xref: /openbmc/linux/mm/hugetlb.c (revision 9fb29c73)
1 /*
2  * Generic hugetlb support.
3  * (C) Nadia Yvette Chambers, April 2004
4  */
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/mm.h>
8 #include <linux/seq_file.h>
9 #include <linux/sysctl.h>
10 #include <linux/highmem.h>
11 #include <linux/mmu_notifier.h>
12 #include <linux/nodemask.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/compiler.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/memblock.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/mmdebug.h>
22 #include <linux/sched/signal.h>
23 #include <linux/rmap.h>
24 #include <linux/string_helpers.h>
25 #include <linux/swap.h>
26 #include <linux/swapops.h>
27 #include <linux/jhash.h>
28 
29 #include <asm/page.h>
30 #include <asm/pgtable.h>
31 #include <asm/tlb.h>
32 
33 #include <linux/io.h>
34 #include <linux/hugetlb.h>
35 #include <linux/hugetlb_cgroup.h>
36 #include <linux/node.h>
37 #include <linux/userfaultfd_k.h>
38 #include <linux/page_owner.h>
39 #include "internal.h"
40 
41 int hugetlb_max_hstate __read_mostly;
42 unsigned int default_hstate_idx;
43 struct hstate hstates[HUGE_MAX_HSTATE];
44 /*
45  * Minimum page order among possible hugepage sizes, set to a proper value
46  * at boot time.
47  */
48 static unsigned int minimum_order __read_mostly = UINT_MAX;
49 
50 __initdata LIST_HEAD(huge_boot_pages);
51 
52 /* for command line parsing */
53 static struct hstate * __initdata parsed_hstate;
54 static unsigned long __initdata default_hstate_max_huge_pages;
55 static unsigned long __initdata default_hstate_size;
56 static bool __initdata parsed_valid_hugepagesz = true;
57 
58 /*
59  * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
60  * free_huge_pages, and surplus_huge_pages.
61  */
62 DEFINE_SPINLOCK(hugetlb_lock);
63 
64 /*
65  * Serializes faults on the same logical page.  This is used to
66  * prevent spurious OOMs when the hugepage pool is fully utilized.
67  */
68 static int num_fault_mutexes;
69 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
70 
71 /* Forward declaration */
72 static int hugetlb_acct_memory(struct hstate *h, long delta);
73 
74 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
75 {
76 	bool free = (spool->count == 0) && (spool->used_hpages == 0);
77 
78 	spin_unlock(&spool->lock);
79 
80 	/* If no pages are used, and no other handles to the subpool
81 	 * remain, give up any reservations mased on minimum size and
82 	 * free the subpool */
83 	if (free) {
84 		if (spool->min_hpages != -1)
85 			hugetlb_acct_memory(spool->hstate,
86 						-spool->min_hpages);
87 		kfree(spool);
88 	}
89 }
90 
91 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
92 						long min_hpages)
93 {
94 	struct hugepage_subpool *spool;
95 
96 	spool = kzalloc(sizeof(*spool), GFP_KERNEL);
97 	if (!spool)
98 		return NULL;
99 
100 	spin_lock_init(&spool->lock);
101 	spool->count = 1;
102 	spool->max_hpages = max_hpages;
103 	spool->hstate = h;
104 	spool->min_hpages = min_hpages;
105 
106 	if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
107 		kfree(spool);
108 		return NULL;
109 	}
110 	spool->rsv_hpages = min_hpages;
111 
112 	return spool;
113 }
114 
115 void hugepage_put_subpool(struct hugepage_subpool *spool)
116 {
117 	spin_lock(&spool->lock);
118 	BUG_ON(!spool->count);
119 	spool->count--;
120 	unlock_or_release_subpool(spool);
121 }
122 
123 /*
124  * Subpool accounting for allocating and reserving pages.
125  * Return -ENOMEM if there are not enough resources to satisfy the
126  * the request.  Otherwise, return the number of pages by which the
127  * global pools must be adjusted (upward).  The returned value may
128  * only be different than the passed value (delta) in the case where
129  * a subpool minimum size must be manitained.
130  */
131 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
132 				      long delta)
133 {
134 	long ret = delta;
135 
136 	if (!spool)
137 		return ret;
138 
139 	spin_lock(&spool->lock);
140 
141 	if (spool->max_hpages != -1) {		/* maximum size accounting */
142 		if ((spool->used_hpages + delta) <= spool->max_hpages)
143 			spool->used_hpages += delta;
144 		else {
145 			ret = -ENOMEM;
146 			goto unlock_ret;
147 		}
148 	}
149 
150 	/* minimum size accounting */
151 	if (spool->min_hpages != -1 && spool->rsv_hpages) {
152 		if (delta > spool->rsv_hpages) {
153 			/*
154 			 * Asking for more reserves than those already taken on
155 			 * behalf of subpool.  Return difference.
156 			 */
157 			ret = delta - spool->rsv_hpages;
158 			spool->rsv_hpages = 0;
159 		} else {
160 			ret = 0;	/* reserves already accounted for */
161 			spool->rsv_hpages -= delta;
162 		}
163 	}
164 
165 unlock_ret:
166 	spin_unlock(&spool->lock);
167 	return ret;
168 }
169 
170 /*
171  * Subpool accounting for freeing and unreserving pages.
172  * Return the number of global page reservations that must be dropped.
173  * The return value may only be different than the passed value (delta)
174  * in the case where a subpool minimum size must be maintained.
175  */
176 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
177 				       long delta)
178 {
179 	long ret = delta;
180 
181 	if (!spool)
182 		return delta;
183 
184 	spin_lock(&spool->lock);
185 
186 	if (spool->max_hpages != -1)		/* maximum size accounting */
187 		spool->used_hpages -= delta;
188 
189 	 /* minimum size accounting */
190 	if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
191 		if (spool->rsv_hpages + delta <= spool->min_hpages)
192 			ret = 0;
193 		else
194 			ret = spool->rsv_hpages + delta - spool->min_hpages;
195 
196 		spool->rsv_hpages += delta;
197 		if (spool->rsv_hpages > spool->min_hpages)
198 			spool->rsv_hpages = spool->min_hpages;
199 	}
200 
201 	/*
202 	 * If hugetlbfs_put_super couldn't free spool due to an outstanding
203 	 * quota reference, free it now.
204 	 */
205 	unlock_or_release_subpool(spool);
206 
207 	return ret;
208 }
209 
210 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
211 {
212 	return HUGETLBFS_SB(inode->i_sb)->spool;
213 }
214 
215 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
216 {
217 	return subpool_inode(file_inode(vma->vm_file));
218 }
219 
220 /*
221  * Region tracking -- allows tracking of reservations and instantiated pages
222  *                    across the pages in a mapping.
223  *
224  * The region data structures are embedded into a resv_map and protected
225  * by a resv_map's lock.  The set of regions within the resv_map represent
226  * reservations for huge pages, or huge pages that have already been
227  * instantiated within the map.  The from and to elements are huge page
228  * indicies into the associated mapping.  from indicates the starting index
229  * of the region.  to represents the first index past the end of  the region.
230  *
231  * For example, a file region structure with from == 0 and to == 4 represents
232  * four huge pages in a mapping.  It is important to note that the to element
233  * represents the first element past the end of the region. This is used in
234  * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
235  *
236  * Interval notation of the form [from, to) will be used to indicate that
237  * the endpoint from is inclusive and to is exclusive.
238  */
239 struct file_region {
240 	struct list_head link;
241 	long from;
242 	long to;
243 };
244 
245 /*
246  * Add the huge page range represented by [f, t) to the reserve
247  * map.  In the normal case, existing regions will be expanded
248  * to accommodate the specified range.  Sufficient regions should
249  * exist for expansion due to the previous call to region_chg
250  * with the same range.  However, it is possible that region_del
251  * could have been called after region_chg and modifed the map
252  * in such a way that no region exists to be expanded.  In this
253  * case, pull a region descriptor from the cache associated with
254  * the map and use that for the new range.
255  *
256  * Return the number of new huge pages added to the map.  This
257  * number is greater than or equal to zero.
258  */
259 static long region_add(struct resv_map *resv, long f, long t)
260 {
261 	struct list_head *head = &resv->regions;
262 	struct file_region *rg, *nrg, *trg;
263 	long add = 0;
264 
265 	spin_lock(&resv->lock);
266 	/* Locate the region we are either in or before. */
267 	list_for_each_entry(rg, head, link)
268 		if (f <= rg->to)
269 			break;
270 
271 	/*
272 	 * If no region exists which can be expanded to include the
273 	 * specified range, the list must have been modified by an
274 	 * interleving call to region_del().  Pull a region descriptor
275 	 * from the cache and use it for this range.
276 	 */
277 	if (&rg->link == head || t < rg->from) {
278 		VM_BUG_ON(resv->region_cache_count <= 0);
279 
280 		resv->region_cache_count--;
281 		nrg = list_first_entry(&resv->region_cache, struct file_region,
282 					link);
283 		list_del(&nrg->link);
284 
285 		nrg->from = f;
286 		nrg->to = t;
287 		list_add(&nrg->link, rg->link.prev);
288 
289 		add += t - f;
290 		goto out_locked;
291 	}
292 
293 	/* Round our left edge to the current segment if it encloses us. */
294 	if (f > rg->from)
295 		f = rg->from;
296 
297 	/* Check for and consume any regions we now overlap with. */
298 	nrg = rg;
299 	list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
300 		if (&rg->link == head)
301 			break;
302 		if (rg->from > t)
303 			break;
304 
305 		/* If this area reaches higher then extend our area to
306 		 * include it completely.  If this is not the first area
307 		 * which we intend to reuse, free it. */
308 		if (rg->to > t)
309 			t = rg->to;
310 		if (rg != nrg) {
311 			/* Decrement return value by the deleted range.
312 			 * Another range will span this area so that by
313 			 * end of routine add will be >= zero
314 			 */
315 			add -= (rg->to - rg->from);
316 			list_del(&rg->link);
317 			kfree(rg);
318 		}
319 	}
320 
321 	add += (nrg->from - f);		/* Added to beginning of region */
322 	nrg->from = f;
323 	add += t - nrg->to;		/* Added to end of region */
324 	nrg->to = t;
325 
326 out_locked:
327 	resv->adds_in_progress--;
328 	spin_unlock(&resv->lock);
329 	VM_BUG_ON(add < 0);
330 	return add;
331 }
332 
333 /*
334  * Examine the existing reserve map and determine how many
335  * huge pages in the specified range [f, t) are NOT currently
336  * represented.  This routine is called before a subsequent
337  * call to region_add that will actually modify the reserve
338  * map to add the specified range [f, t).  region_chg does
339  * not change the number of huge pages represented by the
340  * map.  However, if the existing regions in the map can not
341  * be expanded to represent the new range, a new file_region
342  * structure is added to the map as a placeholder.  This is
343  * so that the subsequent region_add call will have all the
344  * regions it needs and will not fail.
345  *
346  * Upon entry, region_chg will also examine the cache of region descriptors
347  * associated with the map.  If there are not enough descriptors cached, one
348  * will be allocated for the in progress add operation.
349  *
350  * Returns the number of huge pages that need to be added to the existing
351  * reservation map for the range [f, t).  This number is greater or equal to
352  * zero.  -ENOMEM is returned if a new file_region structure or cache entry
353  * is needed and can not be allocated.
354  */
355 static long region_chg(struct resv_map *resv, long f, long t)
356 {
357 	struct list_head *head = &resv->regions;
358 	struct file_region *rg, *nrg = NULL;
359 	long chg = 0;
360 
361 retry:
362 	spin_lock(&resv->lock);
363 retry_locked:
364 	resv->adds_in_progress++;
365 
366 	/*
367 	 * Check for sufficient descriptors in the cache to accommodate
368 	 * the number of in progress add operations.
369 	 */
370 	if (resv->adds_in_progress > resv->region_cache_count) {
371 		struct file_region *trg;
372 
373 		VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
374 		/* Must drop lock to allocate a new descriptor. */
375 		resv->adds_in_progress--;
376 		spin_unlock(&resv->lock);
377 
378 		trg = kmalloc(sizeof(*trg), GFP_KERNEL);
379 		if (!trg) {
380 			kfree(nrg);
381 			return -ENOMEM;
382 		}
383 
384 		spin_lock(&resv->lock);
385 		list_add(&trg->link, &resv->region_cache);
386 		resv->region_cache_count++;
387 		goto retry_locked;
388 	}
389 
390 	/* Locate the region we are before or in. */
391 	list_for_each_entry(rg, head, link)
392 		if (f <= rg->to)
393 			break;
394 
395 	/* If we are below the current region then a new region is required.
396 	 * Subtle, allocate a new region at the position but make it zero
397 	 * size such that we can guarantee to record the reservation. */
398 	if (&rg->link == head || t < rg->from) {
399 		if (!nrg) {
400 			resv->adds_in_progress--;
401 			spin_unlock(&resv->lock);
402 			nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
403 			if (!nrg)
404 				return -ENOMEM;
405 
406 			nrg->from = f;
407 			nrg->to   = f;
408 			INIT_LIST_HEAD(&nrg->link);
409 			goto retry;
410 		}
411 
412 		list_add(&nrg->link, rg->link.prev);
413 		chg = t - f;
414 		goto out_nrg;
415 	}
416 
417 	/* Round our left edge to the current segment if it encloses us. */
418 	if (f > rg->from)
419 		f = rg->from;
420 	chg = t - f;
421 
422 	/* Check for and consume any regions we now overlap with. */
423 	list_for_each_entry(rg, rg->link.prev, link) {
424 		if (&rg->link == head)
425 			break;
426 		if (rg->from > t)
427 			goto out;
428 
429 		/* We overlap with this area, if it extends further than
430 		 * us then we must extend ourselves.  Account for its
431 		 * existing reservation. */
432 		if (rg->to > t) {
433 			chg += rg->to - t;
434 			t = rg->to;
435 		}
436 		chg -= rg->to - rg->from;
437 	}
438 
439 out:
440 	spin_unlock(&resv->lock);
441 	/*  We already know we raced and no longer need the new region */
442 	kfree(nrg);
443 	return chg;
444 out_nrg:
445 	spin_unlock(&resv->lock);
446 	return chg;
447 }
448 
449 /*
450  * Abort the in progress add operation.  The adds_in_progress field
451  * of the resv_map keeps track of the operations in progress between
452  * calls to region_chg and region_add.  Operations are sometimes
453  * aborted after the call to region_chg.  In such cases, region_abort
454  * is called to decrement the adds_in_progress counter.
455  *
456  * NOTE: The range arguments [f, t) are not needed or used in this
457  * routine.  They are kept to make reading the calling code easier as
458  * arguments will match the associated region_chg call.
459  */
460 static void region_abort(struct resv_map *resv, long f, long t)
461 {
462 	spin_lock(&resv->lock);
463 	VM_BUG_ON(!resv->region_cache_count);
464 	resv->adds_in_progress--;
465 	spin_unlock(&resv->lock);
466 }
467 
468 /*
469  * Delete the specified range [f, t) from the reserve map.  If the
470  * t parameter is LONG_MAX, this indicates that ALL regions after f
471  * should be deleted.  Locate the regions which intersect [f, t)
472  * and either trim, delete or split the existing regions.
473  *
474  * Returns the number of huge pages deleted from the reserve map.
475  * In the normal case, the return value is zero or more.  In the
476  * case where a region must be split, a new region descriptor must
477  * be allocated.  If the allocation fails, -ENOMEM will be returned.
478  * NOTE: If the parameter t == LONG_MAX, then we will never split
479  * a region and possibly return -ENOMEM.  Callers specifying
480  * t == LONG_MAX do not need to check for -ENOMEM error.
481  */
482 static long region_del(struct resv_map *resv, long f, long t)
483 {
484 	struct list_head *head = &resv->regions;
485 	struct file_region *rg, *trg;
486 	struct file_region *nrg = NULL;
487 	long del = 0;
488 
489 retry:
490 	spin_lock(&resv->lock);
491 	list_for_each_entry_safe(rg, trg, head, link) {
492 		/*
493 		 * Skip regions before the range to be deleted.  file_region
494 		 * ranges are normally of the form [from, to).  However, there
495 		 * may be a "placeholder" entry in the map which is of the form
496 		 * (from, to) with from == to.  Check for placeholder entries
497 		 * at the beginning of the range to be deleted.
498 		 */
499 		if (rg->to <= f && (rg->to != rg->from || rg->to != f))
500 			continue;
501 
502 		if (rg->from >= t)
503 			break;
504 
505 		if (f > rg->from && t < rg->to) { /* Must split region */
506 			/*
507 			 * Check for an entry in the cache before dropping
508 			 * lock and attempting allocation.
509 			 */
510 			if (!nrg &&
511 			    resv->region_cache_count > resv->adds_in_progress) {
512 				nrg = list_first_entry(&resv->region_cache,
513 							struct file_region,
514 							link);
515 				list_del(&nrg->link);
516 				resv->region_cache_count--;
517 			}
518 
519 			if (!nrg) {
520 				spin_unlock(&resv->lock);
521 				nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
522 				if (!nrg)
523 					return -ENOMEM;
524 				goto retry;
525 			}
526 
527 			del += t - f;
528 
529 			/* New entry for end of split region */
530 			nrg->from = t;
531 			nrg->to = rg->to;
532 			INIT_LIST_HEAD(&nrg->link);
533 
534 			/* Original entry is trimmed */
535 			rg->to = f;
536 
537 			list_add(&nrg->link, &rg->link);
538 			nrg = NULL;
539 			break;
540 		}
541 
542 		if (f <= rg->from && t >= rg->to) { /* Remove entire region */
543 			del += rg->to - rg->from;
544 			list_del(&rg->link);
545 			kfree(rg);
546 			continue;
547 		}
548 
549 		if (f <= rg->from) {	/* Trim beginning of region */
550 			del += t - rg->from;
551 			rg->from = t;
552 		} else {		/* Trim end of region */
553 			del += rg->to - f;
554 			rg->to = f;
555 		}
556 	}
557 
558 	spin_unlock(&resv->lock);
559 	kfree(nrg);
560 	return del;
561 }
562 
563 /*
564  * A rare out of memory error was encountered which prevented removal of
565  * the reserve map region for a page.  The huge page itself was free'ed
566  * and removed from the page cache.  This routine will adjust the subpool
567  * usage count, and the global reserve count if needed.  By incrementing
568  * these counts, the reserve map entry which could not be deleted will
569  * appear as a "reserved" entry instead of simply dangling with incorrect
570  * counts.
571  */
572 void hugetlb_fix_reserve_counts(struct inode *inode)
573 {
574 	struct hugepage_subpool *spool = subpool_inode(inode);
575 	long rsv_adjust;
576 
577 	rsv_adjust = hugepage_subpool_get_pages(spool, 1);
578 	if (rsv_adjust) {
579 		struct hstate *h = hstate_inode(inode);
580 
581 		hugetlb_acct_memory(h, 1);
582 	}
583 }
584 
585 /*
586  * Count and return the number of huge pages in the reserve map
587  * that intersect with the range [f, t).
588  */
589 static long region_count(struct resv_map *resv, long f, long t)
590 {
591 	struct list_head *head = &resv->regions;
592 	struct file_region *rg;
593 	long chg = 0;
594 
595 	spin_lock(&resv->lock);
596 	/* Locate each segment we overlap with, and count that overlap. */
597 	list_for_each_entry(rg, head, link) {
598 		long seg_from;
599 		long seg_to;
600 
601 		if (rg->to <= f)
602 			continue;
603 		if (rg->from >= t)
604 			break;
605 
606 		seg_from = max(rg->from, f);
607 		seg_to = min(rg->to, t);
608 
609 		chg += seg_to - seg_from;
610 	}
611 	spin_unlock(&resv->lock);
612 
613 	return chg;
614 }
615 
616 /*
617  * Convert the address within this vma to the page offset within
618  * the mapping, in pagecache page units; huge pages here.
619  */
620 static pgoff_t vma_hugecache_offset(struct hstate *h,
621 			struct vm_area_struct *vma, unsigned long address)
622 {
623 	return ((address - vma->vm_start) >> huge_page_shift(h)) +
624 			(vma->vm_pgoff >> huge_page_order(h));
625 }
626 
627 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
628 				     unsigned long address)
629 {
630 	return vma_hugecache_offset(hstate_vma(vma), vma, address);
631 }
632 EXPORT_SYMBOL_GPL(linear_hugepage_index);
633 
634 /*
635  * Return the size of the pages allocated when backing a VMA. In the majority
636  * cases this will be same size as used by the page table entries.
