Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/roland...
[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 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 unsigned long nr_overcommit_huge_pages;
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, 1);
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         unsigned int nid;
231
232         /*
233          * Assume we will successfully allocate the surplus page to
234          * prevent racing processes from causing the surplus to exceed
235          * overcommit
236          *
237          * This however introduces a different race, where a process B
238          * tries to grow the static hugepage pool while alloc_pages() is
239          * called by process A. B will only examine the per-node
240          * counters in determining if surplus huge pages can be
241          * converted to normal huge pages in adjust_pool_surplus(). A
242          * won't be able to increment the per-node counter, until the
243          * lock is dropped by B, but B doesn't drop hugetlb_lock until
244          * no more huge pages can be converted from surplus to normal
245          * state (and doesn't try to convert again). Thus, we have a
246          * case where a surplus huge page exists, the pool is grown, and
247          * the surplus huge page still exists after, even though it
248          * should just have been converted to a normal huge page. This
249          * does not leak memory, though, as the hugepage will be freed
250          * once it is out of use. It also does not allow the counters to
251          * go out of whack in adjust_pool_surplus() as we don't modify
252          * the node values until we've gotten the hugepage and only the
253          * per-node value is checked there.
254          */
255         spin_lock(&hugetlb_lock);
256         if (surplus_huge_pages >= nr_overcommit_huge_pages) {
257                 spin_unlock(&hugetlb_lock);
258                 return NULL;
259         } else {
260                 nr_huge_pages++;
261                 surplus_huge_pages++;
262         }
263         spin_unlock(&hugetlb_lock);
264
265         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
266                                         HUGETLB_PAGE_ORDER);
267
268         spin_lock(&hugetlb_lock);
269         if (page) {
270                 nid = page_to_nid(page);
271                 set_compound_page_dtor(page, free_huge_page);
272                 /*
273                  * We incremented the global counters already
274                  */
275                 nr_huge_pages_node[nid]++;
276                 surplus_huge_pages_node[nid]++;
277         } else {
278                 nr_huge_pages--;
279                 surplus_huge_pages--;
280         }
281         spin_unlock(&hugetlb_lock);
282
283         return page;
284 }
285
286 /*
287  * Increase the hugetlb pool such that it can accomodate a reservation
288  * of size 'delta'.
289  */
290 static int gather_surplus_pages(int delta)
291 {
292         struct list_head surplus_list;
293         struct page *page, *tmp;
294         int ret, i;
295         int needed, allocated;
296
297         needed = (resv_huge_pages + delta) - free_huge_pages;
298         if (needed <= 0)
299                 return 0;
300
301         allocated = 0;
302         INIT_LIST_HEAD(&surplus_list);
303
304         ret = -ENOMEM;
305 retry:
306         spin_unlock(&hugetlb_lock);
307         for (i = 0; i < needed; i++) {
308                 page = alloc_buddy_huge_page(NULL, 0);
309                 if (!page) {
310                         /*
311                          * We were not able to allocate enough pages to
312                          * satisfy the entire reservation so we free what
313                          * we've allocated so far.
314                          */
315                         spin_lock(&hugetlb_lock);
316                         needed = 0;
317                         goto free;
318                 }
319
320                 list_add(&page->lru, &surplus_list);
321         }
322         allocated += needed;
323
324         /*
325          * After retaking hugetlb_lock, we need to recalculate 'needed'
326          * because either resv_huge_pages or free_huge_pages may have changed.
327          */
328         spin_lock(&hugetlb_lock);
329         needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
330         if (needed > 0)
331                 goto retry;
332
333         /*
334          * The surplus_list now contains _at_least_ the number of extra pages
335          * needed to accomodate the reservation.  Add the appropriate number
336          * of pages to the hugetlb pool and free the extras back to the buddy
337          * allocator.
