Merge master.kernel.org:/pub/scm/linux/kernel/git/gregkh/usb-2.6
[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 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
31 unsigned long hugepages_treat_as_movable;
32
33 /*
34  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
35  */
36 static DEFINE_SPINLOCK(hugetlb_lock);
37
38 static void clear_huge_page(struct page *page, unsigned long addr)
39 {
40         int i;
41
42         might_sleep();
43         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
44                 cond_resched();
45                 clear_user_highpage(page + i, addr);
46         }
47 }
48
49 static void copy_huge_page(struct page *dst, struct page *src,
50                            unsigned long addr, struct vm_area_struct *vma)
51 {
52         int i;
53
54         might_sleep();
55         for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
56                 cond_resched();
57                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
58         }
59 }
60
61 static void enqueue_huge_page(struct page *page)
62 {
63         int nid = page_to_nid(page);
64         list_add(&page->lru, &hugepage_freelists[nid]);
65         free_huge_pages++;
66         free_huge_pages_node[nid]++;
67 }
68
69 static struct page *dequeue_huge_page(struct vm_area_struct *vma,
70                                 unsigned long address)
71 {
72         int nid;
73         struct page *page = NULL;
74         struct zonelist *zonelist = huge_zonelist(vma, address,
75                                                 htlb_alloc_mask);
76         struct zone **z;
77
78         for (z = zonelist->zones; *z; z++) {
79                 nid = zone_to_nid(*z);
80                 if (cpuset_zone_allowed_softwall(*z, htlb_alloc_mask) &&
81                     !list_empty(&hugepage_freelists[nid])) {
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                         break;
88                 }
89         }
90         return page;
91 }
92
93 static void free_huge_page(struct page *page)
94 {
95         BUG_ON(page_count(page));
96
97         INIT_LIST_HEAD(&page->lru);
98
99         spin_lock(&hugetlb_lock);
100         enqueue_huge_page(page);
101         spin_unlock(&hugetlb_lock);
102 }
103
104 static int alloc_fresh_huge_page(void)
105 {
106         static int prev_nid;
107         struct page *page;
108         int nid;
109
110         /*
111          * Copy static prev_nid to local nid, work on that, then copy it
112          * back to prev_nid afterwards: otherwise there's a window in which
113          * a racer might pass invalid nid MAX_NUMNODES to alloc_pages_node.
114          * But we don't need to use a spin_lock here: it really doesn't
115          * matter if occasionally a racer chooses the same nid as we do.
116          */
117         nid = next_node(prev_nid, node_online_map);
118         if (nid == MAX_NUMNODES)
119                 nid = first_node(node_online_map);
120         prev_nid = nid;
121
122         page = alloc_pages_node(nid, htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
123                                         HUGETLB_PAGE_ORDER);
124         if (page) {
125                 set_compound_page_dtor(page, free_huge_page);
126                 spin_lock(&hugetlb_lock);
127                 nr_huge_pages++;
128                 nr_huge_pages_node[page_to_nid(page)]++;
129                 spin_unlock(&hugetlb_lock);
130                 put_page(page); /* free it into the hugepage allocator */
131                 return 1;
132         }
133         return 0;
134 }
135
136 static struct page *alloc_huge_page(struct vm_area_struct *vma,
137                                     unsigned long addr)
138 {
139         struct page *page;
140
141         spin_lock(&hugetlb_lock);
142         if (vma->vm_flags & VM_MAYSHARE)
143                 resv_huge_pages--;
144         else if (free_huge_pages <= resv_huge_pages)
145                 goto fail;
146
147         page = dequeue_huge_page(vma, addr);
148         if (!page)
149                 goto fail;
150
151         spin_unlock(&hugetlb_lock);
152         set_page_refcounted(page);
153         return page;
154
155 fail:
156         if (vma->vm_flags & VM_MAYSHARE)
157                 resv_huge_pages++;
158         spin_unlock(&hugetlb_lock);
159         return NULL;
160 }
161
162 static int __init hugetlb_init(void)
163 {
164         unsigned long i;
165
166         if (HPAGE_SHIFT == 0)
167                 return 0;
168
169         for (i = 0; i < MAX_NUMNODES; ++i)
170                 INIT_LIST_HEAD(&hugepage_freelists[i]);
171
172         for (i = 0; i < max_huge_pages; ++i) {
173                 if (!alloc_fresh_huge_page())
174                         break;
175         }
176         max_huge_pages = free_huge_pages = nr_huge_pages = i;
177         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
