xref: /openbmc/linux/mm/huge_memory.c (revision ae213c44)
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/sched/coredump.h>
13 #include <linux/sched/numa_balancing.h>
14 #include <linux/highmem.h>
15 #include <linux/hugetlb.h>
16 #include <linux/mmu_notifier.h>
17 #include <linux/rmap.h>
18 #include <linux/swap.h>
19 #include <linux/shrinker.h>
20 #include <linux/mm_inline.h>
21 #include <linux/swapops.h>
22 #include <linux/dax.h>
23 #include <linux/khugepaged.h>
24 #include <linux/freezer.h>
25 #include <linux/pfn_t.h>
26 #include <linux/mman.h>
27 #include <linux/memremap.h>
28 #include <linux/pagemap.h>
29 #include <linux/debugfs.h>
30 #include <linux/migrate.h>
31 #include <linux/hashtable.h>
32 #include <linux/userfaultfd_k.h>
33 #include <linux/page_idle.h>
34 #include <linux/shmem_fs.h>
35 #include <linux/oom.h>
36 #include <linux/numa.h>
37 
38 #include <asm/tlb.h>
39 #include <asm/pgalloc.h>
40 #include "internal.h"
41 
42 /*
43  * By default, transparent hugepage support is disabled in order to avoid
44  * risking an increased memory footprint for applications that are not
45  * guaranteed to benefit from it. When transparent hugepage support is
46  * enabled, it is for all mappings, and khugepaged scans all mappings.
47  * Defrag is invoked by khugepaged hugepage allocations and by page faults
48  * for all hugepage allocations.
49  */
50 unsigned long transparent_hugepage_flags __read_mostly =
51 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
52 	(1<<TRANSPARENT_HUGEPAGE_FLAG)|
53 #endif
54 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
55 	(1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
56 #endif
57 	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
58 	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
59 	(1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
60 
61 static struct shrinker deferred_split_shrinker;
62 
63 static atomic_t huge_zero_refcount;
64 struct page *huge_zero_page __read_mostly;
65 
66 bool transparent_hugepage_enabled(struct vm_area_struct *vma)
67 {
68 	if (vma_is_anonymous(vma))
69 		return __transparent_hugepage_enabled(vma);
70 	if (vma_is_shmem(vma) && shmem_huge_enabled(vma))
71 		return __transparent_hugepage_enabled(vma);
72 
73 	return false;
74 }
75 
76 static struct page *get_huge_zero_page(void)
77 {
78 	struct page *zero_page;
79 retry:
80 	if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
81 		return READ_ONCE(huge_zero_page);
82 
83 	zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
84 			HPAGE_PMD_ORDER);
85 	if (!zero_page) {
86 		count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
87 		return NULL;
88 	}
89 	count_vm_event(THP_ZERO_PAGE_ALLOC);
90 	preempt_disable();
91 	if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
92 		preempt_enable();
93 		__free_pages(zero_page, compound_order(zero_page));
94 		goto retry;
95 	}
96 
97 	/* We take additional reference here. It will be put back by shrinker */
98 	atomic_set(&huge_zero_refcount, 2);
99 	preempt_enable();
100 	return READ_ONCE(huge_zero_page);
101 }
102 
103 static void put_huge_zero_page(void)
104 {
105 	/*
106 	 * Counter should never go to zero here. Only shrinker can put
107 	 * last reference.
108 	 */
109 	BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
110 }
111 
112 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
113 {
114 	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
115 		return READ_ONCE(huge_zero_page);
116 
117 	if (!get_huge_zero_page())
118 		return NULL;
119 
120 	if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
121 		put_huge_zero_page();
122 
123 	return READ_ONCE(huge_zero_page);
124 }
125 
126 void mm_put_huge_zero_page(struct mm_struct *mm)
127 {
128 	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
129 		put_huge_zero_page();
130 }
131 
132 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
133 					struct shrink_control *sc)
134 {
135 	/* we can free zero page only if last reference remains */
136 	return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
137 }
138 
139 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
140 				       struct shrink_control *sc)
141 {
142 	if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
143 		struct page *zero_page = xchg(&huge_zero_page, NULL);
144 		BUG_ON(zero_page == NULL);
145 		__free_pages(zero_page, compound_order(zero_page));
146 		return HPAGE_PMD_NR;
147 	}
148 
149 	return 0;
150 }
151 
152 static struct shrinker huge_zero_page_shrinker = {
153 	.count_objects = shrink_huge_zero_page_count,
154 	.scan_objects = shrink_huge_zero_page_scan,
155 	.seeks = DEFAULT_SEEKS,
156 };
157 
158 #ifdef CONFIG_SYSFS
159 static ssize_t enabled_show(struct kobject *kobj,
160 			    struct kobj_attribute *attr, char *buf)
161 {
162 	if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
163 		return sprintf(buf, "[always] madvise never\n");
164 	else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
165 		return sprintf(buf, "always [madvise] never\n");
166 	else
167 		return sprintf(buf, "always madvise [never]\n");
168 }
169 
170 static ssize_t enabled_store(struct kobject *kobj,
171 			     struct kobj_attribute *attr,
172 			     const char *buf, size_t count)
173 {
174 	ssize_t ret = count;
175 
176 	if (!memcmp("always", buf,
177 		    min(sizeof("always")-1, count))) {
178 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
179 		set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
180 	} else if (!memcmp("madvise", buf,
181 			   min(sizeof("madvise")-1, count))) {
182 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
183 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
184 	} else if (!memcmp("never", buf,
185 			   min(sizeof("never")-1, count))) {
186 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
187 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
188 	} else
189 		ret = -EINVAL;
190 
191 	if (ret > 0) {
192 		int err = start_stop_khugepaged();
193 		if (err)
194 			ret = err;
195 	}
196 	return ret;
197 }
198 static struct kobj_attribute enabled_attr =
199 	__ATTR(enabled, 0644, enabled_show, enabled_store);
200 
201 ssize_t single_hugepage_flag_show(struct kobject *kobj,
202 				struct kobj_attribute *attr, char *buf,
203 				enum transparent_hugepage_flag flag)
204 {
205 	return sprintf(buf, "%d\n",
206 		       !!test_bit(flag, &transparent_hugepage_flags));
207 }
208 
209 ssize_t single_hugepage_flag_store(struct kobject *kobj,
210 				 struct kobj_attribute *attr,
211 				 const char *buf, size_t count,
212 				 enum transparent_hugepage_flag flag)
213 {
214 	unsigned long value;
215 	int ret;
216 
217 	ret = kstrtoul(buf, 10, &value);
218 	if (ret < 0)
219 		return ret;
220 	if (value > 1)
221 		return -EINVAL;
222 
223 	if (value)
224 		set_bit(flag, &transparent_hugepage_flags);
225 	else
226 		clear_bit(flag, &transparent_hugepage_flags);
227 
228 	return count;
229 }
230 
231 static ssize_t defrag_show(struct kobject *kobj,
232 			   struct kobj_attribute *attr, char *buf)
233 {
234 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
235 		return sprintf(buf, "[always] defer defer+madvise madvise never\n");
236 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
237 		return sprintf(buf, "always [defer] defer+madvise madvise never\n");
238 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
239 		return sprintf(buf, "always defer [defer+madvise] madvise never\n");
240 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
241 		return sprintf(buf, "always defer defer+madvise [madvise] never\n");
242 	return sprintf(buf, "always defer defer+madvise madvise [never]\n");
243 }
244 
245 static ssize_t defrag_store(struct kobject *kobj,
246 			    struct kobj_attribute *attr,
247 			    const char *buf, size_t count)
248 {
249 	if (!memcmp("always", buf,
250 		    min(sizeof("always")-1, count))) {
251 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
252 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
253 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
254 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
255 	} else if (!memcmp("defer+madvise", buf,
256 		    min(sizeof("defer+madvise")-1, count))) {
257 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
258 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
259 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
260 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
261 	} else if (!memcmp("defer", buf,
262 		    min(sizeof("defer")-1, count))) {
263 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
264 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
265 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
266 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
267 	} else if (!memcmp("madvise", buf,
268 			   min(sizeof("madvise")-1, count))) {
269 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
270 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
271 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
272 		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
273 	} else if (!memcmp("never", buf,
274 			   min(sizeof("never")-1, count))) {
275 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
276 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
277 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
278 		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
279 	} else
280 		return -EINVAL;
281 
282 	return count;
283 }
284 static struct kobj_attribute defrag_attr =
285 	__ATTR(defrag, 0644, defrag_show, defrag_store);
286 
287 static ssize_t use_zero_page_show(struct kobject *kobj,
288 		struct kobj_attribute *attr, char *buf)
289 {
290 	return single_hugepage_flag_show(kobj, attr, buf,
291 				TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
292 }
293 static ssize_t use_zero_page_store(struct kobject *kobj,
294 		struct kobj_attribute *attr, const char *buf, size_t count)
295 {
296 	return single_hugepage_flag_store(kobj, attr, buf, count,
297 				 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
298 }
299 static struct kobj_attribute use_zero_page_attr =
300 	__ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
301 
302 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
303 		struct kobj_attribute *attr, char *buf)
304 {
305 	return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
306 }
307 static struct kobj_attribute hpage_pmd_size_attr =
308 	__ATTR_RO(hpage_pmd_size);
309 
310 #ifdef CONFIG_DEBUG_VM
311 static ssize_t debug_cow_show(struct kobject *kobj,
312 				struct kobj_attribute *attr, char *buf)
313 {
314 	return single_hugepage_flag_show(kobj, attr, buf,
315 				TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
316 }
317 static ssize_t debug_cow_store(struct kobject *kobj,
318 			       struct kobj_attribute *attr,
319 			       const char *buf, size_t count)
320 {
321 	return single_hugepage_flag_store(kobj, attr, buf, count,
322 				 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
323 }
324 static struct kobj_attribute debug_cow_attr =
325 	__ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
326 #endif /* CONFIG_DEBUG_VM */
327 
328 static struct attribute *hugepage_attr[] = {
329 	&enabled_attr.attr,
330 	&defrag_attr.attr,
331 	&use_zero_page_attr.attr,
332 	&hpage_pmd_size_attr.attr,
333 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
334 	&shmem_enabled_attr.attr,
335 #endif
336 #ifdef CONFIG_DEBUG_VM
337 	&debug_cow_attr.attr,
338 #endif
339 	NULL,
340 };
341 
342 static const struct attribute_group hugepage_attr_group = {
343 	.attrs = hugepage_attr,
344 };
345 
346 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
347 {
348 	int err;
349 
350 	*hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
351 	if (unlikely(!*hugepage_kobj)) {
352 		pr_err("failed to create transparent hugepage kobject\n");
353 		return -ENOMEM;
354 	}
355 
356 	err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
357 	if (err) {
358 		pr_err("failed to register transparent hugepage group\n");
359 		goto delete_obj;
360 	}
361 
362 	err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
363 	if (err) {
364 		pr_err("failed to register transparent hugepage group\n");
365 		goto remove_hp_group;
366 	}
367 
368 	return 0;
369 
370 remove_hp_group:
371 	sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
372 delete_obj:
373 	kobject_put(*hugepage_kobj);
374 	return err;
375 }
376 
377 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
378 {
379 	sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
380 	sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
381 	kobject_put(hugepage_kobj);
382 }
383 #else
384 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
385 {
386 	return 0;
387 }
388 
389 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
390 {
391 }
392 #endif /* CONFIG_SYSFS */
393 
394 static int __init hugepage_init(void)
395 {
396 	int err;
397 	struct kobject *hugepage_kobj;
398 
399 	if (!has_transparent_hugepage()) {
400 		transparent_hugepage_flags = 0;
401 		return -EINVAL;
402 	}
403 
404 	/*
405 	 * hugepages can't be allocated by the buddy allocator
406 	 */
407 	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
408 	/*
409 	 * we use page->mapping and page->index in second tail page
410 	 * as list_head: assuming THP order >= 2
411 	 */
412 	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
413 
414 	err = hugepage_init_sysfs(&hugepage_kobj);
415 	if (err)
416 		goto err_sysfs;
417 
418 	err = khugepaged_init();
419 	if (err)
420 		goto err_slab;
421 
422 	err = register_shrinker(&huge_zero_page_shrinker);
423 	if (err)
424 		goto err_hzp_shrinker;
425 	err = register_shrinker(&deferred_split_shrinker);
426 	if (err)
427 		goto err_split_shrinker;
428 
429 	/*
430 	 * By default disable transparent hugepages on smaller systems,
431 	 * where the extra memory used could hurt more than TLB overhead
432 	 * is likely to save.  The admin can still enable it through /sys.
