xref: /openbmc/linux/mm/hugetlb.c (revision 04def1b9)
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 
35 #include <asm/page.h>
36 #include <asm/pgalloc.h>
37 #include <asm/tlb.h>
38 
39 #include <linux/io.h>
40 #include <linux/hugetlb.h>
41 #include <linux/hugetlb_cgroup.h>
42 #include <linux/node.h>
43 #include <linux/page_owner.h>
44 #include "internal.h"
45 #include "hugetlb_vmemmap.h"
46 
47 int hugetlb_max_hstate __read_mostly;
48 unsigned int default_hstate_idx;
49 struct hstate hstates[HUGE_MAX_HSTATE];
50 
51 #ifdef CONFIG_CMA
52 static struct cma *hugetlb_cma[MAX_NUMNODES];
53 static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata;
54 static bool hugetlb_cma_page(struct page *page, unsigned int order)
55 {
56 	return cma_pages_valid(hugetlb_cma[page_to_nid(page)], page,
57 				1 << order);
58 }
59 #else
60 static bool hugetlb_cma_page(struct page *page, unsigned int order)
61 {
62 	return false;
63 }
64 #endif
65 static unsigned long hugetlb_cma_size __initdata;
66 
67 /*
68  * Minimum page order among possible hugepage sizes, set to a proper value
69  * at boot time.
70  */
71 static unsigned int minimum_order __read_mostly = UINT_MAX;
72 
73 __initdata LIST_HEAD(huge_boot_pages);
74 
75 /* for command line parsing */
76 static struct hstate * __initdata parsed_hstate;
77 static unsigned long __initdata default_hstate_max_huge_pages;
78 static bool __initdata parsed_valid_hugepagesz = true;
79 static bool __initdata parsed_default_hugepagesz;
80 static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata;
81 
82 /*
83  * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
84  * free_huge_pages, and surplus_huge_pages.
85  */
86 DEFINE_SPINLOCK(hugetlb_lock);
87 
88 /*
89  * Serializes faults on the same logical page.  This is used to
90  * prevent spurious OOMs when the hugepage pool is fully utilized.
91  */
92 static int num_fault_mutexes;
93 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
94 
95 /* Forward declaration */
96 static int hugetlb_acct_memory(struct hstate *h, long delta);
97 
98 static inline bool subpool_is_free(struct hugepage_subpool *spool)
99 {
100 	if (spool->count)
101 		return false;
102 	if (spool->max_hpages != -1)
103 		return spool->used_hpages == 0;
104 	if (spool->min_hpages != -1)
105 		return spool->rsv_hpages == spool->min_hpages;
106 
107 	return true;
108 }
109 
110 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
111 						unsigned long irq_flags)
112 {
113 	spin_unlock_irqrestore(&spool->lock, irq_flags);
114 
115 	/* If no pages are used, and no other handles to the subpool
116 	 * remain, give up any reservations based on minimum size and
117 	 * free the subpool */
118 	if (subpool_is_free(spool)) {
119 		if (spool->min_hpages != -1)
120 			hugetlb_acct_memory(spool->hstate,
121 						-spool->min_hpages);
122 		kfree(spool);
123 	}
124 }
125 
126 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
127 						long min_hpages)
128 {
129 	struct hugepage_subpool *spool;
130 
131 	spool = kzalloc(sizeof(*spool), GFP_KERNEL);
132 	if (!spool)
133 		return NULL;
134 
135 	spin_lock_init(&spool->lock);
136 	spool->count = 1;
137 	spool->max_hpages = max_hpages;
138 	spool->hstate = h;
139 	spool->min_hpages = min_hpages;
140 
141 	if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
142 		kfree(spool);
143 		return NULL;
144 	}
145 	spool->rsv_hpages = min_hpages;
146 
147 	return spool;
148 }
149 
150 void hugepage_put_subpool(struct hugepage_subpool *spool)
151 {
152 	unsigned long flags;
153 
154 	spin_lock_irqsave(&spool->lock, flags);
155 	BUG_ON(!spool->count);
156 	spool->count--;
157 	unlock_or_release_subpool(spool, flags);
158 }
159 
160 /*
161  * Subpool accounting for allocating and reserving pages.
162  * Return -ENOMEM if there are not enough resources to satisfy the
163  * request.  Otherwise, return the number of pages by which the
164  * global pools must be adjusted (upward).  The returned value may
165  * only be different than the passed value (delta) in the case where
166  * a subpool minimum size must be maintained.
167  */
168 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
169 				      long delta)
170 {
171 	long ret = delta;
172 
173 	if (!spool)
174 		return ret;
175 
176 	spin_lock_irq(&spool->lock);
177 
178 	if (spool->max_hpages != -1) {		/* maximum size accounting */
179 		if ((spool->used_hpages + delta) <= spool->max_hpages)
180 			spool->used_hpages += delta;
181 		else {
182 			ret = -ENOMEM;
183 			goto unlock_ret;
184 		}
185 	}
186 
187 	/* minimum size accounting */
188 	if (spool->min_hpages != -1 && spool->rsv_hpages) {
189 		if (delta > spool->rsv_hpages) {
190 			/*
191 			 * Asking for more reserves than those already taken on
192 			 * behalf of subpool.  Return difference.
193 			 */
194 			ret = delta - spool->rsv_hpages;
195 			spool->rsv_hpages = 0;
196 		} else {
197 			ret = 0;	/* reserves already accounted for */
198 			spool->rsv_hpages -= delta;
199 		}
200 	}
201 
202 unlock_ret:
203 	spin_unlock_irq(&spool->lock);
204 	return ret;
205 }
206 
207 /*
208  * Subpool accounting for freeing and unreserving pages.
209  * Return the number of global page reservations that must be dropped.
210  * The return value may only be different than the passed value (delta)
211  * in the case where a subpool minimum size must be maintained.
212  */
213 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
214 				       long delta)
215 {
216 	long ret = delta;
217 	unsigned long flags;
218 
219 	if (!spool)
220 		return delta;
221 
222 	spin_lock_irqsave(&spool->lock, flags);
223 
224 	if (spool->max_hpages != -1)		/* maximum size accounting */
225 		spool->used_hpages -= delta;
226 
227 	 /* minimum size accounting */
228 	if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
229 		if (spool->rsv_hpages + delta <= spool->min_hpages)
230 			ret = 0;
231 		else
232 			ret = spool->rsv_hpages + delta - spool->min_hpages;
233 
234 		spool->rsv_hpages += delta;
235 		if (spool->rsv_hpages > spool->min_hpages)
236 			spool->rsv_hpages = spool->min_hpages;
237 	}
238 
239 	/*
240 	 * If hugetlbfs_put_super couldn't free spool due to an outstanding
241 	 * quota reference, free it now.
242 	 */
243 	unlock_or_release_subpool(spool, flags);
244 
245 	return ret;
246 }
247 
248 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
249 {
250 	return HUGETLBFS_SB(inode->i_sb)->spool;
251 }
252 
253 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
254 {
255 	return subpool_inode(file_inode(vma->vm_file));
256 }
257 
258 /* Helper that removes a struct file_region from the resv_map cache and returns
259  * it for use.
260  */
261 static struct file_region *
262 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
263 {
264 	struct file_region *nrg = NULL;
265 
266 	VM_BUG_ON(resv->region_cache_count <= 0);
267 
268 	resv->region_cache_count--;
269 	nrg = list_first_entry(&resv->region_cache, struct file_region, link);
270 	list_del(&nrg->link);
271 
272 	nrg->from = from;
273 	nrg->to = to;
274 
275 	return nrg;
276 }
277 
278 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
279 					      struct file_region *rg)
280 {
281 #ifdef CONFIG_CGROUP_HUGETLB
282 	nrg->reservation_counter = rg->reservation_counter;
283 	nrg->css = rg->css;
284 	if (rg->css)
285 		css_get(rg->css);
286 #endif
287 }
288 
289 /* Helper that records hugetlb_cgroup uncharge info. */
290 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
291 						struct hstate *h,
292 						struct resv_map *resv,
293 						struct file_region *nrg)
294 {
295 #ifdef CONFIG_CGROUP_HUGETLB
296 	if (h_cg) {
297 		nrg->reservation_counter =
298 			&h_cg->rsvd_hugepage[hstate_index(h)];
299 		nrg->css = &h_cg->css;
300 		/*
301 		 * The caller will hold exactly one h_cg->css reference for the
302 		 * whole contiguous reservation region. But this area might be
303 		 * scattered when there are already some file_regions reside in
304 		 * it. As a result, many file_regions may share only one css
305 		 * reference. In order to ensure that one file_region must hold
306 		 * exactly one h_cg->css reference, we should do css_get for
307 		 * each file_region and leave the reference held by caller
308 		 * untouched.
309 		 */
310 		css_get(&h_cg->css);
311 		if (!resv->pages_per_hpage)
312 			resv->pages_per_hpage = pages_per_huge_page(h);
313 		/* pages_per_hpage should be the same for all entries in
314 		 * a resv_map.
315 		 */
316 		VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
317 	} else {
318 		nrg->reservation_counter = NULL;
319 		nrg->css = NULL;
320 	}
321 #endif
322 }
323 
324 static void put_uncharge_info(struct file_region *rg)
325 {
326 #ifdef CONFIG_CGROUP_HUGETLB
327 	if (rg->css)
328 		css_put(rg->css);
329 #endif
330 }
331 
332 static bool has_same_uncharge_info(struct file_region *rg,
333 				   struct file_region *org)
334 {
335 #ifdef CONFIG_CGROUP_HUGETLB
336 	return rg->reservation_counter == org->reservation_counter &&
337 	       rg->css == org->css;
338 
339 #else
340 	return true;
341 #endif
342 }
343 
344 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
345 {
346 	struct file_region *nrg = NULL, *prg = NULL;
347 
348 	prg = list_prev_entry(rg, link);
349 	if (&prg->link != &resv->regions && prg->to == rg->from &&
350 	    has_same_uncharge_info(prg, rg)) {
351 		prg->to = rg->to;
352 
353 		list_del(&rg->link);
354 		put_uncharge_info(rg);
355 		kfree(rg);
356 
357 		rg = prg;
358 	}
359 
360 	nrg = list_next_entry(rg, link);
361 	if (&nrg->link != &resv->regions && nrg->from == rg->to &&
362 	    has_same_uncharge_info(nrg, rg)) {
363 		nrg->from = rg->from;
364 
365 		list_del(&rg->link);
366 		put_uncharge_info(rg);
367 		kfree(rg);
368 	}
369 }
370 
371 static inline long
372 hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from,
373 		     long to, struct hstate *h, struct hugetlb_cgroup *cg,
374 		     long *regions_needed)
375 {
376 	struct file_region *nrg;
377 
378 	if (!regions_needed) {
379 		nrg = get_file_region_entry_from_cache(map, from, to);
380 		record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
381 		list_add(&nrg->link, rg->link.prev);
382 		coalesce_file_region(map, nrg);
383 	} else
384 		*regions_needed += 1;
385 
386 	return to - from;
387 }
388 
389 /*
390  * Must be called with resv->lock held.
391  *
392  * Calling this with regions_needed != NULL will count the number of pages
393  * to be added but will not modify the linked list. And regions_needed will
394  * indicate the number of file_regions needed in the cache to carry out to add
395  * the regions for this range.
396  */
397 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
398 				     struct hugetlb_cgroup *h_cg,
399 				     struct hstate *h, long *regions_needed)
400 {
401 	long add = 0;
402 	struct list_head *head = &resv->regions;
403 	long last_accounted_offset = f;
404 	struct file_region *rg = NULL, *trg = NULL;
405 
406 	if (regions_needed)
407 		*regions_needed = 0;
408 
409 	/* In this loop, we essentially handle an entry for the range
410 	 * [last_accounted_offset, rg->from), at every iteration, with some
411 	 * bounds checking.
412 	 */
413 	list_for_each_entry_safe(rg, trg, head, link) {
414 		/* Skip irrelevant regions that start before our range. */
415 		if (rg->from < f) {
416 			/* If this region ends after the last accounted offset,
417 			 * then we need to update last_accounted_offset.
418 			 */
419 			if (rg->to > last_accounted_offset)
420 				last_accounted_offset = rg->to;
421 			continue;
422 		}
423 
424 		/* When we find a region that starts beyond our range, we've
425 		 * finished.
426 		 */
427 		if (rg->from >= t)
428 			break;
429 
430 		/* Add an entry for last_accounted_offset -> rg->from, and
431 		 * update last_accounted_offset.
432 		 */
433 		if (rg->from > last_accounted_offset)
434 			add += hugetlb_resv_map_add(resv, rg,
435 						    last_accounted_offset,
436 						    rg->from, h, h_cg,
437 						    regions_needed);
438 
439 		last_accounted_offset = rg->to;
440 	}
441 
442 	/* Handle the case where our range extends beyond
443 	 * last_accounted_offset.
444 	 */
445 	if (last_accounted_offset < t)
446 		add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
447 					    t, h, h_cg, regions_needed);
448 
449 	return add;
450 }
451 
452 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
453  */
454 static int allocate_file_region_entries(struct resv_map *resv,
455 					int regions_needed)
456 	__must_hold(&resv->lock)
457 {
458 	struct list_head allocated_regions;
459 	int to_allocate = 0, i = 0;
460 	struct file_region *trg = NULL, *rg = NULL;
461 
462 	VM_BUG_ON(regions_needed < 0);
463 
464 	INIT_LIST_HEAD(&allocated_regions);
465 
466 	/*
467 	 * Check for sufficient descriptors in the cache to accommodate
468 	 * the number of in progress add operations plus regions_needed.
469 	 *
470 	 * This is a while loop because when we drop the lock, some other call
471 	 * to region_add or region_del may have consumed some region_entries,
472 	 * so we keep looping here until we finally have enough entries for
473 	 * (adds_in_progress + regions_needed).
474 	 */
475 	while (resv->region_cache_count <
476 	       (resv->adds_in_progress + regions_needed)) {
477 		to_allocate = resv->adds_in_progress + regions_needed -
478 			      resv->region_cache_count;
479 
480 		/* At this point, we should have enough entries in the cache
481 		 * for all the existing adds_in_progress. We should only be
482 		 * needing to allocate for regions_needed.
483 		 */
484 		VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
485 
486 		spin_unlock(&resv->lock);
487 		for (i = 0; i < to_allocate; i++) {
488 			trg = kmalloc(sizeof(*trg), GFP_KERNEL);
489 			if (!trg)
490 				goto out_of_memory;
491 			list_add(&trg->link, &allocated_regions);
492 		}
493 
494 		spin_lock(&resv->lock);
495 
496 		list_splice(&allocated_regions, &resv->region_cache);
497 		resv->region_cache_count += to_allocate;
498 	}
499 
500 	return 0;
501 
502 out_of_memory:
503 	list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
504 		list_del(&rg->link);
505 		kfree(rg);
506 	}
507 	return -ENOMEM;
508 }
509 
510 /*
511  * Add the huge page range represented by [f, t) to the reserve
512  * map.  Regions will be taken from the cache to fill in this range.
513  * Sufficient regions should exist in the cache due to the previous
514  * call to region_chg with the same range, but in some cases the cache will not
515  * have sufficient entries due to races with other code doing region_add or
516  * region_del.  The extra needed entries will be allocated.
517  *
518  * regions_needed is the out value provided by a previous call to region_chg.
519  *
520  * Return the number of new huge pages added to the map.  This number is greater
521  * than or equal to zero.  If file_region entries needed to be allocated for
522  * this operation and we were not able to allocate, it returns -ENOMEM.
523  * region_add of regions of length 1 never allocate file_regions and cannot
524  * fail; region_chg will always allocate at least 1 entry and a region_add for
525  * 1 page will only require at most 1 entry.
526  */
527 static long region_add(struct resv_map *resv, long f, long t,
528 		       long in_regions_needed, struct hstate *h,
529 		       struct hugetlb_cgroup *h_cg)
530 {
531 	long add = 0, actual_regions_needed = 0;
532 
533 	spin_lock(&resv->lock);
534 retry:
535 
536 	/* Count how many regions are actually needed to execute this add. */
537 	add_reservation_in_range(resv, f, t, NULL, NULL,
538 				 &actual_regions_needed);
539 
540 	/*
541 	 * Check for sufficient descriptors in the cache to accommodate
542 	 * this add operation. Note that actual_regions_needed may be greater
543 	 * than in_regions_needed, as the resv_map may have been modified since
544 	 * the region_chg call. In this case, we need to make sure that we
545 	 * allocate extra entries, such that we have enough for all the
546 	 * existing adds_in_progress, plus the excess needed for this
547 	 * operation.
548 	 */
549 	if (actual_regions_needed > in_regions_needed &&
550 	    resv->region_cache_count <
551 		    resv->adds_in_progress +
552 			    (actual_regions_needed - in_regions_needed)) {
553 		/* region_add operation of range 1 should never need to
554 		 * allocate file_region entries.
555 		 */
556 		VM_BUG_ON(t - f <= 1);
557 
558 		if (allocate_file_region_entries(
559 			    resv, actual_regions_needed - in_regions_needed)) {
560 			return -ENOMEM;
561 		}
562 
563 		goto retry;
564 	}
565 
566 	add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
567 
568 	resv->adds_in_progress -= in_regions_needed;
569 
570 	spin_unlock(&resv->lock);
571 	return add;
572 }
573 
574 /*
575  * Examine the existing reserve map and determine how many
576  * huge pages in the specified range [f, t) are NOT currently
577  * represented.  This routine is called before a subsequent
578  * call to region_add that will actually modify the reserve
579  * map to add the specified range [f, t).  region_chg does
580  * not change the number of huge pages represented by the
581  * map.  A number of new file_region structures is added to the cache as a
582  * placeholder, for the subsequent region_add call to use. At least 1
583  * file_region structure is added.
584  *
585  * out_regions_needed is the number of regions added to the
586  * resv->adds_in_progress.  This value needs to be provided to a follow up call
587  * to region_add or region_abort for proper accounting.
588  *
589  * Returns the number of huge pages that need to be added to the existing
590  * reservation map for the range [f, t).  This number is greater or equal to
591  * zero.  -ENOMEM is returned if a new file_region structure or cache entry
592  * is needed and can not be allocated.
593  */
594 static long region_chg(struct resv_map *resv, long f, long t,
595 		       long *out_regions_needed)
596 {
597 	long chg = 0;
598 
599 	spin_lock(&resv->lock);
600 
601 	/* Count how many hugepages in this range are NOT represented. */
602 	chg = add_reservation_in_range(resv, f, t, NULL, NULL,
603 				       out_regions_needed);
604 
605 	if (*out_regions_needed == 0)
606 		*out_regions_needed = 1;
607 
608 	if (allocate_file_region_entries(resv, *out_regions_needed))
609 		return -ENOMEM;
610 
611 	resv->adds_in_progress += *out_regions_needed;
612 
613 	spin_unlock(&resv->lock);
614 	return chg;
615 }
616 
617 /*
618  * Abort the in progress add operation.  The adds_in_progress field
619  * of the resv_map keeps track of the operations in progress between
620  * calls to region_chg and region_add.  Operations are sometimes
621  * aborted after the call to region_chg.  In such cases, region_abort
622  * is called to decrement the adds_in_progress counter. regions_needed
623  * is the value returned by the region_chg call, it is used to decrement
624  * the adds_in_progress counter.
625  *
626  * NOTE: The range arguments [f, t) are not needed or used in this
627  * routine.  They are kept to make reading the calling code easier as
628  * arguments will match the associated region_chg call.
629  */
630 static void region_abort(struct resv_map *resv, long f, long t,
631 			 long regions_needed)
632 {
633 	spin_lock(&resv->lock);
634 	VM_BUG_ON(!resv->region_cache_count);
635 	resv->adds_in_progress -= regions_needed;
636 	spin_unlock(&resv->lock);
637 }
638 
639 /*
640  * Delete the specified range [f, t) from the reserve map.  If the
641  * t parameter is LONG_MAX, this indicates that ALL regions after f
642  * should be deleted.  Locate the regions which intersect [f, t)
643  * and either trim, delete or split the existing regions.
644  *
645  * Returns the number of huge pages deleted from the reserve map.
646  * In the normal case, the return value is zero or more.  In the
647  * case where a region must be split, a new region descriptor must
648  * be allocated.  If the allocation fails, -ENOMEM will be returned.
649  * NOTE: If the parameter t == LONG_MAX, then we will never split
650  * a region and possibly return -ENOMEM.  Callers specifying
651  * t == LONG_MAX do not need to check for -ENOMEM error.
652  */
653 static long region_del(struct resv_map *resv, long f, long t)
654 {
655 	struct list_head *head = &resv->regions;
656 	struct file_region *rg, *trg;
657 	struct file_region *nrg = NULL;
658 	long del = 0;
659 
660 retry:
661 	spin_lock(&resv->lock);
662 	list_for_each_entry_safe(rg, trg, head, link) {
663 		/*
664 		 * Skip regions before the range to be deleted.  file_region
665 		 * ranges are normally of the form [from, to).  However, there
666 		 * may be a "placeholder" entry in the map which is of the form
667 		 * (from, to) with from == to.  Check for placeholder entries
668 		 * at the beginning of the range to be deleted.
669 		 */
670 		if (rg->to <= f && (rg->to != rg->from || rg->to != f))
671 			continue;
672 
673 		if (rg->from >= t)
674 			break;
675 
676 		if (f > rg->from && t < rg->to) { /* Must split region */
677 			/*
678 			 * Check for an entry in the cache before dropping
679 			 * lock and attempting allocation.
680 			 */
681 			if (!nrg &&
682 			    resv->region_cache_count > resv->adds_in_progress) {
683 				nrg = list_first_entry(&resv->region_cache,
684 							struct file_region,
685 							link);
686 				list_del(&nrg->link);
687 				resv->region_cache_count--;
688 			}
689 
690 			if (!nrg) {
691 				spin_unlock(&resv->lock);
692 				nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
693 				if (!nrg)
694 					return -ENOMEM;
695 				goto retry;
696 			}
697 
698 			del += t - f;
699 			hugetlb_cgroup_uncharge_file_region(
700 				resv, rg, t - f, false);
701 
702 			/* New entry for end of split region */
703 			nrg->from = t;
704 			nrg->to = rg->to;
705 
706 			copy_hugetlb_cgroup_uncharge_info(nrg, rg);
707 
708 			INIT_LIST_HEAD(&nrg->link);
709 
710 			/* Original entry is trimmed */
711 			rg->to = f;
712 
713 			list_add(&nrg->link, &rg->link);
714 			nrg = NULL;
715 			break;
716 		}
717 
718 		if (f <= rg->from && t >= rg->to) { /* Remove entire region */
719 			del += rg->to - rg->from;
720 			hugetlb_cgroup_uncharge_file_region(resv, rg,
721 							    rg->to - rg->from, true);
722 			list_del(&rg->link);
723 			kfree(rg);
724 			continue;
725 		}
726 
727 		if (f <= rg->from) {	/* Trim beginning of region */
728 			hugetlb_cgroup_uncharge_file_region(resv, rg,
729 							    t - rg->from, false);
730 
731 			del += t - rg->from;
732 			rg->from = t;
733 		} else {		/* Trim end of region */
734 			hugetlb_cgroup_uncharge_file_region(resv, rg,
735 							    rg->to - f, false);
736 
737 			del += rg->to - f;
738 			rg->to = f;
739 		}
740 	}
741 
742 	spin_unlock(&resv->lock);
743 	kfree(nrg);
744 	return del;
745 }
746 
747 /*
748  * A rare out of memory error was encountered which prevented removal of
749  * the reserve map region for a page.  The huge page itself was free'ed
750  * and removed from the page cache.  This routine will adjust the subpool
751  * usage count, and the global reserve count if needed.  By incrementing
752  * these counts, the reserve map entry which could not be deleted will
753  * appear as a "reserved" entry instead of simply dangling with incorrect
754  * counts.
755  */
756 void hugetlb_fix_reserve_counts(struct inode *inode)
757 {
758 	struct hugepage_subpool *spool = subpool_inode(inode);
759 	long rsv_adjust;
760 	bool reserved = false;
761 
762 	rsv_adjust = hugepage_subpool_get_pages(spool, 1);
763 	if (rsv_adjust > 0) {
764 		struct hstate *h = hstate_inode(inode);
765 
766 		if (!hugetlb_acct_memory(h, 1))
767 			reserved = true;
768 	} else if (!rsv_adjust) {
769 		reserved = true;
770 	}
771 
772 	if (!reserved)
773 		pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
774 }
775 
776 /*
777  * Count and return the number of huge pages in the reserve map
778  * that intersect with the range [f, t).
779  */
780 static long region_count(struct resv_map *resv, long f, long t)
781 {
782 	struct list_head *head = &resv->regions;
783 	struct file_region *rg;
784 	long chg = 0;
785 
786 	spin_lock(&resv->lock);
787 	/* Locate each segment we overlap with, and count that overlap. */
788 	list_for_each_entry(rg, head, link) {
789 		long seg_from;
790 		long seg_to;
791 
792 		if (rg->to <= f)
793 			continue;
794 		if (rg->from >= t)
795 			break;
796 
797 		seg_from = max(rg->from, f);
798 		seg_to = min(rg->to, t);
799 
800 		chg += seg_to - seg_from;
801 	}
802 	spin_unlock(&resv->lock);
803 
804 	return chg;
805 }
806 
807 /*
808  * Convert the address within this vma to the page offset within
809  * the mapping, in pagecache page units; huge pages here.
810  */
811 static pgoff_t vma_hugecache_offset(struct hstate *h,
812 			struct vm_area_struct *vma, unsigned long address)
813 {
814 	return ((address - vma->vm_start) >> huge_page_shift(h)) +
815 			(vma->vm_pgoff >> huge_page_order(h));
816 }
817 
818 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
819 				     unsigned long address)
820 {
821 	return vma_hugecache_offset(hstate_vma(vma), vma, address);
822 }
823 EXPORT_SYMBOL_GPL(linear_hugepage_index);
824 
825 /*
826  * Return the size of the pages allocated when backing a VMA. In the majority
827  * cases this will be same size as used by the page table entries.
828  */
829 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
830 {
831 	if (vma->vm_ops && vma->vm_ops->pagesize)
832 		return vma->vm_ops->pagesize(vma);
833 	return PAGE_SIZE;
834 }
835 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
836 
837 /*
838  * Return the page size being used by the MMU to back a VMA. In the majority
839  * of cases, the page size used by the kernel matches the MMU size. On
840  * architectures where it differs, an architecture-specific 'strong'
841  * version of this symbol is required.
842  */
843 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
844 {
845 	return vma_kernel_pagesize(vma);
846 }
847 
848 /*
849  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
850  * bits of the reservation map pointer, which are always clear due to
851  * alignment.
852  */
853 #define HPAGE_RESV_OWNER    (1UL << 0)
854 #define HPAGE_RESV_UNMAPPED (1UL << 1)
855 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
856 
857 /*
858  * These helpers are used to track how many pages are reserved for
859  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
860  * is guaranteed to have their future faults succeed.
861  *
862  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
863  * the reserve counters are updated with the hugetlb_lock held. It is safe
864  * to reset the VMA at fork() time as it is not in use yet and there is no
865  * chance of the global counters getting corrupted as a result of the values.
866  *
867  * The private mapping reservation is represented in a subtly different
868  * manner to a shared mapping.  A shared mapping has a region map associated
869  * with the underlying file, this region map represents the backing file
870  * pages which have ever had a reservation assigned which this persists even
871  * after the page is instantiated.  A private mapping has a region map
872  * associated with the original mmap which is attached to all VMAs which
873  * reference it, this region map represents those offsets which have consumed
874  * reservation ie. where pages have been instantiated.
875  */
876 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
877 {
878 	return (unsigned long)vma->vm_private_data;
879 }
880 
881 static void set_vma_private_data(struct vm_area_struct *vma,
882 							unsigned long value)
883 {
884 	vma->vm_private_data = (void *)value;
885 }
886 
887 static void
888 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
889 					  struct hugetlb_cgroup *h_cg,
890 					  struct hstate *h)
891 {
892 #ifdef CONFIG_CGROUP_HUGETLB
893 	if (!h_cg || !h) {
894 		resv_map->reservation_counter = NULL;
895 		resv_map->pages_per_hpage = 0;
896 		resv_map->css = NULL;
897 	} else {
898 		resv_map->reservation_counter =
899 			&h_cg->rsvd_hugepage[hstate_index(h)];
900 		resv_map->pages_per_hpage = pages_per_huge_page(h);
901 		resv_map->css = &h_cg->css;
902 	}
903 #endif
904 }
905 
906 struct resv_map *resv_map_alloc(void)
907 {
908 	struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
909 	struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
910 
911 	if (!resv_map || !rg) {
912 		kfree(resv_map);
913 		kfree(rg);
914 		return NULL;
915 	}
916 
917 	kref_init(&resv_map->refs);
918 	spin_lock_init(&resv_map->lock);
919 	INIT_LIST_HEAD(&resv_map->regions);
920 
921 	resv_map->adds_in_progress = 0;
922 	/*
923 	 * Initialize these to 0. On shared mappings, 0's here indicate these
924 	 * fields don't do cgroup accounting. On private mappings, these will be
925 	 * re-initialized to the proper values, to indicate that hugetlb cgroup
926 	 * reservations are to be un-charged from here.
927 	 */
928 	resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
929 
930 	INIT_LIST_HEAD(&resv_map->region_cache);
931 	list_add(&rg->link, &resv_map->region_cache);
932 	resv_map->region_cache_count = 1;
933 
934 	return resv_map;
935 }
936 
937 void resv_map_release(struct kref *ref)
938 {
939 	struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
940 	struct list_head *head = &resv_map->region_cache;
941 	struct file_region *rg, *trg;
942 
943 	/* Clear out any active regions before we release the map. */
944 	region_del(resv_map, 0, LONG_MAX);
945 
946 	/* ... and any entries left in the cache */
947 	list_for_each_entry_safe(rg, trg, head, link) {
948 		list_del(&rg->link);
949 		kfree(rg);
950 	}
951 
952 	VM_BUG_ON(resv_map->adds_in_progress);
953 
954 	kfree(resv_map);
955 }
956 
957 static inline struct resv_map *inode_resv_map(struct inode *inode)
958 {
959 	/*
960 	 * At inode evict time, i_mapping may not point to the original
961 	 * address space within the inode.  This original address space
962 	 * contains the pointer to the resv_map.  So, always use the
963 	 * address space embedded within the inode.
