source: src/linux/universal/linux-3.18/mm/hugetlb.c @ 31885

Last change on this file since 31885 was 31885, checked in by brainslayer, 3 months ago

update

File size: 100.9 KB
Line 
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
38unsigned long hugepages_treat_as_movable;
39
40int hugetlb_max_hstate __read_mostly;
41unsigned int default_hstate_idx;
42struct hstate hstates[HUGE_MAX_HSTATE];
43
44__initdata LIST_HEAD(huge_boot_pages);
45
46/* for command line parsing */
47static struct hstate * __initdata parsed_hstate;
48static unsigned long __initdata default_hstate_max_huge_pages;
49static unsigned long __initdata default_hstate_size;
50
51/*
52 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
53 * free_huge_pages, and surplus_huge_pages.
54 */
55DEFINE_SPINLOCK(hugetlb_lock);
56
57/*
58 * Serializes faults on the same logical page.  This is used to
59 * prevent spurious OOMs when the hugepage pool is fully utilized.
60 */
61static int num_fault_mutexes;
62static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp;
63
64static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
65{
66        bool free = (spool->count == 0) && (spool->used_hpages == 0);
67
68        spin_unlock(&spool->lock);
69
70        /* If no pages are used, and no other handles to the subpool
71         * remain, free the subpool the subpool remain */
72        if (free)
73                kfree(spool);
74}
75
76struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
77{
78        struct hugepage_subpool *spool;
79
80        spool = kmalloc(sizeof(*spool), GFP_KERNEL);
81        if (!spool)
82                return NULL;
83
84        spin_lock_init(&spool->lock);
85        spool->count = 1;
86        spool->max_hpages = nr_blocks;
87        spool->used_hpages = 0;
88
89        return spool;
90}
91
92void hugepage_put_subpool(struct hugepage_subpool *spool)
93{
94        spin_lock(&spool->lock);
95        BUG_ON(!spool->count);
96        spool->count--;
97        unlock_or_release_subpool(spool);
98}
99
100static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
101                                      long delta)
102{
103        int ret = 0;
104
105        if (!spool)
106                return 0;
107
108        spin_lock(&spool->lock);
109        if ((spool->used_hpages + delta) <= spool->max_hpages) {
110                spool->used_hpages += delta;
111        } else {
112                ret = -ENOMEM;
113        }
114        spin_unlock(&spool->lock);
115
116        return ret;
117}
118
119static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
120                                       long delta)
121{
122        if (!spool)
123                return;
124
125        spin_lock(&spool->lock);
126        spool->used_hpages -= delta;
127        /* If hugetlbfs_put_super couldn't free spool due to
128        * an outstanding quota reference, free it now. */
129        unlock_or_release_subpool(spool);
130}
131
132static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
133{
134        return HUGETLBFS_SB(inode->i_sb)->spool;
135}
136
137static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
138{
139        return subpool_inode(file_inode(vma->vm_file));
140}
141
142/*
143 * Region tracking -- allows tracking of reservations and instantiated pages
144 *                    across the pages in a mapping.
145 *
146 * The region data structures are embedded into a resv_map and
147 * protected by a resv_map's lock
148 */
149struct file_region {
150        struct list_head link;
151        long from;
152        long to;
153};
154
155static long region_add(struct resv_map *resv, long f, long t)
156{
157        struct list_head *head = &resv->regions;
158        struct file_region *rg, *nrg, *trg;
159
160        spin_lock(&resv->lock);
161        /* Locate the region we are either in or before. */
162        list_for_each_entry(rg, head, link)
163                if (f <= rg->to)
164                        break;
165
166        /* Round our left edge to the current segment if it encloses us. */
167        if (f > rg->from)
168                f = rg->from;
169
170        /* Check for and consume any regions we now overlap with. */
171        nrg = rg;
172        list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
173                if (&rg->link == head)
174                        break;
175                if (rg->from > t)
176                        break;
177
178                /* If this area reaches higher then extend our area to
179                 * include it completely.  If this is not the first area
180                 * which we intend to reuse, free it. */
181                if (rg->to > t)
182                        t = rg->to;
183                if (rg != nrg) {
184                        list_del(&rg->link);
185                        kfree(rg);
186                }
187        }
188        nrg->from = f;
189        nrg->to = t;
190        spin_unlock(&resv->lock);
191        return 0;
192}
193
194static long region_chg(struct resv_map *resv, long f, long t)
195{
196        struct list_head *head = &resv->regions;
197        struct file_region *rg, *nrg = NULL;
198        long chg = 0;
199
200retry:
201        spin_lock(&resv->lock);
202        /* Locate the region we are before or in. */
203        list_for_each_entry(rg, head, link)
204                if (f <= rg->to)
205                        break;
206
207        /* If we are below the current region then a new region is required.
208         * Subtle, allocate a new region at the position but make it zero
209         * size such that we can guarantee to record the reservation. */
210        if (&rg->link == head || t < rg->from) {
211                if (!nrg) {
212                        spin_unlock(&resv->lock);
213                        nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
214                        if (!nrg)
215                                return -ENOMEM;
216
217                        nrg->from = f;
218                        nrg->to   = f;
219                        INIT_LIST_HEAD(&nrg->link);
220                        goto retry;
221                }
222
223                list_add(&nrg->link, rg->link.prev);
224                chg = t - f;
225                goto out_nrg;
226        }
227
228        /* Round our left edge to the current segment if it encloses us. */
229        if (f > rg->from)
230                f = rg->from;
231        chg = t - f;
232
233        /* Check for and consume any regions we now overlap with. */
234        list_for_each_entry(rg, rg->link.prev, link) {
235                if (&rg->link == head)
236                        break;
237                if (rg->from > t)
238                        goto out;
239
240                /* We overlap with this area, if it extends further than
241                 * us then we must extend ourselves.  Account for its
242                 * existing reservation. */
243                if (rg->to > t) {
244                        chg += rg->to - t;
245                        t = rg->to;
246                }
247                chg -= rg->to - rg->from;
248        }
249
250out:
251        spin_unlock(&resv->lock);
252        /*  We already know we raced and no longer need the new region */
253        kfree(nrg);
254        return chg;
255out_nrg:
256        spin_unlock(&resv->lock);
257        return chg;
258}
259
260static long region_truncate(struct resv_map *resv, long end)
261{
262        struct list_head *head = &resv->regions;
263        struct file_region *rg, *trg;
264        long chg = 0;
265
266        spin_lock(&resv->lock);
267        /* Locate the region we are either in or before. */
268        list_for_each_entry(rg, head, link)
269                if (end <= rg->to)
270                        break;
271        if (&rg->link == head)
272                goto out;
273
274        /* If we are in the middle of a region then adjust it. */
275        if (end > rg->from) {
276                chg = rg->to - end;
277                rg->to = end;
278                rg = list_entry(rg->link.next, typeof(*rg), link);
279        }
280
281        /* Drop any remaining regions. */
282        list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
283                if (&rg->link == head)
284                        break;
285                chg += rg->to - rg->from;
286                list_del(&rg->link);
287                kfree(rg);
288        }
289
290out:
291        spin_unlock(&resv->lock);
292        return chg;
293}
294
295static long region_count(struct resv_map *resv, long f, long t)
296{
297        struct list_head *head = &resv->regions;
298        struct file_region *rg;
299        long chg = 0;
300
301        spin_lock(&resv->lock);
302        /* Locate each segment we overlap with, and count that overlap. */
303        list_for_each_entry(rg, head, link) {
304                long seg_from;
305                long seg_to;
306
307                if (rg->to <= f)
308                        continue;
309                if (rg->from >= t)
310                        break;
311
312                seg_from = max(rg->from, f);
313                seg_to = min(rg->to, t);
314
315                chg += seg_to - seg_from;
316        }
317        spin_unlock(&resv->lock);
318
319        return chg;
320}
321
322/*
323 * Convert the address within this vma to the page offset within
324 * the mapping, in pagecache page units; huge pages here.
325 */
326static pgoff_t vma_hugecache_offset(struct hstate *h,
327                        struct vm_area_struct *vma, unsigned long address)
328{
329        return ((address - vma->vm_start) >> huge_page_shift(h)) +
330                        (vma->vm_pgoff >> huge_page_order(h));
331}
332
333pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
334                                     unsigned long address)
335{
336        return vma_hugecache_offset(hstate_vma(vma), vma, address);
337}
338
339/*
340 * Return the size of the pages allocated when backing a VMA. In the majority
341 * cases this will be same size as used by the page table entries.
342 */
343unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
344{
345        struct hstate *hstate;
346
347        if (!is_vm_hugetlb_page(vma))
348                return PAGE_SIZE;
349
350        hstate = hstate_vma(vma);
351
352        return 1UL << huge_page_shift(hstate);
353}
354EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
355
356/*
357 * Return the page size being used by the MMU to back a VMA. In the majority
358 * of cases, the page size used by the kernel matches the MMU size. On
359 * architectures where it differs, an architecture-specific version of this
360 * function is required.
361 */
362#ifndef vma_mmu_pagesize
363unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
364{
365        return vma_kernel_pagesize(vma);
366}
367#endif
368
369/*
370 * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
371 * bits of the reservation map pointer, which are always clear due to
372 * alignment.
373 */
374#define HPAGE_RESV_OWNER    (1UL << 0)
375#define HPAGE_RESV_UNMAPPED (1UL << 1)
376#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
377
378/*
379 * These helpers are used to track how many pages are reserved for
380 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
381 * is guaranteed to have their future faults succeed.
382 *
383 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
384 * the reserve counters are updated with the hugetlb_lock held. It is safe
385 * to reset the VMA at fork() time as it is not in use yet and there is no
386 * chance of the global counters getting corrupted as a result of the values.
387 *
388 * The private mapping reservation is represented in a subtly different
389 * manner to a shared mapping.  A shared mapping has a region map associated
390 * with the underlying file, this region map represents the backing file
391 * pages which have ever had a reservation assigned which this persists even
392 * after the page is instantiated.  A private mapping has a region map
393 * associated with the original mmap which is attached to all VMAs which
394 * reference it, this region map represents those offsets which have consumed
395 * reservation ie. where pages have been instantiated.
396 */
397static unsigned long get_vma_private_data(struct vm_area_struct *vma)
398{
399        return (unsigned long)vma->vm_private_data;
400}
401
402static void set_vma_private_data(struct vm_area_struct *vma,
403                                                        unsigned long value)
404{
405        vma->vm_private_data = (void *)value;
406}
407
408struct resv_map *resv_map_alloc(void)
409{
410        struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
411        if (!resv_map)
412                return NULL;
413
414        kref_init(&resv_map->refs);
415        spin_lock_init(&resv_map->lock);
416        INIT_LIST_HEAD(&resv_map->regions);
417
418        return resv_map;
419}
420
421void resv_map_release(struct kref *ref)
422{
423        struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
424
425        /* Clear out any active regions before we release the map. */
426        region_truncate(resv_map, 0);
427        kfree(resv_map);
428}
429
430static inline struct resv_map *inode_resv_map(struct inode *inode)
431{
432        return inode->i_mapping->private_data;
433}
434
435static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
436{
437        VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
438        if (vma->vm_flags & VM_MAYSHARE) {
439                struct address_space *mapping = vma->vm_file->f_mapping;
440                struct inode *inode = mapping->host;
441
442                return inode_resv_map(inode);
443
444        } else {
445                return (struct resv_map *)(get_vma_private_data(vma) &
446                                                        ~HPAGE_RESV_MASK);
447        }
448}
449
450static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
451{
452        VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
453        VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
454
455        set_vma_private_data(vma, (get_vma_private_data(vma) &
456                                HPAGE_RESV_MASK) | (unsigned long)map);
457}
458
459static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
460{
461        VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
462        VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
463
464        set_vma_private_data(vma, get_vma_private_data(vma) | flags);
465}
466
467static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
468{
469        VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
470
471        return (get_vma_private_data(vma) & flag) != 0;
472}
473
474/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
475void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
476{
477        VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
478        if (!(vma->vm_flags & VM_MAYSHARE))
479                vma->vm_private_data = (void *)0;
480}
481
482/* Returns true if the VMA has associated reserve pages */
483static int vma_has_reserves(struct vm_area_struct *vma, long chg)
484{
485        if (vma->vm_flags & VM_NORESERVE) {
486                /*
487                 * This address is already reserved by other process(chg == 0),
488                 * so, we should decrement reserved count. Without decrementing,
489                 * reserve count remains after releasing inode, because this
490                 * allocated page will go into page cache and is regarded as
491                 * coming from reserved pool in releasing step.  Currently, we
492                 * don't have any other solution to deal with this situation
493                 * properly, so add work-around here.
494                 */
495                if (vma->vm_flags & VM_MAYSHARE && chg == 0)
496                        return 1;
497                else
498                        return 0;
499        }
500
501        /* Shared mappings always use reserves */
502        if (vma->vm_flags & VM_MAYSHARE)
503                return 1;
504
505        /*
506         * Only the process that called mmap() has reserves for
507         * private mappings.
508         */
509        if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
510                return 1;
511
512        return 0;
513}
514
515static void enqueue_huge_page(struct hstate *h, struct page *page)
516{
517        int nid = page_to_nid(page);
518        list_move(&page->lru, &h->hugepage_freelists[nid]);
519        h->free_huge_pages++;
520        h->free_huge_pages_node[nid]++;
521}
522
523static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
524{
525        struct page *page;
526
527        list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
528                if (!is_migrate_isolate_page(page))
529                        break;
530        /*
531         * if 'non-isolated free hugepage' not found on the list,
532         * the allocation fails.
533         */
534        if (&h->hugepage_freelists[nid] == &page->lru)
535                return NULL;
536        list_move(&page->lru, &h->hugepage_activelist);
537        set_page_refcounted(page);
538        h->free_huge_pages--;
539        h->free_huge_pages_node[nid]--;
540        return page;
541}
542
543/* Movability of hugepages depends on migration support. */
544static inline gfp_t htlb_alloc_mask(struct hstate *h)
545{
546        if (hugepages_treat_as_movable || hugepage_migration_supported(h))
547                return GFP_HIGHUSER_MOVABLE;
548        else
549                return GFP_HIGHUSER;
550}
551
552static struct page *dequeue_huge_page_vma(struct hstate *h,
553                                struct vm_area_struct *vma,
554                                unsigned long address, int avoid_reserve,
555                                long chg)
556{
557        struct page *page = NULL;
558        struct mempolicy *mpol;
559        nodemask_t *nodemask;
560        struct zonelist *zonelist;
561        struct zone *zone;
562        struct zoneref *z;
563        unsigned int cpuset_mems_cookie;
564
565        /*
566         * A child process with MAP_PRIVATE mappings created by their parent
567         * have no page reserves. This check ensures that reservations are
568         * not "stolen". The child may still get SIGKILLed
569         */
570        if (!vma_has_reserves(vma, chg) &&
571                        h->free_huge_pages - h->resv_huge_pages == 0)
572                goto err;
573
574        /* If reserves cannot be used, ensure enough pages are in the pool */
575        if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
576                goto err;
577
578retry_cpuset:
579        cpuset_mems_cookie = read_mems_allowed_begin();
580        zonelist = huge_zonelist(vma, address,
581                                        htlb_alloc_mask(h), &mpol, &nodemask);
582
583        for_each_zone_zonelist_nodemask(zone, z, zonelist,
584                                                MAX_NR_ZONES - 1, nodemask) {
585                if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask(h))) {
586                        page = dequeue_huge_page_node(h, zone_to_nid(zone));
587                        if (page) {
588                                if (avoid_reserve)
589                                        break;
590                                if (!vma_has_reserves(vma, chg))
591                                        break;
592
593                                SetPagePrivate(page);
594                                h->resv_huge_pages--;
595                                break;
596                        }
597                }
598        }
599
600        mpol_cond_put(mpol);
601        if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
602                goto retry_cpuset;
603        return page;
604
605err:
606        return NULL;
607}
608
609/*
610 * common helper functions for hstate_next_node_to_{alloc|free}.
