Merge branch 'hwpoison' of git://git.kernel.org/pub/scm/linux/kernel/git/ak/linux...
[linux-2.6.git] / mm / hugetlb.c
1 /*
2  * Generic hugetlb support.
3  * (C) William Irwin, 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/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/rmap.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
24
25 #include <asm/page.h>
26 #include <asm/pgtable.h>
27 #include <asm/io.h>
28
29 #include <linux/hugetlb.h>
30 #include <linux/node.h>
31 #include "internal.h"
32
33 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
34 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
35 unsigned long hugepages_treat_as_movable;
36
37 static int max_hstate;
38 unsigned int default_hstate_idx;
39 struct hstate hstates[HUGE_MAX_HSTATE];
40
41 __initdata LIST_HEAD(huge_boot_pages);
42
43 /* for command line parsing */
44 static struct hstate * __initdata parsed_hstate;
45 static unsigned long __initdata default_hstate_max_huge_pages;
46 static unsigned long __initdata default_hstate_size;
47
48 #define for_each_hstate(h) \
49         for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
50
51 /*
52  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
53  */
54 static DEFINE_SPINLOCK(hugetlb_lock);
55
56 /*
57  * Region tracking -- allows tracking of reservations and instantiated pages
58  *                    across the pages in a mapping.
59  *
60  * The region data structures are protected by a combination of the mmap_sem
61  * and the hugetlb_instantion_mutex.  To access or modify a region the caller
62  * must either hold the mmap_sem for write, or the mmap_sem for read and
63  * the hugetlb_instantiation mutex:
64  *
65  *      down_write(&mm->mmap_sem);
66  * or
67  *      down_read(&mm->mmap_sem);
68  *      mutex_lock(&hugetlb_instantiation_mutex);
69  */
70 struct file_region {
71         struct list_head link;
72         long from;
73         long to;
74 };
75
76 static long region_add(struct list_head *head, long f, long t)
77 {
78         struct file_region *rg, *nrg, *trg;
79
80         /* Locate the region we are either in or before. */
81         list_for_each_entry(rg, head, link)
82                 if (f <= rg->to)
83                         break;
84
85         /* Round our left edge to the current segment if it encloses us. */
86         if (f > rg->from)
87                 f = rg->from;
88
89         /* Check for and consume any regions we now overlap with. */
90         nrg = rg;
91         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
92                 if (&rg->link == head)
93                         break;
94                 if (rg->from > t)
95                         break;
96
97                 /* If this area reaches higher then extend our area to
98                  * include it completely.  If this is not the first area
99                  * which we intend to reuse, free it. */
100                 if (rg->to > t)
101                         t = rg->to;
102                 if (rg != nrg) {
103                         list_del(&rg->link);
104                         kfree(rg);
105                 }
106         }
107         nrg->from = f;
108         nrg->to = t;
109         return 0;
110 }
111
112 static long region_chg(struct list_head *head, long f, long t)
113 {
114         struct file_region *rg, *nrg;
115         long chg = 0;
116
117         /* Locate the region we are before or in. */
118         list_for_each_entry(rg, head, link)
119                 if (f <= rg->to)
120                         break;
121
122         /* If we are below the current region then a new region is required.
123          * Subtle, allocate a new region at the position but make it zero
124          * size such that we can guarantee to record the reservation. */
125         if (&rg->link == head || t < rg->from) {
126                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
127                 if (!nrg)
128                         return -ENOMEM;
129                 nrg->from = f;
130                 nrg->to   = f;
131                 INIT_LIST_HEAD(&nrg->link);
132                 list_add(&nrg->link, rg->link.prev);
133
134                 return t - f;
135         }
136
137         /* Round our left edge to the current segment if it encloses us. */
138         if (f > rg->from)
139                 f = rg->from;
140         chg = t - f;
141
142         /* Check for and consume any regions we now overlap with. */
143         list_for_each_entry(rg, rg->link.prev, link) {
144                 if (&rg->link == head)
145                         break;
146                 if (rg->from > t)
147                         return chg;
148
149                 /* We overlap with this area, if it extends futher than
150                  * us then we must extend ourselves.  Account for its
151                  * existing reservation. */
152                 if (rg->to > t) {
153                         chg += rg->to - t;
154                         t = rg->to;
155                 }
156                 chg -= rg->to - rg->from;
157         }
158         return chg;
159 }
160
161 static long region_truncate(struct list_head *head, long end)
162 {
163         struct file_region *rg, *trg;
164         long chg = 0;
165
166         /* Locate the region we are either in or before. */
167         list_for_each_entry(rg, head, link)
168                 if (end <= rg->to)
169                         break;
170         if (&rg->link == head)
171                 return 0;
172
173         /* If we are in the middle of a region then adjust it. */
174         if (end > rg->from) {
175                 chg = rg->to - end;
176                 rg->to = end;
177                 rg = list_entry(rg->link.next, typeof(*rg), link);
178         }
179
180         /* Drop any remaining regions. */
181         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
182                 if (&rg->link == head)
183                         break;
184                 chg += rg->to - rg->from;
185                 list_del(&rg->link);
186                 kfree(rg);
187         }
188         return chg;
189 }
190
191 static long region_count(struct list_head *head, long f, long t)
192 {
193         struct file_region *rg;
194         long chg = 0;
195
196         /* Locate each segment we overlap with, and count that overlap. */
197         list_for_each_entry(rg, head, link) {
198                 int seg_from;
199                 int seg_to;
200
201                 if (rg->to <= f)
202                         continue;
203                 if (rg->from >= t)
204                         break;
205
206                 seg_from = max(rg->from, f);
207                 seg_to = min(rg->to, t);
208
209                 chg += seg_to - seg_from;
210         }
211
212         return chg;
213 }
214
215 /*
216  * Convert the address within this vma to the page offset within
217  * the mapping, in pagecache page units; huge pages here.
218  */
219 static pgoff_t vma_hugecache_offset(struct hstate *h,
220                         struct vm_area_struct *vma, unsigned long address)
221 {
222         return ((address - vma->vm_start) >> huge_page_shift(h)) +
223                         (vma->vm_pgoff >> huge_page_order(h));
224 }
225
226 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
227                                      unsigned long address)
228 {
229         return vma_hugecache_offset(hstate_vma(vma), vma, address);
230 }
231
232 /*
233  * Return the size of the pages allocated when backing a VMA. In the majority
234  * cases this will be same size as used by the page table entries.
235  */
236 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
237 {
238         struct hstate *hstate;
239
240         if (!is_vm_hugetlb_page(vma))
241                 return PAGE_SIZE;
242
243         hstate = hstate_vma(vma);
244
245         return 1UL << (hstate->order + PAGE_SHIFT);
246 }
247 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
248
249 /*
250  * Return the page size being used by the MMU to back a VMA. In the majority
251  * of cases, the page size used by the kernel matches the MMU size. On
252  * architectures where it differs, an architecture-specific version of this
253  * function is required.
254  */
255 #ifndef vma_mmu_pagesize
256 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
257 {
258         return vma_kernel_pagesize(vma);
259 }
260 #endif
261
262 /*
263  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
264  * bits of the reservation map pointer, which are always clear due to
265  * alignment.
266  */
267 #define HPAGE_RESV_OWNER    (1UL << 0)
268 #define HPAGE_RESV_UNMAPPED (1UL << 1)
269 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
270
271 /*
272  * These helpers are used to track how many pages are reserved for
273  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
274  * is guaranteed to have their future faults succeed.
275  *
276  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
277  * the reserve counters are updated with the hugetlb_lock held. It is safe
278  * to reset the VMA at fork() time as it is not in use yet and there is no
279  * chance of the global counters getting corrupted as a result of the values.
280  *
281  * The private mapping reservation is represented in a subtly different
282  * manner to a shared mapping.  A shared mapping has a region map associated
283  * with the underlying file, this region map represents the backing file
284  * pages which have ever had a reservation assigned which this persists even
285  * after the page is instantiated.  A private mapping has a region map
286  * associated with the original mmap which is attached to all VMAs which
287  * reference it, this region map represents those offsets which have consumed
288  * reservation ie. where pages have been instantiated.
289  */
290 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
291 {
292         return (unsigned long)vma->vm_private_data;
293 }
294
295 static void set_vma_private_data(struct vm_area_struct *vma,
296                                                         unsigned long value)
297 {
298         vma->vm_private_data = (void *)value;
299 }
300
301 struct resv_map {
302         struct kref refs;
303         struct list_head regions;
304 };
305
306 static struct resv_map *resv_map_alloc(void)
307 {
308         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
309         if (!resv_map)
310                 return NULL;
311
312         kref_init(&resv_map->refs);
313         INIT_LIST_HEAD(&resv_map->regions);
314
315         return resv_map;
316 }
317
318 static void resv_map_release(struct kref *ref)
319 {
320         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
321
322         /* Clear out any active regions before we release the map. */
323         region_truncate(&resv_map->regions, 0);
324         kfree(resv_map);
325 }
326
327 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
328 {
329         VM_BUG_ON(!is_vm_hugetlb_page(vma));
330         if (!(vma->vm_flags & VM_MAYSHARE))
331                 return (struct resv_map *)(get_vma_private_data(vma) &
332                                                         ~HPAGE_RESV_MASK);
333         return NULL;
334 }
335
336 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
337 {
338         VM_BUG_ON(!is_vm_hugetlb_page(vma));
339         VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
340
341         set_vma_private_data(vma, (get_vma_private_data(vma) &
342                                 HPAGE_RESV_MASK) | (unsigned long)map);
343 }
344
345 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
346 {
347         VM_BUG_ON(!is_vm_hugetlb_page(vma));
348         VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
349
350         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
351 }
352
353 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
354 {
355         VM_BUG_ON(!is_vm_hugetlb_page(vma));
356
357         return (get_vma_private_data(vma) & flag) != 0;
358 }
359
360 /* Decrement the reserved pages in the hugepage pool by one */
361 static void decrement_hugepage_resv_vma(struct hstate *h,
362                         struct vm_area_struct *vma)
363 {
364         if (vma->vm_flags & VM_NORESERVE)
365                 return;
366
367         if (vma->vm_flags & VM_MAYSHARE) {
368                 /* Shared mappings always use reserves */
369                 h->resv_huge_pages--;
370         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
371                 /*
372                  * Only the process that called mmap() has reserves for
373                  * private mappings.
