xref: /linux/mm/huge_memory.c (revision eb057b44dbe35ae14527830236a92f51de8f9184)
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/mm.h>
11 #include <linux/sched/coredump.h>
12 #include <linux/sched/numa_balancing.h>
13 #include <linux/highmem.h>
14 #include <linux/hugetlb.h>
15 #include <linux/mmu_notifier.h>
16 #include <linux/rmap.h>
17 #include <linux/swap.h>
18 #include <linux/shrinker.h>
19 #include <linux/mm_inline.h>
20 #include <linux/swapops.h>
21 #include <linux/dax.h>
22 #include <linux/khugepaged.h>
23 #include <linux/freezer.h>
24 #include <linux/pfn_t.h>
25 #include <linux/mman.h>
26 #include <linux/memremap.h>
27 #include <linux/pagemap.h>
28 #include <linux/debugfs.h>
29 #include <linux/migrate.h>
30 #include <linux/hashtable.h>
31 #include <linux/userfaultfd_k.h>
32 #include <linux/page_idle.h>
33 #include <linux/shmem_fs.h>
34 #include <linux/oom.h>
35 #include <linux/numa.h>
36 #include <linux/page_owner.h>
37 
38 #include <asm/tlb.h>
39 #include <asm/pgalloc.h>
40 #include "internal.h"
41 
42 /*
43  * By default, transparent hugepage support is disabled in order to avoid
44  * risking an increased memory footprint for applications that are not
45  * guaranteed to benefit from it. When transparent hugepage support is
46  * enabled, it is for all mappings, and khugepaged scans all mappings.
47  * Defrag is invoked by khugepaged hugepage allocations and by page faults
48  * for all hugepage allocations.
49  */
50 unsigned long transparent_hugepage_flags __read_mostly =
51 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
52 	(1<<TRANSPARENT_HUGEPAGE_FLAG)|
53 #endif
54 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
55 	(1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
56 #endif
57 	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
58 	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
59 	(1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
60 
61 static struct shrinker deferred_split_shrinker;
62 
63 static atomic_t huge_zero_refcount;
64 struct page *huge_zero_page __read_mostly;
65 unsigned long huge_zero_pfn __read_mostly = ~0UL;
66 
67 static inline bool file_thp_enabled(struct vm_area_struct *vma)
68 {
69 	return transhuge_vma_enabled(vma, vma->vm_flags) && vma->vm_file &&
70 	       !inode_is_open_for_write(vma->vm_file->f_inode) &&
71 	       (vma->vm_flags & VM_EXEC);
72 }
73 
74 bool transparent_hugepage_active(struct vm_area_struct *vma)
75 {
76 	/* The addr is used to check if the vma size fits */
77 	unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
78 
79 	if (!transhuge_vma_suitable(vma, addr))
80 		return false;
81 	if (vma_is_anonymous(vma))
82 		return __transparent_hugepage_enabled(vma);
83 	if (vma_is_shmem(vma))
84 		return shmem_huge_enabled(vma);
85 	if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS))
86 		return file_thp_enabled(vma);
87 
88 	return false;
89 }
90 
91 static bool get_huge_zero_page(void)
92 {
93 	struct page *zero_page;
94 retry:
95 	if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
96 		return true;
97 
98 	zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
99 			HPAGE_PMD_ORDER);
100 	if (!zero_page) {
101 		count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
102 		return false;
103 	}
104 	count_vm_event(THP_ZERO_PAGE_ALLOC);
105 	preempt_disable();
106 	if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
107 		preempt_enable();
108 		__free_pages(zero_page, compound_order(zero_page));
109 		goto retry;
110 	}
111 	WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page));
112 
113 	/* We take additional reference here. It will be put back by shrinker */
114 	atomic_set(&huge_zero_refcount, 2);
115 	preempt_enable();
116 	return true;
117 }
118 
119 static void put_huge_zero_page(void)
120 {
121 	/*
122 	 * Counter should never go to zero here. Only shrinker can put
123 	 * last reference.
124 	 */
125 	BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
126 }
127 
128 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
129 {
130 	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
131 		return READ_ONCE(huge_zero_page);
132 
133 	if (!get_huge_zero_page())
134 		return NULL;
135 
136 	if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
137 		put_huge_zero_page();
138 
139 	return READ_ONCE(huge_zero_page);
140 }
141 
142 void mm_put_huge_zero_page(struct mm_struct *mm)
143 {
144 	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
145 		put_huge_zero_page();
146 }
147 
148 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
149 					struct shrink_control *sc)
150 {
151 	/* we can free zero page only if last reference remains */
152 	return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
153 }
154 
155 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
156 				       struct shrink_control *sc)
157 {
158 	if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
159 		struct page *zero_page = xchg(&huge_zero_page, NULL);
160 		BUG_ON(zero_page == NULL);
161 		WRITE_ONCE(huge_zero_pfn, ~0UL);
162 		__free_pages(zero_page, compound_order(zero_page));
163 		return HPAGE_PMD_NR;
164 	}
165 
166 	return 0;
167 }
168 
169 static struct shrinker huge_zero_page_shrinker = {
170 	.count_objects = shrink_huge_zero_page_count,
171 	.scan_objects = shrink_huge_zero_page_scan,
172 	.seeks = DEFAULT_SEEKS,
173 };
174 
175 #ifdef CONFIG_SYSFS
176 static ssize_t enabled_show(struct kobject *kobj,
177 			    struct kobj_attribute *attr, char *buf)
178 {
179 	const char *output;
180 
181 	if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
182 		output = "[always] madvise never";
183 	else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
184 			  &transparent_hugepage_flags))
185 		output = "always [madvise] never";
186 	else
187 		output = "always madvise [never]";
188 
189 	return sysfs_emit(buf, "%s\n", output);
190 }
191 
192 static ssize_t enabled_store(struct kobject *kobj,
193 			     struct kobj_attribute *attr,
194 			     const char *buf, size_t count)
195 {
196 	ssize_t ret = count;
197 
198 	if (sysfs_streq(buf, "always")) {
199 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
200 		set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
201 	} else if (sysfs_streq(buf, "madvise")) {
202 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
203 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
204 	} else if (sysfs_streq(buf, "never")) {
205 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
206 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
207 	} else
208 		ret = -EINVAL;
209 
210 	if (ret > 0) {
211 		int err = start_stop_khugepaged();
212 		if (err)
213 			ret = err;
214 	}
215 	return ret;
216 }
217 static struct kobj_attribute enabled_attr =
218 	__ATTR(enabled, 0644, enabled_show, enabled_store);
219 
220 ssize_t single_hugepage_flag_show(struct kobject *kobj,
221 				  struct kobj_attribute *attr, char *buf,
222 				  enum transparent_hugepage_flag flag)
223 {
224 	return sysfs_emit(buf, "%d\n",
225 			  !!test_bit(flag, &transparent_hugepage_flags));
226 }
227 
228 ssize_t single_hugepage_flag_store(struct kobject *kobj,
229 				 struct kobj_attribute *attr,
230 				 const char *buf, size_t count,
231 				 enum transparent_hugepage_flag flag)
232 {
233 	unsigned long value;
234 	int ret;
235 
236 	ret = kstrtoul(buf, 10, &value);
237 	if (ret < 0)
238 		return ret;
239 	if (value > 1)
240 		return -EINVAL;
241 
242 	if (value)
243 		set_bit(flag, &transparent_hugepage_flags);
244 	else
245 		clear_bit(flag, &transparent_hugepage_flags);
246 
247 	return count;
248 }
249 
250 static ssize_t defrag_show(struct kobject *kobj,
251 			   struct kobj_attribute *attr, char *buf)
252 {
253 	const char *output;
254 
255 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
256 		     &transparent_hugepage_flags))
257 		output = "[always] defer defer+madvise madvise never";
258 	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
259 			  &transparent_hugepage_flags))
260 		output = "always [defer] defer+madvise madvise never";
261 	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
262 			  &transparent_hugepage_flags))
263 		output = "always defer [defer+madvise] madvise never";
264 	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
265 			  &transparent_hugepage_flags))
266 		output = "always defer defer+madvise [madvise] never";
267 	else
268 		output = "always defer defer+madvise madvise [never]";
269 
270 	return sysfs_emit(buf, "%s\n", output);
271 }
272 
273 static ssize_t defrag_store(struct kobject *kobj,
274 			    struct kobj_attribute *attr,
275 			    const char *buf, size_t count)
276 {
277 	if (sysfs_streq(buf, "always")) {
278 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
279 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
280 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
281 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
282 	} else if (sysfs_streq(buf, "defer+madvise")) {
283 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
284 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
285 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
286 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
287 	} else if (sysfs_streq(buf, "defer")) {
288 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
289 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
290 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
291 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
292 	} else if (sysfs_streq(buf, "madvise")) {
293 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
294 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
295 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
296 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
297 	} else if (sysfs_streq(buf, "never")) {
298 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
299 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
300 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
301 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
302 	} else
303 		return -EINVAL;
304 
305 	return count;
306 }
307 static struct kobj_attribute defrag_attr =
308 	__ATTR(defrag, 0644, defrag_show, defrag_store);
309 
310 static ssize_t use_zero_page_show(struct kobject *kobj,
311 				  struct kobj_attribute *attr, char *buf)
312 {
313 	return single_hugepage_flag_show(kobj, attr, buf,
314 					 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
315 }
316 static ssize_t use_zero_page_store(struct kobject *kobj,
317 		struct kobj_attribute *attr, const char *buf, size_t count)
318 {
319 	return single_hugepage_flag_store(kobj, attr, buf, count,
320 				 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
321 }
322 static struct kobj_attribute use_zero_page_attr =
323 	__ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
324 
325 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
326 				   struct kobj_attribute *attr, char *buf)
327 {
328 	return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
329 }
330 static struct kobj_attribute hpage_pmd_size_attr =
331 	__ATTR_RO(hpage_pmd_size);
332 
333 static struct attribute *hugepage_attr[] = {
334 	&enabled_attr.attr,
335 	&defrag_attr.attr,
336 	&use_zero_page_attr.attr,
337 	&hpage_pmd_size_attr.attr,
338 #ifdef CONFIG_SHMEM
339 	&shmem_enabled_attr.attr,
340 #endif
341 	NULL,
342 };
343 
344 static const struct attribute_group hugepage_attr_group = {
345 	.attrs = hugepage_attr,
346 };
347 
348 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
349 {
350 	int err;
351 
352 	*hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
353 	if (unlikely(!*hugepage_kobj)) {
354 		pr_err("failed to create transparent hugepage kobject\n");
355 		return -ENOMEM;
356 	}
357 
358 	err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
359 	if (err) {
360 		pr_err("failed to register transparent hugepage group\n");
361 		goto delete_obj;
362 	}
363 
364 	err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
365 	if (err) {
366 		pr_err("failed to register transparent hugepage group\n");
367 		goto remove_hp_group;
368 	}
369 
370 	return 0;
371 
372 remove_hp_group:
373 	sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
374 delete_obj:
375 	kobject_put(*hugepage_kobj);
376 	return err;
377 }
378 
379 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
380 {
381 	sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
382 	sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
383 	kobject_put(hugepage_kobj);
384 }
385 #else
386 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
387 {
388 	return 0;
389 }
390 
391 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
392 {
393 }
394 #endif /* CONFIG_SYSFS */
395 
396 static int __init hugepage_init(void)
397 {
398 	int err;
399 	struct kobject *hugepage_kobj;
400 
401 	if (!has_transparent_hugepage()) {
402 		/*
403 		 * Hardware doesn't support hugepages, hence disable
404 		 * DAX PMD support.
405 		 */
406 		transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX;
407 		return -EINVAL;
408 	}
409 
410 	/*
411 	 * hugepages can't be allocated by the buddy allocator
412 	 */
413 	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
414 	/*
415 	 * we use page->mapping and page->index in second tail page
416 	 * as list_head: assuming THP order >= 2
417 	 */
418 	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
419 
420 	err = hugepage_init_sysfs(&hugepage_kobj);
421 	if (err)
422 		goto err_sysfs;
423 
424 	err = khugepaged_init();
425 	if (err)
426 		goto err_slab;
427 
428 	err = register_shrinker(&huge_zero_page_shrinker);
429 	if (err)
430 		goto err_hzp_shrinker;
431 	err = register_shrinker(&deferred_split_shrinker);
432 	if (err)
433 		goto err_split_shrinker;
434 
435 	/*
436 	 * By default disable transparent hugepages on smaller systems,
437 	 * where the extra memory used could hurt more than TLB overhead
438 	 * is likely to save.  The admin can still enable it through /sys.
