Merge by hand (conflicts between pending drivers and kfree cleanups)
[linux-2.6] / 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 static DEFINE_SPINLOCK(hugetlb_lock);
26
27 static void enqueue_huge_page(struct page *page)
28 {
29         int nid = page_to_nid(page);
30         list_add(&page->lru, &hugepage_freelists[nid]);
31         free_huge_pages++;
32         free_huge_pages_node[nid]++;
33 }
34
35 static struct page *dequeue_huge_page(void)
36 {
37         int nid = numa_node_id();
38         struct page *page = NULL;
39
40         if (list_empty(&hugepage_freelists[nid])) {
41                 for (nid = 0; nid < MAX_NUMNODES; ++nid)
42                         if (!list_empty(&hugepage_freelists[nid]))
43                                 break;
44         }
45         if (nid >= 0 && nid < MAX_NUMNODES &&
46             !list_empty(&hugepage_freelists[nid])) {
47                 page = list_entry(hugepage_freelists[nid].next,
48                                   struct page, lru);
49                 list_del(&page->lru);
50                 free_huge_pages--;
51                 free_huge_pages_node[nid]--;
52         }
53         return page;
54 }
55
56 static struct page *alloc_fresh_huge_page(void)
57 {
58         static int nid = 0;
59         struct page *page;
60         page = alloc_pages_node(nid, GFP_HIGHUSER|__GFP_COMP|__GFP_NOWARN,
61                                         HUGETLB_PAGE_ORDER);
62         nid = (nid + 1) % num_online_nodes();
63         if (page) {
64                 nr_huge_pages++;
65                 nr_huge_pages_node[page_to_nid(page)]++;
66         }
67         return page;
68 }
69
70 void free_huge_page(struct page *page)
71 {
72         BUG_ON(page_count(page));
73
74         INIT_LIST_HEAD(&page->lru);
75         page[1].mapping = NULL;
76
77         spin_lock(&hugetlb_lock);
78         enqueue_huge_page(page);
79         spin_unlock(&hugetlb_lock);
80 }
81
82 struct page *alloc_huge_page(void)
83 {
84         struct page *page;
85         int i;
86
87         spin_lock(&hugetlb_lock);
88         page = dequeue_huge_page();
89         if (!page) {
90                 spin_unlock(&hugetlb_lock);
91                 return NULL;
92         }
93         spin_unlock(&hugetlb_lock);
94         set_page_count(page, 1);
95         page[1].mapping = (void *)free_huge_page;
96         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); ++i)
97                 clear_highpage(&page[i]);
98         return page;
99 }
100
101 static int __init hugetlb_init(void)
102 {
103         unsigned long i;
104         struct page *page;
105
106         if (HPAGE_SHIFT == 0)
107                 return 0;
108
109         for (i = 0; i < MAX_NUMNODES; ++i)
110                 INIT_LIST_HEAD(&hugepage_freelists[i]);
111
112         for (i = 0; i < max_huge_pages; ++i) {
113                 page = alloc_fresh_huge_page();
114                 if (!page)
115                         break;
116                 spin_lock(&hugetlb_lock);
117                 enqueue_huge_page(page);
118                 spin_unlock(&hugetlb_lock);
119         }
120         max_huge_pages = free_huge_pages = nr_huge_pages = i;
121         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
122         return 0;
123 }
124 module_init(hugetlb_init);
125
126 static int __init hugetlb_setup(char *s)
127 {
128         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
129                 max_huge_pages = 0;
130         return 1;
131 }
132 __setup("hugepages=", hugetlb_setup);
133
134 #ifdef CONFIG_SYSCTL
135 static void update_and_free_page(struct page *page)
136 {
137         int i;
138         nr_huge_pages--;
139         nr_huge_pages_node[page_zone(page)->zone_pgdat->node_id]--;
140         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
141                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
142                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
143                                 1 << PG_private | 1<< PG_writeback);
144                 set_page_count(&page[i], 0);
145         }
146         set_page_count(page, 1);
147         __free_pages(page, HUGETLB_PAGE_ORDER);
148 }
149
150 #ifdef CONFIG_HIGHMEM
151 static void try_to_free_low(unsigned long count)
152 {
153         int i, nid;
154         for (i = 0; i < MAX_NUMNODES; ++i) {
155                 struct page *page, *next;
156                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
157                         if (PageHighMem(page))
158                                 continue;
159                         list_del(&page->lru);
160                         update_and_free_page(page);
161                         nid = page_zone(page)->zone_pgdat->node_id;
162                         free_huge_pages--;
163                         free_huge_pages_node[nid]--;
164                         if (count >= nr_huge_pages)
165                                 return;
166                 }
167         }
168 }
169 #else
170 static inline void try_to_free_low(unsigned long count)
171 {
172 }
173 #endif
174
175 static unsigned long set_max_huge_pages(unsigned long count)
176 {
177         while (count > nr_huge_pages) {
