hugetlb: allow huge page mappings to be created without reservations
[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 /*
44  * Region tracking -- allows tracking of reservations and instantiated pages
45  *                    across the pages in a mapping.
46  */
47 struct file_region {
48         struct list_head link;
49         long from;
50         long to;
51 };
52
53 static long region_add(struct list_head *head, long f, long t)
54 {
55         struct file_region *rg, *nrg, *trg;
56
57         /* Locate the region we are either in or before. */
58         list_for_each_entry(rg, head, link)
59                 if (f <= rg->to)
60                         break;
61
62         /* Round our left edge to the current segment if it encloses us. */
63         if (f > rg->from)
64                 f = rg->from;
65
66         /* Check for and consume any regions we now overlap with. */
67         nrg = rg;
68         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
69                 if (&rg->link == head)
70                         break;
71                 if (rg->from > t)
72                         break;
73
74                 /* If this area reaches higher then extend our area to
75                  * include it completely.  If this is not the first area
76                  * which we intend to reuse, free it. */
77                 if (rg->to > t)
78                         t = rg->to;
79                 if (rg != nrg) {
80                         list_del(&rg->link);
81                         kfree(rg);
82                 }
83         }
84         nrg->from = f;
85         nrg->to = t;
86         return 0;
87 }
88
89 static long region_chg(struct list_head *head, long f, long t)
90 {
91         struct file_region *rg, *nrg;
92         long chg = 0;
93
94         /* Locate the region we are before or in. */
95         list_for_each_entry(rg, head, link)
96                 if (f <= rg->to)
97                         break;
98
99         /* If we are below the current region then a new region is required.
100          * Subtle, allocate a new region at the position but make it zero
101          * size such that we can guarantee to record the reservation. */
102         if (&rg->link == head || t < rg->from) {
103                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
104                 if (!nrg)
105                         return -ENOMEM;
106                 nrg->from = f;
107                 nrg->to   = f;
108                 INIT_LIST_HEAD(&nrg->link);
109                 list_add(&nrg->link, rg->link.prev);
110
111                 return t - f;
112         }
113
114         /* Round our left edge to the current segment if it encloses us. */
115         if (f > rg->from)
116                 f = rg->from;
117         chg = t - f;
118
119         /* Check for and consume any regions we now overlap with. */
120         list_for_each_entry(rg, rg->link.prev, link) {
121                 if (&rg->link == head)
122                         break;
123                 if (rg->from > t)
124                         return chg;
125
126                 /* We overlap with this area, if it extends futher than
127                  * us then we must extend ourselves.  Account for its
128                  * existing reservation. */
129                 if (rg->to > t) {
130                         chg += rg->to - t;
131                         t = rg->to;
132                 }
133                 chg -= rg->to - rg->from;
134         }
135         return chg;
136 }
137
138 static long region_truncate(struct list_head *head, long end)
139 {
140         struct file_region *rg, *trg;
141         long chg = 0;
142
143         /* Locate the region we are either in or before. */
144         list_for_each_entry(rg, head, link)
145                 if (end <= rg->to)
146                         break;
147         if (&rg->link == head)
148                 return 0;
149
150         /* If we are in the middle of a region then adjust it. */
151         if (end > rg->from) {
152                 chg = rg->to - end;
153                 rg->to = end;
154                 rg = list_entry(rg->link.next, typeof(*rg), link);
155         }
156
157         /* Drop any remaining regions. */
158         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
159                 if (&rg->link == head)
160                         break;
161                 chg += rg->to - rg->from;
162                 list_del(&rg->link);
163                 kfree(rg);
164         }
165         return chg;
166 }
167
168 /*
169  * Convert the address within this vma to the page offset within
170  * the mapping, in base page units.
171  */
172 static pgoff_t vma_page_offset(struct vm_area_struct *vma,
173                                 unsigned long address)
174 {
175         return ((address - vma->vm_start) >> PAGE_SHIFT) +
176                                         (vma->vm_pgoff >> PAGE_SHIFT);
177 }
178
179 /*
180  * Convert the address within this vma to the page offset within
181  * the mapping, in pagecache page units; huge pages here.
182  */
183 static pgoff_t vma_pagecache_offset(struct vm_area_struct *vma,
184                                         unsigned long address)
185 {
186         return ((address - vma->vm_start) >> HPAGE_SHIFT) +
187                         (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
188 }
189
190 #define HPAGE_RESV_OWNER    (1UL << (BITS_PER_LONG - 1))
191 #define HPAGE_RESV_UNMAPPED (1UL << (BITS_PER_LONG - 2))
192 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
193 /*
194  * These helpers are used to track how many pages are reserved for
195  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
196  * is guaranteed to have their future faults succeed.
197  *
198  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
199  * the reserve counters are updated with the hugetlb_lock held. It is safe
200  * to reset the VMA at fork() time as it is not in use yet and there is no
201  * chance of the global counters getting corrupted as a result of the values.
202  */
203 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
204 {
205         return (unsigned long)vma->vm_private_data;
206 }
207
208 static void set_vma_private_data(struct vm_area_struct *vma,
209                                                         unsigned long value)
210 {
211         vma->vm_private_data = (void *)value;
212 }
213
214 static unsigned long vma_resv_huge_pages(struct vm_area_struct *vma)
215 {
216         VM_BUG_ON(!is_vm_hugetlb_page(vma));
217         if (!(vma->vm_flags & VM_SHARED))
218                 return get_vma_private_data(vma) & ~HPAGE_RESV_MASK;
219         return 0;
220 }
221
222 static void set_vma_resv_huge_pages(struct vm_area_struct *vma,
223                                                         unsigned long reserve)
224 {
225         VM_BUG_ON(!is_vm_hugetlb_page(vma));
226         VM_BUG_ON(vma->vm_flags & VM_SHARED);
227
228         set_vma_private_data(vma,
229                 (get_vma_private_data(vma) & HPAGE_RESV_MASK) | reserve);
230 }
231
232 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
233 {
234         VM_BUG_ON(!is_vm_hugetlb_page(vma));
235         VM_BUG_ON(vma->vm_flags & VM_SHARED);
236
237         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
238 }
239
240 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
241 {
242         VM_BUG_ON(!is_vm_hugetlb_page(vma));
243
244         return (get_vma_private_data(vma) & flag) != 0;
245 }
246
247 /* Decrement the reserved pages in the hugepage pool by one */
248 static void decrement_hugepage_resv_vma(struct vm_area_struct *vma)
249 {
250         if (vma->vm_flags & VM_NORESERVE)
251                 return;
252
253         if (vma->vm_flags & VM_SHARED) {
254                 /* Shared mappings always use reserves */
255                 resv_huge_pages--;
256         } else {
257                 /*
258                  * Only the process that called mmap() has reserves for
259                  * private mappings.