611 * We may have allocated or freed a huge page based on a different
612 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
613 * be outside of *nodes_allowed.  Ensure that we use an allowed
614 * node for alloc or free.
615 */
616static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
617{
618        nid = next_node(nid, *nodes_allowed);
619        if (nid == MAX_NUMNODES)
620                nid = first_node(*nodes_allowed);
621        VM_BUG_ON(nid >= MAX_NUMNODES);
622
623        return nid;
624}
625
626static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
627{
628        if (!node_isset(nid, *nodes_allowed))
629                nid = next_node_allowed(nid, nodes_allowed);
630        return nid;
631}
632
633/*
634 * returns the previously saved node ["this node"] from which to
635 * allocate a persistent huge page for the pool and advance the
636 * next node from which to allocate, handling wrap at end of node
637 * mask.
638 */
639static int hstate_next_node_to_alloc(struct hstate *h,
640                                        nodemask_t *nodes_allowed)
641{
642        int nid;
643
644        VM_BUG_ON(!nodes_allowed);
645
646        nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
647        h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
648
649        return nid;
650}
651
652/*
653 * helper for free_pool_huge_page() - return the previously saved
654 * node ["this node"] from which to free a huge page.  Advance the
655 * next node id whether or not we find a free huge page to free so
656 * that the next attempt to free addresses the next node.
657 */
658static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
659{
660        int nid;
661
662        VM_BUG_ON(!nodes_allowed);
663
664        nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
665        h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
666
667        return nid;
668}
669
670#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)           \
671        for (nr_nodes = nodes_weight(*mask);                            \
672                nr_nodes > 0 &&                                         \
673                ((node = hstate_next_node_to_alloc(hs, mask)) || 1);    \
674                nr_nodes--)
675
676#define for_each_node_mask_to_free(hs, nr_nodes, node, mask)            \
677        for (nr_nodes = nodes_weight(*mask);                            \
678                nr_nodes > 0 &&                                         \
679                ((node = hstate_next_node_to_free(hs, mask)) || 1);     \
680                nr_nodes--)
681
682#if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
683static void destroy_compound_gigantic_page(struct page *page,
684                                        unsigned int order)
685{
686        int i;
687        int nr_pages = 1 << order;
688        struct page *p = page + 1;
689
690        for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
691                __ClearPageTail(p);
692                set_page_refcounted(p);
693                p->first_page = NULL;
694        }
695
696        set_compound_order(page, 0);
697        __ClearPageHead(page);
698}
699
700static void free_gigantic_page(struct page *page, unsigned int order)
701{
702        free_contig_range(page_to_pfn(page), 1 << order);
703}
704
705static int __alloc_gigantic_page(unsigned long start_pfn,
706                                unsigned long nr_pages)
707{
708        unsigned long end_pfn = start_pfn + nr_pages;
709        return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
710}
711
712static bool pfn_range_valid_gigantic(unsigned long start_pfn,
713                                unsigned long nr_pages)
714{
715        unsigned long i, end_pfn = start_pfn + nr_pages;
716        struct page *page;
717
718        for (i = start_pfn; i < end_pfn; i++) {
719                if (!pfn_valid(i))
720                        return false;
721
722                page = pfn_to_page(i);
723
724                if (PageReserved(page))
725                        return false;
726
727                if (page_count(page) > 0)
728                        return false;
729
730                if (PageHuge(page))
731                        return false;
732        }
733
734        return true;
735}
736
737static bool zone_spans_last_pfn(const struct zone *zone,
738                        unsigned long start_pfn, unsigned long nr_pages)
739{
740        unsigned long last_pfn = start_pfn + nr_pages - 1;
741        return zone_spans_pfn(zone, last_pfn);
742}
743
744static struct page *alloc_gigantic_page(int nid, unsigned int order)
745{
746        unsigned long nr_pages = 1 << order;
747        unsigned long ret, pfn, flags;
748        struct zone *z;
749
750        z = NODE_DATA(nid)->node_zones;
751        for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
752                spin_lock_irqsave(&z->lock, flags);
753
754                pfn = ALIGN(z->zone_start_pfn, nr_pages);
755                while (zone_spans_last_pfn(z, pfn, nr_pages)) {
756                        if (pfn_range_valid_gigantic(pfn, nr_pages)) {
757                                /*
758                                 * We release the zone lock here because
759                                 * alloc_contig_range() will also lock the zone
760                                 * at some point. If there's an allocation
761                                 * spinning on this lock, it may win the race
762                                 * and cause alloc_contig_range() to fail...
763                                 */
764                                spin_unlock_irqrestore(&z->lock, flags);
765                                ret = __alloc_gigantic_page(pfn, nr_pages);
766                                if (!ret)
767                                        return pfn_to_page(pfn);
768                                spin_lock_irqsave(&z->lock, flags);
769                        }
770                        pfn += nr_pages;
771                }
772
773                spin_unlock_irqrestore(&z->lock, flags);
774        }
775
776        return NULL;
777}
778
779static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
780static void prep_compound_gigantic_page(struct page *page, unsigned int order);
781
782static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
783{
784        struct page *page;
785
786        page = alloc_gigantic_page(nid, huge_page_order(h));
787        if (page) {
788                prep_compound_gigantic_page(page, huge_page_order(h));
789                prep_new_huge_page(h, page, nid);
790        }
791
792        return page;
793}
794
795static int alloc_fresh_gigantic_page(struct hstate *h,
796                                nodemask_t *nodes_allowed)
797{
798        struct page *page = NULL;
799        int nr_nodes, node;
800
801        for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
802                page = alloc_fresh_gigantic_page_node(h, node);
803                if (page)
804                        return 1;
805        }
806
807        return 0;
808}
809
810static inline bool gigantic_page_supported(void) { return true; }
811#else
812static inline bool gigantic_page_supported(void) { return false; }
813static inline void free_gigantic_page(struct page *page, unsigned int order) { }
814static inline void destroy_compound_gigantic_page(struct page *page,
815                                                unsigned int order) { }
816static inline int alloc_fresh_gigantic_page(struct hstate *h,
817                                        nodemask_t *nodes_allowed) { return 0; }
818#endif
819
820static void update_and_free_page(struct hstate *h, struct page *page)
821{
822        int i;
823
824        if (hstate_is_gigantic(h) && !gigantic_page_supported())
825                return;
826
827        h->nr_huge_pages--;
828        h->nr_huge_pages_node[page_to_nid(page)]--;
829        for (i = 0; i < pages_per_huge_page(h); i++) {
830                page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
831                                1 << PG_referenced | 1 << PG_dirty |
832                                1 << PG_active | 1 << PG_private |
833                                1 << PG_writeback);
834        }
835        VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
836        set_compound_page_dtor(page, NULL);
837        set_page_refcounted(page);
838        if (hstate_is_gigantic(h)) {
839                destroy_compound_gigantic_page(page, huge_page_order(h));
840                free_gigantic_page(page, huge_page_order(h));
841        } else {
842                arch_release_hugepage(page);
843                __free_pages(page, huge_page_order(h));
844        }
845}
846
847struct hstate *size_to_hstate(unsigned long size)
848{
849        struct hstate *h;
850
851        for_each_hstate(h) {
852                if (huge_page_size(h) == size)
853                        return h;
854        }
855        return NULL;
856}
857
858/*
859 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
860 * to hstate->hugepage_activelist.)
861 *
862 * This function can be called for tail pages, but never returns true for them.
863 */
864bool page_huge_active(struct page *page)
865{
866        VM_BUG_ON_PAGE(!PageHuge(page), page);
867        return PageHead(page) && PagePrivate(&page[1]);
868}
869
870/* never called for tail page */
871static void set_page_huge_active(struct page *page)
872{
873        VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
874        SetPagePrivate(&page[1]);
875}
876
877static void clear_page_huge_active(struct page *page)
878{
879        VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
880        ClearPagePrivate(&page[1]);
881}
882
883void free_huge_page(struct page *page)
884{
885        /*
886         * Can't pass hstate in here because it is called from the
887         * compound page destructor.
888         */
889        struct hstate *h = page_hstate(page);
890        int nid = page_to_nid(page);
891        struct hugepage_subpool *spool =
892                (struct hugepage_subpool *)page_private(page);
893        bool restore_reserve;
894
895        set_page_private(page, 0);
896        page->mapping = NULL;
897        BUG_ON(page_count(page));
898        BUG_ON(page_mapcount(page));
899        restore_reserve = PagePrivate(page);
900        ClearPagePrivate(page);
901
902        spin_lock(&hugetlb_lock);
903        clear_page_huge_active(page);
904        hugetlb_cgroup_uncharge_page(hstate_index(h),
905                                     pages_per_huge_page(h), page);
906        if (restore_reserve)
907                h->resv_huge_pages++;
908
909        if (h->surplus_huge_pages_node[nid]) {
910                /* remove the page from active list */
911                list_del(&page->lru);
912                update_and_free_page(h, page);
913                h->surplus_huge_pages--;
914                h->surplus_huge_pages_node[nid]--;
915        } else {
916                arch_clear_hugepage_flags(page);
917                enqueue_huge_page(h, page);
918        }
919        spin_unlock(&hugetlb_lock);
920        hugepage_subpool_put_pages(spool, 1);
921}
922
923static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
924{
925        INIT_LIST_HEAD(&page->lru);
926        set_compound_page_dtor(page, free_huge_page);
927        spin_lock(&hugetlb_lock);
928        set_hugetlb_cgroup(page, NULL);
929        h->nr_huge_pages++;
930        h->nr_huge_pages_node[nid]++;
931        spin_unlock(&hugetlb_lock);
932        put_page(page); /* free it into the hugepage allocator */
933}
934
935static void prep_compound_gigantic_page(struct page *page, unsigned int order)
936{
937        int i;
938        int nr_pages = 1 << order;
939        struct page *p = page + 1;
940
941        /* we rely on prep_new_huge_page to set the destructor */
942        set_compound_order(page, order);
943        __SetPageHead(page);
944        __ClearPageReserved(page);
945        for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
946                __SetPageTail(p);
947                /*
948                 * For gigantic hugepages allocated through bootmem at
949                 * boot, it's safer to be consistent with the not-gigantic
950                 * hugepages and clear the PG_reserved bit from all tail pages
951                 * too.  Otherwse drivers using get_user_pages() to access tail
952                 * pages may get the reference counting wrong if they see
953                 * PG_reserved set on a tail page (despite the head page not
954                 * having PG_reserved set).  Enforcing this consistency between
955                 * head and tail pages allows drivers to optimize away a check
956                 * on the head page when they need know if put_page() is needed
957                 * after get_user_pages().
958                 */
959                __ClearPageReserved(p);
960                set_page_count(p, 0);
961                p->first_page = page;
962        }
963}
964
965/*
966 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
967 * transparent huge pages.  See the PageTransHuge() documentation for more
968 * details.
969 */
970int PageHuge(struct page *page)
971{
972        if (!PageCompound(page))
973                return 0;
974
975        page = compound_head(page);
976        return get_compound_page_dtor(page) == free_huge_page;
977}
978EXPORT_SYMBOL_GPL(PageHuge);
979
980/*
981 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
982 * normal or transparent huge pages.
983 */
984int PageHeadHuge(struct page *page_head)
985{
986        if (!PageHead(page_head))
987                return 0;
988
989        return get_compound_page_dtor(page_head) == free_huge_page;
990}
991
992pgoff_t __basepage_index(struct page *page)
993{
994        struct page *page_head = compound_head(page);
995        pgoff_t index = page_index(page_head);
996        unsigned long compound_idx;
997
998        if (!PageHuge(page_head))
999                return page_index(page);
1000
1001        if (compound_order(page_head) >= MAX_ORDER)
1002                compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1003        else
1004                compound_idx = page - page_head;
1005
1006        return (index << compound_order(page_head)) + compound_idx;
1007}
1008
1009static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1010{
1011        struct page *page;
1012
1013        page = alloc_pages_exact_node(nid,
1014                htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1015                                                __GFP_REPEAT|__GFP_NOWARN,
1016                huge_page_order(h));
1017        if (page) {
1018                if (arch_prepare_hugepage(page)) {
1019                        __free_pages(page, huge_page_order(h));
1020                        return NULL;
1021                }
1022                prep_new_huge_page(h, page, nid);
1023        }
1024
1025        return page;
1026}
1027
1028static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1029{
1030        struct page *page;
1031        int nr_nodes, node;
1032        int ret = 0;
1033
1034        for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1035                page = alloc_fresh_huge_page_node(h, node);
1036                if (page) {
1037                        ret = 1;
1038                        break;
1039                }
1040        }
1041
1042        if (ret)
1043                count_vm_event(HTLB_BUDDY_PGALLOC);
1044        else
1045                count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1046
1047        return ret;
1048}
1049
1050/*
1051 * Free huge page from pool from next node to free.
1052 * Attempt to keep persistent huge pages more or less
1053 * balanced over allowed nodes.
1054 * Called with hugetlb_lock locked.
1055 */
1056static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1057                                                         bool acct_surplus)
1058{
1059        int nr_nodes, node;
1060        int ret = 0;
1061
1062        for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1063                /*
1064                 * If we're returning unused surplus pages, only examine
1065                 * nodes with surplus pages.
1066                 */
1067                if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1068                    !list_empty(&h->hugepage_freelists[node])) {
1069                        struct page *page =
1070                                list_entry(h->hugepage_freelists[node].next,
1071                                          struct page, lru);
1072                        list_del(&page->lru);
1073                        h->free_huge_pages--;
1074                        h->free_huge_pages_node[node]--;
1075                        if (acct_surplus) {
1076                                h->surplus_huge_pages--;
1077                                h->surplus_huge_pages_node[node]--;
1078                        }
1079                        update_and_free_page(h, page);
1080                        ret = 1;
1081                        break;
1082                }
1083        }
1084
1085        return ret;
1086}
1087
1088/*
1089 * Dissolve a given free hugepage into free buddy pages. This function does
1090 * nothing for in-use (including surplus) hugepages.
