hugetlb: fix quota management for private mappings
[linux-2.6.git] / mm / hugetlb.c
1 /*
2  * Generic hugetlb support.
3  * (C) William Irwin, April 2004
4  */
5 #include <linux/gfp.h>
6 #include <linux/list.h>
7 #include <linux/init.h>
8 #include <linux/module.h>
9 #include <linux/mm.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/nodemask.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/cpuset.h>
16 #include <linux/mutex.h>
17
18 #include <asm/page.h>
19 #include <asm/pgtable.h>
20
21 #include <linux/hugetlb.h>
22 #include "internal.h"
23
24 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
25 static unsigned long nr_huge_pages, free_huge_pages, resv_huge_pages;
26 static unsigned long surplus_huge_pages;
27 unsigned long max_huge_pages;
28 static struct list_head hugepage_freelists[MAX_NUMNODES];
29 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
30 static unsigned int free_huge_pages_node[MAX_NUMNODES];
31 static unsigned int surplus_huge_pages_node[MAX_NUMNODES];
32 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
33 unsigned long hugepages_treat_as_movable;
34 int hugetlb_dynamic_pool;
35 static int hugetlb_next_nid;
36
37 /*
38  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
39  */
40 static DEFINE_SPINLOCK(hugetlb_lock);
41
42 static void clear_huge_page(struct page *page, unsigned long addr)
43 {
44         int i;
45
46         might_sleep();
47         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
48                 cond_resched();
49                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
50         }
51 }
52
53 static void copy_huge_page(struct page *dst, struct page *src,
54                            unsigned long addr, struct vm_area_struct *vma)
55 {
56         int i;
57
58         might_sleep();
59         for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
60                 cond_resched();
61                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
62         }
63 }
64
65 static void enqueue_huge_page(struct page *page)
66 {
67         int nid = page_to_nid(page);
68         list_add(&page->lru, &hugepage_freelists[nid]);
69         free_huge_pages++;
70         free_huge_pages_node[nid]++;
71 }
72
73 static struct page *dequeue_huge_page(struct vm_area_struct *vma,
74                                 unsigned long address)
75 {
76         int nid;
77         struct page *page = NULL;
78         struct mempolicy *mpol;
79         struct zonelist *zonelist = huge_zonelist(vma, address,
80                                         htlb_alloc_mask, &mpol);
81         struct zone **z;
82
83         for (z = zonelist->zones; *z; z++) {
84                 nid = zone_to_nid(*z);
85                 if (cpuset_zone_allowed_softwall(*z, htlb_alloc_mask) &&
86                     !list_empty(&hugepage_freelists[nid])) {
87                         page = list_entry(hugepage_freelists[nid].next,
88                                           struct page, lru);
89                         list_del(&page->lru);
90                         free_huge_pages--;
91                         free_huge_pages_node[nid]--;
92                         if (vma && vma->vm_flags & VM_MAYSHARE)
93                                 resv_huge_pages--;
94                         break;
95                 }
96         }
97         mpol_free(mpol);        /* unref if mpol !NULL */
98         return page;
99 }
100
101 static void update_and_free_page(struct page *page)
102 {
103         int i;
104         nr_huge_pages--;
105         nr_huge_pages_node[page_to_nid(page)]--;
106         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
107                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
108                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
109                                 1 << PG_private | 1<< PG_writeback);
110         }
111         set_compound_page_dtor(page, NULL);
112         set_page_refcounted(page);
113         __free_pages(page, HUGETLB_PAGE_ORDER);
114 }
115
116 static void free_huge_page(struct page *page)
117 {
118         int nid = page_to_nid(page);
119         struct address_space *mapping;
120
121         mapping = (struct address_space *) page_private(page);
122         BUG_ON(page_count(page));
123         INIT_LIST_HEAD(&page->lru);
124
125         spin_lock(&hugetlb_lock);
126         if (surplus_huge_pages_node[nid]) {
127                 update_and_free_page(page);
128                 surplus_huge_pages--;
129                 surplus_huge_pages_node[nid]--;
130         } else {
131                 enqueue_huge_page(page);
132         }
133         spin_unlock(&hugetlb_lock);
134         if (mapping)
135                 hugetlb_put_quota(mapping);
136         set_page_private(page, 0);
137 }
138
139 /*
140  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
141  * balanced by operating on them in a round-robin fashion.
142  * Returns 1 if an adjustment was made.
