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(); | |
45 | clear_user_highpage(page + i, addr); | |
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; |
396faf03 MG |
74 | struct zonelist *zonelist = huge_zonelist(vma, address, |
75 | htlb_alloc_mask); | |
96df9333 | 76 | struct zone **z; |
1da177e4 | 77 | |
96df9333 | 78 | for (z = zonelist->zones; *z; z++) { |
89fa3024 | 79 | nid = zone_to_nid(*z); |
396faf03 | 80 | if (cpuset_zone_allowed_softwall(*z, htlb_alloc_mask) && |
aea47ff3 | 81 | !list_empty(&hugepage_freelists[nid])) |
96df9333 | 82 | break; |
1da177e4 | 83 | } |
96df9333 CL |
84 | |
85 | if (*z) { | |
1da177e4 LT |
86 | page = list_entry(hugepage_freelists[nid].next, |
87 | struct page, lru); | |
88 | list_del(&page->lru); | |
89 | free_huge_pages--; | |
90 | free_huge_pages_node[nid]--; | |
91 | } | |
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 | static DEFINE_SPINLOCK(nid_lock); |
111 | int nid; | |
112 | ||
113 | spin_lock(&nid_lock); | |
114 | nid = next_node(prev_nid, node_online_map); | |
fdb7cc59 PJ |
115 | if (nid == MAX_NUMNODES) |
116 | nid = first_node(node_online_map); | |
f96efd58 JJ |
117 | prev_nid = nid; |
118 | spin_unlock(&nid_lock); | |
119 | ||
396faf03 | 120 | page = alloc_pages_node(nid, htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN, |
f96efd58 | 121 | HUGETLB_PAGE_ORDER); |
1da177e4 | 122 | if (page) { |
33f2ef89 | 123 | set_compound_page_dtor(page, free_huge_page); |
0bd0f9fb | 124 | spin_lock(&hugetlb_lock); |
1da177e4 LT |
125 | nr_huge_pages++; |
126 | nr_huge_pages_node[page_to_nid(page)]++; | |
0bd0f9fb | 127 | spin_unlock(&hugetlb_lock); |
a482289d NP |
128 | put_page(page); /* free it into the hugepage allocator */ |
129 | return 1; | |
1da177e4 | 130 | } |
a482289d | 131 | return 0; |
1da177e4 LT |
132 | } |
133 | ||
27a85ef1 DG |
134 | static struct page *alloc_huge_page(struct vm_area_struct *vma, |
135 | unsigned long addr) | |
1da177e4 LT |
136 | { |
137 | struct page *page; | |
1da177e4 LT |
138 | |
139 | spin_lock(&hugetlb_lock); | |
a43a8c39 CK |
140 | if (vma->vm_flags & VM_MAYSHARE) |
141 | resv_huge_pages--; | |
142 | else if (free_huge_pages <= resv_huge_pages) | |
143 | goto fail; | |
b45b5bd6 DG |
144 | |
145 | page = dequeue_huge_page(vma, addr); | |
146 | if (!page) | |
147 | goto fail; | |
148 | ||
1da177e4 | 149 | spin_unlock(&hugetlb_lock); |
7835e98b | 150 | set_page_refcounted(page); |
1da177e4 | 151 | return page; |
b45b5bd6 | 152 | |
a43a8c39 | 153 | fail: |
ace4bd29 KC |
154 | if (vma->vm_flags & VM_MAYSHARE) |
155 | resv_huge_pages++; | |
b45b5bd6 DG |
156 | spin_unlock(&hugetlb_lock); |
157 | return NULL; | |
158 | } | |
159 | ||
1da177e4 LT |
160 | static int __init hugetlb_init(void) |
161 | { | |
162 | unsigned long i; | |
1da177e4 | 163 | |
3c726f8d BH |
164 | if (HPAGE_SHIFT == 0) |
165 | return 0; | |
166 | ||
1da177e4 LT |
167 | for (i = 0; i < MAX_NUMNODES; ++i) |
168 | INIT_LIST_HEAD(&hugepage_freelists[i]); | |
169 | ||
170 | for (i = 0; i < max_huge_pages; ++i) { | |
a482289d | 171 | if (!alloc_fresh_huge_page()) |
1da177e4 | 172 | break; |
1da177e4 LT |
173 | } |
174 | max_huge_pages = free_huge_pages = nr_huge_pages = i; | |
175 | printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages); | |
176 | return 0; | |
177 | } | |
178 | module_init(hugetlb_init); | |
179 | ||
180 | static int __init hugetlb_setup(char *s) | |
181 | { | |
182 | if (sscanf(s, "%lu", &max_huge_pages) <= 0) | |
183 | max_huge_pages = 0; | |
184 | return 1; | |
185 | } | |
