xref: /openbmc/linux/mm/huge_memory.c (revision fc28ab18)
1 /*
2  *  Copyright (C) 2009  Red Hat, Inc.
3  *
4  *  This work is licensed under the terms of the GNU GPL, version 2. See
5  *  the COPYING file in the top-level directory.
6  */
7 
8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9 
10 #include <linux/mm.h>
11 #include <linux/sched.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 
34 #include <asm/tlb.h>
35 #include <asm/pgalloc.h>
36 #include "internal.h"
37 
38 /*
39  * By default transparent hugepage support is disabled in order that avoid
40  * to risk increase the memory footprint of applications without a guaranteed
41  * benefit. When transparent hugepage support is enabled, is for all mappings,
42  * and khugepaged scans all mappings.
43  * Defrag is invoked by khugepaged hugepage allocations and by page faults
44  * for all hugepage allocations.
45  */
46 unsigned long transparent_hugepage_flags __read_mostly =
47 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
48 	(1<<TRANSPARENT_HUGEPAGE_FLAG)|
49 #endif
50 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
51 	(1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
52 #endif
53 	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
54 	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
55 	(1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
56 
57 static struct shrinker deferred_split_shrinker;
58 
59 static atomic_t huge_zero_refcount;
60 struct page *huge_zero_page __read_mostly;
61 
62 static struct page *get_huge_zero_page(void)
63 {
64 	struct page *zero_page;
65 retry:
66 	if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
67 		return READ_ONCE(huge_zero_page);
68 
69 	zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
70 			HPAGE_PMD_ORDER);
71 	if (!zero_page) {
72 		count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
73 		return NULL;
74 	}
75 	count_vm_event(THP_ZERO_PAGE_ALLOC);
76 	preempt_disable();
77 	if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
78 		preempt_enable();
79 		__free_pages(zero_page, compound_order(zero_page));
80 		goto retry;
81 	}
82 
83 	/* We take additional reference here. It will be put back by shrinker */
84 	atomic_set(&huge_zero_refcount, 2);
85 	preempt_enable();
86 	return READ_ONCE(huge_zero_page);
87 }
88 
89 static void put_huge_zero_page(void)
90 {
91 	/*
92 	 * Counter should never go to zero here. Only shrinker can put
93 	 * last reference.
94 	 */
95 	BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
96 }
97 
98 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
99 {
100 	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
101 		return READ_ONCE(huge_zero_page);
102 
103 	if (!get_huge_zero_page())
104 		return NULL;
105 
106 	if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
107 		put_huge_zero_page();
108 
109 	return READ_ONCE(huge_zero_page);
110 }
111 
112 void mm_put_huge_zero_page(struct mm_struct *mm)
113 {
114 	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
115 		put_huge_zero_page();
116 }
117 
118 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
119 					struct shrink_control *sc)
120 {
121 	/* we can free zero page only if last reference remains */
122 	return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
123 }
124 
125 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
126 				       struct shrink_control *sc)
127 {
128 	if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
129 		struct page *zero_page = xchg(&huge_zero_page, NULL);
130 		BUG_ON(zero_page == NULL);
131 		__free_pages(zero_page, compound_order(zero_page));
132 		return HPAGE_PMD_NR;
133 	}
134 
135 	return 0;
136 }
137 
138 static struct shrinker huge_zero_page_shrinker = {
139 	.count_objects = shrink_huge_zero_page_count,
140 	.scan_objects = shrink_huge_zero_page_scan,
141 	.seeks = DEFAULT_SEEKS,
142 };
143 
144 #ifdef CONFIG_SYSFS
145 
146 static ssize_t triple_flag_store(struct kobject *kobj,
147 				 struct kobj_attribute *attr,
148 				 const char *buf, size_t count,
149 				 enum transparent_hugepage_flag enabled,
150 				 enum transparent_hugepage_flag deferred,
151 				 enum transparent_hugepage_flag req_madv)
152 {
153 	if (!memcmp("defer", buf,
154 		    min(sizeof("defer")-1, count))) {
155 		if (enabled == deferred)
156 			return -EINVAL;
157 		clear_bit(enabled, &transparent_hugepage_flags);
158 		clear_bit(req_madv, &transparent_hugepage_flags);
159 		set_bit(deferred, &transparent_hugepage_flags);
160 	} else if (!memcmp("always", buf,
161 		    min(sizeof("always")-1, count))) {
162 		clear_bit(deferred, &transparent_hugepage_flags);
163 		clear_bit(req_madv, &transparent_hugepage_flags);
164 		set_bit(enabled, &transparent_hugepage_flags);
165 	} else if (!memcmp("madvise", buf,
166 			   min(sizeof("madvise")-1, count))) {
167 		clear_bit(enabled, &transparent_hugepage_flags);
168 		clear_bit(deferred, &transparent_hugepage_flags);
169 		set_bit(req_madv, &transparent_hugepage_flags);
170 	} else if (!memcmp("never", buf,
171 			   min(sizeof("never")-1, count))) {
172 		clear_bit(enabled, &transparent_hugepage_flags);
173 		clear_bit(req_madv, &transparent_hugepage_flags);
174 		clear_bit(deferred, &transparent_hugepage_flags);
175 	} else
176 		return -EINVAL;
177 
178 	return count;
179 }
180 
181 static ssize_t enabled_show(struct kobject *kobj,
182 			    struct kobj_attribute *attr, char *buf)
183 {
184 	if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
185 		return sprintf(buf, "[always] madvise never\n");
186 	else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
187 		return sprintf(buf, "always [madvise] never\n");
188 	else
189 		return sprintf(buf, "always madvise [never]\n");
190 }
191 
192 static ssize_t enabled_store(struct kobject *kobj,
193 			     struct kobj_attribute *attr,
194 			     const char *buf, size_t count)
195 {
196 	ssize_t ret;
197 
198 	ret = triple_flag_store(kobj, attr, buf, count,
199 				TRANSPARENT_HUGEPAGE_FLAG,
200 				TRANSPARENT_HUGEPAGE_FLAG,
201 				TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
202 
203 	if (ret > 0) {
204 		int err = start_stop_khugepaged();
205 		if (err)
206 			ret = err;
207 	}
208 
209 	return ret;
210 }
211 static struct kobj_attribute enabled_attr =
212 	__ATTR(enabled, 0644, enabled_show, enabled_store);
213 
214 ssize_t single_hugepage_flag_show(struct kobject *kobj,
215 				struct kobj_attribute *attr, char *buf,
216 				enum transparent_hugepage_flag flag)
217 {
218 	return sprintf(buf, "%d\n",
219 		       !!test_bit(flag, &transparent_hugepage_flags));
220 }
221 
222 ssize_t single_hugepage_flag_store(struct kobject *kobj,
223 				 struct kobj_attribute *attr,
224 				 const char *buf, size_t count,
225 				 enum transparent_hugepage_flag flag)
226 {
227 	unsigned long value;
228 	int ret;
229 
230 	ret = kstrtoul(buf, 10, &value);
231 	if (ret < 0)
232 		return ret;
233 	if (value > 1)
234 		return -EINVAL;
235 
236 	if (value)
237 		set_bit(flag, &transparent_hugepage_flags);
238 	else
239 		clear_bit(flag, &transparent_hugepage_flags);
240 
241 	return count;
242 }
243 
244 /*
245  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
246  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
247  * memory just to allocate one more hugepage.
248  */
249 static ssize_t defrag_show(struct kobject *kobj,
250 			   struct kobj_attribute *attr, char *buf)
251 {
252 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
253 		return sprintf(buf, "[always] defer madvise never\n");
254 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
255 		return sprintf(buf, "always [defer] madvise never\n");
256 	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
257 		return sprintf(buf, "always defer [madvise] never\n");
258 	else
259 		return sprintf(buf, "always defer madvise [never]\n");
260 
261 }
262 static ssize_t defrag_store(struct kobject *kobj,
263 			    struct kobj_attribute *attr,
264 			    const char *buf, size_t count)
265 {
266 	return triple_flag_store(kobj, attr, buf, count,
267 				 TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
268 				 TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
269 				 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
270 }
271 static struct kobj_attribute defrag_attr =
272 	__ATTR(defrag, 0644, defrag_show, defrag_store);
273 
274 static ssize_t use_zero_page_show(struct kobject *kobj,
275 		struct kobj_attribute *attr, char *buf)
276 {
277 	return single_hugepage_flag_show(kobj, attr, buf,
278 				TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
279 }
280 static ssize_t use_zero_page_store(struct kobject *kobj,
281 		struct kobj_attribute *attr, const char *buf, size_t count)
282 {
283 	return single_hugepage_flag_store(kobj, attr, buf, count,
284 				 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
285 }
286 static struct kobj_attribute use_zero_page_attr =
287 	__ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
288 
289 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
290 		struct kobj_attribute *attr, char *buf)
291 {
292 	return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
293 }
294 static struct kobj_attribute hpage_pmd_size_attr =
295 	__ATTR_RO(hpage_pmd_size);
296 
297 #ifdef CONFIG_DEBUG_VM
298 static ssize_t debug_cow_show(struct kobject *kobj,
299 				struct kobj_attribute *attr, char *buf)
300 {
301 	return single_hugepage_flag_show(kobj, attr, buf,
302 				TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
303 }
304 static ssize_t debug_cow_store(struct kobject *kobj,
305 			       struct kobj_attribute *attr,
306 			       const char *buf, size_t count)
307 {
308 	return single_hugepage_flag_store(kobj, attr, buf, count,
309 				 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
310 }
311 static struct kobj_attribute debug_cow_attr =
312 	__ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
313 #endif /* CONFIG_DEBUG_VM */
314 
315 static struct attribute *hugepage_attr[] = {
316 	&enabled_attr.attr,
317 	&defrag_attr.attr,
318 	&use_zero_page_attr.attr,
319 	&hpage_pmd_size_attr.attr,
320 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
321 	&shmem_enabled_attr.attr,
322 #endif
323 #ifdef CONFIG_DEBUG_VM
324 	&debug_cow_attr.attr,
325 #endif
326 	NULL,
327 };
328 
329 static struct attribute_group hugepage_attr_group = {
330 	.attrs = hugepage_attr,
331 };
332 
333 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
334 {
335 	int err;
336 
337 	*hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
338 	if (unlikely(!*hugepage_kobj)) {
339 		pr_err("failed to create transparent hugepage kobject\n");
340 		return -ENOMEM;
341 	}
342 
343 	err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
344 	if (err) {
345 		pr_err("failed to register transparent hugepage group\n");
346 		goto delete_obj;
347 	}
348 
349 	err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
350 	if (err) {
351 		pr_err("failed to register transparent hugepage group\n");
352 		goto remove_hp_group;
353 	}
354 
355 	return 0;
356 
357 remove_hp_group:
358 	sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
359 delete_obj:
360 	kobject_put(*hugepage_kobj);
361 	return err;
362 }
363 
364 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
365 {
366 	sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
367 	sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
368 	kobject_put(hugepage_kobj);
369 }
370 #else
371 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
372 {
373 	return 0;
374 }
375 
376 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
377 {
378 }
379 #endif /* CONFIG_SYSFS */
380 
381 static int __init hugepage_init(void)
382 {
383 	int err;
384 	struct kobject *hugepage_kobj;
385 
386 	if (!has_transparent_hugepage()) {
387 		transparent_hugepage_flags = 0;
388 		return -EINVAL;
389 	}
390 
391 	/*
392 	 * hugepages can't be allocated by the buddy allocator
393 	 */
394 	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
395 	/*
396 	 * we use page->mapping and page->index in second tail page
397 	 * as list_head: assuming THP order >= 2
398 	 */
399 	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
400 
401 	err = hugepage_init_sysfs(&hugepage_kobj);
402 	if (err)
403 		goto err_sysfs;
404 
405 	err = khugepaged_init();
406 	if (err)
407 		goto err_slab;
408 
409 	err = register_shrinker(&huge_zero_page_shrinker);
410 	if (err)
411 		goto err_hzp_shrinker;
412 	err = register_shrinker(&deferred_split_shrinker);
413 	if (err)
414 		goto err_split_shrinker;
415 
416 	/*
417 	 * By default disable transparent hugepages on smaller systems,
418 	 * where the extra memory used could hurt more than TLB overhead
419 	 * is likely to save.  The admin can still enable it through /sys.