964 	 * The VERY common case is inode->mapping == &inode->i_data but,
965 	 * this may not be true for device special inodes.
966 	 */
967 	return (struct resv_map *)(&inode->i_data)->private_data;
968 }
969 
970 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
971 {
972 	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
973 	if (vma->vm_flags & VM_MAYSHARE) {
974 		struct address_space *mapping = vma->vm_file->f_mapping;
975 		struct inode *inode = mapping->host;
976 
977 		return inode_resv_map(inode);
978 
979 	} else {
980 		return (struct resv_map *)(get_vma_private_data(vma) &
981 							~HPAGE_RESV_MASK);
982 	}
983 }
984 
985 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
986 {
987 	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
988 	VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
989 
990 	set_vma_private_data(vma, (get_vma_private_data(vma) &
991 				HPAGE_RESV_MASK) | (unsigned long)map);
992 }
993 
994 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
995 {
996 	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
997 	VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
998 
999 	set_vma_private_data(vma, get_vma_private_data(vma) | flags);
1000 }
1001 
1002 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
1003 {
1004 	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1005 
1006 	return (get_vma_private_data(vma) & flag) != 0;
1007 }
1008 
1009 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
1010 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
1011 {
1012 	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1013 	if (!(vma->vm_flags & VM_MAYSHARE))
1014 		vma->vm_private_data = (void *)0;
1015 }
1016 
1017 /*
1018  * Reset and decrement one ref on hugepage private reservation.
1019  * Called with mm->mmap_sem writer semaphore held.
1020  * This function should be only used by move_vma() and operate on
1021  * same sized vma. It should never come here with last ref on the
1022  * reservation.
1023  */
1024 void clear_vma_resv_huge_pages(struct vm_area_struct *vma)
1025 {
1026 	/*
1027 	 * Clear the old hugetlb private page reservation.
1028 	 * It has already been transferred to new_vma.
1029 	 *
1030 	 * During a mremap() operation of a hugetlb vma we call move_vma()
1031 	 * which copies vma into new_vma and unmaps vma. After the copy
1032 	 * operation both new_vma and vma share a reference to the resv_map
1033 	 * struct, and at that point vma is about to be unmapped. We don't
1034 	 * want to return the reservation to the pool at unmap of vma because
1035 	 * the reservation still lives on in new_vma, so simply decrement the
1036 	 * ref here and remove the resv_map reference from this vma.
1037 	 */
1038 	struct resv_map *reservations = vma_resv_map(vma);
1039 
1040 	if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
1041 		kref_put(&reservations->refs, resv_map_release);
1042 
1043 	reset_vma_resv_huge_pages(vma);
1044 }
1045 
1046 /* Returns true if the VMA has associated reserve pages */
1047 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1048 {
1049 	if (vma->vm_flags & VM_NORESERVE) {
1050 		/*
1051 		 * This address is already reserved by other process(chg == 0),
1052 		 * so, we should decrement reserved count. Without decrementing,
1053 		 * reserve count remains after releasing inode, because this
1054 		 * allocated page will go into page cache and is regarded as
1055 		 * coming from reserved pool in releasing step.  Currently, we
1056 		 * don't have any other solution to deal with this situation
1057 		 * properly, so add work-around here.
1058 		 */
1059 		if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1060 			return true;
1061 		else
1062 			return false;
1063 	}
1064 
1065 	/* Shared mappings always use reserves */
1066 	if (vma->vm_flags & VM_MAYSHARE) {
1067 		/*
1068 		 * We know VM_NORESERVE is not set.  Therefore, there SHOULD
1069 		 * be a region map for all pages.  The only situation where
1070 		 * there is no region map is if a hole was punched via
1071 		 * fallocate.  In this case, there really are no reserves to
1072 		 * use.  This situation is indicated if chg != 0.
1073 		 */
1074 		if (chg)
1075 			return false;
1076 		else
1077 			return true;
1078 	}
1079 
1080 	/*
1081 	 * Only the process that called mmap() has reserves for
1082 	 * private mappings.
1083 	 */
1084 	if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1085 		/*
1086 		 * Like the shared case above, a hole punch or truncate
1087 		 * could have been performed on the private mapping.
1088 		 * Examine the value of chg to determine if reserves
1089 		 * actually exist or were previously consumed.
1090 		 * Very Subtle - The value of chg comes from a previous
1091 		 * call to vma_needs_reserves().  The reserve map for
1092 		 * private mappings has different (opposite) semantics
1093 		 * than that of shared mappings.  vma_needs_reserves()
1094 		 * has already taken this difference in semantics into
1095 		 * account.  Therefore, the meaning of chg is the same
1096 		 * as in the shared case above.  Code could easily be
1097 		 * combined, but keeping it separate draws attention to
1098 		 * subtle differences.
1099 		 */
1100 		if (chg)
1101 			return false;
1102 		else
1103 			return true;
1104 	}
1105 
1106 	return false;
1107 }
1108 
1109 static void enqueue_huge_page(struct hstate *h, struct page *page)
1110 {
1111 	int nid = page_to_nid(page);
1112 
1113 	lockdep_assert_held(&hugetlb_lock);
1114 	VM_BUG_ON_PAGE(page_count(page), page);
1115 
1116 	list_move(&page->lru, &h->hugepage_freelists[nid]);
1117 	h->free_huge_pages++;
1118 	h->free_huge_pages_node[nid]++;
1119 	SetHPageFreed(page);
1120 }
1121 
1122 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1123 {
1124 	struct page *page;
1125 	bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1126 
1127 	lockdep_assert_held(&hugetlb_lock);
1128 	list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1129 		if (pin && !is_pinnable_page(page))
1130 			continue;
1131 
1132 		if (PageHWPoison(page))
1133 			continue;
1134 
1135 		list_move(&page->lru, &h->hugepage_activelist);
1136 		set_page_refcounted(page);
1137 		ClearHPageFreed(page);
1138 		h->free_huge_pages--;
1139 		h->free_huge_pages_node[nid]--;
1140 		return page;
1141 	}
1142 
1143 	return NULL;
1144 }
1145 
1146 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1147 		nodemask_t *nmask)
1148 {
1149 	unsigned int cpuset_mems_cookie;
1150 	struct zonelist *zonelist;
1151 	struct zone *zone;
1152 	struct zoneref *z;
1153 	int node = NUMA_NO_NODE;
1154 
1155 	zonelist = node_zonelist(nid, gfp_mask);
1156 
1157 retry_cpuset:
1158 	cpuset_mems_cookie = read_mems_allowed_begin();
1159 	for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1160 		struct page *page;
1161 
1162 		if (!cpuset_zone_allowed(zone, gfp_mask))
1163 			continue;
1164 		/*
1165 		 * no need to ask again on the same node. Pool is node rather than
1166 		 * zone aware
1167 		 */
1168 		if (zone_to_nid(zone) == node)
1169 			continue;
1170 		node = zone_to_nid(zone);
1171 
1172 		page = dequeue_huge_page_node_exact(h, node);
1173 		if (page)
1174 			return page;
1175 	}
1176 	if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1177 		goto retry_cpuset;
1178 
1179 	return NULL;
1180 }
1181 
1182 static struct page *dequeue_huge_page_vma(struct hstate *h,
1183 				struct vm_area_struct *vma,
1184 				unsigned long address, int avoid_reserve,
1185 				long chg)
1186 {
1187 	struct page *page = NULL;
1188 	struct mempolicy *mpol;
1189 	gfp_t gfp_mask;
1190 	nodemask_t *nodemask;
1191 	int nid;
1192 
1193 	/*
1194 	 * A child process with MAP_PRIVATE mappings created by their parent
1195 	 * have no page reserves. This check ensures that reservations are
1196 	 * not "stolen". The child may still get SIGKILLed
1197 	 */
1198 	if (!vma_has_reserves(vma, chg) &&
1199 			h->free_huge_pages - h->resv_huge_pages == 0)
1200 		goto err;
1201 
1202 	/* If reserves cannot be used, ensure enough pages are in the pool */
1203 	if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1204 		goto err;
1205 
1206 	gfp_mask = htlb_alloc_mask(h);
1207 	nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1208 
1209 	if (mpol_is_preferred_many(mpol)) {
1210 		page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1211 
1212 		/* Fallback to all nodes if page==NULL */
1213 		nodemask = NULL;
1214 	}
1215 
1216 	if (!page)
1217 		page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1218 
1219 	if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1220 		SetHPageRestoreReserve(page);
1221 		h->resv_huge_pages--;
1222 	}
1223 
1224 	mpol_cond_put(mpol);
1225 	return page;
1226 
1227 err:
1228 	return NULL;
1229 }
1230 
1231 /*
1232  * common helper functions for hstate_next_node_to_{alloc|free}.
1233  * We may have allocated or freed a huge page based on a different
1234  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1235  * be outside of *nodes_allowed.  Ensure that we use an allowed
1236  * node for alloc or free.
1237  */
1238 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1239 {
1240 	nid = next_node_in(nid, *nodes_allowed);
1241 	VM_BUG_ON(nid >= MAX_NUMNODES);
1242 
1243 	return nid;
1244 }
1245 
1246 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1247 {
1248 	if (!node_isset(nid, *nodes_allowed))
1249 		nid = next_node_allowed(nid, nodes_allowed);
1250 	return nid;
1251 }
1252 
1253 /*
1254  * returns the previously saved node ["this node"] from which to
1255  * allocate a persistent huge page for the pool and advance the
1256  * next node from which to allocate, handling wrap at end of node
1257  * mask.
1258  */
1259 static int hstate_next_node_to_alloc(struct hstate *h,
1260 					nodemask_t *nodes_allowed)
1261 {
1262 	int nid;
1263 
1264 	VM_BUG_ON(!nodes_allowed);
1265 
1266 	nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1267 	h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1268 
1269 	return nid;
1270 }
1271 
1272 /*
1273  * helper for remove_pool_huge_page() - return the previously saved
1274  * node ["this node"] from which to free a huge page.  Advance the
1275  * next node id whether or not we find a free huge page to free so
1276  * that the next attempt to free addresses the next node.
1277  */
1278 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1279 {
1280 	int nid;
1281 
1282 	VM_BUG_ON(!nodes_allowed);
1283 
1284 	nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1285 	h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1286 
1287 	return nid;
1288 }
1289 
1290 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)		\
1291 	for (nr_nodes = nodes_weight(*mask);				\
1292 		nr_nodes > 0 &&						\
1293 		((node = hstate_next_node_to_alloc(hs, mask)) || 1);	\
1294 		nr_nodes--)
1295 
1296 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)		\
1297 	for (nr_nodes = nodes_weight(*mask);				\
1298 		nr_nodes > 0 &&						\
1299 		((node = hstate_next_node_to_free(hs, mask)) || 1);	\
1300 		nr_nodes--)
1301 
1302 /* used to demote non-gigantic_huge pages as well */
1303 static void __destroy_compound_gigantic_page(struct page *page,
1304 					unsigned int order, bool demote)
1305 {
1306 	int i;
1307 	int nr_pages = 1 << order;
1308 	struct page *p = page + 1;
1309 
1310 	atomic_set(compound_mapcount_ptr(page), 0);
1311 	atomic_set(compound_pincount_ptr(page), 0);
1312 
1313 	for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1314 		p->mapping = NULL;
1315 		clear_compound_head(p);
1316 		if (!demote)
1317 			set_page_refcounted(p);
1318 	}
1319 
1320 	set_compound_order(page, 0);
1321 	page[1].compound_nr = 0;
1322 	__ClearPageHead(page);
1323 }
1324 
1325 static void destroy_compound_hugetlb_page_for_demote(struct page *page,
1326 					unsigned int order)
1327 {
1328 	__destroy_compound_gigantic_page(page, order, true);
1329 }
1330 
1331 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1332 static void destroy_compound_gigantic_page(struct page *page,
1333 					unsigned int order)
1334 {
1335 	__destroy_compound_gigantic_page(page, order, false);
1336 }
1337 
1338 static void free_gigantic_page(struct page *page, unsigned int order)
1339 {
1340 	/*
1341 	 * If the page isn't allocated using the cma allocator,
1342 	 * cma_release() returns false.
1343 	 */
1344 #ifdef CONFIG_CMA
1345 	if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1346 		return;
1347 #endif
1348 
1349 	free_contig_range(page_to_pfn(page), 1 << order);
1350 }
1351 
1352 #ifdef CONFIG_CONTIG_ALLOC
1353 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1354 		int nid, nodemask_t *nodemask)
1355 {
1356 	unsigned long nr_pages = pages_per_huge_page(h);
1357 	if (nid == NUMA_NO_NODE)
1358 		nid = numa_mem_id();
1359 
1360 #ifdef CONFIG_CMA
1361 	{
1362 		struct page *page;
1363 		int node;
1364 
1365 		if (hugetlb_cma[nid]) {
1366 			page = cma_alloc(hugetlb_cma[nid], nr_pages,
1367 					huge_page_order(h), true);
1368 			if (page)
1369 				return page;
1370 		}
1371 
1372 		if (!(gfp_mask & __GFP_THISNODE)) {
1373 			for_each_node_mask(node, *nodemask) {
1374 				if (node == nid || !hugetlb_cma[node])
1375 					continue;
1376 
1377 				page = cma_alloc(hugetlb_cma[node], nr_pages,
1378 						huge_page_order(h), true);
1379 				if (page)
1380 					return page;
1381 			}
1382 		}
1383 	}
1384 #endif
1385 
1386 	return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1387 }
1388 
1389 #else /* !CONFIG_CONTIG_ALLOC */
1390 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1391 					int nid, nodemask_t *nodemask)
1392 {
1393 	return NULL;
1394 }
1395 #endif /* CONFIG_CONTIG_ALLOC */
1396 
1397 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1398 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1399 					int nid, nodemask_t *nodemask)
1400 {
1401 	return NULL;
1402 }
1403 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1404 static inline void destroy_compound_gigantic_page(struct page *page,
1405 						unsigned int order) { }
1406 #endif
1407 
1408 /*
1409  * Remove hugetlb page from lists, and update dtor so that page appears
1410  * as just a compound page.
1411  *
1412  * A reference is held on the page, except in the case of demote.
1413  *
1414  * Must be called with hugetlb lock held.
1415  */
1416 static void __remove_hugetlb_page(struct hstate *h, struct page *page,
1417 							bool adjust_surplus,
1418 							bool demote)
1419 {
1420 	int nid = page_to_nid(page);
1421 
1422 	VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1423 	VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1424 
1425 	lockdep_assert_held(&hugetlb_lock);
1426 	if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1427 		return;
1428 
1429 	list_del(&page->lru);
1430 
1431 	if (HPageFreed(page)) {
1432 		h->free_huge_pages--;
1433 		h->free_huge_pages_node[nid]--;
1434 	}
1435 	if (adjust_surplus) {
1436 		h->surplus_huge_pages--;
1437 		h->surplus_huge_pages_node[nid]--;
1438 	}
1439 
1440 	/*
1441 	 * Very subtle
1442 	 *
1443 	 * For non-gigantic pages set the destructor to the normal compound
1444 	 * page dtor.  This is needed in case someone takes an additional
1445 	 * temporary ref to the page, and freeing is delayed until they drop
1446 	 * their reference.
1447 	 *
1448 	 * For gigantic pages set the destructor to the null dtor.  This
1449 	 * destructor will never be called.  Before freeing the gigantic
1450 	 * page destroy_compound_gigantic_page will turn the compound page
1451 	 * into a simple group of pages.  After this the destructor does not
1452 	 * apply.
1453 	 *
1454 	 * This handles the case where more than one ref is held when and
1455 	 * after update_and_free_page is called.
1456 	 *
1457 	 * In the case of demote we do not ref count the page as it will soon
1458 	 * be turned into a page of smaller size.
1459 	 */
1460 	if (!demote)
1461 		set_page_refcounted(page);
1462 	if (hstate_is_gigantic(h))
1463 		set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1464 	else
1465 		set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
1466 
1467 	h->nr_huge_pages--;
1468 	h->nr_huge_pages_node[nid]--;
1469 }
1470 
1471 static void remove_hugetlb_page(struct hstate *h, struct page *page,
1472 							bool adjust_surplus)
1473 {
1474 	__remove_hugetlb_page(h, page, adjust_surplus, false);
1475 }
1476 
1477 static void remove_hugetlb_page_for_demote(struct hstate *h, struct page *page,
1478 							bool adjust_surplus)
1479 {
1480 	__remove_hugetlb_page(h, page, adjust_surplus, true);
1481 }
1482 
1483 static void add_hugetlb_page(struct hstate *h, struct page *page,
1484 			     bool adjust_surplus)
1485 {
1486 	int zeroed;
1487 	int nid = page_to_nid(page);
1488 
1489 	VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page);
1490 
1491 	lockdep_assert_held(&hugetlb_lock);
1492 
1493 	INIT_LIST_HEAD(&page->lru);
1494 	h->nr_huge_pages++;
1495 	h->nr_huge_pages_node[nid]++;
1496 
1497 	if (adjust_surplus) {
1498 		h->surplus_huge_pages++;
1499 		h->surplus_huge_pages_node[nid]++;
1500 	}
1501 
1502 	set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1503 	set_page_private(page, 0);
1504 	SetHPageVmemmapOptimized(page);
1505 
1506 	/*
1507 	 * This page is about to be managed by the hugetlb allocator and
1508 	 * should have no users.  Drop our reference, and check for others
1509 	 * just in case.
1510 	 */
1511 	zeroed = put_page_testzero(page);
1512 	if (!zeroed)
1513 		/*
1514 		 * It is VERY unlikely soneone else has taken a ref on
1515 		 * the page.  In this case, we simply return as the
1516 		 * hugetlb destructor (free_huge_page) will be called
1517 		 * when this other ref is dropped.
1518 		 */
1519 		return;
1520 
1521 	arch_clear_hugepage_flags(page);
1522 	enqueue_huge_page(h, page);
1523 }
1524 
1525 static void __update_and_free_page(struct hstate *h, struct page *page)
1526 {
1527 	int i;
1528 	struct page *subpage = page;
1529 
1530 	if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1531 		return;
1532 
1533 	if (alloc_huge_page_vmemmap(h, page)) {
1534 		spin_lock_irq(&hugetlb_lock);
1535 		/*
1536 		 * If we cannot allocate vmemmap pages, just refuse to free the
1537 		 * page and put the page back on the hugetlb free list and treat
1538 		 * as a surplus page.
1539 		 */
1540 		add_hugetlb_page(h, page, true);
1541 		spin_unlock_irq(&hugetlb_lock);
1542 		return;
1543 	}
1544 
1545 	for (i = 0; i < pages_per_huge_page(h);
1546 	     i++, subpage = mem_map_next(subpage, page, i)) {
1547 		subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1548 				1 << PG_referenced | 1 << PG_dirty |
1549 				1 << PG_active | 1 << PG_private |
1550 				1 << PG_writeback);
1551 	}
1552 
1553 	/*
1554 	 * Non-gigantic pages demoted from CMA allocated gigantic pages
1555 	 * need to be given back to CMA in free_gigantic_page.
1556 	 */
1557 	if (hstate_is_gigantic(h) ||
1558 	    hugetlb_cma_page(page, huge_page_order(h))) {
1559 		destroy_compound_gigantic_page(page, huge_page_order(h));
1560 		free_gigantic_page(page, huge_page_order(h));
1561 	} else {
1562 		__free_pages(page, huge_page_order(h));
1563 	}
1564 }
1565 
1566 /*
1567  * As update_and_free_page() can be called under any context, so we cannot
1568  * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1569  * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1570  * the vmemmap pages.
1571  *
1572  * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1573  * freed and frees them one-by-one. As the page->mapping pointer is going
1574  * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1575  * structure of a lockless linked list of huge pages to be freed.
1576  */
1577 static LLIST_HEAD(hpage_freelist);
1578 
1579 static void free_hpage_workfn(struct work_struct *work)
1580 {
1581 	struct llist_node *node;
1582 
1583 	node = llist_del_all(&hpage_freelist);
1584 
1585 	while (node) {
1586 		struct page *page;
1587 		struct hstate *h;
1588 
1589 		page = container_of((struct address_space **)node,
1590 				     struct page, mapping);
1591 		node = node->next;
1592 		page->mapping = NULL;
1593 		/*
1594 		 * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate()
1595 		 * is going to trigger because a previous call to
1596 		 * remove_hugetlb_page() will set_compound_page_dtor(page,
1597 		 * NULL_COMPOUND_DTOR), so do not use page_hstate() directly.
1598 		 */
1599 		h = size_to_hstate(page_size(page));
1600 
1601 		__update_and_free_page(h, page);
1602 
1603 		cond_resched();
1604 	}
1605 }
1606 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1607 
1608 static inline void flush_free_hpage_work(struct hstate *h)
1609 {
1610 	if (free_vmemmap_pages_per_hpage(h))
1611 		flush_work(&free_hpage_work);
1612 }
1613 
1614 static void update_and_free_page(struct hstate *h, struct page *page,
1615 				 bool atomic)
1616 {
1617 	if (!HPageVmemmapOptimized(page) || !atomic) {
1618 		__update_and_free_page(h, page);
1619 		return;
1620 	}
1621 
1622 	/*
1623 	 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1624 	 *
1625 	 * Only call schedule_work() if hpage_freelist is previously
1626 	 * empty. Otherwise, schedule_work() had been called but the workfn
1627 	 * hasn't retrieved the list yet.
1628 	 */
1629 	if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist))
1630 		schedule_work(&free_hpage_work);
1631 }
1632 
1633 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1634 {
1635 	struct page *page, *t_page;
1636 
1637 	list_for_each_entry_safe(page, t_page, list, lru) {
1638 		update_and_free_page(h, page, false);
1639 		cond_resched();
1640 	}
1641 }
1642 
1643 struct hstate *size_to_hstate(unsigned long size)
1644 {
1645 	struct hstate *h;
1646 
1647 	for_each_hstate(h) {
1648 		if (huge_page_size(h) == size)
1649 			return h;
1650 	}
1651 	return NULL;
1652 }
1653 
1654 void free_huge_page(struct page *page)
1655 {
1656 	/*
1657 	 * Can't pass hstate in here because it is called from the
1658 	 * compound page destructor.
1659 	 */
1660 	struct hstate *h = page_hstate(page);
1661 	int nid = page_to_nid(page);
1662 	struct hugepage_subpool *spool = hugetlb_page_subpool(page);
1663 	bool restore_reserve;
1664 	unsigned long flags;
1665 
1666 	VM_BUG_ON_PAGE(page_count(page), page);
1667 	VM_BUG_ON_PAGE(page_mapcount(page), page);
1668 
1669 	hugetlb_set_page_subpool(page, NULL);
1670 	page->mapping = NULL;
1671 	restore_reserve = HPageRestoreReserve(page);
1672 	ClearHPageRestoreReserve(page);
1673 
1674 	/*
1675 	 * If HPageRestoreReserve was set on page, page allocation consumed a
1676 	 * reservation.  If the page was associated with a subpool, there
1677 	 * would have been a page reserved in the subpool before allocation
1678 	 * via hugepage_subpool_get_pages().  Since we are 'restoring' the
1679 	 * reservation, do not call hugepage_subpool_put_pages() as this will
1680 	 * remove the reserved page from the subpool.
1681 	 */
1682 	if (!restore_reserve) {
1683 		/*
1684 		 * A return code of zero implies that the subpool will be
1685 		 * under its minimum size if the reservation is not restored
1686 		 * after page is free.  Therefore, force restore_reserve
1687 		 * operation.
1688 		 */
1689 		if (hugepage_subpool_put_pages(spool, 1) == 0)
1690 			restore_reserve = true;
1691 	}
1692 
1693 	spin_lock_irqsave(&hugetlb_lock, flags);
1694 	ClearHPageMigratable(page);
1695 	hugetlb_cgroup_uncharge_page(hstate_index(h),
1696 				     pages_per_huge_page(h), page);
1697 	hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1698 					  pages_per_huge_page(h), page);
1699 	if (restore_reserve)
1700 		h->resv_huge_pages++;
1701 
1702 	if (HPageTemporary(page)) {
1703 		remove_hugetlb_page(h, page, false);
1704 		spin_unlock_irqrestore(&hugetlb_lock, flags);
1705 		update_and_free_page(h, page, true);
1706 	} else if (h->surplus_huge_pages_node[nid]) {
1707 		/* remove the page from active list */
1708 		remove_hugetlb_page(h, page, true);
1709 		spin_unlock_irqrestore(&hugetlb_lock, flags);
1710 		update_and_free_page(h, page, true);
1711 	} else {
1712 		arch_clear_hugepage_flags(page);
1713 		enqueue_huge_page(h, page);
1714 		spin_unlock_irqrestore(&hugetlb_lock, flags);
1715 	}
1716 }
1717 
1718 /*
1719  * Must be called with the hugetlb lock held
1720  */
1721 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1722 {
1723 	lockdep_assert_held(&hugetlb_lock);
1724 	h->nr_huge_pages++;
1725 	h->nr_huge_pages_node[nid]++;
1726 }
1727 
1728 static void __prep_new_huge_page(struct hstate *h, struct page *page)
1729 {
1730 	free_huge_page_vmemmap(h, page);
1731 	INIT_LIST_HEAD(&page->lru);
1732 	set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1733 	hugetlb_set_page_subpool(page, NULL);
1734 	set_hugetlb_cgroup(page, NULL);
1735 	set_hugetlb_cgroup_rsvd(page, NULL);
1736 }
1737 
1738 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1739 {
1740 	__prep_new_huge_page(h, page);
1741 	spin_lock_irq(&hugetlb_lock);
1742 	__prep_account_new_huge_page(h, nid);
1743 	spin_unlock_irq(&hugetlb_lock);
1744 }
1745 
1746 static bool __prep_compound_gigantic_page(struct page *page, unsigned int order,
1747 								bool demote)
1748 {
1749 	int i, j;
1750 	int nr_pages = 1 << order;
1751 	struct page *p = page + 1;
1752 
1753 	/* we rely on prep_new_huge_page to set the destructor */
1754 	set_compound_order(page, order);
1755 	__ClearPageReserved(page);
1756 	__SetPageHead(page);
1757 	for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1758 		/*
1759 		 * For gigantic hugepages allocated through bootmem at
1760 		 * boot, it's safer to be consistent with the not-gigantic
1761 		 * hugepages and clear the PG_reserved bit from all tail pages
1762 		 * too.  Otherwise drivers using get_user_pages() to access tail
1763 		 * pages may get the reference counting wrong if they see
1764 		 * PG_reserved set on a tail page (despite the head page not
1765 		 * having PG_reserved set).  Enforcing this consistency between
1766 		 * head and tail pages allows drivers to optimize away a check
1767 		 * on the head page when they need know if put_page() is needed
1768 		 * after get_user_pages().
1769 		 */
1770 		__ClearPageReserved(p);
1771 		/*
1772 		 * Subtle and very unlikely
1773 		 *
1774 		 * Gigantic 'page allocators' such as memblock or cma will
1775 		 * return a set of pages with each page ref counted.  We need
1776 		 * to turn this set of pages into a compound page with tail
1777 		 * page ref counts set to zero.  Code such as speculative page
1778 		 * cache adding could take a ref on a 'to be' tail page.
1779 		 * We need to respect any increased ref count, and only set
1780 		 * the ref count to zero if count is currently 1.  If count
1781 		 * is not 1, we return an error.  An error return indicates
1782 		 * the set of pages can not be converted to a gigantic page.
1783 		 * The caller who allocated the pages should then discard the
1784 		 * pages using the appropriate free interface.
1785 		 *
1786 		 * In the case of demote, the ref count will be zero.
1787 		 */
1788 		if (!demote) {
1789 			if (!page_ref_freeze(p, 1)) {
1790 				pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
1791 				goto out_error;
1792 			}
1793 		} else {
1794 			VM_BUG_ON_PAGE(page_count(p), p);
1795 		}
1796 		set_compound_head(p, page);
1797 	}
1798 	atomic_set(compound_mapcount_ptr(page), -1);
1799 	atomic_set(compound_pincount_ptr(page), 0);
1800 	return true;
1801 
1802 out_error:
1803 	/* undo tail page modifications made above */
1804 	p = page + 1;
1805 	for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) {
1806 		clear_compound_head(p);
1807 		set_page_refcounted(p);
1808 	}
1809 	/* need to clear PG_reserved on remaining tail pages  */
1810 	for (; j < nr_pages; j++, p = mem_map_next(p, page, j))
1811 		__ClearPageReserved(p);
1812 	set_compound_order(page, 0);
1813 	page[1].compound_nr = 0;
1814 	__ClearPageHead(page);
1815 	return false;
1816 }
1817 
1818 static bool prep_compound_gigantic_page(struct page *page, unsigned int order)
1819 {
1820 	return __prep_compound_gigantic_page(page, order, false);
1821 }
1822 
1823 static bool prep_compound_gigantic_page_for_demote(struct page *page,
1824 							unsigned int order)
1825 {
1826 	return __prep_compound_gigantic_page(page, order, true);
1827 }
1828 
1829 /*
1830  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1831  * transparent huge pages.  See the PageTransHuge() documentation for more
1832  * details.
1833  */
1834 int PageHuge(struct page *page)
1835 {
1836 	if (!PageCompound(page))
1837 		return 0;
1838 
1839 	page = compound_head(page);
1840 	return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1841 }
1842 EXPORT_SYMBOL_GPL(PageHuge);
1843 
1844 /*
1845  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1846  * normal or transparent huge pages.
1847  */
1848 int PageHeadHuge(struct page *page_head)
1849 {
1850 	if (!PageHead(page_head))
1851 		return 0;
1852 
1853 	return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1854 }
1855 
1856 /*
1857  * Find and lock address space (mapping) in write mode.