433 	 */
434 	if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
435 		transparent_hugepage_flags = 0;
436 		return 0;
437 	}
438 
439 	err = start_stop_khugepaged();
440 	if (err)
441 		goto err_khugepaged;
442 
443 	return 0;
444 err_khugepaged:
445 	unregister_shrinker(&deferred_split_shrinker);
446 err_split_shrinker:
447 	unregister_shrinker(&huge_zero_page_shrinker);
448 err_hzp_shrinker:
449 	khugepaged_destroy();
450 err_slab:
451 	hugepage_exit_sysfs(hugepage_kobj);
452 err_sysfs:
453 	return err;
454 }
455 subsys_initcall(hugepage_init);
456 
457 static int __init setup_transparent_hugepage(char *str)
458 {
459 	int ret = 0;
460 	if (!str)
461 		goto out;
462 	if (!strcmp(str, "always")) {
463 		set_bit(TRANSPARENT_HUGEPAGE_FLAG,
464 			&transparent_hugepage_flags);
465 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
466 			  &transparent_hugepage_flags);
467 		ret = 1;
468 	} else if (!strcmp(str, "madvise")) {
469 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
470 			  &transparent_hugepage_flags);
471 		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
472 			&transparent_hugepage_flags);
473 		ret = 1;
474 	} else if (!strcmp(str, "never")) {
475 		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
476 			  &transparent_hugepage_flags);
477 		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
478 			  &transparent_hugepage_flags);
479 		ret = 1;
480 	}
481 out:
482 	if (!ret)
483 		pr_warn("transparent_hugepage= cannot parse, ignored\n");
484 	return ret;
485 }
486 __setup("transparent_hugepage=", setup_transparent_hugepage);
487 
488 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
489 {
490 	if (likely(vma->vm_flags & VM_WRITE))
491 		pmd = pmd_mkwrite(pmd);
492 	return pmd;
493 }
494 
495 static inline struct list_head *page_deferred_list(struct page *page)
496 {
497 	/* ->lru in the tail pages is occupied by compound_head. */
498 	return &page[2].deferred_list;
499 }
500 
501 void prep_transhuge_page(struct page *page)
502 {
503 	/*
504 	 * we use page->mapping and page->indexlru in second tail page
505 	 * as list_head: assuming THP order >= 2
506 	 */
507 
508 	INIT_LIST_HEAD(page_deferred_list(page));
509 	set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
510 }
511 
512 static unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
513 		loff_t off, unsigned long flags, unsigned long size)
514 {
515 	unsigned long addr;
516 	loff_t off_end = off + len;
517 	loff_t off_align = round_up(off, size);
518 	unsigned long len_pad;
519 
520 	if (off_end <= off_align || (off_end - off_align) < size)
521 		return 0;
522 
523 	len_pad = len + size;
524 	if (len_pad < len || (off + len_pad) < off)
525 		return 0;
526 
527 	addr = current->mm->get_unmapped_area(filp, 0, len_pad,
528 					      off >> PAGE_SHIFT, flags);
529 	if (IS_ERR_VALUE(addr))
530 		return 0;
531 
532 	addr += (off - addr) & (size - 1);
533 	return addr;
534 }
535 
536 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
537 		unsigned long len, unsigned long pgoff, unsigned long flags)
538 {
539 	loff_t off = (loff_t)pgoff << PAGE_SHIFT;
540 
541 	if (addr)
542 		goto out;
543 	if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
544 		goto out;
545 
546 	addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
547 	if (addr)
548 		return addr;
549 
550  out:
551 	return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
552 }
553 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
554 
555 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
556 			struct page *page, gfp_t gfp)
557 {
558 	struct vm_area_struct *vma = vmf->vma;
559 	struct mem_cgroup *memcg;
560 	pgtable_t pgtable;
561 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
562 	vm_fault_t ret = 0;
563 
564 	VM_BUG_ON_PAGE(!PageCompound(page), page);
565 
566 	if (mem_cgroup_try_charge_delay(page, vma->vm_mm, gfp, &memcg, true)) {
567 		put_page(page);
568 		count_vm_event(THP_FAULT_FALLBACK);
569 		return VM_FAULT_FALLBACK;
570 	}
571 
572 	pgtable = pte_alloc_one(vma->vm_mm);
573 	if (unlikely(!pgtable)) {
574 		ret = VM_FAULT_OOM;
575 		goto release;
576 	}
577 
578 	clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
579 	/*
580 	 * The memory barrier inside __SetPageUptodate makes sure that
581 	 * clear_huge_page writes become visible before the set_pmd_at()
582 	 * write.
583 	 */
584 	__SetPageUptodate(page);
585 
586 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
587 	if (unlikely(!pmd_none(*vmf->pmd))) {
588 		goto unlock_release;
589 	} else {
590 		pmd_t entry;
591 
592 		ret = check_stable_address_space(vma->vm_mm);
593 		if (ret)
594 			goto unlock_release;
595 
596 		/* Deliver the page fault to userland */
597 		if (userfaultfd_missing(vma)) {
598 			vm_fault_t ret2;
599 
600 			spin_unlock(vmf->ptl);
601 			mem_cgroup_cancel_charge(page, memcg, true);
602 			put_page(page);
603 			pte_free(vma->vm_mm, pgtable);
604 			ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
605 			VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
606 			return ret2;
607 		}
608 
609 		entry = mk_huge_pmd(page, vma->vm_page_prot);
610 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
611 		page_add_new_anon_rmap(page, vma, haddr, true);
612 		mem_cgroup_commit_charge(page, memcg, false, true);
613 		lru_cache_add_active_or_unevictable(page, vma);
614 		pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
615 		set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
616 		add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
617 		mm_inc_nr_ptes(vma->vm_mm);
618 		spin_unlock(vmf->ptl);
619 		count_vm_event(THP_FAULT_ALLOC);
620 		count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
621 	}
622 
623 	return 0;
624 unlock_release:
625 	spin_unlock(vmf->ptl);
626 release:
627 	if (pgtable)
628 		pte_free(vma->vm_mm, pgtable);
629 	mem_cgroup_cancel_charge(page, memcg, true);
630 	put_page(page);
631 	return ret;
632 
633 }
634 
635 /*
636  * always: directly stall for all thp allocations
637  * defer: wake kswapd and fail if not immediately available
638  * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
639  *		  fail if not immediately available
640  * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
641  *	    available
642  * never: never stall for any thp allocation
643  */
644 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
645 {
646 	const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
647 
648 	/* Always do synchronous compaction */
649 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
650 		return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
651 
652 	/* Kick kcompactd and fail quickly */
653 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
654 		return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
655 
656 	/* Synchronous compaction if madvised, otherwise kick kcompactd */
657 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
658 		return GFP_TRANSHUGE_LIGHT |
659 			(vma_madvised ? __GFP_DIRECT_RECLAIM :
660 					__GFP_KSWAPD_RECLAIM);
661 
662 	/* Only do synchronous compaction if madvised */
663 	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
664 		return GFP_TRANSHUGE_LIGHT |
665 		       (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
666 
667 	return GFP_TRANSHUGE_LIGHT;
668 }
669 
670 /* Caller must hold page table lock. */
671 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
672 		struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
673 		struct page *zero_page)
674 {
675 	pmd_t entry;
676 	if (!pmd_none(*pmd))
677 		return false;
678 	entry = mk_pmd(zero_page, vma->vm_page_prot);
679 	entry = pmd_mkhuge(entry);
680 	if (pgtable)
681 		pgtable_trans_huge_deposit(mm, pmd, pgtable);
682 	set_pmd_at(mm, haddr, pmd, entry);
683 	mm_inc_nr_ptes(mm);
684 	return true;
685 }
686 
687 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
688 {
689 	struct vm_area_struct *vma = vmf->vma;
690 	gfp_t gfp;
691 	struct page *page;
692 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
693 
694 	if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
695 		return VM_FAULT_FALLBACK;
696 	if (unlikely(anon_vma_prepare(vma)))
697 		return VM_FAULT_OOM;
698 	if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
699 		return VM_FAULT_OOM;
700 	if (!(vmf->flags & FAULT_FLAG_WRITE) &&
701 			!mm_forbids_zeropage(vma->vm_mm) &&
702 			transparent_hugepage_use_zero_page()) {
703 		pgtable_t pgtable;
704 		struct page *zero_page;
705 		bool set;
706 		vm_fault_t ret;
707 		pgtable = pte_alloc_one(vma->vm_mm);
708 		if (unlikely(!pgtable))
709 			return VM_FAULT_OOM;
710 		zero_page = mm_get_huge_zero_page(vma->vm_mm);
711 		if (unlikely(!zero_page)) {
712 			pte_free(vma->vm_mm, pgtable);
713 			count_vm_event(THP_FAULT_FALLBACK);
714 			return VM_FAULT_FALLBACK;
715 		}
716 		vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
717 		ret = 0;
718 		set = false;
719 		if (pmd_none(*vmf->pmd)) {
720 			ret = check_stable_address_space(vma->vm_mm);
721 			if (ret) {
722 				spin_unlock(vmf->ptl);
723 			} else if (userfaultfd_missing(vma)) {
724 				spin_unlock(vmf->ptl);
725 				ret = handle_userfault(vmf, VM_UFFD_MISSING);
726 				VM_BUG_ON(ret & VM_FAULT_FALLBACK);
727 			} else {
728 				set_huge_zero_page(pgtable, vma->vm_mm, vma,
729 						   haddr, vmf->pmd, zero_page);
730 				spin_unlock(vmf->ptl);
731 				set = true;
732 			}
733 		} else
734 			spin_unlock(vmf->ptl);
735 		if (!set)
736 			pte_free(vma->vm_mm, pgtable);
737 		return ret;
738 	}
739 	gfp = alloc_hugepage_direct_gfpmask(vma);
740 	page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
741 	if (unlikely(!page)) {
742 		count_vm_event(THP_FAULT_FALLBACK);
743 		return VM_FAULT_FALLBACK;
744 	}
745 	prep_transhuge_page(page);
746 	return __do_huge_pmd_anonymous_page(vmf, page, gfp);
747 }
748 
749 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
750 		pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
751 		pgtable_t pgtable)
752 {
753 	struct mm_struct *mm = vma->vm_mm;
754 	pmd_t entry;
755 	spinlock_t *ptl;
756 
757 	ptl = pmd_lock(mm, pmd);
758 	if (!pmd_none(*pmd)) {
759 		if (write) {
760 			if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
761 				WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
762 				goto out_unlock;
763 			}
764 			entry = pmd_mkyoung(*pmd);
765 			entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
766 			if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
767 				update_mmu_cache_pmd(vma, addr, pmd);
768 		}
769 
770 		goto out_unlock;
771 	}
772 
773 	entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
774 	if (pfn_t_devmap(pfn))
775 		entry = pmd_mkdevmap(entry);
776 	if (write) {
777 		entry = pmd_mkyoung(pmd_mkdirty(entry));
778 		entry = maybe_pmd_mkwrite(entry, vma);
779 	}
780 
781 	if (pgtable) {
782 		pgtable_trans_huge_deposit(mm, pmd, pgtable);
783 		mm_inc_nr_ptes(mm);
784 		pgtable = NULL;
785 	}
786 
787 	set_pmd_at(mm, addr, pmd, entry);
788 	update_mmu_cache_pmd(vma, addr, pmd);
789 
790 out_unlock:
791 	spin_unlock(ptl);
792 	if (pgtable)
793 		pte_free(mm, pgtable);
794 }
795 
796 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)
797 {
798 	unsigned long addr = vmf->address & PMD_MASK;
799 	struct vm_area_struct *vma = vmf->vma;
800 	pgprot_t pgprot = vma->vm_page_prot;
801 	pgtable_t pgtable = NULL;
802 
803 	/*
804 	 * If we had pmd_special, we could avoid all these restrictions,
805 	 * but we need to be consistent with PTEs and architectures that
806 	 * can't support a 'special' bit.
