hugetlb: indicate surplus huge page counts in per-node meminfo
[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                 /*
290                  * This page is now managed by the hugetlb allocator and has
291                  * no users -- drop the buddy allocator's reference.
292                  */
293                 put_page_testzero(page);
294                 VM_BUG_ON(page_count(page));
295                 nid = page_to_nid(page);
296                 set_compound_page_dtor(page, free_huge_page);
297                 /*
298                  * We incremented the global counters already
299                  */
300                 nr_huge_pages_node[nid]++;
301                 surplus_huge_pages_node[nid]++;
302         } else {
303                 nr_huge_pages--;
304                 surplus_huge_pages--;
305         }
306         spin_unlock(&hugetlb_lock);
307
308         return page;
309 }
310
311 /*
312  * Increase the hugetlb pool such that it can accomodate a reservation
313  * of size 'delta'.
314  */
315 static int gather_surplus_pages(int delta)
316 {
317         struct list_head surplus_list;
318         struct page *page, *tmp;
319         int ret, i;
320         int needed, allocated;
321
322         needed = (resv_huge_pages + delta) - free_huge_pages;
323         if (needed <= 0) {
324                 resv_huge_pages += delta;
325                 return 0;
326         }
327
328         allocated = 0;
329         INIT_LIST_HEAD(&surplus_list);
330
331         ret = -ENOMEM;
332 retry:
333         spin_unlock(&hugetlb_lock);
334         for (i = 0; i < needed; i++) {
335                 page = alloc_buddy_huge_page(NULL, 0);
336                 if (!page) {
337                         /*
338                          * We were not able to allocate enough pages to
339                          * satisfy the entire reservation so we free what
340                          * we've allocated so far.
341                          */
342                         spin_lock(&hugetlb_lock);
343                         needed = 0;
344                         goto free;
345                 }
346
347                 list_add(&page->lru, &surplus_list);
348         }
349         allocated += needed;
350
351         /*
352          * After retaking hugetlb_lock, we need to recalculate 'needed'
353          * because either resv_huge_pages or free_huge_pages may have changed.
354          */
355         spin_lock(&hugetlb_lock);
356         needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
357         if (needed > 0)
358                 goto retry;
359
360         /*
361          * The surplus_list now contains _at_least_ the number of extra pages
362          * needed to accomodate the reservation.  Add the appropriate number
363          * of pages to the hugetlb pool and free the extras back to the buddy
364          * allocator.  Commit the entire reservation here to prevent another
365          * process from stealing the pages as they are added to the pool but
366          * before they are reserved.
367          */
368         needed += allocated;
369         resv_huge_pages += delta;
370         ret = 0;
371 free:
372         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
373                 list_del(&page->lru);
374                 if ((--needed) >= 0)
375                         enqueue_huge_page(page);
376                 else {
377                         /*
378                          * The page has a reference count of zero already, so
379                          * call free_huge_page directly instead of using
380                          * put_page.  This must be done with hugetlb_lock
381                          * unlocked which is safe because free_huge_page takes
382                          * hugetlb_lock before deciding how to free the page.
383                          */
384                         spin_unlock(&hugetlb_lock);
385                         free_huge_page(page);
386                         spin_lock(&hugetlb_lock);
387                 }
388         }
389
390         return ret;
391 }
392
393 /*
394  * When releasing a hugetlb pool reservation, any surplus pages that were
395  * allocated to satisfy the reservation must be explicitly freed if they were
396  * never used.
