xref: /linux/mm/huge_memory.c (revision dec1c62e91ba268ab2a6e339d4d7a59287d5eba1)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  Copyright (C) 2009  Red Hat, Inc.
4  */
5 
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7 
8 #include <linux/mm.h>
9 #include <linux/sched.h>
10 #include <linux/sched/mm.h>
11 #include <linux/sched/coredump.h>
12 #include <linux/sched/numa_balancing.h>
13 #include <linux/highmem.h>
14 #include <linux/hugetlb.h>
15 #include <linux/mmu_notifier.h>
16 #include <linux/rmap.h>
17 #include <linux/swap.h>
18 #include <linux/shrinker.h>
19 #include <linux/mm_inline.h>
20 #include <linux/swapops.h>
21 #include <linux/dax.h>
22 #include <linux/khugepaged.h>
23 #include <linux/freezer.h>
24 #include <linux/pfn_t.h>
25 #include <linux/mman.h>
26 #include <linux/memremap.h>
27 #include <linux/pagemap.h>
28 #include <linux/debugfs.h>
29 #include <linux/migrate.h>
30 #include <linux/hashtable.h>
31 #include <linux/userfaultfd_k.h>
32 #include <linux/page_idle.h>
33 #include <linux/shmem_fs.h>
34 #include <linux/oom.h>
35 #include <linux/numa.h>
36 #include <linux/page_owner.h>
37 #include <linux/sched/sysctl.h>
38 
39 #include <asm/tlb.h>
40 #include <asm/pgalloc.h>
41 #include "internal.h"
42 #include "swap.h"
43 
44 #define CREATE_TRACE_POINTS
45 #include <trace/events/thp.h>
46 
47 /*
48  * By default, transparent hugepage support is disabled in order to avoid
49  * risking an increased memory footprint for applications that are not
50  * guaranteed to benefit from it. When transparent hugepage support is
51  * enabled, it is for all mappings, and khugepaged scans all mappings.
52  * Defrag is invoked by khugepaged hugepage allocations and by page faults
53  * for all hugepage allocations.
54  */
55 unsigned long transparent_hugepage_flags __read_mostly =
56 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
57 	(1<<TRANSPARENT_HUGEPAGE_FLAG)|
58 #endif
59 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
60 	(1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
61 #endif
62 	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
63 	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
64 	(1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
65 
66 static struct shrinker deferred_split_shrinker;
67 
68 static atomic_t huge_zero_refcount;
69 struct page *huge_zero_page __read_mostly;
70 unsigned long huge_zero_pfn __read_mostly = ~0UL;
71 
72 bool transparent_hugepage_active(struct vm_area_struct *vma)
73 {
74 	/* The addr is used to check if the vma size fits */
75 	unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
76 
77 	if (!transhuge_vma_suitable(vma, addr))
78 		return false;
79 	if (vma_is_anonymous(vma))
80 		return __transparent_hugepage_enabled(vma);
81 	if (vma_is_shmem(vma))
82 		return shmem_huge_enabled(vma);
83 	if (transhuge_vma_enabled(vma, vma->vm_flags) && file_thp_enabled(vma))
84 		return true;
85 
86 	return false;
87 }
88 
89 static bool get_huge_zero_page(void)
90 {
91 	struct page *zero_page;
92 retry:
93 	if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
94 		return true;
95 
96 	zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
97 			HPAGE_PMD_ORDER);
98 	if (!zero_page) {
99 		count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
100 		return false;
101 	}
102 	count_vm_event(THP_ZERO_PAGE_ALLOC);
103 	preempt_disable();
104 	if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
105 		preempt_enable();
106 		__free_pages(zero_page, compound_order(zero_page));
107 		goto retry;
108 	}
109 	WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page));
110 
111 	/* We take additional reference here. It will be put back by shrinker */
112 	atomic_set(&huge_zero_refcount, 2);
113 	preempt_enable();
114 	return true;
115 }
116 
117 static void put_huge_zero_page(void)
118 {
119 	/*
120 	 * Counter should never go to zero here. Only shrinker can put
121 	 * last reference.
122 	 */
123 	BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
124 }
125 
126 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
127 {
128 	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
129 		return READ_ONCE(huge_zero_page);
130 
131 	if (!get_huge_zero_page())
132 		return NULL;
133 
134 	if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
135 		put_huge_zero_page();
136 
137 	return READ_ONCE(huge_zero_page);
138 }
139 
140 void mm_put_huge_zero_page(struct mm_struct *mm)
141 {
142 	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
143 		put_huge_zero_page();
144 }
145 
146 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
147 					struct shrink_control *sc)
148 {
149 	/* we can free zero page only if last reference remains */
150 	return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
151 }
152 
153 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
154 				       struct shrink_control *sc)
155 {
156 	if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
157 		struct page *zero_page = xchg(&huge_zero_page, NULL);
158 		BUG_ON(zero_page == NULL);
159 		WRITE_ONCE(huge_zero_pfn, ~0UL);
160 		__free_pages(zero_page, compound_order(zero_page));
161 		return HPAGE_PMD_NR;
162 	}
163 
164 	return 0;
165 }
166 
167 static struct shrinker huge_zero_page_shrinker = {
168 	.count_objects = shrink_huge_zero_page_count,
169 	.scan_objects = shrink_huge_zero_page_scan,
170 	.seeks = DEFAULT_SEEKS,
171 };
172 
173 #ifdef CONFIG_SYSFS
174 static ssize_t enabled_show(struct kobject *kobj,
175 			    struct kobj_attribute *attr, char *buf)
176 {
177 	const char *output;
178 
179 	if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
180 		output = "[always] madvise never";
181 	else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
182 			  &transparent_hugepage_flags))
183 		output = "always [madvise] never";
184 	else
185 		output = "always madvise [never]";
186 
187 	return sysfs_emit(buf, "%s\n", output);
188 }
189 
190 static ssize_t enabled_store(struct kobject *kobj,
191 			     struct kobj_attribute *attr,
192 			     const char *buf, size_t count)
193 {
194 	ssize_t ret = count;
195 
196 	if (sysfs_streq(buf, "always")) {
197 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
198 		set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
199 	} else if (sysfs_streq(buf, "madvise")) {
200 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
201 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
202 	} else if (sysfs_streq(buf, "never")) {
203 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
204 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
205 	} else
206 		ret = -EINVAL;
207 
208 	if (ret > 0) {
209 		int err = start_stop_khugepaged();
210 		if (err)
211 			ret = err;
212 	}
213 	return ret;
214 }
215 static struct kobj_attribute enabled_attr =
216 	__ATTR(enabled, 0644, enabled_show, enabled_store);
217 
218 ssize_t single_hugepage_flag_show(struct kobject *kobj,
219 				  struct kobj_attribute *attr, char *buf,
220 				  enum transparent_hugepage_flag flag)
221 {
222 	return sysfs_emit(buf, "%d\n",
223 			  !!test_bit(flag, &transparent_hugepage_flags));
224 }
225 
226 ssize_t single_hugepage_flag_store(struct kobject *kobj,
227 				 struct kobj_attribute *attr,
228 				 const char *buf, size_t count,
229 				 enum transparent_hugepage_flag flag)
230 {
231 	unsigned long value;
232 	int ret;
233 
234 	ret = kstrtoul(buf, 10, &value);
235 	if (ret < 0)
236 		return ret;
237 	if (value > 1)
238 		return -EINVAL;
239 
240 	if (value)
241 		set_bit(flag, &transparent_hugepage_flags);
242 	else
243 		clear_bit(flag, &transparent_hugepage_flags);
244 
245 	return count;
246 }
247 
248 static ssize_t defrag_show(struct kobject *kobj,
249 			   struct kobj_attribute *attr, char *buf)
250 {
251 	const char *output;
252 
253 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
254 		     &transparent_hugepage_flags))
255 		output = "[always] defer defer+madvise madvise never";
256 	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
257 			  &transparent_hugepage_flags))
258 		output = "always [defer] defer+madvise madvise never";
259 	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
260 			  &transparent_hugepage_flags))
261 		output = "always defer [defer+madvise] madvise never";
262 	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
263 			  &transparent_hugepage_flags))
264 		output = "always defer defer+madvise [madvise] never";
265 	else
266 		output = "always defer defer+madvise madvise [never]";
267 
268 	return sysfs_emit(buf, "%s\n", output);
269 }
270 
271 static ssize_t defrag_store(struct kobject *kobj,
272 			    struct kobj_attribute *attr,
273 			    const char *buf, size_t count)
274 {
275 	if (sysfs_streq(buf, "always")) {
276 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
277 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
278 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
279 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
280 	} else if (sysfs_streq(buf, "defer+madvise")) {
281 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
282 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
283 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
284 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
285 	} else if (sysfs_streq(buf, "defer")) {
286 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
287 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
288 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
289 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
290 	} else if (sysfs_streq(buf, "madvise")) {
291 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
292 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
293 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
294 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
295 	} else if (sysfs_streq(buf, "never")) {
296 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
297 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
298 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
299 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
300 	} else
301 		return -EINVAL;
302 
303 	return count;
304 }
305 static struct kobj_attribute defrag_attr =
306 	__ATTR(defrag, 0644, defrag_show, defrag_store);
307 
308 static ssize_t use_zero_page_show(struct kobject *kobj,
309 				  struct kobj_attribute *attr, char *buf)
310 {
311 	return single_hugepage_flag_show(kobj, attr, buf,
312 					 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
313 }
314 static ssize_t use_zero_page_store(struct kobject *kobj,
315 		struct kobj_attribute *attr, const char *buf, size_t count)
316 {
317 	return single_hugepage_flag_store(kobj, attr, buf, count,
318 				 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
319 }
320 static struct kobj_attribute use_zero_page_attr =
321 	__ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
322 
323 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
324 				   struct kobj_attribute *attr, char *buf)
325 {
326 	return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
327 }
328 static struct kobj_attribute hpage_pmd_size_attr =
329 	__ATTR_RO(hpage_pmd_size);
330 
331 static struct attribute *hugepage_attr[] = {
332 	&enabled_attr.attr,
333 	&defrag_attr.attr,
334 	&use_zero_page_attr.attr,
335 	&hpage_pmd_size_attr.attr,
336 #ifdef CONFIG_SHMEM
337 	&shmem_enabled_attr.attr,
338 #endif
339 	NULL,
340 };
341 
342 static const struct attribute_group hugepage_attr_group = {
343 	.attrs = hugepage_attr,
344 };
345 
346 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
347 {
348 	int err;
349 
350 	*hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
351 	if (unlikely(!*hugepage_kobj)) {
352 		pr_err("failed to create transparent hugepage kobject\n");
353 		return -ENOMEM;
354 	}
355 
356 	err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
357 	if (err) {
358 		pr_err("failed to register transparent hugepage group\n");
359 		goto delete_obj;
360 	}
361 
362 	err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
363 	if (err) {
364 		pr_err("failed to register transparent hugepage group\n");
365 		goto remove_hp_group;
366 	}
367 
368 	return 0;
369 
370 remove_hp_group:
371 	sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
372 delete_obj:
373 	kobject_put(*hugepage_kobj);
374 	return err;
375 }
376 
377 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
378 {
379 	sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
380 	sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
381 	kobject_put(hugepage_kobj);
382 }
383 #else
384 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
385 {
386 	return 0;
387 }
388 
389 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
390 {
391 }
392 #endif /* CONFIG_SYSFS */
393 
394 static int __init hugepage_init(void)
395 {
396 	int err;
397 	struct kobject *hugepage_kobj;
398 
399 	if (!has_transparent_hugepage()) {
400 		/*
401 		 * Hardware doesn't support hugepages, hence disable
402 		 * DAX PMD support.
403 		 */
404 		transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX;
405 		return -EINVAL;
406 	}
407 
408 	/*
409 	 * hugepages can't be allocated by the buddy allocator
410 	 */
411 	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
412 	/*
413 	 * we use page->mapping and page->index in second tail page
414 	 * as list_head: assuming THP order >= 2
415 	 */
416 	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
417 
418 	err = hugepage_init_sysfs(&hugepage_kobj);
419 	if (err)
420 		goto err_sysfs;
421 
422 	err = khugepaged_init();
423 	if (err)
424 		goto err_slab;
425 
426 	err = register_shrinker(&huge_zero_page_shrinker);
427 	if (err)
428 		goto err_hzp_shrinker;
429 	err = register_shrinker(&deferred_split_shrinker);
430 	if (err)
431 		goto err_split_shrinker;
432 
433 	/*
434 	 * By default disable transparent hugepages on smaller systems,
435 	 * where the extra memory used could hurt more than TLB overhead
436 	 * is likely to save.  The admin can still enable it through /sys.