420 	 */
421 	if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
422 		transparent_hugepage_flags = 0;
423 		return 0;
424 	}
425 
426 	err = start_stop_khugepaged();
427 	if (err)
428 		goto err_khugepaged;
429 
430 	return 0;
431 err_khugepaged:
432 	unregister_shrinker(&deferred_split_shrinker);
433 err_split_shrinker:
434 	unregister_shrinker(&huge_zero_page_shrinker);
435 err_hzp_shrinker:
436 	khugepaged_destroy();
437 err_slab:
438 	hugepage_exit_sysfs(hugepage_kobj);
439 err_sysfs:
440 	return err;
441 }
442 subsys_initcall(hugepage_init);
443 
444 static int __init setup_transparent_hugepage(char *str)
445 {
446 	int ret = 0;
447 	if (!str)
448 		goto out;
449 	if (!strcmp(str, "always")) {
450 		set_bit(TRANSPARENT_HUGEPAGE_FLAG,
451 			&transparent_hugepage_flags);
452 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
453 			  &transparent_hugepage_flags);
454 		ret = 1;
455 	} else if (!strcmp(str, "madvise")) {
456 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
457 			  &transparent_hugepage_flags);
458 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
459 			&transparent_hugepage_flags);
460 		ret = 1;
461 	} else if (!strcmp(str, "never")) {
462 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
463 			  &transparent_hugepage_flags);
464 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
465 			  &transparent_hugepage_flags);
466 		ret = 1;
467 	}
468 out:
469 	if (!ret)
470 		pr_warn("transparent_hugepage= cannot parse, ignored\n");
471 	return ret;
472 }
473 __setup("transparent_hugepage=", setup_transparent_hugepage);
474 
475 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
476 {
477 	if (likely(vma->vm_flags & VM_WRITE))
478 		pmd = pmd_mkwrite(pmd);
479 	return pmd;
480 }
481 
482 static inline struct list_head *page_deferred_list(struct page *page)
483 {
484 	/*
485 	 * ->lru in the tail pages is occupied by compound_head.
486 	 * Let's use ->mapping + ->index in the second tail page as list_head.
487 	 */
488 	return (struct list_head *)&page[2].mapping;
489 }
490 
491 void prep_transhuge_page(struct page *page)
492 {
493 	/*
494 	 * we use page->mapping and page->indexlru in second tail page
495 	 * as list_head: assuming THP order >= 2
496 	 */
497 
498 	INIT_LIST_HEAD(page_deferred_list(page));
499 	set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
500 }
501 
502 unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
503 		loff_t off, unsigned long flags, unsigned long size)
504 {
505 	unsigned long addr;
506 	loff_t off_end = off + len;
507 	loff_t off_align = round_up(off, size);
508 	unsigned long len_pad;
509 
510 	if (off_end <= off_align || (off_end - off_align) < size)
511 		return 0;
512 
513 	len_pad = len + size;
514 	if (len_pad < len || (off + len_pad) < off)
515 		return 0;
516 
517 	addr = current->mm->get_unmapped_area(filp, 0, len_pad,
518 					      off >> PAGE_SHIFT, flags);
519 	if (IS_ERR_VALUE(addr))
520 		return 0;
521 
522 	addr += (off - addr) & (size - 1);
523 	return addr;
524 }
525 
526 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
527 		unsigned long len, unsigned long pgoff, unsigned long flags)
528 {
529 	loff_t off = (loff_t)pgoff << PAGE_SHIFT;
530 
531 	if (addr)
532 		goto out;
533 	if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
534 		goto out;
535 
536 	addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
537 	if (addr)
538 		return addr;
539 
540  out:
541 	return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
542 }
543 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
544 
545 static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page,
546 		gfp_t gfp)
547 {
548 	struct vm_area_struct *vma = vmf->vma;
549 	struct mem_cgroup *memcg;
550 	pgtable_t pgtable;
551 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
552 
553 	VM_BUG_ON_PAGE(!PageCompound(page), page);
554 
555 	if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
556 		put_page(page);
557 		count_vm_event(THP_FAULT_FALLBACK);
558 		return VM_FAULT_FALLBACK;
559 	}
560 
561 	pgtable = pte_alloc_one(vma->vm_mm, haddr);
562 	if (unlikely(!pgtable)) {
563 		mem_cgroup_cancel_charge(page, memcg, true);
564 		put_page(page);
565 		return VM_FAULT_OOM;
566 	}
567 
568 	clear_huge_page(page, haddr, HPAGE_PMD_NR);
569 	/*
570 	 * The memory barrier inside __SetPageUptodate makes sure that
571 	 * clear_huge_page writes become visible before the set_pmd_at()
572 	 * write.
