xref: /linux/mm/huge_memory.c (revision b05f8d7e077952d14acb63e3ccdf5f64404b59a4)
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/numa_balancing.h>
12 #include <linux/highmem.h>
13 #include <linux/hugetlb.h>
14 #include <linux/mmu_notifier.h>
15 #include <linux/rmap.h>
16 #include <linux/swap.h>
17 #include <linux/shrinker.h>
18 #include <linux/mm_inline.h>
19 #include <linux/swapops.h>
20 #include <linux/backing-dev.h>
21 #include <linux/dax.h>
22 #include <linux/mm_types.h>
23 #include <linux/khugepaged.h>
24 #include <linux/freezer.h>
25 #include <linux/pfn_t.h>
26 #include <linux/mman.h>
27 #include <linux/memremap.h>
28 #include <linux/pagemap.h>
29 #include <linux/debugfs.h>
30 #include <linux/migrate.h>
31 #include <linux/hashtable.h>
32 #include <linux/userfaultfd_k.h>
33 #include <linux/page_idle.h>
34 #include <linux/shmem_fs.h>
35 #include <linux/oom.h>
36 #include <linux/numa.h>
37 #include <linux/page_owner.h>
38 #include <linux/sched/sysctl.h>
39 #include <linux/memory-tiers.h>
40 #include <linux/compat.h>
41 #include <linux/pgalloc_tag.h>
42 #include <linux/pagewalk.h>
43 
44 #include <asm/tlb.h>
45 #include <asm/pgalloc.h>
46 #include "internal.h"
47 #include "swap.h"
48 
49 #define CREATE_TRACE_POINTS
50 #include <trace/events/thp.h>
51 
52 /*
53  * By default, transparent hugepage support is disabled in order to avoid
54  * risking an increased memory footprint for applications that are not
55  * guaranteed to benefit from it. When transparent hugepage support is
56  * enabled, it is for all mappings, and khugepaged scans all mappings.
57  * Defrag is invoked by khugepaged hugepage allocations and by page faults
58  * for all hugepage allocations.
59  */
60 unsigned long transparent_hugepage_flags __read_mostly =
61 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
62 	(1<<TRANSPARENT_HUGEPAGE_FLAG)|
63 #endif
64 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
65 	(1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
66 #endif
67 	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
68 	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
69 	(1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
70 
71 static struct shrinker *deferred_split_shrinker;
72 static unsigned long deferred_split_count(struct shrinker *shrink,
73 					  struct shrink_control *sc);
74 static unsigned long deferred_split_scan(struct shrinker *shrink,
75 					 struct shrink_control *sc);
76 static bool split_underused_thp = true;
77 
78 static atomic_t huge_zero_refcount;
79 struct folio *huge_zero_folio __read_mostly;
80 unsigned long huge_zero_pfn __read_mostly = ~0UL;
81 unsigned long huge_anon_orders_always __read_mostly;
82 unsigned long huge_anon_orders_madvise __read_mostly;
83 unsigned long huge_anon_orders_inherit __read_mostly;
84 static bool anon_orders_configured __initdata;
85 
86 static inline bool file_thp_enabled(struct vm_area_struct *vma)
87 {
88 	struct inode *inode;
89 
90 	if (!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS))
91 		return false;
92 
93 	if (!vma->vm_file)
94 		return false;
95 
96 	inode = file_inode(vma->vm_file);
97 
98 	return !inode_is_open_for_write(inode) && S_ISREG(inode->i_mode);
99 }
100 
101 unsigned long __thp_vma_allowable_orders(struct vm_area_struct *vma,
102 					 unsigned long vm_flags,
103 					 unsigned long tva_flags,
104 					 unsigned long orders)
105 {
106 	bool smaps = tva_flags & TVA_SMAPS;
107 	bool in_pf = tva_flags & TVA_IN_PF;
108 	bool enforce_sysfs = tva_flags & TVA_ENFORCE_SYSFS;
109 	unsigned long supported_orders;
110 
111 	/* Check the intersection of requested and supported orders. */
112 	if (vma_is_anonymous(vma))
113 		supported_orders = THP_ORDERS_ALL_ANON;
114 	else if (vma_is_special_huge(vma))
115 		supported_orders = THP_ORDERS_ALL_SPECIAL;
116 	else
117 		supported_orders = THP_ORDERS_ALL_FILE_DEFAULT;
118 
119 	orders &= supported_orders;
120 	if (!orders)
121 		return 0;
122 
123 	if (!vma->vm_mm)		/* vdso */
124 		return 0;
125 
126 	if (thp_disabled_by_hw() || vma_thp_disabled(vma, vm_flags))
127 		return 0;
128 
129 	/* khugepaged doesn't collapse DAX vma, but page fault is fine. */
130 	if (vma_is_dax(vma))
131 		return in_pf ? orders : 0;
132 
133 	/*
134 	 * khugepaged special VMA and hugetlb VMA.
135 	 * Must be checked after dax since some dax mappings may have
136 	 * VM_MIXEDMAP set.
137 	 */
138 	if (!in_pf && !smaps && (vm_flags & VM_NO_KHUGEPAGED))
139 		return 0;
140 
141 	/*
142 	 * Check alignment for file vma and size for both file and anon vma by
143 	 * filtering out the unsuitable orders.
144 	 *
145 	 * Skip the check for page fault. Huge fault does the check in fault
146 	 * handlers.
147 	 */
148 	if (!in_pf) {
149 		int order = highest_order(orders);
150 		unsigned long addr;
151 
152 		while (orders) {
153 			addr = vma->vm_end - (PAGE_SIZE << order);
154 			if (thp_vma_suitable_order(vma, addr, order))
155 				break;
156 			order = next_order(&orders, order);
157 		}
158 
159 		if (!orders)
160 			return 0;
161 	}
162 
163 	/*
164 	 * Enabled via shmem mount options or sysfs settings.
165 	 * Must be done before hugepage flags check since shmem has its
166 	 * own flags.
167 	 */
168 	if (!in_pf && shmem_file(vma->vm_file))
169 		return shmem_allowable_huge_orders(file_inode(vma->vm_file),
170 						   vma, vma->vm_pgoff, 0,
171 						   !enforce_sysfs);
172 
173 	if (!vma_is_anonymous(vma)) {
174 		/*
175 		 * Enforce sysfs THP requirements as necessary. Anonymous vmas
176 		 * were already handled in thp_vma_allowable_orders().
177 		 */
178 		if (enforce_sysfs &&
179 		    (!hugepage_global_enabled() || (!(vm_flags & VM_HUGEPAGE) &&
180 						    !hugepage_global_always())))
181 			return 0;
182 
183 		/*
184 		 * Trust that ->huge_fault() handlers know what they are doing
185 		 * in fault path.
186 		 */
187 		if (((in_pf || smaps)) && vma->vm_ops->huge_fault)
188 			return orders;
189 		/* Only regular file is valid in collapse path */
190 		if (((!in_pf || smaps)) && file_thp_enabled(vma))
191 			return orders;
192 		return 0;
193 	}
194 
195 	if (vma_is_temporary_stack(vma))
196 		return 0;
197 
198 	/*
199 	 * THPeligible bit of smaps should show 1 for proper VMAs even
200 	 * though anon_vma is not initialized yet.
201 	 *
202 	 * Allow page fault since anon_vma may be not initialized until
203 	 * the first page fault.
204 	 */
205 	if (!vma->anon_vma)
206 		return (smaps || in_pf) ? orders : 0;
207 
208 	return orders;
209 }
210 
211 static bool get_huge_zero_page(void)
212 {
213 	struct folio *zero_folio;
214 retry:
215 	if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
216 		return true;
217 
218 	zero_folio = folio_alloc((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
219 			HPAGE_PMD_ORDER);
220 	if (!zero_folio) {
221 		count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
222 		return false;
223 	}
224 	/* Ensure zero folio won't have large_rmappable flag set. */
225 	folio_clear_large_rmappable(zero_folio);
226 	preempt_disable();
227 	if (cmpxchg(&huge_zero_folio, NULL, zero_folio)) {
228 		preempt_enable();
229 		folio_put(zero_folio);
230 		goto retry;
231 	}
232 	WRITE_ONCE(huge_zero_pfn, folio_pfn(zero_folio));
233 
234 	/* We take additional reference here. It will be put back by shrinker */
235 	atomic_set(&huge_zero_refcount, 2);
236 	preempt_enable();
237 	count_vm_event(THP_ZERO_PAGE_ALLOC);
238 	return true;
239 }
240 
241 static void put_huge_zero_page(void)
242 {
243 	/*
244 	 * Counter should never go to zero here. Only shrinker can put
245 	 * last reference.
246 	 */
247 	BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
248 }
249 
250 struct folio *mm_get_huge_zero_folio(struct mm_struct *mm)
251 {
252 	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
253 		return READ_ONCE(huge_zero_folio);
254 
255 	if (!get_huge_zero_page())
256 		return NULL;
257 
258 	if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
259 		put_huge_zero_page();
260 
261 	return READ_ONCE(huge_zero_folio);
262 }
263 
264 void mm_put_huge_zero_folio(struct mm_struct *mm)
265 {
266 	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
267 		put_huge_zero_page();
268 }
269 
270 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
271 					struct shrink_control *sc)
272 {
273 	/* we can free zero page only if last reference remains */
274 	return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
275 }
276 
277 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
278 				       struct shrink_control *sc)
279 {
280 	if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
281 		struct folio *zero_folio = xchg(&huge_zero_folio, NULL);
282 		BUG_ON(zero_folio == NULL);
283 		WRITE_ONCE(huge_zero_pfn, ~0UL);
284 		folio_put(zero_folio);
285 		return HPAGE_PMD_NR;
286 	}
287 
288 	return 0;
289 }
290 
291 static struct shrinker *huge_zero_page_shrinker;
292 
293 #ifdef CONFIG_SYSFS
294 static ssize_t enabled_show(struct kobject *kobj,
295 			    struct kobj_attribute *attr, char *buf)
296 {
297 	const char *output;
298 
299 	if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
300 		output = "[always] madvise never";
301 	else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
302 			  &transparent_hugepage_flags))
303 		output = "always [madvise] never";
304 	else
305 		output = "always madvise [never]";
306 
307 	return sysfs_emit(buf, "%s\n", output);
308 }
309 
310 static ssize_t enabled_store(struct kobject *kobj,
311 			     struct kobj_attribute *attr,
312 			     const char *buf, size_t count)
313 {
314 	ssize_t ret = count;
315 
316 	if (sysfs_streq(buf, "always")) {
317 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
318 		set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
319 	} else if (sysfs_streq(buf, "madvise")) {
320 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
321 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
322 	} else if (sysfs_streq(buf, "never")) {
323 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
324 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
325 	} else
326 		ret = -EINVAL;
327 
328 	if (ret > 0) {
329 		int err = start_stop_khugepaged();
330 		if (err)
331 			ret = err;
332 	}
333 	return ret;
334 }
335 
336 static struct kobj_attribute enabled_attr = __ATTR_RW(enabled);
337 
338 ssize_t single_hugepage_flag_show(struct kobject *kobj,
339 				  struct kobj_attribute *attr, char *buf,
340 				  enum transparent_hugepage_flag flag)
341 {
342 	return sysfs_emit(buf, "%d\n",
343 			  !!test_bit(flag, &transparent_hugepage_flags));
344 }
345 
346 ssize_t single_hugepage_flag_store(struct kobject *kobj,
347 				 struct kobj_attribute *attr,
348 				 const char *buf, size_t count,
349 				 enum transparent_hugepage_flag flag)
350 {
351 	unsigned long value;
352 	int ret;
353 
354 	ret = kstrtoul(buf, 10, &value);
355 	if (ret < 0)
356 		return ret;
357 	if (value > 1)
358 		return -EINVAL;
359 
360 	if (value)
361 		set_bit(flag, &transparent_hugepage_flags);
362 	else
363 		clear_bit(flag, &transparent_hugepage_flags);
364 
365 	return count;
366 }
367 
368 static ssize_t defrag_show(struct kobject *kobj,
369 			   struct kobj_attribute *attr, char *buf)
370 {
371 	const char *output;
372 
373 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
374 		     &transparent_hugepage_flags))
375 		output = "[always] defer defer+madvise madvise never";
376 	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
377 			  &transparent_hugepage_flags))
378 		output = "always [defer] defer+madvise madvise never";
379 	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
380 			  &transparent_hugepage_flags))
381 		output = "always defer [defer+madvise] madvise never";
382 	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
383 			  &transparent_hugepage_flags))
384 		output = "always defer defer+madvise [madvise] never";
385 	else
386 		output = "always defer defer+madvise madvise [never]";
387 
388 	return sysfs_emit(buf, "%s\n", output);
389 }
390 
391 static ssize_t defrag_store(struct kobject *kobj,
392 			    struct kobj_attribute *attr,
393 			    const char *buf, size_t count)
394 {
395 	if (sysfs_streq(buf, "always")) {
396 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
397 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
398 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
399 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
400 	} else if (sysfs_streq(buf, "defer+madvise")) {
401 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
402 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
403 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
404 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
405 	} else if (sysfs_streq(buf, "defer")) {
406 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
407 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
408 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
409 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
410 	} else if (sysfs_streq(buf, "madvise")) {
411 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
412 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
413 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
414 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
415 	} else if (sysfs_streq(buf, "never")) {
416 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
417 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
418 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
419 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
420 	} else
421 		return -EINVAL;
422 
423 	return count;
424 }
425 static struct kobj_attribute defrag_attr = __ATTR_RW(defrag);
426 
427 static ssize_t use_zero_page_show(struct kobject *kobj,
428 				  struct kobj_attribute *attr, char *buf)
429 {
430 	return single_hugepage_flag_show(kobj, attr, buf,
431 					 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
432 }
433 static ssize_t use_zero_page_store(struct kobject *kobj,
434 		struct kobj_attribute *attr, const char *buf, size_t count)
435 {
436 	return single_hugepage_flag_store(kobj, attr, buf, count,
437 				 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
438 }
439 static struct kobj_attribute use_zero_page_attr = __ATTR_RW(use_zero_page);
440 
441 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
442 				   struct kobj_attribute *attr, char *buf)
443 {
444 	return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
445 }
446 static struct kobj_attribute hpage_pmd_size_attr =
447 	__ATTR_RO(hpage_pmd_size);
448 
449 static ssize_t split_underused_thp_show(struct kobject *kobj,
450 			    struct kobj_attribute *attr, char *buf)
451 {
452 	return sysfs_emit(buf, "%d\n", split_underused_thp);
453 }
454 
455 static ssize_t split_underused_thp_store(struct kobject *kobj,
456 			     struct kobj_attribute *attr,
457 			     const char *buf, size_t count)
458 {
459 	int err = kstrtobool(buf, &split_underused_thp);
460 
461 	if (err < 0)
462 		return err;
463 
464 	return count;
465 }
466 
467 static struct kobj_attribute split_underused_thp_attr = __ATTR(
468 	shrink_underused, 0644, split_underused_thp_show, split_underused_thp_store);
469 
470 static struct attribute *hugepage_attr[] = {
471 	&enabled_attr.attr,
472 	&defrag_attr.attr,
473 	&use_zero_page_attr.attr,
474 	&hpage_pmd_size_attr.attr,
475 #ifdef CONFIG_SHMEM
476 	&shmem_enabled_attr.attr,
477 #endif
478 	&split_underused_thp_attr.attr,
479 	NULL,
480 };
481 
482 static const struct attribute_group hugepage_attr_group = {
483 	.attrs = hugepage_attr,
484 };
485 
486 static void hugepage_exit_sysfs(struct kobject *hugepage_kobj);
487 static void thpsize_release(struct kobject *kobj);
488 static DEFINE_SPINLOCK(huge_anon_orders_lock);
489 static LIST_HEAD(thpsize_list);
490 
491 static ssize_t anon_enabled_show(struct kobject *kobj,
492 				 struct kobj_attribute *attr, char *buf)
493 {
494 	int order = to_thpsize(kobj)->order;
495 	const char *output;
496 
497 	if (test_bit(order, &huge_anon_orders_always))
498 		output = "[always] inherit madvise never";
499 	else if (test_bit(order, &huge_anon_orders_inherit))
500 		output = "always [inherit] madvise never";
501 	else if (test_bit(order, &huge_anon_orders_madvise))
502 		output = "always inherit [madvise] never";
503 	else
504 		output = "always inherit madvise [never]";
505 
506 	return sysfs_emit(buf, "%s\n", output);
507 }
508 
509 static ssize_t anon_enabled_store(struct kobject *kobj,
510 				  struct kobj_attribute *attr,
511 				  const char *buf, size_t count)
512 {
513 	int order = to_thpsize(kobj)->order;
514 	ssize_t ret = count;
515 
516 	if (sysfs_streq(buf, "always")) {
517 		spin_lock(&huge_anon_orders_lock);
518 		clear_bit(order, &huge_anon_orders_inherit);
519 		clear_bit(order, &huge_anon_orders_madvise);
520 		set_bit(order, &huge_anon_orders_always);
521 		spin_unlock(&huge_anon_orders_lock);
522 	} else if (sysfs_streq(buf, "inherit")) {
523 		spin_lock(&huge_anon_orders_lock);
524 		clear_bit(order, &huge_anon_orders_always);
525 		clear_bit(order, &huge_anon_orders_madvise);
526 		set_bit(order, &huge_anon_orders_inherit);
527 		spin_unlock(&huge_anon_orders_lock);
528 	} else if (sysfs_streq(buf, "madvise")) {
529 		spin_lock(&huge_anon_orders_lock);
530 		clear_bit(order, &huge_anon_orders_always);
531 		clear_bit(order, &huge_anon_orders_inherit);
532 		set_bit(order, &huge_anon_orders_madvise);
533 		spin_unlock(&huge_anon_orders_lock);
534 	} else if (sysfs_streq(buf, "never")) {
535 		spin_lock(&huge_anon_orders_lock);
536 		clear_bit(order, &huge_anon_orders_always);
537 		clear_bit(order, &huge_anon_orders_inherit);
538 		clear_bit(order, &huge_anon_orders_madvise);
539 		spin_unlock(&huge_anon_orders_lock);
540 	} else
541 		ret = -EINVAL;
542 
543 	if (ret > 0) {
544 		int err;
545 
546 		err = start_stop_khugepaged();
547 		if (err)
548 			ret = err;
549 	}
550 	return ret;
551 }
552 
553 static struct kobj_attribute anon_enabled_attr =
554 	__ATTR(enabled, 0644, anon_enabled_show, anon_enabled_store);
555 
556 static struct attribute *anon_ctrl_attrs[] = {
557 	&anon_enabled_attr.attr,
558 	NULL,
559 };
560 
561 static const struct attribute_group anon_ctrl_attr_grp = {
562 	.attrs = anon_ctrl_attrs,
563 };
564 
565 static struct attribute *file_ctrl_attrs[] = {
566 #ifdef CONFIG_SHMEM
567 	&thpsize_shmem_enabled_attr.attr,
568 #endif
569 	NULL,
570 };
571 
572 static const struct attribute_group file_ctrl_attr_grp = {
573 	.attrs = file_ctrl_attrs,
574 };
575 
576 static struct attribute *any_ctrl_attrs[] = {
577 	NULL,
578 };
579 
580 static const struct attribute_group any_ctrl_attr_grp = {
581 	.attrs = any_ctrl_attrs,
582 };
583 
584 static const struct kobj_type thpsize_ktype = {
585 	.release = &thpsize_release,
586 	.sysfs_ops = &kobj_sysfs_ops,
587 };
588 
589 DEFINE_PER_CPU(struct mthp_stat, mthp_stats) = {{{0}}};
590 
591 static unsigned long sum_mthp_stat(int order, enum mthp_stat_item item)
592 {
593 	unsigned long sum = 0;
594 	int cpu;
595 
596 	for_each_possible_cpu(cpu) {
597 		struct mthp_stat *this = &per_cpu(mthp_stats, cpu);
598 
599 		sum += this->stats[order][item];
600 	}
601 
602 	return sum;
603 }
604 
605 #define DEFINE_MTHP_STAT_ATTR(_name, _index)				\
606 static ssize_t _name##_show(struct kobject *kobj,			\
607 			struct kobj_attribute *attr, char *buf)		\
608 {									\
609 	int order = to_thpsize(kobj)->order;				\
610 									\
611 	return sysfs_emit(buf, "%lu\n", sum_mthp_stat(order, _index));	\
612 }									\
613 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
614 
615 DEFINE_MTHP_STAT_ATTR(anon_fault_alloc, MTHP_STAT_ANON_FAULT_ALLOC);
616 DEFINE_MTHP_STAT_ATTR(anon_fault_fallback, MTHP_STAT_ANON_FAULT_FALLBACK);
617 DEFINE_MTHP_STAT_ATTR(anon_fault_fallback_charge, MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE);
618 DEFINE_MTHP_STAT_ATTR(zswpout, MTHP_STAT_ZSWPOUT);
619 DEFINE_MTHP_STAT_ATTR(swpin, MTHP_STAT_SWPIN);
620 DEFINE_MTHP_STAT_ATTR(swpin_fallback, MTHP_STAT_SWPIN_FALLBACK);
621 DEFINE_MTHP_STAT_ATTR(swpin_fallback_charge, MTHP_STAT_SWPIN_FALLBACK_CHARGE);
622 DEFINE_MTHP_STAT_ATTR(swpout, MTHP_STAT_SWPOUT);
623 DEFINE_MTHP_STAT_ATTR(swpout_fallback, MTHP_STAT_SWPOUT_FALLBACK);
624 #ifdef CONFIG_SHMEM
625 DEFINE_MTHP_STAT_ATTR(shmem_alloc, MTHP_STAT_SHMEM_ALLOC);
626 DEFINE_MTHP_STAT_ATTR(shmem_fallback, MTHP_STAT_SHMEM_FALLBACK);
627 DEFINE_MTHP_STAT_ATTR(shmem_fallback_charge, MTHP_STAT_SHMEM_FALLBACK_CHARGE);
628 #endif
629 DEFINE_MTHP_STAT_ATTR(split, MTHP_STAT_SPLIT);
630 DEFINE_MTHP_STAT_ATTR(split_failed, MTHP_STAT_SPLIT_FAILED);
631 DEFINE_MTHP_STAT_ATTR(split_deferred, MTHP_STAT_SPLIT_DEFERRED);
632 DEFINE_MTHP_STAT_ATTR(nr_anon, MTHP_STAT_NR_ANON);
633 DEFINE_MTHP_STAT_ATTR(nr_anon_partially_mapped, MTHP_STAT_NR_ANON_PARTIALLY_MAPPED);
634 
635 static struct attribute *anon_stats_attrs[] = {
636 	&anon_fault_alloc_attr.attr,
637 	&anon_fault_fallback_attr.attr,
638 	&anon_fault_fallback_charge_attr.attr,
639 #ifndef CONFIG_SHMEM
640 	&zswpout_attr.attr,
641 	&swpin_attr.attr,
642 	&swpin_fallback_attr.attr,
643 	&swpin_fallback_charge_attr.attr,
644 	&swpout_attr.attr,
645 	&swpout_fallback_attr.attr,
646 #endif
647 	&split_deferred_attr.attr,
648 	&nr_anon_attr.attr,
649 	&nr_anon_partially_mapped_attr.attr,
650 	NULL,
651 };
652 
653 static struct attribute_group anon_stats_attr_grp = {
654 	.name = "stats",
655 	.attrs = anon_stats_attrs,
656 };
657 
658 static struct attribute *file_stats_attrs[] = {
659 #ifdef CONFIG_SHMEM
660 	&shmem_alloc_attr.attr,
661 	&shmem_fallback_attr.attr,
662 	&shmem_fallback_charge_attr.attr,
663 #endif
664 	NULL,
665 };
666 
667 static struct attribute_group file_stats_attr_grp = {
668 	.name = "stats",
669 	.attrs = file_stats_attrs,
670 };
671 
672 static struct attribute *any_stats_attrs[] = {
673 #ifdef CONFIG_SHMEM
674 	&zswpout_attr.attr,
675 	&swpin_attr.attr,
676 	&swpin_fallback_attr.attr,
677 	&swpin_fallback_charge_attr.attr,
678 	&swpout_attr.attr,
679 	&swpout_fallback_attr.attr,
680 #endif
681 	&split_attr.attr,
682 	&split_failed_attr.attr,
683 	NULL,
684 };
685 
686 static struct attribute_group any_stats_attr_grp = {
687 	.name = "stats",
688 	.attrs = any_stats_attrs,
689 };
690 
691 static int sysfs_add_group(struct kobject *kobj,
692 			   const struct attribute_group *grp)
693 {
694 	int ret = -ENOENT;
695 
696 	/*
697 	 * If the group is named, try to merge first, assuming the subdirectory
698 	 * was already created. This avoids the warning emitted by
699 	 * sysfs_create_group() if the directory already exists.