437 	 */
438 	if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
439 		transparent_hugepage_flags = 0;
440 		return 0;
441 	}
442 
443 	err = start_stop_khugepaged();
444 	if (err)
445 		goto err_khugepaged;
446 
447 	return 0;
448 err_khugepaged:
449 	unregister_shrinker(&deferred_split_shrinker);
450 err_split_shrinker:
451 	unregister_shrinker(&huge_zero_page_shrinker);
452 err_hzp_shrinker:
453 	khugepaged_destroy();
454 err_slab:
455 	hugepage_exit_sysfs(hugepage_kobj);
456 err_sysfs:
457 	return err;
458 }
459 subsys_initcall(hugepage_init);
460 
461 static int __init setup_transparent_hugepage(char *str)
462 {
463 	int ret = 0;
464 	if (!str)
465 		goto out;
466 	if (!strcmp(str, "always")) {
467 		set_bit(TRANSPARENT_HUGEPAGE_FLAG,
468 			&transparent_hugepage_flags);
469 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
470 			  &transparent_hugepage_flags);
471 		ret = 1;
472 	} else if (!strcmp(str, "madvise")) {
473 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
474 			  &transparent_hugepage_flags);
475 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
476 			&transparent_hugepage_flags);
477 		ret = 1;
478 	} else if (!strcmp(str, "never")) {
479 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
480 			  &transparent_hugepage_flags);
481 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
482 			  &transparent_hugepage_flags);
483 		ret = 1;
484 	}
485 out:
486 	if (!ret)
487 		pr_warn("transparent_hugepage= cannot parse, ignored\n");
488 	return ret;
489 }
490 __setup("transparent_hugepage=", setup_transparent_hugepage);
491 
492 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
493 {
494 	if (likely(vma->vm_flags & VM_WRITE))
495 		pmd = pmd_mkwrite(pmd);
496 	return pmd;
497 }
498 
499 #ifdef CONFIG_MEMCG
500 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
501 {
502 	struct mem_cgroup *memcg = page_memcg(compound_head(page));
503 	struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
504 
505 	if (memcg)
506 		return &memcg->deferred_split_queue;
507 	else
508 		return &pgdat->deferred_split_queue;
509 }
510 #else
511 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
512 {
513 	struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
514 
515 	return &pgdat->deferred_split_queue;
516 }
517 #endif
518 
519 void prep_transhuge_page(struct page *page)
520 {
521 	/*
522 	 * we use page->mapping and page->indexlru in second tail page
523 	 * as list_head: assuming THP order >= 2
524 	 */
525 
526 	INIT_LIST_HEAD(page_deferred_list(page));
527 	set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
528 }
529 
530 static inline bool is_transparent_hugepage(struct page *page)
531 {
532 	if (!PageCompound(page))
533 		return false;
534 
535 	page = compound_head(page);
536 	return is_huge_zero_page(page) ||
537 	       page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
538 }
539 
540 static unsigned long __thp_get_unmapped_area(struct file *filp,
541 		unsigned long addr, unsigned long len,
542 		loff_t off, unsigned long flags, unsigned long size)
543 {
544 	loff_t off_end = off + len;
545 	loff_t off_align = round_up(off, size);
546 	unsigned long len_pad, ret;
547 
548 	if (off_end <= off_align || (off_end - off_align) < size)
549 		return 0;
550 
551 	len_pad = len + size;
552 	if (len_pad < len || (off + len_pad) < off)
553 		return 0;
554 
555 	ret = current->mm->get_unmapped_area(filp, addr, len_pad,
556 					      off >> PAGE_SHIFT, flags);
557 
558 	/*
559 	 * The failure might be due to length padding. The caller will retry
560 	 * without the padding.
561 	 */
562 	if (IS_ERR_VALUE(ret))
563 		return 0;
564 
565 	/*
566 	 * Do not try to align to THP boundary if allocation at the address
567 	 * hint succeeds.
568 	 */
569 	if (ret == addr)
570 		return addr;
571 
572 	ret += (off - ret) & (size - 1);
573 	return ret;
574 }
575 
576 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
577 		unsigned long len, unsigned long pgoff, unsigned long flags)
578 {
579 	unsigned long ret;
580 	loff_t off = (loff_t)pgoff << PAGE_SHIFT;
581 
582 	ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
583 	if (ret)
584 		return ret;
585 
586 	return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
587 }
588 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
589 
590 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
591 			struct page *page, gfp_t gfp)
592 {
593 	struct vm_area_struct *vma = vmf->vma;
594 	pgtable_t pgtable;
595 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
596 	vm_fault_t ret = 0;
597 
598 	VM_BUG_ON_PAGE(!PageCompound(page), page);
599 
600 	if (mem_cgroup_charge(page_folio(page), vma->vm_mm, gfp)) {
601 		put_page(page);
602 		count_vm_event(THP_FAULT_FALLBACK);
603 		count_vm_event(THP_FAULT_FALLBACK_CHARGE);
604 		return VM_FAULT_FALLBACK;
605 	}
606 	cgroup_throttle_swaprate(page, gfp);
607 
608 	pgtable = pte_alloc_one(vma->vm_mm);
609 	if (unlikely(!pgtable)) {
610 		ret = VM_FAULT_OOM;
611 		goto release;
612 	}
613 
614 	clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
615 	/*
616 	 * The memory barrier inside __SetPageUptodate makes sure that
617 	 * clear_huge_page writes become visible before the set_pmd_at()
618 	 * write.
619 	 */
620 	__SetPageUptodate(page);
621 
622 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
623 	if (unlikely(!pmd_none(*vmf->pmd))) {
624 		goto unlock_release;
625 	} else {
626 		pmd_t entry;
627 
628 		ret = check_stable_address_space(vma->vm_mm);
629 		if (ret)
630 			goto unlock_release;
631 
632 		/* Deliver the page fault to userland */
633 		if (userfaultfd_missing(vma)) {
634 			spin_unlock(vmf->ptl);
635 			put_page(page);
636 			pte_free(vma->vm_mm, pgtable);
637 			ret = handle_userfault(vmf, VM_UFFD_MISSING);
638 			VM_BUG_ON(ret & VM_FAULT_FALLBACK);
639 			return ret;
640 		}
641 
642 		entry = mk_huge_pmd(page, vma->vm_page_prot);
643 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
644 		page_add_new_anon_rmap(page, vma, haddr);
645 		lru_cache_add_inactive_or_unevictable(page, vma);
646 		pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
647 		set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
648 		update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
649 		add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
650 		mm_inc_nr_ptes(vma->vm_mm);
651 		spin_unlock(vmf->ptl);
652 		count_vm_event(THP_FAULT_ALLOC);
653 		count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
654 	}
655 
656 	return 0;
657 unlock_release:
658 	spin_unlock(vmf->ptl);
659 release:
660 	if (pgtable)
661 		pte_free(vma->vm_mm, pgtable);
662 	put_page(page);
663 	return ret;
664 
665 }
666 
667 /*
668  * always: directly stall for all thp allocations
669  * defer: wake kswapd and fail if not immediately available
670  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
671  *		  fail if not immediately available
672  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
673  *	    available
674  * never: never stall for any thp allocation
675  */
676 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
677 {
678 	const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
679 
680 	/* Always do synchronous compaction */
681 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
682 		return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
683 
684 	/* Kick kcompactd and fail quickly */
685 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
686 		return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
687 
688 	/* Synchronous compaction if madvised, otherwise kick kcompactd */
689 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
690 		return GFP_TRANSHUGE_LIGHT |
691 			(vma_madvised ? __GFP_DIRECT_RECLAIM :
692 					__GFP_KSWAPD_RECLAIM);
693 
694 	/* Only do synchronous compaction if madvised */
695 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
696 		return GFP_TRANSHUGE_LIGHT |
697 		       (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
698 
699 	return GFP_TRANSHUGE_LIGHT;
700 }
701 
702 /* Caller must hold page table lock. */
703 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
704 		struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
705 		struct page *zero_page)
706 {
707 	pmd_t entry;
708 	if (!pmd_none(*pmd))
709 		return;
710 	entry = mk_pmd(zero_page, vma->vm_page_prot);
711 	entry = pmd_mkhuge(entry);
712 	if (pgtable)
713 		pgtable_trans_huge_deposit(mm, pmd, pgtable);
714 	set_pmd_at(mm, haddr, pmd, entry);
715 	mm_inc_nr_ptes(mm);
716 }
717 
718 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
719 {
720 	struct vm_area_struct *vma = vmf->vma;
721 	gfp_t gfp;
722 	struct folio *folio;
723 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
724 
725 	if (!transhuge_vma_suitable(vma, haddr))
726 		return VM_FAULT_FALLBACK;
727 	if (unlikely(anon_vma_prepare(vma)))
728 		return VM_FAULT_OOM;
729 	khugepaged_enter(vma, vma->vm_flags);
730 
731 	if (!(vmf->flags & FAULT_FLAG_WRITE) &&
732 			!mm_forbids_zeropage(vma->vm_mm) &&
733 			transparent_hugepage_use_zero_page()) {
734 		pgtable_t pgtable;
735 		struct page *zero_page;
736 		vm_fault_t ret;
737 		pgtable = pte_alloc_one(vma->vm_mm);
738 		if (unlikely(!pgtable))
739 			return VM_FAULT_OOM;
740 		zero_page = mm_get_huge_zero_page(vma->vm_mm);
741 		if (unlikely(!zero_page)) {
742 			pte_free(vma->vm_mm, pgtable);
743 			count_vm_event(THP_FAULT_FALLBACK);
744 			return VM_FAULT_FALLBACK;
745 		}
746 		vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
747 		ret = 0;
748 		if (pmd_none(*vmf->pmd)) {
749 			ret = check_stable_address_space(vma->vm_mm);
750 			if (ret) {
751 				spin_unlock(vmf->ptl);
752 				pte_free(vma->vm_mm, pgtable);
753 			} else if (userfaultfd_missing(vma)) {
754 				spin_unlock(vmf->ptl);
755 				pte_free(vma->vm_mm, pgtable);
756 				ret = handle_userfault(vmf, VM_UFFD_MISSING);
757 				VM_BUG_ON(ret & VM_FAULT_FALLBACK);
758 			} else {
759 				set_huge_zero_page(pgtable, vma->vm_mm, vma,
760 						   haddr, vmf->pmd, zero_page);
761 				update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
762 				spin_unlock(vmf->ptl);
763 			}
764 		} else {
765 			spin_unlock(vmf->ptl);
766 			pte_free(vma->vm_mm, pgtable);
767 		}
768 		return ret;
769 	}
770 	gfp = vma_thp_gfp_mask(vma);
771 	folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, vma, haddr, true);
772 	if (unlikely(!folio)) {
773 		count_vm_event(THP_FAULT_FALLBACK);
774 		return VM_FAULT_FALLBACK;
775 	}
776 	return __do_huge_pmd_anonymous_page(vmf, &folio->page, gfp);
777 }
778 
779 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
780 		pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
781 		pgtable_t pgtable)
782 {
783 	struct mm_struct *mm = vma->vm_mm;
784 	pmd_t entry;
785 	spinlock_t *ptl;
786 
787 	ptl = pmd_lock(mm, pmd);
788 	if (!pmd_none(*pmd)) {
789 		if (write) {
790 			if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
791 				WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
792 				goto out_unlock;
793 			}
794 			entry = pmd_mkyoung(*pmd);
795 			entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
796 			if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
797 				update_mmu_cache_pmd(vma, addr, pmd);
798 		}
799 
800 		goto out_unlock;
801 	}
802 
803 	entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
804 	if (pfn_t_devmap(pfn))
805 		entry = pmd_mkdevmap(entry);
806 	if (write) {
807 		entry = pmd_mkyoung(pmd_mkdirty(entry));
808 		entry = maybe_pmd_mkwrite(entry, vma);
809 	}
810 
811 	if (pgtable) {
812 		pgtable_trans_huge_deposit(mm, pmd, pgtable);
813 		mm_inc_nr_ptes(mm);
814 		pgtable = NULL;
815 	}
816 
817 	set_pmd_at(mm, addr, pmd, entry);
818 	update_mmu_cache_pmd(vma, addr, pmd);
819 
820 out_unlock:
821 	spin_unlock(ptl);
822 	if (pgtable)
823 		pte_free(mm, pgtable);
824 }
825 
826 /**
827  * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
828  * @vmf: Structure describing the fault
829  * @pfn: pfn to insert
830  * @pgprot: page protection to use
831  * @write: whether it's a write fault
832  *
833  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
834  * also consult the vmf_insert_mixed_prot() documentation when
835  * @pgprot != @vmf->vma->vm_page_prot.
836  *
837  * Return: vm_fault_t value.
838  */
839 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
840 				   pgprot_t pgprot, bool write)
841 {
842 	unsigned long addr = vmf->address & PMD_MASK;
843 	struct vm_area_struct *vma = vmf->vma;
844 	pgtable_t pgtable = NULL;
845 
846 	/*
847 	 * If we had pmd_special, we could avoid all these restrictions,
848 	 * but we need to be consistent with PTEs and architectures that
849 	 * can't support a 'special' bit.