374                  */
375                 h->resv_huge_pages--;
376         }
377 }
378
379 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
380 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
381 {
382         VM_BUG_ON(!is_vm_hugetlb_page(vma));
383         if (!(vma->vm_flags & VM_MAYSHARE))
384                 vma->vm_private_data = (void *)0;
385 }
386
387 /* Returns true if the VMA has associated reserve pages */
388 static int vma_has_reserves(struct vm_area_struct *vma)
389 {
390         if (vma->vm_flags & VM_MAYSHARE)
391                 return 1;
392         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
393                 return 1;
394         return 0;
395 }
396
397 static void clear_gigantic_page(struct page *page,
398                         unsigned long addr, unsigned long sz)
399 {
400         int i;
401         struct page *p = page;
402
403         might_sleep();
404         for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) {
405                 cond_resched();
406                 clear_user_highpage(p, addr + i * PAGE_SIZE);
407         }
408 }
409 static void clear_huge_page(struct page *page,
410                         unsigned long addr, unsigned long sz)
411 {
412         int i;
413
414         if (unlikely(sz/PAGE_SIZE > MAX_ORDER_NR_PAGES)) {
415                 clear_gigantic_page(page, addr, sz);
416                 return;
417         }
418
419         might_sleep();
420         for (i = 0; i < sz/PAGE_SIZE; i++) {
421                 cond_resched();
422                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
423         }
424 }
425
426 static void copy_gigantic_page(struct page *dst, struct page *src,
427                            unsigned long addr, struct vm_area_struct *vma)
428 {
429         int i;
430         struct hstate *h = hstate_vma(vma);
431         struct page *dst_base = dst;
432         struct page *src_base = src;
433         might_sleep();
434         for (i = 0; i < pages_per_huge_page(h); ) {
435                 cond_resched();
436                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
437
438                 i++;
439                 dst = mem_map_next(dst, dst_base, i);
440                 src = mem_map_next(src, src_base, i);
441         }
442 }
443 static void copy_huge_page(struct page *dst, struct page *src,
444                            unsigned long addr, struct vm_area_struct *vma)
445 {
446         int i;
447         struct hstate *h = hstate_vma(vma);
448
449         if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
450                 copy_gigantic_page(dst, src, addr, vma);
451                 return;
452         }
453
454         might_sleep();
455         for (i = 0; i < pages_per_huge_page(h); i++) {
456                 cond_resched();
457                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
458         }
459 }
460
461 static void enqueue_huge_page(struct hstate *h, struct page *page)
462 {
463         int nid = page_to_nid(page);
464         list_add(&page->lru, &h->hugepage_freelists[nid]);
465         h->free_huge_pages++;
466         h->free_huge_pages_node[nid]++;
467 }
468
469 static struct page *dequeue_huge_page_vma(struct hstate *h,
470                                 struct vm_area_struct *vma,
471                                 unsigned long address, int avoid_reserve)
472 {
473         int nid;
474         struct page *page = NULL;
475         struct mempolicy *mpol;
476         nodemask_t *nodemask;
477         struct zonelist *zonelist;
478         struct zone *zone;
479         struct zoneref *z;
480
481         get_mems_allowed();
482         zonelist = huge_zonelist(vma, address,
483                                         htlb_alloc_mask, &mpol, &nodemask);
484         /*
485          * A child process with MAP_PRIVATE mappings created by their parent
486          * have no page reserves. This check ensures that reservations are
487          * not "stolen". The child may still get SIGKILLed
488          */
489         if (!vma_has_reserves(vma) &&
490                         h->free_huge_pages - h->resv_huge_pages == 0)
491                 goto err;
492
493         /* If reserves cannot be used, ensure enough pages are in the pool */
494         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
495                 goto err;;
496
497         for_each_zone_zonelist_nodemask(zone, z, zonelist,
498                                                 MAX_NR_ZONES - 1, nodemask) {
499                 nid = zone_to_nid(zone);
500                 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
501                     !list_empty(&h->hugepage_freelists[nid])) {
502                         page = list_entry(h->hugepage_freelists[nid].next,
503                                           struct page, lru);
504                         list_del(&page->lru);
505                         h->free_huge_pages--;
506                         h->free_huge_pages_node[nid]--;
507
508                         if (!avoid_reserve)
509                                 decrement_hugepage_resv_vma(h, vma);
510
511                         break;
512                 }
513         }
514 err:
515         mpol_cond_put(mpol);
516         put_mems_allowed();
517         return page;
518 }
519
520 static void update_and_free_page(struct hstate *h, struct page *page)
521 {
522         int i;
523
524         VM_BUG_ON(h->order >= MAX_ORDER);
525
526         h->nr_huge_pages--;
527         h->nr_huge_pages_node[page_to_nid(page)]--;
528         for (i = 0; i < pages_per_huge_page(h); i++) {
529                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
530                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
531                                 1 << PG_private | 1<< PG_writeback);
532         }
533         set_compound_page_dtor(page, NULL);
534         set_page_refcounted(page);
535         arch_release_hugepage(page);
536         __free_pages(page, huge_page_order(h));
537 }
538
539 struct hstate *size_to_hstate(unsigned long size)
540 {
541         struct hstate *h;
542
543         for_each_hstate(h) {
544                 if (huge_page_size(h) == size)
545                         return h;
546         }
547         return NULL;
548 }
549
550 static void free_huge_page(struct page *page)
551 {
552         /*
553          * Can't pass hstate in here because it is called from the
554          * compound page destructor.
555          */
556         struct hstate *h = page_hstate(page);
557         int nid = page_to_nid(page);
558         struct address_space *mapping;
559
560         mapping = (struct address_space *) page_private(page);
561         set_page_private(page, 0);
562         page->mapping = NULL;
563         BUG_ON(page_count(page));
564         BUG_ON(page_mapcount(page));
565         INIT_LIST_HEAD(&page->lru);
566
567         spin_lock(&hugetlb_lock);
568         if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
569                 update_and_free_page(h, page);
570                 h->surplus_huge_pages--;
571                 h->surplus_huge_pages_node[nid]--;
572         } else {
573                 enqueue_huge_page(h, page);
574         }
575         spin_unlock(&hugetlb_lock);
576         if (mapping)
577                 hugetlb_put_quota(mapping, 1);
578 }
579
580 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
581 {
582         set_compound_page_dtor(page, free_huge_page);
583         spin_lock(&hugetlb_lock);
584         h->nr_huge_pages++;
585         h->nr_huge_pages_node[nid]++;
586         spin_unlock(&hugetlb_lock);
587         put_page(page); /* free it into the hugepage allocator */
588 }
589
590 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
591 {
592         int i;
593         int nr_pages = 1 << order;
594         struct page *p = page + 1;
595
596         /* we rely on prep_new_huge_page to set the destructor */
597         set_compound_order(page, order);
598         __SetPageHead(page);
599         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
600                 __SetPageTail(p);
601                 p->first_page = page;
602         }
603 }
604
605 int PageHuge(struct page *page)
606 {
607         compound_page_dtor *dtor;
608
609         if (!PageCompound(page))
610                 return 0;
611
612         page = compound_head(page);
613         dtor = get_compound_page_dtor(page);
614
615         return dtor == free_huge_page;
616 }
617
618 EXPORT_SYMBOL_GPL(PageHuge);
619
620 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
621 {
622         struct page *page;
623
624         if (h->order >= MAX_ORDER)
625                 return NULL;
626
627         page = alloc_pages_exact_node(nid,
628                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
629                                                 __GFP_REPEAT|__GFP_NOWARN,
630                 huge_page_order(h));
631         if (page) {
632                 if (arch_prepare_hugepage(page)) {
633                         __free_pages(page, huge_page_order(h));
634                         return NULL;
635                 }
636                 prep_new_huge_page(h, page, nid);
637         }
638
639         return page;
640 }
641
642 /*
643  * common helper functions for hstate_next_node_to_{alloc|free}.
644  * We may have allocated or freed a huge page based on a different
645  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
646  * be outside of *nodes_allowed.  Ensure that we use an allowed
647  * node for alloc or free.
648  */
649 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
650 {
651         nid = next_node(nid, *nodes_allowed);
652         if (nid == MAX_NUMNODES)
653                 nid = first_node(*nodes_allowed);
654         VM_BUG_ON(nid >= MAX_NUMNODES);
655
656         return nid;
657 }
658
659 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
660 {
661         if (!node_isset(nid, *nodes_allowed))
662                 nid = next_node_allowed(nid, nodes_allowed);
663         return nid;
664 }
665
666 /*
667  * returns the previously saved node ["this node"] from which to
668  * allocate a persistent huge page for the pool and advance the
669  * next node from which to allocate, handling wrap at end of node
670  * mask.
671  */
672 static int hstate_next_node_to_alloc(struct hstate *h,
673                                         nodemask_t *nodes_allowed)
674 {
675         int nid;
676
677         VM_BUG_ON(!nodes_allowed);
678
679         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
680         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
681
682         return nid;
683 }
684
685 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
686 {
687         struct page *page;
688         int start_nid;
689         int next_nid;
690         int ret = 0;
691
692         start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
693         next_nid = start_nid;
694
695         do {
696                 page = alloc_fresh_huge_page_node(h, next_nid);
697                 if (page) {
698                         ret = 1;
699                         break;
700                 }
701                 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
702         } while (next_nid != start_nid);
703
704         if (ret)
705                 count_vm_event(HTLB_BUDDY_PGALLOC);
706         else
707                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
708
709         return ret;
710 }
711
712 /*
713  * helper for free_pool_huge_page() - return the previously saved
714  * node ["this node"] from which to free a huge page.  Advance the
715  * next node id whether or not we find a free huge page to free so
716  * that the next attempt to free addresses the next node.
717  */
718 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
719 {
720         int nid;
721
722         VM_BUG_ON(!nodes_allowed);
723
724         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
725         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
726
727         return nid;
728 }
729
730 /*
731  * Free huge page from pool from next node to free.
732  * Attempt to keep persistent huge pages more or less
733  * balanced over allowed nodes.
734  * Called with hugetlb_lock locked.
735  */
736 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
737                                                          bool acct_surplus)
738 {
739         int start_nid;
740         int next_nid;
741         int ret = 0;
742
743         start_nid = hstate_next_node_to_free(h, nodes_allowed);
744         next_nid = start_nid;
745
746         do {
747                 /*
748                  * If we're returning unused surplus pages, only examine
749                  * nodes with surplus pages.
750                  */
751                 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
752                     !list_empty(&h->hugepage_freelists[next_nid])) {
753                         struct page *page =
754                                 list_entry(h->hugepage_freelists[next_nid].next,
755                                           struct page, lru);
756                         list_del(&page->lru);
757                         h->free_huge_pages--;
758                         h->free_huge_pages_node[next_nid]--;
759                         if (acct_surplus) {
760                                 h->surplus_huge_pages--;
761                                 h->surplus_huge_pages_node[next_nid]--;
762                         }
763                         update_and_free_page(h, page);
764                         ret = 1;
765                         break;
766                 }
767                 next_nid = hstate_next_node_to_free(h, nodes_allowed);
768         } while (next_nid != start_nid);
769
770         return ret;
771 }
772
773 static struct page *alloc_buddy_huge_page(struct hstate *h,
774                         struct vm_area_struct *vma, unsigned long address)
775 {
776         struct page *page;
777         unsigned int nid;
778
779         if (h->order >= MAX_ORDER)
780                 return NULL;
781
782         /*
783          * Assume we will successfully allocate the surplus page to
784          * prevent racing processes from causing the surplus to exceed
785          * overcommit
786          *
787          * This however introduces a different race, where a process B
788          * tries to grow the static hugepage pool while alloc_pages() is
789          * called by process A. B will only examine the per-node
790          * counters in determining if surplus huge pages can be
791          * converted to normal huge pages in adjust_pool_surplus(). A
792          * won't be able to increment the per-node counter, until the
793          * lock is dropped by B, but B doesn't drop hugetlb_lock until
794          * no more huge pages can be converted from surplus to normal
795          * state (and doesn't try to convert again). Thus, we have a
796          * case where a surplus huge page exists, the pool is grown, and
797          * the surplus huge page still exists after, even though it
798          * should just have been converted to a normal huge page. This
799          * does not leak memory, though, as the hugepage will be freed
800          * once it is out of use. It also does not allow the counters to
801          * go out of whack in adjust_pool_surplus() as we don't modify
802          * the node values until we've gotten the hugepage and only the
803          * per-node value is checked there.
