xref: /openbmc/linux/mm/huge_memory.c (revision 09b35b41)
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 (!memcmp("always", buf,
181 		    min(sizeof("always")-1, count))) {
182 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
183 		set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
184 	} else if (!memcmp("madvise", buf,
185 			   min(sizeof("madvise")-1, count))) {
186 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
187 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
188 	} else if (!memcmp("never", buf,
189 			   min(sizeof("never")-1, count))) {
190 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
191 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
192 	} else
193 		ret = -EINVAL;
194 
195 	if (ret > 0) {
196 		int err = start_stop_khugepaged();
197 		if (err)
198 			ret = err;
199 	}
200 	return ret;
201 }
202 static struct kobj_attribute enabled_attr =
203 	__ATTR(enabled, 0644, enabled_show, enabled_store);
204 
205 ssize_t single_hugepage_flag_show(struct kobject *kobj,
206 				struct kobj_attribute *attr, char *buf,
207 				enum transparent_hugepage_flag flag)
208 {
209 	return sprintf(buf, "%d\n",
210 		       !!test_bit(flag, &transparent_hugepage_flags));
211 }
212 
213 ssize_t single_hugepage_flag_store(struct kobject *kobj,
214 				 struct kobj_attribute *attr,
215 				 const char *buf, size_t count,
216 				 enum transparent_hugepage_flag flag)
217 {
218 	unsigned long value;
219 	int ret;
220 
221 	ret = kstrtoul(buf, 10, &value);
222 	if (ret < 0)
223 		return ret;
224 	if (value > 1)
225 		return -EINVAL;
226 
227 	if (value)
228 		set_bit(flag, &transparent_hugepage_flags);
229 	else
230 		clear_bit(flag, &transparent_hugepage_flags);
231 
232 	return count;
233 }
234 
235 static ssize_t defrag_show(struct kobject *kobj,
236 			   struct kobj_attribute *attr, char *buf)
237 {
238 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
239 		return sprintf(buf, "[always] defer defer+madvise madvise never\n");
240 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
241 		return sprintf(buf, "always [defer] defer+madvise madvise never\n");
242 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
243 		return sprintf(buf, "always defer [defer+madvise] madvise never\n");
244 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
245 		return sprintf(buf, "always defer defer+madvise [madvise] never\n");
246 	return sprintf(buf, "always defer defer+madvise madvise [never]\n");
247 }
248 
249 static ssize_t defrag_store(struct kobject *kobj,
250 			    struct kobj_attribute *attr,
251 			    const char *buf, size_t count)
252 {
253 	if (!memcmp("always", buf,
254 		    min(sizeof("always")-1, count))) {
255 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
256 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
257 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
258 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
259 	} else if (!memcmp("defer+madvise", buf,
260 		    min(sizeof("defer+madvise")-1, count))) {
261 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
262 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
263 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
264 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
265 	} else if (!memcmp("defer", buf,
266 		    min(sizeof("defer")-1, count))) {
267 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
268 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
269 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
270 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
271 	} else if (!memcmp("madvise", buf,
272 			   min(sizeof("madvise")-1, count))) {
273 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
274 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
275 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
276 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
277 	} else if (!memcmp("never", buf,
278 			   min(sizeof("never")-1, count))) {
279 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
280 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
281 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
282 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
283 	} else
284 		return -EINVAL;
285 
286 	return count;
287 }
288 static struct kobj_attribute defrag_attr =
289 	__ATTR(defrag, 0644, defrag_show, defrag_store);
290 
291 static ssize_t use_zero_page_show(struct kobject *kobj,
292 		struct kobj_attribute *attr, char *buf)
293 {
294 	return single_hugepage_flag_show(kobj, attr, buf,
295 				TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
296 }
297 static ssize_t use_zero_page_store(struct kobject *kobj,
298 		struct kobj_attribute *attr, const char *buf, size_t count)
299 {
300 	return single_hugepage_flag_store(kobj, attr, buf, count,
301 				 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
302 }
303 static struct kobj_attribute use_zero_page_attr =
304 	__ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
305 
306 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
307 		struct kobj_attribute *attr, char *buf)
308 {
309 	return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
310 }
311 static struct kobj_attribute hpage_pmd_size_attr =
312 	__ATTR_RO(hpage_pmd_size);
313 
314 #ifdef CONFIG_DEBUG_VM
315 static ssize_t debug_cow_show(struct kobject *kobj,
316 				struct kobj_attribute *attr, char *buf)
317 {
318 	return single_hugepage_flag_show(kobj, attr, buf,
319 				TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
320 }
321 static ssize_t debug_cow_store(struct kobject *kobj,
322 			       struct kobj_attribute *attr,
323 			       const char *buf, size_t count)
324 {
325 	return single_hugepage_flag_store(kobj, attr, buf, count,
326 				 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
327 }
328 static struct kobj_attribute debug_cow_attr =
329 	__ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
330 #endif /* CONFIG_DEBUG_VM */
331 
332 static struct attribute *hugepage_attr[] = {
333 	&enabled_attr.attr,
334 	&defrag_attr.attr,
335 	&use_zero_page_attr.attr,
336 	&hpage_pmd_size_attr.attr,
337 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
338 	&shmem_enabled_attr.attr,
339 #endif
340 #ifdef CONFIG_DEBUG_VM
341 	&debug_cow_attr.attr,
342 #endif
343 	NULL,
344 };
345 
346 static const struct attribute_group hugepage_attr_group = {
347 	.attrs = hugepage_attr,
348 };
349 
350 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
351 {
352 	int err;
353 
354 	*hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
355 	if (unlikely(!*hugepage_kobj)) {
356 		pr_err("failed to create transparent hugepage kobject\n");
357 		return -ENOMEM;
358 	}
359 
360 	err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
361 	if (err) {
362 		pr_err("failed to register transparent hugepage group\n");
363 		goto delete_obj;
364 	}
365 
366 	err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
367 	if (err) {
368 		pr_err("failed to register transparent hugepage group\n");
369 		goto remove_hp_group;
370 	}
371 
372 	return 0;
373 
374 remove_hp_group:
375 	sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
376 delete_obj:
377 	kobject_put(*hugepage_kobj);
378 	return err;
379 }
380 
381 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
382 {
383 	sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
384 	sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
385 	kobject_put(hugepage_kobj);
386 }
387 #else
388 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
389 {
390 	return 0;
391 }
392 
393 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
394 {
395 }
396 #endif /* CONFIG_SYSFS */
397 
398 static int __init hugepage_init(void)
399 {
400 	int err;
401 	struct kobject *hugepage_kobj;
402 
403 	if (!has_transparent_hugepage()) {
404 		transparent_hugepage_flags = 0;
405 		return -EINVAL;
406 	}
407 
408 	/*
409 	 * hugepages can't be allocated by the buddy allocator
410 	 */
411 	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
412 	/*
413 	 * we use page->mapping and page->index in second tail page
414 	 * as list_head: assuming THP order >= 2
415 	 */
416 	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
417 
418 	err = hugepage_init_sysfs(&hugepage_kobj);
419 	if (err)
420 		goto err_sysfs;
421 
422 	err = khugepaged_init();
423 	if (err)
424 		goto err_slab;
425 
426 	err = register_shrinker(&huge_zero_page_shrinker);
427 	if (err)
428 		goto err_hzp_shrinker;
429 	err = register_shrinker(&deferred_split_shrinker);
430 	if (err)
431 		goto err_split_shrinker;
432 
433 	/*
434 	 * By default disable transparent hugepages on smaller systems,
435 	 * where the extra memory used could hurt more than TLB overhead
436 	 * is likely to save.  The admin can still enable it through /sys.
437 	 */
438 	if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
439 		transparent_hugepage_flags = 0;
440 		return 0;
441 	}
442 
443 	err = start_stop_khugepaged();
444 	if (err)
445 		goto err_khugepaged;
446 
447 	return 0;
448 err_khugepaged:
449 	unregister_shrinker(&deferred_split_shrinker);
450 err_split_shrinker:
451 	unregister_shrinker(&huge_zero_page_shrinker);
452 err_hzp_shrinker:
453 	khugepaged_destroy();
454 err_slab:
455 	hugepage_exit_sysfs(hugepage_kobj);
456 err_sysfs:
457 	return err;
458 }
459 subsys_initcall(hugepage_init);
460 
461 static int __init setup_transparent_hugepage(char *str)
462 {
463 	int ret = 0;
464 	if (!str)
465 		goto out;
466 	if (!strcmp(str, "always")) {
467 		set_bit(TRANSPARENT_HUGEPAGE_FLAG,
468 			&transparent_hugepage_flags);
469 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
470 			  &transparent_hugepage_flags);
471 		ret = 1;
472 	} else if (!strcmp(str, "madvise")) {
473 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
474 			  &transparent_hugepage_flags);
475 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
476 			&transparent_hugepage_flags);
477 		ret = 1;
478 	} else if (!strcmp(str, "never")) {
479 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
480 			  &transparent_hugepage_flags);
481 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
482 			  &transparent_hugepage_flags);
483 		ret = 1;
484 	}
485 out:
486 	if (!ret)
487 		pr_warn("transparent_hugepage= cannot parse, ignored\n");
488 	return ret;
489 }
490 __setup("transparent_hugepage=", setup_transparent_hugepage);
491 
492 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
493 {
494 	if (likely(vma->vm_flags & VM_WRITE))
495 		pmd = pmd_mkwrite(pmd);
496 	return pmd;
497 }
498 
499 #ifdef CONFIG_MEMCG
500 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
501 {
502 	struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
503 	struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
504 
505 	if (memcg)
506 		return &memcg->deferred_split_queue;
507 	else
508 		return &pgdat->deferred_split_queue;
509 }
510 #else
511 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
512 {
513 	struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
514 
515 	return &pgdat->deferred_split_queue;
516 }
517 #endif
518 
519 void prep_transhuge_page(struct page *page)
520 {
521 	/*
522 	 * we use page->mapping and page->indexlru in second tail page
523 	 * as list_head: assuming THP order >= 2
524 	 */
525 
526 	INIT_LIST_HEAD(page_deferred_list(page));
527 	set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
528 }
529 
530 static unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
531 		loff_t off, unsigned long flags, unsigned long size)
532 {
533 	unsigned long addr;
534 	loff_t off_end = off + len;
535 	loff_t off_align = round_up(off, size);
536 	unsigned long len_pad;
537 
538 	if (off_end <= off_align || (off_end - off_align) < size)
539 		return 0;
540 
541 	len_pad = len + size;
542 	if (len_pad < len || (off + len_pad) < off)
543 		return 0;
544 
545 	addr = current->mm->get_unmapped_area(filp, 0, len_pad,
546 					      off >> PAGE_SHIFT, flags);
547 	if (IS_ERR_VALUE(addr))
548 		return 0;
549 
550 	addr += (off - addr) & (size - 1);
551 	return addr;
552 }
553 
554 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
555 		unsigned long len, unsigned long pgoff, unsigned long flags)
556 {
557 	loff_t off = (loff_t)pgoff << PAGE_SHIFT;
558 
559 	if (addr)
560 		goto out;
561 	if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
562 		goto out;
563 
564 	addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
565 	if (addr)
566 		return addr;
567 
568  out:
569 	return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
570 }
571 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
572 
573 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
574 			struct page *page, gfp_t gfp)
575 {
576 	struct vm_area_struct *vma = vmf->vma;
577 	struct mem_cgroup *memcg;
578 	pgtable_t pgtable;
579 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
580 	vm_fault_t ret = 0;
581 
582 	VM_BUG_ON_PAGE(!PageCompound(page), page);
583 
584 	if (mem_cgroup_try_charge_delay(page, vma->vm_mm, gfp, &memcg, true)) {
585 		put_page(page);
586 		count_vm_event(THP_FAULT_FALLBACK);
587 		return VM_FAULT_FALLBACK;
588 	}
589 
590 	pgtable = pte_alloc_one(vma->vm_mm);
591 	if (unlikely(!pgtable)) {
592 		ret = VM_FAULT_OOM;
593 		goto release;
594 	}
595 
596 	clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
597 	/*
598 	 * The memory barrier inside __SetPageUptodate makes sure that
599 	 * clear_huge_page writes become visible before the set_pmd_at()
600 	 * write.