637  */
638 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
639 {
640 	if (vma->vm_ops && vma->vm_ops->pagesize)
641 		return vma->vm_ops->pagesize(vma);
642 	return PAGE_SIZE;
643 }
644 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
645 
646 /*
647  * Return the page size being used by the MMU to back a VMA. In the majority
648  * of cases, the page size used by the kernel matches the MMU size. On
649  * architectures where it differs, an architecture-specific 'strong'
650  * version of this symbol is required.
651  */
652 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
653 {
654 	return vma_kernel_pagesize(vma);
655 }
656 
657 /*
658  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
659  * bits of the reservation map pointer, which are always clear due to
660  * alignment.
661  */
662 #define HPAGE_RESV_OWNER    (1UL << 0)
663 #define HPAGE_RESV_UNMAPPED (1UL << 1)
664 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
665 
666 /*
667  * These helpers are used to track how many pages are reserved for
668  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
669  * is guaranteed to have their future faults succeed.
670  *
671  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
672  * the reserve counters are updated with the hugetlb_lock held. It is safe
673  * to reset the VMA at fork() time as it is not in use yet and there is no
674  * chance of the global counters getting corrupted as a result of the values.
675  *
676  * The private mapping reservation is represented in a subtly different
677  * manner to a shared mapping.  A shared mapping has a region map associated
678  * with the underlying file, this region map represents the backing file
679  * pages which have ever had a reservation assigned which this persists even
680  * after the page is instantiated.  A private mapping has a region map
681  * associated with the original mmap which is attached to all VMAs which
682  * reference it, this region map represents those offsets which have consumed
683  * reservation ie. where pages have been instantiated.
684  */
685 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
686 {
687 	return (unsigned long)vma->vm_private_data;
688 }
689 
690 static void set_vma_private_data(struct vm_area_struct *vma,
691 							unsigned long value)
692 {
693 	vma->vm_private_data = (void *)value;
694 }
695 
696 struct resv_map *resv_map_alloc(void)
697 {
698 	struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
699 	struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
700 
701 	if (!resv_map || !rg) {
702 		kfree(resv_map);
703 		kfree(rg);
704 		return NULL;
705 	}
706 
707 	kref_init(&resv_map->refs);
708 	spin_lock_init(&resv_map->lock);
709 	INIT_LIST_HEAD(&resv_map->regions);
710 
711 	resv_map->adds_in_progress = 0;
712 
713 	INIT_LIST_HEAD(&resv_map->region_cache);
714 	list_add(&rg->link, &resv_map->region_cache);
715 	resv_map->region_cache_count = 1;
716 
717 	return resv_map;
718 }
719 
720 void resv_map_release(struct kref *ref)
721 {
722 	struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
723 	struct list_head *head = &resv_map->region_cache;
724 	struct file_region *rg, *trg;
725 
726 	/* Clear out any active regions before we release the map. */
727 	region_del(resv_map, 0, LONG_MAX);
728 
729 	/* ... and any entries left in the cache */
730 	list_for_each_entry_safe(rg, trg, head, link) {
731 		list_del(&rg->link);
732 		kfree(rg);
733 	}
734 
735 	VM_BUG_ON(resv_map->adds_in_progress);
736 
737 	kfree(resv_map);
738 }
739 
740 static inline struct resv_map *inode_resv_map(struct inode *inode)
741 {
742 	return inode->i_mapping->private_data;
743 }
744 
745 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
746 {
747 	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
748 	if (vma->vm_flags & VM_MAYSHARE) {
749 		struct address_space *mapping = vma->vm_file->f_mapping;
750 		struct inode *inode = mapping->host;
751 
752 		return inode_resv_map(inode);
753 
754 	} else {
755 		return (struct resv_map *)(get_vma_private_data(vma) &
756 							~HPAGE_RESV_MASK);
757 	}
758 }
759 
760 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
761 {
762 	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
763 	VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
764 
765 	set_vma_private_data(vma, (get_vma_private_data(vma) &
766 				HPAGE_RESV_MASK) | (unsigned long)map);
767 }
768 
769 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
770 {
771 	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
772 	VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
773 
774 	set_vma_private_data(vma, get_vma_private_data(vma) | flags);
775 }
776 
777 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
778 {
779 	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
780 
781 	return (get_vma_private_data(vma) & flag) != 0;
782 }
783 
784 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
785 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
786 {
787 	VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
788 	if (!(vma->vm_flags & VM_MAYSHARE))
789 		vma->vm_private_data = (void *)0;
790 }
791 
792 /* Returns true if the VMA has associated reserve pages */
793 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
794 {
795 	if (vma->vm_flags & VM_NORESERVE) {
796 		/*
797 		 * This address is already reserved by other process(chg == 0),
798 		 * so, we should decrement reserved count. Without decrementing,
799 		 * reserve count remains after releasing inode, because this
800 		 * allocated page will go into page cache and is regarded as
801 		 * coming from reserved pool in releasing step.  Currently, we
802 		 * don't have any other solution to deal with this situation
803 		 * properly, so add work-around here.
804 		 */
805 		if (vma->vm_flags & VM_MAYSHARE && chg == 0)
806 			return true;
807 		else
808 			return false;
809 	}
810 
811 	/* Shared mappings always use reserves */
812 	if (vma->vm_flags & VM_MAYSHARE) {
813 		/*
814 		 * We know VM_NORESERVE is not set.  Therefore, there SHOULD
815 		 * be a region map for all pages.  The only situation where
816 		 * there is no region map is if a hole was punched via
817 		 * fallocate.  In this case, there really are no reverves to
818 		 * use.  This situation is indicated if chg != 0.
819 		 */
820 		if (chg)
821 			return false;
822 		else
823 			return true;
824 	}
825 
826 	/*
827 	 * Only the process that called mmap() has reserves for
828 	 * private mappings.
829 	 */
830 	if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
831 		/*
832 		 * Like the shared case above, a hole punch or truncate
833 		 * could have been performed on the private mapping.
834 		 * Examine the value of chg to determine if reserves
835 		 * actually exist or were previously consumed.
836 		 * Very Subtle - The value of chg comes from a previous
837 		 * call to vma_needs_reserves().  The reserve map for
838 		 * private mappings has different (opposite) semantics
839 		 * than that of shared mappings.  vma_needs_reserves()
840 		 * has already taken this difference in semantics into
841 		 * account.  Therefore, the meaning of chg is the same
842 		 * as in the shared case above.  Code could easily be
843 		 * combined, but keeping it separate draws attention to
844 		 * subtle differences.
845 		 */
846 		if (chg)
847 			return false;
848 		else
849 			return true;
850 	}
851 
852 	return false;
853 }
854 
855 static void enqueue_huge_page(struct hstate *h, struct page *page)
856 {
857 	int nid = page_to_nid(page);
858 	list_move(&page->lru, &h->hugepage_freelists[nid]);
859 	h->free_huge_pages++;
860 	h->free_huge_pages_node[nid]++;
861 }
862 
863 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
864 {
865 	struct page *page;
866 
867 	list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
868 		if (!PageHWPoison(page))
869 			break;
870 	/*
871 	 * if 'non-isolated free hugepage' not found on the list,
872 	 * the allocation fails.
873 	 */
874 	if (&h->hugepage_freelists[nid] == &page->lru)
875 		return NULL;
876 	list_move(&page->lru, &h->hugepage_activelist);
877 	set_page_refcounted(page);
878 	h->free_huge_pages--;
879 	h->free_huge_pages_node[nid]--;
880 	return page;
881 }
882 
883 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
884 		nodemask_t *nmask)
885 {
886 	unsigned int cpuset_mems_cookie;
887 	struct zonelist *zonelist;
888 	struct zone *zone;
889 	struct zoneref *z;
890 	int node = -1;
891 
892 	zonelist = node_zonelist(nid, gfp_mask);
893 
894 retry_cpuset:
895 	cpuset_mems_cookie = read_mems_allowed_begin();
896 	for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
897 		struct page *page;
898 
899 		if (!cpuset_zone_allowed(zone, gfp_mask))
900 			continue;
901 		/*
902 		 * no need to ask again on the same node. Pool is node rather than
903 		 * zone aware
904 		 */
905 		if (zone_to_nid(zone) == node)
906 			continue;
907 		node = zone_to_nid(zone);
908 
909 		page = dequeue_huge_page_node_exact(h, node);
910 		if (page)
911 			return page;
912 	}
913 	if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
914 		goto retry_cpuset;
915 
916 	return NULL;
917 }
918 
919 /* Movability of hugepages depends on migration support. */
920 static inline gfp_t htlb_alloc_mask(struct hstate *h)
921 {
922 	if (hugepage_migration_supported(h))
923 		return GFP_HIGHUSER_MOVABLE;
924 	else
925 		return GFP_HIGHUSER;
926 }
927 
928 static struct page *dequeue_huge_page_vma(struct hstate *h,
929 				struct vm_area_struct *vma,
930 				unsigned long address, int avoid_reserve,
931 				long chg)
932 {
933 	struct page *page;
934 	struct mempolicy *mpol;
935 	gfp_t gfp_mask;
936 	nodemask_t *nodemask;
937 	int nid;
938 
939 	/*
940 	 * A child process with MAP_PRIVATE mappings created by their parent
941 	 * have no page reserves. This check ensures that reservations are
942 	 * not "stolen". The child may still get SIGKILLed
943 	 */
944 	if (!vma_has_reserves(vma, chg) &&
945 			h->free_huge_pages - h->resv_huge_pages == 0)
946 		goto err;
947 
948 	/* If reserves cannot be used, ensure enough pages are in the pool */
949 	if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
950 		goto err;
951 
952 	gfp_mask = htlb_alloc_mask(h);
953 	nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
954 	page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
955 	if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
956 		SetPagePrivate(page);
957 		h->resv_huge_pages--;
958 	}
959 
960 	mpol_cond_put(mpol);
961 	return page;
962 
963 err:
964 	return NULL;
965 }
966 
967 /*
968  * common helper functions for hstate_next_node_to_{alloc|free}.
969  * We may have allocated or freed a huge page based on a different
970  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
971  * be outside of *nodes_allowed.  Ensure that we use an allowed
972  * node for alloc or free.
973  */
974 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
975 {
976 	nid = next_node_in(nid, *nodes_allowed);
977 	VM_BUG_ON(nid >= MAX_NUMNODES);
978 
979 	return nid;
980 }
981 
982 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
983 {
984 	if (!node_isset(nid, *nodes_allowed))
985 		nid = next_node_allowed(nid, nodes_allowed);
986 	return nid;
987 }
988 
989 /*
990  * returns the previously saved node ["this node"] from which to
991  * allocate a persistent huge page for the pool and advance the
992  * next node from which to allocate, handling wrap at end of node
993  * mask.
994  */
995 static int hstate_next_node_to_alloc(struct hstate *h,
996 					nodemask_t *nodes_allowed)
997 {
998 	int nid;
999 
1000 	VM_BUG_ON(!nodes_allowed);
1001 
1002 	nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1003 	h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1004 
1005 	return nid;
1006 }
1007 
1008 /*
1009  * helper for free_pool_huge_page() - return the previously saved
1010  * node ["this node"] from which to free a huge page.  Advance the
1011  * next node id whether or not we find a free huge page to free so
1012  * that the next attempt to free addresses the next node.
1013  */
1014 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1015 {
1016 	int nid;
1017 
1018 	VM_BUG_ON(!nodes_allowed);
1019 
1020 	nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1021 	h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1022 
1023 	return nid;
1024 }
1025 
1026 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)		\
1027 	for (nr_nodes = nodes_weight(*mask);				\
1028 		nr_nodes > 0 &&						\
1029 		((node = hstate_next_node_to_alloc(hs, mask)) || 1);	\
1030 		nr_nodes--)
1031 
1032 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)		\
1033 	for (nr_nodes = nodes_weight(*mask);				\
1034 		nr_nodes > 0 &&						\
1035 		((node = hstate_next_node_to_free(hs, mask)) || 1);	\
1036 		nr_nodes--)
1037 
1038 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1039 static void destroy_compound_gigantic_page(struct page *page,
1040 					unsigned int order)
1041 {
1042 	int i;
1043 	int nr_pages = 1 << order;
1044 	struct page *p = page + 1;
1045 
1046 	atomic_set(compound_mapcount_ptr(page), 0);
1047 	for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1048 		clear_compound_head(p);
1049 		set_page_refcounted(p);
1050 	}
1051 
1052 	set_compound_order(page, 0);
1053 	__ClearPageHead(page);
1054 }
1055 
1056 static void free_gigantic_page(struct page *page, unsigned int order)
1057 {
1058 	free_contig_range(page_to_pfn(page), 1 << order);
1059 }
1060 
1061 static int __alloc_gigantic_page(unsigned long start_pfn,
1062 				unsigned long nr_pages, gfp_t gfp_mask)
1063 {
1064 	unsigned long end_pfn = start_pfn + nr_pages;
1065 	return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1066 				  gfp_mask);
1067 }
1068 
1069 static bool pfn_range_valid_gigantic(struct zone *z,
1070 			unsigned long start_pfn, unsigned long nr_pages)
1071 {
1072 	unsigned long i, end_pfn = start_pfn + nr_pages;
1073 	struct page *page;
1074 
1075 	for (i = start_pfn; i < end_pfn; i++) {
1076 		if (!pfn_valid(i))
1077 			return false;
1078 
1079 		page = pfn_to_page(i);
1080 
1081 		if (page_zone(page) != z)
1082 			return false;
1083 
1084 		if (PageReserved(page))
1085 			return false;
1086 
1087 		if (page_count(page) > 0)
1088 			return false;
1089 
1090 		if (PageHuge(page))
1091 			return false;
1092 	}
1093 
1094 	return true;
1095 }
1096 
1097 static bool zone_spans_last_pfn(const struct zone *zone,
1098 			unsigned long start_pfn, unsigned long nr_pages)
1099 {
1100 	unsigned long last_pfn = start_pfn + nr_pages - 1;
1101 	return zone_spans_pfn(zone, last_pfn);
1102 }
1103 
1104 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1105 		int nid, nodemask_t *nodemask)
1106 {
1107 	unsigned int order = huge_page_order(h);
1108 	unsigned long nr_pages = 1 << order;
1109 	unsigned long ret, pfn, flags;
1110 	struct zonelist *zonelist;
1111 	struct zone *zone;
1112 	struct zoneref *z;
1113 
1114 	zonelist = node_zonelist(nid, gfp_mask);
1115 	for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) {
1116 		spin_lock_irqsave(&zone->lock, flags);
1117 
1118 		pfn = ALIGN(zone->zone_start_pfn, nr_pages);
1119 		while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
1120 			if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) {
1121 				/*
1122 				 * We release the zone lock here because
1123 				 * alloc_contig_range() will also lock the zone
1124 				 * at some point. If there's an allocation
1125 				 * spinning on this lock, it may win the race
1126 				 * and cause alloc_contig_range() to fail...
1127 				 */
1128 				spin_unlock_irqrestore(&zone->lock, flags);
1129 				ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask);
1130 				if (!ret)
1131 					return pfn_to_page(pfn);
1132 				spin_lock_irqsave(&zone->lock, flags);
1133 			}
1134 			pfn += nr_pages;
1135 		}
1136 
1137 		spin_unlock_irqrestore(&zone->lock, flags);
1138 	}
1139 
1140 	return NULL;
1141 }
1142 
1143 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1144 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1145 
1146 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1147 static inline bool gigantic_page_supported(void) { return false; }
1148 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1149 		int nid, nodemask_t *nodemask) { return NULL; }
1150 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1151 static inline void destroy_compound_gigantic_page(struct page *page,
1152 						unsigned int order) { }
1153 #endif
1154 
1155 static void update_and_free_page(struct hstate *h, struct page *page)
1156 {
1157 	int i;
1158 
1159 	if (hstate_is_gigantic(h) && !gigantic_page_supported())
1160 		return;
1161 
1162 	h->nr_huge_pages--;
1163 	h->nr_huge_pages_node[page_to_nid(page)]--;
1164 	for (i = 0; i < pages_per_huge_page(h); i++) {
1165 		page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1166 				1 << PG_referenced | 1 << PG_dirty |
1167 				1 << PG_active | 1 << PG_private |
1168 				1 << PG_writeback);
1169 	}
1170 	VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1171 	set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1172 	set_page_refcounted(page);
1173 	if (hstate_is_gigantic(h)) {
1174 		destroy_compound_gigantic_page(page, huge_page_order(h));
1175 		free_gigantic_page(page, huge_page_order(h));
1176 	} else {
1177 		__free_pages(page, huge_page_order(h));
1178 	}
1179 }
1180 
1181 struct hstate *size_to_hstate(unsigned long size)
1182 {
1183 	struct hstate *h;
1184 
1185 	for_each_hstate(h) {
1186 		if (huge_page_size(h) == size)
1187 			return h;
1188 	}
1189 	return NULL;
1190 }
1191 
1192 /*
1193  * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1194  * to hstate->hugepage_activelist.)
1195  *
1196  * This function can be called for tail pages, but never returns true for them.
1197  */
1198 bool page_huge_active(struct page *page)
1199 {
1200 	VM_BUG_ON_PAGE(!PageHuge(page), page);
1201 	return PageHead(page) && PagePrivate(&page[1]);
1202 }
1203 
1204 /* never called for tail page */
1205 static void set_page_huge_active(struct page *page)
1206 {
1207 	VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1208 	SetPagePrivate(&page[1]);
1209 }
1210 
1211 static void clear_page_huge_active(struct page *page)
1212 {
1213 	VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1214 	ClearPagePrivate(&page[1]);
1215 }
1216 
1217 /*
1218  * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1219  * code
1220  */
1221 static inline bool PageHugeTemporary(struct page *page)
1222 {
1223 	if (!PageHuge(page))
1224 		return false;
1225 
1226 	return (unsigned long)page[2].mapping == -1U;
1227 }
1228 
1229 static inline void SetPageHugeTemporary(struct page *page)
1230 {
1231 	page[2].mapping = (void *)-1U;
1232 }
1233 
1234 static inline void ClearPageHugeTemporary(struct page *page)
1235 {
1236 	page[2].mapping = NULL;
1237 }
1238 
1239 void free_huge_page(struct page *page)
1240 {
1241 	/*
1242 	 * Can't pass hstate in here because it is called from the
1243 	 * compound page destructor.
1244 	 */
1245 	struct hstate *h = page_hstate(page);
1246 	int nid = page_to_nid(page);
1247 	struct hugepage_subpool *spool =
1248 		(struct hugepage_subpool *)page_private(page);
1249 	bool restore_reserve;
1250 
1251 	VM_BUG_ON_PAGE(page_count(page), page);
1252 	VM_BUG_ON_PAGE(page_mapcount(page), page);
1253 
1254 	set_page_private(page, 0);
1255 	page->mapping = NULL;
1256 	restore_reserve = PagePrivate(page);
1257 	ClearPagePrivate(page);
1258 
1259 	/*
1260 	 * A return code of zero implies that the subpool will be under its
1261 	 * minimum size if the reservation is not restored after page is free.
1262 	 * Therefore, force restore_reserve operation.
1263 	 */
1264 	if (hugepage_subpool_put_pages(spool, 1) == 0)
1265 		restore_reserve = true;
1266 
1267 	spin_lock(&hugetlb_lock);
1268 	clear_page_huge_active(page);
1269 	hugetlb_cgroup_uncharge_page(hstate_index(h),
1270 				     pages_per_huge_page(h), page);
1271 	if (restore_reserve)
1272 		h->resv_huge_pages++;
1273 
1274 	if (PageHugeTemporary(page)) {
1275 		list_del(&page->lru);
1276 		ClearPageHugeTemporary(page);
1277 		update_and_free_page(h, page);
1278 	} else if (h->surplus_huge_pages_node[nid]) {
1279 		/* remove the page from active list */
1280 		list_del(&page->lru);
1281 		update_and_free_page(h, page);
1282 		h->surplus_huge_pages--;
1283 		h->surplus_huge_pages_node[nid]--;
1284 	} else {
1285 		arch_clear_hugepage_flags(page);
1286 		enqueue_huge_page(h, page);
1287 	}
1288 	spin_unlock(&hugetlb_lock);
1289 }
1290 
1291 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1292 {
1293 	INIT_LIST_HEAD(&page->lru);
1294 	set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1295 	spin_lock(&hugetlb_lock);
1296 	set_hugetlb_cgroup(page, NULL);
1297 	h->nr_huge_pages++;
1298 	h->nr_huge_pages_node[nid]++;
1299 	spin_unlock(&hugetlb_lock);
1300 }
1301 
1302 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1303 {
1304 	int i;
1305 	int nr_pages = 1 << order;
1306 	struct page *p = page + 1;
1307 
1308 	/* we rely on prep_new_huge_page to set the destructor */
1309 	set_compound_order(page, order);
1310 	__ClearPageReserved(page);
1311 	__SetPageHead(page);
1312 	for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1313 		/*
1314 		 * For gigantic hugepages allocated through bootmem at
1315 		 * boot, it's safer to be consistent with the not-gigantic
1316 		 * hugepages and clear the PG_reserved bit from all tail pages
1317 		 * too.  Otherwse drivers using get_user_pages() to access tail
1318 		 * pages may get the reference counting wrong if they see
1319 		 * PG_reserved set on a tail page (despite the head page not
1320 		 * having PG_reserved set).  Enforcing this consistency between
1321 		 * head and tail pages allows drivers to optimize away a check
1322 		 * on the head page when they need know if put_page() is needed
1323 		 * after get_user_pages().