573 	 */
574 	__SetPageUptodate(page);
575 
576 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
577 	if (unlikely(!pmd_none(*vmf->pmd))) {
578 		spin_unlock(vmf->ptl);
579 		mem_cgroup_cancel_charge(page, memcg, true);
580 		put_page(page);
581 		pte_free(vma->vm_mm, pgtable);
582 	} else {
583 		pmd_t entry;
584 
585 		/* Deliver the page fault to userland */
586 		if (userfaultfd_missing(vma)) {
587 			int ret;
588 
589 			spin_unlock(vmf->ptl);
590 			mem_cgroup_cancel_charge(page, memcg, true);
591 			put_page(page);
592 			pte_free(vma->vm_mm, pgtable);
593 			ret = handle_userfault(vmf, VM_UFFD_MISSING);
594 			VM_BUG_ON(ret & VM_FAULT_FALLBACK);
595 			return ret;
596 		}
597 
598 		entry = mk_huge_pmd(page, vma->vm_page_prot);
599 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
600 		page_add_new_anon_rmap(page, vma, haddr, true);
601 		mem_cgroup_commit_charge(page, memcg, false, true);
602 		lru_cache_add_active_or_unevictable(page, vma);
603 		pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
604 		set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
605 		add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
606 		atomic_long_inc(&vma->vm_mm->nr_ptes);
607 		spin_unlock(vmf->ptl);
608 		count_vm_event(THP_FAULT_ALLOC);
609 	}
610 
611 	return 0;
612 }
613 
614 /*
615  * If THP defrag is set to always then directly reclaim/compact as necessary
616  * If set to defer then do only background reclaim/compact and defer to khugepaged
617  * If set to madvise and the VMA is flagged then directly reclaim/compact
618  * When direct reclaim/compact is allowed, don't retry except for flagged VMA's
619  */
620 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
621 {
622 	bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
623 
624 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
625 				&transparent_hugepage_flags) && vma_madvised)
626 		return GFP_TRANSHUGE;
627 	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
628 						&transparent_hugepage_flags))
629 		return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
630 	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
631 						&transparent_hugepage_flags))
632 		return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
633 
634 	return GFP_TRANSHUGE_LIGHT;
635 }
636 
637 /* Caller must hold page table lock. */
638 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
639 		struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
640 		struct page *zero_page)
641 {
642 	pmd_t entry;
643 	if (!pmd_none(*pmd))
644 		return false;
645 	entry = mk_pmd(zero_page, vma->vm_page_prot);
646 	entry = pmd_mkhuge(entry);
647 	if (pgtable)
648 		pgtable_trans_huge_deposit(mm, pmd, pgtable);
649 	set_pmd_at(mm, haddr, pmd, entry);
650 	atomic_long_inc(&mm->nr_ptes);
651 	return true;
652 }
653 
654 int do_huge_pmd_anonymous_page(struct vm_fault *vmf)
655 {
656 	struct vm_area_struct *vma = vmf->vma;
657 	gfp_t gfp;
658 	struct page *page;
659 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
660 
661 	if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
662 		return VM_FAULT_FALLBACK;
663 	if (unlikely(anon_vma_prepare(vma)))
664 		return VM_FAULT_OOM;
665 	if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
666 		return VM_FAULT_OOM;
667 	if (!(vmf->flags & FAULT_FLAG_WRITE) &&
668 			!mm_forbids_zeropage(vma->vm_mm) &&
669 			transparent_hugepage_use_zero_page()) {
670 		pgtable_t pgtable;
671 		struct page *zero_page;
672 		bool set;
673 		int ret;
674 		pgtable = pte_alloc_one(vma->vm_mm, haddr);
675 		if (unlikely(!pgtable))
676 			return VM_FAULT_OOM;
677 		zero_page = mm_get_huge_zero_page(vma->vm_mm);
678 		if (unlikely(!zero_page)) {
679 			pte_free(vma->vm_mm, pgtable);
680 			count_vm_event(THP_FAULT_FALLBACK);
681 			return VM_FAULT_FALLBACK;
682 		}
683 		vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
684 		ret = 0;
685 		set = false;
686 		if (pmd_none(*vmf->pmd)) {
687 			if (userfaultfd_missing(vma)) {
688 				spin_unlock(vmf->ptl);
689 				ret = handle_userfault(vmf, VM_UFFD_MISSING);
690 				VM_BUG_ON(ret & VM_FAULT_FALLBACK);
691 			} else {
692 				set_huge_zero_page(pgtable, vma->vm_mm, vma,
693 						   haddr, vmf->pmd, zero_page);
694 				spin_unlock(vmf->ptl);
695 				set = true;
696 			}
697 		} else
698 			spin_unlock(vmf->ptl);
699 		if (!set)
700 			pte_free(vma->vm_mm, pgtable);
701 		return ret;
702 	}
703 	gfp = alloc_hugepage_direct_gfpmask(vma);
704 	page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
705 	if (unlikely(!page)) {
706 		count_vm_event(THP_FAULT_FALLBACK);
707 		return VM_FAULT_FALLBACK;
708 	}
709 	prep_transhuge_page(page);
710 	return __do_huge_pmd_anonymous_page(vmf, page, gfp);
711 }
712 
713 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
714 		pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
715 {
716 	struct mm_struct *mm = vma->vm_mm;
717 	pmd_t entry;
718 	spinlock_t *ptl;
719 
720 	ptl = pmd_lock(mm, pmd);
721 	entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
722 	if (pfn_t_devmap(pfn))
723 		entry = pmd_mkdevmap(entry);
724 	if (write) {
725 		entry = pmd_mkyoung(pmd_mkdirty(entry));
726 		entry = maybe_pmd_mkwrite(entry, vma);
727 	}
728 	set_pmd_at(mm, addr, pmd, entry);
729 	update_mmu_cache_pmd(vma, addr, pmd);
730 	spin_unlock(ptl);
731 }
732 
733 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
734 			pmd_t *pmd, pfn_t pfn, bool write)
735 {
736 	pgprot_t pgprot = vma->vm_page_prot;
737 	/*
738 	 * If we had pmd_special, we could avoid all these restrictions,
739 	 * but we need to be consistent with PTEs and architectures that
740 	 * can't support a 'special' bit.
741 	 */
742 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
743 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
744 						(VM_PFNMAP|VM_MIXEDMAP));
745 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
746 	BUG_ON(!pfn_t_devmap(pfn));
747 
748 	if (addr < vma->vm_start || addr >= vma->vm_end)
749 		return VM_FAULT_SIGBUS;
750 
751 	track_pfn_insert(vma, &pgprot, pfn);
752 
753 	insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
754 	return VM_FAULT_NOPAGE;
755 }
756 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
757 
758 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
759 		pmd_t *pmd)
760 {
761 	pmd_t _pmd;
762 
763 	/*
764 	 * We should set the dirty bit only for FOLL_WRITE but for now
765 	 * the dirty bit in the pmd is meaningless.  And if the dirty
766 	 * bit will become meaningful and we'll only set it with
767 	 * FOLL_WRITE, an atomic set_bit will be required on the pmd to
768 	 * set the young bit, instead of the current set_pmd_at.
769 	 */
770 	_pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
771 	if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
772 				pmd, _pmd,  1))
773 		update_mmu_cache_pmd(vma, addr, pmd);
774 }
775 
776 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
777 		pmd_t *pmd, int flags)
778 {
779 	unsigned long pfn = pmd_pfn(*pmd);
780 	struct mm_struct *mm = vma->vm_mm;
781 	struct dev_pagemap *pgmap;
782 	struct page *page;
783 
784 	assert_spin_locked(pmd_lockptr(mm, pmd));
785 
786 	if (flags & FOLL_WRITE && !pmd_write(*pmd))
787 		return NULL;
788 
789 	if (pmd_present(*pmd) && pmd_devmap(*pmd))
790 		/* pass */;
791 	else
792 		return NULL;
793 
794 	if (flags & FOLL_TOUCH)
795 		touch_pmd(vma, addr, pmd);
796 
797 	/*
798 	 * device mapped pages can only be returned if the
799 	 * caller will manage the page reference count.
800 	 */
801 	if (!(flags & FOLL_GET))
802 		return ERR_PTR(-EEXIST);
803 
804 	pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
805 	pgmap = get_dev_pagemap(pfn, NULL);
806 	if (!pgmap)
807 		return ERR_PTR(-EFAULT);
808 	page = pfn_to_page(pfn);
809 	get_page(page);
810 	put_dev_pagemap(pgmap);
811 
812 	return page;
813 }
814 
815 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
816 		  pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
817 		  struct vm_area_struct *vma)
818 {
819 	spinlock_t *dst_ptl, *src_ptl;
820 	struct page *src_page;
821 	pmd_t pmd;
822 	pgtable_t pgtable = NULL;
823 	int ret = -ENOMEM;
824 
825 	/* Skip if can be re-fill on fault */
826 	if (!vma_is_anonymous(vma))
827 		return 0;
828 
829 	pgtable = pte_alloc_one(dst_mm, addr);
830 	if (unlikely(!pgtable))
831 		goto out;
832 
833 	dst_ptl = pmd_lock(dst_mm, dst_pmd);
834 	src_ptl = pmd_lockptr(src_mm, src_pmd);
835 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
836 
837 	ret = -EAGAIN;
838 	pmd = *src_pmd;
839 	if (unlikely(!pmd_trans_huge(pmd))) {
840 		pte_free(dst_mm, pgtable);
841 		goto out_unlock;
842 	}
843 	/*
844 	 * When page table lock is held, the huge zero pmd should not be
845 	 * under splitting since we don't split the page itself, only pmd to
846 	 * a page table.
847 	 */
848 	if (is_huge_zero_pmd(pmd)) {
849 		struct page *zero_page;
850 		/*
851 		 * get_huge_zero_page() will never allocate a new page here,
852 		 * since we already have a zero page to copy. It just takes a
853 		 * reference.