143  */
144 static int adjust_pool_surplus(int delta)
145 {
146         static int prev_nid;
147         int nid = prev_nid;
148         int ret = 0;
149
150         VM_BUG_ON(delta != -1 && delta != 1);
151         do {
152                 nid = next_node(nid, node_online_map);
153                 if (nid == MAX_NUMNODES)
154                         nid = first_node(node_online_map);
155
156                 /* To shrink on this node, there must be a surplus page */
157                 if (delta < 0 && !surplus_huge_pages_node[nid])
158                         continue;
159                 /* Surplus cannot exceed the total number of pages */
160                 if (delta > 0 && surplus_huge_pages_node[nid] >=
161                                                 nr_huge_pages_node[nid])
162                         continue;
163
164                 surplus_huge_pages += delta;
165                 surplus_huge_pages_node[nid] += delta;
166                 ret = 1;
167                 break;
168         } while (nid != prev_nid);
169
170         prev_nid = nid;
171         return ret;
172 }
173
174 static struct page *alloc_fresh_huge_page_node(int nid)
175 {
176         struct page *page;
177
178         page = alloc_pages_node(nid,
179                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|__GFP_NOWARN,
180                 HUGETLB_PAGE_ORDER);
181         if (page) {
182                 set_compound_page_dtor(page, free_huge_page);
183                 spin_lock(&hugetlb_lock);
184                 nr_huge_pages++;
185                 nr_huge_pages_node[nid]++;
186                 spin_unlock(&hugetlb_lock);
187                 put_page(page); /* free it into the hugepage allocator */
188         }
189
190         return page;
191 }
192
193 static int alloc_fresh_huge_page(void)
194 {
195         struct page *page;
196         int start_nid;
197         int next_nid;
198         int ret = 0;
199
200         start_nid = hugetlb_next_nid;
201
202         do {
203                 page = alloc_fresh_huge_page_node(hugetlb_next_nid);
204                 if (page)
205                         ret = 1;
206                 /*
207                  * Use a helper variable to find the next node and then
208                  * copy it back to hugetlb_next_nid afterwards:
209                  * otherwise there's a window in which a racer might
210                  * pass invalid nid MAX_NUMNODES to alloc_pages_node.
211                  * But we don't need to use a spin_lock here: it really
212                  * doesn't matter if occasionally a racer chooses the
213                  * same nid as we do.  Move nid forward in the mask even
214                  * if we just successfully allocated a hugepage so that
215                  * the next caller gets hugepages on the next node.
216                  */
217                 next_nid = next_node(hugetlb_next_nid, node_online_map);
218                 if (next_nid == MAX_NUMNODES)
219                         next_nid = first_node(node_online_map);
220                 hugetlb_next_nid = next_nid;
221         } while (!page && hugetlb_next_nid != start_nid);
222
223         return ret;
224 }
225
226 static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
227                                                 unsigned long address)
228 {
229         struct page *page;
230
231         /* Check if the dynamic pool is enabled */
232         if (!hugetlb_dynamic_pool)
233                 return NULL;
234
235         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
236                                         HUGETLB_PAGE_ORDER);
237         if (page) {
238                 set_compound_page_dtor(page, free_huge_page);
239                 spin_lock(&hugetlb_lock);
240                 nr_huge_pages++;
241                 nr_huge_pages_node[page_to_nid(page)]++;
242                 surplus_huge_pages++;
243                 surplus_huge_pages_node[page_to_nid(page)]++;
244                 spin_unlock(&hugetlb_lock);
245         }
246
247         return page;
248 }
249
250 /*
251  * Increase the hugetlb pool such that it can accomodate a reservation
252  * of size 'delta'.
253  */
254 static int gather_surplus_pages(int delta)
255 {
256         struct list_head surplus_list;
257         struct page *page, *tmp;
258         int ret, i;
259         int needed, allocated;
260
261         needed = (resv_huge_pages + delta) - free_huge_pages;
262         if (needed <= 0)
263                 return 0;
264
265         allocated = 0;
266         INIT_LIST_HEAD(&surplus_list);
267
268         ret = -ENOMEM;
269 retry:
270         spin_unlock(&hugetlb_lock);
271         for (i = 0; i < needed; i++) {
272                 page = alloc_buddy_huge_page(NULL, 0);
273                 if (!page) {
274                         /*
275                          * We were not able to allocate enough pages to
276                          * satisfy the entire reservation so we free what
277                          * we've allocated so far.
