xref: /openbmc/linux/mm/hugetlb.c (revision b34e08d5)
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/module.h>
8 #include <linux/mm.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/bootmem.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/rmap.h>
23 #include <linux/swap.h>
24 #include <linux/swapops.h>
25 #include <linux/page-isolation.h>
26 #include <linux/jhash.h>
27 
28 #include <asm/page.h>
29 #include <asm/pgtable.h>
30 #include <asm/tlb.h>
31 
32 #include <linux/io.h>
33 #include <linux/hugetlb.h>
34 #include <linux/hugetlb_cgroup.h>
35 #include <linux/node.h>
36 #include "internal.h"
37 
38 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
39 unsigned long hugepages_treat_as_movable;
40 
41 int hugetlb_max_hstate __read_mostly;
42 unsigned int default_hstate_idx;
43 struct hstate hstates[HUGE_MAX_HSTATE];
44 
45 __initdata LIST_HEAD(huge_boot_pages);
46 
47 /* for command line parsing */
48 static struct hstate * __initdata parsed_hstate;
49 static unsigned long __initdata default_hstate_max_huge_pages;
50 static unsigned long __initdata default_hstate_size;
51 
52 /*
53  * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
54  * free_huge_pages, and surplus_huge_pages.
55  */
56 DEFINE_SPINLOCK(hugetlb_lock);
57 
58 /*
59  * Serializes faults on the same logical page.  This is used to
60  * prevent spurious OOMs when the hugepage pool is fully utilized.
61  */
62 static int num_fault_mutexes;
63 static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp;
64 
65 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
66 {
67 	bool free = (spool->count == 0) && (spool->used_hpages == 0);
68 
69 	spin_unlock(&spool->lock);
70 
71 	/* If no pages are used, and no other handles to the subpool
72 	 * remain, free the subpool the subpool remain */
73 	if (free)
74 		kfree(spool);
75 }
76 
77 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
78 {
79 	struct hugepage_subpool *spool;
80 
81 	spool = kmalloc(sizeof(*spool), GFP_KERNEL);
82 	if (!spool)
83 		return NULL;
84 
85 	spin_lock_init(&spool->lock);
86 	spool->count = 1;
87 	spool->max_hpages = nr_blocks;
88 	spool->used_hpages = 0;
89 
90 	return spool;
91 }
92 
93 void hugepage_put_subpool(struct hugepage_subpool *spool)
94 {
95 	spin_lock(&spool->lock);
96 	BUG_ON(!spool->count);
97 	spool->count--;
98 	unlock_or_release_subpool(spool);
99 }
100 
101 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
102 				      long delta)
103 {
104 	int ret = 0;
105 
106 	if (!spool)
107 		return 0;
108 
109 	spin_lock(&spool->lock);
110 	if ((spool->used_hpages + delta) <= spool->max_hpages) {
111 		spool->used_hpages += delta;
112 	} else {
113 		ret = -ENOMEM;
114 	}
115 	spin_unlock(&spool->lock);
116 
117 	return ret;
118 }
119 
120 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
121 				       long delta)
122 {
123 	if (!spool)
124 		return;
125 
126 	spin_lock(&spool->lock);
127 	spool->used_hpages -= delta;
128 	/* If hugetlbfs_put_super couldn't free spool due to
129 	* an outstanding quota reference, free it now. */
130 	unlock_or_release_subpool(spool);
131 }
132 
133 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
134 {
135 	return HUGETLBFS_SB(inode->i_sb)->spool;
136 }
137 
138 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
139 {
140 	return subpool_inode(file_inode(vma->vm_file));
141 }
142 
143 /*
144  * Region tracking -- allows tracking of reservations and instantiated pages
145  *                    across the pages in a mapping.
146  *
147  * The region data structures are embedded into a resv_map and
148  * protected by a resv_map's lock
149  */
150 struct file_region {
151 	struct list_head link;
152 	long from;
153 	long to;
154 };
155 
156 static long region_add(struct resv_map *resv, long f, long t)
157 {
158 	struct list_head *head = &resv->regions;
159 	struct file_region *rg, *nrg, *trg;
160 
161 	spin_lock(&resv->lock);
162 	/* Locate the region we are either in or before. */
163 	list_for_each_entry(rg, head, link)
164 		if (f <= rg->to)
165 			break;
166 
167 	/* Round our left edge to the current segment if it encloses us. */
168 	if (f > rg->from)
169 		f = rg->from;
170 
171 	/* Check for and consume any regions we now overlap with. */
172 	nrg = rg;
173 	list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
174 		if (&rg->link == head)
175 			break;
176 		if (rg->from > t)
177 			break;
178 
179 		/* If this area reaches higher then extend our area to
180 		 * include it completely.  If this is not the first area
181 		 * which we intend to reuse, free it. */
182 		if (rg->to > t)
183 			t = rg->to;
184 		if (rg != nrg) {
185 			list_del(&rg->link);
186 			kfree(rg);
187 		}
188 	}
189 	nrg->from = f;
190 	nrg->to = t;
191 	spin_unlock(&resv->lock);
192 	return 0;
193 }
194 
195 static long region_chg(struct resv_map *resv, long f, long t)
196 {
197 	struct list_head *head = &resv->regions;
198 	struct file_region *rg, *nrg = NULL;
199 	long chg = 0;
200 
201 retry:
202 	spin_lock(&resv->lock);
203 	/* Locate the region we are before or in. */
204 	list_for_each_entry(rg, head, link)
205 		if (f <= rg->to)
206 			break;
207 
208 	/* If we are below the current region then a new region is required.
209 	 * Subtle, allocate a new region at the position but make it zero
210 	 * size such that we can guarantee to record the reservation. */
211 	if (&rg->link == head || t < rg->from) {
212 		if (!nrg) {
213 			spin_unlock(&resv->lock);
214 			nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
215 			if (!nrg)
216 				return -ENOMEM;
217 
218 			nrg->from = f;
219 			nrg->to   = f;
220 			INIT_LIST_HEAD(&nrg->link);
221 			goto retry;
222 		}
223 
224 		list_add(&nrg->link, rg->link.prev);
225 		chg = t - f;
226 		goto out_nrg;
227 	}
228 
229 	/* Round our left edge to the current segment if it encloses us. */
230 	if (f > rg->from)
231 		f = rg->from;
232 	chg = t - f;
233 
234 	/* Check for and consume any regions we now overlap with. */
235 	list_for_each_entry(rg, rg->link.prev, link) {
236 		if (&rg->link == head)
237 			break;
238 		if (rg->from > t)
239 			goto out;
240 
241 		/* We overlap with this area, if it extends further than
242 		 * us then we must extend ourselves.  Account for its
243 		 * existing reservation. */
244 		if (rg->to > t) {
245 			chg += rg->to - t;
246 			t = rg->to;
247 		}
248 		chg -= rg->to - rg->from;
249 	}
250 
251 out:
252 	spin_unlock(&resv->lock);
253 	/*  We already know we raced and no longer need the new region */
254 	kfree(nrg);
255 	return chg;
256 out_nrg:
257 	spin_unlock(&resv->lock);
258 	return chg;
259 }
260 
261 static long region_truncate(struct resv_map *resv, long end)
262 {
263 	struct list_head *head = &resv->regions;
264 	struct file_region *rg, *trg;
265 	long chg = 0;
266 
267 	spin_lock(&resv->lock);
268 	/* Locate the region we are either in or before. */
269 	list_for_each_entry(rg, head, link)
270 		if (end <= rg->to)
271 			break;
272 	if (&rg->link == head)
273 		goto out;
274 
275 	/* If we are in the middle of a region then adjust it. */
276 	if (end > rg->from) {
277 		chg = rg->to - end;
278 		rg->to = end;
279 		rg = list_entry(rg->link.next, typeof(*rg), link);
280 	}
281 
282 	/* Drop any remaining regions. */
283 	list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
284 		if (&rg->link == head)
285 			break;
286 		chg += rg->to - rg->from;
287 		list_del(&rg->link);
288 		kfree(rg);
289 	}
290 
291 out:
292 	spin_unlock(&resv->lock);
293 	return chg;
294 }
295 
296 static long region_count(struct resv_map *resv, long f, long t)
297 {
298 	struct list_head *head = &resv->regions;
299 	struct file_region *rg;
300 	long chg = 0;
301 
302 	spin_lock(&resv->lock);
303 	/* Locate each segment we overlap with, and count that overlap. */
304 	list_for_each_entry(rg, head, link) {
305 		long seg_from;
306 		long seg_to;
307 
308 		if (rg->to <= f)
309 			continue;
310 		if (rg->from >= t)
311 			break;
312 
313 		seg_from = max(rg->from, f);
314 		seg_to = min(rg->to, t);
315 
316 		chg += seg_to - seg_from;
317 	}
318 	spin_unlock(&resv->lock);
319 
320 	return chg;
321 }
322 
323 /*
324  * Convert the address within this vma to the page offset within
325  * the mapping, in pagecache page units; huge pages here.
326  */
327 static pgoff_t vma_hugecache_offset(struct hstate *h,
328 			struct vm_area_struct *vma, unsigned long address)
329 {
330 	return ((address - vma->vm_start) >> huge_page_shift(h)) +
331 			(vma->vm_pgoff >> huge_page_order(h));
332 }
333 
334 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
335 				     unsigned long address)
336 {
337 	return vma_hugecache_offset(hstate_vma(vma), vma, address);
338 }
339 
340 /*
341  * Return the size of the pages allocated when backing a VMA. In the majority
342  * cases this will be same size as used by the page table entries.
343  */
344 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
345 {
346 	struct hstate *hstate;
347 
348 	if (!is_vm_hugetlb_page(vma))
349 		return PAGE_SIZE;
350 
351 	hstate = hstate_vma(vma);
352 
353 	return 1UL << huge_page_shift(hstate);
354 }
355 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
356 
357 /*
358  * Return the page size being used by the MMU to back a VMA. In the majority
359  * of cases, the page size used by the kernel matches the MMU size. On
360  * architectures where it differs, an architecture-specific version of this
361  * function is required.
362  */
363 #ifndef vma_mmu_pagesize
364 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
365 {
366 	return vma_kernel_pagesize(vma);
367 }
368 #endif
369 
370 /*
371  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
372  * bits of the reservation map pointer, which are always clear due to
373  * alignment.
374  */
375 #define HPAGE_RESV_OWNER    (1UL << 0)
376 #define HPAGE_RESV_UNMAPPED (1UL << 1)
377 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
378 
379 /*
380  * These helpers are used to track how many pages are reserved for
381  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
382  * is guaranteed to have their future faults succeed.
383  *
384  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
385  * the reserve counters are updated with the hugetlb_lock held. It is safe
386  * to reset the VMA at fork() time as it is not in use yet and there is no
387  * chance of the global counters getting corrupted as a result of the values.
388  *
389  * The private mapping reservation is represented in a subtly different
390  * manner to a shared mapping.  A shared mapping has a region map associated
391  * with the underlying file, this region map represents the backing file
392  * pages which have ever had a reservation assigned which this persists even
393  * after the page is instantiated.  A private mapping has a region map
394  * associated with the original mmap which is attached to all VMAs which
395  * reference it, this region map represents those offsets which have consumed
396  * reservation ie. where pages have been instantiated.
397  */
398 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
399 {
400 	return (unsigned long)vma->vm_private_data;
401 }
402 
403 static void set_vma_private_data(struct vm_area_struct *vma,
404 							unsigned long value)
405 {
406 	vma->vm_private_data = (void *)value;
407 }
408 
409 struct resv_map *resv_map_alloc(void)
410 {
411 	struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
412 	if (!resv_map)
413 		return NULL;
414 
415 	kref_init(&resv_map->refs);
416 	spin_lock_init(&resv_map->lock);
417 	INIT_LIST_HEAD(&resv_map->regions);
418 
419 	return resv_map;
420 }
421 
422 void resv_map_release(struct kref *ref)
423 {
424 	struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
425 
426 	/* Clear out any active regions before we release the map. */
427 	region_truncate(resv_map, 0);
428 	kfree(resv_map);
429 }
430 
431 static inline struct resv_map *inode_resv_map(struct inode *inode)
432 {
433 	return inode->i_mapping->private_data;
434 }
435 
436 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
437 {
438 	VM_BUG_ON(!is_vm_hugetlb_page(vma));
439 	if (vma->vm_flags & VM_MAYSHARE) {
440 		struct address_space *mapping = vma->vm_file->f_mapping;
441 		struct inode *inode = mapping->host;
442 
443 		return inode_resv_map(inode);
444 
445 	} else {
446 		return (struct resv_map *)(get_vma_private_data(vma) &
447 							~HPAGE_RESV_MASK);
448 	}
449 }
450 
451 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
452 {
453 	VM_BUG_ON(!is_vm_hugetlb_page(vma));
454 	VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
455 
456 	set_vma_private_data(vma, (get_vma_private_data(vma) &
457 				HPAGE_RESV_MASK) | (unsigned long)map);
458 }
459 
460 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
461 {
462 	VM_BUG_ON(!is_vm_hugetlb_page(vma));
463 	VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
464 
465 	set_vma_private_data(vma, get_vma_private_data(vma) | flags);
466 }
467 
468 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
469 {
470 	VM_BUG_ON(!is_vm_hugetlb_page(vma));
471 
472 	return (get_vma_private_data(vma) & flag) != 0;
473 }
474 
475 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
476 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
477 {
478 	VM_BUG_ON(!is_vm_hugetlb_page(vma));
479 	if (!(vma->vm_flags & VM_MAYSHARE))
480 		vma->vm_private_data = (void *)0;
481 }
482 
483 /* Returns true if the VMA has associated reserve pages */
484 static int vma_has_reserves(struct vm_area_struct *vma, long chg)
485 {
486 	if (vma->vm_flags & VM_NORESERVE) {
487 		/*
488 		 * This address is already reserved by other process(chg == 0),
489 		 * so, we should decrement reserved count. Without decrementing,
490 		 * reserve count remains after releasing inode, because this
491 		 * allocated page will go into page cache and is regarded as
492 		 * coming from reserved pool in releasing step.  Currently, we
493 		 * don't have any other solution to deal with this situation
494 		 * properly, so add work-around here.
495 		 */
496 		if (vma->vm_flags & VM_MAYSHARE && chg == 0)
497 			return 1;
498 		else
499 			return 0;
500 	}
501 
502 	/* Shared mappings always use reserves */
503 	if (vma->vm_flags & VM_MAYSHARE)
504 		return 1;
505 
506 	/*
507 	 * Only the process that called mmap() has reserves for
508 	 * private mappings.
509 	 */
510 	if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
511 		return 1;
512 
513 	return 0;
514 }
515 
516 static void enqueue_huge_page(struct hstate *h, struct page *page)
517 {
518 	int nid = page_to_nid(page);
519 	list_move(&page->lru, &h->hugepage_freelists[nid]);
520 	h->free_huge_pages++;
521 	h->free_huge_pages_node[nid]++;
522 }
523 
524 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
525 {
526 	struct page *page;
527 
528 	list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
529 		if (!is_migrate_isolate_page(page))
530 			break;
531 	/*
532 	 * if 'non-isolated free hugepage' not found on the list,
533 	 * the allocation fails.