807 	 */
808 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
809 			!pfn_t_devmap(pfn));
810 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
811 						(VM_PFNMAP|VM_MIXEDMAP));
812 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
813 
814 	if (addr < vma->vm_start || addr >= vma->vm_end)
815 		return VM_FAULT_SIGBUS;
816 
817 	if (arch_needs_pgtable_deposit()) {
818 		pgtable = pte_alloc_one(vma->vm_mm);
819 		if (!pgtable)
820 			return VM_FAULT_OOM;
821 	}
822 
823 	track_pfn_insert(vma, &pgprot, pfn);
824 
825 	insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
826 	return VM_FAULT_NOPAGE;
827 }
828 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
829 
830 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
831 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
832 {
833 	if (likely(vma->vm_flags & VM_WRITE))
834 		pud = pud_mkwrite(pud);
835 	return pud;
836 }
837 
838 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
839 		pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
840 {
841 	struct mm_struct *mm = vma->vm_mm;
842 	pud_t entry;
843 	spinlock_t *ptl;
844 
845 	ptl = pud_lock(mm, pud);
846 	if (!pud_none(*pud)) {
847 		if (write) {
848 			if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
849 				WARN_ON_ONCE(!is_huge_zero_pud(*pud));
850 				goto out_unlock;
851 			}
852 			entry = pud_mkyoung(*pud);
853 			entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
854 			if (pudp_set_access_flags(vma, addr, pud, entry, 1))
855 				update_mmu_cache_pud(vma, addr, pud);
856 		}
857 		goto out_unlock;
858 	}
859 
860 	entry = pud_mkhuge(pfn_t_pud(pfn, prot));
861 	if (pfn_t_devmap(pfn))
862 		entry = pud_mkdevmap(entry);
863 	if (write) {
864 		entry = pud_mkyoung(pud_mkdirty(entry));
865 		entry = maybe_pud_mkwrite(entry, vma);
866 	}
867 	set_pud_at(mm, addr, pud, entry);
868 	update_mmu_cache_pud(vma, addr, pud);
869 
870 out_unlock:
871 	spin_unlock(ptl);
872 }
873 
874 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)
875 {
876 	unsigned long addr = vmf->address & PUD_MASK;
877 	struct vm_area_struct *vma = vmf->vma;
878 	pgprot_t pgprot = vma->vm_page_prot;
879 
880 	/*
881 	 * If we had pud_special, we could avoid all these restrictions,
882 	 * but we need to be consistent with PTEs and architectures that
883 	 * can't support a 'special' bit.
884 	 */
885 	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
886 			!pfn_t_devmap(pfn));
887 	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
888 						(VM_PFNMAP|VM_MIXEDMAP));
889 	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
890 
891 	if (addr < vma->vm_start || addr >= vma->vm_end)
892 		return VM_FAULT_SIGBUS;
893 
894 	track_pfn_insert(vma, &pgprot, pfn);
895 
896 	insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
897 	return VM_FAULT_NOPAGE;
898 }
899 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
900 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
901 
902 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
903 		pmd_t *pmd, int flags)
904 {
905 	pmd_t _pmd;
906 
907 	_pmd = pmd_mkyoung(*pmd);
908 	if (flags & FOLL_WRITE)
909 		_pmd = pmd_mkdirty(_pmd);
910 	if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
911 				pmd, _pmd, flags & FOLL_WRITE))
912 		update_mmu_cache_pmd(vma, addr, pmd);
913 }
914 
915 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
916 		pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
917 {
918 	unsigned long pfn = pmd_pfn(*pmd);
919 	struct mm_struct *mm = vma->vm_mm;
920 	struct page *page;
921 
922 	assert_spin_locked(pmd_lockptr(mm, pmd));
923 
924 	/*
925 	 * When we COW a devmap PMD entry, we split it into PTEs, so we should
926 	 * not be in this function with `flags & FOLL_COW` set.
927 	 */
928 	WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
929 
930 	if (flags & FOLL_WRITE && !pmd_write(*pmd))
931 		return NULL;
932 
933 	if (pmd_present(*pmd) && pmd_devmap(*pmd))
934 		/* pass */;
935 	else
936 		return NULL;
937 
938 	if (flags & FOLL_TOUCH)
939 		touch_pmd(vma, addr, pmd, flags);
940 
941 	/*
942 	 * device mapped pages can only be returned if the
943 	 * caller will manage the page reference count.
944 	 */
945 	if (!(flags & FOLL_GET))
946 		return ERR_PTR(-EEXIST);
947 
948 	pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
949 	*pgmap = get_dev_pagemap(pfn, *pgmap);
950 	if (!*pgmap)
951 		return ERR_PTR(-EFAULT);
952 	page = pfn_to_page(pfn);
953 	get_page(page);
954 
955 	return page;
956 }
957 
958 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
959 		  pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
960 		  struct vm_area_struct *vma)
961 {
962 	spinlock_t *dst_ptl, *src_ptl;
963 	struct page *src_page;
964 	pmd_t pmd;
965 	pgtable_t pgtable = NULL;
966 	int ret = -ENOMEM;
967 
968 	/* Skip if can be re-fill on fault */
969 	if (!vma_is_anonymous(vma))
970 		return 0;
971 
972 	pgtable = pte_alloc_one(dst_mm);
973 	if (unlikely(!pgtable))
974 		goto out;
975 
976 	dst_ptl = pmd_lock(dst_mm, dst_pmd);
977 	src_ptl = pmd_lockptr(src_mm, src_pmd);
978 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
979 
980 	ret = -EAGAIN;
981 	pmd = *src_pmd;
982 
983 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
984 	if (unlikely(is_swap_pmd(pmd))) {
985 		swp_entry_t entry = pmd_to_swp_entry(pmd);
986 
987 		VM_BUG_ON(!is_pmd_migration_entry(pmd));
988 		if (is_write_migration_entry(entry)) {
989 			make_migration_entry_read(&entry);
990 			pmd = swp_entry_to_pmd(entry);
991 			if (pmd_swp_soft_dirty(*src_pmd))
992 				pmd = pmd_swp_mksoft_dirty(pmd);
993 			set_pmd_at(src_mm, addr, src_pmd, pmd);
994 		}
995 		add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
996 		mm_inc_nr_ptes(dst_mm);
997 		pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
998 		set_pmd_at(dst_mm, addr, dst_pmd, pmd);
999 		ret = 0;
1000 		goto out_unlock;
1001 	}
1002 #endif
1003 
1004 	if (unlikely(!pmd_trans_huge(pmd))) {
1005 		pte_free(dst_mm, pgtable);
1006 		goto out_unlock;
1007 	}
1008 	/*
1009 	 * When page table lock is held, the huge zero pmd should not be
1010 	 * under splitting since we don't split the page itself, only pmd to
1011 	 * a page table.
1012 	 */
1013 	if (is_huge_zero_pmd(pmd)) {
1014 		struct page *zero_page;
1015 		/*
1016 		 * get_huge_zero_page() will never allocate a new page here,
1017 		 * since we already have a zero page to copy. It just takes a
1018 		 * reference.
1019 		 */
1020 		zero_page = mm_get_huge_zero_page(dst_mm);
1021 		set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1022 				zero_page);
1023 		ret = 0;
1024 		goto out_unlock;
1025 	}
1026 
1027 	src_page = pmd_page(pmd);
1028 	VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1029 	get_page(src_page);
1030 	page_dup_rmap(src_page, true);
1031 	add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1032 	mm_inc_nr_ptes(dst_mm);
1033 	pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1034 
1035 	pmdp_set_wrprotect(src_mm, addr, src_pmd);
1036 	pmd = pmd_mkold(pmd_wrprotect(pmd));
1037 	set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1038 
1039 	ret = 0;
1040 out_unlock:
1041 	spin_unlock(src_ptl);
1042 	spin_unlock(dst_ptl);
1043 out:
1044 	return ret;
1045 }
1046 
1047 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1048 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1049 		pud_t *pud, int flags)
1050 {
1051 	pud_t _pud;
1052 
1053 	_pud = pud_mkyoung(*pud);
1054 	if (flags & FOLL_WRITE)
1055 		_pud = pud_mkdirty(_pud);
1056 	if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1057 				pud, _pud, flags & FOLL_WRITE))
1058 		update_mmu_cache_pud(vma, addr, pud);
1059 }
1060 
1061 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1062 		pud_t *pud, int flags, struct dev_pagemap **pgmap)
1063 {
1064 	unsigned long pfn = pud_pfn(*pud);
1065 	struct mm_struct *mm = vma->vm_mm;
1066 	struct page *page;
1067 
1068 	assert_spin_locked(pud_lockptr(mm, pud));
1069 
1070 	if (flags & FOLL_WRITE && !pud_write(*pud))
1071 		return NULL;
1072 
1073 	if (pud_present(*pud) && pud_devmap(*pud))
1074 		/* pass */;
1075 	else
1076 		return NULL;
1077 
1078 	if (flags & FOLL_TOUCH)
1079 		touch_pud(vma, addr, pud, flags);
1080 
1081 	/*
1082 	 * device mapped pages can only be returned if the
1083 	 * caller will manage the page reference count.
1084 	 */
1085 	if (!(flags & FOLL_GET))
1086 		return ERR_PTR(-EEXIST);
1087 
1088 	pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1089 	*pgmap = get_dev_pagemap(pfn, *pgmap);
1090 	if (!*pgmap)
1091 		return ERR_PTR(-EFAULT);
1092 	page = pfn_to_page(pfn);
1093 	get_page(page);
1094 
1095 	return page;
1096 }
1097 
1098 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1099 		  pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1100 		  struct vm_area_struct *vma)
1101 {
1102 	spinlock_t *dst_ptl, *src_ptl;
1103 	pud_t pud;
1104 	int ret;
1105 
1106 	dst_ptl = pud_lock(dst_mm, dst_pud);
1107 	src_ptl = pud_lockptr(src_mm, src_pud);
1108 	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1109 
1110 	ret = -EAGAIN;
1111 	pud = *src_pud;
1112 	if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1113 		goto out_unlock;
1114 
1115 	/*
1116 	 * When page table lock is held, the huge zero pud should not be
1117 	 * under splitting since we don't split the page itself, only pud to
1118 	 * a page table.
1119 	 */
1120 	if (is_huge_zero_pud(pud)) {
1121 		/* No huge zero pud yet */
1122 	}
1123 
1124 	pudp_set_wrprotect(src_mm, addr, src_pud);
1125 	pud = pud_mkold(pud_wrprotect(pud));
1126 	set_pud_at(dst_mm, addr, dst_pud, pud);
1127 
1128 	ret = 0;
1129 out_unlock:
1130 	spin_unlock(src_ptl);
1131 	spin_unlock(dst_ptl);
1132 	return ret;
1133 }
1134 
1135 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1136 {
1137 	pud_t entry;
1138 	unsigned long haddr;
1139 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1140 
1141 	vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1142 	if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1143 		goto unlock;
1144 
1145 	entry = pud_mkyoung(orig_pud);
1146 	if (write)
1147 		entry = pud_mkdirty(entry);
1148 	haddr = vmf->address & HPAGE_PUD_MASK;
1149 	if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1150 		update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1151 
1152 unlock:
1153 	spin_unlock(vmf->ptl);
1154 }
1155 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1156 
1157 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1158 {
1159 	pmd_t entry;
1160 	unsigned long haddr;
1161 	bool write = vmf->flags & FAULT_FLAG_WRITE;
1162 
1163 	vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1164 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1165 		goto unlock;
1166 
1167 	entry = pmd_mkyoung(orig_pmd);
1168 	if (write)
1169 		entry = pmd_mkdirty(entry);
1170 	haddr = vmf->address & HPAGE_PMD_MASK;
1171 	if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1172 		update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1173 
1174 unlock:
1175 	spin_unlock(vmf->ptl);
1176 }
1177 
1178 static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1179 			pmd_t orig_pmd, struct page *page)
1180 {
1181 	struct vm_area_struct *vma = vmf->vma;
1182 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1183 	struct mem_cgroup *memcg;
1184 	pgtable_t pgtable;
1185 	pmd_t _pmd;
1186 	int i;
1187 	vm_fault_t ret = 0;
1188 	struct page **pages;
1189 	struct mmu_notifier_range range;
1190 
1191 	pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1192 			      GFP_KERNEL);
1193 	if (unlikely(!pages)) {
1194 		ret |= VM_FAULT_OOM;
1195 		goto out;
1196 	}
1197 
1198 	for (i = 0; i < HPAGE_PMD_NR; i++) {
1199 		pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1200 					       vmf->address, page_to_nid(page));
1201 		if (unlikely(!pages[i] ||
1202 			     mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
1203 				     GFP_KERNEL, &memcg, false))) {
1204 			if (pages[i])
1205 				put_page(pages[i]);
1206 			while (--i >= 0) {
1207 				memcg = (void *)page_private(pages[i]);
1208 				set_page_private(pages[i], 0);
1209 				mem_cgroup_cancel_charge(pages[i], memcg,
1210 						false);
1211 				put_page(pages[i]);
1212 			}
1213 			kfree(pages);
1214 			ret |= VM_FAULT_OOM;
1215 			goto out;
1216 		}
1217 		set_page_private(pages[i], (unsigned long)memcg);
1218 	}
1219 
1220 	for (i = 0; i < HPAGE_PMD_NR; i++) {
1221 		copy_user_highpage(pages[i], page + i,
1222 				   haddr + PAGE_SIZE * i, vma);
1223 		__SetPageUptodate(pages[i]);
1224 		cond_resched();
1225 	}
1226 
1227 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1228 				haddr, haddr + HPAGE_PMD_SIZE);
1229 	mmu_notifier_invalidate_range_start(&range);
1230 
1231 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1232 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1233 		goto out_free_pages;
1234 	VM_BUG_ON_PAGE(!PageHead(page), page);
1235 
1236 	/*
1237 	 * Leave pmd empty until pte is filled note we must notify here as
1238 	 * concurrent CPU thread might write to new page before the call to
1239 	 * mmu_notifier_invalidate_range_end() happens which can lead to a
1240 	 * device seeing memory write in different order than CPU.