854 		 */
855 		zero_page = mm_get_huge_zero_page(dst_mm);
856 		set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
857 				zero_page);
858 		ret = 0;
859 		goto out_unlock;
860 	}
861 
862 	src_page = pmd_page(pmd);
863 	VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
864 	get_page(src_page);
865 	page_dup_rmap(src_page, true);
866 	add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
867 	atomic_long_inc(&dst_mm->nr_ptes);
868 	pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
869 
870 	pmdp_set_wrprotect(src_mm, addr, src_pmd);
871 	pmd = pmd_mkold(pmd_wrprotect(pmd));
872 	set_pmd_at(dst_mm, addr, dst_pmd, pmd);
873 
874 	ret = 0;
875 out_unlock:
876 	spin_unlock(src_ptl);
877 	spin_unlock(dst_ptl);
878 out:
879 	return ret;
880 }
881 
882 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
883 {
884 	pmd_t entry;
885 	unsigned long haddr;
886 	bool write = vmf->flags & FAULT_FLAG_WRITE;
887 
888 	vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
889 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
890 		goto unlock;
891 
892 	entry = pmd_mkyoung(orig_pmd);
893 	if (write)
894 		entry = pmd_mkdirty(entry);
895 	haddr = vmf->address & HPAGE_PMD_MASK;
896 	if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
897 		update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
898 
899 unlock:
900 	spin_unlock(vmf->ptl);
901 }
902 
903 static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd,
904 		struct page *page)
905 {
906 	struct vm_area_struct *vma = vmf->vma;
907 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
908 	struct mem_cgroup *memcg;
909 	pgtable_t pgtable;
910 	pmd_t _pmd;
911 	int ret = 0, i;
912 	struct page **pages;
913 	unsigned long mmun_start;	/* For mmu_notifiers */
914 	unsigned long mmun_end;		/* For mmu_notifiers */
915 
916 	pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
917 			GFP_KERNEL);
918 	if (unlikely(!pages)) {
919 		ret |= VM_FAULT_OOM;
920 		goto out;
921 	}
922 
923 	for (i = 0; i < HPAGE_PMD_NR; i++) {
924 		pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
925 					       vmf->address, page_to_nid(page));
926 		if (unlikely(!pages[i] ||
927 			     mem_cgroup_try_charge(pages[i], vma->vm_mm,
928 				     GFP_KERNEL, &memcg, false))) {
929 			if (pages[i])
930 				put_page(pages[i]);
931 			while (--i >= 0) {
932 				memcg = (void *)page_private(pages[i]);
933 				set_page_private(pages[i], 0);
934 				mem_cgroup_cancel_charge(pages[i], memcg,
935 						false);
936 				put_page(pages[i]);
937 			}
938 			kfree(pages);
939 			ret |= VM_FAULT_OOM;
940 			goto out;
941 		}
942 		set_page_private(pages[i], (unsigned long)memcg);
943 	}
944 
945 	for (i = 0; i < HPAGE_PMD_NR; i++) {
946 		copy_user_highpage(pages[i], page + i,
947 				   haddr + PAGE_SIZE * i, vma);
948 		__SetPageUptodate(pages[i]);
949 		cond_resched();
950 	}
951 
952 	mmun_start = haddr;
953 	mmun_end   = haddr + HPAGE_PMD_SIZE;
954 	mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
955 
956 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
957 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
958 		goto out_free_pages;
959 	VM_BUG_ON_PAGE(!PageHead(page), page);
960 
961 	pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
962 	/* leave pmd empty until pte is filled */
963 
964 	pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
965 	pmd_populate(vma->vm_mm, &_pmd, pgtable);
966 
967 	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
968 		pte_t entry;
969 		entry = mk_pte(pages[i], vma->vm_page_prot);
970 		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
971 		memcg = (void *)page_private(pages[i]);
972 		set_page_private(pages[i], 0);
973 		page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
974 		mem_cgroup_commit_charge(pages[i], memcg, false, false);
975 		lru_cache_add_active_or_unevictable(pages[i], vma);
976 		vmf->pte = pte_offset_map(&_pmd, haddr);
977 		VM_BUG_ON(!pte_none(*vmf->pte));
978 		set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
979 		pte_unmap(vmf->pte);
980 	}
981 	kfree(pages);
982 
983 	smp_wmb(); /* make pte visible before pmd */
984 	pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
985 	page_remove_rmap(page, true);
986 	spin_unlock(vmf->ptl);
987 
988 	mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
989 
990 	ret |= VM_FAULT_WRITE;
991 	put_page(page);
992 
993 out:
994 	return ret;
995 
996 out_free_pages:
997 	spin_unlock(vmf->ptl);
998 	mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
999 	for (i = 0; i < HPAGE_PMD_NR; i++) {
1000 		memcg = (void *)page_private(pages[i]);
1001 		set_page_private(pages[i], 0);
1002 		mem_cgroup_cancel_charge(pages[i], memcg, false);
1003 		put_page(pages[i]);
1004 	}
1005 	kfree(pages);
1006 	goto out;
1007 }
1008 
1009 int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1010 {
1011 	struct vm_area_struct *vma = vmf->vma;
1012 	struct page *page = NULL, *new_page;
1013 	struct mem_cgroup *memcg;
1014 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1015 	unsigned long mmun_start;	/* For mmu_notifiers */
1016 	unsigned long mmun_end;		/* For mmu_notifiers */
1017 	gfp_t huge_gfp;			/* for allocation and charge */
1018 	int ret = 0;
1019 
1020 	vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1021 	VM_BUG_ON_VMA(!vma->anon_vma, vma);
1022 	if (is_huge_zero_pmd(orig_pmd))
1023 		goto alloc;
1024 	spin_lock(vmf->ptl);
1025 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1026 		goto out_unlock;
1027 
1028 	page = pmd_page(orig_pmd);
1029 	VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1030 	/*
1031 	 * We can only reuse the page if nobody else maps the huge page or it's
1032 	 * part.
1033 	 */
1034 	if (page_trans_huge_mapcount(page, NULL) == 1) {
1035 		pmd_t entry;
1036 		entry = pmd_mkyoung(orig_pmd);
1037 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1038 		if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1039 			update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1040 		ret |= VM_FAULT_WRITE;
1041 		goto out_unlock;
1042 	}
1043 	get_page(page);
1044 	spin_unlock(vmf->ptl);
1045 alloc:
1046 	if (transparent_hugepage_enabled(vma) &&
1047 	    !transparent_hugepage_debug_cow()) {
1048 		huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1049 		new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1050 	} else
1051 		new_page = NULL;
1052 
1053 	if (likely(new_page)) {
1054 		prep_transhuge_page(new_page);
1055 	} else {
1056 		if (!page) {
1057 			split_huge_pmd(vma, vmf->pmd, vmf->address);
1058 			ret |= VM_FAULT_FALLBACK;
1059 		} else {
1060 			ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1061 			if (ret & VM_FAULT_OOM) {
1062 				split_huge_pmd(vma, vmf->pmd, vmf->address);
1063 				ret |= VM_FAULT_FALLBACK;
1064 			}
1065 			put_page(page);
1066 		}
1067 		count_vm_event(THP_FAULT_FALLBACK);
1068 		goto out;
1069 	}
1070 
1071 	if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1072 					huge_gfp, &memcg, true))) {
1073 		put_page(new_page);
1074 		split_huge_pmd(vma, vmf->pmd, vmf->address);
1075 		if (page)
1076 			put_page(page);
1077 		ret |= VM_FAULT_FALLBACK;
1078 		count_vm_event(THP_FAULT_FALLBACK);
1079 		goto out;
1080 	}
1081 
1082 	count_vm_event(THP_FAULT_ALLOC);
1083 
1084 	if (!page)
1085 		clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1086 	else
1087 		copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1088 	__SetPageUptodate(new_page);
1089 
1090 	mmun_start = haddr;
1091 	mmun_end   = haddr + HPAGE_PMD_SIZE;
1092 	mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1093 
1094 	spin_lock(vmf->ptl);
1095 	if (page)
1096 		put_page(page);
1097 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1098 		spin_unlock(vmf->ptl);
1099 		mem_cgroup_cancel_charge(new_page, memcg, true);
1100 		put_page(new_page);
1101 		goto out_mn;
1102 	} else {
1103 		pmd_t entry;
1104 		entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1105 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1106 		pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1107 		page_add_new_anon_rmap(new_page, vma, haddr, true);
1108 		mem_cgroup_commit_charge(new_page, memcg, false, true);
1109 		lru_cache_add_active_or_unevictable(new_page, vma);
1110 		set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1111 		update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1112 		if (!page) {
1113 			add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1114 		} else {
1115 			VM_BUG_ON_PAGE(!PageHead(page), page);
1116 			page_remove_rmap(page, true);
1117 			put_page(page);
1118 		}
1119 		ret |= VM_FAULT_WRITE;
1120 	}
1121 	spin_unlock(vmf->ptl);
1122 out_mn:
1123 	mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1124 out:
1125 	return ret;
1126 out_unlock:
1127 	spin_unlock(vmf->ptl);
1128 	return ret;
1129 }
1130 
1131 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1132 				   unsigned long addr,
1133 				   pmd_t *pmd,
1134 				   unsigned int flags)
1135 {
1136 	struct mm_struct *mm = vma->vm_mm;
1137 	struct page *page = NULL;
1138 
1139 	assert_spin_locked(pmd_lockptr(mm, pmd));
1140 
1141 	if (flags & FOLL_WRITE && !pmd_write(*pmd))
1142 		goto out;
1143 
1144 	/* Avoid dumping huge zero page */
1145 	if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1146 		return ERR_PTR(-EFAULT);
1147 
1148 	/* Full NUMA hinting faults to serialise migration in fault paths */
1149 	if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1150 		goto out;
1151 
1152 	page = pmd_page(*pmd);
1153 	VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1154 	if (flags & FOLL_TOUCH)
1155 		touch_pmd(vma, addr, pmd);
1156 	if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1157 		/*
1158 		 * We don't mlock() pte-mapped THPs. This way we can avoid
1159 		 * leaking mlocked pages into non-VM_LOCKED VMAs.
1160 		 *
1161 		 * For anon THP:
1162 		 *
1163 		 * In most cases the pmd is the only mapping of the page as we
1164 		 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1165 		 * writable private mappings in populate_vma_page_range().
1166 		 *
1167 		 * The only scenario when we have the page shared here is if we
1168 		 * mlocking read-only mapping shared over fork(). We skip
1169 		 * mlocking such pages.
1170 		 *
1171 		 * For file THP:
1172 		 *
1173 		 * We can expect PageDoubleMap() to be stable under page lock:
1174 		 * for file pages we set it in page_add_file_rmap(), which
1175 		 * requires page to be locked.
1176 		 */
1177 
1178 		if (PageAnon(page) && compound_mapcount(page) != 1)
1179 			goto skip_mlock;
1180 		if (PageDoubleMap(page) || !page->mapping)
1181 			goto skip_mlock;
1182 		if (!trylock_page(page))
1183 			goto skip_mlock;
1184 		lru_add_drain();
1185 		if (page->mapping && !PageDoubleMap(page))
1186 			mlock_vma_page(page);
1187 		unlock_page(page);
1188 	}
1189 skip_mlock:
1190 	page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1191 	VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1192 	if (flags & FOLL_GET)
1193 		get_page(page);
1194 
1195 out:
1196 	return page;
1197 }
1198 
1199 /* NUMA hinting page fault entry point for trans huge pmds */
1200 int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1201 {
1202 	struct vm_area_struct *vma = vmf->vma;
1203 	struct anon_vma *anon_vma = NULL;
1204 	struct page *page;
1205 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1206 	int page_nid = -1, this_nid = numa_node_id();
1207 	int target_nid, last_cpupid = -1;
1208 	bool page_locked;
1209 	bool migrated = false;
1210 	bool was_writable;
1211 	int flags = 0;
1212 
1213 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1214 	if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1215 		goto out_unlock;
1216 
1217 	/*
1218 	 * If there are potential migrations, wait for completion and retry
1219 	 * without disrupting NUMA hinting information. Do not relock and
1220 	 * check_same as the page may no longer be mapped.
