089092d152ab3ee80a0e1e17d67ff6bc923a6759
[linux-2.6.git] / mm / hugetlb.c
1 /*
2  * Generic hugetlb support.
3  * (C) William Irwin, April 2004
4  */
5 #include <linux/gfp.h>
6 #include <linux/list.h>
7 #include <linux/init.h>
8 #include <linux/module.h>
9 #include <linux/mm.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/nodemask.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/cpuset.h>
16 #include <linux/mutex.h>
17
18 #include <asm/page.h>
19 #include <asm/pgtable.h>
20
21 #include <linux/hugetlb.h>
22 #include "internal.h"
23
24 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
25 static unsigned long nr_huge_pages, free_huge_pages, resv_huge_pages;
26 unsigned long max_huge_pages;
27 static struct list_head hugepage_freelists[MAX_NUMNODES];
28 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
29 static unsigned int free_huge_pages_node[MAX_NUMNODES];
30 /*
31  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
32  */
33 static DEFINE_SPINLOCK(hugetlb_lock);
34
35 static void clear_huge_page(struct page *page, unsigned long addr)
36 {
37         int i;
38
39         might_sleep();
40         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
41                 cond_resched();
42                 clear_user_highpage(page + i, addr);
43         }
44 }
45
46 static void copy_huge_page(struct page *dst, struct page *src,
47                            unsigned long addr)
48 {
49         int i;
50
51         might_sleep();
52         for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
53                 cond_resched();
54                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE);
55         }
56 }
57
58 static void enqueue_huge_page(struct page *page)
59 {
60         int nid = page_to_nid(page);
61         list_add(&page->lru, &hugepage_freelists[nid]);
62         free_huge_pages++;
63         free_huge_pages_node[nid]++;
64 }
65
66 static struct page *dequeue_huge_page(struct vm_area_struct *vma,
67                                 unsigned long address)
68 {
69         int nid = numa_node_id();
70         struct page *page = NULL;
71         struct zonelist *zonelist = huge_zonelist(vma, address);
72         struct zone **z;
73
74         for (z = zonelist->zones; *z; z++) {
75                 nid = zone_to_nid(*z);
76                 if (cpuset_zone_allowed_softwall(*z, GFP_HIGHUSER) &&
77                     !list_empty(&hugepage_freelists[nid]))
78                         break;
79         }
80
81         if (*z) {
82                 page = list_entry(hugepage_freelists[nid].next,
83                                   struct page, lru);
84                 list_del(&page->lru);
85                 free_huge_pages--;
86                 free_huge_pages_node[nid]--;
87         }
88         return page;
89 }
90
91 static void free_huge_page(struct page *page)
92 {
93         BUG_ON(page_count(page));
94
95         INIT_LIST_HEAD(&page->lru);
96
97         spin_lock(&hugetlb_lock);
98         enqueue_huge_page(page);
99         spin_unlock(&hugetlb_lock);
100 }
101
102 static int alloc_fresh_huge_page(void)
103 {
104         static int nid = 0;
105         struct page *page;
106         page = alloc_pages_node(nid, GFP_HIGHUSER|__GFP_COMP|__GFP_NOWARN,
107                                         HUGETLB_PAGE_ORDER);
108         nid = next_node(nid, node_online_map);
109         if (nid == MAX_NUMNODES)
110                 nid = first_node(node_online_map);
111         if (page) {
112                 set_compound_page_dtor(page, free_huge_page);
113                 spin_lock(&hugetlb_lock);
114                 nr_huge_pages++;
115                 nr_huge_pages_node[page_to_nid(page)]++;
116                 spin_unlock(&hugetlb_lock);
117                 put_page(page); /* free it into the hugepage allocator */
118                 return 1;
119         }
120         return 0;
121 }
122
123 static struct page *alloc_huge_page(struct vm_area_struct *vma,
124                                     unsigned long addr)
125 {
126         struct page *page;
127
128         spin_lock(&hugetlb_lock);
129         if (vma->vm_flags & VM_MAYSHARE)
130                 resv_huge_pages--;
131         else if (free_huge_pages <= resv_huge_pages)
132                 goto fail;
133
134         page = dequeue_huge_page(vma, addr);
135         if (!page)
136                 goto fail;
137
138         spin_unlock(&hugetlb_lock);
139         set_page_refcounted(page);
140         return page;
141
142 fail:
143         spin_unlock(&hugetlb_lock);
144         return NULL;
145 }
146
147 static int __init hugetlb_init(void)
148 {
