xref: /linux/mm/huge_memory.c (revision 2ec41967189cd65a8f79c760dd1b50c4f56e8ac6)
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/mman.h>
26 #include <linux/memremap.h>
27 #include <linux/pagemap.h>
28 #include <linux/debugfs.h>
29 #include <linux/migrate.h>
30 #include <linux/hashtable.h>
31 #include <linux/userfaultfd_k.h>
32 #include <linux/page_idle.h>
33 #include <linux/shmem_fs.h>
34 #include <linux/oom.h>
35 #include <linux/numa.h>
36 #include <linux/page_owner.h>
37 #include <linux/sched/sysctl.h>
38 #include <linux/memory-tiers.h>
39 #include <linux/compat.h>
40 #include <linux/pgalloc.h>
41 #include <linux/pgalloc_tag.h>
42 #include <linux/pagewalk.h>
43 
44 #include <asm/tlb.h>
45 #include "internal.h"
46 #include "swap.h"
47 
48 #define CREATE_TRACE_POINTS
49 #include <trace/events/thp.h>
50 
51 /*
52  * By default, transparent hugepage support is disabled in order to avoid
53  * risking an increased memory footprint for applications that are not
54  * guaranteed to benefit from it. When transparent hugepage support is
55  * enabled, it is for all mappings, and khugepaged scans all mappings.
56  * Defrag is invoked by khugepaged hugepage allocations and by page faults
57  * for all hugepage allocations.
58  */
59 unsigned long transparent_hugepage_flags __read_mostly =
60 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
61 	(1<<TRANSPARENT_HUGEPAGE_FLAG)|
62 #endif
63 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
64 	(1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
65 #endif
66 	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
67 	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
68 	(1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
69 
70 static struct shrinker *deferred_split_shrinker;
71 static unsigned long deferred_split_count(struct shrinker *shrink,
72 					  struct shrink_control *sc);
73 static unsigned long deferred_split_scan(struct shrinker *shrink,
74 					 struct shrink_control *sc);
75 static bool split_underused_thp = true;
76 
77 static atomic_t huge_zero_refcount;
78 struct folio *huge_zero_folio __read_mostly;
79 unsigned long huge_zero_pfn __read_mostly = ~0UL;
80 unsigned long huge_anon_orders_always __read_mostly;
81 unsigned long huge_anon_orders_madvise __read_mostly;
82 unsigned long huge_anon_orders_inherit __read_mostly;
83 static bool anon_orders_configured __initdata;
84 
85 static inline bool file_thp_enabled(struct vm_area_struct *vma)
86 {
87 	struct inode *inode;
88 
89 	if (!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS))
90 		return false;
91 
92 	if (!vma->vm_file)
93 		return false;
94 
95 	inode = file_inode(vma->vm_file);
96 
97 	return !inode_is_open_for_write(inode) && S_ISREG(inode->i_mode);
98 }
99 
100 unsigned long __thp_vma_allowable_orders(struct vm_area_struct *vma,
101 					 vm_flags_t vm_flags,
102 					 enum tva_type type,
103 					 unsigned long orders)
104 {
105 	const bool smaps = type == TVA_SMAPS;
106 	const bool in_pf = type == TVA_PAGEFAULT;
107 	const bool forced_collapse = type == TVA_FORCED_COLLAPSE;
108 	unsigned long supported_orders;
109 
110 	/* Check the intersection of requested and supported orders. */
111 	if (vma_is_anonymous(vma))
112 		supported_orders = THP_ORDERS_ALL_ANON;
113 	else if (vma_is_special_huge(vma))
114 		supported_orders = THP_ORDERS_ALL_SPECIAL;
115 	else
116 		supported_orders = THP_ORDERS_ALL_FILE_DEFAULT;
117 
118 	orders &= supported_orders;
119 	if (!orders)
120 		return 0;
121 
122 	if (!vma->vm_mm)		/* vdso */
123 		return 0;
124 
125 	if (thp_disabled_by_hw() || vma_thp_disabled(vma, vm_flags, forced_collapse))
126 		return 0;
127 
128 	/* khugepaged doesn't collapse DAX vma, but page fault is fine. */
129 	if (vma_is_dax(vma))
130 		return in_pf ? orders : 0;
131 
132 	/*
133 	 * khugepaged special VMA and hugetlb VMA.
134 	 * Must be checked after dax since some dax mappings may have
135 	 * VM_MIXEDMAP set.
136 	 */
137 	if (!in_pf && !smaps && (vm_flags & VM_NO_KHUGEPAGED))
138 		return 0;
139 
140 	/*
141 	 * Check alignment for file vma and size for both file and anon vma by
142 	 * filtering out the unsuitable orders.
143 	 *
144 	 * Skip the check for page fault. Huge fault does the check in fault
145 	 * handlers.
146 	 */
147 	if (!in_pf) {
148 		int order = highest_order(orders);
149 		unsigned long addr;
150 
151 		while (orders) {
152 			addr = vma->vm_end - (PAGE_SIZE << order);
153 			if (thp_vma_suitable_order(vma, addr, order))
154 				break;
155 			order = next_order(&orders, order);
156 		}
157 
158 		if (!orders)
159 			return 0;
160 	}
161 
162 	/*
163 	 * Enabled via shmem mount options or sysfs settings.
164 	 * Must be done before hugepage flags check since shmem has its
165 	 * own flags.
166 	 */
167 	if (!in_pf && shmem_file(vma->vm_file))
168 		return orders & shmem_allowable_huge_orders(file_inode(vma->vm_file),
169 						   vma, vma->vm_pgoff, 0,
170 						   forced_collapse);
171 
172 	if (!vma_is_anonymous(vma)) {
173 		/*
174 		 * Enforce THP collapse requirements as necessary. Anonymous vmas
175 		 * were already handled in thp_vma_allowable_orders().
176 		 */
177 		if (!forced_collapse &&
178 		    (!hugepage_global_enabled() || (!(vm_flags & VM_HUGEPAGE) &&
179 						    !hugepage_global_always())))
180 			return 0;
181 
182 		/*
183 		 * Trust that ->huge_fault() handlers know what they are doing
184 		 * in fault path.
185 		 */
186 		if (((in_pf || smaps)) && vma->vm_ops->huge_fault)
187 			return orders;
188 		/* Only regular file is valid in collapse path */
189 		if (((!in_pf || smaps)) && file_thp_enabled(vma))
190 			return orders;
191 		return 0;
192 	}
193 
194 	if (vma_is_temporary_stack(vma))
195 		return 0;
196 
197 	/*
198 	 * THPeligible bit of smaps should show 1 for proper VMAs even
199 	 * though anon_vma is not initialized yet.
200 	 *
201 	 * Allow page fault since anon_vma may be not initialized until
202 	 * the first page fault.
203 	 */
204 	if (!vma->anon_vma)
205 		return (smaps || in_pf) ? orders : 0;
206 
207 	return orders;
208 }
209 
210 static bool get_huge_zero_folio(void)
211 {
212 	struct folio *zero_folio;
213 retry:
214 	if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
215 		return true;
216 
217 	zero_folio = folio_alloc((GFP_TRANSHUGE | __GFP_ZERO | __GFP_ZEROTAGS) &
218 				 ~__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_folio(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 (IS_ENABLED(CONFIG_PERSISTENT_HUGE_ZERO_FOLIO))
253 		return huge_zero_folio;
254 
255 	if (mm_flags_test(MMF_HUGE_ZERO_FOLIO, mm))
256 		return READ_ONCE(huge_zero_folio);
257 
258 	if (!get_huge_zero_folio())
259 		return NULL;
260 
261 	if (mm_flags_test_and_set(MMF_HUGE_ZERO_FOLIO, mm))
262 		put_huge_zero_folio();
263 
264 	return READ_ONCE(huge_zero_folio);
265 }
266 
267 void mm_put_huge_zero_folio(struct mm_struct *mm)
268 {
269 	if (IS_ENABLED(CONFIG_PERSISTENT_HUGE_ZERO_FOLIO))
270 		return;
271 
272 	if (mm_flags_test(MMF_HUGE_ZERO_FOLIO, mm))
273 		put_huge_zero_folio();
274 }
275 
276 static unsigned long shrink_huge_zero_folio_count(struct shrinker *shrink,
277 						  struct shrink_control *sc)
278 {
279 	/* we can free zero page only if last reference remains */
280 	return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
281 }
282 
283 static unsigned long shrink_huge_zero_folio_scan(struct shrinker *shrink,
284 						 struct shrink_control *sc)
285 {
286 	if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
287 		struct folio *zero_folio = xchg(&huge_zero_folio, NULL);
288 		BUG_ON(zero_folio == NULL);
289 		WRITE_ONCE(huge_zero_pfn, ~0UL);
290 		folio_put(zero_folio);
291 		return HPAGE_PMD_NR;
292 	}
293 
294 	return 0;
295 }
296 
297 static struct shrinker *huge_zero_folio_shrinker;
298 
299 #ifdef CONFIG_SYSFS
300 static ssize_t enabled_show(struct kobject *kobj,
301 			    struct kobj_attribute *attr, char *buf)
302 {
303 	const char *output;
304 
305 	if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
306 		output = "[always] madvise never";
307 	else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
308 			  &transparent_hugepage_flags))
309 		output = "always [madvise] never";
310 	else
311 		output = "always madvise [never]";
312 
313 	return sysfs_emit(buf, "%s\n", output);
314 }
315 
316 static ssize_t enabled_store(struct kobject *kobj,
317 			     struct kobj_attribute *attr,
318 			     const char *buf, size_t count)
319 {
320 	ssize_t ret = count;
321 
322 	if (sysfs_streq(buf, "always")) {
323 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
324 		set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
325 	} else if (sysfs_streq(buf, "madvise")) {
326 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
327 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
328 	} else if (sysfs_streq(buf, "never")) {
329 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
330 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
331 	} else
332 		ret = -EINVAL;
333 
334 	if (ret > 0) {
335 		int err = start_stop_khugepaged();
336 		if (err)
337 			ret = err;
338 	}
339 	return ret;
340 }
341 
342 static struct kobj_attribute enabled_attr = __ATTR_RW(enabled);
343 
344 ssize_t single_hugepage_flag_show(struct kobject *kobj,
345 				  struct kobj_attribute *attr, char *buf,
346 				  enum transparent_hugepage_flag flag)
347 {
348 	return sysfs_emit(buf, "%d\n",
349 			  !!test_bit(flag, &transparent_hugepage_flags));
350 }
351 
352 ssize_t single_hugepage_flag_store(struct kobject *kobj,
353 				 struct kobj_attribute *attr,
354 				 const char *buf, size_t count,
355 				 enum transparent_hugepage_flag flag)
356 {
357 	unsigned long value;
358 	int ret;
359 
360 	ret = kstrtoul(buf, 10, &value);
361 	if (ret < 0)
362 		return ret;
363 	if (value > 1)
364 		return -EINVAL;
365 
366 	if (value)
367 		set_bit(flag, &transparent_hugepage_flags);
368 	else
369 		clear_bit(flag, &transparent_hugepage_flags);
370 
371 	return count;
372 }
373 
374 static ssize_t defrag_show(struct kobject *kobj,
375 			   struct kobj_attribute *attr, char *buf)
376 {
377 	const char *output;
378 
379 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
380 		     &transparent_hugepage_flags))
381 		output = "[always] defer defer+madvise madvise never";
382 	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
383 			  &transparent_hugepage_flags))
384 		output = "always [defer] defer+madvise madvise never";
385 	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
386 			  &transparent_hugepage_flags))
387 		output = "always defer [defer+madvise] madvise never";
388 	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
389 			  &transparent_hugepage_flags))
390 		output = "always defer defer+madvise [madvise] never";
391 	else
392 		output = "always defer defer+madvise madvise [never]";
393 
394 	return sysfs_emit(buf, "%s\n", output);
395 }
396 
397 static ssize_t defrag_store(struct kobject *kobj,
398 			    struct kobj_attribute *attr,
399 			    const char *buf, size_t count)
400 {
401 	if (sysfs_streq(buf, "always")) {
402 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
403 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
404 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
405 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
406 	} else if (sysfs_streq(buf, "defer+madvise")) {
407 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
408 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
409 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
410 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
411 	} else if (sysfs_streq(buf, "defer")) {
412 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
413 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
414 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
415 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
416 	} else if (sysfs_streq(buf, "madvise")) {
417 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
418 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
419 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
420 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
421 	} else if (sysfs_streq(buf, "never")) {
422 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
423 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
424 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
425 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
426 	} else
427 		return -EINVAL;
428 
429 	return count;
430 }
431 static struct kobj_attribute defrag_attr = __ATTR_RW(defrag);
432 
433 static ssize_t use_zero_page_show(struct kobject *kobj,
434 				  struct kobj_attribute *attr, char *buf)
435 {
436 	return single_hugepage_flag_show(kobj, attr, buf,
437 					 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
438 }
439 static ssize_t use_zero_page_store(struct kobject *kobj,
440 		struct kobj_attribute *attr, const char *buf, size_t count)
441 {
442 	return single_hugepage_flag_store(kobj, attr, buf, count,
443 				 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
444 }
445 static struct kobj_attribute use_zero_page_attr = __ATTR_RW(use_zero_page);
446 
447 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
448 				   struct kobj_attribute *attr, char *buf)
449 {
450 	return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
451 }
452 static struct kobj_attribute hpage_pmd_size_attr =
453 	__ATTR_RO(hpage_pmd_size);
454 
455 static ssize_t split_underused_thp_show(struct kobject *kobj,
456 			    struct kobj_attribute *attr, char *buf)
457 {
458 	return sysfs_emit(buf, "%d\n", split_underused_thp);
459 }
460 
461 static ssize_t split_underused_thp_store(struct kobject *kobj,
462 			     struct kobj_attribute *attr,
463 			     const char *buf, size_t count)
464 {
465 	int err = kstrtobool(buf, &split_underused_thp);
466 
467 	if (err < 0)
468 		return err;
469 
470 	return count;
471 }
472 
473 static struct kobj_attribute split_underused_thp_attr = __ATTR(
474 	shrink_underused, 0644, split_underused_thp_show, split_underused_thp_store);
475 
476 static struct attribute *hugepage_attr[] = {
477 	&enabled_attr.attr,
478 	&defrag_attr.attr,
479 	&use_zero_page_attr.attr,
480 	&hpage_pmd_size_attr.attr,
481 #ifdef CONFIG_SHMEM
482 	&shmem_enabled_attr.attr,
483 #endif
484 	&split_underused_thp_attr.attr,
485 	NULL,
486 };
487 
488 static const struct attribute_group hugepage_attr_group = {
489 	.attrs = hugepage_attr,
490 };
491 
492 static void hugepage_exit_sysfs(struct kobject *hugepage_kobj);
493 static void thpsize_release(struct kobject *kobj);
494 static DEFINE_SPINLOCK(huge_anon_orders_lock);
495 static LIST_HEAD(thpsize_list);
496 
497 static ssize_t anon_enabled_show(struct kobject *kobj,
498 				 struct kobj_attribute *attr, char *buf)
499 {
500 	int order = to_thpsize(kobj)->order;
501 	const char *output;
502 
503 	if (test_bit(order, &huge_anon_orders_always))
504 		output = "[always] inherit madvise never";
505 	else if (test_bit(order, &huge_anon_orders_inherit))
506 		output = "always [inherit] madvise never";
507 	else if (test_bit(order, &huge_anon_orders_madvise))
508 		output = "always inherit [madvise] never";
509 	else
510 		output = "always inherit madvise [never]";
511 
512 	return sysfs_emit(buf, "%s\n", output);
513 }
514 
515 static ssize_t anon_enabled_store(struct kobject *kobj,
516 				  struct kobj_attribute *attr,
517 				  const char *buf, size_t count)
518 {
519 	int order = to_thpsize(kobj)->order;
520 	ssize_t ret = count;
521 
522 	if (sysfs_streq(buf, "always")) {
523 		spin_lock(&huge_anon_orders_lock);
524 		clear_bit(order, &huge_anon_orders_inherit);
525 		clear_bit(order, &huge_anon_orders_madvise);
526 		set_bit(order, &huge_anon_orders_always);
527 		spin_unlock(&huge_anon_orders_lock);
528 	} else if (sysfs_streq(buf, "inherit")) {
529 		spin_lock(&huge_anon_orders_lock);
530 		clear_bit(order, &huge_anon_orders_always);
531 		clear_bit(order, &huge_anon_orders_madvise);
532 		set_bit(order, &huge_anon_orders_inherit);
533 		spin_unlock(&huge_anon_orders_lock);
534 	} else if (sysfs_streq(buf, "madvise")) {
535 		spin_lock(&huge_anon_orders_lock);
536 		clear_bit(order, &huge_anon_orders_always);
537 		clear_bit(order, &huge_anon_orders_inherit);
538 		set_bit(order, &huge_anon_orders_madvise);
539 		spin_unlock(&huge_anon_orders_lock);
540 	} else if (sysfs_streq(buf, "never")) {
541 		spin_lock(&huge_anon_orders_lock);
542 		clear_bit(order, &huge_anon_orders_always);
543 		clear_bit(order, &huge_anon_orders_inherit);
544 		clear_bit(order, &huge_anon_orders_madvise);
545 		spin_unlock(&huge_anon_orders_lock);
546 	} else
547 		ret = -EINVAL;
548 
549 	if (ret > 0) {
550 		int err;
551 
552 		err = start_stop_khugepaged();
553 		if (err)
554 			ret = err;
555 	}
556 	return ret;
557 }
558 
559 static struct kobj_attribute anon_enabled_attr =
560 	__ATTR(enabled, 0644, anon_enabled_show, anon_enabled_store);
561 
562 static struct attribute *anon_ctrl_attrs[] = {
563 	&anon_enabled_attr.attr,
564 	NULL,
565 };
566 
567 static const struct attribute_group anon_ctrl_attr_grp = {
568 	.attrs = anon_ctrl_attrs,
569 };
570 
571 static struct attribute *file_ctrl_attrs[] = {
572 #ifdef CONFIG_SHMEM
573 	&thpsize_shmem_enabled_attr.attr,
574 #endif
575 	NULL,
576 };
577 
578 static const struct attribute_group file_ctrl_attr_grp = {
579 	.attrs = file_ctrl_attrs,
580 };
581 
582 static struct attribute *any_ctrl_attrs[] = {
583 	NULL,
584 };
585 
586 static const struct attribute_group any_ctrl_attr_grp = {
587 	.attrs = any_ctrl_attrs,
588 };
589 
590 static const struct kobj_type thpsize_ktype = {
591 	.release = &thpsize_release,
592 	.sysfs_ops = &kobj_sysfs_ops,
593 };
594 
595 DEFINE_PER_CPU(struct mthp_stat, mthp_stats) = {{{0}}};
596 
597 static unsigned long sum_mthp_stat(int order, enum mthp_stat_item item)
598 {
599 	unsigned long sum = 0;
600 	int cpu;
601 
602 	for_each_possible_cpu(cpu) {
603 		struct mthp_stat *this = &per_cpu(mthp_stats, cpu);
604 
605 		sum += this->stats[order][item];
606 	}
607 
608 	return sum;
609 }
610 
611 #define DEFINE_MTHP_STAT_ATTR(_name, _index)				\
612 static ssize_t _name##_show(struct kobject *kobj,			\
613 			struct kobj_attribute *attr, char *buf)		\
614 {									\
615 	int order = to_thpsize(kobj)->order;				\
616 									\
617 	return sysfs_emit(buf, "%lu\n", sum_mthp_stat(order, _index));	\
618 }									\
619 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
620 
621 DEFINE_MTHP_STAT_ATTR(anon_fault_alloc, MTHP_STAT_ANON_FAULT_ALLOC);
622 DEFINE_MTHP_STAT_ATTR(anon_fault_fallback, MTHP_STAT_ANON_FAULT_FALLBACK);
623 DEFINE_MTHP_STAT_ATTR(anon_fault_fallback_charge, MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE);
624 DEFINE_MTHP_STAT_ATTR(zswpout, MTHP_STAT_ZSWPOUT);
625 DEFINE_MTHP_STAT_ATTR(swpin, MTHP_STAT_SWPIN);
626 DEFINE_MTHP_STAT_ATTR(swpin_fallback, MTHP_STAT_SWPIN_FALLBACK);
627 DEFINE_MTHP_STAT_ATTR(swpin_fallback_charge, MTHP_STAT_SWPIN_FALLBACK_CHARGE);
628 DEFINE_MTHP_STAT_ATTR(swpout, MTHP_STAT_SWPOUT);
629 DEFINE_MTHP_STAT_ATTR(swpout_fallback, MTHP_STAT_SWPOUT_FALLBACK);
630 #ifdef CONFIG_SHMEM
631 DEFINE_MTHP_STAT_ATTR(shmem_alloc, MTHP_STAT_SHMEM_ALLOC);
632 DEFINE_MTHP_STAT_ATTR(shmem_fallback, MTHP_STAT_SHMEM_FALLBACK);
633 DEFINE_MTHP_STAT_ATTR(shmem_fallback_charge, MTHP_STAT_SHMEM_FALLBACK_CHARGE);
634 #endif
635 DEFINE_MTHP_STAT_ATTR(split, MTHP_STAT_SPLIT);
636 DEFINE_MTHP_STAT_ATTR(split_failed, MTHP_STAT_SPLIT_FAILED);
637 DEFINE_MTHP_STAT_ATTR(split_deferred, MTHP_STAT_SPLIT_DEFERRED);
638 DEFINE_MTHP_STAT_ATTR(nr_anon, MTHP_STAT_NR_ANON);
639 DEFINE_MTHP_STAT_ATTR(nr_anon_partially_mapped, MTHP_STAT_NR_ANON_PARTIALLY_MAPPED);
640 
641 static struct attribute *anon_stats_attrs[] = {
642 	&anon_fault_alloc_attr.attr,
643 	&anon_fault_fallback_attr.attr,
644 	&anon_fault_fallback_charge_attr.attr,
645 #ifndef CONFIG_SHMEM
646 	&zswpout_attr.attr,
647 	&swpin_attr.attr,
648 	&swpin_fallback_attr.attr,
649 	&swpin_fallback_charge_attr.attr,
650 	&swpout_attr.attr,
651 	&swpout_fallback_attr.attr,
652 #endif
653 	&split_deferred_attr.attr,
654 	&nr_anon_attr.attr,
655 	&nr_anon_partially_mapped_attr.attr,
656 	NULL,
657 };
658 
659 static struct attribute_group anon_stats_attr_grp = {
660 	.name = "stats",
661 	.attrs = anon_stats_attrs,
662 };
663 
664 static struct attribute *file_stats_attrs[] = {
665 #ifdef CONFIG_SHMEM
666 	&shmem_alloc_attr.attr,
667 	&shmem_fallback_attr.attr,
668 	&shmem_fallback_charge_attr.attr,
669 #endif
670 	NULL,
671 };
672 
673 static struct attribute_group file_stats_attr_grp = {
674 	.name = "stats",
675 	.attrs = file_stats_attrs,
676 };
677 
678 static struct attribute *any_stats_attrs[] = {
679 #ifdef CONFIG_SHMEM
680 	&zswpout_attr.attr,
681 	&swpin_attr.attr,
682 	&swpin_fallback_attr.attr,
683 	&swpin_fallback_charge_attr.attr,
684 	&swpout_attr.attr,
685 	&swpout_fallback_attr.attr,
686 #endif
687 	&split_attr.attr,
688 	&split_failed_attr.attr,
689 	NULL,
690 };
691 
692 static struct attribute_group any_stats_attr_grp = {
693 	.name = "stats",
694 	.attrs = any_stats_attrs,
695 };
696 
697 static int sysfs_add_group(struct kobject *kobj,
698 			   const struct attribute_group *grp)
699 {
700 	int ret = -ENOENT;
701 
702 	/*
703 	 * If the group is named, try to merge first, assuming the subdirectory
704 	 * was already created. This avoids the warning emitted by
705 	 * sysfs_create_group() if the directory already exists.
