xref: /openbmc/linux/mm/huge_memory.c (revision 4e95bc26)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  Copyright (C) 2009  Red Hat, Inc.
4  */
5 
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7 
8 #include <linux/mm.h>
9 #include <linux/sched.h>
10 #include <linux/sched/coredump.h>
11 #include <linux/sched/numa_balancing.h>
12 #include <linux/highmem.h>
13 #include <linux/hugetlb.h>
14 #include <linux/mmu_notifier.h>
15 #include <linux/rmap.h>
16 #include <linux/swap.h>
17 #include <linux/shrinker.h>
18 #include <linux/mm_inline.h>
19 #include <linux/swapops.h>
20 #include <linux/dax.h>
21 #include <linux/khugepaged.h>
22 #include <linux/freezer.h>
23 #include <linux/pfn_t.h>
24 #include <linux/mman.h>
25 #include <linux/memremap.h>
26 #include <linux/pagemap.h>
27 #include <linux/debugfs.h>
28 #include <linux/migrate.h>
29 #include <linux/hashtable.h>
30 #include <linux/userfaultfd_k.h>
31 #include <linux/page_idle.h>
32 #include <linux/shmem_fs.h>
33 #include <linux/oom.h>
34 #include <linux/numa.h>
35 
36 #include <asm/tlb.h>
37 #include <asm/pgalloc.h>
38 #include "internal.h"
39 
40 /*
41  * By default, transparent hugepage support is disabled in order to avoid
42  * risking an increased memory footprint for applications that are not
43  * guaranteed to benefit from it. When transparent hugepage support is
44  * enabled, it is for all mappings, and khugepaged scans all mappings.
45  * Defrag is invoked by khugepaged hugepage allocations and by page faults
46  * for all hugepage allocations.
47  */
48 unsigned long transparent_hugepage_flags __read_mostly =
49 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
50 	(1<<TRANSPARENT_HUGEPAGE_FLAG)|
51 #endif
52 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
53 	(1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
54 #endif
55 	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
56 	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
57 	(1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
58 
59 static struct shrinker deferred_split_shrinker;
60 
61 static atomic_t huge_zero_refcount;
62 struct page *huge_zero_page __read_mostly;
63 
64 bool transparent_hugepage_enabled(struct vm_area_struct *vma)
65 {
66 	if (vma_is_anonymous(vma))
67 		return __transparent_hugepage_enabled(vma);
68 	if (vma_is_shmem(vma) && shmem_huge_enabled(vma))
69 		return __transparent_hugepage_enabled(vma);
70 
71 	return false;
72 }
73 
74 static struct page *get_huge_zero_page(void)
75 {
76 	struct page *zero_page;
77 retry:
78 	if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
79 		return READ_ONCE(huge_zero_page);
80 
81 	zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
82 			HPAGE_PMD_ORDER);
83 	if (!zero_page) {
84 		count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
85 		return NULL;
86 	}
87 	count_vm_event(THP_ZERO_PAGE_ALLOC);
88 	preempt_disable();
89 	if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
90 		preempt_enable();
91 		__free_pages(zero_page, compound_order(zero_page));
92 		goto retry;
93 	}
94 
95 	/* We take additional reference here. It will be put back by shrinker */
96 	atomic_set(&huge_zero_refcount, 2);
97 	preempt_enable();
98 	return READ_ONCE(huge_zero_page);
99 }
100 
101 static void put_huge_zero_page(void)
102 {
103 	/*
104 	 * Counter should never go to zero here. Only shrinker can put
105 	 * last reference.
106 	 */
107 	BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
108 }
109 
110 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
111 {
112 	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
113 		return READ_ONCE(huge_zero_page);
114 
115 	if (!get_huge_zero_page())
116 		return NULL;
117 
118 	if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
119 		put_huge_zero_page();
120 
121 	return READ_ONCE(huge_zero_page);
122 }
123 
124 void mm_put_huge_zero_page(struct mm_struct *mm)
125 {
126 	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
127 		put_huge_zero_page();
128 }
129 
130 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
131 					struct shrink_control *sc)
132 {
133 	/* we can free zero page only if last reference remains */
134 	return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
135 }
136 
137 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
138 				       struct shrink_control *sc)
139 {
140 	if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
141 		struct page *zero_page = xchg(&huge_zero_page, NULL);
142 		BUG_ON(zero_page == NULL);
143 		__free_pages(zero_page, compound_order(zero_page));
144 		return HPAGE_PMD_NR;
145 	}
146 
147 	return 0;
148 }
149 
150 static struct shrinker huge_zero_page_shrinker = {
151 	.count_objects = shrink_huge_zero_page_count,
152 	.scan_objects = shrink_huge_zero_page_scan,
153 	.seeks = DEFAULT_SEEKS,
154 };
155 
156 #ifdef CONFIG_SYSFS
157 static ssize_t enabled_show(struct kobject *kobj,
158 			    struct kobj_attribute *attr, char *buf)
159 {
160 	if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
161 		return sprintf(buf, "[always] madvise never\n");
162 	else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
163 		return sprintf(buf, "always [madvise] never\n");
164 	else
165 		return sprintf(buf, "always madvise [never]\n");
166 }
167 
168 static ssize_t enabled_store(struct kobject *kobj,
169 			     struct kobj_attribute *attr,
170 			     const char *buf, size_t count)
171 {
172 	ssize_t ret = count;
173 
174 	if (!memcmp("always", buf,
175 		    min(sizeof("always")-1, count))) {
176 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
177 		set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
178 	} else if (!memcmp("madvise", buf,
179 			   min(sizeof("madvise")-1, count))) {
180 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
181 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
182 	} else if (!memcmp("never", buf,
183 			   min(sizeof("never")-1, count))) {
184 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
185 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
186 	} else
187 		ret = -EINVAL;
188 
189 	if (ret > 0) {
190 		int err = start_stop_khugepaged();
191 		if (err)
192 			ret = err;
193 	}
194 	return ret;
195 }
196 static struct kobj_attribute enabled_attr =
197 	__ATTR(enabled, 0644, enabled_show, enabled_store);
198 
199 ssize_t single_hugepage_flag_show(struct kobject *kobj,
200 				struct kobj_attribute *attr, char *buf,
201 				enum transparent_hugepage_flag flag)
202 {
203 	return sprintf(buf, "%d\n",
204 		       !!test_bit(flag, &transparent_hugepage_flags));
205 }
206 
207 ssize_t single_hugepage_flag_store(struct kobject *kobj,
208 				 struct kobj_attribute *attr,
209 				 const char *buf, size_t count,
210 				 enum transparent_hugepage_flag flag)
211 {
212 	unsigned long value;
213 	int ret;
214 
215 	ret = kstrtoul(buf, 10, &value);
216 	if (ret < 0)
217 		return ret;
218 	if (value > 1)
219 		return -EINVAL;
220 
221 	if (value)
222 		set_bit(flag, &transparent_hugepage_flags);
223 	else
224 		clear_bit(flag, &transparent_hugepage_flags);
225 
226 	return count;
227 }
228 
229 static ssize_t defrag_show(struct kobject *kobj,
230 			   struct kobj_attribute *attr, char *buf)
231 {
232 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
233 		return sprintf(buf, "[always] defer defer+madvise madvise never\n");
234 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
235 		return sprintf(buf, "always [defer] defer+madvise madvise never\n");
236 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
237 		return sprintf(buf, "always defer [defer+madvise] madvise never\n");
238 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
239 		return sprintf(buf, "always defer defer+madvise [madvise] never\n");
240 	return sprintf(buf, "always defer defer+madvise madvise [never]\n");
241 }
242 
243 static ssize_t defrag_store(struct kobject *kobj,
244 			    struct kobj_attribute *attr,
245 			    const char *buf, size_t count)
246 {
247 	if (!memcmp("always", buf,
248 		    min(sizeof("always")-1, count))) {
249 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
250 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
251 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
252 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
253 	} else if (!memcmp("defer+madvise", buf,
254 		    min(sizeof("defer+madvise")-1, count))) {
255 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
256 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
257 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
258 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
259 	} else if (!memcmp("defer", buf,
260 		    min(sizeof("defer")-1, count))) {
261 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
262 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
263 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
264 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
265 	} else if (!memcmp("madvise", buf,
266 			   min(sizeof("madvise")-1, count))) {
267 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
268 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
269 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
270 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
271 	} else if (!memcmp("never", buf,
272 			   min(sizeof("never")-1, count))) {
273 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
274 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
275 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
276 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
277 	} else
278 		return -EINVAL;
279 
280 	return count;
281 }
282 static struct kobj_attribute defrag_attr =
283 	__ATTR(defrag, 0644, defrag_show, defrag_store);
284 
285 static ssize_t use_zero_page_show(struct kobject *kobj,
286 		struct kobj_attribute *attr, char *buf)
287 {
288 	return single_hugepage_flag_show(kobj, attr, buf,
289 				TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
290 }
291 static ssize_t use_zero_page_store(struct kobject *kobj,
292 		struct kobj_attribute *attr, const char *buf, size_t count)
293 {
294 	return single_hugepage_flag_store(kobj, attr, buf, count,
295 				 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
296 }
297 static struct kobj_attribute use_zero_page_attr =
298 	__ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
299 
300 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
301 		struct kobj_attribute *attr, char *buf)
302 {
303 	return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
304 }
305 static struct kobj_attribute hpage_pmd_size_attr =
306 	__ATTR_RO(hpage_pmd_size);
307 
308 #ifdef CONFIG_DEBUG_VM
309 static ssize_t debug_cow_show(struct kobject *kobj,
310 				struct kobj_attribute *attr, char *buf)
311 {
312 	return single_hugepage_flag_show(kobj, attr, buf,
313 				TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
314 }
315 static ssize_t debug_cow_store(struct kobject *kobj,
316 			       struct kobj_attribute *attr,
317 			       const char *buf, size_t count)
318 {
319 	return single_hugepage_flag_store(kobj, attr, buf, count,
320 				 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
321 }
322 static struct kobj_attribute debug_cow_attr =
323 	__ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
324 #endif /* CONFIG_DEBUG_VM */
325 
326 static struct attribute *hugepage_attr[] = {
327 	&enabled_attr.attr,
328 	&defrag_attr.attr,
329 	&use_zero_page_attr.attr,
330 	&hpage_pmd_size_attr.attr,
331 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
332 	&shmem_enabled_attr.attr,
333 #endif
334 #ifdef CONFIG_DEBUG_VM
335 	&debug_cow_attr.attr,
336 #endif
337 	NULL,
338 };
339 
340 static const struct attribute_group hugepage_attr_group = {
341 	.attrs = hugepage_attr,
342 };
343 
344 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
345 {
346 	int err;
347 
348 	*hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
349 	if (unlikely(!*hugepage_kobj)) {
350 		pr_err("failed to create transparent hugepage kobject\n");
351 		return -ENOMEM;
352 	}
353 
354 	err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
355 	if (err) {
356 		pr_err("failed to register transparent hugepage group\n");
357 		goto delete_obj;
358 	}
359 
360 	err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
361 	if (err) {
362 		pr_err("failed to register transparent hugepage group\n");
363 		goto remove_hp_group;
364 	}
365 
366 	return 0;
367 
368 remove_hp_group:
369 	sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
370 delete_obj:
371 	kobject_put(*hugepage_kobj);
372 	return err;
373 }
374 
375 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
376 {
377 	sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
378 	sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
379 	kobject_put(hugepage_kobj);
380 }
381 #else
382 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
383 {
384 	return 0;
385 }
386 
387 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
388 {
389 }
390 #endif /* CONFIG_SYSFS */
391 
392 static int __init hugepage_init(void)
393 {
394 	int err;
395 	struct kobject *hugepage_kobj;
396 
397 	if (!has_transparent_hugepage()) {
398 		transparent_hugepage_flags = 0;
399 		return -EINVAL;
400 	}
401 
402 	/*
403 	 * hugepages can't be allocated by the buddy allocator
404 	 */
405 	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
406 	/*
407 	 * we use page->mapping and page->index in second tail page
408 	 * as list_head: assuming THP order >= 2
409 	 */
410 	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
411 
412 	err = hugepage_init_sysfs(&hugepage_kobj);
413 	if (err)
414 		goto err_sysfs;
415 
416 	err = khugepaged_init();
417 	if (err)
418 		goto err_slab;
419 
420 	err = register_shrinker(&huge_zero_page_shrinker);
421 	if (err)
422 		goto err_hzp_shrinker;
423 	err = register_shrinker(&deferred_split_shrinker);
424 	if (err)
425 		goto err_split_shrinker;
426 
427 	/*
428 	 * By default disable transparent hugepages on smaller systems,
429 	 * where the extra memory used could hurt more than TLB overhead
430 	 * is likely to save.  The admin can still enable it through /sys.