397  */
398 static void return_unused_surplus_pages(unsigned long unused_resv_pages)
399 {
400         static int nid = -1;
401         struct page *page;
402         unsigned long nr_pages;
403
404         /* Uncommit the reservation */
405         resv_huge_pages -= unused_resv_pages;
406
407         nr_pages = min(unused_resv_pages, surplus_huge_pages);
408
409         while (nr_pages) {
410                 nid = next_node(nid, node_online_map);
411                 if (nid == MAX_NUMNODES)
412                         nid = first_node(node_online_map);
413
414                 if (!surplus_huge_pages_node[nid])
415                         continue;
416
417                 if (!list_empty(&hugepage_freelists[nid])) {
418                         page = list_entry(hugepage_freelists[nid].next,
419                                           struct page, lru);
420                         list_del(&page->lru);
421                         update_and_free_page(page);
422                         free_huge_pages--;
423                         free_huge_pages_node[nid]--;
424                         surplus_huge_pages--;
425                         surplus_huge_pages_node[nid]--;
426                         nr_pages--;
427                 }
428         }
429 }
430
431
432 static struct page *alloc_huge_page_shared(struct vm_area_struct *vma,
433                                                 unsigned long addr)
434 {
435         struct page *page;
436
437         spin_lock(&hugetlb_lock);
438         page = dequeue_huge_page_vma(vma, addr);
439         spin_unlock(&hugetlb_lock);
440         return page ? page : ERR_PTR(-VM_FAULT_OOM);
441 }
442
443 static struct page *alloc_huge_page_private(struct vm_area_struct *vma,
444                                                 unsigned long addr)
445 {
446         struct page *page = NULL;
447
448         if (hugetlb_get_quota(vma->vm_file->f_mapping, 1))
449                 return ERR_PTR(-VM_FAULT_SIGBUS);
450
451         spin_lock(&hugetlb_lock);
452         if (free_huge_pages > resv_huge_pages)
453                 page = dequeue_huge_page_vma(vma, addr);
454         spin_unlock(&hugetlb_lock);
455         if (!page) {
456                 page = alloc_buddy_huge_page(vma, addr);
457                 if (!page) {
458                         hugetlb_put_quota(vma->vm_file->f_mapping, 1);
459                         return ERR_PTR(-VM_FAULT_OOM);
460                 }
461         }
462         return page;
463 }
464
465 static struct page *alloc_huge_page(struct vm_area_struct *vma,
466                                     unsigned long addr)
467 {
468         struct page *page;
469         struct address_space *mapping = vma->vm_file->f_mapping;
470
471         if (vma->vm_flags & VM_MAYSHARE)
472                 page = alloc_huge_page_shared(vma, addr);
473         else
474                 page = alloc_huge_page_private(vma, addr);
475
476         if (!IS_ERR(page)) {
477                 set_page_refcounted(page);
478                 set_page_private(page, (unsigned long) mapping);
479         }
480         return page;
481 }
482
483 static int __init hugetlb_init(void)
484 {
485         unsigned long i;
486
487         if (HPAGE_SHIFT == 0)
488                 return 0;
489
490         for (i = 0; i < MAX_NUMNODES; ++i)
491                 INIT_LIST_HEAD(&hugepage_freelists[i]);
492
493         hugetlb_next_nid = first_node(node_online_map);
494
495         for (i = 0; i < max_huge_pages; ++i) {
496                 if (!alloc_fresh_huge_page())
497                         break;
498         }
499         max_huge_pages = free_huge_pages = nr_huge_pages = i;
500         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
501         return 0;
502 }
503 module_init(hugetlb_init);
504
505 static int __init hugetlb_setup(char *s)
506 {
507         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
508                 max_huge_pages = 0;
509         return 1;
510 }
511 __setup("hugepages=", hugetlb_setup);
512
513 static unsigned int cpuset_mems_nr(unsigned int *array)
514 {
515         int node;
516         unsigned int nr = 0;
517
518         for_each_node_mask(node, cpuset_current_mems_allowed)
519                 nr += array[node];
520
521         return nr;
522 }
523
524 #ifdef CONFIG_SYSCTL
525 #ifdef CONFIG_HIGHMEM
526 static void try_to_free_low(unsigned long count)
527 {
528         int i;
529
530         for (i = 0; i < MAX_NUMNODES; ++i) {
531                 struct page *page, *next;
532                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
533                         if (count >= nr_huge_pages)
534                                 return;
535                         if (PageHighMem(page))
536                                 continue;
537                         list_del(&page->lru);
538                         update_and_free_page(page);
539                         free_huge_pages--;
540                         free_huge_pages_node[page_to_nid(page)]--;
541                 }
542         }
543 }
544 #else
545 static inline void try_to_free_low(unsigned long count)
546 {
547 }
548 #endif
549
550 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
551 static unsigned long set_max_huge_pages(unsigned long count)
552 {
553         unsigned long min_count, ret;
554
555         /*
556          * Increase the pool size
557          * First take pages out of surplus state.  Then make up the
558          * remaining difference by allocating fresh huge pages.
