Rework ptep_set_access_flags and fix sun4c
[sfrench/cifs-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, 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 = 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         if (vma->vm_flags & VM_MAYSHARE)
144                 resv_huge_pages++;
145         spin_unlock(&hugetlb_lock);
146         return NULL;
147 }
148
149 static int __init hugetlb_init(void)
150 {
151         unsigned long i;
152
153         if (HPAGE_SHIFT == 0)
154                 return 0;
155
156         for (i = 0; i < MAX_NUMNODES; ++i)
157                 INIT_LIST_HEAD(&hugepage_freelists[i]);
158
159         for (i = 0; i < max_huge_pages; ++i) {
160                 if (!alloc_fresh_huge_page())
161                         break;
162         }
163         max_huge_pages = free_huge_pages = nr_huge_pages = i;
164         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
165         return 0;
166 }
167 module_init(hugetlb_init);
168
169 static int __init hugetlb_setup(char *s)
170 {
171         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
172                 max_huge_pages = 0;
173         return 1;
174 }
175 __setup("hugepages=", hugetlb_setup);
176
177 static unsigned int cpuset_mems_nr(unsigned int *array)
178 {
179         int node;
180         unsigned int nr = 0;
181
182         for_each_node_mask(node, cpuset_current_mems_allowed)
183                 nr += array[node];
184
185         return nr;
186 }
187
188 #ifdef CONFIG_SYSCTL
189 static void update_and_free_page(struct page *page)
190 {
191         int i;
192         nr_huge_pages--;
193         nr_huge_pages_node[page_to_nid(page)]--;
194         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
195                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
196                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
197                                 1 << PG_private | 1<< PG_writeback);
198         }
199         page[1].lru.next = NULL;
200         set_page_refcounted(page);
201         __free_pages(page, HUGETLB_PAGE_ORDER);
202 }
203
204 #ifdef CONFIG_HIGHMEM
205 static void try_to_free_low(unsigned long count)
206 {
207         int i;
208
209         for (i = 0; i < MAX_NUMNODES; ++i) {
210                 struct page *page, *next;
211                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
212                         if (PageHighMem(page))
213                                 continue;
214                         list_del(&page->lru);
215                         update_and_free_page(page);
216                         free_huge_pages--;
217                         free_huge_pages_node[page_to_nid(page)]--;
218                         if (count >= nr_huge_pages)
219                                 return;
220                 }
221         }
222 }
223 #else
224 static inline void try_to_free_low(unsigned long count)
225 {
226 }
227 #endif
228
229 static unsigned long set_max_huge_pages(unsigned long count)
230 {
231         while (count > nr_huge_pages) {
232                 if (!alloc_fresh_huge_page())
233                         return nr_huge_pages;
234         }
235         if (count >= nr_huge_pages)
236                 return nr_huge_pages;
237
238         spin_lock(&hugetlb_lock);
239         count = max(count, resv_huge_pages);
240         try_to_free_low(count);
241         while (count < nr_huge_pages) {
242                 struct page *page = dequeue_huge_page(NULL, 0);
243                 if (!page)
244                         break;
245                 update_and_free_page(page);
246         }
247         spin_unlock(&hugetlb_lock);
248         return nr_huge_pages;
249 }
250
251 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
252                            struct file *file, void __user *buffer,
253                            size_t *length, loff_t *ppos)
254 {
255         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
256         max_huge_pages = set_max_huge_pages(max_huge_pages);
257         return 0;
258 }
259 #endif /* CONFIG_SYSCTL */
260
261 int hugetlb_report_meminfo(char *buf)
262 {
263         return sprintf(buf,
264                         "HugePages_Total: %5lu\n"
265                         "HugePages_Free:  %5lu\n"
266                         "HugePages_Rsvd:  %5lu\n"
267                         "Hugepagesize:    %5lu kB\n",
268                         nr_huge_pages,
269                         free_huge_pages,
270                         resv_huge_pages,
271                         HPAGE_SIZE/1024);
272 }
273
274 int hugetlb_report_node_meminfo(int nid, char *buf)
275 {
276         return sprintf(buf,
277                 "Node %d HugePages_Total: %5u\n"
278                 "Node %d HugePages_Free:  %5u\n",
279                 nid, nr_huge_pages_node[nid],
280                 nid, free_huge_pages_node[nid]);
281 }
282
283 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
284 unsigned long hugetlb_total_pages(void)
285 {
286         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
287 }
288
289 /*
290  * We cannot handle pagefaults against hugetlb pages at all.  They cause
291  * handle_mm_fault() to try to instantiate regular-sized pages in the
292  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
293  * this far.
