xref: /openbmc/linux/mm/huge_memory.c (revision b9df3997)
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)
663 {
664 	const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
665 
666 	/* Always do synchronous compaction */
667 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
668 		return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
669 
670 	/* Kick kcompactd and fail quickly */
671 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
672 		return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
673 
674 	/* Synchronous compaction if madvised, otherwise kick kcompactd */
675 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
676 		return GFP_TRANSHUGE_LIGHT |
677 			(vma_madvised ? __GFP_DIRECT_RECLAIM :
678 					__GFP_KSWAPD_RECLAIM);
679 
680 	/* Only do synchronous compaction if madvised */
681 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
682 		return GFP_TRANSHUGE_LIGHT |
683 		       (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
684 
685 	return GFP_TRANSHUGE_LIGHT;
686 }
687 
688 /* Caller must hold page table lock. */
689 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
690 		struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
691 		struct page *zero_page)
692 {
693 	pmd_t entry;
694 	if (!pmd_none(*pmd))
695 		return false;
696 	entry = mk_pmd(zero_page, vma->vm_page_prot);
697 	entry = pmd_mkhuge(entry);
698 	if (pgtable)
699 		pgtable_trans_huge_deposit(mm, pmd, pgtable);
700 	set_pmd_at(mm, haddr, pmd, entry);
701 	mm_inc_nr_ptes(mm);
702 	return true;
703 }
704 
705 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
706 {
707 	struct vm_area_struct *vma = vmf->vma;
708 	gfp_t gfp;
709 	struct page *page;
710 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
711 
712 	if (!transhuge_vma_suitable(vma, haddr))
713 		return VM_FAULT_FALLBACK;
714 	if (unlikely(anon_vma_prepare(vma)))
715 		return VM_FAULT_OOM;
716 	if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
717 		return VM_FAULT_OOM;
718 	if (!(vmf->flags & FAULT_FLAG_WRITE) &&
719 			!mm_forbids_zeropage(vma->vm_mm) &&
720 			transparent_hugepage_use_zero_page()) {
721 		pgtable_t pgtable;
722 		struct page *zero_page;
723 		bool set;
724 		vm_fault_t ret;
725 		pgtable = pte_alloc_one(vma->vm_mm);
726 		if (unlikely(!pgtable))
727 			return VM_FAULT_OOM;
728 		zero_page = mm_get_huge_zero_page(vma->vm_mm);
729 		if (unlikely(!zero_page)) {
730 			pte_free(vma->vm_mm, pgtable);
731 			count_vm_event(THP_FAULT_FALLBACK);
732 			return VM_FAULT_FALLBACK;
733 		}
734 		vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
735 		ret = 0;
736 		set = false;
737 		if (pmd_none(*vmf->pmd)) {
738 			ret = check_stable_address_space(vma->vm_mm);
739 			if (ret) {
740 				spin_unlock(vmf->ptl);
741 			} else if (userfaultfd_missing(vma)) {
742 				spin_unlock(vmf->ptl);
743 				ret = handle_userfault(vmf, VM_UFFD_MISSING);
744 				VM_BUG_ON(ret & VM_FAULT_FALLBACK);
745 			} else {
746 				set_huge_zero_page(pgtable, vma->vm_mm, vma,
747 						   haddr, vmf->pmd, zero_page);
748 				spin_unlock(vmf->ptl);
749 				set = true;
750 			}
751 		} else
752 			spin_unlock(vmf->ptl);
753 		if (!set)
754 			pte_free(vma->vm_mm, pgtable);
755 		return ret;
756 	}
757 	gfp = alloc_hugepage_direct_gfpmask(vma);
758 	page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
759 	if (unlikely(!page)) {
760 		count_vm_event(THP_FAULT_FALLBACK);
761 		return VM_FAULT_FALLBACK;
762 	}
763 	prep_transhuge_page(page);
764 	return __do_huge_pmd_anonymous_page(vmf, page, gfp);
765 }
766 
767 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
768 		pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
769 		pgtable_t pgtable)
770 {
771 	struct mm_struct *mm = vma->vm_mm;
772 	pmd_t entry;
773 	spinlock_t *ptl;
774 
775 	ptl = pmd_lock(mm, pmd);
776 	if (!pmd_none(*pmd)) {
777 		if (write) {
778 			if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
779 				WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
780 				goto out_unlock;
781 			}
782 			entry = pmd_mkyoung(*pmd);
783 			entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
784 			if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
785 				update_mmu_cache_pmd(vma, addr, pmd);
786 		}
787 
788 		goto out_unlock;
789 	}
790 
791 	entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
792 	if (pfn_t_devmap(pfn))
793 		entry = pmd_mkdevmap(entry);
794 	if (write) {
795 		entry = pmd_mkyoung(pmd_mkdirty(entry));
796 		entry = maybe_pmd_mkwrite(entry, vma);
797 	}
798 
799 	if (pgtable) {
800 		pgtable_trans_huge_deposit(mm, pmd, pgtable);
801 		mm_inc_nr_ptes(mm);
802 		pgtable = NULL;
803 	}
804 
805 	set_pmd_at(mm, addr, pmd, entry);
806 	update_mmu_cache_pmd(vma, addr, pmd);
807 
808 out_unlock:
809 	spin_unlock(ptl);
810 	if (pgtable)
811 		pte_free(mm, pgtable);
812 }
813 
814 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)
815 {
816 	unsigned long addr = vmf->address & PMD_MASK;
817 	struct vm_area_struct *vma = vmf->vma;
818 	pgprot_t pgprot = vma->vm_page_prot;
819 	pgtable_t pgtable = NULL;
820 
821 	/*
822 	 * If we had pmd_special, we could avoid all these restrictions,
823 	 * but we need to be consistent with PTEs and architectures that
824 	 * can't support a 'special' bit.
825 	 */
826 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
827 			!pfn_t_devmap(pfn));
828 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
829 						(VM_PFNMAP|VM_MIXEDMAP));
830 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
831 
832 	if (addr < vma->vm_start || addr >= vma->vm_end)
833 		return VM_FAULT_SIGBUS;
834 
835 	if (arch_needs_pgtable_deposit()) {
836 		pgtable = pte_alloc_one(vma->vm_mm);
837 		if (!pgtable)
838 			return VM_FAULT_OOM;
839 	}
840 
841 	track_pfn_insert(vma, &pgprot, pfn);
842 
843 	insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
844 	return VM_FAULT_NOPAGE;
845 }
846 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
847 
848 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
849 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
850 {
851 	if (likely(vma->vm_flags & VM_WRITE))
852 		pud = pud_mkwrite(pud);
853 	return pud;
854 }
855 
856 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
857 		pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
858 {
859 	struct mm_struct *mm = vma->vm_mm;
860 	pud_t entry;
861 	spinlock_t *ptl;
862 
863 	ptl = pud_lock(mm, pud);
864 	if (!pud_none(*pud)) {
865 		if (write) {
866 			if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
867 				WARN_ON_ONCE(!is_huge_zero_pud(*pud));
868 				goto out_unlock;
869 			}
870 			entry = pud_mkyoung(*pud);
871 			entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
872 			if (pudp_set_access_flags(vma, addr, pud, entry, 1))
873 				update_mmu_cache_pud(vma, addr, pud);
874 		}
875 		goto out_unlock;
876 	}
877 
878 	entry = pud_mkhuge(pfn_t_pud(pfn, prot));
879 	if (pfn_t_devmap(pfn))
880 		entry = pud_mkdevmap(entry);
881 	if (write) {
882 		entry = pud_mkyoung(pud_mkdirty(entry));
883 		entry = maybe_pud_mkwrite(entry, vma);
884 	}
885 	set_pud_at(mm, addr, pud, entry);
886 	update_mmu_cache_pud(vma, addr, pud);
887 
888 out_unlock:
889 	spin_unlock(ptl);
890 }
891 
892 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)
893 {
894 	unsigned long addr = vmf->address & PUD_MASK;
895 	struct vm_area_struct *vma = vmf->vma;
896 	pgprot_t pgprot = vma->vm_page_prot;
897 
898 	/*
899 	 * If we had pud_special, we could avoid all these restrictions,
900 	 * but we need to be consistent with PTEs and architectures that
901 	 * can't support a 'special' bit.
902 	 */
903 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
904 			!pfn_t_devmap(pfn));
905 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
906 						(VM_PFNMAP|VM_MIXEDMAP));
907 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
908 
909 	if (addr < vma->vm_start || addr >= vma->vm_end)
910 		return VM_FAULT_SIGBUS;
911 
912 	track_pfn_insert(vma, &pgprot, pfn);
913 
914 	insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
915 	return VM_FAULT_NOPAGE;
916 }
917 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
918 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
919 
920 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
921 		pmd_t *pmd, int flags)
922 {
923 	pmd_t _pmd;
924 
925 	_pmd = pmd_mkyoung(*pmd);
926 	if (flags & FOLL_WRITE)
927 		_pmd = pmd_mkdirty(_pmd);
928 	if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
929 				pmd, _pmd, flags & FOLL_WRITE))
930 		update_mmu_cache_pmd(vma, addr, pmd);
931 }
932 
933 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
934 		pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
935 {
936 	unsigned long pfn = pmd_pfn(*pmd);
937 	struct mm_struct *mm = vma->vm_mm;
938 	struct page *page;
939 
940 	assert_spin_locked(pmd_lockptr(mm, pmd));
941 
942 	/*
943 	 * When we COW a devmap PMD entry, we split it into PTEs, so we should
944 	 * not be in this function with `flags & FOLL_COW` set.
945 	 */
946 	WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
947 
948 	if (flags & FOLL_WRITE && !pmd_write(*pmd))
949 		return NULL;
950 
951 	if (pmd_present(*pmd) && pmd_devmap(*pmd))
952 		/* pass */;
953 	else
954 		return NULL;
955 
956 	if (flags & FOLL_TOUCH)
957 		touch_pmd(vma, addr, pmd, flags);
958 
959 	/*
960 	 * device mapped pages can only be returned if the
961 	 * caller will manage the page reference count.