186 | __setup("hugepages=", hugetlb_setup); | |
187 | ||
8a630112 KC |
188 | static unsigned int cpuset_mems_nr(unsigned int *array) |
189 | { | |
190 | int node; | |
191 | unsigned int nr = 0; | |
192 | ||
193 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
194 | nr += array[node]; | |
195 | ||
196 | return nr; | |
197 | } | |
198 | ||
1da177e4 LT |
199 | #ifdef CONFIG_SYSCTL |
200 | static void update_and_free_page(struct page *page) | |
201 | { | |
202 | int i; | |
203 | nr_huge_pages--; | |
4415cc8d | 204 | nr_huge_pages_node[page_to_nid(page)]--; |
1da177e4 LT |
205 | for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) { |
206 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced | | |
207 | 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved | | |
208 | 1 << PG_private | 1<< PG_writeback); | |
1da177e4 | 209 | } |
a482289d | 210 | page[1].lru.next = NULL; |
7835e98b | 211 | set_page_refcounted(page); |
1da177e4 LT |
212 | __free_pages(page, HUGETLB_PAGE_ORDER); |
213 | } | |
214 | ||
215 | #ifdef CONFIG_HIGHMEM | |
216 | static void try_to_free_low(unsigned long count) | |
217 | { | |
4415cc8d CL |
218 | int i; |
219 | ||
1da177e4 LT |
220 | for (i = 0; i < MAX_NUMNODES; ++i) { |
221 | struct page *page, *next; | |
222 | list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) { | |
223 | if (PageHighMem(page)) | |
224 | continue; | |
225 | list_del(&page->lru); | |
226 | update_and_free_page(page); | |
1da177e4 | 227 | free_huge_pages--; |
4415cc8d | 228 | free_huge_pages_node[page_to_nid(page)]--; |
1da177e4 LT |
229 | if (count >= nr_huge_pages) |
230 | return; | |
231 | } | |
232 | } | |
233 | } | |
234 | #else | |
235 | static inline void try_to_free_low(unsigned long count) | |
236 | { | |
237 | } | |
238 | #endif | |
239 | ||
240 | static unsigned long set_max_huge_pages(unsigned long count) | |
241 | { | |
242 | while (count > nr_huge_pages) { | |
a482289d | 243 | if (!alloc_fresh_huge_page()) |
1da177e4 | 244 | return nr_huge_pages; |
1da177e4 LT |
245 | } |
246 | if (count >= nr_huge_pages) | |
247 | return nr_huge_pages; | |
248 | ||
249 | spin_lock(&hugetlb_lock); | |
a43a8c39 | 250 | count = max(count, resv_huge_pages); |
1da177e4 LT |
251 | try_to_free_low(count); |
252 | while (count < nr_huge_pages) { | |
5da7ca86 | 253 | struct page *page = dequeue_huge_page(NULL, 0); |
1da177e4 LT |
254 | if (!page) |
255 | break; | |
256 | update_and_free_page(page); | |
257 | } | |
258 | spin_unlock(&hugetlb_lock); | |
259 | return nr_huge_pages; | |
260 | } | |
261 | ||
262 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, | |
263 | struct file *file, void __user *buffer, | |
264 | size_t *length, loff_t *ppos) | |
265 | { | |
266 | proc_doulongvec_minmax(table, write, file, buffer, length, ppos); | |
267 | max_huge_pages = set_max_huge_pages(max_huge_pages); | |
268 | return 0; | |
269 | } | |
396faf03 MG |
270 | |
271 | int hugetlb_treat_movable_handler(struct ctl_table *table, int write, | |
272 | struct file *file, void __user *buffer, | |
273 | size_t *length, loff_t *ppos) | |
274 | { | |
275 | proc_dointvec(table, write, file, buffer, length, ppos); | |
276 | if (hugepages_treat_as_movable) | |
277 | htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; | |
278 | else | |
279 | htlb_alloc_mask = GFP_HIGHUSER; | |
280 | return 0; | |
281 | } | |
282 | ||
1da177e4 LT |
283 | #endif /* CONFIG_SYSCTL */ |
284 | ||
285 | int hugetlb_report_meminfo(char *buf) | |
286 | { | |
287 | return sprintf(buf, | |
288 | "HugePages_Total: %5lu\n" | |
