dcacc811e70ede9cda49d4d2cf56a6d5d5404a56
[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 static unsigned long nr_overcommit_huge_pages;
28 unsigned long max_huge_pages;
29 unsigned long sysctl_overcommit_huge_pages;
30 static struct list_head hugepage_freelists[MAX_NUMNODES];
31 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
32 static unsigned int free_huge_pages_node[MAX_NUMNODES];
33 static unsigned int surplus_huge_pages_node[MAX_NUMNODES];
34 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
35 unsigned long hugepages_treat_as_movable;
36 static int hugetlb_next_nid;
37
38 /*
39  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
40  */
41 static DEFINE_SPINLOCK(hugetlb_lock);
42
43 static void clear_huge_page(struct page *page, unsigned long addr)
44 {
45         int i;
46
47         might_sleep();
48         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
49                 cond_resched();
50                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
51         }
52 }
53
54 static void copy_huge_page(struct page *dst, struct page *src,
55                            unsigned long addr, struct vm_area_struct *vma)
56 {
57         int i;
58
59         might_sleep();
60         for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
61                 cond_resched();
62                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
63         }
64 }
65
66 static void enqueue_huge_page(struct page *page)
67 {
68         int nid = page_to_nid(page);
69         list_add(&page->lru, &hugepage_freelists[nid]);
70         free_huge_pages++;
71         free_huge_pages_node[nid]++;
72 }
73
74 static struct page *dequeue_huge_page(void)
75 {
76         int nid;
77         struct page *page = NULL;
78
79         for (nid = 0; nid < MAX_NUMNODES; ++nid) {
80                 if (!list_empty(&hugepage_freelists[nid])) {
81                         page = list_entry(hugepage_freelists[nid].next,
82                                           struct page, lru);
83                         list_del(&page->lru);
84                         free_huge_pages--;
85                         free_huge_pages_node[nid]--;
86                         break;
87                 }
88         }
89         return page;
90 }
91
92 static struct page *dequeue_huge_page_vma(struct vm_area_struct *vma,
93                                 unsigned long address)
94 {
95         int nid;
96         struct page *page = NULL;
97         struct mempolicy *mpol;
98         struct zonelist *zonelist = huge_zonelist(vma, address,
99                                         htlb_alloc_mask, &mpol);
100         struct zone **z;
101
102         for (z = zonelist->zones; *z; z++) {
103                 nid = zone_to_nid(*z);
104                 if (cpuset_zone_allowed_softwall(*z, htlb_alloc_mask) &&
105                     !list_empty(&hugepage_freelists[nid])) {
106                         page = list_entry(hugepage_freelists[nid].next,
107                                           struct page, lru);
108                         list_del(&page->lru);
109                         free_huge_pages--;
110                         free_huge_pages_node[nid]--;
111                         if (vma && vma->vm_flags & VM_MAYSHARE)
112                                 resv_huge_pages--;
113                         break;
114                 }
115         }
116         mpol_free(mpol);        /* unref if mpol !NULL */
117         return page;
118 }
119
120 static void update_and_free_page(struct page *page)
121 {
122         int i;
123         nr_huge_pages--;
124         nr_huge_pages_node[page_to_nid(page)]--;
125         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
126                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
127                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
128                                 1 << PG_private | 1<< PG_writeback);
129         }
130         set_compound_page_dtor(page, NULL);
131         set_page_refcounted(page);
132         __free_pages(page, HUGETLB_PAGE_ORDER);
133 }
134
135 static void free_huge_page(struct page *page)
136 {
137         int nid = page_to_nid(page);
138         struct address_space *mapping;
139
140         mapping = (struct address_space *) page_private(page);
141         set_page_private(page, 0);
142         BUG_ON(page_count(page));
143         INIT_LIST_HEAD(&page->lru);
144
145         spin_lock(&hugetlb_lock);
146         if (surplus_huge_pages_node[nid]) {
147                 update_and_free_page(page);
148                 surplus_huge_pages--;
149                 surplus_huge_pages_node[nid]--;
150         } else {
151                 enqueue_huge_page(page);
152         }
153         spin_unlock(&hugetlb_lock);
154         if (mapping)
155                 hugetlb_put_quota(mapping, 1);
156 }
157
158 /*
159  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
160  * balanced by operating on them in a round-robin fashion.
161  * Returns 1 if an adjustment was made.
