove_pud_tlb_entry(tlb, pud, addr); | |
1872 | if (vma_is_dax(vma)) { | |
1873 | spin_unlock(ptl); | |
1874 | /* No zero page support yet */ | |
1875 | } else { | |
1876 | /* No support for anonymous PUD pages yet */ | |
1877 | BUG(); | |
1878 | } | |
1879 | return 1; | |
1880 | } | |
1881 | ||
1882 | static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud, | |
1883 | unsigned long haddr) | |
1884 | { | |
1885 | VM_BUG_ON(haddr & ~HPAGE_PUD_MASK); | |
1886 | VM_BUG_ON_VMA(vma->vm_start > haddr, vma); | |
1887 | VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma); | |
1888 | VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud)); | |
1889 | ||
1890 | count_vm_event(THP_SPLIT_PUD); | |
1891 | ||
1892 | pudp_huge_clear_flush_notify(vma, haddr, pud); | |
1893 | } | |
1894 | ||
1895 | void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud, | |
1896 | unsigned long address) | |
1897 | { | |
1898 | spinlock_t *ptl; | |
1899 | struct mm_struct *mm = vma->vm_mm; | |
1900 | unsigned long haddr = address & HPAGE_PUD_MASK; | |
1901 | ||
1902 | mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE); | |
1903 | ptl = pud_lock(mm, pud); | |
1904 | if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud))) | |
1905 | goto out; | |
1906 | __split_huge_pud_locked(vma, pud, haddr); | |
1907 | ||
1908 | out: | |
1909 | spin_unlock(ptl); | |
1910 | mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PUD_SIZE); | |
1911 | } | |
1912 | #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */ | |
1913 | ||
1914 | static void __split_huge_zero_page_pmd(struct vm_area_struct *vma, | |
1915 | unsigned long haddr, pmd_t *pmd) | |
1916 | { | |
1917 | struct mm_struct *mm = vma->vm_mm; | |
1918 | pgtable_t pgtable; | |
1919 | pmd_t _pmd; | |
1920 | int i; | |
1921 | ||
1922 | /* leave pmd empty until pte is filled */ | |
1923 | pmdp_huge_clear_flush_notify(vma, haddr, pmd); | |
1924 | ||
1925 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); | |
1926 | pmd_populate(mm, &_pmd, pgtable); | |
1927 | ||
1928 | for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) { | |
1929 | pte_t *pte, entry; | |
1930 | entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot); | |
1931 | entry = pte_mkspecial(entry); | |
1932 | pte = pte_offset_map(&_pmd, haddr); | |
1933 | VM_BUG_ON(!pte_none(*pte)); | |
1934 | set_pte_at(mm, haddr, pte, entry); | |
1935 | pte_unmap(pte); | |
1936 | } | |
1937 | smp_wmb(); /* make pte visible before pmd */ | |
1938 | pmd_populate(mm, pmd, pgtable); | |
1939 | } | |
1940 | ||
1941 | static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd, | |
1942 | unsigned long haddr, bool freeze) | |
1943 | { | |
1944 | struct mm_struct *mm = vma->vm_mm; | |
1945 | struct page *page; | |
1946 | pgtable_t pgtable; | |
1947 | pmd_t _pmd; | |
1948 | bool young, write, dirty, soft_dirty; | |
1949 | unsigned long addr; | |
1950 | int i; | |
1951 | ||
1952 | VM_BUG_ON(haddr & ~HPAGE_PMD_MASK); | |
1953 | VM_BUG_ON_VMA(vma->vm_start > haddr, vma); | |
1954 | VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma); | |
1955 | VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)); | |
1956 | ||
1957 | count_vm_event(THP_SPLIT_PMD); | |
1958 | ||
1959 | if (!vma_is_anonymous(vma)) { | |
1960 | _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd); | |
1961 | /* | |
1962 | * We are going to unmap this huge page. So | |
1963 | * just go ahead and zap it | |
1964 | */ | |
1965 | if (arch_needs_pgtable_deposit()) | |
1966 | zap_deposited_table(mm, pmd); | |
1967 | if (vma_is_dax(vma)) | |
1968 | return; | |
1969 | page = pmd_page(_pmd); | |
1970 | if (!PageReferenced(page) && pmd_young(_pmd)) | |
1971 | SetPageReferenced(page); | |
1972 | page_remove_rmap(page, true); | |
1973 | put_page(page); | |
1974 | add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR); | |
1975 | return; | |
1976 | } else if (is_huge_zero_pmd(*pmd)) { | |
1977 | return __split_huge_zero_page_pmd(vma, haddr, pmd); | |
1978 | } | |