559          *
560          * We might race with alloc_buddy_huge_page() here and be unable
561          * to convert a surplus huge page to a normal huge page. That is
562          * not critical, though, it just means the overall size of the
563          * pool might be one hugepage larger than it needs to be, but
564          * within all the constraints specified by the sysctls.
565          */
566         spin_lock(&hugetlb_lock);
567         while (surplus_huge_pages && count > persistent_huge_pages) {
568                 if (!adjust_pool_surplus(-1))
569                         break;
570         }
571
572         while (count > persistent_huge_pages) {
573                 int ret;
574                 /*
575                  * If this allocation races such that we no longer need the
576                  * page, free_huge_page will handle it by freeing the page
577                  * and reducing the surplus.
578                  */
579                 spin_unlock(&hugetlb_lock);
580                 ret = alloc_fresh_huge_page();
581                 spin_lock(&hugetlb_lock);
582                 if (!ret)
583                         goto out;
584
585         }
586
587         /*
588          * Decrease the pool size
589          * First return free pages to the buddy allocator (being careful
590          * to keep enough around to satisfy reservations).  Then place
591          * pages into surplus state as needed so the pool will shrink
592          * to the desired size as pages become free.
593          *
594          * By placing pages into the surplus state independent of the
595          * overcommit value, we are allowing the surplus pool size to
596          * exceed overcommit. There are few sane options here. Since
597          * alloc_buddy_huge_page() is checking the global counter,
598          * though, we'll note that we're not allowed to exceed surplus
599          * and won't grow the pool anywhere else. Not until one of the
600          * sysctls are changed, or the surplus pages go out of use.
601          */
602         min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
603         min_count = max(count, min_count);
604         try_to_free_low(min_count);
605         while (min_count < persistent_huge_pages) {
606                 struct page *page = dequeue_huge_page();
607                 if (!page)
608                         break;
609                 update_and_free_page(page);
610         }
611         while (count < persistent_huge_pages) {
612                 if (!adjust_pool_surplus(1))
613                         break;
614         }
615 out:
616         ret = persistent_huge_pages;
617         spin_unlock(&hugetlb_lock);
618         return ret;
619 }
620
621 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
622                            struct file *file, void __user *buffer,
623                            size_t *length, loff_t *ppos)
624 {
625         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
626         max_huge_pages = set_max_huge_pages(max_huge_pages);
627         return 0;
628 }
629
630 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
631                         struct file *file, void __user *buffer,
632                         size_t *length, loff_t *ppos)
633 {
634         proc_dointvec(table, write, file, buffer, length, ppos);
635         if (hugepages_treat_as_movable)
636                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
637         else
638                 htlb_alloc_mask = GFP_HIGHUSER;
639         return 0;
640 }
641
642 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
643                         struct file *file, void __user *buffer,
644                         size_t *length, loff_t *ppos)
645 {
646         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
647         spin_lock(&hugetlb_lock);
648         nr_overcommit_huge_pages = sysctl_overcommit_huge_pages;
649         spin_unlock(&hugetlb_lock);
650         return 0;
651 }
652
653 #endif /* CONFIG_SYSCTL */
654
655 int hugetlb_report_meminfo(char *buf)
656 {
657         return sprintf(buf,
658                         "HugePages_Total: %5lu\n"
659                         "HugePages_Free:  %5lu\n"
660                         "HugePages_Rsvd:  %5lu\n"
661                         "HugePages_Surp:  %5lu\n"
