]>
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 | |
79ac6ba4 DG |
43 | static void clear_huge_page(struct page *page, unsigned long addr) |
44 | { | |
45 | int i; | |
46 | ||
47 | might_sleep(); | |
48 | for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) { | |
49 | cond_resched(); | |
281e0e3b | 50 | clear_user_highpage(page + i, addr + i * PAGE_SIZE); |
79ac6ba4 DG |
51 | } |
52 | } | |
53 | ||
54 | static void copy_huge_page(struct page *dst, struct page *src, | |
9de455b2 | 55 | unsigned long addr, struct vm_area_struct *vma) |
79ac6ba4 DG |
56 | { |
57 | int i; | |
58 | ||
59 | might_sleep(); | |
60 | for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) { | |
61 | cond_resched(); | |
9de455b2 | 62 | copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma); |
79ac6ba4 DG |
63 | } |
64 | } | |
65 | ||
1da177e4 LT |
66 | static void enqueue_huge_page(struct page *page) |
67 | { | |
68 | int nid = page_to_nid(page); | |
69 | list_add(&page->lru, &hugepage_freelists[nid]); | |
70 | free_huge_pages++; | |
71 | free_huge_pages_node[nid]++; | |
72 | } | |
73 | ||
348e1e04 NA |
74 | static struct page *dequeue_huge_page(void) |
75 | { | |
76 | int nid; | |
77 | struct page *page = NULL; | |
78 | ||
79 | for (nid = 0; nid < MAX_NUMNODES; ++nid) { | |
80 | if (!list_empty(&hugepage_freelists[nid])) { | |
81 | page = list_entry(hugepage_freelists[nid].next, | |
82 | struct page, lru); | |
83 | list_del(&page->lru); | |
84 | free_huge_pages--; | |
85 | free_huge_pages_node[nid]--; | |
86 | break; | |
87 | } | |
88 | } | |
89 | return page; | |
90 | } | |
91 | ||
92 | static struct page *dequeue_huge_page_vma(struct vm_area_struct *vma, | |
5da7ca86 | 93 | unsigned long address) |
1da177e4 | 94 | { |
31a5c6e4 | 95 | int nid; |
1da177e4 | 96 | struct page *page = NULL; |
480eccf9 | 97 | struct mempolicy *mpol; |
19770b32 | 98 | nodemask_t *nodemask; |
396faf03 | 99 | struct zonelist *zonelist = huge_zonelist(vma, address, |
19770b32 | 100 | htlb_alloc_mask, &mpol, &nodemask); |
dd1a239f MG |
101 | struct zone *zone; |
102 | struct zoneref *z; | |
1da177e4 | 103 | |
19770b32 MG |
104 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
105 | MAX_NR_ZONES - 1, nodemask) { | |
54a6eb5c MG |
106 | nid = zone_to_nid(zone); |
107 | if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) && | |
3abf7afd AM |
108 | !list_empty(&hugepage_freelists[nid])) { |
109 | page = list_entry(hugepage_freelists[nid].next, | |
110 | struct page, lru); | |
111 | list_del(&page->lru); | |
112 | free_huge_pages--; | |
113 | free_huge_pages_node[nid]--; | |
e4e574b7 AL |
114 | if (vma && vma->vm_flags & VM_MAYSHARE) |
115 | resv_huge_pages--; | |
5ab3ee7b | 116 | break; |
3abf7afd | 117 | } |
1da177e4 | 118 | } |
52cd3b07 | 119 | mpol_cond_put(mpol); |
1da177e4 LT |
120 | return page; |
121 | } | |
122 | ||
6af2acb6 AL |
123 | static void update_and_free_page(struct page *page) |
124 | { | |
125 | int i; | |
126 | nr_huge_pages--; | |
127 | nr_huge_pages_node[page_to_nid(page)]--; | |
128 | for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) { | |
129 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced | | |
130 | 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved | | |
131 | 1 << PG_private | 1<< PG_writeback); | |
132 | } | |
133 | set_compound_page_dtor(page, NULL); | |
134 | set_page_refcounted(page); | |
135 | __free_pages(page, HUGETLB_PAGE_ORDER); | |
136 | } | |
137 | ||
27a85ef1 DG |
138 | static void free_huge_page(struct page *page) |
139 | { | |
7893d1d5 | 140 | int nid = page_to_nid(page); |
c79fb75e | 141 | struct address_space *mapping; |
27a85ef1 | 142 | |
c79fb75e | 143 | mapping = (struct address_space *) page_private(page); |
e5df70ab | 144 | set_page_private(page, 0); |
7893d1d5 | 145 | BUG_ON(page_count(page)); |
27a85ef1 DG |
146 | INIT_LIST_HEAD(&page->lru); |
147 | ||
148 | spin_lock(&hugetlb_lock); | |
7893d1d5 AL |
149 | if (surplus_huge_pages_node[nid]) { |
150 | update_and_free_page(page); | |
151 | surplus_huge_pages--; | |
152 | surplus_huge_pages_node[nid]--; | |
153 | } else { | |
154 | enqueue_huge_page(page); | |
155 | } | |
27a85ef1 | 156 | spin_unlock(&hugetlb_lock); |
c79fb75e | 157 | if (mapping) |
9a119c05 | 158 | hugetlb_put_quota(mapping, 1); |
27a85ef1 DG |
159 | } |
160 | ||
7893d1d5 AL |
161 | /* |
162 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
163 | * balanced by operating on them in a round-robin fashion. | |
164 | * Returns 1 if an adjustment was made. | |
165 | */ | |
166 | static int adjust_pool_surplus(int delta) | |
167 | { | |
168 | static int prev_nid; | |
169 | int nid = prev_nid; | |
170 | int ret = 0; | |
171 | ||
172 | VM_BUG_ON(delta != -1 && delta != 1); | |
173 | do { | |
174 | nid = next_node(nid, node_online_map); | |
175 | if (nid == MAX_NUMNODES) | |
176 | nid = first_node(node_online_map); | |
177 | ||
178 | /* To shrink on this node, there must be a surplus page */ | |
179 | if (delta < 0 && !surplus_huge_pages_node[nid]) | |
180 | continue; | |
181 | /* Surplus cannot exceed the total number of pages */ | |
182 | if (delta > 0 && surplus_huge_pages_node[nid] >= | |
183 | nr_huge_pages_node[nid]) | |
184 | continue; | |
185 | ||
186 | surplus_huge_pages += delta; | |
187 | surplus_huge_pages_node[nid] += delta; | |
188 | ret = 1; | |
189 | break; | |
190 | } while (nid != prev_nid); | |
191 | ||
192 | prev_nid = nid; | |
193 | return ret; | |
194 | } | |
195 | ||
63b4613c | 196 | static struct page *alloc_fresh_huge_page_node(int nid) |
1da177e4 | 197 | { |
1da177e4 | 198 | struct page *page; |
f96efd58 | 199 | |
63b4613c NA |
200 | page = alloc_pages_node(nid, |
201 | htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|__GFP_NOWARN, | |
202 | HUGETLB_PAGE_ORDER); | |
1da177e4 | 203 | if (page) { |
33f2ef89 | 204 | set_compound_page_dtor(page, free_huge_page); |
0bd0f9fb | 205 | spin_lock(&hugetlb_lock); |
