Commit | Line | Data |
---|---|---|
1da177e4 LT |
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> | |
1da177e4 LT |
10 | #include <linux/sysctl.h> |
11 | #include <linux/highmem.h> | |
12 | #include <linux/nodemask.h> | |
63551ae0 | 13 | #include <linux/pagemap.h> |
5da7ca86 | 14 | #include <linux/mempolicy.h> |
aea47ff3 | 15 | #include <linux/cpuset.h> |
3935baa9 | 16 | #include <linux/mutex.h> |
5da7ca86 | 17 | |
63551ae0 DG |
18 | #include <asm/page.h> |
19 | #include <asm/pgtable.h> | |
20 | ||
21 | #include <linux/hugetlb.h> | |
7835e98b | 22 | #include "internal.h" |
1da177e4 LT |
23 | |
24 | const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL; | |
b45b5bd6 | 25 | static unsigned long nr_huge_pages, free_huge_pages, reserved_huge_pages; |
1da177e4 LT |
26 | unsigned long max_huge_pages; |
27 | static struct list_head hugepage_freelists[MAX_NUMNODES]; | |
28 | static unsigned int nr_huge_pages_node[MAX_NUMNODES]; | |
29 | static unsigned int free_huge_pages_node[MAX_NUMNODES]; | |
3935baa9 DG |
30 | /* |
31 | * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages | |
32 | */ | |
33 | static DEFINE_SPINLOCK(hugetlb_lock); | |
0bd0f9fb | 34 | |
79ac6ba4 DG |
35 | static 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 | ||
46 | static 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 | ||
1da177e4 LT |
58 | static 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 | ||
5da7ca86 CL |
66 | static struct page *dequeue_huge_page(struct vm_area_struct *vma, |
67 | unsigned long address) | |
1da177e4 LT |
68 | { |
69 | int nid = numa_node_id(); | |
70 | struct page *page = NULL; | |
5da7ca86 | 71 | struct zonelist *zonelist = huge_zonelist(vma, address); |
96df9333 | 72 | struct zone **z; |
1da177e4 | 73 | |
96df9333 CL |
74 | for (z = zonelist->zones; *z; z++) { |
75 | nid = (*z)->zone_pgdat->node_id; | |
aea47ff3 CL |
76 | if (cpuset_zone_allowed(*z, GFP_HIGHUSER) && |
77 | !list_empty(&hugepage_freelists[nid])) | |
96df9333 | 78 | break; |
1da177e4 | 79 | } |
96df9333 CL |
80 | |
81 | if (*z) { | |
1da177e4 LT |
82 | 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 | ||
27a85ef1 DG |
91 | static 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 | ||
a482289d | 102 | static int alloc_fresh_huge_page(void) |
1da177e4 LT |
103 | { |
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); | |
fdb7cc59 PJ |
108 | nid = next_node(nid, node_online_map); |
109 | if (nid == MAX_NUMNODES) | |
110 | nid = first_node(node_online_map); | |
1da177e4 | 111 | if (page) { |
a482289d | 112 | page[1].lru.next = (void *)free_huge_page; /* dtor */ |
0bd0f9fb | 113 | spin_lock(&hugetlb_lock); |
1da177e4 LT |
114 | nr_huge_pages++; |
115 | nr_huge_pages_node[page_to_nid(page)]++; | |
0bd0f9fb | 116 | spin_unlock(&hugetlb_lock); |
a482289d NP |
117 | put_page(page); /* free it into the hugepage allocator */ |
118 | return 1; | |
1da177e4 | 119 | } |
a482289d | 120 | return 0; |
1da177e4 LT |
121 | } |
122 | ||
27a85ef1 DG |
123 | static struct page *alloc_huge_page(struct vm_area_struct *vma, |
124 | unsigned long addr) | |
1da177e4 | 125 | { |
b45b5bd6 | 126 | struct inode *inode = vma->vm_file->f_dentry->d_inode; |
1da177e4 | 127 | struct page *page; |
b45b5bd6 DG |
128 | int use_reserve = 0; |
129 | unsigned long idx; | |