439 	 */
440 	if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
441 		transparent_hugepage_flags = 0;
442 		return 0;
443 	}
444 
445 	err = start_stop_khugepaged();
446 	if (err)
447 		goto err_khugepaged;
448 
449 	return 0;
450 err_khugepaged:
451 	unregister_shrinker(&deferred_split_shrinker);
452 err_split_shrinker:
453 	unregister_shrinker(&huge_zero_page_shrinker);
454 err_hzp_shrinker:
455 	khugepaged_destroy();
456 err_slab:
457 	hugepage_exit_sysfs(hugepage_kobj);
458 err_sysfs:
459 	return err;
460 }
461 subsys_initcall(hugepage_init);
462 
463 static int __init setup_transparent_hugepage(char *str)
464 {
465 	int ret = 0;
466 	if (!str)
467 		goto out;
468 	if (!strcmp(str, "always")) {
469 		set_bit(TRANSPARENT_HUGEPAGE_FLAG,
470 			&transparent_hugepage_flags);
471 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
472 			  &transparent_hugepage_flags);
473 		ret = 1;
474 	} else if (!strcmp(str, "madvise")) {
475 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
476 			  &transparent_hugepage_flags);
477 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
478 			&transparent_hugepage_flags);
479 		ret = 1;
480 	} else if (!strcmp(str, "never")) {
481 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
482 			  &transparent_hugepage_flags);
483 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
484 			  &transparent_hugepage_flags);
485 		ret = 1;
486 	}
487 out:
488 	if (!ret)
489 		pr_warn("transparent_hugepage= cannot parse, ignored\n");
490 	return ret;
491 }
492 __setup("transparent_hugepage=", setup_transparent_hugepage);
493 
494 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
495 {
496 	if (likely(vma->vm_flags & VM_WRITE))
497 		pmd = pmd_mkwrite(pmd);
498 	return pmd;
499 }
500 
501 #ifdef CONFIG_MEMCG
502 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
503 {
504 	struct mem_cgroup *memcg = page_memcg(compound_head(page));
505 	struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
506 
507 	if (memcg)
508 		return &memcg->deferred_split_queue;
509 	else
510 		return &pgdat->deferred_split_queue;
511 }
512 #else
513 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
514 {
515 	struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
516 
517 	return &pgdat->deferred_split_queue;
518 }
519 #endif
520 
521 void prep_transhuge_page(struct page *page)
522 {
523 	/*
524 	 * we use page->mapping and page->indexlru in second tail page
525 	 * as list_head: assuming THP order >= 2
526 	 */
527 
528 	INIT_LIST_HEAD(page_deferred_list(page));
529 	set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
530 }
531 
532 bool is_transparent_hugepage(struct page *page)
533 {
534 	if (!PageCompound(page))
535 		return false;
536 
537 	page = compound_head(page);
538 	return is_huge_zero_page(page) ||
539 	       page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
540 }
541 EXPORT_SYMBOL_GPL(is_transparent_hugepage);
542 
543 static unsigned long __thp_get_unmapped_area(struct file *filp,
544 		unsigned long addr, unsigned long len,
545 		loff_t off, unsigned long flags, unsigned long size)
546 {
547 	loff_t off_end = off + len;
548 	loff_t off_align = round_up(off, size);
549 	unsigned long len_pad, ret;
550 
551 	if (off_end <= off_align || (off_end - off_align) < size)
552 		return 0;
553 
554 	len_pad = len + size;
555 	if (len_pad < len || (off + len_pad) < off)
556 		return 0;
557 
558 	ret = current->mm->get_unmapped_area(filp, addr, len_pad,
559 					      off >> PAGE_SHIFT, flags);
560 
561 	/*
562 	 * The failure might be due to length padding. The caller will retry
563 	 * without the padding.
564 	 */
565 	if (IS_ERR_VALUE(ret))
566 		return 0;
567 
568 	/*
569 	 * Do not try to align to THP boundary if allocation at the address
570 	 * hint succeeds.
571 	 */
572 	if (ret == addr)
573 		return addr;
574 
575 	ret += (off - ret) & (size - 1);
576 	return ret;
577 }
578 
579 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
580 		unsigned long len, unsigned long pgoff, unsigned long flags)
581 {
582 	unsigned long ret;
583 	loff_t off = (loff_t)pgoff << PAGE_SHIFT;
584 
585 	if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
586 		goto out;
587 
588 	ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
589 	if (ret)
590 		return ret;
591 out:
592 	return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
593 }
594 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
595 
596 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
597 			struct page *page, gfp_t gfp)
598 {
599 	struct vm_area_struct *vma = vmf->vma;
600 	pgtable_t pgtable;
601 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
602 	vm_fault_t ret = 0;
603 
604 	VM_BUG_ON_PAGE(!PageCompound(page), page);
605 
606 	if (mem_cgroup_charge(page_folio(page), vma->vm_mm, gfp)) {
607 		put_page(page);
608 		count_vm_event(THP_FAULT_FALLBACK);
609 		count_vm_event(THP_FAULT_FALLBACK_CHARGE);
610 		return VM_FAULT_FALLBACK;
611 	}
612 	cgroup_throttle_swaprate(page, gfp);
613 
614 	pgtable = pte_alloc_one(vma->vm_mm);
615 	if (unlikely(!pgtable)) {
616 		ret = VM_FAULT_OOM;
617 		goto release;
618 	}
619 
620 	clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
621 	/*
622 	 * The memory barrier inside __SetPageUptodate makes sure that
623 	 * clear_huge_page writes become visible before the set_pmd_at()
624 	 * write.
625 	 */
626 	__SetPageUptodate(page);
627 
628 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
629 	if (unlikely(!pmd_none(*vmf->pmd))) {
630 		goto unlock_release;
631 	} else {
632 		pmd_t entry;
633 
634 		ret = check_stable_address_space(vma->vm_mm);
635 		if (ret)
636 			goto unlock_release;
637 
638 		/* Deliver the page fault to userland */
639 		if (userfaultfd_missing(vma)) {
640 			spin_unlock(vmf->ptl);
641 			put_page(page);
642 			pte_free(vma->vm_mm, pgtable);
643 			ret = handle_userfault(vmf, VM_UFFD_MISSING);
644 			VM_BUG_ON(ret & VM_FAULT_FALLBACK);
645 			return ret;
646 		}
647 
648 		entry = mk_huge_pmd(page, vma->vm_page_prot);
649 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
650 		page_add_new_anon_rmap(page, vma, haddr, true);
651 		lru_cache_add_inactive_or_unevictable(page, vma);
652 		pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
653 		set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
654 		update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
655 		add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
656 		mm_inc_nr_ptes(vma->vm_mm);
657 		spin_unlock(vmf->ptl);
658 		count_vm_event(THP_FAULT_ALLOC);
659 		count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
660 	}
661 
662 	return 0;
663 unlock_release:
664 	spin_unlock(vmf->ptl);
665 release:
666 	if (pgtable)
667 		pte_free(vma->vm_mm, pgtable);
668 	put_page(page);
669 	return ret;
670 
671 }
672 
673 /*
674  * always: directly stall for all thp allocations
675  * defer: wake kswapd and fail if not immediately available
676  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
677  *		  fail if not immediately available
678  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
679  *	    available
680  * never: never stall for any thp allocation
681  */
682 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
683 {
684 	const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
685 
686 	/* Always do synchronous compaction */
687 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
688 		return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
689 
690 	/* Kick kcompactd and fail quickly */
691 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
692 		return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
693 
694 	/* Synchronous compaction if madvised, otherwise kick kcompactd */
695 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
696 		return GFP_TRANSHUGE_LIGHT |
697 			(vma_madvised ? __GFP_DIRECT_RECLAIM :
698 					__GFP_KSWAPD_RECLAIM);
699 
700 	/* Only do synchronous compaction if madvised */
701 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
702 		return GFP_TRANSHUGE_LIGHT |
703 		       (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
704 
705 	return GFP_TRANSHUGE_LIGHT;
706 }
707 
708 /* Caller must hold page table lock. */
709 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
710 		struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
711 		struct page *zero_page)
712 {
713 	pmd_t entry;
714 	if (!pmd_none(*pmd))
715 		return;
716 	entry = mk_pmd(zero_page, vma->vm_page_prot);
717 	entry = pmd_mkhuge(entry);
718 	if (pgtable)
719 		pgtable_trans_huge_deposit(mm, pmd, pgtable);
720 	set_pmd_at(mm, haddr, pmd, entry);
721 	mm_inc_nr_ptes(mm);
722 }
723 
724 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
725 {
726 	struct vm_area_struct *vma = vmf->vma;
727 	gfp_t gfp;
728 	struct page *page;
729 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
730 
731 	if (!transhuge_vma_suitable(vma, haddr))
732 		return VM_FAULT_FALLBACK;
733 	if (unlikely(anon_vma_prepare(vma)))
734 		return VM_FAULT_OOM;
735 	if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
736 		return VM_FAULT_OOM;
737 	if (!(vmf->flags & FAULT_FLAG_WRITE) &&
738 			!mm_forbids_zeropage(vma->vm_mm) &&
739 			transparent_hugepage_use_zero_page()) {
740 		pgtable_t pgtable;
741 		struct page *zero_page;
742 		vm_fault_t ret;
743 		pgtable = pte_alloc_one(vma->vm_mm);
744 		if (unlikely(!pgtable))
745 			return VM_FAULT_OOM;
746 		zero_page = mm_get_huge_zero_page(vma->vm_mm);
747 		if (unlikely(!zero_page)) {
748 			pte_free(vma->vm_mm, pgtable);
749 			count_vm_event(THP_FAULT_FALLBACK);
750 			return VM_FAULT_FALLBACK;
751 		}
752 		vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
753 		ret = 0;
754 		if (pmd_none(*vmf->pmd)) {
755 			ret = check_stable_address_space(vma->vm_mm);
756 			if (ret) {
757 				spin_unlock(vmf->ptl);
758 				pte_free(vma->vm_mm, pgtable);
759 			} else if (userfaultfd_missing(vma)) {
760 				spin_unlock(vmf->ptl);
761 				pte_free(vma->vm_mm, pgtable);
762 				ret = handle_userfault(vmf, VM_UFFD_MISSING);
763 				VM_BUG_ON(ret & VM_FAULT_FALLBACK);
764 			} else {
765 				set_huge_zero_page(pgtable, vma->vm_mm, vma,
766 						   haddr, vmf->pmd, zero_page);
767 				update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
768 				spin_unlock(vmf->ptl);
769 			}
770 		} else {
771 			spin_unlock(vmf->ptl);
772 			pte_free(vma->vm_mm, pgtable);
773 		}
774 		return ret;
775 	}
776 	gfp = vma_thp_gfp_mask(vma);
777 	page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
778 	if (unlikely(!page)) {
779 		count_vm_event(THP_FAULT_FALLBACK);
780 		return VM_FAULT_FALLBACK;
781 	}
782 	prep_transhuge_page(page);
783 	return __do_huge_pmd_anonymous_page(vmf, page, gfp);
784 }
785 
786 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
787 		pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
788 		pgtable_t pgtable)
789 {
790 	struct mm_struct *mm = vma->vm_mm;
791 	pmd_t entry;
792 	spinlock_t *ptl;
793 
794 	ptl = pmd_lock(mm, pmd);
795 	if (!pmd_none(*pmd)) {
796 		if (write) {
797 			if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
798 				WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
799 				goto out_unlock;
800 			}
801 			entry = pmd_mkyoung(*pmd);
802 			entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
803 			if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
804 				update_mmu_cache_pmd(vma, addr, pmd);
805 		}
806 
807 		goto out_unlock;
808 	}
809 
810 	entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
811 	if (pfn_t_devmap(pfn))
812 		entry = pmd_mkdevmap(entry);
813 	if (write) {
814 		entry = pmd_mkyoung(pmd_mkdirty(entry));
815 		entry = maybe_pmd_mkwrite(entry, vma);
816 	}
817 
818 	if (pgtable) {
819 		pgtable_trans_huge_deposit(mm, pmd, pgtable);
820 		mm_inc_nr_ptes(mm);
821 		pgtable = NULL;
822 	}
823 
824 	set_pmd_at(mm, addr, pmd, entry);
825 	update_mmu_cache_pmd(vma, addr, pmd);
826 
827 out_unlock:
828 	spin_unlock(ptl);
829 	if (pgtable)
830 		pte_free(mm, pgtable);
831 }
832 
833 /**
834  * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
835  * @vmf: Structure describing the fault
836  * @pfn: pfn to insert
837  * @pgprot: page protection to use
838  * @write: whether it's a write fault
839  *
840  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
841  * also consult the vmf_insert_mixed_prot() documentation when
842  * @pgprot != @vmf->vma->vm_page_prot.
843  *
844  * Return: vm_fault_t value.
845  */
846 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
847 				   pgprot_t pgprot, bool write)
848 {
849 	unsigned long addr = vmf->address & PMD_MASK;
850 	struct vm_area_struct *vma = vmf->vma;
851 	pgtable_t pgtable = NULL;
852 
853 	/*
854 	 * If we had pmd_special, we could avoid all these restrictions,
855 	 * but we need to be consistent with PTEs and architectures that
856 	 * can't support a 'special' bit.
