[PATCH] Hugepages: Use page_to_nid rather than traversing zone pointers
[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 = (*z)->zone_pgdat->node_id;
76                 if (cpuset_zone_allowed(*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                 page[1].lru.next = (void *)free_huge_page;      /* dtor */
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                         add_mm_counter(dst, file_rss, HPAGE_SIZE / PAGE_SIZE);
348                         set_huge_pte_at(dst, addr, dst_pte, entry);
349                 }
350                 spin_unlock(&src->page_table_lock);
351                 spin_unlock(&dst->page_table_lock);
352         }
353         return 0;
354
355 nomem:
356         return -ENOMEM;
357 }
358
359 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
360                           unsigned long end)
361 {
362         struct mm_struct *mm = vma->vm_mm;
363         unsigned long address;
364         pte_t *ptep;
365         pte_t pte;
366         struct page *page;
367
368         WARN_ON(!is_vm_hugetlb_page(vma));
369         BUG_ON(start & ~HPAGE_MASK);
370         BUG_ON(end & ~HPAGE_MASK);
371
372         spin_lock(&mm->page_table_lock);
373
374         /* Update high watermark before we lower rss */
375         update_hiwater_rss(mm);
376
377         for (address = start; address < end; address += HPAGE_SIZE) {
378                 ptep = huge_pte_offset(mm, address);
379                 if (!ptep)
380                         continue;
381
382                 pte = huge_ptep_get_and_clear(mm, address, ptep);
383                 if (pte_none(pte))
384                         continue;
385
386                 page = pte_page(pte);
387                 put_page(page);
388                 add_mm_counter(mm, file_rss, (int) -(HPAGE_SIZE / PAGE_SIZE));
389         }
390
391         spin_unlock(&mm->page_table_lock);
392         flush_tlb_range(vma, start, end);
393 }
394
395 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
396                         unsigned long address, pte_t *ptep, pte_t pte)
397 {
398         struct page *old_page, *new_page;
399         int avoidcopy;
400
401         old_page = pte_page(pte);
402
403         /* If no-one else is actually using this page, avoid the copy
404          * and just make the page writable */
405         avoidcopy = (page_count(old_page) == 1);
406         if (avoidcopy) {
407                 set_huge_ptep_writable(vma, address, ptep);
408                 return VM_FAULT_MINOR;
409         }
410
411         page_cache_get(old_page);
412         new_page = alloc_huge_page(vma, address);
413
414         if (!new_page) {
415                 page_cache_release(old_page);
416                 return VM_FAULT_OOM;
417         }
418
419         spin_unlock(&mm->page_table_lock);
420         copy_huge_page(new_page, old_page, address);
421         spin_lock(&mm->page_table_lock);
422
423         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
424         if (likely(pte_same(*ptep, pte))) {
425                 /* Break COW */
426                 set_huge_pte_at(mm, address, ptep,
427                                 make_huge_pte(vma, new_page, 1));
428                 /* Make the old page be freed below */
429                 new_page = old_page;
430         }
431         page_cache_release(new_page);
432         page_cache_release(old_page);
433         return VM_FAULT_MINOR;
434 }
435
436 int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
437                         unsigned long address, pte_t *ptep, int write_access)
438 {
439         int ret = VM_FAULT_SIGBUS;
440         unsigned long idx;
441         unsigned long size;
442         struct page *page;
443         struct address_space *mapping;
444         pte_t new_pte;
445
446         mapping = vma->vm_file->f_mapping;
447         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
448                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
449
450         /*
451          * Use page lock to guard against racing truncation
452          * before we get page_table_lock.
