xref: /openbmc/linux/mm/huge_memory.c (revision 3e30a927)
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/coredump.h>
11 #include <linux/sched/numa_balancing.h>
12 #include <linux/highmem.h>
13 #include <linux/hugetlb.h>
14 #include <linux/mmu_notifier.h>
15 #include <linux/rmap.h>
16 #include <linux/swap.h>
17 #include <linux/shrinker.h>
18 #include <linux/mm_inline.h>
19 #include <linux/swapops.h>
20 #include <linux/dax.h>
21 #include <linux/khugepaged.h>
22 #include <linux/freezer.h>
23 #include <linux/pfn_t.h>
24 #include <linux/mman.h>
25 #include <linux/memremap.h>
26 #include <linux/pagemap.h>
27 #include <linux/debugfs.h>
28 #include <linux/migrate.h>
29 #include <linux/hashtable.h>
30 #include <linux/userfaultfd_k.h>
31 #include <linux/page_idle.h>
32 #include <linux/shmem_fs.h>
33 #include <linux/oom.h>
34 #include <linux/numa.h>
35 #include <linux/page_owner.h>
36 
37 #include <asm/tlb.h>
38 #include <asm/pgalloc.h>
39 #include "internal.h"
40 
41 /*
42  * By default, transparent hugepage support is disabled in order to avoid
43  * risking an increased memory footprint for applications that are not
44  * guaranteed to benefit from it. When transparent hugepage support is
45  * enabled, it is for all mappings, and khugepaged scans all mappings.
46  * Defrag is invoked by khugepaged hugepage allocations and by page faults
47  * for all hugepage allocations.
48  */
49 unsigned long transparent_hugepage_flags __read_mostly =
50 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
51 	(1<<TRANSPARENT_HUGEPAGE_FLAG)|
52 #endif
53 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
54 	(1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
55 #endif
56 	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
57 	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
58 	(1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
59 
60 static struct shrinker deferred_split_shrinker;
61 
62 static atomic_t huge_zero_refcount;
63 struct page *huge_zero_page __read_mostly;
64 
65 bool transparent_hugepage_enabled(struct vm_area_struct *vma)
66 {
67 	/* The addr is used to check if the vma size fits */
68 	unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
69 
70 	if (!transhuge_vma_suitable(vma, addr))
71 		return false;
72 	if (vma_is_anonymous(vma))
73 		return __transparent_hugepage_enabled(vma);
74 	if (vma_is_shmem(vma))
75 		return shmem_huge_enabled(vma);
76 
77 	return false;
78 }
79 
80 static struct page *get_huge_zero_page(void)
81 {
82 	struct page *zero_page;
83 retry:
84 	if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
85 		return READ_ONCE(huge_zero_page);
86 
87 	zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
88 			HPAGE_PMD_ORDER);
89 	if (!zero_page) {
90 		count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
91 		return NULL;
92 	}
93 	count_vm_event(THP_ZERO_PAGE_ALLOC);
94 	preempt_disable();
95 	if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
96 		preempt_enable();
97 		__free_pages(zero_page, compound_order(zero_page));
98 		goto retry;
99 	}
100 
101 	/* We take additional reference here. It will be put back by shrinker */
102 	atomic_set(&huge_zero_refcount, 2);
103 	preempt_enable();
104 	return READ_ONCE(huge_zero_page);
105 }
106 
107 static void put_huge_zero_page(void)
108 {
109 	/*
110 	 * Counter should never go to zero here. Only shrinker can put
111 	 * last reference.
112 	 */
113 	BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
114 }
115 
116 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
117 {
118 	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
119 		return READ_ONCE(huge_zero_page);
120 
121 	if (!get_huge_zero_page())
122 		return NULL;
123 
124 	if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
125 		put_huge_zero_page();
126 
127 	return READ_ONCE(huge_zero_page);
128 }
129 
130 void mm_put_huge_zero_page(struct mm_struct *mm)
131 {
132 	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
133 		put_huge_zero_page();
134 }
135 
136 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
137 					struct shrink_control *sc)
138 {
139 	/* we can free zero page only if last reference remains */
140 	return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
141 }
142 
143 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
144 				       struct shrink_control *sc)
145 {
146 	if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
147 		struct page *zero_page = xchg(&huge_zero_page, NULL);
148 		BUG_ON(zero_page == NULL);
149 		__free_pages(zero_page, compound_order(zero_page));
150 		return HPAGE_PMD_NR;
151 	}
152 
153 	return 0;
154 }
155 
156 static struct shrinker huge_zero_page_shrinker = {
157 	.count_objects = shrink_huge_zero_page_count,
158 	.scan_objects = shrink_huge_zero_page_scan,
159 	.seeks = DEFAULT_SEEKS,
160 };
161 
162 #ifdef CONFIG_SYSFS
163 static ssize_t enabled_show(struct kobject *kobj,
164 			    struct kobj_attribute *attr, char *buf)
165 {
166 	if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
167 		return sprintf(buf, "[always] madvise never\n");
168 	else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
169 		return sprintf(buf, "always [madvise] never\n");
170 	else
171 		return sprintf(buf, "always madvise [never]\n");
172 }
173 
174 static ssize_t enabled_store(struct kobject *kobj,
175 			     struct kobj_attribute *attr,
176 			     const char *buf, size_t count)
177 {
178 	ssize_t ret = count;
179 
180 	if (sysfs_streq(buf, "always")) {
181 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
182 		set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
183 	} else if (sysfs_streq(buf, "madvise")) {
184 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
185 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
186 	} else if (sysfs_streq(buf, "never")) {
187 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
188 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
189 	} else
190 		ret = -EINVAL;
191 
192 	if (ret > 0) {
193 		int err = start_stop_khugepaged();
194 		if (err)
195 			ret = err;
196 	}
197 	return ret;
198 }
199 static struct kobj_attribute enabled_attr =
200 	__ATTR(enabled, 0644, enabled_show, enabled_store);
201 
202 ssize_t single_hugepage_flag_show(struct kobject *kobj,
203 				struct kobj_attribute *attr, char *buf,
204 				enum transparent_hugepage_flag flag)
205 {
206 	return sprintf(buf, "%d\n",
207 		       !!test_bit(flag, &transparent_hugepage_flags));
208 }
209 
210 ssize_t single_hugepage_flag_store(struct kobject *kobj,
211 				 struct kobj_attribute *attr,
212 				 const char *buf, size_t count,
213 				 enum transparent_hugepage_flag flag)
214 {
215 	unsigned long value;
216 	int ret;
217 
218 	ret = kstrtoul(buf, 10, &value);
219 	if (ret < 0)
220 		return ret;
221 	if (value > 1)
222 		return -EINVAL;
223 
224 	if (value)
225 		set_bit(flag, &transparent_hugepage_flags);
226 	else
227 		clear_bit(flag, &transparent_hugepage_flags);
228 
229 	return count;
230 }
231 
232 static ssize_t defrag_show(struct kobject *kobj,
233 			   struct kobj_attribute *attr, char *buf)
234 {
235 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
236 		return sprintf(buf, "[always] defer defer+madvise madvise never\n");
237 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
238 		return sprintf(buf, "always [defer] defer+madvise madvise never\n");
239 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
240 		return sprintf(buf, "always defer [defer+madvise] madvise never\n");
241 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
242 		return sprintf(buf, "always defer defer+madvise [madvise] never\n");
243 	return sprintf(buf, "always defer defer+madvise madvise [never]\n");
244 }
245 
246 static ssize_t defrag_store(struct kobject *kobj,
247 			    struct kobj_attribute *attr,
248 			    const char *buf, size_t count)
249 {
250 	if (sysfs_streq(buf, "always")) {
251 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
252 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
253 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
254 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
255 	} else if (sysfs_streq(buf, "defer+madvise")) {
256 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
257 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
258 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
259 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
260 	} else if (sysfs_streq(buf, "defer")) {
261 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
262 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
263 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
264 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
265 	} else if (sysfs_streq(buf, "madvise")) {
266 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
267 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
268 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
269 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
270 	} else if (sysfs_streq(buf, "never")) {
271 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
272 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
273 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
274 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
275 	} else
276 		return -EINVAL;
277 
278 	return count;
279 }
280 static struct kobj_attribute defrag_attr =
281 	__ATTR(defrag, 0644, defrag_show, defrag_store);
282 
283 static ssize_t use_zero_page_show(struct kobject *kobj,
284 		struct kobj_attribute *attr, char *buf)
285 {
286 	return single_hugepage_flag_show(kobj, attr, buf,
287 				TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
288 }
289 static ssize_t use_zero_page_store(struct kobject *kobj,
290 		struct kobj_attribute *attr, const char *buf, size_t count)
291 {
292 	return single_hugepage_flag_store(kobj, attr, buf, count,
293 				 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
294 }
295 static struct kobj_attribute use_zero_page_attr =
296 	__ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
297 
298 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
299 		struct kobj_attribute *attr, char *buf)
300 {
301 	return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
302 }
303 static struct kobj_attribute hpage_pmd_size_attr =
304 	__ATTR_RO(hpage_pmd_size);
305 
306 #ifdef CONFIG_DEBUG_VM
307 static ssize_t debug_cow_show(struct kobject *kobj,
308 				struct kobj_attribute *attr, char *buf)
309 {
310 	return single_hugepage_flag_show(kobj, attr, buf,
311 				TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
312 }
313 static ssize_t debug_cow_store(struct kobject *kobj,
314 			       struct kobj_attribute *attr,
315 			       const char *buf, size_t count)
316 {
317 	return single_hugepage_flag_store(kobj, attr, buf, count,
318 				 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
319 }
320 static struct kobj_attribute debug_cow_attr =
321 	__ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
322 #endif /* CONFIG_DEBUG_VM */
323 
324 static struct attribute *hugepage_attr[] = {
325 	&enabled_attr.attr,
326 	&defrag_attr.attr,
327 	&use_zero_page_attr.attr,
328 	&hpage_pmd_size_attr.attr,
329 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
330 	&shmem_enabled_attr.attr,
331 #endif
332 #ifdef CONFIG_DEBUG_VM
333 	&debug_cow_attr.attr,
334 #endif
335 	NULL,
336 };
337 
338 static const struct attribute_group hugepage_attr_group = {
339 	.attrs = hugepage_attr,
340 };
341 
342 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
343 {
344 	int err;
345 
346 	*hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
347 	if (unlikely(!*hugepage_kobj)) {
348 		pr_err("failed to create transparent hugepage kobject\n");
349 		return -ENOMEM;
350 	}
351 
352 	err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
353 	if (err) {
354 		pr_err("failed to register transparent hugepage group\n");
355 		goto delete_obj;
356 	}
357 
358 	err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
359 	if (err) {
360 		pr_err("failed to register transparent hugepage group\n");
361 		goto remove_hp_group;
362 	}
363 
364 	return 0;
365 
366 remove_hp_group:
367 	sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
368 delete_obj:
369 	kobject_put(*hugepage_kobj);
370 	return err;
371 }
372 
373 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
374 {
375 	sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
376 	sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
377 	kobject_put(hugepage_kobj);
378 }
379 #else
380 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
381 {
382 	return 0;
383 }
384 
385 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
386 {
387 }
388 #endif /* CONFIG_SYSFS */
389 
390 static int __init hugepage_init(void)
391 {
392 	int err;
393 	struct kobject *hugepage_kobj;
394 
395 	if (!has_transparent_hugepage()) {
396 		transparent_hugepage_flags = 0;
397 		return -EINVAL;
398 	}
399 
400 	/*
401 	 * hugepages can't be allocated by the buddy allocator
402 	 */
403 	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
404 	/*
405 	 * we use page->mapping and page->index in second tail page
406 	 * as list_head: assuming THP order >= 2
407 	 */
408 	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
409 
410 	err = hugepage_init_sysfs(&hugepage_kobj);
411 	if (err)
412 		goto err_sysfs;
413 
414 	err = khugepaged_init();
415 	if (err)
416 		goto err_slab;
417 
418 	err = register_shrinker(&huge_zero_page_shrinker);
419 	if (err)
420 		goto err_hzp_shrinker;
421 	err = register_shrinker(&deferred_split_shrinker);
422 	if (err)
423 		goto err_split_shrinker;
424 
425 	/*
426 	 * By default disable transparent hugepages on smaller systems,
427 	 * where the extra memory used could hurt more than TLB overhead
428 	 * is likely to save.  The admin can still enable it through /sys.