601 	 */
602 	__SetPageUptodate(page);
603 
604 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
605 	if (unlikely(!pmd_none(*vmf->pmd))) {
606 		goto unlock_release;
607 	} else {
608 		pmd_t entry;
609 
610 		ret = check_stable_address_space(vma->vm_mm);
611 		if (ret)
612 			goto unlock_release;
613 
614 		/* Deliver the page fault to userland */
615 		if (userfaultfd_missing(vma)) {
616 			vm_fault_t ret2;
617 
618 			spin_unlock(vmf->ptl);
619 			mem_cgroup_cancel_charge(page, memcg, true);
620 			put_page(page);
621 			pte_free(vma->vm_mm, pgtable);
622 			ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
623 			VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
624 			return ret2;
625 		}
626 
627 		entry = mk_huge_pmd(page, vma->vm_page_prot);
628 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
629 		page_add_new_anon_rmap(page, vma, haddr, true);
630 		mem_cgroup_commit_charge(page, memcg, false, true);
631 		lru_cache_add_active_or_unevictable(page, vma);
632 		pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
633 		set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
634 		add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
635 		mm_inc_nr_ptes(vma->vm_mm);
636 		spin_unlock(vmf->ptl);
637 		count_vm_event(THP_FAULT_ALLOC);
638 		count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
639 	}
640 
641 	return 0;
642 unlock_release:
643 	spin_unlock(vmf->ptl);
644 release:
645 	if (pgtable)
646 		pte_free(vma->vm_mm, pgtable);
647 	mem_cgroup_cancel_charge(page, memcg, true);
648 	put_page(page);
649 	return ret;
650 
651 }
652 
653 /*
654  * always: directly stall for all thp allocations
655  * defer: wake kswapd and fail if not immediately available
656  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
657  *		  fail if not immediately available
658  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
659  *	    available
660  * never: never stall for any thp allocation
661  */
662 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma, unsigned long addr)
663 {
664 	const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
665 	gfp_t this_node = 0;
666 
667 #ifdef CONFIG_NUMA
668 	struct mempolicy *pol;
669 	/*
670 	 * __GFP_THISNODE is used only when __GFP_DIRECT_RECLAIM is not
671 	 * specified, to express a general desire to stay on the current
672 	 * node for optimistic allocation attempts. If the defrag mode
673 	 * and/or madvise hint requires the direct reclaim then we prefer
674 	 * to fallback to other node rather than node reclaim because that
675 	 * can lead to excessive reclaim even though there is free memory
676 	 * on other nodes. We expect that NUMA preferences are specified
677 	 * by memory policies.
678 	 */
679 	pol = get_vma_policy(vma, addr);
680 	if (pol->mode != MPOL_BIND)
681 		this_node = __GFP_THISNODE;
682 	mpol_cond_put(pol);
683 #endif
684 
685 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
686 		return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
687 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
688 		return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM | this_node;
689 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
690 		return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
691 							     __GFP_KSWAPD_RECLAIM | this_node);
692 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
693 		return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
694 							     this_node);
695 	return GFP_TRANSHUGE_LIGHT | this_node;
696 }
697 
698 /* Caller must hold page table lock. */
699 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
700 		struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
701 		struct page *zero_page)
702 {
703 	pmd_t entry;
704 	if (!pmd_none(*pmd))
705 		return false;
706 	entry = mk_pmd(zero_page, vma->vm_page_prot);
707 	entry = pmd_mkhuge(entry);
708 	if (pgtable)
709 		pgtable_trans_huge_deposit(mm, pmd, pgtable);
710 	set_pmd_at(mm, haddr, pmd, entry);
711 	mm_inc_nr_ptes(mm);
712 	return true;
713 }
714 
715 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
716 {
717 	struct vm_area_struct *vma = vmf->vma;
718 	gfp_t gfp;
719 	struct page *page;
720 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
721 
722 	if (!transhuge_vma_suitable(vma, haddr))
723 		return VM_FAULT_FALLBACK;
724 	if (unlikely(anon_vma_prepare(vma)))
725 		return VM_FAULT_OOM;
726 	if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
727 		return VM_FAULT_OOM;
728 	if (!(vmf->flags & FAULT_FLAG_WRITE) &&
729 			!mm_forbids_zeropage(vma->vm_mm) &&
730 			transparent_hugepage_use_zero_page()) {
731 		pgtable_t pgtable;
732 		struct page *zero_page;
733 		bool set;
734 		vm_fault_t ret;
735 		pgtable = pte_alloc_one(vma->vm_mm);
736 		if (unlikely(!pgtable))
737 			return VM_FAULT_OOM;
738 		zero_page = mm_get_huge_zero_page(vma->vm_mm);
739 		if (unlikely(!zero_page)) {
740 			pte_free(vma->vm_mm, pgtable);
741 			count_vm_event(THP_FAULT_FALLBACK);
742 			return VM_FAULT_FALLBACK;
743 		}
744 		vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
745 		ret = 0;
746 		set = false;
747 		if (pmd_none(*vmf->pmd)) {
748 			ret = check_stable_address_space(vma->vm_mm);
749 			if (ret) {
750 				spin_unlock(vmf->ptl);
751 			} else if (userfaultfd_missing(vma)) {
752 				spin_unlock(vmf->ptl);
753 				ret = handle_userfault(vmf, VM_UFFD_MISSING);
754 				VM_BUG_ON(ret & VM_FAULT_FALLBACK);
755 			} else {
756 				set_huge_zero_page(pgtable, vma->vm_mm, vma,
757 						   haddr, vmf->pmd, zero_page);
758 				spin_unlock(vmf->ptl);
759 				set = true;
760 			}
761 		} else
762 			spin_unlock(vmf->ptl);
763 		if (!set)
764 			pte_free(vma->vm_mm, pgtable);
765 		return ret;
766 	}
767 	gfp = alloc_hugepage_direct_gfpmask(vma, haddr);
768 	page = alloc_pages_vma(gfp, HPAGE_PMD_ORDER, vma, haddr, numa_node_id());
769 	if (unlikely(!page)) {
770 		count_vm_event(THP_FAULT_FALLBACK);
771 		return VM_FAULT_FALLBACK;
772 	}
773 	prep_transhuge_page(page);
774 	return __do_huge_pmd_anonymous_page(vmf, page, gfp);
775 }
776 
777 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
778 		pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
779 		pgtable_t pgtable)
780 {
781 	struct mm_struct *mm = vma->vm_mm;
782 	pmd_t entry;
783 	spinlock_t *ptl;
784 
785 	ptl = pmd_lock(mm, pmd);
786 	if (!pmd_none(*pmd)) {
787 		if (write) {
788 			if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
789 				WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
790 				goto out_unlock;
791 			}
792 			entry = pmd_mkyoung(*pmd);
793 			entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
794 			if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
795 				update_mmu_cache_pmd(vma, addr, pmd);
796 		}
797 
798 		goto out_unlock;
799 	}
800 
801 	entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
802 	if (pfn_t_devmap(pfn))
803 		entry = pmd_mkdevmap(entry);
804 	if (write) {
805 		entry = pmd_mkyoung(pmd_mkdirty(entry));
806 		entry = maybe_pmd_mkwrite(entry, vma);
807 	}
808 
809 	if (pgtable) {
810 		pgtable_trans_huge_deposit(mm, pmd, pgtable);
811 		mm_inc_nr_ptes(mm);
812 		pgtable = NULL;
813 	}
814 
815 	set_pmd_at(mm, addr, pmd, entry);
816 	update_mmu_cache_pmd(vma, addr, pmd);
817 
818 out_unlock:
819 	spin_unlock(ptl);
820 	if (pgtable)
821 		pte_free(mm, pgtable);
822 }
823 
824 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)
825 {
826 	unsigned long addr = vmf->address & PMD_MASK;
827 	struct vm_area_struct *vma = vmf->vma;
828 	pgprot_t pgprot = vma->vm_page_prot;
829 	pgtable_t pgtable = NULL;
830 
831 	/*
832 	 * If we had pmd_special, we could avoid all these restrictions,
833 	 * but we need to be consistent with PTEs and architectures that
834 	 * can't support a 'special' bit.
835 	 */
836 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
837 			!pfn_t_devmap(pfn));
838 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
839 						(VM_PFNMAP|VM_MIXEDMAP));
840 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
841 
842 	if (addr < vma->vm_start || addr >= vma->vm_end)
843 		return VM_FAULT_SIGBUS;
844 
845 	if (arch_needs_pgtable_deposit()) {
846 		pgtable = pte_alloc_one(vma->vm_mm);
847 		if (!pgtable)
848 			return VM_FAULT_OOM;
849 	}
850 
851 	track_pfn_insert(vma, &pgprot, pfn);
852 
853 	insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
854 	return VM_FAULT_NOPAGE;
855 }
856 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
857 
858 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
859 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
860 {
861 	if (likely(vma->vm_flags & VM_WRITE))
862 		pud = pud_mkwrite(pud);
863 	return pud;
864 }
865 
866 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
867 		pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
868 {
869 	struct mm_struct *mm = vma->vm_mm;
870 	pud_t entry;
871 	spinlock_t *ptl;
872 
873 	ptl = pud_lock(mm, pud);
874 	if (!pud_none(*pud)) {
875 		if (write) {
876 			if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
877 				WARN_ON_ONCE(!is_huge_zero_pud(*pud));
878 				goto out_unlock;
879 			}
880 			entry = pud_mkyoung(*pud);
881 			entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
882 			if (pudp_set_access_flags(vma, addr, pud, entry, 1))
883 				update_mmu_cache_pud(vma, addr, pud);
884 		}
885 		goto out_unlock;
886 	}
887 
888 	entry = pud_mkhuge(pfn_t_pud(pfn, prot));
889 	if (pfn_t_devmap(pfn))
890 		entry = pud_mkdevmap(entry);
891 	if (write) {
892 		entry = pud_mkyoung(pud_mkdirty(entry));
893 		entry = maybe_pud_mkwrite(entry, vma);
894 	}
895 	set_pud_at(mm, addr, pud, entry);
896 	update_mmu_cache_pud(vma, addr, pud);
897 
898 out_unlock:
899 	spin_unlock(ptl);
900 }
901 
902 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)
903 {
904 	unsigned long addr = vmf->address & PUD_MASK;
905 	struct vm_area_struct *vma = vmf->vma;
906 	pgprot_t pgprot = vma->vm_page_prot;
907 
908 	/*
909 	 * If we had pud_special, we could avoid all these restrictions,
910 	 * but we need to be consistent with PTEs and architectures that
911 	 * can't support a 'special' bit.
912 	 */
913 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
914 			!pfn_t_devmap(pfn));
915 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
916 						(VM_PFNMAP|VM_MIXEDMAP));
917 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
918 
919 	if (addr < vma->vm_start || addr >= vma->vm_end)
920 		return VM_FAULT_SIGBUS;
921 
922 	track_pfn_insert(vma, &pgprot, pfn);
923 
924 	insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
925 	return VM_FAULT_NOPAGE;
926 }
927 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
928 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
929 
930 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
931 		pmd_t *pmd, int flags)
932 {
933 	pmd_t _pmd;
934 
935 	_pmd = pmd_mkyoung(*pmd);
936 	if (flags & FOLL_WRITE)
937 		_pmd = pmd_mkdirty(_pmd);
938 	if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
939 				pmd, _pmd, flags & FOLL_WRITE))
940 		update_mmu_cache_pmd(vma, addr, pmd);
941 }
942 
943 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
944 		pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
945 {
946 	unsigned long pfn = pmd_pfn(*pmd);
947 	struct mm_struct *mm = vma->vm_mm;
948 	struct page *page;
949 
950 	assert_spin_locked(pmd_lockptr(mm, pmd));
951 
952 	/*
953 	 * When we COW a devmap PMD entry, we split it into PTEs, so we should
954 	 * not be in this function with `flags & FOLL_COW` set.
955 	 */
956 	WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
957 
958 	if (flags & FOLL_WRITE && !pmd_write(*pmd))
959 		return NULL;
960 
961 	if (pmd_present(*pmd) && pmd_devmap(*pmd))
962 		/* pass */;
963 	else
964 		return NULL;
965 
966 	if (flags & FOLL_TOUCH)
967 		touch_pmd(vma, addr, pmd, flags);
968 
969 	/*
970 	 * device mapped pages can only be returned if the
971 	 * caller will manage the page reference count.