662                         "Hugepagesize:    %5lu kB\n",
663                         nr_huge_pages,
664                         free_huge_pages,
665                         resv_huge_pages,
666                         surplus_huge_pages,
667                         HPAGE_SIZE/1024);
668 }
669
670 int hugetlb_report_node_meminfo(int nid, char *buf)
671 {
672         return sprintf(buf,
673                 "Node %d HugePages_Total: %5u\n"
674                 "Node %d HugePages_Free:  %5u\n"
675                 "Node %d HugePages_Surp:  %5u\n",
676                 nid, nr_huge_pages_node[nid],
677                 nid, free_huge_pages_node[nid],
678                 nid, surplus_huge_pages_node[nid]);
679 }
680
681 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
682 unsigned long hugetlb_total_pages(void)
683 {
684         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
685 }
686
687 /*
688  * We cannot handle pagefaults against hugetlb pages at all.  They cause
689  * handle_mm_fault() to try to instantiate regular-sized pages in the
690  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
691  * this far.
692  */
693 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
694 {
695         BUG();
696         return 0;
697 }
698
699 struct vm_operations_struct hugetlb_vm_ops = {
700         .fault = hugetlb_vm_op_fault,
701 };
702
703 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
704                                 int writable)
705 {
706         pte_t entry;
707
708         if (writable) {
709                 entry =
710                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
711         } else {
712                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
713         }
714         entry = pte_mkyoung(entry);
715         entry = pte_mkhuge(entry);
716
717         return entry;
718 }
719
720 static void set_huge_ptep_writable(struct vm_area_struct *vma,
721                                    unsigned long address, pte_t *ptep)
722 {
723         pte_t entry;
724
725         entry = pte_mkwrite(pte_mkdirty(*ptep));
726         if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
727                 update_mmu_cache(vma, address, entry);
728         }
729 }
730
731
732 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
733                             struct vm_area_struct *vma)
734 {
735         pte_t *src_pte, *dst_pte, entry;
736         struct page *ptepage;
737         unsigned long addr;
738         int cow;
739
740         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
741
742         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
743                 src_pte = huge_pte_offset(src, addr);
744                 if (!src_pte)
745                         continue;
746                 dst_pte = huge_pte_alloc(dst, addr);
747                 if (!dst_pte)
748                         goto nomem;
749
750                 /* If the pagetables are shared don't copy or take references */
751                 if (dst_pte == src_pte)
752                         continue;
753
754                 spin_lock(&dst->page_table_lock);
755                 spin_lock(&src->page_table_lock);
756                 if (!pte_none(*src_pte)) {
757                         if (cow)
758                                 ptep_set_wrprotect(src, addr, src_pte);
759                         entry = *src_pte;
760                         ptepage = pte_page(entry);
761                         get_page(ptepage);
762                         set_huge_pte_at(dst, addr, dst_pte, entry);
763                 }
764                 spin_unlock(&src->page_table_lock);
765                 spin_unlock(&dst->page_table_lock);
766         }
767         return 0;
768
769 nomem:
770         return -ENOMEM;
771 }
772
773 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
774                             unsigned long end)
775 {
776         struct mm_struct *mm = vma->vm_mm;
777         unsigned long address;
778         pte_t *ptep;
779         pte_t pte;
780         struct page *page;
781         struct page *tmp;
782         /*
783          * A page gathering list, protected by per file i_mmap_lock. The
784          * lock is used to avoid list corruption from multiple unmapping
785          * of the same page since we are using page->lru.
