[PATCH] Hugetlb: Copy on Write support
[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 <asm/page.h>
15 #include <asm/pgtable.h>
16
17 #include <linux/hugetlb.h>
18
19 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
20 static unsigned long nr_huge_pages, free_huge_pages;
21 unsigned long max_huge_pages;
22 static struct list_head hugepage_freelists[MAX_NUMNODES];
23 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
24 static unsigned int free_huge_pages_node[MAX_NUMNODES];
25
26 /*
27  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
28  */
29 static DEFINE_SPINLOCK(hugetlb_lock);
30
31 static void enqueue_huge_page(struct page *page)
32 {
33         int nid = page_to_nid(page);
34         list_add(&page->lru, &hugepage_freelists[nid]);
35         free_huge_pages++;
36         free_huge_pages_node[nid]++;
37 }
38
39 static struct page *dequeue_huge_page(void)
40 {
41         int nid = numa_node_id();
42         struct page *page = NULL;
43
44         if (list_empty(&hugepage_freelists[nid])) {
45                 for (nid = 0; nid < MAX_NUMNODES; ++nid)
46                         if (!list_empty(&hugepage_freelists[nid]))
47                                 break;
48         }
49         if (nid >= 0 && nid < MAX_NUMNODES &&
50             !list_empty(&hugepage_freelists[nid])) {
51                 page = list_entry(hugepage_freelists[nid].next,
52                                   struct page, lru);
53                 list_del(&page->lru);
54                 free_huge_pages--;
55                 free_huge_pages_node[nid]--;
56         }
57         return page;
58 }
59
60 static struct page *alloc_fresh_huge_page(void)
61 {
62         static int nid = 0;
63         struct page *page;
64         page = alloc_pages_node(nid, GFP_HIGHUSER|__GFP_COMP|__GFP_NOWARN,
65                                         HUGETLB_PAGE_ORDER);
66         nid = (nid + 1) % num_online_nodes();
67         if (page) {
68                 spin_lock(&hugetlb_lock);
69                 nr_huge_pages++;
70                 nr_huge_pages_node[page_to_nid(page)]++;
71                 spin_unlock(&hugetlb_lock);
72         }
73         return page;
74 }
75
76 void free_huge_page(struct page *page)
77 {
78         BUG_ON(page_count(page));
79
80         INIT_LIST_HEAD(&page->lru);
81         page[1].mapping = NULL;
82
83         spin_lock(&hugetlb_lock);
84         enqueue_huge_page(page);
85         spin_unlock(&hugetlb_lock);
86 }
87
88 struct page *alloc_huge_page(void)
89 {
90         struct page *page;
91         int i;
92
93         spin_lock(&hugetlb_lock);
94         page = dequeue_huge_page();
95         if (!page) {
96                 spin_unlock(&hugetlb_lock);
97                 return NULL;
98         }
99         spin_unlock(&hugetlb_lock);
100         set_page_count(page, 1);
101         page[1].mapping = (void *)free_huge_page;
102         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); ++i)
103                 clear_highpage(&page[i]);
104         return page;
105 }
106
107 static int __init hugetlb_init(void)
108 {
109         unsigned long i;
110         struct page *page;
111
112         if (HPAGE_SHIFT == 0)
113                 return 0;
114
115         for (i = 0; i < MAX_NUMNODES; ++i)
116                 INIT_LIST_HEAD(&hugepage_freelists[i]);
117
118         for (i = 0; i < max_huge_pages; ++i) {
119                 page = alloc_fresh_huge_page();
120                 if (!page)
121                         break;
122                 spin_lock(&hugetlb_lock);
123                 enqueue_huge_page(page);
124                 spin_unlock(&hugetlb_lock);
125         }
126         max_huge_pages = free_huge_pages = nr_huge_pages = i;
127         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
128         return 0;
129 }
130 module_init(hugetlb_init);
131
132 static int __init hugetlb_setup(char *s)
133 {
134         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
135                 max_huge_pages = 0;
136         return 1;
137 }
138 __setup("hugepages=", hugetlb_setup);
139
140 #ifdef CONFIG_SYSCTL
141 static void update_and_free_page(struct page *page)
142 {
143         int i;
144         nr_huge_pages--;
145         nr_huge_pages_node[page_zone(page)->zone_pgdat->node_id]--;
146         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
147                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
148                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
149                                 1 << PG_private | 1<< PG_writeback);
150                 set_page_count(&page[i], 0);
151         }
152         set_page_count(page, 1);
153         __free_pages(page, HUGETLB_PAGE_ORDER);
154 }
155
156 #ifdef CONFIG_HIGHMEM
157 static void try_to_free_low(unsigned long count)
158 {
159         int i, nid;
160         for (i = 0; i < MAX_NUMNODES; ++i) {