972 	 */
973 	if (!(flags & FOLL_GET))
974 		return ERR_PTR(-EEXIST);
975 
976 	pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
977 	*pgmap = get_dev_pagemap(pfn, *pgmap);
978 	if (!*pgmap)
979 		return ERR_PTR(-EFAULT);
980 	page = pfn_to_page(pfn);
981 	get_page(page);
982 
983 	return page;
984 }
985 
986 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
987 		  pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
988 		  struct vm_area_struct *vma)
989 {
990 	spinlock_t *dst_ptl, *src_ptl;
991 	struct page *src_page;
992 	pmd_t pmd;
993 	pgtable_t pgtable = NULL;
994 	int ret = -ENOMEM;
995 
996 	/* Skip if can be re-fill on fault */
997 	if (!vma_is_anonymous(vma))
998 		return 0;
999 
1000 	pgtable = pte_alloc_one(dst_mm);
1001 	if (unlikely(!pgtable))
1002 		goto out;
1003 
1004 	dst_ptl = pmd_lock(dst_mm, dst_pmd);
1005 	src_ptl = pmd_lockptr(src_mm, src_pmd);
1006 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1007 
1008 	ret = -EAGAIN;
1009 	pmd = *src_pmd;
1010 
1011 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1012 	if (unlikely(is_swap_pmd(pmd))) {
1013 		swp_entry_t entry = pmd_to_swp_entry(pmd);
1014 
1015 		VM_BUG_ON(!is_pmd_migration_entry(pmd));
1016 		if (is_write_migration_entry(entry)) {
1017 			make_migration_entry_read(&entry);
1018 			pmd = swp_entry_to_pmd(entry);
1019 			if (pmd_swp_soft_dirty(*src_pmd))
1020 				pmd = pmd_swp_mksoft_dirty(pmd);
1021 			set_pmd_at(src_mm, addr, src_pmd, pmd);
1022 		}
1023 		add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1024 		mm_inc_nr_ptes(dst_mm);
1025 		pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1026 		set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1027 		ret = 0;
1028 		goto out_unlock;
1029 	}
1030 #endif
1031 
1032 	if (unlikely(!pmd_trans_huge(pmd))) {
1033 		pte_free(dst_mm, pgtable);
1034 		goto out_unlock;
1035 	}
1036 	/*
1037 	 * When page table lock is held, the huge zero pmd should not be
1038 	 * under splitting since we don't split the page itself, only pmd to
1039 	 * a page table.
1040 	 */
1041 	if (is_huge_zero_pmd(pmd)) {
1042 		struct page *zero_page;
1043 		/*
1044 		 * get_huge_zero_page() will never allocate a new page here,
1045 		 * since we already have a zero page to copy. It just takes a
1046 		 * reference.
1047 		 */
1048 		zero_page = mm_get_huge_zero_page(dst_mm);
1049 		set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1050 				zero_page);
1051 		ret = 0;
1052 		goto out_unlock;
1053 	}
1054 
1055 	src_page = pmd_page(pmd);
1056 	VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1057 	get_page(src_page);
1058 	page_dup_rmap(src_page, true);
1059 	add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1060 	mm_inc_nr_ptes(dst_mm);
1061 	pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1062 
1063 	pmdp_set_wrprotect(src_mm, addr, src_pmd);
1064 	pmd = pmd_mkold(pmd_wrprotect(pmd));
1065 	set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1066 
1067 	ret = 0;
1068 out_unlock:
1069 	spin_unlock(src_ptl);
1070 	spin_unlock(dst_ptl);
1071 out:
1072 	return ret;
1073 }
1074 
1075 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1076 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1077 		pud_t *pud, int flags)
1078 {
1079 	pud_t _pud;
1080 
1081 	_pud = pud_mkyoung(*pud);
1082 	if (flags & FOLL_WRITE)
1083 		_pud = pud_mkdirty(_pud);
1084 	if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1085 				pud, _pud, flags & FOLL_WRITE))
1086 		update_mmu_cache_pud(vma, addr, pud);
1087 }
1088 
1089 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1090 		pud_t *pud, int flags, struct dev_pagemap **pgmap)
1091 {
1092 	unsigned long pfn = pud_pfn(*pud);
1093 	struct mm_struct *mm = vma->vm_mm;
1094 	struct page *page;
1095 
1096 	assert_spin_locked(pud_lockptr(mm, pud));
1097 
1098 	if (flags & FOLL_WRITE && !pud_write(*pud))
1099 		return NULL;
1100 
1101 	if (pud_present(*pud) && pud_devmap(*pud))
1102 		/* pass */;
1103 	else
1104 		return NULL;
1105 
1106 	if (flags & FOLL_TOUCH)
1107 		touch_pud(vma, addr, pud, flags);
1108 
1109 	/*
1110 	 * device mapped pages can only be returned if the
1111 	 * caller will manage the page reference count.
1112 	 */
1113 	if (!(flags & FOLL_GET))
1114 		return ERR_PTR(-EEXIST);
1115 
1116 	pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1117 	*pgmap = get_dev_pagemap(pfn, *pgmap);
1118 	if (!*pgmap)
1119 		return ERR_PTR(-EFAULT);
1120 	page = pfn_to_page(pfn);
1121 	get_page(page);
1122 
1123 	return page;
1124 }
1125 
1126 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1127 		  pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1128 		  struct vm_area_struct *vma)
1129 {
1130 	spinlock_t *dst_ptl, *src_ptl;
1131 	pud_t pud;
1132 	int ret;
1133 
1134 	dst_ptl = pud_lock(dst_mm, dst_pud);
1135 	src_ptl = pud_lockptr(src_mm, src_pud);
1136 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1137 
1138 	ret = -EAGAIN;
1139 	pud = *src_pud;
1140 	if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1141 		goto out_unlock;
1142 
1143 	/*
1144 	 * When page table lock is held, the huge zero pud should not be
1145 	 * under splitting since we don't split the page itself, only pud to
1146 	 * a page table.
1147 	 */
1148 	if (is_huge_zero_pud(pud)) {
1149 		/* No huge zero pud yet */
1150 	}
1151 
1152 	pudp_set_wrprotect(src_mm, addr, src_pud);
1153 	pud = pud_mkold(pud_wrprotect(pud));
1154 	set_pud_at(dst_mm, addr, dst_pud, pud);
1155 
1156 	ret = 0;
1157 out_unlock:
1158 	spin_unlock(src_ptl);
1159 	spin_unlock(dst_ptl);
1160 	return ret;
1161 }
1162 
1163 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1164 {
1165 	pud_t entry;
1166 	unsigned long haddr;
1167 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1168 
1169 	vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1170 	if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1171 		goto unlock;
1172 
1173 	entry = pud_mkyoung(orig_pud);
1174 	if (write)
1175 		entry = pud_mkdirty(entry);
1176 	haddr = vmf->address & HPAGE_PUD_MASK;
1177 	if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1178 		update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1179 
1180 unlock:
1181 	spin_unlock(vmf->ptl);
1182 }
1183 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1184 
1185 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1186 {
1187 	pmd_t entry;
1188 	unsigned long haddr;
1189 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1190 
1191 	vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1192 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1193 		goto unlock;
1194 
1195 	entry = pmd_mkyoung(orig_pmd);
1196 	if (write)
1197 		entry = pmd_mkdirty(entry);
1198 	haddr = vmf->address & HPAGE_PMD_MASK;
1199 	if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1200 		update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1201 
1202 unlock:
1203 	spin_unlock(vmf->ptl);
1204 }
1205 
1206 static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1207 			pmd_t orig_pmd, struct page *page)
1208 {
1209 	struct vm_area_struct *vma = vmf->vma;
1210 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1211 	struct mem_cgroup *memcg;
1212 	pgtable_t pgtable;
1213 	pmd_t _pmd;
1214 	int i;
1215 	vm_fault_t ret = 0;
1216 	struct page **pages;
1217 	struct mmu_notifier_range range;
1218 
1219 	pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1220 			      GFP_KERNEL);
1221 	if (unlikely(!pages)) {
1222 		ret |= VM_FAULT_OOM;
1223 		goto out;
1224 	}
1225 
1226 	for (i = 0; i < HPAGE_PMD_NR; i++) {
1227 		pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1228 					       vmf->address, page_to_nid(page));
1229 		if (unlikely(!pages[i] ||
1230 			     mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
1231 				     GFP_KERNEL, &memcg, false))) {
1232 			if (pages[i])
1233 				put_page(pages[i]);
1234 			while (--i >= 0) {
1235 				memcg = (void *)page_private(pages[i]);
1236 				set_page_private(pages[i], 0);
1237 				mem_cgroup_cancel_charge(pages[i], memcg,
1238 						false);
1239 				put_page(pages[i]);
1240 			}
1241 			kfree(pages);
1242 			ret |= VM_FAULT_OOM;
1243 			goto out;
1244 		}
1245 		set_page_private(pages[i], (unsigned long)memcg);
1246 	}
1247 
1248 	for (i = 0; i < HPAGE_PMD_NR; i++) {
1249 		copy_user_highpage(pages[i], page + i,
1250 				   haddr + PAGE_SIZE * i, vma);
1251 		__SetPageUptodate(pages[i]);
1252 		cond_resched();
1253 	}
1254 
1255 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1256 				haddr, haddr + HPAGE_PMD_SIZE);
1257 	mmu_notifier_invalidate_range_start(&range);
1258 
1259 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1260 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1261 		goto out_free_pages;
1262 	VM_BUG_ON_PAGE(!PageHead(page), page);
1263 
1264 	/*
1265 	 * Leave pmd empty until pte is filled note we must notify here as
1266 	 * concurrent CPU thread might write to new page before the call to
1267 	 * mmu_notifier_invalidate_range_end() happens which can lead to a
1268 	 * device seeing memory write in different order than CPU.
1269 	 *
1270 	 * See Documentation/vm/mmu_notifier.rst
1271 	 */
1272 	pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1273 
1274 	pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1275 	pmd_populate(vma->vm_mm, &_pmd, pgtable);
1276 
1277 	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1278 		pte_t entry;
1279 		entry = mk_pte(pages[i], vma->vm_page_prot);
1280 		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1281 		memcg = (void *)page_private(pages[i]);
1282 		set_page_private(pages[i], 0);
1283 		page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1284 		mem_cgroup_commit_charge(pages[i], memcg, false, false);
1285 		lru_cache_add_active_or_unevictable(pages[i], vma);
1286 		vmf->pte = pte_offset_map(&_pmd, haddr);
1287 		VM_BUG_ON(!pte_none(*vmf->pte));
1288 		set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1289 		pte_unmap(vmf->pte);
1290 	}
1291 	kfree(pages);
1292 
1293 	smp_wmb(); /* make pte visible before pmd */
1294 	pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1295 	page_remove_rmap(page, true);
1296 	spin_unlock(vmf->ptl);
1297 
1298 	/*
1299 	 * No need to double call mmu_notifier->invalidate_range() callback as
1300 	 * the above pmdp_huge_clear_flush_notify() did already call it.
1301 	 */
1302 	mmu_notifier_invalidate_range_only_end(&range);
1303 
1304 	ret |= VM_FAULT_WRITE;
1305 	put_page(page);
1306 
1307 out:
1308 	return ret;
1309 
1310 out_free_pages:
1311 	spin_unlock(vmf->ptl);
1312 	mmu_notifier_invalidate_range_end(&range);
1313 	for (i = 0; i < HPAGE_PMD_NR; i++) {
1314 		memcg = (void *)page_private(pages[i]);
1315 		set_page_private(pages[i], 0);
1316 		mem_cgroup_cancel_charge(pages[i], memcg, false);
1317 		put_page(pages[i]);
1318 	}
1319 	kfree(pages);
1320 	goto out;
1321 }
1322 
1323 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1324 {
1325 	struct vm_area_struct *vma = vmf->vma;
1326 	struct page *page = NULL, *new_page;
1327 	struct mem_cgroup *memcg;
1328 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1329 	struct mmu_notifier_range range;
1330 	gfp_t huge_gfp;			/* for allocation and charge */
1331 	vm_fault_t ret = 0;
1332 
1333 	vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1334 	VM_BUG_ON_VMA(!vma->anon_vma, vma);
1335 	if (is_huge_zero_pmd(orig_pmd))
1336 		goto alloc;
1337 	spin_lock(vmf->ptl);
1338 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1339 		goto out_unlock;
1340 
1341 	page = pmd_page(orig_pmd);
1342 	VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1343 	/*
1344 	 * We can only reuse the page if nobody else maps the huge page or it's
1345 	 * part.
