xref: /openbmc/linux/mm/huge_memory.c (revision f4356947)
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_UNSUPPORTED))
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_is_huge(file_inode(vma->vm_file), vma->vm_pgoff,
123 				     !enforce_sysfs, vma->vm_mm, vm_flags);
124 
125 	/* Enforce sysfs THP requirements as necessary */
126 	if (enforce_sysfs &&
127 	    (!hugepage_flags_enabled() || (!(vm_flags & VM_HUGEPAGE) &&
128 					   !hugepage_flags_always())))
129 		return false;
130 
131 	/* Only regular file is valid */
132 	if (!in_pf && file_thp_enabled(vma))
133 		return true;
134 
135 	if (!vma_is_anonymous(vma))
136 		return false;
137 
138 	if (vma_is_temporary_stack(vma))
139 		return false;
140 
141 	/*
142 	 * THPeligible bit of smaps should show 1 for proper VMAs even
143 	 * though anon_vma is not initialized yet.
144 	 *
145 	 * Allow page fault since anon_vma may be not initialized until
146 	 * the first page fault.
147 	 */
148 	if (!vma->anon_vma)
149 		return (smaps || in_pf);
150 
151 	return true;
152 }
153 
154 static bool get_huge_zero_page(void)
155 {
156 	struct page *zero_page;
157 retry:
158 	if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
159 		return true;
160 
161 	zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
162 			HPAGE_PMD_ORDER);
163 	if (!zero_page) {
164 		count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
165 		return false;
166 	}
167 	preempt_disable();
168 	if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
169 		preempt_enable();
170 		__free_pages(zero_page, compound_order(zero_page));
171 		goto retry;
172 	}
173 	WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page));
174 
175 	/* We take additional reference here. It will be put back by shrinker */
176 	atomic_set(&huge_zero_refcount, 2);
177 	preempt_enable();
178 	count_vm_event(THP_ZERO_PAGE_ALLOC);
179 	return true;
180 }
181 
182 static void put_huge_zero_page(void)
183 {
184 	/*
185 	 * Counter should never go to zero here. Only shrinker can put
186 	 * last reference.
187 	 */
188 	BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
189 }
190 
191 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
192 {
193 	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
194 		return READ_ONCE(huge_zero_page);
195 
196 	if (!get_huge_zero_page())
197 		return NULL;
198 
199 	if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
200 		put_huge_zero_page();
201 
202 	return READ_ONCE(huge_zero_page);
203 }
204 
205 void mm_put_huge_zero_page(struct mm_struct *mm)
206 {
207 	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
208 		put_huge_zero_page();
209 }
210 
211 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
212 					struct shrink_control *sc)
213 {
214 	/* we can free zero page only if last reference remains */
215 	return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
216 }
217 
218 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
219 				       struct shrink_control *sc)
220 {
221 	if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
222 		struct page *zero_page = xchg(&huge_zero_page, NULL);
223 		BUG_ON(zero_page == NULL);
224 		WRITE_ONCE(huge_zero_pfn, ~0UL);
225 		__free_pages(zero_page, compound_order(zero_page));
226 		return HPAGE_PMD_NR;
227 	}
228 
229 	return 0;
230 }
231 
232 static struct shrinker huge_zero_page_shrinker = {
233 	.count_objects = shrink_huge_zero_page_count,
234 	.scan_objects = shrink_huge_zero_page_scan,
235 	.seeks = DEFAULT_SEEKS,
236 };
237 
238 #ifdef CONFIG_SYSFS
239 static ssize_t enabled_show(struct kobject *kobj,
240 			    struct kobj_attribute *attr, char *buf)
241 {
242 	const char *output;
243 
244 	if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
245 		output = "[always] madvise never";
246 	else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
247 			  &transparent_hugepage_flags))
248 		output = "always [madvise] never";
249 	else
250 		output = "always madvise [never]";
251 
252 	return sysfs_emit(buf, "%s\n", output);
253 }
254 
255 static ssize_t enabled_store(struct kobject *kobj,
256 			     struct kobj_attribute *attr,
257 			     const char *buf, size_t count)
258 {
259 	ssize_t ret = count;
260 
261 	if (sysfs_streq(buf, "always")) {
262 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
263 		set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
264 	} else if (sysfs_streq(buf, "madvise")) {
265 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
266 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
267 	} else if (sysfs_streq(buf, "never")) {
268 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
269 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
270 	} else
271 		ret = -EINVAL;
272 
273 	if (ret > 0) {
274 		int err = start_stop_khugepaged();
275 		if (err)
276 			ret = err;
277 	}
278 	return ret;
279 }
280 
281 static struct kobj_attribute enabled_attr = __ATTR_RW(enabled);
282 
283 ssize_t single_hugepage_flag_show(struct kobject *kobj,
284 				  struct kobj_attribute *attr, char *buf,
285 				  enum transparent_hugepage_flag flag)
286 {
287 	return sysfs_emit(buf, "%d\n",
288 			  !!test_bit(flag, &transparent_hugepage_flags));
289 }
290 
291 ssize_t single_hugepage_flag_store(struct kobject *kobj,
292 				 struct kobj_attribute *attr,
293 				 const char *buf, size_t count,
294 				 enum transparent_hugepage_flag flag)
295 {
296 	unsigned long value;
297 	int ret;
298 
299 	ret = kstrtoul(buf, 10, &value);
300 	if (ret < 0)
301 		return ret;
302 	if (value > 1)
303 		return -EINVAL;
304 
305 	if (value)
306 		set_bit(flag, &transparent_hugepage_flags);
307 	else
308 		clear_bit(flag, &transparent_hugepage_flags);
309 
310 	return count;
311 }
312 
313 static ssize_t defrag_show(struct kobject *kobj,
314 			   struct kobj_attribute *attr, char *buf)
315 {
316 	const char *output;
317 
318 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
319 		     &transparent_hugepage_flags))
320 		output = "[always] defer defer+madvise madvise never";
321 	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
322 			  &transparent_hugepage_flags))
323 		output = "always [defer] defer+madvise madvise never";
324 	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
325 			  &transparent_hugepage_flags))
326 		output = "always defer [defer+madvise] madvise never";
327 	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
328 			  &transparent_hugepage_flags))
329 		output = "always defer defer+madvise [madvise] never";
330 	else
331 		output = "always defer defer+madvise madvise [never]";
332 
333 	return sysfs_emit(buf, "%s\n", output);
334 }
335 
336 static ssize_t defrag_store(struct kobject *kobj,
337 			    struct kobj_attribute *attr,
338 			    const char *buf, size_t count)
339 {
340 	if (sysfs_streq(buf, "always")) {
341 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
342 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
343 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
344 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
345 	} else if (sysfs_streq(buf, "defer+madvise")) {
346 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
347 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
348 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
349 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
350 	} else if (sysfs_streq(buf, "defer")) {
351 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
352 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
353 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
354 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
355 	} else if (sysfs_streq(buf, "madvise")) {
356 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
357 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
358 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
359 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
360 	} else if (sysfs_streq(buf, "never")) {
361 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
362 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
363 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
364 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
365 	} else
366 		return -EINVAL;
367 
368 	return count;
369 }
370 static struct kobj_attribute defrag_attr = __ATTR_RW(defrag);
371 
372 static ssize_t use_zero_page_show(struct kobject *kobj,
373 				  struct kobj_attribute *attr, char *buf)
374 {
375 	return single_hugepage_flag_show(kobj, attr, buf,
376 					 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
377 }
378 static ssize_t use_zero_page_store(struct kobject *kobj,
379 		struct kobj_attribute *attr, const char *buf, size_t count)
380 {
381 	return single_hugepage_flag_store(kobj, attr, buf, count,
382 				 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
383 }
384 static struct kobj_attribute use_zero_page_attr = __ATTR_RW(use_zero_page);
385 
386 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
387 				   struct kobj_attribute *attr, char *buf)
388 {
389 	return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
390 }
391 static struct kobj_attribute hpage_pmd_size_attr =
392 	__ATTR_RO(hpage_pmd_size);
393 
394 static struct attribute *hugepage_attr[] = {
395 	&enabled_attr.attr,
396 	&defrag_attr.attr,
397 	&use_zero_page_attr.attr,
398 	&hpage_pmd_size_attr.attr,
399 #ifdef CONFIG_SHMEM
400 	&shmem_enabled_attr.attr,
401 #endif
402 	NULL,
403 };
404 
405 static const struct attribute_group hugepage_attr_group = {
406 	.attrs = hugepage_attr,
407 };
408 
409 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
410 {
411 	int err;
412 
413 	*hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
414 	if (unlikely(!*hugepage_kobj)) {
415 		pr_err("failed to create transparent hugepage kobject\n");
416 		return -ENOMEM;
417 	}
418 
419 	err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
420 	if (err) {
421 		pr_err("failed to register transparent hugepage group\n");
422 		goto delete_obj;
423 	}
424 
425 	err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
426 	if (err) {
427 		pr_err("failed to register transparent hugepage group\n");
428 		goto remove_hp_group;
429 	}
430 
431 	return 0;
432 
433 remove_hp_group:
434 	sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
435 delete_obj:
436 	kobject_put(*hugepage_kobj);
437 	return err;
438 }
439 
440 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
441 {
442 	sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
443 	sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
444 	kobject_put(hugepage_kobj);
445 }
446 #else
447 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
448 {
449 	return 0;
450 }
451 
452 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
453 {
454 }
455 #endif /* CONFIG_SYSFS */
456 
457 static int __init hugepage_init(void)
458 {
459 	int err;
460 	struct kobject *hugepage_kobj;
461 
462 	if (!has_transparent_hugepage()) {
463 		transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_UNSUPPORTED;
464 		return -EINVAL;
465 	}
466 
467 	/*
468 	 * hugepages can't be allocated by the buddy allocator
469 	 */
470 	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER > MAX_ORDER);
471 	/*
472 	 * we use page->mapping and page->index in second tail page
473 	 * as list_head: assuming THP order >= 2
474 	 */
475 	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
476 
477 	err = hugepage_init_sysfs(&hugepage_kobj);
478 	if (err)
479 		goto err_sysfs;
480 
481 	err = khugepaged_init();
482 	if (err)
483 		goto err_slab;
484 
485 	err = register_shrinker(&huge_zero_page_shrinker, "thp-zero");
486 	if (err)
487 		goto err_hzp_shrinker;
488 	err = register_shrinker(&deferred_split_shrinker, "thp-deferred_split");
489 	if (err)
490 		goto err_split_shrinker;
491 
492 	/*
493 	 * By default disable transparent hugepages on smaller systems,
494 	 * where the extra memory used could hurt more than TLB overhead
495 	 * is likely to save.  The admin can still enable it through /sys.