1da177e4 | 206 | nr_huge_pages++; |
63b4613c | 207 | nr_huge_pages_node[nid]++; |
0bd0f9fb | 208 | spin_unlock(&hugetlb_lock); |
a482289d | 209 | put_page(page); /* free it into the hugepage allocator */ |
1da177e4 | 210 | } |
63b4613c NA |
211 | |
212 | return page; | |
213 | } | |
214 | ||
215 | static int alloc_fresh_huge_page(void) | |
216 | { | |
217 | struct page *page; | |
218 | int start_nid; | |
219 | int next_nid; | |
220 | int ret = 0; | |
221 | ||
222 | start_nid = hugetlb_next_nid; | |
223 | ||
224 | do { | |
225 | page = alloc_fresh_huge_page_node(hugetlb_next_nid); | |
226 | if (page) | |
227 | ret = 1; | |
228 | /* | |
229 | * Use a helper variable to find the next node and then | |
230 | * copy it back to hugetlb_next_nid afterwards: | |
231 | * otherwise there's a window in which a racer might | |
232 | * pass invalid nid MAX_NUMNODES to alloc_pages_node. | |
233 | * But we don't need to use a spin_lock here: it really | |
234 | * doesn't matter if occasionally a racer chooses the | |
235 | * same nid as we do. Move nid forward in the mask even | |
236 | * if we just successfully allocated a hugepage so that | |
237 | * the next caller gets hugepages on the next node. | |
238 | */ | |
239 | next_nid = next_node(hugetlb_next_nid, node_online_map); | |
240 | if (next_nid == MAX_NUMNODES) | |
241 | next_nid = first_node(node_online_map); | |
242 | hugetlb_next_nid = next_nid; | |
243 | } while (!page && hugetlb_next_nid != start_nid); | |
244 | ||
3b116300 AL |
245 | if (ret) |
246 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
247 | else | |
248 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
249 | ||
63b4613c | 250 | return ret; |
1da177e4 LT |
251 | } |
252 | ||
7893d1d5 AL |
253 | static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma, |
254 | unsigned long address) | |
255 | { | |
256 | struct page *page; | |
d1c3fb1f | 257 | unsigned int nid; |
7893d1d5 | 258 | |
d1c3fb1f NA |
259 | /* |
260 | * Assume we will successfully allocate the surplus page to | |
261 | * prevent racing processes from causing the surplus to exceed | |
262 | * overcommit | |
263 | * | |
264 | * This however introduces a different race, where a process B | |
265 | * tries to grow the static hugepage pool while alloc_pages() is | |
266 | * called by process A. B will only examine the per-node | |
267 | * counters in determining if surplus huge pages can be | |
268 | * converted to normal huge pages in adjust_pool_surplus(). A | |
269 | * won't be able to increment the per-node counter, until the | |
270 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
271 | * no more huge pages can be converted from surplus to normal | |
272 | * state (and doesn't try to convert again). Thus, we have a | |
273 | * case where a surplus huge page exists, the pool is grown, and | |
274 | * the surplus huge page still exists after, even though it | |
275 | * should just have been converted to a normal huge page. This | |
276 | * does not leak memory, though, as the hugepage will be freed | |
277 | * once it is out of use. It also does not allow the counters to | |
278 | * go out of whack in adjust_pool_surplus() as we don't modify | |
279 | * the node values until we've gotten the hugepage and only the | |
280 | * per-node value is checked there. | |
281 | */ | |
282 | spin_lock(&hugetlb_lock); | |
283 | if (surplus_huge_pages >= nr_overcommit_huge_pages) { | |
284 | spin_unlock(&hugetlb_lock); | |
285 | return NULL; | |
286 | } else { | |
287 | nr_huge_pages++; | |
288 | surplus_huge_pages++; | |
289 | } | |
290 | spin_unlock(&hugetlb_lock); | |
291 | ||
7893d1d5 AL |
292 | page = alloc_pages(htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN, |
293 | HUGETLB_PAGE_ORDER); | |
d1c3fb1f NA |
294 | |
295 | spin_lock(&hugetlb_lock); | |
7893d1d5 | 296 | if (page) { |
2668db91 AL |
297 | /* |
298 | * This page is now managed by the hugetlb allocator and has | |
299 | * no users -- drop the buddy allocator's reference. | |
300 | */ | |
301 | put_page_testzero(page); | |
302 | VM_BUG_ON(page_count(page)); | |
d1c3fb1f | 303 | nid = page_to_nid(page); |
7893d1d5 | 304 | set_compound_page_dtor(page, free_huge_page); |
d1c3fb1f NA |
305 | /* |
306 | * We incremented the global counters already | |
307 | */ | |
308 | nr_huge_pages_node[nid]++; | |
309 | surplus_huge_pages_node[nid]++; | |
3b116300 | 310 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f NA |
311 | } else { |
312 | nr_huge_pages--; | |
313 | surplus_huge_pages--; | |
3b116300 | 314 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 315 | } |
d1c3fb1f | 316 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
317 | |
318 | return page; | |
319 | } | |
320 | ||
e4e574b7 AL |
321 | /* |
322 | * Increase the hugetlb pool such that it can accomodate a reservation | |
323 | * of size 'delta'. | |
324 | */ | |
325 | static int gather_surplus_pages(int delta) | |
326 | { | |
327 | struct list_head surplus_list; | |
328 | struct page *page, *tmp; | |
329 | int ret, i; | |
330 | int needed, allocated; | |
331 | ||
332 | needed = (resv_huge_pages + delta) - free_huge_pages; | |
ac09b3a1 AL |
333 | if (needed <= 0) { |
334 | resv_huge_pages += delta; | |
e4e574b7 | 335 | return 0; |
ac09b3a1 | 336 | } |
e4e574b7 AL |
337 | |
338 | allocated = 0; | |
339 | INIT_LIST_HEAD(&surplus_list); | |
340 | ||
341 | ret = -ENOMEM; | |
342 | retry: | |
343 | spin_unlock(&hugetlb_lock); | |
344 | for (i = 0; i < needed; i++) { | |
345 | page = alloc_buddy_huge_page(NULL, 0); | |
346 | if (!page) { | |
347 | /* | |
348 | * We were not able to allocate enough pages to | |
349 | * satisfy the entire reservation so we free what | |
350 | * we've allocated so far. | |
351 | */ | |
352 | spin_lock(&hugetlb_lock); | |
353 | needed = 0; | |
354 | goto free; | |
355 | } | |
356 | ||
357 | list_add(&page->lru, &surplus_list); | |
358 | } | |