1da177e4 LT |
130 | |
131 | spin_lock(&hugetlb_lock); | |
b45b5bd6 DG |
132 | |
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; | |
1da177e4 | 150 | } |
b45b5bd6 DG |
151 | |
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 | ||
1da177e4 | 164 | spin_unlock(&hugetlb_lock); |
7835e98b | 165 | set_page_refcounted(page); |
1da177e4 | 166 | return page; |
b45b5bd6 DG |
167 | |
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 | */ | |
183 | int 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 | */ | |
223 | void 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); | |
1da177e4 LT |
256 | } |
257 | ||
258 | static int __init hugetlb_init(void) | |
259 | { | |
260 | unsigned long i; | |
1da177e4 | 261 | |
3c726f8d BH |
262 | if (HPAGE_SHIFT == 0) |
263 | return 0; | |
264 | ||
1da177e4 LT |
265 | for (i = 0; i < MAX_NUMNODES; ++i) |
266 | INIT_LIST_HEAD(&hugepage_freelists[i]); | |
267 | ||
268 | for (i = 0; i < max_huge_pages; ++i) { | |
a482289d | 269 | if (!alloc_fresh_huge_page()) |
1da177e4 | 270 | break; |
1da177e4 LT |
271 | } |
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 | } | |
276 | module_init(hugetlb_init); | |
277 | ||
278 | static 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 | |
287 | static 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); | |
1da177e4 | 296 | } |
a482289d | 297 | page[1].lru.next = NULL; |
7835e98b | 298 | set_page_refcounted(page); |
1da177e4 LT |
299 | __free_pages(page, HUGETLB_PAGE_ORDER); |
300 | } | |
301 | ||
302 | #ifdef CONFIG_HIGHMEM | |
303 | static 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 | |
322 | static inline void try_to_free_low(unsigned long count) | |
323 | { | |
324 | } | |
325 | #endif | |
326 | ||
327 | static unsigned long set_max_huge_pages(unsigned long count) | |
328 | { | |
329 | while (count > nr_huge_pages) { | |
a482289d | 330 | if (!alloc_fresh_huge_page()) |
1da177e4 | 331 | return nr_huge_pages; |
1da177e4 LT |
332 | } |
333 | if (count >= nr_huge_pages) | |
334 | return nr_huge_pages; | |
335 | ||
336 | spin_lock(&hugetlb_lock); | |
337 | try_to_free_low(count); | |
338 | while (count < nr_huge_pages) { | |
5da7ca86 | 339 | struct page *page = dequeue_huge_page(NULL, 0); |
1da177e4 LT |
340 | if (!page) |
341 | break; | |
342 | update_and_free_page(page); | |
343 | } | |
344 | spin_unlock(&hugetlb_lock); | |
345 | return nr_huge_pages; | |
346 | } | |
347 | ||
348 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, | |
349 | struct file *file, void __user *buffer, | |
350 | size_t *length, loff_t *ppos) | |
351 | { | |
352 | proc_doulongvec_minmax(table, write, file, buffer, length, ppos); | |
353 | max_huge_pages = set_max_huge_pages(max_huge_pages); | |
354 | return 0; | |
355 | } | |
356 | #endif /* CONFIG_SYSCTL */ | |
357 | ||
358 | int hugetlb_report_meminfo(char *buf) | |
359 | { | |
360 | return sprintf(buf, | |
361 | "HugePages_Total: %5lu\n" | |
362 | "HugePages_Free: %5lu\n" | |
b45b5bd6 | 363 | "HugePages_Rsvd: %5lu\n" |
1da177e4 LT |
364 | "Hugepagesize: %5lu kB\n", |
365 | nr_huge_pages, | |
366 | free_huge_pages, | |
b45b5bd6 | 367 | reserved_huge_pages, |
1da177e4 LT |
368 | HPAGE_SIZE/1024); |
369 | } | |
370 | ||
371 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
372 | { | |
373 | return sprintf(buf, | |