1324 		 */
1325 		__ClearPageReserved(p);
1326 		set_page_count(p, 0);
1327 		set_compound_head(p, page);
1328 	}
1329 	atomic_set(compound_mapcount_ptr(page), -1);
1330 }
1331 
1332 /*
1333  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1334  * transparent huge pages.  See the PageTransHuge() documentation for more
1335  * details.
1336  */
1337 int PageHuge(struct page *page)
1338 {
1339 	if (!PageCompound(page))
1340 		return 0;
1341 
1342 	page = compound_head(page);
1343 	return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1344 }
1345 EXPORT_SYMBOL_GPL(PageHuge);
1346 
1347 /*
1348  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1349  * normal or transparent huge pages.
1350  */
1351 int PageHeadHuge(struct page *page_head)
1352 {
1353 	if (!PageHead(page_head))
1354 		return 0;
1355 
1356 	return get_compound_page_dtor(page_head) == free_huge_page;
1357 }
1358 
1359 pgoff_t __basepage_index(struct page *page)
1360 {
1361 	struct page *page_head = compound_head(page);
1362 	pgoff_t index = page_index(page_head);
1363 	unsigned long compound_idx;
1364 
1365 	if (!PageHuge(page_head))
1366 		return page_index(page);
1367 
1368 	if (compound_order(page_head) >= MAX_ORDER)
1369 		compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1370 	else
1371 		compound_idx = page - page_head;
1372 
1373 	return (index << compound_order(page_head)) + compound_idx;
1374 }
1375 
1376 static struct page *alloc_buddy_huge_page(struct hstate *h,
1377 		gfp_t gfp_mask, int nid, nodemask_t *nmask)
1378 {
1379 	int order = huge_page_order(h);
1380 	struct page *page;
1381 
1382 	gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
1383 	if (nid == NUMA_NO_NODE)
1384 		nid = numa_mem_id();
1385 	page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1386 	if (page)
1387 		__count_vm_event(HTLB_BUDDY_PGALLOC);
1388 	else
1389 		__count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1390 
1391 	return page;
1392 }
1393 
1394 /*
1395  * Common helper to allocate a fresh hugetlb page. All specific allocators
1396  * should use this function to get new hugetlb pages
1397  */
1398 static struct page *alloc_fresh_huge_page(struct hstate *h,
1399 		gfp_t gfp_mask, int nid, nodemask_t *nmask)
1400 {
1401 	struct page *page;
1402 
1403 	if (hstate_is_gigantic(h))
1404 		page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1405 	else
1406 		page = alloc_buddy_huge_page(h, gfp_mask,
1407 				nid, nmask);
1408 	if (!page)
1409 		return NULL;
1410 
1411 	if (hstate_is_gigantic(h))
1412 		prep_compound_gigantic_page(page, huge_page_order(h));
1413 	prep_new_huge_page(h, page, page_to_nid(page));
1414 
1415 	return page;
1416 }
1417 
1418 /*
1419  * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1420  * manner.
1421  */
1422 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1423 {
1424 	struct page *page;
1425 	int nr_nodes, node;
1426 	gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1427 
1428 	for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1429 		page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed);
1430 		if (page)
1431 			break;
1432 	}
1433 
1434 	if (!page)
1435 		return 0;
1436 
1437 	put_page(page); /* free it into the hugepage allocator */
1438 
1439 	return 1;
1440 }
1441 
1442 /*
1443  * Free huge page from pool from next node to free.
1444  * Attempt to keep persistent huge pages more or less
1445  * balanced over allowed nodes.
1446  * Called with hugetlb_lock locked.
1447  */
1448 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1449 							 bool acct_surplus)
1450 {
1451 	int nr_nodes, node;
1452 	int ret = 0;
1453 
1454 	for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1455 		/*
1456 		 * If we're returning unused surplus pages, only examine
1457 		 * nodes with surplus pages.
1458 		 */
1459 		if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1460 		    !list_empty(&h->hugepage_freelists[node])) {
1461 			struct page *page =
1462 				list_entry(h->hugepage_freelists[node].next,
1463 					  struct page, lru);
1464 			list_del(&page->lru);
1465 			h->free_huge_pages--;
1466 			h->free_huge_pages_node[node]--;
1467 			if (acct_surplus) {
1468 				h->surplus_huge_pages--;
1469 				h->surplus_huge_pages_node[node]--;
1470 			}
1471 			update_and_free_page(h, page);
1472 			ret = 1;
1473 			break;
1474 		}
1475 	}
1476 
1477 	return ret;
1478 }
1479 
1480 /*
1481  * Dissolve a given free hugepage into free buddy pages. This function does
1482  * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1483  * dissolution fails because a give page is not a free hugepage, or because
1484  * free hugepages are fully reserved.
1485  */
1486 int dissolve_free_huge_page(struct page *page)
1487 {
1488 	int rc = -EBUSY;
1489 
1490 	spin_lock(&hugetlb_lock);
1491 	if (PageHuge(page) && !page_count(page)) {
1492 		struct page *head = compound_head(page);
1493 		struct hstate *h = page_hstate(head);
1494 		int nid = page_to_nid(head);
1495 		if (h->free_huge_pages - h->resv_huge_pages == 0)
1496 			goto out;
1497 		/*
1498 		 * Move PageHWPoison flag from head page to the raw error page,
1499 		 * which makes any subpages rather than the error page reusable.
1500 		 */
1501 		if (PageHWPoison(head) && page != head) {
1502 			SetPageHWPoison(page);
1503 			ClearPageHWPoison(head);
1504 		}
1505 		list_del(&head->lru);
1506 		h->free_huge_pages--;
1507 		h->free_huge_pages_node[nid]--;
1508 		h->max_huge_pages--;
1509 		update_and_free_page(h, head);
1510 		rc = 0;
1511 	}
1512 out:
1513 	spin_unlock(&hugetlb_lock);
1514 	return rc;
1515 }
1516 
1517 /*
1518  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1519  * make specified memory blocks removable from the system.
1520  * Note that this will dissolve a free gigantic hugepage completely, if any
1521  * part of it lies within the given range.
1522  * Also note that if dissolve_free_huge_page() returns with an error, all
1523  * free hugepages that were dissolved before that error are lost.
1524  */
1525 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1526 {
1527 	unsigned long pfn;
1528 	struct page *page;
1529 	int rc = 0;
1530 
1531 	if (!hugepages_supported())
1532 		return rc;
1533 
1534 	for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1535 		page = pfn_to_page(pfn);
1536 		if (PageHuge(page) && !page_count(page)) {
1537 			rc = dissolve_free_huge_page(page);
1538 			if (rc)
1539 				break;
1540 		}
1541 	}
1542 
1543 	return rc;
1544 }
1545 
1546 /*
1547  * Allocates a fresh surplus page from the page allocator.
1548  */
1549 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1550 		int nid, nodemask_t *nmask)
1551 {
1552 	struct page *page = NULL;
1553 
1554 	if (hstate_is_gigantic(h))
1555 		return NULL;
1556 
1557 	spin_lock(&hugetlb_lock);
1558 	if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1559 		goto out_unlock;
1560 	spin_unlock(&hugetlb_lock);
1561 
1562 	page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1563 	if (!page)
1564 		return NULL;
1565 
1566 	spin_lock(&hugetlb_lock);
1567 	/*
1568 	 * We could have raced with the pool size change.
1569 	 * Double check that and simply deallocate the new page
1570 	 * if we would end up overcommiting the surpluses. Abuse
1571 	 * temporary page to workaround the nasty free_huge_page
1572 	 * codeflow
1573 	 */
1574 	if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1575 		SetPageHugeTemporary(page);
1576 		put_page(page);
1577 		page = NULL;
1578 	} else {
1579 		h->surplus_huge_pages++;
1580 		h->surplus_huge_pages_node[page_to_nid(page)]++;
1581 	}
1582 
1583 out_unlock:
1584 	spin_unlock(&hugetlb_lock);
1585 
1586 	return page;
1587 }
1588 
1589 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1590 		int nid, nodemask_t *nmask)
1591 {
1592 	struct page *page;
1593 
1594 	if (hstate_is_gigantic(h))
1595 		return NULL;
1596 
1597 	page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1598 	if (!page)
1599 		return NULL;
1600 
1601 	/*
1602 	 * We do not account these pages as surplus because they are only
1603 	 * temporary and will be released properly on the last reference
1604 	 */
1605 	SetPageHugeTemporary(page);
1606 
1607 	return page;
1608 }
1609 
1610 /*
1611  * Use the VMA's mpolicy to allocate a huge page from the buddy.
1612  */
1613 static
1614 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1615 		struct vm_area_struct *vma, unsigned long addr)
1616 {
1617 	struct page *page;
1618 	struct mempolicy *mpol;
1619 	gfp_t gfp_mask = htlb_alloc_mask(h);
1620 	int nid;
1621 	nodemask_t *nodemask;
1622 
1623 	nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1624 	page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1625 	mpol_cond_put(mpol);
1626 
1627 	return page;
1628 }
1629 
1630 /* page migration callback function */
1631 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1632 {
1633 	gfp_t gfp_mask = htlb_alloc_mask(h);
1634 	struct page *page = NULL;
1635 
1636 	if (nid != NUMA_NO_NODE)
1637 		gfp_mask |= __GFP_THISNODE;
1638 
1639 	spin_lock(&hugetlb_lock);
1640 	if (h->free_huge_pages - h->resv_huge_pages > 0)
1641 		page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1642 	spin_unlock(&hugetlb_lock);
1643 
1644 	if (!page)
1645 		page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1646 
1647 	return page;
1648 }
1649 
1650 /* page migration callback function */
1651 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1652 		nodemask_t *nmask)
1653 {
1654 	gfp_t gfp_mask = htlb_alloc_mask(h);
1655 
1656 	spin_lock(&hugetlb_lock);
1657 	if (h->free_huge_pages - h->resv_huge_pages > 0) {
1658 		struct page *page;
1659 
1660 		page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1661 		if (page) {
1662 			spin_unlock(&hugetlb_lock);
1663 			return page;
1664 		}
1665 	}
1666 	spin_unlock(&hugetlb_lock);
1667 
1668 	return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1669 }
1670 
1671 /* mempolicy aware migration callback */
1672 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1673 		unsigned long address)
1674 {
1675 	struct mempolicy *mpol;
1676 	nodemask_t *nodemask;
1677 	struct page *page;
1678 	gfp_t gfp_mask;
1679 	int node;
1680 
1681 	gfp_mask = htlb_alloc_mask(h);
1682 	node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1683 	page = alloc_huge_page_nodemask(h, node, nodemask);
1684 	mpol_cond_put(mpol);
1685 
1686 	return page;
1687 }
1688 
1689 /*
1690  * Increase the hugetlb pool such that it can accommodate a reservation
1691  * of size 'delta'.
1692  */
1693 static int gather_surplus_pages(struct hstate *h, int delta)
1694 {
1695 	struct list_head surplus_list;
1696 	struct page *page, *tmp;
1697 	int ret, i;
1698 	int needed, allocated;
1699 	bool alloc_ok = true;
1700 
1701 	needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1702 	if (needed <= 0) {
1703 		h->resv_huge_pages += delta;
1704 		return 0;
1705 	}
1706 
1707 	allocated = 0;
1708 	INIT_LIST_HEAD(&surplus_list);
1709 
1710 	ret = -ENOMEM;
1711 retry:
1712 	spin_unlock(&hugetlb_lock);
1713 	for (i = 0; i < needed; i++) {
1714 		page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1715 				NUMA_NO_NODE, NULL);
1716 		if (!page) {
1717 			alloc_ok = false;
1718 			break;
1719 		}
1720 		list_add(&page->lru, &surplus_list);
1721 		cond_resched();
1722 	}
1723 	allocated += i;
1724 
1725 	/*
1726 	 * After retaking hugetlb_lock, we need to recalculate 'needed'
1727 	 * because either resv_huge_pages or free_huge_pages may have changed.
1728 	 */
1729 	spin_lock(&hugetlb_lock);
1730 	needed = (h->resv_huge_pages + delta) -
1731 			(h->free_huge_pages + allocated);
1732 	if (needed > 0) {
1733 		if (alloc_ok)
1734 			goto retry;
1735 		/*
1736 		 * We were not able to allocate enough pages to
1737 		 * satisfy the entire reservation so we free what
1738 		 * we've allocated so far.
1739 		 */
1740 		goto free;
1741 	}
1742 	/*
1743 	 * The surplus_list now contains _at_least_ the number of extra pages
1744 	 * needed to accommodate the reservation.  Add the appropriate number
1745 	 * of pages to the hugetlb pool and free the extras back to the buddy
1746 	 * allocator.  Commit the entire reservation here to prevent another
1747 	 * process from stealing the pages as they are added to the pool but
1748 	 * before they are reserved.
1749 	 */
1750 	needed += allocated;
1751 	h->resv_huge_pages += delta;
1752 	ret = 0;
1753 
1754 	/* Free the needed pages to the hugetlb pool */
1755 	list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1756 		if ((--needed) < 0)
1757 			break;
1758 		/*
1759 		 * This page is now managed by the hugetlb allocator and has
1760 		 * no users -- drop the buddy allocator's reference.
1761 		 */
1762 		put_page_testzero(page);
1763 		VM_BUG_ON_PAGE(page_count(page), page);
1764 		enqueue_huge_page(h, page);
1765 	}
1766 free:
1767 	spin_unlock(&hugetlb_lock);
1768 
1769 	/* Free unnecessary surplus pages to the buddy allocator */
1770 	list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1771 		put_page(page);
1772 	spin_lock(&hugetlb_lock);
1773 
1774 	return ret;
1775 }
1776 
1777 /*
1778  * This routine has two main purposes:
1779  * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1780  *    in unused_resv_pages.  This corresponds to the prior adjustments made
1781  *    to the associated reservation map.
1782  * 2) Free any unused surplus pages that may have been allocated to satisfy
1783  *    the reservation.  As many as unused_resv_pages may be freed.
1784  *
1785  * Called with hugetlb_lock held.  However, the lock could be dropped (and
1786  * reacquired) during calls to cond_resched_lock.  Whenever dropping the lock,
1787  * we must make sure nobody else can claim pages we are in the process of
1788  * freeing.  Do this by ensuring resv_huge_page always is greater than the
1789  * number of huge pages we plan to free when dropping the lock.
1790  */
1791 static void return_unused_surplus_pages(struct hstate *h,
1792 					unsigned long unused_resv_pages)
1793 {
1794 	unsigned long nr_pages;
1795 
1796 	/* Cannot return gigantic pages currently */
1797 	if (hstate_is_gigantic(h))
1798 		goto out;
1799 
1800 	/*
1801 	 * Part (or even all) of the reservation could have been backed
1802 	 * by pre-allocated pages. Only free surplus pages.
1803 	 */
1804 	nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1805 
1806 	/*
1807 	 * We want to release as many surplus pages as possible, spread
1808 	 * evenly across all nodes with memory. Iterate across these nodes
1809 	 * until we can no longer free unreserved surplus pages. This occurs
1810 	 * when the nodes with surplus pages have no free pages.
1811 	 * free_pool_huge_page() will balance the the freed pages across the
1812 	 * on-line nodes with memory and will handle the hstate accounting.
1813 	 *
1814 	 * Note that we decrement resv_huge_pages as we free the pages.  If
1815 	 * we drop the lock, resv_huge_pages will still be sufficiently large
1816 	 * to cover subsequent pages we may free.
1817 	 */
1818 	while (nr_pages--) {
1819 		h->resv_huge_pages--;
1820 		unused_resv_pages--;
1821 		if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1822 			goto out;
1823 		cond_resched_lock(&hugetlb_lock);
1824 	}
1825 
1826 out:
1827 	/* Fully uncommit the reservation */
1828 	h->resv_huge_pages -= unused_resv_pages;
1829 }
1830 
1831 
1832 /*
1833  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1834  * are used by the huge page allocation routines to manage reservations.
1835  *
1836  * vma_needs_reservation is called to determine if the huge page at addr
1837  * within the vma has an associated reservation.  If a reservation is
1838  * needed, the value 1 is returned.  The caller is then responsible for
1839  * managing the global reservation and subpool usage counts.  After
1840  * the huge page has been allocated, vma_commit_reservation is called
1841  * to add the page to the reservation map.  If the page allocation fails,
1842  * the reservation must be ended instead of committed.  vma_end_reservation
1843  * is called in such cases.
1844  *
1845  * In the normal case, vma_commit_reservation returns the same value
1846  * as the preceding vma_needs_reservation call.  The only time this
1847  * is not the case is if a reserve map was changed between calls.  It
1848  * is the responsibility of the caller to notice the difference and
1849  * take appropriate action.
1850  *
1851  * vma_add_reservation is used in error paths where a reservation must
1852  * be restored when a newly allocated huge page must be freed.  It is
1853  * to be called after calling vma_needs_reservation to determine if a
1854  * reservation exists.
1855  */
1856 enum vma_resv_mode {
1857 	VMA_NEEDS_RESV,
1858 	VMA_COMMIT_RESV,
1859 	VMA_END_RESV,
1860 	VMA_ADD_RESV,
1861 };
1862 static long __vma_reservation_common(struct hstate *h,
1863 				struct vm_area_struct *vma, unsigned long addr,
1864 				enum vma_resv_mode mode)
1865 {
1866 	struct resv_map *resv;
1867 	pgoff_t idx;
1868 	long ret;
1869 
1870 	resv = vma_resv_map(vma);
1871 	if (!resv)
1872 		return 1;
1873 
1874 	idx = vma_hugecache_offset(h, vma, addr);
1875 	switch (mode) {
1876 	case VMA_NEEDS_RESV:
1877 		ret = region_chg(resv, idx, idx + 1);
1878 		break;
1879 	case VMA_COMMIT_RESV:
1880 		ret = region_add(resv, idx, idx + 1);
1881 		break;
1882 	case VMA_END_RESV:
1883 		region_abort(resv, idx, idx + 1);
1884 		ret = 0;
1885 		break;
1886 	case VMA_ADD_RESV:
1887 		if (vma->vm_flags & VM_MAYSHARE)
1888 			ret = region_add(resv, idx, idx + 1);
1889 		else {
1890 			region_abort(resv, idx, idx + 1);
1891 			ret = region_del(resv, idx, idx + 1);
1892 		}
1893 		break;
1894 	default:
1895 		BUG();
1896 	}
1897 
1898 	if (vma->vm_flags & VM_MAYSHARE)
1899 		return ret;
1900 	else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1901 		/*
1902 		 * In most cases, reserves always exist for private mappings.
1903 		 * However, a file associated with mapping could have been
1904 		 * hole punched or truncated after reserves were consumed.
1905 		 * As subsequent fault on such a range will not use reserves.
1906 		 * Subtle - The reserve map for private mappings has the
1907 		 * opposite meaning than that of shared mappings.  If NO
1908 		 * entry is in the reserve map, it means a reservation exists.
1909 		 * If an entry exists in the reserve map, it means the
1910 		 * reservation has already been consumed.  As a result, the
1911 		 * return value of this routine is the opposite of the
1912 		 * value returned from reserve map manipulation routines above.
1913 		 */
1914 		if (ret)
1915 			return 0;
1916 		else
1917 			return 1;
1918 	}
1919 	else
1920 		return ret < 0 ? ret : 0;
1921 }
1922 
1923 static long vma_needs_reservation(struct hstate *h,
1924 			struct vm_area_struct *vma, unsigned long addr)
1925 {
1926 	return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1927 }
1928 
1929 static long vma_commit_reservation(struct hstate *h,
1930 			struct vm_area_struct *vma, unsigned long addr)
1931 {
1932 	return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1933 }
1934 
1935 static void vma_end_reservation(struct hstate *h,
1936 			struct vm_area_struct *vma, unsigned long addr)
1937 {
1938 	(void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1939 }
1940 
1941 static long vma_add_reservation(struct hstate *h,
1942 			struct vm_area_struct *vma, unsigned long addr)
1943 {
1944 	return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1945 }
1946 
1947 /*
1948  * This routine is called to restore a reservation on error paths.  In the
1949  * specific error paths, a huge page was allocated (via alloc_huge_page)
1950  * and is about to be freed.  If a reservation for the page existed,
1951  * alloc_huge_page would have consumed the reservation and set PagePrivate
1952  * in the newly allocated page.  When the page is freed via free_huge_page,
1953  * the global reservation count will be incremented if PagePrivate is set.
