xref: /linux/mm/huge_memory.c (revision f2527d8f566a45fa00ee5abd04d1c9476d4d704f)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  Copyright (C) 2009  Red Hat, Inc.
4  */
5 
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7 
8 #include <linux/mm.h>
9 #include <linux/sched.h>
10 #include <linux/sched/mm.h>
11 #include <linux/sched/coredump.h>
12 #include <linux/sched/numa_balancing.h>
13 #include <linux/highmem.h>
14 #include <linux/hugetlb.h>
15 #include <linux/mmu_notifier.h>
16 #include <linux/rmap.h>
17 #include <linux/swap.h>
18 #include <linux/shrinker.h>
19 #include <linux/mm_inline.h>
20 #include <linux/swapops.h>
21 #include <linux/backing-dev.h>
22 #include <linux/dax.h>
23 #include <linux/khugepaged.h>
24 #include <linux/freezer.h>
25 #include <linux/pfn_t.h>
26 #include <linux/mman.h>
27 #include <linux/memremap.h>
28 #include <linux/pagemap.h>
29 #include <linux/debugfs.h>
30 #include <linux/migrate.h>
31 #include <linux/hashtable.h>
32 #include <linux/userfaultfd_k.h>
33 #include <linux/page_idle.h>
34 #include <linux/shmem_fs.h>
35 #include <linux/oom.h>
36 #include <linux/numa.h>
37 #include <linux/page_owner.h>
38 #include <linux/sched/sysctl.h>
39 #include <linux/memory-tiers.h>
40 
41 #include <asm/tlb.h>
42 #include <asm/pgalloc.h>
43 #include "internal.h"
44 #include "swap.h"
45 
46 #define CREATE_TRACE_POINTS
47 #include <trace/events/thp.h>
48 
49 /*
50  * By default, transparent hugepage support is disabled in order to avoid
51  * risking an increased memory footprint for applications that are not
52  * guaranteed to benefit from it. When transparent hugepage support is
53  * enabled, it is for all mappings, and khugepaged scans all mappings.
54  * Defrag is invoked by khugepaged hugepage allocations and by page faults
55  * for all hugepage allocations.
56  */
57 unsigned long transparent_hugepage_flags __read_mostly =
58 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
59 	(1<<TRANSPARENT_HUGEPAGE_FLAG)|
60 #endif
61 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
62 	(1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
63 #endif
64 	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
65 	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
66 	(1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
67 
68 static struct shrinker deferred_split_shrinker;
69 
70 static atomic_t huge_zero_refcount;
71 struct page *huge_zero_page __read_mostly;
72 unsigned long huge_zero_pfn __read_mostly = ~0UL;
73 
74 bool hugepage_vma_check(struct vm_area_struct *vma, unsigned long vm_flags,
75 			bool smaps, bool in_pf, bool enforce_sysfs)
76 {
77 	if (!vma->vm_mm)		/* vdso */
78 		return false;
79 
80 	/*
81 	 * Explicitly disabled through madvise or prctl, or some
82 	 * architectures may disable THP for some mappings, for
83 	 * example, s390 kvm.
84 	 * */
85 	if ((vm_flags & VM_NOHUGEPAGE) ||
86 	    test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
87 		return false;
88 	/*
89 	 * If the hardware/firmware marked hugepage support disabled.
90 	 */
91 	if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_NEVER_DAX))
92 		return false;
93 
94 	/* khugepaged doesn't collapse DAX vma, but page fault is fine. */
95 	if (vma_is_dax(vma))
96 		return in_pf;
97 
98 	/*
99 	 * Special VMA and hugetlb VMA.
100 	 * Must be checked after dax since some dax mappings may have
101 	 * VM_MIXEDMAP set.
102 	 */
103 	if (vm_flags & VM_NO_KHUGEPAGED)
104 		return false;
105 
106 	/*
107 	 * Check alignment for file vma and size for both file and anon vma.
108 	 *
109 	 * Skip the check for page fault. Huge fault does the check in fault
110 	 * handlers. And this check is not suitable for huge PUD fault.
111 	 */
112 	if (!in_pf &&
113 	    !transhuge_vma_suitable(vma, (vma->vm_end - HPAGE_PMD_SIZE)))
114 		return false;
115 
116 	/*
117 	 * Enabled via shmem mount options or sysfs settings.
118 	 * Must be done before hugepage flags check since shmem has its
119 	 * own flags.
120 	 */
121 	if (!in_pf && shmem_file(vma->vm_file))
122 		return shmem_huge_enabled(vma, !enforce_sysfs);
123 
124 	/* Enforce sysfs THP requirements as necessary */
125 	if (enforce_sysfs &&
126 	    (!hugepage_flags_enabled() || (!(vm_flags & VM_HUGEPAGE) &&
127 					   !hugepage_flags_always())))
128 		return false;
129 
130 	/* Only regular file is valid */
131 	if (!in_pf && file_thp_enabled(vma))
132 		return true;
133 
134 	if (!vma_is_anonymous(vma))
135 		return false;
136 
137 	if (vma_is_temporary_stack(vma))
138 		return false;
139 
140 	/*
141 	 * THPeligible bit of smaps should show 1 for proper VMAs even
142 	 * though anon_vma is not initialized yet.
143 	 *
144 	 * Allow page fault since anon_vma may be not initialized until
145 	 * the first page fault.
146 	 */
147 	if (!vma->anon_vma)
148 		return (smaps || in_pf);
149 
150 	return true;
151 }
152 
153 static bool get_huge_zero_page(void)
154 {
155 	struct page *zero_page;
156 retry:
157 	if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
158 		return true;
159 
160 	zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
161 			HPAGE_PMD_ORDER);
162 	if (!zero_page) {
163 		count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
164 		return false;
165 	}
166 	preempt_disable();
167 	if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
168 		preempt_enable();
169 		__free_pages(zero_page, compound_order(zero_page));
170 		goto retry;
171 	}
172 	WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page));
173 
174 	/* We take additional reference here. It will be put back by shrinker */
175 	atomic_set(&huge_zero_refcount, 2);
176 	preempt_enable();
177 	count_vm_event(THP_ZERO_PAGE_ALLOC);
178 	return true;
179 }
180 
181 static void put_huge_zero_page(void)
182 {
183 	/*
184 	 * Counter should never go to zero here. Only shrinker can put
185 	 * last reference.
186 	 */
187 	BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
188 }
189 
190 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
191 {
192 	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
193 		return READ_ONCE(huge_zero_page);
194 
195 	if (!get_huge_zero_page())
196 		return NULL;
197 
198 	if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
199 		put_huge_zero_page();
200 
201 	return READ_ONCE(huge_zero_page);
202 }
203 
204 void mm_put_huge_zero_page(struct mm_struct *mm)
205 {
206 	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
207 		put_huge_zero_page();
208 }
209 
210 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
211 					struct shrink_control *sc)
212 {
213 	/* we can free zero page only if last reference remains */
214 	return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
215 }
216 
217 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
218 				       struct shrink_control *sc)
219 {
220 	if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
221 		struct page *zero_page = xchg(&huge_zero_page, NULL);
222 		BUG_ON(zero_page == NULL);
223 		WRITE_ONCE(huge_zero_pfn, ~0UL);
224 		__free_pages(zero_page, compound_order(zero_page));
225 		return HPAGE_PMD_NR;
226 	}
227 
228 	return 0;
229 }
230 
231 static struct shrinker huge_zero_page_shrinker = {
232 	.count_objects = shrink_huge_zero_page_count,
233 	.scan_objects = shrink_huge_zero_page_scan,
234 	.seeks = DEFAULT_SEEKS,
235 };
236 
237 #ifdef CONFIG_SYSFS
238 static ssize_t enabled_show(struct kobject *kobj,
239 			    struct kobj_attribute *attr, char *buf)
240 {
241 	const char *output;
242 
243 	if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
244 		output = "[always] madvise never";
245 	else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
246 			  &transparent_hugepage_flags))
247 		output = "always [madvise] never";
248 	else
249 		output = "always madvise [never]";
250 
251 	return sysfs_emit(buf, "%s\n", output);
252 }
253 
254 static ssize_t enabled_store(struct kobject *kobj,
255 			     struct kobj_attribute *attr,
256 			     const char *buf, size_t count)
257 {
258 	ssize_t ret = count;
259 
260 	if (sysfs_streq(buf, "always")) {
261 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
262 		set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
263 	} else if (sysfs_streq(buf, "madvise")) {
264 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
265 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
266 	} else if (sysfs_streq(buf, "never")) {
267 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
268 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
269 	} else
270 		ret = -EINVAL;
271 
272 	if (ret > 0) {
273 		int err = start_stop_khugepaged();
274 		if (err)
275 			ret = err;
276 	}
277 	return ret;
278 }
279 
280 static struct kobj_attribute enabled_attr = __ATTR_RW(enabled);
281 
282 ssize_t single_hugepage_flag_show(struct kobject *kobj,
283 				  struct kobj_attribute *attr, char *buf,
284 				  enum transparent_hugepage_flag flag)
285 {
286 	return sysfs_emit(buf, "%d\n",
287 			  !!test_bit(flag, &transparent_hugepage_flags));
288 }
289 
290 ssize_t single_hugepage_flag_store(struct kobject *kobj,
291 				 struct kobj_attribute *attr,
292 				 const char *buf, size_t count,
293 				 enum transparent_hugepage_flag flag)
294 {
295 	unsigned long value;
296 	int ret;
297 
298 	ret = kstrtoul(buf, 10, &value);
299 	if (ret < 0)
300 		return ret;
301 	if (value > 1)
302 		return -EINVAL;
303 
304 	if (value)
305 		set_bit(flag, &transparent_hugepage_flags);
306 	else
307 		clear_bit(flag, &transparent_hugepage_flags);
308 
309 	return count;
310 }
311 
312 static ssize_t defrag_show(struct kobject *kobj,
313 			   struct kobj_attribute *attr, char *buf)
314 {
315 	const char *output;
316 
317 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
318 		     &transparent_hugepage_flags))
319 		output = "[always] defer defer+madvise madvise never";
320 	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
321 			  &transparent_hugepage_flags))
322 		output = "always [defer] defer+madvise madvise never";
323 	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
324 			  &transparent_hugepage_flags))
325 		output = "always defer [defer+madvise] madvise never";
326 	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
327 			  &transparent_hugepage_flags))
328 		output = "always defer defer+madvise [madvise] never";
329 	else
330 		output = "always defer defer+madvise madvise [never]";
331 
332 	return sysfs_emit(buf, "%s\n", output);
333 }
334 
335 static ssize_t defrag_store(struct kobject *kobj,
336 			    struct kobj_attribute *attr,
337 			    const char *buf, size_t count)
338 {
339 	if (sysfs_streq(buf, "always")) {
340 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
341 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
342 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
343 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
344 	} else if (sysfs_streq(buf, "defer+madvise")) {
345 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
346 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
347 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
348 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
349 	} else if (sysfs_streq(buf, "defer")) {
350 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
351 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
352 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
353 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
354 	} else if (sysfs_streq(buf, "madvise")) {
355 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
356 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
357 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
358 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
359 	} else if (sysfs_streq(buf, "never")) {
360 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
361 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
362 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
363 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
364 	} else
365 		return -EINVAL;
366 
367 	return count;
368 }
369 static struct kobj_attribute defrag_attr = __ATTR_RW(defrag);
370 
371 static ssize_t use_zero_page_show(struct kobject *kobj,
372 				  struct kobj_attribute *attr, char *buf)
373 {
374 	return single_hugepage_flag_show(kobj, attr, buf,
375 					 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
376 }
377 static ssize_t use_zero_page_store(struct kobject *kobj,
378 		struct kobj_attribute *attr, const char *buf, size_t count)
379 {
380 	return single_hugepage_flag_store(kobj, attr, buf, count,
381 				 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
382 }
383 static struct kobj_attribute use_zero_page_attr = __ATTR_RW(use_zero_page);
384 
385 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
386 				   struct kobj_attribute *attr, char *buf)
387 {
388 	return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
389 }
390 static struct kobj_attribute hpage_pmd_size_attr =
391 	__ATTR_RO(hpage_pmd_size);
392 
393 static struct attribute *hugepage_attr[] = {
394 	&enabled_attr.attr,
395 	&defrag_attr.attr,
396 	&use_zero_page_attr.attr,
397 	&hpage_pmd_size_attr.attr,
398 #ifdef CONFIG_SHMEM
399 	&shmem_enabled_attr.attr,
400 #endif
401 	NULL,
402 };
403 
404 static const struct attribute_group hugepage_attr_group = {
405 	.attrs = hugepage_attr,
406 };
407 
408 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
409 {
410 	int err;
411 
412 	*hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
413 	if (unlikely(!*hugepage_kobj)) {
414 		pr_err("failed to create transparent hugepage kobject\n");
415 		return -ENOMEM;
416 	}
417 
418 	err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
419 	if (err) {
420 		pr_err("failed to register transparent hugepage group\n");
421 		goto delete_obj;
422 	}
423 
424 	err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
425 	if (err) {
426 		pr_err("failed to register transparent hugepage group\n");
427 		goto remove_hp_group;
428 	}
429 
430 	return 0;
431 
432 remove_hp_group:
433 	sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
434 delete_obj:
435 	kobject_put(*hugepage_kobj);
436 	return err;
437 }
438 
439 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
440 {
441 	sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
442 	sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
443 	kobject_put(hugepage_kobj);
444 }
445 #else
446 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
447 {
448 	return 0;
449 }
450 
451 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
452 {
453 }
454 #endif /* CONFIG_SYSFS */
455 
456 static int __init hugepage_init(void)
457 {
458 	int err;
459 	struct kobject *hugepage_kobj;
460 
461 	if (!has_transparent_hugepage()) {
462 		/*
463 		 * Hardware doesn't support hugepages, hence disable
464 		 * DAX PMD support.