1979 | ||
1980 | page = pmd_page(*pmd); | |
1981 | VM_BUG_ON_PAGE(!page_count(page), page); | |
1982 | page_ref_add(page, HPAGE_PMD_NR - 1); | |
1983 | write = pmd_write(*pmd); | |
1984 | young = pmd_young(*pmd); | |
1985 | dirty = pmd_dirty(*pmd); | |
1986 | soft_dirty = pmd_soft_dirty(*pmd); | |
1987 | ||
1988 | pmdp_huge_split_prepare(vma, haddr, pmd); | |
1989 | pgtable = pgtable_trans_huge_withdraw(mm, pmd); | |
1990 | pmd_populate(mm, &_pmd, pgtable); | |
1991 | ||
1992 | for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) { | |
1993 | pte_t entry, *pte; | |
1994 | /* | |
1995 | * Note that NUMA hinting access restrictions are not | |
1996 | * transferred to avoid any possibility of altering | |
1997 | * permissions across VMAs. | |
1998 | */ | |
1999 | if (freeze) { | |
2000 | swp_entry_t swp_entry; | |
2001 | swp_entry = make_migration_entry(page + i, write); | |
2002 | entry = swp_entry_to_pte(swp_entry); | |
2003 | if (soft_dirty) | |
2004 | entry = pte_swp_mksoft_dirty(entry); | |
2005 | } else { | |
2006 | entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot)); | |
2007 | entry = maybe_mkwrite(entry, vma); | |
2008 | if (!write) | |
2009 | entry = pte_wrprotect(entry); | |
2010 | if (!young) | |
2011 | entry = pte_mkold(entry); | |
2012 | if (soft_dirty) | |
2013 | entry = pte_mksoft_dirty(entry); | |
2014 | } | |
2015 | if (dirty) | |
2016 | SetPageDirty(page + i); | |
2017 | pte = pte_offset_map(&_pmd, addr); | |
2018 | BUG_ON(!pte_none(*pte)); | |
2019 | set_pte_at(mm, addr, pte, entry); | |
2020 | atomic_inc(&page[i]._mapcount); | |
2021 | pte_unmap(pte); | |
2022 | } | |
2023 | ||
2024 | /* | |
2025 | * Set PG_double_map before dropping compound_mapcount to avoid | |
2026 | * false-negative page_mapped(). | |
2027 | */ | |
2028 | if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) { | |
2029 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
2030 | atomic_inc(&page[i]._mapcount); | |
2031 | } | |
2032 | ||
2033 | if (atomic_add_negative(-1, compound_mapcount_ptr(page))) { | |
2034 | /* Last compound_mapcount is gone. */ | |
2035 | __dec_node_page_state(page, NR_ANON_THPS); | |
2036 | if (TestClearPageDoubleMap(page)) { | |
2037 | /* No need in mapcount reference anymore */ | |
2038 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
2039 | atomic_dec(&page[i]._mapcount); | |
2040 | } | |
2041 | } | |
2042 | ||
2043 | smp_wmb(); /* make pte visible before pmd */ | |
2044 | /* | |
2045 | * Up to this point the pmd is present and huge and userland has the | |
2046 | * whole access to the hugepage during the split (which happens in | |
2047 | * place). If we overwrite the pmd with the not-huge version pointing | |
2048 | * to the pte here (which of course we could if all CPUs were bug | |
2049 | * free), userland could trigger a small page size TLB miss on the | |
2050 | * small sized TLB while the hugepage TLB entry is still established in | |
2051 | * the huge TLB. Some CPU doesn't like that. | |
2052 | * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum | |
2053 | * 383 on page 93. Intel should be safe but is also warns that it's | |
2054 | * only safe if the permission and cache attributes of the two entries | |
2055 | * loaded in the two TLB is identical (which should be the case here). | |
2056 | * But it is generally safer to never allow small and huge TLB entries | |
2057 | * for the same virtual address to be loaded simultaneously. So instead | |
2058 | * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the | |
2059 | * current pmd notpresent (atomically because here the pmd_trans_huge | |
2060 | * and pmd_trans_splitting must remain set at all times on the pmd | |
2061 | * until the split is complete for this pmd), then we flush the SMP TLB | |
2062 | * and finally we write the non-huge version of the pmd entry with | |
2063 | * pmd_populate. | |
2064 | */ | |
2065 | pmdp_invalidate(vma, haddr, pmd); | |
2066 | pmd_populate(mm, pmd, pgtable); | |
2067 | ||