149         unsigned long i;
150
151         if (HPAGE_SHIFT == 0)
152                 return 0;
153
154         for (i = 0; i < MAX_NUMNODES; ++i)
155                 INIT_LIST_HEAD(&hugepage_freelists[i]);
156
157         for (i = 0; i < max_huge_pages; ++i) {
158                 if (!alloc_fresh_huge_page())
159                         break;
160         }
161         max_huge_pages = free_huge_pages = nr_huge_pages = i;
162         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
163         return 0;
164 }
165 module_init(hugetlb_init);
166
167 static int __init hugetlb_setup(char *s)
168 {
169         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
170                 max_huge_pages = 0;
171         return 1;
172 }
173 __setup("hugepages=", hugetlb_setup);
174
175 #ifdef CONFIG_SYSCTL
176 static void update_and_free_page(struct page *page)
177 {
178         int i;
179         nr_huge_pages--;
180         nr_huge_pages_node[page_to_nid(page)]--;
181         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
182                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
183                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
184                                 1 << PG_private | 1<< PG_writeback);
185         }
186         page[1].lru.next = NULL;
187         set_page_refcounted(page);
188         __free_pages(page, HUGETLB_PAGE_ORDER);
189 }
190
191 #ifdef CONFIG_HIGHMEM
192 static void try_to_free_low(unsigned long count)
193 {
194         int i;
195
196         for (i = 0; i < MAX_NUMNODES; ++i) {
197                 struct page *page, *next;
198                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
199                         if (PageHighMem(page))
200                                 continue;
201                         list_del(&page->lru);
202                         update_and_free_page(page);
203                         free_huge_pages--;
204                         free_huge_pages_node[page_to_nid(page)]--;
205                         if (count >= nr_huge_pages)
206                                 return;
207                 }
208         }
209 }
210 #else
211 static inline void try_to_free_low(unsigned long count)
212 {
213 }
214 #endif
215
216 static unsigned long set_max_huge_pages(unsigned long count)
217 {
218         while (count > nr_huge_pages) {
219                 if (!alloc_fresh_huge_page())
220                         return nr_huge_pages;
221         }
222         if (count >= nr_huge_pages)
223                 return nr_huge_pages;
224
225         spin_lock(&hugetlb_lock);
226         count = max(count, resv_huge_pages);
227         try_to_free_low(count);
228         while (count < nr_huge_pages) {
229                 struct page *page = dequeue_huge_page(NULL, 0);
230                 if (!page)
231                         break;
232                 update_and_free_page(page);
233         }
234         spin_unlock(&hugetlb_lock);
235         return nr_huge_pages;
236 }
237
238 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
239                            struct file *file, void __user *buffer,
240                            size_t *length, loff_t *ppos)
241 {
242         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
243         max_huge_pages = set_max_huge_pages(max_huge_pages);
244         return 0;
245 }
246 #endif /* CONFIG_SYSCTL */
247
248 int hugetlb_report_meminfo(char *buf)
249 {
250         return sprintf(buf,
251                         "HugePages_Total: %5lu\n"
252                         "HugePages_Free:  %5lu\n"
253                         "HugePages_Rsvd:  %5lu\n"
254                         "Hugepagesize:    %5lu kB\n",
255                         nr_huge_pages,
256                         free_huge_pages,
257                         resv_huge_pages,
258                         HPAGE_SIZE/1024);
259 }
260
261 int hugetlb_report_node_meminfo(int nid, char *buf)
262 {
263         return sprintf(buf,
264                 "Node %d HugePages_Total: %5u\n"
265                 "Node %d HugePages_Free:  %5u\n",
266                 nid, nr_huge_pages_node[nid],
267                 nid, free_huge_pages_node[nid]);
268 }
269
270 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
271 unsigned long hugetlb_total_pages(void)
272 {
273         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
274 }
275
276 /*
277  * We cannot handle pagefaults against hugetlb pages at all.  They cause
278  * handle_mm_fault() to try to instantiate regular-sized pages in the
279  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
280  * this far.
