]>
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; |
396faf03 | 98 | struct zonelist *zonelist = huge_zonelist(vma, address, |
480eccf9 | 99 | htlb_alloc_mask, &mpol); |
96df9333 | 100 | struct zone **z; |
1da177e4 | 101 | |
96df9333 | 102 | for (z = zonelist->zones; *z; z++) { |
89fa3024 | 103 | nid = zone_to_nid(*z); |
396faf03 | 104 | if (cpuset_zone_allowed_softwall(*z, htlb_alloc_mask) && |
3abf7afd AM |
105 | !list_empty(&hugepage_freelists[nid])) { |
106 | page = list_entry(hugepage_freelists[nid].next, | |
107 | struct page, lru); | |
108 | list_del(&page->lru); | |
109 | free_huge_pages--; | |
110 | free_huge_pages_node[nid]--; | |
e4e574b7 AL |
111 | if (vma && vma->vm_flags & VM_MAYSHARE) |
112 | resv_huge_pages--; | |
5ab3ee7b | 113 | break; |
3abf7afd | 114 | } |
1da177e4 | 115 | } |
480eccf9 | 116 | mpol_free(mpol); /* unref if mpol !NULL */ |
1da177e4 LT |
117 | return page; |
118 | } | |
119 | ||
6af2acb6 AL |
120 | static void update_and_free_page(struct page *page) |
121 | { | |
122 | int i; | |
123 | nr_huge_pages--; | |
124 | nr_huge_pages_node[page_to_nid(page)]--; | |
125 | for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) { | |
126 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced | | |
127 | 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved | | |
128 | 1 << PG_private | 1<< PG_writeback); | |
129 | } | |
130 | set_compound_page_dtor(page, NULL); | |
131 | set_page_refcounted(page); | |
132 | __free_pages(page, HUGETLB_PAGE_ORDER); | |
133 | } | |
134 | ||
27a85ef1 DG |
135 | static void free_huge_page(struct page *page) |
136 | { | |
7893d1d5 | 137 | int nid = page_to_nid(page); |
c79fb75e | 138 | struct address_space *mapping; |
27a85ef1 | 139 | |
c79fb75e | 140 | mapping = (struct address_space *) page_private(page); |
e5df70ab | 141 | set_page_private(page, 0); |
7893d1d5 | 142 | BUG_ON(page_count(page)); |
27a85ef1 DG |
143 | INIT_LIST_HEAD(&page->lru); |
144 | ||
145 | spin_lock(&hugetlb_lock); | |
7893d1d5 AL |
146 | if (surplus_huge_pages_node[nid]) { |
147 | update_and_free_page(page); | |
148 | surplus_huge_pages--; | |
149 | surplus_huge_pages_node[nid]--; | |
150 | } else { | |
151 | enqueue_huge_page(page); | |
152 | } | |
27a85ef1 | 153 | spin_unlock(&hugetlb_lock); |
c79fb75e | 154 | if (mapping) |
9a119c05 | 155 | hugetlb_put_quota(mapping, 1); |
27a85ef1 DG |
156 | } |
157 | ||
7893d1d5 AL |
158 | /* |
159 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
160 | * balanced by operating on them in a round-robin fashion. | |
161 | * Returns 1 if an adjustment was made. | |
162 | */ | |
163 | static int adjust_pool_surplus(int delta) | |
164 | { | |
165 | static int prev_nid; | |
166 | int nid = prev_nid; | |
167 | int ret = 0; | |
168 | ||
169 | VM_BUG_ON(delta != -1 && delta != 1); | |
170 | do { | |
171 | nid = next_node(nid, node_online_map); | |
172 | if (nid == MAX_NUMNODES) | |
173 | nid = first_node(node_online_map); | |
174 | ||
175 | /* To shrink on this node, there must be a surplus page */ | |
176 | if (delta < 0 && !surplus_huge_pages_node[nid]) | |
177 | continue; | |
178 | /* Surplus cannot exceed the total number of pages */ | |
179 | if (delta > 0 && surplus_huge_pages_node[nid] >= | |
180 | nr_huge_pages_node[nid]) | |
181 | continue; | |
182 | ||
183 | surplus_huge_pages += delta; | |
184 | surplus_huge_pages_node[nid] += delta; | |
185 | ret = 1; | |
186 | break; | |
187 | } while (nid != prev_nid); | |
188 | ||
189 | prev_nid = nid; | |
190 | return ret; | |
191 | } | |
192 | ||
63b4613c | 193 | static struct page *alloc_fresh_huge_page_node(int nid) |
1da177e4 | 194 | { |
1da177e4 | 195 | struct page *page; |
f96efd58 | 196 | |
63b4613c NA |
197 | page = alloc_pages_node(nid, |
198 | htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|__GFP_NOWARN, | |
199 | HUGETLB_PAGE_ORDER); | |
1da177e4 | 200 | if (page) { |
33f2ef89 | 201 | set_compound_page_dtor(page, free_huge_page); |
0bd0f9fb | 202 | spin_lock(&hugetlb_lock); |
1da177e4 | 203 | nr_huge_pages++; |
63b4613c | 204 | nr_huge_pages_node[nid]++; |
0bd0f9fb | 205 | spin_unlock(&hugetlb_lock); |
a482289d | 206 | put_page(page); /* free it into the hugepage allocator */ |
1da177e4 | 207 | } |
63b4613c NA |
208 | |
209 | return page; | |
210 | } | |
211 | ||
212 | static int alloc_fresh_huge_page(void) | |
213 | { | |
214 | struct page *page; | |
215 | int start_nid; | |
216 | int next_nid; | |
217 | int ret = 0; | |
218 | ||
219 | start_nid = hugetlb_next_nid; | |
220 | ||
221 | do { | |
222 | page = alloc_fresh_huge_page_node(hugetlb_next_nid); | |
223 | if (page) | |
224 | ret = 1; | |
225 | /* | |
226 | * Use a helper variable to find the next node and then | |
227 | * copy it back to hugetlb_next_nid afterwards: | |
228 | * otherwise there's a window in which a racer might | |
229 | * pass invalid nid MAX_NUMNODES to alloc_pages_node. | |
230 | * But we don't need to use a spin_lock here: it really | |
231 | * doesn't matter if occasionally a racer chooses the | |
232 | * same nid as we do. Move nid forward in the mask even | |
233 | * if we just successfully allocated a hugepage so that | |
234 | * the next caller gets hugepages on the next node. | |
235 | */ | |
236 | next_nid = next_node(hugetlb_next_nid, node_online_map); | |
237 | if (next_nid == MAX_NUMNODES) | |
238 | next_nid = first_node(node_online_map); | |
239 | hugetlb_next_nid = next_nid; | |
240 | } while (!page && hugetlb_next_nid != start_nid); | |
241 | ||