1241 	 *
1242 	 * See Documentation/vm/mmu_notifier.rst
1243 	 */
1244 	pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1245 
1246 	pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1247 	pmd_populate(vma->vm_mm, &_pmd, pgtable);
1248 
1249 	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1250 		pte_t entry;
1251 		entry = mk_pte(pages[i], vma->vm_page_prot);
1252 		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1253 		memcg = (void *)page_private(pages[i]);
1254 		set_page_private(pages[i], 0);
1255 		page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1256 		mem_cgroup_commit_charge(pages[i], memcg, false, false);
1257 		lru_cache_add_active_or_unevictable(pages[i], vma);
1258 		vmf->pte = pte_offset_map(&_pmd, haddr);
1259 		VM_BUG_ON(!pte_none(*vmf->pte));
1260 		set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1261 		pte_unmap(vmf->pte);
1262 	}
1263 	kfree(pages);
1264 
1265 	smp_wmb(); /* make pte visible before pmd */
1266 	pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1267 	page_remove_rmap(page, true);
1268 	spin_unlock(vmf->ptl);
1269 
1270 	/*
1271 	 * No need to double call mmu_notifier->invalidate_range() callback as
1272 	 * the above pmdp_huge_clear_flush_notify() did already call it.
1273 	 */
1274 	mmu_notifier_invalidate_range_only_end(&range);
1275 
1276 	ret |= VM_FAULT_WRITE;
1277 	put_page(page);
1278 
1279 out:
1280 	return ret;
1281 
1282 out_free_pages:
1283 	spin_unlock(vmf->ptl);
1284 	mmu_notifier_invalidate_range_end(&range);
1285 	for (i = 0; i < HPAGE_PMD_NR; i++) {
1286 		memcg = (void *)page_private(pages[i]);
1287 		set_page_private(pages[i], 0);
1288 		mem_cgroup_cancel_charge(pages[i], memcg, false);
1289 		put_page(pages[i]);
1290 	}
1291 	kfree(pages);
1292 	goto out;
1293 }
1294 
1295 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1296 {
1297 	struct vm_area_struct *vma = vmf->vma;
1298 	struct page *page = NULL, *new_page;
1299 	struct mem_cgroup *memcg;
1300 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1301 	struct mmu_notifier_range range;
1302 	gfp_t huge_gfp;			/* for allocation and charge */
1303 	vm_fault_t ret = 0;
1304 
1305 	vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1306 	VM_BUG_ON_VMA(!vma->anon_vma, vma);
1307 	if (is_huge_zero_pmd(orig_pmd))
1308 		goto alloc;
1309 	spin_lock(vmf->ptl);
1310 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1311 		goto out_unlock;
1312 
1313 	page = pmd_page(orig_pmd);
1314 	VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1315 	/*
1316 	 * We can only reuse the page if nobody else maps the huge page or it's
1317 	 * part.
1318 	 */
1319 	if (!trylock_page(page)) {
1320 		get_page(page);
1321 		spin_unlock(vmf->ptl);
1322 		lock_page(page);
1323 		spin_lock(vmf->ptl);
1324 		if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1325 			unlock_page(page);
1326 			put_page(page);
1327 			goto out_unlock;
1328 		}
1329 		put_page(page);
1330 	}
1331 	if (reuse_swap_page(page, NULL)) {
1332 		pmd_t entry;
1333 		entry = pmd_mkyoung(orig_pmd);
1334 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1335 		if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1336 			update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1337 		ret |= VM_FAULT_WRITE;
1338 		unlock_page(page);
1339 		goto out_unlock;
1340 	}
1341 	unlock_page(page);
1342 	get_page(page);
1343 	spin_unlock(vmf->ptl);
1344 alloc:
1345 	if (__transparent_hugepage_enabled(vma) &&
1346 	    !transparent_hugepage_debug_cow()) {
1347 		huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1348 		new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1349 	} else
1350 		new_page = NULL;
1351 
1352 	if (likely(new_page)) {
1353 		prep_transhuge_page(new_page);
1354 	} else {
1355 		if (!page) {
1356 			split_huge_pmd(vma, vmf->pmd, vmf->address);
1357 			ret |= VM_FAULT_FALLBACK;
1358 		} else {
1359 			ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1360 			if (ret & VM_FAULT_OOM) {
1361 				split_huge_pmd(vma, vmf->pmd, vmf->address);
1362 				ret |= VM_FAULT_FALLBACK;
1363 			}
1364 			put_page(page);
1365 		}
1366 		count_vm_event(THP_FAULT_FALLBACK);
1367 		goto out;
1368 	}
1369 
1370 	if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
1371 					huge_gfp, &memcg, true))) {
1372 		put_page(new_page);
1373 		split_huge_pmd(vma, vmf->pmd, vmf->address);
1374 		if (page)
1375 			put_page(page);
1376 		ret |= VM_FAULT_FALLBACK;
1377 		count_vm_event(THP_FAULT_FALLBACK);
1378 		goto out;
1379 	}
1380 
1381 	count_vm_event(THP_FAULT_ALLOC);
1382 	count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
1383 
1384 	if (!page)
1385 		clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1386 	else
1387 		copy_user_huge_page(new_page, page, vmf->address,
1388 				    vma, HPAGE_PMD_NR);
1389 	__SetPageUptodate(new_page);
1390 
1391 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1392 				haddr, haddr + HPAGE_PMD_SIZE);
1393 	mmu_notifier_invalidate_range_start(&range);
1394 
1395 	spin_lock(vmf->ptl);
1396 	if (page)
1397 		put_page(page);
1398 	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1399 		spin_unlock(vmf->ptl);
1400 		mem_cgroup_cancel_charge(new_page, memcg, true);
1401 		put_page(new_page);
1402 		goto out_mn;
1403 	} else {
1404 		pmd_t entry;
1405 		entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1406 		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1407 		pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1408 		page_add_new_anon_rmap(new_page, vma, haddr, true);
1409 		mem_cgroup_commit_charge(new_page, memcg, false, true);
1410 		lru_cache_add_active_or_unevictable(new_page, vma);
1411 		set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1412 		update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1413 		if (!page) {
1414 			add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1415 		} else {
1416 			VM_BUG_ON_PAGE(!PageHead(page), page);
1417 			page_remove_rmap(page, true);
1418 			put_page(page);
1419 		}
1420 		ret |= VM_FAULT_WRITE;
1421 	}
1422 	spin_unlock(vmf->ptl);
1423 out_mn:
1424 	/*
1425 	 * No need to double call mmu_notifier->invalidate_range() callback as
1426 	 * the above pmdp_huge_clear_flush_notify() did already call it.
1427 	 */
1428 	mmu_notifier_invalidate_range_only_end(&range);
1429 out:
1430 	return ret;
1431 out_unlock:
1432 	spin_unlock(vmf->ptl);
1433 	return ret;
1434 }
1435 
1436 /*
1437  * FOLL_FORCE can write to even unwritable pmd's, but only
1438  * after we've gone through a COW cycle and they are dirty.
1439  */
1440 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1441 {
1442 	return pmd_write(pmd) ||
1443 	       ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1444 }
1445 
1446 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1447 				   unsigned long addr,
1448 				   pmd_t *pmd,
1449 				   unsigned int flags)
1450 {
1451 	struct mm_struct *mm = vma->vm_mm;
1452 	struct page *page = NULL;
1453 
1454 	assert_spin_locked(pmd_lockptr(mm, pmd));
1455 
1456 	if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1457 		goto out;
1458 
1459 	/* Avoid dumping huge zero page */
1460 	if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1461 		return ERR_PTR(-EFAULT);
1462 
1463 	/* Full NUMA hinting faults to serialise migration in fault paths */
1464 	if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1465 		goto out;
1466 
1467 	page = pmd_page(*pmd);
1468 	VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1469 	if (flags & FOLL_TOUCH)
1470 		touch_pmd(vma, addr, pmd, flags);
1471 	if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1472 		/*
1473 		 * We don't mlock() pte-mapped THPs. This way we can avoid
1474 		 * leaking mlocked pages into non-VM_LOCKED VMAs.
1475 		 *
1476 		 * For anon THP:
1477 		 *
1478 		 * In most cases the pmd is the only mapping of the page as we
1479 		 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1480 		 * writable private mappings in populate_vma_page_range().
1481 		 *
1482 		 * The only scenario when we have the page shared here is if we
1483 		 * mlocking read-only mapping shared over fork(). We skip
1484 		 * mlocking such pages.
1485 		 *
1486 		 * For file THP:
1487 		 *
1488 		 * We can expect PageDoubleMap() to be stable under page lock:
1489 		 * for file pages we set it in page_add_file_rmap(), which
1490 		 * requires page to be locked.
1491 		 */
1492 
1493 		if (PageAnon(page) && compound_mapcount(page) != 1)
1494 			goto skip_mlock;
1495 		if (PageDoubleMap(page) || !page->mapping)
1496 			goto skip_mlock;
1497 		if (!trylock_page(page))
1498 			goto skip_mlock;
1499 		lru_add_drain();
1500 		if (page->mapping && !PageDoubleMap(page))
1501 			mlock_vma_page(page);
1502 		unlock_page(page);
1503 	}
1504 skip_mlock:
1505 	page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1506 	VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1507 	if (flags & FOLL_GET)
1508 		get_page(page);
1509 
1510 out:
1511 	return page;
1512 }
1513 
1514 /* NUMA hinting page fault entry point for trans huge pmds */
1515 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1516 {
1517 	struct vm_area_struct *vma = vmf->vma;
1518 	struct anon_vma *anon_vma = NULL;
1519 	struct page *page;
1520 	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1521 	int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1522 	int target_nid, last_cpupid = -1;
1523 	bool page_locked;
1524 	bool migrated = false;
1525 	bool was_writable;
1526 	int flags = 0;
1527 
1528 	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1529 	if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1530 		goto out_unlock;
1531 
1532 	/*
1533 	 * If there are potential migrations, wait for completion and retry
1534 	 * without disrupting NUMA hinting information. Do not relock and
1535 	 * check_same as the page may no longer be mapped.
1536 	 */
1537 	if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1538 		page = pmd_page(*vmf->pmd);
1539 		if (!get_page_unless_zero(page))
1540 			goto out_unlock;
1541 		spin_unlock(vmf->ptl);
1542 		put_and_wait_on_page_locked(page);
1543 		goto out;
1544 	}
1545 
1546 	page = pmd_page(pmd);
1547 	BUG_ON(is_huge_zero_page(page));
1548 	page_nid = page_to_nid(page);
1549 	last_cpupid = page_cpupid_last(page);
1550 	count_vm_numa_event(NUMA_HINT_FAULTS);
1551 	if (page_nid == this_nid) {
1552 		count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1553 		flags |= TNF_FAULT_LOCAL;
1554 	}
1555 
1556 	/* See similar comment in do_numa_page for explanation */
1557 	if (!pmd_savedwrite(pmd))
1558 		flags |= TNF_NO_GROUP;
1559 
1560 	/*
1561 	 * Acquire the page lock to serialise THP migrations but avoid dropping
1562 	 * page_table_lock if at all possible
1563 	 */
1564 	page_locked = trylock_page(page);
1565 	target_nid = mpol_misplaced(page, vma, haddr);
1566 	if (target_nid == NUMA_NO_NODE) {
1567 		/* If the page was locked, there are no parallel migrations */
1568 		if (page_locked)
1569 			goto clear_pmdnuma;
1570 	}
1571 
1572 	/* Migration could have started since the pmd_trans_migrating check */
1573 	if (!page_locked) {
1574 		page_nid = NUMA_NO_NODE;
1575 		if (!get_page_unless_zero(page))
1576 			goto out_unlock;
1577 		spin_unlock(vmf->ptl);
1578 		put_and_wait_on_page_locked(page);
1579 		goto out;
1580 	}
1581 
1582 	/*
1583 	 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1584 	 * to serialises splits
1585 	 */
1586 	get_page(page);
1587 	spin_unlock(vmf->ptl);
1588 	anon_vma = page_lock_anon_vma_read(page);
1589 
1590 	/* Confirm the PMD did not change while page_table_lock was released */
1591 	spin_lock(vmf->ptl);
1592 	if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1593 		unlock_page(page);
1594 		put_page(page);
1595 		page_nid = NUMA_NO_NODE;
1596 		goto out_unlock;
1597 	}
1598 
1599 	/* Bail if we fail to protect against THP splits for any reason */
1600 	if (unlikely(!anon_vma)) {
1601 		put_page(page);
1602 		page_nid = NUMA_NO_NODE;
1603 		goto clear_pmdnuma;
1604 	}
1605 
1606 	/*
1607 	 * Since we took the NUMA fault, we must have observed the !accessible
1608 	 * bit. Make sure all other CPUs agree with that, to avoid them
1609 	 * modifying the page we're about to migrate.