465 		 */
466 		transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX;
467 		return -EINVAL;
468 	}
469 
470 	/*
471 	 * hugepages can't be allocated by the buddy allocator
472 	 */
473 	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
474 	/*
475 	 * we use page->mapping and page->index in second tail page
476 	 * as list_head: assuming THP order >= 2
477 	 */
478 	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
479 
480 	err = hugepage_init_sysfs(&hugepage_kobj);
481 	if (err)
482 		goto err_sysfs;
483 
484 	err = khugepaged_init();
485 	if (err)
486 		goto err_slab;
487 
488 	err = register_shrinker(&huge_zero_page_shrinker, "thp-zero");
489 	if (err)
490 		goto err_hzp_shrinker;
491 	err = register_shrinker(&deferred_split_shrinker, "thp-deferred_split");
492 	if (err)
493 		goto err_split_shrinker;
494 
495 	/*
496 	 * By default disable transparent hugepages on smaller systems,
497 	 * where the extra memory used could hurt more than TLB overhead
498 	 * is likely to save.  The admin can still enable it through /sys.
499 	 */
500 	if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
501 		transparent_hugepage_flags = 0;
502 		return 0;
503 	}
504 
505 	err = start_stop_khugepaged();
506 	if (err)
507 		goto err_khugepaged;
508 
509 	return 0;
510 err_khugepaged:
511 	unregister_shrinker(&deferred_split_shrinker);
512 err_split_shrinker:
513 	unregister_shrinker(&huge_zero_page_shrinker);
514 err_hzp_shrinker:
515 	khugepaged_destroy();
516 err_slab:
517 	hugepage_exit_sysfs(hugepage_kobj);
518 err_sysfs:
519 	return err;
520 }
521 subsys_initcall(hugepage_init);
522 
523 static int __init setup_transparent_hugepage(char *str)
524 {
525 	int ret = 0;
526 	if (!str)
527 		goto out;
528 	if (!strcmp(str, "always")) {
529 		set_bit(TRANSPARENT_HUGEPAGE_FLAG,
530 			&transparent_hugepage_flags);
531 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
532 			  &transparent_hugepage_flags);
533 		ret = 1;
534 	} else if (!strcmp(str, "madvise")) {
535 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
536 			  &transparent_hugepage_flags);
537 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
538 			&transparent_hugepage_flags);
539 		ret = 1;
540 	} else if (!strcmp(str, "never")) {
541 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
542 			  &transparent_hugepage_flags);
543 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
544 			  &transparent_hugepage_flags);
545 		ret = 1;
546 	}
547 out:
548 	if (!ret)
549 		pr_warn("transparent_hugepage= cannot parse, ignored\n");
550 	return ret;
551 }
552 __setup("transparent_hugepage=", setup_transparent_hugepage);
553 
554 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
555 {
556 	if (likely(vma->vm_flags & VM_WRITE))
557 		pmd = pmd_mkwrite(pmd);
558 	return pmd;
559 }
560 
561 #ifdef CONFIG_MEMCG
562 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
563 {
564 	struct mem_cgroup *memcg = page_memcg(compound_head(page));
565 	struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
566 
567 	if (memcg)
568 		return &memcg->deferred_split_queue;
569 	else
570 		return &pgdat->deferred_split_queue;
571 }
572 #else
573 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
574 {
575 	struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
576 
577 	return &pgdat->deferred_split_queue;
578 }
579 #endif
580 
581 void prep_transhuge_page(struct page *page)
582 {
583 	/*
584 	 * we use page->mapping and page->index in second tail page
585 	 * as list_head: assuming THP order >= 2
586 	 */
587 
588 	INIT_LIST_HEAD(page_deferred_list(page));
589 	set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
590 }
591 
592 static inline bool is_transparent_hugepage(struct page *page)
593 {
594 	if (!PageCompound(page))
595 		return false;
596 
597 	page = compound_head(page);
598 	return is_huge_zero_page(page) ||
599 	       page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
600 }
601 
602 static unsigned long __thp_get_unmapped_area(struct file *filp,
603 		unsigned long addr, unsigned long len,
604 		loff_t off, unsigned long flags, unsigned long size)
605 {
606 	loff_t off_end = off + len;
607 	loff_t off_align = round_up(off, size);
608 	unsigned long len_pad, ret;
609 
610 	if (off_end <= off_align || (off_end - off_align) < size)
611 		return 0;
612 
613 	len_pad = len + size;
614 	if (len_pad < len || (off + len_pad) < off)
615 		return 0;
616 
617 	ret = current->mm->get_unmapped_area(filp, addr, len_pad,
618 					      off >> PAGE_SHIFT, flags);
619 
620 	/*
621 	 * The failure might be due to length padding. The caller will retry
622 	 * without the padding.
623 	 */
624 	if (IS_ERR_VALUE(ret))
625 		return 0;
626 
627 	/*
628 	 * Do not try to align to THP boundary if allocation at the address
629 	 * hint succeeds.
630 	 */
631 	if (ret == addr)
632 		return addr;
633 
634 	ret += (off - ret) & (size - 1);
635 	return ret;
636 }
637 
638 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
639 		unsigned long len, unsigned long pgoff, unsigned long flags)
640 {
641 	unsigned long ret;
642 	loff_t off = (loff_t)pgoff << PAGE_SHIFT;
643 
644 	ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
645 	if (ret)
646 		return ret;
647 
648 	return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
649 }
650 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
651 
652 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
653 			struct page *page, gfp_t gfp)
654 {
655 	struct vm_area_struct *vma = vmf->vma;
656 	pgtable_t pgtable;
657 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
658 	vm_fault_t ret = 0;
659 
660 	VM_BUG_ON_PAGE(!PageCompound(page), page);
661 
662 	if (mem_cgroup_charge(page_folio(page), vma->vm_mm, gfp)) {
663 		put_page(page);
664 		count_vm_event(THP_FAULT_FALLBACK);
665 		count_vm_event(THP_FAULT_FALLBACK_CHARGE);
666 		return VM_FAULT_FALLBACK;
667 	}
668 	cgroup_throttle_swaprate(page, gfp);
669 
670 	pgtable = pte_alloc_one(vma->vm_mm);
671 	if (unlikely(!pgtable)) {
672 		ret = VM_FAULT_OOM;
673 		goto release;
674 	}
675 
676 	clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
677 	/*
678 	 * The memory barrier inside __SetPageUptodate makes sure that
679 	 * clear_huge_page writes become visible before the set_pmd_at()
680 	 * write.
681 	 */
682 	__SetPageUptodate(page);
683 
684 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
685 	if (unlikely(!pmd_none(*vmf->pmd))) {
686 		goto unlock_release;
687 	} else {
688 		pmd_t entry;
689 
690 		ret = check_stable_address_space(vma->vm_mm);
691 		if (ret)
692 			goto unlock_release;
693 
694 		/* Deliver the page fault to userland */
695 		if (userfaultfd_missing(vma)) {
696 			spin_unlock(vmf->ptl);
697 			put_page(page);
698 			pte_free(vma->vm_mm, pgtable);
699 			ret = handle_userfault(vmf, VM_UFFD_MISSING);
700 			VM_BUG_ON(ret & VM_FAULT_FALLBACK);
701 			return ret;
702 		}
703 
704 		entry = mk_huge_pmd(page, vma->vm_page_prot);
705 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
706 		page_add_new_anon_rmap(page, vma, haddr);
707 		lru_cache_add_inactive_or_unevictable(page, vma);
708 		pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
709 		set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
710 		update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
711 		add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
712 		mm_inc_nr_ptes(vma->vm_mm);
713 		spin_unlock(vmf->ptl);
714 		count_vm_event(THP_FAULT_ALLOC);
715 		count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
716 	}
717 
718 	return 0;
719 unlock_release:
720 	spin_unlock(vmf->ptl);
721 release:
722 	if (pgtable)
723 		pte_free(vma->vm_mm, pgtable);
724 	put_page(page);
725 	return ret;
726 
727 }
728 
729 /*
730  * always: directly stall for all thp allocations
731  * defer: wake kswapd and fail if not immediately available
732  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
733  *		  fail if not immediately available
734  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
735  *	    available
736  * never: never stall for any thp allocation
737  */
738 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
739 {
740 	const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
741 
742 	/* Always do synchronous compaction */
743 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
744 		return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
745 
746 	/* Kick kcompactd and fail quickly */
747 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
748 		return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
749 
750 	/* Synchronous compaction if madvised, otherwise kick kcompactd */
751 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
752 		return GFP_TRANSHUGE_LIGHT |
753 			(vma_madvised ? __GFP_DIRECT_RECLAIM :
754 					__GFP_KSWAPD_RECLAIM);
755 
756 	/* Only do synchronous compaction if madvised */
757 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
758 		return GFP_TRANSHUGE_LIGHT |
759 		       (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
760 
761 	return GFP_TRANSHUGE_LIGHT;
762 }
763 
764 /* Caller must hold page table lock. */
765 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
766 		struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
767 		struct page *zero_page)
768 {
769 	pmd_t entry;
770 	if (!pmd_none(*pmd))
771 		return;
772 	entry = mk_pmd(zero_page, vma->vm_page_prot);
773 	entry = pmd_mkhuge(entry);
774 	pgtable_trans_huge_deposit(mm, pmd, pgtable);
775 	set_pmd_at(mm, haddr, pmd, entry);
776 	mm_inc_nr_ptes(mm);
777 }
778 
779 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
780 {
781 	struct vm_area_struct *vma = vmf->vma;
782 	gfp_t gfp;
783 	struct folio *folio;
784 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
785 
786 	if (!transhuge_vma_suitable(vma, haddr))
787 		return VM_FAULT_FALLBACK;
788 	if (unlikely(anon_vma_prepare(vma)))
789 		return VM_FAULT_OOM;
790 	khugepaged_enter_vma(vma, vma->vm_flags);
791 
792 	if (!(vmf->flags & FAULT_FLAG_WRITE) &&
793 			!mm_forbids_zeropage(vma->vm_mm) &&
794 			transparent_hugepage_use_zero_page()) {
795 		pgtable_t pgtable;
796 		struct page *zero_page;
797 		vm_fault_t ret;
798 		pgtable = pte_alloc_one(vma->vm_mm);
799 		if (unlikely(!pgtable))
800 			return VM_FAULT_OOM;
801 		zero_page = mm_get_huge_zero_page(vma->vm_mm);
802 		if (unlikely(!zero_page)) {
803 			pte_free(vma->vm_mm, pgtable);
804 			count_vm_event(THP_FAULT_FALLBACK);
805 			return VM_FAULT_FALLBACK;
806 		}
807 		vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
808 		ret = 0;
809 		if (pmd_none(*vmf->pmd)) {
810 			ret = check_stable_address_space(vma->vm_mm);
811 			if (ret) {
812 				spin_unlock(vmf->ptl);
813 				pte_free(vma->vm_mm, pgtable);
814 			} else if (userfaultfd_missing(vma)) {
815 				spin_unlock(vmf->ptl);
816 				pte_free(vma->vm_mm, pgtable);
817 				ret = handle_userfault(vmf, VM_UFFD_MISSING);
818 				VM_BUG_ON(ret & VM_FAULT_FALLBACK);
819 			} else {
820 				set_huge_zero_page(pgtable, vma->vm_mm, vma,
821 						   haddr, vmf->pmd, zero_page);
822 				update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
823 				spin_unlock(vmf->ptl);
824 			}
825 		} else {
826 			spin_unlock(vmf->ptl);
827 			pte_free(vma->vm_mm, pgtable);
828 		}
829 		return ret;
830 	}
831 	gfp = vma_thp_gfp_mask(vma);
832 	folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, vma, haddr, true);
833 	if (unlikely(!folio)) {
834 		count_vm_event(THP_FAULT_FALLBACK);
835 		return VM_FAULT_FALLBACK;
836 	}
837 	return __do_huge_pmd_anonymous_page(vmf, &folio->page, gfp);
838 }
839 
840 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
841 		pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
842 		pgtable_t pgtable)
843 {
844 	struct mm_struct *mm = vma->vm_mm;
845 	pmd_t entry;
846 	spinlock_t *ptl;
847 
848 	ptl = pmd_lock(mm, pmd);
849 	if (!pmd_none(*pmd)) {
850 		if (write) {
851 			if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
852 				WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
853 				goto out_unlock;
854 			}
855 			entry = pmd_mkyoung(*pmd);
856 			entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
857 			if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
858 				update_mmu_cache_pmd(vma, addr, pmd);
859 		}
860 
861 		goto out_unlock;
862 	}
863 
864 	entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
865 	if (pfn_t_devmap(pfn))
866 		entry = pmd_mkdevmap(entry);
867 	if (write) {
868 		entry = pmd_mkyoung(pmd_mkdirty(entry));
869 		entry = maybe_pmd_mkwrite(entry, vma);
870 	}
871 
872 	if (pgtable) {
873 		pgtable_trans_huge_deposit(mm, pmd, pgtable);
874 		mm_inc_nr_ptes(mm);
875 		pgtable = NULL;
876 	}
877 
878 	set_pmd_at(mm, addr, pmd, entry);
879 	update_mmu_cache_pmd(vma, addr, pmd);
880 
881 out_unlock:
882 	spin_unlock(ptl);
883 	if (pgtable)
884 		pte_free(mm, pgtable);
885 }
886 
887 /**
888  * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
889  * @vmf: Structure describing the fault
890  * @pfn: pfn to insert
891  * @pgprot: page protection to use
892  * @write: whether it's a write fault
893  *
894  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
895  * also consult the vmf_insert_mixed_prot() documentation when
896  * @pgprot != @vmf->vma->vm_page_prot.
897  *
898  * Return: vm_fault_t value.
899  */
900 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
901 				   pgprot_t pgprot, bool write)
902 {
903 	unsigned long addr = vmf->address & PMD_MASK;
904 	struct vm_area_struct *vma = vmf->vma;
905 	pgtable_t pgtable = NULL;
906 
907 	/*
908 	 * If we had pmd_special, we could avoid all these restrictions,
909 	 * but we need to be consistent with PTEs and architectures that
910 	 * can't support a 'special' bit.