850 	 */
851 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
852 			!pfn_t_devmap(pfn));
853 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
854 						(VM_PFNMAP|VM_MIXEDMAP));
855 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
856 
857 	if (addr < vma->vm_start || addr >= vma->vm_end)
858 		return VM_FAULT_SIGBUS;
859 
860 	if (arch_needs_pgtable_deposit()) {
861 		pgtable = pte_alloc_one(vma->vm_mm);
862 		if (!pgtable)
863 			return VM_FAULT_OOM;
864 	}
865 
866 	track_pfn_insert(vma, &pgprot, pfn);
867 
868 	insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
869 	return VM_FAULT_NOPAGE;
870 }
871 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
872 
873 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
874 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
875 {
876 	if (likely(vma->vm_flags & VM_WRITE))
877 		pud = pud_mkwrite(pud);
878 	return pud;
879 }
880 
881 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
882 		pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
883 {
884 	struct mm_struct *mm = vma->vm_mm;
885 	pud_t entry;
886 	spinlock_t *ptl;
887 
888 	ptl = pud_lock(mm, pud);
889 	if (!pud_none(*pud)) {
890 		if (write) {
891 			if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
892 				WARN_ON_ONCE(!is_huge_zero_pud(*pud));
893 				goto out_unlock;
894 			}
895 			entry = pud_mkyoung(*pud);
896 			entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
897 			if (pudp_set_access_flags(vma, addr, pud, entry, 1))
898 				update_mmu_cache_pud(vma, addr, pud);
899 		}
900 		goto out_unlock;
901 	}
902 
903 	entry = pud_mkhuge(pfn_t_pud(pfn, prot));
904 	if (pfn_t_devmap(pfn))
905 		entry = pud_mkdevmap(entry);
906 	if (write) {
907 		entry = pud_mkyoung(pud_mkdirty(entry));
908 		entry = maybe_pud_mkwrite(entry, vma);
909 	}
910 	set_pud_at(mm, addr, pud, entry);
911 	update_mmu_cache_pud(vma, addr, pud);
912 
913 out_unlock:
914 	spin_unlock(ptl);
915 }
916 
917 /**
918  * vmf_insert_pfn_pud_prot - insert a pud size pfn
919  * @vmf: Structure describing the fault
920  * @pfn: pfn to insert
921  * @pgprot: page protection to use
922  * @write: whether it's a write fault
923  *
924  * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
925  * also consult the vmf_insert_mixed_prot() documentation when
926  * @pgprot != @vmf->vma->vm_page_prot.
927  *
928  * Return: vm_fault_t value.
929  */
930 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
931 				   pgprot_t pgprot, bool write)
932 {
933 	unsigned long addr = vmf->address & PUD_MASK;
934 	struct vm_area_struct *vma = vmf->vma;
935 
936 	/*
937 	 * If we had pud_special, we could avoid all these restrictions,
938 	 * but we need to be consistent with PTEs and architectures that
939 	 * can't support a 'special' bit.
940 	 */
941 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
942 			!pfn_t_devmap(pfn));
943 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
944 						(VM_PFNMAP|VM_MIXEDMAP));
945 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
946 
947 	if (addr < vma->vm_start || addr >= vma->vm_end)
948 		return VM_FAULT_SIGBUS;
949 
950 	track_pfn_insert(vma, &pgprot, pfn);
951 
952 	insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
953 	return VM_FAULT_NOPAGE;
954 }
955 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
956 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
957 
958 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
959 		pmd_t *pmd, int flags)
960 {
961 	pmd_t _pmd;
962 
963 	_pmd = pmd_mkyoung(*pmd);
964 	if (flags & FOLL_WRITE)
965 		_pmd = pmd_mkdirty(_pmd);
966 	if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
967 				pmd, _pmd, flags & FOLL_WRITE))
968 		update_mmu_cache_pmd(vma, addr, pmd);
969 }
970 
971 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
972 		pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
973 {
974 	unsigned long pfn = pmd_pfn(*pmd);
975 	struct mm_struct *mm = vma->vm_mm;
976 	struct page *page;
977 
978 	assert_spin_locked(pmd_lockptr(mm, pmd));
979 
980 	/*
981 	 * When we COW a devmap PMD entry, we split it into PTEs, so we should
982 	 * not be in this function with `flags & FOLL_COW` set.
983 	 */
984 	WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
985 
986 	/* FOLL_GET and FOLL_PIN are mutually exclusive. */
987 	if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
988 			 (FOLL_PIN | FOLL_GET)))
989 		return NULL;
990 
991 	if (flags & FOLL_WRITE && !pmd_write(*pmd))
992 		return NULL;
993 
994 	if (pmd_present(*pmd) && pmd_devmap(*pmd))
995 		/* pass */;
996 	else
997 		return NULL;
998 
999 	if (flags & FOLL_TOUCH)
1000 		touch_pmd(vma, addr, pmd, flags);
1001 
1002 	/*
1003 	 * device mapped pages can only be returned if the
1004 	 * caller will manage the page reference count.
1005 	 */
1006 	if (!(flags & (FOLL_GET | FOLL_PIN)))
1007 		return ERR_PTR(-EEXIST);
1008 
1009 	pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1010 	*pgmap = get_dev_pagemap(pfn, *pgmap);
1011 	if (!*pgmap)
1012 		return ERR_PTR(-EFAULT);
1013 	page = pfn_to_page(pfn);
1014 	if (!try_grab_page(page, flags))
1015 		page = ERR_PTR(-ENOMEM);
1016 
1017 	return page;
1018 }
1019 
1020 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1021 		  pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1022 		  struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1023 {
1024 	spinlock_t *dst_ptl, *src_ptl;
1025 	struct page *src_page;
1026 	pmd_t pmd;
1027 	pgtable_t pgtable = NULL;
1028 	int ret = -ENOMEM;
1029 
1030 	/* Skip if can be re-fill on fault */
1031 	if (!vma_is_anonymous(dst_vma))
1032 		return 0;
1033 
1034 	pgtable = pte_alloc_one(dst_mm);
1035 	if (unlikely(!pgtable))
1036 		goto out;
1037 
1038 	dst_ptl = pmd_lock(dst_mm, dst_pmd);
1039 	src_ptl = pmd_lockptr(src_mm, src_pmd);
1040 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1041 
1042 	ret = -EAGAIN;
1043 	pmd = *src_pmd;
1044 
1045 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1046 	if (unlikely(is_swap_pmd(pmd))) {
1047 		swp_entry_t entry = pmd_to_swp_entry(pmd);
1048 
1049 		VM_BUG_ON(!is_pmd_migration_entry(pmd));
1050 		if (!is_readable_migration_entry(entry)) {
1051 			entry = make_readable_migration_entry(
1052 							swp_offset(entry));
1053 			pmd = swp_entry_to_pmd(entry);
1054 			if (pmd_swp_soft_dirty(*src_pmd))
1055 				pmd = pmd_swp_mksoft_dirty(pmd);
1056 			if (pmd_swp_uffd_wp(*src_pmd))
1057 				pmd = pmd_swp_mkuffd_wp(pmd);
1058 			set_pmd_at(src_mm, addr, src_pmd, pmd);
1059 		}
1060 		add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1061 		mm_inc_nr_ptes(dst_mm);
1062 		pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1063 		if (!userfaultfd_wp(dst_vma))
1064 			pmd = pmd_swp_clear_uffd_wp(pmd);
1065 		set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1066 		ret = 0;
1067 		goto out_unlock;
1068 	}
1069 #endif
1070 
1071 	if (unlikely(!pmd_trans_huge(pmd))) {
1072 		pte_free(dst_mm, pgtable);
1073 		goto out_unlock;
1074 	}
1075 	/*
1076 	 * When page table lock is held, the huge zero pmd should not be
1077 	 * under splitting since we don't split the page itself, only pmd to
1078 	 * a page table.
1079 	 */
1080 	if (is_huge_zero_pmd(pmd)) {
1081 		/*
1082 		 * get_huge_zero_page() will never allocate a new page here,
1083 		 * since we already have a zero page to copy. It just takes a
1084 		 * reference.
1085 		 */
1086 		mm_get_huge_zero_page(dst_mm);
1087 		goto out_zero_page;
1088 	}
1089 
1090 	src_page = pmd_page(pmd);
1091 	VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1092 
1093 	get_page(src_page);
1094 	if (unlikely(page_try_dup_anon_rmap(src_page, true, src_vma))) {
1095 		/* Page maybe pinned: split and retry the fault on PTEs. */
1096 		put_page(src_page);
1097 		pte_free(dst_mm, pgtable);
1098 		spin_unlock(src_ptl);
1099 		spin_unlock(dst_ptl);
1100 		__split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1101 		return -EAGAIN;
1102 	}
1103 	add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1104 out_zero_page:
1105 	mm_inc_nr_ptes(dst_mm);
1106 	pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1107 	pmdp_set_wrprotect(src_mm, addr, src_pmd);
1108 	if (!userfaultfd_wp(dst_vma))
1109 		pmd = pmd_clear_uffd_wp(pmd);
1110 	pmd = pmd_mkold(pmd_wrprotect(pmd));
1111 	set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1112 
1113 	ret = 0;
1114 out_unlock:
1115 	spin_unlock(src_ptl);
1116 	spin_unlock(dst_ptl);
1117 out:
1118 	return ret;
1119 }
1120 
1121 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1122 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1123 		pud_t *pud, int flags)
1124 {
1125 	pud_t _pud;
1126 
1127 	_pud = pud_mkyoung(*pud);
1128 	if (flags & FOLL_WRITE)
1129 		_pud = pud_mkdirty(_pud);
1130 	if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1131 				pud, _pud, flags & FOLL_WRITE))
1132 		update_mmu_cache_pud(vma, addr, pud);
1133 }
1134 
1135 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1136 		pud_t *pud, int flags, struct dev_pagemap **pgmap)
1137 {
1138 	unsigned long pfn = pud_pfn(*pud);
1139 	struct mm_struct *mm = vma->vm_mm;
1140 	struct page *page;
1141 
1142 	assert_spin_locked(pud_lockptr(mm, pud));
1143 
1144 	if (flags & FOLL_WRITE && !pud_write(*pud))
1145 		return NULL;
1146 
1147 	/* FOLL_GET and FOLL_PIN are mutually exclusive. */
1148 	if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1149 			 (FOLL_PIN | FOLL_GET)))
1150 		return NULL;
1151 
1152 	if (pud_present(*pud) && pud_devmap(*pud))
1153 		/* pass */;
1154 	else
1155 		return NULL;
1156 
1157 	if (flags & FOLL_TOUCH)
1158 		touch_pud(vma, addr, pud, flags);
1159 
1160 	/*
1161 	 * device mapped pages can only be returned if the
1162 	 * caller will manage the page reference count.
1163 	 *
1164 	 * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1165 	 */
1166 	if (!(flags & (FOLL_GET | FOLL_PIN)))
1167 		return ERR_PTR(-EEXIST);
1168 
1169 	pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1170 	*pgmap = get_dev_pagemap(pfn, *pgmap);
1171 	if (!*pgmap)
1172 		return ERR_PTR(-EFAULT);
1173 	page = pfn_to_page(pfn);
1174 	if (!try_grab_page(page, flags))
1175 		page = ERR_PTR(-ENOMEM);
1176 
1177 	return page;
1178 }
1179 
1180 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1181 		  pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1182 		  struct vm_area_struct *vma)
1183 {
1184 	spinlock_t *dst_ptl, *src_ptl;
1185 	pud_t pud;
1186 	int ret;
1187 
1188 	dst_ptl = pud_lock(dst_mm, dst_pud);
1189 	src_ptl = pud_lockptr(src_mm, src_pud);
1190 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1191 
1192 	ret = -EAGAIN;
1193 	pud = *src_pud;
1194 	if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1195 		goto out_unlock;
1196 
1197 	/*
1198 	 * When page table lock is held, the huge zero pud should not be
1199 	 * under splitting since we don't split the page itself, only pud to
1200 	 * a page table.
1201 	 */
1202 	if (is_huge_zero_pud(pud)) {
1203 		/* No huge zero pud yet */
1204 	}
1205 
1206 	/*
1207 	 * TODO: once we support anonymous pages, use page_try_dup_anon_rmap()
1208 	 * and split if duplicating fails.