242 | return ret; | |
1da177e4 LT |
243 | } |
244 | ||
7893d1d5 AL |
245 | static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma, |
246 | unsigned long address) | |
247 | { | |
248 | struct page *page; | |
d1c3fb1f | 249 | unsigned int nid; |
7893d1d5 | 250 | |
d1c3fb1f NA |
251 | /* |
252 | * Assume we will successfully allocate the surplus page to | |
253 | * prevent racing processes from causing the surplus to exceed | |
254 | * overcommit | |
255 | * | |
256 | * This however introduces a different race, where a process B | |
257 | * tries to grow the static hugepage pool while alloc_pages() is | |
258 | * called by process A. B will only examine the per-node | |
259 | * counters in determining if surplus huge pages can be | |
260 | * converted to normal huge pages in adjust_pool_surplus(). A | |
261 | * won't be able to increment the per-node counter, until the | |
262 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
263 | * no more huge pages can be converted from surplus to normal | |
264 | * state (and doesn't try to convert again). Thus, we have a | |
265 | * case where a surplus huge page exists, the pool is grown, and | |
266 | * the surplus huge page still exists after, even though it | |
267 | * should just have been converted to a normal huge page. This | |
268 | * does not leak memory, though, as the hugepage will be freed | |
269 | * once it is out of use. It also does not allow the counters to | |
270 | * go out of whack in adjust_pool_surplus() as we don't modify | |
271 | * the node values until we've gotten the hugepage and only the | |
272 | * per-node value is checked there. | |
273 | */ | |
274 | spin_lock(&hugetlb_lock); | |
275 | if (surplus_huge_pages >= nr_overcommit_huge_pages) { | |
276 | spin_unlock(&hugetlb_lock); | |
277 | return NULL; | |
278 | } else { | |
279 | nr_huge_pages++; | |
280 | surplus_huge_pages++; | |
281 | } | |
282 | spin_unlock(&hugetlb_lock); | |
283 | ||
7893d1d5 AL |
284 | page = alloc_pages(htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN, |
285 | HUGETLB_PAGE_ORDER); | |
d1c3fb1f NA |
286 | |
287 | spin_lock(&hugetlb_lock); | |
7893d1d5 | 288 | if (page) { |
2668db91 AL |
289 | /* |
290 | * This page is now managed by the hugetlb allocator and has | |
291 | * no users -- drop the buddy allocator's reference. | |
292 | */ | |
293 | put_page_testzero(page); | |
294 | VM_BUG_ON(page_count(page)); | |
d1c3fb1f | 295 | nid = page_to_nid(page); |
7893d1d5 | 296 | set_compound_page_dtor(page, free_huge_page); |
d1c3fb1f NA |
297 | /* |
298 | * We incremented the global counters already | |
299 | */ | |
300 | nr_huge_pages_node[nid]++; | |
301 | surplus_huge_pages_node[nid]++; | |
302 | } else { | |
303 | nr_huge_pages--; | |
304 | surplus_huge_pages--; | |
7893d1d5 | 305 | } |
d1c3fb1f | 306 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
307 | |
308 | return page; | |
309 | } | |
310 | ||
e4e574b7 AL |
311 | /* |
312 | * Increase the hugetlb pool such that it can accomodate a reservation | |
313 | * of size 'delta'. | |
314 | */ | |
315 | static int gather_surplus_pages(int delta) | |
316 | { | |
317 | struct list_head surplus_list; | |
318 | struct page *page, *tmp; | |
319 | int ret, i; | |
320 | int needed, allocated; | |
321 | ||
322 | needed = (resv_huge_pages + delta) - free_huge_pages; | |
ac09b3a1 AL |
323 | if (needed <= 0) { |
324 | resv_huge_pages += delta; | |
e4e574b7 | 325 | return 0; |
ac09b3a1 | 326 | } |
e4e574b7 AL |
327 | |
328 | allocated = 0; | |
329 | INIT_LIST_HEAD(&surplus_list); | |
330 | ||
331 | ret = -ENOMEM; | |
332 | retry: | |
333 | spin_unlock(&hugetlb_lock); | |
334 | for (i = 0; i < needed; i++) { | |
335 | page = alloc_buddy_huge_page(NULL, 0); | |
336 | if (!page) { | |
337 | /* | |
338 | * We were not able to allocate enough pages to | |
339 | * satisfy the entire reservation so we free what | |
340 | * we've allocated so far. | |
341 | */ | |
342 | spin_lock(&hugetlb_lock); | |
343 | needed = 0; | |
344 | goto free; | |
345 | } | |
346 | ||
347 | list_add(&page->lru, &surplus_list); | |
348 | } | |
349 | allocated += needed; | |
350 | ||
351 | /* | |
352 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
353 | * because either resv_huge_pages or free_huge_pages may have changed. | |
354 | */ | |
355 | spin_lock(&hugetlb_lock); | |
356 | needed = (resv_huge_pages + delta) - (free_huge_pages + allocated); | |
357 | if (needed > 0) | |
358 | goto retry; | |
359 | ||
360 | /* | |
361 | * The surplus_list now contains _at_least_ the number of extra pages | |
362 | * needed to accomodate the reservation. Add the appropriate number | |
363 | * of pages to the hugetlb pool and free the extras back to the buddy | |
ac09b3a1 AL |
364 | * allocator. Commit the entire reservation here to prevent another |
365 | * process from stealing the pages as they are added to the pool but | |
366 | * before they are reserved. | |
e4e574b7 AL |
367 | */ |
368 | needed += allocated; | |
ac09b3a1 | 369 | resv_huge_pages += delta; |
e4e574b7 AL |
370 | ret = 0; |
371 | free: | |
372 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { | |
373 | list_del(&page->lru); | |
374 | if ((--needed) >= 0) | |
375 | enqueue_huge_page(page); | |
af767cbd AL |
376 | else { |
377 | /* | |
2668db91 AL |
378 | * The page has a reference count of zero already, so |
379 | * call free_huge_page directly instead of using | |
380 | * put_page. This must be done with hugetlb_lock | |
af767cbd AL |
