blob: 6bf720bc662c41983c1208e2aef409d4e0a276ac [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include <linux/sysctl.h>
11#include <linux/highmem.h>
12#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070013#include <linux/pagemap.h>
14#include <asm/page.h>
15#include <asm/pgtable.h>
16
17#include <linux/hugetlb.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070018
19const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
20static unsigned long nr_huge_pages, free_huge_pages;
21unsigned long max_huge_pages;
22static struct list_head hugepage_freelists[MAX_NUMNODES];
23static unsigned int nr_huge_pages_node[MAX_NUMNODES];
24static unsigned int free_huge_pages_node[MAX_NUMNODES];
25static DEFINE_SPINLOCK(hugetlb_lock);
26
27static 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
35static 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
56static 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
70void 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
82struct 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
101static 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}
121module_init(hugetlb_init);
122
123static 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
132static 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
148static 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
167static inline void try_to_free_low(unsigned long count)
168{
169}
170#endif
171
172static 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
197int 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
207int 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
218int 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
227int 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. */
233unsigned long hugetlb_total_pages(void)
234{
235 return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
236}
237EXPORT_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 */
245static struct page *hugetlb_nopage(struct vm_area_struct *vma,
246 unsigned long address, int *unused)
247{
248 BUG();
249 return NULL;
250}
251
252struct vm_operations_struct hugetlb_vm_ops = {
253 .nopage = hugetlb_nopage,
254};
255
David Gibson63551ae2005-06-21 17:14:44 -0700256static 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
272int 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 = vma->vm_start;
278 unsigned long end = vma->vm_end;
279
280 while (addr < end) {
281 dst_pte = huge_pte_alloc(dst, addr);
282 if (!dst_pte)
283 goto nomem;
284 src_pte = huge_pte_offset(src, addr);
285 BUG_ON(!src_pte || pte_none(*src_pte)); /* prefaulted */
286 entry = *src_pte;
287 ptepage = pte_page(entry);
288 get_page(ptepage);
289 add_mm_counter(dst, rss, HPAGE_SIZE / PAGE_SIZE);
290 set_huge_pte_at(dst, addr, dst_pte, entry);
291 addr += HPAGE_SIZE;
292 }
293 return 0;
294
295nomem:
296 return -ENOMEM;
297}
298
299void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
300 unsigned long end)
301{
302 struct mm_struct *mm = vma->vm_mm;
303 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -0700304 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -0700305 pte_t pte;
306 struct page *page;
307
308 WARN_ON(!is_vm_hugetlb_page(vma));
309 BUG_ON(start & ~HPAGE_MASK);
310 BUG_ON(end & ~HPAGE_MASK);
311
312 for (address = start; address < end; address += HPAGE_SIZE) {
David Gibsonc7546f82005-08-05 11:59:35 -0700313 ptep = huge_pte_offset(mm, address);
314 if (! ptep)
315 /* This can happen on truncate, or if an
316 * mmap() is aborted due to an error before
317 * the prefault */
318 continue;
319
320 pte = huge_ptep_get_and_clear(mm, address, ptep);
David Gibson63551ae2005-06-21 17:14:44 -0700321 if (pte_none(pte))
322 continue;
David Gibsonc7546f82005-08-05 11:59:35 -0700323
David Gibson63551ae2005-06-21 17:14:44 -0700324 page = pte_page(pte);
325 put_page(page);
326 }
327 add_mm_counter(mm, rss, -((end - start) >> PAGE_SHIFT));
328 flush_tlb_range(vma, start, end);
329}
330
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331void zap_hugepage_range(struct vm_area_struct *vma,
332 unsigned long start, unsigned long length)
333{
334 struct mm_struct *mm = vma->vm_mm;
335
336 spin_lock(&mm->page_table_lock);
337 unmap_hugepage_range(vma, start, start + length);
338 spin_unlock(&mm->page_table_lock);
339}
David Gibson63551ae2005-06-21 17:14:44 -0700340
341int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
342{
343 struct mm_struct *mm = current->mm;
344 unsigned long addr;
345 int ret = 0;
346
347 WARN_ON(!is_vm_hugetlb_page(vma));
348 BUG_ON(vma->vm_start & ~HPAGE_MASK);
349 BUG_ON(vma->vm_end & ~HPAGE_MASK);
350
351 hugetlb_prefault_arch_hook(mm);
352
353 spin_lock(&mm->page_table_lock);
354 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
355 unsigned long idx;
356 pte_t *pte = huge_pte_alloc(mm, addr);
357 struct page *page;
358
359 if (!pte) {
360 ret = -ENOMEM;
361 goto out;
362 }
363 if (! pte_none(*pte))
364 hugetlb_clean_stale_pgtable(pte);
365
366 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
367 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
368 page = find_get_page(mapping, idx);
369 if (!page) {
370 /* charge the fs quota first */
371 if (hugetlb_get_quota(mapping)) {
372 ret = -ENOMEM;
373 goto out;
374 }
375 page = alloc_huge_page();
376 if (!page) {
377 hugetlb_put_quota(mapping);
378 ret = -ENOMEM;
379 goto out;
380 }
381 ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
382 if (! ret) {
383 unlock_page(page);
384 } else {
385 hugetlb_put_quota(mapping);
386 free_huge_page(page);
387 goto out;
388 }
389 }
390 add_mm_counter(mm, rss, HPAGE_SIZE / PAGE_SIZE);
391 set_huge_pte_at(mm, addr, pte, make_huge_pte(vma, page));
392 }
393out:
394 spin_unlock(&mm->page_table_lock);
395 return ret;
396}
397
398int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
399 struct page **pages, struct vm_area_struct **vmas,
400 unsigned long *position, int *length, int i)
401{
402 unsigned long vpfn, vaddr = *position;
403 int remainder = *length;
404
405 BUG_ON(!is_vm_hugetlb_page(vma));
406
407 vpfn = vaddr/PAGE_SIZE;
408 while (vaddr < vma->vm_end && remainder) {
409
410 if (pages) {
411 pte_t *pte;
412 struct page *page;
413
414 /* Some archs (sparc64, sh*) have multiple
415 * pte_ts to each hugepage. We have to make
416 * sure we get the first, for the page
417 * indexing below to work. */
418 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
419
420 /* hugetlb should be locked, and hence, prefaulted */
421 WARN_ON(!pte || pte_none(*pte));
422
423 page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
424
425 WARN_ON(!PageCompound(page));
426
427 get_page(page);
428 pages[i] = page;
429 }
430
431 if (vmas)
432 vmas[i] = vma;
433
434 vaddr += PAGE_SIZE;
435 ++vpfn;
436 --remainder;
437 ++i;
438 }
439
440 *length = remainder;
441 *position = vaddr;
442
443 return i;
444}