1091 */
1092static void dissolve_free_huge_page(struct page *page)
1093{
1094        spin_lock(&hugetlb_lock);
1095        if (PageHuge(page) && !page_count(page)) {
1096                struct hstate *h = page_hstate(page);
1097                int nid = page_to_nid(page);
1098                list_del(&page->lru);
1099                h->free_huge_pages--;
1100                h->free_huge_pages_node[nid]--;
1101                update_and_free_page(h, page);
1102        }
1103        spin_unlock(&hugetlb_lock);
1104}
1105
1106/*
1107 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1108 * make specified memory blocks removable from the system.
1109 * Note that start_pfn should aligned with (minimum) hugepage size.
1110 */
1111void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1112{
1113        unsigned int order = 8 * sizeof(void *);
1114        unsigned long pfn;
1115        struct hstate *h;
1116
1117        if (!hugepages_supported())
1118                return;
1119
1120        /* Set scan step to minimum hugepage size */
1121        for_each_hstate(h)
1122                if (order > huge_page_order(h))
1123                        order = huge_page_order(h);
1124        VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order));
1125        for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order)
1126                dissolve_free_huge_page(pfn_to_page(pfn));
1127}
1128
1129static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
1130{
1131        struct page *page;
1132        unsigned int r_nid;
1133
1134        if (hstate_is_gigantic(h))
1135                return NULL;
1136
1137        /*
1138         * Assume we will successfully allocate the surplus page to
1139         * prevent racing processes from causing the surplus to exceed
1140         * overcommit
1141         *
1142         * This however introduces a different race, where a process B
1143         * tries to grow the static hugepage pool while alloc_pages() is
1144         * called by process A. B will only examine the per-node
1145         * counters in determining if surplus huge pages can be
1146         * converted to normal huge pages in adjust_pool_surplus(). A
1147         * won't be able to increment the per-node counter, until the
1148         * lock is dropped by B, but B doesn't drop hugetlb_lock until
1149         * no more huge pages can be converted from surplus to normal
1150         * state (and doesn't try to convert again). Thus, we have a
1151         * case where a surplus huge page exists, the pool is grown, and
1152         * the surplus huge page still exists after, even though it
1153         * should just have been converted to a normal huge page. This
1154         * does not leak memory, though, as the hugepage will be freed
1155         * once it is out of use. It also does not allow the counters to
1156         * go out of whack in adjust_pool_surplus() as we don't modify
1157         * the node values until we've gotten the hugepage and only the
1158         * per-node value is checked there.
1159         */
1160        spin_lock(&hugetlb_lock);
1161        if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1162                spin_unlock(&hugetlb_lock);
1163                return NULL;
1164        } else {
1165                h->nr_huge_pages++;
1166                h->surplus_huge_pages++;
1167        }
1168        spin_unlock(&hugetlb_lock);
1169
1170        if (nid == NUMA_NO_NODE)
1171                page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
1172                                   __GFP_REPEAT|__GFP_NOWARN,
1173                                   huge_page_order(h));
1174        else
1175                page = alloc_pages_exact_node(nid,
1176                        htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1177                        __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
1178
1179        if (page && arch_prepare_hugepage(page)) {
1180                __free_pages(page, huge_page_order(h));
1181                page = NULL;
1182        }
1183
1184        spin_lock(&hugetlb_lock);
1185        if (page) {
1186                INIT_LIST_HEAD(&page->lru);
1187                r_nid = page_to_nid(page);
1188                set_compound_page_dtor(page, free_huge_page);
1189                set_hugetlb_cgroup(page, NULL);
1190                /*
1191                 * We incremented the global counters already
1192                 */
1193                h->nr_huge_pages_node[r_nid]++;
1194                h->surplus_huge_pages_node[r_nid]++;
1195                __count_vm_event(HTLB_BUDDY_PGALLOC);
1196        } else {
1197                h->nr_huge_pages--;
1198                h->surplus_huge_pages--;
1199                __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1200        }
1201        spin_unlock(&hugetlb_lock);
1202
1203        return page;
1204}
1205
1206/*
1207 * This allocation function is useful in the context where vma is irrelevant.
1208 * E.g. soft-offlining uses this function because it only cares physical
1209 * address of error page.
1210 */
1211struct page *alloc_huge_page_node(struct hstate *h, int nid)
1212{
1213        struct page *page = NULL;
1214
1215        spin_lock(&hugetlb_lock);
1216        if (h->free_huge_pages - h->resv_huge_pages > 0)
1217                page = dequeue_huge_page_node(h, nid);
1218        spin_unlock(&hugetlb_lock);
1219
1220        if (!page)
1221                page = alloc_buddy_huge_page(h, nid);
1222
1223        return page;
1224}
1225
1226/*
1227 * Increase the hugetlb pool such that it can accommodate a reservation
1228 * of size 'delta'.
1229 */
1230static int gather_surplus_pages(struct hstate *h, int delta)
1231{
1232        struct list_head surplus_list;
1233        struct page *page, *tmp;
1234        int ret, i;
1235        int needed, allocated;
1236        bool alloc_ok = true;
1237
1238        needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1239        if (needed <= 0) {
1240                h->resv_huge_pages += delta;
1241                return 0;
1242        }
1243
1244        allocated = 0;
1245        INIT_LIST_HEAD(&surplus_list);
1246
1247        ret = -ENOMEM;
1248retry:
1249        spin_unlock(&hugetlb_lock);
1250        for (i = 0; i < needed; i++) {
1251                page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1252                if (!page) {
1253                        alloc_ok = false;
1254                        break;
1255                }
1256                list_add(&page->lru, &surplus_list);
1257        }
1258        allocated += i;
1259
1260        /*
1261         * After retaking hugetlb_lock, we need to recalculate 'needed'
1262         * because either resv_huge_pages or free_huge_pages may have changed.
1263         */
1264        spin_lock(&hugetlb_lock);
1265        needed = (h->resv_huge_pages + delta) -
1266                        (h->free_huge_pages + allocated);
1267        if (needed > 0) {
1268                if (alloc_ok)
1269                        goto retry;
1270                /*
1271                 * We were not able to allocate enough pages to
1272                 * satisfy the entire reservation so we free what
1273                 * we've allocated so far.
1274                 */
1275                goto free;
1276        }
1277        /*
1278         * The surplus_list now contains _at_least_ the number of extra pages
1279         * needed to accommodate the reservation.  Add the appropriate number
1280         * of pages to the hugetlb pool and free the extras back to the buddy
1281         * allocator.  Commit the entire reservation here to prevent another
1282         * process from stealing the pages as they are added to the pool but
1283         * before they are reserved.
1284         */
1285        needed += allocated;
1286        h->resv_huge_pages += delta;
1287        ret = 0;
1288
1289        /* Free the needed pages to the hugetlb pool */
1290        list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1291                if ((--needed) < 0)
1292                        break;
1293                /*
1294                 * This page is now managed by the hugetlb allocator and has
1295                 * no users -- drop the buddy allocator's reference.
1296                 */
1297                put_page_testzero(page);
1298                VM_BUG_ON_PAGE(page_count(page), page);
1299                enqueue_huge_page(h, page);
1300        }
1301free:
1302        spin_unlock(&hugetlb_lock);
1303
1304        /* Free unnecessary surplus pages to the buddy allocator */
1305        list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1306                put_page(page);
1307        spin_lock(&hugetlb_lock);
1308
1309        return ret;
1310}
1311
1312/*
1313 * When releasing a hugetlb pool reservation, any surplus pages that were
1314 * allocated to satisfy the reservation must be explicitly freed if they were
1315 * never used.
1316 * Called with hugetlb_lock held.
1317 */
1318static void return_unused_surplus_pages(struct hstate *h,
1319                                        unsigned long unused_resv_pages)
1320{
1321        unsigned long nr_pages;
1322
1323        /* Uncommit the reservation */
1324        h->resv_huge_pages -= unused_resv_pages;
1325
1326        /* Cannot return gigantic pages currently */
1327        if (hstate_is_gigantic(h))
1328                return;
1329
1330        nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1331
1332        /*
1333         * We want to release as many surplus pages as possible, spread
1334         * evenly across all nodes with memory. Iterate across these nodes
1335         * until we can no longer free unreserved surplus pages. This occurs
1336         * when the nodes with surplus pages have no free pages.
1337         * free_pool_huge_page() will balance the the freed pages across the
1338         * on-line nodes with memory and will handle the hstate accounting.
1339         */
1340        while (nr_pages--) {
1341                if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1342                        break;
1343                cond_resched_lock(&hugetlb_lock);
1344        }
1345}
1346
1347/*
1348 * Determine if the huge page at addr within the vma has an associated
1349 * reservation.  Where it does not we will need to logically increase
1350 * reservation and actually increase subpool usage before an allocation
1351 * can occur.  Where any new reservation would be required the
1352 * reservation change is prepared, but not committed.  Once the page
1353 * has been allocated from the subpool and instantiated the change should
1354 * be committed via vma_commit_reservation.  No action is required on
1355 * failure.
1356 */
1357static long vma_needs_reservation(struct hstate *h,
1358                        struct vm_area_struct *vma, unsigned long addr)
1359{
1360        struct resv_map *resv;
1361        pgoff_t idx;
1362        long chg;
1363
1364        resv = vma_resv_map(vma);
1365        if (!resv)
1366                return 1;
1367
1368        idx = vma_hugecache_offset(h, vma, addr);
1369        chg = region_chg(resv, idx, idx + 1);
1370
1371        if (vma->vm_flags & VM_MAYSHARE)
1372                return chg;
1373        else
1374                return chg < 0 ? chg : 0;
1375}
1376static void vma_commit_reservation(struct hstate *h,
1377                        struct vm_area_struct *vma, unsigned long addr)
1378{
1379        struct resv_map *resv;
1380        pgoff_t idx;
1381
1382        resv = vma_resv_map(vma);
1383        if (!resv)
1384                return;
1385
1386        idx = vma_hugecache_offset(h, vma, addr);
1387        region_add(resv, idx, idx + 1);
1388}
1389
1390static struct page *alloc_huge_page(struct vm_area_struct *vma,
1391                                    unsigned long addr, int avoid_reserve)
1392{
1393        struct hugepage_subpool *spool = subpool_vma(vma);
1394        struct hstate *h = hstate_vma(vma);
1395        struct page *page;
1396        long chg;
1397        int ret, idx;
1398        struct hugetlb_cgroup *h_cg;
1399
1400        idx = hstate_index(h);
1401        /*
1402         * Processes that did not create the mapping will have no
1403         * reserves and will not have accounted against subpool
1404         * limit. Check that the subpool limit can be made before
1405         * satisfying the allocation MAP_NORESERVE mappings may also
1406         * need pages and subpool limit allocated allocated if no reserve
1407         * mapping overlaps.
1408         */
1409        chg = vma_needs_reservation(h, vma, addr);
1410        if (chg < 0)
1411                return ERR_PTR(-ENOMEM);
1412        if (chg || avoid_reserve)
1413                if (hugepage_subpool_get_pages(spool, 1))
1414                        return ERR_PTR(-ENOSPC);
1415
1416        ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1417        if (ret)
1418                goto out_subpool_put;
1419
1420        spin_lock(&hugetlb_lock);
1421        page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
1422        if (!page) {
1423                spin_unlock(&hugetlb_lock);
1424                page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1425                if (!page)
1426                        goto out_uncharge_cgroup;
1427
1428                spin_lock(&hugetlb_lock);
1429                list_move(&page->lru, &h->hugepage_activelist);
1430                /* Fall through */
1431        }
1432        hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1433        spin_unlock(&hugetlb_lock);
1434
1435        set_page_private(page, (unsigned long)spool);
1436
1437        vma_commit_reservation(h, vma, addr);
1438        return page;
1439
1440out_uncharge_cgroup:
1441        hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1442out_subpool_put:
1443        if (chg || avoid_reserve)
1444                hugepage_subpool_put_pages(spool, 1);
1445        return ERR_PTR(-ENOSPC);
1446}
1447
1448/*
1449 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1450 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1451 * where no ERR_VALUE is expected to be returned.
1452 */
1453struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1454                                unsigned long addr, int avoid_reserve)
1455{
1456        struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1457        if (IS_ERR(page))
1458                page = NULL;
1459        return page;
1460}
1461
1462int __weak alloc_bootmem_huge_page(struct hstate *h)
1463{
1464        struct huge_bootmem_page *m;
1465        int nr_nodes, node;
1466
1467        for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1468                void *addr;
1469
1470                addr = memblock_virt_alloc_try_nid_nopanic(
1471                                huge_page_size(h), huge_page_size(h),
1472                                0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1473                if (addr) {
1474                        /*
1475                         * Use the beginning of the huge page to store the
1476                         * huge_bootmem_page struct (until gather_bootmem
1477                         * puts them into the mem_map).
1478                         */
1479                        m = addr;
1480                        goto found;
1481                }
1482        }
1483        return 0;
1484
1485found:
1486        BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1487        /* Put them into a private list first because mem_map is not up yet */
1488        list_add(&m->list, &huge_boot_pages);
1489        m->hstate = h;
1490        return 1;
1491}
1492
1493static void __init prep_compound_huge_page(struct page *page,
1494                unsigned int order)
1495{
1496        if (unlikely(order > (MAX_ORDER - 1)))
1497                prep_compound_gigantic_page(page, order);
1498        else
1499                prep_compound_page(page, order);
1500}
1501
1502/* Put bootmem huge pages into the standard lists after mem_map is up */
1503static void __init gather_bootmem_prealloc(void)
1504{
1505        struct huge_bootmem_page *m;
1506
1507        list_for_each_entry(m, &huge_boot_pages, list) {
1508                struct hstate *h = m->hstate;
1509                struct page *page;
1510
1511#ifdef CONFIG_HIGHMEM
1512                page = pfn_to_page(m->phys >> PAGE_SHIFT);
1513                memblock_free_late(__pa(m),
1514                                   sizeof(struct huge_bootmem_page));
1515#else
1516                page = virt_to_page(m);
1517#endif
1518                WARN_ON(page_count(page) != 1);
1519                prep_compound_huge_page(page, h->order);
1520                WARN_ON(PageReserved(page));
1521                prep_new_huge_page(h, page, page_to_nid(page));
1522                /*
1523                 * If we had gigantic hugepages allocated at boot time, we need
1524                 * to restore the 'stolen' pages to totalram_pages in order to
1525                 * fix confusing memory reports from free(1) and another
1526                 * side-effects, like CommitLimit going negative.