381 | * unlocked which is safe because free_huge_page takes |
382 | * hugetlb_lock before deciding how to free the page. | |
383 | */ | |
384 | spin_unlock(&hugetlb_lock); | |
2668db91 | 385 | free_huge_page(page); |
af767cbd AL |
386 | spin_lock(&hugetlb_lock); |
387 | } | |
e4e574b7 AL |
388 | } |
389 | ||
390 | return ret; | |
391 | } | |
392 | ||
393 | /* | |
394 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
395 | * allocated to satisfy the reservation must be explicitly freed if they were | |
396 | * never used. | |
397 | */ | |
8cde045c | 398 | static void return_unused_surplus_pages(unsigned long unused_resv_pages) |
e4e574b7 AL |
399 | { |
400 | static int nid = -1; | |
401 | struct page *page; | |
402 | unsigned long nr_pages; | |
403 | ||
ac09b3a1 AL |
404 | /* Uncommit the reservation */ |
405 | resv_huge_pages -= unused_resv_pages; | |
406 | ||
e4e574b7 AL |
407 | nr_pages = min(unused_resv_pages, surplus_huge_pages); |
408 | ||
409 | while (nr_pages) { | |
410 | nid = next_node(nid, node_online_map); | |
411 | if (nid == MAX_NUMNODES) | |
412 | nid = first_node(node_online_map); | |
413 | ||
414 | if (!surplus_huge_pages_node[nid]) | |
415 | continue; | |
416 | ||
417 | if (!list_empty(&hugepage_freelists[nid])) { | |
418 | page = list_entry(hugepage_freelists[nid].next, | |
419 | struct page, lru); | |
420 | list_del(&page->lru); | |
421 | update_and_free_page(page); | |
422 | free_huge_pages--; | |
423 | free_huge_pages_node[nid]--; | |
424 | surplus_huge_pages--; | |
425 | surplus_huge_pages_node[nid]--; | |
426 | nr_pages--; | |
427 | } | |
428 | } | |
429 | } | |
430 | ||
348ea204 AL |
431 | |
432 | static struct page *alloc_huge_page_shared(struct vm_area_struct *vma, | |
433 | unsigned long addr) | |
1da177e4 | 434 | { |
348ea204 | 435 | struct page *page; |
1da177e4 LT |
436 | |
437 | spin_lock(&hugetlb_lock); | |
348e1e04 | 438 | page = dequeue_huge_page_vma(vma, addr); |
1da177e4 | 439 | spin_unlock(&hugetlb_lock); |
90d8b7e6 | 440 | return page ? page : ERR_PTR(-VM_FAULT_OOM); |
348ea204 | 441 | } |
b45b5bd6 | 442 | |
348ea204 AL |
443 | static struct page *alloc_huge_page_private(struct vm_area_struct *vma, |
444 | unsigned long addr) | |
445 | { | |
446 | struct page *page = NULL; | |
7893d1d5 | 447 | |
90d8b7e6 AL |
448 | if (hugetlb_get_quota(vma->vm_file->f_mapping, 1)) |
449 | return ERR_PTR(-VM_FAULT_SIGBUS); | |
450 | ||
348ea204 AL |
451 | spin_lock(&hugetlb_lock); |
452 | if (free_huge_pages > resv_huge_pages) | |
348e1e04 | 453 | page = dequeue_huge_page_vma(vma, addr); |
348ea204 | 454 | spin_unlock(&hugetlb_lock); |
68842c9b | 455 | if (!page) { |
7893d1d5 | 456 | page = alloc_buddy_huge_page(vma, addr); |
68842c9b KC |
457 | if (!page) { |
458 | hugetlb_put_quota(vma->vm_file->f_mapping, 1); | |
459 | return ERR_PTR(-VM_FAULT_OOM); | |
460 | } | |
461 | } | |
462 | return page; | |
348ea204 AL |
463 | } |
464 | ||
465 | static struct page *alloc_huge_page(struct vm_area_struct *vma, | |
466 | unsigned long addr) | |
467 | { | |
468 | struct page *page; | |
2fc39cec AL |
469 | struct address_space *mapping = vma->vm_file->f_mapping; |
470 | ||
348ea204 AL |
471 | if (vma->vm_flags & VM_MAYSHARE) |
472 | page = alloc_huge_page_shared(vma, addr); | |
473 | else | |
474 | page = alloc_huge_page_private(vma, addr); | |
90d8b7e6 AL |
475 | |
476 | if (!IS_ERR(page)) { | |
348ea204 | 477 | set_page_refcounted(page); |
2fc39cec | 478 | set_page_private(page, (unsigned long) mapping); |
90d8b7e6 AL |
479 | } |
480 | return page; | |
b45b5bd6 DG |
481 | } |
482 | ||
1da177e4 LT |
483 | static int __init hugetlb_init(void) |
484 | { | |
485 | unsigned long i; | |
1da177e4 | 486 | |
3c726f8d BH |
487 | if (HPAGE_SHIFT == 0) |
488 | return 0; | |
489 | ||
1da177e4 LT |
490 | for (i = 0; i < MAX_NUMNODES; ++i) |
491 | INIT_LIST_HEAD(&hugepage_freelists[i]); | |
492 | ||
63b4613c NA |
493 | hugetlb_next_nid = first_node(node_online_map); |
494 | ||
1da177e4 | 495 | for (i = 0; i < max_huge_pages; ++i) { |
a482289d | 496 | if (!alloc_fresh_huge_page()) |
1da177e4 | 497 | break; |
1da177e4 LT |
498 | } |
499 | max_huge_pages = free_huge_pages = nr_huge_pages = i; | |
500 | printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages); | |
501 | return 0; | |
502 | } | |
503 | module_init(hugetlb_init); | |
504 | ||
505 | static int __init hugetlb_setup(char *s) | |
506 | { | |
507 | if (sscanf(s, "%lu", &max_huge_pages) <= 0) | |
508 | max_huge_pages = 0; | |
509 | return 1; | |
510 | } | |
511 | __setup("hugepages=", hugetlb_setup); | |
512 | ||
8a630112 KC |
513 | static unsigned int cpuset_mems_nr(unsigned int *array) |
514 | { | |
515 | int node; | |
516 | unsigned int nr = 0; | |
517 | ||
518 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
519 | nr += array[node]; | |
520 | ||
521 | return nr; | |
522 | } | |
523 | ||
1da177e4 | 524 | #ifdef CONFIG_SYSCTL |
1da177e4 LT |
525 | #ifdef CONFIG_HIGHMEM |
526 | static void try_to_free_low(unsigned long count) | |
527 | { | |
4415cc8d CL |
528 | int i; |
529 | ||
1da177e4 LT |
530 | for (i = 0; i < MAX_NUMNODES; ++i) { |
531 | struct page *page, *next; | |
532 | list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) { | |
6b0c880d AL |
533 | if (count >= nr_huge_pages) |
534 | return; | |
1da177e4 LT |
535 | if (PageHighMem(page)) |
536 | continue; | |
537 | list_del(&page->lru); | |
538 | update_and_free_page(page); | |
1da177e4 | 539 | free_huge_pages--; |
4415cc8d | 540 | free_huge_pages_node[page_to_nid(page)]--; |