911 	 */
912 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
913 			!pfn_t_devmap(pfn));
914 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
915 						(VM_PFNMAP|VM_MIXEDMAP));
916 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
917 
918 	if (addr < vma->vm_start || addr >= vma->vm_end)
919 		return VM_FAULT_SIGBUS;
920 
921 	if (arch_needs_pgtable_deposit()) {
922 		pgtable = pte_alloc_one(vma->vm_mm);
923 		if (!pgtable)
924 			return VM_FAULT_OOM;
925 	}
926 
927 	track_pfn_insert(vma, &pgprot, pfn);
928 
929 	insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
930 	return VM_FAULT_NOPAGE;
931 }
932 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
933 
934 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
935 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
936 {
937 	if (likely(vma->vm_flags & VM_WRITE))
938 		pud = pud_mkwrite(pud);
939 	return pud;
940 }
941 
942 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
943 		pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
944 {
945 	struct mm_struct *mm = vma->vm_mm;
946 	pud_t entry;
947 	spinlock_t *ptl;
948 
949 	ptl = pud_lock(mm, pud);
950 	if (!pud_none(*pud)) {
951 		if (write) {
952 			if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
953 				WARN_ON_ONCE(!is_huge_zero_pud(*pud));
954 				goto out_unlock;
955 			}
956 			entry = pud_mkyoung(*pud);
957 			entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
958 			if (pudp_set_access_flags(vma, addr, pud, entry, 1))
959 				update_mmu_cache_pud(vma, addr, pud);
960 		}
961 		goto out_unlock;
962 	}
963 
964 	entry = pud_mkhuge(pfn_t_pud(pfn, prot));
965 	if (pfn_t_devmap(pfn))
966 		entry = pud_mkdevmap(entry);
967 	if (write) {
968 		entry = pud_mkyoung(pud_mkdirty(entry));
969 		entry = maybe_pud_mkwrite(entry, vma);
970 	}
971 	set_pud_at(mm, addr, pud, entry);
972 	update_mmu_cache_pud(vma, addr, pud);
973 
974 out_unlock:
975 	spin_unlock(ptl);
976 }
977 
978 /**
979  * vmf_insert_pfn_pud_prot - insert a pud size pfn
980  * @vmf: Structure describing the fault
981  * @pfn: pfn to insert
982  * @pgprot: page protection to use
983  * @write: whether it's a write fault
984  *
985  * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
986  * also consult the vmf_insert_mixed_prot() documentation when
987  * @pgprot != @vmf->vma->vm_page_prot.
988  *
989  * Return: vm_fault_t value.
990  */
991 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
992 				   pgprot_t pgprot, bool write)
993 {
994 	unsigned long addr = vmf->address & PUD_MASK;
995 	struct vm_area_struct *vma = vmf->vma;
996 
997 	/*
998 	 * If we had pud_special, we could avoid all these restrictions,
999 	 * but we need to be consistent with PTEs and architectures that
1000 	 * can't support a 'special' bit.
1001 	 */
1002 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
1003 			!pfn_t_devmap(pfn));
1004 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1005 						(VM_PFNMAP|VM_MIXEDMAP));
1006 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1007 
1008 	if (addr < vma->vm_start || addr >= vma->vm_end)
1009 		return VM_FAULT_SIGBUS;
1010 
1011 	track_pfn_insert(vma, &pgprot, pfn);
1012 
1013 	insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
1014 	return VM_FAULT_NOPAGE;
1015 }
1016 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
1017 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1018 
1019 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1020 		      pmd_t *pmd, bool write)
1021 {
1022 	pmd_t _pmd;
1023 
1024 	_pmd = pmd_mkyoung(*pmd);
1025 	if (write)
1026 		_pmd = pmd_mkdirty(_pmd);
1027 	if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1028 				  pmd, _pmd, write))
1029 		update_mmu_cache_pmd(vma, addr, pmd);
1030 }
1031 
1032 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1033 		pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
1034 {
1035 	unsigned long pfn = pmd_pfn(*pmd);
1036 	struct mm_struct *mm = vma->vm_mm;
1037 	struct page *page;
1038 	int ret;
1039 
1040 	assert_spin_locked(pmd_lockptr(mm, pmd));
1041 
1042 	/* FOLL_GET and FOLL_PIN are mutually exclusive. */
1043 	if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1044 			 (FOLL_PIN | FOLL_GET)))
1045 		return NULL;
1046 
1047 	if (flags & FOLL_WRITE && !pmd_write(*pmd))
1048 		return NULL;
1049 
1050 	if (pmd_present(*pmd) && pmd_devmap(*pmd))
1051 		/* pass */;
1052 	else
1053 		return NULL;
1054 
1055 	if (flags & FOLL_TOUCH)
1056 		touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1057 
1058 	/*
1059 	 * device mapped pages can only be returned if the
1060 	 * caller will manage the page reference count.
1061 	 */
1062 	if (!(flags & (FOLL_GET | FOLL_PIN)))
1063 		return ERR_PTR(-EEXIST);
1064 
1065 	pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1066 	*pgmap = get_dev_pagemap(pfn, *pgmap);
1067 	if (!*pgmap)
1068 		return ERR_PTR(-EFAULT);
1069 	page = pfn_to_page(pfn);
1070 	ret = try_grab_page(page, flags);
1071 	if (ret)
1072 		page = ERR_PTR(ret);
1073 
1074 	return page;
1075 }
1076 
1077 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1078 		  pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1079 		  struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1080 {
1081 	spinlock_t *dst_ptl, *src_ptl;
1082 	struct page *src_page;
1083 	pmd_t pmd;
1084 	pgtable_t pgtable = NULL;
1085 	int ret = -ENOMEM;
1086 
1087 	/* Skip if can be re-fill on fault */
1088 	if (!vma_is_anonymous(dst_vma))
1089 		return 0;
1090 
1091 	pgtable = pte_alloc_one(dst_mm);
1092 	if (unlikely(!pgtable))
1093 		goto out;
1094 
1095 	dst_ptl = pmd_lock(dst_mm, dst_pmd);
1096 	src_ptl = pmd_lockptr(src_mm, src_pmd);
1097 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1098 
1099 	ret = -EAGAIN;
1100 	pmd = *src_pmd;
1101 
1102 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1103 	if (unlikely(is_swap_pmd(pmd))) {
1104 		swp_entry_t entry = pmd_to_swp_entry(pmd);
1105 
1106 		VM_BUG_ON(!is_pmd_migration_entry(pmd));
1107 		if (!is_readable_migration_entry(entry)) {
1108 			entry = make_readable_migration_entry(
1109 							swp_offset(entry));
1110 			pmd = swp_entry_to_pmd(entry);
1111 			if (pmd_swp_soft_dirty(*src_pmd))
1112 				pmd = pmd_swp_mksoft_dirty(pmd);
1113 			if (pmd_swp_uffd_wp(*src_pmd))
1114 				pmd = pmd_swp_mkuffd_wp(pmd);
1115 			set_pmd_at(src_mm, addr, src_pmd, pmd);
1116 		}
1117 		add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1118 		mm_inc_nr_ptes(dst_mm);
1119 		pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1120 		if (!userfaultfd_wp(dst_vma))
1121 			pmd = pmd_swp_clear_uffd_wp(pmd);
1122 		set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1123 		ret = 0;
1124 		goto out_unlock;
1125 	}
1126 #endif
1127 
1128 	if (unlikely(!pmd_trans_huge(pmd))) {
1129 		pte_free(dst_mm, pgtable);
1130 		goto out_unlock;
1131 	}
1132 	/*
1133 	 * When page table lock is held, the huge zero pmd should not be
1134 	 * under splitting since we don't split the page itself, only pmd to
1135 	 * a page table.
1136 	 */
1137 	if (is_huge_zero_pmd(pmd)) {
1138 		/*
1139 		 * get_huge_zero_page() will never allocate a new page here,
1140 		 * since we already have a zero page to copy. It just takes a
1141 		 * reference.
1142 		 */
1143 		mm_get_huge_zero_page(dst_mm);
1144 		goto out_zero_page;
1145 	}
1146 
1147 	src_page = pmd_page(pmd);
1148 	VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1149 
1150 	get_page(src_page);
1151 	if (unlikely(page_try_dup_anon_rmap(src_page, true, src_vma))) {
1152 		/* Page maybe pinned: split and retry the fault on PTEs. */
1153 		put_page(src_page);
1154 		pte_free(dst_mm, pgtable);
1155 		spin_unlock(src_ptl);
1156 		spin_unlock(dst_ptl);
1157 		__split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1158 		return -EAGAIN;
1159 	}
1160 	add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1161 out_zero_page:
1162 	mm_inc_nr_ptes(dst_mm);
1163 	pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1164 	pmdp_set_wrprotect(src_mm, addr, src_pmd);
1165 	if (!userfaultfd_wp(dst_vma))
1166 		pmd = pmd_clear_uffd_wp(pmd);
1167 	pmd = pmd_mkold(pmd_wrprotect(pmd));
1168 	set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1169 
1170 	ret = 0;
1171 out_unlock:
1172 	spin_unlock(src_ptl);
1173 	spin_unlock(dst_ptl);
1174 out:
1175 	return ret;
1176 }
1177 
1178 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1179 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1180 		      pud_t *pud, bool write)
1181 {
1182 	pud_t _pud;
1183 
1184 	_pud = pud_mkyoung(*pud);
1185 	if (write)
1186 		_pud = pud_mkdirty(_pud);
1187 	if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1188 				  pud, _pud, write))
1189 		update_mmu_cache_pud(vma, addr, pud);
1190 }
1191 
1192 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1193 		pud_t *pud, int flags, struct dev_pagemap **pgmap)
1194 {
1195 	unsigned long pfn = pud_pfn(*pud);
1196 	struct mm_struct *mm = vma->vm_mm;
1197 	struct page *page;
1198 	int ret;
1199 
1200 	assert_spin_locked(pud_lockptr(mm, pud));
1201 
1202 	if (flags & FOLL_WRITE && !pud_write(*pud))
1203 		return NULL;
1204 
1205 	/* FOLL_GET and FOLL_PIN are mutually exclusive. */
1206 	if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1207 			 (FOLL_PIN | FOLL_GET)))
1208 		return NULL;
1209 
1210 	if (pud_present(*pud) && pud_devmap(*pud))
1211 		/* pass */;
1212 	else
1213 		return NULL;
1214 
1215 	if (flags & FOLL_TOUCH)
1216 		touch_pud(vma, addr, pud, flags & FOLL_WRITE);
1217 
1218 	/*
1219 	 * device mapped pages can only be returned if the
1220 	 * caller will manage the page reference count.
1221 	 *
1222 	 * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1223 	 */
1224 	if (!(flags & (FOLL_GET | FOLL_PIN)))
1225 		return ERR_PTR(-EEXIST);
1226 
1227 	pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1228 	*pgmap = get_dev_pagemap(pfn, *pgmap);
1229 	if (!*pgmap)
1230 		return ERR_PTR(-EFAULT);
1231 	page = pfn_to_page(pfn);
1232 
1233 	ret = try_grab_page(page, flags);
1234 	if (ret)
1235 		page = ERR_PTR(ret);
1236 
1237 	return page;
1238 }
1239 
1240 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1241 		  pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1242 		  struct vm_area_struct *vma)
1243 {
1244 	spinlock_t *dst_ptl, *src_ptl;
1245 	pud_t pud;
1246 	int ret;
1247 
1248 	dst_ptl = pud_lock(dst_mm, dst_pud);
1249 	src_ptl = pud_lockptr(src_mm, src_pud);
1250 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1251 
1252 	ret = -EAGAIN;
1253 	pud = *src_pud;
1254 	if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1255 		goto out_unlock;
1256 
1257 	/*
1258 	 * When page table lock is held, the huge zero pud should not be
1259 	 * under splitting since we don't split the page itself, only pud to
1260 	 * a page table.
1261 	 */
1262 	if (is_huge_zero_pud(pud)) {
1263 		/* No huge zero pud yet */
1264 	}
1265 
1266 	/*
1267 	 * TODO: once we support anonymous pages, use page_try_dup_anon_rmap()
1268 	 * and split if duplicating fails.
1269 	 */
1270 	pudp_set_wrprotect(src_mm, addr, src_pud);
1271 	pud = pud_mkold(pud_wrprotect(pud));
1272 	set_pud_at(dst_mm, addr, dst_pud, pud);
1273 
1274 	ret = 0;
1275 out_unlock:
1276 	spin_unlock(src_ptl);
1277 	spin_unlock(dst_ptl);
1278 	return ret;
1279 }
1280 
1281 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1282 {
1283 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1284 
1285 	vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1286 	if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1287 		goto unlock;
1288 
1289 	touch_pud(vmf->vma, vmf->address, vmf->pud, write);
1290 unlock:
1291 	spin_unlock(vmf->ptl);
1292 }
1293 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1294 
1295 void huge_pmd_set_accessed(struct vm_fault *vmf)
1296 {
1297 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1298 
1299 	vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1300 	if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd)))
1301 		goto unlock;
1302 
1303 	touch_pmd(vmf->vma, vmf->address, vmf->pmd, write);
1304 
1305 unlock:
1306 	spin_unlock(vmf->ptl);
1307 }
1308 
1309 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1310 {
1311 	const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
1312 	struct vm_area_struct *vma = vmf->vma;
1313 	struct folio *folio;
1314 	struct page *page;
1315 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1316 	pmd_t orig_pmd = vmf->orig_pmd;
1317 
1318 	vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1319 	VM_BUG_ON_VMA(!vma->anon_vma, vma);
1320 
1321 	if (is_huge_zero_pmd(orig_pmd))
1322 		goto fallback;
1323 
1324 	spin_lock(vmf->ptl);
1325 
1326 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1327 		spin_unlock(vmf->ptl);
1328 		return 0;
1329 	}
1330 
1331 	page = pmd_page(orig_pmd);
1332 	folio = page_folio(page);
1333 	VM_BUG_ON_PAGE(!PageHead(page), page);
1334 
1335 	/* Early check when only holding the PT lock. */
1336 	if (PageAnonExclusive(page))
1337 		goto reuse;
1338 
1339 	if (!folio_trylock(folio)) {
1340 		folio_get(folio);
1341 		spin_unlock(vmf->ptl);
1342 		folio_lock(folio);
1343 		spin_lock(vmf->ptl);
1344 		if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1345 			spin_unlock(vmf->ptl);
1346 			folio_unlock(folio);
1347 			folio_put(folio);
1348 			return 0;
1349 		}
1350 		folio_put(folio);
1351 	}
1352 
1353 	/* Recheck after temporarily dropping the PT lock. */
1354 	if (PageAnonExclusive(page)) {
1355 		folio_unlock(folio);
1356 		goto reuse;
1357 	}
1358 
1359 	/*
1360 	 * See do_wp_page(): we can only reuse the folio exclusively if
1361 	 * there are no additional references. Note that we always drain
1362 	 * the LRU pagevecs immediately after adding a THP.
