Merge branch 'upstream-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mfashe...
[linux-2.6] / 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, struct vm_area_struct *vma)
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, vma);
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;
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 prev_nid;
105         struct page *page;
106         static DEFINE_SPINLOCK(nid_lock);
107         int nid;
108
109         spin_lock(&nid_lock);
110         nid = next_node(prev_nid, node_online_map);
111         if (nid == MAX_NUMNODES)
112                 nid = first_node(node_online_map);
113         prev_nid = nid;
114         spin_unlock(&nid_lock);
115
116         page = alloc_pages_node(nid, GFP_HIGHUSER|__GFP_COMP|__GFP_NOWARN,
117                                         HUGETLB_PAGE_ORDER);
118         if (page) {
119                 set_compound_page_dtor(page, free_huge_page);
120                 spin_lock(&hugetlb_lock);
121                 nr_huge_pages++;
122                 nr_huge_pages_node[page_to_nid(page)]++;
123                 spin_unlock(&hugetlb_lock);
124                 put_page(page); /* free it into the hugepage allocator */
125                 return 1;
126         }
127         return 0;
128 }
129
130 static struct page *alloc_huge_page(struct vm_area_struct *vma,
131                                     unsigned long addr)
132 {
133         struct page *page;
134
135         spin_lock(&hugetlb_lock);
136         if (vma->vm_flags & VM_MAYSHARE)
137                 resv_huge_pages--;
138         else if (free_huge_pages <= resv_huge_pages)
139                 goto fail;
140
141         page = dequeue_huge_page(vma, addr);
142         if (!page)
143                 goto fail;
144
145         spin_unlock(&hugetlb_lock);
146         set_page_refcounted(page);
147         return page;
148
149 fail:
150         if (vma->vm_flags & VM_MAYSHARE)
151                 resv_huge_pages++;
152         spin_unlock(&hugetlb_lock);
153         return NULL;
154 }
155
156 static int __init hugetlb_init(void)
157 {
158         unsigned long i;
159
160         if (HPAGE_SHIFT == 0)
161                 return 0;
162
163         for (i = 0; i < MAX_NUMNODES; ++i)
164                 INIT_LIST_HEAD(&hugepage_freelists[i]);
165
166         for (i = 0; i < max_huge_pages; ++i) {
167                 if (!alloc_fresh_huge_page())
168                         break;
169         }
170         max_huge_pages = free_huge_pages = nr_huge_pages = i;
171         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
172         return 0;
173 }
174 module_init(hugetlb_init);
175
176 static int __init hugetlb_setup(char *s)
177 {
178         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
179                 max_huge_pages = 0;
180         return 1;
181 }
182 __setup("hugepages=", hugetlb_setup);
183
184 static unsigned int cpuset_mems_nr(unsigned int *array)
185 {
186         int node;
187         unsigned int nr = 0;
188
189         for_each_node_mask(node, cpuset_current_mems_allowed)
190                 nr += array[node];
191
192         return nr;
193 }
194
195 #ifdef CONFIG_SYSCTL
196 static void update_and_free_page(struct page *page)
197 {
198         int i;
199         nr_huge_pages--;
200         nr_huge_pages_node[page_to_nid(page)]--;
201         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
202                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
203                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
204                                 1 << PG_private | 1<< PG_writeback);
205         }
206         page[1].lru.next = NULL;
207         set_page_refcounted(page);
208         __free_pages(page, HUGETLB_PAGE_ORDER);
209 }
210
211 #ifdef CONFIG_HIGHMEM
212 static void try_to_free_low(unsigned long count)
213 {
214         int i;
215
216         for (i = 0; i < MAX_NUMNODES; ++i) {
217                 struct page *page, *next;
218                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
219                         if (PageHighMem(page))
220                                 continue;
221                         list_del(&page->lru);
222                         update_and_free_page(page);
223                         free_huge_pages--;
224                         free_huge_pages_node[page_to_nid(page)]--;
225                         if (count >= nr_huge_pages)
226                                 return;
227                 }
228         }
229 }
230 #else
231 static inline void try_to_free_low(unsigned long count)
232 {
233 }
234 #endif
235
236 static unsigned long set_max_huge_pages(unsigned long count)
237 {
238         while (count > nr_huge_pages) {
239                 if (!alloc_fresh_huge_page())
240                         return nr_huge_pages;
241         }
242         if (count >= nr_huge_pages)
243                 return nr_huge_pages;
244
245         spin_lock(&hugetlb_lock);
246         count = max(count, resv_huge_pages);
247         try_to_free_low(count);