1da177e4 LT |
541 | } |
542 | } | |
543 | } | |
544 | #else | |
545 | static inline void try_to_free_low(unsigned long count) | |
546 | { | |
547 | } | |
548 | #endif | |
549 | ||
7893d1d5 | 550 | #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages) |
1da177e4 LT |
551 | static unsigned long set_max_huge_pages(unsigned long count) |
552 | { | |
7893d1d5 | 553 | unsigned long min_count, ret; |
1da177e4 | 554 | |
7893d1d5 AL |
555 | /* |
556 | * Increase the pool size | |
557 | * First take pages out of surplus state. Then make up the | |
558 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f NA |
559 | * |
560 | * We might race with alloc_buddy_huge_page() here and be unable | |
561 | * to convert a surplus huge page to a normal huge page. That is | |
562 | * not critical, though, it just means the overall size of the | |
563 | * pool might be one hugepage larger than it needs to be, but | |
564 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 565 | */ |
1da177e4 | 566 | spin_lock(&hugetlb_lock); |
7893d1d5 AL |
567 | while (surplus_huge_pages && count > persistent_huge_pages) { |
568 | if (!adjust_pool_surplus(-1)) | |
569 | break; | |
570 | } | |
571 | ||
572 | while (count > persistent_huge_pages) { | |
573 | int ret; | |
574 | /* | |
575 | * If this allocation races such that we no longer need the | |
576 | * page, free_huge_page will handle it by freeing the page | |
577 | * and reducing the surplus. | |
578 | */ | |
579 | spin_unlock(&hugetlb_lock); | |
580 | ret = alloc_fresh_huge_page(); | |
581 | spin_lock(&hugetlb_lock); | |
582 | if (!ret) | |
583 | goto out; | |
584 | ||
585 | } | |
7893d1d5 AL |
586 | |
587 | /* | |
588 | * Decrease the pool size | |
589 | * First return free pages to the buddy allocator (being careful | |
590 | * to keep enough around to satisfy reservations). Then place | |
591 | * pages into surplus state as needed so the pool will shrink | |
592 | * to the desired size as pages become free. | |
d1c3fb1f NA |
593 | * |
594 | * By placing pages into the surplus state independent of the | |
595 | * overcommit value, we are allowing the surplus pool size to | |
596 | * exceed overcommit. There are few sane options here. Since | |
597 | * alloc_buddy_huge_page() is checking the global counter, | |
598 | * though, we'll note that we're not allowed to exceed surplus | |
599 | * and won't grow the pool anywhere else. Not until one of the | |
600 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 601 | */ |
6b0c880d AL |
602 | min_count = resv_huge_pages + nr_huge_pages - free_huge_pages; |
603 | min_count = max(count, min_count); | |
7893d1d5 AL |
604 | try_to_free_low(min_count); |
605 | while (min_count < persistent_huge_pages) { | |
348e1e04 | 606 | struct page *page = dequeue_huge_page(); |
1da177e4 LT |
607 | if (!page) |
608 | break; | |
609 | update_and_free_page(page); | |
610 | } | |
7893d1d5 AL |
611 | while (count < persistent_huge_pages) { |
612 | if (!adjust_pool_surplus(1)) | |
613 | break; | |
614 | } | |
615 | out: | |
616 | ret = persistent_huge_pages; | |
1da177e4 | 617 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 618 | return ret; |
1da177e4 LT |
619 | } |
620 | ||
621 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, | |
622 | struct file *file, void __user *buffer, | |
623 | size_t *length, loff_t *ppos) | |
624 | { | |
625 | proc_doulongvec_minmax(table, write, file, buffer, length, ppos); | |
626 | max_huge_pages = set_max_huge_pages(max_huge_pages); | |
627 | return 0; | |
628 | } | |
396faf03 MG |
629 | |
630 | int hugetlb_treat_movable_handler(struct ctl_table *table, int write, | |
631 | struct file *file, void __user *buffer, | |
632 | size_t *length, loff_t *ppos) | |
633 | { | |
634 | proc_dointvec(table, write, file, buffer, length, ppos); | |
635 | if (hugepages_treat_as_movable) | |
636 | htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; | |
637 | else | |
638 | htlb_alloc_mask = GFP_HIGHUSER; | |
639 | return 0; | |
640 | } | |
641 | ||
a3d0c6aa NA |
642 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
643 | struct file *file, void __user *buffer, | |
644 | size_t *length, loff_t *ppos) | |
645 | { | |
a3d0c6aa | 646 | proc_doulongvec_minmax(table, write, file, buffer, length, ppos); |
064d9efe NA |
647 | spin_lock(&hugetlb_lock); |
648 | nr_overcommit_huge_pages = sysctl_overcommit_huge_pages; | |
a3d0c6aa NA |
649 | spin_unlock(&hugetlb_lock); |
650 | return 0; | |
651 | } | |
652 | ||
1da177e4 LT |
653 | #endif /* CONFIG_SYSCTL */ |
654 | ||
655 | int hugetlb_report_meminfo(char *buf) | |
656 | { | |
657 | return sprintf(buf, | |
658 | "HugePages_Total: %5lu\n" | |
659 | "HugePages_Free: %5lu\n" | |
a43a8c39 | 660 | "HugePages_Rsvd: %5lu\n" |
7893d1d5 | 661 | "HugePages_Surp: %5lu\n" |
1da177e4 LT |
662 | "Hugepagesize: %5lu kB\n", |
663 | nr_huge_pages, | |
664 | free_huge_pages, | |
a43a8c39 | 665 | resv_huge_pages, |
7893d1d5 | 666 | surplus_huge_pages, |
1da177e4 LT |
667 | HPAGE_SIZE/1024); |
668 | } | |
669 | ||
670 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
671 | { | |
672 | return sprintf(buf, | |
673 | "Node %d HugePages_Total: %5u\n" | |
674 | "Node %d HugePages_Free: %5u\n", | |
675 | nid, nr_huge_pages_node[nid], | |
676 | nid, free_huge_pages_node[nid]); | |
677 | } | |
678 | ||
1da177e4 LT |
679 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
680 | unsigned long hugetlb_total_pages(void) | |
681 | { | |
682 | return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE); | |