857 	 */
858 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
859 			!pfn_t_devmap(pfn));
860 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
861 						(VM_PFNMAP|VM_MIXEDMAP));
862 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
863 
864 	if (addr < vma->vm_start || addr >= vma->vm_end)
865 		return VM_FAULT_SIGBUS;
866 
867 	if (arch_needs_pgtable_deposit()) {
868 		pgtable = pte_alloc_one(vma->vm_mm);
869 		if (!pgtable)
870 			return VM_FAULT_OOM;
871 	}
872 
873 	track_pfn_insert(vma, &pgprot, pfn);
874 
875 	insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
876 	return VM_FAULT_NOPAGE;
877 }
878 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
879 
880 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
881 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
882 {
883 	if (likely(vma->vm_flags & VM_WRITE))
884 		pud = pud_mkwrite(pud);
885 	return pud;
886 }
887 
888 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
889 		pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
890 {
891 	struct mm_struct *mm = vma->vm_mm;
892 	pud_t entry;
893 	spinlock_t *ptl;
894 
895 	ptl = pud_lock(mm, pud);
896 	if (!pud_none(*pud)) {
897 		if (write) {
898 			if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
899 				WARN_ON_ONCE(!is_huge_zero_pud(*pud));
900 				goto out_unlock;
901 			}
902 			entry = pud_mkyoung(*pud);
903 			entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
904 			if (pudp_set_access_flags(vma, addr, pud, entry, 1))
905 				update_mmu_cache_pud(vma, addr, pud);
906 		}
907 		goto out_unlock;
908 	}
909 
910 	entry = pud_mkhuge(pfn_t_pud(pfn, prot));
911 	if (pfn_t_devmap(pfn))
912 		entry = pud_mkdevmap(entry);
913 	if (write) {
914 		entry = pud_mkyoung(pud_mkdirty(entry));
915 		entry = maybe_pud_mkwrite(entry, vma);
916 	}
917 	set_pud_at(mm, addr, pud, entry);
918 	update_mmu_cache_pud(vma, addr, pud);
919 
920 out_unlock:
921 	spin_unlock(ptl);
922 }
923 
924 /**
925  * vmf_insert_pfn_pud_prot - insert a pud size pfn
926  * @vmf: Structure describing the fault
927  * @pfn: pfn to insert
928  * @pgprot: page protection to use
929  * @write: whether it's a write fault
930  *
931  * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
932  * also consult the vmf_insert_mixed_prot() documentation when
933  * @pgprot != @vmf->vma->vm_page_prot.
934  *
935  * Return: vm_fault_t value.
936  */
937 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
938 				   pgprot_t pgprot, bool write)
939 {
940 	unsigned long addr = vmf->address & PUD_MASK;
941 	struct vm_area_struct *vma = vmf->vma;
942 
943 	/*
944 	 * If we had pud_special, we could avoid all these restrictions,
945 	 * but we need to be consistent with PTEs and architectures that
946 	 * can't support a 'special' bit.
947 	 */
948 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
949 			!pfn_t_devmap(pfn));
950 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
951 						(VM_PFNMAP|VM_MIXEDMAP));
952 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
953 
954 	if (addr < vma->vm_start || addr >= vma->vm_end)
955 		return VM_FAULT_SIGBUS;
956 
957 	track_pfn_insert(vma, &pgprot, pfn);
958 
959 	insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
960 	return VM_FAULT_NOPAGE;
961 }
962 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
963 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
964 
965 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
966 		pmd_t *pmd, int flags)
967 {
968 	pmd_t _pmd;
969 
970 	_pmd = pmd_mkyoung(*pmd);
971 	if (flags & FOLL_WRITE)
972 		_pmd = pmd_mkdirty(_pmd);
973 	if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
974 				pmd, _pmd, flags & FOLL_WRITE))
975 		update_mmu_cache_pmd(vma, addr, pmd);
976 }
977 
978 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
979 		pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
980 {
981 	unsigned long pfn = pmd_pfn(*pmd);
982 	struct mm_struct *mm = vma->vm_mm;
983 	struct page *page;
984 
985 	assert_spin_locked(pmd_lockptr(mm, pmd));
986 
987 	/*
988 	 * When we COW a devmap PMD entry, we split it into PTEs, so we should
989 	 * not be in this function with `flags & FOLL_COW` set.
990 	 */
991 	WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
992 
993 	/* FOLL_GET and FOLL_PIN are mutually exclusive. */
994 	if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
995 			 (FOLL_PIN | FOLL_GET)))
996 		return NULL;
997 
998 	if (flags & FOLL_WRITE && !pmd_write(*pmd))
999 		return NULL;
1000 
1001 	if (pmd_present(*pmd) && pmd_devmap(*pmd))
1002 		/* pass */;
1003 	else
1004 		return NULL;
1005 
1006 	if (flags & FOLL_TOUCH)
1007 		touch_pmd(vma, addr, pmd, flags);
1008 
1009 	/*
1010 	 * device mapped pages can only be returned if the
1011 	 * caller will manage the page reference count.
1012 	 */
1013 	if (!(flags & (FOLL_GET | FOLL_PIN)))
1014 		return ERR_PTR(-EEXIST);
1015 
1016 	pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1017 	*pgmap = get_dev_pagemap(pfn, *pgmap);
1018 	if (!*pgmap)
1019 		return ERR_PTR(-EFAULT);
1020 	page = pfn_to_page(pfn);
1021 	if (!try_grab_page(page, flags))
1022 		page = ERR_PTR(-ENOMEM);
1023 
1024 	return page;
1025 }
1026 
1027 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1028 		  pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1029 		  struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1030 {
1031 	spinlock_t *dst_ptl, *src_ptl;
1032 	struct page *src_page;
1033 	pmd_t pmd;
1034 	pgtable_t pgtable = NULL;
1035 	int ret = -ENOMEM;
1036 
1037 	/* Skip if can be re-fill on fault */
1038 	if (!vma_is_anonymous(dst_vma))
1039 		return 0;
1040 
1041 	pgtable = pte_alloc_one(dst_mm);
1042 	if (unlikely(!pgtable))
1043 		goto out;
1044 
1045 	dst_ptl = pmd_lock(dst_mm, dst_pmd);
1046 	src_ptl = pmd_lockptr(src_mm, src_pmd);
1047 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1048 
1049 	ret = -EAGAIN;
1050 	pmd = *src_pmd;
1051 
1052 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1053 	if (unlikely(is_swap_pmd(pmd))) {
1054 		swp_entry_t entry = pmd_to_swp_entry(pmd);
1055 
1056 		VM_BUG_ON(!is_pmd_migration_entry(pmd));
1057 		if (is_writable_migration_entry(entry)) {
1058 			entry = make_readable_migration_entry(
1059 							swp_offset(entry));
1060 			pmd = swp_entry_to_pmd(entry);
1061 			if (pmd_swp_soft_dirty(*src_pmd))
1062 				pmd = pmd_swp_mksoft_dirty(pmd);
1063 			if (pmd_swp_uffd_wp(*src_pmd))
1064 				pmd = pmd_swp_mkuffd_wp(pmd);
1065 			set_pmd_at(src_mm, addr, src_pmd, pmd);
1066 		}
1067 		add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1068 		mm_inc_nr_ptes(dst_mm);
1069 		pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1070 		if (!userfaultfd_wp(dst_vma))
1071 			pmd = pmd_swp_clear_uffd_wp(pmd);
1072 		set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1073 		ret = 0;
1074 		goto out_unlock;
1075 	}
1076 #endif
1077 
1078 	if (unlikely(!pmd_trans_huge(pmd))) {
1079 		pte_free(dst_mm, pgtable);
1080 		goto out_unlock;
1081 	}
1082 	/*
1083 	 * When page table lock is held, the huge zero pmd should not be
1084 	 * under splitting since we don't split the page itself, only pmd to
1085 	 * a page table.
1086 	 */
1087 	if (is_huge_zero_pmd(pmd)) {
1088 		/*
1089 		 * get_huge_zero_page() will never allocate a new page here,
1090 		 * since we already have a zero page to copy. It just takes a
1091 		 * reference.
1092 		 */
1093 		mm_get_huge_zero_page(dst_mm);
1094 		goto out_zero_page;
1095 	}
1096 
1097 	src_page = pmd_page(pmd);
1098 	VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1099 
1100 	/*
1101 	 * If this page is a potentially pinned page, split and retry the fault
1102 	 * with smaller page size.  Normally this should not happen because the
1103 	 * userspace should use MADV_DONTFORK upon pinned regions.  This is a
1104 	 * best effort that the pinned pages won't be replaced by another
1105 	 * random page during the coming copy-on-write.
1106 	 */
1107 	if (unlikely(page_needs_cow_for_dma(src_vma, src_page))) {
1108 		pte_free(dst_mm, pgtable);
1109 		spin_unlock(src_ptl);
1110 		spin_unlock(dst_ptl);
1111 		__split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1112 		return -EAGAIN;
1113 	}
1114 
1115 	get_page(src_page);
1116 	page_dup_rmap(src_page, true);
1117 	add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1118 out_zero_page:
1119 	mm_inc_nr_ptes(dst_mm);
1120 	pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1121 	pmdp_set_wrprotect(src_mm, addr, src_pmd);
1122 	if (!userfaultfd_wp(dst_vma))
1123 		pmd = pmd_clear_uffd_wp(pmd);
1124 	pmd = pmd_mkold(pmd_wrprotect(pmd));
1125 	set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1126 
1127 	ret = 0;
1128 out_unlock:
1129 	spin_unlock(src_ptl);
1130 	spin_unlock(dst_ptl);
1131 out:
1132 	return ret;
1133 }
1134 
1135 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1136 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1137 		pud_t *pud, int flags)
1138 {
1139 	pud_t _pud;
1140 
1141 	_pud = pud_mkyoung(*pud);
1142 	if (flags & FOLL_WRITE)
1143 		_pud = pud_mkdirty(_pud);
1144 	if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1145 				pud, _pud, flags & FOLL_WRITE))
1146 		update_mmu_cache_pud(vma, addr, pud);
1147 }
1148 
1149 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1150 		pud_t *pud, int flags, struct dev_pagemap **pgmap)
1151 {
1152 	unsigned long pfn = pud_pfn(*pud);
1153 	struct mm_struct *mm = vma->vm_mm;
1154 	struct page *page;
1155 
1156 	assert_spin_locked(pud_lockptr(mm, pud));
1157 
1158 	if (flags & FOLL_WRITE && !pud_write(*pud))
1159 		return NULL;
1160 
1161 	/* FOLL_GET and FOLL_PIN are mutually exclusive. */
1162 	if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1163 			 (FOLL_PIN | FOLL_GET)))
1164 		return NULL;
1165 
1166 	if (pud_present(*pud) && pud_devmap(*pud))
1167 		/* pass */;
1168 	else
1169 		return NULL;
1170 
1171 	if (flags & FOLL_TOUCH)
1172 		touch_pud(vma, addr, pud, flags);
1173 
1174 	/*
1175 	 * device mapped pages can only be returned if the
1176 	 * caller will manage the page reference count.
1177 	 *
1178 	 * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1179 	 */
1180 	if (!(flags & (FOLL_GET | FOLL_PIN)))
1181 		return ERR_PTR(-EEXIST);
1182 
1183 	pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1184 	*pgmap = get_dev_pagemap(pfn, *pgmap);
1185 	if (!*pgmap)
1186 		return ERR_PTR(-EFAULT);
1187 	page = pfn_to_page(pfn);
1188 	if (!try_grab_page(page, flags))
1189 		page = ERR_PTR(-ENOMEM);
1190 
1191 	return page;
1192 }
1193 
1194 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1195 		  pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1196 		  struct vm_area_struct *vma)
1197 {
1198 	spinlock_t *dst_ptl, *src_ptl;
1199 	pud_t pud;
1200 	int ret;
1201 
1202 	dst_ptl = pud_lock(dst_mm, dst_pud);
1203 	src_ptl = pud_lockptr(src_mm, src_pud);
1204 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1205 
1206 	ret = -EAGAIN;
1207 	pud = *src_pud;
1208 	if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1209 		goto out_unlock;
1210 
1211 	/*
1212 	 * When page table lock is held, the huge zero pud should not be
1213 	 * under splitting since we don't split the page itself, only pud to
1214 	 * a page table.
1215 	 */
1216 	if (is_huge_zero_pud(pud)) {
1217 		/* No huge zero pud yet */
1218 	}
1219 
1220 	/* Please refer to comments in copy_huge_pmd() */
1221 	if (unlikely(page_needs_cow_for_dma(vma, pud_page(pud)))) {
1222 		spin_unlock(src_ptl);
1223 		spin_unlock(dst_ptl);
1224 		__split_huge_pud(vma, src_pud, addr);
1225 		return -EAGAIN;
1226 	}
1227 
1228 	pudp_set_wrprotect(src_mm, addr, src_pud);
1229 	pud = pud_mkold(pud_wrprotect(pud));
1230 	set_pud_at(dst_mm, addr, dst_pud, pud);
1231 
1232 	ret = 0;
1233 out_unlock:
1234 	spin_unlock(src_ptl);
1235 	spin_unlock(dst_ptl);
1236 	return ret;
1237 }
1238 
1239 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1240 {
1241 	pud_t entry;
1242 	unsigned long haddr;
1243 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1244 
1245 	vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1246 	if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1247 		goto unlock;
1248 
1249 	entry = pud_mkyoung(orig_pud);
1250 	if (write)
1251 		entry = pud_mkdirty(entry);
1252 	haddr = vmf->address & HPAGE_PUD_MASK;
1253 	if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1254 		update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1255 
1256 unlock:
1257 	spin_unlock(vmf->ptl);
1258 }
1259 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1260 
1261 void huge_pmd_set_accessed(struct vm_fault *vmf)
1262 {
1263 	pmd_t entry;
1264 	unsigned long haddr;
1265 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1266 	pmd_t orig_pmd = vmf->orig_pmd;
1267 
1268 	vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1269 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1270 		goto unlock;
1271 
1272 	entry = pmd_mkyoung(orig_pmd);
1273 	if (write)
1274 		entry = pmd_mkdirty(entry);
1275 	haddr = vmf->address & HPAGE_PMD_MASK;
1276 	if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1277 		update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1278 
1279 unlock:
1280 	spin_unlock(vmf->ptl);
1281 }
1282 
1283 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1284 {
1285 	struct vm_area_struct *vma = vmf->vma;
1286 	struct page *page;
1287 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1288 	pmd_t orig_pmd = vmf->orig_pmd;
1289 
1290 	vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1291 	VM_BUG_ON_VMA(!vma->anon_vma, vma);
1292 
1293 	if (is_huge_zero_pmd(orig_pmd))
1294 		goto fallback;
1295 
1296 	spin_lock(vmf->ptl);
1297 
1298 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1299 		spin_unlock(vmf->ptl);
1300 		return 0;
1301 	}
1302 
1303 	page = pmd_page(orig_pmd);
1304 	VM_BUG_ON_PAGE(!PageHead(page), page);
1305 
1306 	/* Lock page for reuse_swap_page() */
1307 	if (!trylock_page(page)) {
1308 		get_page(page);
1309 		spin_unlock(vmf->ptl);
1310 		lock_page(page);
1311 		spin_lock(vmf->ptl);
1312 		if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1313 			spin_unlock(vmf->ptl);
1314 			unlock_page(page);
1315 			put_page(page);
1316 			return 0;
1317 		}
1318 		put_page(page);
1319 	}
1320 
1321 	/*
1322 	 * We can only reuse the page if nobody else maps the huge page or it's
1323 	 * part.