359 | allocated += needed; | |
360 | ||
361 | /* | |
362 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
363 | * because either resv_huge_pages or free_huge_pages may have changed. | |
364 | */ | |
365 | spin_lock(&hugetlb_lock); | |
366 | needed = (resv_huge_pages + delta) - (free_huge_pages + allocated); | |
367 | if (needed > 0) | |
368 | goto retry; | |
369 | ||
370 | /* | |
371 | * The surplus_list now contains _at_least_ the number of extra pages | |
372 | * needed to accomodate the reservation. Add the appropriate number | |
373 | * of pages to the hugetlb pool and free the extras back to the buddy | |
ac09b3a1 AL |
374 | * allocator. Commit the entire reservation here to prevent another |
375 | * process from stealing the pages as they are added to the pool but | |
376 | * before they are reserved. | |
e4e574b7 AL |
377 | */ |
378 | needed += allocated; | |
ac09b3a1 | 379 | resv_huge_pages += delta; |
e4e574b7 AL |
380 | ret = 0; |
381 | free: | |
19fc3f0a | 382 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 383 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
384 | if ((--needed) < 0) |
385 | break; | |
e4e574b7 | 386 | list_del(&page->lru); |
19fc3f0a AL |
387 | enqueue_huge_page(page); |
388 | } | |
389 | ||
390 | /* Free unnecessary surplus pages to the buddy allocator */ | |
391 | if (!list_empty(&surplus_list)) { | |
392 | spin_unlock(&hugetlb_lock); | |
393 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { | |
394 | list_del(&page->lru); | |
af767cbd | 395 | /* |
2668db91 AL |
396 | * The page has a reference count of zero already, so |
397 | * call free_huge_page directly instead of using | |
398 | * put_page. This must be done with hugetlb_lock | |
af767cbd AL |
399 | * unlocked which is safe because free_huge_page takes |
400 | * hugetlb_lock before deciding how to free the page. | |
401 | */ | |
2668db91 | 402 | free_huge_page(page); |
af767cbd | 403 | } |
19fc3f0a | 404 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
405 | } |
406 | ||
407 | return ret; | |
408 | } | |
409 | ||
410 | /* | |
411 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
412 | * allocated to satisfy the reservation must be explicitly freed if they were | |
413 | * never used. | |
414 | */ | |
8cde045c | 415 | static void return_unused_surplus_pages(unsigned long unused_resv_pages) |
e4e574b7 AL |
416 | { |
417 | static int nid = -1; | |
418 | struct page *page; | |
419 | unsigned long nr_pages; | |
420 | ||
11320d17 NA |
421 | /* |
422 | * We want to release as many surplus pages as possible, spread | |
423 | * evenly across all nodes. Iterate across all nodes until we | |
424 | * can no longer free unreserved surplus pages. This occurs when | |
425 | * the nodes with surplus pages have no free pages. | |
426 | */ | |
427 | unsigned long remaining_iterations = num_online_nodes(); | |
428 | ||
ac09b3a1 AL |
429 | /* Uncommit the reservation */ |
430 | resv_huge_pages -= unused_resv_pages; | |
431 | ||
e4e574b7 AL |
432 | nr_pages = min(unused_resv_pages, surplus_huge_pages); |
433 | ||
11320d17 | 434 | while (remaining_iterations-- && nr_pages) { |
e4e574b7 AL |
435 | nid = next_node(nid, node_online_map); |
436 | if (nid == MAX_NUMNODES) | |
437 | nid = first_node(node_online_map); | |
438 | ||
439 | if (!surplus_huge_pages_node[nid]) | |
440 | continue; | |
441 | ||
442 | if (!list_empty(&hugepage_freelists[nid])) { | |
443 | page = list_entry(hugepage_freelists[nid].next, | |
444 | struct page, lru); | |
445 | list_del(&page->lru); | |
446 | update_and_free_page(page); | |
447 | free_huge_pages--; | |
448 | free_huge_pages_node[nid]--; | |
449 | surplus_huge_pages--; | |
450 | surplus_huge_pages_node[nid]--; | |
451 | nr_pages--; | |
11320d17 | 452 | remaining_iterations = num_online_nodes(); |
e4e574b7 AL |
453 | } |
454 | } | |
455 | } | |
456 | ||
348ea204 AL |
457 | |
458 | static struct page *alloc_huge_page_shared(struct vm_area_struct *vma, | |
459 | unsigned long addr) | |
1da177e4 | 460 | { |
348ea204 | 461 | struct page *page; |
1da177e4 LT |
462 | |
463 | spin_lock(&hugetlb_lock); | |
348e1e04 | 464 | page = dequeue_huge_page_vma(vma, addr); |
1da177e4 | 465 | spin_unlock(&hugetlb_lock); |
90d8b7e6 | 466 | return page ? page : ERR_PTR(-VM_FAULT_OOM); |
348ea204 | 467 | } |
b45b5bd6 | 468 | |
348ea204 AL |
469 | static struct page *alloc_huge_page_private(struct vm_area_struct *vma, |
470 | unsigned long addr) | |
471 | { | |
472 | struct page *page = NULL; | |
7893d1d5 | 473 | |
90d8b7e6 AL |
474 | if (hugetlb_get_quota(vma->vm_file->f_mapping, 1)) |
475 | return ERR_PTR(-VM_FAULT_SIGBUS); | |
476 | ||
348ea204 AL |
477 | spin_lock(&hugetlb_lock); |
478 | if (free_huge_pages > resv_huge_pages) | |
348e1e04 | 479 | page = dequeue_huge_page_vma(vma, addr); |
348ea204 | 480 | spin_unlock(&hugetlb_lock); |
68842c9b | 481 | if (!page) { |
7893d1d5 | 482 | page = alloc_buddy_huge_page(vma, addr); |
68842c9b KC |
483 | if (!page) { |
484 | hugetlb_put_quota(vma->vm_file->f_mapping, 1); | |
485 | return ERR_PTR(-VM_FAULT_OOM); | |
486 | } | |
487 | } | |
488 | return page; | |
348ea204 AL |
489 | } |
490 | ||
491 | static struct page *alloc_huge_page(struct vm_area_struct *vma, | |
492 | unsigned long addr) | |
493 | { | |
494 | struct page *page; | |
2fc39cec AL |
495 | struct address_space *mapping = vma->vm_file->f_mapping; |
496 | ||
348ea204 AL |
497 | if (vma->vm_flags & VM_MAYSHARE) |
498 | page = alloc_huge_page_shared(vma, addr); | |
499 | else | |
500 | page = alloc_huge_page_private(vma, addr); | |
90d8b7e6 AL |
501 | |
502 | if (!IS_ERR(page)) { | |
348ea204 | 503 | set_page_refcounted(page); |
2fc39cec | 504 | set_page_private(page, (unsigned long) mapping); |
90d8b7e6 AL |
505 | } |
506 | return page; | |
b45b5bd6 DG |
507 | } |
508 | ||
1da177e4 LT |
509 | static int __init hugetlb_init(void) |
510 | { | |
511 | unsigned long i; | |
1da177e4 | 512 | |