496 	 */
497 	if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
498 		transparent_hugepage_flags = 0;
499 		return 0;
500 	}
501 
502 	err = start_stop_khugepaged();
503 	if (err)
504 		goto err_khugepaged;
505 
506 	return 0;
507 err_khugepaged:
508 	unregister_shrinker(&deferred_split_shrinker);
509 err_split_shrinker:
510 	unregister_shrinker(&huge_zero_page_shrinker);
511 err_hzp_shrinker:
512 	khugepaged_destroy();
513 err_slab:
514 	hugepage_exit_sysfs(hugepage_kobj);
515 err_sysfs:
516 	return err;
517 }
518 subsys_initcall(hugepage_init);
519 
520 static int __init setup_transparent_hugepage(char *str)
521 {
522 	int ret = 0;
523 	if (!str)
524 		goto out;
525 	if (!strcmp(str, "always")) {
526 		set_bit(TRANSPARENT_HUGEPAGE_FLAG,
527 			&transparent_hugepage_flags);
528 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
529 			  &transparent_hugepage_flags);
530 		ret = 1;
531 	} else if (!strcmp(str, "madvise")) {
532 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
533 			  &transparent_hugepage_flags);
534 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
535 			&transparent_hugepage_flags);
536 		ret = 1;
537 	} else if (!strcmp(str, "never")) {
538 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
539 			  &transparent_hugepage_flags);
540 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
541 			  &transparent_hugepage_flags);
542 		ret = 1;
543 	}
544 out:
545 	if (!ret)
546 		pr_warn("transparent_hugepage= cannot parse, ignored\n");
547 	return ret;
548 }
549 __setup("transparent_hugepage=", setup_transparent_hugepage);
550 
551 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
552 {
553 	if (likely(vma->vm_flags & VM_WRITE))
554 		pmd = pmd_mkwrite(pmd);
555 	return pmd;
556 }
557 
558 #ifdef CONFIG_MEMCG
559 static inline
560 struct deferred_split *get_deferred_split_queue(struct folio *folio)
561 {
562 	struct mem_cgroup *memcg = folio_memcg(folio);
563 	struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));
564 
565 	if (memcg)
566 		return &memcg->deferred_split_queue;
567 	else
568 		return &pgdat->deferred_split_queue;
569 }
570 #else
571 static inline
572 struct deferred_split *get_deferred_split_queue(struct folio *folio)
573 {
574 	struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));
575 
576 	return &pgdat->deferred_split_queue;
577 }
578 #endif
579 
580 void prep_transhuge_page(struct page *page)
581 {
582 	struct folio *folio = (struct folio *)page;
583 
584 	VM_BUG_ON_FOLIO(folio_order(folio) < 2, folio);
585 	INIT_LIST_HEAD(&folio->_deferred_list);
586 	set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
587 }
588 
589 static inline bool is_transparent_hugepage(struct page *page)
590 {
591 	struct folio *folio;
592 
593 	if (!PageCompound(page))
594 		return false;
595 
596 	folio = page_folio(page);
597 	return is_huge_zero_page(&folio->page) ||
598 	       folio->_folio_dtor == TRANSHUGE_PAGE_DTOR;
599 }
600 
601 static unsigned long __thp_get_unmapped_area(struct file *filp,
602 		unsigned long addr, unsigned long len,
603 		loff_t off, unsigned long flags, unsigned long size)
604 {
605 	loff_t off_end = off + len;
606 	loff_t off_align = round_up(off, size);
607 	unsigned long len_pad, ret;
608 
609 	if (off_end <= off_align || (off_end - off_align) < size)
610 		return 0;
611 
612 	len_pad = len + size;
613 	if (len_pad < len || (off + len_pad) < off)
614 		return 0;
615 
616 	ret = current->mm->get_unmapped_area(filp, addr, len_pad,
617 					      off >> PAGE_SHIFT, flags);
618 
619 	/*
620 	 * The failure might be due to length padding. The caller will retry
621 	 * without the padding.
622 	 */
623 	if (IS_ERR_VALUE(ret))
624 		return 0;
625 
626 	/*
627 	 * Do not try to align to THP boundary if allocation at the address
628 	 * hint succeeds.
629 	 */
630 	if (ret == addr)
631 		return addr;
632 
633 	ret += (off - ret) & (size - 1);
634 	return ret;
635 }
636 
637 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
638 		unsigned long len, unsigned long pgoff, unsigned long flags)
639 {
640 	unsigned long ret;
641 	loff_t off = (loff_t)pgoff << PAGE_SHIFT;
642 
643 	ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
644 	if (ret)
645 		return ret;
646 
647 	return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
648 }
649 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
650 
651 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
652 			struct page *page, gfp_t gfp)
653 {
654 	struct vm_area_struct *vma = vmf->vma;
655 	struct folio *folio = page_folio(page);
656 	pgtable_t pgtable;
657 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
658 	vm_fault_t ret = 0;
659 
660 	VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
661 
662 	if (mem_cgroup_charge(folio, vma->vm_mm, gfp)) {
663 		folio_put(folio);
664 		count_vm_event(THP_FAULT_FALLBACK);
665 		count_vm_event(THP_FAULT_FALLBACK_CHARGE);
666 		return VM_FAULT_FALLBACK;
667 	}
668 	folio_throttle_swaprate(folio, 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 __folio_mark_uptodate makes sure that
679 	 * clear_huge_page writes become visible before the set_pmd_at()
680 	 * write.
681 	 */
682 	__folio_mark_uptodate(folio);
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 			folio_put(folio);
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 		folio_add_new_anon_rmap(folio, vma, haddr);
707 		folio_add_lru_vma(folio, 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 	folio_put(folio);
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 - insert a pmd size pfn
889  * @vmf: Structure describing the fault
890  * @pfn: pfn to insert
891  * @write: whether it's a write fault
892  *
893  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info.
894  *
895  * Return: vm_fault_t value.
896  */
897 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)
898 {
899 	unsigned long addr = vmf->address & PMD_MASK;
900 	struct vm_area_struct *vma = vmf->vma;
901 	pgprot_t pgprot = vma->vm_page_prot;
902 	pgtable_t pgtable = NULL;
903 
904 	/*
905 	 * If we had pmd_special, we could avoid all these restrictions,
906 	 * but we need to be consistent with PTEs and architectures that
907 	 * can't support a 'special' bit.
908 	 */
909 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
910 			!pfn_t_devmap(pfn));
911 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
912 						(VM_PFNMAP|VM_MIXEDMAP));
913 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
914 
915 	if (addr < vma->vm_start || addr >= vma->vm_end)
916 		return VM_FAULT_SIGBUS;
917 
918 	if (arch_needs_pgtable_deposit()) {
919 		pgtable = pte_alloc_one(vma->vm_mm);
920 		if (!pgtable)
921 			return VM_FAULT_OOM;
922 	}
923 
924 	track_pfn_insert(vma, &pgprot, pfn);
925 
926 	insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
927 	return VM_FAULT_NOPAGE;
928 }
929 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
930 
931 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
932 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
933 {
934 	if (likely(vma->vm_flags & VM_WRITE))
935 		pud = pud_mkwrite(pud);
936 	return pud;
937 }
938 
939 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
940 		pud_t *pud, pfn_t pfn, bool write)
941 {
942 	struct mm_struct *mm = vma->vm_mm;
943 	pgprot_t prot = vma->vm_page_prot;
944 	pud_t entry;
945 	spinlock_t *ptl;
946 
947 	ptl = pud_lock(mm, pud);
948 	if (!pud_none(*pud)) {
949 		if (write) {
950 			if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
951 				WARN_ON_ONCE(!is_huge_zero_pud(*pud));
952 				goto out_unlock;
953 			}
954 			entry = pud_mkyoung(*pud);
955 			entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
956 			if (pudp_set_access_flags(vma, addr, pud, entry, 1))
957 				update_mmu_cache_pud(vma, addr, pud);
958 		}
959 		goto out_unlock;
960 	}
961 
962 	entry = pud_mkhuge(pfn_t_pud(pfn, prot));
963 	if (pfn_t_devmap(pfn))
964 		entry = pud_mkdevmap(entry);
965 	if (write) {
966 		entry = pud_mkyoung(pud_mkdirty(entry));
967 		entry = maybe_pud_mkwrite(entry, vma);
968 	}
969 	set_pud_at(mm, addr, pud, entry);
970 	update_mmu_cache_pud(vma, addr, pud);
971 
972 out_unlock:
973 	spin_unlock(ptl);
974 }
975 
976 /**
977  * vmf_insert_pfn_pud - insert a pud size pfn
978  * @vmf: Structure describing the fault
979  * @pfn: pfn to insert
980  * @write: whether it's a write fault
981  *
982  * Insert a pud size pfn. See vmf_insert_pfn() for additional info.
983  *
984  * Return: vm_fault_t value.
985  */
986 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)
987 {
988 	unsigned long addr = vmf->address & PUD_MASK;
989 	struct vm_area_struct *vma = vmf->vma;
990 	pgprot_t pgprot = vma->vm_page_prot;
991 
992 	/*
993 	 * If we had pud_special, we could avoid all these restrictions,
994 	 * but we need to be consistent with PTEs and architectures that
995 	 * can't support a 'special' bit.
996 	 */
997 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
998 			!pfn_t_devmap(pfn));
999 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1000 						(VM_PFNMAP|VM_MIXEDMAP));
1001 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1002 
1003 	if (addr < vma->vm_start || addr >= vma->vm_end)
1004 		return VM_FAULT_SIGBUS;
1005 
1006 	track_pfn_insert(vma, &pgprot, pfn);
1007 
1008 	insert_pfn_pud(vma, addr, vmf->pud, pfn, write);
1009 	return VM_FAULT_NOPAGE;
1010 }
1011 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
1012 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1013 
1014 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1015 		      pmd_t *pmd, bool write)
1016 {
1017 	pmd_t _pmd;
1018 
1019 	_pmd = pmd_mkyoung(*pmd);
1020 	if (write)
1021 		_pmd = pmd_mkdirty(_pmd);
1022 	if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1023 				  pmd, _pmd, write))
1024 		update_mmu_cache_pmd(vma, addr, pmd);
1025 }
1026 
1027 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1028 		pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
1029 {
1030 	unsigned long pfn = pmd_pfn(*pmd);
1031 	struct mm_struct *mm = vma->vm_mm;
1032 	struct page *page;
1033 	int ret;
1034 
1035 	assert_spin_locked(pmd_lockptr(mm, pmd));
1036 
1037 	if (flags & FOLL_WRITE && !pmd_write(*pmd))
1038 		return NULL;
1039 
1040 	if (pmd_present(*pmd) && pmd_devmap(*pmd))
1041 		/* pass */;
1042 	else
1043 		return NULL;
1044 
1045 	if (flags & FOLL_TOUCH)
1046 		touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1047 
1048 	/*
1049 	 * device mapped pages can only be returned if the
1050 	 * caller will manage the page reference count.
1051 	 */
1052 	if (!(flags & (FOLL_GET | FOLL_PIN)))
1053 		return ERR_PTR(-EEXIST);
1054 
1055 	pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1056 	*pgmap = get_dev_pagemap(pfn, *pgmap);
1057 	if (!*pgmap)
1058 		return ERR_PTR(-EFAULT);
1059 	page = pfn_to_page(pfn);
1060 	ret = try_grab_page(page, flags);
1061 	if (ret)
1062 		page = ERR_PTR(ret);
1063 
1064 	return page;
1065 }
1066 
1067 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1068 		  pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1069 		  struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1070 {
1071 	spinlock_t *dst_ptl, *src_ptl;
1072 	struct page *src_page;
1073 	pmd_t pmd;
1074 	pgtable_t pgtable = NULL;
1075 	int ret = -ENOMEM;
1076 
1077 	/* Skip if can be re-fill on fault */
1078 	if (!vma_is_anonymous(dst_vma))
1079 		return 0;
1080 
1081 	pgtable = pte_alloc_one(dst_mm);
1082 	if (unlikely(!pgtable))
1083 		goto out;
1084 
1085 	dst_ptl = pmd_lock(dst_mm, dst_pmd);
1086 	src_ptl = pmd_lockptr(src_mm, src_pmd);
1087 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1088 
1089 	ret = -EAGAIN;
1090 	pmd = *src_pmd;
1091 
1092 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1093 	if (unlikely(is_swap_pmd(pmd))) {
1094 		swp_entry_t entry = pmd_to_swp_entry(pmd);
1095 
1096 		VM_BUG_ON(!is_pmd_migration_entry(pmd));
1097 		if (!is_readable_migration_entry(entry)) {
1098 			entry = make_readable_migration_entry(
1099 							swp_offset(entry));
1100 			pmd = swp_entry_to_pmd(entry);
1101 			if (pmd_swp_soft_dirty(*src_pmd))
1102 				pmd = pmd_swp_mksoft_dirty(pmd);
1103 			if (pmd_swp_uffd_wp(*src_pmd))
1104 				pmd = pmd_swp_mkuffd_wp(pmd);
1105 			set_pmd_at(src_mm, addr, src_pmd, pmd);
1106 		}
1107 		add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1108 		mm_inc_nr_ptes(dst_mm);
1109 		pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1110 		if (!userfaultfd_wp(dst_vma))
1111 			pmd = pmd_swp_clear_uffd_wp(pmd);
1112 		set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1113 		ret = 0;
1114 		goto out_unlock;
1115 	}
1116 #endif
1117 
1118 	if (unlikely(!pmd_trans_huge(pmd))) {
1119 		pte_free(dst_mm, pgtable);
1120 		goto out_unlock;
1121 	}
1122 	/*
1123 	 * When page table lock is held, the huge zero pmd should not be
1124 	 * under splitting since we don't split the page itself, only pmd to
1125 	 * a page table.
1126 	 */
1127 	if (is_huge_zero_pmd(pmd)) {
1128 		/*
1129 		 * get_huge_zero_page() will never allocate a new page here,
1130 		 * since we already have a zero page to copy. It just takes a
1131 		 * reference.