1858  *
1859  * Upon entry, the page is locked which means that page_mapping() is
1860  * stable.  Due to locking order, we can only trylock_write.  If we can
1861  * not get the lock, simply return NULL to caller.
1862  */
1863 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1864 {
1865 	struct address_space *mapping = page_mapping(hpage);
1866 
1867 	if (!mapping)
1868 		return mapping;
1869 
1870 	if (i_mmap_trylock_write(mapping))
1871 		return mapping;
1872 
1873 	return NULL;
1874 }
1875 
1876 pgoff_t hugetlb_basepage_index(struct page *page)
1877 {
1878 	struct page *page_head = compound_head(page);
1879 	pgoff_t index = page_index(page_head);
1880 	unsigned long compound_idx;
1881 
1882 	if (compound_order(page_head) >= MAX_ORDER)
1883 		compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1884 	else
1885 		compound_idx = page - page_head;
1886 
1887 	return (index << compound_order(page_head)) + compound_idx;
1888 }
1889 
1890 static struct page *alloc_buddy_huge_page(struct hstate *h,
1891 		gfp_t gfp_mask, int nid, nodemask_t *nmask,
1892 		nodemask_t *node_alloc_noretry)
1893 {
1894 	int order = huge_page_order(h);
1895 	struct page *page;
1896 	bool alloc_try_hard = true;
1897 
1898 	/*
1899 	 * By default we always try hard to allocate the page with
1900 	 * __GFP_RETRY_MAYFAIL flag.  However, if we are allocating pages in
1901 	 * a loop (to adjust global huge page counts) and previous allocation
1902 	 * failed, do not continue to try hard on the same node.  Use the
1903 	 * node_alloc_noretry bitmap to manage this state information.
1904 	 */
1905 	if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1906 		alloc_try_hard = false;
1907 	gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1908 	if (alloc_try_hard)
1909 		gfp_mask |= __GFP_RETRY_MAYFAIL;
1910 	if (nid == NUMA_NO_NODE)
1911 		nid = numa_mem_id();
1912 	page = __alloc_pages(gfp_mask, order, nid, nmask);
1913 	if (page)
1914 		__count_vm_event(HTLB_BUDDY_PGALLOC);
1915 	else
1916 		__count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1917 
1918 	/*
1919 	 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1920 	 * indicates an overall state change.  Clear bit so that we resume
1921 	 * normal 'try hard' allocations.
1922 	 */
1923 	if (node_alloc_noretry && page && !alloc_try_hard)
1924 		node_clear(nid, *node_alloc_noretry);
1925 
1926 	/*
1927 	 * If we tried hard to get a page but failed, set bit so that
1928 	 * subsequent attempts will not try as hard until there is an
1929 	 * overall state change.
1930 	 */
1931 	if (node_alloc_noretry && !page && alloc_try_hard)
1932 		node_set(nid, *node_alloc_noretry);
1933 
1934 	return page;
1935 }
1936 
1937 /*
1938  * Common helper to allocate a fresh hugetlb page. All specific allocators
1939  * should use this function to get new hugetlb pages
1940  */
1941 static struct page *alloc_fresh_huge_page(struct hstate *h,
1942 		gfp_t gfp_mask, int nid, nodemask_t *nmask,
1943 		nodemask_t *node_alloc_noretry)
1944 {
1945 	struct page *page;
1946 	bool retry = false;
1947 
1948 retry:
1949 	if (hstate_is_gigantic(h))
1950 		page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1951 	else
1952 		page = alloc_buddy_huge_page(h, gfp_mask,
1953 				nid, nmask, node_alloc_noretry);
1954 	if (!page)
1955 		return NULL;
1956 
1957 	if (hstate_is_gigantic(h)) {
1958 		if (!prep_compound_gigantic_page(page, huge_page_order(h))) {
1959 			/*
1960 			 * Rare failure to convert pages to compound page.
1961 			 * Free pages and try again - ONCE!
1962 			 */
1963 			free_gigantic_page(page, huge_page_order(h));
1964 			if (!retry) {
1965 				retry = true;
1966 				goto retry;
1967 			}
1968 			return NULL;
1969 		}
1970 	}
1971 	prep_new_huge_page(h, page, page_to_nid(page));
1972 
1973 	return page;
1974 }
1975 
1976 /*
1977  * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1978  * manner.
1979  */
1980 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1981 				nodemask_t *node_alloc_noretry)
1982 {
1983 	struct page *page;
1984 	int nr_nodes, node;
1985 	gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1986 
1987 	for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1988 		page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1989 						node_alloc_noretry);
1990 		if (page)
1991 			break;
1992 	}
1993 
1994 	if (!page)
1995 		return 0;
1996 
1997 	put_page(page); /* free it into the hugepage allocator */
1998 
1999 	return 1;
2000 }
2001 
2002 /*
2003  * Remove huge page from pool from next node to free.  Attempt to keep
2004  * persistent huge pages more or less balanced over allowed nodes.
2005  * This routine only 'removes' the hugetlb page.  The caller must make
2006  * an additional call to free the page to low level allocators.
2007  * Called with hugetlb_lock locked.
2008  */
2009 static struct page *remove_pool_huge_page(struct hstate *h,
2010 						nodemask_t *nodes_allowed,
2011 						 bool acct_surplus)
2012 {
2013 	int nr_nodes, node;
2014 	struct page *page = NULL;
2015 
2016 	lockdep_assert_held(&hugetlb_lock);
2017 	for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2018 		/*
2019 		 * If we're returning unused surplus pages, only examine
2020 		 * nodes with surplus pages.
2021 		 */
2022 		if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
2023 		    !list_empty(&h->hugepage_freelists[node])) {
2024 			page = list_entry(h->hugepage_freelists[node].next,
2025 					  struct page, lru);
2026 			remove_hugetlb_page(h, page, acct_surplus);
2027 			break;
2028 		}
2029 	}
2030 
2031 	return page;
2032 }
2033 
2034 /*
2035  * Dissolve a given free hugepage into free buddy pages. This function does
2036  * nothing for in-use hugepages and non-hugepages.
2037  * This function returns values like below:
2038  *
2039  *  -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
2040  *           when the system is under memory pressure and the feature of
2041  *           freeing unused vmemmap pages associated with each hugetlb page
2042  *           is enabled.
2043  *  -EBUSY:  failed to dissolved free hugepages or the hugepage is in-use
2044  *           (allocated or reserved.)
2045  *       0:  successfully dissolved free hugepages or the page is not a
2046  *           hugepage (considered as already dissolved)
2047  */
2048 int dissolve_free_huge_page(struct page *page)
2049 {
2050 	int rc = -EBUSY;
2051 
2052 retry:
2053 	/* Not to disrupt normal path by vainly holding hugetlb_lock */
2054 	if (!PageHuge(page))
2055 		return 0;
2056 
2057 	spin_lock_irq(&hugetlb_lock);
2058 	if (!PageHuge(page)) {
2059 		rc = 0;
2060 		goto out;
2061 	}
2062 
2063 	if (!page_count(page)) {
2064 		struct page *head = compound_head(page);
2065 		struct hstate *h = page_hstate(head);
2066 		if (h->free_huge_pages - h->resv_huge_pages == 0)
2067 			goto out;
2068 
2069 		/*
2070 		 * We should make sure that the page is already on the free list
2071 		 * when it is dissolved.
2072 		 */
2073 		if (unlikely(!HPageFreed(head))) {
2074 			spin_unlock_irq(&hugetlb_lock);
2075 			cond_resched();
2076 
2077 			/*
2078 			 * Theoretically, we should return -EBUSY when we
2079 			 * encounter this race. In fact, we have a chance
2080 			 * to successfully dissolve the page if we do a
2081 			 * retry. Because the race window is quite small.
2082 			 * If we seize this opportunity, it is an optimization
2083 			 * for increasing the success rate of dissolving page.
2084 			 */
2085 			goto retry;
2086 		}
2087 
2088 		remove_hugetlb_page(h, head, false);
2089 		h->max_huge_pages--;
2090 		spin_unlock_irq(&hugetlb_lock);
2091 
2092 		/*
2093 		 * Normally update_and_free_page will allocate required vmemmmap
2094 		 * before freeing the page.  update_and_free_page will fail to
2095 		 * free the page if it can not allocate required vmemmap.  We
2096 		 * need to adjust max_huge_pages if the page is not freed.
2097 		 * Attempt to allocate vmemmmap here so that we can take
2098 		 * appropriate action on failure.
2099 		 */
2100 		rc = alloc_huge_page_vmemmap(h, head);
2101 		if (!rc) {
2102 			/*
2103 			 * Move PageHWPoison flag from head page to the raw
2104 			 * error page, which makes any subpages rather than
2105 			 * the error page reusable.
2106 			 */
2107 			if (PageHWPoison(head) && page != head) {
2108 				SetPageHWPoison(page);
2109 				ClearPageHWPoison(head);
2110 			}
2111 			update_and_free_page(h, head, false);
2112 		} else {
2113 			spin_lock_irq(&hugetlb_lock);
2114 			add_hugetlb_page(h, head, false);
2115 			h->max_huge_pages++;
2116 			spin_unlock_irq(&hugetlb_lock);
2117 		}
2118 
2119 		return rc;
2120 	}
2121 out:
2122 	spin_unlock_irq(&hugetlb_lock);
2123 	return rc;
2124 }
2125 
2126 /*
2127  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2128  * make specified memory blocks removable from the system.
2129  * Note that this will dissolve a free gigantic hugepage completely, if any
2130  * part of it lies within the given range.
2131  * Also note that if dissolve_free_huge_page() returns with an error, all
2132  * free hugepages that were dissolved before that error are lost.
2133  */
2134 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2135 {
2136 	unsigned long pfn;
2137 	struct page *page;
2138 	int rc = 0;
2139 
2140 	if (!hugepages_supported())
2141 		return rc;
2142 
2143 	for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
2144 		page = pfn_to_page(pfn);
2145 		rc = dissolve_free_huge_page(page);
2146 		if (rc)
2147 			break;
2148 	}
2149 
2150 	return rc;
2151 }
2152 
2153 /*
2154  * Allocates a fresh surplus page from the page allocator.
2155  */
2156 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
2157 		int nid, nodemask_t *nmask, bool zero_ref)
2158 {
2159 	struct page *page = NULL;
2160 	bool retry = false;
2161 
2162 	if (hstate_is_gigantic(h))
2163 		return NULL;
2164 
2165 	spin_lock_irq(&hugetlb_lock);
2166 	if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2167 		goto out_unlock;
2168 	spin_unlock_irq(&hugetlb_lock);
2169 
2170 retry:
2171 	page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2172 	if (!page)
2173 		return NULL;
2174 
2175 	spin_lock_irq(&hugetlb_lock);
2176 	/*
2177 	 * We could have raced with the pool size change.
2178 	 * Double check that and simply deallocate the new page
2179 	 * if we would end up overcommiting the surpluses. Abuse
2180 	 * temporary page to workaround the nasty free_huge_page
2181 	 * codeflow
2182 	 */
2183 	if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2184 		SetHPageTemporary(page);
2185 		spin_unlock_irq(&hugetlb_lock);
2186 		put_page(page);
2187 		return NULL;
2188 	}
2189 
2190 	if (zero_ref) {
2191 		/*
2192 		 * Caller requires a page with zero ref count.
2193 		 * We will drop ref count here.  If someone else is holding
2194 		 * a ref, the page will be freed when they drop it.  Abuse
2195 		 * temporary page flag to accomplish this.
2196 		 */
2197 		SetHPageTemporary(page);
2198 		if (!put_page_testzero(page)) {
2199 			/*
2200 			 * Unexpected inflated ref count on freshly allocated
2201 			 * huge.  Retry once.
2202 			 */
2203 			pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n");
2204 			spin_unlock_irq(&hugetlb_lock);
2205 			if (retry)
2206 				return NULL;
2207 
2208 			retry = true;
2209 			goto retry;
2210 		}
2211 		ClearHPageTemporary(page);
2212 	}
2213 
2214 	h->surplus_huge_pages++;
2215 	h->surplus_huge_pages_node[page_to_nid(page)]++;
2216 
2217 out_unlock:
2218 	spin_unlock_irq(&hugetlb_lock);
2219 
2220 	return page;
2221 }
2222 
2223 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
2224 				     int nid, nodemask_t *nmask)
2225 {
2226 	struct page *page;
2227 
2228 	if (hstate_is_gigantic(h))
2229 		return NULL;
2230 
2231 	page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2232 	if (!page)
2233 		return NULL;
2234 
2235 	/*
2236 	 * We do not account these pages as surplus because they are only
2237 	 * temporary and will be released properly on the last reference
2238 	 */
2239 	SetHPageTemporary(page);
2240 
2241 	return page;
2242 }
2243 
2244 /*
2245  * Use the VMA's mpolicy to allocate a huge page from the buddy.
2246  */
2247 static
2248 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
2249 		struct vm_area_struct *vma, unsigned long addr)
2250 {
2251 	struct page *page = NULL;
2252 	struct mempolicy *mpol;
2253 	gfp_t gfp_mask = htlb_alloc_mask(h);
2254 	int nid;
2255 	nodemask_t *nodemask;
2256 
2257 	nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2258 	if (mpol_is_preferred_many(mpol)) {
2259 		gfp_t gfp = gfp_mask | __GFP_NOWARN;
2260 
2261 		gfp &=  ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2262 		page = alloc_surplus_huge_page(h, gfp, nid, nodemask, false);
2263 
2264 		/* Fallback to all nodes if page==NULL */
2265 		nodemask = NULL;
2266 	}
2267 
2268 	if (!page)
2269 		page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask, false);
2270 	mpol_cond_put(mpol);
2271 	return page;
2272 }
2273 
2274 /* page migration callback function */
2275 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
2276 		nodemask_t *nmask, gfp_t gfp_mask)
2277 {
2278 	spin_lock_irq(&hugetlb_lock);
2279 	if (h->free_huge_pages - h->resv_huge_pages > 0) {
2280 		struct page *page;
2281 
2282 		page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
2283 		if (page) {
2284 			spin_unlock_irq(&hugetlb_lock);
2285 			return page;
2286 		}
2287 	}
2288 	spin_unlock_irq(&hugetlb_lock);
2289 
2290 	return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2291 }
2292 
2293 /* mempolicy aware migration callback */
2294 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2295 		unsigned long address)
2296 {
2297 	struct mempolicy *mpol;
2298 	nodemask_t *nodemask;
2299 	struct page *page;
2300 	gfp_t gfp_mask;
2301 	int node;
2302 
2303 	gfp_mask = htlb_alloc_mask(h);
2304 	node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2305 	page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
2306 	mpol_cond_put(mpol);
2307 
2308 	return page;
2309 }
2310 
2311 /*
2312  * Increase the hugetlb pool such that it can accommodate a reservation
2313  * of size 'delta'.
2314  */
2315 static int gather_surplus_pages(struct hstate *h, long delta)
2316 	__must_hold(&hugetlb_lock)
2317 {
2318 	struct list_head surplus_list;
2319 	struct page *page, *tmp;
2320 	int ret;
2321 	long i;
2322 	long needed, allocated;
2323 	bool alloc_ok = true;
2324 
2325 	lockdep_assert_held(&hugetlb_lock);
2326 	needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2327 	if (needed <= 0) {
2328 		h->resv_huge_pages += delta;
2329 		return 0;
2330 	}
2331 
2332 	allocated = 0;
2333 	INIT_LIST_HEAD(&surplus_list);
2334 
2335 	ret = -ENOMEM;
2336 retry:
2337 	spin_unlock_irq(&hugetlb_lock);
2338 	for (i = 0; i < needed; i++) {
2339 		page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2340 				NUMA_NO_NODE, NULL, true);
2341 		if (!page) {
2342 			alloc_ok = false;
2343 			break;
2344 		}
2345 		list_add(&page->lru, &surplus_list);
2346 		cond_resched();
2347 	}
2348 	allocated += i;
2349 
2350 	/*
2351 	 * After retaking hugetlb_lock, we need to recalculate 'needed'
2352 	 * because either resv_huge_pages or free_huge_pages may have changed.
2353 	 */
2354 	spin_lock_irq(&hugetlb_lock);
2355 	needed = (h->resv_huge_pages + delta) -
2356 			(h->free_huge_pages + allocated);
2357 	if (needed > 0) {
2358 		if (alloc_ok)
2359 			goto retry;
2360 		/*
2361 		 * We were not able to allocate enough pages to
2362 		 * satisfy the entire reservation so we free what
2363 		 * we've allocated so far.
2364 		 */
2365 		goto free;
2366 	}
2367 	/*
2368 	 * The surplus_list now contains _at_least_ the number of extra pages
2369 	 * needed to accommodate the reservation.  Add the appropriate number
2370 	 * of pages to the hugetlb pool and free the extras back to the buddy
2371 	 * allocator.  Commit the entire reservation here to prevent another
2372 	 * process from stealing the pages as they are added to the pool but
2373 	 * before they are reserved.
2374 	 */
2375 	needed += allocated;
2376 	h->resv_huge_pages += delta;
2377 	ret = 0;
2378 
2379 	/* Free the needed pages to the hugetlb pool */
2380 	list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2381 		if ((--needed) < 0)
2382 			break;
2383 		/* Add the page to the hugetlb allocator */
2384 		enqueue_huge_page(h, page);
2385 	}
2386 free:
2387 	spin_unlock_irq(&hugetlb_lock);
2388 
2389 	/*
2390 	 * Free unnecessary surplus pages to the buddy allocator.
2391 	 * Pages have no ref count, call free_huge_page directly.
2392 	 */
2393 	list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2394 		free_huge_page(page);
2395 	spin_lock_irq(&hugetlb_lock);
2396 
2397 	return ret;
2398 }
2399 
2400 /*
2401  * This routine has two main purposes:
2402  * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2403  *    in unused_resv_pages.  This corresponds to the prior adjustments made
2404  *    to the associated reservation map.
2405  * 2) Free any unused surplus pages that may have been allocated to satisfy
2406  *    the reservation.  As many as unused_resv_pages may be freed.
2407  */
2408 static void return_unused_surplus_pages(struct hstate *h,
2409 					unsigned long unused_resv_pages)
2410 {
2411 	unsigned long nr_pages;
2412 	struct page *page;
2413 	LIST_HEAD(page_list);
2414 
2415 	lockdep_assert_held(&hugetlb_lock);
2416 	/* Uncommit the reservation */
2417 	h->resv_huge_pages -= unused_resv_pages;
2418 
2419 	/* Cannot return gigantic pages currently */
2420 	if (hstate_is_gigantic(h))
2421 		goto out;
2422 
2423 	/*
2424 	 * Part (or even all) of the reservation could have been backed
2425 	 * by pre-allocated pages. Only free surplus pages.
2426 	 */
2427 	nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2428 
2429 	/*
2430 	 * We want to release as many surplus pages as possible, spread
2431 	 * evenly across all nodes with memory. Iterate across these nodes
2432 	 * until we can no longer free unreserved surplus pages. This occurs
2433 	 * when the nodes with surplus pages have no free pages.
2434 	 * remove_pool_huge_page() will balance the freed pages across the
2435 	 * on-line nodes with memory and will handle the hstate accounting.
2436 	 */
2437 	while (nr_pages--) {
2438 		page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2439 		if (!page)
2440 			goto out;
2441 
2442 		list_add(&page->lru, &page_list);
2443 	}
2444 
2445 out:
2446 	spin_unlock_irq(&hugetlb_lock);
2447 	update_and_free_pages_bulk(h, &page_list);
2448 	spin_lock_irq(&hugetlb_lock);
2449 }
2450 
2451 
2452 /*
2453  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2454  * are used by the huge page allocation routines to manage reservations.
2455  *
2456  * vma_needs_reservation is called to determine if the huge page at addr
2457  * within the vma has an associated reservation.  If a reservation is
2458  * needed, the value 1 is returned.  The caller is then responsible for
2459  * managing the global reservation and subpool usage counts.  After
2460  * the huge page has been allocated, vma_commit_reservation is called
2461  * to add the page to the reservation map.  If the page allocation fails,
2462  * the reservation must be ended instead of committed.  vma_end_reservation
2463  * is called in such cases.
2464  *
2465  * In the normal case, vma_commit_reservation returns the same value
2466  * as the preceding vma_needs_reservation call.  The only time this
2467  * is not the case is if a reserve map was changed between calls.  It
2468  * is the responsibility of the caller to notice the difference and
2469  * take appropriate action.
2470  *
2471  * vma_add_reservation is used in error paths where a reservation must
2472  * be restored when a newly allocated huge page must be freed.  It is
2473  * to be called after calling vma_needs_reservation to determine if a
2474  * reservation exists.
2475  *
2476  * vma_del_reservation is used in error paths where an entry in the reserve
2477  * map was created during huge page allocation and must be removed.  It is to
2478  * be called after calling vma_needs_reservation to determine if a reservation
2479  * exists.
2480  */
2481 enum vma_resv_mode {
2482 	VMA_NEEDS_RESV,
2483 	VMA_COMMIT_RESV,
2484 	VMA_END_RESV,
2485 	VMA_ADD_RESV,
2486 	VMA_DEL_RESV,
2487 };
2488 static long __vma_reservation_common(struct hstate *h,
2489 				struct vm_area_struct *vma, unsigned long addr,
2490 				enum vma_resv_mode mode)
2491 {
2492 	struct resv_map *resv;
2493 	pgoff_t idx;
2494 	long ret;
2495 	long dummy_out_regions_needed;
2496 
2497 	resv = vma_resv_map(vma);
2498 	if (!resv)
2499 		return 1;
2500 
2501 	idx = vma_hugecache_offset(h, vma, addr);
2502 	switch (mode) {
2503 	case VMA_NEEDS_RESV:
2504 		ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2505 		/* We assume that vma_reservation_* routines always operate on
2506 		 * 1 page, and that adding to resv map a 1 page entry can only
2507 		 * ever require 1 region.
2508 		 */
2509 		VM_BUG_ON(dummy_out_regions_needed != 1);
2510 		break;
2511 	case VMA_COMMIT_RESV:
2512 		ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2513 		/* region_add calls of range 1 should never fail. */
2514 		VM_BUG_ON(ret < 0);
2515 		break;
2516 	case VMA_END_RESV:
2517 		region_abort(resv, idx, idx + 1, 1);
2518 		ret = 0;
2519 		break;
2520 	case VMA_ADD_RESV:
2521 		if (vma->vm_flags & VM_MAYSHARE) {
2522 			ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2523 			/* region_add calls of range 1 should never fail. */
2524 			VM_BUG_ON(ret < 0);
2525 		} else {
2526 			region_abort(resv, idx, idx + 1, 1);
2527 			ret = region_del(resv, idx, idx + 1);
2528 		}
2529 		break;
2530 	case VMA_DEL_RESV:
2531 		if (vma->vm_flags & VM_MAYSHARE) {
2532 			region_abort(resv, idx, idx + 1, 1);
2533 			ret = region_del(resv, idx, idx + 1);
2534 		} else {
2535 			ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2536 			/* region_add calls of range 1 should never fail. */
2537 			VM_BUG_ON(ret < 0);
2538 		}
2539 		break;
2540 	default:
2541 		BUG();
2542 	}
2543 
2544 	if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2545 		return ret;
2546 	/*
2547 	 * We know private mapping must have HPAGE_RESV_OWNER set.
2548 	 *
2549 	 * In most cases, reserves always exist for private mappings.
2550 	 * However, a file associated with mapping could have been
2551 	 * hole punched or truncated after reserves were consumed.
2552 	 * As subsequent fault on such a range will not use reserves.
2553 	 * Subtle - The reserve map for private mappings has the
2554 	 * opposite meaning than that of shared mappings.  If NO
2555 	 * entry is in the reserve map, it means a reservation exists.
2556 	 * If an entry exists in the reserve map, it means the
2557 	 * reservation has already been consumed.  As a result, the
2558 	 * return value of this routine is the opposite of the
2559 	 * value returned from reserve map manipulation routines above.
2560 	 */
2561 	if (ret > 0)
2562 		return 0;
2563 	if (ret == 0)
2564 		return 1;
2565 	return ret;
2566 }
2567 
2568 static long vma_needs_reservation(struct hstate *h,
2569 			struct vm_area_struct *vma, unsigned long addr)
2570 {
2571 	return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2572 }
2573 
2574 static long vma_commit_reservation(struct hstate *h,
2575 			struct vm_area_struct *vma, unsigned long addr)
2576 {
2577 	return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2578 }
2579 
2580 static void vma_end_reservation(struct hstate *h,
2581 			struct vm_area_struct *vma, unsigned long addr)
2582 {
2583 	(void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2584 }
2585 
2586 static long vma_add_reservation(struct hstate *h,
2587 			struct vm_area_struct *vma, unsigned long addr)
2588 {
2589 	return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2590 }
2591 
2592 static long vma_del_reservation(struct hstate *h,
2593 			struct vm_area_struct *vma, unsigned long addr)
2594 {
2595 	return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2596 }
2597 
2598 /*
2599  * This routine is called to restore reservation information on error paths.
2600  * It should ONLY be called for pages allocated via alloc_huge_page(), and
2601  * the hugetlb mutex should remain held when calling this routine.
2602  *
2603  * It handles two specific cases:
2604  * 1) A reservation was in place and the page consumed the reservation.
2605  *    HPageRestoreReserve is set in the page.
2606  * 2) No reservation was in place for the page, so HPageRestoreReserve is
2607  *    not set.  However, alloc_huge_page always updates the reserve map.
2608  *
2609  * In case 1, free_huge_page later in the error path will increment the
2610  * global reserve count.  But, free_huge_page does not have enough context
2611  * to adjust the reservation map.  This case deals primarily with private
2612  * mappings.  Adjust the reserve map here to be consistent with global
2613  * reserve count adjustments to be made by free_huge_page.  Make sure the
2614  * reserve map indicates there is a reservation present.
2615  *
2616  * In case 2, simply undo reserve map modifications done by alloc_huge_page.
2617  */
2618 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2619 			unsigned long address, struct page *page)
2620 {
2621 	long rc = vma_needs_reservation(h, vma, address);
2622 
2623 	if (HPageRestoreReserve(page)) {
2624 		if (unlikely(rc < 0))
2625 			/*
2626 			 * Rare out of memory condition in reserve map
2627 			 * manipulation.  Clear HPageRestoreReserve so that
2628 			 * global reserve count will not be incremented
2629 			 * by free_huge_page.  This will make it appear
2630 			 * as though the reservation for this page was
2631 			 * consumed.  This may prevent the task from
2632 			 * faulting in the page at a later time.  This
2633 			 * is better than inconsistent global huge page
2634 			 * accounting of reserve counts.
2635 			 */
2636 			ClearHPageRestoreReserve(page);
2637 		else if (rc)
2638 			(void)vma_add_reservation(h, vma, address);
2639 		else
2640 			vma_end_reservation(h, vma, address);
2641 	} else {
2642 		if (!rc) {
2643 			/*
2644 			 * This indicates there is an entry in the reserve map
2645 			 * not added by alloc_huge_page.  We know it was added
2646 			 * before the alloc_huge_page call, otherwise
2647 			 * HPageRestoreReserve would be set on the page.
2648 			 * Remove the entry so that a subsequent allocation
2649 			 * does not consume a reservation.
2650 			 */
2651 			rc = vma_del_reservation(h, vma, address);
2652 			if (rc < 0)
2653 				/*
2654 				 * VERY rare out of memory condition.  Since
2655 				 * we can not delete the entry, set
2656 				 * HPageRestoreReserve so that the reserve
2657 				 * count will be incremented when the page
2658 				 * is freed.  This reserve will be consumed
2659 				 * on a subsequent allocation.
2660 				 */
2661 				SetHPageRestoreReserve(page);
2662 		} else if (rc < 0) {
2663 			/*
2664 			 * Rare out of memory condition from
2665 			 * vma_needs_reservation call.  Memory allocation is
2666 			 * only attempted if a new entry is needed.  Therefore,
2667 			 * this implies there is not an entry in the
2668 			 * reserve map.
2669 			 *
2670 			 * For shared mappings, no entry in the map indicates
2671 			 * no reservation.  We are done.
2672 			 */
2673 			if (!(vma->vm_flags & VM_MAYSHARE))
2674 				/*
2675 				 * For private mappings, no entry indicates
2676 				 * a reservation is present.  Since we can
2677 				 * not add an entry, set SetHPageRestoreReserve
2678 				 * on the page so reserve count will be
2679 				 * incremented when freed.  This reserve will
2680 				 * be consumed on a subsequent allocation.
2681 				 */
2682 				SetHPageRestoreReserve(page);
2683 		} else
2684 			/*
2685 			 * No reservation present, do nothing
2686 			 */
2687 			 vma_end_reservation(h, vma, address);
2688 	}
2689 }
2690 
2691 /*
2692  * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one
2693  * @h: struct hstate old page belongs to
2694  * @old_page: Old page to dissolve
2695  * @list: List to isolate the page in case we need to
2696  * Returns 0 on success, otherwise negated error.
2697  */
2698 static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page,
2699 					struct list_head *list)
2700 {
2701 	gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2702 	int nid = page_to_nid(old_page);
2703 	bool alloc_retry = false;
2704 	struct page *new_page;
2705 	int ret = 0;
2706 
2707 	/*
2708 	 * Before dissolving the page, we need to allocate a new one for the
2709 	 * pool to remain stable.  Here, we allocate the page and 'prep' it
2710 	 * by doing everything but actually updating counters and adding to
2711 	 * the pool.  This simplifies and let us do most of the processing
2712 	 * under the lock.
2713 	 */
2714 alloc_retry:
2715 	new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL);
2716 	if (!new_page)
2717 		return -ENOMEM;
2718 	/*
2719 	 * If all goes well, this page will be directly added to the free
2720 	 * list in the pool.  For this the ref count needs to be zero.
2721 	 * Attempt to drop now, and retry once if needed.  It is VERY
2722 	 * unlikely there is another ref on the page.
2723 	 *
2724 	 * If someone else has a reference to the page, it will be freed
2725 	 * when they drop their ref.  Abuse temporary page flag to accomplish
2726 	 * this.  Retry once if there is an inflated ref count.