3c726f8d BH |
513 | if (HPAGE_SHIFT == 0) |
514 | return 0; | |
515 | ||
1da177e4 LT |
516 | for (i = 0; i < MAX_NUMNODES; ++i) |
517 | INIT_LIST_HEAD(&hugepage_freelists[i]); | |
518 | ||
63b4613c NA |
519 | hugetlb_next_nid = first_node(node_online_map); |
520 | ||
1da177e4 | 521 | for (i = 0; i < max_huge_pages; ++i) { |
a482289d | 522 | if (!alloc_fresh_huge_page()) |
1da177e4 | 523 | break; |
1da177e4 LT |
524 | } |
525 | max_huge_pages = free_huge_pages = nr_huge_pages = i; | |
526 | printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages); | |
527 | return 0; | |
528 | } | |
529 | module_init(hugetlb_init); | |
530 | ||
531 | static int __init hugetlb_setup(char *s) | |
532 | { | |
533 | if (sscanf(s, "%lu", &max_huge_pages) <= 0) | |
534 | max_huge_pages = 0; | |
535 | return 1; | |
536 | } | |
537 | __setup("hugepages=", hugetlb_setup); | |
538 | ||
8a630112 KC |
539 | static unsigned int cpuset_mems_nr(unsigned int *array) |
540 | { | |
541 | int node; | |
542 | unsigned int nr = 0; | |
543 | ||
544 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
545 | nr += array[node]; | |
546 | ||
547 | return nr; | |
548 | } | |
549 | ||
1da177e4 | 550 | #ifdef CONFIG_SYSCTL |
1da177e4 LT |
551 | #ifdef CONFIG_HIGHMEM |
552 | static void try_to_free_low(unsigned long count) | |
553 | { | |
4415cc8d CL |
554 | int i; |
555 | ||
1da177e4 LT |
556 | for (i = 0; i < MAX_NUMNODES; ++i) { |
557 | struct page *page, *next; | |
558 | list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) { | |
6b0c880d AL |
559 | if (count >= nr_huge_pages) |
560 | return; | |
1da177e4 LT |
561 | if (PageHighMem(page)) |
562 | continue; | |
563 | list_del(&page->lru); | |
564 | update_and_free_page(page); | |
1da177e4 | 565 | free_huge_pages--; |
4415cc8d | 566 | free_huge_pages_node[page_to_nid(page)]--; |
1da177e4 LT |
567 | } |
568 | } | |
569 | } | |
570 | #else | |
571 | static inline void try_to_free_low(unsigned long count) | |
572 | { | |
573 | } | |
574 | #endif | |
575 | ||
7893d1d5 | 576 | #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages) |
1da177e4 LT |
577 | static unsigned long set_max_huge_pages(unsigned long count) |
578 | { | |
7893d1d5 | 579 | unsigned long min_count, ret; |
1da177e4 | 580 | |
7893d1d5 AL |
581 | /* |
582 | * Increase the pool size | |
583 | * First take pages out of surplus state. Then make up the | |
584 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f NA |
585 | * |
586 | * We might race with alloc_buddy_huge_page() here and be unable | |
587 | * to convert a surplus huge page to a normal huge page. That is | |
588 | * not critical, though, it just means the overall size of the | |
589 | * pool might be one hugepage larger than it needs to be, but | |
590 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 591 | */ |
1da177e4 | 592 | spin_lock(&hugetlb_lock); |
7893d1d5 AL |
593 | while (surplus_huge_pages && count > persistent_huge_pages) { |
594 | if (!adjust_pool_surplus(-1)) | |
595 | break; | |
596 | } | |
597 | ||
598 | while (count > persistent_huge_pages) { | |
599 | int ret; | |
600 | /* | |
601 | * If this allocation races such that we no longer need the | |
602 | * page, free_huge_page will handle it by freeing the page | |
603 | * and reducing the surplus. | |
604 | */ | |
605 | spin_unlock(&hugetlb_lock); | |
606 | ret = alloc_fresh_huge_page(); | |
607 | spin_lock(&hugetlb_lock); | |
608 | if (!ret) | |
609 | goto out; | |
610 | ||
611 | } | |
7893d1d5 AL |
612 | |
613 | /* | |
614 | * Decrease the pool size | |
615 | * First return free pages to the buddy allocator (being careful | |
616 | * to keep enough around to satisfy reservations). Then place | |
617 | * pages into surplus state as needed so the pool will shrink | |
618 | * to the desired size as pages become free. | |
d1c3fb1f NA |
619 | * |
620 | * By placing pages into the surplus state independent of the | |
621 | * overcommit value, we are allowing the surplus pool size to | |
622 | * exceed overcommit. There are few sane options here. Since | |
623 | * alloc_buddy_huge_page() is checking the global counter, | |
624 | * though, we'll note that we're not allowed to exceed surplus | |
625 | * and won't grow the pool anywhere else. Not until one of the | |
626 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 627 | */ |
6b0c880d AL |
628 | min_count = resv_huge_pages + nr_huge_pages - free_huge_pages; |
629 | min_count = max(count, min_count); | |
7893d1d5 AL |
630 | try_to_free_low(min_count); |
631 | while (min_count < persistent_huge_pages) { | |
348e1e04 | 632 | struct page *page = dequeue_huge_page(); |
1da177e4 LT |
633 | if (!page) |
634 | break; | |
635 | update_and_free_page(page); | |
636 | } | |
7893d1d5 AL |
637 | while (count < persistent_huge_pages) { |
638 | if (!adjust_pool_surplus(1)) | |
639 | break; | |
640 | } | |
641 | out: | |
642 | ret = persistent_huge_pages; | |
1da177e4 | 643 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 644 | return ret; |
1da177e4 LT |
645 | } |
646 | ||
647 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, | |
648 | struct file *file, void __user *buffer, | |
649 | size_t *length, loff_t *ppos) | |
650 | { | |
651 | proc_doulongvec_minmax(table, write, file, buffer, length, ppos); | |
652 | max_huge_pages = set_max_huge_pages(max_huge_pages); | |
653 | return 0; | |
654 | } | |
396faf03 MG |
655 | |
656 | int hugetlb_treat_movable_handler(struct ctl_table *table, int write, | |
657 | struct file *file, void __user *buffer, | |
658 | size_t *length, loff_t *ppos) | |
659 | { | |
660 | proc_dointvec(table, write, file, buffer, length, ppos); | |
661 | if (hugepages_treat_as_movable) | |
662 | htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; | |
663 | else | |
664 | htlb_alloc_mask = GFP_HIGHUSER; | |
665 | return 0; | |
666 | } | |
667 | ||
a3d0c6aa NA |
668 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