289 | "HugePages_Free: %5lu\n" | |
a43a8c39 | 290 | "HugePages_Rsvd: %5lu\n" |
1da177e4 LT |
291 | "Hugepagesize: %5lu kB\n", |
292 | nr_huge_pages, | |
293 | free_huge_pages, | |
a43a8c39 | 294 | resv_huge_pages, |
1da177e4 LT |
295 | HPAGE_SIZE/1024); |
296 | } | |
297 | ||
298 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
299 | { | |
300 | return sprintf(buf, | |
301 | "Node %d HugePages_Total: %5u\n" | |
302 | "Node %d HugePages_Free: %5u\n", | |
303 | nid, nr_huge_pages_node[nid], | |
304 | nid, free_huge_pages_node[nid]); | |
305 | } | |
306 | ||
1da177e4 LT |
307 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
308 | unsigned long hugetlb_total_pages(void) | |
309 | { | |
310 | return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE); | |
311 | } | |
1da177e4 LT |
312 | |
313 | /* | |
314 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
315 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
316 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
317 | * this far. | |
318 | */ | |
d0217ac0 | 319 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
320 | { |
321 | BUG(); | |
d0217ac0 | 322 | return 0; |
1da177e4 LT |
323 | } |
324 | ||
325 | struct vm_operations_struct hugetlb_vm_ops = { | |
d0217ac0 | 326 | .fault = hugetlb_vm_op_fault, |
1da177e4 LT |
327 | }; |
328 | ||
1e8f889b DG |
329 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
330 | int writable) | |
63551ae0 DG |
331 | { |
332 | pte_t entry; | |
333 | ||
1e8f889b | 334 | if (writable) { |
63551ae0 DG |
335 | entry = |
336 | pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); | |
337 | } else { | |
338 | entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot)); | |
339 | } | |
340 | entry = pte_mkyoung(entry); | |
341 | entry = pte_mkhuge(entry); | |
342 | ||
343 | return entry; | |
344 | } | |
345 | ||
1e8f889b DG |
346 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
347 | unsigned long address, pte_t *ptep) | |
348 | { | |
349 | pte_t entry; | |
350 | ||
351 | entry = pte_mkwrite(pte_mkdirty(*ptep)); | |
8dab5241 BH |
352 | if (ptep_set_access_flags(vma, address, ptep, entry, 1)) { |
353 | update_mmu_cache(vma, address, entry); | |
354 | lazy_mmu_prot_update(entry); | |
355 | } | |
1e8f889b DG |
356 | } |
357 | ||
358 | ||
63551ae0 DG |
359 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
360 | struct vm_area_struct *vma) | |
361 | { | |
362 | pte_t *src_pte, *dst_pte, entry; | |
363 | struct page *ptepage; | |
1c59827d | 364 | unsigned long addr; |
1e8f889b DG |
365 | int cow; |
366 | ||
367 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 368 | |
1c59827d | 369 | for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) { |
c74df32c HD |
370 | src_pte = huge_pte_offset(src, addr); |
371 | if (!src_pte) | |
372 | continue; | |
63551ae0 DG |
373 | dst_pte = huge_pte_alloc(dst, addr); |
374 | if (!dst_pte) | |
375 | goto nomem; | |
c74df32c | 376 | spin_lock(&dst->page_table_lock); |
1c59827d | 377 | spin_lock(&src->page_table_lock); |
c74df32c | 378 | if (!pte_none(*src_pte)) { |
1e8f889b DG |
379 | if (cow) |
380 | ptep_set_wrprotect(src, addr, src_pte); | |
1c59827d HD |
381 | entry = *src_pte; |
382 | ptepage = pte_page(entry); | |
383 | get_page(ptepage); | |
1c59827d HD |
384 | set_huge_pte_at(dst, addr, dst_pte, entry); |
385 | } | |
386 | spin_unlock(&src->page_table_lock); | |
c74df32c | 387 | spin_unlock(&dst->page_table_lock); |
63551ae0 DG |
388 | } |
389 | return 0; | |
390 | ||
391 | nomem: | |
392 | return -ENOMEM; | |
393 | } | |