1363 	 */
1364 	if (folio_ref_count(folio) >
1365 			1 + folio_test_swapcache(folio) * folio_nr_pages(folio))
1366 		goto unlock_fallback;
1367 	if (folio_test_swapcache(folio))
1368 		folio_free_swap(folio);
1369 	if (folio_ref_count(folio) == 1) {
1370 		pmd_t entry;
1371 
1372 		page_move_anon_rmap(page, vma);
1373 		folio_unlock(folio);
1374 reuse:
1375 		if (unlikely(unshare)) {
1376 			spin_unlock(vmf->ptl);
1377 			return 0;
1378 		}
1379 		entry = pmd_mkyoung(orig_pmd);
1380 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1381 		if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1382 			update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1383 		spin_unlock(vmf->ptl);
1384 		return 0;
1385 	}
1386 
1387 unlock_fallback:
1388 	folio_unlock(folio);
1389 	spin_unlock(vmf->ptl);
1390 fallback:
1391 	__split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1392 	return VM_FAULT_FALLBACK;
1393 }
1394 
1395 static inline bool can_change_pmd_writable(struct vm_area_struct *vma,
1396 					   unsigned long addr, pmd_t pmd)
1397 {
1398 	struct page *page;
1399 
1400 	if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE)))
1401 		return false;
1402 
1403 	/* Don't touch entries that are not even readable (NUMA hinting). */
1404 	if (pmd_protnone(pmd))
1405 		return false;
1406 
1407 	/* Do we need write faults for softdirty tracking? */
1408 	if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1409 		return false;
1410 
1411 	/* Do we need write faults for uffd-wp tracking? */
1412 	if (userfaultfd_huge_pmd_wp(vma, pmd))
1413 		return false;
1414 
1415 	if (!(vma->vm_flags & VM_SHARED)) {
1416 		/* See can_change_pte_writable(). */
1417 		page = vm_normal_page_pmd(vma, addr, pmd);
1418 		return page && PageAnon(page) && PageAnonExclusive(page);
1419 	}
1420 
1421 	/* See can_change_pte_writable(). */
1422 	return pmd_dirty(pmd);
1423 }
1424 
1425 /* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */
1426 static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page,
1427 					struct vm_area_struct *vma,
1428 					unsigned int flags)
1429 {
1430 	/* If the pmd is writable, we can write to the page. */
1431 	if (pmd_write(pmd))
1432 		return true;
1433 
1434 	/* Maybe FOLL_FORCE is set to override it? */
1435 	if (!(flags & FOLL_FORCE))
1436 		return false;
1437 
1438 	/* But FOLL_FORCE has no effect on shared mappings */
1439 	if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
1440 		return false;
1441 
1442 	/* ... or read-only private ones */
1443 	if (!(vma->vm_flags & VM_MAYWRITE))
1444 		return false;
1445 
1446 	/* ... or already writable ones that just need to take a write fault */
1447 	if (vma->vm_flags & VM_WRITE)
1448 		return false;
1449 
1450 	/*
1451 	 * See can_change_pte_writable(): we broke COW and could map the page
1452 	 * writable if we have an exclusive anonymous page ...
1453 	 */
1454 	if (!page || !PageAnon(page) || !PageAnonExclusive(page))
1455 		return false;
1456 
1457 	/* ... and a write-fault isn't required for other reasons. */
1458 	if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1459 		return false;
1460 	return !userfaultfd_huge_pmd_wp(vma, pmd);
1461 }
1462 
1463 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1464 				   unsigned long addr,
1465 				   pmd_t *pmd,
1466 				   unsigned int flags)
1467 {
1468 	struct mm_struct *mm = vma->vm_mm;
1469 	struct page *page;
1470 	int ret;
1471 
1472 	assert_spin_locked(pmd_lockptr(mm, pmd));
1473 
1474 	page = pmd_page(*pmd);
1475 	VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1476 
1477 	if ((flags & FOLL_WRITE) &&
1478 	    !can_follow_write_pmd(*pmd, page, vma, flags))
1479 		return NULL;
1480 
1481 	/* Avoid dumping huge zero page */
1482 	if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1483 		return ERR_PTR(-EFAULT);
1484 
1485 	/* Full NUMA hinting faults to serialise migration in fault paths */
1486 	if (pmd_protnone(*pmd) && !gup_can_follow_protnone(flags))
1487 		return NULL;
1488 
1489 	if (!pmd_write(*pmd) && gup_must_unshare(vma, flags, page))
1490 		return ERR_PTR(-EMLINK);
1491 
1492 	VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
1493 			!PageAnonExclusive(page), page);
1494 
1495 	ret = try_grab_page(page, flags);
1496 	if (ret)
1497 		return ERR_PTR(ret);
1498 
1499 	if (flags & FOLL_TOUCH)
1500 		touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1501 
1502 	page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1503 	VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1504 
1505 	return page;
1506 }
1507 
1508 /* NUMA hinting page fault entry point for trans huge pmds */
1509 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
1510 {
1511 	struct vm_area_struct *vma = vmf->vma;
1512 	pmd_t oldpmd = vmf->orig_pmd;
1513 	pmd_t pmd;
1514 	struct page *page;
1515 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1516 	int page_nid = NUMA_NO_NODE;
1517 	int target_nid, last_cpupid = (-1 & LAST_CPUPID_MASK);
1518 	bool migrated = false, writable = false;
1519 	int flags = 0;
1520 
1521 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1522 	if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1523 		spin_unlock(vmf->ptl);
1524 		goto out;
1525 	}
1526 
1527 	pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1528 
1529 	/*
1530 	 * Detect now whether the PMD could be writable; this information
1531 	 * is only valid while holding the PT lock.
1532 	 */
1533 	writable = pmd_write(pmd);
1534 	if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
1535 	    can_change_pmd_writable(vma, vmf->address, pmd))
1536 		writable = true;
1537 
1538 	page = vm_normal_page_pmd(vma, haddr, pmd);
1539 	if (!page)
1540 		goto out_map;
1541 
1542 	/* See similar comment in do_numa_page for explanation */
1543 	if (!writable)
1544 		flags |= TNF_NO_GROUP;
1545 
1546 	page_nid = page_to_nid(page);
1547 	/*
1548 	 * For memory tiering mode, cpupid of slow memory page is used
1549 	 * to record page access time.  So use default value.
1550 	 */
1551 	if (node_is_toptier(page_nid))
1552 		last_cpupid = page_cpupid_last(page);
1553 	target_nid = numa_migrate_prep(page, vma, haddr, page_nid,
1554 				       &flags);
1555 
1556 	if (target_nid == NUMA_NO_NODE) {
1557 		put_page(page);
1558 		goto out_map;
1559 	}
1560 
1561 	spin_unlock(vmf->ptl);
1562 	writable = false;
1563 
1564 	migrated = migrate_misplaced_page(page, vma, target_nid);
1565 	if (migrated) {
1566 		flags |= TNF_MIGRATED;
1567 		page_nid = target_nid;
1568 	} else {
1569 		flags |= TNF_MIGRATE_FAIL;
1570 		vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1571 		if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1572 			spin_unlock(vmf->ptl);
1573 			goto out;
1574 		}
1575 		goto out_map;
1576 	}
1577 
1578 out:
1579 	if (page_nid != NUMA_NO_NODE)
1580 		task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1581 				flags);
1582 
1583 	return 0;
1584 
1585 out_map:
1586 	/* Restore the PMD */
1587 	pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1588 	pmd = pmd_mkyoung(pmd);
1589 	if (writable)
1590 		pmd = pmd_mkwrite(pmd);
1591 	set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1592 	update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1593 	spin_unlock(vmf->ptl);
1594 	goto out;
1595 }
1596 
1597 /*
1598  * Return true if we do MADV_FREE successfully on entire pmd page.
1599  * Otherwise, return false.
1600  */
1601 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1602 		pmd_t *pmd, unsigned long addr, unsigned long next)
1603 {
1604 	spinlock_t *ptl;
1605 	pmd_t orig_pmd;
1606 	struct page *page;
1607 	struct mm_struct *mm = tlb->mm;
1608 	bool ret = false;
1609 
1610 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1611 
1612 	ptl = pmd_trans_huge_lock(pmd, vma);
1613 	if (!ptl)
1614 		goto out_unlocked;
1615 
1616 	orig_pmd = *pmd;
1617 	if (is_huge_zero_pmd(orig_pmd))
1618 		goto out;
1619 
1620 	if (unlikely(!pmd_present(orig_pmd))) {
1621 		VM_BUG_ON(thp_migration_supported() &&
1622 				  !is_pmd_migration_entry(orig_pmd));
1623 		goto out;
1624 	}
1625 
1626 	page = pmd_page(orig_pmd);
1627 	/*
1628 	 * If other processes are mapping this page, we couldn't discard
1629 	 * the page unless they all do MADV_FREE so let's skip the page.
1630 	 */
1631 	if (total_mapcount(page) != 1)
1632 		goto out;
1633 
1634 	if (!trylock_page(page))
1635 		goto out;
1636 
1637 	/*
1638 	 * If user want to discard part-pages of THP, split it so MADV_FREE
1639 	 * will deactivate only them.
1640 	 */
1641 	if (next - addr != HPAGE_PMD_SIZE) {
1642 		get_page(page);
1643 		spin_unlock(ptl);
1644 		split_huge_page(page);
1645 		unlock_page(page);
1646 		put_page(page);
1647 		goto out_unlocked;
1648 	}
1649 
1650 	if (PageDirty(page))
1651 		ClearPageDirty(page);
1652 	unlock_page(page);
1653 
1654 	if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1655 		pmdp_invalidate(vma, addr, pmd);
1656 		orig_pmd = pmd_mkold(orig_pmd);
1657 		orig_pmd = pmd_mkclean(orig_pmd);
1658 
1659 		set_pmd_at(mm, addr, pmd, orig_pmd);
1660 		tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1661 	}
1662 
1663 	mark_page_lazyfree(page);
1664 	ret = true;
1665 out:
1666 	spin_unlock(ptl);
1667 out_unlocked:
1668 	return ret;
1669 }
1670 
1671 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1672 {
1673 	pgtable_t pgtable;
1674 
1675 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1676 	pte_free(mm, pgtable);
1677 	mm_dec_nr_ptes(mm);
1678 }
1679 
1680 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1681 		 pmd_t *pmd, unsigned long addr)
1682 {
1683 	pmd_t orig_pmd;
1684 	spinlock_t *ptl;
1685 
1686 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1687 
1688 	ptl = __pmd_trans_huge_lock(pmd, vma);
1689 	if (!ptl)
1690 		return 0;
1691 	/*
1692 	 * For architectures like ppc64 we look at deposited pgtable
1693 	 * when calling pmdp_huge_get_and_clear. So do the
1694 	 * pgtable_trans_huge_withdraw after finishing pmdp related
1695 	 * operations.
1696 	 */
1697 	orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1698 						tlb->fullmm);
1699 	tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1700 	if (vma_is_special_huge(vma)) {
1701 		if (arch_needs_pgtable_deposit())
1702 			zap_deposited_table(tlb->mm, pmd);
1703 		spin_unlock(ptl);
1704 	} else if (is_huge_zero_pmd(orig_pmd)) {
1705 		zap_deposited_table(tlb->mm, pmd);
1706 		spin_unlock(ptl);
1707 	} else {
1708 		struct page *page = NULL;
1709 		int flush_needed = 1;
1710 
1711 		if (pmd_present(orig_pmd)) {
1712 			page = pmd_page(orig_pmd);
1713 			page_remove_rmap(page, vma, true);
1714 			VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1715 			VM_BUG_ON_PAGE(!PageHead(page), page);
1716 		} else if (thp_migration_supported()) {
1717 			swp_entry_t entry;
1718 
1719 			VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1720 			entry = pmd_to_swp_entry(orig_pmd);
1721 			page = pfn_swap_entry_to_page(entry);
1722 			flush_needed = 0;
1723 		} else
1724 			WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1725 
1726 		if (PageAnon(page)) {
1727 			zap_deposited_table(tlb->mm, pmd);
1728 			add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1729 		} else {
1730 			if (arch_needs_pgtable_deposit())
1731 				zap_deposited_table(tlb->mm, pmd);
1732 			add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1733 		}
1734 
1735 		spin_unlock(ptl);
1736 		if (flush_needed)
1737 			tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1738 	}
1739 	return 1;
1740 }
1741 
1742 #ifndef pmd_move_must_withdraw
1743 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1744 					 spinlock_t *old_pmd_ptl,
1745 					 struct vm_area_struct *vma)
1746 {
1747 	/*
1748 	 * With split pmd lock we also need to move preallocated
1749 	 * PTE page table if new_pmd is on different PMD page table.
