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 | |
96822904 AW |
43 | /* |
44 | * Region tracking -- allows tracking of reservations and instantiated pages | |
45 | * across the pages in a mapping. | |
84afd99b AW |
46 | * |
47 | * The region data structures are protected by a combination of the mmap_sem | |
48 | * and the hugetlb_instantion_mutex. To access or modify a region the caller | |
49 | * must either hold the mmap_sem for write, or the mmap_sem for read and | |
50 | * the hugetlb_instantiation mutex: | |
51 | * | |
52 | * down_write(&mm->mmap_sem); | |
53 | * or | |
54 | * down_read(&mm->mmap_sem); | |
55 | * mutex_lock(&hugetlb_instantiation_mutex); | |
96822904 AW |
56 | */ |
57 | struct file_region { | |
58 | struct list_head link; | |
59 | long from; | |
60 | long to; | |
61 | }; | |
62 | ||
63 | static long region_add(struct list_head *head, long f, long t) | |
64 | { | |
65 | struct file_region *rg, *nrg, *trg; | |
66 | ||
67 | /* Locate the region we are either in or before. */ | |
68 | list_for_each_entry(rg, head, link) | |
69 | if (f <= rg->to) | |
70 | break; | |
71 | ||
72 | /* Round our left edge to the current segment if it encloses us. */ | |
73 | if (f > rg->from) | |
74 | f = rg->from; | |
75 | ||
76 | /* Check for and consume any regions we now overlap with. */ | |
77 | nrg = rg; | |
78 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
79 | if (&rg->link == head) | |
80 | break; | |
81 | if (rg->from > t) | |
82 | break; | |
83 | ||
84 | /* If this area reaches higher then extend our area to | |
85 | * include it completely. If this is not the first area | |
86 | * which we intend to reuse, free it. */ | |
87 | if (rg->to > t) | |
88 | t = rg->to; | |
89 | if (rg != nrg) { | |
90 | list_del(&rg->link); | |
91 | kfree(rg); | |
92 | } | |
93 | } | |
94 | nrg->from = f; | |
95 | nrg->to = t; | |
96 | return 0; | |
97 | } | |
98 | ||
99 | static long region_chg(struct list_head *head, long f, long t) | |
100 | { | |
101 | struct file_region *rg, *nrg; | |
102 | long chg = 0; | |
103 | ||
104 | /* Locate the region we are before or in. */ | |
105 | list_for_each_entry(rg, head, link) | |
106 | if (f <= rg->to) | |
107 | break; | |
108 | ||
109 | /* If we are below the current region then a new region is required. | |
110 | * Subtle, allocate a new region at the position but make it zero | |
111 | * size such that we can guarantee to record the reservation. */ | |
112 | if (&rg->link == head || t < rg->from) { | |
113 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
114 | if (!nrg) | |
115 | return -ENOMEM; | |
116 | nrg->from = f; | |
117 | nrg->to = f; | |
118 | INIT_LIST_HEAD(&nrg->link); | |
119 | list_add(&nrg->link, rg->link.prev); | |
120 | ||
121 | return t - f; | |
122 | } | |
123 | ||
124 | /* Round our left edge to the current segment if it encloses us. */ | |
125 | if (f > rg->from) | |
126 | f = rg->from; | |
127 | chg = t - f; | |
128 | ||
129 | /* Check for and consume any regions we now overlap with. */ | |
130 | list_for_each_entry(rg, rg->link.prev, link) { | |
131 | if (&rg->link == head) | |
132 | break; | |
133 | if (rg->from > t) | |
134 | return chg; | |
135 | ||
136 | /* We overlap with this area, if it extends futher than | |
137 | * us then we must extend ourselves. Account for its | |
138 | * existing reservation. */ | |
139 | if (rg->to > t) { | |
140 | chg += rg->to - t; | |
141 | t = rg->to; | |
142 | } | |
143 | chg -= rg->to - rg->from; | |
144 | } | |
145 | return chg; | |
146 | } | |
147 | ||
148 | static long region_truncate(struct list_head *head, long end) | |
149 | { | |
150 | struct file_region *rg, *trg; | |
151 | long chg = 0; | |
152 | ||
153 | /* Locate the region we are either in or before. */ | |
154 | list_for_each_entry(rg, head, link) | |
155 | if (end <= rg->to) | |
156 | break; | |
157 | if (&rg->link == head) | |
158 | return 0; | |
159 | ||
160 | /* If we are in the middle of a region then adjust it. */ | |
161 | if (end > rg->from) { | |
162 | chg = rg->to - end; | |
163 | rg->to = end; | |
164 | rg = list_entry(rg->link.next, typeof(*rg), link); | |
165 | } | |
166 | ||
167 | /* Drop any remaining regions. */ | |
168 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
169 | if (&rg->link == head) | |
170 | break; | |
171 | chg += rg->to - rg->from; | |
172 | list_del(&rg->link); | |
173 | kfree(rg); | |
174 | } | |
175 | return chg; | |
176 | } | |
177 | ||
84afd99b AW |
178 | static long region_count(struct list_head *head, long f, long t) |
179 | { | |
180 | struct file_region *rg; | |
181 | long chg = 0; | |
182 | ||
183 | /* Locate each segment we overlap with, and count that overlap. */ | |
184 | list_for_each_entry(rg, head, link) { | |
185 | int seg_from; | |
186 | int seg_to; | |
187 | ||
188 | if (rg->to <= f) | |
189 | continue; | |
190 | if (rg->from >= t) | |
191 | break; | |
192 | ||
193 | seg_from = max(rg->from, f); | |
194 | seg_to = min(rg->to, t); | |
195 | ||
196 | chg += seg_to - seg_from; | |
197 | } | |
198 | ||
199 | return chg; | |
200 | } | |
201 | ||
e7c4b0bf AW |
202 | /* |
203 | * Convert the address within this vma to the page offset within | |
204 | * the mapping, in pagecache page units; huge pages here. | |
205 | */ | |
a858f7b2 | 206 | static pgoff_t vma_hugecache_offset(struct vm_area_struct *vma, |
e7c4b0bf AW |
207 | unsigned long address) |
208 | { | |
209 | return ((address - vma->vm_start) >> HPAGE_SHIFT) + | |
210 | (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT)); | |
211 | } | |
212 | ||
84afd99b AW |
213 | /* |
214 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
215 | * bits of the reservation map pointer, which are always clear due to | |
216 | * alignment. | |
217 | */ | |
218 | #define HPAGE_RESV_OWNER (1UL << 0) | |
219 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 220 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 221 | |
a1e78772 MG |
222 | /* |
223 | * These helpers are used to track how many pages are reserved for | |
224 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
225 | * is guaranteed to have their future faults succeed. | |
226 | * | |
227 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
228 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
229 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
230 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
231 | * |
232 | * The private mapping reservation is represented in a subtly different | |
233 | * manner to a shared mapping. A shared mapping has a region map associated | |
234 | * with the underlying file, this region map represents the backing file | |
235 | * pages which have ever had a reservation assigned which this persists even | |
236 | * after the page is instantiated. A private mapping has a region map | |
237 | * associated with the original mmap which is attached to all VMAs which | |
238 | * reference it, this region map represents those offsets which have consumed | |
239 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 240 | */ |
e7c4b0bf AW |
241 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
242 | { | |
243 | return (unsigned long)vma->vm_private_data; | |
244 | } | |
245 | ||
246 | static void set_vma_private_data(struct vm_area_struct *vma, | |
247 | unsigned long value) | |
248 | { | |
249 | vma->vm_private_data = (void *)value; | |
250 | } | |
251 | ||
84afd99b AW |
252 | struct resv_map { |
253 | struct kref refs; | |
254 | struct list_head regions; | |
255 | }; | |
256 | ||
257 | struct resv_map *resv_map_alloc(void) | |
258 | { | |
259 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
260 | if (!resv_map) | |
261 | return NULL; | |
262 | ||
263 | kref_init(&resv_map->refs); | |
264 | INIT_LIST_HEAD(&resv_map->regions); | |
265 | ||
266 | return resv_map; | |
267 | } | |
268 | ||
269 | void resv_map_release(struct kref *ref) | |
270 | { | |
271 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
272 | ||
273 | /* Clear out any active regions before we release the map. */ | |
274 | region_truncate(&resv_map->regions, 0); | |
275 | kfree(resv_map); | |
276 | } | |
277 | ||