429 	 */
430 	if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
431 		transparent_hugepage_flags = 0;
432 		return 0;
433 	}
434 
435 	err = start_stop_khugepaged();
436 	if (err)
437 		goto err_khugepaged;
438 
439 	return 0;
440 err_khugepaged:
441 	unregister_shrinker(&deferred_split_shrinker);
442 err_split_shrinker:
443 	unregister_shrinker(&huge_zero_page_shrinker);
444 err_hzp_shrinker:
445 	khugepaged_destroy();
446 err_slab:
447 	hugepage_exit_sysfs(hugepage_kobj);
448 err_sysfs:
449 	return err;
450 }
451 subsys_initcall(hugepage_init);
452 
453 static int __init setup_transparent_hugepage(char *str)
454 {
455 	int ret = 0;
456 	if (!str)
457 		goto out;
458 	if (!strcmp(str, "always")) {
459 		set_bit(TRANSPARENT_HUGEPAGE_FLAG,
460 			&transparent_hugepage_flags);
461 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
462 			  &transparent_hugepage_flags);
463 		ret = 1;
464 	} else if (!strcmp(str, "madvise")) {
465 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
466 			  &transparent_hugepage_flags);
467 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
468 			&transparent_hugepage_flags);
469 		ret = 1;
470 	} else if (!strcmp(str, "never")) {
471 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
472 			  &transparent_hugepage_flags);
473 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
474 			  &transparent_hugepage_flags);
475 		ret = 1;
476 	}
477 out:
478 	if (!ret)
479 		pr_warn("transparent_hugepage= cannot parse, ignored\n");
480 	return ret;
481 }
482 __setup("transparent_hugepage=", setup_transparent_hugepage);
483 
484 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
485 {
486 	if (likely(vma->vm_flags & VM_WRITE))
487 		pmd = pmd_mkwrite(pmd);
488 	return pmd;
489 }
490 
491 #ifdef CONFIG_MEMCG
492 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
493 {
494 	struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
495 	struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
496 
497 	if (memcg)
498 		return &memcg->deferred_split_queue;
499 	else
500 		return &pgdat->deferred_split_queue;
501 }
502 #else
503 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
504 {
505 	struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
506 
507 	return &pgdat->deferred_split_queue;
508 }
509 #endif
510 
511 void prep_transhuge_page(struct page *page)
512 {
513 	/*
514 	 * we use page->mapping and page->indexlru in second tail page
515 	 * as list_head: assuming THP order >= 2
516 	 */
517 
518 	INIT_LIST_HEAD(page_deferred_list(page));
519 	set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
520 }
521 
522 bool is_transparent_hugepage(struct page *page)
523 {
524 	if (!PageCompound(page))
525 		return 0;
526 
527 	page = compound_head(page);
528 	return is_huge_zero_page(page) ||
529 	       page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
530 }
531 EXPORT_SYMBOL_GPL(is_transparent_hugepage);
532 
533 static unsigned long __thp_get_unmapped_area(struct file *filp,
534 		unsigned long addr, unsigned long len,
535 		loff_t off, unsigned long flags, unsigned long size)
536 {
537 	loff_t off_end = off + len;
538 	loff_t off_align = round_up(off, size);
539 	unsigned long len_pad, ret;
540 
541 	if (off_end <= off_align || (off_end - off_align) < size)
542 		return 0;
543 
544 	len_pad = len + size;
545 	if (len_pad < len || (off + len_pad) < off)
546 		return 0;
547 
548 	ret = current->mm->get_unmapped_area(filp, addr, len_pad,
549 					      off >> PAGE_SHIFT, flags);
550 
551 	/*
552 	 * The failure might be due to length padding. The caller will retry
553 	 * without the padding.
554 	 */
555 	if (IS_ERR_VALUE(ret))
556 		return 0;
557 
558 	/*
559 	 * Do not try to align to THP boundary if allocation at the address
560 	 * hint succeeds.
561 	 */
562 	if (ret == addr)
563 		return addr;
564 
565 	ret += (off - ret) & (size - 1);
566 	return ret;
567 }
568 
569 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
570 		unsigned long len, unsigned long pgoff, unsigned long flags)
571 {
572 	unsigned long ret;
573 	loff_t off = (loff_t)pgoff << PAGE_SHIFT;
574 
575 	if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
576 		goto out;
577 
578 	ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
579 	if (ret)
580 		return ret;
581 out:
582 	return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
583 }
584 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
585 
586 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
587 			struct page *page, gfp_t gfp)
588 {
589 	struct vm_area_struct *vma = vmf->vma;
590 	struct mem_cgroup *memcg;
591 	pgtable_t pgtable;
592 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
593 	vm_fault_t ret = 0;
594 
595 	VM_BUG_ON_PAGE(!PageCompound(page), page);
596 
597 	if (mem_cgroup_try_charge_delay(page, vma->vm_mm, gfp, &memcg, true)) {
598 		put_page(page);
599 		count_vm_event(THP_FAULT_FALLBACK);
600 		return VM_FAULT_FALLBACK;
601 	}
602 
603 	pgtable = pte_alloc_one(vma->vm_mm);
604 	if (unlikely(!pgtable)) {
605 		ret = VM_FAULT_OOM;
606 		goto release;
607 	}
608 
609 	clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
610 	/*
611 	 * The memory barrier inside __SetPageUptodate makes sure that
612 	 * clear_huge_page writes become visible before the set_pmd_at()
613 	 * write.
614 	 */
615 	__SetPageUptodate(page);
616 
617 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
618 	if (unlikely(!pmd_none(*vmf->pmd))) {
619 		goto unlock_release;
620 	} else {
621 		pmd_t entry;
622 
623 		ret = check_stable_address_space(vma->vm_mm);
624 		if (ret)
625 			goto unlock_release;
626 
627 		/* Deliver the page fault to userland */
628 		if (userfaultfd_missing(vma)) {
629 			vm_fault_t ret2;
630 
631 			spin_unlock(vmf->ptl);
632 			mem_cgroup_cancel_charge(page, memcg, true);
633 			put_page(page);
634 			pte_free(vma->vm_mm, pgtable);
635 			ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
636 			VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
637 			return ret2;
638 		}
639 
640 		entry = mk_huge_pmd(page, vma->vm_page_prot);
641 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
642 		page_add_new_anon_rmap(page, vma, haddr, true);
643 		mem_cgroup_commit_charge(page, memcg, false, true);
644 		lru_cache_add_active_or_unevictable(page, vma);
645 		pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
646 		set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
647 		add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
648 		mm_inc_nr_ptes(vma->vm_mm);
649 		spin_unlock(vmf->ptl);
650 		count_vm_event(THP_FAULT_ALLOC);
651 		count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
652 	}
653 
654 	return 0;
655 unlock_release:
656 	spin_unlock(vmf->ptl);
657 release:
658 	if (pgtable)
659 		pte_free(vma->vm_mm, pgtable);
660 	mem_cgroup_cancel_charge(page, memcg, true);
661 	put_page(page);
662 	return ret;
663 
664 }
665 
666 /*
667  * always: directly stall for all thp allocations
668  * defer: wake kswapd and fail if not immediately available
669  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
670  *		  fail if not immediately available
671  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
672  *	    available
673  * never: never stall for any thp allocation
674  */
675 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
676 {
677 	const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
678 
679 	/* Always do synchronous compaction */
680 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
681 		return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
682 
683 	/* Kick kcompactd and fail quickly */
684 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
685 		return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
686 
687 	/* Synchronous compaction if madvised, otherwise kick kcompactd */
688 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
689 		return GFP_TRANSHUGE_LIGHT |
690 			(vma_madvised ? __GFP_DIRECT_RECLAIM :
691 					__GFP_KSWAPD_RECLAIM);
692 
693 	/* Only do synchronous compaction if madvised */
694 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
695 		return GFP_TRANSHUGE_LIGHT |
696 		       (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
697 
698 	return GFP_TRANSHUGE_LIGHT;
699 }
700 
701 /* Caller must hold page table lock. */
702 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
703 		struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
704 		struct page *zero_page)
705 {
706 	pmd_t entry;
707 	if (!pmd_none(*pmd))
708 		return false;
709 	entry = mk_pmd(zero_page, vma->vm_page_prot);
710 	entry = pmd_mkhuge(entry);
711 	if (pgtable)
712 		pgtable_trans_huge_deposit(mm, pmd, pgtable);
713 	set_pmd_at(mm, haddr, pmd, entry);
714 	mm_inc_nr_ptes(mm);
715 	return true;
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 page *page;
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 	if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
730 		return VM_FAULT_OOM;
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 		bool set;
737 		vm_fault_t ret;
738 		pgtable = pte_alloc_one(vma->vm_mm);
739 		if (unlikely(!pgtable))
740 			return VM_FAULT_OOM;
741 		zero_page = mm_get_huge_zero_page(vma->vm_mm);
742 		if (unlikely(!zero_page)) {
743 			pte_free(vma->vm_mm, pgtable);
744 			count_vm_event(THP_FAULT_FALLBACK);
745 			return VM_FAULT_FALLBACK;
746 		}
747 		vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
748 		ret = 0;
749 		set = false;
750 		if (pmd_none(*vmf->pmd)) {
751 			ret = check_stable_address_space(vma->vm_mm);
752 			if (ret) {
753 				spin_unlock(vmf->ptl);
754 			} else if (userfaultfd_missing(vma)) {
755 				spin_unlock(vmf->ptl);
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 				spin_unlock(vmf->ptl);
762 				set = true;
763 			}
764 		} else
765 			spin_unlock(vmf->ptl);
766 		if (!set)
767 			pte_free(vma->vm_mm, pgtable);
768 		return ret;
769 	}
770 	gfp = alloc_hugepage_direct_gfpmask(vma);
771 	page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
772 	if (unlikely(!page)) {
773 		count_vm_event(THP_FAULT_FALLBACK);
774 		return VM_FAULT_FALLBACK;
775 	}
776 	prep_transhuge_page(page);
777 	return __do_huge_pmd_anonymous_page(vmf, page, gfp);
778 }
779 
780 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
781 		pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
782 		pgtable_t pgtable)
783 {
784 	struct mm_struct *mm = vma->vm_mm;
785 	pmd_t entry;
786 	spinlock_t *ptl;
787 
788 	ptl = pmd_lock(mm, pmd);
789 	if (!pmd_none(*pmd)) {
790 		if (write) {
791 			if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
792 				WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
793 				goto out_unlock;
794 			}
795 			entry = pmd_mkyoung(*pmd);
796 			entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
797 			if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
798 				update_mmu_cache_pmd(vma, addr, pmd);
799 		}
800 
801 		goto out_unlock;
802 	}
803 
804 	entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
805 	if (pfn_t_devmap(pfn))
806 		entry = pmd_mkdevmap(entry);
807 	if (write) {
808 		entry = pmd_mkyoung(pmd_mkdirty(entry));
809 		entry = maybe_pmd_mkwrite(entry, vma);
810 	}
811 
812 	if (pgtable) {
813 		pgtable_trans_huge_deposit(mm, pmd, pgtable);
814 		mm_inc_nr_ptes(mm);
815 		pgtable = NULL;
816 	}
817 
818 	set_pmd_at(mm, addr, pmd, entry);
819 	update_mmu_cache_pmd(vma, addr, pmd);
820 
821 out_unlock:
822 	spin_unlock(ptl);
823 	if (pgtable)
824 		pte_free(mm, pgtable);
825 }
826 
827 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)
828 {
829 	unsigned long addr = vmf->address & PMD_MASK;
830 	struct vm_area_struct *vma = vmf->vma;
831 	pgprot_t pgprot = vma->vm_page_prot;
832 	pgtable_t pgtable = NULL;
833 
834 	/*
835 	 * If we had pmd_special, we could avoid all these restrictions,
836 	 * but we need to be consistent with PTEs and architectures that
837 	 * can't support a 'special' bit.
838 	 */
839 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
840 			!pfn_t_devmap(pfn));
841 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
842 						(VM_PFNMAP|VM_MIXEDMAP));
843 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
844 
845 	if (addr < vma->vm_start || addr >= vma->vm_end)
846 		return VM_FAULT_SIGBUS;
847 
848 	if (arch_needs_pgtable_deposit()) {
849 		pgtable = pte_alloc_one(vma->vm_mm);
850 		if (!pgtable)
851 			return VM_FAULT_OOM;
852 	}
853 
854 	track_pfn_insert(vma, &pgprot, pfn);
855 
856 	insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
857 	return VM_FAULT_NOPAGE;
858 }
859 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
860 
861 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
862 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
863 {
864 	if (likely(vma->vm_flags & VM_WRITE))
865 		pud = pud_mkwrite(pud);
866 	return pud;
867 }
868 
869 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
870 		pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
871 {
872 	struct mm_struct *mm = vma->vm_mm;
873 	pud_t entry;
874 	spinlock_t *ptl;
875 
876 	ptl = pud_lock(mm, pud);
877 	if (!pud_none(*pud)) {
878 		if (write) {
879 			if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
880 				WARN_ON_ONCE(!is_huge_zero_pud(*pud));
881 				goto out_unlock;
882 			}
883 			entry = pud_mkyoung(*pud);
884 			entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
885 			if (pudp_set_access_flags(vma, addr, pud, entry, 1))
886 				update_mmu_cache_pud(vma, addr, pud);
887 		}
888 		goto out_unlock;
889 	}
890 
891 	entry = pud_mkhuge(pfn_t_pud(pfn, prot));
892 	if (pfn_t_devmap(pfn))
893 		entry = pud_mkdevmap(entry);
894 	if (write) {
895 		entry = pud_mkyoung(pud_mkdirty(entry));
896 		entry = maybe_pud_mkwrite(entry, vma);
897 	}
898 	set_pud_at(mm, addr, pud, entry);
899 	update_mmu_cache_pud(vma, addr, pud);
900 
901 out_unlock:
902 	spin_unlock(ptl);
903 }
904 
905 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)
906 {
907 	unsigned long addr = vmf->address & PUD_MASK;
908 	struct vm_area_struct *vma = vmf->vma;
909 	pgprot_t pgprot = vma->vm_page_prot;
910 
911 	/*
912 	 * If we had pud_special, we could avoid all these restrictions,
913 	 * but we need to be consistent with PTEs and architectures that
914 	 * can't support a 'special' bit.