2068 | if (freeze) { | |
2069 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
2070 | page_remove_rmap(page + i, false); | |
2071 | put_page(page + i); | |
2072 | } | |
2073 | } | |
2074 | } | |
2075 | ||
2076 | void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd, | |
2077 | unsigned long address, bool freeze, struct page *page) | |
2078 | { | |
2079 | spinlock_t *ptl; | |
2080 | struct mm_struct *mm = vma->vm_mm; | |
2081 | unsigned long haddr = address & HPAGE_PMD_MASK; | |
2082 | ||
2083 | mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE); | |
2084 | ptl = pmd_lock(mm, pmd); | |
2085 | ||
2086 | /* | |
2087 | * If caller asks to setup a migration entries, we need a page to check | |
2088 | * pmd against. Otherwise we can end up replacing wrong page. | |
2089 | */ | |
2090 | VM_BUG_ON(freeze && !page); | |
2091 | if (page && page != pmd_page(*pmd)) | |
2092 | goto out; | |
2093 | ||
2094 | if (pmd_trans_huge(*pmd)) { | |
2095 | page = pmd_page(*pmd); | |
2096 | if (PageMlocked(page)) | |
2097 | clear_page_mlock(page); | |
2098 | } else if (!pmd_devmap(*pmd)) | |
2099 | goto out; | |
2100 | __split_huge_pmd_locked(vma, pmd, haddr, freeze); | |
2101 | out: | |
2102 | spin_unlock(ptl); | |
2103 | mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE); | |
2104 | } | |
2105 | ||
2106 | void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address, | |
2107 | bool freeze, struct page *page) | |
2108 | { | |
2109 | pgd_t *pgd; | |
2110 | p4d_t *p4d; | |
2111 | pud_t *pud; | |
2112 | pmd_t *pmd; | |
2113 | ||
2114 | pgd = pgd_offset(vma->vm_mm, address); | |
2115 | if (!pgd_present(*pgd)) | |
2116 | return; | |
2117 | ||
2118 | p4d = p4d_offset(pgd, address); | |
2119 | if (!p4d_present(*p4d)) | |
2120 | return; | |
2121 | ||
2122 | pud = pud_offset(p4d, address); | |
2123 | if (!pud_present(*pud)) | |
2124 | return; | |
2125 | ||
2126 | pmd = pmd_offset(pud, address); | |
2127 | ||
2128 | __split_huge_pmd(vma, pmd, address, freeze, page); | |
2129 | } | |
2130 | ||
2131 | void vma_adjust_trans_huge(struct vm_area_struct *vma, | |
2132 | unsigned long start, | |
2133 | unsigned long end, | |
2134 | long adjust_next) | |
2135 | { | |
2136 | /* | |
2137 | * If the new start address isn't hpage aligned and it could | |
2138 | * previously contain an hugepage: check if we need to split | |
2139 | * an huge pmd. | |
2140 | */ | |
2141 | if (start & ~HPAGE_PMD_MASK && | |
2142 | (start & HPAGE_PMD_MASK) >= vma->vm_start && | |
2143 | (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
2144 | split_huge_pmd_address(vma, start, false, NULL); | |
2145 | ||
2146 | /* | |
2147 | * If the new end address isn't hpage aligned and it could | |
2148 | * previously contain an hugepage: check if we need to split | |
2149 | * an huge pmd. | |
2150 | */ | |
2151 | if (end & ~HPAGE_PMD_MASK && | |
2152 | (end & HPAGE_PMD_MASK) >= vma->vm_start && | |
2153 | (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end) | |
2154 | split_huge_pmd_address(vma, end, false, NULL); | |
2155 | ||
2156 | /* | |
2157 | * If we're also updating the vma->vm_next->vm_start, if the new | |
2158 | * vm_next->vm_start isn't page aligned and it could previously | |
2159 | * contain an hugepage: check if we need to split an huge pmd. | |
2160 | */ | |
2161 | if (adjust_next > 0) { | |
2162 | struct vm_area_struct *next = vma->vm_next; | |
2163 | unsigned long nstart = next->vm_start; | |
2164 | nstart += adjust_next << PAGE_SHIFT; | |
2165 | if (nstart & ~HPAGE_PMD_MASK && | |
2166 | (nstart & HPAGE_PMD_MASK) >= next->vm_start && | |
2167 | (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end) | |
2168 | split_huge_pmd_address(next, nstart, false, NULL); | |
2169 | } | |
2170 | } | |
2171 | ||
2172 | static void freeze_page(struct page *page) | |
2173 | { | |
2174 | enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS | | |
2175 | TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD; | |
2176 | bool unmap_success; | |