278 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) | |
a1e78772 MG |
279 | { |
280 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
281 | if (!(vma->vm_flags & VM_SHARED)) | |
84afd99b AW |
282 | return (struct resv_map *)(get_vma_private_data(vma) & |
283 | ~HPAGE_RESV_MASK); | |
a1e78772 MG |
284 | return 0; |
285 | } | |
286 | ||
84afd99b | 287 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 MG |
288 | { |
289 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
290 | VM_BUG_ON(vma->vm_flags & VM_SHARED); | |
291 | ||
84afd99b AW |
292 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
293 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
294 | } |
295 | ||
296 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
297 | { | |
04f2cbe3 | 298 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); |
e7c4b0bf AW |
299 | VM_BUG_ON(vma->vm_flags & VM_SHARED); |
300 | ||
301 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
302 | } |
303 | ||
304 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
305 | { | |
306 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
e7c4b0bf AW |
307 | |
308 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
309 | } |
310 | ||
311 | /* Decrement the reserved pages in the hugepage pool by one */ | |
312 | static void decrement_hugepage_resv_vma(struct vm_area_struct *vma) | |
313 | { | |
c37f9fb1 AW |
314 | if (vma->vm_flags & VM_NORESERVE) |
315 | return; | |
316 | ||
a1e78772 MG |
317 | if (vma->vm_flags & VM_SHARED) { |
318 | /* Shared mappings always use reserves */ | |
319 | resv_huge_pages--; | |
84afd99b | 320 | } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
a1e78772 MG |
321 | /* |
322 | * Only the process that called mmap() has reserves for | |
323 | * private mappings. | |
324 | */ | |
84afd99b | 325 | resv_huge_pages--; |
a1e78772 MG |
326 | } |
327 | } | |
328 | ||
04f2cbe3 | 329 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
330 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
331 | { | |
332 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
333 | if (!(vma->vm_flags & VM_SHARED)) | |
334 | vma->vm_private_data = (void *)0; | |
335 | } | |
336 | ||
337 | /* Returns true if the VMA has associated reserve pages */ | |
338 | static int vma_has_private_reserves(struct vm_area_struct *vma) | |
339 | { | |
340 | if (vma->vm_flags & VM_SHARED) | |
341 | return 0; | |
84afd99b | 342 | if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
a1e78772 MG |
343 | return 0; |
344 | return 1; | |
345 | } | |
346 | ||
79ac6ba4 DG |
347 | static void clear_huge_page(struct page *page, unsigned long addr) |
348 | { | |
349 | int i; | |
350 | ||
351 | might_sleep(); | |
352 | for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) { | |
353 | cond_resched(); | |
281e0e3b | 354 | clear_user_highpage(page + i, addr + i * PAGE_SIZE); |
79ac6ba4 DG |
355 | } |
356 | } | |
357 | ||
358 | static void copy_huge_page(struct page *dst, struct page *src, | |
9de455b2 | 359 | unsigned long addr, struct vm_area_struct *vma) |
79ac6ba4 DG |
360 | { |
361 | int i; | |
362 | ||
363 | might_sleep(); | |
364 | for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) { | |
365 | cond_resched(); | |
9de455b2 | 366 | copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma); |
79ac6ba4 DG |
367 | } |
368 | } | |
369 | ||
1da177e4 LT |
370 | static void enqueue_huge_page(struct page *page) |
371 | { | |
372 | int nid = page_to_nid(page); | |
373 | list_add(&page->lru, &hugepage_freelists[nid]); | |
374 | free_huge_pages++; | |
375 | free_huge_pages_node[nid]++; | |
376 | } | |
377 | ||
348e1e04 NA |
378 | static struct page *dequeue_huge_page(void) |
379 | { | |
380 | int nid; | |
381 | struct page *page = NULL; | |
382 | ||
383 | for (nid = 0; nid < MAX_NUMNODES; ++nid) { | |
384 | if (!list_empty(&hugepage_freelists[nid])) { | |
385 | page = list_entry(hugepage_freelists[nid].next, | |
386 | struct page, lru); | |
387 | list_del(&page->lru); | |
388 | free_huge_pages--; | |
389 | free_huge_pages_node[nid]--; | |
390 | break; | |
391 | } | |
392 | } | |
393 | return page; | |
394 | } | |
395 | ||
396 | static struct page *dequeue_huge_page_vma(struct vm_area_struct *vma, | |
04f2cbe3 | 397 | unsigned long address, int avoid_reserve) |
1da177e4 | 398 | { |
31a5c6e4 | 399 | int nid; |
1da177e4 | 400 | struct page *page = NULL; |
480eccf9 | 401 | struct mempolicy *mpol; |
19770b32 | 402 | nodemask_t *nodemask; |
396faf03 | 403 | struct zonelist *zonelist = huge_zonelist(vma, address, |
19770b32 | 404 | htlb_alloc_mask, &mpol, &nodemask); |
dd1a239f MG |
405 | struct zone *zone; |
406 | struct zoneref *z; | |
1da177e4 | 407 | |
a1e78772 MG |
408 | /* |
409 | * A child process with MAP_PRIVATE mappings created by their parent | |
410 | * have no page reserves. This check ensures that reservations are | |
411 | * not "stolen". The child may still get SIGKILLed | |
412 | */ | |
413 | if (!vma_has_private_reserves(vma) && | |
414 | free_huge_pages - resv_huge_pages == 0) | |
415 | return NULL; | |
416 | ||
04f2cbe3 MG |
417 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
418 | if (avoid_reserve && free_huge_pages - resv_huge_pages == 0) | |
419 | return NULL; | |
420 | ||
19770b32 MG |
421 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
422 | MAX_NR_ZONES - 1, nodemask) { | |
54a6eb5c MG |
423 | nid = zone_to_nid(zone); |
424 | if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) && | |
3abf7afd AM |
425 | !list_empty(&hugepage_freelists[nid])) { |
426 | page = list_entry(hugepage_freelists[nid].next, | |
427 | struct page, lru); | |
428 | list_del(&page->lru); | |
429 | free_huge_pages--; | |
430 | free_huge_pages_node[nid]--; | |
04f2cbe3 MG |
431 | |
432 | if (!avoid_reserve) | |
433 | decrement_hugepage_resv_vma(vma); | |
a1e78772 | 434 | |
5ab3ee7b | 435 | break; |
3abf7afd | 436 | } |
1da177e4 | 437 | } |
52cd3b07 | 438 | mpol_cond_put(mpol); |
1da177e4 LT |
439 | return page; |
440 | } | |
441 | ||
6af2acb6 AL |
442 | static void update_and_free_page(struct page *page) |
443 | { | |
444 | int i; | |
445 | nr_huge_pages--; | |
446 | nr_huge_pages_node[page_to_nid(page)]--; | |
447 | for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) { | |
448 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced | | |
449 | 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved | | |
450 | 1 << PG_private | 1<< PG_writeback); | |
451 | } | |
452 | set_compound_page_dtor(page, NULL); | |
453 | set_page_refcounted(page); | |
7f2e9525 | 454 | arch_release_hugepage(page); |
6af2acb6 AL |
455 | __free_pages(page, HUGETLB_PAGE_ORDER); |
456 | } | |
457 | ||
27a85ef1 DG |
458 | static void free_huge_page(struct page *page) |
459 | { | |
7893d1d5 | 460 | int nid = page_to_nid(page); |
c79fb75e | 461 | struct address_space *mapping; |
27a85ef1 | 462 | |
c79fb75e | 463 | mapping = (struct address_space *) page_private(page); |
e5df70ab | 464 | set_page_private(page, 0); |
7893d1d5 | 465 | BUG_ON(page_count(page)); |
27a85ef1 DG |
466 | INIT_LIST_HEAD(&page->lru); |
467 | ||
468 | spin_lock(&hugetlb_lock); | |
7893d1d5 AL |
469 | if (surplus_huge_pages_node[nid]) { |
470 | update_and_free_page(page); | |
471 | surplus_huge_pages--; | |
472 | surplus_huge_pages_node[nid]--; | |
473 | } else { | |
474 | enqueue_huge_page(page); | |
475 | } | |
27a85ef1 | 476 | spin_unlock(&hugetlb_lock); |
c79fb75e | 477 | if (mapping) |
9a119c05 | 478 | hugetlb_put_quota(mapping, 1); |
27a85ef1 DG |
479 | } |
480 | ||
7893d1d5 AL |
481 | /* |
482 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
483 | * balanced by operating on them in a round-robin fashion. | |
484 | * Returns 1 if an adjustment was made. | |
485 | */ | |
486 | static int adjust_pool_surplus(int delta) | |
487 | { | |
488 | static int prev_nid; | |
489 | int nid = prev_nid; | |
490 | int ret = 0; | |
491 | ||
492 | VM_BUG_ON(delta != -1 && delta != 1); | |
493 | do { | |
494 | nid = next_node(nid, node_online_map); | |
495 | if (nid == MAX_NUMNODES) | |
496 | nid = first_node(node_online_map); | |
497 | ||
498 | /* To shrink on this node, there must be a surplus page */ | |