804          */
805         spin_lock(&hugetlb_lock);
806         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
807                 spin_unlock(&hugetlb_lock);
808                 return NULL;
809         } else {
810                 h->nr_huge_pages++;
811                 h->surplus_huge_pages++;
812         }
813         spin_unlock(&hugetlb_lock);
814
815         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
816                                         __GFP_REPEAT|__GFP_NOWARN,
817                                         huge_page_order(h));
818
819         if (page && arch_prepare_hugepage(page)) {
820                 __free_pages(page, huge_page_order(h));
821                 return NULL;
822         }
823
824         spin_lock(&hugetlb_lock);
825         if (page) {
826                 /*
827                  * This page is now managed by the hugetlb allocator and has
828                  * no users -- drop the buddy allocator's reference.
829                  */
830                 put_page_testzero(page);
831                 VM_BUG_ON(page_count(page));
832                 nid = page_to_nid(page);
833                 set_compound_page_dtor(page, free_huge_page);
834                 /*
835                  * We incremented the global counters already
836                  */
837                 h->nr_huge_pages_node[nid]++;
838                 h->surplus_huge_pages_node[nid]++;
839                 __count_vm_event(HTLB_BUDDY_PGALLOC);
840         } else {
841                 h->nr_huge_pages--;
842                 h->surplus_huge_pages--;
843                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
844         }
845         spin_unlock(&hugetlb_lock);
846
847         return page;
848 }
849
850 /*
851  * Increase the hugetlb pool such that it can accomodate a reservation
852  * of size 'delta'.
853  */
854 static int gather_surplus_pages(struct hstate *h, int delta)
855 {
856         struct list_head surplus_list;
857         struct page *page, *tmp;
858         int ret, i;
859         int needed, allocated;
860
861         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
862         if (needed <= 0) {
863                 h->resv_huge_pages += delta;
864                 return 0;
865         }
866
867         allocated = 0;
868         INIT_LIST_HEAD(&surplus_list);
869
870         ret = -ENOMEM;
871 retry:
872         spin_unlock(&hugetlb_lock);
873         for (i = 0; i < needed; i++) {
874                 page = alloc_buddy_huge_page(h, NULL, 0);
875                 if (!page) {
876                         /*
877                          * We were not able to allocate enough pages to
878                          * satisfy the entire reservation so we free what
879                          * we've allocated so far.
880                          */
881                         spin_lock(&hugetlb_lock);
882                         needed = 0;
883                         goto free;
884                 }
885
886                 list_add(&page->lru, &surplus_list);
887         }
888         allocated += needed;
889
890         /*
891          * After retaking hugetlb_lock, we need to recalculate 'needed'
892          * because either resv_huge_pages or free_huge_pages may have changed.
893          */
894         spin_lock(&hugetlb_lock);
895         needed = (h->resv_huge_pages + delta) -
896                         (h->free_huge_pages + allocated);
897         if (needed > 0)
898                 goto retry;
899
900         /*
901          * The surplus_list now contains _at_least_ the number of extra pages
902          * needed to accomodate the reservation.  Add the appropriate number
903          * of pages to the hugetlb pool and free the extras back to the buddy
904          * allocator.  Commit the entire reservation here to prevent another
905          * process from stealing the pages as they are added to the pool but
906          * before they are reserved.
907          */
908         needed += allocated;
909         h->resv_huge_pages += delta;
910         ret = 0;
911 free:
912         /* Free the needed pages to the hugetlb pool */
913         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
914                 if ((--needed) < 0)
915                         break;
916                 list_del(&page->lru);
917                 enqueue_huge_page(h, page);
918         }
919
920         /* Free unnecessary surplus pages to the buddy allocator */
921         if (!list_empty(&surplus_list)) {
922                 spin_unlock(&hugetlb_lock);
923                 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
924                         list_del(&page->lru);
925                         /*
926                          * The page has a reference count of zero already, so
927                          * call free_huge_page directly instead of using
928                          * put_page.  This must be done with hugetlb_lock
929                          * unlocked which is safe because free_huge_page takes
930                          * hugetlb_lock before deciding how to free the page.
931                          */
932                         free_huge_page(page);
933                 }
934                 spin_lock(&hugetlb_lock);
935         }
936
937         return ret;
938 }
939
940 /*
941  * When releasing a hugetlb pool reservation, any surplus pages that were
942  * allocated to satisfy the reservation must be explicitly freed if they were
943  * never used.
944  * Called with hugetlb_lock held.
945  */
946 static void return_unused_surplus_pages(struct hstate *h,
947                                         unsigned long unused_resv_pages)
948 {
949         unsigned long nr_pages;
950
951         /* Uncommit the reservation */
952         h->resv_huge_pages -= unused_resv_pages;
953
954         /* Cannot return gigantic pages currently */
955         if (h->order >= MAX_ORDER)
956                 return;
957
958         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
959
960         /*
961          * We want to release as many surplus pages as possible, spread
962          * evenly across all nodes with memory. Iterate across these nodes
963          * until we can no longer free unreserved surplus pages. This occurs
964          * when the nodes with surplus pages have no free pages.
965          * free_pool_huge_page() will balance the the freed pages across the
966          * on-line nodes with memory and will handle the hstate accounting.
967          */
968         while (nr_pages--) {
969                 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
970                         break;
971         }
972 }
973
974 /*
975  * Determine if the huge page at addr within the vma has an associated
976  * reservation.  Where it does not we will need to logically increase
977  * reservation and actually increase quota before an allocation can occur.
978  * Where any new reservation would be required the reservation change is
979  * prepared, but not committed.  Once the page has been quota'd allocated
980  * an instantiated the change should be committed via vma_commit_reservation.
981  * No action is required on failure.
982  */
983 static long vma_needs_reservation(struct hstate *h,
984                         struct vm_area_struct *vma, unsigned long addr)
985 {
986         struct address_space *mapping = vma->vm_file->f_mapping;
987         struct inode *inode = mapping->host;
988
989         if (vma->vm_flags & VM_MAYSHARE) {
990                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
991                 return region_chg(&inode->i_mapping->private_list,
992                                                         idx, idx + 1);
993
994         } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
995                 return 1;
996
997         } else  {
998                 long err;
999                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1000                 struct resv_map *reservations = vma_resv_map(vma);
1001
1002                 err = region_chg(&reservations->regions, idx, idx + 1);
1003                 if (err < 0)
1004                         return err;
1005                 return 0;
1006         }
1007 }
1008 static void vma_commit_reservation(struct hstate *h,
1009                         struct vm_area_struct *vma, unsigned long addr)
1010 {
1011         struct address_space *mapping = vma->vm_file->f_mapping;
1012         struct inode *inode = mapping->host;
1013
1014         if (vma->vm_flags & VM_MAYSHARE) {
1015                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1016                 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1017
1018         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1019                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1020                 struct resv_map *reservations = vma_resv_map(vma);
1021
1022                 /* Mark this page used in the map. */
1023                 region_add(&reservations->regions, idx, idx + 1);
1024         }
1025 }
1026
1027 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1028                                     unsigned long addr, int avoid_reserve)
1029 {
1030         struct hstate *h = hstate_vma(vma);
1031         struct page *page;
1032         struct address_space *mapping = vma->vm_file->f_mapping;
1033         struct inode *inode = mapping->host;
1034         long chg;
1035
1036         /*
1037          * Processes that did not create the mapping will have no reserves and
1038          * will not have accounted against quota. Check that the quota can be
1039          * made before satisfying the allocation
1040          * MAP_NORESERVE mappings may also need pages and quota allocated
1041          * if no reserve mapping overlaps.
1042          */
1043         chg = vma_needs_reservation(h, vma, addr);
1044         if (chg < 0)
1045                 return ERR_PTR(chg);
1046         if (chg)
1047                 if (hugetlb_get_quota(inode->i_mapping, chg))
1048                         return ERR_PTR(-ENOSPC);
1049
1050         spin_lock(&hugetlb_lock);
1051         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1052         spin_unlock(&hugetlb_lock);
1053
1054         if (!page) {
1055                 page = alloc_buddy_huge_page(h, vma, addr);
1056                 if (!page) {
1057                         hugetlb_put_quota(inode->i_mapping, chg);
1058                         return ERR_PTR(-VM_FAULT_SIGBUS);
1059                 }
1060         }
1061
1062         set_page_refcounted(page);
1063         set_page_private(page, (unsigned long) mapping);
1064
1065         vma_commit_reservation(h, vma, addr);
1066
1067         return page;
1068 }
1069
1070 int __weak alloc_bootmem_huge_page(struct hstate *h)
1071 {
1072         struct huge_bootmem_page *m;
1073         int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1074
1075         while (nr_nodes) {
1076                 void *addr;
1077
1078                 addr = __alloc_bootmem_node_nopanic(
1079                                 NODE_DATA(hstate_next_node_to_alloc(h,
1080                                                 &node_states[N_HIGH_MEMORY])),
1081                                 huge_page_size(h), huge_page_size(h), 0);
1082
1083                 if (addr) {
1084                         /*
1085                          * Use the beginning of the huge page to store the
1086                          * huge_bootmem_page struct (until gather_bootmem
1087                          * puts them into the mem_map).
1088                          */
1089                         m = addr;
1090                         goto found;
1091                 }
1092                 nr_nodes--;
1093         }
1094         return 0;
1095
1096 found:
1097         BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1098         /* Put them into a private list first because mem_map is not up yet */
1099         list_add(&m->list, &huge_boot_pages);
1100         m->hstate = h;
1101         return 1;
1102 }
1103
1104 static void prep_compound_huge_page(struct page *page, int order)
1105 {
1106         if (unlikely(order > (MAX_ORDER - 1)))
1107                 prep_compound_gigantic_page(page, order);
1108         else
1109                 prep_compound_page(page, order);
1110 }
1111
1112 /* Put bootmem huge pages into the standard lists after mem_map is up */
1113 static void __init gather_bootmem_prealloc(void)
1114 {
1115         struct huge_bootmem_page *m;
1116
1117         list_for_each_entry(m, &huge_boot_pages, list) {
1118                 struct page *page = virt_to_page(m);
1119                 struct hstate *h = m->hstate;
1120                 __ClearPageReserved(page);
1121                 WARN_ON(page_count(page) != 1);
1122                 prep_compound_huge_page(page, h->order);
1123                 prep_new_huge_page(h, page, page_to_nid(page));
1124         }
1125 }
1126
1127 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1128 {
1129         unsigned long i;
1130
1131         for (i = 0; i < h->max_huge_pages; ++i) {
1132                 if (h->order >= MAX_ORDER) {
1133                         if (!alloc_bootmem_huge_page(h))
1134                                 break;
1135                 } else if (!alloc_fresh_huge_page(h,
1136                                          &node_states[N_HIGH_MEMORY]))
1137                         break;
1138         }
1139         h->max_huge_pages = i;
1140 }
1141
1142 static void __init hugetlb_init_hstates(void)
1143 {
1144         struct hstate *h;
1145
1146         for_each_hstate(h) {
1147                 /* oversize hugepages were init'ed in early boot */
1148                 if (h->order < MAX_ORDER)
1149                         hugetlb_hstate_alloc_pages(h);
1150         }
1151 }
1152
1153 static char * __init memfmt(char *buf, unsigned long n)
1154 {
1155         if (n >= (1UL << 30))
1156                 sprintf(buf, "%lu GB", n >> 30);
1157         else if (n >= (1UL << 20))
1158                 sprintf(buf, "%lu MB", n >> 20);
1159         else
1160                 sprintf(buf, "%lu KB", n >> 10);
1161         return buf;
1162 }
1163
1164 static void __init report_hugepages(void)
1165 {
1166         struct hstate *h;
1167
1168         for_each_hstate(h) {
1169                 char buf[32];
1170                 printk(KERN_INFO "HugeTLB registered %s page size, "
1171                                  "pre-allocated %ld pages\n",
1172                         memfmt(buf, huge_page_size(h)),
1173                         h->free_huge_pages);
1174         }
1175 }
1176
1177 #ifdef CONFIG_HIGHMEM
1178 static void try_to_free_low(struct hstate *h, unsigned long count,
1179                                                 nodemask_t *nodes_allowed)
1180 {
1181         int i;
1182
1183         if (h->order >= MAX_ORDER)
1184                 return;
1185
1186         for_each_node_mask(i, *nodes_allowed) {
1187                 struct page *page, *next;
1188                 struct list_head *freel = &h->hugepage_freelists[i];
1189                 list_for_each_entry_safe(page, next, freel, lru) {
1190                         if (count >= h->nr_huge_pages)
1191                                 return;
1192                         if (PageHighMem(page))
1193                                 continue;
1194                         list_del(&page->lru);
1195                         update_and_free_page(h, page);
1196                         h->free_huge_pages--;
1197                         h->free_huge_pages_node[page_to_nid(page)]--;
1198                 }
1199         }
1200 }
1201 #else
1202 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1203                                                 nodemask_t *nodes_allowed)
1204 {
1205 }
1206 #endif
1207
1208 /*
1209  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
1210  * balanced by operating on them in a round-robin fashion.