2177 | ||
2178 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
2179 | ||
2180 | if (PageAnon(page)) | |
2181 | ttu_flags |= TTU_MIGRATION; | |
2182 | ||
2183 | unmap_success = try_to_unmap(page, ttu_flags); | |
2184 | VM_BUG_ON_PAGE(!unmap_success, page); | |
2185 | } | |
2186 | ||
2187 | static void unfreeze_page(struct page *page) | |
2188 | { | |
2189 | int i; | |
2190 | if (PageTransHuge(page)) { | |
2191 | remove_migration_ptes(page, page, true); | |
2192 | } else { | |
2193 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
2194 | remove_migration_ptes(page + i, page + i, true); | |
2195 | } | |
2196 | } | |
2197 | ||
2198 | static void __split_huge_page_tail(struct page *head, int tail, | |
2199 | struct lruvec *lruvec, struct list_head *list) | |
2200 | { | |
2201 | struct page *page_tail = head + tail; | |
2202 | ||
2203 | VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail); | |
2204 | VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail); | |
2205 | ||
2206 | /* | |
2207 | * tail_page->_refcount is zero and not changing from under us. But | |
2208 | * get_page_unless_zero() may be running from under us on the | |
2209 | * tail_page. If we used atomic_set() below instead of atomic_inc() or | |
2210 | * atomic_add(), we would then run atomic_set() concurrently with | |
2211 | * get_page_unless_zero(), and atomic_set() is implemented in C not | |
2212 | * using locked ops. spin_unlock on x86 sometime uses locked ops | |
2213 | * because of PPro errata 66, 92, so unless somebody can guarantee | |
2214 | * atomic_set() here would be safe on all archs (and not only on x86), | |
2215 | * it's safer to use atomic_inc()/atomic_add(). | |
2216 | */ | |
2217 | if (PageAnon(head) && !PageSwapCache(head)) { | |
2218 | page_ref_inc(page_tail); | |
2219 | } else { | |
2220 | /* Additional pin to radix tree */ | |
2221 | page_ref_add(page_tail, 2); | |
2222 | } | |
2223 | ||
2224 | page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; | |
2225 | page_tail->flags |= (head->flags & | |
2226 | ((1L << PG_referenced) | | |
2227 | (1L << PG_swapbacked) | | |
2228 | (1L << PG_swapcache) | | |
2229 | (1L << PG_mlocked) | | |
2230 | (1L << PG_uptodate) | | |
2231 | (1L << PG_active) | | |
2232 | (1L << PG_locked) | | |
2233 | (1L << PG_unevictable) | | |
2234 | (1L << PG_dirty))); | |
2235 | ||
2236 | /* | |
2237 | * After clearing PageTail the gup refcount can be released. | |
2238 | * Page flags also must be visible before we make the page non-compound. | |
2239 | */ | |
2240 | smp_wmb(); | |
2241 | ||
2242 | clear_compound_head(page_tail); | |
2243 | ||
2244 | if (page_is_young(head)) | |
2245 | set_page_young(page_tail); | |
2246 | if (page_is_idle(head)) | |
2247 | set_page_idle(page_tail); | |
2248 | ||
2249 | /* ->mapping in first tail page is compound_mapcount */ | |
2250 | VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING, | |
2251 | page_tail); | |
2252 | page_tail->mapping = head->mapping; | |
2253 | ||
2254 | page_tail->index = head->index + tail; | |
2255 | page_cpupid_xchg_last(page_tail, page_cpupid_last(head)); | |
2256 | lru_add_page_tail(head, page_tail, lruvec, list); | |
2257 | } | |
2258 | ||
2259 | static void __split_huge_page(struct page *page, struct list_head *list, | |
2260 | unsigned long flags) | |
2261 | { | |
2262 | struct page *head = compound_head(page); | |
2263 | struct zone *zone = page_zone(head); | |
2264 | struct lruvec *lruvec; | |
2265 | pgoff_t end = -1; | |
2266 | int i; | |
2267 | ||
2268 | lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat); | |
2269 | ||
2270 | /* complete memcg works before add pages to LRU */ | |
2271 | mem_cgroup_split_huge_fixup(head); | |
2272 | ||
2273 | if (!PageAnon(page)) | |
2274 | end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE); | |
2275 | ||
2276 | for (i = HPAGE_PMD_NR - 1; i >= 1; i--) { | |
2277 | __split_huge_page_tail(head, i, lruvec, list); | |
2278 | /* Some pages can be beyond i_size: drop them from page cache */ | |