499 | if (delta < 0 && !surplus_huge_pages_node[nid]) | |
500 | continue; | |
501 | /* Surplus cannot exceed the total number of pages */ | |
502 | if (delta > 0 && surplus_huge_pages_node[nid] >= | |
503 | nr_huge_pages_node[nid]) | |
504 | continue; | |
505 | ||
506 | surplus_huge_pages += delta; | |
507 | surplus_huge_pages_node[nid] += delta; | |
508 | ret = 1; | |
509 | break; | |
510 | } while (nid != prev_nid); | |
511 | ||
512 | prev_nid = nid; | |
513 | return ret; | |
514 | } | |
515 | ||
63b4613c | 516 | static struct page *alloc_fresh_huge_page_node(int nid) |
1da177e4 | 517 | { |
1da177e4 | 518 | struct page *page; |
f96efd58 | 519 | |
63b4613c | 520 | page = alloc_pages_node(nid, |
551883ae NA |
521 | htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| |
522 | __GFP_REPEAT|__GFP_NOWARN, | |
63b4613c | 523 | HUGETLB_PAGE_ORDER); |
1da177e4 | 524 | if (page) { |
7f2e9525 GS |
525 | if (arch_prepare_hugepage(page)) { |
526 | __free_pages(page, HUGETLB_PAGE_ORDER); | |
7b8ee84d | 527 | return NULL; |
7f2e9525 | 528 | } |
33f2ef89 | 529 | set_compound_page_dtor(page, free_huge_page); |
0bd0f9fb | 530 | spin_lock(&hugetlb_lock); |
1da177e4 | 531 | nr_huge_pages++; |
63b4613c | 532 | nr_huge_pages_node[nid]++; |
0bd0f9fb | 533 | spin_unlock(&hugetlb_lock); |
a482289d | 534 | put_page(page); /* free it into the hugepage allocator */ |
1da177e4 | 535 | } |
63b4613c NA |
536 | |
537 | return page; | |
538 | } | |
539 | ||
540 | static int alloc_fresh_huge_page(void) | |
541 | { | |
542 | struct page *page; | |
543 | int start_nid; | |
544 | int next_nid; | |
545 | int ret = 0; | |
546 | ||
547 | start_nid = hugetlb_next_nid; | |
548 | ||
549 | do { | |
550 | page = alloc_fresh_huge_page_node(hugetlb_next_nid); | |
551 | if (page) | |
552 | ret = 1; | |
553 | /* | |
554 | * Use a helper variable to find the next node and then | |
555 | * copy it back to hugetlb_next_nid afterwards: | |
556 | * otherwise there's a window in which a racer might | |
557 | * pass invalid nid MAX_NUMNODES to alloc_pages_node. | |
558 | * But we don't need to use a spin_lock here: it really | |
559 | * doesn't matter if occasionally a racer chooses the | |
560 | * same nid as we do. Move nid forward in the mask even | |
561 | * if we just successfully allocated a hugepage so that | |
562 | * the next caller gets hugepages on the next node. | |
563 | */ | |
564 | next_nid = next_node(hugetlb_next_nid, node_online_map); | |
565 | if (next_nid == MAX_NUMNODES) | |
566 | next_nid = first_node(node_online_map); | |
567 | hugetlb_next_nid = next_nid; | |
568 | } while (!page && hugetlb_next_nid != start_nid); | |
569 | ||
3b116300 AL |
570 | if (ret) |
571 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
572 | else | |
573 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
574 | ||
63b4613c | 575 | return ret; |
1da177e4 LT |
576 | } |
577 | ||
7893d1d5 AL |
578 | static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma, |
579 | unsigned long address) | |
580 | { | |
581 | struct page *page; | |
d1c3fb1f | 582 | unsigned int nid; |
7893d1d5 | 583 | |
d1c3fb1f NA |
584 | /* |
585 | * Assume we will successfully allocate the surplus page to | |
586 | * prevent racing processes from causing the surplus to exceed | |
587 | * overcommit | |
588 | * | |
589 | * This however introduces a different race, where a process B | |
590 | * tries to grow the static hugepage pool while alloc_pages() is | |
591 | * called by process A. B will only examine the per-node | |
592 | * counters in determining if surplus huge pages can be | |
593 | * converted to normal huge pages in adjust_pool_surplus(). A | |
594 | * won't be able to increment the per-node counter, until the | |
595 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
596 | * no more huge pages can be converted from surplus to normal | |
597 | * state (and doesn't try to convert again). Thus, we have a | |
598 | * case where a surplus huge page exists, the pool is grown, and | |
599 | * the surplus huge page still exists after, even though it | |
600 | * should just have been converted to a normal huge page. This | |
601 | * does not leak memory, though, as the hugepage will be freed | |
602 | * once it is out of use. It also does not allow the counters to | |
603 | * go out of whack in adjust_pool_surplus() as we don't modify | |
604 | * the node values until we've gotten the hugepage and only the | |
605 | * per-node value is checked there. | |
606 | */ | |
607 | spin_lock(&hugetlb_lock); | |
608 | if (surplus_huge_pages >= nr_overcommit_huge_pages) { | |
609 | spin_unlock(&hugetlb_lock); | |
610 | return NULL; | |
611 | } else { | |
612 | nr_huge_pages++; | |
613 | surplus_huge_pages++; | |
614 | } | |
615 | spin_unlock(&hugetlb_lock); | |
616 | ||
551883ae NA |
617 | page = alloc_pages(htlb_alloc_mask|__GFP_COMP| |
618 | __GFP_REPEAT|__GFP_NOWARN, | |
7893d1d5 | 619 | HUGETLB_PAGE_ORDER); |
d1c3fb1f NA |
620 | |
621 | spin_lock(&hugetlb_lock); | |
7893d1d5 | 622 | if (page) { |
2668db91 AL |
623 | /* |
624 | * This page is now managed by the hugetlb allocator and has | |
625 | * no users -- drop the buddy allocator's reference. | |
626 | */ | |
627 | put_page_testzero(page); | |
628 | VM_BUG_ON(page_count(page)); | |
d1c3fb1f | 629 | nid = page_to_nid(page); |
7893d1d5 | 630 | set_compound_page_dtor(page, free_huge_page); |
d1c3fb1f NA |
631 | /* |
632 | * We incremented the global counters already | |
633 | */ | |
634 | nr_huge_pages_node[nid]++; | |
635 | surplus_huge_pages_node[nid]++; | |
3b116300 | 636 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f NA |
637 | } else { |
638 | nr_huge_pages--; | |
639 | surplus_huge_pages--; | |
3b116300 | 640 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 641 | } |
d1c3fb1f | 642 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
643 | |
644 | return page; | |
645 | } | |
646 | ||
e4e574b7 AL |
647 | /* |
648 | * Increase the hugetlb pool such that it can accomodate a reservation | |
649 | * of size 'delta'. | |
650 | */ | |
651 | static int gather_surplus_pages(int delta) | |
652 | { | |
653 | struct list_head surplus_list; | |
654 | struct page *page, *tmp; | |
655 | int ret, i; | |
656 | int needed, allocated; | |
657 | ||
658 | needed = (resv_huge_pages + delta) - free_huge_pages; | |
ac09b3a1 AL |
659 | if (needed <= 0) { |
660 | resv_huge_pages += delta; | |
e4e574b7 | 661 | return 0; |
ac09b3a1 | 662 | } |
e4e574b7 AL |
663 | |
664 | allocated = 0; | |
665 | INIT_LIST_HEAD(&surplus_list); | |
666 | ||
667 | ret = -ENOMEM; | |
668 | retry: | |
669 | spin_unlock(&hugetlb_lock); | |
670 | for (i = 0; i < needed; i++) { | |
671 | page = alloc_buddy_huge_page(NULL, 0); | |
672 | if (!page) { | |
673 | /* | |
674 | * We were not able to allocate enough pages to | |
675 | * satisfy the entire reservation so we free what | |
676 | * we've allocated so far. | |
677 | */ | |
678 | spin_lock(&hugetlb_lock); | |
679 | needed = 0; | |
680 | goto free; | |
681 | } | |
682 | ||
683 | list_add(&page->lru, &surplus_list); | |
684 | } | |
685 | allocated += needed; | |
686 | ||
687 | /* | |
688 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
689 | * because either resv_huge_pages or free_huge_pages may have changed. | |
690 | */ | |
691 | spin_lock(&hugetlb_lock); | |
692 | needed = (resv_huge_pages + delta) - (free_huge_pages + allocated); | |
693 | if (needed > 0) | |
694 | goto retry; | |
695 | ||
696 | /* | |
697 | * The surplus_list now contains _at_least_ the number of extra pages | |
698 | * needed to accomodate the reservation. Add the appropriate number | |
699 | * of pages to the hugetlb pool and free the extras back to the buddy | |
ac09b3a1 AL |
700 | * allocator. Commit the entire reservation here to prevent another |
701 | * process from stealing the pages as they are added to the pool but | |
702 | * before they are reserved. | |
e4e574b7 AL |
703 | */ |
704 | needed += allocated; | |
ac09b3a1 | 705 | resv_huge_pages += delta; |
e4e574b7 AL |
706 | ret = 0; |
707 | free: | |
19fc3f0a | 708 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 709 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
