]>
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; |
7893d1d5 | 26 | static unsigned long surplus_huge_pages; |
064d9efe | 27 | static unsigned long nr_overcommit_huge_pages; |
1da177e4 | 28 | unsigned long max_huge_pages; |
064d9efe | 29 | unsigned long sysctl_overcommit_huge_pages; |
1da177e4 LT |
30 | static struct list_head hugepage_freelists[MAX_NUMNODES]; |
31 | static unsigned int nr_huge_pages_node[MAX_NUMNODES]; | |
32 | static unsigned int free_huge_pages_node[MAX_NUMNODES]; | |
7893d1d5 | 33 | static unsigned int surplus_huge_pages_node[MAX_NUMNODES]; |
396faf03 MG |
34 | static gfp_t htlb_alloc_mask = GFP_HIGHUSER; |
35 | unsigned long hugepages_treat_as_movable; | |
63b4613c | 36 | static int hugetlb_next_nid; |
396faf03 | 37 | |
3935baa9 DG |
38 | /* |
39 | * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages | |
40 | */ | |
41 | static DEFINE_SPINLOCK(hugetlb_lock); | |
0bd0f9fb | 42 | |
a1e78772 MG |
43 | /* |
44 | * These helpers are used to track how many pages are reserved for | |
45 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
46 | * is guaranteed to have their future faults succeed. | |
47 | * | |
48 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
49 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
50 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
51 | * chance of the global counters getting corrupted as a result of the values. | |
52 | */ | |
53 | static unsigned long vma_resv_huge_pages(struct vm_area_struct *vma) | |
54 | { | |
55 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
56 | if (!(vma->vm_flags & VM_SHARED)) | |
57 | return (unsigned long)vma->vm_private_data; | |
58 | return 0; | |
59 | } | |
60 | ||
61 | static void set_vma_resv_huge_pages(struct vm_area_struct *vma, | |
62 | unsigned long reserve) | |
63 | { | |
64 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
65 | VM_BUG_ON(vma->vm_flags & VM_SHARED); | |
66 | ||
67 | vma->vm_private_data = (void *)reserve; | |
68 | } | |
69 | ||
70 | /* Decrement the reserved pages in the hugepage pool by one */ | |
71 | static void decrement_hugepage_resv_vma(struct vm_area_struct *vma) | |
72 | { | |
73 | if (vma->vm_flags & VM_SHARED) { | |
74 | /* Shared mappings always use reserves */ | |
75 | resv_huge_pages--; | |
76 | } else { | |
77 | /* | |
78 | * Only the process that called mmap() has reserves for | |
79 | * private mappings. | |
80 | */ | |
81 | if (vma_resv_huge_pages(vma)) { | |
82 | resv_huge_pages--; | |
83 | reserve = (unsigned long)vma->vm_private_data - 1; | |
84 | vma->vm_private_data = (void *)reserve; | |
85 | } | |
86 | } | |
87 | } | |
88 | ||
89 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) | |
90 | { | |
91 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
92 | if (!(vma->vm_flags & VM_SHARED)) | |
93 | vma->vm_private_data = (void *)0; | |
94 | } | |
95 | ||
96 | /* Returns true if the VMA has associated reserve pages */ | |
97 | static int vma_has_private_reserves(struct vm_area_struct *vma) | |
98 | { | |
99 | if (vma->vm_flags & VM_SHARED) | |
100 | return 0; | |
101 | if (!vma_resv_huge_pages(vma)) | |
102 | return 0; | |
103 | return 1; | |
104 | } | |
105 | ||
79ac6ba4 DG |
106 | static void clear_huge_page(struct page *page, unsigned long addr) |
107 | { | |
108 | int i; | |
109 | ||
110 | might_sleep(); | |
111 | for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) { | |
112 | cond_resched(); | |
281e0e3b | 113 | clear_user_highpage(page + i, addr + i * PAGE_SIZE); |
79ac6ba4 DG |
114 | } |
115 | } | |
116 | ||
117 | static void copy_huge_page(struct page *dst, struct page *src, | |
9de455b2 | 118 | unsigned long addr, struct vm_area_struct *vma) |
79ac6ba4 DG |
119 | { |
120 | int i; | |
121 | ||
122 | might_sleep(); | |
123 | for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) { | |
124 | cond_resched(); | |
9de455b2 | 125 | copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma); |
79ac6ba4 DG |
126 | } |
127 | } | |
128 | ||
1da177e4 LT |
129 | static void enqueue_huge_page(struct page *page) |
130 | { | |
131 | int nid = page_to_nid(page); | |
132 | list_add(&page->lru, &hugepage_freelists[nid]); | |
133 | free_huge_pages++; | |
134 | free_huge_pages_node[nid]++; | |
135 | } | |
136 | ||
348e1e04 NA |
137 | static struct page *dequeue_huge_page(void) |
138 | { | |
139 | int nid; | |
140 | struct page *page = NULL; | |
141 | ||
142 | for (nid = 0; nid < MAX_NUMNODES; ++nid) { | |
143 | if (!list_empty(&hugepage_freelists[nid])) { | |
144 | page = list_entry(hugepage_freelists[nid].next, | |
145 | struct page, lru); | |
146 | list_del(&page->lru); | |
147 | free_huge_pages--; | |
148 | free_huge_pages_node[nid]--; | |
149 | break; | |
150 | } | |
151 | } | |
152 | return page; | |
153 | } | |
154 | ||
155 | static struct page *dequeue_huge_page_vma(struct vm_area_struct *vma, | |
5da7ca86 | 156 | unsigned long address) |
1da177e4 | 157 | { |
31a5c6e4 | 158 | int nid; |
1da177e4 | 159 | struct page *page = NULL; |
480eccf9 | 160 | struct mempolicy *mpol; |
19770b32 | 161 | nodemask_t *nodemask; |
396faf03 | 162 | struct zonelist *zonelist = huge_zonelist(vma, address, |
19770b32 | 163 | htlb_alloc_mask, &mpol, &nodemask); |
dd1a239f MG |
164 | struct zone *zone; |
165 | struct zoneref *z; | |
1da177e4 | 166 | |
a1e78772 MG |
167 | /* |
168 | * A child process with MAP_PRIVATE mappings created by their parent | |
169 | * have no page reserves. This check ensures that reservations are | |
170 | * not "stolen". The child may still get SIGKILLed | |
171 | */ | |
172 | if (!vma_has_private_reserves(vma) && | |
173 | free_huge_pages - resv_huge_pages == 0) | |
174 | return NULL; | |
175 | ||
19770b32 MG |
176 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
177 | MAX_NR_ZONES - 1, nodemask) { | |
54a6eb5c MG |
178 | nid = zone_to_nid(zone); |
179 | if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) && | |
3abf7afd AM |
180 | !list_empty(&hugepage_freelists[nid])) { |
181 | page = list_entry(hugepage_freelists[nid].next, | |
182 | struct page, lru); | |
183 | list_del(&page->lru); | |
184 | free_huge_pages--; | |
185 | free_huge_pages_node[nid]--; | |
a1e78772 MG |
186 | decrement_hugepage_resv_vma(vma); |
187 | ||
5ab3ee7b | 188 | break; |
3abf7afd | 189 | } |
1da177e4 | 190 | } |
52cd3b07 | 191 | mpol_cond_put(mpol); |
1da177e4 LT |
192 | return page; |
193 | } | |
194 | ||
6af2acb6 AL |
195 | static void update_and_free_page(struct page *page) |