2727 	 */
2728 	SetHPageTemporary(new_page);
2729 	if (!put_page_testzero(new_page)) {
2730 		if (alloc_retry)
2731 			return -EBUSY;
2732 
2733 		alloc_retry = true;
2734 		goto alloc_retry;
2735 	}
2736 	ClearHPageTemporary(new_page);
2737 
2738 	__prep_new_huge_page(h, new_page);
2739 
2740 retry:
2741 	spin_lock_irq(&hugetlb_lock);
2742 	if (!PageHuge(old_page)) {
2743 		/*
2744 		 * Freed from under us. Drop new_page too.
2745 		 */
2746 		goto free_new;
2747 	} else if (page_count(old_page)) {
2748 		/*
2749 		 * Someone has grabbed the page, try to isolate it here.
2750 		 * Fail with -EBUSY if not possible.
2751 		 */
2752 		spin_unlock_irq(&hugetlb_lock);
2753 		if (!isolate_huge_page(old_page, list))
2754 			ret = -EBUSY;
2755 		spin_lock_irq(&hugetlb_lock);
2756 		goto free_new;
2757 	} else if (!HPageFreed(old_page)) {
2758 		/*
2759 		 * Page's refcount is 0 but it has not been enqueued in the
2760 		 * freelist yet. Race window is small, so we can succeed here if
2761 		 * we retry.
2762 		 */
2763 		spin_unlock_irq(&hugetlb_lock);
2764 		cond_resched();
2765 		goto retry;
2766 	} else {
2767 		/*
2768 		 * Ok, old_page is still a genuine free hugepage. Remove it from
2769 		 * the freelist and decrease the counters. These will be
2770 		 * incremented again when calling __prep_account_new_huge_page()
2771 		 * and enqueue_huge_page() for new_page. The counters will remain
2772 		 * stable since this happens under the lock.
2773 		 */
2774 		remove_hugetlb_page(h, old_page, false);
2775 
2776 		/*
2777 		 * Ref count on new page is already zero as it was dropped
2778 		 * earlier.  It can be directly added to the pool free list.
2779 		 */
2780 		__prep_account_new_huge_page(h, nid);
2781 		enqueue_huge_page(h, new_page);
2782 
2783 		/*
2784 		 * Pages have been replaced, we can safely free the old one.
2785 		 */
2786 		spin_unlock_irq(&hugetlb_lock);
2787 		update_and_free_page(h, old_page, false);
2788 	}
2789 
2790 	return ret;
2791 
2792 free_new:
2793 	spin_unlock_irq(&hugetlb_lock);
2794 	/* Page has a zero ref count, but needs a ref to be freed */
2795 	set_page_refcounted(new_page);
2796 	update_and_free_page(h, new_page, false);
2797 
2798 	return ret;
2799 }
2800 
2801 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
2802 {
2803 	struct hstate *h;
2804 	struct page *head;
2805 	int ret = -EBUSY;
2806 
2807 	/*
2808 	 * The page might have been dissolved from under our feet, so make sure
2809 	 * to carefully check the state under the lock.
2810 	 * Return success when racing as if we dissolved the page ourselves.
2811 	 */
2812 	spin_lock_irq(&hugetlb_lock);
2813 	if (PageHuge(page)) {
2814 		head = compound_head(page);
2815 		h = page_hstate(head);
2816 	} else {
2817 		spin_unlock_irq(&hugetlb_lock);
2818 		return 0;
2819 	}
2820 	spin_unlock_irq(&hugetlb_lock);
2821 
2822 	/*
2823 	 * Fence off gigantic pages as there is a cyclic dependency between
2824 	 * alloc_contig_range and them. Return -ENOMEM as this has the effect
2825 	 * of bailing out right away without further retrying.
2826 	 */
2827 	if (hstate_is_gigantic(h))
2828 		return -ENOMEM;
2829 
2830 	if (page_count(head) && isolate_huge_page(head, list))
2831 		ret = 0;
2832 	else if (!page_count(head))
2833 		ret = alloc_and_dissolve_huge_page(h, head, list);
2834 
2835 	return ret;
2836 }
2837 
2838 struct page *alloc_huge_page(struct vm_area_struct *vma,
2839 				    unsigned long addr, int avoid_reserve)
2840 {
2841 	struct hugepage_subpool *spool = subpool_vma(vma);
2842 	struct hstate *h = hstate_vma(vma);
2843 	struct page *page;
2844 	long map_chg, map_commit;
2845 	long gbl_chg;
2846 	int ret, idx;
2847 	struct hugetlb_cgroup *h_cg;
2848 	bool deferred_reserve;
2849 
2850 	idx = hstate_index(h);
2851 	/*
2852 	 * Examine the region/reserve map to determine if the process
2853 	 * has a reservation for the page to be allocated.  A return
2854 	 * code of zero indicates a reservation exists (no change).
2855 	 */
2856 	map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2857 	if (map_chg < 0)
2858 		return ERR_PTR(-ENOMEM);
2859 
2860 	/*
2861 	 * Processes that did not create the mapping will have no
2862 	 * reserves as indicated by the region/reserve map. Check
2863 	 * that the allocation will not exceed the subpool limit.
2864 	 * Allocations for MAP_NORESERVE mappings also need to be
2865 	 * checked against any subpool limit.
2866 	 */
2867 	if (map_chg || avoid_reserve) {
2868 		gbl_chg = hugepage_subpool_get_pages(spool, 1);
2869 		if (gbl_chg < 0) {
2870 			vma_end_reservation(h, vma, addr);
2871 			return ERR_PTR(-ENOSPC);
2872 		}
2873 
2874 		/*
2875 		 * Even though there was no reservation in the region/reserve
2876 		 * map, there could be reservations associated with the
2877 		 * subpool that can be used.  This would be indicated if the
2878 		 * return value of hugepage_subpool_get_pages() is zero.
2879 		 * However, if avoid_reserve is specified we still avoid even
2880 		 * the subpool reservations.
2881 		 */
2882 		if (avoid_reserve)
2883 			gbl_chg = 1;
2884 	}
2885 
2886 	/* If this allocation is not consuming a reservation, charge it now.
2887 	 */
2888 	deferred_reserve = map_chg || avoid_reserve;
2889 	if (deferred_reserve) {
2890 		ret = hugetlb_cgroup_charge_cgroup_rsvd(
2891 			idx, pages_per_huge_page(h), &h_cg);
2892 		if (ret)
2893 			goto out_subpool_put;
2894 	}
2895 
2896 	ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2897 	if (ret)
2898 		goto out_uncharge_cgroup_reservation;
2899 
2900 	spin_lock_irq(&hugetlb_lock);
2901 	/*
2902 	 * glb_chg is passed to indicate whether or not a page must be taken
2903 	 * from the global free pool (global change).  gbl_chg == 0 indicates
2904 	 * a reservation exists for the allocation.
2905 	 */
2906 	page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2907 	if (!page) {
2908 		spin_unlock_irq(&hugetlb_lock);
2909 		page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2910 		if (!page)
2911 			goto out_uncharge_cgroup;
2912 		if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2913 			SetHPageRestoreReserve(page);
2914 			h->resv_huge_pages--;
2915 		}
2916 		spin_lock_irq(&hugetlb_lock);
2917 		list_add(&page->lru, &h->hugepage_activelist);
2918 		/* Fall through */
2919 	}
2920 	hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2921 	/* If allocation is not consuming a reservation, also store the
2922 	 * hugetlb_cgroup pointer on the page.
2923 	 */
2924 	if (deferred_reserve) {
2925 		hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2926 						  h_cg, page);
2927 	}
2928 
2929 	spin_unlock_irq(&hugetlb_lock);
2930 
2931 	hugetlb_set_page_subpool(page, spool);
2932 
2933 	map_commit = vma_commit_reservation(h, vma, addr);
2934 	if (unlikely(map_chg > map_commit)) {
2935 		/*
2936 		 * The page was added to the reservation map between
2937 		 * vma_needs_reservation and vma_commit_reservation.
2938 		 * This indicates a race with hugetlb_reserve_pages.
2939 		 * Adjust for the subpool count incremented above AND
2940 		 * in hugetlb_reserve_pages for the same page.  Also,
2941 		 * the reservation count added in hugetlb_reserve_pages
2942 		 * no longer applies.
2943 		 */
2944 		long rsv_adjust;
2945 
2946 		rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2947 		hugetlb_acct_memory(h, -rsv_adjust);
2948 		if (deferred_reserve)
2949 			hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2950 					pages_per_huge_page(h), page);
2951 	}
2952 	return page;
2953 
2954 out_uncharge_cgroup:
2955 	hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2956 out_uncharge_cgroup_reservation:
2957 	if (deferred_reserve)
2958 		hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2959 						    h_cg);
2960 out_subpool_put:
2961 	if (map_chg || avoid_reserve)
2962 		hugepage_subpool_put_pages(spool, 1);
2963 	vma_end_reservation(h, vma, addr);
2964 	return ERR_PTR(-ENOSPC);
2965 }
2966 
2967 int alloc_bootmem_huge_page(struct hstate *h, int nid)
2968 	__attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2969 int __alloc_bootmem_huge_page(struct hstate *h, int nid)
2970 {
2971 	struct huge_bootmem_page *m = NULL; /* initialize for clang */
2972 	int nr_nodes, node;
2973 
2974 	if (nid >= nr_online_nodes)
2975 		return 0;
2976 	/* do node specific alloc */
2977 	if (nid != NUMA_NO_NODE) {
2978 		m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h),
2979 				0, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
2980 		if (!m)
2981 			return 0;
2982 		goto found;
2983 	}
2984 	/* allocate from next node when distributing huge pages */
2985 	for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2986 		m = memblock_alloc_try_nid_raw(
2987 				huge_page_size(h), huge_page_size(h),
2988 				0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2989 		/*
2990 		 * Use the beginning of the huge page to store the
2991 		 * huge_bootmem_page struct (until gather_bootmem
2992 		 * puts them into the mem_map).
2993 		 */
2994 		if (!m)
2995 			return 0;
2996 		goto found;
2997 	}
2998 
2999 found:
3000 	/* Put them into a private list first because mem_map is not up yet */
3001 	INIT_LIST_HEAD(&m->list);
3002 	list_add(&m->list, &huge_boot_pages);
3003 	m->hstate = h;
3004 	return 1;
3005 }
3006 
3007 /*
3008  * Put bootmem huge pages into the standard lists after mem_map is up.
3009  * Note: This only applies to gigantic (order > MAX_ORDER) pages.
3010  */
3011 static void __init gather_bootmem_prealloc(void)
3012 {
3013 	struct huge_bootmem_page *m;
3014 
3015 	list_for_each_entry(m, &huge_boot_pages, list) {
3016 		struct page *page = virt_to_page(m);
3017 		struct hstate *h = m->hstate;
3018 
3019 		VM_BUG_ON(!hstate_is_gigantic(h));
3020 		WARN_ON(page_count(page) != 1);
3021 		if (prep_compound_gigantic_page(page, huge_page_order(h))) {
3022 			WARN_ON(PageReserved(page));
3023 			prep_new_huge_page(h, page, page_to_nid(page));
3024 			put_page(page); /* add to the hugepage allocator */
3025 		} else {
3026 			/* VERY unlikely inflated ref count on a tail page */
3027 			free_gigantic_page(page, huge_page_order(h));
3028 		}
3029 
3030 		/*
3031 		 * We need to restore the 'stolen' pages to totalram_pages
3032 		 * in order to fix confusing memory reports from free(1) and
3033 		 * other side-effects, like CommitLimit going negative.
3034 		 */
3035 		adjust_managed_page_count(page, pages_per_huge_page(h));
3036 		cond_resched();
3037 	}
3038 }
3039 static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)
3040 {
3041 	unsigned long i;
3042 	char buf[32];
3043 
3044 	for (i = 0; i < h->max_huge_pages_node[nid]; ++i) {
3045 		if (hstate_is_gigantic(h)) {
3046 			if (!alloc_bootmem_huge_page(h, nid))
3047 				break;
3048 		} else {
3049 			struct page *page;
3050 			gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3051 
3052 			page = alloc_fresh_huge_page(h, gfp_mask, nid,
3053 					&node_states[N_MEMORY], NULL);
3054 			if (!page)
3055 				break;
3056 			put_page(page); /* free it into the hugepage allocator */
3057 		}
3058 		cond_resched();
3059 	}
3060 	if (i == h->max_huge_pages_node[nid])
3061 		return;
3062 
3063 	string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3064 	pr_warn("HugeTLB: allocating %u of page size %s failed node%d.  Only allocated %lu hugepages.\n",
3065 		h->max_huge_pages_node[nid], buf, nid, i);
3066 	h->max_huge_pages -= (h->max_huge_pages_node[nid] - i);
3067 	h->max_huge_pages_node[nid] = i;
3068 }
3069 
3070 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
3071 {
3072 	unsigned long i;
3073 	nodemask_t *node_alloc_noretry;
3074 	bool node_specific_alloc = false;
3075 
3076 	/* skip gigantic hugepages allocation if hugetlb_cma enabled */
3077 	if (hstate_is_gigantic(h) && hugetlb_cma_size) {
3078 		pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
3079 		return;
3080 	}
3081 
3082 	/* do node specific alloc */
3083 	for (i = 0; i < nr_online_nodes; i++) {
3084 		if (h->max_huge_pages_node[i] > 0) {
3085 			hugetlb_hstate_alloc_pages_onenode(h, i);
3086 			node_specific_alloc = true;
3087 		}
3088 	}
3089 
3090 	if (node_specific_alloc)
3091 		return;
3092 
3093 	/* below will do all node balanced alloc */
3094 	if (!hstate_is_gigantic(h)) {
3095 		/*
3096 		 * Bit mask controlling how hard we retry per-node allocations.
3097 		 * Ignore errors as lower level routines can deal with
3098 		 * node_alloc_noretry == NULL.  If this kmalloc fails at boot
3099 		 * time, we are likely in bigger trouble.
3100 		 */
3101 		node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
3102 						GFP_KERNEL);
3103 	} else {
3104 		/* allocations done at boot time */
3105 		node_alloc_noretry = NULL;
3106 	}
3107 
3108 	/* bit mask controlling how hard we retry per-node allocations */
3109 	if (node_alloc_noretry)
3110 		nodes_clear(*node_alloc_noretry);
3111 
3112 	for (i = 0; i < h->max_huge_pages; ++i) {
3113 		if (hstate_is_gigantic(h)) {
3114 			if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))
3115 				break;
3116 		} else if (!alloc_pool_huge_page(h,
3117 					 &node_states[N_MEMORY],
3118 					 node_alloc_noretry))
3119 			break;
3120 		cond_resched();
3121 	}
3122 	if (i < h->max_huge_pages) {
3123 		char buf[32];
3124 
3125 		string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3126 		pr_warn("HugeTLB: allocating %lu of page size %s failed.  Only allocated %lu hugepages.\n",
3127 			h->max_huge_pages, buf, i);
3128 		h->max_huge_pages = i;
3129 	}
3130 	kfree(node_alloc_noretry);
3131 }
3132 
3133 static void __init hugetlb_init_hstates(void)
3134 {
3135 	struct hstate *h, *h2;
3136 
3137 	for_each_hstate(h) {
3138 		if (minimum_order > huge_page_order(h))
3139 			minimum_order = huge_page_order(h);
3140 
3141 		/* oversize hugepages were init'ed in early boot */
3142 		if (!hstate_is_gigantic(h))
3143 			hugetlb_hstate_alloc_pages(h);
3144 
3145 		/*
3146 		 * Set demote order for each hstate.  Note that
3147 		 * h->demote_order is initially 0.
3148 		 * - We can not demote gigantic pages if runtime freeing
3149 		 *   is not supported, so skip this.
3150 		 * - If CMA allocation is possible, we can not demote
3151 		 *   HUGETLB_PAGE_ORDER or smaller size pages.
3152 		 */
3153 		if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3154 			continue;
3155 		if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER)
3156 			continue;
3157 		for_each_hstate(h2) {
3158 			if (h2 == h)
3159 				continue;
3160 			if (h2->order < h->order &&
3161 			    h2->order > h->demote_order)
3162 				h->demote_order = h2->order;
3163 		}
3164 	}
3165 	VM_BUG_ON(minimum_order == UINT_MAX);
3166 }
3167 
3168 static void __init report_hugepages(void)
3169 {
3170 	struct hstate *h;
3171 
3172 	for_each_hstate(h) {
3173 		char buf[32];
3174 
3175 		string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3176 		pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
3177 			buf, h->free_huge_pages);
3178 	}
3179 }
3180 
3181 #ifdef CONFIG_HIGHMEM
3182 static void try_to_free_low(struct hstate *h, unsigned long count,
3183 						nodemask_t *nodes_allowed)
3184 {
3185 	int i;
3186 	LIST_HEAD(page_list);
3187 
3188 	lockdep_assert_held(&hugetlb_lock);
3189 	if (hstate_is_gigantic(h))
3190 		return;
3191 
3192 	/*
3193 	 * Collect pages to be freed on a list, and free after dropping lock
3194 	 */
3195 	for_each_node_mask(i, *nodes_allowed) {
3196 		struct page *page, *next;
3197 		struct list_head *freel = &h->hugepage_freelists[i];
3198 		list_for_each_entry_safe(page, next, freel, lru) {
3199 			if (count >= h->nr_huge_pages)
3200 				goto out;
3201 			if (PageHighMem(page))
3202 				continue;
3203 			remove_hugetlb_page(h, page, false);
3204 			list_add(&page->lru, &page_list);
3205 		}
3206 	}
3207 
3208 out:
3209 	spin_unlock_irq(&hugetlb_lock);
3210 	update_and_free_pages_bulk(h, &page_list);
3211 	spin_lock_irq(&hugetlb_lock);
3212 }
3213 #else
3214 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3215 						nodemask_t *nodes_allowed)
3216 {
3217 }
3218 #endif
3219 
3220 /*
3221  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
3222  * balanced by operating on them in a round-robin fashion.
3223  * Returns 1 if an adjustment was made.
3224  */
3225 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3226 				int delta)
3227 {
3228 	int nr_nodes, node;
3229 
3230 	lockdep_assert_held(&hugetlb_lock);
3231 	VM_BUG_ON(delta != -1 && delta != 1);
3232 
3233 	if (delta < 0) {
3234 		for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3235 			if (h->surplus_huge_pages_node[node])
3236 				goto found;
3237 		}
3238 	} else {
3239 		for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3240 			if (h->surplus_huge_pages_node[node] <
3241 					h->nr_huge_pages_node[node])
3242 				goto found;
3243 		}
3244 	}
3245 	return 0;
3246 
3247 found:
3248 	h->surplus_huge_pages += delta;
3249 	h->surplus_huge_pages_node[node] += delta;
3250 	return 1;
3251 }
3252 
3253 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3254 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3255 			      nodemask_t *nodes_allowed)
3256 {
3257 	unsigned long min_count, ret;
3258 	struct page *page;
3259 	LIST_HEAD(page_list);
3260 	NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3261 
3262 	/*
3263 	 * Bit mask controlling how hard we retry per-node allocations.
3264 	 * If we can not allocate the bit mask, do not attempt to allocate
3265 	 * the requested huge pages.
3266 	 */
3267 	if (node_alloc_noretry)
3268 		nodes_clear(*node_alloc_noretry);
3269 	else
3270 		return -ENOMEM;
3271 
3272 	/*
3273 	 * resize_lock mutex prevents concurrent adjustments to number of
3274 	 * pages in hstate via the proc/sysfs interfaces.
3275 	 */
3276 	mutex_lock(&h->resize_lock);
3277 	flush_free_hpage_work(h);
3278 	spin_lock_irq(&hugetlb_lock);
3279 
3280 	/*
3281 	 * Check for a node specific request.
3282 	 * Changing node specific huge page count may require a corresponding
3283 	 * change to the global count.  In any case, the passed node mask
3284 	 * (nodes_allowed) will restrict alloc/free to the specified node.
3285 	 */
3286 	if (nid != NUMA_NO_NODE) {
3287 		unsigned long old_count = count;
3288 
3289 		count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3290 		/*
3291 		 * User may have specified a large count value which caused the
3292 		 * above calculation to overflow.  In this case, they wanted
3293 		 * to allocate as many huge pages as possible.  Set count to
3294 		 * largest possible value to align with their intention.
3295 		 */
3296 		if (count < old_count)
3297 			count = ULONG_MAX;
3298 	}
3299 
3300 	/*
3301 	 * Gigantic pages runtime allocation depend on the capability for large
3302 	 * page range allocation.
3303 	 * If the system does not provide this feature, return an error when
3304 	 * the user tries to allocate gigantic pages but let the user free the
3305 	 * boottime allocated gigantic pages.
3306 	 */
3307 	if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3308 		if (count > persistent_huge_pages(h)) {
3309 			spin_unlock_irq(&hugetlb_lock);
3310 			mutex_unlock(&h->resize_lock);
3311 			NODEMASK_FREE(node_alloc_noretry);
3312 			return -EINVAL;
3313 		}
3314 		/* Fall through to decrease pool */
3315 	}
3316 
3317 	/*
3318 	 * Increase the pool size
3319 	 * First take pages out of surplus state.  Then make up the
3320 	 * remaining difference by allocating fresh huge pages.
3321 	 *
3322 	 * We might race with alloc_surplus_huge_page() here and be unable
3323 	 * to convert a surplus huge page to a normal huge page. That is
3324 	 * not critical, though, it just means the overall size of the
3325 	 * pool might be one hugepage larger than it needs to be, but
3326 	 * within all the constraints specified by the sysctls.
3327 	 */
3328 	while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3329 		if (!adjust_pool_surplus(h, nodes_allowed, -1))
3330 			break;
3331 	}
3332 
3333 	while (count > persistent_huge_pages(h)) {
3334 		/*
3335 		 * If this allocation races such that we no longer need the
3336 		 * page, free_huge_page will handle it by freeing the page
3337 		 * and reducing the surplus.
3338 		 */
3339 		spin_unlock_irq(&hugetlb_lock);
3340 
3341 		/* yield cpu to avoid soft lockup */
3342 		cond_resched();
3343 
3344 		ret = alloc_pool_huge_page(h, nodes_allowed,
3345 						node_alloc_noretry);
3346 		spin_lock_irq(&hugetlb_lock);
3347 		if (!ret)
3348 			goto out;
3349 
3350 		/* Bail for signals. Probably ctrl-c from user */
3351 		if (signal_pending(current))
3352 			goto out;
3353 	}
3354 
3355 	/*
3356 	 * Decrease the pool size
3357 	 * First return free pages to the buddy allocator (being careful
3358 	 * to keep enough around to satisfy reservations).  Then place
3359 	 * pages into surplus state as needed so the pool will shrink
3360 	 * to the desired size as pages become free.
3361 	 *
3362 	 * By placing pages into the surplus state independent of the
3363 	 * overcommit value, we are allowing the surplus pool size to
3364 	 * exceed overcommit. There are few sane options here. Since
3365 	 * alloc_surplus_huge_page() is checking the global counter,
3366 	 * though, we'll note that we're not allowed to exceed surplus
3367 	 * and won't grow the pool anywhere else. Not until one of the
3368 	 * sysctls are changed, or the surplus pages go out of use.
3369 	 */
3370 	min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3371 	min_count = max(count, min_count);
3372 	try_to_free_low(h, min_count, nodes_allowed);
3373 
3374 	/*
3375 	 * Collect pages to be removed on list without dropping lock
3376 	 */
3377 	while (min_count < persistent_huge_pages(h)) {
3378 		page = remove_pool_huge_page(h, nodes_allowed, 0);
3379 		if (!page)
3380 			break;
3381 
3382 		list_add(&page->lru, &page_list);
3383 	}
3384 	/* free the pages after dropping lock */
3385 	spin_unlock_irq(&hugetlb_lock);
3386 	update_and_free_pages_bulk(h, &page_list);
3387 	flush_free_hpage_work(h);
3388 	spin_lock_irq(&hugetlb_lock);
3389 
3390 	while (count < persistent_huge_pages(h)) {
3391 		if (!adjust_pool_surplus(h, nodes_allowed, 1))
3392 			break;
3393 	}
3394 out:
3395 	h->max_huge_pages = persistent_huge_pages(h);
3396 	spin_unlock_irq(&hugetlb_lock);
3397 	mutex_unlock(&h->resize_lock);
3398 
3399 	NODEMASK_FREE(node_alloc_noretry);
3400 
3401 	return 0;
3402 }
3403 
3404 static int demote_free_huge_page(struct hstate *h, struct page *page)
3405 {
3406 	int i, nid = page_to_nid(page);
3407 	struct hstate *target_hstate;
3408 	int rc = 0;
3409 
3410 	target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order);
3411 
3412 	remove_hugetlb_page_for_demote(h, page, false);
3413 	spin_unlock_irq(&hugetlb_lock);
3414 
3415 	rc = alloc_huge_page_vmemmap(h, page);
3416 	if (rc) {
3417 		/* Allocation of vmemmmap failed, we can not demote page */
3418 		spin_lock_irq(&hugetlb_lock);
3419 		set_page_refcounted(page);
3420 		add_hugetlb_page(h, page, false);
3421 		return rc;
3422 	}
3423 
3424 	/*
3425 	 * Use destroy_compound_hugetlb_page_for_demote for all huge page
3426 	 * sizes as it will not ref count pages.
3427 	 */
3428 	destroy_compound_hugetlb_page_for_demote(page, huge_page_order(h));
3429 
3430 	/*
3431 	 * Taking target hstate mutex synchronizes with set_max_huge_pages.
3432 	 * Without the mutex, pages added to target hstate could be marked
3433 	 * as surplus.
3434 	 *
3435 	 * Note that we already hold h->resize_lock.  To prevent deadlock,
3436 	 * use the convention of always taking larger size hstate mutex first.
3437 	 */
3438 	mutex_lock(&target_hstate->resize_lock);
3439 	for (i = 0; i < pages_per_huge_page(h);
3440 				i += pages_per_huge_page(target_hstate)) {
3441 		if (hstate_is_gigantic(target_hstate))
3442 			prep_compound_gigantic_page_for_demote(page + i,
3443 							target_hstate->order);
3444 		else
3445 			prep_compound_page(page + i, target_hstate->order);
3446 		set_page_private(page + i, 0);
3447 		set_page_refcounted(page + i);
3448 		prep_new_huge_page(target_hstate, page + i, nid);
3449 		put_page(page + i);
3450 	}
3451 	mutex_unlock(&target_hstate->resize_lock);
3452 
3453 	spin_lock_irq(&hugetlb_lock);
3454 
3455 	/*
3456 	 * Not absolutely necessary, but for consistency update max_huge_pages
3457 	 * based on pool changes for the demoted page.
3458 	 */
3459 	h->max_huge_pages--;
3460 	target_hstate->max_huge_pages += pages_per_huge_page(h);
3461 
3462 	return rc;
3463 }
3464 
3465 static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
3466 	__must_hold(&hugetlb_lock)
3467 {
3468 	int nr_nodes, node;
3469 	struct page *page;
3470 	int rc = 0;
3471 
3472 	lockdep_assert_held(&hugetlb_lock);
3473 
3474 	/* We should never get here if no demote order */
3475 	if (!h->demote_order) {
3476 		pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n");
3477 		return -EINVAL;		/* internal error */
3478 	}
3479 
3480 	for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3481 		if (!list_empty(&h->hugepage_freelists[node])) {
3482 			page = list_entry(h->hugepage_freelists[node].next,
3483 					struct page, lru);
3484 			rc = demote_free_huge_page(h, page);
3485 			break;
3486 		}
3487 	}
3488 
3489 	return rc;
3490 }
3491 
3492 #define HSTATE_ATTR_RO(_name) \
3493 	static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3494 
3495 #define HSTATE_ATTR_WO(_name) \
3496 	static struct kobj_attribute _name##_attr = __ATTR_WO(_name)
3497 
3498 #define HSTATE_ATTR(_name) \
3499 	static struct kobj_attribute _name##_attr = \
3500 		__ATTR(_name, 0644, _name##_show, _name##_store)
3501 
3502 static struct kobject *hugepages_kobj;
3503 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3504 
3505 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3506 
3507 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3508 {
3509 	int i;
3510 
3511 	for (i = 0; i < HUGE_MAX_HSTATE; i++)
3512 		if (hstate_kobjs[i] == kobj) {
3513 			if (nidp)
3514 				*nidp = NUMA_NO_NODE;
3515 			return &hstates[i];
3516 		}
3517 
3518 	return kobj_to_node_hstate(kobj, nidp);
3519 }
3520 
3521 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3522 					struct kobj_attribute *attr, char *buf)
3523 {
3524 	struct hstate *h;
3525 	unsigned long nr_huge_pages;
3526 	int nid;
3527 
3528 	h = kobj_to_hstate(kobj, &nid);
3529 	if (nid == NUMA_NO_NODE)
3530 		nr_huge_pages = h->nr_huge_pages;
3531 	else
3532 		nr_huge_pages = h->nr_huge_pages_node[nid];
3533 
3534 	return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3535 }
3536 
3537 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3538 					   struct hstate *h, int nid,
3539 					   unsigned long count, size_t len)
3540 {
3541 	int err;
3542 	nodemask_t nodes_allowed, *n_mask;
3543 
3544 	if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3545 		return -EINVAL;
3546 
3547 	if (nid == NUMA_NO_NODE) {
3548 		/*
3549 		 * global hstate attribute
3550 		 */
3551 		if (!(obey_mempolicy &&
3552 				init_nodemask_of_mempolicy(&nodes_allowed)))
3553 			n_mask = &node_states[N_MEMORY];
3554 		else
3555 			n_mask = &nodes_allowed;
3556 	} else {
3557 		/*
3558 		 * Node specific request.  count adjustment happens in
3559 		 * set_max_huge_pages() after acquiring hugetlb_lock.