1209 	 */
1210 	pudp_set_wrprotect(src_mm, addr, src_pud);
1211 	pud = pud_mkold(pud_wrprotect(pud));
1212 	set_pud_at(dst_mm, addr, dst_pud, pud);
1213 
1214 	ret = 0;
1215 out_unlock:
1216 	spin_unlock(src_ptl);
1217 	spin_unlock(dst_ptl);
1218 	return ret;
1219 }
1220 
1221 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1222 {
1223 	pud_t entry;
1224 	unsigned long haddr;
1225 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1226 
1227 	vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1228 	if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1229 		goto unlock;
1230 
1231 	entry = pud_mkyoung(orig_pud);
1232 	if (write)
1233 		entry = pud_mkdirty(entry);
1234 	haddr = vmf->address & HPAGE_PUD_MASK;
1235 	if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1236 		update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1237 
1238 unlock:
1239 	spin_unlock(vmf->ptl);
1240 }
1241 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1242 
1243 void huge_pmd_set_accessed(struct vm_fault *vmf)
1244 {
1245 	pmd_t entry;
1246 	unsigned long haddr;
1247 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1248 	pmd_t orig_pmd = vmf->orig_pmd;
1249 
1250 	vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1251 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1252 		goto unlock;
1253 
1254 	entry = pmd_mkyoung(orig_pmd);
1255 	if (write)
1256 		entry = pmd_mkdirty(entry);
1257 	haddr = vmf->address & HPAGE_PMD_MASK;
1258 	if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1259 		update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1260 
1261 unlock:
1262 	spin_unlock(vmf->ptl);
1263 }
1264 
1265 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)
1266 {
1267 	const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;
1268 	struct vm_area_struct *vma = vmf->vma;
1269 	struct page *page;
1270 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1271 	pmd_t orig_pmd = vmf->orig_pmd;
1272 
1273 	vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1274 	VM_BUG_ON_VMA(!vma->anon_vma, vma);
1275 
1276 	VM_BUG_ON(unshare && (vmf->flags & FAULT_FLAG_WRITE));
1277 	VM_BUG_ON(!unshare && !(vmf->flags & FAULT_FLAG_WRITE));
1278 
1279 	if (is_huge_zero_pmd(orig_pmd))
1280 		goto fallback;
1281 
1282 	spin_lock(vmf->ptl);
1283 
1284 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1285 		spin_unlock(vmf->ptl);
1286 		return 0;
1287 	}
1288 
1289 	page = pmd_page(orig_pmd);
1290 	VM_BUG_ON_PAGE(!PageHead(page), page);
1291 
1292 	/* Early check when only holding the PT lock. */
1293 	if (PageAnonExclusive(page))
1294 		goto reuse;
1295 
1296 	if (!trylock_page(page)) {
1297 		get_page(page);
1298 		spin_unlock(vmf->ptl);
1299 		lock_page(page);
1300 		spin_lock(vmf->ptl);
1301 		if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1302 			spin_unlock(vmf->ptl);
1303 			unlock_page(page);
1304 			put_page(page);
1305 			return 0;
1306 		}
1307 		put_page(page);
1308 	}
1309 
1310 	/* Recheck after temporarily dropping the PT lock. */
1311 	if (PageAnonExclusive(page)) {
1312 		unlock_page(page);
1313 		goto reuse;
1314 	}
1315 
1316 	/*
1317 	 * See do_wp_page(): we can only reuse the page exclusively if there are
1318 	 * no additional references. Note that we always drain the LRU
1319 	 * pagevecs immediately after adding a THP.
1320 	 */
1321 	if (page_count(page) > 1 + PageSwapCache(page) * thp_nr_pages(page))
1322 		goto unlock_fallback;
1323 	if (PageSwapCache(page))
1324 		try_to_free_swap(page);
1325 	if (page_count(page) == 1) {
1326 		pmd_t entry;
1327 
1328 		page_move_anon_rmap(page, vma);
1329 		unlock_page(page);
1330 reuse:
1331 		if (unlikely(unshare)) {
1332 			spin_unlock(vmf->ptl);
1333 			return 0;
1334 		}
1335 		entry = pmd_mkyoung(orig_pmd);
1336 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1337 		if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1338 			update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1339 		spin_unlock(vmf->ptl);
1340 		return VM_FAULT_WRITE;
1341 	}
1342 
1343 unlock_fallback:
1344 	unlock_page(page);
1345 	spin_unlock(vmf->ptl);
1346 fallback:
1347 	__split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1348 	return VM_FAULT_FALLBACK;
1349 }
1350 
1351 /*
1352  * FOLL_FORCE can write to even unwritable pmd's, but only
1353  * after we've gone through a COW cycle and they are dirty.
1354  */
1355 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1356 {
1357 	return pmd_write(pmd) ||
1358 	       ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1359 }
1360 
1361 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1362 				   unsigned long addr,
1363 				   pmd_t *pmd,
1364 				   unsigned int flags)
1365 {
1366 	struct mm_struct *mm = vma->vm_mm;
1367 	struct page *page = NULL;
1368 
1369 	assert_spin_locked(pmd_lockptr(mm, pmd));
1370 
1371 	if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1372 		goto out;
1373 
1374 	/* Avoid dumping huge zero page */
1375 	if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1376 		return ERR_PTR(-EFAULT);
1377 
1378 	/* Full NUMA hinting faults to serialise migration in fault paths */
1379 	if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1380 		goto out;
1381 
1382 	page = pmd_page(*pmd);
1383 	VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1384 
1385 	if (!pmd_write(*pmd) && gup_must_unshare(flags, page))
1386 		return ERR_PTR(-EMLINK);
1387 
1388 	VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
1389 			!PageAnonExclusive(page), page);
1390 
1391 	if (!try_grab_page(page, flags))
1392 		return ERR_PTR(-ENOMEM);
1393 
1394 	if (flags & FOLL_TOUCH)
1395 		touch_pmd(vma, addr, pmd, flags);
1396 
1397 	page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1398 	VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1399 
1400 out:
1401 	return page;
1402 }
1403 
1404 /* NUMA hinting page fault entry point for trans huge pmds */
1405 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
1406 {
1407 	struct vm_area_struct *vma = vmf->vma;
1408 	pmd_t oldpmd = vmf->orig_pmd;
1409 	pmd_t pmd;
1410 	struct page *page;
1411 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1412 	int page_nid = NUMA_NO_NODE;
1413 	int target_nid, last_cpupid = -1;
1414 	bool migrated = false;
1415 	bool was_writable = pmd_savedwrite(oldpmd);
1416 	int flags = 0;
1417 
1418 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1419 	if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1420 		spin_unlock(vmf->ptl);
1421 		goto out;
1422 	}
1423 
1424 	pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1425 	page = vm_normal_page_pmd(vma, haddr, pmd);
1426 	if (!page)
1427 		goto out_map;
1428 
1429 	/* See similar comment in do_numa_page for explanation */
1430 	if (!was_writable)
1431 		flags |= TNF_NO_GROUP;
1432 
1433 	page_nid = page_to_nid(page);
1434 	last_cpupid = page_cpupid_last(page);
1435 	target_nid = numa_migrate_prep(page, vma, haddr, page_nid,
1436 				       &flags);
1437 
1438 	if (target_nid == NUMA_NO_NODE) {
1439 		put_page(page);
1440 		goto out_map;
1441 	}
1442 
1443 	spin_unlock(vmf->ptl);
1444 
1445 	migrated = migrate_misplaced_page(page, vma, target_nid);
1446 	if (migrated) {
1447 		flags |= TNF_MIGRATED;
1448 		page_nid = target_nid;
1449 	} else {
1450 		flags |= TNF_MIGRATE_FAIL;
1451 		vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1452 		if (unlikely(!pmd_same(oldpmd, *vmf->pmd))) {
1453 			spin_unlock(vmf->ptl);
1454 			goto out;
1455 		}
1456 		goto out_map;
1457 	}
1458 
1459 out:
1460 	if (page_nid != NUMA_NO_NODE)
1461 		task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1462 				flags);
1463 
1464 	return 0;
1465 
1466 out_map:
1467 	/* Restore the PMD */
1468 	pmd = pmd_modify(oldpmd, vma->vm_page_prot);
1469 	pmd = pmd_mkyoung(pmd);
1470 	if (was_writable)
1471 		pmd = pmd_mkwrite(pmd);
1472 	set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1473 	update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1474 	spin_unlock(vmf->ptl);
1475 	goto out;
1476 }
1477 
1478 /*
1479  * Return true if we do MADV_FREE successfully on entire pmd page.
1480  * Otherwise, return false.
1481  */
1482 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1483 		pmd_t *pmd, unsigned long addr, unsigned long next)
1484 {
1485 	spinlock_t *ptl;
1486 	pmd_t orig_pmd;
1487 	struct page *page;
1488 	struct mm_struct *mm = tlb->mm;
1489 	bool ret = false;
1490 
1491 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1492 
1493 	ptl = pmd_trans_huge_lock(pmd, vma);
1494 	if (!ptl)
1495 		goto out_unlocked;
1496 
1497 	orig_pmd = *pmd;
1498 	if (is_huge_zero_pmd(orig_pmd))
1499 		goto out;
1500 
1501 	if (unlikely(!pmd_present(orig_pmd))) {
1502 		VM_BUG_ON(thp_migration_supported() &&
1503 				  !is_pmd_migration_entry(orig_pmd));
1504 		goto out;
1505 	}
1506 
1507 	page = pmd_page(orig_pmd);
1508 	/*
1509 	 * If other processes are mapping this page, we couldn't discard
1510 	 * the page unless they all do MADV_FREE so let's skip the page.
1511 	 */
1512 	if (total_mapcount(page) != 1)
1513 		goto out;
1514 
1515 	if (!trylock_page(page))
1516 		goto out;
1517 
1518 	/*
1519 	 * If user want to discard part-pages of THP, split it so MADV_FREE
1520 	 * will deactivate only them.
1521 	 */
1522 	if (next - addr != HPAGE_PMD_SIZE) {
1523 		get_page(page);
1524 		spin_unlock(ptl);
1525 		split_huge_page(page);
1526 		unlock_page(page);
1527 		put_page(page);
1528 		goto out_unlocked;
1529 	}
1530 
1531 	if (PageDirty(page))
1532 		ClearPageDirty(page);
1533 	unlock_page(page);
1534 
1535 	if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1536 		pmdp_invalidate(vma, addr, pmd);
1537 		orig_pmd = pmd_mkold(orig_pmd);
1538 		orig_pmd = pmd_mkclean(orig_pmd);
1539 
1540 		set_pmd_at(mm, addr, pmd, orig_pmd);
1541 		tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1542 	}
1543 
1544 	mark_page_lazyfree(page);
1545 	ret = true;
1546 out:
1547 	spin_unlock(ptl);
1548 out_unlocked:
1549 	return ret;
1550 }
1551 
1552 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1553 {
1554 	pgtable_t pgtable;
1555 
1556 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1557 	pte_free(mm, pgtable);
1558 	mm_dec_nr_ptes(mm);
1559 }
1560 
1561 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1562 		 pmd_t *pmd, unsigned long addr)
1563 {
1564 	pmd_t orig_pmd;
1565 	spinlock_t *ptl;
1566 
1567 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1568 
1569 	ptl = __pmd_trans_huge_lock(pmd, vma);
1570 	if (!ptl)
1571 		return 0;
1572 	/*
1573 	 * For architectures like ppc64 we look at deposited pgtable
1574 	 * when calling pmdp_huge_get_and_clear. So do the
1575 	 * pgtable_trans_huge_withdraw after finishing pmdp related
1576 	 * operations.
1577 	 */
1578 	orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1579 						tlb->fullmm);
1580 	tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1581 	if (vma_is_special_huge(vma)) {
1582 		if (arch_needs_pgtable_deposit())
1583 			zap_deposited_table(tlb->mm, pmd);
1584 		spin_unlock(ptl);
1585 	} else if (is_huge_zero_pmd(orig_pmd)) {
1586 		zap_deposited_table(tlb->mm, pmd);
1587 		spin_unlock(ptl);
1588 	} else {
1589 		struct page *page = NULL;
1590 		int flush_needed = 1;
1591 
1592 		if (pmd_present(orig_pmd)) {
1593 			page = pmd_page(orig_pmd);
1594 			page_remove_rmap(page, vma, true);
1595 			VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1596 			VM_BUG_ON_PAGE(!PageHead(page), page);
1597 		} else if (thp_migration_supported()) {
1598 			swp_entry_t entry;
1599 
1600 			VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1601 			entry = pmd_to_swp_entry(orig_pmd);
1602 			page = pfn_swap_entry_to_page(entry);
1603 			flush_needed = 0;
1604 		} else
1605 			WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1606 
1607 		if (PageAnon(page)) {
1608 			zap_deposited_table(tlb->mm, pmd);
1609 			add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1610 		} else {
1611 			if (arch_needs_pgtable_deposit())
1612 				zap_deposited_table(tlb->mm, pmd);
1613 			add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1614 		}
1615 
1616 		spin_unlock(ptl);
1617 		if (flush_needed)
1618 			tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1619 	}
1620 	return 1;
1621 }
1622 
1623 #ifndef pmd_move_must_withdraw
1624 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1625 					 spinlock_t *old_pmd_ptl,
1626 					 struct vm_area_struct *vma)
1627 {
1628 	/*
1629 	 * With split pmd lock we also need to move preallocated
1630 	 * PTE page table if new_pmd is on different PMD page table.
1631 	 *
1632 	 * We also don't deposit and withdraw tables for file pages.
1633 	 */
1634 	return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1635 }
1636 #endif
1637 
1638 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1639 {
1640 #ifdef CONFIG_MEM_SOFT_DIRTY
1641 	if (unlikely(is_pmd_migration_entry(pmd)))
1642 		pmd = pmd_swp_mksoft_dirty(pmd);
1643 	else if (pmd_present(pmd))
1644 		pmd = pmd_mksoft_dirty(pmd);
1645 #endif
1646 	return pmd;
1647 }
1648 
1649 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1650 		  unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1651 {
1652 	spinlock_t *old_ptl, *new_ptl;
1653 	pmd_t pmd;
1654 	struct mm_struct *mm = vma->vm_mm;
1655 	bool force_flush = false;
1656 
1657 	/*
1658 	 * The destination pmd shouldn't be established, free_pgtables()
1659 	 * should have release it.
1660 	 */
1661 	if (WARN_ON(!pmd_none(*new_pmd))) {
1662 		VM_BUG_ON(pmd_trans_huge(*new_pmd));
1663 		return false;
1664 	}
1665 
1666 	/*
1667 	 * We don't have to worry about the ordering of src and dst
1668 	 * ptlocks because exclusive mmap_lock prevents deadlock.