260                  */
261                 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
262                         unsigned long flags, reserve;
263                         resv_huge_pages--;
264                         flags = (unsigned long)vma->vm_private_data &
265                                                         HPAGE_RESV_MASK;
266                         reserve = (unsigned long)vma->vm_private_data - 1;
267                         vma->vm_private_data = (void *)(reserve | flags);
268                 }
269         }
270 }
271
272 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
273 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
274 {
275         VM_BUG_ON(!is_vm_hugetlb_page(vma));
276         if (!(vma->vm_flags & VM_SHARED))
277                 vma->vm_private_data = (void *)0;
278 }
279
280 /* Returns true if the VMA has associated reserve pages */
281 static int vma_has_private_reserves(struct vm_area_struct *vma)
282 {
283         if (vma->vm_flags & VM_SHARED)
284                 return 0;
285         if (!vma_resv_huge_pages(vma))
286                 return 0;
287         return 1;
288 }
289
290 static void clear_huge_page(struct page *page, unsigned long addr)
291 {
292         int i;
293
294         might_sleep();
295         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
296                 cond_resched();
297                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
298         }
299 }
300
301 static void copy_huge_page(struct page *dst, struct page *src,
302                            unsigned long addr, struct vm_area_struct *vma)
303 {
304         int i;
305
306         might_sleep();
307         for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
308                 cond_resched();
309                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
310         }
311 }
312
313 static void enqueue_huge_page(struct page *page)
314 {
315         int nid = page_to_nid(page);
316         list_add(&page->lru, &hugepage_freelists[nid]);
317         free_huge_pages++;
318         free_huge_pages_node[nid]++;
319 }
320
321 static struct page *dequeue_huge_page(void)
322 {
323         int nid;
324         struct page *page = NULL;
325
326         for (nid = 0; nid < MAX_NUMNODES; ++nid) {
327                 if (!list_empty(&hugepage_freelists[nid])) {
328                         page = list_entry(hugepage_freelists[nid].next,
329                                           struct page, lru);
330                         list_del(&page->lru);
331                         free_huge_pages--;
332                         free_huge_pages_node[nid]--;
333                         break;
334                 }
335         }
336         return page;
337 }
338
339 static struct page *dequeue_huge_page_vma(struct vm_area_struct *vma,
340                                 unsigned long address, int avoid_reserve)
341 {
342         int nid;
343         struct page *page = NULL;
344         struct mempolicy *mpol;
345         nodemask_t *nodemask;
346         struct zonelist *zonelist = huge_zonelist(vma, address,
347                                         htlb_alloc_mask, &mpol, &nodemask);
348         struct zone *zone;
349         struct zoneref *z;
350
351         /*
352          * A child process with MAP_PRIVATE mappings created by their parent
353          * have no page reserves. This check ensures that reservations are
354          * not "stolen". The child may still get SIGKILLed
355          */
356         if (!vma_has_private_reserves(vma) &&
357                         free_huge_pages - resv_huge_pages == 0)
358                 return NULL;
359
360         /* If reserves cannot be used, ensure enough pages are in the pool */
361         if (avoid_reserve && free_huge_pages - resv_huge_pages == 0)
362                 return NULL;
363
364         for_each_zone_zonelist_nodemask(zone, z, zonelist,
365                                                 MAX_NR_ZONES - 1, nodemask) {
366                 nid = zone_to_nid(zone);
367                 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
368                     !list_empty(&hugepage_freelists[nid])) {
369                         page = list_entry(hugepage_freelists[nid].next,
370                                           struct page, lru);
371                         list_del(&page->lru);
372                         free_huge_pages--;
373                         free_huge_pages_node[nid]--;
374
375                         if (!avoid_reserve)
376                                 decrement_hugepage_resv_vma(vma);
377
378                         break;
379                 }
380         }
381         mpol_cond_put(mpol);
382         return page;
383 }
384
385 static void update_and_free_page(struct page *page)
386 {
387         int i;
388         nr_huge_pages--;
389         nr_huge_pages_node[page_to_nid(page)]--;
390         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
391                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
392                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
393                                 1 << PG_private | 1<< PG_writeback);
394         }
395         set_compound_page_dtor(page, NULL);
396         set_page_refcounted(page);
397         arch_release_hugepage(page);
398         __free_pages(page, HUGETLB_PAGE_ORDER);
399 }
400
401 static void free_huge_page(struct page *page)
402 {
403         int nid = page_to_nid(page);
404         struct address_space *mapping;
405
406         mapping = (struct address_space *) page_private(page);
407         set_page_private(page, 0);
408         BUG_ON(page_count(page));
409         INIT_LIST_HEAD(&page->lru);
410
411         spin_lock(&hugetlb_lock);
412         if (surplus_huge_pages_node[nid]) {
413                 update_and_free_page(page);
414                 surplus_huge_pages--;
415                 surplus_huge_pages_node[nid]--;
416         } else {
417                 enqueue_huge_page(page);
418         }
419         spin_unlock(&hugetlb_lock);
420         if (mapping)
421                 hugetlb_put_quota(mapping, 1);
422 }
423
424 /*
425  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
426  * balanced by operating on them in a round-robin fashion.
427  * Returns 1 if an adjustment was made.
428  */
429 static int adjust_pool_surplus(int delta)
430 {
431         static int prev_nid;
432         int nid = prev_nid;
433         int ret = 0;
434
435         VM_BUG_ON(delta != -1 && delta != 1);
436         do {
437                 nid = next_node(nid, node_online_map);
438                 if (nid == MAX_NUMNODES)
439                         nid = first_node(node_online_map);
440
441                 /* To shrink on this node, there must be a surplus page */
442                 if (delta < 0 && !surplus_huge_pages_node[nid])
443                         continue;
444                 /* Surplus cannot exceed the total number of pages */
445                 if (delta > 0 && surplus_huge_pages_node[nid] >=
446                                                 nr_huge_pages_node[nid])
447                         continue;
448
449                 surplus_huge_pages += delta;
450                 surplus_huge_pages_node[nid] += delta;
451                 ret = 1;
452                 break;
453         } while (nid != prev_nid);
454
455         prev_nid = nid;
456         return ret;
457 }
458
459 static struct page *alloc_fresh_huge_page_node(int nid)
460 {
461         struct page *page;
462
463         page = alloc_pages_node(nid,
464                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
465                                                 __GFP_REPEAT|__GFP_NOWARN,
466                 HUGETLB_PAGE_ORDER);
467         if (page) {
468                 if (arch_prepare_hugepage(page)) {
469                         __free_pages(page, HUGETLB_PAGE_ORDER);
470                         return NULL;
471                 }
472                 set_compound_page_dtor(page, free_huge_page);
473                 spin_lock(&hugetlb_lock);
474                 nr_huge_pages++;
475                 nr_huge_pages_node[nid]++;
476                 spin_unlock(&hugetlb_lock);
477                 put_page(page); /* free it into the hugepage allocator */
478         }
479
480         return page;
481 }
482
483 static int alloc_fresh_huge_page(void)
484 {
485         struct page *page;
486         int start_nid;
487         int next_nid;
488         int ret = 0;
489
490         start_nid = hugetlb_next_nid;
491
492         do {
493                 page = alloc_fresh_huge_page_node(hugetlb_next_nid);
494                 if (page)
495                         ret = 1;
496                 /*
497                  * Use a helper variable to find the next node and then
498                  * copy it back to hugetlb_next_nid afterwards:
499                  * otherwise there's a window in which a racer might
500                  * pass invalid nid MAX_NUMNODES to alloc_pages_node.