453          */
454 retry:
455         page = find_lock_page(mapping, idx);
456         if (!page) {
457                 if (hugetlb_get_quota(mapping))
458                         goto out;
459                 page = alloc_huge_page(vma, address);
460                 if (!page) {
461                         hugetlb_put_quota(mapping);
462                         ret = VM_FAULT_OOM;
463                         goto out;
464                 }
465                 clear_huge_page(page, address);
466
467                 if (vma->vm_flags & VM_SHARED) {
468                         int err;
469
470                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
471                         if (err) {
472                                 put_page(page);
473                                 hugetlb_put_quota(mapping);
474                                 if (err == -EEXIST)
475                                         goto retry;
476                                 goto out;
477                         }
478                 } else
479                         lock_page(page);
480         }
481
482         spin_lock(&mm->page_table_lock);
483         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
484         if (idx >= size)
485                 goto backout;
486
487         ret = VM_FAULT_MINOR;
488         if (!pte_none(*ptep))
489                 goto backout;
490
491         add_mm_counter(mm, file_rss, HPAGE_SIZE / PAGE_SIZE);
492         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
493                                 && (vma->vm_flags & VM_SHARED)));
494         set_huge_pte_at(mm, address, ptep, new_pte);
495
496         if (write_access && !(vma->vm_flags & VM_SHARED)) {
497                 /* Optimization, do the COW without a second fault */
498                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
499         }
500
501         spin_unlock(&mm->page_table_lock);
502         unlock_page(page);
503 out:
504         return ret;
505
506 backout:
507         spin_unlock(&mm->page_table_lock);
508         hugetlb_put_quota(mapping);
509         unlock_page(page);
510         put_page(page);
511         goto out;
512 }
513
514 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
515                         unsigned long address, int write_access)
516 {
517         pte_t *ptep;
518         pte_t entry;
519         int ret;
520         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
521
522         ptep = huge_pte_alloc(mm, address);
523         if (!ptep)
524                 return VM_FAULT_OOM;
525
526         /*
527          * Serialize hugepage allocation and instantiation, so that we don't
528          * get spurious allocation failures if two CPUs race to instantiate
529          * the same page in the page cache.
530          */
531         mutex_lock(&hugetlb_instantiation_mutex);
532         entry = *ptep;
533         if (pte_none(entry)) {
534                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
535                 mutex_unlock(&hugetlb_instantiation_mutex);
536                 return ret;
537         }
538
539         ret = VM_FAULT_MINOR;
540
541         spin_lock(&mm->page_table_lock);
542         /* Check for a racing update before calling hugetlb_cow */
543         if (likely(pte_same(entry, *ptep)))
544                 if (write_access && !pte_write(entry))
545                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
546         spin_unlock(&mm->page_table_lock);
547         mutex_unlock(&hugetlb_instantiation_mutex);
548
549         return ret;
550 }
551
552 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
553                         struct page **pages, struct vm_area_struct **vmas,
554                         unsigned long *position, int *length, int i)
555 {
556         unsigned long pfn_offset;
557         unsigned long vaddr = *position;
558         int remainder = *length;
559
560         spin_lock(&mm->page_table_lock);
561         while (vaddr < vma->vm_end && remainder) {
562                 pte_t *pte;
563                 struct page *page;
564
565                 /*
566                  * Some archs (sparc64, sh*) have multiple pte_ts to
567                  * each hugepage.  We have to make * sure we get the
568                  * first, for the page indexing below to work.
569                  */
570                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
571
572                 if (!pte || pte_none(*pte)) {
573                         int ret;
574
575                         spin_unlock(&mm->page_table_lock);
576                         ret = hugetlb_fault(mm, vma, vaddr, 0);
577                         spin_lock(&mm->page_table_lock);
578                         if (ret == VM_FAULT_MINOR)
579                                 continue;
580
581                         remainder = 0;
582                         if (!i)
583                                 i = -EFAULT;
584                         break;
585                 }
586
587                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
588                 page = pte_page(*pte);
589 same_page:
590                 if (pages) {
591                         get_page(page);
592                         pages[i] = page + pfn_offset;
593                 }
594
595                 if (vmas)
596                         vmas[i] = vma;
597
598                 vaddr += PAGE_SIZE;
599                 ++pfn_offset;
600                 --remainder;
601                 ++i;
602                 if (vaddr < vma->vm_end && remainder &&
603                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
604                         /*
605                          * We use pfn_offset to avoid touching the pageframes
606                          * of this compound page.