1750 	 *
1751 	 * We also don't deposit and withdraw tables for file pages.
1752 	 */
1753 	return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1754 }
1755 #endif
1756 
1757 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1758 {
1759 #ifdef CONFIG_MEM_SOFT_DIRTY
1760 	if (unlikely(is_pmd_migration_entry(pmd)))
1761 		pmd = pmd_swp_mksoft_dirty(pmd);
1762 	else if (pmd_present(pmd))
1763 		pmd = pmd_mksoft_dirty(pmd);
1764 #endif
1765 	return pmd;
1766 }
1767 
1768 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1769 		  unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1770 {
1771 	spinlock_t *old_ptl, *new_ptl;
1772 	pmd_t pmd;
1773 	struct mm_struct *mm = vma->vm_mm;
1774 	bool force_flush = false;
1775 
1776 	/*
1777 	 * The destination pmd shouldn't be established, free_pgtables()
1778 	 * should have release it.
1779 	 */
1780 	if (WARN_ON(!pmd_none(*new_pmd))) {
1781 		VM_BUG_ON(pmd_trans_huge(*new_pmd));
1782 		return false;
1783 	}
1784 
1785 	/*
1786 	 * We don't have to worry about the ordering of src and dst
1787 	 * ptlocks because exclusive mmap_lock prevents deadlock.
1788 	 */
1789 	old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1790 	if (old_ptl) {
1791 		new_ptl = pmd_lockptr(mm, new_pmd);
1792 		if (new_ptl != old_ptl)
1793 			spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1794 		pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1795 		if (pmd_present(pmd))
1796 			force_flush = true;
1797 		VM_BUG_ON(!pmd_none(*new_pmd));
1798 
1799 		if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1800 			pgtable_t pgtable;
1801 			pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1802 			pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1803 		}
1804 		pmd = move_soft_dirty_pmd(pmd);
1805 		set_pmd_at(mm, new_addr, new_pmd, pmd);
1806 		if (force_flush)
1807 			flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1808 		if (new_ptl != old_ptl)
1809 			spin_unlock(new_ptl);
1810 		spin_unlock(old_ptl);
1811 		return true;
1812 	}
1813 	return false;
1814 }
1815 
1816 /*
1817  * Returns
1818  *  - 0 if PMD could not be locked
1819  *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1820  *      or if prot_numa but THP migration is not supported
1821  *  - HPAGE_PMD_NR if protections changed and TLB flush necessary
1822  */
1823 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1824 		    pmd_t *pmd, unsigned long addr, pgprot_t newprot,
1825 		    unsigned long cp_flags)
1826 {
1827 	struct mm_struct *mm = vma->vm_mm;
1828 	spinlock_t *ptl;
1829 	pmd_t oldpmd, entry;
1830 	bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1831 	bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1832 	bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1833 	int ret = 1;
1834 
1835 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1836 
1837 	if (prot_numa && !thp_migration_supported())
1838 		return 1;
1839 
1840 	ptl = __pmd_trans_huge_lock(pmd, vma);
1841 	if (!ptl)
1842 		return 0;
1843 
1844 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1845 	if (is_swap_pmd(*pmd)) {
1846 		swp_entry_t entry = pmd_to_swp_entry(*pmd);
1847 		struct page *page = pfn_swap_entry_to_page(entry);
1848 
1849 		VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1850 		if (is_writable_migration_entry(entry)) {
1851 			pmd_t newpmd;
1852 			/*
1853 			 * A protection check is difficult so
1854 			 * just be safe and disable write
1855 			 */
1856 			if (PageAnon(page))
1857 				entry = make_readable_exclusive_migration_entry(swp_offset(entry));
1858 			else
1859 				entry = make_readable_migration_entry(swp_offset(entry));
1860 			newpmd = swp_entry_to_pmd(entry);
1861 			if (pmd_swp_soft_dirty(*pmd))
1862 				newpmd = pmd_swp_mksoft_dirty(newpmd);
1863 			if (pmd_swp_uffd_wp(*pmd))
1864 				newpmd = pmd_swp_mkuffd_wp(newpmd);
1865 			set_pmd_at(mm, addr, pmd, newpmd);
1866 		}
1867 		goto unlock;
1868 	}
1869 #endif
1870 
1871 	if (prot_numa) {
1872 		struct page *page;
1873 		bool toptier;
1874 		/*
1875 		 * Avoid trapping faults against the zero page. The read-only
1876 		 * data is likely to be read-cached on the local CPU and
1877 		 * local/remote hits to the zero page are not interesting.
1878 		 */
1879 		if (is_huge_zero_pmd(*pmd))
1880 			goto unlock;
1881 
1882 		if (pmd_protnone(*pmd))
1883 			goto unlock;
1884 
1885 		page = pmd_page(*pmd);
1886 		toptier = node_is_toptier(page_to_nid(page));
1887 		/*
1888 		 * Skip scanning top tier node if normal numa
1889 		 * balancing is disabled
1890 		 */
1891 		if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
1892 		    toptier)
1893 			goto unlock;
1894 
1895 		if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
1896 		    !toptier)
1897 			xchg_page_access_time(page, jiffies_to_msecs(jiffies));
1898 	}
1899 	/*
1900 	 * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1901 	 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1902 	 * which is also under mmap_read_lock(mm):
1903 	 *
1904 	 *	CPU0:				CPU1:
1905 	 *				change_huge_pmd(prot_numa=1)
1906 	 *				 pmdp_huge_get_and_clear_notify()
1907 	 * madvise_dontneed()
1908 	 *  zap_pmd_range()
1909 	 *   pmd_trans_huge(*pmd) == 0 (without ptl)
1910 	 *   // skip the pmd
1911 	 *				 set_pmd_at();
1912 	 *				 // pmd is re-established
1913 	 *
1914 	 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1915 	 * which may break userspace.
1916 	 *
1917 	 * pmdp_invalidate_ad() is required to make sure we don't miss
1918 	 * dirty/young flags set by hardware.
1919 	 */
1920 	oldpmd = pmdp_invalidate_ad(vma, addr, pmd);
1921 
1922 	entry = pmd_modify(oldpmd, newprot);
1923 	if (uffd_wp) {
1924 		entry = pmd_wrprotect(entry);
1925 		entry = pmd_mkuffd_wp(entry);
1926 	} else if (uffd_wp_resolve) {
1927 		/*
1928 		 * Leave the write bit to be handled by PF interrupt
1929 		 * handler, then things like COW could be properly
1930 		 * handled.
1931 		 */
1932 		entry = pmd_clear_uffd_wp(entry);
1933 	}
1934 
1935 	/* See change_pte_range(). */
1936 	if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) && !pmd_write(entry) &&
1937 	    can_change_pmd_writable(vma, addr, entry))
1938 		entry = pmd_mkwrite(entry);
1939 
1940 	ret = HPAGE_PMD_NR;
1941 	set_pmd_at(mm, addr, pmd, entry);
1942 
1943 	if (huge_pmd_needs_flush(oldpmd, entry))
1944 		tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE);
1945 unlock:
1946 	spin_unlock(ptl);
1947 	return ret;
1948 }
1949 
1950 /*
1951  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1952  *
1953  * Note that if it returns page table lock pointer, this routine returns without
1954  * unlocking page table lock. So callers must unlock it.
1955  */
1956 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1957 {
1958 	spinlock_t *ptl;
1959 	ptl = pmd_lock(vma->vm_mm, pmd);
1960 	if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1961 			pmd_devmap(*pmd)))
1962 		return ptl;
1963 	spin_unlock(ptl);
1964 	return NULL;
1965 }
1966 
1967 /*
1968  * Returns page table lock pointer if a given pud maps a thp, NULL otherwise.
1969  *
1970  * Note that if it returns page table lock pointer, this routine returns without
1971  * unlocking page table lock. So callers must unlock it.
1972  */
1973 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1974 {
1975 	spinlock_t *ptl;
1976 
1977 	ptl = pud_lock(vma->vm_mm, pud);
1978 	if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1979 		return ptl;
1980 	spin_unlock(ptl);
1981 	return NULL;
1982 }
1983 
1984 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1985 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1986 		 pud_t *pud, unsigned long addr)
1987 {
1988 	spinlock_t *ptl;
1989 
1990 	ptl = __pud_trans_huge_lock(pud, vma);
1991 	if (!ptl)
1992 		return 0;
1993 
1994 	pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1995 	tlb_remove_pud_tlb_entry(tlb, pud, addr);
1996 	if (vma_is_special_huge(vma)) {
1997 		spin_unlock(ptl);
1998 		/* No zero page support yet */
1999 	} else {
2000 		/* No support for anonymous PUD pages yet */
2001 		BUG();
2002 	}
2003 	return 1;
2004 }
2005 
2006 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2007 		unsigned long haddr)
2008 {
2009 	VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2010 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2011 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2012 	VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2013 
2014 	count_vm_event(THP_SPLIT_PUD);
2015 
2016 	pudp_huge_clear_flush_notify(vma, haddr, pud);
2017 }
2018 
2019 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2020 		unsigned long address)
2021 {
2022 	spinlock_t *ptl;
2023 	struct mmu_notifier_range range;
2024 
2025 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2026 				address & HPAGE_PUD_MASK,
2027 				(address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2028 	mmu_notifier_invalidate_range_start(&range);
2029 	ptl = pud_lock(vma->vm_mm, pud);
2030 	if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2031 		goto out;
2032 	__split_huge_pud_locked(vma, pud, range.start);
2033 
2034 out:
2035 	spin_unlock(ptl);
2036 	/*
2037 	 * No need to double call mmu_notifier->invalidate_range() callback as
2038 	 * the above pudp_huge_clear_flush_notify() did already call it.
2039 	 */
2040 	mmu_notifier_invalidate_range_only_end(&range);
2041 }
2042 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2043 
2044 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2045 		unsigned long haddr, pmd_t *pmd)
2046 {
2047 	struct mm_struct *mm = vma->vm_mm;
2048 	pgtable_t pgtable;
2049 	pmd_t _pmd;
2050 	int i;
2051 
2052 	/*
2053 	 * Leave pmd empty until pte is filled note that it is fine to delay
2054 	 * notification until mmu_notifier_invalidate_range_end() as we are
2055 	 * replacing a zero pmd write protected page with a zero pte write
2056 	 * protected page.
2057 	 *
2058 	 * See Documentation/mm/mmu_notifier.rst
2059 	 */
2060 	pmdp_huge_clear_flush(vma, haddr, pmd);
2061 
2062 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2063 	pmd_populate(mm, &_pmd, pgtable);
2064 
2065 	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2066 		pte_t *pte, entry;
2067 		entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2068 		entry = pte_mkspecial(entry);
2069 		pte = pte_offset_map(&_pmd, haddr);
2070 		VM_BUG_ON(!pte_none(*pte));
2071 		set_pte_at(mm, haddr, pte, entry);
2072 		pte_unmap(pte);
2073 	}
2074 	smp_wmb(); /* make pte visible before pmd */
2075 	pmd_populate(mm, pmd, pgtable);
2076 }
2077 
2078 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2079 		unsigned long haddr, bool freeze)
2080 {
2081 	struct mm_struct *mm = vma->vm_mm;
2082 	struct page *page;
2083 	pgtable_t pgtable;
2084 	pmd_t old_pmd, _pmd;
2085 	bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2086 	bool anon_exclusive = false, dirty = false;
2087 	unsigned long addr;
2088 	int i;
2089 
2090 	VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2091 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2092 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2093 	VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2094 				&& !pmd_devmap(*pmd));
2095 
2096 	count_vm_event(THP_SPLIT_PMD);
2097 
2098 	if (!vma_is_anonymous(vma)) {
2099 		old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2100 		/*
2101 		 * We are going to unmap this huge page. So
2102 		 * just go ahead and zap it
2103 		 */
2104 		if (arch_needs_pgtable_deposit())
2105 			zap_deposited_table(mm, pmd);
2106 		if (vma_is_special_huge(vma))
2107 			return;
2108 		if (unlikely(is_pmd_migration_entry(old_pmd))) {
2109 			swp_entry_t entry;
2110 
2111 			entry = pmd_to_swp_entry(old_pmd);
2112 			page = pfn_swap_entry_to_page(entry);
2113 		} else {
2114 			page = pmd_page(old_pmd);
2115 			if (!PageDirty(page) && pmd_dirty(old_pmd))
2116 				set_page_dirty(page);
2117 			if (!PageReferenced(page) && pmd_young(old_pmd))
2118 				SetPageReferenced(page);
2119 			page_remove_rmap(page, vma, true);
2120 			put_page(page);
2121 		}
2122 		add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2123 		return;
2124 	}
2125 
2126 	if (is_huge_zero_pmd(*pmd)) {
2127 		/*
2128 		 * FIXME: Do we want to invalidate secondary mmu by calling
2129 		 * mmu_notifier_invalidate_range() see comments below inside
2130 		 * __split_huge_pmd() ?
2131 		 *
2132 		 * We are going from a zero huge page write protected to zero
2133 		 * small page also write protected so it does not seems useful
2134 		 * to invalidate secondary mmu at this time.
2135 		 */
2136 		return __split_huge_zero_page_pmd(vma, haddr, pmd);
2137 	}
2138 
2139 	/*
2140 	 * Up to this point the pmd is present and huge and userland has the
2141 	 * whole access to the hugepage during the split (which happens in
2142 	 * place). If we overwrite the pmd with the not-huge version pointing
2143 	 * to the pte here (which of course we could if all CPUs were bug
2144 	 * free), userland could trigger a small page size TLB miss on the
2145 	 * small sized TLB while the hugepage TLB entry is still established in
2146 	 * the huge TLB. Some CPU doesn't like that.