534 	 */
535 	if (&h->hugepage_freelists[nid] == &page->lru)
536 		return NULL;
537 	list_move(&page->lru, &h->hugepage_activelist);
538 	set_page_refcounted(page);
539 	h->free_huge_pages--;
540 	h->free_huge_pages_node[nid]--;
541 	return page;
542 }
543 
544 /* Movability of hugepages depends on migration support. */
545 static inline gfp_t htlb_alloc_mask(struct hstate *h)
546 {
547 	if (hugepages_treat_as_movable || hugepage_migration_support(h))
548 		return GFP_HIGHUSER_MOVABLE;
549 	else
550 		return GFP_HIGHUSER;
551 }
552 
553 static struct page *dequeue_huge_page_vma(struct hstate *h,
554 				struct vm_area_struct *vma,
555 				unsigned long address, int avoid_reserve,
556 				long chg)
557 {
558 	struct page *page = NULL;
559 	struct mempolicy *mpol;
560 	nodemask_t *nodemask;
561 	struct zonelist *zonelist;
562 	struct zone *zone;
563 	struct zoneref *z;
564 	unsigned int cpuset_mems_cookie;
565 
566 	/*
567 	 * A child process with MAP_PRIVATE mappings created by their parent
568 	 * have no page reserves. This check ensures that reservations are
569 	 * not "stolen". The child may still get SIGKILLed
570 	 */
571 	if (!vma_has_reserves(vma, chg) &&
572 			h->free_huge_pages - h->resv_huge_pages == 0)
573 		goto err;
574 
575 	/* If reserves cannot be used, ensure enough pages are in the pool */
576 	if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
577 		goto err;
578 
579 retry_cpuset:
580 	cpuset_mems_cookie = read_mems_allowed_begin();
581 	zonelist = huge_zonelist(vma, address,
582 					htlb_alloc_mask(h), &mpol, &nodemask);
583 
584 	for_each_zone_zonelist_nodemask(zone, z, zonelist,
585 						MAX_NR_ZONES - 1, nodemask) {
586 		if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask(h))) {
587 			page = dequeue_huge_page_node(h, zone_to_nid(zone));
588 			if (page) {
589 				if (avoid_reserve)
590 					break;
591 				if (!vma_has_reserves(vma, chg))
592 					break;
593 
594 				SetPagePrivate(page);
595 				h->resv_huge_pages--;
596 				break;
597 			}
598 		}
599 	}
600 
601 	mpol_cond_put(mpol);
602 	if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
603 		goto retry_cpuset;
604 	return page;
605 
606 err:
607 	return NULL;
608 }
609 
610 static void update_and_free_page(struct hstate *h, struct page *page)
611 {
612 	int i;
613 
614 	VM_BUG_ON(h->order >= MAX_ORDER);
615 
616 	h->nr_huge_pages--;
617 	h->nr_huge_pages_node[page_to_nid(page)]--;
618 	for (i = 0; i < pages_per_huge_page(h); i++) {
619 		page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
620 				1 << PG_referenced | 1 << PG_dirty |
621 				1 << PG_active | 1 << PG_reserved |
622 				1 << PG_private | 1 << PG_writeback);
623 	}
624 	VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
625 	set_compound_page_dtor(page, NULL);
626 	set_page_refcounted(page);
627 	arch_release_hugepage(page);
628 	__free_pages(page, huge_page_order(h));
629 }
630 
631 struct hstate *size_to_hstate(unsigned long size)
632 {
633 	struct hstate *h;
634 
635 	for_each_hstate(h) {
636 		if (huge_page_size(h) == size)
637 			return h;
638 	}
639 	return NULL;
640 }
641 
642 static void free_huge_page(struct page *page)
643 {
644 	/*
645 	 * Can't pass hstate in here because it is called from the
646 	 * compound page destructor.
647 	 */
648 	struct hstate *h = page_hstate(page);
649 	int nid = page_to_nid(page);
650 	struct hugepage_subpool *spool =
651 		(struct hugepage_subpool *)page_private(page);
652 	bool restore_reserve;
653 
654 	set_page_private(page, 0);
655 	page->mapping = NULL;
656 	BUG_ON(page_count(page));
657 	BUG_ON(page_mapcount(page));
658 	restore_reserve = PagePrivate(page);
659 	ClearPagePrivate(page);
660 
661 	spin_lock(&hugetlb_lock);
662 	hugetlb_cgroup_uncharge_page(hstate_index(h),
663 				     pages_per_huge_page(h), page);
664 	if (restore_reserve)
665 		h->resv_huge_pages++;
666 
667 	if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
668 		/* remove the page from active list */
669 		list_del(&page->lru);
670 		update_and_free_page(h, page);
671 		h->surplus_huge_pages--;
672 		h->surplus_huge_pages_node[nid]--;
673 	} else {
674 		arch_clear_hugepage_flags(page);
675 		enqueue_huge_page(h, page);
676 	}
677 	spin_unlock(&hugetlb_lock);
678 	hugepage_subpool_put_pages(spool, 1);
679 }
680 
681 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
682 {
683 	INIT_LIST_HEAD(&page->lru);
684 	set_compound_page_dtor(page, free_huge_page);
685 	spin_lock(&hugetlb_lock);
686 	set_hugetlb_cgroup(page, NULL);
687 	h->nr_huge_pages++;
688 	h->nr_huge_pages_node[nid]++;
689 	spin_unlock(&hugetlb_lock);
690 	put_page(page); /* free it into the hugepage allocator */
691 }
692 
693 static void __init prep_compound_gigantic_page(struct page *page,
694 					       unsigned long order)
695 {
696 	int i;
697 	int nr_pages = 1 << order;
698 	struct page *p = page + 1;
699 
700 	/* we rely on prep_new_huge_page to set the destructor */
701 	set_compound_order(page, order);
702 	__SetPageHead(page);
703 	__ClearPageReserved(page);
704 	for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
705 		__SetPageTail(p);
706 		/*
707 		 * For gigantic hugepages allocated through bootmem at
708 		 * boot, it's safer to be consistent with the not-gigantic
709 		 * hugepages and clear the PG_reserved bit from all tail pages
710 		 * too.  Otherwse drivers using get_user_pages() to access tail
711 		 * pages may get the reference counting wrong if they see
712 		 * PG_reserved set on a tail page (despite the head page not
713 		 * having PG_reserved set).  Enforcing this consistency between
714 		 * head and tail pages allows drivers to optimize away a check
715 		 * on the head page when they need know if put_page() is needed
716 		 * after get_user_pages().
717 		 */
718 		__ClearPageReserved(p);
719 		set_page_count(p, 0);
720 		p->first_page = page;
721 	}
722 }
723 
724 /*
725  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
726  * transparent huge pages.  See the PageTransHuge() documentation for more
727  * details.
728  */
729 int PageHuge(struct page *page)
730 {
731 	if (!PageCompound(page))
732 		return 0;
733 
734 	page = compound_head(page);
735 	return get_compound_page_dtor(page) == free_huge_page;
736 }
737 EXPORT_SYMBOL_GPL(PageHuge);
738 
739 /*
740  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
741  * normal or transparent huge pages.
742  */
743 int PageHeadHuge(struct page *page_head)
744 {
745 	if (!PageHead(page_head))
746 		return 0;
747 
748 	return get_compound_page_dtor(page_head) == free_huge_page;
749 }
750 
751 pgoff_t __basepage_index(struct page *page)
752 {
753 	struct page *page_head = compound_head(page);
754 	pgoff_t index = page_index(page_head);
755 	unsigned long compound_idx;
756 
757 	if (!PageHuge(page_head))
758 		return page_index(page);
759 
760 	if (compound_order(page_head) >= MAX_ORDER)
761 		compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
762 	else
763 		compound_idx = page - page_head;
764 
765 	return (index << compound_order(page_head)) + compound_idx;
766 }
767 
768 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
769 {
770 	struct page *page;
771 
772 	if (h->order >= MAX_ORDER)
773 		return NULL;
774 
775 	page = alloc_pages_exact_node(nid,
776 		htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
777 						__GFP_REPEAT|__GFP_NOWARN,
778 		huge_page_order(h));
779 	if (page) {
780 		if (arch_prepare_hugepage(page)) {
781 			__free_pages(page, huge_page_order(h));
782 			return NULL;
783 		}
784 		prep_new_huge_page(h, page, nid);
785 	}
786 
787 	return page;
788 }
789 
790 /*
791  * common helper functions for hstate_next_node_to_{alloc|free}.
792  * We may have allocated or freed a huge page based on a different
793  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
794  * be outside of *nodes_allowed.  Ensure that we use an allowed
795  * node for alloc or free.
796  */
797 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
798 {
799 	nid = next_node(nid, *nodes_allowed);
800 	if (nid == MAX_NUMNODES)
801 		nid = first_node(*nodes_allowed);
802 	VM_BUG_ON(nid >= MAX_NUMNODES);
803 
804 	return nid;
805 }
806 
807 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
808 {
809 	if (!node_isset(nid, *nodes_allowed))
810 		nid = next_node_allowed(nid, nodes_allowed);
811 	return nid;
812 }
813 
814 /*
815  * returns the previously saved node ["this node"] from which to
816  * allocate a persistent huge page for the pool and advance the
817  * next node from which to allocate, handling wrap at end of node
818  * mask.
819  */
820 static int hstate_next_node_to_alloc(struct hstate *h,
821 					nodemask_t *nodes_allowed)
822 {
823 	int nid;
824 
825 	VM_BUG_ON(!nodes_allowed);
826 
827 	nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
828 	h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
829 
830 	return nid;
831 }
832 
833 /*
834  * helper for free_pool_huge_page() - return the previously saved
835  * node ["this node"] from which to free a huge page.  Advance the
836  * next node id whether or not we find a free huge page to free so
837  * that the next attempt to free addresses the next node.
838  */
839 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
840 {
841 	int nid;
842 
843 	VM_BUG_ON(!nodes_allowed);
844 
845 	nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
846 	h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
847 
848 	return nid;
849 }
850 
851 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)		\
852 	for (nr_nodes = nodes_weight(*mask);				\
853 		nr_nodes > 0 &&						\
854 		((node = hstate_next_node_to_alloc(hs, mask)) || 1);	\
855 		nr_nodes--)
856 
857 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)		\
858 	for (nr_nodes = nodes_weight(*mask);				\
859 		nr_nodes > 0 &&						\
860 		((node = hstate_next_node_to_free(hs, mask)) || 1);	\
861 		nr_nodes--)
862 
863 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
864 {
865 	struct page *page;
866 	int nr_nodes, node;
867 	int ret = 0;
868 
869 	for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
870 		page = alloc_fresh_huge_page_node(h, node);
871 		if (page) {
872 			ret = 1;
873 			break;
874 		}
875 	}
876 
877 	if (ret)
878 		count_vm_event(HTLB_BUDDY_PGALLOC);
879 	else
880 		count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
881 
882 	return ret;
883 }
884 
885 /*
886  * Free huge page from pool from next node to free.
887  * Attempt to keep persistent huge pages more or less
888  * balanced over allowed nodes.
889  * Called with hugetlb_lock locked.
890  */
891 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
892 							 bool acct_surplus)
893 {
894 	int nr_nodes, node;
895 	int ret = 0;
896 
897 	for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
898 		/*
899 		 * If we're returning unused surplus pages, only examine
900 		 * nodes with surplus pages.
901 		 */
902 		if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
903 		    !list_empty(&h->hugepage_freelists[node])) {
904 			struct page *page =
905 				list_entry(h->hugepage_freelists[node].next,
906 					  struct page, lru);
907 			list_del(&page->lru);
908 			h->free_huge_pages--;
909 			h->free_huge_pages_node[node]--;
910 			if (acct_surplus) {
911 				h->surplus_huge_pages--;
912 				h->surplus_huge_pages_node[node]--;
913 			}
914 			update_and_free_page(h, page);
915 			ret = 1;
916 			break;
917 		}
918 	}
919 
920 	return ret;
921 }
922 
923 /*
924  * Dissolve a given free hugepage into free buddy pages. This function does
925  * nothing for in-use (including surplus) hugepages.
926  */
927 static void dissolve_free_huge_page(struct page *page)
928 {
929 	spin_lock(&hugetlb_lock);
930 	if (PageHuge(page) && !page_count(page)) {
931 		struct hstate *h = page_hstate(page);
932 		int nid = page_to_nid(page);
933 		list_del(&page->lru);
934 		h->free_huge_pages--;
935 		h->free_huge_pages_node[nid]--;
936 		update_and_free_page(h, page);
937 	}
938 	spin_unlock(&hugetlb_lock);
939 }
940 
941 /*
942  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
943  * make specified memory blocks removable from the system.
944  * Note that start_pfn should aligned with (minimum) hugepage size.
945  */
946 void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
947 {
948 	unsigned int order = 8 * sizeof(void *);
949 	unsigned long pfn;
950 	struct hstate *h;
951 
952 	/* Set scan step to minimum hugepage size */
953 	for_each_hstate(h)
954 		if (order > huge_page_order(h))
955 			order = huge_page_order(h);
956 	VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order));
957 	for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order)
958 		dissolve_free_huge_page(pfn_to_page(pfn));
959 }
960 
961 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
962 {
963 	struct page *page;
964 	unsigned int r_nid;
965 
966 	if (h->order >= MAX_ORDER)
967 		return NULL;
968 
969 	/*
970 	 * Assume we will successfully allocate the surplus page to
971 	 * prevent racing processes from causing the surplus to exceed
972 	 * overcommit
973 	 *
974 	 * This however introduces a different race, where a process B
975 	 * tries to grow the static hugepage pool while alloc_pages() is
976 	 * called by process A. B will only examine the per-node
977 	 * counters in determining if surplus huge pages can be
978 	 * converted to normal huge pages in adjust_pool_surplus(). A
979 	 * won't be able to increment the per-node counter, until the
980 	 * lock is dropped by B, but B doesn't drop hugetlb_lock until
981 	 * no more huge pages can be converted from surplus to normal
982 	 * state (and doesn't try to convert again). Thus, we have a
983 	 * case where a surplus huge page exists, the pool is grown, and
984 	 * the surplus huge page still exists after, even though it
985 	 * should just have been converted to a normal huge page. This
986 	 * does not leak memory, though, as the hugepage will be freed
987 	 * once it is out of use. It also does not allow the counters to
988 	 * go out of whack in adjust_pool_surplus() as we don't modify
989 	 * the node values until we've gotten the hugepage and only the
990 	 * per-node value is checked there.