161                 struct page *page, *next;
162                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
163                         if (PageHighMem(page))
164                                 continue;
165                         list_del(&page->lru);
166                         update_and_free_page(page);
167                         nid = page_zone(page)->zone_pgdat->node_id;
168                         free_huge_pages--;
169                         free_huge_pages_node[nid]--;
170                         if (count >= nr_huge_pages)
171                                 return;
172                 }
173         }
174 }
175 #else
176 static inline void try_to_free_low(unsigned long count)
177 {
178 }
179 #endif
180
181 static unsigned long set_max_huge_pages(unsigned long count)
182 {
183         while (count > nr_huge_pages) {
184                 struct page *page = alloc_fresh_huge_page();
185                 if (!page)
186                         return nr_huge_pages;
187                 spin_lock(&hugetlb_lock);
188                 enqueue_huge_page(page);
189                 spin_unlock(&hugetlb_lock);
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();
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 struct page *find_or_alloc_huge_page(struct address_space *mapping,
367                                 unsigned long idx, int shared)
368 {
369         struct page *page;
370         int err;
371
372 retry:
373         page = find_lock_page(mapping, idx);
374         if (page)
375                 goto out;
376
377         if (hugetlb_get_quota(mapping))
378                 goto out;
379         page = alloc_huge_page();
380         if (!page) {
381                 hugetlb_put_quota(mapping);
382                 goto out;
383         }
384
385         if (shared) {
386                 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
387                 if (err) {
388                         put_page(page);
389                         hugetlb_put_quota(mapping);
390                         if (err == -EEXIST)
391                                 goto retry;
392                         page = NULL;
393                 }
394         } else {
395                 /* Caller expects a locked page */
396                 lock_page(page);
397         }
398 out:
399         return page;
400 }
401
402 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
403                         unsigned long address, pte_t *ptep, pte_t pte)
404 {
405         struct page *old_page, *new_page;
406         int i, avoidcopy;
407
408         old_page = pte_page(pte);
409
410         /* If no-one else is actually using this page, avoid the copy
411          * and just make the page writable */
412         avoidcopy = (page_count(old_page) == 1);
413         if (avoidcopy) {
414                 set_huge_ptep_writable(vma, address, ptep);
415                 return VM_FAULT_MINOR;
416         }
417
418         page_cache_get(old_page);
419         new_page = alloc_huge_page();
420
421         if (!new_page) {
422                 page_cache_release(old_page);
423
424                 /* Logically this is OOM, not a SIGBUS, but an OOM
425                  * could cause the kernel to go killing other
426                  * processes which won't help the hugepage situation
427                  * at all (?) */
428                 return VM_FAULT_SIGBUS;
429         }
430
431         spin_unlock(&mm->page_table_lock);
432         for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++)
433                 copy_user_highpage(new_page + i, old_page + i,
434                                    address + i*PAGE_SIZE);
435         spin_lock(&mm->page_table_lock);
436
437         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
438         if (likely(pte_same(*ptep, pte))) {
439                 /* Break COW */
440                 set_huge_pte_at(mm, address, ptep,
441                                 make_huge_pte(vma, new_page, 1));
442                 /* Make the old page be freed below */
443                 new_page = old_page;
444         }
445         page_cache_release(new_page);
446         page_cache_release(old_page);
447         return VM_FAULT_MINOR;
448 }
449
450 int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
451                         unsigned long address, pte_t *ptep, int write_access)
452 {
453         int ret = VM_FAULT_SIGBUS;
454         unsigned long idx;
455         unsigned long size;
456         struct page *page;
457         struct address_space *mapping;
458         pte_t new_pte;
459
460         mapping = vma->vm_file->f_mapping;
461         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
462                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
463
464         /*
465          * Use page lock to guard against racing truncation
466          * before we get page_table_lock.
