xref: /openbmc/linux/mm/hugetlb.c (revision 80d0624d)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Generic hugetlb support.
4  * (C) Nadia Yvette Chambers, April 2004
5  */
6 #include <linux/list.h>
7 #include <linux/init.h>
8 #include <linux/mm.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/sched/mm.h>
23 #include <linux/mmdebug.h>
24 #include <linux/sched/signal.h>
25 #include <linux/rmap.h>
26 #include <linux/string_helpers.h>
27 #include <linux/swap.h>
28 #include <linux/swapops.h>
29 #include <linux/jhash.h>
30 #include <linux/numa.h>
31 #include <linux/llist.h>
32 #include <linux/cma.h>
33 #include <linux/migrate.h>
34 #include <linux/nospec.h>
35 #include <linux/delayacct.h>
36 #include <linux/memory.h>
37 #include <linux/mm_inline.h>
38 
39 #include <asm/page.h>
40 #include <asm/pgalloc.h>
41 #include <asm/tlb.h>
42 
43 #include <linux/io.h>
44 #include <linux/hugetlb.h>
45 #include <linux/hugetlb_cgroup.h>
46 #include <linux/node.h>
47 #include <linux/page_owner.h>
48 #include "internal.h"
49 #include "hugetlb_vmemmap.h"
50 
51 int hugetlb_max_hstate __read_mostly;
52 unsigned int default_hstate_idx;
53 struct hstate hstates[HUGE_MAX_HSTATE];
54 
55 #ifdef CONFIG_CMA
56 static struct cma *hugetlb_cma[MAX_NUMNODES];
57 static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata;
58 static bool hugetlb_cma_folio(struct folio *folio, unsigned int order)
59 {
60 	return cma_pages_valid(hugetlb_cma[folio_nid(folio)], &folio->page,
61 				1 << order);
62 }
63 #else
64 static bool hugetlb_cma_folio(struct folio *folio, unsigned int order)
65 {
66 	return false;
67 }
68 #endif
69 static unsigned long hugetlb_cma_size __initdata;
70 
71 __initdata LIST_HEAD(huge_boot_pages);
72 
73 /* for command line parsing */
74 static struct hstate * __initdata parsed_hstate;
75 static unsigned long __initdata default_hstate_max_huge_pages;
76 static bool __initdata parsed_valid_hugepagesz = true;
77 static bool __initdata parsed_default_hugepagesz;
78 static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata;
79 
80 /*
81  * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
82  * free_huge_pages, and surplus_huge_pages.
83  */
84 DEFINE_SPINLOCK(hugetlb_lock);
85 
86 /*
87  * Serializes faults on the same logical page.  This is used to
88  * prevent spurious OOMs when the hugepage pool is fully utilized.
89  */
90 static int num_fault_mutexes;
91 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
92 
93 /* Forward declaration */
94 static int hugetlb_acct_memory(struct hstate *h, long delta);
95 static void hugetlb_vma_lock_free(struct vm_area_struct *vma);
96 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma);
97 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma);
98 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
99 		unsigned long start, unsigned long end);
100 static struct resv_map *vma_resv_map(struct vm_area_struct *vma);
101 
102 static inline bool subpool_is_free(struct hugepage_subpool *spool)
103 {
104 	if (spool->count)
105 		return false;
106 	if (spool->max_hpages != -1)
107 		return spool->used_hpages == 0;
108 	if (spool->min_hpages != -1)
109 		return spool->rsv_hpages == spool->min_hpages;
110 
111 	return true;
112 }
113 
114 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
115 						unsigned long irq_flags)
116 {
117 	spin_unlock_irqrestore(&spool->lock, irq_flags);
118 
119 	/* If no pages are used, and no other handles to the subpool
120 	 * remain, give up any reservations based on minimum size and
121 	 * free the subpool */
122 	if (subpool_is_free(spool)) {
123 		if (spool->min_hpages != -1)
124 			hugetlb_acct_memory(spool->hstate,
125 						-spool->min_hpages);
126 		kfree(spool);
127 	}
128 }
129 
130 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
131 						long min_hpages)
132 {
133 	struct hugepage_subpool *spool;
134 
135 	spool = kzalloc(sizeof(*spool), GFP_KERNEL);
136 	if (!spool)
137 		return NULL;
138 
139 	spin_lock_init(&spool->lock);
140 	spool->count = 1;
141 	spool->max_hpages = max_hpages;
142 	spool->hstate = h;
143 	spool->min_hpages = min_hpages;
144 
145 	if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
146 		kfree(spool);
147 		return NULL;
148 	}
149 	spool->rsv_hpages = min_hpages;
150 
151 	return spool;
152 }
153 
154 void hugepage_put_subpool(struct hugepage_subpool *spool)
155 {
156 	unsigned long flags;
157 
158 	spin_lock_irqsave(&spool->lock, flags);
159 	BUG_ON(!spool->count);
160 	spool->count--;
161 	unlock_or_release_subpool(spool, flags);
162 }
163 
164 /*
165  * Subpool accounting for allocating and reserving pages.
166  * Return -ENOMEM if there are not enough resources to satisfy the
167  * request.  Otherwise, return the number of pages by which the
168  * global pools must be adjusted (upward).  The returned value may
169  * only be different than the passed value (delta) in the case where
170  * a subpool minimum size must be maintained.
171  */
172 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
173 				      long delta)
174 {
175 	long ret = delta;
176 
177 	if (!spool)
178 		return ret;
179 
180 	spin_lock_irq(&spool->lock);
181 
182 	if (spool->max_hpages != -1) {		/* maximum size accounting */
183 		if ((spool->used_hpages + delta) <= spool->max_hpages)
184 			spool->used_hpages += delta;
185 		else {
186 			ret = -ENOMEM;
187 			goto unlock_ret;
188 		}
189 	}
190 
191 	/* minimum size accounting */
192 	if (spool->min_hpages != -1 && spool->rsv_hpages) {
193 		if (delta > spool->rsv_hpages) {
194 			/*
195 			 * Asking for more reserves than those already taken on
196 			 * behalf of subpool.  Return difference.
197 			 */
198 			ret = delta - spool->rsv_hpages;
199 			spool->rsv_hpages = 0;
200 		} else {
201 			ret = 0;	/* reserves already accounted for */
202 			spool->rsv_hpages -= delta;
203 		}
204 	}
205 
206 unlock_ret:
207 	spin_unlock_irq(&spool->lock);
208 	return ret;
209 }
210 
211 /*
212  * Subpool accounting for freeing and unreserving pages.
213  * Return the number of global page reservations that must be dropped.
214  * The return value may only be different than the passed value (delta)
215  * in the case where a subpool minimum size must be maintained.
216  */
217 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
218 				       long delta)
219 {
220 	long ret = delta;
221 	unsigned long flags;
222 
223 	if (!spool)
224 		return delta;
225 
226 	spin_lock_irqsave(&spool->lock, flags);
227 
228 	if (spool->max_hpages != -1)		/* maximum size accounting */
229 		spool->used_hpages -= delta;
230 
231 	 /* minimum size accounting */
232 	if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
233 		if (spool->rsv_hpages + delta <= spool->min_hpages)
234 			ret = 0;
235 		else
236 			ret = spool->rsv_hpages + delta - spool->min_hpages;
237 
238 		spool->rsv_hpages += delta;
239 		if (spool->rsv_hpages > spool->min_hpages)
240 			spool->rsv_hpages = spool->min_hpages;
241 	}
242 
243 	/*
244 	 * If hugetlbfs_put_super couldn't free spool due to an outstanding
245 	 * quota reference, free it now.
246 	 */
247 	unlock_or_release_subpool(spool, flags);
248 
249 	return ret;
250 }
251 
252 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
253 {
254 	return HUGETLBFS_SB(inode->i_sb)->spool;
255 }
256 
257 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
258 {
259 	return subpool_inode(file_inode(vma->vm_file));
260 }
261 
262 /*
263  * hugetlb vma_lock helper routines
264  */
265 void hugetlb_vma_lock_read(struct vm_area_struct *vma)
266 {
267 	if (__vma_shareable_lock(vma)) {
268 		struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
269 
270 		down_read(&vma_lock->rw_sema);
271 	} else if (__vma_private_lock(vma)) {
272 		struct resv_map *resv_map = vma_resv_map(vma);
273 
274 		down_read(&resv_map->rw_sema);
275 	}
276 }
277 
278 void hugetlb_vma_unlock_read(struct vm_area_struct *vma)
279 {
280 	if (__vma_shareable_lock(vma)) {
281 		struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
282 
283 		up_read(&vma_lock->rw_sema);
284 	} else if (__vma_private_lock(vma)) {
285 		struct resv_map *resv_map = vma_resv_map(vma);
286 
287 		up_read(&resv_map->rw_sema);
288 	}
289 }
290 
291 void hugetlb_vma_lock_write(struct vm_area_struct *vma)
292 {
293 	if (__vma_shareable_lock(vma)) {
294 		struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
295 
296 		down_write(&vma_lock->rw_sema);
297 	} else if (__vma_private_lock(vma)) {
298 		struct resv_map *resv_map = vma_resv_map(vma);
299 
300 		down_write(&resv_map->rw_sema);
301 	}
302 }
303 
304 void hugetlb_vma_unlock_write(struct vm_area_struct *vma)
305 {
306 	if (__vma_shareable_lock(vma)) {
307 		struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
308 
309 		up_write(&vma_lock->rw_sema);
310 	} else if (__vma_private_lock(vma)) {
311 		struct resv_map *resv_map = vma_resv_map(vma);
312 
313 		up_write(&resv_map->rw_sema);
314 	}
315 }
316 
317 int hugetlb_vma_trylock_write(struct vm_area_struct *vma)
318 {
319 
320 	if (__vma_shareable_lock(vma)) {
321 		struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
322 
323 		return down_write_trylock(&vma_lock->rw_sema);
324 	} else if (__vma_private_lock(vma)) {
325 		struct resv_map *resv_map = vma_resv_map(vma);
326 
327 		return down_write_trylock(&resv_map->rw_sema);
328 	}
329 
330 	return 1;
331 }
332 
333 void hugetlb_vma_assert_locked(struct vm_area_struct *vma)
334 {
335 	if (__vma_shareable_lock(vma)) {
336 		struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
337 
338 		lockdep_assert_held(&vma_lock->rw_sema);
339 	} else if (__vma_private_lock(vma)) {
340 		struct resv_map *resv_map = vma_resv_map(vma);
341 
342 		lockdep_assert_held(&resv_map->rw_sema);
343 	}
344 }
345 
346 void hugetlb_vma_lock_release(struct kref *kref)
347 {
348 	struct hugetlb_vma_lock *vma_lock = container_of(kref,
349 			struct hugetlb_vma_lock, refs);
350 
351 	kfree(vma_lock);
352 }
353 
354 static void __hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock *vma_lock)
355 {
356 	struct vm_area_struct *vma = vma_lock->vma;
357 
358 	/*
359 	 * vma_lock structure may or not be released as a result of put,
360 	 * it certainly will no longer be attached to vma so clear pointer.
361 	 * Semaphore synchronizes access to vma_lock->vma field.
362 	 */
363 	vma_lock->vma = NULL;
364 	vma->vm_private_data = NULL;
365 	up_write(&vma_lock->rw_sema);
366 	kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
367 }
368 
369 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma)
370 {
371 	if (__vma_shareable_lock(vma)) {
372 		struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
373 
374 		__hugetlb_vma_unlock_write_put(vma_lock);
375 	} else if (__vma_private_lock(vma)) {
376 		struct resv_map *resv_map = vma_resv_map(vma);
377 
378 		/* no free for anon vmas, but still need to unlock */
379 		up_write(&resv_map->rw_sema);
380 	}
381 }
382 
383 static void hugetlb_vma_lock_free(struct vm_area_struct *vma)
384 {
385 	/*
386 	 * Only present in sharable vmas.
387 	 */
388 	if (!vma || !__vma_shareable_lock(vma))
389 		return;
390 
391 	if (vma->vm_private_data) {
392 		struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
393 
394 		down_write(&vma_lock->rw_sema);
395 		__hugetlb_vma_unlock_write_put(vma_lock);
396 	}
397 }
398 
399 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma)
400 {
401 	struct hugetlb_vma_lock *vma_lock;
402 
403 	/* Only establish in (flags) sharable vmas */
404 	if (!vma || !(vma->vm_flags & VM_MAYSHARE))
405 		return;
406 
407 	/* Should never get here with non-NULL vm_private_data */
408 	if (vma->vm_private_data)
409 		return;
410 
411 	vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL);
412 	if (!vma_lock) {
413 		/*
414 		 * If we can not allocate structure, then vma can not
415 		 * participate in pmd sharing.  This is only a possible
416 		 * performance enhancement and memory saving issue.
417 		 * However, the lock is also used to synchronize page
418 		 * faults with truncation.  If the lock is not present,
419 		 * unlikely races could leave pages in a file past i_size
420 		 * until the file is removed.  Warn in the unlikely case of
421 		 * allocation failure.
422 		 */
423 		pr_warn_once("HugeTLB: unable to allocate vma specific lock\n");
424 		return;
425 	}
426 
427 	kref_init(&vma_lock->refs);
428 	init_rwsem(&vma_lock->rw_sema);
429 	vma_lock->vma = vma;
430 	vma->vm_private_data = vma_lock;
431 }
432 
433 /* Helper that removes a struct file_region from the resv_map cache and returns
434  * it for use.
435  */
436 static struct file_region *
437 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
438 {
439 	struct file_region *nrg;
440 
441 	VM_BUG_ON(resv->region_cache_count <= 0);
442 
443 	resv->region_cache_count--;
444 	nrg = list_first_entry(&resv->region_cache, struct file_region, link);
445 	list_del(&nrg->link);
446 
447 	nrg->from = from;
448 	nrg->to = to;
449 
450 	return nrg;
451 }
452 
453 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
454 					      struct file_region *rg)
455 {
456 #ifdef CONFIG_CGROUP_HUGETLB
457 	nrg->reservation_counter = rg->reservation_counter;
458 	nrg->css = rg->css;
459 	if (rg->css)
460 		css_get(rg->css);
461 #endif
462 }
463 
464 /* Helper that records hugetlb_cgroup uncharge info. */
465 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
466 						struct hstate *h,
467 						struct resv_map *resv,
468 						struct file_region *nrg)
469 {
470 #ifdef CONFIG_CGROUP_HUGETLB
471 	if (h_cg) {
472 		nrg->reservation_counter =
473 			&h_cg->rsvd_hugepage[hstate_index(h)];
474 		nrg->css = &h_cg->css;
475 		/*
476 		 * The caller will hold exactly one h_cg->css reference for the
477 		 * whole contiguous reservation region. But this area might be
478 		 * scattered when there are already some file_regions reside in
479 		 * it. As a result, many file_regions may share only one css
480 		 * reference. In order to ensure that one file_region must hold
481 		 * exactly one h_cg->css reference, we should do css_get for
482 		 * each file_region and leave the reference held by caller
483 		 * untouched.
484 		 */
485 		css_get(&h_cg->css);
486 		if (!resv->pages_per_hpage)
487 			resv->pages_per_hpage = pages_per_huge_page(h);
488 		/* pages_per_hpage should be the same for all entries in
489 		 * a resv_map.
490 		 */
491 		VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
492 	} else {
493 		nrg->reservation_counter = NULL;
494 		nrg->css = NULL;
495 	}
496 #endif
497 }
498 
499 static void put_uncharge_info(struct file_region *rg)
500 {
501 #ifdef CONFIG_CGROUP_HUGETLB
502 	if (rg->css)
503 		css_put(rg->css);
504 #endif
505 }
506 
507 static bool has_same_uncharge_info(struct file_region *rg,
508 				   struct file_region *org)
509 {
510 #ifdef CONFIG_CGROUP_HUGETLB
511 	return rg->reservation_counter == org->reservation_counter &&
512 	       rg->css == org->css;
513 
514 #else
515 	return true;
516 #endif
517 }
518 
519 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
520 {
521 	struct file_region *nrg, *prg;
522 
523 	prg = list_prev_entry(rg, link);
524 	if (&prg->link != &resv->regions && prg->to == rg->from &&
525 	    has_same_uncharge_info(prg, rg)) {
526 		prg->to = rg->to;
527 
528 		list_del(&rg->link);
529 		put_uncharge_info(rg);
530 		kfree(rg);
531 
532 		rg = prg;
533 	}
534 
535 	nrg = list_next_entry(rg, link);
536 	if (&nrg->link != &resv->regions && nrg->from == rg->to &&
537 	    has_same_uncharge_info(nrg, rg)) {
538 		nrg->from = rg->from;
539 
540 		list_del(&rg->link);
541 		put_uncharge_info(rg);
542 		kfree(rg);
543 	}
544 }
545 
546 static inline long
547 hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from,
548 		     long to, struct hstate *h, struct hugetlb_cgroup *cg,
549 		     long *regions_needed)
550 {
551 	struct file_region *nrg;
552 
553 	if (!regions_needed) {
554 		nrg = get_file_region_entry_from_cache(map, from, to);
555 		record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
556 		list_add(&nrg->link, rg);
557 		coalesce_file_region(map, nrg);
558 	} else
559 		*regions_needed += 1;
560 
561 	return to - from;
562 }
563 
564 /*
565  * Must be called with resv->lock held.
566  *
567  * Calling this with regions_needed != NULL will count the number of pages
568  * to be added but will not modify the linked list. And regions_needed will
569  * indicate the number of file_regions needed in the cache to carry out to add
570  * the regions for this range.
571  */
572 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
573 				     struct hugetlb_cgroup *h_cg,
574 				     struct hstate *h, long *regions_needed)
575 {
576 	long add = 0;
577 	struct list_head *head = &resv->regions;
578 	long last_accounted_offset = f;
579 	struct file_region *iter, *trg = NULL;
580 	struct list_head *rg = NULL;
581 
582 	if (regions_needed)
583 		*regions_needed = 0;
584 
585 	/* In this loop, we essentially handle an entry for the range
586 	 * [last_accounted_offset, iter->from), at every iteration, with some
587 	 * bounds checking.
588 	 */
589 	list_for_each_entry_safe(iter, trg, head, link) {
590 		/* Skip irrelevant regions that start before our range. */
591 		if (iter->from < f) {
592 			/* If this region ends after the last accounted offset,
593 			 * then we need to update last_accounted_offset.
594 			 */
595 			if (iter->to > last_accounted_offset)
596 				last_accounted_offset = iter->to;
597 			continue;
598 		}
599 
600 		/* When we find a region that starts beyond our range, we've
601 		 * finished.
602 		 */
603 		if (iter->from >= t) {
604 			rg = iter->link.prev;
605 			break;
606 		}
607 
608 		/* Add an entry for last_accounted_offset -> iter->from, and
609 		 * update last_accounted_offset.
610 		 */
611 		if (iter->from > last_accounted_offset)
612 			add += hugetlb_resv_map_add(resv, iter->link.prev,
613 						    last_accounted_offset,
614 						    iter->from, h, h_cg,
615 						    regions_needed);
616 
617 		last_accounted_offset = iter->to;
618 	}
619 
620 	/* Handle the case where our range extends beyond
621 	 * last_accounted_offset.
622 	 */
623 	if (!rg)
624 		rg = head->prev;
625 	if (last_accounted_offset < t)
626 		add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
627 					    t, h, h_cg, regions_needed);
628 
629 	return add;
630 }
631 
632 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
633  */
634 static int allocate_file_region_entries(struct resv_map *resv,
635 					int regions_needed)
636 	__must_hold(&resv->lock)
637 {
638 	LIST_HEAD(allocated_regions);
639 	int to_allocate = 0, i = 0;
640 	struct file_region *trg = NULL, *rg = NULL;
641 
642 	VM_BUG_ON(regions_needed < 0);
643 
644 	/*
645 	 * Check for sufficient descriptors in the cache to accommodate
646 	 * the number of in progress add operations plus regions_needed.
647 	 *
648 	 * This is a while loop because when we drop the lock, some other call
649 	 * to region_add or region_del may have consumed some region_entries,
650 	 * so we keep looping here until we finally have enough entries for
651 	 * (adds_in_progress + regions_needed).
652 	 */
653 	while (resv->region_cache_count <
654 	       (resv->adds_in_progress + regions_needed)) {
655 		to_allocate = resv->adds_in_progress + regions_needed -
656 			      resv->region_cache_count;
657 
658 		/* At this point, we should have enough entries in the cache
659 		 * for all the existing adds_in_progress. We should only be
660 		 * needing to allocate for regions_needed.
661 		 */
662 		VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
663 
664 		spin_unlock(&resv->lock);
665 		for (i = 0; i < to_allocate; i++) {
666 			trg = kmalloc(sizeof(*trg), GFP_KERNEL);
667 			if (!trg)
668 				goto out_of_memory;
669 			list_add(&trg->link, &allocated_regions);
670 		}
671 
672 		spin_lock(&resv->lock);
673 
674 		list_splice(&allocated_regions, &resv->region_cache);
675 		resv->region_cache_count += to_allocate;
676 	}
677 
678 	return 0;
679 
680 out_of_memory:
681 	list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
682 		list_del(&rg->link);
683 		kfree(rg);
684 	}
685 	return -ENOMEM;
686 }
687 
688 /*
689  * Add the huge page range represented by [f, t) to the reserve
690  * map.  Regions will be taken from the cache to fill in this range.
691  * Sufficient regions should exist in the cache due to the previous
692  * call to region_chg with the same range, but in some cases the cache will not
693  * have sufficient entries due to races with other code doing region_add or
694  * region_del.  The extra needed entries will be allocated.
695  *
696  * regions_needed is the out value provided by a previous call to region_chg.
697  *
698  * Return the number of new huge pages added to the map.  This number is greater
699  * than or equal to zero.  If file_region entries needed to be allocated for
700  * this operation and we were not able to allocate, it returns -ENOMEM.
701  * region_add of regions of length 1 never allocate file_regions and cannot
702  * fail; region_chg will always allocate at least 1 entry and a region_add for
703  * 1 page will only require at most 1 entry.
704  */
705 static long region_add(struct resv_map *resv, long f, long t,
706 		       long in_regions_needed, struct hstate *h,
707 		       struct hugetlb_cgroup *h_cg)
708 {
709 	long add = 0, actual_regions_needed = 0;
710 
711 	spin_lock(&resv->lock);
712 retry:
713 
714 	/* Count how many regions are actually needed to execute this add. */
715 	add_reservation_in_range(resv, f, t, NULL, NULL,
716 				 &actual_regions_needed);
717 
718 	/*
719 	 * Check for sufficient descriptors in the cache to accommodate
720 	 * this add operation. Note that actual_regions_needed may be greater
721 	 * than in_regions_needed, as the resv_map may have been modified since
722 	 * the region_chg call. In this case, we need to make sure that we
723 	 * allocate extra entries, such that we have enough for all the
724 	 * existing adds_in_progress, plus the excess needed for this
725 	 * operation.
726 	 */
727 	if (actual_regions_needed > in_regions_needed &&
728 	    resv->region_cache_count <
729 		    resv->adds_in_progress +
730 			    (actual_regions_needed - in_regions_needed)) {
731 		/* region_add operation of range 1 should never need to
732 		 * allocate file_region entries.
733 		 */
734 		VM_BUG_ON(t - f <= 1);
735 
736 		if (allocate_file_region_entries(
737 			    resv, actual_regions_needed - in_regions_needed)) {
738 			return -ENOMEM;
739 		}
740 
741 		goto retry;
742 	}
743 
744 	add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
745 
746 	resv->adds_in_progress -= in_regions_needed;
747 
748 	spin_unlock(&resv->lock);
749 	return add;
750 }
751 
752 /*
753  * Examine the existing reserve map and determine how many
754  * huge pages in the specified range [f, t) are NOT currently
755  * represented.  This routine is called before a subsequent
756  * call to region_add that will actually modify the reserve
757  * map to add the specified range [f, t).  region_chg does
758  * not change the number of huge pages represented by the
759  * map.  A number of new file_region structures is added to the cache as a
760  * placeholder, for the subsequent region_add call to use. At least 1
761  * file_region structure is added.
762  *
763  * out_regions_needed is the number of regions added to the
764  * resv->adds_in_progress.  This value needs to be provided to a follow up call
765  * to region_add or region_abort for proper accounting.
766  *
767  * Returns the number of huge pages that need to be added to the existing
768  * reservation map for the range [f, t).  This number is greater or equal to
769  * zero.  -ENOMEM is returned if a new file_region structure or cache entry
770  * is needed and can not be allocated.
771  */
772 static long region_chg(struct resv_map *resv, long f, long t,
773 		       long *out_regions_needed)
774 {
775 	long chg = 0;
776 
777 	spin_lock(&resv->lock);
778 
779 	/* Count how many hugepages in this range are NOT represented. */
780 	chg = add_reservation_in_range(resv, f, t, NULL, NULL,
781 				       out_regions_needed);
782 
783 	if (*out_regions_needed == 0)
784 		*out_regions_needed = 1;
785 
786 	if (allocate_file_region_entries(resv, *out_regions_needed))
787 		return -ENOMEM;
788 
789 	resv->adds_in_progress += *out_regions_needed;
790 
791 	spin_unlock(&resv->lock);
792 	return chg;
793 }
794 
795 /*
796  * Abort the in progress add operation.  The adds_in_progress field
797  * of the resv_map keeps track of the operations in progress between
798  * calls to region_chg and region_add.  Operations are sometimes
799  * aborted after the call to region_chg.  In such cases, region_abort
800  * is called to decrement the adds_in_progress counter. regions_needed
801  * is the value returned by the region_chg call, it is used to decrement
802  * the adds_in_progress counter.
803  *
804  * NOTE: The range arguments [f, t) are not needed or used in this
805  * routine.  They are kept to make reading the calling code easier as
806  * arguments will match the associated region_chg call.
807  */
808 static void region_abort(struct resv_map *resv, long f, long t,
809 			 long regions_needed)
810 {
811 	spin_lock(&resv->lock);
812 	VM_BUG_ON(!resv->region_cache_count);
813 	resv->adds_in_progress -= regions_needed;
814 	spin_unlock(&resv->lock);
815 }
816 
817 /*
818  * Delete the specified range [f, t) from the reserve map.  If the
819  * t parameter is LONG_MAX, this indicates that ALL regions after f
820  * should be deleted.  Locate the regions which intersect [f, t)
821  * and either trim, delete or split the existing regions.
822  *
823  * Returns the number of huge pages deleted from the reserve map.
824  * In the normal case, the return value is zero or more.  In the
825  * case where a region must be split, a new region descriptor must
826  * be allocated.  If the allocation fails, -ENOMEM will be returned.
827  * NOTE: If the parameter t == LONG_MAX, then we will never split
828  * a region and possibly return -ENOMEM.  Callers specifying
829  * t == LONG_MAX do not need to check for -ENOMEM error.
830  */
831 static long region_del(struct resv_map *resv, long f, long t)
832 {
833 	struct list_head *head = &resv->regions;
834 	struct file_region *rg, *trg;
835 	struct file_region *nrg = NULL;
836 	long del = 0;
837 
838 retry:
839 	spin_lock(&resv->lock);
840 	list_for_each_entry_safe(rg, trg, head, link) {
841 		/*
842 		 * Skip regions before the range to be deleted.  file_region
843 		 * ranges are normally of the form [from, to).  However, there
844 		 * may be a "placeholder" entry in the map which is of the form
845 		 * (from, to) with from == to.  Check for placeholder entries
846 		 * at the beginning of the range to be deleted.
847 		 */
848 		if (rg->to <= f && (rg->to != rg->from || rg->to != f))
849 			continue;
850 
851 		if (rg->from >= t)
852 			break;
853 
854 		if (f > rg->from && t < rg->to) { /* Must split region */
855 			/*
856 			 * Check for an entry in the cache before dropping
857 			 * lock and attempting allocation.
858 			 */
859 			if (!nrg &&
860 			    resv->region_cache_count > resv->adds_in_progress) {
861 				nrg = list_first_entry(&resv->region_cache,
862 							struct file_region,
863 							link);
864 				list_del(&nrg->link);
865 				resv->region_cache_count--;
866 			}
867 
868 			if (!nrg) {
869 				spin_unlock(&resv->lock);
870 				nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
871 				if (!nrg)
872 					return -ENOMEM;
873 				goto retry;
874 			}
875 
876 			del += t - f;
877 			hugetlb_cgroup_uncharge_file_region(
878 				resv, rg, t - f, false);
879 
880 			/* New entry for end of split region */
881 			nrg->from = t;
882 			nrg->to = rg->to;
883 
884 			copy_hugetlb_cgroup_uncharge_info(nrg, rg);
885 
886 			INIT_LIST_HEAD(&nrg->link);
887 
888 			/* Original entry is trimmed */
889 			rg->to = f;
890 
891 			list_add(&nrg->link, &rg->link);
892 			nrg = NULL;
893 			break;
894 		}
895 
896 		if (f <= rg->from && t >= rg->to) { /* Remove entire region */
897 			del += rg->to - rg->from;
898 			hugetlb_cgroup_uncharge_file_region(resv, rg,
899 							    rg->to - rg->from, true);
900 			list_del(&rg->link);
901 			kfree(rg);
902 			continue;
903 		}
904 
905 		if (f <= rg->from) {	/* Trim beginning of region */
906 			hugetlb_cgroup_uncharge_file_region(resv, rg,
907 							    t - rg->from, false);
908 
909 			del += t - rg->from;
910 			rg->from = t;
911 		} else {		/* Trim end of region */
912 			hugetlb_cgroup_uncharge_file_region(resv, rg,
913 							    rg->to - f, false);
914 
915 			del += rg->to - f;
916 			rg->to = f;
917 		}
918 	}
919 
920 	spin_unlock(&resv->lock);
921 	kfree(nrg);
922 	return del;
923 }
924 
925 /*
926  * A rare out of memory error was encountered which prevented removal of
927  * the reserve map region for a page.  The huge page itself was free'ed
928  * and removed from the page cache.  This routine will adjust the subpool
929  * usage count, and the global reserve count if needed.  By incrementing
930  * these counts, the reserve map entry which could not be deleted will
931  * appear as a "reserved" entry instead of simply dangling with incorrect
932  * counts.
933  */
934 void hugetlb_fix_reserve_counts(struct inode *inode)
935 {
936 	struct hugepage_subpool *spool = subpool_inode(inode);
937 	long rsv_adjust;
938 	bool reserved = false;
939 
940 	rsv_adjust = hugepage_subpool_get_pages(spool, 1);
941 	if (rsv_adjust > 0) {
942 		struct hstate *h = hstate_inode(inode);
943 
944 		if (!hugetlb_acct_memory(h, 1))
945 			reserved = true;
946 	} else if (!rsv_adjust) {
947 		reserved = true;
948 	}
949 
950 	if (!reserved)
951 		pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
952 }
953 
954 /*
955  * Count and return the number of huge pages in the reserve map
956  * that intersect with the range [f, t).
957  */
958 static long region_count(struct resv_map *resv, long f, long t)
959 {
960 	struct list_head *head = &resv->regions;
961 	struct file_region *rg;
962 	long chg = 0;
963 
964 	spin_lock(&resv->lock);
965 	/* Locate each segment we overlap with, and count that overlap. */
966 	list_for_each_entry(rg, head, link) {
967 		long seg_from;
968 		long seg_to;
969 
970 		if (rg->to <= f)
971 			continue;
972 		if (rg->from >= t)
973 			break;
974 
975 		seg_from = max(rg->from, f);
976 		seg_to = min(rg->to, t);
977 
978 		chg += seg_to - seg_from;
979 	}
980 	spin_unlock(&resv->lock);
981 
982 	return chg;
983 }
984 
985 /*
986  * Convert the address within this vma to the page offset within
987  * the mapping, in pagecache page units; huge pages here.
988  */
989 static pgoff_t vma_hugecache_offset(struct hstate *h,
990 			struct vm_area_struct *vma, unsigned long address)
991 {
992 	return ((address - vma->vm_start) >> huge_page_shift(h)) +
993 			(vma->vm_pgoff >> huge_page_order(h));
994 }
995 
996 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
997 				     unsigned long address)
998 {
999 	return vma_hugecache_offset(hstate_vma(vma), vma, address);
1000 }
1001 EXPORT_SYMBOL_GPL(linear_hugepage_index);
1002 
1003 /**
1004  * vma_kernel_pagesize - Page size granularity for this VMA.
1005  * @vma: The user mapping.
1006  *
1007  * Folios in this VMA will be aligned to, and at least the size of the
1008  * number of bytes returned by this function.
1009  *
1010  * Return: The default size of the folios allocated when backing a VMA.
1011  */
1012 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
1013 {
1014 	if (vma->vm_ops && vma->vm_ops->pagesize)
1015 		return vma->vm_ops->pagesize(vma);
1016 	return PAGE_SIZE;
1017 }
1018 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
1019 
1020 /*
1021  * Return the page size being used by the MMU to back a VMA. In the majority
1022  * of cases, the page size used by the kernel matches the MMU size. On
1023  * architectures where it differs, an architecture-specific 'strong'
1024  * version of this symbol is required.
1025  */
1026 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
1027 {
1028 	return vma_kernel_pagesize(vma);
1029 }
1030 
1031 /*
1032  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
1033  * bits of the reservation map pointer, which are always clear due to
1034  * alignment.