394 | ||
502717f4 CK |
395 | void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
396 | unsigned long end) | |
63551ae0 DG |
397 | { |
398 | struct mm_struct *mm = vma->vm_mm; | |
399 | unsigned long address; | |
c7546f8f | 400 | pte_t *ptep; |
63551ae0 DG |
401 | pte_t pte; |
402 | struct page *page; | |
fe1668ae | 403 | struct page *tmp; |
c0a499c2 CK |
404 | /* |
405 | * A page gathering list, protected by per file i_mmap_lock. The | |
406 | * lock is used to avoid list corruption from multiple unmapping | |
407 | * of the same page since we are using page->lru. | |
408 | */ | |
fe1668ae | 409 | LIST_HEAD(page_list); |
63551ae0 DG |
410 | |
411 | WARN_ON(!is_vm_hugetlb_page(vma)); | |
412 | BUG_ON(start & ~HPAGE_MASK); | |
413 | BUG_ON(end & ~HPAGE_MASK); | |
414 | ||
508034a3 | 415 | spin_lock(&mm->page_table_lock); |
63551ae0 | 416 | for (address = start; address < end; address += HPAGE_SIZE) { |
c7546f8f | 417 | ptep = huge_pte_offset(mm, address); |
4c887265 | 418 | if (!ptep) |
c7546f8f DG |
419 | continue; |
420 | ||
39dde65c CK |
421 | if (huge_pmd_unshare(mm, &address, ptep)) |
422 | continue; | |
423 | ||
c7546f8f | 424 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
63551ae0 DG |
425 | if (pte_none(pte)) |
426 | continue; | |
c7546f8f | 427 | |
63551ae0 | 428 | page = pte_page(pte); |
6649a386 KC |
429 | if (pte_dirty(pte)) |
430 | set_page_dirty(page); | |
fe1668ae | 431 | list_add(&page->lru, &page_list); |
63551ae0 | 432 | } |
1da177e4 | 433 | spin_unlock(&mm->page_table_lock); |
508034a3 | 434 | flush_tlb_range(vma, start, end); |
fe1668ae CK |
435 | list_for_each_entry_safe(page, tmp, &page_list, lru) { |
436 | list_del(&page->lru); | |
437 | put_page(page); | |
438 | } | |
1da177e4 | 439 | } |
63551ae0 | 440 | |
502717f4 CK |
441 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
442 | unsigned long end) | |
443 | { | |
444 | /* | |
445 | * It is undesirable to test vma->vm_file as it should be non-null | |
446 | * for valid hugetlb area. However, vm_file will be NULL in the error | |
447 | * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails, | |
448 | * do_mmap_pgoff() nullifies vma->vm_file before calling this function | |
449 | * to clean up. Since no pte has actually been setup, it is safe to | |
450 | * do nothing in this case. | |
451 | */ | |
452 | if (vma->vm_file) { | |
453 | spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); | |
454 | __unmap_hugepage_range(vma, start, end); | |
455 | spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); | |
456 | } | |
457 | } | |
458 | ||
1e8f889b DG |
459 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
460 | unsigned long address, pte_t *ptep, pte_t pte) | |
461 | { | |
462 | struct page *old_page, *new_page; | |
79ac6ba4 | 463 | int avoidcopy; |
1e8f889b DG |
464 | |
465 | old_page = pte_page(pte); | |
466 | ||
467 | /* If no-one else is actually using this page, avoid the copy | |
468 | * and just make the page writable */ | |
469 | avoidcopy = (page_count(old_page) == 1); | |
470 | if (avoidcopy) { | |
471 | set_huge_ptep_writable(vma, address, ptep); | |
83c54070 | 472 | return 0; |
1e8f889b DG |
473 | } |
474 | ||
475 | page_cache_get(old_page); | |
5da7ca86 | 476 | new_page = alloc_huge_page(vma, address); |
1e8f889b DG |
477 | |
478 | if (!new_page) { | |
479 | page_cache_release(old_page); | |
0df420d8 | 480 | return VM_FAULT_OOM; |
1e8f889b DG |
481 | } |
482 | ||
483 | spin_unlock(&mm->page_table_lock); | |
9de455b2 | 484 | copy_huge_page(new_page, old_page, address, vma); |