1610 	 *
1611 	 * Must be done under PTL such that we'll observe the relevant
1612 	 * inc_tlb_flush_pending().
1613 	 *
1614 	 * We are not sure a pending tlb flush here is for a huge page
1615 	 * mapping or not. Hence use the tlb range variant
1616 	 */
1617 	if (mm_tlb_flush_pending(vma->vm_mm)) {
1618 		flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1619 		/*
1620 		 * change_huge_pmd() released the pmd lock before
1621 		 * invalidating the secondary MMUs sharing the primary
1622 		 * MMU pagetables (with ->invalidate_range()). The
1623 		 * mmu_notifier_invalidate_range_end() (which
1624 		 * internally calls ->invalidate_range()) in
1625 		 * change_pmd_range() will run after us, so we can't
1626 		 * rely on it here and we need an explicit invalidate.
1627 		 */
1628 		mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1629 					      haddr + HPAGE_PMD_SIZE);
1630 	}
1631 
1632 	/*
1633 	 * Migrate the THP to the requested node, returns with page unlocked
1634 	 * and access rights restored.
1635 	 */
1636 	spin_unlock(vmf->ptl);
1637 
1638 	migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1639 				vmf->pmd, pmd, vmf->address, page, target_nid);
1640 	if (migrated) {
1641 		flags |= TNF_MIGRATED;
1642 		page_nid = target_nid;
1643 	} else
1644 		flags |= TNF_MIGRATE_FAIL;
1645 
1646 	goto out;
1647 clear_pmdnuma:
1648 	BUG_ON(!PageLocked(page));
1649 	was_writable = pmd_savedwrite(pmd);
1650 	pmd = pmd_modify(pmd, vma->vm_page_prot);
1651 	pmd = pmd_mkyoung(pmd);
1652 	if (was_writable)
1653 		pmd = pmd_mkwrite(pmd);
1654 	set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1655 	update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1656 	unlock_page(page);
1657 out_unlock:
1658 	spin_unlock(vmf->ptl);
1659 
1660 out:
1661 	if (anon_vma)
1662 		page_unlock_anon_vma_read(anon_vma);
1663 
1664 	if (page_nid != NUMA_NO_NODE)
1665 		task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1666 				flags);
1667 
1668 	return 0;
1669 }
1670 
1671 /*
1672  * Return true if we do MADV_FREE successfully on entire pmd page.
1673  * Otherwise, return false.
1674  */
1675 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1676 		pmd_t *pmd, unsigned long addr, unsigned long next)
1677 {
1678 	spinlock_t *ptl;
1679 	pmd_t orig_pmd;
1680 	struct page *page;
1681 	struct mm_struct *mm = tlb->mm;
1682 	bool ret = false;
1683 
1684 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1685 
1686 	ptl = pmd_trans_huge_lock(pmd, vma);
1687 	if (!ptl)
1688 		goto out_unlocked;
1689 
1690 	orig_pmd = *pmd;
1691 	if (is_huge_zero_pmd(orig_pmd))
1692 		goto out;
1693 
1694 	if (unlikely(!pmd_present(orig_pmd))) {
1695 		VM_BUG_ON(thp_migration_supported() &&
1696 				  !is_pmd_migration_entry(orig_pmd));
1697 		goto out;
1698 	}
1699 
1700 	page = pmd_page(orig_pmd);
1701 	/*
1702 	 * If other processes are mapping this page, we couldn't discard
1703 	 * the page unless they all do MADV_FREE so let's skip the page.
1704 	 */
1705 	if (page_mapcount(page) != 1)
1706 		goto out;
1707 
1708 	if (!trylock_page(page))
1709 		goto out;
1710 
1711 	/*
1712 	 * If user want to discard part-pages of THP, split it so MADV_FREE
1713 	 * will deactivate only them.
1714 	 */
1715 	if (next - addr != HPAGE_PMD_SIZE) {
1716 		get_page(page);
1717 		spin_unlock(ptl);
1718 		split_huge_page(page);
1719 		unlock_page(page);
1720 		put_page(page);
1721 		goto out_unlocked;
1722 	}
1723 
1724 	if (PageDirty(page))
1725 		ClearPageDirty(page);
1726 	unlock_page(page);
1727 
1728 	if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1729 		pmdp_invalidate(vma, addr, pmd);
1730 		orig_pmd = pmd_mkold(orig_pmd);
1731 		orig_pmd = pmd_mkclean(orig_pmd);
1732 
1733 		set_pmd_at(mm, addr, pmd, orig_pmd);
1734 		tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1735 	}
1736 
1737 	mark_page_lazyfree(page);
1738 	ret = true;
1739 out:
1740 	spin_unlock(ptl);
1741 out_unlocked:
1742 	return ret;
1743 }
1744 
1745 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1746 {
1747 	pgtable_t pgtable;
1748 
1749 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1750 	pte_free(mm, pgtable);
1751 	mm_dec_nr_ptes(mm);
1752 }
1753 
1754 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1755 		 pmd_t *pmd, unsigned long addr)
1756 {
1757 	pmd_t orig_pmd;
1758 	spinlock_t *ptl;
1759 
1760 	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1761 
1762 	ptl = __pmd_trans_huge_lock(pmd, vma);
1763 	if (!ptl)
1764 		return 0;
1765 	/*
1766 	 * For architectures like ppc64 we look at deposited pgtable
1767 	 * when calling pmdp_huge_get_and_clear. So do the
1768 	 * pgtable_trans_huge_withdraw after finishing pmdp related
1769 	 * operations.
1770 	 */
1771 	orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1772 			tlb->fullmm);
1773 	tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1774 	if (vma_is_dax(vma)) {
1775 		if (arch_needs_pgtable_deposit())
1776 			zap_deposited_table(tlb->mm, pmd);
1777 		spin_unlock(ptl);
1778 		if (is_huge_zero_pmd(orig_pmd))
1779 			tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1780 	} else if (is_huge_zero_pmd(orig_pmd)) {
1781 		zap_deposited_table(tlb->mm, pmd);
1782 		spin_unlock(ptl);
1783 		tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1784 	} else {
1785 		struct page *page = NULL;
1786 		int flush_needed = 1;
1787 
1788 		if (pmd_present(orig_pmd)) {
1789 			page = pmd_page(orig_pmd);
1790 			page_remove_rmap(page, true);
1791 			VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1792 			VM_BUG_ON_PAGE(!PageHead(page), page);
1793 		} else if (thp_migration_supported()) {
1794 			swp_entry_t entry;
1795 
1796 			VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1797 			entry = pmd_to_swp_entry(orig_pmd);
1798 			page = pfn_to_page(swp_offset(entry));
1799 			flush_needed = 0;
1800 		} else
1801 			WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1802 
1803 		if (PageAnon(page)) {
1804 			zap_deposited_table(tlb->mm, pmd);
1805 			add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1806 		} else {
1807 			if (arch_needs_pgtable_deposit())
1808 				zap_deposited_table(tlb->mm, pmd);
1809 			add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1810 		}
1811 
1812 		spin_unlock(ptl);
1813 		if (flush_needed)
1814 			tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1815 	}
1816 	return 1;
1817 }
1818 
1819 #ifndef pmd_move_must_withdraw
1820 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1821 					 spinlock_t *old_pmd_ptl,
1822 					 struct vm_area_struct *vma)
1823 {
1824 	/*
1825 	 * With split pmd lock we also need to move preallocated
1826 	 * PTE page table if new_pmd is on different PMD page table.
1827 	 *
1828 	 * We also don't deposit and withdraw tables for file pages.
1829 	 */
1830 	return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1831 }
1832 #endif
1833 
1834 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1835 {
1836 #ifdef CONFIG_MEM_SOFT_DIRTY
1837 	if (unlikely(is_pmd_migration_entry(pmd)))
1838 		pmd = pmd_swp_mksoft_dirty(pmd);
1839 	else if (pmd_present(pmd))
1840 		pmd = pmd_mksoft_dirty(pmd);
1841 #endif
1842 	return pmd;
1843 }
1844 
1845 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1846 		  unsigned long new_addr, unsigned long old_end,
1847 		  pmd_t *old_pmd, pmd_t *new_pmd)
1848 {
1849 	spinlock_t *old_ptl, *new_ptl;
1850 	pmd_t pmd;
1851 	struct mm_struct *mm = vma->vm_mm;
1852 	bool force_flush = false;
1853 
1854 	if ((old_addr & ~HPAGE_PMD_MASK) ||
1855 	    (new_addr & ~HPAGE_PMD_MASK) ||
1856 	    old_end - old_addr < HPAGE_PMD_SIZE)
1857 		return false;
1858 
1859 	/*
1860 	 * The destination pmd shouldn't be established, free_pgtables()
1861 	 * should have release it.
1862 	 */
1863 	if (WARN_ON(!pmd_none(*new_pmd))) {
1864 		VM_BUG_ON(pmd_trans_huge(*new_pmd));
1865 		return false;
1866 	}
1867 
1868 	/*
1869 	 * We don't have to worry about the ordering of src and dst
1870 	 * ptlocks because exclusive mmap_sem prevents deadlock.
1871 	 */
1872 	old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1873 	if (old_ptl) {
1874 		new_ptl = pmd_lockptr(mm, new_pmd);
1875 		if (new_ptl != old_ptl)
1876 			spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1877 		pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1878 		if (pmd_present(pmd))
1879 			force_flush = true;
1880 		VM_BUG_ON(!pmd_none(*new_pmd));
1881 
1882 		if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1883 			pgtable_t pgtable;
1884 			pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1885 			pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1886 		}
1887 		pmd = move_soft_dirty_pmd(pmd);
1888 		set_pmd_at(mm, new_addr, new_pmd, pmd);
1889 		if (force_flush)
1890 			flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1891 		if (new_ptl != old_ptl)
1892 			spin_unlock(new_ptl);
1893 		spin_unlock(old_ptl);
1894 		return true;
1895 	}
1896 	return false;
1897 }
1898 
1899 /*
1900  * Returns
1901  *  - 0 if PMD could not be locked
1902  *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1903  *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1904  */
1905 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1906 		unsigned long addr, pgprot_t newprot, int prot_numa)
1907 {
1908 	struct mm_struct *mm = vma->vm_mm;
1909 	spinlock_t *ptl;
1910 	pmd_t entry;
1911 	bool preserve_write;
1912 	int ret;
1913 
1914 	ptl = __pmd_trans_huge_lock(pmd, vma);
1915 	if (!ptl)
1916 		return 0;
1917 
1918 	preserve_write = prot_numa && pmd_write(*pmd);
1919 	ret = 1;
1920 
1921 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1922 	if (is_swap_pmd(*pmd)) {
1923 		swp_entry_t entry = pmd_to_swp_entry(*pmd);
1924 
1925 		VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1926 		if (is_write_migration_entry(entry)) {
1927 			pmd_t newpmd;
1928 			/*
1929 			 * A protection check is difficult so
1930 			 * just be safe and disable write
1931 			 */
1932 			make_migration_entry_read(&entry);
1933 			newpmd = swp_entry_to_pmd(entry);
1934 			if (pmd_swp_soft_dirty(*pmd))
1935 				newpmd = pmd_swp_mksoft_dirty(newpmd);
1936 			set_pmd_at(mm, addr, pmd, newpmd);
1937 		}
1938 		goto unlock;
1939 	}
1940 #endif
1941 
1942 	/*
1943 	 * Avoid trapping faults against the zero page. The read-only
1944 	 * data is likely to be read-cached on the local CPU and
1945 	 * local/remote hits to the zero page are not interesting.