1346 	 */
1347 	if (!trylock_page(page)) {
1348 		get_page(page);
1349 		spin_unlock(vmf->ptl);
1350 		lock_page(page);
1351 		spin_lock(vmf->ptl);
1352 		if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1353 			unlock_page(page);
1354 			put_page(page);
1355 			goto out_unlock;
1356 		}
1357 		put_page(page);
1358 	}
1359 	if (reuse_swap_page(page, NULL)) {
1360 		pmd_t entry;
1361 		entry = pmd_mkyoung(orig_pmd);
1362 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1363 		if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1364 			update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1365 		ret |= VM_FAULT_WRITE;
1366 		unlock_page(page);
1367 		goto out_unlock;
1368 	}
1369 	unlock_page(page);
1370 	get_page(page);
1371 	spin_unlock(vmf->ptl);
1372 alloc:
1373 	if (__transparent_hugepage_enabled(vma) &&
1374 	    !transparent_hugepage_debug_cow()) {
1375 		huge_gfp = alloc_hugepage_direct_gfpmask(vma, haddr);
1376 		new_page = alloc_pages_vma(huge_gfp, HPAGE_PMD_ORDER, vma,
1377 				haddr, numa_node_id());
1378 	} else
1379 		new_page = NULL;
1380 
1381 	if (likely(new_page)) {
1382 		prep_transhuge_page(new_page);
1383 	} else {
1384 		if (!page) {
1385 			split_huge_pmd(vma, vmf->pmd, vmf->address);
1386 			ret |= VM_FAULT_FALLBACK;
1387 		} else {
1388 			ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1389 			if (ret & VM_FAULT_OOM) {
1390 				split_huge_pmd(vma, vmf->pmd, vmf->address);
1391 				ret |= VM_FAULT_FALLBACK;
1392 			}
1393 			put_page(page);
1394 		}
1395 		count_vm_event(THP_FAULT_FALLBACK);
1396 		goto out;
1397 	}
1398 
1399 	if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
1400 					huge_gfp, &memcg, true))) {
1401 		put_page(new_page);
1402 		split_huge_pmd(vma, vmf->pmd, vmf->address);
1403 		if (page)
1404 			put_page(page);
1405 		ret |= VM_FAULT_FALLBACK;
1406 		count_vm_event(THP_FAULT_FALLBACK);
1407 		goto out;
1408 	}
1409 
1410 	count_vm_event(THP_FAULT_ALLOC);
1411 	count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
1412 
1413 	if (!page)
1414 		clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1415 	else
1416 		copy_user_huge_page(new_page, page, vmf->address,
1417 				    vma, HPAGE_PMD_NR);
1418 	__SetPageUptodate(new_page);
1419 
1420 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1421 				haddr, haddr + HPAGE_PMD_SIZE);
1422 	mmu_notifier_invalidate_range_start(&range);
1423 
1424 	spin_lock(vmf->ptl);
1425 	if (page)
1426 		put_page(page);
1427 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1428 		spin_unlock(vmf->ptl);
1429 		mem_cgroup_cancel_charge(new_page, memcg, true);
1430 		put_page(new_page);
1431 		goto out_mn;
1432 	} else {
1433 		pmd_t entry;
1434 		entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1435 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1436 		pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1437 		page_add_new_anon_rmap(new_page, vma, haddr, true);
1438 		mem_cgroup_commit_charge(new_page, memcg, false, true);
1439 		lru_cache_add_active_or_unevictable(new_page, vma);
1440 		set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1441 		update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1442 		if (!page) {
1443 			add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1444 		} else {
1445 			VM_BUG_ON_PAGE(!PageHead(page), page);
1446 			page_remove_rmap(page, true);
1447 			put_page(page);
1448 		}
1449 		ret |= VM_FAULT_WRITE;
1450 	}
1451 	spin_unlock(vmf->ptl);
1452 out_mn:
1453 	/*
1454 	 * No need to double call mmu_notifier->invalidate_range() callback as
1455 	 * the above pmdp_huge_clear_flush_notify() did already call it.
1456 	 */
1457 	mmu_notifier_invalidate_range_only_end(&range);
1458 out:
1459 	return ret;
1460 out_unlock:
1461 	spin_unlock(vmf->ptl);
1462 	return ret;
1463 }
1464 
1465 /*
1466  * FOLL_FORCE can write to even unwritable pmd's, but only
1467  * after we've gone through a COW cycle and they are dirty.
1468  */
1469 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1470 {
1471 	return pmd_write(pmd) ||
1472 	       ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1473 }
1474 
1475 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1476 				   unsigned long addr,
1477 				   pmd_t *pmd,
1478 				   unsigned int flags)
1479 {
1480 	struct mm_struct *mm = vma->vm_mm;
1481 	struct page *page = NULL;
1482 
1483 	assert_spin_locked(pmd_lockptr(mm, pmd));
1484 
1485 	if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1486 		goto out;
1487 
1488 	/* Avoid dumping huge zero page */
1489 	if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1490 		return ERR_PTR(-EFAULT);
1491 
1492 	/* Full NUMA hinting faults to serialise migration in fault paths */
1493 	if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1494 		goto out;
1495 
1496 	page = pmd_page(*pmd);
1497 	VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1498 	if (flags & FOLL_TOUCH)
1499 		touch_pmd(vma, addr, pmd, flags);
1500 	if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1501 		/*
1502 		 * We don't mlock() pte-mapped THPs. This way we can avoid
1503 		 * leaking mlocked pages into non-VM_LOCKED VMAs.
1504 		 *
1505 		 * For anon THP:
1506 		 *
1507 		 * In most cases the pmd is the only mapping of the page as we
1508 		 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1509 		 * writable private mappings in populate_vma_page_range().
1510 		 *
1511 		 * The only scenario when we have the page shared here is if we
1512 		 * mlocking read-only mapping shared over fork(). We skip
1513 		 * mlocking such pages.
1514 		 *
1515 		 * For file THP:
1516 		 *
1517 		 * We can expect PageDoubleMap() to be stable under page lock:
1518 		 * for file pages we set it in page_add_file_rmap(), which
1519 		 * requires page to be locked.
1520 		 */
1521 
1522 		if (PageAnon(page) && compound_mapcount(page) != 1)
1523 			goto skip_mlock;
1524 		if (PageDoubleMap(page) || !page->mapping)
1525 			goto skip_mlock;
1526 		if (!trylock_page(page))
1527 			goto skip_mlock;
1528 		lru_add_drain();
1529 		if (page->mapping && !PageDoubleMap(page))
1530 			mlock_vma_page(page);
1531 		unlock_page(page);
1532 	}
1533 skip_mlock:
1534 	page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1535 	VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1536 	if (flags & FOLL_GET)
1537 		get_page(page);
1538 
1539 out:
1540 	return page;
1541 }
1542 
1543 /* NUMA hinting page fault entry point for trans huge pmds */
1544 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1545 {
1546 	struct vm_area_struct *vma = vmf->vma;
1547 	struct anon_vma *anon_vma = NULL;
1548 	struct page *page;
1549 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1550 	int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1551 	int target_nid, last_cpupid = -1;
1552 	bool page_locked;
1553 	bool migrated = false;
1554 	bool was_writable;
1555 	int flags = 0;
1556 
1557 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1558 	if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1559 		goto out_unlock;
1560 
1561 	/*
1562 	 * If there are potential migrations, wait for completion and retry
1563 	 * without disrupting NUMA hinting information. Do not relock and
1564 	 * check_same as the page may no longer be mapped.
1565 	 */
1566 	if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1567 		page = pmd_page(*vmf->pmd);
1568 		if (!get_page_unless_zero(page))
1569 			goto out_unlock;
1570 		spin_unlock(vmf->ptl);
1571 		put_and_wait_on_page_locked(page);
1572 		goto out;
1573 	}
1574 
1575 	page = pmd_page(pmd);
1576 	BUG_ON(is_huge_zero_page(page));
1577 	page_nid = page_to_nid(page);
1578 	last_cpupid = page_cpupid_last(page);
1579 	count_vm_numa_event(NUMA_HINT_FAULTS);
1580 	if (page_nid == this_nid) {
1581 		count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1582 		flags |= TNF_FAULT_LOCAL;
1583 	}
1584 
1585 	/* See similar comment in do_numa_page for explanation */
1586 	if (!pmd_savedwrite(pmd))
1587 		flags |= TNF_NO_GROUP;
1588 
1589 	/*
1590 	 * Acquire the page lock to serialise THP migrations but avoid dropping
1591 	 * page_table_lock if at all possible
1592 	 */
1593 	page_locked = trylock_page(page);
1594 	target_nid = mpol_misplaced(page, vma, haddr);
1595 	if (target_nid == NUMA_NO_NODE) {
1596 		/* If the page was locked, there are no parallel migrations */
1597 		if (page_locked)
1598 			goto clear_pmdnuma;
1599 	}
1600 
1601 	/* Migration could have started since the pmd_trans_migrating check */
1602 	if (!page_locked) {
1603 		page_nid = NUMA_NO_NODE;
1604 		if (!get_page_unless_zero(page))
1605 			goto out_unlock;
1606 		spin_unlock(vmf->ptl);
1607 		put_and_wait_on_page_locked(page);
1608 		goto out;
1609 	}
1610 
1611 	/*
1612 	 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1613 	 * to serialises splits
1614 	 */
1615 	get_page(page);
1616 	spin_unlock(vmf->ptl);
1617 	anon_vma = page_lock_anon_vma_read(page);
1618 
1619 	/* Confirm the PMD did not change while page_table_lock was released */
1620 	spin_lock(vmf->ptl);
1621 	if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1622 		unlock_page(page);
1623 		put_page(page);
1624 		page_nid = NUMA_NO_NODE;
1625 		goto out_unlock;
1626 	}
1627 
1628 	/* Bail if we fail to protect against THP splits for any reason */
1629 	if (unlikely(!anon_vma)) {
1630 		put_page(page);
1631 		page_nid = NUMA_NO_NODE;
1632 		goto clear_pmdnuma;
1633 	}
1634 
1635 	/*
1636 	 * Since we took the NUMA fault, we must have observed the !accessible
1637 	 * bit. Make sure all other CPUs agree with that, to avoid them
1638 	 * modifying the page we're about to migrate.
1639 	 *
1640 	 * Must be done under PTL such that we'll observe the relevant
1641 	 * inc_tlb_flush_pending().
1642 	 *
1643 	 * We are not sure a pending tlb flush here is for a huge page
1644 	 * mapping or not. Hence use the tlb range variant
1645 	 */
1646 	if (mm_tlb_flush_pending(vma->vm_mm)) {
1647 		flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1648 		/*
1649 		 * change_huge_pmd() released the pmd lock before
1650 		 * invalidating the secondary MMUs sharing the primary
1651 		 * MMU pagetables (with ->invalidate_range()). The
1652 		 * mmu_notifier_invalidate_range_end() (which
1653 		 * internally calls ->invalidate_range()) in
1654 		 * change_pmd_range() will run after us, so we can't
1655 		 * rely on it here and we need an explicit invalidate.
1656 		 */
1657 		mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1658 					      haddr + HPAGE_PMD_SIZE);
1659 	}
1660 
1661 	/*
1662 	 * Migrate the THP to the requested node, returns with page unlocked
1663 	 * and access rights restored.
1664 	 */
1665 	spin_unlock(vmf->ptl);
1666 
1667 	migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1668 				vmf->pmd, pmd, vmf->address, page, target_nid);
1669 	if (migrated) {
1670 		flags |= TNF_MIGRATED;
1671 		page_nid = target_nid;
1672 	} else
1673 		flags |= TNF_MIGRATE_FAIL;
1674 
1675 	goto out;
1676 clear_pmdnuma:
1677 	BUG_ON(!PageLocked(page));
1678 	was_writable = pmd_savedwrite(pmd);
1679 	pmd = pmd_modify(pmd, vma->vm_page_prot);
1680 	pmd = pmd_mkyoung(pmd);
1681 	if (was_writable)
1682 		pmd = pmd_mkwrite(pmd);
1683 	set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1684 	update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1685 	unlock_page(page);
1686 out_unlock:
1687 	spin_unlock(vmf->ptl);
1688 
1689 out:
1690 	if (anon_vma)
1691 		page_unlock_anon_vma_read(anon_vma);
1692 
1693 	if (page_nid != NUMA_NO_NODE)
1694 		task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1695 				flags);
1696 
1697 	return 0;
1698 }
1699 
1700 /*
1701  * Return true if we do MADV_FREE successfully on entire pmd page.