706 	 */
707 	if (grp->name)
708 		ret = sysfs_merge_group(kobj, grp);
709 	if (ret)
710 		ret = sysfs_create_group(kobj, grp);
711 
712 	return ret;
713 }
714 
715 static struct thpsize *thpsize_create(int order, struct kobject *parent)
716 {
717 	unsigned long size = (PAGE_SIZE << order) / SZ_1K;
718 	struct thpsize *thpsize;
719 	int ret = -ENOMEM;
720 
721 	thpsize = kzalloc(sizeof(*thpsize), GFP_KERNEL);
722 	if (!thpsize)
723 		goto err;
724 
725 	thpsize->order = order;
726 
727 	ret = kobject_init_and_add(&thpsize->kobj, &thpsize_ktype, parent,
728 				   "hugepages-%lukB", size);
729 	if (ret) {
730 		kfree(thpsize);
731 		goto err;
732 	}
733 
734 
735 	ret = sysfs_add_group(&thpsize->kobj, &any_ctrl_attr_grp);
736 	if (ret)
737 		goto err_put;
738 
739 	ret = sysfs_add_group(&thpsize->kobj, &any_stats_attr_grp);
740 	if (ret)
741 		goto err_put;
742 
743 	if (BIT(order) & THP_ORDERS_ALL_ANON) {
744 		ret = sysfs_add_group(&thpsize->kobj, &anon_ctrl_attr_grp);
745 		if (ret)
746 			goto err_put;
747 
748 		ret = sysfs_add_group(&thpsize->kobj, &anon_stats_attr_grp);
749 		if (ret)
750 			goto err_put;
751 	}
752 
753 	if (BIT(order) & THP_ORDERS_ALL_FILE_DEFAULT) {
754 		ret = sysfs_add_group(&thpsize->kobj, &file_ctrl_attr_grp);
755 		if (ret)
756 			goto err_put;
757 
758 		ret = sysfs_add_group(&thpsize->kobj, &file_stats_attr_grp);
759 		if (ret)
760 			goto err_put;
761 	}
762 
763 	return thpsize;
764 err_put:
765 	kobject_put(&thpsize->kobj);
766 err:
767 	return ERR_PTR(ret);
768 }
769 
770 static void thpsize_release(struct kobject *kobj)
771 {
772 	kfree(to_thpsize(kobj));
773 }
774 
775 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
776 {
777 	int err;
778 	struct thpsize *thpsize;
779 	unsigned long orders;
780 	int order;
781 
782 	/*
783 	 * Default to setting PMD-sized THP to inherit the global setting and
784 	 * disable all other sizes. powerpc's PMD_ORDER isn't a compile-time
785 	 * constant so we have to do this here.
786 	 */
787 	if (!anon_orders_configured)
788 		huge_anon_orders_inherit = BIT(PMD_ORDER);
789 
790 	*hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
791 	if (unlikely(!*hugepage_kobj)) {
792 		pr_err("failed to create transparent hugepage kobject\n");
793 		return -ENOMEM;
794 	}
795 
796 	err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
797 	if (err) {
798 		pr_err("failed to register transparent hugepage group\n");
799 		goto delete_obj;
800 	}
801 
802 	err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
803 	if (err) {
804 		pr_err("failed to register transparent hugepage group\n");
805 		goto remove_hp_group;
806 	}
807 
808 	orders = THP_ORDERS_ALL_ANON | THP_ORDERS_ALL_FILE_DEFAULT;
809 	order = highest_order(orders);
810 	while (orders) {
811 		thpsize = thpsize_create(order, *hugepage_kobj);
812 		if (IS_ERR(thpsize)) {
813 			pr_err("failed to create thpsize for order %d\n", order);
814 			err = PTR_ERR(thpsize);
815 			goto remove_all;
816 		}
817 		list_add(&thpsize->node, &thpsize_list);
818 		order = next_order(&orders, order);
819 	}
820 
821 	return 0;
822 
823 remove_all:
824 	hugepage_exit_sysfs(*hugepage_kobj);
825 	return err;
826 remove_hp_group:
827 	sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
828 delete_obj:
829 	kobject_put(*hugepage_kobj);
830 	return err;
831 }
832 
833 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
834 {
835 	struct thpsize *thpsize, *tmp;
836 
837 	list_for_each_entry_safe(thpsize, tmp, &thpsize_list, node) {
838 		list_del(&thpsize->node);
839 		kobject_put(&thpsize->kobj);
840 	}
841 
842 	sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
843 	sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
844 	kobject_put(hugepage_kobj);
845 }
846 #else
847 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
848 {
849 	return 0;
850 }
851 
852 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
853 {
854 }
855 #endif /* CONFIG_SYSFS */
856 
857 static int __init thp_shrinker_init(void)
858 {
859 	deferred_split_shrinker = shrinker_alloc(SHRINKER_NUMA_AWARE |
860 						 SHRINKER_MEMCG_AWARE |
861 						 SHRINKER_NONSLAB,
862 						 "thp-deferred_split");
863 	if (!deferred_split_shrinker)
864 		return -ENOMEM;
865 
866 	deferred_split_shrinker->count_objects = deferred_split_count;
867 	deferred_split_shrinker->scan_objects = deferred_split_scan;
868 	shrinker_register(deferred_split_shrinker);
869 
870 	if (IS_ENABLED(CONFIG_PERSISTENT_HUGE_ZERO_FOLIO)) {
871 		/*
872 		 * Bump the reference of the huge_zero_folio and do not
873 		 * initialize the shrinker.
874 		 *
875 		 * huge_zero_folio will always be NULL on failure. We assume
876 		 * that get_huge_zero_folio() will most likely not fail as
877 		 * thp_shrinker_init() is invoked early on during boot.
878 		 */
879 		if (!get_huge_zero_folio())
880 			pr_warn("Allocating persistent huge zero folio failed\n");
881 		return 0;
882 	}
883 
884 	huge_zero_folio_shrinker = shrinker_alloc(0, "thp-zero");
885 	if (!huge_zero_folio_shrinker) {
886 		shrinker_free(deferred_split_shrinker);
887 		return -ENOMEM;
888 	}
889 
890 	huge_zero_folio_shrinker->count_objects = shrink_huge_zero_folio_count;
891 	huge_zero_folio_shrinker->scan_objects = shrink_huge_zero_folio_scan;
892 	shrinker_register(huge_zero_folio_shrinker);
893 
894 	return 0;
895 }
896 
897 static void __init thp_shrinker_exit(void)
898 {
899 	shrinker_free(huge_zero_folio_shrinker);
900 	shrinker_free(deferred_split_shrinker);
901 }
902 
903 static int __init hugepage_init(void)
904 {
905 	int err;
906 	struct kobject *hugepage_kobj;
907 
908 	if (!has_transparent_hugepage()) {
909 		transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_UNSUPPORTED;
910 		return -EINVAL;
911 	}
912 
913 	/*
914 	 * hugepages can't be allocated by the buddy allocator
915 	 */
916 	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER > MAX_PAGE_ORDER);
917 
918 	err = hugepage_init_sysfs(&hugepage_kobj);
919 	if (err)
920 		goto err_sysfs;
921 
922 	err = khugepaged_init();
923 	if (err)
924 		goto err_slab;
925 
926 	err = thp_shrinker_init();
927 	if (err)
928 		goto err_shrinker;
929 
930 	/*
931 	 * By default disable transparent hugepages on smaller systems,
932 	 * where the extra memory used could hurt more than TLB overhead
933 	 * is likely to save.  The admin can still enable it through /sys.
934 	 */
935 	if (totalram_pages() < MB_TO_PAGES(512)) {
936 		transparent_hugepage_flags = 0;
937 		return 0;
938 	}
939 
940 	err = start_stop_khugepaged();
941 	if (err)
942 		goto err_khugepaged;
943 
944 	return 0;
945 err_khugepaged:
946 	thp_shrinker_exit();
947 err_shrinker:
948 	khugepaged_destroy();
949 err_slab:
950 	hugepage_exit_sysfs(hugepage_kobj);
951 err_sysfs:
952 	return err;
953 }
954 subsys_initcall(hugepage_init);
955 
956 static int __init setup_transparent_hugepage(char *str)
957 {
958 	int ret = 0;
959 	if (!str)
960 		goto out;
961 	if (!strcmp(str, "always")) {
962 		set_bit(TRANSPARENT_HUGEPAGE_FLAG,
963 			&transparent_hugepage_flags);
964 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
965 			  &transparent_hugepage_flags);
966 		ret = 1;
967 	} else if (!strcmp(str, "madvise")) {
968 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
969 			  &transparent_hugepage_flags);
970 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
971 			&transparent_hugepage_flags);
972 		ret = 1;
973 	} else if (!strcmp(str, "never")) {
974 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
975 			  &transparent_hugepage_flags);
976 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
977 			  &transparent_hugepage_flags);
978 		ret = 1;
979 	}
980 out:
981 	if (!ret)
982 		pr_warn("transparent_hugepage= cannot parse, ignored\n");
983 	return ret;
984 }
985 __setup("transparent_hugepage=", setup_transparent_hugepage);
986 
987 static char str_dup[PAGE_SIZE] __initdata;
988 static int __init setup_thp_anon(char *str)
989 {
990 	char *token, *range, *policy, *subtoken;
991 	unsigned long always, inherit, madvise;
992 	char *start_size, *end_size;
993 	int start, end, nr;
994 	char *p;
995 
996 	if (!str || strlen(str) + 1 > PAGE_SIZE)
997 		goto err;
998 	strscpy(str_dup, str);
999 
1000 	always = huge_anon_orders_always;
1001 	madvise = huge_anon_orders_madvise;
1002 	inherit = huge_anon_orders_inherit;
1003 	p = str_dup;
1004 	while ((token = strsep(&p, ";")) != NULL) {
1005 		range = strsep(&token, ":");
1006 		policy = token;
1007 
1008 		if (!policy)
1009 			goto err;
1010 
1011 		while ((subtoken = strsep(&range, ",")) != NULL) {
1012 			if (strchr(subtoken, '-')) {
1013 				start_size = strsep(&subtoken, "-");
1014 				end_size = subtoken;
1015 
1016 				start = get_order_from_str(start_size, THP_ORDERS_ALL_ANON);
1017 				end = get_order_from_str(end_size, THP_ORDERS_ALL_ANON);
1018 			} else {
1019 				start_size = end_size = subtoken;
1020 				start = end = get_order_from_str(subtoken,
1021 								 THP_ORDERS_ALL_ANON);
1022 			}
1023 
1024 			if (start == -EINVAL) {
1025 				pr_err("invalid size %s in thp_anon boot parameter\n", start_size);
1026 				goto err;
1027 			}
1028 
1029 			if (end == -EINVAL) {
1030 				pr_err("invalid size %s in thp_anon boot parameter\n", end_size);
1031 				goto err;
1032 			}
1033 
1034 			if (start < 0 || end < 0 || start > end)
1035 				goto err;
1036 
1037 			nr = end - start + 1;
1038 			if (!strcmp(policy, "always")) {
1039 				bitmap_set(&always, start, nr);
1040 				bitmap_clear(&inherit, start, nr);
1041 				bitmap_clear(&madvise, start, nr);
1042 			} else if (!strcmp(policy, "madvise")) {
1043 				bitmap_set(&madvise, start, nr);
1044 				bitmap_clear(&inherit, start, nr);
1045 				bitmap_clear(&always, start, nr);
1046 			} else if (!strcmp(policy, "inherit")) {
1047 				bitmap_set(&inherit, start, nr);
1048 				bitmap_clear(&madvise, start, nr);
1049 				bitmap_clear(&always, start, nr);
1050 			} else if (!strcmp(policy, "never")) {
1051 				bitmap_clear(&inherit, start, nr);
1052 				bitmap_clear(&madvise, start, nr);
1053 				bitmap_clear(&always, start, nr);
1054 			} else {
1055 				pr_err("invalid policy %s in thp_anon boot parameter\n", policy);
1056 				goto err;
1057 			}
1058 		}
1059 	}
1060 
1061 	huge_anon_orders_always = always;
1062 	huge_anon_orders_madvise = madvise;
1063 	huge_anon_orders_inherit = inherit;
1064 	anon_orders_configured = true;
1065 	return 1;
1066 
1067 err:
1068 	pr_warn("thp_anon=%s: error parsing string, ignoring setting\n", str);
1069 	return 0;
1070 }
1071 __setup("thp_anon=", setup_thp_anon);
1072 
1073 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
1074 {
1075 	if (likely(vma->vm_flags & VM_WRITE))
1076 		pmd = pmd_mkwrite(pmd, vma);
1077 	return pmd;
1078 }
1079 
1080 #ifdef CONFIG_MEMCG
1081 static inline
1082 struct deferred_split *get_deferred_split_queue(struct folio *folio)
1083 {
1084 	struct mem_cgroup *memcg = folio_memcg(folio);
1085 	struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));
1086 
1087 	if (memcg)
1088 		return &memcg->deferred_split_queue;
1089 	else
1090 		return &pgdat->deferred_split_queue;
1091 }
1092 #else
1093 static inline
1094 struct deferred_split *get_deferred_split_queue(struct folio *folio)
1095 {
1096 	struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));
1097 
1098 	return &pgdat->deferred_split_queue;
1099 }
1100 #endif
1101 
1102 static inline bool is_transparent_hugepage(const struct folio *folio)
1103 {
1104 	if (!folio_test_large(folio))
1105 		return false;
1106 
1107 	return is_huge_zero_folio(folio) ||
1108 		folio_test_large_rmappable(folio);
1109 }
1110 
1111 static unsigned long __thp_get_unmapped_area(struct file *filp,
1112 		unsigned long addr, unsigned long len,
1113 		loff_t off, unsigned long flags, unsigned long size,
1114 		vm_flags_t vm_flags)
1115 {
1116 	loff_t off_end = off + len;
1117 	loff_t off_align = round_up(off, size);
1118 	unsigned long len_pad, ret, off_sub;
1119 
1120 	if (!IS_ENABLED(CONFIG_64BIT) || in_compat_syscall())
1121 		return 0;
1122 
1123 	if (off_end <= off_align || (off_end - off_align) < size)
1124 		return 0;
1125 
1126 	len_pad = len + size;
1127 	if (len_pad < len || (off + len_pad) < off)
1128 		return 0;
1129 
1130 	ret = mm_get_unmapped_area_vmflags(filp, addr, len_pad,
1131 					   off >> PAGE_SHIFT, flags, vm_flags);
1132 
1133 	/*
1134 	 * The failure might be due to length padding. The caller will retry
1135 	 * without the padding.
1136 	 */
1137 	if (IS_ERR_VALUE(ret))
1138 		return 0;
1139 
1140 	/*
1141 	 * Do not try to align to THP boundary if allocation at the address
1142 	 * hint succeeds.
1143 	 */
1144 	if (ret == addr)
1145 		return addr;
1146 
1147 	off_sub = (off - ret) & (size - 1);
1148 
1149 	if (mm_flags_test(MMF_TOPDOWN, current->mm) && !off_sub)
1150 		return ret + size;
1151 
1152 	ret += off_sub;
1153 	return ret;
1154 }
1155 
1156 unsigned long thp_get_unmapped_area_vmflags(struct file *filp, unsigned long addr,
1157 		unsigned long len, unsigned long pgoff, unsigned long flags,
1158 		vm_flags_t vm_flags)
1159 {
1160 	unsigned long ret;
1161 	loff_t off = (loff_t)pgoff << PAGE_SHIFT;
1162 
1163 	ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE, vm_flags);
1164 	if (ret)
1165 		return ret;
1166 
1167 	return mm_get_unmapped_area_vmflags(filp, addr, len, pgoff, flags,
1168 					    vm_flags);
1169 }
1170 
1171 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
1172 		unsigned long len, unsigned long pgoff, unsigned long flags)
1173 {
1174 	return thp_get_unmapped_area_vmflags(filp, addr, len, pgoff, flags, 0);
1175 }
1176 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
1177 
1178 static struct folio *vma_alloc_anon_folio_pmd(struct vm_area_struct *vma,
1179 		unsigned long addr)
1180 {
1181 	gfp_t gfp = vma_thp_gfp_mask(vma);
1182 	const int order = HPAGE_PMD_ORDER;
1183 	struct folio *folio;
1184 
1185 	folio = vma_alloc_folio(gfp, order, vma, addr & HPAGE_PMD_MASK);
1186 
1187 	if (unlikely(!folio)) {
1188 		count_vm_event(THP_FAULT_FALLBACK);
1189 		count_mthp_stat(order, MTHP_STAT_ANON_FAULT_FALLBACK);
1190 		return NULL;
1191 	}
1192 
1193 	VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
1194 	if (mem_cgroup_charge(folio, vma->vm_mm, gfp)) {
1195 		folio_put(folio);
1196 		count_vm_event(THP_FAULT_FALLBACK);
1197 		count_vm_event(THP_FAULT_FALLBACK_CHARGE);
1198 		count_mthp_stat(order, MTHP_STAT_ANON_FAULT_FALLBACK);
1199 		count_mthp_stat(order, MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE);
1200 		return NULL;
1201 	}
1202 	folio_throttle_swaprate(folio, gfp);
1203 
1204        /*
1205 	* When a folio is not zeroed during allocation (__GFP_ZERO not used)
1206 	* or user folios require special handling, folio_zero_user() is used to
1207 	* make sure that the page corresponding to the faulting address will be
1208 	* hot in the cache after zeroing.
1209 	*/
1210 	if (user_alloc_needs_zeroing())
1211 		folio_zero_user(folio, addr);
1212 	/*
1213 	 * The memory barrier inside __folio_mark_uptodate makes sure that
1214 	 * folio_zero_user writes become visible before the set_pmd_at()
1215 	 * write.