1132 		 */
1133 		mm_get_huge_zero_page(dst_mm);
1134 		goto out_zero_page;
1135 	}
1136 
1137 	src_page = pmd_page(pmd);
1138 	VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1139 
1140 	get_page(src_page);
1141 	if (unlikely(page_try_dup_anon_rmap(src_page, true, src_vma))) {
1142 		/* Page maybe pinned: split and retry the fault on PTEs. */
1143 		put_page(src_page);
1144 		pte_free(dst_mm, pgtable);
1145 		spin_unlock(src_ptl);
1146 		spin_unlock(dst_ptl);
1147 		__split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1148 		return -EAGAIN;
1149 	}
1150 	add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1151 out_zero_page:
1152 	mm_inc_nr_ptes(dst_mm);
1153 	pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1154 	pmdp_set_wrprotect(src_mm, addr, src_pmd);
1155 	if (!userfaultfd_wp(dst_vma))
1156 		pmd = pmd_clear_uffd_wp(pmd);
1157 	pmd = pmd_mkold(pmd_wrprotect(pmd));
1158 	set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1159 
1160 	ret = 0;
1161 out_unlock:
1162 	spin_unlock(src_ptl);
1163 	spin_unlock(dst_ptl);
1164 out:
1165 	return ret;
1166 }
1167 
1168 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1169 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1170 		      pud_t *pud, bool write)
1171 {
1172 	pud_t _pud;
1173 
1174 	_pud = pud_mkyoung(*pud);
1175 	if (write)
1176 		_pud = pud_mkdirty(_pud);
1177 	if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1178 				  pud, _pud, write))
1179 		update_mmu_cache_pud(vma, addr, pud);
1180 }
1181 
1182 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1183 		pud_t *pud, int flags, struct dev_pagemap **pgmap)
1184 {
1185 	unsigned long pfn = pud_pfn(*pud);
1186 	struct mm_struct *mm = vma->vm_mm;
1187 	struct page *page;
1188 	int ret;
1189 
1190 	assert_spin_locked(pud_lockptr(mm, pud));
1191 
1192 	if (flags & FOLL_WRITE && !pud_write(*pud))
1193 		return NULL;
1194 
1195 	if (pud_present(*pud) && pud_devmap(*pud))
1196 		/* pass */;
1197 	else
1198 		return NULL;
1199 
1200 	if (flags & FOLL_TOUCH)
1201 		touch_pud(vma, addr, pud, flags & FOLL_WRITE);
1202 
1203 	/*
1204 	 * device mapped pages can only be returned if the
1205 	 * caller will manage the page reference count.
1206 	 *
1207 	 * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1208 	 */
1209 	if (!(flags & (FOLL_GET | FOLL_PIN)))
1210 		return ERR_PTR(-EEXIST);
1211 
1212 	pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1213 	*pgmap = get_dev_pagemap(pfn, *pgmap);
1214 	if (!*pgmap)
1215 		return ERR_PTR(-EFAULT);
1216 	page = pfn_to_page(pfn);
1217 
1218 	ret = try_grab_page(page, flags);
1219 	if (ret)
1220 		page = ERR_PTR(ret);
1221 
1222 	return page;
1223 }
1224 
1225 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1226 		  pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1227 		  struct vm_area_struct *vma)
1228 {
1229 	spinlock_t *dst_ptl, *src_ptl;
1230 	pud_t pud;
1231 	int ret;
1232 
1233 	dst_ptl = pud_lock(dst_mm, dst_pud);
1234 	src_ptl = pud_lockptr(src_mm, src_pud);
1235 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1236 
1237 	ret = -EAGAIN;
1238 	pud = *src_pud;
1239 	if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1240 		goto out_unlock;
1241 
1242 	/*
1243 	 * When page table lock is held, the huge zero pud should not be
1244 	 * under splitting since we don't split the page itself, only pud to
1245 	 * a page table.
1246 	 */
1247 	if (is_huge_zero_pud(pud)) {
1248 		/* No huge zero pud yet */
1249 	}
1250 
1251 	/*
1252 	 * TODO: once we support anonymous pages, use page_try_dup_anon_rmap()
1253 	 * and split if duplicating fails.
1254 	 */
1255 	pudp_set_wrprotect(src_mm, addr, src_pud);
1256 	pud = pud_mkold(pud_wrprotect(pud));
1257 	set_pud_at(dst_mm, addr, dst_pud, pud);
1258 
1259 	ret = 0;
1260 out_unlock:
1261 	spin_unlock(src_ptl);
1262 	spin_unlock(dst_ptl);
1263 	return ret;
1264 }
1265 
1266 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1267 {
1268 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1269 
1270 	vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1271 	if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1272 		goto unlock;
1273 
1274 	touch_pud(vmf->vma, vmf->address, vmf->pud, write);
1275 unlock:
1276 	spin_unlock(vmf->ptl);
1277 }
1278 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1279 
1280 void huge_pmd_set_accessed(struct vm_fault *vmf)
1281 {
1282 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1283 
1284 	vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1285 	if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd)))
1286 		goto unlock;
1287 
1288 	touch_pmd(vmf->vma, vmf->address, vmf->pmd, write);
1289 
1290 unlock:
1291 	spin_unlock(vmf->ptl);
1292 }
1293 
1294 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1295 {
1296 	const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
1297 	struct vm_area_struct *vma = vmf->vma;
1298 	struct folio *folio;
1299 	struct page *page;
1300 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1301 	pmd_t orig_pmd = vmf->orig_pmd;
1302 
1303 	vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1304 	VM_BUG_ON_VMA(!vma->anon_vma, vma);
1305 
1306 	if (is_huge_zero_pmd(orig_pmd))
1307 		goto fallback;
1308 
1309 	spin_lock(vmf->ptl);
1310 
1311 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1312 		spin_unlock(vmf->ptl);
1313 		return 0;
1314 	}
1315 
1316 	page = pmd_page(orig_pmd);
1317 	folio = page_folio(page);
1318 	VM_BUG_ON_PAGE(!PageHead(page), page);
1319 
1320 	/* Early check when only holding the PT lock. */
1321 	if (PageAnonExclusive(page))
1322 		goto reuse;
1323 
1324 	if (!folio_trylock(folio)) {
1325 		folio_get(folio);
1326 		spin_unlock(vmf->ptl);
1327 		folio_lock(folio);
1328 		spin_lock(vmf->ptl);
1329 		if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1330 			spin_unlock(vmf->ptl);
1331 			folio_unlock(folio);
1332 			folio_put(folio);
1333 			return 0;
1334 		}
1335 		folio_put(folio);
1336 	}
1337 
1338 	/* Recheck after temporarily dropping the PT lock. */
1339 	if (PageAnonExclusive(page)) {
1340 		folio_unlock(folio);
1341 		goto reuse;
1342 	}
1343 
1344 	/*
1345 	 * See do_wp_page(): we can only reuse the folio exclusively if
1346 	 * there are no additional references. Note that we always drain
1347 	 * the LRU pagevecs immediately after adding a THP.
1348 	 */
1349 	if (folio_ref_count(folio) >
1350 			1 + folio_test_swapcache(folio) * folio_nr_pages(folio))
1351 		goto unlock_fallback;
1352 	if (folio_test_swapcache(folio))
1353 		folio_free_swap(folio);
1354 	if (folio_ref_count(folio) == 1) {
1355 		pmd_t entry;
1356 
1357 		page_move_anon_rmap(page, vma);
1358 		folio_unlock(folio);
1359 reuse:
1360 		if (unlikely(unshare)) {
1361 			spin_unlock(vmf->ptl);
1362 			return 0;
1363 		}
1364 		entry = pmd_mkyoung(orig_pmd);
1365 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1366 		if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1367 			update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1368 		spin_unlock(vmf->ptl);
1369 		return 0;
1370 	}
1371 
1372 unlock_fallback:
1373 	folio_unlock(folio);
1374 	spin_unlock(vmf->ptl);
1375 fallback:
1376 	__split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1377 	return VM_FAULT_FALLBACK;
1378 }
1379 
1380 static inline bool can_change_pmd_writable(struct vm_area_struct *vma,
1381 					   unsigned long addr, pmd_t pmd)
1382 {
1383 	struct page *page;
1384 
1385 	if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE)))
1386 		return false;
1387 
1388 	/* Don't touch entries that are not even readable (NUMA hinting). */
1389 	if (pmd_protnone(pmd))
1390 		return false;
1391 
1392 	/* Do we need write faults for softdirty tracking? */
1393 	if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1394 		return false;
1395 
1396 	/* Do we need write faults for uffd-wp tracking? */
1397 	if (userfaultfd_huge_pmd_wp(vma, pmd))
1398 		return false;
1399 
1400 	if (!(vma->vm_flags & VM_SHARED)) {
1401 		/* See can_change_pte_writable(). */
1402 		page = vm_normal_page_pmd(vma, addr, pmd);
1403 		return page && PageAnon(page) && PageAnonExclusive(page);
1404 	}
1405 
1406 	/* See can_change_pte_writable(). */
1407 	return pmd_dirty(pmd);
1408 }
1409 
1410 /* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */
1411 static inline bool can_follow_write_pmd(pmd_t pmd, struct page *page,
1412 					struct vm_area_struct *vma,
1413 					unsigned int flags)
1414 {
1415 	/* If the pmd is writable, we can write to the page. */
1416 	if (pmd_write(pmd))
1417 		return true;
1418 
1419 	/* Maybe FOLL_FORCE is set to override it? */
1420 	if (!(flags & FOLL_FORCE))
1421 		return false;
1422 
1423 	/* But FOLL_FORCE has no effect on shared mappings */
1424 	if (vma->vm_flags & (VM_MAYSHARE | VM_SHARED))
1425 		return false;
1426 
1427 	/* ... or read-only private ones */
1428 	if (!(vma->vm_flags & VM_MAYWRITE))
1429 		return false;
1430 
1431 	/* ... or already writable ones that just need to take a write fault */
1432 	if (vma->vm_flags & VM_WRITE)
1433 		return false;
1434 
1435 	/*
1436 	 * See can_change_pte_writable(): we broke COW and could map the page
1437 	 * writable if we have an exclusive anonymous page ...
1438 	 */
1439 	if (!page || !PageAnon(page) || !PageAnonExclusive(page))
1440 		return false;
1441 
1442 	/* ... and a write-fault isn't required for other reasons. */
1443 	if (vma_soft_dirty_enabled(vma) && !pmd_soft_dirty(pmd))
1444 		return false;
1445 	return !userfaultfd_huge_pmd_wp(vma, pmd);
1446 }
1447 
1448 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1449 				   unsigned long addr,
1450 				   pmd_t *pmd,
1451 				   unsigned int flags)
1452 {
1453 	struct mm_struct *mm = vma->vm_mm;
1454 	struct page *page;
1455 	int ret;
1456 
1457 	assert_spin_locked(pmd_lockptr(mm, pmd));
1458 
1459 	page = pmd_page(*pmd);
1460 	VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1461 
1462 	if ((flags & FOLL_WRITE) &&
1463 	    !can_follow_write_pmd(*pmd, page, vma, flags))
1464 		return NULL;
1465 
1466 	/* Avoid dumping huge zero page */
1467 	if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1468 		return ERR_PTR(-EFAULT);
1469 
1470 	/* Full NUMA hinting faults to serialise migration in fault paths */
1471 	if (pmd_protnone(*pmd) && !gup_can_follow_protnone(flags))
1472 		return NULL;
1473 
1474 	if (!pmd_write(*pmd) && gup_must_unshare(vma, flags, page))
1475 		return ERR_PTR(-EMLINK);
1476 
1477 	VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
1478 			!PageAnonExclusive(page), page);
1479 
1480 	ret = try_grab_page(page, flags);
1481 	if (ret)
1482 		return ERR_PTR(ret);
1483 
1484 	if (flags & FOLL_TOUCH)
1485 		touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);
1486 
1487 	page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1488 	VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1489 
1490 	return page;
1491 }
1492 
1493 /* NUMA hinting page fault entry point for trans huge pmds */
1494 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
1495 {
1496 	struct vm_area_struct *vma = vmf->vma;
1497 	pmd_t oldpmd = vmf->orig_pmd;
1498 	pmd_t pmd;
1499 	struct page *page;
1500 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1501 	int page_nid = NUMA_NO_NODE;
1502 	int target_nid, last_cpupid = (-1 & LAST_CPUPID_MASK);
1503 	bool migrated = false, writable = false;
1504 	int flags = 0;
1505 
1506 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1507 	if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1508 		spin_unlock(vmf->ptl);
1509 		goto out;
1510 	}
1511 
1512 	pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1513 
1514 	/*
1515 	 * Detect now whether the PMD could be writable; this information
1516 	 * is only valid while holding the PT lock.