1e8f889b DG |
485 | spin_lock(&mm->page_table_lock); |
486 | ||
487 | ptep = huge_pte_offset(mm, address & HPAGE_MASK); | |
488 | if (likely(pte_same(*ptep, pte))) { | |
489 | /* Break COW */ | |
490 | set_huge_pte_at(mm, address, ptep, | |
491 | make_huge_pte(vma, new_page, 1)); | |
492 | /* Make the old page be freed below */ | |
493 | new_page = old_page; | |
494 | } | |
495 | page_cache_release(new_page); | |
496 | page_cache_release(old_page); | |
83c54070 | 497 | return 0; |
1e8f889b DG |
498 | } |
499 | ||
a1ed3dda | 500 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
1e8f889b | 501 | unsigned long address, pte_t *ptep, int write_access) |
ac9b9c66 HD |
502 | { |
503 | int ret = VM_FAULT_SIGBUS; | |
4c887265 AL |
504 | unsigned long idx; |
505 | unsigned long size; | |
4c887265 AL |
506 | struct page *page; |
507 | struct address_space *mapping; | |
1e8f889b | 508 | pte_t new_pte; |
4c887265 | 509 | |
4c887265 AL |
510 | mapping = vma->vm_file->f_mapping; |
511 | idx = ((address - vma->vm_start) >> HPAGE_SHIFT) | |
512 | + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT)); | |
513 | ||
514 | /* | |
515 | * Use page lock to guard against racing truncation | |
516 | * before we get page_table_lock. | |
517 | */ | |
6bda666a CL |
518 | retry: |
519 | page = find_lock_page(mapping, idx); | |
520 | if (!page) { | |
ebed4bfc HD |
521 | size = i_size_read(mapping->host) >> HPAGE_SHIFT; |
522 | if (idx >= size) | |
523 | goto out; | |
6bda666a CL |
524 | if (hugetlb_get_quota(mapping)) |
525 | goto out; | |
526 | page = alloc_huge_page(vma, address); | |
527 | if (!page) { | |
528 | hugetlb_put_quota(mapping); | |
0df420d8 | 529 | ret = VM_FAULT_OOM; |
6bda666a CL |
530 | goto out; |
531 | } | |
79ac6ba4 | 532 | clear_huge_page(page, address); |
ac9b9c66 | 533 | |
6bda666a CL |
534 | if (vma->vm_flags & VM_SHARED) { |
535 | int err; | |
536 | ||
537 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
538 | if (err) { | |
539 | put_page(page); | |
540 | hugetlb_put_quota(mapping); | |
541 | if (err == -EEXIST) | |
542 | goto retry; | |
543 | goto out; | |
544 | } | |
545 | } else | |
546 | lock_page(page); | |
547 | } | |
1e8f889b | 548 | |
ac9b9c66 | 549 | spin_lock(&mm->page_table_lock); |
4c887265 AL |
550 | size = i_size_read(mapping->host) >> HPAGE_SHIFT; |
551 | if (idx >= size) | |
552 | goto backout; | |
553 | ||
83c54070 | 554 | ret = 0; |
86e5216f | 555 | if (!pte_none(*ptep)) |
4c887265 AL |
556 | goto backout; |
557 | ||
1e8f889b DG |
558 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
559 | && (vma->vm_flags & VM_SHARED))); | |
560 | set_huge_pte_at(mm, address, ptep, new_pte); | |
561 | ||
562 | if (write_access && !(vma->vm_flags & VM_SHARED)) { | |
563 | /* Optimization, do the COW without a second fault */ | |
564 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte); | |
565 | } | |
566 | ||
ac9b9c66 | 567 | spin_unlock(&mm->page_table_lock); |
4c887265 AL |
568 | unlock_page(page); |
569 | out: | |
ac9b9c66 | 570 | return ret; |
4c887265 AL |
571 | |
572 | backout: | |
573 | spin_unlock(&mm->page_table_lock); | |
574 | hugetlb_put_quota(mapping); | |
575 | unlock_page(page); | |
576 | put_page(page); | |
577 | goto out; | |
ac9b9c66 HD |
578 | } |
579 | ||
86e5216f AL |
580 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
581 | unsigned long address, int write_access) | |
582 | { | |
583 | pte_t *ptep; | |
584 | pte_t entry; | |
1e8f889b | 585 | int ret; |
3935baa9 | 586 | static DEFINE_MUTEX(hugetlb_instantiation_mutex); |