501                  * But we don't need to use a spin_lock here: it really
502                  * doesn't matter if occasionally a racer chooses the
503                  * same nid as we do.  Move nid forward in the mask even
504                  * if we just successfully allocated a hugepage so that
505                  * the next caller gets hugepages on the next node.
506                  */
507                 next_nid = next_node(hugetlb_next_nid, node_online_map);
508                 if (next_nid == MAX_NUMNODES)
509                         next_nid = first_node(node_online_map);
510                 hugetlb_next_nid = next_nid;
511         } while (!page && hugetlb_next_nid != start_nid);
512
513         if (ret)
514                 count_vm_event(HTLB_BUDDY_PGALLOC);
515         else
516                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
517
518         return ret;
519 }
520
521 static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
522                                                 unsigned long address)
523 {
524         struct page *page;
525         unsigned int nid;
526
527         /*
528          * Assume we will successfully allocate the surplus page to
529          * prevent racing processes from causing the surplus to exceed
530          * overcommit
531          *
532          * This however introduces a different race, where a process B
533          * tries to grow the static hugepage pool while alloc_pages() is
534          * called by process A. B will only examine the per-node
535          * counters in determining if surplus huge pages can be
536          * converted to normal huge pages in adjust_pool_surplus(). A
537          * won't be able to increment the per-node counter, until the
538          * lock is dropped by B, but B doesn't drop hugetlb_lock until
539          * no more huge pages can be converted from surplus to normal
540          * state (and doesn't try to convert again). Thus, we have a
541          * case where a surplus huge page exists, the pool is grown, and
542          * the surplus huge page still exists after, even though it
543          * should just have been converted to a normal huge page. This
544          * does not leak memory, though, as the hugepage will be freed
545          * once it is out of use. It also does not allow the counters to
546          * go out of whack in adjust_pool_surplus() as we don't modify
547          * the node values until we've gotten the hugepage and only the
548          * per-node value is checked there.
549          */
550         spin_lock(&hugetlb_lock);
551         if (surplus_huge_pages >= nr_overcommit_huge_pages) {
552                 spin_unlock(&hugetlb_lock);
553                 return NULL;
554         } else {
555                 nr_huge_pages++;
556                 surplus_huge_pages++;
557         }
558         spin_unlock(&hugetlb_lock);
559
560         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
561                                         __GFP_REPEAT|__GFP_NOWARN,
562                                         HUGETLB_PAGE_ORDER);
563
564         spin_lock(&hugetlb_lock);
565         if (page) {
566                 /*
567                  * This page is now managed by the hugetlb allocator and has
568                  * no users -- drop the buddy allocator's reference.
569                  */
570                 put_page_testzero(page);
571                 VM_BUG_ON(page_count(page));
572                 nid = page_to_nid(page);
573                 set_compound_page_dtor(page, free_huge_page);
574                 /*
575                  * We incremented the global counters already
576                  */
577                 nr_huge_pages_node[nid]++;
578                 surplus_huge_pages_node[nid]++;
579                 __count_vm_event(HTLB_BUDDY_PGALLOC);
580         } else {
581                 nr_huge_pages--;
582                 surplus_huge_pages--;
583                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
584         }
585         spin_unlock(&hugetlb_lock);
586
587         return page;
588 }
589
590 /*
591  * Increase the hugetlb pool such that it can accomodate a reservation
592  * of size 'delta'.
593  */
594 static int gather_surplus_pages(int delta)
595 {
596         struct list_head surplus_list;
597         struct page *page, *tmp;
598         int ret, i;
599         int needed, allocated;
600
601         needed = (resv_huge_pages + delta) - free_huge_pages;
602         if (needed <= 0) {
603                 resv_huge_pages += delta;
604                 return 0;
605         }
606
607         allocated = 0;
608         INIT_LIST_HEAD(&surplus_list);
609
610         ret = -ENOMEM;
611 retry:
612         spin_unlock(&hugetlb_lock);
613         for (i = 0; i < needed; i++) {
614                 page = alloc_buddy_huge_page(NULL, 0);
615                 if (!page) {
616                         /*
617                          * We were not able to allocate enough pages to
618                          * satisfy the entire reservation so we free what
619                          * we've allocated so far.
620                          */
621                         spin_lock(&hugetlb_lock);
622                         needed = 0;
623                         goto free;
624                 }
625
626                 list_add(&page->lru, &surplus_list);
627         }
628         allocated += needed;
629
630         /*
631          * After retaking hugetlb_lock, we need to recalculate 'needed'
632          * because either resv_huge_pages or free_huge_pages may have changed.
633          */
634         spin_lock(&hugetlb_lock);
635         needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
636         if (needed > 0)
637                 goto retry;
638
639         /*
640          * The surplus_list now contains _at_least_ the number of extra pages
641          * needed to accomodate the reservation.  Add the appropriate number
642          * of pages to the hugetlb pool and free the extras back to the buddy
643          * allocator.  Commit the entire reservation here to prevent another
644          * process from stealing the pages as they are added to the pool but
645          * before they are reserved.