915 	 */
916 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
917 			!pfn_t_devmap(pfn));
918 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
919 						(VM_PFNMAP|VM_MIXEDMAP));
920 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
921 
922 	if (addr < vma->vm_start || addr >= vma->vm_end)
923 		return VM_FAULT_SIGBUS;
924 
925 	track_pfn_insert(vma, &pgprot, pfn);
926 
927 	insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
928 	return VM_FAULT_NOPAGE;
929 }
930 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
931 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
932 
933 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
934 		pmd_t *pmd, int flags)
935 {
936 	pmd_t _pmd;
937 
938 	_pmd = pmd_mkyoung(*pmd);
939 	if (flags & FOLL_WRITE)
940 		_pmd = pmd_mkdirty(_pmd);
941 	if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
942 				pmd, _pmd, flags & FOLL_WRITE))
943 		update_mmu_cache_pmd(vma, addr, pmd);
944 }
945 
946 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
947 		pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
948 {
949 	unsigned long pfn = pmd_pfn(*pmd);
950 	struct mm_struct *mm = vma->vm_mm;
951 	struct page *page;
952 
953 	assert_spin_locked(pmd_lockptr(mm, pmd));
954 
955 	/*
956 	 * When we COW a devmap PMD entry, we split it into PTEs, so we should
957 	 * not be in this function with `flags & FOLL_COW` set.
958 	 */
959 	WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
960 
961 	if (flags & FOLL_WRITE && !pmd_write(*pmd))
962 		return NULL;
963 
964 	if (pmd_present(*pmd) && pmd_devmap(*pmd))
965 		/* pass */;
966 	else
967 		return NULL;
968 
969 	if (flags & FOLL_TOUCH)
970 		touch_pmd(vma, addr, pmd, flags);
971 
972 	/*
973 	 * device mapped pages can only be returned if the
974 	 * caller will manage the page reference count.
975 	 */
976 	if (!(flags & FOLL_GET))
977 		return ERR_PTR(-EEXIST);
978 
979 	pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
980 	*pgmap = get_dev_pagemap(pfn, *pgmap);
981 	if (!*pgmap)
982 		return ERR_PTR(-EFAULT);
983 	page = pfn_to_page(pfn);
984 	get_page(page);
985 
986 	return page;
987 }
988 
989 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
990 		  pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
991 		  struct vm_area_struct *vma)
992 {
993 	spinlock_t *dst_ptl, *src_ptl;
994 	struct page *src_page;
995 	pmd_t pmd;
996 	pgtable_t pgtable = NULL;
997 	int ret = -ENOMEM;
998 
999 	/* Skip if can be re-fill on fault */
1000 	if (!vma_is_anonymous(vma))
1001 		return 0;
1002 
1003 	pgtable = pte_alloc_one(dst_mm);
1004 	if (unlikely(!pgtable))
1005 		goto out;
1006 
1007 	dst_ptl = pmd_lock(dst_mm, dst_pmd);
1008 	src_ptl = pmd_lockptr(src_mm, src_pmd);
1009 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1010 
1011 	ret = -EAGAIN;
1012 	pmd = *src_pmd;
1013 
1014 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1015 	if (unlikely(is_swap_pmd(pmd))) {
1016 		swp_entry_t entry = pmd_to_swp_entry(pmd);
1017 
1018 		VM_BUG_ON(!is_pmd_migration_entry(pmd));
1019 		if (is_write_migration_entry(entry)) {
1020 			make_migration_entry_read(&entry);
1021 			pmd = swp_entry_to_pmd(entry);
1022 			if (pmd_swp_soft_dirty(*src_pmd))
1023 				pmd = pmd_swp_mksoft_dirty(pmd);
1024 			set_pmd_at(src_mm, addr, src_pmd, pmd);
1025 		}
1026 		add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1027 		mm_inc_nr_ptes(dst_mm);
1028 		pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1029 		set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1030 		ret = 0;
1031 		goto out_unlock;
1032 	}
1033 #endif
1034 
1035 	if (unlikely(!pmd_trans_huge(pmd))) {
1036 		pte_free(dst_mm, pgtable);
1037 		goto out_unlock;
1038 	}
1039 	/*
1040 	 * When page table lock is held, the huge zero pmd should not be
1041 	 * under splitting since we don't split the page itself, only pmd to
1042 	 * a page table.
1043 	 */
1044 	if (is_huge_zero_pmd(pmd)) {
1045 		struct page *zero_page;
1046 		/*
1047 		 * get_huge_zero_page() will never allocate a new page here,
1048 		 * since we already have a zero page to copy. It just takes a
1049 		 * reference.
1050 		 */
1051 		zero_page = mm_get_huge_zero_page(dst_mm);
1052 		set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1053 				zero_page);
1054 		ret = 0;
1055 		goto out_unlock;
1056 	}
1057 
1058 	src_page = pmd_page(pmd);
1059 	VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1060 	get_page(src_page);
1061 	page_dup_rmap(src_page, true);
1062 	add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1063 	mm_inc_nr_ptes(dst_mm);
1064 	pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1065 
1066 	pmdp_set_wrprotect(src_mm, addr, src_pmd);
1067 	pmd = pmd_mkold(pmd_wrprotect(pmd));
1068 	set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1069 
1070 	ret = 0;
1071 out_unlock:
1072 	spin_unlock(src_ptl);
1073 	spin_unlock(dst_ptl);
1074 out:
1075 	return ret;
1076 }
1077 
1078 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1079 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1080 		pud_t *pud, int flags)
1081 {
1082 	pud_t _pud;
1083 
1084 	_pud = pud_mkyoung(*pud);
1085 	if (flags & FOLL_WRITE)
1086 		_pud = pud_mkdirty(_pud);
1087 	if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1088 				pud, _pud, flags & FOLL_WRITE))
1089 		update_mmu_cache_pud(vma, addr, pud);
1090 }
1091 
1092 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1093 		pud_t *pud, int flags, struct dev_pagemap **pgmap)
1094 {
1095 	unsigned long pfn = pud_pfn(*pud);
1096 	struct mm_struct *mm = vma->vm_mm;
1097 	struct page *page;
1098 
1099 	assert_spin_locked(pud_lockptr(mm, pud));
1100 
1101 	if (flags & FOLL_WRITE && !pud_write(*pud))
1102 		return NULL;
1103 
1104 	if (pud_present(*pud) && pud_devmap(*pud))
1105 		/* pass */;
1106 	else
1107 		return NULL;
1108 
1109 	if (flags & FOLL_TOUCH)
1110 		touch_pud(vma, addr, pud, flags);
1111 
1112 	/*
1113 	 * device mapped pages can only be returned if the
1114 	 * caller will manage the page reference count.
1115 	 */
1116 	if (!(flags & FOLL_GET))
1117 		return ERR_PTR(-EEXIST);
1118 
1119 	pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1120 	*pgmap = get_dev_pagemap(pfn, *pgmap);
1121 	if (!*pgmap)
1122 		return ERR_PTR(-EFAULT);
1123 	page = pfn_to_page(pfn);
1124 	get_page(page);
1125 
1126 	return page;
1127 }
1128 
1129 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1130 		  pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1131 		  struct vm_area_struct *vma)
1132 {
1133 	spinlock_t *dst_ptl, *src_ptl;
1134 	pud_t pud;
1135 	int ret;
1136 
1137 	dst_ptl = pud_lock(dst_mm, dst_pud);
1138 	src_ptl = pud_lockptr(src_mm, src_pud);
1139 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1140 
1141 	ret = -EAGAIN;
1142 	pud = *src_pud;
1143 	if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1144 		goto out_unlock;
1145 
1146 	/*
1147 	 * When page table lock is held, the huge zero pud should not be
1148 	 * under splitting since we don't split the page itself, only pud to
1149 	 * a page table.
1150 	 */
1151 	if (is_huge_zero_pud(pud)) {
1152 		/* No huge zero pud yet */
1153 	}
1154 
1155 	pudp_set_wrprotect(src_mm, addr, src_pud);
1156 	pud = pud_mkold(pud_wrprotect(pud));
1157 	set_pud_at(dst_mm, addr, dst_pud, pud);
1158 
1159 	ret = 0;
1160 out_unlock:
1161 	spin_unlock(src_ptl);
1162 	spin_unlock(dst_ptl);
1163 	return ret;
1164 }
1165 
1166 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1167 {
1168 	pud_t entry;
1169 	unsigned long haddr;
1170 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1171 
1172 	vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1173 	if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1174 		goto unlock;
1175 
1176 	entry = pud_mkyoung(orig_pud);
1177 	if (write)
1178 		entry = pud_mkdirty(entry);
1179 	haddr = vmf->address & HPAGE_PUD_MASK;
1180 	if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1181 		update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1182 
1183 unlock:
1184 	spin_unlock(vmf->ptl);
1185 }
1186 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1187 
1188 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1189 {
1190 	pmd_t entry;
1191 	unsigned long haddr;
1192 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1193 
1194 	vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1195 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1196 		goto unlock;
1197 
1198 	entry = pmd_mkyoung(orig_pmd);
1199 	if (write)
1200 		entry = pmd_mkdirty(entry);
1201 	haddr = vmf->address & HPAGE_PMD_MASK;
1202 	if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1203 		update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1204 
1205 unlock:
1206 	spin_unlock(vmf->ptl);
1207 }
1208 
1209 static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1210 			pmd_t orig_pmd, struct page *page)
1211 {
1212 	struct vm_area_struct *vma = vmf->vma;
1213 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1214 	struct mem_cgroup *memcg;
1215 	pgtable_t pgtable;
1216 	pmd_t _pmd;
1217 	int i;
1218 	vm_fault_t ret = 0;
1219 	struct page **pages;
1220 	struct mmu_notifier_range range;
1221 
1222 	pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1223 			      GFP_KERNEL);
1224 	if (unlikely(!pages)) {
1225 		ret |= VM_FAULT_OOM;
1226 		goto out;
1227 	}
1228 
1229 	for (i = 0; i < HPAGE_PMD_NR; i++) {
1230 		pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1231 					       vmf->address, page_to_nid(page));
1232 		if (unlikely(!pages[i] ||
1233 			     mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
1234 				     GFP_KERNEL, &memcg, false))) {
1235 			if (pages[i])
1236 				put_page(pages[i]);
1237 			while (--i >= 0) {
1238 				memcg = (void *)page_private(pages[i]);
1239 				set_page_private(pages[i], 0);
1240 				mem_cgroup_cancel_charge(pages[i], memcg,
1241 						false);
1242 				put_page(pages[i]);
1243 			}
1244 			kfree(pages);
1245 			ret |= VM_FAULT_OOM;
1246 			goto out;
1247 		}
1248 		set_page_private(pages[i], (unsigned long)memcg);
1249 	}
1250 
1251 	for (i = 0; i < HPAGE_PMD_NR; i++) {
1252 		copy_user_highpage(pages[i], page + i,
1253 				   haddr + PAGE_SIZE * i, vma);
1254 		__SetPageUptodate(pages[i]);
1255 		cond_resched();
1256 	}
1257 
1258 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1259 				haddr, haddr + HPAGE_PMD_SIZE);
1260 	mmu_notifier_invalidate_range_start(&range);
1261 
1262 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1263 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1264 		goto out_free_pages;
1265 	VM_BUG_ON_PAGE(!PageHead(page), page);
1266 
1267 	/*
1268 	 * Leave pmd empty until pte is filled note we must notify here as
1269 	 * concurrent CPU thread might write to new page before the call to
1270 	 * mmu_notifier_invalidate_range_end() happens which can lead to a
1271 	 * device seeing memory write in different order than CPU.