3560 		 */
3561 		init_nodemask_of_node(&nodes_allowed, nid);
3562 		n_mask = &nodes_allowed;
3563 	}
3564 
3565 	err = set_max_huge_pages(h, count, nid, n_mask);
3566 
3567 	return err ? err : len;
3568 }
3569 
3570 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3571 					 struct kobject *kobj, const char *buf,
3572 					 size_t len)
3573 {
3574 	struct hstate *h;
3575 	unsigned long count;
3576 	int nid;
3577 	int err;
3578 
3579 	err = kstrtoul(buf, 10, &count);
3580 	if (err)
3581 		return err;
3582 
3583 	h = kobj_to_hstate(kobj, &nid);
3584 	return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3585 }
3586 
3587 static ssize_t nr_hugepages_show(struct kobject *kobj,
3588 				       struct kobj_attribute *attr, char *buf)
3589 {
3590 	return nr_hugepages_show_common(kobj, attr, buf);
3591 }
3592 
3593 static ssize_t nr_hugepages_store(struct kobject *kobj,
3594 	       struct kobj_attribute *attr, const char *buf, size_t len)
3595 {
3596 	return nr_hugepages_store_common(false, kobj, buf, len);
3597 }
3598 HSTATE_ATTR(nr_hugepages);
3599 
3600 #ifdef CONFIG_NUMA
3601 
3602 /*
3603  * hstate attribute for optionally mempolicy-based constraint on persistent
3604  * huge page alloc/free.
3605  */
3606 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3607 					   struct kobj_attribute *attr,
3608 					   char *buf)
3609 {
3610 	return nr_hugepages_show_common(kobj, attr, buf);
3611 }
3612 
3613 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3614 	       struct kobj_attribute *attr, const char *buf, size_t len)
3615 {
3616 	return nr_hugepages_store_common(true, kobj, buf, len);
3617 }
3618 HSTATE_ATTR(nr_hugepages_mempolicy);
3619 #endif
3620 
3621 
3622 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3623 					struct kobj_attribute *attr, char *buf)
3624 {
3625 	struct hstate *h = kobj_to_hstate(kobj, NULL);
3626 	return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3627 }
3628 
3629 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3630 		struct kobj_attribute *attr, const char *buf, size_t count)
3631 {
3632 	int err;
3633 	unsigned long input;
3634 	struct hstate *h = kobj_to_hstate(kobj, NULL);
3635 
3636 	if (hstate_is_gigantic(h))
3637 		return -EINVAL;
3638 
3639 	err = kstrtoul(buf, 10, &input);
3640 	if (err)
3641 		return err;
3642 
3643 	spin_lock_irq(&hugetlb_lock);
3644 	h->nr_overcommit_huge_pages = input;
3645 	spin_unlock_irq(&hugetlb_lock);
3646 
3647 	return count;
3648 }
3649 HSTATE_ATTR(nr_overcommit_hugepages);
3650 
3651 static ssize_t free_hugepages_show(struct kobject *kobj,
3652 					struct kobj_attribute *attr, char *buf)
3653 {
3654 	struct hstate *h;
3655 	unsigned long free_huge_pages;
3656 	int nid;
3657 
3658 	h = kobj_to_hstate(kobj, &nid);
3659 	if (nid == NUMA_NO_NODE)
3660 		free_huge_pages = h->free_huge_pages;
3661 	else
3662 		free_huge_pages = h->free_huge_pages_node[nid];
3663 
3664 	return sysfs_emit(buf, "%lu\n", free_huge_pages);
3665 }
3666 HSTATE_ATTR_RO(free_hugepages);
3667 
3668 static ssize_t resv_hugepages_show(struct kobject *kobj,
3669 					struct kobj_attribute *attr, char *buf)
3670 {
3671 	struct hstate *h = kobj_to_hstate(kobj, NULL);
3672 	return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3673 }
3674 HSTATE_ATTR_RO(resv_hugepages);
3675 
3676 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3677 					struct kobj_attribute *attr, char *buf)
3678 {
3679 	struct hstate *h;
3680 	unsigned long surplus_huge_pages;
3681 	int nid;
3682 
3683 	h = kobj_to_hstate(kobj, &nid);
3684 	if (nid == NUMA_NO_NODE)
3685 		surplus_huge_pages = h->surplus_huge_pages;
3686 	else
3687 		surplus_huge_pages = h->surplus_huge_pages_node[nid];
3688 
3689 	return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3690 }
3691 HSTATE_ATTR_RO(surplus_hugepages);
3692 
3693 static ssize_t demote_store(struct kobject *kobj,
3694 	       struct kobj_attribute *attr, const char *buf, size_t len)
3695 {
3696 	unsigned long nr_demote;
3697 	unsigned long nr_available;
3698 	nodemask_t nodes_allowed, *n_mask;
3699 	struct hstate *h;
3700 	int err = 0;
3701 	int nid;
3702 
3703 	err = kstrtoul(buf, 10, &nr_demote);
3704 	if (err)
3705 		return err;
3706 	h = kobj_to_hstate(kobj, &nid);
3707 
3708 	if (nid != NUMA_NO_NODE) {
3709 		init_nodemask_of_node(&nodes_allowed, nid);
3710 		n_mask = &nodes_allowed;
3711 	} else {
3712 		n_mask = &node_states[N_MEMORY];
3713 	}
3714 
3715 	/* Synchronize with other sysfs operations modifying huge pages */
3716 	mutex_lock(&h->resize_lock);
3717 	spin_lock_irq(&hugetlb_lock);
3718 
3719 	while (nr_demote) {
3720 		/*
3721 		 * Check for available pages to demote each time thorough the
3722 		 * loop as demote_pool_huge_page will drop hugetlb_lock.
3723 		 */
3724 		if (nid != NUMA_NO_NODE)
3725 			nr_available = h->free_huge_pages_node[nid];
3726 		else
3727 			nr_available = h->free_huge_pages;
3728 		nr_available -= h->resv_huge_pages;
3729 		if (!nr_available)
3730 			break;
3731 
3732 		err = demote_pool_huge_page(h, n_mask);
3733 		if (err)
3734 			break;
3735 
3736 		nr_demote--;
3737 	}
3738 
3739 	spin_unlock_irq(&hugetlb_lock);
3740 	mutex_unlock(&h->resize_lock);
3741 
3742 	if (err)
3743 		return err;
3744 	return len;
3745 }
3746 HSTATE_ATTR_WO(demote);
3747 
3748 static ssize_t demote_size_show(struct kobject *kobj,
3749 					struct kobj_attribute *attr, char *buf)
3750 {
3751 	int nid;
3752 	struct hstate *h = kobj_to_hstate(kobj, &nid);
3753 	unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K;
3754 
3755 	return sysfs_emit(buf, "%lukB\n", demote_size);
3756 }
3757 
3758 static ssize_t demote_size_store(struct kobject *kobj,
3759 					struct kobj_attribute *attr,
3760 					const char *buf, size_t count)
3761 {
3762 	struct hstate *h, *demote_hstate;
3763 	unsigned long demote_size;
3764 	unsigned int demote_order;
3765 	int nid;
3766 
3767 	demote_size = (unsigned long)memparse(buf, NULL);
3768 
3769 	demote_hstate = size_to_hstate(demote_size);
3770 	if (!demote_hstate)
3771 		return -EINVAL;
3772 	demote_order = demote_hstate->order;
3773 	if (demote_order < HUGETLB_PAGE_ORDER)
3774 		return -EINVAL;
3775 
3776 	/* demote order must be smaller than hstate order */
3777 	h = kobj_to_hstate(kobj, &nid);
3778 	if (demote_order >= h->order)
3779 		return -EINVAL;
3780 
3781 	/* resize_lock synchronizes access to demote size and writes */
3782 	mutex_lock(&h->resize_lock);
3783 	h->demote_order = demote_order;
3784 	mutex_unlock(&h->resize_lock);
3785 
3786 	return count;
3787 }
3788 HSTATE_ATTR(demote_size);
3789 
3790 static struct attribute *hstate_attrs[] = {
3791 	&nr_hugepages_attr.attr,
3792 	&nr_overcommit_hugepages_attr.attr,
3793 	&free_hugepages_attr.attr,
3794 	&resv_hugepages_attr.attr,
3795 	&surplus_hugepages_attr.attr,
3796 #ifdef CONFIG_NUMA
3797 	&nr_hugepages_mempolicy_attr.attr,
3798 #endif
3799 	NULL,
3800 };
3801 
3802 static const struct attribute_group hstate_attr_group = {
3803 	.attrs = hstate_attrs,
3804 };
3805 
3806 static struct attribute *hstate_demote_attrs[] = {
3807 	&demote_size_attr.attr,
3808 	&demote_attr.attr,
3809 	NULL,
3810 };
3811 
3812 static const struct attribute_group hstate_demote_attr_group = {
3813 	.attrs = hstate_demote_attrs,
3814 };
3815 
3816 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3817 				    struct kobject **hstate_kobjs,
3818 				    const struct attribute_group *hstate_attr_group)
3819 {
3820 	int retval;
3821 	int hi = hstate_index(h);
3822 
3823 	hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3824 	if (!hstate_kobjs[hi])
3825 		return -ENOMEM;
3826 
3827 	retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3828 	if (retval) {
3829 		kobject_put(hstate_kobjs[hi]);
3830 		hstate_kobjs[hi] = NULL;
3831 	}
3832 
3833 	if (h->demote_order) {
3834 		if (sysfs_create_group(hstate_kobjs[hi],
3835 					&hstate_demote_attr_group))
3836 			pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name);
3837 	}
3838 
3839 	return retval;
3840 }
3841 
3842 static void __init hugetlb_sysfs_init(void)
3843 {
3844 	struct hstate *h;
3845 	int err;
3846 
3847 	hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3848 	if (!hugepages_kobj)
3849 		return;
3850 
3851 	for_each_hstate(h) {
3852 		err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3853 					 hstate_kobjs, &hstate_attr_group);
3854 		if (err)
3855 			pr_err("HugeTLB: Unable to add hstate %s", h->name);
3856 	}
3857 }
3858 
3859 #ifdef CONFIG_NUMA
3860 
3861 /*
3862  * node_hstate/s - associate per node hstate attributes, via their kobjects,
3863  * with node devices in node_devices[] using a parallel array.  The array
3864  * index of a node device or _hstate == node id.
3865  * This is here to avoid any static dependency of the node device driver, in
3866  * the base kernel, on the hugetlb module.
3867  */
3868 struct node_hstate {
3869 	struct kobject		*hugepages_kobj;
3870 	struct kobject		*hstate_kobjs[HUGE_MAX_HSTATE];
3871 };
3872 static struct node_hstate node_hstates[MAX_NUMNODES];
3873 
3874 /*
3875  * A subset of global hstate attributes for node devices
3876  */
3877 static struct attribute *per_node_hstate_attrs[] = {
3878 	&nr_hugepages_attr.attr,
3879 	&free_hugepages_attr.attr,
3880 	&surplus_hugepages_attr.attr,
3881 	NULL,
3882 };
3883 
3884 static const struct attribute_group per_node_hstate_attr_group = {
3885 	.attrs = per_node_hstate_attrs,
3886 };
3887 
3888 /*
3889  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3890  * Returns node id via non-NULL nidp.
3891  */
3892 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3893 {
3894 	int nid;
3895 
3896 	for (nid = 0; nid < nr_node_ids; nid++) {
3897 		struct node_hstate *nhs = &node_hstates[nid];
3898 		int i;
3899 		for (i = 0; i < HUGE_MAX_HSTATE; i++)
3900 			if (nhs->hstate_kobjs[i] == kobj) {
3901 				if (nidp)
3902 					*nidp = nid;
3903 				return &hstates[i];
3904 			}
3905 	}
3906 
3907 	BUG();
3908 	return NULL;
3909 }
3910 
3911 /*
3912  * Unregister hstate attributes from a single node device.
3913  * No-op if no hstate attributes attached.
3914  */
3915 static void hugetlb_unregister_node(struct node *node)
3916 {
3917 	struct hstate *h;
3918 	struct node_hstate *nhs = &node_hstates[node->dev.id];
3919 
3920 	if (!nhs->hugepages_kobj)
3921 		return;		/* no hstate attributes */
3922 
3923 	for_each_hstate(h) {
3924 		int idx = hstate_index(h);
3925 		if (nhs->hstate_kobjs[idx]) {
3926 			kobject_put(nhs->hstate_kobjs[idx]);
3927 			nhs->hstate_kobjs[idx] = NULL;
3928 		}
3929 	}
3930 
3931 	kobject_put(nhs->hugepages_kobj);
3932 	nhs->hugepages_kobj = NULL;
3933 }
3934 
3935 
3936 /*
3937  * Register hstate attributes for a single node device.
3938  * No-op if attributes already registered.
3939  */
3940 static void hugetlb_register_node(struct node *node)
3941 {
3942 	struct hstate *h;
3943 	struct node_hstate *nhs = &node_hstates[node->dev.id];
3944 	int err;
3945 
3946 	if (nhs->hugepages_kobj)
3947 		return;		/* already allocated */
3948 
3949 	nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3950 							&node->dev.kobj);
3951 	if (!nhs->hugepages_kobj)
3952 		return;
3953 
3954 	for_each_hstate(h) {
3955 		err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3956 						nhs->hstate_kobjs,
3957 						&per_node_hstate_attr_group);
3958 		if (err) {
3959 			pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3960 				h->name, node->dev.id);
3961 			hugetlb_unregister_node(node);
3962 			break;
3963 		}
3964 	}
3965 }
3966 
3967 /*
3968  * hugetlb init time:  register hstate attributes for all registered node
3969  * devices of nodes that have memory.  All on-line nodes should have
3970  * registered their associated device by this time.
3971  */
3972 static void __init hugetlb_register_all_nodes(void)
3973 {
3974 	int nid;
3975 
3976 	for_each_node_state(nid, N_MEMORY) {
3977 		struct node *node = node_devices[nid];
3978 		if (node->dev.id == nid)
3979 			hugetlb_register_node(node);
3980 	}
3981 
3982 	/*
3983 	 * Let the node device driver know we're here so it can
3984 	 * [un]register hstate attributes on node hotplug.
3985 	 */
3986 	register_hugetlbfs_with_node(hugetlb_register_node,
3987 				     hugetlb_unregister_node);
3988 }
3989 #else	/* !CONFIG_NUMA */
3990 
3991 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3992 {
3993 	BUG();
3994 	if (nidp)
3995 		*nidp = -1;
3996 	return NULL;
3997 }
3998 
3999 static void hugetlb_register_all_nodes(void) { }
4000 
4001 #endif
4002 
4003 static int __init hugetlb_init(void)
4004 {
4005 	int i;
4006 
4007 	BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
4008 			__NR_HPAGEFLAGS);
4009 
4010 	if (!hugepages_supported()) {
4011 		if (hugetlb_max_hstate || default_hstate_max_huge_pages)
4012 			pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
4013 		return 0;
4014 	}
4015 
4016 	/*
4017 	 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists.  Some
4018 	 * architectures depend on setup being done here.
4019 	 */
4020 	hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
4021 	if (!parsed_default_hugepagesz) {
4022 		/*
4023 		 * If we did not parse a default huge page size, set
4024 		 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
4025 		 * number of huge pages for this default size was implicitly
4026 		 * specified, set that here as well.
4027 		 * Note that the implicit setting will overwrite an explicit
4028 		 * setting.  A warning will be printed in this case.
4029 		 */
4030 		default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
4031 		if (default_hstate_max_huge_pages) {
4032 			if (default_hstate.max_huge_pages) {
4033 				char buf[32];
4034 
4035 				string_get_size(huge_page_size(&default_hstate),
4036 					1, STRING_UNITS_2, buf, 32);
4037 				pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
4038 					default_hstate.max_huge_pages, buf);
4039 				pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
4040 					default_hstate_max_huge_pages);
4041 			}
4042 			default_hstate.max_huge_pages =
4043 				default_hstate_max_huge_pages;
4044 
4045 			for (i = 0; i < nr_online_nodes; i++)
4046 				default_hstate.max_huge_pages_node[i] =
4047 					default_hugepages_in_node[i];
4048 		}
4049 	}
4050 
4051 	hugetlb_cma_check();
4052 	hugetlb_init_hstates();
4053 	gather_bootmem_prealloc();
4054 	report_hugepages();
4055 
4056 	hugetlb_sysfs_init();
4057 	hugetlb_register_all_nodes();
4058 	hugetlb_cgroup_file_init();
4059 
4060 #ifdef CONFIG_SMP
4061 	num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
4062 #else
4063 	num_fault_mutexes = 1;
4064 #endif
4065 	hugetlb_fault_mutex_table =
4066 		kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
4067 			      GFP_KERNEL);
4068 	BUG_ON(!hugetlb_fault_mutex_table);
4069 
4070 	for (i = 0; i < num_fault_mutexes; i++)
4071 		mutex_init(&hugetlb_fault_mutex_table[i]);
4072 	return 0;
4073 }
4074 subsys_initcall(hugetlb_init);
4075 
4076 /* Overwritten by architectures with more huge page sizes */
4077 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
4078 {
4079 	return size == HPAGE_SIZE;
4080 }
4081 
4082 void __init hugetlb_add_hstate(unsigned int order)
4083 {
4084 	struct hstate *h;
4085 	unsigned long i;
4086 
4087 	if (size_to_hstate(PAGE_SIZE << order)) {
4088 		return;
4089 	}
4090 	BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
4091 	BUG_ON(order == 0);
4092 	h = &hstates[hugetlb_max_hstate++];
4093 	mutex_init(&h->resize_lock);
4094 	h->order = order;
4095 	h->mask = ~(huge_page_size(h) - 1);
4096 	for (i = 0; i < MAX_NUMNODES; ++i)
4097 		INIT_LIST_HEAD(&h->hugepage_freelists[i]);
4098 	INIT_LIST_HEAD(&h->hugepage_activelist);
4099 	h->next_nid_to_alloc = first_memory_node;
4100 	h->next_nid_to_free = first_memory_node;
4101 	snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
4102 					huge_page_size(h)/1024);
4103 	hugetlb_vmemmap_init(h);
4104 
4105 	parsed_hstate = h;
4106 }
4107 
4108 bool __init __weak hugetlb_node_alloc_supported(void)
4109 {
4110 	return true;
4111 }
4112 /*
4113  * hugepages command line processing
4114  * hugepages normally follows a valid hugepagsz or default_hugepagsz
4115  * specification.  If not, ignore the hugepages value.  hugepages can also
4116  * be the first huge page command line  option in which case it implicitly
4117  * specifies the number of huge pages for the default size.
4118  */
4119 static int __init hugepages_setup(char *s)
4120 {
4121 	unsigned long *mhp;
4122 	static unsigned long *last_mhp;
4123 	int node = NUMA_NO_NODE;
4124 	int count;
4125 	unsigned long tmp;
4126 	char *p = s;
4127 
4128 	if (!parsed_valid_hugepagesz) {
4129 		pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
4130 		parsed_valid_hugepagesz = true;
4131 		return 0;
4132 	}
4133 
4134 	/*
4135 	 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
4136 	 * yet, so this hugepages= parameter goes to the "default hstate".
4137 	 * Otherwise, it goes with the previously parsed hugepagesz or
4138 	 * default_hugepagesz.
4139 	 */
4140 	else if (!hugetlb_max_hstate)
4141 		mhp = &default_hstate_max_huge_pages;
4142 	else
4143 		mhp = &parsed_hstate->max_huge_pages;
4144 
4145 	if (mhp == last_mhp) {
4146 		pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
4147 		return 0;
4148 	}
4149 
4150 	while (*p) {
4151 		count = 0;
4152 		if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4153 			goto invalid;
4154 		/* Parameter is node format */
4155 		if (p[count] == ':') {
4156 			if (!hugetlb_node_alloc_supported()) {
4157 				pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n");
4158 				return 0;
4159 			}
4160 			node = tmp;
4161 			p += count + 1;
4162 			if (node < 0 || node >= nr_online_nodes)
4163 				goto invalid;
4164 			/* Parse hugepages */
4165 			if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4166 				goto invalid;
4167 			if (!hugetlb_max_hstate)
4168 				default_hugepages_in_node[node] = tmp;
4169 			else
4170 				parsed_hstate->max_huge_pages_node[node] = tmp;
4171 			*mhp += tmp;
4172 			/* Go to parse next node*/
4173 			if (p[count] == ',')
4174 				p += count + 1;
4175 			else
4176 				break;
4177 		} else {
4178 			if (p != s)
4179 				goto invalid;
4180 			*mhp = tmp;
4181 			break;
4182 		}
4183 	}
4184 
4185 	/*
4186 	 * Global state is always initialized later in hugetlb_init.
4187 	 * But we need to allocate gigantic hstates here early to still
4188 	 * use the bootmem allocator.
4189 	 */
4190 	if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
4191 		hugetlb_hstate_alloc_pages(parsed_hstate);
4192 
4193 	last_mhp = mhp;
4194 
4195 	return 1;
4196 
4197 invalid:
4198 	pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p);
4199 	return 0;
4200 }
4201 __setup("hugepages=", hugepages_setup);
4202 
4203 /*
4204  * hugepagesz command line processing
4205  * A specific huge page size can only be specified once with hugepagesz.
4206  * hugepagesz is followed by hugepages on the command line.  The global
4207  * variable 'parsed_valid_hugepagesz' is used to determine if prior
4208  * hugepagesz argument was valid.
4209  */
4210 static int __init hugepagesz_setup(char *s)
4211 {
4212 	unsigned long size;
4213 	struct hstate *h;
4214 
4215 	parsed_valid_hugepagesz = false;
4216 	size = (unsigned long)memparse(s, NULL);
4217 
4218 	if (!arch_hugetlb_valid_size(size)) {
4219 		pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
4220 		return 0;
4221 	}
4222 
4223 	h = size_to_hstate(size);
4224 	if (h) {
4225 		/*
4226 		 * hstate for this size already exists.  This is normally
4227 		 * an error, but is allowed if the existing hstate is the
4228 		 * default hstate.  More specifically, it is only allowed if
4229 		 * the number of huge pages for the default hstate was not
4230 		 * previously specified.
4231 		 */
4232 		if (!parsed_default_hugepagesz ||  h != &default_hstate ||
4233 		    default_hstate.max_huge_pages) {
4234 			pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
4235 			return 0;
4236 		}
4237 
4238 		/*
4239 		 * No need to call hugetlb_add_hstate() as hstate already
4240 		 * exists.  But, do set parsed_hstate so that a following
4241 		 * hugepages= parameter will be applied to this hstate.
4242 		 */
4243 		parsed_hstate = h;
4244 		parsed_valid_hugepagesz = true;
4245 		return 1;
4246 	}
4247 
4248 	hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4249 	parsed_valid_hugepagesz = true;
4250 	return 1;
4251 }
4252 __setup("hugepagesz=", hugepagesz_setup);
4253 
4254 /*
4255  * default_hugepagesz command line input
4256  * Only one instance of default_hugepagesz allowed on command line.
4257  */
4258 static int __init default_hugepagesz_setup(char *s)
4259 {
4260 	unsigned long size;
4261 	int i;
4262 
4263 	parsed_valid_hugepagesz = false;
4264 	if (parsed_default_hugepagesz) {
4265 		pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
4266 		return 0;
4267 	}
4268 
4269 	size = (unsigned long)memparse(s, NULL);
4270 
4271 	if (!arch_hugetlb_valid_size(size)) {
4272 		pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
4273 		return 0;
4274 	}
4275 
4276 	hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4277 	parsed_valid_hugepagesz = true;
4278 	parsed_default_hugepagesz = true;
4279 	default_hstate_idx = hstate_index(size_to_hstate(size));
4280 
4281 	/*
4282 	 * The number of default huge pages (for this size) could have been
4283 	 * specified as the first hugetlb parameter: hugepages=X.  If so,
4284 	 * then default_hstate_max_huge_pages is set.  If the default huge
4285 	 * page size is gigantic (>= MAX_ORDER), then the pages must be
4286 	 * allocated here from bootmem allocator.
4287 	 */
4288 	if (default_hstate_max_huge_pages) {
4289 		default_hstate.max_huge_pages = default_hstate_max_huge_pages;
4290 		for (i = 0; i < nr_online_nodes; i++)
4291 			default_hstate.max_huge_pages_node[i] =
4292 				default_hugepages_in_node[i];
4293 		if (hstate_is_gigantic(&default_hstate))
4294 			hugetlb_hstate_alloc_pages(&default_hstate);
4295 		default_hstate_max_huge_pages = 0;
4296 	}
4297 
4298 	return 1;
4299 }
4300 __setup("default_hugepagesz=", default_hugepagesz_setup);
4301 
4302 static unsigned int allowed_mems_nr(struct hstate *h)
4303 {
4304 	int node;
4305 	unsigned int nr = 0;
4306 	nodemask_t *mpol_allowed;
4307 	unsigned int *array = h->free_huge_pages_node;
4308 	gfp_t gfp_mask = htlb_alloc_mask(h);
4309 
4310 	mpol_allowed = policy_nodemask_current(gfp_mask);
4311 
4312 	for_each_node_mask(node, cpuset_current_mems_allowed) {
4313 		if (!mpol_allowed || node_isset(node, *mpol_allowed))
4314 			nr += array[node];
4315 	}
4316 
4317 	return nr;
4318 }
4319 
4320 #ifdef CONFIG_SYSCTL
4321 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
4322 					  void *buffer, size_t *length,
4323 					  loff_t *ppos, unsigned long *out)
4324 {
4325 	struct ctl_table dup_table;
4326 
4327 	/*
4328 	 * In order to avoid races with __do_proc_doulongvec_minmax(), we
4329 	 * can duplicate the @table and alter the duplicate of it.
4330 	 */
4331 	dup_table = *table;
4332 	dup_table.data = out;
4333 
4334 	return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
4335 }
4336 
4337 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
4338 			 struct ctl_table *table, int write,
4339 			 void *buffer, size_t *length, loff_t *ppos)
4340 {
4341 	struct hstate *h = &default_hstate;
4342 	unsigned long tmp = h->max_huge_pages;
4343 	int ret;
4344 
4345 	if (!hugepages_supported())
4346 		return -EOPNOTSUPP;
4347 
4348 	ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4349 					     &tmp);
4350 	if (ret)
4351 		goto out;
4352 
4353 	if (write)
4354 		ret = __nr_hugepages_store_common(obey_mempolicy, h,
4355 						  NUMA_NO_NODE, tmp, *length);
4356 out:
4357 	return ret;
4358 }
4359 
4360 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
4361 			  void *buffer, size_t *length, loff_t *ppos)
4362 {
4363 
4364 	return hugetlb_sysctl_handler_common(false, table, write,
4365 							buffer, length, ppos);
4366 }
4367 
4368 #ifdef CONFIG_NUMA
4369 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
4370 			  void *buffer, size_t *length, loff_t *ppos)
4371 {
4372 	return hugetlb_sysctl_handler_common(true, table, write,
4373 							buffer, length, ppos);
4374 }
4375 #endif /* CONFIG_NUMA */
4376 
4377 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
4378 		void *buffer, size_t *length, loff_t *ppos)
4379 {
4380 	struct hstate *h = &default_hstate;
4381 	unsigned long tmp;
4382 	int ret;
4383 
4384 	if (!hugepages_supported())
4385 		return -EOPNOTSUPP;
4386 
4387 	tmp = h->nr_overcommit_huge_pages;
4388 
4389 	if (write && hstate_is_gigantic(h))
4390 		return -EINVAL;
4391 
4392 	ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4393 					     &tmp);
4394 	if (ret)
4395 		goto out;
4396 
4397 	if (write) {
4398 		spin_lock_irq(&hugetlb_lock);
4399 		h->nr_overcommit_huge_pages = tmp;
4400 		spin_unlock_irq(&hugetlb_lock);
4401 	}
4402 out:
4403 	return ret;
4404 }
4405 
4406 #endif /* CONFIG_SYSCTL */
4407 
4408 void hugetlb_report_meminfo(struct seq_file *m)
4409 {
4410 	struct hstate *h;
4411 	unsigned long total = 0;
4412 
4413 	if (!hugepages_supported())
4414 		return;
4415 
4416 	for_each_hstate(h) {
4417 		unsigned long count = h->nr_huge_pages;
4418 
4419 		total += huge_page_size(h) * count;
4420 
4421 		if (h == &default_hstate)
4422 			seq_printf(m,
4423 				   "HugePages_Total:   %5lu\n"
4424 				   "HugePages_Free:    %5lu\n"
4425 				   "HugePages_Rsvd:    %5lu\n"
4426 				   "HugePages_Surp:    %5lu\n"
4427 				   "Hugepagesize:   %8lu kB\n",
4428 				   count,
4429 				   h->free_huge_pages,
4430 				   h->resv_huge_pages,
4431 				   h->surplus_huge_pages,
4432 				   huge_page_size(h) / SZ_1K);
4433 	}
4434 
4435 	seq_printf(m, "Hugetlb:        %8lu kB\n", total / SZ_1K);
4436 }
4437 
4438 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
4439 {
4440 	struct hstate *h = &default_hstate;
4441 
4442 	if (!hugepages_supported())
4443 		return 0;
4444 
4445 	return sysfs_emit_at(buf, len,
4446 			     "Node %d HugePages_Total: %5u\n"
4447 			     "Node %d HugePages_Free:  %5u\n"
4448 			     "Node %d HugePages_Surp:  %5u\n",
4449 			     nid, h->nr_huge_pages_node[nid],
4450 			     nid, h->free_huge_pages_node[nid],
4451 			     nid, h->surplus_huge_pages_node[nid]);
4452 }
4453 
4454 void hugetlb_show_meminfo(void)
4455 {
4456 	struct hstate *h;
4457 	int nid;
4458 
4459 	if (!hugepages_supported())
4460 		return;
4461 
4462 	for_each_node_state(nid, N_MEMORY)
4463 		for_each_hstate(h)
4464 			pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
4465 				nid,
4466 				h->nr_huge_pages_node[nid],
4467 				h->free_huge_pages_node[nid],
4468 				h->surplus_huge_pages_node[nid],
4469 				huge_page_size(h) / SZ_1K);
4470 }
4471 
4472 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
4473 {
4474 	seq_printf(m, "HugetlbPages:\t%8lu kB\n",
4475 		   atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
4476 }
4477 
4478 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
4479 unsigned long hugetlb_total_pages(void)
4480 {
4481 	struct hstate *h;
4482 	unsigned long nr_total_pages = 0;
4483 
4484 	for_each_hstate(h)
4485 		nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
4486 	return nr_total_pages;
4487 }
4488 
4489 static int hugetlb_acct_memory(struct hstate *h, long delta)
4490 {
4491 	int ret = -ENOMEM;
4492 
4493 	if (!delta)
4494 		return 0;
4495 
4496 	spin_lock_irq(&hugetlb_lock);
4497 	/*
4498 	 * When cpuset is configured, it breaks the strict hugetlb page
4499 	 * reservation as the accounting is done on a global variable. Such
4500 	 * reservation is completely rubbish in the presence of cpuset because
4501 	 * the reservation is not checked against page availability for the
4502 	 * current cpuset. Application can still potentially OOM'ed by kernel
4503 	 * with lack of free htlb page in cpuset that the task is in.