646          */
647         needed += allocated;
648         resv_huge_pages += delta;
649         ret = 0;
650 free:
651         /* Free the needed pages to the hugetlb pool */
652         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
653                 if ((--needed) < 0)
654                         break;
655                 list_del(&page->lru);
656                 enqueue_huge_page(page);
657         }
658
659         /* Free unnecessary surplus pages to the buddy allocator */
660         if (!list_empty(&surplus_list)) {
661                 spin_unlock(&hugetlb_lock);
662                 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
663                         list_del(&page->lru);
664                         /*
665                          * The page has a reference count of zero already, so
666                          * call free_huge_page directly instead of using
667                          * put_page.  This must be done with hugetlb_lock
668                          * unlocked which is safe because free_huge_page takes
669                          * hugetlb_lock before deciding how to free the page.
670                          */
671                         free_huge_page(page);
672                 }
673                 spin_lock(&hugetlb_lock);
674         }
675
676         return ret;
677 }
678
679 /*
680  * When releasing a hugetlb pool reservation, any surplus pages that were
681  * allocated to satisfy the reservation must be explicitly freed if they were
682  * never used.
683  */
684 static void return_unused_surplus_pages(unsigned long unused_resv_pages)
685 {
686         static int nid = -1;
687         struct page *page;
688         unsigned long nr_pages;
689
690         /*
691          * We want to release as many surplus pages as possible, spread
692          * evenly across all nodes. Iterate across all nodes until we
693          * can no longer free unreserved surplus pages. This occurs when
694          * the nodes with surplus pages have no free pages.
695          */
696         unsigned long remaining_iterations = num_online_nodes();
697
698         /* Uncommit the reservation */
699         resv_huge_pages -= unused_resv_pages;
700
701         nr_pages = min(unused_resv_pages, surplus_huge_pages);
702
703         while (remaining_iterations-- && nr_pages) {
704                 nid = next_node(nid, node_online_map);
705                 if (nid == MAX_NUMNODES)
706                         nid = first_node(node_online_map);
707
708                 if (!surplus_huge_pages_node[nid])
709                         continue;
710
711                 if (!list_empty(&hugepage_freelists[nid])) {
712                         page = list_entry(hugepage_freelists[nid].next,
713                                           struct page, lru);
714                         list_del(&page->lru);
715                         update_and_free_page(page);
716                         free_huge_pages--;
717                         free_huge_pages_node[nid]--;
718                         surplus_huge_pages--;
719                         surplus_huge_pages_node[nid]--;
720                         nr_pages--;
721                         remaining_iterations = num_online_nodes();
722                 }
723         }
724 }
725
726 /*
727  * Determine if the huge page at addr within the vma has an associated
728  * reservation.  Where it does not we will need to logically increase
729  * reservation and actually increase quota before an allocation can occur.
730  * Where any new reservation would be required the reservation change is
731  * prepared, but not committed.  Once the page has been quota'd allocated
732  * an instantiated the change should be committed via vma_commit_reservation.
733  * No action is required on failure.
734  */
735 static int vma_needs_reservation(struct vm_area_struct *vma, unsigned long addr)
736 {
737         struct address_space *mapping = vma->vm_file->f_mapping;
738         struct inode *inode = mapping->host;
739
740         if (vma->vm_flags & VM_SHARED) {
741                 pgoff_t idx = vma_pagecache_offset(vma, addr);
742                 return region_chg(&inode->i_mapping->private_list,
743                                                         idx, idx + 1);
744
745         } else {
746                 if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER))
747                         return 1;
748         }
749
750         return 0;
751 }
752 static void vma_commit_reservation(struct vm_area_struct *vma,
753                                                         unsigned long addr)
754 {
755         struct address_space *mapping = vma->vm_file->f_mapping;
756         struct inode *inode = mapping->host;
757
758         if (vma->vm_flags & VM_SHARED) {
759                 pgoff_t idx = vma_pagecache_offset(vma, addr);
760                 region_add(&inode->i_mapping->private_list, idx, idx + 1);
761         }
762 }
763
764 static struct page *alloc_huge_page(struct vm_area_struct *vma,
765                                     unsigned long addr, int avoid_reserve)
766 {
767         struct page *page;
768         struct address_space *mapping = vma->vm_file->f_mapping;
769         struct inode *inode = mapping->host;
770         unsigned int chg;
771
772         /*
773          * Processes that did not create the mapping will have no reserves and
774          * will not have accounted against quota. Check that the quota can be
775          * made before satisfying the allocation
776          * MAP_NORESERVE mappings may also need pages and quota allocated
777          * if no reserve mapping overlaps.
778          */
779         chg = vma_needs_reservation(vma, addr);
780         if (chg < 0)
781                 return ERR_PTR(chg);
782         if (chg)
783                 if (hugetlb_get_quota(inode->i_mapping, chg))
784                         return ERR_PTR(-ENOSPC);
785
786         spin_lock(&hugetlb_lock);
787         page = dequeue_huge_page_vma(vma, addr, avoid_reserve);
788         spin_unlock(&hugetlb_lock);
789
790         if (!page) {
791                 page = alloc_buddy_huge_page(vma, addr);
792                 if (!page) {
793                         hugetlb_put_quota(inode->i_mapping, chg);
794                         return ERR_PTR(-VM_FAULT_OOM);
795                 }
796         }
797
798         set_page_refcounted(page);
799         set_page_private(page, (unsigned long) mapping);
800
801         vma_commit_reservation(vma, addr);
802
803         return page;
804 }
805
806 static int __init hugetlb_init(void)
807 {
808         unsigned long i;
809
810         if (HPAGE_SHIFT == 0)
811                 return 0;
812
813         for (i = 0; i < MAX_NUMNODES; ++i)
814                 INIT_LIST_HEAD(&hugepage_freelists[i]);
815
816         hugetlb_next_nid = first_node(node_online_map);
817
818         for (i = 0; i < max_huge_pages; ++i) {
819                 if (!alloc_fresh_huge_page())
820                         break;
821         }
822         max_huge_pages = free_huge_pages = nr_huge_pages = i;
823         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
824         return 0;
825 }
826 module_init(hugetlb_init);
827
828 static int __init hugetlb_setup(char *s)
829 {
830         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
831                 max_huge_pages = 0;
832         return 1;
833 }
834 __setup("hugepages=", hugetlb_setup);
835
836 static unsigned int cpuset_mems_nr(unsigned int *array)
837 {
838         int node;
839         unsigned int nr = 0;
840
841         for_each_node_mask(node, cpuset_current_mems_allowed)
842                 nr += array[node];
843
844         return nr;
845 }
846
847 #ifdef CONFIG_SYSCTL
848 #ifdef CONFIG_HIGHMEM
849 static void try_to_free_low(unsigned long count)
850 {
851         int i;
852
853         for (i = 0; i < MAX_NUMNODES; ++i) {
854                 struct page *page, *next;
855                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
856                         if (count >= nr_huge_pages)
857                                 return;
858                         if (PageHighMem(page))
859                                 continue;
860                         list_del(&page->lru);
861                         update_and_free_page(page);
862                         free_huge_pages--;
863                         free_huge_pages_node[page_to_nid(page)]--;
864                 }
865         }
866 }
867 #else
868 static inline void try_to_free_low(unsigned long count)
869 {
870 }
871 #endif
872
873 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
874 static unsigned long set_max_huge_pages(unsigned long count)
875 {
876         unsigned long min_count, ret;
877
878         /*
879          * Increase the pool size
880          * First take pages out of surplus state.  Then make up the
881          * remaining difference by allocating fresh huge pages.