1216 	 */
1217 	__folio_mark_uptodate(folio);
1218 	return folio;
1219 }
1220 
1221 static void map_anon_folio_pmd(struct folio *folio, pmd_t *pmd,
1222 		struct vm_area_struct *vma, unsigned long haddr)
1223 {
1224 	pmd_t entry;
1225 
1226 	entry = folio_mk_pmd(folio, vma->vm_page_prot);
1227 	entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1228 	folio_add_new_anon_rmap(folio, vma, haddr, RMAP_EXCLUSIVE);
1229 	folio_add_lru_vma(folio, vma);
1230 	set_pmd_at(vma->vm_mm, haddr, pmd, entry);
1231 	update_mmu_cache_pmd(vma, haddr, pmd);
1232 	add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1233 	count_vm_event(THP_FAULT_ALLOC);
1234 	count_mthp_stat(HPAGE_PMD_ORDER, MTHP_STAT_ANON_FAULT_ALLOC);
1235 	count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
1236 }
1237 
1238 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf)
1239 {
1240 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1241 	struct vm_area_struct *vma = vmf->vma;
1242 	struct folio *folio;
1243 	pgtable_t pgtable;
1244 	vm_fault_t ret = 0;
1245 
1246 	folio = vma_alloc_anon_folio_pmd(vma, vmf->address);
1247 	if (unlikely(!folio))
1248 		return VM_FAULT_FALLBACK;
1249 
1250 	pgtable = pte_alloc_one(vma->vm_mm);
1251 	if (unlikely(!pgtable)) {
1252 		ret = VM_FAULT_OOM;
1253 		goto release;
1254 	}
1255 
1256 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1257 	if (unlikely(!pmd_none(*vmf->pmd))) {
1258 		goto unlock_release;
1259 	} else {
1260 		ret = check_stable_address_space(vma->vm_mm);
1261 		if (ret)
1262 			goto unlock_release;
1263 
1264 		/* Deliver the page fault to userland */
1265 		if (userfaultfd_missing(vma)) {
1266 			spin_unlock(vmf->ptl);
1267 			folio_put(folio);
1268 			pte_free(vma->vm_mm, pgtable);
1269 			ret = handle_userfault(vmf, VM_UFFD_MISSING);
1270 			VM_BUG_ON(ret & VM_FAULT_FALLBACK);
1271 			return ret;
1272 		}
1273 		pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
1274 		map_anon_folio_pmd(folio, vmf->pmd, vma, haddr);
1275 		mm_inc_nr_ptes(vma->vm_mm);
1276 		deferred_split_folio(folio, false);
1277 		spin_unlock(vmf->ptl);
1278 	}
1279 
1280 	return 0;
1281 unlock_release:
1282 	spin_unlock(vmf->ptl);
1283 release:
1284 	if (pgtable)
1285 		pte_free(vma->vm_mm, pgtable);
1286 	folio_put(folio);
1287 	return ret;
1288 
1289 }
1290 
1291 /*
1292  * always: directly stall for all thp allocations
1293  * defer: wake kswapd and fail if not immediately available
1294  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
1295  *		  fail if not immediately available
1296  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
1297  *	    available
1298  * never: never stall for any thp allocation
1299  */
1300 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
1301 {
1302 	const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
1303 
1304 	/* Always do synchronous compaction */
1305 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
1306 		return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
1307 
1308 	/* Kick kcompactd and fail quickly */
1309 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
1310 		return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
1311 
1312 	/* Synchronous compaction if madvised, otherwise kick kcompactd */
1313 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
1314 		return GFP_TRANSHUGE_LIGHT |
1315 			(vma_madvised ? __GFP_DIRECT_RECLAIM :
1316 					__GFP_KSWAPD_RECLAIM);
1317 
1318 	/* Only do synchronous compaction if madvised */
1319 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
1320 		return GFP_TRANSHUGE_LIGHT |
1321 		       (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
1322 
1323 	return GFP_TRANSHUGE_LIGHT;
1324 }
1325 
1326 /* Caller must hold page table lock. */
1327 static void set_huge_zero_folio(pgtable_t pgtable, struct mm_struct *mm,
1328 		struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
1329 		struct folio *zero_folio)
1330 {
1331 	pmd_t entry;
1332 	entry = folio_mk_pmd(zero_folio, vma->vm_page_prot);
1333 	entry = pmd_mkspecial(entry);
1334 	pgtable_trans_huge_deposit(mm, pmd, pgtable);
1335 	set_pmd_at(mm, haddr, pmd, entry);
1336 	mm_inc_nr_ptes(mm);
1337 }
1338 
1339 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
1340 {
1341 	struct vm_area_struct *vma = vmf->vma;
1342 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1343 	vm_fault_t ret;
1344 
1345 	if (!thp_vma_suitable_order(vma, haddr, PMD_ORDER))
1346 		return VM_FAULT_FALLBACK;
1347 	ret = vmf_anon_prepare(vmf);
1348 	if (ret)
1349 		return ret;
1350 	khugepaged_enter_vma(vma, vma->vm_flags);
1351 
1352 	if (!(vmf->flags & FAULT_FLAG_WRITE) &&
1353 			!mm_forbids_zeropage(vma->vm_mm) &&
1354 			transparent_hugepage_use_zero_page()) {
1355 		pgtable_t pgtable;
1356 		struct folio *zero_folio;
1357 		vm_fault_t ret;
1358 
1359 		pgtable = pte_alloc_one(vma->vm_mm);
1360 		if (unlikely(!pgtable))
1361 			return VM_FAULT_OOM;
1362 		zero_folio = mm_get_huge_zero_folio(vma->vm_mm);
1363 		if (unlikely(!zero_folio)) {
1364 			pte_free(vma->vm_mm, pgtable);
1365 			count_vm_event(THP_FAULT_FALLBACK);
1366 			return VM_FAULT_FALLBACK;
1367 		}
1368 		vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1369 		ret = 0;
1370 		if (pmd_none(*vmf->pmd)) {
1371 			ret = check_stable_address_space(vma->vm_mm);
1372 			if (ret) {
1373 				spin_unlock(vmf->ptl);
1374 				pte_free(vma->vm_mm, pgtable);
1375 			} else if (userfaultfd_missing(vma)) {
1376 				spin_unlock(vmf->ptl);
1377 				pte_free(vma->vm_mm, pgtable);
1378 				ret = handle_userfault(vmf, VM_UFFD_MISSING);
1379 				VM_BUG_ON(ret & VM_FAULT_FALLBACK);
1380 			} else {
1381 				set_huge_zero_folio(pgtable, vma->vm_mm, vma,
1382 						   haddr, vmf->pmd, zero_folio);
1383 				update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1384 				spin_unlock(vmf->ptl);
1385 			}
1386 		} else {
1387 			spin_unlock(vmf->ptl);
1388 			pte_free(vma->vm_mm, pgtable);
1389 		}
1390 		return ret;
1391 	}
1392 
1393 	return __do_huge_pmd_anonymous_page(vmf);
1394 }
1395 
1396 struct folio_or_pfn {
1397 	union {
1398 		struct folio *folio;
1399 		unsigned long pfn;
1400 	};
1401 	bool is_folio;
1402 };
1403 
1404 static vm_fault_t insert_pmd(struct vm_area_struct *vma, unsigned long addr,
1405 		pmd_t *pmd, struct folio_or_pfn fop, pgprot_t prot,
1406 		bool write)
1407 {
1408 	struct mm_struct *mm = vma->vm_mm;
1409 	pgtable_t pgtable = NULL;
1410 	spinlock_t *ptl;
1411 	pmd_t entry;
1412 
1413 	if (addr < vma->vm_start || addr >= vma->vm_end)
1414 		return VM_FAULT_SIGBUS;
1415 
1416 	if (arch_needs_pgtable_deposit()) {
1417 		pgtable = pte_alloc_one(vma->vm_mm);
1418 		if (!pgtable)
1419 			return VM_FAULT_OOM;
1420 	}
1421 
1422 	ptl = pmd_lock(mm, pmd);
1423 	if (!pmd_none(*pmd)) {
1424 		const unsigned long pfn = fop.is_folio ? folio_pfn(fop.folio) :
1425 					  fop.pfn;
1426 
1427 		if (write) {
1428 			if (pmd_pfn(*pmd) != pfn) {
1429 				WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
1430 				goto out_unlock;
1431 			}
1432 			entry = pmd_mkyoung(*pmd);
1433 			entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1434 			if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
1435 				update_mmu_cache_pmd(vma, addr, pmd);
1436 		}
1437 		goto out_unlock;
1438 	}
1439 
1440 	if (fop.is_folio) {
1441 		entry = folio_mk_pmd(fop.folio, vma->vm_page_prot);
1442 
1443 		if (is_huge_zero_folio(fop.folio)) {
1444 			entry = pmd_mkspecial(entry);
1445 		} else {
1446 			folio_get(fop.folio);
1447 			folio_add_file_rmap_pmd(fop.folio, &fop.folio->page, vma);
1448 			add_mm_counter(mm, mm_counter_file(fop.folio), HPAGE_PMD_NR);
1449 		}
1450 	} else {
1451 		entry = pmd_mkhuge(pfn_pmd(fop.pfn, prot));
1452 		entry = pmd_mkspecial(entry);
1453 	}
1454 	if (write) {
1455 		entry = pmd_mkyoung(pmd_mkdirty(entry));
1456 		entry = maybe_pmd_mkwrite(entry, vma);
1457 	}
1458 
1459 	if (pgtable) {
1460 		pgtable_trans_huge_deposit(mm, pmd, pgtable);
1461 		mm_inc_nr_ptes(mm);
1462 		pgtable = NULL;
1463 	}
1464 
1465 	set_pmd_at(mm, addr, pmd, entry);
1466 	update_mmu_cache_pmd(vma, addr, pmd);
1467 
1468 out_unlock:
1469 	spin_unlock(ptl);
1470 	if (pgtable)
1471 		pte_free(mm, pgtable);
1472 	return VM_FAULT_NOPAGE;
1473 }
1474 
1475 /**
1476  * vmf_insert_pfn_pmd - insert a pmd size pfn
1477  * @vmf: Structure describing the fault
1478  * @pfn: pfn to insert
1479  * @write: whether it's a write fault
1480  *
1481  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info.
1482  *
1483  * Return: vm_fault_t value.
1484  */
1485 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, unsigned long pfn,
1486 			      bool write)
1487 {
1488 	unsigned long addr = vmf->address & PMD_MASK;
1489 	struct vm_area_struct *vma = vmf->vma;
1490 	pgprot_t pgprot = vma->vm_page_prot;
1491 	struct folio_or_pfn fop = {
1492 		.pfn = pfn,
1493 	};
1494 
1495 	/*
1496 	 * If we had pmd_special, we could avoid all these restrictions,
1497 	 * but we need to be consistent with PTEs and architectures that
1498 	 * can't support a 'special' bit.
1499 	 */
1500 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1501 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1502 						(VM_PFNMAP|VM_MIXEDMAP));
1503 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1504 
1505 	pfnmap_setup_cachemode_pfn(pfn, &pgprot);
1506 
1507 	return insert_pmd(vma, addr, vmf->pmd, fop, pgprot, write);
1508 }
1509 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
1510 
1511 vm_fault_t vmf_insert_folio_pmd(struct vm_fault *vmf, struct folio *folio,
1512 				bool write)
1513 {
1514 	struct vm_area_struct *vma = vmf->vma;
1515 	unsigned long addr = vmf->address & PMD_MASK;
1516 	struct folio_or_pfn fop = {
1517 		.folio = folio,
1518 		.is_folio = true,
1519 	};
1520 
1521 	if (WARN_ON_ONCE(folio_order(folio) != PMD_ORDER))
1522 		return VM_FAULT_SIGBUS;
1523 
1524 	return insert_pmd(vma, addr, vmf->pmd, fop, vma->vm_page_prot, write);
1525 }
1526 EXPORT_SYMBOL_GPL(vmf_insert_folio_pmd);
1527 
1528 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1529 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
1530 {
1531 	if (likely(vma->vm_flags & VM_WRITE))
1532 		pud = pud_mkwrite(pud);
1533 	return pud;
1534 }
1535 
1536 static vm_fault_t insert_pud(struct vm_area_struct *vma, unsigned long addr,
1537 		pud_t *pud, struct folio_or_pfn fop, pgprot_t prot, bool write)
1538 {
1539 	struct mm_struct *mm = vma->vm_mm;
1540 	spinlock_t *ptl;
1541 	pud_t entry;
1542 
1543 	if (addr < vma->vm_start || addr >= vma->vm_end)
1544 		return VM_FAULT_SIGBUS;
1545 
1546 	ptl = pud_lock(mm, pud);
1547 	if (!pud_none(*pud)) {
1548 		const unsigned long pfn = fop.is_folio ? folio_pfn(fop.folio) :
1549 					  fop.pfn;
1550 
1551 		if (write) {
1552 			if (WARN_ON_ONCE(pud_pfn(*pud) != pfn))
1553 				goto out_unlock;
1554 			entry = pud_mkyoung(*pud);
1555 			entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
1556 			if (pudp_set_access_flags(vma, addr, pud, entry, 1))
1557 				update_mmu_cache_pud(vma, addr, pud);
1558 		}
1559 		goto out_unlock;
1560 	}
1561 
1562 	if (fop.is_folio) {
1563 		entry = folio_mk_pud(fop.folio, vma->vm_page_prot);
1564 
1565 		folio_get(fop.folio);
1566 		folio_add_file_rmap_pud(fop.folio, &fop.folio->page, vma);
1567 		add_mm_counter(mm, mm_counter_file(fop.folio), HPAGE_PUD_NR);
1568 	} else {
1569 		entry = pud_mkhuge(pfn_pud(fop.pfn, prot));
1570 		entry = pud_mkspecial(entry);
1571 	}
1572 	if (write) {
1573 		entry = pud_mkyoung(pud_mkdirty(entry));
1574 		entry = maybe_pud_mkwrite(entry, vma);
1575 	}
1576 	set_pud_at(mm, addr, pud, entry);
1577 	update_mmu_cache_pud(vma, addr, pud);
1578 out_unlock:
1579 	spin_unlock(ptl);
1580 	return VM_FAULT_NOPAGE;
1581 }
1582 
1583 /**
1584  * vmf_insert_pfn_pud - insert a pud size pfn
1585  * @vmf: Structure describing the fault
1586  * @pfn: pfn to insert
1587  * @write: whether it's a write fault
1588  *
1589  * Insert a pud size pfn. See vmf_insert_pfn() for additional info.
1590  *
1591  * Return: vm_fault_t value.
1592  */
1593 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, unsigned long pfn,
1594 			      bool write)
1595 {
1596 	unsigned long addr = vmf->address & PUD_MASK;
1597 	struct vm_area_struct *vma = vmf->vma;
1598 	pgprot_t pgprot = vma->vm_page_prot;
1599 	struct folio_or_pfn fop = {
1600 		.pfn = pfn,
1601 	};
1602 
1603 	/*
1604 	 * If we had pud_special, we could avoid all these restrictions,
1605 	 * but we need to be consistent with PTEs and architectures that
1606 	 * can't support a 'special' bit.
1607 	 */
1608 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1609 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1610 						(VM_PFNMAP|VM_MIXEDMAP));
1611 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1612 
1613 	pfnmap_setup_cachemode_pfn(pfn, &pgprot);
1614 
1615 	return insert_pud(vma, addr, vmf->pud, fop, pgprot, write);
1616 }
1617 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
1618 
1619 /**
1620  * vmf_insert_folio_pud - insert a pud size folio mapped by a pud entry
1621  * @vmf: Structure describing the fault
1622  * @folio: folio to insert
1623  * @write: whether it's a write fault
1624  *
1625  * Return: vm_fault_t value.
1626  */
1627 vm_fault_t vmf_insert_folio_pud(struct vm_fault *vmf, struct folio *folio,
1628 				bool write)
1629 {
1630 	struct vm_area_struct *vma = vmf->vma;
1631 	unsigned long addr = vmf->address & PUD_MASK;
1632 	struct folio_or_pfn fop = {
1633 		.folio = folio,
1634 		.is_folio = true,
1635 	};
1636 
1637 	if (WARN_ON_ONCE(folio_order(folio) != PUD_ORDER))
1638 		return VM_FAULT_SIGBUS;
1639 
1640 	return insert_pud(vma, addr, vmf->pud, fop, vma->vm_page_prot, write);
1641 }
1642 EXPORT_SYMBOL_GPL(vmf_insert_folio_pud);
1643 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1644 
1645 void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1646 	       pmd_t *pmd, bool write)
1647 {
1648 	pmd_t _pmd;
1649 
1650 	_pmd = pmd_mkyoung(*pmd);
1651 	if (write)
1652 		_pmd = pmd_mkdirty(_pmd);
1653 	if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1654 				  pmd, _pmd, write))
1655 		update_mmu_cache_pmd(vma, addr, pmd);
1656 }
1657 
1658 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1659 		  pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1660 		  struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1661 {
1662 	spinlock_t *dst_ptl, *src_ptl;
1663 	struct page *src_page;
1664 	struct folio *src_folio;
1665 	pmd_t pmd;
1666 	pgtable_t pgtable = NULL;
1667 	int ret = -ENOMEM;
1668 
1669 	pmd = pmdp_get_lockless(src_pmd);
1670 	if (unlikely(pmd_present(pmd) && pmd_special(pmd) &&
1671 		     !is_huge_zero_pmd(pmd))) {
1672 		dst_ptl = pmd_lock(dst_mm, dst_pmd);
1673 		src_ptl = pmd_lockptr(src_mm, src_pmd);
1674 		spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1675 		/*
1676 		 * No need to recheck the pmd, it can't change with write
1677 		 * mmap lock held here.
1678 		 *
1679 		 * Meanwhile, making sure it's not a CoW VMA with writable
1680 		 * mapping, otherwise it means either the anon page wrongly
1681 		 * applied special bit, or we made the PRIVATE mapping be
1682 		 * able to wrongly write to the backend MMIO.
1683 		 */
1684 		VM_WARN_ON_ONCE(is_cow_mapping(src_vma->vm_flags) && pmd_write(pmd));
1685 		goto set_pmd;
1686 	}
1687 
1688 	/* Skip if can be re-fill on fault */
1689 	if (!vma_is_anonymous(dst_vma))
1690 		return 0;
1691 
1692 	pgtable = pte_alloc_one(dst_mm);
1693 	if (unlikely(!pgtable))
1694 		goto out;
1695 
1696 	dst_ptl = pmd_lock(dst_mm, dst_pmd);
1697 	src_ptl = pmd_lockptr(src_mm, src_pmd);
1698 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1699 
1700 	ret = -EAGAIN;
1701 	pmd = *src_pmd;
1702 
1703 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1704 	if (unlikely(is_swap_pmd(pmd))) {
1705 		swp_entry_t entry = pmd_to_swp_entry(pmd);
1706 
1707 		VM_BUG_ON(!is_pmd_migration_entry(pmd));
1708 		if (!is_readable_migration_entry(entry)) {
1709 			entry = make_readable_migration_entry(
1710 							swp_offset(entry));
1711 			pmd = swp_entry_to_pmd(entry);
1712 			if (pmd_swp_soft_dirty(*src_pmd))
1713 				pmd = pmd_swp_mksoft_dirty(pmd);
1714 			if (pmd_swp_uffd_wp(*src_pmd))
1715 				pmd = pmd_swp_mkuffd_wp(pmd);
1716 			set_pmd_at(src_mm, addr, src_pmd, pmd);
1717 		}
1718 		add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1719 		mm_inc_nr_ptes(dst_mm);
1720 		pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1721 		if (!userfaultfd_wp(dst_vma))
1722 			pmd = pmd_swp_clear_uffd_wp(pmd);
1723 		set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1724 		ret = 0;
1725 		goto out_unlock;
1726 	}
1727 #endif
1728 
1729 	if (unlikely(!pmd_trans_huge(pmd))) {
1730 		pte_free(dst_mm, pgtable);
1731 		goto out_unlock;
1732 	}
1733 	/*
1734 	 * When page table lock is held, the huge zero pmd should not be
1735 	 * under splitting since we don't split the page itself, only pmd to
1736 	 * a page table.
1737 	 */
1738 	if (is_huge_zero_pmd(pmd)) {
1739 		/*
1740 		 * mm_get_huge_zero_folio() will never allocate a new
1741 		 * folio here, since we already have a zero page to
1742 		 * copy. It just takes a reference.
1743 		 */
1744 		mm_get_huge_zero_folio(dst_mm);
1745 		goto out_zero_page;
1746 	}
1747 
1748 	src_page = pmd_page(pmd);
1749 	VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1750 	src_folio = page_folio(src_page);
1751 
1752 	folio_get(src_folio);
1753 	if (unlikely(folio_try_dup_anon_rmap_pmd(src_folio, src_page, dst_vma, src_vma))) {
1754 		/* Page maybe pinned: split and retry the fault on PTEs. */
1755 		folio_put(src_folio);
1756 		pte_free(dst_mm, pgtable);
1757 		spin_unlock(src_ptl);
1758 		spin_unlock(dst_ptl);
1759 		__split_huge_pmd(src_vma, src_pmd, addr, false);
1760 		return -EAGAIN;
1761 	}
1762 	add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1763 out_zero_page:
1764 	mm_inc_nr_ptes(dst_mm);
1765 	pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1766 	pmdp_set_wrprotect(src_mm, addr, src_pmd);
1767 	if (!userfaultfd_wp(dst_vma))
1768 		pmd = pmd_clear_uffd_wp(pmd);
1769 	pmd = pmd_wrprotect(pmd);
1770 set_pmd:
1771 	pmd = pmd_mkold(pmd);
1772 	set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1773 
1774 	ret = 0;
1775 out_unlock:
1776 	spin_unlock(src_ptl);
1777 	spin_unlock(dst_ptl);
1778 out:
1779 	return ret;
1780 }
1781 
1782 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1783 void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1784 	       pud_t *pud, bool write)
1785 {
1786 	pud_t _pud;
1787 
1788 	_pud = pud_mkyoung(*pud);
1789 	if (write)
1790 		_pud = pud_mkdirty(_pud);
1791 	if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1792 				  pud, _pud, write))
1793 		update_mmu_cache_pud(vma, addr, pud);
1794 }
1795 
1796 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1797 		  pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1798 		  struct vm_area_struct *vma)
1799 {
1800 	spinlock_t *dst_ptl, *src_ptl;
1801 	pud_t pud;
1802 	int ret;
1803 
1804 	dst_ptl = pud_lock(dst_mm, dst_pud);
1805 	src_ptl = pud_lockptr(src_mm, src_pud);
1806 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1807 
1808 	ret = -EAGAIN;
1809 	pud = *src_pud;
1810 	if (unlikely(!pud_trans_huge(pud)))
1811 		goto out_unlock;
1812 
1813 	/*
1814 	 * TODO: once we support anonymous pages, use
1815 	 * folio_try_dup_anon_rmap_*() and split if duplicating fails.