4504 	 * Attempt to enforce strict accounting with cpuset is almost
4505 	 * impossible (or too ugly) because cpuset is too fluid that
4506 	 * task or memory node can be dynamically moved between cpusets.
4507 	 *
4508 	 * The change of semantics for shared hugetlb mapping with cpuset is
4509 	 * undesirable. However, in order to preserve some of the semantics,
4510 	 * we fall back to check against current free page availability as
4511 	 * a best attempt and hopefully to minimize the impact of changing
4512 	 * semantics that cpuset has.
4513 	 *
4514 	 * Apart from cpuset, we also have memory policy mechanism that
4515 	 * also determines from which node the kernel will allocate memory
4516 	 * in a NUMA system. So similar to cpuset, we also should consider
4517 	 * the memory policy of the current task. Similar to the description
4518 	 * above.
4519 	 */
4520 	if (delta > 0) {
4521 		if (gather_surplus_pages(h, delta) < 0)
4522 			goto out;
4523 
4524 		if (delta > allowed_mems_nr(h)) {
4525 			return_unused_surplus_pages(h, delta);
4526 			goto out;
4527 		}
4528 	}
4529 
4530 	ret = 0;
4531 	if (delta < 0)
4532 		return_unused_surplus_pages(h, (unsigned long) -delta);
4533 
4534 out:
4535 	spin_unlock_irq(&hugetlb_lock);
4536 	return ret;
4537 }
4538 
4539 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4540 {
4541 	struct resv_map *resv = vma_resv_map(vma);
4542 
4543 	/*
4544 	 * This new VMA should share its siblings reservation map if present.
4545 	 * The VMA will only ever have a valid reservation map pointer where
4546 	 * it is being copied for another still existing VMA.  As that VMA
4547 	 * has a reference to the reservation map it cannot disappear until
4548 	 * after this open call completes.  It is therefore safe to take a
4549 	 * new reference here without additional locking.
4550 	 */
4551 	if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
4552 		resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
4553 		kref_get(&resv->refs);
4554 	}
4555 }
4556 
4557 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4558 {
4559 	struct hstate *h = hstate_vma(vma);
4560 	struct resv_map *resv = vma_resv_map(vma);
4561 	struct hugepage_subpool *spool = subpool_vma(vma);
4562 	unsigned long reserve, start, end;
4563 	long gbl_reserve;
4564 
4565 	if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4566 		return;
4567 
4568 	start = vma_hugecache_offset(h, vma, vma->vm_start);
4569 	end = vma_hugecache_offset(h, vma, vma->vm_end);
4570 
4571 	reserve = (end - start) - region_count(resv, start, end);
4572 	hugetlb_cgroup_uncharge_counter(resv, start, end);
4573 	if (reserve) {
4574 		/*
4575 		 * Decrement reserve counts.  The global reserve count may be
4576 		 * adjusted if the subpool has a minimum size.
4577 		 */
4578 		gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4579 		hugetlb_acct_memory(h, -gbl_reserve);
4580 	}
4581 
4582 	kref_put(&resv->refs, resv_map_release);
4583 }
4584 
4585 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4586 {
4587 	if (addr & ~(huge_page_mask(hstate_vma(vma))))
4588 		return -EINVAL;
4589 	return 0;
4590 }
4591 
4592 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4593 {
4594 	return huge_page_size(hstate_vma(vma));
4595 }
4596 
4597 /*
4598  * We cannot handle pagefaults against hugetlb pages at all.  They cause
4599  * handle_mm_fault() to try to instantiate regular-sized pages in the
4600  * hugepage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
4601  * this far.
4602  */
4603 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4604 {
4605 	BUG();
4606 	return 0;
4607 }
4608 
4609 /*
4610  * When a new function is introduced to vm_operations_struct and added
4611  * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4612  * This is because under System V memory model, mappings created via
4613  * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4614  * their original vm_ops are overwritten with shm_vm_ops.
4615  */
4616 const struct vm_operations_struct hugetlb_vm_ops = {
4617 	.fault = hugetlb_vm_op_fault,
4618 	.open = hugetlb_vm_op_open,
4619 	.close = hugetlb_vm_op_close,
4620 	.may_split = hugetlb_vm_op_split,
4621 	.pagesize = hugetlb_vm_op_pagesize,
4622 };
4623 
4624 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4625 				int writable)
4626 {
4627 	pte_t entry;
4628 	unsigned int shift = huge_page_shift(hstate_vma(vma));
4629 
4630 	if (writable) {
4631 		entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4632 					 vma->vm_page_prot)));
4633 	} else {
4634 		entry = huge_pte_wrprotect(mk_huge_pte(page,
4635 					   vma->vm_page_prot));
4636 	}
4637 	entry = pte_mkyoung(entry);
4638 	entry = pte_mkhuge(entry);
4639 	entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4640 
4641 	return entry;
4642 }
4643 
4644 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4645 				   unsigned long address, pte_t *ptep)
4646 {
4647 	pte_t entry;
4648 
4649 	entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4650 	if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4651 		update_mmu_cache(vma, address, ptep);
4652 }
4653 
4654 bool is_hugetlb_entry_migration(pte_t pte)
4655 {
4656 	swp_entry_t swp;
4657 
4658 	if (huge_pte_none(pte) || pte_present(pte))
4659 		return false;
4660 	swp = pte_to_swp_entry(pte);
4661 	if (is_migration_entry(swp))
4662 		return true;
4663 	else
4664 		return false;
4665 }
4666 
4667 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4668 {
4669 	swp_entry_t swp;
4670 
4671 	if (huge_pte_none(pte) || pte_present(pte))
4672 		return false;
4673 	swp = pte_to_swp_entry(pte);
4674 	if (is_hwpoison_entry(swp))
4675 		return true;
4676 	else
4677 		return false;
4678 }
4679 
4680 static void
4681 hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
4682 		     struct page *new_page)
4683 {
4684 	__SetPageUptodate(new_page);
4685 	set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
4686 	hugepage_add_new_anon_rmap(new_page, vma, addr);
4687 	hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
4688 	ClearHPageRestoreReserve(new_page);
4689 	SetHPageMigratable(new_page);
4690 }
4691 
4692 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
4693 			    struct vm_area_struct *vma)
4694 {
4695 	pte_t *src_pte, *dst_pte, entry, dst_entry;
4696 	struct page *ptepage;
4697 	unsigned long addr;
4698 	bool cow = is_cow_mapping(vma->vm_flags);
4699 	struct hstate *h = hstate_vma(vma);
4700 	unsigned long sz = huge_page_size(h);
4701 	unsigned long npages = pages_per_huge_page(h);
4702 	struct address_space *mapping = vma->vm_file->f_mapping;
4703 	struct mmu_notifier_range range;
4704 	int ret = 0;
4705 
4706 	if (cow) {
4707 		mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
4708 					vma->vm_start,
4709 					vma->vm_end);
4710 		mmu_notifier_invalidate_range_start(&range);
4711 	} else {
4712 		/*
4713 		 * For shared mappings i_mmap_rwsem must be held to call
4714 		 * huge_pte_alloc, otherwise the returned ptep could go
4715 		 * away if part of a shared pmd and another thread calls
4716 		 * huge_pmd_unshare.
4717 		 */
4718 		i_mmap_lock_read(mapping);
4719 	}
4720 
4721 	for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
4722 		spinlock_t *src_ptl, *dst_ptl;
4723 		src_pte = huge_pte_offset(src, addr, sz);
4724 		if (!src_pte)
4725 			continue;
4726 		dst_pte = huge_pte_alloc(dst, vma, addr, sz);
4727 		if (!dst_pte) {
4728 			ret = -ENOMEM;
4729 			break;
4730 		}
4731 
4732 		/*
4733 		 * If the pagetables are shared don't copy or take references.
4734 		 * dst_pte == src_pte is the common case of src/dest sharing.
4735 		 *
4736 		 * However, src could have 'unshared' and dst shares with
4737 		 * another vma.  If dst_pte !none, this implies sharing.
4738 		 * Check here before taking page table lock, and once again
4739 		 * after taking the lock below.
4740 		 */
4741 		dst_entry = huge_ptep_get(dst_pte);
4742 		if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
4743 			continue;
4744 
4745 		dst_ptl = huge_pte_lock(h, dst, dst_pte);
4746 		src_ptl = huge_pte_lockptr(h, src, src_pte);
4747 		spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4748 		entry = huge_ptep_get(src_pte);
4749 		dst_entry = huge_ptep_get(dst_pte);
4750 again:
4751 		if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
4752 			/*
4753 			 * Skip if src entry none.  Also, skip in the
4754 			 * unlikely case dst entry !none as this implies
4755 			 * sharing with another vma.
4756 			 */
4757 			;
4758 		} else if (unlikely(is_hugetlb_entry_migration(entry) ||
4759 				    is_hugetlb_entry_hwpoisoned(entry))) {
4760 			swp_entry_t swp_entry = pte_to_swp_entry(entry);
4761 
4762 			if (is_writable_migration_entry(swp_entry) && cow) {
4763 				/*
4764 				 * COW mappings require pages in both
4765 				 * parent and child to be set to read.
4766 				 */
4767 				swp_entry = make_readable_migration_entry(
4768 							swp_offset(swp_entry));
4769 				entry = swp_entry_to_pte(swp_entry);
4770 				set_huge_swap_pte_at(src, addr, src_pte,
4771 						     entry, sz);
4772 			}
4773 			set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
4774 		} else {
4775 			entry = huge_ptep_get(src_pte);
4776 			ptepage = pte_page(entry);
4777 			get_page(ptepage);
4778 
4779 			/*
4780 			 * This is a rare case where we see pinned hugetlb
4781 			 * pages while they're prone to COW.  We need to do the
4782 			 * COW earlier during fork.
4783 			 *
4784 			 * When pre-allocating the page or copying data, we
4785 			 * need to be without the pgtable locks since we could
4786 			 * sleep during the process.
4787 			 */
4788 			if (unlikely(page_needs_cow_for_dma(vma, ptepage))) {
4789 				pte_t src_pte_old = entry;
4790 				struct page *new;
4791 
4792 				spin_unlock(src_ptl);
4793 				spin_unlock(dst_ptl);
4794 				/* Do not use reserve as it's private owned */
4795 				new = alloc_huge_page(vma, addr, 1);
4796 				if (IS_ERR(new)) {
4797 					put_page(ptepage);
4798 					ret = PTR_ERR(new);
4799 					break;
4800 				}
4801 				copy_user_huge_page(new, ptepage, addr, vma,
4802 						    npages);
4803 				put_page(ptepage);
4804 
4805 				/* Install the new huge page if src pte stable */
4806 				dst_ptl = huge_pte_lock(h, dst, dst_pte);
4807 				src_ptl = huge_pte_lockptr(h, src, src_pte);
4808 				spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4809 				entry = huge_ptep_get(src_pte);
4810 				if (!pte_same(src_pte_old, entry)) {
4811 					restore_reserve_on_error(h, vma, addr,
4812 								new);
4813 					put_page(new);
4814 					/* dst_entry won't change as in child */
4815 					goto again;
4816 				}
4817 				hugetlb_install_page(vma, dst_pte, addr, new);
4818 				spin_unlock(src_ptl);
4819 				spin_unlock(dst_ptl);
4820 				continue;
4821 			}
4822 
4823 			if (cow) {
4824 				/*
4825 				 * No need to notify as we are downgrading page
4826 				 * table protection not changing it to point
4827 				 * to a new page.
4828 				 *
4829 				 * See Documentation/vm/mmu_notifier.rst
4830 				 */
4831 				huge_ptep_set_wrprotect(src, addr, src_pte);
4832 				entry = huge_pte_wrprotect(entry);
4833 			}
4834 
4835 			page_dup_rmap(ptepage, true);
4836 			set_huge_pte_at(dst, addr, dst_pte, entry);
4837 			hugetlb_count_add(npages, dst);
4838 		}
4839 		spin_unlock(src_ptl);
4840 		spin_unlock(dst_ptl);
4841 	}
4842 
4843 	if (cow)
4844 		mmu_notifier_invalidate_range_end(&range);
4845 	else
4846 		i_mmap_unlock_read(mapping);
4847 
4848 	return ret;
4849 }
4850 
4851 static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
4852 			  unsigned long new_addr, pte_t *src_pte)
4853 {
4854 	struct hstate *h = hstate_vma(vma);
4855 	struct mm_struct *mm = vma->vm_mm;
4856 	pte_t *dst_pte, pte;
4857 	spinlock_t *src_ptl, *dst_ptl;
4858 
4859 	dst_pte = huge_pte_offset(mm, new_addr, huge_page_size(h));
4860 	dst_ptl = huge_pte_lock(h, mm, dst_pte);
4861 	src_ptl = huge_pte_lockptr(h, mm, src_pte);
4862 
4863 	/*
4864 	 * We don't have to worry about the ordering of src and dst ptlocks
4865 	 * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock.
4866 	 */
4867 	if (src_ptl != dst_ptl)
4868 		spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4869 
4870 	pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);
4871 	set_huge_pte_at(mm, new_addr, dst_pte, pte);
4872 
4873 	if (src_ptl != dst_ptl)
4874 		spin_unlock(src_ptl);
4875 	spin_unlock(dst_ptl);
4876 }
4877 
4878 int move_hugetlb_page_tables(struct vm_area_struct *vma,
4879 			     struct vm_area_struct *new_vma,
4880 			     unsigned long old_addr, unsigned long new_addr,
4881 			     unsigned long len)
4882 {
4883 	struct hstate *h = hstate_vma(vma);
4884 	struct address_space *mapping = vma->vm_file->f_mapping;
4885 	unsigned long sz = huge_page_size(h);
4886 	struct mm_struct *mm = vma->vm_mm;
4887 	unsigned long old_end = old_addr + len;
4888 	unsigned long old_addr_copy;
4889 	pte_t *src_pte, *dst_pte;
4890 	struct mmu_notifier_range range;
4891 
4892 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr,
4893 				old_end);
4894 	adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4895 	mmu_notifier_invalidate_range_start(&range);
4896 	/* Prevent race with file truncation */
4897 	i_mmap_lock_write(mapping);
4898 	for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
4899 		src_pte = huge_pte_offset(mm, old_addr, sz);
4900 		if (!src_pte)
4901 			continue;
4902 		if (huge_pte_none(huge_ptep_get(src_pte)))
4903 			continue;
4904 
4905 		/* old_addr arg to huge_pmd_unshare() is a pointer and so the
4906 		 * arg may be modified. Pass a copy instead to preserve the
4907 		 * value in old_addr.
4908 		 */
4909 		old_addr_copy = old_addr;
4910 
4911 		if (huge_pmd_unshare(mm, vma, &old_addr_copy, src_pte))
4912 			continue;
4913 
4914 		dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz);
4915 		if (!dst_pte)
4916 			break;
4917 
4918 		move_huge_pte(vma, old_addr, new_addr, src_pte);
4919 	}
4920 	i_mmap_unlock_write(mapping);
4921 	flush_tlb_range(vma, old_end - len, old_end);
4922 	mmu_notifier_invalidate_range_end(&range);
4923 
4924 	return len + old_addr - old_end;
4925 }
4926 
4927 static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
4928 				   unsigned long start, unsigned long end,
4929 				   struct page *ref_page)
4930 {
4931 	struct mm_struct *mm = vma->vm_mm;
4932 	unsigned long address;
4933 	pte_t *ptep;
4934 	pte_t pte;
4935 	spinlock_t *ptl;
4936 	struct page *page;
4937 	struct hstate *h = hstate_vma(vma);
4938 	unsigned long sz = huge_page_size(h);
4939 	struct mmu_notifier_range range;
4940 
4941 	WARN_ON(!is_vm_hugetlb_page(vma));
4942 	BUG_ON(start & ~huge_page_mask(h));
4943 	BUG_ON(end & ~huge_page_mask(h));
4944 
4945 	/*
4946 	 * This is a hugetlb vma, all the pte entries should point
4947 	 * to huge page.
4948 	 */
4949 	tlb_change_page_size(tlb, sz);
4950 	tlb_start_vma(tlb, vma);
4951 
4952 	/*
4953 	 * If sharing possible, alert mmu notifiers of worst case.
4954 	 */
4955 	mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
4956 				end);
4957 	adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4958 	mmu_notifier_invalidate_range_start(&range);
4959 	address = start;
4960 	for (; address < end; address += sz) {
4961 		ptep = huge_pte_offset(mm, address, sz);
4962 		if (!ptep)
4963 			continue;
4964 
4965 		ptl = huge_pte_lock(h, mm, ptep);
4966 		if (huge_pmd_unshare(mm, vma, &address, ptep)) {
4967 			spin_unlock(ptl);
4968 			/*
4969 			 * We just unmapped a page of PMDs by clearing a PUD.
4970 			 * The caller's TLB flush range should cover this area.
4971 			 */
4972 			continue;
4973 		}
4974 
4975 		pte = huge_ptep_get(ptep);
4976 		if (huge_pte_none(pte)) {
4977 			spin_unlock(ptl);
4978 			continue;
4979 		}
4980 
4981 		/*
4982 		 * Migrating hugepage or HWPoisoned hugepage is already
4983 		 * unmapped and its refcount is dropped, so just clear pte here.
4984 		 */
4985 		if (unlikely(!pte_present(pte))) {
4986 			huge_pte_clear(mm, address, ptep, sz);
4987 			spin_unlock(ptl);
4988 			continue;
4989 		}
4990 
4991 		page = pte_page(pte);
4992 		/*
4993 		 * If a reference page is supplied, it is because a specific
4994 		 * page is being unmapped, not a range. Ensure the page we
4995 		 * are about to unmap is the actual page of interest.
4996 		 */
4997 		if (ref_page) {
4998 			if (page != ref_page) {
4999 				spin_unlock(ptl);
5000 				continue;
5001 			}
5002 			/*
5003 			 * Mark the VMA as having unmapped its page so that
5004 			 * future faults in this VMA will fail rather than
5005 			 * looking like data was lost
5006 			 */
5007 			set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
5008 		}
5009 
5010 		pte = huge_ptep_get_and_clear(mm, address, ptep);
5011 		tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
5012 		if (huge_pte_dirty(pte))
5013 			set_page_dirty(page);
5014 
5015 		hugetlb_count_sub(pages_per_huge_page(h), mm);
5016 		page_remove_rmap(page, true);
5017 
5018 		spin_unlock(ptl);
5019 		tlb_remove_page_size(tlb, page, huge_page_size(h));
5020 		/*
5021 		 * Bail out after unmapping reference page if supplied
5022 		 */
5023 		if (ref_page)
5024 			break;
5025 	}
5026 	mmu_notifier_invalidate_range_end(&range);
5027 	tlb_end_vma(tlb, vma);
5028 }
5029 
5030 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
5031 			  struct vm_area_struct *vma, unsigned long start,
5032 			  unsigned long end, struct page *ref_page)
5033 {
5034 	__unmap_hugepage_range(tlb, vma, start, end, ref_page);
5035 
5036 	/*
5037 	 * Clear this flag so that x86's huge_pmd_share page_table_shareable
5038 	 * test will fail on a vma being torn down, and not grab a page table
5039 	 * on its way out.  We're lucky that the flag has such an appropriate
5040 	 * name, and can in fact be safely cleared here. We could clear it
5041 	 * before the __unmap_hugepage_range above, but all that's necessary
5042 	 * is to clear it before releasing the i_mmap_rwsem. This works
5043 	 * because in the context this is called, the VMA is about to be
5044 	 * destroyed and the i_mmap_rwsem is held.
5045 	 */
5046 	vma->vm_flags &= ~VM_MAYSHARE;
5047 }
5048 
5049 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
5050 			  unsigned long end, struct page *ref_page)
5051 {
5052 	struct mmu_gather tlb;
5053 
5054 	tlb_gather_mmu(&tlb, vma->vm_mm);
5055 	__unmap_hugepage_range(&tlb, vma, start, end, ref_page);
5056 	tlb_finish_mmu(&tlb);
5057 }
5058 
5059 /*
5060  * This is called when the original mapper is failing to COW a MAP_PRIVATE
5061  * mapping it owns the reserve page for. The intention is to unmap the page
5062  * from other VMAs and let the children be SIGKILLed if they are faulting the
5063  * same region.
5064  */
5065 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
5066 			      struct page *page, unsigned long address)
5067 {
5068 	struct hstate *h = hstate_vma(vma);
5069 	struct vm_area_struct *iter_vma;
5070 	struct address_space *mapping;
5071 	pgoff_t pgoff;
5072 
5073 	/*
5074 	 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
5075 	 * from page cache lookup which is in HPAGE_SIZE units.
5076 	 */
5077 	address = address & huge_page_mask(h);
5078 	pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
5079 			vma->vm_pgoff;
5080 	mapping = vma->vm_file->f_mapping;
5081 
5082 	/*
5083 	 * Take the mapping lock for the duration of the table walk. As
5084 	 * this mapping should be shared between all the VMAs,
5085 	 * __unmap_hugepage_range() is called as the lock is already held
5086 	 */
5087 	i_mmap_lock_write(mapping);
5088 	vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
5089 		/* Do not unmap the current VMA */
5090 		if (iter_vma == vma)
5091 			continue;
5092 
5093 		/*
5094 		 * Shared VMAs have their own reserves and do not affect
5095 		 * MAP_PRIVATE accounting but it is possible that a shared
5096 		 * VMA is using the same page so check and skip such VMAs.
5097 		 */
5098 		if (iter_vma->vm_flags & VM_MAYSHARE)
5099 			continue;
5100 
5101 		/*
5102 		 * Unmap the page from other VMAs without their own reserves.
5103 		 * They get marked to be SIGKILLed if they fault in these
5104 		 * areas. This is because a future no-page fault on this VMA
5105 		 * could insert a zeroed page instead of the data existing
5106 		 * from the time of fork. This would look like data corruption
5107 		 */
5108 		if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
5109 			unmap_hugepage_range(iter_vma, address,
5110 					     address + huge_page_size(h), page);
5111 	}
5112 	i_mmap_unlock_write(mapping);
5113 }
5114 
5115 /*
5116  * Hugetlb_cow() should be called with page lock of the original hugepage held.
5117  * Called with hugetlb_fault_mutex_table held and pte_page locked so we
5118  * cannot race with other handlers or page migration.
5119  * Keep the pte_same checks anyway to make transition from the mutex easier.
5120  */
5121 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
5122 		       unsigned long address, pte_t *ptep,
5123 		       struct page *pagecache_page, spinlock_t *ptl)
5124 {
5125 	pte_t pte;
5126 	struct hstate *h = hstate_vma(vma);
5127 	struct page *old_page, *new_page;
5128 	int outside_reserve = 0;
5129 	vm_fault_t ret = 0;
5130 	unsigned long haddr = address & huge_page_mask(h);
5131 	struct mmu_notifier_range range;
5132 
5133 	pte = huge_ptep_get(ptep);
5134 	old_page = pte_page(pte);
5135 
5136 retry_avoidcopy:
5137 	/* If no-one else is actually using this page, avoid the copy
5138 	 * and just make the page writable */
5139 	if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
5140 		page_move_anon_rmap(old_page, vma);
5141 		set_huge_ptep_writable(vma, haddr, ptep);
5142 		return 0;
5143 	}
5144 
5145 	/*
5146 	 * If the process that created a MAP_PRIVATE mapping is about to
5147 	 * perform a COW due to a shared page count, attempt to satisfy
5148 	 * the allocation without using the existing reserves. The pagecache
5149 	 * page is used to determine if the reserve at this address was
5150 	 * consumed or not. If reserves were used, a partial faulted mapping
5151 	 * at the time of fork() could consume its reserves on COW instead
5152 	 * of the full address range.
5153 	 */
5154 	if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
5155 			old_page != pagecache_page)
5156 		outside_reserve = 1;
5157 
5158 	get_page(old_page);
5159 
5160 	/*
5161 	 * Drop page table lock as buddy allocator may be called. It will
5162 	 * be acquired again before returning to the caller, as expected.
5163 	 */
5164 	spin_unlock(ptl);
5165 	new_page = alloc_huge_page(vma, haddr, outside_reserve);
5166 
5167 	if (IS_ERR(new_page)) {
5168 		/*
5169 		 * If a process owning a MAP_PRIVATE mapping fails to COW,
5170 		 * it is due to references held by a child and an insufficient
5171 		 * huge page pool. To guarantee the original mappers
5172 		 * reliability, unmap the page from child processes. The child
5173 		 * may get SIGKILLed if it later faults.
5174 		 */
5175 		if (outside_reserve) {
5176 			struct address_space *mapping = vma->vm_file->f_mapping;
5177 			pgoff_t idx;
5178 			u32 hash;
5179 
5180 			put_page(old_page);
5181 			BUG_ON(huge_pte_none(pte));
5182 			/*
5183 			 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
5184 			 * unmapping.  unmapping needs to hold i_mmap_rwsem
5185 			 * in write mode.  Dropping i_mmap_rwsem in read mode
5186 			 * here is OK as COW mappings do not interact with
5187 			 * PMD sharing.
5188 			 *
5189 			 * Reacquire both after unmap operation.
5190 			 */
5191 			idx = vma_hugecache_offset(h, vma, haddr);
5192 			hash = hugetlb_fault_mutex_hash(mapping, idx);
5193 			mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5194 			i_mmap_unlock_read(mapping);
5195 
5196 			unmap_ref_private(mm, vma, old_page, haddr);
5197 
5198 			i_mmap_lock_read(mapping);
5199 			mutex_lock(&hugetlb_fault_mutex_table[hash]);
5200 			spin_lock(ptl);
5201 			ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5202 			if (likely(ptep &&
5203 				   pte_same(huge_ptep_get(ptep), pte)))
5204 				goto retry_avoidcopy;
5205 			/*
5206 			 * race occurs while re-acquiring page table
5207 			 * lock, and our job is done.
5208 			 */
5209 			return 0;
5210 		}
5211 
5212 		ret = vmf_error(PTR_ERR(new_page));
5213 		goto out_release_old;
5214 	}
5215 
5216 	/*
5217 	 * When the original hugepage is shared one, it does not have
5218 	 * anon_vma prepared.
5219 	 */
5220 	if (unlikely(anon_vma_prepare(vma))) {
5221 		ret = VM_FAULT_OOM;
5222 		goto out_release_all;
5223 	}
5224 
5225 	copy_user_huge_page(new_page, old_page, address, vma,
5226 			    pages_per_huge_page(h));
5227 	__SetPageUptodate(new_page);
5228 
5229 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
5230 				haddr + huge_page_size(h));
5231 	mmu_notifier_invalidate_range_start(&range);
5232 
5233 	/*
5234 	 * Retake the page table lock to check for racing updates
5235 	 * before the page tables are altered
5236 	 */
5237 	spin_lock(ptl);
5238 	ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5239 	if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
5240 		ClearHPageRestoreReserve(new_page);
5241 
5242 		/* Break COW */
5243 		huge_ptep_clear_flush(vma, haddr, ptep);
5244 		mmu_notifier_invalidate_range(mm, range.start, range.end);
5245 		set_huge_pte_at(mm, haddr, ptep,
5246 				make_huge_pte(vma, new_page, 1));
5247 		page_remove_rmap(old_page, true);
5248 		hugepage_add_new_anon_rmap(new_page, vma, haddr);
5249 		SetHPageMigratable(new_page);
5250 		/* Make the old page be freed below */
5251 		new_page = old_page;
5252 	}
5253 	spin_unlock(ptl);
5254 	mmu_notifier_invalidate_range_end(&range);
5255 out_release_all:
5256 	/* No restore in case of successful pagetable update (Break COW) */
5257 	if (new_page != old_page)
5258 		restore_reserve_on_error(h, vma, haddr, new_page);
5259 	put_page(new_page);
5260 out_release_old:
5261 	put_page(old_page);
5262 
5263 	spin_lock(ptl); /* Caller expects lock to be held */
5264 	return ret;
5265 }
5266 
5267 /* Return the pagecache page at a given address within a VMA */
5268 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
5269 			struct vm_area_struct *vma, unsigned long address)
5270 {
5271 	struct address_space *mapping;
5272 	pgoff_t idx;
5273 
5274 	mapping = vma->vm_file->f_mapping;
5275 	idx = vma_hugecache_offset(h, vma, address);
5276 
5277 	return find_lock_page(mapping, idx);
5278 }
5279 
5280 /*
5281  * Return whether there is a pagecache page to back given address within VMA.
5282  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
5283  */
5284 static bool hugetlbfs_pagecache_present(struct hstate *h,
5285 			struct vm_area_struct *vma, unsigned long address)
5286 {
5287 	struct address_space *mapping;
5288 	pgoff_t idx;
5289 	struct page *page;
5290 
5291 	mapping = vma->vm_file->f_mapping;
5292 	idx = vma_hugecache_offset(h, vma, address);
5293 
5294 	page = find_get_page(mapping, idx);
5295 	if (page)
5296 		put_page(page);
5297 	return page != NULL;
5298 }
5299 
5300 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
5301 			   pgoff_t idx)
5302 {
5303 	struct inode *inode = mapping->host;
5304 	struct hstate *h = hstate_inode(inode);
5305 	int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
5306 
5307 	if (err)
5308 		return err;
5309 	ClearHPageRestoreReserve(page);
5310 
5311 	/*
5312 	 * set page dirty so that it will not be removed from cache/file
5313 	 * by non-hugetlbfs specific code paths.