281  */
282 static struct page *hugetlb_nopage(struct vm_area_struct *vma,
283                                 unsigned long address, int *unused)
284 {
285         BUG();
286         return NULL;
287 }
288
289 struct vm_operations_struct hugetlb_vm_ops = {
290         .nopage = hugetlb_nopage,
291 };
292
293 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
294                                 int writable)
295 {
296         pte_t entry;
297
298         if (writable) {
299                 entry =
300                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
301         } else {
302                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
303         }
304         entry = pte_mkyoung(entry);
305         entry = pte_mkhuge(entry);
306
307         return entry;
308 }
309
310 static void set_huge_ptep_writable(struct vm_area_struct *vma,
311                                    unsigned long address, pte_t *ptep)
312 {
313         pte_t entry;
314
315         entry = pte_mkwrite(pte_mkdirty(*ptep));
316         ptep_set_access_flags(vma, address, ptep, entry, 1);
317         update_mmu_cache(vma, address, entry);
318         lazy_mmu_prot_update(entry);
319 }
320
321
322 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
323                             struct vm_area_struct *vma)
324 {
325         pte_t *src_pte, *dst_pte, entry;
326         struct page *ptepage;
327         unsigned long addr;
328         int cow;
329
330         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
331
332         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
333                 src_pte = huge_pte_offset(src, addr);
334                 if (!src_pte)
335                         continue;
336                 dst_pte = huge_pte_alloc(dst, addr);
337                 if (!dst_pte)
338                         goto nomem;
339                 spin_lock(&dst->page_table_lock);
340                 spin_lock(&src->page_table_lock);
341                 if (!pte_none(*src_pte)) {
342                         if (cow)
343                                 ptep_set_wrprotect(src, addr, src_pte);
344                         entry = *src_pte;
345                         ptepage = pte_page(entry);
346                         get_page(ptepage);
347                         set_huge_pte_at(dst, addr, dst_pte, entry);
348                 }
349                 spin_unlock(&src->page_table_lock);
350                 spin_unlock(&dst->page_table_lock);
351         }
352         return 0;
353
354 nomem:
355         return -ENOMEM;
356 }
357
358 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
359                             unsigned long end)
360 {
361         struct mm_struct *mm = vma->vm_mm;
362         unsigned long address;
363         pte_t *ptep;
364         pte_t pte;
365         struct page *page;
366         struct page *tmp;
367         /*
368          * A page gathering list, protected by per file i_mmap_lock. The
369          * lock is used to avoid list corruption from multiple unmapping
370          * of the same page since we are using page->lru.
371          */
372         LIST_HEAD(page_list);
373
374         WARN_ON(!is_vm_hugetlb_page(vma));
375         BUG_ON(start & ~HPAGE_MASK);
376         BUG_ON(end & ~HPAGE_MASK);
377
378         spin_lock(&mm->page_table_lock);
379         for (address = start; address < end; address += HPAGE_SIZE) {
380                 ptep = huge_pte_offset(mm, address);
381                 if (!ptep)
382                         continue;
383
384                 if (huge_pmd_unshare(mm, &address, ptep))
385                         continue;
386
387                 pte = huge_ptep_get_and_clear(mm, address, ptep);
388                 if (pte_none(pte))
389                         continue;
390
391                 page = pte_page(pte);
392                 list_add(&page->lru, &page_list);
393         }
394         spin_unlock(&mm->page_table_lock);
395         flush_tlb_range(vma, start, end);
396         list_for_each_entry_safe(page, tmp, &page_list, lru) {
397                 list_del(&page->lru);
398                 put_page(page);
399         }
400 }
401
402 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
403                           unsigned long end)
404 {
405         /*
406          * It is undesirable to test vma->vm_file as it should be non-null
407          * for valid hugetlb area. However, vm_file will be NULL in the error
408          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
409          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
410          * to clean up. Since no pte has actually been setup, it is safe to
411          * do nothing in this case.