710 | if ((--needed) < 0) |
711 | break; | |
e4e574b7 | 712 | list_del(&page->lru); |
19fc3f0a AL |
713 | enqueue_huge_page(page); |
714 | } | |
715 | ||
716 | /* Free unnecessary surplus pages to the buddy allocator */ | |
717 | if (!list_empty(&surplus_list)) { | |
718 | spin_unlock(&hugetlb_lock); | |
719 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { | |
720 | list_del(&page->lru); | |
af767cbd | 721 | /* |
2668db91 AL |
722 | * The page has a reference count of zero already, so |
723 | * call free_huge_page directly instead of using | |
724 | * put_page. This must be done with hugetlb_lock | |
af767cbd AL |
725 | * unlocked which is safe because free_huge_page takes |
726 | * hugetlb_lock before deciding how to free the page. | |
727 | */ | |
2668db91 | 728 | free_huge_page(page); |
af767cbd | 729 | } |
19fc3f0a | 730 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
731 | } |
732 | ||
733 | return ret; | |
734 | } | |
735 | ||
736 | /* | |
737 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
738 | * allocated to satisfy the reservation must be explicitly freed if they were | |
739 | * never used. | |
740 | */ | |
8cde045c | 741 | static void return_unused_surplus_pages(unsigned long unused_resv_pages) |
e4e574b7 AL |
742 | { |
743 | static int nid = -1; | |
744 | struct page *page; | |
745 | unsigned long nr_pages; | |
746 | ||
11320d17 NA |
747 | /* |
748 | * We want to release as many surplus pages as possible, spread | |
749 | * evenly across all nodes. Iterate across all nodes until we | |
750 | * can no longer free unreserved surplus pages. This occurs when | |
751 | * the nodes with surplus pages have no free pages. | |
752 | */ | |
753 | unsigned long remaining_iterations = num_online_nodes(); | |
754 | ||
ac09b3a1 AL |
755 | /* Uncommit the reservation */ |
756 | resv_huge_pages -= unused_resv_pages; | |
757 | ||
e4e574b7 AL |
758 | nr_pages = min(unused_resv_pages, surplus_huge_pages); |
759 | ||
11320d17 | 760 | while (remaining_iterations-- && nr_pages) { |
e4e574b7 AL |
761 | nid = next_node(nid, node_online_map); |
762 | if (nid == MAX_NUMNODES) | |
763 | nid = first_node(node_online_map); | |
764 | ||
765 | if (!surplus_huge_pages_node[nid]) | |
766 | continue; | |
767 | ||
768 | if (!list_empty(&hugepage_freelists[nid])) { | |
769 | page = list_entry(hugepage_freelists[nid].next, | |
770 | struct page, lru); | |
771 | list_del(&page->lru); | |
772 | update_and_free_page(page); | |
773 | free_huge_pages--; | |
774 | free_huge_pages_node[nid]--; | |
775 | surplus_huge_pages--; | |
776 | surplus_huge_pages_node[nid]--; | |
777 | nr_pages--; | |
11320d17 | 778 | remaining_iterations = num_online_nodes(); |
e4e574b7 AL |
779 | } |
780 | } | |
781 | } | |
782 | ||
c37f9fb1 AW |
783 | /* |
784 | * Determine if the huge page at addr within the vma has an associated | |
785 | * reservation. Where it does not we will need to logically increase | |
786 | * reservation and actually increase quota before an allocation can occur. | |
787 | * Where any new reservation would be required the reservation change is | |
788 | * prepared, but not committed. Once the page has been quota'd allocated | |
789 | * an instantiated the change should be committed via vma_commit_reservation. | |
790 | * No action is required on failure. | |
791 | */ | |
792 | static int vma_needs_reservation(struct vm_area_struct *vma, unsigned long addr) | |
793 | { | |
794 | struct address_space *mapping = vma->vm_file->f_mapping; | |
795 | struct inode *inode = mapping->host; | |
796 | ||
797 | if (vma->vm_flags & VM_SHARED) { | |
a858f7b2 | 798 | pgoff_t idx = vma_hugecache_offset(vma, addr); |
c37f9fb1 AW |
799 | return region_chg(&inode->i_mapping->private_list, |
800 | idx, idx + 1); | |
801 | ||
84afd99b AW |
802 | } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
803 | return 1; | |
c37f9fb1 | 804 | |
84afd99b AW |
805 | } else { |
806 | int err; | |
a858f7b2 | 807 | pgoff_t idx = vma_hugecache_offset(vma, addr); |
84afd99b AW |
808 | struct resv_map *reservations = vma_resv_map(vma); |
809 | ||
810 | err = region_chg(&reservations->regions, idx, idx + 1); | |
811 | if (err < 0) | |
812 | return err; | |
813 | return 0; | |
814 | } | |
c37f9fb1 AW |
815 | } |
816 | static void vma_commit_reservation(struct vm_area_struct *vma, | |
817 | unsigned long addr) | |
818 | { | |
819 | struct address_space *mapping = vma->vm_file->f_mapping; | |
820 | struct inode *inode = mapping->host; | |
821 | ||
822 | if (vma->vm_flags & VM_SHARED) { | |
a858f7b2 | 823 | pgoff_t idx = vma_hugecache_offset(vma, addr); |
c37f9fb1 | 824 | region_add(&inode->i_mapping->private_list, idx, idx + 1); |
84afd99b AW |
825 | |
826 | } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { | |
a858f7b2 | 827 | pgoff_t idx = vma_hugecache_offset(vma, addr); |
84afd99b AW |
828 | struct resv_map *reservations = vma_resv_map(vma); |
829 | ||
830 | /* Mark this page used in the map. */ | |
831 | region_add(&reservations->regions, idx, idx + 1); | |
c37f9fb1 AW |
832 | } |
833 | } | |
834 | ||
a1e78772 | 835 | static struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 836 | unsigned long addr, int avoid_reserve) |
1da177e4 | 837 | { |
348ea204 | 838 | struct page *page; |
a1e78772 MG |
839 | struct address_space *mapping = vma->vm_file->f_mapping; |
840 | struct inode *inode = mapping->host; | |
c37f9fb1 | 841 | unsigned int chg; |
a1e78772 MG |
842 | |
843 | /* | |
844 | * Processes that did not create the mapping will have no reserves and | |
845 | * will not have accounted against quota. Check that the quota can be | |
846 | * made before satisfying the allocation | |
c37f9fb1 AW |
847 | * MAP_NORESERVE mappings may also need pages and quota allocated |
848 | * if no reserve mapping overlaps. | |
a1e78772 | 849 | */ |
c37f9fb1 AW |
850 | chg = vma_needs_reservation(vma, addr); |
851 | if (chg < 0) | |
852 | return ERR_PTR(chg); | |
853 | if (chg) | |
a1e78772 MG |
854 | if (hugetlb_get_quota(inode->i_mapping, chg)) |
855 | return ERR_PTR(-ENOSPC); | |
1da177e4 LT |
856 | |
857 | spin_lock(&hugetlb_lock); | |
04f2cbe3 | 858 | page = dequeue_huge_page_vma(vma, addr, avoid_reserve); |
1da177e4 | 859 | spin_unlock(&hugetlb_lock); |
b45b5bd6 | 860 | |
68842c9b | 861 | if (!page) { |
7893d1d5 | 862 | page = alloc_buddy_huge_page(vma, addr); |
68842c9b | 863 | if (!page) { |
a1e78772 | 864 | hugetlb_put_quota(inode->i_mapping, chg); |
68842c9b KC |
865 | return ERR_PTR(-VM_FAULT_OOM); |
866 | } | |
867 | } | |
348ea204 | 868 | |
a1e78772 MG |
869 | set_page_refcounted(page); |
870 | set_page_private(page, (unsigned long) mapping); | |
90d8b7e6 | 871 | |
c37f9fb1 AW |
872 | vma_commit_reservation(vma, addr); |
873 | ||
90d8b7e6 | 874 | return page; |
b45b5bd6 DG |
875 | } |
876 | ||
1da177e4 LT |
877 | static int __init hugetlb_init(void) |
878 | { | |
879 | unsigned long i; | |
1da177e4 | 880 | |
3c726f8d BH |
881 | if (HPAGE_SHIFT == 0) |
882 | return 0; | |
883 | ||
1da177e4 LT |
884 | for (i = 0; i < MAX_NUMNODES; ++i) |
885 | INIT_LIST_HEAD(&hugepage_freelists[i]); | |
886 | ||
63b4613c NA |
887 | hugetlb_next_nid = first_node(node_online_map); |
888 | ||
1da177e4 | 889 | for (i = 0; i < max_huge_pages; ++i) { |
a482289d | 890 | if (!alloc_fresh_huge_page()) |
1da177e4 | 891 | break; |
1da177e4 LT |
892 | } |
893 | max_huge_pages = free_huge_pages = nr_huge_pages = i; | |
894 | printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages); | |
895 | return 0; | |
896 | } | |
897 | module_init(hugetlb_init); | |
898 | ||
899 | static int __init hugetlb_setup(char *s) | |
900 | { | |
901 | if (sscanf(s, "%lu", &max_huge_pages) <= 0) | |
902 | max_huge_pages = 0; | |
903 | return 1; | |
904 | } | |
905 | __setup("hugepages=", hugetlb_setup); | |
906 | ||
8a630112 KC |
907 | static unsigned int cpuset_mems_nr(unsigned int *array) |
908 | { | |
909 | int node; | |
910 | unsigned int nr = 0; | |
911 | ||
912 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
913 | nr += array[node]; | |
914 | ||
915 | return nr; | |
916 | } | |
917 | ||
1da177e4 | 918 | #ifdef CONFIG_SYSCTL |