1946 	 */
1947 	if (prot_numa && is_huge_zero_pmd(*pmd))
1948 		goto unlock;
1949 
1950 	if (prot_numa && pmd_protnone(*pmd))
1951 		goto unlock;
1952 
1953 	/*
1954 	 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1955 	 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1956 	 * which is also under down_read(mmap_sem):
1957 	 *
1958 	 *	CPU0:				CPU1:
1959 	 *				change_huge_pmd(prot_numa=1)
1960 	 *				 pmdp_huge_get_and_clear_notify()
1961 	 * madvise_dontneed()
1962 	 *  zap_pmd_range()
1963 	 *   pmd_trans_huge(*pmd) == 0 (without ptl)
1964 	 *   // skip the pmd
1965 	 *				 set_pmd_at();
1966 	 *				 // pmd is re-established
1967 	 *
1968 	 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1969 	 * which may break userspace.
1970 	 *
1971 	 * pmdp_invalidate() is required to make sure we don't miss
1972 	 * dirty/young flags set by hardware.
1973 	 */
1974 	entry = pmdp_invalidate(vma, addr, pmd);
1975 
1976 	entry = pmd_modify(entry, newprot);
1977 	if (preserve_write)
1978 		entry = pmd_mk_savedwrite(entry);
1979 	ret = HPAGE_PMD_NR;
1980 	set_pmd_at(mm, addr, pmd, entry);
1981 	BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1982 unlock:
1983 	spin_unlock(ptl);
1984 	return ret;
1985 }
1986 
1987 /*
1988  * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1989  *
1990  * Note that if it returns page table lock pointer, this routine returns without
1991  * unlocking page table lock. So callers must unlock it.
1992  */
1993 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1994 {
1995 	spinlock_t *ptl;
1996 	ptl = pmd_lock(vma->vm_mm, pmd);
1997 	if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1998 			pmd_devmap(*pmd)))
1999 		return ptl;
2000 	spin_unlock(ptl);
2001 	return NULL;
2002 }
2003 
2004 /*
2005  * Returns true if a given pud maps a thp, false otherwise.
2006  *
2007  * Note that if it returns true, this routine returns without unlocking page
2008  * table lock. So callers must unlock it.
2009  */
2010 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2011 {
2012 	spinlock_t *ptl;
2013 
2014 	ptl = pud_lock(vma->vm_mm, pud);
2015 	if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2016 		return ptl;
2017 	spin_unlock(ptl);
2018 	return NULL;
2019 }
2020 
2021 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2022 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2023 		 pud_t *pud, unsigned long addr)
2024 {
2025 	spinlock_t *ptl;
2026 
2027 	ptl = __pud_trans_huge_lock(pud, vma);
2028 	if (!ptl)
2029 		return 0;
2030 	/*
2031 	 * For architectures like ppc64 we look at deposited pgtable
2032 	 * when calling pudp_huge_get_and_clear. So do the
2033 	 * pgtable_trans_huge_withdraw after finishing pudp related
2034 	 * operations.
2035 	 */
2036 	pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
2037 	tlb_remove_pud_tlb_entry(tlb, pud, addr);
2038 	if (vma_is_dax(vma)) {
2039 		spin_unlock(ptl);
2040 		/* No zero page support yet */
2041 	} else {
2042 		/* No support for anonymous PUD pages yet */
2043 		BUG();
2044 	}
2045 	return 1;
2046 }
2047 
2048 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2049 		unsigned long haddr)
2050 {
2051 	VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2052 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2053 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2054 	VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2055 
2056 	count_vm_event(THP_SPLIT_PUD);
2057 
2058 	pudp_huge_clear_flush_notify(vma, haddr, pud);
2059 }
2060 
2061 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2062 		unsigned long address)
2063 {
2064 	spinlock_t *ptl;
2065 	struct mmu_notifier_range range;
2066 
2067 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2068 				address & HPAGE_PUD_MASK,
2069 				(address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2070 	mmu_notifier_invalidate_range_start(&range);
2071 	ptl = pud_lock(vma->vm_mm, pud);
2072 	if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2073 		goto out;
2074 	__split_huge_pud_locked(vma, pud, range.start);
2075 
2076 out:
2077 	spin_unlock(ptl);
2078 	/*
2079 	 * No need to double call mmu_notifier->invalidate_range() callback as
2080 	 * the above pudp_huge_clear_flush_notify() did already call it.
2081 	 */
2082 	mmu_notifier_invalidate_range_only_end(&range);
2083 }
2084 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2085 
2086 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2087 		unsigned long haddr, pmd_t *pmd)
2088 {
2089 	struct mm_struct *mm = vma->vm_mm;
2090 	pgtable_t pgtable;
2091 	pmd_t _pmd;
2092 	int i;
2093 
2094 	/*
2095 	 * Leave pmd empty until pte is filled note that it is fine to delay
2096 	 * notification until mmu_notifier_invalidate_range_end() as we are
2097 	 * replacing a zero pmd write protected page with a zero pte write
2098 	 * protected page.
2099 	 *
2100 	 * See Documentation/vm/mmu_notifier.rst
2101 	 */
2102 	pmdp_huge_clear_flush(vma, haddr, pmd);
2103 
2104 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2105 	pmd_populate(mm, &_pmd, pgtable);
2106 
2107 	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2108 		pte_t *pte, entry;
2109 		entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2110 		entry = pte_mkspecial(entry);
2111 		pte = pte_offset_map(&_pmd, haddr);
2112 		VM_BUG_ON(!pte_none(*pte));
2113 		set_pte_at(mm, haddr, pte, entry);
2114 		pte_unmap(pte);
2115 	}
2116 	smp_wmb(); /* make pte visible before pmd */
2117 	pmd_populate(mm, pmd, pgtable);
2118 }
2119 
2120 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2121 		unsigned long haddr, bool freeze)
2122 {
2123 	struct mm_struct *mm = vma->vm_mm;
2124 	struct page *page;
2125 	pgtable_t pgtable;
2126 	pmd_t old_pmd, _pmd;
2127 	bool young, write, soft_dirty, pmd_migration = false;
2128 	unsigned long addr;
2129 	int i;
2130 
2131 	VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2132 	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2133 	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2134 	VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2135 				&& !pmd_devmap(*pmd));
2136 
2137 	count_vm_event(THP_SPLIT_PMD);
2138 
2139 	if (!vma_is_anonymous(vma)) {
2140 		_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2141 		/*
2142 		 * We are going to unmap this huge page. So
2143 		 * just go ahead and zap it
2144 		 */
2145 		if (arch_needs_pgtable_deposit())
2146 			zap_deposited_table(mm, pmd);
2147 		if (vma_is_dax(vma))
2148 			return;
2149 		page = pmd_page(_pmd);
2150 		if (!PageDirty(page) && pmd_dirty(_pmd))
2151 			set_page_dirty(page);
2152 		if (!PageReferenced(page) && pmd_young(_pmd))
2153 			SetPageReferenced(page);
2154 		page_remove_rmap(page, true);
2155 		put_page(page);
2156 		add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2157 		return;
2158 	} else if (is_huge_zero_pmd(*pmd)) {
2159 		/*
2160 		 * FIXME: Do we want to invalidate secondary mmu by calling
2161 		 * mmu_notifier_invalidate_range() see comments below inside
2162 		 * __split_huge_pmd() ?
2163 		 *
2164 		 * We are going from a zero huge page write protected to zero
2165 		 * small page also write protected so it does not seems useful
2166 		 * to invalidate secondary mmu at this time.
2167 		 */
2168 		return __split_huge_zero_page_pmd(vma, haddr, pmd);
2169 	}
2170 
2171 	/*
2172 	 * Up to this point the pmd is present and huge and userland has the
2173 	 * whole access to the hugepage during the split (which happens in
2174 	 * place). If we overwrite the pmd with the not-huge version pointing
2175 	 * to the pte here (which of course we could if all CPUs were bug
2176 	 * free), userland could trigger a small page size TLB miss on the
2177 	 * small sized TLB while the hugepage TLB entry is still established in
2178 	 * the huge TLB. Some CPU doesn't like that.
2179 	 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2180 	 * 383 on page 93. Intel should be safe but is also warns that it's
2181 	 * only safe if the permission and cache attributes of the two entries
2182 	 * loaded in the two TLB is identical (which should be the case here).
2183 	 * But it is generally safer to never allow small and huge TLB entries
2184 	 * for the same virtual address to be loaded simultaneously. So instead
2185 	 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2186 	 * current pmd notpresent (atomically because here the pmd_trans_huge
2187 	 * must remain set at all times on the pmd until the split is complete
2188 	 * for this pmd), then we flush the SMP TLB and finally we write the
2189 	 * non-huge version of the pmd entry with pmd_populate.
2190 	 */
2191 	old_pmd = pmdp_invalidate(vma, haddr, pmd);
2192 
2193 	pmd_migration = is_pmd_migration_entry(old_pmd);
2194 	if (unlikely(pmd_migration)) {
2195 		swp_entry_t entry;
2196 
2197 		entry = pmd_to_swp_entry(old_pmd);
2198 		page = pfn_to_page(swp_offset(entry));
2199 		write = is_write_migration_entry(entry);
2200 		young = false;
2201 		soft_dirty = pmd_swp_soft_dirty(old_pmd);
2202 	} else {
2203 		page = pmd_page(old_pmd);
2204 		if (pmd_dirty(old_pmd))
2205 			SetPageDirty(page);
2206 		write = pmd_write(old_pmd);
2207 		young = pmd_young(old_pmd);
2208 		soft_dirty = pmd_soft_dirty(old_pmd);
2209 	}
2210 	VM_BUG_ON_PAGE(!page_count(page), page);
2211 	page_ref_add(page, HPAGE_PMD_NR - 1);
2212 
2213 	/*
2214 	 * Withdraw the table only after we mark the pmd entry invalid.
2215 	 * This's critical for some architectures (Power).
2216 	 */
2217 	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2218 	pmd_populate(mm, &_pmd, pgtable);
2219 
2220 	for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2221 		pte_t entry, *pte;
2222 		/*
2223 		 * Note that NUMA hinting access restrictions are not
2224 		 * transferred to avoid any possibility of altering
2225 		 * permissions across VMAs.
2226 		 */
2227 		if (freeze || pmd_migration) {
2228 			swp_entry_t swp_entry;
2229 			swp_entry = make_migration_entry(page + i, write);
2230 			entry = swp_entry_to_pte(swp_entry);
2231 			if (soft_dirty)
2232 				entry = pte_swp_mksoft_dirty(entry);
2233 		} else {
2234 			entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2235 			entry = maybe_mkwrite(entry, vma);
2236 			if (!write)
2237 				entry = pte_wrprotect(entry);
2238 			if (!young)
2239 				entry = pte_mkold(entry);
2240 			if (soft_dirty)
2241 				entry = pte_mksoft_dirty(entry);
2242 		}
2243 		pte = pte_offset_map(&_pmd, addr);
2244 		BUG_ON(!pte_none(*pte));
2245 		set_pte_at(mm, addr, pte, entry);
2246 		atomic_inc(&page[i]._mapcount);
2247 		pte_unmap(pte);
2248 	}
2249 
2250 	/*
2251 	 * Set PG_double_map before dropping compound_mapcount to avoid
2252 	 * false-negative page_mapped().
2253 	 */
2254 	if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2255 		for (i = 0; i < HPAGE_PMD_NR; i++)
2256 			atomic_inc(&page[i]._mapcount);
2257 	}
2258 
2259 	if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2260 		/* Last compound_mapcount is gone. */
2261 		__dec_node_page_state(page, NR_ANON_THPS);
2262 		if (TestClearPageDoubleMap(page)) {
2263 			/* No need in mapcount reference anymore */
2264 			for (i = 0; i < HPAGE_PMD_NR; i++)
2265 				atomic_dec(&page[i]._mapcount);
2266 		}
2267 	}
2268 
2269 	smp_wmb(); /* make pte visible before pmd */
2270 	pmd_populate(mm, pmd, pgtable);
2271 
2272 	if (freeze) {
2273 		for (i = 0; i < HPAGE_PMD_NR; i++) {
2274 			page_remove_rmap(page + i, false);
2275 			put_page(page + i);
2276 		}
2277 	}
2278 }
2279 
2280 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2281 		unsigned long address, bool freeze, struct page *page)
2282 {
2283 	spinlock_t *ptl;
2284 	struct mmu_notifier_range range;
2285 
2286 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2287 				address & HPAGE_PMD_MASK,
2288 				(address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2289 	mmu_notifier_invalidate_range_start(&range);
2290 	ptl = pmd_lock(vma->vm_mm, pmd);
2291 
2292 	/*
2293 	 * If caller asks to setup a migration entries, we need a page to check
2294 	 * pmd against. Otherwise we can end up replacing wrong page.