1702  * Otherwise, return false.
1703  */
1704 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1705 		pmd_t *pmd, unsigned long addr, unsigned long next)
1706 {
1707 	spinlock_t *ptl;
1708 	pmd_t orig_pmd;
1709 	struct page *page;
1710 	struct mm_struct *mm = tlb->mm;
1711 	bool ret = false;
1712 
1713 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1714 
1715 	ptl = pmd_trans_huge_lock(pmd, vma);
1716 	if (!ptl)
1717 		goto out_unlocked;
1718 
1719 	orig_pmd = *pmd;
1720 	if (is_huge_zero_pmd(orig_pmd))
1721 		goto out;
1722 
1723 	if (unlikely(!pmd_present(orig_pmd))) {
1724 		VM_BUG_ON(thp_migration_supported() &&
1725 				  !is_pmd_migration_entry(orig_pmd));
1726 		goto out;
1727 	}
1728 
1729 	page = pmd_page(orig_pmd);
1730 	/*
1731 	 * If other processes are mapping this page, we couldn't discard
1732 	 * the page unless they all do MADV_FREE so let's skip the page.
1733 	 */
1734 	if (page_mapcount(page) != 1)
1735 		goto out;
1736 
1737 	if (!trylock_page(page))
1738 		goto out;
1739 
1740 	/*
1741 	 * If user want to discard part-pages of THP, split it so MADV_FREE
1742 	 * will deactivate only them.
1743 	 */
1744 	if (next - addr != HPAGE_PMD_SIZE) {
1745 		get_page(page);
1746 		spin_unlock(ptl);
1747 		split_huge_page(page);
1748 		unlock_page(page);
1749 		put_page(page);
1750 		goto out_unlocked;
1751 	}
1752 
1753 	if (PageDirty(page))
1754 		ClearPageDirty(page);
1755 	unlock_page(page);
1756 
1757 	if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1758 		pmdp_invalidate(vma, addr, pmd);
1759 		orig_pmd = pmd_mkold(orig_pmd);
1760 		orig_pmd = pmd_mkclean(orig_pmd);
1761 
1762 		set_pmd_at(mm, addr, pmd, orig_pmd);
1763 		tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1764 	}
1765 
1766 	mark_page_lazyfree(page);
1767 	ret = true;
1768 out:
1769 	spin_unlock(ptl);
1770 out_unlocked:
1771 	return ret;
1772 }
1773 
1774 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1775 {
1776 	pgtable_t pgtable;
1777 
1778 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1779 	pte_free(mm, pgtable);
1780 	mm_dec_nr_ptes(mm);
1781 }
1782 
1783 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1784 		 pmd_t *pmd, unsigned long addr)
1785 {
1786 	pmd_t orig_pmd;
1787 	spinlock_t *ptl;
1788 
1789 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1790 
1791 	ptl = __pmd_trans_huge_lock(pmd, vma);
1792 	if (!ptl)
1793 		return 0;
1794 	/*
1795 	 * For architectures like ppc64 we look at deposited pgtable
1796 	 * when calling pmdp_huge_get_and_clear. So do the
1797 	 * pgtable_trans_huge_withdraw after finishing pmdp related
1798 	 * operations.
1799 	 */
1800 	orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1801 			tlb->fullmm);
1802 	tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1803 	if (vma_is_dax(vma)) {
1804 		if (arch_needs_pgtable_deposit())
1805 			zap_deposited_table(tlb->mm, pmd);
1806 		spin_unlock(ptl);
1807 		if (is_huge_zero_pmd(orig_pmd))
1808 			tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1809 	} else if (is_huge_zero_pmd(orig_pmd)) {
1810 		zap_deposited_table(tlb->mm, pmd);
1811 		spin_unlock(ptl);
1812 		tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1813 	} else {
1814 		struct page *page = NULL;
1815 		int flush_needed = 1;
1816 
1817 		if (pmd_present(orig_pmd)) {
1818 			page = pmd_page(orig_pmd);
1819 			page_remove_rmap(page, true);
1820 			VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1821 			VM_BUG_ON_PAGE(!PageHead(page), page);
1822 		} else if (thp_migration_supported()) {
1823 			swp_entry_t entry;
1824 
1825 			VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1826 			entry = pmd_to_swp_entry(orig_pmd);
1827 			page = pfn_to_page(swp_offset(entry));
1828 			flush_needed = 0;
1829 		} else
1830 			WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1831 
1832 		if (PageAnon(page)) {
1833 			zap_deposited_table(tlb->mm, pmd);
1834 			add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1835 		} else {
1836 			if (arch_needs_pgtable_deposit())
1837 				zap_deposited_table(tlb->mm, pmd);
1838 			add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1839 		}
1840 
1841 		spin_unlock(ptl);
1842 		if (flush_needed)
1843 			tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1844 	}
1845 	return 1;
1846 }
1847 
1848 #ifndef pmd_move_must_withdraw
1849 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1850 					 spinlock_t *old_pmd_ptl,
1851 					 struct vm_area_struct *vma)
1852 {
1853 	/*
1854 	 * With split pmd lock we also need to move preallocated
1855 	 * PTE page table if new_pmd is on different PMD page table.
1856 	 *
1857 	 * We also don't deposit and withdraw tables for file pages.
1858 	 */
1859 	return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1860 }
1861 #endif
1862 
1863 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1864 {
1865 #ifdef CONFIG_MEM_SOFT_DIRTY
1866 	if (unlikely(is_pmd_migration_entry(pmd)))
1867 		pmd = pmd_swp_mksoft_dirty(pmd);
1868 	else if (pmd_present(pmd))
1869 		pmd = pmd_mksoft_dirty(pmd);
1870 #endif
1871 	return pmd;
1872 }
1873 
1874 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1875 		  unsigned long new_addr, unsigned long old_end,
1876 		  pmd_t *old_pmd, pmd_t *new_pmd)
1877 {
1878 	spinlock_t *old_ptl, *new_ptl;
1879 	pmd_t pmd;
1880 	struct mm_struct *mm = vma->vm_mm;
1881 	bool force_flush = false;
1882 
1883 	if ((old_addr & ~HPAGE_PMD_MASK) ||
1884 	    (new_addr & ~HPAGE_PMD_MASK) ||
1885 	    old_end - old_addr < HPAGE_PMD_SIZE)
1886 		return false;
1887 
1888 	/*
1889 	 * The destination pmd shouldn't be established, free_pgtables()
1890 	 * should have release it.
1891 	 */
1892 	if (WARN_ON(!pmd_none(*new_pmd))) {
1893 		VM_BUG_ON(pmd_trans_huge(*new_pmd));
1894 		return false;
1895 	}
1896 
1897 	/*
1898 	 * We don't have to worry about the ordering of src and dst
1899 	 * ptlocks because exclusive mmap_sem prevents deadlock.
1900 	 */
1901 	old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1902 	if (old_ptl) {
1903 		new_ptl = pmd_lockptr(mm, new_pmd);
1904 		if (new_ptl != old_ptl)
1905 			spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1906 		pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1907 		if (pmd_present(pmd))
1908 			force_flush = true;
1909 		VM_BUG_ON(!pmd_none(*new_pmd));
1910 
1911 		if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1912 			pgtable_t pgtable;
1913 			pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1914 			pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1915 		}
1916 		pmd = move_soft_dirty_pmd(pmd);
1917 		set_pmd_at(mm, new_addr, new_pmd, pmd);
1918 		if (force_flush)
1919 			flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1920 		if (new_ptl != old_ptl)
1921 			spin_unlock(new_ptl);
1922 		spin_unlock(old_ptl);
1923 		return true;
1924 	}
1925 	return false;
1926 }
1927 
1928 /*
1929  * Returns
1930  *  - 0 if PMD could not be locked
1931  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1932  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1933  */
1934 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1935 		unsigned long addr, pgprot_t newprot, int prot_numa)
1936 {
1937 	struct mm_struct *mm = vma->vm_mm;
1938 	spinlock_t *ptl;
1939 	pmd_t entry;
1940 	bool preserve_write;
1941 	int ret;
1942 
1943 	ptl = __pmd_trans_huge_lock(pmd, vma);
1944 	if (!ptl)
1945 		return 0;
1946 
1947 	preserve_write = prot_numa && pmd_write(*pmd);
1948 	ret = 1;
1949 
1950 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1951 	if (is_swap_pmd(*pmd)) {
1952 		swp_entry_t entry = pmd_to_swp_entry(*pmd);
1953 
1954 		VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1955 		if (is_write_migration_entry(entry)) {
1956 			pmd_t newpmd;
1957 			/*
1958 			 * A protection check is difficult so
1959 			 * just be safe and disable write
1960 			 */
1961 			make_migration_entry_read(&entry);
1962 			newpmd = swp_entry_to_pmd(entry);
1963 			if (pmd_swp_soft_dirty(*pmd))
1964 				newpmd = pmd_swp_mksoft_dirty(newpmd);
1965 			set_pmd_at(mm, addr, pmd, newpmd);
1966 		}
1967 		goto unlock;
1968 	}
1969 #endif
1970 
1971 	/*
1972 	 * Avoid trapping faults against the zero page. The read-only
1973 	 * data is likely to be read-cached on the local CPU and
1974 	 * local/remote hits to the zero page are not interesting.
1975 	 */
1976 	if (prot_numa && is_huge_zero_pmd(*pmd))
1977 		goto unlock;
1978 
1979 	if (prot_numa && pmd_protnone(*pmd))
1980 		goto unlock;
1981 
1982 	/*
1983 	 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1984 	 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1985 	 * which is also under down_read(mmap_sem):
1986 	 *
1987 	 *	CPU0:				CPU1:
1988 	 *				change_huge_pmd(prot_numa=1)
1989 	 *				 pmdp_huge_get_and_clear_notify()
1990 	 * madvise_dontneed()
1991 	 *  zap_pmd_range()
1992 	 *   pmd_trans_huge(*pmd) == 0 (without ptl)
1993 	 *   // skip the pmd
1994 	 *				 set_pmd_at();
1995 	 *				 // pmd is re-established
1996 	 *
1997 	 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1998 	 * which may break userspace.
1999 	 *
2000 	 * pmdp_invalidate() is required to make sure we don't miss
2001 	 * dirty/young flags set by hardware.
2002 	 */
2003 	entry = pmdp_invalidate(vma, addr, pmd);
2004 
2005 	entry = pmd_modify(entry, newprot);
2006 	if (preserve_write)
2007 		entry = pmd_mk_savedwrite(entry);
2008 	ret = HPAGE_PMD_NR;
2009 	set_pmd_at(mm, addr, pmd, entry);
2010 	BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
2011 unlock:
2012 	spin_unlock(ptl);
2013 	return ret;
2014 }
2015 
2016 /*
2017  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
2018  *
2019  * Note that if it returns page table lock pointer, this routine returns without
2020  * unlocking page table lock. So callers must unlock it.
2021  */
2022 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
2023 {
2024 	spinlock_t *ptl;
2025 	ptl = pmd_lock(vma->vm_mm, pmd);
2026 	if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
2027 			pmd_devmap(*pmd)))
2028 		return ptl;
2029 	spin_unlock(ptl);
2030 	return NULL;
2031 }
2032 
2033 /*
2034  * Returns true if a given pud maps a thp, false otherwise.
2035  *
2036  * Note that if it returns true, this routine returns without unlocking page
2037  * table lock. So callers must unlock it.
2038  */
2039 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2040 {
2041 	spinlock_t *ptl;
2042 
2043 	ptl = pud_lock(vma->vm_mm, pud);
2044 	if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2045 		return ptl;
2046 	spin_unlock(ptl);
2047 	return NULL;
2048 }
2049 
2050 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2051 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2052 		 pud_t *pud, unsigned long addr)
2053 {
2054 	spinlock_t *ptl;
2055 
2056 	ptl = __pud_trans_huge_lock(pud, vma);
2057 	if (!ptl)
2058 		return 0;
2059 	/*
2060 	 * For architectures like ppc64 we look at deposited pgtable
2061 	 * when calling pudp_huge_get_and_clear. So do the
2062 	 * pgtable_trans_huge_withdraw after finishing pudp related
2063 	 * operations.