669 | struct file *file, void __user *buffer, | |
670 | size_t *length, loff_t *ppos) | |
671 | { | |
a3d0c6aa | 672 | proc_doulongvec_minmax(table, write, file, buffer, length, ppos); |
064d9efe NA |
673 | spin_lock(&hugetlb_lock); |
674 | nr_overcommit_huge_pages = sysctl_overcommit_huge_pages; | |
a3d0c6aa NA |
675 | spin_unlock(&hugetlb_lock); |
676 | return 0; | |
677 | } | |
678 | ||
1da177e4 LT |
679 | #endif /* CONFIG_SYSCTL */ |
680 | ||
681 | int hugetlb_report_meminfo(char *buf) | |
682 | { | |
683 | return sprintf(buf, | |
684 | "HugePages_Total: %5lu\n" | |
685 | "HugePages_Free: %5lu\n" | |
a43a8c39 | 686 | "HugePages_Rsvd: %5lu\n" |
7893d1d5 | 687 | "HugePages_Surp: %5lu\n" |
1da177e4 LT |
688 | "Hugepagesize: %5lu kB\n", |
689 | nr_huge_pages, | |
690 | free_huge_pages, | |
a43a8c39 | 691 | resv_huge_pages, |
7893d1d5 | 692 | surplus_huge_pages, |
1da177e4 LT |
693 | HPAGE_SIZE/1024); |
694 | } | |
695 | ||
696 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
697 | { | |
698 | return sprintf(buf, | |
699 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
700 | "Node %d HugePages_Free: %5u\n" |
701 | "Node %d HugePages_Surp: %5u\n", | |
1da177e4 | 702 | nid, nr_huge_pages_node[nid], |
a1de0919 NA |
703 | nid, free_huge_pages_node[nid], |
704 | nid, surplus_huge_pages_node[nid]); | |
1da177e4 LT |
705 | } |
706 | ||
1da177e4 LT |
707 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
708 | unsigned long hugetlb_total_pages(void) | |
709 | { | |
710 | return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE); | |
711 | } | |
1da177e4 LT |
712 | |
713 | /* | |
714 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
715 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
716 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
717 | * this far. | |
718 | */ | |
d0217ac0 | 719 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
720 | { |
721 | BUG(); | |
d0217ac0 | 722 | return 0; |
1da177e4 LT |
723 | } |
724 | ||
725 | struct vm_operations_struct hugetlb_vm_ops = { | |
d0217ac0 | 726 | .fault = hugetlb_vm_op_fault, |
1da177e4 LT |
727 | }; |
728 | ||
1e8f889b DG |
729 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
730 | int writable) | |
63551ae0 DG |
731 | { |
732 | pte_t entry; | |
733 | ||
1e8f889b | 734 | if (writable) { |
63551ae0 DG |
735 | entry = |
736 | pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); | |
737 | } else { | |
738 | entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot)); | |
739 | } | |
740 | entry = pte_mkyoung(entry); | |
741 | entry = pte_mkhuge(entry); | |
742 | ||
743 | return entry; | |
744 | } | |
745 | ||
1e8f889b DG |
746 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
747 | unsigned long address, pte_t *ptep) | |
748 | { | |
749 | pte_t entry; | |
750 | ||
751 | entry = pte_mkwrite(pte_mkdirty(*ptep)); | |
8dab5241 BH |
752 | if (ptep_set_access_flags(vma, address, ptep, entry, 1)) { |
753 | update_mmu_cache(vma, address, entry); | |
8dab5241 | 754 | } |
1e8f889b DG |
755 | } |
756 | ||
757 | ||
63551ae0 DG |
758 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
759 | struct vm_area_struct *vma) | |
760 | { | |
761 | pte_t *src_pte, *dst_pte, entry; | |
762 | struct page *ptepage; | |
1c59827d | 763 | unsigned long addr; |
1e8f889b DG |
764 | int cow; |
765 | ||
766 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 767 | |
1c59827d | 768 | for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) { |
c74df32c HD |
769 | src_pte = huge_pte_offset(src, addr); |
770 | if (!src_pte) | |
771 | continue; | |
63551ae0 DG |
772 | dst_pte = huge_pte_alloc(dst, addr); |
773 | if (!dst_pte) | |
774 | goto nomem; | |
c5c99429 LW |
775 | |
776 | /* If the pagetables are shared don't copy or take references */ | |
777 | if (dst_pte == src_pte) | |
778 | continue; | |
779 | ||
c74df32c | 780 | spin_lock(&dst->page_table_lock); |
1c59827d | 781 | spin_lock(&src->page_table_lock); |
c74df32c | 782 | if (!pte_none(*src_pte)) { |
1e8f889b DG |
783 | if (cow) |
784 | ptep_set_wrprotect(src, addr, src_pte); | |
1c59827d HD |
785 | entry = *src_pte; |
786 | ptepage = pte_page(entry); | |
787 | get_page(ptepage); | |
1c59827d HD |
788 | set_huge_pte_at(dst, addr, dst_pte, entry); |
789 | } | |
790 | spin_unlock(&src->page_table_lock); | |
c74df32c | 791 | spin_unlock(&dst->page_table_lock); |
63551ae0 DG |
792 | } |
793 | return 0; | |
794 | ||
795 | nomem: | |
796 | return -ENOMEM; | |
797 | } | |
798 | ||
502717f4 KC |
799 | void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
800 | unsigned long end) | |
63551ae0 DG |
801 | { |
802 | struct mm_struct *mm = vma->vm_mm; | |
803 | unsigned long address; | |
c7546f8f | 804 | pte_t *ptep; |
63551ae0 DG |
805 | pte_t pte; |
806 | struct page *page; | |
fe1668ae | 807 | struct page *tmp; |
c0a499c2 KC |
808 | /* |
809 | * A page gathering list, protected by per file i_mmap_lock. The | |
810 | * lock is used to avoid list corruption from multiple unmapping | |
811 | * of the same page since we are using page->lru. | |
812 | */ | |
fe1668ae | 813 | LIST_HEAD(page_list); |
63551ae0 DG |
814 | |
815 | WARN_ON(!is_vm_hugetlb_page(vma)); | |
816 | BUG_ON(start & ~HPAGE_MASK); | |
817 | BUG_ON(end & ~HPAGE_MASK); | |
818 | ||
508034a3 | 819 | spin_lock(&mm->page_table_lock); |
63551ae0 | 820 | for (address = start; address < end; address += HPAGE_SIZE) { |
c7546f8f | 821 | ptep = huge_pte_offset(mm, address); |
4c887265 | 822 | if (!ptep) |
c7546f8f DG |
823 | continue; |
824 | ||
39dde65c KC |
825 | if (huge_pmd_unshare(mm, &address, ptep)) |
826 | continue; | |
827 | ||
c7546f8f | 828 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
63551ae0 DG |
829 | if (pte_none(pte)) |
830 | continue; | |
c7546f8f | 831 | |
63551ae0 | 832 | page = pte_page(pte); |
6649a386 KC |
833 | if (pte_dirty(pte)) |
834 | set_page_dirty(page); | |