162  */
163 static int adjust_pool_surplus(int delta)
164 {
165         static int prev_nid;
166         int nid = prev_nid;
167         int ret = 0;
168
169         VM_BUG_ON(delta != -1 && delta != 1);
170         do {
171                 nid = next_node(nid, node_online_map);
172                 if (nid == MAX_NUMNODES)
173                         nid = first_node(node_online_map);
174
175                 /* To shrink on this node, there must be a surplus page */
176                 if (delta < 0 && !surplus_huge_pages_node[nid])
177                         continue;
178                 /* Surplus cannot exceed the total number of pages */
179                 if (delta > 0 && surplus_huge_pages_node[nid] >=
180                                                 nr_huge_pages_node[nid])
181                         continue;
182
183                 surplus_huge_pages += delta;
184                 surplus_huge_pages_node[nid] += delta;
185                 ret = 1;
186                 break;
187         } while (nid != prev_nid);
188
189         prev_nid = nid;
190         return ret;
191 }
192
193 static struct page *alloc_fresh_huge_page_node(int nid)
194 {
195         struct page *page;
196
197         page = alloc_pages_node(nid,
198                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|__GFP_NOWARN,
199                 HUGETLB_PAGE_ORDER);
200         if (page) {
201                 set_compound_page_dtor(page, free_huge_page);
202                 spin_lock(&hugetlb_lock);
203                 nr_huge_pages++;
204                 nr_huge_pages_node[nid]++;
205                 spin_unlock(&hugetlb_lock);
206                 put_page(page); /* free it into the hugepage allocator */
207         }
208
209         return page;
210 }
211
212 static int alloc_fresh_huge_page(void)
213 {
214         struct page *page;
215         int start_nid;
216         int next_nid;
217         int ret = 0;
218
219         start_nid = hugetlb_next_nid;
220
221         do {
222                 page = alloc_fresh_huge_page_node(hugetlb_next_nid);
223                 if (page)
224                         ret = 1;
225                 /*
226                  * Use a helper variable to find the next node and then
227                  * copy it back to hugetlb_next_nid afterwards:
228                  * otherwise there's a window in which a racer might
229                  * pass invalid nid MAX_NUMNODES to alloc_pages_node.
230                  * But we don't need to use a spin_lock here: it really
231                  * doesn't matter if occasionally a racer chooses the
232                  * same nid as we do.  Move nid forward in the mask even
233                  * if we just successfully allocated a hugepage so that
234                  * the next caller gets hugepages on the next node.
235                  */
236                 next_nid = next_node(hugetlb_next_nid, node_online_map);
237                 if (next_nid == MAX_NUMNODES)
238                         next_nid = first_node(node_online_map);
239                 hugetlb_next_nid = next_nid;
240         } while (!page && hugetlb_next_nid != start_nid);
241
242         return ret;
243 }
244
245 static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
246                                                 unsigned long address)
247 {
248         struct page *page;
249         unsigned int nid;
250
251         /*
252          * Assume we will successfully allocate the surplus page to
253          * prevent racing processes from causing the surplus to exceed
254          * overcommit
255          *
256          * This however introduces a different race, where a process B
257          * tries to grow the static hugepage pool while alloc_pages() is
258          * called by process A. B will only examine the per-node
259          * counters in determining if surplus huge pages can be
260          * converted to normal huge pages in adjust_pool_surplus(). A
261          * won't be able to increment the per-node counter, until the
262          * lock is dropped by B, but B doesn't drop hugetlb_lock until
263          * no more huge pages can be converted from surplus to normal
264          * state (and doesn't try to convert again). Thus, we have a
265          * case where a surplus huge page exists, the pool is grown, and
266          * the surplus huge page still exists after, even though it
267          * should just have been converted to a normal huge page. This
268          * does not leak memory, though, as the hugepage will be freed
269          * once it is out of use. It also does not allow the counters to
270          * go out of whack in adjust_pool_surplus() as we don't modify
271          * the node values until we've gotten the hugepage and only the
272          * per-node value is checked there.
273          */
274         spin_lock(&hugetlb_lock);
275         if (surplus_huge_pages >= nr_overcommit_huge_pages) {
276                 spin_unlock(&hugetlb_lock);
277                 return NULL;
278         } else {
279                 nr_huge_pages++;
280                 surplus_huge_pages++;
281         }
282         spin_unlock(&hugetlb_lock);
283
284         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
285                                         HUGETLB_PAGE_ORDER);
286
287         spin_lock(&hugetlb_lock);
288         if (page) {
289                 nid = page_to_nid(page);
290                 set_compound_page_dtor(page, free_huge_page);
291                 /*
292                  * We incremented the global counters already
293                  */
294                 nr_huge_pages_node[nid]++;
295                 surplus_huge_pages_node[nid]++;
296         } else {
297                 nr_huge_pages--;
298                 surplus_huge_pages--;
299         }
300         spin_unlock(&hugetlb_lock);
301
302         return page;
303 }
304
305 /*
306  * Increase the hugetlb pool such that it can accomodate a reservation
307  * of size 'delta'.
308  */
309 static int gather_surplus_pages(int delta)
310 {
311         struct list_head surplus_list;
312         struct page *page, *tmp;
313         int ret, i;
314         int needed, allocated;
315
316         needed = (resv_huge_pages + delta) - free_huge_pages;
317         if (needed <= 0) {
318                 resv_huge_pages += delta;
319                 return 0;
320         }
321
322         allocated = 0;
323         INIT_LIST_HEAD(&surplus_list);
324
325         ret = -ENOMEM;
326 retry:
327         spin_unlock(&hugetlb_lock);
328         for (i = 0; i < needed; i++) {
329                 page = alloc_buddy_huge_page(NULL, 0);
330                 if (!page) {
331                         /*
332                          * We were not able to allocate enough pages to
333                          * satisfy the entire reservation so we free what
334                          * we've allocated so far.