882          *
883          * We might race with alloc_buddy_huge_page() here and be unable
884          * to convert a surplus huge page to a normal huge page. That is
885          * not critical, though, it just means the overall size of the
886          * pool might be one hugepage larger than it needs to be, but
887          * within all the constraints specified by the sysctls.
888          */
889         spin_lock(&hugetlb_lock);
890         while (surplus_huge_pages && count > persistent_huge_pages) {
891                 if (!adjust_pool_surplus(-1))
892                         break;
893         }
894
895         while (count > persistent_huge_pages) {
896                 /*
897                  * If this allocation races such that we no longer need the
898                  * page, free_huge_page will handle it by freeing the page
899                  * and reducing the surplus.
900                  */
901                 spin_unlock(&hugetlb_lock);
902                 ret = alloc_fresh_huge_page();
903                 spin_lock(&hugetlb_lock);
904                 if (!ret)
905                         goto out;
906
907         }
908
909         /*
910          * Decrease the pool size
911          * First return free pages to the buddy allocator (being careful
912          * to keep enough around to satisfy reservations).  Then place
913          * pages into surplus state as needed so the pool will shrink
914          * to the desired size as pages become free.
915          *
916          * By placing pages into the surplus state independent of the
917          * overcommit value, we are allowing the surplus pool size to
918          * exceed overcommit. There are few sane options here. Since
919          * alloc_buddy_huge_page() is checking the global counter,
920          * though, we'll note that we're not allowed to exceed surplus
921          * and won't grow the pool anywhere else. Not until one of the
922          * sysctls are changed, or the surplus pages go out of use.
923          */
924         min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
925         min_count = max(count, min_count);
926         try_to_free_low(min_count);
927         while (min_count < persistent_huge_pages) {
928                 struct page *page = dequeue_huge_page();
929                 if (!page)
930                         break;
931                 update_and_free_page(page);
932         }
933         while (count < persistent_huge_pages) {
934                 if (!adjust_pool_surplus(1))
935                         break;
936         }
937 out:
938         ret = persistent_huge_pages;
939         spin_unlock(&hugetlb_lock);
940         return ret;
941 }
942
943 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
944                            struct file *file, void __user *buffer,
945                            size_t *length, loff_t *ppos)
946 {
947         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
948         max_huge_pages = set_max_huge_pages(max_huge_pages);
949         return 0;
950 }
951
952 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
953                         struct file *file, void __user *buffer,
954                         size_t *length, loff_t *ppos)
955 {
956         proc_dointvec(table, write, file, buffer, length, ppos);
957         if (hugepages_treat_as_movable)
958                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
959         else
960                 htlb_alloc_mask = GFP_HIGHUSER;
961         return 0;
962 }
963
964 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
965                         struct file *file, void __user *buffer,
966                         size_t *length, loff_t *ppos)
967 {
968         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
969         spin_lock(&hugetlb_lock);
970         nr_overcommit_huge_pages = sysctl_overcommit_huge_pages;
971         spin_unlock(&hugetlb_lock);
972         return 0;
973 }
974
975 #endif /* CONFIG_SYSCTL */
976
977 int hugetlb_report_meminfo(char *buf)
978 {
979         return sprintf(buf,
980                         "HugePages_Total: %5lu\n"
981                         "HugePages_Free:  %5lu\n"
982                         "HugePages_Rsvd:  %5lu\n"
983                         "HugePages_Surp:  %5lu\n"
984                         "Hugepagesize:    %5lu kB\n",
985                         nr_huge_pages,
986                         free_huge_pages,
987                         resv_huge_pages,
988                         surplus_huge_pages,
989                         HPAGE_SIZE/1024);
990 }
991
992 int hugetlb_report_node_meminfo(int nid, char *buf)
993 {
994         return sprintf(buf,
995                 "Node %d HugePages_Total: %5u\n"
996                 "Node %d HugePages_Free:  %5u\n"
997                 "Node %d HugePages_Surp:  %5u\n",
998                 nid, nr_huge_pages_node[nid],
999                 nid, free_huge_pages_node[nid],
1000                 nid, surplus_huge_pages_node[nid]);
1001 }
1002
1003 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1004 unsigned long hugetlb_total_pages(void)
1005 {
1006         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
1007 }
1008
1009 static int hugetlb_acct_memory(long delta)
1010 {
1011         int ret = -ENOMEM;
1012
1013         spin_lock(&hugetlb_lock);
1014         /*
1015          * When cpuset is configured, it breaks the strict hugetlb page
1016          * reservation as the accounting is done on a global variable. Such
1017          * reservation is completely rubbish in the presence of cpuset because
1018          * the reservation is not checked against page availability for the
1019          * current cpuset. Application can still potentially OOM'ed by kernel
1020          * with lack of free htlb page in cpuset that the task is in.
1021          * Attempt to enforce strict accounting with cpuset is almost
1022          * impossible (or too ugly) because cpuset is too fluid that
1023          * task or memory node can be dynamically moved between cpusets.
1024          *
1025          * The change of semantics for shared hugetlb mapping with cpuset is
1026          * undesirable. However, in order to preserve some of the semantics,
1027          * we fall back to check against current free page availability as
1028          * a best attempt and hopefully to minimize the impact of changing
1029          * semantics that cpuset has.
1030          */
1031         if (delta > 0) {
1032                 if (gather_surplus_pages(delta) < 0)
1033                         goto out;
1034
1035                 if (delta > cpuset_mems_nr(free_huge_pages_node)) {
1036                         return_unused_surplus_pages(delta);
1037                         goto out;
1038                 }
1039         }
1040
1041         ret = 0;
1042         if (delta < 0)
1043                 return_unused_surplus_pages((unsigned long) -delta);
1044
1045 out:
1046         spin_unlock(&hugetlb_lock);
1047         return ret;
1048 }
1049
1050 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
1051 {
1052         unsigned long reserve = vma_resv_huge_pages(vma);
1053         if (reserve)
1054                 hugetlb_acct_memory(-reserve);
1055 }
1056
1057 /*
1058  * We cannot handle pagefaults against hugetlb pages at all.  They cause
1059  * handle_mm_fault() to try to instantiate regular-sized pages in the
1060  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
1061  * this far.