178         return 0;
179 }
180 module_init(hugetlb_init);
181
182 static int __init hugetlb_setup(char *s)
183 {
184         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
185                 max_huge_pages = 0;
186         return 1;
187 }
188 __setup("hugepages=", hugetlb_setup);
189
190 static unsigned int cpuset_mems_nr(unsigned int *array)
191 {
192         int node;
193         unsigned int nr = 0;
194
195         for_each_node_mask(node, cpuset_current_mems_allowed)
196                 nr += array[node];
197
198         return nr;
199 }
200
201 #ifdef CONFIG_SYSCTL
202 static void update_and_free_page(struct page *page)
203 {
204         int i;
205         nr_huge_pages--;
206         nr_huge_pages_node[page_to_nid(page)]--;
207         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
208                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
209                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
210                                 1 << PG_private | 1<< PG_writeback);
211         }
212         set_compound_page_dtor(page, NULL);
213         set_page_refcounted(page);
214         __free_pages(page, HUGETLB_PAGE_ORDER);
215 }
216
217 #ifdef CONFIG_HIGHMEM
218 static void try_to_free_low(unsigned long count)
219 {
220         int i;
221
222         for (i = 0; i < MAX_NUMNODES; ++i) {
223                 struct page *page, *next;
224                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
225                         if (PageHighMem(page))
226                                 continue;
227                         list_del(&page->lru);
228                         update_and_free_page(page);
229                         free_huge_pages--;
230                         free_huge_pages_node[page_to_nid(page)]--;
231                         if (count >= nr_huge_pages)
232                                 return;
233                 }
234         }
235 }
236 #else
237 static inline void try_to_free_low(unsigned long count)
238 {
239 }
240 #endif
241
242 static unsigned long set_max_huge_pages(unsigned long count)
243 {
244         while (count > nr_huge_pages) {
245                 if (!alloc_fresh_huge_page())
246                         return nr_huge_pages;
247         }
248         if (count >= nr_huge_pages)
249                 return nr_huge_pages;
250
251         spin_lock(&hugetlb_lock);
252         count = max(count, resv_huge_pages);
253         try_to_free_low(count);
254         while (count < nr_huge_pages) {
255                 struct page *page = dequeue_huge_page(NULL, 0);
256                 if (!page)
257                         break;
258                 update_and_free_page(page);
259         }
260         spin_unlock(&hugetlb_lock);
261         return nr_huge_pages;
262 }
263
264 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
265                            struct file *file, void __user *buffer,
266                            size_t *length, loff_t *ppos)
267 {
268         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
269         max_huge_pages = set_max_huge_pages(max_huge_pages);
270         return 0;
271 }
272
273 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
274                         struct file *file, void __user *buffer,
275                         size_t *length, loff_t *ppos)
276 {
277         proc_dointvec(table, write, file, buffer, length, ppos);
278         if (hugepages_treat_as_movable)
279                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
280         else
281                 htlb_alloc_mask = GFP_HIGHUSER;
282         return 0;
283 }
284
285 #endif /* CONFIG_SYSCTL */
286
287 int hugetlb_report_meminfo(char *buf)
288 {
289         return sprintf(buf,
290                         "HugePages_Total: %5lu\n"
291                         "HugePages_Free:  %5lu\n"
292                         "HugePages_Rsvd:  %5lu\n"
293                         "Hugepagesize:    %5lu kB\n",
294                         nr_huge_pages,
295                         free_huge_pages,
296                         resv_huge_pages,
297                         HPAGE_SIZE/1024);
298 }
299
300 int hugetlb_report_node_meminfo(int nid, char *buf)
301 {
302         return sprintf(buf,
303                 "Node %d HugePages_Total: %5u\n"
304                 "Node %d HugePages_Free:  %5u\n",
305                 nid, nr_huge_pages_node[nid],
306                 nid, free_huge_pages_node[nid]);
307 }
308
309 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
310 unsigned long hugetlb_total_pages(void)
311 {
312         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
313 }
314
315 /*
316  * We cannot handle pagefaults against hugetlb pages at all.  They cause
317  * handle_mm_fault() to try to instantiate regular-sized pages in the
318  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
319  * this far.