1816 	 */
1817 	if (is_cow_mapping(vma->vm_flags) && pud_write(pud)) {
1818 		pudp_set_wrprotect(src_mm, addr, src_pud);
1819 		pud = pud_wrprotect(pud);
1820 	}
1821 	pud = pud_mkold(pud);
1822 	set_pud_at(dst_mm, addr, dst_pud, pud);
1823 
1824 	ret = 0;
1825 out_unlock:
1826 	spin_unlock(src_ptl);
1827 	spin_unlock(dst_ptl);
1828 	return ret;
1829 }
1830 
1831 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1832 {
1833 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1834 
1835 	vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1836 	if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1837 		goto unlock;
1838 
1839 	touch_pud(vmf->vma, vmf->address, vmf->pud, write);
1840 unlock:
1841 	spin_unlock(vmf->ptl);
1842 }
1843 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1844 
1845 void huge_pmd_set_accessed(struct vm_fault *vmf)
1846 {
1847 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1848 
1849 	vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1850 	if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd)))
1851 		goto unlock;
1852 
1853 	touch_pmd(vmf->vma, vmf->address, vmf->pmd, write);
1854 
1855 unlock:
1856 	spin_unlock(vmf->ptl);
1857 }
1858 
1859 static vm_fault_t do_huge_zero_wp_pmd(struct vm_fault *vmf)
1860 {
1861 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1862 	struct vm_area_struct *vma = vmf->vma;
1863 	struct mmu_notifier_range range;
1864 	struct folio *folio;
1865 	vm_fault_t ret = 0;
1866 
1867 	folio = vma_alloc_anon_folio_pmd(vma, vmf->address);
1868 	if (unlikely(!folio))
1869 		return VM_FAULT_FALLBACK;
1870 
1871 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, haddr,
1872 				haddr + HPAGE_PMD_SIZE);
1873 	mmu_notifier_invalidate_range_start(&range);
1874 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1875 	if (unlikely(!pmd_same(pmdp_get(vmf->pmd), vmf->orig_pmd)))
1876 		goto release;
1877 	ret = check_stable_address_space(vma->vm_mm);
1878 	if (ret)
1879 		goto release;
1880 	(void)pmdp_huge_clear_flush(vma, haddr, vmf->pmd);
1881 	map_anon_folio_pmd(folio, vmf->pmd, vma, haddr);
1882 	goto unlock;
1883 release:
1884 	folio_put(folio);
1885 unlock:
1886 	spin_unlock(vmf->ptl);
1887 	mmu_notifier_invalidate_range_end(&range);
1888 	return ret;
1889 }
1890 
1891 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1892 {
1893 	const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
1894 	struct vm_area_struct *vma = vmf->vma;
1895 	struct folio *folio;
1896 	struct page *page;
1897 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1898 	pmd_t orig_pmd = vmf->orig_pmd;
1899 
1900 	vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1901 	VM_BUG_ON_VMA(!vma->anon_vma, vma);
1902 
1903 	if (is_huge_zero_pmd(orig_pmd)) {
1904 		vm_fault_t ret = do_huge_zero_wp_pmd(vmf);
1905 
1906 		if (!(ret & VM_FAULT_FALLBACK))
1907 			return ret;
1908 
1909 		/* Fallback to splitting PMD if THP cannot be allocated */
1910 		goto fallback;
1911 	}
1912 
1913 	spin_lock(vmf->ptl);
1914 
1915 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1916 		spin_unlock(vmf->ptl);
1917 		return 0;
1918 	}
1919 
1920 	page = pmd_page(orig_pmd);
1921 	folio = page_folio(page);
1922 	VM_BUG_ON_PAGE(!PageHead(page), page);
1923 
1924 	/* Early check when only holding the PT lock. */
1925 	if (PageAnonExclusive(page))
1926 		goto reuse;
1927 
1928 	if (!folio_trylock(folio)) {
1929 		folio_get(folio);
1930 		spin_unlock(vmf->ptl);
1931 		folio_lock(folio);
1932 		spin_lock(vmf->ptl);
1933 		if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1934 			spin_unlock(vmf->ptl);
1935 			folio_unlock(folio);
1936 			folio_put(folio);
1937 			return 0;
1938 		}
1939 		folio_put(folio);
1940 	}
1941 
1942 	/* Recheck after temporarily dropping the PT lock. */
1943 	if (PageAnonExclusive(page)) {
1944 		folio_unlock(folio);
1945 		goto reuse;
1946 	}
1947 
1948 	/*
1949 	 * See do_wp_page(): we can only reuse the folio exclusively if
1950 	 * there are no additional references. Note that we always drain
1951 	 * the LRU cache immediately after adding a THP.
1952 	 */
1953 	if (folio_ref_count(folio) >
1954 			1 + folio_test_swapcache(folio) * folio_nr_pages(folio))
1955 		goto unlock_fallback;
1956 	if (folio_test_swapcache(folio))
1957 		folio_free_swap(folio);
1958 	if (folio_ref_count(folio) == 1) {
1959 		pmd_t entry;
1960 
1961 		folio_move_anon_rmap(folio, vma);
1962 		SetPageAnonExclusive(page);
1963 		folio_unlock(folio);
1964 reuse:
1965 		if (unlikely(unshare)) {
1966 			spin_unlock(vmf->ptl);
1967 			return 0;
1968 		}
1969 		entry = pmd_mkyoung(orig_pmd);
1970 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1971 		if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1972 			update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1973 		spin_unlock(vmf->ptl);
1974 		return 0;
1975 	}
1976 
1977 unlock_fallback:
1978 	folio_unlock(folio);
1979 	spin_unlock(vmf->ptl);
1980 fallback:
1981 	__split_huge_pmd(vma, vmf->pmd, vmf->address, false);
1982 	return VM_FAULT_FALLBACK;
1983 }
1984 
1985 static inline bool can_change_pmd_writable(struct vm_area_struct *vma,
1986 					   unsigned long addr, pmd_t pmd)
1987 {
1988 	struct page *page;
1989 
1990 	if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE)))
1991 		return false;
1992 
1993 	/* Don't touch entries that are not even readable (NUMA hinting). */
1994 	if (pmd_protnone(pmd))
1995 		return false;
1996 
1997 	/* Do we need write faults for softdirty tracking? */
1998 	if (pmd_needs_soft_dirty_wp(vma, pmd))
1999 		return false;
2000 
2001 	/* Do we need write faults for uffd-wp tracking? */
2002 	if (userfaultfd_huge_pmd_wp(vma, pmd))
2003 		return false;
2004 
2005 	if (!(vma->vm_flags & VM_SHARED)) {
2006 		/* See can_change_pte_writable(). */
2007 		page = vm_normal_page_pmd(vma, addr, pmd);
2008 		return page && PageAnon(page) && PageAnonExclusive(page);
2009 	}
2010 
2011 	/* See can_change_pte_writable(). */
2012 	return pmd_dirty(pmd);
2013 }
2014 
2015 /* NUMA hinting page fault entry point for trans huge pmds */
2016 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
2017 {
2018 	struct vm_area_struct *vma = vmf->vma;
2019 	struct folio *folio;
2020 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
2021 	int nid = NUMA_NO_NODE;
2022 	int target_nid, last_cpupid;
2023 	pmd_t pmd, old_pmd;
2024 	bool writable = false;
2025 	int flags = 0;
2026 
2027 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
2028 	old_pmd = pmdp_get(vmf->pmd);
2029 
2030 	if (unlikely(!pmd_same(old_pmd, vmf->orig_pmd))) {
2031 		spin_unlock(vmf->ptl);
2032 		return 0;
2033 	}
2034 
2035 	pmd = pmd_modify(old_pmd, vma->vm_page_prot);
2036 
2037 	/*
2038 	 * Detect now whether the PMD could be writable; this information
2039 	 * is only valid while holding the PT lock.
2040 	 */
2041 	writable = pmd_write(pmd);
2042 	if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
2043 	    can_change_pmd_writable(vma, vmf->address, pmd))
2044 		writable = true;
2045 
2046 	folio = vm_normal_folio_pmd(vma, haddr, pmd);
2047 	if (!folio)
2048 		goto out_map;
2049 
2050 	nid = folio_nid(folio);
2051 
2052 	target_nid = numa_migrate_check(folio, vmf, haddr, &flags, writable,
2053 					&last_cpupid);
2054 	if (target_nid == NUMA_NO_NODE)
2055 		goto out_map;
2056 	if (migrate_misplaced_folio_prepare(folio, vma, target_nid)) {
2057 		flags |= TNF_MIGRATE_FAIL;
2058 		goto out_map;
2059 	}
2060 	/* The folio is isolated and isolation code holds a folio reference. */
2061 	spin_unlock(vmf->ptl);
2062 	writable = false;
2063 
2064 	if (!migrate_misplaced_folio(folio, target_nid)) {
2065 		flags |= TNF_MIGRATED;
2066 		nid = target_nid;
2067 		task_numa_fault(last_cpupid, nid, HPAGE_PMD_NR, flags);
2068 		return 0;
2069 	}
2070 
2071 	flags |= TNF_MIGRATE_FAIL;
2072 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
2073 	if (unlikely(!pmd_same(pmdp_get(vmf->pmd), vmf->orig_pmd))) {
2074 		spin_unlock(vmf->ptl);
2075 		return 0;
2076 	}
2077 out_map:
2078 	/* Restore the PMD */
2079 	pmd = pmd_modify(pmdp_get(vmf->pmd), vma->vm_page_prot);
2080 	pmd = pmd_mkyoung(pmd);
2081 	if (writable)
2082 		pmd = pmd_mkwrite(pmd, vma);
2083 	set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
2084 	update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
2085 	spin_unlock(vmf->ptl);
2086 
2087 	if (nid != NUMA_NO_NODE)
2088 		task_numa_fault(last_cpupid, nid, HPAGE_PMD_NR, flags);
2089 	return 0;
2090 }
2091 
2092 /*
2093  * Return true if we do MADV_FREE successfully on entire pmd page.
2094  * Otherwise, return false.
2095  */
2096 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
2097 		pmd_t *pmd, unsigned long addr, unsigned long next)
2098 {
2099 	spinlock_t *ptl;
2100 	pmd_t orig_pmd;
2101 	struct folio *folio;
2102 	struct mm_struct *mm = tlb->mm;
2103 	bool ret = false;
2104 
2105 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
2106 
2107 	ptl = pmd_trans_huge_lock(pmd, vma);
2108 	if (!ptl)
2109 		goto out_unlocked;
2110 
2111 	orig_pmd = *pmd;
2112 	if (is_huge_zero_pmd(orig_pmd))
2113 		goto out;
2114 
2115 	if (unlikely(!pmd_present(orig_pmd))) {
2116 		VM_BUG_ON(thp_migration_supported() &&
2117 				  !is_pmd_migration_entry(orig_pmd));
2118 		goto out;
2119 	}
2120 
2121 	folio = pmd_folio(orig_pmd);
2122 	/*
2123 	 * If other processes are mapping this folio, we couldn't discard
2124 	 * the folio unless they all do MADV_FREE so let's skip the folio.
2125 	 */
2126 	if (folio_maybe_mapped_shared(folio))
2127 		goto out;
2128 
2129 	if (!folio_trylock(folio))
2130 		goto out;
2131 
2132 	/*
2133 	 * If user want to discard part-pages of THP, split it so MADV_FREE
2134 	 * will deactivate only them.
2135 	 */
2136 	if (next - addr != HPAGE_PMD_SIZE) {
2137 		folio_get(folio);
2138 		spin_unlock(ptl);
2139 		split_folio(folio);
2140 		folio_unlock(folio);
2141 		folio_put(folio);
2142 		goto out_unlocked;
2143 	}
2144 
2145 	if (folio_test_dirty(folio))
2146 		folio_clear_dirty(folio);
2147 	folio_unlock(folio);
2148 
2149 	if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
2150 		pmdp_invalidate(vma, addr, pmd);
2151 		orig_pmd = pmd_mkold(orig_pmd);
2152 		orig_pmd = pmd_mkclean(orig_pmd);
2153 
2154 		set_pmd_at(mm, addr, pmd, orig_pmd);
2155 		tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
2156 	}
2157 
2158 	folio_mark_lazyfree(folio);
2159 	ret = true;
2160 out:
2161 	spin_unlock(ptl);
2162 out_unlocked:
2163 	return ret;
2164 }
2165 
2166 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
2167 {
2168 	pgtable_t pgtable;
2169 
2170 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2171 	pte_free(mm, pgtable);
2172 	mm_dec_nr_ptes(mm);
2173 }
2174 
2175 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
2176 		 pmd_t *pmd, unsigned long addr)
2177 {
2178 	pmd_t orig_pmd;
2179 	spinlock_t *ptl;
2180 
2181 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
2182 
2183 	ptl = __pmd_trans_huge_lock(pmd, vma);
2184 	if (!ptl)
2185 		return 0;
2186 	/*
2187 	 * For architectures like ppc64 we look at deposited pgtable
2188 	 * when calling pmdp_huge_get_and_clear. So do the
2189 	 * pgtable_trans_huge_withdraw after finishing pmdp related
2190 	 * operations.
2191 	 */
2192 	orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
2193 						tlb->fullmm);
2194 	arch_check_zapped_pmd(vma, orig_pmd);
2195 	tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
2196 	if (!vma_is_dax(vma) && vma_is_special_huge(vma)) {
2197 		if (arch_needs_pgtable_deposit())
2198 			zap_deposited_table(tlb->mm, pmd);
2199 		spin_unlock(ptl);
2200 	} else if (is_huge_zero_pmd(orig_pmd)) {
2201 		if (!vma_is_dax(vma) || arch_needs_pgtable_deposit())
2202 			zap_deposited_table(tlb->mm, pmd);
2203 		spin_unlock(ptl);
2204 	} else {
2205 		struct folio *folio = NULL;
2206 		int flush_needed = 1;
2207 
2208 		if (pmd_present(orig_pmd)) {
2209 			struct page *page = pmd_page(orig_pmd);
2210 
2211 			folio = page_folio(page);
2212 			folio_remove_rmap_pmd(folio, page, vma);
2213 			WARN_ON_ONCE(folio_mapcount(folio) < 0);
2214 			VM_BUG_ON_PAGE(!PageHead(page), page);
2215 		} else if (thp_migration_supported()) {
2216 			swp_entry_t entry;
2217 
2218 			VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
2219 			entry = pmd_to_swp_entry(orig_pmd);
2220 			folio = pfn_swap_entry_folio(entry);
2221 			flush_needed = 0;
2222 		} else
2223 			WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
2224 
2225 		if (folio_test_anon(folio)) {
2226 			zap_deposited_table(tlb->mm, pmd);
2227 			add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
2228 		} else {
2229 			if (arch_needs_pgtable_deposit())
2230 				zap_deposited_table(tlb->mm, pmd);
2231 			add_mm_counter(tlb->mm, mm_counter_file(folio),
2232 				       -HPAGE_PMD_NR);
2233 
2234 			/*
2235 			 * Use flush_needed to indicate whether the PMD entry
2236 			 * is present, instead of checking pmd_present() again.
2237 			 */
2238 			if (flush_needed && pmd_young(orig_pmd) &&
2239 			    likely(vma_has_recency(vma)))
2240 				folio_mark_accessed(folio);
2241 		}
2242 
2243 		spin_unlock(ptl);
2244 		if (flush_needed)
2245 			tlb_remove_page_size(tlb, &folio->page, HPAGE_PMD_SIZE);
2246 	}
2247 	return 1;
2248 }
2249 
2250 #ifndef pmd_move_must_withdraw
2251 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
2252 					 spinlock_t *old_pmd_ptl,
2253 					 struct vm_area_struct *vma)
2254 {
2255 	/*
2256 	 * With split pmd lock we also need to move preallocated
2257 	 * PTE page table if new_pmd is on different PMD page table.
2258 	 *
2259 	 * We also don't deposit and withdraw tables for file pages.
2260 	 */
2261 	return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
2262 }
2263 #endif
2264 
2265 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
2266 {
2267 #ifdef CONFIG_MEM_SOFT_DIRTY
2268 	if (unlikely(is_pmd_migration_entry(pmd)))
2269 		pmd = pmd_swp_mksoft_dirty(pmd);
2270 	else if (pmd_present(pmd))
2271 		pmd = pmd_mksoft_dirty(pmd);
2272 #endif
2273 	return pmd;
2274 }
2275 
2276 static pmd_t clear_uffd_wp_pmd(pmd_t pmd)
2277 {
2278 	if (pmd_present(pmd))
2279 		pmd = pmd_clear_uffd_wp(pmd);
2280 	else if (is_swap_pmd(pmd))
2281 		pmd = pmd_swp_clear_uffd_wp(pmd);
2282 
2283 	return pmd;
2284 }
2285 
2286 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
2287 		  unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
2288 {
2289 	spinlock_t *old_ptl, *new_ptl;
2290 	pmd_t pmd;
2291 	struct mm_struct *mm = vma->vm_mm;
2292 	bool force_flush = false;
2293 
2294 	/*
2295 	 * The destination pmd shouldn't be established, free_pgtables()
2296 	 * should have released it; but move_page_tables() might have already
2297 	 * inserted a page table, if racing against shmem/file collapse.
2298 	 */
2299 	if (!pmd_none(*new_pmd)) {
2300 		VM_BUG_ON(pmd_trans_huge(*new_pmd));
2301 		return false;
2302 	}
2303 
2304 	/*
2305 	 * We don't have to worry about the ordering of src and dst
2306 	 * ptlocks because exclusive mmap_lock prevents deadlock.
2307 	 */
2308 	old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
2309 	if (old_ptl) {
2310 		new_ptl = pmd_lockptr(mm, new_pmd);
2311 		if (new_ptl != old_ptl)
2312 			spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
2313 		pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
2314 		if (pmd_present(pmd))
2315 			force_flush = true;
2316 		VM_BUG_ON(!pmd_none(*new_pmd));
2317 
2318 		if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
2319 			pgtable_t pgtable;
2320 			pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
2321 			pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
2322 		}
2323 		pmd = move_soft_dirty_pmd(pmd);
2324 		if (vma_has_uffd_without_event_remap(vma))
2325 			pmd = clear_uffd_wp_pmd(pmd);
2326 		set_pmd_at(mm, new_addr, new_pmd, pmd);
2327 		if (force_flush)
2328 			flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
2329 		if (new_ptl != old_ptl)
2330 			spin_unlock(new_ptl);
2331 		spin_unlock(old_ptl);
2332 		return true;
2333 	}
2334 	return false;
2335 }
2336 
2337 /*
2338  * Returns
2339  *  - 0 if PMD could not be locked
2340  *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
2341  *      or if prot_numa but THP migration is not supported
2342  *  - HPAGE_PMD_NR if protections changed and TLB flush necessary
2343  */
2344 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
2345 		    pmd_t *pmd, unsigned long addr, pgprot_t newprot,
2346 		    unsigned long cp_flags)
2347 {
2348 	struct mm_struct *mm = vma->vm_mm;
2349 	spinlock_t *ptl;
2350 	pmd_t oldpmd, entry;
2351 	bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
2352 	bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
2353 	bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
2354 	int ret = 1;
2355 
2356 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
2357 
2358 	if (prot_numa && !thp_migration_supported())
2359 		return 1;
2360 
2361 	ptl = __pmd_trans_huge_lock(pmd, vma);
2362 	if (!ptl)
2363 		return 0;
2364 
2365 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2366 	if (is_swap_pmd(*pmd)) {
2367 		swp_entry_t entry = pmd_to_swp_entry(*pmd);
2368 		struct folio *folio = pfn_swap_entry_folio(entry);
2369 		pmd_t newpmd;
2370 
2371 		VM_BUG_ON(!is_pmd_migration_entry(*pmd));
2372 		if (is_writable_migration_entry(entry)) {
2373 			/*
2374 			 * A protection check is difficult so
2375 			 * just be safe and disable write
2376 			 */
2377 			if (folio_test_anon(folio))
2378 				entry = make_readable_exclusive_migration_entry(swp_offset(entry));
2379 			else
2380 				entry = make_readable_migration_entry(swp_offset(entry));
2381 			newpmd = swp_entry_to_pmd(entry);
2382 			if (pmd_swp_soft_dirty(*pmd))
2383 				newpmd = pmd_swp_mksoft_dirty(newpmd);
2384 		} else {
2385 			newpmd = *pmd;
2386 		}
2387 
2388 		if (uffd_wp)
2389 			newpmd = pmd_swp_mkuffd_wp(newpmd);
2390 		else if (uffd_wp_resolve)
2391 			newpmd = pmd_swp_clear_uffd_wp(newpmd);
2392 		if (!pmd_same(*pmd, newpmd))
2393 			set_pmd_at(mm, addr, pmd, newpmd);
2394 		goto unlock;
2395 	}
2396 #endif
2397 
2398 	if (prot_numa) {
2399 
2400 		/*
2401 		 * Avoid trapping faults against the zero page. The read-only
2402 		 * data is likely to be read-cached on the local CPU and
2403 		 * local/remote hits to the zero page are not interesting.
2404 		 */
2405 		if (is_huge_zero_pmd(*pmd))
2406 			goto unlock;
2407 
2408 		if (pmd_protnone(*pmd))
2409 			goto unlock;
2410 
2411 		if (!folio_can_map_prot_numa(pmd_folio(*pmd), vma,
2412 					     vma_is_single_threaded_private(vma)))
2413 			goto unlock;
2414 	}
2415 	/*
2416 	 * In case prot_numa, we are under mmap_read_lock(mm). It's critical
2417 	 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
2418 	 * which is also under mmap_read_lock(mm):
2419 	 *
2420 	 *	CPU0:				CPU1:
2421 	 *				change_huge_pmd(prot_numa=1)
2422 	 *				 pmdp_huge_get_and_clear_notify()
2423 	 * madvise_dontneed()
2424 	 *  zap_pmd_range()
2425 	 *   pmd_trans_huge(*pmd) == 0 (without ptl)
2426 	 *   // skip the pmd
2427 	 *				 set_pmd_at();
2428 	 *				 // pmd is re-established
2429 	 *
2430 	 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
2431 	 * which may break userspace.
