[PATCH] Enable mprotect on huge pages
[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
17 #include <asm/page.h>
18 #include <asm/pgtable.h>
19
20 #include <linux/hugetlb.h>
21 #include "internal.h"
22
23 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
24 static unsigned long nr_huge_pages, free_huge_pages;
25 unsigned long max_huge_pages;
26 static struct list_head hugepage_freelists[MAX_NUMNODES];
27 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
28 static unsigned int free_huge_pages_node[MAX_NUMNODES];
29
30 /*
31  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
32  */
33 static DEFINE_SPINLOCK(hugetlb_lock);
34
35 static void enqueue_huge_page(struct page *page)
36 {
37         int nid = page_to_nid(page);
38         list_add(&page->lru, &hugepage_freelists[nid]);
39         free_huge_pages++;
40         free_huge_pages_node[nid]++;
41 }
42
43 static struct page *dequeue_huge_page(struct vm_area_struct *vma,
44                                 unsigned long address)
45 {
46         int nid = numa_node_id();
47         struct page *page = NULL;
48         struct zonelist *zonelist = huge_zonelist(vma, address);
49         struct zone **z;
50
51         for (z = zonelist->zones; *z; z++) {
52                 nid = (*z)->zone_pgdat->node_id;
53                 if (cpuset_zone_allowed(*z, GFP_HIGHUSER) &&
54                     !list_empty(&hugepage_freelists[nid]))
55                         break;
56         }
57
58         if (*z) {
59                 page = list_entry(hugepage_freelists[nid].next,
60                                   struct page, lru);
61                 list_del(&page->lru);
62                 free_huge_pages--;
63                 free_huge_pages_node[nid]--;
64         }
65         return page;
66 }
67
68 static int alloc_fresh_huge_page(void)
69 {
70         static int nid = 0;
71         struct page *page;
72         page = alloc_pages_node(nid, GFP_HIGHUSER|__GFP_COMP|__GFP_NOWARN,
73                                         HUGETLB_PAGE_ORDER);
74         nid = (nid + 1) % num_online_nodes();
75         if (page) {
76                 page[1].lru.next = (void *)free_huge_page;      /* dtor */
77                 spin_lock(&hugetlb_lock);
78                 nr_huge_pages++;
79                 nr_huge_pages_node[page_to_nid(page)]++;
80                 spin_unlock(&hugetlb_lock);
81                 put_page(page); /* free it into the hugepage allocator */
82                 return 1;
83         }
84         return 0;
85 }
86
87 void free_huge_page(struct page *page)
88 {
89         BUG_ON(page_count(page));
90
91         INIT_LIST_HEAD(&page->lru);
92
93         spin_lock(&hugetlb_lock);
94         enqueue_huge_page(page);
95         spin_unlock(&hugetlb_lock);
96 }
97
98 struct page *alloc_huge_page(struct vm_area_struct *vma, unsigned long addr)
99 {
100         struct page *page;
101         int i;
102
103         spin_lock(&hugetlb_lock);
104         page = dequeue_huge_page(vma, addr);
105         if (!page) {
106                 spin_unlock(&hugetlb_lock);
107                 return NULL;
108         }
109         spin_unlock(&hugetlb_lock);
110         set_page_refcounted(page);
111         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); ++i)
112                 clear_user_highpage(&page[i], addr);
113         return page;
114 }
115
116 static int __init hugetlb_init(void)
117 {
118         unsigned long i;
119
120         if (HPAGE_SHIFT == 0)
121                 return 0;
122
123         for (i = 0; i < MAX_NUMNODES; ++i)
124                 INIT_LIST_HEAD(&hugepage_freelists[i]);
125
126         for (i = 0; i < max_huge_pages; ++i) {
127                 if (!alloc_fresh_huge_page())
128                         break;
129         }
130         max_huge_pages = free_huge_pages = nr_huge_pages = i;
131         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
132         return 0;
133 }
134 module_init(hugetlb_init);
135
136 static int __init hugetlb_setup(char *s)
137 {
138         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
139                 max_huge_pages = 0;
140         return 1;
141 }
142 __setup("hugepages=", hugetlb_setup);
143
144 #ifdef CONFIG_SYSCTL
145 static void update_and_free_page(struct page *page)
146 {
147         int i;
148         nr_huge_pages--;
149         nr_huge_pages_node[page_zone(page)->zone_pgdat->node_id]--;
150         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
151                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
152                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
153                                 1 << PG_private | 1<< PG_writeback);
154         }
155         page[1].lru.next = NULL;