1da177e4 LT |
919 | #ifdef CONFIG_HIGHMEM |
920 | static void try_to_free_low(unsigned long count) | |
921 | { | |
4415cc8d CL |
922 | int i; |
923 | ||
1da177e4 LT |
924 | for (i = 0; i < MAX_NUMNODES; ++i) { |
925 | struct page *page, *next; | |
926 | list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) { | |
6b0c880d AL |
927 | if (count >= nr_huge_pages) |
928 | return; | |
1da177e4 LT |
929 | if (PageHighMem(page)) |
930 | continue; | |
931 | list_del(&page->lru); | |
932 | update_and_free_page(page); | |
1da177e4 | 933 | free_huge_pages--; |
4415cc8d | 934 | free_huge_pages_node[page_to_nid(page)]--; |
1da177e4 LT |
935 | } |
936 | } | |
937 | } | |
938 | #else | |
939 | static inline void try_to_free_low(unsigned long count) | |
940 | { | |
941 | } | |
942 | #endif | |
943 | ||
7893d1d5 | 944 | #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages) |
1da177e4 LT |
945 | static unsigned long set_max_huge_pages(unsigned long count) |
946 | { | |
7893d1d5 | 947 | unsigned long min_count, ret; |
1da177e4 | 948 | |
7893d1d5 AL |
949 | /* |
950 | * Increase the pool size | |
951 | * First take pages out of surplus state. Then make up the | |
952 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f NA |
953 | * |
954 | * We might race with alloc_buddy_huge_page() here and be unable | |
955 | * to convert a surplus huge page to a normal huge page. That is | |
956 | * not critical, though, it just means the overall size of the | |
957 | * pool might be one hugepage larger than it needs to be, but | |
958 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 959 | */ |
1da177e4 | 960 | spin_lock(&hugetlb_lock); |
7893d1d5 AL |
961 | while (surplus_huge_pages && count > persistent_huge_pages) { |
962 | if (!adjust_pool_surplus(-1)) | |
963 | break; | |
964 | } | |
965 | ||
966 | while (count > persistent_huge_pages) { | |
7893d1d5 AL |
967 | /* |
968 | * If this allocation races such that we no longer need the | |
969 | * page, free_huge_page will handle it by freeing the page | |
970 | * and reducing the surplus. | |
971 | */ | |
972 | spin_unlock(&hugetlb_lock); | |
973 | ret = alloc_fresh_huge_page(); | |
974 | spin_lock(&hugetlb_lock); | |
975 | if (!ret) | |
976 | goto out; | |
977 | ||
978 | } | |
7893d1d5 AL |
979 | |
980 | /* | |
981 | * Decrease the pool size | |
982 | * First return free pages to the buddy allocator (being careful | |
983 | * to keep enough around to satisfy reservations). Then place | |
984 | * pages into surplus state as needed so the pool will shrink | |
985 | * to the desired size as pages become free. | |
d1c3fb1f NA |
986 | * |
987 | * By placing pages into the surplus state independent of the | |
988 | * overcommit value, we are allowing the surplus pool size to | |
989 | * exceed overcommit. There are few sane options here. Since | |
990 | * alloc_buddy_huge_page() is checking the global counter, | |
991 | * though, we'll note that we're not allowed to exceed surplus | |
992 | * and won't grow the pool anywhere else. Not until one of the | |
993 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 994 | */ |
6b0c880d AL |
995 | min_count = resv_huge_pages + nr_huge_pages - free_huge_pages; |
996 | min_count = max(count, min_count); | |
7893d1d5 AL |
997 | try_to_free_low(min_count); |
998 | while (min_count < persistent_huge_pages) { | |
348e1e04 | 999 | struct page *page = dequeue_huge_page(); |
1da177e4 LT |
1000 | if (!page) |
1001 | break; | |
1002 | update_and_free_page(page); | |
1003 | } | |
7893d1d5 AL |
1004 | while (count < persistent_huge_pages) { |
1005 | if (!adjust_pool_surplus(1)) | |
1006 | break; | |
1007 | } | |
1008 | out: | |
1009 | ret = persistent_huge_pages; | |
1da177e4 | 1010 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 1011 | return ret; |
1da177e4 LT |
1012 | } |
1013 | ||
1014 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, | |
1015 | struct file *file, void __user *buffer, | |
1016 | size_t *length, loff_t *ppos) | |
1017 | { | |
1018 | proc_doulongvec_minmax(table, write, file, buffer, length, ppos); | |
1019 | max_huge_pages = set_max_huge_pages(max_huge_pages); | |
1020 | return 0; | |
1021 | } | |
396faf03 MG |
1022 | |
1023 | int hugetlb_treat_movable_handler(struct ctl_table *table, int write, | |
1024 | struct file *file, void __user *buffer, | |
1025 | size_t *length, loff_t *ppos) | |
1026 | { | |
1027 | proc_dointvec(table, write, file, buffer, length, ppos); | |
1028 | if (hugepages_treat_as_movable) | |
1029 | htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; | |
1030 | else | |
1031 | htlb_alloc_mask = GFP_HIGHUSER; | |
1032 | return 0; | |
1033 | } | |
1034 | ||
a3d0c6aa NA |
1035 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
1036 | struct file *file, void __user *buffer, | |
1037 | size_t *length, loff_t *ppos) | |
1038 | { | |
a3d0c6aa | 1039 | proc_doulongvec_minmax(table, write, file, buffer, length, ppos); |
064d9efe NA |
1040 | spin_lock(&hugetlb_lock); |
1041 | nr_overcommit_huge_pages = sysctl_overcommit_huge_pages; | |
a3d0c6aa NA |
1042 | spin_unlock(&hugetlb_lock); |
1043 | return 0; | |
1044 | } | |
1045 | ||
1da177e4 LT |
1046 | #endif /* CONFIG_SYSCTL */ |
1047 | ||
1048 | int hugetlb_report_meminfo(char *buf) | |
1049 | { | |
1050 | return sprintf(buf, | |
1051 | "HugePages_Total: %5lu\n" | |
1052 | "HugePages_Free: %5lu\n" | |
a43a8c39 | 1053 | "HugePages_Rsvd: %5lu\n" |
7893d1d5 | 1054 | "HugePages_Surp: %5lu\n" |
1da177e4 LT |
1055 | "Hugepagesize: %5lu kB\n", |
1056 | nr_huge_pages, | |
1057 | free_huge_pages, | |
a43a8c39 | 1058 | resv_huge_pages, |
7893d1d5 | 1059 | surplus_huge_pages, |
1da177e4 LT |
1060 | HPAGE_SIZE/1024); |
1061 | } | |
1062 | ||
1063 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
1064 | { | |
1065 | return sprintf(buf, | |
1066 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
1067 | "Node %d HugePages_Free: %5u\n" |
1068 | "Node %d HugePages_Surp: %5u\n", | |
1da177e4 | 1069 | nid, nr_huge_pages_node[nid], |
a1de0919 NA |
1070 | nid, free_huge_pages_node[nid], |
1071 | nid, surplus_huge_pages_node[nid]); | |
1da177e4 LT |
1072 | } |
1073 | ||
1da177e4 LT |
1074 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
1075 | unsigned long hugetlb_total_pages(void) | |
1076 | { | |
1077 | return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE); | |
1078 | } | |
1da177e4 | 1079 | |
fc1b8a73 MG |
1080 | static int hugetlb_acct_memory(long delta) |
1081 | { | |
1082 | int ret = -ENOMEM; | |
1083 | ||
1084 | spin_lock(&hugetlb_lock); | |
1085 | /* | |
1086 | * When cpuset is configured, it breaks the strict hugetlb page | |
1087 | * reservation as the accounting is done on a global variable. Such | |
1088 | * reservation is completely rubbish in the presence of cpuset because | |
1089 | * the reservation is not checked against page availability for the | |
1090 | * current cpuset. Application can still potentially OOM'ed by kernel | |
1091 | * with lack of free htlb page in cpuset that the task is in. | |
1092 | * Attempt to enforce strict accounting with cpuset is almost | |
1093 | * impossible (or too ugly) because cpuset is too fluid that | |
1094 | * task or memory node can be dynamically moved between cpusets. | |
1095 | * | |
1096 | * The change of semantics for shared hugetlb mapping with cpuset is | |
1097 | * undesirable. However, in order to preserve some of the semantics, | |
1098 | * we fall back to check against current free page availability as | |
1099 | * a best attempt and hopefully to minimize the impact of changing | |
1100 | * semantics that cpuset has. | |
1101 | */ | |
1102 | if (delta > 0) { | |
1103 | if (gather_surplus_pages(delta) < 0) | |
1104 | goto out; | |
1105 | ||
1106 | if (delta > cpuset_mems_nr(free_huge_pages_node)) { | |
1107 | return_unused_surplus_pages(delta); | |
1108 | goto out; | |
1109 | } | |
1110 | } | |
1111 | ||
1112 | ret = 0; | |
1113 | if (delta < 0) | |
1114 | return_unused_surplus_pages((unsigned long) -delta); | |
1115 | ||
1116 | out: | |
1117 | spin_unlock(&hugetlb_lock); | |
1118 | return ret; | |
1119 | } | |
1120 | ||
84afd99b AW |