294  */
295 static struct page *hugetlb_nopage(struct vm_area_struct *vma,
296                                 unsigned long address, int *unused)
297 {
298         BUG();
299         return NULL;
300 }
301
302 struct vm_operations_struct hugetlb_vm_ops = {
303         .nopage = hugetlb_nopage,
304 };
305
306 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
307                                 int writable)
308 {
309         pte_t entry;
310
311         if (writable) {
312                 entry =
313                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
314         } else {
315                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
316         }
317         entry = pte_mkyoung(entry);
318         entry = pte_mkhuge(entry);
319
320         return entry;
321 }
322
323 static void set_huge_ptep_writable(struct vm_area_struct *vma,
324                                    unsigned long address, pte_t *ptep)
325 {
326         pte_t entry;
327
328         entry = pte_mkwrite(pte_mkdirty(*ptep));
329         if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
330                 update_mmu_cache(vma, address, entry);
331                 lazy_mmu_prot_update(entry);
332         }
333 }
334
335
336 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
337                             struct vm_area_struct *vma)
338 {
339         pte_t *src_pte, *dst_pte, entry;
340         struct page *ptepage;
341         unsigned long addr;
342         int cow;
343
344         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
345
346         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
347                 src_pte = huge_pte_offset(src, addr);
348                 if (!src_pte)
349                         continue;
350                 dst_pte = huge_pte_alloc(dst, addr);
351                 if (!dst_pte)
352                         goto nomem;
353                 spin_lock(&dst->page_table_lock);
354                 spin_lock(&src->page_table_lock);
355                 if (!pte_none(*src_pte)) {
356                         if (cow)
357                                 ptep_set_wrprotect(src, addr, src_pte);
358                         entry = *src_pte;
359                         ptepage = pte_page(entry);
360                         get_page(ptepage);
361                         set_huge_pte_at(dst, addr, dst_pte, entry);
362                 }
363                 spin_unlock(&src->page_table_lock);
364                 spin_unlock(&dst->page_table_lock);
365         }
366         return 0;
367
368 nomem:
369         return -ENOMEM;
370 }
371
372 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
373                             unsigned long end)
374 {
375         struct mm_struct *mm = vma->vm_mm;
376         unsigned long address;
377         pte_t *ptep;
378         pte_t pte;
379         struct page *page;
380         struct page *tmp;
381         /*
382          * A page gathering list, protected by per file i_mmap_lock. The
383          * lock is used to avoid list corruption from multiple unmapping
384          * of the same page since we are using page->lru.
385          */
386         LIST_HEAD(page_list);
387
388         WARN_ON(!is_vm_hugetlb_page(vma));
389         BUG_ON(start & ~HPAGE_MASK);
390         BUG_ON(end & ~HPAGE_MASK);
391
392         spin_lock(&mm->page_table_lock);
393         for (address = start; address < end; address += HPAGE_SIZE) {
394                 ptep = huge_pte_offset(mm, address);
395                 if (!ptep)
396                         continue;
397
398                 if (huge_pmd_unshare(mm, &address, ptep))
399                         continue;
400
401                 pte = huge_ptep_get_and_clear(mm, address, ptep);
402                 if (pte_none(pte))
403                         continue;
404
405                 page = pte_page(pte);
406                 if (pte_dirty(pte))
407                         set_page_dirty(page);
408                 list_add(&page->lru, &page_list);
409         }
410         spin_unlock(&mm->page_table_lock);
411         flush_tlb_range(vma, start, end);
412         list_for_each_entry_safe(page, tmp, &page_list, lru) {
413                 list_del(&page->lru);
414                 put_page(page);
415         }
416 }
417
418 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
419                           unsigned long end)
420 {
421         /*
422          * It is undesirable to test vma->vm_file as it should be non-null
423          * for valid hugetlb area. However, vm_file will be NULL in the error
424          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
425          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
426          * to clean up. Since no pte has actually been setup, it is safe to
427          * do nothing in this case.