374 | "Node %d HugePages_Total: %5u\n" | |
375 | "Node %d HugePages_Free: %5u\n", | |
376 | nid, nr_huge_pages_node[nid], | |
377 | nid, free_huge_pages_node[nid]); | |
378 | } | |
379 | ||
1da177e4 LT |
380 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
381 | unsigned long hugetlb_total_pages(void) | |
382 | { | |
383 | return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE); | |
384 | } | |
1da177e4 LT |
385 | |
386 | /* | |
387 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
388 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
389 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
390 | * this far. | |
391 | */ | |
392 | static struct page *hugetlb_nopage(struct vm_area_struct *vma, | |
393 | unsigned long address, int *unused) | |
394 | { | |
395 | BUG(); | |
396 | return NULL; | |
397 | } | |
398 | ||
399 | struct vm_operations_struct hugetlb_vm_ops = { | |
400 | .nopage = hugetlb_nopage, | |
401 | }; | |
402 | ||
1e8f889b DG |
403 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
404 | int writable) | |
63551ae0 DG |
405 | { |
406 | pte_t entry; | |
407 | ||
1e8f889b | 408 | if (writable) { |
63551ae0 DG |
409 | entry = |
410 | pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); | |
411 | } else { | |
412 | entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot)); | |
413 | } | |
414 | entry = pte_mkyoung(entry); | |
415 | entry = pte_mkhuge(entry); | |
416 | ||
417 | return entry; | |
418 | } | |
419 | ||
1e8f889b DG |
420 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
421 | unsigned long address, pte_t *ptep) | |
422 | { | |
423 | pte_t entry; | |
424 | ||
425 | entry = pte_mkwrite(pte_mkdirty(*ptep)); | |
426 | ptep_set_access_flags(vma, address, ptep, entry, 1); | |
427 | update_mmu_cache(vma, address, entry); | |
428 | lazy_mmu_prot_update(entry); | |
429 | } | |
430 | ||
431 | ||
63551ae0 DG |
432 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
433 | struct vm_area_struct *vma) | |
434 | { | |
435 | pte_t *src_pte, *dst_pte, entry; | |
436 | struct page *ptepage; | |
1c59827d | 437 | unsigned long addr; |
1e8f889b DG |
438 | int cow; |
439 | ||
440 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 441 | |
1c59827d | 442 | for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) { |
c74df32c HD |
443 | src_pte = huge_pte_offset(src, addr); |
444 | if (!src_pte) | |
445 | continue; | |
63551ae0 DG |
446 | dst_pte = huge_pte_alloc(dst, addr); |
447 | if (!dst_pte) | |
448 | goto nomem; | |
c74df32c | 449 | spin_lock(&dst->page_table_lock); |
1c59827d | 450 | spin_lock(&src->page_table_lock); |
c74df32c | 451 | if (!pte_none(*src_pte)) { |
1e8f889b DG |
452 | if (cow) |
453 | ptep_set_wrprotect(src, addr, src_pte); | |
1c59827d HD |
454 | entry = *src_pte; |
455 | ptepage = pte_page(entry); | |
456 | get_page(ptepage); | |
4294621f | 457 | add_mm_counter(dst, file_rss, HPAGE_SIZE / PAGE_SIZE); |
1c59827d HD |
458 | set_huge_pte_at(dst, addr, dst_pte, entry); |
459 | } | |
460 | spin_unlock(&src->page_table_lock); | |
c74df32c | 461 | spin_unlock(&dst->page_table_lock); |
63551ae0 DG |
462 | } |
463 | return 0; | |
464 | ||
465 | nomem: | |
466 | return -ENOMEM; | |
467 | } | |
468 | ||
469 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, | |
470 | unsigned long end) | |