248         while (count < nr_huge_pages) {
249                 struct page *page = dequeue_huge_page(NULL, 0);
250                 if (!page)
251                         break;
252                 update_and_free_page(page);
253         }
254         spin_unlock(&hugetlb_lock);
255         return nr_huge_pages;
256 }
257
258 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
259                            struct file *file, void __user *buffer,
260                            size_t *length, loff_t *ppos)
261 {
262         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
263         max_huge_pages = set_max_huge_pages(max_huge_pages);
264         return 0;
265 }
266 #endif /* CONFIG_SYSCTL */
267
268 int hugetlb_report_meminfo(char *buf)
269 {
270         return sprintf(buf,
271                         "HugePages_Total: %5lu\n"
272                         "HugePages_Free:  %5lu\n"
273                         "HugePages_Rsvd:  %5lu\n"
274                         "Hugepagesize:    %5lu kB\n",
275                         nr_huge_pages,
276                         free_huge_pages,
277                         resv_huge_pages,
278                         HPAGE_SIZE/1024);
279 }
280
281 int hugetlb_report_node_meminfo(int nid, char *buf)
282 {
283         return sprintf(buf,
284                 "Node %d HugePages_Total: %5u\n"
285                 "Node %d HugePages_Free:  %5u\n",
286                 nid, nr_huge_pages_node[nid],
287                 nid, free_huge_pages_node[nid]);
288 }
289
290 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
291 unsigned long hugetlb_total_pages(void)
292 {
293         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
294 }
295
296 /*
297  * We cannot handle pagefaults against hugetlb pages at all.  They cause
298  * handle_mm_fault() to try to instantiate regular-sized pages in the
299  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
300  * this far.
301  */
302 static struct page *hugetlb_nopage(struct vm_area_struct *vma,
303                                 unsigned long address, int *unused)
304 {
305         BUG();
306         return NULL;
307 }
308
309 struct vm_operations_struct hugetlb_vm_ops = {
310         .nopage = hugetlb_nopage,
311 };
312
313 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
314                                 int writable)
315 {
316         pte_t entry;
317
318         if (writable) {
319                 entry =
320                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
321         } else {
322                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
323         }
324         entry = pte_mkyoung(entry);
325         entry = pte_mkhuge(entry);
326
327         return entry;
328 }
329
330 static void set_huge_ptep_writable(struct vm_area_struct *vma,
331                                    unsigned long address, pte_t *ptep)
332 {
333         pte_t entry;
334
335         entry = pte_mkwrite(pte_mkdirty(*ptep));
336         if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
337                 update_mmu_cache(vma, address, entry);
338                 lazy_mmu_prot_update(entry);
339         }
340 }
341
342
343 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
344                             struct vm_area_struct *vma)
345 {
346         pte_t *src_pte, *dst_pte, entry;
347         struct page *ptepage;
348         unsigned long addr;
349         int cow;
350
351         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
352
353         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
354                 src_pte = huge_pte_offset(src, addr);
355                 if (!src_pte)
356                         continue;
357                 dst_pte = huge_pte_alloc(dst, addr);
358                 if (!dst_pte)
359                         goto nomem;
360                 spin_lock(&dst->page_table_lock);
361                 spin_lock(&src->page_table_lock);
362                 if (!pte_none(*src_pte)) {
363                         if (cow)
364                                 ptep_set_wrprotect(src, addr, src_pte);
365                         entry = *src_pte;
366                         ptepage = pte_page(entry);
367                         get_page(ptepage);
368                         set_huge_pte_at(dst, addr, dst_pte, entry);
369                 }
370                 spin_unlock(&src->page_table_lock);
371                 spin_unlock(&dst->page_table_lock);
372         }
373         return 0;
374
375 nomem:
376         return -ENOMEM;
377 }
378
379 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
380                             unsigned long end)
381 {
382         struct mm_struct *mm = vma->vm_mm;
383         unsigned long address;
384         pte_t *ptep;
385         pte_t pte;
386         struct page *page;
387         struct page *tmp;
388         /*
389          * A page gathering list, protected by per file i_mmap_lock. The
390          * lock is used to avoid list corruption from multiple unmapping
391          * of the same page since we are using page->lru.
