blob: 832f676ca038837ab87c7544d4c3344b45cdc264 [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>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080014#include <linux/mempolicy.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080015#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080016#include <linux/mutex.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080017
David Gibson63551ae2005-06-21 17:14:44 -070018#include <asm/page.h>
19#include <asm/pgtable.h>
20
21#include <linux/hugetlb.h>
Nick Piggin7835e982006-03-22 00:08:40 -080022#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070023
24const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
David Gibsonb45b5bd2006-03-22 00:08:55 -080025static unsigned long nr_huge_pages, free_huge_pages, reserved_huge_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -070026unsigned long max_huge_pages;
27static struct list_head hugepage_freelists[MAX_NUMNODES];
28static unsigned int nr_huge_pages_node[MAX_NUMNODES];
29static unsigned int free_huge_pages_node[MAX_NUMNODES];
David Gibson3935baa2006-03-22 00:08:53 -080030/*
31 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
32 */
33static DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080034
David Gibson79ac6ba2006-03-22 00:08:51 -080035static void clear_huge_page(struct page *page, unsigned long addr)
36{
37 int i;
38
39 might_sleep();
40 for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
41 cond_resched();
42 clear_user_highpage(page + i, addr);
43 }
44}
45
46static void copy_huge_page(struct page *dst, struct page *src,
47 unsigned long addr)
48{
49 int i;
50
51 might_sleep();
52 for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
53 cond_resched();
54 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE);
55 }
56}
57
Linus Torvalds1da177e2005-04-16 15:20:36 -070058static void enqueue_huge_page(struct page *page)
59{
60 int nid = page_to_nid(page);
61 list_add(&page->lru, &hugepage_freelists[nid]);
62 free_huge_pages++;
63 free_huge_pages_node[nid]++;
64}
65
Christoph Lameter5da7ca82006-01-06 00:10:46 -080066static struct page *dequeue_huge_page(struct vm_area_struct *vma,
67 unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -070068{
69 int nid = numa_node_id();
70 struct page *page = NULL;
Christoph Lameter5da7ca82006-01-06 00:10:46 -080071 struct zonelist *zonelist = huge_zonelist(vma, address);
Christoph Lameter96df9332006-01-06 00:10:45 -080072 struct zone **z;
Linus Torvalds1da177e2005-04-16 15:20:36 -070073
Christoph Lameter96df9332006-01-06 00:10:45 -080074 for (z = zonelist->zones; *z; z++) {
75 nid = (*z)->zone_pgdat->node_id;
Christoph Lameteraea47ff2006-01-08 01:00:57 -080076 if (cpuset_zone_allowed(*z, GFP_HIGHUSER) &&
77 !list_empty(&hugepage_freelists[nid]))
Christoph Lameter96df9332006-01-06 00:10:45 -080078 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -070079 }
Christoph Lameter96df9332006-01-06 00:10:45 -080080
81 if (*z) {
Linus Torvalds1da177e2005-04-16 15:20:36 -070082 page = list_entry(hugepage_freelists[nid].next,
83 struct page, lru);
84 list_del(&page->lru);
85 free_huge_pages--;
86 free_huge_pages_node[nid]--;
87 }
88 return page;
89}
90
David Gibson27a85ef2006-03-22 00:08:56 -080091static void free_huge_page(struct page *page)
92{
93 BUG_ON(page_count(page));
94
95 INIT_LIST_HEAD(&page->lru);
96
97 spin_lock(&hugetlb_lock);
98 enqueue_huge_page(page);
99 spin_unlock(&hugetlb_lock);
100}
101
Nick Piggina4822892006-03-22 00:08:08 -0800102static int alloc_fresh_huge_page(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103{
104 static int nid = 0;
105 struct page *page;
106 page = alloc_pages_node(nid, GFP_HIGHUSER|__GFP_COMP|__GFP_NOWARN,
107 HUGETLB_PAGE_ORDER);
Paul Jacksonfdb7cc52006-03-22 00:09:10 -0800108 nid = next_node(nid, node_online_map);
109 if (nid == MAX_NUMNODES)
110 nid = first_node(node_online_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111 if (page) {
Nick Piggina4822892006-03-22 00:08:08 -0800112 page[1].lru.next = (void *)free_huge_page; /* dtor */
Eric Paris0bd0f9f2005-11-21 21:32:28 -0800113 spin_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114 nr_huge_pages++;