471 | { | |
472 | struct mm_struct *mm = vma->vm_mm; | |
473 | unsigned long address; | |
c7546f8f | 474 | pte_t *ptep; |
63551ae0 DG |
475 | pte_t pte; |
476 | struct page *page; | |
477 | ||
478 | WARN_ON(!is_vm_hugetlb_page(vma)); | |
479 | BUG_ON(start & ~HPAGE_MASK); | |
480 | BUG_ON(end & ~HPAGE_MASK); | |
481 | ||
508034a3 HD |
482 | spin_lock(&mm->page_table_lock); |
483 | ||
365e9c87 HD |
484 | /* Update high watermark before we lower rss */ |
485 | update_hiwater_rss(mm); | |
486 | ||
63551ae0 | 487 | for (address = start; address < end; address += HPAGE_SIZE) { |
c7546f8f | 488 | ptep = huge_pte_offset(mm, address); |
4c887265 | 489 | if (!ptep) |
c7546f8f DG |
490 | continue; |
491 | ||
492 | pte = huge_ptep_get_and_clear(mm, address, ptep); | |
63551ae0 DG |
493 | if (pte_none(pte)) |
494 | continue; | |
c7546f8f | 495 | |
63551ae0 DG |
496 | page = pte_page(pte); |
497 | put_page(page); | |
4294621f | 498 | add_mm_counter(mm, file_rss, (int) -(HPAGE_SIZE / PAGE_SIZE)); |
63551ae0 | 499 | } |
63551ae0 | 500 | |
1da177e4 | 501 | spin_unlock(&mm->page_table_lock); |
508034a3 | 502 | flush_tlb_range(vma, start, end); |
1da177e4 | 503 | } |
63551ae0 | 504 | |
1e8f889b DG |
505 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
506 | unsigned long address, pte_t *ptep, pte_t pte) | |
507 | { | |
508 | struct page *old_page, *new_page; | |
79ac6ba4 | 509 | int avoidcopy; |
1e8f889b DG |
510 | |
511 | old_page = pte_page(pte); | |
512 | ||
513 | /* If no-one else is actually using this page, avoid the copy | |
514 | * and just make the page writable */ | |
515 | avoidcopy = (page_count(old_page) == 1); | |
516 | if (avoidcopy) { | |
517 | set_huge_ptep_writable(vma, address, ptep); | |
518 | return VM_FAULT_MINOR; | |
519 | } | |
520 | ||
521 | page_cache_get(old_page); | |
5da7ca86 | 522 | new_page = alloc_huge_page(vma, address); |
1e8f889b DG |
523 | |
524 | if (!new_page) { | |
525 | page_cache_release(old_page); | |
0df420d8 | 526 | return VM_FAULT_OOM; |
1e8f889b DG |
527 | } |
528 | ||
529 | spin_unlock(&mm->page_table_lock); | |
79ac6ba4 | 530 | copy_huge_page(new_page, old_page, address); |
1e8f889b DG |
531 | spin_lock(&mm->page_table_lock); |
532 | ||
533 | ptep = huge_pte_offset(mm, address & HPAGE_MASK); | |
534 | if (likely(pte_same(*ptep, pte))) { | |
535 | /* Break COW */ | |
536 | set_huge_pte_at(mm, address, ptep, | |
537 | make_huge_pte(vma, new_page, 1)); | |
538 | /* Make the old page be freed below */ | |
539 | new_page = old_page; | |
540 | } | |
541 | page_cache_release(new_page); | |
542 | page_cache_release(old_page); | |
543 | return VM_FAULT_MINOR; | |
544 | } | |
545 | ||
86e5216f | 546 | int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
1e8f889b | 547 | unsigned long address, pte_t *ptep, int write_access) |
ac9b9c66 HD |
548 | { |
549 | int ret = VM_FAULT_SIGBUS; | |
4c887265 AL |
550 | unsigned long idx; |
551 | unsigned long size; | |
4c887265 AL |
552 | struct page *page; |
553 | struct address_space *mapping; | |
1e8f889b | 554 | pte_t new_pte; |
4c887265 | 555 | |
4c887265 AL |
556 | mapping = vma->vm_file->f_mapping; |
557 | idx = ((address - vma->vm_start) >> HPAGE_SHIFT) | |
558 | + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT)); | |
559 | ||
560 | /* | |
