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