mm: filter based on a nodemask as well as a gfp_mask
[linux-2.6.git] / mm / hugetlb.c
1 /*
2  * Generic hugetlb support.
3  * (C) William Irwin, April 2004
4  */
5 #include <linux/gfp.h>
6 #include <linux/list.h>
7 #include <linux/init.h>
8 #include <linux/module.h>
9 #include <linux/mm.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/nodemask.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/cpuset.h>
16 #include <linux/mutex.h>
17
18 #include <asm/page.h>
19 #include <asm/pgtable.h>
20
21 #include <linux/hugetlb.h>
22 #include "internal.h"
23
24 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
25 static unsigned long nr_huge_pages, free_huge_pages, resv_huge_pages;
26 static unsigned long surplus_huge_pages;
27 static unsigned long nr_overcommit_huge_pages;
28 unsigned long max_huge_pages;
29 unsigned long sysctl_overcommit_huge_pages;
30 static struct list_head hugepage_freelists[MAX_NUMNODES];
31 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
32 static unsigned int free_huge_pages_node[MAX_NUMNODES];
33 static unsigned int surplus_huge_pages_node[MAX_NUMNODES];
34 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
35 unsigned long hugepages_treat_as_movable;
36 static int hugetlb_next_nid;
37
38 /*
39  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
40  */
41 static DEFINE_SPINLOCK(hugetlb_lock);
42
43 static void clear_huge_page(struct page *page, unsigned long addr)
44 {
45         int i;
46
47         might_sleep();
48         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
49                 cond_resched();
50                 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
51         }
52 }
53
54 static void copy_huge_page(struct page *dst, struct page *src,
55                            unsigned long addr, struct vm_area_struct *vma)
56 {
57         int i;
58
59         might_sleep();
60         for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
61                 cond_resched();
62                 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
63         }
64 }
65
66 static void enqueue_huge_page(struct page *page)
67 {
68         int nid = page_to_nid(page);
69         list_add(&page->lru, &hugepage_freelists[nid]);
70         free_huge_pages++;
71         free_huge_pages_node[nid]++;
72 }
73
74 static struct page *dequeue_huge_page(void)
75 {
76         int nid;
77         struct page *page = NULL;
78
79         for (nid = 0; nid < MAX_NUMNODES; ++nid) {
80                 if (!list_empty(&hugepage_freelists[nid])) {
81                         page = list_entry(hugepage_freelists[nid].next,
82                                           struct page, lru);
83                         list_del(&page->lru);
84                         free_huge_pages--;
85                         free_huge_pages_node[nid]--;
86                         break;
87                 }
88         }
89         return page;
90 }
91
92 static struct page *dequeue_huge_page_vma(struct vm_area_struct *vma,
93                                 unsigned long address)
94 {
95         int nid;
96         struct page *page = NULL;
97         struct mempolicy *mpol;
98         nodemask_t *nodemask;
99         struct zonelist *zonelist = huge_zonelist(vma, address,
100                                         htlb_alloc_mask, &mpol, &nodemask);
101         struct zone *zone;
102         struct zoneref *z;
103
104         for_each_zone_zonelist_nodemask(zone, z, zonelist,
105                                                 MAX_NR_ZONES - 1, nodemask) {
106                 nid = zone_to_nid(zone);
107                 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
108                     !list_empty(&hugepage_freelists[nid])) {
109                         page = list_entry(hugepage_freelists[nid].next,
110                                           struct page, lru);
111                         list_del(&page->lru);
112                         free_huge_pages--;
113                         free_huge_pages_node[nid]--;
114                         if (vma && vma->vm_flags & VM_MAYSHARE)
115                                 resv_huge_pages--;
116                         break;
117                 }
118         }
119         mpol_free(mpol);        /* unref if mpol !NULL */
120         return page;
121 }
122
123 static void update_and_free_page(struct page *page)
124 {
125         int i;
126         nr_huge_pages--;
127         nr_huge_pages_node[page_to_nid(page)]--;
128         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
129                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
130                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
131                                 1 << PG_private | 1<< PG_writeback);
132         }
133         set_compound_page_dtor(page, NULL);
134         set_page_refcounted(page);
135         __free_pages(page, HUGETLB_PAGE_ORDER);
136 }
137
138 static void free_huge_page(struct page *page)
139 {
140         int nid = page_to_nid(page);
141         struct address_space *mapping;
142
143         mapping = (struct address_space *) page_private(page);
144         set_page_private(page, 0);
145         BUG_ON(page_count(page));
146         INIT_LIST_HEAD(&page->lru);
147
148         spin_lock(&hugetlb_lock);
149         if (surplus_huge_pages_node[nid]) {
150                 update_and_free_page(page);
151                 surplus_huge_pages--;
152                 surplus_huge_pages_node[nid]--;
153         } else {
154                 enqueue_huge_page(page);
155         }
156         spin_unlock(&hugetlb_lock);
157         if (mapping)
158                 hugetlb_put_quota(mapping, 1);
159 }
160
161 /*
162  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
163  * balanced by operating on them in a round-robin fashion.
164  * Returns 1 if an adjustment was made.