2064 	 */
2065 	pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
2066 	tlb_remove_pud_tlb_entry(tlb, pud, addr);
2067 	if (vma_is_dax(vma)) {
2068 		spin_unlock(ptl);
2069 		/* No zero page support yet */
2070 	} else {
2071 		/* No support for anonymous PUD pages yet */
2072 		BUG();
2073 	}
2074 	return 1;
2075 }
2076 
2077 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2078 		unsigned long haddr)
2079 {
2080 	VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2081 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2082 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2083 	VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2084 
2085 	count_vm_event(THP_SPLIT_PUD);
2086 
2087 	pudp_huge_clear_flush_notify(vma, haddr, pud);
2088 }
2089 
2090 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2091 		unsigned long address)
2092 {
2093 	spinlock_t *ptl;
2094 	struct mmu_notifier_range range;
2095 
2096 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2097 				address & HPAGE_PUD_MASK,
2098 				(address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2099 	mmu_notifier_invalidate_range_start(&range);
2100 	ptl = pud_lock(vma->vm_mm, pud);
2101 	if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2102 		goto out;
2103 	__split_huge_pud_locked(vma, pud, range.start);
2104 
2105 out:
2106 	spin_unlock(ptl);
2107 	/*
2108 	 * No need to double call mmu_notifier->invalidate_range() callback as
2109 	 * the above pudp_huge_clear_flush_notify() did already call it.
2110 	 */
2111 	mmu_notifier_invalidate_range_only_end(&range);
2112 }
2113 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2114 
2115 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2116 		unsigned long haddr, pmd_t *pmd)
2117 {
2118 	struct mm_struct *mm = vma->vm_mm;
2119 	pgtable_t pgtable;
2120 	pmd_t _pmd;
2121 	int i;
2122 
2123 	/*
2124 	 * Leave pmd empty until pte is filled note that it is fine to delay
2125 	 * notification until mmu_notifier_invalidate_range_end() as we are
2126 	 * replacing a zero pmd write protected page with a zero pte write
2127 	 * protected page.
2128 	 *
2129 	 * See Documentation/vm/mmu_notifier.rst
2130 	 */
2131 	pmdp_huge_clear_flush(vma, haddr, pmd);
2132 
2133 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2134 	pmd_populate(mm, &_pmd, pgtable);
2135 
2136 	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2137 		pte_t *pte, entry;
2138 		entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2139 		entry = pte_mkspecial(entry);
2140 		pte = pte_offset_map(&_pmd, haddr);
2141 		VM_BUG_ON(!pte_none(*pte));
2142 		set_pte_at(mm, haddr, pte, entry);
2143 		pte_unmap(pte);
2144 	}
2145 	smp_wmb(); /* make pte visible before pmd */
2146 	pmd_populate(mm, pmd, pgtable);
2147 }
2148 
2149 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2150 		unsigned long haddr, bool freeze)
2151 {
2152 	struct mm_struct *mm = vma->vm_mm;
2153 	struct page *page;
2154 	pgtable_t pgtable;
2155 	pmd_t old_pmd, _pmd;
2156 	bool young, write, soft_dirty, pmd_migration = false;
2157 	unsigned long addr;
2158 	int i;
2159 
2160 	VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2161 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2162 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2163 	VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2164 				&& !pmd_devmap(*pmd));
2165 
2166 	count_vm_event(THP_SPLIT_PMD);
2167 
2168 	if (!vma_is_anonymous(vma)) {
2169 		_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2170 		/*
2171 		 * We are going to unmap this huge page. So
2172 		 * just go ahead and zap it
2173 		 */
2174 		if (arch_needs_pgtable_deposit())
2175 			zap_deposited_table(mm, pmd);
2176 		if (vma_is_dax(vma))
2177 			return;
2178 		page = pmd_page(_pmd);
2179 		if (!PageDirty(page) && pmd_dirty(_pmd))
2180 			set_page_dirty(page);
2181 		if (!PageReferenced(page) && pmd_young(_pmd))
2182 			SetPageReferenced(page);
2183 		page_remove_rmap(page, true);
2184 		put_page(page);
2185 		add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2186 		return;
2187 	} else if (is_huge_zero_pmd(*pmd)) {
2188 		/*
2189 		 * FIXME: Do we want to invalidate secondary mmu by calling
2190 		 * mmu_notifier_invalidate_range() see comments below inside
2191 		 * __split_huge_pmd() ?
2192 		 *
2193 		 * We are going from a zero huge page write protected to zero
2194 		 * small page also write protected so it does not seems useful
2195 		 * to invalidate secondary mmu at this time.
2196 		 */
2197 		return __split_huge_zero_page_pmd(vma, haddr, pmd);
2198 	}
2199 
2200 	/*
2201 	 * Up to this point the pmd is present and huge and userland has the
2202 	 * whole access to the hugepage during the split (which happens in
2203 	 * place). If we overwrite the pmd with the not-huge version pointing
2204 	 * to the pte here (which of course we could if all CPUs were bug
2205 	 * free), userland could trigger a small page size TLB miss on the
2206 	 * small sized TLB while the hugepage TLB entry is still established in
2207 	 * the huge TLB. Some CPU doesn't like that.
2208 	 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2209 	 * 383 on page 93. Intel should be safe but is also warns that it's
2210 	 * only safe if the permission and cache attributes of the two entries
2211 	 * loaded in the two TLB is identical (which should be the case here).
2212 	 * But it is generally safer to never allow small and huge TLB entries
2213 	 * for the same virtual address to be loaded simultaneously. So instead
2214 	 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2215 	 * current pmd notpresent (atomically because here the pmd_trans_huge
2216 	 * must remain set at all times on the pmd until the split is complete
2217 	 * for this pmd), then we flush the SMP TLB and finally we write the
2218 	 * non-huge version of the pmd entry with pmd_populate.
2219 	 */
2220 	old_pmd = pmdp_invalidate(vma, haddr, pmd);
2221 
2222 	pmd_migration = is_pmd_migration_entry(old_pmd);
2223 	if (unlikely(pmd_migration)) {
2224 		swp_entry_t entry;
2225 
2226 		entry = pmd_to_swp_entry(old_pmd);
2227 		page = pfn_to_page(swp_offset(entry));
2228 		write = is_write_migration_entry(entry);
2229 		young = false;
2230 		soft_dirty = pmd_swp_soft_dirty(old_pmd);
2231 	} else {
2232 		page = pmd_page(old_pmd);
2233 		if (pmd_dirty(old_pmd))
2234 			SetPageDirty(page);
2235 		write = pmd_write(old_pmd);
2236 		young = pmd_young(old_pmd);
2237 		soft_dirty = pmd_soft_dirty(old_pmd);
2238 	}
2239 	VM_BUG_ON_PAGE(!page_count(page), page);
2240 	page_ref_add(page, HPAGE_PMD_NR - 1);
2241 
2242 	/*
2243 	 * Withdraw the table only after we mark the pmd entry invalid.
2244 	 * This's critical for some architectures (Power).
2245 	 */
2246 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2247 	pmd_populate(mm, &_pmd, pgtable);
2248 
2249 	for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2250 		pte_t entry, *pte;
2251 		/*
2252 		 * Note that NUMA hinting access restrictions are not
2253 		 * transferred to avoid any possibility of altering
2254 		 * permissions across VMAs.
2255 		 */
2256 		if (freeze || pmd_migration) {
2257 			swp_entry_t swp_entry;
2258 			swp_entry = make_migration_entry(page + i, write);
2259 			entry = swp_entry_to_pte(swp_entry);
2260 			if (soft_dirty)
2261 				entry = pte_swp_mksoft_dirty(entry);
2262 		} else {
2263 			entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2264 			entry = maybe_mkwrite(entry, vma);
2265 			if (!write)
2266 				entry = pte_wrprotect(entry);
2267 			if (!young)
2268 				entry = pte_mkold(entry);
2269 			if (soft_dirty)
2270 				entry = pte_mksoft_dirty(entry);
2271 		}
2272 		pte = pte_offset_map(&_pmd, addr);
2273 		BUG_ON(!pte_none(*pte));
2274 		set_pte_at(mm, addr, pte, entry);
2275 		atomic_inc(&page[i]._mapcount);
2276 		pte_unmap(pte);
2277 	}
2278 
2279 	/*
2280 	 * Set PG_double_map before dropping compound_mapcount to avoid
2281 	 * false-negative page_mapped().
2282 	 */
2283 	if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2284 		for (i = 0; i < HPAGE_PMD_NR; i++)
2285 			atomic_inc(&page[i]._mapcount);
2286 	}
2287 
2288 	if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2289 		/* Last compound_mapcount is gone. */
2290 		__dec_node_page_state(page, NR_ANON_THPS);
2291 		if (TestClearPageDoubleMap(page)) {
2292 			/* No need in mapcount reference anymore */
2293 			for (i = 0; i < HPAGE_PMD_NR; i++)
2294 				atomic_dec(&page[i]._mapcount);
2295 		}
2296 	}
2297 
2298 	smp_wmb(); /* make pte visible before pmd */
2299 	pmd_populate(mm, pmd, pgtable);
2300 
2301 	if (freeze) {
2302 		for (i = 0; i < HPAGE_PMD_NR; i++) {
2303 			page_remove_rmap(page + i, false);
2304 			put_page(page + i);
2305 		}
2306 	}
2307 }
2308 
2309 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2310 		unsigned long address, bool freeze, struct page *page)
2311 {
2312 	spinlock_t *ptl;
2313 	struct mmu_notifier_range range;
2314 
2315 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2316 				address & HPAGE_PMD_MASK,
2317 				(address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2318 	mmu_notifier_invalidate_range_start(&range);
2319 	ptl = pmd_lock(vma->vm_mm, pmd);
2320 
2321 	/*
2322 	 * If caller asks to setup a migration entries, we need a page to check
2323 	 * pmd against. Otherwise we can end up replacing wrong page.
2324 	 */
2325 	VM_BUG_ON(freeze && !page);
2326 	if (page && page != pmd_page(*pmd))
2327 	        goto out;
2328 
2329 	if (pmd_trans_huge(*pmd)) {
2330 		page = pmd_page(*pmd);
2331 		if (PageMlocked(page))
2332 			clear_page_mlock(page);
2333 	} else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2334 		goto out;
2335 	__split_huge_pmd_locked(vma, pmd, range.start, freeze);
2336 out:
2337 	spin_unlock(ptl);
2338 	/*
2339 	 * No need to double call mmu_notifier->invalidate_range() callback.
2340 	 * They are 3 cases to consider inside __split_huge_pmd_locked():
2341 	 *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2342 	 *  2) __split_huge_zero_page_pmd() read only zero page and any write
2343 	 *    fault will trigger a flush_notify before pointing to a new page
2344 	 *    (it is fine if the secondary mmu keeps pointing to the old zero
2345 	 *    page in the meantime)
2346 	 *  3) Split a huge pmd into pte pointing to the same page. No need
2347 	 *     to invalidate secondary tlb entry they are all still valid.
2348 	 *     any further changes to individual pte will notify. So no need
2349 	 *     to call mmu_notifier->invalidate_range()
2350 	 */
2351 	mmu_notifier_invalidate_range_only_end(&range);
2352 }
2353 
2354 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2355 		bool freeze, struct page *page)
2356 {
2357 	pgd_t *pgd;
2358 	p4d_t *p4d;
2359 	pud_t *pud;
2360 	pmd_t *pmd;
2361 
2362 	pgd = pgd_offset(vma->vm_mm, address);
2363 	if (!pgd_present(*pgd))
2364 		return;
2365 
2366 	p4d = p4d_offset(pgd, address);
2367 	if (!p4d_present(*p4d))
2368 		return;
2369 
2370 	pud = pud_offset(p4d, address);
2371 	if (!pud_present(*pud))
2372 		return;
2373 
2374 	pmd = pmd_offset(pud, address);
2375 
2376 	__split_huge_pmd(vma, pmd, address, freeze, page);
2377 }
2378 
2379 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2380 			     unsigned long start,
2381 			     unsigned long end,
2382 			     long adjust_next)
2383 {
2384 	/*
2385 	 * If the new start address isn't hpage aligned and it could
2386 	 * previously contain an hugepage: check if we need to split
2387 	 * an huge pmd.