683 | } | |
1da177e4 LT |
684 | |
685 | /* | |
686 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
687 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
688 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
689 | * this far. | |
690 | */ | |
d0217ac0 | 691 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
692 | { |
693 | BUG(); | |
d0217ac0 | 694 | return 0; |
1da177e4 LT |
695 | } |
696 | ||
697 | struct vm_operations_struct hugetlb_vm_ops = { | |
d0217ac0 | 698 | .fault = hugetlb_vm_op_fault, |
1da177e4 LT |
699 | }; |
700 | ||
1e8f889b DG |
701 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
702 | int writable) | |
63551ae0 DG |
703 | { |
704 | pte_t entry; | |
705 | ||
1e8f889b | 706 | if (writable) { |
63551ae0 DG |
707 | entry = |
708 | pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); | |
709 | } else { | |
710 | entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot)); | |
711 | } | |
712 | entry = pte_mkyoung(entry); | |
713 | entry = pte_mkhuge(entry); | |
714 | ||
715 | return entry; | |
716 | } | |
717 | ||
1e8f889b DG |
718 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
719 | unsigned long address, pte_t *ptep) | |
720 | { | |
721 | pte_t entry; | |
722 | ||
723 | entry = pte_mkwrite(pte_mkdirty(*ptep)); | |
8dab5241 BH |
724 | if (ptep_set_access_flags(vma, address, ptep, entry, 1)) { |
725 | update_mmu_cache(vma, address, entry); | |
8dab5241 | 726 | } |
1e8f889b DG |
727 | } |
728 | ||
729 | ||
63551ae0 DG |
730 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
731 | struct vm_area_struct *vma) | |
732 | { | |
733 | pte_t *src_pte, *dst_pte, entry; | |
734 | struct page *ptepage; | |
1c59827d | 735 | unsigned long addr; |
1e8f889b DG |
736 | int cow; |
737 | ||
738 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 739 | |
1c59827d | 740 | for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) { |
c74df32c HD |
741 | src_pte = huge_pte_offset(src, addr); |
742 | if (!src_pte) | |
743 | continue; | |
63551ae0 DG |
744 | dst_pte = huge_pte_alloc(dst, addr); |
745 | if (!dst_pte) | |
746 | goto nomem; | |
c5c99429 LW |
747 | |
748 | /* If the pagetables are shared don't copy or take references */ | |
749 | if (dst_pte == src_pte) | |
750 | continue; | |
751 | ||
c74df32c | 752 | spin_lock(&dst->page_table_lock); |
1c59827d | 753 | spin_lock(&src->page_table_lock); |
c74df32c | 754 | if (!pte_none(*src_pte)) { |
1e8f889b DG |
755 | if (cow) |
756 | ptep_set_wrprotect(src, addr, src_pte); | |
1c59827d HD |
757 | entry = *src_pte; |
758 | ptepage = pte_page(entry); | |
759 | get_page(ptepage); | |
1c59827d HD |
760 | set_huge_pte_at(dst, addr, dst_pte, entry); |
761 | } | |
762 | spin_unlock(&src->page_table_lock); | |
c74df32c | 763 | spin_unlock(&dst->page_table_lock); |
63551ae0 DG |
764 | } |
765 | return 0; | |
766 | ||
767 | nomem: | |
768 | return -ENOMEM; | |
769 | } | |
770 | ||
502717f4 KC |
771 | void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
772 | unsigned long end) | |
63551ae0 DG |
773 | { |
774 | struct mm_struct *mm = vma->vm_mm; | |
775 | unsigned long address; | |
c7546f8f | 776 | pte_t *ptep; |
63551ae0 DG |
777 | pte_t pte; |
778 | struct page *page; | |
fe1668ae | 779 | struct page *tmp; |
c0a499c2 KC |
780 | /* |
781 | * A page gathering list, protected by per file i_mmap_lock. The | |
782 | * lock is used to avoid list corruption from multiple unmapping | |
783 | * of the same page since we are using page->lru. | |
784 | */ | |
fe1668ae | 785 | LIST_HEAD(page_list); |
63551ae0 DG |
786 | |
787 | WARN_ON(!is_vm_hugetlb_page(vma)); | |
788 | BUG_ON(start & ~HPAGE_MASK); | |
789 | BUG_ON(end & ~HPAGE_MASK); | |
790 | ||
508034a3 | 791 | spin_lock(&mm->page_table_lock); |
63551ae0 | 792 | for (address = start; address < end; address += HPAGE_SIZE) { |
c7546f8f | 793 | ptep = huge_pte_offset(mm, address); |
4c887265 | 794 | if (!ptep) |
c7546f8f DG |
795 | continue; |
796 | ||
39dde65c KC |
797 | if (huge_pmd_unshare(mm, &address, ptep)) |
798 | continue; | |
799 | ||
c7546f8f | 800 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
63551ae0 DG |
801 | if (pte_none(pte)) |
802 | continue; | |
c7546f8f | 803 | |
63551ae0 | 804 | page = pte_page(pte); |
6649a386 KC |
805 | if (pte_dirty(pte)) |
806 | set_page_dirty(page); | |
fe1668ae | 807 | list_add(&page->lru, &page_list); |
63551ae0 | 808 | } |
1da177e4 | 809 | spin_unlock(&mm->page_table_lock); |
508034a3 | 810 | flush_tlb_range(vma, start, end); |
fe1668ae KC |
811 | list_for_each_entry_safe(page, tmp, &page_list, lru) { |
812 | list_del(&page->lru); | |
813 | put_page(page); | |
814 | } | |
1da177e4 | 815 | } |
63551ae0 | 816 | |
502717f4 KC |
817 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
818 | unsigned long end) | |
819 | { | |
820 | /* | |
821 | * It is undesirable to test vma->vm_file as it should be non-null | |
822 | * for valid hugetlb area. However, vm_file will be NULL in the error | |
823 | * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails, | |
824 | * do_mmap_pgoff() nullifies vma->vm_file before calling this function | |
825 | * to clean up. Since no pte has actually been setup, it is safe to | |
826 | * do nothing in this case. | |
827 | */ | |
828 | if (vma->vm_file) { | |
829 | spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); | |
830 | __unmap_hugepage_range(vma, start, end); | |
831 | spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); | |
832 | } | |
