]>
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; | |
a43a8c39 | 25 | static unsigned long nr_huge_pages, free_huge_pages, resv_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]; | |
396faf03 MG |
30 | static gfp_t htlb_alloc_mask = GFP_HIGHUSER; |
31 | unsigned long hugepages_treat_as_movable; | |
32 | ||
3935baa9 DG |
33 | /* |
34 | * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages | |
35 | */ | |
36 | static DEFINE_SPINLOCK(hugetlb_lock); | |
0bd0f9fb | 37 | |
79ac6ba4 DG |
38 | static void clear_huge_page(struct page *page, unsigned long addr) |
39 | { | |
40 | int i; | |
41 | ||
42 | might_sleep(); | |
43 | for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) { | |
44 | cond_resched(); | |
281e0e3b | 45 | clear_user_highpage(page + i, addr + i * PAGE_SIZE); |
79ac6ba4 DG |
46 | } |
47 | } | |
48 | ||
49 | static void copy_huge_page(struct page *dst, struct page *src, | |
9de455b2 | 50 | unsigned long addr, struct vm_area_struct *vma) |
79ac6ba4 DG |
51 | { |
52 | int i; | |
53 | ||
54 | might_sleep(); | |
55 | for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) { | |
56 | cond_resched(); | |
9de455b2 | 57 | copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma); |
79ac6ba4 DG |
58 | } |
59 | } | |
60 | ||
1da177e4 LT |
61 | static void enqueue_huge_page(struct page *page) |
62 | { | |
63 | int nid = page_to_nid(page); | |
64 | list_add(&page->lru, &hugepage_freelists[nid]); | |
65 | free_huge_pages++; | |
66 | free_huge_pages_node[nid]++; | |
67 | } | |
68 | ||
5da7ca86 CL |
69 | static struct page *dequeue_huge_page(struct vm_area_struct *vma, |
70 | unsigned long address) | |
1da177e4 | 71 | { |
31a5c6e4 | 72 | int nid; |
1da177e4 | 73 | struct page *page = NULL; |
480eccf9 | 74 | struct mempolicy *mpol; |
396faf03 | 75 | struct zonelist *zonelist = huge_zonelist(vma, address, |
480eccf9 | 76 | htlb_alloc_mask, &mpol); |
96df9333 | 77 | struct zone **z; |
1da177e4 | 78 | |
96df9333 | 79 | for (z = zonelist->zones; *z; z++) { |
89fa3024 | 80 | nid = zone_to_nid(*z); |
396faf03 | 81 | if (cpuset_zone_allowed_softwall(*z, htlb_alloc_mask) && |
3abf7afd AM |
82 | !list_empty(&hugepage_freelists[nid])) { |
83 | page = list_entry(hugepage_freelists[nid].next, | |
84 | struct page, lru); | |
85 | list_del(&page->lru); | |
86 | free_huge_pages--; | |
87 | free_huge_pages_node[nid]--; | |
5ab3ee7b | 88 | break; |
3abf7afd | 89 | } |
1da177e4 | 90 | } |
480eccf9 | 91 | mpol_free(mpol); /* unref if mpol !NULL */ |
1da177e4 LT |
92 | return page; |
93 | } | |
94 | ||
27a85ef1 DG |
95 | static void free_huge_page(struct page *page) |
96 | { | |
97 | BUG_ON(page_count(page)); | |
98 | ||
99 | INIT_LIST_HEAD(&page->lru); | |
100 | ||
101 | spin_lock(&hugetlb_lock); | |
102 | enqueue_huge_page(page); | |
103 | spin_unlock(&hugetlb_lock); | |
104 | } | |
105 | ||
a482289d | 106 | static int alloc_fresh_huge_page(void) |
1da177e4 | 107 | { |
f96efd58 | 108 | static int prev_nid; |
1da177e4 | 109 | struct page *page; |
f96efd58 JJ |
110 | int nid; |
111 | ||
7ed5cb2b HD |
112 | /* |
113 | * Copy static prev_nid to local nid, work on that, then copy it | |
114 | * back to prev_nid afterwards: otherwise there's a window in which | |
115 | * a racer might pass invalid nid MAX_NUMNODES to alloc_pages_node. | |
116 | * But we don't need to use a spin_lock here: it really doesn't | |
117 | * matter if occasionally a racer chooses the same nid as we do. | |
118 | */ | |
f96efd58 | 119 | nid = next_node(prev_nid, node_online_map); |
fdb7cc59 PJ |
120 | if (nid == MAX_NUMNODES) |
121 | nid = first_node(node_online_map); | |
f96efd58 | 122 | prev_nid = nid; |
f96efd58 | 123 | |
396faf03 | 124 | page = alloc_pages_node(nid, htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN, |
f96efd58 | 125 | HUGETLB_PAGE_ORDER); |
1da177e4 | 126 | if (page) { |