165  */
166 static int adjust_pool_surplus(int delta)
167 {
168         static int prev_nid;
169         int nid = prev_nid;
170         int ret = 0;
171
172         VM_BUG_ON(delta != -1 && delta != 1);
173         do {
174                 nid = next_node(nid, node_online_map);
175                 if (nid == MAX_NUMNODES)
176                         nid = first_node(node_online_map);
177
178                 /* To shrink on this node, there must be a surplus page */
179                 if (delta < 0 && !surplus_huge_pages_node[nid])
180                         continue;
181                 /* Surplus cannot exceed the total number of pages */
182                 if (delta > 0 && surplus_huge_pages_node[nid] >=
183                                                 nr_huge_pages_node[nid])
184                         continue;
185
186                 surplus_huge_pages += delta;
187                 surplus_huge_pages_node[nid] += delta;
188                 ret = 1;
189                 break;
190         } while (nid != prev_nid);
191
192         prev_nid = nid;
193         return ret;
194 }
195
196 static struct page *alloc_fresh_huge_page_node(int nid)
197 {
198         struct page *page;
199
200         page = alloc_pages_node(nid,
201                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|__GFP_NOWARN,
202                 HUGETLB_PAGE_ORDER);
203         if (page) {
204                 set_compound_page_dtor(page, free_huge_page);
205                 spin_lock(&hugetlb_lock);
206                 nr_huge_pages++;
207                 nr_huge_pages_node[nid]++;
208                 spin_unlock(&hugetlb_lock);
209                 put_page(page); /* free it into the hugepage allocator */
210         }
211
212         return page;
213 }
214
215 static int alloc_fresh_huge_page(void)
216 {
217         struct page *page;
218         int start_nid;
219         int next_nid;
220         int ret = 0;
221
222         start_nid = hugetlb_next_nid;
223
224         do {
225                 page = alloc_fresh_huge_page_node(hugetlb_next_nid);
226                 if (page)
227                         ret = 1;
228                 /*
229                  * Use a helper variable to find the next node and then
230                  * copy it back to hugetlb_next_nid afterwards:
231                  * otherwise there's a window in which a racer might
232                  * pass invalid nid MAX_NUMNODES to alloc_pages_node.
233                  * But we don't need to use a spin_lock here: it really
234                  * doesn't matter if occasionally a racer chooses the
235                  * same nid as we do.  Move nid forward in the mask even
236                  * if we just successfully allocated a hugepage so that
237                  * the next caller gets hugepages on the next node.
238                  */
239                 next_nid = next_node(hugetlb_next_nid, node_online_map);
240                 if (next_nid == MAX_NUMNODES)
241                         next_nid = first_node(node_online_map);
242                 hugetlb_next_nid = next_nid;
243         } while (!page && hugetlb_next_nid != start_nid);
244
245         return ret;
246 }
247
248 static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
249                                                 unsigned long address)
250 {
251         struct page *page;
252         unsigned int nid;
253
254         /*
255          * Assume we will successfully allocate the surplus page to
256          * prevent racing processes from causing the surplus to exceed
257          * overcommit
258          *
259          * This however introduces a different race, where a process B
260          * tries to grow the static hugepage pool while alloc_pages() is
261          * called by process A. B will only examine the per-node
262          * counters in determining if surplus huge pages can be
263          * converted to normal huge pages in adjust_pool_surplus(). A
264          * won't be able to increment the per-node counter, until the
265          * lock is dropped by B, but B doesn't drop hugetlb_lock until
266          * no more huge pages can be converted from surplus to normal
267          * state (and doesn't try to convert again). Thus, we have a
268          * case where a surplus huge page exists, the pool is grown, and
269          * the surplus huge page still exists after, even though it
270          * should just have been converted to a normal huge page. This
271          * does not leak memory, though, as the hugepage will be freed
272          * once it is out of use. It also does not allow the counters to
273          * go out of whack in adjust_pool_surplus() as we don't modify
274          * the node values until we've gotten the hugepage and only the
275          * per-node value is checked there.
276          */
277         spin_lock(&hugetlb_lock);
278         if (surplus_huge_pages >= nr_overcommit_huge_pages) {
279                 spin_unlock(&hugetlb_lock);
280                 return NULL;
281         } else {
282                 nr_huge_pages++;
283                 surplus_huge_pages++;
284         }
285         spin_unlock(&hugetlb_lock);
286
287         page = alloc_pages(htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
288                                         HUGETLB_PAGE_ORDER);
289
290         spin_lock(&hugetlb_lock);
291         if (page) {
292                 /*
293                  * This page is now managed by the hugetlb allocator and has
294                  * no users -- drop the buddy allocator's reference.
295                  */
296                 put_page_testzero(page);
297                 VM_BUG_ON(page_count(page));
298                 nid = page_to_nid(page);
299                 set_compound_page_dtor(page, free_huge_page);
300                 /*
301                  * We incremented the global counters already
302                  */
303                 nr_huge_pages_node[nid]++;
304                 surplus_huge_pages_node[nid]++;
305         } else {
306                 nr_huge_pages--;
307                 surplus_huge_pages--;
308         }
309         spin_unlock(&hugetlb_lock);
310
311         return page;
312 }
313
314 /*
315  * Increase the hugetlb pool such that it can accomodate a reservation
316  * of size 'delta'.