1527                 */
1528                if (hstate_is_gigantic(h))
1529                        adjust_managed_page_count(page, 1 << h->order);
1530        }
1531}
1532
1533static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1534{
1535        unsigned long i;
1536
1537        for (i = 0; i < h->max_huge_pages; ++i) {
1538                if (hstate_is_gigantic(h)) {
1539                        if (!alloc_bootmem_huge_page(h))
1540                                break;
1541                } else if (!alloc_fresh_huge_page(h,
1542                                         &node_states[N_MEMORY]))
1543                        break;
1544        }
1545        h->max_huge_pages = i;
1546}
1547
1548static void __init hugetlb_init_hstates(void)
1549{
1550        struct hstate *h;
1551
1552        for_each_hstate(h) {
1553                /* oversize hugepages were init'ed in early boot */
1554                if (!hstate_is_gigantic(h))
1555                        hugetlb_hstate_alloc_pages(h);
1556        }
1557}
1558
1559static char * __init memfmt(char *buf, unsigned long n)
1560{
1561        if (n >= (1UL << 30))
1562                sprintf(buf, "%lu GB", n >> 30);
1563        else if (n >= (1UL << 20))
1564                sprintf(buf, "%lu MB", n >> 20);
1565        else
1566                sprintf(buf, "%lu KB", n >> 10);
1567        return buf;
1568}
1569
1570static void __init report_hugepages(void)
1571{
1572        struct hstate *h;
1573
1574        for_each_hstate(h) {
1575                char buf[32];
1576                pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1577                        memfmt(buf, huge_page_size(h)),
1578                        h->free_huge_pages);
1579        }
1580}
1581
1582#ifdef CONFIG_HIGHMEM
1583static void try_to_free_low(struct hstate *h, unsigned long count,
1584                                                nodemask_t *nodes_allowed)
1585{
1586        int i;
1587
1588        if (hstate_is_gigantic(h))
1589                return;
1590
1591        for_each_node_mask(i, *nodes_allowed) {
1592                struct page *page, *next;
1593                struct list_head *freel = &h->hugepage_freelists[i];
1594                list_for_each_entry_safe(page, next, freel, lru) {
1595                        if (count >= h->nr_huge_pages)
1596                                return;
1597                        if (PageHighMem(page))
1598                                continue;
1599                        list_del(&page->lru);
1600                        update_and_free_page(h, page);
1601                        h->free_huge_pages--;
1602                        h->free_huge_pages_node[page_to_nid(page)]--;
1603                }
1604        }
1605}
1606#else
1607static inline void try_to_free_low(struct hstate *h, unsigned long count,
1608                                                nodemask_t *nodes_allowed)
1609{
1610}
1611#endif
1612
1613/*
1614 * Increment or decrement surplus_huge_pages.  Keep node-specific counters
1615 * balanced by operating on them in a round-robin fashion.
1616 * Returns 1 if an adjustment was made.
1617 */
1618static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1619                                int delta)
1620{
1621        int nr_nodes, node;
1622
1623        VM_BUG_ON(delta != -1 && delta != 1);
1624
1625        if (delta < 0) {
1626                for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1627                        if (h->surplus_huge_pages_node[node])
1628                                goto found;
1629                }
1630        } else {
1631                for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1632                        if (h->surplus_huge_pages_node[node] <
1633                                        h->nr_huge_pages_node[node])
1634                                goto found;
1635                }
1636        }
1637        return 0;
1638
1639found:
1640        h->surplus_huge_pages += delta;
1641        h->surplus_huge_pages_node[node] += delta;
1642        return 1;
1643}
1644
1645#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1646static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1647                                                nodemask_t *nodes_allowed)
1648{
1649        unsigned long min_count, ret;
1650
1651        if (hstate_is_gigantic(h) && !gigantic_page_supported())
1652                return h->max_huge_pages;
1653
1654        /*
1655         * Increase the pool size
1656         * First take pages out of surplus state.  Then make up the
1657         * remaining difference by allocating fresh huge pages.
1658         *
1659         * We might race with alloc_buddy_huge_page() here and be unable
1660         * to convert a surplus huge page to a normal huge page. That is
1661         * not critical, though, it just means the overall size of the
1662         * pool might be one hugepage larger than it needs to be, but
1663         * within all the constraints specified by the sysctls.
1664         */
1665        spin_lock(&hugetlb_lock);
1666        while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1667                if (!adjust_pool_surplus(h, nodes_allowed, -1))
1668                        break;
1669        }
1670
1671        while (count > persistent_huge_pages(h)) {
1672                /*
1673                 * If this allocation races such that we no longer need the
1674                 * page, free_huge_page will handle it by freeing the page
1675                 * and reducing the surplus.
1676                 */
1677                spin_unlock(&hugetlb_lock);
1678
1679                /* yield cpu to avoid soft lockup */
1680                cond_resched();
1681
1682                if (hstate_is_gigantic(h))
1683                        ret = alloc_fresh_gigantic_page(h, nodes_allowed);
1684                else
1685                        ret = alloc_fresh_huge_page(h, nodes_allowed);
1686                spin_lock(&hugetlb_lock);
1687                if (!ret)
1688                        goto out;
1689
1690                /* Bail for signals. Probably ctrl-c from user */
1691                if (signal_pending(current))
1692                        goto out;
1693        }
1694
1695        /*
1696         * Decrease the pool size
1697         * First return free pages to the buddy allocator (being careful
1698         * to keep enough around to satisfy reservations).  Then place
1699         * pages into surplus state as needed so the pool will shrink
1700         * to the desired size as pages become free.
1701         *
1702         * By placing pages into the surplus state independent of the
1703         * overcommit value, we are allowing the surplus pool size to
1704         * exceed overcommit. There are few sane options here. Since
1705         * alloc_buddy_huge_page() is checking the global counter,
1706         * though, we'll note that we're not allowed to exceed surplus
1707         * and won't grow the pool anywhere else. Not until one of the
1708         * sysctls are changed, or the surplus pages go out of use.
1709         */
1710        min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1711        min_count = max(count, min_count);
1712        try_to_free_low(h, min_count, nodes_allowed);
1713        while (min_count < persistent_huge_pages(h)) {
1714                if (!free_pool_huge_page(h, nodes_allowed, 0))
1715                        break;
1716                cond_resched_lock(&hugetlb_lock);
1717        }
1718        while (count < persistent_huge_pages(h)) {
1719                if (!adjust_pool_surplus(h, nodes_allowed, 1))
1720                        break;
1721        }
1722out:
1723        ret = persistent_huge_pages(h);
1724        spin_unlock(&hugetlb_lock);
1725        return ret;
1726}
1727
1728#define HSTATE_ATTR_RO(_name) \
1729        static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1730
1731#define HSTATE_ATTR(_name) \
1732        static struct kobj_attribute _name##_attr = \
1733                __ATTR(_name, 0644, _name##_show, _name##_store)
1734
1735static struct kobject *hugepages_kobj;
1736static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1737
1738static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1739
1740static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1741{
1742        int i;
1743
1744        for (i = 0; i < HUGE_MAX_HSTATE; i++)
1745                if (hstate_kobjs[i] == kobj) {
1746                        if (nidp)
1747                                *nidp = NUMA_NO_NODE;
1748                        return &hstates[i];
1749                }
1750
1751        return kobj_to_node_hstate(kobj, nidp);
1752}
1753
1754static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1755                                        struct kobj_attribute *attr, char *buf)
1756{
1757        struct hstate *h;
1758        unsigned long nr_huge_pages;
1759        int nid;
1760
1761        h = kobj_to_hstate(kobj, &nid);
1762        if (nid == NUMA_NO_NODE)
1763                nr_huge_pages = h->nr_huge_pages;
1764        else
1765                nr_huge_pages = h->nr_huge_pages_node[nid];
1766
1767        return sprintf(buf, "%lu\n", nr_huge_pages);
1768}
1769
1770static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
1771                                           struct hstate *h, int nid,
1772                                           unsigned long count, size_t len)
1773{
1774        int err;
1775        NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1776
1777        if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
1778                err = -EINVAL;
1779                goto out;
1780        }
1781
1782        if (nid == NUMA_NO_NODE) {
1783                /*
1784                 * global hstate attribute
1785                 */
1786                if (!(obey_mempolicy &&
1787                                init_nodemask_of_mempolicy(nodes_allowed))) {
1788                        NODEMASK_FREE(nodes_allowed);
1789                        nodes_allowed = &node_states[N_MEMORY];
1790                }
1791        } else if (nodes_allowed) {
1792                /*
1793                 * per node hstate attribute: adjust count to global,
1794                 * but restrict alloc/free to the specified node.
1795                 */
1796                count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1797                init_nodemask_of_node(nodes_allowed, nid);
1798        } else
1799                nodes_allowed = &node_states[N_MEMORY];
1800
1801        h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1802
1803        if (nodes_allowed != &node_states[N_MEMORY])
1804                NODEMASK_FREE(nodes_allowed);
1805
1806        return len;
1807out:
1808        NODEMASK_FREE(nodes_allowed);
1809        return err;
1810}
1811
1812static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1813                                         struct kobject *kobj, const char *buf,
1814                                         size_t len)
1815{
1816        struct hstate *h;
1817        unsigned long count;
1818        int nid;
1819        int err;
1820
1821        err = kstrtoul(buf, 10, &count);
1822        if (err)
1823                return err;
1824
1825        h = kobj_to_hstate(kobj, &nid);
1826        return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
1827}
1828
1829static ssize_t nr_hugepages_show(struct kobject *kobj,
1830                                       struct kobj_attribute *attr, char *buf)
1831{
1832        return nr_hugepages_show_common(kobj, attr, buf);
1833}
1834
1835static ssize_t nr_hugepages_store(struct kobject *kobj,
1836               struct kobj_attribute *attr, const char *buf, size_t len)
1837{
1838        return nr_hugepages_store_common(false, kobj, buf, len);
1839}
1840HSTATE_ATTR(nr_hugepages);
1841
1842#ifdef CONFIG_NUMA
1843
1844/*
1845 * hstate attribute for optionally mempolicy-based constraint on persistent
1846 * huge page alloc/free.
1847 */
1848static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1849                                       struct kobj_attribute *attr, char *buf)
1850{
1851        return nr_hugepages_show_common(kobj, attr, buf);
1852}
1853
1854static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1855               struct kobj_attribute *attr, const char *buf, size_t len)
1856{
1857        return nr_hugepages_store_common(true, kobj, buf, len);
1858}
1859HSTATE_ATTR(nr_hugepages_mempolicy);
1860#endif
1861
1862
1863static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1864                                        struct kobj_attribute *attr, char *buf)
1865{
1866        struct hstate *h = kobj_to_hstate(kobj, NULL);
1867        return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1868}
1869
1870static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1871                struct kobj_attribute *attr, const char *buf, size_t count)
1872{
1873        int err;
1874        unsigned long input;
1875        struct hstate *h = kobj_to_hstate(kobj, NULL);
1876
1877        if (hstate_is_gigantic(h))
1878                return -EINVAL;
1879
1880        err = kstrtoul(buf, 10, &input);
1881        if (err)
1882                return err;
1883
1884        spin_lock(&hugetlb_lock);
1885        h->nr_overcommit_huge_pages = input;
1886        spin_unlock(&hugetlb_lock);
1887
1888        return count;
1889}
1890HSTATE_ATTR(nr_overcommit_hugepages);
1891
1892static ssize_t free_hugepages_show(struct kobject *kobj,
1893                                        struct kobj_attribute *attr, char *buf)
1894{
1895        struct hstate *h;
1896        unsigned long free_huge_pages;
1897        int nid;
1898
1899        h = kobj_to_hstate(kobj, &nid);
1900        if (nid == NUMA_NO_NODE)
1901                free_huge_pages = h->free_huge_pages;
1902        else
1903                free_huge_pages = h->free_huge_pages_node[nid];
1904
1905        return sprintf(buf, "%lu\n", free_huge_pages);
1906}
1907HSTATE_ATTR_RO(free_hugepages);
1908
1909static ssize_t resv_hugepages_show(struct kobject *kobj,
1910                                        struct kobj_attribute *attr, char *buf)
1911{
1912        struct hstate *h = kobj_to_hstate(kobj, NULL);
1913        return sprintf(buf, "%lu\n", h->resv_huge_pages);
1914}
1915HSTATE_ATTR_RO(resv_hugepages);
1916
1917static ssize_t surplus_hugepages_show(struct kobject *kobj,
1918                                        struct kobj_attribute *attr, char *buf)
1919{
1920        struct hstate *h;
1921        unsigned long surplus_huge_pages;
1922        int nid;
1923
1924        h = kobj_to_hstate(kobj, &nid);
1925        if (nid == NUMA_NO_NODE)
1926                surplus_huge_pages = h->surplus_huge_pages;
1927        else
1928                surplus_huge_pages = h->surplus_huge_pages_node[nid];
1929
1930        return sprintf(buf, "%lu\n", surplus_huge_pages);
1931}
1932HSTATE_ATTR_RO(surplus_hugepages);
1933
1934static struct attribute *hstate_attrs[] = {
1935        &nr_hugepages_attr.attr,
1936        &nr_overcommit_hugepages_attr.attr,
1937        &free_hugepages_attr.attr,
1938        &resv_hugepages_attr.attr,
1939        &surplus_hugepages_attr.attr,
1940#ifdef CONFIG_NUMA
1941        &nr_hugepages_mempolicy_attr.attr,
1942#endif
1943        NULL,
1944};
1945
1946static struct attribute_group hstate_attr_group = {
1947        .attrs = hstate_attrs,
1948};
1949
1950static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1951                                    struct kobject **hstate_kobjs,
1952                                    struct attribute_group *hstate_attr_group)
1953{
1954        int retval;
1955        int hi = hstate_index(h);
1956
1957        hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1958        if (!hstate_kobjs[hi])
1959                return -ENOMEM;
1960
1961        retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1962        if (retval)
1963                kobject_put(hstate_kobjs[hi]);
1964
1965        return retval;
1966}
1967
1968static void __init hugetlb_sysfs_init(void)
1969{
1970        struct hstate *h;
1971        int err;
1972
1973        hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1974        if (!hugepages_kobj)
1975                return;
1976
1977        for_each_hstate(h) {
1978                err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1979                                         hstate_kobjs, &hstate_attr_group);
1980                if (err)
1981                        pr_err("Hugetlb: Unable to add hstate %s", h->name);
1982        }
1983}
1984
1985#ifdef CONFIG_NUMA
1986
1987/*
1988 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1989 * with node devices in node_devices[] using a parallel array.  The array
1990 * index of a node device or _hstate == node id.
1991 * This is here to avoid any static dependency of the node device driver, in
1992 * the base kernel, on the hugetlb module.
1993 */
1994struct node_hstate {
1995        struct kobject          *hugepages_kobj;
1996        struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
1997};
1998struct node_hstate node_hstates[MAX_NUMNODES];
1999
2000/*
2001 * A subset of global hstate attributes for node devices
2002 */
2003static struct attribute *per_node_hstate_attrs[] = {
2004        &nr_hugepages_attr.attr,
2005        &free_hugepages_attr.attr,
2006        &surplus_hugepages_attr.attr,
2007        NULL,
2008};
2009
2010static struct attribute_group per_node_hstate_attr_group = {
2011        .attrs = per_node_hstate_attrs,
2012};
2013
2014/*
2015 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2016 * Returns node id via non-NULL nidp.
2017 */
2018static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2019{
2020        int nid;
2021
2022        for (nid = 0; nid < nr_node_ids; nid++) {
2023                struct node_hstate *nhs = &node_hstates[nid];
2024                int i;
2025                for (i = 0; i < HUGE_MAX_HSTATE; i++)
2026                        if (nhs->hstate_kobjs[i] == kobj) {
2027                                if (nidp)
2028                                        *nidp = nid;
2029                                return &hstates[i];
2030                        }
2031        }
2032
2033        BUG();
2034        return NULL;
2035}
2036
2037/*
2038 * Unregister hstate attributes from a single node device.