86e5216f AL |
587 | |
588 | ptep = huge_pte_alloc(mm, address); | |
589 | if (!ptep) | |
590 | return VM_FAULT_OOM; | |
591 | ||
3935baa9 DG |
592 | /* |
593 | * Serialize hugepage allocation and instantiation, so that we don't | |
594 | * get spurious allocation failures if two CPUs race to instantiate | |
595 | * the same page in the page cache. | |
596 | */ | |
597 | mutex_lock(&hugetlb_instantiation_mutex); | |
86e5216f | 598 | entry = *ptep; |
3935baa9 DG |
599 | if (pte_none(entry)) { |
600 | ret = hugetlb_no_page(mm, vma, address, ptep, write_access); | |
601 | mutex_unlock(&hugetlb_instantiation_mutex); | |
602 | return ret; | |
603 | } | |
86e5216f | 604 | |
83c54070 | 605 | ret = 0; |
1e8f889b DG |
606 | |
607 | spin_lock(&mm->page_table_lock); | |
608 | /* Check for a racing update before calling hugetlb_cow */ | |
609 | if (likely(pte_same(entry, *ptep))) | |
610 | if (write_access && !pte_write(entry)) | |
611 | ret = hugetlb_cow(mm, vma, address, ptep, entry); | |
612 | spin_unlock(&mm->page_table_lock); | |
3935baa9 | 613 | mutex_unlock(&hugetlb_instantiation_mutex); |
1e8f889b DG |
614 | |
615 | return ret; | |
86e5216f AL |
616 | } |
617 | ||
63551ae0 DG |
618 | int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
619 | struct page **pages, struct vm_area_struct **vmas, | |
620 | unsigned long *position, int *length, int i) | |
621 | { | |
d5d4b0aa CK |
622 | unsigned long pfn_offset; |
623 | unsigned long vaddr = *position; | |
63551ae0 DG |
624 | int remainder = *length; |
625 | ||
1c59827d | 626 | spin_lock(&mm->page_table_lock); |
63551ae0 | 627 | while (vaddr < vma->vm_end && remainder) { |
4c887265 AL |
628 | pte_t *pte; |
629 | struct page *page; | |
63551ae0 | 630 | |
4c887265 AL |
631 | /* |
632 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
633 | * each hugepage. We have to make * sure we get the | |
634 | * first, for the page indexing below to work. | |
635 | */ | |
636 | pte = huge_pte_offset(mm, vaddr & HPAGE_MASK); | |
63551ae0 | 637 | |
4c887265 AL |
638 | if (!pte || pte_none(*pte)) { |
639 | int ret; | |
63551ae0 | 640 | |
4c887265 AL |
641 | spin_unlock(&mm->page_table_lock); |
642 | ret = hugetlb_fault(mm, vma, vaddr, 0); | |
643 | spin_lock(&mm->page_table_lock); | |
83c54070 | 644 | if (!(ret & VM_FAULT_MAJOR)) |
4c887265 | 645 | continue; |
63551ae0 | 646 | |
4c887265 AL |
647 | remainder = 0; |
648 | if (!i) | |
649 | i = -EFAULT; | |
650 | break; | |
651 | } | |
652 | ||
d5d4b0aa CK |
653 | pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT; |
654 | page = pte_page(*pte); | |
655 | same_page: | |
d6692183 CK |
656 | if (pages) { |
657 | get_page(page); | |
d5d4b0aa | 658 | pages[i] = page + pfn_offset; |
d6692183 | 659 | } |
63551ae0 DG |
660 | |
661 | if (vmas) | |
662 | vmas[i] = vma; | |
663 | ||
664 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 665 | ++pfn_offset; |
63551ae0 DG |
666 | --remainder; |
667 | ++i; | |
d5d4b0aa CK |
668 | if (vaddr < vma->vm_end && remainder && |
669 | pfn_offset < HPAGE_SIZE/PAGE_SIZE) { | |
670 | /* | |
671 | * We use pfn_offset to avoid touching the pageframes | |
672 | * of this compound page. | |
673 | */ | |
674 | goto same_page; | |
675 | } | |
63551ae0 | 676 | } |
1c59827d | 677 | spin_unlock(&mm->page_table_lock); |
63551ae0 DG |
678 | *length = remainder; |
679 | *position = vaddr; | |
680 | ||
681 | return i; | |
682 | } | |
8f860591 ZY |
683 | |
684 | void hugetlb_change_protection(struct vm_area_struct *vma, | |
685 | unsigned long address, unsigned long end, pgprot_t newprot) | |