338          */
339         needed += allocated;
340         ret = 0;
341 free:
342         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
343                 list_del(&page->lru);
344                 if ((--needed) >= 0)
345                         enqueue_huge_page(page);
346                 else {
347                         /*
348                          * Decrement the refcount and free the page using its
349                          * destructor.  This must be done with hugetlb_lock
350                          * unlocked which is safe because free_huge_page takes
351                          * hugetlb_lock before deciding how to free the page.
352                          */
353                         spin_unlock(&hugetlb_lock);
354                         put_page(page);
355                         spin_lock(&hugetlb_lock);
356                 }
357         }
358
359         return ret;
360 }
361
362 /*
363  * When releasing a hugetlb pool reservation, any surplus pages that were
364  * allocated to satisfy the reservation must be explicitly freed if they were
365  * never used.
366  */
367 static void return_unused_surplus_pages(unsigned long unused_resv_pages)
368 {
369         static int nid = -1;
370         struct page *page;
371         unsigned long nr_pages;
372
373         nr_pages = min(unused_resv_pages, surplus_huge_pages);
374
375         while (nr_pages) {
376                 nid = next_node(nid, node_online_map);
377                 if (nid == MAX_NUMNODES)
378                         nid = first_node(node_online_map);
379
380                 if (!surplus_huge_pages_node[nid])
381                         continue;
382
383                 if (!list_empty(&hugepage_freelists[nid])) {
384                         page = list_entry(hugepage_freelists[nid].next,
385                                           struct page, lru);
386                         list_del(&page->lru);
387                         update_and_free_page(page);
388                         free_huge_pages--;
389                         free_huge_pages_node[nid]--;
390                         surplus_huge_pages--;
391                         surplus_huge_pages_node[nid]--;
392                         nr_pages--;
393                 }
394         }
395 }
396
397
398 static struct page *alloc_huge_page_shared(struct vm_area_struct *vma,
399                                                 unsigned long addr)
400 {
401         struct page *page;
402
403         spin_lock(&hugetlb_lock);
404         page = dequeue_huge_page(vma, addr);
405         spin_unlock(&hugetlb_lock);
406         return page ? page : ERR_PTR(-VM_FAULT_OOM);
407 }
408
409 static struct page *alloc_huge_page_private(struct vm_area_struct *vma,
410                                                 unsigned long addr)
411 {
412         struct page *page = NULL;
413
414         if (hugetlb_get_quota(vma->vm_file->f_mapping, 1))
415                 return ERR_PTR(-VM_FAULT_SIGBUS);
416
417         spin_lock(&hugetlb_lock);
418         if (free_huge_pages > resv_huge_pages)
419                 page = dequeue_huge_page(vma, addr);
420         spin_unlock(&hugetlb_lock);
421         if (!page) {
422                 page = alloc_buddy_huge_page(vma, addr);
423                 if (!page) {
424                         hugetlb_put_quota(vma->vm_file->f_mapping, 1);
425                         return ERR_PTR(-VM_FAULT_OOM);
426                 }
427         }
428         return page;
429 }
430
431 static struct page *alloc_huge_page(struct vm_area_struct *vma,
432                                     unsigned long addr)
433 {
434         struct page *page;
435         struct address_space *mapping = vma->vm_file->f_mapping;
436
437         if (vma->vm_flags & VM_MAYSHARE)
438                 page = alloc_huge_page_shared(vma, addr);
439         else
440                 page = alloc_huge_page_private(vma, addr);
441
442         if (!IS_ERR(page)) {
443                 set_page_refcounted(page);
444                 set_page_private(page, (unsigned long) mapping);
445         }
446         return page;
447 }
448
449 static int __init hugetlb_init(void)
450 {
451         unsigned long i;
452
453         if (HPAGE_SHIFT == 0)
454                 return 0;
455