2432 	 *
2433 	 * pmdp_invalidate_ad() is required to make sure we don't miss
2434 	 * dirty/young flags set by hardware.
2435 	 */
2436 	oldpmd = pmdp_invalidate_ad(vma, addr, pmd);
2437 
2438 	entry = pmd_modify(oldpmd, newprot);
2439 	if (uffd_wp)
2440 		entry = pmd_mkuffd_wp(entry);
2441 	else if (uffd_wp_resolve)
2442 		/*
2443 		 * Leave the write bit to be handled by PF interrupt
2444 		 * handler, then things like COW could be properly
2445 		 * handled.
2446 		 */
2447 		entry = pmd_clear_uffd_wp(entry);
2448 
2449 	/* See change_pte_range(). */
2450 	if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) && !pmd_write(entry) &&
2451 	    can_change_pmd_writable(vma, addr, entry))
2452 		entry = pmd_mkwrite(entry, vma);
2453 
2454 	ret = HPAGE_PMD_NR;
2455 	set_pmd_at(mm, addr, pmd, entry);
2456 
2457 	if (huge_pmd_needs_flush(oldpmd, entry))
2458 		tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE);
2459 unlock:
2460 	spin_unlock(ptl);
2461 	return ret;
2462 }
2463 
2464 /*
2465  * Returns:
2466  *
2467  * - 0: if pud leaf changed from under us
2468  * - 1: if pud can be skipped
2469  * - HPAGE_PUD_NR: if pud was successfully processed
2470  */
2471 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2472 int change_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2473 		    pud_t *pudp, unsigned long addr, pgprot_t newprot,
2474 		    unsigned long cp_flags)
2475 {
2476 	struct mm_struct *mm = vma->vm_mm;
2477 	pud_t oldpud, entry;
2478 	spinlock_t *ptl;
2479 
2480 	tlb_change_page_size(tlb, HPAGE_PUD_SIZE);
2481 
2482 	/* NUMA balancing doesn't apply to dax */
2483 	if (cp_flags & MM_CP_PROT_NUMA)
2484 		return 1;
2485 
2486 	/*
2487 	 * Huge entries on userfault-wp only works with anonymous, while we
2488 	 * don't have anonymous PUDs yet.
2489 	 */
2490 	if (WARN_ON_ONCE(cp_flags & MM_CP_UFFD_WP_ALL))
2491 		return 1;
2492 
2493 	ptl = __pud_trans_huge_lock(pudp, vma);
2494 	if (!ptl)
2495 		return 0;
2496 
2497 	/*
2498 	 * Can't clear PUD or it can race with concurrent zapping.  See
2499 	 * change_huge_pmd().
2500 	 */
2501 	oldpud = pudp_invalidate(vma, addr, pudp);
2502 	entry = pud_modify(oldpud, newprot);
2503 	set_pud_at(mm, addr, pudp, entry);
2504 	tlb_flush_pud_range(tlb, addr, HPAGE_PUD_SIZE);
2505 
2506 	spin_unlock(ptl);
2507 	return HPAGE_PUD_NR;
2508 }
2509 #endif
2510 
2511 #ifdef CONFIG_USERFAULTFD
2512 /*
2513  * The PT lock for src_pmd and dst_vma/src_vma (for reading) are locked by
2514  * the caller, but it must return after releasing the page_table_lock.
2515  * Just move the page from src_pmd to dst_pmd if possible.
2516  * Return zero if succeeded in moving the page, -EAGAIN if it needs to be
2517  * repeated by the caller, or other errors in case of failure.
2518  */
2519 int move_pages_huge_pmd(struct mm_struct *mm, pmd_t *dst_pmd, pmd_t *src_pmd, pmd_t dst_pmdval,
2520 			struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
2521 			unsigned long dst_addr, unsigned long src_addr)
2522 {
2523 	pmd_t _dst_pmd, src_pmdval;
2524 	struct page *src_page;
2525 	struct folio *src_folio;
2526 	spinlock_t *src_ptl, *dst_ptl;
2527 	pgtable_t src_pgtable;
2528 	struct mmu_notifier_range range;
2529 	int err = 0;
2530 
2531 	src_pmdval = *src_pmd;
2532 	src_ptl = pmd_lockptr(mm, src_pmd);
2533 
2534 	lockdep_assert_held(src_ptl);
2535 	vma_assert_locked(src_vma);
2536 	vma_assert_locked(dst_vma);
2537 
2538 	/* Sanity checks before the operation */
2539 	if (WARN_ON_ONCE(!pmd_none(dst_pmdval)) || WARN_ON_ONCE(src_addr & ~HPAGE_PMD_MASK) ||
2540 	    WARN_ON_ONCE(dst_addr & ~HPAGE_PMD_MASK)) {
2541 		spin_unlock(src_ptl);
2542 		return -EINVAL;
2543 	}
2544 
2545 	if (!pmd_trans_huge(src_pmdval)) {
2546 		spin_unlock(src_ptl);
2547 		if (is_pmd_migration_entry(src_pmdval)) {
2548 			pmd_migration_entry_wait(mm, &src_pmdval);
2549 			return -EAGAIN;
2550 		}
2551 		return -ENOENT;
2552 	}
2553 
2554 	src_page = pmd_page(src_pmdval);
2555 
2556 	if (!is_huge_zero_pmd(src_pmdval)) {
2557 		if (unlikely(!PageAnonExclusive(src_page))) {
2558 			spin_unlock(src_ptl);
2559 			return -EBUSY;
2560 		}
2561 
2562 		src_folio = page_folio(src_page);
2563 		folio_get(src_folio);
2564 	} else
2565 		src_folio = NULL;
2566 
2567 	spin_unlock(src_ptl);
2568 
2569 	flush_cache_range(src_vma, src_addr, src_addr + HPAGE_PMD_SIZE);
2570 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, src_addr,
2571 				src_addr + HPAGE_PMD_SIZE);
2572 	mmu_notifier_invalidate_range_start(&range);
2573 
2574 	if (src_folio)
2575 		folio_lock(src_folio);
2576 
2577 	dst_ptl = pmd_lockptr(mm, dst_pmd);
2578 	double_pt_lock(src_ptl, dst_ptl);
2579 	if (unlikely(!pmd_same(*src_pmd, src_pmdval) ||
2580 		     !pmd_same(*dst_pmd, dst_pmdval))) {
2581 		err = -EAGAIN;
2582 		goto unlock_ptls;
2583 	}
2584 	if (src_folio) {
2585 		if (folio_maybe_dma_pinned(src_folio) ||
2586 		    !PageAnonExclusive(&src_folio->page)) {
2587 			err = -EBUSY;
2588 			goto unlock_ptls;
2589 		}
2590 
2591 		if (WARN_ON_ONCE(!folio_test_head(src_folio)) ||
2592 		    WARN_ON_ONCE(!folio_test_anon(src_folio))) {
2593 			err = -EBUSY;
2594 			goto unlock_ptls;
2595 		}
2596 
2597 		src_pmdval = pmdp_huge_clear_flush(src_vma, src_addr, src_pmd);
2598 		/* Folio got pinned from under us. Put it back and fail the move. */
2599 		if (folio_maybe_dma_pinned(src_folio)) {
2600 			set_pmd_at(mm, src_addr, src_pmd, src_pmdval);
2601 			err = -EBUSY;
2602 			goto unlock_ptls;
2603 		}
2604 
2605 		folio_move_anon_rmap(src_folio, dst_vma);
2606 		src_folio->index = linear_page_index(dst_vma, dst_addr);
2607 
2608 		_dst_pmd = folio_mk_pmd(src_folio, dst_vma->vm_page_prot);
2609 		/* Follow mremap() behavior and treat the entry dirty after the move */
2610 		_dst_pmd = pmd_mkwrite(pmd_mkdirty(_dst_pmd), dst_vma);
2611 	} else {
2612 		src_pmdval = pmdp_huge_clear_flush(src_vma, src_addr, src_pmd);
2613 		_dst_pmd = folio_mk_pmd(src_folio, dst_vma->vm_page_prot);
2614 	}
2615 	set_pmd_at(mm, dst_addr, dst_pmd, _dst_pmd);
2616 
2617 	src_pgtable = pgtable_trans_huge_withdraw(mm, src_pmd);
2618 	pgtable_trans_huge_deposit(mm, dst_pmd, src_pgtable);
2619 unlock_ptls:
2620 	double_pt_unlock(src_ptl, dst_ptl);
2621 	/* unblock rmap walks */
2622 	if (src_folio)
2623 		folio_unlock(src_folio);
2624 	mmu_notifier_invalidate_range_end(&range);
2625 	if (src_folio)
2626 		folio_put(src_folio);
2627 	return err;
2628 }
2629 #endif /* CONFIG_USERFAULTFD */
2630 
2631 /*
2632  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
2633  *
2634  * Note that if it returns page table lock pointer, this routine returns without
2635  * unlocking page table lock. So callers must unlock it.
2636  */
2637 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
2638 {
2639 	spinlock_t *ptl;
2640 	ptl = pmd_lock(vma->vm_mm, pmd);
2641 	if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd)))
2642 		return ptl;
2643 	spin_unlock(ptl);
2644 	return NULL;
2645 }
2646 
2647 /*
2648  * Returns page table lock pointer if a given pud maps a thp, NULL otherwise.
2649  *
2650  * Note that if it returns page table lock pointer, this routine returns without
2651  * unlocking page table lock. So callers must unlock it.
2652  */
2653 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2654 {
2655 	spinlock_t *ptl;
2656 
2657 	ptl = pud_lock(vma->vm_mm, pud);
2658 	if (likely(pud_trans_huge(*pud)))
2659 		return ptl;
2660 	spin_unlock(ptl);
2661 	return NULL;
2662 }
2663 
2664 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2665 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2666 		 pud_t *pud, unsigned long addr)
2667 {
2668 	spinlock_t *ptl;
2669 	pud_t orig_pud;
2670 
2671 	ptl = __pud_trans_huge_lock(pud, vma);
2672 	if (!ptl)
2673 		return 0;
2674 
2675 	orig_pud = pudp_huge_get_and_clear_full(vma, addr, pud, tlb->fullmm);
2676 	arch_check_zapped_pud(vma, orig_pud);
2677 	tlb_remove_pud_tlb_entry(tlb, pud, addr);
2678 	if (!vma_is_dax(vma) && vma_is_special_huge(vma)) {
2679 		spin_unlock(ptl);
2680 		/* No zero page support yet */
2681 	} else {
2682 		struct page *page = NULL;
2683 		struct folio *folio;
2684 
2685 		/* No support for anonymous PUD pages or migration yet */
2686 		VM_WARN_ON_ONCE(vma_is_anonymous(vma) ||
2687 				!pud_present(orig_pud));
2688 
2689 		page = pud_page(orig_pud);
2690 		folio = page_folio(page);
2691 		folio_remove_rmap_pud(folio, page, vma);
2692 		add_mm_counter(tlb->mm, mm_counter_file(folio), -HPAGE_PUD_NR);
2693 
2694 		spin_unlock(ptl);
2695 		tlb_remove_page_size(tlb, page, HPAGE_PUD_SIZE);
2696 	}
2697 	return 1;
2698 }
2699 
2700 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2701 		unsigned long haddr)
2702 {
2703 	struct folio *folio;
2704 	struct page *page;
2705 	pud_t old_pud;
2706 
2707 	VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2708 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2709 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2710 	VM_BUG_ON(!pud_trans_huge(*pud));
2711 
2712 	count_vm_event(THP_SPLIT_PUD);
2713 
2714 	old_pud = pudp_huge_clear_flush(vma, haddr, pud);
2715 
2716 	if (!vma_is_dax(vma))
2717 		return;
2718 
2719 	page = pud_page(old_pud);
2720 	folio = page_folio(page);
2721 
2722 	if (!folio_test_dirty(folio) && pud_dirty(old_pud))
2723 		folio_mark_dirty(folio);
2724 	if (!folio_test_referenced(folio) && pud_young(old_pud))
2725 		folio_set_referenced(folio);
2726 	folio_remove_rmap_pud(folio, page, vma);
2727 	folio_put(folio);
2728 	add_mm_counter(vma->vm_mm, mm_counter_file(folio),
2729 		-HPAGE_PUD_NR);
2730 }
2731 
2732 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2733 		unsigned long address)
2734 {
2735 	spinlock_t *ptl;
2736 	struct mmu_notifier_range range;
2737 
2738 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2739 				address & HPAGE_PUD_MASK,
2740 				(address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2741 	mmu_notifier_invalidate_range_start(&range);
2742 	ptl = pud_lock(vma->vm_mm, pud);
2743 	if (unlikely(!pud_trans_huge(*pud)))
2744 		goto out;
2745 	__split_huge_pud_locked(vma, pud, range.start);
2746 
2747 out:
2748 	spin_unlock(ptl);
2749 	mmu_notifier_invalidate_range_end(&range);
2750 }
2751 #else
2752 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2753 		unsigned long address)
2754 {
2755 }
2756 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2757 
2758 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2759 		unsigned long haddr, pmd_t *pmd)
2760 {
2761 	struct mm_struct *mm = vma->vm_mm;
2762 	pgtable_t pgtable;
2763 	pmd_t _pmd, old_pmd;
2764 	unsigned long addr;
2765 	pte_t *pte;
2766 	int i;
2767 
2768 	/*
2769 	 * Leave pmd empty until pte is filled note that it is fine to delay
2770 	 * notification until mmu_notifier_invalidate_range_end() as we are
2771 	 * replacing a zero pmd write protected page with a zero pte write
2772 	 * protected page.
2773 	 *
2774 	 * See Documentation/mm/mmu_notifier.rst
2775 	 */
2776 	old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd);
2777 
2778 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2779 	pmd_populate(mm, &_pmd, pgtable);
2780 
2781 	pte = pte_offset_map(&_pmd, haddr);
2782 	VM_BUG_ON(!pte);
2783 	for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2784 		pte_t entry;
2785 
2786 		entry = pfn_pte(my_zero_pfn(addr), vma->vm_page_prot);
2787 		entry = pte_mkspecial(entry);
2788 		if (pmd_uffd_wp(old_pmd))
2789 			entry = pte_mkuffd_wp(entry);
2790 		VM_BUG_ON(!pte_none(ptep_get(pte)));
2791 		set_pte_at(mm, addr, pte, entry);
2792 		pte++;
2793 	}
2794 	pte_unmap(pte - 1);
2795 	smp_wmb(); /* make pte visible before pmd */
2796 	pmd_populate(mm, pmd, pgtable);
2797 }
2798 
2799 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2800 		unsigned long haddr, bool freeze)
2801 {
2802 	struct mm_struct *mm = vma->vm_mm;
2803 	struct folio *folio;
2804 	struct page *page;
2805 	pgtable_t pgtable;
2806 	pmd_t old_pmd, _pmd;
2807 	bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2808 	bool anon_exclusive = false, dirty = false;
2809 	unsigned long addr;
2810 	pte_t *pte;
2811 	int i;
2812 
2813 	VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2814 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2815 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2816 	VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd));
2817 
2818 	count_vm_event(THP_SPLIT_PMD);
2819 
2820 	if (!vma_is_anonymous(vma)) {
2821 		old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd);
2822 		/*
2823 		 * We are going to unmap this huge page. So
2824 		 * just go ahead and zap it
2825 		 */
2826 		if (arch_needs_pgtable_deposit())
2827 			zap_deposited_table(mm, pmd);
2828 		if (!vma_is_dax(vma) && vma_is_special_huge(vma))
2829 			return;
2830 		if (unlikely(is_pmd_migration_entry(old_pmd))) {
2831 			swp_entry_t entry;
2832 
2833 			entry = pmd_to_swp_entry(old_pmd);
2834 			folio = pfn_swap_entry_folio(entry);
2835 		} else if (is_huge_zero_pmd(old_pmd)) {
2836 			return;
2837 		} else {
2838 			page = pmd_page(old_pmd);
2839 			folio = page_folio(page);
2840 			if (!folio_test_dirty(folio) && pmd_dirty(old_pmd))
2841 				folio_mark_dirty(folio);
2842 			if (!folio_test_referenced(folio) && pmd_young(old_pmd))
2843 				folio_set_referenced(folio);
2844 			folio_remove_rmap_pmd(folio, page, vma);
2845 			folio_put(folio);
2846 		}
2847 		add_mm_counter(mm, mm_counter_file(folio), -HPAGE_PMD_NR);
2848 		return;
2849 	}
2850 
2851 	if (is_huge_zero_pmd(*pmd)) {
2852 		/*
2853 		 * FIXME: Do we want to invalidate secondary mmu by calling
2854 		 * mmu_notifier_arch_invalidate_secondary_tlbs() see comments below
2855 		 * inside __split_huge_pmd() ?
2856 		 *
2857 		 * We are going from a zero huge page write protected to zero
2858 		 * small page also write protected so it does not seems useful
2859 		 * to invalidate secondary mmu at this time.
2860 		 */
2861 		return __split_huge_zero_page_pmd(vma, haddr, pmd);
2862 	}
2863 
2864 	pmd_migration = is_pmd_migration_entry(*pmd);
2865 	if (unlikely(pmd_migration)) {
2866 		swp_entry_t entry;
2867 
2868 		old_pmd = *pmd;
2869 		entry = pmd_to_swp_entry(old_pmd);
2870 		page = pfn_swap_entry_to_page(entry);
2871 		write = is_writable_migration_entry(entry);
2872 		if (PageAnon(page))
2873 			anon_exclusive = is_readable_exclusive_migration_entry(entry);
2874 		young = is_migration_entry_young(entry);
2875 		dirty = is_migration_entry_dirty(entry);
2876 		soft_dirty = pmd_swp_soft_dirty(old_pmd);
2877 		uffd_wp = pmd_swp_uffd_wp(old_pmd);
2878 	} else {
2879 		/*
2880 		 * Up to this point the pmd is present and huge and userland has
2881 		 * the whole access to the hugepage during the split (which
2882 		 * happens in place). If we overwrite the pmd with the not-huge
2883 		 * version pointing to the pte here (which of course we could if
2884 		 * all CPUs were bug free), userland could trigger a small page
2885 		 * size TLB miss on the small sized TLB while the hugepage TLB
2886 		 * entry is still established in the huge TLB. Some CPU doesn't
2887 		 * like that. See
2888 		 * http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2889 		 * 383 on page 105. Intel should be safe but is also warns that
2890 		 * it's only safe if the permission and cache attributes of the
2891 		 * two entries loaded in the two TLB is identical (which should
2892 		 * be the case here). But it is generally safer to never allow
2893 		 * small and huge TLB entries for the same virtual address to be
2894 		 * loaded simultaneously. So instead of doing "pmd_populate();
2895 		 * flush_pmd_tlb_range();" we first mark the current pmd
2896 		 * notpresent (atomically because here the pmd_trans_huge must
2897 		 * remain set at all times on the pmd until the split is
2898 		 * complete for this pmd), then we flush the SMP TLB and finally
2899 		 * we write the non-huge version of the pmd entry with
2900 		 * pmd_populate.
2901 		 */
2902 		old_pmd = pmdp_invalidate(vma, haddr, pmd);
2903 		page = pmd_page(old_pmd);
2904 		folio = page_folio(page);
2905 		if (pmd_dirty(old_pmd)) {
2906 			dirty = true;
2907 			folio_set_dirty(folio);
2908 		}
2909 		write = pmd_write(old_pmd);
2910 		young = pmd_young(old_pmd);
2911 		soft_dirty = pmd_soft_dirty(old_pmd);
2912 		uffd_wp = pmd_uffd_wp(old_pmd);
2913 
2914 		VM_WARN_ON_FOLIO(!folio_ref_count(folio), folio);
2915 		VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio);
2916 
2917 		/*
2918 		 * Without "freeze", we'll simply split the PMD, propagating the
2919 		 * PageAnonExclusive() flag for each PTE by setting it for
2920 		 * each subpage -- no need to (temporarily) clear.
2921 		 *
2922 		 * With "freeze" we want to replace mapped pages by
2923 		 * migration entries right away. This is only possible if we
2924 		 * managed to clear PageAnonExclusive() -- see
2925 		 * set_pmd_migration_entry().
2926 		 *
2927 		 * In case we cannot clear PageAnonExclusive(), split the PMD
2928 		 * only and let try_to_migrate_one() fail later.
2929 		 *
2930 		 * See folio_try_share_anon_rmap_pmd(): invalidate PMD first.
2931 		 */
2932 		anon_exclusive = PageAnonExclusive(page);
2933 		if (freeze && anon_exclusive &&
2934 		    folio_try_share_anon_rmap_pmd(folio, page))
2935 			freeze = false;
2936 		if (!freeze) {
2937 			rmap_t rmap_flags = RMAP_NONE;
2938 
2939 			folio_ref_add(folio, HPAGE_PMD_NR - 1);
2940 			if (anon_exclusive)
2941 				rmap_flags |= RMAP_EXCLUSIVE;
2942 			folio_add_anon_rmap_ptes(folio, page, HPAGE_PMD_NR,
2943 						 vma, haddr, rmap_flags);
2944 		}
2945 	}
2946 
2947 	/*
2948 	 * Withdraw the table only after we mark the pmd entry invalid.
2949 	 * This's critical for some architectures (Power).
2950 	 */
2951 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2952 	pmd_populate(mm, &_pmd, pgtable);
2953 
2954 	pte = pte_offset_map(&_pmd, haddr);
2955 	VM_BUG_ON(!pte);
2956 
2957 	/*
2958 	 * Note that NUMA hinting access restrictions are not transferred to
2959 	 * avoid any possibility of altering permissions across VMAs.