2147 	 * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2148 	 * 383 on page 105. Intel should be safe but is also warns that it's
2149 	 * only safe if the permission and cache attributes of the two entries
2150 	 * loaded in the two TLB is identical (which should be the case here).
2151 	 * But it is generally safer to never allow small and huge TLB entries
2152 	 * for the same virtual address to be loaded simultaneously. So instead
2153 	 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2154 	 * current pmd notpresent (atomically because here the pmd_trans_huge
2155 	 * must remain set at all times on the pmd until the split is complete
2156 	 * for this pmd), then we flush the SMP TLB and finally we write the
2157 	 * non-huge version of the pmd entry with pmd_populate.
2158 	 */
2159 	old_pmd = pmdp_invalidate(vma, haddr, pmd);
2160 
2161 	pmd_migration = is_pmd_migration_entry(old_pmd);
2162 	if (unlikely(pmd_migration)) {
2163 		swp_entry_t entry;
2164 
2165 		entry = pmd_to_swp_entry(old_pmd);
2166 		page = pfn_swap_entry_to_page(entry);
2167 		write = is_writable_migration_entry(entry);
2168 		if (PageAnon(page))
2169 			anon_exclusive = is_readable_exclusive_migration_entry(entry);
2170 		young = is_migration_entry_young(entry);
2171 		dirty = is_migration_entry_dirty(entry);
2172 		soft_dirty = pmd_swp_soft_dirty(old_pmd);
2173 		uffd_wp = pmd_swp_uffd_wp(old_pmd);
2174 	} else {
2175 		page = pmd_page(old_pmd);
2176 		if (pmd_dirty(old_pmd)) {
2177 			dirty = true;
2178 			SetPageDirty(page);
2179 		}
2180 		write = pmd_write(old_pmd);
2181 		young = pmd_young(old_pmd);
2182 		soft_dirty = pmd_soft_dirty(old_pmd);
2183 		uffd_wp = pmd_uffd_wp(old_pmd);
2184 
2185 		VM_BUG_ON_PAGE(!page_count(page), page);
2186 
2187 		/*
2188 		 * Without "freeze", we'll simply split the PMD, propagating the
2189 		 * PageAnonExclusive() flag for each PTE by setting it for
2190 		 * each subpage -- no need to (temporarily) clear.
2191 		 *
2192 		 * With "freeze" we want to replace mapped pages by
2193 		 * migration entries right away. This is only possible if we
2194 		 * managed to clear PageAnonExclusive() -- see
2195 		 * set_pmd_migration_entry().
2196 		 *
2197 		 * In case we cannot clear PageAnonExclusive(), split the PMD
2198 		 * only and let try_to_migrate_one() fail later.
2199 		 *
2200 		 * See page_try_share_anon_rmap(): invalidate PMD first.
2201 		 */
2202 		anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
2203 		if (freeze && anon_exclusive && page_try_share_anon_rmap(page))
2204 			freeze = false;
2205 		if (!freeze)
2206 			page_ref_add(page, HPAGE_PMD_NR - 1);
2207 	}
2208 
2209 	/*
2210 	 * Withdraw the table only after we mark the pmd entry invalid.
2211 	 * This's critical for some architectures (Power).
2212 	 */
2213 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2214 	pmd_populate(mm, &_pmd, pgtable);
2215 
2216 	for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2217 		pte_t entry, *pte;
2218 		/*
2219 		 * Note that NUMA hinting access restrictions are not
2220 		 * transferred to avoid any possibility of altering
2221 		 * permissions across VMAs.
2222 		 */
2223 		if (freeze || pmd_migration) {
2224 			swp_entry_t swp_entry;
2225 			if (write)
2226 				swp_entry = make_writable_migration_entry(
2227 							page_to_pfn(page + i));
2228 			else if (anon_exclusive)
2229 				swp_entry = make_readable_exclusive_migration_entry(
2230 							page_to_pfn(page + i));
2231 			else
2232 				swp_entry = make_readable_migration_entry(
2233 							page_to_pfn(page + i));
2234 			if (young)
2235 				swp_entry = make_migration_entry_young(swp_entry);
2236 			if (dirty)
2237 				swp_entry = make_migration_entry_dirty(swp_entry);
2238 			entry = swp_entry_to_pte(swp_entry);
2239 			if (soft_dirty)
2240 				entry = pte_swp_mksoft_dirty(entry);
2241 			if (uffd_wp)
2242 				entry = pte_swp_mkuffd_wp(entry);
2243 		} else {
2244 			entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2245 			entry = maybe_mkwrite(entry, vma);
2246 			if (anon_exclusive)
2247 				SetPageAnonExclusive(page + i);
2248 			if (!young)
2249 				entry = pte_mkold(entry);
2250 			/* NOTE: this may set soft-dirty too on some archs */
2251 			if (dirty)
2252 				entry = pte_mkdirty(entry);
2253 			/*
2254 			 * NOTE: this needs to happen after pte_mkdirty,
2255 			 * because some archs (sparc64, loongarch) could
2256 			 * set hw write bit when mkdirty.
2257 			 */
2258 			if (!write)
2259 				entry = pte_wrprotect(entry);
2260 			if (soft_dirty)
2261 				entry = pte_mksoft_dirty(entry);
2262 			if (uffd_wp)
2263 				entry = pte_mkuffd_wp(entry);
2264 			page_add_anon_rmap(page + i, vma, addr, false);
2265 		}
2266 		pte = pte_offset_map(&_pmd, addr);
2267 		BUG_ON(!pte_none(*pte));
2268 		set_pte_at(mm, addr, pte, entry);
2269 		pte_unmap(pte);
2270 	}
2271 
2272 	if (!pmd_migration)
2273 		page_remove_rmap(page, vma, true);
2274 	if (freeze)
2275 		put_page(page);
2276 
2277 	smp_wmb(); /* make pte visible before pmd */
2278 	pmd_populate(mm, pmd, pgtable);
2279 }
2280 
2281 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2282 		unsigned long address, bool freeze, struct folio *folio)
2283 {
2284 	spinlock_t *ptl;
2285 	struct mmu_notifier_range range;
2286 
2287 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2288 				address & HPAGE_PMD_MASK,
2289 				(address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2290 	mmu_notifier_invalidate_range_start(&range);
2291 	ptl = pmd_lock(vma->vm_mm, pmd);
2292 
2293 	/*
2294 	 * If caller asks to setup a migration entry, we need a folio to check
2295 	 * pmd against. Otherwise we can end up replacing wrong folio.
2296 	 */
2297 	VM_BUG_ON(freeze && !folio);
2298 	VM_WARN_ON_ONCE(folio && !folio_test_locked(folio));
2299 
2300 	if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) ||
2301 	    is_pmd_migration_entry(*pmd)) {
2302 		/*
2303 		 * It's safe to call pmd_page when folio is set because it's
2304 		 * guaranteed that pmd is present.
2305 		 */
2306 		if (folio && folio != page_folio(pmd_page(*pmd)))
2307 			goto out;
2308 		__split_huge_pmd_locked(vma, pmd, range.start, freeze);
2309 	}
2310 
2311 out:
2312 	spin_unlock(ptl);
2313 	/*
2314 	 * No need to double call mmu_notifier->invalidate_range() callback.
2315 	 * They are 3 cases to consider inside __split_huge_pmd_locked():
2316 	 *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2317 	 *  2) __split_huge_zero_page_pmd() read only zero page and any write
2318 	 *    fault will trigger a flush_notify before pointing to a new page
2319 	 *    (it is fine if the secondary mmu keeps pointing to the old zero
2320 	 *    page in the meantime)
2321 	 *  3) Split a huge pmd into pte pointing to the same page. No need
2322 	 *     to invalidate secondary tlb entry they are all still valid.
2323 	 *     any further changes to individual pte will notify. So no need
2324 	 *     to call mmu_notifier->invalidate_range()
2325 	 */
2326 	mmu_notifier_invalidate_range_only_end(&range);
2327 }
2328 
2329 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2330 		bool freeze, struct folio *folio)
2331 {
2332 	pmd_t *pmd = mm_find_pmd(vma->vm_mm, address);
2333 
2334 	if (!pmd)
2335 		return;
2336 
2337 	__split_huge_pmd(vma, pmd, address, freeze, folio);
2338 }
2339 
2340 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2341 {
2342 	/*
2343 	 * If the new address isn't hpage aligned and it could previously
2344 	 * contain an hugepage: check if we need to split an huge pmd.
2345 	 */
2346 	if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2347 	    range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2348 			 ALIGN(address, HPAGE_PMD_SIZE)))
2349 		split_huge_pmd_address(vma, address, false, NULL);
2350 }
2351 
2352 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2353 			     unsigned long start,
2354 			     unsigned long end,
2355 			     long adjust_next)
2356 {
2357 	/* Check if we need to split start first. */
2358 	split_huge_pmd_if_needed(vma, start);
2359 
2360 	/* Check if we need to split end next. */
2361 	split_huge_pmd_if_needed(vma, end);
2362 
2363 	/*
2364 	 * If we're also updating the next vma vm_start,
2365 	 * check if we need to split it.
2366 	 */
2367 	if (adjust_next > 0) {
2368 		struct vm_area_struct *next = find_vma(vma->vm_mm, vma->vm_end);
2369 		unsigned long nstart = next->vm_start;
2370 		nstart += adjust_next;
2371 		split_huge_pmd_if_needed(next, nstart);
2372 	}
2373 }
2374 
2375 static void unmap_folio(struct folio *folio)
2376 {
2377 	enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2378 		TTU_SYNC;
2379 
2380 	VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2381 
2382 	/*
2383 	 * Anon pages need migration entries to preserve them, but file
2384 	 * pages can simply be left unmapped, then faulted back on demand.
2385 	 * If that is ever changed (perhaps for mlock), update remap_page().
2386 	 */
2387 	if (folio_test_anon(folio))
2388 		try_to_migrate(folio, ttu_flags);
2389 	else
2390 		try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK);
2391 }
2392 
2393 static void remap_page(struct folio *folio, unsigned long nr)
2394 {
2395 	int i = 0;
2396 
2397 	/* If unmap_folio() uses try_to_migrate() on file, remove this check */
2398 	if (!folio_test_anon(folio))
2399 		return;
2400 	for (;;) {
2401 		remove_migration_ptes(folio, folio, true);
2402 		i += folio_nr_pages(folio);
2403 		if (i >= nr)
2404 			break;
2405 		folio = folio_next(folio);
2406 	}
2407 }
2408 
2409 static void lru_add_page_tail(struct page *head, struct page *tail,
2410 		struct lruvec *lruvec, struct list_head *list)
2411 {
2412 	VM_BUG_ON_PAGE(!PageHead(head), head);
2413 	VM_BUG_ON_PAGE(PageCompound(tail), head);
2414 	VM_BUG_ON_PAGE(PageLRU(tail), head);
2415 	lockdep_assert_held(&lruvec->lru_lock);
2416 
2417 	if (list) {
2418 		/* page reclaim is reclaiming a huge page */
2419 		VM_WARN_ON(PageLRU(head));
2420 		get_page(tail);
2421 		list_add_tail(&tail->lru, list);
2422 	} else {
2423 		/* head is still on lru (and we have it frozen) */
2424 		VM_WARN_ON(!PageLRU(head));
2425 		if (PageUnevictable(tail))
2426 			tail->mlock_count = 0;
2427 		else
2428 			list_add_tail(&tail->lru, &head->lru);
2429 		SetPageLRU(tail);
2430 	}
2431 }
2432 
2433 static void __split_huge_page_tail(struct page *head, int tail,
2434 		struct lruvec *lruvec, struct list_head *list)
2435 {
2436 	struct page *page_tail = head + tail;
2437 
2438 	VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2439 
2440 	/*
2441 	 * Clone page flags before unfreezing refcount.
2442 	 *
2443 	 * After successful get_page_unless_zero() might follow flags change,
2444 	 * for example lock_page() which set PG_waiters.
2445 	 *
2446 	 * Note that for mapped sub-pages of an anonymous THP,
2447 	 * PG_anon_exclusive has been cleared in unmap_folio() and is stored in
2448 	 * the migration entry instead from where remap_page() will restore it.
2449 	 * We can still have PG_anon_exclusive set on effectively unmapped and
2450 	 * unreferenced sub-pages of an anonymous THP: we can simply drop
2451 	 * PG_anon_exclusive (-> PG_mappedtodisk) for these here.
2452 	 */
2453 	page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2454 	page_tail->flags |= (head->flags &
2455 			((1L << PG_referenced) |
2456 			 (1L << PG_swapbacked) |
2457 			 (1L << PG_swapcache) |
2458 			 (1L << PG_mlocked) |
2459 			 (1L << PG_uptodate) |
2460 			 (1L << PG_active) |
2461 			 (1L << PG_workingset) |
2462 			 (1L << PG_locked) |
2463 			 (1L << PG_unevictable) |
2464 #ifdef CONFIG_ARCH_USES_PG_ARCH_X
2465 			 (1L << PG_arch_2) |
2466 			 (1L << PG_arch_3) |
2467 #endif
2468 			 (1L << PG_dirty) |
2469 			 LRU_GEN_MASK | LRU_REFS_MASK));
2470 
2471 	/* ->mapping in first and second tail page is replaced by other uses */
2472 	VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2473 			page_tail);
2474 	page_tail->mapping = head->mapping;
2475 	page_tail->index = head->index + tail;
2476 
2477 	/*
2478 	 * page->private should not be set in tail pages with the exception
2479 	 * of swap cache pages that store the swp_entry_t in tail pages.
2480 	 * Fix up and warn once if private is unexpectedly set.
2481 	 *
2482 	 * What of 32-bit systems, on which head[1].compound_pincount overlays
2483 	 * head[1].private?  No problem: THP_SWAP is not enabled on 32-bit, and
2484 	 * compound_pincount must be 0 for folio_ref_freeze() to have succeeded.