786          */
787         LIST_HEAD(page_list);
788
789         WARN_ON(!is_vm_hugetlb_page(vma));
790         BUG_ON(start & ~HPAGE_MASK);
791         BUG_ON(end & ~HPAGE_MASK);
792
793         spin_lock(&mm->page_table_lock);
794         for (address = start; address < end; address += HPAGE_SIZE) {
795                 ptep = huge_pte_offset(mm, address);
796                 if (!ptep)
797                         continue;
798
799                 if (huge_pmd_unshare(mm, &address, ptep))
800                         continue;
801
802                 pte = huge_ptep_get_and_clear(mm, address, ptep);
803                 if (pte_none(pte))
804                         continue;
805
806                 page = pte_page(pte);
807                 if (pte_dirty(pte))
808                         set_page_dirty(page);
809                 list_add(&page->lru, &page_list);
810         }
811         spin_unlock(&mm->page_table_lock);
812         flush_tlb_range(vma, start, end);
813         list_for_each_entry_safe(page, tmp, &page_list, lru) {
814                 list_del(&page->lru);
815                 put_page(page);
816         }
817 }
818
819 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
820                           unsigned long end)
821 {
822         /*
823          * It is undesirable to test vma->vm_file as it should be non-null
824          * for valid hugetlb area. However, vm_file will be NULL in the error
825          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
826          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
827          * to clean up. Since no pte has actually been setup, it is safe to
828          * do nothing in this case.
829          */
830         if (vma->vm_file) {
831                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
832                 __unmap_hugepage_range(vma, start, end);
833                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
834         }
835 }
836
837 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
838                         unsigned long address, pte_t *ptep, pte_t pte)
839 {
840         struct page *old_page, *new_page;
841         int avoidcopy;
842
843         old_page = pte_page(pte);
844
845         /* If no-one else is actually using this page, avoid the copy
846          * and just make the page writable */
847         avoidcopy = (page_count(old_page) == 1);
848         if (avoidcopy) {
849                 set_huge_ptep_writable(vma, address, ptep);
850                 return 0;
851         }
852
853         page_cache_get(old_page);
854         new_page = alloc_huge_page(vma, address);
855
856         if (IS_ERR(new_page)) {
857                 page_cache_release(old_page);
858                 return -PTR_ERR(new_page);
859         }
860
861         spin_unlock(&mm->page_table_lock);
862         copy_huge_page(new_page, old_page, address, vma);
863         __SetPageUptodate(new_page);
864         spin_lock(&mm->page_table_lock);
865
866         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
867         if (likely(pte_same(*ptep, pte))) {
868                 /* Break COW */
869                 set_huge_pte_at(mm, address, ptep,
870                                 make_huge_pte(vma, new_page, 1));
871                 /* Make the old page be freed below */
872                 new_page = old_page;
873         }
874         page_cache_release(new_page);
875         page_cache_release(old_page);
876         return 0;
877 }
878
879 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
880                         unsigned long address, pte_t *ptep, int write_access)
881 {
882         int ret = VM_FAULT_SIGBUS;
883         unsigned long idx;
884         unsigned long size;
885         struct page *page;
886         struct address_space *mapping;
887         pte_t new_pte;
888
889         mapping = vma->vm_file->f_mapping;
890         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
891                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
892
893         /*
894          * Use page lock to guard against racing truncation
895          * before we get page_table_lock.