456         for (i = 0; i < MAX_NUMNODES; ++i)
457                 INIT_LIST_HEAD(&hugepage_freelists[i]);
458
459         hugetlb_next_nid = first_node(node_online_map);
460
461         for (i = 0; i < max_huge_pages; ++i) {
462                 if (!alloc_fresh_huge_page())
463                         break;
464         }
465         max_huge_pages = free_huge_pages = nr_huge_pages = i;
466         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
467         return 0;
468 }
469 module_init(hugetlb_init);
470
471 static int __init hugetlb_setup(char *s)
472 {
473         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
474                 max_huge_pages = 0;
475         return 1;
476 }
477 __setup("hugepages=", hugetlb_setup);
478
479 static unsigned int cpuset_mems_nr(unsigned int *array)
480 {
481         int node;
482         unsigned int nr = 0;
483
484         for_each_node_mask(node, cpuset_current_mems_allowed)
485                 nr += array[node];
486
487         return nr;
488 }
489
490 #ifdef CONFIG_SYSCTL
491 #ifdef CONFIG_HIGHMEM
492 static void try_to_free_low(unsigned long count)
493 {
494         int i;
495
496         for (i = 0; i < MAX_NUMNODES; ++i) {
497                 struct page *page, *next;
498                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
499                         if (count >= nr_huge_pages)
500                                 return;
501                         if (PageHighMem(page))
502                                 continue;
503                         list_del(&page->lru);
504                         update_and_free_page(page);
505                         free_huge_pages--;
506                         free_huge_pages_node[page_to_nid(page)]--;
507                 }
508         }
509 }
510 #else
511 static inline void try_to_free_low(unsigned long count)
512 {
513 }
514 #endif
515
516 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
517 static unsigned long set_max_huge_pages(unsigned long count)
518 {
519         unsigned long min_count, ret;
520
521         /*
522          * Increase the pool size
523          * First take pages out of surplus state.  Then make up the
524          * remaining difference by allocating fresh huge pages.
525          *
526          * We might race with alloc_buddy_huge_page() here and be unable
527          * to convert a surplus huge page to a normal huge page. That is
528          * not critical, though, it just means the overall size of the
529          * pool might be one hugepage larger than it needs to be, but
530          * within all the constraints specified by the sysctls.
531          */
532         spin_lock(&hugetlb_lock);
533         while (surplus_huge_pages && count > persistent_huge_pages) {
534                 if (!adjust_pool_surplus(-1))
535                         break;
536         }
537
538         while (count > persistent_huge_pages) {
539                 int ret;
540                 /*
541                  * If this allocation races such that we no longer need the
542                  * page, free_huge_page will handle it by freeing the page
543                  * and reducing the surplus.
544                  */
545                 spin_unlock(&hugetlb_lock);
546                 ret = alloc_fresh_huge_page();
547                 spin_lock(&hugetlb_lock);
548                 if (!ret)
549                         goto out;
550
551         }
552
553         /*
554          * Decrease the pool size
555          * First return free pages to the buddy allocator (being careful
556          * to keep enough around to satisfy reservations).  Then place
557          * pages into surplus state as needed so the pool will shrink
558          * to the desired size as pages become free.
559          *
560          * By placing pages into the surplus state independent of the
561          * overcommit value, we are allowing the surplus pool size to
562          * exceed overcommit. There are few sane options here. Since
563          * alloc_buddy_huge_page() is checking the global counter,
564          * though, we'll note that we're not allowed to exceed surplus
565          * and won't grow the pool anywhere else. Not until one of the
566          * sysctls are changed, or the surplus pages go out of use.