1035  */
1036 #define HPAGE_RESV_OWNER    (1UL << 0)
1037 #define HPAGE_RESV_UNMAPPED (1UL << 1)
1038 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
1039 
1040 /*
1041  * These helpers are used to track how many pages are reserved for
1042  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
1043  * is guaranteed to have their future faults succeed.
1044  *
1045  * With the exception of hugetlb_dup_vma_private() which is called at fork(),
1046  * the reserve counters are updated with the hugetlb_lock held. It is safe
1047  * to reset the VMA at fork() time as it is not in use yet and there is no
1048  * chance of the global counters getting corrupted as a result of the values.
1049  *
1050  * The private mapping reservation is represented in a subtly different
1051  * manner to a shared mapping.  A shared mapping has a region map associated
1052  * with the underlying file, this region map represents the backing file
1053  * pages which have ever had a reservation assigned which this persists even
1054  * after the page is instantiated.  A private mapping has a region map
1055  * associated with the original mmap which is attached to all VMAs which
1056  * reference it, this region map represents those offsets which have consumed
1057  * reservation ie. where pages have been instantiated.
1058  */
1059 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
1060 {
1061 	return (unsigned long)vma->vm_private_data;
1062 }
1063 
1064 static void set_vma_private_data(struct vm_area_struct *vma,
1065 							unsigned long value)
1066 {
1067 	vma->vm_private_data = (void *)value;
1068 }
1069 
1070 static void
1071 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
1072 					  struct hugetlb_cgroup *h_cg,
1073 					  struct hstate *h)
1074 {
1075 #ifdef CONFIG_CGROUP_HUGETLB
1076 	if (!h_cg || !h) {
1077 		resv_map->reservation_counter = NULL;
1078 		resv_map->pages_per_hpage = 0;
1079 		resv_map->css = NULL;
1080 	} else {
1081 		resv_map->reservation_counter =
1082 			&h_cg->rsvd_hugepage[hstate_index(h)];
1083 		resv_map->pages_per_hpage = pages_per_huge_page(h);
1084 		resv_map->css = &h_cg->css;
1085 	}
1086 #endif
1087 }
1088 
1089 struct resv_map *resv_map_alloc(void)
1090 {
1091 	struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
1092 	struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
1093 
1094 	if (!resv_map || !rg) {
1095 		kfree(resv_map);
1096 		kfree(rg);
1097 		return NULL;
1098 	}
1099 
1100 	kref_init(&resv_map->refs);
1101 	spin_lock_init(&resv_map->lock);
1102 	INIT_LIST_HEAD(&resv_map->regions);
1103 	init_rwsem(&resv_map->rw_sema);
1104 
1105 	resv_map->adds_in_progress = 0;
1106 	/*
1107 	 * Initialize these to 0. On shared mappings, 0's here indicate these
1108 	 * fields don't do cgroup accounting. On private mappings, these will be
1109 	 * re-initialized to the proper values, to indicate that hugetlb cgroup
1110 	 * reservations are to be un-charged from here.
1111 	 */
1112 	resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
1113 
1114 	INIT_LIST_HEAD(&resv_map->region_cache);
1115 	list_add(&rg->link, &resv_map->region_cache);
1116 	resv_map->region_cache_count = 1;
1117 
1118 	return resv_map;
1119 }
1120 
1121 void resv_map_release(struct kref *ref)
1122 {
1123 	struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
1124 	struct list_head *head = &resv_map->region_cache;
1125 	struct file_region *rg, *trg;
1126 
1127 	/* Clear out any active regions before we release the map. */
1128 	region_del(resv_map, 0, LONG_MAX);
1129 
1130 	/* ... and any entries left in the cache */
1131 	list_for_each_entry_safe(rg, trg, head, link) {
1132 		list_del(&rg->link);
1133 		kfree(rg);
1134 	}
1135 
1136 	VM_BUG_ON(resv_map->adds_in_progress);
1137 
1138 	kfree(resv_map);
1139 }
1140 
1141 static inline struct resv_map *inode_resv_map(struct inode *inode)
1142 {
1143 	/*
1144 	 * At inode evict time, i_mapping may not point to the original
1145 	 * address space within the inode.  This original address space
1146 	 * contains the pointer to the resv_map.  So, always use the
1147 	 * address space embedded within the inode.
1148 	 * The VERY common case is inode->mapping == &inode->i_data but,
1149 	 * this may not be true for device special inodes.
1150 	 */
1151 	return (struct resv_map *)(&inode->i_data)->private_data;
1152 }
1153 
1154 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
1155 {
1156 	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1157 	if (vma->vm_flags & VM_MAYSHARE) {
1158 		struct address_space *mapping = vma->vm_file->f_mapping;
1159 		struct inode *inode = mapping->host;
1160 
1161 		return inode_resv_map(inode);
1162 
1163 	} else {
1164 		return (struct resv_map *)(get_vma_private_data(vma) &
1165 							~HPAGE_RESV_MASK);
1166 	}
1167 }
1168 
1169 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
1170 {
1171 	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1172 	VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1173 
1174 	set_vma_private_data(vma, (unsigned long)map);
1175 }
1176 
1177 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
1178 {
1179 	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1180 	VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1181 
1182 	set_vma_private_data(vma, get_vma_private_data(vma) | flags);
1183 }
1184 
1185 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
1186 {
1187 	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1188 
1189 	return (get_vma_private_data(vma) & flag) != 0;
1190 }
1191 
1192 bool __vma_private_lock(struct vm_area_struct *vma)
1193 {
1194 	return !(vma->vm_flags & VM_MAYSHARE) &&
1195 		get_vma_private_data(vma) & ~HPAGE_RESV_MASK &&
1196 		is_vma_resv_set(vma, HPAGE_RESV_OWNER);
1197 }
1198 
1199 void hugetlb_dup_vma_private(struct vm_area_struct *vma)
1200 {
1201 	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1202 	/*
1203 	 * Clear vm_private_data
1204 	 * - For shared mappings this is a per-vma semaphore that may be
1205 	 *   allocated in a subsequent call to hugetlb_vm_op_open.
1206 	 *   Before clearing, make sure pointer is not associated with vma
1207 	 *   as this will leak the structure.  This is the case when called
1208 	 *   via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already
1209 	 *   been called to allocate a new structure.
1210 	 * - For MAP_PRIVATE mappings, this is the reserve map which does
1211 	 *   not apply to children.  Faults generated by the children are
1212 	 *   not guaranteed to succeed, even if read-only.
1213 	 */
1214 	if (vma->vm_flags & VM_MAYSHARE) {
1215 		struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
1216 
1217 		if (vma_lock && vma_lock->vma != vma)
1218 			vma->vm_private_data = NULL;
1219 	} else
1220 		vma->vm_private_data = NULL;
1221 }
1222 
1223 /*
1224  * Reset and decrement one ref on hugepage private reservation.
1225  * Called with mm->mmap_lock writer semaphore held.
1226  * This function should be only used by move_vma() and operate on
1227  * same sized vma. It should never come here with last ref on the
1228  * reservation.
1229  */
1230 void clear_vma_resv_huge_pages(struct vm_area_struct *vma)
1231 {
1232 	/*
1233 	 * Clear the old hugetlb private page reservation.
1234 	 * It has already been transferred to new_vma.
1235 	 *
1236 	 * During a mremap() operation of a hugetlb vma we call move_vma()
1237 	 * which copies vma into new_vma and unmaps vma. After the copy
1238 	 * operation both new_vma and vma share a reference to the resv_map
1239 	 * struct, and at that point vma is about to be unmapped. We don't
1240 	 * want to return the reservation to the pool at unmap of vma because
1241 	 * the reservation still lives on in new_vma, so simply decrement the
1242 	 * ref here and remove the resv_map reference from this vma.
1243 	 */
1244 	struct resv_map *reservations = vma_resv_map(vma);
1245 
1246 	if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1247 		resv_map_put_hugetlb_cgroup_uncharge_info(reservations);
1248 		kref_put(&reservations->refs, resv_map_release);
1249 	}
1250 
1251 	hugetlb_dup_vma_private(vma);
1252 }
1253 
1254 /* Returns true if the VMA has associated reserve pages */
1255 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1256 {
1257 	if (vma->vm_flags & VM_NORESERVE) {
1258 		/*
1259 		 * This address is already reserved by other process(chg == 0),
1260 		 * so, we should decrement reserved count. Without decrementing,
1261 		 * reserve count remains after releasing inode, because this
1262 		 * allocated page will go into page cache and is regarded as
1263 		 * coming from reserved pool in releasing step.  Currently, we
1264 		 * don't have any other solution to deal with this situation
1265 		 * properly, so add work-around here.
1266 		 */
1267 		if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1268 			return true;
1269 		else
1270 			return false;
1271 	}
1272 
1273 	/* Shared mappings always use reserves */
1274 	if (vma->vm_flags & VM_MAYSHARE) {
1275 		/*
1276 		 * We know VM_NORESERVE is not set.  Therefore, there SHOULD
1277 		 * be a region map for all pages.  The only situation where
1278 		 * there is no region map is if a hole was punched via
1279 		 * fallocate.  In this case, there really are no reserves to
1280 		 * use.  This situation is indicated if chg != 0.
1281 		 */
1282 		if (chg)
1283 			return false;
1284 		else
1285 			return true;
1286 	}
1287 
1288 	/*
1289 	 * Only the process that called mmap() has reserves for
1290 	 * private mappings.
1291 	 */
1292 	if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1293 		/*
1294 		 * Like the shared case above, a hole punch or truncate
1295 		 * could have been performed on the private mapping.
1296 		 * Examine the value of chg to determine if reserves
1297 		 * actually exist or were previously consumed.
1298 		 * Very Subtle - The value of chg comes from a previous
1299 		 * call to vma_needs_reserves().  The reserve map for
1300 		 * private mappings has different (opposite) semantics
1301 		 * than that of shared mappings.  vma_needs_reserves()
1302 		 * has already taken this difference in semantics into
1303 		 * account.  Therefore, the meaning of chg is the same
1304 		 * as in the shared case above.  Code could easily be
1305 		 * combined, but keeping it separate draws attention to
1306 		 * subtle differences.
1307 		 */
1308 		if (chg)
1309 			return false;
1310 		else
1311 			return true;
1312 	}
1313 
1314 	return false;
1315 }
1316 
1317 static void enqueue_hugetlb_folio(struct hstate *h, struct folio *folio)
1318 {
1319 	int nid = folio_nid(folio);
1320 
1321 	lockdep_assert_held(&hugetlb_lock);
1322 	VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1323 
1324 	list_move(&folio->lru, &h->hugepage_freelists[nid]);
1325 	h->free_huge_pages++;
1326 	h->free_huge_pages_node[nid]++;
1327 	folio_set_hugetlb_freed(folio);
1328 }
1329 
1330 static struct folio *dequeue_hugetlb_folio_node_exact(struct hstate *h,
1331 								int nid)
1332 {
1333 	struct folio *folio;
1334 	bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1335 
1336 	lockdep_assert_held(&hugetlb_lock);
1337 	list_for_each_entry(folio, &h->hugepage_freelists[nid], lru) {
1338 		if (pin && !folio_is_longterm_pinnable(folio))
1339 			continue;
1340 
1341 		if (folio_test_hwpoison(folio))
1342 			continue;
1343 
1344 		list_move(&folio->lru, &h->hugepage_activelist);
1345 		folio_ref_unfreeze(folio, 1);
1346 		folio_clear_hugetlb_freed(folio);
1347 		h->free_huge_pages--;
1348 		h->free_huge_pages_node[nid]--;
1349 		return folio;
1350 	}
1351 
1352 	return NULL;
1353 }
1354 
1355 static struct folio *dequeue_hugetlb_folio_nodemask(struct hstate *h, gfp_t gfp_mask,
1356 							int nid, nodemask_t *nmask)
1357 {
1358 	unsigned int cpuset_mems_cookie;
1359 	struct zonelist *zonelist;
1360 	struct zone *zone;
1361 	struct zoneref *z;
1362 	int node = NUMA_NO_NODE;
1363 
1364 	zonelist = node_zonelist(nid, gfp_mask);
1365 
1366 retry_cpuset:
1367 	cpuset_mems_cookie = read_mems_allowed_begin();
1368 	for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1369 		struct folio *folio;
1370 
1371 		if (!cpuset_zone_allowed(zone, gfp_mask))
1372 			continue;
1373 		/*
1374 		 * no need to ask again on the same node. Pool is node rather than
1375 		 * zone aware
1376 		 */
1377 		if (zone_to_nid(zone) == node)
1378 			continue;
1379 		node = zone_to_nid(zone);
1380 
1381 		folio = dequeue_hugetlb_folio_node_exact(h, node);
1382 		if (folio)
1383 			return folio;
1384 	}
1385 	if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1386 		goto retry_cpuset;
1387 
1388 	return NULL;
1389 }
1390 
1391 static unsigned long available_huge_pages(struct hstate *h)
1392 {
1393 	return h->free_huge_pages - h->resv_huge_pages;
1394 }
1395 
1396 static struct folio *dequeue_hugetlb_folio_vma(struct hstate *h,
1397 				struct vm_area_struct *vma,
1398 				unsigned long address, int avoid_reserve,
1399 				long chg)
1400 {
1401 	struct folio *folio = NULL;
1402 	struct mempolicy *mpol;
1403 	gfp_t gfp_mask;
1404 	nodemask_t *nodemask;
1405 	int nid;
1406 
1407 	/*
1408 	 * A child process with MAP_PRIVATE mappings created by their parent
1409 	 * have no page reserves. This check ensures that reservations are
1410 	 * not "stolen". The child may still get SIGKILLed
1411 	 */
1412 	if (!vma_has_reserves(vma, chg) && !available_huge_pages(h))
1413 		goto err;
1414 
1415 	/* If reserves cannot be used, ensure enough pages are in the pool */
1416 	if (avoid_reserve && !available_huge_pages(h))
1417 		goto err;
1418 
1419 	gfp_mask = htlb_alloc_mask(h);
1420 	nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1421 
1422 	if (mpol_is_preferred_many(mpol)) {
1423 		folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
1424 							nid, nodemask);
1425 
1426 		/* Fallback to all nodes if page==NULL */
1427 		nodemask = NULL;
1428 	}
1429 
1430 	if (!folio)
1431 		folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
1432 							nid, nodemask);
1433 
1434 	if (folio && !avoid_reserve && vma_has_reserves(vma, chg)) {
1435 		folio_set_hugetlb_restore_reserve(folio);
1436 		h->resv_huge_pages--;
1437 	}
1438 
1439 	mpol_cond_put(mpol);
1440 	return folio;
1441 
1442 err:
1443 	return NULL;
1444 }
1445 
1446 /*
1447  * common helper functions for hstate_next_node_to_{alloc|free}.
1448  * We may have allocated or freed a huge page based on a different
1449  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1450  * be outside of *nodes_allowed.  Ensure that we use an allowed
1451  * node for alloc or free.
1452  */
1453 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1454 {
1455 	nid = next_node_in(nid, *nodes_allowed);
1456 	VM_BUG_ON(nid >= MAX_NUMNODES);
1457 
1458 	return nid;
1459 }
1460 
1461 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1462 {
1463 	if (!node_isset(nid, *nodes_allowed))
1464 		nid = next_node_allowed(nid, nodes_allowed);
1465 	return nid;
1466 }
1467 
1468 /*
1469  * returns the previously saved node ["this node"] from which to
1470  * allocate a persistent huge page for the pool and advance the
1471  * next node from which to allocate, handling wrap at end of node
1472  * mask.
1473  */
1474 static int hstate_next_node_to_alloc(struct hstate *h,
1475 					nodemask_t *nodes_allowed)
1476 {
1477 	int nid;
1478 
1479 	VM_BUG_ON(!nodes_allowed);
1480 
1481 	nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1482 	h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1483 
1484 	return nid;
1485 }
1486 
1487 /*
1488  * helper for remove_pool_huge_page() - return the previously saved
1489  * node ["this node"] from which to free a huge page.  Advance the
1490  * next node id whether or not we find a free huge page to free so
1491  * that the next attempt to free addresses the next node.
1492  */
1493 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1494 {
1495 	int nid;
1496 
1497 	VM_BUG_ON(!nodes_allowed);
1498 
1499 	nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1500 	h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1501 
1502 	return nid;
1503 }
1504 
1505 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)		\
1506 	for (nr_nodes = nodes_weight(*mask);				\
1507 		nr_nodes > 0 &&						\
1508 		((node = hstate_next_node_to_alloc(hs, mask)) || 1);	\
1509 		nr_nodes--)
1510 
1511 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)		\
1512 	for (nr_nodes = nodes_weight(*mask);				\
1513 		nr_nodes > 0 &&						\
1514 		((node = hstate_next_node_to_free(hs, mask)) || 1);	\
1515 		nr_nodes--)
1516 
1517 /* used to demote non-gigantic_huge pages as well */
1518 static void __destroy_compound_gigantic_folio(struct folio *folio,
1519 					unsigned int order, bool demote)
1520 {
1521 	int i;
1522 	int nr_pages = 1 << order;
1523 	struct page *p;
1524 
1525 	atomic_set(&folio->_entire_mapcount, 0);
1526 	atomic_set(&folio->_nr_pages_mapped, 0);
1527 	atomic_set(&folio->_pincount, 0);
1528 
1529 	for (i = 1; i < nr_pages; i++) {
1530 		p = folio_page(folio, i);
1531 		p->flags &= ~PAGE_FLAGS_CHECK_AT_FREE;
1532 		p->mapping = NULL;
1533 		clear_compound_head(p);
1534 		if (!demote)
1535 			set_page_refcounted(p);
1536 	}
1537 
1538 	__folio_clear_head(folio);
1539 }
1540 
1541 static void destroy_compound_hugetlb_folio_for_demote(struct folio *folio,
1542 					unsigned int order)
1543 {
1544 	__destroy_compound_gigantic_folio(folio, order, true);
1545 }
1546 
1547 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1548 static void destroy_compound_gigantic_folio(struct folio *folio,
1549 					unsigned int order)
1550 {
1551 	__destroy_compound_gigantic_folio(folio, order, false);
1552 }
1553 
1554 static void free_gigantic_folio(struct folio *folio, unsigned int order)
1555 {
1556 	/*
1557 	 * If the page isn't allocated using the cma allocator,
1558 	 * cma_release() returns false.
1559 	 */
1560 #ifdef CONFIG_CMA
1561 	int nid = folio_nid(folio);
1562 
1563 	if (cma_release(hugetlb_cma[nid], &folio->page, 1 << order))
1564 		return;
1565 #endif
1566 
1567 	free_contig_range(folio_pfn(folio), 1 << order);
1568 }
1569 
1570 #ifdef CONFIG_CONTIG_ALLOC
1571 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1572 		int nid, nodemask_t *nodemask)
1573 {
1574 	struct page *page;
1575 	unsigned long nr_pages = pages_per_huge_page(h);
1576 	if (nid == NUMA_NO_NODE)
1577 		nid = numa_mem_id();
1578 
1579 #ifdef CONFIG_CMA
1580 	{
1581 		int node;
1582 
1583 		if (hugetlb_cma[nid]) {
1584 			page = cma_alloc(hugetlb_cma[nid], nr_pages,
1585 					huge_page_order(h), true);
1586 			if (page)
1587 				return page_folio(page);
1588 		}
1589 
1590 		if (!(gfp_mask & __GFP_THISNODE)) {
1591 			for_each_node_mask(node, *nodemask) {
1592 				if (node == nid || !hugetlb_cma[node])
1593 					continue;
1594 
1595 				page = cma_alloc(hugetlb_cma[node], nr_pages,
1596 						huge_page_order(h), true);
1597 				if (page)
1598 					return page_folio(page);
1599 			}
1600 		}
1601 	}
1602 #endif
1603 
1604 	page = alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1605 	return page ? page_folio(page) : NULL;
1606 }
1607 
1608 #else /* !CONFIG_CONTIG_ALLOC */
1609 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1610 					int nid, nodemask_t *nodemask)
1611 {
1612 	return NULL;
1613 }
1614 #endif /* CONFIG_CONTIG_ALLOC */
1615 
1616 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1617 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1618 					int nid, nodemask_t *nodemask)
1619 {
1620 	return NULL;
1621 }
1622 static inline void free_gigantic_folio(struct folio *folio,
1623 						unsigned int order) { }
1624 static inline void destroy_compound_gigantic_folio(struct folio *folio,
1625 						unsigned int order) { }
1626 #endif
1627 
1628 static inline void __clear_hugetlb_destructor(struct hstate *h,
1629 						struct folio *folio)
1630 {
1631 	lockdep_assert_held(&hugetlb_lock);
1632 
1633 	__folio_clear_hugetlb(folio);
1634 }
1635 
1636 /*
1637  * Remove hugetlb folio from lists.
1638  * If vmemmap exists for the folio, update dtor so that the folio appears
1639  * as just a compound page.  Otherwise, wait until after allocating vmemmap
1640  * to update dtor.
1641  *
1642  * A reference is held on the folio, except in the case of demote.
1643  *
1644  * Must be called with hugetlb lock held.
1645  */
1646 static void __remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1647 							bool adjust_surplus,
1648 							bool demote)
1649 {
1650 	int nid = folio_nid(folio);
1651 
1652 	VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio);
1653 	VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio);
1654 
1655 	lockdep_assert_held(&hugetlb_lock);
1656 	if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1657 		return;
1658 
1659 	list_del(&folio->lru);
1660 
1661 	if (folio_test_hugetlb_freed(folio)) {
1662 		h->free_huge_pages--;
1663 		h->free_huge_pages_node[nid]--;
1664 	}
1665 	if (adjust_surplus) {
1666 		h->surplus_huge_pages--;
1667 		h->surplus_huge_pages_node[nid]--;
1668 	}
1669 
1670 	/*
1671 	 * We can only clear the hugetlb destructor after allocating vmemmap
1672 	 * pages.  Otherwise, someone (memory error handling) may try to write
1673 	 * to tail struct pages.
1674 	 */
1675 	if (!folio_test_hugetlb_vmemmap_optimized(folio))
1676 		__clear_hugetlb_destructor(h, folio);
1677 
1678 	 /*
1679 	  * In the case of demote we do not ref count the page as it will soon
1680 	  * be turned into a page of smaller size.
1681 	 */
1682 	if (!demote)
1683 		folio_ref_unfreeze(folio, 1);
1684 
1685 	h->nr_huge_pages--;
1686 	h->nr_huge_pages_node[nid]--;
1687 }
1688 
1689 static void remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1690 							bool adjust_surplus)
1691 {
1692 	__remove_hugetlb_folio(h, folio, adjust_surplus, false);
1693 }
1694 
1695 static void remove_hugetlb_folio_for_demote(struct hstate *h, struct folio *folio,
1696 							bool adjust_surplus)
1697 {
1698 	__remove_hugetlb_folio(h, folio, adjust_surplus, true);
1699 }
1700 
1701 static void add_hugetlb_folio(struct hstate *h, struct folio *folio,
1702 			     bool adjust_surplus)
1703 {
1704 	int zeroed;
1705 	int nid = folio_nid(folio);
1706 
1707 	VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio);
1708 
1709 	lockdep_assert_held(&hugetlb_lock);
1710 
1711 	INIT_LIST_HEAD(&folio->lru);
1712 	h->nr_huge_pages++;
1713 	h->nr_huge_pages_node[nid]++;
1714 
1715 	if (adjust_surplus) {
1716 		h->surplus_huge_pages++;
1717 		h->surplus_huge_pages_node[nid]++;
1718 	}
1719 
1720 	__folio_set_hugetlb(folio);
1721 	folio_change_private(folio, NULL);
1722 	/*
1723 	 * We have to set hugetlb_vmemmap_optimized again as above
1724 	 * folio_change_private(folio, NULL) cleared it.
1725 	 */
1726 	folio_set_hugetlb_vmemmap_optimized(folio);
1727 
1728 	/*
1729 	 * This folio is about to be managed by the hugetlb allocator and
1730 	 * should have no users.  Drop our reference, and check for others
1731 	 * just in case.
1732 	 */
1733 	zeroed = folio_put_testzero(folio);
1734 	if (unlikely(!zeroed))
1735 		/*
1736 		 * It is VERY unlikely soneone else has taken a ref
1737 		 * on the folio.  In this case, we simply return as
1738 		 * free_huge_folio() will be called when this other ref
1739 		 * is dropped.
1740 		 */
1741 		return;
1742 
1743 	arch_clear_hugepage_flags(&folio->page);
1744 	enqueue_hugetlb_folio(h, folio);
1745 }
1746 
1747 static void __update_and_free_hugetlb_folio(struct hstate *h,
1748 						struct folio *folio)
1749 {
1750 	if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1751 		return;
1752 
1753 	/*
1754 	 * If we don't know which subpages are hwpoisoned, we can't free
1755 	 * the hugepage, so it's leaked intentionally.
1756 	 */
1757 	if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1758 		return;
1759 
1760 	if (hugetlb_vmemmap_restore(h, &folio->page)) {
1761 		spin_lock_irq(&hugetlb_lock);
1762 		/*
1763 		 * If we cannot allocate vmemmap pages, just refuse to free the
1764 		 * page and put the page back on the hugetlb free list and treat
1765 		 * as a surplus page.
1766 		 */
1767 		add_hugetlb_folio(h, folio, true);
1768 		spin_unlock_irq(&hugetlb_lock);
1769 		return;
1770 	}
1771 
1772 	/*
1773 	 * Move PageHWPoison flag from head page to the raw error pages,
1774 	 * which makes any healthy subpages reusable.
1775 	 */
1776 	if (unlikely(folio_test_hwpoison(folio)))
1777 		folio_clear_hugetlb_hwpoison(folio);
1778 
1779 	/*
1780 	 * If vmemmap pages were allocated above, then we need to clear the
1781 	 * hugetlb destructor under the hugetlb lock.
1782 	 */
1783 	if (folio_test_hugetlb(folio)) {
1784 		spin_lock_irq(&hugetlb_lock);
1785 		__clear_hugetlb_destructor(h, folio);
1786 		spin_unlock_irq(&hugetlb_lock);
1787 	}
1788 
1789 	/*
1790 	 * Non-gigantic pages demoted from CMA allocated gigantic pages
1791 	 * need to be given back to CMA in free_gigantic_folio.
1792 	 */
1793 	if (hstate_is_gigantic(h) ||
1794 	    hugetlb_cma_folio(folio, huge_page_order(h))) {
1795 		destroy_compound_gigantic_folio(folio, huge_page_order(h));
1796 		free_gigantic_folio(folio, huge_page_order(h));
1797 	} else {
1798 		__free_pages(&folio->page, huge_page_order(h));
1799 	}
1800 }
1801 
1802 /*
1803  * As update_and_free_hugetlb_folio() can be called under any context, so we cannot
1804  * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1805  * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1806  * the vmemmap pages.
1807  *
1808  * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1809  * freed and frees them one-by-one. As the page->mapping pointer is going
1810  * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1811  * structure of a lockless linked list of huge pages to be freed.
1812  */
1813 static LLIST_HEAD(hpage_freelist);
1814 
1815 static void free_hpage_workfn(struct work_struct *work)
1816 {
1817 	struct llist_node *node;
1818 
1819 	node = llist_del_all(&hpage_freelist);
1820 
1821 	while (node) {
1822 		struct page *page;
1823 		struct hstate *h;
1824 
1825 		page = container_of((struct address_space **)node,
1826 				     struct page, mapping);
1827 		node = node->next;
1828 		page->mapping = NULL;
1829 		/*
1830 		 * The VM_BUG_ON_FOLIO(!folio_test_hugetlb(folio), folio) in
1831 		 * folio_hstate() is going to trigger because a previous call to
1832 		 * remove_hugetlb_folio() will clear the hugetlb bit, so do
1833 		 * not use folio_hstate() directly.
1834 		 */
1835 		h = size_to_hstate(page_size(page));
1836 
1837 		__update_and_free_hugetlb_folio(h, page_folio(page));
1838 
1839 		cond_resched();
1840 	}
1841 }
1842 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1843 
1844 static inline void flush_free_hpage_work(struct hstate *h)
1845 {
1846 	if (hugetlb_vmemmap_optimizable(h))
1847 		flush_work(&free_hpage_work);
1848 }
1849 
1850 static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio,
1851 				 bool atomic)
1852 {
1853 	if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) {
1854 		__update_and_free_hugetlb_folio(h, folio);
1855 		return;
1856 	}
1857 
1858 	/*
1859 	 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1860 	 *
1861 	 * Only call schedule_work() if hpage_freelist is previously
1862 	 * empty. Otherwise, schedule_work() had been called but the workfn
1863 	 * hasn't retrieved the list yet.
1864 	 */
1865 	if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist))
1866 		schedule_work(&free_hpage_work);
1867 }
1868 
1869 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1870 {
1871 	struct page *page, *t_page;
1872 	struct folio *folio;
1873 
1874 	list_for_each_entry_safe(page, t_page, list, lru) {
1875 		folio = page_folio(page);
1876 		update_and_free_hugetlb_folio(h, folio, false);
1877 		cond_resched();
1878 	}
1879 }
1880 
1881 struct hstate *size_to_hstate(unsigned long size)
1882 {
1883 	struct hstate *h;
1884 
1885 	for_each_hstate(h) {
1886 		if (huge_page_size(h) == size)
1887 			return h;
1888 	}
1889 	return NULL;
1890 }
1891 
1892 void free_huge_folio(struct folio *folio)
1893 {
1894 	/*
1895 	 * Can't pass hstate in here because it is called from the
1896 	 * compound page destructor.
1897 	 */
1898 	struct hstate *h = folio_hstate(folio);
1899 	int nid = folio_nid(folio);
1900 	struct hugepage_subpool *spool = hugetlb_folio_subpool(folio);
1901 	bool restore_reserve;
1902 	unsigned long flags;
1903 
1904 	VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1905 	VM_BUG_ON_FOLIO(folio_mapcount(folio), folio);
1906 
1907 	hugetlb_set_folio_subpool(folio, NULL);
1908 	if (folio_test_anon(folio))
1909 		__ClearPageAnonExclusive(&folio->page);
1910 	folio->mapping = NULL;
1911 	restore_reserve = folio_test_hugetlb_restore_reserve(folio);
1912 	folio_clear_hugetlb_restore_reserve(folio);
1913 
1914 	/*
1915 	 * If HPageRestoreReserve was set on page, page allocation consumed a
1916 	 * reservation.  If the page was associated with a subpool, there
1917 	 * would have been a page reserved in the subpool before allocation
1918 	 * via hugepage_subpool_get_pages().  Since we are 'restoring' the
1919 	 * reservation, do not call hugepage_subpool_put_pages() as this will
1920 	 * remove the reserved page from the subpool.
1921 	 */
1922 	if (!restore_reserve) {
1923 		/*
1924 		 * A return code of zero implies that the subpool will be
1925 		 * under its minimum size if the reservation is not restored
1926 		 * after page is free.  Therefore, force restore_reserve
1927 		 * operation.
1928 		 */
1929 		if (hugepage_subpool_put_pages(spool, 1) == 0)
1930 			restore_reserve = true;
1931 	}
1932 
1933 	spin_lock_irqsave(&hugetlb_lock, flags);
1934 	folio_clear_hugetlb_migratable(folio);
1935 	hugetlb_cgroup_uncharge_folio(hstate_index(h),
1936 				     pages_per_huge_page(h), folio);
1937 	hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
1938 					  pages_per_huge_page(h), folio);
1939 	if (restore_reserve)
1940 		h->resv_huge_pages++;
1941 
1942 	if (folio_test_hugetlb_temporary(folio)) {
1943 		remove_hugetlb_folio(h, folio, false);
1944 		spin_unlock_irqrestore(&hugetlb_lock, flags);
1945 		update_and_free_hugetlb_folio(h, folio, true);
1946 	} else if (h->surplus_huge_pages_node[nid]) {
1947 		/* remove the page from active list */
1948 		remove_hugetlb_folio(h, folio, true);
1949 		spin_unlock_irqrestore(&hugetlb_lock, flags);
1950 		update_and_free_hugetlb_folio(h, folio, true);
1951 	} else {
1952 		arch_clear_hugepage_flags(&folio->page);
1953 		enqueue_hugetlb_folio(h, folio);
1954 		spin_unlock_irqrestore(&hugetlb_lock, flags);
1955 	}
1956 }
1957 
1958 /*
1959  * Must be called with the hugetlb lock held
1960  */
1961 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1962 {
1963 	lockdep_assert_held(&hugetlb_lock);
1964 	h->nr_huge_pages++;
1965 	h->nr_huge_pages_node[nid]++;
1966 }
1967 
1968 static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio)
1969 {
1970 	hugetlb_vmemmap_optimize(h, &folio->page);
1971 	INIT_LIST_HEAD(&folio->lru);
1972 	__folio_set_hugetlb(folio);
1973 	hugetlb_set_folio_subpool(folio, NULL);
1974 	set_hugetlb_cgroup(folio, NULL);
1975 	set_hugetlb_cgroup_rsvd(folio, NULL);
1976 }
1977 
1978 static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid)
1979 {
1980 	__prep_new_hugetlb_folio(h, folio);
1981 	spin_lock_irq(&hugetlb_lock);
1982 	__prep_account_new_huge_page(h, nid);
1983 	spin_unlock_irq(&hugetlb_lock);
1984 }
1985 
1986 static bool __prep_compound_gigantic_folio(struct folio *folio,
1987 					unsigned int order, bool demote)
1988 {
1989 	int i, j;
1990 	int nr_pages = 1 << order;
1991 	struct page *p;
1992 
1993 	__folio_clear_reserved(folio);
1994 	for (i = 0; i < nr_pages; i++) {
1995 		p = folio_page(folio, i);
1996 
1997 		/*
1998 		 * For gigantic hugepages allocated through bootmem at
1999 		 * boot, it's safer to be consistent with the not-gigantic
2000 		 * hugepages and clear the PG_reserved bit from all tail pages
2001 		 * too.  Otherwise drivers using get_user_pages() to access tail
2002 		 * pages may get the reference counting wrong if they see
2003 		 * PG_reserved set on a tail page (despite the head page not
2004 		 * having PG_reserved set).  Enforcing this consistency between
2005 		 * head and tail pages allows drivers to optimize away a check
2006 		 * on the head page when they need know if put_page() is needed
2007 		 * after get_user_pages().