700 	 */
701 	if (grp->name)
702 		ret = sysfs_merge_group(kobj, grp);
703 	if (ret)
704 		ret = sysfs_create_group(kobj, grp);
705 
706 	return ret;
707 }
708 
709 static struct thpsize *thpsize_create(int order, struct kobject *parent)
710 {
711 	unsigned long size = (PAGE_SIZE << order) / SZ_1K;
712 	struct thpsize *thpsize;
713 	int ret = -ENOMEM;
714 
715 	thpsize = kzalloc(sizeof(*thpsize), GFP_KERNEL);
716 	if (!thpsize)
717 		goto err;
718 
719 	thpsize->order = order;
720 
721 	ret = kobject_init_and_add(&thpsize->kobj, &thpsize_ktype, parent,
722 				   "hugepages-%lukB", size);
723 	if (ret) {
724 		kfree(thpsize);
725 		goto err;
726 	}
727 
728 
729 	ret = sysfs_add_group(&thpsize->kobj, &any_ctrl_attr_grp);
730 	if (ret)
731 		goto err_put;
732 
733 	ret = sysfs_add_group(&thpsize->kobj, &any_stats_attr_grp);
734 	if (ret)
735 		goto err_put;
736 
737 	if (BIT(order) & THP_ORDERS_ALL_ANON) {
738 		ret = sysfs_add_group(&thpsize->kobj, &anon_ctrl_attr_grp);
739 		if (ret)
740 			goto err_put;
741 
742 		ret = sysfs_add_group(&thpsize->kobj, &anon_stats_attr_grp);
743 		if (ret)
744 			goto err_put;
745 	}
746 
747 	if (BIT(order) & THP_ORDERS_ALL_FILE_DEFAULT) {
748 		ret = sysfs_add_group(&thpsize->kobj, &file_ctrl_attr_grp);
749 		if (ret)
750 			goto err_put;
751 
752 		ret = sysfs_add_group(&thpsize->kobj, &file_stats_attr_grp);
753 		if (ret)
754 			goto err_put;
755 	}
756 
757 	return thpsize;
758 err_put:
759 	kobject_put(&thpsize->kobj);
760 err:
761 	return ERR_PTR(ret);
762 }
763 
764 static void thpsize_release(struct kobject *kobj)
765 {
766 	kfree(to_thpsize(kobj));
767 }
768 
769 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
770 {
771 	int err;
772 	struct thpsize *thpsize;
773 	unsigned long orders;
774 	int order;
775 
776 	/*
777 	 * Default to setting PMD-sized THP to inherit the global setting and
778 	 * disable all other sizes. powerpc's PMD_ORDER isn't a compile-time
779 	 * constant so we have to do this here.
780 	 */
781 	if (!anon_orders_configured)
782 		huge_anon_orders_inherit = BIT(PMD_ORDER);
783 
784 	*hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
785 	if (unlikely(!*hugepage_kobj)) {
786 		pr_err("failed to create transparent hugepage kobject\n");
787 		return -ENOMEM;
788 	}
789 
790 	err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
791 	if (err) {
792 		pr_err("failed to register transparent hugepage group\n");
793 		goto delete_obj;
794 	}
795 
796 	err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
797 	if (err) {
798 		pr_err("failed to register transparent hugepage group\n");
799 		goto remove_hp_group;
800 	}
801 
802 	orders = THP_ORDERS_ALL_ANON | THP_ORDERS_ALL_FILE_DEFAULT;
803 	order = highest_order(orders);
804 	while (orders) {
805 		thpsize = thpsize_create(order, *hugepage_kobj);
806 		if (IS_ERR(thpsize)) {
807 			pr_err("failed to create thpsize for order %d\n", order);
808 			err = PTR_ERR(thpsize);
809 			goto remove_all;
810 		}
811 		list_add(&thpsize->node, &thpsize_list);
812 		order = next_order(&orders, order);
813 	}
814 
815 	return 0;
816 
817 remove_all:
818 	hugepage_exit_sysfs(*hugepage_kobj);
819 	return err;
820 remove_hp_group:
821 	sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
822 delete_obj:
823 	kobject_put(*hugepage_kobj);
824 	return err;
825 }
826 
827 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
828 {
829 	struct thpsize *thpsize, *tmp;
830 
831 	list_for_each_entry_safe(thpsize, tmp, &thpsize_list, node) {
832 		list_del(&thpsize->node);
833 		kobject_put(&thpsize->kobj);
834 	}
835 
836 	sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
837 	sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
838 	kobject_put(hugepage_kobj);
839 }
840 #else
841 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
842 {
843 	return 0;
844 }
845 
846 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
847 {
848 }
849 #endif /* CONFIG_SYSFS */
850 
851 static int __init thp_shrinker_init(void)
852 {
853 	huge_zero_page_shrinker = shrinker_alloc(0, "thp-zero");
854 	if (!huge_zero_page_shrinker)
855 		return -ENOMEM;
856 
857 	deferred_split_shrinker = shrinker_alloc(SHRINKER_NUMA_AWARE |
858 						 SHRINKER_MEMCG_AWARE |
859 						 SHRINKER_NONSLAB,
860 						 "thp-deferred_split");
861 	if (!deferred_split_shrinker) {
862 		shrinker_free(huge_zero_page_shrinker);
863 		return -ENOMEM;
864 	}
865 
866 	huge_zero_page_shrinker->count_objects = shrink_huge_zero_page_count;
867 	huge_zero_page_shrinker->scan_objects = shrink_huge_zero_page_scan;
868 	shrinker_register(huge_zero_page_shrinker);
869 
870 	deferred_split_shrinker->count_objects = deferred_split_count;
871 	deferred_split_shrinker->scan_objects = deferred_split_scan;
872 	shrinker_register(deferred_split_shrinker);
873 
874 	return 0;
875 }
876 
877 static void __init thp_shrinker_exit(void)
878 {
879 	shrinker_free(huge_zero_page_shrinker);
880 	shrinker_free(deferred_split_shrinker);
881 }
882 
883 static int __init hugepage_init(void)
884 {
885 	int err;
886 	struct kobject *hugepage_kobj;
887 
888 	if (!has_transparent_hugepage()) {
889 		transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_UNSUPPORTED;
890 		return -EINVAL;
891 	}
892 
893 	/*
894 	 * hugepages can't be allocated by the buddy allocator
895 	 */
896 	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER > MAX_PAGE_ORDER);
897 
898 	err = hugepage_init_sysfs(&hugepage_kobj);
899 	if (err)
900 		goto err_sysfs;
901 
902 	err = khugepaged_init();
903 	if (err)
904 		goto err_slab;
905 
906 	err = thp_shrinker_init();
907 	if (err)
908 		goto err_shrinker;
909 
910 	/*
911 	 * By default disable transparent hugepages on smaller systems,
912 	 * where the extra memory used could hurt more than TLB overhead
913 	 * is likely to save.  The admin can still enable it through /sys.
914 	 */
915 	if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
916 		transparent_hugepage_flags = 0;
917 		return 0;
918 	}
919 
920 	err = start_stop_khugepaged();
921 	if (err)
922 		goto err_khugepaged;
923 
924 	return 0;
925 err_khugepaged:
926 	thp_shrinker_exit();
927 err_shrinker:
928 	khugepaged_destroy();
929 err_slab:
930 	hugepage_exit_sysfs(hugepage_kobj);
931 err_sysfs:
932 	return err;
933 }
934 subsys_initcall(hugepage_init);
935 
936 static int __init setup_transparent_hugepage(char *str)
937 {
938 	int ret = 0;
939 	if (!str)
940 		goto out;
941 	if (!strcmp(str, "always")) {
942 		set_bit(TRANSPARENT_HUGEPAGE_FLAG,
943 			&transparent_hugepage_flags);
944 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
945 			  &transparent_hugepage_flags);
946 		ret = 1;
947 	} else if (!strcmp(str, "madvise")) {
948 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
949 			  &transparent_hugepage_flags);
950 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
951 			&transparent_hugepage_flags);
952 		ret = 1;
953 	} else if (!strcmp(str, "never")) {
954 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
955 			  &transparent_hugepage_flags);
956 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
957 			  &transparent_hugepage_flags);
958 		ret = 1;
959 	}
960 out:
961 	if (!ret)
962 		pr_warn("transparent_hugepage= cannot parse, ignored\n");
963 	return ret;
964 }
965 __setup("transparent_hugepage=", setup_transparent_hugepage);
966 
967 static char str_dup[PAGE_SIZE] __initdata;
968 static int __init setup_thp_anon(char *str)
969 {
970 	char *token, *range, *policy, *subtoken;
971 	unsigned long always, inherit, madvise;
972 	char *start_size, *end_size;
973 	int start, end, nr;
974 	char *p;
975 
976 	if (!str || strlen(str) + 1 > PAGE_SIZE)
977 		goto err;
978 	strscpy(str_dup, str);
979 
980 	always = huge_anon_orders_always;
981 	madvise = huge_anon_orders_madvise;
982 	inherit = huge_anon_orders_inherit;
983 	p = str_dup;
984 	while ((token = strsep(&p, ";")) != NULL) {
985 		range = strsep(&token, ":");
986 		policy = token;
987 
988 		if (!policy)
989 			goto err;
990 
991 		while ((subtoken = strsep(&range, ",")) != NULL) {
992 			if (strchr(subtoken, '-')) {
993 				start_size = strsep(&subtoken, "-");
994 				end_size = subtoken;
995 
996 				start = get_order_from_str(start_size, THP_ORDERS_ALL_ANON);
997 				end = get_order_from_str(end_size, THP_ORDERS_ALL_ANON);
998 			} else {
999 				start_size = end_size = subtoken;
1000 				start = end = get_order_from_str(subtoken,
1001 								 THP_ORDERS_ALL_ANON);
1002 			}
1003 
1004 			if (start == -EINVAL) {
1005 				pr_err("invalid size %s in thp_anon boot parameter\n", start_size);
1006 				goto err;
1007 			}
1008 
1009 			if (end == -EINVAL) {
1010 				pr_err("invalid size %s in thp_anon boot parameter\n", end_size);
1011 				goto err;
1012 			}
1013 
1014 			if (start < 0 || end < 0 || start > end)
1015 				goto err;
1016 
1017 			nr = end - start + 1;
1018 			if (!strcmp(policy, "always")) {
1019 				bitmap_set(&always, start, nr);
1020 				bitmap_clear(&inherit, start, nr);
1021 				bitmap_clear(&madvise, start, nr);
1022 			} else if (!strcmp(policy, "madvise")) {
1023 				bitmap_set(&madvise, start, nr);
1024 				bitmap_clear(&inherit, start, nr);
1025 				bitmap_clear(&always, start, nr);
1026 			} else if (!strcmp(policy, "inherit")) {
1027 				bitmap_set(&inherit, start, nr);
1028 				bitmap_clear(&madvise, start, nr);
1029 				bitmap_clear(&always, start, nr);
1030 			} else if (!strcmp(policy, "never")) {
1031 				bitmap_clear(&inherit, start, nr);
1032 				bitmap_clear(&madvise, start, nr);
1033 				bitmap_clear(&always, start, nr);
1034 			} else {
1035 				pr_err("invalid policy %s in thp_anon boot parameter\n", policy);
1036 				goto err;
1037 			}
1038 		}
1039 	}
1040 
1041 	huge_anon_orders_always = always;
1042 	huge_anon_orders_madvise = madvise;
1043 	huge_anon_orders_inherit = inherit;
1044 	anon_orders_configured = true;
1045 	return 1;
1046 
1047 err:
1048 	pr_warn("thp_anon=%s: error parsing string, ignoring setting\n", str);
1049 	return 0;
1050 }
1051 __setup("thp_anon=", setup_thp_anon);
1052 
1053 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
1054 {
1055 	if (likely(vma->vm_flags & VM_WRITE))
1056 		pmd = pmd_mkwrite(pmd, vma);
1057 	return pmd;
1058 }
1059 
1060 #ifdef CONFIG_MEMCG
1061 static inline
1062 struct deferred_split *get_deferred_split_queue(struct folio *folio)
1063 {
1064 	struct mem_cgroup *memcg = folio_memcg(folio);
1065 	struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));
1066 
1067 	if (memcg)
1068 		return &memcg->deferred_split_queue;
1069 	else
1070 		return &pgdat->deferred_split_queue;
1071 }
1072 #else
1073 static inline
1074 struct deferred_split *get_deferred_split_queue(struct folio *folio)
1075 {
1076 	struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));
1077 
1078 	return &pgdat->deferred_split_queue;
1079 }
1080 #endif
1081 
1082 static inline bool is_transparent_hugepage(const struct folio *folio)
1083 {
1084 	if (!folio_test_large(folio))
1085 		return false;
1086 
1087 	return is_huge_zero_folio(folio) ||
1088 		folio_test_large_rmappable(folio);
1089 }
1090 
1091 static unsigned long __thp_get_unmapped_area(struct file *filp,
1092 		unsigned long addr, unsigned long len,
1093 		loff_t off, unsigned long flags, unsigned long size,
1094 		vm_flags_t vm_flags)
1095 {
1096 	loff_t off_end = off + len;
1097 	loff_t off_align = round_up(off, size);
1098 	unsigned long len_pad, ret, off_sub;
1099 
1100 	if (!IS_ENABLED(CONFIG_64BIT) || in_compat_syscall())
1101 		return 0;
1102 
1103 	if (off_end <= off_align || (off_end - off_align) < size)
1104 		return 0;
1105 
1106 	len_pad = len + size;
1107 	if (len_pad < len || (off + len_pad) < off)
1108 		return 0;
1109 
1110 	ret = mm_get_unmapped_area_vmflags(current->mm, filp, addr, len_pad,
1111 					   off >> PAGE_SHIFT, flags, vm_flags);
1112 
1113 	/*
1114 	 * The failure might be due to length padding. The caller will retry
1115 	 * without the padding.
1116 	 */
1117 	if (IS_ERR_VALUE(ret))
1118 		return 0;
1119 
1120 	/*
1121 	 * Do not try to align to THP boundary if allocation at the address
1122 	 * hint succeeds.
1123 	 */
1124 	if (ret == addr)
1125 		return addr;
1126 
1127 	off_sub = (off - ret) & (size - 1);
1128 
1129 	if (test_bit(MMF_TOPDOWN, &current->mm->flags) && !off_sub)
1130 		return ret + size;
1131 
1132 	ret += off_sub;
1133 	return ret;
1134 }
1135 
1136 unsigned long thp_get_unmapped_area_vmflags(struct file *filp, unsigned long addr,
1137 		unsigned long len, unsigned long pgoff, unsigned long flags,
1138 		vm_flags_t vm_flags)
1139 {
1140 	unsigned long ret;
1141 	loff_t off = (loff_t)pgoff << PAGE_SHIFT;
1142 
1143 	ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE, vm_flags);
1144 	if (ret)
1145 		return ret;
1146 
1147 	return mm_get_unmapped_area_vmflags(current->mm, filp, addr, len, pgoff, flags,
1148 					    vm_flags);
1149 }
1150 
1151 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
1152 		unsigned long len, unsigned long pgoff, unsigned long flags)
1153 {
1154 	return thp_get_unmapped_area_vmflags(filp, addr, len, pgoff, flags, 0);
1155 }
1156 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
1157 
1158 static struct folio *vma_alloc_anon_folio_pmd(struct vm_area_struct *vma,
1159 		unsigned long addr)
1160 {
1161 	gfp_t gfp = vma_thp_gfp_mask(vma);
1162 	const int order = HPAGE_PMD_ORDER;
1163 	struct folio *folio;
1164 
1165 	folio = vma_alloc_folio(gfp, order, vma, addr & HPAGE_PMD_MASK);
1166 
1167 	if (unlikely(!folio)) {
1168 		count_vm_event(THP_FAULT_FALLBACK);
1169 		count_mthp_stat(order, MTHP_STAT_ANON_FAULT_FALLBACK);
1170 		return NULL;
1171 	}
1172 
1173 	VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
1174 	if (mem_cgroup_charge(folio, vma->vm_mm, gfp)) {
1175 		folio_put(folio);
1176 		count_vm_event(THP_FAULT_FALLBACK);
1177 		count_vm_event(THP_FAULT_FALLBACK_CHARGE);
1178 		count_mthp_stat(order, MTHP_STAT_ANON_FAULT_FALLBACK);
1179 		count_mthp_stat(order, MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE);
1180 		return NULL;
1181 	}
1182 	folio_throttle_swaprate(folio, gfp);
1183 
1184        /*
1185 	* When a folio is not zeroed during allocation (__GFP_ZERO not used)
1186 	* or user folios require special handling, folio_zero_user() is used to
1187 	* make sure that the page corresponding to the faulting address will be
1188 	* hot in the cache after zeroing.
1189 	*/
1190 	if (user_alloc_needs_zeroing())
1191 		folio_zero_user(folio, addr);
1192 	/*
1193 	 * The memory barrier inside __folio_mark_uptodate makes sure that
1194 	 * folio_zero_user writes become visible before the set_pmd_at()
1195 	 * write.