561 | * Use page lock to guard against racing truncation | |
562 | * before we get page_table_lock. | |
563 | */ | |
6bda666a CL |
564 | retry: |
565 | page = find_lock_page(mapping, idx); | |
566 | if (!page) { | |
567 | if (hugetlb_get_quota(mapping)) | |
568 | goto out; | |
569 | page = alloc_huge_page(vma, address); | |
570 | if (!page) { | |
571 | hugetlb_put_quota(mapping); | |
0df420d8 | 572 | ret = VM_FAULT_OOM; |
6bda666a CL |
573 | goto out; |
574 | } | |
79ac6ba4 | 575 | clear_huge_page(page, address); |
ac9b9c66 | 576 | |
6bda666a CL |
577 | if (vma->vm_flags & VM_SHARED) { |
578 | int err; | |
579 | ||
580 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
581 | if (err) { | |
582 | put_page(page); | |
583 | hugetlb_put_quota(mapping); | |
584 | if (err == -EEXIST) | |
585 | goto retry; | |
586 | goto out; | |
587 | } | |
588 | } else | |
589 | lock_page(page); | |
590 | } | |
1e8f889b | 591 | |
ac9b9c66 | 592 | spin_lock(&mm->page_table_lock); |
4c887265 AL |
593 | size = i_size_read(mapping->host) >> HPAGE_SHIFT; |
594 | if (idx >= size) | |
595 | goto backout; | |
596 | ||
597 | ret = VM_FAULT_MINOR; | |
86e5216f | 598 | if (!pte_none(*ptep)) |
4c887265 AL |
599 | goto backout; |
600 | ||
601 | add_mm_counter(mm, file_rss, HPAGE_SIZE / PAGE_SIZE); | |
1e8f889b DG |
602 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
603 | && (vma->vm_flags & VM_SHARED))); | |
604 | set_huge_pte_at(mm, address, ptep, new_pte); | |
605 | ||
606 | if (write_access && !(vma->vm_flags & VM_SHARED)) { | |
607 | /* Optimization, do the COW without a second fault */ | |
608 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte); | |
609 | } | |
610 | ||
ac9b9c66 | 611 | spin_unlock(&mm->page_table_lock); |
4c887265 AL |
612 | unlock_page(page); |
613 | out: | |
ac9b9c66 | 614 | return ret; |
4c887265 AL |
615 | |
616 | backout: | |
617 | spin_unlock(&mm->page_table_lock); | |
618 | hugetlb_put_quota(mapping); | |
619 | unlock_page(page); | |
620 | put_page(page); | |
621 | goto out; | |
ac9b9c66 HD |
622 | } |
623 | ||
86e5216f AL |
624 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
625 | unsigned long address, int write_access) | |
626 | { | |
627 | pte_t *ptep; | |
628 | pte_t entry; | |
1e8f889b | 629 | int ret; |
3935baa9 | 630 | static DEFINE_MUTEX(hugetlb_instantiation_mutex); |
86e5216f AL |
631 | |
632 | ptep = huge_pte_alloc(mm, address); | |
633 | if (!ptep) | |
634 | return VM_FAULT_OOM; | |
635 | ||
3935baa9 DG |
636 | /* |
637 | * Serialize hugepage allocation and instantiation, so that we don't | |
638 | * get spurious allocation failures if two CPUs race to instantiate | |
639 | * the same page in the page cache. | |
640 | */ | |
641 | mutex_lock(&hugetlb_instantiation_mutex); | |
86e5216f | 642 | entry = *ptep; |
3935baa9 DG |
643 | if (pte_none(entry)) { |
644 | ret = hugetlb_no_page(mm, vma, address, ptep, write_access); | |
645 | mutex_unlock(&hugetlb_instantiation_mutex); | |
646 | return ret; | |
647 | } | |
86e5216f | 648 | |
1e8f889b DG |
649 | ret = VM_FAULT_MINOR; |
650 | ||
651 | spin_lock(&mm->page_table_lock); | |
652 | /* Check for a racing update before calling hugetlb_cow */ | |
653 | if (likely(pte_same(entry, *ptep))) | |
654 | if (write_access && !pte_write(entry)) | |
655 | ret = hugetlb_cow(mm, vma, address, ptep, entry); | |