1517 	 */
1518 	writable = pmd_write(pmd);
1519 	if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&
1520 	    can_change_pmd_writable(vma, vmf->address, pmd))
1521 		writable = true;
1522 
1523 	page = vm_normal_page_pmd(vma, haddr, pmd);
1524 	if (!page)
1525 		goto out_map;
1526 
1527 	/* See similar comment in do_numa_page for explanation */
1528 	if (!writable)
1529 		flags |= TNF_NO_GROUP;
1530 
1531 	page_nid = page_to_nid(page);
1532 	/*
1533 	 * For memory tiering mode, cpupid of slow memory page is used
1534 	 * to record page access time.  So use default value.
1535 	 */
1536 	if (node_is_toptier(page_nid))
1537 		last_cpupid = page_cpupid_last(page);
1538 	target_nid = numa_migrate_prep(page, vma, haddr, page_nid,
1539 				       &flags);
1540 
1541 	if (target_nid == NUMA_NO_NODE) {
1542 		put_page(page);
1543 		goto out_map;
1544 	}
1545 
1546 	spin_unlock(vmf->ptl);
1547 	writable = false;
1548 
1549 	migrated = migrate_misplaced_page(page, vma, target_nid);
1550 	if (migrated) {
1551 		flags |= TNF_MIGRATED;
1552 		page_nid = target_nid;
1553 	} else {
1554 		flags |= TNF_MIGRATE_FAIL;
1555 		vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1556 		if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1557 			spin_unlock(vmf->ptl);
1558 			goto out;
1559 		}
1560 		goto out_map;
1561 	}
1562 
1563 out:
1564 	if (page_nid != NUMA_NO_NODE)
1565 		task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1566 				flags);
1567 
1568 	return 0;
1569 
1570 out_map:
1571 	/* Restore the PMD */
1572 	pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1573 	pmd = pmd_mkyoung(pmd);
1574 	if (writable)
1575 		pmd = pmd_mkwrite(pmd);
1576 	set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1577 	update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1578 	spin_unlock(vmf->ptl);
1579 	goto out;
1580 }
1581 
1582 /*
1583  * Return true if we do MADV_FREE successfully on entire pmd page.
1584  * Otherwise, return false.
1585  */
1586 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1587 		pmd_t *pmd, unsigned long addr, unsigned long next)
1588 {
1589 	spinlock_t *ptl;
1590 	pmd_t orig_pmd;
1591 	struct folio *folio;
1592 	struct mm_struct *mm = tlb->mm;
1593 	bool ret = false;
1594 
1595 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1596 
1597 	ptl = pmd_trans_huge_lock(pmd, vma);
1598 	if (!ptl)
1599 		goto out_unlocked;
1600 
1601 	orig_pmd = *pmd;
1602 	if (is_huge_zero_pmd(orig_pmd))
1603 		goto out;
1604 
1605 	if (unlikely(!pmd_present(orig_pmd))) {
1606 		VM_BUG_ON(thp_migration_supported() &&
1607 				  !is_pmd_migration_entry(orig_pmd));
1608 		goto out;
1609 	}
1610 
1611 	folio = pfn_folio(pmd_pfn(orig_pmd));
1612 	/*
1613 	 * If other processes are mapping this folio, we couldn't discard
1614 	 * the folio unless they all do MADV_FREE so let's skip the folio.
1615 	 */
1616 	if (folio_mapcount(folio) != 1)
1617 		goto out;
1618 
1619 	if (!folio_trylock(folio))
1620 		goto out;
1621 
1622 	/*
1623 	 * If user want to discard part-pages of THP, split it so MADV_FREE
1624 	 * will deactivate only them.
1625 	 */
1626 	if (next - addr != HPAGE_PMD_SIZE) {
1627 		folio_get(folio);
1628 		spin_unlock(ptl);
1629 		split_folio(folio);
1630 		folio_unlock(folio);
1631 		folio_put(folio);
1632 		goto out_unlocked;
1633 	}
1634 
1635 	if (folio_test_dirty(folio))
1636 		folio_clear_dirty(folio);
1637 	folio_unlock(folio);
1638 
1639 	if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1640 		pmdp_invalidate(vma, addr, pmd);
1641 		orig_pmd = pmd_mkold(orig_pmd);
1642 		orig_pmd = pmd_mkclean(orig_pmd);
1643 
1644 		set_pmd_at(mm, addr, pmd, orig_pmd);
1645 		tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1646 	}
1647 
1648 	folio_mark_lazyfree(folio);
1649 	ret = true;
1650 out:
1651 	spin_unlock(ptl);
1652 out_unlocked:
1653 	return ret;
1654 }
1655 
1656 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1657 {
1658 	pgtable_t pgtable;
1659 
1660 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1661 	pte_free(mm, pgtable);
1662 	mm_dec_nr_ptes(mm);
1663 }
1664 
1665 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1666 		 pmd_t *pmd, unsigned long addr)
1667 {
1668 	pmd_t orig_pmd;
1669 	spinlock_t *ptl;
1670 
1671 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1672 
1673 	ptl = __pmd_trans_huge_lock(pmd, vma);
1674 	if (!ptl)
1675 		return 0;
1676 	/*
1677 	 * For architectures like ppc64 we look at deposited pgtable
1678 	 * when calling pmdp_huge_get_and_clear. So do the
1679 	 * pgtable_trans_huge_withdraw after finishing pmdp related
1680 	 * operations.
1681 	 */
1682 	orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1683 						tlb->fullmm);
1684 	tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1685 	if (vma_is_special_huge(vma)) {
1686 		if (arch_needs_pgtable_deposit())
1687 			zap_deposited_table(tlb->mm, pmd);
1688 		spin_unlock(ptl);
1689 	} else if (is_huge_zero_pmd(orig_pmd)) {
1690 		zap_deposited_table(tlb->mm, pmd);
1691 		spin_unlock(ptl);
1692 	} else {
1693 		struct page *page = NULL;
1694 		int flush_needed = 1;
1695 
1696 		if (pmd_present(orig_pmd)) {
1697 			page = pmd_page(orig_pmd);
1698 			page_remove_rmap(page, vma, true);
1699 			VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1700 			VM_BUG_ON_PAGE(!PageHead(page), page);
1701 		} else if (thp_migration_supported()) {
1702 			swp_entry_t entry;
1703 
1704 			VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1705 			entry = pmd_to_swp_entry(orig_pmd);
1706 			page = pfn_swap_entry_to_page(entry);
1707 			flush_needed = 0;
1708 		} else
1709 			WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1710 
1711 		if (PageAnon(page)) {
1712 			zap_deposited_table(tlb->mm, pmd);
1713 			add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1714 		} else {
1715 			if (arch_needs_pgtable_deposit())
1716 				zap_deposited_table(tlb->mm, pmd);
1717 			add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1718 		}
1719 
1720 		spin_unlock(ptl);
1721 		if (flush_needed)
1722 			tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1723 	}
1724 	return 1;
1725 }
1726 
1727 #ifndef pmd_move_must_withdraw
1728 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1729 					 spinlock_t *old_pmd_ptl,
1730 					 struct vm_area_struct *vma)
1731 {
1732 	/*
1733 	 * With split pmd lock we also need to move preallocated
1734 	 * PTE page table if new_pmd is on different PMD page table.
1735 	 *
1736 	 * We also don't deposit and withdraw tables for file pages.
1737 	 */
1738 	return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1739 }
1740 #endif
1741 
1742 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1743 {
1744 #ifdef CONFIG_MEM_SOFT_DIRTY
1745 	if (unlikely(is_pmd_migration_entry(pmd)))
1746 		pmd = pmd_swp_mksoft_dirty(pmd);
1747 	else if (pmd_present(pmd))
1748 		pmd = pmd_mksoft_dirty(pmd);
1749 #endif
1750 	return pmd;
1751 }
1752 
1753 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1754 		  unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1755 {
1756 	spinlock_t *old_ptl, *new_ptl;
1757 	pmd_t pmd;
1758 	struct mm_struct *mm = vma->vm_mm;
1759 	bool force_flush = false;
1760 
1761 	/*
1762 	 * The destination pmd shouldn't be established, free_pgtables()
1763 	 * should have release it.
1764 	 */
1765 	if (WARN_ON(!pmd_none(*new_pmd))) {
1766 		VM_BUG_ON(pmd_trans_huge(*new_pmd));
1767 		return false;
1768 	}
1769 
1770 	/*
1771 	 * We don't have to worry about the ordering of src and dst
1772 	 * ptlocks because exclusive mmap_lock prevents deadlock.
1773 	 */
1774 	old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1775 	if (old_ptl) {
1776 		new_ptl = pmd_lockptr(mm, new_pmd);
1777 		if (new_ptl != old_ptl)
1778 			spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1779 		pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1780 		if (pmd_present(pmd))
1781 			force_flush = true;
1782 		VM_BUG_ON(!pmd_none(*new_pmd));
1783 
1784 		if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1785 			pgtable_t pgtable;
1786 			pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1787 			pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1788 		}
1789 		pmd = move_soft_dirty_pmd(pmd);
1790 		set_pmd_at(mm, new_addr, new_pmd, pmd);
1791 		if (force_flush)
1792 			flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1793 		if (new_ptl != old_ptl)
1794 			spin_unlock(new_ptl);
1795 		spin_unlock(old_ptl);
1796 		return true;
1797 	}
1798 	return false;
1799 }
1800 
1801 /*
1802  * Returns
1803  *  - 0 if PMD could not be locked
1804  *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1805  *      or if prot_numa but THP migration is not supported
1806  *  - HPAGE_PMD_NR if protections changed and TLB flush necessary
1807  */
1808 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1809 		    pmd_t *pmd, unsigned long addr, pgprot_t newprot,
1810 		    unsigned long cp_flags)
1811 {
1812 	struct mm_struct *mm = vma->vm_mm;
1813 	spinlock_t *ptl;
1814 	pmd_t oldpmd, entry;
1815 	bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1816 	bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1817 	bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1818 	int ret = 1;
1819 
1820 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1821 
1822 	if (prot_numa && !thp_migration_supported())
1823 		return 1;
1824 
1825 	ptl = __pmd_trans_huge_lock(pmd, vma);
1826 	if (!ptl)
1827 		return 0;
1828 
1829 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1830 	if (is_swap_pmd(*pmd)) {
1831 		swp_entry_t entry = pmd_to_swp_entry(*pmd);
1832 		struct page *page = pfn_swap_entry_to_page(entry);
1833 		pmd_t newpmd;
1834 
1835 		VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1836 		if (is_writable_migration_entry(entry)) {
1837 			/*
1838 			 * A protection check is difficult so
1839 			 * just be safe and disable write
1840 			 */
1841 			if (PageAnon(page))
1842 				entry = make_readable_exclusive_migration_entry(swp_offset(entry));
1843 			else
1844 				entry = make_readable_migration_entry(swp_offset(entry));
1845 			newpmd = swp_entry_to_pmd(entry);
1846 			if (pmd_swp_soft_dirty(*pmd))
1847 				newpmd = pmd_swp_mksoft_dirty(newpmd);
1848 		} else {
1849 			newpmd = *pmd;
1850 		}
1851 
1852 		if (uffd_wp)
1853 			newpmd = pmd_swp_mkuffd_wp(newpmd);
1854 		else if (uffd_wp_resolve)
1855 			newpmd = pmd_swp_clear_uffd_wp(newpmd);
1856 		if (!pmd_same(*pmd, newpmd))
1857 			set_pmd_at(mm, addr, pmd, newpmd);
1858 		goto unlock;
1859 	}
1860 #endif
1861 
1862 	if (prot_numa) {
1863 		struct page *page;
1864 		bool toptier;
1865 		/*
1866 		 * Avoid trapping faults against the zero page. The read-only
1867 		 * data is likely to be read-cached on the local CPU and
1868 		 * local/remote hits to the zero page are not interesting.
1869 		 */
1870 		if (is_huge_zero_pmd(*pmd))
1871 			goto unlock;
1872 
1873 		if (pmd_protnone(*pmd))
1874 			goto unlock;
1875 
1876 		page = pmd_page(*pmd);
1877 		toptier = node_is_toptier(page_to_nid(page));
1878 		/*
1879 		 * Skip scanning top tier node if normal numa
1880 		 * balancing is disabled
1881 		 */
1882 		if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
1883 		    toptier)
1884 			goto unlock;
1885 
1886 		if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
1887 		    !toptier)
1888 			xchg_page_access_time(page, jiffies_to_msecs(jiffies));
1889 	}
1890 	/*
1891 	 * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1892 	 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1893 	 * which is also under mmap_read_lock(mm):
1894 	 *
1895 	 *	CPU0:				CPU1:
1896 	 *				change_huge_pmd(prot_numa=1)
1897 	 *				 pmdp_huge_get_and_clear_notify()
1898 	 * madvise_dontneed()
1899 	 *  zap_pmd_range()
1900 	 *   pmd_trans_huge(*pmd) == 0 (without ptl)
1901 	 *   // skip the pmd
1902 	 *				 set_pmd_at();
1903 	 *				 // pmd is re-established
1904 	 *
1905 	 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1906 	 * which may break userspace.