2485 	 */
2486 	if (!folio_test_swapcache(page_folio(head))) {
2487 		VM_WARN_ON_ONCE_PAGE(page_tail->private != 0, page_tail);
2488 		page_tail->private = 0;
2489 	}
2490 
2491 	/* Page flags must be visible before we make the page non-compound. */
2492 	smp_wmb();
2493 
2494 	/*
2495 	 * Clear PageTail before unfreezing page refcount.
2496 	 *
2497 	 * After successful get_page_unless_zero() might follow put_page()
2498 	 * which needs correct compound_head().
2499 	 */
2500 	clear_compound_head(page_tail);
2501 
2502 	/* Finally unfreeze refcount. Additional reference from page cache. */
2503 	page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2504 					  PageSwapCache(head)));
2505 
2506 	if (page_is_young(head))
2507 		set_page_young(page_tail);
2508 	if (page_is_idle(head))
2509 		set_page_idle(page_tail);
2510 
2511 	page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2512 
2513 	/*
2514 	 * always add to the tail because some iterators expect new
2515 	 * pages to show after the currently processed elements - e.g.
2516 	 * migrate_pages
2517 	 */
2518 	lru_add_page_tail(head, page_tail, lruvec, list);
2519 }
2520 
2521 static void __split_huge_page(struct page *page, struct list_head *list,
2522 		pgoff_t end)
2523 {
2524 	struct folio *folio = page_folio(page);
2525 	struct page *head = &folio->page;
2526 	struct lruvec *lruvec;
2527 	struct address_space *swap_cache = NULL;
2528 	unsigned long offset = 0;
2529 	unsigned int nr = thp_nr_pages(head);
2530 	int i;
2531 
2532 	/* complete memcg works before add pages to LRU */
2533 	split_page_memcg(head, nr);
2534 
2535 	if (PageAnon(head) && PageSwapCache(head)) {
2536 		swp_entry_t entry = { .val = page_private(head) };
2537 
2538 		offset = swp_offset(entry);
2539 		swap_cache = swap_address_space(entry);
2540 		xa_lock(&swap_cache->i_pages);
2541 	}
2542 
2543 	/* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2544 	lruvec = folio_lruvec_lock(folio);
2545 
2546 	ClearPageHasHWPoisoned(head);
2547 
2548 	for (i = nr - 1; i >= 1; i--) {
2549 		__split_huge_page_tail(head, i, lruvec, list);
2550 		/* Some pages can be beyond EOF: drop them from page cache */
2551 		if (head[i].index >= end) {
2552 			struct folio *tail = page_folio(head + i);
2553 
2554 			if (shmem_mapping(head->mapping))
2555 				shmem_uncharge(head->mapping->host, 1);
2556 			else if (folio_test_clear_dirty(tail))
2557 				folio_account_cleaned(tail,
2558 					inode_to_wb(folio->mapping->host));
2559 			__filemap_remove_folio(tail, NULL);
2560 			folio_put(tail);
2561 		} else if (!PageAnon(page)) {
2562 			__xa_store(&head->mapping->i_pages, head[i].index,
2563 					head + i, 0);
2564 		} else if (swap_cache) {
2565 			__xa_store(&swap_cache->i_pages, offset + i,
2566 					head + i, 0);
2567 		}
2568 	}
2569 
2570 	ClearPageCompound(head);
2571 	unlock_page_lruvec(lruvec);
2572 	/* Caller disabled irqs, so they are still disabled here */
2573 
2574 	split_page_owner(head, nr);
2575 
2576 	/* See comment in __split_huge_page_tail() */
2577 	if (PageAnon(head)) {
2578 		/* Additional pin to swap cache */
2579 		if (PageSwapCache(head)) {
2580 			page_ref_add(head, 2);
2581 			xa_unlock(&swap_cache->i_pages);
2582 		} else {
2583 			page_ref_inc(head);
2584 		}
2585 	} else {
2586 		/* Additional pin to page cache */
2587 		page_ref_add(head, 2);
2588 		xa_unlock(&head->mapping->i_pages);
2589 	}
2590 	local_irq_enable();
2591 
2592 	remap_page(folio, nr);
2593 
2594 	if (PageSwapCache(head)) {
2595 		swp_entry_t entry = { .val = page_private(head) };
2596 
2597 		split_swap_cluster(entry);
2598 	}
2599 
2600 	for (i = 0; i < nr; i++) {
2601 		struct page *subpage = head + i;
2602 		if (subpage == page)
2603 			continue;
2604 		unlock_page(subpage);
2605 
2606 		/*
2607 		 * Subpages may be freed if there wasn't any mapping
2608 		 * like if add_to_swap() is running on a lru page that
2609 		 * had its mapping zapped. And freeing these pages
2610 		 * requires taking the lru_lock so we do the put_page
2611 		 * of the tail pages after the split is complete.
2612 		 */
2613 		free_page_and_swap_cache(subpage);
2614 	}
2615 }
2616 
2617 /* Racy check whether the huge page can be split */
2618 bool can_split_folio(struct folio *folio, int *pextra_pins)
2619 {
2620 	int extra_pins;
2621 
2622 	/* Additional pins from page cache */
2623 	if (folio_test_anon(folio))
2624 		extra_pins = folio_test_swapcache(folio) ?
2625 				folio_nr_pages(folio) : 0;
2626 	else
2627 		extra_pins = folio_nr_pages(folio);
2628 	if (pextra_pins)
2629 		*pextra_pins = extra_pins;
2630 	return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 1;
2631 }
2632 
2633 /*
2634  * This function splits huge page into normal pages. @page can point to any
2635  * subpage of huge page to split. Split doesn't change the position of @page.
2636  *
2637  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2638  * The huge page must be locked.
2639  *
2640  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2641  *
2642  * Both head page and tail pages will inherit mapping, flags, and so on from
2643  * the hugepage.
2644  *
2645  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2646  * they are not mapped.
2647  *
2648  * Returns 0 if the hugepage is split successfully.
2649  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2650  * us.
2651  */
2652 int split_huge_page_to_list(struct page *page, struct list_head *list)
2653 {
2654 	struct folio *folio = page_folio(page);
2655 	struct deferred_split *ds_queue = get_deferred_split_queue(&folio->page);
2656 	XA_STATE(xas, &folio->mapping->i_pages, folio->index);
2657 	struct anon_vma *anon_vma = NULL;
2658 	struct address_space *mapping = NULL;
2659 	int extra_pins, ret;
2660 	pgoff_t end;
2661 	bool is_hzp;
2662 
2663 	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2664 	VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2665 
2666 	is_hzp = is_huge_zero_page(&folio->page);
2667 	VM_WARN_ON_ONCE_FOLIO(is_hzp, folio);
2668 	if (is_hzp)
2669 		return -EBUSY;
2670 
2671 	if (folio_test_writeback(folio))
2672 		return -EBUSY;
2673 
2674 	if (folio_test_anon(folio)) {
2675 		/*
2676 		 * The caller does not necessarily hold an mmap_lock that would
2677 		 * prevent the anon_vma disappearing so we first we take a
2678 		 * reference to it and then lock the anon_vma for write. This
2679 		 * is similar to folio_lock_anon_vma_read except the write lock
2680 		 * is taken to serialise against parallel split or collapse
2681 		 * operations.
2682 		 */
2683 		anon_vma = folio_get_anon_vma(folio);
2684 		if (!anon_vma) {
2685 			ret = -EBUSY;
2686 			goto out;
2687 		}
2688 		end = -1;
2689 		mapping = NULL;
2690 		anon_vma_lock_write(anon_vma);
2691 	} else {
2692 		gfp_t gfp;
2693 
2694 		mapping = folio->mapping;
2695 
2696 		/* Truncated ? */
2697 		if (!mapping) {
2698 			ret = -EBUSY;
2699 			goto out;
2700 		}
2701 
2702 		gfp = current_gfp_context(mapping_gfp_mask(mapping) &
2703 							GFP_RECLAIM_MASK);
2704 
2705 		if (folio_test_private(folio) &&
2706 				!filemap_release_folio(folio, gfp)) {
2707 			ret = -EBUSY;
2708 			goto out;
2709 		}
2710 
2711 		xas_split_alloc(&xas, folio, folio_order(folio), gfp);
2712 		if (xas_error(&xas)) {
2713 			ret = xas_error(&xas);
2714 			goto out;
2715 		}
2716 
2717 		anon_vma = NULL;
2718 		i_mmap_lock_read(mapping);
2719 
2720 		/*
2721 		 *__split_huge_page() may need to trim off pages beyond EOF:
2722 		 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2723 		 * which cannot be nested inside the page tree lock. So note
2724 		 * end now: i_size itself may be changed at any moment, but
2725 		 * folio lock is good enough to serialize the trimming.
2726 		 */
2727 		end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2728 		if (shmem_mapping(mapping))
2729 			end = shmem_fallocend(mapping->host, end);
2730 	}
2731 
2732 	/*
2733 	 * Racy check if we can split the page, before unmap_folio() will
2734 	 * split PMDs
2735 	 */
2736 	if (!can_split_folio(folio, &extra_pins)) {
2737 		ret = -EAGAIN;
2738 		goto out_unlock;
2739 	}
2740 
2741 	unmap_folio(folio);
2742 
2743 	/* block interrupt reentry in xa_lock and spinlock */
2744 	local_irq_disable();
2745 	if (mapping) {
2746 		/*
2747 		 * Check if the folio is present in page cache.
2748 		 * We assume all tail are present too, if folio is there.
2749 		 */
2750 		xas_lock(&xas);
2751 		xas_reset(&xas);
2752 		if (xas_load(&xas) != folio)
2753 			goto fail;
2754 	}
2755 
2756 	/* Prevent deferred_split_scan() touching ->_refcount */
2757 	spin_lock(&ds_queue->split_queue_lock);
2758 	if (folio_ref_freeze(folio, 1 + extra_pins)) {
2759 		if (!list_empty(page_deferred_list(&folio->page))) {
2760 			ds_queue->split_queue_len--;
2761 			list_del(page_deferred_list(&folio->page));
2762 		}
2763 		spin_unlock(&ds_queue->split_queue_lock);
2764 		if (mapping) {
2765 			int nr = folio_nr_pages(folio);
2766 
2767 			xas_split(&xas, folio, folio_order(folio));
2768 			if (folio_test_swapbacked(folio)) {
2769 				__lruvec_stat_mod_folio(folio, NR_SHMEM_THPS,
2770 							-nr);
2771 			} else {
2772 				__lruvec_stat_mod_folio(folio, NR_FILE_THPS,
2773 							-nr);
2774 				filemap_nr_thps_dec(mapping);
2775 			}
2776 		}
2777 
2778 		__split_huge_page(page, list, end);
2779 		ret = 0;
2780 	} else {
2781 		spin_unlock(&ds_queue->split_queue_lock);
2782 fail:
2783 		if (mapping)
2784 			xas_unlock(&xas);
2785 		local_irq_enable();
2786 		remap_page(folio, folio_nr_pages(folio));
2787 		ret = -EAGAIN;
2788 	}
2789 
2790 out_unlock:
2791 	if (anon_vma) {
2792 		anon_vma_unlock_write(anon_vma);
2793 		put_anon_vma(anon_vma);
2794 	}
2795 	if (mapping)
2796 		i_mmap_unlock_read(mapping);
2797 out:
2798 	xas_destroy(&xas);
2799 	count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2800 	return ret;
2801 }
2802 
2803 void free_transhuge_page(struct page *page)
2804 {
2805 	struct deferred_split *ds_queue = get_deferred_split_queue(page);
2806 	unsigned long flags;
2807 
2808 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2809 	if (!list_empty(page_deferred_list(page))) {
2810 		ds_queue->split_queue_len--;
2811 		list_del(page_deferred_list(page));
2812 	}
2813 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2814 	free_compound_page(page);
2815 }
2816 
2817 void deferred_split_huge_page(struct page *page)
2818 {
2819 	struct deferred_split *ds_queue = get_deferred_split_queue(page);
2820 #ifdef CONFIG_MEMCG
2821 	struct mem_cgroup *memcg = page_memcg(compound_head(page));
2822 #endif
2823 	unsigned long flags;
2824 
2825 	VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2826 
2827 	/*
2828 	 * The try_to_unmap() in page reclaim path might reach here too,
2829 	 * this may cause a race condition to corrupt deferred split queue.
2830 	 * And, if page reclaim is already handling the same page, it is
2831 	 * unnecessary to handle it again in shrinker.
2832 	 *
2833 	 * Check PageSwapCache to determine if the page is being
2834 	 * handled by page reclaim since THP swap would add the page into
2835 	 * swap cache before calling try_to_unmap().
2836 	 */
2837 	if (PageSwapCache(page))
2838 		return;
2839 
2840 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2841 	if (list_empty(page_deferred_list(page))) {
2842 		count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2843 		list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2844 		ds_queue->split_queue_len++;
2845 #ifdef CONFIG_MEMCG
2846 		if (memcg)
2847 			set_shrinker_bit(memcg, page_to_nid(page),
2848 					 deferred_split_shrinker.id);
2849 #endif
2850 	}
2851 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2852 }
2853 
2854 static unsigned long deferred_split_count(struct shrinker *shrink,
2855 		struct shrink_control *sc)
2856 {
2857 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2858 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2859 
2860 #ifdef CONFIG_MEMCG
2861 	if (sc->memcg)
2862 		ds_queue = &sc->memcg->deferred_split_queue;
2863 #endif
2864 	return READ_ONCE(ds_queue->split_queue_len);
2865 }
2866 
2867 static unsigned long deferred_split_scan(struct shrinker *shrink,
2868 		struct shrink_control *sc)
2869 {
2870 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2871 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2872 	unsigned long flags;
2873 	LIST_HEAD(list), *pos, *next;
2874 	struct page *page;
2875 	int split = 0;
2876 
2877 #ifdef CONFIG_MEMCG
2878 	if (sc->memcg)
2879 		ds_queue = &sc->memcg->deferred_split_queue;
2880 #endif
2881 
2882 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2883 	/* Take pin on all head pages to avoid freeing them under us */
2884 	list_for_each_safe(pos, next, &ds_queue->split_queue) {
2885 		page = list_entry((void *)pos, struct page, deferred_list);
2886 		page = compound_head(page);
2887 		if (get_page_unless_zero(page)) {
2888 			list_move(page_deferred_list(page), &list);
2889 		} else {
2890 			/* We lost race with put_compound_page() */
2891 			list_del_init(page_deferred_list(page));
2892 			ds_queue->split_queue_len--;
2893 		}
2894 		if (!--sc->nr_to_scan)
2895 			break;
2896 	}
2897 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2898 
2899 	list_for_each_safe(pos, next, &list) {
2900 		page = list_entry((void *)pos, struct page, deferred_list);
2901 		if (!trylock_page(page))
2902 			goto next;
2903 		/* split_huge_page() removes page from list on success */
2904 		if (!split_huge_page(page))
2905 			split++;
2906 		unlock_page(page);
2907 next:
2908 		put_page(page);
2909 	}
2910 
2911 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2912 	list_splice_tail(&list, &ds_queue->split_queue);
2913 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2914 
2915 	/*
2916 	 * Stop shrinker if we didn't split any page, but the queue is empty.
