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