fe1668ae | 835 | list_add(&page->lru, &page_list); |
63551ae0 | 836 | } |
1da177e4 | 837 | spin_unlock(&mm->page_table_lock); |
508034a3 | 838 | flush_tlb_range(vma, start, end); |
fe1668ae KC |
839 | list_for_each_entry_safe(page, tmp, &page_list, lru) { |
840 | list_del(&page->lru); | |
841 | put_page(page); | |
842 | } | |
1da177e4 | 843 | } |
63551ae0 | 844 | |
502717f4 KC |
845 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
846 | unsigned long end) | |
847 | { | |
848 | /* | |
849 | * It is undesirable to test vma->vm_file as it should be non-null | |
850 | * for valid hugetlb area. However, vm_file will be NULL in the error | |
851 | * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails, | |
852 | * do_mmap_pgoff() nullifies vma->vm_file before calling this function | |
853 | * to clean up. Since no pte has actually been setup, it is safe to | |
854 | * do nothing in this case. | |
855 | */ | |
856 | if (vma->vm_file) { | |
857 | spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); | |
858 | __unmap_hugepage_range(vma, start, end); | |
859 | spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); | |
860 | } | |
861 | } | |
862 | ||
1e8f889b DG |
863 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
864 | unsigned long address, pte_t *ptep, pte_t pte) | |
865 | { | |
866 | struct page *old_page, *new_page; | |
79ac6ba4 | 867 | int avoidcopy; |
1e8f889b DG |
868 | |
869 | old_page = pte_page(pte); | |
870 | ||
871 | /* If no-one else is actually using this page, avoid the copy | |
872 | * and just make the page writable */ | |
873 | avoidcopy = (page_count(old_page) == 1); | |
874 | if (avoidcopy) { | |
875 | set_huge_ptep_writable(vma, address, ptep); | |
83c54070 | 876 | return 0; |
1e8f889b DG |
877 | } |
878 | ||
879 | page_cache_get(old_page); | |
5da7ca86 | 880 | new_page = alloc_huge_page(vma, address); |
1e8f889b | 881 | |
2fc39cec | 882 | if (IS_ERR(new_page)) { |
1e8f889b | 883 | page_cache_release(old_page); |
2fc39cec | 884 | return -PTR_ERR(new_page); |
1e8f889b DG |
885 | } |
886 | ||
887 | spin_unlock(&mm->page_table_lock); | |
9de455b2 | 888 | copy_huge_page(new_page, old_page, address, vma); |
0ed361de | 889 | __SetPageUptodate(new_page); |
1e8f889b DG |
890 | spin_lock(&mm->page_table_lock); |
891 | ||
892 | ptep = huge_pte_offset(mm, address & HPAGE_MASK); | |
893 | if (likely(pte_same(*ptep, pte))) { | |
894 | /* Break COW */ | |
8fe627ec | 895 | huge_ptep_clear_flush(vma, address, ptep); |
1e8f889b DG |
896 | set_huge_pte_at(mm, address, ptep, |
897 | make_huge_pte(vma, new_page, 1)); | |
898 | /* Make the old page be freed below */ | |
899 | new_page = old_page; | |
900 | } | |
901 | page_cache_release(new_page); | |
902 | page_cache_release(old_page); | |
83c54070 | 903 | return 0; |
1e8f889b DG |
904 | } |
905 | ||
a1ed3dda | 906 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
1e8f889b | 907 | unsigned long address, pte_t *ptep, int write_access) |
ac9b9c66 HD |
908 | { |
909 | int ret = VM_FAULT_SIGBUS; | |
4c887265 AL |
910 | unsigned long idx; |
911 | unsigned long size; | |
4c887265 AL |
912 | struct page *page; |
913 | struct address_space *mapping; | |
1e8f889b | 914 | pte_t new_pte; |
4c887265 | 915 | |
4c887265 AL |
916 | mapping = vma->vm_file->f_mapping; |
917 | idx = ((address - vma->vm_start) >> HPAGE_SHIFT) | |
918 | + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT)); | |
919 | ||
920 | /* | |
921 | * Use page lock to guard against racing truncation | |
922 | * before we get page_table_lock. | |
923 | */ | |
6bda666a CL |
924 | retry: |
925 | page = find_lock_page(mapping, idx); | |
926 | if (!page) { | |
ebed4bfc HD |
927 | size = i_size_read(mapping->host) >> HPAGE_SHIFT; |
928 | if (idx >= size) | |
929 | goto out; | |
6bda666a | 930 | page = alloc_huge_page(vma, address); |
2fc39cec AL |
931 | if (IS_ERR(page)) { |
932 | ret = -PTR_ERR(page); | |
6bda666a CL |
933 | goto out; |
934 | } | |
79ac6ba4 | 935 | clear_huge_page(page, address); |
0ed361de | 936 | __SetPageUptodate(page); |
ac9b9c66 | 937 | |
6bda666a CL |
938 | if (vma->vm_flags & VM_SHARED) { |
939 | int err; | |
45c682a6 | 940 | struct inode *inode = mapping->host; |
6bda666a CL |
941 | |
942 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
943 | if (err) { | |
944 | put_page(page); | |
6bda666a CL |
945 | if (err == -EEXIST) |
946 | goto retry; | |
947 | goto out; | |
948 | } | |
45c682a6 KC |
949 | |
950 | spin_lock(&inode->i_lock); | |
951 | inode->i_blocks += BLOCKS_PER_HUGEPAGE; | |
952 | spin_unlock(&inode->i_lock); | |
6bda666a CL |
953 | } else |
954 | lock_page(page); | |
955 | } | |
1e8f889b | 956 | |
ac9b9c66 | 957 | spin_lock(&mm->page_table_lock); |
4c887265 AL |
958 | size = i_size_read(mapping->host) >> HPAGE_SHIFT; |
959 | if (idx >= size) | |
960 | goto backout; | |
961 | ||
83c54070 | 962 | ret = 0; |
86e5216f | 963 | if (!pte_none(*ptep)) |
4c887265 AL |
964 | goto backout; |
965 | ||
1e8f889b DG |
966 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
967 | && (vma->vm_flags & VM_SHARED))); | |
968 | set_huge_pte_at(mm, address, ptep, new_pte); | |
969 | ||
970 | if (write_access && !(vma->vm_flags & VM_SHARED)) { | |
971 | /* Optimization, do the COW without a second fault */ | |
972 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte); | |
973 | } | |
974 | ||
ac9b9c66 | 975 | spin_unlock(&mm->page_table_lock); |
4c887265 AL |
976 | unlock_page(page); |
977 | out: | |
ac9b9c66 | 978 | return ret; |
4c887265 AL |
979 | |
980 | backout: | |
981 | spin_unlock(&mm->page_table_lock); | |
4c887265 AL |
982 | unlock_page(page); |
983 | put_page(page); | |
984 | goto out; | |
ac9b9c66 HD |
985 | } |
986 | ||
86e5216f AL |
987 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
988 | unsigned long address, int write_access) | |
989 | { | |
990 | pte_t *ptep; | |
991 | pte_t entry; | |
1e8f889b | 992 | int ret; |