392          */
393         LIST_HEAD(page_list);
394
395         WARN_ON(!is_vm_hugetlb_page(vma));
396         BUG_ON(start & ~HPAGE_MASK);
397         BUG_ON(end & ~HPAGE_MASK);
398
399         spin_lock(&mm->page_table_lock);
400         for (address = start; address < end; address += HPAGE_SIZE) {
401                 ptep = huge_pte_offset(mm, address);
402                 if (!ptep)
403                         continue;
404
405                 if (huge_pmd_unshare(mm, &address, ptep))
406                         continue;
407
408                 pte = huge_ptep_get_and_clear(mm, address, ptep);
409                 if (pte_none(pte))
410                         continue;
411
412                 page = pte_page(pte);
413                 if (pte_dirty(pte))
414                         set_page_dirty(page);
415                 list_add(&page->lru, &page_list);
416         }
417         spin_unlock(&mm->page_table_lock);
418         flush_tlb_range(vma, start, end);
419         list_for_each_entry_safe(page, tmp, &page_list, lru) {
420                 list_del(&page->lru);
421                 put_page(page);
422         }
423 }
424
425 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
426                           unsigned long end)
427 {
428         /*
429          * It is undesirable to test vma->vm_file as it should be non-null
430          * for valid hugetlb area. However, vm_file will be NULL in the error
431          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
432          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
433          * to clean up. Since no pte has actually been setup, it is safe to
434          * do nothing in this case.
435          */
436         if (vma->vm_file) {
437                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
438                 __unmap_hugepage_range(vma, start, end);
439                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
440         }
441 }
442
443 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
444                         unsigned long address, pte_t *ptep, pte_t pte)
445 {
446         struct page *old_page, *new_page;
447         int avoidcopy;
448
449         old_page = pte_page(pte);
450
451         /* If no-one else is actually using this page, avoid the copy
452          * and just make the page writable */
453         avoidcopy = (page_count(old_page) == 1);
454         if (avoidcopy) {
455                 set_huge_ptep_writable(vma, address, ptep);
456                 return VM_FAULT_MINOR;
457         }
458
459         page_cache_get(old_page);
460         new_page = alloc_huge_page(vma, address);
461
462         if (!new_page) {
463                 page_cache_release(old_page);
464                 return VM_FAULT_OOM;
465         }
466
467         spin_unlock(&mm->page_table_lock);
468         copy_huge_page(new_page, old_page, address, vma);
469         spin_lock(&mm->page_table_lock);
470
471         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
472         if (likely(pte_same(*ptep, pte))) {
473                 /* Break COW */
474                 set_huge_pte_at(mm, address, ptep,
475                                 make_huge_pte(vma, new_page, 1));
476                 /* Make the old page be freed below */
477                 new_page = old_page;
478         }
479         page_cache_release(new_page);
480         page_cache_release(old_page);
481         return VM_FAULT_MINOR;
482 }
483
484 int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
485                         unsigned long address, pte_t *ptep, int write_access)
486 {
487         int ret = VM_FAULT_SIGBUS;
488         unsigned long idx;
489         unsigned long size;
490         struct page *page;
491         struct address_space *mapping;
492         pte_t new_pte;
493
494         mapping = vma->vm_file->f_mapping;
495         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
496                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
497
498         /*
499          * Use page lock to guard against racing truncation
500          * before we get page_table_lock.