33f2ef89 | 127 | set_compound_page_dtor(page, free_huge_page); |
0bd0f9fb | 128 | spin_lock(&hugetlb_lock); |
1da177e4 LT |
129 | nr_huge_pages++; |
130 | nr_huge_pages_node[page_to_nid(page)]++; | |
0bd0f9fb | 131 | spin_unlock(&hugetlb_lock); |
a482289d NP |
132 | put_page(page); /* free it into the hugepage allocator */ |
133 | return 1; | |
1da177e4 | 134 | } |
a482289d | 135 | return 0; |
1da177e4 LT |
136 | } |
137 | ||
27a85ef1 DG |
138 | static struct page *alloc_huge_page(struct vm_area_struct *vma, |
139 | unsigned long addr) | |
1da177e4 LT |
140 | { |
141 | struct page *page; | |
1da177e4 LT |
142 | |
143 | spin_lock(&hugetlb_lock); | |
a43a8c39 KC |
144 | if (vma->vm_flags & VM_MAYSHARE) |
145 | resv_huge_pages--; | |
146 | else if (free_huge_pages <= resv_huge_pages) | |
147 | goto fail; | |
b45b5bd6 DG |
148 | |
149 | page = dequeue_huge_page(vma, addr); | |
150 | if (!page) | |
151 | goto fail; | |
152 | ||
1da177e4 | 153 | spin_unlock(&hugetlb_lock); |
7835e98b | 154 | set_page_refcounted(page); |
1da177e4 | 155 | return page; |
b45b5bd6 | 156 | |
a43a8c39 | 157 | fail: |
ace4bd29 KC |
158 | if (vma->vm_flags & VM_MAYSHARE) |
159 | resv_huge_pages++; | |
b45b5bd6 DG |
160 | spin_unlock(&hugetlb_lock); |
161 | return NULL; | |
162 | } | |
163 | ||
1da177e4 LT |
164 | static int __init hugetlb_init(void) |
165 | { | |
166 | unsigned long i; | |
1da177e4 | 167 | |
3c726f8d BH |
168 | if (HPAGE_SHIFT == 0) |
169 | return 0; | |
170 | ||
1da177e4 LT |
171 | for (i = 0; i < MAX_NUMNODES; ++i) |
172 | INIT_LIST_HEAD(&hugepage_freelists[i]); | |
173 | ||
174 | for (i = 0; i < max_huge_pages; ++i) { | |
a482289d | 175 | if (!alloc_fresh_huge_page()) |
1da177e4 | 176 | break; |
1da177e4 LT |
177 | } |
178 | max_huge_pages = free_huge_pages = nr_huge_pages = i; | |
179 | printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages); | |
180 | return 0; | |
181 | } | |
182 | module_init(hugetlb_init); | |
183 | ||
184 | static int __init hugetlb_setup(char *s) | |
185 | { | |
186 | if (sscanf(s, "%lu", &max_huge_pages) <= 0) | |
187 | max_huge_pages = 0; | |
188 | return 1; | |
189 | } | |
190 | __setup("hugepages=", hugetlb_setup); | |
191 | ||
8a630112 KC |
192 | static unsigned int cpuset_mems_nr(unsigned int *array) |
193 | { | |
194 | int node; | |
195 | unsigned int nr = 0; | |
196 | ||
197 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
198 | nr += array[node]; | |
199 | ||
200 | return nr; | |
201 | } | |
202 | ||
1da177e4 LT |
203 | #ifdef CONFIG_SYSCTL |
204 | static void update_and_free_page(struct page *page) | |
205 | { | |
206 | int i; | |
207 | nr_huge_pages--; | |
4415cc8d | 208 | nr_huge_pages_node[page_to_nid(page)]--; |
1da177e4 LT |
209 | for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) { |
210 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced | | |
211 | 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved | | |
212 | 1 << PG_private | 1<< PG_writeback); | |
1da177e4 | 213 | } |
f8af0bb8 | 214 | set_compound_page_dtor(page, NULL); |
7835e98b | 215 | set_page_refcounted(page); |
1da177e4 LT |
216 | __free_pages(page, HUGETLB_PAGE_ORDER); |
217 | } | |
218 | ||
219 | #ifdef CONFIG_HIGHMEM | |
220 | static void try_to_free_low(unsigned long count) | |
221 | { | |
4415cc8d CL |
222 | int i; |
223 | ||
1da177e4 LT |
224 | for (i = 0; i < MAX_NUMNODES; ++i) { |
225 | struct page *page, *next; | |
226 | list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) { | |
227 | if (PageHighMem(page)) | |
228 | continue; | |
229 | list_del(&page->lru); | |
230 | update_and_free_page(page); | |
1da177e4 | 231 | free_huge_pages--; |
4415cc8d | 232 | free_huge_pages_node[page_to_nid(page)]--; |
1da177e4 LT |
233 | if (count >= nr_huge_pages) |
234 | return; | |
235 | } | |
236 | } | |
237 | } | |
238 | #else | |