1669 	 */
1670 	old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1671 	if (old_ptl) {
1672 		new_ptl = pmd_lockptr(mm, new_pmd);
1673 		if (new_ptl != old_ptl)
1674 			spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1675 		pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1676 		if (pmd_present(pmd))
1677 			force_flush = true;
1678 		VM_BUG_ON(!pmd_none(*new_pmd));
1679 
1680 		if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1681 			pgtable_t pgtable;
1682 			pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1683 			pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1684 		}
1685 		pmd = move_soft_dirty_pmd(pmd);
1686 		set_pmd_at(mm, new_addr, new_pmd, pmd);
1687 		if (force_flush)
1688 			flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1689 		if (new_ptl != old_ptl)
1690 			spin_unlock(new_ptl);
1691 		spin_unlock(old_ptl);
1692 		return true;
1693 	}
1694 	return false;
1695 }
1696 
1697 /*
1698  * Returns
1699  *  - 0 if PMD could not be locked
1700  *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1701  *      or if prot_numa but THP migration is not supported
1702  *  - HPAGE_PMD_NR if protections changed and TLB flush necessary
1703  */
1704 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1705 		    pmd_t *pmd, unsigned long addr, pgprot_t newprot,
1706 		    unsigned long cp_flags)
1707 {
1708 	struct mm_struct *mm = vma->vm_mm;
1709 	spinlock_t *ptl;
1710 	pmd_t oldpmd, entry;
1711 	bool preserve_write;
1712 	int ret;
1713 	bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1714 	bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1715 	bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1716 
1717 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1718 
1719 	if (prot_numa && !thp_migration_supported())
1720 		return 1;
1721 
1722 	ptl = __pmd_trans_huge_lock(pmd, vma);
1723 	if (!ptl)
1724 		return 0;
1725 
1726 	preserve_write = prot_numa && pmd_write(*pmd);
1727 	ret = 1;
1728 
1729 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1730 	if (is_swap_pmd(*pmd)) {
1731 		swp_entry_t entry = pmd_to_swp_entry(*pmd);
1732 		struct page *page = pfn_swap_entry_to_page(entry);
1733 
1734 		VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1735 		if (is_writable_migration_entry(entry)) {
1736 			pmd_t newpmd;
1737 			/*
1738 			 * A protection check is difficult so
1739 			 * just be safe and disable write
1740 			 */
1741 			if (PageAnon(page))
1742 				entry = make_readable_exclusive_migration_entry(swp_offset(entry));
1743 			else
1744 				entry = make_readable_migration_entry(swp_offset(entry));
1745 			newpmd = swp_entry_to_pmd(entry);
1746 			if (pmd_swp_soft_dirty(*pmd))
1747 				newpmd = pmd_swp_mksoft_dirty(newpmd);
1748 			if (pmd_swp_uffd_wp(*pmd))
1749 				newpmd = pmd_swp_mkuffd_wp(newpmd);
1750 			set_pmd_at(mm, addr, pmd, newpmd);
1751 		}
1752 		goto unlock;
1753 	}
1754 #endif
1755 
1756 	if (prot_numa) {
1757 		struct page *page;
1758 		/*
1759 		 * Avoid trapping faults against the zero page. The read-only
1760 		 * data is likely to be read-cached on the local CPU and
1761 		 * local/remote hits to the zero page are not interesting.
1762 		 */
1763 		if (is_huge_zero_pmd(*pmd))
1764 			goto unlock;
1765 
1766 		if (pmd_protnone(*pmd))
1767 			goto unlock;
1768 
1769 		page = pmd_page(*pmd);
1770 		/*
1771 		 * Skip scanning top tier node if normal numa
1772 		 * balancing is disabled
1773 		 */
1774 		if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
1775 		    node_is_toptier(page_to_nid(page)))
1776 			goto unlock;
1777 	}
1778 	/*
1779 	 * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1780 	 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1781 	 * which is also under mmap_read_lock(mm):
1782 	 *
1783 	 *	CPU0:				CPU1:
1784 	 *				change_huge_pmd(prot_numa=1)
1785 	 *				 pmdp_huge_get_and_clear_notify()
1786 	 * madvise_dontneed()
1787 	 *  zap_pmd_range()
1788 	 *   pmd_trans_huge(*pmd) == 0 (without ptl)
1789 	 *   // skip the pmd
1790 	 *				 set_pmd_at();
1791 	 *				 // pmd is re-established
1792 	 *
1793 	 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1794 	 * which may break userspace.
1795 	 *
1796 	 * pmdp_invalidate_ad() is required to make sure we don't miss
1797 	 * dirty/young flags set by hardware.
1798 	 */
1799 	oldpmd = pmdp_invalidate_ad(vma, addr, pmd);
1800 
1801 	entry = pmd_modify(oldpmd, newprot);
1802 	if (preserve_write)
1803 		entry = pmd_mk_savedwrite(entry);
1804 	if (uffd_wp) {
1805 		entry = pmd_wrprotect(entry);
1806 		entry = pmd_mkuffd_wp(entry);
1807 	} else if (uffd_wp_resolve) {
1808 		/*
1809 		 * Leave the write bit to be handled by PF interrupt
1810 		 * handler, then things like COW could be properly
1811 		 * handled.
1812 		 */
1813 		entry = pmd_clear_uffd_wp(entry);
1814 	}
1815 	ret = HPAGE_PMD_NR;
1816 	set_pmd_at(mm, addr, pmd, entry);
1817 
1818 	if (huge_pmd_needs_flush(oldpmd, entry))
1819 		tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE);
1820 
1821 	BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1822 unlock:
1823 	spin_unlock(ptl);
1824 	return ret;
1825 }
1826 
1827 /*
1828  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1829  *
1830  * Note that if it returns page table lock pointer, this routine returns without
1831  * unlocking page table lock. So callers must unlock it.
1832  */
1833 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1834 {
1835 	spinlock_t *ptl;
1836 	ptl = pmd_lock(vma->vm_mm, pmd);
1837 	if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1838 			pmd_devmap(*pmd)))
1839 		return ptl;
1840 	spin_unlock(ptl);
1841 	return NULL;
1842 }
1843 
1844 /*
1845  * Returns true if a given pud maps a thp, false otherwise.
1846  *
1847  * Note that if it returns true, this routine returns without unlocking page
1848  * table lock. So callers must unlock it.
1849  */
1850 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1851 {
1852 	spinlock_t *ptl;
1853 
1854 	ptl = pud_lock(vma->vm_mm, pud);
1855 	if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1856 		return ptl;
1857 	spin_unlock(ptl);
1858 	return NULL;
1859 }
1860 
1861 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1862 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1863 		 pud_t *pud, unsigned long addr)
1864 {
1865 	spinlock_t *ptl;
1866 
1867 	ptl = __pud_trans_huge_lock(pud, vma);
1868 	if (!ptl)
1869 		return 0;
1870 	/*
1871 	 * For architectures like ppc64 we look at deposited pgtable
1872 	 * when calling pudp_huge_get_and_clear. So do the
1873 	 * pgtable_trans_huge_withdraw after finishing pudp related
1874 	 * operations.
1875 	 */
1876 	pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1877 	tlb_remove_pud_tlb_entry(tlb, pud, addr);
1878 	if (vma_is_special_huge(vma)) {
1879 		spin_unlock(ptl);
1880 		/* No zero page support yet */
1881 	} else {
1882 		/* No support for anonymous PUD pages yet */
1883 		BUG();
1884 	}
1885 	return 1;
1886 }
1887 
1888 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1889 		unsigned long haddr)
1890 {
1891 	VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1892 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1893 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1894 	VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1895 
1896 	count_vm_event(THP_SPLIT_PUD);
1897 
1898 	pudp_huge_clear_flush_notify(vma, haddr, pud);
1899 }
1900 
1901 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1902 		unsigned long address)
1903 {
1904 	spinlock_t *ptl;
1905 	struct mmu_notifier_range range;
1906 
1907 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1908 				address & HPAGE_PUD_MASK,
1909 				(address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1910 	mmu_notifier_invalidate_range_start(&range);
1911 	ptl = pud_lock(vma->vm_mm, pud);
1912 	if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1913 		goto out;
1914 	__split_huge_pud_locked(vma, pud, range.start);
1915 
1916 out:
1917 	spin_unlock(ptl);
1918 	/*
1919 	 * No need to double call mmu_notifier->invalidate_range() callback as
1920 	 * the above pudp_huge_clear_flush_notify() did already call it.
1921 	 */
1922 	mmu_notifier_invalidate_range_only_end(&range);
1923 }
1924 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1925 
1926 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1927 		unsigned long haddr, pmd_t *pmd)
1928 {
1929 	struct mm_struct *mm = vma->vm_mm;
1930 	pgtable_t pgtable;
1931 	pmd_t _pmd;
1932 	int i;
1933 
1934 	/*
1935 	 * Leave pmd empty until pte is filled note that it is fine to delay
1936 	 * notification until mmu_notifier_invalidate_range_end() as we are
1937 	 * replacing a zero pmd write protected page with a zero pte write
1938 	 * protected page.
1939 	 *
1940 	 * See Documentation/vm/mmu_notifier.rst
1941 	 */
1942 	pmdp_huge_clear_flush(vma, haddr, pmd);
1943 
1944 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1945 	pmd_populate(mm, &_pmd, pgtable);
1946 
1947 	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1948 		pte_t *pte, entry;
1949 		entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1950 		entry = pte_mkspecial(entry);
1951 		pte = pte_offset_map(&_pmd, haddr);
1952 		VM_BUG_ON(!pte_none(*pte));
1953 		set_pte_at(mm, haddr, pte, entry);
1954 		pte_unmap(pte);
1955 	}
1956 	smp_wmb(); /* make pte visible before pmd */
1957 	pmd_populate(mm, pmd, pgtable);
1958 }
1959 
1960 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
1961 		unsigned long haddr, bool freeze)
1962 {
1963 	struct mm_struct *mm = vma->vm_mm;
1964 	struct page *page;
1965 	pgtable_t pgtable;
1966 	pmd_t old_pmd, _pmd;
1967 	bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
1968 	bool anon_exclusive = false;
1969 	unsigned long addr;
1970 	int i;
1971 
1972 	VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1973 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1974 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
1975 	VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
1976 				&& !pmd_devmap(*pmd));
1977 
1978 	count_vm_event(THP_SPLIT_PMD);
1979 
1980 	if (!vma_is_anonymous(vma)) {
1981 		old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1982 		/*
1983 		 * We are going to unmap this huge page. So
1984 		 * just go ahead and zap it
1985 		 */
1986 		if (arch_needs_pgtable_deposit())
1987 			zap_deposited_table(mm, pmd);
1988 		if (vma_is_special_huge(vma))
1989 			return;
1990 		if (unlikely(is_pmd_migration_entry(old_pmd))) {
1991 			swp_entry_t entry;
1992 
1993 			entry = pmd_to_swp_entry(old_pmd);
1994 			page = pfn_swap_entry_to_page(entry);
1995 		} else {
1996 			page = pmd_page(old_pmd);
1997 			if (!PageDirty(page) && pmd_dirty(old_pmd))
1998 				set_page_dirty(page);
1999 			if (!PageReferenced(page) && pmd_young(old_pmd))
2000 				SetPageReferenced(page);
2001 			page_remove_rmap(page, vma, true);
2002 			put_page(page);
2003 		}
2004 		add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2005 		return;
2006 	}
2007 
2008 	if (is_huge_zero_pmd(*pmd)) {
2009 		/*
2010 		 * FIXME: Do we want to invalidate secondary mmu by calling
2011 		 * mmu_notifier_invalidate_range() see comments below inside
2012 		 * __split_huge_pmd() ?
2013 		 *
2014 		 * We are going from a zero huge page write protected to zero
2015 		 * small page also write protected so it does not seems useful
2016 		 * to invalidate secondary mmu at this time.
2017 		 */
2018 		return __split_huge_zero_page_pmd(vma, haddr, pmd);
2019 	}
2020 
2021 	/*
2022 	 * Up to this point the pmd is present and huge and userland has the
2023 	 * whole access to the hugepage during the split (which happens in
2024 	 * place). If we overwrite the pmd with the not-huge version pointing
2025 	 * to the pte here (which of course we could if all CPUs were bug
2026 	 * free), userland could trigger a small page size TLB miss on the
2027 	 * small sized TLB while the hugepage TLB entry is still established in
2028 	 * the huge TLB. Some CPU doesn't like that.
2029 	 * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2030 	 * 383 on page 105. Intel should be safe but is also warns that it's
2031 	 * only safe if the permission and cache attributes of the two entries
2032 	 * loaded in the two TLB is identical (which should be the case here).
2033 	 * But it is generally safer to never allow small and huge TLB entries
2034 	 * for the same virtual address to be loaded simultaneously. So instead
2035 	 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2036 	 * current pmd notpresent (atomically because here the pmd_trans_huge
2037 	 * must remain set at all times on the pmd until the split is complete
2038 	 * for this pmd), then we flush the SMP TLB and finally we write the
2039 	 * non-huge version of the pmd entry with pmd_populate.