1907 	 *
1908 	 * pmdp_invalidate_ad() is required to make sure we don't miss
1909 	 * dirty/young flags set by hardware.
1910 	 */
1911 	oldpmd = pmdp_invalidate_ad(vma, addr, pmd);
1912 
1913 	entry = pmd_modify(oldpmd, newprot);
1914 	if (uffd_wp)
1915 		entry = pmd_mkuffd_wp(entry);
1916 	else if (uffd_wp_resolve)
1917 		/*
1918 		 * Leave the write bit to be handled by PF interrupt
1919 		 * handler, then things like COW could be properly
1920 		 * handled.
1921 		 */
1922 		entry = pmd_clear_uffd_wp(entry);
1923 
1924 	/* See change_pte_range(). */
1925 	if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) && !pmd_write(entry) &&
1926 	    can_change_pmd_writable(vma, addr, entry))
1927 		entry = pmd_mkwrite(entry);
1928 
1929 	ret = HPAGE_PMD_NR;
1930 	set_pmd_at(mm, addr, pmd, entry);
1931 
1932 	if (huge_pmd_needs_flush(oldpmd, entry))
1933 		tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE);
1934 unlock:
1935 	spin_unlock(ptl);
1936 	return ret;
1937 }
1938 
1939 /*
1940  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1941  *
1942  * Note that if it returns page table lock pointer, this routine returns without
1943  * unlocking page table lock. So callers must unlock it.
1944  */
1945 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1946 {
1947 	spinlock_t *ptl;
1948 	ptl = pmd_lock(vma->vm_mm, pmd);
1949 	if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1950 			pmd_devmap(*pmd)))
1951 		return ptl;
1952 	spin_unlock(ptl);
1953 	return NULL;
1954 }
1955 
1956 /*
1957  * Returns page table lock pointer if a given pud maps a thp, NULL otherwise.
1958  *
1959  * Note that if it returns page table lock pointer, this routine returns without
1960  * unlocking page table lock. So callers must unlock it.
1961  */
1962 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1963 {
1964 	spinlock_t *ptl;
1965 
1966 	ptl = pud_lock(vma->vm_mm, pud);
1967 	if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1968 		return ptl;
1969 	spin_unlock(ptl);
1970 	return NULL;
1971 }
1972 
1973 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1974 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1975 		 pud_t *pud, unsigned long addr)
1976 {
1977 	spinlock_t *ptl;
1978 
1979 	ptl = __pud_trans_huge_lock(pud, vma);
1980 	if (!ptl)
1981 		return 0;
1982 
1983 	pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1984 	tlb_remove_pud_tlb_entry(tlb, pud, addr);
1985 	if (vma_is_special_huge(vma)) {
1986 		spin_unlock(ptl);
1987 		/* No zero page support yet */
1988 	} else {
1989 		/* No support for anonymous PUD pages yet */
1990 		BUG();
1991 	}
1992 	return 1;
1993 }
1994 
1995 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1996 		unsigned long haddr)
1997 {
1998 	VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1999 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2000 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2001 	VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2002 
2003 	count_vm_event(THP_SPLIT_PUD);
2004 
2005 	pudp_huge_clear_flush_notify(vma, haddr, pud);
2006 }
2007 
2008 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2009 		unsigned long address)
2010 {
2011 	spinlock_t *ptl;
2012 	struct mmu_notifier_range range;
2013 
2014 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2015 				address & HPAGE_PUD_MASK,
2016 				(address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2017 	mmu_notifier_invalidate_range_start(&range);
2018 	ptl = pud_lock(vma->vm_mm, pud);
2019 	if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2020 		goto out;
2021 	__split_huge_pud_locked(vma, pud, range.start);
2022 
2023 out:
2024 	spin_unlock(ptl);
2025 	/*
2026 	 * No need to double call mmu_notifier->invalidate_range() callback as
2027 	 * the above pudp_huge_clear_flush_notify() did already call it.
2028 	 */
2029 	mmu_notifier_invalidate_range_only_end(&range);
2030 }
2031 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2032 
2033 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2034 		unsigned long haddr, pmd_t *pmd)
2035 {
2036 	struct mm_struct *mm = vma->vm_mm;
2037 	pgtable_t pgtable;
2038 	pmd_t _pmd, old_pmd;
2039 	int i;
2040 
2041 	/*
2042 	 * Leave pmd empty until pte is filled note that it is fine to delay
2043 	 * notification until mmu_notifier_invalidate_range_end() as we are
2044 	 * replacing a zero pmd write protected page with a zero pte write
2045 	 * protected page.
2046 	 *
2047 	 * See Documentation/mm/mmu_notifier.rst
2048 	 */
2049 	old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd);
2050 
2051 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2052 	pmd_populate(mm, &_pmd, pgtable);
2053 
2054 	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2055 		pte_t *pte, entry;
2056 		entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2057 		entry = pte_mkspecial(entry);
2058 		if (pmd_uffd_wp(old_pmd))
2059 			entry = pte_mkuffd_wp(entry);
2060 		pte = pte_offset_map(&_pmd, haddr);
2061 		VM_BUG_ON(!pte_none(*pte));
2062 		set_pte_at(mm, haddr, pte, entry);
2063 		pte_unmap(pte);
2064 	}
2065 	smp_wmb(); /* make pte visible before pmd */
2066 	pmd_populate(mm, pmd, pgtable);
2067 }
2068 
2069 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2070 		unsigned long haddr, bool freeze)
2071 {
2072 	struct mm_struct *mm = vma->vm_mm;
2073 	struct page *page;
2074 	pgtable_t pgtable;
2075 	pmd_t old_pmd, _pmd;
2076 	bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2077 	bool anon_exclusive = false, dirty = false;
2078 	unsigned long addr;
2079 	int i;
2080 
2081 	VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2082 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2083 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2084 	VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2085 				&& !pmd_devmap(*pmd));
2086 
2087 	count_vm_event(THP_SPLIT_PMD);
2088 
2089 	if (!vma_is_anonymous(vma)) {
2090 		old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2091 		/*
2092 		 * We are going to unmap this huge page. So
2093 		 * just go ahead and zap it
2094 		 */
2095 		if (arch_needs_pgtable_deposit())
2096 			zap_deposited_table(mm, pmd);
2097 		if (vma_is_special_huge(vma))
2098 			return;
2099 		if (unlikely(is_pmd_migration_entry(old_pmd))) {
2100 			swp_entry_t entry;
2101 
2102 			entry = pmd_to_swp_entry(old_pmd);
2103 			page = pfn_swap_entry_to_page(entry);
2104 		} else {
2105 			page = pmd_page(old_pmd);
2106 			if (!PageDirty(page) && pmd_dirty(old_pmd))
2107 				set_page_dirty(page);
2108 			if (!PageReferenced(page) && pmd_young(old_pmd))
2109 				SetPageReferenced(page);
2110 			page_remove_rmap(page, vma, true);
2111 			put_page(page);
2112 		}
2113 		add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2114 		return;
2115 	}
2116 
2117 	if (is_huge_zero_pmd(*pmd)) {
2118 		/*
2119 		 * FIXME: Do we want to invalidate secondary mmu by calling
2120 		 * mmu_notifier_invalidate_range() see comments below inside
2121 		 * __split_huge_pmd() ?
2122 		 *
2123 		 * We are going from a zero huge page write protected to zero
2124 		 * small page also write protected so it does not seems useful
2125 		 * to invalidate secondary mmu at this time.
2126 		 */
2127 		return __split_huge_zero_page_pmd(vma, haddr, pmd);
2128 	}
2129 
2130 	/*
2131 	 * Up to this point the pmd is present and huge and userland has the
2132 	 * whole access to the hugepage during the split (which happens in
2133 	 * place). If we overwrite the pmd with the not-huge version pointing
2134 	 * to the pte here (which of course we could if all CPUs were bug
2135 	 * free), userland could trigger a small page size TLB miss on the
2136 	 * small sized TLB while the hugepage TLB entry is still established in
2137 	 * the huge TLB. Some CPU doesn't like that.
2138 	 * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2139 	 * 383 on page 105. Intel should be safe but is also warns that it's
2140 	 * only safe if the permission and cache attributes of the two entries
2141 	 * loaded in the two TLB is identical (which should be the case here).
2142 	 * But it is generally safer to never allow small and huge TLB entries
2143 	 * for the same virtual address to be loaded simultaneously. So instead
2144 	 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2145 	 * current pmd notpresent (atomically because here the pmd_trans_huge
2146 	 * must remain set at all times on the pmd until the split is complete
2147 	 * for this pmd), then we flush the SMP TLB and finally we write the
2148 	 * non-huge version of the pmd entry with pmd_populate.
2149 	 */
2150 	old_pmd = pmdp_invalidate(vma, haddr, pmd);
2151 
2152 	pmd_migration = is_pmd_migration_entry(old_pmd);
2153 	if (unlikely(pmd_migration)) {
2154 		swp_entry_t entry;
2155 
2156 		entry = pmd_to_swp_entry(old_pmd);
2157 		page = pfn_swap_entry_to_page(entry);
2158 		write = is_writable_migration_entry(entry);
2159 		if (PageAnon(page))
2160 			anon_exclusive = is_readable_exclusive_migration_entry(entry);
2161 		young = is_migration_entry_young(entry);
2162 		dirty = is_migration_entry_dirty(entry);
2163 		soft_dirty = pmd_swp_soft_dirty(old_pmd);
2164 		uffd_wp = pmd_swp_uffd_wp(old_pmd);
2165 	} else {
2166 		page = pmd_page(old_pmd);
2167 		if (pmd_dirty(old_pmd)) {
2168 			dirty = true;
2169 			SetPageDirty(page);
2170 		}
2171 		write = pmd_write(old_pmd);
2172 		young = pmd_young(old_pmd);
2173 		soft_dirty = pmd_soft_dirty(old_pmd);
2174 		uffd_wp = pmd_uffd_wp(old_pmd);
2175 
2176 		VM_BUG_ON_PAGE(!page_count(page), page);
2177 
2178 		/*
2179 		 * Without "freeze", we'll simply split the PMD, propagating the
2180 		 * PageAnonExclusive() flag for each PTE by setting it for
2181 		 * each subpage -- no need to (temporarily) clear.
2182 		 *
2183 		 * With "freeze" we want to replace mapped pages by
2184 		 * migration entries right away. This is only possible if we
2185 		 * managed to clear PageAnonExclusive() -- see
2186 		 * set_pmd_migration_entry().
2187 		 *
2188 		 * In case we cannot clear PageAnonExclusive(), split the PMD
2189 		 * only and let try_to_migrate_one() fail later.
2190 		 *
2191 		 * See page_try_share_anon_rmap(): invalidate PMD first.
2192 		 */
2193 		anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
2194 		if (freeze && anon_exclusive && page_try_share_anon_rmap(page))
2195 			freeze = false;
2196 		if (!freeze)
2197 			page_ref_add(page, HPAGE_PMD_NR - 1);
2198 	}
2199 
2200 	/*
2201 	 * Withdraw the table only after we mark the pmd entry invalid.
2202 	 * This's critical for some architectures (Power).
2203 	 */
2204 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2205 	pmd_populate(mm, &_pmd, pgtable);
2206 
2207 	for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2208 		pte_t entry, *pte;
2209 		/*
2210 		 * Note that NUMA hinting access restrictions are not
2211 		 * transferred to avoid any possibility of altering
2212 		 * permissions across VMAs.