115 nr_huge_pages_node[page_to_nid(page)]++;
Eric Paris0bd0f9f2005-11-21 21:32:28 -0800116 spin_unlock(&hugetlb_lock);
Nick Piggina4822892006-03-22 00:08:08 -0800117 put_page(page); /* free it into the hugepage allocator */
118 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119 }
Nick Piggina4822892006-03-22 00:08:08 -0800120 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121}
122
David Gibson27a85ef2006-03-22 00:08:56 -0800123static struct page *alloc_huge_page(struct vm_area_struct *vma,
124 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125{
David Gibsonb45b5bd2006-03-22 00:08:55 -0800126 struct inode *inode = vma->vm_file->f_dentry->d_inode;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127 struct page *page;
David Gibsonb45b5bd2006-03-22 00:08:55 -0800128 int use_reserve = 0;
129 unsigned long idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700130
131 spin_lock(&hugetlb_lock);
David Gibsonb45b5bd2006-03-22 00:08:55 -0800132
133 if (vma->vm_flags & VM_MAYSHARE) {
134
135 /* idx = radix tree index, i.e. offset into file in
136 * HPAGE_SIZE units */
137 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
138 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
139
140 /* The hugetlbfs specific inode info stores the number
141 * of "guaranteed available" (huge) pages. That is,
142 * the first 'prereserved_hpages' pages of the inode
143 * are either already instantiated, or have been
144 * pre-reserved (by hugetlb_reserve_for_inode()). Here
145 * we're in the process of instantiating the page, so
146 * we use this to determine whether to draw from the
147 * pre-reserved pool or the truly free pool. */
148 if (idx < HUGETLBFS_I(inode)->prereserved_hpages)
149 use_reserve = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150 }
David Gibsonb45b5bd2006-03-22 00:08:55 -0800151
152 if (!use_reserve) {
153 if (free_huge_pages <= reserved_huge_pages)
154 goto fail;
155 } else {
156 BUG_ON(reserved_huge_pages == 0);
157 reserved_huge_pages--;
158 }
159
160 page = dequeue_huge_page(vma, addr);
161 if (!page)
162 goto fail;
163
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164 spin_unlock(&hugetlb_lock);
Nick Piggin7835e982006-03-22 00:08:40 -0800165 set_page_refcounted(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -0800167
168 fail:
169 WARN_ON(use_reserve); /* reserved allocations shouldn't fail */
170 spin_unlock(&hugetlb_lock);
171 return NULL;
172}
173
174/* hugetlb_extend_reservation()
175 *
176 * Ensure that at least 'atleast' hugepages are, and will remain,
177 * available to instantiate the first 'atleast' pages of the given
178 * inode. If the inode doesn't already have this many pages reserved
179 * or instantiated, set aside some hugepages in the reserved pool to
180 * satisfy later faults (or fail now if there aren't enough, rather
181 * than getting the SIGBUS later).
182 */
183int hugetlb_extend_reservation(struct hugetlbfs_inode_info *info,
184 unsigned long atleast)
185{
186 struct inode *inode = &info->vfs_inode;
187 unsigned long change_in_reserve = 0;
188 int ret = 0;
189
190 spin_lock(&hugetlb_lock);
191 read_lock_irq(&inode->i_mapping->tree_lock);
192
193 if (info->prereserved_hpages >= atleast)
194 goto out;
195
196 /* Because we always call this on shared mappings, none of the
197 * pages beyond info->prereserved_hpages can have been
198 * instantiated, so we need to reserve all of them now. */
199 change_in_reserve = atleast - info->prereserved_hpages;
200
201 if ((reserved_huge_pages + change_in_reserve) > free_huge_pages) {
202 ret = -ENOMEM;
203 goto out;
204 }
205
206 reserved_huge_pages += change_in_reserve;
207 info->prereserved_hpages = atleast;
208
209 out:
210 read_unlock_irq(&inode->i_mapping->tree_lock);
211 spin_unlock(&hugetlb_lock);
212
213 return ret;
214}
215
216/* hugetlb_truncate_reservation()
217 *
218 * This returns pages reserved for the given inode to the general free
219 * hugepage pool. If the inode has any pages prereserved, but not
220 * instantiated, beyond offset (atmost << HPAGE_SIZE), then release
221 * them.