2279 | if (head[i].index >= end) { | |
2280 | __ClearPageDirty(head + i); | |
2281 | __delete_from_page_cache(head + i, NULL); | |
2282 | if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head)) | |
2283 | shmem_uncharge(head->mapping->host, 1); | |
2284 | put_page(head + i); | |
2285 | } | |
2286 | } | |
2287 | ||
2288 | ClearPageCompound(head); | |
2289 | /* See comment in __split_huge_page_tail() */ | |
2290 | if (PageAnon(head)) { | |
2291 | /* Additional pin to radix tree of swap cache */ | |
2292 | if (PageSwapCache(head)) | |
2293 | page_ref_add(head, 2); | |
2294 | else | |
2295 | page_ref_inc(head); | |
2296 | } else { | |
2297 | /* Additional pin to radix tree */ | |
2298 | page_ref_add(head, 2); | |
2299 | spin_unlock(&head->mapping->tree_lock); | |
2300 | } | |
2301 | ||
2302 | spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags); | |
2303 | ||
2304 | unfreeze_page(head); | |
2305 | ||
2306 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
2307 | struct page *subpage = head + i; | |
2308 | if (subpage == page) | |
2309 | continue; | |
2310 | unlock_page(subpage); | |
2311 | ||
2312 | /* | |
2313 | * Subpages may be freed if there wasn't any mapping | |
2314 | * like if add_to_swap() is running on a lru page that | |
2315 | * had its mapping zapped. And freeing these pages | |
2316 | * requires taking the lru_lock so we do the put_page | |
2317 | * of the tail pages after the split is complete. | |
2318 | */ | |
2319 | put_page(subpage); | |
2320 | } | |
2321 | } | |
2322 | ||
2323 | int total_mapcount(struct page *page) | |
2324 | { | |
2325 | int i, compound, ret; | |
2326 | ||
2327 | VM_BUG_ON_PAGE(PageTail(page), page); | |
2328 | ||
2329 | if (likely(!PageCompound(page))) | |
2330 | return atomic_read(&page->_mapcount) + 1; | |
2331 | ||
2332 | compound = compound_mapcount(page); | |
2333 | if (PageHuge(page)) | |
2334 | return compound; | |
2335 | ret = compound; | |
2336 | for (i = 0; i < HPAGE_PMD_NR; i++) | |
2337 | ret += atomic_read(&page[i]._mapcount) + 1; | |
2338 | /* File pages has compound_mapcount included in _mapcount */ | |
2339 | if (!PageAnon(page)) | |
2340 | return ret - compound * HPAGE_PMD_NR; | |
2341 | if (PageDoubleMap(page)) | |
2342 | ret -= HPAGE_PMD_NR; | |
2343 | return ret; | |
2344 | } | |
2345 | ||
2346 | /* | |
2347 | * This calculates accurately how many mappings a transparent hugepage | |
2348 | * has (unlike page_mapcount() which isn't fully accurate). This full | |
2349 | * accuracy is primarily needed to know if copy-on-write faults can | |
2350 | * reuse the page and change the mapping to read-write instead of | |
2351 | * copying them. At the same time this returns the total_mapcount too. | |
2352 | * | |
2353 | * The function returns the highest mapcount any one of the subpages | |
2354 | * has. If the return value is one, even if different processes are | |
2355 | * mapping different subpages of the transparent hugepage, they can | |
2356 | * all reuse it, because each process is reusing a different subpage. | |
2357 | * | |
2358 | * The total_mapcount is instead counting all virtual mappings of the | |
2359 | * subpages. If the total_mapcount is equal to "one", it tells the | |
2360 | * caller all mappings belong to the same "mm" and in turn the | |
2361 | * anon_vma of the transparent hugepage can become the vma->anon_vma | |
2362 | * local one as no other process may be mapping any of the subpages. | |
2363 | * | |
2364 | * It would be more accurate to replace page_mapcount() with | |
2365 | * page_trans_huge_mapcount(), however we only use | |
2366 | * page_trans_huge_mapcount() in the copy-on-write faults where we | |
2367 | * need full accuracy to avoid breaking page pinning, because | |
2368 | * page_trans_huge_mapcount() is slower than page_mapcount(). | |
2369 | */ | |
2370 | int page_trans_huge_mapcount(struct page *page, int *total_mapcount) | |
2371 | { | |
2372 | int i, ret, _total_mapcount, mapcount; | |