991 	 */
992 	spin_lock(&hugetlb_lock);
993 	if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
994 		spin_unlock(&hugetlb_lock);
995 		return NULL;
996 	} else {
997 		h->nr_huge_pages++;
998 		h->surplus_huge_pages++;
999 	}
1000 	spin_unlock(&hugetlb_lock);
1001 
1002 	if (nid == NUMA_NO_NODE)
1003 		page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
1004 				   __GFP_REPEAT|__GFP_NOWARN,
1005 				   huge_page_order(h));
1006 	else
1007 		page = alloc_pages_exact_node(nid,
1008 			htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1009 			__GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
1010 
1011 	if (page && arch_prepare_hugepage(page)) {
1012 		__free_pages(page, huge_page_order(h));
1013 		page = NULL;
1014 	}
1015 
1016 	spin_lock(&hugetlb_lock);
1017 	if (page) {
1018 		INIT_LIST_HEAD(&page->lru);
1019 		r_nid = page_to_nid(page);
1020 		set_compound_page_dtor(page, free_huge_page);
1021 		set_hugetlb_cgroup(page, NULL);
1022 		/*
1023 		 * We incremented the global counters already
1024 		 */
1025 		h->nr_huge_pages_node[r_nid]++;
1026 		h->surplus_huge_pages_node[r_nid]++;
1027 		__count_vm_event(HTLB_BUDDY_PGALLOC);
1028 	} else {
1029 		h->nr_huge_pages--;
1030 		h->surplus_huge_pages--;
1031 		__count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1032 	}
1033 	spin_unlock(&hugetlb_lock);
1034 
1035 	return page;
1036 }
1037 
1038 /*
1039  * This allocation function is useful in the context where vma is irrelevant.
1040  * E.g. soft-offlining uses this function because it only cares physical
1041  * address of error page.
1042  */
1043 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1044 {
1045 	struct page *page = NULL;
1046 
1047 	spin_lock(&hugetlb_lock);
1048 	if (h->free_huge_pages - h->resv_huge_pages > 0)
1049 		page = dequeue_huge_page_node(h, nid);
1050 	spin_unlock(&hugetlb_lock);
1051 
1052 	if (!page)
1053 		page = alloc_buddy_huge_page(h, nid);
1054 
1055 	return page;
1056 }
1057 
1058 /*
1059  * Increase the hugetlb pool such that it can accommodate a reservation
1060  * of size 'delta'.
1061  */
1062 static int gather_surplus_pages(struct hstate *h, int delta)
1063 {
1064 	struct list_head surplus_list;
1065 	struct page *page, *tmp;
1066 	int ret, i;
1067 	int needed, allocated;
1068 	bool alloc_ok = true;
1069 
1070 	needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1071 	if (needed <= 0) {
1072 		h->resv_huge_pages += delta;
1073 		return 0;
1074 	}
1075 
1076 	allocated = 0;
1077 	INIT_LIST_HEAD(&surplus_list);
1078 
1079 	ret = -ENOMEM;
1080 retry:
1081 	spin_unlock(&hugetlb_lock);
1082 	for (i = 0; i < needed; i++) {
1083 		page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1084 		if (!page) {
1085 			alloc_ok = false;
1086 			break;
1087 		}
1088 		list_add(&page->lru, &surplus_list);
1089 	}
1090 	allocated += i;
1091 
1092 	/*
1093 	 * After retaking hugetlb_lock, we need to recalculate 'needed'
1094 	 * because either resv_huge_pages or free_huge_pages may have changed.
1095 	 */
1096 	spin_lock(&hugetlb_lock);
1097 	needed = (h->resv_huge_pages + delta) -
1098 			(h->free_huge_pages + allocated);
1099 	if (needed > 0) {
1100 		if (alloc_ok)
1101 			goto retry;
1102 		/*
1103 		 * We were not able to allocate enough pages to
1104 		 * satisfy the entire reservation so we free what
1105 		 * we've allocated so far.
1106 		 */
1107 		goto free;
1108 	}
1109 	/*
1110 	 * The surplus_list now contains _at_least_ the number of extra pages
1111 	 * needed to accommodate the reservation.  Add the appropriate number
1112 	 * of pages to the hugetlb pool and free the extras back to the buddy
1113 	 * allocator.  Commit the entire reservation here to prevent another
1114 	 * process from stealing the pages as they are added to the pool but
1115 	 * before they are reserved.
1116 	 */
1117 	needed += allocated;
1118 	h->resv_huge_pages += delta;
1119 	ret = 0;
1120 
1121 	/* Free the needed pages to the hugetlb pool */
1122 	list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1123 		if ((--needed) < 0)
1124 			break;
1125 		/*
1126 		 * This page is now managed by the hugetlb allocator and has
1127 		 * no users -- drop the buddy allocator's reference.
1128 		 */
1129 		put_page_testzero(page);
1130 		VM_BUG_ON_PAGE(page_count(page), page);
1131 		enqueue_huge_page(h, page);
1132 	}
1133 free:
1134 	spin_unlock(&hugetlb_lock);
1135 
1136 	/* Free unnecessary surplus pages to the buddy allocator */
1137 	list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1138 		put_page(page);
1139 	spin_lock(&hugetlb_lock);
1140 
1141 	return ret;
1142 }
1143 
1144 /*
1145  * When releasing a hugetlb pool reservation, any surplus pages that were
1146  * allocated to satisfy the reservation must be explicitly freed if they were
1147  * never used.
1148  * Called with hugetlb_lock held.
1149  */
1150 static void return_unused_surplus_pages(struct hstate *h,
1151 					unsigned long unused_resv_pages)
1152 {
1153 	unsigned long nr_pages;
1154 
1155 	/* Uncommit the reservation */
1156 	h->resv_huge_pages -= unused_resv_pages;
1157 
1158 	/* Cannot return gigantic pages currently */
1159 	if (h->order >= MAX_ORDER)
1160 		return;
1161 
1162 	nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1163 
1164 	/*
1165 	 * We want to release as many surplus pages as possible, spread
1166 	 * evenly across all nodes with memory. Iterate across these nodes
1167 	 * until we can no longer free unreserved surplus pages. This occurs
1168 	 * when the nodes with surplus pages have no free pages.
1169 	 * free_pool_huge_page() will balance the the freed pages across the
1170 	 * on-line nodes with memory and will handle the hstate accounting.
1171 	 */
1172 	while (nr_pages--) {
1173 		if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1174 			break;
1175 	}
1176 }
1177 
1178 /*
1179  * Determine if the huge page at addr within the vma has an associated
1180  * reservation.  Where it does not we will need to logically increase
1181  * reservation and actually increase subpool usage before an allocation
1182  * can occur.  Where any new reservation would be required the
1183  * reservation change is prepared, but not committed.  Once the page
1184  * has been allocated from the subpool and instantiated the change should
1185  * be committed via vma_commit_reservation.  No action is required on
1186  * failure.
1187  */
1188 static long vma_needs_reservation(struct hstate *h,
1189 			struct vm_area_struct *vma, unsigned long addr)
1190 {
1191 	struct resv_map *resv;
1192 	pgoff_t idx;
1193 	long chg;
1194 
1195 	resv = vma_resv_map(vma);
1196 	if (!resv)
1197 		return 1;
1198 
1199 	idx = vma_hugecache_offset(h, vma, addr);
1200 	chg = region_chg(resv, idx, idx + 1);
1201 
1202 	if (vma->vm_flags & VM_MAYSHARE)
1203 		return chg;
1204 	else
1205 		return chg < 0 ? chg : 0;
1206 }
1207 static void vma_commit_reservation(struct hstate *h,
1208 			struct vm_area_struct *vma, unsigned long addr)
1209 {
1210 	struct resv_map *resv;
1211 	pgoff_t idx;
1212 
1213 	resv = vma_resv_map(vma);
1214 	if (!resv)
1215 		return;
1216 
1217 	idx = vma_hugecache_offset(h, vma, addr);
1218 	region_add(resv, idx, idx + 1);
1219 }
1220 
1221 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1222 				    unsigned long addr, int avoid_reserve)
1223 {
1224 	struct hugepage_subpool *spool = subpool_vma(vma);
1225 	struct hstate *h = hstate_vma(vma);
1226 	struct page *page;
1227 	long chg;
1228 	int ret, idx;
1229 	struct hugetlb_cgroup *h_cg;
1230 
1231 	idx = hstate_index(h);
1232 	/*
1233 	 * Processes that did not create the mapping will have no
1234 	 * reserves and will not have accounted against subpool
1235 	 * limit. Check that the subpool limit can be made before
1236 	 * satisfying the allocation MAP_NORESERVE mappings may also
1237 	 * need pages and subpool limit allocated allocated if no reserve
1238 	 * mapping overlaps.
1239 	 */
1240 	chg = vma_needs_reservation(h, vma, addr);
1241 	if (chg < 0)
1242 		return ERR_PTR(-ENOMEM);
1243 	if (chg || avoid_reserve)
1244 		if (hugepage_subpool_get_pages(spool, 1))
1245 			return ERR_PTR(-ENOSPC);
1246 
1247 	ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1248 	if (ret) {
1249 		if (chg || avoid_reserve)
1250 			hugepage_subpool_put_pages(spool, 1);
1251 		return ERR_PTR(-ENOSPC);
1252 	}
1253 	spin_lock(&hugetlb_lock);
1254 	page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
1255 	if (!page) {
1256 		spin_unlock(&hugetlb_lock);
1257 		page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1258 		if (!page) {
1259 			hugetlb_cgroup_uncharge_cgroup(idx,
1260 						       pages_per_huge_page(h),
1261 						       h_cg);
1262 			if (chg || avoid_reserve)
1263 				hugepage_subpool_put_pages(spool, 1);
1264 			return ERR_PTR(-ENOSPC);
1265 		}
1266 		spin_lock(&hugetlb_lock);
1267 		list_move(&page->lru, &h->hugepage_activelist);
1268 		/* Fall through */
1269 	}
1270 	hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1271 	spin_unlock(&hugetlb_lock);
1272 
1273 	set_page_private(page, (unsigned long)spool);
1274 
1275 	vma_commit_reservation(h, vma, addr);
1276 	return page;
1277 }
1278 
1279 /*
1280  * alloc_huge_page()'s wrapper which simply returns the page if allocation
1281  * succeeds, otherwise NULL. This function is called from new_vma_page(),
1282  * where no ERR_VALUE is expected to be returned.
1283  */
1284 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1285 				unsigned long addr, int avoid_reserve)
1286 {
1287 	struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1288 	if (IS_ERR(page))
1289 		page = NULL;
1290 	return page;
1291 }
1292 
1293 int __weak alloc_bootmem_huge_page(struct hstate *h)
1294 {
1295 	struct huge_bootmem_page *m;
1296 	int nr_nodes, node;
1297 
1298 	for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1299 		void *addr;
1300 
1301 		addr = memblock_virt_alloc_try_nid_nopanic(
1302 				huge_page_size(h), huge_page_size(h),
1303 				0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1304 		if (addr) {
1305 			/*
1306 			 * Use the beginning of the huge page to store the
1307 			 * huge_bootmem_page struct (until gather_bootmem
1308 			 * puts them into the mem_map).
1309 			 */
1310 			m = addr;
1311 			goto found;
1312 		}
1313 	}
1314 	return 0;
1315 
1316 found:
1317 	BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1318 	/* Put them into a private list first because mem_map is not up yet */
1319 	list_add(&m->list, &huge_boot_pages);
1320 	m->hstate = h;
1321 	return 1;
1322 }
1323 
1324 static void __init prep_compound_huge_page(struct page *page, int order)
1325 {
1326 	if (unlikely(order > (MAX_ORDER - 1)))
1327 		prep_compound_gigantic_page(page, order);
1328 	else
1329 		prep_compound_page(page, order);
1330 }
1331 
1332 /* Put bootmem huge pages into the standard lists after mem_map is up */
1333 static void __init gather_bootmem_prealloc(void)
1334 {
1335 	struct huge_bootmem_page *m;
1336 
1337 	list_for_each_entry(m, &huge_boot_pages, list) {
1338 		struct hstate *h = m->hstate;
1339 		struct page *page;
1340 
1341 #ifdef CONFIG_HIGHMEM
1342 		page = pfn_to_page(m->phys >> PAGE_SHIFT);
1343 		memblock_free_late(__pa(m),
1344 				   sizeof(struct huge_bootmem_page));
1345 #else
1346 		page = virt_to_page(m);
1347 #endif
1348 		WARN_ON(page_count(page) != 1);
1349 		prep_compound_huge_page(page, h->order);
1350 		WARN_ON(PageReserved(page));
1351 		prep_new_huge_page(h, page, page_to_nid(page));
1352 		/*
1353 		 * If we had gigantic hugepages allocated at boot time, we need
1354 		 * to restore the 'stolen' pages to totalram_pages in order to
1355 		 * fix confusing memory reports from free(1) and another
1356 		 * side-effects, like CommitLimit going negative.
1357 		 */
1358 		if (h->order > (MAX_ORDER - 1))
1359 			adjust_managed_page_count(page, 1 << h->order);
1360 	}
1361 }
1362 
1363 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1364 {
1365 	unsigned long i;
1366 
1367 	for (i = 0; i < h->max_huge_pages; ++i) {
1368 		if (h->order >= MAX_ORDER) {
1369 			if (!alloc_bootmem_huge_page(h))
1370 				break;
1371 		} else if (!alloc_fresh_huge_page(h,
1372 					 &node_states[N_MEMORY]))
1373 			break;
1374 	}
1375 	h->max_huge_pages = i;
1376 }
1377 
1378 static void __init hugetlb_init_hstates(void)
1379 {
1380 	struct hstate *h;
1381 
1382 	for_each_hstate(h) {
1383 		/* oversize hugepages were init'ed in early boot */
1384 		if (h->order < MAX_ORDER)
1385 			hugetlb_hstate_alloc_pages(h);
1386 	}
1387 }
1388 
1389 static char * __init memfmt(char *buf, unsigned long n)
1390 {
1391 	if (n >= (1UL << 30))
1392 		sprintf(buf, "%lu GB", n >> 30);
1393 	else if (n >= (1UL << 20))
1394 		sprintf(buf, "%lu MB", n >> 20);
1395 	else
1396 		sprintf(buf, "%lu KB", n >> 10);
1397 	return buf;
1398 }
1399 
1400 static void __init report_hugepages(void)
1401 {
1402 	struct hstate *h;
1403 
1404 	for_each_hstate(h) {
1405 		char buf[32];
1406 		pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1407 			memfmt(buf, huge_page_size(h)),
1408 			h->free_huge_pages);
1409 	}
1410 }
1411 
1412 #ifdef CONFIG_HIGHMEM
1413 static void try_to_free_low(struct hstate *h, unsigned long count,
1414 						nodemask_t *nodes_allowed)
1415 {
1416 	int i;
1417 
1418 	if (h->order >= MAX_ORDER)
1419 		return;
1420 
1421 	for_each_node_mask(i, *nodes_allowed) {
1422 		struct page *page, *next;
1423 		struct list_head *freel = &h->hugepage_freelists[i];
1424 		list_for_each_entry_safe(page, next, freel, lru) {
1425 			if (count >= h->nr_huge_pages)
1426 				return;
1427 			if (PageHighMem(page))
1428 				continue;
1429 			list_del(&page->lru);
1430 			update_and_free_page(h, page);
1431 			h->free_huge_pages--;
1432 			h->free_huge_pages_node[page_to_nid(page)]--;
1433 		}
1434 	}
1435 }
1436 #else
1437 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1438 						nodemask_t *nodes_allowed)
1439 {
1440 }
1441 #endif
1442 
1443 /*
1444  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
1445  * balanced by operating on them in a round-robin fashion.