2960 	 */
2961 	if (freeze || pmd_migration) {
2962 		for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2963 			pte_t entry;
2964 			swp_entry_t swp_entry;
2965 
2966 			if (write)
2967 				swp_entry = make_writable_migration_entry(
2968 							page_to_pfn(page + i));
2969 			else if (anon_exclusive)
2970 				swp_entry = make_readable_exclusive_migration_entry(
2971 							page_to_pfn(page + i));
2972 			else
2973 				swp_entry = make_readable_migration_entry(
2974 							page_to_pfn(page + i));
2975 			if (young)
2976 				swp_entry = make_migration_entry_young(swp_entry);
2977 			if (dirty)
2978 				swp_entry = make_migration_entry_dirty(swp_entry);
2979 			entry = swp_entry_to_pte(swp_entry);
2980 			if (soft_dirty)
2981 				entry = pte_swp_mksoft_dirty(entry);
2982 			if (uffd_wp)
2983 				entry = pte_swp_mkuffd_wp(entry);
2984 
2985 			VM_WARN_ON(!pte_none(ptep_get(pte + i)));
2986 			set_pte_at(mm, addr, pte + i, entry);
2987 		}
2988 	} else {
2989 		pte_t entry;
2990 
2991 		entry = mk_pte(page, READ_ONCE(vma->vm_page_prot));
2992 		if (write)
2993 			entry = pte_mkwrite(entry, vma);
2994 		if (!young)
2995 			entry = pte_mkold(entry);
2996 		/* NOTE: this may set soft-dirty too on some archs */
2997 		if (dirty)
2998 			entry = pte_mkdirty(entry);
2999 		if (soft_dirty)
3000 			entry = pte_mksoft_dirty(entry);
3001 		if (uffd_wp)
3002 			entry = pte_mkuffd_wp(entry);
3003 
3004 		for (i = 0; i < HPAGE_PMD_NR; i++)
3005 			VM_WARN_ON(!pte_none(ptep_get(pte + i)));
3006 
3007 		set_ptes(mm, haddr, pte, entry, HPAGE_PMD_NR);
3008 	}
3009 	pte_unmap(pte);
3010 
3011 	if (!pmd_migration)
3012 		folio_remove_rmap_pmd(folio, page, vma);
3013 	if (freeze)
3014 		put_page(page);
3015 
3016 	smp_wmb(); /* make pte visible before pmd */
3017 	pmd_populate(mm, pmd, pgtable);
3018 }
3019 
3020 void split_huge_pmd_locked(struct vm_area_struct *vma, unsigned long address,
3021 			   pmd_t *pmd, bool freeze)
3022 {
3023 	VM_WARN_ON_ONCE(!IS_ALIGNED(address, HPAGE_PMD_SIZE));
3024 	if (pmd_trans_huge(*pmd) || is_pmd_migration_entry(*pmd))
3025 		__split_huge_pmd_locked(vma, pmd, address, freeze);
3026 }
3027 
3028 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
3029 		unsigned long address, bool freeze)
3030 {
3031 	spinlock_t *ptl;
3032 	struct mmu_notifier_range range;
3033 
3034 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
3035 				address & HPAGE_PMD_MASK,
3036 				(address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
3037 	mmu_notifier_invalidate_range_start(&range);
3038 	ptl = pmd_lock(vma->vm_mm, pmd);
3039 	split_huge_pmd_locked(vma, range.start, pmd, freeze);
3040 	spin_unlock(ptl);
3041 	mmu_notifier_invalidate_range_end(&range);
3042 }
3043 
3044 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
3045 		bool freeze)
3046 {
3047 	pmd_t *pmd = mm_find_pmd(vma->vm_mm, address);
3048 
3049 	if (!pmd)
3050 		return;
3051 
3052 	__split_huge_pmd(vma, pmd, address, freeze);
3053 }
3054 
3055 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
3056 {
3057 	/*
3058 	 * If the new address isn't hpage aligned and it could previously
3059 	 * contain an hugepage: check if we need to split an huge pmd.
3060 	 */
3061 	if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
3062 	    range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
3063 			 ALIGN(address, HPAGE_PMD_SIZE)))
3064 		split_huge_pmd_address(vma, address, false);
3065 }
3066 
3067 void vma_adjust_trans_huge(struct vm_area_struct *vma,
3068 			   unsigned long start,
3069 			   unsigned long end,
3070 			   struct vm_area_struct *next)
3071 {
3072 	/* Check if we need to split start first. */
3073 	split_huge_pmd_if_needed(vma, start);
3074 
3075 	/* Check if we need to split end next. */
3076 	split_huge_pmd_if_needed(vma, end);
3077 
3078 	/* If we're incrementing next->vm_start, we might need to split it. */
3079 	if (next)
3080 		split_huge_pmd_if_needed(next, end);
3081 }
3082 
3083 static void unmap_folio(struct folio *folio)
3084 {
3085 	enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SYNC |
3086 		TTU_BATCH_FLUSH;
3087 
3088 	VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
3089 
3090 	if (folio_test_pmd_mappable(folio))
3091 		ttu_flags |= TTU_SPLIT_HUGE_PMD;
3092 
3093 	/*
3094 	 * Anon pages need migration entries to preserve them, but file
3095 	 * pages can simply be left unmapped, then faulted back on demand.
3096 	 * If that is ever changed (perhaps for mlock), update remap_page().
3097 	 */
3098 	if (folio_test_anon(folio))
3099 		try_to_migrate(folio, ttu_flags);
3100 	else
3101 		try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK);
3102 
3103 	try_to_unmap_flush();
3104 }
3105 
3106 static bool __discard_anon_folio_pmd_locked(struct vm_area_struct *vma,
3107 					    unsigned long addr, pmd_t *pmdp,
3108 					    struct folio *folio)
3109 {
3110 	struct mm_struct *mm = vma->vm_mm;
3111 	int ref_count, map_count;
3112 	pmd_t orig_pmd = *pmdp;
3113 
3114 	if (pmd_dirty(orig_pmd))
3115 		folio_set_dirty(folio);
3116 	if (folio_test_dirty(folio) && !(vma->vm_flags & VM_DROPPABLE)) {
3117 		folio_set_swapbacked(folio);
3118 		return false;
3119 	}
3120 
3121 	orig_pmd = pmdp_huge_clear_flush(vma, addr, pmdp);
3122 
3123 	/*
3124 	 * Syncing against concurrent GUP-fast:
3125 	 * - clear PMD; barrier; read refcount
3126 	 * - inc refcount; barrier; read PMD
3127 	 */
3128 	smp_mb();
3129 
3130 	ref_count = folio_ref_count(folio);
3131 	map_count = folio_mapcount(folio);
3132 
3133 	/*
3134 	 * Order reads for folio refcount and dirty flag
3135 	 * (see comments in __remove_mapping()).
3136 	 */
3137 	smp_rmb();
3138 
3139 	/*
3140 	 * If the folio or its PMD is redirtied at this point, or if there
3141 	 * are unexpected references, we will give up to discard this folio
3142 	 * and remap it.
3143 	 *
3144 	 * The only folio refs must be one from isolation plus the rmap(s).
3145 	 */
3146 	if (pmd_dirty(orig_pmd))
3147 		folio_set_dirty(folio);
3148 	if (folio_test_dirty(folio) && !(vma->vm_flags & VM_DROPPABLE)) {
3149 		folio_set_swapbacked(folio);
3150 		set_pmd_at(mm, addr, pmdp, orig_pmd);
3151 		return false;
3152 	}
3153 
3154 	if (ref_count != map_count + 1) {
3155 		set_pmd_at(mm, addr, pmdp, orig_pmd);
3156 		return false;
3157 	}
3158 
3159 	folio_remove_rmap_pmd(folio, pmd_page(orig_pmd), vma);
3160 	zap_deposited_table(mm, pmdp);
3161 	add_mm_counter(mm, MM_ANONPAGES, -HPAGE_PMD_NR);
3162 	if (vma->vm_flags & VM_LOCKED)
3163 		mlock_drain_local();
3164 	folio_put(folio);
3165 
3166 	return true;
3167 }
3168 
3169 bool unmap_huge_pmd_locked(struct vm_area_struct *vma, unsigned long addr,
3170 			   pmd_t *pmdp, struct folio *folio)
3171 {
3172 	VM_WARN_ON_FOLIO(!folio_test_pmd_mappable(folio), folio);
3173 	VM_WARN_ON_FOLIO(!folio_test_locked(folio), folio);
3174 	VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio);
3175 	VM_WARN_ON_FOLIO(folio_test_swapbacked(folio), folio);
3176 	VM_WARN_ON_ONCE(!IS_ALIGNED(addr, HPAGE_PMD_SIZE));
3177 
3178 	return __discard_anon_folio_pmd_locked(vma, addr, pmdp, folio);
3179 }
3180 
3181 static void remap_page(struct folio *folio, unsigned long nr, int flags)
3182 {
3183 	int i = 0;
3184 
3185 	/* If unmap_folio() uses try_to_migrate() on file, remove this check */
3186 	if (!folio_test_anon(folio))
3187 		return;
3188 	for (;;) {
3189 		remove_migration_ptes(folio, folio, RMP_LOCKED | flags);
3190 		i += folio_nr_pages(folio);
3191 		if (i >= nr)
3192 			break;
3193 		folio = folio_next(folio);
3194 	}
3195 }
3196 
3197 static void lru_add_split_folio(struct folio *folio, struct folio *new_folio,
3198 		struct lruvec *lruvec, struct list_head *list)
3199 {
3200 	VM_BUG_ON_FOLIO(folio_test_lru(new_folio), folio);
3201 	lockdep_assert_held(&lruvec->lru_lock);
3202 
3203 	if (list) {
3204 		/* page reclaim is reclaiming a huge page */
3205 		VM_WARN_ON(folio_test_lru(folio));
3206 		folio_get(new_folio);
3207 		list_add_tail(&new_folio->lru, list);
3208 	} else {
3209 		/* head is still on lru (and we have it frozen) */
3210 		VM_WARN_ON(!folio_test_lru(folio));
3211 		if (folio_test_unevictable(folio))
3212 			new_folio->mlock_count = 0;
3213 		else
3214 			list_add_tail(&new_folio->lru, &folio->lru);
3215 		folio_set_lru(new_folio);
3216 	}
3217 }
3218 
3219 /* Racy check whether the huge page can be split */
3220 bool can_split_folio(struct folio *folio, int caller_pins, int *pextra_pins)
3221 {
3222 	int extra_pins;
3223 
3224 	/* Additional pins from page cache */
3225 	if (folio_test_anon(folio))
3226 		extra_pins = folio_test_swapcache(folio) ?
3227 				folio_nr_pages(folio) : 0;
3228 	else
3229 		extra_pins = folio_nr_pages(folio);
3230 	if (pextra_pins)
3231 		*pextra_pins = extra_pins;
3232 	return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins -
3233 					caller_pins;
3234 }
3235 
3236 static bool page_range_has_hwpoisoned(struct page *page, long nr_pages)
3237 {
3238 	for (; nr_pages; page++, nr_pages--)
3239 		if (PageHWPoison(page))
3240 			return true;
3241 	return false;
3242 }
3243 
3244 /*
3245  * It splits @folio into @new_order folios and copies the @folio metadata to
3246  * all the resulting folios.
3247  */
3248 static void __split_folio_to_order(struct folio *folio, int old_order,
3249 		int new_order)
3250 {
3251 	/* Scan poisoned pages when split a poisoned folio to large folios */
3252 	const bool handle_hwpoison = folio_test_has_hwpoisoned(folio) && new_order;
3253 	long new_nr_pages = 1 << new_order;
3254 	long nr_pages = 1 << old_order;
3255 	long i;
3256 
3257 	folio_clear_has_hwpoisoned(folio);
3258 
3259 	/* Check first new_nr_pages since the loop below skips them */
3260 	if (handle_hwpoison &&
3261 	    page_range_has_hwpoisoned(folio_page(folio, 0), new_nr_pages))
3262 		folio_set_has_hwpoisoned(folio);
3263 	/*
3264 	 * Skip the first new_nr_pages, since the new folio from them have all
3265 	 * the flags from the original folio.
3266 	 */
3267 	for (i = new_nr_pages; i < nr_pages; i += new_nr_pages) {
3268 		struct page *new_head = &folio->page + i;
3269 		/*
3270 		 * Careful: new_folio is not a "real" folio before we cleared PageTail.
3271 		 * Don't pass it around before clear_compound_head().
3272 		 */
3273 		struct folio *new_folio = (struct folio *)new_head;
3274 
3275 		VM_BUG_ON_PAGE(atomic_read(&new_folio->_mapcount) != -1, new_head);
3276 
3277 		/*
3278 		 * Clone page flags before unfreezing refcount.
3279 		 *
3280 		 * After successful get_page_unless_zero() might follow flags change,
3281 		 * for example lock_page() which set PG_waiters.
3282 		 *
3283 		 * Note that for mapped sub-pages of an anonymous THP,
3284 		 * PG_anon_exclusive has been cleared in unmap_folio() and is stored in
3285 		 * the migration entry instead from where remap_page() will restore it.
3286 		 * We can still have PG_anon_exclusive set on effectively unmapped and
3287 		 * unreferenced sub-pages of an anonymous THP: we can simply drop
3288 		 * PG_anon_exclusive (-> PG_mappedtodisk) for these here.
3289 		 */
3290 		new_folio->flags.f &= ~PAGE_FLAGS_CHECK_AT_PREP;
3291 		new_folio->flags.f |= (folio->flags.f &
3292 				((1L << PG_referenced) |
3293 				 (1L << PG_swapbacked) |
3294 				 (1L << PG_swapcache) |
3295 				 (1L << PG_mlocked) |
3296 				 (1L << PG_uptodate) |
3297 				 (1L << PG_active) |
3298 				 (1L << PG_workingset) |
3299 				 (1L << PG_locked) |
3300 				 (1L << PG_unevictable) |
3301 #ifdef CONFIG_ARCH_USES_PG_ARCH_2
3302 				 (1L << PG_arch_2) |
3303 #endif
3304 #ifdef CONFIG_ARCH_USES_PG_ARCH_3
3305 				 (1L << PG_arch_3) |
3306 #endif
3307 				 (1L << PG_dirty) |
3308 				 LRU_GEN_MASK | LRU_REFS_MASK));
3309 
3310 		if (handle_hwpoison &&
3311 		    page_range_has_hwpoisoned(new_head, new_nr_pages))
3312 			folio_set_has_hwpoisoned(new_folio);
3313 
3314 		new_folio->mapping = folio->mapping;
3315 		new_folio->index = folio->index + i;
3316 
3317 		/*
3318 		 * page->private should not be set in tail pages. Fix up and warn once
3319 		 * if private is unexpectedly set.
3320 		 */
3321 		if (unlikely(new_folio->private)) {
3322 			VM_WARN_ON_ONCE_PAGE(true, new_head);
3323 			new_folio->private = NULL;
3324 		}
3325 
3326 		if (folio_test_swapcache(folio))
3327 			new_folio->swap.val = folio->swap.val + i;
3328 
3329 		/* Page flags must be visible before we make the page non-compound. */
3330 		smp_wmb();
3331 
3332 		/*
3333 		 * Clear PageTail before unfreezing page refcount.
3334 		 *
3335 		 * After successful get_page_unless_zero() might follow put_page()
3336 		 * which needs correct compound_head().
3337 		 */
3338 		clear_compound_head(new_head);
3339 		if (new_order) {
3340 			prep_compound_page(new_head, new_order);
3341 			folio_set_large_rmappable(new_folio);
3342 		}
3343 
3344 		if (folio_test_young(folio))
3345 			folio_set_young(new_folio);
3346 		if (folio_test_idle(folio))
3347 			folio_set_idle(new_folio);
3348 #ifdef CONFIG_MEMCG
3349 		new_folio->memcg_data = folio->memcg_data;
3350 #endif
3351 
3352 		folio_xchg_last_cpupid(new_folio, folio_last_cpupid(folio));
3353 	}
3354 
3355 	if (new_order)
3356 		folio_set_order(folio, new_order);
3357 	else
3358 		ClearPageCompound(&folio->page);
3359 }
3360 
3361 /*
3362  * It splits an unmapped @folio to lower order smaller folios in two ways.
3363  * @folio: the to-be-split folio
3364  * @new_order: the smallest order of the after split folios (since buddy
3365  *             allocator like split generates folios with orders from @folio's
3366  *             order - 1 to new_order).
3367  * @split_at: in buddy allocator like split, the folio containing @split_at
3368  *            will be split until its order becomes @new_order.
3369  * @xas: xa_state pointing to folio->mapping->i_pages and locked by caller
3370  * @mapping: @folio->mapping
3371  * @uniform_split: if the split is uniform or not (buddy allocator like split)
3372  *
3373  *
3374  * 1. uniform split: the given @folio into multiple @new_order small folios,
3375  *    where all small folios have the same order. This is done when
3376  *    uniform_split is true.
3377  * 2. buddy allocator like (non-uniform) split: the given @folio is split into
3378  *    half and one of the half (containing the given page) is split into half
3379  *    until the given @page's order becomes @new_order. This is done when
3380  *    uniform_split is false.
3381  *
3382  * The high level flow for these two methods are:
3383  * 1. uniform split: a single __split_folio_to_order() is called to split the
3384  *    @folio into @new_order, then we traverse all the resulting folios one by
3385  *    one in PFN ascending order and perform stats, unfreeze, adding to list,
3386  *    and file mapping index operations.
3387  * 2. non-uniform split: in general, folio_order - @new_order calls to
3388  *    __split_folio_to_order() are made in a for loop to split the @folio
3389  *    to one lower order at a time. The resulting small folios are processed
3390  *    like what is done during the traversal in 1, except the one containing
3391  *    @page, which is split in next for loop.
3392  *
3393  * After splitting, the caller's folio reference will be transferred to the
3394  * folio containing @page. The caller needs to unlock and/or free after-split
3395  * folios if necessary.
3396  *
3397  * For !uniform_split, when -ENOMEM is returned, the original folio might be
3398  * split. The caller needs to check the input folio.
3399  */
3400 static int __split_unmapped_folio(struct folio *folio, int new_order,
3401 		struct page *split_at, struct xa_state *xas,
3402 		struct address_space *mapping, bool uniform_split)
3403 {
3404 	const bool is_anon = folio_test_anon(folio);
3405 	int old_order = folio_order(folio);
3406 	int start_order = uniform_split ? new_order : old_order - 1;
3407 	int split_order;
3408 
3409 	/*
3410 	 * split to new_order one order at a time. For uniform split,
3411 	 * folio is split to new_order directly.
3412 	 */
3413 	for (split_order = start_order;
3414 	     split_order >= new_order;
3415 	     split_order--) {
3416 		int nr_new_folios = 1UL << (old_order - split_order);
3417 
3418 		/* order-1 anonymous folio is not supported */
3419 		if (is_anon && split_order == 1)
3420 			continue;
3421 
3422 		if (mapping) {
3423 			/*
3424 			 * uniform split has xas_split_alloc() called before
3425 			 * irq is disabled to allocate enough memory, whereas
3426 			 * non-uniform split can handle ENOMEM.
3427 			 */
3428 			if (uniform_split)
3429 				xas_split(xas, folio, old_order);
3430 			else {
3431 				xas_set_order(xas, folio->index, split_order);
3432 				xas_try_split(xas, folio, old_order);
3433 				if (xas_error(xas))
3434 					return xas_error(xas);
3435 			}
3436 		}
3437 
3438 		folio_split_memcg_refs(folio, old_order, split_order);
3439 		split_page_owner(&folio->page, old_order, split_order);
3440 		pgalloc_tag_split(folio, old_order, split_order);
3441 		__split_folio_to_order(folio, old_order, split_order);
3442 
3443 		if (is_anon) {
3444 			mod_mthp_stat(old_order, MTHP_STAT_NR_ANON, -1);
3445 			mod_mthp_stat(split_order, MTHP_STAT_NR_ANON, nr_new_folios);
3446 		}
3447 		/*
3448 		 * If uniform split, the process is complete.
3449 		 * If non-uniform, continue splitting the folio at @split_at
3450 		 * as long as the next @split_order is >= @new_order.
3451 		 */
3452 		folio = page_folio(split_at);
3453 		old_order = split_order;
3454 	}
3455 
3456 	return 0;
3457 }
3458 
3459 bool non_uniform_split_supported(struct folio *folio, unsigned int new_order,
3460 		bool warns)
3461 {
3462 	if (folio_test_anon(folio)) {
3463 		/* order-1 is not supported for anonymous THP. */
3464 		VM_WARN_ONCE(warns && new_order == 1,
3465 				"Cannot split to order-1 folio");
3466 		if (new_order == 1)
3467 			return false;
3468 	} else if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&
3469 	    !mapping_large_folio_support(folio->mapping)) {
3470 		/*
3471 		 * No split if the file system does not support large folio.
3472 		 * Note that we might still have THPs in such mappings due to
3473 		 * CONFIG_READ_ONLY_THP_FOR_FS. But in that case, the mapping
3474 		 * does not actually support large folios properly.