2039 * No-op if no hstate attributes attached.
2040 */
2041static void hugetlb_unregister_node(struct node *node)
2042{
2043        struct hstate *h;
2044        struct node_hstate *nhs = &node_hstates[node->dev.id];
2045
2046        if (!nhs->hugepages_kobj)
2047                return;         /* no hstate attributes */
2048
2049        for_each_hstate(h) {
2050                int idx = hstate_index(h);
2051                if (nhs->hstate_kobjs[idx]) {
2052                        kobject_put(nhs->hstate_kobjs[idx]);
2053                        nhs->hstate_kobjs[idx] = NULL;
2054                }
2055        }
2056
2057        kobject_put(nhs->hugepages_kobj);
2058        nhs->hugepages_kobj = NULL;
2059}
2060
2061/*
2062 * hugetlb module exit:  unregister hstate attributes from node devices
2063 * that have them.
2064 */
2065static void hugetlb_unregister_all_nodes(void)
2066{
2067        int nid;
2068
2069        /*
2070         * disable node device registrations.
2071         */
2072        register_hugetlbfs_with_node(NULL, NULL);
2073
2074        /*
2075         * remove hstate attributes from any nodes that have them.
2076         */
2077        for (nid = 0; nid < nr_node_ids; nid++)
2078                hugetlb_unregister_node(node_devices[nid]);
2079}
2080
2081/*
2082 * Register hstate attributes for a single node device.
2083 * No-op if attributes already registered.
2084 */
2085static void hugetlb_register_node(struct node *node)
2086{
2087        struct hstate *h;
2088        struct node_hstate *nhs = &node_hstates[node->dev.id];
2089        int err;
2090
2091        if (nhs->hugepages_kobj)
2092                return;         /* already allocated */
2093
2094        nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2095                                                        &node->dev.kobj);
2096        if (!nhs->hugepages_kobj)
2097                return;
2098
2099        for_each_hstate(h) {
2100                err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2101                                                nhs->hstate_kobjs,
2102                                                &per_node_hstate_attr_group);
2103                if (err) {
2104                        pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2105                                h->name, node->dev.id);
2106                        hugetlb_unregister_node(node);
2107                        break;
2108                }
2109        }
2110}
2111
2112/*
2113 * hugetlb init time:  register hstate attributes for all registered node
2114 * devices of nodes that have memory.  All on-line nodes should have
2115 * registered their associated device by this time.
2116 */
2117static void hugetlb_register_all_nodes(void)
2118{
2119        int nid;
2120
2121        for_each_node_state(nid, N_MEMORY) {
2122                struct node *node = node_devices[nid];
2123                if (node->dev.id == nid)
2124                        hugetlb_register_node(node);
2125        }
2126
2127        /*
2128         * Let the node device driver know we're here so it can
2129         * [un]register hstate attributes on node hotplug.
2130         */
2131        register_hugetlbfs_with_node(hugetlb_register_node,
2132                                     hugetlb_unregister_node);
2133}
2134#else   /* !CONFIG_NUMA */
2135
2136static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2137{
2138        BUG();
2139        if (nidp)
2140                *nidp = -1;
2141        return NULL;
2142}
2143
2144static void hugetlb_unregister_all_nodes(void) { }
2145
2146static void hugetlb_register_all_nodes(void) { }
2147
2148#endif
2149
2150static void __exit hugetlb_exit(void)
2151{
2152        struct hstate *h;
2153
2154        hugetlb_unregister_all_nodes();
2155
2156        for_each_hstate(h) {
2157                kobject_put(hstate_kobjs[hstate_index(h)]);
2158        }
2159
2160        kobject_put(hugepages_kobj);
2161        kfree(htlb_fault_mutex_table);
2162}
2163module_exit(hugetlb_exit);
2164
2165static int __init hugetlb_init(void)
2166{
2167        int i;
2168
2169        if (!hugepages_supported())
2170                return 0;
2171
2172        if (!size_to_hstate(default_hstate_size)) {
2173                default_hstate_size = HPAGE_SIZE;
2174                if (!size_to_hstate(default_hstate_size))
2175                        hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2176        }
2177        default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2178        if (default_hstate_max_huge_pages)
2179                default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2180
2181        hugetlb_init_hstates();
2182        gather_bootmem_prealloc();
2183        report_hugepages();
2184
2185        hugetlb_sysfs_init();
2186        hugetlb_register_all_nodes();
2187        hugetlb_cgroup_file_init();
2188
2189#ifdef CONFIG_SMP
2190        num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2191#else
2192        num_fault_mutexes = 1;
2193#endif
2194        htlb_fault_mutex_table =
2195                kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2196        BUG_ON(!htlb_fault_mutex_table);
2197
2198        for (i = 0; i < num_fault_mutexes; i++)
2199                mutex_init(&htlb_fault_mutex_table[i]);
2200        return 0;
2201}
2202module_init(hugetlb_init);
2203
2204/* Should be called on processing a hugepagesz=... option */
2205void __init hugetlb_add_hstate(unsigned int order)
2206{
2207        struct hstate *h;
2208        unsigned long i;
2209
2210        if (size_to_hstate(PAGE_SIZE << order)) {
2211                pr_warning("hugepagesz= specified twice, ignoring\n");
2212                return;
2213        }
2214        BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2215        BUG_ON(order == 0);
2216        h = &hstates[hugetlb_max_hstate++];
2217        h->order = order;
2218        h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2219        h->nr_huge_pages = 0;
2220        h->free_huge_pages = 0;
2221        for (i = 0; i < MAX_NUMNODES; ++i)
2222                INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2223        INIT_LIST_HEAD(&h->hugepage_activelist);
2224        h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2225        h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2226        snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2227                                        huge_page_size(h)/1024);
2228
2229        parsed_hstate = h;
2230}
2231
2232static int __init hugetlb_nrpages_setup(char *s)
2233{
2234        unsigned long *mhp;
2235        static unsigned long *last_mhp;
2236
2237        /*
2238         * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2239         * so this hugepages= parameter goes to the "default hstate".
2240         */
2241        if (!hugetlb_max_hstate)
2242                mhp = &default_hstate_max_huge_pages;
2243        else
2244                mhp = &parsed_hstate->max_huge_pages;
2245
2246        if (mhp == last_mhp) {
2247                pr_warning("hugepages= specified twice without "
2248                           "interleaving hugepagesz=, ignoring\n");
2249                return 1;
2250        }
2251
2252        if (sscanf(s, "%lu", mhp) <= 0)
2253                *mhp = 0;
2254
2255        /*
2256         * Global state is always initialized later in hugetlb_init.
2257         * But we need to allocate >= MAX_ORDER hstates here early to still
2258         * use the bootmem allocator.
2259         */
2260        if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2261                hugetlb_hstate_alloc_pages(parsed_hstate);
2262
2263        last_mhp = mhp;
2264
2265        return 1;
2266}
2267__setup("hugepages=", hugetlb_nrpages_setup);
2268
2269static int __init hugetlb_default_setup(char *s)
2270{
2271        default_hstate_size = memparse(s, &s);
2272        return 1;
2273}
2274__setup("default_hugepagesz=", hugetlb_default_setup);
2275
2276static unsigned int cpuset_mems_nr(unsigned int *array)
2277{
2278        int node;
2279        unsigned int nr = 0;
2280
2281        for_each_node_mask(node, cpuset_current_mems_allowed)
2282                nr += array[node];
2283
2284        return nr;
2285}
2286
2287#ifdef CONFIG_SYSCTL
2288static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2289                         struct ctl_table *table, int write,
2290                         void __user *buffer, size_t *length, loff_t *ppos)
2291{
2292        struct hstate *h = &default_hstate;
2293        unsigned long tmp = h->max_huge_pages;
2294        int ret;
2295
2296        if (!hugepages_supported())
2297                return -ENOTSUPP;
2298
2299        table->data = &tmp;
2300        table->maxlen = sizeof(unsigned long);
2301        ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2302        if (ret)
2303                goto out;
2304
2305        if (write)
2306                ret = __nr_hugepages_store_common(obey_mempolicy, h,
2307                                                  NUMA_NO_NODE, tmp, *length);
2308out:
2309        return ret;
2310}
2311
2312int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2313                          void __user *buffer, size_t *length, loff_t *ppos)
2314{
2315
2316        return hugetlb_sysctl_handler_common(false, table, write,
2317                                                        buffer, length, ppos);
2318}
2319
2320#ifdef CONFIG_NUMA
2321int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2322                          void __user *buffer, size_t *length, loff_t *ppos)
2323{
2324        return hugetlb_sysctl_handler_common(true, table, write,
2325                                                        buffer, length, ppos);
2326}
2327#endif /* CONFIG_NUMA */
2328
2329int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2330                        void __user *buffer,
2331                        size_t *length, loff_t *ppos)
2332{
2333        struct hstate *h = &default_hstate;
2334        unsigned long tmp;
2335        int ret;
2336
2337        if (!hugepages_supported())
2338                return -ENOTSUPP;
2339
2340        tmp = h->nr_overcommit_huge_pages;
2341
2342        if (write && hstate_is_gigantic(h))
2343                return -EINVAL;
2344
2345        table->data = &tmp;
2346        table->maxlen = sizeof(unsigned long);
2347        ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2348        if (ret)
2349                goto out;
2350
2351        if (write) {
2352                spin_lock(&hugetlb_lock);
2353                h->nr_overcommit_huge_pages = tmp;
2354                spin_unlock(&hugetlb_lock);
2355        }
2356out:
2357        return ret;
2358}
2359
2360#endif /* CONFIG_SYSCTL */
2361
2362void hugetlb_report_meminfo(struct seq_file *m)
2363{
2364        struct hstate *h = &default_hstate;
2365        if (!hugepages_supported())
2366                return;
2367        seq_printf(m,
2368                        "HugePages_Total:   %5lu\n"
2369                        "HugePages_Free:    %5lu\n"
2370                        "HugePages_Rsvd:    %5lu\n"
2371                        "HugePages_Surp:    %5lu\n"
2372                        "Hugepagesize:   %8lu kB\n",
2373                        h->nr_huge_pages,
2374                        h->free_huge_pages,
2375                        h->resv_huge_pages,
2376                        h->surplus_huge_pages,
2377                        1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2378}
2379
2380int hugetlb_report_node_meminfo(int nid, char *buf)
2381{
2382        struct hstate *h = &default_hstate;
2383        if (!hugepages_supported())
2384                return 0;
2385        return sprintf(buf,
2386                "Node %d HugePages_Total: %5u\n"
2387                "Node %d HugePages_Free:  %5u\n"
2388                "Node %d HugePages_Surp:  %5u\n",
2389                nid, h->nr_huge_pages_node[nid],
2390                nid, h->free_huge_pages_node[nid],
2391                nid, h->surplus_huge_pages_node[nid]);
2392}
2393
2394void hugetlb_show_meminfo(void)
2395{
2396        struct hstate *h;
2397        int nid;
2398
2399        if (!hugepages_supported())
2400                return;
2401
2402        for_each_node_state(nid, N_MEMORY)
2403                for_each_hstate(h)
2404                        pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2405                                nid,
2406                                h->nr_huge_pages_node[nid],
2407                                h->free_huge_pages_node[nid],
2408                                h->surplus_huge_pages_node[nid],
2409                                1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2410}
2411
2412/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2413unsigned long hugetlb_total_pages(void)
2414{
2415        struct hstate *h;
2416        unsigned long nr_total_pages = 0;
2417
2418        for_each_hstate(h)
2419                nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2420        return nr_total_pages;
2421}
2422
2423static int hugetlb_acct_memory(struct hstate *h, long delta)
2424{
2425        int ret = -ENOMEM;
2426
2427        spin_lock(&hugetlb_lock);
2428        /*
2429         * When cpuset is configured, it breaks the strict hugetlb page
2430         * reservation as the accounting is done on a global variable. Such
2431         * reservation is completely rubbish in the presence of cpuset because
2432         * the reservation is not checked against page availability for the
2433         * current cpuset. Application can still potentially OOM'ed by kernel
2434         * with lack of free htlb page in cpuset that the task is in.
2435         * Attempt to enforce strict accounting with cpuset is almost
2436         * impossible (or too ugly) because cpuset is too fluid that
2437         * task or memory node can be dynamically moved between cpusets.
2438         *
2439         * The change of semantics for shared hugetlb mapping with cpuset is
2440         * undesirable. However, in order to preserve some of the semantics,
2441         * we fall back to check against current free page availability as
2442         * a best attempt and hopefully to minimize the impact of changing
2443         * semantics that cpuset has.
2444         */
2445        if (delta > 0) {
2446                if (gather_surplus_pages(h, delta) < 0)
2447                        goto out;
2448
2449                if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2450                        return_unused_surplus_pages(h, delta);
2451                        goto out;
2452                }
2453        }
2454
2455        ret = 0;
2456        if (delta < 0)
2457                return_unused_surplus_pages(h, (unsigned long) -delta);
2458
2459out:
2460        spin_unlock(&hugetlb_lock);
2461        return ret;
2462}
2463
2464static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2465{
2466        struct resv_map *resv = vma_resv_map(vma);
2467
2468        /*
2469         * This new VMA should share its siblings reservation map if present.
2470         * The VMA will only ever have a valid reservation map pointer where
2471         * it is being copied for another still existing VMA.  As that VMA
2472         * has a reference to the reservation map it cannot disappear until
2473         * after this open call completes.  It is therefore safe to take a
2474         * new reference here without additional locking.
2475         */
2476        if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2477                kref_get(&resv->refs);
2478}
2479
2480static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2481{
2482        struct hstate *h = hstate_vma(vma);
2483        struct resv_map *resv = vma_resv_map(vma);
2484        struct hugepage_subpool *spool = subpool_vma(vma);
2485        unsigned long reserve, start, end;
2486
2487        if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2488                return;
2489
2490        start = vma_hugecache_offset(h, vma, vma->vm_start);
2491        end = vma_hugecache_offset(h, vma, vma->vm_end);
2492
2493        reserve = (end - start) - region_count(resv, start, end);
2494
2495        kref_put(&resv->refs, resv_map_release);
2496
2497        if (reserve) {
2498                hugetlb_acct_memory(h, -reserve);
2499                hugepage_subpool_put_pages(spool, reserve);
2500        }
2501}
2502
2503/*
2504 * We cannot handle pagefaults against hugetlb pages at all.  They cause
2505 * handle_mm_fault() to try to instantiate regular-sized pages in the
2506 * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
2507 * this far.