1196 	 */
1197 	__folio_mark_uptodate(folio);
1198 	return folio;
1199 }
1200 
1201 static void map_anon_folio_pmd(struct folio *folio, pmd_t *pmd,
1202 		struct vm_area_struct *vma, unsigned long haddr)
1203 {
1204 	pmd_t entry;
1205 
1206 	entry = folio_mk_pmd(folio, vma->vm_page_prot);
1207 	entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1208 	folio_add_new_anon_rmap(folio, vma, haddr, RMAP_EXCLUSIVE);
1209 	folio_add_lru_vma(folio, vma);
1210 	set_pmd_at(vma->vm_mm, haddr, pmd, entry);
1211 	update_mmu_cache_pmd(vma, haddr, pmd);
1212 	add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1213 	count_vm_event(THP_FAULT_ALLOC);
1214 	count_mthp_stat(HPAGE_PMD_ORDER, MTHP_STAT_ANON_FAULT_ALLOC);
1215 	count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
1216 }
1217 
1218 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf)
1219 {
1220 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1221 	struct vm_area_struct *vma = vmf->vma;
1222 	struct folio *folio;
1223 	pgtable_t pgtable;
1224 	vm_fault_t ret = 0;
1225 
1226 	folio = vma_alloc_anon_folio_pmd(vma, vmf->address);
1227 	if (unlikely(!folio))
1228 		return VM_FAULT_FALLBACK;
1229 
1230 	pgtable = pte_alloc_one(vma->vm_mm);
1231 	if (unlikely(!pgtable)) {
1232 		ret = VM_FAULT_OOM;
1233 		goto release;
1234 	}
1235 
1236 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1237 	if (unlikely(!pmd_none(*vmf->pmd))) {
1238 		goto unlock_release;
1239 	} else {
1240 		ret = check_stable_address_space(vma->vm_mm);
1241 		if (ret)
1242 			goto unlock_release;
1243 
1244 		/* Deliver the page fault to userland */
1245 		if (userfaultfd_missing(vma)) {
1246 			spin_unlock(vmf->ptl);
1247 			folio_put(folio);
1248 			pte_free(vma->vm_mm, pgtable);
1249 			ret = handle_userfault(vmf, VM_UFFD_MISSING);
1250 			VM_BUG_ON(ret & VM_FAULT_FALLBACK);
1251 			return ret;
1252 		}
1253 		pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
1254 		map_anon_folio_pmd(folio, vmf->pmd, vma, haddr);
1255 		mm_inc_nr_ptes(vma->vm_mm);
1256 		deferred_split_folio(folio, false);
1257 		spin_unlock(vmf->ptl);
1258 	}
1259 
1260 	return 0;
1261 unlock_release:
1262 	spin_unlock(vmf->ptl);
1263 release:
1264 	if (pgtable)
1265 		pte_free(vma->vm_mm, pgtable);
1266 	folio_put(folio);
1267 	return ret;
1268 
1269 }
1270 
1271 /*
1272  * always: directly stall for all thp allocations
1273  * defer: wake kswapd and fail if not immediately available
1274  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
1275  *		  fail if not immediately available
1276  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
1277  *	    available
1278  * never: never stall for any thp allocation
1279  */
1280 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
1281 {
1282 	const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
1283 
1284 	/* Always do synchronous compaction */
1285 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
1286 		return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
1287 
1288 	/* Kick kcompactd and fail quickly */
1289 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
1290 		return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
1291 
1292 	/* Synchronous compaction if madvised, otherwise kick kcompactd */
1293 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
1294 		return GFP_TRANSHUGE_LIGHT |
1295 			(vma_madvised ? __GFP_DIRECT_RECLAIM :
1296 					__GFP_KSWAPD_RECLAIM);
1297 
1298 	/* Only do synchronous compaction if madvised */
1299 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
1300 		return GFP_TRANSHUGE_LIGHT |
1301 		       (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
1302 
1303 	return GFP_TRANSHUGE_LIGHT;
1304 }
1305 
1306 /* Caller must hold page table lock. */
1307 static void set_huge_zero_folio(pgtable_t pgtable, struct mm_struct *mm,
1308 		struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
1309 		struct folio *zero_folio)
1310 {
1311 	pmd_t entry;
1312 	entry = folio_mk_pmd(zero_folio, vma->vm_page_prot);
1313 	pgtable_trans_huge_deposit(mm, pmd, pgtable);
1314 	set_pmd_at(mm, haddr, pmd, entry);
1315 	mm_inc_nr_ptes(mm);
1316 }
1317 
1318 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
1319 {
1320 	struct vm_area_struct *vma = vmf->vma;
1321 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1322 	vm_fault_t ret;
1323 
1324 	if (!thp_vma_suitable_order(vma, haddr, PMD_ORDER))
1325 		return VM_FAULT_FALLBACK;
1326 	ret = vmf_anon_prepare(vmf);
1327 	if (ret)
1328 		return ret;
1329 	khugepaged_enter_vma(vma, vma->vm_flags);
1330 
1331 	if (!(vmf->flags & FAULT_FLAG_WRITE) &&
1332 			!mm_forbids_zeropage(vma->vm_mm) &&
1333 			transparent_hugepage_use_zero_page()) {
1334 		pgtable_t pgtable;
1335 		struct folio *zero_folio;
1336 		vm_fault_t ret;
1337 
1338 		pgtable = pte_alloc_one(vma->vm_mm);
1339 		if (unlikely(!pgtable))
1340 			return VM_FAULT_OOM;
1341 		zero_folio = mm_get_huge_zero_folio(vma->vm_mm);
1342 		if (unlikely(!zero_folio)) {
1343 			pte_free(vma->vm_mm, pgtable);
1344 			count_vm_event(THP_FAULT_FALLBACK);
1345 			return VM_FAULT_FALLBACK;
1346 		}
1347 		vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1348 		ret = 0;
1349 		if (pmd_none(*vmf->pmd)) {
1350 			ret = check_stable_address_space(vma->vm_mm);
1351 			if (ret) {
1352 				spin_unlock(vmf->ptl);
1353 				pte_free(vma->vm_mm, pgtable);
1354 			} else if (userfaultfd_missing(vma)) {
1355 				spin_unlock(vmf->ptl);
1356 				pte_free(vma->vm_mm, pgtable);
1357 				ret = handle_userfault(vmf, VM_UFFD_MISSING);
1358 				VM_BUG_ON(ret & VM_FAULT_FALLBACK);
1359 			} else {
1360 				set_huge_zero_folio(pgtable, vma->vm_mm, vma,
1361 						   haddr, vmf->pmd, zero_folio);
1362 				update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1363 				spin_unlock(vmf->ptl);
1364 			}
1365 		} else {
1366 			spin_unlock(vmf->ptl);
1367 			pte_free(vma->vm_mm, pgtable);
1368 		}
1369 		return ret;
1370 	}
1371 
1372 	return __do_huge_pmd_anonymous_page(vmf);
1373 }
1374 
1375 static int insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
1376 		pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
1377 		pgtable_t pgtable)
1378 {
1379 	struct mm_struct *mm = vma->vm_mm;
1380 	pmd_t entry;
1381 
1382 	lockdep_assert_held(pmd_lockptr(mm, pmd));
1383 
1384 	if (!pmd_none(*pmd)) {
1385 		if (write) {
1386 			if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
1387 				WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
1388 				return -EEXIST;
1389 			}
1390 			entry = pmd_mkyoung(*pmd);
1391 			entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1392 			if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
1393 				update_mmu_cache_pmd(vma, addr, pmd);
1394 		}
1395 
1396 		return -EEXIST;
1397 	}
1398 
1399 	entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
1400 	if (pfn_t_devmap(pfn))
1401 		entry = pmd_mkdevmap(entry);
1402 	else
1403 		entry = pmd_mkspecial(entry);
1404 	if (write) {
1405 		entry = pmd_mkyoung(pmd_mkdirty(entry));
1406 		entry = maybe_pmd_mkwrite(entry, vma);
1407 	}
1408 
1409 	if (pgtable) {
1410 		pgtable_trans_huge_deposit(mm, pmd, pgtable);
1411 		mm_inc_nr_ptes(mm);
1412 	}
1413 
1414 	set_pmd_at(mm, addr, pmd, entry);
1415 	update_mmu_cache_pmd(vma, addr, pmd);
1416 	return 0;
1417 }
1418 
1419 /**
1420  * vmf_insert_pfn_pmd - insert a pmd size pfn
1421  * @vmf: Structure describing the fault
1422  * @pfn: pfn to insert
1423  * @write: whether it's a write fault
1424  *
1425  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info.
1426  *
1427  * Return: vm_fault_t value.
1428  */
1429 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)
1430 {
1431 	unsigned long addr = vmf->address & PMD_MASK;
1432 	struct vm_area_struct *vma = vmf->vma;
1433 	pgprot_t pgprot = vma->vm_page_prot;
1434 	pgtable_t pgtable = NULL;
1435 	spinlock_t *ptl;
1436 	int error;
1437 
1438 	/*
1439 	 * If we had pmd_special, we could avoid all these restrictions,
1440 	 * but we need to be consistent with PTEs and architectures that
1441 	 * can't support a 'special' bit.
1442 	 */
1443 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
1444 			!pfn_t_devmap(pfn));
1445 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1446 						(VM_PFNMAP|VM_MIXEDMAP));
1447 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1448 
1449 	if (addr < vma->vm_start || addr >= vma->vm_end)
1450 		return VM_FAULT_SIGBUS;
1451 
1452 	if (arch_needs_pgtable_deposit()) {
1453 		pgtable = pte_alloc_one(vma->vm_mm);
1454 		if (!pgtable)
1455 			return VM_FAULT_OOM;
1456 	}
1457 
1458 	track_pfn_insert(vma, &pgprot, pfn);
1459 	ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1460 	error = insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write,
1461 			pgtable);
1462 	spin_unlock(ptl);
1463 	if (error && pgtable)
1464 		pte_free(vma->vm_mm, pgtable);
1465 
1466 	return VM_FAULT_NOPAGE;
1467 }
1468 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
1469 
1470 vm_fault_t vmf_insert_folio_pmd(struct vm_fault *vmf, struct folio *folio,
1471 				bool write)
1472 {
1473 	struct vm_area_struct *vma = vmf->vma;
1474 	unsigned long addr = vmf->address & PMD_MASK;
1475 	struct mm_struct *mm = vma->vm_mm;
1476 	spinlock_t *ptl;
1477 	pgtable_t pgtable = NULL;
1478 	int error;
1479 
1480 	if (addr < vma->vm_start || addr >= vma->vm_end)
1481 		return VM_FAULT_SIGBUS;
1482 
1483 	if (WARN_ON_ONCE(folio_order(folio) != PMD_ORDER))
1484 		return VM_FAULT_SIGBUS;
1485 
1486 	if (arch_needs_pgtable_deposit()) {
1487 		pgtable = pte_alloc_one(vma->vm_mm);
1488 		if (!pgtable)
1489 			return VM_FAULT_OOM;
1490 	}
1491 
1492 	ptl = pmd_lock(mm, vmf->pmd);
1493 	if (pmd_none(*vmf->pmd)) {
1494 		folio_get(folio);
1495 		folio_add_file_rmap_pmd(folio, &folio->page, vma);
1496 		add_mm_counter(mm, mm_counter_file(folio), HPAGE_PMD_NR);
1497 	}
1498 	error = insert_pfn_pmd(vma, addr, vmf->pmd,
1499 			pfn_to_pfn_t(folio_pfn(folio)), vma->vm_page_prot,
1500 			write, pgtable);
1501 	spin_unlock(ptl);
1502 	if (error && pgtable)
1503 		pte_free(mm, pgtable);
1504 
1505 	return VM_FAULT_NOPAGE;
1506 }
1507 EXPORT_SYMBOL_GPL(vmf_insert_folio_pmd);
1508 
1509 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1510 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
1511 {
1512 	if (likely(vma->vm_flags & VM_WRITE))
1513 		pud = pud_mkwrite(pud);
1514 	return pud;
1515 }
1516 
1517 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
1518 		pud_t *pud, pfn_t pfn, bool write)
1519 {
1520 	struct mm_struct *mm = vma->vm_mm;
1521 	pgprot_t prot = vma->vm_page_prot;
1522 	pud_t entry;
1523 
1524 	if (!pud_none(*pud)) {
1525 		if (write) {
1526 			if (WARN_ON_ONCE(pud_pfn(*pud) != pfn_t_to_pfn(pfn)))
1527 				return;
1528 			entry = pud_mkyoung(*pud);
1529 			entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
1530 			if (pudp_set_access_flags(vma, addr, pud, entry, 1))
1531 				update_mmu_cache_pud(vma, addr, pud);
1532 		}
1533 		return;
1534 	}
1535 
1536 	entry = pud_mkhuge(pfn_t_pud(pfn, prot));
1537 	if (pfn_t_devmap(pfn))
1538 		entry = pud_mkdevmap(entry);
1539 	else
1540 		entry = pud_mkspecial(entry);
1541 	if (write) {
1542 		entry = pud_mkyoung(pud_mkdirty(entry));
1543 		entry = maybe_pud_mkwrite(entry, vma);
1544 	}
1545 	set_pud_at(mm, addr, pud, entry);
1546 	update_mmu_cache_pud(vma, addr, pud);
1547 }
1548 
1549 /**
1550  * vmf_insert_pfn_pud - insert a pud size pfn
1551  * @vmf: Structure describing the fault
1552  * @pfn: pfn to insert
1553  * @write: whether it's a write fault
1554  *
1555  * Insert a pud size pfn. See vmf_insert_pfn() for additional info.
1556  *
1557  * Return: vm_fault_t value.
1558  */
1559 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)
1560 {
1561 	unsigned long addr = vmf->address & PUD_MASK;
1562 	struct vm_area_struct *vma = vmf->vma;
1563 	pgprot_t pgprot = vma->vm_page_prot;
1564 	spinlock_t *ptl;
1565 
1566 	/*
1567 	 * If we had pud_special, we could avoid all these restrictions,
1568 	 * but we need to be consistent with PTEs and architectures that
1569 	 * can't support a 'special' bit.
1570 	 */
1571 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
1572 			!pfn_t_devmap(pfn));
1573 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1574 						(VM_PFNMAP|VM_MIXEDMAP));
1575 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1576 
1577 	if (addr < vma->vm_start || addr >= vma->vm_end)
1578 		return VM_FAULT_SIGBUS;
1579 
1580 	track_pfn_insert(vma, &pgprot, pfn);
1581 
1582 	ptl = pud_lock(vma->vm_mm, vmf->pud);
1583 	insert_pfn_pud(vma, addr, vmf->pud, pfn, write);
1584 	spin_unlock(ptl);
1585 
1586 	return VM_FAULT_NOPAGE;
1587 }
1588 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
1589 
1590 /**
1591  * vmf_insert_folio_pud - insert a pud size folio mapped by a pud entry
1592  * @vmf: Structure describing the fault
1593  * @folio: folio to insert
1594  * @write: whether it's a write fault
1595  *
1596  * Return: vm_fault_t value.
1597  */
1598 vm_fault_t vmf_insert_folio_pud(struct vm_fault *vmf, struct folio *folio,
1599 				bool write)
1600 {
1601 	struct vm_area_struct *vma = vmf->vma;
1602 	unsigned long addr = vmf->address & PUD_MASK;
1603 	pud_t *pud = vmf->pud;
1604 	struct mm_struct *mm = vma->vm_mm;
1605 	spinlock_t *ptl;
1606 
1607 	if (addr < vma->vm_start || addr >= vma->vm_end)
1608 		return VM_FAULT_SIGBUS;
1609 
1610 	if (WARN_ON_ONCE(folio_order(folio) != PUD_ORDER))
1611 		return VM_FAULT_SIGBUS;
1612 
1613 	ptl = pud_lock(mm, pud);
1614 
1615 	/*
1616 	 * If there is already an entry present we assume the folio is
1617 	 * already mapped, hence no need to take another reference. We
1618 	 * still call insert_pfn_pud() though in case the mapping needs
1619 	 * upgrading to writeable.
1620 	 */
1621 	if (pud_none(*vmf->pud)) {
1622 		folio_get(folio);
1623 		folio_add_file_rmap_pud(folio, &folio->page, vma);
1624 		add_mm_counter(mm, mm_counter_file(folio), HPAGE_PUD_NR);
1625 	}
1626 	insert_pfn_pud(vma, addr, vmf->pud, pfn_to_pfn_t(folio_pfn(folio)),
1627 		write);
1628 	spin_unlock(ptl);
1629 
1630 	return VM_FAULT_NOPAGE;
1631 }
1632 EXPORT_SYMBOL_GPL(vmf_insert_folio_pud);
1633 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1634 
1635 void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1636 	       pmd_t *pmd, bool write)
1637 {
1638 	pmd_t _pmd;
1639 
1640 	_pmd = pmd_mkyoung(*pmd);
1641 	if (write)
1642 		_pmd = pmd_mkdirty(_pmd);
1643 	if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1644 				  pmd, _pmd, write))
1645 		update_mmu_cache_pmd(vma, addr, pmd);
1646 }
1647 
1648 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1649 		pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
1650 {
1651 	unsigned long pfn = pmd_pfn(*pmd);
1652 	struct mm_struct *mm = vma->vm_mm;
1653 	struct page *page;
1654 	int ret;
1655 
1656 	assert_spin_locked(pmd_lockptr(mm, pmd));
1657 
1658 	if (flags & FOLL_WRITE && !pmd_write(*pmd))
1659 		return NULL;
1660 
1661 	if (pmd_present(*pmd) && pmd_devmap(*pmd))
1662 		/* pass */;
1663 	else
1664 		return NULL;
1665 
1666 	if (flags & FOLL_TOUCH)
1667 		touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1668 
1669 	/*
1670 	 * device mapped pages can only be returned if the
1671 	 * caller will manage the page reference count.
1672 	 */
1673 	if (!(flags & (FOLL_GET | FOLL_PIN)))
1674 		return ERR_PTR(-EEXIST);
1675 
1676 	pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1677 	*pgmap = get_dev_pagemap(pfn, *pgmap);
1678 	if (!*pgmap)
1679 		return ERR_PTR(-EFAULT);
1680 	page = pfn_to_page(pfn);
1681 	ret = try_grab_folio(page_folio(page), 1, flags);
1682 	if (ret)
1683 		page = ERR_PTR(ret);
1684 
1685 	return page;
1686 }
1687 
1688 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1689 		  pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1690 		  struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1691 {
1692 	spinlock_t *dst_ptl, *src_ptl;
1693 	struct page *src_page;
1694 	struct folio *src_folio;
1695 	pmd_t pmd;
1696 	pgtable_t pgtable = NULL;
1697 	int ret = -ENOMEM;
1698 
1699 	pmd = pmdp_get_lockless(src_pmd);
1700 	if (unlikely(pmd_present(pmd) && pmd_special(pmd))) {
1701 		dst_ptl = pmd_lock(dst_mm, dst_pmd);
1702 		src_ptl = pmd_lockptr(src_mm, src_pmd);
1703 		spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1704 		/*
1705 		 * No need to recheck the pmd, it can't change with write
1706 		 * mmap lock held here.
1707 		 *
1708 		 * Meanwhile, making sure it's not a CoW VMA with writable
1709 		 * mapping, otherwise it means either the anon page wrongly
1710 		 * applied special bit, or we made the PRIVATE mapping be
1711 		 * able to wrongly write to the backend MMIO.
1712 		 */
1713 		VM_WARN_ON_ONCE(is_cow_mapping(src_vma->vm_flags) && pmd_write(pmd));
1714 		goto set_pmd;
1715 	}
1716 
1717 	/* Skip if can be re-fill on fault */
1718 	if (!vma_is_anonymous(dst_vma))
1719 		return 0;
1720 
1721 	pgtable = pte_alloc_one(dst_mm);
1722 	if (unlikely(!pgtable))
1723 		goto out;
1724 
1725 	dst_ptl = pmd_lock(dst_mm, dst_pmd);
1726 	src_ptl = pmd_lockptr(src_mm, src_pmd);
1727 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1728 
1729 	ret = -EAGAIN;
1730 	pmd = *src_pmd;
1731 
1732 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1733 	if (unlikely(is_swap_pmd(pmd))) {
1734 		swp_entry_t entry = pmd_to_swp_entry(pmd);
1735 
1736 		VM_BUG_ON(!is_pmd_migration_entry(pmd));
1737 		if (!is_readable_migration_entry(entry)) {
1738 			entry = make_readable_migration_entry(
1739 							swp_offset(entry));
1740 			pmd = swp_entry_to_pmd(entry);
1741 			if (pmd_swp_soft_dirty(*src_pmd))
1742 				pmd = pmd_swp_mksoft_dirty(pmd);
1743 			if (pmd_swp_uffd_wp(*src_pmd))
1744 				pmd = pmd_swp_mkuffd_wp(pmd);
1745 			set_pmd_at(src_mm, addr, src_pmd, pmd);
1746 		}
1747 		add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1748 		mm_inc_nr_ptes(dst_mm);
1749 		pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1750 		if (!userfaultfd_wp(dst_vma))
1751 			pmd = pmd_swp_clear_uffd_wp(pmd);
1752 		set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1753 		ret = 0;
1754 		goto out_unlock;
1755 	}
1756 #endif
1757 
1758 	if (unlikely(!pmd_trans_huge(pmd))) {
1759 		pte_free(dst_mm, pgtable);
1760 		goto out_unlock;
1761 	}
1762 	/*
1763 	 * When page table lock is held, the huge zero pmd should not be
1764 	 * under splitting since we don't split the page itself, only pmd to
1765 	 * a page table.
1766 	 */
1767 	if (is_huge_zero_pmd(pmd)) {
1768 		/*
1769 		 * mm_get_huge_zero_folio() will never allocate a new
1770 		 * folio here, since we already have a zero page to
1771 		 * copy. It just takes a reference.
1772 		 */
1773 		mm_get_huge_zero_folio(dst_mm);
1774 		goto out_zero_page;
1775 	}
1776 
1777 	src_page = pmd_page(pmd);
1778 	VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1779 	src_folio = page_folio(src_page);
1780 
1781 	folio_get(src_folio);
1782 	if (unlikely(folio_try_dup_anon_rmap_pmd(src_folio, src_page, dst_vma, src_vma))) {
1783 		/* Page maybe pinned: split and retry the fault on PTEs. */
1784 		folio_put(src_folio);
1785 		pte_free(dst_mm, pgtable);
1786 		spin_unlock(src_ptl);
1787 		spin_unlock(dst_ptl);
1788 		__split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1789 		return -EAGAIN;
1790 	}
1791 	add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1792 out_zero_page:
1793 	mm_inc_nr_ptes(dst_mm);
1794 	pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1795 	pmdp_set_wrprotect(src_mm, addr, src_pmd);
1796 	if (!userfaultfd_wp(dst_vma))
1797 		pmd = pmd_clear_uffd_wp(pmd);
1798 	pmd = pmd_wrprotect(pmd);
1799 set_pmd:
1800 	pmd = pmd_mkold(pmd);
1801 	set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1802 
1803 	ret = 0;
1804 out_unlock:
1805 	spin_unlock(src_ptl);
1806 	spin_unlock(dst_ptl);
1807 out:
1808 	return ret;
1809 }
1810 
1811 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1812 void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1813 	       pud_t *pud, bool write)
1814 {
1815 	pud_t _pud;
1816 
1817 	_pud = pud_mkyoung(*pud);
1818 	if (write)
1819 		_pud = pud_mkdirty(_pud);
1820 	if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1821 				  pud, _pud, write))
1822 		update_mmu_cache_pud(vma, addr, pud);
1823 }
1824 
1825 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1826 		  pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1827 		  struct vm_area_struct *vma)
1828 {
1829 	spinlock_t *dst_ptl, *src_ptl;
1830 	pud_t pud;
1831 	int ret;
1832 
1833 	dst_ptl = pud_lock(dst_mm, dst_pud);
1834 	src_ptl = pud_lockptr(src_mm, src_pud);
1835 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1836 
1837 	ret = -EAGAIN;
1838 	pud = *src_pud;
1839 	if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1840 		goto out_unlock;
1841 
1842 	/*
1843 	 * TODO: once we support anonymous pages, use
1844 	 * folio_try_dup_anon_rmap_*() and split if duplicating fails.