686 | { | |
687 | struct mm_struct *mm = vma->vm_mm; | |
688 | unsigned long start = address; | |
689 | pte_t *ptep; | |
690 | pte_t pte; | |
691 | ||
692 | BUG_ON(address >= end); | |
693 | flush_cache_range(vma, address, end); | |
694 | ||
39dde65c | 695 | spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); |
8f860591 ZY |
696 | spin_lock(&mm->page_table_lock); |
697 | for (; address < end; address += HPAGE_SIZE) { | |
698 | ptep = huge_pte_offset(mm, address); | |
699 | if (!ptep) | |
700 | continue; | |
39dde65c CK |
701 | if (huge_pmd_unshare(mm, &address, ptep)) |
702 | continue; | |
8f860591 ZY |
703 | if (!pte_none(*ptep)) { |
704 | pte = huge_ptep_get_and_clear(mm, address, ptep); | |
705 | pte = pte_mkhuge(pte_modify(pte, newprot)); | |
706 | set_huge_pte_at(mm, address, ptep, pte); | |
707 | lazy_mmu_prot_update(pte); | |
708 | } | |
709 | } | |
710 | spin_unlock(&mm->page_table_lock); | |
39dde65c | 711 | spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); |
8f860591 ZY |
712 | |
713 | flush_tlb_range(vma, start, end); | |
714 | } | |
715 | ||
a43a8c39 CK |
716 | struct file_region { |
717 | struct list_head link; | |
718 | long from; | |
719 | long to; | |
720 | }; | |
721 | ||
722 | static long region_add(struct list_head *head, long f, long t) | |
723 | { | |
724 | struct file_region *rg, *nrg, *trg; | |
725 | ||
726 | /* Locate the region we are either in or before. */ | |
727 | list_for_each_entry(rg, head, link) | |
728 | if (f <= rg->to) | |
729 | break; | |
730 | ||
731 | /* Round our left edge to the current segment if it encloses us. */ | |
732 | if (f > rg->from) | |
733 | f = rg->from; | |
734 | ||
735 | /* Check for and consume any regions we now overlap with. */ | |
736 | nrg = rg; | |
737 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
738 | if (&rg->link == head) | |
739 | break; | |
740 | if (rg->from > t) | |
741 | break; | |
742 | ||
743 | /* If this area reaches higher then extend our area to | |
744 | * include it completely. If this is not the first area | |
745 | * which we intend to reuse, free it. */ | |
746 | if (rg->to > t) | |
747 | t = rg->to; | |
748 | if (rg != nrg) { | |
749 | list_del(&rg->link); | |
750 | kfree(rg); | |
751 | } | |
752 | } | |
753 | nrg->from = f; | |
754 | nrg->to = t; | |
755 | return 0; | |
756 | } | |
757 | ||
758 | static long region_chg(struct list_head *head, long f, long t) | |
759 | { | |
760 | struct file_region *rg, *nrg; | |
761 | long chg = 0; | |
762 | ||
763 | /* Locate the region we are before or in. */ | |
764 | list_for_each_entry(rg, head, link) | |
765 | if (f <= rg->to) | |
766 | break; | |
767 | ||
768 | /* If we are below the current region then a new region is required. | |
769 | * Subtle, allocate a new region at the position but make it zero | |
770 | * size such that we can guarentee to record the reservation. */ | |
771 | if (&rg->link == head || t < rg->from) { | |
772 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
773 | if (nrg == 0) | |
774 | return -ENOMEM; | |
775 | nrg->from = f; | |
776 | nrg->to = f; | |
777 | INIT_LIST_HEAD(&nrg->link); | |
778 | list_add(&nrg->link, rg->link.prev); | |
779 | ||
780 | return t - f; | |
781 | } | |
782 | ||
783 | /* Round our left edge to the current segment if it encloses us. */ | |
784 | if (f > rg->from) | |
785 | f = rg->from; | |
786 | chg = t - f; | |
787 | ||
788 | /* Check for and consume any regions we now overlap with. */ | |
789 | list_for_each_entry(rg, rg->link.prev, link) { | |