833 | } | |
834 | ||
1e8f889b DG |
835 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
836 | unsigned long address, pte_t *ptep, pte_t pte) | |
837 | { | |
838 | struct page *old_page, *new_page; | |
79ac6ba4 | 839 | int avoidcopy; |
1e8f889b DG |
840 | |
841 | old_page = pte_page(pte); | |
842 | ||
843 | /* If no-one else is actually using this page, avoid the copy | |
844 | * and just make the page writable */ | |
845 | avoidcopy = (page_count(old_page) == 1); | |
846 | if (avoidcopy) { | |
847 | set_huge_ptep_writable(vma, address, ptep); | |
83c54070 | 848 | return 0; |
1e8f889b DG |
849 | } |
850 | ||
851 | page_cache_get(old_page); | |
5da7ca86 | 852 | new_page = alloc_huge_page(vma, address); |
1e8f889b | 853 | |
2fc39cec | 854 | if (IS_ERR(new_page)) { |
1e8f889b | 855 | page_cache_release(old_page); |
2fc39cec | 856 | return -PTR_ERR(new_page); |
1e8f889b DG |
857 | } |
858 | ||
859 | spin_unlock(&mm->page_table_lock); | |
9de455b2 | 860 | copy_huge_page(new_page, old_page, address, vma); |
0ed361de | 861 | __SetPageUptodate(new_page); |
1e8f889b DG |
862 | spin_lock(&mm->page_table_lock); |
863 | ||
864 | ptep = huge_pte_offset(mm, address & HPAGE_MASK); | |
865 | if (likely(pte_same(*ptep, pte))) { | |
866 | /* Break COW */ | |
867 | set_huge_pte_at(mm, address, ptep, | |
868 | make_huge_pte(vma, new_page, 1)); | |
869 | /* Make the old page be freed below */ | |
870 | new_page = old_page; | |
871 | } | |
872 | page_cache_release(new_page); | |
873 | page_cache_release(old_page); | |
83c54070 | 874 | return 0; |
1e8f889b DG |
875 | } |
876 | ||
a1ed3dda | 877 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
1e8f889b | 878 | unsigned long address, pte_t *ptep, int write_access) |
ac9b9c66 HD |
879 | { |
880 | int ret = VM_FAULT_SIGBUS; | |
4c887265 AL |
881 | unsigned long idx; |
882 | unsigned long size; | |
4c887265 AL |
883 | struct page *page; |
884 | struct address_space *mapping; | |
1e8f889b | 885 | pte_t new_pte; |
4c887265 | 886 | |
4c887265 AL |
887 | mapping = vma->vm_file->f_mapping; |
888 | idx = ((address - vma->vm_start) >> HPAGE_SHIFT) | |
889 | + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT)); | |
890 | ||
891 | /* | |
892 | * Use page lock to guard against racing truncation | |
893 | * before we get page_table_lock. | |
894 | */ | |
6bda666a CL |
895 | retry: |
896 | page = find_lock_page(mapping, idx); | |
897 | if (!page) { | |
ebed4bfc HD |
898 | size = i_size_read(mapping->host) >> HPAGE_SHIFT; |
899 | if (idx >= size) | |
900 | goto out; | |
6bda666a | 901 | page = alloc_huge_page(vma, address); |
2fc39cec AL |
902 | if (IS_ERR(page)) { |
903 | ret = -PTR_ERR(page); | |
6bda666a CL |
904 | goto out; |
905 | } | |
79ac6ba4 | 906 | clear_huge_page(page, address); |
0ed361de | 907 | __SetPageUptodate(page); |
ac9b9c66 | 908 | |
6bda666a CL |
909 | if (vma->vm_flags & VM_SHARED) { |
910 | int err; | |
45c682a6 | 911 | struct inode *inode = mapping->host; |
6bda666a CL |
912 | |
913 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
914 | if (err) { | |
915 | put_page(page); | |
6bda666a CL |
916 | if (err == -EEXIST) |
917 | goto retry; | |
918 | goto out; | |
919 | } | |
45c682a6 KC |
920 | |
921 | spin_lock(&inode->i_lock); | |
922 | inode->i_blocks += BLOCKS_PER_HUGEPAGE; | |
923 | spin_unlock(&inode->i_lock); | |
6bda666a CL |
924 | } else |
925 | lock_page(page); | |
926 | } | |
1e8f889b | 927 | |
ac9b9c66 | 928 | spin_lock(&mm->page_table_lock); |
4c887265 AL |
929 | size = i_size_read(mapping->host) >> HPAGE_SHIFT; |
930 | if (idx >= size) | |
931 | goto backout; | |
932 | ||
83c54070 | 933 | ret = 0; |
86e5216f | 934 | if (!pte_none(*ptep)) |
4c887265 AL |
935 | goto backout; |
936 | ||
1e8f889b DG |
937 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
938 | && (vma->vm_flags & VM_SHARED))); | |
939 | set_huge_pte_at(mm, address, ptep, new_pte); | |
940 | ||
941 | if (write_access && !(vma->vm_flags & VM_SHARED)) { | |
942 | /* Optimization, do the COW without a second fault */ | |
943 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte); | |
944 | } | |
945 | ||
ac9b9c66 | 946 | spin_unlock(&mm->page_table_lock); |
4c887265 AL |
947 | unlock_page(page); |
948 | out: | |
ac9b9c66 | 949 | return ret; |
4c887265 AL |
950 | |
951 | backout: | |
952 | spin_unlock(&mm->page_table_lock); | |
4c887265 AL |
953 | unlock_page(page); |
954 | put_page(page); | |
955 | goto out; | |
ac9b9c66 HD |
956 | } |
957 | ||
86e5216f AL |
958 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
959 | unsigned long address, int write_access) | |
960 | { | |
961 | pte_t *ptep; | |
962 | pte_t entry; | |
1e8f889b | 963 | int ret; |
3935baa9 | 964 | static DEFINE_MUTEX(hugetlb_instantiation_mutex); |
86e5216f AL |
965 | |
966 | ptep = huge_pte_alloc(mm, address); | |
967 | if (!ptep) | |
968 | return VM_FAULT_OOM; | |
969 | ||
3935baa9 DG |
970 | /* |
971 | * Serialize hugepage allocation and instantiation, so that we don't | |
972 | * get spurious allocation failures if two CPUs race to instantiate | |
973 | * the same page in the page cache. | |
974 | */ | |
975 | mutex_lock(&hugetlb_instantiation_mutex); | |
86e5216f | 976 | entry = *ptep; |
3935baa9 DG |
977 | if (pte_none(entry)) { |
978 | ret = hugetlb_no_page(mm, vma, address, ptep, write_access); | |
979 | mutex_unlock(&hugetlb_instantiation_mutex); | |
980 | return ret; | |
981 | } | |
86e5216f | 982 | |
83c54070 | 983 | ret = 0; |
1e8f889b DG |
984 | |
985 | spin_lock(&mm->page_table_lock); | |