1446  * Returns 1 if an adjustment was made.
1447  */
1448 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1449 				int delta)
1450 {
1451 	int nr_nodes, node;
1452 
1453 	VM_BUG_ON(delta != -1 && delta != 1);
1454 
1455 	if (delta < 0) {
1456 		for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1457 			if (h->surplus_huge_pages_node[node])
1458 				goto found;
1459 		}
1460 	} else {
1461 		for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1462 			if (h->surplus_huge_pages_node[node] <
1463 					h->nr_huge_pages_node[node])
1464 				goto found;
1465 		}
1466 	}
1467 	return 0;
1468 
1469 found:
1470 	h->surplus_huge_pages += delta;
1471 	h->surplus_huge_pages_node[node] += delta;
1472 	return 1;
1473 }
1474 
1475 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1476 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1477 						nodemask_t *nodes_allowed)
1478 {
1479 	unsigned long min_count, ret;
1480 
1481 	if (h->order >= MAX_ORDER)
1482 		return h->max_huge_pages;
1483 
1484 	/*
1485 	 * Increase the pool size
1486 	 * First take pages out of surplus state.  Then make up the
1487 	 * remaining difference by allocating fresh huge pages.
1488 	 *
1489 	 * We might race with alloc_buddy_huge_page() here and be unable
1490 	 * to convert a surplus huge page to a normal huge page. That is
1491 	 * not critical, though, it just means the overall size of the
1492 	 * pool might be one hugepage larger than it needs to be, but
1493 	 * within all the constraints specified by the sysctls.
1494 	 */
1495 	spin_lock(&hugetlb_lock);
1496 	while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1497 		if (!adjust_pool_surplus(h, nodes_allowed, -1))
1498 			break;
1499 	}
1500 
1501 	while (count > persistent_huge_pages(h)) {
1502 		/*
1503 		 * If this allocation races such that we no longer need the
1504 		 * page, free_huge_page will handle it by freeing the page
1505 		 * and reducing the surplus.
1506 		 */
1507 		spin_unlock(&hugetlb_lock);
1508 		ret = alloc_fresh_huge_page(h, nodes_allowed);
1509 		spin_lock(&hugetlb_lock);
1510 		if (!ret)
1511 			goto out;
1512 
1513 		/* Bail for signals. Probably ctrl-c from user */
1514 		if (signal_pending(current))
1515 			goto out;
1516 	}
1517 
1518 	/*
1519 	 * Decrease the pool size
1520 	 * First return free pages to the buddy allocator (being careful
1521 	 * to keep enough around to satisfy reservations).  Then place
1522 	 * pages into surplus state as needed so the pool will shrink
1523 	 * to the desired size as pages become free.
1524 	 *
1525 	 * By placing pages into the surplus state independent of the
1526 	 * overcommit value, we are allowing the surplus pool size to
1527 	 * exceed overcommit. There are few sane options here. Since
1528 	 * alloc_buddy_huge_page() is checking the global counter,
1529 	 * though, we'll note that we're not allowed to exceed surplus
1530 	 * and won't grow the pool anywhere else. Not until one of the
1531 	 * sysctls are changed, or the surplus pages go out of use.
1532 	 */
1533 	min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1534 	min_count = max(count, min_count);
1535 	try_to_free_low(h, min_count, nodes_allowed);
1536 	while (min_count < persistent_huge_pages(h)) {
1537 		if (!free_pool_huge_page(h, nodes_allowed, 0))
1538 			break;
1539 		cond_resched_lock(&hugetlb_lock);
1540 	}
1541 	while (count < persistent_huge_pages(h)) {
1542 		if (!adjust_pool_surplus(h, nodes_allowed, 1))
1543 			break;
1544 	}
1545 out:
1546 	ret = persistent_huge_pages(h);
1547 	spin_unlock(&hugetlb_lock);
1548 	return ret;
1549 }
1550 
1551 #define HSTATE_ATTR_RO(_name) \
1552 	static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1553 
1554 #define HSTATE_ATTR(_name) \
1555 	static struct kobj_attribute _name##_attr = \
1556 		__ATTR(_name, 0644, _name##_show, _name##_store)
1557 
1558 static struct kobject *hugepages_kobj;
1559 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1560 
1561 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1562 
1563 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1564 {
1565 	int i;
1566 
1567 	for (i = 0; i < HUGE_MAX_HSTATE; i++)
1568 		if (hstate_kobjs[i] == kobj) {
1569 			if (nidp)
1570 				*nidp = NUMA_NO_NODE;
1571 			return &hstates[i];
1572 		}
1573 
1574 	return kobj_to_node_hstate(kobj, nidp);
1575 }
1576 
1577 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1578 					struct kobj_attribute *attr, char *buf)
1579 {
1580 	struct hstate *h;
1581 	unsigned long nr_huge_pages;
1582 	int nid;
1583 
1584 	h = kobj_to_hstate(kobj, &nid);
1585 	if (nid == NUMA_NO_NODE)
1586 		nr_huge_pages = h->nr_huge_pages;
1587 	else
1588 		nr_huge_pages = h->nr_huge_pages_node[nid];
1589 
1590 	return sprintf(buf, "%lu\n", nr_huge_pages);
1591 }
1592 
1593 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1594 			struct kobject *kobj, struct kobj_attribute *attr,
1595 			const char *buf, size_t len)
1596 {
1597 	int err;
1598 	int nid;
1599 	unsigned long count;
1600 	struct hstate *h;
1601 	NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1602 
1603 	err = kstrtoul(buf, 10, &count);
1604 	if (err)
1605 		goto out;
1606 
1607 	h = kobj_to_hstate(kobj, &nid);
1608 	if (h->order >= MAX_ORDER) {
1609 		err = -EINVAL;
1610 		goto out;
1611 	}
1612 
1613 	if (nid == NUMA_NO_NODE) {
1614 		/*
1615 		 * global hstate attribute
1616 		 */
1617 		if (!(obey_mempolicy &&
1618 				init_nodemask_of_mempolicy(nodes_allowed))) {
1619 			NODEMASK_FREE(nodes_allowed);
1620 			nodes_allowed = &node_states[N_MEMORY];
1621 		}
1622 	} else if (nodes_allowed) {
1623 		/*
1624 		 * per node hstate attribute: adjust count to global,
1625 		 * but restrict alloc/free to the specified node.
1626 		 */
1627 		count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1628 		init_nodemask_of_node(nodes_allowed, nid);
1629 	} else
1630 		nodes_allowed = &node_states[N_MEMORY];
1631 
1632 	h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1633 
1634 	if (nodes_allowed != &node_states[N_MEMORY])
1635 		NODEMASK_FREE(nodes_allowed);
1636 
1637 	return len;
1638 out:
1639 	NODEMASK_FREE(nodes_allowed);
1640 	return err;
1641 }
1642 
1643 static ssize_t nr_hugepages_show(struct kobject *kobj,
1644 				       struct kobj_attribute *attr, char *buf)
1645 {
1646 	return nr_hugepages_show_common(kobj, attr, buf);
1647 }
1648 
1649 static ssize_t nr_hugepages_store(struct kobject *kobj,
1650 	       struct kobj_attribute *attr, const char *buf, size_t len)
1651 {
1652 	return nr_hugepages_store_common(false, kobj, attr, buf, len);
1653 }
1654 HSTATE_ATTR(nr_hugepages);
1655 
1656 #ifdef CONFIG_NUMA
1657 
1658 /*
1659  * hstate attribute for optionally mempolicy-based constraint on persistent
1660  * huge page alloc/free.
1661  */
1662 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1663 				       struct kobj_attribute *attr, char *buf)
1664 {
1665 	return nr_hugepages_show_common(kobj, attr, buf);
1666 }
1667 
1668 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1669 	       struct kobj_attribute *attr, const char *buf, size_t len)
1670 {
1671 	return nr_hugepages_store_common(true, kobj, attr, buf, len);
1672 }
1673 HSTATE_ATTR(nr_hugepages_mempolicy);
1674 #endif
1675 
1676 
1677 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1678 					struct kobj_attribute *attr, char *buf)
1679 {
1680 	struct hstate *h = kobj_to_hstate(kobj, NULL);
1681 	return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1682 }
1683 
1684 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1685 		struct kobj_attribute *attr, const char *buf, size_t count)
1686 {
1687 	int err;
1688 	unsigned long input;
1689 	struct hstate *h = kobj_to_hstate(kobj, NULL);
1690 
1691 	if (h->order >= MAX_ORDER)
1692 		return -EINVAL;
1693 
1694 	err = kstrtoul(buf, 10, &input);
1695 	if (err)
1696 		return err;
1697 
1698 	spin_lock(&hugetlb_lock);
1699 	h->nr_overcommit_huge_pages = input;
1700 	spin_unlock(&hugetlb_lock);
1701 
1702 	return count;
1703 }
1704 HSTATE_ATTR(nr_overcommit_hugepages);
1705 
1706 static ssize_t free_hugepages_show(struct kobject *kobj,
1707 					struct kobj_attribute *attr, char *buf)
1708 {
1709 	struct hstate *h;
1710 	unsigned long free_huge_pages;
1711 	int nid;
1712 
1713 	h = kobj_to_hstate(kobj, &nid);
1714 	if (nid == NUMA_NO_NODE)
1715 		free_huge_pages = h->free_huge_pages;
1716 	else
1717 		free_huge_pages = h->free_huge_pages_node[nid];
1718 
1719 	return sprintf(buf, "%lu\n", free_huge_pages);
1720 }
1721 HSTATE_ATTR_RO(free_hugepages);
1722 
1723 static ssize_t resv_hugepages_show(struct kobject *kobj,
1724 					struct kobj_attribute *attr, char *buf)
1725 {
1726 	struct hstate *h = kobj_to_hstate(kobj, NULL);
1727 	return sprintf(buf, "%lu\n", h->resv_huge_pages);
1728 }
1729 HSTATE_ATTR_RO(resv_hugepages);
1730 
1731 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1732 					struct kobj_attribute *attr, char *buf)
1733 {
1734 	struct hstate *h;
1735 	unsigned long surplus_huge_pages;
1736 	int nid;
1737 
1738 	h = kobj_to_hstate(kobj, &nid);
1739 	if (nid == NUMA_NO_NODE)
1740 		surplus_huge_pages = h->surplus_huge_pages;
1741 	else
1742 		surplus_huge_pages = h->surplus_huge_pages_node[nid];
1743 
1744 	return sprintf(buf, "%lu\n", surplus_huge_pages);
1745 }
1746 HSTATE_ATTR_RO(surplus_hugepages);
1747 
1748 static struct attribute *hstate_attrs[] = {
1749 	&nr_hugepages_attr.attr,
1750 	&nr_overcommit_hugepages_attr.attr,
1751 	&free_hugepages_attr.attr,
1752 	&resv_hugepages_attr.attr,
1753 	&surplus_hugepages_attr.attr,
1754 #ifdef CONFIG_NUMA
1755 	&nr_hugepages_mempolicy_attr.attr,
1756 #endif
1757 	NULL,
1758 };
1759 
1760 static struct attribute_group hstate_attr_group = {
1761 	.attrs = hstate_attrs,
1762 };
1763 
1764 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1765 				    struct kobject **hstate_kobjs,
1766 				    struct attribute_group *hstate_attr_group)
1767 {
1768 	int retval;
1769 	int hi = hstate_index(h);
1770 
1771 	hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1772 	if (!hstate_kobjs[hi])
1773 		return -ENOMEM;
1774 
1775 	retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1776 	if (retval)
1777 		kobject_put(hstate_kobjs[hi]);
1778 
1779 	return retval;
1780 }
1781 
1782 static void __init hugetlb_sysfs_init(void)
1783 {
1784 	struct hstate *h;
1785 	int err;
1786 
1787 	hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1788 	if (!hugepages_kobj)
1789 		return;
1790 
1791 	for_each_hstate(h) {
1792 		err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1793 					 hstate_kobjs, &hstate_attr_group);
1794 		if (err)
1795 			pr_err("Hugetlb: Unable to add hstate %s", h->name);
1796 	}
1797 }
1798 
1799 #ifdef CONFIG_NUMA
1800 
1801 /*
1802  * node_hstate/s - associate per node hstate attributes, via their kobjects,
1803  * with node devices in node_devices[] using a parallel array.  The array
1804  * index of a node device or _hstate == node id.
1805  * This is here to avoid any static dependency of the node device driver, in
1806  * the base kernel, on the hugetlb module.
1807  */
1808 struct node_hstate {
1809 	struct kobject		*hugepages_kobj;
1810 	struct kobject		*hstate_kobjs[HUGE_MAX_HSTATE];
1811 };
1812 struct node_hstate node_hstates[MAX_NUMNODES];
1813 
1814 /*
1815  * A subset of global hstate attributes for node devices
1816  */
1817 static struct attribute *per_node_hstate_attrs[] = {
1818 	&nr_hugepages_attr.attr,
1819 	&free_hugepages_attr.attr,
1820 	&surplus_hugepages_attr.attr,
1821 	NULL,
1822 };
1823 
1824 static struct attribute_group per_node_hstate_attr_group = {
1825 	.attrs = per_node_hstate_attrs,
1826 };
1827 
1828 /*
1829  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1830  * Returns node id via non-NULL nidp.
1831  */
1832 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1833 {
1834 	int nid;
1835 
1836 	for (nid = 0; nid < nr_node_ids; nid++) {
1837 		struct node_hstate *nhs = &node_hstates[nid];
1838 		int i;
1839 		for (i = 0; i < HUGE_MAX_HSTATE; i++)
1840 			if (nhs->hstate_kobjs[i] == kobj) {
1841 				if (nidp)
1842 					*nidp = nid;
1843 				return &hstates[i];
1844 			}
1845 	}
1846 
1847 	BUG();
1848 	return NULL;
1849 }
1850 
1851 /*
1852  * Unregister hstate attributes from a single node device.
1853  * No-op if no hstate attributes attached.
1854  */
1855 static void hugetlb_unregister_node(struct node *node)
1856 {
1857 	struct hstate *h;
1858 	struct node_hstate *nhs = &node_hstates[node->dev.id];
1859 
1860 	if (!nhs->hugepages_kobj)
1861 		return;		/* no hstate attributes */
1862 
1863 	for_each_hstate(h) {
1864 		int idx = hstate_index(h);
1865 		if (nhs->hstate_kobjs[idx]) {
1866 			kobject_put(nhs->hstate_kobjs[idx]);
1867 			nhs->hstate_kobjs[idx] = NULL;
1868 		}
1869 	}
1870 
1871 	kobject_put(nhs->hugepages_kobj);
1872 	nhs->hugepages_kobj = NULL;
1873 }
1874 
1875 /*
1876  * hugetlb module exit:  unregister hstate attributes from node devices
1877  * that have them.