1221 	 */
1222 	if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1223 		page = pmd_page(*vmf->pmd);
1224 		spin_unlock(vmf->ptl);
1225 		wait_on_page_locked(page);
1226 		goto out;
1227 	}
1228 
1229 	page = pmd_page(pmd);
1230 	BUG_ON(is_huge_zero_page(page));
1231 	page_nid = page_to_nid(page);
1232 	last_cpupid = page_cpupid_last(page);
1233 	count_vm_numa_event(NUMA_HINT_FAULTS);
1234 	if (page_nid == this_nid) {
1235 		count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1236 		flags |= TNF_FAULT_LOCAL;
1237 	}
1238 
1239 	/* See similar comment in do_numa_page for explanation */
1240 	if (!pmd_write(pmd))
1241 		flags |= TNF_NO_GROUP;
1242 
1243 	/*
1244 	 * Acquire the page lock to serialise THP migrations but avoid dropping
1245 	 * page_table_lock if at all possible
1246 	 */
1247 	page_locked = trylock_page(page);
1248 	target_nid = mpol_misplaced(page, vma, haddr);
1249 	if (target_nid == -1) {
1250 		/* If the page was locked, there are no parallel migrations */
1251 		if (page_locked)
1252 			goto clear_pmdnuma;
1253 	}
1254 
1255 	/* Migration could have started since the pmd_trans_migrating check */
1256 	if (!page_locked) {
1257 		spin_unlock(vmf->ptl);
1258 		wait_on_page_locked(page);
1259 		page_nid = -1;
1260 		goto out;
1261 	}
1262 
1263 	/*
1264 	 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1265 	 * to serialises splits
1266 	 */
1267 	get_page(page);
1268 	spin_unlock(vmf->ptl);
1269 	anon_vma = page_lock_anon_vma_read(page);
1270 
1271 	/* Confirm the PMD did not change while page_table_lock was released */
1272 	spin_lock(vmf->ptl);
1273 	if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1274 		unlock_page(page);
1275 		put_page(page);
1276 		page_nid = -1;
1277 		goto out_unlock;
1278 	}
1279 
1280 	/* Bail if we fail to protect against THP splits for any reason */
1281 	if (unlikely(!anon_vma)) {
1282 		put_page(page);
1283 		page_nid = -1;
1284 		goto clear_pmdnuma;
1285 	}
1286 
1287 	/*
1288 	 * Migrate the THP to the requested node, returns with page unlocked
1289 	 * and access rights restored.
1290 	 */
1291 	spin_unlock(vmf->ptl);
1292 	migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1293 				vmf->pmd, pmd, vmf->address, page, target_nid);
1294 	if (migrated) {
1295 		flags |= TNF_MIGRATED;
1296 		page_nid = target_nid;
1297 	} else
1298 		flags |= TNF_MIGRATE_FAIL;
1299 
1300 	goto out;
1301 clear_pmdnuma:
1302 	BUG_ON(!PageLocked(page));
1303 	was_writable = pmd_write(pmd);
1304 	pmd = pmd_modify(pmd, vma->vm_page_prot);
1305 	pmd = pmd_mkyoung(pmd);
1306 	if (was_writable)
1307 		pmd = pmd_mkwrite(pmd);
1308 	set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1309 	update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1310 	unlock_page(page);
1311 out_unlock:
1312 	spin_unlock(vmf->ptl);
1313 
1314 out:
1315 	if (anon_vma)
1316 		page_unlock_anon_vma_read(anon_vma);
1317 
1318 	if (page_nid != -1)
1319 		task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1320 				vmf->flags);
1321 
1322 	return 0;
1323 }
1324 
1325 /*
1326  * Return true if we do MADV_FREE successfully on entire pmd page.
1327  * Otherwise, return false.
1328  */
1329 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1330 		pmd_t *pmd, unsigned long addr, unsigned long next)
1331 {
1332 	spinlock_t *ptl;
1333 	pmd_t orig_pmd;
1334 	struct page *page;
1335 	struct mm_struct *mm = tlb->mm;
1336 	bool ret = false;
1337 
1338 	tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1339 
1340 	ptl = pmd_trans_huge_lock(pmd, vma);
1341 	if (!ptl)
1342 		goto out_unlocked;
1343 
1344 	orig_pmd = *pmd;
1345 	if (is_huge_zero_pmd(orig_pmd))
1346 		goto out;
1347 
1348 	page = pmd_page(orig_pmd);
1349 	/*
1350 	 * If other processes are mapping this page, we couldn't discard
1351 	 * the page unless they all do MADV_FREE so let's skip the page.
1352 	 */
1353 	if (page_mapcount(page) != 1)
1354 		goto out;
1355 
1356 	if (!trylock_page(page))
1357 		goto out;
1358 
1359 	/*
1360 	 * If user want to discard part-pages of THP, split it so MADV_FREE
1361 	 * will deactivate only them.
1362 	 */
1363 	if (next - addr != HPAGE_PMD_SIZE) {
1364 		get_page(page);
1365 		spin_unlock(ptl);
1366 		split_huge_page(page);
1367 		put_page(page);
1368 		unlock_page(page);
1369 		goto out_unlocked;
1370 	}
1371 
1372 	if (PageDirty(page))
1373 		ClearPageDirty(page);
1374 	unlock_page(page);
1375 
1376 	if (PageActive(page))
1377 		deactivate_page(page);
1378 
1379 	if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1380 		orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1381 			tlb->fullmm);
1382 		orig_pmd = pmd_mkold(orig_pmd);
1383 		orig_pmd = pmd_mkclean(orig_pmd);
1384 
1385 		set_pmd_at(mm, addr, pmd, orig_pmd);
1386 		tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1387 	}
1388 	ret = true;
1389 out:
1390 	spin_unlock(ptl);
1391 out_unlocked:
1392 	return ret;
1393 }
1394 
1395 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1396 {
1397 	pgtable_t pgtable;
1398 
1399 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1400 	pte_free(mm, pgtable);
1401 	atomic_long_dec(&mm->nr_ptes);
1402 }
1403 
1404 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1405 		 pmd_t *pmd, unsigned long addr)
1406 {
1407 	pmd_t orig_pmd;
1408 	spinlock_t *ptl;
1409 
1410 	tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1411 
1412 	ptl = __pmd_trans_huge_lock(pmd, vma);
1413 	if (!ptl)
1414 		return 0;
1415 	/*
1416 	 * For architectures like ppc64 we look at deposited pgtable
1417 	 * when calling pmdp_huge_get_and_clear. So do the
1418 	 * pgtable_trans_huge_withdraw after finishing pmdp related
1419 	 * operations.
1420 	 */
1421 	orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1422 			tlb->fullmm);
1423 	tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1424 	if (vma_is_dax(vma)) {
1425 		spin_unlock(ptl);
1426 		if (is_huge_zero_pmd(orig_pmd))
1427 			tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1428 	} else if (is_huge_zero_pmd(orig_pmd)) {
1429 		pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1430 		atomic_long_dec(&tlb->mm->nr_ptes);
1431 		spin_unlock(ptl);
1432 		tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1433 	} else {
1434 		struct page *page = pmd_page(orig_pmd);
1435 		page_remove_rmap(page, true);
1436 		VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1437 		VM_BUG_ON_PAGE(!PageHead(page), page);
1438 		if (PageAnon(page)) {
1439 			pgtable_t pgtable;
1440 			pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1441 			pte_free(tlb->mm, pgtable);
1442 			atomic_long_dec(&tlb->mm->nr_ptes);
1443 			add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1444 		} else {
1445 			if (arch_needs_pgtable_deposit())
1446 				zap_deposited_table(tlb->mm, pmd);
1447 			add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1448 		}
1449 		spin_unlock(ptl);
1450 		tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1451 	}
1452 	return 1;
1453 }
1454 
1455 #ifndef pmd_move_must_withdraw
1456 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1457 					 spinlock_t *old_pmd_ptl,
1458 					 struct vm_area_struct *vma)
1459 {
1460 	/*
1461 	 * With split pmd lock we also need to move preallocated
1462 	 * PTE page table if new_pmd is on different PMD page table.
1463 	 *
1464 	 * We also don't deposit and withdraw tables for file pages.
1465 	 */
1466 	return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1467 }
1468 #endif
1469 
1470 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1471 		  unsigned long new_addr, unsigned long old_end,
1472 		  pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
1473 {
1474 	spinlock_t *old_ptl, *new_ptl;
1475 	pmd_t pmd;
1476 	struct mm_struct *mm = vma->vm_mm;
1477 	bool force_flush = false;
1478 
1479 	if ((old_addr & ~HPAGE_PMD_MASK) ||
1480 	    (new_addr & ~HPAGE_PMD_MASK) ||
1481 	    old_end - old_addr < HPAGE_PMD_SIZE)
1482 		return false;
1483 
1484 	/*
1485 	 * The destination pmd shouldn't be established, free_pgtables()
1486 	 * should have release it.
1487 	 */
1488 	if (WARN_ON(!pmd_none(*new_pmd))) {
1489 		VM_BUG_ON(pmd_trans_huge(*new_pmd));
1490 		return false;
1491 	}
1492 
1493 	/*
1494 	 * We don't have to worry about the ordering of src and dst
1495 	 * ptlocks because exclusive mmap_sem prevents deadlock.
1496 	 */
1497 	old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1498 	if (old_ptl) {
1499 		new_ptl = pmd_lockptr(mm, new_pmd);
1500 		if (new_ptl != old_ptl)
1501 			spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1502 		pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1503 		if (pmd_present(pmd) && pmd_dirty(pmd))
1504 			force_flush = true;
1505 		VM_BUG_ON(!pmd_none(*new_pmd));
1506 
1507 		if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1508 			pgtable_t pgtable;
1509 			pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1510 			pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1511 		}
1512 		set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1513 		if (new_ptl != old_ptl)
1514 			spin_unlock(new_ptl);
1515 		if (force_flush)
1516 			flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1517 		else
1518 			*need_flush = true;
1519 		spin_unlock(old_ptl);
1520 		return true;
1521 	}
1522 	return false;
1523 }
1524 
1525 /*
1526  * Returns
1527  *  - 0 if PMD could not be locked
1528  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1529  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1530  */
1531 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1532 		unsigned long addr, pgprot_t newprot, int prot_numa)
1533 {
1534 	struct mm_struct *mm = vma->vm_mm;
1535 	spinlock_t *ptl;
1536 	int ret = 0;
1537 
1538 	ptl = __pmd_trans_huge_lock(pmd, vma);
1539 	if (ptl) {
1540 		pmd_t entry;
1541 		bool preserve_write = prot_numa && pmd_write(*pmd);
1542 		ret = 1;
1543 
1544 		/*
1545 		 * Avoid trapping faults against the zero page. The read-only
1546 		 * data is likely to be read-cached on the local CPU and
1547 		 * local/remote hits to the zero page are not interesting.