2040 	 */
2041 	old_pmd = pmdp_invalidate(vma, haddr, pmd);
2042 
2043 	pmd_migration = is_pmd_migration_entry(old_pmd);
2044 	if (unlikely(pmd_migration)) {
2045 		swp_entry_t entry;
2046 
2047 		entry = pmd_to_swp_entry(old_pmd);
2048 		page = pfn_swap_entry_to_page(entry);
2049 		write = is_writable_migration_entry(entry);
2050 		if (PageAnon(page))
2051 			anon_exclusive = is_readable_exclusive_migration_entry(entry);
2052 		young = false;
2053 		soft_dirty = pmd_swp_soft_dirty(old_pmd);
2054 		uffd_wp = pmd_swp_uffd_wp(old_pmd);
2055 	} else {
2056 		page = pmd_page(old_pmd);
2057 		if (pmd_dirty(old_pmd))
2058 			SetPageDirty(page);
2059 		write = pmd_write(old_pmd);
2060 		young = pmd_young(old_pmd);
2061 		soft_dirty = pmd_soft_dirty(old_pmd);
2062 		uffd_wp = pmd_uffd_wp(old_pmd);
2063 
2064 		VM_BUG_ON_PAGE(!page_count(page), page);
2065 		page_ref_add(page, HPAGE_PMD_NR - 1);
2066 
2067 		/*
2068 		 * Without "freeze", we'll simply split the PMD, propagating the
2069 		 * PageAnonExclusive() flag for each PTE by setting it for
2070 		 * each subpage -- no need to (temporarily) clear.
2071 		 *
2072 		 * With "freeze" we want to replace mapped pages by
2073 		 * migration entries right away. This is only possible if we
2074 		 * managed to clear PageAnonExclusive() -- see
2075 		 * set_pmd_migration_entry().
2076 		 *
2077 		 * In case we cannot clear PageAnonExclusive(), split the PMD
2078 		 * only and let try_to_migrate_one() fail later.
2079 		 */
2080 		anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
2081 		if (freeze && anon_exclusive && page_try_share_anon_rmap(page))
2082 			freeze = false;
2083 	}
2084 
2085 	/*
2086 	 * Withdraw the table only after we mark the pmd entry invalid.
2087 	 * This's critical for some architectures (Power).
2088 	 */
2089 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2090 	pmd_populate(mm, &_pmd, pgtable);
2091 
2092 	for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2093 		pte_t entry, *pte;
2094 		/*
2095 		 * Note that NUMA hinting access restrictions are not
2096 		 * transferred to avoid any possibility of altering
2097 		 * permissions across VMAs.
2098 		 */
2099 		if (freeze || pmd_migration) {
2100 			swp_entry_t swp_entry;
2101 			if (write)
2102 				swp_entry = make_writable_migration_entry(
2103 							page_to_pfn(page + i));
2104 			else if (anon_exclusive)
2105 				swp_entry = make_readable_exclusive_migration_entry(
2106 							page_to_pfn(page + i));
2107 			else
2108 				swp_entry = make_readable_migration_entry(
2109 							page_to_pfn(page + i));
2110 			entry = swp_entry_to_pte(swp_entry);
2111 			if (soft_dirty)
2112 				entry = pte_swp_mksoft_dirty(entry);
2113 			if (uffd_wp)
2114 				entry = pte_swp_mkuffd_wp(entry);
2115 		} else {
2116 			entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2117 			entry = maybe_mkwrite(entry, vma);
2118 			if (anon_exclusive)
2119 				SetPageAnonExclusive(page + i);
2120 			if (!write)
2121 				entry = pte_wrprotect(entry);
2122 			if (!young)
2123 				entry = pte_mkold(entry);
2124 			if (soft_dirty)
2125 				entry = pte_mksoft_dirty(entry);
2126 			if (uffd_wp)
2127 				entry = pte_mkuffd_wp(entry);
2128 		}
2129 		pte = pte_offset_map(&_pmd, addr);
2130 		BUG_ON(!pte_none(*pte));
2131 		set_pte_at(mm, addr, pte, entry);
2132 		if (!pmd_migration)
2133 			atomic_inc(&page[i]._mapcount);
2134 		pte_unmap(pte);
2135 	}
2136 
2137 	if (!pmd_migration) {
2138 		/*
2139 		 * Set PG_double_map before dropping compound_mapcount to avoid
2140 		 * false-negative page_mapped().
2141 		 */
2142 		if (compound_mapcount(page) > 1 &&
2143 		    !TestSetPageDoubleMap(page)) {
2144 			for (i = 0; i < HPAGE_PMD_NR; i++)
2145 				atomic_inc(&page[i]._mapcount);
2146 		}
2147 
2148 		lock_page_memcg(page);
2149 		if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2150 			/* Last compound_mapcount is gone. */
2151 			__mod_lruvec_page_state(page, NR_ANON_THPS,
2152 						-HPAGE_PMD_NR);
2153 			if (TestClearPageDoubleMap(page)) {
2154 				/* No need in mapcount reference anymore */
2155 				for (i = 0; i < HPAGE_PMD_NR; i++)
2156 					atomic_dec(&page[i]._mapcount);
2157 			}
2158 		}
2159 		unlock_page_memcg(page);
2160 
2161 		/* Above is effectively page_remove_rmap(page, vma, true) */
2162 		munlock_vma_page(page, vma, true);
2163 	}
2164 
2165 	smp_wmb(); /* make pte visible before pmd */
2166 	pmd_populate(mm, pmd, pgtable);
2167 
2168 	if (freeze) {
2169 		for (i = 0; i < HPAGE_PMD_NR; i++) {
2170 			page_remove_rmap(page + i, vma, false);
2171 			put_page(page + i);
2172 		}
2173 	}
2174 }
2175 
2176 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2177 		unsigned long address, bool freeze, struct folio *folio)
2178 {
2179 	spinlock_t *ptl;
2180 	struct mmu_notifier_range range;
2181 
2182 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2183 				address & HPAGE_PMD_MASK,
2184 				(address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2185 	mmu_notifier_invalidate_range_start(&range);
2186 	ptl = pmd_lock(vma->vm_mm, pmd);
2187 
2188 	/*
2189 	 * If caller asks to setup a migration entry, we need a folio to check
2190 	 * pmd against. Otherwise we can end up replacing wrong folio.
2191 	 */
2192 	VM_BUG_ON(freeze && !folio);
2193 	VM_WARN_ON_ONCE(folio && !folio_test_locked(folio));
2194 
2195 	if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) ||
2196 	    is_pmd_migration_entry(*pmd)) {
2197 		if (folio && folio != page_folio(pmd_page(*pmd)))
2198 			goto out;
2199 		__split_huge_pmd_locked(vma, pmd, range.start, freeze);
2200 	}
2201 
2202 out:
2203 	spin_unlock(ptl);
2204 	/*
2205 	 * No need to double call mmu_notifier->invalidate_range() callback.
2206 	 * They are 3 cases to consider inside __split_huge_pmd_locked():
2207 	 *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2208 	 *  2) __split_huge_zero_page_pmd() read only zero page and any write
2209 	 *    fault will trigger a flush_notify before pointing to a new page
2210 	 *    (it is fine if the secondary mmu keeps pointing to the old zero
2211 	 *    page in the meantime)
2212 	 *  3) Split a huge pmd into pte pointing to the same page. No need
2213 	 *     to invalidate secondary tlb entry they are all still valid.
2214 	 *     any further changes to individual pte will notify. So no need
2215 	 *     to call mmu_notifier->invalidate_range()
2216 	 */
2217 	mmu_notifier_invalidate_range_only_end(&range);
2218 }
2219 
2220 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2221 		bool freeze, struct folio *folio)
2222 {
2223 	pgd_t *pgd;
2224 	p4d_t *p4d;
2225 	pud_t *pud;
2226 	pmd_t *pmd;
2227 
2228 	pgd = pgd_offset(vma->vm_mm, address);
2229 	if (!pgd_present(*pgd))
2230 		return;
2231 
2232 	p4d = p4d_offset(pgd, address);
2233 	if (!p4d_present(*p4d))
2234 		return;
2235 
2236 	pud = pud_offset(p4d, address);
2237 	if (!pud_present(*pud))
2238 		return;
2239 
2240 	pmd = pmd_offset(pud, address);
2241 
2242 	__split_huge_pmd(vma, pmd, address, freeze, folio);
2243 }
2244 
2245 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2246 {
2247 	/*
2248 	 * If the new address isn't hpage aligned and it could previously
2249 	 * contain an hugepage: check if we need to split an huge pmd.
2250 	 */
2251 	if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2252 	    range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2253 			 ALIGN(address, HPAGE_PMD_SIZE)))
2254 		split_huge_pmd_address(vma, address, false, NULL);
2255 }
2256 
2257 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2258 			     unsigned long start,
2259 			     unsigned long end,
2260 			     long adjust_next)
2261 {
2262 	/* Check if we need to split start first. */
2263 	split_huge_pmd_if_needed(vma, start);
2264 
2265 	/* Check if we need to split end next. */
2266 	split_huge_pmd_if_needed(vma, end);
2267 
2268 	/*
2269 	 * If we're also updating the vma->vm_next->vm_start,
2270 	 * check if we need to split it.
2271 	 */
2272 	if (adjust_next > 0) {
2273 		struct vm_area_struct *next = vma->vm_next;
2274 		unsigned long nstart = next->vm_start;
2275 		nstart += adjust_next;
2276 		split_huge_pmd_if_needed(next, nstart);
2277 	}
2278 }
2279 
2280 static void unmap_page(struct page *page)
2281 {
2282 	struct folio *folio = page_folio(page);
2283 	enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
2284 		TTU_SYNC;
2285 
2286 	VM_BUG_ON_PAGE(!PageHead(page), page);
2287 
2288 	/*
2289 	 * Anon pages need migration entries to preserve them, but file
2290 	 * pages can simply be left unmapped, then faulted back on demand.
2291 	 * If that is ever changed (perhaps for mlock), update remap_page().
2292 	 */
2293 	if (folio_test_anon(folio))
2294 		try_to_migrate(folio, ttu_flags);
2295 	else
2296 		try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK);
2297 }
2298 
2299 static void remap_page(struct folio *folio, unsigned long nr)
2300 {
2301 	int i = 0;
2302 
2303 	/* If unmap_page() uses try_to_migrate() on file, remove this check */
2304 	if (!folio_test_anon(folio))
2305 		return;
2306 	for (;;) {
2307 		remove_migration_ptes(folio, folio, true);
2308 		i += folio_nr_pages(folio);
2309 		if (i >= nr)
2310 			break;
2311 		folio = folio_next(folio);
2312 	}
2313 }
2314 
2315 static void lru_add_page_tail(struct page *head, struct page *tail,
2316 		struct lruvec *lruvec, struct list_head *list)
2317 {
2318 	VM_BUG_ON_PAGE(!PageHead(head), head);
2319 	VM_BUG_ON_PAGE(PageCompound(tail), head);
2320 	VM_BUG_ON_PAGE(PageLRU(tail), head);
2321 	lockdep_assert_held(&lruvec->lru_lock);
2322 
2323 	if (list) {
2324 		/* page reclaim is reclaiming a huge page */
2325 		VM_WARN_ON(PageLRU(head));
2326 		get_page(tail);
2327 		list_add_tail(&tail->lru, list);
2328 	} else {
2329 		/* head is still on lru (and we have it frozen) */
2330 		VM_WARN_ON(!PageLRU(head));
2331 		if (PageUnevictable(tail))
2332 			tail->mlock_count = 0;
2333 		else
2334 			list_add_tail(&tail->lru, &head->lru);
2335 		SetPageLRU(tail);
2336 	}
2337 }
2338 
2339 static void __split_huge_page_tail(struct page *head, int tail,
2340 		struct lruvec *lruvec, struct list_head *list)
2341 {
2342 	struct page *page_tail = head + tail;
2343 
2344 	VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2345 
2346 	/*
2347 	 * Clone page flags before unfreezing refcount.
2348 	 *
2349 	 * After successful get_page_unless_zero() might follow flags change,
2350 	 * for example lock_page() which set PG_waiters.
2351 	 *
2352 	 * Note that for mapped sub-pages of an anonymous THP,
2353 	 * PG_anon_exclusive has been cleared in unmap_page() and is stored in
2354 	 * the migration entry instead from where remap_page() will restore it.
2355 	 * We can still have PG_anon_exclusive set on effectively unmapped and
2356 	 * unreferenced sub-pages of an anonymous THP: we can simply drop
2357 	 * PG_anon_exclusive (-> PG_mappedtodisk) for these here.
2358 	 */
2359 	page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2360 	page_tail->flags |= (head->flags &
2361 			((1L << PG_referenced) |
2362 			 (1L << PG_swapbacked) |
2363 			 (1L << PG_swapcache) |
2364 			 (1L << PG_mlocked) |
2365 			 (1L << PG_uptodate) |
2366 			 (1L << PG_active) |
2367 			 (1L << PG_workingset) |
2368 			 (1L << PG_locked) |
2369 			 (1L << PG_unevictable) |
2370 #ifdef CONFIG_64BIT
2371 			 (1L << PG_arch_2) |
2372 #endif
2373 			 (1L << PG_dirty)));
2374 
2375 	/* ->mapping in first tail page is compound_mapcount */
2376 	VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2377 			page_tail);
2378 	page_tail->mapping = head->mapping;
2379 	page_tail->index = head->index + tail;
2380 
2381 	/* Page flags must be visible before we make the page non-compound. */
2382 	smp_wmb();
2383 
2384 	/*
2385 	 * Clear PageTail before unfreezing page refcount.