335                          */
336                         spin_lock(&hugetlb_lock);
337                         needed = 0;
338                         goto free;
339                 }
340
341                 list_add(&page->lru, &surplus_list);
342         }
343         allocated += needed;
344
345         /*
346          * After retaking hugetlb_lock, we need to recalculate 'needed'
347          * because either resv_huge_pages or free_huge_pages may have changed.
348          */
349         spin_lock(&hugetlb_lock);
350         needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
351         if (needed > 0)
352                 goto retry;
353
354         /*
355          * The surplus_list now contains _at_least_ the number of extra pages
356          * needed to accomodate the reservation.  Add the appropriate number
357          * of pages to the hugetlb pool and free the extras back to the buddy
358          * allocator.  Commit the entire reservation here to prevent another
359          * process from stealing the pages as they are added to the pool but
360          * before they are reserved.
361          */
362         needed += allocated;
363         resv_huge_pages += delta;
364         ret = 0;
365 free:
366         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
367                 list_del(&page->lru);
368                 if ((--needed) >= 0)
369                         enqueue_huge_page(page);
370                 else {
371                         /*
372                          * Decrement the refcount and free the page using its
373                          * destructor.  This must be done with hugetlb_lock
374                          * unlocked which is safe because free_huge_page takes
375                          * hugetlb_lock before deciding how to free the page.
376                          */
377                         spin_unlock(&hugetlb_lock);
378                         put_page(page);
379                         spin_lock(&hugetlb_lock);
380                 }
381         }
382
383         return ret;
384 }
385
386 /*
387  * When releasing a hugetlb pool reservation, any surplus pages that were
388  * allocated to satisfy the reservation must be explicitly freed if they were
389  * never used.
390  */
391 static void return_unused_surplus_pages(unsigned long unused_resv_pages)
392 {
393         static int nid = -1;
394         struct page *page;
395         unsigned long nr_pages;
396
397         /* Uncommit the reservation */
398         resv_huge_pages -= unused_resv_pages;
399
400         nr_pages = min(unused_resv_pages, surplus_huge_pages);
401
402         while (nr_pages) {
403                 nid = next_node(nid, node_online_map);
404                 if (nid == MAX_NUMNODES)
405                         nid = first_node(node_online_map);
406
407                 if (!surplus_huge_pages_node[nid])
408                         continue;
409
410                 if (!list_empty(&hugepage_freelists[nid])) {
411                         page = list_entry(hugepage_freelists[nid].next,
412                                           struct page, lru);
413                         list_del(&page->lru);
414                         update_and_free_page(page);
415                         free_huge_pages--;
416                         free_huge_pages_node[nid]--;
417                         surplus_huge_pages--;
418                         surplus_huge_pages_node[nid]--;
419                         nr_pages--;
420                 }
421         }
422 }
423
424
425 static struct page *alloc_huge_page_shared(struct vm_area_struct *vma,
426                                                 unsigned long addr)
427 {
428         struct page *page;
429
430         spin_lock(&hugetlb_lock);
431         page = dequeue_huge_page_vma(vma, addr);
432         spin_unlock(&hugetlb_lock);
433         return page ? page : ERR_PTR(-VM_FAULT_OOM);
434 }
435
436 static struct page *alloc_huge_page_private(struct vm_area_struct *vma,
437                                                 unsigned long addr)
438 {
439         struct page *page = NULL;
440
441         if (hugetlb_get_quota(vma->vm_file->f_mapping, 1))
442                 return ERR_PTR(-VM_FAULT_SIGBUS);
443
444         spin_lock(&hugetlb_lock);
445         if (free_huge_pages > resv_huge_pages)
446                 page = dequeue_huge_page_vma(vma, addr);
447         spin_unlock(&hugetlb_lock);
448         if (!page) {
449                 page = alloc_buddy_huge_page(vma, addr);
450                 if (!page) {
451                         hugetlb_put_quota(vma->vm_file->f_mapping, 1);
452                         return ERR_PTR(-VM_FAULT_OOM);
453                 }
454         }
455         return page;
456 }
457
458 static struct page *alloc_huge_page(struct vm_area_struct *vma,
459                                     unsigned long addr)
460 {
461         struct page *page;