278                          */
279                         spin_lock(&hugetlb_lock);
280                         needed = 0;
281                         goto free;
282                 }
283
284                 list_add(&page->lru, &surplus_list);
285         }
286         allocated += needed;
287
288         /*
289          * After retaking hugetlb_lock, we need to recalculate 'needed'
290          * because either resv_huge_pages or free_huge_pages may have changed.
291          */
292         spin_lock(&hugetlb_lock);
293         needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
294         if (needed > 0)
295                 goto retry;
296
297         /*
298          * The surplus_list now contains _at_least_ the number of extra pages
299          * needed to accomodate the reservation.  Add the appropriate number
300          * of pages to the hugetlb pool and free the extras back to the buddy
301          * allocator.
302          */
303         needed += allocated;
304         ret = 0;
305 free:
306         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
307                 list_del(&page->lru);
308                 if ((--needed) >= 0)
309                         enqueue_huge_page(page);
310                 else {
311                         /*
312                          * Decrement the refcount and free the page using its
313                          * destructor.  This must be done with hugetlb_lock
314                          * unlocked which is safe because free_huge_page takes
315                          * hugetlb_lock before deciding how to free the page.
316                          */
317                         spin_unlock(&hugetlb_lock);
318                         put_page(page);
319                         spin_lock(&hugetlb_lock);
320                 }
321         }
322
323         return ret;
324 }
325
326 /*
327  * When releasing a hugetlb pool reservation, any surplus pages that were
328  * allocated to satisfy the reservation must be explicitly freed if they were
329  * never used.
330  */
331 void return_unused_surplus_pages(unsigned long unused_resv_pages)
332 {
333         static int nid = -1;
334         struct page *page;
335         unsigned long nr_pages;
336
337         nr_pages = min(unused_resv_pages, surplus_huge_pages);
338
339         while (nr_pages) {
340                 nid = next_node(nid, node_online_map);
341                 if (nid == MAX_NUMNODES)
342                         nid = first_node(node_online_map);
343
344                 if (!surplus_huge_pages_node[nid])
345                         continue;
346
347                 if (!list_empty(&hugepage_freelists[nid])) {
348                         page = list_entry(hugepage_freelists[nid].next,
349                                           struct page, lru);
350                         list_del(&page->lru);
351                         update_and_free_page(page);
352                         free_huge_pages--;
353                         free_huge_pages_node[nid]--;
354                         surplus_huge_pages--;
355                         surplus_huge_pages_node[nid]--;
356                         nr_pages--;
357                 }
358         }
359 }
360
361
362 static struct page *alloc_huge_page_shared(struct vm_area_struct *vma,
363                                                 unsigned long addr)
364 {
365         struct page *page;
366
367         spin_lock(&hugetlb_lock);
368         page = dequeue_huge_page(vma, addr);
369         spin_unlock(&hugetlb_lock);
370         return page;
371 }
372
373 static struct page *alloc_huge_page_private(struct vm_area_struct *vma,
374                                                 unsigned long addr)
375 {
376         struct page *page = NULL;
377
378         spin_lock(&hugetlb_lock);
379         if (free_huge_pages > resv_huge_pages)
380                 page = dequeue_huge_page(vma, addr);
381         spin_unlock(&hugetlb_lock);
382         if (!page)
383                 page = alloc_buddy_huge_page(vma, addr);
384         return page;
385 }
386
387 static struct page *alloc_huge_page(struct vm_area_struct *vma,
388                                     unsigned long addr)
389 {
390         struct page *page;
391
392         if (vma->vm_flags & VM_MAYSHARE)
393                 page = alloc_huge_page_shared(vma, addr);
394         else
395                 page = alloc_huge_page_private(vma, addr);
396         if (page) {
397                 set_page_refcounted(page);
398                 set_page_private(page, (unsigned long) vma->vm_file->f_mapping);
399         }
400         return page;
401 }
402
403 static int __init hugetlb_init(void)
404 {
405         unsigned long i;
406
407         if (HPAGE_SHIFT == 0)
408                 return 0;
409
410         for (i = 0; i < MAX_NUMNODES; ++i)
411                 INIT_LIST_HEAD(&hugepage_freelists[i]);
412
413         hugetlb_next_nid = first_node(node_online_map);
414
415         for (i = 0; i < max_huge_pages; ++i) {
416                 if (!alloc_fresh_huge_page())
417                         break;
418         }
419         max_huge_pages = free_huge_pages = nr_huge_pages = i;
420         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
421         return 0;
422 }
423 module_init(hugetlb_init);
424
425 static int __init hugetlb_setup(char *s)
426 {
427         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
428                 max_huge_pages = 0;
429         return 1;
430 }
431 __setup("hugepages=", hugetlb_setup);