222 */
223void hugetlb_truncate_reservation(struct hugetlbfs_inode_info *info,
224 unsigned long atmost)
225{
226 struct inode *inode = &info->vfs_inode;
227 struct address_space *mapping = inode->i_mapping;
228 unsigned long idx;
229 unsigned long change_in_reserve = 0;
230 struct page *page;
231
232 spin_lock(&hugetlb_lock);
233 read_lock_irq(&inode->i_mapping->tree_lock);
234
235 if (info->prereserved_hpages <= atmost)
236 goto out;
237
238 /* Count pages which were reserved, but not instantiated, and
239 * which we can now release. */
240 for (idx = atmost; idx < info->prereserved_hpages; idx++) {
241 page = radix_tree_lookup(&mapping->page_tree, idx);
242 if (!page)
243 /* Pages which are already instantiated can't
244 * be unreserved (and in fact have already
245 * been removed from the reserved pool) */
246 change_in_reserve++;
247 }
248
249 BUG_ON(reserved_huge_pages < change_in_reserve);
250 reserved_huge_pages -= change_in_reserve;
251 info->prereserved_hpages = atmost;
252
253 out:
254 read_unlock_irq(&inode->i_mapping->tree_lock);
255 spin_unlock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256}
257
258static int __init hugetlb_init(void)
259{
260 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261
Benjamin Herrenschmidt3c726f82005-11-07 11:06:55 +1100262 if (HPAGE_SHIFT == 0)
263 return 0;
264
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265 for (i = 0; i < MAX_NUMNODES; ++i)
266 INIT_LIST_HEAD(&hugepage_freelists[i]);
267
268 for (i = 0; i < max_huge_pages; ++i) {
Nick Piggina4822892006-03-22 00:08:08 -0800269 if (!alloc_fresh_huge_page())
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271 }
272 max_huge_pages = free_huge_pages = nr_huge_pages = i;
273 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
274 return 0;
275}
276module_init(hugetlb_init);
277
278static int __init hugetlb_setup(char *s)
279{
280 if (sscanf(s, "%lu", &max_huge_pages) <= 0)
281 max_huge_pages = 0;
282 return 1;
283}
284__setup("hugepages=", hugetlb_setup);
285
286#ifdef CONFIG_SYSCTL
287static void update_and_free_page(struct page *page)
288{
289 int i;
290 nr_huge_pages--;
291 nr_huge_pages_node[page_zone(page)->zone_pgdat->node_id]--;
292 for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
293 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
294 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
295 1 << PG_private | 1<< PG_writeback);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296 }
Nick Piggina4822892006-03-22 00:08:08 -0800297 page[1].lru.next = NULL;
Nick Piggin7835e982006-03-22 00:08:40 -0800298 set_page_refcounted(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700299 __free_pages(page, HUGETLB_PAGE_ORDER);
300}
301
302#ifdef CONFIG_HIGHMEM
303static void try_to_free_low(unsigned long count)
304{
305 int i, nid;
306 for (i = 0; i < MAX_NUMNODES; ++i) {
307 struct page *page, *next;
308 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
309 if (PageHighMem(page))
310 continue;
311 list_del(&page->lru);
312 update_and_free_page(page);
313 nid = page_zone(page)->zone_pgdat->node_id;
314 free_huge_pages--;
315 free_huge_pages_node[nid]--;
316 if (count >= nr_huge_pages)
317 return;
318 }
319 }
320}
321#else
322static inline void try_to_free_low(unsigned long count)
323{
324}
325#endif
326
327static unsigned long set_max_huge_pages(unsigned long count)
328{
329 while (count > nr_huge_pages) {
Nick Piggina4822892006-03-22 00:08:08 -0800330 if (!alloc_fresh_huge_page())
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331 return nr_huge_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700332 }
333 if (count >= nr_huge_pages)
334 return nr_huge_pages;
335
336 spin_lock(&hugetlb_lock);
Chen, Kenneth W78c997a2006-03-31 02:30:01 -0800337 count = max(count, reserved_huge_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700338 try_to_free_low(count);
339 while (count < nr_huge_pages) {
Christoph Lameter5da7ca82006-01-06 00:10:46 -0800340 struct page *page = dequeue_huge_page(NULL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341 if (!page)
342 break;
343 update_and_free_page(page);
344 }
345 spin_unlock(&hugetlb_lock);
346 return nr_huge_pages;