317  */
318 static int gather_surplus_pages(int delta)
319 {
320         struct list_head surplus_list;
321         struct page *page, *tmp;
322         int ret, i;
323         int needed, allocated;
324
325         needed = (resv_huge_pages + delta) - free_huge_pages;
326         if (needed <= 0) {
327                 resv_huge_pages += delta;
328                 return 0;
329         }
330
331         allocated = 0;
332         INIT_LIST_HEAD(&surplus_list);
333
334         ret = -ENOMEM;
335 retry:
336         spin_unlock(&hugetlb_lock);
337         for (i = 0; i < needed; i++) {
338                 page = alloc_buddy_huge_page(NULL, 0);
339                 if (!page) {
340                         /*
341                          * We were not able to allocate enough pages to
342                          * satisfy the entire reservation so we free what
343                          * we've allocated so far.
344                          */
345                         spin_lock(&hugetlb_lock);
346                         needed = 0;
347                         goto free;
348                 }
349
350                 list_add(&page->lru, &surplus_list);
351         }
352         allocated += needed;
353
354         /*
355          * After retaking hugetlb_lock, we need to recalculate 'needed'
356          * because either resv_huge_pages or free_huge_pages may have changed.
357          */
358         spin_lock(&hugetlb_lock);
359         needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
360         if (needed > 0)
361                 goto retry;
362
363         /*
364          * The surplus_list now contains _at_least_ the number of extra pages
365          * needed to accomodate the reservation.  Add the appropriate number
366          * of pages to the hugetlb pool and free the extras back to the buddy
367          * allocator.  Commit the entire reservation here to prevent another
368          * process from stealing the pages as they are added to the pool but
369          * before they are reserved.
370          */
371         needed += allocated;
372         resv_huge_pages += delta;
373         ret = 0;
374 free:
375         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
376                 list_del(&page->lru);
377                 if ((--needed) >= 0)
378                         enqueue_huge_page(page);
379                 else {
380                         /*
381                          * The page has a reference count of zero already, so
382                          * call free_huge_page directly instead of using
383                          * put_page.  This must be done with hugetlb_lock
384                          * unlocked which is safe because free_huge_page takes
385                          * hugetlb_lock before deciding how to free the page.
386                          */
387                         spin_unlock(&hugetlb_lock);
388                         free_huge_page(page);
389                         spin_lock(&hugetlb_lock);
390                 }
391         }
392
393         return ret;
394 }
395
396 /*
397  * When releasing a hugetlb pool reservation, any surplus pages that were
398  * allocated to satisfy the reservation must be explicitly freed if they were
399  * never used.
400  */
401 static void return_unused_surplus_pages(unsigned long unused_resv_pages)
402 {
403         static int nid = -1;
404         struct page *page;
405         unsigned long nr_pages;
406
407         /*
408          * We want to release as many surplus pages as possible, spread
409          * evenly across all nodes. Iterate across all nodes until we
410          * can no longer free unreserved surplus pages. This occurs when
411          * the nodes with surplus pages have no free pages.
412          */
413         unsigned long remaining_iterations = num_online_nodes();
414
415         /* Uncommit the reservation */
416         resv_huge_pages -= unused_resv_pages;
417
418         nr_pages = min(unused_resv_pages, surplus_huge_pages);
419
420         while (remaining_iterations-- && nr_pages) {
421                 nid = next_node(nid, node_online_map);
422                 if (nid == MAX_NUMNODES)
423                         nid = first_node(node_online_map);
424
425                 if (!surplus_huge_pages_node[nid])
426                         continue;
427
428                 if (!list_empty(&hugepage_freelists[nid])) {
429                         page = list_entry(hugepage_freelists[nid].next,
430                                           struct page, lru);
431                         list_del(&page->lru);
432                         update_and_free_page(page);
433                         free_huge_pages--;
434                         free_huge_pages_node[nid]--;
435                         surplus_huge_pages--;
436                         surplus_huge_pages_node[nid]--;
437                         nr_pages--;
438                         remaining_iterations = num_online_nodes();
439                 }
440         }
441 }
442
443
444 static struct page *alloc_huge_page_shared(struct vm_area_struct *vma,
445                                                 unsigned long addr)
446 {
447         struct page *page;
448
449         spin_lock(&hugetlb_lock);
450         page = dequeue_huge_page_vma(vma, addr);
451         spin_unlock(&hugetlb_lock);
452         return page ? page : ERR_PTR(-VM_FAULT_OOM);
453 }
454
455 static struct page *alloc_huge_page_private(struct vm_area_struct *vma,
456                                                 unsigned long addr)
457 {
458         struct page *page = NULL;
459
460         if (hugetlb_get_quota(vma->vm_file->f_mapping, 1))