431 	 */
432 	if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
433 		transparent_hugepage_flags = 0;
434 		return 0;
435 	}
436 
437 	err = start_stop_khugepaged();
438 	if (err)
439 		goto err_khugepaged;
440 
441 	return 0;
442 err_khugepaged:
443 	unregister_shrinker(&deferred_split_shrinker);
444 err_split_shrinker:
445 	unregister_shrinker(&huge_zero_page_shrinker);
446 err_hzp_shrinker:
447 	khugepaged_destroy();
448 err_slab:
449 	hugepage_exit_sysfs(hugepage_kobj);
450 err_sysfs:
451 	return err;
452 }
453 subsys_initcall(hugepage_init);
454 
455 static int __init setup_transparent_hugepage(char *str)
456 {
457 	int ret = 0;
458 	if (!str)
459 		goto out;
460 	if (!strcmp(str, "always")) {
461 		set_bit(TRANSPARENT_HUGEPAGE_FLAG,
462 			&transparent_hugepage_flags);
463 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
464 			  &transparent_hugepage_flags);
465 		ret = 1;
466 	} else if (!strcmp(str, "madvise")) {
467 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
468 			  &transparent_hugepage_flags);
469 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
470 			&transparent_hugepage_flags);
471 		ret = 1;
472 	} else if (!strcmp(str, "never")) {
473 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
474 			  &transparent_hugepage_flags);
475 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
476 			  &transparent_hugepage_flags);
477 		ret = 1;
478 	}
479 out:
480 	if (!ret)
481 		pr_warn("transparent_hugepage= cannot parse, ignored\n");
482 	return ret;
483 }
484 __setup("transparent_hugepage=", setup_transparent_hugepage);
485 
486 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
487 {
488 	if (likely(vma->vm_flags & VM_WRITE))
489 		pmd = pmd_mkwrite(pmd);
490 	return pmd;
491 }
492 
493 static inline struct list_head *page_deferred_list(struct page *page)
494 {
495 	/* ->lru in the tail pages is occupied by compound_head. */
496 	return &page[2].deferred_list;
497 }
498 
499 void prep_transhuge_page(struct page *page)
500 {
501 	/*
502 	 * we use page->mapping and page->indexlru in second tail page
503 	 * as list_head: assuming THP order >= 2
504 	 */
505 
506 	INIT_LIST_HEAD(page_deferred_list(page));
507 	set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
508 }
509 
510 static unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
511 		loff_t off, unsigned long flags, unsigned long size)
512 {
513 	unsigned long addr;
514 	loff_t off_end = off + len;
515 	loff_t off_align = round_up(off, size);
516 	unsigned long len_pad;
517 
518 	if (off_end <= off_align || (off_end - off_align) < size)
519 		return 0;
520 
521 	len_pad = len + size;
522 	if (len_pad < len || (off + len_pad) < off)
523 		return 0;
524 
525 	addr = current->mm->get_unmapped_area(filp, 0, len_pad,
526 					      off >> PAGE_SHIFT, flags);
527 	if (IS_ERR_VALUE(addr))
528 		return 0;
529 
530 	addr += (off - addr) & (size - 1);
531 	return addr;
532 }
533 
534 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
535 		unsigned long len, unsigned long pgoff, unsigned long flags)
536 {
537 	loff_t off = (loff_t)pgoff << PAGE_SHIFT;
538 
539 	if (addr)
540 		goto out;
541 	if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
542 		goto out;
543 
544 	addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
545 	if (addr)
546 		return addr;
547 
548  out:
549 	return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
550 }
551 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
552 
553 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
554 			struct page *page, gfp_t gfp)
555 {
556 	struct vm_area_struct *vma = vmf->vma;
557 	struct mem_cgroup *memcg;
558 	pgtable_t pgtable;
559 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
560 	vm_fault_t ret = 0;
561 
562 	VM_BUG_ON_PAGE(!PageCompound(page), page);
563 
564 	if (mem_cgroup_try_charge_delay(page, vma->vm_mm, gfp, &memcg, true)) {
565 		put_page(page);
566 		count_vm_event(THP_FAULT_FALLBACK);
567 		return VM_FAULT_FALLBACK;
568 	}
569 
570 	pgtable = pte_alloc_one(vma->vm_mm);
571 	if (unlikely(!pgtable)) {
572 		ret = VM_FAULT_OOM;
573 		goto release;
574 	}
575 
576 	clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
577 	/*
578 	 * The memory barrier inside __SetPageUptodate makes sure that
579 	 * clear_huge_page writes become visible before the set_pmd_at()
580 	 * write.
581 	 */
582 	__SetPageUptodate(page);
583 
584 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
585 	if (unlikely(!pmd_none(*vmf->pmd))) {
586 		goto unlock_release;
587 	} else {
588 		pmd_t entry;
589 
590 		ret = check_stable_address_space(vma->vm_mm);
591 		if (ret)
592 			goto unlock_release;
593 
594 		/* Deliver the page fault to userland */
595 		if (userfaultfd_missing(vma)) {
596 			vm_fault_t ret2;
597 
598 			spin_unlock(vmf->ptl);
599 			mem_cgroup_cancel_charge(page, memcg, true);
600 			put_page(page);
601 			pte_free(vma->vm_mm, pgtable);
602 			ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
603 			VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
604 			return ret2;
605 		}
606 
607 		entry = mk_huge_pmd(page, vma->vm_page_prot);
608 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
609 		page_add_new_anon_rmap(page, vma, haddr, true);
610 		mem_cgroup_commit_charge(page, memcg, false, true);
611 		lru_cache_add_active_or_unevictable(page, vma);
612 		pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
613 		set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
614 		add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
615 		mm_inc_nr_ptes(vma->vm_mm);
616 		spin_unlock(vmf->ptl);
617 		count_vm_event(THP_FAULT_ALLOC);
618 		count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
619 	}
620 
621 	return 0;
622 unlock_release:
623 	spin_unlock(vmf->ptl);
624 release:
625 	if (pgtable)
626 		pte_free(vma->vm_mm, pgtable);
627 	mem_cgroup_cancel_charge(page, memcg, true);
628 	put_page(page);
629 	return ret;
630 
631 }
632 
633 /*
634  * always: directly stall for all thp allocations
635  * defer: wake kswapd and fail if not immediately available
636  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
637  *		  fail if not immediately available
638  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
639  *	    available
640  * never: never stall for any thp allocation
641  */
642 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
643 {
644 	const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
645 
646 	/* Always do synchronous compaction */
647 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
648 		return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
649 
650 	/* Kick kcompactd and fail quickly */
651 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
652 		return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
653 
654 	/* Synchronous compaction if madvised, otherwise kick kcompactd */
655 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
656 		return GFP_TRANSHUGE_LIGHT |
657 			(vma_madvised ? __GFP_DIRECT_RECLAIM :
658 					__GFP_KSWAPD_RECLAIM);
659 
660 	/* Only do synchronous compaction if madvised */
661 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
662 		return GFP_TRANSHUGE_LIGHT |
663 		       (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
664 
665 	return GFP_TRANSHUGE_LIGHT;
666 }
667 
668 /* Caller must hold page table lock. */
669 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
670 		struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
671 		struct page *zero_page)
672 {
673 	pmd_t entry;
674 	if (!pmd_none(*pmd))
675 		return false;
676 	entry = mk_pmd(zero_page, vma->vm_page_prot);
677 	entry = pmd_mkhuge(entry);
678 	if (pgtable)
679 		pgtable_trans_huge_deposit(mm, pmd, pgtable);
680 	set_pmd_at(mm, haddr, pmd, entry);
681 	mm_inc_nr_ptes(mm);
682 	return true;
683 }
684 
685 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
686 {
687 	struct vm_area_struct *vma = vmf->vma;
688 	gfp_t gfp;
689 	struct page *page;
690 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
691 
692 	if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
693 		return VM_FAULT_FALLBACK;
694 	if (unlikely(anon_vma_prepare(vma)))
695 		return VM_FAULT_OOM;
696 	if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
697 		return VM_FAULT_OOM;
698 	if (!(vmf->flags & FAULT_FLAG_WRITE) &&
699 			!mm_forbids_zeropage(vma->vm_mm) &&
700 			transparent_hugepage_use_zero_page()) {
701 		pgtable_t pgtable;
702 		struct page *zero_page;
703 		bool set;
704 		vm_fault_t ret;
705 		pgtable = pte_alloc_one(vma->vm_mm);
706 		if (unlikely(!pgtable))
707 			return VM_FAULT_OOM;
708 		zero_page = mm_get_huge_zero_page(vma->vm_mm);
709 		if (unlikely(!zero_page)) {
710 			pte_free(vma->vm_mm, pgtable);
711 			count_vm_event(THP_FAULT_FALLBACK);
712 			return VM_FAULT_FALLBACK;
713 		}
714 		vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
715 		ret = 0;
716 		set = false;
717 		if (pmd_none(*vmf->pmd)) {
718 			ret = check_stable_address_space(vma->vm_mm);
719 			if (ret) {
720 				spin_unlock(vmf->ptl);
721 			} else if (userfaultfd_missing(vma)) {
722 				spin_unlock(vmf->ptl);
723 				ret = handle_userfault(vmf, VM_UFFD_MISSING);
724 				VM_BUG_ON(ret & VM_FAULT_FALLBACK);
725 			} else {
726 				set_huge_zero_page(pgtable, vma->vm_mm, vma,
727 						   haddr, vmf->pmd, zero_page);
728 				spin_unlock(vmf->ptl);
729 				set = true;
730 			}
731 		} else
732 			spin_unlock(vmf->ptl);
733 		if (!set)
734 			pte_free(vma->vm_mm, pgtable);
735 		return ret;
736 	}
737 	gfp = alloc_hugepage_direct_gfpmask(vma);
738 	page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
739 	if (unlikely(!page)) {
740 		count_vm_event(THP_FAULT_FALLBACK);
741 		return VM_FAULT_FALLBACK;
742 	}
743 	prep_transhuge_page(page);
744 	return __do_huge_pmd_anonymous_page(vmf, page, gfp);
745 }
746 
747 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
748 		pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
749 		pgtable_t pgtable)
750 {
751 	struct mm_struct *mm = vma->vm_mm;
752 	pmd_t entry;
753 	spinlock_t *ptl;
754 
755 	ptl = pmd_lock(mm, pmd);
756 	if (!pmd_none(*pmd)) {
757 		if (write) {
758 			if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
759 				WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
760 				goto out_unlock;
761 			}
762 			entry = pmd_mkyoung(*pmd);
763 			entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
764 			if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
765 				update_mmu_cache_pmd(vma, addr, pmd);
766 		}
767 
768 		goto out_unlock;
769 	}
770 
771 	entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
772 	if (pfn_t_devmap(pfn))
773 		entry = pmd_mkdevmap(entry);
774 	if (write) {
775 		entry = pmd_mkyoung(pmd_mkdirty(entry));
776 		entry = maybe_pmd_mkwrite(entry, vma);
777 	}
778 
779 	if (pgtable) {
780 		pgtable_trans_huge_deposit(mm, pmd, pgtable);
781 		mm_inc_nr_ptes(mm);
782 		pgtable = NULL;
783 	}
784 
785 	set_pmd_at(mm, addr, pmd, entry);
786 	update_mmu_cache_pmd(vma, addr, pmd);
787 
788 out_unlock:
789 	spin_unlock(ptl);
790 	if (pgtable)
791 		pte_free(mm, pgtable);
792 }
793 
794 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)
795 {
796 	unsigned long addr = vmf->address & PMD_MASK;
797 	struct vm_area_struct *vma = vmf->vma;
798 	pgprot_t pgprot = vma->vm_page_prot;
799 	pgtable_t pgtable = NULL;
800 
801 	/*
802 	 * If we had pmd_special, we could avoid all these restrictions,
803 	 * but we need to be consistent with PTEs and architectures that
804 	 * can't support a 'special' bit.