2373 | ||
2374 | /* hugetlbfs shouldn't call it */ | |
2375 | VM_BUG_ON_PAGE(PageHuge(page), page); | |
2376 | ||
2377 | if (likely(!PageTransCompound(page))) { | |
2378 | mapcount = atomic_read(&page->_mapcount) + 1; | |
2379 | if (total_mapcount) | |
2380 | *total_mapcount = mapcount; | |
2381 | return mapcount; | |
2382 | } | |
2383 | ||
2384 | page = compound_head(page); | |
2385 | ||
2386 | _total_mapcount = ret = 0; | |
2387 | for (i = 0; i < HPAGE_PMD_NR; i++) { | |
2388 | mapcount = atomic_read(&page[i]._mapcount) + 1; | |
2389 | ret = max(ret, mapcount); | |
2390 | _total_mapcount += mapcount; | |
2391 | } | |
2392 | if (PageDoubleMap(page)) { | |
2393 | ret -= 1; | |
2394 | _total_mapcount -= HPAGE_PMD_NR; | |
2395 | } | |
2396 | mapcount = compound_mapcount(page); | |
2397 | ret += mapcount; | |
2398 | _total_mapcount += mapcount; | |
2399 | if (total_mapcount) | |
2400 | *total_mapcount = _total_mapcount; | |
2401 | return ret; | |
2402 | } | |
2403 | ||
2404 | /* Racy check whether the huge page can be split */ | |
2405 | bool can_split_huge_page(struct page *page, int *pextra_pins) | |
2406 | { | |
2407 | int extra_pins; | |
2408 | ||
2409 | /* Additional pins from radix tree */ | |
2410 | if (PageAnon(page)) | |
2411 | extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0; | |
2412 | else | |
2413 | extra_pins = HPAGE_PMD_NR; | |
2414 | if (pextra_pins) | |
2415 | *pextra_pins = extra_pins; | |
2416 | return total_mapcount(page) == page_count(page) - extra_pins - 1; | |
2417 | } | |
2418 | ||
2419 | /* | |
2420 | * This function splits huge page into normal pages. @page can point to any | |
2421 | * subpage of huge page to split. Split doesn't change the position of @page. | |
2422 | * | |
2423 | * Only caller must hold pin on the @page, otherwise split fails with -EBUSY. | |
2424 | * The huge page must be locked. | |
2425 | * | |
2426 | * If @list is null, tail pages will be added to LRU list, otherwise, to @list. | |
2427 | * | |
2428 | * Both head page and tail pages will inherit mapping, flags, and so on from | |
2429 | * the hugepage. | |
2430 | * | |
2431 | * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if | |
2432 | * they are not mapped. | |
2433 | * | |
2434 | * Returns 0 if the hugepage is split successfully. | |
2435 | * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under | |
2436 | * us. | |
2437 | */ | |
2438 | int split_huge_page_to_list(struct page *page, struct list_head *list) | |
2439 | { | |
2440 | struct page *head = compound_head(page); | |
2441 | struct pglist_data *pgdata = NODE_DATA(page_to_nid(head)); | |
2442 | struct anon_vma *anon_vma = NULL; | |
2443 | struct address_space *mapping = NULL; | |
2444 | int count, mapcount, extra_pins, ret; | |
2445 | bool mlocked; | |
2446 | unsigned long flags; | |
2447 | ||
2448 | VM_BUG_ON_PAGE(is_huge_zero_page(page), page); | |
2449 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
2450 | VM_BUG_ON_PAGE(!PageCompound(page), page); | |
2451 | ||
2452 | if (PageAnon(head)) { | |
2453 | /* | |
2454 | * The caller does not necessarily hold an mmap_sem that would | |
2455 | * prevent the anon_vma disappearing so we first we take a | |
2456 | * reference to it and then lock the anon_vma for write. This | |
2457 | * is similar to page_lock_anon_vma_read except the write lock | |
2458 | * is taken to serialise against parallel split or collapse | |
2459 | * operations. | |
2460 | */ | |
2461 | anon_vma = page_get_anon_vma(head); | |
2462 | if (!anon_vma) { | |
2463 | ret = -EBUSY; | |
2464 | goto out; | |
2465 | } | |
2466 | mapping = NULL; | |
2467 | anon_vma_lock_write(anon_vma); | |
2468 | } else { | |
2469 | mapping = head->mapping; | |
2470 | ||
2471 | /* Truncated ? */ | |
2472 | if (!mapping) { | |
2473 | ret = -EBUSY; | |
2474 | goto out; | |
2475 | } | |
2476 | ||
2477 | anon_vma = NULL; | |
2478 | i_mmap_lock_read(mapping); | |
2479 | } | |
2480 | ||
2481 | /* | |
2482 | * Racy check if we can split the page, before freeze_page() will | |