239 | static inline void try_to_free_low(unsigned long count) | |
240 | { | |
241 | } | |
242 | #endif | |
243 | ||
244 | static unsigned long set_max_huge_pages(unsigned long count) | |
245 | { | |
246 | while (count > nr_huge_pages) { | |
a482289d | 247 | if (!alloc_fresh_huge_page()) |
1da177e4 | 248 | return nr_huge_pages; |
1da177e4 LT |
249 | } |
250 | if (count >= nr_huge_pages) | |
251 | return nr_huge_pages; | |
252 | ||
253 | spin_lock(&hugetlb_lock); | |
a43a8c39 | 254 | count = max(count, resv_huge_pages); |
1da177e4 LT |
255 | try_to_free_low(count); |
256 | while (count < nr_huge_pages) { | |
5da7ca86 | 257 | struct page *page = dequeue_huge_page(NULL, 0); |
1da177e4 LT |
258 | if (!page) |
259 | break; | |
260 | update_and_free_page(page); | |
261 | } | |
262 | spin_unlock(&hugetlb_lock); | |
263 | return nr_huge_pages; | |
264 | } | |
265 | ||
266 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, | |
267 | struct file *file, void __user *buffer, | |
268 | size_t *length, loff_t *ppos) | |
269 | { | |
270 | proc_doulongvec_minmax(table, write, file, buffer, length, ppos); | |
271 | max_huge_pages = set_max_huge_pages(max_huge_pages); | |
272 | return 0; | |
273 | } | |
396faf03 MG |
274 | |
275 | int hugetlb_treat_movable_handler(struct ctl_table *table, int write, | |
276 | struct file *file, void __user *buffer, | |
277 | size_t *length, loff_t *ppos) | |
278 | { | |
279 | proc_dointvec(table, write, file, buffer, length, ppos); | |
280 | if (hugepages_treat_as_movable) | |
281 | htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; | |
282 | else | |
283 | htlb_alloc_mask = GFP_HIGHUSER; | |
284 | return 0; | |
285 | } | |
286 | ||
1da177e4 LT |
287 | #endif /* CONFIG_SYSCTL */ |
288 | ||
289 | int hugetlb_report_meminfo(char *buf) | |
290 | { | |
291 | return sprintf(buf, | |
292 | "HugePages_Total: %5lu\n" | |
293 | "HugePages_Free: %5lu\n" | |
a43a8c39 | 294 | "HugePages_Rsvd: %5lu\n" |
1da177e4 LT |
295 | "Hugepagesize: %5lu kB\n", |
296 | nr_huge_pages, | |
297 | free_huge_pages, | |
a43a8c39 | 298 | resv_huge_pages, |
1da177e4 LT |
299 | HPAGE_SIZE/1024); |
300 | } | |
301 | ||
302 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
303 | { | |
304 | return sprintf(buf, | |
305 | "Node %d HugePages_Total: %5u\n" | |
306 | "Node %d HugePages_Free: %5u\n", | |
307 | nid, nr_huge_pages_node[nid], | |
308 | nid, free_huge_pages_node[nid]); | |
309 | } | |
310 | ||
1da177e4 LT |
311 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
312 | unsigned long hugetlb_total_pages(void) | |
313 | { | |
314 | return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE); | |
315 | } | |
1da177e4 LT |
316 | |
317 | /* | |
318 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
319 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
320 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
321 | * this far. | |
322 | */ | |
d0217ac0 | 323 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
324 | { |
325 | BUG(); | |
d0217ac0 | 326 | return 0; |
1da177e4 LT |
327 | } |
328 | ||
329 | struct vm_operations_struct hugetlb_vm_ops = { | |
d0217ac0 | 330 | .fault = hugetlb_vm_op_fault, |
1da177e4 LT |
331 | }; |
332 | ||
1e8f889b DG |
333 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
334 | int writable) | |
63551ae0 DG |
335 | { |
336 | pte_t entry; | |
337 | ||
1e8f889b | 338 | if (writable) { |
63551ae0 DG |
339 | entry = |
340 | pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); | |
341 | } else { | |
342 | entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot)); | |
343 | } | |
344 | entry = pte_mkyoung(entry); | |
345 | entry = pte_mkhuge(entry); | |
346 | ||
347 | return entry; | |
348 | } | |
349 | ||
1e8f889b DG |
350 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
351 | unsigned long address, pte_t *ptep) | |
352 | { | |
353 | pte_t entry; | |
354 | ||