2508 */
2509static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2510{
2511        BUG();
2512        return 0;
2513}
2514
2515const struct vm_operations_struct hugetlb_vm_ops = {
2516        .fault = hugetlb_vm_op_fault,
2517        .open = hugetlb_vm_op_open,
2518        .close = hugetlb_vm_op_close,
2519};
2520
2521static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2522                                int writable)
2523{
2524        pte_t entry;
2525
2526        if (writable) {
2527                entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2528                                         vma->vm_page_prot)));
2529        } else {
2530                entry = huge_pte_wrprotect(mk_huge_pte(page,
2531                                           vma->vm_page_prot));
2532        }
2533        entry = pte_mkyoung(entry);
2534        entry = pte_mkhuge(entry);
2535        entry = arch_make_huge_pte(entry, vma, page, writable);
2536
2537        return entry;
2538}
2539
2540static void set_huge_ptep_writable(struct vm_area_struct *vma,
2541                                   unsigned long address, pte_t *ptep)
2542{
2543        pte_t entry;
2544
2545        entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2546        if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2547                update_mmu_cache(vma, address, ptep);
2548}
2549
2550static int is_hugetlb_entry_migration(pte_t pte)
2551{
2552        swp_entry_t swp;
2553
2554        if (huge_pte_none(pte) || pte_present(pte))
2555                return 0;
2556        swp = pte_to_swp_entry(pte);
2557        if (non_swap_entry(swp) && is_migration_entry(swp))
2558                return 1;
2559        else
2560                return 0;
2561}
2562
2563static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2564{
2565        swp_entry_t swp;
2566
2567        if (huge_pte_none(pte) || pte_present(pte))
2568                return 0;
2569        swp = pte_to_swp_entry(pte);
2570        if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2571                return 1;
2572        else
2573                return 0;
2574}
2575
2576int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2577                            struct vm_area_struct *vma)
2578{
2579        pte_t *src_pte, *dst_pte, entry;
2580        struct page *ptepage;
2581        unsigned long addr;
2582        int cow;
2583        struct hstate *h = hstate_vma(vma);
2584        unsigned long sz = huge_page_size(h);
2585        unsigned long mmun_start;       /* For mmu_notifiers */
2586        unsigned long mmun_end;         /* For mmu_notifiers */
2587        int ret = 0;
2588
2589        cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2590
2591        mmun_start = vma->vm_start;
2592        mmun_end = vma->vm_end;
2593        if (cow)
2594                mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2595
2596        for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2597                spinlock_t *src_ptl, *dst_ptl;
2598                src_pte = huge_pte_offset(src, addr);
2599                if (!src_pte)
2600                        continue;
2601                dst_pte = huge_pte_alloc(dst, addr, sz);
2602                if (!dst_pte) {
2603                        ret = -ENOMEM;
2604                        break;
2605                }
2606
2607                /* If the pagetables are shared don't copy or take references */
2608                if (dst_pte == src_pte)
2609                        continue;
2610
2611                dst_ptl = huge_pte_lock(h, dst, dst_pte);
2612                src_ptl = huge_pte_lockptr(h, src, src_pte);
2613                spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
2614                entry = huge_ptep_get(src_pte);
2615                if (huge_pte_none(entry)) { /* skip none entry */
2616                        ;
2617                } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2618                                    is_hugetlb_entry_hwpoisoned(entry))) {
2619                        swp_entry_t swp_entry = pte_to_swp_entry(entry);
2620
2621                        if (is_write_migration_entry(swp_entry) && cow) {
2622                                /*
2623                                 * COW mappings require pages in both
2624                                 * parent and child to be set to read.
2625                                 */
2626                                make_migration_entry_read(&swp_entry);
2627                                entry = swp_entry_to_pte(swp_entry);
2628                                set_huge_pte_at(src, addr, src_pte, entry);
2629                        }
2630                        set_huge_pte_at(dst, addr, dst_pte, entry);
2631                } else {
2632                        if (cow)
2633                                huge_ptep_set_wrprotect(src, addr, src_pte);
2634                        entry = huge_ptep_get(src_pte);
2635                        ptepage = pte_page(entry);
2636                        get_page(ptepage);
2637                        page_dup_rmap(ptepage);
2638                        set_huge_pte_at(dst, addr, dst_pte, entry);
2639                }
2640                spin_unlock(src_ptl);
2641                spin_unlock(dst_ptl);
2642        }
2643
2644        if (cow)
2645                mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
2646
2647        return ret;
2648}
2649
2650void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2651                            unsigned long start, unsigned long end,
2652                            struct page *ref_page)
2653{
2654        int force_flush = 0;
2655        struct mm_struct *mm = vma->vm_mm;
2656        unsigned long address;
2657        pte_t *ptep;
2658        pte_t pte;
2659        spinlock_t *ptl;
2660        struct page *page;
2661        struct hstate *h = hstate_vma(vma);
2662        unsigned long sz = huge_page_size(h);
2663        const unsigned long mmun_start = start; /* For mmu_notifiers */
2664        const unsigned long mmun_end   = end;   /* For mmu_notifiers */
2665
2666        WARN_ON(!is_vm_hugetlb_page(vma));
2667        BUG_ON(start & ~huge_page_mask(h));
2668        BUG_ON(end & ~huge_page_mask(h));
2669
2670        tlb_start_vma(tlb, vma);
2671        mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2672again:
2673        for (address = start; address < end; address += sz) {
2674                ptep = huge_pte_offset(mm, address);
2675                if (!ptep)
2676                        continue;
2677
2678                ptl = huge_pte_lock(h, mm, ptep);
2679                if (huge_pmd_unshare(mm, &address, ptep))
2680                        goto unlock;
2681
2682                pte = huge_ptep_get(ptep);
2683                if (huge_pte_none(pte))
2684                        goto unlock;
2685
2686                /*
2687                 * Migrating hugepage or HWPoisoned hugepage is already
2688                 * unmapped and its refcount is dropped, so just clear pte here.
2689                 */
2690                if (unlikely(!pte_present(pte))) {
2691                        huge_pte_clear(mm, address, ptep);
2692                        goto unlock;
2693                }
2694
2695                page = pte_page(pte);
2696                /*
2697                 * If a reference page is supplied, it is because a specific
2698                 * page is being unmapped, not a range. Ensure the page we
2699                 * are about to unmap is the actual page of interest.
2700                 */
2701                if (ref_page) {
2702                        if (page != ref_page)
2703                                goto unlock;
2704
2705                        /*
2706                         * Mark the VMA as having unmapped its page so that
2707                         * future faults in this VMA will fail rather than
2708                         * looking like data was lost
2709                         */
2710                        set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2711                }
2712
2713                pte = huge_ptep_get_and_clear(mm, address, ptep);
2714                tlb_remove_tlb_entry(tlb, ptep, address);
2715                if (huge_pte_dirty(pte))
2716                        set_page_dirty(page);
2717
2718                page_remove_rmap(page);
2719                force_flush = !__tlb_remove_page(tlb, page);
2720                if (force_flush) {
2721                        spin_unlock(ptl);
2722                        break;
2723                }
2724                /* Bail out after unmapping reference page if supplied */
2725                if (ref_page) {
2726                        spin_unlock(ptl);
2727                        break;
2728                }
2729unlock:
2730                spin_unlock(ptl);
2731        }
2732        /*
2733         * mmu_gather ran out of room to batch pages, we break out of
2734         * the PTE lock to avoid doing the potential expensive TLB invalidate
2735         * and page-free while holding it.
2736         */
2737        if (force_flush) {
2738                force_flush = 0;
2739                tlb_flush_mmu(tlb);
2740                if (address < end && !ref_page)
2741                        goto again;
2742        }
2743        mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2744        tlb_end_vma(tlb, vma);
2745}
2746
2747void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2748                          struct vm_area_struct *vma, unsigned long start,
2749                          unsigned long end, struct page *ref_page)
2750{
2751        __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2752
2753        /*
2754         * Clear this flag so that x86's huge_pmd_share page_table_shareable
2755         * test will fail on a vma being torn down, and not grab a page table
2756         * on its way out.  We're lucky that the flag has such an appropriate
2757         * name, and can in fact be safely cleared here. We could clear it
2758         * before the __unmap_hugepage_range above, but all that's necessary
2759         * is to clear it before releasing the i_mmap_mutex. This works
2760         * because in the context this is called, the VMA is about to be
2761         * destroyed and the i_mmap_mutex is held.
2762         */
2763        vma->vm_flags &= ~VM_MAYSHARE;
2764}
2765
2766void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2767                          unsigned long end, struct page *ref_page)
2768{
2769        struct mm_struct *mm;
2770        struct mmu_gather tlb;
2771
2772        mm = vma->vm_mm;
2773
2774        tlb_gather_mmu(&tlb, mm, start, end);
2775        __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2776        tlb_finish_mmu(&tlb, start, end);
2777}
2778
2779/*
2780 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2781 * mappping it owns the reserve page for. The intention is to unmap the page
2782 * from other VMAs and let the children be SIGKILLed if they are faulting the
2783 * same region.
2784 */
2785static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2786                              struct page *page, unsigned long address)
2787{
2788        struct hstate *h = hstate_vma(vma);
2789        struct vm_area_struct *iter_vma;
2790        struct address_space *mapping;
2791        pgoff_t pgoff;
2792
2793        /*
2794         * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2795         * from page cache lookup which is in HPAGE_SIZE units.
2796         */
2797        address = address & huge_page_mask(h);
2798        pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2799                        vma->vm_pgoff;
2800        mapping = file_inode(vma->vm_file)->i_mapping;
2801
2802        /*
2803         * Take the mapping lock for the duration of the table walk. As
2804         * this mapping should be shared between all the VMAs,
2805         * __unmap_hugepage_range() is called as the lock is already held
2806         */
2807        mutex_lock(&mapping->i_mmap_mutex);
2808        vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2809                /* Do not unmap the current VMA */
2810                if (iter_vma == vma)
2811                        continue;
2812
2813                /*
2814                 * Shared VMAs have their own reserves and do not affect
2815                 * MAP_PRIVATE accounting but it is possible that a shared
2816                 * VMA is using the same page so check and skip such VMAs.
2817                 */
2818                if (iter_vma->vm_flags & VM_MAYSHARE)
2819                        continue;
2820
2821                /*
2822                 * Unmap the page from other VMAs without their own reserves.
2823                 * They get marked to be SIGKILLed if they fault in these
2824                 * areas. This is because a future no-page fault on this VMA
2825                 * could insert a zeroed page instead of the data existing
2826                 * from the time of fork. This would look like data corruption
2827                 */
2828                if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2829                        unmap_hugepage_range(iter_vma, address,
2830                                             address + huge_page_size(h), page);
2831        }
2832        mutex_unlock(&mapping->i_mmap_mutex);
2833}
2834
2835/*
2836 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2837 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2838 * cannot race with other handlers or page migration.
2839 * Keep the pte_same checks anyway to make transition from the mutex easier.
2840 */
2841static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2842                        unsigned long address, pte_t *ptep, pte_t pte,
2843                        struct page *pagecache_page, spinlock_t *ptl)
2844{
2845        struct hstate *h = hstate_vma(vma);
2846        struct page *old_page, *new_page;
2847        int ret = 0, outside_reserve = 0;
2848        unsigned long mmun_start;       /* For mmu_notifiers */
2849        unsigned long mmun_end;         /* For mmu_notifiers */
2850
2851        old_page = pte_page(pte);
2852
2853retry_avoidcopy:
2854        /* If no-one else is actually using this page, avoid the copy
2855         * and just make the page writable */
2856        if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2857                page_move_anon_rmap(old_page, vma, address);
2858                set_huge_ptep_writable(vma, address, ptep);
2859                return 0;
2860        }
2861
2862        /*
2863         * If the process that created a MAP_PRIVATE mapping is about to
2864         * perform a COW due to a shared page count, attempt to satisfy
2865         * the allocation without using the existing reserves. The pagecache
2866         * page is used to determine if the reserve at this address was
2867         * consumed or not. If reserves were used, a partial faulted mapping
2868         * at the time of fork() could consume its reserves on COW instead
2869         * of the full address range.
2870         */
2871        if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2872                        old_page != pagecache_page)
2873                outside_reserve = 1;
2874
2875        page_cache_get(old_page);
2876
2877        /*
2878         * Drop page table lock as buddy allocator may be called. It will
2879         * be acquired again before returning to the caller, as expected.
2880         */
2881        spin_unlock(ptl);
2882        new_page = alloc_huge_page(vma, address, outside_reserve);
2883
2884        if (IS_ERR(new_page)) {
2885                /*
2886                 * If a process owning a MAP_PRIVATE mapping fails to COW,
2887                 * it is due to references held by a child and an insufficient
2888                 * huge page pool. To guarantee the original mappers
2889                 * reliability, unmap the page from child processes. The child
2890                 * may get SIGKILLed if it later faults.
2891                 */
2892                if (outside_reserve) {
2893                        page_cache_release(old_page);
2894                        BUG_ON(huge_pte_none(pte));
2895                        unmap_ref_private(mm, vma, old_page, address);
2896                        BUG_ON(huge_pte_none(pte));
2897                        spin_lock(ptl);
2898                        ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2899                        if (likely(ptep &&
2900                                   pte_same(huge_ptep_get(ptep), pte)))
2901                                goto retry_avoidcopy;
2902                        /*
2903                         * race occurs while re-acquiring page table
2904                         * lock, and our job is done.
2905                         */
2906                        return 0;
2907                }
2908
2909                ret = (PTR_ERR(new_page) == -ENOMEM) ?
2910                        VM_FAULT_OOM : VM_FAULT_SIGBUS;
2911                goto out_release_old;
2912        }
2913
2914        /*
2915         * When the original hugepage is shared one, it does not have
2916         * anon_vma prepared.
2917         */
2918        if (unlikely(anon_vma_prepare(vma))) {
2919                ret = VM_FAULT_OOM;
2920                goto out_release_all;
2921        }
2922
2923        copy_user_huge_page(new_page, old_page, address, vma,
2924                            pages_per_huge_page(h));
2925        __SetPageUptodate(new_page);
2926        set_page_huge_active(new_page);
2927
2928        mmun_start = address & huge_page_mask(h);
2929        mmun_end = mmun_start + huge_page_size(h);
2930        mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2931
2932        /*
2933         * Retake the page table lock to check for racing updates
2934         * before the page tables are altered
2935         */
2936        spin_lock(ptl);
2937        ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2938        if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
2939                ClearPagePrivate(new_page);
2940
2941                /* Break COW */
2942                huge_ptep_clear_flush(vma, address, ptep);
2943                set_huge_pte_at(mm, address, ptep,
2944                                make_huge_pte(vma, new_page, 1));
2945                page_remove_rmap(old_page);
2946                hugepage_add_new_anon_rmap(new_page, vma, address);
2947                /* Make the old page be freed below */
2948                new_page = old_page;
2949        }
2950        spin_unlock(ptl);
2951        mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2952out_release_all:
2953        page_cache_release(new_page);
2954out_release_old:
2955        page_cache_release(old_page);
2956
2957        spin_lock(ptl); /* Caller expects lock to be held */
2958        return ret;
2959}
2960
2961/* Return the pagecache page at a given address within a VMA */
2962static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2963                        struct vm_area_struct *vma, unsigned long address)
2964{
2965        struct address_space *mapping;
2966        pgoff_t idx;
2967
2968        mapping = vma->vm_file->f_mapping;
2969        idx = vma_hugecache_offset(h, vma, address);
2970
2971        return find_lock_page(mapping, idx);
2972}
2973
2974/*
2975 * Return whether there is a pagecache page to back given address within VMA.