2483 | * split PMDs | |
2484 | */ | |
2485 | if (!can_split_huge_page(head, &extra_pins)) { | |
2486 | ret = -EBUSY; | |
2487 | goto out_unlock; | |
2488 | } | |
2489 | ||
2490 | mlocked = PageMlocked(page); | |
2491 | freeze_page(head); | |
2492 | VM_BUG_ON_PAGE(compound_mapcount(head), head); | |
2493 | ||
2494 | /* Make sure the page is not on per-CPU pagevec as it takes pin */ | |
2495 | if (mlocked) | |
2496 | lru_add_drain(); | |
2497 | ||
2498 | /* prevent PageLRU to go away from under us, and freeze lru stats */ | |
2499 | spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags); | |
2500 | ||
2501 | if (mapping) { | |
2502 | void **pslot; | |
2503 | ||
2504 | spin_lock(&mapping->tree_lock); | |
2505 | pslot = radix_tree_lookup_slot(&mapping->page_tree, | |
2506 | page_index(head)); | |
2507 | /* | |
2508 | * Check if the head page is present in radix tree. | |
2509 | * We assume all tail are present too, if head is there. | |
2510 | */ | |
2511 | if (radix_tree_deref_slot_protected(pslot, | |
2512 | &mapping->tree_lock) != head) | |
2513 | goto fail; | |
2514 | } | |
2515 | ||
2516 | /* Prevent deferred_split_scan() touching ->_refcount */ | |
2517 | spin_lock(&pgdata->split_queue_lock); | |
2518 | count = page_count(head); | |
2519 | mapcount = total_mapcount(head); | |
2520 | if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) { | |
2521 | if (!list_empty(page_deferred_list(head))) { | |
2522 | pgdata->split_queue_len--; | |
2523 | list_del(page_deferred_list(head)); | |
2524 | } | |
2525 | if (mapping) | |
2526 | __dec_node_page_state(page, NR_SHMEM_THPS); | |
2527 | spin_unlock(&pgdata->split_queue_lock); | |
2528 | __split_huge_page(page, list, flags); | |
2529 | ret = 0; | |
2530 | } else { | |
2531 | if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) { | |
2532 | pr_alert("total_mapcount: %u, page_count(): %u\n", | |
2533 | mapcount, count); | |
2534 | if (PageTail(page)) | |
2535 | dump_page(head, NULL); | |
2536 | dump_page(page, "total_mapcount(head) > 0"); | |
2537 | BUG(); | |
2538 | } | |
2539 | spin_unlock(&pgdata->split_queue_lock); | |
2540 | fail: if (mapping) | |
2541 | spin_unlock(&mapping->tree_lock); | |
2542 | spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags); | |
2543 | unfreeze_page(head); | |
2544 | ret = -EBUSY; | |
2545 | } | |
2546 | ||
2547 | out_unlock: | |
2548 | if (anon_vma) { | |
2549 | anon_vma_unlock_write(anon_vma); | |
2550 | put_anon_vma(anon_vma); | |
2551 | } | |
2552 | if (mapping) | |
2553 | i_mmap_unlock_read(mapping); | |
2554 | out: | |
2555 | count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED); | |
2556 | return ret; | |
2557 | } | |
2558 | ||
2559 | void free_transhuge_page(struct page *page) | |
2560 | { | |
2561 | struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); | |
2562 | unsigned long flags; | |
2563 | ||
2564 | spin_lock_irqsave(&pgdata->split_queue_lock, flags); | |
2565 | if (!list_empty(page_deferred_list(page))) { | |
2566 | pgdata->split_queue_len--; | |
2567 | list_del(page_deferred_list(page)); | |
2568 | } | |
2569 | spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); | |
2570 | free_compound_page(page); | |
2571 | } | |
2572 | ||
2573 | void deferred_split_huge_page(struct page *page) | |
2574 | { | |
2575 | struct pglist_data *pgdata = NODE_DATA(page_to_nid(page)); | |
2576 | unsigned long flags; | |
2577 | ||
2578 | VM_BUG_ON_PAGE(!PageTransHuge(page), page); | |
2579 | ||
2580 | spin_lock_irqsave(&pgdata->split_queue_lock, flags); | |
2581 | if (list_empty(page_deferred_list(page))) { | |
2582 | count_vm_event(THP_DEFERRED_SPLIT_PAGE); | |
2583 | list_add_tail(page_deferred_list(page), &pgdata->split_queue); | |
2584 | pgdata->split_queue_len++; | |
2585 | } | |
2586 | spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); | |
2587 | } | |
2588 | ||
2589 | static unsigned long deferred_split_count(struct shrinker *shrink, | |
2590 | struct shrink_control *sc) | |
