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