567          */
568         min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
569         min_count = max(count, min_count);
570         try_to_free_low(min_count);
571         while (min_count < persistent_huge_pages) {
572                 struct page *page = dequeue_huge_page(NULL, 0);
573                 if (!page)
574                         break;
575                 update_and_free_page(page);
576         }
577         while (count < persistent_huge_pages) {
578                 if (!adjust_pool_surplus(1))
579                         break;
580         }
581 out:
582         ret = persistent_huge_pages;
583         spin_unlock(&hugetlb_lock);
584         return ret;
585 }
586
587 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
588                            struct file *file, void __user *buffer,
589                            size_t *length, loff_t *ppos)
590 {
591         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
592         max_huge_pages = set_max_huge_pages(max_huge_pages);
593         return 0;
594 }
595
596 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
597                         struct file *file, void __user *buffer,
598                         size_t *length, loff_t *ppos)
599 {
600         proc_dointvec(table, write, file, buffer, length, ppos);
601         if (hugepages_treat_as_movable)
602                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
603         else
604                 htlb_alloc_mask = GFP_HIGHUSER;
605         return 0;
606 }
607
608 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
609                         struct file *file, void __user *buffer,
610                         size_t *length, loff_t *ppos)
611 {
612         spin_lock(&hugetlb_lock);
613         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
614         spin_unlock(&hugetlb_lock);
615         return 0;
616 }
617
618 #endif /* CONFIG_SYSCTL */
619
620 int hugetlb_report_meminfo(char *buf)
621 {
622         return sprintf(buf,
623                         "HugePages_Total: %5lu\n"
624                         "HugePages_Free:  %5lu\n"
625                         "HugePages_Rsvd:  %5lu\n"
626                         "HugePages_Surp:  %5lu\n"
627                         "Hugepagesize:    %5lu kB\n",
628                         nr_huge_pages,
629                         free_huge_pages,
630                         resv_huge_pages,
631                         surplus_huge_pages,
632                         HPAGE_SIZE/1024);
633 }
634
635 int hugetlb_report_node_meminfo(int nid, char *buf)
636 {
637         return sprintf(buf,
638                 "Node %d HugePages_Total: %5u\n"
639                 "Node %d HugePages_Free:  %5u\n",
640                 nid, nr_huge_pages_node[nid],
641                 nid, free_huge_pages_node[nid]);
642 }
643
644 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
645 unsigned long hugetlb_total_pages(void)
646 {
647         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
648 }
649
650 /*
651  * We cannot handle pagefaults against hugetlb pages at all.  They cause
652  * handle_mm_fault() to try to instantiate regular-sized pages in the
653  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
654  * this far.
655  */
656 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
657 {
658         BUG();
659         return 0;
660 }
661
662 struct vm_operations_struct hugetlb_vm_ops = {
663         .fault = hugetlb_vm_op_fault,
664 };
665
666 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
667                                 int writable)
668 {
669         pte_t entry;
670
671         if (writable) {
672                 entry =
673                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
674         } else {
675                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
676         }
677         entry = pte_mkyoung(entry);
678         entry = pte_mkhuge(entry);
679
680         return entry;
681 }
682
683 static void set_huge_ptep_writable(struct vm_area_struct *vma,
684                                    unsigned long address, pte_t *ptep)
685 {
686         pte_t entry;
687
688         entry = pte_mkwrite(pte_mkdirty(*ptep));
689         if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
690                 update_mmu_cache(vma, address, entry);
691         }
692 }
693
694
695 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
696                             struct vm_area_struct *vma)
697 {
698         pte_t *src_pte, *dst_pte, entry;
699         struct page *ptepage;
700         unsigned long addr;
701         int cow;
702
703         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
704
705         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
706                 src_pte = huge_pte_offset(src, addr);
707                 if (!src_pte)
708                         continue;
709                 dst_pte = huge_pte_alloc(dst, addr);
710                 if (!dst_pte)
711                         goto nomem;
712
713                 /* If the pagetables are shared don't copy or take references */
714                 if (dst_pte == src_pte)
715                         continue;
716
717                 spin_lock(&dst->page_table_lock);
718                 spin_lock(&src->page_table_lock);
719                 if (!pte_none(*src_pte)) {
720                         if (cow)
721                                 ptep_set_wrprotect(src, addr, src_pte);
722                         entry = *src_pte;
723                         ptepage = pte_page(entry);
724                         get_page(ptepage);
725                         set_huge_pte_at(dst, addr, dst_pte, entry);
726                 }
727                 spin_unlock(&src->page_table_lock);
728                 spin_unlock(&dst->page_table_lock);
729         }
730         return 0;
731
732 nomem:
733         return -ENOMEM;
734 }
735
736 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
737                             unsigned long end)
738 {
739         struct mm_struct *mm = vma->vm_mm;
740         unsigned long address;
741         pte_t *ptep;
742         pte_t pte;
743         struct page *page;
744         struct page *tmp;
745         /*
746          * A page gathering list, protected by per file i_mmap_lock. The
747          * lock is used to avoid list corruption from multiple unmapping
748          * of the same page since we are using page->lru.