1548 		 */
1549 		if (prot_numa && is_huge_zero_pmd(*pmd)) {
1550 			spin_unlock(ptl);
1551 			return ret;
1552 		}
1553 
1554 		if (!prot_numa || !pmd_protnone(*pmd)) {
1555 			entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1556 			entry = pmd_modify(entry, newprot);
1557 			if (preserve_write)
1558 				entry = pmd_mkwrite(entry);
1559 			ret = HPAGE_PMD_NR;
1560 			set_pmd_at(mm, addr, pmd, entry);
1561 			BUG_ON(vma_is_anonymous(vma) && !preserve_write &&
1562 					pmd_write(entry));
1563 		}
1564 		spin_unlock(ptl);
1565 	}
1566 
1567 	return ret;
1568 }
1569 
1570 /*
1571  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1572  *
1573  * Note that if it returns page table lock pointer, this routine returns without
1574  * unlocking page table lock. So callers must unlock it.
1575  */
1576 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1577 {
1578 	spinlock_t *ptl;
1579 	ptl = pmd_lock(vma->vm_mm, pmd);
1580 	if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1581 		return ptl;
1582 	spin_unlock(ptl);
1583 	return NULL;
1584 }
1585 
1586 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1587 		unsigned long haddr, pmd_t *pmd)
1588 {
1589 	struct mm_struct *mm = vma->vm_mm;
1590 	pgtable_t pgtable;
1591 	pmd_t _pmd;
1592 	int i;
1593 
1594 	/* leave pmd empty until pte is filled */
1595 	pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1596 
1597 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1598 	pmd_populate(mm, &_pmd, pgtable);
1599 
1600 	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1601 		pte_t *pte, entry;
1602 		entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1603 		entry = pte_mkspecial(entry);
1604 		pte = pte_offset_map(&_pmd, haddr);
1605 		VM_BUG_ON(!pte_none(*pte));
1606 		set_pte_at(mm, haddr, pte, entry);
1607 		pte_unmap(pte);
1608 	}
1609 	smp_wmb(); /* make pte visible before pmd */
1610 	pmd_populate(mm, pmd, pgtable);
1611 }
1612 
1613 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
1614 		unsigned long haddr, bool freeze)
1615 {
1616 	struct mm_struct *mm = vma->vm_mm;
1617 	struct page *page;
1618 	pgtable_t pgtable;
1619 	pmd_t _pmd;
1620 	bool young, write, dirty, soft_dirty;
1621 	unsigned long addr;
1622 	int i;
1623 
1624 	VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1625 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1626 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
1627 	VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
1628 
1629 	count_vm_event(THP_SPLIT_PMD);
1630 
1631 	if (!vma_is_anonymous(vma)) {
1632 		_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1633 		/*
1634 		 * We are going to unmap this huge page. So
1635 		 * just go ahead and zap it
1636 		 */
1637 		if (arch_needs_pgtable_deposit())
1638 			zap_deposited_table(mm, pmd);
1639 		if (vma_is_dax(vma))
1640 			return;
1641 		page = pmd_page(_pmd);
1642 		if (!PageReferenced(page) && pmd_young(_pmd))
1643 			SetPageReferenced(page);
1644 		page_remove_rmap(page, true);
1645 		put_page(page);
1646 		add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1647 		return;
1648 	} else if (is_huge_zero_pmd(*pmd)) {
1649 		return __split_huge_zero_page_pmd(vma, haddr, pmd);
1650 	}
1651 
1652 	page = pmd_page(*pmd);
1653 	VM_BUG_ON_PAGE(!page_count(page), page);
1654 	page_ref_add(page, HPAGE_PMD_NR - 1);
1655 	write = pmd_write(*pmd);
1656 	young = pmd_young(*pmd);
1657 	dirty = pmd_dirty(*pmd);
1658 	soft_dirty = pmd_soft_dirty(*pmd);
1659 
1660 	pmdp_huge_split_prepare(vma, haddr, pmd);
1661 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1662 	pmd_populate(mm, &_pmd, pgtable);
1663 
1664 	for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
1665 		pte_t entry, *pte;
1666 		/*
1667 		 * Note that NUMA hinting access restrictions are not
1668 		 * transferred to avoid any possibility of altering
1669 		 * permissions across VMAs.
1670 		 */
1671 		if (freeze) {
1672 			swp_entry_t swp_entry;
1673 			swp_entry = make_migration_entry(page + i, write);
1674 			entry = swp_entry_to_pte(swp_entry);
1675 			if (soft_dirty)
1676 				entry = pte_swp_mksoft_dirty(entry);
1677 		} else {
1678 			entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
1679 			entry = maybe_mkwrite(entry, vma);
1680 			if (!write)
1681 				entry = pte_wrprotect(entry);
1682 			if (!young)
1683 				entry = pte_mkold(entry);
1684 			if (soft_dirty)
1685 				entry = pte_mksoft_dirty(entry);
1686 		}
1687 		if (dirty)
1688 			SetPageDirty(page + i);
1689 		pte = pte_offset_map(&_pmd, addr);
1690 		BUG_ON(!pte_none(*pte));
1691 		set_pte_at(mm, addr, pte, entry);
1692 		atomic_inc(&page[i]._mapcount);
1693 		pte_unmap(pte);
1694 	}
1695 
1696 	/*
1697 	 * Set PG_double_map before dropping compound_mapcount to avoid
1698 	 * false-negative page_mapped().
1699 	 */
1700 	if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
1701 		for (i = 0; i < HPAGE_PMD_NR; i++)
1702 			atomic_inc(&page[i]._mapcount);
1703 	}
1704 
1705 	if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
1706 		/* Last compound_mapcount is gone. */
1707 		__dec_node_page_state(page, NR_ANON_THPS);
1708 		if (TestClearPageDoubleMap(page)) {
1709 			/* No need in mapcount reference anymore */
1710 			for (i = 0; i < HPAGE_PMD_NR; i++)
1711 				atomic_dec(&page[i]._mapcount);
1712 		}
1713 	}
1714 
1715 	smp_wmb(); /* make pte visible before pmd */
1716 	/*
1717 	 * Up to this point the pmd is present and huge and userland has the
1718 	 * whole access to the hugepage during the split (which happens in
1719 	 * place). If we overwrite the pmd with the not-huge version pointing
1720 	 * to the pte here (which of course we could if all CPUs were bug
1721 	 * free), userland could trigger a small page size TLB miss on the
1722 	 * small sized TLB while the hugepage TLB entry is still established in
1723 	 * the huge TLB. Some CPU doesn't like that.
1724 	 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
1725 	 * 383 on page 93. Intel should be safe but is also warns that it's
1726 	 * only safe if the permission and cache attributes of the two entries
1727 	 * loaded in the two TLB is identical (which should be the case here).
1728 	 * But it is generally safer to never allow small and huge TLB entries
1729 	 * for the same virtual address to be loaded simultaneously. So instead
1730 	 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
1731 	 * current pmd notpresent (atomically because here the pmd_trans_huge
1732 	 * and pmd_trans_splitting must remain set at all times on the pmd
1733 	 * until the split is complete for this pmd), then we flush the SMP TLB
1734 	 * and finally we write the non-huge version of the pmd entry with
1735 	 * pmd_populate.
1736 	 */
1737 	pmdp_invalidate(vma, haddr, pmd);
1738 	pmd_populate(mm, pmd, pgtable);
1739 
1740 	if (freeze) {
1741 		for (i = 0; i < HPAGE_PMD_NR; i++) {
1742 			page_remove_rmap(page + i, false);
1743 			put_page(page + i);
1744 		}
1745 	}
1746 }
1747 
1748 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1749 		unsigned long address, bool freeze, struct page *page)
1750 {
1751 	spinlock_t *ptl;
1752 	struct mm_struct *mm = vma->vm_mm;
1753 	unsigned long haddr = address & HPAGE_PMD_MASK;
1754 
1755 	mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
1756 	ptl = pmd_lock(mm, pmd);
1757 
1758 	/*
1759 	 * If caller asks to setup a migration entries, we need a page to check
1760 	 * pmd against. Otherwise we can end up replacing wrong page.
1761 	 */
1762 	VM_BUG_ON(freeze && !page);
1763 	if (page && page != pmd_page(*pmd))
1764 	        goto out;
1765 
1766 	if (pmd_trans_huge(*pmd)) {
1767 		page = pmd_page(*pmd);
1768 		if (PageMlocked(page))
1769 			clear_page_mlock(page);
1770 	} else if (!pmd_devmap(*pmd))
1771 		goto out;
1772 	__split_huge_pmd_locked(vma, pmd, haddr, freeze);
1773 out:
1774 	spin_unlock(ptl);
1775 	mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
1776 }
1777 
1778 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
1779 		bool freeze, struct page *page)
1780 {
1781 	pgd_t *pgd;
1782 	pud_t *pud;
1783 	pmd_t *pmd;
1784 
1785 	pgd = pgd_offset(vma->vm_mm, address);
1786 	if (!pgd_present(*pgd))
1787 		return;
1788 
1789 	pud = pud_offset(pgd, address);
1790 	if (!pud_present(*pud))
1791 		return;
1792 
1793 	pmd = pmd_offset(pud, address);
1794 
1795 	__split_huge_pmd(vma, pmd, address, freeze, page);
1796 }
1797 
1798 void vma_adjust_trans_huge(struct vm_area_struct *vma,
1799 			     unsigned long start,
1800 			     unsigned long end,
1801 			     long adjust_next)
1802 {
1803 	/*
1804 	 * If the new start address isn't hpage aligned and it could
1805 	 * previously contain an hugepage: check if we need to split
1806 	 * an huge pmd.