1324 	 */
1325 	if (reuse_swap_page(page)) {
1326 		pmd_t entry;
1327 		entry = pmd_mkyoung(orig_pmd);
1328 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1329 		if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1330 			update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1331 		unlock_page(page);
1332 		spin_unlock(vmf->ptl);
1333 		return VM_FAULT_WRITE;
1334 	}
1335 
1336 	unlock_page(page);
1337 	spin_unlock(vmf->ptl);
1338 fallback:
1339 	__split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1340 	return VM_FAULT_FALLBACK;
1341 }
1342 
1343 /*
1344  * FOLL_FORCE can write to even unwritable pmd's, but only
1345  * after we've gone through a COW cycle and they are dirty.
1346  */
1347 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1348 {
1349 	return pmd_write(pmd) ||
1350 	       ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1351 }
1352 
1353 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1354 				   unsigned long addr,
1355 				   pmd_t *pmd,
1356 				   unsigned int flags)
1357 {
1358 	struct mm_struct *mm = vma->vm_mm;
1359 	struct page *page = NULL;
1360 
1361 	assert_spin_locked(pmd_lockptr(mm, pmd));
1362 
1363 	if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1364 		goto out;
1365 
1366 	/* Avoid dumping huge zero page */
1367 	if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1368 		return ERR_PTR(-EFAULT);
1369 
1370 	/* Full NUMA hinting faults to serialise migration in fault paths */
1371 	if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1372 		goto out;
1373 
1374 	page = pmd_page(*pmd);
1375 	VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1376 
1377 	if (!try_grab_page(page, flags))
1378 		return ERR_PTR(-ENOMEM);
1379 
1380 	if (flags & FOLL_TOUCH)
1381 		touch_pmd(vma, addr, pmd, flags);
1382 
1383 	if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1384 		/*
1385 		 * We don't mlock() pte-mapped THPs. This way we can avoid
1386 		 * leaking mlocked pages into non-VM_LOCKED VMAs.
1387 		 *
1388 		 * For anon THP:
1389 		 *
1390 		 * In most cases the pmd is the only mapping of the page as we
1391 		 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1392 		 * writable private mappings in populate_vma_page_range().
1393 		 *
1394 		 * The only scenario when we have the page shared here is if we
1395 		 * mlocking read-only mapping shared over fork(). We skip
1396 		 * mlocking such pages.
1397 		 *
1398 		 * For file THP:
1399 		 *
1400 		 * We can expect PageDoubleMap() to be stable under page lock:
1401 		 * for file pages we set it in page_add_file_rmap(), which
1402 		 * requires page to be locked.
1403 		 */
1404 
1405 		if (PageAnon(page) && compound_mapcount(page) != 1)
1406 			goto skip_mlock;
1407 		if (PageDoubleMap(page) || !page->mapping)
1408 			goto skip_mlock;
1409 		if (!trylock_page(page))
1410 			goto skip_mlock;
1411 		if (page->mapping && !PageDoubleMap(page))
1412 			mlock_vma_page(page);
1413 		unlock_page(page);
1414 	}
1415 skip_mlock:
1416 	page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1417 	VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1418 
1419 out:
1420 	return page;
1421 }
1422 
1423 /* NUMA hinting page fault entry point for trans huge pmds */
1424 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
1425 {
1426 	struct vm_area_struct *vma = vmf->vma;
1427 	pmd_t oldpmd = vmf->orig_pmd;
1428 	pmd_t pmd;
1429 	struct page *page;
1430 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1431 	int page_nid = NUMA_NO_NODE;
1432 	int target_nid, last_cpupid = -1;
1433 	bool migrated = false;
1434 	bool was_writable = pmd_savedwrite(oldpmd);
1435 	int flags = 0;
1436 
1437 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1438 	if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1439 		spin_unlock(vmf->ptl);
1440 		goto out;
1441 	}
1442 
1443 	pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1444 	page = vm_normal_page_pmd(vma, haddr, pmd);
1445 	if (!page)
1446 		goto out_map;
1447 
1448 	/* See similar comment in do_numa_page for explanation */
1449 	if (!was_writable)
1450 		flags |= TNF_NO_GROUP;
1451 
1452 	page_nid = page_to_nid(page);
1453 	last_cpupid = page_cpupid_last(page);
1454 	target_nid = numa_migrate_prep(page, vma, haddr, page_nid,
1455 				       &flags);
1456 
1457 	if (target_nid == NUMA_NO_NODE) {
1458 		put_page(page);
1459 		goto out_map;
1460 	}
1461 
1462 	spin_unlock(vmf->ptl);
1463 
1464 	migrated = migrate_misplaced_page(page, vma, target_nid);
1465 	if (migrated) {
1466 		flags |= TNF_MIGRATED;
1467 		page_nid = target_nid;
1468 	} else {
1469 		flags |= TNF_MIGRATE_FAIL;
1470 		vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1471 		if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1472 			spin_unlock(vmf->ptl);
1473 			goto out;
1474 		}
1475 		goto out_map;
1476 	}
1477 
1478 out:
1479 	if (page_nid != NUMA_NO_NODE)
1480 		task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1481 				flags);
1482 
1483 	return 0;
1484 
1485 out_map:
1486 	/* Restore the PMD */
1487 	pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1488 	pmd = pmd_mkyoung(pmd);
1489 	if (was_writable)
1490 		pmd = pmd_mkwrite(pmd);
1491 	set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1492 	update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1493 	spin_unlock(vmf->ptl);
1494 	goto out;
1495 }
1496 
1497 /*
1498  * Return true if we do MADV_FREE successfully on entire pmd page.
1499  * Otherwise, return false.
1500  */
1501 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1502 		pmd_t *pmd, unsigned long addr, unsigned long next)
1503 {
1504 	spinlock_t *ptl;
1505 	pmd_t orig_pmd;
1506 	struct page *page;
1507 	struct mm_struct *mm = tlb->mm;
1508 	bool ret = false;
1509 
1510 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1511 
1512 	ptl = pmd_trans_huge_lock(pmd, vma);
1513 	if (!ptl)
1514 		goto out_unlocked;
1515 
1516 	orig_pmd = *pmd;
1517 	if (is_huge_zero_pmd(orig_pmd))
1518 		goto out;
1519 
1520 	if (unlikely(!pmd_present(orig_pmd))) {
1521 		VM_BUG_ON(thp_migration_supported() &&
1522 				  !is_pmd_migration_entry(orig_pmd));
1523 		goto out;
1524 	}
1525 
1526 	page = pmd_page(orig_pmd);
1527 	/*
1528 	 * If other processes are mapping this page, we couldn't discard
1529 	 * the page unless they all do MADV_FREE so let's skip the page.
1530 	 */
1531 	if (total_mapcount(page) != 1)
1532 		goto out;
1533 
1534 	if (!trylock_page(page))
1535 		goto out;
1536 
1537 	/*
1538 	 * If user want to discard part-pages of THP, split it so MADV_FREE
1539 	 * will deactivate only them.
1540 	 */
1541 	if (next - addr != HPAGE_PMD_SIZE) {
1542 		get_page(page);
1543 		spin_unlock(ptl);
1544 		split_huge_page(page);
1545 		unlock_page(page);
1546 		put_page(page);
1547 		goto out_unlocked;
1548 	}
1549 
1550 	if (PageDirty(page))
1551 		ClearPageDirty(page);
1552 	unlock_page(page);
1553 
1554 	if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1555 		pmdp_invalidate(vma, addr, pmd);
1556 		orig_pmd = pmd_mkold(orig_pmd);
1557 		orig_pmd = pmd_mkclean(orig_pmd);
1558 
1559 		set_pmd_at(mm, addr, pmd, orig_pmd);
1560 		tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1561 	}
1562 
1563 	mark_page_lazyfree(page);
1564 	ret = true;
1565 out:
1566 	spin_unlock(ptl);
1567 out_unlocked:
1568 	return ret;
1569 }
1570 
1571 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1572 {
1573 	pgtable_t pgtable;
1574 
1575 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1576 	pte_free(mm, pgtable);
1577 	mm_dec_nr_ptes(mm);
1578 }
1579 
1580 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1581 		 pmd_t *pmd, unsigned long addr)
1582 {
1583 	pmd_t orig_pmd;
1584 	spinlock_t *ptl;
1585 
1586 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1587 
1588 	ptl = __pmd_trans_huge_lock(pmd, vma);
1589 	if (!ptl)
1590 		return 0;
1591 	/*
1592 	 * For architectures like ppc64 we look at deposited pgtable
1593 	 * when calling pmdp_huge_get_and_clear. So do the
1594 	 * pgtable_trans_huge_withdraw after finishing pmdp related
1595 	 * operations.
1596 	 */
1597 	orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1598 						tlb->fullmm);
1599 	tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1600 	if (vma_is_special_huge(vma)) {
1601 		if (arch_needs_pgtable_deposit())
1602 			zap_deposited_table(tlb->mm, pmd);
1603 		spin_unlock(ptl);
1604 	} else if (is_huge_zero_pmd(orig_pmd)) {
1605 		zap_deposited_table(tlb->mm, pmd);
1606 		spin_unlock(ptl);
1607 	} else {
1608 		struct page *page = NULL;
1609 		int flush_needed = 1;
1610 
1611 		if (pmd_present(orig_pmd)) {
1612 			page = pmd_page(orig_pmd);
1613 			page_remove_rmap(page, true);
1614 			VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1615 			VM_BUG_ON_PAGE(!PageHead(page), page);
1616 		} else if (thp_migration_supported()) {
1617 			swp_entry_t entry;
1618 
1619 			VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1620 			entry = pmd_to_swp_entry(orig_pmd);
1621 			page = pfn_swap_entry_to_page(entry);
1622 			flush_needed = 0;
1623 		} else
1624 			WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1625 
1626 		if (PageAnon(page)) {
1627 			zap_deposited_table(tlb->mm, pmd);
1628 			add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1629 		} else {
1630 			if (arch_needs_pgtable_deposit())
1631 				zap_deposited_table(tlb->mm, pmd);
1632 			add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1633 		}
1634 
1635 		spin_unlock(ptl);
1636 		if (flush_needed)
1637 			tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1638 	}
1639 	return 1;
1640 }
1641 
1642 #ifndef pmd_move_must_withdraw
1643 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1644 					 spinlock_t *old_pmd_ptl,
1645 					 struct vm_area_struct *vma)
1646 {
1647 	/*
1648 	 * With split pmd lock we also need to move preallocated
1649 	 * PTE page table if new_pmd is on different PMD page table.
1650 	 *
1651 	 * We also don't deposit and withdraw tables for file pages.
1652 	 */
1653 	return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1654 }
1655 #endif
1656 
1657 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1658 {
1659 #ifdef CONFIG_MEM_SOFT_DIRTY
1660 	if (unlikely(is_pmd_migration_entry(pmd)))
1661 		pmd = pmd_swp_mksoft_dirty(pmd);
1662 	else if (pmd_present(pmd))
1663 		pmd = pmd_mksoft_dirty(pmd);
1664 #endif
1665 	return pmd;
1666 }
1667 
1668 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1669 		  unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1670 {
1671 	spinlock_t *old_ptl, *new_ptl;
1672 	pmd_t pmd;
1673 	struct mm_struct *mm = vma->vm_mm;
1674 	bool force_flush = false;
1675 
1676 	/*
1677 	 * The destination pmd shouldn't be established, free_pgtables()
1678 	 * should have release it.
1679 	 */
1680 	if (WARN_ON(!pmd_none(*new_pmd))) {
1681 		VM_BUG_ON(pmd_trans_huge(*new_pmd));
1682 		return false;
1683 	}
1684 
1685 	/*
1686 	 * We don't have to worry about the ordering of src and dst
1687 	 * ptlocks because exclusive mmap_lock prevents deadlock.
