xref: /openbmc/linux/mm/huge_memory.c (revision 711aab1d)
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 const 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, vmf->address, 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)
846 {
847 	pmd_t _pmd;
848 
849 	/*
850 	 * We should set the dirty bit only for FOLL_WRITE but for now
851 	 * the dirty bit in the pmd is meaningless.  And if the dirty
852 	 * bit will become meaningful and we'll only set it with
853 	 * FOLL_WRITE, an atomic set_bit will be required on the pmd to
854 	 * set the young bit, instead of the current set_pmd_at.
855 	 */
856 	_pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
857 	if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
858 				pmd, _pmd,  1))
859 		update_mmu_cache_pmd(vma, addr, pmd);
860 }
861 
862 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
863 		pmd_t *pmd, int flags)
864 {
865 	unsigned long pfn = pmd_pfn(*pmd);
866 	struct mm_struct *mm = vma->vm_mm;
867 	struct dev_pagemap *pgmap;
868 	struct page *page;
869 
870 	assert_spin_locked(pmd_lockptr(mm, pmd));
871 
872 	/*
873 	 * When we COW a devmap PMD entry, we split it into PTEs, so we should
874 	 * not be in this function with `flags & FOLL_COW` set.
875 	 */
876 	WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
877 
878 	if (flags & FOLL_WRITE && !pmd_write(*pmd))
879 		return NULL;
880 
881 	if (pmd_present(*pmd) && pmd_devmap(*pmd))
882 		/* pass */;
883 	else
884 		return NULL;
885 
886 	if (flags & FOLL_TOUCH)
887 		touch_pmd(vma, addr, pmd);
888 
889 	/*
890 	 * device mapped pages can only be returned if the
891 	 * caller will manage the page reference count.
892 	 */
893 	if (!(flags & FOLL_GET))
894 		return ERR_PTR(-EEXIST);
895 
896 	pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
897 	pgmap = get_dev_pagemap(pfn, NULL);
898 	if (!pgmap)
899 		return ERR_PTR(-EFAULT);
900 	page = pfn_to_page(pfn);
901 	get_page(page);
902 	put_dev_pagemap(pgmap);
903 
904 	return page;
905 }
906 
907 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
908 		  pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
909 		  struct vm_area_struct *vma)
910 {
911 	spinlock_t *dst_ptl, *src_ptl;
912 	struct page *src_page;
913 	pmd_t pmd;
914 	pgtable_t pgtable = NULL;
915 	int ret = -ENOMEM;
916 
917 	/* Skip if can be re-fill on fault */
918 	if (!vma_is_anonymous(vma))
919 		return 0;
920 
921 	pgtable = pte_alloc_one(dst_mm, addr);
922 	if (unlikely(!pgtable))
923 		goto out;
924 
925 	dst_ptl = pmd_lock(dst_mm, dst_pmd);
926 	src_ptl = pmd_lockptr(src_mm, src_pmd);
927 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
928 
929 	ret = -EAGAIN;
930 	pmd = *src_pmd;
931 
932 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
933 	if (unlikely(is_swap_pmd(pmd))) {
934 		swp_entry_t entry = pmd_to_swp_entry(pmd);
935 
936 		VM_BUG_ON(!is_pmd_migration_entry(pmd));
937 		if (is_write_migration_entry(entry)) {
938 			make_migration_entry_read(&entry);
939 			pmd = swp_entry_to_pmd(entry);
940 			if (pmd_swp_soft_dirty(*src_pmd))
941 				pmd = pmd_swp_mksoft_dirty(pmd);
942 			set_pmd_at(src_mm, addr, src_pmd, pmd);
943 		}
944 		set_pmd_at(dst_mm, addr, dst_pmd, pmd);
945 		ret = 0;
946 		goto out_unlock;
947 	}
948 #endif
949 
950 	if (unlikely(!pmd_trans_huge(pmd))) {
951 		pte_free(dst_mm, pgtable);
952 		goto out_unlock;
953 	}
954 	/*
955 	 * When page table lock is held, the huge zero pmd should not be
956 	 * under splitting since we don't split the page itself, only pmd to
957 	 * a page table.
958 	 */
959 	if (is_huge_zero_pmd(pmd)) {
960 		struct page *zero_page;
961 		/*
962 		 * get_huge_zero_page() will never allocate a new page here,
963 		 * since we already have a zero page to copy. It just takes a
964 		 * reference.
965 		 */
966 		zero_page = mm_get_huge_zero_page(dst_mm);
967 		set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
968 				zero_page);
969 		ret = 0;
970 		goto out_unlock;
971 	}
972 
973 	src_page = pmd_page(pmd);
974 	VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
975 	get_page(src_page);
976 	page_dup_rmap(src_page, true);
977 	add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
978 	atomic_long_inc(&dst_mm->nr_ptes);
979 	pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
980 
981 	pmdp_set_wrprotect(src_mm, addr, src_pmd);
982 	pmd = pmd_mkold(pmd_wrprotect(pmd));
983 	set_pmd_at(dst_mm, addr, dst_pmd, pmd);
984 
985 	ret = 0;
986 out_unlock:
987 	spin_unlock(src_ptl);
988 	spin_unlock(dst_ptl);
989 out:
990 	return ret;
991 }
992 
993 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
994 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
995 		pud_t *pud)
996 {
997 	pud_t _pud;
998 
999 	/*
1000 	 * We should set the dirty bit only for FOLL_WRITE but for now
1001 	 * the dirty bit in the pud is meaningless.  And if the dirty
1002 	 * bit will become meaningful and we'll only set it with
1003 	 * FOLL_WRITE, an atomic set_bit will be required on the pud to
1004 	 * set the young bit, instead of the current set_pud_at.
1005 	 */
1006 	_pud = pud_mkyoung(pud_mkdirty(*pud));
1007 	if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1008 				pud, _pud,  1))
1009 		update_mmu_cache_pud(vma, addr, pud);
1010 }
1011 
1012 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1013 		pud_t *pud, int flags)
1014 {
1015 	unsigned long pfn = pud_pfn(*pud);
1016 	struct mm_struct *mm = vma->vm_mm;
1017 	struct dev_pagemap *pgmap;
1018 	struct page *page;
1019 
1020 	assert_spin_locked(pud_lockptr(mm, pud));
1021 
1022 	if (flags & FOLL_WRITE && !pud_write(*pud))
1023 		return NULL;
1024 
1025 	if (pud_present(*pud) && pud_devmap(*pud))
1026 		/* pass */;
1027 	else
1028 		return NULL;
1029 
1030 	if (flags & FOLL_TOUCH)
1031 		touch_pud(vma, addr, pud);
1032 
1033 	/*
1034 	 * device mapped pages can only be returned if the
1035 	 * caller will manage the page reference count.
1036 	 */
1037 	if (!(flags & FOLL_GET))
1038 		return ERR_PTR(-EEXIST);
1039 
1040 	pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1041 	pgmap = get_dev_pagemap(pfn, NULL);
1042 	if (!pgmap)
1043 		return ERR_PTR(-EFAULT);
1044 	page = pfn_to_page(pfn);
1045 	get_page(page);
1046 	put_dev_pagemap(pgmap);
1047 
1048 	return page;
1049 }
1050 
1051 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1052 		  pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1053 		  struct vm_area_struct *vma)
1054 {
1055 	spinlock_t *dst_ptl, *src_ptl;
1056 	pud_t pud;
1057 	int ret;
1058 
1059 	dst_ptl = pud_lock(dst_mm, dst_pud);
1060 	src_ptl = pud_lockptr(src_mm, src_pud);
1061 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1062 
1063 	ret = -EAGAIN;
1064 	pud = *src_pud;
1065 	if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1066 		goto out_unlock;
1067 
1068 	/*
1069 	 * When page table lock is held, the huge zero pud should not be
1070 	 * under splitting since we don't split the page itself, only pud to
1071 	 * a page table.