355 | entry = pte_mkwrite(pte_mkdirty(*ptep)); | |
8dab5241 BH |
356 | if (ptep_set_access_flags(vma, address, ptep, entry, 1)) { |
357 | update_mmu_cache(vma, address, entry); | |
358 | lazy_mmu_prot_update(entry); | |
359 | } | |
1e8f889b DG |
360 | } |
361 | ||
362 | ||
63551ae0 DG |
363 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
364 | struct vm_area_struct *vma) | |
365 | { | |
366 | pte_t *src_pte, *dst_pte, entry; | |
367 | struct page *ptepage; | |
1c59827d | 368 | unsigned long addr; |
1e8f889b DG |
369 | int cow; |
370 | ||
371 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 372 | |
1c59827d | 373 | for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) { |
c74df32c HD |
374 | src_pte = huge_pte_offset(src, addr); |
375 | if (!src_pte) | |
376 | continue; | |
63551ae0 DG |
377 | dst_pte = huge_pte_alloc(dst, addr); |
378 | if (!dst_pte) | |
379 | goto nomem; | |
c74df32c | 380 | spin_lock(&dst->page_table_lock); |
1c59827d | 381 | spin_lock(&src->page_table_lock); |
c74df32c | 382 | if (!pte_none(*src_pte)) { |
1e8f889b DG |
383 | if (cow) |
384 | ptep_set_wrprotect(src, addr, src_pte); | |
1c59827d HD |
385 | entry = *src_pte; |
386 | ptepage = pte_page(entry); | |
387 | get_page(ptepage); | |
1c59827d HD |
388 | set_huge_pte_at(dst, addr, dst_pte, entry); |
389 | } | |
390 | spin_unlock(&src->page_table_lock); | |
c74df32c | 391 | spin_unlock(&dst->page_table_lock); |
63551ae0 DG |
392 | } |
393 | return 0; | |
394 | ||
395 | nomem: | |
396 | return -ENOMEM; | |
397 | } | |
398 | ||
502717f4 KC |
399 | void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
400 | unsigned long end) | |
63551ae0 DG |
401 | { |
402 | struct mm_struct *mm = vma->vm_mm; | |
403 | unsigned long address; | |
c7546f8f | 404 | pte_t *ptep; |
63551ae0 DG |
405 | pte_t pte; |
406 | struct page *page; | |
fe1668ae | 407 | struct page *tmp; |
c0a499c2 KC |
408 | /* |
409 | * A page gathering list, protected by per file i_mmap_lock. The | |
410 | * lock is used to avoid list corruption from multiple unmapping | |
411 | * of the same page since we are using page->lru. | |
412 | */ | |
fe1668ae | 413 | LIST_HEAD(page_list); |
63551ae0 DG |
414 | |
415 | WARN_ON(!is_vm_hugetlb_page(vma)); | |
416 | BUG_ON(start & ~HPAGE_MASK); | |
417 | BUG_ON(end & ~HPAGE_MASK); | |
418 | ||
508034a3 | 419 | spin_lock(&mm->page_table_lock); |
63551ae0 | 420 | for (address = start; address < end; address += HPAGE_SIZE) { |
c7546f8f | 421 | ptep = huge_pte_offset(mm, address); |
4c887265 | 422 | if (!ptep) |
c7546f8f DG |
423 | continue; |
424 | ||
39dde65c KC |
425 | if (huge_pmd_unshare(mm, &address, ptep)) |
426 | continue; | |
427 | ||
c7546f8f | 428 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
63551ae0 DG |
429 | if (pte_none(pte)) |
430 | continue; | |
c7546f8f | 431 | |
63551ae0 | 432 | page = pte_page(pte); |
6649a386 KC |
433 | if (pte_dirty(pte)) |
434 | set_page_dirty(page); | |
fe1668ae | 435 | list_add(&page->lru, &page_list); |
63551ae0 | 436 | } |
1da177e4 | 437 | spin_unlock(&mm->page_table_lock); |
508034a3 | 438 | flush_tlb_range(vma, start, end); |
fe1668ae KC |
439 | list_for_each_entry_safe(page, tmp, &page_list, lru) { |
440 | list_del(&page->lru); | |
441 | put_page(page); | |
442 | } | |
1da177e4 | 443 | } |
63551ae0 | 444 | |
502717f4 KC |
445 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
446 | unsigned long end) | |
447 | { | |
448 | /* | |
449 | * It is undesirable to test vma->vm_file as it should be non-null | |
450 | * for valid hugetlb area. However, vm_file will be NULL in the error | |
451 | * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails, | |
452 | * do_mmap_pgoff() nullifies vma->vm_file before calling this function | |
453 | * to clean up. Since no pte has actually been setup, it is safe to | |
454 | * do nothing in this case. | |