749          */
750         LIST_HEAD(page_list);
751
752         WARN_ON(!is_vm_hugetlb_page(vma));
753         BUG_ON(start & ~HPAGE_MASK);
754         BUG_ON(end & ~HPAGE_MASK);
755
756         spin_lock(&mm->page_table_lock);
757         for (address = start; address < end; address += HPAGE_SIZE) {
758                 ptep = huge_pte_offset(mm, address);
759                 if (!ptep)
760                         continue;
761
762                 if (huge_pmd_unshare(mm, &address, ptep))
763                         continue;
764
765                 pte = huge_ptep_get_and_clear(mm, address, ptep);
766                 if (pte_none(pte))
767                         continue;
768
769                 page = pte_page(pte);
770                 if (pte_dirty(pte))
771                         set_page_dirty(page);
772                 list_add(&page->lru, &page_list);
773         }
774         spin_unlock(&mm->page_table_lock);
775         flush_tlb_range(vma, start, end);
776         list_for_each_entry_safe(page, tmp, &page_list, lru) {
777                 list_del(&page->lru);
778                 put_page(page);
779         }
780 }
781
782 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
783                           unsigned long end)
784 {
785         /*
786          * It is undesirable to test vma->vm_file as it should be non-null
787          * for valid hugetlb area. However, vm_file will be NULL in the error
788          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
789          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
790          * to clean up. Since no pte has actually been setup, it is safe to
791          * do nothing in this case.
792          */
793         if (vma->vm_file) {
794                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
795                 __unmap_hugepage_range(vma, start, end);
796                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
797         }
798 }
799
800 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
801                         unsigned long address, pte_t *ptep, pte_t pte)
802 {
803         struct page *old_page, *new_page;
804         int avoidcopy;
805
806         old_page = pte_page(pte);
807
808         /* If no-one else is actually using this page, avoid the copy
809          * and just make the page writable */
810         avoidcopy = (page_count(old_page) == 1);
811         if (avoidcopy) {
812                 set_huge_ptep_writable(vma, address, ptep);
813                 return 0;
814         }
815
816         page_cache_get(old_page);
817         new_page = alloc_huge_page(vma, address);
818
819         if (IS_ERR(new_page)) {
820                 page_cache_release(old_page);
821                 return -PTR_ERR(new_page);
822         }
823
824         spin_unlock(&mm->page_table_lock);
825         copy_huge_page(new_page, old_page, address, vma);
826         __SetPageUptodate(new_page);
827         spin_lock(&mm->page_table_lock);
828
829         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
830         if (likely(pte_same(*ptep, pte))) {
831                 /* Break COW */
832                 set_huge_pte_at(mm, address, ptep,
833                                 make_huge_pte(vma, new_page, 1));
834                 /* Make the old page be freed below */
835                 new_page = old_page;
836         }
837         page_cache_release(new_page);
838         page_cache_release(old_page);
839         return 0;
840 }
841
842 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
843                         unsigned long address, pte_t *ptep, int write_access)
844 {
845         int ret = VM_FAULT_SIGBUS;
846         unsigned long idx;
847         unsigned long size;
848         struct page *page;
849         struct address_space *mapping;
850         pte_t new_pte;
851
852         mapping = vma->vm_file->f_mapping;
853         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
854                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
855
856         /*
857          * Use page lock to guard against racing truncation
858          * before we get page_table_lock.