1954  * However, free_huge_page can not adjust the reserve map.  Adjust the
1955  * reserve map here to be consistent with global reserve count adjustments
1956  * to be made by free_huge_page.
1957  */
1958 static void restore_reserve_on_error(struct hstate *h,
1959 			struct vm_area_struct *vma, unsigned long address,
1960 			struct page *page)
1961 {
1962 	if (unlikely(PagePrivate(page))) {
1963 		long rc = vma_needs_reservation(h, vma, address);
1964 
1965 		if (unlikely(rc < 0)) {
1966 			/*
1967 			 * Rare out of memory condition in reserve map
1968 			 * manipulation.  Clear PagePrivate so that
1969 			 * global reserve count will not be incremented
1970 			 * by free_huge_page.  This will make it appear
1971 			 * as though the reservation for this page was
1972 			 * consumed.  This may prevent the task from
1973 			 * faulting in the page at a later time.  This
1974 			 * is better than inconsistent global huge page
1975 			 * accounting of reserve counts.
1976 			 */
1977 			ClearPagePrivate(page);
1978 		} else if (rc) {
1979 			rc = vma_add_reservation(h, vma, address);
1980 			if (unlikely(rc < 0))
1981 				/*
1982 				 * See above comment about rare out of
1983 				 * memory condition.
1984 				 */
1985 				ClearPagePrivate(page);
1986 		} else
1987 			vma_end_reservation(h, vma, address);
1988 	}
1989 }
1990 
1991 struct page *alloc_huge_page(struct vm_area_struct *vma,
1992 				    unsigned long addr, int avoid_reserve)
1993 {
1994 	struct hugepage_subpool *spool = subpool_vma(vma);
1995 	struct hstate *h = hstate_vma(vma);
1996 	struct page *page;
1997 	long map_chg, map_commit;
1998 	long gbl_chg;
1999 	int ret, idx;
2000 	struct hugetlb_cgroup *h_cg;
2001 
2002 	idx = hstate_index(h);
2003 	/*
2004 	 * Examine the region/reserve map to determine if the process
2005 	 * has a reservation for the page to be allocated.  A return
2006 	 * code of zero indicates a reservation exists (no change).
2007 	 */
2008 	map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2009 	if (map_chg < 0)
2010 		return ERR_PTR(-ENOMEM);
2011 
2012 	/*
2013 	 * Processes that did not create the mapping will have no
2014 	 * reserves as indicated by the region/reserve map. Check
2015 	 * that the allocation will not exceed the subpool limit.
2016 	 * Allocations for MAP_NORESERVE mappings also need to be
2017 	 * checked against any subpool limit.
2018 	 */
2019 	if (map_chg || avoid_reserve) {
2020 		gbl_chg = hugepage_subpool_get_pages(spool, 1);
2021 		if (gbl_chg < 0) {
2022 			vma_end_reservation(h, vma, addr);
2023 			return ERR_PTR(-ENOSPC);
2024 		}
2025 
2026 		/*
2027 		 * Even though there was no reservation in the region/reserve
2028 		 * map, there could be reservations associated with the
2029 		 * subpool that can be used.  This would be indicated if the
2030 		 * return value of hugepage_subpool_get_pages() is zero.
2031 		 * However, if avoid_reserve is specified we still avoid even
2032 		 * the subpool reservations.
2033 		 */
2034 		if (avoid_reserve)
2035 			gbl_chg = 1;
2036 	}
2037 
2038 	ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2039 	if (ret)
2040 		goto out_subpool_put;
2041 
2042 	spin_lock(&hugetlb_lock);
2043 	/*
2044 	 * glb_chg is passed to indicate whether or not a page must be taken
2045 	 * from the global free pool (global change).  gbl_chg == 0 indicates
2046 	 * a reservation exists for the allocation.
2047 	 */
2048 	page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2049 	if (!page) {
2050 		spin_unlock(&hugetlb_lock);
2051 		page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2052 		if (!page)
2053 			goto out_uncharge_cgroup;
2054 		if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2055 			SetPagePrivate(page);
2056 			h->resv_huge_pages--;
2057 		}
2058 		spin_lock(&hugetlb_lock);
2059 		list_move(&page->lru, &h->hugepage_activelist);
2060 		/* Fall through */
2061 	}
2062 	hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2063 	spin_unlock(&hugetlb_lock);
2064 
2065 	set_page_private(page, (unsigned long)spool);
2066 
2067 	map_commit = vma_commit_reservation(h, vma, addr);
2068 	if (unlikely(map_chg > map_commit)) {
2069 		/*
2070 		 * The page was added to the reservation map between
2071 		 * vma_needs_reservation and vma_commit_reservation.
2072 		 * This indicates a race with hugetlb_reserve_pages.
2073 		 * Adjust for the subpool count incremented above AND
2074 		 * in hugetlb_reserve_pages for the same page.  Also,
2075 		 * the reservation count added in hugetlb_reserve_pages
2076 		 * no longer applies.
2077 		 */
2078 		long rsv_adjust;
2079 
2080 		rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2081 		hugetlb_acct_memory(h, -rsv_adjust);
2082 	}
2083 	return page;
2084 
2085 out_uncharge_cgroup:
2086 	hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2087 out_subpool_put:
2088 	if (map_chg || avoid_reserve)
2089 		hugepage_subpool_put_pages(spool, 1);
2090 	vma_end_reservation(h, vma, addr);
2091 	return ERR_PTR(-ENOSPC);
2092 }
2093 
2094 int alloc_bootmem_huge_page(struct hstate *h)
2095 	__attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2096 int __alloc_bootmem_huge_page(struct hstate *h)
2097 {
2098 	struct huge_bootmem_page *m;
2099 	int nr_nodes, node;
2100 
2101 	for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2102 		void *addr;
2103 
2104 		addr = memblock_alloc_try_nid_raw(
2105 				huge_page_size(h), huge_page_size(h),
2106 				0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2107 		if (addr) {
2108 			/*
2109 			 * Use the beginning of the huge page to store the
2110 			 * huge_bootmem_page struct (until gather_bootmem
2111 			 * puts them into the mem_map).
2112 			 */
2113 			m = addr;
2114 			goto found;
2115 		}
2116 	}
2117 	return 0;
2118 
2119 found:
2120 	BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2121 	/* Put them into a private list first because mem_map is not up yet */
2122 	INIT_LIST_HEAD(&m->list);
2123 	list_add(&m->list, &huge_boot_pages);
2124 	m->hstate = h;
2125 	return 1;
2126 }
2127 
2128 static void __init prep_compound_huge_page(struct page *page,
2129 		unsigned int order)
2130 {
2131 	if (unlikely(order > (MAX_ORDER - 1)))
2132 		prep_compound_gigantic_page(page, order);
2133 	else
2134 		prep_compound_page(page, order);
2135 }
2136 
2137 /* Put bootmem huge pages into the standard lists after mem_map is up */
2138 static void __init gather_bootmem_prealloc(void)
2139 {
2140 	struct huge_bootmem_page *m;
2141 
2142 	list_for_each_entry(m, &huge_boot_pages, list) {
2143 		struct page *page = virt_to_page(m);
2144 		struct hstate *h = m->hstate;
2145 
2146 		WARN_ON(page_count(page) != 1);
2147 		prep_compound_huge_page(page, h->order);
2148 		WARN_ON(PageReserved(page));
2149 		prep_new_huge_page(h, page, page_to_nid(page));
2150 		put_page(page); /* free it into the hugepage allocator */
2151 
2152 		/*
2153 		 * If we had gigantic hugepages allocated at boot time, we need
2154 		 * to restore the 'stolen' pages to totalram_pages in order to
2155 		 * fix confusing memory reports from free(1) and another
2156 		 * side-effects, like CommitLimit going negative.
2157 		 */
2158 		if (hstate_is_gigantic(h))
2159 			adjust_managed_page_count(page, 1 << h->order);
2160 		cond_resched();
2161 	}
2162 }
2163 
2164 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2165 {
2166 	unsigned long i;
2167 
2168 	for (i = 0; i < h->max_huge_pages; ++i) {
2169 		if (hstate_is_gigantic(h)) {
2170 			if (!alloc_bootmem_huge_page(h))
2171 				break;
2172 		} else if (!alloc_pool_huge_page(h,
2173 					 &node_states[N_MEMORY]))
2174 			break;
2175 		cond_resched();
2176 	}
2177 	if (i < h->max_huge_pages) {
2178 		char buf[32];
2179 
2180 		string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2181 		pr_warn("HugeTLB: allocating %lu of page size %s failed.  Only allocated %lu hugepages.\n",
2182 			h->max_huge_pages, buf, i);
2183 		h->max_huge_pages = i;
2184 	}
2185 }
2186 
2187 static void __init hugetlb_init_hstates(void)
2188 {
2189 	struct hstate *h;
2190 
2191 	for_each_hstate(h) {
2192 		if (minimum_order > huge_page_order(h))
2193 			minimum_order = huge_page_order(h);
2194 
2195 		/* oversize hugepages were init'ed in early boot */
2196 		if (!hstate_is_gigantic(h))
2197 			hugetlb_hstate_alloc_pages(h);
2198 	}
2199 	VM_BUG_ON(minimum_order == UINT_MAX);
2200 }
2201 
2202 static void __init report_hugepages(void)
2203 {
2204 	struct hstate *h;
2205 
2206 	for_each_hstate(h) {
2207 		char buf[32];
2208 
2209 		string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2210 		pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2211 			buf, h->free_huge_pages);
2212 	}
2213 }
2214 
2215 #ifdef CONFIG_HIGHMEM
2216 static void try_to_free_low(struct hstate *h, unsigned long count,
2217 						nodemask_t *nodes_allowed)
2218 {
2219 	int i;
2220 
2221 	if (hstate_is_gigantic(h))
2222 		return;
2223 
2224 	for_each_node_mask(i, *nodes_allowed) {
2225 		struct page *page, *next;
2226 		struct list_head *freel = &h->hugepage_freelists[i];
2227 		list_for_each_entry_safe(page, next, freel, lru) {
2228 			if (count >= h->nr_huge_pages)
2229 				return;
2230 			if (PageHighMem(page))
2231 				continue;
2232 			list_del(&page->lru);
2233 			update_and_free_page(h, page);
2234 			h->free_huge_pages--;
2235 			h->free_huge_pages_node[page_to_nid(page)]--;
2236 		}
2237 	}
2238 }
2239 #else
2240 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2241 						nodemask_t *nodes_allowed)
2242 {
2243 }
2244 #endif
2245 
2246 /*
2247  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
2248  * balanced by operating on them in a round-robin fashion.
2249  * Returns 1 if an adjustment was made.
2250  */
2251 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2252 				int delta)
2253 {
2254 	int nr_nodes, node;
2255 
2256 	VM_BUG_ON(delta != -1 && delta != 1);
2257 
2258 	if (delta < 0) {
2259 		for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2260 			if (h->surplus_huge_pages_node[node])
2261 				goto found;
2262 		}
2263 	} else {
2264 		for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2265 			if (h->surplus_huge_pages_node[node] <
2266 					h->nr_huge_pages_node[node])
2267 				goto found;
2268 		}
2269 	}
2270 	return 0;
2271 
2272 found:
2273 	h->surplus_huge_pages += delta;
2274 	h->surplus_huge_pages_node[node] += delta;
2275 	return 1;
2276 }
2277 
2278 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2279 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2280 						nodemask_t *nodes_allowed)
2281 {
2282 	unsigned long min_count, ret;
2283 
2284 	if (hstate_is_gigantic(h) && !gigantic_page_supported())
2285 		return h->max_huge_pages;
2286 
2287 	/*
2288 	 * Increase the pool size
2289 	 * First take pages out of surplus state.  Then make up the
2290 	 * remaining difference by allocating fresh huge pages.
2291 	 *
2292 	 * We might race with alloc_surplus_huge_page() here and be unable
2293 	 * to convert a surplus huge page to a normal huge page. That is
2294 	 * not critical, though, it just means the overall size of the
2295 	 * pool might be one hugepage larger than it needs to be, but
2296 	 * within all the constraints specified by the sysctls.
2297 	 */
2298 	spin_lock(&hugetlb_lock);
2299 	while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2300 		if (!adjust_pool_surplus(h, nodes_allowed, -1))
2301 			break;
2302 	}
2303 
2304 	while (count > persistent_huge_pages(h)) {
2305 		/*
2306 		 * If this allocation races such that we no longer need the
2307 		 * page, free_huge_page will handle it by freeing the page
2308 		 * and reducing the surplus.
2309 		 */
2310 		spin_unlock(&hugetlb_lock);
2311 
2312 		/* yield cpu to avoid soft lockup */
2313 		cond_resched();
2314 
2315 		ret = alloc_pool_huge_page(h, nodes_allowed);
2316 		spin_lock(&hugetlb_lock);
2317 		if (!ret)
2318 			goto out;
2319 
2320 		/* Bail for signals. Probably ctrl-c from user */
2321 		if (signal_pending(current))
2322 			goto out;
2323 	}
2324 
2325 	/*
2326 	 * Decrease the pool size
2327 	 * First return free pages to the buddy allocator (being careful
2328 	 * to keep enough around to satisfy reservations).  Then place
2329 	 * pages into surplus state as needed so the pool will shrink
2330 	 * to the desired size as pages become free.
2331 	 *
2332 	 * By placing pages into the surplus state independent of the
2333 	 * overcommit value, we are allowing the surplus pool size to
2334 	 * exceed overcommit. There are few sane options here. Since
2335 	 * alloc_surplus_huge_page() is checking the global counter,
2336 	 * though, we'll note that we're not allowed to exceed surplus
2337 	 * and won't grow the pool anywhere else. Not until one of the
2338 	 * sysctls are changed, or the surplus pages go out of use.
2339 	 */
2340 	min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2341 	min_count = max(count, min_count);
2342 	try_to_free_low(h, min_count, nodes_allowed);
2343 	while (min_count < persistent_huge_pages(h)) {
2344 		if (!free_pool_huge_page(h, nodes_allowed, 0))
2345 			break;
2346 		cond_resched_lock(&hugetlb_lock);
2347 	}
2348 	while (count < persistent_huge_pages(h)) {
2349 		if (!adjust_pool_surplus(h, nodes_allowed, 1))
2350 			break;
2351 	}
2352 out:
2353 	ret = persistent_huge_pages(h);
2354 	spin_unlock(&hugetlb_lock);
2355 	return ret;
2356 }
2357 
2358 #define HSTATE_ATTR_RO(_name) \
2359 	static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2360 
2361 #define HSTATE_ATTR(_name) \
2362 	static struct kobj_attribute _name##_attr = \
2363 		__ATTR(_name, 0644, _name##_show, _name##_store)
2364 
2365 static struct kobject *hugepages_kobj;
2366 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2367 
2368 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2369 
2370 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2371 {
2372 	int i;
2373 
2374 	for (i = 0; i < HUGE_MAX_HSTATE; i++)
2375 		if (hstate_kobjs[i] == kobj) {
2376 			if (nidp)
2377 				*nidp = NUMA_NO_NODE;
2378 			return &hstates[i];
2379 		}
2380 
2381 	return kobj_to_node_hstate(kobj, nidp);
2382 }
2383 
2384 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2385 					struct kobj_attribute *attr, char *buf)
2386 {
2387 	struct hstate *h;
2388 	unsigned long nr_huge_pages;
2389 	int nid;
2390 
2391 	h = kobj_to_hstate(kobj, &nid);
2392 	if (nid == NUMA_NO_NODE)
2393 		nr_huge_pages = h->nr_huge_pages;
2394 	else
2395 		nr_huge_pages = h->nr_huge_pages_node[nid];
2396 
2397 	return sprintf(buf, "%lu\n", nr_huge_pages);
2398 }
2399 
2400 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2401 					   struct hstate *h, int nid,
2402 					   unsigned long count, size_t len)
2403 {
2404 	int err;
2405 	NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2406 
2407 	if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2408 		err = -EINVAL;
2409 		goto out;
2410 	}
2411 
2412 	if (nid == NUMA_NO_NODE) {
2413 		/*
2414 		 * global hstate attribute
2415 		 */
2416 		if (!(obey_mempolicy &&
2417 				init_nodemask_of_mempolicy(nodes_allowed))) {
2418 			NODEMASK_FREE(nodes_allowed);
2419 			nodes_allowed = &node_states[N_MEMORY];
2420 		}
2421 	} else if (nodes_allowed) {
2422 		/*
2423 		 * per node hstate attribute: adjust count to global,
2424 		 * but restrict alloc/free to the specified node.
2425 		 */
2426 		count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2427 		init_nodemask_of_node(nodes_allowed, nid);
2428 	} else
2429 		nodes_allowed = &node_states[N_MEMORY];
2430 
2431 	h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2432 
2433 	if (nodes_allowed != &node_states[N_MEMORY])
2434 		NODEMASK_FREE(nodes_allowed);
2435 
2436 	return len;
2437 out:
2438 	NODEMASK_FREE(nodes_allowed);
2439 	return err;
2440 }
2441 
2442 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2443 					 struct kobject *kobj, const char *buf,
2444 					 size_t len)
2445 {
2446 	struct hstate *h;
2447 	unsigned long count;
2448 	int nid;
2449 	int err;
2450 
2451 	err = kstrtoul(buf, 10, &count);
2452 	if (err)
2453 		return err;
2454 
2455 	h = kobj_to_hstate(kobj, &nid);
2456 	return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2457 }
2458 
2459 static ssize_t nr_hugepages_show(struct kobject *kobj,
2460 				       struct kobj_attribute *attr, char *buf)
2461 {
2462 	return nr_hugepages_show_common(kobj, attr, buf);
2463 }
2464 
2465 static ssize_t nr_hugepages_store(struct kobject *kobj,
2466 	       struct kobj_attribute *attr, const char *buf, size_t len)
2467 {
2468 	return nr_hugepages_store_common(false, kobj, buf, len);
2469 }
2470 HSTATE_ATTR(nr_hugepages);
2471 
2472 #ifdef CONFIG_NUMA
2473 
2474 /*
2475  * hstate attribute for optionally mempolicy-based constraint on persistent
2476  * huge page alloc/free.