896          */
897 retry:
898         page = find_lock_page(mapping, idx);
899         if (!page) {
900                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
901                 if (idx >= size)
902                         goto out;
903                 page = alloc_huge_page(vma, address);
904                 if (IS_ERR(page)) {
905                         ret = -PTR_ERR(page);
906                         goto out;
907                 }
908                 clear_huge_page(page, address);
909                 __SetPageUptodate(page);
910
911                 if (vma->vm_flags & VM_SHARED) {
912                         int err;
913                         struct inode *inode = mapping->host;
914
915                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
916                         if (err) {
917                                 put_page(page);
918                                 if (err == -EEXIST)
919                                         goto retry;
920                                 goto out;
921                         }
922
923                         spin_lock(&inode->i_lock);
924                         inode->i_blocks += BLOCKS_PER_HUGEPAGE;
925                         spin_unlock(&inode->i_lock);
926                 } else
927                         lock_page(page);
928         }
929
930         spin_lock(&mm->page_table_lock);
931         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
932         if (idx >= size)
933                 goto backout;
934
935         ret = 0;
936         if (!pte_none(*ptep))
937                 goto backout;
938
939         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
940                                 && (vma->vm_flags & VM_SHARED)));
941         set_huge_pte_at(mm, address, ptep, new_pte);
942
943         if (write_access && !(vma->vm_flags & VM_SHARED)) {
944                 /* Optimization, do the COW without a second fault */
945                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
946         }
947
948         spin_unlock(&mm->page_table_lock);
949         unlock_page(page);
950 out:
951         return ret;
952
953 backout:
954         spin_unlock(&mm->page_table_lock);
955         unlock_page(page);
956         put_page(page);
957         goto out;
958 }
959
960 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
961                         unsigned long address, int write_access)
962 {
963         pte_t *ptep;
964         pte_t entry;
965         int ret;
966         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
967
968         ptep = huge_pte_alloc(mm, address);
969         if (!ptep)
970                 return VM_FAULT_OOM;
971
972         /*
973          * Serialize hugepage allocation and instantiation, so that we don't
974          * get spurious allocation failures if two CPUs race to instantiate
975          * the same page in the page cache.
976          */
977         mutex_lock(&hugetlb_instantiation_mutex);
978         entry = *ptep;
979         if (pte_none(entry)) {
980                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
981                 mutex_unlock(&hugetlb_instantiation_mutex);
982                 return ret;
983         }
984
985         ret = 0;
986
987         spin_lock(&mm->page_table_lock);
988         /* Check for a racing update before calling hugetlb_cow */
989         if (likely(pte_same(entry, *ptep)))
990                 if (write_access && !pte_write(entry))
991                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
992         spin_unlock(&mm->page_table_lock);
993         mutex_unlock(&hugetlb_instantiation_mutex);
994
995         return ret;
996 }
997
998 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
999                         struct page **pages, struct vm_area_struct **vmas,
1000                         unsigned long *position, int *length, int i,
1001                         int write)
1002 {
1003         unsigned long pfn_offset;
1004         unsigned long vaddr = *position;
1005         int remainder = *length;
1006
1007         spin_lock(&mm->page_table_lock);
1008         while (vaddr < vma->vm_end && remainder) {
1009                 pte_t *pte;
1010                 struct page *page;
1011
1012                 /*
1013                  * Some archs (sparc64, sh*) have multiple pte_ts to
1014                  * each hugepage.  We have to make * sure we get the
1015                  * first, for the page indexing below to work.
1016                  */
1017                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
1018
1019                 if (!pte || pte_none(*pte) || (write && !pte_write(*pte))) {
1020                         int ret;
1021
1022                         spin_unlock(&mm->page_table_lock);
1023                         ret = hugetlb_fault(mm, vma, vaddr, write);
1024                         spin_lock(&mm->page_table_lock);
1025                         if (!(ret & VM_FAULT_ERROR))
1026                                 continue;
1027
1028                         remainder = 0;
1029                         if (!i)
1030                                 i = -EFAULT;
1031                         break;
1032                 }
1033
1034                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
1035                 page = pte_page(*pte);
1036 same_page:
1037                 if (pages) {
1038                         get_page(page);
1039                         pages[i] = page + pfn_offset;
1040                 }
1041
1042                 if (vmas)
1043                         vmas[i] = vma;
1044
1045                 vaddr += PAGE_SIZE;
1046                 ++pfn_offset;
1047                 --remainder;
1048                 ++i;
1049                 if (vaddr < vma->vm_end && remainder &&
1050                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
1051                         /*
1052                          * We use pfn_offset to avoid touching the pageframes
1053                          * of this compound page.