1878  */
1879 static void hugetlb_unregister_all_nodes(void)
1880 {
1881 	int nid;
1882 
1883 	/*
1884 	 * disable node device registrations.
1885 	 */
1886 	register_hugetlbfs_with_node(NULL, NULL);
1887 
1888 	/*
1889 	 * remove hstate attributes from any nodes that have them.
1890 	 */
1891 	for (nid = 0; nid < nr_node_ids; nid++)
1892 		hugetlb_unregister_node(node_devices[nid]);
1893 }
1894 
1895 /*
1896  * Register hstate attributes for a single node device.
1897  * No-op if attributes already registered.
1898  */
1899 static void hugetlb_register_node(struct node *node)
1900 {
1901 	struct hstate *h;
1902 	struct node_hstate *nhs = &node_hstates[node->dev.id];
1903 	int err;
1904 
1905 	if (nhs->hugepages_kobj)
1906 		return;		/* already allocated */
1907 
1908 	nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1909 							&node->dev.kobj);
1910 	if (!nhs->hugepages_kobj)
1911 		return;
1912 
1913 	for_each_hstate(h) {
1914 		err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1915 						nhs->hstate_kobjs,
1916 						&per_node_hstate_attr_group);
1917 		if (err) {
1918 			pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1919 				h->name, node->dev.id);
1920 			hugetlb_unregister_node(node);
1921 			break;
1922 		}
1923 	}
1924 }
1925 
1926 /*
1927  * hugetlb init time:  register hstate attributes for all registered node
1928  * devices of nodes that have memory.  All on-line nodes should have
1929  * registered their associated device by this time.
1930  */
1931 static void hugetlb_register_all_nodes(void)
1932 {
1933 	int nid;
1934 
1935 	for_each_node_state(nid, N_MEMORY) {
1936 		struct node *node = node_devices[nid];
1937 		if (node->dev.id == nid)
1938 			hugetlb_register_node(node);
1939 	}
1940 
1941 	/*
1942 	 * Let the node device driver know we're here so it can
1943 	 * [un]register hstate attributes on node hotplug.
1944 	 */
1945 	register_hugetlbfs_with_node(hugetlb_register_node,
1946 				     hugetlb_unregister_node);
1947 }
1948 #else	/* !CONFIG_NUMA */
1949 
1950 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1951 {
1952 	BUG();
1953 	if (nidp)
1954 		*nidp = -1;
1955 	return NULL;
1956 }
1957 
1958 static void hugetlb_unregister_all_nodes(void) { }
1959 
1960 static void hugetlb_register_all_nodes(void) { }
1961 
1962 #endif
1963 
1964 static void __exit hugetlb_exit(void)
1965 {
1966 	struct hstate *h;
1967 
1968 	hugetlb_unregister_all_nodes();
1969 
1970 	for_each_hstate(h) {
1971 		kobject_put(hstate_kobjs[hstate_index(h)]);
1972 	}
1973 
1974 	kobject_put(hugepages_kobj);
1975 	kfree(htlb_fault_mutex_table);
1976 }
1977 module_exit(hugetlb_exit);
1978 
1979 static int __init hugetlb_init(void)
1980 {
1981 	int i;
1982 
1983 	/* Some platform decide whether they support huge pages at boot
1984 	 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1985 	 * there is no such support
1986 	 */
1987 	if (HPAGE_SHIFT == 0)
1988 		return 0;
1989 
1990 	if (!size_to_hstate(default_hstate_size)) {
1991 		default_hstate_size = HPAGE_SIZE;
1992 		if (!size_to_hstate(default_hstate_size))
1993 			hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1994 	}
1995 	default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1996 	if (default_hstate_max_huge_pages)
1997 		default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1998 
1999 	hugetlb_init_hstates();
2000 	gather_bootmem_prealloc();
2001 	report_hugepages();
2002 
2003 	hugetlb_sysfs_init();
2004 	hugetlb_register_all_nodes();
2005 	hugetlb_cgroup_file_init();
2006 
2007 #ifdef CONFIG_SMP
2008 	num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2009 #else
2010 	num_fault_mutexes = 1;
2011 #endif
2012 	htlb_fault_mutex_table =
2013 		kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2014 	BUG_ON(!htlb_fault_mutex_table);
2015 
2016 	for (i = 0; i < num_fault_mutexes; i++)
2017 		mutex_init(&htlb_fault_mutex_table[i]);
2018 	return 0;
2019 }
2020 module_init(hugetlb_init);
2021 
2022 /* Should be called on processing a hugepagesz=... option */
2023 void __init hugetlb_add_hstate(unsigned order)
2024 {
2025 	struct hstate *h;
2026 	unsigned long i;
2027 
2028 	if (size_to_hstate(PAGE_SIZE << order)) {
2029 		pr_warning("hugepagesz= specified twice, ignoring\n");
2030 		return;
2031 	}
2032 	BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2033 	BUG_ON(order == 0);
2034 	h = &hstates[hugetlb_max_hstate++];
2035 	h->order = order;
2036 	h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2037 	h->nr_huge_pages = 0;
2038 	h->free_huge_pages = 0;
2039 	for (i = 0; i < MAX_NUMNODES; ++i)
2040 		INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2041 	INIT_LIST_HEAD(&h->hugepage_activelist);
2042 	h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2043 	h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2044 	snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2045 					huge_page_size(h)/1024);
2046 
2047 	parsed_hstate = h;
2048 }
2049 
2050 static int __init hugetlb_nrpages_setup(char *s)
2051 {
2052 	unsigned long *mhp;
2053 	static unsigned long *last_mhp;
2054 
2055 	/*
2056 	 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2057 	 * so this hugepages= parameter goes to the "default hstate".
2058 	 */
2059 	if (!hugetlb_max_hstate)
2060 		mhp = &default_hstate_max_huge_pages;
2061 	else
2062 		mhp = &parsed_hstate->max_huge_pages;
2063 
2064 	if (mhp == last_mhp) {
2065 		pr_warning("hugepages= specified twice without "
2066 			   "interleaving hugepagesz=, ignoring\n");
2067 		return 1;
2068 	}
2069 
2070 	if (sscanf(s, "%lu", mhp) <= 0)
2071 		*mhp = 0;
2072 
2073 	/*
2074 	 * Global state is always initialized later in hugetlb_init.
2075 	 * But we need to allocate >= MAX_ORDER hstates here early to still
2076 	 * use the bootmem allocator.
2077 	 */
2078 	if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2079 		hugetlb_hstate_alloc_pages(parsed_hstate);
2080 
2081 	last_mhp = mhp;
2082 
2083 	return 1;
2084 }
2085 __setup("hugepages=", hugetlb_nrpages_setup);
2086 
2087 static int __init hugetlb_default_setup(char *s)
2088 {
2089 	default_hstate_size = memparse(s, &s);
2090 	return 1;
2091 }
2092 __setup("default_hugepagesz=", hugetlb_default_setup);
2093 
2094 static unsigned int cpuset_mems_nr(unsigned int *array)
2095 {
2096 	int node;
2097 	unsigned int nr = 0;
2098 
2099 	for_each_node_mask(node, cpuset_current_mems_allowed)
2100 		nr += array[node];
2101 
2102 	return nr;
2103 }
2104 
2105 #ifdef CONFIG_SYSCTL
2106 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2107 			 struct ctl_table *table, int write,
2108 			 void __user *buffer, size_t *length, loff_t *ppos)
2109 {
2110 	struct hstate *h = &default_hstate;
2111 	unsigned long tmp;
2112 	int ret;
2113 
2114 	tmp = h->max_huge_pages;
2115 
2116 	if (write && h->order >= MAX_ORDER)
2117 		return -EINVAL;
2118 
2119 	table->data = &tmp;
2120 	table->maxlen = sizeof(unsigned long);
2121 	ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2122 	if (ret)
2123 		goto out;
2124 
2125 	if (write) {
2126 		NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2127 						GFP_KERNEL | __GFP_NORETRY);
2128 		if (!(obey_mempolicy &&
2129 			       init_nodemask_of_mempolicy(nodes_allowed))) {
2130 			NODEMASK_FREE(nodes_allowed);
2131 			nodes_allowed = &node_states[N_MEMORY];
2132 		}
2133 		h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2134 
2135 		if (nodes_allowed != &node_states[N_MEMORY])
2136 			NODEMASK_FREE(nodes_allowed);
2137 	}
2138 out:
2139 	return ret;
2140 }
2141 
2142 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2143 			  void __user *buffer, size_t *length, loff_t *ppos)
2144 {
2145 
2146 	return hugetlb_sysctl_handler_common(false, table, write,
2147 							buffer, length, ppos);
2148 }
2149 
2150 #ifdef CONFIG_NUMA
2151 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2152 			  void __user *buffer, size_t *length, loff_t *ppos)
2153 {
2154 	return hugetlb_sysctl_handler_common(true, table, write,
2155 							buffer, length, ppos);
2156 }
2157 #endif /* CONFIG_NUMA */
2158 
2159 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2160 			void __user *buffer,
2161 			size_t *length, loff_t *ppos)
2162 {
2163 	struct hstate *h = &default_hstate;
2164 	unsigned long tmp;
2165 	int ret;
2166 
2167 	tmp = h->nr_overcommit_huge_pages;
2168 
2169 	if (write && h->order >= MAX_ORDER)
2170 		return -EINVAL;
2171 
2172 	table->data = &tmp;
2173 	table->maxlen = sizeof(unsigned long);
2174 	ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2175 	if (ret)
2176 		goto out;
2177 
2178 	if (write) {
2179 		spin_lock(&hugetlb_lock);
2180 		h->nr_overcommit_huge_pages = tmp;
2181 		spin_unlock(&hugetlb_lock);
2182 	}
2183 out:
2184 	return ret;
2185 }
2186 
2187 #endif /* CONFIG_SYSCTL */
2188 
2189 void hugetlb_report_meminfo(struct seq_file *m)
2190 {
2191 	struct hstate *h = &default_hstate;
2192 	seq_printf(m,
2193 			"HugePages_Total:   %5lu\n"
2194 			"HugePages_Free:    %5lu\n"
2195 			"HugePages_Rsvd:    %5lu\n"
2196 			"HugePages_Surp:    %5lu\n"
2197 			"Hugepagesize:   %8lu kB\n",
2198 			h->nr_huge_pages,
2199 			h->free_huge_pages,
2200 			h->resv_huge_pages,
2201 			h->surplus_huge_pages,
2202 			1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2203 }
2204 
2205 int hugetlb_report_node_meminfo(int nid, char *buf)
2206 {
2207 	struct hstate *h = &default_hstate;
2208 	return sprintf(buf,
2209 		"Node %d HugePages_Total: %5u\n"
2210 		"Node %d HugePages_Free:  %5u\n"
2211 		"Node %d HugePages_Surp:  %5u\n",
2212 		nid, h->nr_huge_pages_node[nid],
2213 		nid, h->free_huge_pages_node[nid],
2214 		nid, h->surplus_huge_pages_node[nid]);
2215 }
2216 
2217 void hugetlb_show_meminfo(void)
2218 {
2219 	struct hstate *h;
2220 	int nid;
2221 
2222 	for_each_node_state(nid, N_MEMORY)
2223 		for_each_hstate(h)
2224 			pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2225 				nid,
2226 				h->nr_huge_pages_node[nid],
2227 				h->free_huge_pages_node[nid],
2228 				h->surplus_huge_pages_node[nid],
2229 				1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2230 }
2231 
2232 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2233 unsigned long hugetlb_total_pages(void)
2234 {
2235 	struct hstate *h;
2236 	unsigned long nr_total_pages = 0;
2237 
2238 	for_each_hstate(h)
2239 		nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2240 	return nr_total_pages;
2241 }
2242 
2243 static int hugetlb_acct_memory(struct hstate *h, long delta)
2244 {
2245 	int ret = -ENOMEM;
2246 
2247 	spin_lock(&hugetlb_lock);
2248 	/*
2249 	 * When cpuset is configured, it breaks the strict hugetlb page
2250 	 * reservation as the accounting is done on a global variable. Such
2251 	 * reservation is completely rubbish in the presence of cpuset because
2252 	 * the reservation is not checked against page availability for the
2253 	 * current cpuset. Application can still potentially OOM'ed by kernel
2254 	 * with lack of free htlb page in cpuset that the task is in.
2255 	 * Attempt to enforce strict accounting with cpuset is almost
2256 	 * impossible (or too ugly) because cpuset is too fluid that
2257 	 * task or memory node can be dynamically moved between cpusets.
2258 	 *
2259 	 * The change of semantics for shared hugetlb mapping with cpuset is
2260 	 * undesirable. However, in order to preserve some of the semantics,
2261 	 * we fall back to check against current free page availability as
2262 	 * a best attempt and hopefully to minimize the impact of changing
2263 	 * semantics that cpuset has.
2264 	 */
2265 	if (delta > 0) {
2266 		if (gather_surplus_pages(h, delta) < 0)
2267 			goto out;
2268 
2269 		if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2270 			return_unused_surplus_pages(h, delta);
2271 			goto out;
2272 		}
2273 	}
2274 
2275 	ret = 0;
2276 	if (delta < 0)
2277 		return_unused_surplus_pages(h, (unsigned long) -delta);
2278 
2279 out:
2280 	spin_unlock(&hugetlb_lock);
2281 	return ret;
2282 }
2283 
2284 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2285 {
2286 	struct resv_map *resv = vma_resv_map(vma);
2287 
2288 	/*
2289 	 * This new VMA should share its siblings reservation map if present.
2290 	 * The VMA will only ever have a valid reservation map pointer where
2291 	 * it is being copied for another still existing VMA.  As that VMA
2292 	 * has a reference to the reservation map it cannot disappear until
2293 	 * after this open call completes.  It is therefore safe to take a
2294 	 * new reference here without additional locking.
2295 	 */
2296 	if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2297 		kref_get(&resv->refs);
2298 }
2299 
2300 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2301 {
2302 	struct hstate *h = hstate_vma(vma);
2303 	struct resv_map *resv = vma_resv_map(vma);
2304 	struct hugepage_subpool *spool = subpool_vma(vma);
2305 	unsigned long reserve, start, end;
2306 
2307 	if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2308 		return;
2309 
2310 	start = vma_hugecache_offset(h, vma, vma->vm_start);
2311 	end = vma_hugecache_offset(h, vma, vma->vm_end);
2312 
2313 	reserve = (end - start) - region_count(resv, start, end);
2314 
2315 	kref_put(&resv->refs, resv_map_release);
2316 
2317 	if (reserve) {
2318 		hugetlb_acct_memory(h, -reserve);
2319 		hugepage_subpool_put_pages(spool, reserve);
2320 	}
2321 }
2322 
2323 /*
2324  * We cannot handle pagefaults against hugetlb pages at all.  They cause
2325  * handle_mm_fault() to try to instantiate regular-sized pages in the
2326  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
2327  * this far.