1688 	 */
1689 	old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1690 	if (old_ptl) {
1691 		new_ptl = pmd_lockptr(mm, new_pmd);
1692 		if (new_ptl != old_ptl)
1693 			spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1694 		pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1695 		if (pmd_present(pmd))
1696 			force_flush = true;
1697 		VM_BUG_ON(!pmd_none(*new_pmd));
1698 
1699 		if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1700 			pgtable_t pgtable;
1701 			pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1702 			pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1703 		}
1704 		pmd = move_soft_dirty_pmd(pmd);
1705 		set_pmd_at(mm, new_addr, new_pmd, pmd);
1706 		if (force_flush)
1707 			flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1708 		if (new_ptl != old_ptl)
1709 			spin_unlock(new_ptl);
1710 		spin_unlock(old_ptl);
1711 		return true;
1712 	}
1713 	return false;
1714 }
1715 
1716 /*
1717  * Returns
1718  *  - 0 if PMD could not be locked
1719  *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1720  *      or if prot_numa but THP migration is not supported
1721  *  - HPAGE_PMD_NR if protections changed and TLB flush necessary
1722  */
1723 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1724 		unsigned long addr, pgprot_t newprot, unsigned long cp_flags)
1725 {
1726 	struct mm_struct *mm = vma->vm_mm;
1727 	spinlock_t *ptl;
1728 	pmd_t entry;
1729 	bool preserve_write;
1730 	int ret;
1731 	bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1732 	bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1733 	bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1734 
1735 	if (prot_numa && !thp_migration_supported())
1736 		return 1;
1737 
1738 	ptl = __pmd_trans_huge_lock(pmd, vma);
1739 	if (!ptl)
1740 		return 0;
1741 
1742 	preserve_write = prot_numa && pmd_write(*pmd);
1743 	ret = 1;
1744 
1745 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1746 	if (is_swap_pmd(*pmd)) {
1747 		swp_entry_t entry = pmd_to_swp_entry(*pmd);
1748 
1749 		VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1750 		if (is_writable_migration_entry(entry)) {
1751 			pmd_t newpmd;
1752 			/*
1753 			 * A protection check is difficult so
1754 			 * just be safe and disable write
1755 			 */
1756 			entry = make_readable_migration_entry(
1757 							swp_offset(entry));
1758 			newpmd = swp_entry_to_pmd(entry);
1759 			if (pmd_swp_soft_dirty(*pmd))
1760 				newpmd = pmd_swp_mksoft_dirty(newpmd);
1761 			if (pmd_swp_uffd_wp(*pmd))
1762 				newpmd = pmd_swp_mkuffd_wp(newpmd);
1763 			set_pmd_at(mm, addr, pmd, newpmd);
1764 		}
1765 		goto unlock;
1766 	}
1767 #endif
1768 
1769 	/*
1770 	 * Avoid trapping faults against the zero page. The read-only
1771 	 * data is likely to be read-cached on the local CPU and
1772 	 * local/remote hits to the zero page are not interesting.
1773 	 */
1774 	if (prot_numa && is_huge_zero_pmd(*pmd))
1775 		goto unlock;
1776 
1777 	if (prot_numa && pmd_protnone(*pmd))
1778 		goto unlock;
1779 
1780 	/*
1781 	 * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1782 	 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1783 	 * which is also under mmap_read_lock(mm):
1784 	 *
1785 	 *	CPU0:				CPU1:
1786 	 *				change_huge_pmd(prot_numa=1)
1787 	 *				 pmdp_huge_get_and_clear_notify()
1788 	 * madvise_dontneed()
1789 	 *  zap_pmd_range()
1790 	 *   pmd_trans_huge(*pmd) == 0 (without ptl)
1791 	 *   // skip the pmd
1792 	 *				 set_pmd_at();
1793 	 *				 // pmd is re-established
1794 	 *
1795 	 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1796 	 * which may break userspace.
1797 	 *
1798 	 * pmdp_invalidate() is required to make sure we don't miss
1799 	 * dirty/young flags set by hardware.
1800 	 */
1801 	entry = pmdp_invalidate(vma, addr, pmd);
1802 
1803 	entry = pmd_modify(entry, newprot);
1804 	if (preserve_write)
1805 		entry = pmd_mk_savedwrite(entry);
1806 	if (uffd_wp) {
1807 		entry = pmd_wrprotect(entry);
1808 		entry = pmd_mkuffd_wp(entry);
1809 	} else if (uffd_wp_resolve) {
1810 		/*
1811 		 * Leave the write bit to be handled by PF interrupt
1812 		 * handler, then things like COW could be properly
1813 		 * handled.
1814 		 */
1815 		entry = pmd_clear_uffd_wp(entry);
1816 	}
1817 	ret = HPAGE_PMD_NR;
1818 	set_pmd_at(mm, addr, pmd, entry);
1819 	BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1820 unlock:
1821 	spin_unlock(ptl);
1822 	return ret;
1823 }
1824 
1825 /*
1826  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1827  *
1828  * Note that if it returns page table lock pointer, this routine returns without
1829  * unlocking page table lock. So callers must unlock it.
1830  */
1831 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1832 {
1833 	spinlock_t *ptl;
1834 	ptl = pmd_lock(vma->vm_mm, pmd);
1835 	if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1836 			pmd_devmap(*pmd)))
1837 		return ptl;
1838 	spin_unlock(ptl);
1839 	return NULL;
1840 }
1841 
1842 /*
1843  * Returns true if a given pud maps a thp, false otherwise.
1844  *
1845  * Note that if it returns true, this routine returns without unlocking page
1846  * table lock. So callers must unlock it.
1847  */
1848 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1849 {
1850 	spinlock_t *ptl;
1851 
1852 	ptl = pud_lock(vma->vm_mm, pud);
1853 	if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1854 		return ptl;
1855 	spin_unlock(ptl);
1856 	return NULL;
1857 }
1858 
1859 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1860 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1861 		 pud_t *pud, unsigned long addr)
1862 {
1863 	spinlock_t *ptl;
1864 
1865 	ptl = __pud_trans_huge_lock(pud, vma);
1866 	if (!ptl)
1867 		return 0;
1868 	/*
1869 	 * For architectures like ppc64 we look at deposited pgtable
1870 	 * when calling pudp_huge_get_and_clear. So do the
1871 	 * pgtable_trans_huge_withdraw after finishing pudp related
1872 	 * operations.
1873 	 */
1874 	pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1875 	tlb_remove_pud_tlb_entry(tlb, pud, addr);
1876 	if (vma_is_special_huge(vma)) {
1877 		spin_unlock(ptl);
1878 		/* No zero page support yet */
1879 	} else {
1880 		/* No support for anonymous PUD pages yet */
1881 		BUG();
1882 	}
1883 	return 1;
1884 }
1885 
1886 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1887 		unsigned long haddr)
1888 {
1889 	VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1890 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1891 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1892 	VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1893 
1894 	count_vm_event(THP_SPLIT_PUD);
1895 
1896 	pudp_huge_clear_flush_notify(vma, haddr, pud);
1897 }
1898 
1899 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1900 		unsigned long address)
1901 {
1902 	spinlock_t *ptl;
1903 	struct mmu_notifier_range range;
1904 
1905 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1906 				address & HPAGE_PUD_MASK,
1907 				(address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1908 	mmu_notifier_invalidate_range_start(&range);
1909 	ptl = pud_lock(vma->vm_mm, pud);
1910 	if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1911 		goto out;
1912 	__split_huge_pud_locked(vma, pud, range.start);
1913 
1914 out:
1915 	spin_unlock(ptl);
1916 	/*
1917 	 * No need to double call mmu_notifier->invalidate_range() callback as
1918 	 * the above pudp_huge_clear_flush_notify() did already call it.
1919 	 */
1920 	mmu_notifier_invalidate_range_only_end(&range);
1921 }
1922 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1923 
1924 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1925 		unsigned long haddr, pmd_t *pmd)
1926 {
1927 	struct mm_struct *mm = vma->vm_mm;
1928 	pgtable_t pgtable;
1929 	pmd_t _pmd;
1930 	int i;
1931 
1932 	/*
1933 	 * Leave pmd empty until pte is filled note that it is fine to delay
1934 	 * notification until mmu_notifier_invalidate_range_end() as we are
1935 	 * replacing a zero pmd write protected page with a zero pte write
1936 	 * protected page.
1937 	 *
1938 	 * See Documentation/vm/mmu_notifier.rst
1939 	 */
1940 	pmdp_huge_clear_flush(vma, haddr, pmd);
1941 
1942 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1943 	pmd_populate(mm, &_pmd, pgtable);
1944 
1945 	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1946 		pte_t *pte, entry;
1947 		entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1948 		entry = pte_mkspecial(entry);
1949 		pte = pte_offset_map(&_pmd, haddr);
1950 		VM_BUG_ON(!pte_none(*pte));
1951 		set_pte_at(mm, haddr, pte, entry);
1952 		pte_unmap(pte);
1953 	}
1954 	smp_wmb(); /* make pte visible before pmd */
1955 	pmd_populate(mm, pmd, pgtable);
1956 }
1957 
1958 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
1959 		unsigned long haddr, bool freeze)
1960 {
1961 	struct mm_struct *mm = vma->vm_mm;
1962 	struct page *page;
1963 	pgtable_t pgtable;
1964 	pmd_t old_pmd, _pmd;
1965 	bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
1966 	unsigned long addr;
1967 	int i;
1968 
1969 	VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1970 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1971 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
1972 	VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
1973 				&& !pmd_devmap(*pmd));
1974 
1975 	count_vm_event(THP_SPLIT_PMD);
1976 
1977 	if (!vma_is_anonymous(vma)) {
1978 		old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1979 		/*
1980 		 * We are going to unmap this huge page. So
1981 		 * just go ahead and zap it
1982 		 */
1983 		if (arch_needs_pgtable_deposit())
1984 			zap_deposited_table(mm, pmd);
1985 		if (vma_is_special_huge(vma))
1986 			return;
1987 		if (unlikely(is_pmd_migration_entry(old_pmd))) {
1988 			swp_entry_t entry;
1989 
1990 			entry = pmd_to_swp_entry(old_pmd);
1991 			page = pfn_swap_entry_to_page(entry);
1992 		} else {
1993 			page = pmd_page(old_pmd);
1994 			if (!PageDirty(page) && pmd_dirty(old_pmd))
1995 				set_page_dirty(page);
1996 			if (!PageReferenced(page) && pmd_young(old_pmd))
1997 				SetPageReferenced(page);
1998 			page_remove_rmap(page, true);
1999 			put_page(page);
2000 		}
2001 		add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2002 		return;
2003 	}
2004 
2005 	if (is_huge_zero_pmd(*pmd)) {
2006 		/*
2007 		 * FIXME: Do we want to invalidate secondary mmu by calling
2008 		 * mmu_notifier_invalidate_range() see comments below inside
2009 		 * __split_huge_pmd() ?
2010 		 *
2011 		 * We are going from a zero huge page write protected to zero
2012 		 * small page also write protected so it does not seems useful
2013 		 * to invalidate secondary mmu at this time.
2014 		 */
2015 		return __split_huge_zero_page_pmd(vma, haddr, pmd);
2016 	}
2017 
2018 	/*
2019 	 * Up to this point the pmd is present and huge and userland has the
2020 	 * whole access to the hugepage during the split (which happens in
2021 	 * place). If we overwrite the pmd with the not-huge version pointing
2022 	 * to the pte here (which of course we could if all CPUs were bug
2023 	 * free), userland could trigger a small page size TLB miss on the
2024 	 * small sized TLB while the hugepage TLB entry is still established in
2025 	 * the huge TLB. Some CPU doesn't like that.
2026 	 * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2027 	 * 383 on page 105. Intel should be safe but is also warns that it's
2028 	 * only safe if the permission and cache attributes of the two entries
2029 	 * loaded in the two TLB is identical (which should be the case here).
2030 	 * But it is generally safer to never allow small and huge TLB entries
2031 	 * for the same virtual address to be loaded simultaneously. So instead
2032 	 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2033 	 * current pmd notpresent (atomically because here the pmd_trans_huge
2034 	 * must remain set at all times on the pmd until the split is complete
2035 	 * for this pmd), then we flush the SMP TLB and finally we write the
2036 	 * non-huge version of the pmd entry with pmd_populate.
2037 	 */
2038 	old_pmd = pmdp_invalidate(vma, haddr, pmd);
2039 
2040 	pmd_migration = is_pmd_migration_entry(old_pmd);
2041 	if (unlikely(pmd_migration)) {
2042 		swp_entry_t entry;
2043 
2044 		entry = pmd_to_swp_entry(old_pmd);
2045 		page = pfn_swap_entry_to_page(entry);
2046 		write = is_writable_migration_entry(entry);
2047 		young = false;
2048 		soft_dirty = pmd_swp_soft_dirty(old_pmd);
2049 		uffd_wp = pmd_swp_uffd_wp(old_pmd);
2050 	} else {
2051 		page = pmd_page(old_pmd);
2052 		if (pmd_dirty(old_pmd))
2053 			SetPageDirty(page);
2054 		write = pmd_write(old_pmd);
2055 		young = pmd_young(old_pmd);
2056 		soft_dirty = pmd_soft_dirty(old_pmd);
2057 		uffd_wp = pmd_uffd_wp(old_pmd);
2058 	}
2059 	VM_BUG_ON_PAGE(!page_count(page), page);
2060 	page_ref_add(page, HPAGE_PMD_NR - 1);
2061 
2062 	/*
2063 	 * Withdraw the table only after we mark the pmd entry invalid.