196 | { | |
197 | int i; | |
198 | nr_huge_pages--; | |
199 | nr_huge_pages_node[page_to_nid(page)]--; | |
200 | for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) { | |
201 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced | | |
202 | 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved | | |
203 | 1 << PG_private | 1<< PG_writeback); | |
204 | } | |
205 | set_compound_page_dtor(page, NULL); | |
206 | set_page_refcounted(page); | |
7f2e9525 | 207 | arch_release_hugepage(page); |
6af2acb6 AL |
208 | __free_pages(page, HUGETLB_PAGE_ORDER); |
209 | } | |
210 | ||
27a85ef1 DG |
211 | static void free_huge_page(struct page *page) |
212 | { | |
7893d1d5 | 213 | int nid = page_to_nid(page); |
c79fb75e | 214 | struct address_space *mapping; |
27a85ef1 | 215 | |
c79fb75e | 216 | mapping = (struct address_space *) page_private(page); |
e5df70ab | 217 | set_page_private(page, 0); |
7893d1d5 | 218 | BUG_ON(page_count(page)); |
27a85ef1 DG |
219 | INIT_LIST_HEAD(&page->lru); |
220 | ||
221 | spin_lock(&hugetlb_lock); | |
7893d1d5 AL |
222 | if (surplus_huge_pages_node[nid]) { |
223 | update_and_free_page(page); | |
224 | surplus_huge_pages--; | |
225 | surplus_huge_pages_node[nid]--; | |
226 | } else { | |
227 | enqueue_huge_page(page); | |
228 | } | |
27a85ef1 | 229 | spin_unlock(&hugetlb_lock); |
c79fb75e | 230 | if (mapping) |
9a119c05 | 231 | hugetlb_put_quota(mapping, 1); |
27a85ef1 DG |
232 | } |
233 | ||
7893d1d5 AL |
234 | /* |
235 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
236 | * balanced by operating on them in a round-robin fashion. | |
237 | * Returns 1 if an adjustment was made. | |
238 | */ | |
239 | static int adjust_pool_surplus(int delta) | |
240 | { | |
241 | static int prev_nid; | |
242 | int nid = prev_nid; | |
243 | int ret = 0; | |
244 | ||
245 | VM_BUG_ON(delta != -1 && delta != 1); | |
246 | do { | |
247 | nid = next_node(nid, node_online_map); | |
248 | if (nid == MAX_NUMNODES) | |
249 | nid = first_node(node_online_map); | |
250 | ||
251 | /* To shrink on this node, there must be a surplus page */ | |
252 | if (delta < 0 && !surplus_huge_pages_node[nid]) | |
253 | continue; | |
254 | /* Surplus cannot exceed the total number of pages */ | |
255 | if (delta > 0 && surplus_huge_pages_node[nid] >= | |
256 | nr_huge_pages_node[nid]) | |
257 | continue; | |
258 | ||
259 | surplus_huge_pages += delta; | |
260 | surplus_huge_pages_node[nid] += delta; | |
261 | ret = 1; | |
262 | break; | |
263 | } while (nid != prev_nid); | |
264 | ||
265 | prev_nid = nid; | |
266 | return ret; | |
267 | } | |
268 | ||
63b4613c | 269 | static struct page *alloc_fresh_huge_page_node(int nid) |
1da177e4 | 270 | { |
1da177e4 | 271 | struct page *page; |
f96efd58 | 272 | |
63b4613c | 273 | page = alloc_pages_node(nid, |
551883ae NA |
274 | htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| |
275 | __GFP_REPEAT|__GFP_NOWARN, | |
63b4613c | 276 | HUGETLB_PAGE_ORDER); |
1da177e4 | 277 | if (page) { |
7f2e9525 GS |
278 | if (arch_prepare_hugepage(page)) { |
279 | __free_pages(page, HUGETLB_PAGE_ORDER); | |
7b8ee84d | 280 | return NULL; |
7f2e9525 | 281 | } |
33f2ef89 | 282 | set_compound_page_dtor(page, free_huge_page); |
0bd0f9fb | 283 | spin_lock(&hugetlb_lock); |
1da177e4 | 284 | nr_huge_pages++; |
63b4613c | 285 | nr_huge_pages_node[nid]++; |
0bd0f9fb | 286 | spin_unlock(&hugetlb_lock); |
a482289d | 287 | put_page(page); /* free it into the hugepage allocator */ |
1da177e4 | 288 | } |
63b4613c NA |
289 | |
290 | return page; | |
291 | } | |
292 | ||
293 | static int alloc_fresh_huge_page(void) | |
294 | { | |
295 | struct page *page; | |
296 | int start_nid; | |
297 | int next_nid; | |
298 | int ret = 0; | |
299 | ||
300 | start_nid = hugetlb_next_nid; | |
301 | ||
302 | do { | |
303 | page = alloc_fresh_huge_page_node(hugetlb_next_nid); | |
304 | if (page) | |
305 | ret = 1; | |
306 | /* | |
307 | * Use a helper variable to find the next node and then | |
308 | * copy it back to hugetlb_next_nid afterwards: | |
309 | * otherwise there's a window in which a racer might | |
310 | * pass invalid nid MAX_NUMNODES to alloc_pages_node. | |
311 | * But we don't need to use a spin_lock here: it really | |
312 | * doesn't matter if occasionally a racer chooses the | |
313 | * same nid as we do. Move nid forward in the mask even | |
314 | * if we just successfully allocated a hugepage so that | |
315 | * the next caller gets hugepages on the next node. | |
316 | */ | |
317 | next_nid = next_node(hugetlb_next_nid, node_online_map); | |
318 | if (next_nid == MAX_NUMNODES) | |
319 | next_nid = first_node(node_online_map); | |
320 | hugetlb_next_nid = next_nid; | |
321 | } while (!page && hugetlb_next_nid != start_nid); | |
322 | ||
3b116300 AL |
323 | if (ret) |
324 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
325 | else | |
326 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
327 | ||
63b4613c | 328 | return ret; |
1da177e4 LT |
329 | } |
330 | ||
7893d1d5 AL |
331 | static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma, |
332 | unsigned long address) | |
333 | { | |
334 | struct page *page; | |
d1c3fb1f | 335 | unsigned int nid; |
7893d1d5 | 336 | |
d1c3fb1f NA |
337 | /* |
338 | * Assume we will successfully allocate the surplus page to | |
339 | * prevent racing processes from causing the surplus to exceed | |
340 | * overcommit | |
341 | * | |
342 | * This however introduces a different race, where a process B | |
343 | * tries to grow the static hugepage pool while alloc_pages() is | |
344 | * called by process A. B will only examine the per-node | |
345 | * counters in determining if surplus huge pages can be | |
346 | * converted to normal huge pages in adjust_pool_surplus(). A | |
347 | * won't be able to increment the per-node counter, until the | |
348 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
349 | * no more huge pages can be converted from surplus to normal | |
350 | * state (and doesn't try to convert again). Thus, we have a | |
351 | * case where a surplus huge page exists, the pool is grown, and | |
352 | * the surplus huge page still exists after, even though it | |
353 | * should just have been converted to a normal huge page. This | |
354 | * does not leak memory, though, as the hugepage will be freed | |
355 | * once it is out of use. It also does not allow the counters to | |
356 | * go out of whack in adjust_pool_surplus() as we don't modify | |
357 | * the node values until we've gotten the hugepage and only the | |