1807 	 */
1808 	if (start & ~HPAGE_PMD_MASK &&
1809 	    (start & HPAGE_PMD_MASK) >= vma->vm_start &&
1810 	    (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1811 		split_huge_pmd_address(vma, start, false, NULL);
1812 
1813 	/*
1814 	 * If the new end address isn't hpage aligned and it could
1815 	 * previously contain an hugepage: check if we need to split
1816 	 * an huge pmd.
1817 	 */
1818 	if (end & ~HPAGE_PMD_MASK &&
1819 	    (end & HPAGE_PMD_MASK) >= vma->vm_start &&
1820 	    (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1821 		split_huge_pmd_address(vma, end, false, NULL);
1822 
1823 	/*
1824 	 * If we're also updating the vma->vm_next->vm_start, if the new
1825 	 * vm_next->vm_start isn't page aligned and it could previously
1826 	 * contain an hugepage: check if we need to split an huge pmd.
1827 	 */
1828 	if (adjust_next > 0) {
1829 		struct vm_area_struct *next = vma->vm_next;
1830 		unsigned long nstart = next->vm_start;
1831 		nstart += adjust_next << PAGE_SHIFT;
1832 		if (nstart & ~HPAGE_PMD_MASK &&
1833 		    (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
1834 		    (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
1835 			split_huge_pmd_address(next, nstart, false, NULL);
1836 	}
1837 }
1838 
1839 static void freeze_page(struct page *page)
1840 {
1841 	enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
1842 		TTU_RMAP_LOCKED;
1843 	int i, ret;
1844 
1845 	VM_BUG_ON_PAGE(!PageHead(page), page);
1846 
1847 	if (PageAnon(page))
1848 		ttu_flags |= TTU_MIGRATION;
1849 
1850 	/* We only need TTU_SPLIT_HUGE_PMD once */
1851 	ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
1852 	for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
1853 		/* Cut short if the page is unmapped */
1854 		if (page_count(page) == 1)
1855 			return;
1856 
1857 		ret = try_to_unmap(page + i, ttu_flags);
1858 	}
1859 	VM_BUG_ON_PAGE(ret, page + i - 1);
1860 }
1861 
1862 static void unfreeze_page(struct page *page)
1863 {
1864 	int i;
1865 
1866 	for (i = 0; i < HPAGE_PMD_NR; i++)
1867 		remove_migration_ptes(page + i, page + i, true);
1868 }
1869 
1870 static void __split_huge_page_tail(struct page *head, int tail,
1871 		struct lruvec *lruvec, struct list_head *list)
1872 {
1873 	struct page *page_tail = head + tail;
1874 
1875 	VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
1876 	VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
1877 
1878 	/*
1879 	 * tail_page->_refcount is zero and not changing from under us. But
1880 	 * get_page_unless_zero() may be running from under us on the
1881 	 * tail_page. If we used atomic_set() below instead of atomic_inc() or
1882 	 * atomic_add(), we would then run atomic_set() concurrently with
1883 	 * get_page_unless_zero(), and atomic_set() is implemented in C not
1884 	 * using locked ops. spin_unlock on x86 sometime uses locked ops
1885 	 * because of PPro errata 66, 92, so unless somebody can guarantee
1886 	 * atomic_set() here would be safe on all archs (and not only on x86),
1887 	 * it's safer to use atomic_inc()/atomic_add().
1888 	 */
1889 	if (PageAnon(head)) {
1890 		page_ref_inc(page_tail);
1891 	} else {
1892 		/* Additional pin to radix tree */
1893 		page_ref_add(page_tail, 2);
1894 	}
1895 
1896 	page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1897 	page_tail->flags |= (head->flags &
1898 			((1L << PG_referenced) |
1899 			 (1L << PG_swapbacked) |
1900 			 (1L << PG_mlocked) |
1901 			 (1L << PG_uptodate) |
1902 			 (1L << PG_active) |
1903 			 (1L << PG_locked) |
1904 			 (1L << PG_unevictable) |
1905 			 (1L << PG_dirty)));
1906 
1907 	/*
1908 	 * After clearing PageTail the gup refcount can be released.
1909 	 * Page flags also must be visible before we make the page non-compound.
1910 	 */
1911 	smp_wmb();
1912 
1913 	clear_compound_head(page_tail);
1914 
1915 	if (page_is_young(head))
1916 		set_page_young(page_tail);
1917 	if (page_is_idle(head))
1918 		set_page_idle(page_tail);
1919 
1920 	/* ->mapping in first tail page is compound_mapcount */
1921 	VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
1922 			page_tail);
1923 	page_tail->mapping = head->mapping;
1924 
1925 	page_tail->index = head->index + tail;
1926 	page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
1927 	lru_add_page_tail(head, page_tail, lruvec, list);
1928 }
1929 
1930 static void __split_huge_page(struct page *page, struct list_head *list,
1931 		unsigned long flags)
1932 {
1933 	struct page *head = compound_head(page);
1934 	struct zone *zone = page_zone(head);
1935 	struct lruvec *lruvec;
1936 	pgoff_t end = -1;
1937 	int i;
1938 
1939 	lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
1940 
1941 	/* complete memcg works before add pages to LRU */
1942 	mem_cgroup_split_huge_fixup(head);
1943 
1944 	if (!PageAnon(page))
1945 		end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
1946 
1947 	for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1948 		__split_huge_page_tail(head, i, lruvec, list);
1949 		/* Some pages can be beyond i_size: drop them from page cache */
1950 		if (head[i].index >= end) {
1951 			__ClearPageDirty(head + i);
1952 			__delete_from_page_cache(head + i, NULL);
1953 			if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
1954 				shmem_uncharge(head->mapping->host, 1);
1955 			put_page(head + i);
1956 		}
1957 	}
1958 
1959 	ClearPageCompound(head);
1960 	/* See comment in __split_huge_page_tail() */
1961 	if (PageAnon(head)) {
1962 		page_ref_inc(head);
1963 	} else {
1964 		/* Additional pin to radix tree */
1965 		page_ref_add(head, 2);
1966 		spin_unlock(&head->mapping->tree_lock);
1967 	}
1968 
1969 	spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
1970 
1971 	unfreeze_page(head);
1972 
1973 	for (i = 0; i < HPAGE_PMD_NR; i++) {
1974 		struct page *subpage = head + i;
1975 		if (subpage == page)
1976 			continue;
1977 		unlock_page(subpage);
1978 
1979 		/*
1980 		 * Subpages may be freed if there wasn't any mapping
1981 		 * like if add_to_swap() is running on a lru page that
1982 		 * had its mapping zapped. And freeing these pages
1983 		 * requires taking the lru_lock so we do the put_page
1984 		 * of the tail pages after the split is complete.
1985 		 */
1986 		put_page(subpage);
1987 	}
1988 }
1989 
1990 int total_mapcount(struct page *page)
1991 {
1992 	int i, compound, ret;
1993 
1994 	VM_BUG_ON_PAGE(PageTail(page), page);
1995 
1996 	if (likely(!PageCompound(page)))
1997 		return atomic_read(&page->_mapcount) + 1;
1998 
1999 	compound = compound_mapcount(page);
2000 	if (PageHuge(page))
2001 		return compound;
2002 	ret = compound;
2003 	for (i = 0; i < HPAGE_PMD_NR; i++)
2004 		ret += atomic_read(&page[i]._mapcount) + 1;
2005 	/* File pages has compound_mapcount included in _mapcount */
2006 	if (!PageAnon(page))
2007 		return ret - compound * HPAGE_PMD_NR;
2008 	if (PageDoubleMap(page))
2009 		ret -= HPAGE_PMD_NR;
2010 	return ret;
2011 }
2012 
2013 /*
2014  * This calculates accurately how many mappings a transparent hugepage
2015  * has (unlike page_mapcount() which isn't fully accurate). This full
2016  * accuracy is primarily needed to know if copy-on-write faults can
2017  * reuse the page and change the mapping to read-write instead of
2018  * copying them. At the same time this returns the total_mapcount too.
2019  *
2020  * The function returns the highest mapcount any one of the subpages
2021  * has. If the return value is one, even if different processes are
2022  * mapping different subpages of the transparent hugepage, they can
2023  * all reuse it, because each process is reusing a different subpage.
2024  *
2025  * The total_mapcount is instead counting all virtual mappings of the
2026  * subpages. If the total_mapcount is equal to "one", it tells the
2027  * caller all mappings belong to the same "mm" and in turn the
2028  * anon_vma of the transparent hugepage can become the vma->anon_vma
2029  * local one as no other process may be mapping any of the subpages.
2030  *
2031  * It would be more accurate to replace page_mapcount() with
2032  * page_trans_huge_mapcount(), however we only use
2033  * page_trans_huge_mapcount() in the copy-on-write faults where we
2034  * need full accuracy to avoid breaking page pinning, because
2035  * page_trans_huge_mapcount() is slower than page_mapcount().
2036  */
2037 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2038 {
2039 	int i, ret, _total_mapcount, mapcount;
2040 
2041 	/* hugetlbfs shouldn't call it */
2042 	VM_BUG_ON_PAGE(PageHuge(page), page);
2043 
2044 	if (likely(!PageTransCompound(page))) {
2045 		mapcount = atomic_read(&page->_mapcount) + 1;
2046 		if (total_mapcount)
2047 			*total_mapcount = mapcount;
2048 		return mapcount;
2049 	}
2050 
2051 	page = compound_head(page);
2052 
2053 	_total_mapcount = ret = 0;
2054 	for (i = 0; i < HPAGE_PMD_NR; i++) {
2055 		mapcount = atomic_read(&page[i]._mapcount) + 1;
2056 		ret = max(ret, mapcount);
2057 		_total_mapcount += mapcount;
2058 	}
2059 	if (PageDoubleMap(page)) {
2060 		ret -= 1;
2061 		_total_mapcount -= HPAGE_PMD_NR;
2062 	}
2063 	mapcount = compound_mapcount(page);
2064 	ret += mapcount;
2065 	_total_mapcount += mapcount;
2066 	if (total_mapcount)
2067 		*total_mapcount = _total_mapcount;
2068 	return ret;
2069 }
2070 
2071 /*
2072  * This function splits huge page into normal pages. @page can point to any
2073  * subpage of huge page to split. Split doesn't change the position of @page.