1272 	 *
1273 	 * See Documentation/vm/mmu_notifier.rst
1274 	 */
1275 	pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1276 
1277 	pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1278 	pmd_populate(vma->vm_mm, &_pmd, pgtable);
1279 
1280 	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1281 		pte_t entry;
1282 		entry = mk_pte(pages[i], vma->vm_page_prot);
1283 		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1284 		memcg = (void *)page_private(pages[i]);
1285 		set_page_private(pages[i], 0);
1286 		page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1287 		mem_cgroup_commit_charge(pages[i], memcg, false, false);
1288 		lru_cache_add_active_or_unevictable(pages[i], vma);
1289 		vmf->pte = pte_offset_map(&_pmd, haddr);
1290 		VM_BUG_ON(!pte_none(*vmf->pte));
1291 		set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1292 		pte_unmap(vmf->pte);
1293 	}
1294 	kfree(pages);
1295 
1296 	smp_wmb(); /* make pte visible before pmd */
1297 	pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1298 	page_remove_rmap(page, true);
1299 	spin_unlock(vmf->ptl);
1300 
1301 	/*
1302 	 * No need to double call mmu_notifier->invalidate_range() callback as
1303 	 * the above pmdp_huge_clear_flush_notify() did already call it.
1304 	 */
1305 	mmu_notifier_invalidate_range_only_end(&range);
1306 
1307 	ret |= VM_FAULT_WRITE;
1308 	put_page(page);
1309 
1310 out:
1311 	return ret;
1312 
1313 out_free_pages:
1314 	spin_unlock(vmf->ptl);
1315 	mmu_notifier_invalidate_range_end(&range);
1316 	for (i = 0; i < HPAGE_PMD_NR; i++) {
1317 		memcg = (void *)page_private(pages[i]);
1318 		set_page_private(pages[i], 0);
1319 		mem_cgroup_cancel_charge(pages[i], memcg, false);
1320 		put_page(pages[i]);
1321 	}
1322 	kfree(pages);
1323 	goto out;
1324 }
1325 
1326 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1327 {
1328 	struct vm_area_struct *vma = vmf->vma;
1329 	struct page *page = NULL, *new_page;
1330 	struct mem_cgroup *memcg;
1331 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1332 	struct mmu_notifier_range range;
1333 	gfp_t huge_gfp;			/* for allocation and charge */
1334 	vm_fault_t ret = 0;
1335 
1336 	vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1337 	VM_BUG_ON_VMA(!vma->anon_vma, vma);
1338 	if (is_huge_zero_pmd(orig_pmd))
1339 		goto alloc;
1340 	spin_lock(vmf->ptl);
1341 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1342 		goto out_unlock;
1343 
1344 	page = pmd_page(orig_pmd);
1345 	VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1346 	/*
1347 	 * We can only reuse the page if nobody else maps the huge page or it's
1348 	 * part.
1349 	 */
1350 	if (!trylock_page(page)) {
1351 		get_page(page);
1352 		spin_unlock(vmf->ptl);
1353 		lock_page(page);
1354 		spin_lock(vmf->ptl);
1355 		if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1356 			unlock_page(page);
1357 			put_page(page);
1358 			goto out_unlock;
1359 		}
1360 		put_page(page);
1361 	}
1362 	if (reuse_swap_page(page, NULL)) {
1363 		pmd_t entry;
1364 		entry = pmd_mkyoung(orig_pmd);
1365 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1366 		if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1367 			update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1368 		ret |= VM_FAULT_WRITE;
1369 		unlock_page(page);
1370 		goto out_unlock;
1371 	}
1372 	unlock_page(page);
1373 	get_page(page);
1374 	spin_unlock(vmf->ptl);
1375 alloc:
1376 	if (__transparent_hugepage_enabled(vma) &&
1377 	    !transparent_hugepage_debug_cow()) {
1378 		huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1379 		new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1380 	} else
1381 		new_page = NULL;
1382 
1383 	if (likely(new_page)) {
1384 		prep_transhuge_page(new_page);
1385 	} else {
1386 		if (!page) {
1387 			split_huge_pmd(vma, vmf->pmd, vmf->address);
1388 			ret |= VM_FAULT_FALLBACK;
1389 		} else {
1390 			ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1391 			if (ret & VM_FAULT_OOM) {
1392 				split_huge_pmd(vma, vmf->pmd, vmf->address);
1393 				ret |= VM_FAULT_FALLBACK;
1394 			}
1395 			put_page(page);
1396 		}
1397 		count_vm_event(THP_FAULT_FALLBACK);
1398 		goto out;
1399 	}
1400 
1401 	if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
1402 					huge_gfp, &memcg, true))) {
1403 		put_page(new_page);
1404 		split_huge_pmd(vma, vmf->pmd, vmf->address);
1405 		if (page)
1406 			put_page(page);
1407 		ret |= VM_FAULT_FALLBACK;
1408 		count_vm_event(THP_FAULT_FALLBACK);
1409 		goto out;
1410 	}
1411 
1412 	count_vm_event(THP_FAULT_ALLOC);
1413 	count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
1414 
1415 	if (!page)
1416 		clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1417 	else
1418 		copy_user_huge_page(new_page, page, vmf->address,
1419 				    vma, HPAGE_PMD_NR);
1420 	__SetPageUptodate(new_page);
1421 
1422 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1423 				haddr, haddr + HPAGE_PMD_SIZE);
1424 	mmu_notifier_invalidate_range_start(&range);
1425 
1426 	spin_lock(vmf->ptl);
1427 	if (page)
1428 		put_page(page);
1429 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1430 		spin_unlock(vmf->ptl);
1431 		mem_cgroup_cancel_charge(new_page, memcg, true);
1432 		put_page(new_page);
1433 		goto out_mn;
1434 	} else {
1435 		pmd_t entry;
1436 		entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1437 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1438 		pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1439 		page_add_new_anon_rmap(new_page, vma, haddr, true);
1440 		mem_cgroup_commit_charge(new_page, memcg, false, true);
1441 		lru_cache_add_active_or_unevictable(new_page, vma);
1442 		set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1443 		update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1444 		if (!page) {
1445 			add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1446 		} else {
1447 			VM_BUG_ON_PAGE(!PageHead(page), page);
1448 			page_remove_rmap(page, true);
1449 			put_page(page);
1450 		}
1451 		ret |= VM_FAULT_WRITE;
1452 	}
1453 	spin_unlock(vmf->ptl);
1454 out_mn:
1455 	/*
1456 	 * No need to double call mmu_notifier->invalidate_range() callback as
1457 	 * the above pmdp_huge_clear_flush_notify() did already call it.
1458 	 */
1459 	mmu_notifier_invalidate_range_only_end(&range);
1460 out:
1461 	return ret;
1462 out_unlock:
1463 	spin_unlock(vmf->ptl);
1464 	return ret;
1465 }
1466 
1467 /*
1468  * FOLL_FORCE can write to even unwritable pmd's, but only
1469  * after we've gone through a COW cycle and they are dirty.
1470  */
1471 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1472 {
1473 	return pmd_write(pmd) ||
1474 	       ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1475 }
1476 
1477 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1478 				   unsigned long addr,
1479 				   pmd_t *pmd,
1480 				   unsigned int flags)
1481 {
1482 	struct mm_struct *mm = vma->vm_mm;
1483 	struct page *page = NULL;
1484 
1485 	assert_spin_locked(pmd_lockptr(mm, pmd));
1486 
1487 	if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1488 		goto out;
1489 
1490 	/* Avoid dumping huge zero page */
1491 	if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1492 		return ERR_PTR(-EFAULT);
1493 
1494 	/* Full NUMA hinting faults to serialise migration in fault paths */
1495 	if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1496 		goto out;
1497 
1498 	page = pmd_page(*pmd);
1499 	VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1500 	if (flags & FOLL_TOUCH)
1501 		touch_pmd(vma, addr, pmd, flags);
1502 	if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1503 		/*
1504 		 * We don't mlock() pte-mapped THPs. This way we can avoid
1505 		 * leaking mlocked pages into non-VM_LOCKED VMAs.
1506 		 *
1507 		 * For anon THP:
1508 		 *
1509 		 * In most cases the pmd is the only mapping of the page as we
1510 		 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1511 		 * writable private mappings in populate_vma_page_range().
1512 		 *
1513 		 * The only scenario when we have the page shared here is if we
1514 		 * mlocking read-only mapping shared over fork(). We skip
1515 		 * mlocking such pages.
1516 		 *
1517 		 * For file THP:
1518 		 *
1519 		 * We can expect PageDoubleMap() to be stable under page lock:
1520 		 * for file pages we set it in page_add_file_rmap(), which
1521 		 * requires page to be locked.
1522 		 */
1523 
1524 		if (PageAnon(page) && compound_mapcount(page) != 1)
1525 			goto skip_mlock;
1526 		if (PageDoubleMap(page) || !page->mapping)
1527 			goto skip_mlock;
1528 		if (!trylock_page(page))
1529 			goto skip_mlock;
1530 		lru_add_drain();
1531 		if (page->mapping && !PageDoubleMap(page))
1532 			mlock_vma_page(page);
1533 		unlock_page(page);
1534 	}
1535 skip_mlock:
1536 	page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1537 	VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1538 	if (flags & FOLL_GET)
1539 		get_page(page);
1540 
1541 out:
1542 	return page;
1543 }
1544 
1545 /* NUMA hinting page fault entry point for trans huge pmds */
1546 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1547 {
1548 	struct vm_area_struct *vma = vmf->vma;
1549 	struct anon_vma *anon_vma = NULL;
1550 	struct page *page;
1551 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1552 	int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1553 	int target_nid, last_cpupid = -1;
1554 	bool page_locked;
1555 	bool migrated = false;
1556 	bool was_writable;
1557 	int flags = 0;
1558 
1559 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1560 	if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1561 		goto out_unlock;
1562 
1563 	/*
1564 	 * If there are potential migrations, wait for completion and retry
1565 	 * without disrupting NUMA hinting information. Do not relock and
1566 	 * check_same as the page may no longer be mapped.
1567 	 */
1568 	if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1569 		page = pmd_page(*vmf->pmd);
1570 		if (!get_page_unless_zero(page))
1571 			goto out_unlock;
1572 		spin_unlock(vmf->ptl);
1573 		put_and_wait_on_page_locked(page);
1574 		goto out;
1575 	}
1576 
1577 	page = pmd_page(pmd);
1578 	BUG_ON(is_huge_zero_page(page));
1579 	page_nid = page_to_nid(page);
1580 	last_cpupid = page_cpupid_last(page);
1581 	count_vm_numa_event(NUMA_HINT_FAULTS);
1582 	if (page_nid == this_nid) {
1583 		count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1584 		flags |= TNF_FAULT_LOCAL;
1585 	}
1586 
1587 	/* See similar comment in do_numa_page for explanation */
1588 	if (!pmd_savedwrite(pmd))
1589 		flags |= TNF_NO_GROUP;
1590 
1591 	/*
1592 	 * Acquire the page lock to serialise THP migrations but avoid dropping
1593 	 * page_table_lock if at all possible
1594 	 */
1595 	page_locked = trylock_page(page);
1596 	target_nid = mpol_misplaced(page, vma, haddr);
1597 	if (target_nid == NUMA_NO_NODE) {
1598 		/* If the page was locked, there are no parallel migrations */
1599 		if (page_locked)
1600 			goto clear_pmdnuma;
1601 	}
1602 
1603 	/* Migration could have started since the pmd_trans_migrating check */
1604 	if (!page_locked) {
1605 		page_nid = NUMA_NO_NODE;
1606 		if (!get_page_unless_zero(page))
1607 			goto out_unlock;
1608 		spin_unlock(vmf->ptl);
1609 		put_and_wait_on_page_locked(page);
1610 		goto out;
1611 	}
1612 
1613 	/*
1614 	 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1615 	 * to serialises splits
1616 	 */
1617 	get_page(page);
1618 	spin_unlock(vmf->ptl);
1619 	anon_vma = page_lock_anon_vma_read(page);
1620 
1621 	/* Confirm the PMD did not change while page_table_lock was released */
1622 	spin_lock(vmf->ptl);
1623 	if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1624 		unlock_page(page);
1625 		put_page(page);
1626 		page_nid = NUMA_NO_NODE;
1627 		goto out_unlock;
1628 	}
1629 
1630 	/* Bail if we fail to protect against THP splits for any reason */
1631 	if (unlikely(!anon_vma)) {
1632 		put_page(page);
1633 		page_nid = NUMA_NO_NODE;
1634 		goto clear_pmdnuma;
1635 	}
1636 
1637 	/*
1638 	 * Since we took the NUMA fault, we must have observed the !accessible
1639 	 * bit. Make sure all other CPUs agree with that, to avoid them
1640 	 * modifying the page we're about to migrate.
1641 	 *
1642 	 * Must be done under PTL such that we'll observe the relevant
1643 	 * inc_tlb_flush_pending().