358 | * per-node value is checked there. | |
359 | */ | |
360 | spin_lock(&hugetlb_lock); | |
361 | if (surplus_huge_pages >= nr_overcommit_huge_pages) { | |
362 | spin_unlock(&hugetlb_lock); | |
363 | return NULL; | |
364 | } else { | |
365 | nr_huge_pages++; | |
366 | surplus_huge_pages++; | |
367 | } | |
368 | spin_unlock(&hugetlb_lock); | |
369 | ||
551883ae NA |
370 | page = alloc_pages(htlb_alloc_mask|__GFP_COMP| |
371 | __GFP_REPEAT|__GFP_NOWARN, | |
7893d1d5 | 372 | HUGETLB_PAGE_ORDER); |
d1c3fb1f NA |
373 | |
374 | spin_lock(&hugetlb_lock); | |
7893d1d5 | 375 | if (page) { |
2668db91 AL |
376 | /* |
377 | * This page is now managed by the hugetlb allocator and has | |
378 | * no users -- drop the buddy allocator's reference. | |
379 | */ | |
380 | put_page_testzero(page); | |
381 | VM_BUG_ON(page_count(page)); | |
d1c3fb1f | 382 | nid = page_to_nid(page); |
7893d1d5 | 383 | set_compound_page_dtor(page, free_huge_page); |
d1c3fb1f NA |
384 | /* |
385 | * We incremented the global counters already | |
386 | */ | |
387 | nr_huge_pages_node[nid]++; | |
388 | surplus_huge_pages_node[nid]++; | |
3b116300 | 389 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f NA |
390 | } else { |
391 | nr_huge_pages--; | |
392 | surplus_huge_pages--; | |
3b116300 | 393 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 394 | } |
d1c3fb1f | 395 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
396 | |
397 | return page; | |
398 | } | |
399 | ||
e4e574b7 AL |
400 | /* |
401 | * Increase the hugetlb pool such that it can accomodate a reservation | |
402 | * of size 'delta'. | |
403 | */ | |
404 | static int gather_surplus_pages(int delta) | |
405 | { | |
406 | struct list_head surplus_list; | |
407 | struct page *page, *tmp; | |
408 | int ret, i; | |
409 | int needed, allocated; | |
410 | ||
411 | needed = (resv_huge_pages + delta) - free_huge_pages; | |
ac09b3a1 AL |
412 | if (needed <= 0) { |
413 | resv_huge_pages += delta; | |
e4e574b7 | 414 | return 0; |
ac09b3a1 | 415 | } |
e4e574b7 AL |
416 | |
417 | allocated = 0; | |
418 | INIT_LIST_HEAD(&surplus_list); | |
419 | ||
420 | ret = -ENOMEM; | |
421 | retry: | |
422 | spin_unlock(&hugetlb_lock); | |
423 | for (i = 0; i < needed; i++) { | |
424 | page = alloc_buddy_huge_page(NULL, 0); | |
425 | if (!page) { | |
426 | /* | |
427 | * We were not able to allocate enough pages to | |
428 | * satisfy the entire reservation so we free what | |
429 | * we've allocated so far. | |
430 | */ | |
431 | spin_lock(&hugetlb_lock); | |
432 | needed = 0; | |
433 | goto free; | |
434 | } | |
435 | ||
436 | list_add(&page->lru, &surplus_list); | |
437 | } | |
438 | allocated += needed; | |
439 | ||
440 | /* | |
441 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
442 | * because either resv_huge_pages or free_huge_pages may have changed. | |
443 | */ | |
444 | spin_lock(&hugetlb_lock); | |
445 | needed = (resv_huge_pages + delta) - (free_huge_pages + allocated); | |
446 | if (needed > 0) | |
447 | goto retry; | |
448 | ||
449 | /* | |
450 | * The surplus_list now contains _at_least_ the number of extra pages | |
451 | * needed to accomodate the reservation. Add the appropriate number | |
452 | * of pages to the hugetlb pool and free the extras back to the buddy | |
ac09b3a1 AL |
453 | * allocator. Commit the entire reservation here to prevent another |
454 | * process from stealing the pages as they are added to the pool but | |
455 | * before they are reserved. | |
e4e574b7 AL |
456 | */ |
457 | needed += allocated; | |
ac09b3a1 | 458 | resv_huge_pages += delta; |
e4e574b7 AL |
459 | ret = 0; |
460 | free: | |
19fc3f0a | 461 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 462 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
463 | if ((--needed) < 0) |
464 | break; | |
e4e574b7 | 465 | list_del(&page->lru); |
19fc3f0a AL |
466 | enqueue_huge_page(page); |
467 | } | |
468 | ||
469 | /* Free unnecessary surplus pages to the buddy allocator */ | |
470 | if (!list_empty(&surplus_list)) { | |
471 | spin_unlock(&hugetlb_lock); | |
472 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { | |
473 | list_del(&page->lru); | |
af767cbd | 474 | /* |
2668db91 AL |
475 | * The page has a reference count of zero already, so |
476 | * call free_huge_page directly instead of using | |
477 | * put_page. This must be done with hugetlb_lock | |
af767cbd AL |
478 | * unlocked which is safe because free_huge_page takes |
479 | * hugetlb_lock before deciding how to free the page. | |
480 | */ | |
2668db91 | 481 | free_huge_page(page); |
af767cbd | 482 | } |
19fc3f0a | 483 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
484 | } |
485 | ||
486 | return ret; | |
487 | } | |
488 | ||
489 | /* | |
490 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
491 | * allocated to satisfy the reservation must be explicitly freed if they were | |
492 | * never used. | |
493 | */ | |
8cde045c | 494 | static void return_unused_surplus_pages(unsigned long unused_resv_pages) |
e4e574b7 AL |
495 | { |
496 | static int nid = -1; | |
497 | struct page *page; | |
498 | unsigned long nr_pages; | |
499 | ||
11320d17 NA |
500 | /* |
501 | * We want to release as many surplus pages as possible, spread | |
502 | * evenly across all nodes. Iterate across all nodes until we | |
503 | * can no longer free unreserved surplus pages. This occurs when | |
504 | * the nodes with surplus pages have no free pages. | |
505 | */ | |
506 | unsigned long remaining_iterations = num_online_nodes(); | |
507 | ||
ac09b3a1 AL |
508 | /* Uncommit the reservation */ |
509 | resv_huge_pages -= unused_resv_pages; | |
510 | ||
e4e574b7 AL |
511 | nr_pages = min(unused_resv_pages, surplus_huge_pages); |
512 | ||
11320d17 | 513 | while (remaining_iterations-- && nr_pages) { |
e4e574b7 AL |
514 | nid = next_node(nid, node_online_map); |
515 | if (nid == MAX_NUMNODES) | |
516 | nid = first_node(node_online_map); | |
517 | ||
518 | if (!surplus_huge_pages_node[nid]) | |
519 | continue; | |
520 | ||
521 | if (!list_empty(&hugepage_freelists[nid])) { | |
522 | page = list_entry(hugepage_freelists[nid].next, | |
523 | struct page, lru); | |
524 | list_del(&page->lru); | |
525 | update_and_free_page(page); | |
526 | free_huge_pages--; | |
527 | free_huge_pages_node[nid]--; | |
528 | surplus_huge_pages--; | |
529 | surplus_huge_pages_node[nid]--; | |
530 | nr_pages--; | |
11320d17 | 531 | remaining_iterations = num_online_nodes(); |
e4e574b7 AL |
