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