1845 	 */
1846 	if (is_cow_mapping(vma->vm_flags) && pud_write(pud)) {
1847 		pudp_set_wrprotect(src_mm, addr, src_pud);
1848 		pud = pud_wrprotect(pud);
1849 	}
1850 	pud = pud_mkold(pud);
1851 	set_pud_at(dst_mm, addr, dst_pud, pud);
1852 
1853 	ret = 0;
1854 out_unlock:
1855 	spin_unlock(src_ptl);
1856 	spin_unlock(dst_ptl);
1857 	return ret;
1858 }
1859 
1860 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1861 {
1862 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1863 
1864 	vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1865 	if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1866 		goto unlock;
1867 
1868 	touch_pud(vmf->vma, vmf->address, vmf->pud, write);
1869 unlock:
1870 	spin_unlock(vmf->ptl);
1871 }
1872 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1873 
1874 void huge_pmd_set_accessed(struct vm_fault *vmf)
1875 {
1876 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1877 
1878 	vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1879 	if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd)))
1880 		goto unlock;
1881 
1882 	touch_pmd(vmf->vma, vmf->address, vmf->pmd, write);
1883 
1884 unlock:
1885 	spin_unlock(vmf->ptl);
1886 }
1887 
1888 static vm_fault_t do_huge_zero_wp_pmd(struct vm_fault *vmf)
1889 {
1890 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1891 	struct vm_area_struct *vma = vmf->vma;
1892 	struct mmu_notifier_range range;
1893 	struct folio *folio;
1894 	vm_fault_t ret = 0;
1895 
1896 	folio = vma_alloc_anon_folio_pmd(vma, vmf->address);
1897 	if (unlikely(!folio))
1898 		return VM_FAULT_FALLBACK;
1899 
1900 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, haddr,
1901 				haddr + HPAGE_PMD_SIZE);
1902 	mmu_notifier_invalidate_range_start(&range);
1903 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1904 	if (unlikely(!pmd_same(pmdp_get(vmf->pmd), vmf->orig_pmd)))
1905 		goto release;
1906 	ret = check_stable_address_space(vma->vm_mm);
1907 	if (ret)
1908 		goto release;
1909 	(void)pmdp_huge_clear_flush(vma, haddr, vmf->pmd);
1910 	map_anon_folio_pmd(folio, vmf->pmd, vma, haddr);
1911 	goto unlock;
1912 release:
1913 	folio_put(folio);
1914 unlock:
1915 	spin_unlock(vmf->ptl);
1916 	mmu_notifier_invalidate_range_end(&range);
1917 	return ret;
1918 }
1919 
1920 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1921 {
1922 	const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
1923 	struct vm_area_struct *vma = vmf->vma;
1924 	struct folio *folio;
1925 	struct page *page;
1926 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1927 	pmd_t orig_pmd = vmf->orig_pmd;
1928 
1929 	vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1930 	VM_BUG_ON_VMA(!vma->anon_vma, vma);
1931 
1932 	if (is_huge_zero_pmd(orig_pmd)) {
1933 		vm_fault_t ret = do_huge_zero_wp_pmd(vmf);
1934 
1935 		if (!(ret & VM_FAULT_FALLBACK))
1936 			return ret;
1937 
1938 		/* Fallback to splitting PMD if THP cannot be allocated */
1939 		goto fallback;
1940 	}
1941 
1942 	spin_lock(vmf->ptl);
1943 
1944 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1945 		spin_unlock(vmf->ptl);
1946 		return 0;
1947 	}
1948 
1949 	page = pmd_page(orig_pmd);
1950 	folio = page_folio(page);
1951 	VM_BUG_ON_PAGE(!PageHead(page), page);
1952 
1953 	/* Early check when only holding the PT lock. */
1954 	if (PageAnonExclusive(page))
1955 		goto reuse;
1956 
1957 	if (!folio_trylock(folio)) {
1958 		folio_get(folio);
1959 		spin_unlock(vmf->ptl);
1960 		folio_lock(folio);
1961 		spin_lock(vmf->ptl);
1962 		if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1963 			spin_unlock(vmf->ptl);
1964 			folio_unlock(folio);
1965 			folio_put(folio);
1966 			return 0;
1967 		}
1968 		folio_put(folio);
1969 	}
1970 
1971 	/* Recheck after temporarily dropping the PT lock. */
1972 	if (PageAnonExclusive(page)) {
1973 		folio_unlock(folio);
1974 		goto reuse;
1975 	}
1976 
1977 	/*
1978 	 * See do_wp_page(): we can only reuse the folio exclusively if
1979 	 * there are no additional references. Note that we always drain
1980 	 * the LRU cache immediately after adding a THP.
1981 	 */
1982 	if (folio_ref_count(folio) >
1983 			1 + folio_test_swapcache(folio) * folio_nr_pages(folio))
1984 		goto unlock_fallback;
1985 	if (folio_test_swapcache(folio))
1986 		folio_free_swap(folio);
1987 	if (folio_ref_count(folio) == 1) {
1988 		pmd_t entry;
1989 
1990 		folio_move_anon_rmap(folio, vma);
1991 		SetPageAnonExclusive(page);
1992 		folio_unlock(folio);
1993 reuse:
1994 		if (unlikely(unshare)) {
1995 			spin_unlock(vmf->ptl);
1996 			return 0;
1997 		}
1998 		entry = pmd_mkyoung(orig_pmd);
1999 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2000 		if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
2001 			update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
2002 		spin_unlock(vmf->ptl);
2003 		return 0;
2004 	}
2005 
2006 unlock_fallback:
2007 	folio_unlock(folio);
2008 	spin_unlock(vmf->ptl);
2009 fallback:
2010 	__split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
2011 	return VM_FAULT_FALLBACK;
2012 }
2013 
2014 static inline bool can_change_pmd_writable(struct vm_area_struct *vma,
2015 					   unsigned long addr, pmd_t pmd)
2016 {
2017 	struct page *page;
2018 
2019 	if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE)))
2020 		return false;
2021 
2022 	/* Don't touch entries that are not even readable (NUMA hinting). */
2023 	if (pmd_protnone(pmd))
2024 		return false;
2025 
2026 	/* Do we need write faults for softdirty tracking? */
2027 	if (pmd_needs_soft_dirty_wp(vma, pmd))
2028 		return false;
2029 
2030 	/* Do we need write faults for uffd-wp tracking? */
2031 	if (userfaultfd_huge_pmd_wp(vma, pmd))
2032 		return false;
2033 
2034 	if (!(vma->vm_flags & VM_SHARED)) {
2035 		/* See can_change_pte_writable(). */
2036 		page = vm_normal_page_pmd(vma, addr, pmd);
2037 		return page && PageAnon(page) && PageAnonExclusive(page);
2038 	}
2039 
2040 	/* See can_change_pte_writable(). */
2041 	return pmd_dirty(pmd);
2042 }
2043 
2044 /* NUMA hinting page fault entry point for trans huge pmds */
2045 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
2046 {
2047 	struct vm_area_struct *vma = vmf->vma;
2048 	struct folio *folio;
2049 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
2050 	int nid = NUMA_NO_NODE;
2051 	int target_nid, last_cpupid;
2052 	pmd_t pmd, old_pmd;
2053 	bool writable = false;
2054 	int flags = 0;
2055 
2056 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
2057 	old_pmd = pmdp_get(vmf->pmd);
2058 
2059 	if (unlikely(!pmd_same(old_pmd, vmf->orig_pmd))) {
2060 		spin_unlock(vmf->ptl);
2061 		return 0;
2062 	}
2063 
2064 	pmd = pmd_modify(old_pmd, vma->vm_page_prot);
2065 
2066 	/*
2067 	 * Detect now whether the PMD could be writable; this information
2068 	 * is only valid while holding the PT lock.
2069 	 */
2070 	writable = pmd_write(pmd);
2071 	if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
2072 	    can_change_pmd_writable(vma, vmf->address, pmd))
2073 		writable = true;
2074 
2075 	folio = vm_normal_folio_pmd(vma, haddr, pmd);
2076 	if (!folio)
2077 		goto out_map;
2078 
2079 	nid = folio_nid(folio);
2080 
2081 	target_nid = numa_migrate_check(folio, vmf, haddr, &flags, writable,
2082 					&last_cpupid);
2083 	if (target_nid == NUMA_NO_NODE)
2084 		goto out_map;
2085 	if (migrate_misplaced_folio_prepare(folio, vma, target_nid)) {
2086 		flags |= TNF_MIGRATE_FAIL;
2087 		goto out_map;
2088 	}
2089 	/* The folio is isolated and isolation code holds a folio reference. */
2090 	spin_unlock(vmf->ptl);
2091 	writable = false;
2092 
2093 	if (!migrate_misplaced_folio(folio, target_nid)) {
2094 		flags |= TNF_MIGRATED;
2095 		nid = target_nid;
2096 		task_numa_fault(last_cpupid, nid, HPAGE_PMD_NR, flags);
2097 		return 0;
2098 	}
2099 
2100 	flags |= TNF_MIGRATE_FAIL;
2101 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
2102 	if (unlikely(!pmd_same(pmdp_get(vmf->pmd), vmf->orig_pmd))) {
2103 		spin_unlock(vmf->ptl);
2104 		return 0;
2105 	}
2106 out_map:
2107 	/* Restore the PMD */
2108 	pmd = pmd_modify(pmdp_get(vmf->pmd), vma->vm_page_prot);
2109 	pmd = pmd_mkyoung(pmd);
2110 	if (writable)
2111 		pmd = pmd_mkwrite(pmd, vma);
2112 	set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
2113 	update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
2114 	spin_unlock(vmf->ptl);
2115 
2116 	if (nid != NUMA_NO_NODE)
2117 		task_numa_fault(last_cpupid, nid, HPAGE_PMD_NR, flags);
2118 	return 0;
2119 }
2120 
2121 /*
2122  * Return true if we do MADV_FREE successfully on entire pmd page.
2123  * Otherwise, return false.
2124  */
2125 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
2126 		pmd_t *pmd, unsigned long addr, unsigned long next)
2127 {
2128 	spinlock_t *ptl;
2129 	pmd_t orig_pmd;
2130 	struct folio *folio;
2131 	struct mm_struct *mm = tlb->mm;
2132 	bool ret = false;
2133 
2134 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
2135 
2136 	ptl = pmd_trans_huge_lock(pmd, vma);
2137 	if (!ptl)
2138 		goto out_unlocked;
2139 
2140 	orig_pmd = *pmd;
2141 	if (is_huge_zero_pmd(orig_pmd))
2142 		goto out;
2143 
2144 	if (unlikely(!pmd_present(orig_pmd))) {
2145 		VM_BUG_ON(thp_migration_supported() &&
2146 				  !is_pmd_migration_entry(orig_pmd));
2147 		goto out;
2148 	}
2149 
2150 	folio = pmd_folio(orig_pmd);
2151 	/*
2152 	 * If other processes are mapping this folio, we couldn't discard
2153 	 * the folio unless they all do MADV_FREE so let's skip the folio.
2154 	 */
2155 	if (folio_maybe_mapped_shared(folio))
2156 		goto out;
2157 
2158 	if (!folio_trylock(folio))
2159 		goto out;
2160 
2161 	/*
2162 	 * If user want to discard part-pages of THP, split it so MADV_FREE
2163 	 * will deactivate only them.
2164 	 */
2165 	if (next - addr != HPAGE_PMD_SIZE) {
2166 		folio_get(folio);
2167 		spin_unlock(ptl);
2168 		split_folio(folio);
2169 		folio_unlock(folio);
2170 		folio_put(folio);
2171 		goto out_unlocked;
2172 	}
2173 
2174 	if (folio_test_dirty(folio))
2175 		folio_clear_dirty(folio);
2176 	folio_unlock(folio);
2177 
2178 	if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
2179 		pmdp_invalidate(vma, addr, pmd);
2180 		orig_pmd = pmd_mkold(orig_pmd);
2181 		orig_pmd = pmd_mkclean(orig_pmd);
2182 
2183 		set_pmd_at(mm, addr, pmd, orig_pmd);
2184 		tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
2185 	}
2186 
2187 	folio_mark_lazyfree(folio);
2188 	ret = true;
2189 out:
2190 	spin_unlock(ptl);
2191 out_unlocked:
2192 	return ret;
2193 }
2194 
2195 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
2196 {
2197 	pgtable_t pgtable;
2198 
2199 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2200 	pte_free(mm, pgtable);
2201 	mm_dec_nr_ptes(mm);
2202 }
2203 
2204 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
2205 		 pmd_t *pmd, unsigned long addr)
2206 {
2207 	pmd_t orig_pmd;
2208 	spinlock_t *ptl;
2209 
2210 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
2211 
2212 	ptl = __pmd_trans_huge_lock(pmd, vma);
2213 	if (!ptl)
2214 		return 0;
2215 	/*
2216 	 * For architectures like ppc64 we look at deposited pgtable
2217 	 * when calling pmdp_huge_get_and_clear. So do the
2218 	 * pgtable_trans_huge_withdraw after finishing pmdp related
2219 	 * operations.
2220 	 */
2221 	orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
2222 						tlb->fullmm);
2223 	arch_check_zapped_pmd(vma, orig_pmd);
2224 	tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
2225 	if (!vma_is_dax(vma) && vma_is_special_huge(vma)) {
2226 		if (arch_needs_pgtable_deposit())
2227 			zap_deposited_table(tlb->mm, pmd);
2228 		spin_unlock(ptl);
2229 	} else if (is_huge_zero_pmd(orig_pmd)) {
2230 		if (!vma_is_dax(vma) || arch_needs_pgtable_deposit())
2231 			zap_deposited_table(tlb->mm, pmd);
2232 		spin_unlock(ptl);
2233 	} else {
2234 		struct folio *folio = NULL;
2235 		int flush_needed = 1;
2236 
2237 		if (pmd_present(orig_pmd)) {
2238 			struct page *page = pmd_page(orig_pmd);
2239 
2240 			folio = page_folio(page);
2241 			folio_remove_rmap_pmd(folio, page, vma);
2242 			WARN_ON_ONCE(folio_mapcount(folio) < 0);
2243 			VM_BUG_ON_PAGE(!PageHead(page), page);
2244 		} else if (thp_migration_supported()) {
2245 			swp_entry_t entry;
2246 
2247 			VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
2248 			entry = pmd_to_swp_entry(orig_pmd);
2249 			folio = pfn_swap_entry_folio(entry);
2250 			flush_needed = 0;
2251 		} else
2252 			WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
2253 
2254 		if (folio_test_anon(folio)) {
2255 			zap_deposited_table(tlb->mm, pmd);
2256 			add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
2257 		} else {
2258 			if (arch_needs_pgtable_deposit())
2259 				zap_deposited_table(tlb->mm, pmd);
2260 			add_mm_counter(tlb->mm, mm_counter_file(folio),
2261 				       -HPAGE_PMD_NR);
2262 
2263 			/*
2264 			 * Use flush_needed to indicate whether the PMD entry
2265 			 * is present, instead of checking pmd_present() again.
2266 			 */
2267 			if (flush_needed && pmd_young(orig_pmd) &&
2268 			    likely(vma_has_recency(vma)))
2269 				folio_mark_accessed(folio);
2270 		}
2271 
2272 		spin_unlock(ptl);
2273 		if (flush_needed)
2274 			tlb_remove_page_size(tlb, &folio->page, HPAGE_PMD_SIZE);
2275 	}
2276 	return 1;
2277 }
2278 
2279 #ifndef pmd_move_must_withdraw
2280 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
2281 					 spinlock_t *old_pmd_ptl,
2282 					 struct vm_area_struct *vma)
2283 {
2284 	/*
2285 	 * With split pmd lock we also need to move preallocated
2286 	 * PTE page table if new_pmd is on different PMD page table.
2287 	 *
2288 	 * We also don't deposit and withdraw tables for file pages.
2289 	 */
2290 	return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
2291 }
2292 #endif
2293 
2294 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
2295 {
2296 #ifdef CONFIG_MEM_SOFT_DIRTY
2297 	if (unlikely(is_pmd_migration_entry(pmd)))
2298 		pmd = pmd_swp_mksoft_dirty(pmd);
2299 	else if (pmd_present(pmd))
2300 		pmd = pmd_mksoft_dirty(pmd);
2301 #endif
2302 	return pmd;
2303 }
2304 
2305 static pmd_t clear_uffd_wp_pmd(pmd_t pmd)
2306 {
2307 	if (pmd_present(pmd))
2308 		pmd = pmd_clear_uffd_wp(pmd);
2309 	else if (is_swap_pmd(pmd))
2310 		pmd = pmd_swp_clear_uffd_wp(pmd);
2311 
2312 	return pmd;
2313 }
2314 
2315 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
2316 		  unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
2317 {
2318 	spinlock_t *old_ptl, *new_ptl;
2319 	pmd_t pmd;
2320 	struct mm_struct *mm = vma->vm_mm;
2321 	bool force_flush = false;
2322 
2323 	/*
2324 	 * The destination pmd shouldn't be established, free_pgtables()
2325 	 * should have released it; but move_page_tables() might have already
2326 	 * inserted a page table, if racing against shmem/file collapse.
2327 	 */
2328 	if (!pmd_none(*new_pmd)) {
2329 		VM_BUG_ON(pmd_trans_huge(*new_pmd));
2330 		return false;
2331 	}
2332 
2333 	/*
2334 	 * We don't have to worry about the ordering of src and dst
2335 	 * ptlocks because exclusive mmap_lock prevents deadlock.
2336 	 */
2337 	old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
2338 	if (old_ptl) {
2339 		new_ptl = pmd_lockptr(mm, new_pmd);
2340 		if (new_ptl != old_ptl)
2341 			spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
2342 		pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
2343 		if (pmd_present(pmd))
2344 			force_flush = true;
2345 		VM_BUG_ON(!pmd_none(*new_pmd));
2346 
2347 		if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
2348 			pgtable_t pgtable;
2349 			pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
2350 			pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
2351 		}
2352 		pmd = move_soft_dirty_pmd(pmd);
2353 		if (vma_has_uffd_without_event_remap(vma))
2354 			pmd = clear_uffd_wp_pmd(pmd);
2355 		set_pmd_at(mm, new_addr, new_pmd, pmd);
2356 		if (force_flush)
2357 			flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
2358 		if (new_ptl != old_ptl)
2359 			spin_unlock(new_ptl);
2360 		spin_unlock(old_ptl);
2361 		return true;
2362 	}
2363 	return false;
2364 }
2365 
2366 /*
2367  * Returns
2368  *  - 0 if PMD could not be locked
2369  *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
2370  *      or if prot_numa but THP migration is not supported
2371  *  - HPAGE_PMD_NR if protections changed and TLB flush necessary
2372  */
2373 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
2374 		    pmd_t *pmd, unsigned long addr, pgprot_t newprot,
2375 		    unsigned long cp_flags)
2376 {
2377 	struct mm_struct *mm = vma->vm_mm;
2378 	spinlock_t *ptl;
2379 	pmd_t oldpmd, entry;
2380 	bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
2381 	bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
2382 	bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
2383 	int ret = 1;
2384 
2385 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
2386 
2387 	if (prot_numa && !thp_migration_supported())
2388 		return 1;
2389 
2390 	ptl = __pmd_trans_huge_lock(pmd, vma);
2391 	if (!ptl)
2392 		return 0;
2393 
2394 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2395 	if (is_swap_pmd(*pmd)) {
2396 		swp_entry_t entry = pmd_to_swp_entry(*pmd);
2397 		struct folio *folio = pfn_swap_entry_folio(entry);
2398 		pmd_t newpmd;
2399 
2400 		VM_BUG_ON(!is_pmd_migration_entry(*pmd));
2401 		if (is_writable_migration_entry(entry)) {
2402 			/*
2403 			 * A protection check is difficult so
2404 			 * just be safe and disable write
2405 			 */
2406 			if (folio_test_anon(folio))
2407 				entry = make_readable_exclusive_migration_entry(swp_offset(entry));
2408 			else
2409 				entry = make_readable_migration_entry(swp_offset(entry));
2410 			newpmd = swp_entry_to_pmd(entry);
2411 			if (pmd_swp_soft_dirty(*pmd))
2412 				newpmd = pmd_swp_mksoft_dirty(newpmd);
2413 		} else {
2414 			newpmd = *pmd;
2415 		}
2416 
2417 		if (uffd_wp)
2418 			newpmd = pmd_swp_mkuffd_wp(newpmd);
2419 		else if (uffd_wp_resolve)
2420 			newpmd = pmd_swp_clear_uffd_wp(newpmd);
2421 		if (!pmd_same(*pmd, newpmd))
2422 			set_pmd_at(mm, addr, pmd, newpmd);
2423 		goto unlock;
2424 	}
2425 #endif
2426 
2427 	if (prot_numa) {
2428 		struct folio *folio;
2429 		bool toptier;
2430 		/*
2431 		 * Avoid trapping faults against the zero page. The read-only
2432 		 * data is likely to be read-cached on the local CPU and
2433 		 * local/remote hits to the zero page are not interesting.
2434 		 */
2435 		if (is_huge_zero_pmd(*pmd))
2436 			goto unlock;
2437 
2438 		if (pmd_protnone(*pmd))
2439 			goto unlock;
2440 
2441 		folio = pmd_folio(*pmd);
2442 		toptier = node_is_toptier(folio_nid(folio));
2443 		/*
2444 		 * Skip scanning top tier node if normal numa
2445 		 * balancing is disabled
2446 		 */
2447 		if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
2448 		    toptier)
2449 			goto unlock;
2450 
2451 		if (folio_use_access_time(folio))
2452 			folio_xchg_access_time(folio,
2453 					       jiffies_to_msecs(jiffies));
2454 	}
2455 	/*
2456 	 * In case prot_numa, we are under mmap_read_lock(mm). It's critical
2457 	 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
2458 	 * which is also under mmap_read_lock(mm):
2459 	 *
2460 	 *	CPU0:				CPU1:
2461 	 *				change_huge_pmd(prot_numa=1)
2462 	 *				 pmdp_huge_get_and_clear_notify()
2463 	 * madvise_dontneed()
2464 	 *  zap_pmd_range()
2465 	 *   pmd_trans_huge(*pmd) == 0 (without ptl)
2466 	 *   // skip the pmd
2467 	 *				 set_pmd_at();
2468 	 *				 // pmd is re-established
2469 	 *
2470 	 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
2471 	 * which may break userspace.