412          */
413         if (vma->vm_file) {
414                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
415                 __unmap_hugepage_range(vma, start, end);
416                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
417         }
418 }
419
420 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
421                         unsigned long address, pte_t *ptep, pte_t pte)
422 {
423         struct page *old_page, *new_page;
424         int avoidcopy;
425
426         old_page = pte_page(pte);
427
428         /* If no-one else is actually using this page, avoid the copy
429          * and just make the page writable */
430         avoidcopy = (page_count(old_page) == 1);
431         if (avoidcopy) {
432                 set_huge_ptep_writable(vma, address, ptep);
433                 return VM_FAULT_MINOR;
434         }
435
436         page_cache_get(old_page);
437         new_page = alloc_huge_page(vma, address);
438
439         if (!new_page) {
440                 page_cache_release(old_page);
441                 return VM_FAULT_OOM;
442         }
443
444         spin_unlock(&mm->page_table_lock);
445         copy_huge_page(new_page, old_page, address);
446         spin_lock(&mm->page_table_lock);
447
448         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
449         if (likely(pte_same(*ptep, pte))) {
450                 /* Break COW */
451                 set_huge_pte_at(mm, address, ptep,
452                                 make_huge_pte(vma, new_page, 1));
453                 /* Make the old page be freed below */
454                 new_page = old_page;
455         }
456         page_cache_release(new_page);
457         page_cache_release(old_page);
458         return VM_FAULT_MINOR;
459 }
460
461 int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
462                         unsigned long address, pte_t *ptep, int write_access)
463 {
464         int ret = VM_FAULT_SIGBUS;
465         unsigned long idx;
466         unsigned long size;
467         struct page *page;
468         struct address_space *mapping;
469         pte_t new_pte;
470
471         mapping = vma->vm_file->f_mapping;
472         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
473                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
474
475         /*
476          * Use page lock to guard against racing truncation
477          * before we get page_table_lock.
478          */
479 retry:
480         page = find_lock_page(mapping, idx);
481         if (!page) {
482                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
483                 if (idx >= size)
484                         goto out;
485                 if (hugetlb_get_quota(mapping))
486                         goto out;
487                 page = alloc_huge_page(vma, address);
488                 if (!page) {
489                         hugetlb_put_quota(mapping);
490                         ret = VM_FAULT_OOM;
491                         goto out;
492                 }
493                 clear_huge_page(page, address);
494
495                 if (vma->vm_flags & VM_SHARED) {
496                         int err;
497
498                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
499                         if (err) {
500                                 put_page(page);
501                                 hugetlb_put_quota(mapping);
502                                 if (err == -EEXIST)
503                                         goto retry;
504                                 goto out;
505                         }
506                 } else
507                         lock_page(page);
508         }
509
510         spin_lock(&mm->page_table_lock);
511         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
512         if (idx >= size)
513                 goto backout;
514
515         ret = VM_FAULT_MINOR;
516         if (!pte_none(*ptep))
517                 goto backout;
518
519         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
520                                 && (vma->vm_flags & VM_SHARED)));
521         set_huge_pte_at(mm, address, ptep, new_pte);
522
523         if (write_access && !(vma->vm_flags & VM_SHARED)) {
524                 /* Optimization, do the COW without a second fault */
525                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
526         }
527
528         spin_unlock(&mm->page_table_lock);
529         unlock_page(page);
530 out:
531         return ret;
532
533 backout:
534         spin_unlock(&mm->page_table_lock);
535         hugetlb_put_quota(mapping);
536         unlock_page(page);
537         put_page(page);
538         goto out;
539 }
540
541 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
542                         unsigned long address, int write_access)
543 {
544         pte_t *ptep;
545         pte_t entry;
546         int ret;
547         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
548
549         ptep = huge_pte_alloc(mm, address);
550         if (!ptep)
551                 return VM_FAULT_OOM;
552
553         /*
554          * Serialize hugepage allocation and instantiation, so that we don't
555          * get spurious allocation failures if two CPUs race to instantiate
556          * the same page in the page cache.