2295 	 */
2296 	VM_BUG_ON(freeze && !page);
2297 	if (page && page != pmd_page(*pmd))
2298 	        goto out;
2299 
2300 	if (pmd_trans_huge(*pmd)) {
2301 		page = pmd_page(*pmd);
2302 		if (PageMlocked(page))
2303 			clear_page_mlock(page);
2304 	} else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2305 		goto out;
2306 	__split_huge_pmd_locked(vma, pmd, range.start, freeze);
2307 out:
2308 	spin_unlock(ptl);
2309 	/*
2310 	 * No need to double call mmu_notifier->invalidate_range() callback.
2311 	 * They are 3 cases to consider inside __split_huge_pmd_locked():
2312 	 *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2313 	 *  2) __split_huge_zero_page_pmd() read only zero page and any write
2314 	 *    fault will trigger a flush_notify before pointing to a new page
2315 	 *    (it is fine if the secondary mmu keeps pointing to the old zero
2316 	 *    page in the meantime)
2317 	 *  3) Split a huge pmd into pte pointing to the same page. No need
2318 	 *     to invalidate secondary tlb entry they are all still valid.
2319 	 *     any further changes to individual pte will notify. So no need
2320 	 *     to call mmu_notifier->invalidate_range()
2321 	 */
2322 	mmu_notifier_invalidate_range_only_end(&range);
2323 }
2324 
2325 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2326 		bool freeze, struct page *page)
2327 {
2328 	pgd_t *pgd;
2329 	p4d_t *p4d;
2330 	pud_t *pud;
2331 	pmd_t *pmd;
2332 
2333 	pgd = pgd_offset(vma->vm_mm, address);
2334 	if (!pgd_present(*pgd))
2335 		return;
2336 
2337 	p4d = p4d_offset(pgd, address);
2338 	if (!p4d_present(*p4d))
2339 		return;
2340 
2341 	pud = pud_offset(p4d, address);
2342 	if (!pud_present(*pud))
2343 		return;
2344 
2345 	pmd = pmd_offset(pud, address);
2346 
2347 	__split_huge_pmd(vma, pmd, address, freeze, page);
2348 }
2349 
2350 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2351 			     unsigned long start,
2352 			     unsigned long end,
2353 			     long adjust_next)
2354 {
2355 	/*
2356 	 * If the new start address isn't hpage aligned and it could
2357 	 * previously contain an hugepage: check if we need to split
2358 	 * an huge pmd.
2359 	 */
2360 	if (start & ~HPAGE_PMD_MASK &&
2361 	    (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2362 	    (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2363 		split_huge_pmd_address(vma, start, false, NULL);
2364 
2365 	/*
2366 	 * If the new end address isn't hpage aligned and it could
2367 	 * previously contain an hugepage: check if we need to split
2368 	 * an huge pmd.
2369 	 */
2370 	if (end & ~HPAGE_PMD_MASK &&
2371 	    (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2372 	    (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2373 		split_huge_pmd_address(vma, end, false, NULL);
2374 
2375 	/*
2376 	 * If we're also updating the vma->vm_next->vm_start, if the new
2377 	 * vm_next->vm_start isn't page aligned and it could previously
2378 	 * contain an hugepage: check if we need to split an huge pmd.
2379 	 */
2380 	if (adjust_next > 0) {
2381 		struct vm_area_struct *next = vma->vm_next;
2382 		unsigned long nstart = next->vm_start;
2383 		nstart += adjust_next << PAGE_SHIFT;
2384 		if (nstart & ~HPAGE_PMD_MASK &&
2385 		    (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2386 		    (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2387 			split_huge_pmd_address(next, nstart, false, NULL);
2388 	}
2389 }
2390 
2391 static void unmap_page(struct page *page)
2392 {
2393 	enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2394 		TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2395 	bool unmap_success;
2396 
2397 	VM_BUG_ON_PAGE(!PageHead(page), page);
2398 
2399 	if (PageAnon(page))
2400 		ttu_flags |= TTU_SPLIT_FREEZE;
2401 
2402 	unmap_success = try_to_unmap(page, ttu_flags);
2403 	VM_BUG_ON_PAGE(!unmap_success, page);
2404 }
2405 
2406 static void remap_page(struct page *page)
2407 {
2408 	int i;
2409 	if (PageTransHuge(page)) {
2410 		remove_migration_ptes(page, page, true);
2411 	} else {
2412 		for (i = 0; i < HPAGE_PMD_NR; i++)
2413 			remove_migration_ptes(page + i, page + i, true);
2414 	}
2415 }
2416 
2417 static void __split_huge_page_tail(struct page *head, int tail,
2418 		struct lruvec *lruvec, struct list_head *list)
2419 {
2420 	struct page *page_tail = head + tail;
2421 
2422 	VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2423 
2424 	/*
2425 	 * Clone page flags before unfreezing refcount.
2426 	 *
2427 	 * After successful get_page_unless_zero() might follow flags change,
2428 	 * for exmaple lock_page() which set PG_waiters.
2429 	 */
2430 	page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2431 	page_tail->flags |= (head->flags &
2432 			((1L << PG_referenced) |
2433 			 (1L << PG_swapbacked) |
2434 			 (1L << PG_swapcache) |
2435 			 (1L << PG_mlocked) |
2436 			 (1L << PG_uptodate) |
2437 			 (1L << PG_active) |
2438 			 (1L << PG_workingset) |
2439 			 (1L << PG_locked) |
2440 			 (1L << PG_unevictable) |
2441 			 (1L << PG_dirty)));
2442 
2443 	/* ->mapping in first tail page is compound_mapcount */
2444 	VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2445 			page_tail);
2446 	page_tail->mapping = head->mapping;
2447 	page_tail->index = head->index + tail;
2448 
2449 	/* Page flags must be visible before we make the page non-compound. */
2450 	smp_wmb();
2451 
2452 	/*
2453 	 * Clear PageTail before unfreezing page refcount.
2454 	 *
2455 	 * After successful get_page_unless_zero() might follow put_page()
2456 	 * which needs correct compound_head().
2457 	 */
2458 	clear_compound_head(page_tail);
2459 
2460 	/* Finally unfreeze refcount. Additional reference from page cache. */
2461 	page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2462 					  PageSwapCache(head)));
2463 
2464 	if (page_is_young(head))
2465 		set_page_young(page_tail);
2466 	if (page_is_idle(head))
2467 		set_page_idle(page_tail);
2468 
2469 	page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2470 
2471 	/*
2472 	 * always add to the tail because some iterators expect new
2473 	 * pages to show after the currently processed elements - e.g.
2474 	 * migrate_pages
2475 	 */
2476 	lru_add_page_tail(head, page_tail, lruvec, list);
2477 }
2478 
2479 static void __split_huge_page(struct page *page, struct list_head *list,
2480 		pgoff_t end, unsigned long flags)
2481 {
2482 	struct page *head = compound_head(page);
2483 	pg_data_t *pgdat = page_pgdat(head);
2484 	struct lruvec *lruvec;
2485 	int i;
2486 
2487 	lruvec = mem_cgroup_page_lruvec(head, pgdat);
2488 
2489 	/* complete memcg works before add pages to LRU */
2490 	mem_cgroup_split_huge_fixup(head);
2491 
2492 	for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2493 		__split_huge_page_tail(head, i, lruvec, list);
2494 		/* Some pages can be beyond i_size: drop them from page cache */
2495 		if (head[i].index >= end) {
2496 			ClearPageDirty(head + i);
2497 			__delete_from_page_cache(head + i, NULL);
2498 			if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2499 				shmem_uncharge(head->mapping->host, 1);
2500 			put_page(head + i);
2501 		} else if (!PageAnon(page)) {
2502 			__xa_store(&head->mapping->i_pages, head[i].index,
2503 					head + i, 0);
2504 		}
2505 	}
2506 
2507 	ClearPageCompound(head);
2508 	/* See comment in __split_huge_page_tail() */
2509 	if (PageAnon(head)) {
2510 		/* Additional pin to swap cache */
2511 		if (PageSwapCache(head))
2512 			page_ref_add(head, 2);
2513 		else
2514 			page_ref_inc(head);
2515 	} else {
2516 		/* Additional pin to page cache */
2517 		page_ref_add(head, 2);
2518 		xa_unlock(&head->mapping->i_pages);
2519 	}
2520 
2521 	spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2522 
2523 	remap_page(head);
2524 
2525 	for (i = 0; i < HPAGE_PMD_NR; i++) {
2526 		struct page *subpage = head + i;
2527 		if (subpage == page)
2528 			continue;
2529 		unlock_page(subpage);
2530 
2531 		/*
2532 		 * Subpages may be freed if there wasn't any mapping
2533 		 * like if add_to_swap() is running on a lru page that
2534 		 * had its mapping zapped. And freeing these pages
2535 		 * requires taking the lru_lock so we do the put_page
2536 		 * of the tail pages after the split is complete.
2537 		 */
2538 		put_page(subpage);
2539 	}
2540 }
2541 
2542 int total_mapcount(struct page *page)
2543 {
2544 	int i, compound, ret;
2545 
2546 	VM_BUG_ON_PAGE(PageTail(page), page);
2547 
2548 	if (likely(!PageCompound(page)))
2549 		return atomic_read(&page->_mapcount) + 1;
2550 
2551 	compound = compound_mapcount(page);
2552 	if (PageHuge(page))
2553 		return compound;
2554 	ret = compound;
2555 	for (i = 0; i < HPAGE_PMD_NR; i++)
2556 		ret += atomic_read(&page[i]._mapcount) + 1;
2557 	/* File pages has compound_mapcount included in _mapcount */
2558 	if (!PageAnon(page))
2559 		return ret - compound * HPAGE_PMD_NR;
2560 	if (PageDoubleMap(page))
2561 		ret -= HPAGE_PMD_NR;
2562 	return ret;
2563 }
2564 
2565 /*
2566  * This calculates accurately how many mappings a transparent hugepage
2567  * has (unlike page_mapcount() which isn't fully accurate). This full
2568  * accuracy is primarily needed to know if copy-on-write faults can
2569  * reuse the page and change the mapping to read-write instead of
2570  * copying them. At the same time this returns the total_mapcount too.
2571  *
2572  * The function returns the highest mapcount any one of the subpages
2573  * has. If the return value is one, even if different processes are
2574  * mapping different subpages of the transparent hugepage, they can
2575  * all reuse it, because each process is reusing a different subpage.
2576  *
2577  * The total_mapcount is instead counting all virtual mappings of the
2578  * subpages. If the total_mapcount is equal to "one", it tells the
2579  * caller all mappings belong to the same "mm" and in turn the
2580  * anon_vma of the transparent hugepage can become the vma->anon_vma
2581  * local one as no other process may be mapping any of the subpages.
2582  *
2583  * It would be more accurate to replace page_mapcount() with
2584  * page_trans_huge_mapcount(), however we only use
2585  * page_trans_huge_mapcount() in the copy-on-write faults where we
2586  * need full accuracy to avoid breaking page pinning, because
2587  * page_trans_huge_mapcount() is slower than page_mapcount().
2588  */
2589 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2590 {
2591 	int i, ret, _total_mapcount, mapcount;
2592 
2593 	/* hugetlbfs shouldn't call it */
2594 	VM_BUG_ON_PAGE(PageHuge(page), page);
2595 
2596 	if (likely(!PageTransCompound(page))) {
2597 		mapcount = atomic_read(&page->_mapcount) + 1;
2598 		if (total_mapcount)
2599 			*total_mapcount = mapcount;
2600 		return mapcount;
2601 	}
2602 
2603 	page = compound_head(page);
2604 
2605 	_total_mapcount = ret = 0;
2606 	for (i = 0; i < HPAGE_PMD_NR; i++) {
2607 		mapcount = atomic_read(&page[i]._mapcount) + 1;
2608 		ret = max(ret, mapcount);
2609 		_total_mapcount += mapcount;
2610 	}
2611 	if (PageDoubleMap(page)) {
2612 		ret -= 1;
2613 		_total_mapcount -= HPAGE_PMD_NR;
2614 	}
2615 	mapcount = compound_mapcount(page);
2616 	ret += mapcount;
2617 	_total_mapcount += mapcount;
2618 	if (total_mapcount)
2619 		*total_mapcount = _total_mapcount;
2620 	return ret;
2621 }
2622 
2623 /* Racy check whether the huge page can be split */
2624 bool can_split_huge_page(struct page *page, int *pextra_pins)
2625 {
2626 	int extra_pins;
2627 
2628 	/* Additional pins from page cache */
2629 	if (PageAnon(page))
2630 		extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2631 	else
2632 		extra_pins = HPAGE_PMD_NR;
2633 	if (pextra_pins)
2634 		*pextra_pins = extra_pins;
2635 	return total_mapcount(page) == page_count(page) - extra_pins - 1;
2636 }
2637 
2638 /*
2639  * This function splits huge page into normal pages. @page can point to any
2640  * subpage of huge page to split. Split doesn't change the position of @page.