2388 	 */
2389 	if (start & ~HPAGE_PMD_MASK &&
2390 	    (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2391 	    (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2392 		split_huge_pmd_address(vma, start, false, NULL);
2393 
2394 	/*
2395 	 * If the new end address isn't hpage aligned and it could
2396 	 * previously contain an hugepage: check if we need to split
2397 	 * an huge pmd.
2398 	 */
2399 	if (end & ~HPAGE_PMD_MASK &&
2400 	    (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2401 	    (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2402 		split_huge_pmd_address(vma, end, false, NULL);
2403 
2404 	/*
2405 	 * If we're also updating the vma->vm_next->vm_start, if the new
2406 	 * vm_next->vm_start isn't page aligned and it could previously
2407 	 * contain an hugepage: check if we need to split an huge pmd.
2408 	 */
2409 	if (adjust_next > 0) {
2410 		struct vm_area_struct *next = vma->vm_next;
2411 		unsigned long nstart = next->vm_start;
2412 		nstart += adjust_next << PAGE_SHIFT;
2413 		if (nstart & ~HPAGE_PMD_MASK &&
2414 		    (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2415 		    (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2416 			split_huge_pmd_address(next, nstart, false, NULL);
2417 	}
2418 }
2419 
2420 static void unmap_page(struct page *page)
2421 {
2422 	enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2423 		TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2424 	bool unmap_success;
2425 
2426 	VM_BUG_ON_PAGE(!PageHead(page), page);
2427 
2428 	if (PageAnon(page))
2429 		ttu_flags |= TTU_SPLIT_FREEZE;
2430 
2431 	unmap_success = try_to_unmap(page, ttu_flags);
2432 	VM_BUG_ON_PAGE(!unmap_success, page);
2433 }
2434 
2435 static void remap_page(struct page *page)
2436 {
2437 	int i;
2438 	if (PageTransHuge(page)) {
2439 		remove_migration_ptes(page, page, true);
2440 	} else {
2441 		for (i = 0; i < HPAGE_PMD_NR; i++)
2442 			remove_migration_ptes(page + i, page + i, true);
2443 	}
2444 }
2445 
2446 static void __split_huge_page_tail(struct page *head, int tail,
2447 		struct lruvec *lruvec, struct list_head *list)
2448 {
2449 	struct page *page_tail = head + tail;
2450 
2451 	VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2452 
2453 	/*
2454 	 * Clone page flags before unfreezing refcount.
2455 	 *
2456 	 * After successful get_page_unless_zero() might follow flags change,
2457 	 * for exmaple lock_page() which set PG_waiters.
2458 	 */
2459 	page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2460 	page_tail->flags |= (head->flags &
2461 			((1L << PG_referenced) |
2462 			 (1L << PG_swapbacked) |
2463 			 (1L << PG_swapcache) |
2464 			 (1L << PG_mlocked) |
2465 			 (1L << PG_uptodate) |
2466 			 (1L << PG_active) |
2467 			 (1L << PG_workingset) |
2468 			 (1L << PG_locked) |
2469 			 (1L << PG_unevictable) |
2470 			 (1L << PG_dirty)));
2471 
2472 	/* ->mapping in first tail page is compound_mapcount */
2473 	VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2474 			page_tail);
2475 	page_tail->mapping = head->mapping;
2476 	page_tail->index = head->index + tail;
2477 
2478 	/* Page flags must be visible before we make the page non-compound. */
2479 	smp_wmb();
2480 
2481 	/*
2482 	 * Clear PageTail before unfreezing page refcount.
2483 	 *
2484 	 * After successful get_page_unless_zero() might follow put_page()
2485 	 * which needs correct compound_head().
2486 	 */
2487 	clear_compound_head(page_tail);
2488 
2489 	/* Finally unfreeze refcount. Additional reference from page cache. */
2490 	page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2491 					  PageSwapCache(head)));
2492 
2493 	if (page_is_young(head))
2494 		set_page_young(page_tail);
2495 	if (page_is_idle(head))
2496 		set_page_idle(page_tail);
2497 
2498 	page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2499 
2500 	/*
2501 	 * always add to the tail because some iterators expect new
2502 	 * pages to show after the currently processed elements - e.g.
2503 	 * migrate_pages
2504 	 */
2505 	lru_add_page_tail(head, page_tail, lruvec, list);
2506 }
2507 
2508 static void __split_huge_page(struct page *page, struct list_head *list,
2509 		pgoff_t end, unsigned long flags)
2510 {
2511 	struct page *head = compound_head(page);
2512 	pg_data_t *pgdat = page_pgdat(head);
2513 	struct lruvec *lruvec;
2514 	struct address_space *swap_cache = NULL;
2515 	unsigned long offset = 0;
2516 	int i;
2517 
2518 	lruvec = mem_cgroup_page_lruvec(head, pgdat);
2519 
2520 	/* complete memcg works before add pages to LRU */
2521 	mem_cgroup_split_huge_fixup(head);
2522 
2523 	if (PageAnon(head) && PageSwapCache(head)) {
2524 		swp_entry_t entry = { .val = page_private(head) };
2525 
2526 		offset = swp_offset(entry);
2527 		swap_cache = swap_address_space(entry);
2528 		xa_lock(&swap_cache->i_pages);
2529 	}
2530 
2531 	for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2532 		__split_huge_page_tail(head, i, lruvec, list);
2533 		/* Some pages can be beyond i_size: drop them from page cache */
2534 		if (head[i].index >= end) {
2535 			ClearPageDirty(head + i);
2536 			__delete_from_page_cache(head + i, NULL);
2537 			if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2538 				shmem_uncharge(head->mapping->host, 1);
2539 			put_page(head + i);
2540 		} else if (!PageAnon(page)) {
2541 			__xa_store(&head->mapping->i_pages, head[i].index,
2542 					head + i, 0);
2543 		} else if (swap_cache) {
2544 			__xa_store(&swap_cache->i_pages, offset + i,
2545 					head + i, 0);
2546 		}
2547 	}
2548 
2549 	ClearPageCompound(head);
2550 
2551 	split_page_owner(head, HPAGE_PMD_ORDER);
2552 
2553 	/* See comment in __split_huge_page_tail() */
2554 	if (PageAnon(head)) {
2555 		/* Additional pin to swap cache */
2556 		if (PageSwapCache(head)) {
2557 			page_ref_add(head, 2);
2558 			xa_unlock(&swap_cache->i_pages);
2559 		} else {
2560 			page_ref_inc(head);
2561 		}
2562 	} else {
2563 		/* Additional pin to page cache */
2564 		page_ref_add(head, 2);
2565 		xa_unlock(&head->mapping->i_pages);
2566 	}
2567 
2568 	spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2569 
2570 	remap_page(head);
2571 
2572 	for (i = 0; i < HPAGE_PMD_NR; i++) {
2573 		struct page *subpage = head + i;
2574 		if (subpage == page)
2575 			continue;
2576 		unlock_page(subpage);
2577 
2578 		/*
2579 		 * Subpages may be freed if there wasn't any mapping
2580 		 * like if add_to_swap() is running on a lru page that
2581 		 * had its mapping zapped. And freeing these pages
2582 		 * requires taking the lru_lock so we do the put_page
2583 		 * of the tail pages after the split is complete.
2584 		 */
2585 		put_page(subpage);
2586 	}
2587 }
2588 
2589 int total_mapcount(struct page *page)
2590 {
2591 	int i, compound, ret;
2592 
2593 	VM_BUG_ON_PAGE(PageTail(page), page);
2594 
2595 	if (likely(!PageCompound(page)))
2596 		return atomic_read(&page->_mapcount) + 1;
2597 
2598 	compound = compound_mapcount(page);
2599 	if (PageHuge(page))
2600 		return compound;
2601 	ret = compound;
2602 	for (i = 0; i < HPAGE_PMD_NR; i++)
2603 		ret += atomic_read(&page[i]._mapcount) + 1;
2604 	/* File pages has compound_mapcount included in _mapcount */
2605 	if (!PageAnon(page))
2606 		return ret - compound * HPAGE_PMD_NR;
2607 	if (PageDoubleMap(page))
2608 		ret -= HPAGE_PMD_NR;
2609 	return ret;
2610 }
2611 
2612 /*
2613  * This calculates accurately how many mappings a transparent hugepage
2614  * has (unlike page_mapcount() which isn't fully accurate). This full
2615  * accuracy is primarily needed to know if copy-on-write faults can
2616  * reuse the page and change the mapping to read-write instead of
2617  * copying them. At the same time this returns the total_mapcount too.
2618  *
2619  * The function returns the highest mapcount any one of the subpages
2620  * has. If the return value is one, even if different processes are
2621  * mapping different subpages of the transparent hugepage, they can
2622  * all reuse it, because each process is reusing a different subpage.
2623  *
2624  * The total_mapcount is instead counting all virtual mappings of the
2625  * subpages. If the total_mapcount is equal to "one", it tells the
2626  * caller all mappings belong to the same "mm" and in turn the
2627  * anon_vma of the transparent hugepage can become the vma->anon_vma
2628  * local one as no other process may be mapping any of the subpages.
2629  *
2630  * It would be more accurate to replace page_mapcount() with
2631  * page_trans_huge_mapcount(), however we only use
2632  * page_trans_huge_mapcount() in the copy-on-write faults where we
2633  * need full accuracy to avoid breaking page pinning, because
2634  * page_trans_huge_mapcount() is slower than page_mapcount().
2635  */
2636 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2637 {
2638 	int i, ret, _total_mapcount, mapcount;
2639 
2640 	/* hugetlbfs shouldn't call it */
2641 	VM_BUG_ON_PAGE(PageHuge(page), page);
2642 
2643 	if (likely(!PageTransCompound(page))) {
2644 		mapcount = atomic_read(&page->_mapcount) + 1;
2645 		if (total_mapcount)
2646 			*total_mapcount = mapcount;
2647 		return mapcount;
2648 	}
2649 
2650 	page = compound_head(page);
2651 
2652 	_total_mapcount = ret = 0;
2653 	for (i = 0; i < HPAGE_PMD_NR; i++) {
2654 		mapcount = atomic_read(&page[i]._mapcount) + 1;
2655 		ret = max(ret, mapcount);
2656 		_total_mapcount += mapcount;
2657 	}
2658 	if (PageDoubleMap(page)) {
2659 		ret -= 1;
2660 		_total_mapcount -= HPAGE_PMD_NR;
2661 	}
2662 	mapcount = compound_mapcount(page);
2663 	ret += mapcount;
2664 	_total_mapcount += mapcount;
2665 	if (total_mapcount)
2666 		*total_mapcount = _total_mapcount;
2667 	return ret;
2668 }
2669 
2670 /* Racy check whether the huge page can be split */
2671 bool can_split_huge_page(struct page *page, int *pextra_pins)
2672 {
2673 	int extra_pins;
2674 
2675 	/* Additional pins from page cache */
2676 	if (PageAnon(page))
2677 		extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2678 	else
2679 		extra_pins = HPAGE_PMD_NR;
2680 	if (pextra_pins)
2681 		*pextra_pins = extra_pins;
2682 	return total_mapcount(page) == page_count(page) - extra_pins - 1;
2683 }
2684 
2685 /*
2686  * This function splits huge page into normal pages. @page can point to any
2687  * subpage of huge page to split. Split doesn't change the position of @page.
2688  *
2689  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2690  * The huge page must be locked.
2691  *
2692  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2693  *
2694  * Both head page and tail pages will inherit mapping, flags, and so on from
2695  * the hugepage.
2696  *
2697  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2698  * they are not mapped.
2699  *
2700  * Returns 0 if the hugepage is split successfully.
2701  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2702  * us.
2703  */
2704 int split_huge_page_to_list(struct page *page, struct list_head *list)
2705 {
2706 	struct page *head = compound_head(page);
2707 	struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2708 	struct deferred_split *ds_queue = get_deferred_split_queue(page);
2709 	struct anon_vma *anon_vma = NULL;
2710 	struct address_space *mapping = NULL;
2711 	int count, mapcount, extra_pins, ret;
2712 	bool mlocked;
2713 	unsigned long flags;
2714 	pgoff_t end;
2715 
2716 	VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2717 	VM_BUG_ON_PAGE(!PageLocked(page), page);
2718 	VM_BUG_ON_PAGE(!PageCompound(page), page);
2719 
2720 	if (PageWriteback(page))
2721 		return -EBUSY;
2722 
2723 	if (PageAnon(head)) {
2724 		/*
2725 		 * The caller does not necessarily hold an mmap_sem that would
2726 		 * prevent the anon_vma disappearing so we first we take a
2727 		 * reference to it and then lock the anon_vma for write. This
2728 		 * is similar to page_lock_anon_vma_read except the write lock
2729 		 * is taken to serialise against parallel split or collapse
2730 		 * operations.