2213 		 */
2214 		if (freeze || pmd_migration) {
2215 			swp_entry_t swp_entry;
2216 			if (write)
2217 				swp_entry = make_writable_migration_entry(
2218 							page_to_pfn(page + i));
2219 			else if (anon_exclusive)
2220 				swp_entry = make_readable_exclusive_migration_entry(
2221 							page_to_pfn(page + i));
2222 			else
2223 				swp_entry = make_readable_migration_entry(
2224 							page_to_pfn(page + i));
2225 			if (young)
2226 				swp_entry = make_migration_entry_young(swp_entry);
2227 			if (dirty)
2228 				swp_entry = make_migration_entry_dirty(swp_entry);
2229 			entry = swp_entry_to_pte(swp_entry);
2230 			if (soft_dirty)
2231 				entry = pte_swp_mksoft_dirty(entry);
2232 			if (uffd_wp)
2233 				entry = pte_swp_mkuffd_wp(entry);
2234 		} else {
2235 			entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2236 			if (write)
2237 				entry = pte_mkwrite(entry);
2238 			if (anon_exclusive)
2239 				SetPageAnonExclusive(page + i);
2240 			if (!young)
2241 				entry = pte_mkold(entry);
2242 			/* NOTE: this may set soft-dirty too on some archs */
2243 			if (dirty)
2244 				entry = pte_mkdirty(entry);
2245 			if (soft_dirty)
2246 				entry = pte_mksoft_dirty(entry);
2247 			if (uffd_wp)
2248 				entry = pte_mkuffd_wp(entry);
2249 			page_add_anon_rmap(page + i, vma, addr, false);
2250 		}
2251 		pte = pte_offset_map(&_pmd, addr);
2252 		BUG_ON(!pte_none(*pte));
2253 		set_pte_at(mm, addr, pte, entry);
2254 		pte_unmap(pte);
2255 	}
2256 
2257 	if (!pmd_migration)
2258 		page_remove_rmap(page, vma, true);
2259 	if (freeze)
2260 		put_page(page);
2261 
2262 	smp_wmb(); /* make pte visible before pmd */
2263 	pmd_populate(mm, pmd, pgtable);
2264 }
2265 
2266 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2267 		unsigned long address, bool freeze, struct folio *folio)
2268 {
2269 	spinlock_t *ptl;
2270 	struct mmu_notifier_range range;
2271 
2272 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
2273 				address & HPAGE_PMD_MASK,
2274 				(address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2275 	mmu_notifier_invalidate_range_start(&range);
2276 	ptl = pmd_lock(vma->vm_mm, pmd);
2277 
2278 	/*
2279 	 * If caller asks to setup a migration entry, we need a folio to check
2280 	 * pmd against. Otherwise we can end up replacing wrong folio.
2281 	 */
2282 	VM_BUG_ON(freeze && !folio);
2283 	VM_WARN_ON_ONCE(folio && !folio_test_locked(folio));
2284 
2285 	if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) ||
2286 	    is_pmd_migration_entry(*pmd)) {
2287 		/*
2288 		 * It's safe to call pmd_page when folio is set because it's
2289 		 * guaranteed that pmd is present.
2290 		 */
2291 		if (folio && folio != page_folio(pmd_page(*pmd)))
2292 			goto out;
2293 		__split_huge_pmd_locked(vma, pmd, range.start, freeze);
2294 	}
2295 
2296 out:
2297 	spin_unlock(ptl);
2298 	/*
2299 	 * No need to double call mmu_notifier->invalidate_range() callback.
2300 	 * They are 3 cases to consider inside __split_huge_pmd_locked():
2301 	 *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2302 	 *  2) __split_huge_zero_page_pmd() read only zero page and any write
2303 	 *    fault will trigger a flush_notify before pointing to a new page
2304 	 *    (it is fine if the secondary mmu keeps pointing to the old zero
2305 	 *    page in the meantime)
2306 	 *  3) Split a huge pmd into pte pointing to the same page. No need
2307 	 *     to invalidate secondary tlb entry they are all still valid.
2308 	 *     any further changes to individual pte will notify. So no need
2309 	 *     to call mmu_notifier->invalidate_range()
2310 	 */
2311 	mmu_notifier_invalidate_range_only_end(&range);
2312 }
2313 
2314 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2315 		bool freeze, struct folio *folio)
2316 {
2317 	pmd_t *pmd = mm_find_pmd(vma->vm_mm, address);
2318 
2319 	if (!pmd)
2320 		return;
2321 
2322 	__split_huge_pmd(vma, pmd, address, freeze, folio);
2323 }
2324 
2325 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2326 {
2327 	/*
2328 	 * If the new address isn't hpage aligned and it could previously
2329 	 * contain an hugepage: check if we need to split an huge pmd.
2330 	 */
2331 	if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2332 	    range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2333 			 ALIGN(address, HPAGE_PMD_SIZE)))
2334 		split_huge_pmd_address(vma, address, false, NULL);
2335 }
2336 
2337 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2338 			     unsigned long start,
2339 			     unsigned long end,
2340 			     long adjust_next)
2341 {
2342 	/* Check if we need to split start first. */
2343 	split_huge_pmd_if_needed(vma, start);
2344 
2345 	/* Check if we need to split end next. */
2346 	split_huge_pmd_if_needed(vma, end);
2347 
2348 	/*
2349 	 * If we're also updating the next vma vm_start,
2350 	 * check if we need to split it.
2351 	 */
2352 	if (adjust_next > 0) {
2353 		struct vm_area_struct *next = find_vma(vma->vm_mm, vma->vm_end);
2354 		unsigned long nstart = next->vm_start;
2355 		nstart += adjust_next;
2356 		split_huge_pmd_if_needed(next, nstart);
2357 	}
2358 }
2359 
2360 static void unmap_folio(struct folio *folio)
2361 {
2362 	enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2363 		TTU_SYNC;
2364 
2365 	VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2366 
2367 	/*
2368 	 * Anon pages need migration entries to preserve them, but file
2369 	 * pages can simply be left unmapped, then faulted back on demand.
2370 	 * If that is ever changed (perhaps for mlock), update remap_page().
2371 	 */
2372 	if (folio_test_anon(folio))
2373 		try_to_migrate(folio, ttu_flags);
2374 	else
2375 		try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK);
2376 }
2377 
2378 static void remap_page(struct folio *folio, unsigned long nr)
2379 {
2380 	int i = 0;
2381 
2382 	/* If unmap_folio() uses try_to_migrate() on file, remove this check */
2383 	if (!folio_test_anon(folio))
2384 		return;
2385 	for (;;) {
2386 		remove_migration_ptes(folio, folio, true);
2387 		i += folio_nr_pages(folio);
2388 		if (i >= nr)
2389 			break;
2390 		folio = folio_next(folio);
2391 	}
2392 }
2393 
2394 static void lru_add_page_tail(struct page *head, struct page *tail,
2395 		struct lruvec *lruvec, struct list_head *list)
2396 {
2397 	VM_BUG_ON_PAGE(!PageHead(head), head);
2398 	VM_BUG_ON_PAGE(PageCompound(tail), head);
2399 	VM_BUG_ON_PAGE(PageLRU(tail), head);
2400 	lockdep_assert_held(&lruvec->lru_lock);
2401 
2402 	if (list) {
2403 		/* page reclaim is reclaiming a huge page */
2404 		VM_WARN_ON(PageLRU(head));
2405 		get_page(tail);
2406 		list_add_tail(&tail->lru, list);
2407 	} else {
2408 		/* head is still on lru (and we have it frozen) */
2409 		VM_WARN_ON(!PageLRU(head));
2410 		if (PageUnevictable(tail))
2411 			tail->mlock_count = 0;
2412 		else
2413 			list_add_tail(&tail->lru, &head->lru);
2414 		SetPageLRU(tail);
2415 	}
2416 }
2417 
2418 static void __split_huge_page_tail(struct page *head, int tail,
2419 		struct lruvec *lruvec, struct list_head *list)
2420 {
2421 	struct page *page_tail = head + tail;
2422 
2423 	VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2424 
2425 	/*
2426 	 * Clone page flags before unfreezing refcount.
2427 	 *
2428 	 * After successful get_page_unless_zero() might follow flags change,
2429 	 * for example lock_page() which set PG_waiters.
2430 	 *
2431 	 * Note that for mapped sub-pages of an anonymous THP,
2432 	 * PG_anon_exclusive has been cleared in unmap_folio() and is stored in
2433 	 * the migration entry instead from where remap_page() will restore it.
2434 	 * We can still have PG_anon_exclusive set on effectively unmapped and
2435 	 * unreferenced sub-pages of an anonymous THP: we can simply drop
2436 	 * PG_anon_exclusive (-> PG_mappedtodisk) for these here.
2437 	 */
2438 	page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2439 	page_tail->flags |= (head->flags &
2440 			((1L << PG_referenced) |
2441 			 (1L << PG_swapbacked) |
2442 			 (1L << PG_swapcache) |
2443 			 (1L << PG_mlocked) |
2444 			 (1L << PG_uptodate) |
2445 			 (1L << PG_active) |
2446 			 (1L << PG_workingset) |
2447 			 (1L << PG_locked) |
2448 			 (1L << PG_unevictable) |
2449 #ifdef CONFIG_ARCH_USES_PG_ARCH_X
2450 			 (1L << PG_arch_2) |
2451 			 (1L << PG_arch_3) |
2452 #endif
2453 			 (1L << PG_dirty) |
2454 			 LRU_GEN_MASK | LRU_REFS_MASK));
2455 
2456 	/* ->mapping in first and second tail page is replaced by other uses */
2457 	VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2458 			page_tail);
2459 	page_tail->mapping = head->mapping;
2460 	page_tail->index = head->index + tail;
2461 
2462 	/*
2463 	 * page->private should not be set in tail pages with the exception
2464 	 * of swap cache pages that store the swp_entry_t in tail pages.
2465 	 * Fix up and warn once if private is unexpectedly set.
2466 	 *
2467 	 * What of 32-bit systems, on which folio->_pincount overlays
2468 	 * head[1].private?  No problem: THP_SWAP is not enabled on 32-bit, and
2469 	 * pincount must be 0 for folio_ref_freeze() to have succeeded.
2470 	 */
2471 	if (!folio_test_swapcache(page_folio(head))) {
2472 		VM_WARN_ON_ONCE_PAGE(page_tail->private != 0, page_tail);
2473 		page_tail->private = 0;
2474 	}
2475 
2476 	/* Page flags must be visible before we make the page non-compound. */
2477 	smp_wmb();
2478 
2479 	/*
2480 	 * Clear PageTail before unfreezing page refcount.
2481 	 *
2482 	 * After successful get_page_unless_zero() might follow put_page()
2483 	 * which needs correct compound_head().
2484 	 */
2485 	clear_compound_head(page_tail);
2486 
2487 	/* Finally unfreeze refcount. Additional reference from page cache. */
2488 	page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2489 					  PageSwapCache(head)));
2490 
2491 	if (page_is_young(head))
2492 		set_page_young(page_tail);
2493 	if (page_is_idle(head))
2494 		set_page_idle(page_tail);
2495 
2496 	page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2497 
2498 	/*
2499 	 * always add to the tail because some iterators expect new
2500 	 * pages to show after the currently processed elements - e.g.
2501 	 * migrate_pages
2502 	 */
2503 	lru_add_page_tail(head, page_tail, lruvec, list);
2504 }
2505 
2506 static void __split_huge_page(struct page *page, struct list_head *list,
2507 		pgoff_t end)
2508 {
2509 	struct folio *folio = page_folio(page);
2510 	struct page *head = &folio->page;
2511 	struct lruvec *lruvec;
2512 	struct address_space *swap_cache = NULL;
2513 	unsigned long offset = 0;
2514 	unsigned int nr = thp_nr_pages(head);
2515 	int i;
2516 
2517 	/* complete memcg works before add pages to LRU */
2518 	split_page_memcg(head, nr);
2519 
2520 	if (PageAnon(head) && PageSwapCache(head)) {
2521 		swp_entry_t entry = { .val = page_private(head) };
2522 
2523 		offset = swp_offset(entry);
2524 		swap_cache = swap_address_space(entry);
2525 		xa_lock(&swap_cache->i_pages);
2526 	}
2527 
2528 	/* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2529 	lruvec = folio_lruvec_lock(folio);
2530 
2531 	ClearPageHasHWPoisoned(head);
2532 
2533 	for (i = nr - 1; i >= 1; i--) {
2534 		__split_huge_page_tail(head, i, lruvec, list);
2535 		/* Some pages can be beyond EOF: drop them from page cache */
2536 		if (head[i].index >= end) {
2537 			struct folio *tail = page_folio(head + i);
2538 
2539 			if (shmem_mapping(head->mapping))
2540 				shmem_uncharge(head->mapping->host, 1);
2541 			else if (folio_test_clear_dirty(tail))
2542 				folio_account_cleaned(tail,
2543 					inode_to_wb(folio->mapping->host));
2544 			__filemap_remove_folio(tail, NULL);
2545 			folio_put(tail);
2546 		} else if (!PageAnon(page)) {
2547 			__xa_store(&head->mapping->i_pages, head[i].index,
2548 					head + i, 0);
2549 		} else if (swap_cache) {
2550 			__xa_store(&swap_cache->i_pages, offset + i,
2551 					head + i, 0);
2552 		}
2553 	}
2554 
2555 	ClearPageCompound(head);
2556 	unlock_page_lruvec(lruvec);
2557 	/* Caller disabled irqs, so they are still disabled here */
2558 
2559 	split_page_owner(head, nr);
2560 
2561 	/* See comment in __split_huge_page_tail() */
2562 	if (PageAnon(head)) {
2563 		/* Additional pin to swap cache */
2564 		if (PageSwapCache(head)) {
2565 			page_ref_add(head, 2);
2566 			xa_unlock(&swap_cache->i_pages);
2567 		} else {
2568 			page_ref_inc(head);
2569 		}
2570 	} else {
2571 		/* Additional pin to page cache */
2572 		page_ref_add(head, 2);
2573 		xa_unlock(&head->mapping->i_pages);
2574 	}
2575 	local_irq_enable();
2576 
2577 	remap_page(folio, nr);
2578 
2579 	if (PageSwapCache(head)) {
2580 		swp_entry_t entry = { .val = page_private(head) };
2581 
2582 		split_swap_cluster(entry);
2583 	}
2584 
2585 	for (i = 0; i < nr; i++) {
2586 		struct page *subpage = head + i;
2587 		if (subpage == page)
2588 			continue;
2589 		unlock_page(subpage);
2590 
2591 		/*
2592 		 * Subpages may be freed if there wasn't any mapping
2593 		 * like if add_to_swap() is running on a lru page that
2594 		 * had its mapping zapped. And freeing these pages
2595 		 * requires taking the lru_lock so we do the put_page
2596 		 * of the tail pages after the split is complete.