2477  */
2478 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2479 				       struct kobj_attribute *attr, char *buf)
2480 {
2481 	return nr_hugepages_show_common(kobj, attr, buf);
2482 }
2483 
2484 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2485 	       struct kobj_attribute *attr, const char *buf, size_t len)
2486 {
2487 	return nr_hugepages_store_common(true, kobj, buf, len);
2488 }
2489 HSTATE_ATTR(nr_hugepages_mempolicy);
2490 #endif
2491 
2492 
2493 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2494 					struct kobj_attribute *attr, char *buf)
2495 {
2496 	struct hstate *h = kobj_to_hstate(kobj, NULL);
2497 	return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2498 }
2499 
2500 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2501 		struct kobj_attribute *attr, const char *buf, size_t count)
2502 {
2503 	int err;
2504 	unsigned long input;
2505 	struct hstate *h = kobj_to_hstate(kobj, NULL);
2506 
2507 	if (hstate_is_gigantic(h))
2508 		return -EINVAL;
2509 
2510 	err = kstrtoul(buf, 10, &input);
2511 	if (err)
2512 		return err;
2513 
2514 	spin_lock(&hugetlb_lock);
2515 	h->nr_overcommit_huge_pages = input;
2516 	spin_unlock(&hugetlb_lock);
2517 
2518 	return count;
2519 }
2520 HSTATE_ATTR(nr_overcommit_hugepages);
2521 
2522 static ssize_t free_hugepages_show(struct kobject *kobj,
2523 					struct kobj_attribute *attr, char *buf)
2524 {
2525 	struct hstate *h;
2526 	unsigned long free_huge_pages;
2527 	int nid;
2528 
2529 	h = kobj_to_hstate(kobj, &nid);
2530 	if (nid == NUMA_NO_NODE)
2531 		free_huge_pages = h->free_huge_pages;
2532 	else
2533 		free_huge_pages = h->free_huge_pages_node[nid];
2534 
2535 	return sprintf(buf, "%lu\n", free_huge_pages);
2536 }
2537 HSTATE_ATTR_RO(free_hugepages);
2538 
2539 static ssize_t resv_hugepages_show(struct kobject *kobj,
2540 					struct kobj_attribute *attr, char *buf)
2541 {
2542 	struct hstate *h = kobj_to_hstate(kobj, NULL);
2543 	return sprintf(buf, "%lu\n", h->resv_huge_pages);
2544 }
2545 HSTATE_ATTR_RO(resv_hugepages);
2546 
2547 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2548 					struct kobj_attribute *attr, char *buf)
2549 {
2550 	struct hstate *h;
2551 	unsigned long surplus_huge_pages;
2552 	int nid;
2553 
2554 	h = kobj_to_hstate(kobj, &nid);
2555 	if (nid == NUMA_NO_NODE)
2556 		surplus_huge_pages = h->surplus_huge_pages;
2557 	else
2558 		surplus_huge_pages = h->surplus_huge_pages_node[nid];
2559 
2560 	return sprintf(buf, "%lu\n", surplus_huge_pages);
2561 }
2562 HSTATE_ATTR_RO(surplus_hugepages);
2563 
2564 static struct attribute *hstate_attrs[] = {
2565 	&nr_hugepages_attr.attr,
2566 	&nr_overcommit_hugepages_attr.attr,
2567 	&free_hugepages_attr.attr,
2568 	&resv_hugepages_attr.attr,
2569 	&surplus_hugepages_attr.attr,
2570 #ifdef CONFIG_NUMA
2571 	&nr_hugepages_mempolicy_attr.attr,
2572 #endif
2573 	NULL,
2574 };
2575 
2576 static const struct attribute_group hstate_attr_group = {
2577 	.attrs = hstate_attrs,
2578 };
2579 
2580 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2581 				    struct kobject **hstate_kobjs,
2582 				    const struct attribute_group *hstate_attr_group)
2583 {
2584 	int retval;
2585 	int hi = hstate_index(h);
2586 
2587 	hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2588 	if (!hstate_kobjs[hi])
2589 		return -ENOMEM;
2590 
2591 	retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2592 	if (retval)
2593 		kobject_put(hstate_kobjs[hi]);
2594 
2595 	return retval;
2596 }
2597 
2598 static void __init hugetlb_sysfs_init(void)
2599 {
2600 	struct hstate *h;
2601 	int err;
2602 
2603 	hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2604 	if (!hugepages_kobj)
2605 		return;
2606 
2607 	for_each_hstate(h) {
2608 		err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2609 					 hstate_kobjs, &hstate_attr_group);
2610 		if (err)
2611 			pr_err("Hugetlb: Unable to add hstate %s", h->name);
2612 	}
2613 }
2614 
2615 #ifdef CONFIG_NUMA
2616 
2617 /*
2618  * node_hstate/s - associate per node hstate attributes, via their kobjects,
2619  * with node devices in node_devices[] using a parallel array.  The array
2620  * index of a node device or _hstate == node id.
2621  * This is here to avoid any static dependency of the node device driver, in
2622  * the base kernel, on the hugetlb module.
2623  */
2624 struct node_hstate {
2625 	struct kobject		*hugepages_kobj;
2626 	struct kobject		*hstate_kobjs[HUGE_MAX_HSTATE];
2627 };
2628 static struct node_hstate node_hstates[MAX_NUMNODES];
2629 
2630 /*
2631  * A subset of global hstate attributes for node devices
2632  */
2633 static struct attribute *per_node_hstate_attrs[] = {
2634 	&nr_hugepages_attr.attr,
2635 	&free_hugepages_attr.attr,
2636 	&surplus_hugepages_attr.attr,
2637 	NULL,
2638 };
2639 
2640 static const struct attribute_group per_node_hstate_attr_group = {
2641 	.attrs = per_node_hstate_attrs,
2642 };
2643 
2644 /*
2645  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2646  * Returns node id via non-NULL nidp.
2647  */
2648 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2649 {
2650 	int nid;
2651 
2652 	for (nid = 0; nid < nr_node_ids; nid++) {
2653 		struct node_hstate *nhs = &node_hstates[nid];
2654 		int i;
2655 		for (i = 0; i < HUGE_MAX_HSTATE; i++)
2656 			if (nhs->hstate_kobjs[i] == kobj) {
2657 				if (nidp)
2658 					*nidp = nid;
2659 				return &hstates[i];
2660 			}
2661 	}
2662 
2663 	BUG();
2664 	return NULL;
2665 }
2666 
2667 /*
2668  * Unregister hstate attributes from a single node device.
2669  * No-op if no hstate attributes attached.
2670  */
2671 static void hugetlb_unregister_node(struct node *node)
2672 {
2673 	struct hstate *h;
2674 	struct node_hstate *nhs = &node_hstates[node->dev.id];
2675 
2676 	if (!nhs->hugepages_kobj)
2677 		return;		/* no hstate attributes */
2678 
2679 	for_each_hstate(h) {
2680 		int idx = hstate_index(h);
2681 		if (nhs->hstate_kobjs[idx]) {
2682 			kobject_put(nhs->hstate_kobjs[idx]);
2683 			nhs->hstate_kobjs[idx] = NULL;
2684 		}
2685 	}
2686 
2687 	kobject_put(nhs->hugepages_kobj);
2688 	nhs->hugepages_kobj = NULL;
2689 }
2690 
2691 
2692 /*
2693  * Register hstate attributes for a single node device.
2694  * No-op if attributes already registered.
2695  */
2696 static void hugetlb_register_node(struct node *node)
2697 {
2698 	struct hstate *h;
2699 	struct node_hstate *nhs = &node_hstates[node->dev.id];
2700 	int err;
2701 
2702 	if (nhs->hugepages_kobj)
2703 		return;		/* already allocated */
2704 
2705 	nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2706 							&node->dev.kobj);
2707 	if (!nhs->hugepages_kobj)
2708 		return;
2709 
2710 	for_each_hstate(h) {
2711 		err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2712 						nhs->hstate_kobjs,
2713 						&per_node_hstate_attr_group);
2714 		if (err) {
2715 			pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2716 				h->name, node->dev.id);
2717 			hugetlb_unregister_node(node);
2718 			break;
2719 		}
2720 	}
2721 }
2722 
2723 /*
2724  * hugetlb init time:  register hstate attributes for all registered node
2725  * devices of nodes that have memory.  All on-line nodes should have
2726  * registered their associated device by this time.
2727  */
2728 static void __init hugetlb_register_all_nodes(void)
2729 {
2730 	int nid;
2731 
2732 	for_each_node_state(nid, N_MEMORY) {
2733 		struct node *node = node_devices[nid];
2734 		if (node->dev.id == nid)
2735 			hugetlb_register_node(node);
2736 	}
2737 
2738 	/*
2739 	 * Let the node device driver know we're here so it can
2740 	 * [un]register hstate attributes on node hotplug.
2741 	 */
2742 	register_hugetlbfs_with_node(hugetlb_register_node,
2743 				     hugetlb_unregister_node);
2744 }
2745 #else	/* !CONFIG_NUMA */
2746 
2747 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2748 {
2749 	BUG();
2750 	if (nidp)
2751 		*nidp = -1;
2752 	return NULL;
2753 }
2754 
2755 static void hugetlb_register_all_nodes(void) { }
2756 
2757 #endif
2758 
2759 static int __init hugetlb_init(void)
2760 {
2761 	int i;
2762 
2763 	if (!hugepages_supported())
2764 		return 0;
2765 
2766 	if (!size_to_hstate(default_hstate_size)) {
2767 		if (default_hstate_size != 0) {
2768 			pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2769 			       default_hstate_size, HPAGE_SIZE);
2770 		}
2771 
2772 		default_hstate_size = HPAGE_SIZE;
2773 		if (!size_to_hstate(default_hstate_size))
2774 			hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2775 	}
2776 	default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2777 	if (default_hstate_max_huge_pages) {
2778 		if (!default_hstate.max_huge_pages)
2779 			default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2780 	}
2781 
2782 	hugetlb_init_hstates();
2783 	gather_bootmem_prealloc();
2784 	report_hugepages();
2785 
2786 	hugetlb_sysfs_init();
2787 	hugetlb_register_all_nodes();
2788 	hugetlb_cgroup_file_init();
2789 
2790 #ifdef CONFIG_SMP
2791 	num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2792 #else
2793 	num_fault_mutexes = 1;
2794 #endif
2795 	hugetlb_fault_mutex_table =
2796 		kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
2797 			      GFP_KERNEL);
2798 	BUG_ON(!hugetlb_fault_mutex_table);
2799 
2800 	for (i = 0; i < num_fault_mutexes; i++)
2801 		mutex_init(&hugetlb_fault_mutex_table[i]);
2802 	return 0;
2803 }
2804 subsys_initcall(hugetlb_init);
2805 
2806 /* Should be called on processing a hugepagesz=... option */
2807 void __init hugetlb_bad_size(void)
2808 {
2809 	parsed_valid_hugepagesz = false;
2810 }
2811 
2812 void __init hugetlb_add_hstate(unsigned int order)
2813 {
2814 	struct hstate *h;
2815 	unsigned long i;
2816 
2817 	if (size_to_hstate(PAGE_SIZE << order)) {
2818 		pr_warn("hugepagesz= specified twice, ignoring\n");
2819 		return;
2820 	}
2821 	BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2822 	BUG_ON(order == 0);
2823 	h = &hstates[hugetlb_max_hstate++];
2824 	h->order = order;
2825 	h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2826 	h->nr_huge_pages = 0;
2827 	h->free_huge_pages = 0;
2828 	for (i = 0; i < MAX_NUMNODES; ++i)
2829 		INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2830 	INIT_LIST_HEAD(&h->hugepage_activelist);
2831 	h->next_nid_to_alloc = first_memory_node;
2832 	h->next_nid_to_free = first_memory_node;
2833 	snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2834 					huge_page_size(h)/1024);
2835 
2836 	parsed_hstate = h;
2837 }
2838 
2839 static int __init hugetlb_nrpages_setup(char *s)
2840 {
2841 	unsigned long *mhp;
2842 	static unsigned long *last_mhp;
2843 
2844 	if (!parsed_valid_hugepagesz) {
2845 		pr_warn("hugepages = %s preceded by "
2846 			"an unsupported hugepagesz, ignoring\n", s);
2847 		parsed_valid_hugepagesz = true;
2848 		return 1;
2849 	}
2850 	/*
2851 	 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2852 	 * so this hugepages= parameter goes to the "default hstate".
2853 	 */
2854 	else if (!hugetlb_max_hstate)
2855 		mhp = &default_hstate_max_huge_pages;
2856 	else
2857 		mhp = &parsed_hstate->max_huge_pages;
2858 
2859 	if (mhp == last_mhp) {
2860 		pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2861 		return 1;
2862 	}
2863 
2864 	if (sscanf(s, "%lu", mhp) <= 0)
2865 		*mhp = 0;
2866 
2867 	/*
2868 	 * Global state is always initialized later in hugetlb_init.
2869 	 * But we need to allocate >= MAX_ORDER hstates here early to still
2870 	 * use the bootmem allocator.
2871 	 */
2872 	if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2873 		hugetlb_hstate_alloc_pages(parsed_hstate);
2874 
2875 	last_mhp = mhp;
2876 
2877 	return 1;
2878 }
2879 __setup("hugepages=", hugetlb_nrpages_setup);
2880 
2881 static int __init hugetlb_default_setup(char *s)
2882 {
2883 	default_hstate_size = memparse(s, &s);
2884 	return 1;
2885 }
2886 __setup("default_hugepagesz=", hugetlb_default_setup);
2887 
2888 static unsigned int cpuset_mems_nr(unsigned int *array)
2889 {
2890 	int node;
2891 	unsigned int nr = 0;
2892 
2893 	for_each_node_mask(node, cpuset_current_mems_allowed)
2894 		nr += array[node];
2895 
2896 	return nr;
2897 }
2898 
2899 #ifdef CONFIG_SYSCTL
2900 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2901 			 struct ctl_table *table, int write,
2902 			 void __user *buffer, size_t *length, loff_t *ppos)
2903 {
2904 	struct hstate *h = &default_hstate;
2905 	unsigned long tmp = h->max_huge_pages;
2906 	int ret;
2907 
2908 	if (!hugepages_supported())
2909 		return -EOPNOTSUPP;
2910 
2911 	table->data = &tmp;
2912 	table->maxlen = sizeof(unsigned long);
2913 	ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2914 	if (ret)
2915 		goto out;
2916 
2917 	if (write)
2918 		ret = __nr_hugepages_store_common(obey_mempolicy, h,
2919 						  NUMA_NO_NODE, tmp, *length);
2920 out:
2921 	return ret;
2922 }
2923 
2924 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2925 			  void __user *buffer, size_t *length, loff_t *ppos)
2926 {
2927 
2928 	return hugetlb_sysctl_handler_common(false, table, write,
2929 							buffer, length, ppos);
2930 }
2931 
2932 #ifdef CONFIG_NUMA
2933 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2934 			  void __user *buffer, size_t *length, loff_t *ppos)
2935 {
2936 	return hugetlb_sysctl_handler_common(true, table, write,
2937 							buffer, length, ppos);
2938 }
2939 #endif /* CONFIG_NUMA */
2940 
2941 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2942 			void __user *buffer,
2943 			size_t *length, loff_t *ppos)
2944 {
2945 	struct hstate *h = &default_hstate;
2946 	unsigned long tmp;
2947 	int ret;
2948 
2949 	if (!hugepages_supported())
2950 		return -EOPNOTSUPP;
2951 
2952 	tmp = h->nr_overcommit_huge_pages;
2953 
2954 	if (write && hstate_is_gigantic(h))
2955 		return -EINVAL;
2956 
2957 	table->data = &tmp;
2958 	table->maxlen = sizeof(unsigned long);
2959 	ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2960 	if (ret)
2961 		goto out;
2962 
2963 	if (write) {
2964 		spin_lock(&hugetlb_lock);
2965 		h->nr_overcommit_huge_pages = tmp;
2966 		spin_unlock(&hugetlb_lock);
2967 	}
2968 out:
2969 	return ret;
2970 }
2971 
2972 #endif /* CONFIG_SYSCTL */
2973 
2974 void hugetlb_report_meminfo(struct seq_file *m)
2975 {
2976 	struct hstate *h;
2977 	unsigned long total = 0;
2978 
2979 	if (!hugepages_supported())
2980 		return;
2981 
2982 	for_each_hstate(h) {
2983 		unsigned long count = h->nr_huge_pages;
2984 
2985 		total += (PAGE_SIZE << huge_page_order(h)) * count;
2986 
2987 		if (h == &default_hstate)
2988 			seq_printf(m,
2989 				   "HugePages_Total:   %5lu\n"
2990 				   "HugePages_Free:    %5lu\n"
2991 				   "HugePages_Rsvd:    %5lu\n"
2992 				   "HugePages_Surp:    %5lu\n"
2993 				   "Hugepagesize:   %8lu kB\n",
2994 				   count,
2995 				   h->free_huge_pages,
2996 				   h->resv_huge_pages,
2997 				   h->surplus_huge_pages,
2998 				   (PAGE_SIZE << huge_page_order(h)) / 1024);
2999 	}
3000 
3001 	seq_printf(m, "Hugetlb:        %8lu kB\n", total / 1024);
3002 }
3003 
3004 int hugetlb_report_node_meminfo(int nid, char *buf)
3005 {
3006 	struct hstate *h = &default_hstate;
3007 	if (!hugepages_supported())
3008 		return 0;
3009 	return sprintf(buf,
3010 		"Node %d HugePages_Total: %5u\n"
3011 		"Node %d HugePages_Free:  %5u\n"
3012 		"Node %d HugePages_Surp:  %5u\n",
3013 		nid, h->nr_huge_pages_node[nid],
3014 		nid, h->free_huge_pages_node[nid],
3015 		nid, h->surplus_huge_pages_node[nid]);
3016 }
3017 
3018 void hugetlb_show_meminfo(void)
3019 {
3020 	struct hstate *h;
3021 	int nid;
3022 
3023 	if (!hugepages_supported())
3024 		return;
3025 
3026 	for_each_node_state(nid, N_MEMORY)
3027 		for_each_hstate(h)
3028 			pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3029 				nid,
3030 				h->nr_huge_pages_node[nid],
3031 				h->free_huge_pages_node[nid],
3032 				h->surplus_huge_pages_node[nid],
3033 				1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3034 }
3035 
3036 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3037 {
3038 	seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3039 		   atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3040 }
3041 
3042 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3043 unsigned long hugetlb_total_pages(void)
3044 {
3045 	struct hstate *h;
3046 	unsigned long nr_total_pages = 0;
3047 
3048 	for_each_hstate(h)
3049 		nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3050 	return nr_total_pages;
3051 }
3052 
3053 static int hugetlb_acct_memory(struct hstate *h, long delta)
3054 {
3055 	int ret = -ENOMEM;
3056 
3057 	spin_lock(&hugetlb_lock);
3058 	/*
3059 	 * When cpuset is configured, it breaks the strict hugetlb page
3060 	 * reservation as the accounting is done on a global variable. Such
3061 	 * reservation is completely rubbish in the presence of cpuset because
3062 	 * the reservation is not checked against page availability for the
3063 	 * current cpuset. Application can still potentially OOM'ed by kernel
3064 	 * with lack of free htlb page in cpuset that the task is in.
3065 	 * Attempt to enforce strict accounting with cpuset is almost
3066 	 * impossible (or too ugly) because cpuset is too fluid that
3067 	 * task or memory node can be dynamically moved between cpusets.
3068 	 *
3069 	 * The change of semantics for shared hugetlb mapping with cpuset is
3070 	 * undesirable. However, in order to preserve some of the semantics,
3071 	 * we fall back to check against current free page availability as
3072 	 * a best attempt and hopefully to minimize the impact of changing
3073 	 * semantics that cpuset has.
3074 	 */
3075 	if (delta > 0) {
3076 		if (gather_surplus_pages(h, delta) < 0)
3077 			goto out;
3078 
3079 		if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3080 			return_unused_surplus_pages(h, delta);
3081 			goto out;
3082 		}
3083 	}
3084 
3085 	ret = 0;
3086 	if (delta < 0)
3087 		return_unused_surplus_pages(h, (unsigned long) -delta);
3088 
3089 out:
3090 	spin_unlock(&hugetlb_lock);
3091 	return ret;
3092 }
3093 
3094 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3095 {
3096 	struct resv_map *resv = vma_resv_map(vma);
3097 
3098 	/*
3099 	 * This new VMA should share its siblings reservation map if present.
3100 	 * The VMA will only ever have a valid reservation map pointer where
3101 	 * it is being copied for another still existing VMA.  As that VMA
3102 	 * has a reference to the reservation map it cannot disappear until
3103 	 * after this open call completes.  It is therefore safe to take a
3104 	 * new reference here without additional locking.
3105 	 */
3106 	if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3107 		kref_get(&resv->refs);
3108 }
3109 
3110 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3111 {
3112 	struct hstate *h = hstate_vma(vma);
3113 	struct resv_map *resv = vma_resv_map(vma);
3114 	struct hugepage_subpool *spool = subpool_vma(vma);
3115 	unsigned long reserve, start, end;
3116 	long gbl_reserve;
3117 
3118 	if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3119 		return;
3120 
3121 	start = vma_hugecache_offset(h, vma, vma->vm_start);
3122 	end = vma_hugecache_offset(h, vma, vma->vm_end);
3123 
3124 	reserve = (end - start) - region_count(resv, start, end);
3125 
3126 	kref_put(&resv->refs, resv_map_release);
3127 
3128 	if (reserve) {
3129 		/*
3130 		 * Decrement reserve counts.  The global reserve count may be
3131 		 * adjusted if the subpool has a minimum size.
3132 		 */
3133 		gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3134 		hugetlb_acct_memory(h, -gbl_reserve);
3135 	}
3136 }
3137 
3138 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3139 {
3140 	if (addr & ~(huge_page_mask(hstate_vma(vma))))
3141 		return -EINVAL;
3142 	return 0;
3143 }
3144 
3145 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3146 {
3147 	struct hstate *hstate = hstate_vma(vma);
3148 
3149 	return 1UL << huge_page_shift(hstate);
3150 }
3151 
3152 /*
3153  * We cannot handle pagefaults against hugetlb pages at all.  They cause
3154  * handle_mm_fault() to try to instantiate regular-sized pages in the
3155  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
3156  * this far.
3157  */
3158 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3159 {
3160 	BUG();
3161 	return 0;
3162 }
3163 
3164 /*
3165  * When a new function is introduced to vm_operations_struct and added
3166  * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3167  * This is because under System V memory model, mappings created via
3168  * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3169  * their original vm_ops are overwritten with shm_vm_ops.