607                          */
608                         goto same_page;
609                 }
610         }
611         spin_unlock(&mm->page_table_lock);
612         *length = remainder;
613         *position = vaddr;
614
615         return i;
616 }
617
618 void hugetlb_change_protection(struct vm_area_struct *vma,
619                 unsigned long address, unsigned long end, pgprot_t newprot)
620 {
621         struct mm_struct *mm = vma->vm_mm;
622         unsigned long start = address;
623         pte_t *ptep;
624         pte_t pte;
625
626         BUG_ON(address >= end);
627         flush_cache_range(vma, address, end);
628
629         spin_lock(&mm->page_table_lock);
630         for (; address < end; address += HPAGE_SIZE) {
631                 ptep = huge_pte_offset(mm, address);
632                 if (!ptep)
633                         continue;
634                 if (!pte_none(*ptep)) {
635                         pte = huge_ptep_get_and_clear(mm, address, ptep);
636                         pte = pte_mkhuge(pte_modify(pte, newprot));
637                         set_huge_pte_at(mm, address, ptep, pte);
638                         lazy_mmu_prot_update(pte);
639                 }
640         }
641         spin_unlock(&mm->page_table_lock);
642
643         flush_tlb_range(vma, start, end);
644 }
645
646 struct file_region {
647         struct list_head link;
648         long from;
649         long to;
650 };
651
652 static long region_add(struct list_head *head, long f, long t)
653 {
654         struct file_region *rg, *nrg, *trg;
655
656         /* Locate the region we are either in or before. */
657         list_for_each_entry(rg, head, link)
658                 if (f <= rg->to)
659                         break;
660
661         /* Round our left edge to the current segment if it encloses us. */
662         if (f > rg->from)
663                 f = rg->from;
664
665         /* Check for and consume any regions we now overlap with. */
666         nrg = rg;
667         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
668                 if (&rg->link == head)
669                         break;
670                 if (rg->from > t)
671                         break;
672
673                 /* If this area reaches higher then extend our area to
674                  * include it completely.  If this is not the first area
675                  * which we intend to reuse, free it. */
676                 if (rg->to > t)
677                         t = rg->to;
678                 if (rg != nrg) {
679                         list_del(&rg->link);
680                         kfree(rg);
681                 }
682         }
683         nrg->from = f;
684         nrg->to = t;
685         return 0;
686 }
687
688 static long region_chg(struct list_head *head, long f, long t)
689 {
690         struct file_region *rg, *nrg;
691         long chg = 0;
692
693         /* Locate the region we are before or in. */
694         list_for_each_entry(rg, head, link)
695                 if (f <= rg->to)
696                         break;
697
698         /* If we are below the current region then a new region is required.
699          * Subtle, allocate a new region at the position but make it zero
700          * size such that we can guarentee to record the reservation. */
701         if (&rg->link == head || t < rg->from) {
702                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
703                 if (nrg == 0)
704                         return -ENOMEM;
705                 nrg->from = f;
706                 nrg->to   = f;
707                 INIT_LIST_HEAD(&nrg->link);
708                 list_add(&nrg->link, rg->link.prev);
709
710                 return t - f;
711         }
712
713         /* Round our left edge to the current segment if it encloses us. */
714         if (f > rg->from)
715                 f = rg->from;
716         chg = t - f;
717
718         /* Check for and consume any regions we now overlap with. */
719         list_for_each_entry(rg, rg->link.prev, link) {
720                 if (&rg->link == head)
721                         break;
722                 if (rg->from > t)
723                         return chg;
724
725                 /* We overlap with this area, if it extends futher than
726                  * us then we must extend ourselves.  Account for its
727                  * existing reservation. */
728                 if (rg->to > t) {
729                         chg += rg->to - t;
730                         t = rg->to;
731                 }
732                 chg -= rg->to - rg->from;
733         }
734         return chg;
735 }
736
737 static long region_truncate(struct list_head *head, long end)
738 {
739         struct file_region *rg, *trg;
740         long chg = 0;
741
742         /* Locate the region we are either in or before. */
743         list_for_each_entry(rg, head, link)
744                 if (end <= rg->to)
745                         break;
746         if (&rg->link == head)
747                 return 0;
748
749         /* If we are in the middle of a region then adjust it. */
750         if (end > rg->from) {
751                 chg = rg->to - end;
752                 rg->to = end;
753                 rg = list_entry(rg->link.next, typeof(*rg), link);
754         }
755
756         /* Drop any remaining regions. */
757         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
758                 if (&rg->link == head)
759                         break;
760                 chg += rg->to - rg->from;
761                 list_del(&rg->link);
762                 kfree(rg);
763         }
764         return chg;
765 }
766
767 static int hugetlb_acct_memory(long delta)
768 {
769         int ret = -ENOMEM;
770
771         spin_lock(&hugetlb_lock);
772         if ((delta + resv_huge_pages) <= free_huge_pages) {
773                 resv_huge_pages += delta;
774                 ret = 0;
775         }
776         spin_unlock(&hugetlb_lock);
777         return ret;
778 }
779
780 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
781 {
782         long ret, chg;
783
784         chg = region_chg(&inode->i_mapping->private_list, from, to);
785         if (chg < 0)
786                 return chg;
787         ret = hugetlb_acct_memory(chg);
788         if (ret < 0)
789                 return ret;
790         region_add(&inode->i_mapping->private_list, from, to);
791         return 0;
792 }
793
794 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
795 {
796         long chg = region_truncate(&inode->i_mapping->private_list, offset);
797         hugetlb_acct_memory(freed - chg);
798 }