2074  *
2075  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2076  * The huge page must be locked.
2077  *
2078  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2079  *
2080  * Both head page and tail pages will inherit mapping, flags, and so on from
2081  * the hugepage.
2082  *
2083  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2084  * they are not mapped.
2085  *
2086  * Returns 0 if the hugepage is split successfully.
2087  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2088  * us.
2089  */
2090 int split_huge_page_to_list(struct page *page, struct list_head *list)
2091 {
2092 	struct page *head = compound_head(page);
2093 	struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2094 	struct anon_vma *anon_vma = NULL;
2095 	struct address_space *mapping = NULL;
2096 	int count, mapcount, extra_pins, ret;
2097 	bool mlocked;
2098 	unsigned long flags;
2099 
2100 	VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2101 	VM_BUG_ON_PAGE(!PageLocked(page), page);
2102 	VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2103 	VM_BUG_ON_PAGE(!PageCompound(page), page);
2104 
2105 	if (PageAnon(head)) {
2106 		/*
2107 		 * The caller does not necessarily hold an mmap_sem that would
2108 		 * prevent the anon_vma disappearing so we first we take a
2109 		 * reference to it and then lock the anon_vma for write. This
2110 		 * is similar to page_lock_anon_vma_read except the write lock
2111 		 * is taken to serialise against parallel split or collapse
2112 		 * operations.
2113 		 */
2114 		anon_vma = page_get_anon_vma(head);
2115 		if (!anon_vma) {
2116 			ret = -EBUSY;
2117 			goto out;
2118 		}
2119 		extra_pins = 0;
2120 		mapping = NULL;
2121 		anon_vma_lock_write(anon_vma);
2122 	} else {
2123 		mapping = head->mapping;
2124 
2125 		/* Truncated ? */
2126 		if (!mapping) {
2127 			ret = -EBUSY;
2128 			goto out;
2129 		}
2130 
2131 		/* Addidional pins from radix tree */
2132 		extra_pins = HPAGE_PMD_NR;
2133 		anon_vma = NULL;
2134 		i_mmap_lock_read(mapping);
2135 	}
2136 
2137 	/*
2138 	 * Racy check if we can split the page, before freeze_page() will
2139 	 * split PMDs
2140 	 */
2141 	if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
2142 		ret = -EBUSY;
2143 		goto out_unlock;
2144 	}
2145 
2146 	mlocked = PageMlocked(page);
2147 	freeze_page(head);
2148 	VM_BUG_ON_PAGE(compound_mapcount(head), head);
2149 
2150 	/* Make sure the page is not on per-CPU pagevec as it takes pin */
2151 	if (mlocked)
2152 		lru_add_drain();
2153 
2154 	/* prevent PageLRU to go away from under us, and freeze lru stats */
2155 	spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2156 
2157 	if (mapping) {
2158 		void **pslot;
2159 
2160 		spin_lock(&mapping->tree_lock);
2161 		pslot = radix_tree_lookup_slot(&mapping->page_tree,
2162 				page_index(head));
2163 		/*
2164 		 * Check if the head page is present in radix tree.
2165 		 * We assume all tail are present too, if head is there.
2166 		 */
2167 		if (radix_tree_deref_slot_protected(pslot,
2168 					&mapping->tree_lock) != head)
2169 			goto fail;
2170 	}
2171 
2172 	/* Prevent deferred_split_scan() touching ->_refcount */
2173 	spin_lock(&pgdata->split_queue_lock);
2174 	count = page_count(head);
2175 	mapcount = total_mapcount(head);
2176 	if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2177 		if (!list_empty(page_deferred_list(head))) {
2178 			pgdata->split_queue_len--;
2179 			list_del(page_deferred_list(head));
2180 		}
2181 		if (mapping)
2182 			__dec_node_page_state(page, NR_SHMEM_THPS);
2183 		spin_unlock(&pgdata->split_queue_lock);
2184 		__split_huge_page(page, list, flags);
2185 		ret = 0;
2186 	} else {
2187 		if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2188 			pr_alert("total_mapcount: %u, page_count(): %u\n",
2189 					mapcount, count);
2190 			if (PageTail(page))
2191 				dump_page(head, NULL);
2192 			dump_page(page, "total_mapcount(head) > 0");
2193 			BUG();
2194 		}
2195 		spin_unlock(&pgdata->split_queue_lock);
2196 fail:		if (mapping)
2197 			spin_unlock(&mapping->tree_lock);
2198 		spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2199 		unfreeze_page(head);
2200 		ret = -EBUSY;
2201 	}
2202 
2203 out_unlock:
2204 	if (anon_vma) {
2205 		anon_vma_unlock_write(anon_vma);
2206 		put_anon_vma(anon_vma);
2207 	}
2208 	if (mapping)
2209 		i_mmap_unlock_read(mapping);
2210 out:
2211 	count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2212 	return ret;
2213 }
2214 
2215 void free_transhuge_page(struct page *page)
2216 {
2217 	struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2218 	unsigned long flags;
2219 
2220 	spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2221 	if (!list_empty(page_deferred_list(page))) {
2222 		pgdata->split_queue_len--;
2223 		list_del(page_deferred_list(page));
2224 	}
2225 	spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2226 	free_compound_page(page);
2227 }
2228 
2229 void deferred_split_huge_page(struct page *page)
2230 {
2231 	struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2232 	unsigned long flags;
2233 
2234 	VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2235 
2236 	spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2237 	if (list_empty(page_deferred_list(page))) {
2238 		count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2239 		list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2240 		pgdata->split_queue_len++;
2241 	}
2242 	spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2243 }
2244 
2245 static unsigned long deferred_split_count(struct shrinker *shrink,
2246 		struct shrink_control *sc)
2247 {
2248 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2249 	return ACCESS_ONCE(pgdata->split_queue_len);
2250 }
2251 
2252 static unsigned long deferred_split_scan(struct shrinker *shrink,
2253 		struct shrink_control *sc)
2254 {
2255 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2256 	unsigned long flags;
2257 	LIST_HEAD(list), *pos, *next;
2258 	struct page *page;
2259 	int split = 0;
2260 
2261 	spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2262 	/* Take pin on all head pages to avoid freeing them under us */
2263 	list_for_each_safe(pos, next, &pgdata->split_queue) {
2264 		page = list_entry((void *)pos, struct page, mapping);
2265 		page = compound_head(page);
2266 		if (get_page_unless_zero(page)) {
2267 			list_move(page_deferred_list(page), &list);
2268 		} else {
2269 			/* We lost race with put_compound_page() */
2270 			list_del_init(page_deferred_list(page));
2271 			pgdata->split_queue_len--;
2272 		}
2273 		if (!--sc->nr_to_scan)
2274 			break;
2275 	}
2276 	spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2277 
2278 	list_for_each_safe(pos, next, &list) {
2279 		page = list_entry((void *)pos, struct page, mapping);
2280 		lock_page(page);
2281 		/* split_huge_page() removes page from list on success */
2282 		if (!split_huge_page(page))
2283 			split++;
2284 		unlock_page(page);
2285 		put_page(page);
2286 	}
2287 
2288 	spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2289 	list_splice_tail(&list, &pgdata->split_queue);
2290 	spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2291 
2292 	/*
2293 	 * Stop shrinker if we didn't split any page, but the queue is empty.
2294 	 * This can happen if pages were freed under us.
2295 	 */
2296 	if (!split && list_empty(&pgdata->split_queue))
2297 		return SHRINK_STOP;
2298 	return split;
2299 }
2300 
2301 static struct shrinker deferred_split_shrinker = {
2302 	.count_objects = deferred_split_count,
2303 	.scan_objects = deferred_split_scan,
2304 	.seeks = DEFAULT_SEEKS,
2305 	.flags = SHRINKER_NUMA_AWARE,
2306 };
2307 
2308 #ifdef CONFIG_DEBUG_FS
2309 static int split_huge_pages_set(void *data, u64 val)
2310 {
2311 	struct zone *zone;
2312 	struct page *page;
2313 	unsigned long pfn, max_zone_pfn;
2314 	unsigned long total = 0, split = 0;
2315 
2316 	if (val != 1)
2317 		return -EINVAL;
2318 
2319 	for_each_populated_zone(zone) {
2320 		max_zone_pfn = zone_end_pfn(zone);
2321 		for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2322 			if (!pfn_valid(pfn))
2323 				continue;
2324 
2325 			page = pfn_to_page(pfn);
2326 			if (!get_page_unless_zero(page))
2327 				continue;
2328 
2329 			if (zone != page_zone(page))
2330 				goto next;
2331 
2332 			if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2333 				goto next;
2334 
2335 			total++;
2336 			lock_page(page);
2337 			if (!split_huge_page(page))
2338 				split++;
2339 			unlock_page(page);
2340 next:
2341 			put_page(page);
2342 		}
2343 	}
2344 
2345 	pr_info("%lu of %lu THP split\n", split, total);
2346 
2347 	return 0;
2348 }
2349 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2350 		"%llu\n");
2351 
2352 static int __init split_huge_pages_debugfs(void)
2353 {
2354 	void *ret;
2355 
2356 	ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2357 			&split_huge_pages_fops);
2358 	if (!ret)
2359 		pr_warn("Failed to create split_huge_pages in debugfs");
2360 	return 0;
2361 }
2362 late_initcall(split_huge_pages_debugfs);
2363 #endif
2364