156         set_page_refcounted(page);
157         __free_pages(page, HUGETLB_PAGE_ORDER);
158 }
159
160 #ifdef CONFIG_HIGHMEM
161 static void try_to_free_low(unsigned long count)
162 {
163         int i, nid;
164         for (i = 0; i < MAX_NUMNODES; ++i) {
165                 struct page *page, *next;
166                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
167                         if (PageHighMem(page))
168                                 continue;
169                         list_del(&page->lru);
170                         update_and_free_page(page);
171                         nid = page_zone(page)->zone_pgdat->node_id;
172                         free_huge_pages--;
173                         free_huge_pages_node[nid]--;
174                         if (count >= nr_huge_pages)
175                                 return;
176                 }
177         }
178 }
179 #else
180 static inline void try_to_free_low(unsigned long count)
181 {
182 }
183 #endif
184
185 static unsigned long set_max_huge_pages(unsigned long count)
186 {
187         while (count > nr_huge_pages) {
188                 if (!alloc_fresh_huge_page())
189                         return nr_huge_pages;
190         }
191         if (count >= nr_huge_pages)
192                 return nr_huge_pages;
193
194         spin_lock(&hugetlb_lock);
195         try_to_free_low(count);
196         while (count < nr_huge_pages) {
197                 struct page *page = dequeue_huge_page(NULL, 0);
198                 if (!page)
199                         break;
200                 update_and_free_page(page);
201         }
202         spin_unlock(&hugetlb_lock);
203         return nr_huge_pages;
204 }
205
206 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
207                            struct file *file, void __user *buffer,
208                            size_t *length, loff_t *ppos)
209 {
210         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
211         max_huge_pages = set_max_huge_pages(max_huge_pages);
212         return 0;
213 }
214 #endif /* CONFIG_SYSCTL */
215
216 int hugetlb_report_meminfo(char *buf)
217 {
218         return sprintf(buf,
219                         "HugePages_Total: %5lu\n"
220                         "HugePages_Free:  %5lu\n"
221                         "Hugepagesize:    %5lu kB\n",
222                         nr_huge_pages,
223                         free_huge_pages,
224                         HPAGE_SIZE/1024);
225 }
226
227 int hugetlb_report_node_meminfo(int nid, char *buf)
228 {
229         return sprintf(buf,
230                 "Node %d HugePages_Total: %5u\n"
231                 "Node %d HugePages_Free:  %5u\n",
232                 nid, nr_huge_pages_node[nid],
233                 nid, free_huge_pages_node[nid]);
234 }
235
236 int is_hugepage_mem_enough(size_t size)
237 {
238         return (size + ~HPAGE_MASK)/HPAGE_SIZE <= free_huge_pages;
239 }
240
241 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
242 unsigned long hugetlb_total_pages(void)
243 {
244         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
245 }
246
247 /*
248  * We cannot handle pagefaults against hugetlb pages at all.  They cause
249  * handle_mm_fault() to try to instantiate regular-sized pages in the
250  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
251  * this far.
252  */
253 static struct page *hugetlb_nopage(struct vm_area_struct *vma,
254                                 unsigned long address, int *unused)
255 {
256         BUG();
257         return NULL;
258 }
259
260 struct vm_operations_struct hugetlb_vm_ops = {
261         .nopage = hugetlb_nopage,
262 };
263
264 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
265                                 int writable)
266 {
267         pte_t entry;
268
269         if (writable) {
270                 entry =
271                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
272         } else {
273                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
274         }
275         entry = pte_mkyoung(entry);
276         entry = pte_mkhuge(entry);
277
278         return entry;
279 }
280
281 static void set_huge_ptep_writable(struct vm_area_struct *vma,
282                                    unsigned long address, pte_t *ptep)
283 {
284         pte_t entry;
285
286         entry = pte_mkwrite(pte_mkdirty(*ptep));
287         ptep_set_access_flags(vma, address, ptep, entry, 1);
288         update_mmu_cache(vma, address, entry);
289         lazy_mmu_prot_update(entry);
290 }
291
292
293 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
294                             struct vm_area_struct *vma)
295 {
296         pte_t *src_pte, *dst_pte, entry;
297         struct page *ptepage;