2976 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2977 */
2978static bool hugetlbfs_pagecache_present(struct hstate *h,
2979                        struct vm_area_struct *vma, unsigned long address)
2980{
2981        struct address_space *mapping;
2982        pgoff_t idx;
2983        struct page *page;
2984
2985        mapping = vma->vm_file->f_mapping;
2986        idx = vma_hugecache_offset(h, vma, address);
2987
2988        page = find_get_page(mapping, idx);
2989        if (page)
2990                put_page(page);
2991        return page != NULL;
2992}
2993
2994static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2995                           struct address_space *mapping, pgoff_t idx,
2996                           unsigned long address, pte_t *ptep, unsigned int flags)
2997{
2998        struct hstate *h = hstate_vma(vma);
2999        int ret = VM_FAULT_SIGBUS;
3000        int anon_rmap = 0;
3001        unsigned long size;
3002        struct page *page;
3003        pte_t new_pte;
3004        spinlock_t *ptl;
3005
3006        /*
3007         * Currently, we are forced to kill the process in the event the
3008         * original mapper has unmapped pages from the child due to a failed
3009         * COW. Warn that such a situation has occurred as it may not be obvious
3010         */
3011        if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3012                pr_warning("PID %d killed due to inadequate hugepage pool\n",
3013                           current->pid);
3014                return ret;
3015        }
3016
3017        /*
3018         * Use page lock to guard against racing truncation
3019         * before we get page_table_lock.
3020         */
3021retry:
3022        page = find_lock_page(mapping, idx);
3023        if (!page) {
3024                size = i_size_read(mapping->host) >> huge_page_shift(h);
3025                if (idx >= size)
3026                        goto out;
3027                page = alloc_huge_page(vma, address, 0);
3028                if (IS_ERR(page)) {
3029                        ret = PTR_ERR(page);
3030                        if (ret == -ENOMEM)
3031                                ret = VM_FAULT_OOM;
3032                        else
3033                                ret = VM_FAULT_SIGBUS;
3034                        goto out;
3035                }
3036                clear_huge_page(page, address, pages_per_huge_page(h));
3037                __SetPageUptodate(page);
3038                set_page_huge_active(page);
3039
3040                if (vma->vm_flags & VM_MAYSHARE) {
3041                        int err;
3042                        struct inode *inode = mapping->host;
3043
3044                        err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3045                        if (err) {
3046                                put_page(page);
3047                                if (err == -EEXIST)
3048                                        goto retry;
3049                                goto out;
3050                        }
3051                        ClearPagePrivate(page);
3052
3053                        spin_lock(&inode->i_lock);
3054                        inode->i_blocks += blocks_per_huge_page(h);
3055                        spin_unlock(&inode->i_lock);
3056                } else {
3057                        lock_page(page);
3058                        if (unlikely(anon_vma_prepare(vma))) {
3059                                ret = VM_FAULT_OOM;
3060                                goto backout_unlocked;
3061                        }
3062                        anon_rmap = 1;
3063                }
3064        } else {
3065                /*
3066                 * If memory error occurs between mmap() and fault, some process
3067                 * don't have hwpoisoned swap entry for errored virtual address.
3068                 * So we need to block hugepage fault by PG_hwpoison bit check.
3069                 */
3070                if (unlikely(PageHWPoison(page))) {
3071                        ret = VM_FAULT_HWPOISON |
3072                                VM_FAULT_SET_HINDEX(hstate_index(h));
3073                        goto backout_unlocked;
3074                }
3075        }
3076
3077        /*
3078         * If we are going to COW a private mapping later, we examine the
3079         * pending reservations for this page now. This will ensure that
3080         * any allocations necessary to record that reservation occur outside
3081         * the spinlock.
3082         */
3083        if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
3084                if (vma_needs_reservation(h, vma, address) < 0) {
3085                        ret = VM_FAULT_OOM;
3086                        goto backout_unlocked;
3087                }
3088
3089        ptl = huge_pte_lockptr(h, mm, ptep);
3090        spin_lock(ptl);
3091        size = i_size_read(mapping->host) >> huge_page_shift(h);
3092        if (idx >= size)
3093                goto backout;
3094
3095        ret = 0;
3096        if (!huge_pte_none(huge_ptep_get(ptep)))
3097                goto backout;
3098
3099        if (anon_rmap) {
3100                ClearPagePrivate(page);
3101                hugepage_add_new_anon_rmap(page, vma, address);
3102        } else
3103                page_dup_rmap(page);
3104        new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3105                                && (vma->vm_flags & VM_SHARED)));
3106        set_huge_pte_at(mm, address, ptep, new_pte);
3107
3108        if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3109                /* Optimization, do the COW without a second fault */
3110                ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
3111        }
3112
3113        spin_unlock(ptl);
3114        unlock_page(page);
3115out:
3116        return ret;
3117
3118backout:
3119        spin_unlock(ptl);
3120backout_unlocked:
3121        unlock_page(page);
3122        put_page(page);
3123        goto out;
3124}
3125
3126#ifdef CONFIG_SMP
3127static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3128                            struct vm_area_struct *vma,
3129                            struct address_space *mapping,
3130                            pgoff_t idx, unsigned long address)
3131{
3132        unsigned long key[2];
3133        u32 hash;
3134
3135        if (vma->vm_flags & VM_SHARED) {
3136                key[0] = (unsigned long) mapping;
3137                key[1] = idx;
3138        } else {
3139                key[0] = (unsigned long) mm;
3140                key[1] = address >> huge_page_shift(h);
3141        }
3142
3143        hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3144
3145        return hash & (num_fault_mutexes - 1);
3146}
3147#else
3148/*
3149 * For uniprocesor systems we always use a single mutex, so just
3150 * return 0 and avoid the hashing overhead.
3151 */
3152static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3153                            struct vm_area_struct *vma,
3154                            struct address_space *mapping,
3155                            pgoff_t idx, unsigned long address)
3156{
3157        return 0;
3158}
3159#endif
3160
3161int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3162                        unsigned long address, unsigned int flags)
3163{
3164        pte_t *ptep, entry;
3165        spinlock_t *ptl;
3166        int ret;
3167        u32 hash;
3168        pgoff_t idx;
3169        struct page *page = NULL;
3170        struct page *pagecache_page = NULL;
3171        struct hstate *h = hstate_vma(vma);
3172        struct address_space *mapping;
3173        int need_wait_lock = 0;
3174
3175        address &= huge_page_mask(h);
3176
3177        ptep = huge_pte_offset(mm, address);
3178        if (ptep) {
3179                entry = huge_ptep_get(ptep);
3180                if (unlikely(is_hugetlb_entry_migration(entry))) {
3181                        migration_entry_wait_huge(vma, mm, ptep);
3182                        return 0;
3183                } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3184                        return VM_FAULT_HWPOISON_LARGE |
3185                                VM_FAULT_SET_HINDEX(hstate_index(h));
3186        }
3187
3188        ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3189        if (!ptep)
3190                return VM_FAULT_OOM;
3191
3192        mapping = vma->vm_file->f_mapping;
3193        idx = vma_hugecache_offset(h, vma, address);
3194
3195        /*
3196         * Serialize hugepage allocation and instantiation, so that we don't
3197         * get spurious allocation failures if two CPUs race to instantiate
3198         * the same page in the page cache.
3199         */
3200        hash = fault_mutex_hash(h, mm, vma, mapping, idx, address);
3201        mutex_lock(&htlb_fault_mutex_table[hash]);
3202
3203        entry = huge_ptep_get(ptep);
3204        if (huge_pte_none(entry)) {
3205                ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3206                goto out_mutex;
3207        }
3208
3209        ret = 0;
3210
3211        /*
3212         * entry could be a migration/hwpoison entry at this point, so this
3213         * check prevents the kernel from going below assuming that we have
3214         * a active hugepage in pagecache. This goto expects the 2nd page fault,
3215         * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3216         * handle it.
3217         */
3218        if (!pte_present(entry))
3219                goto out_mutex;
3220
3221        /*
3222         * If we are going to COW the mapping later, we examine the pending
3223         * reservations for this page now. This will ensure that any
3224         * allocations necessary to record that reservation occur outside the
3225         * spinlock. For private mappings, we also lookup the pagecache
3226         * page now as it is used to determine if a reservation has been
3227         * consumed.
3228         */
3229        if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3230                if (vma_needs_reservation(h, vma, address) < 0) {
3231                        ret = VM_FAULT_OOM;
3232                        goto out_mutex;
3233                }
3234
3235                if (!(vma->vm_flags & VM_MAYSHARE))
3236                        pagecache_page = hugetlbfs_pagecache_page(h,
3237                                                                vma, address);
3238        }
3239
3240        ptl = huge_pte_lock(h, mm, ptep);
3241
3242        /* Check for a racing update before calling hugetlb_cow */
3243        if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3244                goto out_ptl;
3245
3246        /*
3247         * hugetlb_cow() requires page locks of pte_page(entry) and
3248         * pagecache_page, so here we need take the former one
3249         * when page != pagecache_page or !pagecache_page.
3250         */
3251        page = pte_page(entry);
3252        if (page != pagecache_page)
3253                if (!trylock_page(page)) {
3254                        need_wait_lock = 1;
3255                        goto out_ptl;
3256                }
3257
3258        get_page(page);
3259
3260        if (flags & FAULT_FLAG_WRITE) {
3261                if (!huge_pte_write(entry)) {
3262                        ret = hugetlb_cow(mm, vma, address, ptep, entry,
3263                                        pagecache_page, ptl);
3264                        goto out_put_page;
3265                }
3266                entry = huge_pte_mkdirty(entry);
3267        }
3268        entry = pte_mkyoung(entry);
3269        if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3270                                                flags & FAULT_FLAG_WRITE))
3271                update_mmu_cache(vma, address, ptep);
3272out_put_page:
3273        if (page != pagecache_page)
3274                unlock_page(page);
3275        put_page(page);
3276out_ptl:
3277        spin_unlock(ptl);
3278
3279        if (pagecache_page) {
3280                unlock_page(pagecache_page);
3281                put_page(pagecache_page);
3282        }
3283out_mutex:
3284        mutex_unlock(&htlb_fault_mutex_table[hash]);
3285        /*
3286         * Generally it's safe to hold refcount during waiting page lock. But
3287         * here we just wait to defer the next page fault to avoid busy loop and
3288         * the page is not used after unlocked before returning from the current
3289         * page fault. So we are safe from accessing freed page, even if we wait
3290         * here without taking refcount.
3291         */
3292        if (need_wait_lock)
3293                wait_on_page_locked(page);
3294        return ret;
3295}
3296
3297long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3298                         struct page **pages, struct vm_area_struct **vmas,
3299                         unsigned long *position, unsigned long *nr_pages,
3300                         long i, unsigned int flags)
3301{
3302        unsigned long pfn_offset;
3303        unsigned long vaddr = *position;
3304        unsigned long remainder = *nr_pages;
3305        struct hstate *h = hstate_vma(vma);
3306
3307        while (vaddr < vma->vm_end && remainder) {
3308                pte_t *pte;
3309                spinlock_t *ptl = NULL;
3310                int absent;
3311                struct page *page;
3312
3313                /*
3314                 * Some archs (sparc64, sh*) have multiple pte_ts to
3315                 * each hugepage.  We have to make sure we get the
3316                 * first, for the page indexing below to work.
3317                 *
3318                 * Note that page table lock is not held when pte is null.
3319                 */
3320                pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3321                if (pte)
3322                        ptl = huge_pte_lock(h, mm, pte);
3323                absent = !pte || huge_pte_none(huge_ptep_get(pte));
3324
3325                /*
3326                 * When coredumping, it suits get_dump_page if we just return
3327                 * an error where there's an empty slot with no huge pagecache
3328                 * to back it.  This way, we avoid allocating a hugepage, and
3329                 * the sparse dumpfile avoids allocating disk blocks, but its
3330                 * huge holes still show up with zeroes where they need to be.
3331                 */
3332                if (absent && (flags & FOLL_DUMP) &&
3333                    !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3334                        if (pte)
3335                                spin_unlock(ptl);
3336                        remainder = 0;
3337                        break;
3338                }
3339
3340                /*
3341                 * We need call hugetlb_fault for both hugepages under migration
3342                 * (in which case hugetlb_fault waits for the migration,) and
3343                 * hwpoisoned hugepages (in which case we need to prevent the
3344                 * caller from accessing to them.) In order to do this, we use
3345                 * here is_swap_pte instead of is_hugetlb_entry_migration and
3346                 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3347                 * both cases, and because we can't follow correct pages
3348                 * directly from any kind of swap entries.
3349                 */
3350                if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3351                    ((flags & FOLL_WRITE) &&
3352                      !huge_pte_write(huge_ptep_get(pte)))) {
3353                        int ret;
3354
3355                        if (pte)
3356                                spin_unlock(ptl);
3357                        ret = hugetlb_fault(mm, vma, vaddr,
3358                                (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3359                        if (!(ret & VM_FAULT_ERROR))
3360                                continue;
3361
3362                        remainder = 0;
3363                        break;
3364                }
3365
3366                pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3367                page = pte_page(huge_ptep_get(pte));
3368same_page:
3369                if (pages) {
3370                        pages[i] = mem_map_offset(page, pfn_offset);
3371                        get_page_foll(pages[i]);
3372                }
3373
3374                if (vmas)
3375                        vmas[i] = vma;
3376
3377                vaddr += PAGE_SIZE;
3378                ++pfn_offset;
3379                --remainder;
3380                ++i;
3381                if (vaddr < vma->vm_end && remainder &&
3382                                pfn_offset < pages_per_huge_page(h)) {
3383                        /*
3384                         * We use pfn_offset to avoid touching the pageframes
3385                         * of this compound page.
3386                         */
3387                        goto same_page;
3388                }
3389                spin_unlock(ptl);
3390        }
3391        *nr_pages = remainder;
3392        *position = vaddr;
3393
3394        return i ? i : -EFAULT;
3395}
3396
3397unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3398                unsigned long address, unsigned long end, pgprot_t newprot)
3399{
3400        struct mm_struct *mm = vma->vm_mm;
3401        unsigned long start = address;
3402        pte_t *ptep;
3403        pte_t pte;
3404        struct hstate *h = hstate_vma(vma);
3405        unsigned long pages = 0;
3406
3407        BUG_ON(address >= end);
3408        flush_cache_range(vma, address, end);
3409
3410        mmu_notifier_invalidate_range_start(mm, start, end);
3411        mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3412        for (; address < end; address += huge_page_size(h)) {
3413                spinlock_t *ptl;
3414                ptep = huge_pte_offset(mm, address);
3415                if (!ptep)
3416                        continue;
3417                ptl = huge_pte_lock(h, mm, ptep);
3418                if (huge_pmd_unshare(mm, &address, ptep)) {
3419                        pages++;
3420                        spin_unlock(ptl);
3421                        continue;
3422                }
3423                pte = huge_ptep_get(ptep);
3424                if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3425                        spin_unlock(ptl);
3426                        continue;
3427                }
3428                if (unlikely(is_hugetlb_entry_migration(pte))) {
3429                        swp_entry_t entry = pte_to_swp_entry(pte);
3430
3431                        if (is_write_migration_entry(entry)) {
3432                                pte_t newpte;
3433
3434                                make_migration_entry_read(&entry);
3435                                newpte = swp_entry_to_pte(entry);
3436                                set_huge_pte_at(mm, address, ptep, newpte);
3437                                pages++;
3438                        }
3439                        spin_unlock(ptl);
3440                        continue;
3441                }
3442                if (!huge_pte_none(pte)) {
3443                        pte = huge_ptep_get_and_clear(mm, address, ptep);
3444                        pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3445                        pte = arch_make_huge_pte(pte, vma, NULL, 0);
3446                        set_huge_pte_at(mm, address, ptep, pte);
3447                        pages++;
3448                }
3449                spin_unlock(ptl);
3450        }
3451        /*
3452         * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3453         * may have cleared our pud entry and done put_page on the page table:
3454         * once we release i_mmap_mutex, another task can do the final put_page
3455         * and that page table be reused and filled with junk.