320  */
321 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
322 {
323         BUG();
324         return 0;
325 }
326
327 struct vm_operations_struct hugetlb_vm_ops = {
328         .fault = hugetlb_vm_op_fault,
329 };
330
331 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
332                                 int writable)
333 {
334         pte_t entry;
335
336         if (writable) {
337                 entry =
338                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
339         } else {
340                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
341         }
342         entry = pte_mkyoung(entry);
343         entry = pte_mkhuge(entry);
344
345         return entry;
346 }
347
348 static void set_huge_ptep_writable(struct vm_area_struct *vma,
349                                    unsigned long address, pte_t *ptep)
350 {
351         pte_t entry;
352
353         entry = pte_mkwrite(pte_mkdirty(*ptep));
354         if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
355                 update_mmu_cache(vma, address, entry);
356                 lazy_mmu_prot_update(entry);
357         }
358 }
359
360
361 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
362                             struct vm_area_struct *vma)
363 {
364         pte_t *src_pte, *dst_pte, entry;
365         struct page *ptepage;
366         unsigned long addr;
367         int cow;
368
369         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
370
371         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
372                 src_pte = huge_pte_offset(src, addr);
373                 if (!src_pte)
374                         continue;
375                 dst_pte = huge_pte_alloc(dst, addr);
376                 if (!dst_pte)
377                         goto nomem;
378                 spin_lock(&dst->page_table_lock);
379                 spin_lock(&src->page_table_lock);
380                 if (!pte_none(*src_pte)) {
381                         if (cow)
382                                 ptep_set_wrprotect(src, addr, src_pte);
383                         entry = *src_pte;
384                         ptepage = pte_page(entry);
385                         get_page(ptepage);
386                         set_huge_pte_at(dst, addr, dst_pte, entry);
387                 }
388                 spin_unlock(&src->page_table_lock);
389                 spin_unlock(&dst->page_table_lock);
390         }
391         return 0;
392
393 nomem:
394         return -ENOMEM;
395 }
396
397 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
398                             unsigned long end)
399 {
400         struct mm_struct *mm = vma->vm_mm;
401         unsigned long address;
402         pte_t *ptep;
403         pte_t pte;
404         struct page *page;
405         struct page *tmp;
406         /*
407          * A page gathering list, protected by per file i_mmap_lock. The
408          * lock is used to avoid list corruption from multiple unmapping
409          * of the same page since we are using page->lru.
410          */
411         LIST_HEAD(page_list);
412
413         WARN_ON(!is_vm_hugetlb_page(vma));
414         BUG_ON(start & ~HPAGE_MASK);
415         BUG_ON(end & ~HPAGE_MASK);
416
417         spin_lock(&mm->page_table_lock);
418         for (address = start; address < end; address += HPAGE_SIZE) {
419                 ptep = huge_pte_offset(mm, address);
420                 if (!ptep)
421                         continue;
422
423                 if (huge_pmd_unshare(mm, &address, ptep))
424                         continue;
425
426                 pte = huge_ptep_get_and_clear(mm, address, ptep);
427                 if (pte_none(pte))
428                         continue;
429
430                 page = pte_page(pte);
431                 if (pte_dirty(pte))
432                         set_page_dirty(page);
433                 list_add(&page->lru, &page_list);
434         }
435         spin_unlock(&mm->page_table_lock);
436         flush_tlb_range(vma, start, end);
437         list_for_each_entry_safe(page, tmp, &page_list, lru) {
438                 list_del(&page->lru);
439                 put_page(page);
440         }
441 }
442
443 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
444                           unsigned long end)
445 {
446         /*
447          * It is undesirable to test vma->vm_file as it should be non-null
448          * for valid hugetlb area. However, vm_file will be NULL in the error
449          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
450          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
451          * to clean up. Since no pte has actually been setup, it is safe to
452          * do nothing in this case.