298         unsigned long addr;
299         int cow;
300
301         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
302
303         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
304                 src_pte = huge_pte_offset(src, addr);
305                 if (!src_pte)
306                         continue;
307                 dst_pte = huge_pte_alloc(dst, addr);
308                 if (!dst_pte)
309                         goto nomem;
310                 spin_lock(&dst->page_table_lock);
311                 spin_lock(&src->page_table_lock);
312                 if (!pte_none(*src_pte)) {
313                         if (cow)
314                                 ptep_set_wrprotect(src, addr, src_pte);
315                         entry = *src_pte;
316                         ptepage = pte_page(entry);
317                         get_page(ptepage);
318                         add_mm_counter(dst, file_rss, HPAGE_SIZE / PAGE_SIZE);
319                         set_huge_pte_at(dst, addr, dst_pte, entry);
320                 }
321                 spin_unlock(&src->page_table_lock);
322                 spin_unlock(&dst->page_table_lock);
323         }
324         return 0;
325
326 nomem:
327         return -ENOMEM;
328 }
329
330 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
331                           unsigned long end)
332 {
333         struct mm_struct *mm = vma->vm_mm;
334         unsigned long address;
335         pte_t *ptep;
336         pte_t pte;
337         struct page *page;
338
339         WARN_ON(!is_vm_hugetlb_page(vma));
340         BUG_ON(start & ~HPAGE_MASK);
341         BUG_ON(end & ~HPAGE_MASK);
342
343         spin_lock(&mm->page_table_lock);
344
345         /* Update high watermark before we lower rss */
346         update_hiwater_rss(mm);
347
348         for (address = start; address < end; address += HPAGE_SIZE) {
349                 ptep = huge_pte_offset(mm, address);
350                 if (!ptep)
351                         continue;
352
353                 pte = huge_ptep_get_and_clear(mm, address, ptep);
354                 if (pte_none(pte))
355                         continue;
356
357                 page = pte_page(pte);
358                 put_page(page);
359                 add_mm_counter(mm, file_rss, (int) -(HPAGE_SIZE / PAGE_SIZE));
360         }
361
362         spin_unlock(&mm->page_table_lock);
363         flush_tlb_range(vma, start, end);
364 }
365
366 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
367                         unsigned long address, pte_t *ptep, pte_t pte)
368 {
369         struct page *old_page, *new_page;
370         int i, avoidcopy;
371
372         old_page = pte_page(pte);
373
374         /* If no-one else is actually using this page, avoid the copy
375          * and just make the page writable */
376         avoidcopy = (page_count(old_page) == 1);
377         if (avoidcopy) {
378                 set_huge_ptep_writable(vma, address, ptep);
379                 return VM_FAULT_MINOR;
380         }
381
382         page_cache_get(old_page);
383         new_page = alloc_huge_page(vma, address);
384
385         if (!new_page) {
386                 page_cache_release(old_page);
387                 return VM_FAULT_OOM;
388         }
389
390         spin_unlock(&mm->page_table_lock);
391         for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++)
392                 copy_user_highpage(new_page + i, old_page + i,
393                                    address + i*PAGE_SIZE);
394         spin_lock(&mm->page_table_lock);
395
396         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
397         if (likely(pte_same(*ptep, pte))) {
398                 /* Break COW */
399                 set_huge_pte_at(mm, address, ptep,
400                                 make_huge_pte(vma, new_page, 1));
401                 /* Make the old page be freed below */
402                 new_page = old_page;
403         }
404         page_cache_release(new_page);
405         page_cache_release(old_page);
406         return VM_FAULT_MINOR;
407 }
408
409 int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
410                         unsigned long address, pte_t *ptep, int write_access)
411 {
412         int ret = VM_FAULT_SIGBUS;
413         unsigned long idx;
414         unsigned long size;
415         struct page *page;
416         struct address_space *mapping;
417         pte_t new_pte;
418
419         mapping = vma->vm_file->f_mapping;
420         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
421                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
422
423         /*
424          * Use page lock to guard against racing truncation
425          * before we get page_table_lock.