3475 		 */
3476 		VM_WARN_ONCE(warns,
3477 			"Cannot split file folio to non-0 order");
3478 		return false;
3479 	}
3480 
3481 	/* Only swapping a whole PMD-mapped folio is supported */
3482 	if (folio_test_swapcache(folio)) {
3483 		VM_WARN_ONCE(warns,
3484 			"Cannot split swapcache folio to non-0 order");
3485 		return false;
3486 	}
3487 
3488 	return true;
3489 }
3490 
3491 /* See comments in non_uniform_split_supported() */
3492 bool uniform_split_supported(struct folio *folio, unsigned int new_order,
3493 		bool warns)
3494 {
3495 	if (folio_test_anon(folio)) {
3496 		VM_WARN_ONCE(warns && new_order == 1,
3497 				"Cannot split to order-1 folio");
3498 		if (new_order == 1)
3499 			return false;
3500 	} else  if (new_order) {
3501 		if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&
3502 		    !mapping_large_folio_support(folio->mapping)) {
3503 			VM_WARN_ONCE(warns,
3504 				"Cannot split file folio to non-0 order");
3505 			return false;
3506 		}
3507 	}
3508 
3509 	if (new_order && folio_test_swapcache(folio)) {
3510 		VM_WARN_ONCE(warns,
3511 			"Cannot split swapcache folio to non-0 order");
3512 		return false;
3513 	}
3514 
3515 	return true;
3516 }
3517 
3518 /*
3519  * __folio_split: split a folio at @split_at to a @new_order folio
3520  * @folio: folio to split
3521  * @new_order: the order of the new folio
3522  * @split_at: a page within the new folio
3523  * @lock_at: a page within @folio to be left locked to caller
3524  * @list: after-split folios will be put on it if non NULL
3525  * @uniform_split: perform uniform split or not (non-uniform split)
3526  *
3527  * It calls __split_unmapped_folio() to perform uniform and non-uniform split.
3528  * It is in charge of checking whether the split is supported or not and
3529  * preparing @folio for __split_unmapped_folio().
3530  *
3531  * After splitting, the after-split folio containing @lock_at remains locked
3532  * and others are unlocked:
3533  * 1. for uniform split, @lock_at points to one of @folio's subpages;
3534  * 2. for buddy allocator like (non-uniform) split, @lock_at points to @folio.
3535  *
3536  * return: 0: successful, <0 failed (if -ENOMEM is returned, @folio might be
3537  * split but not to @new_order, the caller needs to check)
3538  */
3539 static int __folio_split(struct folio *folio, unsigned int new_order,
3540 		struct page *split_at, struct page *lock_at,
3541 		struct list_head *list, bool uniform_split)
3542 {
3543 	struct deferred_split *ds_queue = get_deferred_split_queue(folio);
3544 	XA_STATE(xas, &folio->mapping->i_pages, folio->index);
3545 	struct folio *end_folio = folio_next(folio);
3546 	bool is_anon = folio_test_anon(folio);
3547 	struct address_space *mapping = NULL;
3548 	struct anon_vma *anon_vma = NULL;
3549 	int order = folio_order(folio);
3550 	struct folio *new_folio, *next;
3551 	int nr_shmem_dropped = 0;
3552 	int remap_flags = 0;
3553 	int extra_pins, ret;
3554 	pgoff_t end;
3555 	bool is_hzp;
3556 
3557 	VM_WARN_ON_ONCE_FOLIO(!folio_test_locked(folio), folio);
3558 	VM_WARN_ON_ONCE_FOLIO(!folio_test_large(folio), folio);
3559 
3560 	if (folio != page_folio(split_at) || folio != page_folio(lock_at))
3561 		return -EINVAL;
3562 
3563 	if (new_order >= folio_order(folio))
3564 		return -EINVAL;
3565 
3566 	if (uniform_split && !uniform_split_supported(folio, new_order, true))
3567 		return -EINVAL;
3568 
3569 	if (!uniform_split &&
3570 	    !non_uniform_split_supported(folio, new_order, true))
3571 		return -EINVAL;
3572 
3573 	is_hzp = is_huge_zero_folio(folio);
3574 	if (is_hzp) {
3575 		pr_warn_ratelimited("Called split_huge_page for huge zero page\n");
3576 		return -EBUSY;
3577 	}
3578 
3579 	if (folio_test_writeback(folio))
3580 		return -EBUSY;
3581 
3582 	if (is_anon) {
3583 		/*
3584 		 * The caller does not necessarily hold an mmap_lock that would
3585 		 * prevent the anon_vma disappearing so we first we take a
3586 		 * reference to it and then lock the anon_vma for write. This
3587 		 * is similar to folio_lock_anon_vma_read except the write lock
3588 		 * is taken to serialise against parallel split or collapse
3589 		 * operations.
3590 		 */
3591 		anon_vma = folio_get_anon_vma(folio);
3592 		if (!anon_vma) {
3593 			ret = -EBUSY;
3594 			goto out;
3595 		}
3596 		mapping = NULL;
3597 		anon_vma_lock_write(anon_vma);
3598 	} else {
3599 		unsigned int min_order;
3600 		gfp_t gfp;
3601 
3602 		mapping = folio->mapping;
3603 
3604 		/* Truncated ? */
3605 		/*
3606 		 * TODO: add support for large shmem folio in swap cache.
3607 		 * When shmem is in swap cache, mapping is NULL and
3608 		 * folio_test_swapcache() is true.
3609 		 */
3610 		if (!mapping) {
3611 			ret = -EBUSY;
3612 			goto out;
3613 		}
3614 
3615 		min_order = mapping_min_folio_order(folio->mapping);
3616 		if (new_order < min_order) {
3617 			ret = -EINVAL;
3618 			goto out;
3619 		}
3620 
3621 		gfp = current_gfp_context(mapping_gfp_mask(mapping) &
3622 							GFP_RECLAIM_MASK);
3623 
3624 		if (!filemap_release_folio(folio, gfp)) {
3625 			ret = -EBUSY;
3626 			goto out;
3627 		}
3628 
3629 		if (uniform_split) {
3630 			xas_set_order(&xas, folio->index, new_order);
3631 			xas_split_alloc(&xas, folio, folio_order(folio), gfp);
3632 			if (xas_error(&xas)) {
3633 				ret = xas_error(&xas);
3634 				goto out;
3635 			}
3636 		}
3637 
3638 		anon_vma = NULL;
3639 		i_mmap_lock_read(mapping);
3640 
3641 		/*
3642 		 *__split_unmapped_folio() may need to trim off pages beyond
3643 		 * EOF: but on 32-bit, i_size_read() takes an irq-unsafe
3644 		 * seqlock, which cannot be nested inside the page tree lock.
3645 		 * So note end now: i_size itself may be changed at any moment,
3646 		 * but folio lock is good enough to serialize the trimming.
3647 		 */
3648 		end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
3649 		if (shmem_mapping(mapping))
3650 			end = shmem_fallocend(mapping->host, end);
3651 	}
3652 
3653 	/*
3654 	 * Racy check if we can split the page, before unmap_folio() will
3655 	 * split PMDs
3656 	 */
3657 	if (!can_split_folio(folio, 1, &extra_pins)) {
3658 		ret = -EAGAIN;
3659 		goto out_unlock;
3660 	}
3661 
3662 	unmap_folio(folio);
3663 
3664 	/* block interrupt reentry in xa_lock and spinlock */
3665 	local_irq_disable();
3666 	if (mapping) {
3667 		/*
3668 		 * Check if the folio is present in page cache.
3669 		 * We assume all tail are present too, if folio is there.
3670 		 */
3671 		xas_lock(&xas);
3672 		xas_reset(&xas);
3673 		if (xas_load(&xas) != folio) {
3674 			ret = -EAGAIN;
3675 			goto fail;
3676 		}
3677 	}
3678 
3679 	/* Prevent deferred_split_scan() touching ->_refcount */
3680 	spin_lock(&ds_queue->split_queue_lock);
3681 	if (folio_ref_freeze(folio, 1 + extra_pins)) {
3682 		struct swap_cluster_info *ci = NULL;
3683 		struct lruvec *lruvec;
3684 		int expected_refs;
3685 
3686 		if (folio_order(folio) > 1 &&
3687 		    !list_empty(&folio->_deferred_list)) {
3688 			ds_queue->split_queue_len--;
3689 			if (folio_test_partially_mapped(folio)) {
3690 				folio_clear_partially_mapped(folio);
3691 				mod_mthp_stat(folio_order(folio),
3692 					      MTHP_STAT_NR_ANON_PARTIALLY_MAPPED, -1);
3693 			}
3694 			/*
3695 			 * Reinitialize page_deferred_list after removing the
3696 			 * page from the split_queue, otherwise a subsequent
3697 			 * split will see list corruption when checking the
3698 			 * page_deferred_list.
3699 			 */
3700 			list_del_init(&folio->_deferred_list);
3701 		}
3702 		spin_unlock(&ds_queue->split_queue_lock);
3703 		if (mapping) {
3704 			int nr = folio_nr_pages(folio);
3705 
3706 			if (folio_test_pmd_mappable(folio) &&
3707 			    new_order < HPAGE_PMD_ORDER) {
3708 				if (folio_test_swapbacked(folio)) {
3709 					__lruvec_stat_mod_folio(folio,
3710 							NR_SHMEM_THPS, -nr);
3711 				} else {
3712 					__lruvec_stat_mod_folio(folio,
3713 							NR_FILE_THPS, -nr);
3714 					filemap_nr_thps_dec(mapping);
3715 				}
3716 			}
3717 		}
3718 
3719 		if (folio_test_swapcache(folio)) {
3720 			if (mapping) {
3721 				VM_WARN_ON_ONCE_FOLIO(mapping, folio);
3722 				ret = -EINVAL;
3723 				goto fail;
3724 			}
3725 
3726 			ci = swap_cluster_get_and_lock(folio);
3727 		}
3728 
3729 		/* lock lru list/PageCompound, ref frozen by page_ref_freeze */
3730 		lruvec = folio_lruvec_lock(folio);
3731 
3732 		ret = __split_unmapped_folio(folio, new_order, split_at, &xas,
3733 					     mapping, uniform_split);
3734 
3735 		/*
3736 		 * Unfreeze after-split folios and put them back to the right
3737 		 * list. @folio should be kept frozon until page cache
3738 		 * entries are updated with all the other after-split folios
3739 		 * to prevent others seeing stale page cache entries.
3740 		 * As a result, new_folio starts from the next folio of
3741 		 * @folio.
3742 		 */
3743 		for (new_folio = folio_next(folio); new_folio != end_folio;
3744 		     new_folio = next) {
3745 			unsigned long nr_pages = folio_nr_pages(new_folio);
3746 
3747 			next = folio_next(new_folio);
3748 
3749 			expected_refs = folio_expected_ref_count(new_folio) + 1;
3750 			folio_ref_unfreeze(new_folio, expected_refs);
3751 
3752 			lru_add_split_folio(folio, new_folio, lruvec, list);
3753 
3754 			/*
3755 			 * Anonymous folio with swap cache.
3756 			 * NOTE: shmem in swap cache is not supported yet.
3757 			 */
3758 			if (ci) {
3759 				__swap_cache_replace_folio(ci, folio, new_folio);
3760 				continue;
3761 			}
3762 
3763 			/* Anonymous folio without swap cache */
3764 			if (!mapping)
3765 				continue;
3766 
3767 			/* Add the new folio to the page cache. */
3768 			if (new_folio->index < end) {
3769 				__xa_store(&mapping->i_pages, new_folio->index,
3770 					   new_folio, 0);
3771 				continue;
3772 			}
3773 
3774 			/* Drop folio beyond EOF: ->index >= end */
3775 			if (shmem_mapping(mapping))
3776 				nr_shmem_dropped += nr_pages;
3777 			else if (folio_test_clear_dirty(new_folio))
3778 				folio_account_cleaned(
3779 					new_folio, inode_to_wb(mapping->host));
3780 			__filemap_remove_folio(new_folio, NULL);
3781 			folio_put_refs(new_folio, nr_pages);
3782 		}
3783 		/*
3784 		 * Unfreeze @folio only after all page cache entries, which
3785 		 * used to point to it, have been updated with new folios.
3786 		 * Otherwise, a parallel folio_try_get() can grab @folio
3787 		 * and its caller can see stale page cache entries.
3788 		 */
3789 		expected_refs = folio_expected_ref_count(folio) + 1;
3790 		folio_ref_unfreeze(folio, expected_refs);
3791 
3792 		unlock_page_lruvec(lruvec);
3793 
3794 		if (ci)
3795 			swap_cluster_unlock(ci);
3796 	} else {
3797 		spin_unlock(&ds_queue->split_queue_lock);
3798 		ret = -EAGAIN;
3799 	}
3800 fail:
3801 	if (mapping)
3802 		xas_unlock(&xas);
3803 
3804 	local_irq_enable();
3805 
3806 	if (nr_shmem_dropped)
3807 		shmem_uncharge(mapping->host, nr_shmem_dropped);
3808 
3809 	if (!ret && is_anon)
3810 		remap_flags = RMP_USE_SHARED_ZEROPAGE;
3811 	remap_page(folio, 1 << order, remap_flags);
3812 
3813 	/*
3814 	 * Unlock all after-split folios except the one containing
3815 	 * @lock_at page. If @folio is not split, it will be kept locked.
3816 	 */
3817 	for (new_folio = folio; new_folio != end_folio; new_folio = next) {
3818 		next = folio_next(new_folio);
3819 		if (new_folio == page_folio(lock_at))
3820 			continue;
3821 
3822 		folio_unlock(new_folio);
3823 		/*
3824 		 * Subpages may be freed if there wasn't any mapping
3825 		 * like if add_to_swap() is running on a lru page that
3826 		 * had its mapping zapped. And freeing these pages
3827 		 * requires taking the lru_lock so we do the put_page
3828 		 * of the tail pages after the split is complete.
3829 		 */
3830 		free_folio_and_swap_cache(new_folio);
3831 	}
3832 
3833 out_unlock:
3834 	if (anon_vma) {
3835 		anon_vma_unlock_write(anon_vma);
3836 		put_anon_vma(anon_vma);
3837 	}
3838 	if (mapping)
3839 		i_mmap_unlock_read(mapping);
3840 out:
3841 	xas_destroy(&xas);
3842 	if (order == HPAGE_PMD_ORDER)
3843 		count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
3844 	count_mthp_stat(order, !ret ? MTHP_STAT_SPLIT : MTHP_STAT_SPLIT_FAILED);
3845 	return ret;
3846 }
3847 
3848 /*
3849  * This function splits a large folio into smaller folios of order @new_order.
3850  * @page can point to any page of the large folio to split. The split operation
3851  * does not change the position of @page.
3852  *
3853  * Prerequisites:
3854  *
3855  * 1) The caller must hold a reference on the @page's owning folio, also known
3856  *    as the large folio.
3857  *
3858  * 2) The large folio must be locked.
3859  *
3860  * 3) The folio must not be pinned. Any unexpected folio references, including
3861  *    GUP pins, will result in the folio not getting split; instead, the caller
3862  *    will receive an -EAGAIN.
3863  *
3864  * 4) @new_order > 1, usually. Splitting to order-1 anonymous folios is not
3865  *    supported for non-file-backed folios, because folio->_deferred_list, which
3866  *    is used by partially mapped folios, is stored in subpage 2, but an order-1
3867  *    folio only has subpages 0 and 1. File-backed order-1 folios are supported,
3868  *    since they do not use _deferred_list.
3869  *
3870  * After splitting, the caller's folio reference will be transferred to @page,
3871  * resulting in a raised refcount of @page after this call. The other pages may
3872  * be freed if they are not mapped.
3873  *
3874  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3875  *
3876  * Pages in @new_order will inherit the mapping, flags, and so on from the
3877  * huge page.
3878  *
3879  * Returns 0 if the huge page was split successfully.
3880  *
3881  * Returns -EAGAIN if the folio has unexpected reference (e.g., GUP) or if
3882  * the folio was concurrently removed from the page cache.
3883  *
3884  * Returns -EBUSY when trying to split the huge zeropage, if the folio is
3885  * under writeback, if fs-specific folio metadata cannot currently be
3886  * released, or if some unexpected race happened (e.g., anon VMA disappeared,
3887  * truncation).
3888  *
3889  * Callers should ensure that the order respects the address space mapping
3890  * min-order if one is set for non-anonymous folios.
3891  *
3892  * Returns -EINVAL when trying to split to an order that is incompatible
3893  * with the folio. Splitting to order 0 is compatible with all folios.
3894  */
3895 int split_huge_page_to_list_to_order(struct page *page, struct list_head *list,
3896 				     unsigned int new_order)
3897 {
3898 	struct folio *folio = page_folio(page);
3899 
3900 	return __folio_split(folio, new_order, &folio->page, page, list, true);
3901 }
3902 
3903 /*
3904  * folio_split: split a folio at @split_at to a @new_order folio
3905  * @folio: folio to split
3906  * @new_order: the order of the new folio
3907  * @split_at: a page within the new folio
3908  *
3909  * return: 0: successful, <0 failed (if -ENOMEM is returned, @folio might be
3910  * split but not to @new_order, the caller needs to check)
3911  *
3912  * It has the same prerequisites and returns as
3913  * split_huge_page_to_list_to_order().
3914  *
3915  * Split a folio at @split_at to a new_order folio, leave the
3916  * remaining subpages of the original folio as large as possible. For example,
3917  * in the case of splitting an order-9 folio at its third order-3 subpages to
3918  * an order-3 folio, there are 2^(9-3)=64 order-3 subpages in the order-9 folio.
3919  * After the split, there will be a group of folios with different orders and
3920  * the new folio containing @split_at is marked in bracket:
3921  * [order-4, {order-3}, order-3, order-5, order-6, order-7, order-8].
3922  *
3923  * After split, folio is left locked for caller.
3924  */
3925 int folio_split(struct folio *folio, unsigned int new_order,
3926 		struct page *split_at, struct list_head *list)
3927 {
3928 	return __folio_split(folio, new_order, split_at, &folio->page, list,
3929 			false);
3930 }
3931 
3932 int min_order_for_split(struct folio *folio)
3933 {
3934 	if (folio_test_anon(folio))
3935 		return 0;
3936 
3937 	if (!folio->mapping) {
3938 		if (folio_test_pmd_mappable(folio))
3939 			count_vm_event(THP_SPLIT_PAGE_FAILED);
3940 		return -EBUSY;
3941 	}
3942 
3943 	return mapping_min_folio_order(folio->mapping);
3944 }
3945 
3946 int split_folio_to_list(struct folio *folio, struct list_head *list)
3947 {
3948 	return split_huge_page_to_list_to_order(&folio->page, list, 0);
3949 }
3950 
3951 /*
3952  * __folio_unqueue_deferred_split() is not to be called directly:
3953  * the folio_unqueue_deferred_split() inline wrapper in mm/internal.h
3954  * limits its calls to those folios which may have a _deferred_list for
3955  * queueing THP splits, and that list is (racily observed to be) non-empty.
3956  *
3957  * It is unsafe to call folio_unqueue_deferred_split() until folio refcount is
3958  * zero: because even when split_queue_lock is held, a non-empty _deferred_list
3959  * might be in use on deferred_split_scan()'s unlocked on-stack list.
3960  *
3961  * If memory cgroups are enabled, split_queue_lock is in the mem_cgroup: it is
3962  * therefore important to unqueue deferred split before changing folio memcg.
3963  */
3964 bool __folio_unqueue_deferred_split(struct folio *folio)
3965 {
3966 	struct deferred_split *ds_queue;
3967 	unsigned long flags;
3968 	bool unqueued = false;
3969 
3970 	WARN_ON_ONCE(folio_ref_count(folio));
3971 	WARN_ON_ONCE(!mem_cgroup_disabled() && !folio_memcg(folio));
3972 
3973 	ds_queue = get_deferred_split_queue(folio);
3974 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
3975 	if (!list_empty(&folio->_deferred_list)) {
3976 		ds_queue->split_queue_len--;
3977 		if (folio_test_partially_mapped(folio)) {
3978 			folio_clear_partially_mapped(folio);
3979 			mod_mthp_stat(folio_order(folio),
3980 				      MTHP_STAT_NR_ANON_PARTIALLY_MAPPED, -1);
3981 		}
3982 		list_del_init(&folio->_deferred_list);
3983 		unqueued = true;
3984 	}
3985 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
3986 
3987 	return unqueued;	/* useful for debug warnings */
3988 }
3989 
3990 /* partially_mapped=false won't clear PG_partially_mapped folio flag */
3991 void deferred_split_folio(struct folio *folio, bool partially_mapped)
3992 {
3993 	struct deferred_split *ds_queue = get_deferred_split_queue(folio);
3994 #ifdef CONFIG_MEMCG
3995 	struct mem_cgroup *memcg = folio_memcg(folio);
3996 #endif
3997 	unsigned long flags;
3998 
3999 	/*
4000 	 * Order 1 folios have no space for a deferred list, but we also
4001 	 * won't waste much memory by not adding them to the deferred list.
4002 	 */
4003 	if (folio_order(folio) <= 1)
4004 		return;
4005 
4006 	if (!partially_mapped && !split_underused_thp)
4007 		return;
4008 
4009 	/*
4010 	 * Exclude swapcache: originally to avoid a corrupt deferred split
4011 	 * queue. Nowadays that is fully prevented by memcg1_swapout();
4012 	 * but if page reclaim is already handling the same folio, it is
4013 	 * unnecessary to handle it again in the shrinker, so excluding
4014 	 * swapcache here may still be a useful optimization.