461                 return ERR_PTR(-VM_FAULT_SIGBUS);
462
463         spin_lock(&hugetlb_lock);
464         if (free_huge_pages > resv_huge_pages)
465                 page = dequeue_huge_page_vma(vma, addr);
466         spin_unlock(&hugetlb_lock);
467         if (!page) {
468                 page = alloc_buddy_huge_page(vma, addr);
469                 if (!page) {
470                         hugetlb_put_quota(vma->vm_file->f_mapping, 1);
471                         return ERR_PTR(-VM_FAULT_OOM);
472                 }
473         }
474         return page;
475 }
476
477 static struct page *alloc_huge_page(struct vm_area_struct *vma,
478                                     unsigned long addr)
479 {
480         struct page *page;
481         struct address_space *mapping = vma->vm_file->f_mapping;
482
483         if (vma->vm_flags & VM_MAYSHARE)
484                 page = alloc_huge_page_shared(vma, addr);
485         else
486                 page = alloc_huge_page_private(vma, addr);
487
488         if (!IS_ERR(page)) {
489                 set_page_refcounted(page);
490                 set_page_private(page, (unsigned long) mapping);
491         }
492         return page;
493 }
494
495 static int __init hugetlb_init(void)
496 {
497         unsigned long i;
498
499         if (HPAGE_SHIFT == 0)
500                 return 0;
501
502         for (i = 0; i < MAX_NUMNODES; ++i)
503                 INIT_LIST_HEAD(&hugepage_freelists[i]);
504
505         hugetlb_next_nid = first_node(node_online_map);
506
507         for (i = 0; i < max_huge_pages; ++i) {
508                 if (!alloc_fresh_huge_page())
509                         break;
510         }
511         max_huge_pages = free_huge_pages = nr_huge_pages = i;
512         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
513         return 0;
514 }
515 module_init(hugetlb_init);
516
517 static int __init hugetlb_setup(char *s)
518 {
519         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
520                 max_huge_pages = 0;
521         return 1;
522 }
523 __setup("hugepages=", hugetlb_setup);
524
525 static unsigned int cpuset_mems_nr(unsigned int *array)
526 {
527         int node;
528         unsigned int nr = 0;
529
530         for_each_node_mask(node, cpuset_current_mems_allowed)
531                 nr += array[node];
532
533         return nr;
534 }
535
536 #ifdef CONFIG_SYSCTL
537 #ifdef CONFIG_HIGHMEM
538 static void try_to_free_low(unsigned long count)
539 {
540         int i;
541
542         for (i = 0; i < MAX_NUMNODES; ++i) {
543                 struct page *page, *next;
544                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
545                         if (count >= nr_huge_pages)
546                                 return;
547                         if (PageHighMem(page))
548                                 continue;
549                         list_del(&page->lru);
550                         update_and_free_page(page);
551                         free_huge_pages--;
552                         free_huge_pages_node[page_to_nid(page)]--;
553                 }
554         }
555 }
556 #else
557 static inline void try_to_free_low(unsigned long count)
558 {
559 }
560 #endif
561
562 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
563 static unsigned long set_max_huge_pages(unsigned long count)
564 {
565         unsigned long min_count, ret;
566
567         /*
568          * Increase the pool size
569          * First take pages out of surplus state.  Then make up the
570          * remaining difference by allocating fresh huge pages.
571          *
572          * We might race with alloc_buddy_huge_page() here and be unable
573          * to convert a surplus huge page to a normal huge page. That is
574          * not critical, though, it just means the overall size of the
575          * pool might be one hugepage larger than it needs to be, but
576          * within all the constraints specified by the sysctls.
577          */
578         spin_lock(&hugetlb_lock);
579         while (surplus_huge_pages && count > persistent_huge_pages) {
580                 if (!adjust_pool_surplus(-1))
581                         break;
582         }
583
584         while (count > persistent_huge_pages) {
585                 int ret;
586                 /*
587                  * If this allocation races such that we no longer need the
588                  * page, free_huge_page will handle it by freeing the page
589                  * and reducing the surplus.
590                  */
591                 spin_unlock(&hugetlb_lock);
592                 ret = alloc_fresh_huge_page();
593                 spin_lock(&hugetlb_lock);
594                 if (!ret)
595                         goto out;
596
597         }
598
599         /*
600          * Decrease the pool size
601          * First return free pages to the buddy allocator (being careful
602          * to keep enough around to satisfy reservations).  Then place
603          * pages into surplus state as needed so the pool will shrink
604          * to the desired size as pages become free.
605          *
606          * By placing pages into the surplus state independent of the
607          * overcommit value, we are allowing the surplus pool size to
608          * exceed overcommit. There are few sane options here. Since
609          * alloc_buddy_huge_page() is checking the global counter,
610          * though, we'll note that we're not allowed to exceed surplus
611          * and won't grow the pool anywhere else. Not until one of the
612          * sysctls are changed, or the surplus pages go out of use.