462         struct address_space *mapping = vma->vm_file->f_mapping;
463
464         if (vma->vm_flags & VM_MAYSHARE)
465                 page = alloc_huge_page_shared(vma, addr);
466         else
467                 page = alloc_huge_page_private(vma, addr);
468
469         if (!IS_ERR(page)) {
470                 set_page_refcounted(page);
471                 set_page_private(page, (unsigned long) mapping);
472         }
473         return page;
474 }
475
476 static int __init hugetlb_init(void)
477 {
478         unsigned long i;
479
480         if (HPAGE_SHIFT == 0)
481                 return 0;
482
483         for (i = 0; i < MAX_NUMNODES; ++i)
484                 INIT_LIST_HEAD(&hugepage_freelists[i]);
485
486         hugetlb_next_nid = first_node(node_online_map);
487
488         for (i = 0; i < max_huge_pages; ++i) {
489                 if (!alloc_fresh_huge_page())
490                         break;
491         }
492         max_huge_pages = free_huge_pages = nr_huge_pages = i;
493         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
494         return 0;
495 }
496 module_init(hugetlb_init);
497
498 static int __init hugetlb_setup(char *s)
499 {
500         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
501                 max_huge_pages = 0;
502         return 1;
503 }
504 __setup("hugepages=", hugetlb_setup);
505
506 static unsigned int cpuset_mems_nr(unsigned int *array)
507 {
508         int node;
509         unsigned int nr = 0;
510
511         for_each_node_mask(node, cpuset_current_mems_allowed)
512                 nr += array[node];
513
514         return nr;
515 }
516
517 #ifdef CONFIG_SYSCTL
518 #ifdef CONFIG_HIGHMEM
519 static void try_to_free_low(unsigned long count)
520 {
521         int i;
522
523         for (i = 0; i < MAX_NUMNODES; ++i) {
524                 struct page *page, *next;
525                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
526                         if (count >= nr_huge_pages)
527                                 return;
528                         if (PageHighMem(page))
529                                 continue;
530                         list_del(&page->lru);
531                         update_and_free_page(page);
532                         free_huge_pages--;
533                         free_huge_pages_node[page_to_nid(page)]--;
534                 }
535         }
536 }
537 #else
538 static inline void try_to_free_low(unsigned long count)
539 {
540 }
541 #endif
542
543 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
544 static unsigned long set_max_huge_pages(unsigned long count)
545 {
546         unsigned long min_count, ret;
547
548         /*
549          * Increase the pool size
550          * First take pages out of surplus state.  Then make up the
551          * remaining difference by allocating fresh huge pages.
552          *
553          * We might race with alloc_buddy_huge_page() here and be unable
554          * to convert a surplus huge page to a normal huge page. That is
555          * not critical, though, it just means the overall size of the
556          * pool might be one hugepage larger than it needs to be, but
557          * within all the constraints specified by the sysctls.
558          */
559         spin_lock(&hugetlb_lock);
560         while (surplus_huge_pages && count > persistent_huge_pages) {
561                 if (!adjust_pool_surplus(-1))
562                         break;
563         }
564
565         while (count > persistent_huge_pages) {
566                 int ret;
567                 /*
568                  * If this allocation races such that we no longer need the
569                  * page, free_huge_page will handle it by freeing the page
570                  * and reducing the surplus.
571                  */
572                 spin_unlock(&hugetlb_lock);
573                 ret = alloc_fresh_huge_page();
574                 spin_lock(&hugetlb_lock);
575                 if (!ret)
576                         goto out;
577
578         }
579
580         /*
581          * Decrease the pool size
582          * First return free pages to the buddy allocator (being careful
583          * to keep enough around to satisfy reservations).  Then place
584          * pages into surplus state as needed so the pool will shrink
585          * to the desired size as pages become free.
586          *
587          * By placing pages into the surplus state independent of the
588          * overcommit value, we are allowing the surplus pool size to
589          * exceed overcommit. There are few sane options here. Since
590          * alloc_buddy_huge_page() is checking the global counter,
591          * though, we'll note that we're not allowed to exceed surplus
592          * and won't grow the pool anywhere else. Not until one of the
593          * sysctls are changed, or the surplus pages go out of use.