5314 	 */
5315 	set_page_dirty(page);
5316 
5317 	spin_lock(&inode->i_lock);
5318 	inode->i_blocks += blocks_per_huge_page(h);
5319 	spin_unlock(&inode->i_lock);
5320 	return 0;
5321 }
5322 
5323 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
5324 						  struct address_space *mapping,
5325 						  pgoff_t idx,
5326 						  unsigned int flags,
5327 						  unsigned long haddr,
5328 						  unsigned long reason)
5329 {
5330 	vm_fault_t ret;
5331 	u32 hash;
5332 	struct vm_fault vmf = {
5333 		.vma = vma,
5334 		.address = haddr,
5335 		.flags = flags,
5336 
5337 		/*
5338 		 * Hard to debug if it ends up being
5339 		 * used by a callee that assumes
5340 		 * something about the other
5341 		 * uninitialized fields... same as in
5342 		 * memory.c
5343 		 */
5344 	};
5345 
5346 	/*
5347 	 * hugetlb_fault_mutex and i_mmap_rwsem must be
5348 	 * dropped before handling userfault.  Reacquire
5349 	 * after handling fault to make calling code simpler.
5350 	 */
5351 	hash = hugetlb_fault_mutex_hash(mapping, idx);
5352 	mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5353 	i_mmap_unlock_read(mapping);
5354 	ret = handle_userfault(&vmf, reason);
5355 	i_mmap_lock_read(mapping);
5356 	mutex_lock(&hugetlb_fault_mutex_table[hash]);
5357 
5358 	return ret;
5359 }
5360 
5361 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
5362 			struct vm_area_struct *vma,
5363 			struct address_space *mapping, pgoff_t idx,
5364 			unsigned long address, pte_t *ptep, unsigned int flags)
5365 {
5366 	struct hstate *h = hstate_vma(vma);
5367 	vm_fault_t ret = VM_FAULT_SIGBUS;
5368 	int anon_rmap = 0;
5369 	unsigned long size;
5370 	struct page *page;
5371 	pte_t new_pte;
5372 	spinlock_t *ptl;
5373 	unsigned long haddr = address & huge_page_mask(h);
5374 	bool new_page, new_pagecache_page = false;
5375 
5376 	/*
5377 	 * Currently, we are forced to kill the process in the event the
5378 	 * original mapper has unmapped pages from the child due to a failed
5379 	 * COW. Warn that such a situation has occurred as it may not be obvious
5380 	 */
5381 	if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
5382 		pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
5383 			   current->pid);
5384 		return ret;
5385 	}
5386 
5387 	/*
5388 	 * We can not race with truncation due to holding i_mmap_rwsem.
5389 	 * i_size is modified when holding i_mmap_rwsem, so check here
5390 	 * once for faults beyond end of file.
5391 	 */
5392 	size = i_size_read(mapping->host) >> huge_page_shift(h);
5393 	if (idx >= size)
5394 		goto out;
5395 
5396 retry:
5397 	new_page = false;
5398 	page = find_lock_page(mapping, idx);
5399 	if (!page) {
5400 		/* Check for page in userfault range */
5401 		if (userfaultfd_missing(vma)) {
5402 			ret = hugetlb_handle_userfault(vma, mapping, idx,
5403 						       flags, haddr,
5404 						       VM_UFFD_MISSING);
5405 			goto out;
5406 		}
5407 
5408 		page = alloc_huge_page(vma, haddr, 0);
5409 		if (IS_ERR(page)) {
5410 			/*
5411 			 * Returning error will result in faulting task being
5412 			 * sent SIGBUS.  The hugetlb fault mutex prevents two
5413 			 * tasks from racing to fault in the same page which
5414 			 * could result in false unable to allocate errors.
5415 			 * Page migration does not take the fault mutex, but
5416 			 * does a clear then write of pte's under page table
5417 			 * lock.  Page fault code could race with migration,
5418 			 * notice the clear pte and try to allocate a page
5419 			 * here.  Before returning error, get ptl and make
5420 			 * sure there really is no pte entry.
5421 			 */
5422 			ptl = huge_pte_lock(h, mm, ptep);
5423 			ret = 0;
5424 			if (huge_pte_none(huge_ptep_get(ptep)))
5425 				ret = vmf_error(PTR_ERR(page));
5426 			spin_unlock(ptl);
5427 			goto out;
5428 		}
5429 		clear_huge_page(page, address, pages_per_huge_page(h));
5430 		__SetPageUptodate(page);
5431 		new_page = true;
5432 
5433 		if (vma->vm_flags & VM_MAYSHARE) {
5434 			int err = huge_add_to_page_cache(page, mapping, idx);
5435 			if (err) {
5436 				put_page(page);
5437 				if (err == -EEXIST)
5438 					goto retry;
5439 				goto out;
5440 			}
5441 			new_pagecache_page = true;
5442 		} else {
5443 			lock_page(page);
5444 			if (unlikely(anon_vma_prepare(vma))) {
5445 				ret = VM_FAULT_OOM;
5446 				goto backout_unlocked;
5447 			}
5448 			anon_rmap = 1;
5449 		}
5450 	} else {
5451 		/*
5452 		 * If memory error occurs between mmap() and fault, some process
5453 		 * don't have hwpoisoned swap entry for errored virtual address.
5454 		 * So we need to block hugepage fault by PG_hwpoison bit check.
5455 		 */
5456 		if (unlikely(PageHWPoison(page))) {
5457 			ret = VM_FAULT_HWPOISON_LARGE |
5458 				VM_FAULT_SET_HINDEX(hstate_index(h));
5459 			goto backout_unlocked;
5460 		}
5461 
5462 		/* Check for page in userfault range. */
5463 		if (userfaultfd_minor(vma)) {
5464 			unlock_page(page);
5465 			put_page(page);
5466 			ret = hugetlb_handle_userfault(vma, mapping, idx,
5467 						       flags, haddr,
5468 						       VM_UFFD_MINOR);
5469 			goto out;
5470 		}
5471 	}
5472 
5473 	/*
5474 	 * If we are going to COW a private mapping later, we examine the
5475 	 * pending reservations for this page now. This will ensure that
5476 	 * any allocations necessary to record that reservation occur outside
5477 	 * the spinlock.
5478 	 */
5479 	if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5480 		if (vma_needs_reservation(h, vma, haddr) < 0) {
5481 			ret = VM_FAULT_OOM;
5482 			goto backout_unlocked;
5483 		}
5484 		/* Just decrements count, does not deallocate */
5485 		vma_end_reservation(h, vma, haddr);
5486 	}
5487 
5488 	ptl = huge_pte_lock(h, mm, ptep);
5489 	ret = 0;
5490 	if (!huge_pte_none(huge_ptep_get(ptep)))
5491 		goto backout;
5492 
5493 	if (anon_rmap) {
5494 		ClearHPageRestoreReserve(page);
5495 		hugepage_add_new_anon_rmap(page, vma, haddr);
5496 	} else
5497 		page_dup_rmap(page, true);
5498 	new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
5499 				&& (vma->vm_flags & VM_SHARED)));
5500 	set_huge_pte_at(mm, haddr, ptep, new_pte);
5501 
5502 	hugetlb_count_add(pages_per_huge_page(h), mm);
5503 	if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5504 		/* Optimization, do the COW without a second fault */
5505 		ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
5506 	}
5507 
5508 	spin_unlock(ptl);
5509 
5510 	/*
5511 	 * Only set HPageMigratable in newly allocated pages.  Existing pages
5512 	 * found in the pagecache may not have HPageMigratableset if they have
5513 	 * been isolated for migration.
5514 	 */
5515 	if (new_page)
5516 		SetHPageMigratable(page);
5517 
5518 	unlock_page(page);
5519 out:
5520 	return ret;
5521 
5522 backout:
5523 	spin_unlock(ptl);
5524 backout_unlocked:
5525 	unlock_page(page);
5526 	/* restore reserve for newly allocated pages not in page cache */
5527 	if (new_page && !new_pagecache_page)
5528 		restore_reserve_on_error(h, vma, haddr, page);
5529 	put_page(page);
5530 	goto out;
5531 }
5532 
5533 #ifdef CONFIG_SMP
5534 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5535 {
5536 	unsigned long key[2];
5537 	u32 hash;
5538 
5539 	key[0] = (unsigned long) mapping;
5540 	key[1] = idx;
5541 
5542 	hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
5543 
5544 	return hash & (num_fault_mutexes - 1);
5545 }
5546 #else
5547 /*
5548  * For uniprocessor systems we always use a single mutex, so just
5549  * return 0 and avoid the hashing overhead.
5550  */
5551 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5552 {
5553 	return 0;
5554 }
5555 #endif
5556 
5557 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
5558 			unsigned long address, unsigned int flags)
5559 {
5560 	pte_t *ptep, entry;
5561 	spinlock_t *ptl;
5562 	vm_fault_t ret;
5563 	u32 hash;
5564 	pgoff_t idx;
5565 	struct page *page = NULL;
5566 	struct page *pagecache_page = NULL;
5567 	struct hstate *h = hstate_vma(vma);
5568 	struct address_space *mapping;
5569 	int need_wait_lock = 0;
5570 	unsigned long haddr = address & huge_page_mask(h);
5571 
5572 	ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5573 	if (ptep) {
5574 		/*
5575 		 * Since we hold no locks, ptep could be stale.  That is
5576 		 * OK as we are only making decisions based on content and
5577 		 * not actually modifying content here.
5578 		 */
5579 		entry = huge_ptep_get(ptep);
5580 		if (unlikely(is_hugetlb_entry_migration(entry))) {
5581 			migration_entry_wait_huge(vma, mm, ptep);
5582 			return 0;
5583 		} else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
5584 			return VM_FAULT_HWPOISON_LARGE |
5585 				VM_FAULT_SET_HINDEX(hstate_index(h));
5586 	}
5587 
5588 	/*
5589 	 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
5590 	 * until finished with ptep.  This serves two purposes:
5591 	 * 1) It prevents huge_pmd_unshare from being called elsewhere
5592 	 *    and making the ptep no longer valid.
5593 	 * 2) It synchronizes us with i_size modifications during truncation.
5594 	 *
5595 	 * ptep could have already be assigned via huge_pte_offset.  That
5596 	 * is OK, as huge_pte_alloc will return the same value unless
5597 	 * something has changed.
5598 	 */
5599 	mapping = vma->vm_file->f_mapping;
5600 	i_mmap_lock_read(mapping);
5601 	ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
5602 	if (!ptep) {
5603 		i_mmap_unlock_read(mapping);
5604 		return VM_FAULT_OOM;
5605 	}
5606 
5607 	/*
5608 	 * Serialize hugepage allocation and instantiation, so that we don't
5609 	 * get spurious allocation failures if two CPUs race to instantiate
5610 	 * the same page in the page cache.
5611 	 */
5612 	idx = vma_hugecache_offset(h, vma, haddr);
5613 	hash = hugetlb_fault_mutex_hash(mapping, idx);
5614 	mutex_lock(&hugetlb_fault_mutex_table[hash]);
5615 
5616 	entry = huge_ptep_get(ptep);
5617 	if (huge_pte_none(entry)) {
5618 		ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
5619 		goto out_mutex;
5620 	}
5621 
5622 	ret = 0;
5623 
5624 	/*
5625 	 * entry could be a migration/hwpoison entry at this point, so this
5626 	 * check prevents the kernel from going below assuming that we have
5627 	 * an active hugepage in pagecache. This goto expects the 2nd page
5628 	 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
5629 	 * properly handle it.
5630 	 */
5631 	if (!pte_present(entry))
5632 		goto out_mutex;
5633 
5634 	/*
5635 	 * If we are going to COW the mapping later, we examine the pending
5636 	 * reservations for this page now. This will ensure that any
5637 	 * allocations necessary to record that reservation occur outside the
5638 	 * spinlock. For private mappings, we also lookup the pagecache
5639 	 * page now as it is used to determine if a reservation has been
5640 	 * consumed.
5641 	 */
5642 	if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
5643 		if (vma_needs_reservation(h, vma, haddr) < 0) {
5644 			ret = VM_FAULT_OOM;
5645 			goto out_mutex;
5646 		}
5647 		/* Just decrements count, does not deallocate */
5648 		vma_end_reservation(h, vma, haddr);
5649 
5650 		if (!(vma->vm_flags & VM_MAYSHARE))
5651 			pagecache_page = hugetlbfs_pagecache_page(h,
5652 								vma, haddr);
5653 	}
5654 
5655 	ptl = huge_pte_lock(h, mm, ptep);
5656 
5657 	/* Check for a racing update before calling hugetlb_cow */
5658 	if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
5659 		goto out_ptl;
5660 
5661 	/*
5662 	 * hugetlb_cow() requires page locks of pte_page(entry) and
5663 	 * pagecache_page, so here we need take the former one
5664 	 * when page != pagecache_page or !pagecache_page.
5665 	 */
5666 	page = pte_page(entry);
5667 	if (page != pagecache_page)
5668 		if (!trylock_page(page)) {
5669 			need_wait_lock = 1;
5670 			goto out_ptl;
5671 		}
5672 
5673 	get_page(page);
5674 
5675 	if (flags & FAULT_FLAG_WRITE) {
5676 		if (!huge_pte_write(entry)) {
5677 			ret = hugetlb_cow(mm, vma, address, ptep,
5678 					  pagecache_page, ptl);
5679 			goto out_put_page;
5680 		}
5681 		entry = huge_pte_mkdirty(entry);
5682 	}
5683 	entry = pte_mkyoung(entry);
5684 	if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
5685 						flags & FAULT_FLAG_WRITE))
5686 		update_mmu_cache(vma, haddr, ptep);
5687 out_put_page:
5688 	if (page != pagecache_page)
5689 		unlock_page(page);
5690 	put_page(page);
5691 out_ptl:
5692 	spin_unlock(ptl);
5693 
5694 	if (pagecache_page) {
5695 		unlock_page(pagecache_page);
5696 		put_page(pagecache_page);
5697 	}
5698 out_mutex:
5699 	mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5700 	i_mmap_unlock_read(mapping);
5701 	/*
5702 	 * Generally it's safe to hold refcount during waiting page lock. But
5703 	 * here we just wait to defer the next page fault to avoid busy loop and
5704 	 * the page is not used after unlocked before returning from the current
5705 	 * page fault. So we are safe from accessing freed page, even if we wait
5706 	 * here without taking refcount.
5707 	 */
5708 	if (need_wait_lock)
5709 		wait_on_page_locked(page);
5710 	return ret;
5711 }
5712 
5713 #ifdef CONFIG_USERFAULTFD
5714 /*
5715  * Used by userfaultfd UFFDIO_COPY.  Based on mcopy_atomic_pte with
5716  * modifications for huge pages.
5717  */
5718 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
5719 			    pte_t *dst_pte,
5720 			    struct vm_area_struct *dst_vma,
5721 			    unsigned long dst_addr,
5722 			    unsigned long src_addr,
5723 			    enum mcopy_atomic_mode mode,
5724 			    struct page **pagep)
5725 {
5726 	bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE);
5727 	struct hstate *h = hstate_vma(dst_vma);
5728 	struct address_space *mapping = dst_vma->vm_file->f_mapping;
5729 	pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
5730 	unsigned long size;
5731 	int vm_shared = dst_vma->vm_flags & VM_SHARED;
5732 	pte_t _dst_pte;
5733 	spinlock_t *ptl;
5734 	int ret = -ENOMEM;
5735 	struct page *page;
5736 	int writable;
5737 	bool new_pagecache_page = false;
5738 
5739 	if (is_continue) {
5740 		ret = -EFAULT;
5741 		page = find_lock_page(mapping, idx);
5742 		if (!page)
5743 			goto out;
5744 	} else if (!*pagep) {
5745 		/* If a page already exists, then it's UFFDIO_COPY for
5746 		 * a non-missing case. Return -EEXIST.
5747 		 */
5748 		if (vm_shared &&
5749 		    hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5750 			ret = -EEXIST;
5751 			goto out;
5752 		}
5753 
5754 		page = alloc_huge_page(dst_vma, dst_addr, 0);
5755 		if (IS_ERR(page)) {
5756 			ret = -ENOMEM;
5757 			goto out;
5758 		}
5759 
5760 		ret = copy_huge_page_from_user(page,
5761 						(const void __user *) src_addr,
5762 						pages_per_huge_page(h), false);
5763 
5764 		/* fallback to copy_from_user outside mmap_lock */
5765 		if (unlikely(ret)) {
5766 			ret = -ENOENT;
5767 			/* Free the allocated page which may have
5768 			 * consumed a reservation.
5769 			 */
5770 			restore_reserve_on_error(h, dst_vma, dst_addr, page);
5771 			put_page(page);
5772 
5773 			/* Allocate a temporary page to hold the copied
5774 			 * contents.
5775 			 */
5776 			page = alloc_huge_page_vma(h, dst_vma, dst_addr);
5777 			if (!page) {
5778 				ret = -ENOMEM;
5779 				goto out;
5780 			}
5781 			*pagep = page;
5782 			/* Set the outparam pagep and return to the caller to
5783 			 * copy the contents outside the lock. Don't free the
5784 			 * page.
5785 			 */
5786 			goto out;
5787 		}
5788 	} else {
5789 		if (vm_shared &&
5790 		    hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5791 			put_page(*pagep);
5792 			ret = -EEXIST;
5793 			*pagep = NULL;
5794 			goto out;
5795 		}
5796 
5797 		page = alloc_huge_page(dst_vma, dst_addr, 0);
5798 		if (IS_ERR(page)) {
5799 			ret = -ENOMEM;
5800 			*pagep = NULL;
5801 			goto out;
5802 		}
5803 		folio_copy(page_folio(page), page_folio(*pagep));
5804 		put_page(*pagep);
5805 		*pagep = NULL;
5806 	}
5807 
5808 	/*
5809 	 * The memory barrier inside __SetPageUptodate makes sure that
5810 	 * preceding stores to the page contents become visible before
5811 	 * the set_pte_at() write.
5812 	 */
5813 	__SetPageUptodate(page);
5814 
5815 	/* Add shared, newly allocated pages to the page cache. */
5816 	if (vm_shared && !is_continue) {
5817 		size = i_size_read(mapping->host) >> huge_page_shift(h);
5818 		ret = -EFAULT;
5819 		if (idx >= size)
5820 			goto out_release_nounlock;
5821 
5822 		/*
5823 		 * Serialization between remove_inode_hugepages() and
5824 		 * huge_add_to_page_cache() below happens through the
5825 		 * hugetlb_fault_mutex_table that here must be hold by
5826 		 * the caller.
5827 		 */
5828 		ret = huge_add_to_page_cache(page, mapping, idx);
5829 		if (ret)
5830 			goto out_release_nounlock;
5831 		new_pagecache_page = true;
5832 	}
5833 
5834 	ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
5835 	spin_lock(ptl);
5836 
5837 	/*
5838 	 * Recheck the i_size after holding PT lock to make sure not
5839 	 * to leave any page mapped (as page_mapped()) beyond the end
5840 	 * of the i_size (remove_inode_hugepages() is strict about
5841 	 * enforcing that). If we bail out here, we'll also leave a
5842 	 * page in the radix tree in the vm_shared case beyond the end
5843 	 * of the i_size, but remove_inode_hugepages() will take care
5844 	 * of it as soon as we drop the hugetlb_fault_mutex_table.
5845 	 */
5846 	size = i_size_read(mapping->host) >> huge_page_shift(h);
5847 	ret = -EFAULT;
5848 	if (idx >= size)
5849 		goto out_release_unlock;
5850 
5851 	ret = -EEXIST;
5852 	if (!huge_pte_none(huge_ptep_get(dst_pte)))
5853 		goto out_release_unlock;
5854 
5855 	if (vm_shared) {
5856 		page_dup_rmap(page, true);
5857 	} else {
5858 		ClearHPageRestoreReserve(page);
5859 		hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
5860 	}
5861 
5862 	/* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */
5863 	if (is_continue && !vm_shared)
5864 		writable = 0;
5865 	else
5866 		writable = dst_vma->vm_flags & VM_WRITE;
5867 
5868 	_dst_pte = make_huge_pte(dst_vma, page, writable);
5869 	if (writable)
5870 		_dst_pte = huge_pte_mkdirty(_dst_pte);
5871 	_dst_pte = pte_mkyoung(_dst_pte);
5872 
5873 	set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
5874 
5875 	(void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
5876 					dst_vma->vm_flags & VM_WRITE);
5877 	hugetlb_count_add(pages_per_huge_page(h), dst_mm);
5878 
5879 	/* No need to invalidate - it was non-present before */
5880 	update_mmu_cache(dst_vma, dst_addr, dst_pte);
5881 
5882 	spin_unlock(ptl);
5883 	if (!is_continue)
5884 		SetHPageMigratable(page);
5885 	if (vm_shared || is_continue)
5886 		unlock_page(page);
5887 	ret = 0;
5888 out:
5889 	return ret;
5890 out_release_unlock:
5891 	spin_unlock(ptl);
5892 	if (vm_shared || is_continue)
5893 		unlock_page(page);
5894 out_release_nounlock:
5895 	if (!new_pagecache_page)
5896 		restore_reserve_on_error(h, dst_vma, dst_addr, page);
5897 	put_page(page);
5898 	goto out;
5899 }
5900 #endif /* CONFIG_USERFAULTFD */
5901 
5902 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
5903 				 int refs, struct page **pages,
5904 				 struct vm_area_struct **vmas)
5905 {
5906 	int nr;
5907 
5908 	for (nr = 0; nr < refs; nr++) {
5909 		if (likely(pages))
5910 			pages[nr] = mem_map_offset(page, nr);
5911 		if (vmas)
5912 			vmas[nr] = vma;
5913 	}
5914 }
5915 
5916 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
5917 			 struct page **pages, struct vm_area_struct **vmas,
5918 			 unsigned long *position, unsigned long *nr_pages,
5919 			 long i, unsigned int flags, int *locked)
5920 {
5921 	unsigned long pfn_offset;
5922 	unsigned long vaddr = *position;
5923 	unsigned long remainder = *nr_pages;
5924 	struct hstate *h = hstate_vma(vma);
5925 	int err = -EFAULT, refs;
5926 
5927 	while (vaddr < vma->vm_end && remainder) {
5928 		pte_t *pte;
5929 		spinlock_t *ptl = NULL;
5930 		int absent;
5931 		struct page *page;
5932 
5933 		/*
5934 		 * If we have a pending SIGKILL, don't keep faulting pages and
5935 		 * potentially allocating memory.
5936 		 */
5937 		if (fatal_signal_pending(current)) {
5938 			remainder = 0;
5939 			break;
5940 		}
5941 
5942 		/*
5943 		 * Some archs (sparc64, sh*) have multiple pte_ts to
5944 		 * each hugepage.  We have to make sure we get the
5945 		 * first, for the page indexing below to work.
5946 		 *
5947 		 * Note that page table lock is not held when pte is null.
5948 		 */
5949 		pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
5950 				      huge_page_size(h));
5951 		if (pte)
5952 			ptl = huge_pte_lock(h, mm, pte);
5953 		absent = !pte || huge_pte_none(huge_ptep_get(pte));
5954 
5955 		/*
5956 		 * When coredumping, it suits get_dump_page if we just return
5957 		 * an error where there's an empty slot with no huge pagecache
5958 		 * to back it.  This way, we avoid allocating a hugepage, and
5959 		 * the sparse dumpfile avoids allocating disk blocks, but its
5960 		 * huge holes still show up with zeroes where they need to be.
5961 		 */
5962 		if (absent && (flags & FOLL_DUMP) &&
5963 		    !hugetlbfs_pagecache_present(h, vma, vaddr)) {
5964 			if (pte)
5965 				spin_unlock(ptl);
5966 			remainder = 0;
5967 			break;
5968 		}
5969 
5970 		/*
5971 		 * We need call hugetlb_fault for both hugepages under migration
5972 		 * (in which case hugetlb_fault waits for the migration,) and
5973 		 * hwpoisoned hugepages (in which case we need to prevent the
5974 		 * caller from accessing to them.) In order to do this, we use
5975 		 * here is_swap_pte instead of is_hugetlb_entry_migration and
5976 		 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
5977 		 * both cases, and because we can't follow correct pages
5978 		 * directly from any kind of swap entries.
5979 		 */
5980 		if (absent || is_swap_pte(huge_ptep_get(pte)) ||
5981 		    ((flags & FOLL_WRITE) &&
5982 		      !huge_pte_write(huge_ptep_get(pte)))) {
5983 			vm_fault_t ret;
5984 			unsigned int fault_flags = 0;
5985 
5986 			if (pte)
5987 				spin_unlock(ptl);
5988 			if (flags & FOLL_WRITE)
5989 				fault_flags |= FAULT_FLAG_WRITE;
5990 			if (locked)
5991 				fault_flags |= FAULT_FLAG_ALLOW_RETRY |
5992 					FAULT_FLAG_KILLABLE;
5993 			if (flags & FOLL_NOWAIT)
5994 				fault_flags |= FAULT_FLAG_ALLOW_RETRY |
5995 					FAULT_FLAG_RETRY_NOWAIT;
5996 			if (flags & FOLL_TRIED) {
5997 				/*
5998 				 * Note: FAULT_FLAG_ALLOW_RETRY and
5999 				 * FAULT_FLAG_TRIED can co-exist
6000 				 */
6001 				fault_flags |= FAULT_FLAG_TRIED;
6002 			}
6003 			ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
6004 			if (ret & VM_FAULT_ERROR) {
6005 				err = vm_fault_to_errno(ret, flags);
6006 				remainder = 0;
6007 				break;
6008 			}
6009 			if (ret & VM_FAULT_RETRY) {
6010 				if (locked &&
6011 				    !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
6012 					*locked = 0;
6013 				*nr_pages = 0;
6014 				/*
6015 				 * VM_FAULT_RETRY must not return an
6016 				 * error, it will return zero
6017 				 * instead.
6018 				 *
6019 				 * No need to update "position" as the
6020 				 * caller will not check it after
6021 				 * *nr_pages is set to 0.
6022 				 */
6023 				return i;
6024 			}
6025 			continue;
6026 		}
6027 
6028 		pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
6029 		page = pte_page(huge_ptep_get(pte));
6030 
6031 		/*
6032 		 * If subpage information not requested, update counters
6033 		 * and skip the same_page loop below.
6034 		 */
6035 		if (!pages && !vmas && !pfn_offset &&
6036 		    (vaddr + huge_page_size(h) < vma->vm_end) &&
6037 		    (remainder >= pages_per_huge_page(h))) {
6038 			vaddr += huge_page_size(h);
6039 			remainder -= pages_per_huge_page(h);
6040 			i += pages_per_huge_page(h);
6041 			spin_unlock(ptl);
6042 			continue;
6043 		}
6044 
6045 		/* vaddr may not be aligned to PAGE_SIZE */
6046 		refs = min3(pages_per_huge_page(h) - pfn_offset, remainder,
6047 		    (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT);
6048 
6049 		if (pages || vmas)
6050 			record_subpages_vmas(mem_map_offset(page, pfn_offset),
6051 					     vma, refs,
6052 					     likely(pages) ? pages + i : NULL,
6053 					     vmas ? vmas + i : NULL);
6054 
6055 		if (pages) {
6056 			/*
6057 			 * try_grab_compound_head() should always succeed here,
6058 			 * because: a) we hold the ptl lock, and b) we've just
6059 			 * checked that the huge page is present in the page
6060 			 * tables. If the huge page is present, then the tail
6061 			 * pages must also be present. The ptl prevents the
6062 			 * head page and tail pages from being rearranged in
6063 			 * any way. So this page must be available at this
6064 			 * point, unless the page refcount overflowed:
6065 			 */
6066 			if (WARN_ON_ONCE(!try_grab_compound_head(pages[i],
6067 								 refs,
6068 								 flags))) {
6069 				spin_unlock(ptl);
6070 				remainder = 0;
6071 				err = -ENOMEM;
6072 				break;
6073 			}
6074 		}
6075 
6076 		vaddr += (refs << PAGE_SHIFT);
6077 		remainder -= refs;
6078 		i += refs;
6079 
6080 		spin_unlock(ptl);
6081 	}
6082 	*nr_pages = remainder;
6083 	/*
6084 	 * setting position is actually required only if remainder is
6085 	 * not zero but it's faster not to add a "if (remainder)"
6086 	 * branch.
6087 	 */
6088 	*position = vaddr;
6089 
6090 	return i ? i : err;
6091 }
6092 
6093 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
6094 		unsigned long address, unsigned long end, pgprot_t newprot)
6095 {
6096 	struct mm_struct *mm = vma->vm_mm;
6097 	unsigned long start = address;
6098 	pte_t *ptep;
6099 	pte_t pte;
6100 	struct hstate *h = hstate_vma(vma);
6101 	unsigned long pages = 0;
6102 	bool shared_pmd = false;
6103 	struct mmu_notifier_range range;
6104 
6105 	/*
6106 	 * In the case of shared PMDs, the area to flush could be beyond
6107 	 * start/end.  Set range.start/range.end to cover the maximum possible
6108 	 * range if PMD sharing is possible.