986 | /* Check for a racing update before calling hugetlb_cow */ | |
987 | if (likely(pte_same(entry, *ptep))) | |
988 | if (write_access && !pte_write(entry)) | |
989 | ret = hugetlb_cow(mm, vma, address, ptep, entry); | |
990 | spin_unlock(&mm->page_table_lock); | |
3935baa9 | 991 | mutex_unlock(&hugetlb_instantiation_mutex); |
1e8f889b DG |
992 | |
993 | return ret; | |
86e5216f AL |
994 | } |
995 | ||
63551ae0 DG |
996 | int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
997 | struct page **pages, struct vm_area_struct **vmas, | |
5b23dbe8 AL |
998 | unsigned long *position, int *length, int i, |
999 | int write) | |
63551ae0 | 1000 | { |
d5d4b0aa KC |
1001 | unsigned long pfn_offset; |
1002 | unsigned long vaddr = *position; | |
63551ae0 DG |
1003 | int remainder = *length; |
1004 | ||
1c59827d | 1005 | spin_lock(&mm->page_table_lock); |
63551ae0 | 1006 | while (vaddr < vma->vm_end && remainder) { |
4c887265 AL |
1007 | pte_t *pte; |
1008 | struct page *page; | |
63551ae0 | 1009 | |
4c887265 AL |
1010 | /* |
1011 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
1012 | * each hugepage. We have to make * sure we get the | |
1013 | * first, for the page indexing below to work. | |
1014 | */ | |
1015 | pte = huge_pte_offset(mm, vaddr & HPAGE_MASK); | |
63551ae0 | 1016 | |
72fad713 | 1017 | if (!pte || pte_none(*pte) || (write && !pte_write(*pte))) { |
4c887265 | 1018 | int ret; |
63551ae0 | 1019 | |
4c887265 | 1020 | spin_unlock(&mm->page_table_lock); |
5b23dbe8 | 1021 | ret = hugetlb_fault(mm, vma, vaddr, write); |
4c887265 | 1022 | spin_lock(&mm->page_table_lock); |
a89182c7 | 1023 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 1024 | continue; |
63551ae0 | 1025 | |
4c887265 AL |
1026 | remainder = 0; |
1027 | if (!i) | |
1028 | i = -EFAULT; | |
1029 | break; | |
1030 | } | |
1031 | ||
d5d4b0aa KC |
1032 | pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT; |
1033 | page = pte_page(*pte); | |
1034 | same_page: | |
d6692183 KC |
1035 | if (pages) { |
1036 | get_page(page); | |
d5d4b0aa | 1037 | pages[i] = page + pfn_offset; |
d6692183 | 1038 | } |
63551ae0 DG |
1039 | |
1040 | if (vmas) | |
1041 | vmas[i] = vma; | |
1042 | ||
1043 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 1044 | ++pfn_offset; |
63551ae0 DG |
1045 | --remainder; |
1046 | ++i; | |
d5d4b0aa KC |
1047 | if (vaddr < vma->vm_end && remainder && |
1048 | pfn_offset < HPAGE_SIZE/PAGE_SIZE) { | |
1049 | /* | |
1050 | * We use pfn_offset to avoid touching the pageframes | |
1051 | * of this compound page. | |
1052 | */ | |
1053 | goto same_page; | |
1054 | } | |
63551ae0 | 1055 | } |
1c59827d | 1056 | spin_unlock(&mm->page_table_lock); |
63551ae0 DG |
1057 | *length = remainder; |
1058 | *position = vaddr; | |
1059 | ||
1060 | return i; | |
1061 | } | |
8f860591 ZY |
1062 | |
1063 | void hugetlb_change_protection(struct vm_area_struct *vma, | |
1064 | unsigned long address, unsigned long end, pgprot_t newprot) | |
1065 | { | |
1066 | struct mm_struct *mm = vma->vm_mm; | |
1067 | unsigned long start = address; | |
1068 | pte_t *ptep; | |
1069 | pte_t pte; | |
1070 | ||
1071 | BUG_ON(address >= end); | |
1072 | flush_cache_range(vma, address, end); | |
1073 | ||
39dde65c | 1074 | spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); |
8f860591 ZY |
1075 | spin_lock(&mm->page_table_lock); |
1076 | for (; address < end; address += HPAGE_SIZE) { | |
1077 | ptep = huge_pte_offset(mm, address); | |
1078 | if (!ptep) | |
1079 | continue; | |
39dde65c KC |
1080 | if (huge_pmd_unshare(mm, &address, ptep)) |
1081 | continue; | |
8f860591 ZY |
1082 | if (!pte_none(*ptep)) { |
1083 | pte = huge_ptep_get_and_clear(mm, address, ptep); | |
1084 | pte = pte_mkhuge(pte_modify(pte, newprot)); | |
1085 | set_huge_pte_at(mm, address, ptep, pte); | |
8f860591 ZY |
1086 | } |
1087 | } | |
1088 | spin_unlock(&mm->page_table_lock); | |
39dde65c | 1089 | spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); |
8f860591 ZY |
1090 | |
1091 | flush_tlb_range(vma, start, end); | |
1092 | } | |
1093 | ||
a43a8c39 KC |
1094 | struct file_region { |
1095 | struct list_head link; | |
1096 | long from; | |
1097 | long to; | |
1098 | }; | |
1099 | ||
1100 | static long region_add(struct list_head *head, long f, long t) | |
1101 | { | |
1102 | struct file_region *rg, *nrg, *trg; | |
1103 | ||
1104 | /* Locate the region we are either in or before. */ | |
1105 | list_for_each_entry(rg, head, link) | |
1106 | if (f <= rg->to) | |
1107 | break; | |
1108 | ||
1109 | /* Round our left edge to the current segment if it encloses us. */ | |
1110 | if (f > rg->from) | |
1111 | f = rg->from; | |
1112 | ||
1113 | /* Check for and consume any regions we now overlap with. */ | |
1114 | nrg = rg; | |
1115 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
1116 | if (&rg->link == head) | |
1117 | break; | |
1118 | if (rg->from > t) | |
1119 | break; | |
1120 | ||
1121 | /* If this area reaches higher then extend our area to | |
1122 | * include it completely. If this is not the first area | |
1123 | * which we intend to reuse, free it. */ | |
1124 | if (rg->to > t) | |
1125 | t = rg->to; | |
1126 | if (rg != nrg) { | |
1127 | list_del(&rg->link); | |
1128 | kfree(rg); | |
1129 | } | |
1130 | } | |
1131 | nrg->from = f; | |
1132 | nrg->to = t; | |
1133 | return 0; | |
1134 | } | |
1135 | ||
1136 | static long region_chg(struct list_head *head, long f, long t) | |
1137 | { | |
1138 | struct file_region *rg, *nrg; | |