859          */
860 retry:
861         page = find_lock_page(mapping, idx);
862         if (!page) {
863                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
864                 if (idx >= size)
865                         goto out;
866                 page = alloc_huge_page(vma, address);
867                 if (IS_ERR(page)) {
868                         ret = -PTR_ERR(page);
869                         goto out;
870                 }
871                 clear_huge_page(page, address);
872                 __SetPageUptodate(page);
873
874                 if (vma->vm_flags & VM_SHARED) {
875                         int err;
876                         struct inode *inode = mapping->host;
877
878                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
879                         if (err) {
880                                 put_page(page);
881                                 if (err == -EEXIST)
882                                         goto retry;
883                                 goto out;
884                         }
885
886                         spin_lock(&inode->i_lock);
887                         inode->i_blocks += BLOCKS_PER_HUGEPAGE;
888                         spin_unlock(&inode->i_lock);
889                 } else
890                         lock_page(page);
891         }
892
893         spin_lock(&mm->page_table_lock);
894         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
895         if (idx >= size)
896                 goto backout;
897
898         ret = 0;
899         if (!pte_none(*ptep))
900                 goto backout;
901
902         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
903                                 && (vma->vm_flags & VM_SHARED)));
904         set_huge_pte_at(mm, address, ptep, new_pte);
905
906         if (write_access && !(vma->vm_flags & VM_SHARED)) {
907                 /* Optimization, do the COW without a second fault */
908                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
909         }
910
911         spin_unlock(&mm->page_table_lock);
912         unlock_page(page);
913 out:
914         return ret;
915
916 backout:
917         spin_unlock(&mm->page_table_lock);
918         unlock_page(page);
919         put_page(page);
920         goto out;
921 }
922
923 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
924                         unsigned long address, int write_access)
925 {
926         pte_t *ptep;
927         pte_t entry;
928         int ret;
929         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
930
931         ptep = huge_pte_alloc(mm, address);
932         if (!ptep)
933                 return VM_FAULT_OOM;
934
935         /*
936          * Serialize hugepage allocation and instantiation, so that we don't
937          * get spurious allocation failures if two CPUs race to instantiate
938          * the same page in the page cache.
939          */
940         mutex_lock(&hugetlb_instantiation_mutex);
941         entry = *ptep;
942         if (pte_none(entry)) {
943                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
944                 mutex_unlock(&hugetlb_instantiation_mutex);
945                 return ret;
946         }
947
948         ret = 0;
949
950         spin_lock(&mm->page_table_lock);
951         /* Check for a racing update before calling hugetlb_cow */
952         if (likely(pte_same(entry, *ptep)))
953                 if (write_access && !pte_write(entry))
954                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
955         spin_unlock(&mm->page_table_lock);
956         mutex_unlock(&hugetlb_instantiation_mutex);
957
958         return ret;
959 }
960
961 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
962                         struct page **pages, struct vm_area_struct **vmas,
963                         unsigned long *position, int *length, int i,
964                         int write)
965 {
966         unsigned long pfn_offset;
967         unsigned long vaddr = *position;
968         int remainder = *length;
969
970         spin_lock(&mm->page_table_lock);
971         while (vaddr < vma->vm_end && remainder) {
972                 pte_t *pte;
973                 struct page *page;
974
975                 /*
976                  * Some archs (sparc64, sh*) have multiple pte_ts to
977                  * each hugepage.  We have to make * sure we get the
978                  * first, for the page indexing below to work.
979                  */
980                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
981
982                 if (!pte || pte_none(*pte) || (write && !pte_write(*pte))) {
983                         int ret;
984
985                         spin_unlock(&mm->page_table_lock);
986                         ret = hugetlb_fault(mm, vma, vaddr, write);
987                         spin_lock(&mm->page_table_lock);
988                         if (!(ret & VM_FAULT_ERROR))
989                                 continue;
990
991                         remainder = 0;
992                         if (!i)
993                                 i = -EFAULT;
994                         break;
995                 }
996
997                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
998                 page = pte_page(*pte);
999 same_page:
1000                 if (pages) {
1001                         get_page(page);
1002                         pages[i] = page + pfn_offset;
1003                 }
1004
1005                 if (vmas)
1006                         vmas[i] = vma;
1007
1008                 vaddr += PAGE_SIZE;
1009                 ++pfn_offset;
1010                 --remainder;
1011                 ++i;
1012                 if (vaddr < vma->vm_end && remainder &&
1013                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
1014                         /*
1015                          * We use pfn_offset to avoid touching the pageframes
1016                          * of this compound page.