613          */
614         min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
615         min_count = max(count, min_count);
616         try_to_free_low(min_count);
617         while (min_count < persistent_huge_pages) {
618                 struct page *page = dequeue_huge_page();
619                 if (!page)
620                         break;
621                 update_and_free_page(page);
622         }
623         while (count < persistent_huge_pages) {
624                 if (!adjust_pool_surplus(1))
625                         break;
626         }
627 out:
628         ret = persistent_huge_pages;
629         spin_unlock(&hugetlb_lock);
630         return ret;
631 }
632
633 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
634                            struct file *file, void __user *buffer,
635                            size_t *length, loff_t *ppos)
636 {
637         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
638         max_huge_pages = set_max_huge_pages(max_huge_pages);
639         return 0;
640 }
641
642 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
643                         struct file *file, void __user *buffer,
644                         size_t *length, loff_t *ppos)
645 {
646         proc_dointvec(table, write, file, buffer, length, ppos);
647         if (hugepages_treat_as_movable)
648                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
649         else
650                 htlb_alloc_mask = GFP_HIGHUSER;
651         return 0;
652 }
653
654 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
655                         struct file *file, void __user *buffer,
656                         size_t *length, loff_t *ppos)
657 {
658         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
659         spin_lock(&hugetlb_lock);
660         nr_overcommit_huge_pages = sysctl_overcommit_huge_pages;
661         spin_unlock(&hugetlb_lock);
662         return 0;
663 }
664
665 #endif /* CONFIG_SYSCTL */
666
667 int hugetlb_report_meminfo(char *buf)
668 {
669         return sprintf(buf,
670                         "HugePages_Total: %5lu\n"
671                         "HugePages_Free:  %5lu\n"
672                         "HugePages_Rsvd:  %5lu\n"
673                         "HugePages_Surp:  %5lu\n"
674                         "Hugepagesize:    %5lu kB\n",
675                         nr_huge_pages,
676                         free_huge_pages,
677                         resv_huge_pages,
678                         surplus_huge_pages,
679                         HPAGE_SIZE/1024);
680 }
681
682 int hugetlb_report_node_meminfo(int nid, char *buf)
683 {
684         return sprintf(buf,
685                 "Node %d HugePages_Total: %5u\n"
686                 "Node %d HugePages_Free:  %5u\n"
687                 "Node %d HugePages_Surp:  %5u\n",
688                 nid, nr_huge_pages_node[nid],
689                 nid, free_huge_pages_node[nid],
690                 nid, surplus_huge_pages_node[nid]);
691 }
692
693 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
694 unsigned long hugetlb_total_pages(void)
695 {
696         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
697 }
698
699 /*
700  * We cannot handle pagefaults against hugetlb pages at all.  They cause
701  * handle_mm_fault() to try to instantiate regular-sized pages in the
702  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
703  * this far.
704  */
705 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
706 {
707         BUG();
708         return 0;
709 }
710
711 struct vm_operations_struct hugetlb_vm_ops = {
712         .fault = hugetlb_vm_op_fault,
713 };
714
715 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
716                                 int writable)
717 {
718         pte_t entry;
719
720         if (writable) {
721                 entry =
722                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
723         } else {
724                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
725         }
726         entry = pte_mkyoung(entry);
727         entry = pte_mkhuge(entry);
728
729         return entry;
730 }
731
732 static void set_huge_ptep_writable(struct vm_area_struct *vma,
733                                    unsigned long address, pte_t *ptep)
734 {
735         pte_t entry;
736
737         entry = pte_mkwrite(pte_mkdirty(*ptep));
738         if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
739                 update_mmu_cache(vma, address, entry);
740         }
741 }
742
743
744 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
745                             struct vm_area_struct *vma)
746 {
747         pte_t *src_pte, *dst_pte, entry;
748         struct page *ptepage;
749         unsigned long addr;
750         int cow;
751
752         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
753
754         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
755                 src_pte = huge_pte_offset(src, addr);
756                 if (!src_pte)
757                         continue;
758                 dst_pte = huge_pte_alloc(dst, addr);
759                 if (!dst_pte)
760                         goto nomem;
761
762                 /* If the pagetables are shared don't copy or take references */
763                 if (dst_pte == src_pte)
764                         continue;
765
766                 spin_lock(&dst->page_table_lock);
767                 spin_lock(&src->page_table_lock);
768                 if (!pte_none(*src_pte)) {
769                         if (cow)
770                                 ptep_set_wrprotect(src, addr, src_pte);
771                         entry = *src_pte;
772                         ptepage = pte_page(entry);
773                         get_page(ptepage);
774                         set_huge_pte_at(dst, addr, dst_pte, entry);
775                 }
776                 spin_unlock(&src->page_table_lock);
777                 spin_unlock(&dst->page_table_lock);
778         }
779         return 0;
780
781 nomem:
782         return -ENOMEM;
783 }
784
785 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
786                             unsigned long end)
787 {
788         struct mm_struct *mm = vma->vm_mm;
789         unsigned long address;
790         pte_t *ptep;
791         pte_t pte;
792         struct page *page;
793         struct page *tmp;
794         /*
795          * A page gathering list, protected by per file i_mmap_lock. The
796          * lock is used to avoid list corruption from multiple unmapping
797          * of the same page since we are using page->lru.