3456         */
3457        flush_tlb_range(vma, start, end);
3458        mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3459        mmu_notifier_invalidate_range_end(mm, start, end);
3460
3461        return pages << h->order;
3462}
3463
3464int hugetlb_reserve_pages(struct inode *inode,
3465                                        long from, long to,
3466                                        struct vm_area_struct *vma,
3467                                        vm_flags_t vm_flags)
3468{
3469        long ret, chg;
3470        struct hstate *h = hstate_inode(inode);
3471        struct hugepage_subpool *spool = subpool_inode(inode);
3472        struct resv_map *resv_map;
3473
3474        /*
3475         * Only apply hugepage reservation if asked. At fault time, an
3476         * attempt will be made for VM_NORESERVE to allocate a page
3477         * without using reserves
3478         */
3479        if (vm_flags & VM_NORESERVE)
3480                return 0;
3481
3482        /*
3483         * Shared mappings base their reservation on the number of pages that
3484         * are already allocated on behalf of the file. Private mappings need
3485         * to reserve the full area even if read-only as mprotect() may be
3486         * called to make the mapping read-write. Assume !vma is a shm mapping
3487         */
3488        if (!vma || vma->vm_flags & VM_MAYSHARE) {
3489                resv_map = inode_resv_map(inode);
3490
3491                chg = region_chg(resv_map, from, to);
3492
3493        } else {
3494                resv_map = resv_map_alloc();
3495                if (!resv_map)
3496                        return -ENOMEM;
3497
3498                chg = to - from;
3499
3500                set_vma_resv_map(vma, resv_map);
3501                set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3502        }
3503
3504        if (chg < 0) {
3505                ret = chg;
3506                goto out_err;
3507        }
3508
3509        /* There must be enough pages in the subpool for the mapping */
3510        if (hugepage_subpool_get_pages(spool, chg)) {
3511                ret = -ENOSPC;
3512                goto out_err;
3513        }
3514
3515        /*
3516         * Check enough hugepages are available for the reservation.
3517         * Hand the pages back to the subpool if there are not
3518         */
3519        ret = hugetlb_acct_memory(h, chg);
3520        if (ret < 0) {
3521                hugepage_subpool_put_pages(spool, chg);
3522                goto out_err;
3523        }
3524
3525        /*
3526         * Account for the reservations made. Shared mappings record regions
3527         * that have reservations as they are shared by multiple VMAs.
3528         * When the last VMA disappears, the region map says how much
3529         * the reservation was and the page cache tells how much of
3530         * the reservation was consumed. Private mappings are per-VMA and
3531         * only the consumed reservations are tracked. When the VMA
3532         * disappears, the original reservation is the VMA size and the
3533         * consumed reservations are stored in the map. Hence, nothing
3534         * else has to be done for private mappings here
3535         */
3536        if (!vma || vma->vm_flags & VM_MAYSHARE)
3537                region_add(resv_map, from, to);
3538        return 0;
3539out_err:
3540        if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3541                kref_put(&resv_map->refs, resv_map_release);
3542        return ret;
3543}
3544
3545void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3546{
3547        struct hstate *h = hstate_inode(inode);
3548        struct resv_map *resv_map = inode_resv_map(inode);
3549        long chg = 0;
3550        struct hugepage_subpool *spool = subpool_inode(inode);
3551
3552        if (resv_map)
3553                chg = region_truncate(resv_map, offset);
3554        spin_lock(&inode->i_lock);
3555        inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3556        spin_unlock(&inode->i_lock);
3557
3558        hugepage_subpool_put_pages(spool, (chg - freed));
3559        hugetlb_acct_memory(h, -(chg - freed));
3560}
3561
3562#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3563static unsigned long page_table_shareable(struct vm_area_struct *svma,
3564                                struct vm_area_struct *vma,
3565                                unsigned long addr, pgoff_t idx)
3566{
3567        unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3568                                svma->vm_start;
3569        unsigned long sbase = saddr & PUD_MASK;
3570        unsigned long s_end = sbase + PUD_SIZE;
3571
3572        /* Allow segments to share if only one is marked locked */
3573        unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3574        unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3575
3576        /*
3577         * match the virtual addresses, permission and the alignment of the
3578         * page table page.
3579         */
3580        if (pmd_index(addr) != pmd_index(saddr) ||
3581            vm_flags != svm_flags ||
3582            sbase < svma->vm_start || svma->vm_end < s_end)
3583                return 0;
3584
3585        return saddr;
3586}
3587
3588static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3589{
3590        unsigned long base = addr & PUD_MASK;
3591        unsigned long end = base + PUD_SIZE;
3592
3593        /*
3594         * check on proper vm_flags and page table alignment
3595         */
3596        if (vma->vm_flags & VM_MAYSHARE &&
3597            vma->vm_start <= base && end <= vma->vm_end)
3598                return 1;
3599        return 0;
3600}
3601
3602/*
3603 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3604 * and returns the corresponding pte. While this is not necessary for the
3605 * !shared pmd case because we can allocate the pmd later as well, it makes the
3606 * code much cleaner. pmd allocation is essential for the shared case because
3607 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3608 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3609 * bad pmd for sharing.
3610 */
3611pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3612{
3613        struct vm_area_struct *vma = find_vma(mm, addr);
3614        struct address_space *mapping = vma->vm_file->f_mapping;
3615        pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3616                        vma->vm_pgoff;
3617        struct vm_area_struct *svma;
3618        unsigned long saddr;
3619        pte_t *spte = NULL;
3620        pte_t *pte;
3621        spinlock_t *ptl;
3622
3623        if (!vma_shareable(vma, addr))
3624                return (pte_t *)pmd_alloc(mm, pud, addr);
3625
3626        mutex_lock(&mapping->i_mmap_mutex);
3627        vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3628                if (svma == vma)
3629                        continue;
3630
3631                saddr = page_table_shareable(svma, vma, addr, idx);
3632                if (saddr) {
3633                        spte = huge_pte_offset(svma->vm_mm, saddr);
3634                        if (spte) {
3635                                get_page(virt_to_page(spte));
3636                                break;
3637                        }
3638                }
3639        }
3640
3641        if (!spte)
3642                goto out;
3643
3644        ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
3645        spin_lock(ptl);
3646        if (pud_none(*pud))
3647                pud_populate(mm, pud,
3648                                (pmd_t *)((unsigned long)spte & PAGE_MASK));
3649        else
3650                put_page(virt_to_page(spte));
3651        spin_unlock(ptl);
3652out:
3653        pte = (pte_t *)pmd_alloc(mm, pud, addr);
3654        mutex_unlock(&mapping->i_mmap_mutex);
3655        return pte;
3656}
3657
3658/*
3659 * unmap huge page backed by shared pte.
3660 *
3661 * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
3662 * indicated by page_count > 1, unmap is achieved by clearing pud and
3663 * decrementing the ref count. If count == 1, the pte page is not shared.
3664 *
3665 * called with page table lock held.
3666 *
3667 * returns: 1 successfully unmapped a shared pte page
3668 *          0 the underlying pte page is not shared, or it is the last user
3669 */
3670int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3671{
3672        pgd_t *pgd = pgd_offset(mm, *addr);
3673        pud_t *pud = pud_offset(pgd, *addr);
3674
3675        BUG_ON(page_count(virt_to_page(ptep)) == 0);
3676        if (page_count(virt_to_page(ptep)) == 1)
3677                return 0;
3678
3679        pud_clear(pud);
3680        put_page(virt_to_page(ptep));
3681        *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3682        return 1;
3683}
3684#define want_pmd_share()        (1)
3685#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3686pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3687{
3688        return NULL;
3689}
3690#define want_pmd_share()        (0)
3691#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3692
3693#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3694pte_t *huge_pte_alloc(struct mm_struct *mm,
3695                        unsigned long addr, unsigned long sz)
3696{
3697        pgd_t *pgd;
3698        pud_t *pud;
3699        pte_t *pte = NULL;
3700
3701        pgd = pgd_offset(mm, addr);
3702        pud = pud_alloc(mm, pgd, addr);
3703        if (pud) {
3704                if (sz == PUD_SIZE) {
3705                        pte = (pte_t *)pud;
3706                } else {
3707                        BUG_ON(sz != PMD_SIZE);
3708                        if (want_pmd_share() && pud_none(*pud))
3709                                pte = huge_pmd_share(mm, addr, pud);
3710                        else
3711                                pte = (pte_t *)pmd_alloc(mm, pud, addr);
3712                }
3713        }
3714        BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3715
3716        return pte;
3717}
3718
3719pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3720{
3721        pgd_t *pgd;
3722        pud_t *pud;
3723        pmd_t *pmd = NULL;
3724
3725        pgd = pgd_offset(mm, addr);
3726        if (pgd_present(*pgd)) {
3727                pud = pud_offset(pgd, addr);
3728                if (pud_present(*pud)) {
3729                        if (pud_huge(*pud))
3730                                return (pte_t *)pud;
3731                        pmd = pmd_offset(pud, addr);
3732                }
3733        }
3734        return (pte_t *) pmd;
3735}
3736
3737#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3738
3739/*
3740 * These functions are overwritable if your architecture needs its own
3741 * behavior.
3742 */
3743struct page * __weak
3744follow_huge_addr(struct mm_struct *mm, unsigned long address,
3745                              int write)
3746{
3747        return ERR_PTR(-EINVAL);
3748}
3749
3750struct page * __weak
3751follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3752                pmd_t *pmd, int flags)
3753{
3754        struct page *page = NULL;
3755        spinlock_t *ptl;
3756        pte_t pte;
3757retry:
3758        ptl = pmd_lockptr(mm, pmd);
3759        spin_lock(ptl);
3760        /*
3761         * make sure that the address range covered by this pmd is not
3762         * unmapped from other threads.
3763         */
3764        if (!pmd_huge(*pmd))
3765                goto out;
3766        pte = huge_ptep_get((pte_t *)pmd);
3767        if (pte_present(pte)) {
3768                page = pte_page(*(pte_t *)pmd) +
3769                        ((address & ~PMD_MASK) >> PAGE_SHIFT);
3770                if (flags & FOLL_GET)
3771                        get_page(page);
3772        } else {
3773                if (is_hugetlb_entry_migration(pte)) {
3774                        spin_unlock(ptl);
3775                        __migration_entry_wait(mm, (pte_t *)pmd, ptl);
3776                        goto retry;
3777                }
3778                /*
3779                 * hwpoisoned entry is treated as no_page_table in
3780                 * follow_page_mask().
3781                 */
3782        }
3783out:
3784        spin_unlock(ptl);
3785        return page;
3786}
3787
3788struct page * __weak
3789follow_huge_pud(struct mm_struct *mm, unsigned long address,
3790                pud_t *pud, int flags)
3791{
3792        if (flags & FOLL_GET)
3793                return NULL;
3794
3795        return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
3796}
3797
3798#ifdef CONFIG_MEMORY_FAILURE
3799
3800/* Should be called in hugetlb_lock */
3801static int is_hugepage_on_freelist(struct page *hpage)
3802{
3803        struct page *page;
3804        struct page *tmp;
3805        struct hstate *h = page_hstate(hpage);
3806        int nid = page_to_nid(hpage);
3807
3808        list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3809                if (page == hpage)
3810                        return 1;
3811        return 0;
3812}
3813
3814/*
3815 * This function is called from memory failure code.
3816 * Assume the caller holds page lock of the head page.
3817 */
3818int dequeue_hwpoisoned_huge_page(struct page *hpage)
3819{
3820        struct hstate *h = page_hstate(hpage);
3821        int nid = page_to_nid(hpage);
3822        int ret = -EBUSY;
3823
3824        spin_lock(&hugetlb_lock);
3825        if (is_hugepage_on_freelist(hpage)) {
3826                /*
3827                 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3828                 * but dangling hpage->lru can trigger list-debug warnings
3829                 * (this happens when we call unpoison_memory() on it),
3830                 * so let it point to itself with list_del_init().
3831                 */
3832                list_del_init(&hpage->lru);
3833                set_page_refcounted(hpage);
3834                h->free_huge_pages--;
3835                h->free_huge_pages_node[nid]--;
3836                ret = 0;
3837        }
3838        spin_unlock(&hugetlb_lock);
3839        return ret;
3840}
3841#endif
3842
3843bool isolate_huge_page(struct page *page, struct list_head *list)
3844{
3845        bool ret = true;
3846
3847        VM_BUG_ON_PAGE(!PageHead(page), page);
3848        spin_lock(&hugetlb_lock);
3849        if (!page_huge_active(page) || !get_page_unless_zero(page)) {
3850                ret = false;
3851                goto unlock;
3852        }
3853        clear_page_huge_active(page);
3854        list_move_tail(&page->lru, list);
3855unlock:
3856        spin_unlock(&hugetlb_lock);
3857        return ret;
3858}
3859
3860void putback_active_hugepage(struct page *page)
3861{
3862        VM_BUG_ON_PAGE(!PageHead(page), page);
3863        spin_lock(&hugetlb_lock);
3864        set_page_huge_active(page);
3865        list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
3866        spin_unlock(&hugetlb_lock);
3867        put_page(page);
3868}
3869
3870bool is_hugepage_active(struct page *page)
3871{
3872        VM_BUG_ON_PAGE(!PageHuge(page), page);
3873        /*
3874         * This function can be called for a tail page because the caller,
3875         * scan_movable_pages, scans through a given pfn-range which typically
3876         * covers one memory block. In systems using gigantic hugepage (1GB
3877         * for x86_64,) a hugepage is larger than a memory block, and we don't
3878         * support migrating such large hugepages for now, so return false
3879         * when called for tail pages.
3880         */
3881        if (PageTail(page))
3882                return false;
3883        /*
3884         * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3885         * so we should return false for them.
3886         */
3887        if (unlikely(PageHWPoison(page)))
3888                return false;
3889        return page_count(page) > 0;
3890}
Note: See TracBrowser for help on using the repository browser.