962 	 */
963 	if (!(flags & FOLL_GET))
964 		return ERR_PTR(-EEXIST);
965 
966 	pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
967 	*pgmap = get_dev_pagemap(pfn, *pgmap);
968 	if (!*pgmap)
969 		return ERR_PTR(-EFAULT);
970 	page = pfn_to_page(pfn);
971 	get_page(page);
972 
973 	return page;
974 }
975 
976 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
977 		  pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
978 		  struct vm_area_struct *vma)
979 {
980 	spinlock_t *dst_ptl, *src_ptl;
981 	struct page *src_page;
982 	pmd_t pmd;
983 	pgtable_t pgtable = NULL;
984 	int ret = -ENOMEM;
985 
986 	/* Skip if can be re-fill on fault */
987 	if (!vma_is_anonymous(vma))
988 		return 0;
989 
990 	pgtable = pte_alloc_one(dst_mm);
991 	if (unlikely(!pgtable))
992 		goto out;
993 
994 	dst_ptl = pmd_lock(dst_mm, dst_pmd);
995 	src_ptl = pmd_lockptr(src_mm, src_pmd);
996 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
997 
998 	ret = -EAGAIN;
999 	pmd = *src_pmd;
1000 
1001 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1002 	if (unlikely(is_swap_pmd(pmd))) {
1003 		swp_entry_t entry = pmd_to_swp_entry(pmd);
1004 
1005 		VM_BUG_ON(!is_pmd_migration_entry(pmd));
1006 		if (is_write_migration_entry(entry)) {
1007 			make_migration_entry_read(&entry);
1008 			pmd = swp_entry_to_pmd(entry);
1009 			if (pmd_swp_soft_dirty(*src_pmd))
1010 				pmd = pmd_swp_mksoft_dirty(pmd);
1011 			set_pmd_at(src_mm, addr, src_pmd, pmd);
1012 		}
1013 		add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1014 		mm_inc_nr_ptes(dst_mm);
1015 		pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1016 		set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1017 		ret = 0;
1018 		goto out_unlock;
1019 	}
1020 #endif
1021 
1022 	if (unlikely(!pmd_trans_huge(pmd))) {
1023 		pte_free(dst_mm, pgtable);
1024 		goto out_unlock;
1025 	}
1026 	/*
1027 	 * When page table lock is held, the huge zero pmd should not be
1028 	 * under splitting since we don't split the page itself, only pmd to
1029 	 * a page table.
1030 	 */
1031 	if (is_huge_zero_pmd(pmd)) {
1032 		struct page *zero_page;
1033 		/*
1034 		 * get_huge_zero_page() will never allocate a new page here,
1035 		 * since we already have a zero page to copy. It just takes a
1036 		 * reference.
1037 		 */
1038 		zero_page = mm_get_huge_zero_page(dst_mm);
1039 		set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1040 				zero_page);
1041 		ret = 0;
1042 		goto out_unlock;
1043 	}
1044 
1045 	src_page = pmd_page(pmd);
1046 	VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1047 	get_page(src_page);
1048 	page_dup_rmap(src_page, true);
1049 	add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1050 	mm_inc_nr_ptes(dst_mm);
1051 	pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1052 
1053 	pmdp_set_wrprotect(src_mm, addr, src_pmd);
1054 	pmd = pmd_mkold(pmd_wrprotect(pmd));
1055 	set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1056 
1057 	ret = 0;
1058 out_unlock:
1059 	spin_unlock(src_ptl);
1060 	spin_unlock(dst_ptl);
1061 out:
1062 	return ret;
1063 }
1064 
1065 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1066 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1067 		pud_t *pud, int flags)
1068 {
1069 	pud_t _pud;
1070 
1071 	_pud = pud_mkyoung(*pud);
1072 	if (flags & FOLL_WRITE)
1073 		_pud = pud_mkdirty(_pud);
1074 	if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1075 				pud, _pud, flags & FOLL_WRITE))
1076 		update_mmu_cache_pud(vma, addr, pud);
1077 }
1078 
1079 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1080 		pud_t *pud, int flags, struct dev_pagemap **pgmap)
1081 {
1082 	unsigned long pfn = pud_pfn(*pud);
1083 	struct mm_struct *mm = vma->vm_mm;
1084 	struct page *page;
1085 
1086 	assert_spin_locked(pud_lockptr(mm, pud));
1087 
1088 	if (flags & FOLL_WRITE && !pud_write(*pud))
1089 		return NULL;
1090 
1091 	if (pud_present(*pud) && pud_devmap(*pud))
1092 		/* pass */;
1093 	else
1094 		return NULL;
1095 
1096 	if (flags & FOLL_TOUCH)
1097 		touch_pud(vma, addr, pud, flags);
1098 
1099 	/*
1100 	 * device mapped pages can only be returned if the
1101 	 * caller will manage the page reference count.
1102 	 */
1103 	if (!(flags & FOLL_GET))
1104 		return ERR_PTR(-EEXIST);
1105 
1106 	pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1107 	*pgmap = get_dev_pagemap(pfn, *pgmap);
1108 	if (!*pgmap)
1109 		return ERR_PTR(-EFAULT);
1110 	page = pfn_to_page(pfn);
1111 	get_page(page);
1112 
1113 	return page;
1114 }
1115 
1116 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1117 		  pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1118 		  struct vm_area_struct *vma)
1119 {
1120 	spinlock_t *dst_ptl, *src_ptl;
1121 	pud_t pud;
1122 	int ret;
1123 
1124 	dst_ptl = pud_lock(dst_mm, dst_pud);
1125 	src_ptl = pud_lockptr(src_mm, src_pud);
1126 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1127 
1128 	ret = -EAGAIN;
1129 	pud = *src_pud;
1130 	if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1131 		goto out_unlock;
1132 
1133 	/*
1134 	 * When page table lock is held, the huge zero pud should not be
1135 	 * under splitting since we don't split the page itself, only pud to
1136 	 * a page table.
1137 	 */
1138 	if (is_huge_zero_pud(pud)) {
1139 		/* No huge zero pud yet */
1140 	}
1141 
1142 	pudp_set_wrprotect(src_mm, addr, src_pud);
1143 	pud = pud_mkold(pud_wrprotect(pud));
1144 	set_pud_at(dst_mm, addr, dst_pud, pud);
1145 
1146 	ret = 0;
1147 out_unlock:
1148 	spin_unlock(src_ptl);
1149 	spin_unlock(dst_ptl);
1150 	return ret;
1151 }
1152 
1153 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1154 {
1155 	pud_t entry;
1156 	unsigned long haddr;
1157 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1158 
1159 	vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1160 	if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1161 		goto unlock;
1162 
1163 	entry = pud_mkyoung(orig_pud);
1164 	if (write)
1165 		entry = pud_mkdirty(entry);
1166 	haddr = vmf->address & HPAGE_PUD_MASK;
1167 	if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1168 		update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1169 
1170 unlock:
1171 	spin_unlock(vmf->ptl);
1172 }
1173 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1174 
1175 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1176 {
1177 	pmd_t entry;
1178 	unsigned long haddr;
1179 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1180 
1181 	vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1182 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1183 		goto unlock;
1184 
1185 	entry = pmd_mkyoung(orig_pmd);
1186 	if (write)
1187 		entry = pmd_mkdirty(entry);
1188 	haddr = vmf->address & HPAGE_PMD_MASK;
1189 	if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1190 		update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1191 
1192 unlock:
1193 	spin_unlock(vmf->ptl);
1194 }
1195 
1196 static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1197 			pmd_t orig_pmd, struct page *page)
1198 {
1199 	struct vm_area_struct *vma = vmf->vma;
1200 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1201 	struct mem_cgroup *memcg;
1202 	pgtable_t pgtable;
1203 	pmd_t _pmd;
1204 	int i;
1205 	vm_fault_t ret = 0;
1206 	struct page **pages;
1207 	struct mmu_notifier_range range;
1208 
1209 	pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1210 			      GFP_KERNEL);
1211 	if (unlikely(!pages)) {
1212 		ret |= VM_FAULT_OOM;
1213 		goto out;
1214 	}
1215 
1216 	for (i = 0; i < HPAGE_PMD_NR; i++) {
1217 		pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1218 					       vmf->address, page_to_nid(page));
1219 		if (unlikely(!pages[i] ||
1220 			     mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
1221 				     GFP_KERNEL, &memcg, false))) {
1222 			if (pages[i])
1223 				put_page(pages[i]);
1224 			while (--i >= 0) {
1225 				memcg = (void *)page_private(pages[i]);
1226 				set_page_private(pages[i], 0);
1227 				mem_cgroup_cancel_charge(pages[i], memcg,
1228 						false);
1229 				put_page(pages[i]);
1230 			}
1231 			kfree(pages);
1232 			ret |= VM_FAULT_OOM;
1233 			goto out;
1234 		}
1235 		set_page_private(pages[i], (unsigned long)memcg);
1236 	}
1237 
1238 	for (i = 0; i < HPAGE_PMD_NR; i++) {
1239 		copy_user_highpage(pages[i], page + i,
1240 				   haddr + PAGE_SIZE * i, vma);
1241 		__SetPageUptodate(pages[i]);
1242 		cond_resched();
1243 	}
1244 
1245 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1246 				haddr, haddr + HPAGE_PMD_SIZE);
1247 	mmu_notifier_invalidate_range_start(&range);
1248 
1249 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1250 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1251 		goto out_free_pages;
1252 	VM_BUG_ON_PAGE(!PageHead(page), page);
1253 
1254 	/*
1255 	 * Leave pmd empty until pte is filled note we must notify here as
1256 	 * concurrent CPU thread might write to new page before the call to
1257 	 * mmu_notifier_invalidate_range_end() happens which can lead to a
1258 	 * device seeing memory write in different order than CPU.
1259 	 *
1260 	 * See Documentation/vm/mmu_notifier.rst
1261 	 */
1262 	pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1263 
1264 	pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1265 	pmd_populate(vma->vm_mm, &_pmd, pgtable);
1266 
1267 	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1268 		pte_t entry;
1269 		entry = mk_pte(pages[i], vma->vm_page_prot);
1270 		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1271 		memcg = (void *)page_private(pages[i]);
1272 		set_page_private(pages[i], 0);
1273 		page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1274 		mem_cgroup_commit_charge(pages[i], memcg, false, false);
1275 		lru_cache_add_active_or_unevictable(pages[i], vma);
1276 		vmf->pte = pte_offset_map(&_pmd, haddr);
1277 		VM_BUG_ON(!pte_none(*vmf->pte));
1278 		set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1279 		pte_unmap(vmf->pte);
1280 	}
1281 	kfree(pages);
1282 
1283 	smp_wmb(); /* make pte visible before pmd */
1284 	pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1285 	page_remove_rmap(page, true);
1286 	spin_unlock(vmf->ptl);
1287 
1288 	/*
1289 	 * No need to double call mmu_notifier->invalidate_range() callback as
1290 	 * the above pmdp_huge_clear_flush_notify() did already call it.