1139 | long chg = 0; | |
1140 | ||
1141 | /* Locate the region we are before or in. */ | |
1142 | list_for_each_entry(rg, head, link) | |
1143 | if (f <= rg->to) | |
1144 | break; | |
1145 | ||
1146 | /* If we are below the current region then a new region is required. | |
1147 | * Subtle, allocate a new region at the position but make it zero | |
183ff22b | 1148 | * size such that we can guarantee to record the reservation. */ |
a43a8c39 KC |
1149 | if (&rg->link == head || t < rg->from) { |
1150 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
c80544dc | 1151 | if (!nrg) |
a43a8c39 KC |
1152 | return -ENOMEM; |
1153 | nrg->from = f; | |
1154 | nrg->to = f; | |
1155 | INIT_LIST_HEAD(&nrg->link); | |
1156 | list_add(&nrg->link, rg->link.prev); | |
1157 | ||
1158 | return t - f; | |
1159 | } | |
1160 | ||
1161 | /* Round our left edge to the current segment if it encloses us. */ | |
1162 | if (f > rg->from) | |
1163 | f = rg->from; | |
1164 | chg = t - f; | |
1165 | ||
1166 | /* Check for and consume any regions we now overlap with. */ | |
1167 | list_for_each_entry(rg, rg->link.prev, link) { | |
1168 | if (&rg->link == head) | |
1169 | break; | |
1170 | if (rg->from > t) | |
1171 | return chg; | |
1172 | ||
1173 | /* We overlap with this area, if it extends futher than | |
1174 | * us then we must extend ourselves. Account for its | |
1175 | * existing reservation. */ | |
1176 | if (rg->to > t) { | |
1177 | chg += rg->to - t; | |
1178 | t = rg->to; | |
1179 | } | |
1180 | chg -= rg->to - rg->from; | |
1181 | } | |
1182 | return chg; | |
1183 | } | |
1184 | ||
1185 | static long region_truncate(struct list_head *head, long end) | |
1186 | { | |
1187 | struct file_region *rg, *trg; | |
1188 | long chg = 0; | |
1189 | ||
1190 | /* Locate the region we are either in or before. */ | |
1191 | list_for_each_entry(rg, head, link) | |
1192 | if (end <= rg->to) | |
1193 | break; | |
1194 | if (&rg->link == head) | |
1195 | return 0; | |
1196 | ||
1197 | /* If we are in the middle of a region then adjust it. */ | |
1198 | if (end > rg->from) { | |
1199 | chg = rg->to - end; | |
1200 | rg->to = end; | |
1201 | rg = list_entry(rg->link.next, typeof(*rg), link); | |
1202 | } | |
1203 | ||
1204 | /* Drop any remaining regions. */ | |
1205 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
1206 | if (&rg->link == head) | |
1207 | break; | |
1208 | chg += rg->to - rg->from; | |
1209 | list_del(&rg->link); | |
1210 | kfree(rg); | |
1211 | } | |
1212 | return chg; | |
1213 | } | |
1214 | ||
1215 | static int hugetlb_acct_memory(long delta) | |
1216 | { | |
1217 | int ret = -ENOMEM; | |
1218 | ||
1219 | spin_lock(&hugetlb_lock); | |
8a630112 KC |
1220 | /* |
1221 | * When cpuset is configured, it breaks the strict hugetlb page | |
1222 | * reservation as the accounting is done on a global variable. Such | |
1223 | * reservation is completely rubbish in the presence of cpuset because | |
1224 | * the reservation is not checked against page availability for the | |
1225 | * current cpuset. Application can still potentially OOM'ed by kernel | |
1226 | * with lack of free htlb page in cpuset that the task is in. | |
1227 | * Attempt to enforce strict accounting with cpuset is almost | |
1228 | * impossible (or too ugly) because cpuset is too fluid that | |
1229 | * task or memory node can be dynamically moved between cpusets. | |
1230 | * | |
1231 | * The change of semantics for shared hugetlb mapping with cpuset is | |
1232 | * undesirable. However, in order to preserve some of the semantics, | |
1233 | * we fall back to check against current free page availability as | |
1234 | * a best attempt and hopefully to minimize the impact of changing | |
1235 | * semantics that cpuset has. | |
1236 | */ | |
e4e574b7 AL |
1237 | if (delta > 0) { |
1238 | if (gather_surplus_pages(delta) < 0) | |
1239 | goto out; | |
1240 | ||
ac09b3a1 AL |
1241 | if (delta > cpuset_mems_nr(free_huge_pages_node)) { |
1242 | return_unused_surplus_pages(delta); | |
e4e574b7 | 1243 | goto out; |
ac09b3a1 | 1244 | } |
e4e574b7 AL |
1245 | } |
1246 | ||
1247 | ret = 0; | |
e4e574b7 AL |
1248 | if (delta < 0) |
1249 | return_unused_surplus_pages((unsigned long) -delta); | |
1250 | ||
1251 | out: | |
1252 | spin_unlock(&hugetlb_lock); | |
1253 | return ret; | |
1254 | } | |
1255 | ||
1256 | int hugetlb_reserve_pages(struct inode *inode, long from, long to) | |
1257 | { | |
1258 | long ret, chg; | |
1259 | ||
1260 | chg = region_chg(&inode->i_mapping->private_list, from, to); | |
1261 | if (chg < 0) | |
1262 | return chg; | |
8a630112 | 1263 | |
90d8b7e6 AL |
1264 | if (hugetlb_get_quota(inode->i_mapping, chg)) |
1265 | return -ENOSPC; | |
a43a8c39 | 1266 | ret = hugetlb_acct_memory(chg); |
68842c9b KC |
1267 | if (ret < 0) { |
1268 | hugetlb_put_quota(inode->i_mapping, chg); | |
a43a8c39 | 1269 | return ret; |
68842c9b | 1270 | } |
a43a8c39 KC |
1271 | region_add(&inode->i_mapping->private_list, from, to); |
1272 | return 0; | |
1273 | } | |
1274 | ||
1275 | void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) | |
1276 | { | |
1277 | long chg = region_truncate(&inode->i_mapping->private_list, offset); | |
45c682a6 KC |
1278 | |
1279 | spin_lock(&inode->i_lock); | |
1280 | inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed; | |
1281 | spin_unlock(&inode->i_lock); | |
1282 | ||
90d8b7e6 AL |
1283 | hugetlb_put_quota(inode->i_mapping, (chg - freed)); |
1284 | hugetlb_acct_memory(-(chg - freed)); | |
a43a8c39 | 1285 | } |