455 | */ | |
456 | if (vma->vm_file) { | |
457 | spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); | |
458 | __unmap_hugepage_range(vma, start, end); | |
459 | spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); | |
460 | } | |
461 | } | |
462 | ||
1e8f889b DG |
463 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
464 | unsigned long address, pte_t *ptep, pte_t pte) | |
465 | { | |
466 | struct page *old_page, *new_page; | |
79ac6ba4 | 467 | int avoidcopy; |
1e8f889b DG |
468 | |
469 | old_page = pte_page(pte); | |
470 | ||
471 | /* If no-one else is actually using this page, avoid the copy | |
472 | * and just make the page writable */ | |
473 | avoidcopy = (page_count(old_page) == 1); | |
474 | if (avoidcopy) { | |
475 | set_huge_ptep_writable(vma, address, ptep); | |
83c54070 | 476 | return 0; |
1e8f889b DG |
477 | } |
478 | ||
479 | page_cache_get(old_page); | |
5da7ca86 | 480 | new_page = alloc_huge_page(vma, address); |
1e8f889b DG |
481 | |
482 | if (!new_page) { | |
483 | page_cache_release(old_page); | |
0df420d8 | 484 | return VM_FAULT_OOM; |
1e8f889b DG |
485 | } |
486 | ||
487 | spin_unlock(&mm->page_table_lock); | |
9de455b2 | 488 | copy_huge_page(new_page, old_page, address, vma); |
1e8f889b DG |
489 | spin_lock(&mm->page_table_lock); |
490 | ||
491 | ptep = huge_pte_offset(mm, address & HPAGE_MASK); | |
492 | if (likely(pte_same(*ptep, pte))) { | |
493 | /* Break COW */ | |
494 | set_huge_pte_at(mm, address, ptep, | |
495 | make_huge_pte(vma, new_page, 1)); | |
496 | /* Make the old page be freed below */ | |
497 | new_page = old_page; | |
498 | } | |
499 | page_cache_release(new_page); | |
500 | page_cache_release(old_page); | |
83c54070 | 501 | return 0; |
1e8f889b DG |
502 | } |
503 | ||
a1ed3dda | 504 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
1e8f889b | 505 | unsigned long address, pte_t *ptep, int write_access) |
ac9b9c66 HD |
506 | { |
507 | int ret = VM_FAULT_SIGBUS; | |
4c887265 AL |
508 | unsigned long idx; |
509 | unsigned long size; | |
4c887265 AL |
510 | struct page *page; |
511 | struct address_space *mapping; | |
1e8f889b | 512 | pte_t new_pte; |
4c887265 | 513 | |
4c887265 AL |
514 | mapping = vma->vm_file->f_mapping; |
515 | idx = ((address - vma->vm_start) >> HPAGE_SHIFT) | |
516 | + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT)); | |
517 | ||
518 | /* | |
519 | * Use page lock to guard against racing truncation | |
520 | * before we get page_table_lock. | |
521 | */ | |
6bda666a CL |
522 | retry: |
523 | page = find_lock_page(mapping, idx); | |
524 | if (!page) { | |
ebed4bfc HD |
525 | size = i_size_read(mapping->host) >> HPAGE_SHIFT; |
526 | if (idx >= size) | |
527 | goto out; | |
6bda666a CL |
528 | if (hugetlb_get_quota(mapping)) |
529 | goto out; | |
530 | page = alloc_huge_page(vma, address); | |
531 | if (!page) { | |
532 | hugetlb_put_quota(mapping); | |
0df420d8 | 533 | ret = VM_FAULT_OOM; |
6bda666a CL |
534 | goto out; |
535 | } | |
79ac6ba4 | 536 | clear_huge_page(page, address); |
ac9b9c66 | 537 | |
6bda666a CL |
538 | if (vma->vm_flags & VM_SHARED) { |
539 | int err; | |
540 | ||
541 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
542 | if (err) { | |
543 | put_page(page); | |
544 | hugetlb_put_quota(mapping); | |
545 | if (err == -EEXIST) | |
546 | goto retry; | |
547 | goto out; | |
548 | } | |
549 | } else | |
550 | lock_page(page); | |
551 | } | |
1e8f889b | 552 | |
ac9b9c66 | 553 | spin_lock(&mm->page_table_lock); |
4c887265 AL |
554 | size = i_size_read(mapping->host) >> HPAGE_SHIFT; |
555 | if (idx >= size) | |
556 | goto backout; | |
557 | ||
83c54070 | 558 | ret = 0; |
86e5216f | 559 | if (!pte_none(*ptep)) |
4c887265 AL |
560 | goto backout; |
561 | ||
1e8f889b DG |
562 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
563 | && (vma->vm_flags & VM_SHARED))); | |