432
433 static unsigned int cpuset_mems_nr(unsigned int *array)
434 {
435         int node;
436         unsigned int nr = 0;
437
438         for_each_node_mask(node, cpuset_current_mems_allowed)
439                 nr += array[node];
440
441         return nr;
442 }
443
444 #ifdef CONFIG_SYSCTL
445 #ifdef CONFIG_HIGHMEM
446 static void try_to_free_low(unsigned long count)
447 {
448         int i;
449
450         for (i = 0; i < MAX_NUMNODES; ++i) {
451                 struct page *page, *next;
452                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
453                         if (count >= nr_huge_pages)
454                                 return;
455                         if (PageHighMem(page))
456                                 continue;
457                         list_del(&page->lru);
458                         update_and_free_page(page);
459                         free_huge_pages--;
460                         free_huge_pages_node[page_to_nid(page)]--;
461                 }
462         }
463 }
464 #else
465 static inline void try_to_free_low(unsigned long count)
466 {
467 }
468 #endif
469
470 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
471 static unsigned long set_max_huge_pages(unsigned long count)
472 {
473         unsigned long min_count, ret;
474
475         /*
476          * Increase the pool size
477          * First take pages out of surplus state.  Then make up the
478          * remaining difference by allocating fresh huge pages.
479          */
480         spin_lock(&hugetlb_lock);
481         while (surplus_huge_pages && count > persistent_huge_pages) {
482                 if (!adjust_pool_surplus(-1))
483                         break;
484         }
485
486         while (count > persistent_huge_pages) {
487                 int ret;
488                 /*
489                  * If this allocation races such that we no longer need the
490                  * page, free_huge_page will handle it by freeing the page
491                  * and reducing the surplus.
492                  */
493                 spin_unlock(&hugetlb_lock);
494                 ret = alloc_fresh_huge_page();
495                 spin_lock(&hugetlb_lock);
496                 if (!ret)
497                         goto out;
498
499         }
500
501         /*
502          * Decrease the pool size
503          * First return free pages to the buddy allocator (being careful
504          * to keep enough around to satisfy reservations).  Then place
505          * pages into surplus state as needed so the pool will shrink
506          * to the desired size as pages become free.
507          */
508         min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
509         min_count = max(count, min_count);
510         try_to_free_low(min_count);
511         while (min_count < persistent_huge_pages) {
512                 struct page *page = dequeue_huge_page(NULL, 0);
513                 if (!page)
514                         break;
515                 update_and_free_page(page);
516         }
517         while (count < persistent_huge_pages) {
518                 if (!adjust_pool_surplus(1))
519                         break;
520         }
521 out:
522         ret = persistent_huge_pages;
523         spin_unlock(&hugetlb_lock);
524         return ret;
525 }
526
527 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
528                            struct file *file, void __user *buffer,
529                            size_t *length, loff_t *ppos)
530 {
531         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
532         max_huge_pages = set_max_huge_pages(max_huge_pages);
533         return 0;
534 }
535
536 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
537                         struct file *file, void __user *buffer,
538                         size_t *length, loff_t *ppos)
539 {
540         proc_dointvec(table, write, file, buffer, length, ppos);
541         if (hugepages_treat_as_movable)
542                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
543         else
544                 htlb_alloc_mask = GFP_HIGHUSER;
545         return 0;
546 }
547
548 #endif /* CONFIG_SYSCTL */
549
550 int hugetlb_report_meminfo(char *buf)
551 {
552         return sprintf(buf,
553                         "HugePages_Total: %5lu\n"
554                         "HugePages_Free:  %5lu\n"
555                         "HugePages_Rsvd:  %5lu\n"
556                         "HugePages_Surp:  %5lu\n"
557                         "Hugepagesize:    %5lu kB\n",
558                         nr_huge_pages,
559                         free_huge_pages,
560                         resv_huge_pages,
561                         surplus_huge_pages,
562                         HPAGE_SIZE/1024);
563 }
564
565 int hugetlb_report_node_meminfo(int nid, char *buf)
566 {
567         return sprintf(buf,
568                 "Node %d HugePages_Total: %5u\n"
569                 "Node %d HugePages_Free:  %5u\n",
570                 nid, nr_huge_pages_node[nid],
571                 nid, free_huge_pages_node[nid]);
572 }
573
574 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
575 unsigned long hugetlb_total_pages(void)
576 {
577         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
578 }
579
580 /*
581  * We cannot handle pagefaults against hugetlb pages at all.  They cause
582  * handle_mm_fault() to try to instantiate regular-sized pages in the
583  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
584  * this far.