1644 	 *
1645 	 * We are not sure a pending tlb flush here is for a huge page
1646 	 * mapping or not. Hence use the tlb range variant
1647 	 */
1648 	if (mm_tlb_flush_pending(vma->vm_mm)) {
1649 		flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1650 		/*
1651 		 * change_huge_pmd() released the pmd lock before
1652 		 * invalidating the secondary MMUs sharing the primary
1653 		 * MMU pagetables (with ->invalidate_range()). The
1654 		 * mmu_notifier_invalidate_range_end() (which
1655 		 * internally calls ->invalidate_range()) in
1656 		 * change_pmd_range() will run after us, so we can't
1657 		 * rely on it here and we need an explicit invalidate.
1658 		 */
1659 		mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1660 					      haddr + HPAGE_PMD_SIZE);
1661 	}
1662 
1663 	/*
1664 	 * Migrate the THP to the requested node, returns with page unlocked
1665 	 * and access rights restored.
1666 	 */
1667 	spin_unlock(vmf->ptl);
1668 
1669 	migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1670 				vmf->pmd, pmd, vmf->address, page, target_nid);
1671 	if (migrated) {
1672 		flags |= TNF_MIGRATED;
1673 		page_nid = target_nid;
1674 	} else
1675 		flags |= TNF_MIGRATE_FAIL;
1676 
1677 	goto out;
1678 clear_pmdnuma:
1679 	BUG_ON(!PageLocked(page));
1680 	was_writable = pmd_savedwrite(pmd);
1681 	pmd = pmd_modify(pmd, vma->vm_page_prot);
1682 	pmd = pmd_mkyoung(pmd);
1683 	if (was_writable)
1684 		pmd = pmd_mkwrite(pmd);
1685 	set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1686 	update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1687 	unlock_page(page);
1688 out_unlock:
1689 	spin_unlock(vmf->ptl);
1690 
1691 out:
1692 	if (anon_vma)
1693 		page_unlock_anon_vma_read(anon_vma);
1694 
1695 	if (page_nid != NUMA_NO_NODE)
1696 		task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1697 				flags);
1698 
1699 	return 0;
1700 }
1701 
1702 /*
1703  * Return true if we do MADV_FREE successfully on entire pmd page.
1704  * Otherwise, return false.
1705  */
1706 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1707 		pmd_t *pmd, unsigned long addr, unsigned long next)
1708 {
1709 	spinlock_t *ptl;
1710 	pmd_t orig_pmd;
1711 	struct page *page;
1712 	struct mm_struct *mm = tlb->mm;
1713 	bool ret = false;
1714 
1715 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1716 
1717 	ptl = pmd_trans_huge_lock(pmd, vma);
1718 	if (!ptl)
1719 		goto out_unlocked;
1720 
1721 	orig_pmd = *pmd;
1722 	if (is_huge_zero_pmd(orig_pmd))
1723 		goto out;
1724 
1725 	if (unlikely(!pmd_present(orig_pmd))) {
1726 		VM_BUG_ON(thp_migration_supported() &&
1727 				  !is_pmd_migration_entry(orig_pmd));
1728 		goto out;
1729 	}
1730 
1731 	page = pmd_page(orig_pmd);
1732 	/*
1733 	 * If other processes are mapping this page, we couldn't discard
1734 	 * the page unless they all do MADV_FREE so let's skip the page.
1735 	 */
1736 	if (page_mapcount(page) != 1)
1737 		goto out;
1738 
1739 	if (!trylock_page(page))
1740 		goto out;
1741 
1742 	/*
1743 	 * If user want to discard part-pages of THP, split it so MADV_FREE
1744 	 * will deactivate only them.
1745 	 */
1746 	if (next - addr != HPAGE_PMD_SIZE) {
1747 		get_page(page);
1748 		spin_unlock(ptl);
1749 		split_huge_page(page);
1750 		unlock_page(page);
1751 		put_page(page);
1752 		goto out_unlocked;
1753 	}
1754 
1755 	if (PageDirty(page))
1756 		ClearPageDirty(page);
1757 	unlock_page(page);
1758 
1759 	if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1760 		pmdp_invalidate(vma, addr, pmd);
1761 		orig_pmd = pmd_mkold(orig_pmd);
1762 		orig_pmd = pmd_mkclean(orig_pmd);
1763 
1764 		set_pmd_at(mm, addr, pmd, orig_pmd);
1765 		tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1766 	}
1767 
1768 	mark_page_lazyfree(page);
1769 	ret = true;
1770 out:
1771 	spin_unlock(ptl);
1772 out_unlocked:
1773 	return ret;
1774 }
1775 
1776 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1777 {
1778 	pgtable_t pgtable;
1779 
1780 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1781 	pte_free(mm, pgtable);
1782 	mm_dec_nr_ptes(mm);
1783 }
1784 
1785 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1786 		 pmd_t *pmd, unsigned long addr)
1787 {
1788 	pmd_t orig_pmd;
1789 	spinlock_t *ptl;
1790 
1791 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1792 
1793 	ptl = __pmd_trans_huge_lock(pmd, vma);
1794 	if (!ptl)
1795 		return 0;
1796 	/*
1797 	 * For architectures like ppc64 we look at deposited pgtable
1798 	 * when calling pmdp_huge_get_and_clear. So do the
1799 	 * pgtable_trans_huge_withdraw after finishing pmdp related
1800 	 * operations.
1801 	 */
1802 	orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1803 			tlb->fullmm);
1804 	tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1805 	if (vma_is_dax(vma)) {
1806 		if (arch_needs_pgtable_deposit())
1807 			zap_deposited_table(tlb->mm, pmd);
1808 		spin_unlock(ptl);
1809 		if (is_huge_zero_pmd(orig_pmd))
1810 			tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1811 	} else if (is_huge_zero_pmd(orig_pmd)) {
1812 		zap_deposited_table(tlb->mm, pmd);
1813 		spin_unlock(ptl);
1814 		tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1815 	} else {
1816 		struct page *page = NULL;
1817 		int flush_needed = 1;
1818 
1819 		if (pmd_present(orig_pmd)) {
1820 			page = pmd_page(orig_pmd);
1821 			page_remove_rmap(page, true);
1822 			VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1823 			VM_BUG_ON_PAGE(!PageHead(page), page);
1824 		} else if (thp_migration_supported()) {
1825 			swp_entry_t entry;
1826 
1827 			VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1828 			entry = pmd_to_swp_entry(orig_pmd);
1829 			page = pfn_to_page(swp_offset(entry));
1830 			flush_needed = 0;
1831 		} else
1832 			WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1833 
1834 		if (PageAnon(page)) {
1835 			zap_deposited_table(tlb->mm, pmd);
1836 			add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1837 		} else {
1838 			if (arch_needs_pgtable_deposit())
1839 				zap_deposited_table(tlb->mm, pmd);
1840 			add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1841 		}
1842 
1843 		spin_unlock(ptl);
1844 		if (flush_needed)
1845 			tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1846 	}
1847 	return 1;
1848 }
1849 
1850 #ifndef pmd_move_must_withdraw
1851 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1852 					 spinlock_t *old_pmd_ptl,
1853 					 struct vm_area_struct *vma)
1854 {
1855 	/*
1856 	 * With split pmd lock we also need to move preallocated
1857 	 * PTE page table if new_pmd is on different PMD page table.
1858 	 *
1859 	 * We also don't deposit and withdraw tables for file pages.
1860 	 */
1861 	return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1862 }
1863 #endif
1864 
1865 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1866 {
1867 #ifdef CONFIG_MEM_SOFT_DIRTY
1868 	if (unlikely(is_pmd_migration_entry(pmd)))
1869 		pmd = pmd_swp_mksoft_dirty(pmd);
1870 	else if (pmd_present(pmd))
1871 		pmd = pmd_mksoft_dirty(pmd);
1872 #endif
1873 	return pmd;
1874 }
1875 
1876 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1877 		  unsigned long new_addr, unsigned long old_end,
1878 		  pmd_t *old_pmd, pmd_t *new_pmd)
1879 {
1880 	spinlock_t *old_ptl, *new_ptl;
1881 	pmd_t pmd;
1882 	struct mm_struct *mm = vma->vm_mm;
1883 	bool force_flush = false;
1884 
1885 	if ((old_addr & ~HPAGE_PMD_MASK) ||
1886 	    (new_addr & ~HPAGE_PMD_MASK) ||
1887 	    old_end - old_addr < HPAGE_PMD_SIZE)
1888 		return false;
1889 
1890 	/*
1891 	 * The destination pmd shouldn't be established, free_pgtables()
1892 	 * should have release it.
1893 	 */
1894 	if (WARN_ON(!pmd_none(*new_pmd))) {
1895 		VM_BUG_ON(pmd_trans_huge(*new_pmd));
1896 		return false;
1897 	}
1898 
1899 	/*
1900 	 * We don't have to worry about the ordering of src and dst
1901 	 * ptlocks because exclusive mmap_sem prevents deadlock.
1902 	 */
1903 	old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1904 	if (old_ptl) {
1905 		new_ptl = pmd_lockptr(mm, new_pmd);
1906 		if (new_ptl != old_ptl)
1907 			spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1908 		pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1909 		if (pmd_present(pmd))
1910 			force_flush = true;
1911 		VM_BUG_ON(!pmd_none(*new_pmd));
1912 
1913 		if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1914 			pgtable_t pgtable;
1915 			pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1916 			pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1917 		}
1918 		pmd = move_soft_dirty_pmd(pmd);
1919 		set_pmd_at(mm, new_addr, new_pmd, pmd);
1920 		if (force_flush)
1921 			flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1922 		if (new_ptl != old_ptl)
1923 			spin_unlock(new_ptl);
1924 		spin_unlock(old_ptl);
1925 		return true;
1926 	}
1927 	return false;
1928 }
1929 
1930 /*
1931  * Returns
1932  *  - 0 if PMD could not be locked
1933  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1934  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1935  */
1936 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1937 		unsigned long addr, pgprot_t newprot, int prot_numa)
1938 {
1939 	struct mm_struct *mm = vma->vm_mm;
1940 	spinlock_t *ptl;
1941 	pmd_t entry;
1942 	bool preserve_write;
1943 	int ret;
1944 
1945 	ptl = __pmd_trans_huge_lock(pmd, vma);
1946 	if (!ptl)
1947 		return 0;
1948 
1949 	preserve_write = prot_numa && pmd_write(*pmd);
1950 	ret = 1;
1951 
1952 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1953 	if (is_swap_pmd(*pmd)) {
1954 		swp_entry_t entry = pmd_to_swp_entry(*pmd);
1955 
1956 		VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1957 		if (is_write_migration_entry(entry)) {
1958 			pmd_t newpmd;
1959 			/*
1960 			 * A protection check is difficult so
1961 			 * just be safe and disable write
1962 			 */
1963 			make_migration_entry_read(&entry);
1964 			newpmd = swp_entry_to_pmd(entry);
1965 			if (pmd_swp_soft_dirty(*pmd))
1966 				newpmd = pmd_swp_mksoft_dirty(newpmd);
1967 			set_pmd_at(mm, addr, pmd, newpmd);
1968 		}
1969 		goto unlock;
1970 	}
1971 #endif
1972 
1973 	/*
1974 	 * Avoid trapping faults against the zero page. The read-only
1975 	 * data is likely to be read-cached on the local CPU and
1976 	 * local/remote hits to the zero page are not interesting.
1977 	 */
1978 	if (prot_numa && is_huge_zero_pmd(*pmd))
1979 		goto unlock;
1980 
1981 	if (prot_numa && pmd_protnone(*pmd))
1982 		goto unlock;
1983 
1984 	/*
1985 	 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1986 	 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1987 	 * which is also under down_read(mmap_sem):
1988 	 *
1989 	 *	CPU0:				CPU1:
1990 	 *				change_huge_pmd(prot_numa=1)
1991 	 *				 pmdp_huge_get_and_clear_notify()
1992 	 * madvise_dontneed()
1993 	 *  zap_pmd_range()
1994 	 *   pmd_trans_huge(*pmd) == 0 (without ptl)
1995 	 *   // skip the pmd
1996 	 *				 set_pmd_at();
1997 	 *				 // pmd is re-established
1998 	 *
1999 	 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
2000 	 * which may break userspace.
2001 	 *
2002 	 * pmdp_invalidate() is required to make sure we don't miss
2003 	 * dirty/young flags set by hardware.
2004 	 */
2005 	entry = pmdp_invalidate(vma, addr, pmd);
2006 
2007 	entry = pmd_modify(entry, newprot);
2008 	if (preserve_write)
2009 		entry = pmd_mk_savedwrite(entry);
2010 	ret = HPAGE_PMD_NR;
2011 	set_pmd_at(mm, addr, pmd, entry);
2012 	BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
2013 unlock:
2014 	spin_unlock(ptl);
2015 	return ret;
2016 }
2017 
2018 /*
2019  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
2020  *
2021  * Note that if it returns page table lock pointer, this routine returns without
2022  * unlocking page table lock. So callers must unlock it.