1054                          */
1055                         goto same_page;
1056                 }
1057         }
1058         spin_unlock(&mm->page_table_lock);
1059         *length = remainder;
1060         *position = vaddr;
1061
1062         return i;
1063 }
1064
1065 void hugetlb_change_protection(struct vm_area_struct *vma,
1066                 unsigned long address, unsigned long end, pgprot_t newprot)
1067 {
1068         struct mm_struct *mm = vma->vm_mm;
1069         unsigned long start = address;
1070         pte_t *ptep;
1071         pte_t pte;
1072
1073         BUG_ON(address >= end);
1074         flush_cache_range(vma, address, end);
1075
1076         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1077         spin_lock(&mm->page_table_lock);
1078         for (; address < end; address += HPAGE_SIZE) {
1079                 ptep = huge_pte_offset(mm, address);
1080                 if (!ptep)
1081                         continue;
1082                 if (huge_pmd_unshare(mm, &address, ptep))
1083                         continue;
1084                 if (!pte_none(*ptep)) {
1085                         pte = huge_ptep_get_and_clear(mm, address, ptep);
1086                         pte = pte_mkhuge(pte_modify(pte, newprot));
1087                         set_huge_pte_at(mm, address, ptep, pte);
1088                 }
1089         }
1090         spin_unlock(&mm->page_table_lock);
1091         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1092
1093         flush_tlb_range(vma, start, end);
1094 }
1095
1096 struct file_region {
1097         struct list_head link;
1098         long from;
1099         long to;
1100 };
1101
1102 static long region_add(struct list_head *head, long f, long t)
1103 {
1104         struct file_region *rg, *nrg, *trg;
1105
1106         /* Locate the region we are either in or before. */
1107         list_for_each_entry(rg, head, link)
1108                 if (f <= rg->to)
1109                         break;
1110
1111         /* Round our left edge to the current segment if it encloses us. */
1112         if (f > rg->from)
1113                 f = rg->from;
1114
1115         /* Check for and consume any regions we now overlap with. */
1116         nrg = rg;
1117         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1118                 if (&rg->link == head)
1119                         break;
1120                 if (rg->from > t)
1121                         break;
1122
1123                 /* If this area reaches higher then extend our area to
1124                  * include it completely.  If this is not the first area
1125                  * which we intend to reuse, free it. */
1126                 if (rg->to > t)
1127                         t = rg->to;
1128                 if (rg != nrg) {
1129                         list_del(&rg->link);
1130                         kfree(rg);
1131                 }
1132         }
1133         nrg->from = f;
1134         nrg->to = t;
1135         return 0;
1136 }
1137
1138 static long region_chg(struct list_head *head, long f, long t)
1139 {
1140         struct file_region *rg, *nrg;
1141         long chg = 0;
1142
1143         /* Locate the region we are before or in. */
1144         list_for_each_entry(rg, head, link)
1145                 if (f <= rg->to)
1146                         break;
1147
1148         /* If we are below the current region then a new region is required.