2008 		 */
2009 		if (i != 0)	/* head page cleared above */
2010 			__ClearPageReserved(p);
2011 		/*
2012 		 * Subtle and very unlikely
2013 		 *
2014 		 * Gigantic 'page allocators' such as memblock or cma will
2015 		 * return a set of pages with each page ref counted.  We need
2016 		 * to turn this set of pages into a compound page with tail
2017 		 * page ref counts set to zero.  Code such as speculative page
2018 		 * cache adding could take a ref on a 'to be' tail page.
2019 		 * We need to respect any increased ref count, and only set
2020 		 * the ref count to zero if count is currently 1.  If count
2021 		 * is not 1, we return an error.  An error return indicates
2022 		 * the set of pages can not be converted to a gigantic page.
2023 		 * The caller who allocated the pages should then discard the
2024 		 * pages using the appropriate free interface.
2025 		 *
2026 		 * In the case of demote, the ref count will be zero.
2027 		 */
2028 		if (!demote) {
2029 			if (!page_ref_freeze(p, 1)) {
2030 				pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
2031 				goto out_error;
2032 			}
2033 		} else {
2034 			VM_BUG_ON_PAGE(page_count(p), p);
2035 		}
2036 		if (i != 0)
2037 			set_compound_head(p, &folio->page);
2038 	}
2039 	__folio_set_head(folio);
2040 	/* we rely on prep_new_hugetlb_folio to set the destructor */
2041 	folio_set_order(folio, order);
2042 	atomic_set(&folio->_entire_mapcount, -1);
2043 	atomic_set(&folio->_nr_pages_mapped, 0);
2044 	atomic_set(&folio->_pincount, 0);
2045 	return true;
2046 
2047 out_error:
2048 	/* undo page modifications made above */
2049 	for (j = 0; j < i; j++) {
2050 		p = folio_page(folio, j);
2051 		if (j != 0)
2052 			clear_compound_head(p);
2053 		set_page_refcounted(p);
2054 	}
2055 	/* need to clear PG_reserved on remaining tail pages  */
2056 	for (; j < nr_pages; j++) {
2057 		p = folio_page(folio, j);
2058 		__ClearPageReserved(p);
2059 	}
2060 	return false;
2061 }
2062 
2063 static bool prep_compound_gigantic_folio(struct folio *folio,
2064 							unsigned int order)
2065 {
2066 	return __prep_compound_gigantic_folio(folio, order, false);
2067 }
2068 
2069 static bool prep_compound_gigantic_folio_for_demote(struct folio *folio,
2070 							unsigned int order)
2071 {
2072 	return __prep_compound_gigantic_folio(folio, order, true);
2073 }
2074 
2075 /*
2076  * Find and lock address space (mapping) in write mode.
2077  *
2078  * Upon entry, the page is locked which means that page_mapping() is
2079  * stable.  Due to locking order, we can only trylock_write.  If we can
2080  * not get the lock, simply return NULL to caller.
2081  */
2082 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
2083 {
2084 	struct address_space *mapping = page_mapping(hpage);
2085 
2086 	if (!mapping)
2087 		return mapping;
2088 
2089 	if (i_mmap_trylock_write(mapping))
2090 		return mapping;
2091 
2092 	return NULL;
2093 }
2094 
2095 pgoff_t hugetlb_basepage_index(struct page *page)
2096 {
2097 	struct page *page_head = compound_head(page);
2098 	pgoff_t index = page_index(page_head);
2099 	unsigned long compound_idx;
2100 
2101 	if (compound_order(page_head) > MAX_ORDER)
2102 		compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
2103 	else
2104 		compound_idx = page - page_head;
2105 
2106 	return (index << compound_order(page_head)) + compound_idx;
2107 }
2108 
2109 static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h,
2110 		gfp_t gfp_mask, int nid, nodemask_t *nmask,
2111 		nodemask_t *node_alloc_noretry)
2112 {
2113 	int order = huge_page_order(h);
2114 	struct page *page;
2115 	bool alloc_try_hard = true;
2116 	bool retry = true;
2117 
2118 	/*
2119 	 * By default we always try hard to allocate the page with
2120 	 * __GFP_RETRY_MAYFAIL flag.  However, if we are allocating pages in
2121 	 * a loop (to adjust global huge page counts) and previous allocation
2122 	 * failed, do not continue to try hard on the same node.  Use the
2123 	 * node_alloc_noretry bitmap to manage this state information.
2124 	 */
2125 	if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
2126 		alloc_try_hard = false;
2127 	gfp_mask |= __GFP_COMP|__GFP_NOWARN;
2128 	if (alloc_try_hard)
2129 		gfp_mask |= __GFP_RETRY_MAYFAIL;
2130 	if (nid == NUMA_NO_NODE)
2131 		nid = numa_mem_id();
2132 retry:
2133 	page = __alloc_pages(gfp_mask, order, nid, nmask);
2134 
2135 	/* Freeze head page */
2136 	if (page && !page_ref_freeze(page, 1)) {
2137 		__free_pages(page, order);
2138 		if (retry) {	/* retry once */
2139 			retry = false;
2140 			goto retry;
2141 		}
2142 		/* WOW!  twice in a row. */
2143 		pr_warn("HugeTLB head page unexpected inflated ref count\n");
2144 		page = NULL;
2145 	}
2146 
2147 	/*
2148 	 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
2149 	 * indicates an overall state change.  Clear bit so that we resume
2150 	 * normal 'try hard' allocations.
2151 	 */
2152 	if (node_alloc_noretry && page && !alloc_try_hard)
2153 		node_clear(nid, *node_alloc_noretry);
2154 
2155 	/*
2156 	 * If we tried hard to get a page but failed, set bit so that
2157 	 * subsequent attempts will not try as hard until there is an
2158 	 * overall state change.
2159 	 */
2160 	if (node_alloc_noretry && !page && alloc_try_hard)
2161 		node_set(nid, *node_alloc_noretry);
2162 
2163 	if (!page) {
2164 		__count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
2165 		return NULL;
2166 	}
2167 
2168 	__count_vm_event(HTLB_BUDDY_PGALLOC);
2169 	return page_folio(page);
2170 }
2171 
2172 /*
2173  * Common helper to allocate a fresh hugetlb page. All specific allocators
2174  * should use this function to get new hugetlb pages
2175  *
2176  * Note that returned page is 'frozen':  ref count of head page and all tail
2177  * pages is zero.
2178  */
2179 static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h,
2180 		gfp_t gfp_mask, int nid, nodemask_t *nmask,
2181 		nodemask_t *node_alloc_noretry)
2182 {
2183 	struct folio *folio;
2184 	bool retry = false;
2185 
2186 retry:
2187 	if (hstate_is_gigantic(h))
2188 		folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask);
2189 	else
2190 		folio = alloc_buddy_hugetlb_folio(h, gfp_mask,
2191 				nid, nmask, node_alloc_noretry);
2192 	if (!folio)
2193 		return NULL;
2194 	if (hstate_is_gigantic(h)) {
2195 		if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) {
2196 			/*
2197 			 * Rare failure to convert pages to compound page.
2198 			 * Free pages and try again - ONCE!
2199 			 */
2200 			free_gigantic_folio(folio, huge_page_order(h));
2201 			if (!retry) {
2202 				retry = true;
2203 				goto retry;
2204 			}
2205 			return NULL;
2206 		}
2207 	}
2208 	prep_new_hugetlb_folio(h, folio, folio_nid(folio));
2209 
2210 	return folio;
2211 }
2212 
2213 /*
2214  * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
2215  * manner.
2216  */
2217 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
2218 				nodemask_t *node_alloc_noretry)
2219 {
2220 	struct folio *folio;
2221 	int nr_nodes, node;
2222 	gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2223 
2224 	for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2225 		folio = alloc_fresh_hugetlb_folio(h, gfp_mask, node,
2226 					nodes_allowed, node_alloc_noretry);
2227 		if (folio) {
2228 			free_huge_folio(folio); /* free it into the hugepage allocator */
2229 			return 1;
2230 		}
2231 	}
2232 
2233 	return 0;
2234 }
2235 
2236 /*
2237  * Remove huge page from pool from next node to free.  Attempt to keep
2238  * persistent huge pages more or less balanced over allowed nodes.
2239  * This routine only 'removes' the hugetlb page.  The caller must make
2240  * an additional call to free the page to low level allocators.
2241  * Called with hugetlb_lock locked.
2242  */
2243 static struct page *remove_pool_huge_page(struct hstate *h,
2244 						nodemask_t *nodes_allowed,
2245 						 bool acct_surplus)
2246 {
2247 	int nr_nodes, node;
2248 	struct page *page = NULL;
2249 	struct folio *folio;
2250 
2251 	lockdep_assert_held(&hugetlb_lock);
2252 	for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2253 		/*
2254 		 * If we're returning unused surplus pages, only examine
2255 		 * nodes with surplus pages.
2256 		 */
2257 		if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
2258 		    !list_empty(&h->hugepage_freelists[node])) {
2259 			page = list_entry(h->hugepage_freelists[node].next,
2260 					  struct page, lru);
2261 			folio = page_folio(page);
2262 			remove_hugetlb_folio(h, folio, acct_surplus);
2263 			break;
2264 		}
2265 	}
2266 
2267 	return page;
2268 }
2269 
2270 /*
2271  * Dissolve a given free hugepage into free buddy pages. This function does
2272  * nothing for in-use hugepages and non-hugepages.
2273  * This function returns values like below:
2274  *
2275  *  -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
2276  *           when the system is under memory pressure and the feature of
2277  *           freeing unused vmemmap pages associated with each hugetlb page
2278  *           is enabled.
2279  *  -EBUSY:  failed to dissolved free hugepages or the hugepage is in-use
2280  *           (allocated or reserved.)
2281  *       0:  successfully dissolved free hugepages or the page is not a
2282  *           hugepage (considered as already dissolved)
2283  */
2284 int dissolve_free_huge_page(struct page *page)
2285 {
2286 	int rc = -EBUSY;
2287 	struct folio *folio = page_folio(page);
2288 
2289 retry:
2290 	/* Not to disrupt normal path by vainly holding hugetlb_lock */
2291 	if (!folio_test_hugetlb(folio))
2292 		return 0;
2293 
2294 	spin_lock_irq(&hugetlb_lock);
2295 	if (!folio_test_hugetlb(folio)) {
2296 		rc = 0;
2297 		goto out;
2298 	}
2299 
2300 	if (!folio_ref_count(folio)) {
2301 		struct hstate *h = folio_hstate(folio);
2302 		if (!available_huge_pages(h))
2303 			goto out;
2304 
2305 		/*
2306 		 * We should make sure that the page is already on the free list
2307 		 * when it is dissolved.
2308 		 */
2309 		if (unlikely(!folio_test_hugetlb_freed(folio))) {
2310 			spin_unlock_irq(&hugetlb_lock);
2311 			cond_resched();
2312 
2313 			/*
2314 			 * Theoretically, we should return -EBUSY when we
2315 			 * encounter this race. In fact, we have a chance
2316 			 * to successfully dissolve the page if we do a
2317 			 * retry. Because the race window is quite small.
2318 			 * If we seize this opportunity, it is an optimization
2319 			 * for increasing the success rate of dissolving page.
2320 			 */
2321 			goto retry;
2322 		}
2323 
2324 		remove_hugetlb_folio(h, folio, false);
2325 		h->max_huge_pages--;
2326 		spin_unlock_irq(&hugetlb_lock);
2327 
2328 		/*
2329 		 * Normally update_and_free_hugtlb_folio will allocate required vmemmmap
2330 		 * before freeing the page.  update_and_free_hugtlb_folio will fail to
2331 		 * free the page if it can not allocate required vmemmap.  We
2332 		 * need to adjust max_huge_pages if the page is not freed.
2333 		 * Attempt to allocate vmemmmap here so that we can take
2334 		 * appropriate action on failure.
2335 		 */
2336 		rc = hugetlb_vmemmap_restore(h, &folio->page);
2337 		if (!rc) {
2338 			update_and_free_hugetlb_folio(h, folio, false);
2339 		} else {
2340 			spin_lock_irq(&hugetlb_lock);
2341 			add_hugetlb_folio(h, folio, false);
2342 			h->max_huge_pages++;
2343 			spin_unlock_irq(&hugetlb_lock);
2344 		}
2345 
2346 		return rc;
2347 	}
2348 out:
2349 	spin_unlock_irq(&hugetlb_lock);
2350 	return rc;
2351 }
2352 
2353 /*
2354  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2355  * make specified memory blocks removable from the system.
2356  * Note that this will dissolve a free gigantic hugepage completely, if any
2357  * part of it lies within the given range.
2358  * Also note that if dissolve_free_huge_page() returns with an error, all
2359  * free hugepages that were dissolved before that error are lost.
2360  */
2361 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2362 {
2363 	unsigned long pfn;
2364 	struct page *page;
2365 	int rc = 0;
2366 	unsigned int order;
2367 	struct hstate *h;
2368 
2369 	if (!hugepages_supported())
2370 		return rc;
2371 
2372 	order = huge_page_order(&default_hstate);
2373 	for_each_hstate(h)
2374 		order = min(order, huge_page_order(h));
2375 
2376 	for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) {
2377 		page = pfn_to_page(pfn);
2378 		rc = dissolve_free_huge_page(page);
2379 		if (rc)
2380 			break;
2381 	}
2382 
2383 	return rc;
2384 }
2385 
2386 /*
2387  * Allocates a fresh surplus page from the page allocator.
2388  */
2389 static struct folio *alloc_surplus_hugetlb_folio(struct hstate *h,
2390 				gfp_t gfp_mask,	int nid, nodemask_t *nmask)
2391 {
2392 	struct folio *folio = NULL;
2393 
2394 	if (hstate_is_gigantic(h))
2395 		return NULL;
2396 
2397 	spin_lock_irq(&hugetlb_lock);
2398 	if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2399 		goto out_unlock;
2400 	spin_unlock_irq(&hugetlb_lock);
2401 
2402 	folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2403 	if (!folio)
2404 		return NULL;
2405 
2406 	spin_lock_irq(&hugetlb_lock);
2407 	/*
2408 	 * We could have raced with the pool size change.
2409 	 * Double check that and simply deallocate the new page
2410 	 * if we would end up overcommiting the surpluses. Abuse
2411 	 * temporary page to workaround the nasty free_huge_folio
2412 	 * codeflow
2413 	 */
2414 	if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2415 		folio_set_hugetlb_temporary(folio);
2416 		spin_unlock_irq(&hugetlb_lock);
2417 		free_huge_folio(folio);
2418 		return NULL;
2419 	}
2420 
2421 	h->surplus_huge_pages++;
2422 	h->surplus_huge_pages_node[folio_nid(folio)]++;
2423 
2424 out_unlock:
2425 	spin_unlock_irq(&hugetlb_lock);
2426 
2427 	return folio;
2428 }
2429 
2430 static struct folio *alloc_migrate_hugetlb_folio(struct hstate *h, gfp_t gfp_mask,
2431 				     int nid, nodemask_t *nmask)
2432 {
2433 	struct folio *folio;
2434 
2435 	if (hstate_is_gigantic(h))
2436 		return NULL;
2437 
2438 	folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2439 	if (!folio)
2440 		return NULL;
2441 
2442 	/* fresh huge pages are frozen */
2443 	folio_ref_unfreeze(folio, 1);
2444 	/*
2445 	 * We do not account these pages as surplus because they are only
2446 	 * temporary and will be released properly on the last reference
2447 	 */
2448 	folio_set_hugetlb_temporary(folio);
2449 
2450 	return folio;
2451 }
2452 
2453 /*
2454  * Use the VMA's mpolicy to allocate a huge page from the buddy.
2455  */
2456 static
2457 struct folio *alloc_buddy_hugetlb_folio_with_mpol(struct hstate *h,
2458 		struct vm_area_struct *vma, unsigned long addr)
2459 {
2460 	struct folio *folio = NULL;
2461 	struct mempolicy *mpol;
2462 	gfp_t gfp_mask = htlb_alloc_mask(h);
2463 	int nid;
2464 	nodemask_t *nodemask;
2465 
2466 	nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2467 	if (mpol_is_preferred_many(mpol)) {
2468 		gfp_t gfp = gfp_mask | __GFP_NOWARN;
2469 
2470 		gfp &=  ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2471 		folio = alloc_surplus_hugetlb_folio(h, gfp, nid, nodemask);
2472 
2473 		/* Fallback to all nodes if page==NULL */
2474 		nodemask = NULL;
2475 	}
2476 
2477 	if (!folio)
2478 		folio = alloc_surplus_hugetlb_folio(h, gfp_mask, nid, nodemask);
2479 	mpol_cond_put(mpol);
2480 	return folio;
2481 }
2482 
2483 /* folio migration callback function */
2484 struct folio *alloc_hugetlb_folio_nodemask(struct hstate *h, int preferred_nid,
2485 		nodemask_t *nmask, gfp_t gfp_mask)
2486 {
2487 	spin_lock_irq(&hugetlb_lock);
2488 	if (available_huge_pages(h)) {
2489 		struct folio *folio;
2490 
2491 		folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
2492 						preferred_nid, nmask);
2493 		if (folio) {
2494 			spin_unlock_irq(&hugetlb_lock);
2495 			return folio;
2496 		}
2497 	}
2498 	spin_unlock_irq(&hugetlb_lock);
2499 
2500 	return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask);
2501 }
2502 
2503 /* mempolicy aware migration callback */
2504 struct folio *alloc_hugetlb_folio_vma(struct hstate *h, struct vm_area_struct *vma,
2505 		unsigned long address)
2506 {
2507 	struct mempolicy *mpol;
2508 	nodemask_t *nodemask;
2509 	struct folio *folio;
2510 	gfp_t gfp_mask;
2511 	int node;
2512 
2513 	gfp_mask = htlb_alloc_mask(h);
2514 	node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2515 	folio = alloc_hugetlb_folio_nodemask(h, node, nodemask, gfp_mask);
2516 	mpol_cond_put(mpol);
2517 
2518 	return folio;
2519 }
2520 
2521 static nodemask_t *policy_mbind_nodemask(gfp_t gfp)
2522 {
2523 #ifdef CONFIG_NUMA
2524 	struct mempolicy *mpol = get_task_policy(current);
2525 
2526 	/*
2527 	 * Only enforce MPOL_BIND policy which overlaps with cpuset policy
2528 	 * (from policy_nodemask) specifically for hugetlb case
2529 	 */
2530 	if (mpol->mode == MPOL_BIND &&
2531 		(apply_policy_zone(mpol, gfp_zone(gfp)) &&
2532 		 cpuset_nodemask_valid_mems_allowed(&mpol->nodes)))
2533 		return &mpol->nodes;
2534 #endif
2535 	return NULL;
2536 }
2537 
2538 /*
2539  * Increase the hugetlb pool such that it can accommodate a reservation
2540  * of size 'delta'.
2541  */
2542 static int gather_surplus_pages(struct hstate *h, long delta)
2543 	__must_hold(&hugetlb_lock)
2544 {
2545 	LIST_HEAD(surplus_list);
2546 	struct folio *folio, *tmp;
2547 	int ret;
2548 	long i;
2549 	long needed, allocated;
2550 	bool alloc_ok = true;
2551 	int node;
2552 	nodemask_t *mbind_nodemask = policy_mbind_nodemask(htlb_alloc_mask(h));
2553 
2554 	lockdep_assert_held(&hugetlb_lock);
2555 	needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2556 	if (needed <= 0) {
2557 		h->resv_huge_pages += delta;
2558 		return 0;
2559 	}
2560 
2561 	allocated = 0;
2562 
2563 	ret = -ENOMEM;
2564 retry:
2565 	spin_unlock_irq(&hugetlb_lock);
2566 	for (i = 0; i < needed; i++) {
2567 		folio = NULL;
2568 		for_each_node_mask(node, cpuset_current_mems_allowed) {
2569 			if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) {
2570 				folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h),
2571 						node, NULL);
2572 				if (folio)
2573 					break;
2574 			}
2575 		}
2576 		if (!folio) {
2577 			alloc_ok = false;
2578 			break;
2579 		}
2580 		list_add(&folio->lru, &surplus_list);
2581 		cond_resched();
2582 	}
2583 	allocated += i;
2584 
2585 	/*
2586 	 * After retaking hugetlb_lock, we need to recalculate 'needed'
2587 	 * because either resv_huge_pages or free_huge_pages may have changed.
2588 	 */
2589 	spin_lock_irq(&hugetlb_lock);
2590 	needed = (h->resv_huge_pages + delta) -
2591 			(h->free_huge_pages + allocated);
2592 	if (needed > 0) {
2593 		if (alloc_ok)
2594 			goto retry;
2595 		/*
2596 		 * We were not able to allocate enough pages to
2597 		 * satisfy the entire reservation so we free what
2598 		 * we've allocated so far.
2599 		 */
2600 		goto free;
2601 	}
2602 	/*
2603 	 * The surplus_list now contains _at_least_ the number of extra pages
2604 	 * needed to accommodate the reservation.  Add the appropriate number
2605 	 * of pages to the hugetlb pool and free the extras back to the buddy
2606 	 * allocator.  Commit the entire reservation here to prevent another
2607 	 * process from stealing the pages as they are added to the pool but
2608 	 * before they are reserved.
2609 	 */
2610 	needed += allocated;
2611 	h->resv_huge_pages += delta;
2612 	ret = 0;
2613 
2614 	/* Free the needed pages to the hugetlb pool */
2615 	list_for_each_entry_safe(folio, tmp, &surplus_list, lru) {
2616 		if ((--needed) < 0)
2617 			break;
2618 		/* Add the page to the hugetlb allocator */
2619 		enqueue_hugetlb_folio(h, folio);
2620 	}
2621 free:
2622 	spin_unlock_irq(&hugetlb_lock);
2623 
2624 	/*
2625 	 * Free unnecessary surplus pages to the buddy allocator.
2626 	 * Pages have no ref count, call free_huge_folio directly.
2627 	 */
2628 	list_for_each_entry_safe(folio, tmp, &surplus_list, lru)
2629 		free_huge_folio(folio);
2630 	spin_lock_irq(&hugetlb_lock);
2631 
2632 	return ret;
2633 }
2634 
2635 /*
2636  * This routine has two main purposes:
2637  * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2638  *    in unused_resv_pages.  This corresponds to the prior adjustments made
2639  *    to the associated reservation map.
2640  * 2) Free any unused surplus pages that may have been allocated to satisfy
2641  *    the reservation.  As many as unused_resv_pages may be freed.
2642  */
2643 static void return_unused_surplus_pages(struct hstate *h,
2644 					unsigned long unused_resv_pages)
2645 {
2646 	unsigned long nr_pages;
2647 	struct page *page;
2648 	LIST_HEAD(page_list);
2649 
2650 	lockdep_assert_held(&hugetlb_lock);
2651 	/* Uncommit the reservation */
2652 	h->resv_huge_pages -= unused_resv_pages;
2653 
2654 	if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2655 		goto out;
2656 
2657 	/*
2658 	 * Part (or even all) of the reservation could have been backed
2659 	 * by pre-allocated pages. Only free surplus pages.
2660 	 */
2661 	nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2662 
2663 	/*
2664 	 * We want to release as many surplus pages as possible, spread
2665 	 * evenly across all nodes with memory. Iterate across these nodes
2666 	 * until we can no longer free unreserved surplus pages. This occurs
2667 	 * when the nodes with surplus pages have no free pages.
2668 	 * remove_pool_huge_page() will balance the freed pages across the
2669 	 * on-line nodes with memory and will handle the hstate accounting.
2670 	 */
2671 	while (nr_pages--) {
2672 		page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2673 		if (!page)
2674 			goto out;
2675 
2676 		list_add(&page->lru, &page_list);
2677 	}
2678 
2679 out:
2680 	spin_unlock_irq(&hugetlb_lock);
2681 	update_and_free_pages_bulk(h, &page_list);
2682 	spin_lock_irq(&hugetlb_lock);
2683 }
2684 
2685 
2686 /*
2687  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2688  * are used by the huge page allocation routines to manage reservations.
2689  *
2690  * vma_needs_reservation is called to determine if the huge page at addr
2691  * within the vma has an associated reservation.  If a reservation is
2692  * needed, the value 1 is returned.  The caller is then responsible for
2693  * managing the global reservation and subpool usage counts.  After
2694  * the huge page has been allocated, vma_commit_reservation is called
2695  * to add the page to the reservation map.  If the page allocation fails,
2696  * the reservation must be ended instead of committed.  vma_end_reservation
2697  * is called in such cases.
2698  *
2699  * In the normal case, vma_commit_reservation returns the same value
2700  * as the preceding vma_needs_reservation call.  The only time this
2701  * is not the case is if a reserve map was changed between calls.  It
2702  * is the responsibility of the caller to notice the difference and
2703  * take appropriate action.
2704  *
2705  * vma_add_reservation is used in error paths where a reservation must
2706  * be restored when a newly allocated huge page must be freed.  It is
2707  * to be called after calling vma_needs_reservation to determine if a
2708  * reservation exists.
2709  *
2710  * vma_del_reservation is used in error paths where an entry in the reserve
2711  * map was created during huge page allocation and must be removed.  It is to
2712  * be called after calling vma_needs_reservation to determine if a reservation
2713  * exists.
2714  */
2715 enum vma_resv_mode {
2716 	VMA_NEEDS_RESV,
2717 	VMA_COMMIT_RESV,
2718 	VMA_END_RESV,
2719 	VMA_ADD_RESV,
2720 	VMA_DEL_RESV,
2721 };
2722 static long __vma_reservation_common(struct hstate *h,
2723 				struct vm_area_struct *vma, unsigned long addr,
2724 				enum vma_resv_mode mode)
2725 {
2726 	struct resv_map *resv;
2727 	pgoff_t idx;
2728 	long ret;
2729 	long dummy_out_regions_needed;
2730 
2731 	resv = vma_resv_map(vma);
2732 	if (!resv)
2733 		return 1;
2734 
2735 	idx = vma_hugecache_offset(h, vma, addr);
2736 	switch (mode) {
2737 	case VMA_NEEDS_RESV:
2738 		ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2739 		/* We assume that vma_reservation_* routines always operate on
2740 		 * 1 page, and that adding to resv map a 1 page entry can only
2741 		 * ever require 1 region.
2742 		 */
2743 		VM_BUG_ON(dummy_out_regions_needed != 1);
2744 		break;
2745 	case VMA_COMMIT_RESV:
2746 		ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2747 		/* region_add calls of range 1 should never fail. */
2748 		VM_BUG_ON(ret < 0);
2749 		break;
2750 	case VMA_END_RESV:
2751 		region_abort(resv, idx, idx + 1, 1);
2752 		ret = 0;
2753 		break;
2754 	case VMA_ADD_RESV:
2755 		if (vma->vm_flags & VM_MAYSHARE) {
2756 			ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2757 			/* region_add calls of range 1 should never fail. */
2758 			VM_BUG_ON(ret < 0);
2759 		} else {
2760 			region_abort(resv, idx, idx + 1, 1);
2761 			ret = region_del(resv, idx, idx + 1);
2762 		}
2763 		break;
2764 	case VMA_DEL_RESV:
2765 		if (vma->vm_flags & VM_MAYSHARE) {
2766 			region_abort(resv, idx, idx + 1, 1);
2767 			ret = region_del(resv, idx, idx + 1);
2768 		} else {
2769 			ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2770 			/* region_add calls of range 1 should never fail. */
2771 			VM_BUG_ON(ret < 0);
2772 		}
2773 		break;
2774 	default:
2775 		BUG();
2776 	}
2777 
2778 	if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2779 		return ret;
2780 	/*
2781 	 * We know private mapping must have HPAGE_RESV_OWNER set.
2782 	 *
2783 	 * In most cases, reserves always exist for private mappings.
2784 	 * However, a file associated with mapping could have been
2785 	 * hole punched or truncated after reserves were consumed.
2786 	 * As subsequent fault on such a range will not use reserves.
2787 	 * Subtle - The reserve map for private mappings has the
2788 	 * opposite meaning than that of shared mappings.  If NO
2789 	 * entry is in the reserve map, it means a reservation exists.
2790 	 * If an entry exists in the reserve map, it means the
2791 	 * reservation has already been consumed.  As a result, the
2792 	 * return value of this routine is the opposite of the
2793 	 * value returned from reserve map manipulation routines above.
2794 	 */
2795 	if (ret > 0)
2796 		return 0;
2797 	if (ret == 0)
2798 		return 1;
2799 	return ret;
2800 }
2801 
2802 static long vma_needs_reservation(struct hstate *h,
2803 			struct vm_area_struct *vma, unsigned long addr)
2804 {
2805 	return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2806 }
2807 
2808 static long vma_commit_reservation(struct hstate *h,
2809 			struct vm_area_struct *vma, unsigned long addr)
2810 {
2811 	return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2812 }
2813 
2814 static void vma_end_reservation(struct hstate *h,
2815 			struct vm_area_struct *vma, unsigned long addr)
2816 {
2817 	(void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2818 }
2819 
2820 static long vma_add_reservation(struct hstate *h,
2821 			struct vm_area_struct *vma, unsigned long addr)
2822 {
2823 	return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2824 }
2825 
2826 static long vma_del_reservation(struct hstate *h,
2827 			struct vm_area_struct *vma, unsigned long addr)
2828 {
2829 	return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2830 }
2831 
2832 /*
2833  * This routine is called to restore reservation information on error paths.
2834  * It should ONLY be called for folios allocated via alloc_hugetlb_folio(),
2835  * and the hugetlb mutex should remain held when calling this routine.
2836  *
2837  * It handles two specific cases:
2838  * 1) A reservation was in place and the folio consumed the reservation.
2839  *    hugetlb_restore_reserve is set in the folio.
2840  * 2) No reservation was in place for the page, so hugetlb_restore_reserve is
2841  *    not set.  However, alloc_hugetlb_folio always updates the reserve map.
2842  *
2843  * In case 1, free_huge_folio later in the error path will increment the
2844  * global reserve count.  But, free_huge_folio does not have enough context
2845  * to adjust the reservation map.  This case deals primarily with private
2846  * mappings.  Adjust the reserve map here to be consistent with global
2847  * reserve count adjustments to be made by free_huge_folio.  Make sure the
2848  * reserve map indicates there is a reservation present.
2849  *
2850  * In case 2, simply undo reserve map modifications done by alloc_hugetlb_folio.
2851  */
2852 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2853 			unsigned long address, struct folio *folio)
2854 {
2855 	long rc = vma_needs_reservation(h, vma, address);
2856 
2857 	if (folio_test_hugetlb_restore_reserve(folio)) {
2858 		if (unlikely(rc < 0))
2859 			/*
2860 			 * Rare out of memory condition in reserve map
2861 			 * manipulation.  Clear hugetlb_restore_reserve so
2862 			 * that global reserve count will not be incremented
2863 			 * by free_huge_folio.  This will make it appear
2864 			 * as though the reservation for this folio was
2865 			 * consumed.  This may prevent the task from
2866 			 * faulting in the folio at a later time.  This
2867 			 * is better than inconsistent global huge page
2868 			 * accounting of reserve counts.
2869 			 */
2870 			folio_clear_hugetlb_restore_reserve(folio);
2871 		else if (rc)
2872 			(void)vma_add_reservation(h, vma, address);
2873 		else
2874 			vma_end_reservation(h, vma, address);
2875 	} else {
2876 		if (!rc) {
2877 			/*
2878 			 * This indicates there is an entry in the reserve map
2879 			 * not added by alloc_hugetlb_folio.  We know it was added
2880 			 * before the alloc_hugetlb_folio call, otherwise
2881 			 * hugetlb_restore_reserve would be set on the folio.
2882 			 * Remove the entry so that a subsequent allocation
2883 			 * does not consume a reservation.
2884 			 */
2885 			rc = vma_del_reservation(h, vma, address);
2886 			if (rc < 0)
2887 				/*
2888 				 * VERY rare out of memory condition.  Since
2889 				 * we can not delete the entry, set
2890 				 * hugetlb_restore_reserve so that the reserve
2891 				 * count will be incremented when the folio
2892 				 * is freed.  This reserve will be consumed
2893 				 * on a subsequent allocation.
2894 				 */
2895 				folio_set_hugetlb_restore_reserve(folio);
2896 		} else if (rc < 0) {
2897 			/*
2898 			 * Rare out of memory condition from
2899 			 * vma_needs_reservation call.  Memory allocation is
2900 			 * only attempted if a new entry is needed.  Therefore,
2901 			 * this implies there is not an entry in the
2902 			 * reserve map.