428          */
429         if (vma->vm_file) {
430                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
431                 __unmap_hugepage_range(vma, start, end);
432                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
433         }
434 }
435
436 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
437                         unsigned long address, pte_t *ptep, pte_t pte)
438 {
439         struct page *old_page, *new_page;
440         int avoidcopy;
441
442         old_page = pte_page(pte);
443
444         /* If no-one else is actually using this page, avoid the copy
445          * and just make the page writable */
446         avoidcopy = (page_count(old_page) == 1);
447         if (avoidcopy) {
448                 set_huge_ptep_writable(vma, address, ptep);
449                 return VM_FAULT_MINOR;
450         }
451
452         page_cache_get(old_page);
453         new_page = alloc_huge_page(vma, address);
454
455         if (!new_page) {
456                 page_cache_release(old_page);
457                 return VM_FAULT_OOM;
458         }
459
460         spin_unlock(&mm->page_table_lock);
461         copy_huge_page(new_page, old_page, address, vma);
462         spin_lock(&mm->page_table_lock);
463
464         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
465         if (likely(pte_same(*ptep, pte))) {
466                 /* Break COW */
467                 set_huge_pte_at(mm, address, ptep,
468                                 make_huge_pte(vma, new_page, 1));
469                 /* Make the old page be freed below */
470                 new_page = old_page;
471         }
472         page_cache_release(new_page);
473         page_cache_release(old_page);
474         return VM_FAULT_MINOR;
475 }
476
477 int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
478                         unsigned long address, pte_t *ptep, int write_access)
479 {
480         int ret = VM_FAULT_SIGBUS;
481         unsigned long idx;
482         unsigned long size;
483         struct page *page;
484         struct address_space *mapping;
485         pte_t new_pte;
486
487         mapping = vma->vm_file->f_mapping;
488         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
489                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
490
491         /*
492          * Use page lock to guard against racing truncation
493          * before we get page_table_lock.
494          */
495 retry:
496         page = find_lock_page(mapping, idx);
497         if (!page) {
498                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
499                 if (idx >= size)
500                         goto out;
501                 if (hugetlb_get_quota(mapping))
502                         goto out;
503                 page = alloc_huge_page(vma, address);
504                 if (!page) {
505                         hugetlb_put_quota(mapping);
506                         ret = VM_FAULT_OOM;
507                         goto out;
508                 }
509                 clear_huge_page(page, address);
510
511                 if (vma->vm_flags & VM_SHARED) {
512                         int err;
513
514                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
515                         if (err) {
516                                 put_page(page);
517                                 hugetlb_put_quota(mapping);
518                                 if (err == -EEXIST)
519                                         goto retry;
520                                 goto out;
521                         }
522                 } else
523                         lock_page(page);
524         }
525
526         spin_lock(&mm->page_table_lock);
527         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
528         if (idx >= size)
529                 goto backout;
530
531         ret = VM_FAULT_MINOR;
532         if (!pte_none(*ptep))
533                 goto backout;
534
535         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
536                                 && (vma->vm_flags & VM_SHARED)));
537         set_huge_pte_at(mm, address, ptep, new_pte);
538
539         if (write_access && !(vma->vm_flags & VM_SHARED)) {
540                 /* Optimization, do the COW without a second fault */
541                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
542         }
543
544         spin_unlock(&mm->page_table_lock);
545         unlock_page(page);
546 out:
547         return ret;
548
549 backout:
550         spin_unlock(&mm->page_table_lock);
551         hugetlb_put_quota(mapping);
552         unlock_page(page);
553         put_page(page);
554         goto out;
555 }
556
557 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
558                         unsigned long address, int write_access)
559 {
560         pte_t *ptep;
561         pte_t entry;
562         int ret;
563         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
564
565         ptep = huge_pte_alloc(mm, address);
566         if (!ptep)
567                 return VM_FAULT_OOM;
568
569         /*
570          * Serialize hugepage allocation and instantiation, so that we don't
571          * get spurious allocation failures if two CPUs race to instantiate
572          * the same page in the page cache.