1211  * Returns 1 if an adjustment was made.
1212  */
1213 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1214                                 int delta)
1215 {
1216         int start_nid, next_nid;
1217         int ret = 0;
1218
1219         VM_BUG_ON(delta != -1 && delta != 1);
1220
1221         if (delta < 0)
1222                 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1223         else
1224                 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1225         next_nid = start_nid;
1226
1227         do {
1228                 int nid = next_nid;
1229                 if (delta < 0)  {
1230                         /*
1231                          * To shrink on this node, there must be a surplus page
1232                          */
1233                         if (!h->surplus_huge_pages_node[nid]) {
1234                                 next_nid = hstate_next_node_to_alloc(h,
1235                                                                 nodes_allowed);
1236                                 continue;
1237                         }
1238                 }
1239                 if (delta > 0) {
1240                         /*
1241                          * Surplus cannot exceed the total number of pages
1242                          */
1243                         if (h->surplus_huge_pages_node[nid] >=
1244                                                 h->nr_huge_pages_node[nid]) {
1245                                 next_nid = hstate_next_node_to_free(h,
1246                                                                 nodes_allowed);
1247                                 continue;
1248                         }
1249                 }
1250
1251                 h->surplus_huge_pages += delta;
1252                 h->surplus_huge_pages_node[nid] += delta;
1253                 ret = 1;
1254                 break;
1255         } while (next_nid != start_nid);
1256
1257         return ret;
1258 }
1259
1260 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1261 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1262                                                 nodemask_t *nodes_allowed)
1263 {
1264         unsigned long min_count, ret;
1265
1266         if (h->order >= MAX_ORDER)
1267                 return h->max_huge_pages;
1268
1269         /*
1270          * Increase the pool size
1271          * First take pages out of surplus state.  Then make up the
1272          * remaining difference by allocating fresh huge pages.
1273          *
1274          * We might race with alloc_buddy_huge_page() here and be unable
1275          * to convert a surplus huge page to a normal huge page. That is
1276          * not critical, though, it just means the overall size of the
1277          * pool might be one hugepage larger than it needs to be, but
1278          * within all the constraints specified by the sysctls.
1279          */
1280         spin_lock(&hugetlb_lock);
1281         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1282                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1283                         break;
1284         }
1285
1286         while (count > persistent_huge_pages(h)) {
1287                 /*
1288                  * If this allocation races such that we no longer need the
1289                  * page, free_huge_page will handle it by freeing the page
1290                  * and reducing the surplus.
1291                  */
1292                 spin_unlock(&hugetlb_lock);
1293                 ret = alloc_fresh_huge_page(h, nodes_allowed);
1294                 spin_lock(&hugetlb_lock);
1295                 if (!ret)
1296                         goto out;
1297
1298                 /* Bail for signals. Probably ctrl-c from user */
1299                 if (signal_pending(current))
1300                         goto out;
1301         }
1302
1303         /*
1304          * Decrease the pool size
1305          * First return free pages to the buddy allocator (being careful
1306          * to keep enough around to satisfy reservations).  Then place
1307          * pages into surplus state as needed so the pool will shrink
1308          * to the desired size as pages become free.
1309          *
1310          * By placing pages into the surplus state independent of the
1311          * overcommit value, we are allowing the surplus pool size to
1312          * exceed overcommit. There are few sane options here. Since
1313          * alloc_buddy_huge_page() is checking the global counter,
1314          * though, we'll note that we're not allowed to exceed surplus
1315          * and won't grow the pool anywhere else. Not until one of the
1316          * sysctls are changed, or the surplus pages go out of use.
1317          */
1318         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1319         min_count = max(count, min_count);
1320         try_to_free_low(h, min_count, nodes_allowed);
1321         while (min_count < persistent_huge_pages(h)) {
1322                 if (!free_pool_huge_page(h, nodes_allowed, 0))
1323                         break;
1324         }
1325         while (count < persistent_huge_pages(h)) {
1326                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1327                         break;
1328         }
1329 out:
1330         ret = persistent_huge_pages(h);
1331         spin_unlock(&hugetlb_lock);
1332         return ret;
1333 }
1334
1335 #define HSTATE_ATTR_RO(_name) \
1336         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1337
1338 #define HSTATE_ATTR(_name) \
1339         static struct kobj_attribute _name##_attr = \
1340                 __ATTR(_name, 0644, _name##_show, _name##_store)
1341
1342 static struct kobject *hugepages_kobj;
1343 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1344
1345 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1346
1347 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1348 {
1349         int i;
1350
1351         for (i = 0; i < HUGE_MAX_HSTATE; i++)
1352                 if (hstate_kobjs[i] == kobj) {
1353                         if (nidp)
1354                                 *nidp = NUMA_NO_NODE;
1355                         return &hstates[i];
1356                 }
1357
1358         return kobj_to_node_hstate(kobj, nidp);
1359 }
1360
1361 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1362                                         struct kobj_attribute *attr, char *buf)
1363 {
1364         struct hstate *h;
1365         unsigned long nr_huge_pages;
1366         int nid;
1367
1368         h = kobj_to_hstate(kobj, &nid);
1369         if (nid == NUMA_NO_NODE)
1370                 nr_huge_pages = h->nr_huge_pages;
1371         else
1372                 nr_huge_pages = h->nr_huge_pages_node[nid];
1373
1374         return sprintf(buf, "%lu\n", nr_huge_pages);
1375 }
1376 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1377                         struct kobject *kobj, struct kobj_attribute *attr,
1378                         const char *buf, size_t len)
1379 {
1380         int err;
1381         int nid;
1382         unsigned long count;
1383         struct hstate *h;
1384         NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1385
1386         err = strict_strtoul(buf, 10, &count);
1387         if (err)
1388                 return 0;
1389
1390         h = kobj_to_hstate(kobj, &nid);
1391         if (nid == NUMA_NO_NODE) {
1392                 /*
1393                  * global hstate attribute
1394                  */
1395                 if (!(obey_mempolicy &&
1396                                 init_nodemask_of_mempolicy(nodes_allowed))) {
1397                         NODEMASK_FREE(nodes_allowed);
1398                         nodes_allowed = &node_states[N_HIGH_MEMORY];
1399                 }
1400         } else if (nodes_allowed) {
1401                 /*
1402                  * per node hstate attribute: adjust count to global,
1403                  * but restrict alloc/free to the specified node.
1404                  */
1405                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1406                 init_nodemask_of_node(nodes_allowed, nid);
1407         } else
1408                 nodes_allowed = &node_states[N_HIGH_MEMORY];
1409
1410         h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1411
1412         if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1413                 NODEMASK_FREE(nodes_allowed);
1414
1415         return len;
1416 }
1417
1418 static ssize_t nr_hugepages_show(struct kobject *kobj,
1419                                        struct kobj_attribute *attr, char *buf)
1420 {
1421         return nr_hugepages_show_common(kobj, attr, buf);
1422 }
1423
1424 static ssize_t nr_hugepages_store(struct kobject *kobj,
1425                struct kobj_attribute *attr, const char *buf, size_t len)
1426 {
1427         return nr_hugepages_store_common(false, kobj, attr, buf, len);
1428 }
1429 HSTATE_ATTR(nr_hugepages);
1430
1431 #ifdef CONFIG_NUMA
1432
1433 /*
1434  * hstate attribute for optionally mempolicy-based constraint on persistent
1435  * huge page alloc/free.