2064 	 * This's critical for some architectures (Power).
2065 	 */
2066 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2067 	pmd_populate(mm, &_pmd, pgtable);
2068 
2069 	for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2070 		pte_t entry, *pte;
2071 		/*
2072 		 * Note that NUMA hinting access restrictions are not
2073 		 * transferred to avoid any possibility of altering
2074 		 * permissions across VMAs.
2075 		 */
2076 		if (freeze || pmd_migration) {
2077 			swp_entry_t swp_entry;
2078 			if (write)
2079 				swp_entry = make_writable_migration_entry(
2080 							page_to_pfn(page + i));
2081 			else
2082 				swp_entry = make_readable_migration_entry(
2083 							page_to_pfn(page + i));
2084 			entry = swp_entry_to_pte(swp_entry);
2085 			if (soft_dirty)
2086 				entry = pte_swp_mksoft_dirty(entry);
2087 			if (uffd_wp)
2088 				entry = pte_swp_mkuffd_wp(entry);
2089 		} else {
2090 			entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2091 			entry = maybe_mkwrite(entry, vma);
2092 			if (!write)
2093 				entry = pte_wrprotect(entry);
2094 			if (!young)
2095 				entry = pte_mkold(entry);
2096 			if (soft_dirty)
2097 				entry = pte_mksoft_dirty(entry);
2098 			if (uffd_wp)
2099 				entry = pte_mkuffd_wp(entry);
2100 		}
2101 		pte = pte_offset_map(&_pmd, addr);
2102 		BUG_ON(!pte_none(*pte));
2103 		set_pte_at(mm, addr, pte, entry);
2104 		if (!pmd_migration)
2105 			atomic_inc(&page[i]._mapcount);
2106 		pte_unmap(pte);
2107 	}
2108 
2109 	if (!pmd_migration) {
2110 		/*
2111 		 * Set PG_double_map before dropping compound_mapcount to avoid
2112 		 * false-negative page_mapped().
2113 		 */
2114 		if (compound_mapcount(page) > 1 &&
2115 		    !TestSetPageDoubleMap(page)) {
2116 			for (i = 0; i < HPAGE_PMD_NR; i++)
2117 				atomic_inc(&page[i]._mapcount);
2118 		}
2119 
2120 		lock_page_memcg(page);
2121 		if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2122 			/* Last compound_mapcount is gone. */
2123 			__mod_lruvec_page_state(page, NR_ANON_THPS,
2124 						-HPAGE_PMD_NR);
2125 			if (TestClearPageDoubleMap(page)) {
2126 				/* No need in mapcount reference anymore */
2127 				for (i = 0; i < HPAGE_PMD_NR; i++)
2128 					atomic_dec(&page[i]._mapcount);
2129 			}
2130 		}
2131 		unlock_page_memcg(page);
2132 	}
2133 
2134 	smp_wmb(); /* make pte visible before pmd */
2135 	pmd_populate(mm, pmd, pgtable);
2136 
2137 	if (freeze) {
2138 		for (i = 0; i < HPAGE_PMD_NR; i++) {
2139 			page_remove_rmap(page + i, false);
2140 			put_page(page + i);
2141 		}
2142 	}
2143 }
2144 
2145 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2146 		unsigned long address, bool freeze, struct page *page)
2147 {
2148 	spinlock_t *ptl;
2149 	struct mmu_notifier_range range;
2150 	bool do_unlock_page = false;
2151 	pmd_t _pmd;
2152 
2153 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2154 				address & HPAGE_PMD_MASK,
2155 				(address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2156 	mmu_notifier_invalidate_range_start(&range);
2157 	ptl = pmd_lock(vma->vm_mm, pmd);
2158 
2159 	/*
2160 	 * If caller asks to setup a migration entries, we need a page to check
2161 	 * pmd against. Otherwise we can end up replacing wrong page.
2162 	 */
2163 	VM_BUG_ON(freeze && !page);
2164 	if (page) {
2165 		VM_WARN_ON_ONCE(!PageLocked(page));
2166 		if (page != pmd_page(*pmd))
2167 			goto out;
2168 	}
2169 
2170 repeat:
2171 	if (pmd_trans_huge(*pmd)) {
2172 		if (!page) {
2173 			page = pmd_page(*pmd);
2174 			/*
2175 			 * An anonymous page must be locked, to ensure that a
2176 			 * concurrent reuse_swap_page() sees stable mapcount;
2177 			 * but reuse_swap_page() is not used on shmem or file,
2178 			 * and page lock must not be taken when zap_pmd_range()
2179 			 * calls __split_huge_pmd() while i_mmap_lock is held.
2180 			 */
2181 			if (PageAnon(page)) {
2182 				if (unlikely(!trylock_page(page))) {
2183 					get_page(page);
2184 					_pmd = *pmd;
2185 					spin_unlock(ptl);
2186 					lock_page(page);
2187 					spin_lock(ptl);
2188 					if (unlikely(!pmd_same(*pmd, _pmd))) {
2189 						unlock_page(page);
2190 						put_page(page);
2191 						page = NULL;
2192 						goto repeat;
2193 					}
2194 					put_page(page);
2195 				}
2196 				do_unlock_page = true;
2197 			}
2198 		}
2199 		if (PageMlocked(page))
2200 			clear_page_mlock(page);
2201 	} else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2202 		goto out;
2203 	__split_huge_pmd_locked(vma, pmd, range.start, freeze);
2204 out:
2205 	spin_unlock(ptl);
2206 	if (do_unlock_page)
2207 		unlock_page(page);
2208 	/*
2209 	 * No need to double call mmu_notifier->invalidate_range() callback.
2210 	 * They are 3 cases to consider inside __split_huge_pmd_locked():
2211 	 *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2212 	 *  2) __split_huge_zero_page_pmd() read only zero page and any write
2213 	 *    fault will trigger a flush_notify before pointing to a new page
2214 	 *    (it is fine if the secondary mmu keeps pointing to the old zero
2215 	 *    page in the meantime)
2216 	 *  3) Split a huge pmd into pte pointing to the same page. No need
2217 	 *     to invalidate secondary tlb entry they are all still valid.
2218 	 *     any further changes to individual pte will notify. So no need
2219 	 *     to call mmu_notifier->invalidate_range()
2220 	 */
2221 	mmu_notifier_invalidate_range_only_end(&range);
2222 }
2223 
2224 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2225 		bool freeze, struct page *page)
2226 {
2227 	pgd_t *pgd;
2228 	p4d_t *p4d;
2229 	pud_t *pud;
2230 	pmd_t *pmd;
2231 
2232 	pgd = pgd_offset(vma->vm_mm, address);
2233 	if (!pgd_present(*pgd))
2234 		return;
2235 
2236 	p4d = p4d_offset(pgd, address);
2237 	if (!p4d_present(*p4d))
2238 		return;
2239 
2240 	pud = pud_offset(p4d, address);
2241 	if (!pud_present(*pud))
2242 		return;
2243 
2244 	pmd = pmd_offset(pud, address);
2245 
2246 	__split_huge_pmd(vma, pmd, address, freeze, page);
2247 }
2248 
2249 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2250 {
2251 	/*
2252 	 * If the new address isn't hpage aligned and it could previously
2253 	 * contain an hugepage: check if we need to split an huge pmd.
2254 	 */
2255 	if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2256 	    range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2257 			 ALIGN(address, HPAGE_PMD_SIZE)))
2258 		split_huge_pmd_address(vma, address, false, NULL);
2259 }
2260 
2261 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2262 			     unsigned long start,
2263 			     unsigned long end,
2264 			     long adjust_next)
2265 {
2266 	/* Check if we need to split start first. */
2267 	split_huge_pmd_if_needed(vma, start);
2268 
2269 	/* Check if we need to split end next. */
2270 	split_huge_pmd_if_needed(vma, end);
2271 
2272 	/*
2273 	 * If we're also updating the vma->vm_next->vm_start,
2274 	 * check if we need to split it.
2275 	 */
2276 	if (adjust_next > 0) {
2277 		struct vm_area_struct *next = vma->vm_next;
2278 		unsigned long nstart = next->vm_start;
2279 		nstart += adjust_next;
2280 		split_huge_pmd_if_needed(next, nstart);
2281 	}
2282 }
2283 
2284 static void unmap_page(struct page *page)
2285 {
2286 	enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2287 		TTU_SYNC;
2288 
2289 	VM_BUG_ON_PAGE(!PageHead(page), page);
2290 
2291 	/*
2292 	 * Anon pages need migration entries to preserve them, but file
2293 	 * pages can simply be left unmapped, then faulted back on demand.
2294 	 * If that is ever changed (perhaps for mlock), update remap_page().
2295 	 */
2296 	if (PageAnon(page))
2297 		try_to_migrate(page, ttu_flags);
2298 	else
2299 		try_to_unmap(page, ttu_flags | TTU_IGNORE_MLOCK);
2300 
2301 	VM_WARN_ON_ONCE_PAGE(page_mapped(page), page);
2302 }
2303 
2304 static void remap_page(struct page *page, unsigned int nr)
2305 {
2306 	int i;
2307 
2308 	/* If unmap_page() uses try_to_migrate() on file, remove this check */
2309 	if (!PageAnon(page))
2310 		return;
2311 	if (PageTransHuge(page)) {
2312 		remove_migration_ptes(page, page, true);
2313 	} else {
2314 		for (i = 0; i < nr; i++)
2315 			remove_migration_ptes(page + i, page + i, true);
2316 	}
2317 }
2318 
2319 static void lru_add_page_tail(struct page *head, struct page *tail,
2320 		struct lruvec *lruvec, struct list_head *list)
2321 {
2322 	VM_BUG_ON_PAGE(!PageHead(head), head);
2323 	VM_BUG_ON_PAGE(PageCompound(tail), head);
2324 	VM_BUG_ON_PAGE(PageLRU(tail), head);
2325 	lockdep_assert_held(&lruvec->lru_lock);
2326 
2327 	if (list) {
2328 		/* page reclaim is reclaiming a huge page */
2329 		VM_WARN_ON(PageLRU(head));
2330 		get_page(tail);
2331 		list_add_tail(&tail->lru, list);
2332 	} else {
2333 		/* head is still on lru (and we have it frozen) */
2334 		VM_WARN_ON(!PageLRU(head));
2335 		SetPageLRU(tail);
2336 		list_add_tail(&tail->lru, &head->lru);
2337 	}
2338 }
2339 
2340 static void __split_huge_page_tail(struct page *head, int tail,
2341 		struct lruvec *lruvec, struct list_head *list)
2342 {
2343 	struct page *page_tail = head + tail;
2344 
2345 	VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2346 
2347 	/*
2348 	 * Clone page flags before unfreezing refcount.
2349 	 *
2350 	 * After successful get_page_unless_zero() might follow flags change,
2351 	 * for example lock_page() which set PG_waiters.
2352 	 */
2353 	page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2354 	page_tail->flags |= (head->flags &
2355 			((1L << PG_referenced) |
2356 			 (1L << PG_swapbacked) |
2357 			 (1L << PG_swapcache) |
2358 			 (1L << PG_mlocked) |
2359 			 (1L << PG_uptodate) |
2360 			 (1L << PG_active) |
2361 			 (1L << PG_workingset) |
2362 			 (1L << PG_locked) |
2363 			 (1L << PG_unevictable) |
2364 #ifdef CONFIG_64BIT
2365 			 (1L << PG_arch_2) |
2366 #endif
2367 			 (1L << PG_dirty)));
2368 
2369 	/* ->mapping in first tail page is compound_mapcount */
2370 	VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2371 			page_tail);
2372 	page_tail->mapping = head->mapping;
2373 	page_tail->index = head->index + tail;
2374 
2375 	/* Page flags must be visible before we make the page non-compound. */
2376 	smp_wmb();
2377 
2378 	/*
2379 	 * Clear PageTail before unfreezing page refcount.
2380 	 *
2381 	 * After successful get_page_unless_zero() might follow put_page()
2382 	 * which needs correct compound_head().
2383 	 */
2384 	clear_compound_head(page_tail);
2385 
2386 	/* Finally unfreeze refcount. Additional reference from page cache. */
2387 	page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2388 					  PageSwapCache(head)));
2389 
2390 	if (page_is_young(head))
2391 		set_page_young(page_tail);
2392 	if (page_is_idle(head))
2393 		set_page_idle(page_tail);
2394 
2395 	page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2396 
2397 	/*
2398 	 * always add to the tail because some iterators expect new
2399 	 * pages to show after the currently processed elements - e.g.