501          */
502 retry:
503         page = find_lock_page(mapping, idx);
504         if (!page) {
505                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
506                 if (idx >= size)
507                         goto out;
508                 if (hugetlb_get_quota(mapping))
509                         goto out;
510                 page = alloc_huge_page(vma, address);
511                 if (!page) {
512                         hugetlb_put_quota(mapping);
513                         ret = VM_FAULT_OOM;
514                         goto out;
515                 }
516                 clear_huge_page(page, address);
517
518                 if (vma->vm_flags & VM_SHARED) {
519                         int err;
520
521                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
522                         if (err) {
523                                 put_page(page);
524                                 hugetlb_put_quota(mapping);
525                                 if (err == -EEXIST)
526                                         goto retry;
527                                 goto out;
528                         }
529                 } else
530                         lock_page(page);
531         }
532
533         spin_lock(&mm->page_table_lock);
534         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
535         if (idx >= size)
536                 goto backout;
537
538         ret = VM_FAULT_MINOR;
539         if (!pte_none(*ptep))
540                 goto backout;
541
542         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
543                                 && (vma->vm_flags & VM_SHARED)));
544         set_huge_pte_at(mm, address, ptep, new_pte);
545
546         if (write_access && !(vma->vm_flags & VM_SHARED)) {
547                 /* Optimization, do the COW without a second fault */
548                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
549         }
550
551         spin_unlock(&mm->page_table_lock);
552         unlock_page(page);
553 out:
554         return ret;
555
556 backout:
557         spin_unlock(&mm->page_table_lock);
558         hugetlb_put_quota(mapping);
559         unlock_page(page);
560         put_page(page);
561         goto out;
562 }
563
564 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
565                         unsigned long address, int write_access)
566 {
567         pte_t *ptep;
568         pte_t entry;
569         int ret;
570         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
571
572         ptep = huge_pte_alloc(mm, address);
573         if (!ptep)
574                 return VM_FAULT_OOM;
575
576         /*
577          * Serialize hugepage allocation and instantiation, so that we don't
578          * get spurious allocation failures if two CPUs race to instantiate
579          * the same page in the page cache.
580          */
581         mutex_lock(&hugetlb_instantiation_mutex);
582         entry = *ptep;
583         if (pte_none(entry)) {
584                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
585                 mutex_unlock(&hugetlb_instantiation_mutex);
586                 return ret;
587         }
588
589         ret = VM_FAULT_MINOR;
590
591         spin_lock(&mm->page_table_lock);
592         /* Check for a racing update before calling hugetlb_cow */
593         if (likely(pte_same(entry, *ptep)))
594                 if (write_access && !pte_write(entry))
595                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
596         spin_unlock(&mm->page_table_lock);
597         mutex_unlock(&hugetlb_instantiation_mutex);
598
599         return ret;
600 }
601
602 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
603                         struct page **pages, struct vm_area_struct **vmas,
604                         unsigned long *position, int *length, int i)
605 {
606         unsigned long pfn_offset;
607         unsigned long vaddr = *position;
608         int remainder = *length;
609
610         spin_lock(&mm->page_table_lock);
611         while (vaddr < vma->vm_end && remainder) {
612                 pte_t *pte;
613                 struct page *page;
614
615                 /*
616                  * Some archs (sparc64, sh*) have multiple pte_ts to
617                  * each hugepage.  We have to make * sure we get the
618                  * first, for the page indexing below to work.