2023  */
2024 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
2025 {
2026 	spinlock_t *ptl;
2027 	ptl = pmd_lock(vma->vm_mm, pmd);
2028 	if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
2029 			pmd_devmap(*pmd)))
2030 		return ptl;
2031 	spin_unlock(ptl);
2032 	return NULL;
2033 }
2034 
2035 /*
2036  * Returns true if a given pud maps a thp, false otherwise.
2037  *
2038  * Note that if it returns true, this routine returns without unlocking page
2039  * table lock. So callers must unlock it.
2040  */
2041 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2042 {
2043 	spinlock_t *ptl;
2044 
2045 	ptl = pud_lock(vma->vm_mm, pud);
2046 	if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2047 		return ptl;
2048 	spin_unlock(ptl);
2049 	return NULL;
2050 }
2051 
2052 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2053 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2054 		 pud_t *pud, unsigned long addr)
2055 {
2056 	spinlock_t *ptl;
2057 
2058 	ptl = __pud_trans_huge_lock(pud, vma);
2059 	if (!ptl)
2060 		return 0;
2061 	/*
2062 	 * For architectures like ppc64 we look at deposited pgtable
2063 	 * when calling pudp_huge_get_and_clear. So do the
2064 	 * pgtable_trans_huge_withdraw after finishing pudp related
2065 	 * operations.
2066 	 */
2067 	pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
2068 	tlb_remove_pud_tlb_entry(tlb, pud, addr);
2069 	if (vma_is_dax(vma)) {
2070 		spin_unlock(ptl);
2071 		/* No zero page support yet */
2072 	} else {
2073 		/* No support for anonymous PUD pages yet */
2074 		BUG();
2075 	}
2076 	return 1;
2077 }
2078 
2079 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2080 		unsigned long haddr)
2081 {
2082 	VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2083 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2084 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2085 	VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2086 
2087 	count_vm_event(THP_SPLIT_PUD);
2088 
2089 	pudp_huge_clear_flush_notify(vma, haddr, pud);
2090 }
2091 
2092 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2093 		unsigned long address)
2094 {
2095 	spinlock_t *ptl;
2096 	struct mmu_notifier_range range;
2097 
2098 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2099 				address & HPAGE_PUD_MASK,
2100 				(address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2101 	mmu_notifier_invalidate_range_start(&range);
2102 	ptl = pud_lock(vma->vm_mm, pud);
2103 	if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2104 		goto out;
2105 	__split_huge_pud_locked(vma, pud, range.start);
2106 
2107 out:
2108 	spin_unlock(ptl);
2109 	/*
2110 	 * No need to double call mmu_notifier->invalidate_range() callback as
2111 	 * the above pudp_huge_clear_flush_notify() did already call it.
2112 	 */
2113 	mmu_notifier_invalidate_range_only_end(&range);
2114 }
2115 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2116 
2117 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2118 		unsigned long haddr, pmd_t *pmd)
2119 {
2120 	struct mm_struct *mm = vma->vm_mm;
2121 	pgtable_t pgtable;
2122 	pmd_t _pmd;
2123 	int i;
2124 
2125 	/*
2126 	 * Leave pmd empty until pte is filled note that it is fine to delay
2127 	 * notification until mmu_notifier_invalidate_range_end() as we are
2128 	 * replacing a zero pmd write protected page with a zero pte write
2129 	 * protected page.
2130 	 *
2131 	 * See Documentation/vm/mmu_notifier.rst
2132 	 */
2133 	pmdp_huge_clear_flush(vma, haddr, pmd);
2134 
2135 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2136 	pmd_populate(mm, &_pmd, pgtable);
2137 
2138 	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2139 		pte_t *pte, entry;
2140 		entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2141 		entry = pte_mkspecial(entry);
2142 		pte = pte_offset_map(&_pmd, haddr);
2143 		VM_BUG_ON(!pte_none(*pte));
2144 		set_pte_at(mm, haddr, pte, entry);
2145 		pte_unmap(pte);
2146 	}
2147 	smp_wmb(); /* make pte visible before pmd */
2148 	pmd_populate(mm, pmd, pgtable);
2149 }
2150 
2151 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2152 		unsigned long haddr, bool freeze)
2153 {
2154 	struct mm_struct *mm = vma->vm_mm;
2155 	struct page *page;
2156 	pgtable_t pgtable;
2157 	pmd_t old_pmd, _pmd;
2158 	bool young, write, soft_dirty, pmd_migration = false;
2159 	unsigned long addr;
2160 	int i;
2161 
2162 	VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2163 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2164 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2165 	VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2166 				&& !pmd_devmap(*pmd));
2167 
2168 	count_vm_event(THP_SPLIT_PMD);
2169 
2170 	if (!vma_is_anonymous(vma)) {
2171 		_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2172 		/*
2173 		 * We are going to unmap this huge page. So
2174 		 * just go ahead and zap it
2175 		 */
2176 		if (arch_needs_pgtable_deposit())
2177 			zap_deposited_table(mm, pmd);
2178 		if (vma_is_dax(vma))
2179 			return;
2180 		page = pmd_page(_pmd);
2181 		if (!PageDirty(page) && pmd_dirty(_pmd))
2182 			set_page_dirty(page);
2183 		if (!PageReferenced(page) && pmd_young(_pmd))
2184 			SetPageReferenced(page);
2185 		page_remove_rmap(page, true);
2186 		put_page(page);
2187 		add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2188 		return;
2189 	} else if (is_huge_zero_pmd(*pmd)) {
2190 		/*
2191 		 * FIXME: Do we want to invalidate secondary mmu by calling
2192 		 * mmu_notifier_invalidate_range() see comments below inside
2193 		 * __split_huge_pmd() ?
2194 		 *
2195 		 * We are going from a zero huge page write protected to zero
2196 		 * small page also write protected so it does not seems useful
2197 		 * to invalidate secondary mmu at this time.
2198 		 */
2199 		return __split_huge_zero_page_pmd(vma, haddr, pmd);
2200 	}
2201 
2202 	/*
2203 	 * Up to this point the pmd is present and huge and userland has the
2204 	 * whole access to the hugepage during the split (which happens in
2205 	 * place). If we overwrite the pmd with the not-huge version pointing
2206 	 * to the pte here (which of course we could if all CPUs were bug
2207 	 * free), userland could trigger a small page size TLB miss on the
2208 	 * small sized TLB while the hugepage TLB entry is still established in
2209 	 * the huge TLB. Some CPU doesn't like that.
2210 	 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2211 	 * 383 on page 93. Intel should be safe but is also warns that it's
2212 	 * only safe if the permission and cache attributes of the two entries
2213 	 * loaded in the two TLB is identical (which should be the case here).
2214 	 * But it is generally safer to never allow small and huge TLB entries
2215 	 * for the same virtual address to be loaded simultaneously. So instead
2216 	 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2217 	 * current pmd notpresent (atomically because here the pmd_trans_huge
2218 	 * must remain set at all times on the pmd until the split is complete
2219 	 * for this pmd), then we flush the SMP TLB and finally we write the
2220 	 * non-huge version of the pmd entry with pmd_populate.
2221 	 */
2222 	old_pmd = pmdp_invalidate(vma, haddr, pmd);
2223 
2224 	pmd_migration = is_pmd_migration_entry(old_pmd);
2225 	if (unlikely(pmd_migration)) {
2226 		swp_entry_t entry;
2227 
2228 		entry = pmd_to_swp_entry(old_pmd);
2229 		page = pfn_to_page(swp_offset(entry));
2230 		write = is_write_migration_entry(entry);
2231 		young = false;
2232 		soft_dirty = pmd_swp_soft_dirty(old_pmd);
2233 	} else {
2234 		page = pmd_page(old_pmd);
2235 		if (pmd_dirty(old_pmd))
2236 			SetPageDirty(page);
2237 		write = pmd_write(old_pmd);
2238 		young = pmd_young(old_pmd);
2239 		soft_dirty = pmd_soft_dirty(old_pmd);
2240 	}
2241 	VM_BUG_ON_PAGE(!page_count(page), page);
2242 	page_ref_add(page, HPAGE_PMD_NR - 1);
2243 
2244 	/*
2245 	 * Withdraw the table only after we mark the pmd entry invalid.
2246 	 * This's critical for some architectures (Power).
2247 	 */
2248 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2249 	pmd_populate(mm, &_pmd, pgtable);
2250 
2251 	for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2252 		pte_t entry, *pte;
2253 		/*
2254 		 * Note that NUMA hinting access restrictions are not
2255 		 * transferred to avoid any possibility of altering
2256 		 * permissions across VMAs.
2257 		 */
2258 		if (freeze || pmd_migration) {
2259 			swp_entry_t swp_entry;
2260 			swp_entry = make_migration_entry(page + i, write);
2261 			entry = swp_entry_to_pte(swp_entry);
2262 			if (soft_dirty)
2263 				entry = pte_swp_mksoft_dirty(entry);
2264 		} else {
2265 			entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2266 			entry = maybe_mkwrite(entry, vma);
2267 			if (!write)
2268 				entry = pte_wrprotect(entry);
2269 			if (!young)
2270 				entry = pte_mkold(entry);
2271 			if (soft_dirty)
2272 				entry = pte_mksoft_dirty(entry);
2273 		}
2274 		pte = pte_offset_map(&_pmd, addr);
2275 		BUG_ON(!pte_none(*pte));
2276 		set_pte_at(mm, addr, pte, entry);
2277 		atomic_inc(&page[i]._mapcount);
2278 		pte_unmap(pte);
2279 	}
2280 
2281 	/*
2282 	 * Set PG_double_map before dropping compound_mapcount to avoid
2283 	 * false-negative page_mapped().
2284 	 */
2285 	if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2286 		for (i = 0; i < HPAGE_PMD_NR; i++)
2287 			atomic_inc(&page[i]._mapcount);
2288 	}
2289 
2290 	if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2291 		/* Last compound_mapcount is gone. */
2292 		__dec_node_page_state(page, NR_ANON_THPS);
2293 		if (TestClearPageDoubleMap(page)) {
2294 			/* No need in mapcount reference anymore */
2295 			for (i = 0; i < HPAGE_PMD_NR; i++)
2296 				atomic_dec(&page[i]._mapcount);
2297 		}
2298 	}
2299 
2300 	smp_wmb(); /* make pte visible before pmd */
2301 	pmd_populate(mm, pmd, pgtable);
2302 
2303 	if (freeze) {
2304 		for (i = 0; i < HPAGE_PMD_NR; i++) {
2305 			page_remove_rmap(page + i, false);
2306 			put_page(page + i);
2307 		}
2308 	}
2309 }
2310 
2311 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2312 		unsigned long address, bool freeze, struct page *page)
2313 {
2314 	spinlock_t *ptl;
2315 	struct mmu_notifier_range range;
2316 
2317 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2318 				address & HPAGE_PMD_MASK,
2319 				(address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2320 	mmu_notifier_invalidate_range_start(&range);
2321 	ptl = pmd_lock(vma->vm_mm, pmd);
2322 
2323 	/*
2324 	 * If caller asks to setup a migration entries, we need a page to check
2325 	 * pmd against. Otherwise we can end up replacing wrong page.
2326 	 */
2327 	VM_BUG_ON(freeze && !page);
2328 	if (page && page != pmd_page(*pmd))
2329 	        goto out;
2330 
2331 	if (pmd_trans_huge(*pmd)) {
2332 		page = pmd_page(*pmd);
2333 		if (PageMlocked(page))
2334 			clear_page_mlock(page);
2335 	} else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2336 		goto out;
2337 	__split_huge_pmd_locked(vma, pmd, range.start, freeze);
2338 out:
2339 	spin_unlock(ptl);
2340 	/*
2341 	 * No need to double call mmu_notifier->invalidate_range() callback.
2342 	 * They are 3 cases to consider inside __split_huge_pmd_locked():
2343 	 *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2344 	 *  2) __split_huge_zero_page_pmd() read only zero page and any write
2345 	 *    fault will trigger a flush_notify before pointing to a new page
2346 	 *    (it is fine if the secondary mmu keeps pointing to the old zero
2347 	 *    page in the meantime)
2348 	 *  3) Split a huge pmd into pte pointing to the same page. No need
2349 	 *     to invalidate secondary tlb entry they are all still valid.