1436  */
1437 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1438                                        struct kobj_attribute *attr, char *buf)
1439 {
1440         return nr_hugepages_show_common(kobj, attr, buf);
1441 }
1442
1443 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1444                struct kobj_attribute *attr, const char *buf, size_t len)
1445 {
1446         return nr_hugepages_store_common(true, kobj, attr, buf, len);
1447 }
1448 HSTATE_ATTR(nr_hugepages_mempolicy);
1449 #endif
1450
1451
1452 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1453                                         struct kobj_attribute *attr, char *buf)
1454 {
1455         struct hstate *h = kobj_to_hstate(kobj, NULL);
1456         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1457 }
1458 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1459                 struct kobj_attribute *attr, const char *buf, size_t count)
1460 {
1461         int err;
1462         unsigned long input;
1463         struct hstate *h = kobj_to_hstate(kobj, NULL);
1464
1465         err = strict_strtoul(buf, 10, &input);
1466         if (err)
1467                 return 0;
1468
1469         spin_lock(&hugetlb_lock);
1470         h->nr_overcommit_huge_pages = input;
1471         spin_unlock(&hugetlb_lock);
1472
1473         return count;
1474 }
1475 HSTATE_ATTR(nr_overcommit_hugepages);
1476
1477 static ssize_t free_hugepages_show(struct kobject *kobj,
1478                                         struct kobj_attribute *attr, char *buf)
1479 {
1480         struct hstate *h;
1481         unsigned long free_huge_pages;
1482         int nid;
1483
1484         h = kobj_to_hstate(kobj, &nid);
1485         if (nid == NUMA_NO_NODE)
1486                 free_huge_pages = h->free_huge_pages;
1487         else
1488                 free_huge_pages = h->free_huge_pages_node[nid];
1489
1490         return sprintf(buf, "%lu\n", free_huge_pages);
1491 }
1492 HSTATE_ATTR_RO(free_hugepages);
1493
1494 static ssize_t resv_hugepages_show(struct kobject *kobj,
1495                                         struct kobj_attribute *attr, char *buf)
1496 {
1497         struct hstate *h = kobj_to_hstate(kobj, NULL);
1498         return sprintf(buf, "%lu\n", h->resv_huge_pages);
1499 }
1500 HSTATE_ATTR_RO(resv_hugepages);
1501
1502 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1503                                         struct kobj_attribute *attr, char *buf)
1504 {
1505         struct hstate *h;
1506         unsigned long surplus_huge_pages;
1507         int nid;
1508
1509         h = kobj_to_hstate(kobj, &nid);
1510         if (nid == NUMA_NO_NODE)
1511                 surplus_huge_pages = h->surplus_huge_pages;
1512         else
1513                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1514
1515         return sprintf(buf, "%lu\n", surplus_huge_pages);
1516 }
1517 HSTATE_ATTR_RO(surplus_hugepages);
1518
1519 static struct attribute *hstate_attrs[] = {
1520         &nr_hugepages_attr.attr,
1521         &nr_overcommit_hugepages_attr.attr,
1522         &free_hugepages_attr.attr,
1523         &resv_hugepages_attr.attr,
1524         &surplus_hugepages_attr.attr,
1525 #ifdef CONFIG_NUMA
1526         &nr_hugepages_mempolicy_attr.attr,
1527 #endif
1528         NULL,
1529 };
1530
1531 static struct attribute_group hstate_attr_group = {
1532         .attrs = hstate_attrs,
1533 };
1534
1535 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1536                                     struct kobject **hstate_kobjs,
1537                                     struct attribute_group *hstate_attr_group)
1538 {
1539         int retval;
1540         int hi = h - hstates;
1541
1542         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1543         if (!hstate_kobjs[hi])
1544                 return -ENOMEM;
1545
1546         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1547         if (retval)
1548                 kobject_put(hstate_kobjs[hi]);
1549
1550         return retval;
1551 }
1552
1553 static void __init hugetlb_sysfs_init(void)
1554 {
1555         struct hstate *h;
1556         int err;
1557
1558         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1559         if (!hugepages_kobj)
1560                 return;
1561
1562         for_each_hstate(h) {
1563                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1564                                          hstate_kobjs, &hstate_attr_group);
1565                 if (err)
1566                         printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1567                                                                 h->name);
1568         }
1569 }
1570
1571 #ifdef CONFIG_NUMA
1572
1573 /*
1574  * node_hstate/s - associate per node hstate attributes, via their kobjects,
1575  * with node sysdevs in node_devices[] using a parallel array.  The array
1576  * index of a node sysdev or _hstate == node id.
1577  * This is here to avoid any static dependency of the node sysdev driver, in
1578  * the base kernel, on the hugetlb module.
1579  */
1580 struct node_hstate {
1581         struct kobject          *hugepages_kobj;
1582         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
1583 };
1584 struct node_hstate node_hstates[MAX_NUMNODES];
1585
1586 /*
1587  * A subset of global hstate attributes for node sysdevs
1588  */
1589 static struct attribute *per_node_hstate_attrs[] = {
1590         &nr_hugepages_attr.attr,
1591         &free_hugepages_attr.attr,
1592         &surplus_hugepages_attr.attr,
1593         NULL,
1594 };
1595
1596 static struct attribute_group per_node_hstate_attr_group = {
1597         .attrs = per_node_hstate_attrs,
1598 };
1599
1600 /*
1601  * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj.
1602  * Returns node id via non-NULL nidp.
1603  */
1604 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1605 {
1606         int nid;
1607
1608         for (nid = 0; nid < nr_node_ids; nid++) {
1609                 struct node_hstate *nhs = &node_hstates[nid];
1610                 int i;
1611                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1612                         if (nhs->hstate_kobjs[i] == kobj) {
1613                                 if (nidp)
1614                                         *nidp = nid;
1615                                 return &hstates[i];
1616                         }
1617         }
1618
1619         BUG();
1620         return NULL;
1621 }
1622
1623 /*
1624  * Unregister hstate attributes from a single node sysdev.
1625  * No-op if no hstate attributes attached.
1626  */
1627 void hugetlb_unregister_node(struct node *node)
1628 {
1629         struct hstate *h;
1630         struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1631
1632         if (!nhs->hugepages_kobj)
1633                 return;         /* no hstate attributes */
1634
1635         for_each_hstate(h)
1636                 if (nhs->hstate_kobjs[h - hstates]) {
1637                         kobject_put(nhs->hstate_kobjs[h - hstates]);
1638                         nhs->hstate_kobjs[h - hstates] = NULL;
1639                 }
1640
1641         kobject_put(nhs->hugepages_kobj);
1642         nhs->hugepages_kobj = NULL;
1643 }
1644
1645 /*
1646  * hugetlb module exit:  unregister hstate attributes from node sysdevs
1647  * that have them.
1648  */
1649 static void hugetlb_unregister_all_nodes(void)
1650 {
1651         int nid;
1652
1653         /*
1654          * disable node sysdev registrations.
1655          */
1656         register_hugetlbfs_with_node(NULL, NULL);
1657
1658         /*
1659          * remove hstate attributes from any nodes that have them.
1660          */
1661         for (nid = 0; nid < nr_node_ids; nid++)
1662                 hugetlb_unregister_node(&node_devices[nid]);
1663 }
1664
1665 /*
1666  * Register hstate attributes for a single node sysdev.
1667  * No-op if attributes already registered.
1668  */
1669 void hugetlb_register_node(struct node *node)
1670 {
1671         struct hstate *h;
1672         struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1673         int err;
1674
1675         if (nhs->hugepages_kobj)
1676                 return;         /* already allocated */
1677
1678         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1679                                                         &node->sysdev.kobj);
1680         if (!nhs->hugepages_kobj)
1681                 return;
1682
1683         for_each_hstate(h) {
1684                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1685                                                 nhs->hstate_kobjs,
1686                                                 &per_node_hstate_attr_group);
1687                 if (err) {
1688                         printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1689                                         " for node %d\n",
1690                                                 h->name, node->sysdev.id);
1691                         hugetlb_unregister_node(node);
1692                         break;
1693                 }
1694         }
1695 }
1696
1697 /*
1698  * hugetlb init time:  register hstate attributes for all registered node
1699  * sysdevs of nodes that have memory.  All on-line nodes should have
1700  * registered their associated sysdev by this time.
1701  */
1702 static void hugetlb_register_all_nodes(void)
1703 {
1704         int nid;
1705
1706         for_each_node_state(nid, N_HIGH_MEMORY) {
1707                 struct node *node = &node_devices[nid];
1708                 if (node->sysdev.id == nid)
1709                         hugetlb_register_node(node);
1710         }
1711
1712         /*
1713          * Let the node sysdev driver know we're here so it can
1714          * [un]register hstate attributes on node hotplug.
1715          */
1716         register_hugetlbfs_with_node(hugetlb_register_node,
1717                                      hugetlb_unregister_node);
1718 }
1719 #else   /* !CONFIG_NUMA */
1720
1721 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1722 {
1723         BUG();
1724         if (nidp)
1725                 *nidp = -1;
1726         return NULL;
1727 }
1728
1729 static void hugetlb_unregister_all_nodes(void) { }
1730
1731 static void hugetlb_register_all_nodes(void) { }
1732
1733 #endif
1734
1735 static void __exit hugetlb_exit(void)
1736 {
1737         struct hstate *h;
1738
1739         hugetlb_unregister_all_nodes();
1740
1741         for_each_hstate(h) {
1742                 kobject_put(hstate_kobjs[h - hstates]);
1743         }
1744
1745         kobject_put(hugepages_kobj);
1746 }
1747 module_exit(hugetlb_exit);
1748
1749 static int __init hugetlb_init(void)
1750 {
1751         /* Some platform decide whether they support huge pages at boot
1752          * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1753          * there is no such support
1754          */
1755         if (HPAGE_SHIFT == 0)
1756                 return 0;
1757
1758         if (!size_to_hstate(default_hstate_size)) {
1759                 default_hstate_size = HPAGE_SIZE;
1760                 if (!size_to_hstate(default_hstate_size))
1761                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1762         }
1763         default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1764         if (default_hstate_max_huge_pages)
1765                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1766
1767         hugetlb_init_hstates();
1768
1769         gather_bootmem_prealloc();
1770
1771         report_hugepages();
1772
1773         hugetlb_sysfs_init();
1774
1775         hugetlb_register_all_nodes();
1776
1777         return 0;
1778 }
1779 module_init(hugetlb_init);
1780
1781 /* Should be called on processing a hugepagesz=... option */
1782 void __init hugetlb_add_hstate(unsigned order)
1783 {
1784         struct hstate *h;
1785         unsigned long i;
1786
1787         if (size_to_hstate(PAGE_SIZE << order)) {
1788                 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1789                 return;
1790         }
1791         BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1792         BUG_ON(order == 0);
1793         h = &hstates[max_hstate++];
1794         h->order = order;
1795         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1796         h->nr_huge_pages = 0;
1797         h->free_huge_pages = 0;
1798         for (i = 0; i < MAX_NUMNODES; ++i)
1799                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1800         h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1801         h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1802         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1803                                         huge_page_size(h)/1024);
1804
1805         parsed_hstate = h;
1806 }
1807
1808 static int __init hugetlb_nrpages_setup(char *s)
1809 {
1810         unsigned long *mhp;
1811         static unsigned long *last_mhp;
1812
1813         /*
1814          * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1815          * so this hugepages= parameter goes to the "default hstate".
1816          */
1817         if (!max_hstate)
1818                 mhp = &default_hstate_max_huge_pages;
1819         else
1820                 mhp = &parsed_hstate->max_huge_pages;
1821
1822         if (mhp == last_mhp) {
1823                 printk(KERN_WARNING "hugepages= specified twice without "
1824                         "interleaving hugepagesz=, ignoring\n");
1825                 return 1;
1826         }
1827
1828         if (sscanf(s, "%lu", mhp) <= 0)
1829                 *mhp = 0;
1830
1831         /*
1832          * Global state is always initialized later in hugetlb_init.
1833          * But we need to allocate >= MAX_ORDER hstates here early to still
1834          * use the bootmem allocator.