2472 	 *
2473 	 * pmdp_invalidate_ad() is required to make sure we don't miss
2474 	 * dirty/young flags set by hardware.
2475 	 */
2476 	oldpmd = pmdp_invalidate_ad(vma, addr, pmd);
2477 
2478 	entry = pmd_modify(oldpmd, newprot);
2479 	if (uffd_wp)
2480 		entry = pmd_mkuffd_wp(entry);
2481 	else if (uffd_wp_resolve)
2482 		/*
2483 		 * Leave the write bit to be handled by PF interrupt
2484 		 * handler, then things like COW could be properly
2485 		 * handled.
2486 		 */
2487 		entry = pmd_clear_uffd_wp(entry);
2488 
2489 	/* See change_pte_range(). */
2490 	if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) && !pmd_write(entry) &&
2491 	    can_change_pmd_writable(vma, addr, entry))
2492 		entry = pmd_mkwrite(entry, vma);
2493 
2494 	ret = HPAGE_PMD_NR;
2495 	set_pmd_at(mm, addr, pmd, entry);
2496 
2497 	if (huge_pmd_needs_flush(oldpmd, entry))
2498 		tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE);
2499 unlock:
2500 	spin_unlock(ptl);
2501 	return ret;
2502 }
2503 
2504 /*
2505  * Returns:
2506  *
2507  * - 0: if pud leaf changed from under us
2508  * - 1: if pud can be skipped
2509  * - HPAGE_PUD_NR: if pud was successfully processed
2510  */
2511 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2512 int change_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2513 		    pud_t *pudp, unsigned long addr, pgprot_t newprot,
2514 		    unsigned long cp_flags)
2515 {
2516 	struct mm_struct *mm = vma->vm_mm;
2517 	pud_t oldpud, entry;
2518 	spinlock_t *ptl;
2519 
2520 	tlb_change_page_size(tlb, HPAGE_PUD_SIZE);
2521 
2522 	/* NUMA balancing doesn't apply to dax */
2523 	if (cp_flags & MM_CP_PROT_NUMA)
2524 		return 1;
2525 
2526 	/*
2527 	 * Huge entries on userfault-wp only works with anonymous, while we
2528 	 * don't have anonymous PUDs yet.
2529 	 */
2530 	if (WARN_ON_ONCE(cp_flags & MM_CP_UFFD_WP_ALL))
2531 		return 1;
2532 
2533 	ptl = __pud_trans_huge_lock(pudp, vma);
2534 	if (!ptl)
2535 		return 0;
2536 
2537 	/*
2538 	 * Can't clear PUD or it can race with concurrent zapping.  See
2539 	 * change_huge_pmd().
2540 	 */
2541 	oldpud = pudp_invalidate(vma, addr, pudp);
2542 	entry = pud_modify(oldpud, newprot);
2543 	set_pud_at(mm, addr, pudp, entry);
2544 	tlb_flush_pud_range(tlb, addr, HPAGE_PUD_SIZE);
2545 
2546 	spin_unlock(ptl);
2547 	return HPAGE_PUD_NR;
2548 }
2549 #endif
2550 
2551 #ifdef CONFIG_USERFAULTFD
2552 /*
2553  * The PT lock for src_pmd and dst_vma/src_vma (for reading) are locked by
2554  * the caller, but it must return after releasing the page_table_lock.
2555  * Just move the page from src_pmd to dst_pmd if possible.
2556  * Return zero if succeeded in moving the page, -EAGAIN if it needs to be
2557  * repeated by the caller, or other errors in case of failure.
2558  */
2559 int move_pages_huge_pmd(struct mm_struct *mm, pmd_t *dst_pmd, pmd_t *src_pmd, pmd_t dst_pmdval,
2560 			struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
2561 			unsigned long dst_addr, unsigned long src_addr)
2562 {
2563 	pmd_t _dst_pmd, src_pmdval;
2564 	struct page *src_page;
2565 	struct folio *src_folio;
2566 	struct anon_vma *src_anon_vma;
2567 	spinlock_t *src_ptl, *dst_ptl;
2568 	pgtable_t src_pgtable;
2569 	struct mmu_notifier_range range;
2570 	int err = 0;
2571 
2572 	src_pmdval = *src_pmd;
2573 	src_ptl = pmd_lockptr(mm, src_pmd);
2574 
2575 	lockdep_assert_held(src_ptl);
2576 	vma_assert_locked(src_vma);
2577 	vma_assert_locked(dst_vma);
2578 
2579 	/* Sanity checks before the operation */
2580 	if (WARN_ON_ONCE(!pmd_none(dst_pmdval)) || WARN_ON_ONCE(src_addr & ~HPAGE_PMD_MASK) ||
2581 	    WARN_ON_ONCE(dst_addr & ~HPAGE_PMD_MASK)) {
2582 		spin_unlock(src_ptl);
2583 		return -EINVAL;
2584 	}
2585 
2586 	if (!pmd_trans_huge(src_pmdval)) {
2587 		spin_unlock(src_ptl);
2588 		if (is_pmd_migration_entry(src_pmdval)) {
2589 			pmd_migration_entry_wait(mm, &src_pmdval);
2590 			return -EAGAIN;
2591 		}
2592 		return -ENOENT;
2593 	}
2594 
2595 	src_page = pmd_page(src_pmdval);
2596 
2597 	if (!is_huge_zero_pmd(src_pmdval)) {
2598 		if (unlikely(!PageAnonExclusive(src_page))) {
2599 			spin_unlock(src_ptl);
2600 			return -EBUSY;
2601 		}
2602 
2603 		src_folio = page_folio(src_page);
2604 		folio_get(src_folio);
2605 	} else
2606 		src_folio = NULL;
2607 
2608 	spin_unlock(src_ptl);
2609 
2610 	flush_cache_range(src_vma, src_addr, src_addr + HPAGE_PMD_SIZE);
2611 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, src_addr,
2612 				src_addr + HPAGE_PMD_SIZE);
2613 	mmu_notifier_invalidate_range_start(&range);
2614 
2615 	if (src_folio) {
2616 		folio_lock(src_folio);
2617 
2618 		/*
2619 		 * split_huge_page walks the anon_vma chain without the page
2620 		 * lock. Serialize against it with the anon_vma lock, the page
2621 		 * lock is not enough.
2622 		 */
2623 		src_anon_vma = folio_get_anon_vma(src_folio);
2624 		if (!src_anon_vma) {
2625 			err = -EAGAIN;
2626 			goto unlock_folio;
2627 		}
2628 		anon_vma_lock_write(src_anon_vma);
2629 	} else
2630 		src_anon_vma = NULL;
2631 
2632 	dst_ptl = pmd_lockptr(mm, dst_pmd);
2633 	double_pt_lock(src_ptl, dst_ptl);
2634 	if (unlikely(!pmd_same(*src_pmd, src_pmdval) ||
2635 		     !pmd_same(*dst_pmd, dst_pmdval))) {
2636 		err = -EAGAIN;
2637 		goto unlock_ptls;
2638 	}
2639 	if (src_folio) {
2640 		if (folio_maybe_dma_pinned(src_folio) ||
2641 		    !PageAnonExclusive(&src_folio->page)) {
2642 			err = -EBUSY;
2643 			goto unlock_ptls;
2644 		}
2645 
2646 		if (WARN_ON_ONCE(!folio_test_head(src_folio)) ||
2647 		    WARN_ON_ONCE(!folio_test_anon(src_folio))) {
2648 			err = -EBUSY;
2649 			goto unlock_ptls;
2650 		}
2651 
2652 		src_pmdval = pmdp_huge_clear_flush(src_vma, src_addr, src_pmd);
2653 		/* Folio got pinned from under us. Put it back and fail the move. */
2654 		if (folio_maybe_dma_pinned(src_folio)) {
2655 			set_pmd_at(mm, src_addr, src_pmd, src_pmdval);
2656 			err = -EBUSY;
2657 			goto unlock_ptls;
2658 		}
2659 
2660 		folio_move_anon_rmap(src_folio, dst_vma);
2661 		src_folio->index = linear_page_index(dst_vma, dst_addr);
2662 
2663 		_dst_pmd = folio_mk_pmd(src_folio, dst_vma->vm_page_prot);
2664 		/* Follow mremap() behavior and treat the entry dirty after the move */
2665 		_dst_pmd = pmd_mkwrite(pmd_mkdirty(_dst_pmd), dst_vma);
2666 	} else {
2667 		src_pmdval = pmdp_huge_clear_flush(src_vma, src_addr, src_pmd);
2668 		_dst_pmd = folio_mk_pmd(src_folio, dst_vma->vm_page_prot);
2669 	}
2670 	set_pmd_at(mm, dst_addr, dst_pmd, _dst_pmd);
2671 
2672 	src_pgtable = pgtable_trans_huge_withdraw(mm, src_pmd);
2673 	pgtable_trans_huge_deposit(mm, dst_pmd, src_pgtable);
2674 unlock_ptls:
2675 	double_pt_unlock(src_ptl, dst_ptl);
2676 	if (src_anon_vma) {
2677 		anon_vma_unlock_write(src_anon_vma);
2678 		put_anon_vma(src_anon_vma);
2679 	}
2680 unlock_folio:
2681 	/* unblock rmap walks */
2682 	if (src_folio)
2683 		folio_unlock(src_folio);
2684 	mmu_notifier_invalidate_range_end(&range);
2685 	if (src_folio)
2686 		folio_put(src_folio);
2687 	return err;
2688 }
2689 #endif /* CONFIG_USERFAULTFD */
2690 
2691 /*
2692  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
2693  *
2694  * Note that if it returns page table lock pointer, this routine returns without
2695  * unlocking page table lock. So callers must unlock it.
2696  */
2697 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
2698 {
2699 	spinlock_t *ptl;
2700 	ptl = pmd_lock(vma->vm_mm, pmd);
2701 	if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
2702 			pmd_devmap(*pmd)))
2703 		return ptl;
2704 	spin_unlock(ptl);
2705 	return NULL;
2706 }
2707 
2708 /*
2709  * Returns page table lock pointer if a given pud maps a thp, NULL otherwise.
2710  *
2711  * Note that if it returns page table lock pointer, this routine returns without
2712  * unlocking page table lock. So callers must unlock it.
2713  */
2714 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2715 {
2716 	spinlock_t *ptl;
2717 
2718 	ptl = pud_lock(vma->vm_mm, pud);
2719 	if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2720 		return ptl;
2721 	spin_unlock(ptl);
2722 	return NULL;
2723 }
2724 
2725 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2726 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2727 		 pud_t *pud, unsigned long addr)
2728 {
2729 	spinlock_t *ptl;
2730 	pud_t orig_pud;
2731 
2732 	ptl = __pud_trans_huge_lock(pud, vma);
2733 	if (!ptl)
2734 		return 0;
2735 
2736 	orig_pud = pudp_huge_get_and_clear_full(vma, addr, pud, tlb->fullmm);
2737 	arch_check_zapped_pud(vma, orig_pud);
2738 	tlb_remove_pud_tlb_entry(tlb, pud, addr);
2739 	if (!vma_is_dax(vma) && vma_is_special_huge(vma)) {
2740 		spin_unlock(ptl);
2741 		/* No zero page support yet */
2742 	} else {
2743 		struct page *page = NULL;
2744 		struct folio *folio;
2745 
2746 		/* No support for anonymous PUD pages or migration yet */
2747 		VM_WARN_ON_ONCE(vma_is_anonymous(vma) ||
2748 				!pud_present(orig_pud));
2749 
2750 		page = pud_page(orig_pud);
2751 		folio = page_folio(page);
2752 		folio_remove_rmap_pud(folio, page, vma);
2753 		add_mm_counter(tlb->mm, mm_counter_file(folio), -HPAGE_PUD_NR);
2754 
2755 		spin_unlock(ptl);
2756 		tlb_remove_page_size(tlb, page, HPAGE_PUD_SIZE);
2757 	}
2758 	return 1;
2759 }
2760 
2761 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2762 		unsigned long haddr)
2763 {
2764 	struct folio *folio;
2765 	struct page *page;
2766 	pud_t old_pud;
2767 
2768 	VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2769 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2770 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2771 	VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2772 
2773 	count_vm_event(THP_SPLIT_PUD);
2774 
2775 	old_pud = pudp_huge_clear_flush(vma, haddr, pud);
2776 
2777 	if (!vma_is_dax(vma))
2778 		return;
2779 
2780 	page = pud_page(old_pud);
2781 	folio = page_folio(page);
2782 
2783 	if (!folio_test_dirty(folio) && pud_dirty(old_pud))
2784 		folio_mark_dirty(folio);
2785 	if (!folio_test_referenced(folio) && pud_young(old_pud))
2786 		folio_set_referenced(folio);
2787 	folio_remove_rmap_pud(folio, page, vma);
2788 	folio_put(folio);
2789 	add_mm_counter(vma->vm_mm, mm_counter_file(folio),
2790 		-HPAGE_PUD_NR);
2791 }
2792 
2793 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2794 		unsigned long address)
2795 {
2796 	spinlock_t *ptl;
2797 	struct mmu_notifier_range range;
2798 
2799 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2800 				address & HPAGE_PUD_MASK,
2801 				(address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2802 	mmu_notifier_invalidate_range_start(&range);
2803 	ptl = pud_lock(vma->vm_mm, pud);
2804 	if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2805 		goto out;
2806 	__split_huge_pud_locked(vma, pud, range.start);
2807 
2808 out:
2809 	spin_unlock(ptl);
2810 	mmu_notifier_invalidate_range_end(&range);
2811 }
2812 #else
2813 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2814 		unsigned long address)
2815 {
2816 }
2817 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2818 
2819 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2820 		unsigned long haddr, pmd_t *pmd)
2821 {
2822 	struct mm_struct *mm = vma->vm_mm;
2823 	pgtable_t pgtable;
2824 	pmd_t _pmd, old_pmd;
2825 	unsigned long addr;
2826 	pte_t *pte;
2827 	int i;
2828 
2829 	/*
2830 	 * Leave pmd empty until pte is filled note that it is fine to delay
2831 	 * notification until mmu_notifier_invalidate_range_end() as we are
2832 	 * replacing a zero pmd write protected page with a zero pte write
2833 	 * protected page.
2834 	 *
2835 	 * See Documentation/mm/mmu_notifier.rst
2836 	 */
2837 	old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd);
2838 
2839 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2840 	pmd_populate(mm, &_pmd, pgtable);
2841 
2842 	pte = pte_offset_map(&_pmd, haddr);
2843 	VM_BUG_ON(!pte);
2844 	for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2845 		pte_t entry;
2846 
2847 		entry = pfn_pte(my_zero_pfn(addr), vma->vm_page_prot);
2848 		entry = pte_mkspecial(entry);
2849 		if (pmd_uffd_wp(old_pmd))
2850 			entry = pte_mkuffd_wp(entry);
2851 		VM_BUG_ON(!pte_none(ptep_get(pte)));
2852 		set_pte_at(mm, addr, pte, entry);
2853 		pte++;
2854 	}
2855 	pte_unmap(pte - 1);
2856 	smp_wmb(); /* make pte visible before pmd */
2857 	pmd_populate(mm, pmd, pgtable);
2858 }
2859 
2860 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2861 		unsigned long haddr, bool freeze)
2862 {
2863 	struct mm_struct *mm = vma->vm_mm;
2864 	struct folio *folio;
2865 	struct page *page;
2866 	pgtable_t pgtable;
2867 	pmd_t old_pmd, _pmd;
2868 	bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2869 	bool anon_exclusive = false, dirty = false;
2870 	unsigned long addr;
2871 	pte_t *pte;
2872 	int i;
2873 
2874 	VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2875 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2876 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2877 	VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2878 				&& !pmd_devmap(*pmd));
2879 
2880 	count_vm_event(THP_SPLIT_PMD);
2881 
2882 	if (!vma_is_anonymous(vma)) {
2883 		old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd);
2884 		/*
2885 		 * We are going to unmap this huge page. So
2886 		 * just go ahead and zap it
2887 		 */
2888 		if (arch_needs_pgtable_deposit())
2889 			zap_deposited_table(mm, pmd);
2890 		if (!vma_is_dax(vma) && vma_is_special_huge(vma))
2891 			return;
2892 		if (unlikely(is_pmd_migration_entry(old_pmd))) {
2893 			swp_entry_t entry;
2894 
2895 			entry = pmd_to_swp_entry(old_pmd);
2896 			folio = pfn_swap_entry_folio(entry);
2897 		} else if (is_huge_zero_pmd(old_pmd)) {
2898 			return;
2899 		} else {
2900 			page = pmd_page(old_pmd);
2901 			folio = page_folio(page);
2902 			if (!folio_test_dirty(folio) && pmd_dirty(old_pmd))
2903 				folio_mark_dirty(folio);
2904 			if (!folio_test_referenced(folio) && pmd_young(old_pmd))
2905 				folio_set_referenced(folio);
2906 			folio_remove_rmap_pmd(folio, page, vma);
2907 			folio_put(folio);
2908 		}
2909 		add_mm_counter(mm, mm_counter_file(folio), -HPAGE_PMD_NR);
2910 		return;
2911 	}
2912 
2913 	if (is_huge_zero_pmd(*pmd)) {
2914 		/*
2915 		 * FIXME: Do we want to invalidate secondary mmu by calling
2916 		 * mmu_notifier_arch_invalidate_secondary_tlbs() see comments below
2917 		 * inside __split_huge_pmd() ?
2918 		 *
2919 		 * We are going from a zero huge page write protected to zero
2920 		 * small page also write protected so it does not seems useful
2921 		 * to invalidate secondary mmu at this time.
2922 		 */
2923 		return __split_huge_zero_page_pmd(vma, haddr, pmd);
2924 	}
2925 
2926 	pmd_migration = is_pmd_migration_entry(*pmd);
2927 	if (unlikely(pmd_migration)) {
2928 		swp_entry_t entry;
2929 
2930 		old_pmd = *pmd;
2931 		entry = pmd_to_swp_entry(old_pmd);
2932 		page = pfn_swap_entry_to_page(entry);
2933 		write = is_writable_migration_entry(entry);
2934 		if (PageAnon(page))
2935 			anon_exclusive = is_readable_exclusive_migration_entry(entry);
2936 		young = is_migration_entry_young(entry);
2937 		dirty = is_migration_entry_dirty(entry);
2938 		soft_dirty = pmd_swp_soft_dirty(old_pmd);
2939 		uffd_wp = pmd_swp_uffd_wp(old_pmd);
2940 	} else {
2941 		/*
2942 		 * Up to this point the pmd is present and huge and userland has
2943 		 * the whole access to the hugepage during the split (which
2944 		 * happens in place). If we overwrite the pmd with the not-huge
2945 		 * version pointing to the pte here (which of course we could if
2946 		 * all CPUs were bug free), userland could trigger a small page
2947 		 * size TLB miss on the small sized TLB while the hugepage TLB
2948 		 * entry is still established in the huge TLB. Some CPU doesn't
2949 		 * like that. See
2950 		 * http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2951 		 * 383 on page 105. Intel should be safe but is also warns that
2952 		 * it's only safe if the permission and cache attributes of the
2953 		 * two entries loaded in the two TLB is identical (which should
2954 		 * be the case here). But it is generally safer to never allow
2955 		 * small and huge TLB entries for the same virtual address to be
2956 		 * loaded simultaneously. So instead of doing "pmd_populate();
2957 		 * flush_pmd_tlb_range();" we first mark the current pmd
2958 		 * notpresent (atomically because here the pmd_trans_huge must
2959 		 * remain set at all times on the pmd until the split is
2960 		 * complete for this pmd), then we flush the SMP TLB and finally
2961 		 * we write the non-huge version of the pmd entry with
2962 		 * pmd_populate.
2963 		 */
2964 		old_pmd = pmdp_invalidate(vma, haddr, pmd);
2965 		page = pmd_page(old_pmd);
2966 		folio = page_folio(page);
2967 		if (pmd_dirty(old_pmd)) {
2968 			dirty = true;
2969 			folio_set_dirty(folio);
2970 		}
2971 		write = pmd_write(old_pmd);
2972 		young = pmd_young(old_pmd);
2973 		soft_dirty = pmd_soft_dirty(old_pmd);
2974 		uffd_wp = pmd_uffd_wp(old_pmd);
2975 
2976 		VM_WARN_ON_FOLIO(!folio_ref_count(folio), folio);
2977 		VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio);
2978 
2979 		/*
2980 		 * Without "freeze", we'll simply split the PMD, propagating the
2981 		 * PageAnonExclusive() flag for each PTE by setting it for
2982 		 * each subpage -- no need to (temporarily) clear.
2983 		 *
2984 		 * With "freeze" we want to replace mapped pages by
2985 		 * migration entries right away. This is only possible if we
2986 		 * managed to clear PageAnonExclusive() -- see
2987 		 * set_pmd_migration_entry().
2988 		 *
2989 		 * In case we cannot clear PageAnonExclusive(), split the PMD
2990 		 * only and let try_to_migrate_one() fail later.
2991 		 *
2992 		 * See folio_try_share_anon_rmap_pmd(): invalidate PMD first.
2993 		 */
2994 		anon_exclusive = PageAnonExclusive(page);
2995 		if (freeze && anon_exclusive &&
2996 		    folio_try_share_anon_rmap_pmd(folio, page))
2997 			freeze = false;
2998 		if (!freeze) {
2999 			rmap_t rmap_flags = RMAP_NONE;
3000 
3001 			folio_ref_add(folio, HPAGE_PMD_NR - 1);
3002 			if (anon_exclusive)
3003 				rmap_flags |= RMAP_EXCLUSIVE;
3004 			folio_add_anon_rmap_ptes(folio, page, HPAGE_PMD_NR,
3005 						 vma, haddr, rmap_flags);
3006 		}
3007 	}
3008 
3009 	/*
3010 	 * Withdraw the table only after we mark the pmd entry invalid.