453          */
454         if (vma->vm_file) {
455                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
456                 __unmap_hugepage_range(vma, start, end);
457                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
458         }
459 }
460
461 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
462                         unsigned long address, pte_t *ptep, pte_t pte)
463 {
464         struct page *old_page, *new_page;
465         int avoidcopy;
466
467         old_page = pte_page(pte);
468
469         /* If no-one else is actually using this page, avoid the copy
470          * and just make the page writable */
471         avoidcopy = (page_count(old_page) == 1);
472         if (avoidcopy) {
473                 set_huge_ptep_writable(vma, address, ptep);
474                 return 0;
475         }
476
477         page_cache_get(old_page);
478         new_page = alloc_huge_page(vma, address);
479
480         if (!new_page) {
481                 page_cache_release(old_page);
482                 return VM_FAULT_OOM;
483         }
484
485         spin_unlock(&mm->page_table_lock);
486         copy_huge_page(new_page, old_page, address, vma);
487         spin_lock(&mm->page_table_lock);
488
489         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
490         if (likely(pte_same(*ptep, pte))) {
491                 /* Break COW */
492                 set_huge_pte_at(mm, address, ptep,
493                                 make_huge_pte(vma, new_page, 1));
494                 /* Make the old page be freed below */
495                 new_page = old_page;
496         }
497         page_cache_release(new_page);
498         page_cache_release(old_page);
499         return 0;
500 }
501
502 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
503                         unsigned long address, pte_t *ptep, int write_access)
504 {
505         int ret = VM_FAULT_SIGBUS;
506         unsigned long idx;
507         unsigned long size;
508         struct page *page;
509         struct address_space *mapping;
510         pte_t new_pte;
511
512         mapping = vma->vm_file->f_mapping;
513         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
514                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
515
516         /*
517          * Use page lock to guard against racing truncation
518          * before we get page_table_lock.
519          */
520 retry:
521         page = find_lock_page(mapping, idx);
522         if (!page) {
523                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
524                 if (idx >= size)
525                         goto out;
526                 if (hugetlb_get_quota(mapping))
527                         goto out;
528                 page = alloc_huge_page(vma, address);
529                 if (!page) {
530                         hugetlb_put_quota(mapping);
531                         ret = VM_FAULT_OOM;
532                         goto out;
533                 }
534                 clear_huge_page(page, address);
535
536                 if (vma->vm_flags & VM_SHARED) {
537                         int err;
538
539                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
540                         if (err) {
541                                 put_page(page);
542                                 hugetlb_put_quota(mapping);
543                                 if (err == -EEXIST)
544                                         goto retry;
545                                 goto out;
546                         }
547                 } else
548                         lock_page(page);
549         }
550
551         spin_lock(&mm->page_table_lock);
552         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
553         if (idx >= size)
554                 goto backout;
555
556         ret = 0;
557         if (!pte_none(*ptep))
558                 goto backout;
559
560         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
561                                 && (vma->vm_flags & VM_SHARED)));
562         set_huge_pte_at(mm, address, ptep, new_pte);
563
564         if (write_access && !(vma->vm_flags & VM_SHARED)) {
565                 /* Optimization, do the COW without a second fault */
566                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
567         }
568
569         spin_unlock(&mm->page_table_lock);
570         unlock_page(page);
571 out:
572         return ret;
573
574 backout:
575         spin_unlock(&mm->page_table_lock);
576         hugetlb_put_quota(mapping);
577         unlock_page(page);
578         put_page(page);
579         goto out;
580 }
581
582 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
583                         unsigned long address, int write_access)
584 {
585         pte_t *ptep;
586         pte_t entry;
587         int ret;
588         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
589
590         ptep = huge_pte_alloc(mm, address);
591         if (!ptep)
592                 return VM_FAULT_OOM;
593
594         /*
595          * Serialize hugepage allocation and instantiation, so that we don't
596          * get spurious allocation failures if two CPUs race to instantiate
597          * the same page in the page cache.