4015 	 */
4016 	if (folio_test_swapcache(folio))
4017 		return;
4018 
4019 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
4020 	if (partially_mapped) {
4021 		if (!folio_test_partially_mapped(folio)) {
4022 			folio_set_partially_mapped(folio);
4023 			if (folio_test_pmd_mappable(folio))
4024 				count_vm_event(THP_DEFERRED_SPLIT_PAGE);
4025 			count_mthp_stat(folio_order(folio), MTHP_STAT_SPLIT_DEFERRED);
4026 			mod_mthp_stat(folio_order(folio), MTHP_STAT_NR_ANON_PARTIALLY_MAPPED, 1);
4027 
4028 		}
4029 	} else {
4030 		/* partially mapped folios cannot become non-partially mapped */
4031 		VM_WARN_ON_FOLIO(folio_test_partially_mapped(folio), folio);
4032 	}
4033 	if (list_empty(&folio->_deferred_list)) {
4034 		list_add_tail(&folio->_deferred_list, &ds_queue->split_queue);
4035 		ds_queue->split_queue_len++;
4036 #ifdef CONFIG_MEMCG
4037 		if (memcg)
4038 			set_shrinker_bit(memcg, folio_nid(folio),
4039 					 deferred_split_shrinker->id);
4040 #endif
4041 	}
4042 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
4043 }
4044 
4045 static unsigned long deferred_split_count(struct shrinker *shrink,
4046 		struct shrink_control *sc)
4047 {
4048 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
4049 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
4050 
4051 #ifdef CONFIG_MEMCG
4052 	if (sc->memcg)
4053 		ds_queue = &sc->memcg->deferred_split_queue;
4054 #endif
4055 	return READ_ONCE(ds_queue->split_queue_len);
4056 }
4057 
4058 static bool thp_underused(struct folio *folio)
4059 {
4060 	int num_zero_pages = 0, num_filled_pages = 0;
4061 	int i;
4062 
4063 	if (khugepaged_max_ptes_none == HPAGE_PMD_NR - 1)
4064 		return false;
4065 
4066 	if (folio_contain_hwpoisoned_page(folio))
4067 		return false;
4068 
4069 	for (i = 0; i < folio_nr_pages(folio); i++) {
4070 		if (pages_identical(folio_page(folio, i), ZERO_PAGE(0))) {
4071 			if (++num_zero_pages > khugepaged_max_ptes_none)
4072 				return true;
4073 		} else {
4074 			/*
4075 			 * Another path for early exit once the number
4076 			 * of non-zero filled pages exceeds threshold.
4077 			 */
4078 			if (++num_filled_pages >= HPAGE_PMD_NR - khugepaged_max_ptes_none)
4079 				return false;
4080 		}
4081 	}
4082 	return false;
4083 }
4084 
4085 static unsigned long deferred_split_scan(struct shrinker *shrink,
4086 		struct shrink_control *sc)
4087 {
4088 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
4089 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
4090 	unsigned long flags;
4091 	LIST_HEAD(list);
4092 	struct folio *folio, *next, *prev = NULL;
4093 	int split = 0, removed = 0;
4094 
4095 #ifdef CONFIG_MEMCG
4096 	if (sc->memcg)
4097 		ds_queue = &sc->memcg->deferred_split_queue;
4098 #endif
4099 
4100 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
4101 	/* Take pin on all head pages to avoid freeing them under us */
4102 	list_for_each_entry_safe(folio, next, &ds_queue->split_queue,
4103 							_deferred_list) {
4104 		if (folio_try_get(folio)) {
4105 			list_move(&folio->_deferred_list, &list);
4106 		} else {
4107 			/* We lost race with folio_put() */
4108 			if (folio_test_partially_mapped(folio)) {
4109 				folio_clear_partially_mapped(folio);
4110 				mod_mthp_stat(folio_order(folio),
4111 					      MTHP_STAT_NR_ANON_PARTIALLY_MAPPED, -1);
4112 			}
4113 			list_del_init(&folio->_deferred_list);
4114 			ds_queue->split_queue_len--;
4115 		}
4116 		if (!--sc->nr_to_scan)
4117 			break;
4118 	}
4119 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
4120 
4121 	list_for_each_entry_safe(folio, next, &list, _deferred_list) {
4122 		bool did_split = false;
4123 		bool underused = false;
4124 
4125 		if (!folio_test_partially_mapped(folio)) {
4126 			/*
4127 			 * See try_to_map_unused_to_zeropage(): we cannot
4128 			 * optimize zero-filled pages after splitting an
4129 			 * mlocked folio.
4130 			 */
4131 			if (folio_test_mlocked(folio))
4132 				goto next;
4133 			underused = thp_underused(folio);
4134 			if (!underused)
4135 				goto next;
4136 		}
4137 		if (!folio_trylock(folio))
4138 			goto next;
4139 		if (!split_folio(folio)) {
4140 			did_split = true;
4141 			if (underused)
4142 				count_vm_event(THP_UNDERUSED_SPLIT_PAGE);
4143 			split++;
4144 		}
4145 		folio_unlock(folio);
4146 next:
4147 		/*
4148 		 * split_folio() removes folio from list on success.
4149 		 * Only add back to the queue if folio is partially mapped.
4150 		 * If thp_underused returns false, or if split_folio fails
4151 		 * in the case it was underused, then consider it used and
4152 		 * don't add it back to split_queue.
4153 		 */
4154 		if (did_split) {
4155 			; /* folio already removed from list */
4156 		} else if (!folio_test_partially_mapped(folio)) {
4157 			list_del_init(&folio->_deferred_list);
4158 			removed++;
4159 		} else {
4160 			/*
4161 			 * That unlocked list_del_init() above would be unsafe,
4162 			 * unless its folio is separated from any earlier folios
4163 			 * left on the list (which may be concurrently unqueued)
4164 			 * by one safe folio with refcount still raised.
4165 			 */
4166 			swap(folio, prev);
4167 		}
4168 		if (folio)
4169 			folio_put(folio);
4170 	}
4171 
4172 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
4173 	list_splice_tail(&list, &ds_queue->split_queue);
4174 	ds_queue->split_queue_len -= removed;
4175 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
4176 
4177 	if (prev)
4178 		folio_put(prev);
4179 
4180 	/*
4181 	 * Stop shrinker if we didn't split any page, but the queue is empty.
4182 	 * This can happen if pages were freed under us.
4183 	 */
4184 	if (!split && list_empty(&ds_queue->split_queue))
4185 		return SHRINK_STOP;
4186 	return split;
4187 }
4188 
4189 #ifdef CONFIG_DEBUG_FS
4190 static void split_huge_pages_all(void)
4191 {
4192 	struct zone *zone;
4193 	struct page *page;
4194 	struct folio *folio;
4195 	unsigned long pfn, max_zone_pfn;
4196 	unsigned long total = 0, split = 0;
4197 
4198 	pr_debug("Split all THPs\n");
4199 	for_each_zone(zone) {
4200 		if (!managed_zone(zone))
4201 			continue;
4202 		max_zone_pfn = zone_end_pfn(zone);
4203 		for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
4204 			int nr_pages;
4205 
4206 			page = pfn_to_online_page(pfn);
4207 			if (!page || PageTail(page))
4208 				continue;
4209 			folio = page_folio(page);
4210 			if (!folio_try_get(folio))
4211 				continue;
4212 
4213 			if (unlikely(page_folio(page) != folio))
4214 				goto next;
4215 
4216 			if (zone != folio_zone(folio))
4217 				goto next;
4218 
4219 			if (!folio_test_large(folio)
4220 				|| folio_test_hugetlb(folio)
4221 				|| !folio_test_lru(folio))
4222 				goto next;
4223 
4224 			total++;
4225 			folio_lock(folio);
4226 			nr_pages = folio_nr_pages(folio);
4227 			if (!split_folio(folio))
4228 				split++;
4229 			pfn += nr_pages - 1;
4230 			folio_unlock(folio);
4231 next:
4232 			folio_put(folio);
4233 			cond_resched();
4234 		}
4235 	}
4236 
4237 	pr_debug("%lu of %lu THP split\n", split, total);
4238 }
4239 
4240 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
4241 {
4242 	return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
4243 		    is_vm_hugetlb_page(vma);
4244 }
4245 
4246 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
4247 				unsigned long vaddr_end, unsigned int new_order,
4248 				long in_folio_offset)
4249 {
4250 	int ret = 0;
4251 	struct task_struct *task;
4252 	struct mm_struct *mm;
4253 	unsigned long total = 0, split = 0;
4254 	unsigned long addr;
4255 
4256 	vaddr_start &= PAGE_MASK;
4257 	vaddr_end &= PAGE_MASK;
4258 
4259 	task = find_get_task_by_vpid(pid);
4260 	if (!task) {
4261 		ret = -ESRCH;
4262 		goto out;
4263 	}
4264 
4265 	/* Find the mm_struct */
4266 	mm = get_task_mm(task);
4267 	put_task_struct(task);
4268 
4269 	if (!mm) {
4270 		ret = -EINVAL;
4271 		goto out;
4272 	}
4273 
4274 	pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx], new_order: %u, in_folio_offset: %ld\n",
4275 		 pid, vaddr_start, vaddr_end, new_order, in_folio_offset);
4276 
4277 	mmap_read_lock(mm);
4278 	/*
4279 	 * always increase addr by PAGE_SIZE, since we could have a PTE page
4280 	 * table filled with PTE-mapped THPs, each of which is distinct.
4281 	 */
4282 	for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
4283 		struct vm_area_struct *vma = vma_lookup(mm, addr);
4284 		struct folio_walk fw;
4285 		struct folio *folio;
4286 		struct address_space *mapping;
4287 		unsigned int target_order = new_order;
4288 
4289 		if (!vma)
4290 			break;
4291 
4292 		/* skip special VMA and hugetlb VMA */
4293 		if (vma_not_suitable_for_thp_split(vma)) {
4294 			addr = vma->vm_end;
4295 			continue;
4296 		}
4297 
4298 		folio = folio_walk_start(&fw, vma, addr, 0);
4299 		if (!folio)
4300 			continue;
4301 
4302 		if (!is_transparent_hugepage(folio))
4303 			goto next;
4304 
4305 		if (!folio_test_anon(folio)) {
4306 			mapping = folio->mapping;
4307 			target_order = max(new_order,
4308 					   mapping_min_folio_order(mapping));
4309 		}
4310 
4311 		if (target_order >= folio_order(folio))
4312 			goto next;
4313 
4314 		total++;
4315 		/*
4316 		 * For folios with private, split_huge_page_to_list_to_order()
4317 		 * will try to drop it before split and then check if the folio
4318 		 * can be split or not. So skip the check here.
4319 		 */
4320 		if (!folio_test_private(folio) &&
4321 		    !can_split_folio(folio, 0, NULL))
4322 			goto next;
4323 
4324 		if (!folio_trylock(folio))
4325 			goto next;
4326 		folio_get(folio);
4327 		folio_walk_end(&fw, vma);
4328 
4329 		if (!folio_test_anon(folio) && folio->mapping != mapping)
4330 			goto unlock;
4331 
4332 		if (in_folio_offset < 0 ||
4333 		    in_folio_offset >= folio_nr_pages(folio)) {
4334 			if (!split_folio_to_order(folio, target_order))
4335 				split++;
4336 		} else {
4337 			struct page *split_at = folio_page(folio,
4338 							   in_folio_offset);
4339 			if (!folio_split(folio, target_order, split_at, NULL))
4340 				split++;
4341 		}
4342 
4343 unlock:
4344 
4345 		folio_unlock(folio);
4346 		folio_put(folio);
4347 
4348 		cond_resched();
4349 		continue;
4350 next:
4351 		folio_walk_end(&fw, vma);
4352 		cond_resched();
4353 	}
4354 	mmap_read_unlock(mm);
4355 	mmput(mm);
4356 
4357 	pr_debug("%lu of %lu THP split\n", split, total);
4358 
4359 out:
4360 	return ret;
4361 }
4362 
4363 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
4364 				pgoff_t off_end, unsigned int new_order,
4365 				long in_folio_offset)
4366 {
4367 	struct filename *file;
4368 	struct file *candidate;
4369 	struct address_space *mapping;
4370 	int ret = -EINVAL;
4371 	pgoff_t index;
4372 	int nr_pages = 1;
4373 	unsigned long total = 0, split = 0;
4374 	unsigned int min_order;
4375 	unsigned int target_order;
4376 
4377 	file = getname_kernel(file_path);
4378 	if (IS_ERR(file))
4379 		return ret;
4380 
4381 	candidate = file_open_name(file, O_RDONLY, 0);
4382 	if (IS_ERR(candidate))
4383 		goto out;
4384 
4385 	pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx], new_order: %u, in_folio_offset: %ld\n",
4386 		 file_path, off_start, off_end, new_order, in_folio_offset);
4387 
4388 	mapping = candidate->f_mapping;
4389 	min_order = mapping_min_folio_order(mapping);
4390 	target_order = max(new_order, min_order);
4391 
4392 	for (index = off_start; index < off_end; index += nr_pages) {
4393 		struct folio *folio = filemap_get_folio(mapping, index);
4394 
4395 		nr_pages = 1;
4396 		if (IS_ERR(folio))
4397 			continue;
4398 
4399 		if (!folio_test_large(folio))
4400 			goto next;
4401 
4402 		total++;
4403 		nr_pages = folio_nr_pages(folio);
4404 
4405 		if (target_order >= folio_order(folio))
4406 			goto next;
4407 
4408 		if (!folio_trylock(folio))
4409 			goto next;
4410 
4411 		if (folio->mapping != mapping)
4412 			goto unlock;
4413 
4414 		if (in_folio_offset < 0 || in_folio_offset >= nr_pages) {
4415 			if (!split_folio_to_order(folio, target_order))
4416 				split++;
4417 		} else {
4418 			struct page *split_at = folio_page(folio,
4419 							   in_folio_offset);
4420 			if (!folio_split(folio, target_order, split_at, NULL))
4421 				split++;
4422 		}
4423 
4424 unlock:
4425 		folio_unlock(folio);
4426 next:
4427 		folio_put(folio);
4428 		cond_resched();
4429 	}
4430 
4431 	filp_close(candidate, NULL);
4432 	ret = 0;
4433 
4434 	pr_debug("%lu of %lu file-backed THP split\n", split, total);
4435 out:
4436 	putname(file);
4437 	return ret;
4438 }
4439 
4440 #define MAX_INPUT_BUF_SZ 255
4441 
4442 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
4443 				size_t count, loff_t *ppops)
4444 {
4445 	static DEFINE_MUTEX(split_debug_mutex);
4446 	ssize_t ret;
4447 	/*
4448 	 * hold pid, start_vaddr, end_vaddr, new_order or
4449 	 * file_path, off_start, off_end, new_order
4450 	 */
4451 	char input_buf[MAX_INPUT_BUF_SZ];
4452 	int pid;
4453 	unsigned long vaddr_start, vaddr_end;
4454 	unsigned int new_order = 0;
4455 	long in_folio_offset = -1;
4456 
4457 	ret = mutex_lock_interruptible(&split_debug_mutex);
4458 	if (ret)
4459 		return ret;
4460 
4461 	ret = -EFAULT;
4462 
4463 	memset(input_buf, 0, MAX_INPUT_BUF_SZ);
4464 	if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
4465 		goto out;
4466 
4467 	input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
4468 
4469 	if (input_buf[0] == '/') {
4470 		char *tok;
4471 		char *tok_buf = input_buf;
4472 		char file_path[MAX_INPUT_BUF_SZ];
4473 		pgoff_t off_start = 0, off_end = 0;
4474 		size_t input_len = strlen(input_buf);
4475 
4476 		tok = strsep(&tok_buf, ",");
4477 		if (tok && tok_buf) {
4478 			strscpy(file_path, tok);
4479 		} else {
4480 			ret = -EINVAL;
4481 			goto out;
4482 		}
4483 
4484 		ret = sscanf(tok_buf, "0x%lx,0x%lx,%d,%ld", &off_start, &off_end,
4485 				&new_order, &in_folio_offset);
4486 		if (ret != 2 && ret != 3 && ret != 4) {
4487 			ret = -EINVAL;
4488 			goto out;
4489 		}
4490 		ret = split_huge_pages_in_file(file_path, off_start, off_end,
4491 				new_order, in_folio_offset);
4492 		if (!ret)
4493 			ret = input_len;
4494 
4495 		goto out;
4496 	}
4497 
4498 	ret = sscanf(input_buf, "%d,0x%lx,0x%lx,%d,%ld", &pid, &vaddr_start,
4499 			&vaddr_end, &new_order, &in_folio_offset);
4500 	if (ret == 1 && pid == 1) {
4501 		split_huge_pages_all();
4502 		ret = strlen(input_buf);
4503 		goto out;
4504 	} else if (ret != 3 && ret != 4 && ret != 5) {
4505 		ret = -EINVAL;
4506 		goto out;
4507 	}
4508 
4509 	ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end, new_order,
4510 			in_folio_offset);
4511 	if (!ret)
4512 		ret = strlen(input_buf);
4513 out:
4514 	mutex_unlock(&split_debug_mutex);
4515 	return ret;
4516 
4517 }
4518 
4519 static const struct file_operations split_huge_pages_fops = {
4520 	.owner	 = THIS_MODULE,
4521 	.write	 = split_huge_pages_write,
4522 };
4523 
4524 static int __init split_huge_pages_debugfs(void)
4525 {
4526 	debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
4527 			    &split_huge_pages_fops);
4528 	return 0;
4529 }
4530 late_initcall(split_huge_pages_debugfs);
4531 #endif
4532 
4533 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
4534 int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
4535 		struct page *page)
4536 {
4537 	struct folio *folio = page_folio(page);
4538 	struct vm_area_struct *vma = pvmw->vma;
4539 	struct mm_struct *mm = vma->vm_mm;
4540 	unsigned long address = pvmw->address;
4541 	bool anon_exclusive;
4542 	pmd_t pmdval;
4543 	swp_entry_t entry;
4544 	pmd_t pmdswp;
4545 
4546 	if (!(pvmw->pmd && !pvmw->pte))
4547 		return 0;
4548 
4549 	flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
4550 	pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
4551 
4552 	/* See folio_try_share_anon_rmap_pmd(): invalidate PMD first. */
4553 	anon_exclusive = folio_test_anon(folio) && PageAnonExclusive(page);
4554 	if (anon_exclusive && folio_try_share_anon_rmap_pmd(folio, page)) {
4555 		set_pmd_at(mm, address, pvmw->pmd, pmdval);
4556 		return -EBUSY;
4557 	}
4558 
4559 	if (pmd_dirty(pmdval))
4560 		folio_mark_dirty(folio);
4561 	if (pmd_write(pmdval))
4562 		entry = make_writable_migration_entry(page_to_pfn(page));
4563 	else if (anon_exclusive)
4564 		entry = make_readable_exclusive_migration_entry(page_to_pfn(page));
4565 	else
4566 		entry = make_readable_migration_entry(page_to_pfn(page));
4567 	if (pmd_young(pmdval))
4568 		entry = make_migration_entry_young(entry);
4569 	if (pmd_dirty(pmdval))
4570 		entry = make_migration_entry_dirty(entry);
4571 	pmdswp = swp_entry_to_pmd(entry);
4572 	if (pmd_soft_dirty(pmdval))
4573 		pmdswp = pmd_swp_mksoft_dirty(pmdswp);
4574 	if (pmd_uffd_wp(pmdval))
4575 		pmdswp = pmd_swp_mkuffd_wp(pmdswp);
4576 	set_pmd_at(mm, address, pvmw->pmd, pmdswp);
4577 	folio_remove_rmap_pmd(folio, page, vma);
4578 	folio_put(folio);
4579 	trace_set_migration_pmd(address, pmd_val(pmdswp));
4580 
4581 	return 0;
4582 }
4583 
4584 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
4585 {
4586 	struct folio *folio = page_folio(new);
4587 	struct vm_area_struct *vma = pvmw->vma;
4588 	struct mm_struct *mm = vma->vm_mm;
4589 	unsigned long address = pvmw->address;
4590 	unsigned long haddr = address & HPAGE_PMD_MASK;
4591 	pmd_t pmde;
4592 	swp_entry_t entry;
4593 
4594 	if (!(pvmw->pmd && !pvmw->pte))
4595 		return;
4596 
4597 	entry = pmd_to_swp_entry(*pvmw->pmd);
4598 	folio_get(folio);
4599 	pmde = folio_mk_pmd(folio, READ_ONCE(vma->vm_page_prot));
4600 	if (pmd_swp_soft_dirty(*pvmw->pmd))
4601 		pmde = pmd_mksoft_dirty(pmde);
4602 	if (is_writable_migration_entry(entry))
4603 		pmde = pmd_mkwrite(pmde, vma);
4604 	if (pmd_swp_uffd_wp(*pvmw->pmd))
4605 		pmde = pmd_mkuffd_wp(pmde);
4606 	if (!is_migration_entry_young(entry))
4607 		pmde = pmd_mkold(pmde);
4608 	/* NOTE: this may contain setting soft-dirty on some archs */
4609 	if (folio_test_dirty(folio) && is_migration_entry_dirty(entry))
4610 		pmde = pmd_mkdirty(pmde);
4611 
4612 	if (folio_test_anon(folio)) {
4613 		rmap_t rmap_flags = RMAP_NONE;
4614 
4615 		if (!is_readable_migration_entry(entry))
4616 			rmap_flags |= RMAP_EXCLUSIVE;
4617 
4618 		folio_add_anon_rmap_pmd(folio, new, vma, haddr, rmap_flags);
4619 	} else {
4620 		folio_add_file_rmap_pmd(folio, new, vma);
4621 	}
4622 	VM_BUG_ON(pmd_write(pmde) && folio_test_anon(folio) && !PageAnonExclusive(new));
4623 	set_pmd_at(mm, haddr, pvmw->pmd, pmde);
4624 
4625 	/* No need to invalidate - it was non-present before */
4626 	update_mmu_cache_pmd(vma, address, pvmw->pmd);
4627 	trace_remove_migration_pmd(address, pmd_val(pmde));
4628 }
4629 #endif
4630