1291 	 */
1292 	mmu_notifier_invalidate_range_only_end(&range);
1293 
1294 	ret |= VM_FAULT_WRITE;
1295 	put_page(page);
1296 
1297 out:
1298 	return ret;
1299 
1300 out_free_pages:
1301 	spin_unlock(vmf->ptl);
1302 	mmu_notifier_invalidate_range_end(&range);
1303 	for (i = 0; i < HPAGE_PMD_NR; i++) {
1304 		memcg = (void *)page_private(pages[i]);
1305 		set_page_private(pages[i], 0);
1306 		mem_cgroup_cancel_charge(pages[i], memcg, false);
1307 		put_page(pages[i]);
1308 	}
1309 	kfree(pages);
1310 	goto out;
1311 }
1312 
1313 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1314 {
1315 	struct vm_area_struct *vma = vmf->vma;
1316 	struct page *page = NULL, *new_page;
1317 	struct mem_cgroup *memcg;
1318 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1319 	struct mmu_notifier_range range;
1320 	gfp_t huge_gfp;			/* for allocation and charge */
1321 	vm_fault_t ret = 0;
1322 
1323 	vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1324 	VM_BUG_ON_VMA(!vma->anon_vma, vma);
1325 	if (is_huge_zero_pmd(orig_pmd))
1326 		goto alloc;
1327 	spin_lock(vmf->ptl);
1328 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1329 		goto out_unlock;
1330 
1331 	page = pmd_page(orig_pmd);
1332 	VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1333 	/*
1334 	 * We can only reuse the page if nobody else maps the huge page or it's
1335 	 * part.
1336 	 */
1337 	if (!trylock_page(page)) {
1338 		get_page(page);
1339 		spin_unlock(vmf->ptl);
1340 		lock_page(page);
1341 		spin_lock(vmf->ptl);
1342 		if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1343 			unlock_page(page);
1344 			put_page(page);
1345 			goto out_unlock;
1346 		}
1347 		put_page(page);
1348 	}
1349 	if (reuse_swap_page(page, NULL)) {
1350 		pmd_t entry;
1351 		entry = pmd_mkyoung(orig_pmd);
1352 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1353 		if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1354 			update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1355 		ret |= VM_FAULT_WRITE;
1356 		unlock_page(page);
1357 		goto out_unlock;
1358 	}
1359 	unlock_page(page);
1360 	get_page(page);
1361 	spin_unlock(vmf->ptl);
1362 alloc:
1363 	if (__transparent_hugepage_enabled(vma) &&
1364 	    !transparent_hugepage_debug_cow()) {
1365 		huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1366 		new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1367 	} else
1368 		new_page = NULL;
1369 
1370 	if (likely(new_page)) {
1371 		prep_transhuge_page(new_page);
1372 	} else {
1373 		if (!page) {
1374 			split_huge_pmd(vma, vmf->pmd, vmf->address);
1375 			ret |= VM_FAULT_FALLBACK;
1376 		} else {
1377 			ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1378 			if (ret & VM_FAULT_OOM) {
1379 				split_huge_pmd(vma, vmf->pmd, vmf->address);
1380 				ret |= VM_FAULT_FALLBACK;
1381 			}
1382 			put_page(page);
1383 		}
1384 		count_vm_event(THP_FAULT_FALLBACK);
1385 		goto out;
1386 	}
1387 
1388 	if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
1389 					huge_gfp, &memcg, true))) {
1390 		put_page(new_page);
1391 		split_huge_pmd(vma, vmf->pmd, vmf->address);
1392 		if (page)
1393 			put_page(page);
1394 		ret |= VM_FAULT_FALLBACK;
1395 		count_vm_event(THP_FAULT_FALLBACK);
1396 		goto out;
1397 	}
1398 
1399 	count_vm_event(THP_FAULT_ALLOC);
1400 	count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
1401 
1402 	if (!page)
1403 		clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1404 	else
1405 		copy_user_huge_page(new_page, page, vmf->address,
1406 				    vma, HPAGE_PMD_NR);
1407 	__SetPageUptodate(new_page);
1408 
1409 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1410 				haddr, haddr + HPAGE_PMD_SIZE);
1411 	mmu_notifier_invalidate_range_start(&range);
1412 
1413 	spin_lock(vmf->ptl);
1414 	if (page)
1415 		put_page(page);
1416 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1417 		spin_unlock(vmf->ptl);
1418 		mem_cgroup_cancel_charge(new_page, memcg, true);
1419 		put_page(new_page);
1420 		goto out_mn;
1421 	} else {
1422 		pmd_t entry;
1423 		entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1424 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1425 		pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1426 		page_add_new_anon_rmap(new_page, vma, haddr, true);
1427 		mem_cgroup_commit_charge(new_page, memcg, false, true);
1428 		lru_cache_add_active_or_unevictable(new_page, vma);
1429 		set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1430 		update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1431 		if (!page) {
1432 			add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1433 		} else {
1434 			VM_BUG_ON_PAGE(!PageHead(page), page);
1435 			page_remove_rmap(page, true);
1436 			put_page(page);
1437 		}
1438 		ret |= VM_FAULT_WRITE;
1439 	}
1440 	spin_unlock(vmf->ptl);
1441 out_mn:
1442 	/*
1443 	 * No need to double call mmu_notifier->invalidate_range() callback as
1444 	 * the above pmdp_huge_clear_flush_notify() did already call it.
1445 	 */
1446 	mmu_notifier_invalidate_range_only_end(&range);
1447 out:
1448 	return ret;
1449 out_unlock:
1450 	spin_unlock(vmf->ptl);
1451 	return ret;
1452 }
1453 
1454 /*
1455  * FOLL_FORCE can write to even unwritable pmd's, but only
1456  * after we've gone through a COW cycle and they are dirty.
1457  */
1458 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1459 {
1460 	return pmd_write(pmd) ||
1461 	       ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1462 }
1463 
1464 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1465 				   unsigned long addr,
1466 				   pmd_t *pmd,
1467 				   unsigned int flags)
1468 {
1469 	struct mm_struct *mm = vma->vm_mm;
1470 	struct page *page = NULL;
1471 
1472 	assert_spin_locked(pmd_lockptr(mm, pmd));
1473 
1474 	if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1475 		goto out;
1476 
1477 	/* Avoid dumping huge zero page */
1478 	if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1479 		return ERR_PTR(-EFAULT);
1480 
1481 	/* Full NUMA hinting faults to serialise migration in fault paths */
1482 	if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1483 		goto out;
1484 
1485 	page = pmd_page(*pmd);
1486 	VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1487 	if (flags & FOLL_TOUCH)
1488 		touch_pmd(vma, addr, pmd, flags);
1489 	if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1490 		/*
1491 		 * We don't mlock() pte-mapped THPs. This way we can avoid
1492 		 * leaking mlocked pages into non-VM_LOCKED VMAs.
1493 		 *
1494 		 * For anon THP:
1495 		 *
1496 		 * In most cases the pmd is the only mapping of the page as we
1497 		 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1498 		 * writable private mappings in populate_vma_page_range().
1499 		 *
1500 		 * The only scenario when we have the page shared here is if we
1501 		 * mlocking read-only mapping shared over fork(). We skip
1502 		 * mlocking such pages.
1503 		 *
1504 		 * For file THP:
1505 		 *
1506 		 * We can expect PageDoubleMap() to be stable under page lock:
1507 		 * for file pages we set it in page_add_file_rmap(), which
1508 		 * requires page to be locked.
1509 		 */
1510 
1511 		if (PageAnon(page) && compound_mapcount(page) != 1)
1512 			goto skip_mlock;
1513 		if (PageDoubleMap(page) || !page->mapping)
1514 			goto skip_mlock;
1515 		if (!trylock_page(page))
1516 			goto skip_mlock;
1517 		lru_add_drain();
1518 		if (page->mapping && !PageDoubleMap(page))
1519 			mlock_vma_page(page);
1520 		unlock_page(page);
1521 	}
1522 skip_mlock:
1523 	page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1524 	VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1525 	if (flags & FOLL_GET)
1526 		get_page(page);
1527 
1528 out:
1529 	return page;
1530 }
1531 
1532 /* NUMA hinting page fault entry point for trans huge pmds */
1533 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1534 {
1535 	struct vm_area_struct *vma = vmf->vma;
1536 	struct anon_vma *anon_vma = NULL;
1537 	struct page *page;
1538 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1539 	int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1540 	int target_nid, last_cpupid = -1;
1541 	bool page_locked;
1542 	bool migrated = false;
1543 	bool was_writable;
1544 	int flags = 0;
1545 
1546 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1547 	if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1548 		goto out_unlock;
1549 
1550 	/*
1551 	 * If there are potential migrations, wait for completion and retry
1552 	 * without disrupting NUMA hinting information. Do not relock and
1553 	 * check_same as the page may no longer be mapped.
1554 	 */
1555 	if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1556 		page = pmd_page(*vmf->pmd);
1557 		if (!get_page_unless_zero(page))
1558 			goto out_unlock;
1559 		spin_unlock(vmf->ptl);
1560 		put_and_wait_on_page_locked(page);
1561 		goto out;
1562 	}
1563 
1564 	page = pmd_page(pmd);
1565 	BUG_ON(is_huge_zero_page(page));
1566 	page_nid = page_to_nid(page);
1567 	last_cpupid = page_cpupid_last(page);
1568 	count_vm_numa_event(NUMA_HINT_FAULTS);
1569 	if (page_nid == this_nid) {
1570 		count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1571 		flags |= TNF_FAULT_LOCAL;
1572 	}
1573 
1574 	/* See similar comment in do_numa_page for explanation */
1575 	if (!pmd_savedwrite(pmd))
1576 		flags |= TNF_NO_GROUP;
1577 
1578 	/*
1579 	 * Acquire the page lock to serialise THP migrations but avoid dropping
1580 	 * page_table_lock if at all possible
1581 	 */
1582 	page_locked = trylock_page(page);
1583 	target_nid = mpol_misplaced(page, vma, haddr);
1584 	if (target_nid == NUMA_NO_NODE) {
1585 		/* If the page was locked, there are no parallel migrations */
1586 		if (page_locked)
1587 			goto clear_pmdnuma;
1588 	}
1589 
1590 	/* Migration could have started since the pmd_trans_migrating check */
1591 	if (!page_locked) {
1592 		page_nid = NUMA_NO_NODE;
1593 		if (!get_page_unless_zero(page))
1594 			goto out_unlock;
1595 		spin_unlock(vmf->ptl);
1596 		put_and_wait_on_page_locked(page);
1597 		goto out;
1598 	}
1599 
1600 	/*
1601 	 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1602 	 * to serialises splits
1603 	 */
1604 	get_page(page);
1605 	spin_unlock(vmf->ptl);
1606 	anon_vma = page_lock_anon_vma_read(page);
1607 
1608 	/* Confirm the PMD did not change while page_table_lock was released */
1609 	spin_lock(vmf->ptl);
1610 	if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1611 		unlock_page(page);
1612 		put_page(page);
1613 		page_nid = NUMA_NO_NODE;
1614 		goto out_unlock;
1615 	}
1616 
1617 	/* Bail if we fail to protect against THP splits for any reason */
1618 	if (unlikely(!anon_vma)) {
1619 		put_page(page);
1620 		page_nid = NUMA_NO_NODE;
1621 		goto clear_pmdnuma;
1622 	}
1623 
1624 	/*
1625 	 * Since we took the NUMA fault, we must have observed the !accessible
1626 	 * bit. Make sure all other CPUs agree with that, to avoid them
1627 	 * modifying the page we're about to migrate.