2591 | { | |
2592 | struct pglist_data *pgdata = NODE_DATA(sc->nid); | |
2593 | return ACCESS_ONCE(pgdata->split_queue_len); | |
2594 | } | |
2595 | ||
2596 | static unsigned long deferred_split_scan(struct shrinker *shrink, | |
2597 | struct shrink_control *sc) | |
2598 | { | |
2599 | struct pglist_data *pgdata = NODE_DATA(sc->nid); | |
2600 | unsigned long flags; | |
2601 | LIST_HEAD(list), *pos, *next; | |
2602 | struct page *page; | |
2603 | int split = 0; | |
2604 | ||
2605 | spin_lock_irqsave(&pgdata->split_queue_lock, flags); | |
2606 | /* Take pin on all head pages to avoid freeing them under us */ | |
2607 | list_for_each_safe(pos, next, &pgdata->split_queue) { | |
2608 | page = list_entry((void *)pos, struct page, mapping); | |
2609 | page = compound_head(page); | |
2610 | if (get_page_unless_zero(page)) { | |
2611 | list_move(page_deferred_list(page), &list); | |
2612 | } else { | |
2613 | /* We lost race with put_compound_page() */ | |
2614 | list_del_init(page_deferred_list(page)); | |
2615 | pgdata->split_queue_len--; | |
2616 | } | |
2617 | if (!--sc->nr_to_scan) | |
2618 | break; | |
2619 | } | |
2620 | spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); | |
2621 | ||
2622 | list_for_each_safe(pos, next, &list) { | |
2623 | page = list_entry((void *)pos, struct page, mapping); | |
2624 | lock_page(page); | |
2625 | /* split_huge_page() removes page from list on success */ | |
2626 | if (!split_huge_page(page)) | |
2627 | split++; | |
2628 | unlock_page(page); | |
2629 | put_page(page); | |
2630 | } | |
2631 | ||
2632 | spin_lock_irqsave(&pgdata->split_queue_lock, flags); | |
2633 | list_splice_tail(&list, &pgdata->split_queue); | |
2634 | spin_unlock_irqrestore(&pgdata->split_queue_lock, flags); | |
2635 | ||
2636 | /* | |
2637 | * Stop shrinker if we didn't split any page, but the queue is empty. | |
2638 | * This can happen if pages were freed under us. | |
2639 | */ | |
2640 | if (!split && list_empty(&pgdata->split_queue)) | |
2641 | return SHRINK_STOP; | |
2642 | return split; | |
2643 | } | |
2644 | ||
2645 | static struct shrinker deferred_split_shrinker = { | |
2646 | .count_objects = deferred_split_count, | |
2647 | .scan_objects = deferred_split_scan, | |
2648 | .seeks = DEFAULT_SEEKS, | |
2649 | .flags = SHRINKER_NUMA_AWARE, | |
2650 | }; | |
2651 | ||
2652 | #ifdef CONFIG_DEBUG_FS | |
2653 | static int split_huge_pages_set(void *data, u64 val) | |
2654 | { | |
2655 | struct zone *zone; | |
2656 | struct page *page; | |
2657 | unsigned long pfn, max_zone_pfn; | |
2658 | unsigned long total = 0, split = 0; | |
2659 | ||
2660 | if (val != 1) | |
2661 | return -EINVAL; | |
2662 | ||
2663 | for_each_populated_zone(zone) { | |
2664 | max_zone_pfn = zone_end_pfn(zone); | |
2665 | for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) { | |
2666 | if (!pfn_valid(pfn)) | |
2667 | continue; | |
2668 | ||
2669 | page = pfn_to_page(pfn); | |
2670 | if (!get_page_unless_zero(page)) | |
2671 | continue; | |
2672 | ||
2673 | if (zone != page_zone(page)) | |
2674 | goto next; | |
2675 | ||
2676 | if (!PageHead(page) || PageHuge(page) || !PageLRU(page)) | |
2677 | goto next; | |
2678 | ||
2679 | total++; | |
2680 | lock_page(page); | |
2681 | if (!split_huge_page(page)) | |
2682 | split++; | |
2683 | unlock_page(page); | |
2684 | next: | |
2685 | put_page(page); | |
2686 | } | |
2687 | } | |
2688 | ||
2689 | pr_info("%lu of %lu THP split\n", split, total); | |
2690 | ||
2691 | return 0; | |
2692 | } | |
2693 | DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set, | |
2694 | "%llu\n"); | |
2695 | ||
2696 | static int __init split_huge_pages_debugfs(void) | |
2697 | { | |
2698 | void *ret; | |
2699 | ||
2700 | ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL, | |
2701 | &split_huge_pages_fops); | |
2702 | if (!ret) | |
2703 | pr_warn("Failed to create split_huge_pages in debugfs"); | |
2704 | return 0; | |
2705 | } | |
2706 | late_initcall(split_huge_pages_debugfs); | |
2707 | #endif |