557          */
558         mutex_lock(&hugetlb_instantiation_mutex);
559         entry = *ptep;
560         if (pte_none(entry)) {
561                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
562                 mutex_unlock(&hugetlb_instantiation_mutex);
563                 return ret;
564         }
565
566         ret = VM_FAULT_MINOR;
567
568         spin_lock(&mm->page_table_lock);
569         /* Check for a racing update before calling hugetlb_cow */
570         if (likely(pte_same(entry, *ptep)))
571                 if (write_access && !pte_write(entry))
572                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
573         spin_unlock(&mm->page_table_lock);
574         mutex_unlock(&hugetlb_instantiation_mutex);
575
576         return ret;
577 }
578
579 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
580                         struct page **pages, struct vm_area_struct **vmas,
581                         unsigned long *position, int *length, int i)
582 {
583         unsigned long pfn_offset;
584         unsigned long vaddr = *position;
585         int remainder = *length;
586
587         spin_lock(&mm->page_table_lock);
588         while (vaddr < vma->vm_end && remainder) {
589                 pte_t *pte;
590                 struct page *page;
591
592                 /*
593                  * Some archs (sparc64, sh*) have multiple pte_ts to
594                  * each hugepage.  We have to make * sure we get the
595                  * first, for the page indexing below to work.
596                  */
597                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
598
599                 if (!pte || pte_none(*pte)) {
600                         int ret;
601
602                         spin_unlock(&mm->page_table_lock);
603                         ret = hugetlb_fault(mm, vma, vaddr, 0);
604                         spin_lock(&mm->page_table_lock);
605                         if (ret == VM_FAULT_MINOR)
606                                 continue;
607
608                         remainder = 0;
609                         if (!i)
610                                 i = -EFAULT;
611                         break;
612                 }
613
614                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
615                 page = pte_page(*pte);
616 same_page:
617                 if (pages) {
618                         get_page(page);
619                         pages[i] = page + pfn_offset;
620                 }
621
622                 if (vmas)
623                         vmas[i] = vma;
624
625                 vaddr += PAGE_SIZE;
626                 ++pfn_offset;
627                 --remainder;
628                 ++i;
629                 if (vaddr < vma->vm_end && remainder &&
630                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
631                         /*
632                          * We use pfn_offset to avoid touching the pageframes
633                          * of this compound page.
634                          */
635                         goto same_page;
636                 }
637         }
638         spin_unlock(&mm->page_table_lock);
639         *length = remainder;
640         *position = vaddr;
641
642         return i;
643 }
644
645 void hugetlb_change_protection(struct vm_area_struct *vma,
646                 unsigned long address, unsigned long end, pgprot_t newprot)
647 {
648         struct mm_struct *mm = vma->vm_mm;
649         unsigned long start = address;
650         pte_t *ptep;
651         pte_t pte;
652
653         BUG_ON(address >= end);
654         flush_cache_range(vma, address, end);
655
656         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
657         spin_lock(&mm->page_table_lock);
658         for (; address < end; address += HPAGE_SIZE) {
659                 ptep = huge_pte_offset(mm, address);
660                 if (!ptep)
661                         continue;
662                 if (huge_pmd_unshare(mm, &address, ptep))
663                         continue;
664                 if (!pte_none(*ptep)) {
665                         pte = huge_ptep_get_and_clear(mm, address, ptep);
666                         pte = pte_mkhuge(pte_modify(pte, newprot));
667                         set_huge_pte_at(mm, address, ptep, pte);
668                         lazy_mmu_prot_update(pte);
669                 }
670         }
671         spin_unlock(&mm->page_table_lock);
672         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
673
674         flush_tlb_range(vma, start, end);
675 }
676
677 struct file_region {
678         struct list_head link;
679         long from;
680         long to;
681 };
682
683 static long region_add(struct list_head *head, long f, long t)
684 {
685         struct file_region *rg, *nrg, *trg;
686
687         /* Locate the region we are either in or before. */
688         list_for_each_entry(rg, head, link)
689                 if (f <= rg->to)
690                         break;
691
692         /* Round our left edge to the current segment if it encloses us. */
693         if (f > rg->from)
694                 f = rg->from;
695
696         /* Check for and consume any regions we now overlap with. */
697         nrg = rg;
698         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
699                 if (&rg->link == head)
700                         break;
701                 if (rg->from > t)
702                         break;
703
704                 /* If this area reaches higher then extend our area to
705                  * include it completely.  If this is not the first area
706                  * which we intend to reuse, free it. */
707                 if (rg->to > t)
708                         t = rg->to;
709                 if (rg != nrg) {
710                         list_del(&rg->link);
711                         kfree(rg);
712                 }
713         }
714         nrg->from = f;
715         nrg->to = t;
716         return 0;
717 }
718
719 static long region_chg(struct list_head *head, long f, long t)
720 {
721         struct file_region *rg, *nrg;
722         long chg = 0;
723
724         /* Locate the region we are before or in. */
725         list_for_each_entry(rg, head, link)
726                 if (f <= rg->to)
727                         break;
728
729         /* If we are below the current region then a new region is required.