426          */
427 retry:
428         page = find_lock_page(mapping, idx);
429         if (!page) {
430                 if (hugetlb_get_quota(mapping))
431                         goto out;
432                 page = alloc_huge_page(vma, address);
433                 if (!page) {
434                         hugetlb_put_quota(mapping);
435                         ret = VM_FAULT_OOM;
436                         goto out;
437                 }
438
439                 if (vma->vm_flags & VM_SHARED) {
440                         int err;
441
442                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
443                         if (err) {
444                                 put_page(page);
445                                 hugetlb_put_quota(mapping);
446                                 if (err == -EEXIST)
447                                         goto retry;
448                                 goto out;
449                         }
450                 } else
451                         lock_page(page);
452         }
453
454         spin_lock(&mm->page_table_lock);
455         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
456         if (idx >= size)
457                 goto backout;
458
459         ret = VM_FAULT_MINOR;
460         if (!pte_none(*ptep))
461                 goto backout;
462
463         add_mm_counter(mm, file_rss, HPAGE_SIZE / PAGE_SIZE);
464         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
465                                 && (vma->vm_flags & VM_SHARED)));
466         set_huge_pte_at(mm, address, ptep, new_pte);
467
468         if (write_access && !(vma->vm_flags & VM_SHARED)) {
469                 /* Optimization, do the COW without a second fault */
470                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
471         }
472
473         spin_unlock(&mm->page_table_lock);
474         unlock_page(page);
475 out:
476         return ret;
477
478 backout:
479         spin_unlock(&mm->page_table_lock);
480         hugetlb_put_quota(mapping);
481         unlock_page(page);
482         put_page(page);
483         goto out;
484 }
485
486 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
487                         unsigned long address, int write_access)
488 {
489         pte_t *ptep;
490         pte_t entry;
491         int ret;
492
493         ptep = huge_pte_alloc(mm, address);
494         if (!ptep)
495                 return VM_FAULT_OOM;
496
497         entry = *ptep;
498         if (pte_none(entry))
499                 return hugetlb_no_page(mm, vma, address, ptep, write_access);
500
501         ret = VM_FAULT_MINOR;
502
503         spin_lock(&mm->page_table_lock);
504         /* Check for a racing update before calling hugetlb_cow */
505         if (likely(pte_same(entry, *ptep)))
506                 if (write_access && !pte_write(entry))
507                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
508         spin_unlock(&mm->page_table_lock);
509
510         return ret;
511 }
512
513 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
514                         struct page **pages, struct vm_area_struct **vmas,
515                         unsigned long *position, int *length, int i)
516 {
517         unsigned long vpfn, vaddr = *position;
518         int remainder = *length;
519
520         vpfn = vaddr/PAGE_SIZE;
521         spin_lock(&mm->page_table_lock);
522         while (vaddr < vma->vm_end && remainder) {
523                 pte_t *pte;
524                 struct page *page;
525
526                 /*
527                  * Some archs (sparc64, sh*) have multiple pte_ts to
528                  * each hugepage.  We have to make * sure we get the
529                  * first, for the page indexing below to work.
530                  */
531                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
532
533                 if (!pte || pte_none(*pte)) {
534                         int ret;
535
536                         spin_unlock(&mm->page_table_lock);
537                         ret = hugetlb_fault(mm, vma, vaddr, 0);
538                         spin_lock(&mm->page_table_lock);
539                         if (ret == VM_FAULT_MINOR)
540                                 continue;
541
542                         remainder = 0;
543                         if (!i)
544                                 i = -EFAULT;
545                         break;
546                 }
547
548                 if (pages) {
549                         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
550                         get_page(page);
551                         pages[i] = page;
552                 }
553
554                 if (vmas)
555                         vmas[i] = vma;
556
557                 vaddr += PAGE_SIZE;
558                 ++vpfn;
559                 --remainder;
560                 ++i;
561         }
562         spin_unlock(&mm->page_table_lock);
563         *length = remainder;
564         *position = vaddr;
565
566         return i;
567 }
568
569 void hugetlb_change_protection(struct vm_area_struct *vma,
570                 unsigned long address, unsigned long end, pgprot_t newprot)
571 {
572         struct mm_struct *mm = vma->vm_mm;
573         unsigned long start = address;
574         pte_t *ptep;
575         pte_t pte;
576
577         BUG_ON(address >= end);
578         flush_cache_range(vma, address, end);
579
580         spin_lock(&mm->page_table_lock);
581         for (; address < end; address += HPAGE_SIZE) {
582                 ptep = huge_pte_offset(mm, address);
583                 if (!ptep)
584                         continue;
585                 if (!pte_none(*ptep)) {
586                         pte = huge_ptep_get_and_clear(mm, address, ptep);
587                         pte = pte_mkhuge(pte_modify(pte, newprot));
588                         set_huge_pte_at(mm, address, ptep, pte);
589                         lazy_mmu_prot_update(pte);
590                 }
591         }
592         spin_unlock(&mm->page_table_lock);
593
594         flush_tlb_range(vma, start, end);
595 }
596