1628 	 *
1629 	 * Must be done under PTL such that we'll observe the relevant
1630 	 * inc_tlb_flush_pending().
1631 	 *
1632 	 * We are not sure a pending tlb flush here is for a huge page
1633 	 * mapping or not. Hence use the tlb range variant
1634 	 */
1635 	if (mm_tlb_flush_pending(vma->vm_mm)) {
1636 		flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1637 		/*
1638 		 * change_huge_pmd() released the pmd lock before
1639 		 * invalidating the secondary MMUs sharing the primary
1640 		 * MMU pagetables (with ->invalidate_range()). The
1641 		 * mmu_notifier_invalidate_range_end() (which
1642 		 * internally calls ->invalidate_range()) in
1643 		 * change_pmd_range() will run after us, so we can't
1644 		 * rely on it here and we need an explicit invalidate.
1645 		 */
1646 		mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1647 					      haddr + HPAGE_PMD_SIZE);
1648 	}
1649 
1650 	/*
1651 	 * Migrate the THP to the requested node, returns with page unlocked
1652 	 * and access rights restored.
1653 	 */
1654 	spin_unlock(vmf->ptl);
1655 
1656 	migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1657 				vmf->pmd, pmd, vmf->address, page, target_nid);
1658 	if (migrated) {
1659 		flags |= TNF_MIGRATED;
1660 		page_nid = target_nid;
1661 	} else
1662 		flags |= TNF_MIGRATE_FAIL;
1663 
1664 	goto out;
1665 clear_pmdnuma:
1666 	BUG_ON(!PageLocked(page));
1667 	was_writable = pmd_savedwrite(pmd);
1668 	pmd = pmd_modify(pmd, vma->vm_page_prot);
1669 	pmd = pmd_mkyoung(pmd);
1670 	if (was_writable)
1671 		pmd = pmd_mkwrite(pmd);
1672 	set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1673 	update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1674 	unlock_page(page);
1675 out_unlock:
1676 	spin_unlock(vmf->ptl);
1677 
1678 out:
1679 	if (anon_vma)
1680 		page_unlock_anon_vma_read(anon_vma);
1681 
1682 	if (page_nid != NUMA_NO_NODE)
1683 		task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1684 				flags);
1685 
1686 	return 0;
1687 }
1688 
1689 /*
1690  * Return true if we do MADV_FREE successfully on entire pmd page.
1691  * Otherwise, return false.
1692  */
1693 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1694 		pmd_t *pmd, unsigned long addr, unsigned long next)
1695 {
1696 	spinlock_t *ptl;
1697 	pmd_t orig_pmd;
1698 	struct page *page;
1699 	struct mm_struct *mm = tlb->mm;
1700 	bool ret = false;
1701 
1702 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1703 
1704 	ptl = pmd_trans_huge_lock(pmd, vma);
1705 	if (!ptl)
1706 		goto out_unlocked;
1707 
1708 	orig_pmd = *pmd;
1709 	if (is_huge_zero_pmd(orig_pmd))
1710 		goto out;
1711 
1712 	if (unlikely(!pmd_present(orig_pmd))) {
1713 		VM_BUG_ON(thp_migration_supported() &&
1714 				  !is_pmd_migration_entry(orig_pmd));
1715 		goto out;
1716 	}
1717 
1718 	page = pmd_page(orig_pmd);
1719 	/*
1720 	 * If other processes are mapping this page, we couldn't discard
1721 	 * the page unless they all do MADV_FREE so let's skip the page.
1722 	 */
1723 	if (page_mapcount(page) != 1)
1724 		goto out;
1725 
1726 	if (!trylock_page(page))
1727 		goto out;
1728 
1729 	/*
1730 	 * If user want to discard part-pages of THP, split it so MADV_FREE
1731 	 * will deactivate only them.
1732 	 */
1733 	if (next - addr != HPAGE_PMD_SIZE) {
1734 		get_page(page);
1735 		spin_unlock(ptl);
1736 		split_huge_page(page);
1737 		unlock_page(page);
1738 		put_page(page);
1739 		goto out_unlocked;
1740 	}
1741 
1742 	if (PageDirty(page))
1743 		ClearPageDirty(page);
1744 	unlock_page(page);
1745 
1746 	if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1747 		pmdp_invalidate(vma, addr, pmd);
1748 		orig_pmd = pmd_mkold(orig_pmd);
1749 		orig_pmd = pmd_mkclean(orig_pmd);
1750 
1751 		set_pmd_at(mm, addr, pmd, orig_pmd);
1752 		tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1753 	}
1754 
1755 	mark_page_lazyfree(page);
1756 	ret = true;
1757 out:
1758 	spin_unlock(ptl);
1759 out_unlocked:
1760 	return ret;
1761 }
1762 
1763 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1764 {
1765 	pgtable_t pgtable;
1766 
1767 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1768 	pte_free(mm, pgtable);
1769 	mm_dec_nr_ptes(mm);
1770 }
1771 
1772 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1773 		 pmd_t *pmd, unsigned long addr)
1774 {
1775 	pmd_t orig_pmd;
1776 	spinlock_t *ptl;
1777 
1778 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1779 
1780 	ptl = __pmd_trans_huge_lock(pmd, vma);
1781 	if (!ptl)
1782 		return 0;
1783 	/*
1784 	 * For architectures like ppc64 we look at deposited pgtable
1785 	 * when calling pmdp_huge_get_and_clear. So do the
1786 	 * pgtable_trans_huge_withdraw after finishing pmdp related
1787 	 * operations.
1788 	 */
1789 	orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1790 			tlb->fullmm);
1791 	tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1792 	if (vma_is_dax(vma)) {
1793 		if (arch_needs_pgtable_deposit())
1794 			zap_deposited_table(tlb->mm, pmd);
1795 		spin_unlock(ptl);
1796 		if (is_huge_zero_pmd(orig_pmd))
1797 			tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1798 	} else if (is_huge_zero_pmd(orig_pmd)) {
1799 		zap_deposited_table(tlb->mm, pmd);
1800 		spin_unlock(ptl);
1801 		tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1802 	} else {
1803 		struct page *page = NULL;
1804 		int flush_needed = 1;
1805 
1806 		if (pmd_present(orig_pmd)) {
1807 			page = pmd_page(orig_pmd);
1808 			page_remove_rmap(page, true);
1809 			VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1810 			VM_BUG_ON_PAGE(!PageHead(page), page);
1811 		} else if (thp_migration_supported()) {
1812 			swp_entry_t entry;
1813 
1814 			VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1815 			entry = pmd_to_swp_entry(orig_pmd);
1816 			page = pfn_to_page(swp_offset(entry));
1817 			flush_needed = 0;
1818 		} else
1819 			WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1820 
1821 		if (PageAnon(page)) {
1822 			zap_deposited_table(tlb->mm, pmd);
1823 			add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1824 		} else {
1825 			if (arch_needs_pgtable_deposit())
1826 				zap_deposited_table(tlb->mm, pmd);
1827 			add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1828 		}
1829 
1830 		spin_unlock(ptl);
1831 		if (flush_needed)
1832 			tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1833 	}
1834 	return 1;
1835 }
1836 
1837 #ifndef pmd_move_must_withdraw
1838 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1839 					 spinlock_t *old_pmd_ptl,
1840 					 struct vm_area_struct *vma)
1841 {
1842 	/*
1843 	 * With split pmd lock we also need to move preallocated
1844 	 * PTE page table if new_pmd is on different PMD page table.
1845 	 *
1846 	 * We also don't deposit and withdraw tables for file pages.
1847 	 */
1848 	return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1849 }
1850 #endif
1851 
1852 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1853 {
1854 #ifdef CONFIG_MEM_SOFT_DIRTY
1855 	if (unlikely(is_pmd_migration_entry(pmd)))
1856 		pmd = pmd_swp_mksoft_dirty(pmd);
1857 	else if (pmd_present(pmd))
1858 		pmd = pmd_mksoft_dirty(pmd);
1859 #endif
1860 	return pmd;
1861 }
1862 
1863 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1864 		  unsigned long new_addr, unsigned long old_end,
1865 		  pmd_t *old_pmd, pmd_t *new_pmd)
1866 {
1867 	spinlock_t *old_ptl, *new_ptl;
1868 	pmd_t pmd;
1869 	struct mm_struct *mm = vma->vm_mm;
1870 	bool force_flush = false;
1871 
1872 	if ((old_addr & ~HPAGE_PMD_MASK) ||
1873 	    (new_addr & ~HPAGE_PMD_MASK) ||
1874 	    old_end - old_addr < HPAGE_PMD_SIZE)
1875 		return false;
1876 
1877 	/*
1878 	 * The destination pmd shouldn't be established, free_pgtables()
1879 	 * should have release it.
1880 	 */
1881 	if (WARN_ON(!pmd_none(*new_pmd))) {
1882 		VM_BUG_ON(pmd_trans_huge(*new_pmd));
1883 		return false;
1884 	}
1885 
1886 	/*
1887 	 * We don't have to worry about the ordering of src and dst
1888 	 * ptlocks because exclusive mmap_sem prevents deadlock.