2641  *
2642  * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2643  * The huge page must be locked.
2644  *
2645  * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2646  *
2647  * Both head page and tail pages will inherit mapping, flags, and so on from
2648  * the hugepage.
2649  *
2650  * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2651  * they are not mapped.
2652  *
2653  * Returns 0 if the hugepage is split successfully.
2654  * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2655  * us.
2656  */
2657 int split_huge_page_to_list(struct page *page, struct list_head *list)
2658 {
2659 	struct page *head = compound_head(page);
2660 	struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2661 	struct anon_vma *anon_vma = NULL;
2662 	struct address_space *mapping = NULL;
2663 	int count, mapcount, extra_pins, ret;
2664 	bool mlocked;
2665 	unsigned long flags;
2666 	pgoff_t end;
2667 
2668 	VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2669 	VM_BUG_ON_PAGE(!PageLocked(page), page);
2670 	VM_BUG_ON_PAGE(!PageCompound(page), page);
2671 
2672 	if (PageWriteback(page))
2673 		return -EBUSY;
2674 
2675 	if (PageAnon(head)) {
2676 		/*
2677 		 * The caller does not necessarily hold an mmap_sem that would
2678 		 * prevent the anon_vma disappearing so we first we take a
2679 		 * reference to it and then lock the anon_vma for write. This
2680 		 * is similar to page_lock_anon_vma_read except the write lock
2681 		 * is taken to serialise against parallel split or collapse
2682 		 * operations.
2683 		 */
2684 		anon_vma = page_get_anon_vma(head);
2685 		if (!anon_vma) {
2686 			ret = -EBUSY;
2687 			goto out;
2688 		}
2689 		end = -1;
2690 		mapping = NULL;
2691 		anon_vma_lock_write(anon_vma);
2692 	} else {
2693 		mapping = head->mapping;
2694 
2695 		/* Truncated ? */
2696 		if (!mapping) {
2697 			ret = -EBUSY;
2698 			goto out;
2699 		}
2700 
2701 		anon_vma = NULL;
2702 		i_mmap_lock_read(mapping);
2703 
2704 		/*
2705 		 *__split_huge_page() may need to trim off pages beyond EOF:
2706 		 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2707 		 * which cannot be nested inside the page tree lock. So note
2708 		 * end now: i_size itself may be changed at any moment, but
2709 		 * head page lock is good enough to serialize the trimming.
2710 		 */
2711 		end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2712 	}
2713 
2714 	/*
2715 	 * Racy check if we can split the page, before unmap_page() will
2716 	 * split PMDs
2717 	 */
2718 	if (!can_split_huge_page(head, &extra_pins)) {
2719 		ret = -EBUSY;
2720 		goto out_unlock;
2721 	}
2722 
2723 	mlocked = PageMlocked(page);
2724 	unmap_page(head);
2725 	VM_BUG_ON_PAGE(compound_mapcount(head), head);
2726 
2727 	/* Make sure the page is not on per-CPU pagevec as it takes pin */
2728 	if (mlocked)
2729 		lru_add_drain();
2730 
2731 	/* prevent PageLRU to go away from under us, and freeze lru stats */
2732 	spin_lock_irqsave(&pgdata->lru_lock, flags);
2733 
2734 	if (mapping) {
2735 		XA_STATE(xas, &mapping->i_pages, page_index(head));
2736 
2737 		/*
2738 		 * Check if the head page is present in page cache.
2739 		 * We assume all tail are present too, if head is there.
2740 		 */
2741 		xa_lock(&mapping->i_pages);
2742 		if (xas_load(&xas) != head)
2743 			goto fail;
2744 	}
2745 
2746 	/* Prevent deferred_split_scan() touching ->_refcount */
2747 	spin_lock(&pgdata->split_queue_lock);
2748 	count = page_count(head);
2749 	mapcount = total_mapcount(head);
2750 	if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2751 		if (!list_empty(page_deferred_list(head))) {
2752 			pgdata->split_queue_len--;
2753 			list_del(page_deferred_list(head));
2754 		}
2755 		if (mapping)
2756 			__dec_node_page_state(page, NR_SHMEM_THPS);
2757 		spin_unlock(&pgdata->split_queue_lock);
2758 		__split_huge_page(page, list, end, flags);
2759 		if (PageSwapCache(head)) {
2760 			swp_entry_t entry = { .val = page_private(head) };
2761 
2762 			ret = split_swap_cluster(entry);
2763 		} else
2764 			ret = 0;
2765 	} else {
2766 		if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2767 			pr_alert("total_mapcount: %u, page_count(): %u\n",
2768 					mapcount, count);
2769 			if (PageTail(page))
2770 				dump_page(head, NULL);
2771 			dump_page(page, "total_mapcount(head) > 0");
2772 			BUG();
2773 		}
2774 		spin_unlock(&pgdata->split_queue_lock);
2775 fail:		if (mapping)
2776 			xa_unlock(&mapping->i_pages);
2777 		spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2778 		remap_page(head);
2779 		ret = -EBUSY;
2780 	}
2781 
2782 out_unlock:
2783 	if (anon_vma) {
2784 		anon_vma_unlock_write(anon_vma);
2785 		put_anon_vma(anon_vma);
2786 	}
2787 	if (mapping)
2788 		i_mmap_unlock_read(mapping);
2789 out:
2790 	count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2791 	return ret;
2792 }
2793 
2794 void free_transhuge_page(struct page *page)
2795 {
2796 	struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2797 	unsigned long flags;
2798 
2799 	spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2800 	if (!list_empty(page_deferred_list(page))) {
2801 		pgdata->split_queue_len--;
2802 		list_del(page_deferred_list(page));
2803 	}
2804 	spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2805 	free_compound_page(page);
2806 }
2807 
2808 void deferred_split_huge_page(struct page *page)
2809 {
2810 	struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2811 	unsigned long flags;
2812 
2813 	VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2814 
2815 	spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2816 	if (list_empty(page_deferred_list(page))) {
2817 		count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2818 		list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2819 		pgdata->split_queue_len++;
2820 	}
2821 	spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2822 }
2823 
2824 static unsigned long deferred_split_count(struct shrinker *shrink,
2825 		struct shrink_control *sc)
2826 {
2827 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2828 	return READ_ONCE(pgdata->split_queue_len);
2829 }
2830 
2831 static unsigned long deferred_split_scan(struct shrinker *shrink,
2832 		struct shrink_control *sc)
2833 {
2834 	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2835 	unsigned long flags;
2836 	LIST_HEAD(list), *pos, *next;
2837 	struct page *page;
2838 	int split = 0;
2839 
2840 	spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2841 	/* Take pin on all head pages to avoid freeing them under us */
2842 	list_for_each_safe(pos, next, &pgdata->split_queue) {
2843 		page = list_entry((void *)pos, struct page, mapping);
2844 		page = compound_head(page);
2845 		if (get_page_unless_zero(page)) {
2846 			list_move(page_deferred_list(page), &list);
2847 		} else {
2848 			/* We lost race with put_compound_page() */
2849 			list_del_init(page_deferred_list(page));
2850 			pgdata->split_queue_len--;
2851 		}
2852 		if (!--sc->nr_to_scan)
2853 			break;
2854 	}
2855 	spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2856 
2857 	list_for_each_safe(pos, next, &list) {
2858 		page = list_entry((void *)pos, struct page, mapping);
2859 		if (!trylock_page(page))
2860 			goto next;
2861 		/* split_huge_page() removes page from list on success */
2862 		if (!split_huge_page(page))
2863 			split++;
2864 		unlock_page(page);
2865 next:
2866 		put_page(page);
2867 	}
2868 
2869 	spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2870 	list_splice_tail(&list, &pgdata->split_queue);
2871 	spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2872 
2873 	/*
2874 	 * Stop shrinker if we didn't split any page, but the queue is empty.
2875 	 * This can happen if pages were freed under us.
2876 	 */
2877 	if (!split && list_empty(&pgdata->split_queue))
2878 		return SHRINK_STOP;
2879 	return split;
2880 }
2881 
2882 static struct shrinker deferred_split_shrinker = {
2883 	.count_objects = deferred_split_count,
2884 	.scan_objects = deferred_split_scan,
2885 	.seeks = DEFAULT_SEEKS,
2886 	.flags = SHRINKER_NUMA_AWARE,
2887 };
2888 
2889 #ifdef CONFIG_DEBUG_FS
2890 static int split_huge_pages_set(void *data, u64 val)
2891 {
2892 	struct zone *zone;
2893 	struct page *page;
2894 	unsigned long pfn, max_zone_pfn;
2895 	unsigned long total = 0, split = 0;
2896 
2897 	if (val != 1)
2898 		return -EINVAL;
2899 
2900 	for_each_populated_zone(zone) {
2901 		max_zone_pfn = zone_end_pfn(zone);
2902 		for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2903 			if (!pfn_valid(pfn))
2904 				continue;
2905 
2906 			page = pfn_to_page(pfn);
2907 			if (!get_page_unless_zero(page))
2908 				continue;
2909 
2910 			if (zone != page_zone(page))
2911 				goto next;
2912 
2913 			if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2914 				goto next;
2915 
2916 			total++;
2917 			lock_page(page);
2918 			if (!split_huge_page(page))
2919 				split++;
2920 			unlock_page(page);
2921 next:
2922 			put_page(page);
2923 		}
2924 	}
2925 
2926 	pr_info("%lu of %lu THP split\n", split, total);
2927 
2928 	return 0;
2929 }
2930 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2931 		"%llu\n");
2932 
2933 static int __init split_huge_pages_debugfs(void)
2934 {
2935 	debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2936 			    &split_huge_pages_fops);
2937 	return 0;
2938 }
2939 late_initcall(split_huge_pages_debugfs);
2940 #endif
2941 
2942 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2943 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2944 		struct page *page)
2945 {
2946 	struct vm_area_struct *vma = pvmw->vma;
2947 	struct mm_struct *mm = vma->vm_mm;
2948 	unsigned long address = pvmw->address;
2949 	pmd_t pmdval;
2950 	swp_entry_t entry;
2951 	pmd_t pmdswp;
2952 
2953 	if (!(pvmw->pmd && !pvmw->pte))
2954 		return;
2955 
2956 	flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2957 	pmdval = *pvmw->pmd;
2958 	pmdp_invalidate(vma, address, pvmw->pmd);
2959 	if (pmd_dirty(pmdval))
2960 		set_page_dirty(page);
2961 	entry = make_migration_entry(page, pmd_write(pmdval));
2962 	pmdswp = swp_entry_to_pmd(entry);
2963 	if (pmd_soft_dirty(pmdval))
2964 		pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2965 	set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2966 	page_remove_rmap(page, true);
2967 	put_page(page);
2968 }
2969 
2970 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2971 {
2972 	struct vm_area_struct *vma = pvmw->vma;
2973 	struct mm_struct *mm = vma->vm_mm;
2974 	unsigned long address = pvmw->address;
2975 	unsigned long mmun_start = address & HPAGE_PMD_MASK;
2976 	pmd_t pmde;
2977 	swp_entry_t entry;
2978 
2979 	if (!(pvmw->pmd && !pvmw->pte))
2980 		return;
2981 
2982 	entry = pmd_to_swp_entry(*pvmw->pmd);
2983 	get_page(new);
2984 	pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2985 	if (pmd_swp_soft_dirty(*pvmw->pmd))
2986 		pmde = pmd_mksoft_dirty(pmde);
2987 	if (is_write_migration_entry(entry))
2988 		pmde = maybe_pmd_mkwrite(pmde, vma);
2989 
2990 	flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2991 	if (PageAnon(new))
2992 		page_add_anon_rmap(new, vma, mmun_start, true);
2993 	else
2994 		page_add_file_rmap(new, true);
2995 	set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2996 	if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
2997 		mlock_vma_page(new);
2998 	update_mmu_cache_pmd(vma, address, pvmw->pmd);
2999 }
3000 #endif
3001