347}
348
349int hugetlb_sysctl_handler(struct ctl_table *table, int write,
350 struct file *file, void __user *buffer,
351 size_t *length, loff_t *ppos)
352{
353 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
354 max_huge_pages = set_max_huge_pages(max_huge_pages);
355 return 0;
356}
357#endif /* CONFIG_SYSCTL */
358
359int hugetlb_report_meminfo(char *buf)
360{
361 return sprintf(buf,
362 "HugePages_Total: %5lu\n"
363 "HugePages_Free: %5lu\n"
David Gibsonb45b5bd2006-03-22 00:08:55 -0800364 "HugePages_Rsvd: %5lu\n"
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365 "Hugepagesize: %5lu kB\n",
366 nr_huge_pages,
367 free_huge_pages,
David Gibsonb45b5bd2006-03-22 00:08:55 -0800368 reserved_huge_pages,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369 HPAGE_SIZE/1024);
370}
371
372int hugetlb_report_node_meminfo(int nid, char *buf)
373{
374 return sprintf(buf,
375 "Node %d HugePages_Total: %5u\n"
376 "Node %d HugePages_Free: %5u\n",
377 nid, nr_huge_pages_node[nid],
378 nid, free_huge_pages_node[nid]);
379}
380
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
382unsigned long hugetlb_total_pages(void)
383{
384 return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
385}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386
387/*
388 * We cannot handle pagefaults against hugetlb pages at all. They cause
389 * handle_mm_fault() to try to instantiate regular-sized pages in the
390 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
391 * this far.
392 */
393static struct page *hugetlb_nopage(struct vm_area_struct *vma,
394 unsigned long address, int *unused)
395{
396 BUG();
397 return NULL;
398}
399
400struct vm_operations_struct hugetlb_vm_ops = {
401 .nopage = hugetlb_nopage,
402};
403
David Gibson1e8f8892006-01-06 00:10:44 -0800404static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
405 int writable)
David Gibson63551ae2005-06-21 17:14:44 -0700406{
407 pte_t entry;
408
David Gibson1e8f8892006-01-06 00:10:44 -0800409 if (writable) {
David Gibson63551ae2005-06-21 17:14:44 -0700410 entry =
411 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
412 } else {
413 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
414 }
415 entry = pte_mkyoung(entry);
416 entry = pte_mkhuge(entry);
417
418 return entry;
419}
420
David Gibson1e8f8892006-01-06 00:10:44 -0800421static void set_huge_ptep_writable(struct vm_area_struct *vma,
422 unsigned long address, pte_t *ptep)
423{
424 pte_t entry;
425
426 entry = pte_mkwrite(pte_mkdirty(*ptep));
427 ptep_set_access_flags(vma, address, ptep, entry, 1);
428 update_mmu_cache(vma, address, entry);
429 lazy_mmu_prot_update(entry);
430}
431
432
David Gibson63551ae2005-06-21 17:14:44 -0700433int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
434 struct vm_area_struct *vma)
435{
436 pte_t *src_pte, *dst_pte, entry;
437 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -0700438 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -0800439 int cow;
440
441 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -0700442
Hugh Dickins1c598272005-10-19 21:23:43 -0700443 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
Hugh Dickinsc74df322005-10-29 18:16:23 -0700444 src_pte = huge_pte_offset(src, addr);
445 if (!src_pte)
446 continue;
David Gibson63551ae2005-06-21 17:14:44 -0700447 dst_pte = huge_pte_alloc(dst, addr);
448 if (!dst_pte)
449 goto nomem;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700450 spin_lock(&dst->page_table_lock);
Hugh Dickins1c598272005-10-19 21:23:43 -0700451 spin_lock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700452 if (!pte_none(*src_pte)) {
David Gibson1e8f8892006-01-06 00:10:44 -0800453 if (cow)
454 ptep_set_wrprotect(src, addr, src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -0700455 entry = *src_pte;
456 ptepage = pte_page(entry);
457 get_page(ptepage);
Hugh Dickins42946212005-10-29 18:16:05 -0700458 add_mm_counter(dst, file_rss, HPAGE_SIZE / PAGE_SIZE);
Hugh Dickins1c598272005-10-19 21:23:43 -0700459 set_huge_pte_at(dst, addr, dst_pte, entry);
460 }
461 spin_unlock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700462 spin_unlock(&dst->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700463 }