594          */
595         min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
596         min_count = max(count, min_count);
597         try_to_free_low(min_count);
598         while (min_count < persistent_huge_pages) {
599                 struct page *page = dequeue_huge_page();
600                 if (!page)
601                         break;
602                 update_and_free_page(page);
603         }
604         while (count < persistent_huge_pages) {
605                 if (!adjust_pool_surplus(1))
606                         break;
607         }
608 out:
609         ret = persistent_huge_pages;
610         spin_unlock(&hugetlb_lock);
611         return ret;
612 }
613
614 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
615                            struct file *file, void __user *buffer,
616                            size_t *length, loff_t *ppos)
617 {
618         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
619         max_huge_pages = set_max_huge_pages(max_huge_pages);
620         return 0;
621 }
622
623 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
624                         struct file *file, void __user *buffer,
625                         size_t *length, loff_t *ppos)
626 {
627         proc_dointvec(table, write, file, buffer, length, ppos);
628         if (hugepages_treat_as_movable)
629                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
630         else
631                 htlb_alloc_mask = GFP_HIGHUSER;
632         return 0;
633 }
634
635 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
636                         struct file *file, void __user *buffer,
637                         size_t *length, loff_t *ppos)
638 {
639         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
640         spin_lock(&hugetlb_lock);
641         nr_overcommit_huge_pages = sysctl_overcommit_huge_pages;
642         spin_unlock(&hugetlb_lock);
643         return 0;
644 }
645
646 #endif /* CONFIG_SYSCTL */
647
648 int hugetlb_report_meminfo(char *buf)
649 {
650         return sprintf(buf,
651                         "HugePages_Total: %5lu\n"
652                         "HugePages_Free:  %5lu\n"
653                         "HugePages_Rsvd:  %5lu\n"
654                         "HugePages_Surp:  %5lu\n"
655                         "Hugepagesize:    %5lu kB\n",
656                         nr_huge_pages,
657                         free_huge_pages,
658                         resv_huge_pages,
659                         surplus_huge_pages,
660                         HPAGE_SIZE/1024);
661 }
662
663 int hugetlb_report_node_meminfo(int nid, char *buf)
664 {
665         return sprintf(buf,
666                 "Node %d HugePages_Total: %5u\n"
667                 "Node %d HugePages_Free:  %5u\n",
668                 nid, nr_huge_pages_node[nid],
669                 nid, free_huge_pages_node[nid]);
670 }
671
672 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
673 unsigned long hugetlb_total_pages(void)
674 {
675         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
676 }
677
678 /*
679  * We cannot handle pagefaults against hugetlb pages at all.  They cause
680  * handle_mm_fault() to try to instantiate regular-sized pages in the
681  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
682  * this far.
683  */
684 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
685 {
686         BUG();
687         return 0;
688 }
689
690 struct vm_operations_struct hugetlb_vm_ops = {
691         .fault = hugetlb_vm_op_fault,
692 };
693
694 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
695                                 int writable)
696 {
697         pte_t entry;
698
699         if (writable) {
700                 entry =
701                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
702         } else {
703                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
704         }
705         entry = pte_mkyoung(entry);
706         entry = pte_mkhuge(entry);
707
708         return entry;
709 }
710
711 static void set_huge_ptep_writable(struct vm_area_struct *vma,
712                                    unsigned long address, pte_t *ptep)
713 {
714         pte_t entry;
715
716         entry = pte_mkwrite(pte_mkdirty(*ptep));
717         if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
718                 update_mmu_cache(vma, address, entry);
719         }
720 }
721
722
723 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
724                             struct vm_area_struct *vma)
725 {
726         pte_t *src_pte, *dst_pte, entry;
727         struct page *ptepage;
728         unsigned long addr;
729         int cow;
730
731         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
732
733         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
734                 src_pte = huge_pte_offset(src, addr);
735                 if (!src_pte)
736                         continue;
737                 dst_pte = huge_pte_alloc(dst, addr);
738                 if (!dst_pte)
739                         goto nomem;
740
741                 /* If the pagetables are shared don't copy or take references */
742                 if (dst_pte == src_pte)
743                         continue;
744
745                 spin_lock(&dst->page_table_lock);
746                 spin_lock(&src->page_table_lock);
747                 if (!pte_none(*src_pte)) {
748                         if (cow)
749                                 ptep_set_wrprotect(src, addr, src_pte);
750                         entry = *src_pte;
751                         ptepage = pte_page(entry);
752                         get_page(ptepage);
753                         set_huge_pte_at(dst, addr, dst_pte, entry);
754                 }
755                 spin_unlock(&src->page_table_lock);
756                 spin_unlock(&dst->page_table_lock);
757         }
758         return 0;
759
760 nomem:
761         return -ENOMEM;
762 }
763
764 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
765                             unsigned long end)
766 {
767         struct mm_struct *mm = vma->vm_mm;
768         unsigned long address;
769         pte_t *ptep;
770         pte_t pte;
771         struct page *page;
772         struct page *tmp;
773         /*
774          * A page gathering list, protected by per file i_mmap_lock. The
775          * lock is used to avoid list corruption from multiple unmapping
776          * of the same page since we are using page->lru.
777          */
778         LIST_HEAD(page_list);
779
780         WARN_ON(!is_vm_hugetlb_page(vma));
781         BUG_ON(start & ~HPAGE_MASK);
782         BUG_ON(end & ~HPAGE_MASK);
783
784         spin_lock(&mm->page_table_lock);
785         for (address = start; address < end; address += HPAGE_SIZE) {
786                 ptep = huge_pte_offset(mm, address);
787                 if (!ptep)
788                         continue;
789
790                 if (huge_pmd_unshare(mm, &address, ptep))
791                         continue;
792
793                 pte = huge_ptep_get_and_clear(mm, address, ptep);
794                 if (pte_none(pte))
795                         continue;
796
797                 page = pte_page(pte);
798                 if (pte_dirty(pte))
799                         set_page_dirty(page);
800                 list_add(&page->lru, &page_list);
801         }
802         spin_unlock(&mm->page_table_lock);
803         flush_tlb_range(vma, start, end);
804         list_for_each_entry_safe(page, tmp, &page_list, lru) {
805                 list_del(&page->lru);
806                 put_page(page);
807         }
808 }
809
810 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
811                           unsigned long end)
812 {
813         /*
814          * It is undesirable to test vma->vm_file as it should be non-null
815          * for valid hugetlb area. However, vm_file will be NULL in the error
816          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
817          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
818          * to clean up. Since no pte has actually been setup, it is safe to
819          * do nothing in this case.