1072 	 */
1073 	if (is_huge_zero_pud(pud)) {
1074 		/* No huge zero pud yet */
1075 	}
1076 
1077 	pudp_set_wrprotect(src_mm, addr, src_pud);
1078 	pud = pud_mkold(pud_wrprotect(pud));
1079 	set_pud_at(dst_mm, addr, dst_pud, pud);
1080 
1081 	ret = 0;
1082 out_unlock:
1083 	spin_unlock(src_ptl);
1084 	spin_unlock(dst_ptl);
1085 	return ret;
1086 }
1087 
1088 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1089 {
1090 	pud_t entry;
1091 	unsigned long haddr;
1092 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1093 
1094 	vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1095 	if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1096 		goto unlock;
1097 
1098 	entry = pud_mkyoung(orig_pud);
1099 	if (write)
1100 		entry = pud_mkdirty(entry);
1101 	haddr = vmf->address & HPAGE_PUD_MASK;
1102 	if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1103 		update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1104 
1105 unlock:
1106 	spin_unlock(vmf->ptl);
1107 }
1108 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1109 
1110 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1111 {
1112 	pmd_t entry;
1113 	unsigned long haddr;
1114 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1115 
1116 	vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1117 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1118 		goto unlock;
1119 
1120 	entry = pmd_mkyoung(orig_pmd);
1121 	if (write)
1122 		entry = pmd_mkdirty(entry);
1123 	haddr = vmf->address & HPAGE_PMD_MASK;
1124 	if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1125 		update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1126 
1127 unlock:
1128 	spin_unlock(vmf->ptl);
1129 }
1130 
1131 static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd,
1132 		struct page *page)
1133 {
1134 	struct vm_area_struct *vma = vmf->vma;
1135 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1136 	struct mem_cgroup *memcg;
1137 	pgtable_t pgtable;
1138 	pmd_t _pmd;
1139 	int ret = 0, i;
1140 	struct page **pages;
1141 	unsigned long mmun_start;	/* For mmu_notifiers */
1142 	unsigned long mmun_end;		/* For mmu_notifiers */
1143 
1144 	pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1145 			GFP_KERNEL);
1146 	if (unlikely(!pages)) {
1147 		ret |= VM_FAULT_OOM;
1148 		goto out;
1149 	}
1150 
1151 	for (i = 0; i < HPAGE_PMD_NR; i++) {
1152 		pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1153 					       vmf->address, page_to_nid(page));
1154 		if (unlikely(!pages[i] ||
1155 			     mem_cgroup_try_charge(pages[i], vma->vm_mm,
1156 				     GFP_KERNEL, &memcg, false))) {
1157 			if (pages[i])
1158 				put_page(pages[i]);
1159 			while (--i >= 0) {
1160 				memcg = (void *)page_private(pages[i]);
1161 				set_page_private(pages[i], 0);
1162 				mem_cgroup_cancel_charge(pages[i], memcg,
1163 						false);
1164 				put_page(pages[i]);
1165 			}
1166 			kfree(pages);
1167 			ret |= VM_FAULT_OOM;
1168 			goto out;
1169 		}
1170 		set_page_private(pages[i], (unsigned long)memcg);
1171 	}
1172 
1173 	for (i = 0; i < HPAGE_PMD_NR; i++) {
1174 		copy_user_highpage(pages[i], page + i,
1175 				   haddr + PAGE_SIZE * i, vma);
1176 		__SetPageUptodate(pages[i]);
1177 		cond_resched();
1178 	}
1179 
1180 	mmun_start = haddr;
1181 	mmun_end   = haddr + HPAGE_PMD_SIZE;
1182 	mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1183 
1184 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1185 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1186 		goto out_free_pages;
1187 	VM_BUG_ON_PAGE(!PageHead(page), page);
1188 
1189 	pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1190 	/* leave pmd empty until pte is filled */
1191 
1192 	pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1193 	pmd_populate(vma->vm_mm, &_pmd, pgtable);
1194 
1195 	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1196 		pte_t entry;
1197 		entry = mk_pte(pages[i], vma->vm_page_prot);
1198 		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1199 		memcg = (void *)page_private(pages[i]);
1200 		set_page_private(pages[i], 0);
1201 		page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1202 		mem_cgroup_commit_charge(pages[i], memcg, false, false);
1203 		lru_cache_add_active_or_unevictable(pages[i], vma);
1204 		vmf->pte = pte_offset_map(&_pmd, haddr);
1205 		VM_BUG_ON(!pte_none(*vmf->pte));
1206 		set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1207 		pte_unmap(vmf->pte);
1208 	}
1209 	kfree(pages);
1210 
1211 	smp_wmb(); /* make pte visible before pmd */
1212 	pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1213 	page_remove_rmap(page, true);
1214 	spin_unlock(vmf->ptl);
1215 
1216 	mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1217 
1218 	ret |= VM_FAULT_WRITE;
1219 	put_page(page);
1220 
1221 out:
1222 	return ret;
1223 
1224 out_free_pages:
1225 	spin_unlock(vmf->ptl);
1226 	mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1227 	for (i = 0; i < HPAGE_PMD_NR; i++) {
1228 		memcg = (void *)page_private(pages[i]);
1229 		set_page_private(pages[i], 0);
1230 		mem_cgroup_cancel_charge(pages[i], memcg, false);
1231 		put_page(pages[i]);
1232 	}
1233 	kfree(pages);
1234 	goto out;
1235 }
1236 
1237 int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1238 {
1239 	struct vm_area_struct *vma = vmf->vma;
1240 	struct page *page = NULL, *new_page;
1241 	struct mem_cgroup *memcg;
1242 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1243 	unsigned long mmun_start;	/* For mmu_notifiers */
1244 	unsigned long mmun_end;		/* For mmu_notifiers */
1245 	gfp_t huge_gfp;			/* for allocation and charge */
1246 	int ret = 0;
1247 
1248 	vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1249 	VM_BUG_ON_VMA(!vma->anon_vma, vma);
1250 	if (is_huge_zero_pmd(orig_pmd))
1251 		goto alloc;
1252 	spin_lock(vmf->ptl);
1253 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1254 		goto out_unlock;
1255 
1256 	page = pmd_page(orig_pmd);
1257 	VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1258 	/*
1259 	 * We can only reuse the page if nobody else maps the huge page or it's
1260 	 * part.
1261 	 */
1262 	if (!trylock_page(page)) {
1263 		get_page(page);
1264 		spin_unlock(vmf->ptl);
1265 		lock_page(page);
1266 		spin_lock(vmf->ptl);
1267 		if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1268 			unlock_page(page);
1269 			put_page(page);
1270 			goto out_unlock;
1271 		}
1272 		put_page(page);
1273 	}
1274 	if (reuse_swap_page(page, NULL)) {
1275 		pmd_t entry;
1276 		entry = pmd_mkyoung(orig_pmd);
1277 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1278 		if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1279 			update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1280 		ret |= VM_FAULT_WRITE;
1281 		unlock_page(page);
1282 		goto out_unlock;
1283 	}
1284 	unlock_page(page);
1285 	get_page(page);
1286 	spin_unlock(vmf->ptl);
1287 alloc:
1288 	if (transparent_hugepage_enabled(vma) &&
1289 	    !transparent_hugepage_debug_cow()) {
1290 		huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1291 		new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1292 	} else
1293 		new_page = NULL;
1294 
1295 	if (likely(new_page)) {
1296 		prep_transhuge_page(new_page);
1297 	} else {
1298 		if (!page) {
1299 			split_huge_pmd(vma, vmf->pmd, vmf->address);
1300 			ret |= VM_FAULT_FALLBACK;
1301 		} else {
1302 			ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1303 			if (ret & VM_FAULT_OOM) {
1304 				split_huge_pmd(vma, vmf->pmd, vmf->address);
1305 				ret |= VM_FAULT_FALLBACK;
1306 			}
1307 			put_page(page);
1308 		}
1309 		count_vm_event(THP_FAULT_FALLBACK);
1310 		goto out;
1311 	}
1312 
1313 	if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1314 					huge_gfp, &memcg, true))) {
1315 		put_page(new_page);
1316 		split_huge_pmd(vma, vmf->pmd, vmf->address);
1317 		if (page)
1318 			put_page(page);
1319 		ret |= VM_FAULT_FALLBACK;
1320 		count_vm_event(THP_FAULT_FALLBACK);
1321 		goto out;
1322 	}
1323 
1324 	count_vm_event(THP_FAULT_ALLOC);
1325 
1326 	if (!page)
1327 		clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1328 	else
1329 		copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1330 	__SetPageUptodate(new_page);
1331 
1332 	mmun_start = haddr;
1333 	mmun_end   = haddr + HPAGE_PMD_SIZE;
1334 	mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1335 
1336 	spin_lock(vmf->ptl);
1337 	if (page)
1338 		put_page(page);
1339 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1340 		spin_unlock(vmf->ptl);
1341 		mem_cgroup_cancel_charge(new_page, memcg, true);
1342 		put_page(new_page);
1343 		goto out_mn;
1344 	} else {
1345 		pmd_t entry;
1346 		entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1347 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1348 		pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1349 		page_add_new_anon_rmap(new_page, vma, haddr, true);
1350 		mem_cgroup_commit_charge(new_page, memcg, false, true);
1351 		lru_cache_add_active_or_unevictable(new_page, vma);
1352 		set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1353 		update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1354 		if (!page) {
1355 			add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1356 		} else {
1357 			VM_BUG_ON_PAGE(!PageHead(page), page);
1358 			page_remove_rmap(page, true);
1359 			put_page(page);
1360 		}
1361 		ret |= VM_FAULT_WRITE;
1362 	}
1363 	spin_unlock(vmf->ptl);
1364 out_mn:
1365 	mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1366 out:
1367 	return ret;
1368 out_unlock:
1369 	spin_unlock(vmf->ptl);
1370 	return ret;
1371 }
1372 
1373 /*
1374  * FOLL_FORCE can write to even unwritable pmd's, but only
1375  * after we've gone through a COW cycle and they are dirty.
