blob: 094455bcbbf73c4476162336806acef804e7d449 [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;
Hugh Dickins1c598272005-10-19 21:23:43 -0700277 unsigned long addr;
David Gibson63551ae2005-06-21 17:14:44 -0700278
Hugh Dickins1c598272005-10-19 21:23:43 -0700279 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
David Gibson63551ae2005-06-21 17:14:44 -0700280 dst_pte = huge_pte_alloc(dst, addr);
281 if (!dst_pte)
282 goto nomem;
Hugh Dickins1c598272005-10-19 21:23:43 -0700283 spin_lock(&src->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700284 src_pte = huge_pte_offset(src, addr);
Hugh Dickins1c598272005-10-19 21:23:43 -0700285 if (src_pte && !pte_none(*src_pte)) {
286 entry = *src_pte;
287 ptepage = pte_page(entry);
288 get_page(ptepage);
Hugh Dickins42946212005-10-29 18:16:05 -0700289 add_mm_counter(dst, file_rss, HPAGE_SIZE / PAGE_SIZE);
Hugh Dickins1c598272005-10-19 21:23:43 -0700290 set_huge_pte_at(dst, addr, dst_pte, entry);
291 }
292 spin_unlock(&src->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700293 }
294 return 0;
295
296nomem:
297 return -ENOMEM;
298}
299
300void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
301 unsigned long end)
302{
303 struct mm_struct *mm = vma->vm_mm;
304 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -0700305 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -0700306 pte_t pte;
307 struct page *page;
308
309 WARN_ON(!is_vm_hugetlb_page(vma));
310 BUG_ON(start & ~HPAGE_MASK);
311 BUG_ON(end & ~HPAGE_MASK);
312
313 for (address = start; address < end; address += HPAGE_SIZE) {
David Gibsonc7546f82005-08-05 11:59:35 -0700314 ptep = huge_pte_offset(mm, address);
315 if (! ptep)
316 /* This can happen on truncate, or if an
317 * mmap() is aborted due to an error before
318 * the prefault */
319 continue;
320
321 pte = huge_ptep_get_and_clear(mm, address, ptep);
David Gibson63551ae2005-06-21 17:14:44 -0700322 if (pte_none(pte))
323 continue;
David Gibsonc7546f82005-08-05 11:59:35 -0700324
David Gibson63551ae2005-06-21 17:14:44 -0700325 page = pte_page(pte);
326 put_page(page);
Hugh Dickins42946212005-10-29 18:16:05 -0700327 add_mm_counter(mm, file_rss, (int) -(HPAGE_SIZE / PAGE_SIZE));
David Gibson63551ae2005-06-21 17:14:44 -0700328 }
David Gibson63551ae2005-06-21 17:14:44 -0700329 flush_tlb_range(vma, start, end);
330}
331
Linus Torvalds1da177e2005-04-16 15:20:36 -0700332void zap_hugepage_range(struct vm_area_struct *vma,
333 unsigned long start, unsigned long length)
334{
335 struct mm_struct *mm = vma->vm_mm;
336
337 spin_lock(&mm->page_table_lock);
338 unmap_hugepage_range(vma, start, start + length);
339 spin_unlock(&mm->page_table_lock);
340}
David Gibson63551ae2005-06-21 17:14:44 -0700341
342int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
343{
344 struct mm_struct *mm = current->mm;
345 unsigned long addr;
346 int ret = 0;
347
348 WARN_ON(!is_vm_hugetlb_page(vma));
349 BUG_ON(vma->vm_start & ~HPAGE_MASK);
350 BUG_ON(vma->vm_end & ~HPAGE_MASK);
351
352 hugetlb_prefault_arch_hook(mm);
353
354 spin_lock(&mm->page_table_lock);
355 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
356 unsigned long idx;
357 pte_t *pte = huge_pte_alloc(mm, addr);
358 struct page *page;
359
360 if (!pte) {
361 ret = -ENOMEM;
362 goto out;
363 }
David Gibson63551ae2005-06-21 17:14:44 -0700364
365 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
366 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
367 page = find_get_page(mapping, idx);
368 if (!page) {
369 /* charge the fs quota first */
370 if (hugetlb_get_quota(mapping)) {
371 ret = -ENOMEM;
372 goto out;
373 }
374 page = alloc_huge_page();
375 if (!page) {
376 hugetlb_put_quota(mapping);
377 ret = -ENOMEM;
378 goto out;
379 }
380 ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
381 if (! ret) {
382 unlock_page(page);
383 } else {
384 hugetlb_put_quota(mapping);
385 free_huge_page(page);
386 goto out;
387 }
388 }
Hugh Dickins42946212005-10-29 18:16:05 -0700389 add_mm_counter(mm, file_rss, HPAGE_SIZE / PAGE_SIZE);
David Gibson63551ae2005-06-21 17:14:44 -0700390 set_huge_pte_at(mm, addr, pte, make_huge_pte(vma, page));
391 }
392out:
393 spin_unlock(&mm->page_table_lock);
394 return ret;
395}
396
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100397/*
398 * On ia64 at least, it is possible to receive a hugetlb fault from a
399 * stale zero entry left in the TLB from earlier hardware prefetching.
400 * Low-level arch code should already have flushed the stale entry as
401 * part of its fault handling, but we do need to accept this minor fault
402 * and return successfully. Whereas the "normal" case is that this is
403 * an access to a hugetlb page which has been truncated off since mmap.
404 */
405int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
406 unsigned long address, int write_access)
407{
408 int ret = VM_FAULT_SIGBUS;
409 pte_t *pte;
410
411 spin_lock(&mm->page_table_lock);
412 pte = huge_pte_offset(mm, address);
413 if (pte && !pte_none(*pte))
414 ret = VM_FAULT_MINOR;
415 spin_unlock(&mm->page_table_lock);
416 return ret;
417}
418
David Gibson63551ae2005-06-21 17:14:44 -0700419int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
420 struct page **pages, struct vm_area_struct **vmas,
421 unsigned long *position, int *length, int i)
422{
423 unsigned long vpfn, vaddr = *position;
424 int remainder = *length;
425
426 BUG_ON(!is_vm_hugetlb_page(vma));
427
428 vpfn = vaddr/PAGE_SIZE;
Hugh Dickins1c598272005-10-19 21:23:43 -0700429 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700430 while (vaddr < vma->vm_end && remainder) {
431
432 if (pages) {
433 pte_t *pte;
434 struct page *page;
435
436 /* Some archs (sparc64, sh*) have multiple
437 * pte_ts to each hugepage. We have to make
438 * sure we get the first, for the page
439 * indexing below to work. */
440 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
441
Hugh Dickins1c598272005-10-19 21:23:43 -0700442 /* the hugetlb file might have been truncated */
443 if (!pte || pte_none(*pte)) {
444 remainder = 0;
445 if (!i)
446 i = -EFAULT;
447 break;
448 }
David Gibson63551ae2005-06-21 17:14:44 -0700449
450 page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
451
452 WARN_ON(!PageCompound(page));
453
454 get_page(page);
455 pages[i] = page;
456 }
457
458 if (vmas)
459 vmas[i] = vma;
460
461 vaddr += PAGE_SIZE;
462 ++vpfn;
463 --remainder;
464 ++i;
465 }
Hugh Dickins1c598272005-10-19 21:23:43 -0700466 spin_unlock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700467 *length = remainder;
468 *position = vaddr;
469
470 return i;
471}