805 	 */
806 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
807 			!pfn_t_devmap(pfn));
808 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
809 						(VM_PFNMAP|VM_MIXEDMAP));
810 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
811 
812 	if (addr < vma->vm_start || addr >= vma->vm_end)
813 		return VM_FAULT_SIGBUS;
814 
815 	if (arch_needs_pgtable_deposit()) {
816 		pgtable = pte_alloc_one(vma->vm_mm);
817 		if (!pgtable)
818 			return VM_FAULT_OOM;
819 	}
820 
821 	track_pfn_insert(vma, &pgprot, pfn);
822 
823 	insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
824 	return VM_FAULT_NOPAGE;
825 }
826 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
827 
828 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
829 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
830 {
831 	if (likely(vma->vm_flags & VM_WRITE))
832 		pud = pud_mkwrite(pud);
833 	return pud;
834 }
835 
836 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
837 		pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
838 {
839 	struct mm_struct *mm = vma->vm_mm;
840 	pud_t entry;
841 	spinlock_t *ptl;
842 
843 	ptl = pud_lock(mm, pud);
844 	if (!pud_none(*pud)) {
845 		if (write) {
846 			if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
847 				WARN_ON_ONCE(!is_huge_zero_pud(*pud));
848 				goto out_unlock;
849 			}
850 			entry = pud_mkyoung(*pud);
851 			entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
852 			if (pudp_set_access_flags(vma, addr, pud, entry, 1))
853 				update_mmu_cache_pud(vma, addr, pud);
854 		}
855 		goto out_unlock;
856 	}
857 
858 	entry = pud_mkhuge(pfn_t_pud(pfn, prot));
859 	if (pfn_t_devmap(pfn))
860 		entry = pud_mkdevmap(entry);
861 	if (write) {
862 		entry = pud_mkyoung(pud_mkdirty(entry));
863 		entry = maybe_pud_mkwrite(entry, vma);
864 	}
865 	set_pud_at(mm, addr, pud, entry);
866 	update_mmu_cache_pud(vma, addr, pud);
867 
868 out_unlock:
869 	spin_unlock(ptl);
870 }
871 
872 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)
873 {
874 	unsigned long addr = vmf->address & PUD_MASK;
875 	struct vm_area_struct *vma = vmf->vma;
876 	pgprot_t pgprot = vma->vm_page_prot;
877 
878 	/*
879 	 * If we had pud_special, we could avoid all these restrictions,
880 	 * but we need to be consistent with PTEs and architectures that
881 	 * can't support a 'special' bit.
882 	 */
883 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
884 			!pfn_t_devmap(pfn));
885 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
886 						(VM_PFNMAP|VM_MIXEDMAP));
887 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
888 
889 	if (addr < vma->vm_start || addr >= vma->vm_end)
890 		return VM_FAULT_SIGBUS;
891 
892 	track_pfn_insert(vma, &pgprot, pfn);
893 
894 	insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
895 	return VM_FAULT_NOPAGE;
896 }
897 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
898 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
899 
900 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
901 		pmd_t *pmd, int flags)
902 {
903 	pmd_t _pmd;
904 
905 	_pmd = pmd_mkyoung(*pmd);
906 	if (flags & FOLL_WRITE)
907 		_pmd = pmd_mkdirty(_pmd);
908 	if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
909 				pmd, _pmd, flags & FOLL_WRITE))
910 		update_mmu_cache_pmd(vma, addr, pmd);
911 }
912 
913 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
914 		pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
915 {
916 	unsigned long pfn = pmd_pfn(*pmd);
917 	struct mm_struct *mm = vma->vm_mm;
918 	struct page *page;
919 
920 	assert_spin_locked(pmd_lockptr(mm, pmd));
921 
922 	/*
923 	 * When we COW a devmap PMD entry, we split it into PTEs, so we should
924 	 * not be in this function with `flags & FOLL_COW` set.
925 	 */
926 	WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
927 
928 	if (flags & FOLL_WRITE && !pmd_write(*pmd))
929 		return NULL;
930 
931 	if (pmd_present(*pmd) && pmd_devmap(*pmd))
932 		/* pass */;
933 	else
934 		return NULL;
935 
936 	if (flags & FOLL_TOUCH)
937 		touch_pmd(vma, addr, pmd, flags);
938 
939 	/*
940 	 * device mapped pages can only be returned if the
941 	 * caller will manage the page reference count.
942 	 */
943 	if (!(flags & FOLL_GET))
944 		return ERR_PTR(-EEXIST);
945 
946 	pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
947 	*pgmap = get_dev_pagemap(pfn, *pgmap);
948 	if (!*pgmap)
949 		return ERR_PTR(-EFAULT);
950 	page = pfn_to_page(pfn);
951 	get_page(page);
952 
953 	return page;
954 }
955 
956 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
957 		  pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
958 		  struct vm_area_struct *vma)
959 {
960 	spinlock_t *dst_ptl, *src_ptl;
961 	struct page *src_page;
962 	pmd_t pmd;
963 	pgtable_t pgtable = NULL;
964 	int ret = -ENOMEM;
965 
966 	/* Skip if can be re-fill on fault */
967 	if (!vma_is_anonymous(vma))
968 		return 0;
969 
970 	pgtable = pte_alloc_one(dst_mm);
971 	if (unlikely(!pgtable))
972 		goto out;
973 
974 	dst_ptl = pmd_lock(dst_mm, dst_pmd);
975 	src_ptl = pmd_lockptr(src_mm, src_pmd);
976 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
977 
978 	ret = -EAGAIN;
979 	pmd = *src_pmd;
980 
981 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
982 	if (unlikely(is_swap_pmd(pmd))) {
983 		swp_entry_t entry = pmd_to_swp_entry(pmd);
984 
985 		VM_BUG_ON(!is_pmd_migration_entry(pmd));
986 		if (is_write_migration_entry(entry)) {
987 			make_migration_entry_read(&entry);
988 			pmd = swp_entry_to_pmd(entry);
989 			if (pmd_swp_soft_dirty(*src_pmd))
990 				pmd = pmd_swp_mksoft_dirty(pmd);
991 			set_pmd_at(src_mm, addr, src_pmd, pmd);
992 		}
993 		add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
994 		mm_inc_nr_ptes(dst_mm);
995 		pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
996 		set_pmd_at(dst_mm, addr, dst_pmd, pmd);
997 		ret = 0;
998 		goto out_unlock;
999 	}
1000 #endif
1001 
1002 	if (unlikely(!pmd_trans_huge(pmd))) {
1003 		pte_free(dst_mm, pgtable);
1004 		goto out_unlock;
1005 	}
1006 	/*
1007 	 * When page table lock is held, the huge zero pmd should not be
1008 	 * under splitting since we don't split the page itself, only pmd to
1009 	 * a page table.
1010 	 */
1011 	if (is_huge_zero_pmd(pmd)) {
1012 		struct page *zero_page;
1013 		/*
1014 		 * get_huge_zero_page() will never allocate a new page here,
1015 		 * since we already have a zero page to copy. It just takes a
1016 		 * reference.
1017 		 */
1018 		zero_page = mm_get_huge_zero_page(dst_mm);
1019 		set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1020 				zero_page);
1021 		ret = 0;
1022 		goto out_unlock;
1023 	}
1024 
1025 	src_page = pmd_page(pmd);
1026 	VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1027 	get_page(src_page);
1028 	page_dup_rmap(src_page, true);
1029 	add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1030 	mm_inc_nr_ptes(dst_mm);
1031 	pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1032 
1033 	pmdp_set_wrprotect(src_mm, addr, src_pmd);
1034 	pmd = pmd_mkold(pmd_wrprotect(pmd));
1035 	set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1036 
1037 	ret = 0;
1038 out_unlock:
1039 	spin_unlock(src_ptl);
1040 	spin_unlock(dst_ptl);
1041 out:
1042 	return ret;
1043 }
1044 
1045 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1046 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1047 		pud_t *pud, int flags)
1048 {
1049 	pud_t _pud;
1050 
1051 	_pud = pud_mkyoung(*pud);
1052 	if (flags & FOLL_WRITE)
1053 		_pud = pud_mkdirty(_pud);
1054 	if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1055 				pud, _pud, flags & FOLL_WRITE))
1056 		update_mmu_cache_pud(vma, addr, pud);
1057 }
1058 
1059 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1060 		pud_t *pud, int flags, struct dev_pagemap **pgmap)
1061 {
1062 	unsigned long pfn = pud_pfn(*pud);
1063 	struct mm_struct *mm = vma->vm_mm;
1064 	struct page *page;
1065 
1066 	assert_spin_locked(pud_lockptr(mm, pud));
1067 
1068 	if (flags & FOLL_WRITE && !pud_write(*pud))
1069 		return NULL;
1070 
1071 	if (pud_present(*pud) && pud_devmap(*pud))
1072 		/* pass */;
1073 	else
1074 		return NULL;
1075 
1076 	if (flags & FOLL_TOUCH)
1077 		touch_pud(vma, addr, pud, flags);
1078 
1079 	/*
1080 	 * device mapped pages can only be returned if the
1081 	 * caller will manage the page reference count.
1082 	 */
1083 	if (!(flags & FOLL_GET))
1084 		return ERR_PTR(-EEXIST);
1085 
1086 	pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1087 	*pgmap = get_dev_pagemap(pfn, *pgmap);
1088 	if (!*pgmap)
1089 		return ERR_PTR(-EFAULT);
1090 	page = pfn_to_page(pfn);
1091 	get_page(page);
1092 
1093 	return page;
1094 }
1095 
1096 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1097 		  pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1098 		  struct vm_area_struct *vma)
1099 {
1100 	spinlock_t *dst_ptl, *src_ptl;
1101 	pud_t pud;
1102 	int ret;
1103 
1104 	dst_ptl = pud_lock(dst_mm, dst_pud);
1105 	src_ptl = pud_lockptr(src_mm, src_pud);
1106 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1107 
1108 	ret = -EAGAIN;
1109 	pud = *src_pud;
1110 	if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1111 		goto out_unlock;
1112 
1113 	/*
1114 	 * When page table lock is held, the huge zero pud should not be
1115 	 * under splitting since we don't split the page itself, only pud to
1116 	 * a page table.
1117 	 */
1118 	if (is_huge_zero_pud(pud)) {
1119 		/* No huge zero pud yet */
1120 	}
1121 
1122 	pudp_set_wrprotect(src_mm, addr, src_pud);
1123 	pud = pud_mkold(pud_wrprotect(pud));
1124 	set_pud_at(dst_mm, addr, dst_pud, pud);
1125 
1126 	ret = 0;
1127 out_unlock:
1128 	spin_unlock(src_ptl);
1129 	spin_unlock(dst_ptl);
1130 	return ret;
1131 }
1132 
1133 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1134 {
1135 	pud_t entry;
1136 	unsigned long haddr;
1137 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1138 
1139 	vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1140 	if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1141 		goto unlock;
1142 
1143 	entry = pud_mkyoung(orig_pud);
1144 	if (write)
1145 		entry = pud_mkdirty(entry);
1146 	haddr = vmf->address & HPAGE_PUD_MASK;
1147 	if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1148 		update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1149 
1150 unlock:
1151 	spin_unlock(vmf->ptl);
1152 }
1153 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1154 
1155 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1156 {
1157 	pmd_t entry;
1158 	unsigned long haddr;
1159 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1160 
1161 	vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1162 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1163 		goto unlock;
1164 
1165 	entry = pmd_mkyoung(orig_pmd);
1166 	if (write)
1167 		entry = pmd_mkdirty(entry);
1168 	haddr = vmf->address & HPAGE_PMD_MASK;
1169 	if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1170 		update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1171 
1172 unlock:
1173 	spin_unlock(vmf->ptl);
1174 }
1175 
1176 static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1177 			pmd_t orig_pmd, struct page *page)
1178 {
1179 	struct vm_area_struct *vma = vmf->vma;
1180 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1181 	struct mem_cgroup *memcg;
1182 	pgtable_t pgtable;
1183 	pmd_t _pmd;
1184 	int i;
1185 	vm_fault_t ret = 0;
1186 	struct page **pages;
1187 	struct mmu_notifier_range range;
1188 
1189 	pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1190 			      GFP_KERNEL);
1191 	if (unlikely(!pages)) {
1192 		ret |= VM_FAULT_OOM;
1193 		goto out;
1194 	}
1195 
1196 	for (i = 0; i < HPAGE_PMD_NR; i++) {
1197 		pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1198 					       vmf->address, page_to_nid(page));
1199 		if (unlikely(!pages[i] ||
1200 			     mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
1201 				     GFP_KERNEL, &memcg, false))) {
1202 			if (pages[i])
1203 				put_page(pages[i]);
1204 			while (--i >= 0) {
1205 				memcg = (void *)page_private(pages[i]);
1206 				set_page_private(pages[i], 0);
1207 				mem_cgroup_cancel_charge(pages[i], memcg,
1208 						false);
1209 				put_page(pages[i]);
1210 			}
1211 			kfree(pages);
1212 			ret |= VM_FAULT_OOM;
1213 			goto out;
1214 		}
1215 		set_page_private(pages[i], (unsigned long)memcg);
1216 	}
1217 
1218 	for (i = 0; i < HPAGE_PMD_NR; i++) {
1219 		copy_user_highpage(pages[i], page + i,
1220 				   haddr + PAGE_SIZE * i, vma);
1221 		__SetPageUptodate(pages[i]);
1222 		cond_resched();
1223 	}
1224 
1225 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1226 				haddr, haddr + HPAGE_PMD_SIZE);
1227 	mmu_notifier_invalidate_range_start(&range);
1228 
1229 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1230 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1231 		goto out_free_pages;
1232 	VM_BUG_ON_PAGE(!PageHead(page), page);
1233 
1234 	/*
1235 	 * Leave pmd empty until pte is filled note we must notify here as
1236 	 * concurrent CPU thread might write to new page before the call to
1237 	 * mmu_notifier_invalidate_range_end() happens which can lead to a
1238 	 * device seeing memory write in different order than CPU.
