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