3011 	 * This's critical for some architectures (Power).
3012 	 */
3013 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
3014 	pmd_populate(mm, &_pmd, pgtable);
3015 
3016 	pte = pte_offset_map(&_pmd, haddr);
3017 	VM_BUG_ON(!pte);
3018 
3019 	/*
3020 	 * Note that NUMA hinting access restrictions are not transferred to
3021 	 * avoid any possibility of altering permissions across VMAs.
3022 	 */
3023 	if (freeze || pmd_migration) {
3024 		for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
3025 			pte_t entry;
3026 			swp_entry_t swp_entry;
3027 
3028 			if (write)
3029 				swp_entry = make_writable_migration_entry(
3030 							page_to_pfn(page + i));
3031 			else if (anon_exclusive)
3032 				swp_entry = make_readable_exclusive_migration_entry(
3033 							page_to_pfn(page + i));
3034 			else
3035 				swp_entry = make_readable_migration_entry(
3036 							page_to_pfn(page + i));
3037 			if (young)
3038 				swp_entry = make_migration_entry_young(swp_entry);
3039 			if (dirty)
3040 				swp_entry = make_migration_entry_dirty(swp_entry);
3041 			entry = swp_entry_to_pte(swp_entry);
3042 			if (soft_dirty)
3043 				entry = pte_swp_mksoft_dirty(entry);
3044 			if (uffd_wp)
3045 				entry = pte_swp_mkuffd_wp(entry);
3046 
3047 			VM_WARN_ON(!pte_none(ptep_get(pte + i)));
3048 			set_pte_at(mm, addr, pte + i, entry);
3049 		}
3050 	} else {
3051 		pte_t entry;
3052 
3053 		entry = mk_pte(page, READ_ONCE(vma->vm_page_prot));
3054 		if (write)
3055 			entry = pte_mkwrite(entry, vma);
3056 		if (!young)
3057 			entry = pte_mkold(entry);
3058 		/* NOTE: this may set soft-dirty too on some archs */
3059 		if (dirty)
3060 			entry = pte_mkdirty(entry);
3061 		if (soft_dirty)
3062 			entry = pte_mksoft_dirty(entry);
3063 		if (uffd_wp)
3064 			entry = pte_mkuffd_wp(entry);
3065 
3066 		for (i = 0; i < HPAGE_PMD_NR; i++)
3067 			VM_WARN_ON(!pte_none(ptep_get(pte + i)));
3068 
3069 		set_ptes(mm, haddr, pte, entry, HPAGE_PMD_NR);
3070 	}
3071 	pte_unmap(pte);
3072 
3073 	if (!pmd_migration)
3074 		folio_remove_rmap_pmd(folio, page, vma);
3075 	if (freeze)
3076 		put_page(page);
3077 
3078 	smp_wmb(); /* make pte visible before pmd */
3079 	pmd_populate(mm, pmd, pgtable);
3080 }
3081 
3082 void split_huge_pmd_locked(struct vm_area_struct *vma, unsigned long address,
3083 			   pmd_t *pmd, bool freeze, struct folio *folio)
3084 {
3085 	bool pmd_migration = is_pmd_migration_entry(*pmd);
3086 
3087 	VM_WARN_ON_ONCE(folio && !folio_test_pmd_mappable(folio));
3088 	VM_WARN_ON_ONCE(!IS_ALIGNED(address, HPAGE_PMD_SIZE));
3089 	VM_WARN_ON_ONCE(folio && !folio_test_locked(folio));
3090 	VM_BUG_ON(freeze && !folio);
3091 
3092 	/*
3093 	 * When the caller requests to set up a migration entry, we
3094 	 * require a folio to check the PMD against. Otherwise, there
3095 	 * is a risk of replacing the wrong folio.
3096 	 */
3097 	if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) || pmd_migration) {
3098 		/*
3099 		 * Do not apply pmd_folio() to a migration entry; and folio lock
3100 		 * guarantees that it must be of the wrong folio anyway.
3101 		 */
3102 		if (folio && (pmd_migration || folio != pmd_folio(*pmd)))
3103 			return;
3104 		__split_huge_pmd_locked(vma, pmd, address, freeze);
3105 	}
3106 }
3107 
3108 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
3109 		unsigned long address, bool freeze, struct folio *folio)
3110 {
3111 	spinlock_t *ptl;
3112 	struct mmu_notifier_range range;
3113 
3114 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
3115 				address & HPAGE_PMD_MASK,
3116 				(address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
3117 	mmu_notifier_invalidate_range_start(&range);
3118 	ptl = pmd_lock(vma->vm_mm, pmd);
3119 	split_huge_pmd_locked(vma, range.start, pmd, freeze, folio);
3120 	spin_unlock(ptl);
3121 	mmu_notifier_invalidate_range_end(&range);
3122 }
3123 
3124 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
3125 		bool freeze, struct folio *folio)
3126 {
3127 	pmd_t *pmd = mm_find_pmd(vma->vm_mm, address);
3128 
3129 	if (!pmd)
3130 		return;
3131 
3132 	__split_huge_pmd(vma, pmd, address, freeze, folio);
3133 }
3134 
3135 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
3136 {
3137 	/*
3138 	 * If the new address isn't hpage aligned and it could previously
3139 	 * contain an hugepage: check if we need to split an huge pmd.
3140 	 */
3141 	if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
3142 	    range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
3143 			 ALIGN(address, HPAGE_PMD_SIZE)))
3144 		split_huge_pmd_address(vma, address, false, NULL);
3145 }
3146 
3147 void vma_adjust_trans_huge(struct vm_area_struct *vma,
3148 			   unsigned long start,
3149 			   unsigned long end,
3150 			   struct vm_area_struct *next)
3151 {
3152 	/* Check if we need to split start first. */
3153 	split_huge_pmd_if_needed(vma, start);
3154 
3155 	/* Check if we need to split end next. */
3156 	split_huge_pmd_if_needed(vma, end);
3157 
3158 	/* If we're incrementing next->vm_start, we might need to split it. */
3159 	if (next)
3160 		split_huge_pmd_if_needed(next, end);
3161 }
3162 
3163 static void unmap_folio(struct folio *folio)
3164 {
3165 	enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SYNC |
3166 		TTU_BATCH_FLUSH;
3167 
3168 	VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
3169 
3170 	if (folio_test_pmd_mappable(folio))
3171 		ttu_flags |= TTU_SPLIT_HUGE_PMD;
3172 
3173 	/*
3174 	 * Anon pages need migration entries to preserve them, but file
3175 	 * pages can simply be left unmapped, then faulted back on demand.
3176 	 * If that is ever changed (perhaps for mlock), update remap_page().
3177 	 */
3178 	if (folio_test_anon(folio))
3179 		try_to_migrate(folio, ttu_flags);
3180 	else
3181 		try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK);
3182 
3183 	try_to_unmap_flush();
3184 }
3185 
3186 static bool __discard_anon_folio_pmd_locked(struct vm_area_struct *vma,
3187 					    unsigned long addr, pmd_t *pmdp,
3188 					    struct folio *folio)
3189 {
3190 	struct mm_struct *mm = vma->vm_mm;
3191 	int ref_count, map_count;
3192 	pmd_t orig_pmd = *pmdp;
3193 
3194 	if (pmd_dirty(orig_pmd))
3195 		folio_set_dirty(folio);
3196 	if (folio_test_dirty(folio) && !(vma->vm_flags & VM_DROPPABLE)) {
3197 		folio_set_swapbacked(folio);
3198 		return false;
3199 	}
3200 
3201 	orig_pmd = pmdp_huge_clear_flush(vma, addr, pmdp);
3202 
3203 	/*
3204 	 * Syncing against concurrent GUP-fast:
3205 	 * - clear PMD; barrier; read refcount
3206 	 * - inc refcount; barrier; read PMD
3207 	 */
3208 	smp_mb();
3209 
3210 	ref_count = folio_ref_count(folio);
3211 	map_count = folio_mapcount(folio);
3212 
3213 	/*
3214 	 * Order reads for folio refcount and dirty flag
3215 	 * (see comments in __remove_mapping()).
3216 	 */
3217 	smp_rmb();
3218 
3219 	/*
3220 	 * If the folio or its PMD is redirtied at this point, or if there
3221 	 * are unexpected references, we will give up to discard this folio
3222 	 * and remap it.
3223 	 *
3224 	 * The only folio refs must be one from isolation plus the rmap(s).
3225 	 */
3226 	if (pmd_dirty(orig_pmd))
3227 		folio_set_dirty(folio);
3228 	if (folio_test_dirty(folio) && !(vma->vm_flags & VM_DROPPABLE)) {
3229 		folio_set_swapbacked(folio);
3230 		set_pmd_at(mm, addr, pmdp, orig_pmd);
3231 		return false;
3232 	}
3233 
3234 	if (ref_count != map_count + 1) {
3235 		set_pmd_at(mm, addr, pmdp, orig_pmd);
3236 		return false;
3237 	}
3238 
3239 	folio_remove_rmap_pmd(folio, pmd_page(orig_pmd), vma);
3240 	zap_deposited_table(mm, pmdp);
3241 	add_mm_counter(mm, MM_ANONPAGES, -HPAGE_PMD_NR);
3242 	if (vma->vm_flags & VM_LOCKED)
3243 		mlock_drain_local();
3244 	folio_put(folio);
3245 
3246 	return true;
3247 }
3248 
3249 bool unmap_huge_pmd_locked(struct vm_area_struct *vma, unsigned long addr,
3250 			   pmd_t *pmdp, struct folio *folio)
3251 {
3252 	VM_WARN_ON_FOLIO(!folio_test_pmd_mappable(folio), folio);
3253 	VM_WARN_ON_FOLIO(!folio_test_locked(folio), folio);
3254 	VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio);
3255 	VM_WARN_ON_FOLIO(folio_test_swapbacked(folio), folio);
3256 	VM_WARN_ON_ONCE(!IS_ALIGNED(addr, HPAGE_PMD_SIZE));
3257 
3258 	return __discard_anon_folio_pmd_locked(vma, addr, pmdp, folio);
3259 }
3260 
3261 static void remap_page(struct folio *folio, unsigned long nr, int flags)
3262 {
3263 	int i = 0;
3264 
3265 	/* If unmap_folio() uses try_to_migrate() on file, remove this check */
3266 	if (!folio_test_anon(folio))
3267 		return;
3268 	for (;;) {
3269 		remove_migration_ptes(folio, folio, RMP_LOCKED | flags);
3270 		i += folio_nr_pages(folio);
3271 		if (i >= nr)
3272 			break;
3273 		folio = folio_next(folio);
3274 	}
3275 }
3276 
3277 static void lru_add_split_folio(struct folio *folio, struct folio *new_folio,
3278 		struct lruvec *lruvec, struct list_head *list)
3279 {
3280 	VM_BUG_ON_FOLIO(folio_test_lru(new_folio), folio);
3281 	lockdep_assert_held(&lruvec->lru_lock);
3282 
3283 	if (list) {
3284 		/* page reclaim is reclaiming a huge page */
3285 		VM_WARN_ON(folio_test_lru(folio));
3286 		folio_get(new_folio);
3287 		list_add_tail(&new_folio->lru, list);
3288 	} else {
3289 		/* head is still on lru (and we have it frozen) */
3290 		VM_WARN_ON(!folio_test_lru(folio));
3291 		if (folio_test_unevictable(folio))
3292 			new_folio->mlock_count = 0;
3293 		else
3294 			list_add_tail(&new_folio->lru, &folio->lru);
3295 		folio_set_lru(new_folio);
3296 	}
3297 }
3298 
3299 /* Racy check whether the huge page can be split */
3300 bool can_split_folio(struct folio *folio, int caller_pins, int *pextra_pins)
3301 {
3302 	int extra_pins;
3303 
3304 	/* Additional pins from page cache */
3305 	if (folio_test_anon(folio))
3306 		extra_pins = folio_test_swapcache(folio) ?
3307 				folio_nr_pages(folio) : 0;
3308 	else
3309 		extra_pins = folio_nr_pages(folio);
3310 	if (pextra_pins)
3311 		*pextra_pins = extra_pins;
3312 	return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins -
3313 					caller_pins;
3314 }
3315 
3316 /*
3317  * It splits @folio into @new_order folios and copies the @folio metadata to
3318  * all the resulting folios.
3319  */
3320 static void __split_folio_to_order(struct folio *folio, int old_order,
3321 		int new_order)
3322 {
3323 	long new_nr_pages = 1 << new_order;
3324 	long nr_pages = 1 << old_order;
3325 	long i;
3326 
3327 	/*
3328 	 * Skip the first new_nr_pages, since the new folio from them have all
3329 	 * the flags from the original folio.
3330 	 */
3331 	for (i = new_nr_pages; i < nr_pages; i += new_nr_pages) {
3332 		struct page *new_head = &folio->page + i;
3333 
3334 		/*
3335 		 * Careful: new_folio is not a "real" folio before we cleared PageTail.
3336 		 * Don't pass it around before clear_compound_head().
3337 		 */
3338 		struct folio *new_folio = (struct folio *)new_head;
3339 
3340 		VM_BUG_ON_PAGE(atomic_read(&new_folio->_mapcount) != -1, new_head);
3341 
3342 		/*
3343 		 * Clone page flags before unfreezing refcount.
3344 		 *
3345 		 * After successful get_page_unless_zero() might follow flags change,
3346 		 * for example lock_page() which set PG_waiters.
3347 		 *
3348 		 * Note that for mapped sub-pages of an anonymous THP,
3349 		 * PG_anon_exclusive has been cleared in unmap_folio() and is stored in
3350 		 * the migration entry instead from where remap_page() will restore it.
3351 		 * We can still have PG_anon_exclusive set on effectively unmapped and
3352 		 * unreferenced sub-pages of an anonymous THP: we can simply drop
3353 		 * PG_anon_exclusive (-> PG_mappedtodisk) for these here.
3354 		 */
3355 		new_folio->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
3356 		new_folio->flags |= (folio->flags &
3357 				((1L << PG_referenced) |
3358 				 (1L << PG_swapbacked) |
3359 				 (1L << PG_swapcache) |
3360 				 (1L << PG_mlocked) |
3361 				 (1L << PG_uptodate) |
3362 				 (1L << PG_active) |
3363 				 (1L << PG_workingset) |
3364 				 (1L << PG_locked) |
3365 				 (1L << PG_unevictable) |
3366 #ifdef CONFIG_ARCH_USES_PG_ARCH_2
3367 				 (1L << PG_arch_2) |
3368 #endif
3369 #ifdef CONFIG_ARCH_USES_PG_ARCH_3
3370 				 (1L << PG_arch_3) |
3371 #endif
3372 				 (1L << PG_dirty) |
3373 				 LRU_GEN_MASK | LRU_REFS_MASK));
3374 
3375 		new_folio->mapping = folio->mapping;
3376 		new_folio->index = folio->index + i;
3377 
3378 		/*
3379 		 * page->private should not be set in tail pages. Fix up and warn once
3380 		 * if private is unexpectedly set.
3381 		 */
3382 		if (unlikely(new_folio->private)) {
3383 			VM_WARN_ON_ONCE_PAGE(true, new_head);
3384 			new_folio->private = NULL;
3385 		}
3386 
3387 		if (folio_test_swapcache(folio))
3388 			new_folio->swap.val = folio->swap.val + i;
3389 
3390 		/* Page flags must be visible before we make the page non-compound. */
3391 		smp_wmb();
3392 
3393 		/*
3394 		 * Clear PageTail before unfreezing page refcount.
3395 		 *
3396 		 * After successful get_page_unless_zero() might follow put_page()
3397 		 * which needs correct compound_head().
3398 		 */
3399 		clear_compound_head(new_head);
3400 		if (new_order) {
3401 			prep_compound_page(new_head, new_order);
3402 			folio_set_large_rmappable(new_folio);
3403 		}
3404 
3405 		if (folio_test_young(folio))
3406 			folio_set_young(new_folio);
3407 		if (folio_test_idle(folio))
3408 			folio_set_idle(new_folio);
3409 #ifdef CONFIG_MEMCG
3410 		new_folio->memcg_data = folio->memcg_data;
3411 #endif
3412 
3413 		folio_xchg_last_cpupid(new_folio, folio_last_cpupid(folio));
3414 	}
3415 
3416 	if (new_order)
3417 		folio_set_order(folio, new_order);
3418 	else
3419 		ClearPageCompound(&folio->page);
3420 }
3421 
3422 /*
3423  * It splits an unmapped @folio to lower order smaller folios in two ways.
3424  * @folio: the to-be-split folio
3425  * @new_order: the smallest order of the after split folios (since buddy
3426  *             allocator like split generates folios with orders from @folio's
3427  *             order - 1 to new_order).
3428  * @split_at: in buddy allocator like split, the folio containing @split_at
3429  *            will be split until its order becomes @new_order.
3430  * @lock_at: the folio containing @lock_at is left locked for caller.
3431  * @list: the after split folios will be added to @list if it is not NULL,
3432  *        otherwise to LRU lists.
3433  * @end: the end of the file @folio maps to. -1 if @folio is anonymous memory.
3434  * @xas: xa_state pointing to folio->mapping->i_pages and locked by caller
3435  * @mapping: @folio->mapping
3436  * @uniform_split: if the split is uniform or not (buddy allocator like split)
3437  *
3438  *
3439  * 1. uniform split: the given @folio into multiple @new_order small folios,
3440  *    where all small folios have the same order. This is done when
3441  *    uniform_split is true.
3442  * 2. buddy allocator like (non-uniform) split: the given @folio is split into
3443  *    half and one of the half (containing the given page) is split into half
3444  *    until the given @page's order becomes @new_order. This is done when
3445  *    uniform_split is false.
3446  *
3447  * The high level flow for these two methods are:
3448  * 1. uniform split: a single __split_folio_to_order() is called to split the
3449  *    @folio into @new_order, then we traverse all the resulting folios one by
3450  *    one in PFN ascending order and perform stats, unfreeze, adding to list,
3451  *    and file mapping index operations.
3452  * 2. non-uniform split: in general, folio_order - @new_order calls to
3453  *    __split_folio_to_order() are made in a for loop to split the @folio
3454  *    to one lower order at a time. The resulting small folios are processed
3455  *    like what is done during the traversal in 1, except the one containing
3456  *    @page, which is split in next for loop.
3457  *
3458  * After splitting, the caller's folio reference will be transferred to the
3459  * folio containing @page. The other folios may be freed if they are not mapped.
3460  *
3461  * In terms of locking, after splitting,
3462  * 1. uniform split leaves @page (or the folio contains it) locked;
3463  * 2. buddy allocator like (non-uniform) split leaves @folio locked.
3464  *
3465  *
3466  * For !uniform_split, when -ENOMEM is returned, the original folio might be
3467  * split. The caller needs to check the input folio.
3468  */
3469 static int __split_unmapped_folio(struct folio *folio, int new_order,
3470 		struct page *split_at, struct page *lock_at,
3471 		struct list_head *list, pgoff_t end,
3472 		struct xa_state *xas, struct address_space *mapping,
3473 		bool uniform_split)
3474 {
3475 	struct lruvec *lruvec;
3476 	struct address_space *swap_cache = NULL;
3477 	struct folio *origin_folio = folio;
3478 	struct folio *next_folio = folio_next(folio);
3479 	struct folio *new_folio;
3480 	struct folio *next;
3481 	int order = folio_order(folio);
3482 	int split_order;
3483 	int start_order = uniform_split ? new_order : order - 1;
3484 	int nr_dropped = 0;
3485 	int ret = 0;
3486 	bool stop_split = false;
3487 
3488 	if (folio_test_swapcache(folio)) {
3489 		VM_BUG_ON(mapping);
3490 
3491 		/* a swapcache folio can only be uniformly split to order-0 */
3492 		if (!uniform_split || new_order != 0)
3493 			return -EINVAL;
3494 
3495 		swap_cache = swap_address_space(folio->swap);
3496 		xa_lock(&swap_cache->i_pages);
3497 	}
3498 
3499 	if (folio_test_anon(folio))
3500 		mod_mthp_stat(order, MTHP_STAT_NR_ANON, -1);
3501 
3502 	/* lock lru list/PageCompound, ref frozen by page_ref_freeze */
3503 	lruvec = folio_lruvec_lock(folio);
3504 
3505 	folio_clear_has_hwpoisoned(folio);
3506 
3507 	/*
3508 	 * split to new_order one order at a time. For uniform split,
3509 	 * folio is split to new_order directly.
3510 	 */
3511 	for (split_order = start_order;
3512 	     split_order >= new_order && !stop_split;
3513 	     split_order--) {
3514 		int old_order = folio_order(folio);
3515 		struct folio *release;
3516 		struct folio *end_folio = folio_next(folio);
3517 
3518 		/* order-1 anonymous folio is not supported */
3519 		if (folio_test_anon(folio) && split_order == 1)
3520 			continue;
3521 		if (uniform_split && split_order != new_order)
3522 			continue;
3523 
3524 		if (mapping) {
3525 			/*
3526 			 * uniform split has xas_split_alloc() called before
3527 			 * irq is disabled to allocate enough memory, whereas
3528 			 * non-uniform split can handle ENOMEM.
3529 			 */
3530 			if (uniform_split)
3531 				xas_split(xas, folio, old_order);
3532 			else {
3533 				xas_set_order(xas, folio->index, split_order);
3534 				xas_try_split(xas, folio, old_order);
3535 				if (xas_error(xas)) {
3536 					ret = xas_error(xas);
3537 					stop_split = true;
3538 					goto after_split;
3539 				}
3540 			}
3541 		}
3542 
3543 		folio_split_memcg_refs(folio, old_order, split_order);
3544 		split_page_owner(&folio->page, old_order, split_order);
3545 		pgalloc_tag_split(folio, old_order, split_order);
3546 
3547 		__split_folio_to_order(folio, old_order, split_order);
3548 
3549 after_split:
3550 		/*
3551 		 * Iterate through after-split folios and perform related
3552 		 * operations. But in buddy allocator like split, the folio
3553 		 * containing the specified page is skipped until its order
3554 		 * is new_order, since the folio will be worked on in next
3555 		 * iteration.