656 | spin_unlock(&mm->page_table_lock); | |
3935baa9 | 657 | mutex_unlock(&hugetlb_instantiation_mutex); |
1e8f889b DG |
658 | |
659 | return ret; | |
86e5216f AL |
660 | } |
661 | ||
63551ae0 DG |
662 | int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
663 | struct page **pages, struct vm_area_struct **vmas, | |
664 | unsigned long *position, int *length, int i) | |
665 | { | |
d5d4b0aa CK |
666 | unsigned long pfn_offset; |
667 | unsigned long vaddr = *position; | |
63551ae0 DG |
668 | int remainder = *length; |
669 | ||
1c59827d | 670 | spin_lock(&mm->page_table_lock); |
63551ae0 | 671 | while (vaddr < vma->vm_end && remainder) { |
4c887265 AL |
672 | pte_t *pte; |
673 | struct page *page; | |
63551ae0 | 674 | |
4c887265 AL |
675 | /* |
676 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
677 | * each hugepage. We have to make * sure we get the | |
678 | * first, for the page indexing below to work. | |
679 | */ | |
680 | pte = huge_pte_offset(mm, vaddr & HPAGE_MASK); | |
63551ae0 | 681 | |
4c887265 AL |
682 | if (!pte || pte_none(*pte)) { |
683 | int ret; | |
63551ae0 | 684 | |
4c887265 AL |
685 | spin_unlock(&mm->page_table_lock); |
686 | ret = hugetlb_fault(mm, vma, vaddr, 0); | |
687 | spin_lock(&mm->page_table_lock); | |
688 | if (ret == VM_FAULT_MINOR) | |
689 | continue; | |
63551ae0 | 690 | |
4c887265 AL |
691 | remainder = 0; |
692 | if (!i) | |
693 | i = -EFAULT; | |
694 | break; | |
695 | } | |
696 | ||
d5d4b0aa CK |
697 | pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT; |
698 | page = pte_page(*pte); | |
699 | same_page: | |
700 | get_page(page); | |
701 | if (pages) | |
702 | pages[i] = page + pfn_offset; | |
63551ae0 DG |
703 | |
704 | if (vmas) | |
705 | vmas[i] = vma; | |
706 | ||
707 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 708 | ++pfn_offset; |
63551ae0 DG |
709 | --remainder; |
710 | ++i; | |
d5d4b0aa CK |
711 | if (vaddr < vma->vm_end && remainder && |
712 | pfn_offset < HPAGE_SIZE/PAGE_SIZE) { | |
713 | /* | |
714 | * We use pfn_offset to avoid touching the pageframes | |
715 | * of this compound page. | |
716 | */ | |
717 | goto same_page; | |
718 | } | |
63551ae0 | 719 | } |
1c59827d | 720 | spin_unlock(&mm->page_table_lock); |
63551ae0 DG |
721 | *length = remainder; |
722 | *position = vaddr; | |
723 | ||
724 | return i; | |
725 | } | |
8f860591 ZY |
726 | |
727 | void hugetlb_change_protection(struct vm_area_struct *vma, | |
728 | unsigned long address, unsigned long end, pgprot_t newprot) | |
729 | { | |
730 | struct mm_struct *mm = vma->vm_mm; | |
731 | unsigned long start = address; | |
732 | pte_t *ptep; | |
733 | pte_t pte; | |
734 | ||
735 | BUG_ON(address >= end); | |
736 | flush_cache_range(vma, address, end); | |
737 | ||
738 | spin_lock(&mm->page_table_lock); | |
739 | for (; address < end; address += HPAGE_SIZE) { | |
740 | ptep = huge_pte_offset(mm, address); | |
741 | if (!ptep) | |
742 | continue; | |
743 | if (!pte_none(*ptep)) { | |
744 | pte = huge_ptep_get_and_clear(mm, address, ptep); | |
745 | pte = pte_mkhuge(pte_modify(pte, newprot)); | |
746 | set_huge_pte_at(mm, address, ptep, pte); | |
747 | lazy_mmu_prot_update(pte); | |
748 | } | |
749 | } | |
750 | spin_unlock(&mm->page_table_lock); | |
751 | ||
752 | flush_tlb_range(vma, start, end); | |
753 | } | |
754 |