2386 	 *
2387 	 * After successful get_page_unless_zero() might follow put_page()
2388 	 * which needs correct compound_head().
2389 	 */
2390 	clear_compound_head(page_tail);
2391 
2392 	/* Finally unfreeze refcount. Additional reference from page cache. */
2393 	page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2394 					  PageSwapCache(head)));
2395 
2396 	if (page_is_young(head))
2397 		set_page_young(page_tail);
2398 	if (page_is_idle(head))
2399 		set_page_idle(page_tail);
2400 
2401 	page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2402 
2403 	/*
2404 	 * always add to the tail because some iterators expect new
2405 	 * pages to show after the currently processed elements - e.g.
2406 	 * migrate_pages
2407 	 */
2408 	lru_add_page_tail(head, page_tail, lruvec, list);
2409 }
2410 
2411 static void __split_huge_page(struct page *page, struct list_head *list,
2412 		pgoff_t end)
2413 {
2414 	struct folio *folio = page_folio(page);
2415 	struct page *head = &folio->page;
2416 	struct lruvec *lruvec;
2417 	struct address_space *swap_cache = NULL;
2418 	unsigned long offset = 0;
2419 	unsigned int nr = thp_nr_pages(head);
2420 	int i;
2421 
2422 	/* complete memcg works before add pages to LRU */
2423 	split_page_memcg(head, nr);
2424 
2425 	if (PageAnon(head) && PageSwapCache(head)) {
2426 		swp_entry_t entry = { .val = page_private(head) };
2427 
2428 		offset = swp_offset(entry);
2429 		swap_cache = swap_address_space(entry);
2430 		xa_lock(&swap_cache->i_pages);
2431 	}
2432 
2433 	/* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2434 	lruvec = folio_lruvec_lock(folio);
2435 
2436 	ClearPageHasHWPoisoned(head);
2437 
2438 	for (i = nr - 1; i >= 1; i--) {
2439 		__split_huge_page_tail(head, i, lruvec, list);
2440 		/* Some pages can be beyond EOF: drop them from page cache */
2441 		if (head[i].index >= end) {
2442 			ClearPageDirty(head + i);
2443 			__delete_from_page_cache(head + i, NULL);
2444 			if (shmem_mapping(head->mapping))
2445 				shmem_uncharge(head->mapping->host, 1);
2446 			put_page(head + i);
2447 		} else if (!PageAnon(page)) {
2448 			__xa_store(&head->mapping->i_pages, head[i].index,
2449 					head + i, 0);
2450 		} else if (swap_cache) {
2451 			__xa_store(&swap_cache->i_pages, offset + i,
2452 					head + i, 0);
2453 		}
2454 	}
2455 
2456 	ClearPageCompound(head);
2457 	unlock_page_lruvec(lruvec);
2458 	/* Caller disabled irqs, so they are still disabled here */
2459 
2460 	split_page_owner(head, nr);
2461 
2462 	/* See comment in __split_huge_page_tail() */
2463 	if (PageAnon(head)) {
2464 		/* Additional pin to swap cache */
2465 		if (PageSwapCache(head)) {
2466 			page_ref_add(head, 2);
2467 			xa_unlock(&swap_cache->i_pages);
2468 		} else {
2469 			page_ref_inc(head);
2470 		}
2471 	} else {
2472 		/* Additional pin to page cache */
2473 		page_ref_add(head, 2);
2474 		xa_unlock(&head->mapping->i_pages);
2475 	}
2476 	local_irq_enable();
2477 
2478 	remap_page(folio, nr);
2479 
2480 	if (PageSwapCache(head)) {
2481 		swp_entry_t entry = { .val = page_private(head) };
2482 
2483 		split_swap_cluster(entry);
2484 	}
2485 
2486 	for (i = 0; i < nr; i++) {
2487 		struct page *subpage = head + i;
2488 		if (subpage == page)
2489 			continue;
2490 		unlock_page(subpage);
2491 
2492 		/*
2493 		 * Subpages may be freed if there wasn't any mapping
2494 		 * like if add_to_swap() is running on a lru page that
2495 		 * had its mapping zapped. And freeing these pages
2496 		 * requires taking the lru_lock so we do the put_page
2497 		 * of the tail pages after the split is complete.
2498 		 */
2499 		put_page(subpage);
2500 	}
2501 }
2502 
2503 /* Racy check whether the huge page can be split */
2504 bool can_split_folio(struct folio *folio, int *pextra_pins)
2505 {
2506 	int extra_pins;
2507 
2508 	/* Additional pins from page cache */
2509 	if (folio_test_anon(folio))
2510 		extra_pins = folio_test_swapcache(folio) ?
2511 				folio_nr_pages(folio) : 0;
2512 	else
2513 		extra_pins = folio_nr_pages(folio);
2514 	if (pextra_pins)
2515 		*pextra_pins = extra_pins;
2516 	return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins - 1;
2517 }
2518 
2519 /*
2520  * This function splits huge page into normal pages. @page can point to any
2521  * subpage of huge page to split. Split doesn't change the position of @page.
2522  *
2523  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2524  * The huge page must be locked.
2525  *
2526  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2527  *
2528  * Both head page and tail pages will inherit mapping, flags, and so on from
2529  * the hugepage.
2530  *
2531  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2532  * they are not mapped.
2533  *
2534  * Returns 0 if the hugepage is split successfully.
2535  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2536  * us.
2537  */
2538 int split_huge_page_to_list(struct page *page, struct list_head *list)
2539 {
2540 	struct folio *folio = page_folio(page);
2541 	struct page *head = &folio->page;
2542 	struct deferred_split *ds_queue = get_deferred_split_queue(head);
2543 	XA_STATE(xas, &head->mapping->i_pages, head->index);
2544 	struct anon_vma *anon_vma = NULL;
2545 	struct address_space *mapping = NULL;
2546 	int extra_pins, ret;
2547 	pgoff_t end;
2548 	bool is_hzp;
2549 
2550 	VM_BUG_ON_PAGE(!PageLocked(head), head);
2551 	VM_BUG_ON_PAGE(!PageCompound(head), head);
2552 
2553 	is_hzp = is_huge_zero_page(head);
2554 	VM_WARN_ON_ONCE_PAGE(is_hzp, head);
2555 	if (is_hzp)
2556 		return -EBUSY;
2557 
2558 	if (PageWriteback(head))
2559 		return -EBUSY;
2560 
2561 	if (PageAnon(head)) {
2562 		/*
2563 		 * The caller does not necessarily hold an mmap_lock that would
2564 		 * prevent the anon_vma disappearing so we first we take a
2565 		 * reference to it and then lock the anon_vma for write. This
2566 		 * is similar to folio_lock_anon_vma_read except the write lock
2567 		 * is taken to serialise against parallel split or collapse
2568 		 * operations.
2569 		 */
2570 		anon_vma = page_get_anon_vma(head);
2571 		if (!anon_vma) {
2572 			ret = -EBUSY;
2573 			goto out;
2574 		}
2575 		end = -1;
2576 		mapping = NULL;
2577 		anon_vma_lock_write(anon_vma);
2578 	} else {
2579 		mapping = head->mapping;
2580 
2581 		/* Truncated ? */
2582 		if (!mapping) {
2583 			ret = -EBUSY;
2584 			goto out;
2585 		}
2586 
2587 		xas_split_alloc(&xas, head, compound_order(head),
2588 				mapping_gfp_mask(mapping) & GFP_RECLAIM_MASK);
2589 		if (xas_error(&xas)) {
2590 			ret = xas_error(&xas);
2591 			goto out;
2592 		}
2593 
2594 		anon_vma = NULL;
2595 		i_mmap_lock_read(mapping);
2596 
2597 		/*
2598 		 *__split_huge_page() may need to trim off pages beyond EOF:
2599 		 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2600 		 * which cannot be nested inside the page tree lock. So note
2601 		 * end now: i_size itself may be changed at any moment, but
2602 		 * head page lock is good enough to serialize the trimming.
2603 		 */
2604 		end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2605 		if (shmem_mapping(mapping))
2606 			end = shmem_fallocend(mapping->host, end);
2607 	}
2608 
2609 	/*
2610 	 * Racy check if we can split the page, before unmap_page() will
2611 	 * split PMDs
2612 	 */
2613 	if (!can_split_folio(folio, &extra_pins)) {
2614 		ret = -EBUSY;
2615 		goto out_unlock;
2616 	}
2617 
2618 	unmap_page(head);
2619 
2620 	/* block interrupt reentry in xa_lock and spinlock */
2621 	local_irq_disable();
2622 	if (mapping) {
2623 		/*
2624 		 * Check if the head page is present in page cache.
2625 		 * We assume all tail are present too, if head is there.
2626 		 */
2627 		xas_lock(&xas);
2628 		xas_reset(&xas);
2629 		if (xas_load(&xas) != head)
2630 			goto fail;
2631 	}
2632 
2633 	/* Prevent deferred_split_scan() touching ->_refcount */
2634 	spin_lock(&ds_queue->split_queue_lock);
2635 	if (page_ref_freeze(head, 1 + extra_pins)) {
2636 		if (!list_empty(page_deferred_list(head))) {
2637 			ds_queue->split_queue_len--;
2638 			list_del(page_deferred_list(head));
2639 		}
2640 		spin_unlock(&ds_queue->split_queue_lock);
2641 		if (mapping) {
2642 			int nr = thp_nr_pages(head);
2643 
2644 			xas_split(&xas, head, thp_order(head));
2645 			if (PageSwapBacked(head)) {
2646 				__mod_lruvec_page_state(head, NR_SHMEM_THPS,
2647 							-nr);
2648 			} else {
2649 				__mod_lruvec_page_state(head, NR_FILE_THPS,
2650 							-nr);
2651 				filemap_nr_thps_dec(mapping);
2652 			}
2653 		}
2654 
2655 		__split_huge_page(page, list, end);
2656 		ret = 0;
2657 	} else {
2658 		spin_unlock(&ds_queue->split_queue_lock);
2659 fail:
2660 		if (mapping)
2661 			xas_unlock(&xas);
2662 		local_irq_enable();
2663 		remap_page(folio, folio_nr_pages(folio));
2664 		ret = -EBUSY;
2665 	}
2666 
2667 out_unlock:
2668 	if (anon_vma) {
2669 		anon_vma_unlock_write(anon_vma);
2670 		put_anon_vma(anon_vma);
2671 	}
2672 	if (mapping)
2673 		i_mmap_unlock_read(mapping);
2674 out:
2675 	/* Free any memory we didn't use */
2676 	xas_nomem(&xas, 0);
2677 	count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2678 	return ret;
2679 }
2680 
2681 void free_transhuge_page(struct page *page)
2682 {
2683 	struct deferred_split *ds_queue = get_deferred_split_queue(page);
2684 	unsigned long flags;
2685 
2686 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2687 	if (!list_empty(page_deferred_list(page))) {
2688 		ds_queue->split_queue_len--;
2689 		list_del(page_deferred_list(page));
2690 	}
2691 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2692 	free_compound_page(page);
2693 }
2694 
2695 void deferred_split_huge_page(struct page *page)
2696 {
2697 	struct deferred_split *ds_queue = get_deferred_split_queue(page);
2698 #ifdef CONFIG_MEMCG
2699 	struct mem_cgroup *memcg = page_memcg(compound_head(page));
2700 #endif
2701 	unsigned long flags;
2702 
2703 	VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2704 
2705 	/*
2706 	 * The try_to_unmap() in page reclaim path might reach here too,
2707 	 * this may cause a race condition to corrupt deferred split queue.
2708 	 * And, if page reclaim is already handling the same page, it is
2709 	 * unnecessary to handle it again in shrinker.
2710 	 *
2711 	 * Check PageSwapCache to determine if the page is being
2712 	 * handled by page reclaim since THP swap would add the page into
2713 	 * swap cache before calling try_to_unmap().