2903 			 *
2904 			 * For shared mappings, no entry in the map indicates
2905 			 * no reservation.  We are done.
2906 			 */
2907 			if (!(vma->vm_flags & VM_MAYSHARE))
2908 				/*
2909 				 * For private mappings, no entry indicates
2910 				 * a reservation is present.  Since we can
2911 				 * not add an entry, set hugetlb_restore_reserve
2912 				 * on the folio so reserve count will be
2913 				 * incremented when freed.  This reserve will
2914 				 * be consumed on a subsequent allocation.
2915 				 */
2916 				folio_set_hugetlb_restore_reserve(folio);
2917 		} else
2918 			/*
2919 			 * No reservation present, do nothing
2920 			 */
2921 			 vma_end_reservation(h, vma, address);
2922 	}
2923 }
2924 
2925 /*
2926  * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve
2927  * the old one
2928  * @h: struct hstate old page belongs to
2929  * @old_folio: Old folio to dissolve
2930  * @list: List to isolate the page in case we need to
2931  * Returns 0 on success, otherwise negated error.
2932  */
2933 static int alloc_and_dissolve_hugetlb_folio(struct hstate *h,
2934 			struct folio *old_folio, struct list_head *list)
2935 {
2936 	gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2937 	int nid = folio_nid(old_folio);
2938 	struct folio *new_folio;
2939 	int ret = 0;
2940 
2941 	/*
2942 	 * Before dissolving the folio, we need to allocate a new one for the
2943 	 * pool to remain stable.  Here, we allocate the folio and 'prep' it
2944 	 * by doing everything but actually updating counters and adding to
2945 	 * the pool.  This simplifies and let us do most of the processing
2946 	 * under the lock.
2947 	 */
2948 	new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, NULL, NULL);
2949 	if (!new_folio)
2950 		return -ENOMEM;
2951 	__prep_new_hugetlb_folio(h, new_folio);
2952 
2953 retry:
2954 	spin_lock_irq(&hugetlb_lock);
2955 	if (!folio_test_hugetlb(old_folio)) {
2956 		/*
2957 		 * Freed from under us. Drop new_folio too.
2958 		 */
2959 		goto free_new;
2960 	} else if (folio_ref_count(old_folio)) {
2961 		bool isolated;
2962 
2963 		/*
2964 		 * Someone has grabbed the folio, try to isolate it here.
2965 		 * Fail with -EBUSY if not possible.
2966 		 */
2967 		spin_unlock_irq(&hugetlb_lock);
2968 		isolated = isolate_hugetlb(old_folio, list);
2969 		ret = isolated ? 0 : -EBUSY;
2970 		spin_lock_irq(&hugetlb_lock);
2971 		goto free_new;
2972 	} else if (!folio_test_hugetlb_freed(old_folio)) {
2973 		/*
2974 		 * Folio's refcount is 0 but it has not been enqueued in the
2975 		 * freelist yet. Race window is small, so we can succeed here if
2976 		 * we retry.
2977 		 */
2978 		spin_unlock_irq(&hugetlb_lock);
2979 		cond_resched();
2980 		goto retry;
2981 	} else {
2982 		/*
2983 		 * Ok, old_folio is still a genuine free hugepage. Remove it from
2984 		 * the freelist and decrease the counters. These will be
2985 		 * incremented again when calling __prep_account_new_huge_page()
2986 		 * and enqueue_hugetlb_folio() for new_folio. The counters will
2987 		 * remain stable since this happens under the lock.
2988 		 */
2989 		remove_hugetlb_folio(h, old_folio, false);
2990 
2991 		/*
2992 		 * Ref count on new_folio is already zero as it was dropped
2993 		 * earlier.  It can be directly added to the pool free list.
2994 		 */
2995 		__prep_account_new_huge_page(h, nid);
2996 		enqueue_hugetlb_folio(h, new_folio);
2997 
2998 		/*
2999 		 * Folio has been replaced, we can safely free the old one.
3000 		 */
3001 		spin_unlock_irq(&hugetlb_lock);
3002 		update_and_free_hugetlb_folio(h, old_folio, false);
3003 	}
3004 
3005 	return ret;
3006 
3007 free_new:
3008 	spin_unlock_irq(&hugetlb_lock);
3009 	/* Folio has a zero ref count, but needs a ref to be freed */
3010 	folio_ref_unfreeze(new_folio, 1);
3011 	update_and_free_hugetlb_folio(h, new_folio, false);
3012 
3013 	return ret;
3014 }
3015 
3016 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
3017 {
3018 	struct hstate *h;
3019 	struct folio *folio = page_folio(page);
3020 	int ret = -EBUSY;
3021 
3022 	/*
3023 	 * The page might have been dissolved from under our feet, so make sure
3024 	 * to carefully check the state under the lock.
3025 	 * Return success when racing as if we dissolved the page ourselves.
3026 	 */
3027 	spin_lock_irq(&hugetlb_lock);
3028 	if (folio_test_hugetlb(folio)) {
3029 		h = folio_hstate(folio);
3030 	} else {
3031 		spin_unlock_irq(&hugetlb_lock);
3032 		return 0;
3033 	}
3034 	spin_unlock_irq(&hugetlb_lock);
3035 
3036 	/*
3037 	 * Fence off gigantic pages as there is a cyclic dependency between
3038 	 * alloc_contig_range and them. Return -ENOMEM as this has the effect
3039 	 * of bailing out right away without further retrying.
3040 	 */
3041 	if (hstate_is_gigantic(h))
3042 		return -ENOMEM;
3043 
3044 	if (folio_ref_count(folio) && isolate_hugetlb(folio, list))
3045 		ret = 0;
3046 	else if (!folio_ref_count(folio))
3047 		ret = alloc_and_dissolve_hugetlb_folio(h, folio, list);
3048 
3049 	return ret;
3050 }
3051 
3052 struct folio *alloc_hugetlb_folio(struct vm_area_struct *vma,
3053 				    unsigned long addr, int avoid_reserve)
3054 {
3055 	struct hugepage_subpool *spool = subpool_vma(vma);
3056 	struct hstate *h = hstate_vma(vma);
3057 	struct folio *folio;
3058 	long map_chg, map_commit;
3059 	long gbl_chg;
3060 	int ret, idx;
3061 	struct hugetlb_cgroup *h_cg = NULL;
3062 	bool deferred_reserve;
3063 
3064 	idx = hstate_index(h);
3065 	/*
3066 	 * Examine the region/reserve map to determine if the process
3067 	 * has a reservation for the page to be allocated.  A return
3068 	 * code of zero indicates a reservation exists (no change).
3069 	 */
3070 	map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
3071 	if (map_chg < 0)
3072 		return ERR_PTR(-ENOMEM);
3073 
3074 	/*
3075 	 * Processes that did not create the mapping will have no
3076 	 * reserves as indicated by the region/reserve map. Check
3077 	 * that the allocation will not exceed the subpool limit.
3078 	 * Allocations for MAP_NORESERVE mappings also need to be
3079 	 * checked against any subpool limit.
3080 	 */
3081 	if (map_chg || avoid_reserve) {
3082 		gbl_chg = hugepage_subpool_get_pages(spool, 1);
3083 		if (gbl_chg < 0) {
3084 			vma_end_reservation(h, vma, addr);
3085 			return ERR_PTR(-ENOSPC);
3086 		}
3087 
3088 		/*
3089 		 * Even though there was no reservation in the region/reserve
3090 		 * map, there could be reservations associated with the
3091 		 * subpool that can be used.  This would be indicated if the
3092 		 * return value of hugepage_subpool_get_pages() is zero.
3093 		 * However, if avoid_reserve is specified we still avoid even
3094 		 * the subpool reservations.
3095 		 */
3096 		if (avoid_reserve)
3097 			gbl_chg = 1;
3098 	}
3099 
3100 	/* If this allocation is not consuming a reservation, charge it now.
3101 	 */
3102 	deferred_reserve = map_chg || avoid_reserve;
3103 	if (deferred_reserve) {
3104 		ret = hugetlb_cgroup_charge_cgroup_rsvd(
3105 			idx, pages_per_huge_page(h), &h_cg);
3106 		if (ret)
3107 			goto out_subpool_put;
3108 	}
3109 
3110 	ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
3111 	if (ret)
3112 		goto out_uncharge_cgroup_reservation;
3113 
3114 	spin_lock_irq(&hugetlb_lock);
3115 	/*
3116 	 * glb_chg is passed to indicate whether or not a page must be taken
3117 	 * from the global free pool (global change).  gbl_chg == 0 indicates
3118 	 * a reservation exists for the allocation.
3119 	 */
3120 	folio = dequeue_hugetlb_folio_vma(h, vma, addr, avoid_reserve, gbl_chg);
3121 	if (!folio) {
3122 		spin_unlock_irq(&hugetlb_lock);
3123 		folio = alloc_buddy_hugetlb_folio_with_mpol(h, vma, addr);
3124 		if (!folio)
3125 			goto out_uncharge_cgroup;
3126 		spin_lock_irq(&hugetlb_lock);
3127 		if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
3128 			folio_set_hugetlb_restore_reserve(folio);
3129 			h->resv_huge_pages--;
3130 		}
3131 		list_add(&folio->lru, &h->hugepage_activelist);
3132 		folio_ref_unfreeze(folio, 1);
3133 		/* Fall through */
3134 	}
3135 
3136 	hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, folio);
3137 	/* If allocation is not consuming a reservation, also store the
3138 	 * hugetlb_cgroup pointer on the page.
3139 	 */
3140 	if (deferred_reserve) {
3141 		hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
3142 						  h_cg, folio);
3143 	}
3144 
3145 	spin_unlock_irq(&hugetlb_lock);
3146 
3147 	hugetlb_set_folio_subpool(folio, spool);
3148 
3149 	map_commit = vma_commit_reservation(h, vma, addr);
3150 	if (unlikely(map_chg > map_commit)) {
3151 		/*
3152 		 * The page was added to the reservation map between
3153 		 * vma_needs_reservation and vma_commit_reservation.
3154 		 * This indicates a race with hugetlb_reserve_pages.
3155 		 * Adjust for the subpool count incremented above AND
3156 		 * in hugetlb_reserve_pages for the same page.  Also,
3157 		 * the reservation count added in hugetlb_reserve_pages
3158 		 * no longer applies.
3159 		 */
3160 		long rsv_adjust;
3161 
3162 		rsv_adjust = hugepage_subpool_put_pages(spool, 1);
3163 		hugetlb_acct_memory(h, -rsv_adjust);
3164 		if (deferred_reserve) {
3165 			spin_lock_irq(&hugetlb_lock);
3166 			hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
3167 					pages_per_huge_page(h), folio);
3168 			spin_unlock_irq(&hugetlb_lock);
3169 		}
3170 	}
3171 	return folio;
3172 
3173 out_uncharge_cgroup:
3174 	hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
3175 out_uncharge_cgroup_reservation:
3176 	if (deferred_reserve)
3177 		hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
3178 						    h_cg);
3179 out_subpool_put:
3180 	if (map_chg || avoid_reserve)
3181 		hugepage_subpool_put_pages(spool, 1);
3182 	vma_end_reservation(h, vma, addr);
3183 	return ERR_PTR(-ENOSPC);
3184 }
3185 
3186 int alloc_bootmem_huge_page(struct hstate *h, int nid)
3187 	__attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
3188 int __alloc_bootmem_huge_page(struct hstate *h, int nid)
3189 {
3190 	struct huge_bootmem_page *m = NULL; /* initialize for clang */
3191 	int nr_nodes, node;
3192 
3193 	/* do node specific alloc */
3194 	if (nid != NUMA_NO_NODE) {
3195 		m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h),
3196 				0, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
3197 		if (!m)
3198 			return 0;
3199 		goto found;
3200 	}
3201 	/* allocate from next node when distributing huge pages */
3202 	for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
3203 		m = memblock_alloc_try_nid_raw(
3204 				huge_page_size(h), huge_page_size(h),
3205 				0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
3206 		/*
3207 		 * Use the beginning of the huge page to store the
3208 		 * huge_bootmem_page struct (until gather_bootmem
3209 		 * puts them into the mem_map).
3210 		 */
3211 		if (!m)
3212 			return 0;
3213 		goto found;
3214 	}
3215 
3216 found:
3217 	/* Put them into a private list first because mem_map is not up yet */
3218 	INIT_LIST_HEAD(&m->list);
3219 	list_add(&m->list, &huge_boot_pages);
3220 	m->hstate = h;
3221 	return 1;
3222 }
3223 
3224 /*
3225  * Put bootmem huge pages into the standard lists after mem_map is up.
3226  * Note: This only applies to gigantic (order > MAX_ORDER) pages.
3227  */
3228 static void __init gather_bootmem_prealloc(void)
3229 {
3230 	struct huge_bootmem_page *m;
3231 
3232 	list_for_each_entry(m, &huge_boot_pages, list) {
3233 		struct page *page = virt_to_page(m);
3234 		struct folio *folio = page_folio(page);
3235 		struct hstate *h = m->hstate;
3236 
3237 		VM_BUG_ON(!hstate_is_gigantic(h));
3238 		WARN_ON(folio_ref_count(folio) != 1);
3239 		if (prep_compound_gigantic_folio(folio, huge_page_order(h))) {
3240 			WARN_ON(folio_test_reserved(folio));
3241 			prep_new_hugetlb_folio(h, folio, folio_nid(folio));
3242 			free_huge_folio(folio); /* add to the hugepage allocator */
3243 		} else {
3244 			/* VERY unlikely inflated ref count on a tail page */
3245 			free_gigantic_folio(folio, huge_page_order(h));
3246 		}
3247 
3248 		/*
3249 		 * We need to restore the 'stolen' pages to totalram_pages
3250 		 * in order to fix confusing memory reports from free(1) and
3251 		 * other side-effects, like CommitLimit going negative.
3252 		 */
3253 		adjust_managed_page_count(page, pages_per_huge_page(h));
3254 		cond_resched();
3255 	}
3256 }
3257 static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)
3258 {
3259 	unsigned long i;
3260 	char buf[32];
3261 
3262 	for (i = 0; i < h->max_huge_pages_node[nid]; ++i) {
3263 		if (hstate_is_gigantic(h)) {
3264 			if (!alloc_bootmem_huge_page(h, nid))
3265 				break;
3266 		} else {
3267 			struct folio *folio;
3268 			gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3269 
3270 			folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid,
3271 					&node_states[N_MEMORY], NULL);
3272 			if (!folio)
3273 				break;
3274 			free_huge_folio(folio); /* free it into the hugepage allocator */
3275 		}
3276 		cond_resched();
3277 	}
3278 	if (i == h->max_huge_pages_node[nid])
3279 		return;
3280 
3281 	string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3282 	pr_warn("HugeTLB: allocating %u of page size %s failed node%d.  Only allocated %lu hugepages.\n",
3283 		h->max_huge_pages_node[nid], buf, nid, i);
3284 	h->max_huge_pages -= (h->max_huge_pages_node[nid] - i);
3285 	h->max_huge_pages_node[nid] = i;
3286 }
3287 
3288 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
3289 {
3290 	unsigned long i;
3291 	nodemask_t *node_alloc_noretry;
3292 	bool node_specific_alloc = false;
3293 
3294 	/* skip gigantic hugepages allocation if hugetlb_cma enabled */
3295 	if (hstate_is_gigantic(h) && hugetlb_cma_size) {
3296 		pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
3297 		return;
3298 	}
3299 
3300 	/* do node specific alloc */
3301 	for_each_online_node(i) {
3302 		if (h->max_huge_pages_node[i] > 0) {
3303 			hugetlb_hstate_alloc_pages_onenode(h, i);
3304 			node_specific_alloc = true;
3305 		}
3306 	}
3307 
3308 	if (node_specific_alloc)
3309 		return;
3310 
3311 	/* below will do all node balanced alloc */
3312 	if (!hstate_is_gigantic(h)) {
3313 		/*
3314 		 * Bit mask controlling how hard we retry per-node allocations.
3315 		 * Ignore errors as lower level routines can deal with
3316 		 * node_alloc_noretry == NULL.  If this kmalloc fails at boot
3317 		 * time, we are likely in bigger trouble.
3318 		 */
3319 		node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
3320 						GFP_KERNEL);
3321 	} else {
3322 		/* allocations done at boot time */
3323 		node_alloc_noretry = NULL;
3324 	}
3325 
3326 	/* bit mask controlling how hard we retry per-node allocations */
3327 	if (node_alloc_noretry)
3328 		nodes_clear(*node_alloc_noretry);
3329 
3330 	for (i = 0; i < h->max_huge_pages; ++i) {
3331 		if (hstate_is_gigantic(h)) {
3332 			if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))
3333 				break;
3334 		} else if (!alloc_pool_huge_page(h,
3335 					 &node_states[N_MEMORY],
3336 					 node_alloc_noretry))
3337 			break;
3338 		cond_resched();
3339 	}
3340 	if (i < h->max_huge_pages) {
3341 		char buf[32];
3342 
3343 		string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3344 		pr_warn("HugeTLB: allocating %lu of page size %s failed.  Only allocated %lu hugepages.\n",
3345 			h->max_huge_pages, buf, i);
3346 		h->max_huge_pages = i;
3347 	}
3348 	kfree(node_alloc_noretry);
3349 }
3350 
3351 static void __init hugetlb_init_hstates(void)
3352 {
3353 	struct hstate *h, *h2;
3354 
3355 	for_each_hstate(h) {
3356 		/* oversize hugepages were init'ed in early boot */
3357 		if (!hstate_is_gigantic(h))
3358 			hugetlb_hstate_alloc_pages(h);
3359 
3360 		/*
3361 		 * Set demote order for each hstate.  Note that
3362 		 * h->demote_order is initially 0.
3363 		 * - We can not demote gigantic pages if runtime freeing
3364 		 *   is not supported, so skip this.
3365 		 * - If CMA allocation is possible, we can not demote
3366 		 *   HUGETLB_PAGE_ORDER or smaller size pages.
3367 		 */
3368 		if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3369 			continue;
3370 		if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER)
3371 			continue;
3372 		for_each_hstate(h2) {
3373 			if (h2 == h)
3374 				continue;
3375 			if (h2->order < h->order &&
3376 			    h2->order > h->demote_order)
3377 				h->demote_order = h2->order;
3378 		}
3379 	}
3380 }
3381 
3382 static void __init report_hugepages(void)
3383 {
3384 	struct hstate *h;
3385 
3386 	for_each_hstate(h) {
3387 		char buf[32];
3388 
3389 		string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3390 		pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n",
3391 			buf, h->free_huge_pages);
3392 		pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n",
3393 			hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf);
3394 	}
3395 }
3396 
3397 #ifdef CONFIG_HIGHMEM
3398 static void try_to_free_low(struct hstate *h, unsigned long count,
3399 						nodemask_t *nodes_allowed)
3400 {
3401 	int i;
3402 	LIST_HEAD(page_list);
3403 
3404 	lockdep_assert_held(&hugetlb_lock);
3405 	if (hstate_is_gigantic(h))
3406 		return;
3407 
3408 	/*
3409 	 * Collect pages to be freed on a list, and free after dropping lock
3410 	 */
3411 	for_each_node_mask(i, *nodes_allowed) {
3412 		struct page *page, *next;
3413 		struct list_head *freel = &h->hugepage_freelists[i];
3414 		list_for_each_entry_safe(page, next, freel, lru) {
3415 			if (count >= h->nr_huge_pages)
3416 				goto out;
3417 			if (PageHighMem(page))
3418 				continue;
3419 			remove_hugetlb_folio(h, page_folio(page), false);
3420 			list_add(&page->lru, &page_list);
3421 		}
3422 	}
3423 
3424 out:
3425 	spin_unlock_irq(&hugetlb_lock);
3426 	update_and_free_pages_bulk(h, &page_list);
3427 	spin_lock_irq(&hugetlb_lock);
3428 }
3429 #else
3430 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3431 						nodemask_t *nodes_allowed)
3432 {
3433 }
3434 #endif
3435 
3436 /*
3437  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
3438  * balanced by operating on them in a round-robin fashion.
3439  * Returns 1 if an adjustment was made.
3440  */
3441 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3442 				int delta)
3443 {
3444 	int nr_nodes, node;
3445 
3446 	lockdep_assert_held(&hugetlb_lock);
3447 	VM_BUG_ON(delta != -1 && delta != 1);
3448 
3449 	if (delta < 0) {
3450 		for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3451 			if (h->surplus_huge_pages_node[node])
3452 				goto found;
3453 		}
3454 	} else {
3455 		for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3456 			if (h->surplus_huge_pages_node[node] <
3457 					h->nr_huge_pages_node[node])
3458 				goto found;
3459 		}
3460 	}
3461 	return 0;
3462 
3463 found:
3464 	h->surplus_huge_pages += delta;
3465 	h->surplus_huge_pages_node[node] += delta;
3466 	return 1;
3467 }
3468 
3469 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3470 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3471 			      nodemask_t *nodes_allowed)
3472 {
3473 	unsigned long min_count, ret;
3474 	struct page *page;
3475 	LIST_HEAD(page_list);
3476 	NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3477 
3478 	/*
3479 	 * Bit mask controlling how hard we retry per-node allocations.
3480 	 * If we can not allocate the bit mask, do not attempt to allocate
3481 	 * the requested huge pages.
3482 	 */
3483 	if (node_alloc_noretry)
3484 		nodes_clear(*node_alloc_noretry);
3485 	else
3486 		return -ENOMEM;
3487 
3488 	/*
3489 	 * resize_lock mutex prevents concurrent adjustments to number of
3490 	 * pages in hstate via the proc/sysfs interfaces.
3491 	 */
3492 	mutex_lock(&h->resize_lock);
3493 	flush_free_hpage_work(h);
3494 	spin_lock_irq(&hugetlb_lock);
3495 
3496 	/*
3497 	 * Check for a node specific request.
3498 	 * Changing node specific huge page count may require a corresponding
3499 	 * change to the global count.  In any case, the passed node mask
3500 	 * (nodes_allowed) will restrict alloc/free to the specified node.
3501 	 */
3502 	if (nid != NUMA_NO_NODE) {
3503 		unsigned long old_count = count;
3504 
3505 		count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3506 		/*
3507 		 * User may have specified a large count value which caused the
3508 		 * above calculation to overflow.  In this case, they wanted
3509 		 * to allocate as many huge pages as possible.  Set count to
3510 		 * largest possible value to align with their intention.
3511 		 */
3512 		if (count < old_count)
3513 			count = ULONG_MAX;
3514 	}
3515 
3516 	/*
3517 	 * Gigantic pages runtime allocation depend on the capability for large
3518 	 * page range allocation.
3519 	 * If the system does not provide this feature, return an error when
3520 	 * the user tries to allocate gigantic pages but let the user free the
3521 	 * boottime allocated gigantic pages.
3522 	 */
3523 	if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3524 		if (count > persistent_huge_pages(h)) {
3525 			spin_unlock_irq(&hugetlb_lock);
3526 			mutex_unlock(&h->resize_lock);
3527 			NODEMASK_FREE(node_alloc_noretry);
3528 			return -EINVAL;
3529 		}
3530 		/* Fall through to decrease pool */
3531 	}
3532 
3533 	/*
3534 	 * Increase the pool size
3535 	 * First take pages out of surplus state.  Then make up the
3536 	 * remaining difference by allocating fresh huge pages.
3537 	 *
3538 	 * We might race with alloc_surplus_hugetlb_folio() here and be unable
3539 	 * to convert a surplus huge page to a normal huge page. That is
3540 	 * not critical, though, it just means the overall size of the
3541 	 * pool might be one hugepage larger than it needs to be, but
3542 	 * within all the constraints specified by the sysctls.
3543 	 */
3544 	while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3545 		if (!adjust_pool_surplus(h, nodes_allowed, -1))
3546 			break;
3547 	}
3548 
3549 	while (count > persistent_huge_pages(h)) {
3550 		/*
3551 		 * If this allocation races such that we no longer need the
3552 		 * page, free_huge_folio will handle it by freeing the page
3553 		 * and reducing the surplus.
3554 		 */
3555 		spin_unlock_irq(&hugetlb_lock);
3556 
3557 		/* yield cpu to avoid soft lockup */
3558 		cond_resched();
3559 
3560 		ret = alloc_pool_huge_page(h, nodes_allowed,
3561 						node_alloc_noretry);
3562 		spin_lock_irq(&hugetlb_lock);
3563 		if (!ret)
3564 			goto out;
3565 
3566 		/* Bail for signals. Probably ctrl-c from user */
3567 		if (signal_pending(current))
3568 			goto out;
3569 	}
3570 
3571 	/*
3572 	 * Decrease the pool size
3573 	 * First return free pages to the buddy allocator (being careful
3574 	 * to keep enough around to satisfy reservations).  Then place
3575 	 * pages into surplus state as needed so the pool will shrink
3576 	 * to the desired size as pages become free.
3577 	 *
3578 	 * By placing pages into the surplus state independent of the
3579 	 * overcommit value, we are allowing the surplus pool size to
3580 	 * exceed overcommit. There are few sane options here. Since
3581 	 * alloc_surplus_hugetlb_folio() is checking the global counter,
3582 	 * though, we'll note that we're not allowed to exceed surplus
3583 	 * and won't grow the pool anywhere else. Not until one of the
3584 	 * sysctls are changed, or the surplus pages go out of use.
3585 	 */
3586 	min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3587 	min_count = max(count, min_count);
3588 	try_to_free_low(h, min_count, nodes_allowed);
3589 
3590 	/*
3591 	 * Collect pages to be removed on list without dropping lock
3592 	 */
3593 	while (min_count < persistent_huge_pages(h)) {
3594 		page = remove_pool_huge_page(h, nodes_allowed, 0);
3595 		if (!page)
3596 			break;
3597 
3598 		list_add(&page->lru, &page_list);
3599 	}
3600 	/* free the pages after dropping lock */
3601 	spin_unlock_irq(&hugetlb_lock);
3602 	update_and_free_pages_bulk(h, &page_list);
3603 	flush_free_hpage_work(h);
3604 	spin_lock_irq(&hugetlb_lock);
3605 
3606 	while (count < persistent_huge_pages(h)) {
3607 		if (!adjust_pool_surplus(h, nodes_allowed, 1))
3608 			break;
3609 	}
3610 out:
3611 	h->max_huge_pages = persistent_huge_pages(h);
3612 	spin_unlock_irq(&hugetlb_lock);
3613 	mutex_unlock(&h->resize_lock);
3614 
3615 	NODEMASK_FREE(node_alloc_noretry);
3616 
3617 	return 0;
3618 }
3619 
3620 static int demote_free_hugetlb_folio(struct hstate *h, struct folio *folio)
3621 {
3622 	int i, nid = folio_nid(folio);
3623 	struct hstate *target_hstate;
3624 	struct page *subpage;
3625 	struct folio *inner_folio;
3626 	int rc = 0;
3627 
3628 	target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order);
3629 
3630 	remove_hugetlb_folio_for_demote(h, folio, false);
3631 	spin_unlock_irq(&hugetlb_lock);
3632 
3633 	rc = hugetlb_vmemmap_restore(h, &folio->page);
3634 	if (rc) {
3635 		/* Allocation of vmemmmap failed, we can not demote folio */
3636 		spin_lock_irq(&hugetlb_lock);
3637 		folio_ref_unfreeze(folio, 1);
3638 		add_hugetlb_folio(h, folio, false);
3639 		return rc;
3640 	}
3641 
3642 	/*
3643 	 * Use destroy_compound_hugetlb_folio_for_demote for all huge page
3644 	 * sizes as it will not ref count folios.
3645 	 */
3646 	destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h));
3647 
3648 	/*
3649 	 * Taking target hstate mutex synchronizes with set_max_huge_pages.
3650 	 * Without the mutex, pages added to target hstate could be marked
3651 	 * as surplus.
3652 	 *
3653 	 * Note that we already hold h->resize_lock.  To prevent deadlock,
3654 	 * use the convention of always taking larger size hstate mutex first.
3655 	 */
3656 	mutex_lock(&target_hstate->resize_lock);
3657 	for (i = 0; i < pages_per_huge_page(h);
3658 				i += pages_per_huge_page(target_hstate)) {
3659 		subpage = folio_page(folio, i);
3660 		inner_folio = page_folio(subpage);
3661 		if (hstate_is_gigantic(target_hstate))
3662 			prep_compound_gigantic_folio_for_demote(inner_folio,
3663 							target_hstate->order);
3664 		else
3665 			prep_compound_page(subpage, target_hstate->order);
3666 		folio_change_private(inner_folio, NULL);
3667 		prep_new_hugetlb_folio(target_hstate, inner_folio, nid);
3668 		free_huge_folio(inner_folio);
3669 	}
3670 	mutex_unlock(&target_hstate->resize_lock);
3671 
3672 	spin_lock_irq(&hugetlb_lock);
3673 
3674 	/*
3675 	 * Not absolutely necessary, but for consistency update max_huge_pages
3676 	 * based on pool changes for the demoted page.
3677 	 */
3678 	h->max_huge_pages--;
3679 	target_hstate->max_huge_pages +=
3680 		pages_per_huge_page(h) / pages_per_huge_page(target_hstate);
3681 
3682 	return rc;
3683 }
3684 
3685 static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
3686 	__must_hold(&hugetlb_lock)
3687 {
3688 	int nr_nodes, node;
3689 	struct folio *folio;
3690 
3691 	lockdep_assert_held(&hugetlb_lock);
3692 
3693 	/* We should never get here if no demote order */
3694 	if (!h->demote_order) {
3695 		pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n");
3696 		return -EINVAL;		/* internal error */
3697 	}
3698 
3699 	for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3700 		list_for_each_entry(folio, &h->hugepage_freelists[node], lru) {
3701 			if (folio_test_hwpoison(folio))
3702 				continue;
3703 			return demote_free_hugetlb_folio(h, folio);
3704 		}
3705 	}
3706 
3707 	/*
3708 	 * Only way to get here is if all pages on free lists are poisoned.
3709 	 * Return -EBUSY so that caller will not retry.
3710 	 */
3711 	return -EBUSY;
3712 }
3713 
3714 #define HSTATE_ATTR_RO(_name) \
3715 	static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3716 
3717 #define HSTATE_ATTR_WO(_name) \
3718 	static struct kobj_attribute _name##_attr = __ATTR_WO(_name)
3719 
3720 #define HSTATE_ATTR(_name) \
3721 	static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
3722 
3723 static struct kobject *hugepages_kobj;
3724 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3725 
3726 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3727 
3728 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3729 {
3730 	int i;
3731 
3732 	for (i = 0; i < HUGE_MAX_HSTATE; i++)
3733 		if (hstate_kobjs[i] == kobj) {
3734 			if (nidp)
3735 				*nidp = NUMA_NO_NODE;
3736 			return &hstates[i];
3737 		}
3738 
3739 	return kobj_to_node_hstate(kobj, nidp);
3740 }
3741 
3742 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3743 					struct kobj_attribute *attr, char *buf)
3744 {
3745 	struct hstate *h;
3746 	unsigned long nr_huge_pages;
3747 	int nid;
3748 
3749 	h = kobj_to_hstate(kobj, &nid);
3750 	if (nid == NUMA_NO_NODE)
3751 		nr_huge_pages = h->nr_huge_pages;
3752 	else
3753 		nr_huge_pages = h->nr_huge_pages_node[nid];
3754 
3755 	return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3756 }
3757 
3758 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3759 					   struct hstate *h, int nid,
3760 					   unsigned long count, size_t len)
3761 {
3762 	int err;
3763 	nodemask_t nodes_allowed, *n_mask;
3764 
3765 	if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3766 		return -EINVAL;
3767 
3768 	if (nid == NUMA_NO_NODE) {
3769 		/*
3770 		 * global hstate attribute
3771 		 */
3772 		if (!(obey_mempolicy &&
3773 				init_nodemask_of_mempolicy(&nodes_allowed)))
3774 			n_mask = &node_states[N_MEMORY];
3775 		else
3776 			n_mask = &nodes_allowed;
3777 	} else {
3778 		/*
3779 		 * Node specific request.  count adjustment happens in
3780 		 * set_max_huge_pages() after acquiring hugetlb_lock.
3781 		 */
3782 		init_nodemask_of_node(&nodes_allowed, nid);
3783 		n_mask = &nodes_allowed;
3784 	}
3785 
3786 	err = set_max_huge_pages(h, count, nid, n_mask);
3787 
3788 	return err ? err : len;
3789 }
3790 
3791 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3792 					 struct kobject *kobj, const char *buf,
3793 					 size_t len)
3794 {
3795 	struct hstate *h;
3796 	unsigned long count;
3797 	int nid;
3798 	int err;
3799 
3800 	err = kstrtoul(buf, 10, &count);
3801 	if (err)
3802 		return err;
3803 
3804 	h = kobj_to_hstate(kobj, &nid);
3805 	return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3806 }
3807 
3808 static ssize_t nr_hugepages_show(struct kobject *kobj,
3809 				       struct kobj_attribute *attr, char *buf)
3810 {
3811 	return nr_hugepages_show_common(kobj, attr, buf);
3812 }
3813 
3814 static ssize_t nr_hugepages_store(struct kobject *kobj,
3815 	       struct kobj_attribute *attr, const char *buf, size_t len)
3816 {
3817 	return nr_hugepages_store_common(false, kobj, buf, len);
3818 }
3819 HSTATE_ATTR(nr_hugepages);
3820 
3821 #ifdef CONFIG_NUMA
3822 
3823 /*
3824  * hstate attribute for optionally mempolicy-based constraint on persistent
3825  * huge page alloc/free.