1376  */
1377 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1378 {
1379 	return pmd_write(pmd) ||
1380 	       ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1381 }
1382 
1383 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1384 				   unsigned long addr,
1385 				   pmd_t *pmd,
1386 				   unsigned int flags)
1387 {
1388 	struct mm_struct *mm = vma->vm_mm;
1389 	struct page *page = NULL;
1390 
1391 	assert_spin_locked(pmd_lockptr(mm, pmd));
1392 
1393 	if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1394 		goto out;
1395 
1396 	/* Avoid dumping huge zero page */
1397 	if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1398 		return ERR_PTR(-EFAULT);
1399 
1400 	/* Full NUMA hinting faults to serialise migration in fault paths */
1401 	if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1402 		goto out;
1403 
1404 	page = pmd_page(*pmd);
1405 	VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1406 	if (flags & FOLL_TOUCH)
1407 		touch_pmd(vma, addr, pmd);
1408 	if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1409 		/*
1410 		 * We don't mlock() pte-mapped THPs. This way we can avoid
1411 		 * leaking mlocked pages into non-VM_LOCKED VMAs.
1412 		 *
1413 		 * For anon THP:
1414 		 *
1415 		 * In most cases the pmd is the only mapping of the page as we
1416 		 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1417 		 * writable private mappings in populate_vma_page_range().
1418 		 *
1419 		 * The only scenario when we have the page shared here is if we
1420 		 * mlocking read-only mapping shared over fork(). We skip
1421 		 * mlocking such pages.
1422 		 *
1423 		 * For file THP:
1424 		 *
1425 		 * We can expect PageDoubleMap() to be stable under page lock:
1426 		 * for file pages we set it in page_add_file_rmap(), which
1427 		 * requires page to be locked.
1428 		 */
1429 
1430 		if (PageAnon(page) && compound_mapcount(page) != 1)
1431 			goto skip_mlock;
1432 		if (PageDoubleMap(page) || !page->mapping)
1433 			goto skip_mlock;
1434 		if (!trylock_page(page))
1435 			goto skip_mlock;
1436 		lru_add_drain();
1437 		if (page->mapping && !PageDoubleMap(page))
1438 			mlock_vma_page(page);
1439 		unlock_page(page);
1440 	}
1441 skip_mlock:
1442 	page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1443 	VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1444 	if (flags & FOLL_GET)
1445 		get_page(page);
1446 
1447 out:
1448 	return page;
1449 }
1450 
1451 /* NUMA hinting page fault entry point for trans huge pmds */
1452 int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1453 {
1454 	struct vm_area_struct *vma = vmf->vma;
1455 	struct anon_vma *anon_vma = NULL;
1456 	struct page *page;
1457 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1458 	int page_nid = -1, this_nid = numa_node_id();
1459 	int target_nid, last_cpupid = -1;
1460 	bool page_locked;
1461 	bool migrated = false;
1462 	bool was_writable;
1463 	int flags = 0;
1464 
1465 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1466 	if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1467 		goto out_unlock;
1468 
1469 	/*
1470 	 * If there are potential migrations, wait for completion and retry
1471 	 * without disrupting NUMA hinting information. Do not relock and
1472 	 * check_same as the page may no longer be mapped.
1473 	 */
1474 	if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1475 		page = pmd_page(*vmf->pmd);
1476 		if (!get_page_unless_zero(page))
1477 			goto out_unlock;
1478 		spin_unlock(vmf->ptl);
1479 		wait_on_page_locked(page);
1480 		put_page(page);
1481 		goto out;
1482 	}
1483 
1484 	page = pmd_page(pmd);
1485 	BUG_ON(is_huge_zero_page(page));
1486 	page_nid = page_to_nid(page);
1487 	last_cpupid = page_cpupid_last(page);
1488 	count_vm_numa_event(NUMA_HINT_FAULTS);
1489 	if (page_nid == this_nid) {
1490 		count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1491 		flags |= TNF_FAULT_LOCAL;
1492 	}
1493 
1494 	/* See similar comment in do_numa_page for explanation */
1495 	if (!pmd_savedwrite(pmd))
1496 		flags |= TNF_NO_GROUP;
1497 
1498 	/*
1499 	 * Acquire the page lock to serialise THP migrations but avoid dropping
1500 	 * page_table_lock if at all possible
1501 	 */
1502 	page_locked = trylock_page(page);
1503 	target_nid = mpol_misplaced(page, vma, haddr);
1504 	if (target_nid == -1) {
1505 		/* If the page was locked, there are no parallel migrations */
1506 		if (page_locked)
1507 			goto clear_pmdnuma;
1508 	}
1509 
1510 	/* Migration could have started since the pmd_trans_migrating check */
1511 	if (!page_locked) {
1512 		page_nid = -1;
1513 		if (!get_page_unless_zero(page))
1514 			goto out_unlock;
1515 		spin_unlock(vmf->ptl);
1516 		wait_on_page_locked(page);
1517 		put_page(page);
1518 		goto out;
1519 	}
1520 
1521 	/*
1522 	 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1523 	 * to serialises splits
1524 	 */
1525 	get_page(page);
1526 	spin_unlock(vmf->ptl);
1527 	anon_vma = page_lock_anon_vma_read(page);
1528 
1529 	/* Confirm the PMD did not change while page_table_lock was released */
1530 	spin_lock(vmf->ptl);
1531 	if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1532 		unlock_page(page);
1533 		put_page(page);
1534 		page_nid = -1;
1535 		goto out_unlock;
1536 	}
1537 
1538 	/* Bail if we fail to protect against THP splits for any reason */
1539 	if (unlikely(!anon_vma)) {
1540 		put_page(page);
1541 		page_nid = -1;
1542 		goto clear_pmdnuma;
1543 	}
1544 
1545 	/*
1546 	 * Since we took the NUMA fault, we must have observed the !accessible
1547 	 * bit. Make sure all other CPUs agree with that, to avoid them
1548 	 * modifying the page we're about to migrate.
1549 	 *
1550 	 * Must be done under PTL such that we'll observe the relevant
1551 	 * inc_tlb_flush_pending().
1552 	 *
1553 	 * We are not sure a pending tlb flush here is for a huge page
1554 	 * mapping or not. Hence use the tlb range variant
1555 	 */
1556 	if (mm_tlb_flush_pending(vma->vm_mm))
1557 		flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1558 
1559 	/*
1560 	 * Migrate the THP to the requested node, returns with page unlocked
1561 	 * and access rights restored.
1562 	 */
1563 	spin_unlock(vmf->ptl);
1564 
1565 	migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1566 				vmf->pmd, pmd, vmf->address, page, target_nid);
1567 	if (migrated) {
1568 		flags |= TNF_MIGRATED;
1569 		page_nid = target_nid;
1570 	} else
1571 		flags |= TNF_MIGRATE_FAIL;
1572 
1573 	goto out;
1574 clear_pmdnuma:
1575 	BUG_ON(!PageLocked(page));
1576 	was_writable = pmd_savedwrite(pmd);
1577 	pmd = pmd_modify(pmd, vma->vm_page_prot);
1578 	pmd = pmd_mkyoung(pmd);
1579 	if (was_writable)
1580 		pmd = pmd_mkwrite(pmd);
1581 	set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1582 	update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1583 	unlock_page(page);
1584 out_unlock:
1585 	spin_unlock(vmf->ptl);
1586 
1587 out:
1588 	if (anon_vma)
1589 		page_unlock_anon_vma_read(anon_vma);
1590 
1591 	if (page_nid != -1)
1592 		task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1593 				flags);
1594 
1595 	return 0;
1596 }
1597 
1598 /*
1599  * Return true if we do MADV_FREE successfully on entire pmd page.
1600  * Otherwise, return false.
1601  */
1602 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1603 		pmd_t *pmd, unsigned long addr, unsigned long next)
1604 {
1605 	spinlock_t *ptl;
1606 	pmd_t orig_pmd;
1607 	struct page *page;
1608 	struct mm_struct *mm = tlb->mm;
1609 	bool ret = false;
1610 
1611 	tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1612 
1613 	ptl = pmd_trans_huge_lock(pmd, vma);
1614 	if (!ptl)
1615 		goto out_unlocked;
1616 
1617 	orig_pmd = *pmd;
1618 	if (is_huge_zero_pmd(orig_pmd))
1619 		goto out;
1620 
1621 	if (unlikely(!pmd_present(orig_pmd))) {
1622 		VM_BUG_ON(thp_migration_supported() &&
1623 				  !is_pmd_migration_entry(orig_pmd));
1624 		goto out;
1625 	}
1626 
1627 	page = pmd_page(orig_pmd);
1628 	/*
1629 	 * If other processes are mapping this page, we couldn't discard
1630 	 * the page unless they all do MADV_FREE so let's skip the page.