532 | } |
533 | } | |
534 | } | |
535 | ||
a1e78772 MG |
536 | static struct page *alloc_huge_page(struct vm_area_struct *vma, |
537 | unsigned long addr) | |
1da177e4 | 538 | { |
348ea204 | 539 | struct page *page; |
a1e78772 MG |
540 | struct address_space *mapping = vma->vm_file->f_mapping; |
541 | struct inode *inode = mapping->host; | |
542 | unsigned int chg = 0; | |
543 | ||
544 | /* | |
545 | * Processes that did not create the mapping will have no reserves and | |
546 | * will not have accounted against quota. Check that the quota can be | |
547 | * made before satisfying the allocation | |
548 | */ | |
549 | if (!vma_has_private_reserves(vma)) { | |
550 | chg = 1; | |
551 | if (hugetlb_get_quota(inode->i_mapping, chg)) | |
552 | return ERR_PTR(-ENOSPC); | |
553 | } | |
1da177e4 LT |
554 | |
555 | spin_lock(&hugetlb_lock); | |
348e1e04 | 556 | page = dequeue_huge_page_vma(vma, addr); |
1da177e4 | 557 | spin_unlock(&hugetlb_lock); |
b45b5bd6 | 558 | |
68842c9b | 559 | if (!page) { |
7893d1d5 | 560 | page = alloc_buddy_huge_page(vma, addr); |
68842c9b | 561 | if (!page) { |
a1e78772 | 562 | hugetlb_put_quota(inode->i_mapping, chg); |
68842c9b KC |
563 | return ERR_PTR(-VM_FAULT_OOM); |
564 | } | |
565 | } | |
348ea204 | 566 | |
a1e78772 MG |
567 | set_page_refcounted(page); |
568 | set_page_private(page, (unsigned long) mapping); | |
90d8b7e6 | 569 | |
90d8b7e6 | 570 | return page; |
b45b5bd6 DG |
571 | } |
572 | ||
1da177e4 LT |
573 | static int __init hugetlb_init(void) |
574 | { | |
575 | unsigned long i; | |
1da177e4 | 576 | |
3c726f8d BH |
577 | if (HPAGE_SHIFT == 0) |
578 | return 0; | |
579 | ||
1da177e4 LT |
580 | for (i = 0; i < MAX_NUMNODES; ++i) |
581 | INIT_LIST_HEAD(&hugepage_freelists[i]); | |
582 | ||
63b4613c NA |
583 | hugetlb_next_nid = first_node(node_online_map); |
584 | ||
1da177e4 | 585 | for (i = 0; i < max_huge_pages; ++i) { |
a482289d | 586 | if (!alloc_fresh_huge_page()) |
1da177e4 | 587 | break; |
1da177e4 LT |
588 | } |
589 | max_huge_pages = free_huge_pages = nr_huge_pages = i; | |
590 | printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages); | |
591 | return 0; | |
592 | } | |
593 | module_init(hugetlb_init); | |
594 | ||
595 | static int __init hugetlb_setup(char *s) | |
596 | { | |
597 | if (sscanf(s, "%lu", &max_huge_pages) <= 0) | |
598 | max_huge_pages = 0; | |
599 | return 1; | |
600 | } | |
601 | __setup("hugepages=", hugetlb_setup); | |
602 | ||
8a630112 KC |
603 | static unsigned int cpuset_mems_nr(unsigned int *array) |
604 | { | |
605 | int node; | |
606 | unsigned int nr = 0; | |
607 | ||
608 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
609 | nr += array[node]; | |
610 | ||
611 | return nr; | |
612 | } | |
613 | ||
1da177e4 | 614 | #ifdef CONFIG_SYSCTL |
1da177e4 LT |
615 | #ifdef CONFIG_HIGHMEM |
616 | static void try_to_free_low(unsigned long count) | |
617 | { | |
4415cc8d CL |
618 | int i; |
619 | ||
1da177e4 LT |
620 | for (i = 0; i < MAX_NUMNODES; ++i) { |
621 | struct page *page, *next; | |
622 | list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) { | |
6b0c880d AL |
623 | if (count >= nr_huge_pages) |
624 | return; | |
1da177e4 LT |
625 | if (PageHighMem(page)) |
626 | continue; | |
627 | list_del(&page->lru); | |
628 | update_and_free_page(page); | |
1da177e4 | 629 | free_huge_pages--; |
4415cc8d | 630 | free_huge_pages_node[page_to_nid(page)]--; |
1da177e4 LT |
631 | } |
632 | } | |
633 | } | |
634 | #else | |
635 | static inline void try_to_free_low(unsigned long count) | |
636 | { | |
637 | } | |
638 | #endif | |
639 | ||
7893d1d5 | 640 | #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages) |
1da177e4 LT |
641 | static unsigned long set_max_huge_pages(unsigned long count) |
642 | { | |
7893d1d5 | 643 | unsigned long min_count, ret; |
1da177e4 | 644 | |
7893d1d5 AL |
645 | /* |
646 | * Increase the pool size | |
647 | * First take pages out of surplus state. Then make up the | |
648 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f NA |
649 | * |
650 | * We might race with alloc_buddy_huge_page() here and be unable | |
651 | * to convert a surplus huge page to a normal huge page. That is | |
652 | * not critical, though, it just means the overall size of the | |
653 | * pool might be one hugepage larger than it needs to be, but | |
654 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 655 | */ |
1da177e4 | 656 | spin_lock(&hugetlb_lock); |
7893d1d5 AL |
657 | while (surplus_huge_pages && count > persistent_huge_pages) { |
658 | if (!adjust_pool_surplus(-1)) | |
659 | break; | |
660 | } | |
661 | ||
662 | while (count > persistent_huge_pages) { | |
7893d1d5 AL |
663 | /* |
664 | * If this allocation races such that we no longer need the | |
665 | * page, free_huge_page will handle it by freeing the page | |
666 | * and reducing the surplus. | |
667 | */ | |
668 | spin_unlock(&hugetlb_lock); | |
669 | ret = alloc_fresh_huge_page(); | |
670 | spin_lock(&hugetlb_lock); | |
671 | if (!ret) | |
672 | goto out; | |
673 | ||
674 | } | |
7893d1d5 AL |
675 | |
676 | /* | |
677 | * Decrease the pool size | |
678 | * First return free pages to the buddy allocator (being careful | |
679 | * to keep enough around to satisfy reservations). Then place | |
680 | * pages into surplus state as needed so the pool will shrink | |
681 | * to the desired size as pages become free. | |
d1c3fb1f NA |
682 | * |
683 | * By placing pages into the surplus state independent of the | |
684 | * overcommit value, we are allowing the surplus pool size to | |
685 | * exceed overcommit. There are few sane options here. Since | |
686 | * alloc_buddy_huge_page() is checking the global counter, | |
687 | * though, we'll note that we're not allowed to exceed surplus | |
688 | * and won't grow the pool anywhere else. Not until one of the | |
689 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 690 | */ |
6b0c880d AL |
691 | min_count = resv_huge_pages + nr_huge_pages - free_huge_pages; |
692 | min_count = max(count, min_count); | |
7893d1d5 AL |
693 | try_to_free_low(min_count); |
694 | while (min_count < persistent_huge_pages) { | |
348e1e04 | 695 | struct page *page = dequeue_huge_page(); |
1da177e4 LT |
696 | if (!page) |
697 | break; | |
698 | update_and_free_page(page); | |
699 | } | |
7893d1d5 AL |
700 | while (count < persistent_huge_pages) { |
701 | if (!adjust_pool_surplus(1)) | |
702 | break; | |
703 | } | |
704 | out: | |
705 | ret = persistent_huge_pages; | |