3826  */
3827 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3828 					   struct kobj_attribute *attr,
3829 					   char *buf)
3830 {
3831 	return nr_hugepages_show_common(kobj, attr, buf);
3832 }
3833 
3834 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3835 	       struct kobj_attribute *attr, const char *buf, size_t len)
3836 {
3837 	return nr_hugepages_store_common(true, kobj, buf, len);
3838 }
3839 HSTATE_ATTR(nr_hugepages_mempolicy);
3840 #endif
3841 
3842 
3843 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3844 					struct kobj_attribute *attr, char *buf)
3845 {
3846 	struct hstate *h = kobj_to_hstate(kobj, NULL);
3847 	return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3848 }
3849 
3850 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3851 		struct kobj_attribute *attr, const char *buf, size_t count)
3852 {
3853 	int err;
3854 	unsigned long input;
3855 	struct hstate *h = kobj_to_hstate(kobj, NULL);
3856 
3857 	if (hstate_is_gigantic(h))
3858 		return -EINVAL;
3859 
3860 	err = kstrtoul(buf, 10, &input);
3861 	if (err)
3862 		return err;
3863 
3864 	spin_lock_irq(&hugetlb_lock);
3865 	h->nr_overcommit_huge_pages = input;
3866 	spin_unlock_irq(&hugetlb_lock);
3867 
3868 	return count;
3869 }
3870 HSTATE_ATTR(nr_overcommit_hugepages);
3871 
3872 static ssize_t free_hugepages_show(struct kobject *kobj,
3873 					struct kobj_attribute *attr, char *buf)
3874 {
3875 	struct hstate *h;
3876 	unsigned long free_huge_pages;
3877 	int nid;
3878 
3879 	h = kobj_to_hstate(kobj, &nid);
3880 	if (nid == NUMA_NO_NODE)
3881 		free_huge_pages = h->free_huge_pages;
3882 	else
3883 		free_huge_pages = h->free_huge_pages_node[nid];
3884 
3885 	return sysfs_emit(buf, "%lu\n", free_huge_pages);
3886 }
3887 HSTATE_ATTR_RO(free_hugepages);
3888 
3889 static ssize_t resv_hugepages_show(struct kobject *kobj,
3890 					struct kobj_attribute *attr, char *buf)
3891 {
3892 	struct hstate *h = kobj_to_hstate(kobj, NULL);
3893 	return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3894 }
3895 HSTATE_ATTR_RO(resv_hugepages);
3896 
3897 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3898 					struct kobj_attribute *attr, char *buf)
3899 {
3900 	struct hstate *h;
3901 	unsigned long surplus_huge_pages;
3902 	int nid;
3903 
3904 	h = kobj_to_hstate(kobj, &nid);
3905 	if (nid == NUMA_NO_NODE)
3906 		surplus_huge_pages = h->surplus_huge_pages;
3907 	else
3908 		surplus_huge_pages = h->surplus_huge_pages_node[nid];
3909 
3910 	return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3911 }
3912 HSTATE_ATTR_RO(surplus_hugepages);
3913 
3914 static ssize_t demote_store(struct kobject *kobj,
3915 	       struct kobj_attribute *attr, const char *buf, size_t len)
3916 {
3917 	unsigned long nr_demote;
3918 	unsigned long nr_available;
3919 	nodemask_t nodes_allowed, *n_mask;
3920 	struct hstate *h;
3921 	int err;
3922 	int nid;
3923 
3924 	err = kstrtoul(buf, 10, &nr_demote);
3925 	if (err)
3926 		return err;
3927 	h = kobj_to_hstate(kobj, &nid);
3928 
3929 	if (nid != NUMA_NO_NODE) {
3930 		init_nodemask_of_node(&nodes_allowed, nid);
3931 		n_mask = &nodes_allowed;
3932 	} else {
3933 		n_mask = &node_states[N_MEMORY];
3934 	}
3935 
3936 	/* Synchronize with other sysfs operations modifying huge pages */
3937 	mutex_lock(&h->resize_lock);
3938 	spin_lock_irq(&hugetlb_lock);
3939 
3940 	while (nr_demote) {
3941 		/*
3942 		 * Check for available pages to demote each time thorough the
3943 		 * loop as demote_pool_huge_page will drop hugetlb_lock.
3944 		 */
3945 		if (nid != NUMA_NO_NODE)
3946 			nr_available = h->free_huge_pages_node[nid];
3947 		else
3948 			nr_available = h->free_huge_pages;
3949 		nr_available -= h->resv_huge_pages;
3950 		if (!nr_available)
3951 			break;
3952 
3953 		err = demote_pool_huge_page(h, n_mask);
3954 		if (err)
3955 			break;
3956 
3957 		nr_demote--;
3958 	}
3959 
3960 	spin_unlock_irq(&hugetlb_lock);
3961 	mutex_unlock(&h->resize_lock);
3962 
3963 	if (err)
3964 		return err;
3965 	return len;
3966 }
3967 HSTATE_ATTR_WO(demote);
3968 
3969 static ssize_t demote_size_show(struct kobject *kobj,
3970 					struct kobj_attribute *attr, char *buf)
3971 {
3972 	struct hstate *h = kobj_to_hstate(kobj, NULL);
3973 	unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K;
3974 
3975 	return sysfs_emit(buf, "%lukB\n", demote_size);
3976 }
3977 
3978 static ssize_t demote_size_store(struct kobject *kobj,
3979 					struct kobj_attribute *attr,
3980 					const char *buf, size_t count)
3981 {
3982 	struct hstate *h, *demote_hstate;
3983 	unsigned long demote_size;
3984 	unsigned int demote_order;
3985 
3986 	demote_size = (unsigned long)memparse(buf, NULL);
3987 
3988 	demote_hstate = size_to_hstate(demote_size);
3989 	if (!demote_hstate)
3990 		return -EINVAL;
3991 	demote_order = demote_hstate->order;
3992 	if (demote_order < HUGETLB_PAGE_ORDER)
3993 		return -EINVAL;
3994 
3995 	/* demote order must be smaller than hstate order */
3996 	h = kobj_to_hstate(kobj, NULL);
3997 	if (demote_order >= h->order)
3998 		return -EINVAL;
3999 
4000 	/* resize_lock synchronizes access to demote size and writes */
4001 	mutex_lock(&h->resize_lock);
4002 	h->demote_order = demote_order;
4003 	mutex_unlock(&h->resize_lock);
4004 
4005 	return count;
4006 }
4007 HSTATE_ATTR(demote_size);
4008 
4009 static struct attribute *hstate_attrs[] = {
4010 	&nr_hugepages_attr.attr,
4011 	&nr_overcommit_hugepages_attr.attr,
4012 	&free_hugepages_attr.attr,
4013 	&resv_hugepages_attr.attr,
4014 	&surplus_hugepages_attr.attr,
4015 #ifdef CONFIG_NUMA
4016 	&nr_hugepages_mempolicy_attr.attr,
4017 #endif
4018 	NULL,
4019 };
4020 
4021 static const struct attribute_group hstate_attr_group = {
4022 	.attrs = hstate_attrs,
4023 };
4024 
4025 static struct attribute *hstate_demote_attrs[] = {
4026 	&demote_size_attr.attr,
4027 	&demote_attr.attr,
4028 	NULL,
4029 };
4030 
4031 static const struct attribute_group hstate_demote_attr_group = {
4032 	.attrs = hstate_demote_attrs,
4033 };
4034 
4035 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
4036 				    struct kobject **hstate_kobjs,
4037 				    const struct attribute_group *hstate_attr_group)
4038 {
4039 	int retval;
4040 	int hi = hstate_index(h);
4041 
4042 	hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
4043 	if (!hstate_kobjs[hi])
4044 		return -ENOMEM;
4045 
4046 	retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
4047 	if (retval) {
4048 		kobject_put(hstate_kobjs[hi]);
4049 		hstate_kobjs[hi] = NULL;
4050 		return retval;
4051 	}
4052 
4053 	if (h->demote_order) {
4054 		retval = sysfs_create_group(hstate_kobjs[hi],
4055 					    &hstate_demote_attr_group);
4056 		if (retval) {
4057 			pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name);
4058 			sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group);
4059 			kobject_put(hstate_kobjs[hi]);
4060 			hstate_kobjs[hi] = NULL;
4061 			return retval;
4062 		}
4063 	}
4064 
4065 	return 0;
4066 }
4067 
4068 #ifdef CONFIG_NUMA
4069 static bool hugetlb_sysfs_initialized __ro_after_init;
4070 
4071 /*
4072  * node_hstate/s - associate per node hstate attributes, via their kobjects,
4073  * with node devices in node_devices[] using a parallel array.  The array
4074  * index of a node device or _hstate == node id.
4075  * This is here to avoid any static dependency of the node device driver, in
4076  * the base kernel, on the hugetlb module.
4077  */
4078 struct node_hstate {
4079 	struct kobject		*hugepages_kobj;
4080 	struct kobject		*hstate_kobjs[HUGE_MAX_HSTATE];
4081 };
4082 static struct node_hstate node_hstates[MAX_NUMNODES];
4083 
4084 /*
4085  * A subset of global hstate attributes for node devices
4086  */
4087 static struct attribute *per_node_hstate_attrs[] = {
4088 	&nr_hugepages_attr.attr,
4089 	&free_hugepages_attr.attr,
4090 	&surplus_hugepages_attr.attr,
4091 	NULL,
4092 };
4093 
4094 static const struct attribute_group per_node_hstate_attr_group = {
4095 	.attrs = per_node_hstate_attrs,
4096 };
4097 
4098 /*
4099  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
4100  * Returns node id via non-NULL nidp.
4101  */
4102 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4103 {
4104 	int nid;
4105 
4106 	for (nid = 0; nid < nr_node_ids; nid++) {
4107 		struct node_hstate *nhs = &node_hstates[nid];
4108 		int i;
4109 		for (i = 0; i < HUGE_MAX_HSTATE; i++)
4110 			if (nhs->hstate_kobjs[i] == kobj) {
4111 				if (nidp)
4112 					*nidp = nid;
4113 				return &hstates[i];
4114 			}
4115 	}
4116 
4117 	BUG();
4118 	return NULL;
4119 }
4120 
4121 /*
4122  * Unregister hstate attributes from a single node device.
4123  * No-op if no hstate attributes attached.
4124  */
4125 void hugetlb_unregister_node(struct node *node)
4126 {
4127 	struct hstate *h;
4128 	struct node_hstate *nhs = &node_hstates[node->dev.id];
4129 
4130 	if (!nhs->hugepages_kobj)
4131 		return;		/* no hstate attributes */
4132 
4133 	for_each_hstate(h) {
4134 		int idx = hstate_index(h);
4135 		struct kobject *hstate_kobj = nhs->hstate_kobjs[idx];
4136 
4137 		if (!hstate_kobj)
4138 			continue;
4139 		if (h->demote_order)
4140 			sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group);
4141 		sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group);
4142 		kobject_put(hstate_kobj);
4143 		nhs->hstate_kobjs[idx] = NULL;
4144 	}
4145 
4146 	kobject_put(nhs->hugepages_kobj);
4147 	nhs->hugepages_kobj = NULL;
4148 }
4149 
4150 
4151 /*
4152  * Register hstate attributes for a single node device.
4153  * No-op if attributes already registered.
4154  */
4155 void hugetlb_register_node(struct node *node)
4156 {
4157 	struct hstate *h;
4158 	struct node_hstate *nhs = &node_hstates[node->dev.id];
4159 	int err;
4160 
4161 	if (!hugetlb_sysfs_initialized)
4162 		return;
4163 
4164 	if (nhs->hugepages_kobj)
4165 		return;		/* already allocated */
4166 
4167 	nhs->hugepages_kobj = kobject_create_and_add("hugepages",
4168 							&node->dev.kobj);
4169 	if (!nhs->hugepages_kobj)
4170 		return;
4171 
4172 	for_each_hstate(h) {
4173 		err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
4174 						nhs->hstate_kobjs,
4175 						&per_node_hstate_attr_group);
4176 		if (err) {
4177 			pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
4178 				h->name, node->dev.id);
4179 			hugetlb_unregister_node(node);
4180 			break;
4181 		}
4182 	}
4183 }
4184 
4185 /*
4186  * hugetlb init time:  register hstate attributes for all registered node
4187  * devices of nodes that have memory.  All on-line nodes should have
4188  * registered their associated device by this time.
4189  */
4190 static void __init hugetlb_register_all_nodes(void)
4191 {
4192 	int nid;
4193 
4194 	for_each_online_node(nid)
4195 		hugetlb_register_node(node_devices[nid]);
4196 }
4197 #else	/* !CONFIG_NUMA */
4198 
4199 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4200 {
4201 	BUG();
4202 	if (nidp)
4203 		*nidp = -1;
4204 	return NULL;
4205 }
4206 
4207 static void hugetlb_register_all_nodes(void) { }
4208 
4209 #endif
4210 
4211 #ifdef CONFIG_CMA
4212 static void __init hugetlb_cma_check(void);
4213 #else
4214 static inline __init void hugetlb_cma_check(void)
4215 {
4216 }
4217 #endif
4218 
4219 static void __init hugetlb_sysfs_init(void)
4220 {
4221 	struct hstate *h;
4222 	int err;
4223 
4224 	hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
4225 	if (!hugepages_kobj)
4226 		return;
4227 
4228 	for_each_hstate(h) {
4229 		err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
4230 					 hstate_kobjs, &hstate_attr_group);
4231 		if (err)
4232 			pr_err("HugeTLB: Unable to add hstate %s", h->name);
4233 	}
4234 
4235 #ifdef CONFIG_NUMA
4236 	hugetlb_sysfs_initialized = true;
4237 #endif
4238 	hugetlb_register_all_nodes();
4239 }
4240 
4241 #ifdef CONFIG_SYSCTL
4242 static void hugetlb_sysctl_init(void);
4243 #else
4244 static inline void hugetlb_sysctl_init(void) { }
4245 #endif
4246 
4247 static int __init hugetlb_init(void)
4248 {
4249 	int i;
4250 
4251 	BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
4252 			__NR_HPAGEFLAGS);
4253 
4254 	if (!hugepages_supported()) {
4255 		if (hugetlb_max_hstate || default_hstate_max_huge_pages)
4256 			pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
4257 		return 0;
4258 	}
4259 
4260 	/*
4261 	 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists.  Some
4262 	 * architectures depend on setup being done here.
4263 	 */
4264 	hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
4265 	if (!parsed_default_hugepagesz) {
4266 		/*
4267 		 * If we did not parse a default huge page size, set
4268 		 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
4269 		 * number of huge pages for this default size was implicitly
4270 		 * specified, set that here as well.
4271 		 * Note that the implicit setting will overwrite an explicit
4272 		 * setting.  A warning will be printed in this case.
4273 		 */
4274 		default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
4275 		if (default_hstate_max_huge_pages) {
4276 			if (default_hstate.max_huge_pages) {
4277 				char buf[32];
4278 
4279 				string_get_size(huge_page_size(&default_hstate),
4280 					1, STRING_UNITS_2, buf, 32);
4281 				pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
4282 					default_hstate.max_huge_pages, buf);
4283 				pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
4284 					default_hstate_max_huge_pages);
4285 			}
4286 			default_hstate.max_huge_pages =
4287 				default_hstate_max_huge_pages;
4288 
4289 			for_each_online_node(i)
4290 				default_hstate.max_huge_pages_node[i] =
4291 					default_hugepages_in_node[i];
4292 		}
4293 	}
4294 
4295 	hugetlb_cma_check();
4296 	hugetlb_init_hstates();
4297 	gather_bootmem_prealloc();
4298 	report_hugepages();
4299 
4300 	hugetlb_sysfs_init();
4301 	hugetlb_cgroup_file_init();
4302 	hugetlb_sysctl_init();
4303 
4304 #ifdef CONFIG_SMP
4305 	num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
4306 #else
4307 	num_fault_mutexes = 1;
4308 #endif
4309 	hugetlb_fault_mutex_table =
4310 		kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
4311 			      GFP_KERNEL);
4312 	BUG_ON(!hugetlb_fault_mutex_table);
4313 
4314 	for (i = 0; i < num_fault_mutexes; i++)
4315 		mutex_init(&hugetlb_fault_mutex_table[i]);
4316 	return 0;
4317 }
4318 subsys_initcall(hugetlb_init);
4319 
4320 /* Overwritten by architectures with more huge page sizes */
4321 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
4322 {
4323 	return size == HPAGE_SIZE;
4324 }
4325 
4326 void __init hugetlb_add_hstate(unsigned int order)
4327 {
4328 	struct hstate *h;
4329 	unsigned long i;
4330 
4331 	if (size_to_hstate(PAGE_SIZE << order)) {
4332 		return;
4333 	}
4334 	BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
4335 	BUG_ON(order == 0);
4336 	h = &hstates[hugetlb_max_hstate++];
4337 	__mutex_init(&h->resize_lock, "resize mutex", &h->resize_key);
4338 	h->order = order;
4339 	h->mask = ~(huge_page_size(h) - 1);
4340 	for (i = 0; i < MAX_NUMNODES; ++i)
4341 		INIT_LIST_HEAD(&h->hugepage_freelists[i]);
4342 	INIT_LIST_HEAD(&h->hugepage_activelist);
4343 	h->next_nid_to_alloc = first_memory_node;
4344 	h->next_nid_to_free = first_memory_node;
4345 	snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
4346 					huge_page_size(h)/SZ_1K);
4347 
4348 	parsed_hstate = h;
4349 }
4350 
4351 bool __init __weak hugetlb_node_alloc_supported(void)
4352 {
4353 	return true;
4354 }
4355 
4356 static void __init hugepages_clear_pages_in_node(void)
4357 {
4358 	if (!hugetlb_max_hstate) {
4359 		default_hstate_max_huge_pages = 0;
4360 		memset(default_hugepages_in_node, 0,
4361 			sizeof(default_hugepages_in_node));
4362 	} else {
4363 		parsed_hstate->max_huge_pages = 0;
4364 		memset(parsed_hstate->max_huge_pages_node, 0,
4365 			sizeof(parsed_hstate->max_huge_pages_node));
4366 	}
4367 }
4368 
4369 /*
4370  * hugepages command line processing
4371  * hugepages normally follows a valid hugepagsz or default_hugepagsz
4372  * specification.  If not, ignore the hugepages value.  hugepages can also
4373  * be the first huge page command line  option in which case it implicitly
4374  * specifies the number of huge pages for the default size.
4375  */
4376 static int __init hugepages_setup(char *s)
4377 {
4378 	unsigned long *mhp;
4379 	static unsigned long *last_mhp;
4380 	int node = NUMA_NO_NODE;
4381 	int count;
4382 	unsigned long tmp;
4383 	char *p = s;
4384 
4385 	if (!parsed_valid_hugepagesz) {
4386 		pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
4387 		parsed_valid_hugepagesz = true;
4388 		return 1;
4389 	}
4390 
4391 	/*
4392 	 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
4393 	 * yet, so this hugepages= parameter goes to the "default hstate".
4394 	 * Otherwise, it goes with the previously parsed hugepagesz or
4395 	 * default_hugepagesz.
4396 	 */
4397 	else if (!hugetlb_max_hstate)
4398 		mhp = &default_hstate_max_huge_pages;
4399 	else
4400 		mhp = &parsed_hstate->max_huge_pages;
4401 
4402 	if (mhp == last_mhp) {
4403 		pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
4404 		return 1;
4405 	}
4406 
4407 	while (*p) {
4408 		count = 0;
4409 		if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4410 			goto invalid;
4411 		/* Parameter is node format */
4412 		if (p[count] == ':') {
4413 			if (!hugetlb_node_alloc_supported()) {
4414 				pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n");
4415 				return 1;
4416 			}
4417 			if (tmp >= MAX_NUMNODES || !node_online(tmp))
4418 				goto invalid;
4419 			node = array_index_nospec(tmp, MAX_NUMNODES);
4420 			p += count + 1;
4421 			/* Parse hugepages */
4422 			if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4423 				goto invalid;
4424 			if (!hugetlb_max_hstate)
4425 				default_hugepages_in_node[node] = tmp;
4426 			else
4427 				parsed_hstate->max_huge_pages_node[node] = tmp;
4428 			*mhp += tmp;
4429 			/* Go to parse next node*/
4430 			if (p[count] == ',')
4431 				p += count + 1;
4432 			else
4433 				break;
4434 		} else {
4435 			if (p != s)
4436 				goto invalid;
4437 			*mhp = tmp;
4438 			break;
4439 		}
4440 	}
4441 
4442 	/*
4443 	 * Global state is always initialized later in hugetlb_init.
4444 	 * But we need to allocate gigantic hstates here early to still
4445 	 * use the bootmem allocator.
4446 	 */
4447 	if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
4448 		hugetlb_hstate_alloc_pages(parsed_hstate);
4449 
4450 	last_mhp = mhp;
4451 
4452 	return 1;
4453 
4454 invalid:
4455 	pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p);
4456 	hugepages_clear_pages_in_node();
4457 	return 1;
4458 }
4459 __setup("hugepages=", hugepages_setup);
4460 
4461 /*
4462  * hugepagesz command line processing
4463  * A specific huge page size can only be specified once with hugepagesz.
4464  * hugepagesz is followed by hugepages on the command line.  The global
4465  * variable 'parsed_valid_hugepagesz' is used to determine if prior
4466  * hugepagesz argument was valid.
4467  */
4468 static int __init hugepagesz_setup(char *s)
4469 {
4470 	unsigned long size;
4471 	struct hstate *h;
4472 
4473 	parsed_valid_hugepagesz = false;
4474 	size = (unsigned long)memparse(s, NULL);
4475 
4476 	if (!arch_hugetlb_valid_size(size)) {
4477 		pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
4478 		return 1;
4479 	}
4480 
4481 	h = size_to_hstate(size);
4482 	if (h) {
4483 		/*
4484 		 * hstate for this size already exists.  This is normally
4485 		 * an error, but is allowed if the existing hstate is the
4486 		 * default hstate.  More specifically, it is only allowed if
4487 		 * the number of huge pages for the default hstate was not
4488 		 * previously specified.
4489 		 */
4490 		if (!parsed_default_hugepagesz ||  h != &default_hstate ||
4491 		    default_hstate.max_huge_pages) {
4492 			pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
4493 			return 1;
4494 		}
4495 
4496 		/*
4497 		 * No need to call hugetlb_add_hstate() as hstate already
4498 		 * exists.  But, do set parsed_hstate so that a following
4499 		 * hugepages= parameter will be applied to this hstate.
4500 		 */
4501 		parsed_hstate = h;
4502 		parsed_valid_hugepagesz = true;
4503 		return 1;
4504 	}
4505 
4506 	hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4507 	parsed_valid_hugepagesz = true;
4508 	return 1;
4509 }
4510 __setup("hugepagesz=", hugepagesz_setup);
4511 
4512 /*
4513  * default_hugepagesz command line input
4514  * Only one instance of default_hugepagesz allowed on command line.
4515  */
4516 static int __init default_hugepagesz_setup(char *s)
4517 {
4518 	unsigned long size;
4519 	int i;
4520 
4521 	parsed_valid_hugepagesz = false;
4522 	if (parsed_default_hugepagesz) {
4523 		pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
4524 		return 1;
4525 	}
4526 
4527 	size = (unsigned long)memparse(s, NULL);
4528 
4529 	if (!arch_hugetlb_valid_size(size)) {
4530 		pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
4531 		return 1;
4532 	}
4533 
4534 	hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4535 	parsed_valid_hugepagesz = true;
4536 	parsed_default_hugepagesz = true;
4537 	default_hstate_idx = hstate_index(size_to_hstate(size));
4538 
4539 	/*
4540 	 * The number of default huge pages (for this size) could have been
4541 	 * specified as the first hugetlb parameter: hugepages=X.  If so,
4542 	 * then default_hstate_max_huge_pages is set.  If the default huge
4543 	 * page size is gigantic (> MAX_ORDER), then the pages must be
4544 	 * allocated here from bootmem allocator.
4545 	 */
4546 	if (default_hstate_max_huge_pages) {
4547 		default_hstate.max_huge_pages = default_hstate_max_huge_pages;
4548 		for_each_online_node(i)
4549 			default_hstate.max_huge_pages_node[i] =
4550 				default_hugepages_in_node[i];
4551 		if (hstate_is_gigantic(&default_hstate))
4552 			hugetlb_hstate_alloc_pages(&default_hstate);
4553 		default_hstate_max_huge_pages = 0;
4554 	}
4555 
4556 	return 1;
4557 }
4558 __setup("default_hugepagesz=", default_hugepagesz_setup);
4559 
4560 static unsigned int allowed_mems_nr(struct hstate *h)
4561 {
4562 	int node;
4563 	unsigned int nr = 0;
4564 	nodemask_t *mbind_nodemask;
4565 	unsigned int *array = h->free_huge_pages_node;
4566 	gfp_t gfp_mask = htlb_alloc_mask(h);
4567 
4568 	mbind_nodemask = policy_mbind_nodemask(gfp_mask);
4569 	for_each_node_mask(node, cpuset_current_mems_allowed) {
4570 		if (!mbind_nodemask || node_isset(node, *mbind_nodemask))
4571 			nr += array[node];
4572 	}
4573 
4574 	return nr;
4575 }
4576 
4577 #ifdef CONFIG_SYSCTL
4578 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
4579 					  void *buffer, size_t *length,
4580 					  loff_t *ppos, unsigned long *out)
4581 {
4582 	struct ctl_table dup_table;
4583 
4584 	/*
4585 	 * In order to avoid races with __do_proc_doulongvec_minmax(), we
4586 	 * can duplicate the @table and alter the duplicate of it.
4587 	 */
4588 	dup_table = *table;
4589 	dup_table.data = out;
4590 
4591 	return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
4592 }
4593 
4594 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
4595 			 struct ctl_table *table, int write,
4596 			 void *buffer, size_t *length, loff_t *ppos)
4597 {
4598 	struct hstate *h = &default_hstate;
4599 	unsigned long tmp = h->max_huge_pages;
4600 	int ret;
4601 
4602 	if (!hugepages_supported())
4603 		return -EOPNOTSUPP;
4604 
4605 	ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4606 					     &tmp);
4607 	if (ret)
4608 		goto out;
4609 
4610 	if (write)
4611 		ret = __nr_hugepages_store_common(obey_mempolicy, h,
4612 						  NUMA_NO_NODE, tmp, *length);
4613 out:
4614 	return ret;
4615 }
4616 
4617 static int hugetlb_sysctl_handler(struct ctl_table *table, int write,
4618 			  void *buffer, size_t *length, loff_t *ppos)
4619 {
4620 
4621 	return hugetlb_sysctl_handler_common(false, table, write,
4622 							buffer, length, ppos);
4623 }
4624 
4625 #ifdef CONFIG_NUMA
4626 static int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
4627 			  void *buffer, size_t *length, loff_t *ppos)
4628 {
4629 	return hugetlb_sysctl_handler_common(true, table, write,
4630 							buffer, length, ppos);
4631 }
4632 #endif /* CONFIG_NUMA */
4633 
4634 static int hugetlb_overcommit_handler(struct ctl_table *table, int write,
4635 		void *buffer, size_t *length, loff_t *ppos)
4636 {
4637 	struct hstate *h = &default_hstate;
4638 	unsigned long tmp;
4639 	int ret;
4640 
4641 	if (!hugepages_supported())
4642 		return -EOPNOTSUPP;
4643 
4644 	tmp = h->nr_overcommit_huge_pages;
4645 
4646 	if (write && hstate_is_gigantic(h))
4647 		return -EINVAL;
4648 
4649 	ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4650 					     &tmp);
4651 	if (ret)
4652 		goto out;
4653 
4654 	if (write) {
4655 		spin_lock_irq(&hugetlb_lock);
4656 		h->nr_overcommit_huge_pages = tmp;
4657 		spin_unlock_irq(&hugetlb_lock);
4658 	}
4659 out:
4660 	return ret;
4661 }
4662 
4663 static struct ctl_table hugetlb_table[] = {
4664 	{
4665 		.procname	= "nr_hugepages",
4666 		.data		= NULL,
4667 		.maxlen		= sizeof(unsigned long),
4668 		.mode		= 0644,
4669 		.proc_handler	= hugetlb_sysctl_handler,
4670 	},
4671 #ifdef CONFIG_NUMA
4672 	{
4673 		.procname       = "nr_hugepages_mempolicy",
4674 		.data           = NULL,
4675 		.maxlen         = sizeof(unsigned long),
4676 		.mode           = 0644,
4677 		.proc_handler   = &hugetlb_mempolicy_sysctl_handler,
4678 	},
4679 #endif
4680 	{
4681 		.procname	= "hugetlb_shm_group",
4682 		.data		= &sysctl_hugetlb_shm_group,
4683 		.maxlen		= sizeof(gid_t),
4684 		.mode		= 0644,
4685 		.proc_handler	= proc_dointvec,
4686 	},
4687 	{
4688 		.procname	= "nr_overcommit_hugepages",
4689 		.data		= NULL,
4690 		.maxlen		= sizeof(unsigned long),
4691 		.mode		= 0644,
4692 		.proc_handler	= hugetlb_overcommit_handler,
4693 	},
4694 	{ }
4695 };
4696 
4697 static void hugetlb_sysctl_init(void)
4698 {
4699 	register_sysctl_init("vm", hugetlb_table);
4700 }
4701 #endif /* CONFIG_SYSCTL */
4702 
4703 void hugetlb_report_meminfo(struct seq_file *m)
4704 {
4705 	struct hstate *h;
4706 	unsigned long total = 0;
4707 
4708 	if (!hugepages_supported())
4709 		return;
4710 
4711 	for_each_hstate(h) {
4712 		unsigned long count = h->nr_huge_pages;
4713 
4714 		total += huge_page_size(h) * count;
4715 
4716 		if (h == &default_hstate)
4717 			seq_printf(m,
4718 				   "HugePages_Total:   %5lu\n"
4719 				   "HugePages_Free:    %5lu\n"
4720 				   "HugePages_Rsvd:    %5lu\n"
4721 				   "HugePages_Surp:    %5lu\n"
4722 				   "Hugepagesize:   %8lu kB\n",
4723 				   count,
4724 				   h->free_huge_pages,
4725 				   h->resv_huge_pages,
4726 				   h->surplus_huge_pages,
4727 				   huge_page_size(h) / SZ_1K);
4728 	}
4729 
4730 	seq_printf(m, "Hugetlb:        %8lu kB\n", total / SZ_1K);
4731 }
4732 
4733 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
4734 {
4735 	struct hstate *h = &default_hstate;
4736 
4737 	if (!hugepages_supported())
4738 		return 0;
4739 
4740 	return sysfs_emit_at(buf, len,
4741 			     "Node %d HugePages_Total: %5u\n"
4742 			     "Node %d HugePages_Free:  %5u\n"
4743 			     "Node %d HugePages_Surp:  %5u\n",
4744 			     nid, h->nr_huge_pages_node[nid],
4745 			     nid, h->free_huge_pages_node[nid],
4746 			     nid, h->surplus_huge_pages_node[nid]);
4747 }
4748 
4749 void hugetlb_show_meminfo_node(int nid)
4750 {
4751 	struct hstate *h;
4752 
4753 	if (!hugepages_supported())
4754 		return;
4755 
4756 	for_each_hstate(h)
4757 		printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
4758 			nid,
4759 			h->nr_huge_pages_node[nid],
4760 			h->free_huge_pages_node[nid],
4761 			h->surplus_huge_pages_node[nid],
4762 			huge_page_size(h) / SZ_1K);
4763 }
4764 
4765 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
4766 {
4767 	seq_printf(m, "HugetlbPages:\t%8lu kB\n",
4768 		   K(atomic_long_read(&mm->hugetlb_usage)));
4769 }
4770 
4771 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
4772 unsigned long hugetlb_total_pages(void)
4773 {
4774 	struct hstate *h;
4775 	unsigned long nr_total_pages = 0;
4776 
4777 	for_each_hstate(h)
4778 		nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
4779 	return nr_total_pages;
4780 }
4781 
4782 static int hugetlb_acct_memory(struct hstate *h, long delta)
4783 {
4784 	int ret = -ENOMEM;
4785 
4786 	if (!delta)
4787 		return 0;
4788 
4789 	spin_lock_irq(&hugetlb_lock);
4790 	/*
4791 	 * When cpuset is configured, it breaks the strict hugetlb page
4792 	 * reservation as the accounting is done on a global variable. Such
4793 	 * reservation is completely rubbish in the presence of cpuset because
4794 	 * the reservation is not checked against page availability for the
4795 	 * current cpuset. Application can still potentially OOM'ed by kernel
4796 	 * with lack of free htlb page in cpuset that the task is in.
4797 	 * Attempt to enforce strict accounting with cpuset is almost
4798 	 * impossible (or too ugly) because cpuset is too fluid that
4799 	 * task or memory node can be dynamically moved between cpusets.
4800 	 *
4801 	 * The change of semantics for shared hugetlb mapping with cpuset is
4802 	 * undesirable. However, in order to preserve some of the semantics,
4803 	 * we fall back to check against current free page availability as
4804 	 * a best attempt and hopefully to minimize the impact of changing
4805 	 * semantics that cpuset has.