178                 struct page *page = alloc_fresh_huge_page();
179                 if (!page)
180                         return nr_huge_pages;
181                 spin_lock(&hugetlb_lock);
182                 enqueue_huge_page(page);
183                 spin_unlock(&hugetlb_lock);
184         }
185         if (count >= nr_huge_pages)
186                 return nr_huge_pages;
187
188         spin_lock(&hugetlb_lock);
189         try_to_free_low(count);
190         while (count < nr_huge_pages) {
191                 struct page *page = dequeue_huge_page();
192                 if (!page)
193                         break;
194                 update_and_free_page(page);
195         }
196         spin_unlock(&hugetlb_lock);
197         return nr_huge_pages;
198 }
199
200 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
201                            struct file *file, void __user *buffer,
202                            size_t *length, loff_t *ppos)
203 {
204         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
205         max_huge_pages = set_max_huge_pages(max_huge_pages);
206         return 0;
207 }
208 #endif /* CONFIG_SYSCTL */
209
210 int hugetlb_report_meminfo(char *buf)
211 {
212         return sprintf(buf,
213                         "HugePages_Total: %5lu\n"
214                         "HugePages_Free:  %5lu\n"
215                         "Hugepagesize:    %5lu kB\n",
216                         nr_huge_pages,
217                         free_huge_pages,
218                         HPAGE_SIZE/1024);
219 }
220
221 int hugetlb_report_node_meminfo(int nid, char *buf)
222 {
223         return sprintf(buf,
224                 "Node %d HugePages_Total: %5u\n"
225                 "Node %d HugePages_Free:  %5u\n",
226                 nid, nr_huge_pages_node[nid],
227                 nid, free_huge_pages_node[nid]);
228 }
229
230 int is_hugepage_mem_enough(size_t size)
231 {
232         return (size + ~HPAGE_MASK)/HPAGE_SIZE <= free_huge_pages;
233 }
234
235 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
236 unsigned long hugetlb_total_pages(void)
237 {
238         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
239 }
240
241 /*
242  * We cannot handle pagefaults against hugetlb pages at all.  They cause
243  * handle_mm_fault() to try to instantiate regular-sized pages in the
244  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
245  * this far.
246  */
247 static struct page *hugetlb_nopage(struct vm_area_struct *vma,
248                                 unsigned long address, int *unused)
249 {
250         BUG();
251         return NULL;
252 }
253
254 struct vm_operations_struct hugetlb_vm_ops = {
255         .nopage = hugetlb_nopage,
256 };
257
258 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page)
259 {
260         pte_t entry;
261
262         if (vma->vm_flags & VM_WRITE) {
263                 entry =
264                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
265         } else {
266                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
267         }
268         entry = pte_mkyoung(entry);
269         entry = pte_mkhuge(entry);
270
271         return entry;
272 }
273
274 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
275                             struct vm_area_struct *vma)
276 {
277         pte_t *src_pte, *dst_pte, entry;
278         struct page *ptepage;
279         unsigned long addr;
280
281         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
282                 src_pte = huge_pte_offset(src, addr);
283                 if (!src_pte)
284                         continue;
285                 dst_pte = huge_pte_alloc(dst, addr);
286                 if (!dst_pte)
287                         goto nomem;
288                 spin_lock(&dst->page_table_lock);
289                 spin_lock(&src->page_table_lock);
290                 if (!pte_none(*src_pte)) {
291                         entry = *src_pte;
292                         ptepage = pte_page(entry);
293                         get_page(ptepage);
294                         add_mm_counter(dst, file_rss, HPAGE_SIZE / PAGE_SIZE);
295                         set_huge_pte_at(dst, addr, dst_pte, entry);
296                 }
297                 spin_unlock(&src->page_table_lock);
298                 spin_unlock(&dst->page_table_lock);
299         }
300         return 0;
301
302 nomem:
303         return -ENOMEM;
304 }
305
306 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
307                           unsigned long end)
308 {
309         struct mm_struct *mm = vma->vm_mm;
310         unsigned long address;
311         pte_t *ptep;
312         pte_t pte;
313         struct page *page;
314
315         WARN_ON(!is_vm_hugetlb_page(vma));
316         BUG_ON(start & ~HPAGE_MASK);
317         BUG_ON(end & ~HPAGE_MASK);
318
319         spin_lock(&mm->page_table_lock);
320
321         /* Update high watermark before we lower rss */
322         update_hiwater_rss(mm);
323
324         for (address = start; address < end; address += HPAGE_SIZE) {
325                 ptep = huge_pte_offset(mm, address);
326                 if (!ptep)
327                         continue;
328