798          */
799         LIST_HEAD(page_list);
800
801         WARN_ON(!is_vm_hugetlb_page(vma));
802         BUG_ON(start & ~HPAGE_MASK);
803         BUG_ON(end & ~HPAGE_MASK);
804
805         spin_lock(&mm->page_table_lock);
806         for (address = start; address < end; address += HPAGE_SIZE) {
807                 ptep = huge_pte_offset(mm, address);
808                 if (!ptep)
809                         continue;
810
811                 if (huge_pmd_unshare(mm, &address, ptep))
812                         continue;
813
814                 pte = huge_ptep_get_and_clear(mm, address, ptep);
815                 if (pte_none(pte))
816                         continue;
817
818                 page = pte_page(pte);
819                 if (pte_dirty(pte))
820                         set_page_dirty(page);
821                 list_add(&page->lru, &page_list);
822         }
823         spin_unlock(&mm->page_table_lock);
824         flush_tlb_range(vma, start, end);
825         list_for_each_entry_safe(page, tmp, &page_list, lru) {
826                 list_del(&page->lru);
827                 put_page(page);
828         }
829 }
830
831 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
832                           unsigned long end)
833 {
834         /*
835          * It is undesirable to test vma->vm_file as it should be non-null
836          * for valid hugetlb area. However, vm_file will be NULL in the error
837          * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
838          * do_mmap_pgoff() nullifies vma->vm_file before calling this function
839          * to clean up. Since no pte has actually been setup, it is safe to
840          * do nothing in this case.
841          */
842         if (vma->vm_file) {
843                 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
844                 __unmap_hugepage_range(vma, start, end);
845                 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
846         }
847 }
848
849 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
850                         unsigned long address, pte_t *ptep, pte_t pte)
851 {
852         struct page *old_page, *new_page;
853         int avoidcopy;
854
855         old_page = pte_page(pte);
856
857         /* If no-one else is actually using this page, avoid the copy
858          * and just make the page writable */
859         avoidcopy = (page_count(old_page) == 1);
860         if (avoidcopy) {
861                 set_huge_ptep_writable(vma, address, ptep);
862                 return 0;
863         }
864
865         page_cache_get(old_page);
866         new_page = alloc_huge_page(vma, address);
867
868         if (IS_ERR(new_page)) {
869                 page_cache_release(old_page);
870                 return -PTR_ERR(new_page);
871         }
872
873         spin_unlock(&mm->page_table_lock);
874         copy_huge_page(new_page, old_page, address, vma);
875         __SetPageUptodate(new_page);
876         spin_lock(&mm->page_table_lock);
877
878         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
879         if (likely(pte_same(*ptep, pte))) {
880                 /* Break COW */
881                 set_huge_pte_at(mm, address, ptep,
882                                 make_huge_pte(vma, new_page, 1));
883                 /* Make the old page be freed below */
884                 new_page = old_page;
885         }
886         page_cache_release(new_page);
887         page_cache_release(old_page);
888         return 0;
889 }
890
891 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
892                         unsigned long address, pte_t *ptep, int write_access)
893 {
894         int ret = VM_FAULT_SIGBUS;
895         unsigned long idx;
896         unsigned long size;
897         struct page *page;
898         struct address_space *mapping;
899         pte_t new_pte;
900
901         mapping = vma->vm_file->f_mapping;
902         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
903                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
904
905         /*
906          * Use page lock to guard against racing truncation
907          * before we get page_table_lock.
908          */
909 retry:
910         page = find_lock_page(mapping, idx);
911         if (!page) {
912                 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
913                 if (idx >= size)
914                         goto out;
915                 page = alloc_huge_page(vma, address);
916                 if (IS_ERR(page)) {
917                         ret = -PTR_ERR(page);
918                         goto out;
919                 }
920                 clear_huge_page(page, address);
921                 __SetPageUptodate(page);
922
923                 if (vma->vm_flags & VM_SHARED) {
924                         int err;
925                         struct inode *inode = mapping->host;
926
927                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
928                         if (err) {
929                                 put_page(page);
930                                 if (err == -EEXIST)
931                                         goto retry;
932                                 goto out;
933                         }
934
935                         spin_lock(&inode->i_lock);
936                         inode->i_blocks += BLOCKS_PER_HUGEPAGE;
937                         spin_unlock(&inode->i_lock);
938                 } else
939                         lock_page(page);
940         }
941
942         spin_lock(&mm->page_table_lock);
943         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
944         if (idx >= size)
945                 goto backout;
946
947         ret = 0;
948         if (!pte_none(*ptep))
949                 goto backout;
950
951         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
952                                 && (vma->vm_flags & VM_SHARED)));
953         set_huge_pte_at(mm, address, ptep, new_pte);
954
955         if (write_access && !(vma->vm_flags & VM_SHARED)) {
956                 /* Optimization, do the COW without a second fault */
957                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
958         }
959
960         spin_unlock(&mm->page_table_lock);
961         unlock_page(page);
962 out:
963         return ret;
964
965 backout:
966         spin_unlock(&mm->page_table_lock);
967         unlock_page(page);
968         put_page(page);
969         goto out;
970 }
971
972 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
973                         unsigned long address, int write_access)
974 {
975         pte_t *ptep;
976         pte_t entry;
977         int ret;
978         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
979
980         ptep = huge_pte_alloc(mm, address);
981         if (!ptep)
982                 return VM_FAULT_OOM;
983
984         /*
985          * Serialize hugepage allocation and instantiation, so that we don't
986          * get spurious allocation failures if two CPUs race to instantiate
987          * the same page in the page cache.