1835          */
1836         if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1837                 hugetlb_hstate_alloc_pages(parsed_hstate);
1838
1839         last_mhp = mhp;
1840
1841         return 1;
1842 }
1843 __setup("hugepages=", hugetlb_nrpages_setup);
1844
1845 static int __init hugetlb_default_setup(char *s)
1846 {
1847         default_hstate_size = memparse(s, &s);
1848         return 1;
1849 }
1850 __setup("default_hugepagesz=", hugetlb_default_setup);
1851
1852 static unsigned int cpuset_mems_nr(unsigned int *array)
1853 {
1854         int node;
1855         unsigned int nr = 0;
1856
1857         for_each_node_mask(node, cpuset_current_mems_allowed)
1858                 nr += array[node];
1859
1860         return nr;
1861 }
1862
1863 #ifdef CONFIG_SYSCTL
1864 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1865                          struct ctl_table *table, int write,
1866                          void __user *buffer, size_t *length, loff_t *ppos)
1867 {
1868         struct hstate *h = &default_hstate;
1869         unsigned long tmp;
1870
1871         if (!write)
1872                 tmp = h->max_huge_pages;
1873
1874         table->data = &tmp;
1875         table->maxlen = sizeof(unsigned long);
1876         proc_doulongvec_minmax(table, write, buffer, length, ppos);
1877
1878         if (write) {
1879                 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1880                                                 GFP_KERNEL | __GFP_NORETRY);
1881                 if (!(obey_mempolicy &&
1882                                init_nodemask_of_mempolicy(nodes_allowed))) {
1883                         NODEMASK_FREE(nodes_allowed);
1884                         nodes_allowed = &node_states[N_HIGH_MEMORY];
1885                 }
1886                 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1887
1888                 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1889                         NODEMASK_FREE(nodes_allowed);
1890         }
1891
1892         return 0;
1893 }
1894
1895 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1896                           void __user *buffer, size_t *length, loff_t *ppos)
1897 {
1898
1899         return hugetlb_sysctl_handler_common(false, table, write,
1900                                                         buffer, length, ppos);
1901 }
1902
1903 #ifdef CONFIG_NUMA
1904 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
1905                           void __user *buffer, size_t *length, loff_t *ppos)
1906 {
1907         return hugetlb_sysctl_handler_common(true, table, write,
1908                                                         buffer, length, ppos);
1909 }
1910 #endif /* CONFIG_NUMA */
1911
1912 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1913                         void __user *buffer,
1914                         size_t *length, loff_t *ppos)
1915 {
1916         proc_dointvec(table, write, buffer, length, ppos);
1917         if (hugepages_treat_as_movable)
1918                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1919         else
1920                 htlb_alloc_mask = GFP_HIGHUSER;
1921         return 0;
1922 }
1923
1924 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1925                         void __user *buffer,
1926                         size_t *length, loff_t *ppos)
1927 {
1928         struct hstate *h = &default_hstate;
1929         unsigned long tmp;
1930
1931         if (!write)
1932                 tmp = h->nr_overcommit_huge_pages;
1933
1934         table->data = &tmp;
1935         table->maxlen = sizeof(unsigned long);
1936         proc_doulongvec_minmax(table, write, buffer, length, ppos);
1937
1938         if (write) {
1939                 spin_lock(&hugetlb_lock);
1940                 h->nr_overcommit_huge_pages = tmp;
1941                 spin_unlock(&hugetlb_lock);
1942         }
1943
1944         return 0;
1945 }
1946
1947 #endif /* CONFIG_SYSCTL */
1948
1949 void hugetlb_report_meminfo(struct seq_file *m)
1950 {
1951         struct hstate *h = &default_hstate;
1952         seq_printf(m,
1953                         "HugePages_Total:   %5lu\n"
1954                         "HugePages_Free:    %5lu\n"
1955                         "HugePages_Rsvd:    %5lu\n"
1956                         "HugePages_Surp:    %5lu\n"
1957                         "Hugepagesize:   %8lu kB\n",
1958                         h->nr_huge_pages,
1959                         h->free_huge_pages,
1960                         h->resv_huge_pages,
1961                         h->surplus_huge_pages,
1962                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1963 }
1964
1965 int hugetlb_report_node_meminfo(int nid, char *buf)
1966 {
1967         struct hstate *h = &default_hstate;
1968         return sprintf(buf,
1969                 "Node %d HugePages_Total: %5u\n"
1970                 "Node %d HugePages_Free:  %5u\n"
1971                 "Node %d HugePages_Surp:  %5u\n",
1972                 nid, h->nr_huge_pages_node[nid],
1973                 nid, h->free_huge_pages_node[nid],
1974                 nid, h->surplus_huge_pages_node[nid]);
1975 }
1976
1977 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1978 unsigned long hugetlb_total_pages(void)
1979 {
1980         struct hstate *h = &default_hstate;
1981         return h->nr_huge_pages * pages_per_huge_page(h);
1982 }
1983
1984 static int hugetlb_acct_memory(struct hstate *h, long delta)
1985 {
1986         int ret = -ENOMEM;
1987
1988         spin_lock(&hugetlb_lock);
1989         /*
1990          * When cpuset is configured, it breaks the strict hugetlb page
1991          * reservation as the accounting is done on a global variable. Such
1992          * reservation is completely rubbish in the presence of cpuset because
1993          * the reservation is not checked against page availability for the
1994          * current cpuset. Application can still potentially OOM'ed by kernel
1995          * with lack of free htlb page in cpuset that the task is in.
1996          * Attempt to enforce strict accounting with cpuset is almost
1997          * impossible (or too ugly) because cpuset is too fluid that
1998          * task or memory node can be dynamically moved between cpusets.
1999          *
2000          * The change of semantics for shared hugetlb mapping with cpuset is
2001          * undesirable. However, in order to preserve some of the semantics,
2002          * we fall back to check against current free page availability as
2003          * a best attempt and hopefully to minimize the impact of changing
2004          * semantics that cpuset has.
2005          */
2006         if (delta > 0) {
2007                 if (gather_surplus_pages(h, delta) < 0)
2008                         goto out;
2009
2010                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2011                         return_unused_surplus_pages(h, delta);
2012                         goto out;
2013                 }
2014         }
2015
2016         ret = 0;
2017         if (delta < 0)
2018                 return_unused_surplus_pages(h, (unsigned long) -delta);
2019
2020 out:
2021         spin_unlock(&hugetlb_lock);
2022         return ret;
2023 }
2024
2025 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2026 {
2027         struct resv_map *reservations = vma_resv_map(vma);
2028
2029         /*
2030          * This new VMA should share its siblings reservation map if present.
2031          * The VMA will only ever have a valid reservation map pointer where
2032          * it is being copied for another still existing VMA.  As that VMA
2033          * has a reference to the reservation map it cannot dissappear until
2034          * after this open call completes.  It is therefore safe to take a
2035          * new reference here without additional locking.
2036          */
2037         if (reservations)
2038                 kref_get(&reservations->refs);
2039 }
2040
2041 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2042 {
2043         struct hstate *h = hstate_vma(vma);
2044         struct resv_map *reservations = vma_resv_map(vma);
2045         unsigned long reserve;
2046         unsigned long start;
2047         unsigned long end;
2048
2049         if (reservations) {
2050                 start = vma_hugecache_offset(h, vma, vma->vm_start);
2051                 end = vma_hugecache_offset(h, vma, vma->vm_end);
2052
2053                 reserve = (end - start) -
2054                         region_count(&reservations->regions, start, end);
2055
2056                 kref_put(&reservations->refs, resv_map_release);
2057
2058                 if (reserve) {
2059                         hugetlb_acct_memory(h, -reserve);
2060                         hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
2061                 }
2062         }
2063 }
2064
2065 /*
2066  * We cannot handle pagefaults against hugetlb pages at all.  They cause
2067  * handle_mm_fault() to try to instantiate regular-sized pages in the
2068  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
2069  * this far.
2070  */
2071 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2072 {
2073         BUG();
2074         return 0;
2075 }
2076
2077 const struct vm_operations_struct hugetlb_vm_ops = {
2078         .fault = hugetlb_vm_op_fault,
2079         .open = hugetlb_vm_op_open,
2080         .close = hugetlb_vm_op_close,
2081 };
2082
2083 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2084                                 int writable)
2085 {
2086         pte_t entry;
2087
2088         if (writable) {
2089                 entry =
2090                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2091         } else {
2092                 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2093         }
2094         entry = pte_mkyoung(entry);
2095         entry = pte_mkhuge(entry);
2096
2097         return entry;
2098 }
2099
2100 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2101                                    unsigned long address, pte_t *ptep)
2102 {
2103         pte_t entry;
2104
2105         entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2106         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
2107                 update_mmu_cache(vma, address, ptep);
2108         }
2109 }
2110
2111
2112 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2113                             struct vm_area_struct *vma)
2114 {
2115         pte_t *src_pte, *dst_pte, entry;
2116         struct page *ptepage;
2117         unsigned long addr;
2118         int cow;
2119         struct hstate *h = hstate_vma(vma);
2120         unsigned long sz = huge_page_size(h);
2121
2122         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2123
2124         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2125                 src_pte = huge_pte_offset(src, addr);
2126                 if (!src_pte)
2127                         continue;
2128                 dst_pte = huge_pte_alloc(dst, addr, sz);
2129                 if (!dst_pte)
2130                         goto nomem;
2131
2132                 /* If the pagetables are shared don't copy or take references */
2133                 if (dst_pte == src_pte)
2134                         continue;
2135
2136                 spin_lock(&dst->page_table_lock);
2137                 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2138                 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2139                         if (cow)
2140                                 huge_ptep_set_wrprotect(src, addr, src_pte);
2141                         entry = huge_ptep_get(src_pte);
2142                         ptepage = pte_page(entry);
2143                         get_page(ptepage);
2144                         page_dup_rmap(ptepage);
2145                         set_huge_pte_at(dst, addr, dst_pte, entry);
2146                 }
2147                 spin_unlock(&src->page_table_lock);
2148                 spin_unlock(&dst->page_table_lock);
2149         }
2150         return 0;
2151
2152 nomem:
2153         return -ENOMEM;
2154 }
2155
2156 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2157 {
2158         swp_entry_t swp;
2159
2160         if (huge_pte_none(pte) || pte_present(pte))
2161                 return 0;
2162         swp = pte_to_swp_entry(pte);
2163         if (non_swap_entry(swp) && is_hwpoison_entry(swp)) {
2164                 return 1;
2165         } else
2166                 return 0;
2167 }
2168
2169 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2170                             unsigned long end, struct page *ref_page)
2171 {
2172         struct mm_struct *mm = vma->vm_mm;
2173         unsigned long address;
2174         pte_t *ptep;
2175         pte_t pte;
2176         struct page *page;
2177         struct page *tmp;
2178         struct hstate *h = hstate_vma(vma);
2179         unsigned long sz = huge_page_size(h);
2180
2181         /*
2182          * A page gathering list, protected by per file i_mmap_lock. The
2183          * lock is used to avoid list corruption from multiple unmapping
2184          * of the same page since we are using page->lru.
2185          */
2186         LIST_HEAD(page_list);
2187
2188         WARN_ON(!is_vm_hugetlb_page(vma));
2189         BUG_ON(start & ~huge_page_mask(h));
2190         BUG_ON(end & ~huge_page_mask(h));
2191
2192         mmu_notifier_invalidate_range_start(mm, start, end);
2193         spin_lock(&mm->page_table_lock);
2194         for (address = start; address < end; address += sz) {
2195                 ptep = huge_pte_offset(mm, address);
2196                 if (!ptep)
2197                         continue;
2198
2199                 if (huge_pmd_unshare(mm, &address, ptep))
2200                         continue;
2201
2202                 /*
2203                  * If a reference page is supplied, it is because a specific
2204                  * page is being unmapped, not a range. Ensure the page we
2205                  * are about to unmap is the actual page of interest.