1da177e4 | 706 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 707 | return ret; |
1da177e4 LT |
708 | } |
709 | ||
710 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, | |
711 | struct file *file, void __user *buffer, | |
712 | size_t *length, loff_t *ppos) | |
713 | { | |
714 | proc_doulongvec_minmax(table, write, file, buffer, length, ppos); | |
715 | max_huge_pages = set_max_huge_pages(max_huge_pages); | |
716 | return 0; | |
717 | } | |
396faf03 MG |
718 | |
719 | int hugetlb_treat_movable_handler(struct ctl_table *table, int write, | |
720 | struct file *file, void __user *buffer, | |
721 | size_t *length, loff_t *ppos) | |
722 | { | |
723 | proc_dointvec(table, write, file, buffer, length, ppos); | |
724 | if (hugepages_treat_as_movable) | |
725 | htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; | |
726 | else | |
727 | htlb_alloc_mask = GFP_HIGHUSER; | |
728 | return 0; | |
729 | } | |
730 | ||
a3d0c6aa NA |
731 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
732 | struct file *file, void __user *buffer, | |
733 | size_t *length, loff_t *ppos) | |
734 | { | |
a3d0c6aa | 735 | proc_doulongvec_minmax(table, write, file, buffer, length, ppos); |
064d9efe NA |
736 | spin_lock(&hugetlb_lock); |
737 | nr_overcommit_huge_pages = sysctl_overcommit_huge_pages; | |
a3d0c6aa NA |
738 | spin_unlock(&hugetlb_lock); |
739 | return 0; | |
740 | } | |
741 | ||
1da177e4 LT |
742 | #endif /* CONFIG_SYSCTL */ |
743 | ||
744 | int hugetlb_report_meminfo(char *buf) | |
745 | { | |
746 | return sprintf(buf, | |
747 | "HugePages_Total: %5lu\n" | |
748 | "HugePages_Free: %5lu\n" | |
a43a8c39 | 749 | "HugePages_Rsvd: %5lu\n" |
7893d1d5 | 750 | "HugePages_Surp: %5lu\n" |
1da177e4 LT |
751 | "Hugepagesize: %5lu kB\n", |
752 | nr_huge_pages, | |
753 | free_huge_pages, | |
a43a8c39 | 754 | resv_huge_pages, |
7893d1d5 | 755 | surplus_huge_pages, |
1da177e4 LT |
756 | HPAGE_SIZE/1024); |
757 | } | |
758 | ||
759 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
760 | { | |
761 | return sprintf(buf, | |
762 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
763 | "Node %d HugePages_Free: %5u\n" |
764 | "Node %d HugePages_Surp: %5u\n", | |
1da177e4 | 765 | nid, nr_huge_pages_node[nid], |
a1de0919 NA |
766 | nid, free_huge_pages_node[nid], |
767 | nid, surplus_huge_pages_node[nid]); | |
1da177e4 LT |
768 | } |
769 | ||
1da177e4 LT |
770 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
771 | unsigned long hugetlb_total_pages(void) | |
772 | { | |
773 | return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE); | |
774 | } | |
1da177e4 | 775 | |
fc1b8a73 MG |
776 | static int hugetlb_acct_memory(long delta) |
777 | { | |
778 | int ret = -ENOMEM; | |
779 | ||
780 | spin_lock(&hugetlb_lock); | |
781 | /* | |
782 | * When cpuset is configured, it breaks the strict hugetlb page | |
783 | * reservation as the accounting is done on a global variable. Such | |
784 | * reservation is completely rubbish in the presence of cpuset because | |
785 | * the reservation is not checked against page availability for the | |
786 | * current cpuset. Application can still potentially OOM'ed by kernel | |
787 | * with lack of free htlb page in cpuset that the task is in. | |
788 | * Attempt to enforce strict accounting with cpuset is almost | |
789 | * impossible (or too ugly) because cpuset is too fluid that | |
790 | * task or memory node can be dynamically moved between cpusets. | |
791 | * | |
792 | * The change of semantics for shared hugetlb mapping with cpuset is | |
793 | * undesirable. However, in order to preserve some of the semantics, | |
794 | * we fall back to check against current free page availability as | |
795 | * a best attempt and hopefully to minimize the impact of changing | |
796 | * semantics that cpuset has. | |
797 | */ | |
798 | if (delta > 0) { | |
799 | if (gather_surplus_pages(delta) < 0) | |
800 | goto out; | |
801 | ||
802 | if (delta > cpuset_mems_nr(free_huge_pages_node)) { | |
803 | return_unused_surplus_pages(delta); | |
804 | goto out; | |
805 | } | |
806 | } | |
807 | ||
808 | ret = 0; | |
809 | if (delta < 0) | |
810 | return_unused_surplus_pages((unsigned long) -delta); | |
811 | ||
812 | out: | |
813 | spin_unlock(&hugetlb_lock); | |
814 | return ret; | |
815 | } | |
816 | ||
a1e78772 MG |
817 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
818 | { | |
819 | unsigned long reserve = vma_resv_huge_pages(vma); | |
820 | if (reserve) | |
821 | hugetlb_acct_memory(-reserve); | |
822 | } | |
823 | ||
1da177e4 LT |
824 | /* |
825 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
826 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
827 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
828 | * this far. | |
829 | */ | |
d0217ac0 | 830 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
831 | { |
832 | BUG(); | |
d0217ac0 | 833 | return 0; |
1da177e4 LT |
834 | } |
835 | ||
836 | struct vm_operations_struct hugetlb_vm_ops = { | |
d0217ac0 | 837 | .fault = hugetlb_vm_op_fault, |
a1e78772 | 838 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
839 | }; |
840 | ||
1e8f889b DG |
841 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
842 | int writable) | |
63551ae0 DG |
843 | { |
844 | pte_t entry; | |
845 | ||
1e8f889b | 846 | if (writable) { |
63551ae0 DG |
847 | entry = |
848 | pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); | |
849 | } else { | |
7f2e9525 | 850 | entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot)); |
63551ae0 DG |
851 | } |
852 | entry = pte_mkyoung(entry); | |
853 | entry = pte_mkhuge(entry); | |
854 | ||
855 | return entry; | |
856 | } | |
857 | ||
1e8f889b DG |
858 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
859 | unsigned long address, pte_t *ptep) | |
860 | { | |
861 | pte_t entry; | |
862 | ||
7f2e9525 GS |
863 | entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep))); |
864 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) { | |
8dab5241 | 865 | update_mmu_cache(vma, address, entry); |
8dab5241 | 866 | } |
1e8f889b DG |
867 | } |
868 | ||
869 | ||
63551ae0 DG |
870 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
871 | struct vm_area_struct *vma) | |
872 | { | |
873 | pte_t *src_pte, *dst_pte, entry; | |
874 | struct page *ptepage; | |
1c59827d | 875 | unsigned long addr; |
1e8f889b DG |
876 | int cow; |
877 | ||
878 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 879 | |
1c59827d | 880 | for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) { |