2731 		 */
2732 		anon_vma = page_get_anon_vma(head);
2733 		if (!anon_vma) {
2734 			ret = -EBUSY;
2735 			goto out;
2736 		}
2737 		end = -1;
2738 		mapping = NULL;
2739 		anon_vma_lock_write(anon_vma);
2740 	} else {
2741 		mapping = head->mapping;
2742 
2743 		/* Truncated ? */
2744 		if (!mapping) {
2745 			ret = -EBUSY;
2746 			goto out;
2747 		}
2748 
2749 		anon_vma = NULL;
2750 		i_mmap_lock_read(mapping);
2751 
2752 		/*
2753 		 *__split_huge_page() may need to trim off pages beyond EOF:
2754 		 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2755 		 * which cannot be nested inside the page tree lock. So note
2756 		 * end now: i_size itself may be changed at any moment, but
2757 		 * head page lock is good enough to serialize the trimming.
2758 		 */
2759 		end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2760 	}
2761 
2762 	/*
2763 	 * Racy check if we can split the page, before unmap_page() will
2764 	 * split PMDs
2765 	 */
2766 	if (!can_split_huge_page(head, &extra_pins)) {
2767 		ret = -EBUSY;
2768 		goto out_unlock;
2769 	}
2770 
2771 	mlocked = PageMlocked(page);
2772 	unmap_page(head);
2773 	VM_BUG_ON_PAGE(compound_mapcount(head), head);
2774 
2775 	/* Make sure the page is not on per-CPU pagevec as it takes pin */
2776 	if (mlocked)
2777 		lru_add_drain();
2778 
2779 	/* prevent PageLRU to go away from under us, and freeze lru stats */
2780 	spin_lock_irqsave(&pgdata->lru_lock, flags);
2781 
2782 	if (mapping) {
2783 		XA_STATE(xas, &mapping->i_pages, page_index(head));
2784 
2785 		/*
2786 		 * Check if the head page is present in page cache.
2787 		 * We assume all tail are present too, if head is there.
2788 		 */
2789 		xa_lock(&mapping->i_pages);
2790 		if (xas_load(&xas) != head)
2791 			goto fail;
2792 	}
2793 
2794 	/* Prevent deferred_split_scan() touching ->_refcount */
2795 	spin_lock(&ds_queue->split_queue_lock);
2796 	count = page_count(head);
2797 	mapcount = total_mapcount(head);
2798 	if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2799 		if (!list_empty(page_deferred_list(head))) {
2800 			ds_queue->split_queue_len--;
2801 			list_del(page_deferred_list(head));
2802 		}
2803 		if (mapping)
2804 			__dec_node_page_state(page, NR_SHMEM_THPS);
2805 		spin_unlock(&ds_queue->split_queue_lock);
2806 		__split_huge_page(page, list, end, flags);
2807 		if (PageSwapCache(head)) {
2808 			swp_entry_t entry = { .val = page_private(head) };
2809 
2810 			ret = split_swap_cluster(entry);
2811 		} else
2812 			ret = 0;
2813 	} else {
2814 		if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2815 			pr_alert("total_mapcount: %u, page_count(): %u\n",
2816 					mapcount, count);
2817 			if (PageTail(page))
2818 				dump_page(head, NULL);
2819 			dump_page(page, "total_mapcount(head) > 0");
2820 			BUG();
2821 		}
2822 		spin_unlock(&ds_queue->split_queue_lock);
2823 fail:		if (mapping)
2824 			xa_unlock(&mapping->i_pages);
2825 		spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2826 		remap_page(head);
2827 		ret = -EBUSY;
2828 	}
2829 
2830 out_unlock:
2831 	if (anon_vma) {
2832 		anon_vma_unlock_write(anon_vma);
2833 		put_anon_vma(anon_vma);
2834 	}
2835 	if (mapping)
2836 		i_mmap_unlock_read(mapping);
2837 out:
2838 	count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2839 	return ret;
2840 }
2841 
2842 void free_transhuge_page(struct page *page)
2843 {
2844 	struct deferred_split *ds_queue = get_deferred_split_queue(page);
2845 	unsigned long flags;
2846 
2847 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2848 	if (!list_empty(page_deferred_list(page))) {
2849 		ds_queue->split_queue_len--;
2850 		list_del(page_deferred_list(page));
2851 	}
2852 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2853 	free_compound_page(page);
2854 }
2855 
2856 void deferred_split_huge_page(struct page *page)
2857 {
2858 	struct deferred_split *ds_queue = get_deferred_split_queue(page);
2859 #ifdef CONFIG_MEMCG
2860 	struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
2861 #endif
2862 	unsigned long flags;
2863 
2864 	VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2865 
2866 	/*
2867 	 * The try_to_unmap() in page reclaim path might reach here too,
2868 	 * this may cause a race condition to corrupt deferred split queue.
2869 	 * And, if page reclaim is already handling the same page, it is
2870 	 * unnecessary to handle it again in shrinker.
2871 	 *
2872 	 * Check PageSwapCache to determine if the page is being
2873 	 * handled by page reclaim since THP swap would add the page into
2874 	 * swap cache before calling try_to_unmap().
2875 	 */
2876 	if (PageSwapCache(page))
2877 		return;
2878 
2879 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2880 	if (list_empty(page_deferred_list(page))) {
2881 		count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2882 		list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2883 		ds_queue->split_queue_len++;
2884 #ifdef CONFIG_MEMCG
2885 		if (memcg)
2886 			memcg_set_shrinker_bit(memcg, page_to_nid(page),
2887 					       deferred_split_shrinker.id);
2888 #endif
2889 	}
2890 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2891 }
2892 
2893 static unsigned long deferred_split_count(struct shrinker *shrink,
2894 		struct shrink_control *sc)
2895 {
2896 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2897 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2898 
2899 #ifdef CONFIG_MEMCG
2900 	if (sc->memcg)
2901 		ds_queue = &sc->memcg->deferred_split_queue;
2902 #endif
2903 	return READ_ONCE(ds_queue->split_queue_len);
2904 }
2905 
2906 static unsigned long deferred_split_scan(struct shrinker *shrink,
2907 		struct shrink_control *sc)
2908 {
2909 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2910 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2911 	unsigned long flags;
2912 	LIST_HEAD(list), *pos, *next;
2913 	struct page *page;
2914 	int split = 0;
2915 
2916 #ifdef CONFIG_MEMCG
2917 	if (sc->memcg)
2918 		ds_queue = &sc->memcg->deferred_split_queue;
2919 #endif
2920 
2921 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2922 	/* Take pin on all head pages to avoid freeing them under us */
2923 	list_for_each_safe(pos, next, &ds_queue->split_queue) {
2924 		page = list_entry((void *)pos, struct page, mapping);
2925 		page = compound_head(page);
2926 		if (get_page_unless_zero(page)) {
2927 			list_move(page_deferred_list(page), &list);
2928 		} else {
2929 			/* We lost race with put_compound_page() */
2930 			list_del_init(page_deferred_list(page));
2931 			ds_queue->split_queue_len--;
2932 		}
2933 		if (!--sc->nr_to_scan)
2934 			break;
2935 	}
2936 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2937 
2938 	list_for_each_safe(pos, next, &list) {
2939 		page = list_entry((void *)pos, struct page, mapping);
2940 		if (!trylock_page(page))
2941 			goto next;
2942 		/* split_huge_page() removes page from list on success */
2943 		if (!split_huge_page(page))
2944 			split++;
2945 		unlock_page(page);
2946 next:
2947 		put_page(page);
2948 	}
2949 
2950 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2951 	list_splice_tail(&list, &ds_queue->split_queue);
2952 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2953 
2954 	/*
2955 	 * Stop shrinker if we didn't split any page, but the queue is empty.
2956 	 * This can happen if pages were freed under us.
2957 	 */
2958 	if (!split && list_empty(&ds_queue->split_queue))
2959 		return SHRINK_STOP;
2960 	return split;
2961 }
2962 
2963 static struct shrinker deferred_split_shrinker = {
2964 	.count_objects = deferred_split_count,
2965 	.scan_objects = deferred_split_scan,
2966 	.seeks = DEFAULT_SEEKS,
2967 	.flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2968 		 SHRINKER_NONSLAB,
2969 };
2970 
2971 #ifdef CONFIG_DEBUG_FS
2972 static int split_huge_pages_set(void *data, u64 val)
2973 {
2974 	struct zone *zone;
2975 	struct page *page;
2976 	unsigned long pfn, max_zone_pfn;
2977 	unsigned long total = 0, split = 0;
2978 
2979 	if (val != 1)
2980 		return -EINVAL;
2981 
2982 	for_each_populated_zone(zone) {
2983 		max_zone_pfn = zone_end_pfn(zone);
2984 		for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2985 			if (!pfn_valid(pfn))
2986 				continue;
2987 
2988 			page = pfn_to_page(pfn);
2989 			if (!get_page_unless_zero(page))
2990 				continue;
2991 
2992 			if (zone != page_zone(page))
2993 				goto next;
2994 
2995 			if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2996 				goto next;
2997 
2998 			total++;
2999 			lock_page(page);
3000 			if (!split_huge_page(page))
3001 				split++;
3002 			unlock_page(page);
3003 next:
3004 			put_page(page);
3005 		}
3006 	}
3007 
3008 	pr_info("%lu of %lu THP split\n", split, total);
3009 
3010 	return 0;
3011 }
3012 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3013 		"%llu\n");
3014 
3015 static int __init split_huge_pages_debugfs(void)
3016 {
3017 	debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3018 			    &split_huge_pages_fops);
3019 	return 0;
3020 }
3021 late_initcall(split_huge_pages_debugfs);
3022 #endif
3023 
3024 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3025 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3026 		struct page *page)
3027 {
3028 	struct vm_area_struct *vma = pvmw->vma;
3029 	struct mm_struct *mm = vma->vm_mm;
3030 	unsigned long address = pvmw->address;
3031 	pmd_t pmdval;
3032 	swp_entry_t entry;
3033 	pmd_t pmdswp;
3034 
3035 	if (!(pvmw->pmd && !pvmw->pte))
3036 		return;
3037 
3038 	flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3039 	pmdval = *pvmw->pmd;
3040 	pmdp_invalidate(vma, address, pvmw->pmd);
3041 	if (pmd_dirty(pmdval))
3042 		set_page_dirty(page);
3043 	entry = make_migration_entry(page, pmd_write(pmdval));
3044 	pmdswp = swp_entry_to_pmd(entry);
3045 	if (pmd_soft_dirty(pmdval))
3046 		pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3047 	set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3048 	page_remove_rmap(page, true);
3049 	put_page(page);
3050 }
3051 
3052 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3053 {
3054 	struct vm_area_struct *vma = pvmw->vma;
3055 	struct mm_struct *mm = vma->vm_mm;
3056 	unsigned long address = pvmw->address;
3057 	unsigned long mmun_start = address & HPAGE_PMD_MASK;
3058 	pmd_t pmde;
3059 	swp_entry_t entry;
3060 
3061 	if (!(pvmw->pmd && !pvmw->pte))
3062 		return;
3063 
3064 	entry = pmd_to_swp_entry(*pvmw->pmd);
3065 	get_page(new);
3066 	pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3067 	if (pmd_swp_soft_dirty(*pvmw->pmd))
3068 		pmde = pmd_mksoft_dirty(pmde);
3069 	if (is_write_migration_entry(entry))
3070 		pmde = maybe_pmd_mkwrite(pmde, vma);
3071 
3072 	flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3073 	if (PageAnon(new))
3074 		page_add_anon_rmap(new, vma, mmun_start, true);
3075 	else
3076 		page_add_file_rmap(new, true);
3077 	set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3078 	if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3079 		mlock_vma_page(new);
3080 	update_mmu_cache_pmd(vma, address, pvmw->pmd);
3081 }
3082 #endif
3083