3556 		 */
3557 		for (release = folio; release != end_folio; release = next) {
3558 			next = folio_next(release);
3559 			/*
3560 			 * for buddy allocator like split, the folio containing
3561 			 * page will be split next and should not be released,
3562 			 * until the folio's order is new_order or stop_split
3563 			 * is set to true by the above xas_split() failure.
3564 			 */
3565 			if (release == page_folio(split_at)) {
3566 				folio = release;
3567 				if (split_order != new_order && !stop_split)
3568 					continue;
3569 			}
3570 			if (folio_test_anon(release)) {
3571 				mod_mthp_stat(folio_order(release),
3572 						MTHP_STAT_NR_ANON, 1);
3573 			}
3574 
3575 			/*
3576 			 * origin_folio should be kept frozon until page cache
3577 			 * entries are updated with all the other after-split
3578 			 * folios to prevent others seeing stale page cache
3579 			 * entries.
3580 			 */
3581 			if (release == origin_folio)
3582 				continue;
3583 
3584 			folio_ref_unfreeze(release, 1 +
3585 					((mapping || swap_cache) ?
3586 						folio_nr_pages(release) : 0));
3587 
3588 			lru_add_split_folio(origin_folio, release, lruvec,
3589 					list);
3590 
3591 			/* Some pages can be beyond EOF: drop them from cache */
3592 			if (release->index >= end) {
3593 				if (shmem_mapping(mapping))
3594 					nr_dropped += folio_nr_pages(release);
3595 				else if (folio_test_clear_dirty(release))
3596 					folio_account_cleaned(release,
3597 						inode_to_wb(mapping->host));
3598 				__filemap_remove_folio(release, NULL);
3599 				folio_put_refs(release, folio_nr_pages(release));
3600 			} else if (mapping) {
3601 				__xa_store(&mapping->i_pages,
3602 						release->index, release, 0);
3603 			} else if (swap_cache) {
3604 				__xa_store(&swap_cache->i_pages,
3605 						swap_cache_index(release->swap),
3606 						release, 0);
3607 			}
3608 		}
3609 	}
3610 
3611 	/*
3612 	 * Unfreeze origin_folio only after all page cache entries, which used
3613 	 * to point to it, have been updated with new folios. Otherwise,
3614 	 * a parallel folio_try_get() can grab origin_folio and its caller can
3615 	 * see stale page cache entries.
3616 	 */
3617 	folio_ref_unfreeze(origin_folio, 1 +
3618 		((mapping || swap_cache) ? folio_nr_pages(origin_folio) : 0));
3619 
3620 	unlock_page_lruvec(lruvec);
3621 
3622 	if (swap_cache)
3623 		xa_unlock(&swap_cache->i_pages);
3624 	if (mapping)
3625 		xa_unlock(&mapping->i_pages);
3626 
3627 	/* Caller disabled irqs, so they are still disabled here */
3628 	local_irq_enable();
3629 
3630 	if (nr_dropped)
3631 		shmem_uncharge(mapping->host, nr_dropped);
3632 
3633 	remap_page(origin_folio, 1 << order,
3634 			folio_test_anon(origin_folio) ?
3635 				RMP_USE_SHARED_ZEROPAGE : 0);
3636 
3637 	/*
3638 	 * At this point, folio should contain the specified page.
3639 	 * For uniform split, it is left for caller to unlock.
3640 	 * For buddy allocator like split, the first after-split folio is left
3641 	 * for caller to unlock.
3642 	 */
3643 	for (new_folio = origin_folio; new_folio != next_folio; new_folio = next) {
3644 		next = folio_next(new_folio);
3645 		if (new_folio == page_folio(lock_at))
3646 			continue;
3647 
3648 		folio_unlock(new_folio);
3649 		/*
3650 		 * Subpages may be freed if there wasn't any mapping
3651 		 * like if add_to_swap() is running on a lru page that
3652 		 * had its mapping zapped. And freeing these pages
3653 		 * requires taking the lru_lock so we do the put_page
3654 		 * of the tail pages after the split is complete.
3655 		 */
3656 		free_folio_and_swap_cache(new_folio);
3657 	}
3658 	return ret;
3659 }
3660 
3661 bool non_uniform_split_supported(struct folio *folio, unsigned int new_order,
3662 		bool warns)
3663 {
3664 	if (folio_test_anon(folio)) {
3665 		/* order-1 is not supported for anonymous THP. */
3666 		VM_WARN_ONCE(warns && new_order == 1,
3667 				"Cannot split to order-1 folio");
3668 		return new_order != 1;
3669 	} else if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&
3670 	    !mapping_large_folio_support(folio->mapping)) {
3671 		/*
3672 		 * No split if the file system does not support large folio.
3673 		 * Note that we might still have THPs in such mappings due to
3674 		 * CONFIG_READ_ONLY_THP_FOR_FS. But in that case, the mapping
3675 		 * does not actually support large folios properly.
3676 		 */
3677 		VM_WARN_ONCE(warns,
3678 			"Cannot split file folio to non-0 order");
3679 		return false;
3680 	}
3681 
3682 	/* Only swapping a whole PMD-mapped folio is supported */
3683 	if (folio_test_swapcache(folio)) {
3684 		VM_WARN_ONCE(warns,
3685 			"Cannot split swapcache folio to non-0 order");
3686 		return false;
3687 	}
3688 
3689 	return true;
3690 }
3691 
3692 /* See comments in non_uniform_split_supported() */
3693 bool uniform_split_supported(struct folio *folio, unsigned int new_order,
3694 		bool warns)
3695 {
3696 	if (folio_test_anon(folio)) {
3697 		VM_WARN_ONCE(warns && new_order == 1,
3698 				"Cannot split to order-1 folio");
3699 		return new_order != 1;
3700 	} else  if (new_order) {
3701 		if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&
3702 		    !mapping_large_folio_support(folio->mapping)) {
3703 			VM_WARN_ONCE(warns,
3704 				"Cannot split file folio to non-0 order");
3705 			return false;
3706 		}
3707 	}
3708 
3709 	if (new_order && folio_test_swapcache(folio)) {
3710 		VM_WARN_ONCE(warns,
3711 			"Cannot split swapcache folio to non-0 order");
3712 		return false;
3713 	}
3714 
3715 	return true;
3716 }
3717 
3718 /*
3719  * __folio_split: split a folio at @split_at to a @new_order folio
3720  * @folio: folio to split
3721  * @new_order: the order of the new folio
3722  * @split_at: a page within the new folio
3723  * @lock_at: a page within @folio to be left locked to caller
3724  * @list: after-split folios will be put on it if non NULL
3725  * @uniform_split: perform uniform split or not (non-uniform split)
3726  *
3727  * It calls __split_unmapped_folio() to perform uniform and non-uniform split.
3728  * It is in charge of checking whether the split is supported or not and
3729  * preparing @folio for __split_unmapped_folio().
3730  *
3731  * return: 0: successful, <0 failed (if -ENOMEM is returned, @folio might be
3732  * split but not to @new_order, the caller needs to check)
3733  */
3734 static int __folio_split(struct folio *folio, unsigned int new_order,
3735 		struct page *split_at, struct page *lock_at,
3736 		struct list_head *list, bool uniform_split)
3737 {
3738 	struct deferred_split *ds_queue = get_deferred_split_queue(folio);
3739 	XA_STATE(xas, &folio->mapping->i_pages, folio->index);
3740 	bool is_anon = folio_test_anon(folio);
3741 	struct address_space *mapping = NULL;
3742 	struct anon_vma *anon_vma = NULL;
3743 	int order = folio_order(folio);
3744 	int extra_pins, ret;
3745 	pgoff_t end;
3746 	bool is_hzp;
3747 
3748 	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
3749 	VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
3750 
3751 	if (folio != page_folio(split_at) || folio != page_folio(lock_at))
3752 		return -EINVAL;
3753 
3754 	if (new_order >= folio_order(folio))
3755 		return -EINVAL;
3756 
3757 	if (uniform_split && !uniform_split_supported(folio, new_order, true))
3758 		return -EINVAL;
3759 
3760 	if (!uniform_split &&
3761 	    !non_uniform_split_supported(folio, new_order, true))
3762 		return -EINVAL;
3763 
3764 	is_hzp = is_huge_zero_folio(folio);
3765 	if (is_hzp) {
3766 		pr_warn_ratelimited("Called split_huge_page for huge zero page\n");
3767 		return -EBUSY;
3768 	}
3769 
3770 	if (folio_test_writeback(folio))
3771 		return -EBUSY;
3772 
3773 	if (is_anon) {
3774 		/*
3775 		 * The caller does not necessarily hold an mmap_lock that would
3776 		 * prevent the anon_vma disappearing so we first we take a
3777 		 * reference to it and then lock the anon_vma for write. This
3778 		 * is similar to folio_lock_anon_vma_read except the write lock
3779 		 * is taken to serialise against parallel split or collapse
3780 		 * operations.
3781 		 */
3782 		anon_vma = folio_get_anon_vma(folio);
3783 		if (!anon_vma) {
3784 			ret = -EBUSY;
3785 			goto out;
3786 		}
3787 		end = -1;
3788 		mapping = NULL;
3789 		anon_vma_lock_write(anon_vma);
3790 	} else {
3791 		unsigned int min_order;
3792 		gfp_t gfp;
3793 
3794 		mapping = folio->mapping;
3795 
3796 		/* Truncated ? */
3797 		/*
3798 		 * TODO: add support for large shmem folio in swap cache.
3799 		 * When shmem is in swap cache, mapping is NULL and
3800 		 * folio_test_swapcache() is true.
3801 		 */
3802 		if (!mapping) {
3803 			ret = -EBUSY;
3804 			goto out;
3805 		}
3806 
3807 		min_order = mapping_min_folio_order(folio->mapping);
3808 		if (new_order < min_order) {
3809 			VM_WARN_ONCE(1, "Cannot split mapped folio below min-order: %u",
3810 				     min_order);
3811 			ret = -EINVAL;
3812 			goto out;
3813 		}
3814 
3815 		gfp = current_gfp_context(mapping_gfp_mask(mapping) &
3816 							GFP_RECLAIM_MASK);
3817 
3818 		if (!filemap_release_folio(folio, gfp)) {
3819 			ret = -EBUSY;
3820 			goto out;
3821 		}
3822 
3823 		if (uniform_split) {
3824 			xas_set_order(&xas, folio->index, new_order);
3825 			xas_split_alloc(&xas, folio, folio_order(folio), gfp);
3826 			if (xas_error(&xas)) {
3827 				ret = xas_error(&xas);
3828 				goto out;
3829 			}
3830 		}
3831 
3832 		anon_vma = NULL;
3833 		i_mmap_lock_read(mapping);
3834 
3835 		/*
3836 		 *__split_unmapped_folio() may need to trim off pages beyond
3837 		 * EOF: but on 32-bit, i_size_read() takes an irq-unsafe
3838 		 * seqlock, which cannot be nested inside the page tree lock.
3839 		 * So note end now: i_size itself may be changed at any moment,
3840 		 * but folio lock is good enough to serialize the trimming.
3841 		 */
3842 		end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
3843 		if (shmem_mapping(mapping))
3844 			end = shmem_fallocend(mapping->host, end);
3845 	}
3846 
3847 	/*
3848 	 * Racy check if we can split the page, before unmap_folio() will
3849 	 * split PMDs
3850 	 */
3851 	if (!can_split_folio(folio, 1, &extra_pins)) {
3852 		ret = -EAGAIN;
3853 		goto out_unlock;
3854 	}
3855 
3856 	unmap_folio(folio);
3857 
3858 	/* block interrupt reentry in xa_lock and spinlock */
3859 	local_irq_disable();
3860 	if (mapping) {
3861 		/*
3862 		 * Check if the folio is present in page cache.
3863 		 * We assume all tail are present too, if folio is there.
3864 		 */
3865 		xas_lock(&xas);
3866 		xas_reset(&xas);
3867 		if (xas_load(&xas) != folio)
3868 			goto fail;
3869 	}
3870 
3871 	/* Prevent deferred_split_scan() touching ->_refcount */
3872 	spin_lock(&ds_queue->split_queue_lock);
3873 	if (folio_ref_freeze(folio, 1 + extra_pins)) {
3874 		if (folio_order(folio) > 1 &&
3875 		    !list_empty(&folio->_deferred_list)) {
3876 			ds_queue->split_queue_len--;
3877 			if (folio_test_partially_mapped(folio)) {
3878 				folio_clear_partially_mapped(folio);
3879 				mod_mthp_stat(folio_order(folio),
3880 					      MTHP_STAT_NR_ANON_PARTIALLY_MAPPED, -1);
3881 			}
3882 			/*
3883 			 * Reinitialize page_deferred_list after removing the
3884 			 * page from the split_queue, otherwise a subsequent
3885 			 * split will see list corruption when checking the
3886 			 * page_deferred_list.
3887 			 */
3888 			list_del_init(&folio->_deferred_list);
3889 		}
3890 		spin_unlock(&ds_queue->split_queue_lock);
3891 		if (mapping) {
3892 			int nr = folio_nr_pages(folio);
3893 
3894 			if (folio_test_pmd_mappable(folio) &&
3895 			    new_order < HPAGE_PMD_ORDER) {
3896 				if (folio_test_swapbacked(folio)) {
3897 					__lruvec_stat_mod_folio(folio,
3898 							NR_SHMEM_THPS, -nr);
3899 				} else {
3900 					__lruvec_stat_mod_folio(folio,
3901 							NR_FILE_THPS, -nr);
3902 					filemap_nr_thps_dec(mapping);
3903 				}
3904 			}
3905 		}
3906 
3907 		ret = __split_unmapped_folio(folio, new_order,
3908 				split_at, lock_at, list, end, &xas, mapping,
3909 				uniform_split);
3910 	} else {
3911 		spin_unlock(&ds_queue->split_queue_lock);
3912 fail:
3913 		if (mapping)
3914 			xas_unlock(&xas);
3915 		local_irq_enable();
3916 		remap_page(folio, folio_nr_pages(folio), 0);
3917 		ret = -EAGAIN;
3918 	}
3919 
3920 out_unlock:
3921 	if (anon_vma) {
3922 		anon_vma_unlock_write(anon_vma);
3923 		put_anon_vma(anon_vma);
3924 	}
3925 	if (mapping)
3926 		i_mmap_unlock_read(mapping);
3927 out:
3928 	xas_destroy(&xas);
3929 	if (order == HPAGE_PMD_ORDER)
3930 		count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
3931 	count_mthp_stat(order, !ret ? MTHP_STAT_SPLIT : MTHP_STAT_SPLIT_FAILED);
3932 	return ret;
3933 }
3934 
3935 /*
3936  * This function splits a large folio into smaller folios of order @new_order.
3937  * @page can point to any page of the large folio to split. The split operation
3938  * does not change the position of @page.
3939  *
3940  * Prerequisites:
3941  *
3942  * 1) The caller must hold a reference on the @page's owning folio, also known
3943  *    as the large folio.
3944  *
3945  * 2) The large folio must be locked.
3946  *
3947  * 3) The folio must not be pinned. Any unexpected folio references, including
3948  *    GUP pins, will result in the folio not getting split; instead, the caller
3949  *    will receive an -EAGAIN.
3950  *
3951  * 4) @new_order > 1, usually. Splitting to order-1 anonymous folios is not
3952  *    supported for non-file-backed folios, because folio->_deferred_list, which
3953  *    is used by partially mapped folios, is stored in subpage 2, but an order-1
3954  *    folio only has subpages 0 and 1. File-backed order-1 folios are supported,
3955  *    since they do not use _deferred_list.
3956  *
3957  * After splitting, the caller's folio reference will be transferred to @page,
3958  * resulting in a raised refcount of @page after this call. The other pages may
3959  * be freed if they are not mapped.
3960  *
3961  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3962  *
3963  * Pages in @new_order will inherit the mapping, flags, and so on from the
3964  * huge page.
3965  *
3966  * Returns 0 if the huge page was split successfully.
3967  *
3968  * Returns -EAGAIN if the folio has unexpected reference (e.g., GUP) or if
3969  * the folio was concurrently removed from the page cache.
3970  *
3971  * Returns -EBUSY when trying to split the huge zeropage, if the folio is
3972  * under writeback, if fs-specific folio metadata cannot currently be
3973  * released, or if some unexpected race happened (e.g., anon VMA disappeared,
3974  * truncation).
3975  *
3976  * Callers should ensure that the order respects the address space mapping
3977  * min-order if one is set for non-anonymous folios.
3978  *
3979  * Returns -EINVAL when trying to split to an order that is incompatible
3980  * with the folio. Splitting to order 0 is compatible with all folios.
3981  */
3982 int split_huge_page_to_list_to_order(struct page *page, struct list_head *list,
3983 				     unsigned int new_order)
3984 {
3985 	struct folio *folio = page_folio(page);
3986 
3987 	return __folio_split(folio, new_order, &folio->page, page, list, true);
3988 }
3989 
3990 /*
3991  * folio_split: split a folio at @split_at to a @new_order folio
3992  * @folio: folio to split
3993  * @new_order: the order of the new folio
3994  * @split_at: a page within the new folio
3995  *
3996  * return: 0: successful, <0 failed (if -ENOMEM is returned, @folio might be
3997  * split but not to @new_order, the caller needs to check)
3998  *
3999  * It has the same prerequisites and returns as
4000  * split_huge_page_to_list_to_order().
4001  *
4002  * Split a folio at @split_at to a new_order folio, leave the
4003  * remaining subpages of the original folio as large as possible. For example,
4004  * in the case of splitting an order-9 folio at its third order-3 subpages to
4005  * an order-3 folio, there are 2^(9-3)=64 order-3 subpages in the order-9 folio.
4006  * After the split, there will be a group of folios with different orders and
4007  * the new folio containing @split_at is marked in bracket:
4008  * [order-4, {order-3}, order-3, order-5, order-6, order-7, order-8].
4009  *
4010  * After split, folio is left locked for caller.
4011  */
4012 int folio_split(struct folio *folio, unsigned int new_order,
4013 		struct page *split_at, struct list_head *list)
4014 {
4015 	return __folio_split(folio, new_order, split_at, &folio->page, list,
4016 			false);
4017 }
4018 
4019 int min_order_for_split(struct folio *folio)
4020 {
4021 	if (folio_test_anon(folio))
4022 		return 0;
4023 
4024 	if (!folio->mapping) {
4025 		if (folio_test_pmd_mappable(folio))
4026 			count_vm_event(THP_SPLIT_PAGE_FAILED);
4027 		return -EBUSY;
4028 	}
4029 
4030 	return mapping_min_folio_order(folio->mapping);
4031 }
4032 
4033 int split_folio_to_list(struct folio *folio, struct list_head *list)
4034 {
4035 	int ret = min_order_for_split(folio);
4036 
4037 	if (ret < 0)
4038 		return ret;
4039 
4040 	return split_huge_page_to_list_to_order(&folio->page, list, ret);
4041 }
4042 
4043 /*
4044  * __folio_unqueue_deferred_split() is not to be called directly:
4045  * the folio_unqueue_deferred_split() inline wrapper in mm/internal.h
4046  * limits its calls to those folios which may have a _deferred_list for
4047  * queueing THP splits, and that list is (racily observed to be) non-empty.
4048  *
4049  * It is unsafe to call folio_unqueue_deferred_split() until folio refcount is
4050  * zero: because even when split_queue_lock is held, a non-empty _deferred_list
4051  * might be in use on deferred_split_scan()'s unlocked on-stack list.
4052  *
4053  * If memory cgroups are enabled, split_queue_lock is in the mem_cgroup: it is
4054  * therefore important to unqueue deferred split before changing folio memcg.
4055  */
4056 bool __folio_unqueue_deferred_split(struct folio *folio)
4057 {
4058 	struct deferred_split *ds_queue;
4059 	unsigned long flags;
4060 	bool unqueued = false;
4061 
4062 	WARN_ON_ONCE(folio_ref_count(folio));
4063 	WARN_ON_ONCE(!mem_cgroup_disabled() && !folio_memcg(folio));
4064 
4065 	ds_queue = get_deferred_split_queue(folio);
4066 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
4067 	if (!list_empty(&folio->_deferred_list)) {
4068 		ds_queue->split_queue_len--;
4069 		if (folio_test_partially_mapped(folio)) {
4070 			folio_clear_partially_mapped(folio);
4071 			mod_mthp_stat(folio_order(folio),
4072 				      MTHP_STAT_NR_ANON_PARTIALLY_MAPPED, -1);
4073 		}
4074 		list_del_init(&folio->_deferred_list);
4075 		unqueued = true;
4076 	}
4077 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
4078 
4079 	return unqueued;	/* useful for debug warnings */
4080 }
4081 
4082 /* partially_mapped=false won't clear PG_partially_mapped folio flag */
4083 void deferred_split_folio(struct folio *folio, bool partially_mapped)
4084 {
4085 	struct deferred_split *ds_queue = get_deferred_split_queue(folio);
4086 #ifdef CONFIG_MEMCG
4087 	struct mem_cgroup *memcg = folio_memcg(folio);
4088 #endif
4089 	unsigned long flags;
4090 
4091 	/*
4092 	 * Order 1 folios have no space for a deferred list, but we also
4093 	 * won't waste much memory by not adding them to the deferred list.
4094 	 */
4095 	if (folio_order(folio) <= 1)
4096 		return;
4097 
4098 	if (!partially_mapped && !split_underused_thp)
4099 		return;
4100 
4101 	/*
4102 	 * Exclude swapcache: originally to avoid a corrupt deferred split
4103 	 * queue. Nowadays that is fully prevented by memcg1_swapout();
4104 	 * but if page reclaim is already handling the same folio, it is
4105 	 * unnecessary to handle it again in the shrinker, so excluding
4106 	 * swapcache here may still be a useful optimization.