2206                  */
2207                 if (ref_page) {
2208                         pte = huge_ptep_get(ptep);
2209                         if (huge_pte_none(pte))
2210                                 continue;
2211                         page = pte_page(pte);
2212                         if (page != ref_page)
2213                                 continue;
2214
2215                         /*
2216                          * Mark the VMA as having unmapped its page so that
2217                          * future faults in this VMA will fail rather than
2218                          * looking like data was lost
2219                          */
2220                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2221                 }
2222
2223                 pte = huge_ptep_get_and_clear(mm, address, ptep);
2224                 if (huge_pte_none(pte))
2225                         continue;
2226
2227                 /*
2228                  * HWPoisoned hugepage is already unmapped and dropped reference
2229                  */
2230                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2231                         continue;
2232
2233                 page = pte_page(pte);
2234                 if (pte_dirty(pte))
2235                         set_page_dirty(page);
2236                 list_add(&page->lru, &page_list);
2237         }
2238         spin_unlock(&mm->page_table_lock);
2239         flush_tlb_range(vma, start, end);
2240         mmu_notifier_invalidate_range_end(mm, start, end);
2241         list_for_each_entry_safe(page, tmp, &page_list, lru) {
2242                 page_remove_rmap(page);
2243                 list_del(&page->lru);
2244                 put_page(page);
2245         }
2246 }
2247
2248 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2249                           unsigned long end, struct page *ref_page)
2250 {
2251         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2252         __unmap_hugepage_range(vma, start, end, ref_page);
2253         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2254 }
2255
2256 /*
2257  * This is called when the original mapper is failing to COW a MAP_PRIVATE
2258  * mappping it owns the reserve page for. The intention is to unmap the page
2259  * from other VMAs and let the children be SIGKILLed if they are faulting the
2260  * same region.
2261  */
2262 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2263                                 struct page *page, unsigned long address)
2264 {
2265         struct hstate *h = hstate_vma(vma);
2266         struct vm_area_struct *iter_vma;
2267         struct address_space *mapping;
2268         struct prio_tree_iter iter;
2269         pgoff_t pgoff;
2270
2271         /*
2272          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2273          * from page cache lookup which is in HPAGE_SIZE units.
2274          */
2275         address = address & huge_page_mask(h);
2276         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
2277                 + (vma->vm_pgoff >> PAGE_SHIFT);
2278         mapping = (struct address_space *)page_private(page);
2279
2280         /*
2281          * Take the mapping lock for the duration of the table walk. As
2282          * this mapping should be shared between all the VMAs,
2283          * __unmap_hugepage_range() is called as the lock is already held
2284          */
2285         spin_lock(&mapping->i_mmap_lock);
2286         vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2287                 /* Do not unmap the current VMA */
2288                 if (iter_vma == vma)
2289                         continue;
2290
2291                 /*
2292                  * Unmap the page from other VMAs without their own reserves.
2293                  * They get marked to be SIGKILLed if they fault in these
2294                  * areas. This is because a future no-page fault on this VMA
2295                  * could insert a zeroed page instead of the data existing
2296                  * from the time of fork. This would look like data corruption
2297                  */
2298                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2299                         __unmap_hugepage_range(iter_vma,
2300                                 address, address + huge_page_size(h),
2301                                 page);
2302         }
2303         spin_unlock(&mapping->i_mmap_lock);
2304
2305         return 1;
2306 }
2307
2308 /*
2309  * Hugetlb_cow() should be called with page lock of the original hugepage held.
2310  */
2311 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2312                         unsigned long address, pte_t *ptep, pte_t pte,
2313                         struct page *pagecache_page)
2314 {
2315         struct hstate *h = hstate_vma(vma);
2316         struct page *old_page, *new_page;
2317         int avoidcopy;
2318         int outside_reserve = 0;
2319
2320         old_page = pte_page(pte);
2321
2322 retry_avoidcopy:
2323         /* If no-one else is actually using this page, avoid the copy
2324          * and just make the page writable */
2325         avoidcopy = (page_mapcount(old_page) == 1);
2326         if (avoidcopy) {
2327                 if (!trylock_page(old_page)) {
2328                         if (PageAnon(old_page))
2329                                 page_move_anon_rmap(old_page, vma, address);
2330                 } else
2331                         unlock_page(old_page);
2332                 set_huge_ptep_writable(vma, address, ptep);
2333                 return 0;
2334         }
2335
2336         /*
2337          * If the process that created a MAP_PRIVATE mapping is about to
2338          * perform a COW due to a shared page count, attempt to satisfy
2339          * the allocation without using the existing reserves. The pagecache
2340          * page is used to determine if the reserve at this address was
2341          * consumed or not. If reserves were used, a partial faulted mapping
2342          * at the time of fork() could consume its reserves on COW instead
2343          * of the full address range.
2344          */
2345         if (!(vma->vm_flags & VM_MAYSHARE) &&
2346                         is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2347                         old_page != pagecache_page)
2348                 outside_reserve = 1;
2349
2350         page_cache_get(old_page);
2351
2352         /* Drop page_table_lock as buddy allocator may be called */
2353         spin_unlock(&mm->page_table_lock);
2354         new_page = alloc_huge_page(vma, address, outside_reserve);
2355
2356         if (IS_ERR(new_page)) {
2357                 page_cache_release(old_page);
2358
2359                 /*
2360                  * If a process owning a MAP_PRIVATE mapping fails to COW,
2361                  * it is due to references held by a child and an insufficient
2362                  * huge page pool. To guarantee the original mappers
2363                  * reliability, unmap the page from child processes. The child
2364                  * may get SIGKILLed if it later faults.
2365                  */
2366                 if (outside_reserve) {
2367                         BUG_ON(huge_pte_none(pte));
2368                         if (unmap_ref_private(mm, vma, old_page, address)) {
2369                                 BUG_ON(page_count(old_page) != 1);
2370                                 BUG_ON(huge_pte_none(pte));
2371                                 spin_lock(&mm->page_table_lock);
2372                                 goto retry_avoidcopy;
2373                         }
2374                         WARN_ON_ONCE(1);
2375                 }
2376
2377                 /* Caller expects lock to be held */
2378                 spin_lock(&mm->page_table_lock);
2379                 return -PTR_ERR(new_page);
2380         }
2381
2382         /*
2383          * When the original hugepage is shared one, it does not have
2384          * anon_vma prepared.
2385          */
2386         if (unlikely(anon_vma_prepare(vma)))
2387                 return VM_FAULT_OOM;
2388
2389         copy_huge_page(new_page, old_page, address, vma);
2390         __SetPageUptodate(new_page);
2391
2392         /*
2393          * Retake the page_table_lock to check for racing updates
2394          * before the page tables are altered
2395          */
2396         spin_lock(&mm->page_table_lock);
2397         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2398         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2399                 /* Break COW */
2400                 mmu_notifier_invalidate_range_start(mm,
2401                         address & huge_page_mask(h),
2402                         (address & huge_page_mask(h)) + huge_page_size(h));
2403                 huge_ptep_clear_flush(vma, address, ptep);
2404                 set_huge_pte_at(mm, address, ptep,
2405                                 make_huge_pte(vma, new_page, 1));
2406                 page_remove_rmap(old_page);
2407                 hugepage_add_anon_rmap(new_page, vma, address);
2408                 /* Make the old page be freed below */
2409                 new_page = old_page;
2410                 mmu_notifier_invalidate_range_end(mm,
2411                         address & huge_page_mask(h),
2412                         (address & huge_page_mask(h)) + huge_page_size(h));
2413         }
2414         page_cache_release(new_page);
2415         page_cache_release(old_page);
2416         return 0;
2417 }
2418
2419 /* Return the pagecache page at a given address within a VMA */
2420 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2421                         struct vm_area_struct *vma, unsigned long address)
2422 {
2423         struct address_space *mapping;
2424         pgoff_t idx;
2425
2426         mapping = vma->vm_file->f_mapping;
2427         idx = vma_hugecache_offset(h, vma, address);
2428
2429         return find_lock_page(mapping, idx);
2430 }
2431
2432 /*
2433  * Return whether there is a pagecache page to back given address within VMA.
2434  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2435  */
2436 static bool hugetlbfs_pagecache_present(struct hstate *h,
2437                         struct vm_area_struct *vma, unsigned long address)
2438 {
2439         struct address_space *mapping;
2440         pgoff_t idx;
2441         struct page *page;
2442
2443         mapping = vma->vm_file->f_mapping;
2444         idx = vma_hugecache_offset(h, vma, address);
2445
2446         page = find_get_page(mapping, idx);
2447         if (page)
2448                 put_page(page);
2449         return page != NULL;
2450 }
2451
2452 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2453                         unsigned long address, pte_t *ptep, unsigned int flags)
2454 {
2455         struct hstate *h = hstate_vma(vma);
2456         int ret = VM_FAULT_SIGBUS;
2457         pgoff_t idx;
2458         unsigned long size;
2459         struct page *page;
2460         struct address_space *mapping;
2461         pte_t new_pte;
2462
2463         /*
2464          * Currently, we are forced to kill the process in the event the
2465          * original mapper has unmapped pages from the child due to a failed
2466          * COW. Warn that such a situation has occured as it may not be obvious
2467          */
2468         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2469                 printk(KERN_WARNING
2470                         "PID %d killed due to inadequate hugepage pool\n",
2471                         current->pid);
2472                 return ret;
2473         }
2474
2475         mapping = vma->vm_file->f_mapping;
2476         idx = vma_hugecache_offset(h, vma, address);
2477
2478         /*
2479          * Use page lock to guard against racing truncation
2480          * before we get page_table_lock.
2481          */
2482 retry:
2483         page = find_lock_page(mapping, idx);
2484         if (!page) {
2485                 size = i_size_read(mapping->host) >> huge_page_shift(h);
2486                 if (idx >= size)
2487                         goto out;
2488                 page = alloc_huge_page(vma, address, 0);
2489                 if (IS_ERR(page)) {
2490                         ret = -PTR_ERR(page);
2491                         goto out;
2492                 }
2493                 clear_huge_page(page, address, huge_page_size(h));
2494                 __SetPageUptodate(page);
2495
2496                 if (vma->vm_flags & VM_MAYSHARE) {
2497                         int err;
2498                         struct inode *inode = mapping->host;
2499
2500                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2501                         if (err) {
2502                                 put_page(page);
2503                                 if (err == -EEXIST)
2504                                         goto retry;
2505                                 goto out;
2506                         }
2507
2508                         spin_lock(&inode->i_lock);
2509                         inode->i_blocks += blocks_per_huge_page(h);
2510                         spin_unlock(&inode->i_lock);
2511                         page_dup_rmap(page);
2512                 } else {
2513                         lock_page(page);
2514                         if (unlikely(anon_vma_prepare(vma))) {
2515                                 ret = VM_FAULT_OOM;
2516                                 goto backout_unlocked;
2517                         }
2518                         hugepage_add_new_anon_rmap(page, vma, address);
2519                 }
2520         } else {
2521                 page_dup_rmap(page);
2522         }
2523
2524         /*
2525          * Since memory error handler replaces pte into hwpoison swap entry
2526          * at the time of error handling, a process which reserved but not have
2527          * the mapping to the error hugepage does not have hwpoison swap entry.
2528          * So we need to block accesses from such a process by checking
2529          * PG_hwpoison bit here.
2530          */
2531         if (unlikely(PageHWPoison(page))) {
2532                 ret = VM_FAULT_HWPOISON;
2533                 goto backout_unlocked;
2534         }
2535
2536         /*
2537          * If we are going to COW a private mapping later, we examine the
2538          * pending reservations for this page now. This will ensure that
2539          * any allocations necessary to record that reservation occur outside
2540          * the spinlock.