1889 	 */
1890 	old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1891 	if (old_ptl) {
1892 		new_ptl = pmd_lockptr(mm, new_pmd);
1893 		if (new_ptl != old_ptl)
1894 			spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1895 		pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1896 		if (pmd_present(pmd))
1897 			force_flush = true;
1898 		VM_BUG_ON(!pmd_none(*new_pmd));
1899 
1900 		if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1901 			pgtable_t pgtable;
1902 			pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1903 			pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1904 		}
1905 		pmd = move_soft_dirty_pmd(pmd);
1906 		set_pmd_at(mm, new_addr, new_pmd, pmd);
1907 		if (force_flush)
1908 			flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1909 		if (new_ptl != old_ptl)
1910 			spin_unlock(new_ptl);
1911 		spin_unlock(old_ptl);
1912 		return true;
1913 	}
1914 	return false;
1915 }
1916 
1917 /*
1918  * Returns
1919  *  - 0 if PMD could not be locked
1920  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1921  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1922  */
1923 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1924 		unsigned long addr, pgprot_t newprot, int prot_numa)
1925 {
1926 	struct mm_struct *mm = vma->vm_mm;
1927 	spinlock_t *ptl;
1928 	pmd_t entry;
1929 	bool preserve_write;
1930 	int ret;
1931 
1932 	ptl = __pmd_trans_huge_lock(pmd, vma);
1933 	if (!ptl)
1934 		return 0;
1935 
1936 	preserve_write = prot_numa && pmd_write(*pmd);
1937 	ret = 1;
1938 
1939 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1940 	if (is_swap_pmd(*pmd)) {
1941 		swp_entry_t entry = pmd_to_swp_entry(*pmd);
1942 
1943 		VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1944 		if (is_write_migration_entry(entry)) {
1945 			pmd_t newpmd;
1946 			/*
1947 			 * A protection check is difficult so
1948 			 * just be safe and disable write
1949 			 */
1950 			make_migration_entry_read(&entry);
1951 			newpmd = swp_entry_to_pmd(entry);
1952 			if (pmd_swp_soft_dirty(*pmd))
1953 				newpmd = pmd_swp_mksoft_dirty(newpmd);
1954 			set_pmd_at(mm, addr, pmd, newpmd);
1955 		}
1956 		goto unlock;
1957 	}
1958 #endif
1959 
1960 	/*
1961 	 * Avoid trapping faults against the zero page. The read-only
1962 	 * data is likely to be read-cached on the local CPU and
1963 	 * local/remote hits to the zero page are not interesting.
1964 	 */
1965 	if (prot_numa && is_huge_zero_pmd(*pmd))
1966 		goto unlock;
1967 
1968 	if (prot_numa && pmd_protnone(*pmd))
1969 		goto unlock;
1970 
1971 	/*
1972 	 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1973 	 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1974 	 * which is also under down_read(mmap_sem):
1975 	 *
1976 	 *	CPU0:				CPU1:
1977 	 *				change_huge_pmd(prot_numa=1)
1978 	 *				 pmdp_huge_get_and_clear_notify()
1979 	 * madvise_dontneed()
1980 	 *  zap_pmd_range()
1981 	 *   pmd_trans_huge(*pmd) == 0 (without ptl)
1982 	 *   // skip the pmd
1983 	 *				 set_pmd_at();
1984 	 *				 // pmd is re-established
1985 	 *
1986 	 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1987 	 * which may break userspace.
1988 	 *
1989 	 * pmdp_invalidate() is required to make sure we don't miss
1990 	 * dirty/young flags set by hardware.
1991 	 */
1992 	entry = pmdp_invalidate(vma, addr, pmd);
1993 
1994 	entry = pmd_modify(entry, newprot);
1995 	if (preserve_write)
1996 		entry = pmd_mk_savedwrite(entry);
1997 	ret = HPAGE_PMD_NR;
1998 	set_pmd_at(mm, addr, pmd, entry);
1999 	BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
2000 unlock:
2001 	spin_unlock(ptl);
2002 	return ret;
2003 }
2004 
2005 /*
2006  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
2007  *
2008  * Note that if it returns page table lock pointer, this routine returns without
2009  * unlocking page table lock. So callers must unlock it.
2010  */
2011 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
2012 {
2013 	spinlock_t *ptl;
2014 	ptl = pmd_lock(vma->vm_mm, pmd);
2015 	if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
2016 			pmd_devmap(*pmd)))
2017 		return ptl;
2018 	spin_unlock(ptl);
2019 	return NULL;
2020 }
2021 
2022 /*
2023  * Returns true if a given pud maps a thp, false otherwise.
2024  *
2025  * Note that if it returns true, this routine returns without unlocking page
2026  * table lock. So callers must unlock it.
2027  */
2028 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2029 {
2030 	spinlock_t *ptl;
2031 
2032 	ptl = pud_lock(vma->vm_mm, pud);
2033 	if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2034 		return ptl;
2035 	spin_unlock(ptl);
2036 	return NULL;
2037 }
2038 
2039 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2040 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2041 		 pud_t *pud, unsigned long addr)
2042 {
2043 	spinlock_t *ptl;
2044 
2045 	ptl = __pud_trans_huge_lock(pud, vma);
2046 	if (!ptl)
2047 		return 0;
2048 	/*
2049 	 * For architectures like ppc64 we look at deposited pgtable
2050 	 * when calling pudp_huge_get_and_clear. So do the
2051 	 * pgtable_trans_huge_withdraw after finishing pudp related
2052 	 * operations.
2053 	 */
2054 	pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
2055 	tlb_remove_pud_tlb_entry(tlb, pud, addr);
2056 	if (vma_is_dax(vma)) {
2057 		spin_unlock(ptl);
2058 		/* No zero page support yet */
2059 	} else {
2060 		/* No support for anonymous PUD pages yet */
2061 		BUG();
2062 	}
2063 	return 1;
2064 }
2065 
2066 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2067 		unsigned long haddr)
2068 {
2069 	VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2070 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2071 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2072 	VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2073 
2074 	count_vm_event(THP_SPLIT_PUD);
2075 
2076 	pudp_huge_clear_flush_notify(vma, haddr, pud);
2077 }
2078 
2079 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2080 		unsigned long address)
2081 {
2082 	spinlock_t *ptl;
2083 	struct mmu_notifier_range range;
2084 
2085 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2086 				address & HPAGE_PUD_MASK,
2087 				(address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2088 	mmu_notifier_invalidate_range_start(&range);
2089 	ptl = pud_lock(vma->vm_mm, pud);
2090 	if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2091 		goto out;
2092 	__split_huge_pud_locked(vma, pud, range.start);
2093 
2094 out:
2095 	spin_unlock(ptl);
2096 	/*
2097 	 * No need to double call mmu_notifier->invalidate_range() callback as
2098 	 * the above pudp_huge_clear_flush_notify() did already call it.
2099 	 */
2100 	mmu_notifier_invalidate_range_only_end(&range);
2101 }
2102 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2103 
2104 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2105 		unsigned long haddr, pmd_t *pmd)
2106 {
2107 	struct mm_struct *mm = vma->vm_mm;
2108 	pgtable_t pgtable;
2109 	pmd_t _pmd;
2110 	int i;
2111 
2112 	/*
2113 	 * Leave pmd empty until pte is filled note that it is fine to delay
2114 	 * notification until mmu_notifier_invalidate_range_end() as we are
2115 	 * replacing a zero pmd write protected page with a zero pte write
2116 	 * protected page.
2117 	 *
2118 	 * See Documentation/vm/mmu_notifier.rst
2119 	 */
2120 	pmdp_huge_clear_flush(vma, haddr, pmd);
2121 
2122 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2123 	pmd_populate(mm, &_pmd, pgtable);
2124 
2125 	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2126 		pte_t *pte, entry;
2127 		entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2128 		entry = pte_mkspecial(entry);
2129 		pte = pte_offset_map(&_pmd, haddr);
2130 		VM_BUG_ON(!pte_none(*pte));
2131 		set_pte_at(mm, haddr, pte, entry);
2132 		pte_unmap(pte);
2133 	}
2134 	smp_wmb(); /* make pte visible before pmd */
2135 	pmd_populate(mm, pmd, pgtable);
2136 }
2137 
2138 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2139 		unsigned long haddr, bool freeze)
2140 {
2141 	struct mm_struct *mm = vma->vm_mm;
2142 	struct page *page;
2143 	pgtable_t pgtable;
2144 	pmd_t old_pmd, _pmd;
2145 	bool young, write, soft_dirty, pmd_migration = false;
2146 	unsigned long addr;
2147 	int i;
2148 
2149 	VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2150 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2151 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2152 	VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2153 				&& !pmd_devmap(*pmd));
2154 
2155 	count_vm_event(THP_SPLIT_PMD);
2156 
2157 	if (!vma_is_anonymous(vma)) {
2158 		_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2159 		/*
2160 		 * We are going to unmap this huge page. So
2161 		 * just go ahead and zap it
2162 		 */
2163 		if (arch_needs_pgtable_deposit())
2164 			zap_deposited_table(mm, pmd);
2165 		if (vma_is_dax(vma))
2166 			return;
2167 		page = pmd_page(_pmd);
2168 		if (!PageDirty(page) && pmd_dirty(_pmd))
2169 			set_page_dirty(page);
2170 		if (!PageReferenced(page) && pmd_young(_pmd))
2171 			SetPageReferenced(page);
2172 		page_remove_rmap(page, true);
2173 		put_page(page);
2174 		add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2175 		return;
2176 	} else if (is_huge_zero_pmd(*pmd)) {
2177 		/*
2178 		 * FIXME: Do we want to invalidate secondary mmu by calling
2179 		 * mmu_notifier_invalidate_range() see comments below inside
2180 		 * __split_huge_pmd() ?
2181 		 *
2182 		 * We are going from a zero huge page write protected to zero
2183 		 * small page also write protected so it does not seems useful
2184 		 * to invalidate secondary mmu at this time.
2185 		 */
2186 		return __split_huge_zero_page_pmd(vma, haddr, pmd);
2187 	}
2188 
2189 	/*
2190 	 * Up to this point the pmd is present and huge and userland has the
2191 	 * whole access to the hugepage during the split (which happens in
2192 	 * place). If we overwrite the pmd with the not-huge version pointing
2193 	 * to the pte here (which of course we could if all CPUs were bug
2194 	 * free), userland could trigger a small page size TLB miss on the
2195 	 * small sized TLB while the hugepage TLB entry is still established in
2196 	 * the huge TLB. Some CPU doesn't like that.
2197 	 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2198 	 * 383 on page 93. Intel should be safe but is also warns that it's
2199 	 * only safe if the permission and cache attributes of the two entries
2200 	 * loaded in the two TLB is identical (which should be the case here).