598          */
599         mutex_lock(&hugetlb_instantiation_mutex);
600         entry = *ptep;
601         if (pte_none(entry)) {
602                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
603                 mutex_unlock(&hugetlb_instantiation_mutex);
604                 return ret;
605         }
606
607         ret = 0;
608
609         spin_lock(&mm->page_table_lock);
610         /* Check for a racing update before calling hugetlb_cow */
611         if (likely(pte_same(entry, *ptep)))
612                 if (write_access && !pte_write(entry))
613                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
614         spin_unlock(&mm->page_table_lock);
615         mutex_unlock(&hugetlb_instantiation_mutex);
616
617         return ret;
618 }
619
620 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
621                         struct page **pages, struct vm_area_struct **vmas,
622                         unsigned long *position, int *length, int i)
623 {
624         unsigned long pfn_offset;
625         unsigned long vaddr = *position;
626         int remainder = *length;
627
628         spin_lock(&mm->page_table_lock);
629         while (vaddr < vma->vm_end && remainder) {
630                 pte_t *pte;
631                 struct page *page;
632
633                 /*
634                  * Some archs (sparc64, sh*) have multiple pte_ts to
635                  * each hugepage.  We have to make * sure we get the
636                  * first, for the page indexing below to work.
637                  */
638                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
639
640                 if (!pte || pte_none(*pte)) {
641                         int ret;
642
643                         spin_unlock(&mm->page_table_lock);
644                         ret = hugetlb_fault(mm, vma, vaddr, 0);
645                         spin_lock(&mm->page_table_lock);
646                         if (!(ret & VM_FAULT_ERROR))
647                                 continue;
648
649                         remainder = 0;
650                         if (!i)
651                                 i = -EFAULT;
652                         break;
653                 }
654
655                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
656                 page = pte_page(*pte);
657 same_page:
658                 if (pages) {
659                         get_page(page);
660                         pages[i] = page + pfn_offset;
661                 }
662
663                 if (vmas)
664                         vmas[i] = vma;
665
666                 vaddr += PAGE_SIZE;
667                 ++pfn_offset;
668                 --remainder;
669                 ++i;
670                 if (vaddr < vma->vm_end && remainder &&
671                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
672                         /*
673                          * We use pfn_offset to avoid touching the pageframes
674                          * of this compound page.
675                          */
676                         goto same_page;
677                 }
678         }
679         spin_unlock(&mm->page_table_lock);
680         *length = remainder;
681         *position = vaddr;
682
683         return i;
684 }
685
686 void hugetlb_change_protection(struct vm_area_struct *vma,
687                 unsigned long address, unsigned long end, pgprot_t newprot)
688 {
689         struct mm_struct *mm = vma->vm_mm;
690         unsigned long start = address;
691         pte_t *ptep;
692         pte_t pte;
693
694         BUG_ON(address >= end);
695         flush_cache_range(vma, address, end);
696
697         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
698         spin_lock(&mm->page_table_lock);
699         for (; address < end; address += HPAGE_SIZE) {
700                 ptep = huge_pte_offset(mm, address);
701                 if (!ptep)
702                         continue;
703                 if (huge_pmd_unshare(mm, &address, ptep))
704                         continue;
705                 if (!pte_none(*ptep)) {
706                         pte = huge_ptep_get_and_clear(mm, address, ptep);
707                         pte = pte_mkhuge(pte_modify(pte, newprot));
708                         set_huge_pte_at(mm, address, ptep, pte);
709                         lazy_mmu_prot_update(pte);
710                 }
711         }
712         spin_unlock(&mm->page_table_lock);
713         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
714
715         flush_tlb_range(vma, start, end);
716 }
717
718 struct file_region {
719         struct list_head link;
720         long from;
721         long to;
722 };
723
724 static long region_add(struct list_head *head, long f, long t)
725 {
726         struct file_region *rg, *nrg, *trg;
727
728         /* Locate the region we are either in or before. */
729         list_for_each_entry(rg, head, link)
730                 if (f <= rg->to)
731                         break;
732
733         /* Round our left edge to the current segment if it encloses us. */
734         if (f > rg->from)
735                 f = rg->from;
736
737         /* Check for and consume any regions we now overlap with. */
738         nrg = rg;
739         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
740                 if (&rg->link == head)
741                         break;
742                 if (rg->from > t)
743                         break;
744
745                 /* If this area reaches higher then extend our area to
746                  * include it completely.  If this is not the first area
747                  * which we intend to reuse, free it. */
748                 if (rg->to > t)
749                         t = rg->to;
750                 if (rg != nrg) {
751                         list_del(&rg->link);
752                         kfree(rg);
753                 }
754         }
755         nrg->from = f;
756         nrg->to = t;
757         return 0;
758 }
759
760 static long region_chg(struct list_head *head, long f, long t)
761 {
762         struct file_region *rg, *nrg;
763         long chg = 0;
764
765         /* Locate the region we are before or in. */
766         list_for_each_entry(rg, head, link)
767                 if (f <= rg->to)
768                         break;
769
770         /* If we are below the current region then a new region is required.