464 return 0;
465
466nomem:
467 return -ENOMEM;
468}
469
470void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
471 unsigned long end)
472{
473 struct mm_struct *mm = vma->vm_mm;
474 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -0700475 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -0700476 pte_t pte;
477 struct page *page;
478
479 WARN_ON(!is_vm_hugetlb_page(vma));
480 BUG_ON(start & ~HPAGE_MASK);
481 BUG_ON(end & ~HPAGE_MASK);
482
Hugh Dickins508034a2005-10-29 18:16:30 -0700483 spin_lock(&mm->page_table_lock);
484
Hugh Dickins365e9c872005-10-29 18:16:18 -0700485 /* Update high watermark before we lower rss */
486 update_hiwater_rss(mm);
487
David Gibson63551ae2005-06-21 17:14:44 -0700488 for (address = start; address < end; address += HPAGE_SIZE) {
David Gibsonc7546f82005-08-05 11:59:35 -0700489 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -0700490 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -0700491 continue;
492
493 pte = huge_ptep_get_and_clear(mm, address, ptep);
David Gibson63551ae2005-06-21 17:14:44 -0700494 if (pte_none(pte))
495 continue;
David Gibsonc7546f82005-08-05 11:59:35 -0700496
David Gibson63551ae2005-06-21 17:14:44 -0700497 page = pte_page(pte);
498 put_page(page);
Hugh Dickins42946212005-10-29 18:16:05 -0700499 add_mm_counter(mm, file_rss, (int) -(HPAGE_SIZE / PAGE_SIZE));
David Gibson63551ae2005-06-21 17:14:44 -0700500 }
David Gibson63551ae2005-06-21 17:14:44 -0700501
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 spin_unlock(&mm->page_table_lock);
Hugh Dickins508034a2005-10-29 18:16:30 -0700503 flush_tlb_range(vma, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504}
David Gibson63551ae2005-06-21 17:14:44 -0700505
David Gibson1e8f8892006-01-06 00:10:44 -0800506static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
507 unsigned long address, pte_t *ptep, pte_t pte)
508{
509 struct page *old_page, *new_page;
David Gibson79ac6ba2006-03-22 00:08:51 -0800510 int avoidcopy;
David Gibson1e8f8892006-01-06 00:10:44 -0800511
512 old_page = pte_page(pte);
513
514 /* If no-one else is actually using this page, avoid the copy
515 * and just make the page writable */
516 avoidcopy = (page_count(old_page) == 1);
517 if (avoidcopy) {
518 set_huge_ptep_writable(vma, address, ptep);
519 return VM_FAULT_MINOR;
520 }
521
522 page_cache_get(old_page);
Christoph Lameter5da7ca82006-01-06 00:10:46 -0800523 new_page = alloc_huge_page(vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -0800524
525 if (!new_page) {
526 page_cache_release(old_page);
Christoph Lameter0df420d2006-02-07 12:58:30 -0800527 return VM_FAULT_OOM;
David Gibson1e8f8892006-01-06 00:10:44 -0800528 }
529
530 spin_unlock(&mm->page_table_lock);
David Gibson79ac6ba2006-03-22 00:08:51 -0800531 copy_huge_page(new_page, old_page, address);
David Gibson1e8f8892006-01-06 00:10:44 -0800532 spin_lock(&mm->page_table_lock);
533
534 ptep = huge_pte_offset(mm, address & HPAGE_MASK);
535 if (likely(pte_same(*ptep, pte))) {
536 /* Break COW */
537 set_huge_pte_at(mm, address, ptep,
538 make_huge_pte(vma, new_page, 1));
539 /* Make the old page be freed below */
540 new_page = old_page;
541 }
542 page_cache_release(new_page);
543 page_cache_release(old_page);
544 return VM_FAULT_MINOR;
545}
546
Adam Litke86e52162006-01-06 00:10:43 -0800547int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
David Gibson1e8f8892006-01-06 00:10:44 -0800548 unsigned long address, pte_t *ptep, int write_access)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100549{
550 int ret = VM_FAULT_SIGBUS;
Adam Litke4c887262005-10-29 18:16:46 -0700551 unsigned long idx;
552 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -0700553 struct page *page;
554 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -0800555 pte_t new_pte;
Adam Litke4c887262005-10-29 18:16:46 -0700556
Adam Litke4c887262005-10-29 18:16:46 -0700557 mapping = vma->vm_file->f_mapping;
558 idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
559 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
560
561 /*
562 * Use page lock to guard against racing truncation
563 * before we get page_table_lock.