988          */
989         mutex_lock(&hugetlb_instantiation_mutex);
990         entry = *ptep;
991         if (pte_none(entry)) {
992                 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
993                 mutex_unlock(&hugetlb_instantiation_mutex);
994                 return ret;
995         }
996
997         ret = 0;
998
999         spin_lock(&mm->page_table_lock);
1000         /* Check for a racing update before calling hugetlb_cow */
1001         if (likely(pte_same(entry, *ptep)))
1002                 if (write_access && !pte_write(entry))
1003                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
1004         spin_unlock(&mm->page_table_lock);
1005         mutex_unlock(&hugetlb_instantiation_mutex);
1006
1007         return ret;
1008 }
1009
1010 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
1011                         struct page **pages, struct vm_area_struct **vmas,
1012                         unsigned long *position, int *length, int i,
1013                         int write)
1014 {
1015         unsigned long pfn_offset;
1016         unsigned long vaddr = *position;
1017         int remainder = *length;
1018
1019         spin_lock(&mm->page_table_lock);
1020         while (vaddr < vma->vm_end && remainder) {
1021                 pte_t *pte;
1022                 struct page *page;
1023
1024                 /*
1025                  * Some archs (sparc64, sh*) have multiple pte_ts to
1026                  * each hugepage.  We have to make * sure we get the
1027                  * first, for the page indexing below to work.
1028                  */
1029                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
1030
1031                 if (!pte || pte_none(*pte) || (write && !pte_write(*pte))) {
1032                         int ret;
1033
1034                         spin_unlock(&mm->page_table_lock);
1035                         ret = hugetlb_fault(mm, vma, vaddr, write);
1036                         spin_lock(&mm->page_table_lock);
1037                         if (!(ret & VM_FAULT_ERROR))
1038                                 continue;
1039
1040                         remainder = 0;
1041                         if (!i)
1042                                 i = -EFAULT;
1043                         break;
1044                 }
1045
1046                 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
1047                 page = pte_page(*pte);
1048 same_page:
1049                 if (pages) {
1050                         get_page(page);
1051                         pages[i] = page + pfn_offset;
1052                 }
1053
1054                 if (vmas)
1055                         vmas[i] = vma;
1056
1057                 vaddr += PAGE_SIZE;
1058                 ++pfn_offset;
1059                 --remainder;
1060                 ++i;
1061                 if (vaddr < vma->vm_end && remainder &&
1062                                 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
1063                         /*
1064                          * We use pfn_offset to avoid touching the pageframes
1065                          * of this compound page.
1066                          */
1067                         goto same_page;
1068                 }
1069         }
1070         spin_unlock(&mm->page_table_lock);
1071         *length = remainder;
1072         *position = vaddr;
1073
1074         return i;
1075 }
1076
1077 void hugetlb_change_protection(struct vm_area_struct *vma,
1078                 unsigned long address, unsigned long end, pgprot_t newprot)
1079 {
1080         struct mm_struct *mm = vma->vm_mm;
1081         unsigned long start = address;
1082         pte_t *ptep;
1083         pte_t pte;
1084
1085         BUG_ON(address >= end);
1086         flush_cache_range(vma, address, end);
1087
1088         spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1089         spin_lock(&mm->page_table_lock);
1090         for (; address < end; address += HPAGE_SIZE) {
1091                 ptep = huge_pte_offset(mm, address);
1092                 if (!ptep)
1093                         continue;
1094                 if (huge_pmd_unshare(mm, &address, ptep))
1095                         continue;
1096                 if (!pte_none(*ptep)) {
1097                         pte = huge_ptep_get_and_clear(mm, address, ptep);
1098                         pte = pte_mkhuge(pte_modify(pte, newprot));
1099                         set_huge_pte_at(mm, address, ptep, pte);
1100                 }
1101         }
1102         spin_unlock(&mm->page_table_lock);
1103         spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1104
1105         flush_tlb_range(vma, start, end);
1106 }
1107
1108 struct file_region {
1109         struct list_head link;
1110         long from;
1111         long to;
1112 };
1113
1114 static long region_add(struct list_head *head, long f, long t)
1115 {
1116         struct file_region *rg, *nrg, *trg;
1117
1118         /* Locate the region we are either in or before. */
1119         list_for_each_entry(rg, head, link)
1120                 if (f <= rg->to)
1121                         break;
1122
1123         /* Round our left edge to the current segment if it encloses us. */
1124         if (f > rg->from)
1125                 f = rg->from;
1126
1127         /* Check for and consume any regions we now overlap with. */
1128         nrg = rg;
1129         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1130                 if (&rg->link == head)
1131                         break;
1132                 if (rg->from > t)
1133                         break;
1134
1135                 /* If this area reaches higher then extend our area to
1136                  * include it completely.  If this is not the first area
1137                  * which we intend to reuse, free it. */
1138                 if (rg->to > t)
1139                         t = rg->to;
1140                 if (rg != nrg) {
1141                         list_del(&rg->link);
1142                         kfree(rg);
1143                 }
1144         }
1145         nrg->from = f;
1146         nrg->to = t;
1147         return 0;
1148 }
1149
1150 static long region_chg(struct list_head *head, long f, long t)
1151 {
1152         struct file_region *rg, *nrg;
1153         long chg = 0;
1154
1155         /* Locate the region we are before or in. */
1156         list_for_each_entry(rg, head, link)
1157                 if (f <= rg->to)
1158                         break;
1159
1160         /* If we are below the current region then a new region is required.