2201 	 * But it is generally safer to never allow small and huge TLB entries
2202 	 * for the same virtual address to be loaded simultaneously. So instead
2203 	 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2204 	 * current pmd notpresent (atomically because here the pmd_trans_huge
2205 	 * must remain set at all times on the pmd until the split is complete
2206 	 * for this pmd), then we flush the SMP TLB and finally we write the
2207 	 * non-huge version of the pmd entry with pmd_populate.
2208 	 */
2209 	old_pmd = pmdp_invalidate(vma, haddr, pmd);
2210 
2211 	pmd_migration = is_pmd_migration_entry(old_pmd);
2212 	if (unlikely(pmd_migration)) {
2213 		swp_entry_t entry;
2214 
2215 		entry = pmd_to_swp_entry(old_pmd);
2216 		page = pfn_to_page(swp_offset(entry));
2217 		write = is_write_migration_entry(entry);
2218 		young = false;
2219 		soft_dirty = pmd_swp_soft_dirty(old_pmd);
2220 	} else {
2221 		page = pmd_page(old_pmd);
2222 		if (pmd_dirty(old_pmd))
2223 			SetPageDirty(page);
2224 		write = pmd_write(old_pmd);
2225 		young = pmd_young(old_pmd);
2226 		soft_dirty = pmd_soft_dirty(old_pmd);
2227 	}
2228 	VM_BUG_ON_PAGE(!page_count(page), page);
2229 	page_ref_add(page, HPAGE_PMD_NR - 1);
2230 
2231 	/*
2232 	 * Withdraw the table only after we mark the pmd entry invalid.
2233 	 * This's critical for some architectures (Power).
2234 	 */
2235 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2236 	pmd_populate(mm, &_pmd, pgtable);
2237 
2238 	for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2239 		pte_t entry, *pte;
2240 		/*
2241 		 * Note that NUMA hinting access restrictions are not
2242 		 * transferred to avoid any possibility of altering
2243 		 * permissions across VMAs.
2244 		 */
2245 		if (freeze || pmd_migration) {
2246 			swp_entry_t swp_entry;
2247 			swp_entry = make_migration_entry(page + i, write);
2248 			entry = swp_entry_to_pte(swp_entry);
2249 			if (soft_dirty)
2250 				entry = pte_swp_mksoft_dirty(entry);
2251 		} else {
2252 			entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2253 			entry = maybe_mkwrite(entry, vma);
2254 			if (!write)
2255 				entry = pte_wrprotect(entry);
2256 			if (!young)
2257 				entry = pte_mkold(entry);
2258 			if (soft_dirty)
2259 				entry = pte_mksoft_dirty(entry);
2260 		}
2261 		pte = pte_offset_map(&_pmd, addr);
2262 		BUG_ON(!pte_none(*pte));
2263 		set_pte_at(mm, addr, pte, entry);
2264 		atomic_inc(&page[i]._mapcount);
2265 		pte_unmap(pte);
2266 	}
2267 
2268 	/*
2269 	 * Set PG_double_map before dropping compound_mapcount to avoid
2270 	 * false-negative page_mapped().
2271 	 */
2272 	if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2273 		for (i = 0; i < HPAGE_PMD_NR; i++)
2274 			atomic_inc(&page[i]._mapcount);
2275 	}
2276 
2277 	if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2278 		/* Last compound_mapcount is gone. */
2279 		__dec_node_page_state(page, NR_ANON_THPS);
2280 		if (TestClearPageDoubleMap(page)) {
2281 			/* No need in mapcount reference anymore */
2282 			for (i = 0; i < HPAGE_PMD_NR; i++)
2283 				atomic_dec(&page[i]._mapcount);
2284 		}
2285 	}
2286 
2287 	smp_wmb(); /* make pte visible before pmd */
2288 	pmd_populate(mm, pmd, pgtable);
2289 
2290 	if (freeze) {
2291 		for (i = 0; i < HPAGE_PMD_NR; i++) {
2292 			page_remove_rmap(page + i, false);
2293 			put_page(page + i);
2294 		}
2295 	}
2296 }
2297 
2298 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2299 		unsigned long address, bool freeze, struct page *page)
2300 {
2301 	spinlock_t *ptl;
2302 	struct mmu_notifier_range range;
2303 
2304 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2305 				address & HPAGE_PMD_MASK,
2306 				(address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2307 	mmu_notifier_invalidate_range_start(&range);
2308 	ptl = pmd_lock(vma->vm_mm, pmd);
2309 
2310 	/*
2311 	 * If caller asks to setup a migration entries, we need a page to check
2312 	 * pmd against. Otherwise we can end up replacing wrong page.
2313 	 */
2314 	VM_BUG_ON(freeze && !page);
2315 	if (page && page != pmd_page(*pmd))
2316 	        goto out;
2317 
2318 	if (pmd_trans_huge(*pmd)) {
2319 		page = pmd_page(*pmd);
2320 		if (PageMlocked(page))
2321 			clear_page_mlock(page);
2322 	} else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2323 		goto out;
2324 	__split_huge_pmd_locked(vma, pmd, range.start, freeze);
2325 out:
2326 	spin_unlock(ptl);
2327 	/*
2328 	 * No need to double call mmu_notifier->invalidate_range() callback.
2329 	 * They are 3 cases to consider inside __split_huge_pmd_locked():
2330 	 *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2331 	 *  2) __split_huge_zero_page_pmd() read only zero page and any write
2332 	 *    fault will trigger a flush_notify before pointing to a new page
2333 	 *    (it is fine if the secondary mmu keeps pointing to the old zero
2334 	 *    page in the meantime)
2335 	 *  3) Split a huge pmd into pte pointing to the same page. No need
2336 	 *     to invalidate secondary tlb entry they are all still valid.
2337 	 *     any further changes to individual pte will notify. So no need
2338 	 *     to call mmu_notifier->invalidate_range()
2339 	 */
2340 	mmu_notifier_invalidate_range_only_end(&range);
2341 }
2342 
2343 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2344 		bool freeze, struct page *page)
2345 {
2346 	pgd_t *pgd;
2347 	p4d_t *p4d;
2348 	pud_t *pud;
2349 	pmd_t *pmd;
2350 
2351 	pgd = pgd_offset(vma->vm_mm, address);
2352 	if (!pgd_present(*pgd))
2353 		return;
2354 
2355 	p4d = p4d_offset(pgd, address);
2356 	if (!p4d_present(*p4d))
2357 		return;
2358 
2359 	pud = pud_offset(p4d, address);
2360 	if (!pud_present(*pud))
2361 		return;
2362 
2363 	pmd = pmd_offset(pud, address);
2364 
2365 	__split_huge_pmd(vma, pmd, address, freeze, page);
2366 }
2367 
2368 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2369 			     unsigned long start,
2370 			     unsigned long end,
2371 			     long adjust_next)
2372 {
2373 	/*
2374 	 * If the new start address isn't hpage aligned and it could
2375 	 * previously contain an hugepage: check if we need to split
2376 	 * an huge pmd.
2377 	 */
2378 	if (start & ~HPAGE_PMD_MASK &&
2379 	    (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2380 	    (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2381 		split_huge_pmd_address(vma, start, false, NULL);
2382 
2383 	/*
2384 	 * If the new end address isn't hpage aligned and it could
2385 	 * previously contain an hugepage: check if we need to split
2386 	 * an huge pmd.
2387 	 */
2388 	if (end & ~HPAGE_PMD_MASK &&
2389 	    (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2390 	    (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2391 		split_huge_pmd_address(vma, end, false, NULL);
2392 
2393 	/*
2394 	 * If we're also updating the vma->vm_next->vm_start, if the new
2395 	 * vm_next->vm_start isn't page aligned and it could previously
2396 	 * contain an hugepage: check if we need to split an huge pmd.
2397 	 */
2398 	if (adjust_next > 0) {
2399 		struct vm_area_struct *next = vma->vm_next;
2400 		unsigned long nstart = next->vm_start;
2401 		nstart += adjust_next << PAGE_SHIFT;
2402 		if (nstart & ~HPAGE_PMD_MASK &&
2403 		    (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2404 		    (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2405 			split_huge_pmd_address(next, nstart, false, NULL);
2406 	}
2407 }
2408 
2409 static void unmap_page(struct page *page)
2410 {
2411 	enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2412 		TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2413 	bool unmap_success;
2414 
2415 	VM_BUG_ON_PAGE(!PageHead(page), page);
2416 
2417 	if (PageAnon(page))
2418 		ttu_flags |= TTU_SPLIT_FREEZE;
2419 
2420 	unmap_success = try_to_unmap(page, ttu_flags);
2421 	VM_BUG_ON_PAGE(!unmap_success, page);
2422 }
2423 
2424 static void remap_page(struct page *page)
2425 {
2426 	int i;
2427 	if (PageTransHuge(page)) {
2428 		remove_migration_ptes(page, page, true);
2429 	} else {
2430 		for (i = 0; i < HPAGE_PMD_NR; i++)
2431 			remove_migration_ptes(page + i, page + i, true);
2432 	}
2433 }
2434 
2435 static void __split_huge_page_tail(struct page *head, int tail,
2436 		struct lruvec *lruvec, struct list_head *list)
2437 {
2438 	struct page *page_tail = head + tail;
2439 
2440 	VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2441 
2442 	/*
2443 	 * Clone page flags before unfreezing refcount.
2444 	 *
2445 	 * After successful get_page_unless_zero() might follow flags change,
2446 	 * for exmaple lock_page() which set PG_waiters.
2447 	 */
2448 	page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2449 	page_tail->flags |= (head->flags &
2450 			((1L << PG_referenced) |
2451 			 (1L << PG_swapbacked) |
2452 			 (1L << PG_swapcache) |
2453 			 (1L << PG_mlocked) |
2454 			 (1L << PG_uptodate) |
2455 			 (1L << PG_active) |
2456 			 (1L << PG_workingset) |
2457 			 (1L << PG_locked) |
2458 			 (1L << PG_unevictable) |
2459 			 (1L << PG_dirty)));
2460 
2461 	/* ->mapping in first tail page is compound_mapcount */
2462 	VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2463 			page_tail);
2464 	page_tail->mapping = head->mapping;
2465 	page_tail->index = head->index + tail;
2466 
2467 	/* Page flags must be visible before we make the page non-compound. */
2468 	smp_wmb();
2469 
2470 	/*
2471 	 * Clear PageTail before unfreezing page refcount.
2472 	 *
2473 	 * After successful get_page_unless_zero() might follow put_page()
2474 	 * which needs correct compound_head().