1062  */
1063 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1064 {
1065         BUG();
1066         return 0;
1067 }
1068
1069 struct vm_operations_struct hugetlb_vm_ops = {
1070         .fault = hugetlb_vm_op_fault,
1071         .close = hugetlb_vm_op_close,
1072 };
1073
1074 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
1075                                 int writable)
1076 {
1077         pte_t entry;
1078
1079         if (writable) {
1080                 entry =
1081                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1082         } else {
1083                 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
1084         }
1085         entry = pte_mkyoung(entry);
1086         entry = pte_mkhuge(entry);
1087
1088         return entry;
1089 }
1090
1091 static void set_huge_ptep_writable(struct vm_area_struct *vma,
1092                                    unsigned long address, pte_t *ptep)
1093 {
1094         pte_t entry;
1095
1096         entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
1097         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
1098                 update_mmu_cache(vma, address, entry);
1099         }
1100 }
1101
1102
1103 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
1104                             struct vm_area_struct *vma)
1105 {
1106         pte_t *src_pte, *dst_pte, entry;
1107         struct page *ptepage;
1108         unsigned long addr;
1109         int cow;
1110
1111         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
1112
1113         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
1114                 src_pte = huge_pte_offset(src, addr);
1115                 if (!src_pte)
1116                         continue;
1117                 dst_pte = huge_pte_alloc(dst, addr);
1118                 if (!dst_pte)
1119                         goto nomem;
1120
1121                 /* If the pagetables are shared don't copy or take references */
1122                 if (dst_pte == src_pte)
1123                         continue;
1124
1125                 spin_lock(&dst->page_table_lock);
1126                 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
1127                 if (!huge_pte_none(huge_ptep_get(src_pte))) {
1128                         if (cow)
1129                                 huge_ptep_set_wrprotect(src, addr, src_pte);
1130                         entry = huge_ptep_get(src_pte);
1131                         ptepage = pte_page(entry);
1132                         get_page(ptepage);
1133                         set_huge_pte_at(dst, addr, dst_pte, entry);
1134                 }
1135                 spin_unlock(&src->page_table_lock);
1136                 spin_unlock(&dst->page_table_lock);
1137         }
1138         return 0;
1139
1140 nomem:
1141         return -ENOMEM;
1142 }
1143
1144 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1145                             unsigned long end, struct page *ref_page)
1146 {
1147         struct mm_struct *mm = vma->vm_mm;
1148         unsigned long address;
1149         pte_t *ptep;
1150         pte_t pte;
1151         struct page *page;
1152         struct page *tmp;
1153         /*
1154          * A page gathering list, protected by per file i_mmap_lock. The
1155          * lock is used to avoid list corruption from multiple unmapping
1156          * of the same page since we are using page->lru.
1157          */
1158         LIST_HEAD(page_list);
1159
1160         WARN_ON(!is_vm_hugetlb_page(vma));
1161         BUG_ON(start & ~HPAGE_MASK);
1162         BUG_ON(end & ~HPAGE_MASK);
1163
1164         spin_lock(&mm->page_table_lock);
1165         for (address = start; address < end; address += HPAGE_SIZE) {
1166                 ptep = huge_pte_offset(mm, address);
1167                 if (!ptep)
1168                         continue;
1169
1170                 if (huge_pmd_unshare(mm, &address, ptep))
1171                         continue;
1172
1173                 /*
1174                  * If a reference page is supplied, it is because a specific
1175                  * page is being unmapped, not a range. Ensure the page we
1176                  * are about to unmap is the actual page of interest.
1177                  */
1178                 if (ref_page) {
1179                         pte = huge_ptep_get(ptep);
1180                         if (huge_pte_none(pte))
1181                                 continue;
1182                         page = pte_page(pte);
1183                         if (page != ref_page)
1184                                 continue;
1185
1186                         /*
1187                          * Mark the VMA as having unmapped its page so that
1188                          * future faults in this VMA will fail rather than
1189                          * looking like data was lost
1190                          */
1191                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
1192                 }
1193
1194                 pte = huge_ptep_get_and_clear(mm, address, ptep);
1195                 if (huge_pte_none(pte))
1196                         continue;
1197
1198                 page = pte_page(pte);
1199                 if (pte_dirty(pte))
1200                         set_page_dirty(page);
1201                 list_add(&page->lru, &page_list);
1202         }
1203         spin_unlock(&mm->page_table_lock);
1204         flush_tlb_range(vma, start, end);
1205         list_for_each_entry_safe(page, tmp, &page_list, lru) {
1206                 list_del(&page->lru);
1207                 put_page(page);
1208         }
1209 }
1210
1211 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1212                           unsigned long end, struct page *ref_page)
1213 {
1214         /*
1215          * It is undesirable to test vma->vm_file as it should be non-null
1216          * for valid hugetlb area. However, vm_file will be NULL in the error
1217          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
1218          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
1219          * to clean up. Since no pte has actually been setup, it is safe to
1220          * do nothing in this case.
1221          */
1222         if (vma->vm_file) {
1223                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1224                 __unmap_hugepage_range(vma, start, end, ref_page);
1225                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1226         }
1227 }
1228
1229 /*
1230  * This is called when the original mapper is failing to COW a MAP_PRIVATE
1231  * mappping it owns the reserve page for. The intention is to unmap the page
1232  * from other VMAs and let the children be SIGKILLed if they are faulting the
1233  * same region.
1234  */
1235 int unmap_ref_private(struct mm_struct *mm,
1236                                         struct vm_area_struct *vma,
1237                                         struct page *page,
1238                                         unsigned long address)
1239 {
1240         struct vm_area_struct *iter_vma;
1241         struct address_space *mapping;
1242         struct prio_tree_iter iter;
1243         pgoff_t pgoff;
1244
1245         /*
1246          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1247          * from page cache lookup which is in HPAGE_SIZE units.
1248          */
1249         address = address & huge_page_mask(hstate_vma(vma));
1250         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
1251                 + (vma->vm_pgoff >> PAGE_SHIFT);
1252         mapping = (struct address_space *)page_private(page);
1253
1254         vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1255                 /* Do not unmap the current VMA */
1256                 if (iter_vma == vma)
1257                         continue;
1258
1259                 /*
1260                  * Unmap the page from other VMAs without their own reserves.