2917 	 * This can happen if pages were freed under us.
2918 	 */
2919 	if (!split && list_empty(&ds_queue->split_queue))
2920 		return SHRINK_STOP;
2921 	return split;
2922 }
2923 
2924 static struct shrinker deferred_split_shrinker = {
2925 	.count_objects = deferred_split_count,
2926 	.scan_objects = deferred_split_scan,
2927 	.seeks = DEFAULT_SEEKS,
2928 	.flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2929 		 SHRINKER_NONSLAB,
2930 };
2931 
2932 #ifdef CONFIG_DEBUG_FS
2933 static void split_huge_pages_all(void)
2934 {
2935 	struct zone *zone;
2936 	struct page *page;
2937 	unsigned long pfn, max_zone_pfn;
2938 	unsigned long total = 0, split = 0;
2939 
2940 	pr_debug("Split all THPs\n");
2941 	for_each_zone(zone) {
2942 		if (!managed_zone(zone))
2943 			continue;
2944 		max_zone_pfn = zone_end_pfn(zone);
2945 		for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2946 			int nr_pages;
2947 
2948 			page = pfn_to_online_page(pfn);
2949 			if (!page || !get_page_unless_zero(page))
2950 				continue;
2951 
2952 			if (zone != page_zone(page))
2953 				goto next;
2954 
2955 			if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2956 				goto next;
2957 
2958 			total++;
2959 			lock_page(page);
2960 			nr_pages = thp_nr_pages(page);
2961 			if (!split_huge_page(page))
2962 				split++;
2963 			pfn += nr_pages - 1;
2964 			unlock_page(page);
2965 next:
2966 			put_page(page);
2967 			cond_resched();
2968 		}
2969 	}
2970 
2971 	pr_debug("%lu of %lu THP split\n", split, total);
2972 }
2973 
2974 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
2975 {
2976 	return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
2977 		    is_vm_hugetlb_page(vma);
2978 }
2979 
2980 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
2981 				unsigned long vaddr_end)
2982 {
2983 	int ret = 0;
2984 	struct task_struct *task;
2985 	struct mm_struct *mm;
2986 	unsigned long total = 0, split = 0;
2987 	unsigned long addr;
2988 
2989 	vaddr_start &= PAGE_MASK;
2990 	vaddr_end &= PAGE_MASK;
2991 
2992 	/* Find the task_struct from pid */
2993 	rcu_read_lock();
2994 	task = find_task_by_vpid(pid);
2995 	if (!task) {
2996 		rcu_read_unlock();
2997 		ret = -ESRCH;
2998 		goto out;
2999 	}
3000 	get_task_struct(task);
3001 	rcu_read_unlock();
3002 
3003 	/* Find the mm_struct */
3004 	mm = get_task_mm(task);
3005 	put_task_struct(task);
3006 
3007 	if (!mm) {
3008 		ret = -EINVAL;
3009 		goto out;
3010 	}
3011 
3012 	pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
3013 		 pid, vaddr_start, vaddr_end);
3014 
3015 	mmap_read_lock(mm);
3016 	/*
3017 	 * always increase addr by PAGE_SIZE, since we could have a PTE page
3018 	 * table filled with PTE-mapped THPs, each of which is distinct.
3019 	 */
3020 	for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
3021 		struct vm_area_struct *vma = vma_lookup(mm, addr);
3022 		struct page *page;
3023 
3024 		if (!vma)
3025 			break;
3026 
3027 		/* skip special VMA and hugetlb VMA */
3028 		if (vma_not_suitable_for_thp_split(vma)) {
3029 			addr = vma->vm_end;
3030 			continue;
3031 		}
3032 
3033 		/* FOLL_DUMP to ignore special (like zero) pages */
3034 		page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
3035 
3036 		if (IS_ERR_OR_NULL(page))
3037 			continue;
3038 
3039 		if (!is_transparent_hugepage(page))
3040 			goto next;
3041 
3042 		total++;
3043 		if (!can_split_folio(page_folio(page), NULL))
3044 			goto next;
3045 
3046 		if (!trylock_page(page))
3047 			goto next;
3048 
3049 		if (!split_huge_page(page))
3050 			split++;
3051 
3052 		unlock_page(page);
3053 next:
3054 		put_page(page);
3055 		cond_resched();
3056 	}
3057 	mmap_read_unlock(mm);
3058 	mmput(mm);
3059 
3060 	pr_debug("%lu of %lu THP split\n", split, total);
3061 
3062 out:
3063 	return ret;
3064 }
3065 
3066 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
3067 				pgoff_t off_end)
3068 {
3069 	struct filename *file;
3070 	struct file *candidate;
3071 	struct address_space *mapping;
3072 	int ret = -EINVAL;
3073 	pgoff_t index;
3074 	int nr_pages = 1;
3075 	unsigned long total = 0, split = 0;
3076 
3077 	file = getname_kernel(file_path);
3078 	if (IS_ERR(file))
3079 		return ret;
3080 
3081 	candidate = file_open_name(file, O_RDONLY, 0);
3082 	if (IS_ERR(candidate))
3083 		goto out;
3084 
3085 	pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3086 		 file_path, off_start, off_end);
3087 
3088 	mapping = candidate->f_mapping;
3089 
3090 	for (index = off_start; index < off_end; index += nr_pages) {
3091 		struct folio *folio = __filemap_get_folio(mapping, index,
3092 						FGP_ENTRY, 0);
3093 
3094 		nr_pages = 1;
3095 		if (xa_is_value(folio) || !folio)
3096 			continue;
3097 
3098 		if (!folio_test_large(folio))
3099 			goto next;
3100 
3101 		total++;
3102 		nr_pages = folio_nr_pages(folio);
3103 
3104 		if (!folio_trylock(folio))
3105 			goto next;
3106 
3107 		if (!split_folio(folio))
3108 			split++;
3109 
3110 		folio_unlock(folio);
3111 next:
3112 		folio_put(folio);
3113 		cond_resched();
3114 	}
3115 
3116 	filp_close(candidate, NULL);
3117 	ret = 0;
3118 
3119 	pr_debug("%lu of %lu file-backed THP split\n", split, total);
3120 out:
3121 	putname(file);
3122 	return ret;
3123 }
3124 
3125 #define MAX_INPUT_BUF_SZ 255
3126 
3127 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3128 				size_t count, loff_t *ppops)
3129 {
3130 	static DEFINE_MUTEX(split_debug_mutex);
3131 	ssize_t ret;
3132 	/* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3133 	char input_buf[MAX_INPUT_BUF_SZ];
3134 	int pid;
3135 	unsigned long vaddr_start, vaddr_end;
3136 
3137 	ret = mutex_lock_interruptible(&split_debug_mutex);
3138 	if (ret)
3139 		return ret;
3140 
3141 	ret = -EFAULT;
3142 
3143 	memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3144 	if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3145 		goto out;
3146 
3147 	input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3148 
3149 	if (input_buf[0] == '/') {
3150 		char *tok;
3151 		char *buf = input_buf;
3152 		char file_path[MAX_INPUT_BUF_SZ];
3153 		pgoff_t off_start = 0, off_end = 0;
3154 		size_t input_len = strlen(input_buf);
3155 
3156 		tok = strsep(&buf, ",");
3157 		if (tok) {
3158 			strcpy(file_path, tok);
3159 		} else {
3160 			ret = -EINVAL;
3161 			goto out;
3162 		}
3163 
3164 		ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3165 		if (ret != 2) {
3166 			ret = -EINVAL;
3167 			goto out;
3168 		}
3169 		ret = split_huge_pages_in_file(file_path, off_start, off_end);
3170 		if (!ret)
3171 			ret = input_len;
3172 
3173 		goto out;
3174 	}
3175 
3176 	ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3177 	if (ret == 1 && pid == 1) {
3178 		split_huge_pages_all();
3179 		ret = strlen(input_buf);
3180 		goto out;
3181 	} else if (ret != 3) {
3182 		ret = -EINVAL;
3183 		goto out;
3184 	}
3185 
3186 	ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3187 	if (!ret)
3188 		ret = strlen(input_buf);
3189 out:
3190 	mutex_unlock(&split_debug_mutex);
3191 	return ret;
3192 
3193 }
3194 
3195 static const struct file_operations split_huge_pages_fops = {
3196 	.owner	 = THIS_MODULE,
3197 	.write	 = split_huge_pages_write,
3198 	.llseek  = no_llseek,
3199 };
3200 
3201 static int __init split_huge_pages_debugfs(void)
3202 {
3203 	debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3204 			    &split_huge_pages_fops);
3205 	return 0;
3206 }
3207 late_initcall(split_huge_pages_debugfs);
3208 #endif
3209 
3210 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3211 int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3212 		struct page *page)
3213 {
3214 	struct vm_area_struct *vma = pvmw->vma;
3215 	struct mm_struct *mm = vma->vm_mm;
3216 	unsigned long address = pvmw->address;
3217 	bool anon_exclusive;
3218 	pmd_t pmdval;
3219 	swp_entry_t entry;
3220 	pmd_t pmdswp;
3221 
3222 	if (!(pvmw->pmd && !pvmw->pte))
3223 		return 0;
3224 
3225 	flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3226 	pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3227 
3228 	/* See page_try_share_anon_rmap(): invalidate PMD first. */
3229 	anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
3230 	if (anon_exclusive && page_try_share_anon_rmap(page)) {
3231 		set_pmd_at(mm, address, pvmw->pmd, pmdval);
3232 		return -EBUSY;
3233 	}
3234 
3235 	if (pmd_dirty(pmdval))
3236 		set_page_dirty(page);
3237 	if (pmd_write(pmdval))
3238 		entry = make_writable_migration_entry(page_to_pfn(page));
3239 	else if (anon_exclusive)
3240 		entry = make_readable_exclusive_migration_entry(page_to_pfn(page));
3241 	else
3242 		entry = make_readable_migration_entry(page_to_pfn(page));
3243 	if (pmd_young(pmdval))
3244 		entry = make_migration_entry_young(entry);
3245 	if (pmd_dirty(pmdval))
3246 		entry = make_migration_entry_dirty(entry);
3247 	pmdswp = swp_entry_to_pmd(entry);
3248 	if (pmd_soft_dirty(pmdval))
3249 		pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3250 	set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3251 	page_remove_rmap(page, vma, true);
3252 	put_page(page);
3253 	trace_set_migration_pmd(address, pmd_val(pmdswp));
3254 
3255 	return 0;
3256 }
3257 
3258 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3259 {
3260 	struct vm_area_struct *vma = pvmw->vma;
3261 	struct mm_struct *mm = vma->vm_mm;
3262 	unsigned long address = pvmw->address;
3263 	unsigned long haddr = address & HPAGE_PMD_MASK;
3264 	pmd_t pmde;
3265 	swp_entry_t entry;
3266 
3267 	if (!(pvmw->pmd && !pvmw->pte))
3268 		return;
3269 
3270 	entry = pmd_to_swp_entry(*pvmw->pmd);
3271 	get_page(new);
3272 	pmde = mk_huge_pmd(new, READ_ONCE(vma->vm_page_prot));
3273 	if (pmd_swp_soft_dirty(*pvmw->pmd))
3274 		pmde = pmd_mksoft_dirty(pmde);
3275 	if (is_writable_migration_entry(entry))
3276 		pmde = maybe_pmd_mkwrite(pmde, vma);
3277 	if (pmd_swp_uffd_wp(*pvmw->pmd))
3278 		pmde = pmd_wrprotect(pmd_mkuffd_wp(pmde));
3279 	if (!is_migration_entry_young(entry))
3280 		pmde = pmd_mkold(pmde);
3281 	/* NOTE: this may contain setting soft-dirty on some archs */
3282 	if (PageDirty(new) && is_migration_entry_dirty(entry))
3283 		pmde = pmd_mkdirty(pmde);
3284 
3285 	if (PageAnon(new)) {
3286 		rmap_t rmap_flags = RMAP_COMPOUND;
3287 
3288 		if (!is_readable_migration_entry(entry))
3289 			rmap_flags |= RMAP_EXCLUSIVE;
3290 
3291 		page_add_anon_rmap(new, vma, haddr, rmap_flags);
3292 	} else {
3293 		page_add_file_rmap(new, vma, true);
3294 	}
3295 	VM_BUG_ON(pmd_write(pmde) && PageAnon(new) && !PageAnonExclusive(new));
3296 	set_pmd_at(mm, haddr, pvmw->pmd, pmde);
3297 
3298 	/* No need to invalidate - it was non-present before */
3299 	update_mmu_cache_pmd(vma, address, pvmw->pmd);
3300 	trace_remove_migration_pmd(address, pmd_val(pmde));
3301 }
3302 #endif
3303