c74df32c HD |
881 | src_pte = huge_pte_offset(src, addr); |
882 | if (!src_pte) | |
883 | continue; | |
63551ae0 DG |
884 | dst_pte = huge_pte_alloc(dst, addr); |
885 | if (!dst_pte) | |
886 | goto nomem; | |
c5c99429 LW |
887 | |
888 | /* If the pagetables are shared don't copy or take references */ | |
889 | if (dst_pte == src_pte) | |
890 | continue; | |
891 | ||
c74df32c | 892 | spin_lock(&dst->page_table_lock); |
46478758 | 893 | spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); |
7f2e9525 | 894 | if (!huge_pte_none(huge_ptep_get(src_pte))) { |
1e8f889b | 895 | if (cow) |
7f2e9525 GS |
896 | huge_ptep_set_wrprotect(src, addr, src_pte); |
897 | entry = huge_ptep_get(src_pte); | |
1c59827d HD |
898 | ptepage = pte_page(entry); |
899 | get_page(ptepage); | |
1c59827d HD |
900 | set_huge_pte_at(dst, addr, dst_pte, entry); |
901 | } | |
902 | spin_unlock(&src->page_table_lock); | |
c74df32c | 903 | spin_unlock(&dst->page_table_lock); |
63551ae0 DG |
904 | } |
905 | return 0; | |
906 | ||
907 | nomem: | |
908 | return -ENOMEM; | |
909 | } | |
910 | ||
502717f4 KC |
911 | void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
912 | unsigned long end) | |
63551ae0 DG |
913 | { |
914 | struct mm_struct *mm = vma->vm_mm; | |
915 | unsigned long address; | |
c7546f8f | 916 | pte_t *ptep; |
63551ae0 DG |
917 | pte_t pte; |
918 | struct page *page; | |
fe1668ae | 919 | struct page *tmp; |
c0a499c2 KC |
920 | /* |
921 | * A page gathering list, protected by per file i_mmap_lock. The | |
922 | * lock is used to avoid list corruption from multiple unmapping | |
923 | * of the same page since we are using page->lru. | |
924 | */ | |
fe1668ae | 925 | LIST_HEAD(page_list); |
63551ae0 DG |
926 | |
927 | WARN_ON(!is_vm_hugetlb_page(vma)); | |
928 | BUG_ON(start & ~HPAGE_MASK); | |
929 | BUG_ON(end & ~HPAGE_MASK); | |
930 | ||
508034a3 | 931 | spin_lock(&mm->page_table_lock); |
63551ae0 | 932 | for (address = start; address < end; address += HPAGE_SIZE) { |
c7546f8f | 933 | ptep = huge_pte_offset(mm, address); |
4c887265 | 934 | if (!ptep) |
c7546f8f DG |
935 | continue; |
936 | ||
39dde65c KC |
937 | if (huge_pmd_unshare(mm, &address, ptep)) |
938 | continue; | |
939 | ||
c7546f8f | 940 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
7f2e9525 | 941 | if (huge_pte_none(pte)) |
63551ae0 | 942 | continue; |
c7546f8f | 943 | |
63551ae0 | 944 | page = pte_page(pte); |
6649a386 KC |
945 | if (pte_dirty(pte)) |
946 | set_page_dirty(page); | |
fe1668ae | 947 | list_add(&page->lru, &page_list); |
63551ae0 | 948 | } |
1da177e4 | 949 | spin_unlock(&mm->page_table_lock); |
508034a3 | 950 | flush_tlb_range(vma, start, end); |
fe1668ae KC |
951 | list_for_each_entry_safe(page, tmp, &page_list, lru) { |
952 | list_del(&page->lru); | |
953 | put_page(page); | |
954 | } | |
1da177e4 | 955 | } |
63551ae0 | 956 | |
502717f4 KC |
957 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
958 | unsigned long end) | |
959 | { | |
960 | /* | |
961 | * It is undesirable to test vma->vm_file as it should be non-null | |
962 | * for valid hugetlb area. However, vm_file will be NULL in the error | |
963 | * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails, | |
964 | * do_mmap_pgoff() nullifies vma->vm_file before calling this function | |
965 | * to clean up. Since no pte has actually been setup, it is safe to | |
966 | * do nothing in this case. | |
967 | */ | |
968 | if (vma->vm_file) { | |
969 | spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); | |
970 | __unmap_hugepage_range(vma, start, end); | |
971 | spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); | |
972 | } | |
973 | } | |
974 | ||
1e8f889b DG |
975 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
976 | unsigned long address, pte_t *ptep, pte_t pte) | |
977 | { | |
978 | struct page *old_page, *new_page; | |
79ac6ba4 | 979 | int avoidcopy; |
1e8f889b DG |
980 | |
981 | old_page = pte_page(pte); | |
982 | ||
983 | /* If no-one else is actually using this page, avoid the copy | |
984 | * and just make the page writable */ | |
985 | avoidcopy = (page_count(old_page) == 1); | |
986 | if (avoidcopy) { | |
987 | set_huge_ptep_writable(vma, address, ptep); | |
83c54070 | 988 | return 0; |
1e8f889b DG |
989 | } |
990 | ||
991 | page_cache_get(old_page); | |
5da7ca86 | 992 | new_page = alloc_huge_page(vma, address); |
1e8f889b | 993 | |
2fc39cec | 994 | if (IS_ERR(new_page)) { |
1e8f889b | 995 | page_cache_release(old_page); |
2fc39cec | 996 | return -PTR_ERR(new_page); |
1e8f889b DG |
997 | } |
998 | ||
999 | spin_unlock(&mm->page_table_lock); | |
9de455b2 | 1000 | copy_huge_page(new_page, old_page, address, vma); |
0ed361de | 1001 | __SetPageUptodate(new_page); |
1e8f889b DG |
1002 | spin_lock(&mm->page_table_lock); |
1003 | ||
1004 | ptep = huge_pte_offset(mm, address & HPAGE_MASK); | |
7f2e9525 | 1005 | if (likely(pte_same(huge_ptep_get(ptep), pte))) { |
1e8f889b | 1006 | /* Break COW */ |
8fe627ec | 1007 | huge_ptep_clear_flush(vma, address, ptep); |
1e8f889b DG |
1008 | set_huge_pte_at(mm, address, ptep, |
1009 | make_huge_pte(vma, new_page, 1)); | |
1010 | /* Make the old page be freed below */ | |
1011 | new_page = old_page; | |
1012 | } | |
1013 | page_cache_release(new_page); | |
1014 | page_cache_release(old_page); | |
83c54070 | 1015 | return 0; |
1e8f889b DG |
1016 | } |
1017 | ||
a1ed3dda | 1018 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
1e8f889b | 1019 | unsigned long address, pte_t *ptep, int write_access) |
ac9b9c66 HD |
1020 | { |
1021 | int ret = VM_FAULT_SIGBUS; | |
4c887265 AL |
1022 | unsigned long idx; |
1023 | unsigned long size; | |
4c887265 AL |
1024 | struct page *page; |
1025 | struct address_space *mapping; | |
1e8f889b | 1026 | pte_t new_pte; |
4c887265 | 1027 | |
4c887265 AL |
1028 | mapping = vma->vm_file->f_mapping; |
1029 | idx = ((address - vma->vm_start) >> HPAGE_SHIFT) | |
1030 | + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT)); | |
1031 | ||
1032 | /* | |
1033 | * Use page lock to guard against racing truncation | |
1034 | * before we get page_table_lock. | |
1035 | */ | |
6bda666a CL |
1036 | retry: |
1037 | page = find_lock_page(mapping, idx); | |
1038 | if (!page) { | |
ebed4bfc HD |
1039 | size = i_size_read(mapping->host) >> HPAGE_SHIFT; |
1040 | if (idx >= size) | |
1041 | goto out; | |
6bda666a | 1042 | page = alloc_huge_page(vma, address); |
2fc39cec AL |
1043 | if (IS_ERR(page)) { |
1044 | ret = -PTR_ERR(page); | |