1149          * Subtle, allocate a new region at the position but make it zero
1150          * size such that we can guarantee to record the reservation. */
1151         if (&rg->link == head || t < rg->from) {
1152                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
1153                 if (!nrg)
1154                         return -ENOMEM;
1155                 nrg->from = f;
1156                 nrg->to   = f;
1157                 INIT_LIST_HEAD(&nrg->link);
1158                 list_add(&nrg->link, rg->link.prev);
1159
1160                 return t - f;
1161         }
1162
1163         /* Round our left edge to the current segment if it encloses us. */
1164         if (f > rg->from)
1165                 f = rg->from;
1166         chg = t - f;
1167
1168         /* Check for and consume any regions we now overlap with. */
1169         list_for_each_entry(rg, rg->link.prev, link) {
1170                 if (&rg->link == head)
1171                         break;
1172                 if (rg->from > t)
1173                         return chg;
1174
1175                 /* We overlap with this area, if it extends futher than
1176                  * us then we must extend ourselves.  Account for its
1177                  * existing reservation. */
1178                 if (rg->to > t) {
1179                         chg += rg->to - t;
1180                         t = rg->to;
1181                 }
1182                 chg -= rg->to - rg->from;
1183         }
1184         return chg;
1185 }
1186
1187 static long region_truncate(struct list_head *head, long end)
1188 {
1189         struct file_region *rg, *trg;
1190         long chg = 0;
1191
1192         /* Locate the region we are either in or before. */
1193         list_for_each_entry(rg, head, link)
1194                 if (end <= rg->to)
1195                         break;
1196         if (&rg->link == head)
1197                 return 0;
1198
1199         /* If we are in the middle of a region then adjust it. */
1200         if (end > rg->from) {
1201                 chg = rg->to - end;
1202                 rg->to = end;
1203                 rg = list_entry(rg->link.next, typeof(*rg), link);
1204         }
1205
1206         /* Drop any remaining regions. */
1207         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1208                 if (&rg->link == head)
1209                         break;
1210                 chg += rg->to - rg->from;
1211                 list_del(&rg->link);
1212                 kfree(rg);
1213         }
1214         return chg;
1215 }
1216
1217 static int hugetlb_acct_memory(long delta)
1218 {
1219         int ret = -ENOMEM;
1220
1221         spin_lock(&hugetlb_lock);
1222         /*
1223          * When cpuset is configured, it breaks the strict hugetlb page
1224          * reservation as the accounting is done on a global variable. Such
1225          * reservation is completely rubbish in the presence of cpuset because
1226          * the reservation is not checked against page availability for the
1227          * current cpuset. Application can still potentially OOM'ed by kernel
1228          * with lack of free htlb page in cpuset that the task is in.
1229          * Attempt to enforce strict accounting with cpuset is almost
1230          * impossible (or too ugly) because cpuset is too fluid that
1231          * task or memory node can be dynamically moved between cpusets.
1232          *
1233          * The change of semantics for shared hugetlb mapping with cpuset is
1234          * undesirable. However, in order to preserve some of the semantics,
1235          * we fall back to check against current free page availability as
1236          * a best attempt and hopefully to minimize the impact of changing
1237          * semantics that cpuset has.
1238          */
1239         if (delta > 0) {
1240                 if (gather_surplus_pages(delta) < 0)
1241                         goto out;
1242
1243                 if (delta > cpuset_mems_nr(free_huge_pages_node)) {
1244                         return_unused_surplus_pages(delta);
1245                         goto out;
1246                 }
1247         }
1248
1249         ret = 0;
1250         if (delta < 0)
1251                 return_unused_surplus_pages((unsigned long) -delta);
1252
1253 out:
1254         spin_unlock(&hugetlb_lock);
1255         return ret;
1256 }
1257
1258 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
1259 {
1260         long ret, chg;
1261
1262         chg = region_chg(&inode->i_mapping->private_list, from, to);
1263         if (chg < 0)
1264                 return chg;
1265
1266         if (hugetlb_get_quota(inode->i_mapping, chg))
1267                 return -ENOSPC;
1268         ret = hugetlb_acct_memory(chg);
1269         if (ret < 0) {
1270                 hugetlb_put_quota(inode->i_mapping, chg);
1271                 return ret;
1272         }
1273         region_add(&inode->i_mapping->private_list, from, to);
1274         return 0;
1275 }
1276
1277 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1278 {
1279         long chg = region_truncate(&inode->i_mapping->private_list, offset);
1280
1281         spin_lock(&inode->i_lock);
1282         inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed;
1283         spin_unlock(&inode->i_lock);
1284
1285         hugetlb_put_quota(inode->i_mapping, (chg - freed));
1286         hugetlb_acct_memory(-(chg - freed));
1287 }