585  */
586 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
587 {
588         BUG();
589         return 0;
590 }
591
592 struct vm_operations_struct hugetlb_vm_ops = {
593         .fault = hugetlb_vm_op_fault,
594 };
595
596 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
597                                 int writable)
598 {
599         pte_t entry;
600
601         if (writable) {
602                 entry =
603                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
604         } else {
605                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
606         }
607         entry = pte_mkyoung(entry);
608         entry = pte_mkhuge(entry);
609
610         return entry;
611 }
612
613 static void set_huge_ptep_writable(struct vm_area_struct *vma,
614                                    unsigned long address, pte_t *ptep)
615 {
616         pte_t entry;
617
618         entry = pte_mkwrite(pte_mkdirty(*ptep));
619         if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
620                 update_mmu_cache(vma, address, entry);
621         }
622 }
623
624
625 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
626                             struct vm_area_struct *vma)
627 {
628         pte_t *src_pte, *dst_pte, entry;
629         struct page *ptepage;
630         unsigned long addr;
631         int cow;
632
633         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
634
635         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
636                 src_pte = huge_pte_offset(src, addr);
637                 if (!src_pte)
638                         continue;
639                 dst_pte = huge_pte_alloc(dst, addr);
640                 if (!dst_pte)
641                         goto nomem;
642                 spin_lock(&dst->page_table_lock);
643                 spin_lock(&src->page_table_lock);
644                 if (!pte_none(*src_pte)) {
645                         if (cow)
646                                 ptep_set_wrprotect(src, addr, src_pte);
647                         entry = *src_pte;
648                         ptepage = pte_page(entry);
649                         get_page(ptepage);
650                         set_huge_pte_at(dst, addr, dst_pte, entry);
651                 }
652                 spin_unlock(&src->page_table_lock);
653                 spin_unlock(&dst->page_table_lock);
654         }
655         return 0;
656
657 nomem:
658         return -ENOMEM;
659 }
660
661 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
662                             unsigned long end)
663 {
664         struct mm_struct *mm = vma->vm_mm;
665         unsigned long address;
666         pte_t *ptep;
667         pte_t pte;
668         struct page *page;
669         struct page *tmp;
670         /*
671          * A page gathering list, protected by per file i_mmap_lock. The
672          * lock is used to avoid list corruption from multiple unmapping
673          * of the same page since we are using page->lru.
674          */
675         LIST_HEAD(page_list);
676
677         WARN_ON(!is_vm_hugetlb_page(vma));
678         BUG_ON(start & ~HPAGE_MASK);
679         BUG_ON(end & ~HPAGE_MASK);
680
681         spin_lock(&mm->page_table_lock);
682         for (address = start; address < end; address += HPAGE_SIZE) {
683                 ptep = huge_pte_offset(mm, address);
684                 if (!ptep)
685                         continue;
686
687                 if (huge_pmd_unshare(mm, &address, ptep))
688                         continue;
689
690                 pte = huge_ptep_get_and_clear(mm, address, ptep);
691                 if (pte_none(pte))
692                         continue;
693
694                 page = pte_page(pte);
695                 if (pte_dirty(pte))
696                         set_page_dirty(page);
697                 list_add(&page->lru, &page_list);
698         }
699         spin_unlock(&mm->page_table_lock);
700         flush_tlb_range(vma, start, end);
701         list_for_each_entry_safe(page, tmp, &page_list, lru) {
702                 list_del(&page->lru);
703                 put_page(page);
704         }
705 }
706
707 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
708                           unsigned long end)
709 {
710         /*
711          * It is undesirable to test vma->vm_file as it should be non-null
712          * for valid hugetlb area. However, vm_file will be NULL in the error
713          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
714          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
715          * to clean up. Since no pte has actually been setup, it is safe to
716          * do nothing in this case.