3935baa9 | 993 | static DEFINE_MUTEX(hugetlb_instantiation_mutex); |
86e5216f AL |
994 | |
995 | ptep = huge_pte_alloc(mm, address); | |
996 | if (!ptep) | |
997 | return VM_FAULT_OOM; | |
998 | ||
3935baa9 DG |
999 | /* |
1000 | * Serialize hugepage allocation and instantiation, so that we don't | |
1001 | * get spurious allocation failures if two CPUs race to instantiate | |
1002 | * the same page in the page cache. | |
1003 | */ | |
1004 | mutex_lock(&hugetlb_instantiation_mutex); | |
86e5216f | 1005 | entry = *ptep; |
3935baa9 DG |
1006 | if (pte_none(entry)) { |
1007 | ret = hugetlb_no_page(mm, vma, address, ptep, write_access); | |
1008 | mutex_unlock(&hugetlb_instantiation_mutex); | |
1009 | return ret; | |
1010 | } | |
86e5216f | 1011 | |
83c54070 | 1012 | ret = 0; |
1e8f889b DG |
1013 | |
1014 | spin_lock(&mm->page_table_lock); | |
1015 | /* Check for a racing update before calling hugetlb_cow */ | |
1016 | if (likely(pte_same(entry, *ptep))) | |
1017 | if (write_access && !pte_write(entry)) | |
1018 | ret = hugetlb_cow(mm, vma, address, ptep, entry); | |
1019 | spin_unlock(&mm->page_table_lock); | |
3935baa9 | 1020 | mutex_unlock(&hugetlb_instantiation_mutex); |
1e8f889b DG |
1021 | |
1022 | return ret; | |
86e5216f AL |
1023 | } |
1024 | ||
63551ae0 DG |
1025 | int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
1026 | struct page **pages, struct vm_area_struct **vmas, | |
5b23dbe8 AL |
1027 | unsigned long *position, int *length, int i, |
1028 | int write) | |
63551ae0 | 1029 | { |
d5d4b0aa KC |
1030 | unsigned long pfn_offset; |
1031 | unsigned long vaddr = *position; | |
63551ae0 DG |
1032 | int remainder = *length; |
1033 | ||
1c59827d | 1034 | spin_lock(&mm->page_table_lock); |
63551ae0 | 1035 | while (vaddr < vma->vm_end && remainder) { |
4c887265 AL |
1036 | pte_t *pte; |
1037 | struct page *page; | |
63551ae0 | 1038 | |
4c887265 AL |
1039 | /* |
1040 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
1041 | * each hugepage. We have to make * sure we get the | |
1042 | * first, for the page indexing below to work. | |
1043 | */ | |
1044 | pte = huge_pte_offset(mm, vaddr & HPAGE_MASK); | |
63551ae0 | 1045 | |
72fad713 | 1046 | if (!pte || pte_none(*pte) || (write && !pte_write(*pte))) { |
4c887265 | 1047 | int ret; |
63551ae0 | 1048 | |
4c887265 | 1049 | spin_unlock(&mm->page_table_lock); |
5b23dbe8 | 1050 | ret = hugetlb_fault(mm, vma, vaddr, write); |
4c887265 | 1051 | spin_lock(&mm->page_table_lock); |
a89182c7 | 1052 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 1053 | continue; |
63551ae0 | 1054 | |
4c887265 AL |
1055 | remainder = 0; |
1056 | if (!i) | |
1057 | i = -EFAULT; | |
1058 | break; | |
1059 | } | |
1060 | ||
d5d4b0aa KC |
1061 | pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT; |
1062 | page = pte_page(*pte); | |
1063 | same_page: | |
d6692183 KC |
1064 | if (pages) { |
1065 | get_page(page); | |
d5d4b0aa | 1066 | pages[i] = page + pfn_offset; |
d6692183 | 1067 | } |
63551ae0 DG |
1068 | |
1069 | if (vmas) | |
1070 | vmas[i] = vma; | |
1071 | ||
1072 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 1073 | ++pfn_offset; |
63551ae0 DG |
1074 | --remainder; |
1075 | ++i; | |
d5d4b0aa KC |
1076 | if (vaddr < vma->vm_end && remainder && |
1077 | pfn_offset < HPAGE_SIZE/PAGE_SIZE) { | |
1078 | /* | |
1079 | * We use pfn_offset to avoid touching the pageframes | |
1080 | * of this compound page. | |
1081 | */ | |
1082 | goto same_page; | |
1083 | } | |
63551ae0 | 1084 | } |
1c59827d | 1085 | spin_unlock(&mm->page_table_lock); |
63551ae0 DG |
1086 | *length = remainder; |
1087 | *position = vaddr; | |
1088 | ||
1089 | return i; | |
1090 | } | |
8f860591 ZY |
1091 | |
1092 | void hugetlb_change_protection(struct vm_area_struct *vma, | |
1093 | unsigned long address, unsigned long end, pgprot_t newprot) | |
1094 | { | |
1095 | struct mm_struct *mm = vma->vm_mm; | |
1096 | unsigned long start = address; | |
1097 | pte_t *ptep; | |
1098 | pte_t pte; | |
1099 | ||
1100 | BUG_ON(address >= end); | |
1101 | flush_cache_range(vma, address, end); | |
1102 | ||
39dde65c | 1103 | spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); |
8f860591 ZY |
1104 | spin_lock(&mm->page_table_lock); |
1105 | for (; address < end; address += HPAGE_SIZE) { | |
1106 | ptep = huge_pte_offset(mm, address); | |
1107 | if (!ptep) | |
1108 | continue; | |
39dde65c KC |
1109 | if (huge_pmd_unshare(mm, &address, ptep)) |
1110 | continue; | |
8f860591 ZY |
1111 | if (!pte_none(*ptep)) { |
1112 | pte = huge_ptep_get_and_clear(mm, address, ptep); | |
1113 | pte = pte_mkhuge(pte_modify(pte, newprot)); | |
1114 | set_huge_pte_at(mm, address, ptep, pte); | |
8f860591 ZY |
1115 | } |
1116 | } | |
1117 | spin_unlock(&mm->page_table_lock); | |
39dde65c | 1118 | spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); |
8f860591 ZY |
1119 | |
1120 | flush_tlb_range(vma, start, end); | |
1121 | } | |
1122 | ||
a43a8c39 KC |
1123 | struct file_region { |
1124 | struct list_head link; | |
1125 | long from; | |
1126 | long to; | |
1127 | }; | |
1128 | ||
1129 | static long region_add(struct list_head *head, long f, long t) | |
1130 | { | |
1131 | struct file_region *rg, *nrg, *trg; | |
1132 | ||
1133 | /* Locate the region we are either in or before. */ | |
1134 | list_for_each_entry(rg, head, link) | |
1135 | if (f <= rg->to) | |
1136 | break; | |
1137 | ||
1138 | /* Round our left edge to the current segment if it encloses us. */ | |
1139 | if (f > rg->from) | |
1140 | f = rg->from; | |
1141 | ||
1142 | /* Check for and consume any regions we now overlap with. */ | |
1143 | nrg = rg; | |
1144 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
1145 | if (&rg->link == head) | |
1146 | break; | |
1147 | if (rg->from > t) | |
1148 | break; | |
1149 | ||
1150 | /* If this area reaches higher then extend our area to | |
1151 | * include it completely. If this is not the first area | |