2350 	 *     any further changes to individual pte will notify. So no need
2351 	 *     to call mmu_notifier->invalidate_range()
2352 	 */
2353 	mmu_notifier_invalidate_range_only_end(&range);
2354 }
2355 
2356 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2357 		bool freeze, struct page *page)
2358 {
2359 	pgd_t *pgd;
2360 	p4d_t *p4d;
2361 	pud_t *pud;
2362 	pmd_t *pmd;
2363 
2364 	pgd = pgd_offset(vma->vm_mm, address);
2365 	if (!pgd_present(*pgd))
2366 		return;
2367 
2368 	p4d = p4d_offset(pgd, address);
2369 	if (!p4d_present(*p4d))
2370 		return;
2371 
2372 	pud = pud_offset(p4d, address);
2373 	if (!pud_present(*pud))
2374 		return;
2375 
2376 	pmd = pmd_offset(pud, address);
2377 
2378 	__split_huge_pmd(vma, pmd, address, freeze, page);
2379 }
2380 
2381 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2382 			     unsigned long start,
2383 			     unsigned long end,
2384 			     long adjust_next)
2385 {
2386 	/*
2387 	 * If the new start address isn't hpage aligned and it could
2388 	 * previously contain an hugepage: check if we need to split
2389 	 * an huge pmd.
2390 	 */
2391 	if (start & ~HPAGE_PMD_MASK &&
2392 	    (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2393 	    (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2394 		split_huge_pmd_address(vma, start, false, NULL);
2395 
2396 	/*
2397 	 * If the new end address isn't hpage aligned and it could
2398 	 * previously contain an hugepage: check if we need to split
2399 	 * an huge pmd.
2400 	 */
2401 	if (end & ~HPAGE_PMD_MASK &&
2402 	    (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2403 	    (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2404 		split_huge_pmd_address(vma, end, false, NULL);
2405 
2406 	/*
2407 	 * If we're also updating the vma->vm_next->vm_start, if the new
2408 	 * vm_next->vm_start isn't page aligned and it could previously
2409 	 * contain an hugepage: check if we need to split an huge pmd.
2410 	 */
2411 	if (adjust_next > 0) {
2412 		struct vm_area_struct *next = vma->vm_next;
2413 		unsigned long nstart = next->vm_start;
2414 		nstart += adjust_next << PAGE_SHIFT;
2415 		if (nstart & ~HPAGE_PMD_MASK &&
2416 		    (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2417 		    (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2418 			split_huge_pmd_address(next, nstart, false, NULL);
2419 	}
2420 }
2421 
2422 static void unmap_page(struct page *page)
2423 {
2424 	enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2425 		TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2426 	bool unmap_success;
2427 
2428 	VM_BUG_ON_PAGE(!PageHead(page), page);
2429 
2430 	if (PageAnon(page))
2431 		ttu_flags |= TTU_SPLIT_FREEZE;
2432 
2433 	unmap_success = try_to_unmap(page, ttu_flags);
2434 	VM_BUG_ON_PAGE(!unmap_success, page);
2435 }
2436 
2437 static void remap_page(struct page *page)
2438 {
2439 	int i;
2440 	if (PageTransHuge(page)) {
2441 		remove_migration_ptes(page, page, true);
2442 	} else {
2443 		for (i = 0; i < HPAGE_PMD_NR; i++)
2444 			remove_migration_ptes(page + i, page + i, true);
2445 	}
2446 }
2447 
2448 static void __split_huge_page_tail(struct page *head, int tail,
2449 		struct lruvec *lruvec, struct list_head *list)
2450 {
2451 	struct page *page_tail = head + tail;
2452 
2453 	VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2454 
2455 	/*
2456 	 * Clone page flags before unfreezing refcount.
2457 	 *
2458 	 * After successful get_page_unless_zero() might follow flags change,
2459 	 * for exmaple lock_page() which set PG_waiters.
2460 	 */
2461 	page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2462 	page_tail->flags |= (head->flags &
2463 			((1L << PG_referenced) |
2464 			 (1L << PG_swapbacked) |
2465 			 (1L << PG_swapcache) |
2466 			 (1L << PG_mlocked) |
2467 			 (1L << PG_uptodate) |
2468 			 (1L << PG_active) |
2469 			 (1L << PG_workingset) |
2470 			 (1L << PG_locked) |
2471 			 (1L << PG_unevictable) |
2472 			 (1L << PG_dirty)));
2473 
2474 	/* ->mapping in first tail page is compound_mapcount */
2475 	VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2476 			page_tail);
2477 	page_tail->mapping = head->mapping;
2478 	page_tail->index = head->index + tail;
2479 
2480 	/* Page flags must be visible before we make the page non-compound. */
2481 	smp_wmb();
2482 
2483 	/*
2484 	 * Clear PageTail before unfreezing page refcount.
2485 	 *
2486 	 * After successful get_page_unless_zero() might follow put_page()
2487 	 * which needs correct compound_head().
2488 	 */
2489 	clear_compound_head(page_tail);
2490 
2491 	/* Finally unfreeze refcount. Additional reference from page cache. */
2492 	page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2493 					  PageSwapCache(head)));
2494 
2495 	if (page_is_young(head))
2496 		set_page_young(page_tail);
2497 	if (page_is_idle(head))
2498 		set_page_idle(page_tail);
2499 
2500 	page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2501 
2502 	/*
2503 	 * always add to the tail because some iterators expect new
2504 	 * pages to show after the currently processed elements - e.g.
2505 	 * migrate_pages
2506 	 */
2507 	lru_add_page_tail(head, page_tail, lruvec, list);
2508 }
2509 
2510 static void __split_huge_page(struct page *page, struct list_head *list,
2511 		pgoff_t end, unsigned long flags)
2512 {
2513 	struct page *head = compound_head(page);
2514 	pg_data_t *pgdat = page_pgdat(head);
2515 	struct lruvec *lruvec;
2516 	struct address_space *swap_cache = NULL;
2517 	unsigned long offset = 0;
2518 	int i;
2519 
2520 	lruvec = mem_cgroup_page_lruvec(head, pgdat);
2521 
2522 	/* complete memcg works before add pages to LRU */
2523 	mem_cgroup_split_huge_fixup(head);
2524 
2525 	if (PageAnon(head) && PageSwapCache(head)) {
2526 		swp_entry_t entry = { .val = page_private(head) };
2527 
2528 		offset = swp_offset(entry);
2529 		swap_cache = swap_address_space(entry);
2530 		xa_lock(&swap_cache->i_pages);
2531 	}
2532 
2533 	for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2534 		__split_huge_page_tail(head, i, lruvec, list);
2535 		/* Some pages can be beyond i_size: drop them from page cache */
2536 		if (head[i].index >= end) {
2537 			ClearPageDirty(head + i);
2538 			__delete_from_page_cache(head + i, NULL);
2539 			if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2540 				shmem_uncharge(head->mapping->host, 1);
2541 			put_page(head + i);
2542 		} else if (!PageAnon(page)) {
2543 			__xa_store(&head->mapping->i_pages, head[i].index,
2544 					head + i, 0);
2545 		} else if (swap_cache) {
2546 			__xa_store(&swap_cache->i_pages, offset + i,
2547 					head + i, 0);
2548 		}
2549 	}
2550 
2551 	ClearPageCompound(head);
2552 
2553 	split_page_owner(head, HPAGE_PMD_ORDER);
2554 
2555 	/* See comment in __split_huge_page_tail() */
2556 	if (PageAnon(head)) {
2557 		/* Additional pin to swap cache */
2558 		if (PageSwapCache(head)) {
2559 			page_ref_add(head, 2);
2560 			xa_unlock(&swap_cache->i_pages);
2561 		} else {
2562 			page_ref_inc(head);
2563 		}
2564 	} else {
2565 		/* Additional pin to page cache */
2566 		page_ref_add(head, 2);
2567 		xa_unlock(&head->mapping->i_pages);
2568 	}
2569 
2570 	spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2571 
2572 	remap_page(head);
2573 
2574 	for (i = 0; i < HPAGE_PMD_NR; i++) {
2575 		struct page *subpage = head + i;
2576 		if (subpage == page)
2577 			continue;
2578 		unlock_page(subpage);
2579 
2580 		/*
2581 		 * Subpages may be freed if there wasn't any mapping
2582 		 * like if add_to_swap() is running on a lru page that
2583 		 * had its mapping zapped. And freeing these pages
2584 		 * requires taking the lru_lock so we do the put_page
2585 		 * of the tail pages after the split is complete.
2586 		 */
2587 		put_page(subpage);
2588 	}
2589 }
2590 
2591 int total_mapcount(struct page *page)
2592 {
2593 	int i, compound, ret;
2594 
2595 	VM_BUG_ON_PAGE(PageTail(page), page);
2596 
2597 	if (likely(!PageCompound(page)))
2598 		return atomic_read(&page->_mapcount) + 1;
2599 
2600 	compound = compound_mapcount(page);
2601 	if (PageHuge(page))
2602 		return compound;
2603 	ret = compound;
2604 	for (i = 0; i < HPAGE_PMD_NR; i++)
2605 		ret += atomic_read(&page[i]._mapcount) + 1;
2606 	/* File pages has compound_mapcount included in _mapcount */
2607 	if (!PageAnon(page))
2608 		return ret - compound * HPAGE_PMD_NR;
2609 	if (PageDoubleMap(page))
2610 		ret -= HPAGE_PMD_NR;
2611 	return ret;
2612 }
2613 
2614 /*
2615  * This calculates accurately how many mappings a transparent hugepage
2616  * has (unlike page_mapcount() which isn't fully accurate). This full
2617  * accuracy is primarily needed to know if copy-on-write faults can
2618  * reuse the page and change the mapping to read-write instead of
2619  * copying them. At the same time this returns the total_mapcount too.
2620  *
2621  * The function returns the highest mapcount any one of the subpages
2622  * has. If the return value is one, even if different processes are
2623  * mapping different subpages of the transparent hugepage, they can
2624  * all reuse it, because each process is reusing a different subpage.
2625  *
2626  * The total_mapcount is instead counting all virtual mappings of the
2627  * subpages. If the total_mapcount is equal to "one", it tells the
2628  * caller all mappings belong to the same "mm" and in turn the
2629  * anon_vma of the transparent hugepage can become the vma->anon_vma
2630  * local one as no other process may be mapping any of the subpages.
2631  *
2632  * It would be more accurate to replace page_mapcount() with
2633  * page_trans_huge_mapcount(), however we only use
2634  * page_trans_huge_mapcount() in the copy-on-write faults where we
2635  * need full accuracy to avoid breaking page pinning, because
2636  * page_trans_huge_mapcount() is slower than page_mapcount().
2637  */
2638 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2639 {
2640 	int i, ret, _total_mapcount, mapcount;
2641 
2642 	/* hugetlbfs shouldn't call it */
2643 	VM_BUG_ON_PAGE(PageHuge(page), page);
2644 
2645 	if (likely(!PageTransCompound(page))) {
2646 		mapcount = atomic_read(&page->_mapcount) + 1;
2647 		if (total_mapcount)
2648 			*total_mapcount = mapcount;
2649 		return mapcount;
2650 	}
2651 
2652 	page = compound_head(page);
2653 
2654 	_total_mapcount = ret = 0;
2655 	for (i = 0; i < HPAGE_PMD_NR; i++) {
2656 		mapcount = atomic_read(&page[i]._mapcount) + 1;
2657 		ret = max(ret, mapcount);
2658 		_total_mapcount += mapcount;
2659 	}
2660 	if (PageDoubleMap(page)) {
2661 		ret -= 1;
2662 		_total_mapcount -= HPAGE_PMD_NR;
2663 	}
2664 	mapcount = compound_mapcount(page);
2665 	ret += mapcount;
2666 	_total_mapcount += mapcount;
2667 	if (total_mapcount)
2668 		*total_mapcount = _total_mapcount;
2669 	return ret;
2670 }
2671 
2672 /* Racy check whether the huge page can be split */
2673 bool can_split_huge_page(struct page *page, int *pextra_pins)
2674 {
2675 	int extra_pins;
2676 
2677 	/* Additional pins from page cache */
2678 	if (PageAnon(page))
2679 		extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2680 	else
2681 		extra_pins = HPAGE_PMD_NR;
2682 	if (pextra_pins)
2683 		*pextra_pins = extra_pins;
2684 	return total_mapcount(page) == page_count(page) - extra_pins - 1;
2685 }
2686 
2687 /*
2688  * This function splits huge page into normal pages. @page can point to any
2689  * subpage of huge page to split. Split doesn't change the position of @page.
2690  *
2691  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2692  * The huge page must be locked.
2693  *
2694  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2695  *
2696  * Both head page and tail pages will inherit mapping, flags, and so on from
2697  * the hugepage.
2698  *
2699  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2700  * they are not mapped.
2701  *
2702  * Returns 0 if the hugepage is split successfully.
2703  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2704  * us.