1121 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
1122 | { | |
1123 | struct resv_map *reservations = vma_resv_map(vma); | |
1124 | ||
1125 | /* | |
1126 | * This new VMA should share its siblings reservation map if present. | |
1127 | * The VMA will only ever have a valid reservation map pointer where | |
1128 | * it is being copied for another still existing VMA. As that VMA | |
1129 | * has a reference to the reservation map it cannot dissappear until | |
1130 | * after this open call completes. It is therefore safe to take a | |
1131 | * new reference here without additional locking. | |
1132 | */ | |
1133 | if (reservations) | |
1134 | kref_get(&reservations->refs); | |
1135 | } | |
1136 | ||
a1e78772 MG |
1137 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
1138 | { | |
84afd99b AW |
1139 | struct resv_map *reservations = vma_resv_map(vma); |
1140 | unsigned long reserve; | |
1141 | unsigned long start; | |
1142 | unsigned long end; | |
1143 | ||
1144 | if (reservations) { | |
a858f7b2 JW |
1145 | start = vma_hugecache_offset(vma, vma->vm_start); |
1146 | end = vma_hugecache_offset(vma, vma->vm_end); | |
84afd99b AW |
1147 | |
1148 | reserve = (end - start) - | |
1149 | region_count(&reservations->regions, start, end); | |
1150 | ||
1151 | kref_put(&reservations->refs, resv_map_release); | |
1152 | ||
1153 | if (reserve) | |
1154 | hugetlb_acct_memory(-reserve); | |
1155 | } | |
a1e78772 MG |
1156 | } |
1157 | ||
1da177e4 LT |
1158 | /* |
1159 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
1160 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
1161 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
1162 | * this far. | |
1163 | */ | |
d0217ac0 | 1164 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
1165 | { |
1166 | BUG(); | |
d0217ac0 | 1167 | return 0; |
1da177e4 LT |
1168 | } |
1169 | ||
1170 | struct vm_operations_struct hugetlb_vm_ops = { | |
d0217ac0 | 1171 | .fault = hugetlb_vm_op_fault, |
84afd99b | 1172 | .open = hugetlb_vm_op_open, |
a1e78772 | 1173 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
1174 | }; |
1175 | ||
1e8f889b DG |
1176 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
1177 | int writable) | |
63551ae0 DG |
1178 | { |
1179 | pte_t entry; | |
1180 | ||
1e8f889b | 1181 | if (writable) { |
63551ae0 DG |
1182 | entry = |
1183 | pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); | |
1184 | } else { | |
7f2e9525 | 1185 | entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot)); |
63551ae0 DG |
1186 | } |
1187 | entry = pte_mkyoung(entry); | |
1188 | entry = pte_mkhuge(entry); | |
1189 | ||
1190 | return entry; | |
1191 | } | |
1192 | ||
1e8f889b DG |
1193 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
1194 | unsigned long address, pte_t *ptep) | |
1195 | { | |
1196 | pte_t entry; | |
1197 | ||
7f2e9525 GS |
1198 | entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep))); |
1199 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) { | |
8dab5241 | 1200 | update_mmu_cache(vma, address, entry); |
8dab5241 | 1201 | } |
1e8f889b DG |
1202 | } |
1203 | ||
1204 | ||
63551ae0 DG |
1205 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
1206 | struct vm_area_struct *vma) | |
1207 | { | |
1208 | pte_t *src_pte, *dst_pte, entry; | |
1209 | struct page *ptepage; | |
1c59827d | 1210 | unsigned long addr; |
1e8f889b DG |
1211 | int cow; |
1212 | ||
1213 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 1214 | |
1c59827d | 1215 | for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) { |
c74df32c HD |
1216 | src_pte = huge_pte_offset(src, addr); |
1217 | if (!src_pte) | |
1218 | continue; | |
63551ae0 DG |
1219 | dst_pte = huge_pte_alloc(dst, addr); |
1220 | if (!dst_pte) | |
1221 | goto nomem; | |
c5c99429 LW |
1222 | |
1223 | /* If the pagetables are shared don't copy or take references */ | |
1224 | if (dst_pte == src_pte) | |
1225 | continue; | |
1226 | ||
c74df32c | 1227 | spin_lock(&dst->page_table_lock); |
46478758 | 1228 | spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); |
7f2e9525 | 1229 | if (!huge_pte_none(huge_ptep_get(src_pte))) { |
1e8f889b | 1230 | if (cow) |
7f2e9525 GS |
1231 | huge_ptep_set_wrprotect(src, addr, src_pte); |
1232 | entry = huge_ptep_get(src_pte); | |
1c59827d HD |
1233 | ptepage = pte_page(entry); |
1234 | get_page(ptepage); | |
1c59827d HD |
1235 | set_huge_pte_at(dst, addr, dst_pte, entry); |
1236 | } | |
1237 | spin_unlock(&src->page_table_lock); | |
c74df32c | 1238 | spin_unlock(&dst->page_table_lock); |
63551ae0 DG |
1239 | } |
1240 | return 0; | |
1241 | ||
1242 | nomem: | |
1243 | return -ENOMEM; | |
1244 | } | |
1245 | ||
502717f4 | 1246 | void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 1247 | unsigned long end, struct page *ref_page) |
63551ae0 DG |
1248 | { |
1249 | struct mm_struct *mm = vma->vm_mm; | |
1250 | unsigned long address; | |
c7546f8f | 1251 | pte_t *ptep; |
63551ae0 DG |
1252 | pte_t pte; |
1253 | struct page *page; | |
fe1668ae | 1254 | struct page *tmp; |
c0a499c2 CK |
1255 | /* |
1256 | * A page gathering list, protected by per file i_mmap_lock. The | |
1257 | * lock is used to avoid list corruption from multiple unmapping | |
1258 | * of the same page since we are using page->lru. | |
1259 | */ | |
fe1668ae | 1260 | LIST_HEAD(page_list); |
63551ae0 DG |
1261 | |
1262 | WARN_ON(!is_vm_hugetlb_page(vma)); | |
1263 | BUG_ON(start & ~HPAGE_MASK); | |
1264 | BUG_ON(end & ~HPAGE_MASK); | |
1265 | ||
508034a3 | 1266 | spin_lock(&mm->page_table_lock); |
63551ae0 | 1267 | for (address = start; address < end; address += HPAGE_SIZE) { |
c7546f8f | 1268 | ptep = huge_pte_offset(mm, address); |
4c887265 | 1269 | if (!ptep) |
c7546f8f DG |
1270 | continue; |
1271 | ||
39dde65c CK |
1272 | if (huge_pmd_unshare(mm, &address, ptep)) |
1273 | continue; | |
1274 | ||
04f2cbe3 MG |
1275 | /* |
1276 | * If a reference page is supplied, it is because a specific | |
1277 | * page is being unmapped, not a range. Ensure the page we | |
1278 | * are about to unmap is the actual page of interest. | |
1279 | */ | |
1280 | if (ref_page) { | |
1281 | pte = huge_ptep_get(ptep); | |
1282 | if (huge_pte_none(pte)) | |
1283 | continue; | |
1284 | page = pte_page(pte); | |
1285 | if (page != ref_page) | |
1286 | continue; | |
1287 | ||
1288 | /* | |
1289 | * Mark the VMA as having unmapped its page so that | |
1290 | * future faults in this VMA will fail rather than | |
1291 | * looking like data was lost | |
1292 | */ | |
1293 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
1294 | } | |
1295 | ||
c7546f8f | 1296 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
7f2e9525 | 1297 | if (huge_pte_none(pte)) |
63551ae0 | 1298 | continue; |
c7546f8f | 1299 | |
63551ae0 | 1300 | page = pte_page(pte); |
6649a386 KC |
1301 | if (pte_dirty(pte)) |
1302 | set_page_dirty(page); | |
fe1668ae | 1303 | list_add(&page->lru, &page_list); |
63551ae0 | 1304 | } |
1da177e4 | 1305 | spin_unlock(&mm->page_table_lock); |
508034a3 | 1306 | flush_tlb_range(vma, start, end); |
fe1668ae CK |
1307 | list_for_each_entry_safe(page, tmp, &page_list, lru) { |
1308 | list_del(&page->lru); | |
1309 | put_page(page); | |
1310 | } | |
1da177e4 | 1311 | } |
63551ae0 | 1312 | |
502717f4 | 1313 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 1314 | unsigned long end, struct page *ref_page) |
502717f4 CK |
1315 | { |
1316 | /* | |
1317 | * It is undesirable to test vma->vm_file as it should be non-null | |
1318 | * for valid hugetlb area. However, vm_file will be NULL in the error | |
1319 | * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails, | |
1320 | * do_mmap_pgoff() nullifies vma->vm_file before calling this function | |
1321 | * to clean up. Since no pte has actually been setup, it is safe to | |
1322 | * do nothing in this case. | |
1323 | */ | |
1324 | if (vma->vm_file) { | |
1325 | spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); | |
04f2cbe3 | 1326 | __unmap_hugepage_range(vma, start, end, ref_page); |
502717f4 CK |
1327 | spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); |
1328 | } | |
1329 | } | |
1330 | ||
04f2cbe3 MG |
1331 | /* |