3170  */
3171 const struct vm_operations_struct hugetlb_vm_ops = {
3172 	.fault = hugetlb_vm_op_fault,
3173 	.open = hugetlb_vm_op_open,
3174 	.close = hugetlb_vm_op_close,
3175 	.split = hugetlb_vm_op_split,
3176 	.pagesize = hugetlb_vm_op_pagesize,
3177 };
3178 
3179 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3180 				int writable)
3181 {
3182 	pte_t entry;
3183 
3184 	if (writable) {
3185 		entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3186 					 vma->vm_page_prot)));
3187 	} else {
3188 		entry = huge_pte_wrprotect(mk_huge_pte(page,
3189 					   vma->vm_page_prot));
3190 	}
3191 	entry = pte_mkyoung(entry);
3192 	entry = pte_mkhuge(entry);
3193 	entry = arch_make_huge_pte(entry, vma, page, writable);
3194 
3195 	return entry;
3196 }
3197 
3198 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3199 				   unsigned long address, pte_t *ptep)
3200 {
3201 	pte_t entry;
3202 
3203 	entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3204 	if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3205 		update_mmu_cache(vma, address, ptep);
3206 }
3207 
3208 bool is_hugetlb_entry_migration(pte_t pte)
3209 {
3210 	swp_entry_t swp;
3211 
3212 	if (huge_pte_none(pte) || pte_present(pte))
3213 		return false;
3214 	swp = pte_to_swp_entry(pte);
3215 	if (non_swap_entry(swp) && is_migration_entry(swp))
3216 		return true;
3217 	else
3218 		return false;
3219 }
3220 
3221 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3222 {
3223 	swp_entry_t swp;
3224 
3225 	if (huge_pte_none(pte) || pte_present(pte))
3226 		return 0;
3227 	swp = pte_to_swp_entry(pte);
3228 	if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3229 		return 1;
3230 	else
3231 		return 0;
3232 }
3233 
3234 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3235 			    struct vm_area_struct *vma)
3236 {
3237 	pte_t *src_pte, *dst_pte, entry, dst_entry;
3238 	struct page *ptepage;
3239 	unsigned long addr;
3240 	int cow;
3241 	struct hstate *h = hstate_vma(vma);
3242 	unsigned long sz = huge_page_size(h);
3243 	struct mmu_notifier_range range;
3244 	int ret = 0;
3245 
3246 	cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3247 
3248 	if (cow) {
3249 		mmu_notifier_range_init(&range, src, vma->vm_start,
3250 					vma->vm_end);
3251 		mmu_notifier_invalidate_range_start(&range);
3252 	}
3253 
3254 	for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3255 		spinlock_t *src_ptl, *dst_ptl;
3256 		src_pte = huge_pte_offset(src, addr, sz);
3257 		if (!src_pte)
3258 			continue;
3259 		dst_pte = huge_pte_alloc(dst, addr, sz);
3260 		if (!dst_pte) {
3261 			ret = -ENOMEM;
3262 			break;
3263 		}
3264 
3265 		/*
3266 		 * If the pagetables are shared don't copy or take references.
3267 		 * dst_pte == src_pte is the common case of src/dest sharing.
3268 		 *
3269 		 * However, src could have 'unshared' and dst shares with
3270 		 * another vma.  If dst_pte !none, this implies sharing.
3271 		 * Check here before taking page table lock, and once again
3272 		 * after taking the lock below.
3273 		 */
3274 		dst_entry = huge_ptep_get(dst_pte);
3275 		if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3276 			continue;
3277 
3278 		dst_ptl = huge_pte_lock(h, dst, dst_pte);
3279 		src_ptl = huge_pte_lockptr(h, src, src_pte);
3280 		spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3281 		entry = huge_ptep_get(src_pte);
3282 		dst_entry = huge_ptep_get(dst_pte);
3283 		if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3284 			/*
3285 			 * Skip if src entry none.  Also, skip in the
3286 			 * unlikely case dst entry !none as this implies
3287 			 * sharing with another vma.
3288 			 */
3289 			;
3290 		} else if (unlikely(is_hugetlb_entry_migration(entry) ||
3291 				    is_hugetlb_entry_hwpoisoned(entry))) {
3292 			swp_entry_t swp_entry = pte_to_swp_entry(entry);
3293 
3294 			if (is_write_migration_entry(swp_entry) && cow) {
3295 				/*
3296 				 * COW mappings require pages in both
3297 				 * parent and child to be set to read.
3298 				 */
3299 				make_migration_entry_read(&swp_entry);
3300 				entry = swp_entry_to_pte(swp_entry);
3301 				set_huge_swap_pte_at(src, addr, src_pte,
3302 						     entry, sz);
3303 			}
3304 			set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3305 		} else {
3306 			if (cow) {
3307 				/*
3308 				 * No need to notify as we are downgrading page
3309 				 * table protection not changing it to point
3310 				 * to a new page.
3311 				 *
3312 				 * See Documentation/vm/mmu_notifier.rst
3313 				 */
3314 				huge_ptep_set_wrprotect(src, addr, src_pte);
3315 			}
3316 			entry = huge_ptep_get(src_pte);
3317 			ptepage = pte_page(entry);
3318 			get_page(ptepage);
3319 			page_dup_rmap(ptepage, true);
3320 			set_huge_pte_at(dst, addr, dst_pte, entry);
3321 			hugetlb_count_add(pages_per_huge_page(h), dst);
3322 		}
3323 		spin_unlock(src_ptl);
3324 		spin_unlock(dst_ptl);
3325 	}
3326 
3327 	if (cow)
3328 		mmu_notifier_invalidate_range_end(&range);
3329 
3330 	return ret;
3331 }
3332 
3333 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3334 			    unsigned long start, unsigned long end,
3335 			    struct page *ref_page)
3336 {
3337 	struct mm_struct *mm = vma->vm_mm;
3338 	unsigned long address;
3339 	pte_t *ptep;
3340 	pte_t pte;
3341 	spinlock_t *ptl;
3342 	struct page *page;
3343 	struct hstate *h = hstate_vma(vma);
3344 	unsigned long sz = huge_page_size(h);
3345 	struct mmu_notifier_range range;
3346 
3347 	WARN_ON(!is_vm_hugetlb_page(vma));
3348 	BUG_ON(start & ~huge_page_mask(h));
3349 	BUG_ON(end & ~huge_page_mask(h));
3350 
3351 	/*
3352 	 * This is a hugetlb vma, all the pte entries should point
3353 	 * to huge page.
3354 	 */
3355 	tlb_remove_check_page_size_change(tlb, sz);
3356 	tlb_start_vma(tlb, vma);
3357 
3358 	/*
3359 	 * If sharing possible, alert mmu notifiers of worst case.
3360 	 */
3361 	mmu_notifier_range_init(&range, mm, start, end);
3362 	adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3363 	mmu_notifier_invalidate_range_start(&range);
3364 	address = start;
3365 	for (; address < end; address += sz) {
3366 		ptep = huge_pte_offset(mm, address, sz);
3367 		if (!ptep)
3368 			continue;
3369 
3370 		ptl = huge_pte_lock(h, mm, ptep);
3371 		if (huge_pmd_unshare(mm, &address, ptep)) {
3372 			spin_unlock(ptl);
3373 			/*
3374 			 * We just unmapped a page of PMDs by clearing a PUD.
3375 			 * The caller's TLB flush range should cover this area.
3376 			 */
3377 			continue;
3378 		}
3379 
3380 		pte = huge_ptep_get(ptep);
3381 		if (huge_pte_none(pte)) {
3382 			spin_unlock(ptl);
3383 			continue;
3384 		}
3385 
3386 		/*
3387 		 * Migrating hugepage or HWPoisoned hugepage is already
3388 		 * unmapped and its refcount is dropped, so just clear pte here.
3389 		 */
3390 		if (unlikely(!pte_present(pte))) {
3391 			huge_pte_clear(mm, address, ptep, sz);
3392 			spin_unlock(ptl);
3393 			continue;
3394 		}
3395 
3396 		page = pte_page(pte);
3397 		/*
3398 		 * If a reference page is supplied, it is because a specific
3399 		 * page is being unmapped, not a range. Ensure the page we
3400 		 * are about to unmap is the actual page of interest.
3401 		 */
3402 		if (ref_page) {
3403 			if (page != ref_page) {
3404 				spin_unlock(ptl);
3405 				continue;
3406 			}
3407 			/*
3408 			 * Mark the VMA as having unmapped its page so that
3409 			 * future faults in this VMA will fail rather than
3410 			 * looking like data was lost
3411 			 */
3412 			set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3413 		}
3414 
3415 		pte = huge_ptep_get_and_clear(mm, address, ptep);
3416 		tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3417 		if (huge_pte_dirty(pte))
3418 			set_page_dirty(page);
3419 
3420 		hugetlb_count_sub(pages_per_huge_page(h), mm);
3421 		page_remove_rmap(page, true);
3422 
3423 		spin_unlock(ptl);
3424 		tlb_remove_page_size(tlb, page, huge_page_size(h));
3425 		/*
3426 		 * Bail out after unmapping reference page if supplied
3427 		 */
3428 		if (ref_page)
3429 			break;
3430 	}
3431 	mmu_notifier_invalidate_range_end(&range);
3432 	tlb_end_vma(tlb, vma);
3433 }
3434 
3435 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3436 			  struct vm_area_struct *vma, unsigned long start,
3437 			  unsigned long end, struct page *ref_page)
3438 {
3439 	__unmap_hugepage_range(tlb, vma, start, end, ref_page);
3440 
3441 	/*
3442 	 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3443 	 * test will fail on a vma being torn down, and not grab a page table
3444 	 * on its way out.  We're lucky that the flag has such an appropriate
3445 	 * name, and can in fact be safely cleared here. We could clear it
3446 	 * before the __unmap_hugepage_range above, but all that's necessary
3447 	 * is to clear it before releasing the i_mmap_rwsem. This works
3448 	 * because in the context this is called, the VMA is about to be
3449 	 * destroyed and the i_mmap_rwsem is held.
3450 	 */
3451 	vma->vm_flags &= ~VM_MAYSHARE;
3452 }
3453 
3454 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3455 			  unsigned long end, struct page *ref_page)
3456 {
3457 	struct mm_struct *mm;
3458 	struct mmu_gather tlb;
3459 	unsigned long tlb_start = start;
3460 	unsigned long tlb_end = end;
3461 
3462 	/*
3463 	 * If shared PMDs were possibly used within this vma range, adjust
3464 	 * start/end for worst case tlb flushing.
3465 	 * Note that we can not be sure if PMDs are shared until we try to
3466 	 * unmap pages.  However, we want to make sure TLB flushing covers
3467 	 * the largest possible range.
3468 	 */
3469 	adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
3470 
3471 	mm = vma->vm_mm;
3472 
3473 	tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
3474 	__unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3475 	tlb_finish_mmu(&tlb, tlb_start, tlb_end);
3476 }
3477 
3478 /*
3479  * This is called when the original mapper is failing to COW a MAP_PRIVATE
3480  * mappping it owns the reserve page for. The intention is to unmap the page
3481  * from other VMAs and let the children be SIGKILLed if they are faulting the
3482  * same region.
3483  */
3484 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3485 			      struct page *page, unsigned long address)
3486 {
3487 	struct hstate *h = hstate_vma(vma);
3488 	struct vm_area_struct *iter_vma;
3489 	struct address_space *mapping;
3490 	pgoff_t pgoff;
3491 
3492 	/*
3493 	 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3494 	 * from page cache lookup which is in HPAGE_SIZE units.
3495 	 */
3496 	address = address & huge_page_mask(h);
3497 	pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3498 			vma->vm_pgoff;
3499 	mapping = vma->vm_file->f_mapping;
3500 
3501 	/*
3502 	 * Take the mapping lock for the duration of the table walk. As
3503 	 * this mapping should be shared between all the VMAs,
3504 	 * __unmap_hugepage_range() is called as the lock is already held
3505 	 */
3506 	i_mmap_lock_write(mapping);
3507 	vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3508 		/* Do not unmap the current VMA */
3509 		if (iter_vma == vma)
3510 			continue;
3511 
3512 		/*
3513 		 * Shared VMAs have their own reserves and do not affect
3514 		 * MAP_PRIVATE accounting but it is possible that a shared
3515 		 * VMA is using the same page so check and skip such VMAs.
3516 		 */
3517 		if (iter_vma->vm_flags & VM_MAYSHARE)
3518 			continue;
3519 
3520 		/*
3521 		 * Unmap the page from other VMAs without their own reserves.
3522 		 * They get marked to be SIGKILLed if they fault in these
3523 		 * areas. This is because a future no-page fault on this VMA
3524 		 * could insert a zeroed page instead of the data existing
3525 		 * from the time of fork. This would look like data corruption
3526 		 */
3527 		if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3528 			unmap_hugepage_range(iter_vma, address,
3529 					     address + huge_page_size(h), page);
3530 	}
3531 	i_mmap_unlock_write(mapping);
3532 }
3533 
3534 /*
3535  * Hugetlb_cow() should be called with page lock of the original hugepage held.
3536  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3537  * cannot race with other handlers or page migration.
3538  * Keep the pte_same checks anyway to make transition from the mutex easier.
3539  */
3540 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3541 		       unsigned long address, pte_t *ptep,
3542 		       struct page *pagecache_page, spinlock_t *ptl)
3543 {
3544 	pte_t pte;
3545 	struct hstate *h = hstate_vma(vma);
3546 	struct page *old_page, *new_page;
3547 	int outside_reserve = 0;
3548 	vm_fault_t ret = 0;
3549 	unsigned long haddr = address & huge_page_mask(h);
3550 	struct mmu_notifier_range range;
3551 
3552 	pte = huge_ptep_get(ptep);
3553 	old_page = pte_page(pte);
3554 
3555 retry_avoidcopy:
3556 	/* If no-one else is actually using this page, avoid the copy
3557 	 * and just make the page writable */
3558 	if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3559 		page_move_anon_rmap(old_page, vma);
3560 		set_huge_ptep_writable(vma, haddr, ptep);
3561 		return 0;
3562 	}
3563 
3564 	/*
3565 	 * If the process that created a MAP_PRIVATE mapping is about to
3566 	 * perform a COW due to a shared page count, attempt to satisfy
3567 	 * the allocation without using the existing reserves. The pagecache
3568 	 * page is used to determine if the reserve at this address was
3569 	 * consumed or not. If reserves were used, a partial faulted mapping
3570 	 * at the time of fork() could consume its reserves on COW instead
3571 	 * of the full address range.
3572 	 */
3573 	if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3574 			old_page != pagecache_page)
3575 		outside_reserve = 1;
3576 
3577 	get_page(old_page);
3578 
3579 	/*
3580 	 * Drop page table lock as buddy allocator may be called. It will
3581 	 * be acquired again before returning to the caller, as expected.
3582 	 */
3583 	spin_unlock(ptl);
3584 	new_page = alloc_huge_page(vma, haddr, outside_reserve);
3585 
3586 	if (IS_ERR(new_page)) {
3587 		/*
3588 		 * If a process owning a MAP_PRIVATE mapping fails to COW,
3589 		 * it is due to references held by a child and an insufficient
3590 		 * huge page pool. To guarantee the original mappers
3591 		 * reliability, unmap the page from child processes. The child
3592 		 * may get SIGKILLed if it later faults.
3593 		 */
3594 		if (outside_reserve) {
3595 			put_page(old_page);
3596 			BUG_ON(huge_pte_none(pte));
3597 			unmap_ref_private(mm, vma, old_page, haddr);
3598 			BUG_ON(huge_pte_none(pte));
3599 			spin_lock(ptl);
3600 			ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3601 			if (likely(ptep &&
3602 				   pte_same(huge_ptep_get(ptep), pte)))
3603 				goto retry_avoidcopy;
3604 			/*
3605 			 * race occurs while re-acquiring page table
3606 			 * lock, and our job is done.
3607 			 */
3608 			return 0;
3609 		}
3610 
3611 		ret = vmf_error(PTR_ERR(new_page));
3612 		goto out_release_old;
3613 	}
3614 
3615 	/*
3616 	 * When the original hugepage is shared one, it does not have
3617 	 * anon_vma prepared.
3618 	 */
3619 	if (unlikely(anon_vma_prepare(vma))) {
3620 		ret = VM_FAULT_OOM;
3621 		goto out_release_all;
3622 	}
3623 
3624 	copy_user_huge_page(new_page, old_page, address, vma,
3625 			    pages_per_huge_page(h));
3626 	__SetPageUptodate(new_page);
3627 	set_page_huge_active(new_page);
3628 
3629 	mmu_notifier_range_init(&range, mm, haddr, haddr + huge_page_size(h));
3630 	mmu_notifier_invalidate_range_start(&range);
3631 
3632 	/*
3633 	 * Retake the page table lock to check for racing updates
3634 	 * before the page tables are altered
3635 	 */
3636 	spin_lock(ptl);
3637 	ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3638 	if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3639 		ClearPagePrivate(new_page);
3640 
3641 		/* Break COW */
3642 		huge_ptep_clear_flush(vma, haddr, ptep);
3643 		mmu_notifier_invalidate_range(mm, range.start, range.end);
3644 		set_huge_pte_at(mm, haddr, ptep,
3645 				make_huge_pte(vma, new_page, 1));
3646 		page_remove_rmap(old_page, true);
3647 		hugepage_add_new_anon_rmap(new_page, vma, haddr);
3648 		/* Make the old page be freed below */
3649 		new_page = old_page;
3650 	}
3651 	spin_unlock(ptl);
3652 	mmu_notifier_invalidate_range_end(&range);
3653 out_release_all:
3654 	restore_reserve_on_error(h, vma, haddr, new_page);
3655 	put_page(new_page);
3656 out_release_old:
3657 	put_page(old_page);
3658 
3659 	spin_lock(ptl); /* Caller expects lock to be held */
3660 	return ret;
3661 }
3662 
3663 /* Return the pagecache page at a given address within a VMA */
3664 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3665 			struct vm_area_struct *vma, unsigned long address)
3666 {
3667 	struct address_space *mapping;
3668 	pgoff_t idx;
3669 
3670 	mapping = vma->vm_file->f_mapping;
3671 	idx = vma_hugecache_offset(h, vma, address);
3672 
3673 	return find_lock_page(mapping, idx);
3674 }
3675 
3676 /*
3677  * Return whether there is a pagecache page to back given address within VMA.
3678  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3679  */
3680 static bool hugetlbfs_pagecache_present(struct hstate *h,
3681 			struct vm_area_struct *vma, unsigned long address)
3682 {
3683 	struct address_space *mapping;
3684 	pgoff_t idx;
3685 	struct page *page;
3686 
3687 	mapping = vma->vm_file->f_mapping;
3688 	idx = vma_hugecache_offset(h, vma, address);
3689 
3690 	page = find_get_page(mapping, idx);
3691 	if (page)
3692 		put_page(page);
3693 	return page != NULL;
3694 }
3695 
3696 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3697 			   pgoff_t idx)
3698 {
3699 	struct inode *inode = mapping->host;
3700 	struct hstate *h = hstate_inode(inode);
3701 	int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3702 
3703 	if (err)
3704 		return err;
3705 	ClearPagePrivate(page);
3706 
3707 	/*
3708 	 * set page dirty so that it will not be removed from cache/file
3709 	 * by non-hugetlbfs specific code paths.
3710 	 */
3711 	set_page_dirty(page);
3712 
3713 	spin_lock(&inode->i_lock);
3714 	inode->i_blocks += blocks_per_huge_page(h);
3715 	spin_unlock(&inode->i_lock);
3716 	return 0;
3717 }
3718 
3719 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
3720 			struct vm_area_struct *vma,
3721 			struct address_space *mapping, pgoff_t idx,
3722 			unsigned long address, pte_t *ptep, unsigned int flags)
3723 {
3724 	struct hstate *h = hstate_vma(vma);
3725 	vm_fault_t ret = VM_FAULT_SIGBUS;
3726 	int anon_rmap = 0;
3727 	unsigned long size;
3728 	struct page *page;
3729 	pte_t new_pte;
3730 	spinlock_t *ptl;
3731 	unsigned long haddr = address & huge_page_mask(h);
3732 
3733 	/*
3734 	 * Currently, we are forced to kill the process in the event the
3735 	 * original mapper has unmapped pages from the child due to a failed
3736 	 * COW. Warn that such a situation has occurred as it may not be obvious
3737 	 */
3738 	if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3739 		pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3740 			   current->pid);
3741 		return ret;
3742 	}
3743 
3744 	/*
3745 	 * Use page lock to guard against racing truncation
3746 	 * before we get page_table_lock.