1239 	 *
1240 	 * See Documentation/vm/mmu_notifier.rst
1241 	 */
1242 	pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1243 
1244 	pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1245 	pmd_populate(vma->vm_mm, &_pmd, pgtable);
1246 
1247 	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1248 		pte_t entry;
1249 		entry = mk_pte(pages[i], vma->vm_page_prot);
1250 		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1251 		memcg = (void *)page_private(pages[i]);
1252 		set_page_private(pages[i], 0);
1253 		page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1254 		mem_cgroup_commit_charge(pages[i], memcg, false, false);
1255 		lru_cache_add_active_or_unevictable(pages[i], vma);
1256 		vmf->pte = pte_offset_map(&_pmd, haddr);
1257 		VM_BUG_ON(!pte_none(*vmf->pte));
1258 		set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1259 		pte_unmap(vmf->pte);
1260 	}
1261 	kfree(pages);
1262 
1263 	smp_wmb(); /* make pte visible before pmd */
1264 	pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1265 	page_remove_rmap(page, true);
1266 	spin_unlock(vmf->ptl);
1267 
1268 	/*
1269 	 * No need to double call mmu_notifier->invalidate_range() callback as
1270 	 * the above pmdp_huge_clear_flush_notify() did already call it.
1271 	 */
1272 	mmu_notifier_invalidate_range_only_end(&range);
1273 
1274 	ret |= VM_FAULT_WRITE;
1275 	put_page(page);
1276 
1277 out:
1278 	return ret;
1279 
1280 out_free_pages:
1281 	spin_unlock(vmf->ptl);
1282 	mmu_notifier_invalidate_range_end(&range);
1283 	for (i = 0; i < HPAGE_PMD_NR; i++) {
1284 		memcg = (void *)page_private(pages[i]);
1285 		set_page_private(pages[i], 0);
1286 		mem_cgroup_cancel_charge(pages[i], memcg, false);
1287 		put_page(pages[i]);
1288 	}
1289 	kfree(pages);
1290 	goto out;
1291 }
1292 
1293 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1294 {
1295 	struct vm_area_struct *vma = vmf->vma;
1296 	struct page *page = NULL, *new_page;
1297 	struct mem_cgroup *memcg;
1298 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1299 	struct mmu_notifier_range range;
1300 	gfp_t huge_gfp;			/* for allocation and charge */
1301 	vm_fault_t ret = 0;
1302 
1303 	vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1304 	VM_BUG_ON_VMA(!vma->anon_vma, vma);
1305 	if (is_huge_zero_pmd(orig_pmd))
1306 		goto alloc;
1307 	spin_lock(vmf->ptl);
1308 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1309 		goto out_unlock;
1310 
1311 	page = pmd_page(orig_pmd);
1312 	VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1313 	/*
1314 	 * We can only reuse the page if nobody else maps the huge page or it's
1315 	 * part.
1316 	 */
1317 	if (!trylock_page(page)) {
1318 		get_page(page);
1319 		spin_unlock(vmf->ptl);
1320 		lock_page(page);
1321 		spin_lock(vmf->ptl);
1322 		if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1323 			unlock_page(page);
1324 			put_page(page);
1325 			goto out_unlock;
1326 		}
1327 		put_page(page);
1328 	}
1329 	if (reuse_swap_page(page, NULL)) {
1330 		pmd_t entry;
1331 		entry = pmd_mkyoung(orig_pmd);
1332 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1333 		if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1334 			update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1335 		ret |= VM_FAULT_WRITE;
1336 		unlock_page(page);
1337 		goto out_unlock;
1338 	}
1339 	unlock_page(page);
1340 	get_page(page);
1341 	spin_unlock(vmf->ptl);
1342 alloc:
1343 	if (__transparent_hugepage_enabled(vma) &&
1344 	    !transparent_hugepage_debug_cow()) {
1345 		huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1346 		new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1347 	} else
1348 		new_page = NULL;
1349 
1350 	if (likely(new_page)) {
1351 		prep_transhuge_page(new_page);
1352 	} else {
1353 		if (!page) {
1354 			split_huge_pmd(vma, vmf->pmd, vmf->address);
1355 			ret |= VM_FAULT_FALLBACK;
1356 		} else {
1357 			ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1358 			if (ret & VM_FAULT_OOM) {
1359 				split_huge_pmd(vma, vmf->pmd, vmf->address);
1360 				ret |= VM_FAULT_FALLBACK;
1361 			}
1362 			put_page(page);
1363 		}
1364 		count_vm_event(THP_FAULT_FALLBACK);
1365 		goto out;
1366 	}
1367 
1368 	if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
1369 					huge_gfp, &memcg, true))) {
1370 		put_page(new_page);
1371 		split_huge_pmd(vma, vmf->pmd, vmf->address);
1372 		if (page)
1373 			put_page(page);
1374 		ret |= VM_FAULT_FALLBACK;
1375 		count_vm_event(THP_FAULT_FALLBACK);
1376 		goto out;
1377 	}
1378 
1379 	count_vm_event(THP_FAULT_ALLOC);
1380 	count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
1381 
1382 	if (!page)
1383 		clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1384 	else
1385 		copy_user_huge_page(new_page, page, vmf->address,
1386 				    vma, HPAGE_PMD_NR);
1387 	__SetPageUptodate(new_page);
1388 
1389 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1390 				haddr, haddr + HPAGE_PMD_SIZE);
1391 	mmu_notifier_invalidate_range_start(&range);
1392 
1393 	spin_lock(vmf->ptl);
1394 	if (page)
1395 		put_page(page);
1396 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1397 		spin_unlock(vmf->ptl);
1398 		mem_cgroup_cancel_charge(new_page, memcg, true);
1399 		put_page(new_page);
1400 		goto out_mn;
1401 	} else {
1402 		pmd_t entry;
1403 		entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1404 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1405 		pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1406 		page_add_new_anon_rmap(new_page, vma, haddr, true);
1407 		mem_cgroup_commit_charge(new_page, memcg, false, true);
1408 		lru_cache_add_active_or_unevictable(new_page, vma);
1409 		set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1410 		update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1411 		if (!page) {
1412 			add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1413 		} else {
1414 			VM_BUG_ON_PAGE(!PageHead(page), page);
1415 			page_remove_rmap(page, true);
1416 			put_page(page);
1417 		}
1418 		ret |= VM_FAULT_WRITE;
1419 	}
1420 	spin_unlock(vmf->ptl);
1421 out_mn:
1422 	/*
1423 	 * No need to double call mmu_notifier->invalidate_range() callback as
1424 	 * the above pmdp_huge_clear_flush_notify() did already call it.
1425 	 */
1426 	mmu_notifier_invalidate_range_only_end(&range);
1427 out:
1428 	return ret;
1429 out_unlock:
1430 	spin_unlock(vmf->ptl);
1431 	return ret;
1432 }
1433 
1434 /*
1435  * FOLL_FORCE can write to even unwritable pmd's, but only
1436  * after we've gone through a COW cycle and they are dirty.
1437  */
1438 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1439 {
1440 	return pmd_write(pmd) ||
1441 	       ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1442 }
1443 
1444 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1445 				   unsigned long addr,
1446 				   pmd_t *pmd,
1447 				   unsigned int flags)
1448 {
1449 	struct mm_struct *mm = vma->vm_mm;
1450 	struct page *page = NULL;
1451 
1452 	assert_spin_locked(pmd_lockptr(mm, pmd));
1453 
1454 	if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1455 		goto out;
1456 
1457 	/* Avoid dumping huge zero page */
1458 	if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1459 		return ERR_PTR(-EFAULT);
1460 
1461 	/* Full NUMA hinting faults to serialise migration in fault paths */
1462 	if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1463 		goto out;
1464 
1465 	page = pmd_page(*pmd);
1466 	VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1467 	if (flags & FOLL_TOUCH)
1468 		touch_pmd(vma, addr, pmd, flags);
1469 	if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1470 		/*
1471 		 * We don't mlock() pte-mapped THPs. This way we can avoid
1472 		 * leaking mlocked pages into non-VM_LOCKED VMAs.
1473 		 *
1474 		 * For anon THP:
1475 		 *
1476 		 * In most cases the pmd is the only mapping of the page as we
1477 		 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1478 		 * writable private mappings in populate_vma_page_range().
1479 		 *
1480 		 * The only scenario when we have the page shared here is if we
1481 		 * mlocking read-only mapping shared over fork(). We skip
1482 		 * mlocking such pages.
1483 		 *
1484 		 * For file THP:
1485 		 *
1486 		 * We can expect PageDoubleMap() to be stable under page lock:
1487 		 * for file pages we set it in page_add_file_rmap(), which
1488 		 * requires page to be locked.
1489 		 */
1490 
1491 		if (PageAnon(page) && compound_mapcount(page) != 1)
1492 			goto skip_mlock;
1493 		if (PageDoubleMap(page) || !page->mapping)
1494 			goto skip_mlock;
1495 		if (!trylock_page(page))
1496 			goto skip_mlock;
1497 		lru_add_drain();
1498 		if (page->mapping && !PageDoubleMap(page))
1499 			mlock_vma_page(page);
1500 		unlock_page(page);
1501 	}
1502 skip_mlock:
1503 	page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1504 	VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1505 	if (flags & FOLL_GET)
1506 		get_page(page);
1507 
1508 out:
1509 	return page;
1510 }
1511 
1512 /* NUMA hinting page fault entry point for trans huge pmds */
1513 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1514 {
1515 	struct vm_area_struct *vma = vmf->vma;
1516 	struct anon_vma *anon_vma = NULL;
1517 	struct page *page;
1518 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1519 	int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1520 	int target_nid, last_cpupid = -1;
1521 	bool page_locked;
1522 	bool migrated = false;
1523 	bool was_writable;
1524 	int flags = 0;
1525 
1526 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1527 	if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1528 		goto out_unlock;
1529 
1530 	/*
1531 	 * If there are potential migrations, wait for completion and retry
1532 	 * without disrupting NUMA hinting information. Do not relock and
1533 	 * check_same as the page may no longer be mapped.
1534 	 */
1535 	if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1536 		page = pmd_page(*vmf->pmd);
1537 		if (!get_page_unless_zero(page))
1538 			goto out_unlock;
1539 		spin_unlock(vmf->ptl);
1540 		put_and_wait_on_page_locked(page);
1541 		goto out;
1542 	}
1543 
1544 	page = pmd_page(pmd);
1545 	BUG_ON(is_huge_zero_page(page));
1546 	page_nid = page_to_nid(page);
1547 	last_cpupid = page_cpupid_last(page);
1548 	count_vm_numa_event(NUMA_HINT_FAULTS);
1549 	if (page_nid == this_nid) {
1550 		count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1551 		flags |= TNF_FAULT_LOCAL;
1552 	}
1553 
1554 	/* See similar comment in do_numa_page for explanation */
1555 	if (!pmd_savedwrite(pmd))
1556 		flags |= TNF_NO_GROUP;
1557 
1558 	/*
1559 	 * Acquire the page lock to serialise THP migrations but avoid dropping
1560 	 * page_table_lock if at all possible
1561 	 */
1562 	page_locked = trylock_page(page);
1563 	target_nid = mpol_misplaced(page, vma, haddr);
1564 	if (target_nid == NUMA_NO_NODE) {
1565 		/* If the page was locked, there are no parallel migrations */
1566 		if (page_locked)
1567 			goto clear_pmdnuma;
1568 	}
1569 
1570 	/* Migration could have started since the pmd_trans_migrating check */
1571 	if (!page_locked) {
1572 		page_nid = NUMA_NO_NODE;
1573 		if (!get_page_unless_zero(page))
1574 			goto out_unlock;
1575 		spin_unlock(vmf->ptl);
1576 		put_and_wait_on_page_locked(page);
1577 		goto out;
1578 	}
1579 
1580 	/*
1581 	 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1582 	 * to serialises splits
1583 	 */
1584 	get_page(page);
1585 	spin_unlock(vmf->ptl);
1586 	anon_vma = page_lock_anon_vma_read(page);
1587 
1588 	/* Confirm the PMD did not change while page_table_lock was released */
1589 	spin_lock(vmf->ptl);
1590 	if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1591 		unlock_page(page);
1592 		put_page(page);
1593 		page_nid = NUMA_NO_NODE;
1594 		goto out_unlock;
1595 	}
1596 
1597 	/* Bail if we fail to protect against THP splits for any reason */
1598 	if (unlikely(!anon_vma)) {
1599 		put_page(page);
1600 		page_nid = NUMA_NO_NODE;
1601 		goto clear_pmdnuma;
1602 	}
1603 
1604 	/*
1605 	 * Since we took the NUMA fault, we must have observed the !accessible
1606 	 * bit. Make sure all other CPUs agree with that, to avoid them
1607 	 * modifying the page we're about to migrate.