2328  */
2329 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2330 {
2331 	BUG();
2332 	return 0;
2333 }
2334 
2335 const struct vm_operations_struct hugetlb_vm_ops = {
2336 	.fault = hugetlb_vm_op_fault,
2337 	.open = hugetlb_vm_op_open,
2338 	.close = hugetlb_vm_op_close,
2339 };
2340 
2341 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2342 				int writable)
2343 {
2344 	pte_t entry;
2345 
2346 	if (writable) {
2347 		entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2348 					 vma->vm_page_prot)));
2349 	} else {
2350 		entry = huge_pte_wrprotect(mk_huge_pte(page,
2351 					   vma->vm_page_prot));
2352 	}
2353 	entry = pte_mkyoung(entry);
2354 	entry = pte_mkhuge(entry);
2355 	entry = arch_make_huge_pte(entry, vma, page, writable);
2356 
2357 	return entry;
2358 }
2359 
2360 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2361 				   unsigned long address, pte_t *ptep)
2362 {
2363 	pte_t entry;
2364 
2365 	entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2366 	if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2367 		update_mmu_cache(vma, address, ptep);
2368 }
2369 
2370 
2371 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2372 			    struct vm_area_struct *vma)
2373 {
2374 	pte_t *src_pte, *dst_pte, entry;
2375 	struct page *ptepage;
2376 	unsigned long addr;
2377 	int cow;
2378 	struct hstate *h = hstate_vma(vma);
2379 	unsigned long sz = huge_page_size(h);
2380 	unsigned long mmun_start;	/* For mmu_notifiers */
2381 	unsigned long mmun_end;		/* For mmu_notifiers */
2382 	int ret = 0;
2383 
2384 	cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2385 
2386 	mmun_start = vma->vm_start;
2387 	mmun_end = vma->vm_end;
2388 	if (cow)
2389 		mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2390 
2391 	for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2392 		spinlock_t *src_ptl, *dst_ptl;
2393 		src_pte = huge_pte_offset(src, addr);
2394 		if (!src_pte)
2395 			continue;
2396 		dst_pte = huge_pte_alloc(dst, addr, sz);
2397 		if (!dst_pte) {
2398 			ret = -ENOMEM;
2399 			break;
2400 		}
2401 
2402 		/* If the pagetables are shared don't copy or take references */
2403 		if (dst_pte == src_pte)
2404 			continue;
2405 
2406 		dst_ptl = huge_pte_lock(h, dst, dst_pte);
2407 		src_ptl = huge_pte_lockptr(h, src, src_pte);
2408 		spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
2409 		if (!huge_pte_none(huge_ptep_get(src_pte))) {
2410 			if (cow)
2411 				huge_ptep_set_wrprotect(src, addr, src_pte);
2412 			entry = huge_ptep_get(src_pte);
2413 			ptepage = pte_page(entry);
2414 			get_page(ptepage);
2415 			page_dup_rmap(ptepage);
2416 			set_huge_pte_at(dst, addr, dst_pte, entry);
2417 		}
2418 		spin_unlock(src_ptl);
2419 		spin_unlock(dst_ptl);
2420 	}
2421 
2422 	if (cow)
2423 		mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
2424 
2425 	return ret;
2426 }
2427 
2428 static int is_hugetlb_entry_migration(pte_t pte)
2429 {
2430 	swp_entry_t swp;
2431 
2432 	if (huge_pte_none(pte) || pte_present(pte))
2433 		return 0;
2434 	swp = pte_to_swp_entry(pte);
2435 	if (non_swap_entry(swp) && is_migration_entry(swp))
2436 		return 1;
2437 	else
2438 		return 0;
2439 }
2440 
2441 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2442 {
2443 	swp_entry_t swp;
2444 
2445 	if (huge_pte_none(pte) || pte_present(pte))
2446 		return 0;
2447 	swp = pte_to_swp_entry(pte);
2448 	if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2449 		return 1;
2450 	else
2451 		return 0;
2452 }
2453 
2454 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2455 			    unsigned long start, unsigned long end,
2456 			    struct page *ref_page)
2457 {
2458 	int force_flush = 0;
2459 	struct mm_struct *mm = vma->vm_mm;
2460 	unsigned long address;
2461 	pte_t *ptep;
2462 	pte_t pte;
2463 	spinlock_t *ptl;
2464 	struct page *page;
2465 	struct hstate *h = hstate_vma(vma);
2466 	unsigned long sz = huge_page_size(h);
2467 	const unsigned long mmun_start = start;	/* For mmu_notifiers */
2468 	const unsigned long mmun_end   = end;	/* For mmu_notifiers */
2469 
2470 	WARN_ON(!is_vm_hugetlb_page(vma));
2471 	BUG_ON(start & ~huge_page_mask(h));
2472 	BUG_ON(end & ~huge_page_mask(h));
2473 
2474 	tlb_start_vma(tlb, vma);
2475 	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2476 again:
2477 	for (address = start; address < end; address += sz) {
2478 		ptep = huge_pte_offset(mm, address);
2479 		if (!ptep)
2480 			continue;
2481 
2482 		ptl = huge_pte_lock(h, mm, ptep);
2483 		if (huge_pmd_unshare(mm, &address, ptep))
2484 			goto unlock;
2485 
2486 		pte = huge_ptep_get(ptep);
2487 		if (huge_pte_none(pte))
2488 			goto unlock;
2489 
2490 		/*
2491 		 * HWPoisoned hugepage is already unmapped and dropped reference
2492 		 */
2493 		if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
2494 			huge_pte_clear(mm, address, ptep);
2495 			goto unlock;
2496 		}
2497 
2498 		page = pte_page(pte);
2499 		/*
2500 		 * If a reference page is supplied, it is because a specific
2501 		 * page is being unmapped, not a range. Ensure the page we
2502 		 * are about to unmap is the actual page of interest.
2503 		 */
2504 		if (ref_page) {
2505 			if (page != ref_page)
2506 				goto unlock;
2507 
2508 			/*
2509 			 * Mark the VMA as having unmapped its page so that
2510 			 * future faults in this VMA will fail rather than
2511 			 * looking like data was lost
2512 			 */
2513 			set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2514 		}
2515 
2516 		pte = huge_ptep_get_and_clear(mm, address, ptep);
2517 		tlb_remove_tlb_entry(tlb, ptep, address);
2518 		if (huge_pte_dirty(pte))
2519 			set_page_dirty(page);
2520 
2521 		page_remove_rmap(page);
2522 		force_flush = !__tlb_remove_page(tlb, page);
2523 		if (force_flush) {
2524 			spin_unlock(ptl);
2525 			break;
2526 		}
2527 		/* Bail out after unmapping reference page if supplied */
2528 		if (ref_page) {
2529 			spin_unlock(ptl);
2530 			break;
2531 		}
2532 unlock:
2533 		spin_unlock(ptl);
2534 	}
2535 	/*
2536 	 * mmu_gather ran out of room to batch pages, we break out of
2537 	 * the PTE lock to avoid doing the potential expensive TLB invalidate
2538 	 * and page-free while holding it.
2539 	 */
2540 	if (force_flush) {
2541 		force_flush = 0;
2542 		tlb_flush_mmu(tlb);
2543 		if (address < end && !ref_page)
2544 			goto again;
2545 	}
2546 	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2547 	tlb_end_vma(tlb, vma);
2548 }
2549 
2550 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2551 			  struct vm_area_struct *vma, unsigned long start,
2552 			  unsigned long end, struct page *ref_page)
2553 {
2554 	__unmap_hugepage_range(tlb, vma, start, end, ref_page);
2555 
2556 	/*
2557 	 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2558 	 * test will fail on a vma being torn down, and not grab a page table
2559 	 * on its way out.  We're lucky that the flag has such an appropriate
2560 	 * name, and can in fact be safely cleared here. We could clear it
2561 	 * before the __unmap_hugepage_range above, but all that's necessary
2562 	 * is to clear it before releasing the i_mmap_mutex. This works
2563 	 * because in the context this is called, the VMA is about to be
2564 	 * destroyed and the i_mmap_mutex is held.
2565 	 */
2566 	vma->vm_flags &= ~VM_MAYSHARE;
2567 }
2568 
2569 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2570 			  unsigned long end, struct page *ref_page)
2571 {
2572 	struct mm_struct *mm;
2573 	struct mmu_gather tlb;
2574 
2575 	mm = vma->vm_mm;
2576 
2577 	tlb_gather_mmu(&tlb, mm, start, end);
2578 	__unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2579 	tlb_finish_mmu(&tlb, start, end);
2580 }
2581 
2582 /*
2583  * This is called when the original mapper is failing to COW a MAP_PRIVATE
2584  * mappping it owns the reserve page for. The intention is to unmap the page
2585  * from other VMAs and let the children be SIGKILLed if they are faulting the
2586  * same region.
2587  */
2588 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2589 				struct page *page, unsigned long address)
2590 {
2591 	struct hstate *h = hstate_vma(vma);
2592 	struct vm_area_struct *iter_vma;
2593 	struct address_space *mapping;
2594 	pgoff_t pgoff;
2595 
2596 	/*
2597 	 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2598 	 * from page cache lookup which is in HPAGE_SIZE units.
2599 	 */
2600 	address = address & huge_page_mask(h);
2601 	pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2602 			vma->vm_pgoff;
2603 	mapping = file_inode(vma->vm_file)->i_mapping;
2604 
2605 	/*
2606 	 * Take the mapping lock for the duration of the table walk. As
2607 	 * this mapping should be shared between all the VMAs,
2608 	 * __unmap_hugepage_range() is called as the lock is already held
2609 	 */
2610 	mutex_lock(&mapping->i_mmap_mutex);
2611 	vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2612 		/* Do not unmap the current VMA */
2613 		if (iter_vma == vma)
2614 			continue;
2615 
2616 		/*
2617 		 * Unmap the page from other VMAs without their own reserves.
2618 		 * They get marked to be SIGKILLed if they fault in these
2619 		 * areas. This is because a future no-page fault on this VMA
2620 		 * could insert a zeroed page instead of the data existing
2621 		 * from the time of fork. This would look like data corruption
2622 		 */
2623 		if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2624 			unmap_hugepage_range(iter_vma, address,
2625 					     address + huge_page_size(h), page);
2626 	}
2627 	mutex_unlock(&mapping->i_mmap_mutex);
2628 
2629 	return 1;
2630 }
2631 
2632 /*
2633  * Hugetlb_cow() should be called with page lock of the original hugepage held.
2634  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2635  * cannot race with other handlers or page migration.
2636  * Keep the pte_same checks anyway to make transition from the mutex easier.
2637  */
2638 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2639 			unsigned long address, pte_t *ptep, pte_t pte,
2640 			struct page *pagecache_page, spinlock_t *ptl)
2641 {
2642 	struct hstate *h = hstate_vma(vma);
2643 	struct page *old_page, *new_page;
2644 	int outside_reserve = 0;
2645 	unsigned long mmun_start;	/* For mmu_notifiers */
2646 	unsigned long mmun_end;		/* For mmu_notifiers */
2647 
2648 	old_page = pte_page(pte);
2649 
2650 retry_avoidcopy:
2651 	/* If no-one else is actually using this page, avoid the copy
2652 	 * and just make the page writable */
2653 	if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2654 		page_move_anon_rmap(old_page, vma, address);
2655 		set_huge_ptep_writable(vma, address, ptep);
2656 		return 0;
2657 	}
2658 
2659 	/*
2660 	 * If the process that created a MAP_PRIVATE mapping is about to
2661 	 * perform a COW due to a shared page count, attempt to satisfy
2662 	 * the allocation without using the existing reserves. The pagecache
2663 	 * page is used to determine if the reserve at this address was
2664 	 * consumed or not. If reserves were used, a partial faulted mapping
2665 	 * at the time of fork() could consume its reserves on COW instead
2666 	 * of the full address range.
2667 	 */
2668 	if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2669 			old_page != pagecache_page)
2670 		outside_reserve = 1;
2671 
2672 	page_cache_get(old_page);
2673 
2674 	/* Drop page table lock as buddy allocator may be called */
2675 	spin_unlock(ptl);
2676 	new_page = alloc_huge_page(vma, address, outside_reserve);
2677 
2678 	if (IS_ERR(new_page)) {
2679 		long err = PTR_ERR(new_page);
2680 		page_cache_release(old_page);
2681 
2682 		/*
2683 		 * If a process owning a MAP_PRIVATE mapping fails to COW,
2684 		 * it is due to references held by a child and an insufficient
2685 		 * huge page pool. To guarantee the original mappers
2686 		 * reliability, unmap the page from child processes. The child
2687 		 * may get SIGKILLed if it later faults.
2688 		 */
2689 		if (outside_reserve) {
2690 			BUG_ON(huge_pte_none(pte));
2691 			if (unmap_ref_private(mm, vma, old_page, address)) {
2692 				BUG_ON(huge_pte_none(pte));
2693 				spin_lock(ptl);
2694 				ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2695 				if (likely(ptep &&
2696 					   pte_same(huge_ptep_get(ptep), pte)))
2697 					goto retry_avoidcopy;
2698 				/*
2699 				 * race occurs while re-acquiring page table
2700 				 * lock, and our job is done.
2701 				 */
2702 				return 0;
2703 			}
2704 			WARN_ON_ONCE(1);
2705 		}
2706 
2707 		/* Caller expects lock to be held */
2708 		spin_lock(ptl);
2709 		if (err == -ENOMEM)
2710 			return VM_FAULT_OOM;
2711 		else
2712 			return VM_FAULT_SIGBUS;
2713 	}
2714 
2715 	/*
2716 	 * When the original hugepage is shared one, it does not have
2717 	 * anon_vma prepared.
2718 	 */
2719 	if (unlikely(anon_vma_prepare(vma))) {
2720 		page_cache_release(new_page);
2721 		page_cache_release(old_page);
2722 		/* Caller expects lock to be held */
2723 		spin_lock(ptl);
2724 		return VM_FAULT_OOM;
2725 	}
2726 
2727 	copy_user_huge_page(new_page, old_page, address, vma,
2728 			    pages_per_huge_page(h));
2729 	__SetPageUptodate(new_page);
2730 
2731 	mmun_start = address & huge_page_mask(h);
2732 	mmun_end = mmun_start + huge_page_size(h);
2733 	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2734 	/*
2735 	 * Retake the page table lock to check for racing updates
2736 	 * before the page tables are altered
2737 	 */
2738 	spin_lock(ptl);
2739 	ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2740 	if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
2741 		ClearPagePrivate(new_page);
2742 
2743 		/* Break COW */
2744 		huge_ptep_clear_flush(vma, address, ptep);
2745 		set_huge_pte_at(mm, address, ptep,
2746 				make_huge_pte(vma, new_page, 1));
2747 		page_remove_rmap(old_page);
2748 		hugepage_add_new_anon_rmap(new_page, vma, address);
2749 		/* Make the old page be freed below */
2750 		new_page = old_page;
2751 	}
2752 	spin_unlock(ptl);
2753 	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2754 	page_cache_release(new_page);
2755 	page_cache_release(old_page);
2756 
2757 	/* Caller expects lock to be held */
2758 	spin_lock(ptl);
2759 	return 0;
2760 }
2761 
2762 /* Return the pagecache page at a given address within a VMA */
2763 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2764 			struct vm_area_struct *vma, unsigned long address)
2765 {
2766 	struct address_space *mapping;
2767 	pgoff_t idx;
2768 
2769 	mapping = vma->vm_file->f_mapping;
2770 	idx = vma_hugecache_offset(h, vma, address);
2771 
2772 	return find_lock_page(mapping, idx);
2773 }
2774 
2775 /*
2776  * Return whether there is a pagecache page to back given address within VMA.