2400 	 * migrate_pages
2401 	 */
2402 	lru_add_page_tail(head, page_tail, lruvec, list);
2403 }
2404 
2405 static void __split_huge_page(struct page *page, struct list_head *list,
2406 		pgoff_t end)
2407 {
2408 	struct folio *folio = page_folio(page);
2409 	struct page *head = &folio->page;
2410 	struct lruvec *lruvec;
2411 	struct address_space *swap_cache = NULL;
2412 	unsigned long offset = 0;
2413 	unsigned int nr = thp_nr_pages(head);
2414 	int i;
2415 
2416 	/* complete memcg works before add pages to LRU */
2417 	split_page_memcg(head, nr);
2418 
2419 	if (PageAnon(head) && PageSwapCache(head)) {
2420 		swp_entry_t entry = { .val = page_private(head) };
2421 
2422 		offset = swp_offset(entry);
2423 		swap_cache = swap_address_space(entry);
2424 		xa_lock(&swap_cache->i_pages);
2425 	}
2426 
2427 	/* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2428 	lruvec = folio_lruvec_lock(folio);
2429 
2430 	ClearPageHasHWPoisoned(head);
2431 
2432 	for (i = nr - 1; i >= 1; i--) {
2433 		__split_huge_page_tail(head, i, lruvec, list);
2434 		/* Some pages can be beyond EOF: drop them from page cache */
2435 		if (head[i].index >= end) {
2436 			ClearPageDirty(head + i);
2437 			__delete_from_page_cache(head + i, NULL);
2438 			if (shmem_mapping(head->mapping))
2439 				shmem_uncharge(head->mapping->host, 1);
2440 			put_page(head + i);
2441 		} else if (!PageAnon(page)) {
2442 			__xa_store(&head->mapping->i_pages, head[i].index,
2443 					head + i, 0);
2444 		} else if (swap_cache) {
2445 			__xa_store(&swap_cache->i_pages, offset + i,
2446 					head + i, 0);
2447 		}
2448 	}
2449 
2450 	ClearPageCompound(head);
2451 	unlock_page_lruvec(lruvec);
2452 	/* Caller disabled irqs, so they are still disabled here */
2453 
2454 	split_page_owner(head, nr);
2455 
2456 	/* See comment in __split_huge_page_tail() */
2457 	if (PageAnon(head)) {
2458 		/* Additional pin to swap cache */
2459 		if (PageSwapCache(head)) {
2460 			page_ref_add(head, 2);
2461 			xa_unlock(&swap_cache->i_pages);
2462 		} else {
2463 			page_ref_inc(head);
2464 		}
2465 	} else {
2466 		/* Additional pin to page cache */
2467 		page_ref_add(head, 2);
2468 		xa_unlock(&head->mapping->i_pages);
2469 	}
2470 	local_irq_enable();
2471 
2472 	remap_page(head, nr);
2473 
2474 	if (PageSwapCache(head)) {
2475 		swp_entry_t entry = { .val = page_private(head) };
2476 
2477 		split_swap_cluster(entry);
2478 	}
2479 
2480 	for (i = 0; i < nr; i++) {
2481 		struct page *subpage = head + i;
2482 		if (subpage == page)
2483 			continue;
2484 		unlock_page(subpage);
2485 
2486 		/*
2487 		 * Subpages may be freed if there wasn't any mapping
2488 		 * like if add_to_swap() is running on a lru page that
2489 		 * had its mapping zapped. And freeing these pages
2490 		 * requires taking the lru_lock so we do the put_page
2491 		 * of the tail pages after the split is complete.
2492 		 */
2493 		put_page(subpage);
2494 	}
2495 }
2496 
2497 int total_mapcount(struct page *page)
2498 {
2499 	int i, compound, nr, ret;
2500 
2501 	VM_BUG_ON_PAGE(PageTail(page), page);
2502 
2503 	if (likely(!PageCompound(page)))
2504 		return atomic_read(&page->_mapcount) + 1;
2505 
2506 	compound = compound_mapcount(page);
2507 	nr = compound_nr(page);
2508 	if (PageHuge(page))
2509 		return compound;
2510 	ret = compound;
2511 	for (i = 0; i < nr; i++)
2512 		ret += atomic_read(&page[i]._mapcount) + 1;
2513 	/* File pages has compound_mapcount included in _mapcount */
2514 	if (!PageAnon(page))
2515 		return ret - compound * nr;
2516 	if (PageDoubleMap(page))
2517 		ret -= nr;
2518 	return ret;
2519 }
2520 
2521 /*
2522  * This calculates accurately how many mappings a transparent hugepage
2523  * has (unlike page_mapcount() which isn't fully accurate). This full
2524  * accuracy is primarily needed to know if copy-on-write faults can
2525  * reuse the page and change the mapping to read-write instead of
2526  * copying them. At the same time this returns the total_mapcount too.
2527  *
2528  * The function returns the highest mapcount any one of the subpages
2529  * has. If the return value is one, even if different processes are
2530  * mapping different subpages of the transparent hugepage, they can
2531  * all reuse it, because each process is reusing a different subpage.
2532  *
2533  * The total_mapcount is instead counting all virtual mappings of the
2534  * subpages. If the total_mapcount is equal to "one", it tells the
2535  * caller all mappings belong to the same "mm" and in turn the
2536  * anon_vma of the transparent hugepage can become the vma->anon_vma
2537  * local one as no other process may be mapping any of the subpages.
2538  *
2539  * It would be more accurate to replace page_mapcount() with
2540  * page_trans_huge_mapcount(), however we only use
2541  * page_trans_huge_mapcount() in the copy-on-write faults where we
2542  * need full accuracy to avoid breaking page pinning, because
2543  * page_trans_huge_mapcount() is slower than page_mapcount().
2544  */
2545 int page_trans_huge_mapcount(struct page *page)
2546 {
2547 	int i, ret;
2548 
2549 	/* hugetlbfs shouldn't call it */
2550 	VM_BUG_ON_PAGE(PageHuge(page), page);
2551 
2552 	if (likely(!PageTransCompound(page)))
2553 		return atomic_read(&page->_mapcount) + 1;
2554 
2555 	page = compound_head(page);
2556 
2557 	ret = 0;
2558 	for (i = 0; i < thp_nr_pages(page); i++) {
2559 		int mapcount = atomic_read(&page[i]._mapcount) + 1;
2560 		ret = max(ret, mapcount);
2561 	}
2562 
2563 	if (PageDoubleMap(page))
2564 		ret -= 1;
2565 
2566 	return ret + compound_mapcount(page);
2567 }
2568 
2569 /* Racy check whether the huge page can be split */
2570 bool can_split_huge_page(struct page *page, int *pextra_pins)
2571 {
2572 	int extra_pins;
2573 
2574 	/* Additional pins from page cache */
2575 	if (PageAnon(page))
2576 		extra_pins = PageSwapCache(page) ? thp_nr_pages(page) : 0;
2577 	else
2578 		extra_pins = thp_nr_pages(page);
2579 	if (pextra_pins)
2580 		*pextra_pins = extra_pins;
2581 	return total_mapcount(page) == page_count(page) - extra_pins - 1;
2582 }
2583 
2584 /*
2585  * This function splits huge page into normal pages. @page can point to any
2586  * subpage of huge page to split. Split doesn't change the position of @page.
2587  *
2588  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2589  * The huge page must be locked.
2590  *
2591  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2592  *
2593  * Both head page and tail pages will inherit mapping, flags, and so on from
2594  * the hugepage.
2595  *
2596  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2597  * they are not mapped.
2598  *
2599  * Returns 0 if the hugepage is split successfully.
2600  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2601  * us.
2602  */
2603 int split_huge_page_to_list(struct page *page, struct list_head *list)
2604 {
2605 	struct page *head = compound_head(page);
2606 	struct deferred_split *ds_queue = get_deferred_split_queue(head);
2607 	XA_STATE(xas, &head->mapping->i_pages, head->index);
2608 	struct anon_vma *anon_vma = NULL;
2609 	struct address_space *mapping = NULL;
2610 	int extra_pins, ret;
2611 	pgoff_t end;
2612 
2613 	VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2614 	VM_BUG_ON_PAGE(!PageLocked(head), head);
2615 	VM_BUG_ON_PAGE(!PageCompound(head), head);
2616 
2617 	if (PageWriteback(head))
2618 		return -EBUSY;
2619 
2620 	if (PageAnon(head)) {
2621 		/*
2622 		 * The caller does not necessarily hold an mmap_lock that would
2623 		 * prevent the anon_vma disappearing so we first we take a
2624 		 * reference to it and then lock the anon_vma for write. This
2625 		 * is similar to page_lock_anon_vma_read except the write lock
2626 		 * is taken to serialise against parallel split or collapse
2627 		 * operations.
2628 		 */
2629 		anon_vma = page_get_anon_vma(head);
2630 		if (!anon_vma) {
2631 			ret = -EBUSY;
2632 			goto out;
2633 		}
2634 		end = -1;
2635 		mapping = NULL;
2636 		anon_vma_lock_write(anon_vma);
2637 	} else {
2638 		mapping = head->mapping;
2639 
2640 		/* Truncated ? */
2641 		if (!mapping) {
2642 			ret = -EBUSY;
2643 			goto out;
2644 		}
2645 
2646 		xas_split_alloc(&xas, head, compound_order(head),
2647 				mapping_gfp_mask(mapping) & GFP_RECLAIM_MASK);
2648 		if (xas_error(&xas)) {
2649 			ret = xas_error(&xas);
2650 			goto out;
2651 		}
2652 
2653 		anon_vma = NULL;
2654 		i_mmap_lock_read(mapping);
2655 
2656 		/*
2657 		 *__split_huge_page() may need to trim off pages beyond EOF:
2658 		 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2659 		 * which cannot be nested inside the page tree lock. So note
2660 		 * end now: i_size itself may be changed at any moment, but
2661 		 * head page lock is good enough to serialize the trimming.
2662 		 */
2663 		end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2664 		if (shmem_mapping(mapping))
2665 			end = shmem_fallocend(mapping->host, end);
2666 	}
2667 
2668 	/*
2669 	 * Racy check if we can split the page, before unmap_page() will
2670 	 * split PMDs
2671 	 */
2672 	if (!can_split_huge_page(head, &extra_pins)) {
2673 		ret = -EBUSY;
2674 		goto out_unlock;
2675 	}
2676 
2677 	unmap_page(head);
2678 
2679 	/* block interrupt reentry in xa_lock and spinlock */
2680 	local_irq_disable();
2681 	if (mapping) {
2682 		/*
2683 		 * Check if the head page is present in page cache.
2684 		 * We assume all tail are present too, if head is there.
2685 		 */
2686 		xas_lock(&xas);
2687 		xas_reset(&xas);
2688 		if (xas_load(&xas) != head)
2689 			goto fail;
2690 	}
2691 
2692 	/* Prevent deferred_split_scan() touching ->_refcount */
2693 	spin_lock(&ds_queue->split_queue_lock);
2694 	if (page_ref_freeze(head, 1 + extra_pins)) {
2695 		if (!list_empty(page_deferred_list(head))) {
2696 			ds_queue->split_queue_len--;
2697 			list_del(page_deferred_list(head));
2698 		}
2699 		spin_unlock(&ds_queue->split_queue_lock);
2700 		if (mapping) {
2701 			int nr = thp_nr_pages(head);
2702 
2703 			xas_split(&xas, head, thp_order(head));
2704 			if (PageSwapBacked(head)) {
2705 				__mod_lruvec_page_state(head, NR_SHMEM_THPS,
2706 							-nr);
2707 			} else {
2708 				__mod_lruvec_page_state(head, NR_FILE_THPS,
2709 							-nr);
2710 				filemap_nr_thps_dec(mapping);
2711 			}
2712 		}
2713 
2714 		__split_huge_page(page, list, end);
2715 		ret = 0;
2716 	} else {
2717 		spin_unlock(&ds_queue->split_queue_lock);
2718 fail:
2719 		if (mapping)
2720 			xas_unlock(&xas);
2721 		local_irq_enable();
2722 		remap_page(head, thp_nr_pages(head));
2723 		ret = -EBUSY;
2724 	}
2725 
2726 out_unlock:
2727 	if (anon_vma) {
2728 		anon_vma_unlock_write(anon_vma);
2729 		put_anon_vma(anon_vma);
2730 	}
2731 	if (mapping)
2732 		i_mmap_unlock_read(mapping);
2733 out:
2734 	/* Free any memory we didn't use */
2735 	xas_nomem(&xas, 0);
2736 	count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2737 	return ret;
2738 }
2739 
2740 void free_transhuge_page(struct page *page)
2741 {
2742 	struct deferred_split *ds_queue = get_deferred_split_queue(page);
2743 	unsigned long flags;
2744 
2745 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2746 	if (!list_empty(page_deferred_list(page))) {
2747 		ds_queue->split_queue_len--;
2748 		list_del(page_deferred_list(page));
2749 	}
2750 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2751 	free_compound_page(page);
2752 }
2753 
2754 void deferred_split_huge_page(struct page *page)
2755 {
2756 	struct deferred_split *ds_queue = get_deferred_split_queue(page);
2757 #ifdef CONFIG_MEMCG
2758 	struct mem_cgroup *memcg = page_memcg(compound_head(page));
2759 #endif
2760 	unsigned long flags;
2761 
2762 	VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2763 
2764 	/*
2765 	 * The try_to_unmap() in page reclaim path might reach here too,
2766 	 * this may cause a race condition to corrupt deferred split queue.
2767 	 * And, if page reclaim is already handling the same page, it is
2768 	 * unnecessary to handle it again in shrinker.
2769 	 *
2770 	 * Check PageSwapCache to determine if the page is being
2771 	 * handled by page reclaim since THP swap would add the page into
2772 	 * swap cache before calling try_to_unmap().