1608 	 *
1609 	 * Must be done under PTL such that we'll observe the relevant
1610 	 * inc_tlb_flush_pending().
1611 	 *
1612 	 * We are not sure a pending tlb flush here is for a huge page
1613 	 * mapping or not. Hence use the tlb range variant
1614 	 */
1615 	if (mm_tlb_flush_pending(vma->vm_mm)) {
1616 		flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1617 		/*
1618 		 * change_huge_pmd() released the pmd lock before
1619 		 * invalidating the secondary MMUs sharing the primary
1620 		 * MMU pagetables (with ->invalidate_range()). The
1621 		 * mmu_notifier_invalidate_range_end() (which
1622 		 * internally calls ->invalidate_range()) in
1623 		 * change_pmd_range() will run after us, so we can't
1624 		 * rely on it here and we need an explicit invalidate.
1625 		 */
1626 		mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1627 					      haddr + HPAGE_PMD_SIZE);
1628 	}
1629 
1630 	/*
1631 	 * Migrate the THP to the requested node, returns with page unlocked
1632 	 * and access rights restored.
1633 	 */
1634 	spin_unlock(vmf->ptl);
1635 
1636 	migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1637 				vmf->pmd, pmd, vmf->address, page, target_nid);
1638 	if (migrated) {
1639 		flags |= TNF_MIGRATED;
1640 		page_nid = target_nid;
1641 	} else
1642 		flags |= TNF_MIGRATE_FAIL;
1643 
1644 	goto out;
1645 clear_pmdnuma:
1646 	BUG_ON(!PageLocked(page));
1647 	was_writable = pmd_savedwrite(pmd);
1648 	pmd = pmd_modify(pmd, vma->vm_page_prot);
1649 	pmd = pmd_mkyoung(pmd);
1650 	if (was_writable)
1651 		pmd = pmd_mkwrite(pmd);
1652 	set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1653 	update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1654 	unlock_page(page);
1655 out_unlock:
1656 	spin_unlock(vmf->ptl);
1657 
1658 out:
1659 	if (anon_vma)
1660 		page_unlock_anon_vma_read(anon_vma);
1661 
1662 	if (page_nid != NUMA_NO_NODE)
1663 		task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1664 				flags);
1665 
1666 	return 0;
1667 }
1668 
1669 /*
1670  * Return true if we do MADV_FREE successfully on entire pmd page.
1671  * Otherwise, return false.
1672  */
1673 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1674 		pmd_t *pmd, unsigned long addr, unsigned long next)
1675 {
1676 	spinlock_t *ptl;
1677 	pmd_t orig_pmd;
1678 	struct page *page;
1679 	struct mm_struct *mm = tlb->mm;
1680 	bool ret = false;
1681 
1682 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1683 
1684 	ptl = pmd_trans_huge_lock(pmd, vma);
1685 	if (!ptl)
1686 		goto out_unlocked;
1687 
1688 	orig_pmd = *pmd;
1689 	if (is_huge_zero_pmd(orig_pmd))
1690 		goto out;
1691 
1692 	if (unlikely(!pmd_present(orig_pmd))) {
1693 		VM_BUG_ON(thp_migration_supported() &&
1694 				  !is_pmd_migration_entry(orig_pmd));
1695 		goto out;
1696 	}
1697 
1698 	page = pmd_page(orig_pmd);
1699 	/*
1700 	 * If other processes are mapping this page, we couldn't discard
1701 	 * the page unless they all do MADV_FREE so let's skip the page.
1702 	 */
1703 	if (page_mapcount(page) != 1)
1704 		goto out;
1705 
1706 	if (!trylock_page(page))
1707 		goto out;
1708 
1709 	/*
1710 	 * If user want to discard part-pages of THP, split it so MADV_FREE
1711 	 * will deactivate only them.
1712 	 */
1713 	if (next - addr != HPAGE_PMD_SIZE) {
1714 		get_page(page);
1715 		spin_unlock(ptl);
1716 		split_huge_page(page);
1717 		unlock_page(page);
1718 		put_page(page);
1719 		goto out_unlocked;
1720 	}
1721 
1722 	if (PageDirty(page))
1723 		ClearPageDirty(page);
1724 	unlock_page(page);
1725 
1726 	if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1727 		pmdp_invalidate(vma, addr, pmd);
1728 		orig_pmd = pmd_mkold(orig_pmd);
1729 		orig_pmd = pmd_mkclean(orig_pmd);
1730 
1731 		set_pmd_at(mm, addr, pmd, orig_pmd);
1732 		tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1733 	}
1734 
1735 	mark_page_lazyfree(page);
1736 	ret = true;
1737 out:
1738 	spin_unlock(ptl);
1739 out_unlocked:
1740 	return ret;
1741 }
1742 
1743 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1744 {
1745 	pgtable_t pgtable;
1746 
1747 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1748 	pte_free(mm, pgtable);
1749 	mm_dec_nr_ptes(mm);
1750 }
1751 
1752 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1753 		 pmd_t *pmd, unsigned long addr)
1754 {
1755 	pmd_t orig_pmd;
1756 	spinlock_t *ptl;
1757 
1758 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1759 
1760 	ptl = __pmd_trans_huge_lock(pmd, vma);
1761 	if (!ptl)
1762 		return 0;
1763 	/*
1764 	 * For architectures like ppc64 we look at deposited pgtable
1765 	 * when calling pmdp_huge_get_and_clear. So do the
1766 	 * pgtable_trans_huge_withdraw after finishing pmdp related
1767 	 * operations.
1768 	 */
1769 	orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1770 			tlb->fullmm);
1771 	tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1772 	if (vma_is_dax(vma)) {
1773 		if (arch_needs_pgtable_deposit())
1774 			zap_deposited_table(tlb->mm, pmd);
1775 		spin_unlock(ptl);
1776 		if (is_huge_zero_pmd(orig_pmd))
1777 			tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1778 	} else if (is_huge_zero_pmd(orig_pmd)) {
1779 		zap_deposited_table(tlb->mm, pmd);
1780 		spin_unlock(ptl);
1781 		tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1782 	} else {
1783 		struct page *page = NULL;
1784 		int flush_needed = 1;
1785 
1786 		if (pmd_present(orig_pmd)) {
1787 			page = pmd_page(orig_pmd);
1788 			page_remove_rmap(page, true);
1789 			VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1790 			VM_BUG_ON_PAGE(!PageHead(page), page);
1791 		} else if (thp_migration_supported()) {
1792 			swp_entry_t entry;
1793 
1794 			VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1795 			entry = pmd_to_swp_entry(orig_pmd);
1796 			page = pfn_to_page(swp_offset(entry));
1797 			flush_needed = 0;
1798 		} else
1799 			WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1800 
1801 		if (PageAnon(page)) {
1802 			zap_deposited_table(tlb->mm, pmd);
1803 			add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1804 		} else {
1805 			if (arch_needs_pgtable_deposit())
1806 				zap_deposited_table(tlb->mm, pmd);
1807 			add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1808 		}
1809 
1810 		spin_unlock(ptl);
1811 		if (flush_needed)
1812 			tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1813 	}
1814 	return 1;
1815 }
1816 
1817 #ifndef pmd_move_must_withdraw
1818 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1819 					 spinlock_t *old_pmd_ptl,
1820 					 struct vm_area_struct *vma)
1821 {
1822 	/*
1823 	 * With split pmd lock we also need to move preallocated
1824 	 * PTE page table if new_pmd is on different PMD page table.
1825 	 *
1826 	 * We also don't deposit and withdraw tables for file pages.
1827 	 */
1828 	return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1829 }
1830 #endif
1831 
1832 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1833 {
1834 #ifdef CONFIG_MEM_SOFT_DIRTY
1835 	if (unlikely(is_pmd_migration_entry(pmd)))
1836 		pmd = pmd_swp_mksoft_dirty(pmd);
1837 	else if (pmd_present(pmd))
1838 		pmd = pmd_mksoft_dirty(pmd);
1839 #endif
1840 	return pmd;
1841 }
1842 
1843 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1844 		  unsigned long new_addr, unsigned long old_end,
1845 		  pmd_t *old_pmd, pmd_t *new_pmd)
1846 {
1847 	spinlock_t *old_ptl, *new_ptl;
1848 	pmd_t pmd;
1849 	struct mm_struct *mm = vma->vm_mm;
1850 	bool force_flush = false;
1851 
1852 	if ((old_addr & ~HPAGE_PMD_MASK) ||
1853 	    (new_addr & ~HPAGE_PMD_MASK) ||
1854 	    old_end - old_addr < HPAGE_PMD_SIZE)
1855 		return false;
1856 
1857 	/*
1858 	 * The destination pmd shouldn't be established, free_pgtables()
1859 	 * should have release it.
1860 	 */
1861 	if (WARN_ON(!pmd_none(*new_pmd))) {
1862 		VM_BUG_ON(pmd_trans_huge(*new_pmd));
1863 		return false;
1864 	}
1865 
1866 	/*
1867 	 * We don't have to worry about the ordering of src and dst
1868 	 * ptlocks because exclusive mmap_sem prevents deadlock.
1869 	 */
1870 	old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1871 	if (old_ptl) {
1872 		new_ptl = pmd_lockptr(mm, new_pmd);
1873 		if (new_ptl != old_ptl)
1874 			spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1875 		pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1876 		if (pmd_present(pmd))
1877 			force_flush = true;
1878 		VM_BUG_ON(!pmd_none(*new_pmd));
1879 
1880 		if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1881 			pgtable_t pgtable;
1882 			pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1883 			pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1884 		}
1885 		pmd = move_soft_dirty_pmd(pmd);
1886 		set_pmd_at(mm, new_addr, new_pmd, pmd);
1887 		if (force_flush)
1888 			flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1889 		if (new_ptl != old_ptl)
1890 			spin_unlock(new_ptl);
1891 		spin_unlock(old_ptl);
1892 		return true;
1893 	}
1894 	return false;
1895 }
1896 
1897 /*
1898  * Returns
1899  *  - 0 if PMD could not be locked
1900  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1901  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1902  */
1903 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1904 		unsigned long addr, pgprot_t newprot, int prot_numa)
1905 {
1906 	struct mm_struct *mm = vma->vm_mm;
1907 	spinlock_t *ptl;
1908 	pmd_t entry;
1909 	bool preserve_write;
1910 	int ret;
1911 
1912 	ptl = __pmd_trans_huge_lock(pmd, vma);
1913 	if (!ptl)
1914 		return 0;
1915 
1916 	preserve_write = prot_numa && pmd_write(*pmd);
1917 	ret = 1;
1918 
1919 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1920 	if (is_swap_pmd(*pmd)) {
1921 		swp_entry_t entry = pmd_to_swp_entry(*pmd);
1922 
1923 		VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1924 		if (is_write_migration_entry(entry)) {
1925 			pmd_t newpmd;
1926 			/*
1927 			 * A protection check is difficult so
1928 			 * just be safe and disable write
1929 			 */
1930 			make_migration_entry_read(&entry);
1931 			newpmd = swp_entry_to_pmd(entry);
1932 			if (pmd_swp_soft_dirty(*pmd))
1933 				newpmd = pmd_swp_mksoft_dirty(newpmd);
1934 			set_pmd_at(mm, addr, pmd, newpmd);
1935 		}
1936 		goto unlock;
1937 	}
1938 #endif
1939 
1940 	/*
1941 	 * Avoid trapping faults against the zero page. The read-only
1942 	 * data is likely to be read-cached on the local CPU and
1943 	 * local/remote hits to the zero page are not interesting.