6bda666a CL |
1045 | goto out; |
1046 | } | |
79ac6ba4 | 1047 | clear_huge_page(page, address); |
0ed361de | 1048 | __SetPageUptodate(page); |
ac9b9c66 | 1049 | |
6bda666a CL |
1050 | if (vma->vm_flags & VM_SHARED) { |
1051 | int err; | |
45c682a6 | 1052 | struct inode *inode = mapping->host; |
6bda666a CL |
1053 | |
1054 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
1055 | if (err) { | |
1056 | put_page(page); | |
6bda666a CL |
1057 | if (err == -EEXIST) |
1058 | goto retry; | |
1059 | goto out; | |
1060 | } | |
45c682a6 KC |
1061 | |
1062 | spin_lock(&inode->i_lock); | |
1063 | inode->i_blocks += BLOCKS_PER_HUGEPAGE; | |
1064 | spin_unlock(&inode->i_lock); | |
6bda666a CL |
1065 | } else |
1066 | lock_page(page); | |
1067 | } | |
1e8f889b | 1068 | |
ac9b9c66 | 1069 | spin_lock(&mm->page_table_lock); |
4c887265 AL |
1070 | size = i_size_read(mapping->host) >> HPAGE_SHIFT; |
1071 | if (idx >= size) | |
1072 | goto backout; | |
1073 | ||
83c54070 | 1074 | ret = 0; |
7f2e9525 | 1075 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
1076 | goto backout; |
1077 | ||
1e8f889b DG |
1078 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
1079 | && (vma->vm_flags & VM_SHARED))); | |
1080 | set_huge_pte_at(mm, address, ptep, new_pte); | |
1081 | ||
1082 | if (write_access && !(vma->vm_flags & VM_SHARED)) { | |
1083 | /* Optimization, do the COW without a second fault */ | |
1084 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte); | |
1085 | } | |
1086 | ||
ac9b9c66 | 1087 | spin_unlock(&mm->page_table_lock); |
4c887265 AL |
1088 | unlock_page(page); |
1089 | out: | |
ac9b9c66 | 1090 | return ret; |
4c887265 AL |
1091 | |
1092 | backout: | |
1093 | spin_unlock(&mm->page_table_lock); | |
4c887265 AL |
1094 | unlock_page(page); |
1095 | put_page(page); | |
1096 | goto out; | |
ac9b9c66 HD |
1097 | } |
1098 | ||
86e5216f AL |
1099 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
1100 | unsigned long address, int write_access) | |
1101 | { | |
1102 | pte_t *ptep; | |
1103 | pte_t entry; | |
1e8f889b | 1104 | int ret; |
3935baa9 | 1105 | static DEFINE_MUTEX(hugetlb_instantiation_mutex); |
86e5216f AL |
1106 | |
1107 | ptep = huge_pte_alloc(mm, address); | |
1108 | if (!ptep) | |
1109 | return VM_FAULT_OOM; | |
1110 | ||
3935baa9 DG |
1111 | /* |
1112 | * Serialize hugepage allocation and instantiation, so that we don't | |
1113 | * get spurious allocation failures if two CPUs race to instantiate | |
1114 | * the same page in the page cache. | |
1115 | */ | |
1116 | mutex_lock(&hugetlb_instantiation_mutex); | |
7f2e9525 GS |
1117 | entry = huge_ptep_get(ptep); |
1118 | if (huge_pte_none(entry)) { | |
3935baa9 DG |
1119 | ret = hugetlb_no_page(mm, vma, address, ptep, write_access); |
1120 | mutex_unlock(&hugetlb_instantiation_mutex); | |
1121 | return ret; | |
1122 | } | |
86e5216f | 1123 | |
83c54070 | 1124 | ret = 0; |
1e8f889b DG |
1125 | |
1126 | spin_lock(&mm->page_table_lock); | |
1127 | /* Check for a racing update before calling hugetlb_cow */ | |
7f2e9525 | 1128 | if (likely(pte_same(entry, huge_ptep_get(ptep)))) |
1e8f889b DG |
1129 | if (write_access && !pte_write(entry)) |
1130 | ret = hugetlb_cow(mm, vma, address, ptep, entry); | |
1131 | spin_unlock(&mm->page_table_lock); | |
3935baa9 | 1132 | mutex_unlock(&hugetlb_instantiation_mutex); |
1e8f889b DG |
1133 | |
1134 | return ret; | |
86e5216f AL |
1135 | } |
1136 | ||
63551ae0 DG |
1137 | int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
1138 | struct page **pages, struct vm_area_struct **vmas, | |
5b23dbe8 AL |
1139 | unsigned long *position, int *length, int i, |
1140 | int write) | |
63551ae0 | 1141 | { |
d5d4b0aa KC |
1142 | unsigned long pfn_offset; |
1143 | unsigned long vaddr = *position; | |
63551ae0 DG |
1144 | int remainder = *length; |
1145 | ||
1c59827d | 1146 | spin_lock(&mm->page_table_lock); |
63551ae0 | 1147 | while (vaddr < vma->vm_end && remainder) { |
4c887265 AL |
1148 | pte_t *pte; |
1149 | struct page *page; | |
63551ae0 | 1150 | |
4c887265 AL |
1151 | /* |
1152 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
1153 | * each hugepage. We have to make * sure we get the | |
1154 | * first, for the page indexing below to work. | |
1155 | */ | |
1156 | pte = huge_pte_offset(mm, vaddr & HPAGE_MASK); | |
63551ae0 | 1157 | |
7f2e9525 GS |
1158 | if (!pte || huge_pte_none(huge_ptep_get(pte)) || |
1159 | (write && !pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 1160 | int ret; |
63551ae0 | 1161 | |
4c887265 | 1162 | spin_unlock(&mm->page_table_lock); |
5b23dbe8 | 1163 | ret = hugetlb_fault(mm, vma, vaddr, write); |
4c887265 | 1164 | spin_lock(&mm->page_table_lock); |
a89182c7 | 1165 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 1166 | continue; |
63551ae0 | 1167 | |
4c887265 AL |
1168 | remainder = 0; |
1169 | if (!i) | |
1170 | i = -EFAULT; | |
1171 | break; | |
1172 | } | |
1173 | ||
d5d4b0aa | 1174 | pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT; |
7f2e9525 | 1175 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 1176 | same_page: |
d6692183 KC |
1177 | if (pages) { |
1178 | get_page(page); | |
d5d4b0aa | 1179 | pages[i] = page + pfn_offset; |
d6692183 | 1180 | } |
63551ae0 DG |
1181 | |
1182 | if (vmas) | |
1183 | vmas[i] = vma; | |
1184 | ||
1185 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 1186 | ++pfn_offset; |
63551ae0 DG |
1187 | --remainder; |
1188 | ++i; | |
d5d4b0aa KC |
1189 | if (vaddr < vma->vm_end && remainder && |
1190 | pfn_offset < HPAGE_SIZE/PAGE_SIZE) { | |
1191 | /* | |
1192 | * We use pfn_offset to avoid touching the pageframes | |
1193 | * of this compound page. | |
1194 | */ | |
1195 | goto same_page; | |
1196 | } | |
63551ae0 | 1197 | } |
1c59827d | 1198 | spin_unlock(&mm->page_table_lock); |
63551ae0 DG |
1199 | *length = remainder; |
1200 | *position = vaddr; | |
1201 | ||
1202 | return i; | |
1203 | } | |
8f860591 ZY |
1204 | |
1205 | void hugetlb_change_protection(struct vm_area_struct *vma, | |
1206 | unsigned long address, unsigned long end, pgprot_t newprot) | |
1207 | { | |
1208 | struct mm_struct *mm = vma->vm_mm; | |
1209 | unsigned long start = address; | |
1210 | pte_t *ptep; | |
1211 | pte_t pte; | |
1212 | ||
1213 | BUG_ON(address >= end); | |
1214 | flush_cache_range(vma, address, end); | |
1215 | ||
39dde65c | 1216 | spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); |
8f860591 ZY |
1217 | spin_lock(&mm->page_table_lock); |