730          * Subtle, allocate a new region at the position but make it zero
731          * size such that we can guarentee to record the reservation. */
732         if (&rg->link == head || t < rg->from) {
733                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
734                 if (nrg == 0)
735                         return -ENOMEM;
736                 nrg->from = f;
737                 nrg->to   = f;
738                 INIT_LIST_HEAD(&nrg->link);
739                 list_add(&nrg->link, rg->link.prev);
740
741                 return t - f;
742         }
743
744         /* Round our left edge to the current segment if it encloses us. */
745         if (f > rg->from)
746                 f = rg->from;
747         chg = t - f;
748
749         /* Check for and consume any regions we now overlap with. */
750         list_for_each_entry(rg, rg->link.prev, link) {
751                 if (&rg->link == head)
752                         break;
753                 if (rg->from > t)
754                         return chg;
755
756                 /* We overlap with this area, if it extends futher than
757                  * us then we must extend ourselves.  Account for its
758                  * existing reservation. */
759                 if (rg->to > t) {
760                         chg += rg->to - t;
761                         t = rg->to;
762                 }
763                 chg -= rg->to - rg->from;
764         }
765         return chg;
766 }
767
768 static long region_truncate(struct list_head *head, long end)
769 {
770         struct file_region *rg, *trg;
771         long chg = 0;
772
773         /* Locate the region we are either in or before. */
774         list_for_each_entry(rg, head, link)
775                 if (end <= rg->to)
776                         break;
777         if (&rg->link == head)
778                 return 0;
779
780         /* If we are in the middle of a region then adjust it. */
781         if (end > rg->from) {
782                 chg = rg->to - end;
783                 rg->to = end;
784                 rg = list_entry(rg->link.next, typeof(*rg), link);
785         }
786
787         /* Drop any remaining regions. */
788         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
789                 if (&rg->link == head)
790                         break;
791                 chg += rg->to - rg->from;
792                 list_del(&rg->link);
793                 kfree(rg);
794         }
795         return chg;
796 }
797
798 static int hugetlb_acct_memory(long delta)
799 {
800         int ret = -ENOMEM;
801
802         spin_lock(&hugetlb_lock);
803         if ((delta + resv_huge_pages) <= free_huge_pages) {
804                 resv_huge_pages += delta;
805                 ret = 0;
806         }
807         spin_unlock(&hugetlb_lock);
808         return ret;
809 }
810
811 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
812 {
813         long ret, chg;
814
815         chg = region_chg(&inode->i_mapping->private_list, from, to);
816         if (chg < 0)
817                 return chg;
818         ret = hugetlb_acct_memory(chg);
819         if (ret < 0)
820                 return ret;
821         region_add(&inode->i_mapping->private_list, from, to);
822         return 0;
823 }
824
825 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
826 {
827         long chg = region_truncate(&inode->i_mapping->private_list, offset);
828         hugetlb_acct_memory(freed - chg);
829 }