573          */
574         mutex_lock(&hugetlb_instantiation_mutex);
575         entry = *ptep;
576         if (pte_none(entry)) {
577                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
578                 mutex_unlock(&hugetlb_instantiation_mutex);
579                 return ret;
580         }
581
582         ret = VM_FAULT_MINOR;
583
584         spin_lock(&mm->page_table_lock);
585         /* Check for a racing update before calling hugetlb_cow */
586         if (likely(pte_same(entry, *ptep)))
587                 if (write_access && !pte_write(entry))
588                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
589         spin_unlock(&mm->page_table_lock);
590         mutex_unlock(&hugetlb_instantiation_mutex);
591
592         return ret;
593 }
594
595 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
596                         struct page **pages, struct vm_area_struct **vmas,
597                         unsigned long *position, int *length, int i)
598 {
599         unsigned long pfn_offset;
600         unsigned long vaddr = *position;
601         int remainder = *length;
602
603         spin_lock(&mm->page_table_lock);
604         while (vaddr < vma->vm_end && remainder) {
605                 pte_t *pte;
606                 struct page *page;
607
608                 /*
609                  * Some archs (sparc64, sh*) have multiple pte_ts to
610                  * each hugepage.  We have to make * sure we get the
611                  * first, for the page indexing below to work.
612                  */
613                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
614
615                 if (!pte || pte_none(*pte)) {
616                         int ret;
617
618                         spin_unlock(&mm->page_table_lock);
619                         ret = hugetlb_fault(mm, vma, vaddr, 0);
620                         spin_lock(&mm->page_table_lock);
621                         if (ret == VM_FAULT_MINOR)
622                                 continue;
623
624                         remainder = 0;
625                         if (!i)
626                                 i = -EFAULT;
627                         break;
628                 }
629
630                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
631                 page = pte_page(*pte);
632 same_page:
633                 if (pages) {
634                         get_page(page);
635                         pages[i] = page + pfn_offset;
636                 }
637
638                 if (vmas)
639                         vmas[i] = vma;
640
641                 vaddr += PAGE_SIZE;
642                 ++pfn_offset;
643                 --remainder;
644                 ++i;
645                 if (vaddr < vma->vm_end && remainder &&
646                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
647                         /*
648                          * We use pfn_offset to avoid touching the pageframes
649                          * of this compound page.
650                          */
651                         goto same_page;
652                 }
653         }
654         spin_unlock(&mm->page_table_lock);
655         *length = remainder;
656         *position = vaddr;
657
658         return i;
659 }
660
661 void hugetlb_change_protection(struct vm_area_struct *vma,
662                 unsigned long address, unsigned long end, pgprot_t newprot)
663 {
664         struct mm_struct *mm = vma->vm_mm;
665         unsigned long start = address;
666         pte_t *ptep;
667         pte_t pte;
668
669         BUG_ON(address >= end);
670         flush_cache_range(vma, address, end);
671
672         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
673         spin_lock(&mm->page_table_lock);
674         for (; address < end; address += HPAGE_SIZE) {
675                 ptep = huge_pte_offset(mm, address);
676                 if (!ptep)
677                         continue;
678                 if (huge_pmd_unshare(mm, &address, ptep))
679                         continue;
680                 if (!pte_none(*ptep)) {
681                         pte = huge_ptep_get_and_clear(mm, address, ptep);
682                         pte = pte_mkhuge(pte_modify(pte, newprot));
683                         set_huge_pte_at(mm, address, ptep, pte);
684                         lazy_mmu_prot_update(pte);
685                 }
686         }
687         spin_unlock(&mm->page_table_lock);
688         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
689
690         flush_tlb_range(vma, start, end);
691 }
692
693 struct file_region {
694         struct list_head link;
695         long from;
696         long to;
697 };
698
699 static long region_add(struct list_head *head, long f, long t)
700 {
701         struct file_region *rg, *nrg, *trg;
702
703         /* Locate the region we are either in or before. */
704         list_for_each_entry(rg, head, link)
705                 if (f <= rg->to)
706                         break;
707
708         /* Round our left edge to the current segment if it encloses us. */
709         if (f > rg->from)
710                 f = rg->from;
711
712         /* Check for and consume any regions we now overlap with. */
713         nrg = rg;
714         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
715                 if (&rg->link == head)
716                         break;
717                 if (rg->from > t)
718                         break;
719
720                 /* If this area reaches higher then extend our area to
721                  * include it completely.  If this is not the first area
722                  * which we intend to reuse, free it. */
723                 if (rg->to > t)
724                         t = rg->to;
725                 if (rg != nrg) {
726                         list_del(&rg->link);
727                         kfree(rg);
728                 }
729         }
730         nrg->from = f;
731         nrg->to = t;
732         return 0;
733 }
734
735 static long region_chg(struct list_head *head, long f, long t)
736 {
737         struct file_region *rg, *nrg;
738         long chg = 0;
739
740         /* Locate the region we are before or in. */
741         list_for_each_entry(rg, head, link)
742                 if (f <= rg->to)
743                         break;
744
745         /* If we are below the current region then a new region is required.