1152 | * which we intend to reuse, free it. */ | |
1153 | if (rg->to > t) | |
1154 | t = rg->to; | |
1155 | if (rg != nrg) { | |
1156 | list_del(&rg->link); | |
1157 | kfree(rg); | |
1158 | } | |
1159 | } | |
1160 | nrg->from = f; | |
1161 | nrg->to = t; | |
1162 | return 0; | |
1163 | } | |
1164 | ||
1165 | static long region_chg(struct list_head *head, long f, long t) | |
1166 | { | |
1167 | struct file_region *rg, *nrg; | |
1168 | long chg = 0; | |
1169 | ||
1170 | /* Locate the region we are before or in. */ | |
1171 | list_for_each_entry(rg, head, link) | |
1172 | if (f <= rg->to) | |
1173 | break; | |
1174 | ||
1175 | /* If we are below the current region then a new region is required. | |
1176 | * Subtle, allocate a new region at the position but make it zero | |
183ff22b | 1177 | * size such that we can guarantee to record the reservation. */ |
a43a8c39 KC |
1178 | if (&rg->link == head || t < rg->from) { |
1179 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
c80544dc | 1180 | if (!nrg) |
a43a8c39 KC |
1181 | return -ENOMEM; |
1182 | nrg->from = f; | |
1183 | nrg->to = f; | |
1184 | INIT_LIST_HEAD(&nrg->link); | |
1185 | list_add(&nrg->link, rg->link.prev); | |
1186 | ||
1187 | return t - f; | |
1188 | } | |
1189 | ||
1190 | /* Round our left edge to the current segment if it encloses us. */ | |
1191 | if (f > rg->from) | |
1192 | f = rg->from; | |
1193 | chg = t - f; | |
1194 | ||
1195 | /* Check for and consume any regions we now overlap with. */ | |
1196 | list_for_each_entry(rg, rg->link.prev, link) { | |
1197 | if (&rg->link == head) | |
1198 | break; | |
1199 | if (rg->from > t) | |
1200 | return chg; | |
1201 | ||
1202 | /* We overlap with this area, if it extends futher than | |
1203 | * us then we must extend ourselves. Account for its | |
1204 | * existing reservation. */ | |
1205 | if (rg->to > t) { | |
1206 | chg += rg->to - t; | |
1207 | t = rg->to; | |
1208 | } | |
1209 | chg -= rg->to - rg->from; | |
1210 | } | |
1211 | return chg; | |
1212 | } | |
1213 | ||
1214 | static long region_truncate(struct list_head *head, long end) | |
1215 | { | |
1216 | struct file_region *rg, *trg; | |
1217 | long chg = 0; | |
1218 | ||
1219 | /* Locate the region we are either in or before. */ | |
1220 | list_for_each_entry(rg, head, link) | |
1221 | if (end <= rg->to) | |
1222 | break; | |
1223 | if (&rg->link == head) | |
1224 | return 0; | |
1225 | ||
1226 | /* If we are in the middle of a region then adjust it. */ | |
1227 | if (end > rg->from) { | |
1228 | chg = rg->to - end; | |
1229 | rg->to = end; | |
1230 | rg = list_entry(rg->link.next, typeof(*rg), link); | |
1231 | } | |
1232 | ||
1233 | /* Drop any remaining regions. */ | |
1234 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
1235 | if (&rg->link == head) | |
1236 | break; | |
1237 | chg += rg->to - rg->from; | |
1238 | list_del(&rg->link); | |
1239 | kfree(rg); | |
1240 | } | |
1241 | return chg; | |
1242 | } | |
1243 | ||
1244 | static int hugetlb_acct_memory(long delta) | |
1245 | { | |
1246 | int ret = -ENOMEM; | |
1247 | ||
1248 | spin_lock(&hugetlb_lock); | |
8a630112 KC |
1249 | /* |
1250 | * When cpuset is configured, it breaks the strict hugetlb page | |
1251 | * reservation as the accounting is done on a global variable. Such | |
1252 | * reservation is completely rubbish in the presence of cpuset because | |
1253 | * the reservation is not checked against page availability for the | |
1254 | * current cpuset. Application can still potentially OOM'ed by kernel | |
1255 | * with lack of free htlb page in cpuset that the task is in. | |
1256 | * Attempt to enforce strict accounting with cpuset is almost | |
1257 | * impossible (or too ugly) because cpuset is too fluid that | |
1258 | * task or memory node can be dynamically moved between cpusets. | |
1259 | * | |
1260 | * The change of semantics for shared hugetlb mapping with cpuset is | |
1261 | * undesirable. However, in order to preserve some of the semantics, | |
1262 | * we fall back to check against current free page availability as | |
1263 | * a best attempt and hopefully to minimize the impact of changing | |
1264 | * semantics that cpuset has. | |
1265 | */ | |
e4e574b7 AL |
1266 | if (delta > 0) { |
1267 | if (gather_surplus_pages(delta) < 0) | |
1268 | goto out; | |
1269 | ||
ac09b3a1 AL |
1270 | if (delta > cpuset_mems_nr(free_huge_pages_node)) { |
1271 | return_unused_surplus_pages(delta); | |
e4e574b7 | 1272 | goto out; |
ac09b3a1 | 1273 | } |
e4e574b7 AL |
1274 | } |
1275 | ||
1276 | ret = 0; | |
e4e574b7 AL |
1277 | if (delta < 0) |
1278 | return_unused_surplus_pages((unsigned long) -delta); | |
1279 | ||
1280 | out: | |
1281 | spin_unlock(&hugetlb_lock); | |
1282 | return ret; | |
1283 | } | |
1284 | ||
1285 | int hugetlb_reserve_pages(struct inode *inode, long from, long to) | |
1286 | { | |
1287 | long ret, chg; | |
1288 | ||
1289 | chg = region_chg(&inode->i_mapping->private_list, from, to); | |
1290 | if (chg < 0) | |
1291 | return chg; | |
8a630112 | 1292 | |
90d8b7e6 AL |
1293 | if (hugetlb_get_quota(inode->i_mapping, chg)) |
1294 | return -ENOSPC; | |
a43a8c39 | 1295 | ret = hugetlb_acct_memory(chg); |
68842c9b KC |
1296 | if (ret < 0) { |
1297 | hugetlb_put_quota(inode->i_mapping, chg); | |
a43a8c39 | 1298 | return ret; |
68842c9b | 1299 | } |
a43a8c39 KC |
1300 | region_add(&inode->i_mapping->private_list, from, to); |
1301 | return 0; | |
1302 | } | |
1303 | ||
1304 | void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) | |
1305 | { | |
1306 | long chg = region_truncate(&inode->i_mapping->private_list, offset); | |
45c682a6 KC |
1307 | |
1308 | spin_lock(&inode->i_lock); | |
1309 | inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed; | |
1310 | spin_unlock(&inode->i_lock); | |
1311 | ||
90d8b7e6 AL |
1312 | hugetlb_put_quota(inode->i_mapping, (chg - freed)); |
1313 | hugetlb_acct_memory(-(chg - freed)); | |
a43a8c39 | 1314 | } |