2475 	 */
2476 	clear_compound_head(page_tail);
2477 
2478 	/* Finally unfreeze refcount. Additional reference from page cache. */
2479 	page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2480 					  PageSwapCache(head)));
2481 
2482 	if (page_is_young(head))
2483 		set_page_young(page_tail);
2484 	if (page_is_idle(head))
2485 		set_page_idle(page_tail);
2486 
2487 	page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2488 
2489 	/*
2490 	 * always add to the tail because some iterators expect new
2491 	 * pages to show after the currently processed elements - e.g.
2492 	 * migrate_pages
2493 	 */
2494 	lru_add_page_tail(head, page_tail, lruvec, list);
2495 }
2496 
2497 static void __split_huge_page(struct page *page, struct list_head *list,
2498 		pgoff_t end, unsigned long flags)
2499 {
2500 	struct page *head = compound_head(page);
2501 	pg_data_t *pgdat = page_pgdat(head);
2502 	struct lruvec *lruvec;
2503 	struct address_space *swap_cache = NULL;
2504 	unsigned long offset = 0;
2505 	int i;
2506 
2507 	lruvec = mem_cgroup_page_lruvec(head, pgdat);
2508 
2509 	/* complete memcg works before add pages to LRU */
2510 	mem_cgroup_split_huge_fixup(head);
2511 
2512 	if (PageAnon(head) && PageSwapCache(head)) {
2513 		swp_entry_t entry = { .val = page_private(head) };
2514 
2515 		offset = swp_offset(entry);
2516 		swap_cache = swap_address_space(entry);
2517 		xa_lock(&swap_cache->i_pages);
2518 	}
2519 
2520 	for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2521 		__split_huge_page_tail(head, i, lruvec, list);
2522 		/* Some pages can be beyond i_size: drop them from page cache */
2523 		if (head[i].index >= end) {
2524 			ClearPageDirty(head + i);
2525 			__delete_from_page_cache(head + i, NULL);
2526 			if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2527 				shmem_uncharge(head->mapping->host, 1);
2528 			put_page(head + i);
2529 		} else if (!PageAnon(page)) {
2530 			__xa_store(&head->mapping->i_pages, head[i].index,
2531 					head + i, 0);
2532 		} else if (swap_cache) {
2533 			__xa_store(&swap_cache->i_pages, offset + i,
2534 					head + i, 0);
2535 		}
2536 	}
2537 
2538 	ClearPageCompound(head);
2539 
2540 	split_page_owner(head, HPAGE_PMD_ORDER);
2541 
2542 	/* See comment in __split_huge_page_tail() */
2543 	if (PageAnon(head)) {
2544 		/* Additional pin to swap cache */
2545 		if (PageSwapCache(head)) {
2546 			page_ref_add(head, 2);
2547 			xa_unlock(&swap_cache->i_pages);
2548 		} else {
2549 			page_ref_inc(head);
2550 		}
2551 	} else {
2552 		/* Additional pin to page cache */
2553 		page_ref_add(head, 2);
2554 		xa_unlock(&head->mapping->i_pages);
2555 	}
2556 
2557 	spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2558 
2559 	remap_page(head);
2560 
2561 	for (i = 0; i < HPAGE_PMD_NR; i++) {
2562 		struct page *subpage = head + i;
2563 		if (subpage == page)
2564 			continue;
2565 		unlock_page(subpage);
2566 
2567 		/*
2568 		 * Subpages may be freed if there wasn't any mapping
2569 		 * like if add_to_swap() is running on a lru page that
2570 		 * had its mapping zapped. And freeing these pages
2571 		 * requires taking the lru_lock so we do the put_page
2572 		 * of the tail pages after the split is complete.
2573 		 */
2574 		put_page(subpage);
2575 	}
2576 }
2577 
2578 int total_mapcount(struct page *page)
2579 {
2580 	int i, compound, ret;
2581 
2582 	VM_BUG_ON_PAGE(PageTail(page), page);
2583 
2584 	if (likely(!PageCompound(page)))
2585 		return atomic_read(&page->_mapcount) + 1;
2586 
2587 	compound = compound_mapcount(page);
2588 	if (PageHuge(page))
2589 		return compound;
2590 	ret = compound;
2591 	for (i = 0; i < HPAGE_PMD_NR; i++)
2592 		ret += atomic_read(&page[i]._mapcount) + 1;
2593 	/* File pages has compound_mapcount included in _mapcount */
2594 	if (!PageAnon(page))
2595 		return ret - compound * HPAGE_PMD_NR;
2596 	if (PageDoubleMap(page))
2597 		ret -= HPAGE_PMD_NR;
2598 	return ret;
2599 }
2600 
2601 /*
2602  * This calculates accurately how many mappings a transparent hugepage
2603  * has (unlike page_mapcount() which isn't fully accurate). This full
2604  * accuracy is primarily needed to know if copy-on-write faults can
2605  * reuse the page and change the mapping to read-write instead of
2606  * copying them. At the same time this returns the total_mapcount too.
2607  *
2608  * The function returns the highest mapcount any one of the subpages
2609  * has. If the return value is one, even if different processes are
2610  * mapping different subpages of the transparent hugepage, they can
2611  * all reuse it, because each process is reusing a different subpage.
2612  *
2613  * The total_mapcount is instead counting all virtual mappings of the
2614  * subpages. If the total_mapcount is equal to "one", it tells the
2615  * caller all mappings belong to the same "mm" and in turn the
2616  * anon_vma of the transparent hugepage can become the vma->anon_vma
2617  * local one as no other process may be mapping any of the subpages.
2618  *
2619  * It would be more accurate to replace page_mapcount() with
2620  * page_trans_huge_mapcount(), however we only use
2621  * page_trans_huge_mapcount() in the copy-on-write faults where we
2622  * need full accuracy to avoid breaking page pinning, because
2623  * page_trans_huge_mapcount() is slower than page_mapcount().
2624  */
2625 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2626 {
2627 	int i, ret, _total_mapcount, mapcount;
2628 
2629 	/* hugetlbfs shouldn't call it */
2630 	VM_BUG_ON_PAGE(PageHuge(page), page);
2631 
2632 	if (likely(!PageTransCompound(page))) {
2633 		mapcount = atomic_read(&page->_mapcount) + 1;
2634 		if (total_mapcount)
2635 			*total_mapcount = mapcount;
2636 		return mapcount;
2637 	}
2638 
2639 	page = compound_head(page);
2640 
2641 	_total_mapcount = ret = 0;
2642 	for (i = 0; i < HPAGE_PMD_NR; i++) {
2643 		mapcount = atomic_read(&page[i]._mapcount) + 1;
2644 		ret = max(ret, mapcount);
2645 		_total_mapcount += mapcount;
2646 	}
2647 	if (PageDoubleMap(page)) {
2648 		ret -= 1;
2649 		_total_mapcount -= HPAGE_PMD_NR;
2650 	}
2651 	mapcount = compound_mapcount(page);
2652 	ret += mapcount;
2653 	_total_mapcount += mapcount;
2654 	if (total_mapcount)
2655 		*total_mapcount = _total_mapcount;
2656 	return ret;
2657 }
2658 
2659 /* Racy check whether the huge page can be split */
2660 bool can_split_huge_page(struct page *page, int *pextra_pins)
2661 {
2662 	int extra_pins;
2663 
2664 	/* Additional pins from page cache */
2665 	if (PageAnon(page))
2666 		extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2667 	else
2668 		extra_pins = HPAGE_PMD_NR;
2669 	if (pextra_pins)
2670 		*pextra_pins = extra_pins;
2671 	return total_mapcount(page) == page_count(page) - extra_pins - 1;
2672 }
2673 
2674 /*
2675  * This function splits huge page into normal pages. @page can point to any
2676  * subpage of huge page to split. Split doesn't change the position of @page.
2677  *
2678  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2679  * The huge page must be locked.
2680  *
2681  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2682  *
2683  * Both head page and tail pages will inherit mapping, flags, and so on from
2684  * the hugepage.
2685  *
2686  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2687  * they are not mapped.
2688  *
2689  * Returns 0 if the hugepage is split successfully.
2690  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2691  * us.
2692  */
2693 int split_huge_page_to_list(struct page *page, struct list_head *list)
2694 {
2695 	struct page *head = compound_head(page);
2696 	struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2697 	struct deferred_split *ds_queue = get_deferred_split_queue(page);
2698 	struct anon_vma *anon_vma = NULL;
2699 	struct address_space *mapping = NULL;
2700 	int count, mapcount, extra_pins, ret;
2701 	bool mlocked;
2702 	unsigned long flags;
2703 	pgoff_t end;
2704 
2705 	VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2706 	VM_BUG_ON_PAGE(!PageLocked(page), page);
2707 	VM_BUG_ON_PAGE(!PageCompound(page), page);
2708 
2709 	if (PageWriteback(page))
2710 		return -EBUSY;
2711 
2712 	if (PageAnon(head)) {
2713 		/*
2714 		 * The caller does not necessarily hold an mmap_sem that would
2715 		 * prevent the anon_vma disappearing so we first we take a
2716 		 * reference to it and then lock the anon_vma for write. This
2717 		 * is similar to page_lock_anon_vma_read except the write lock
2718 		 * is taken to serialise against parallel split or collapse
2719 		 * operations.
2720 		 */
2721 		anon_vma = page_get_anon_vma(head);
2722 		if (!anon_vma) {
2723 			ret = -EBUSY;
2724 			goto out;
2725 		}
2726 		end = -1;
2727 		mapping = NULL;
2728 		anon_vma_lock_write(anon_vma);
2729 	} else {
2730 		mapping = head->mapping;
2731 
2732 		/* Truncated ? */
2733 		if (!mapping) {
2734 			ret = -EBUSY;
2735 			goto out;
2736 		}
2737 
2738 		anon_vma = NULL;
2739 		i_mmap_lock_read(mapping);
2740 
2741 		/*
2742 		 *__split_huge_page() may need to trim off pages beyond EOF:
2743 		 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2744 		 * which cannot be nested inside the page tree lock. So note
2745 		 * end now: i_size itself may be changed at any moment, but
2746 		 * head page lock is good enough to serialize the trimming.