1332 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
1333 | * mappping it owns the reserve page for. The intention is to unmap the page | |
1334 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
1335 | * same region. | |
1336 | */ | |
1337 | int unmap_ref_private(struct mm_struct *mm, | |
1338 | struct vm_area_struct *vma, | |
1339 | struct page *page, | |
1340 | unsigned long address) | |
1341 | { | |
1342 | struct vm_area_struct *iter_vma; | |
1343 | struct address_space *mapping; | |
1344 | struct prio_tree_iter iter; | |
1345 | pgoff_t pgoff; | |
1346 | ||
1347 | /* | |
1348 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
1349 | * from page cache lookup which is in HPAGE_SIZE units. | |
1350 | */ | |
1351 | address = address & huge_page_mask(hstate_vma(vma)); | |
1352 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) | |
1353 | + (vma->vm_pgoff >> PAGE_SHIFT); | |
1354 | mapping = (struct address_space *)page_private(page); | |
1355 | ||
1356 | vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
1357 | /* Do not unmap the current VMA */ | |
1358 | if (iter_vma == vma) | |
1359 | continue; | |
1360 | ||
1361 | /* | |
1362 | * Unmap the page from other VMAs without their own reserves. | |
1363 | * They get marked to be SIGKILLed if they fault in these | |
1364 | * areas. This is because a future no-page fault on this VMA | |
1365 | * could insert a zeroed page instead of the data existing | |
1366 | * from the time of fork. This would look like data corruption | |
1367 | */ | |
1368 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
1369 | unmap_hugepage_range(iter_vma, | |
1370 | address, address + HPAGE_SIZE, | |
1371 | page); | |
1372 | } | |
1373 | ||
1374 | return 1; | |
1375 | } | |
1376 | ||
1e8f889b | 1377 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
04f2cbe3 MG |
1378 | unsigned long address, pte_t *ptep, pte_t pte, |
1379 | struct page *pagecache_page) | |
1e8f889b DG |
1380 | { |
1381 | struct page *old_page, *new_page; | |
79ac6ba4 | 1382 | int avoidcopy; |
04f2cbe3 | 1383 | int outside_reserve = 0; |
1e8f889b DG |
1384 | |
1385 | old_page = pte_page(pte); | |
1386 | ||
04f2cbe3 | 1387 | retry_avoidcopy: |
1e8f889b DG |
1388 | /* If no-one else is actually using this page, avoid the copy |
1389 | * and just make the page writable */ | |
1390 | avoidcopy = (page_count(old_page) == 1); | |
1391 | if (avoidcopy) { | |
1392 | set_huge_ptep_writable(vma, address, ptep); | |
83c54070 | 1393 | return 0; |
1e8f889b DG |
1394 | } |
1395 | ||
04f2cbe3 MG |
1396 | /* |
1397 | * If the process that created a MAP_PRIVATE mapping is about to | |
1398 | * perform a COW due to a shared page count, attempt to satisfy | |
1399 | * the allocation without using the existing reserves. The pagecache | |
1400 | * page is used to determine if the reserve at this address was | |
1401 | * consumed or not. If reserves were used, a partial faulted mapping | |
1402 | * at the time of fork() could consume its reserves on COW instead | |
1403 | * of the full address range. | |
1404 | */ | |
1405 | if (!(vma->vm_flags & VM_SHARED) && | |
1406 | is_vma_resv_set(vma, HPAGE_RESV_OWNER) && | |
1407 | old_page != pagecache_page) | |
1408 | outside_reserve = 1; | |
1409 | ||
1e8f889b | 1410 | page_cache_get(old_page); |
04f2cbe3 | 1411 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 1412 | |
2fc39cec | 1413 | if (IS_ERR(new_page)) { |
1e8f889b | 1414 | page_cache_release(old_page); |
04f2cbe3 MG |
1415 | |
1416 | /* | |
1417 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
1418 | * it is due to references held by a child and an insufficient | |
1419 | * huge page pool. To guarantee the original mappers | |
1420 | * reliability, unmap the page from child processes. The child | |
1421 | * may get SIGKILLed if it later faults. | |
1422 | */ | |
1423 | if (outside_reserve) { | |
1424 | BUG_ON(huge_pte_none(pte)); | |
1425 | if (unmap_ref_private(mm, vma, old_page, address)) { | |
1426 | BUG_ON(page_count(old_page) != 1); | |
1427 | BUG_ON(huge_pte_none(pte)); | |
1428 | goto retry_avoidcopy; | |
1429 | } | |
1430 | WARN_ON_ONCE(1); | |
1431 | } | |
1432 | ||
2fc39cec | 1433 | return -PTR_ERR(new_page); |
1e8f889b DG |
1434 | } |
1435 | ||
1436 | spin_unlock(&mm->page_table_lock); | |
9de455b2 | 1437 | copy_huge_page(new_page, old_page, address, vma); |
0ed361de | 1438 | __SetPageUptodate(new_page); |
1e8f889b DG |
1439 | spin_lock(&mm->page_table_lock); |
1440 | ||
1441 | ptep = huge_pte_offset(mm, address & HPAGE_MASK); | |
7f2e9525 | 1442 | if (likely(pte_same(huge_ptep_get(ptep), pte))) { |
1e8f889b | 1443 | /* Break COW */ |
8fe627ec | 1444 | huge_ptep_clear_flush(vma, address, ptep); |
1e8f889b DG |
1445 | set_huge_pte_at(mm, address, ptep, |
1446 | make_huge_pte(vma, new_page, 1)); | |
1447 | /* Make the old page be freed below */ | |
1448 | new_page = old_page; | |
1449 | } | |
1450 | page_cache_release(new_page); | |
1451 | page_cache_release(old_page); | |
83c54070 | 1452 | return 0; |
1e8f889b DG |
1453 | } |
1454 | ||
04f2cbe3 MG |
1455 | /* Return the pagecache page at a given address within a VMA */ |
1456 | static struct page *hugetlbfs_pagecache_page(struct vm_area_struct *vma, | |
1457 | unsigned long address) | |
1458 | { | |
1459 | struct address_space *mapping; | |
e7c4b0bf | 1460 | pgoff_t idx; |
04f2cbe3 MG |
1461 | |
1462 | mapping = vma->vm_file->f_mapping; | |
a858f7b2 | 1463 | idx = vma_hugecache_offset(vma, address); |
04f2cbe3 MG |
1464 | |
1465 | return find_lock_page(mapping, idx); | |
1466 | } | |
1467 | ||
a1ed3dda | 1468 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
1e8f889b | 1469 | unsigned long address, pte_t *ptep, int write_access) |
ac9b9c66 HD |
1470 | { |
1471 | int ret = VM_FAULT_SIGBUS; | |
e7c4b0bf | 1472 | pgoff_t idx; |
4c887265 | 1473 | unsigned long size; |
4c887265 AL |
1474 | struct page *page; |
1475 | struct address_space *mapping; | |
1e8f889b | 1476 | pte_t new_pte; |
4c887265 | 1477 | |
04f2cbe3 MG |
1478 | /* |
1479 | * Currently, we are forced to kill the process in the event the | |
1480 | * original mapper has unmapped pages from the child due to a failed | |
1481 | * COW. Warn that such a situation has occured as it may not be obvious | |
1482 | */ | |
1483 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
1484 | printk(KERN_WARNING | |
1485 | "PID %d killed due to inadequate hugepage pool\n", | |
1486 | current->pid); | |
1487 | return ret; | |
1488 | } | |
1489 | ||
4c887265 | 1490 | mapping = vma->vm_file->f_mapping; |
a858f7b2 | 1491 | idx = vma_hugecache_offset(vma, address); |
4c887265 AL |
1492 | |
1493 | /* | |
1494 | * Use page lock to guard against racing truncation | |
1495 | * before we get page_table_lock. | |
1496 | */ | |
6bda666a CL |
1497 | retry: |
1498 | page = find_lock_page(mapping, idx); | |
1499 | if (!page) { | |
ebed4bfc HD |
1500 | size = i_size_read(mapping->host) >> HPAGE_SHIFT; |
1501 | if (idx >= size) | |
1502 | goto out; | |
04f2cbe3 | 1503 | page = alloc_huge_page(vma, address, 0); |
2fc39cec AL |
1504 | if (IS_ERR(page)) { |
1505 | ret = -PTR_ERR(page); | |
6bda666a CL |
1506 | goto out; |
1507 | } | |
79ac6ba4 | 1508 | clear_huge_page(page, address); |
0ed361de | 1509 | __SetPageUptodate(page); |
ac9b9c66 | 1510 | |
6bda666a CL |
1511 | if (vma->vm_flags & VM_SHARED) { |
1512 | int err; | |
45c682a6 | 1513 | struct inode *inode = mapping->host; |
6bda666a CL |
1514 | |
1515 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
1516 | if (err) { | |
1517 | put_page(page); | |
6bda666a CL |
1518 | if (err == -EEXIST) |
1519 | goto retry; | |
1520 | goto out; | |
1521 | } | |
45c682a6 KC |
1522 | |
1523 | spin_lock(&inode->i_lock); | |
1524 | inode->i_blocks += BLOCKS_PER_HUGEPAGE; | |
1525 | spin_unlock(&inode->i_lock); | |
6bda666a CL |
1526 | } else |
1527 | lock_page(page); | |
1528 | } | |
1e8f889b | 1529 | |
ac9b9c66 | 1530 | spin_lock(&mm->page_table_lock); |
4c887265 AL |
1531 | size = i_size_read(mapping->host) >> HPAGE_SHIFT; |
1532 | if (idx >= size) | |
1533 | goto backout; | |
1534 | ||
83c54070 | 1535 | ret = 0; |