1944 	 */
1945 	if (prot_numa && is_huge_zero_pmd(*pmd))
1946 		goto unlock;
1947 
1948 	if (prot_numa && pmd_protnone(*pmd))
1949 		goto unlock;
1950 
1951 	/*
1952 	 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1953 	 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1954 	 * which is also under down_read(mmap_sem):
1955 	 *
1956 	 *	CPU0:				CPU1:
1957 	 *				change_huge_pmd(prot_numa=1)
1958 	 *				 pmdp_huge_get_and_clear_notify()
1959 	 * madvise_dontneed()
1960 	 *  zap_pmd_range()
1961 	 *   pmd_trans_huge(*pmd) == 0 (without ptl)
1962 	 *   // skip the pmd
1963 	 *				 set_pmd_at();
1964 	 *				 // pmd is re-established
1965 	 *
1966 	 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1967 	 * which may break userspace.
1968 	 *
1969 	 * pmdp_invalidate() is required to make sure we don't miss
1970 	 * dirty/young flags set by hardware.
1971 	 */
1972 	entry = pmdp_invalidate(vma, addr, pmd);
1973 
1974 	entry = pmd_modify(entry, newprot);
1975 	if (preserve_write)
1976 		entry = pmd_mk_savedwrite(entry);
1977 	ret = HPAGE_PMD_NR;
1978 	set_pmd_at(mm, addr, pmd, entry);
1979 	BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1980 unlock:
1981 	spin_unlock(ptl);
1982 	return ret;
1983 }
1984 
1985 /*
1986  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1987  *
1988  * Note that if it returns page table lock pointer, this routine returns without
1989  * unlocking page table lock. So callers must unlock it.
1990  */
1991 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1992 {
1993 	spinlock_t *ptl;
1994 	ptl = pmd_lock(vma->vm_mm, pmd);
1995 	if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1996 			pmd_devmap(*pmd)))
1997 		return ptl;
1998 	spin_unlock(ptl);
1999 	return NULL;
2000 }
2001 
2002 /*
2003  * Returns true if a given pud maps a thp, false otherwise.
2004  *
2005  * Note that if it returns true, this routine returns without unlocking page
2006  * table lock. So callers must unlock it.
2007  */
2008 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2009 {
2010 	spinlock_t *ptl;
2011 
2012 	ptl = pud_lock(vma->vm_mm, pud);
2013 	if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2014 		return ptl;
2015 	spin_unlock(ptl);
2016 	return NULL;
2017 }
2018 
2019 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2020 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2021 		 pud_t *pud, unsigned long addr)
2022 {
2023 	spinlock_t *ptl;
2024 
2025 	ptl = __pud_trans_huge_lock(pud, vma);
2026 	if (!ptl)
2027 		return 0;
2028 	/*
2029 	 * For architectures like ppc64 we look at deposited pgtable
2030 	 * when calling pudp_huge_get_and_clear. So do the
2031 	 * pgtable_trans_huge_withdraw after finishing pudp related
2032 	 * operations.
2033 	 */
2034 	pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
2035 	tlb_remove_pud_tlb_entry(tlb, pud, addr);
2036 	if (vma_is_dax(vma)) {
2037 		spin_unlock(ptl);
2038 		/* No zero page support yet */
2039 	} else {
2040 		/* No support for anonymous PUD pages yet */
2041 		BUG();
2042 	}
2043 	return 1;
2044 }
2045 
2046 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2047 		unsigned long haddr)
2048 {
2049 	VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2050 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2051 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2052 	VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2053 
2054 	count_vm_event(THP_SPLIT_PUD);
2055 
2056 	pudp_huge_clear_flush_notify(vma, haddr, pud);
2057 }
2058 
2059 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2060 		unsigned long address)
2061 {
2062 	spinlock_t *ptl;
2063 	struct mmu_notifier_range range;
2064 
2065 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2066 				address & HPAGE_PUD_MASK,
2067 				(address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2068 	mmu_notifier_invalidate_range_start(&range);
2069 	ptl = pud_lock(vma->vm_mm, pud);
2070 	if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2071 		goto out;
2072 	__split_huge_pud_locked(vma, pud, range.start);
2073 
2074 out:
2075 	spin_unlock(ptl);
2076 	/*
2077 	 * No need to double call mmu_notifier->invalidate_range() callback as
2078 	 * the above pudp_huge_clear_flush_notify() did already call it.
2079 	 */
2080 	mmu_notifier_invalidate_range_only_end(&range);
2081 }
2082 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2083 
2084 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2085 		unsigned long haddr, pmd_t *pmd)
2086 {
2087 	struct mm_struct *mm = vma->vm_mm;
2088 	pgtable_t pgtable;
2089 	pmd_t _pmd;
2090 	int i;
2091 
2092 	/*
2093 	 * Leave pmd empty until pte is filled note that it is fine to delay
2094 	 * notification until mmu_notifier_invalidate_range_end() as we are
2095 	 * replacing a zero pmd write protected page with a zero pte write
2096 	 * protected page.
2097 	 *
2098 	 * See Documentation/vm/mmu_notifier.rst
2099 	 */
2100 	pmdp_huge_clear_flush(vma, haddr, pmd);
2101 
2102 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2103 	pmd_populate(mm, &_pmd, pgtable);
2104 
2105 	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2106 		pte_t *pte, entry;
2107 		entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2108 		entry = pte_mkspecial(entry);
2109 		pte = pte_offset_map(&_pmd, haddr);
2110 		VM_BUG_ON(!pte_none(*pte));
2111 		set_pte_at(mm, haddr, pte, entry);
2112 		pte_unmap(pte);
2113 	}
2114 	smp_wmb(); /* make pte visible before pmd */
2115 	pmd_populate(mm, pmd, pgtable);
2116 }
2117 
2118 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2119 		unsigned long haddr, bool freeze)
2120 {
2121 	struct mm_struct *mm = vma->vm_mm;
2122 	struct page *page;
2123 	pgtable_t pgtable;
2124 	pmd_t old_pmd, _pmd;
2125 	bool young, write, soft_dirty, pmd_migration = false;
2126 	unsigned long addr;
2127 	int i;
2128 
2129 	VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2130 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2131 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2132 	VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2133 				&& !pmd_devmap(*pmd));
2134 
2135 	count_vm_event(THP_SPLIT_PMD);
2136 
2137 	if (!vma_is_anonymous(vma)) {
2138 		_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2139 		/*
2140 		 * We are going to unmap this huge page. So
2141 		 * just go ahead and zap it
2142 		 */
2143 		if (arch_needs_pgtable_deposit())
2144 			zap_deposited_table(mm, pmd);
2145 		if (vma_is_dax(vma))
2146 			return;
2147 		page = pmd_page(_pmd);
2148 		if (!PageDirty(page) && pmd_dirty(_pmd))
2149 			set_page_dirty(page);
2150 		if (!PageReferenced(page) && pmd_young(_pmd))
2151 			SetPageReferenced(page);
2152 		page_remove_rmap(page, true);
2153 		put_page(page);
2154 		add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2155 		return;
2156 	} else if (is_huge_zero_pmd(*pmd)) {
2157 		/*
2158 		 * FIXME: Do we want to invalidate secondary mmu by calling
2159 		 * mmu_notifier_invalidate_range() see comments below inside
2160 		 * __split_huge_pmd() ?
2161 		 *
2162 		 * We are going from a zero huge page write protected to zero
2163 		 * small page also write protected so it does not seems useful
2164 		 * to invalidate secondary mmu at this time.
2165 		 */
2166 		return __split_huge_zero_page_pmd(vma, haddr, pmd);
2167 	}
2168 
2169 	/*
2170 	 * Up to this point the pmd is present and huge and userland has the
2171 	 * whole access to the hugepage during the split (which happens in
2172 	 * place). If we overwrite the pmd with the not-huge version pointing
2173 	 * to the pte here (which of course we could if all CPUs were bug
2174 	 * free), userland could trigger a small page size TLB miss on the
2175 	 * small sized TLB while the hugepage TLB entry is still established in
2176 	 * the huge TLB. Some CPU doesn't like that.
2177 	 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2178 	 * 383 on page 93. Intel should be safe but is also warns that it's
2179 	 * only safe if the permission and cache attributes of the two entries
2180 	 * loaded in the two TLB is identical (which should be the case here).
2181 	 * But it is generally safer to never allow small and huge TLB entries
2182 	 * for the same virtual address to be loaded simultaneously. So instead
2183 	 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2184 	 * current pmd notpresent (atomically because here the pmd_trans_huge
2185 	 * must remain set at all times on the pmd until the split is complete
2186 	 * for this pmd), then we flush the SMP TLB and finally we write the
2187 	 * non-huge version of the pmd entry with pmd_populate.
2188 	 */
2189 	old_pmd = pmdp_invalidate(vma, haddr, pmd);
2190 
2191 	pmd_migration = is_pmd_migration_entry(old_pmd);
2192 	if (unlikely(pmd_migration)) {
2193 		swp_entry_t entry;
2194 
2195 		entry = pmd_to_swp_entry(old_pmd);
2196 		page = pfn_to_page(swp_offset(entry));
2197 		write = is_write_migration_entry(entry);
2198 		young = false;
2199 		soft_dirty = pmd_swp_soft_dirty(old_pmd);
2200 	} else {
2201 		page = pmd_page(old_pmd);
2202 		if (pmd_dirty(old_pmd))
2203 			SetPageDirty(page);
2204 		write = pmd_write(old_pmd);
2205 		young = pmd_young(old_pmd);
2206 		soft_dirty = pmd_soft_dirty(old_pmd);
2207 	}
2208 	VM_BUG_ON_PAGE(!page_count(page), page);
2209 	page_ref_add(page, HPAGE_PMD_NR - 1);
2210 
2211 	/*
2212 	 * Withdraw the table only after we mark the pmd entry invalid.
2213 	 * This's critical for some architectures (Power).
2214 	 */
2215 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2216 	pmd_populate(mm, &_pmd, pgtable);
2217 
2218 	for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2219 		pte_t entry, *pte;
2220 		/*
2221 		 * Note that NUMA hinting access restrictions are not
2222 		 * transferred to avoid any possibility of altering
2223 		 * permissions across VMAs.
2224 		 */
2225 		if (freeze || pmd_migration) {
2226 			swp_entry_t swp_entry;
2227 			swp_entry = make_migration_entry(page + i, write);
2228 			entry = swp_entry_to_pte(swp_entry);
2229 			if (soft_dirty)
2230 				entry = pte_swp_mksoft_dirty(entry);
2231 		} else {
2232 			entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2233 			entry = maybe_mkwrite(entry, vma);
2234 			if (!write)
2235 				entry = pte_wrprotect(entry);
2236 			if (!young)
2237 				entry = pte_mkold(entry);
2238 			if (soft_dirty)
2239 				entry = pte_mksoft_dirty(entry);
2240 		}
2241 		pte = pte_offset_map(&_pmd, addr);
2242 		BUG_ON(!pte_none(*pte));
2243 		set_pte_at(mm, addr, pte, entry);
2244 		atomic_inc(&page[i]._mapcount);
2245 		pte_unmap(pte);
2246 	}
2247 
2248 	/*
2249 	 * Set PG_double_map before dropping compound_mapcount to avoid
2250 	 * false-negative page_mapped().
2251 	 */
2252 	if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2253 		for (i = 0; i < HPAGE_PMD_NR; i++)
2254 			atomic_inc(&page[i]._mapcount);
2255 	}
2256 
2257 	if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2258 		/* Last compound_mapcount is gone. */
2259 		__dec_node_page_state(page, NR_ANON_THPS);
2260 		if (TestClearPageDoubleMap(page)) {
2261 			/* No need in mapcount reference anymore */
2262 			for (i = 0; i < HPAGE_PMD_NR; i++)
2263 				atomic_dec(&page[i]._mapcount);
2264 		}
2265 	}
2266 
2267 	smp_wmb(); /* make pte visible before pmd */
2268 	pmd_populate(mm, pmd, pgtable);
2269 
2270 	if (freeze) {
2271 		for (i = 0; i < HPAGE_PMD_NR; i++) {
2272 			page_remove_rmap(page + i, false);
2273 			put_page(page + i);
2274 		}
2275 	}
2276 }
2277 
2278 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2279 		unsigned long address, bool freeze, struct page *page)
2280 {
2281 	spinlock_t *ptl;
2282 	struct mmu_notifier_range range;
2283 
2284 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2285 				address & HPAGE_PMD_MASK,
2286 				(address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2287 	mmu_notifier_invalidate_range_start(&range);
2288 	ptl = pmd_lock(vma->vm_mm, pmd);
2289 
2290 	/*
2291 	 * If caller asks to setup a migration entries, we need a page to check
2292 	 * pmd against. Otherwise we can end up replacing wrong page.