2714 	 */
2715 	if (PageSwapCache(page))
2716 		return;
2717 
2718 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2719 	if (list_empty(page_deferred_list(page))) {
2720 		count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2721 		list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2722 		ds_queue->split_queue_len++;
2723 #ifdef CONFIG_MEMCG
2724 		if (memcg)
2725 			set_shrinker_bit(memcg, page_to_nid(page),
2726 					 deferred_split_shrinker.id);
2727 #endif
2728 	}
2729 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2730 }
2731 
2732 static unsigned long deferred_split_count(struct shrinker *shrink,
2733 		struct shrink_control *sc)
2734 {
2735 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2736 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2737 
2738 #ifdef CONFIG_MEMCG
2739 	if (sc->memcg)
2740 		ds_queue = &sc->memcg->deferred_split_queue;
2741 #endif
2742 	return READ_ONCE(ds_queue->split_queue_len);
2743 }
2744 
2745 static unsigned long deferred_split_scan(struct shrinker *shrink,
2746 		struct shrink_control *sc)
2747 {
2748 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2749 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2750 	unsigned long flags;
2751 	LIST_HEAD(list), *pos, *next;
2752 	struct page *page;
2753 	int split = 0;
2754 
2755 #ifdef CONFIG_MEMCG
2756 	if (sc->memcg)
2757 		ds_queue = &sc->memcg->deferred_split_queue;
2758 #endif
2759 
2760 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2761 	/* Take pin on all head pages to avoid freeing them under us */
2762 	list_for_each_safe(pos, next, &ds_queue->split_queue) {
2763 		page = list_entry((void *)pos, struct page, deferred_list);
2764 		page = compound_head(page);
2765 		if (get_page_unless_zero(page)) {
2766 			list_move(page_deferred_list(page), &list);
2767 		} else {
2768 			/* We lost race with put_compound_page() */
2769 			list_del_init(page_deferred_list(page));
2770 			ds_queue->split_queue_len--;
2771 		}
2772 		if (!--sc->nr_to_scan)
2773 			break;
2774 	}
2775 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2776 
2777 	list_for_each_safe(pos, next, &list) {
2778 		page = list_entry((void *)pos, struct page, deferred_list);
2779 		if (!trylock_page(page))
2780 			goto next;
2781 		/* split_huge_page() removes page from list on success */
2782 		if (!split_huge_page(page))
2783 			split++;
2784 		unlock_page(page);
2785 next:
2786 		put_page(page);
2787 	}
2788 
2789 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2790 	list_splice_tail(&list, &ds_queue->split_queue);
2791 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2792 
2793 	/*
2794 	 * Stop shrinker if we didn't split any page, but the queue is empty.
2795 	 * This can happen if pages were freed under us.
2796 	 */
2797 	if (!split && list_empty(&ds_queue->split_queue))
2798 		return SHRINK_STOP;
2799 	return split;
2800 }
2801 
2802 static struct shrinker deferred_split_shrinker = {
2803 	.count_objects = deferred_split_count,
2804 	.scan_objects = deferred_split_scan,
2805 	.seeks = DEFAULT_SEEKS,
2806 	.flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2807 		 SHRINKER_NONSLAB,
2808 };
2809 
2810 #ifdef CONFIG_DEBUG_FS
2811 static void split_huge_pages_all(void)
2812 {
2813 	struct zone *zone;
2814 	struct page *page;
2815 	unsigned long pfn, max_zone_pfn;
2816 	unsigned long total = 0, split = 0;
2817 
2818 	pr_debug("Split all THPs\n");
2819 	for_each_populated_zone(zone) {
2820 		max_zone_pfn = zone_end_pfn(zone);
2821 		for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2822 			if (!pfn_valid(pfn))
2823 				continue;
2824 
2825 			page = pfn_to_page(pfn);
2826 			if (!get_page_unless_zero(page))
2827 				continue;
2828 
2829 			if (zone != page_zone(page))
2830 				goto next;
2831 
2832 			if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2833 				goto next;
2834 
2835 			total++;
2836 			lock_page(page);
2837 			if (!split_huge_page(page))
2838 				split++;
2839 			unlock_page(page);
2840 next:
2841 			put_page(page);
2842 			cond_resched();
2843 		}
2844 	}
2845 
2846 	pr_debug("%lu of %lu THP split\n", split, total);
2847 }
2848 
2849 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
2850 {
2851 	return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
2852 		    is_vm_hugetlb_page(vma);
2853 }
2854 
2855 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
2856 				unsigned long vaddr_end)
2857 {
2858 	int ret = 0;
2859 	struct task_struct *task;
2860 	struct mm_struct *mm;
2861 	unsigned long total = 0, split = 0;
2862 	unsigned long addr;
2863 
2864 	vaddr_start &= PAGE_MASK;
2865 	vaddr_end &= PAGE_MASK;
2866 
2867 	/* Find the task_struct from pid */
2868 	rcu_read_lock();
2869 	task = find_task_by_vpid(pid);
2870 	if (!task) {
2871 		rcu_read_unlock();
2872 		ret = -ESRCH;
2873 		goto out;
2874 	}
2875 	get_task_struct(task);
2876 	rcu_read_unlock();
2877 
2878 	/* Find the mm_struct */
2879 	mm = get_task_mm(task);
2880 	put_task_struct(task);
2881 
2882 	if (!mm) {
2883 		ret = -EINVAL;
2884 		goto out;
2885 	}
2886 
2887 	pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
2888 		 pid, vaddr_start, vaddr_end);
2889 
2890 	mmap_read_lock(mm);
2891 	/*
2892 	 * always increase addr by PAGE_SIZE, since we could have a PTE page
2893 	 * table filled with PTE-mapped THPs, each of which is distinct.
2894 	 */
2895 	for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
2896 		struct vm_area_struct *vma = find_vma(mm, addr);
2897 		struct page *page;
2898 
2899 		if (!vma || addr < vma->vm_start)
2900 			break;
2901 
2902 		/* skip special VMA and hugetlb VMA */
2903 		if (vma_not_suitable_for_thp_split(vma)) {
2904 			addr = vma->vm_end;
2905 			continue;
2906 		}
2907 
2908 		/* FOLL_DUMP to ignore special (like zero) pages */
2909 		page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
2910 
2911 		if (IS_ERR(page))
2912 			continue;
2913 		if (!page)
2914 			continue;
2915 
2916 		if (!is_transparent_hugepage(page))
2917 			goto next;
2918 
2919 		total++;
2920 		if (!can_split_folio(page_folio(page), NULL))
2921 			goto next;
2922 
2923 		if (!trylock_page(page))
2924 			goto next;
2925 
2926 		if (!split_huge_page(page))
2927 			split++;
2928 
2929 		unlock_page(page);
2930 next:
2931 		put_page(page);
2932 		cond_resched();
2933 	}
2934 	mmap_read_unlock(mm);
2935 	mmput(mm);
2936 
2937 	pr_debug("%lu of %lu THP split\n", split, total);
2938 
2939 out:
2940 	return ret;
2941 }
2942 
2943 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
2944 				pgoff_t off_end)
2945 {
2946 	struct filename *file;
2947 	struct file *candidate;
2948 	struct address_space *mapping;
2949 	int ret = -EINVAL;
2950 	pgoff_t index;
2951 	int nr_pages = 1;
2952 	unsigned long total = 0, split = 0;
2953 
2954 	file = getname_kernel(file_path);
2955 	if (IS_ERR(file))
2956 		return ret;
2957 
2958 	candidate = file_open_name(file, O_RDONLY, 0);
2959 	if (IS_ERR(candidate))
2960 		goto out;
2961 
2962 	pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
2963 		 file_path, off_start, off_end);
2964 
2965 	mapping = candidate->f_mapping;
2966 
2967 	for (index = off_start; index < off_end; index += nr_pages) {
2968 		struct page *fpage = pagecache_get_page(mapping, index,
2969 						FGP_ENTRY | FGP_HEAD, 0);
2970 
2971 		nr_pages = 1;
2972 		if (xa_is_value(fpage) || !fpage)
2973 			continue;
2974 
2975 		if (!is_transparent_hugepage(fpage))
2976 			goto next;
2977 
2978 		total++;
2979 		nr_pages = thp_nr_pages(fpage);
2980 
2981 		if (!trylock_page(fpage))
2982 			goto next;
2983 
2984 		if (!split_huge_page(fpage))
2985 			split++;
2986 
2987 		unlock_page(fpage);
2988 next:
2989 		put_page(fpage);
2990 		cond_resched();
2991 	}
2992 
2993 	filp_close(candidate, NULL);
2994 	ret = 0;
2995 
2996 	pr_debug("%lu of %lu file-backed THP split\n", split, total);
2997 out:
2998 	putname(file);
2999 	return ret;
3000 }
3001 
3002 #define MAX_INPUT_BUF_SZ 255
3003 
3004 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3005 				size_t count, loff_t *ppops)
3006 {
3007 	static DEFINE_MUTEX(split_debug_mutex);
3008 	ssize_t ret;
3009 	/* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3010 	char input_buf[MAX_INPUT_BUF_SZ];
3011 	int pid;
3012 	unsigned long vaddr_start, vaddr_end;
3013 
3014 	ret = mutex_lock_interruptible(&split_debug_mutex);
3015 	if (ret)
3016 		return ret;
3017 
3018 	ret = -EFAULT;
3019 
3020 	memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3021 	if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3022 		goto out;
3023 
3024 	input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3025 
3026 	if (input_buf[0] == '/') {
3027 		char *tok;
3028 		char *buf = input_buf;
3029 		char file_path[MAX_INPUT_BUF_SZ];
3030 		pgoff_t off_start = 0, off_end = 0;
3031 		size_t input_len = strlen(input_buf);
3032 
3033 		tok = strsep(&buf, ",");
3034 		if (tok) {
3035 			strcpy(file_path, tok);
3036 		} else {
3037 			ret = -EINVAL;
3038 			goto out;
3039 		}
3040 
3041 		ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3042 		if (ret != 2) {
3043 			ret = -EINVAL;
3044 			goto out;
3045 		}
3046 		ret = split_huge_pages_in_file(file_path, off_start, off_end);
3047 		if (!ret)
3048 			ret = input_len;
3049 
3050 		goto out;
3051 	}
3052 
3053 	ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3054 	if (ret == 1 && pid == 1) {
3055 		split_huge_pages_all();
3056 		ret = strlen(input_buf);
3057 		goto out;
3058 	} else if (ret != 3) {
3059 		ret = -EINVAL;
3060 		goto out;
3061 	}
3062 
3063 	ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3064 	if (!ret)
3065 		ret = strlen(input_buf);
3066 out:
3067 	mutex_unlock(&split_debug_mutex);
3068 	return ret;
3069 
3070 }
3071 
3072 static const struct file_operations split_huge_pages_fops = {
3073 	.owner	 = THIS_MODULE,
3074 	.write	 = split_huge_pages_write,
3075 	.llseek  = no_llseek,
3076 };
3077 
3078 static int __init split_huge_pages_debugfs(void)
3079 {
3080 	debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3081 			    &split_huge_pages_fops);
3082 	return 0;
3083 }
3084 late_initcall(split_huge_pages_debugfs);
3085 #endif
3086 
3087 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3088 int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3089 		struct page *page)
3090 {
3091 	struct vm_area_struct *vma = pvmw->vma;
3092 	struct mm_struct *mm = vma->vm_mm;
3093 	unsigned long address = pvmw->address;
3094 	bool anon_exclusive;
3095 	pmd_t pmdval;
3096 	swp_entry_t entry;
3097 	pmd_t pmdswp;
3098 
3099 	if (!(pvmw->pmd && !pvmw->pte))
3100 		return 0;
3101 
3102 	flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3103 	pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3104 
3105 	anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
3106 	if (anon_exclusive && page_try_share_anon_rmap(page)) {
3107 		set_pmd_at(mm, address, pvmw->pmd, pmdval);
3108 		return -EBUSY;
3109 	}
3110 
3111 	if (pmd_dirty(pmdval))
3112 		set_page_dirty(page);
3113 	if (pmd_write(pmdval))
3114 		entry = make_writable_migration_entry(page_to_pfn(page));
3115 	else if (anon_exclusive)
3116 		entry = make_readable_exclusive_migration_entry(page_to_pfn(page));
3117 	else
3118 		entry = make_readable_migration_entry(page_to_pfn(page));
3119 	pmdswp = swp_entry_to_pmd(entry);
3120 	if (pmd_soft_dirty(pmdval))
3121 		pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3122 	set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3123 	page_remove_rmap(page, vma, true);
3124 	put_page(page);
3125 	trace_set_migration_pmd(address, pmd_val(pmdswp));
3126 
3127 	return 0;
3128 }
3129 
3130 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3131 {
3132 	struct vm_area_struct *vma = pvmw->vma;
3133 	struct mm_struct *mm = vma->vm_mm;
3134 	unsigned long address = pvmw->address;
3135 	unsigned long mmun_start = address & HPAGE_PMD_MASK;
3136 	pmd_t pmde;
3137 	swp_entry_t entry;
3138 
3139 	if (!(pvmw->pmd && !pvmw->pte))
3140 		return;
3141 
3142 	entry = pmd_to_swp_entry(*pvmw->pmd);
3143 	get_page(new);
3144 	pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3145 	if (pmd_swp_soft_dirty(*pvmw->pmd))
3146 		pmde = pmd_mksoft_dirty(pmde);
3147 	if (is_writable_migration_entry(entry))
3148 		pmde = maybe_pmd_mkwrite(pmde, vma);
3149 	if (pmd_swp_uffd_wp(*pvmw->pmd))
3150 		pmde = pmd_wrprotect(pmd_mkuffd_wp(pmde));
3151 
3152 	if (PageAnon(new)) {
3153 		rmap_t rmap_flags = RMAP_COMPOUND;
3154 
3155 		if (!is_readable_migration_entry(entry))
3156 			rmap_flags |= RMAP_EXCLUSIVE;
3157 
3158 		page_add_anon_rmap(new, vma, mmun_start, rmap_flags);
3159 	} else {
3160 		page_add_file_rmap(new, vma, true);
3161 	}
3162 	VM_BUG_ON(pmd_write(pmde) && PageAnon(new) && !PageAnonExclusive(new));
3163 	set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3164 
3165 	/* No need to invalidate - it was non-present before */
3166 	update_mmu_cache_pmd(vma, address, pvmw->pmd);
3167 	trace_remove_migration_pmd(address, pmd_val(pmde));
3168 }
3169 #endif
3170