771          * Subtle, allocate a new region at the position but make it zero
772          * size such that we can guarentee to record the reservation. */
773         if (&rg->link == head || t < rg->from) {
774                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
775                 if (nrg == 0)
776                         return -ENOMEM;
777                 nrg->from = f;
778                 nrg->to   = f;
779                 INIT_LIST_HEAD(&nrg->link);
780                 list_add(&nrg->link, rg->link.prev);
781
782                 return t - f;
783         }
784
785         /* Round our left edge to the current segment if it encloses us. */
786         if (f > rg->from)
787                 f = rg->from;
788         chg = t - f;
789
790         /* Check for and consume any regions we now overlap with. */
791         list_for_each_entry(rg, rg->link.prev, link) {
792                 if (&rg->link == head)
793                         break;
794                 if (rg->from > t)
795                         return chg;
796
797                 /* We overlap with this area, if it extends futher than
798                  * us then we must extend ourselves.  Account for its
799                  * existing reservation. */
800                 if (rg->to > t) {
801                         chg += rg->to - t;
802                         t = rg->to;
803                 }
804                 chg -= rg->to - rg->from;
805         }
806         return chg;
807 }
808
809 static long region_truncate(struct list_head *head, long end)
810 {
811         struct file_region *rg, *trg;
812         long chg = 0;
813
814         /* Locate the region we are either in or before. */
815         list_for_each_entry(rg, head, link)
816                 if (end <= rg->to)
817                         break;
818         if (&rg->link == head)
819                 return 0;
820
821         /* If we are in the middle of a region then adjust it. */
822         if (end > rg->from) {
823                 chg = rg->to - end;
824                 rg->to = end;
825                 rg = list_entry(rg->link.next, typeof(*rg), link);
826         }
827
828         /* Drop any remaining regions. */
829         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
830                 if (&rg->link == head)
831                         break;
832                 chg += rg->to - rg->from;
833                 list_del(&rg->link);
834                 kfree(rg);
835         }
836         return chg;
837 }
838
839 static int hugetlb_acct_memory(long delta)
840 {
841         int ret = -ENOMEM;
842
843         spin_lock(&hugetlb_lock);
844         if ((delta + resv_huge_pages) <= free_huge_pages) {
845                 resv_huge_pages += delta;
846                 ret = 0;
847         }
848         spin_unlock(&hugetlb_lock);
849         return ret;
850 }
851
852 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
853 {
854         long ret, chg;
855
856         chg = region_chg(&inode->i_mapping->private_list, from, to);
857         if (chg < 0)
858                 return chg;
859         /*
860          * When cpuset is configured, it breaks the strict hugetlb page
861          * reservation as the accounting is done on a global variable. Such
862          * reservation is completely rubbish in the presence of cpuset because
863          * the reservation is not checked against page availability for the
864          * current cpuset. Application can still potentially OOM'ed by kernel
865          * with lack of free htlb page in cpuset that the task is in.
866          * Attempt to enforce strict accounting with cpuset is almost
867          * impossible (or too ugly) because cpuset is too fluid that
868          * task or memory node can be dynamically moved between cpusets.
869          *
870          * The change of semantics for shared hugetlb mapping with cpuset is
871          * undesirable. However, in order to preserve some of the semantics,
872          * we fall back to check against current free page availability as
873          * a best attempt and hopefully to minimize the impact of changing
874          * semantics that cpuset has.
875          */
876         if (chg > cpuset_mems_nr(free_huge_pages_node))
877                 return -ENOMEM;
878
879         ret = hugetlb_acct_memory(chg);
880         if (ret < 0)
881                 return ret;
882         region_add(&inode->i_mapping->private_list, from, to);
883         return 0;
884 }
885
886 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
887 {
888         long chg = region_truncate(&inode->i_mapping->private_list, offset);
889         hugetlb_acct_memory(freed - chg);
890 }