1218 | for (; address < end; address += HPAGE_SIZE) { | |
1219 | ptep = huge_pte_offset(mm, address); | |
1220 | if (!ptep) | |
1221 | continue; | |
39dde65c KC |
1222 | if (huge_pmd_unshare(mm, &address, ptep)) |
1223 | continue; | |
7f2e9525 | 1224 | if (!huge_pte_none(huge_ptep_get(ptep))) { |
8f860591 ZY |
1225 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
1226 | pte = pte_mkhuge(pte_modify(pte, newprot)); | |
1227 | set_huge_pte_at(mm, address, ptep, pte); | |
8f860591 ZY |
1228 | } |
1229 | } | |
1230 | spin_unlock(&mm->page_table_lock); | |
39dde65c | 1231 | spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); |
8f860591 ZY |
1232 | |
1233 | flush_tlb_range(vma, start, end); | |
1234 | } | |
1235 | ||
a43a8c39 KC |
1236 | struct file_region { |
1237 | struct list_head link; | |
1238 | long from; | |
1239 | long to; | |
1240 | }; | |
1241 | ||
1242 | static long region_add(struct list_head *head, long f, long t) | |
1243 | { | |
1244 | struct file_region *rg, *nrg, *trg; | |
1245 | ||
1246 | /* Locate the region we are either in or before. */ | |
1247 | list_for_each_entry(rg, head, link) | |
1248 | if (f <= rg->to) | |
1249 | break; | |
1250 | ||
1251 | /* Round our left edge to the current segment if it encloses us. */ | |
1252 | if (f > rg->from) | |
1253 | f = rg->from; | |
1254 | ||
1255 | /* Check for and consume any regions we now overlap with. */ | |
1256 | nrg = rg; | |
1257 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
1258 | if (&rg->link == head) | |
1259 | break; | |
1260 | if (rg->from > t) | |
1261 | break; | |
1262 | ||
1263 | /* If this area reaches higher then extend our area to | |
1264 | * include it completely. If this is not the first area | |
1265 | * which we intend to reuse, free it. */ | |
1266 | if (rg->to > t) | |
1267 | t = rg->to; | |
1268 | if (rg != nrg) { | |
1269 | list_del(&rg->link); | |
1270 | kfree(rg); | |
1271 | } | |
1272 | } | |
1273 | nrg->from = f; | |
1274 | nrg->to = t; | |
1275 | return 0; | |
1276 | } | |
1277 | ||
1278 | static long region_chg(struct list_head *head, long f, long t) | |
1279 | { | |
1280 | struct file_region *rg, *nrg; | |
1281 | long chg = 0; | |
1282 | ||
1283 | /* Locate the region we are before or in. */ | |
1284 | list_for_each_entry(rg, head, link) | |
1285 | if (f <= rg->to) | |
1286 | break; | |
1287 | ||
1288 | /* If we are below the current region then a new region is required. | |
1289 | * Subtle, allocate a new region at the position but make it zero | |
183ff22b | 1290 | * size such that we can guarantee to record the reservation. */ |
a43a8c39 KC |
1291 | if (&rg->link == head || t < rg->from) { |
1292 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
c80544dc | 1293 | if (!nrg) |
a43a8c39 KC |
1294 | return -ENOMEM; |
1295 | nrg->from = f; | |
1296 | nrg->to = f; | |
1297 | INIT_LIST_HEAD(&nrg->link); | |
1298 | list_add(&nrg->link, rg->link.prev); | |
1299 | ||
1300 | return t - f; | |
1301 | } | |
1302 | ||
1303 | /* Round our left edge to the current segment if it encloses us. */ | |
1304 | if (f > rg->from) | |
1305 | f = rg->from; | |
1306 | chg = t - f; | |
1307 | ||
1308 | /* Check for and consume any regions we now overlap with. */ | |
1309 | list_for_each_entry(rg, rg->link.prev, link) { | |
1310 | if (&rg->link == head) | |
1311 | break; | |
1312 | if (rg->from > t) | |
1313 | return chg; | |
1314 | ||
1315 | /* We overlap with this area, if it extends futher than | |
1316 | * us then we must extend ourselves. Account for its | |
1317 | * existing reservation. */ | |
1318 | if (rg->to > t) { | |
1319 | chg += rg->to - t; | |
1320 | t = rg->to; | |
1321 | } | |
1322 | chg -= rg->to - rg->from; | |
1323 | } | |
1324 | return chg; | |
1325 | } | |
1326 | ||
1327 | static long region_truncate(struct list_head *head, long end) | |
1328 | { | |
1329 | struct file_region *rg, *trg; | |
1330 | long chg = 0; | |
1331 | ||
1332 | /* Locate the region we are either in or before. */ | |
1333 | list_for_each_entry(rg, head, link) | |
1334 | if (end <= rg->to) | |
1335 | break; | |
1336 | if (&rg->link == head) | |
1337 | return 0; | |
1338 | ||
1339 | /* If we are in the middle of a region then adjust it. */ | |
1340 | if (end > rg->from) { | |
1341 | chg = rg->to - end; | |
1342 | rg->to = end; | |
1343 | rg = list_entry(rg->link.next, typeof(*rg), link); | |
1344 | } | |
1345 | ||
1346 | /* Drop any remaining regions. */ | |
1347 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
1348 | if (&rg->link == head) | |
1349 | break; | |
1350 | chg += rg->to - rg->from; | |
1351 | list_del(&rg->link); | |
1352 | kfree(rg); | |
1353 | } | |
1354 | return chg; | |
1355 | } | |
1356 | ||
a1e78772 MG |
1357 | int hugetlb_reserve_pages(struct inode *inode, |
1358 | long from, long to, | |
1359 | struct vm_area_struct *vma) | |
e4e574b7 AL |
1360 | { |
1361 | long ret, chg; | |
1362 | ||
a1e78772 MG |
1363 | /* |
1364 | * Shared mappings base their reservation on the number of pages that | |
1365 | * are already allocated on behalf of the file. Private mappings need | |
1366 | * to reserve the full area even if read-only as mprotect() may be | |
1367 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
1368 | */ | |
1369 | if (!vma || vma->vm_flags & VM_SHARED) | |
1370 | chg = region_chg(&inode->i_mapping->private_list, from, to); | |
1371 | else { | |
1372 | chg = to - from; | |
1373 | set_vma_resv_huge_pages(vma, chg); | |
1374 | } | |
1375 | ||
e4e574b7 AL |
1376 | if (chg < 0) |
1377 | return chg; | |
8a630112 | 1378 | |
90d8b7e6 AL |
1379 | if (hugetlb_get_quota(inode->i_mapping, chg)) |
1380 | return -ENOSPC; | |
a43a8c39 | 1381 | ret = hugetlb_acct_memory(chg); |
68842c9b KC |
1382 | if (ret < 0) { |
1383 | hugetlb_put_quota(inode->i_mapping, chg); | |
a43a8c39 | 1384 | return ret; |
68842c9b | 1385 | } |
a1e78772 MG |
1386 | if (!vma || vma->vm_flags & VM_SHARED) |
1387 | region_add(&inode->i_mapping->private_list, from, to); | |
a43a8c39 KC |
1388 | return 0; |
1389 | } | |
1390 | ||
1391 | void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) | |
1392 | { | |
1393 | long chg = region_truncate(&inode->i_mapping->private_list, offset); | |
45c682a6 KC |
1394 | |
1395 | spin_lock(&inode->i_lock); | |
1396 | inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed; | |
1397 | spin_unlock(&inode->i_lock); | |
1398 | ||
90d8b7e6 AL |
1399 | hugetlb_put_quota(inode->i_mapping, (chg - freed)); |
1400 | hugetlb_acct_memory(-(chg - freed)); | |
a43a8c39 | 1401 | } |