1017                          */
1018                         goto same_page;
1019                 }
1020         }
1021         spin_unlock(&mm->page_table_lock);
1022         *length = remainder;
1023         *position = vaddr;
1024
1025         return i;
1026 }
1027
1028 void hugetlb_change_protection(struct vm_area_struct *vma,
1029                 unsigned long address, unsigned long end, pgprot_t newprot)
1030 {
1031         struct mm_struct *mm = vma->vm_mm;
1032         unsigned long start = address;
1033         pte_t *ptep;
1034         pte_t pte;
1035
1036         BUG_ON(address >= end);
1037         flush_cache_range(vma, address, end);
1038
1039         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1040         spin_lock(&mm->page_table_lock);
1041         for (; address < end; address += HPAGE_SIZE) {
1042                 ptep = huge_pte_offset(mm, address);
1043                 if (!ptep)
1044                         continue;
1045                 if (huge_pmd_unshare(mm, &address, ptep))
1046                         continue;
1047                 if (!pte_none(*ptep)) {
1048                         pte = huge_ptep_get_and_clear(mm, address, ptep);
1049                         pte = pte_mkhuge(pte_modify(pte, newprot));
1050                         set_huge_pte_at(mm, address, ptep, pte);
1051                 }
1052         }
1053         spin_unlock(&mm->page_table_lock);
1054         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1055
1056         flush_tlb_range(vma, start, end);
1057 }
1058
1059 struct file_region {
1060         struct list_head link;
1061         long from;
1062         long to;
1063 };
1064
1065 static long region_add(struct list_head *head, long f, long t)
1066 {
1067         struct file_region *rg, *nrg, *trg;
1068
1069         /* Locate the region we are either in or before. */
1070         list_for_each_entry(rg, head, link)
1071                 if (f <= rg->to)
1072                         break;
1073
1074         /* Round our left edge to the current segment if it encloses us. */
1075         if (f > rg->from)
1076                 f = rg->from;
1077
1078         /* Check for and consume any regions we now overlap with. */
1079         nrg = rg;
1080         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1081                 if (&rg->link == head)
1082                         break;
1083                 if (rg->from > t)
1084                         break;
1085
1086                 /* If this area reaches higher then extend our area to
1087                  * include it completely.  If this is not the first area
1088                  * which we intend to reuse, free it. */
1089                 if (rg->to > t)
1090                         t = rg->to;
1091                 if (rg != nrg) {
1092                         list_del(&rg->link);
1093                         kfree(rg);
1094                 }
1095         }
1096         nrg->from = f;
1097         nrg->to = t;
1098         return 0;
1099 }
1100
1101 static long region_chg(struct list_head *head, long f, long t)
1102 {
1103         struct file_region *rg, *nrg;
1104         long chg = 0;
1105
1106         /* Locate the region we are before or in. */
1107         list_for_each_entry(rg, head, link)
1108                 if (f <= rg->to)
1109                         break;
1110
1111         /* If we are below the current region then a new region is required.