329                 pte = huge_ptep_get_and_clear(mm, address, ptep);
330                 if (pte_none(pte))
331                         continue;
332
333                 page = pte_page(pte);
334                 put_page(page);
335                 add_mm_counter(mm, file_rss, (int) -(HPAGE_SIZE / PAGE_SIZE));
336         }
337
338         spin_unlock(&mm->page_table_lock);
339         flush_tlb_range(vma, start, end);
340 }
341
342 static struct page *find_lock_huge_page(struct address_space *mapping,
343                         unsigned long idx)
344 {
345         struct page *page;
346         int err;
347         struct inode *inode = mapping->host;
348         unsigned long size;
349
350 retry:
351         page = find_lock_page(mapping, idx);
352         if (page)
353                 goto out;
354
355         /* Check to make sure the mapping hasn't been truncated */
356         size = i_size_read(inode) >> HPAGE_SHIFT;
357         if (idx >= size)
358                 goto out;
359
360         if (hugetlb_get_quota(mapping))
361                 goto out;
362         page = alloc_huge_page();
363         if (!page) {
364                 hugetlb_put_quota(mapping);
365                 goto out;
366         }
367
368         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
369         if (err) {
370                 put_page(page);
371                 hugetlb_put_quota(mapping);
372                 if (err == -EEXIST)
373                         goto retry;
374                 page = NULL;
375         }
376 out:
377         return page;
378 }
379
380 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
381                         unsigned long address, int write_access)
382 {
383         int ret = VM_FAULT_SIGBUS;
384         unsigned long idx;
385         unsigned long size;
386         pte_t *pte;
387         struct page *page;
388         struct address_space *mapping;
389
390         pte = huge_pte_alloc(mm, address);
391         if (!pte)
392                 goto out;
393
394         mapping = vma->vm_file->f_mapping;
395         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
396                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
397
398         /*
399          * Use page lock to guard against racing truncation
400          * before we get page_table_lock.
401          */
402         page = find_lock_huge_page(mapping, idx);
403         if (!page)
404                 goto out;
405
406         spin_lock(&mm->page_table_lock);
407         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
408         if (idx >= size)
409                 goto backout;
410
411         ret = VM_FAULT_MINOR;
412         if (!pte_none(*pte))
413                 goto backout;
414
415         add_mm_counter(mm, file_rss, HPAGE_SIZE / PAGE_SIZE);
416         set_huge_pte_at(mm, address, pte, make_huge_pte(vma, page));
417         spin_unlock(&mm->page_table_lock);
418         unlock_page(page);
419 out:
420         return ret;
421
422 backout:
423         spin_unlock(&mm->page_table_lock);
424         hugetlb_put_quota(mapping);
425         unlock_page(page);
426         put_page(page);
427         goto out;
428 }
429
430 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
431                         struct page **pages, struct vm_area_struct **vmas,
432                         unsigned long *position, int *length, int i)
433 {
434         unsigned long vpfn, vaddr = *position;
435         int remainder = *length;
436
437         vpfn = vaddr/PAGE_SIZE;
438         spin_lock(&mm->page_table_lock);
439         while (vaddr < vma->vm_end && remainder) {
440                 pte_t *pte;
441                 struct page *page;
442
443                 /*
444                  * Some archs (sparc64, sh*) have multiple pte_ts to
445                  * each hugepage.  We have to make * sure we get the
446                  * first, for the page indexing below to work.
447                  */
448                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
449
450                 if (!pte || pte_none(*pte)) {
451                         int ret;
452
453                         spin_unlock(&mm->page_table_lock);
454                         ret = hugetlb_fault(mm, vma, vaddr, 0);
455                         spin_lock(&mm->page_table_lock);
456                         if (ret == VM_FAULT_MINOR)
457                                 continue;
458
459                         remainder = 0;
460                         if (!i)
461                                 i = -EFAULT;
462                         break;
463                 }
464
465                 if (pages) {
466                         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
467                         get_page(page);
468                         pages[i] = page;
469                 }
470
471                 if (vmas)
472                         vmas[i] = vma;
473
474                 vaddr += PAGE_SIZE;
475                 ++vpfn;
476                 --remainder;
477                 ++i;
478         }
479         spin_unlock(&mm->page_table_lock);
480         *length = remainder;
481         *position = vaddr;
482
483         return i;
484 }