3747 	 */
3748 retry:
3749 	page = find_lock_page(mapping, idx);
3750 	if (!page) {
3751 		size = i_size_read(mapping->host) >> huge_page_shift(h);
3752 		if (idx >= size)
3753 			goto out;
3754 
3755 		/*
3756 		 * Check for page in userfault range
3757 		 */
3758 		if (userfaultfd_missing(vma)) {
3759 			u32 hash;
3760 			struct vm_fault vmf = {
3761 				.vma = vma,
3762 				.address = haddr,
3763 				.flags = flags,
3764 				/*
3765 				 * Hard to debug if it ends up being
3766 				 * used by a callee that assumes
3767 				 * something about the other
3768 				 * uninitialized fields... same as in
3769 				 * memory.c
3770 				 */
3771 			};
3772 
3773 			/*
3774 			 * hugetlb_fault_mutex must be dropped before
3775 			 * handling userfault.  Reacquire after handling
3776 			 * fault to make calling code simpler.
3777 			 */
3778 			hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping,
3779 							idx, haddr);
3780 			mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3781 			ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3782 			mutex_lock(&hugetlb_fault_mutex_table[hash]);
3783 			goto out;
3784 		}
3785 
3786 		page = alloc_huge_page(vma, haddr, 0);
3787 		if (IS_ERR(page)) {
3788 			ret = vmf_error(PTR_ERR(page));
3789 			goto out;
3790 		}
3791 		clear_huge_page(page, address, pages_per_huge_page(h));
3792 		__SetPageUptodate(page);
3793 		set_page_huge_active(page);
3794 
3795 		if (vma->vm_flags & VM_MAYSHARE) {
3796 			int err = huge_add_to_page_cache(page, mapping, idx);
3797 			if (err) {
3798 				put_page(page);
3799 				if (err == -EEXIST)
3800 					goto retry;
3801 				goto out;
3802 			}
3803 		} else {
3804 			lock_page(page);
3805 			if (unlikely(anon_vma_prepare(vma))) {
3806 				ret = VM_FAULT_OOM;
3807 				goto backout_unlocked;
3808 			}
3809 			anon_rmap = 1;
3810 		}
3811 	} else {
3812 		/*
3813 		 * If memory error occurs between mmap() and fault, some process
3814 		 * don't have hwpoisoned swap entry for errored virtual address.
3815 		 * So we need to block hugepage fault by PG_hwpoison bit check.
3816 		 */
3817 		if (unlikely(PageHWPoison(page))) {
3818 			ret = VM_FAULT_HWPOISON |
3819 				VM_FAULT_SET_HINDEX(hstate_index(h));
3820 			goto backout_unlocked;
3821 		}
3822 	}
3823 
3824 	/*
3825 	 * If we are going to COW a private mapping later, we examine the
3826 	 * pending reservations for this page now. This will ensure that
3827 	 * any allocations necessary to record that reservation occur outside
3828 	 * the spinlock.
3829 	 */
3830 	if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3831 		if (vma_needs_reservation(h, vma, haddr) < 0) {
3832 			ret = VM_FAULT_OOM;
3833 			goto backout_unlocked;
3834 		}
3835 		/* Just decrements count, does not deallocate */
3836 		vma_end_reservation(h, vma, haddr);
3837 	}
3838 
3839 	ptl = huge_pte_lock(h, mm, ptep);
3840 	size = i_size_read(mapping->host) >> huge_page_shift(h);
3841 	if (idx >= size)
3842 		goto backout;
3843 
3844 	ret = 0;
3845 	if (!huge_pte_none(huge_ptep_get(ptep)))
3846 		goto backout;
3847 
3848 	if (anon_rmap) {
3849 		ClearPagePrivate(page);
3850 		hugepage_add_new_anon_rmap(page, vma, haddr);
3851 	} else
3852 		page_dup_rmap(page, true);
3853 	new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3854 				&& (vma->vm_flags & VM_SHARED)));
3855 	set_huge_pte_at(mm, haddr, ptep, new_pte);
3856 
3857 	hugetlb_count_add(pages_per_huge_page(h), mm);
3858 	if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3859 		/* Optimization, do the COW without a second fault */
3860 		ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
3861 	}
3862 
3863 	spin_unlock(ptl);
3864 	unlock_page(page);
3865 out:
3866 	return ret;
3867 
3868 backout:
3869 	spin_unlock(ptl);
3870 backout_unlocked:
3871 	unlock_page(page);
3872 	restore_reserve_on_error(h, vma, haddr, page);
3873 	put_page(page);
3874 	goto out;
3875 }
3876 
3877 #ifdef CONFIG_SMP
3878 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3879 			    struct vm_area_struct *vma,
3880 			    struct address_space *mapping,
3881 			    pgoff_t idx, unsigned long address)
3882 {
3883 	unsigned long key[2];
3884 	u32 hash;
3885 
3886 	if (vma->vm_flags & VM_SHARED) {
3887 		key[0] = (unsigned long) mapping;
3888 		key[1] = idx;
3889 	} else {
3890 		key[0] = (unsigned long) mm;
3891 		key[1] = address >> huge_page_shift(h);
3892 	}
3893 
3894 	hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3895 
3896 	return hash & (num_fault_mutexes - 1);
3897 }
3898 #else
3899 /*
3900  * For uniprocesor systems we always use a single mutex, so just
3901  * return 0 and avoid the hashing overhead.
3902  */
3903 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3904 			    struct vm_area_struct *vma,
3905 			    struct address_space *mapping,
3906 			    pgoff_t idx, unsigned long address)
3907 {
3908 	return 0;
3909 }
3910 #endif
3911 
3912 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3913 			unsigned long address, unsigned int flags)
3914 {
3915 	pte_t *ptep, entry;
3916 	spinlock_t *ptl;
3917 	vm_fault_t ret;
3918 	u32 hash;
3919 	pgoff_t idx;
3920 	struct page *page = NULL;
3921 	struct page *pagecache_page = NULL;
3922 	struct hstate *h = hstate_vma(vma);
3923 	struct address_space *mapping;
3924 	int need_wait_lock = 0;
3925 	unsigned long haddr = address & huge_page_mask(h);
3926 
3927 	ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3928 	if (ptep) {
3929 		entry = huge_ptep_get(ptep);
3930 		if (unlikely(is_hugetlb_entry_migration(entry))) {
3931 			migration_entry_wait_huge(vma, mm, ptep);
3932 			return 0;
3933 		} else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3934 			return VM_FAULT_HWPOISON_LARGE |
3935 				VM_FAULT_SET_HINDEX(hstate_index(h));
3936 	} else {
3937 		ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
3938 		if (!ptep)
3939 			return VM_FAULT_OOM;
3940 	}
3941 
3942 	mapping = vma->vm_file->f_mapping;
3943 	idx = vma_hugecache_offset(h, vma, haddr);
3944 
3945 	/*
3946 	 * Serialize hugepage allocation and instantiation, so that we don't
3947 	 * get spurious allocation failures if two CPUs race to instantiate
3948 	 * the same page in the page cache.
3949 	 */
3950 	hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, haddr);
3951 	mutex_lock(&hugetlb_fault_mutex_table[hash]);
3952 
3953 	entry = huge_ptep_get(ptep);
3954 	if (huge_pte_none(entry)) {
3955 		ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3956 		goto out_mutex;
3957 	}
3958 
3959 	ret = 0;
3960 
3961 	/*
3962 	 * entry could be a migration/hwpoison entry at this point, so this
3963 	 * check prevents the kernel from going below assuming that we have
3964 	 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3965 	 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3966 	 * handle it.
3967 	 */
3968 	if (!pte_present(entry))
3969 		goto out_mutex;
3970 
3971 	/*
3972 	 * If we are going to COW the mapping later, we examine the pending
3973 	 * reservations for this page now. This will ensure that any
3974 	 * allocations necessary to record that reservation occur outside the
3975 	 * spinlock. For private mappings, we also lookup the pagecache
3976 	 * page now as it is used to determine if a reservation has been
3977 	 * consumed.
3978 	 */
3979 	if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3980 		if (vma_needs_reservation(h, vma, haddr) < 0) {
3981 			ret = VM_FAULT_OOM;
3982 			goto out_mutex;
3983 		}
3984 		/* Just decrements count, does not deallocate */
3985 		vma_end_reservation(h, vma, haddr);
3986 
3987 		if (!(vma->vm_flags & VM_MAYSHARE))
3988 			pagecache_page = hugetlbfs_pagecache_page(h,
3989 								vma, haddr);
3990 	}
3991 
3992 	ptl = huge_pte_lock(h, mm, ptep);
3993 
3994 	/* Check for a racing update before calling hugetlb_cow */
3995 	if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3996 		goto out_ptl;
3997 
3998 	/*
3999 	 * hugetlb_cow() requires page locks of pte_page(entry) and
4000 	 * pagecache_page, so here we need take the former one
4001 	 * when page != pagecache_page or !pagecache_page.
4002 	 */
4003 	page = pte_page(entry);
4004 	if (page != pagecache_page)
4005 		if (!trylock_page(page)) {
4006 			need_wait_lock = 1;
4007 			goto out_ptl;
4008 		}
4009 
4010 	get_page(page);
4011 
4012 	if (flags & FAULT_FLAG_WRITE) {
4013 		if (!huge_pte_write(entry)) {
4014 			ret = hugetlb_cow(mm, vma, address, ptep,
4015 					  pagecache_page, ptl);
4016 			goto out_put_page;
4017 		}
4018 		entry = huge_pte_mkdirty(entry);
4019 	}
4020 	entry = pte_mkyoung(entry);
4021 	if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4022 						flags & FAULT_FLAG_WRITE))
4023 		update_mmu_cache(vma, haddr, ptep);
4024 out_put_page:
4025 	if (page != pagecache_page)
4026 		unlock_page(page);
4027 	put_page(page);
4028 out_ptl:
4029 	spin_unlock(ptl);
4030 
4031 	if (pagecache_page) {
4032 		unlock_page(pagecache_page);
4033 		put_page(pagecache_page);
4034 	}
4035 out_mutex:
4036 	mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4037 	/*
4038 	 * Generally it's safe to hold refcount during waiting page lock. But
4039 	 * here we just wait to defer the next page fault to avoid busy loop and
4040 	 * the page is not used after unlocked before returning from the current
4041 	 * page fault. So we are safe from accessing freed page, even if we wait
4042 	 * here without taking refcount.
4043 	 */
4044 	if (need_wait_lock)
4045 		wait_on_page_locked(page);
4046 	return ret;
4047 }
4048 
4049 /*
4050  * Used by userfaultfd UFFDIO_COPY.  Based on mcopy_atomic_pte with
4051  * modifications for huge pages.
4052  */
4053 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4054 			    pte_t *dst_pte,
4055 			    struct vm_area_struct *dst_vma,
4056 			    unsigned long dst_addr,
4057 			    unsigned long src_addr,
4058 			    struct page **pagep)
4059 {
4060 	struct address_space *mapping;
4061 	pgoff_t idx;
4062 	unsigned long size;
4063 	int vm_shared = dst_vma->vm_flags & VM_SHARED;
4064 	struct hstate *h = hstate_vma(dst_vma);
4065 	pte_t _dst_pte;
4066 	spinlock_t *ptl;
4067 	int ret;
4068 	struct page *page;
4069 
4070 	if (!*pagep) {
4071 		ret = -ENOMEM;
4072 		page = alloc_huge_page(dst_vma, dst_addr, 0);
4073 		if (IS_ERR(page))
4074 			goto out;
4075 
4076 		ret = copy_huge_page_from_user(page,
4077 						(const void __user *) src_addr,
4078 						pages_per_huge_page(h), false);
4079 
4080 		/* fallback to copy_from_user outside mmap_sem */
4081 		if (unlikely(ret)) {
4082 			ret = -ENOENT;
4083 			*pagep = page;
4084 			/* don't free the page */
4085 			goto out;
4086 		}
4087 	} else {
4088 		page = *pagep;
4089 		*pagep = NULL;
4090 	}
4091 
4092 	/*
4093 	 * The memory barrier inside __SetPageUptodate makes sure that
4094 	 * preceding stores to the page contents become visible before
4095 	 * the set_pte_at() write.
4096 	 */
4097 	__SetPageUptodate(page);
4098 	set_page_huge_active(page);
4099 
4100 	mapping = dst_vma->vm_file->f_mapping;
4101 	idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4102 
4103 	/*
4104 	 * If shared, add to page cache
4105 	 */
4106 	if (vm_shared) {
4107 		size = i_size_read(mapping->host) >> huge_page_shift(h);
4108 		ret = -EFAULT;
4109 		if (idx >= size)
4110 			goto out_release_nounlock;
4111 
4112 		/*
4113 		 * Serialization between remove_inode_hugepages() and
4114 		 * huge_add_to_page_cache() below happens through the
4115 		 * hugetlb_fault_mutex_table that here must be hold by
4116 		 * the caller.
4117 		 */
4118 		ret = huge_add_to_page_cache(page, mapping, idx);
4119 		if (ret)
4120 			goto out_release_nounlock;
4121 	}
4122 
4123 	ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4124 	spin_lock(ptl);
4125 
4126 	/*
4127 	 * Recheck the i_size after holding PT lock to make sure not
4128 	 * to leave any page mapped (as page_mapped()) beyond the end
4129 	 * of the i_size (remove_inode_hugepages() is strict about
4130 	 * enforcing that). If we bail out here, we'll also leave a
4131 	 * page in the radix tree in the vm_shared case beyond the end
4132 	 * of the i_size, but remove_inode_hugepages() will take care
4133 	 * of it as soon as we drop the hugetlb_fault_mutex_table.
4134 	 */
4135 	size = i_size_read(mapping->host) >> huge_page_shift(h);
4136 	ret = -EFAULT;
4137 	if (idx >= size)
4138 		goto out_release_unlock;
4139 
4140 	ret = -EEXIST;
4141 	if (!huge_pte_none(huge_ptep_get(dst_pte)))
4142 		goto out_release_unlock;
4143 
4144 	if (vm_shared) {
4145 		page_dup_rmap(page, true);
4146 	} else {
4147 		ClearPagePrivate(page);
4148 		hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4149 	}
4150 
4151 	_dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4152 	if (dst_vma->vm_flags & VM_WRITE)
4153 		_dst_pte = huge_pte_mkdirty(_dst_pte);
4154 	_dst_pte = pte_mkyoung(_dst_pte);
4155 
4156 	set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4157 
4158 	(void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4159 					dst_vma->vm_flags & VM_WRITE);
4160 	hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4161 
4162 	/* No need to invalidate - it was non-present before */
4163 	update_mmu_cache(dst_vma, dst_addr, dst_pte);
4164 
4165 	spin_unlock(ptl);
4166 	if (vm_shared)
4167 		unlock_page(page);
4168 	ret = 0;
4169 out:
4170 	return ret;
4171 out_release_unlock:
4172 	spin_unlock(ptl);
4173 	if (vm_shared)
4174 		unlock_page(page);
4175 out_release_nounlock:
4176 	put_page(page);
4177 	goto out;
4178 }
4179 
4180 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4181 			 struct page **pages, struct vm_area_struct **vmas,
4182 			 unsigned long *position, unsigned long *nr_pages,
4183 			 long i, unsigned int flags, int *nonblocking)
4184 {
4185 	unsigned long pfn_offset;
4186 	unsigned long vaddr = *position;
4187 	unsigned long remainder = *nr_pages;
4188 	struct hstate *h = hstate_vma(vma);
4189 	int err = -EFAULT;
4190 
4191 	while (vaddr < vma->vm_end && remainder) {
4192 		pte_t *pte;
4193 		spinlock_t *ptl = NULL;
4194 		int absent;
4195 		struct page *page;
4196 
4197 		/*
4198 		 * If we have a pending SIGKILL, don't keep faulting pages and
4199 		 * potentially allocating memory.
4200 		 */
4201 		if (fatal_signal_pending(current)) {
4202 			remainder = 0;
4203 			break;
4204 		}
4205 
4206 		/*
4207 		 * Some archs (sparc64, sh*) have multiple pte_ts to
4208 		 * each hugepage.  We have to make sure we get the
4209 		 * first, for the page indexing below to work.
4210 		 *
4211 		 * Note that page table lock is not held when pte is null.
4212 		 */
4213 		pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4214 				      huge_page_size(h));
4215 		if (pte)
4216 			ptl = huge_pte_lock(h, mm, pte);
4217 		absent = !pte || huge_pte_none(huge_ptep_get(pte));
4218 
4219 		/*
4220 		 * When coredumping, it suits get_dump_page if we just return
4221 		 * an error where there's an empty slot with no huge pagecache
4222 		 * to back it.  This way, we avoid allocating a hugepage, and
4223 		 * the sparse dumpfile avoids allocating disk blocks, but its
4224 		 * huge holes still show up with zeroes where they need to be.
4225 		 */
4226 		if (absent && (flags & FOLL_DUMP) &&
4227 		    !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4228 			if (pte)
4229 				spin_unlock(ptl);
4230 			remainder = 0;
4231 			break;
4232 		}
4233 
4234 		/*
4235 		 * We need call hugetlb_fault for both hugepages under migration
4236 		 * (in which case hugetlb_fault waits for the migration,) and
4237 		 * hwpoisoned hugepages (in which case we need to prevent the
4238 		 * caller from accessing to them.) In order to do this, we use
4239 		 * here is_swap_pte instead of is_hugetlb_entry_migration and
4240 		 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4241 		 * both cases, and because we can't follow correct pages
4242 		 * directly from any kind of swap entries.
4243 		 */
4244 		if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4245 		    ((flags & FOLL_WRITE) &&
4246 		      !huge_pte_write(huge_ptep_get(pte)))) {
4247 			vm_fault_t ret;
4248 			unsigned int fault_flags = 0;
4249 
4250 			if (pte)
4251 				spin_unlock(ptl);
4252 			if (flags & FOLL_WRITE)
4253 				fault_flags |= FAULT_FLAG_WRITE;
4254 			if (nonblocking)
4255 				fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4256 			if (flags & FOLL_NOWAIT)
4257 				fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4258 					FAULT_FLAG_RETRY_NOWAIT;
4259 			if (flags & FOLL_TRIED) {
4260 				VM_WARN_ON_ONCE(fault_flags &
4261 						FAULT_FLAG_ALLOW_RETRY);
4262 				fault_flags |= FAULT_FLAG_TRIED;
4263 			}
4264 			ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4265 			if (ret & VM_FAULT_ERROR) {
4266 				err = vm_fault_to_errno(ret, flags);
4267 				remainder = 0;
4268 				break;
4269 			}
4270 			if (ret & VM_FAULT_RETRY) {
4271 				if (nonblocking)
4272 					*nonblocking = 0;
4273 				*nr_pages = 0;
4274 				/*
4275 				 * VM_FAULT_RETRY must not return an
4276 				 * error, it will return zero
4277 				 * instead.
4278 				 *
4279 				 * No need to update "position" as the
4280 				 * caller will not check it after
4281 				 * *nr_pages is set to 0.
4282 				 */
4283 				return i;
4284 			}
4285 			continue;
4286 		}
4287 
4288 		pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4289 		page = pte_page(huge_ptep_get(pte));
4290 same_page:
4291 		if (pages) {
4292 			pages[i] = mem_map_offset(page, pfn_offset);
4293 			get_page(pages[i]);
4294 		}
4295 
4296 		if (vmas)
4297 			vmas[i] = vma;
4298 
4299 		vaddr += PAGE_SIZE;
4300 		++pfn_offset;
4301 		--remainder;
4302 		++i;
4303 		if (vaddr < vma->vm_end && remainder &&
4304 				pfn_offset < pages_per_huge_page(h)) {
4305 			/*
4306 			 * We use pfn_offset to avoid touching the pageframes
4307 			 * of this compound page.
4308 			 */
4309 			goto same_page;
4310 		}
4311 		spin_unlock(ptl);
4312 	}
4313 	*nr_pages = remainder;
4314 	/*
4315 	 * setting position is actually required only if remainder is
4316 	 * not zero but it's faster not to add a "if (remainder)"
4317 	 * branch.
4318 	 */
4319 	*position = vaddr;
4320 
4321 	return i ? i : err;
4322 }
4323 
4324 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4325 /*
4326  * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4327  * implement this.
4328  */
4329 #define flush_hugetlb_tlb_range(vma, addr, end)	flush_tlb_range(vma, addr, end)
4330 #endif
4331 
4332 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4333 		unsigned long address, unsigned long end, pgprot_t newprot)
4334 {
4335 	struct mm_struct *mm = vma->vm_mm;
4336 	unsigned long start = address;
4337 	pte_t *ptep;
4338 	pte_t pte;
4339 	struct hstate *h = hstate_vma(vma);
4340 	unsigned long pages = 0;
4341 	bool shared_pmd = false;
4342 	struct mmu_notifier_range range;
4343 
4344 	/*
4345 	 * In the case of shared PMDs, the area to flush could be beyond
4346 	 * start/end.  Set range.start/range.end to cover the maximum possible
4347 	 * range if PMD sharing is possible.