564 */
Christoph Lameter6bda6662006-01-06 00:10:49 -0800565retry:
566 page = find_lock_page(mapping, idx);
567 if (!page) {
568 if (hugetlb_get_quota(mapping))
569 goto out;
570 page = alloc_huge_page(vma, address);
571 if (!page) {
572 hugetlb_put_quota(mapping);
Christoph Lameter0df420d2006-02-07 12:58:30 -0800573 ret = VM_FAULT_OOM;
Christoph Lameter6bda6662006-01-06 00:10:49 -0800574 goto out;
575 }
David Gibson79ac6ba2006-03-22 00:08:51 -0800576 clear_huge_page(page, address);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100577
Christoph Lameter6bda6662006-01-06 00:10:49 -0800578 if (vma->vm_flags & VM_SHARED) {
579 int err;
580
581 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
582 if (err) {
583 put_page(page);
584 hugetlb_put_quota(mapping);
585 if (err == -EEXIST)
586 goto retry;
587 goto out;
588 }
589 } else
590 lock_page(page);
591 }
David Gibson1e8f8892006-01-06 00:10:44 -0800592
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100593 spin_lock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -0700594 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
595 if (idx >= size)
596 goto backout;
597
598 ret = VM_FAULT_MINOR;
Adam Litke86e52162006-01-06 00:10:43 -0800599 if (!pte_none(*ptep))
Adam Litke4c887262005-10-29 18:16:46 -0700600 goto backout;
601
602 add_mm_counter(mm, file_rss, HPAGE_SIZE / PAGE_SIZE);
David Gibson1e8f8892006-01-06 00:10:44 -0800603 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
604 && (vma->vm_flags & VM_SHARED)));
605 set_huge_pte_at(mm, address, ptep, new_pte);
606
607 if (write_access && !(vma->vm_flags & VM_SHARED)) {
608 /* Optimization, do the COW without a second fault */
609 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
610 }
611
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100612 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -0700613 unlock_page(page);
614out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100615 return ret;
Adam Litke4c887262005-10-29 18:16:46 -0700616
617backout:
618 spin_unlock(&mm->page_table_lock);
619 hugetlb_put_quota(mapping);
620 unlock_page(page);
621 put_page(page);
622 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100623}
624
Adam Litke86e52162006-01-06 00:10:43 -0800625int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
626 unsigned long address, int write_access)
627{
628 pte_t *ptep;
629 pte_t entry;
David Gibson1e8f8892006-01-06 00:10:44 -0800630 int ret;
David Gibson3935baa2006-03-22 00:08:53 -0800631 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Adam Litke86e52162006-01-06 00:10:43 -0800632
633 ptep = huge_pte_alloc(mm, address);
634 if (!ptep)
635 return VM_FAULT_OOM;
636
David Gibson3935baa2006-03-22 00:08:53 -0800637 /*
638 * Serialize hugepage allocation and instantiation, so that we don't
639 * get spurious allocation failures if two CPUs race to instantiate
640 * the same page in the page cache.