467          */
468         page = find_or_alloc_huge_page(mapping, idx,
469                         vma->vm_flags & VM_SHARED);
470         if (!page)
471                 goto out;
472
473         BUG_ON(!PageLocked(page));
474
475         spin_lock(&mm->page_table_lock);
476         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
477         if (idx >= size)
478                 goto backout;
479
480         ret = VM_FAULT_MINOR;
481         if (!pte_none(*ptep))
482                 goto backout;
483
484         add_mm_counter(mm, file_rss, HPAGE_SIZE / PAGE_SIZE);
485         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
486                                 && (vma->vm_flags & VM_SHARED)));
487         set_huge_pte_at(mm, address, ptep, new_pte);
488
489         if (write_access && !(vma->vm_flags & VM_SHARED)) {
490                 /* Optimization, do the COW without a second fault */
491                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
492         }
493
494         spin_unlock(&mm->page_table_lock);
495         unlock_page(page);
496 out:
497         return ret;
498
499 backout:
500         spin_unlock(&mm->page_table_lock);
501         hugetlb_put_quota(mapping);
502         unlock_page(page);
503         put_page(page);
504         goto out;
505 }
506
507 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
508                         unsigned long address, int write_access)
509 {
510         pte_t *ptep;
511         pte_t entry;
512         int ret;
513
514         ptep = huge_pte_alloc(mm, address);
515         if (!ptep)
516                 return VM_FAULT_OOM;
517
518         entry = *ptep;
519         if (pte_none(entry))
520                 return hugetlb_no_page(mm, vma, address, ptep, write_access);
521
522         ret = VM_FAULT_MINOR;
523
524         spin_lock(&mm->page_table_lock);
525         /* Check for a racing update before calling hugetlb_cow */
526         if (likely(pte_same(entry, *ptep)))
527                 if (write_access && !pte_write(entry))
528                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
529         spin_unlock(&mm->page_table_lock);
530
531         return ret;
532 }
533
534 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
535                         struct page **pages, struct vm_area_struct **vmas,
536                         unsigned long *position, int *length, int i)
537 {
538         unsigned long vpfn, vaddr = *position;
539         int remainder = *length;
540
541         vpfn = vaddr/PAGE_SIZE;
542         spin_lock(&mm->page_table_lock);
543         while (vaddr < vma->vm_end && remainder) {
544                 pte_t *pte;
545                 struct page *page;
546
547                 /*
548                  * Some archs (sparc64, sh*) have multiple pte_ts to
549                  * each hugepage.  We have to make * sure we get the
550                  * first, for the page indexing below to work.
551                  */
552                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
553
554                 if (!pte || pte_none(*pte)) {
555                         int ret;
556
557                         spin_unlock(&mm->page_table_lock);
558                         ret = hugetlb_fault(mm, vma, vaddr, 0);
559                         spin_lock(&mm->page_table_lock);
560                         if (ret == VM_FAULT_MINOR)
561                                 continue;
562
563                         remainder = 0;
564                         if (!i)
565                                 i = -EFAULT;
566                         break;
567                 }
568
569                 if (pages) {
570                         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
571                         get_page(page);
572                         pages[i] = page;
573                 }
574
575                 if (vmas)
576                         vmas[i] = vma;
577
578                 vaddr += PAGE_SIZE;
579                 ++vpfn;
580                 --remainder;
581                 ++i;
582         }
583         spin_unlock(&mm->page_table_lock);
584         *length = remainder;
585         *position = vaddr;
586
587         return i;
588 }