2597 		 */
2598 		free_page_and_swap_cache(subpage);
2599 	}
2600 }
2601 
2602 /* Racy check whether the huge page can be split */
2603 bool can_split_folio(struct folio *folio, int *pextra_pins)
2604 {
2605 	int extra_pins;
2606 
2607 	/* Additional pins from page cache */
2608 	if (folio_test_anon(folio))
2609 		extra_pins = folio_test_swapcache(folio) ?
2610 				folio_nr_pages(folio) : 0;
2611 	else
2612 		extra_pins = folio_nr_pages(folio);
2613 	if (pextra_pins)
2614 		*pextra_pins = extra_pins;
2615 	return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 1;
2616 }
2617 
2618 /*
2619  * This function splits huge page into normal pages. @page can point to any
2620  * subpage of huge page to split. Split doesn't change the position of @page.
2621  *
2622  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2623  * The huge page must be locked.
2624  *
2625  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2626  *
2627  * Both head page and tail pages will inherit mapping, flags, and so on from
2628  * the hugepage.
2629  *
2630  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2631  * they are not mapped.
2632  *
2633  * Returns 0 if the hugepage is split successfully.
2634  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2635  * us.
2636  */
2637 int split_huge_page_to_list(struct page *page, struct list_head *list)
2638 {
2639 	struct folio *folio = page_folio(page);
2640 	struct deferred_split *ds_queue = get_deferred_split_queue(folio);
2641 	XA_STATE(xas, &folio->mapping->i_pages, folio->index);
2642 	struct anon_vma *anon_vma = NULL;
2643 	struct address_space *mapping = NULL;
2644 	int extra_pins, ret;
2645 	pgoff_t end;
2646 	bool is_hzp;
2647 
2648 	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
2649 	VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
2650 
2651 	is_hzp = is_huge_zero_page(&folio->page);
2652 	if (is_hzp) {
2653 		pr_warn_ratelimited("Called split_huge_page for huge zero page\n");
2654 		return -EBUSY;
2655 	}
2656 
2657 	if (folio_test_writeback(folio))
2658 		return -EBUSY;
2659 
2660 	if (folio_test_anon(folio)) {
2661 		/*
2662 		 * The caller does not necessarily hold an mmap_lock that would
2663 		 * prevent the anon_vma disappearing so we first we take a
2664 		 * reference to it and then lock the anon_vma for write. This
2665 		 * is similar to folio_lock_anon_vma_read except the write lock
2666 		 * is taken to serialise against parallel split or collapse
2667 		 * operations.
2668 		 */
2669 		anon_vma = folio_get_anon_vma(folio);
2670 		if (!anon_vma) {
2671 			ret = -EBUSY;
2672 			goto out;
2673 		}
2674 		end = -1;
2675 		mapping = NULL;
2676 		anon_vma_lock_write(anon_vma);
2677 	} else {
2678 		gfp_t gfp;
2679 
2680 		mapping = folio->mapping;
2681 
2682 		/* Truncated ? */
2683 		if (!mapping) {
2684 			ret = -EBUSY;
2685 			goto out;
2686 		}
2687 
2688 		gfp = current_gfp_context(mapping_gfp_mask(mapping) &
2689 							GFP_RECLAIM_MASK);
2690 
2691 		if (folio_test_private(folio) &&
2692 				!filemap_release_folio(folio, gfp)) {
2693 			ret = -EBUSY;
2694 			goto out;
2695 		}
2696 
2697 		xas_split_alloc(&xas, folio, folio_order(folio), gfp);
2698 		if (xas_error(&xas)) {
2699 			ret = xas_error(&xas);
2700 			goto out;
2701 		}
2702 
2703 		anon_vma = NULL;
2704 		i_mmap_lock_read(mapping);
2705 
2706 		/*
2707 		 *__split_huge_page() may need to trim off pages beyond EOF:
2708 		 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2709 		 * which cannot be nested inside the page tree lock. So note
2710 		 * end now: i_size itself may be changed at any moment, but
2711 		 * folio lock is good enough to serialize the trimming.
2712 		 */
2713 		end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2714 		if (shmem_mapping(mapping))
2715 			end = shmem_fallocend(mapping->host, end);
2716 	}
2717 
2718 	/*
2719 	 * Racy check if we can split the page, before unmap_folio() will
2720 	 * split PMDs
2721 	 */
2722 	if (!can_split_folio(folio, &extra_pins)) {
2723 		ret = -EAGAIN;
2724 		goto out_unlock;
2725 	}
2726 
2727 	unmap_folio(folio);
2728 
2729 	/* block interrupt reentry in xa_lock and spinlock */
2730 	local_irq_disable();
2731 	if (mapping) {
2732 		/*
2733 		 * Check if the folio is present in page cache.
2734 		 * We assume all tail are present too, if folio is there.
2735 		 */
2736 		xas_lock(&xas);
2737 		xas_reset(&xas);
2738 		if (xas_load(&xas) != folio)
2739 			goto fail;
2740 	}
2741 
2742 	/* Prevent deferred_split_scan() touching ->_refcount */
2743 	spin_lock(&ds_queue->split_queue_lock);
2744 	if (folio_ref_freeze(folio, 1 + extra_pins)) {
2745 		if (!list_empty(&folio->_deferred_list)) {
2746 			ds_queue->split_queue_len--;
2747 			list_del(&folio->_deferred_list);
2748 		}
2749 		spin_unlock(&ds_queue->split_queue_lock);
2750 		if (mapping) {
2751 			int nr = folio_nr_pages(folio);
2752 
2753 			xas_split(&xas, folio, folio_order(folio));
2754 			if (folio_test_swapbacked(folio)) {
2755 				__lruvec_stat_mod_folio(folio, NR_SHMEM_THPS,
2756 							-nr);
2757 			} else {
2758 				__lruvec_stat_mod_folio(folio, NR_FILE_THPS,
2759 							-nr);
2760 				filemap_nr_thps_dec(mapping);
2761 			}
2762 		}
2763 
2764 		__split_huge_page(page, list, end);
2765 		ret = 0;
2766 	} else {
2767 		spin_unlock(&ds_queue->split_queue_lock);
2768 fail:
2769 		if (mapping)
2770 			xas_unlock(&xas);
2771 		local_irq_enable();
2772 		remap_page(folio, folio_nr_pages(folio));
2773 		ret = -EAGAIN;
2774 	}
2775 
2776 out_unlock:
2777 	if (anon_vma) {
2778 		anon_vma_unlock_write(anon_vma);
2779 		put_anon_vma(anon_vma);
2780 	}
2781 	if (mapping)
2782 		i_mmap_unlock_read(mapping);
2783 out:
2784 	xas_destroy(&xas);
2785 	count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2786 	return ret;
2787 }
2788 
2789 void free_transhuge_page(struct page *page)
2790 {
2791 	struct folio *folio = (struct folio *)page;
2792 	struct deferred_split *ds_queue = get_deferred_split_queue(folio);
2793 	unsigned long flags;
2794 
2795 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2796 	if (!list_empty(&folio->_deferred_list)) {
2797 		ds_queue->split_queue_len--;
2798 		list_del(&folio->_deferred_list);
2799 	}
2800 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2801 	free_compound_page(page);
2802 }
2803 
2804 void deferred_split_folio(struct folio *folio)
2805 {
2806 	struct deferred_split *ds_queue = get_deferred_split_queue(folio);
2807 #ifdef CONFIG_MEMCG
2808 	struct mem_cgroup *memcg = folio_memcg(folio);
2809 #endif
2810 	unsigned long flags;
2811 
2812 	VM_BUG_ON_FOLIO(folio_order(folio) < 2, folio);
2813 
2814 	/*
2815 	 * The try_to_unmap() in page reclaim path might reach here too,
2816 	 * this may cause a race condition to corrupt deferred split queue.
2817 	 * And, if page reclaim is already handling the same folio, it is
2818 	 * unnecessary to handle it again in shrinker.
2819 	 *
2820 	 * Check the swapcache flag to determine if the folio is being
2821 	 * handled by page reclaim since THP swap would add the folio into
2822 	 * swap cache before calling try_to_unmap().
2823 	 */
2824 	if (folio_test_swapcache(folio))
2825 		return;
2826 
2827 	if (!list_empty(&folio->_deferred_list))
2828 		return;
2829 
2830 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2831 	if (list_empty(&folio->_deferred_list)) {
2832 		count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2833 		list_add_tail(&folio->_deferred_list, &ds_queue->split_queue);
2834 		ds_queue->split_queue_len++;
2835 #ifdef CONFIG_MEMCG
2836 		if (memcg)
2837 			set_shrinker_bit(memcg, folio_nid(folio),
2838 					 deferred_split_shrinker.id);
2839 #endif
2840 	}
2841 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2842 }
2843 
2844 static unsigned long deferred_split_count(struct shrinker *shrink,
2845 		struct shrink_control *sc)
2846 {
2847 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2848 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2849 
2850 #ifdef CONFIG_MEMCG
2851 	if (sc->memcg)
2852 		ds_queue = &sc->memcg->deferred_split_queue;
2853 #endif
2854 	return READ_ONCE(ds_queue->split_queue_len);
2855 }
2856 
2857 static unsigned long deferred_split_scan(struct shrinker *shrink,
2858 		struct shrink_control *sc)
2859 {
2860 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2861 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2862 	unsigned long flags;
2863 	LIST_HEAD(list);
2864 	struct folio *folio, *next;
2865 	int split = 0;
2866 
2867 #ifdef CONFIG_MEMCG
2868 	if (sc->memcg)
2869 		ds_queue = &sc->memcg->deferred_split_queue;
2870 #endif
2871 
2872 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2873 	/* Take pin on all head pages to avoid freeing them under us */
2874 	list_for_each_entry_safe(folio, next, &ds_queue->split_queue,
2875 							_deferred_list) {
2876 		if (folio_try_get(folio)) {
2877 			list_move(&folio->_deferred_list, &list);
2878 		} else {
2879 			/* We lost race with folio_put() */
2880 			list_del_init(&folio->_deferred_list);
2881 			ds_queue->split_queue_len--;
2882 		}
2883 		if (!--sc->nr_to_scan)
2884 			break;
2885 	}
2886 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2887 
2888 	list_for_each_entry_safe(folio, next, &list, _deferred_list) {
2889 		if (!folio_trylock(folio))
2890 			goto next;
2891 		/* split_huge_page() removes page from list on success */
2892 		if (!split_folio(folio))
2893 			split++;
2894 		folio_unlock(folio);
2895 next:
2896 		folio_put(folio);
2897 	}
2898 
2899 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2900 	list_splice_tail(&list, &ds_queue->split_queue);
2901 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2902 
2903 	/*
2904 	 * Stop shrinker if we didn't split any page, but the queue is empty.
2905 	 * This can happen if pages were freed under us.