6109 	 */
6110 	mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
6111 				0, vma, mm, start, end);
6112 	adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
6113 
6114 	BUG_ON(address >= end);
6115 	flush_cache_range(vma, range.start, range.end);
6116 
6117 	mmu_notifier_invalidate_range_start(&range);
6118 	i_mmap_lock_write(vma->vm_file->f_mapping);
6119 	for (; address < end; address += huge_page_size(h)) {
6120 		spinlock_t *ptl;
6121 		ptep = huge_pte_offset(mm, address, huge_page_size(h));
6122 		if (!ptep)
6123 			continue;
6124 		ptl = huge_pte_lock(h, mm, ptep);
6125 		if (huge_pmd_unshare(mm, vma, &address, ptep)) {
6126 			pages++;
6127 			spin_unlock(ptl);
6128 			shared_pmd = true;
6129 			continue;
6130 		}
6131 		pte = huge_ptep_get(ptep);
6132 		if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
6133 			spin_unlock(ptl);
6134 			continue;
6135 		}
6136 		if (unlikely(is_hugetlb_entry_migration(pte))) {
6137 			swp_entry_t entry = pte_to_swp_entry(pte);
6138 
6139 			if (is_writable_migration_entry(entry)) {
6140 				pte_t newpte;
6141 
6142 				entry = make_readable_migration_entry(
6143 							swp_offset(entry));
6144 				newpte = swp_entry_to_pte(entry);
6145 				set_huge_swap_pte_at(mm, address, ptep,
6146 						     newpte, huge_page_size(h));
6147 				pages++;
6148 			}
6149 			spin_unlock(ptl);
6150 			continue;
6151 		}
6152 		if (!huge_pte_none(pte)) {
6153 			pte_t old_pte;
6154 			unsigned int shift = huge_page_shift(hstate_vma(vma));
6155 
6156 			old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
6157 			pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
6158 			pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
6159 			huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
6160 			pages++;
6161 		}
6162 		spin_unlock(ptl);
6163 	}
6164 	/*
6165 	 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
6166 	 * may have cleared our pud entry and done put_page on the page table:
6167 	 * once we release i_mmap_rwsem, another task can do the final put_page
6168 	 * and that page table be reused and filled with junk.  If we actually
6169 	 * did unshare a page of pmds, flush the range corresponding to the pud.
6170 	 */
6171 	if (shared_pmd)
6172 		flush_hugetlb_tlb_range(vma, range.start, range.end);
6173 	else
6174 		flush_hugetlb_tlb_range(vma, start, end);
6175 	/*
6176 	 * No need to call mmu_notifier_invalidate_range() we are downgrading
6177 	 * page table protection not changing it to point to a new page.
6178 	 *
6179 	 * See Documentation/vm/mmu_notifier.rst
6180 	 */
6181 	i_mmap_unlock_write(vma->vm_file->f_mapping);
6182 	mmu_notifier_invalidate_range_end(&range);
6183 
6184 	return pages << h->order;
6185 }
6186 
6187 /* Return true if reservation was successful, false otherwise.  */
6188 bool hugetlb_reserve_pages(struct inode *inode,
6189 					long from, long to,
6190 					struct vm_area_struct *vma,
6191 					vm_flags_t vm_flags)
6192 {
6193 	long chg, add = -1;
6194 	struct hstate *h = hstate_inode(inode);
6195 	struct hugepage_subpool *spool = subpool_inode(inode);
6196 	struct resv_map *resv_map;
6197 	struct hugetlb_cgroup *h_cg = NULL;
6198 	long gbl_reserve, regions_needed = 0;
6199 
6200 	/* This should never happen */
6201 	if (from > to) {
6202 		VM_WARN(1, "%s called with a negative range\n", __func__);
6203 		return false;
6204 	}
6205 
6206 	/*
6207 	 * Only apply hugepage reservation if asked. At fault time, an
6208 	 * attempt will be made for VM_NORESERVE to allocate a page
6209 	 * without using reserves
6210 	 */
6211 	if (vm_flags & VM_NORESERVE)
6212 		return true;
6213 
6214 	/*
6215 	 * Shared mappings base their reservation on the number of pages that
6216 	 * are already allocated on behalf of the file. Private mappings need
6217 	 * to reserve the full area even if read-only as mprotect() may be
6218 	 * called to make the mapping read-write. Assume !vma is a shm mapping
6219 	 */
6220 	if (!vma || vma->vm_flags & VM_MAYSHARE) {
6221 		/*
6222 		 * resv_map can not be NULL as hugetlb_reserve_pages is only
6223 		 * called for inodes for which resv_maps were created (see
6224 		 * hugetlbfs_get_inode).
6225 		 */
6226 		resv_map = inode_resv_map(inode);
6227 
6228 		chg = region_chg(resv_map, from, to, &regions_needed);
6229 
6230 	} else {
6231 		/* Private mapping. */
6232 		resv_map = resv_map_alloc();
6233 		if (!resv_map)
6234 			return false;
6235 
6236 		chg = to - from;
6237 
6238 		set_vma_resv_map(vma, resv_map);
6239 		set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
6240 	}
6241 
6242 	if (chg < 0)
6243 		goto out_err;
6244 
6245 	if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
6246 				chg * pages_per_huge_page(h), &h_cg) < 0)
6247 		goto out_err;
6248 
6249 	if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
6250 		/* For private mappings, the hugetlb_cgroup uncharge info hangs
6251 		 * of the resv_map.
6252 		 */
6253 		resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
6254 	}
6255 
6256 	/*
6257 	 * There must be enough pages in the subpool for the mapping. If
6258 	 * the subpool has a minimum size, there may be some global
6259 	 * reservations already in place (gbl_reserve).
6260 	 */
6261 	gbl_reserve = hugepage_subpool_get_pages(spool, chg);
6262 	if (gbl_reserve < 0)
6263 		goto out_uncharge_cgroup;
6264 
6265 	/*
6266 	 * Check enough hugepages are available for the reservation.
6267 	 * Hand the pages back to the subpool if there are not
6268 	 */
6269 	if (hugetlb_acct_memory(h, gbl_reserve) < 0)
6270 		goto out_put_pages;
6271 
6272 	/*
6273 	 * Account for the reservations made. Shared mappings record regions
6274 	 * that have reservations as they are shared by multiple VMAs.
6275 	 * When the last VMA disappears, the region map says how much
6276 	 * the reservation was and the page cache tells how much of
6277 	 * the reservation was consumed. Private mappings are per-VMA and
6278 	 * only the consumed reservations are tracked. When the VMA
6279 	 * disappears, the original reservation is the VMA size and the
6280 	 * consumed reservations are stored in the map. Hence, nothing
6281 	 * else has to be done for private mappings here
6282 	 */
6283 	if (!vma || vma->vm_flags & VM_MAYSHARE) {
6284 		add = region_add(resv_map, from, to, regions_needed, h, h_cg);
6285 
6286 		if (unlikely(add < 0)) {
6287 			hugetlb_acct_memory(h, -gbl_reserve);
6288 			goto out_put_pages;
6289 		} else if (unlikely(chg > add)) {
6290 			/*
6291 			 * pages in this range were added to the reserve
6292 			 * map between region_chg and region_add.  This
6293 			 * indicates a race with alloc_huge_page.  Adjust
6294 			 * the subpool and reserve counts modified above
6295 			 * based on the difference.
6296 			 */
6297 			long rsv_adjust;
6298 
6299 			/*
6300 			 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
6301 			 * reference to h_cg->css. See comment below for detail.
6302 			 */
6303 			hugetlb_cgroup_uncharge_cgroup_rsvd(
6304 				hstate_index(h),
6305 				(chg - add) * pages_per_huge_page(h), h_cg);
6306 
6307 			rsv_adjust = hugepage_subpool_put_pages(spool,
6308 								chg - add);
6309 			hugetlb_acct_memory(h, -rsv_adjust);
6310 		} else if (h_cg) {
6311 			/*
6312 			 * The file_regions will hold their own reference to
6313 			 * h_cg->css. So we should release the reference held
6314 			 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
6315 			 * done.
6316 			 */
6317 			hugetlb_cgroup_put_rsvd_cgroup(h_cg);
6318 		}
6319 	}
6320 	return true;
6321 
6322 out_put_pages:
6323 	/* put back original number of pages, chg */
6324 	(void)hugepage_subpool_put_pages(spool, chg);
6325 out_uncharge_cgroup:
6326 	hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
6327 					    chg * pages_per_huge_page(h), h_cg);
6328 out_err:
6329 	if (!vma || vma->vm_flags & VM_MAYSHARE)
6330 		/* Only call region_abort if the region_chg succeeded but the
6331 		 * region_add failed or didn't run.
6332 		 */
6333 		if (chg >= 0 && add < 0)
6334 			region_abort(resv_map, from, to, regions_needed);
6335 	if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
6336 		kref_put(&resv_map->refs, resv_map_release);
6337 	return false;
6338 }
6339 
6340 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
6341 								long freed)
6342 {
6343 	struct hstate *h = hstate_inode(inode);
6344 	struct resv_map *resv_map = inode_resv_map(inode);
6345 	long chg = 0;
6346 	struct hugepage_subpool *spool = subpool_inode(inode);
6347 	long gbl_reserve;
6348 
6349 	/*
6350 	 * Since this routine can be called in the evict inode path for all
6351 	 * hugetlbfs inodes, resv_map could be NULL.
6352 	 */
6353 	if (resv_map) {
6354 		chg = region_del(resv_map, start, end);
6355 		/*
6356 		 * region_del() can fail in the rare case where a region
6357 		 * must be split and another region descriptor can not be
6358 		 * allocated.  If end == LONG_MAX, it will not fail.
6359 		 */
6360 		if (chg < 0)
6361 			return chg;
6362 	}
6363 
6364 	spin_lock(&inode->i_lock);
6365 	inode->i_blocks -= (blocks_per_huge_page(h) * freed);
6366 	spin_unlock(&inode->i_lock);
6367 
6368 	/*
6369 	 * If the subpool has a minimum size, the number of global
6370 	 * reservations to be released may be adjusted.
6371 	 *
6372 	 * Note that !resv_map implies freed == 0. So (chg - freed)
6373 	 * won't go negative.
6374 	 */
6375 	gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
6376 	hugetlb_acct_memory(h, -gbl_reserve);
6377 
6378 	return 0;
6379 }
6380 
6381 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
6382 static unsigned long page_table_shareable(struct vm_area_struct *svma,
6383 				struct vm_area_struct *vma,
6384 				unsigned long addr, pgoff_t idx)
6385 {
6386 	unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
6387 				svma->vm_start;
6388 	unsigned long sbase = saddr & PUD_MASK;
6389 	unsigned long s_end = sbase + PUD_SIZE;
6390 
6391 	/* Allow segments to share if only one is marked locked */
6392 	unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
6393 	unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
6394 
6395 	/*
6396 	 * match the virtual addresses, permission and the alignment of the
6397 	 * page table page.
6398 	 */
6399 	if (pmd_index(addr) != pmd_index(saddr) ||
6400 	    vm_flags != svm_flags ||
6401 	    !range_in_vma(svma, sbase, s_end))
6402 		return 0;
6403 
6404 	return saddr;
6405 }
6406 
6407 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
6408 {
6409 	unsigned long base = addr & PUD_MASK;
6410 	unsigned long end = base + PUD_SIZE;
6411 
6412 	/*
6413 	 * check on proper vm_flags and page table alignment
6414 	 */
6415 	if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
6416 		return true;
6417 	return false;
6418 }
6419 
6420 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6421 {
6422 #ifdef CONFIG_USERFAULTFD
6423 	if (uffd_disable_huge_pmd_share(vma))
6424 		return false;
6425 #endif
6426 	return vma_shareable(vma, addr);
6427 }
6428 
6429 /*
6430  * Determine if start,end range within vma could be mapped by shared pmd.
6431  * If yes, adjust start and end to cover range associated with possible
6432  * shared pmd mappings.
6433  */
6434 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
6435 				unsigned long *start, unsigned long *end)
6436 {
6437 	unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
6438 		v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6439 
6440 	/*
6441 	 * vma needs to span at least one aligned PUD size, and the range
6442 	 * must be at least partially within in.
6443 	 */
6444 	if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
6445 		(*end <= v_start) || (*start >= v_end))
6446 		return;
6447 
6448 	/* Extend the range to be PUD aligned for a worst case scenario */
6449 	if (*start > v_start)
6450 		*start = ALIGN_DOWN(*start, PUD_SIZE);
6451 
6452 	if (*end < v_end)
6453 		*end = ALIGN(*end, PUD_SIZE);
6454 }
6455 
6456 /*
6457  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
6458  * and returns the corresponding pte. While this is not necessary for the
6459  * !shared pmd case because we can allocate the pmd later as well, it makes the
6460  * code much cleaner.
6461  *
6462  * This routine must be called with i_mmap_rwsem held in at least read mode if
6463  * sharing is possible.  For hugetlbfs, this prevents removal of any page
6464  * table entries associated with the address space.  This is important as we
6465  * are setting up sharing based on existing page table entries (mappings).
6466  */
6467 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
6468 		      unsigned long addr, pud_t *pud)
6469 {
6470 	struct address_space *mapping = vma->vm_file->f_mapping;
6471 	pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
6472 			vma->vm_pgoff;
6473 	struct vm_area_struct *svma;
6474 	unsigned long saddr;
6475 	pte_t *spte = NULL;
6476 	pte_t *pte;
6477 	spinlock_t *ptl;
6478 
6479 	i_mmap_assert_locked(mapping);
6480 	vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
6481 		if (svma == vma)
6482 			continue;
6483 
6484 		saddr = page_table_shareable(svma, vma, addr, idx);
6485 		if (saddr) {
6486 			spte = huge_pte_offset(svma->vm_mm, saddr,
6487 					       vma_mmu_pagesize(svma));
6488 			if (spte) {
6489 				get_page(virt_to_page(spte));
6490 				break;
6491 			}
6492 		}
6493 	}
6494 
6495 	if (!spte)
6496 		goto out;
6497 
6498 	ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
6499 	if (pud_none(*pud)) {
6500 		pud_populate(mm, pud,
6501 				(pmd_t *)((unsigned long)spte & PAGE_MASK));
6502 		mm_inc_nr_pmds(mm);
6503 	} else {
6504 		put_page(virt_to_page(spte));
6505 	}
6506 	spin_unlock(ptl);
6507 out:
6508 	pte = (pte_t *)pmd_alloc(mm, pud, addr);
6509 	return pte;
6510 }
6511 
6512 /*
6513  * unmap huge page backed by shared pte.
6514  *
6515  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
6516  * indicated by page_count > 1, unmap is achieved by clearing pud and
6517  * decrementing the ref count. If count == 1, the pte page is not shared.
6518  *
6519  * Called with page table lock held and i_mmap_rwsem held in write mode.
6520  *
6521  * returns: 1 successfully unmapped a shared pte page
6522  *	    0 the underlying pte page is not shared, or it is the last user
6523  */
6524 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
6525 					unsigned long *addr, pte_t *ptep)
6526 {
6527 	pgd_t *pgd = pgd_offset(mm, *addr);
6528 	p4d_t *p4d = p4d_offset(pgd, *addr);
6529 	pud_t *pud = pud_offset(p4d, *addr);
6530 
6531 	i_mmap_assert_write_locked(vma->vm_file->f_mapping);
6532 	BUG_ON(page_count(virt_to_page(ptep)) == 0);
6533 	if (page_count(virt_to_page(ptep)) == 1)
6534 		return 0;
6535 
6536 	pud_clear(pud);
6537 	put_page(virt_to_page(ptep));
6538 	mm_dec_nr_pmds(mm);
6539 	*addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
6540 	return 1;
6541 }
6542 
6543 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
6544 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
6545 		      unsigned long addr, pud_t *pud)
6546 {
6547 	return NULL;
6548 }
6549 
6550 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
6551 				unsigned long *addr, pte_t *ptep)
6552 {
6553 	return 0;
6554 }
6555 
6556 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
6557 				unsigned long *start, unsigned long *end)
6558 {
6559 }
6560 
6561 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6562 {
6563 	return false;
6564 }
6565 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
6566 
6567 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
6568 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
6569 			unsigned long addr, unsigned long sz)
6570 {
6571 	pgd_t *pgd;
6572 	p4d_t *p4d;
6573 	pud_t *pud;
6574 	pte_t *pte = NULL;
6575 
6576 	pgd = pgd_offset(mm, addr);
6577 	p4d = p4d_alloc(mm, pgd, addr);
6578 	if (!p4d)
6579 		return NULL;
6580 	pud = pud_alloc(mm, p4d, addr);
6581 	if (pud) {
6582 		if (sz == PUD_SIZE) {
6583 			pte = (pte_t *)pud;
6584 		} else {
6585 			BUG_ON(sz != PMD_SIZE);
6586 			if (want_pmd_share(vma, addr) && pud_none(*pud))
6587 				pte = huge_pmd_share(mm, vma, addr, pud);
6588 			else
6589 				pte = (pte_t *)pmd_alloc(mm, pud, addr);
6590 		}
6591 	}
6592 	BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
6593 
6594 	return pte;
6595 }
6596 
6597 /*
6598  * huge_pte_offset() - Walk the page table to resolve the hugepage
6599  * entry at address @addr
6600  *
6601  * Return: Pointer to page table entry (PUD or PMD) for
6602  * address @addr, or NULL if a !p*d_present() entry is encountered and the
6603  * size @sz doesn't match the hugepage size at this level of the page
6604  * table.
6605  */
6606 pte_t *huge_pte_offset(struct mm_struct *mm,
6607 		       unsigned long addr, unsigned long sz)
6608 {
6609 	pgd_t *pgd;
6610 	p4d_t *p4d;
6611 	pud_t *pud;
6612 	pmd_t *pmd;
6613 
6614 	pgd = pgd_offset(mm, addr);
6615 	if (!pgd_present(*pgd))
6616 		return NULL;
6617 	p4d = p4d_offset(pgd, addr);
6618 	if (!p4d_present(*p4d))
6619 		return NULL;
6620 
6621 	pud = pud_offset(p4d, addr);
6622 	if (sz == PUD_SIZE)
6623 		/* must be pud huge, non-present or none */
6624 		return (pte_t *)pud;
6625 	if (!pud_present(*pud))
6626 		return NULL;
6627 	/* must have a valid entry and size to go further */
6628 
6629 	pmd = pmd_offset(pud, addr);
6630 	/* must be pmd huge, non-present or none */
6631 	return (pte_t *)pmd;
6632 }
6633 
6634 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
6635 
6636 /*
6637  * These functions are overwritable if your architecture needs its own
6638  * behavior.
6639  */
6640 struct page * __weak
6641 follow_huge_addr(struct mm_struct *mm, unsigned long address,
6642 			      int write)
6643 {
6644 	return ERR_PTR(-EINVAL);
6645 }
6646 
6647 struct page * __weak
6648 follow_huge_pd(struct vm_area_struct *vma,
6649 	       unsigned long address, hugepd_t hpd, int flags, int pdshift)
6650 {
6651 	WARN(1, "hugepd follow called with no support for hugepage directory format\n");
6652 	return NULL;
6653 }
6654 
6655 struct page * __weak
6656 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
6657 		pmd_t *pmd, int flags)
6658 {
6659 	struct page *page = NULL;
6660 	spinlock_t *ptl;
6661 	pte_t pte;
6662 
6663 	/* FOLL_GET and FOLL_PIN are mutually exclusive. */
6664 	if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
6665 			 (FOLL_PIN | FOLL_GET)))
6666 		return NULL;
6667 
6668 retry:
6669 	ptl = pmd_lockptr(mm, pmd);
6670 	spin_lock(ptl);
6671 	/*
6672 	 * make sure that the address range covered by this pmd is not
6673 	 * unmapped from other threads.
6674 	 */
6675 	if (!pmd_huge(*pmd))
6676 		goto out;
6677 	pte = huge_ptep_get((pte_t *)pmd);
6678 	if (pte_present(pte)) {
6679 		page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
6680 		/*
6681 		 * try_grab_page() should always succeed here, because: a) we
6682 		 * hold the pmd (ptl) lock, and b) we've just checked that the
6683 		 * huge pmd (head) page is present in the page tables. The ptl
6684 		 * prevents the head page and tail pages from being rearranged
6685 		 * in any way. So this page must be available at this point,
6686 		 * unless the page refcount overflowed:
6687 		 */
6688 		if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
6689 			page = NULL;
6690 			goto out;
6691 		}
6692 	} else {
6693 		if (is_hugetlb_entry_migration(pte)) {
6694 			spin_unlock(ptl);
6695 			__migration_entry_wait(mm, (pte_t *)pmd, ptl);
6696 			goto retry;
6697 		}
6698 		/*
6699 		 * hwpoisoned entry is treated as no_page_table in
6700 		 * follow_page_mask().
6701 		 */
6702 	}
6703 out:
6704 	spin_unlock(ptl);
6705 	return page;
6706 }
6707 
6708 struct page * __weak
6709 follow_huge_pud(struct mm_struct *mm, unsigned long address,
6710 		pud_t *pud, int flags)
6711 {
6712 	if (flags & (FOLL_GET | FOLL_PIN))
6713 		return NULL;
6714 
6715 	return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
6716 }
6717 
6718 struct page * __weak
6719 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
6720 {
6721 	if (flags & (FOLL_GET | FOLL_PIN))
6722 		return NULL;
6723 
6724 	return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
6725 }
6726 
6727 bool isolate_huge_page(struct page *page, struct list_head *list)
6728 {
6729 	bool ret = true;
6730 
6731 	spin_lock_irq(&hugetlb_lock);
6732 	if (!PageHeadHuge(page) ||
6733 	    !HPageMigratable(page) ||
6734 	    !get_page_unless_zero(page)) {
6735 		ret = false;
6736 		goto unlock;
6737 	}
6738 	ClearHPageMigratable(page);
6739 	list_move_tail(&page->lru, list);
6740 unlock:
6741 	spin_unlock_irq(&hugetlb_lock);
6742 	return ret;
6743 }
6744 
6745 int get_hwpoison_huge_page(struct page *page, bool *hugetlb)
6746 {
6747 	int ret = 0;
6748 
6749 	*hugetlb = false;
6750 	spin_lock_irq(&hugetlb_lock);
6751 	if (PageHeadHuge(page)) {
6752 		*hugetlb = true;
6753 		if (HPageFreed(page) || HPageMigratable(page))
6754 			ret = get_page_unless_zero(page);
6755 		else
6756 			ret = -EBUSY;
6757 	}
6758 	spin_unlock_irq(&hugetlb_lock);
6759 	return ret;
6760 }
6761 
6762 void putback_active_hugepage(struct page *page)
6763 {
6764 	spin_lock_irq(&hugetlb_lock);
6765 	SetHPageMigratable(page);
6766 	list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
6767 	spin_unlock_irq(&hugetlb_lock);
6768 	put_page(page);
6769 }
6770 
6771 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
6772 {
6773 	struct hstate *h = page_hstate(oldpage);
6774 
6775 	hugetlb_cgroup_migrate(oldpage, newpage);
6776 	set_page_owner_migrate_reason(newpage, reason);
6777 
6778 	/*
6779 	 * transfer temporary state of the new huge page. This is
6780 	 * reverse to other transitions because the newpage is going to
6781 	 * be final while the old one will be freed so it takes over
6782 	 * the temporary status.
6783 	 *
6784 	 * Also note that we have to transfer the per-node surplus state
6785 	 * here as well otherwise the global surplus count will not match
6786 	 * the per-node's.
6787 	 */
6788 	if (HPageTemporary(newpage)) {
6789 		int old_nid = page_to_nid(oldpage);
6790 		int new_nid = page_to_nid(newpage);
6791 
6792 		SetHPageTemporary(oldpage);
6793 		ClearHPageTemporary(newpage);
6794 
6795 		/*
6796 		 * There is no need to transfer the per-node surplus state
6797 		 * when we do not cross the node.
6798 		 */
6799 		if (new_nid == old_nid)
6800 			return;
6801 		spin_lock_irq(&hugetlb_lock);
6802 		if (h->surplus_huge_pages_node[old_nid]) {
6803 			h->surplus_huge_pages_node[old_nid]--;
6804 			h->surplus_huge_pages_node[new_nid]++;
6805 		}
6806 		spin_unlock_irq(&hugetlb_lock);
6807 	}
6808 }
6809 
6810 /*
6811  * This function will unconditionally remove all the shared pmd pgtable entries
6812  * within the specific vma for a hugetlbfs memory range.
6813  */
6814 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
6815 {
6816 	struct hstate *h = hstate_vma(vma);
6817 	unsigned long sz = huge_page_size(h);
6818 	struct mm_struct *mm = vma->vm_mm;
6819 	struct mmu_notifier_range range;
6820 	unsigned long address, start, end;
6821 	spinlock_t *ptl;
6822 	pte_t *ptep;
6823 
6824 	if (!(vma->vm_flags & VM_MAYSHARE))
6825 		return;
6826 
6827 	start = ALIGN(vma->vm_start, PUD_SIZE);
6828 	end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6829 
6830 	if (start >= end)
6831 		return;
6832 
6833 	/*
6834 	 * No need to call adjust_range_if_pmd_sharing_possible(), because
6835 	 * we have already done the PUD_SIZE alignment.
6836 	 */
6837 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
6838 				start, end);
6839 	mmu_notifier_invalidate_range_start(&range);
6840 	i_mmap_lock_write(vma->vm_file->f_mapping);
6841 	for (address = start; address < end; address += PUD_SIZE) {
6842 		unsigned long tmp = address;
6843 
6844 		ptep = huge_pte_offset(mm, address, sz);
6845 		if (!ptep)
6846 			continue;
6847 		ptl = huge_pte_lock(h, mm, ptep);
6848 		/* We don't want 'address' to be changed */
6849 		huge_pmd_unshare(mm, vma, &tmp, ptep);
6850 		spin_unlock(ptl);
6851 	}
6852 	flush_hugetlb_tlb_range(vma, start, end);
6853 	i_mmap_unlock_write(vma->vm_file->f_mapping);
6854 	/*
6855 	 * No need to call mmu_notifier_invalidate_range(), see
6856 	 * Documentation/vm/mmu_notifier.rst.
6857 	 */
6858 	mmu_notifier_invalidate_range_end(&range);
6859 }
6860 
6861 #ifdef CONFIG_CMA
6862 static bool cma_reserve_called __initdata;
6863 
6864 static int __init cmdline_parse_hugetlb_cma(char *p)
6865 {
6866 	int nid, count = 0;
6867 	unsigned long tmp;
6868 	char *s = p;
6869 
6870 	while (*s) {
6871 		if (sscanf(s, "%lu%n", &tmp, &count) != 1)
6872 			break;
6873 
6874 		if (s[count] == ':') {
6875 			nid = tmp;
6876 			if (nid < 0 || nid >= MAX_NUMNODES)
6877 				break;
6878 
6879 			s += count + 1;
6880 			tmp = memparse(s, &s);
6881 			hugetlb_cma_size_in_node[nid] = tmp;
6882 			hugetlb_cma_size += tmp;
6883 
6884 			/*
6885 			 * Skip the separator if have one, otherwise
6886 			 * break the parsing.
6887 			 */
6888 			if (*s == ',')
6889 				s++;
6890 			else
6891 				break;
6892 		} else {
6893 			hugetlb_cma_size = memparse(p, &p);
6894 			break;
6895 		}
6896 	}
6897 
6898 	return 0;
6899 }
6900 
6901 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
6902 
6903 void __init hugetlb_cma_reserve(int order)
6904 {
6905 	unsigned long size, reserved, per_node;
6906 	bool node_specific_cma_alloc = false;
6907 	int nid;
6908 
6909 	cma_reserve_called = true;
6910 
6911 	if (!hugetlb_cma_size)
6912 		return;
6913 
6914 	for (nid = 0; nid < MAX_NUMNODES; nid++) {
6915 		if (hugetlb_cma_size_in_node[nid] == 0)
6916 			continue;
6917 
6918 		if (!node_state(nid, N_ONLINE)) {
6919 			pr_warn("hugetlb_cma: invalid node %d specified\n", nid);
6920 			hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
6921 			hugetlb_cma_size_in_node[nid] = 0;
6922 			continue;
6923 		}
6924 
6925 		if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) {
6926 			pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n",
6927 				nid, (PAGE_SIZE << order) / SZ_1M);
6928 			hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
6929 			hugetlb_cma_size_in_node[nid] = 0;
6930 		} else {
6931 			node_specific_cma_alloc = true;
6932 		}
6933 	}
6934 
6935 	/* Validate the CMA size again in case some invalid nodes specified. */
6936 	if (!hugetlb_cma_size)
6937 		return;
6938 
6939 	if (hugetlb_cma_size < (PAGE_SIZE << order)) {
6940 		pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
6941 			(PAGE_SIZE << order) / SZ_1M);
6942 		hugetlb_cma_size = 0;
6943 		return;
6944 	}
6945 
6946 	if (!node_specific_cma_alloc) {
6947 		/*
6948 		 * If 3 GB area is requested on a machine with 4 numa nodes,
6949 		 * let's allocate 1 GB on first three nodes and ignore the last one.
6950 		 */
6951 		per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
6952 		pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
6953 			hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
6954 	}
6955 
6956 	reserved = 0;
6957 	for_each_node_state(nid, N_ONLINE) {
6958 		int res;
6959 		char name[CMA_MAX_NAME];
6960 
6961 		if (node_specific_cma_alloc) {
6962 			if (hugetlb_cma_size_in_node[nid] == 0)
6963 				continue;
6964 
6965 			size = hugetlb_cma_size_in_node[nid];
6966 		} else {
6967 			size = min(per_node, hugetlb_cma_size - reserved);
6968 		}
6969 
6970 		size = round_up(size, PAGE_SIZE << order);
6971 
6972 		snprintf(name, sizeof(name), "hugetlb%d", nid);
6973 		/*
6974 		 * Note that 'order per bit' is based on smallest size that
6975 		 * may be returned to CMA allocator in the case of
6976 		 * huge page demotion.
6977 		 */
6978 		res = cma_declare_contiguous_nid(0, size, 0,
6979 						PAGE_SIZE << HUGETLB_PAGE_ORDER,
6980 						 0, false, name,
6981 						 &hugetlb_cma[nid], nid);
6982 		if (res) {
6983 			pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
6984 				res, nid);
6985 			continue;
6986 		}
6987 
6988 		reserved += size;
6989 		pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
6990 			size / SZ_1M, nid);
6991 
6992 		if (reserved >= hugetlb_cma_size)
6993 			break;
6994 	}
6995 
6996 	if (!reserved)
6997 		/*
6998 		 * hugetlb_cma_size is used to determine if allocations from
6999 		 * cma are possible.  Set to zero if no cma regions are set up.
7000 		 */
7001 		hugetlb_cma_size = 0;
7002 }
7003 
7004 void __init hugetlb_cma_check(void)
7005 {
7006 	if (!hugetlb_cma_size || cma_reserve_called)
7007 		return;
7008 
7009 	pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
7010 }
7011 
7012 #endif /* CONFIG_CMA */
7013