4806 	 *
4807 	 * Apart from cpuset, we also have memory policy mechanism that
4808 	 * also determines from which node the kernel will allocate memory
4809 	 * in a NUMA system. So similar to cpuset, we also should consider
4810 	 * the memory policy of the current task. Similar to the description
4811 	 * above.
4812 	 */
4813 	if (delta > 0) {
4814 		if (gather_surplus_pages(h, delta) < 0)
4815 			goto out;
4816 
4817 		if (delta > allowed_mems_nr(h)) {
4818 			return_unused_surplus_pages(h, delta);
4819 			goto out;
4820 		}
4821 	}
4822 
4823 	ret = 0;
4824 	if (delta < 0)
4825 		return_unused_surplus_pages(h, (unsigned long) -delta);
4826 
4827 out:
4828 	spin_unlock_irq(&hugetlb_lock);
4829 	return ret;
4830 }
4831 
4832 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4833 {
4834 	struct resv_map *resv = vma_resv_map(vma);
4835 
4836 	/*
4837 	 * HPAGE_RESV_OWNER indicates a private mapping.
4838 	 * This new VMA should share its siblings reservation map if present.
4839 	 * The VMA will only ever have a valid reservation map pointer where
4840 	 * it is being copied for another still existing VMA.  As that VMA
4841 	 * has a reference to the reservation map it cannot disappear until
4842 	 * after this open call completes.  It is therefore safe to take a
4843 	 * new reference here without additional locking.
4844 	 */
4845 	if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
4846 		resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
4847 		kref_get(&resv->refs);
4848 	}
4849 
4850 	/*
4851 	 * vma_lock structure for sharable mappings is vma specific.
4852 	 * Clear old pointer (if copied via vm_area_dup) and allocate
4853 	 * new structure.  Before clearing, make sure vma_lock is not
4854 	 * for this vma.
4855 	 */
4856 	if (vma->vm_flags & VM_MAYSHARE) {
4857 		struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
4858 
4859 		if (vma_lock) {
4860 			if (vma_lock->vma != vma) {
4861 				vma->vm_private_data = NULL;
4862 				hugetlb_vma_lock_alloc(vma);
4863 			} else
4864 				pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__);
4865 		} else
4866 			hugetlb_vma_lock_alloc(vma);
4867 	}
4868 }
4869 
4870 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4871 {
4872 	struct hstate *h = hstate_vma(vma);
4873 	struct resv_map *resv;
4874 	struct hugepage_subpool *spool = subpool_vma(vma);
4875 	unsigned long reserve, start, end;
4876 	long gbl_reserve;
4877 
4878 	hugetlb_vma_lock_free(vma);
4879 
4880 	resv = vma_resv_map(vma);
4881 	if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4882 		return;
4883 
4884 	start = vma_hugecache_offset(h, vma, vma->vm_start);
4885 	end = vma_hugecache_offset(h, vma, vma->vm_end);
4886 
4887 	reserve = (end - start) - region_count(resv, start, end);
4888 	hugetlb_cgroup_uncharge_counter(resv, start, end);
4889 	if (reserve) {
4890 		/*
4891 		 * Decrement reserve counts.  The global reserve count may be
4892 		 * adjusted if the subpool has a minimum size.
4893 		 */
4894 		gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4895 		hugetlb_acct_memory(h, -gbl_reserve);
4896 	}
4897 
4898 	kref_put(&resv->refs, resv_map_release);
4899 }
4900 
4901 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4902 {
4903 	if (addr & ~(huge_page_mask(hstate_vma(vma))))
4904 		return -EINVAL;
4905 
4906 	/*
4907 	 * PMD sharing is only possible for PUD_SIZE-aligned address ranges
4908 	 * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this
4909 	 * split, unshare PMDs in the PUD_SIZE interval surrounding addr now.
4910 	 */
4911 	if (addr & ~PUD_MASK) {
4912 		/*
4913 		 * hugetlb_vm_op_split is called right before we attempt to
4914 		 * split the VMA. We will need to unshare PMDs in the old and
4915 		 * new VMAs, so let's unshare before we split.
4916 		 */
4917 		unsigned long floor = addr & PUD_MASK;
4918 		unsigned long ceil = floor + PUD_SIZE;
4919 
4920 		if (floor >= vma->vm_start && ceil <= vma->vm_end)
4921 			hugetlb_unshare_pmds(vma, floor, ceil);
4922 	}
4923 
4924 	return 0;
4925 }
4926 
4927 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4928 {
4929 	return huge_page_size(hstate_vma(vma));
4930 }
4931 
4932 /*
4933  * We cannot handle pagefaults against hugetlb pages at all.  They cause
4934  * handle_mm_fault() to try to instantiate regular-sized pages in the
4935  * hugepage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
4936  * this far.
4937  */
4938 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4939 {
4940 	BUG();
4941 	return 0;
4942 }
4943 
4944 /*
4945  * When a new function is introduced to vm_operations_struct and added
4946  * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4947  * This is because under System V memory model, mappings created via
4948  * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4949  * their original vm_ops are overwritten with shm_vm_ops.
4950  */
4951 const struct vm_operations_struct hugetlb_vm_ops = {
4952 	.fault = hugetlb_vm_op_fault,
4953 	.open = hugetlb_vm_op_open,
4954 	.close = hugetlb_vm_op_close,
4955 	.may_split = hugetlb_vm_op_split,
4956 	.pagesize = hugetlb_vm_op_pagesize,
4957 };
4958 
4959 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4960 				int writable)
4961 {
4962 	pte_t entry;
4963 	unsigned int shift = huge_page_shift(hstate_vma(vma));
4964 
4965 	if (writable) {
4966 		entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4967 					 vma->vm_page_prot)));
4968 	} else {
4969 		entry = huge_pte_wrprotect(mk_huge_pte(page,
4970 					   vma->vm_page_prot));
4971 	}
4972 	entry = pte_mkyoung(entry);
4973 	entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4974 
4975 	return entry;
4976 }
4977 
4978 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4979 				   unsigned long address, pte_t *ptep)
4980 {
4981 	pte_t entry;
4982 
4983 	entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4984 	if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4985 		update_mmu_cache(vma, address, ptep);
4986 }
4987 
4988 bool is_hugetlb_entry_migration(pte_t pte)
4989 {
4990 	swp_entry_t swp;
4991 
4992 	if (huge_pte_none(pte) || pte_present(pte))
4993 		return false;
4994 	swp = pte_to_swp_entry(pte);
4995 	if (is_migration_entry(swp))
4996 		return true;
4997 	else
4998 		return false;
4999 }
5000 
5001 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
5002 {
5003 	swp_entry_t swp;
5004 
5005 	if (huge_pte_none(pte) || pte_present(pte))
5006 		return false;
5007 	swp = pte_to_swp_entry(pte);
5008 	if (is_hwpoison_entry(swp))
5009 		return true;
5010 	else
5011 		return false;
5012 }
5013 
5014 static void
5015 hugetlb_install_folio(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
5016 		      struct folio *new_folio, pte_t old, unsigned long sz)
5017 {
5018 	pte_t newpte = make_huge_pte(vma, &new_folio->page, 1);
5019 
5020 	__folio_mark_uptodate(new_folio);
5021 	hugepage_add_new_anon_rmap(new_folio, vma, addr);
5022 	if (userfaultfd_wp(vma) && huge_pte_uffd_wp(old))
5023 		newpte = huge_pte_mkuffd_wp(newpte);
5024 	set_huge_pte_at(vma->vm_mm, addr, ptep, newpte, sz);
5025 	hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
5026 	folio_set_hugetlb_migratable(new_folio);
5027 }
5028 
5029 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
5030 			    struct vm_area_struct *dst_vma,
5031 			    struct vm_area_struct *src_vma)
5032 {
5033 	pte_t *src_pte, *dst_pte, entry;
5034 	struct folio *pte_folio;
5035 	unsigned long addr;
5036 	bool cow = is_cow_mapping(src_vma->vm_flags);
5037 	struct hstate *h = hstate_vma(src_vma);
5038 	unsigned long sz = huge_page_size(h);
5039 	unsigned long npages = pages_per_huge_page(h);
5040 	struct mmu_notifier_range range;
5041 	unsigned long last_addr_mask;
5042 	int ret = 0;
5043 
5044 	if (cow) {
5045 		mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src,
5046 					src_vma->vm_start,
5047 					src_vma->vm_end);
5048 		mmu_notifier_invalidate_range_start(&range);
5049 		vma_assert_write_locked(src_vma);
5050 		raw_write_seqcount_begin(&src->write_protect_seq);
5051 	} else {
5052 		/*
5053 		 * For shared mappings the vma lock must be held before
5054 		 * calling hugetlb_walk() in the src vma. Otherwise, the
5055 		 * returned ptep could go away if part of a shared pmd and
5056 		 * another thread calls huge_pmd_unshare.
5057 		 */
5058 		hugetlb_vma_lock_read(src_vma);
5059 	}
5060 
5061 	last_addr_mask = hugetlb_mask_last_page(h);
5062 	for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) {
5063 		spinlock_t *src_ptl, *dst_ptl;
5064 		src_pte = hugetlb_walk(src_vma, addr, sz);
5065 		if (!src_pte) {
5066 			addr |= last_addr_mask;
5067 			continue;
5068 		}
5069 		dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz);
5070 		if (!dst_pte) {
5071 			ret = -ENOMEM;
5072 			break;
5073 		}
5074 
5075 		/*
5076 		 * If the pagetables are shared don't copy or take references.
5077 		 *
5078 		 * dst_pte == src_pte is the common case of src/dest sharing.
5079 		 * However, src could have 'unshared' and dst shares with
5080 		 * another vma. So page_count of ptep page is checked instead
5081 		 * to reliably determine whether pte is shared.
5082 		 */
5083 		if (page_count(virt_to_page(dst_pte)) > 1) {
5084 			addr |= last_addr_mask;
5085 			continue;
5086 		}
5087 
5088 		dst_ptl = huge_pte_lock(h, dst, dst_pte);
5089 		src_ptl = huge_pte_lockptr(h, src, src_pte);
5090 		spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5091 		entry = huge_ptep_get(src_pte);
5092 again:
5093 		if (huge_pte_none(entry)) {
5094 			/*
5095 			 * Skip if src entry none.
5096 			 */
5097 			;
5098 		} else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) {
5099 			if (!userfaultfd_wp(dst_vma))
5100 				entry = huge_pte_clear_uffd_wp(entry);
5101 			set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5102 		} else if (unlikely(is_hugetlb_entry_migration(entry))) {
5103 			swp_entry_t swp_entry = pte_to_swp_entry(entry);
5104 			bool uffd_wp = pte_swp_uffd_wp(entry);
5105 
5106 			if (!is_readable_migration_entry(swp_entry) && cow) {
5107 				/*
5108 				 * COW mappings require pages in both
5109 				 * parent and child to be set to read.
5110 				 */
5111 				swp_entry = make_readable_migration_entry(
5112 							swp_offset(swp_entry));
5113 				entry = swp_entry_to_pte(swp_entry);
5114 				if (userfaultfd_wp(src_vma) && uffd_wp)
5115 					entry = pte_swp_mkuffd_wp(entry);
5116 				set_huge_pte_at(src, addr, src_pte, entry, sz);
5117 			}
5118 			if (!userfaultfd_wp(dst_vma))
5119 				entry = huge_pte_clear_uffd_wp(entry);
5120 			set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5121 		} else if (unlikely(is_pte_marker(entry))) {
5122 			pte_marker marker = copy_pte_marker(
5123 				pte_to_swp_entry(entry), dst_vma);
5124 
5125 			if (marker)
5126 				set_huge_pte_at(dst, addr, dst_pte,
5127 						make_pte_marker(marker), sz);
5128 		} else {
5129 			entry = huge_ptep_get(src_pte);
5130 			pte_folio = page_folio(pte_page(entry));
5131 			folio_get(pte_folio);
5132 
5133 			/*
5134 			 * Failing to duplicate the anon rmap is a rare case
5135 			 * where we see pinned hugetlb pages while they're
5136 			 * prone to COW. We need to do the COW earlier during
5137 			 * fork.
5138 			 *
5139 			 * When pre-allocating the page or copying data, we
5140 			 * need to be without the pgtable locks since we could
5141 			 * sleep during the process.
5142 			 */
5143 			if (!folio_test_anon(pte_folio)) {
5144 				page_dup_file_rmap(&pte_folio->page, true);
5145 			} else if (page_try_dup_anon_rmap(&pte_folio->page,
5146 							  true, src_vma)) {
5147 				pte_t src_pte_old = entry;
5148 				struct folio *new_folio;
5149 
5150 				spin_unlock(src_ptl);
5151 				spin_unlock(dst_ptl);
5152 				/* Do not use reserve as it's private owned */
5153 				new_folio = alloc_hugetlb_folio(dst_vma, addr, 1);
5154 				if (IS_ERR(new_folio)) {
5155 					folio_put(pte_folio);
5156 					ret = PTR_ERR(new_folio);
5157 					break;
5158 				}
5159 				ret = copy_user_large_folio(new_folio,
5160 							    pte_folio,
5161 							    addr, dst_vma);
5162 				folio_put(pte_folio);
5163 				if (ret) {
5164 					folio_put(new_folio);
5165 					break;
5166 				}
5167 
5168 				/* Install the new hugetlb folio if src pte stable */
5169 				dst_ptl = huge_pte_lock(h, dst, dst_pte);
5170 				src_ptl = huge_pte_lockptr(h, src, src_pte);
5171 				spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5172 				entry = huge_ptep_get(src_pte);
5173 				if (!pte_same(src_pte_old, entry)) {
5174 					restore_reserve_on_error(h, dst_vma, addr,
5175 								new_folio);
5176 					folio_put(new_folio);
5177 					/* huge_ptep of dst_pte won't change as in child */
5178 					goto again;
5179 				}
5180 				hugetlb_install_folio(dst_vma, dst_pte, addr,
5181 						      new_folio, src_pte_old, sz);
5182 				spin_unlock(src_ptl);
5183 				spin_unlock(dst_ptl);
5184 				continue;
5185 			}
5186 
5187 			if (cow) {
5188 				/*
5189 				 * No need to notify as we are downgrading page
5190 				 * table protection not changing it to point
5191 				 * to a new page.
5192 				 *
5193 				 * See Documentation/mm/mmu_notifier.rst
5194 				 */
5195 				huge_ptep_set_wrprotect(src, addr, src_pte);
5196 				entry = huge_pte_wrprotect(entry);
5197 			}
5198 
5199 			if (!userfaultfd_wp(dst_vma))
5200 				entry = huge_pte_clear_uffd_wp(entry);
5201 
5202 			set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5203 			hugetlb_count_add(npages, dst);
5204 		}
5205 		spin_unlock(src_ptl);
5206 		spin_unlock(dst_ptl);
5207 	}
5208 
5209 	if (cow) {
5210 		raw_write_seqcount_end(&src->write_protect_seq);
5211 		mmu_notifier_invalidate_range_end(&range);
5212 	} else {
5213 		hugetlb_vma_unlock_read(src_vma);
5214 	}
5215 
5216 	return ret;
5217 }
5218 
5219 static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
5220 			  unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte,
5221 			  unsigned long sz)
5222 {
5223 	struct hstate *h = hstate_vma(vma);
5224 	struct mm_struct *mm = vma->vm_mm;
5225 	spinlock_t *src_ptl, *dst_ptl;
5226 	pte_t pte;
5227 
5228 	dst_ptl = huge_pte_lock(h, mm, dst_pte);
5229 	src_ptl = huge_pte_lockptr(h, mm, src_pte);
5230 
5231 	/*
5232 	 * We don't have to worry about the ordering of src and dst ptlocks
5233 	 * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock.
5234 	 */
5235 	if (src_ptl != dst_ptl)
5236 		spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5237 
5238 	pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);
5239 	set_huge_pte_at(mm, new_addr, dst_pte, pte, sz);
5240 
5241 	if (src_ptl != dst_ptl)
5242 		spin_unlock(src_ptl);
5243 	spin_unlock(dst_ptl);
5244 }
5245 
5246 int move_hugetlb_page_tables(struct vm_area_struct *vma,
5247 			     struct vm_area_struct *new_vma,
5248 			     unsigned long old_addr, unsigned long new_addr,
5249 			     unsigned long len)
5250 {
5251 	struct hstate *h = hstate_vma(vma);
5252 	struct address_space *mapping = vma->vm_file->f_mapping;
5253 	unsigned long sz = huge_page_size(h);
5254 	struct mm_struct *mm = vma->vm_mm;
5255 	unsigned long old_end = old_addr + len;
5256 	unsigned long last_addr_mask;
5257 	pte_t *src_pte, *dst_pte;
5258 	struct mmu_notifier_range range;
5259 	bool shared_pmd = false;
5260 
5261 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, old_addr,
5262 				old_end);
5263 	adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5264 	/*
5265 	 * In case of shared PMDs, we should cover the maximum possible
5266 	 * range.
5267 	 */
5268 	flush_cache_range(vma, range.start, range.end);
5269 
5270 	mmu_notifier_invalidate_range_start(&range);
5271 	last_addr_mask = hugetlb_mask_last_page(h);
5272 	/* Prevent race with file truncation */
5273 	hugetlb_vma_lock_write(vma);
5274 	i_mmap_lock_write(mapping);
5275 	for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
5276 		src_pte = hugetlb_walk(vma, old_addr, sz);
5277 		if (!src_pte) {
5278 			old_addr |= last_addr_mask;
5279 			new_addr |= last_addr_mask;
5280 			continue;
5281 		}
5282 		if (huge_pte_none(huge_ptep_get(src_pte)))
5283 			continue;
5284 
5285 		if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) {
5286 			shared_pmd = true;
5287 			old_addr |= last_addr_mask;
5288 			new_addr |= last_addr_mask;
5289 			continue;
5290 		}
5291 
5292 		dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz);
5293 		if (!dst_pte)
5294 			break;
5295 
5296 		move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte, sz);
5297 	}
5298 
5299 	if (shared_pmd)
5300 		flush_hugetlb_tlb_range(vma, range.start, range.end);
5301 	else
5302 		flush_hugetlb_tlb_range(vma, old_end - len, old_end);
5303 	mmu_notifier_invalidate_range_end(&range);
5304 	i_mmap_unlock_write(mapping);
5305 	hugetlb_vma_unlock_write(vma);
5306 
5307 	return len + old_addr - old_end;
5308 }
5309 
5310 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
5311 			    unsigned long start, unsigned long end,
5312 			    struct page *ref_page, zap_flags_t zap_flags)
5313 {
5314 	struct mm_struct *mm = vma->vm_mm;
5315 	unsigned long address;
5316 	pte_t *ptep;
5317 	pte_t pte;
5318 	spinlock_t *ptl;
5319 	struct page *page;
5320 	struct hstate *h = hstate_vma(vma);
5321 	unsigned long sz = huge_page_size(h);
5322 	unsigned long last_addr_mask;
5323 	bool force_flush = false;
5324 
5325 	WARN_ON(!is_vm_hugetlb_page(vma));
5326 	BUG_ON(start & ~huge_page_mask(h));
5327 	BUG_ON(end & ~huge_page_mask(h));
5328 
5329 	/*
5330 	 * This is a hugetlb vma, all the pte entries should point
5331 	 * to huge page.
5332 	 */
5333 	tlb_change_page_size(tlb, sz);
5334 	tlb_start_vma(tlb, vma);
5335 
5336 	last_addr_mask = hugetlb_mask_last_page(h);
5337 	address = start;
5338 	for (; address < end; address += sz) {
5339 		ptep = hugetlb_walk(vma, address, sz);
5340 		if (!ptep) {
5341 			address |= last_addr_mask;
5342 			continue;
5343 		}
5344 
5345 		ptl = huge_pte_lock(h, mm, ptep);
5346 		if (huge_pmd_unshare(mm, vma, address, ptep)) {
5347 			spin_unlock(ptl);
5348 			tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
5349 			force_flush = true;
5350 			address |= last_addr_mask;
5351 			continue;
5352 		}
5353 
5354 		pte = huge_ptep_get(ptep);
5355 		if (huge_pte_none(pte)) {
5356 			spin_unlock(ptl);
5357 			continue;
5358 		}
5359 
5360 		/*
5361 		 * Migrating hugepage or HWPoisoned hugepage is already
5362 		 * unmapped and its refcount is dropped, so just clear pte here.
5363 		 */
5364 		if (unlikely(!pte_present(pte))) {
5365 			/*
5366 			 * If the pte was wr-protected by uffd-wp in any of the
5367 			 * swap forms, meanwhile the caller does not want to
5368 			 * drop the uffd-wp bit in this zap, then replace the
5369 			 * pte with a marker.
5370 			 */
5371 			if (pte_swp_uffd_wp_any(pte) &&
5372 			    !(zap_flags & ZAP_FLAG_DROP_MARKER))
5373 				set_huge_pte_at(mm, address, ptep,
5374 						make_pte_marker(PTE_MARKER_UFFD_WP),
5375 						sz);
5376 			else
5377 				huge_pte_clear(mm, address, ptep, sz);
5378 			spin_unlock(ptl);
5379 			continue;
5380 		}
5381 
5382 		page = pte_page(pte);
5383 		/*
5384 		 * If a reference page is supplied, it is because a specific
5385 		 * page is being unmapped, not a range. Ensure the page we
5386 		 * are about to unmap is the actual page of interest.
5387 		 */
5388 		if (ref_page) {
5389 			if (page != ref_page) {
5390 				spin_unlock(ptl);
5391 				continue;
5392 			}
5393 			/*
5394 			 * Mark the VMA as having unmapped its page so that
5395 			 * future faults in this VMA will fail rather than
5396 			 * looking like data was lost
5397 			 */
5398 			set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
5399 		}
5400 
5401 		pte = huge_ptep_get_and_clear(mm, address, ptep);
5402 		tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
5403 		if (huge_pte_dirty(pte))
5404 			set_page_dirty(page);
5405 		/* Leave a uffd-wp pte marker if needed */
5406 		if (huge_pte_uffd_wp(pte) &&
5407 		    !(zap_flags & ZAP_FLAG_DROP_MARKER))
5408 			set_huge_pte_at(mm, address, ptep,
5409 					make_pte_marker(PTE_MARKER_UFFD_WP),
5410 					sz);
5411 		hugetlb_count_sub(pages_per_huge_page(h), mm);
5412 		page_remove_rmap(page, vma, true);
5413 
5414 		spin_unlock(ptl);
5415 		tlb_remove_page_size(tlb, page, huge_page_size(h));
5416 		/*
5417 		 * Bail out after unmapping reference page if supplied
5418 		 */
5419 		if (ref_page)
5420 			break;
5421 	}
5422 	tlb_end_vma(tlb, vma);
5423 
5424 	/*
5425 	 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
5426 	 * could defer the flush until now, since by holding i_mmap_rwsem we
5427 	 * guaranteed that the last refernece would not be dropped. But we must
5428 	 * do the flushing before we return, as otherwise i_mmap_rwsem will be
5429 	 * dropped and the last reference to the shared PMDs page might be
5430 	 * dropped as well.
5431 	 *
5432 	 * In theory we could defer the freeing of the PMD pages as well, but
5433 	 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
5434 	 * detect sharing, so we cannot defer the release of the page either.
5435 	 * Instead, do flush now.
5436 	 */
5437 	if (force_flush)
5438 		tlb_flush_mmu_tlbonly(tlb);
5439 }
5440 
5441 void __hugetlb_zap_begin(struct vm_area_struct *vma,
5442 			 unsigned long *start, unsigned long *end)
5443 {
5444 	if (!vma->vm_file)	/* hugetlbfs_file_mmap error */
5445 		return;
5446 
5447 	adjust_range_if_pmd_sharing_possible(vma, start, end);
5448 	hugetlb_vma_lock_write(vma);
5449 	if (vma->vm_file)
5450 		i_mmap_lock_write(vma->vm_file->f_mapping);
5451 }
5452 
5453 void __hugetlb_zap_end(struct vm_area_struct *vma,
5454 		       struct zap_details *details)
5455 {
5456 	zap_flags_t zap_flags = details ? details->zap_flags : 0;
5457 
5458 	if (!vma->vm_file)	/* hugetlbfs_file_mmap error */
5459 		return;
5460 
5461 	if (zap_flags & ZAP_FLAG_UNMAP) {	/* final unmap */
5462 		/*
5463 		 * Unlock and free the vma lock before releasing i_mmap_rwsem.
5464 		 * When the vma_lock is freed, this makes the vma ineligible
5465 		 * for pmd sharing.  And, i_mmap_rwsem is required to set up
5466 		 * pmd sharing.  This is important as page tables for this
5467 		 * unmapped range will be asynchrously deleted.  If the page
5468 		 * tables are shared, there will be issues when accessed by
5469 		 * someone else.
5470 		 */
5471 		__hugetlb_vma_unlock_write_free(vma);
5472 	} else {
5473 		hugetlb_vma_unlock_write(vma);
5474 	}
5475 
5476 	if (vma->vm_file)
5477 		i_mmap_unlock_write(vma->vm_file->f_mapping);
5478 }
5479 
5480 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
5481 			  unsigned long end, struct page *ref_page,
5482 			  zap_flags_t zap_flags)
5483 {
5484 	struct mmu_notifier_range range;
5485 	struct mmu_gather tlb;
5486 
5487 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
5488 				start, end);
5489 	adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5490 	mmu_notifier_invalidate_range_start(&range);
5491 	tlb_gather_mmu(&tlb, vma->vm_mm);
5492 
5493 	__unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags);
5494 
5495 	mmu_notifier_invalidate_range_end(&range);
5496 	tlb_finish_mmu(&tlb);
5497 }
5498 
5499 /*
5500  * This is called when the original mapper is failing to COW a MAP_PRIVATE
5501  * mapping it owns the reserve page for. The intention is to unmap the page
5502  * from other VMAs and let the children be SIGKILLed if they are faulting the
5503  * same region.
5504  */
5505 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
5506 			      struct page *page, unsigned long address)
5507 {
5508 	struct hstate *h = hstate_vma(vma);
5509 	struct vm_area_struct *iter_vma;
5510 	struct address_space *mapping;
5511 	pgoff_t pgoff;
5512 
5513 	/*
5514 	 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
5515 	 * from page cache lookup which is in HPAGE_SIZE units.
5516 	 */
5517 	address = address & huge_page_mask(h);
5518 	pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
5519 			vma->vm_pgoff;
5520 	mapping = vma->vm_file->f_mapping;
5521 
5522 	/*
5523 	 * Take the mapping lock for the duration of the table walk. As
5524 	 * this mapping should be shared between all the VMAs,
5525 	 * __unmap_hugepage_range() is called as the lock is already held
5526 	 */
5527 	i_mmap_lock_write(mapping);
5528 	vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
5529 		/* Do not unmap the current VMA */
5530 		if (iter_vma == vma)
5531 			continue;
5532 
5533 		/*
5534 		 * Shared VMAs have their own reserves and do not affect
5535 		 * MAP_PRIVATE accounting but it is possible that a shared
5536 		 * VMA is using the same page so check and skip such VMAs.
5537 		 */
5538 		if (iter_vma->vm_flags & VM_MAYSHARE)
5539 			continue;
5540 
5541 		/*
5542 		 * Unmap the page from other VMAs without their own reserves.
5543 		 * They get marked to be SIGKILLed if they fault in these
5544 		 * areas. This is because a future no-page fault on this VMA
5545 		 * could insert a zeroed page instead of the data existing
5546 		 * from the time of fork. This would look like data corruption
5547 		 */
5548 		if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
5549 			unmap_hugepage_range(iter_vma, address,
5550 					     address + huge_page_size(h), page, 0);
5551 	}
5552 	i_mmap_unlock_write(mapping);
5553 }
5554 
5555 /*
5556  * hugetlb_wp() should be called with page lock of the original hugepage held.
5557  * Called with hugetlb_fault_mutex_table held and pte_page locked so we
5558  * cannot race with other handlers or page migration.
5559  * Keep the pte_same checks anyway to make transition from the mutex easier.
5560  */
5561 static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma,
5562 		       unsigned long address, pte_t *ptep, unsigned int flags,
5563 		       struct folio *pagecache_folio, spinlock_t *ptl)
5564 {
5565 	const bool unshare = flags & FAULT_FLAG_UNSHARE;
5566 	pte_t pte = huge_ptep_get(ptep);
5567 	struct hstate *h = hstate_vma(vma);
5568 	struct folio *old_folio;
5569 	struct folio *new_folio;
5570 	int outside_reserve = 0;
5571 	vm_fault_t ret = 0;
5572 	unsigned long haddr = address & huge_page_mask(h);
5573 	struct mmu_notifier_range range;
5574 
5575 	/*
5576 	 * Never handle CoW for uffd-wp protected pages.  It should be only
5577 	 * handled when the uffd-wp protection is removed.
5578 	 *
5579 	 * Note that only the CoW optimization path (in hugetlb_no_page())
5580 	 * can trigger this, because hugetlb_fault() will always resolve
5581 	 * uffd-wp bit first.
5582 	 */
5583 	if (!unshare && huge_pte_uffd_wp(pte))
5584 		return 0;
5585 
5586 	/*
5587 	 * hugetlb does not support FOLL_FORCE-style write faults that keep the
5588 	 * PTE mapped R/O such as maybe_mkwrite() would do.
5589 	 */
5590 	if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE)))
5591 		return VM_FAULT_SIGSEGV;
5592 
5593 	/* Let's take out MAP_SHARED mappings first. */
5594 	if (vma->vm_flags & VM_MAYSHARE) {
5595 		set_huge_ptep_writable(vma, haddr, ptep);
5596 		return 0;
5597 	}
5598 
5599 	old_folio = page_folio(pte_page(pte));
5600 
5601 	delayacct_wpcopy_start();
5602 
5603 retry_avoidcopy:
5604 	/*
5605 	 * If no-one else is actually using this page, we're the exclusive
5606 	 * owner and can reuse this page.
5607 	 */
5608 	if (folio_mapcount(old_folio) == 1 && folio_test_anon(old_folio)) {
5609 		if (!PageAnonExclusive(&old_folio->page))
5610 			page_move_anon_rmap(&old_folio->page, vma);
5611 		if (likely(!unshare))
5612 			set_huge_ptep_writable(vma, haddr, ptep);
5613 
5614 		delayacct_wpcopy_end();
5615 		return 0;
5616 	}
5617 	VM_BUG_ON_PAGE(folio_test_anon(old_folio) &&
5618 		       PageAnonExclusive(&old_folio->page), &old_folio->page);
5619 
5620 	/*
5621 	 * If the process that created a MAP_PRIVATE mapping is about to
5622 	 * perform a COW due to a shared page count, attempt to satisfy
5623 	 * the allocation without using the existing reserves. The pagecache
5624 	 * page is used to determine if the reserve at this address was
5625 	 * consumed or not. If reserves were used, a partial faulted mapping
5626 	 * at the time of fork() could consume its reserves on COW instead
5627 	 * of the full address range.
5628 	 */
5629 	if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
5630 			old_folio != pagecache_folio)
5631 		outside_reserve = 1;
5632 
5633 	folio_get(old_folio);
5634 
5635 	/*
5636 	 * Drop page table lock as buddy allocator may be called. It will
5637 	 * be acquired again before returning to the caller, as expected.
5638 	 */
5639 	spin_unlock(ptl);
5640 	new_folio = alloc_hugetlb_folio(vma, haddr, outside_reserve);
5641 
5642 	if (IS_ERR(new_folio)) {
5643 		/*
5644 		 * If a process owning a MAP_PRIVATE mapping fails to COW,
5645 		 * it is due to references held by a child and an insufficient
5646 		 * huge page pool. To guarantee the original mappers
5647 		 * reliability, unmap the page from child processes. The child
5648 		 * may get SIGKILLed if it later faults.
5649 		 */
5650 		if (outside_reserve) {
5651 			struct address_space *mapping = vma->vm_file->f_mapping;
5652 			pgoff_t idx;
5653 			u32 hash;
5654 
5655 			folio_put(old_folio);
5656 			/*
5657 			 * Drop hugetlb_fault_mutex and vma_lock before
5658 			 * unmapping.  unmapping needs to hold vma_lock
5659 			 * in write mode.  Dropping vma_lock in read mode
5660 			 * here is OK as COW mappings do not interact with
5661 			 * PMD sharing.
5662 			 *
5663 			 * Reacquire both after unmap operation.
5664 			 */
5665 			idx = vma_hugecache_offset(h, vma, haddr);
5666 			hash = hugetlb_fault_mutex_hash(mapping, idx);
5667 			hugetlb_vma_unlock_read(vma);
5668 			mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5669 
5670 			unmap_ref_private(mm, vma, &old_folio->page, haddr);
5671 
5672 			mutex_lock(&hugetlb_fault_mutex_table[hash]);
5673 			hugetlb_vma_lock_read(vma);
5674 			spin_lock(ptl);
5675 			ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
5676 			if (likely(ptep &&
5677 				   pte_same(huge_ptep_get(ptep), pte)))
5678 				goto retry_avoidcopy;
5679 			/*
5680 			 * race occurs while re-acquiring page table
5681 			 * lock, and our job is done.
5682 			 */
5683 			delayacct_wpcopy_end();
5684 			return 0;
5685 		}
5686 
5687 		ret = vmf_error(PTR_ERR(new_folio));
5688 		goto out_release_old;
5689 	}
5690 
5691 	/*
5692 	 * When the original hugepage is shared one, it does not have
5693 	 * anon_vma prepared.