1112          * Subtle, allocate a new region at the position but make it zero
1113          * size such that we can guarantee to record the reservation. */
1114         if (&rg->link == head || t < rg->from) {
1115                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
1116                 if (!nrg)
1117                         return -ENOMEM;
1118                 nrg->from = f;
1119                 nrg->to   = f;
1120                 INIT_LIST_HEAD(&nrg->link);
1121                 list_add(&nrg->link, rg->link.prev);
1122
1123                 return t - f;
1124         }
1125
1126         /* Round our left edge to the current segment if it encloses us. */
1127         if (f > rg->from)
1128                 f = rg->from;
1129         chg = t - f;
1130
1131         /* Check for and consume any regions we now overlap with. */
1132         list_for_each_entry(rg, rg->link.prev, link) {
1133                 if (&rg->link == head)
1134                         break;
1135                 if (rg->from > t)
1136                         return chg;
1137
1138                 /* We overlap with this area, if it extends futher than
1139                  * us then we must extend ourselves.  Account for its
1140                  * existing reservation. */
1141                 if (rg->to > t) {
1142                         chg += rg->to - t;
1143                         t = rg->to;
1144                 }
1145                 chg -= rg->to - rg->from;
1146         }
1147         return chg;
1148 }
1149
1150 static long region_truncate(struct list_head *head, long end)
1151 {
1152         struct file_region *rg, *trg;
1153         long chg = 0;
1154
1155         /* Locate the region we are either in or before. */
1156         list_for_each_entry(rg, head, link)
1157                 if (end <= rg->to)
1158                         break;
1159         if (&rg->link == head)
1160                 return 0;
1161
1162         /* If we are in the middle of a region then adjust it. */
1163         if (end > rg->from) {
1164                 chg = rg->to - end;
1165                 rg->to = end;
1166                 rg = list_entry(rg->link.next, typeof(*rg), link);
1167         }
1168
1169         /* Drop any remaining regions. */
1170         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1171                 if (&rg->link == head)
1172                         break;
1173                 chg += rg->to - rg->from;
1174                 list_del(&rg->link);
1175                 kfree(rg);
1176         }
1177         return chg;
1178 }
1179
1180 static int hugetlb_acct_memory(long delta)
1181 {
1182         int ret = -ENOMEM;
1183
1184         spin_lock(&hugetlb_lock);
1185         /*
1186          * When cpuset is configured, it breaks the strict hugetlb page
1187          * reservation as the accounting is done on a global variable. Such
1188          * reservation is completely rubbish in the presence of cpuset because
1189          * the reservation is not checked against page availability for the
1190          * current cpuset. Application can still potentially OOM'ed by kernel
1191          * with lack of free htlb page in cpuset that the task is in.
1192          * Attempt to enforce strict accounting with cpuset is almost
1193          * impossible (or too ugly) because cpuset is too fluid that
1194          * task or memory node can be dynamically moved between cpusets.
1195          *
1196          * The change of semantics for shared hugetlb mapping with cpuset is
1197          * undesirable. However, in order to preserve some of the semantics,
1198          * we fall back to check against current free page availability as
1199          * a best attempt and hopefully to minimize the impact of changing
1200          * semantics that cpuset has.
1201          */
1202         if (delta > 0) {
1203                 if (gather_surplus_pages(delta) < 0)
1204                         goto out;
1205
1206                 if (delta > cpuset_mems_nr(free_huge_pages_node))
1207                         goto out;
1208         }
1209
1210         ret = 0;
1211         resv_huge_pages += delta;
1212         if (delta < 0)
1213                 return_unused_surplus_pages((unsigned long) -delta);
1214
1215 out:
1216         spin_unlock(&hugetlb_lock);
1217         return ret;
1218 }
1219
1220 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
1221 {
1222         long ret, chg;
1223
1224         chg = region_chg(&inode->i_mapping->private_list, from, to);
1225         if (chg < 0)
1226                 return chg;
1227
1228         if (hugetlb_get_quota(inode->i_mapping, chg))
1229                 return -ENOSPC;
1230         ret = hugetlb_acct_memory(chg);
1231         if (ret < 0) {
1232                 hugetlb_put_quota(inode->i_mapping, chg);
1233                 return ret;
1234         }
1235         region_add(&inode->i_mapping->private_list, from, to);
1236         return 0;
1237 }
1238
1239 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1240 {
1241         long chg = region_truncate(&inode->i_mapping->private_list, offset);
1242
1243         spin_lock(&inode->i_lock);
1244         inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed;
1245         spin_unlock(&inode->i_lock);
1246
1247         hugetlb_put_quota(inode->i_mapping, (chg - freed));
1248         hugetlb_acct_memory(-(chg - freed));
1249 }