717          */
718         if (vma->vm_file) {
719                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
720                 __unmap_hugepage_range(vma, start, end);
721                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
722         }
723 }
724
725 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
726                         unsigned long address, pte_t *ptep, pte_t pte)
727 {
728         struct page *old_page, *new_page;
729         int avoidcopy;
730
731         old_page = pte_page(pte);
732
733         /* If no-one else is actually using this page, avoid the copy
734          * and just make the page writable */
735         avoidcopy = (page_count(old_page) == 1);
736         if (avoidcopy) {
737                 set_huge_ptep_writable(vma, address, ptep);
738                 return 0;
739         }
740         if (hugetlb_get_quota(vma->vm_file->f_mapping))
741                 return VM_FAULT_SIGBUS;
742
743         page_cache_get(old_page);
744         new_page = alloc_huge_page(vma, address);
745
746         if (!new_page) {
747                 page_cache_release(old_page);
748                 return VM_FAULT_OOM;
749         }
750
751         spin_unlock(&mm->page_table_lock);
752         copy_huge_page(new_page, old_page, address, vma);
753         spin_lock(&mm->page_table_lock);
754
755         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
756         if (likely(pte_same(*ptep, pte))) {
757                 /* Break COW */
758                 set_huge_pte_at(mm, address, ptep,
759                                 make_huge_pte(vma, new_page, 1));
760                 /* Make the old page be freed below */
761                 new_page = old_page;
762         }
763         page_cache_release(new_page);
764         page_cache_release(old_page);
765         return 0;
766 }
767
768 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
769                         unsigned long address, pte_t *ptep, int write_access)
770 {
771         int ret = VM_FAULT_SIGBUS;
772         unsigned long idx;
773         unsigned long size;
774         struct page *page;
775         struct address_space *mapping;
776         pte_t new_pte;
777
778         mapping = vma->vm_file->f_mapping;
779         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
780                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
781
782         /*
783          * Use page lock to guard against racing truncation
784          * before we get page_table_lock.
785          */
786 retry:
787         page = find_lock_page(mapping, idx);
788         if (!page) {
789                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
790                 if (idx >= size)
791                         goto out;
792                 if (hugetlb_get_quota(mapping))
793                         goto out;
794                 page = alloc_huge_page(vma, address);
795                 if (!page) {
796                         hugetlb_put_quota(mapping);
797                         ret = VM_FAULT_OOM;
798                         goto out;
799                 }
800                 clear_huge_page(page, address);
801
802                 if (vma->vm_flags & VM_SHARED) {
803                         int err;
804
805                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
806                         if (err) {
807                                 put_page(page);
808                                 if (err == -EEXIST)
809                                         goto retry;
810                                 goto out;
811                         }
812                 } else
813                         lock_page(page);
814         }
815
816         spin_lock(&mm->page_table_lock);
817         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
818         if (idx >= size)
819                 goto backout;
820
821         ret = 0;
822         if (!pte_none(*ptep))
823                 goto backout;
824
825         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
826                                 && (vma->vm_flags & VM_SHARED)));
827         set_huge_pte_at(mm, address, ptep, new_pte);
828
829         if (write_access && !(vma->vm_flags & VM_SHARED)) {
830                 /* Optimization, do the COW without a second fault */
831                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
832         }
833
834         spin_unlock(&mm->page_table_lock);
835         unlock_page(page);
836 out:
837         return ret;
838
839 backout:
840         spin_unlock(&mm->page_table_lock);
841         unlock_page(page);
842         put_page(page);
843         goto out;
844 }
845
846 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
847                         unsigned long address, int write_access)
848 {
849         pte_t *ptep;
850         pte_t entry;
851         int ret;
852         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
853
854         ptep = huge_pte_alloc(mm, address);
855         if (!ptep)
856                 return VM_FAULT_OOM;
857
858         /*
859          * Serialize hugepage allocation and instantiation, so that we don't
860          * get spurious allocation failures if two CPUs race to instantiate
861          * the same page in the page cache.