2773 	 */
2774 	if (PageSwapCache(page))
2775 		return;
2776 
2777 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2778 	if (list_empty(page_deferred_list(page))) {
2779 		count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2780 		list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2781 		ds_queue->split_queue_len++;
2782 #ifdef CONFIG_MEMCG
2783 		if (memcg)
2784 			set_shrinker_bit(memcg, page_to_nid(page),
2785 					 deferred_split_shrinker.id);
2786 #endif
2787 	}
2788 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2789 }
2790 
2791 static unsigned long deferred_split_count(struct shrinker *shrink,
2792 		struct shrink_control *sc)
2793 {
2794 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2795 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2796 
2797 #ifdef CONFIG_MEMCG
2798 	if (sc->memcg)
2799 		ds_queue = &sc->memcg->deferred_split_queue;
2800 #endif
2801 	return READ_ONCE(ds_queue->split_queue_len);
2802 }
2803 
2804 static unsigned long deferred_split_scan(struct shrinker *shrink,
2805 		struct shrink_control *sc)
2806 {
2807 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2808 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2809 	unsigned long flags;
2810 	LIST_HEAD(list), *pos, *next;
2811 	struct page *page;
2812 	int split = 0;
2813 
2814 #ifdef CONFIG_MEMCG
2815 	if (sc->memcg)
2816 		ds_queue = &sc->memcg->deferred_split_queue;
2817 #endif
2818 
2819 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2820 	/* Take pin on all head pages to avoid freeing them under us */
2821 	list_for_each_safe(pos, next, &ds_queue->split_queue) {
2822 		page = list_entry((void *)pos, struct page, deferred_list);
2823 		page = compound_head(page);
2824 		if (get_page_unless_zero(page)) {
2825 			list_move(page_deferred_list(page), &list);
2826 		} else {
2827 			/* We lost race with put_compound_page() */
2828 			list_del_init(page_deferred_list(page));
2829 			ds_queue->split_queue_len--;
2830 		}
2831 		if (!--sc->nr_to_scan)
2832 			break;
2833 	}
2834 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2835 
2836 	list_for_each_safe(pos, next, &list) {
2837 		page = list_entry((void *)pos, struct page, deferred_list);
2838 		if (!trylock_page(page))
2839 			goto next;
2840 		/* split_huge_page() removes page from list on success */
2841 		if (!split_huge_page(page))
2842 			split++;
2843 		unlock_page(page);
2844 next:
2845 		put_page(page);
2846 	}
2847 
2848 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2849 	list_splice_tail(&list, &ds_queue->split_queue);
2850 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2851 
2852 	/*
2853 	 * Stop shrinker if we didn't split any page, but the queue is empty.
2854 	 * This can happen if pages were freed under us.
2855 	 */
2856 	if (!split && list_empty(&ds_queue->split_queue))
2857 		return SHRINK_STOP;
2858 	return split;
2859 }
2860 
2861 static struct shrinker deferred_split_shrinker = {
2862 	.count_objects = deferred_split_count,
2863 	.scan_objects = deferred_split_scan,
2864 	.seeks = DEFAULT_SEEKS,
2865 	.flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2866 		 SHRINKER_NONSLAB,
2867 };
2868 
2869 #ifdef CONFIG_DEBUG_FS
2870 static void split_huge_pages_all(void)
2871 {
2872 	struct zone *zone;
2873 	struct page *page;
2874 	unsigned long pfn, max_zone_pfn;
2875 	unsigned long total = 0, split = 0;
2876 
2877 	pr_debug("Split all THPs\n");
2878 	for_each_populated_zone(zone) {
2879 		max_zone_pfn = zone_end_pfn(zone);
2880 		for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2881 			if (!pfn_valid(pfn))
2882 				continue;
2883 
2884 			page = pfn_to_page(pfn);
2885 			if (!get_page_unless_zero(page))
2886 				continue;
2887 
2888 			if (zone != page_zone(page))
2889 				goto next;
2890 
2891 			if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2892 				goto next;
2893 
2894 			total++;
2895 			lock_page(page);
2896 			if (!split_huge_page(page))
2897 				split++;
2898 			unlock_page(page);
2899 next:
2900 			put_page(page);
2901 			cond_resched();
2902 		}
2903 	}
2904 
2905 	pr_debug("%lu of %lu THP split\n", split, total);
2906 }
2907 
2908 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
2909 {
2910 	return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
2911 		    is_vm_hugetlb_page(vma);
2912 }
2913 
2914 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
2915 				unsigned long vaddr_end)
2916 {
2917 	int ret = 0;
2918 	struct task_struct *task;
2919 	struct mm_struct *mm;
2920 	unsigned long total = 0, split = 0;
2921 	unsigned long addr;
2922 
2923 	vaddr_start &= PAGE_MASK;
2924 	vaddr_end &= PAGE_MASK;
2925 
2926 	/* Find the task_struct from pid */
2927 	rcu_read_lock();
2928 	task = find_task_by_vpid(pid);
2929 	if (!task) {
2930 		rcu_read_unlock();
2931 		ret = -ESRCH;
2932 		goto out;
2933 	}
2934 	get_task_struct(task);
2935 	rcu_read_unlock();
2936 
2937 	/* Find the mm_struct */
2938 	mm = get_task_mm(task);
2939 	put_task_struct(task);
2940 
2941 	if (!mm) {
2942 		ret = -EINVAL;
2943 		goto out;
2944 	}
2945 
2946 	pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
2947 		 pid, vaddr_start, vaddr_end);
2948 
2949 	mmap_read_lock(mm);
2950 	/*
2951 	 * always increase addr by PAGE_SIZE, since we could have a PTE page
2952 	 * table filled with PTE-mapped THPs, each of which is distinct.
2953 	 */
2954 	for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
2955 		struct vm_area_struct *vma = find_vma(mm, addr);
2956 		unsigned int follflags;
2957 		struct page *page;
2958 
2959 		if (!vma || addr < vma->vm_start)
2960 			break;
2961 
2962 		/* skip special VMA and hugetlb VMA */
2963 		if (vma_not_suitable_for_thp_split(vma)) {
2964 			addr = vma->vm_end;
2965 			continue;
2966 		}
2967 
2968 		/* FOLL_DUMP to ignore special (like zero) pages */
2969 		follflags = FOLL_GET | FOLL_DUMP;
2970 		page = follow_page(vma, addr, follflags);
2971 
2972 		if (IS_ERR(page))
2973 			continue;
2974 		if (!page)
2975 			continue;
2976 
2977 		if (!is_transparent_hugepage(page))
2978 			goto next;
2979 
2980 		total++;
2981 		if (!can_split_huge_page(compound_head(page), NULL))
2982 			goto next;
2983 
2984 		if (!trylock_page(page))
2985 			goto next;
2986 
2987 		if (!split_huge_page(page))
2988 			split++;
2989 
2990 		unlock_page(page);
2991 next:
2992 		put_page(page);
2993 		cond_resched();
2994 	}
2995 	mmap_read_unlock(mm);
2996 	mmput(mm);
2997 
2998 	pr_debug("%lu of %lu THP split\n", split, total);
2999 
3000 out:
3001 	return ret;
3002 }
3003 
3004 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
3005 				pgoff_t off_end)
3006 {
3007 	struct filename *file;
3008 	struct file *candidate;
3009 	struct address_space *mapping;
3010 	int ret = -EINVAL;
3011 	pgoff_t index;
3012 	int nr_pages = 1;
3013 	unsigned long total = 0, split = 0;
3014 
3015 	file = getname_kernel(file_path);
3016 	if (IS_ERR(file))
3017 		return ret;
3018 
3019 	candidate = file_open_name(file, O_RDONLY, 0);
3020 	if (IS_ERR(candidate))
3021 		goto out;
3022 
3023 	pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3024 		 file_path, off_start, off_end);
3025 
3026 	mapping = candidate->f_mapping;
3027 
3028 	for (index = off_start; index < off_end; index += nr_pages) {
3029 		struct page *fpage = pagecache_get_page(mapping, index,
3030 						FGP_ENTRY | FGP_HEAD, 0);
3031 
3032 		nr_pages = 1;
3033 		if (xa_is_value(fpage) || !fpage)
3034 			continue;
3035 
3036 		if (!is_transparent_hugepage(fpage))
3037 			goto next;
3038 
3039 		total++;
3040 		nr_pages = thp_nr_pages(fpage);
3041 
3042 		if (!trylock_page(fpage))
3043 			goto next;
3044 
3045 		if (!split_huge_page(fpage))
3046 			split++;
3047 
3048 		unlock_page(fpage);
3049 next:
3050 		put_page(fpage);
3051 		cond_resched();
3052 	}
3053 
3054 	filp_close(candidate, NULL);
3055 	ret = 0;
3056 
3057 	pr_debug("%lu of %lu file-backed THP split\n", split, total);
3058 out:
3059 	putname(file);
3060 	return ret;
3061 }
3062 
3063 #define MAX_INPUT_BUF_SZ 255
3064 
3065 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3066 				size_t count, loff_t *ppops)
3067 {
3068 	static DEFINE_MUTEX(split_debug_mutex);
3069 	ssize_t ret;
3070 	/* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3071 	char input_buf[MAX_INPUT_BUF_SZ];
3072 	int pid;
3073 	unsigned long vaddr_start, vaddr_end;
3074 
3075 	ret = mutex_lock_interruptible(&split_debug_mutex);
3076 	if (ret)
3077 		return ret;
3078 
3079 	ret = -EFAULT;
3080 
3081 	memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3082 	if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3083 		goto out;
3084 
3085 	input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3086 
3087 	if (input_buf[0] == '/') {
3088 		char *tok;
3089 		char *buf = input_buf;
3090 		char file_path[MAX_INPUT_BUF_SZ];
3091 		pgoff_t off_start = 0, off_end = 0;
3092 		size_t input_len = strlen(input_buf);
3093 
3094 		tok = strsep(&buf, ",");
3095 		if (tok) {
3096 			strcpy(file_path, tok);
3097 		} else {
3098 			ret = -EINVAL;
3099 			goto out;
3100 		}
3101 
3102 		ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3103 		if (ret != 2) {
3104 			ret = -EINVAL;
3105 			goto out;
3106 		}
3107 		ret = split_huge_pages_in_file(file_path, off_start, off_end);
3108 		if (!ret)
3109 			ret = input_len;
3110 
3111 		goto out;
3112 	}
3113 
3114 	ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3115 	if (ret == 1 && pid == 1) {
3116 		split_huge_pages_all();
3117 		ret = strlen(input_buf);
3118 		goto out;
3119 	} else if (ret != 3) {
3120 		ret = -EINVAL;
3121 		goto out;
3122 	}
3123 
3124 	ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3125 	if (!ret)
3126 		ret = strlen(input_buf);
3127 out:
3128 	mutex_unlock(&split_debug_mutex);
3129 	return ret;
3130 
3131 }
3132 
3133 static const struct file_operations split_huge_pages_fops = {
3134 	.owner	 = THIS_MODULE,
3135 	.write	 = split_huge_pages_write,
3136 	.llseek  = no_llseek,
3137 };
3138 
3139 static int __init split_huge_pages_debugfs(void)
3140 {
3141 	debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3142 			    &split_huge_pages_fops);
3143 	return 0;
3144 }
3145 late_initcall(split_huge_pages_debugfs);
3146 #endif
3147 
3148 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3149 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3150 		struct page *page)
3151 {
3152 	struct vm_area_struct *vma = pvmw->vma;
3153 	struct mm_struct *mm = vma->vm_mm;
3154 	unsigned long address = pvmw->address;
3155 	pmd_t pmdval;
3156 	swp_entry_t entry;
3157 	pmd_t pmdswp;
3158 
3159 	if (!(pvmw->pmd && !pvmw->pte))
3160 		return;
3161 
3162 	flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3163 	pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3164 	if (pmd_dirty(pmdval))
3165 		set_page_dirty(page);
3166 	if (pmd_write(pmdval))
3167 		entry = make_writable_migration_entry(page_to_pfn(page));
3168 	else
3169 		entry = make_readable_migration_entry(page_to_pfn(page));
3170 	pmdswp = swp_entry_to_pmd(entry);
3171 	if (pmd_soft_dirty(pmdval))
3172 		pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3173 	set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3174 	page_remove_rmap(page, true);
3175 	put_page(page);
3176 }
3177 
3178 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3179 {
3180 	struct vm_area_struct *vma = pvmw->vma;
3181 	struct mm_struct *mm = vma->vm_mm;
3182 	unsigned long address = pvmw->address;
3183 	unsigned long mmun_start = address & HPAGE_PMD_MASK;
3184 	pmd_t pmde;
3185 	swp_entry_t entry;
3186 
3187 	if (!(pvmw->pmd && !pvmw->pte))
3188 		return;
3189 
3190 	entry = pmd_to_swp_entry(*pvmw->pmd);
3191 	get_page(new);
3192 	pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3193 	if (pmd_swp_soft_dirty(*pvmw->pmd))
3194 		pmde = pmd_mksoft_dirty(pmde);
3195 	if (is_writable_migration_entry(entry))
3196 		pmde = maybe_pmd_mkwrite(pmde, vma);
3197 	if (pmd_swp_uffd_wp(*pvmw->pmd))
3198 		pmde = pmd_wrprotect(pmd_mkuffd_wp(pmde));
3199 
3200 	flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3201 	if (PageAnon(new))
3202 		page_add_anon_rmap(new, vma, mmun_start, true);
3203 	else
3204 		page_add_file_rmap(new, true);
3205 	set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3206 	if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3207 		mlock_vma_page(new);
3208 	update_mmu_cache_pmd(vma, address, pvmw->pmd);
3209 }
3210 #endif
3211