619                  */
620                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
621
622                 if (!pte || pte_none(*pte)) {
623                         int ret;
624
625                         spin_unlock(&mm->page_table_lock);
626                         ret = hugetlb_fault(mm, vma, vaddr, 0);
627                         spin_lock(&mm->page_table_lock);
628                         if (ret == VM_FAULT_MINOR)
629                                 continue;
630
631                         remainder = 0;
632                         if (!i)
633                                 i = -EFAULT;
634                         break;
635                 }
636
637                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
638                 page = pte_page(*pte);
639 same_page:
640                 if (pages) {
641                         get_page(page);
642                         pages[i] = page + pfn_offset;
643                 }
644
645                 if (vmas)
646                         vmas[i] = vma;
647
648                 vaddr += PAGE_SIZE;
649                 ++pfn_offset;
650                 --remainder;
651                 ++i;
652                 if (vaddr < vma->vm_end && remainder &&
653                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
654                         /*
655                          * We use pfn_offset to avoid touching the pageframes
656                          * of this compound page.
657                          */
658                         goto same_page;
659                 }
660         }
661         spin_unlock(&mm->page_table_lock);
662         *length = remainder;
663         *position = vaddr;
664
665         return i;
666 }
667
668 void hugetlb_change_protection(struct vm_area_struct *vma,
669                 unsigned long address, unsigned long end, pgprot_t newprot)
670 {
671         struct mm_struct *mm = vma->vm_mm;
672         unsigned long start = address;
673         pte_t *ptep;
674         pte_t pte;
675
676         BUG_ON(address >= end);
677         flush_cache_range(vma, address, end);
678
679         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
680         spin_lock(&mm->page_table_lock);
681         for (; address < end; address += HPAGE_SIZE) {
682                 ptep = huge_pte_offset(mm, address);
683                 if (!ptep)
684                         continue;
685                 if (huge_pmd_unshare(mm, &address, ptep))
686                         continue;
687                 if (!pte_none(*ptep)) {
688                         pte = huge_ptep_get_and_clear(mm, address, ptep);
689                         pte = pte_mkhuge(pte_modify(pte, newprot));
690                         set_huge_pte_at(mm, address, ptep, pte);
691                         lazy_mmu_prot_update(pte);
692                 }
693         }
694         spin_unlock(&mm->page_table_lock);
695         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
696
697         flush_tlb_range(vma, start, end);
698 }
699
700 struct file_region {
701         struct list_head link;
702         long from;
703         long to;
704 };
705
706 static long region_add(struct list_head *head, long f, long t)
707 {
708         struct file_region *rg, *nrg, *trg;
709
710         /* Locate the region we are either in or before. */
711         list_for_each_entry(rg, head, link)
712                 if (f <= rg->to)
713                         break;
714
715         /* Round our left edge to the current segment if it encloses us. */
716         if (f > rg->from)
717                 f = rg->from;
718
719         /* Check for and consume any regions we now overlap with. */
720         nrg = rg;
721         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
722                 if (&rg->link == head)
723                         break;
724                 if (rg->from > t)
725                         break;
726
727                 /* If this area reaches higher then extend our area to
728                  * include it completely.  If this is not the first area
729                  * which we intend to reuse, free it. */
730                 if (rg->to > t)
731                         t = rg->to;
732                 if (rg != nrg) {
733                         list_del(&rg->link);
734                         kfree(rg);
735                 }
736         }
737         nrg->from = f;
738         nrg->to = t;
739         return 0;
740 }
741
742 static long region_chg(struct list_head *head, long f, long t)
743 {
744         struct file_region *rg, *nrg;
745         long chg = 0;
746
747         /* Locate the region we are before or in. */
748         list_for_each_entry(rg, head, link)
749                 if (f <= rg->to)
750                         break;
751
752         /* If we are below the current region then a new region is required.