4107 	 */
4108 	if (folio_test_swapcache(folio))
4109 		return;
4110 
4111 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
4112 	if (partially_mapped) {
4113 		if (!folio_test_partially_mapped(folio)) {
4114 			folio_set_partially_mapped(folio);
4115 			if (folio_test_pmd_mappable(folio))
4116 				count_vm_event(THP_DEFERRED_SPLIT_PAGE);
4117 			count_mthp_stat(folio_order(folio), MTHP_STAT_SPLIT_DEFERRED);
4118 			mod_mthp_stat(folio_order(folio), MTHP_STAT_NR_ANON_PARTIALLY_MAPPED, 1);
4119 
4120 		}
4121 	} else {
4122 		/* partially mapped folios cannot become non-partially mapped */
4123 		VM_WARN_ON_FOLIO(folio_test_partially_mapped(folio), folio);
4124 	}
4125 	if (list_empty(&folio->_deferred_list)) {
4126 		list_add_tail(&folio->_deferred_list, &ds_queue->split_queue);
4127 		ds_queue->split_queue_len++;
4128 #ifdef CONFIG_MEMCG
4129 		if (memcg)
4130 			set_shrinker_bit(memcg, folio_nid(folio),
4131 					 deferred_split_shrinker->id);
4132 #endif
4133 	}
4134 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
4135 }
4136 
4137 static unsigned long deferred_split_count(struct shrinker *shrink,
4138 		struct shrink_control *sc)
4139 {
4140 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
4141 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
4142 
4143 #ifdef CONFIG_MEMCG
4144 	if (sc->memcg)
4145 		ds_queue = &sc->memcg->deferred_split_queue;
4146 #endif
4147 	return READ_ONCE(ds_queue->split_queue_len);
4148 }
4149 
4150 static bool thp_underused(struct folio *folio)
4151 {
4152 	int num_zero_pages = 0, num_filled_pages = 0;
4153 	void *kaddr;
4154 	int i;
4155 
4156 	if (khugepaged_max_ptes_none == HPAGE_PMD_NR - 1)
4157 		return false;
4158 
4159 	for (i = 0; i < folio_nr_pages(folio); i++) {
4160 		kaddr = kmap_local_folio(folio, i * PAGE_SIZE);
4161 		if (!memchr_inv(kaddr, 0, PAGE_SIZE)) {
4162 			num_zero_pages++;
4163 			if (num_zero_pages > khugepaged_max_ptes_none) {
4164 				kunmap_local(kaddr);
4165 				return true;
4166 			}
4167 		} else {
4168 			/*
4169 			 * Another path for early exit once the number
4170 			 * of non-zero filled pages exceeds threshold.
4171 			 */
4172 			num_filled_pages++;
4173 			if (num_filled_pages >= HPAGE_PMD_NR - khugepaged_max_ptes_none) {
4174 				kunmap_local(kaddr);
4175 				return false;
4176 			}
4177 		}
4178 		kunmap_local(kaddr);
4179 	}
4180 	return false;
4181 }
4182 
4183 static unsigned long deferred_split_scan(struct shrinker *shrink,
4184 		struct shrink_control *sc)
4185 {
4186 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
4187 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
4188 	unsigned long flags;
4189 	LIST_HEAD(list);
4190 	struct folio *folio, *next, *prev = NULL;
4191 	int split = 0, removed = 0;
4192 
4193 #ifdef CONFIG_MEMCG
4194 	if (sc->memcg)
4195 		ds_queue = &sc->memcg->deferred_split_queue;
4196 #endif
4197 
4198 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
4199 	/* Take pin on all head pages to avoid freeing them under us */
4200 	list_for_each_entry_safe(folio, next, &ds_queue->split_queue,
4201 							_deferred_list) {
4202 		if (folio_try_get(folio)) {
4203 			list_move(&folio->_deferred_list, &list);
4204 		} else {
4205 			/* We lost race with folio_put() */
4206 			if (folio_test_partially_mapped(folio)) {
4207 				folio_clear_partially_mapped(folio);
4208 				mod_mthp_stat(folio_order(folio),
4209 					      MTHP_STAT_NR_ANON_PARTIALLY_MAPPED, -1);
4210 			}
4211 			list_del_init(&folio->_deferred_list);
4212 			ds_queue->split_queue_len--;
4213 		}
4214 		if (!--sc->nr_to_scan)
4215 			break;
4216 	}
4217 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
4218 
4219 	list_for_each_entry_safe(folio, next, &list, _deferred_list) {
4220 		bool did_split = false;
4221 		bool underused = false;
4222 
4223 		if (!folio_test_partially_mapped(folio)) {
4224 			underused = thp_underused(folio);
4225 			if (!underused)
4226 				goto next;
4227 		}
4228 		if (!folio_trylock(folio))
4229 			goto next;
4230 		if (!split_folio(folio)) {
4231 			did_split = true;
4232 			if (underused)
4233 				count_vm_event(THP_UNDERUSED_SPLIT_PAGE);
4234 			split++;
4235 		}
4236 		folio_unlock(folio);
4237 next:
4238 		/*
4239 		 * split_folio() removes folio from list on success.
4240 		 * Only add back to the queue if folio is partially mapped.
4241 		 * If thp_underused returns false, or if split_folio fails
4242 		 * in the case it was underused, then consider it used and
4243 		 * don't add it back to split_queue.
4244 		 */
4245 		if (did_split) {
4246 			; /* folio already removed from list */
4247 		} else if (!folio_test_partially_mapped(folio)) {
4248 			list_del_init(&folio->_deferred_list);
4249 			removed++;
4250 		} else {
4251 			/*
4252 			 * That unlocked list_del_init() above would be unsafe,
4253 			 * unless its folio is separated from any earlier folios
4254 			 * left on the list (which may be concurrently unqueued)
4255 			 * by one safe folio with refcount still raised.
4256 			 */
4257 			swap(folio, prev);
4258 		}
4259 		if (folio)
4260 			folio_put(folio);
4261 	}
4262 
4263 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
4264 	list_splice_tail(&list, &ds_queue->split_queue);
4265 	ds_queue->split_queue_len -= removed;
4266 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
4267 
4268 	if (prev)
4269 		folio_put(prev);
4270 
4271 	/*
4272 	 * Stop shrinker if we didn't split any page, but the queue is empty.
4273 	 * This can happen if pages were freed under us.
4274 	 */
4275 	if (!split && list_empty(&ds_queue->split_queue))
4276 		return SHRINK_STOP;
4277 	return split;
4278 }
4279 
4280 #ifdef CONFIG_DEBUG_FS
4281 static void split_huge_pages_all(void)
4282 {
4283 	struct zone *zone;
4284 	struct page *page;
4285 	struct folio *folio;
4286 	unsigned long pfn, max_zone_pfn;
4287 	unsigned long total = 0, split = 0;
4288 
4289 	pr_debug("Split all THPs\n");
4290 	for_each_zone(zone) {
4291 		if (!managed_zone(zone))
4292 			continue;
4293 		max_zone_pfn = zone_end_pfn(zone);
4294 		for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
4295 			int nr_pages;
4296 
4297 			page = pfn_to_online_page(pfn);
4298 			if (!page || PageTail(page))
4299 				continue;
4300 			folio = page_folio(page);
4301 			if (!folio_try_get(folio))
4302 				continue;
4303 
4304 			if (unlikely(page_folio(page) != folio))
4305 				goto next;
4306 
4307 			if (zone != folio_zone(folio))
4308 				goto next;
4309 
4310 			if (!folio_test_large(folio)
4311 				|| folio_test_hugetlb(folio)
4312 				|| !folio_test_lru(folio))
4313 				goto next;
4314 
4315 			total++;
4316 			folio_lock(folio);
4317 			nr_pages = folio_nr_pages(folio);
4318 			if (!split_folio(folio))
4319 				split++;
4320 			pfn += nr_pages - 1;
4321 			folio_unlock(folio);
4322 next:
4323 			folio_put(folio);
4324 			cond_resched();
4325 		}
4326 	}
4327 
4328 	pr_debug("%lu of %lu THP split\n", split, total);
4329 }
4330 
4331 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
4332 {
4333 	return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
4334 		    is_vm_hugetlb_page(vma);
4335 }
4336 
4337 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
4338 				unsigned long vaddr_end, unsigned int new_order,
4339 				long in_folio_offset)
4340 {
4341 	int ret = 0;
4342 	struct task_struct *task;
4343 	struct mm_struct *mm;
4344 	unsigned long total = 0, split = 0;
4345 	unsigned long addr;
4346 
4347 	vaddr_start &= PAGE_MASK;
4348 	vaddr_end &= PAGE_MASK;
4349 
4350 	task = find_get_task_by_vpid(pid);
4351 	if (!task) {
4352 		ret = -ESRCH;
4353 		goto out;
4354 	}
4355 
4356 	/* Find the mm_struct */
4357 	mm = get_task_mm(task);
4358 	put_task_struct(task);
4359 
4360 	if (!mm) {
4361 		ret = -EINVAL;
4362 		goto out;
4363 	}
4364 
4365 	pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
4366 		 pid, vaddr_start, vaddr_end);
4367 
4368 	mmap_read_lock(mm);
4369 	/*
4370 	 * always increase addr by PAGE_SIZE, since we could have a PTE page
4371 	 * table filled with PTE-mapped THPs, each of which is distinct.
4372 	 */
4373 	for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
4374 		struct vm_area_struct *vma = vma_lookup(mm, addr);
4375 		struct folio_walk fw;
4376 		struct folio *folio;
4377 		struct address_space *mapping;
4378 		unsigned int target_order = new_order;
4379 
4380 		if (!vma)
4381 			break;
4382 
4383 		/* skip special VMA and hugetlb VMA */
4384 		if (vma_not_suitable_for_thp_split(vma)) {
4385 			addr = vma->vm_end;
4386 			continue;
4387 		}
4388 
4389 		folio = folio_walk_start(&fw, vma, addr, 0);
4390 		if (!folio)
4391 			continue;
4392 
4393 		if (!is_transparent_hugepage(folio))
4394 			goto next;
4395 
4396 		if (!folio_test_anon(folio)) {
4397 			mapping = folio->mapping;
4398 			target_order = max(new_order,
4399 					   mapping_min_folio_order(mapping));
4400 		}
4401 
4402 		if (target_order >= folio_order(folio))
4403 			goto next;
4404 
4405 		total++;
4406 		/*
4407 		 * For folios with private, split_huge_page_to_list_to_order()
4408 		 * will try to drop it before split and then check if the folio
4409 		 * can be split or not. So skip the check here.
4410 		 */
4411 		if (!folio_test_private(folio) &&
4412 		    !can_split_folio(folio, 0, NULL))
4413 			goto next;
4414 
4415 		if (!folio_trylock(folio))
4416 			goto next;
4417 		folio_get(folio);
4418 		folio_walk_end(&fw, vma);
4419 
4420 		if (!folio_test_anon(folio) && folio->mapping != mapping)
4421 			goto unlock;
4422 
4423 		if (in_folio_offset < 0 ||
4424 		    in_folio_offset >= folio_nr_pages(folio)) {
4425 			if (!split_folio_to_order(folio, target_order))
4426 				split++;
4427 		} else {
4428 			struct page *split_at = folio_page(folio,
4429 							   in_folio_offset);
4430 			if (!folio_split(folio, target_order, split_at, NULL))
4431 				split++;
4432 		}
4433 
4434 unlock:
4435 
4436 		folio_unlock(folio);
4437 		folio_put(folio);
4438 
4439 		cond_resched();
4440 		continue;
4441 next:
4442 		folio_walk_end(&fw, vma);
4443 		cond_resched();
4444 	}
4445 	mmap_read_unlock(mm);
4446 	mmput(mm);
4447 
4448 	pr_debug("%lu of %lu THP split\n", split, total);
4449 
4450 out:
4451 	return ret;
4452 }
4453 
4454 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
4455 				pgoff_t off_end, unsigned int new_order,
4456 				long in_folio_offset)
4457 {
4458 	struct filename *file;
4459 	struct file *candidate;
4460 	struct address_space *mapping;
4461 	int ret = -EINVAL;
4462 	pgoff_t index;
4463 	int nr_pages = 1;
4464 	unsigned long total = 0, split = 0;
4465 	unsigned int min_order;
4466 	unsigned int target_order;
4467 
4468 	file = getname_kernel(file_path);
4469 	if (IS_ERR(file))
4470 		return ret;
4471 
4472 	candidate = file_open_name(file, O_RDONLY, 0);
4473 	if (IS_ERR(candidate))
4474 		goto out;
4475 
4476 	pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
4477 		 file_path, off_start, off_end);
4478 
4479 	mapping = candidate->f_mapping;
4480 	min_order = mapping_min_folio_order(mapping);
4481 	target_order = max(new_order, min_order);
4482 
4483 	for (index = off_start; index < off_end; index += nr_pages) {
4484 		struct folio *folio = filemap_get_folio(mapping, index);
4485 
4486 		nr_pages = 1;
4487 		if (IS_ERR(folio))
4488 			continue;
4489 
4490 		if (!folio_test_large(folio))
4491 			goto next;
4492 
4493 		total++;
4494 		nr_pages = folio_nr_pages(folio);
4495 
4496 		if (target_order >= folio_order(folio))
4497 			goto next;
4498 
4499 		if (!folio_trylock(folio))
4500 			goto next;
4501 
4502 		if (folio->mapping != mapping)
4503 			goto unlock;
4504 
4505 		if (in_folio_offset < 0 || in_folio_offset >= nr_pages) {
4506 			if (!split_folio_to_order(folio, target_order))
4507 				split++;
4508 		} else {
4509 			struct page *split_at = folio_page(folio,
4510 							   in_folio_offset);
4511 			if (!folio_split(folio, target_order, split_at, NULL))
4512 				split++;
4513 		}
4514 
4515 unlock:
4516 		folio_unlock(folio);
4517 next:
4518 		folio_put(folio);
4519 		cond_resched();
4520 	}
4521 
4522 	filp_close(candidate, NULL);
4523 	ret = 0;
4524 
4525 	pr_debug("%lu of %lu file-backed THP split\n", split, total);
4526 out:
4527 	putname(file);
4528 	return ret;
4529 }
4530 
4531 #define MAX_INPUT_BUF_SZ 255
4532 
4533 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
4534 				size_t count, loff_t *ppops)
4535 {
4536 	static DEFINE_MUTEX(split_debug_mutex);
4537 	ssize_t ret;
4538 	/*
4539 	 * hold pid, start_vaddr, end_vaddr, new_order or
4540 	 * file_path, off_start, off_end, new_order
4541 	 */
4542 	char input_buf[MAX_INPUT_BUF_SZ];
4543 	int pid;
4544 	unsigned long vaddr_start, vaddr_end;
4545 	unsigned int new_order = 0;
4546 	long in_folio_offset = -1;
4547 
4548 	ret = mutex_lock_interruptible(&split_debug_mutex);
4549 	if (ret)
4550 		return ret;
4551 
4552 	ret = -EFAULT;
4553 
4554 	memset(input_buf, 0, MAX_INPUT_BUF_SZ);
4555 	if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
4556 		goto out;
4557 
4558 	input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
4559 
4560 	if (input_buf[0] == '/') {
4561 		char *tok;
4562 		char *tok_buf = input_buf;
4563 		char file_path[MAX_INPUT_BUF_SZ];
4564 		pgoff_t off_start = 0, off_end = 0;
4565 		size_t input_len = strlen(input_buf);
4566 
4567 		tok = strsep(&tok_buf, ",");
4568 		if (tok && tok_buf) {
4569 			strscpy(file_path, tok);
4570 		} else {
4571 			ret = -EINVAL;
4572 			goto out;
4573 		}
4574 
4575 		ret = sscanf(tok_buf, "0x%lx,0x%lx,%d,%ld", &off_start, &off_end,
4576 				&new_order, &in_folio_offset);
4577 		if (ret != 2 && ret != 3 && ret != 4) {
4578 			ret = -EINVAL;
4579 			goto out;
4580 		}
4581 		ret = split_huge_pages_in_file(file_path, off_start, off_end,
4582 				new_order, in_folio_offset);
4583 		if (!ret)
4584 			ret = input_len;
4585 
4586 		goto out;
4587 	}
4588 
4589 	ret = sscanf(input_buf, "%d,0x%lx,0x%lx,%d,%ld", &pid, &vaddr_start,
4590 			&vaddr_end, &new_order, &in_folio_offset);
4591 	if (ret == 1 && pid == 1) {
4592 		split_huge_pages_all();
4593 		ret = strlen(input_buf);
4594 		goto out;
4595 	} else if (ret != 3 && ret != 4 && ret != 5) {
4596 		ret = -EINVAL;
4597 		goto out;
4598 	}
4599 
4600 	ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end, new_order,
4601 			in_folio_offset);
4602 	if (!ret)
4603 		ret = strlen(input_buf);
4604 out:
4605 	mutex_unlock(&split_debug_mutex);
4606 	return ret;
4607 
4608 }
4609 
4610 static const struct file_operations split_huge_pages_fops = {
4611 	.owner	 = THIS_MODULE,
4612 	.write	 = split_huge_pages_write,
4613 };
4614 
4615 static int __init split_huge_pages_debugfs(void)
4616 {
4617 	debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
4618 			    &split_huge_pages_fops);
4619 	return 0;
4620 }
4621 late_initcall(split_huge_pages_debugfs);
4622 #endif
4623 
4624 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
4625 int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
4626 		struct page *page)
4627 {
4628 	struct folio *folio = page_folio(page);
4629 	struct vm_area_struct *vma = pvmw->vma;
4630 	struct mm_struct *mm = vma->vm_mm;
4631 	unsigned long address = pvmw->address;
4632 	bool anon_exclusive;
4633 	pmd_t pmdval;
4634 	swp_entry_t entry;
4635 	pmd_t pmdswp;
4636 
4637 	if (!(pvmw->pmd && !pvmw->pte))
4638 		return 0;
4639 
4640 	flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
4641 	pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
4642 
4643 	/* See folio_try_share_anon_rmap_pmd(): invalidate PMD first. */
4644 	anon_exclusive = folio_test_anon(folio) && PageAnonExclusive(page);
4645 	if (anon_exclusive && folio_try_share_anon_rmap_pmd(folio, page)) {
4646 		set_pmd_at(mm, address, pvmw->pmd, pmdval);
4647 		return -EBUSY;
4648 	}
4649 
4650 	if (pmd_dirty(pmdval))
4651 		folio_mark_dirty(folio);
4652 	if (pmd_write(pmdval))
4653 		entry = make_writable_migration_entry(page_to_pfn(page));
4654 	else if (anon_exclusive)
4655 		entry = make_readable_exclusive_migration_entry(page_to_pfn(page));
4656 	else
4657 		entry = make_readable_migration_entry(page_to_pfn(page));
4658 	if (pmd_young(pmdval))
4659 		entry = make_migration_entry_young(entry);
4660 	if (pmd_dirty(pmdval))
4661 		entry = make_migration_entry_dirty(entry);
4662 	pmdswp = swp_entry_to_pmd(entry);
4663 	if (pmd_soft_dirty(pmdval))
4664 		pmdswp = pmd_swp_mksoft_dirty(pmdswp);
4665 	if (pmd_uffd_wp(pmdval))
4666 		pmdswp = pmd_swp_mkuffd_wp(pmdswp);
4667 	set_pmd_at(mm, address, pvmw->pmd, pmdswp);
4668 	folio_remove_rmap_pmd(folio, page, vma);
4669 	folio_put(folio);
4670 	trace_set_migration_pmd(address, pmd_val(pmdswp));
4671 
4672 	return 0;
4673 }
4674 
4675 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
4676 {
4677 	struct folio *folio = page_folio(new);
4678 	struct vm_area_struct *vma = pvmw->vma;
4679 	struct mm_struct *mm = vma->vm_mm;
4680 	unsigned long address = pvmw->address;
4681 	unsigned long haddr = address & HPAGE_PMD_MASK;
4682 	pmd_t pmde;
4683 	swp_entry_t entry;
4684 
4685 	if (!(pvmw->pmd && !pvmw->pte))
4686 		return;
4687 
4688 	entry = pmd_to_swp_entry(*pvmw->pmd);
4689 	folio_get(folio);
4690 	pmde = folio_mk_pmd(folio, READ_ONCE(vma->vm_page_prot));
4691 	if (pmd_swp_soft_dirty(*pvmw->pmd))
4692 		pmde = pmd_mksoft_dirty(pmde);
4693 	if (is_writable_migration_entry(entry))
4694 		pmde = pmd_mkwrite(pmde, vma);
4695 	if (pmd_swp_uffd_wp(*pvmw->pmd))
4696 		pmde = pmd_mkuffd_wp(pmde);
4697 	if (!is_migration_entry_young(entry))
4698 		pmde = pmd_mkold(pmde);
4699 	/* NOTE: this may contain setting soft-dirty on some archs */
4700 	if (folio_test_dirty(folio) && is_migration_entry_dirty(entry))
4701 		pmde = pmd_mkdirty(pmde);
4702 
4703 	if (folio_test_anon(folio)) {
4704 		rmap_t rmap_flags = RMAP_NONE;
4705 
4706 		if (!is_readable_migration_entry(entry))
4707 			rmap_flags |= RMAP_EXCLUSIVE;
4708 
4709 		folio_add_anon_rmap_pmd(folio, new, vma, haddr, rmap_flags);
4710 	} else {
4711 		folio_add_file_rmap_pmd(folio, new, vma);
4712 	}
4713 	VM_BUG_ON(pmd_write(pmde) && folio_test_anon(folio) && !PageAnonExclusive(new));
4714 	set_pmd_at(mm, haddr, pvmw->pmd, pmde);
4715 
4716 	/* No need to invalidate - it was non-present before */
4717 	update_mmu_cache_pmd(vma, address, pvmw->pmd);
4718 	trace_remove_migration_pmd(address, pmd_val(pmde));
4719 }
4720 #endif
4721