2541          */
2542         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2543                 if (vma_needs_reservation(h, vma, address) < 0) {
2544                         ret = VM_FAULT_OOM;
2545                         goto backout_unlocked;
2546                 }
2547
2548         spin_lock(&mm->page_table_lock);
2549         size = i_size_read(mapping->host) >> huge_page_shift(h);
2550         if (idx >= size)
2551                 goto backout;
2552
2553         ret = 0;
2554         if (!huge_pte_none(huge_ptep_get(ptep)))
2555                 goto backout;
2556
2557         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2558                                 && (vma->vm_flags & VM_SHARED)));
2559         set_huge_pte_at(mm, address, ptep, new_pte);
2560
2561         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2562                 /* Optimization, do the COW without a second fault */
2563                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2564         }
2565
2566         spin_unlock(&mm->page_table_lock);
2567         unlock_page(page);
2568 out:
2569         return ret;
2570
2571 backout:
2572         spin_unlock(&mm->page_table_lock);
2573 backout_unlocked:
2574         unlock_page(page);
2575         put_page(page);
2576         goto out;
2577 }
2578
2579 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2580                         unsigned long address, unsigned int flags)
2581 {
2582         pte_t *ptep;
2583         pte_t entry;
2584         int ret;
2585         struct page *page = NULL;
2586         struct page *pagecache_page = NULL;
2587         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2588         struct hstate *h = hstate_vma(vma);
2589
2590         ptep = huge_pte_offset(mm, address);
2591         if (ptep) {
2592                 entry = huge_ptep_get(ptep);
2593                 if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2594                         return VM_FAULT_HWPOISON;
2595         }
2596
2597         ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2598         if (!ptep)
2599                 return VM_FAULT_OOM;
2600
2601         /*
2602          * Serialize hugepage allocation and instantiation, so that we don't
2603          * get spurious allocation failures if two CPUs race to instantiate
2604          * the same page in the page cache.
2605          */
2606         mutex_lock(&hugetlb_instantiation_mutex);
2607         entry = huge_ptep_get(ptep);
2608         if (huge_pte_none(entry)) {
2609                 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2610                 goto out_mutex;
2611         }
2612
2613         ret = 0;
2614
2615         /*
2616          * If we are going to COW the mapping later, we examine the pending
2617          * reservations for this page now. This will ensure that any
2618          * allocations necessary to record that reservation occur outside the
2619          * spinlock. For private mappings, we also lookup the pagecache
2620          * page now as it is used to determine if a reservation has been
2621          * consumed.
2622          */
2623         if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2624                 if (vma_needs_reservation(h, vma, address) < 0) {
2625                         ret = VM_FAULT_OOM;
2626                         goto out_mutex;
2627                 }
2628
2629                 if (!(vma->vm_flags & VM_MAYSHARE))
2630                         pagecache_page = hugetlbfs_pagecache_page(h,
2631                                                                 vma, address);
2632         }
2633
2634         if (!pagecache_page) {
2635                 page = pte_page(entry);
2636                 lock_page(page);
2637         }
2638
2639         spin_lock(&mm->page_table_lock);
2640         /* Check for a racing update before calling hugetlb_cow */
2641         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2642                 goto out_page_table_lock;
2643
2644
2645         if (flags & FAULT_FLAG_WRITE) {
2646                 if (!pte_write(entry)) {
2647                         ret = hugetlb_cow(mm, vma, address, ptep, entry,
2648                                                         pagecache_page);
2649                         goto out_page_table_lock;
2650                 }
2651                 entry = pte_mkdirty(entry);
2652         }
2653         entry = pte_mkyoung(entry);
2654         if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2655                                                 flags & FAULT_FLAG_WRITE))
2656                 update_mmu_cache(vma, address, ptep);
2657
2658 out_page_table_lock:
2659         spin_unlock(&mm->page_table_lock);
2660
2661         if (pagecache_page) {
2662                 unlock_page(pagecache_page);
2663                 put_page(pagecache_page);
2664         } else {
2665                 unlock_page(page);
2666         }
2667
2668 out_mutex:
2669         mutex_unlock(&hugetlb_instantiation_mutex);
2670
2671         return ret;
2672 }
2673
2674 /* Can be overriden by architectures */
2675 __attribute__((weak)) struct page *
2676 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2677                pud_t *pud, int write)
2678 {
2679         BUG();
2680         return NULL;
2681 }
2682
2683 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2684                         struct page **pages, struct vm_area_struct **vmas,
2685                         unsigned long *position, int *length, int i,
2686                         unsigned int flags)
2687 {
2688         unsigned long pfn_offset;
2689         unsigned long vaddr = *position;
2690         int remainder = *length;
2691         struct hstate *h = hstate_vma(vma);
2692
2693         spin_lock(&mm->page_table_lock);
2694         while (vaddr < vma->vm_end && remainder) {
2695                 pte_t *pte;
2696                 int absent;
2697                 struct page *page;
2698
2699                 /*
2700                  * Some archs (sparc64, sh*) have multiple pte_ts to
2701                  * each hugepage.  We have to make sure we get the
2702                  * first, for the page indexing below to work.
2703                  */
2704                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2705                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2706
2707                 /*
2708                  * When coredumping, it suits get_dump_page if we just return
2709                  * an error where there's an empty slot with no huge pagecache
2710                  * to back it.  This way, we avoid allocating a hugepage, and
2711                  * the sparse dumpfile avoids allocating disk blocks, but its
2712                  * huge holes still show up with zeroes where they need to be.
2713                  */
2714                 if (absent && (flags & FOLL_DUMP) &&
2715                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2716                         remainder = 0;
2717                         break;
2718                 }
2719
2720                 if (absent ||
2721                     ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2722                         int ret;
2723
2724                         spin_unlock(&mm->page_table_lock);
2725                         ret = hugetlb_fault(mm, vma, vaddr,
2726                                 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2727                         spin_lock(&mm->page_table_lock);
2728                         if (!(ret & VM_FAULT_ERROR))
2729                                 continue;
2730
2731                         remainder = 0;
2732                         break;
2733                 }
2734
2735                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2736                 page = pte_page(huge_ptep_get(pte));
2737 same_page:
2738                 if (pages) {
2739                         pages[i] = mem_map_offset(page, pfn_offset);
2740                         get_page(pages[i]);
2741                 }
2742
2743                 if (vmas)
2744                         vmas[i] = vma;
2745
2746                 vaddr += PAGE_SIZE;
2747                 ++pfn_offset;
2748                 --remainder;
2749                 ++i;
2750                 if (vaddr < vma->vm_end && remainder &&
2751                                 pfn_offset < pages_per_huge_page(h)) {
2752                         /*
2753                          * We use pfn_offset to avoid touching the pageframes
2754                          * of this compound page.
2755                          */
2756                         goto same_page;
2757                 }
2758         }
2759         spin_unlock(&mm->page_table_lock);
2760         *length = remainder;
2761         *position = vaddr;
2762
2763         return i ? i : -EFAULT;
2764 }
2765
2766 void hugetlb_change_protection(struct vm_area_struct *vma,
2767                 unsigned long address, unsigned long end, pgprot_t newprot)
2768 {
2769         struct mm_struct *mm = vma->vm_mm;
2770         unsigned long start = address;
2771         pte_t *ptep;
2772         pte_t pte;
2773         struct hstate *h = hstate_vma(vma);
2774
2775         BUG_ON(address >= end);
2776         flush_cache_range(vma, address, end);
2777
2778         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2779         spin_lock(&mm->page_table_lock);
2780         for (; address < end; address += huge_page_size(h)) {
2781                 ptep = huge_pte_offset(mm, address);
2782                 if (!ptep)
2783                         continue;
2784                 if (huge_pmd_unshare(mm, &address, ptep))
2785                         continue;
2786                 if (!huge_pte_none(huge_ptep_get(ptep))) {
2787                         pte = huge_ptep_get_and_clear(mm, address, ptep);
2788                         pte = pte_mkhuge(pte_modify(pte, newprot));
2789                         set_huge_pte_at(mm, address, ptep, pte);
2790                 }
2791         }
2792         spin_unlock(&mm->page_table_lock);
2793         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2794
2795         flush_tlb_range(vma, start, end);
2796 }
2797
2798 int hugetlb_reserve_pages(struct inode *inode,
2799                                         long from, long to,
2800                                         struct vm_area_struct *vma,
2801                                         int acctflag)
2802 {
2803         long ret, chg;
2804         struct hstate *h = hstate_inode(inode);
2805
2806         /*
2807          * Only apply hugepage reservation if asked. At fault time, an
2808          * attempt will be made for VM_NORESERVE to allocate a page
2809          * and filesystem quota without using reserves
2810          */
2811         if (acctflag & VM_NORESERVE)
2812                 return 0;
2813
2814         /*
2815          * Shared mappings base their reservation on the number of pages that
2816          * are already allocated on behalf of the file. Private mappings need
2817          * to reserve the full area even if read-only as mprotect() may be
2818          * called to make the mapping read-write. Assume !vma is a shm mapping
2819          */
2820         if (!vma || vma->vm_flags & VM_MAYSHARE)
2821                 chg = region_chg(&inode->i_mapping->private_list, from, to);
2822         else {
2823                 struct resv_map *resv_map = resv_map_alloc();
2824                 if (!resv_map)
2825                         return -ENOMEM;
2826
2827                 chg = to - from;
2828
2829                 set_vma_resv_map(vma, resv_map);
2830                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2831         }
2832
2833         if (chg < 0)
2834                 return chg;
2835
2836         /* There must be enough filesystem quota for the mapping */
2837         if (hugetlb_get_quota(inode->i_mapping, chg))
2838                 return -ENOSPC;
2839
2840         /*
2841          * Check enough hugepages are available for the reservation.
2842          * Hand back the quota if there are not
2843          */
2844         ret = hugetlb_acct_memory(h, chg);
2845         if (ret < 0) {
2846                 hugetlb_put_quota(inode->i_mapping, chg);
2847                 return ret;
2848         }
2849
2850         /*
2851          * Account for the reservations made. Shared mappings record regions
2852          * that have reservations as they are shared by multiple VMAs.
2853          * When the last VMA disappears, the region map says how much
2854          * the reservation was and the page cache tells how much of
2855          * the reservation was consumed. Private mappings are per-VMA and
2856          * only the consumed reservations are tracked. When the VMA
2857          * disappears, the original reservation is the VMA size and the
2858          * consumed reservations are stored in the map. Hence, nothing
2859          * else has to be done for private mappings here
2860          */
2861         if (!vma || vma->vm_flags & VM_MAYSHARE)
2862                 region_add(&inode->i_mapping->private_list, from, to);
2863         return 0;
2864 }
2865
2866 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2867 {
2868         struct hstate *h = hstate_inode(inode);
2869         long chg = region_truncate(&inode->i_mapping->private_list, offset);
2870
2871         spin_lock(&inode->i_lock);
2872         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
2873         spin_unlock(&inode->i_lock);
2874
2875         hugetlb_put_quota(inode->i_mapping, (chg - freed));
2876         hugetlb_acct_memory(h, -(chg - freed));
2877 }
2878
2879 /*
2880  * This function is called from memory failure code.
2881  * Assume the caller holds page lock of the head page.
2882  */
2883 void __isolate_hwpoisoned_huge_page(struct page *hpage)
2884 {
2885         struct hstate *h = page_hstate(hpage);
2886         int nid = page_to_nid(hpage);
2887
2888         spin_lock(&hugetlb_lock);
2889         list_del(&hpage->lru);
2890         h->free_huge_pages--;
2891         h->free_huge_pages_node[nid]--;
2892         spin_unlock(&hugetlb_lock);
2893 }