2705  */
2706 int split_huge_page_to_list(struct page *page, struct list_head *list)
2707 {
2708 	struct page *head = compound_head(page);
2709 	struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2710 	struct deferred_split *ds_queue = get_deferred_split_queue(head);
2711 	struct anon_vma *anon_vma = NULL;
2712 	struct address_space *mapping = NULL;
2713 	int count, mapcount, extra_pins, ret;
2714 	bool mlocked;
2715 	unsigned long flags;
2716 	pgoff_t end;
2717 
2718 	VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2719 	VM_BUG_ON_PAGE(!PageLocked(head), head);
2720 	VM_BUG_ON_PAGE(!PageCompound(head), head);
2721 
2722 	if (PageWriteback(head))
2723 		return -EBUSY;
2724 
2725 	if (PageAnon(head)) {
2726 		/*
2727 		 * The caller does not necessarily hold an mmap_sem that would
2728 		 * prevent the anon_vma disappearing so we first we take a
2729 		 * reference to it and then lock the anon_vma for write. This
2730 		 * is similar to page_lock_anon_vma_read except the write lock
2731 		 * is taken to serialise against parallel split or collapse
2732 		 * operations.
2733 		 */
2734 		anon_vma = page_get_anon_vma(head);
2735 		if (!anon_vma) {
2736 			ret = -EBUSY;
2737 			goto out;
2738 		}
2739 		end = -1;
2740 		mapping = NULL;
2741 		anon_vma_lock_write(anon_vma);
2742 	} else {
2743 		mapping = head->mapping;
2744 
2745 		/* Truncated ? */
2746 		if (!mapping) {
2747 			ret = -EBUSY;
2748 			goto out;
2749 		}
2750 
2751 		anon_vma = NULL;
2752 		i_mmap_lock_read(mapping);
2753 
2754 		/*
2755 		 *__split_huge_page() may need to trim off pages beyond EOF:
2756 		 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2757 		 * which cannot be nested inside the page tree lock. So note
2758 		 * end now: i_size itself may be changed at any moment, but
2759 		 * head page lock is good enough to serialize the trimming.
2760 		 */
2761 		end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2762 	}
2763 
2764 	/*
2765 	 * Racy check if we can split the page, before unmap_page() will
2766 	 * split PMDs
2767 	 */
2768 	if (!can_split_huge_page(head, &extra_pins)) {
2769 		ret = -EBUSY;
2770 		goto out_unlock;
2771 	}
2772 
2773 	mlocked = PageMlocked(head);
2774 	unmap_page(head);
2775 	VM_BUG_ON_PAGE(compound_mapcount(head), head);
2776 
2777 	/* Make sure the page is not on per-CPU pagevec as it takes pin */
2778 	if (mlocked)
2779 		lru_add_drain();
2780 
2781 	/* prevent PageLRU to go away from under us, and freeze lru stats */
2782 	spin_lock_irqsave(&pgdata->lru_lock, flags);
2783 
2784 	if (mapping) {
2785 		XA_STATE(xas, &mapping->i_pages, page_index(head));
2786 
2787 		/*
2788 		 * Check if the head page is present in page cache.
2789 		 * We assume all tail are present too, if head is there.
2790 		 */
2791 		xa_lock(&mapping->i_pages);
2792 		if (xas_load(&xas) != head)
2793 			goto fail;
2794 	}
2795 
2796 	/* Prevent deferred_split_scan() touching ->_refcount */
2797 	spin_lock(&ds_queue->split_queue_lock);
2798 	count = page_count(head);
2799 	mapcount = total_mapcount(head);
2800 	if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2801 		if (!list_empty(page_deferred_list(head))) {
2802 			ds_queue->split_queue_len--;
2803 			list_del(page_deferred_list(head));
2804 		}
2805 		spin_unlock(&ds_queue->split_queue_lock);
2806 		if (mapping) {
2807 			if (PageSwapBacked(head))
2808 				__dec_node_page_state(head, NR_SHMEM_THPS);
2809 			else
2810 				__dec_node_page_state(head, NR_FILE_THPS);
2811 		}
2812 
2813 		__split_huge_page(page, list, end, flags);
2814 		if (PageSwapCache(head)) {
2815 			swp_entry_t entry = { .val = page_private(head) };
2816 
2817 			ret = split_swap_cluster(entry);
2818 		} else
2819 			ret = 0;
2820 	} else {
2821 		if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2822 			pr_alert("total_mapcount: %u, page_count(): %u\n",
2823 					mapcount, count);
2824 			if (PageTail(page))
2825 				dump_page(head, NULL);
2826 			dump_page(page, "total_mapcount(head) > 0");
2827 			BUG();
2828 		}
2829 		spin_unlock(&ds_queue->split_queue_lock);
2830 fail:		if (mapping)
2831 			xa_unlock(&mapping->i_pages);
2832 		spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2833 		remap_page(head);
2834 		ret = -EBUSY;
2835 	}
2836 
2837 out_unlock:
2838 	if (anon_vma) {
2839 		anon_vma_unlock_write(anon_vma);
2840 		put_anon_vma(anon_vma);
2841 	}
2842 	if (mapping)
2843 		i_mmap_unlock_read(mapping);
2844 out:
2845 	count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2846 	return ret;
2847 }
2848 
2849 void free_transhuge_page(struct page *page)
2850 {
2851 	struct deferred_split *ds_queue = get_deferred_split_queue(page);
2852 	unsigned long flags;
2853 
2854 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2855 	if (!list_empty(page_deferred_list(page))) {
2856 		ds_queue->split_queue_len--;
2857 		list_del(page_deferred_list(page));
2858 	}
2859 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2860 	free_compound_page(page);
2861 }
2862 
2863 void deferred_split_huge_page(struct page *page)
2864 {
2865 	struct deferred_split *ds_queue = get_deferred_split_queue(page);
2866 #ifdef CONFIG_MEMCG
2867 	struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
2868 #endif
2869 	unsigned long flags;
2870 
2871 	VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2872 
2873 	/*
2874 	 * The try_to_unmap() in page reclaim path might reach here too,
2875 	 * this may cause a race condition to corrupt deferred split queue.
2876 	 * And, if page reclaim is already handling the same page, it is
2877 	 * unnecessary to handle it again in shrinker.
2878 	 *
2879 	 * Check PageSwapCache to determine if the page is being
2880 	 * handled by page reclaim since THP swap would add the page into
2881 	 * swap cache before calling try_to_unmap().
2882 	 */
2883 	if (PageSwapCache(page))
2884 		return;
2885 
2886 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2887 	if (list_empty(page_deferred_list(page))) {
2888 		count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2889 		list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2890 		ds_queue->split_queue_len++;
2891 #ifdef CONFIG_MEMCG
2892 		if (memcg)
2893 			memcg_set_shrinker_bit(memcg, page_to_nid(page),
2894 					       deferred_split_shrinker.id);
2895 #endif
2896 	}
2897 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2898 }
2899 
2900 static unsigned long deferred_split_count(struct shrinker *shrink,
2901 		struct shrink_control *sc)
2902 {
2903 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2904 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2905 
2906 #ifdef CONFIG_MEMCG
2907 	if (sc->memcg)
2908 		ds_queue = &sc->memcg->deferred_split_queue;
2909 #endif
2910 	return READ_ONCE(ds_queue->split_queue_len);
2911 }
2912 
2913 static unsigned long deferred_split_scan(struct shrinker *shrink,
2914 		struct shrink_control *sc)
2915 {
2916 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2917 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2918 	unsigned long flags;
2919 	LIST_HEAD(list), *pos, *next;
2920 	struct page *page;
2921 	int split = 0;
2922 
2923 #ifdef CONFIG_MEMCG
2924 	if (sc->memcg)
2925 		ds_queue = &sc->memcg->deferred_split_queue;
2926 #endif
2927 
2928 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2929 	/* Take pin on all head pages to avoid freeing them under us */
2930 	list_for_each_safe(pos, next, &ds_queue->split_queue) {
2931 		page = list_entry((void *)pos, struct page, mapping);
2932 		page = compound_head(page);
2933 		if (get_page_unless_zero(page)) {
2934 			list_move(page_deferred_list(page), &list);
2935 		} else {
2936 			/* We lost race with put_compound_page() */
2937 			list_del_init(page_deferred_list(page));
2938 			ds_queue->split_queue_len--;
2939 		}
2940 		if (!--sc->nr_to_scan)
2941 			break;
2942 	}
2943 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2944 
2945 	list_for_each_safe(pos, next, &list) {
2946 		page = list_entry((void *)pos, struct page, mapping);
2947 		if (!trylock_page(page))
2948 			goto next;
2949 		/* split_huge_page() removes page from list on success */
2950 		if (!split_huge_page(page))
2951 			split++;
2952 		unlock_page(page);
2953 next:
2954 		put_page(page);
2955 	}
2956 
2957 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2958 	list_splice_tail(&list, &ds_queue->split_queue);
2959 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2960 
2961 	/*
2962 	 * Stop shrinker if we didn't split any page, but the queue is empty.
2963 	 * This can happen if pages were freed under us.
2964 	 */
2965 	if (!split && list_empty(&ds_queue->split_queue))
2966 		return SHRINK_STOP;
2967 	return split;
2968 }
2969 
2970 static struct shrinker deferred_split_shrinker = {
2971 	.count_objects = deferred_split_count,
2972 	.scan_objects = deferred_split_scan,
2973 	.seeks = DEFAULT_SEEKS,
2974 	.flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2975 		 SHRINKER_NONSLAB,
2976 };
2977 
2978 #ifdef CONFIG_DEBUG_FS
2979 static int split_huge_pages_set(void *data, u64 val)
2980 {
2981 	struct zone *zone;
2982 	struct page *page;
2983 	unsigned long pfn, max_zone_pfn;
2984 	unsigned long total = 0, split = 0;
2985 
2986 	if (val != 1)
2987 		return -EINVAL;
2988 
2989 	for_each_populated_zone(zone) {
2990 		max_zone_pfn = zone_end_pfn(zone);
2991 		for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2992 			if (!pfn_valid(pfn))
2993 				continue;
2994 
2995 			page = pfn_to_page(pfn);
2996 			if (!get_page_unless_zero(page))
2997 				continue;
2998 
2999 			if (zone != page_zone(page))
3000 				goto next;
3001 
3002 			if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
3003 				goto next;
3004 
3005 			total++;
3006 			lock_page(page);
3007 			if (!split_huge_page(page))
3008 				split++;
3009 			unlock_page(page);
3010 next:
3011 			put_page(page);
3012 		}
3013 	}
3014 
3015 	pr_info("%lu of %lu THP split\n", split, total);
3016 
3017 	return 0;
3018 }
3019 DEFINE_DEBUGFS_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3020 		"%llu\n");
3021 
3022 static int __init split_huge_pages_debugfs(void)
3023 {
3024 	debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3025 			    &split_huge_pages_fops);
3026 	return 0;
3027 }
3028 late_initcall(split_huge_pages_debugfs);
3029 #endif
3030 
3031 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3032 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3033 		struct page *page)
3034 {
3035 	struct vm_area_struct *vma = pvmw->vma;
3036 	struct mm_struct *mm = vma->vm_mm;
3037 	unsigned long address = pvmw->address;
3038 	pmd_t pmdval;
3039 	swp_entry_t entry;
3040 	pmd_t pmdswp;
3041 
3042 	if (!(pvmw->pmd && !pvmw->pte))
3043 		return;
3044 
3045 	flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3046 	pmdval = *pvmw->pmd;
3047 	pmdp_invalidate(vma, address, pvmw->pmd);
3048 	if (pmd_dirty(pmdval))
3049 		set_page_dirty(page);
3050 	entry = make_migration_entry(page, pmd_write(pmdval));
3051 	pmdswp = swp_entry_to_pmd(entry);
3052 	if (pmd_soft_dirty(pmdval))
3053 		pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3054 	set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3055 	page_remove_rmap(page, true);
3056 	put_page(page);
3057 }
3058 
3059 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3060 {
3061 	struct vm_area_struct *vma = pvmw->vma;
3062 	struct mm_struct *mm = vma->vm_mm;
3063 	unsigned long address = pvmw->address;
3064 	unsigned long mmun_start = address & HPAGE_PMD_MASK;
3065 	pmd_t pmde;
3066 	swp_entry_t entry;
3067 
3068 	if (!(pvmw->pmd && !pvmw->pte))
3069 		return;
3070 
3071 	entry = pmd_to_swp_entry(*pvmw->pmd);
3072 	get_page(new);
3073 	pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3074 	if (pmd_swp_soft_dirty(*pvmw->pmd))
3075 		pmde = pmd_mksoft_dirty(pmde);
3076 	if (is_write_migration_entry(entry))
3077 		pmde = maybe_pmd_mkwrite(pmde, vma);
3078 
3079 	flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3080 	if (PageAnon(new))
3081 		page_add_anon_rmap(new, vma, mmun_start, true);
3082 	else
3083 		page_add_file_rmap(new, true);
3084 	set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3085 	if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3086 		mlock_vma_page(new);
3087 	update_mmu_cache_pmd(vma, address, pvmw->pmd);
3088 }
3089 #endif
3090