790 | if (&rg->link == head) | |
791 | break; | |
792 | if (rg->from > t) | |
793 | return chg; | |
794 | ||
795 | /* We overlap with this area, if it extends futher than | |
796 | * us then we must extend ourselves. Account for its | |
797 | * existing reservation. */ | |
798 | if (rg->to > t) { | |
799 | chg += rg->to - t; | |
800 | t = rg->to; | |
801 | } | |
802 | chg -= rg->to - rg->from; | |
803 | } | |
804 | return chg; | |
805 | } | |
806 | ||
807 | static long region_truncate(struct list_head *head, long end) | |
808 | { | |
809 | struct file_region *rg, *trg; | |
810 | long chg = 0; | |
811 | ||
812 | /* Locate the region we are either in or before. */ | |
813 | list_for_each_entry(rg, head, link) | |
814 | if (end <= rg->to) | |
815 | break; | |
816 | if (&rg->link == head) | |
817 | return 0; | |
818 | ||
819 | /* If we are in the middle of a region then adjust it. */ | |
820 | if (end > rg->from) { | |
821 | chg = rg->to - end; | |
822 | rg->to = end; | |
823 | rg = list_entry(rg->link.next, typeof(*rg), link); | |
824 | } | |
825 | ||
826 | /* Drop any remaining regions. */ | |
827 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
828 | if (&rg->link == head) | |
829 | break; | |
830 | chg += rg->to - rg->from; | |
831 | list_del(&rg->link); | |
832 | kfree(rg); | |
833 | } | |
834 | return chg; | |
835 | } | |
836 | ||
837 | static int hugetlb_acct_memory(long delta) | |
838 | { | |
839 | int ret = -ENOMEM; | |
840 | ||
841 | spin_lock(&hugetlb_lock); | |
842 | if ((delta + resv_huge_pages) <= free_huge_pages) { | |
843 | resv_huge_pages += delta; | |
844 | ret = 0; | |
845 | } | |
846 | spin_unlock(&hugetlb_lock); | |
847 | return ret; | |
848 | } | |
849 | ||
850 | int hugetlb_reserve_pages(struct inode *inode, long from, long to) | |
851 | { | |
852 | long ret, chg; | |
853 | ||
854 | chg = region_chg(&inode->i_mapping->private_list, from, to); | |
855 | if (chg < 0) | |
856 | return chg; | |
8a630112 KC |
857 | /* |
858 | * When cpuset is configured, it breaks the strict hugetlb page | |
859 | * reservation as the accounting is done on a global variable. Such | |
860 | * reservation is completely rubbish in the presence of cpuset because | |
861 | * the reservation is not checked against page availability for the | |
862 | * current cpuset. Application can still potentially OOM'ed by kernel | |
863 | * with lack of free htlb page in cpuset that the task is in. | |
864 | * Attempt to enforce strict accounting with cpuset is almost | |
865 | * impossible (or too ugly) because cpuset is too fluid that | |
866 | * task or memory node can be dynamically moved between cpusets. | |
867 | * | |
868 | * The change of semantics for shared hugetlb mapping with cpuset is | |
869 | * undesirable. However, in order to preserve some of the semantics, | |
870 | * we fall back to check against current free page availability as | |
871 | * a best attempt and hopefully to minimize the impact of changing | |
872 | * semantics that cpuset has. | |
873 | */ | |
874 | if (chg > cpuset_mems_nr(free_huge_pages_node)) | |
875 | return -ENOMEM; | |
876 | ||
a43a8c39 CK |
877 | ret = hugetlb_acct_memory(chg); |
878 | if (ret < 0) | |
879 | return ret; | |
880 | region_add(&inode->i_mapping->private_list, from, to); | |
881 | return 0; | |
882 | } | |
883 | ||
884 | void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) | |
885 | { | |
886 | long chg = region_truncate(&inode->i_mapping->private_list, offset); | |
887 | hugetlb_acct_memory(freed - chg); | |
888 | } |