1631 	 */
1632 	if (page_mapcount(page) != 1)
1633 		goto out;
1634 
1635 	if (!trylock_page(page))
1636 		goto out;
1637 
1638 	/*
1639 	 * If user want to discard part-pages of THP, split it so MADV_FREE
1640 	 * will deactivate only them.
1641 	 */
1642 	if (next - addr != HPAGE_PMD_SIZE) {
1643 		get_page(page);
1644 		spin_unlock(ptl);
1645 		split_huge_page(page);
1646 		unlock_page(page);
1647 		put_page(page);
1648 		goto out_unlocked;
1649 	}
1650 
1651 	if (PageDirty(page))
1652 		ClearPageDirty(page);
1653 	unlock_page(page);
1654 
1655 	if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1656 		pmdp_invalidate(vma, addr, pmd);
1657 		orig_pmd = pmd_mkold(orig_pmd);
1658 		orig_pmd = pmd_mkclean(orig_pmd);
1659 
1660 		set_pmd_at(mm, addr, pmd, orig_pmd);
1661 		tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1662 	}
1663 
1664 	mark_page_lazyfree(page);
1665 	ret = true;
1666 out:
1667 	spin_unlock(ptl);
1668 out_unlocked:
1669 	return ret;
1670 }
1671 
1672 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1673 {
1674 	pgtable_t pgtable;
1675 
1676 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1677 	pte_free(mm, pgtable);
1678 	atomic_long_dec(&mm->nr_ptes);
1679 }
1680 
1681 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1682 		 pmd_t *pmd, unsigned long addr)
1683 {
1684 	pmd_t orig_pmd;
1685 	spinlock_t *ptl;
1686 
1687 	tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1688 
1689 	ptl = __pmd_trans_huge_lock(pmd, vma);
1690 	if (!ptl)
1691 		return 0;
1692 	/*
1693 	 * For architectures like ppc64 we look at deposited pgtable
1694 	 * when calling pmdp_huge_get_and_clear. So do the
1695 	 * pgtable_trans_huge_withdraw after finishing pmdp related
1696 	 * operations.
1697 	 */
1698 	orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1699 			tlb->fullmm);
1700 	tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1701 	if (vma_is_dax(vma)) {
1702 		if (arch_needs_pgtable_deposit())
1703 			zap_deposited_table(tlb->mm, pmd);
1704 		spin_unlock(ptl);
1705 		if (is_huge_zero_pmd(orig_pmd))
1706 			tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1707 	} else if (is_huge_zero_pmd(orig_pmd)) {
1708 		zap_deposited_table(tlb->mm, pmd);
1709 		spin_unlock(ptl);
1710 		tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1711 	} else {
1712 		struct page *page = NULL;
1713 		int flush_needed = 1;
1714 
1715 		if (pmd_present(orig_pmd)) {
1716 			page = pmd_page(orig_pmd);
1717 			page_remove_rmap(page, true);
1718 			VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1719 			VM_BUG_ON_PAGE(!PageHead(page), page);
1720 		} else if (thp_migration_supported()) {
1721 			swp_entry_t entry;
1722 
1723 			VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1724 			entry = pmd_to_swp_entry(orig_pmd);
1725 			page = pfn_to_page(swp_offset(entry));
1726 			flush_needed = 0;
1727 		} else
1728 			WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1729 
1730 		if (PageAnon(page)) {
1731 			zap_deposited_table(tlb->mm, pmd);
1732 			add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1733 		} else {
1734 			if (arch_needs_pgtable_deposit())
1735 				zap_deposited_table(tlb->mm, pmd);
1736 			add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1737 		}
1738 
1739 		spin_unlock(ptl);
1740 		if (flush_needed)
1741 			tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1742 	}
1743 	return 1;
1744 }
1745 
1746 #ifndef pmd_move_must_withdraw
1747 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1748 					 spinlock_t *old_pmd_ptl,
1749 					 struct vm_area_struct *vma)
1750 {
1751 	/*
1752 	 * With split pmd lock we also need to move preallocated
1753 	 * PTE page table if new_pmd is on different PMD page table.
1754 	 *
1755 	 * We also don't deposit and withdraw tables for file pages.
1756 	 */
1757 	return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1758 }
1759 #endif
1760 
1761 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1762 {
1763 #ifdef CONFIG_MEM_SOFT_DIRTY
1764 	if (unlikely(is_pmd_migration_entry(pmd)))
1765 		pmd = pmd_swp_mksoft_dirty(pmd);
1766 	else if (pmd_present(pmd))
1767 		pmd = pmd_mksoft_dirty(pmd);
1768 #endif
1769 	return pmd;
1770 }
1771 
1772 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1773 		  unsigned long new_addr, unsigned long old_end,
1774 		  pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
1775 {
1776 	spinlock_t *old_ptl, *new_ptl;
1777 	pmd_t pmd;
1778 	struct mm_struct *mm = vma->vm_mm;
1779 	bool force_flush = false;
1780 
1781 	if ((old_addr & ~HPAGE_PMD_MASK) ||
1782 	    (new_addr & ~HPAGE_PMD_MASK) ||
1783 	    old_end - old_addr < HPAGE_PMD_SIZE)
1784 		return false;
1785 
1786 	/*
1787 	 * The destination pmd shouldn't be established, free_pgtables()
1788 	 * should have release it.
1789 	 */
1790 	if (WARN_ON(!pmd_none(*new_pmd))) {
1791 		VM_BUG_ON(pmd_trans_huge(*new_pmd));
1792 		return false;
1793 	}
1794 
1795 	/*
1796 	 * We don't have to worry about the ordering of src and dst
1797 	 * ptlocks because exclusive mmap_sem prevents deadlock.
1798 	 */
1799 	old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1800 	if (old_ptl) {
1801 		new_ptl = pmd_lockptr(mm, new_pmd);
1802 		if (new_ptl != old_ptl)
1803 			spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1804 		pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1805 		if (pmd_present(pmd) && pmd_dirty(pmd))
1806 			force_flush = true;
1807 		VM_BUG_ON(!pmd_none(*new_pmd));
1808 
1809 		if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1810 			pgtable_t pgtable;
1811 			pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1812 			pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1813 		}
1814 		pmd = move_soft_dirty_pmd(pmd);
1815 		set_pmd_at(mm, new_addr, new_pmd, pmd);
1816 		if (new_ptl != old_ptl)
1817 			spin_unlock(new_ptl);
1818 		if (force_flush)
1819 			flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1820 		else
1821 			*need_flush = true;
1822 		spin_unlock(old_ptl);
1823 		return true;
1824 	}
1825 	return false;
1826 }
1827 
1828 /*
1829  * Returns
1830  *  - 0 if PMD could not be locked
1831  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1832  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1833  */
1834 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1835 		unsigned long addr, pgprot_t newprot, int prot_numa)
1836 {
1837 	struct mm_struct *mm = vma->vm_mm;
1838 	spinlock_t *ptl;
1839 	pmd_t entry;
1840 	bool preserve_write;
1841 	int ret;
1842 
1843 	ptl = __pmd_trans_huge_lock(pmd, vma);
1844 	if (!ptl)
1845 		return 0;
1846 
1847 	preserve_write = prot_numa && pmd_write(*pmd);
1848 	ret = 1;
1849 
1850 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1851 	if (is_swap_pmd(*pmd)) {
1852 		swp_entry_t entry = pmd_to_swp_entry(*pmd);
1853 
1854 		VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1855 		if (is_write_migration_entry(entry)) {
1856 			pmd_t newpmd;
1857 			/*
1858 			 * A protection check is difficult so
1859 			 * just be safe and disable write
1860 			 */
1861 			make_migration_entry_read(&entry);
1862 			newpmd = swp_entry_to_pmd(entry);
1863 			if (pmd_swp_soft_dirty(*pmd))
1864 				newpmd = pmd_swp_mksoft_dirty(newpmd);
1865 			set_pmd_at(mm, addr, pmd, newpmd);
1866 		}
1867 		goto unlock;
1868 	}
1869 #endif
1870 
1871 	/*
1872 	 * Avoid trapping faults against the zero page. The read-only
1873 	 * data is likely to be read-cached on the local CPU and
1874 	 * local/remote hits to the zero page are not interesting.
1875 	 */
1876 	if (prot_numa && is_huge_zero_pmd(*pmd))
1877 		goto unlock;
1878 
1879 	if (prot_numa && pmd_protnone(*pmd))
1880 		goto unlock;
1881 
1882 	/*
1883 	 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1884 	 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1885 	 * which is also under down_read(mmap_sem):
1886 	 *
1887 	 *	CPU0:				CPU1:
1888 	 *				change_huge_pmd(prot_numa=1)
1889 	 *				 pmdp_huge_get_and_clear_notify()
1890 	 * madvise_dontneed()
1891 	 *  zap_pmd_range()
1892 	 *   pmd_trans_huge(*pmd) == 0 (without ptl)
1893 	 *   // skip the pmd
1894 	 *				 set_pmd_at();
1895 	 *				 // pmd is re-established
1896 	 *
1897 	 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1898 	 * which may break userspace.