2777  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2778  */
2779 static bool hugetlbfs_pagecache_present(struct hstate *h,
2780 			struct vm_area_struct *vma, unsigned long address)
2781 {
2782 	struct address_space *mapping;
2783 	pgoff_t idx;
2784 	struct page *page;
2785 
2786 	mapping = vma->vm_file->f_mapping;
2787 	idx = vma_hugecache_offset(h, vma, address);
2788 
2789 	page = find_get_page(mapping, idx);
2790 	if (page)
2791 		put_page(page);
2792 	return page != NULL;
2793 }
2794 
2795 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2796 			   struct address_space *mapping, pgoff_t idx,
2797 			   unsigned long address, pte_t *ptep, unsigned int flags)
2798 {
2799 	struct hstate *h = hstate_vma(vma);
2800 	int ret = VM_FAULT_SIGBUS;
2801 	int anon_rmap = 0;
2802 	unsigned long size;
2803 	struct page *page;
2804 	pte_t new_pte;
2805 	spinlock_t *ptl;
2806 
2807 	/*
2808 	 * Currently, we are forced to kill the process in the event the
2809 	 * original mapper has unmapped pages from the child due to a failed
2810 	 * COW. Warn that such a situation has occurred as it may not be obvious
2811 	 */
2812 	if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2813 		pr_warning("PID %d killed due to inadequate hugepage pool\n",
2814 			   current->pid);
2815 		return ret;
2816 	}
2817 
2818 	/*
2819 	 * Use page lock to guard against racing truncation
2820 	 * before we get page_table_lock.
2821 	 */
2822 retry:
2823 	page = find_lock_page(mapping, idx);
2824 	if (!page) {
2825 		size = i_size_read(mapping->host) >> huge_page_shift(h);
2826 		if (idx >= size)
2827 			goto out;
2828 		page = alloc_huge_page(vma, address, 0);
2829 		if (IS_ERR(page)) {
2830 			ret = PTR_ERR(page);
2831 			if (ret == -ENOMEM)
2832 				ret = VM_FAULT_OOM;
2833 			else
2834 				ret = VM_FAULT_SIGBUS;
2835 			goto out;
2836 		}
2837 		clear_huge_page(page, address, pages_per_huge_page(h));
2838 		__SetPageUptodate(page);
2839 
2840 		if (vma->vm_flags & VM_MAYSHARE) {
2841 			int err;
2842 			struct inode *inode = mapping->host;
2843 
2844 			err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2845 			if (err) {
2846 				put_page(page);
2847 				if (err == -EEXIST)
2848 					goto retry;
2849 				goto out;
2850 			}
2851 			ClearPagePrivate(page);
2852 
2853 			spin_lock(&inode->i_lock);
2854 			inode->i_blocks += blocks_per_huge_page(h);
2855 			spin_unlock(&inode->i_lock);
2856 		} else {
2857 			lock_page(page);
2858 			if (unlikely(anon_vma_prepare(vma))) {
2859 				ret = VM_FAULT_OOM;
2860 				goto backout_unlocked;
2861 			}
2862 			anon_rmap = 1;
2863 		}
2864 	} else {
2865 		/*
2866 		 * If memory error occurs between mmap() and fault, some process
2867 		 * don't have hwpoisoned swap entry for errored virtual address.
2868 		 * So we need to block hugepage fault by PG_hwpoison bit check.
2869 		 */
2870 		if (unlikely(PageHWPoison(page))) {
2871 			ret = VM_FAULT_HWPOISON |
2872 				VM_FAULT_SET_HINDEX(hstate_index(h));
2873 			goto backout_unlocked;
2874 		}
2875 	}
2876 
2877 	/*
2878 	 * If we are going to COW a private mapping later, we examine the
2879 	 * pending reservations for this page now. This will ensure that
2880 	 * any allocations necessary to record that reservation occur outside
2881 	 * the spinlock.
2882 	 */
2883 	if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2884 		if (vma_needs_reservation(h, vma, address) < 0) {
2885 			ret = VM_FAULT_OOM;
2886 			goto backout_unlocked;
2887 		}
2888 
2889 	ptl = huge_pte_lockptr(h, mm, ptep);
2890 	spin_lock(ptl);
2891 	size = i_size_read(mapping->host) >> huge_page_shift(h);
2892 	if (idx >= size)
2893 		goto backout;
2894 
2895 	ret = 0;
2896 	if (!huge_pte_none(huge_ptep_get(ptep)))
2897 		goto backout;
2898 
2899 	if (anon_rmap) {
2900 		ClearPagePrivate(page);
2901 		hugepage_add_new_anon_rmap(page, vma, address);
2902 	} else
2903 		page_dup_rmap(page);
2904 	new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2905 				&& (vma->vm_flags & VM_SHARED)));
2906 	set_huge_pte_at(mm, address, ptep, new_pte);
2907 
2908 	if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2909 		/* Optimization, do the COW without a second fault */
2910 		ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
2911 	}
2912 
2913 	spin_unlock(ptl);
2914 	unlock_page(page);
2915 out:
2916 	return ret;
2917 
2918 backout:
2919 	spin_unlock(ptl);
2920 backout_unlocked:
2921 	unlock_page(page);
2922 	put_page(page);
2923 	goto out;
2924 }
2925 
2926 #ifdef CONFIG_SMP
2927 static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
2928 			    struct vm_area_struct *vma,
2929 			    struct address_space *mapping,
2930 			    pgoff_t idx, unsigned long address)
2931 {
2932 	unsigned long key[2];
2933 	u32 hash;
2934 
2935 	if (vma->vm_flags & VM_SHARED) {
2936 		key[0] = (unsigned long) mapping;
2937 		key[1] = idx;
2938 	} else {
2939 		key[0] = (unsigned long) mm;
2940 		key[1] = address >> huge_page_shift(h);
2941 	}
2942 
2943 	hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
2944 
2945 	return hash & (num_fault_mutexes - 1);
2946 }
2947 #else
2948 /*
2949  * For uniprocesor systems we always use a single mutex, so just
2950  * return 0 and avoid the hashing overhead.
2951  */
2952 static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
2953 			    struct vm_area_struct *vma,
2954 			    struct address_space *mapping,
2955 			    pgoff_t idx, unsigned long address)
2956 {
2957 	return 0;
2958 }
2959 #endif
2960 
2961 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2962 			unsigned long address, unsigned int flags)
2963 {
2964 	pte_t *ptep, entry;
2965 	spinlock_t *ptl;
2966 	int ret;
2967 	u32 hash;
2968 	pgoff_t idx;
2969 	struct page *page = NULL;
2970 	struct page *pagecache_page = NULL;
2971 	struct hstate *h = hstate_vma(vma);
2972 	struct address_space *mapping;
2973 
2974 	address &= huge_page_mask(h);
2975 
2976 	ptep = huge_pte_offset(mm, address);
2977 	if (ptep) {
2978 		entry = huge_ptep_get(ptep);
2979 		if (unlikely(is_hugetlb_entry_migration(entry))) {
2980 			migration_entry_wait_huge(vma, mm, ptep);
2981 			return 0;
2982 		} else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2983 			return VM_FAULT_HWPOISON_LARGE |
2984 				VM_FAULT_SET_HINDEX(hstate_index(h));
2985 	}
2986 
2987 	ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2988 	if (!ptep)
2989 		return VM_FAULT_OOM;
2990 
2991 	mapping = vma->vm_file->f_mapping;
2992 	idx = vma_hugecache_offset(h, vma, address);
2993 
2994 	/*
2995 	 * Serialize hugepage allocation and instantiation, so that we don't
2996 	 * get spurious allocation failures if two CPUs race to instantiate
2997 	 * the same page in the page cache.
2998 	 */
2999 	hash = fault_mutex_hash(h, mm, vma, mapping, idx, address);
3000 	mutex_lock(&htlb_fault_mutex_table[hash]);
3001 
3002 	entry = huge_ptep_get(ptep);
3003 	if (huge_pte_none(entry)) {
3004 		ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3005 		goto out_mutex;
3006 	}
3007 
3008 	ret = 0;
3009 
3010 	/*
3011 	 * If we are going to COW the mapping later, we examine the pending
3012 	 * reservations for this page now. This will ensure that any
3013 	 * allocations necessary to record that reservation occur outside the
3014 	 * spinlock. For private mappings, we also lookup the pagecache
3015 	 * page now as it is used to determine if a reservation has been
3016 	 * consumed.
3017 	 */
3018 	if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3019 		if (vma_needs_reservation(h, vma, address) < 0) {
3020 			ret = VM_FAULT_OOM;
3021 			goto out_mutex;
3022 		}
3023 
3024 		if (!(vma->vm_flags & VM_MAYSHARE))
3025 			pagecache_page = hugetlbfs_pagecache_page(h,
3026 								vma, address);
3027 	}
3028 
3029 	/*
3030 	 * hugetlb_cow() requires page locks of pte_page(entry) and
3031 	 * pagecache_page, so here we need take the former one
3032 	 * when page != pagecache_page or !pagecache_page.
3033 	 * Note that locking order is always pagecache_page -> page,
3034 	 * so no worry about deadlock.
3035 	 */
3036 	page = pte_page(entry);
3037 	get_page(page);
3038 	if (page != pagecache_page)
3039 		lock_page(page);
3040 
3041 	ptl = huge_pte_lockptr(h, mm, ptep);
3042 	spin_lock(ptl);
3043 	/* Check for a racing update before calling hugetlb_cow */
3044 	if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3045 		goto out_ptl;
3046 
3047 
3048 	if (flags & FAULT_FLAG_WRITE) {
3049 		if (!huge_pte_write(entry)) {
3050 			ret = hugetlb_cow(mm, vma, address, ptep, entry,
3051 					pagecache_page, ptl);
3052 			goto out_ptl;
3053 		}
3054 		entry = huge_pte_mkdirty(entry);
3055 	}
3056 	entry = pte_mkyoung(entry);
3057 	if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3058 						flags & FAULT_FLAG_WRITE))
3059 		update_mmu_cache(vma, address, ptep);
3060 
3061 out_ptl:
3062 	spin_unlock(ptl);
3063 
3064 	if (pagecache_page) {
3065 		unlock_page(pagecache_page);
3066 		put_page(pagecache_page);
3067 	}
3068 	if (page != pagecache_page)
3069 		unlock_page(page);
3070 	put_page(page);
3071 
3072 out_mutex:
3073 	mutex_unlock(&htlb_fault_mutex_table[hash]);
3074 	return ret;
3075 }
3076 
3077 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3078 			 struct page **pages, struct vm_area_struct **vmas,
3079 			 unsigned long *position, unsigned long *nr_pages,
3080 			 long i, unsigned int flags)
3081 {
3082 	unsigned long pfn_offset;
3083 	unsigned long vaddr = *position;
3084 	unsigned long remainder = *nr_pages;
3085 	struct hstate *h = hstate_vma(vma);
3086 
3087 	while (vaddr < vma->vm_end && remainder) {
3088 		pte_t *pte;
3089 		spinlock_t *ptl = NULL;
3090 		int absent;
3091 		struct page *page;
3092 
3093 		/*
3094 		 * Some archs (sparc64, sh*) have multiple pte_ts to
3095 		 * each hugepage.  We have to make sure we get the
3096 		 * first, for the page indexing below to work.
3097 		 *
3098 		 * Note that page table lock is not held when pte is null.
3099 		 */
3100 		pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3101 		if (pte)
3102 			ptl = huge_pte_lock(h, mm, pte);
3103 		absent = !pte || huge_pte_none(huge_ptep_get(pte));
3104 
3105 		/*
3106 		 * When coredumping, it suits get_dump_page if we just return
3107 		 * an error where there's an empty slot with no huge pagecache
3108 		 * to back it.  This way, we avoid allocating a hugepage, and
3109 		 * the sparse dumpfile avoids allocating disk blocks, but its
3110 		 * huge holes still show up with zeroes where they need to be.
3111 		 */
3112 		if (absent && (flags & FOLL_DUMP) &&
3113 		    !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3114 			if (pte)
3115 				spin_unlock(ptl);
3116 			remainder = 0;
3117 			break;
3118 		}
3119 
3120 		/*
3121 		 * We need call hugetlb_fault for both hugepages under migration
3122 		 * (in which case hugetlb_fault waits for the migration,) and
3123 		 * hwpoisoned hugepages (in which case we need to prevent the
3124 		 * caller from accessing to them.) In order to do this, we use
3125 		 * here is_swap_pte instead of is_hugetlb_entry_migration and
3126 		 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3127 		 * both cases, and because we can't follow correct pages
3128 		 * directly from any kind of swap entries.
3129 		 */
3130 		if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3131 		    ((flags & FOLL_WRITE) &&
3132 		      !huge_pte_write(huge_ptep_get(pte)))) {
3133 			int ret;
3134 
3135 			if (pte)
3136 				spin_unlock(ptl);
3137 			ret = hugetlb_fault(mm, vma, vaddr,
3138 				(flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3139 			if (!(ret & VM_FAULT_ERROR))
3140 				continue;
3141 
3142 			remainder = 0;
3143 			break;
3144 		}
3145 
3146 		pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3147 		page = pte_page(huge_ptep_get(pte));
3148 same_page:
3149 		if (pages) {
3150 			pages[i] = mem_map_offset(page, pfn_offset);
3151 			get_page_foll(pages[i]);
3152 		}
3153 
3154 		if (vmas)
3155 			vmas[i] = vma;
3156 
3157 		vaddr += PAGE_SIZE;
3158 		++pfn_offset;
3159 		--remainder;
3160 		++i;
3161 		if (vaddr < vma->vm_end && remainder &&
3162 				pfn_offset < pages_per_huge_page(h)) {
3163 			/*
3164 			 * We use pfn_offset to avoid touching the pageframes
3165 			 * of this compound page.