2293 	 */
2294 	VM_BUG_ON(freeze && !page);
2295 	if (page && page != pmd_page(*pmd))
2296 	        goto out;
2297 
2298 	if (pmd_trans_huge(*pmd)) {
2299 		page = pmd_page(*pmd);
2300 		if (PageMlocked(page))
2301 			clear_page_mlock(page);
2302 	} else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2303 		goto out;
2304 	__split_huge_pmd_locked(vma, pmd, range.start, freeze);
2305 out:
2306 	spin_unlock(ptl);
2307 	/*
2308 	 * No need to double call mmu_notifier->invalidate_range() callback.
2309 	 * They are 3 cases to consider inside __split_huge_pmd_locked():
2310 	 *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2311 	 *  2) __split_huge_zero_page_pmd() read only zero page and any write
2312 	 *    fault will trigger a flush_notify before pointing to a new page
2313 	 *    (it is fine if the secondary mmu keeps pointing to the old zero
2314 	 *    page in the meantime)
2315 	 *  3) Split a huge pmd into pte pointing to the same page. No need
2316 	 *     to invalidate secondary tlb entry they are all still valid.
2317 	 *     any further changes to individual pte will notify. So no need
2318 	 *     to call mmu_notifier->invalidate_range()
2319 	 */
2320 	mmu_notifier_invalidate_range_only_end(&range);
2321 }
2322 
2323 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2324 		bool freeze, struct page *page)
2325 {
2326 	pgd_t *pgd;
2327 	p4d_t *p4d;
2328 	pud_t *pud;
2329 	pmd_t *pmd;
2330 
2331 	pgd = pgd_offset(vma->vm_mm, address);
2332 	if (!pgd_present(*pgd))
2333 		return;
2334 
2335 	p4d = p4d_offset(pgd, address);
2336 	if (!p4d_present(*p4d))
2337 		return;
2338 
2339 	pud = pud_offset(p4d, address);
2340 	if (!pud_present(*pud))
2341 		return;
2342 
2343 	pmd = pmd_offset(pud, address);
2344 
2345 	__split_huge_pmd(vma, pmd, address, freeze, page);
2346 }
2347 
2348 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2349 			     unsigned long start,
2350 			     unsigned long end,
2351 			     long adjust_next)
2352 {
2353 	/*
2354 	 * If the new start address isn't hpage aligned and it could
2355 	 * previously contain an hugepage: check if we need to split
2356 	 * an huge pmd.
2357 	 */
2358 	if (start & ~HPAGE_PMD_MASK &&
2359 	    (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2360 	    (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2361 		split_huge_pmd_address(vma, start, false, NULL);
2362 
2363 	/*
2364 	 * If the new end address isn't hpage aligned and it could
2365 	 * previously contain an hugepage: check if we need to split
2366 	 * an huge pmd.
2367 	 */
2368 	if (end & ~HPAGE_PMD_MASK &&
2369 	    (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2370 	    (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2371 		split_huge_pmd_address(vma, end, false, NULL);
2372 
2373 	/*
2374 	 * If we're also updating the vma->vm_next->vm_start, if the new
2375 	 * vm_next->vm_start isn't page aligned and it could previously
2376 	 * contain an hugepage: check if we need to split an huge pmd.
2377 	 */
2378 	if (adjust_next > 0) {
2379 		struct vm_area_struct *next = vma->vm_next;
2380 		unsigned long nstart = next->vm_start;
2381 		nstart += adjust_next << PAGE_SHIFT;
2382 		if (nstart & ~HPAGE_PMD_MASK &&
2383 		    (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2384 		    (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2385 			split_huge_pmd_address(next, nstart, false, NULL);
2386 	}
2387 }
2388 
2389 static void unmap_page(struct page *page)
2390 {
2391 	enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2392 		TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2393 	bool unmap_success;
2394 
2395 	VM_BUG_ON_PAGE(!PageHead(page), page);
2396 
2397 	if (PageAnon(page))
2398 		ttu_flags |= TTU_SPLIT_FREEZE;
2399 
2400 	unmap_success = try_to_unmap(page, ttu_flags);
2401 	VM_BUG_ON_PAGE(!unmap_success, page);
2402 }
2403 
2404 static void remap_page(struct page *page)
2405 {
2406 	int i;
2407 	if (PageTransHuge(page)) {
2408 		remove_migration_ptes(page, page, true);
2409 	} else {
2410 		for (i = 0; i < HPAGE_PMD_NR; i++)
2411 			remove_migration_ptes(page + i, page + i, true);
2412 	}
2413 }
2414 
2415 static void __split_huge_page_tail(struct page *head, int tail,
2416 		struct lruvec *lruvec, struct list_head *list)
2417 {
2418 	struct page *page_tail = head + tail;
2419 
2420 	VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2421 
2422 	/*
2423 	 * Clone page flags before unfreezing refcount.
2424 	 *
2425 	 * After successful get_page_unless_zero() might follow flags change,
2426 	 * for exmaple lock_page() which set PG_waiters.
2427 	 */
2428 	page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2429 	page_tail->flags |= (head->flags &
2430 			((1L << PG_referenced) |
2431 			 (1L << PG_swapbacked) |
2432 			 (1L << PG_swapcache) |
2433 			 (1L << PG_mlocked) |
2434 			 (1L << PG_uptodate) |
2435 			 (1L << PG_active) |
2436 			 (1L << PG_workingset) |
2437 			 (1L << PG_locked) |
2438 			 (1L << PG_unevictable) |
2439 			 (1L << PG_dirty)));
2440 
2441 	/* ->mapping in first tail page is compound_mapcount */
2442 	VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2443 			page_tail);
2444 	page_tail->mapping = head->mapping;
2445 	page_tail->index = head->index + tail;
2446 
2447 	/* Page flags must be visible before we make the page non-compound. */
2448 	smp_wmb();
2449 
2450 	/*
2451 	 * Clear PageTail before unfreezing page refcount.
2452 	 *
2453 	 * After successful get_page_unless_zero() might follow put_page()
2454 	 * which needs correct compound_head().
2455 	 */
2456 	clear_compound_head(page_tail);
2457 
2458 	/* Finally unfreeze refcount. Additional reference from page cache. */
2459 	page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2460 					  PageSwapCache(head)));
2461 
2462 	if (page_is_young(head))
2463 		set_page_young(page_tail);
2464 	if (page_is_idle(head))
2465 		set_page_idle(page_tail);
2466 
2467 	page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2468 
2469 	/*
2470 	 * always add to the tail because some iterators expect new
2471 	 * pages to show after the currently processed elements - e.g.
2472 	 * migrate_pages
2473 	 */
2474 	lru_add_page_tail(head, page_tail, lruvec, list);
2475 }
2476 
2477 static void __split_huge_page(struct page *page, struct list_head *list,
2478 		pgoff_t end, unsigned long flags)
2479 {
2480 	struct page *head = compound_head(page);
2481 	pg_data_t *pgdat = page_pgdat(head);
2482 	struct lruvec *lruvec;
2483 	int i;
2484 
2485 	lruvec = mem_cgroup_page_lruvec(head, pgdat);
2486 
2487 	/* complete memcg works before add pages to LRU */
2488 	mem_cgroup_split_huge_fixup(head);
2489 
2490 	for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2491 		__split_huge_page_tail(head, i, lruvec, list);
2492 		/* Some pages can be beyond i_size: drop them from page cache */
2493 		if (head[i].index >= end) {
2494 			ClearPageDirty(head + i);
2495 			__delete_from_page_cache(head + i, NULL);
2496 			if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2497 				shmem_uncharge(head->mapping->host, 1);
2498 			put_page(head + i);
2499 		} else if (!PageAnon(page)) {
2500 			__xa_store(&head->mapping->i_pages, head[i].index,
2501 					head + i, 0);
2502 		}
2503 	}
2504 
2505 	ClearPageCompound(head);
2506 	/* See comment in __split_huge_page_tail() */
2507 	if (PageAnon(head)) {
2508 		/* Additional pin to swap cache */
2509 		if (PageSwapCache(head))
2510 			page_ref_add(head, 2);
2511 		else
2512 			page_ref_inc(head);
2513 	} else {
2514 		/* Additional pin to page cache */
2515 		page_ref_add(head, 2);
2516 		xa_unlock(&head->mapping->i_pages);
2517 	}
2518 
2519 	spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2520 
2521 	remap_page(head);
2522 
2523 	for (i = 0; i < HPAGE_PMD_NR; i++) {
2524 		struct page *subpage = head + i;
2525 		if (subpage == page)
2526 			continue;
2527 		unlock_page(subpage);
2528 
2529 		/*
2530 		 * Subpages may be freed if there wasn't any mapping
2531 		 * like if add_to_swap() is running on a lru page that
2532 		 * had its mapping zapped. And freeing these pages
2533 		 * requires taking the lru_lock so we do the put_page
2534 		 * of the tail pages after the split is complete.
2535 		 */
2536 		put_page(subpage);
2537 	}
2538 }
2539 
2540 int total_mapcount(struct page *page)
2541 {
2542 	int i, compound, ret;
2543 
2544 	VM_BUG_ON_PAGE(PageTail(page), page);
2545 
2546 	if (likely(!PageCompound(page)))
2547 		return atomic_read(&page->_mapcount) + 1;
2548 
2549 	compound = compound_mapcount(page);
2550 	if (PageHuge(page))
2551 		return compound;
2552 	ret = compound;
2553 	for (i = 0; i < HPAGE_PMD_NR; i++)
2554 		ret += atomic_read(&page[i]._mapcount) + 1;
2555 	/* File pages has compound_mapcount included in _mapcount */
2556 	if (!PageAnon(page))
2557 		return ret - compound * HPAGE_PMD_NR;
2558 	if (PageDoubleMap(page))
2559 		ret -= HPAGE_PMD_NR;
2560 	return ret;
2561 }
2562 
2563 /*
2564  * This calculates accurately how many mappings a transparent hugepage
2565  * has (unlike page_mapcount() which isn't fully accurate). This full
2566  * accuracy is primarily needed to know if copy-on-write faults can
2567  * reuse the page and change the mapping to read-write instead of
2568  * copying them. At the same time this returns the total_mapcount too.
2569  *
2570  * The function returns the highest mapcount any one of the subpages
2571  * has. If the return value is one, even if different processes are
2572  * mapping different subpages of the transparent hugepage, they can
2573  * all reuse it, because each process is reusing a different subpage.
2574  *
2575  * The total_mapcount is instead counting all virtual mappings of the
2576  * subpages. If the total_mapcount is equal to "one", it tells the
2577  * caller all mappings belong to the same "mm" and in turn the
2578  * anon_vma of the transparent hugepage can become the vma->anon_vma
2579  * local one as no other process may be mapping any of the subpages.
2580  *
2581  * It would be more accurate to replace page_mapcount() with
2582  * page_trans_huge_mapcount(), however we only use
2583  * page_trans_huge_mapcount() in the copy-on-write faults where we
2584  * need full accuracy to avoid breaking page pinning, because
2585  * page_trans_huge_mapcount() is slower than page_mapcount().
2586  */
2587 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2588 {
2589 	int i, ret, _total_mapcount, mapcount;
2590 
2591 	/* hugetlbfs shouldn't call it */
2592 	VM_BUG_ON_PAGE(PageHuge(page), page);
2593 
2594 	if (likely(!PageTransCompound(page))) {
2595 		mapcount = atomic_read(&page->_mapcount) + 1;
2596 		if (total_mapcount)
2597 			*total_mapcount = mapcount;
2598 		return mapcount;
2599 	}
2600 
2601 	page = compound_head(page);
2602 
2603 	_total_mapcount = ret = 0;
2604 	for (i = 0; i < HPAGE_PMD_NR; i++) {
2605 		mapcount = atomic_read(&page[i]._mapcount) + 1;
2606 		ret = max(ret, mapcount);
2607 		_total_mapcount += mapcount;
2608 	}
2609 	if (PageDoubleMap(page)) {
2610 		ret -= 1;
2611 		_total_mapcount -= HPAGE_PMD_NR;
2612 	}
2613 	mapcount = compound_mapcount(page);
2614 	ret += mapcount;
2615 	_total_mapcount += mapcount;
2616 	if (total_mapcount)
2617 		*total_mapcount = _total_mapcount;
2618 	return ret;
2619 }
2620 
2621 /* Racy check whether the huge page can be split */
2622 bool can_split_huge_page(struct page *page, int *pextra_pins)
2623 {
2624 	int extra_pins;
2625 
2626 	/* Additional pins from page cache */
2627 	if (PageAnon(page))
2628 		extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2629 	else
2630 		extra_pins = HPAGE_PMD_NR;
2631 	if (pextra_pins)
2632 		*pextra_pins = extra_pins;
2633 	return total_mapcount(page) == page_count(page) - extra_pins - 1;
2634 }
2635 
2636 /*
2637  * This function splits huge page into normal pages. @page can point to any
2638  * subpage of huge page to split. Split doesn't change the position of @page.