1899 	 *
1900 	 * pmdp_invalidate() is required to make sure we don't miss
1901 	 * dirty/young flags set by hardware.
1902 	 */
1903 	entry = *pmd;
1904 	pmdp_invalidate(vma, addr, pmd);
1905 
1906 	/*
1907 	 * Recover dirty/young flags.  It relies on pmdp_invalidate to not
1908 	 * corrupt them.
1909 	 */
1910 	if (pmd_dirty(*pmd))
1911 		entry = pmd_mkdirty(entry);
1912 	if (pmd_young(*pmd))
1913 		entry = pmd_mkyoung(entry);
1914 
1915 	entry = pmd_modify(entry, newprot);
1916 	if (preserve_write)
1917 		entry = pmd_mk_savedwrite(entry);
1918 	ret = HPAGE_PMD_NR;
1919 	set_pmd_at(mm, addr, pmd, entry);
1920 	BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1921 unlock:
1922 	spin_unlock(ptl);
1923 	return ret;
1924 }
1925 
1926 /*
1927  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1928  *
1929  * Note that if it returns page table lock pointer, this routine returns without
1930  * unlocking page table lock. So callers must unlock it.
1931  */
1932 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1933 {
1934 	spinlock_t *ptl;
1935 	ptl = pmd_lock(vma->vm_mm, pmd);
1936 	if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1937 			pmd_devmap(*pmd)))
1938 		return ptl;
1939 	spin_unlock(ptl);
1940 	return NULL;
1941 }
1942 
1943 /*
1944  * Returns true if a given pud maps a thp, false otherwise.
1945  *
1946  * Note that if it returns true, this routine returns without unlocking page
1947  * table lock. So callers must unlock it.
1948  */
1949 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1950 {
1951 	spinlock_t *ptl;
1952 
1953 	ptl = pud_lock(vma->vm_mm, pud);
1954 	if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1955 		return ptl;
1956 	spin_unlock(ptl);
1957 	return NULL;
1958 }
1959 
1960 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1961 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1962 		 pud_t *pud, unsigned long addr)
1963 {
1964 	pud_t orig_pud;
1965 	spinlock_t *ptl;
1966 
1967 	ptl = __pud_trans_huge_lock(pud, vma);
1968 	if (!ptl)
1969 		return 0;
1970 	/*
1971 	 * For architectures like ppc64 we look at deposited pgtable
1972 	 * when calling pudp_huge_get_and_clear. So do the
1973 	 * pgtable_trans_huge_withdraw after finishing pudp related
1974 	 * operations.
1975 	 */
1976 	orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud,
1977 			tlb->fullmm);
1978 	tlb_remove_pud_tlb_entry(tlb, pud, addr);
1979 	if (vma_is_dax(vma)) {
1980 		spin_unlock(ptl);
1981 		/* No zero page support yet */
1982 	} else {
1983 		/* No support for anonymous PUD pages yet */
1984 		BUG();
1985 	}
1986 	return 1;
1987 }
1988 
1989 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1990 		unsigned long haddr)
1991 {
1992 	VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1993 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1994 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1995 	VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1996 
1997 	count_vm_event(THP_SPLIT_PUD);
1998 
1999 	pudp_huge_clear_flush_notify(vma, haddr, pud);
2000 }
2001 
2002 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2003 		unsigned long address)
2004 {
2005 	spinlock_t *ptl;
2006 	struct mm_struct *mm = vma->vm_mm;
2007 	unsigned long haddr = address & HPAGE_PUD_MASK;
2008 
2009 	mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE);
2010 	ptl = pud_lock(mm, pud);
2011 	if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2012 		goto out;
2013 	__split_huge_pud_locked(vma, pud, haddr);
2014 
2015 out:
2016 	spin_unlock(ptl);
2017 	mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PUD_SIZE);
2018 }
2019 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2020 
2021 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2022 		unsigned long haddr, pmd_t *pmd)
2023 {
2024 	struct mm_struct *mm = vma->vm_mm;
2025 	pgtable_t pgtable;
2026 	pmd_t _pmd;
2027 	int i;
2028 
2029 	/* leave pmd empty until pte is filled */
2030 	pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2031 
2032 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2033 	pmd_populate(mm, &_pmd, pgtable);
2034 
2035 	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2036 		pte_t *pte, entry;
2037 		entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2038 		entry = pte_mkspecial(entry);
2039 		pte = pte_offset_map(&_pmd, haddr);
2040 		VM_BUG_ON(!pte_none(*pte));
2041 		set_pte_at(mm, haddr, pte, entry);
2042 		pte_unmap(pte);
2043 	}
2044 	smp_wmb(); /* make pte visible before pmd */
2045 	pmd_populate(mm, pmd, pgtable);
2046 }
2047 
2048 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2049 		unsigned long haddr, bool freeze)
2050 {
2051 	struct mm_struct *mm = vma->vm_mm;
2052 	struct page *page;
2053 	pgtable_t pgtable;
2054 	pmd_t _pmd;
2055 	bool young, write, dirty, soft_dirty, pmd_migration = false;
2056 	unsigned long addr;
2057 	int i;
2058 
2059 	VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2060 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2061 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2062 	VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2063 				&& !pmd_devmap(*pmd));
2064 
2065 	count_vm_event(THP_SPLIT_PMD);
2066 
2067 	if (!vma_is_anonymous(vma)) {
2068 		_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2069 		/*
2070 		 * We are going to unmap this huge page. So
2071 		 * just go ahead and zap it
2072 		 */
2073 		if (arch_needs_pgtable_deposit())
2074 			zap_deposited_table(mm, pmd);
2075 		if (vma_is_dax(vma))
2076 			return;
2077 		page = pmd_page(_pmd);
2078 		if (!PageReferenced(page) && pmd_young(_pmd))
2079 			SetPageReferenced(page);
2080 		page_remove_rmap(page, true);
2081 		put_page(page);
2082 		add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
2083 		return;
2084 	} else if (is_huge_zero_pmd(*pmd)) {
2085 		return __split_huge_zero_page_pmd(vma, haddr, pmd);
2086 	}
2087 
2088 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2089 	pmd_migration = is_pmd_migration_entry(*pmd);
2090 	if (pmd_migration) {
2091 		swp_entry_t entry;
2092 
2093 		entry = pmd_to_swp_entry(*pmd);
2094 		page = pfn_to_page(swp_offset(entry));
2095 	} else
2096 #endif
2097 		page = pmd_page(*pmd);
2098 	VM_BUG_ON_PAGE(!page_count(page), page);
2099 	page_ref_add(page, HPAGE_PMD_NR - 1);
2100 	write = pmd_write(*pmd);
2101 	young = pmd_young(*pmd);
2102 	dirty = pmd_dirty(*pmd);
2103 	soft_dirty = pmd_soft_dirty(*pmd);
2104 
2105 	pmdp_huge_split_prepare(vma, haddr, pmd);
2106 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2107 	pmd_populate(mm, &_pmd, pgtable);
2108 
2109 	for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2110 		pte_t entry, *pte;
2111 		/*
2112 		 * Note that NUMA hinting access restrictions are not
2113 		 * transferred to avoid any possibility of altering
2114 		 * permissions across VMAs.
2115 		 */
2116 		if (freeze || pmd_migration) {
2117 			swp_entry_t swp_entry;
2118 			swp_entry = make_migration_entry(page + i, write);
2119 			entry = swp_entry_to_pte(swp_entry);
2120 			if (soft_dirty)
2121 				entry = pte_swp_mksoft_dirty(entry);
2122 		} else {
2123 			entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2124 			entry = maybe_mkwrite(entry, vma);
2125 			if (!write)
2126 				entry = pte_wrprotect(entry);
2127 			if (!young)
2128 				entry = pte_mkold(entry);
2129 			if (soft_dirty)
2130 				entry = pte_mksoft_dirty(entry);
2131 		}
2132 		if (dirty)
2133 			SetPageDirty(page + i);
2134 		pte = pte_offset_map(&_pmd, addr);
2135 		BUG_ON(!pte_none(*pte));
2136 		set_pte_at(mm, addr, pte, entry);
2137 		atomic_inc(&page[i]._mapcount);
2138 		pte_unmap(pte);
2139 	}
2140 
2141 	/*
2142 	 * Set PG_double_map before dropping compound_mapcount to avoid
2143 	 * false-negative page_mapped().