862          */
863         mutex_lock(&hugetlb_instantiation_mutex);
864         entry = *ptep;
865         if (pte_none(entry)) {
866                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
867                 mutex_unlock(&hugetlb_instantiation_mutex);
868                 return ret;
869         }
870
871         ret = 0;
872
873         spin_lock(&mm->page_table_lock);
874         /* Check for a racing update before calling hugetlb_cow */
875         if (likely(pte_same(entry, *ptep)))
876                 if (write_access && !pte_write(entry))
877                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
878         spin_unlock(&mm->page_table_lock);
879         mutex_unlock(&hugetlb_instantiation_mutex);
880
881         return ret;
882 }
883
884 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
885                         struct page **pages, struct vm_area_struct **vmas,
886                         unsigned long *position, int *length, int i,
887                         int write)
888 {
889         unsigned long pfn_offset;
890         unsigned long vaddr = *position;
891         int remainder = *length;
892
893         spin_lock(&mm->page_table_lock);
894         while (vaddr < vma->vm_end && remainder) {
895                 pte_t *pte;
896                 struct page *page;
897
898                 /*
899                  * Some archs (sparc64, sh*) have multiple pte_ts to
900                  * each hugepage.  We have to make * sure we get the
901                  * first, for the page indexing below to work.
902                  */
903                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
904
905                 if (!pte || pte_none(*pte)) {
906                         int ret;
907
908                         spin_unlock(&mm->page_table_lock);
909                         ret = hugetlb_fault(mm, vma, vaddr, write);
910                         spin_lock(&mm->page_table_lock);
911                         if (!(ret & VM_FAULT_ERROR))
912                                 continue;
913
914                         remainder = 0;
915                         if (!i)
916                                 i = -EFAULT;
917                         break;
918                 }
919
920                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
921                 page = pte_page(*pte);
922 same_page:
923                 if (pages) {
924                         get_page(page);
925                         pages[i] = page + pfn_offset;
926                 }
927
928                 if (vmas)
929                         vmas[i] = vma;
930
931                 vaddr += PAGE_SIZE;
932                 ++pfn_offset;
933                 --remainder;
934                 ++i;
935                 if (vaddr < vma->vm_end && remainder &&
936                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
937                         /*
938                          * We use pfn_offset to avoid touching the pageframes
939                          * of this compound page.
940                          */
941                         goto same_page;
942                 }
943         }
944         spin_unlock(&mm->page_table_lock);
945         *length = remainder;
946         *position = vaddr;
947
948         return i;
949 }
950
951 void hugetlb_change_protection(struct vm_area_struct *vma,
952                 unsigned long address, unsigned long end, pgprot_t newprot)
953 {
954         struct mm_struct *mm = vma->vm_mm;
955         unsigned long start = address;
956         pte_t *ptep;
957         pte_t pte;
958
959         BUG_ON(address >= end);
960         flush_cache_range(vma, address, end);
961
962         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
963         spin_lock(&mm->page_table_lock);
964         for (; address < end; address += HPAGE_SIZE) {
965                 ptep = huge_pte_offset(mm, address);
966                 if (!ptep)
967                         continue;
968                 if (huge_pmd_unshare(mm, &address, ptep))
969                         continue;
970                 if (!pte_none(*ptep)) {
971                         pte = huge_ptep_get_and_clear(mm, address, ptep);
972                         pte = pte_mkhuge(pte_modify(pte, newprot));
973                         set_huge_pte_at(mm, address, ptep, pte);
974                 }
975         }
976         spin_unlock(&mm->page_table_lock);
977         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
978
979         flush_tlb_range(vma, start, end);
980 }
981
982 struct file_region {
983         struct list_head link;
984         long from;
985         long to;
986 };
987
988 static long region_add(struct list_head *head, long f, long t)
989 {
990         struct file_region *rg, *nrg, *trg;
991
992         /* Locate the region we are either in or before. */
993         list_for_each_entry(rg, head, link)
994                 if (f <= rg->to)
995                         break;
996
997         /* Round our left edge to the current segment if it encloses us. */
998         if (f > rg->from)
999                 f = rg->from;
1000
1001         /* Check for and consume any regions we now overlap with. */
1002         nrg = rg;
1003         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1004                 if (&rg->link == head)
1005                         break;
1006                 if (rg->from > t)
1007                         break;
1008
1009                 /* If this area reaches higher then extend our area to
1010                  * include it completely.  If this is not the first area
1011                  * which we intend to reuse, free it. */
1012                 if (rg->to > t)
1013                         t = rg->to;
1014                 if (rg != nrg) {
1015                         list_del(&rg->link);
1016                         kfree(rg);
1017                 }
1018         }
1019         nrg->from = f;
1020         nrg->to = t;
1021         return 0;
1022 }
1023
1024 static long region_chg(struct list_head *head, long f, long t)
1025 {
1026         struct file_region *rg, *nrg;
1027         long chg = 0;
1028
1029         /* Locate the region we are before or in. */
1030         list_for_each_entry(rg, head, link)
1031                 if (f <= rg->to)
1032                         break;
1033
1034         /* If we are below the current region then a new region is required.