2639  *
2640  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2641  * The huge page must be locked.
2642  *
2643  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2644  *
2645  * Both head page and tail pages will inherit mapping, flags, and so on from
2646  * the hugepage.
2647  *
2648  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2649  * they are not mapped.
2650  *
2651  * Returns 0 if the hugepage is split successfully.
2652  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2653  * us.
2654  */
2655 int split_huge_page_to_list(struct page *page, struct list_head *list)
2656 {
2657 	struct page *head = compound_head(page);
2658 	struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2659 	struct anon_vma *anon_vma = NULL;
2660 	struct address_space *mapping = NULL;
2661 	int count, mapcount, extra_pins, ret;
2662 	bool mlocked;
2663 	unsigned long flags;
2664 	pgoff_t end;
2665 
2666 	VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2667 	VM_BUG_ON_PAGE(!PageLocked(page), page);
2668 	VM_BUG_ON_PAGE(!PageCompound(page), page);
2669 
2670 	if (PageWriteback(page))
2671 		return -EBUSY;
2672 
2673 	if (PageAnon(head)) {
2674 		/*
2675 		 * The caller does not necessarily hold an mmap_sem that would
2676 		 * prevent the anon_vma disappearing so we first we take a
2677 		 * reference to it and then lock the anon_vma for write. This
2678 		 * is similar to page_lock_anon_vma_read except the write lock
2679 		 * is taken to serialise against parallel split or collapse
2680 		 * operations.
2681 		 */
2682 		anon_vma = page_get_anon_vma(head);
2683 		if (!anon_vma) {
2684 			ret = -EBUSY;
2685 			goto out;
2686 		}
2687 		end = -1;
2688 		mapping = NULL;
2689 		anon_vma_lock_write(anon_vma);
2690 	} else {
2691 		mapping = head->mapping;
2692 
2693 		/* Truncated ? */
2694 		if (!mapping) {
2695 			ret = -EBUSY;
2696 			goto out;
2697 		}
2698 
2699 		anon_vma = NULL;
2700 		i_mmap_lock_read(mapping);
2701 
2702 		/*
2703 		 *__split_huge_page() may need to trim off pages beyond EOF:
2704 		 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2705 		 * which cannot be nested inside the page tree lock. So note
2706 		 * end now: i_size itself may be changed at any moment, but
2707 		 * head page lock is good enough to serialize the trimming.
2708 		 */
2709 		end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2710 	}
2711 
2712 	/*
2713 	 * Racy check if we can split the page, before unmap_page() will
2714 	 * split PMDs
2715 	 */
2716 	if (!can_split_huge_page(head, &extra_pins)) {
2717 		ret = -EBUSY;
2718 		goto out_unlock;
2719 	}
2720 
2721 	mlocked = PageMlocked(page);
2722 	unmap_page(head);
2723 	VM_BUG_ON_PAGE(compound_mapcount(head), head);
2724 
2725 	/* Make sure the page is not on per-CPU pagevec as it takes pin */
2726 	if (mlocked)
2727 		lru_add_drain();
2728 
2729 	/* prevent PageLRU to go away from under us, and freeze lru stats */
2730 	spin_lock_irqsave(&pgdata->lru_lock, flags);
2731 
2732 	if (mapping) {
2733 		XA_STATE(xas, &mapping->i_pages, page_index(head));
2734 
2735 		/*
2736 		 * Check if the head page is present in page cache.
2737 		 * We assume all tail are present too, if head is there.
2738 		 */
2739 		xa_lock(&mapping->i_pages);
2740 		if (xas_load(&xas) != head)
2741 			goto fail;
2742 	}
2743 
2744 	/* Prevent deferred_split_scan() touching ->_refcount */
2745 	spin_lock(&pgdata->split_queue_lock);
2746 	count = page_count(head);
2747 	mapcount = total_mapcount(head);
2748 	if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2749 		if (!list_empty(page_deferred_list(head))) {
2750 			pgdata->split_queue_len--;
2751 			list_del(page_deferred_list(head));
2752 		}
2753 		if (mapping)
2754 			__dec_node_page_state(page, NR_SHMEM_THPS);
2755 		spin_unlock(&pgdata->split_queue_lock);
2756 		__split_huge_page(page, list, end, flags);
2757 		if (PageSwapCache(head)) {
2758 			swp_entry_t entry = { .val = page_private(head) };
2759 
2760 			ret = split_swap_cluster(entry);
2761 		} else
2762 			ret = 0;
2763 	} else {
2764 		if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2765 			pr_alert("total_mapcount: %u, page_count(): %u\n",
2766 					mapcount, count);
2767 			if (PageTail(page))
2768 				dump_page(head, NULL);
2769 			dump_page(page, "total_mapcount(head) > 0");
2770 			BUG();
2771 		}
2772 		spin_unlock(&pgdata->split_queue_lock);
2773 fail:		if (mapping)
2774 			xa_unlock(&mapping->i_pages);
2775 		spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2776 		remap_page(head);
2777 		ret = -EBUSY;
2778 	}
2779 
2780 out_unlock:
2781 	if (anon_vma) {
2782 		anon_vma_unlock_write(anon_vma);
2783 		put_anon_vma(anon_vma);
2784 	}
2785 	if (mapping)
2786 		i_mmap_unlock_read(mapping);
2787 out:
2788 	count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2789 	return ret;
2790 }
2791 
2792 void free_transhuge_page(struct page *page)
2793 {
2794 	struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2795 	unsigned long flags;
2796 
2797 	spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2798 	if (!list_empty(page_deferred_list(page))) {
2799 		pgdata->split_queue_len--;
2800 		list_del(page_deferred_list(page));
2801 	}
2802 	spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2803 	free_compound_page(page);
2804 }
2805 
2806 void deferred_split_huge_page(struct page *page)
2807 {
2808 	struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2809 	unsigned long flags;
2810 
2811 	VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2812 
2813 	spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2814 	if (list_empty(page_deferred_list(page))) {
2815 		count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2816 		list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2817 		pgdata->split_queue_len++;
2818 	}
2819 	spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2820 }
2821 
2822 static unsigned long deferred_split_count(struct shrinker *shrink,
2823 		struct shrink_control *sc)
2824 {
2825 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2826 	return READ_ONCE(pgdata->split_queue_len);
2827 }
2828 
2829 static unsigned long deferred_split_scan(struct shrinker *shrink,
2830 		struct shrink_control *sc)
2831 {
2832 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2833 	unsigned long flags;
2834 	LIST_HEAD(list), *pos, *next;
2835 	struct page *page;
2836 	int split = 0;
2837 
2838 	spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2839 	/* Take pin on all head pages to avoid freeing them under us */
2840 	list_for_each_safe(pos, next, &pgdata->split_queue) {
2841 		page = list_entry((void *)pos, struct page, mapping);
2842 		page = compound_head(page);
2843 		if (get_page_unless_zero(page)) {
2844 			list_move(page_deferred_list(page), &list);
2845 		} else {
2846 			/* We lost race with put_compound_page() */
2847 			list_del_init(page_deferred_list(page));
2848 			pgdata->split_queue_len--;
2849 		}
2850 		if (!--sc->nr_to_scan)
2851 			break;
2852 	}
2853 	spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2854 
2855 	list_for_each_safe(pos, next, &list) {
2856 		page = list_entry((void *)pos, struct page, mapping);
2857 		if (!trylock_page(page))
2858 			goto next;
2859 		/* split_huge_page() removes page from list on success */
2860 		if (!split_huge_page(page))
2861 			split++;
2862 		unlock_page(page);
2863 next:
2864 		put_page(page);
2865 	}
2866 
2867 	spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2868 	list_splice_tail(&list, &pgdata->split_queue);
2869 	spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2870 
2871 	/*
2872 	 * Stop shrinker if we didn't split any page, but the queue is empty.
2873 	 * This can happen if pages were freed under us.
2874 	 */
2875 	if (!split && list_empty(&pgdata->split_queue))
2876 		return SHRINK_STOP;
2877 	return split;
2878 }
2879 
2880 static struct shrinker deferred_split_shrinker = {
2881 	.count_objects = deferred_split_count,
2882 	.scan_objects = deferred_split_scan,
2883 	.seeks = DEFAULT_SEEKS,
2884 	.flags = SHRINKER_NUMA_AWARE,
2885 };
2886 
2887 #ifdef CONFIG_DEBUG_FS
2888 static int split_huge_pages_set(void *data, u64 val)
2889 {
2890 	struct zone *zone;
2891 	struct page *page;
2892 	unsigned long pfn, max_zone_pfn;
2893 	unsigned long total = 0, split = 0;
2894 
2895 	if (val != 1)
2896 		return -EINVAL;
2897 
2898 	for_each_populated_zone(zone) {
2899 		max_zone_pfn = zone_end_pfn(zone);
2900 		for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2901 			if (!pfn_valid(pfn))
2902 				continue;
2903 
2904 			page = pfn_to_page(pfn);
2905 			if (!get_page_unless_zero(page))
2906 				continue;
2907 
2908 			if (zone != page_zone(page))
2909 				goto next;
2910 
2911 			if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2912 				goto next;
2913 
2914 			total++;
2915 			lock_page(page);
2916 			if (!split_huge_page(page))
2917 				split++;
2918 			unlock_page(page);
2919 next:
2920 			put_page(page);
2921 		}
2922 	}
2923 
2924 	pr_info("%lu of %lu THP split\n", split, total);
2925 
2926 	return 0;
2927 }
2928 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2929 		"%llu\n");
2930 
2931 static int __init split_huge_pages_debugfs(void)
2932 {
2933 	debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2934 			    &split_huge_pages_fops);
2935 	return 0;
2936 }
2937 late_initcall(split_huge_pages_debugfs);
2938 #endif
2939 
2940 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2941 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2942 		struct page *page)
2943 {
2944 	struct vm_area_struct *vma = pvmw->vma;
2945 	struct mm_struct *mm = vma->vm_mm;
2946 	unsigned long address = pvmw->address;
2947 	pmd_t pmdval;
2948 	swp_entry_t entry;
2949 	pmd_t pmdswp;
2950 
2951 	if (!(pvmw->pmd && !pvmw->pte))
2952 		return;
2953 
2954 	flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2955 	pmdval = *pvmw->pmd;
2956 	pmdp_invalidate(vma, address, pvmw->pmd);
2957 	if (pmd_dirty(pmdval))
2958 		set_page_dirty(page);
2959 	entry = make_migration_entry(page, pmd_write(pmdval));
2960 	pmdswp = swp_entry_to_pmd(entry);
2961 	if (pmd_soft_dirty(pmdval))
2962 		pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2963 	set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2964 	page_remove_rmap(page, true);
2965 	put_page(page);
2966 }
2967 
2968 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2969 {
2970 	struct vm_area_struct *vma = pvmw->vma;
2971 	struct mm_struct *mm = vma->vm_mm;
2972 	unsigned long address = pvmw->address;
2973 	unsigned long mmun_start = address & HPAGE_PMD_MASK;
2974 	pmd_t pmde;
2975 	swp_entry_t entry;
2976 
2977 	if (!(pvmw->pmd && !pvmw->pte))
2978 		return;
2979 
2980 	entry = pmd_to_swp_entry(*pvmw->pmd);
2981 	get_page(new);
2982 	pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2983 	if (pmd_swp_soft_dirty(*pvmw->pmd))
2984 		pmde = pmd_mksoft_dirty(pmde);
2985 	if (is_write_migration_entry(entry))
2986 		pmde = maybe_pmd_mkwrite(pmde, vma);
2987 
2988 	flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2989 	if (PageAnon(new))
2990 		page_add_anon_rmap(new, vma, mmun_start, true);
2991 	else
2992 		page_add_file_rmap(new, true);
2993 	set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2994 	if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
2995 		mlock_vma_page(new);
2996 	update_mmu_cache_pmd(vma, address, pvmw->pmd);
2997 }
2998 #endif
2999