5694 	 */
5695 	if (unlikely(anon_vma_prepare(vma))) {
5696 		ret = VM_FAULT_OOM;
5697 		goto out_release_all;
5698 	}
5699 
5700 	if (copy_user_large_folio(new_folio, old_folio, address, vma)) {
5701 		ret = VM_FAULT_HWPOISON_LARGE;
5702 		goto out_release_all;
5703 	}
5704 	__folio_mark_uptodate(new_folio);
5705 
5706 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, haddr,
5707 				haddr + huge_page_size(h));
5708 	mmu_notifier_invalidate_range_start(&range);
5709 
5710 	/*
5711 	 * Retake the page table lock to check for racing updates
5712 	 * before the page tables are altered
5713 	 */
5714 	spin_lock(ptl);
5715 	ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
5716 	if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
5717 		pte_t newpte = make_huge_pte(vma, &new_folio->page, !unshare);
5718 
5719 		/* Break COW or unshare */
5720 		huge_ptep_clear_flush(vma, haddr, ptep);
5721 		page_remove_rmap(&old_folio->page, vma, true);
5722 		hugepage_add_new_anon_rmap(new_folio, vma, haddr);
5723 		if (huge_pte_uffd_wp(pte))
5724 			newpte = huge_pte_mkuffd_wp(newpte);
5725 		set_huge_pte_at(mm, haddr, ptep, newpte, huge_page_size(h));
5726 		folio_set_hugetlb_migratable(new_folio);
5727 		/* Make the old page be freed below */
5728 		new_folio = old_folio;
5729 	}
5730 	spin_unlock(ptl);
5731 	mmu_notifier_invalidate_range_end(&range);
5732 out_release_all:
5733 	/*
5734 	 * No restore in case of successful pagetable update (Break COW or
5735 	 * unshare)
5736 	 */
5737 	if (new_folio != old_folio)
5738 		restore_reserve_on_error(h, vma, haddr, new_folio);
5739 	folio_put(new_folio);
5740 out_release_old:
5741 	folio_put(old_folio);
5742 
5743 	spin_lock(ptl); /* Caller expects lock to be held */
5744 
5745 	delayacct_wpcopy_end();
5746 	return ret;
5747 }
5748 
5749 /*
5750  * Return whether there is a pagecache page to back given address within VMA.
5751  */
5752 static bool hugetlbfs_pagecache_present(struct hstate *h,
5753 			struct vm_area_struct *vma, unsigned long address)
5754 {
5755 	struct address_space *mapping = vma->vm_file->f_mapping;
5756 	pgoff_t idx = vma_hugecache_offset(h, vma, address);
5757 	struct folio *folio;
5758 
5759 	folio = filemap_get_folio(mapping, idx);
5760 	if (IS_ERR(folio))
5761 		return false;
5762 	folio_put(folio);
5763 	return true;
5764 }
5765 
5766 int hugetlb_add_to_page_cache(struct folio *folio, struct address_space *mapping,
5767 			   pgoff_t idx)
5768 {
5769 	struct inode *inode = mapping->host;
5770 	struct hstate *h = hstate_inode(inode);
5771 	int err;
5772 
5773 	__folio_set_locked(folio);
5774 	err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL);
5775 
5776 	if (unlikely(err)) {
5777 		__folio_clear_locked(folio);
5778 		return err;
5779 	}
5780 	folio_clear_hugetlb_restore_reserve(folio);
5781 
5782 	/*
5783 	 * mark folio dirty so that it will not be removed from cache/file
5784 	 * by non-hugetlbfs specific code paths.
5785 	 */
5786 	folio_mark_dirty(folio);
5787 
5788 	spin_lock(&inode->i_lock);
5789 	inode->i_blocks += blocks_per_huge_page(h);
5790 	spin_unlock(&inode->i_lock);
5791 	return 0;
5792 }
5793 
5794 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
5795 						  struct address_space *mapping,
5796 						  pgoff_t idx,
5797 						  unsigned int flags,
5798 						  unsigned long haddr,
5799 						  unsigned long addr,
5800 						  unsigned long reason)
5801 {
5802 	u32 hash;
5803 	struct vm_fault vmf = {
5804 		.vma = vma,
5805 		.address = haddr,
5806 		.real_address = addr,
5807 		.flags = flags,
5808 
5809 		/*
5810 		 * Hard to debug if it ends up being
5811 		 * used by a callee that assumes
5812 		 * something about the other
5813 		 * uninitialized fields... same as in
5814 		 * memory.c
5815 		 */
5816 	};
5817 
5818 	/*
5819 	 * vma_lock and hugetlb_fault_mutex must be dropped before handling
5820 	 * userfault. Also mmap_lock could be dropped due to handling
5821 	 * userfault, any vma operation should be careful from here.
5822 	 */
5823 	hugetlb_vma_unlock_read(vma);
5824 	hash = hugetlb_fault_mutex_hash(mapping, idx);
5825 	mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5826 	return handle_userfault(&vmf, reason);
5827 }
5828 
5829 /*
5830  * Recheck pte with pgtable lock.  Returns true if pte didn't change, or
5831  * false if pte changed or is changing.
5832  */
5833 static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm,
5834 			       pte_t *ptep, pte_t old_pte)
5835 {
5836 	spinlock_t *ptl;
5837 	bool same;
5838 
5839 	ptl = huge_pte_lock(h, mm, ptep);
5840 	same = pte_same(huge_ptep_get(ptep), old_pte);
5841 	spin_unlock(ptl);
5842 
5843 	return same;
5844 }
5845 
5846 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
5847 			struct vm_area_struct *vma,
5848 			struct address_space *mapping, pgoff_t idx,
5849 			unsigned long address, pte_t *ptep,
5850 			pte_t old_pte, unsigned int flags)
5851 {
5852 	struct hstate *h = hstate_vma(vma);
5853 	vm_fault_t ret = VM_FAULT_SIGBUS;
5854 	int anon_rmap = 0;
5855 	unsigned long size;
5856 	struct folio *folio;
5857 	pte_t new_pte;
5858 	spinlock_t *ptl;
5859 	unsigned long haddr = address & huge_page_mask(h);
5860 	bool new_folio, new_pagecache_folio = false;
5861 	u32 hash = hugetlb_fault_mutex_hash(mapping, idx);
5862 
5863 	/*
5864 	 * Currently, we are forced to kill the process in the event the
5865 	 * original mapper has unmapped pages from the child due to a failed
5866 	 * COW/unsharing. Warn that such a situation has occurred as it may not
5867 	 * be obvious.
5868 	 */
5869 	if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
5870 		pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
5871 			   current->pid);
5872 		goto out;
5873 	}
5874 
5875 	/*
5876 	 * Use page lock to guard against racing truncation
5877 	 * before we get page_table_lock.
5878 	 */
5879 	new_folio = false;
5880 	folio = filemap_lock_folio(mapping, idx);
5881 	if (IS_ERR(folio)) {
5882 		size = i_size_read(mapping->host) >> huge_page_shift(h);
5883 		if (idx >= size)
5884 			goto out;
5885 		/* Check for page in userfault range */
5886 		if (userfaultfd_missing(vma)) {
5887 			/*
5888 			 * Since hugetlb_no_page() was examining pte
5889 			 * without pgtable lock, we need to re-test under
5890 			 * lock because the pte may not be stable and could
5891 			 * have changed from under us.  Try to detect
5892 			 * either changed or during-changing ptes and retry
5893 			 * properly when needed.
5894 			 *
5895 			 * Note that userfaultfd is actually fine with
5896 			 * false positives (e.g. caused by pte changed),
5897 			 * but not wrong logical events (e.g. caused by
5898 			 * reading a pte during changing).  The latter can
5899 			 * confuse the userspace, so the strictness is very
5900 			 * much preferred.  E.g., MISSING event should
5901 			 * never happen on the page after UFFDIO_COPY has
5902 			 * correctly installed the page and returned.
5903 			 */
5904 			if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5905 				ret = 0;
5906 				goto out;
5907 			}
5908 
5909 			return hugetlb_handle_userfault(vma, mapping, idx, flags,
5910 							haddr, address,
5911 							VM_UFFD_MISSING);
5912 		}
5913 
5914 		folio = alloc_hugetlb_folio(vma, haddr, 0);
5915 		if (IS_ERR(folio)) {
5916 			/*
5917 			 * Returning error will result in faulting task being
5918 			 * sent SIGBUS.  The hugetlb fault mutex prevents two
5919 			 * tasks from racing to fault in the same page which
5920 			 * could result in false unable to allocate errors.
5921 			 * Page migration does not take the fault mutex, but
5922 			 * does a clear then write of pte's under page table
5923 			 * lock.  Page fault code could race with migration,
5924 			 * notice the clear pte and try to allocate a page
5925 			 * here.  Before returning error, get ptl and make
5926 			 * sure there really is no pte entry.
5927 			 */
5928 			if (hugetlb_pte_stable(h, mm, ptep, old_pte))
5929 				ret = vmf_error(PTR_ERR(folio));
5930 			else
5931 				ret = 0;
5932 			goto out;
5933 		}
5934 		clear_huge_page(&folio->page, address, pages_per_huge_page(h));
5935 		__folio_mark_uptodate(folio);
5936 		new_folio = true;
5937 
5938 		if (vma->vm_flags & VM_MAYSHARE) {
5939 			int err = hugetlb_add_to_page_cache(folio, mapping, idx);
5940 			if (err) {
5941 				/*
5942 				 * err can't be -EEXIST which implies someone
5943 				 * else consumed the reservation since hugetlb
5944 				 * fault mutex is held when add a hugetlb page
5945 				 * to the page cache. So it's safe to call
5946 				 * restore_reserve_on_error() here.
5947 				 */
5948 				restore_reserve_on_error(h, vma, haddr, folio);
5949 				folio_put(folio);
5950 				goto out;
5951 			}
5952 			new_pagecache_folio = true;
5953 		} else {
5954 			folio_lock(folio);
5955 			if (unlikely(anon_vma_prepare(vma))) {
5956 				ret = VM_FAULT_OOM;
5957 				goto backout_unlocked;
5958 			}
5959 			anon_rmap = 1;
5960 		}
5961 	} else {
5962 		/*
5963 		 * If memory error occurs between mmap() and fault, some process
5964 		 * don't have hwpoisoned swap entry for errored virtual address.
5965 		 * So we need to block hugepage fault by PG_hwpoison bit check.
5966 		 */
5967 		if (unlikely(folio_test_hwpoison(folio))) {
5968 			ret = VM_FAULT_HWPOISON_LARGE |
5969 				VM_FAULT_SET_HINDEX(hstate_index(h));
5970 			goto backout_unlocked;
5971 		}
5972 
5973 		/* Check for page in userfault range. */
5974 		if (userfaultfd_minor(vma)) {
5975 			folio_unlock(folio);
5976 			folio_put(folio);
5977 			/* See comment in userfaultfd_missing() block above */
5978 			if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5979 				ret = 0;
5980 				goto out;
5981 			}
5982 			return hugetlb_handle_userfault(vma, mapping, idx, flags,
5983 							haddr, address,
5984 							VM_UFFD_MINOR);
5985 		}
5986 	}
5987 
5988 	/*
5989 	 * If we are going to COW a private mapping later, we examine the
5990 	 * pending reservations for this page now. This will ensure that
5991 	 * any allocations necessary to record that reservation occur outside
5992 	 * the spinlock.
5993 	 */
5994 	if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5995 		if (vma_needs_reservation(h, vma, haddr) < 0) {
5996 			ret = VM_FAULT_OOM;
5997 			goto backout_unlocked;
5998 		}
5999 		/* Just decrements count, does not deallocate */
6000 		vma_end_reservation(h, vma, haddr);
6001 	}
6002 
6003 	ptl = huge_pte_lock(h, mm, ptep);
6004 	ret = 0;
6005 	/* If pte changed from under us, retry */
6006 	if (!pte_same(huge_ptep_get(ptep), old_pte))
6007 		goto backout;
6008 
6009 	if (anon_rmap)
6010 		hugepage_add_new_anon_rmap(folio, vma, haddr);
6011 	else
6012 		page_dup_file_rmap(&folio->page, true);
6013 	new_pte = make_huge_pte(vma, &folio->page, ((vma->vm_flags & VM_WRITE)
6014 				&& (vma->vm_flags & VM_SHARED)));
6015 	/*
6016 	 * If this pte was previously wr-protected, keep it wr-protected even
6017 	 * if populated.
6018 	 */
6019 	if (unlikely(pte_marker_uffd_wp(old_pte)))
6020 		new_pte = huge_pte_mkuffd_wp(new_pte);
6021 	set_huge_pte_at(mm, haddr, ptep, new_pte, huge_page_size(h));
6022 
6023 	hugetlb_count_add(pages_per_huge_page(h), mm);
6024 	if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
6025 		/* Optimization, do the COW without a second fault */
6026 		ret = hugetlb_wp(mm, vma, address, ptep, flags, folio, ptl);
6027 	}
6028 
6029 	spin_unlock(ptl);
6030 
6031 	/*
6032 	 * Only set hugetlb_migratable in newly allocated pages.  Existing pages
6033 	 * found in the pagecache may not have hugetlb_migratable if they have
6034 	 * been isolated for migration.
6035 	 */
6036 	if (new_folio)
6037 		folio_set_hugetlb_migratable(folio);
6038 
6039 	folio_unlock(folio);
6040 out:
6041 	hugetlb_vma_unlock_read(vma);
6042 	mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6043 	return ret;
6044 
6045 backout:
6046 	spin_unlock(ptl);
6047 backout_unlocked:
6048 	if (new_folio && !new_pagecache_folio)
6049 		restore_reserve_on_error(h, vma, haddr, folio);
6050 
6051 	folio_unlock(folio);
6052 	folio_put(folio);
6053 	goto out;
6054 }
6055 
6056 #ifdef CONFIG_SMP
6057 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
6058 {
6059 	unsigned long key[2];
6060 	u32 hash;
6061 
6062 	key[0] = (unsigned long) mapping;
6063 	key[1] = idx;
6064 
6065 	hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
6066 
6067 	return hash & (num_fault_mutexes - 1);
6068 }
6069 #else
6070 /*
6071  * For uniprocessor systems we always use a single mutex, so just
6072  * return 0 and avoid the hashing overhead.
6073  */
6074 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
6075 {
6076 	return 0;
6077 }
6078 #endif
6079 
6080 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
6081 			unsigned long address, unsigned int flags)
6082 {
6083 	pte_t *ptep, entry;
6084 	spinlock_t *ptl;
6085 	vm_fault_t ret;
6086 	u32 hash;
6087 	pgoff_t idx;
6088 	struct folio *folio = NULL;
6089 	struct folio *pagecache_folio = NULL;
6090 	struct hstate *h = hstate_vma(vma);
6091 	struct address_space *mapping;
6092 	int need_wait_lock = 0;
6093 	unsigned long haddr = address & huge_page_mask(h);
6094 
6095 	/* TODO: Handle faults under the VMA lock */
6096 	if (flags & FAULT_FLAG_VMA_LOCK) {
6097 		vma_end_read(vma);
6098 		return VM_FAULT_RETRY;
6099 	}
6100 
6101 	/*
6102 	 * Serialize hugepage allocation and instantiation, so that we don't
6103 	 * get spurious allocation failures if two CPUs race to instantiate
6104 	 * the same page in the page cache.
6105 	 */
6106 	mapping = vma->vm_file->f_mapping;
6107 	idx = vma_hugecache_offset(h, vma, haddr);
6108 	hash = hugetlb_fault_mutex_hash(mapping, idx);
6109 	mutex_lock(&hugetlb_fault_mutex_table[hash]);
6110 
6111 	/*
6112 	 * Acquire vma lock before calling huge_pte_alloc and hold
6113 	 * until finished with ptep.  This prevents huge_pmd_unshare from
6114 	 * being called elsewhere and making the ptep no longer valid.
6115 	 */
6116 	hugetlb_vma_lock_read(vma);
6117 	ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
6118 	if (!ptep) {
6119 		hugetlb_vma_unlock_read(vma);
6120 		mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6121 		return VM_FAULT_OOM;
6122 	}
6123 
6124 	entry = huge_ptep_get(ptep);
6125 	if (huge_pte_none_mostly(entry)) {
6126 		if (is_pte_marker(entry)) {
6127 			pte_marker marker =
6128 				pte_marker_get(pte_to_swp_entry(entry));
6129 
6130 			if (marker & PTE_MARKER_POISONED) {
6131 				ret = VM_FAULT_HWPOISON_LARGE;
6132 				goto out_mutex;
6133 			}
6134 		}
6135 
6136 		/*
6137 		 * Other PTE markers should be handled the same way as none PTE.
6138 		 *
6139 		 * hugetlb_no_page will drop vma lock and hugetlb fault
6140 		 * mutex internally, which make us return immediately.
6141 		 */
6142 		return hugetlb_no_page(mm, vma, mapping, idx, address, ptep,
6143 				      entry, flags);
6144 	}
6145 
6146 	ret = 0;
6147 
6148 	/*
6149 	 * entry could be a migration/hwpoison entry at this point, so this
6150 	 * check prevents the kernel from going below assuming that we have
6151 	 * an active hugepage in pagecache. This goto expects the 2nd page
6152 	 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
6153 	 * properly handle it.
6154 	 */
6155 	if (!pte_present(entry)) {
6156 		if (unlikely(is_hugetlb_entry_migration(entry))) {
6157 			/*
6158 			 * Release the hugetlb fault lock now, but retain
6159 			 * the vma lock, because it is needed to guard the
6160 			 * huge_pte_lockptr() later in
6161 			 * migration_entry_wait_huge(). The vma lock will
6162 			 * be released there.
6163 			 */
6164 			mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6165 			migration_entry_wait_huge(vma, ptep);
6166 			return 0;
6167 		} else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
6168 			ret = VM_FAULT_HWPOISON_LARGE |
6169 			    VM_FAULT_SET_HINDEX(hstate_index(h));
6170 		goto out_mutex;
6171 	}
6172 
6173 	/*
6174 	 * If we are going to COW/unshare the mapping later, we examine the
6175 	 * pending reservations for this page now. This will ensure that any
6176 	 * allocations necessary to record that reservation occur outside the
6177 	 * spinlock. Also lookup the pagecache page now as it is used to
6178 	 * determine if a reservation has been consumed.
6179 	 */
6180 	if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
6181 	    !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) {
6182 		if (vma_needs_reservation(h, vma, haddr) < 0) {
6183 			ret = VM_FAULT_OOM;
6184 			goto out_mutex;
6185 		}
6186 		/* Just decrements count, does not deallocate */
6187 		vma_end_reservation(h, vma, haddr);
6188 
6189 		pagecache_folio = filemap_lock_folio(mapping, idx);
6190 		if (IS_ERR(pagecache_folio))
6191 			pagecache_folio = NULL;
6192 	}
6193 
6194 	ptl = huge_pte_lock(h, mm, ptep);
6195 
6196 	/* Check for a racing update before calling hugetlb_wp() */
6197 	if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
6198 		goto out_ptl;
6199 
6200 	/* Handle userfault-wp first, before trying to lock more pages */
6201 	if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) &&
6202 	    (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
6203 		struct vm_fault vmf = {
6204 			.vma = vma,
6205 			.address = haddr,
6206 			.real_address = address,
6207 			.flags = flags,
6208 		};
6209 
6210 		spin_unlock(ptl);
6211 		if (pagecache_folio) {
6212 			folio_unlock(pagecache_folio);
6213 			folio_put(pagecache_folio);
6214 		}
6215 		hugetlb_vma_unlock_read(vma);
6216 		mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6217 		return handle_userfault(&vmf, VM_UFFD_WP);
6218 	}
6219 
6220 	/*
6221 	 * hugetlb_wp() requires page locks of pte_page(entry) and
6222 	 * pagecache_folio, so here we need take the former one
6223 	 * when folio != pagecache_folio or !pagecache_folio.
6224 	 */
6225 	folio = page_folio(pte_page(entry));
6226 	if (folio != pagecache_folio)
6227 		if (!folio_trylock(folio)) {
6228 			need_wait_lock = 1;
6229 			goto out_ptl;
6230 		}
6231 
6232 	folio_get(folio);
6233 
6234 	if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
6235 		if (!huge_pte_write(entry)) {
6236 			ret = hugetlb_wp(mm, vma, address, ptep, flags,
6237 					 pagecache_folio, ptl);
6238 			goto out_put_page;
6239 		} else if (likely(flags & FAULT_FLAG_WRITE)) {
6240 			entry = huge_pte_mkdirty(entry);
6241 		}
6242 	}
6243 	entry = pte_mkyoung(entry);
6244 	if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
6245 						flags & FAULT_FLAG_WRITE))
6246 		update_mmu_cache(vma, haddr, ptep);
6247 out_put_page:
6248 	if (folio != pagecache_folio)
6249 		folio_unlock(folio);
6250 	folio_put(folio);
6251 out_ptl:
6252 	spin_unlock(ptl);
6253 
6254 	if (pagecache_folio) {
6255 		folio_unlock(pagecache_folio);
6256 		folio_put(pagecache_folio);
6257 	}
6258 out_mutex:
6259 	hugetlb_vma_unlock_read(vma);
6260 	mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6261 	/*
6262 	 * Generally it's safe to hold refcount during waiting page lock. But
6263 	 * here we just wait to defer the next page fault to avoid busy loop and
6264 	 * the page is not used after unlocked before returning from the current
6265 	 * page fault. So we are safe from accessing freed page, even if we wait
6266 	 * here without taking refcount.
6267 	 */
6268 	if (need_wait_lock)
6269 		folio_wait_locked(folio);
6270 	return ret;
6271 }
6272 
6273 #ifdef CONFIG_USERFAULTFD
6274 /*
6275  * Used by userfaultfd UFFDIO_* ioctls. Based on userfaultfd's mfill_atomic_pte
6276  * with modifications for hugetlb pages.
6277  */
6278 int hugetlb_mfill_atomic_pte(pte_t *dst_pte,
6279 			     struct vm_area_struct *dst_vma,
6280 			     unsigned long dst_addr,
6281 			     unsigned long src_addr,
6282 			     uffd_flags_t flags,
6283 			     struct folio **foliop)
6284 {
6285 	struct mm_struct *dst_mm = dst_vma->vm_mm;
6286 	bool is_continue = uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE);
6287 	bool wp_enabled = (flags & MFILL_ATOMIC_WP);
6288 	struct hstate *h = hstate_vma(dst_vma);
6289 	struct address_space *mapping = dst_vma->vm_file->f_mapping;
6290 	pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
6291 	unsigned long size;
6292 	int vm_shared = dst_vma->vm_flags & VM_SHARED;
6293 	pte_t _dst_pte;
6294 	spinlock_t *ptl;
6295 	int ret = -ENOMEM;
6296 	struct folio *folio;
6297 	int writable;
6298 	bool folio_in_pagecache = false;
6299 
6300 	if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) {
6301 		ptl = huge_pte_lock(h, dst_mm, dst_pte);
6302 
6303 		/* Don't overwrite any existing PTEs (even markers) */
6304 		if (!huge_pte_none(huge_ptep_get(dst_pte))) {
6305 			spin_unlock(ptl);
6306 			return -EEXIST;
6307 		}
6308 
6309 		_dst_pte = make_pte_marker(PTE_MARKER_POISONED);
6310 		set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte,
6311 				huge_page_size(h));
6312 
6313 		/* No need to invalidate - it was non-present before */
6314 		update_mmu_cache(dst_vma, dst_addr, dst_pte);
6315 
6316 		spin_unlock(ptl);
6317 		return 0;
6318 	}
6319 
6320 	if (is_continue) {
6321 		ret = -EFAULT;
6322 		folio = filemap_lock_folio(mapping, idx);
6323 		if (IS_ERR(folio))
6324 			goto out;
6325 		folio_in_pagecache = true;
6326 	} else if (!*foliop) {
6327 		/* If a folio already exists, then it's UFFDIO_COPY for
6328 		 * a non-missing case. Return -EEXIST.
6329 		 */
6330 		if (vm_shared &&
6331 		    hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6332 			ret = -EEXIST;
6333 			goto out;
6334 		}
6335 
6336 		folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);
6337 		if (IS_ERR(folio)) {
6338 			ret = -ENOMEM;
6339 			goto out;
6340 		}
6341 
6342 		ret = copy_folio_from_user(folio, (const void __user *) src_addr,
6343 					   false);
6344 
6345 		/* fallback to copy_from_user outside mmap_lock */
6346 		if (unlikely(ret)) {
6347 			ret = -ENOENT;
6348 			/* Free the allocated folio which may have
6349 			 * consumed a reservation.
6350 			 */
6351 			restore_reserve_on_error(h, dst_vma, dst_addr, folio);
6352 			folio_put(folio);
6353 
6354 			/* Allocate a temporary folio to hold the copied
6355 			 * contents.
6356 			 */
6357 			folio = alloc_hugetlb_folio_vma(h, dst_vma, dst_addr);
6358 			if (!folio) {
6359 				ret = -ENOMEM;
6360 				goto out;
6361 			}
6362 			*foliop = folio;
6363 			/* Set the outparam foliop and return to the caller to
6364 			 * copy the contents outside the lock. Don't free the
6365 			 * folio.
6366 			 */
6367 			goto out;
6368 		}
6369 	} else {
6370 		if (vm_shared &&
6371 		    hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6372 			folio_put(*foliop);
6373 			ret = -EEXIST;
6374 			*foliop = NULL;
6375 			goto out;
6376 		}
6377 
6378 		folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);
6379 		if (IS_ERR(folio)) {
6380 			folio_put(*foliop);
6381 			ret = -ENOMEM;
6382 			*foliop = NULL;
6383 			goto out;
6384 		}
6385 		ret = copy_user_large_folio(folio, *foliop, dst_addr, dst_vma);
6386 		folio_put(*foliop);
6387 		*foliop = NULL;
6388 		if (ret) {
6389 			folio_put(folio);
6390 			goto out;
6391 		}
6392 	}
6393 
6394 	/*
6395 	 * The memory barrier inside __folio_mark_uptodate makes sure that
6396 	 * preceding stores to the page contents become visible before
6397 	 * the set_pte_at() write.
6398 	 */
6399 	__folio_mark_uptodate(folio);
6400 
6401 	/* Add shared, newly allocated pages to the page cache. */
6402 	if (vm_shared && !is_continue) {
6403 		size = i_size_read(mapping->host) >> huge_page_shift(h);
6404 		ret = -EFAULT;
6405 		if (idx >= size)
6406 			goto out_release_nounlock;
6407 
6408 		/*
6409 		 * Serialization between remove_inode_hugepages() and
6410 		 * hugetlb_add_to_page_cache() below happens through the
6411 		 * hugetlb_fault_mutex_table that here must be hold by
6412 		 * the caller.
6413 		 */
6414 		ret = hugetlb_add_to_page_cache(folio, mapping, idx);
6415 		if (ret)
6416 			goto out_release_nounlock;
6417 		folio_in_pagecache = true;
6418 	}
6419 
6420 	ptl = huge_pte_lock(h, dst_mm, dst_pte);
6421 
6422 	ret = -EIO;
6423 	if (folio_test_hwpoison(folio))
6424 		goto out_release_unlock;
6425 
6426 	/*
6427 	 * We allow to overwrite a pte marker: consider when both MISSING|WP
6428 	 * registered, we firstly wr-protect a none pte which has no page cache
6429 	 * page backing it, then access the page.
6430 	 */
6431 	ret = -EEXIST;
6432 	if (!huge_pte_none_mostly(huge_ptep_get(dst_pte)))
6433 		goto out_release_unlock;
6434 
6435 	if (folio_in_pagecache)
6436 		page_dup_file_rmap(&folio->page, true);
6437 	else
6438 		hugepage_add_new_anon_rmap(folio, dst_vma, dst_addr);
6439 
6440 	/*
6441 	 * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY
6442 	 * with wp flag set, don't set pte write bit.
6443 	 */
6444 	if (wp_enabled || (is_continue && !vm_shared))
6445 		writable = 0;
6446 	else
6447 		writable = dst_vma->vm_flags & VM_WRITE;
6448 
6449 	_dst_pte = make_huge_pte(dst_vma, &folio->page, writable);
6450 	/*
6451 	 * Always mark UFFDIO_COPY page dirty; note that this may not be
6452 	 * extremely important for hugetlbfs for now since swapping is not
6453 	 * supported, but we should still be clear in that this page cannot be
6454 	 * thrown away at will, even if write bit not set.
6455 	 */
6456 	_dst_pte = huge_pte_mkdirty(_dst_pte);
6457 	_dst_pte = pte_mkyoung(_dst_pte);
6458 
6459 	if (wp_enabled)
6460 		_dst_pte = huge_pte_mkuffd_wp(_dst_pte);
6461 
6462 	set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte, huge_page_size(h));
6463 
6464 	hugetlb_count_add(pages_per_huge_page(h), dst_mm);
6465 
6466 	/* No need to invalidate - it was non-present before */
6467 	update_mmu_cache(dst_vma, dst_addr, dst_pte);
6468 
6469 	spin_unlock(ptl);
6470 	if (!is_continue)
6471 		folio_set_hugetlb_migratable(folio);
6472 	if (vm_shared || is_continue)
6473 		folio_unlock(folio);
6474 	ret = 0;
6475 out:
6476 	return ret;
6477 out_release_unlock:
6478 	spin_unlock(ptl);
6479 	if (vm_shared || is_continue)
6480 		folio_unlock(folio);
6481 out_release_nounlock:
6482 	if (!folio_in_pagecache)
6483 		restore_reserve_on_error(h, dst_vma, dst_addr, folio);
6484 	folio_put(folio);
6485 	goto out;
6486 }
6487 #endif /* CONFIG_USERFAULTFD */
6488 
6489 struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma,
6490 				      unsigned long address, unsigned int flags,
6491 				      unsigned int *page_mask)
6492 {
6493 	struct hstate *h = hstate_vma(vma);
6494 	struct mm_struct *mm = vma->vm_mm;
6495 	unsigned long haddr = address & huge_page_mask(h);
6496 	struct page *page = NULL;
6497 	spinlock_t *ptl;
6498 	pte_t *pte, entry;
6499 	int ret;
6500 
6501 	hugetlb_vma_lock_read(vma);
6502 	pte = hugetlb_walk(vma, haddr, huge_page_size(h));
6503 	if (!pte)
6504 		goto out_unlock;
6505 
6506 	ptl = huge_pte_lock(h, mm, pte);
6507 	entry = huge_ptep_get(pte);
6508 	if (pte_present(entry)) {
6509 		page = pte_page(entry);
6510 
6511 		if (!huge_pte_write(entry)) {
6512 			if (flags & FOLL_WRITE) {
6513 				page = NULL;
6514 				goto out;
6515 			}
6516 
6517 			if (gup_must_unshare(vma, flags, page)) {
6518 				/* Tell the caller to do unsharing */
6519 				page = ERR_PTR(-EMLINK);
6520 				goto out;
6521 			}
6522 		}
6523 
6524 		page = nth_page(page, ((address & ~huge_page_mask(h)) >> PAGE_SHIFT));
6525 
6526 		/*
6527 		 * Note that page may be a sub-page, and with vmemmap
6528 		 * optimizations the page struct may be read only.
6529 		 * try_grab_page() will increase the ref count on the
6530 		 * head page, so this will be OK.
6531 		 *
6532 		 * try_grab_page() should always be able to get the page here,
6533 		 * because we hold the ptl lock and have verified pte_present().
6534 		 */
6535 		ret = try_grab_page(page, flags);
6536 
6537 		if (WARN_ON_ONCE(ret)) {
6538 			page = ERR_PTR(ret);
6539 			goto out;
6540 		}
6541 
6542 		*page_mask = (1U << huge_page_order(h)) - 1;
6543 	}
6544 out:
6545 	spin_unlock(ptl);
6546 out_unlock:
6547 	hugetlb_vma_unlock_read(vma);
6548 
6549 	/*
6550 	 * Fixup retval for dump requests: if pagecache doesn't exist,
6551 	 * don't try to allocate a new page but just skip it.
6552 	 */
6553 	if (!page && (flags & FOLL_DUMP) &&
6554 	    !hugetlbfs_pagecache_present(h, vma, address))
6555 		page = ERR_PTR(-EFAULT);
6556 
6557 	return page;
6558 }
6559 
6560 long hugetlb_change_protection(struct vm_area_struct *vma,
6561 		unsigned long address, unsigned long end,
6562 		pgprot_t newprot, unsigned long cp_flags)
6563 {
6564 	struct mm_struct *mm = vma->vm_mm;
6565 	unsigned long start = address;
6566 	pte_t *ptep;
6567 	pte_t pte;
6568 	struct hstate *h = hstate_vma(vma);
6569 	long pages = 0, psize = huge_page_size(h);
6570 	bool shared_pmd = false;
6571 	struct mmu_notifier_range range;
6572 	unsigned long last_addr_mask;
6573 	bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
6574 	bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
6575 
6576 	/*
6577 	 * In the case of shared PMDs, the area to flush could be beyond
6578 	 * start/end.  Set range.start/range.end to cover the maximum possible
6579 	 * range if PMD sharing is possible.