2747 		 */
2748 		end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2749 	}
2750 
2751 	/*
2752 	 * Racy check if we can split the page, before unmap_page() will
2753 	 * split PMDs
2754 	 */
2755 	if (!can_split_huge_page(head, &extra_pins)) {
2756 		ret = -EBUSY;
2757 		goto out_unlock;
2758 	}
2759 
2760 	mlocked = PageMlocked(page);
2761 	unmap_page(head);
2762 	VM_BUG_ON_PAGE(compound_mapcount(head), head);
2763 
2764 	/* Make sure the page is not on per-CPU pagevec as it takes pin */
2765 	if (mlocked)
2766 		lru_add_drain();
2767 
2768 	/* prevent PageLRU to go away from under us, and freeze lru stats */
2769 	spin_lock_irqsave(&pgdata->lru_lock, flags);
2770 
2771 	if (mapping) {
2772 		XA_STATE(xas, &mapping->i_pages, page_index(head));
2773 
2774 		/*
2775 		 * Check if the head page is present in page cache.
2776 		 * We assume all tail are present too, if head is there.
2777 		 */
2778 		xa_lock(&mapping->i_pages);
2779 		if (xas_load(&xas) != head)
2780 			goto fail;
2781 	}
2782 
2783 	/* Prevent deferred_split_scan() touching ->_refcount */
2784 	spin_lock(&ds_queue->split_queue_lock);
2785 	count = page_count(head);
2786 	mapcount = total_mapcount(head);
2787 	if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2788 		if (!list_empty(page_deferred_list(head))) {
2789 			ds_queue->split_queue_len--;
2790 			list_del(page_deferred_list(head));
2791 		}
2792 		if (mapping)
2793 			__dec_node_page_state(page, NR_SHMEM_THPS);
2794 		spin_unlock(&ds_queue->split_queue_lock);
2795 		__split_huge_page(page, list, end, flags);
2796 		if (PageSwapCache(head)) {
2797 			swp_entry_t entry = { .val = page_private(head) };
2798 
2799 			ret = split_swap_cluster(entry);
2800 		} else
2801 			ret = 0;
2802 	} else {
2803 		if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2804 			pr_alert("total_mapcount: %u, page_count(): %u\n",
2805 					mapcount, count);
2806 			if (PageTail(page))
2807 				dump_page(head, NULL);
2808 			dump_page(page, "total_mapcount(head) > 0");
2809 			BUG();
2810 		}
2811 		spin_unlock(&ds_queue->split_queue_lock);
2812 fail:		if (mapping)
2813 			xa_unlock(&mapping->i_pages);
2814 		spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2815 		remap_page(head);
2816 		ret = -EBUSY;
2817 	}
2818 
2819 out_unlock:
2820 	if (anon_vma) {
2821 		anon_vma_unlock_write(anon_vma);
2822 		put_anon_vma(anon_vma);
2823 	}
2824 	if (mapping)
2825 		i_mmap_unlock_read(mapping);
2826 out:
2827 	count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2828 	return ret;
2829 }
2830 
2831 void free_transhuge_page(struct page *page)
2832 {
2833 	struct deferred_split *ds_queue = get_deferred_split_queue(page);
2834 	unsigned long flags;
2835 
2836 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2837 	if (!list_empty(page_deferred_list(page))) {
2838 		ds_queue->split_queue_len--;
2839 		list_del(page_deferred_list(page));
2840 	}
2841 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2842 	free_compound_page(page);
2843 }
2844 
2845 void deferred_split_huge_page(struct page *page)
2846 {
2847 	struct deferred_split *ds_queue = get_deferred_split_queue(page);
2848 #ifdef CONFIG_MEMCG
2849 	struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
2850 #endif
2851 	unsigned long flags;
2852 
2853 	VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2854 
2855 	/*
2856 	 * The try_to_unmap() in page reclaim path might reach here too,
2857 	 * this may cause a race condition to corrupt deferred split queue.
2858 	 * And, if page reclaim is already handling the same page, it is
2859 	 * unnecessary to handle it again in shrinker.
2860 	 *
2861 	 * Check PageSwapCache to determine if the page is being
2862 	 * handled by page reclaim since THP swap would add the page into
2863 	 * swap cache before calling try_to_unmap().
2864 	 */
2865 	if (PageSwapCache(page))
2866 		return;
2867 
2868 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2869 	if (list_empty(page_deferred_list(page))) {
2870 		count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2871 		list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2872 		ds_queue->split_queue_len++;
2873 #ifdef CONFIG_MEMCG
2874 		if (memcg)
2875 			memcg_set_shrinker_bit(memcg, page_to_nid(page),
2876 					       deferred_split_shrinker.id);
2877 #endif
2878 	}
2879 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2880 }
2881 
2882 static unsigned long deferred_split_count(struct shrinker *shrink,
2883 		struct shrink_control *sc)
2884 {
2885 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2886 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2887 
2888 #ifdef CONFIG_MEMCG
2889 	if (sc->memcg)
2890 		ds_queue = &sc->memcg->deferred_split_queue;
2891 #endif
2892 	return READ_ONCE(ds_queue->split_queue_len);
2893 }
2894 
2895 static unsigned long deferred_split_scan(struct shrinker *shrink,
2896 		struct shrink_control *sc)
2897 {
2898 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2899 	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2900 	unsigned long flags;
2901 	LIST_HEAD(list), *pos, *next;
2902 	struct page *page;
2903 	int split = 0;
2904 
2905 #ifdef CONFIG_MEMCG
2906 	if (sc->memcg)
2907 		ds_queue = &sc->memcg->deferred_split_queue;
2908 #endif
2909 
2910 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2911 	/* Take pin on all head pages to avoid freeing them under us */
2912 	list_for_each_safe(pos, next, &ds_queue->split_queue) {
2913 		page = list_entry((void *)pos, struct page, mapping);
2914 		page = compound_head(page);
2915 		if (get_page_unless_zero(page)) {
2916 			list_move(page_deferred_list(page), &list);
2917 		} else {
2918 			/* We lost race with put_compound_page() */
2919 			list_del_init(page_deferred_list(page));
2920 			ds_queue->split_queue_len--;
2921 		}
2922 		if (!--sc->nr_to_scan)
2923 			break;
2924 	}
2925 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2926 
2927 	list_for_each_safe(pos, next, &list) {
2928 		page = list_entry((void *)pos, struct page, mapping);
2929 		if (!trylock_page(page))
2930 			goto next;
2931 		/* split_huge_page() removes page from list on success */
2932 		if (!split_huge_page(page))
2933 			split++;
2934 		unlock_page(page);
2935 next:
2936 		put_page(page);
2937 	}
2938 
2939 	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2940 	list_splice_tail(&list, &ds_queue->split_queue);
2941 	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2942 
2943 	/*
2944 	 * Stop shrinker if we didn't split any page, but the queue is empty.
2945 	 * This can happen if pages were freed under us.
2946 	 */
2947 	if (!split && list_empty(&ds_queue->split_queue))
2948 		return SHRINK_STOP;
2949 	return split;
2950 }
2951 
2952 static struct shrinker deferred_split_shrinker = {
2953 	.count_objects = deferred_split_count,
2954 	.scan_objects = deferred_split_scan,
2955 	.seeks = DEFAULT_SEEKS,
2956 	.flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2957 		 SHRINKER_NONSLAB,
2958 };
2959 
2960 #ifdef CONFIG_DEBUG_FS
2961 static int split_huge_pages_set(void *data, u64 val)
2962 {
2963 	struct zone *zone;
2964 	struct page *page;
2965 	unsigned long pfn, max_zone_pfn;
2966 	unsigned long total = 0, split = 0;
2967 
2968 	if (val != 1)
2969 		return -EINVAL;
2970 
2971 	for_each_populated_zone(zone) {
2972 		max_zone_pfn = zone_end_pfn(zone);
2973 		for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2974 			if (!pfn_valid(pfn))
2975 				continue;
2976 
2977 			page = pfn_to_page(pfn);
2978 			if (!get_page_unless_zero(page))
2979 				continue;
2980 
2981 			if (zone != page_zone(page))
2982 				goto next;
2983 
2984 			if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2985 				goto next;
2986 
2987 			total++;
2988 			lock_page(page);
2989 			if (!split_huge_page(page))
2990 				split++;
2991 			unlock_page(page);
2992 next:
2993 			put_page(page);
2994 		}
2995 	}
2996 
2997 	pr_info("%lu of %lu THP split\n", split, total);
2998 
2999 	return 0;
3000 }
3001 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3002 		"%llu\n");
3003 
3004 static int __init split_huge_pages_debugfs(void)
3005 {
3006 	debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3007 			    &split_huge_pages_fops);
3008 	return 0;
3009 }
3010 late_initcall(split_huge_pages_debugfs);
3011 #endif
3012 
3013 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3014 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3015 		struct page *page)
3016 {
3017 	struct vm_area_struct *vma = pvmw->vma;
3018 	struct mm_struct *mm = vma->vm_mm;
3019 	unsigned long address = pvmw->address;
3020 	pmd_t pmdval;
3021 	swp_entry_t entry;
3022 	pmd_t pmdswp;
3023 
3024 	if (!(pvmw->pmd && !pvmw->pte))
3025 		return;
3026 
3027 	flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3028 	pmdval = *pvmw->pmd;
3029 	pmdp_invalidate(vma, address, pvmw->pmd);
3030 	if (pmd_dirty(pmdval))
3031 		set_page_dirty(page);
3032 	entry = make_migration_entry(page, pmd_write(pmdval));
3033 	pmdswp = swp_entry_to_pmd(entry);
3034 	if (pmd_soft_dirty(pmdval))
3035 		pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3036 	set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3037 	page_remove_rmap(page, true);
3038 	put_page(page);
3039 }
3040 
3041 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3042 {
3043 	struct vm_area_struct *vma = pvmw->vma;
3044 	struct mm_struct *mm = vma->vm_mm;
3045 	unsigned long address = pvmw->address;
3046 	unsigned long mmun_start = address & HPAGE_PMD_MASK;
3047 	pmd_t pmde;
3048 	swp_entry_t entry;
3049 
3050 	if (!(pvmw->pmd && !pvmw->pte))
3051 		return;
3052 
3053 	entry = pmd_to_swp_entry(*pvmw->pmd);
3054 	get_page(new);
3055 	pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3056 	if (pmd_swp_soft_dirty(*pvmw->pmd))
3057 		pmde = pmd_mksoft_dirty(pmde);
3058 	if (is_write_migration_entry(entry))
3059 		pmde = maybe_pmd_mkwrite(pmde, vma);
3060 
3061 	flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3062 	if (PageAnon(new))
3063 		page_add_anon_rmap(new, vma, mmun_start, true);
3064 	else
3065 		page_add_file_rmap(new, true);
3066 	set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3067 	if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3068 		mlock_vma_page(new);
3069 	update_mmu_cache_pmd(vma, address, pvmw->pmd);
3070 }
3071 #endif
3072