753          * Subtle, allocate a new region at the position but make it zero
754          * size such that we can guarentee to record the reservation. */
755         if (&rg->link == head || t < rg->from) {
756                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
757                 if (nrg == 0)
758                         return -ENOMEM;
759                 nrg->from = f;
760                 nrg->to   = f;
761                 INIT_LIST_HEAD(&nrg->link);
762                 list_add(&nrg->link, rg->link.prev);
763
764                 return t - f;
765         }
766
767         /* Round our left edge to the current segment if it encloses us. */
768         if (f > rg->from)
769                 f = rg->from;
770         chg = t - f;
771
772         /* Check for and consume any regions we now overlap with. */
773         list_for_each_entry(rg, rg->link.prev, link) {
774                 if (&rg->link == head)
775                         break;
776                 if (rg->from > t)
777                         return chg;
778
779                 /* We overlap with this area, if it extends futher than
780                  * us then we must extend ourselves.  Account for its
781                  * existing reservation. */
782                 if (rg->to > t) {
783                         chg += rg->to - t;
784                         t = rg->to;
785                 }
786                 chg -= rg->to - rg->from;
787         }
788         return chg;
789 }
790
791 static long region_truncate(struct list_head *head, long end)
792 {
793         struct file_region *rg, *trg;
794         long chg = 0;
795
796         /* Locate the region we are either in or before. */
797         list_for_each_entry(rg, head, link)
798                 if (end <= rg->to)
799                         break;
800         if (&rg->link == head)
801                 return 0;
802
803         /* If we are in the middle of a region then adjust it. */
804         if (end > rg->from) {
805                 chg = rg->to - end;
806                 rg->to = end;
807                 rg = list_entry(rg->link.next, typeof(*rg), link);
808         }
809
810         /* Drop any remaining regions. */
811         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
812                 if (&rg->link == head)
813                         break;
814                 chg += rg->to - rg->from;
815                 list_del(&rg->link);
816                 kfree(rg);
817         }
818         return chg;
819 }
820
821 static int hugetlb_acct_memory(long delta)
822 {
823         int ret = -ENOMEM;
824
825         spin_lock(&hugetlb_lock);
826         if ((delta + resv_huge_pages) <= free_huge_pages) {
827                 resv_huge_pages += delta;
828                 ret = 0;
829         }
830         spin_unlock(&hugetlb_lock);
831         return ret;
832 }
833
834 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
835 {
836         long ret, chg;
837
838         chg = region_chg(&inode->i_mapping->private_list, from, to);
839         if (chg < 0)
840                 return chg;
841         /*
842          * When cpuset is configured, it breaks the strict hugetlb page
843          * reservation as the accounting is done on a global variable. Such
844          * reservation is completely rubbish in the presence of cpuset because
845          * the reservation is not checked against page availability for the
846          * current cpuset. Application can still potentially OOM'ed by kernel
847          * with lack of free htlb page in cpuset that the task is in.
848          * Attempt to enforce strict accounting with cpuset is almost
849          * impossible (or too ugly) because cpuset is too fluid that
850          * task or memory node can be dynamically moved between cpusets.
851          *
852          * The change of semantics for shared hugetlb mapping with cpuset is
853          * undesirable. However, in order to preserve some of the semantics,
854          * we fall back to check against current free page availability as
855          * a best attempt and hopefully to minimize the impact of changing
856          * semantics that cpuset has.
857          */
858         if (chg > cpuset_mems_nr(free_huge_pages_node))
859                 return -ENOMEM;
860
861         ret = hugetlb_acct_memory(chg);
862         if (ret < 0)
863                 return ret;
864         region_add(&inode->i_mapping->private_list, from, to);
865         return 0;
866 }
867
868 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
869 {
870         long chg = region_truncate(&inode->i_mapping->private_list, offset);
871         hugetlb_acct_memory(freed - chg);
872 }