2144 	 */
2145 	if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2146 		for (i = 0; i < HPAGE_PMD_NR; i++)
2147 			atomic_inc(&page[i]._mapcount);
2148 	}
2149 
2150 	if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2151 		/* Last compound_mapcount is gone. */
2152 		__dec_node_page_state(page, NR_ANON_THPS);
2153 		if (TestClearPageDoubleMap(page)) {
2154 			/* No need in mapcount reference anymore */
2155 			for (i = 0; i < HPAGE_PMD_NR; i++)
2156 				atomic_dec(&page[i]._mapcount);
2157 		}
2158 	}
2159 
2160 	smp_wmb(); /* make pte visible before pmd */
2161 	/*
2162 	 * Up to this point the pmd is present and huge and userland has the
2163 	 * whole access to the hugepage during the split (which happens in
2164 	 * place). If we overwrite the pmd with the not-huge version pointing
2165 	 * to the pte here (which of course we could if all CPUs were bug
2166 	 * free), userland could trigger a small page size TLB miss on the
2167 	 * small sized TLB while the hugepage TLB entry is still established in
2168 	 * the huge TLB. Some CPU doesn't like that.
2169 	 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2170 	 * 383 on page 93. Intel should be safe but is also warns that it's
2171 	 * only safe if the permission and cache attributes of the two entries
2172 	 * loaded in the two TLB is identical (which should be the case here).
2173 	 * But it is generally safer to never allow small and huge TLB entries
2174 	 * for the same virtual address to be loaded simultaneously. So instead
2175 	 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2176 	 * current pmd notpresent (atomically because here the pmd_trans_huge
2177 	 * and pmd_trans_splitting must remain set at all times on the pmd
2178 	 * until the split is complete for this pmd), then we flush the SMP TLB
2179 	 * and finally we write the non-huge version of the pmd entry with
2180 	 * pmd_populate.
2181 	 */
2182 	pmdp_invalidate(vma, haddr, pmd);
2183 	pmd_populate(mm, pmd, pgtable);
2184 
2185 	if (freeze) {
2186 		for (i = 0; i < HPAGE_PMD_NR; i++) {
2187 			page_remove_rmap(page + i, false);
2188 			put_page(page + i);
2189 		}
2190 	}
2191 }
2192 
2193 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2194 		unsigned long address, bool freeze, struct page *page)
2195 {
2196 	spinlock_t *ptl;
2197 	struct mm_struct *mm = vma->vm_mm;
2198 	unsigned long haddr = address & HPAGE_PMD_MASK;
2199 
2200 	mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2201 	ptl = pmd_lock(mm, pmd);
2202 
2203 	/*
2204 	 * If caller asks to setup a migration entries, we need a page to check
2205 	 * pmd against. Otherwise we can end up replacing wrong page.
2206 	 */
2207 	VM_BUG_ON(freeze && !page);
2208 	if (page && page != pmd_page(*pmd))
2209 	        goto out;
2210 
2211 	if (pmd_trans_huge(*pmd)) {
2212 		page = pmd_page(*pmd);
2213 		if (PageMlocked(page))
2214 			clear_page_mlock(page);
2215 	} else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2216 		goto out;
2217 	__split_huge_pmd_locked(vma, pmd, haddr, freeze);
2218 out:
2219 	spin_unlock(ptl);
2220 	mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
2221 }
2222 
2223 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2224 		bool freeze, struct page *page)
2225 {
2226 	pgd_t *pgd;
2227 	p4d_t *p4d;
2228 	pud_t *pud;
2229 	pmd_t *pmd;
2230 
2231 	pgd = pgd_offset(vma->vm_mm, address);
2232 	if (!pgd_present(*pgd))
2233 		return;
2234 
2235 	p4d = p4d_offset(pgd, address);
2236 	if (!p4d_present(*p4d))
2237 		return;
2238 
2239 	pud = pud_offset(p4d, address);
2240 	if (!pud_present(*pud))
2241 		return;
2242 
2243 	pmd = pmd_offset(pud, address);
2244 
2245 	__split_huge_pmd(vma, pmd, address, freeze, page);
2246 }
2247 
2248 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2249 			     unsigned long start,
2250 			     unsigned long end,
2251 			     long adjust_next)
2252 {
2253 	/*
2254 	 * If the new start address isn't hpage aligned and it could
2255 	 * previously contain an hugepage: check if we need to split
2256 	 * an huge pmd.
2257 	 */
2258 	if (start & ~HPAGE_PMD_MASK &&
2259 	    (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2260 	    (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2261 		split_huge_pmd_address(vma, start, false, NULL);
2262 
2263 	/*
2264 	 * If the new end address isn't hpage aligned and it could
2265 	 * previously contain an hugepage: check if we need to split
2266 	 * an huge pmd.
2267 	 */
2268 	if (end & ~HPAGE_PMD_MASK &&
2269 	    (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2270 	    (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2271 		split_huge_pmd_address(vma, end, false, NULL);
2272 
2273 	/*
2274 	 * If we're also updating the vma->vm_next->vm_start, if the new
2275 	 * vm_next->vm_start isn't page aligned and it could previously
2276 	 * contain an hugepage: check if we need to split an huge pmd.
2277 	 */
2278 	if (adjust_next > 0) {
2279 		struct vm_area_struct *next = vma->vm_next;
2280 		unsigned long nstart = next->vm_start;
2281 		nstart += adjust_next << PAGE_SHIFT;
2282 		if (nstart & ~HPAGE_PMD_MASK &&
2283 		    (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2284 		    (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2285 			split_huge_pmd_address(next, nstart, false, NULL);
2286 	}
2287 }
2288 
2289 static void freeze_page(struct page *page)
2290 {
2291 	enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2292 		TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2293 	bool unmap_success;
2294 
2295 	VM_BUG_ON_PAGE(!PageHead(page), page);
2296 
2297 	if (PageAnon(page))
2298 		ttu_flags |= TTU_SPLIT_FREEZE;
2299 
2300 	unmap_success = try_to_unmap(page, ttu_flags);
2301 	VM_BUG_ON_PAGE(!unmap_success, page);
2302 }
2303 
2304 static void unfreeze_page(struct page *page)
2305 {
2306 	int i;
2307 	if (PageTransHuge(page)) {
2308 		remove_migration_ptes(page, page, true);
2309 	} else {
2310 		for (i = 0; i < HPAGE_PMD_NR; i++)
2311 			remove_migration_ptes(page + i, page + i, true);
2312 	}
2313 }
2314 
2315 static void __split_huge_page_tail(struct page *head, int tail,
2316 		struct lruvec *lruvec, struct list_head *list)
2317 {
2318 	struct page *page_tail = head + tail;
2319 
2320 	VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2321 	VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
2322 
2323 	/*
2324 	 * tail_page->_refcount is zero and not changing from under us. But
2325 	 * get_page_unless_zero() may be running from under us on the
2326 	 * tail_page. If we used atomic_set() below instead of atomic_inc() or
2327 	 * atomic_add(), we would then run atomic_set() concurrently with
2328 	 * get_page_unless_zero(), and atomic_set() is implemented in C not
2329 	 * using locked ops. spin_unlock on x86 sometime uses locked ops
2330 	 * because of PPro errata 66, 92, so unless somebody can guarantee
2331 	 * atomic_set() here would be safe on all archs (and not only on x86),
2332 	 * it's safer to use atomic_inc()/atomic_add().
2333 	 */
2334 	if (PageAnon(head) && !PageSwapCache(head)) {
2335 		page_ref_inc(page_tail);
2336 	} else {
2337 		/* Additional pin to radix tree */
2338 		page_ref_add(page_tail, 2);
2339 	}
2340 
2341 	page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2342 	page_tail->flags |= (head->flags &
2343 			((1L << PG_referenced) |
2344 			 (1L << PG_swapbacked) |
2345 			 (1L << PG_swapcache) |
2346 			 (1L << PG_mlocked) |
2347 			 (1L << PG_uptodate) |
2348 			 (1L << PG_active) |
2349 			 (1L << PG_locked) |
2350 			 (1L << PG_unevictable) |
2351 			 (1L << PG_dirty)));
2352 
2353 	/*
2354 	 * After clearing PageTail the gup refcount can be released.
2355 	 * Page flags also must be visible before we make the page non-compound.