1035          * Subtle, allocate a new region at the position but make it zero
1036          * size such that we can guarantee to record the reservation. */
1037         if (&rg->link == head || t < rg->from) {
1038                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
1039                 if (!nrg)
1040                         return -ENOMEM;
1041                 nrg->from = f;
1042                 nrg->to   = f;
1043                 INIT_LIST_HEAD(&nrg->link);
1044                 list_add(&nrg->link, rg->link.prev);
1045
1046                 return t - f;
1047         }
1048
1049         /* Round our left edge to the current segment if it encloses us. */
1050         if (f > rg->from)
1051                 f = rg->from;
1052         chg = t - f;
1053
1054         /* Check for and consume any regions we now overlap with. */
1055         list_for_each_entry(rg, rg->link.prev, link) {
1056                 if (&rg->link == head)
1057                         break;
1058                 if (rg->from > t)
1059                         return chg;
1060
1061                 /* We overlap with this area, if it extends futher than
1062                  * us then we must extend ourselves.  Account for its
1063                  * existing reservation. */
1064                 if (rg->to > t) {
1065                         chg += rg->to - t;
1066                         t = rg->to;
1067                 }
1068                 chg -= rg->to - rg->from;
1069         }
1070         return chg;
1071 }
1072
1073 static long region_truncate(struct list_head *head, long end)
1074 {
1075         struct file_region *rg, *trg;
1076         long chg = 0;
1077
1078         /* Locate the region we are either in or before. */
1079         list_for_each_entry(rg, head, link)
1080                 if (end <= rg->to)
1081                         break;
1082         if (&rg->link == head)
1083                 return 0;
1084
1085         /* If we are in the middle of a region then adjust it. */
1086         if (end > rg->from) {
1087                 chg = rg->to - end;
1088                 rg->to = end;
1089                 rg = list_entry(rg->link.next, typeof(*rg), link);
1090         }
1091
1092         /* Drop any remaining regions. */
1093         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1094                 if (&rg->link == head)
1095                         break;
1096                 chg += rg->to - rg->from;
1097                 list_del(&rg->link);
1098                 kfree(rg);
1099         }
1100         return chg;
1101 }
1102
1103 static int hugetlb_acct_memory(long delta)
1104 {
1105         int ret = -ENOMEM;
1106
1107         spin_lock(&hugetlb_lock);
1108         /*
1109          * When cpuset is configured, it breaks the strict hugetlb page
1110          * reservation as the accounting is done on a global variable. Such
1111          * reservation is completely rubbish in the presence of cpuset because
1112          * the reservation is not checked against page availability for the
1113          * current cpuset. Application can still potentially OOM'ed by kernel
1114          * with lack of free htlb page in cpuset that the task is in.
1115          * Attempt to enforce strict accounting with cpuset is almost
1116          * impossible (or too ugly) because cpuset is too fluid that
1117          * task or memory node can be dynamically moved between cpusets.
1118          *
1119          * The change of semantics for shared hugetlb mapping with cpuset is
1120          * undesirable. However, in order to preserve some of the semantics,
1121          * we fall back to check against current free page availability as
1122          * a best attempt and hopefully to minimize the impact of changing
1123          * semantics that cpuset has.
1124          */
1125         if (delta > 0) {
1126                 if (gather_surplus_pages(delta) < 0)
1127                         goto out;
1128
1129                 if (delta > cpuset_mems_nr(free_huge_pages_node))
1130                         goto out;
1131         }
1132
1133         ret = 0;
1134         resv_huge_pages += delta;
1135         if (delta < 0)
1136                 return_unused_surplus_pages((unsigned long) -delta);
1137
1138 out:
1139         spin_unlock(&hugetlb_lock);
1140         return ret;
1141 }
1142
1143 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
1144 {
1145         long ret, chg;
1146
1147         chg = region_chg(&inode->i_mapping->private_list, from, to);
1148         if (chg < 0)
1149                 return chg;
1150
1151         ret = hugetlb_acct_memory(chg);
1152         if (ret < 0)
1153                 return ret;
1154         region_add(&inode->i_mapping->private_list, from, to);
1155         return 0;
1156 }
1157
1158 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1159 {
1160         long chg = region_truncate(&inode->i_mapping->private_list, offset);
1161         hugetlb_acct_memory(freed - chg);
1162 }