6580 	 */
6581 	mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
6582 				0, mm, start, end);
6583 	adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
6584 
6585 	BUG_ON(address >= end);
6586 	flush_cache_range(vma, range.start, range.end);
6587 
6588 	mmu_notifier_invalidate_range_start(&range);
6589 	hugetlb_vma_lock_write(vma);
6590 	i_mmap_lock_write(vma->vm_file->f_mapping);
6591 	last_addr_mask = hugetlb_mask_last_page(h);
6592 	for (; address < end; address += psize) {
6593 		spinlock_t *ptl;
6594 		ptep = hugetlb_walk(vma, address, psize);
6595 		if (!ptep) {
6596 			if (!uffd_wp) {
6597 				address |= last_addr_mask;
6598 				continue;
6599 			}
6600 			/*
6601 			 * Userfaultfd wr-protect requires pgtable
6602 			 * pre-allocations to install pte markers.
6603 			 */
6604 			ptep = huge_pte_alloc(mm, vma, address, psize);
6605 			if (!ptep) {
6606 				pages = -ENOMEM;
6607 				break;
6608 			}
6609 		}
6610 		ptl = huge_pte_lock(h, mm, ptep);
6611 		if (huge_pmd_unshare(mm, vma, address, ptep)) {
6612 			/*
6613 			 * When uffd-wp is enabled on the vma, unshare
6614 			 * shouldn't happen at all.  Warn about it if it
6615 			 * happened due to some reason.
6616 			 */
6617 			WARN_ON_ONCE(uffd_wp || uffd_wp_resolve);
6618 			pages++;
6619 			spin_unlock(ptl);
6620 			shared_pmd = true;
6621 			address |= last_addr_mask;
6622 			continue;
6623 		}
6624 		pte = huge_ptep_get(ptep);
6625 		if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
6626 			/* Nothing to do. */
6627 		} else if (unlikely(is_hugetlb_entry_migration(pte))) {
6628 			swp_entry_t entry = pte_to_swp_entry(pte);
6629 			struct page *page = pfn_swap_entry_to_page(entry);
6630 			pte_t newpte = pte;
6631 
6632 			if (is_writable_migration_entry(entry)) {
6633 				if (PageAnon(page))
6634 					entry = make_readable_exclusive_migration_entry(
6635 								swp_offset(entry));
6636 				else
6637 					entry = make_readable_migration_entry(
6638 								swp_offset(entry));
6639 				newpte = swp_entry_to_pte(entry);
6640 				pages++;
6641 			}
6642 
6643 			if (uffd_wp)
6644 				newpte = pte_swp_mkuffd_wp(newpte);
6645 			else if (uffd_wp_resolve)
6646 				newpte = pte_swp_clear_uffd_wp(newpte);
6647 			if (!pte_same(pte, newpte))
6648 				set_huge_pte_at(mm, address, ptep, newpte, psize);
6649 		} else if (unlikely(is_pte_marker(pte))) {
6650 			/*
6651 			 * Do nothing on a poison marker; page is
6652 			 * corrupted, permissons do not apply.  Here
6653 			 * pte_marker_uffd_wp()==true implies !poison
6654 			 * because they're mutual exclusive.
6655 			 */
6656 			if (pte_marker_uffd_wp(pte) && uffd_wp_resolve)
6657 				/* Safe to modify directly (non-present->none). */
6658 				huge_pte_clear(mm, address, ptep, psize);
6659 		} else if (!huge_pte_none(pte)) {
6660 			pte_t old_pte;
6661 			unsigned int shift = huge_page_shift(hstate_vma(vma));
6662 
6663 			old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
6664 			pte = huge_pte_modify(old_pte, newprot);
6665 			pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
6666 			if (uffd_wp)
6667 				pte = huge_pte_mkuffd_wp(pte);
6668 			else if (uffd_wp_resolve)
6669 				pte = huge_pte_clear_uffd_wp(pte);
6670 			huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
6671 			pages++;
6672 		} else {
6673 			/* None pte */
6674 			if (unlikely(uffd_wp))
6675 				/* Safe to modify directly (none->non-present). */
6676 				set_huge_pte_at(mm, address, ptep,
6677 						make_pte_marker(PTE_MARKER_UFFD_WP),
6678 						psize);
6679 		}
6680 		spin_unlock(ptl);
6681 	}
6682 	/*
6683 	 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
6684 	 * may have cleared our pud entry and done put_page on the page table:
6685 	 * once we release i_mmap_rwsem, another task can do the final put_page
6686 	 * and that page table be reused and filled with junk.  If we actually
6687 	 * did unshare a page of pmds, flush the range corresponding to the pud.
6688 	 */
6689 	if (shared_pmd)
6690 		flush_hugetlb_tlb_range(vma, range.start, range.end);
6691 	else
6692 		flush_hugetlb_tlb_range(vma, start, end);
6693 	/*
6694 	 * No need to call mmu_notifier_arch_invalidate_secondary_tlbs() we are
6695 	 * downgrading page table protection not changing it to point to a new
6696 	 * page.
6697 	 *
6698 	 * See Documentation/mm/mmu_notifier.rst
6699 	 */
6700 	i_mmap_unlock_write(vma->vm_file->f_mapping);
6701 	hugetlb_vma_unlock_write(vma);
6702 	mmu_notifier_invalidate_range_end(&range);
6703 
6704 	return pages > 0 ? (pages << h->order) : pages;
6705 }
6706 
6707 /* Return true if reservation was successful, false otherwise.  */
6708 bool hugetlb_reserve_pages(struct inode *inode,
6709 					long from, long to,
6710 					struct vm_area_struct *vma,
6711 					vm_flags_t vm_flags)
6712 {
6713 	long chg = -1, add = -1;
6714 	struct hstate *h = hstate_inode(inode);
6715 	struct hugepage_subpool *spool = subpool_inode(inode);
6716 	struct resv_map *resv_map;
6717 	struct hugetlb_cgroup *h_cg = NULL;
6718 	long gbl_reserve, regions_needed = 0;
6719 
6720 	/* This should never happen */
6721 	if (from > to) {
6722 		VM_WARN(1, "%s called with a negative range\n", __func__);
6723 		return false;
6724 	}
6725 
6726 	/*
6727 	 * vma specific semaphore used for pmd sharing and fault/truncation
6728 	 * synchronization
6729 	 */
6730 	hugetlb_vma_lock_alloc(vma);
6731 
6732 	/*
6733 	 * Only apply hugepage reservation if asked. At fault time, an
6734 	 * attempt will be made for VM_NORESERVE to allocate a page
6735 	 * without using reserves
6736 	 */
6737 	if (vm_flags & VM_NORESERVE)
6738 		return true;
6739 
6740 	/*
6741 	 * Shared mappings base their reservation on the number of pages that
6742 	 * are already allocated on behalf of the file. Private mappings need
6743 	 * to reserve the full area even if read-only as mprotect() may be
6744 	 * called to make the mapping read-write. Assume !vma is a shm mapping
6745 	 */
6746 	if (!vma || vma->vm_flags & VM_MAYSHARE) {
6747 		/*
6748 		 * resv_map can not be NULL as hugetlb_reserve_pages is only
6749 		 * called for inodes for which resv_maps were created (see
6750 		 * hugetlbfs_get_inode).
6751 		 */
6752 		resv_map = inode_resv_map(inode);
6753 
6754 		chg = region_chg(resv_map, from, to, &regions_needed);
6755 	} else {
6756 		/* Private mapping. */
6757 		resv_map = resv_map_alloc();
6758 		if (!resv_map)
6759 			goto out_err;
6760 
6761 		chg = to - from;
6762 
6763 		set_vma_resv_map(vma, resv_map);
6764 		set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
6765 	}
6766 
6767 	if (chg < 0)
6768 		goto out_err;
6769 
6770 	if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
6771 				chg * pages_per_huge_page(h), &h_cg) < 0)
6772 		goto out_err;
6773 
6774 	if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
6775 		/* For private mappings, the hugetlb_cgroup uncharge info hangs
6776 		 * of the resv_map.
6777 		 */
6778 		resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
6779 	}
6780 
6781 	/*
6782 	 * There must be enough pages in the subpool for the mapping. If
6783 	 * the subpool has a minimum size, there may be some global
6784 	 * reservations already in place (gbl_reserve).
6785 	 */
6786 	gbl_reserve = hugepage_subpool_get_pages(spool, chg);
6787 	if (gbl_reserve < 0)
6788 		goto out_uncharge_cgroup;
6789 
6790 	/*
6791 	 * Check enough hugepages are available for the reservation.
6792 	 * Hand the pages back to the subpool if there are not
6793 	 */
6794 	if (hugetlb_acct_memory(h, gbl_reserve) < 0)
6795 		goto out_put_pages;
6796 
6797 	/*
6798 	 * Account for the reservations made. Shared mappings record regions
6799 	 * that have reservations as they are shared by multiple VMAs.
6800 	 * When the last VMA disappears, the region map says how much
6801 	 * the reservation was and the page cache tells how much of
6802 	 * the reservation was consumed. Private mappings are per-VMA and
6803 	 * only the consumed reservations are tracked. When the VMA
6804 	 * disappears, the original reservation is the VMA size and the
6805 	 * consumed reservations are stored in the map. Hence, nothing
6806 	 * else has to be done for private mappings here
6807 	 */
6808 	if (!vma || vma->vm_flags & VM_MAYSHARE) {
6809 		add = region_add(resv_map, from, to, regions_needed, h, h_cg);
6810 
6811 		if (unlikely(add < 0)) {
6812 			hugetlb_acct_memory(h, -gbl_reserve);
6813 			goto out_put_pages;
6814 		} else if (unlikely(chg > add)) {
6815 			/*
6816 			 * pages in this range were added to the reserve
6817 			 * map between region_chg and region_add.  This
6818 			 * indicates a race with alloc_hugetlb_folio.  Adjust
6819 			 * the subpool and reserve counts modified above
6820 			 * based on the difference.
6821 			 */
6822 			long rsv_adjust;
6823 
6824 			/*
6825 			 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
6826 			 * reference to h_cg->css. See comment below for detail.
6827 			 */
6828 			hugetlb_cgroup_uncharge_cgroup_rsvd(
6829 				hstate_index(h),
6830 				(chg - add) * pages_per_huge_page(h), h_cg);
6831 
6832 			rsv_adjust = hugepage_subpool_put_pages(spool,
6833 								chg - add);
6834 			hugetlb_acct_memory(h, -rsv_adjust);
6835 		} else if (h_cg) {
6836 			/*
6837 			 * The file_regions will hold their own reference to
6838 			 * h_cg->css. So we should release the reference held
6839 			 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
6840 			 * done.
6841 			 */
6842 			hugetlb_cgroup_put_rsvd_cgroup(h_cg);
6843 		}
6844 	}
6845 	return true;
6846 
6847 out_put_pages:
6848 	/* put back original number of pages, chg */
6849 	(void)hugepage_subpool_put_pages(spool, chg);
6850 out_uncharge_cgroup:
6851 	hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
6852 					    chg * pages_per_huge_page(h), h_cg);
6853 out_err:
6854 	hugetlb_vma_lock_free(vma);
6855 	if (!vma || vma->vm_flags & VM_MAYSHARE)
6856 		/* Only call region_abort if the region_chg succeeded but the
6857 		 * region_add failed or didn't run.
6858 		 */
6859 		if (chg >= 0 && add < 0)
6860 			region_abort(resv_map, from, to, regions_needed);
6861 	if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
6862 		kref_put(&resv_map->refs, resv_map_release);
6863 		set_vma_resv_map(vma, NULL);
6864 	}
6865 	return false;
6866 }
6867 
6868 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
6869 								long freed)
6870 {
6871 	struct hstate *h = hstate_inode(inode);
6872 	struct resv_map *resv_map = inode_resv_map(inode);
6873 	long chg = 0;
6874 	struct hugepage_subpool *spool = subpool_inode(inode);
6875 	long gbl_reserve;
6876 
6877 	/*
6878 	 * Since this routine can be called in the evict inode path for all
6879 	 * hugetlbfs inodes, resv_map could be NULL.
6880 	 */
6881 	if (resv_map) {
6882 		chg = region_del(resv_map, start, end);
6883 		/*
6884 		 * region_del() can fail in the rare case where a region
6885 		 * must be split and another region descriptor can not be
6886 		 * allocated.  If end == LONG_MAX, it will not fail.
6887 		 */
6888 		if (chg < 0)
6889 			return chg;
6890 	}
6891 
6892 	spin_lock(&inode->i_lock);
6893 	inode->i_blocks -= (blocks_per_huge_page(h) * freed);
6894 	spin_unlock(&inode->i_lock);
6895 
6896 	/*
6897 	 * If the subpool has a minimum size, the number of global
6898 	 * reservations to be released may be adjusted.
6899 	 *
6900 	 * Note that !resv_map implies freed == 0. So (chg - freed)
6901 	 * won't go negative.
6902 	 */
6903 	gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
6904 	hugetlb_acct_memory(h, -gbl_reserve);
6905 
6906 	return 0;
6907 }
6908 
6909 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
6910 static unsigned long page_table_shareable(struct vm_area_struct *svma,
6911 				struct vm_area_struct *vma,
6912 				unsigned long addr, pgoff_t idx)
6913 {
6914 	unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
6915 				svma->vm_start;
6916 	unsigned long sbase = saddr & PUD_MASK;
6917 	unsigned long s_end = sbase + PUD_SIZE;
6918 
6919 	/* Allow segments to share if only one is marked locked */
6920 	unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED_MASK;
6921 	unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED_MASK;
6922 
6923 	/*
6924 	 * match the virtual addresses, permission and the alignment of the
6925 	 * page table page.
6926 	 *
6927 	 * Also, vma_lock (vm_private_data) is required for sharing.
6928 	 */
6929 	if (pmd_index(addr) != pmd_index(saddr) ||
6930 	    vm_flags != svm_flags ||
6931 	    !range_in_vma(svma, sbase, s_end) ||
6932 	    !svma->vm_private_data)
6933 		return 0;
6934 
6935 	return saddr;
6936 }
6937 
6938 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6939 {
6940 	unsigned long start = addr & PUD_MASK;
6941 	unsigned long end = start + PUD_SIZE;
6942 
6943 #ifdef CONFIG_USERFAULTFD
6944 	if (uffd_disable_huge_pmd_share(vma))
6945 		return false;
6946 #endif
6947 	/*
6948 	 * check on proper vm_flags and page table alignment
6949 	 */
6950 	if (!(vma->vm_flags & VM_MAYSHARE))
6951 		return false;
6952 	if (!vma->vm_private_data)	/* vma lock required for sharing */
6953 		return false;
6954 	if (!range_in_vma(vma, start, end))
6955 		return false;
6956 	return true;
6957 }
6958 
6959 /*
6960  * Determine if start,end range within vma could be mapped by shared pmd.
6961  * If yes, adjust start and end to cover range associated with possible
6962  * shared pmd mappings.
6963  */
6964 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
6965 				unsigned long *start, unsigned long *end)
6966 {
6967 	unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
6968 		v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6969 
6970 	/*
6971 	 * vma needs to span at least one aligned PUD size, and the range
6972 	 * must be at least partially within in.
6973 	 */
6974 	if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
6975 		(*end <= v_start) || (*start >= v_end))
6976 		return;
6977 
6978 	/* Extend the range to be PUD aligned for a worst case scenario */
6979 	if (*start > v_start)
6980 		*start = ALIGN_DOWN(*start, PUD_SIZE);
6981 
6982 	if (*end < v_end)
6983 		*end = ALIGN(*end, PUD_SIZE);
6984 }
6985 
6986 /*
6987  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
6988  * and returns the corresponding pte. While this is not necessary for the
6989  * !shared pmd case because we can allocate the pmd later as well, it makes the
6990  * code much cleaner. pmd allocation is essential for the shared case because
6991  * pud has to be populated inside the same i_mmap_rwsem section - otherwise
6992  * racing tasks could either miss the sharing (see huge_pte_offset) or select a
6993  * bad pmd for sharing.
6994  */
6995 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
6996 		      unsigned long addr, pud_t *pud)
6997 {
6998 	struct address_space *mapping = vma->vm_file->f_mapping;
6999 	pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
7000 			vma->vm_pgoff;
7001 	struct vm_area_struct *svma;
7002 	unsigned long saddr;
7003 	pte_t *spte = NULL;
7004 	pte_t *pte;
7005 
7006 	i_mmap_lock_read(mapping);
7007 	vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
7008 		if (svma == vma)
7009 			continue;
7010 
7011 		saddr = page_table_shareable(svma, vma, addr, idx);
7012 		if (saddr) {
7013 			spte = hugetlb_walk(svma, saddr,
7014 					    vma_mmu_pagesize(svma));
7015 			if (spte) {
7016 				get_page(virt_to_page(spte));
7017 				break;
7018 			}
7019 		}
7020 	}
7021 
7022 	if (!spte)
7023 		goto out;
7024 
7025 	spin_lock(&mm->page_table_lock);
7026 	if (pud_none(*pud)) {
7027 		pud_populate(mm, pud,
7028 				(pmd_t *)((unsigned long)spte & PAGE_MASK));
7029 		mm_inc_nr_pmds(mm);
7030 	} else {
7031 		put_page(virt_to_page(spte));
7032 	}
7033 	spin_unlock(&mm->page_table_lock);
7034 out:
7035 	pte = (pte_t *)pmd_alloc(mm, pud, addr);
7036 	i_mmap_unlock_read(mapping);
7037 	return pte;
7038 }
7039 
7040 /*
7041  * unmap huge page backed by shared pte.
7042  *
7043  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
7044  * indicated by page_count > 1, unmap is achieved by clearing pud and
7045  * decrementing the ref count. If count == 1, the pte page is not shared.
7046  *
7047  * Called with page table lock held.
7048  *
7049  * returns: 1 successfully unmapped a shared pte page
7050  *	    0 the underlying pte page is not shared, or it is the last user
7051  */
7052 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7053 					unsigned long addr, pte_t *ptep)
7054 {
7055 	pgd_t *pgd = pgd_offset(mm, addr);
7056 	p4d_t *p4d = p4d_offset(pgd, addr);
7057 	pud_t *pud = pud_offset(p4d, addr);
7058 
7059 	i_mmap_assert_write_locked(vma->vm_file->f_mapping);
7060 	hugetlb_vma_assert_locked(vma);
7061 	BUG_ON(page_count(virt_to_page(ptep)) == 0);
7062 	if (page_count(virt_to_page(ptep)) == 1)
7063 		return 0;
7064 
7065 	pud_clear(pud);
7066 	put_page(virt_to_page(ptep));
7067 	mm_dec_nr_pmds(mm);
7068 	return 1;
7069 }
7070 
7071 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7072 
7073 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7074 		      unsigned long addr, pud_t *pud)
7075 {
7076 	return NULL;
7077 }
7078 
7079 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7080 				unsigned long addr, pte_t *ptep)
7081 {
7082 	return 0;
7083 }
7084 
7085 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7086 				unsigned long *start, unsigned long *end)
7087 {
7088 }
7089 
7090 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
7091 {
7092 	return false;
7093 }
7094 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7095 
7096 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
7097 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
7098 			unsigned long addr, unsigned long sz)
7099 {
7100 	pgd_t *pgd;
7101 	p4d_t *p4d;
7102 	pud_t *pud;
7103 	pte_t *pte = NULL;
7104 
7105 	pgd = pgd_offset(mm, addr);
7106 	p4d = p4d_alloc(mm, pgd, addr);
7107 	if (!p4d)
7108 		return NULL;
7109 	pud = pud_alloc(mm, p4d, addr);
7110 	if (pud) {
7111 		if (sz == PUD_SIZE) {
7112 			pte = (pte_t *)pud;
7113 		} else {
7114 			BUG_ON(sz != PMD_SIZE);
7115 			if (want_pmd_share(vma, addr) && pud_none(*pud))
7116 				pte = huge_pmd_share(mm, vma, addr, pud);
7117 			else
7118 				pte = (pte_t *)pmd_alloc(mm, pud, addr);
7119 		}
7120 	}
7121 
7122 	if (pte) {
7123 		pte_t pteval = ptep_get_lockless(pte);
7124 
7125 		BUG_ON(pte_present(pteval) && !pte_huge(pteval));
7126 	}
7127 
7128 	return pte;
7129 }
7130 
7131 /*
7132  * huge_pte_offset() - Walk the page table to resolve the hugepage
7133  * entry at address @addr
7134  *
7135  * Return: Pointer to page table entry (PUD or PMD) for
7136  * address @addr, or NULL if a !p*d_present() entry is encountered and the
7137  * size @sz doesn't match the hugepage size at this level of the page
7138  * table.
7139  */
7140 pte_t *huge_pte_offset(struct mm_struct *mm,
7141 		       unsigned long addr, unsigned long sz)
7142 {
7143 	pgd_t *pgd;
7144 	p4d_t *p4d;
7145 	pud_t *pud;
7146 	pmd_t *pmd;
7147 
7148 	pgd = pgd_offset(mm, addr);
7149 	if (!pgd_present(*pgd))
7150 		return NULL;
7151 	p4d = p4d_offset(pgd, addr);
7152 	if (!p4d_present(*p4d))
7153 		return NULL;
7154 
7155 	pud = pud_offset(p4d, addr);
7156 	if (sz == PUD_SIZE)
7157 		/* must be pud huge, non-present or none */
7158 		return (pte_t *)pud;
7159 	if (!pud_present(*pud))
7160 		return NULL;
7161 	/* must have a valid entry and size to go further */
7162 
7163 	pmd = pmd_offset(pud, addr);
7164 	/* must be pmd huge, non-present or none */
7165 	return (pte_t *)pmd;
7166 }
7167 
7168 /*
7169  * Return a mask that can be used to update an address to the last huge
7170  * page in a page table page mapping size.  Used to skip non-present
7171  * page table entries when linearly scanning address ranges.  Architectures
7172  * with unique huge page to page table relationships can define their own
7173  * version of this routine.
7174  */
7175 unsigned long hugetlb_mask_last_page(struct hstate *h)
7176 {
7177 	unsigned long hp_size = huge_page_size(h);
7178 
7179 	if (hp_size == PUD_SIZE)
7180 		return P4D_SIZE - PUD_SIZE;
7181 	else if (hp_size == PMD_SIZE)
7182 		return PUD_SIZE - PMD_SIZE;
7183 	else
7184 		return 0UL;
7185 }
7186 
7187 #else
7188 
7189 /* See description above.  Architectures can provide their own version. */
7190 __weak unsigned long hugetlb_mask_last_page(struct hstate *h)
7191 {
7192 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
7193 	if (huge_page_size(h) == PMD_SIZE)
7194 		return PUD_SIZE - PMD_SIZE;
7195 #endif
7196 	return 0UL;
7197 }
7198 
7199 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
7200 
7201 /*
7202  * These functions are overwritable if your architecture needs its own
7203  * behavior.
7204  */
7205 bool isolate_hugetlb(struct folio *folio, struct list_head *list)
7206 {
7207 	bool ret = true;
7208 
7209 	spin_lock_irq(&hugetlb_lock);
7210 	if (!folio_test_hugetlb(folio) ||
7211 	    !folio_test_hugetlb_migratable(folio) ||
7212 	    !folio_try_get(folio)) {
7213 		ret = false;
7214 		goto unlock;
7215 	}
7216 	folio_clear_hugetlb_migratable(folio);
7217 	list_move_tail(&folio->lru, list);
7218 unlock:
7219 	spin_unlock_irq(&hugetlb_lock);
7220 	return ret;
7221 }
7222 
7223 int get_hwpoison_hugetlb_folio(struct folio *folio, bool *hugetlb, bool unpoison)
7224 {
7225 	int ret = 0;
7226 
7227 	*hugetlb = false;
7228 	spin_lock_irq(&hugetlb_lock);
7229 	if (folio_test_hugetlb(folio)) {
7230 		*hugetlb = true;
7231 		if (folio_test_hugetlb_freed(folio))
7232 			ret = 0;
7233 		else if (folio_test_hugetlb_migratable(folio) || unpoison)
7234 			ret = folio_try_get(folio);
7235 		else
7236 			ret = -EBUSY;
7237 	}
7238 	spin_unlock_irq(&hugetlb_lock);
7239 	return ret;
7240 }
7241 
7242 int get_huge_page_for_hwpoison(unsigned long pfn, int flags,
7243 				bool *migratable_cleared)
7244 {
7245 	int ret;
7246 
7247 	spin_lock_irq(&hugetlb_lock);
7248 	ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared);
7249 	spin_unlock_irq(&hugetlb_lock);
7250 	return ret;
7251 }
7252 
7253 void folio_putback_active_hugetlb(struct folio *folio)
7254 {
7255 	spin_lock_irq(&hugetlb_lock);
7256 	folio_set_hugetlb_migratable(folio);
7257 	list_move_tail(&folio->lru, &(folio_hstate(folio))->hugepage_activelist);
7258 	spin_unlock_irq(&hugetlb_lock);
7259 	folio_put(folio);
7260 }
7261 
7262 void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason)
7263 {
7264 	struct hstate *h = folio_hstate(old_folio);
7265 
7266 	hugetlb_cgroup_migrate(old_folio, new_folio);
7267 	set_page_owner_migrate_reason(&new_folio->page, reason);
7268 
7269 	/*
7270 	 * transfer temporary state of the new hugetlb folio. This is
7271 	 * reverse to other transitions because the newpage is going to
7272 	 * be final while the old one will be freed so it takes over
7273 	 * the temporary status.
7274 	 *
7275 	 * Also note that we have to transfer the per-node surplus state
7276 	 * here as well otherwise the global surplus count will not match
7277 	 * the per-node's.
7278 	 */
7279 	if (folio_test_hugetlb_temporary(new_folio)) {
7280 		int old_nid = folio_nid(old_folio);
7281 		int new_nid = folio_nid(new_folio);
7282 
7283 		folio_set_hugetlb_temporary(old_folio);
7284 		folio_clear_hugetlb_temporary(new_folio);
7285 
7286 
7287 		/*
7288 		 * There is no need to transfer the per-node surplus state
7289 		 * when we do not cross the node.
7290 		 */
7291 		if (new_nid == old_nid)
7292 			return;
7293 		spin_lock_irq(&hugetlb_lock);
7294 		if (h->surplus_huge_pages_node[old_nid]) {
7295 			h->surplus_huge_pages_node[old_nid]--;
7296 			h->surplus_huge_pages_node[new_nid]++;
7297 		}
7298 		spin_unlock_irq(&hugetlb_lock);
7299 	}
7300 }
7301 
7302 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
7303 				   unsigned long start,
7304 				   unsigned long end)
7305 {
7306 	struct hstate *h = hstate_vma(vma);
7307 	unsigned long sz = huge_page_size(h);
7308 	struct mm_struct *mm = vma->vm_mm;
7309 	struct mmu_notifier_range range;
7310 	unsigned long address;
7311 	spinlock_t *ptl;
7312 	pte_t *ptep;
7313 
7314 	if (!(vma->vm_flags & VM_MAYSHARE))
7315 		return;
7316 
7317 	if (start >= end)
7318 		return;
7319 
7320 	flush_cache_range(vma, start, end);
7321 	/*
7322 	 * No need to call adjust_range_if_pmd_sharing_possible(), because
7323 	 * we have already done the PUD_SIZE alignment.
7324 	 */
7325 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
7326 				start, end);
7327 	mmu_notifier_invalidate_range_start(&range);
7328 	hugetlb_vma_lock_write(vma);
7329 	i_mmap_lock_write(vma->vm_file->f_mapping);
7330 	for (address = start; address < end; address += PUD_SIZE) {
7331 		ptep = hugetlb_walk(vma, address, sz);
7332 		if (!ptep)
7333 			continue;
7334 		ptl = huge_pte_lock(h, mm, ptep);
7335 		huge_pmd_unshare(mm, vma, address, ptep);
7336 		spin_unlock(ptl);
7337 	}
7338 	flush_hugetlb_tlb_range(vma, start, end);
7339 	i_mmap_unlock_write(vma->vm_file->f_mapping);
7340 	hugetlb_vma_unlock_write(vma);
7341 	/*
7342 	 * No need to call mmu_notifier_arch_invalidate_secondary_tlbs(), see
7343 	 * Documentation/mm/mmu_notifier.rst.
7344 	 */
7345 	mmu_notifier_invalidate_range_end(&range);
7346 }
7347 
7348 /*
7349  * This function will unconditionally remove all the shared pmd pgtable entries
7350  * within the specific vma for a hugetlbfs memory range.
7351  */
7352 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
7353 {
7354 	hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE),
7355 			ALIGN_DOWN(vma->vm_end, PUD_SIZE));
7356 }
7357 
7358 #ifdef CONFIG_CMA
7359 static bool cma_reserve_called __initdata;
7360 
7361 static int __init cmdline_parse_hugetlb_cma(char *p)
7362 {
7363 	int nid, count = 0;
7364 	unsigned long tmp;
7365 	char *s = p;
7366 
7367 	while (*s) {
7368 		if (sscanf(s, "%lu%n", &tmp, &count) != 1)
7369 			break;
7370 
7371 		if (s[count] == ':') {
7372 			if (tmp >= MAX_NUMNODES)
7373 				break;
7374 			nid = array_index_nospec(tmp, MAX_NUMNODES);
7375 
7376 			s += count + 1;
7377 			tmp = memparse(s, &s);
7378 			hugetlb_cma_size_in_node[nid] = tmp;
7379 			hugetlb_cma_size += tmp;
7380 
7381 			/*
7382 			 * Skip the separator if have one, otherwise
7383 			 * break the parsing.
7384 			 */
7385 			if (*s == ',')
7386 				s++;
7387 			else
7388 				break;
7389 		} else {
7390 			hugetlb_cma_size = memparse(p, &p);
7391 			break;
7392 		}
7393 	}
7394 
7395 	return 0;
7396 }
7397 
7398 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
7399 
7400 void __init hugetlb_cma_reserve(int order)
7401 {
7402 	unsigned long size, reserved, per_node;
7403 	bool node_specific_cma_alloc = false;
7404 	int nid;
7405 
7406 	cma_reserve_called = true;
7407 
7408 	if (!hugetlb_cma_size)
7409 		return;
7410 
7411 	for (nid = 0; nid < MAX_NUMNODES; nid++) {
7412 		if (hugetlb_cma_size_in_node[nid] == 0)
7413 			continue;
7414 
7415 		if (!node_online(nid)) {
7416 			pr_warn("hugetlb_cma: invalid node %d specified\n", nid);
7417 			hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7418 			hugetlb_cma_size_in_node[nid] = 0;
7419 			continue;
7420 		}
7421 
7422 		if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) {
7423 			pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n",
7424 				nid, (PAGE_SIZE << order) / SZ_1M);
7425 			hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7426 			hugetlb_cma_size_in_node[nid] = 0;
7427 		} else {
7428 			node_specific_cma_alloc = true;
7429 		}
7430 	}
7431 
7432 	/* Validate the CMA size again in case some invalid nodes specified. */
7433 	if (!hugetlb_cma_size)
7434 		return;
7435 
7436 	if (hugetlb_cma_size < (PAGE_SIZE << order)) {
7437 		pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
7438 			(PAGE_SIZE << order) / SZ_1M);
7439 		hugetlb_cma_size = 0;
7440 		return;
7441 	}
7442 
7443 	if (!node_specific_cma_alloc) {
7444 		/*
7445 		 * If 3 GB area is requested on a machine with 4 numa nodes,
7446 		 * let's allocate 1 GB on first three nodes and ignore the last one.
7447 		 */
7448 		per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
7449 		pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
7450 			hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
7451 	}
7452 
7453 	reserved = 0;
7454 	for_each_online_node(nid) {
7455 		int res;
7456 		char name[CMA_MAX_NAME];
7457 
7458 		if (node_specific_cma_alloc) {
7459 			if (hugetlb_cma_size_in_node[nid] == 0)
7460 				continue;
7461 
7462 			size = hugetlb_cma_size_in_node[nid];
7463 		} else {
7464 			size = min(per_node, hugetlb_cma_size - reserved);
7465 		}
7466 
7467 		size = round_up(size, PAGE_SIZE << order);
7468 
7469 		snprintf(name, sizeof(name), "hugetlb%d", nid);
7470 		/*
7471 		 * Note that 'order per bit' is based on smallest size that
7472 		 * may be returned to CMA allocator in the case of
7473 		 * huge page demotion.
7474 		 */
7475 		res = cma_declare_contiguous_nid(0, size, 0,
7476 					PAGE_SIZE << HUGETLB_PAGE_ORDER,
7477 					HUGETLB_PAGE_ORDER, false, name,
7478 					&hugetlb_cma[nid], nid);
7479 		if (res) {
7480 			pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
7481 				res, nid);
7482 			continue;
7483 		}
7484 
7485 		reserved += size;
7486 		pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
7487 			size / SZ_1M, nid);
7488 
7489 		if (reserved >= hugetlb_cma_size)
7490 			break;
7491 	}
7492 
7493 	if (!reserved)
7494 		/*
7495 		 * hugetlb_cma_size is used to determine if allocations from
7496 		 * cma are possible.  Set to zero if no cma regions are set up.
7497 		 */
7498 		hugetlb_cma_size = 0;
7499 }
7500 
7501 static void __init hugetlb_cma_check(void)
7502 {
7503 	if (!hugetlb_cma_size || cma_reserve_called)
7504 		return;
7505 
7506 	pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
7507 }
7508 
7509 #endif /* CONFIG_CMA */
7510