1261                  * They get marked to be SIGKILLed if they fault in these
1262                  * areas. This is because a future no-page fault on this VMA
1263                  * could insert a zeroed page instead of the data existing
1264                  * from the time of fork. This would look like data corruption
1265                  */
1266                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
1267                         unmap_hugepage_range(iter_vma,
1268                                 address, address + HPAGE_SIZE,
1269                                 page);
1270         }
1271
1272         return 1;
1273 }
1274
1275 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
1276                         unsigned long address, pte_t *ptep, pte_t pte,
1277                         struct page *pagecache_page)
1278 {
1279         struct page *old_page, *new_page;
1280         int avoidcopy;
1281         int outside_reserve = 0;
1282
1283         old_page = pte_page(pte);
1284
1285 retry_avoidcopy:
1286         /* If no-one else is actually using this page, avoid the copy
1287          * and just make the page writable */
1288         avoidcopy = (page_count(old_page) == 1);
1289         if (avoidcopy) {
1290                 set_huge_ptep_writable(vma, address, ptep);
1291                 return 0;
1292         }
1293
1294         /*
1295          * If the process that created a MAP_PRIVATE mapping is about to
1296          * perform a COW due to a shared page count, attempt to satisfy
1297          * the allocation without using the existing reserves. The pagecache
1298          * page is used to determine if the reserve at this address was
1299          * consumed or not. If reserves were used, a partial faulted mapping
1300          * at the time of fork() could consume its reserves on COW instead
1301          * of the full address range.
1302          */
1303         if (!(vma->vm_flags & VM_SHARED) &&
1304                         is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
1305                         old_page != pagecache_page)
1306                 outside_reserve = 1;
1307
1308         page_cache_get(old_page);
1309         new_page = alloc_huge_page(vma, address, outside_reserve);
1310
1311         if (IS_ERR(new_page)) {
1312                 page_cache_release(old_page);
1313
1314                 /*
1315                  * If a process owning a MAP_PRIVATE mapping fails to COW,
1316                  * it is due to references held by a child and an insufficient
1317                  * huge page pool. To guarantee the original mappers
1318                  * reliability, unmap the page from child processes. The child
1319                  * may get SIGKILLed if it later faults.
1320                  */
1321                 if (outside_reserve) {
1322                         BUG_ON(huge_pte_none(pte));
1323                         if (unmap_ref_private(mm, vma, old_page, address)) {
1324                                 BUG_ON(page_count(old_page) != 1);
1325                                 BUG_ON(huge_pte_none(pte));
1326                                 goto retry_avoidcopy;
1327                         }
1328                         WARN_ON_ONCE(1);
1329                 }
1330
1331                 return -PTR_ERR(new_page);
1332         }
1333
1334         spin_unlock(&mm->page_table_lock);
1335         copy_huge_page(new_page, old_page, address, vma);
1336         __SetPageUptodate(new_page);
1337         spin_lock(&mm->page_table_lock);
1338
1339         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
1340         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
1341                 /* Break COW */
1342                 huge_ptep_clear_flush(vma, address, ptep);
1343                 set_huge_pte_at(mm, address, ptep,
1344                                 make_huge_pte(vma, new_page, 1));
1345                 /* Make the old page be freed below */
1346                 new_page = old_page;
1347         }
1348         page_cache_release(new_page);
1349         page_cache_release(old_page);
1350         return 0;
1351 }
1352
1353 /* Return the pagecache page at a given address within a VMA */
1354 static struct page *hugetlbfs_pagecache_page(struct vm_area_struct *vma,
1355                         unsigned long address)
1356 {
1357         struct address_space *mapping;
1358         pgoff_t idx;
1359
1360         mapping = vma->vm_file->f_mapping;
1361         idx = vma_pagecache_offset(vma, address);
1362
1363         return find_lock_page(mapping, idx);
1364 }
1365
1366 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1367                         unsigned long address, pte_t *ptep, int write_access)
1368 {
1369         int ret = VM_FAULT_SIGBUS;
1370         pgoff_t idx;
1371         unsigned long size;
1372         struct page *page;
1373         struct address_space *mapping;
1374         pte_t new_pte;
1375
1376         /*
1377          * Currently, we are forced to kill the process in the event the
1378          * original mapper has unmapped pages from the child due to a failed
1379          * COW. Warn that such a situation has occured as it may not be obvious
1380          */
1381         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
1382                 printk(KERN_WARNING
1383                         "PID %d killed due to inadequate hugepage pool\n",
1384                         current->pid);
1385                 return ret;
1386         }
1387
1388         mapping = vma->vm_file->f_mapping;
1389         idx = vma_pagecache_offset(vma, address);
1390
1391         /*
1392          * Use page lock to guard against racing truncation
1393          * before we get page_table_lock.
1394          */
1395 retry:
1396         page = find_lock_page(mapping, idx);
1397         if (!page) {
1398                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
1399                 if (idx >= size)
1400                         goto out;
1401                 page = alloc_huge_page(vma, address, 0);
1402                 if (IS_ERR(page)) {
1403                         ret = -PTR_ERR(page);
1404                         goto out;
1405                 }
1406                 clear_huge_page(page, address);
1407                 __SetPageUptodate(page);
1408
1409                 if (vma->vm_flags & VM_SHARED) {
1410                         int err;
1411                         struct inode *inode = mapping->host;
1412
1413                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
1414                         if (err) {
1415                                 put_page(page);
1416                                 if (err == -EEXIST)
1417                                         goto retry;
1418                                 goto out;
1419                         }
1420
1421                         spin_lock(&inode->i_lock);
1422                         inode->i_blocks += BLOCKS_PER_HUGEPAGE;
1423                         spin_unlock(&inode->i_lock);
1424                 } else
1425                         lock_page(page);
1426         }
1427
1428         spin_lock(&mm->page_table_lock);
1429         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
1430         if (idx >= size)
1431                 goto backout;
1432
1433         ret = 0;
1434         if (!huge_pte_none(huge_ptep_get(ptep)))
1435                 goto backout;
1436
1437         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
1438                                 && (vma->vm_flags & VM_SHARED)));
1439         set_huge_pte_at(mm, address, ptep, new_pte);
1440
1441         if (write_access && !(vma->vm_flags & VM_SHARED)) {
1442                 /* Optimization, do the COW without a second fault */
1443                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
1444         }
1445
1446         spin_unlock(&mm->page_table_lock);
1447         unlock_page(page);
1448 out:
1449         return ret;
1450
1451 backout:
1452         spin_unlock(&mm->page_table_lock);
1453         unlock_page(page);
1454         put_page(page);
1455         goto out;
1456 }
1457
1458 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
1459                         unsigned long address, int write_access)
1460 {
1461         pte_t *ptep;
1462         pte_t entry;
1463         int ret;
1464         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
1465
1466         ptep = huge_pte_alloc(mm, address);
1467         if (!ptep)
1468                 return VM_FAULT_OOM;
1469
1470         /*
1471          * Serialize hugepage allocation and instantiation, so that we don't
1472          * get spurious allocation failures if two CPUs race to instantiate
1473          * the same page in the page cache.