564 | set_huge_pte_at(mm, address, ptep, new_pte); | |
565 | ||
566 | if (write_access && !(vma->vm_flags & VM_SHARED)) { | |
567 | /* Optimization, do the COW without a second fault */ | |
568 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte); | |
569 | } | |
570 | ||
ac9b9c66 | 571 | spin_unlock(&mm->page_table_lock); |
4c887265 AL |
572 | unlock_page(page); |
573 | out: | |
ac9b9c66 | 574 | return ret; |
4c887265 AL |
575 | |
576 | backout: | |
577 | spin_unlock(&mm->page_table_lock); | |
578 | hugetlb_put_quota(mapping); | |
579 | unlock_page(page); | |
580 | put_page(page); | |
581 | goto out; | |
ac9b9c66 HD |
582 | } |
583 | ||
86e5216f AL |
584 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
585 | unsigned long address, int write_access) | |
586 | { | |
587 | pte_t *ptep; | |
588 | pte_t entry; | |
1e8f889b | 589 | int ret; |
3935baa9 | 590 | static DEFINE_MUTEX(hugetlb_instantiation_mutex); |
86e5216f AL |
591 | |
592 | ptep = huge_pte_alloc(mm, address); | |
593 | if (!ptep) | |
594 | return VM_FAULT_OOM; | |
595 | ||
3935baa9 DG |
596 | /* |
597 | * Serialize hugepage allocation and instantiation, so that we don't | |
598 | * get spurious allocation failures if two CPUs race to instantiate | |
599 | * the same page in the page cache. | |
600 | */ | |
601 | mutex_lock(&hugetlb_instantiation_mutex); | |
86e5216f | 602 | entry = *ptep; |
3935baa9 DG |
603 | if (pte_none(entry)) { |
604 | ret = hugetlb_no_page(mm, vma, address, ptep, write_access); | |
605 | mutex_unlock(&hugetlb_instantiation_mutex); | |
606 | return ret; | |
607 | } | |
86e5216f | 608 | |
83c54070 | 609 | ret = 0; |
1e8f889b DG |
610 | |
611 | spin_lock(&mm->page_table_lock); | |
612 | /* Check for a racing update before calling hugetlb_cow */ | |
613 | if (likely(pte_same(entry, *ptep))) | |
614 | if (write_access && !pte_write(entry)) | |
615 | ret = hugetlb_cow(mm, vma, address, ptep, entry); | |
616 | spin_unlock(&mm->page_table_lock); | |
3935baa9 | 617 | mutex_unlock(&hugetlb_instantiation_mutex); |
1e8f889b DG |
618 | |
619 | return ret; | |
86e5216f AL |
620 | } |
621 | ||
63551ae0 DG |
622 | int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
623 | struct page **pages, struct vm_area_struct **vmas, | |
624 | unsigned long *position, int *length, int i) | |
625 | { | |
d5d4b0aa KC |
626 | unsigned long pfn_offset; |
627 | unsigned long vaddr = *position; | |
63551ae0 DG |
628 | int remainder = *length; |
629 | ||
1c59827d | 630 | spin_lock(&mm->page_table_lock); |
63551ae0 | 631 | while (vaddr < vma->vm_end && remainder) { |
4c887265 AL |
632 | pte_t *pte; |
633 | struct page *page; | |
63551ae0 | 634 | |
4c887265 AL |
635 | /* |
636 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
637 | * each hugepage. We have to make * sure we get the | |
638 | * first, for the page indexing below to work. | |
639 | */ | |
640 | pte = huge_pte_offset(mm, vaddr & HPAGE_MASK); | |
63551ae0 | 641 | |
4c887265 AL |
642 | if (!pte || pte_none(*pte)) { |
643 | int ret; | |
63551ae0 | 644 | |
4c887265 AL |
645 | spin_unlock(&mm->page_table_lock); |
646 | ret = hugetlb_fault(mm, vma, vaddr, 0); | |
647 | spin_lock(&mm->page_table_lock); | |
a89182c7 | 648 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 649 | continue; |
63551ae0 | 650 | |
4c887265 AL |
651 | remainder = 0; |
652 | if (!i) | |
653 | i = -EFAULT; | |
654 | break; | |
655 | } | |
656 | ||
d5d4b0aa KC |
657 | pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT; |
658 | page = pte_page(*pte); | |
659 | same_page: | |
d6692183 KC |
660 | if (pages) { |
661 | get_page(page); | |
d5d4b0aa | 662 | pages[i] = page + pfn_offset; |
d6692183 | 663 | } |
63551ae0 DG |
664 | |
665 | if (vmas) | |
666 | vmas[i] = vma; | |
667 | ||
668 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 669 | ++pfn_offset; |
63551ae0 DG |