820          */
821         if (vma->vm_file) {
822                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
823                 __unmap_hugepage_range(vma, start, end);
824                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
825         }
826 }
827
828 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
829                         unsigned long address, pte_t *ptep, pte_t pte)
830 {
831         struct page *old_page, *new_page;
832         int avoidcopy;
833
834         old_page = pte_page(pte);
835
836         /* If no-one else is actually using this page, avoid the copy
837          * and just make the page writable */
838         avoidcopy = (page_count(old_page) == 1);
839         if (avoidcopy) {
840                 set_huge_ptep_writable(vma, address, ptep);
841                 return 0;
842         }
843
844         page_cache_get(old_page);
845         new_page = alloc_huge_page(vma, address);
846
847         if (IS_ERR(new_page)) {
848                 page_cache_release(old_page);
849                 return -PTR_ERR(new_page);
850         }
851
852         spin_unlock(&mm->page_table_lock);
853         copy_huge_page(new_page, old_page, address, vma);
854         __SetPageUptodate(new_page);
855         spin_lock(&mm->page_table_lock);
856
857         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
858         if (likely(pte_same(*ptep, pte))) {
859                 /* Break COW */
860                 set_huge_pte_at(mm, address, ptep,
861                                 make_huge_pte(vma, new_page, 1));
862                 /* Make the old page be freed below */
863                 new_page = old_page;
864         }
865         page_cache_release(new_page);
866         page_cache_release(old_page);
867         return 0;
868 }
869
870 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
871                         unsigned long address, pte_t *ptep, int write_access)
872 {
873         int ret = VM_FAULT_SIGBUS;
874         unsigned long idx;
875         unsigned long size;
876         struct page *page;
877         struct address_space *mapping;
878         pte_t new_pte;
879
880         mapping = vma->vm_file->f_mapping;
881         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
882                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
883
884         /*
885          * Use page lock to guard against racing truncation
886          * before we get page_table_lock.
887          */
888 retry:
889         page = find_lock_page(mapping, idx);
890         if (!page) {
891                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
892                 if (idx >= size)
893                         goto out;
894                 page = alloc_huge_page(vma, address);
895                 if (IS_ERR(page)) {
896                         ret = -PTR_ERR(page);
897                         goto out;
898                 }
899                 clear_huge_page(page, address);
900                 __SetPageUptodate(page);
901
902                 if (vma->vm_flags & VM_SHARED) {
903                         int err;
904                         struct inode *inode = mapping->host;
905
906                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
907                         if (err) {
908                                 put_page(page);
909                                 if (err == -EEXIST)
910                                         goto retry;
911                                 goto out;
912                         }
913
914                         spin_lock(&inode->i_lock);
915                         inode->i_blocks += BLOCKS_PER_HUGEPAGE;
916                         spin_unlock(&inode->i_lock);
917                 } else
918                         lock_page(page);
919         }
920
921         spin_lock(&mm->page_table_lock);
922         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
923         if (idx >= size)
924                 goto backout;
925
926         ret = 0;
927         if (!pte_none(*ptep))
928                 goto backout;
929
930         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
931                                 && (vma->vm_flags & VM_SHARED)));
932         set_huge_pte_at(mm, address, ptep, new_pte);
933
934         if (write_access && !(vma->vm_flags & VM_SHARED)) {
935                 /* Optimization, do the COW without a second fault */
936                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
937         }
938
939         spin_unlock(&mm->page_table_lock);
940         unlock_page(page);
941 out:
942         return ret;
943
944 backout:
945         spin_unlock(&mm->page_table_lock);
946         unlock_page(page);
947         put_page(page);
948         goto out;
949 }
950
951 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
952                         unsigned long address, int write_access)
953 {
954         pte_t *ptep;
955         pte_t entry;
956         int ret;
957         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
958
959         ptep = huge_pte_alloc(mm, address);
960         if (!ptep)
961                 return VM_FAULT_OOM;
962
963         /*
964          * Serialize hugepage allocation and instantiation, so that we don't
965          * get spurious allocation failures if two CPUs race to instantiate
966          * the same page in the page cache.