746          * Subtle, allocate a new region at the position but make it zero
747          * size such that we can guarentee to record the reservation. */
748         if (&rg->link == head || t < rg->from) {
749                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
750                 if (nrg == 0)
751                         return -ENOMEM;
752                 nrg->from = f;
753                 nrg->to   = f;
754                 INIT_LIST_HEAD(&nrg->link);
755                 list_add(&nrg->link, rg->link.prev);
756
757                 return t - f;
758         }
759
760         /* Round our left edge to the current segment if it encloses us. */
761         if (f > rg->from)
762                 f = rg->from;
763         chg = t - f;
764
765         /* Check for and consume any regions we now overlap with. */
766         list_for_each_entry(rg, rg->link.prev, link) {
767                 if (&rg->link == head)
768                         break;
769                 if (rg->from > t)
770                         return chg;
771
772                 /* We overlap with this area, if it extends futher than
773                  * us then we must extend ourselves.  Account for its
774                  * existing reservation. */
775                 if (rg->to > t) {
776                         chg += rg->to - t;
777                         t = rg->to;
778                 }
779                 chg -= rg->to - rg->from;
780         }
781         return chg;
782 }
783
784 static long region_truncate(struct list_head *head, long end)
785 {
786         struct file_region *rg, *trg;
787         long chg = 0;
788
789         /* Locate the region we are either in or before. */
790         list_for_each_entry(rg, head, link)
791                 if (end <= rg->to)
792                         break;
793         if (&rg->link == head)
794                 return 0;
795
796         /* If we are in the middle of a region then adjust it. */
797         if (end > rg->from) {
798                 chg = rg->to - end;
799                 rg->to = end;
800                 rg = list_entry(rg->link.next, typeof(*rg), link);
801         }
802
803         /* Drop any remaining regions. */
804         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
805                 if (&rg->link == head)
806                         break;
807                 chg += rg->to - rg->from;
808                 list_del(&rg->link);
809                 kfree(rg);
810         }
811         return chg;
812 }
813
814 static int hugetlb_acct_memory(long delta)
815 {
816         int ret = -ENOMEM;
817
818         spin_lock(&hugetlb_lock);
819         if ((delta + resv_huge_pages) <= free_huge_pages) {
820                 resv_huge_pages += delta;
821                 ret = 0;
822         }
823         spin_unlock(&hugetlb_lock);
824         return ret;
825 }
826
827 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
828 {
829         long ret, chg;
830
831         chg = region_chg(&inode->i_mapping->private_list, from, to);
832         if (chg < 0)
833                 return chg;
834         /*
835          * When cpuset is configured, it breaks the strict hugetlb page
836          * reservation as the accounting is done on a global variable. Such
837          * reservation is completely rubbish in the presence of cpuset because
838          * the reservation is not checked against page availability for the
839          * current cpuset. Application can still potentially OOM'ed by kernel
840          * with lack of free htlb page in cpuset that the task is in.
841          * Attempt to enforce strict accounting with cpuset is almost
842          * impossible (or too ugly) because cpuset is too fluid that
843          * task or memory node can be dynamically moved between cpusets.
844          *
845          * The change of semantics for shared hugetlb mapping with cpuset is
846          * undesirable. However, in order to preserve some of the semantics,
847          * we fall back to check against current free page availability as
848          * a best attempt and hopefully to minimize the impact of changing
849          * semantics that cpuset has.
850          */
851         if (chg > cpuset_mems_nr(free_huge_pages_node))
852                 return -ENOMEM;
853
854         ret = hugetlb_acct_memory(chg);
855         if (ret < 0)
856                 return ret;
857         region_add(&inode->i_mapping->private_list, from, to);
858         return 0;
859 }
860
861 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
862 {
863         long chg = region_truncate(&inode->i_mapping->private_list, offset);
864         hugetlb_acct_memory(freed - chg);
865 }