2906 	 */
2907 	if (!split && list_empty(&ds_queue->split_queue))
2908 		return SHRINK_STOP;
2909 	return split;
2910 }
2911 
2912 static struct shrinker deferred_split_shrinker = {
2913 	.count_objects = deferred_split_count,
2914 	.scan_objects = deferred_split_scan,
2915 	.seeks = DEFAULT_SEEKS,
2916 	.flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2917 		 SHRINKER_NONSLAB,
2918 };
2919 
2920 #ifdef CONFIG_DEBUG_FS
2921 static void split_huge_pages_all(void)
2922 {
2923 	struct zone *zone;
2924 	struct page *page;
2925 	struct folio *folio;
2926 	unsigned long pfn, max_zone_pfn;
2927 	unsigned long total = 0, split = 0;
2928 
2929 	pr_debug("Split all THPs\n");
2930 	for_each_zone(zone) {
2931 		if (!managed_zone(zone))
2932 			continue;
2933 		max_zone_pfn = zone_end_pfn(zone);
2934 		for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2935 			int nr_pages;
2936 
2937 			page = pfn_to_online_page(pfn);
2938 			if (!page || PageTail(page))
2939 				continue;
2940 			folio = page_folio(page);
2941 			if (!folio_try_get(folio))
2942 				continue;
2943 
2944 			if (unlikely(page_folio(page) != folio))
2945 				goto next;
2946 
2947 			if (zone != folio_zone(folio))
2948 				goto next;
2949 
2950 			if (!folio_test_large(folio)
2951 				|| folio_test_hugetlb(folio)
2952 				|| !folio_test_lru(folio))
2953 				goto next;
2954 
2955 			total++;
2956 			folio_lock(folio);
2957 			nr_pages = folio_nr_pages(folio);
2958 			if (!split_folio(folio))
2959 				split++;
2960 			pfn += nr_pages - 1;
2961 			folio_unlock(folio);
2962 next:
2963 			folio_put(folio);
2964 			cond_resched();
2965 		}
2966 	}
2967 
2968 	pr_debug("%lu of %lu THP split\n", split, total);
2969 }
2970 
2971 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
2972 {
2973 	return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
2974 		    is_vm_hugetlb_page(vma);
2975 }
2976 
2977 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
2978 				unsigned long vaddr_end)
2979 {
2980 	int ret = 0;
2981 	struct task_struct *task;
2982 	struct mm_struct *mm;
2983 	unsigned long total = 0, split = 0;
2984 	unsigned long addr;
2985 
2986 	vaddr_start &= PAGE_MASK;
2987 	vaddr_end &= PAGE_MASK;
2988 
2989 	/* Find the task_struct from pid */
2990 	rcu_read_lock();
2991 	task = find_task_by_vpid(pid);
2992 	if (!task) {
2993 		rcu_read_unlock();
2994 		ret = -ESRCH;
2995 		goto out;
2996 	}
2997 	get_task_struct(task);
2998 	rcu_read_unlock();
2999 
3000 	/* Find the mm_struct */
3001 	mm = get_task_mm(task);
3002 	put_task_struct(task);
3003 
3004 	if (!mm) {
3005 		ret = -EINVAL;
3006 		goto out;
3007 	}
3008 
3009 	pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
3010 		 pid, vaddr_start, vaddr_end);
3011 
3012 	mmap_read_lock(mm);
3013 	/*
3014 	 * always increase addr by PAGE_SIZE, since we could have a PTE page
3015 	 * table filled with PTE-mapped THPs, each of which is distinct.
3016 	 */
3017 	for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
3018 		struct vm_area_struct *vma = vma_lookup(mm, addr);
3019 		struct page *page;
3020 
3021 		if (!vma)
3022 			break;
3023 
3024 		/* skip special VMA and hugetlb VMA */
3025 		if (vma_not_suitable_for_thp_split(vma)) {
3026 			addr = vma->vm_end;
3027 			continue;
3028 		}
3029 
3030 		/* FOLL_DUMP to ignore special (like zero) pages */
3031 		page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
3032 
3033 		if (IS_ERR_OR_NULL(page))
3034 			continue;
3035 
3036 		if (!is_transparent_hugepage(page))
3037 			goto next;
3038 
3039 		total++;
3040 		if (!can_split_folio(page_folio(page), NULL))
3041 			goto next;
3042 
3043 		if (!trylock_page(page))
3044 			goto next;
3045 
3046 		if (!split_huge_page(page))
3047 			split++;
3048 
3049 		unlock_page(page);
3050 next:
3051 		put_page(page);
3052 		cond_resched();
3053 	}
3054 	mmap_read_unlock(mm);
3055 	mmput(mm);
3056 
3057 	pr_debug("%lu of %lu THP split\n", split, total);
3058 
3059 out:
3060 	return ret;
3061 }
3062 
3063 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
3064 				pgoff_t off_end)
3065 {
3066 	struct filename *file;
3067 	struct file *candidate;
3068 	struct address_space *mapping;
3069 	int ret = -EINVAL;
3070 	pgoff_t index;
3071 	int nr_pages = 1;
3072 	unsigned long total = 0, split = 0;
3073 
3074 	file = getname_kernel(file_path);
3075 	if (IS_ERR(file))
3076 		return ret;
3077 
3078 	candidate = file_open_name(file, O_RDONLY, 0);
3079 	if (IS_ERR(candidate))
3080 		goto out;
3081 
3082 	pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3083 		 file_path, off_start, off_end);
3084 
3085 	mapping = candidate->f_mapping;
3086 
3087 	for (index = off_start; index < off_end; index += nr_pages) {
3088 		struct folio *folio = filemap_get_folio(mapping, index);
3089 
3090 		nr_pages = 1;
3091 		if (IS_ERR(folio))
3092 			continue;
3093 
3094 		if (!folio_test_large(folio))
3095 			goto next;
3096 
3097 		total++;
3098 		nr_pages = folio_nr_pages(folio);
3099 
3100 		if (!folio_trylock(folio))
3101 			goto next;
3102 
3103 		if (!split_folio(folio))
3104 			split++;
3105 
3106 		folio_unlock(folio);
3107 next:
3108 		folio_put(folio);
3109 		cond_resched();
3110 	}
3111 
3112 	filp_close(candidate, NULL);
3113 	ret = 0;
3114 
3115 	pr_debug("%lu of %lu file-backed THP split\n", split, total);
3116 out:
3117 	putname(file);
3118 	return ret;
3119 }
3120 
3121 #define MAX_INPUT_BUF_SZ 255
3122 
3123 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3124 				size_t count, loff_t *ppops)
3125 {
3126 	static DEFINE_MUTEX(split_debug_mutex);
3127 	ssize_t ret;
3128 	/* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3129 	char input_buf[MAX_INPUT_BUF_SZ];
3130 	int pid;
3131 	unsigned long vaddr_start, vaddr_end;
3132 
3133 	ret = mutex_lock_interruptible(&split_debug_mutex);
3134 	if (ret)
3135 		return ret;
3136 
3137 	ret = -EFAULT;
3138 
3139 	memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3140 	if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3141 		goto out;
3142 
3143 	input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3144 
3145 	if (input_buf[0] == '/') {
3146 		char *tok;
3147 		char *buf = input_buf;
3148 		char file_path[MAX_INPUT_BUF_SZ];
3149 		pgoff_t off_start = 0, off_end = 0;
3150 		size_t input_len = strlen(input_buf);
3151 
3152 		tok = strsep(&buf, ",");
3153 		if (tok) {
3154 			strcpy(file_path, tok);
3155 		} else {
3156 			ret = -EINVAL;
3157 			goto out;
3158 		}
3159 
3160 		ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3161 		if (ret != 2) {
3162 			ret = -EINVAL;
3163 			goto out;
3164 		}
3165 		ret = split_huge_pages_in_file(file_path, off_start, off_end);
3166 		if (!ret)
3167 			ret = input_len;
3168 
3169 		goto out;
3170 	}
3171 
3172 	ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3173 	if (ret == 1 && pid == 1) {
3174 		split_huge_pages_all();
3175 		ret = strlen(input_buf);
3176 		goto out;
3177 	} else if (ret != 3) {
3178 		ret = -EINVAL;
3179 		goto out;
3180 	}
3181 
3182 	ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3183 	if (!ret)
3184 		ret = strlen(input_buf);
3185 out:
3186 	mutex_unlock(&split_debug_mutex);
3187 	return ret;
3188 
3189 }
3190 
3191 static const struct file_operations split_huge_pages_fops = {
3192 	.owner	 = THIS_MODULE,
3193 	.write	 = split_huge_pages_write,
3194 	.llseek  = no_llseek,
3195 };
3196 
3197 static int __init split_huge_pages_debugfs(void)
3198 {
3199 	debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3200 			    &split_huge_pages_fops);
3201 	return 0;
3202 }
3203 late_initcall(split_huge_pages_debugfs);
3204 #endif
3205 
3206 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3207 int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3208 		struct page *page)
3209 {
3210 	struct vm_area_struct *vma = pvmw->vma;
3211 	struct mm_struct *mm = vma->vm_mm;
3212 	unsigned long address = pvmw->address;
3213 	bool anon_exclusive;
3214 	pmd_t pmdval;
3215 	swp_entry_t entry;
3216 	pmd_t pmdswp;
3217 
3218 	if (!(pvmw->pmd && !pvmw->pte))
3219 		return 0;
3220 
3221 	flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3222 	pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3223 
3224 	/* See page_try_share_anon_rmap(): invalidate PMD first. */
3225 	anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
3226 	if (anon_exclusive && page_try_share_anon_rmap(page)) {
3227 		set_pmd_at(mm, address, pvmw->pmd, pmdval);
3228 		return -EBUSY;
3229 	}
3230 
3231 	if (pmd_dirty(pmdval))
3232 		set_page_dirty(page);
3233 	if (pmd_write(pmdval))
3234 		entry = make_writable_migration_entry(page_to_pfn(page));
3235 	else if (anon_exclusive)
3236 		entry = make_readable_exclusive_migration_entry(page_to_pfn(page));
3237 	else
3238 		entry = make_readable_migration_entry(page_to_pfn(page));
3239 	if (pmd_young(pmdval))
3240 		entry = make_migration_entry_young(entry);
3241 	if (pmd_dirty(pmdval))
3242 		entry = make_migration_entry_dirty(entry);
3243 	pmdswp = swp_entry_to_pmd(entry);
3244 	if (pmd_soft_dirty(pmdval))
3245 		pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3246 	if (pmd_uffd_wp(pmdval))
3247 		pmdswp = pmd_swp_mkuffd_wp(pmdswp);
3248 	set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3249 	page_remove_rmap(page, vma, true);
3250 	put_page(page);
3251 	trace_set_migration_pmd(address, pmd_val(pmdswp));
3252 
3253 	return 0;
3254 }
3255 
3256 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3257 {
3258 	struct vm_area_struct *vma = pvmw->vma;
3259 	struct mm_struct *mm = vma->vm_mm;
3260 	unsigned long address = pvmw->address;
3261 	unsigned long haddr = address & HPAGE_PMD_MASK;
3262 	pmd_t pmde;
3263 	swp_entry_t entry;
3264 
3265 	if (!(pvmw->pmd && !pvmw->pte))
3266 		return;
3267 
3268 	entry = pmd_to_swp_entry(*pvmw->pmd);
3269 	get_page(new);
3270 	pmde = mk_huge_pmd(new, READ_ONCE(vma->vm_page_prot));
3271 	if (pmd_swp_soft_dirty(*pvmw->pmd))
3272 		pmde = pmd_mksoft_dirty(pmde);
3273 	if (is_writable_migration_entry(entry))
3274 		pmde = pmd_mkwrite(pmde);
3275 	if (pmd_swp_uffd_wp(*pvmw->pmd))
3276 		pmde = pmd_mkuffd_wp(pmde);
3277 	if (!is_migration_entry_young(entry))
3278 		pmde = pmd_mkold(pmde);
3279 	/* NOTE: this may contain setting soft-dirty on some archs */
3280 	if (PageDirty(new) && is_migration_entry_dirty(entry))
3281 		pmde = pmd_mkdirty(pmde);
3282 
3283 	if (PageAnon(new)) {
3284 		rmap_t rmap_flags = RMAP_COMPOUND;
3285 
3286 		if (!is_readable_migration_entry(entry))
3287 			rmap_flags |= RMAP_EXCLUSIVE;
3288 
3289 		page_add_anon_rmap(new, vma, haddr, rmap_flags);
3290 	} else {
3291 		page_add_file_rmap(new, vma, true);
3292 	}
3293 	VM_BUG_ON(pmd_write(pmde) && PageAnon(new) && !PageAnonExclusive(new));
3294 	set_pmd_at(mm, haddr, pvmw->pmd, pmde);
3295 
3296 	/* No need to invalidate - it was non-present before */
3297 	update_mmu_cache_pmd(vma, address, pvmw->pmd);
3298 	trace_remove_migration_pmd(address, pmd_val(pmde));
3299 }
3300 #endif
3301