3166 			 */
3167 			goto same_page;
3168 		}
3169 		spin_unlock(ptl);
3170 	}
3171 	*nr_pages = remainder;
3172 	*position = vaddr;
3173 
3174 	return i ? i : -EFAULT;
3175 }
3176 
3177 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3178 		unsigned long address, unsigned long end, pgprot_t newprot)
3179 {
3180 	struct mm_struct *mm = vma->vm_mm;
3181 	unsigned long start = address;
3182 	pte_t *ptep;
3183 	pte_t pte;
3184 	struct hstate *h = hstate_vma(vma);
3185 	unsigned long pages = 0;
3186 
3187 	BUG_ON(address >= end);
3188 	flush_cache_range(vma, address, end);
3189 
3190 	mmu_notifier_invalidate_range_start(mm, start, end);
3191 	mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3192 	for (; address < end; address += huge_page_size(h)) {
3193 		spinlock_t *ptl;
3194 		ptep = huge_pte_offset(mm, address);
3195 		if (!ptep)
3196 			continue;
3197 		ptl = huge_pte_lock(h, mm, ptep);
3198 		if (huge_pmd_unshare(mm, &address, ptep)) {
3199 			pages++;
3200 			spin_unlock(ptl);
3201 			continue;
3202 		}
3203 		if (!huge_pte_none(huge_ptep_get(ptep))) {
3204 			pte = huge_ptep_get_and_clear(mm, address, ptep);
3205 			pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3206 			pte = arch_make_huge_pte(pte, vma, NULL, 0);
3207 			set_huge_pte_at(mm, address, ptep, pte);
3208 			pages++;
3209 		}
3210 		spin_unlock(ptl);
3211 	}
3212 	/*
3213 	 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3214 	 * may have cleared our pud entry and done put_page on the page table:
3215 	 * once we release i_mmap_mutex, another task can do the final put_page
3216 	 * and that page table be reused and filled with junk.
3217 	 */
3218 	flush_tlb_range(vma, start, end);
3219 	mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3220 	mmu_notifier_invalidate_range_end(mm, start, end);
3221 
3222 	return pages << h->order;
3223 }
3224 
3225 int hugetlb_reserve_pages(struct inode *inode,
3226 					long from, long to,
3227 					struct vm_area_struct *vma,
3228 					vm_flags_t vm_flags)
3229 {
3230 	long ret, chg;
3231 	struct hstate *h = hstate_inode(inode);
3232 	struct hugepage_subpool *spool = subpool_inode(inode);
3233 	struct resv_map *resv_map;
3234 
3235 	/*
3236 	 * Only apply hugepage reservation if asked. At fault time, an
3237 	 * attempt will be made for VM_NORESERVE to allocate a page
3238 	 * without using reserves
3239 	 */
3240 	if (vm_flags & VM_NORESERVE)
3241 		return 0;
3242 
3243 	/*
3244 	 * Shared mappings base their reservation on the number of pages that
3245 	 * are already allocated on behalf of the file. Private mappings need
3246 	 * to reserve the full area even if read-only as mprotect() may be
3247 	 * called to make the mapping read-write. Assume !vma is a shm mapping
3248 	 */
3249 	if (!vma || vma->vm_flags & VM_MAYSHARE) {
3250 		resv_map = inode_resv_map(inode);
3251 
3252 		chg = region_chg(resv_map, from, to);
3253 
3254 	} else {
3255 		resv_map = resv_map_alloc();
3256 		if (!resv_map)
3257 			return -ENOMEM;
3258 
3259 		chg = to - from;
3260 
3261 		set_vma_resv_map(vma, resv_map);
3262 		set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3263 	}
3264 
3265 	if (chg < 0) {
3266 		ret = chg;
3267 		goto out_err;
3268 	}
3269 
3270 	/* There must be enough pages in the subpool for the mapping */
3271 	if (hugepage_subpool_get_pages(spool, chg)) {
3272 		ret = -ENOSPC;
3273 		goto out_err;
3274 	}
3275 
3276 	/*
3277 	 * Check enough hugepages are available for the reservation.
3278 	 * Hand the pages back to the subpool if there are not
3279 	 */
3280 	ret = hugetlb_acct_memory(h, chg);
3281 	if (ret < 0) {
3282 		hugepage_subpool_put_pages(spool, chg);
3283 		goto out_err;
3284 	}
3285 
3286 	/*
3287 	 * Account for the reservations made. Shared mappings record regions
3288 	 * that have reservations as they are shared by multiple VMAs.
3289 	 * When the last VMA disappears, the region map says how much
3290 	 * the reservation was and the page cache tells how much of
3291 	 * the reservation was consumed. Private mappings are per-VMA and
3292 	 * only the consumed reservations are tracked. When the VMA
3293 	 * disappears, the original reservation is the VMA size and the
3294 	 * consumed reservations are stored in the map. Hence, nothing
3295 	 * else has to be done for private mappings here
3296 	 */
3297 	if (!vma || vma->vm_flags & VM_MAYSHARE)
3298 		region_add(resv_map, from, to);
3299 	return 0;
3300 out_err:
3301 	if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3302 		kref_put(&resv_map->refs, resv_map_release);
3303 	return ret;
3304 }
3305 
3306 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3307 {
3308 	struct hstate *h = hstate_inode(inode);
3309 	struct resv_map *resv_map = inode_resv_map(inode);
3310 	long chg = 0;
3311 	struct hugepage_subpool *spool = subpool_inode(inode);
3312 
3313 	if (resv_map)
3314 		chg = region_truncate(resv_map, offset);
3315 	spin_lock(&inode->i_lock);
3316 	inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3317 	spin_unlock(&inode->i_lock);
3318 
3319 	hugepage_subpool_put_pages(spool, (chg - freed));
3320 	hugetlb_acct_memory(h, -(chg - freed));
3321 }
3322 
3323 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3324 static unsigned long page_table_shareable(struct vm_area_struct *svma,
3325 				struct vm_area_struct *vma,
3326 				unsigned long addr, pgoff_t idx)
3327 {
3328 	unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3329 				svma->vm_start;
3330 	unsigned long sbase = saddr & PUD_MASK;
3331 	unsigned long s_end = sbase + PUD_SIZE;
3332 
3333 	/* Allow segments to share if only one is marked locked */
3334 	unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3335 	unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3336 
3337 	/*
3338 	 * match the virtual addresses, permission and the alignment of the
3339 	 * page table page.
3340 	 */
3341 	if (pmd_index(addr) != pmd_index(saddr) ||
3342 	    vm_flags != svm_flags ||
3343 	    sbase < svma->vm_start || svma->vm_end < s_end)
3344 		return 0;
3345 
3346 	return saddr;
3347 }
3348 
3349 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3350 {
3351 	unsigned long base = addr & PUD_MASK;
3352 	unsigned long end = base + PUD_SIZE;
3353 
3354 	/*
3355 	 * check on proper vm_flags and page table alignment
3356 	 */
3357 	if (vma->vm_flags & VM_MAYSHARE &&
3358 	    vma->vm_start <= base && end <= vma->vm_end)
3359 		return 1;
3360 	return 0;
3361 }
3362 
3363 /*
3364  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3365  * and returns the corresponding pte. While this is not necessary for the
3366  * !shared pmd case because we can allocate the pmd later as well, it makes the
3367  * code much cleaner. pmd allocation is essential for the shared case because
3368  * pud has to be populated inside the same i_mmap_mutex section - otherwise
3369  * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3370  * bad pmd for sharing.
3371  */
3372 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3373 {
3374 	struct vm_area_struct *vma = find_vma(mm, addr);
3375 	struct address_space *mapping = vma->vm_file->f_mapping;
3376 	pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3377 			vma->vm_pgoff;
3378 	struct vm_area_struct *svma;
3379 	unsigned long saddr;
3380 	pte_t *spte = NULL;
3381 	pte_t *pte;
3382 	spinlock_t *ptl;
3383 
3384 	if (!vma_shareable(vma, addr))
3385 		return (pte_t *)pmd_alloc(mm, pud, addr);
3386 
3387 	mutex_lock(&mapping->i_mmap_mutex);
3388 	vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3389 		if (svma == vma)
3390 			continue;
3391 
3392 		saddr = page_table_shareable(svma, vma, addr, idx);
3393 		if (saddr) {
3394 			spte = huge_pte_offset(svma->vm_mm, saddr);
3395 			if (spte) {
3396 				get_page(virt_to_page(spte));
3397 				break;
3398 			}
3399 		}
3400 	}
3401 
3402 	if (!spte)
3403 		goto out;
3404 
3405 	ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
3406 	spin_lock(ptl);
3407 	if (pud_none(*pud))
3408 		pud_populate(mm, pud,
3409 				(pmd_t *)((unsigned long)spte & PAGE_MASK));
3410 	else
3411 		put_page(virt_to_page(spte));
3412 	spin_unlock(ptl);
3413 out:
3414 	pte = (pte_t *)pmd_alloc(mm, pud, addr);
3415 	mutex_unlock(&mapping->i_mmap_mutex);
3416 	return pte;
3417 }
3418 
3419 /*
3420  * unmap huge page backed by shared pte.
3421  *
3422  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
3423  * indicated by page_count > 1, unmap is achieved by clearing pud and
3424  * decrementing the ref count. If count == 1, the pte page is not shared.
3425  *
3426  * called with page table lock held.
3427  *
3428  * returns: 1 successfully unmapped a shared pte page
3429  *	    0 the underlying pte page is not shared, or it is the last user
3430  */
3431 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3432 {
3433 	pgd_t *pgd = pgd_offset(mm, *addr);
3434 	pud_t *pud = pud_offset(pgd, *addr);
3435 
3436 	BUG_ON(page_count(virt_to_page(ptep)) == 0);
3437 	if (page_count(virt_to_page(ptep)) == 1)
3438 		return 0;
3439 
3440 	pud_clear(pud);
3441 	put_page(virt_to_page(ptep));
3442 	*addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3443 	return 1;
3444 }
3445 #define want_pmd_share()	(1)
3446 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3447 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3448 {
3449 	return NULL;
3450 }
3451 #define want_pmd_share()	(0)
3452 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3453 
3454 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3455 pte_t *huge_pte_alloc(struct mm_struct *mm,
3456 			unsigned long addr, unsigned long sz)
3457 {
3458 	pgd_t *pgd;
3459 	pud_t *pud;
3460 	pte_t *pte = NULL;
3461 
3462 	pgd = pgd_offset(mm, addr);
3463 	pud = pud_alloc(mm, pgd, addr);
3464 	if (pud) {
3465 		if (sz == PUD_SIZE) {
3466 			pte = (pte_t *)pud;
3467 		} else {
3468 			BUG_ON(sz != PMD_SIZE);
3469 			if (want_pmd_share() && pud_none(*pud))
3470 				pte = huge_pmd_share(mm, addr, pud);
3471 			else
3472 				pte = (pte_t *)pmd_alloc(mm, pud, addr);
3473 		}
3474 	}
3475 	BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3476 
3477 	return pte;
3478 }
3479 
3480 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3481 {
3482 	pgd_t *pgd;
3483 	pud_t *pud;
3484 	pmd_t *pmd = NULL;
3485 
3486 	pgd = pgd_offset(mm, addr);
3487 	if (pgd_present(*pgd)) {
3488 		pud = pud_offset(pgd, addr);
3489 		if (pud_present(*pud)) {
3490 			if (pud_huge(*pud))
3491 				return (pte_t *)pud;
3492 			pmd = pmd_offset(pud, addr);
3493 		}
3494 	}
3495 	return (pte_t *) pmd;
3496 }
3497 
3498 struct page *
3499 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3500 		pmd_t *pmd, int write)
3501 {
3502 	struct page *page;
3503 
3504 	page = pte_page(*(pte_t *)pmd);
3505 	if (page)
3506 		page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
3507 	return page;
3508 }
3509 
3510 struct page *
3511 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3512 		pud_t *pud, int write)
3513 {
3514 	struct page *page;
3515 
3516 	page = pte_page(*(pte_t *)pud);
3517 	if (page)
3518 		page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
3519 	return page;
3520 }
3521 
3522 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3523 
3524 /* Can be overriden by architectures */
3525 struct page * __weak
3526 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3527 	       pud_t *pud, int write)
3528 {
3529 	BUG();
3530 	return NULL;
3531 }
3532 
3533 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3534 
3535 #ifdef CONFIG_MEMORY_FAILURE
3536 
3537 /* Should be called in hugetlb_lock */
3538 static int is_hugepage_on_freelist(struct page *hpage)
3539 {
3540 	struct page *page;
3541 	struct page *tmp;
3542 	struct hstate *h = page_hstate(hpage);
3543 	int nid = page_to_nid(hpage);
3544 
3545 	list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3546 		if (page == hpage)
3547 			return 1;
3548 	return 0;
3549 }
3550 
3551 /*
3552  * This function is called from memory failure code.
3553  * Assume the caller holds page lock of the head page.
3554  */
3555 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3556 {
3557 	struct hstate *h = page_hstate(hpage);
3558 	int nid = page_to_nid(hpage);
3559 	int ret = -EBUSY;
3560 
3561 	spin_lock(&hugetlb_lock);
3562 	if (is_hugepage_on_freelist(hpage)) {
3563 		/*
3564 		 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3565 		 * but dangling hpage->lru can trigger list-debug warnings
3566 		 * (this happens when we call unpoison_memory() on it),
3567 		 * so let it point to itself with list_del_init().
3568 		 */
3569 		list_del_init(&hpage->lru);
3570 		set_page_refcounted(hpage);
3571 		h->free_huge_pages--;
3572 		h->free_huge_pages_node[nid]--;
3573 		ret = 0;
3574 	}
3575 	spin_unlock(&hugetlb_lock);
3576 	return ret;
3577 }
3578 #endif
3579 
3580 bool isolate_huge_page(struct page *page, struct list_head *list)
3581 {
3582 	VM_BUG_ON_PAGE(!PageHead(page), page);
3583 	if (!get_page_unless_zero(page))
3584 		return false;
3585 	spin_lock(&hugetlb_lock);
3586 	list_move_tail(&page->lru, list);
3587 	spin_unlock(&hugetlb_lock);
3588 	return true;
3589 }
3590 
3591 void putback_active_hugepage(struct page *page)
3592 {
3593 	VM_BUG_ON_PAGE(!PageHead(page), page);
3594 	spin_lock(&hugetlb_lock);
3595 	list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
3596 	spin_unlock(&hugetlb_lock);
3597 	put_page(page);
3598 }
3599 
3600 bool is_hugepage_active(struct page *page)
3601 {
3602 	VM_BUG_ON_PAGE(!PageHuge(page), page);
3603 	/*
3604 	 * This function can be called for a tail page because the caller,
3605 	 * scan_movable_pages, scans through a given pfn-range which typically
3606 	 * covers one memory block. In systems using gigantic hugepage (1GB
3607 	 * for x86_64,) a hugepage is larger than a memory block, and we don't
3608 	 * support migrating such large hugepages for now, so return false
3609 	 * when called for tail pages.
3610 	 */
3611 	if (PageTail(page))
3612 		return false;
3613 	/*
3614 	 * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3615 	 * so we should return false for them.
3616 	 */
3617 	if (unlikely(PageHWPoison(page)))
3618 		return false;
3619 	return page_count(page) > 0;
3620 }
3621