4348 	 */
4349 	mmu_notifier_range_init(&range, mm, start, end);
4350 	adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4351 
4352 	BUG_ON(address >= end);
4353 	flush_cache_range(vma, range.start, range.end);
4354 
4355 	mmu_notifier_invalidate_range_start(&range);
4356 	i_mmap_lock_write(vma->vm_file->f_mapping);
4357 	for (; address < end; address += huge_page_size(h)) {
4358 		spinlock_t *ptl;
4359 		ptep = huge_pte_offset(mm, address, huge_page_size(h));
4360 		if (!ptep)
4361 			continue;
4362 		ptl = huge_pte_lock(h, mm, ptep);
4363 		if (huge_pmd_unshare(mm, &address, ptep)) {
4364 			pages++;
4365 			spin_unlock(ptl);
4366 			shared_pmd = true;
4367 			continue;
4368 		}
4369 		pte = huge_ptep_get(ptep);
4370 		if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4371 			spin_unlock(ptl);
4372 			continue;
4373 		}
4374 		if (unlikely(is_hugetlb_entry_migration(pte))) {
4375 			swp_entry_t entry = pte_to_swp_entry(pte);
4376 
4377 			if (is_write_migration_entry(entry)) {
4378 				pte_t newpte;
4379 
4380 				make_migration_entry_read(&entry);
4381 				newpte = swp_entry_to_pte(entry);
4382 				set_huge_swap_pte_at(mm, address, ptep,
4383 						     newpte, huge_page_size(h));
4384 				pages++;
4385 			}
4386 			spin_unlock(ptl);
4387 			continue;
4388 		}
4389 		if (!huge_pte_none(pte)) {
4390 			pte = huge_ptep_get_and_clear(mm, address, ptep);
4391 			pte = pte_mkhuge(huge_pte_modify(pte, newprot));
4392 			pte = arch_make_huge_pte(pte, vma, NULL, 0);
4393 			set_huge_pte_at(mm, address, ptep, pte);
4394 			pages++;
4395 		}
4396 		spin_unlock(ptl);
4397 	}
4398 	/*
4399 	 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4400 	 * may have cleared our pud entry and done put_page on the page table:
4401 	 * once we release i_mmap_rwsem, another task can do the final put_page
4402 	 * and that page table be reused and filled with junk.  If we actually
4403 	 * did unshare a page of pmds, flush the range corresponding to the pud.
4404 	 */
4405 	if (shared_pmd)
4406 		flush_hugetlb_tlb_range(vma, range.start, range.end);
4407 	else
4408 		flush_hugetlb_tlb_range(vma, start, end);
4409 	/*
4410 	 * No need to call mmu_notifier_invalidate_range() we are downgrading
4411 	 * page table protection not changing it to point to a new page.
4412 	 *
4413 	 * See Documentation/vm/mmu_notifier.rst
4414 	 */
4415 	i_mmap_unlock_write(vma->vm_file->f_mapping);
4416 	mmu_notifier_invalidate_range_end(&range);
4417 
4418 	return pages << h->order;
4419 }
4420 
4421 int hugetlb_reserve_pages(struct inode *inode,
4422 					long from, long to,
4423 					struct vm_area_struct *vma,
4424 					vm_flags_t vm_flags)
4425 {
4426 	long ret, chg;
4427 	struct hstate *h = hstate_inode(inode);
4428 	struct hugepage_subpool *spool = subpool_inode(inode);
4429 	struct resv_map *resv_map;
4430 	long gbl_reserve;
4431 
4432 	/* This should never happen */
4433 	if (from > to) {
4434 		VM_WARN(1, "%s called with a negative range\n", __func__);
4435 		return -EINVAL;
4436 	}
4437 
4438 	/*
4439 	 * Only apply hugepage reservation if asked. At fault time, an
4440 	 * attempt will be made for VM_NORESERVE to allocate a page
4441 	 * without using reserves
4442 	 */
4443 	if (vm_flags & VM_NORESERVE)
4444 		return 0;
4445 
4446 	/*
4447 	 * Shared mappings base their reservation on the number of pages that
4448 	 * are already allocated on behalf of the file. Private mappings need
4449 	 * to reserve the full area even if read-only as mprotect() may be
4450 	 * called to make the mapping read-write. Assume !vma is a shm mapping
4451 	 */
4452 	if (!vma || vma->vm_flags & VM_MAYSHARE) {
4453 		resv_map = inode_resv_map(inode);
4454 
4455 		chg = region_chg(resv_map, from, to);
4456 
4457 	} else {
4458 		resv_map = resv_map_alloc();
4459 		if (!resv_map)
4460 			return -ENOMEM;
4461 
4462 		chg = to - from;
4463 
4464 		set_vma_resv_map(vma, resv_map);
4465 		set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4466 	}
4467 
4468 	if (chg < 0) {
4469 		ret = chg;
4470 		goto out_err;
4471 	}
4472 
4473 	/*
4474 	 * There must be enough pages in the subpool for the mapping. If
4475 	 * the subpool has a minimum size, there may be some global
4476 	 * reservations already in place (gbl_reserve).
4477 	 */
4478 	gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4479 	if (gbl_reserve < 0) {
4480 		ret = -ENOSPC;
4481 		goto out_err;
4482 	}
4483 
4484 	/*
4485 	 * Check enough hugepages are available for the reservation.
4486 	 * Hand the pages back to the subpool if there are not
4487 	 */
4488 	ret = hugetlb_acct_memory(h, gbl_reserve);
4489 	if (ret < 0) {
4490 		/* put back original number of pages, chg */
4491 		(void)hugepage_subpool_put_pages(spool, chg);
4492 		goto out_err;
4493 	}
4494 
4495 	/*
4496 	 * Account for the reservations made. Shared mappings record regions
4497 	 * that have reservations as they are shared by multiple VMAs.
4498 	 * When the last VMA disappears, the region map says how much
4499 	 * the reservation was and the page cache tells how much of
4500 	 * the reservation was consumed. Private mappings are per-VMA and
4501 	 * only the consumed reservations are tracked. When the VMA
4502 	 * disappears, the original reservation is the VMA size and the
4503 	 * consumed reservations are stored in the map. Hence, nothing
4504 	 * else has to be done for private mappings here
4505 	 */
4506 	if (!vma || vma->vm_flags & VM_MAYSHARE) {
4507 		long add = region_add(resv_map, from, to);
4508 
4509 		if (unlikely(chg > add)) {
4510 			/*
4511 			 * pages in this range were added to the reserve
4512 			 * map between region_chg and region_add.  This
4513 			 * indicates a race with alloc_huge_page.  Adjust
4514 			 * the subpool and reserve counts modified above
4515 			 * based on the difference.
4516 			 */
4517 			long rsv_adjust;
4518 
4519 			rsv_adjust = hugepage_subpool_put_pages(spool,
4520 								chg - add);
4521 			hugetlb_acct_memory(h, -rsv_adjust);
4522 		}
4523 	}
4524 	return 0;
4525 out_err:
4526 	if (!vma || vma->vm_flags & VM_MAYSHARE)
4527 		/* Don't call region_abort if region_chg failed */
4528 		if (chg >= 0)
4529 			region_abort(resv_map, from, to);
4530 	if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4531 		kref_put(&resv_map->refs, resv_map_release);
4532 	return ret;
4533 }
4534 
4535 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4536 								long freed)
4537 {
4538 	struct hstate *h = hstate_inode(inode);
4539 	struct resv_map *resv_map = inode_resv_map(inode);
4540 	long chg = 0;
4541 	struct hugepage_subpool *spool = subpool_inode(inode);
4542 	long gbl_reserve;
4543 
4544 	if (resv_map) {
4545 		chg = region_del(resv_map, start, end);
4546 		/*
4547 		 * region_del() can fail in the rare case where a region
4548 		 * must be split and another region descriptor can not be
4549 		 * allocated.  If end == LONG_MAX, it will not fail.
4550 		 */
4551 		if (chg < 0)
4552 			return chg;
4553 	}
4554 
4555 	spin_lock(&inode->i_lock);
4556 	inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4557 	spin_unlock(&inode->i_lock);
4558 
4559 	/*
4560 	 * If the subpool has a minimum size, the number of global
4561 	 * reservations to be released may be adjusted.
4562 	 */
4563 	gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4564 	hugetlb_acct_memory(h, -gbl_reserve);
4565 
4566 	return 0;
4567 }
4568 
4569 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4570 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4571 				struct vm_area_struct *vma,
4572 				unsigned long addr, pgoff_t idx)
4573 {
4574 	unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4575 				svma->vm_start;
4576 	unsigned long sbase = saddr & PUD_MASK;
4577 	unsigned long s_end = sbase + PUD_SIZE;
4578 
4579 	/* Allow segments to share if only one is marked locked */
4580 	unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4581 	unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4582 
4583 	/*
4584 	 * match the virtual addresses, permission and the alignment of the
4585 	 * page table page.
4586 	 */
4587 	if (pmd_index(addr) != pmd_index(saddr) ||
4588 	    vm_flags != svm_flags ||
4589 	    sbase < svma->vm_start || svma->vm_end < s_end)
4590 		return 0;
4591 
4592 	return saddr;
4593 }
4594 
4595 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4596 {
4597 	unsigned long base = addr & PUD_MASK;
4598 	unsigned long end = base + PUD_SIZE;
4599 
4600 	/*
4601 	 * check on proper vm_flags and page table alignment
4602 	 */
4603 	if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
4604 		return true;
4605 	return false;
4606 }
4607 
4608 /*
4609  * Determine if start,end range within vma could be mapped by shared pmd.
4610  * If yes, adjust start and end to cover range associated with possible
4611  * shared pmd mappings.
4612  */
4613 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4614 				unsigned long *start, unsigned long *end)
4615 {
4616 	unsigned long check_addr = *start;
4617 
4618 	if (!(vma->vm_flags & VM_MAYSHARE))
4619 		return;
4620 
4621 	for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
4622 		unsigned long a_start = check_addr & PUD_MASK;
4623 		unsigned long a_end = a_start + PUD_SIZE;
4624 
4625 		/*
4626 		 * If sharing is possible, adjust start/end if necessary.
4627 		 */
4628 		if (range_in_vma(vma, a_start, a_end)) {
4629 			if (a_start < *start)
4630 				*start = a_start;
4631 			if (a_end > *end)
4632 				*end = a_end;
4633 		}
4634 	}
4635 }
4636 
4637 /*
4638  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4639  * and returns the corresponding pte. While this is not necessary for the
4640  * !shared pmd case because we can allocate the pmd later as well, it makes the
4641  * code much cleaner. pmd allocation is essential for the shared case because
4642  * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4643  * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4644  * bad pmd for sharing.
4645  */
4646 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4647 {
4648 	struct vm_area_struct *vma = find_vma(mm, addr);
4649 	struct address_space *mapping = vma->vm_file->f_mapping;
4650 	pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4651 			vma->vm_pgoff;
4652 	struct vm_area_struct *svma;
4653 	unsigned long saddr;
4654 	pte_t *spte = NULL;
4655 	pte_t *pte;
4656 	spinlock_t *ptl;
4657 
4658 	if (!vma_shareable(vma, addr))
4659 		return (pte_t *)pmd_alloc(mm, pud, addr);
4660 
4661 	i_mmap_lock_write(mapping);
4662 	vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4663 		if (svma == vma)
4664 			continue;
4665 
4666 		saddr = page_table_shareable(svma, vma, addr, idx);
4667 		if (saddr) {
4668 			spte = huge_pte_offset(svma->vm_mm, saddr,
4669 					       vma_mmu_pagesize(svma));
4670 			if (spte) {
4671 				get_page(virt_to_page(spte));
4672 				break;
4673 			}
4674 		}
4675 	}
4676 
4677 	if (!spte)
4678 		goto out;
4679 
4680 	ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4681 	if (pud_none(*pud)) {
4682 		pud_populate(mm, pud,
4683 				(pmd_t *)((unsigned long)spte & PAGE_MASK));
4684 		mm_inc_nr_pmds(mm);
4685 	} else {
4686 		put_page(virt_to_page(spte));
4687 	}
4688 	spin_unlock(ptl);
4689 out:
4690 	pte = (pte_t *)pmd_alloc(mm, pud, addr);
4691 	i_mmap_unlock_write(mapping);
4692 	return pte;
4693 }
4694 
4695 /*
4696  * unmap huge page backed by shared pte.
4697  *
4698  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
4699  * indicated by page_count > 1, unmap is achieved by clearing pud and
4700  * decrementing the ref count. If count == 1, the pte page is not shared.
4701  *
4702  * called with page table lock held.
4703  *
4704  * returns: 1 successfully unmapped a shared pte page
4705  *	    0 the underlying pte page is not shared, or it is the last user
4706  */
4707 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4708 {
4709 	pgd_t *pgd = pgd_offset(mm, *addr);
4710 	p4d_t *p4d = p4d_offset(pgd, *addr);
4711 	pud_t *pud = pud_offset(p4d, *addr);
4712 
4713 	BUG_ON(page_count(virt_to_page(ptep)) == 0);
4714 	if (page_count(virt_to_page(ptep)) == 1)
4715 		return 0;
4716 
4717 	pud_clear(pud);
4718 	put_page(virt_to_page(ptep));
4719 	mm_dec_nr_pmds(mm);
4720 	*addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4721 	return 1;
4722 }
4723 #define want_pmd_share()	(1)
4724 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4725 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4726 {
4727 	return NULL;
4728 }
4729 
4730 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4731 {
4732 	return 0;
4733 }
4734 
4735 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4736 				unsigned long *start, unsigned long *end)
4737 {
4738 }
4739 #define want_pmd_share()	(0)
4740 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4741 
4742 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4743 pte_t *huge_pte_alloc(struct mm_struct *mm,
4744 			unsigned long addr, unsigned long sz)
4745 {
4746 	pgd_t *pgd;
4747 	p4d_t *p4d;
4748 	pud_t *pud;
4749 	pte_t *pte = NULL;
4750 
4751 	pgd = pgd_offset(mm, addr);
4752 	p4d = p4d_alloc(mm, pgd, addr);
4753 	if (!p4d)
4754 		return NULL;
4755 	pud = pud_alloc(mm, p4d, addr);
4756 	if (pud) {
4757 		if (sz == PUD_SIZE) {
4758 			pte = (pte_t *)pud;
4759 		} else {
4760 			BUG_ON(sz != PMD_SIZE);
4761 			if (want_pmd_share() && pud_none(*pud))
4762 				pte = huge_pmd_share(mm, addr, pud);
4763 			else
4764 				pte = (pte_t *)pmd_alloc(mm, pud, addr);
4765 		}
4766 	}
4767 	BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4768 
4769 	return pte;
4770 }
4771 
4772 /*
4773  * huge_pte_offset() - Walk the page table to resolve the hugepage
4774  * entry at address @addr
4775  *
4776  * Return: Pointer to page table or swap entry (PUD or PMD) for
4777  * address @addr, or NULL if a p*d_none() entry is encountered and the
4778  * size @sz doesn't match the hugepage size at this level of the page
4779  * table.
4780  */
4781 pte_t *huge_pte_offset(struct mm_struct *mm,
4782 		       unsigned long addr, unsigned long sz)
4783 {
4784 	pgd_t *pgd;
4785 	p4d_t *p4d;
4786 	pud_t *pud;
4787 	pmd_t *pmd;
4788 
4789 	pgd = pgd_offset(mm, addr);
4790 	if (!pgd_present(*pgd))
4791 		return NULL;
4792 	p4d = p4d_offset(pgd, addr);
4793 	if (!p4d_present(*p4d))
4794 		return NULL;
4795 
4796 	pud = pud_offset(p4d, addr);
4797 	if (sz != PUD_SIZE && pud_none(*pud))
4798 		return NULL;
4799 	/* hugepage or swap? */
4800 	if (pud_huge(*pud) || !pud_present(*pud))
4801 		return (pte_t *)pud;
4802 
4803 	pmd = pmd_offset(pud, addr);
4804 	if (sz != PMD_SIZE && pmd_none(*pmd))
4805 		return NULL;
4806 	/* hugepage or swap? */
4807 	if (pmd_huge(*pmd) || !pmd_present(*pmd))
4808 		return (pte_t *)pmd;
4809 
4810 	return NULL;
4811 }
4812 
4813 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4814 
4815 /*
4816  * These functions are overwritable if your architecture needs its own
4817  * behavior.
4818  */
4819 struct page * __weak
4820 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4821 			      int write)
4822 {
4823 	return ERR_PTR(-EINVAL);
4824 }
4825 
4826 struct page * __weak
4827 follow_huge_pd(struct vm_area_struct *vma,
4828 	       unsigned long address, hugepd_t hpd, int flags, int pdshift)
4829 {
4830 	WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4831 	return NULL;
4832 }
4833 
4834 struct page * __weak
4835 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4836 		pmd_t *pmd, int flags)
4837 {
4838 	struct page *page = NULL;
4839 	spinlock_t *ptl;
4840 	pte_t pte;
4841 retry:
4842 	ptl = pmd_lockptr(mm, pmd);
4843 	spin_lock(ptl);
4844 	/*
4845 	 * make sure that the address range covered by this pmd is not
4846 	 * unmapped from other threads.
4847 	 */
4848 	if (!pmd_huge(*pmd))
4849 		goto out;
4850 	pte = huge_ptep_get((pte_t *)pmd);
4851 	if (pte_present(pte)) {
4852 		page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4853 		if (flags & FOLL_GET)
4854 			get_page(page);
4855 	} else {
4856 		if (is_hugetlb_entry_migration(pte)) {
4857 			spin_unlock(ptl);
4858 			__migration_entry_wait(mm, (pte_t *)pmd, ptl);
4859 			goto retry;
4860 		}
4861 		/*
4862 		 * hwpoisoned entry is treated as no_page_table in
4863 		 * follow_page_mask().
4864 		 */
4865 	}
4866 out:
4867 	spin_unlock(ptl);
4868 	return page;
4869 }
4870 
4871 struct page * __weak
4872 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4873 		pud_t *pud, int flags)
4874 {
4875 	if (flags & FOLL_GET)
4876 		return NULL;
4877 
4878 	return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4879 }
4880 
4881 struct page * __weak
4882 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
4883 {
4884 	if (flags & FOLL_GET)
4885 		return NULL;
4886 
4887 	return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
4888 }
4889 
4890 bool isolate_huge_page(struct page *page, struct list_head *list)
4891 {
4892 	bool ret = true;
4893 
4894 	VM_BUG_ON_PAGE(!PageHead(page), page);
4895 	spin_lock(&hugetlb_lock);
4896 	if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4897 		ret = false;
4898 		goto unlock;
4899 	}
4900 	clear_page_huge_active(page);
4901 	list_move_tail(&page->lru, list);
4902 unlock:
4903 	spin_unlock(&hugetlb_lock);
4904 	return ret;
4905 }
4906 
4907 void putback_active_hugepage(struct page *page)
4908 {
4909 	VM_BUG_ON_PAGE(!PageHead(page), page);
4910 	spin_lock(&hugetlb_lock);
4911 	set_page_huge_active(page);
4912 	list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4913 	spin_unlock(&hugetlb_lock);
4914 	put_page(page);
4915 }
4916 
4917 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
4918 {
4919 	struct hstate *h = page_hstate(oldpage);
4920 
4921 	hugetlb_cgroup_migrate(oldpage, newpage);
4922 	set_page_owner_migrate_reason(newpage, reason);
4923 
4924 	/*
4925 	 * transfer temporary state of the new huge page. This is
4926 	 * reverse to other transitions because the newpage is going to
4927 	 * be final while the old one will be freed so it takes over
4928 	 * the temporary status.
4929 	 *
4930 	 * Also note that we have to transfer the per-node surplus state
4931 	 * here as well otherwise the global surplus count will not match
4932 	 * the per-node's.
4933 	 */
4934 	if (PageHugeTemporary(newpage)) {
4935 		int old_nid = page_to_nid(oldpage);
4936 		int new_nid = page_to_nid(newpage);
4937 
4938 		SetPageHugeTemporary(oldpage);
4939 		ClearPageHugeTemporary(newpage);
4940 
4941 		spin_lock(&hugetlb_lock);
4942 		if (h->surplus_huge_pages_node[old_nid]) {
4943 			h->surplus_huge_pages_node[old_nid]--;
4944 			h->surplus_huge_pages_node[new_nid]++;
4945 		}
4946 		spin_unlock(&hugetlb_lock);
4947 	}
4948 }
4949