2356 	 */
2357 	smp_wmb();
2358 
2359 	clear_compound_head(page_tail);
2360 
2361 	if (page_is_young(head))
2362 		set_page_young(page_tail);
2363 	if (page_is_idle(head))
2364 		set_page_idle(page_tail);
2365 
2366 	/* ->mapping in first tail page is compound_mapcount */
2367 	VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2368 			page_tail);
2369 	page_tail->mapping = head->mapping;
2370 
2371 	page_tail->index = head->index + tail;
2372 	page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2373 	lru_add_page_tail(head, page_tail, lruvec, list);
2374 }
2375 
2376 static void __split_huge_page(struct page *page, struct list_head *list,
2377 		unsigned long flags)
2378 {
2379 	struct page *head = compound_head(page);
2380 	struct zone *zone = page_zone(head);
2381 	struct lruvec *lruvec;
2382 	pgoff_t end = -1;
2383 	int i;
2384 
2385 	lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
2386 
2387 	/* complete memcg works before add pages to LRU */
2388 	mem_cgroup_split_huge_fixup(head);
2389 
2390 	if (!PageAnon(page))
2391 		end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
2392 
2393 	for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2394 		__split_huge_page_tail(head, i, lruvec, list);
2395 		/* Some pages can be beyond i_size: drop them from page cache */
2396 		if (head[i].index >= end) {
2397 			__ClearPageDirty(head + i);
2398 			__delete_from_page_cache(head + i, NULL);
2399 			if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2400 				shmem_uncharge(head->mapping->host, 1);
2401 			put_page(head + i);
2402 		}
2403 	}
2404 
2405 	ClearPageCompound(head);
2406 	/* See comment in __split_huge_page_tail() */
2407 	if (PageAnon(head)) {
2408 		/* Additional pin to radix tree of swap cache */
2409 		if (PageSwapCache(head))
2410 			page_ref_add(head, 2);
2411 		else
2412 			page_ref_inc(head);
2413 	} else {
2414 		/* Additional pin to radix tree */
2415 		page_ref_add(head, 2);
2416 		spin_unlock(&head->mapping->tree_lock);
2417 	}
2418 
2419 	spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2420 
2421 	unfreeze_page(head);
2422 
2423 	for (i = 0; i < HPAGE_PMD_NR; i++) {
2424 		struct page *subpage = head + i;
2425 		if (subpage == page)
2426 			continue;
2427 		unlock_page(subpage);
2428 
2429 		/*
2430 		 * Subpages may be freed if there wasn't any mapping
2431 		 * like if add_to_swap() is running on a lru page that
2432 		 * had its mapping zapped. And freeing these pages
2433 		 * requires taking the lru_lock so we do the put_page
2434 		 * of the tail pages after the split is complete.
2435 		 */
2436 		put_page(subpage);
2437 	}
2438 }
2439 
2440 int total_mapcount(struct page *page)
2441 {
2442 	int i, compound, ret;
2443 
2444 	VM_BUG_ON_PAGE(PageTail(page), page);
2445 
2446 	if (likely(!PageCompound(page)))
2447 		return atomic_read(&page->_mapcount) + 1;
2448 
2449 	compound = compound_mapcount(page);
2450 	if (PageHuge(page))
2451 		return compound;
2452 	ret = compound;
2453 	for (i = 0; i < HPAGE_PMD_NR; i++)
2454 		ret += atomic_read(&page[i]._mapcount) + 1;
2455 	/* File pages has compound_mapcount included in _mapcount */
2456 	if (!PageAnon(page))
2457 		return ret - compound * HPAGE_PMD_NR;
2458 	if (PageDoubleMap(page))
2459 		ret -= HPAGE_PMD_NR;
2460 	return ret;
2461 }
2462 
2463 /*
2464  * This calculates accurately how many mappings a transparent hugepage
2465  * has (unlike page_mapcount() which isn't fully accurate). This full
2466  * accuracy is primarily needed to know if copy-on-write faults can
2467  * reuse the page and change the mapping to read-write instead of
2468  * copying them. At the same time this returns the total_mapcount too.
2469  *
2470  * The function returns the highest mapcount any one of the subpages
2471  * has. If the return value is one, even if different processes are
2472  * mapping different subpages of the transparent hugepage, they can
2473  * all reuse it, because each process is reusing a different subpage.
2474  *
2475  * The total_mapcount is instead counting all virtual mappings of the
2476  * subpages. If the total_mapcount is equal to "one", it tells the
2477  * caller all mappings belong to the same "mm" and in turn the
2478  * anon_vma of the transparent hugepage can become the vma->anon_vma
2479  * local one as no other process may be mapping any of the subpages.
2480  *
2481  * It would be more accurate to replace page_mapcount() with
2482  * page_trans_huge_mapcount(), however we only use
2483  * page_trans_huge_mapcount() in the copy-on-write faults where we
2484  * need full accuracy to avoid breaking page pinning, because
2485  * page_trans_huge_mapcount() is slower than page_mapcount().
2486  */
2487 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2488 {
2489 	int i, ret, _total_mapcount, mapcount;
2490 
2491 	/* hugetlbfs shouldn't call it */
2492 	VM_BUG_ON_PAGE(PageHuge(page), page);
2493 
2494 	if (likely(!PageTransCompound(page))) {
2495 		mapcount = atomic_read(&page->_mapcount) + 1;
2496 		if (total_mapcount)
2497 			*total_mapcount = mapcount;
2498 		return mapcount;
2499 	}
2500 
2501 	page = compound_head(page);
2502 
2503 	_total_mapcount = ret = 0;
2504 	for (i = 0; i < HPAGE_PMD_NR; i++) {
2505 		mapcount = atomic_read(&page[i]._mapcount) + 1;
2506 		ret = max(ret, mapcount);
2507 		_total_mapcount += mapcount;
2508 	}
2509 	if (PageDoubleMap(page)) {
2510 		ret -= 1;
2511 		_total_mapcount -= HPAGE_PMD_NR;
2512 	}
2513 	mapcount = compound_mapcount(page);
2514 	ret += mapcount;
2515 	_total_mapcount += mapcount;
2516 	if (total_mapcount)
2517 		*total_mapcount = _total_mapcount;
2518 	return ret;
2519 }
2520 
2521 /* Racy check whether the huge page can be split */
2522 bool can_split_huge_page(struct page *page, int *pextra_pins)
2523 {
2524 	int extra_pins;
2525 
2526 	/* Additional pins from radix tree */
2527 	if (PageAnon(page))
2528 		extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2529 	else
2530 		extra_pins = HPAGE_PMD_NR;
2531 	if (pextra_pins)
2532 		*pextra_pins = extra_pins;
2533 	return total_mapcount(page) == page_count(page) - extra_pins - 1;
2534 }
2535 
2536 /*
2537  * This function splits huge page into normal pages. @page can point to any
2538  * subpage of huge page to split. Split doesn't change the position of @page.
2539  *
2540  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2541  * The huge page must be locked.
2542  *
2543  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2544  *
2545  * Both head page and tail pages will inherit mapping, flags, and so on from
2546  * the hugepage.
2547  *
2548  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2549  * they are not mapped.
2550  *
2551  * Returns 0 if the hugepage is split successfully.
2552  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2553  * us.
2554  */
2555 int split_huge_page_to_list(struct page *page, struct list_head *list)
2556 {
2557 	struct page *head = compound_head(page);
2558 	struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2559 	struct anon_vma *anon_vma = NULL;
2560 	struct address_space *mapping = NULL;
2561 	int count, mapcount, extra_pins, ret;
2562 	bool mlocked;
2563 	unsigned long flags;
2564 
2565 	VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2566 	VM_BUG_ON_PAGE(!PageLocked(page), page);
2567 	VM_BUG_ON_PAGE(!PageCompound(page), page);
2568 
2569 	if (PageWriteback(page))
2570 		return -EBUSY;
2571 
2572 	if (PageAnon(head)) {
2573 		/*
2574 		 * The caller does not necessarily hold an mmap_sem that would
2575 		 * prevent the anon_vma disappearing so we first we take a
2576 		 * reference to it and then lock the anon_vma for write. This
2577 		 * is similar to page_lock_anon_vma_read except the write lock
2578 		 * is taken to serialise against parallel split or collapse
2579 		 * operations.
2580 		 */
2581 		anon_vma = page_get_anon_vma(head);
2582 		if (!anon_vma) {
2583 			ret = -EBUSY;
2584 			goto out;
2585 		}
2586 		mapping = NULL;
2587 		anon_vma_lock_write(anon_vma);
2588 	} else {
2589 		mapping = head->mapping;
2590 
2591 		/* Truncated ? */
2592 		if (!mapping) {
2593 			ret = -EBUSY;
2594 			goto out;
2595 		}
2596 
2597 		anon_vma = NULL;
2598 		i_mmap_lock_read(mapping);
2599 	}
2600 
2601 	/*
2602 	 * Racy check if we can split the page, before freeze_page() will
2603 	 * split PMDs
2604 	 */
2605 	if (!can_split_huge_page(head, &extra_pins)) {
2606 		ret = -EBUSY;
2607 		goto out_unlock;
2608 	}
2609 
2610 	mlocked = PageMlocked(page);
2611 	freeze_page(head);
2612 	VM_BUG_ON_PAGE(compound_mapcount(head), head);
2613 
2614 	/* Make sure the page is not on per-CPU pagevec as it takes pin */
2615 	if (mlocked)
2616 		lru_add_drain();
2617 
2618 	/* prevent PageLRU to go away from under us, and freeze lru stats */
2619 	spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2620 
2621 	if (mapping) {
2622 		void **pslot;
2623 
2624 		spin_lock(&mapping->tree_lock);
2625 		pslot = radix_tree_lookup_slot(&mapping->page_tree,
2626 				page_index(head));
2627 		/*
2628 		 * Check if the head page is present in radix tree.
2629 		 * We assume all tail are present too, if head is there.