670 | --remainder; |
671 | ++i; | |
d5d4b0aa KC |
672 | if (vaddr < vma->vm_end && remainder && |
673 | pfn_offset < HPAGE_SIZE/PAGE_SIZE) { | |
674 | /* | |
675 | * We use pfn_offset to avoid touching the pageframes | |
676 | * of this compound page. | |
677 | */ | |
678 | goto same_page; | |
679 | } | |
63551ae0 | 680 | } |
1c59827d | 681 | spin_unlock(&mm->page_table_lock); |
63551ae0 DG |
682 | *length = remainder; |
683 | *position = vaddr; | |
684 | ||
685 | return i; | |
686 | } | |
8f860591 ZY |
687 | |
688 | void hugetlb_change_protection(struct vm_area_struct *vma, | |
689 | unsigned long address, unsigned long end, pgprot_t newprot) | |
690 | { | |
691 | struct mm_struct *mm = vma->vm_mm; | |
692 | unsigned long start = address; | |
693 | pte_t *ptep; | |
694 | pte_t pte; | |
695 | ||
696 | BUG_ON(address >= end); | |
697 | flush_cache_range(vma, address, end); | |
698 | ||
39dde65c | 699 | spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); |
8f860591 ZY |
700 | spin_lock(&mm->page_table_lock); |
701 | for (; address < end; address += HPAGE_SIZE) { | |
702 | ptep = huge_pte_offset(mm, address); | |
703 | if (!ptep) | |
704 | continue; | |
39dde65c KC |
705 | if (huge_pmd_unshare(mm, &address, ptep)) |
706 | continue; | |
8f860591 ZY |
707 | if (!pte_none(*ptep)) { |
708 | pte = huge_ptep_get_and_clear(mm, address, ptep); | |
709 | pte = pte_mkhuge(pte_modify(pte, newprot)); | |
710 | set_huge_pte_at(mm, address, ptep, pte); | |
711 | lazy_mmu_prot_update(pte); | |
712 | } | |
713 | } | |
714 | spin_unlock(&mm->page_table_lock); | |
39dde65c | 715 | spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); |
8f860591 ZY |
716 | |
717 | flush_tlb_range(vma, start, end); | |
718 | } | |
719 | ||
a43a8c39 KC |
720 | struct file_region { |
721 | struct list_head link; | |
722 | long from; | |
723 | long to; | |
724 | }; | |
725 | ||
726 | static long region_add(struct list_head *head, long f, long t) | |
727 | { | |
728 | struct file_region *rg, *nrg, *trg; | |
729 | ||
730 | /* Locate the region we are either in or before. */ | |
731 | list_for_each_entry(rg, head, link) | |
732 | if (f <= rg->to) | |
733 | break; | |
734 | ||
735 | /* Round our left edge to the current segment if it encloses us. */ | |
736 | if (f > rg->from) | |
737 | f = rg->from; | |
738 | ||
739 | /* Check for and consume any regions we now overlap with. */ | |
740 | nrg = rg; | |
741 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
742 | if (&rg->link == head) | |
743 | break; | |
744 | if (rg->from > t) | |
745 | break; | |
746 | ||
747 | /* If this area reaches higher then extend our area to | |
748 | * include it completely. If this is not the first area | |
749 | * which we intend to reuse, free it. */ | |
750 | if (rg->to > t) | |
751 | t = rg->to; | |
752 | if (rg != nrg) { | |
753 | list_del(&rg->link); | |
754 | kfree(rg); | |
755 | } | |
756 | } | |
757 | nrg->from = f; | |
758 | nrg->to = t; | |
759 | return 0; | |
760 | } | |
761 | ||
762 | static long region_chg(struct list_head *head, long f, long t) | |
763 | { | |
764 | struct file_region *rg, *nrg; | |
765 | long chg = 0; | |
766 | ||
767 | /* Locate the region we are before or in. */ | |
768 | list_for_each_entry(rg, head, link) | |
769 | if (f <= rg->to) | |
770 | break; | |
771 | ||
772 | /* If we are below the current region then a new region is required. | |
773 | * Subtle, allocate a new region at the position but make it zero | |
774 | * size such that we can guarentee to record the reservation. */ | |
775 | if (&rg->link == head || t < rg->from) { | |
776 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
777 | if (nrg == 0) | |
778 | return -ENOMEM; | |
779 | nrg->from = f; | |
780 | nrg->to = f; | |
781 | INIT_LIST_HEAD(&nrg->link); | |
782 | list_add(&nrg->link, rg->link.prev); | |
783 | ||
784 | return t - f; | |