1161          * Subtle, allocate a new region at the position but make it zero
1162          * size such that we can guarantee to record the reservation. */
1163         if (&rg->link == head || t < rg->from) {
1164                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
1165                 if (!nrg)
1166                         return -ENOMEM;
1167                 nrg->from = f;
1168                 nrg->to   = f;
1169                 INIT_LIST_HEAD(&nrg->link);
1170                 list_add(&nrg->link, rg->link.prev);
1171
1172                 return t - f;
1173         }
1174
1175         /* Round our left edge to the current segment if it encloses us. */
1176         if (f > rg->from)
1177                 f = rg->from;
1178         chg = t - f;
1179
1180         /* Check for and consume any regions we now overlap with. */
1181         list_for_each_entry(rg, rg->link.prev, link) {
1182                 if (&rg->link == head)
1183                         break;
1184                 if (rg->from > t)
1185                         return chg;
1186
1187                 /* We overlap with this area, if it extends futher than
1188                  * us then we must extend ourselves.  Account for its
1189                  * existing reservation. */
1190                 if (rg->to > t) {
1191                         chg += rg->to - t;
1192                         t = rg->to;
1193                 }
1194                 chg -= rg->to - rg->from;
1195         }
1196         return chg;
1197 }
1198
1199 static long region_truncate(struct list_head *head, long end)
1200 {
1201         struct file_region *rg, *trg;
1202         long chg = 0;
1203
1204         /* Locate the region we are either in or before. */
1205         list_for_each_entry(rg, head, link)
1206                 if (end <= rg->to)
1207                         break;
1208         if (&rg->link == head)
1209                 return 0;
1210
1211         /* If we are in the middle of a region then adjust it. */
1212         if (end > rg->from) {
1213                 chg = rg->to - end;
1214                 rg->to = end;
1215                 rg = list_entry(rg->link.next, typeof(*rg), link);
1216         }
1217
1218         /* Drop any remaining regions. */
1219         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
1220                 if (&rg->link == head)
1221                         break;
1222                 chg += rg->to - rg->from;
1223                 list_del(&rg->link);
1224                 kfree(rg);
1225         }
1226         return chg;
1227 }
1228
1229 static int hugetlb_acct_memory(long delta)
1230 {
1231         int ret = -ENOMEM;
1232
1233         spin_lock(&hugetlb_lock);
1234         /*
1235          * When cpuset is configured, it breaks the strict hugetlb page
1236          * reservation as the accounting is done on a global variable. Such
1237          * reservation is completely rubbish in the presence of cpuset because
1238          * the reservation is not checked against page availability for the
1239          * current cpuset. Application can still potentially OOM'ed by kernel
1240          * with lack of free htlb page in cpuset that the task is in.
1241          * Attempt to enforce strict accounting with cpuset is almost
1242          * impossible (or too ugly) because cpuset is too fluid that
1243          * task or memory node can be dynamically moved between cpusets.
1244          *
1245          * The change of semantics for shared hugetlb mapping with cpuset is
1246          * undesirable. However, in order to preserve some of the semantics,
1247          * we fall back to check against current free page availability as
1248          * a best attempt and hopefully to minimize the impact of changing
1249          * semantics that cpuset has.
1250          */
1251         if (delta > 0) {
1252                 if (gather_surplus_pages(delta) < 0)
1253                         goto out;
1254
1255                 if (delta > cpuset_mems_nr(free_huge_pages_node)) {
1256                         return_unused_surplus_pages(delta);
1257                         goto out;
1258                 }
1259         }
1260
1261         ret = 0;
1262         if (delta < 0)
1263                 return_unused_surplus_pages((unsigned long) -delta);
1264
1265 out:
1266         spin_unlock(&hugetlb_lock);
1267         return ret;
1268 }
1269
1270 int hugetlb_reserve_pages(struct inode *inode, long from, long to)
1271 {
1272         long ret, chg;
1273
1274         chg = region_chg(&inode->i_mapping->private_list, from, to);
1275         if (chg < 0)
1276                 return chg;
1277
1278         if (hugetlb_get_quota(inode->i_mapping, chg))
1279                 return -ENOSPC;
1280         ret = hugetlb_acct_memory(chg);
1281         if (ret < 0) {
1282                 hugetlb_put_quota(inode->i_mapping, chg);
1283                 return ret;
1284         }
1285         region_add(&inode->i_mapping->private_list, from, to);
1286         return 0;
1287 }
1288
1289 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1290 {
1291         long chg = region_truncate(&inode->i_mapping->private_list, offset);
1292
1293         spin_lock(&inode->i_lock);
1294         inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed;
1295         spin_unlock(&inode->i_lock);
1296
1297         hugetlb_put_quota(inode->i_mapping, (chg - freed));
1298         hugetlb_acct_memory(-(chg - freed));
1299 }