641 */
642 mutex_lock(&hugetlb_instantiation_mutex);
Adam Litke86e52162006-01-06 00:10:43 -0800643 entry = *ptep;
David Gibson3935baa2006-03-22 00:08:53 -0800644 if (pte_none(entry)) {
645 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
646 mutex_unlock(&hugetlb_instantiation_mutex);
647 return ret;
648 }
Adam Litke86e52162006-01-06 00:10:43 -0800649
David Gibson1e8f8892006-01-06 00:10:44 -0800650 ret = VM_FAULT_MINOR;
651
652 spin_lock(&mm->page_table_lock);
653 /* Check for a racing update before calling hugetlb_cow */
654 if (likely(pte_same(entry, *ptep)))
655 if (write_access && !pte_write(entry))
656 ret = hugetlb_cow(mm, vma, address, ptep, entry);
657 spin_unlock(&mm->page_table_lock);
David Gibson3935baa2006-03-22 00:08:53 -0800658 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -0800659
660 return ret;
Adam Litke86e52162006-01-06 00:10:43 -0800661}
662
David Gibson63551ae2005-06-21 17:14:44 -0700663int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
664 struct page **pages, struct vm_area_struct **vmas,
665 unsigned long *position, int *length, int i)
666{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800667 unsigned long pfn_offset;
668 unsigned long vaddr = *position;
David Gibson63551ae2005-06-21 17:14:44 -0700669 int remainder = *length;
670
Hugh Dickins1c598272005-10-19 21:23:43 -0700671 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700672 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -0700673 pte_t *pte;
674 struct page *page;
675
676 /*
677 * Some archs (sparc64, sh*) have multiple pte_ts to
678 * each hugepage. We have to make * sure we get the
679 * first, for the page indexing below to work.
680 */
681 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
682
683 if (!pte || pte_none(*pte)) {
684 int ret;
685
686 spin_unlock(&mm->page_table_lock);
687 ret = hugetlb_fault(mm, vma, vaddr, 0);
688 spin_lock(&mm->page_table_lock);
689 if (ret == VM_FAULT_MINOR)
690 continue;
691
692 remainder = 0;
693 if (!i)
694 i = -EFAULT;
695 break;
696 }
David Gibson63551ae2005-06-21 17:14:44 -0700697
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800698 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
699 page = pte_page(*pte);
700same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -0800701 if (pages) {
702 get_page(page);
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800703 pages[i] = page + pfn_offset;
Chen, Kenneth Wd6692182006-03-31 02:29:57 -0800704 }
David Gibson63551ae2005-06-21 17:14:44 -0700705
706 if (vmas)
707 vmas[i] = vma;
708
709 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800710 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -0700711 --remainder;
712 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800713 if (vaddr < vma->vm_end && remainder &&
714 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
715 /*
716 * We use pfn_offset to avoid touching the pageframes
717 * of this compound page.
718 */
719 goto same_page;
720 }
David Gibson63551ae2005-06-21 17:14:44 -0700721 }
Hugh Dickins1c598272005-10-19 21:23:43 -0700722 spin_unlock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700723 *length = remainder;
724 *position = vaddr;
725
726 return i;
727}
Zhang, Yanmin8f860592006-03-22 00:08:50 -0800728
729void hugetlb_change_protection(struct vm_area_struct *vma,
730 unsigned long address, unsigned long end, pgprot_t newprot)
731{
732 struct mm_struct *mm = vma->vm_mm;
733 unsigned long start = address;
734 pte_t *ptep;
735 pte_t pte;
736
737 BUG_ON(address >= end);
738 flush_cache_range(vma, address, end);
739
740 spin_lock(&mm->page_table_lock);
741 for (; address < end; address += HPAGE_SIZE) {
742 ptep = huge_pte_offset(mm, address);
743 if (!ptep)
744 continue;
745 if (!pte_none(*ptep)) {
746 pte = huge_ptep_get_and_clear(mm, address, ptep);
747 pte = pte_mkhuge(pte_modify(pte, newprot));
748 set_huge_pte_at(mm, address, ptep, pte);
749 lazy_mmu_prot_update(pte);
750 }
751 }
752 spin_unlock(&mm->page_table_lock);
753
754 flush_tlb_range(vma, start, end);
755}
756