2630 		 */
2631 		if (radix_tree_deref_slot_protected(pslot,
2632 					&mapping->tree_lock) != head)
2633 			goto fail;
2634 	}
2635 
2636 	/* Prevent deferred_split_scan() touching ->_refcount */
2637 	spin_lock(&pgdata->split_queue_lock);
2638 	count = page_count(head);
2639 	mapcount = total_mapcount(head);
2640 	if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2641 		if (!list_empty(page_deferred_list(head))) {
2642 			pgdata->split_queue_len--;
2643 			list_del(page_deferred_list(head));
2644 		}
2645 		if (mapping)
2646 			__dec_node_page_state(page, NR_SHMEM_THPS);
2647 		spin_unlock(&pgdata->split_queue_lock);
2648 		__split_huge_page(page, list, flags);
2649 		if (PageSwapCache(head)) {
2650 			swp_entry_t entry = { .val = page_private(head) };
2651 
2652 			ret = split_swap_cluster(entry);
2653 		} else
2654 			ret = 0;
2655 	} else {
2656 		if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2657 			pr_alert("total_mapcount: %u, page_count(): %u\n",
2658 					mapcount, count);
2659 			if (PageTail(page))
2660 				dump_page(head, NULL);
2661 			dump_page(page, "total_mapcount(head) > 0");
2662 			BUG();
2663 		}
2664 		spin_unlock(&pgdata->split_queue_lock);
2665 fail:		if (mapping)
2666 			spin_unlock(&mapping->tree_lock);
2667 		spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2668 		unfreeze_page(head);
2669 		ret = -EBUSY;
2670 	}
2671 
2672 out_unlock:
2673 	if (anon_vma) {
2674 		anon_vma_unlock_write(anon_vma);
2675 		put_anon_vma(anon_vma);
2676 	}
2677 	if (mapping)
2678 		i_mmap_unlock_read(mapping);
2679 out:
2680 	count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2681 	return ret;
2682 }
2683 
2684 void free_transhuge_page(struct page *page)
2685 {
2686 	struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2687 	unsigned long flags;
2688 
2689 	spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2690 	if (!list_empty(page_deferred_list(page))) {
2691 		pgdata->split_queue_len--;
2692 		list_del(page_deferred_list(page));
2693 	}
2694 	spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2695 	free_compound_page(page);
2696 }
2697 
2698 void deferred_split_huge_page(struct page *page)
2699 {
2700 	struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2701 	unsigned long flags;
2702 
2703 	VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2704 
2705 	spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2706 	if (list_empty(page_deferred_list(page))) {
2707 		count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2708 		list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2709 		pgdata->split_queue_len++;
2710 	}
2711 	spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2712 }
2713 
2714 static unsigned long deferred_split_count(struct shrinker *shrink,
2715 		struct shrink_control *sc)
2716 {
2717 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2718 	return ACCESS_ONCE(pgdata->split_queue_len);
2719 }
2720 
2721 static unsigned long deferred_split_scan(struct shrinker *shrink,
2722 		struct shrink_control *sc)
2723 {
2724 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2725 	unsigned long flags;
2726 	LIST_HEAD(list), *pos, *next;
2727 	struct page *page;
2728 	int split = 0;
2729 
2730 	spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2731 	/* Take pin on all head pages to avoid freeing them under us */
2732 	list_for_each_safe(pos, next, &pgdata->split_queue) {
2733 		page = list_entry((void *)pos, struct page, mapping);
2734 		page = compound_head(page);
2735 		if (get_page_unless_zero(page)) {
2736 			list_move(page_deferred_list(page), &list);
2737 		} else {
2738 			/* We lost race with put_compound_page() */
2739 			list_del_init(page_deferred_list(page));
2740 			pgdata->split_queue_len--;
2741 		}
2742 		if (!--sc->nr_to_scan)
2743 			break;
2744 	}
2745 	spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2746 
2747 	list_for_each_safe(pos, next, &list) {
2748 		page = list_entry((void *)pos, struct page, mapping);
2749 		lock_page(page);
2750 		/* split_huge_page() removes page from list on success */
2751 		if (!split_huge_page(page))
2752 			split++;
2753 		unlock_page(page);
2754 		put_page(page);
2755 	}
2756 
2757 	spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2758 	list_splice_tail(&list, &pgdata->split_queue);
2759 	spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2760 
2761 	/*
2762 	 * Stop shrinker if we didn't split any page, but the queue is empty.
2763 	 * This can happen if pages were freed under us.
2764 	 */
2765 	if (!split && list_empty(&pgdata->split_queue))
2766 		return SHRINK_STOP;
2767 	return split;
2768 }
2769 
2770 static struct shrinker deferred_split_shrinker = {
2771 	.count_objects = deferred_split_count,
2772 	.scan_objects = deferred_split_scan,
2773 	.seeks = DEFAULT_SEEKS,
2774 	.flags = SHRINKER_NUMA_AWARE,
2775 };
2776 
2777 #ifdef CONFIG_DEBUG_FS
2778 static int split_huge_pages_set(void *data, u64 val)
2779 {
2780 	struct zone *zone;
2781 	struct page *page;
2782 	unsigned long pfn, max_zone_pfn;
2783 	unsigned long total = 0, split = 0;
2784 
2785 	if (val != 1)
2786 		return -EINVAL;
2787 
2788 	for_each_populated_zone(zone) {
2789 		max_zone_pfn = zone_end_pfn(zone);
2790 		for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2791 			if (!pfn_valid(pfn))
2792 				continue;
2793 
2794 			page = pfn_to_page(pfn);
2795 			if (!get_page_unless_zero(page))
2796 				continue;
2797 
2798 			if (zone != page_zone(page))
2799 				goto next;
2800 
2801 			if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2802 				goto next;
2803 
2804 			total++;
2805 			lock_page(page);
2806 			if (!split_huge_page(page))
2807 				split++;
2808 			unlock_page(page);
2809 next:
2810 			put_page(page);
2811 		}
2812 	}
2813 
2814 	pr_info("%lu of %lu THP split\n", split, total);
2815 
2816 	return 0;
2817 }
2818 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2819 		"%llu\n");
2820 
2821 static int __init split_huge_pages_debugfs(void)
2822 {
2823 	void *ret;
2824 
2825 	ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2826 			&split_huge_pages_fops);
2827 	if (!ret)
2828 		pr_warn("Failed to create split_huge_pages in debugfs");
2829 	return 0;
2830 }
2831 late_initcall(split_huge_pages_debugfs);
2832 #endif
2833 
2834 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2835 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2836 		struct page *page)
2837 {
2838 	struct vm_area_struct *vma = pvmw->vma;
2839 	struct mm_struct *mm = vma->vm_mm;
2840 	unsigned long address = pvmw->address;
2841 	pmd_t pmdval;
2842 	swp_entry_t entry;
2843 	pmd_t pmdswp;
2844 
2845 	if (!(pvmw->pmd && !pvmw->pte))
2846 		return;
2847 
2848 	mmu_notifier_invalidate_range_start(mm, address,
2849 			address + HPAGE_PMD_SIZE);
2850 
2851 	flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2852 	pmdval = *pvmw->pmd;
2853 	pmdp_invalidate(vma, address, pvmw->pmd);
2854 	if (pmd_dirty(pmdval))
2855 		set_page_dirty(page);
2856 	entry = make_migration_entry(page, pmd_write(pmdval));
2857 	pmdswp = swp_entry_to_pmd(entry);
2858 	if (pmd_soft_dirty(pmdval))
2859 		pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2860 	set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2861 	page_remove_rmap(page, true);
2862 	put_page(page);
2863 
2864 	mmu_notifier_invalidate_range_end(mm, address,
2865 			address + HPAGE_PMD_SIZE);
2866 }
2867 
2868 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2869 {
2870 	struct vm_area_struct *vma = pvmw->vma;
2871 	struct mm_struct *mm = vma->vm_mm;
2872 	unsigned long address = pvmw->address;
2873 	unsigned long mmun_start = address & HPAGE_PMD_MASK;
2874 	pmd_t pmde;
2875 	swp_entry_t entry;
2876 
2877 	if (!(pvmw->pmd && !pvmw->pte))
2878 		return;
2879 
2880 	entry = pmd_to_swp_entry(*pvmw->pmd);
2881 	get_page(new);
2882 	pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2883 	if (pmd_swp_soft_dirty(*pvmw->pmd))
2884 		pmde = pmd_mksoft_dirty(pmde);
2885 	if (is_write_migration_entry(entry))
2886 		pmde = maybe_pmd_mkwrite(pmde, vma);
2887 
2888 	flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2889 	page_add_anon_rmap(new, vma, mmun_start, true);
2890 	set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2891 	if (vma->vm_flags & VM_LOCKED)
2892 		mlock_vma_page(new);
2893 	update_mmu_cache_pmd(vma, address, pvmw->pmd);
2894 }
2895 #endif
2896