967          */
968         mutex_lock(&hugetlb_instantiation_mutex);
969         entry = *ptep;
970         if (pte_none(entry)) {
971                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
972                 mutex_unlock(&hugetlb_instantiation_mutex);
973                 return ret;
974         }
975
976         ret = 0;
977
978         spin_lock(&mm->page_table_lock);
979         /* Check for a racing update before calling hugetlb_cow */
980         if (likely(pte_same(entry, *ptep)))
981                 if (write_access && !pte_write(entry))
982                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
983         spin_unlock(&mm->page_table_lock);
984         mutex_unlock(&hugetlb_instantiation_mutex);
985
986         return ret;
987 }
988
989 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
990                         struct page **pages, struct vm_area_struct **vmas,
991                         unsigned long *position, int *length, int i,
992                         int write)
993 {
994         unsigned long pfn_offset;
995         unsigned long vaddr = *position;
996         int remainder = *length;
997
998         spin_lock(&mm->page_table_lock);
999         while (vaddr < vma->vm_end && remainder) {
1000                 pte_t *pte;
1001                 struct page *page;
1002
1003                 /*
1004                  * Some archs (sparc64, sh*) have multiple pte_ts to
1005                  * each hugepage.  We have to make * sure we get the
1006                  * first, for the page indexing below to work.
1007                  */
1008                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
1009
1010                 if (!pte || pte_none(*pte) || (write && !pte_write(*pte))) {
1011                         int ret;
1012
1013                         spin_unlock(&mm->page_table_lock);
1014                         ret = hugetlb_fault(mm, vma, vaddr, write);
1015                         spin_lock(&mm->page_table_lock);
1016                         if (!(ret & VM_FAULT_ERROR))
1017                                 continue;
1018
1019                         remainder = 0;
1020                         if (!i)
1021                                 i = -EFAULT;
1022                         break;
1023                 }
1024
1025                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
1026                 page = pte_page(*pte);
1027 same_page:
1028                 if (pages) {
1029                         get_page(page);
1030                         pages[i] = page + pfn_offset;
1031                 }
1032
1033                 if (vmas)
1034                         vmas[i] = vma;
1035
1036                 vaddr += PAGE_SIZE;
1037                 ++pfn_offset;
1038                 --remainder;
1039                 ++i;
1040                 if (vaddr < vma->vm_end && remainder &&
1041                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
1042                         /*
1043                          * We use pfn_offset to avoid touching the pageframes
1044                          * of this compound page.
1045                          */
1046                         goto same_page;
1047                 }
1048         }
1049         spin_unlock(&mm->page_table_lock);
1050         *length = remainder;
1051         *position = vaddr;
1052
1053         return i;
1054 }
1055
1056 void hugetlb_change_protection(struct vm_area_struct *vma,
1057                 unsigned long address, unsigned long end, pgprot_t newprot)
1058 {
1059         struct mm_struct *mm = vma->vm_mm;
1060         unsigned long start = address;
1061         pte_t *ptep;
1062         pte_t pte;
1063
1064         BUG_ON(address >= end);
1065         flush_cache_range(vma, address, end);
1066
1067         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1068         spin_lock(&mm->page_table_lock);
1069         for (; address < end; address += HPAGE_SIZE) {
1070                 ptep = huge_pte_offset(mm, address);
1071                 if (!ptep)
1072                         continue;
1073                 if (huge_pmd_unshare(mm, &address, ptep))
1074                         continue;
1075                 if (!pte_none(*ptep)) {
1076                         pte = huge_ptep_get_and_clear(mm, address, ptep);
1077                         pte = pte_mkhuge(pte_modify(pte, newprot));
1078                         set_huge_pte_at(mm, address, ptep, pte);
1079                 }
1080         }
1081         spin_unlock(&mm->page_table_lock);
1082         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1083
1084         flush_tlb_range(vma, start, end);
1085 }
1086
1087 struct file_region {
1088         struct list_head link;
1089         long from;
1090         long to;
1091 };
1092
1093 static long region_add(struct list_head *head, long f, long t)
1094 {
1095         struct file_region *rg, *nrg, *trg;
1096
1097         /* Locate the region we are either in or before. */
1098         list_for_each_entry(rg, head, link)
1099                 if (f <= rg->to)
1100                         break;
1101
1102         /* Round our left edge to the current segment if it encloses us. */
1103         if (f > rg->from)
1104                 f = rg->from;
1105
1106         /* Check for and consume any regions we now overlap with. */
1107         nrg = rg;
1108         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1109                 if (&rg->link == head)
1110                         break;
1111                 if (rg->from > t)
1112                         break;
1113
1114                 /* If this area reaches higher then extend our area to
1115                  * include it completely.  If this is not the first area
1116                  * which we intend to reuse, free it. */
1117                 if (rg->to > t)
1118                         t = rg->to;
1119                 if (rg != nrg) {
1120                         list_del(&rg->link);
1121                         kfree(rg);
1122                 }
1123         }
1124         nrg->from = f;
1125         nrg->to = t;
1126         return 0;
1127 }
1128
1129 static long region_chg(struct list_head *head, long f, long t)
1130 {
1131         struct file_region *rg, *nrg;
1132         long chg = 0;
1133
1134         /* Locate the region we are before or in. */
1135         list_for_each_entry(rg, head, link)
1136                 if (f <= rg->to)
1137                         break;
1138
1139         /* If we are below the current region then a new region is required.