785 | } | |
786 | ||
787 | /* Round our left edge to the current segment if it encloses us. */ | |
788 | if (f > rg->from) | |
789 | f = rg->from; | |
790 | chg = t - f; | |
791 | ||
792 | /* Check for and consume any regions we now overlap with. */ | |
793 | list_for_each_entry(rg, rg->link.prev, link) { | |
794 | if (&rg->link == head) | |
795 | break; | |
796 | if (rg->from > t) | |
797 | return chg; | |
798 | ||
799 | /* We overlap with this area, if it extends futher than | |
800 | * us then we must extend ourselves. Account for its | |
801 | * existing reservation. */ | |
802 | if (rg->to > t) { | |
803 | chg += rg->to - t; | |
804 | t = rg->to; | |
805 | } | |
806 | chg -= rg->to - rg->from; | |
807 | } | |
808 | return chg; | |
809 | } | |
810 | ||
811 | static long region_truncate(struct list_head *head, long end) | |
812 | { | |
813 | struct file_region *rg, *trg; | |
814 | long chg = 0; | |
815 | ||
816 | /* Locate the region we are either in or before. */ | |
817 | list_for_each_entry(rg, head, link) | |
818 | if (end <= rg->to) | |
819 | break; | |
820 | if (&rg->link == head) | |
821 | return 0; | |
822 | ||
823 | /* If we are in the middle of a region then adjust it. */ | |
824 | if (end > rg->from) { | |
825 | chg = rg->to - end; | |
826 | rg->to = end; | |
827 | rg = list_entry(rg->link.next, typeof(*rg), link); | |
828 | } | |
829 | ||
830 | /* Drop any remaining regions. */ | |
831 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
832 | if (&rg->link == head) | |
833 | break; | |
834 | chg += rg->to - rg->from; | |
835 | list_del(&rg->link); | |
836 | kfree(rg); | |
837 | } | |
838 | return chg; | |
839 | } | |
840 | ||
841 | static int hugetlb_acct_memory(long delta) | |
842 | { | |
843 | int ret = -ENOMEM; | |
844 | ||
845 | spin_lock(&hugetlb_lock); | |
846 | if ((delta + resv_huge_pages) <= free_huge_pages) { | |
847 | resv_huge_pages += delta; | |
848 | ret = 0; | |
849 | } | |
850 | spin_unlock(&hugetlb_lock); | |
851 | return ret; | |
852 | } | |
853 | ||
854 | int hugetlb_reserve_pages(struct inode *inode, long from, long to) | |
855 | { | |
856 | long ret, chg; | |
857 | ||
858 | chg = region_chg(&inode->i_mapping->private_list, from, to); | |
859 | if (chg < 0) | |
860 | return chg; | |
8a630112 KC |
861 | /* |
862 | * When cpuset is configured, it breaks the strict hugetlb page | |
863 | * reservation as the accounting is done on a global variable. Such | |
864 | * reservation is completely rubbish in the presence of cpuset because | |
865 | * the reservation is not checked against page availability for the | |
866 | * current cpuset. Application can still potentially OOM'ed by kernel | |
867 | * with lack of free htlb page in cpuset that the task is in. | |
868 | * Attempt to enforce strict accounting with cpuset is almost | |
869 | * impossible (or too ugly) because cpuset is too fluid that | |
870 | * task or memory node can be dynamically moved between cpusets. | |
871 | * | |
872 | * The change of semantics for shared hugetlb mapping with cpuset is | |
873 | * undesirable. However, in order to preserve some of the semantics, | |
874 | * we fall back to check against current free page availability as | |
875 | * a best attempt and hopefully to minimize the impact of changing | |
876 | * semantics that cpuset has. | |
877 | */ | |
878 | if (chg > cpuset_mems_nr(free_huge_pages_node)) | |
879 | return -ENOMEM; | |
880 | ||
a43a8c39 KC |
881 | ret = hugetlb_acct_memory(chg); |
882 | if (ret < 0) | |
883 | return ret; | |
884 | region_add(&inode->i_mapping->private_list, from, to); | |
885 | return 0; | |
886 | } | |
887 | ||
888 | void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) | |
889 | { | |
890 | long chg = region_truncate(&inode->i_mapping->private_list, offset); | |
891 | hugetlb_acct_memory(freed - chg); | |
892 | } |