1474          */
1475         mutex_lock(&hugetlb_instantiation_mutex);
1476         entry = huge_ptep_get(ptep);
1477         if (huge_pte_none(entry)) {
1478                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
1479                 mutex_unlock(&hugetlb_instantiation_mutex);
1480                 return ret;
1481         }
1482
1483         ret = 0;
1484
1485         spin_lock(&mm->page_table_lock);
1486         /* Check for a racing update before calling hugetlb_cow */
1487         if (likely(pte_same(entry, huge_ptep_get(ptep))))
1488                 if (write_access && !pte_write(entry)) {
1489                         struct page *page;
1490                         page = hugetlbfs_pagecache_page(vma, address);
1491                         ret = hugetlb_cow(mm, vma, address, ptep, entry, page);
1492                         if (page) {
1493                                 unlock_page(page);
1494                                 put_page(page);
1495                         }
1496                 }
1497         spin_unlock(&mm->page_table_lock);
1498         mutex_unlock(&hugetlb_instantiation_mutex);
1499
1500         return ret;
1501 }
1502
1503 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
1504                         struct page **pages, struct vm_area_struct **vmas,
1505                         unsigned long *position, int *length, int i,
1506                         int write)
1507 {
1508         unsigned long pfn_offset;
1509         unsigned long vaddr = *position;
1510         int remainder = *length;
1511
1512         spin_lock(&mm->page_table_lock);
1513         while (vaddr < vma->vm_end && remainder) {
1514                 pte_t *pte;
1515                 struct page *page;
1516
1517                 /*
1518                  * Some archs (sparc64, sh*) have multiple pte_ts to
1519                  * each hugepage.  We have to make * sure we get the
1520                  * first, for the page indexing below to work.
1521                  */
1522                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
1523
1524                 if (!pte || huge_pte_none(huge_ptep_get(pte)) ||
1525                     (write && !pte_write(huge_ptep_get(pte)))) {
1526                         int ret;
1527
1528                         spin_unlock(&mm->page_table_lock);
1529                         ret = hugetlb_fault(mm, vma, vaddr, write);
1530                         spin_lock(&mm->page_table_lock);
1531                         if (!(ret & VM_FAULT_ERROR))
1532                                 continue;
1533
1534                         remainder = 0;
1535                         if (!i)
1536                                 i = -EFAULT;
1537                         break;
1538                 }
1539
1540                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
1541                 page = pte_page(huge_ptep_get(pte));
1542 same_page:
1543                 if (pages) {
1544                         get_page(page);
1545                         pages[i] = page + pfn_offset;
1546                 }
1547
1548                 if (vmas)
1549                         vmas[i] = vma;
1550
1551                 vaddr += PAGE_SIZE;
1552                 ++pfn_offset;
1553                 --remainder;
1554                 ++i;
1555                 if (vaddr < vma->vm_end && remainder &&
1556                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
1557                         /*
1558                          * We use pfn_offset to avoid touching the pageframes
1559                          * of this compound page.
1560                          */
1561                         goto same_page;
1562                 }
1563         }
1564         spin_unlock(&mm->page_table_lock);
1565         *length = remainder;
1566         *position = vaddr;
1567
1568         return i;
1569 }
1570
1571 void hugetlb_change_protection(struct vm_area_struct *vma,
1572                 unsigned long address, unsigned long end, pgprot_t newprot)
1573 {
1574         struct mm_struct *mm = vma->vm_mm;
1575         unsigned long start = address;
1576         pte_t *ptep;
1577         pte_t pte;
1578
1579         BUG_ON(address >= end);
1580         flush_cache_range(vma, address, end);
1581
1582         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1583         spin_lock(&mm->page_table_lock);
1584         for (; address < end; address += HPAGE_SIZE) {
1585                 ptep = huge_pte_offset(mm, address);
1586                 if (!ptep)
1587                         continue;
1588                 if (huge_pmd_unshare(mm, &address, ptep))
1589                         continue;
1590                 if (!huge_pte_none(huge_ptep_get(ptep))) {
1591                         pte = huge_ptep_get_and_clear(mm, address, ptep);
1592                         pte = pte_mkhuge(pte_modify(pte, newprot));
1593                         set_huge_pte_at(mm, address, ptep, pte);
1594                 }
1595         }
1596         spin_unlock(&mm->page_table_lock);
1597         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1598
1599         flush_tlb_range(vma, start, end);
1600 }
1601
1602 int hugetlb_reserve_pages(struct inode *inode,
1603                                         long from, long to,
1604                                         struct vm_area_struct *vma)
1605 {
1606         long ret, chg;
1607
1608         if (vma && vma->vm_flags & VM_NORESERVE)
1609                 return 0;
1610
1611         /*
1612          * Shared mappings base their reservation on the number of pages that
1613          * are already allocated on behalf of the file. Private mappings need
1614          * to reserve the full area even if read-only as mprotect() may be
1615          * called to make the mapping read-write. Assume !vma is a shm mapping
1616          */
1617         if (!vma || vma->vm_flags & VM_SHARED)
1618                 chg = region_chg(&inode->i_mapping->private_list, from, to);
1619         else {
1620                 chg = to - from;
1621                 set_vma_resv_huge_pages(vma, chg);
1622                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
1623         }
1624
1625         if (chg < 0)
1626                 return chg;
1627
1628         if (hugetlb_get_quota(inode->i_mapping, chg))
1629                 return -ENOSPC;
1630         ret = hugetlb_acct_memory(chg);
1631         if (ret < 0) {
1632                 hugetlb_put_quota(inode->i_mapping, chg);
1633                 return ret;
1634         }
1635         if (!vma || vma->vm_flags & VM_SHARED)
1636                 region_add(&inode->i_mapping->private_list, from, to);
1637         return 0;
1638 }
1639
1640 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1641 {
1642         long chg = region_truncate(&inode->i_mapping->private_list, offset);
1643
1644         spin_lock(&inode->i_lock);
1645         inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed;
1646         spin_unlock(&inode->i_lock);
1647
1648         hugetlb_put_quota(inode->i_mapping, (chg - freed));
1649         hugetlb_acct_memory(-(chg - freed));
1650 }