1140          * Subtle, allocate a new region at the position but make it zero
1141          * size such that we can guarantee to record the reservation. */
1142         if (&rg->link == head || t < rg->from) {
1143                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
1144                 if (!nrg)
1145                         return -ENOMEM;
1146                 nrg->from = f;
1147                 nrg->to   = f;
1148                 INIT_LIST_HEAD(&nrg->link);
1149                 list_add(&nrg->link, rg->link.prev);
1150
1151                 return t - f;
1152         }
1153
1154         /* Round our left edge to the current segment if it encloses us. */
1155         if (f > rg->from)
1156                 f = rg->from;
1157         chg = t - f;
1158
1159         /* Check for and consume any regions we now overlap with. */
1160         list_for_each_entry(rg, rg->link.prev, link) {
1161                 if (&rg->link == head)
1162                         break;
1163                 if (rg->from > t)
1164                         return chg;
1165
1166                 /* We overlap with this area, if it extends futher than
1167                  * us then we must extend ourselves.  Account for its
1168                  * existing reservation. */
1169                 if (rg->to > t) {
1170                         chg += rg->to - t;
1171                         t = rg->to;
1172                 }
1173                 chg -= rg->to - rg->from;
1174         }
1175         return chg;
1176 }
1177
1178 static long region_truncate(struct list_head *head, long end)
1179 {
1180         struct file_region *rg, *trg;
1181         long chg = 0;
1182
1183         /* Locate the region we are either in or before. */
1184         list_for_each_entry(rg, head, link)
1185                 if (end <= rg->to)
1186                         break;
1187         if (&rg->link == head)
1188                 return 0;
1189
1190         /* If we are in the middle of a region then adjust it. */
1191         if (end > rg->from) {
1192                 chg = rg->to - end;
1193                 rg->to = end;
1194                 rg = list_entry(rg->link.next, typeof(*rg), link);
1195         }
1196
1197         /* Drop any remaining regions. */
1198         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1199                 if (&rg->link == head)
1200                         break;
1201                 chg += rg->to - rg->from;
1202                 list_del(&rg->link);
1203                 kfree(rg);
1204         }
1205         return chg;
1206 }
1207
1208 static int hugetlb_acct_memory(long delta)
1209 {
1210         int ret = -ENOMEM;
1211
1212         spin_lock(&hugetlb_lock);
1213         /*
1214          * When cpuset is configured, it breaks the strict hugetlb page
1215          * reservation as the accounting is done on a global variable. Such
1216          * reservation is completely rubbish in the presence of cpuset because
1217          * the reservation is not checked against page availability for the
1218          * current cpuset. Application can still potentially OOM'ed by kernel
1219          * with lack of free htlb page in cpuset that the task is in.
1220          * Attempt to enforce strict accounting with cpuset is almost
1221          * impossible (or too ugly) because cpuset is too fluid that
1222          * task or memory node can be dynamically moved between cpusets.
1223          *
1224          * The change of semantics for shared hugetlb mapping with cpuset is
1225          * undesirable. However, in order to preserve some of the semantics,
1226          * we fall back to check against current free page availability as
1227          * a best attempt and hopefully to minimize the impact of changing
1228          * semantics that cpuset has.
1229          */
1230         if (delta > 0) {
1231                 if (gather_surplus_pages(delta) < 0)
1232                         goto out;
1233
1234                 if (delta > cpuset_mems_nr(free_huge_pages_node)) {
1235                         return_unused_surplus_pages(delta);
1236                         goto out;
1237                 }
1238         }
1239
1240         ret = 0;
1241         if (delta < 0)
1242                 return_unused_surplus_pages((unsigned long) -delta);
1243
1244 out:
1245         spin_unlock(&hugetlb_lock);
1246         return ret;
1247 }
1248
1249 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
1250 {
1251         long ret, chg;
1252
1253         chg = region_chg(&inode->i_mapping->private_list, from, to);
1254         if (chg < 0)
1255                 return chg;
1256
1257         if (hugetlb_get_quota(inode->i_mapping, chg))
1258                 return -ENOSPC;
1259         ret = hugetlb_acct_memory(chg);
1260         if (ret < 0) {
1261                 hugetlb_put_quota(inode->i_mapping, chg);
1262                 return ret;
1263         }
1264         region_add(&inode->i_mapping->private_list, from, to);
1265         return 0;
1266 }
1267
1268 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1269 {
1270         long chg = region_truncate(&inode->i_mapping->private_list, offset);
1271
1272         spin_lock(&inode->i_lock);
1273         inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed;
1274         spin_unlock(&inode->i_lock);
1275
1276         hugetlb_put_quota(inode->i_mapping, (chg - freed));
1277         hugetlb_acct_memory(-(chg - freed));
1278 }