7f2e9525 | 1536 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
1537 | goto backout; |
1538 | ||
1e8f889b DG |
1539 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
1540 | && (vma->vm_flags & VM_SHARED))); | |
1541 | set_huge_pte_at(mm, address, ptep, new_pte); | |
1542 | ||
1543 | if (write_access && !(vma->vm_flags & VM_SHARED)) { | |
1544 | /* Optimization, do the COW without a second fault */ | |
04f2cbe3 | 1545 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); |
1e8f889b DG |
1546 | } |
1547 | ||
ac9b9c66 | 1548 | spin_unlock(&mm->page_table_lock); |
4c887265 AL |
1549 | unlock_page(page); |
1550 | out: | |
ac9b9c66 | 1551 | return ret; |
4c887265 AL |
1552 | |
1553 | backout: | |
1554 | spin_unlock(&mm->page_table_lock); | |
4c887265 AL |
1555 | unlock_page(page); |
1556 | put_page(page); | |
1557 | goto out; | |
ac9b9c66 HD |
1558 | } |
1559 | ||
86e5216f AL |
1560 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
1561 | unsigned long address, int write_access) | |
1562 | { | |
1563 | pte_t *ptep; | |
1564 | pte_t entry; | |
1e8f889b | 1565 | int ret; |
3935baa9 | 1566 | static DEFINE_MUTEX(hugetlb_instantiation_mutex); |
86e5216f AL |
1567 | |
1568 | ptep = huge_pte_alloc(mm, address); | |
1569 | if (!ptep) | |
1570 | return VM_FAULT_OOM; | |
1571 | ||
3935baa9 DG |
1572 | /* |
1573 | * Serialize hugepage allocation and instantiation, so that we don't | |
1574 | * get spurious allocation failures if two CPUs race to instantiate | |
1575 | * the same page in the page cache. | |
1576 | */ | |
1577 | mutex_lock(&hugetlb_instantiation_mutex); | |
7f2e9525 GS |
1578 | entry = huge_ptep_get(ptep); |
1579 | if (huge_pte_none(entry)) { | |
3935baa9 DG |
1580 | ret = hugetlb_no_page(mm, vma, address, ptep, write_access); |
1581 | mutex_unlock(&hugetlb_instantiation_mutex); | |
1582 | return ret; | |
1583 | } | |
86e5216f | 1584 | |
83c54070 | 1585 | ret = 0; |
1e8f889b DG |
1586 | |
1587 | spin_lock(&mm->page_table_lock); | |
1588 | /* Check for a racing update before calling hugetlb_cow */ | |
7f2e9525 | 1589 | if (likely(pte_same(entry, huge_ptep_get(ptep)))) |
04f2cbe3 MG |
1590 | if (write_access && !pte_write(entry)) { |
1591 | struct page *page; | |
1592 | page = hugetlbfs_pagecache_page(vma, address); | |
1593 | ret = hugetlb_cow(mm, vma, address, ptep, entry, page); | |
1594 | if (page) { | |
1595 | unlock_page(page); | |
1596 | put_page(page); | |
1597 | } | |
1598 | } | |
1e8f889b | 1599 | spin_unlock(&mm->page_table_lock); |
3935baa9 | 1600 | mutex_unlock(&hugetlb_instantiation_mutex); |
1e8f889b DG |
1601 | |
1602 | return ret; | |
86e5216f AL |
1603 | } |
1604 | ||
63551ae0 DG |
1605 | int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
1606 | struct page **pages, struct vm_area_struct **vmas, | |
5b23dbe8 AL |
1607 | unsigned long *position, int *length, int i, |
1608 | int write) | |
63551ae0 | 1609 | { |
d5d4b0aa CK |
1610 | unsigned long pfn_offset; |
1611 | unsigned long vaddr = *position; | |
63551ae0 DG |
1612 | int remainder = *length; |
1613 | ||
1c59827d | 1614 | spin_lock(&mm->page_table_lock); |
63551ae0 | 1615 | while (vaddr < vma->vm_end && remainder) { |
4c887265 AL |
1616 | pte_t *pte; |
1617 | struct page *page; | |
63551ae0 | 1618 | |
4c887265 AL |
1619 | /* |
1620 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
1621 | * each hugepage. We have to make * sure we get the | |
1622 | * first, for the page indexing below to work. | |
1623 | */ | |
1624 | pte = huge_pte_offset(mm, vaddr & HPAGE_MASK); | |
63551ae0 | 1625 | |
7f2e9525 GS |
1626 | if (!pte || huge_pte_none(huge_ptep_get(pte)) || |
1627 | (write && !pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 1628 | int ret; |
63551ae0 | 1629 | |
4c887265 | 1630 | spin_unlock(&mm->page_table_lock); |
5b23dbe8 | 1631 | ret = hugetlb_fault(mm, vma, vaddr, write); |
4c887265 | 1632 | spin_lock(&mm->page_table_lock); |
a89182c7 | 1633 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 1634 | continue; |
63551ae0 | 1635 | |
4c887265 AL |
1636 | remainder = 0; |
1637 | if (!i) | |
1638 | i = -EFAULT; | |
1639 | break; | |
1640 | } | |
1641 | ||
d5d4b0aa | 1642 | pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT; |
7f2e9525 | 1643 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 1644 | same_page: |
d6692183 CK |
1645 | if (pages) { |
1646 | get_page(page); | |
d5d4b0aa | 1647 | pages[i] = page + pfn_offset; |
d6692183 | 1648 | } |
63551ae0 DG |
1649 | |
1650 | if (vmas) | |
1651 | vmas[i] = vma; | |
1652 | ||
1653 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 1654 | ++pfn_offset; |
63551ae0 DG |
1655 | --remainder; |
1656 | ++i; | |
d5d4b0aa CK |
1657 | if (vaddr < vma->vm_end && remainder && |
1658 | pfn_offset < HPAGE_SIZE/PAGE_SIZE) { | |
1659 | /* | |
1660 | * We use pfn_offset to avoid touching the pageframes | |
1661 | * of this compound page. | |
1662 | */ | |
1663 | goto same_page; | |
1664 | } | |
63551ae0 | 1665 | } |
1c59827d | 1666 | spin_unlock(&mm->page_table_lock); |
63551ae0 DG |
1667 | *length = remainder; |
1668 | *position = vaddr; | |
1669 | ||
1670 | return i; | |
1671 | } | |
8f860591 ZY |
1672 | |
1673 | void hugetlb_change_protection(struct vm_area_struct *vma, | |
1674 | unsigned long address, unsigned long end, pgprot_t newprot) | |
1675 | { | |
1676 | struct mm_struct *mm = vma->vm_mm; | |
1677 | unsigned long start = address; | |
1678 | pte_t *ptep; | |
1679 | pte_t pte; | |
1680 | ||
1681 | BUG_ON(address >= end); | |
1682 | flush_cache_range(vma, address, end); | |
1683 | ||
39dde65c | 1684 | spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); |
8f860591 ZY |
1685 | spin_lock(&mm->page_table_lock); |
1686 | for (; address < end; address += HPAGE_SIZE) { | |
1687 | ptep = huge_pte_offset(mm, address); | |
1688 | if (!ptep) | |
1689 | continue; | |
39dde65c CK |
1690 | if (huge_pmd_unshare(mm, &address, ptep)) |
1691 | continue; | |
7f2e9525 | 1692 | if (!huge_pte_none(huge_ptep_get(ptep))) { |
8f860591 ZY |
1693 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
1694 | pte = pte_mkhuge(pte_modify(pte, newprot)); | |
1695 | set_huge_pte_at(mm, address, ptep, pte); | |
8f860591 ZY |
1696 | } |
1697 | } | |
1698 | spin_unlock(&mm->page_table_lock); | |
39dde65c | 1699 | spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); |
8f860591 ZY |
1700 | |
1701 | flush_tlb_range(vma, start, end); | |
1702 | } | |
1703 | ||
a1e78772 MG |
1704 | int hugetlb_reserve_pages(struct inode *inode, |
1705 | long from, long to, | |
1706 | struct vm_area_struct *vma) | |
e4e574b7 AL |
1707 | { |
1708 | long ret, chg; | |
1709 | ||
c37f9fb1 AW |
1710 | if (vma && vma->vm_flags & VM_NORESERVE) |
1711 | return 0; | |
1712 | ||
a1e78772 MG |
1713 | /* |
1714 | * Shared mappings base their reservation on the number of pages that | |
1715 | * are already allocated on behalf of the file. Private mappings need | |
1716 | * to reserve the full area even if read-only as mprotect() may be | |
1717 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
1718 | */ | |
1719 | if (!vma || vma->vm_flags & VM_SHARED) | |
1720 | chg = region_chg(&inode->i_mapping->private_list, from, to); | |
1721 | else { | |
84afd99b AW |
1722 | struct resv_map *resv_map = resv_map_alloc(); |
1723 | if (!resv_map) | |
1724 | return -ENOMEM; | |
1725 | ||
a1e78772 | 1726 | chg = to - from; |
84afd99b AW |
1727 | |
1728 | set_vma_resv_map(vma, resv_map); | |
04f2cbe3 | 1729 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); |
a1e78772 MG |
1730 | } |
1731 | ||
e4e574b7 AL |
1732 | if (chg < 0) |
1733 | return chg; | |
8a630112 | 1734 | |
90d8b7e6 AL |
1735 | if (hugetlb_get_quota(inode->i_mapping, chg)) |
1736 | return -ENOSPC; | |
a43a8c39 | 1737 | ret = hugetlb_acct_memory(chg); |
68842c9b KC |
1738 | if (ret < 0) { |
1739 | hugetlb_put_quota(inode->i_mapping, chg); | |
a43a8c39 | 1740 | return ret; |
68842c9b | 1741 | } |
a1e78772 MG |
1742 | if (!vma || vma->vm_flags & VM_SHARED) |
1743 | region_add(&inode->i_mapping->private_list, from, to); | |
a43a8c39 CK |
1744 | return 0; |
1745 | } | |
1746 | ||
1747 | void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) | |
1748 | { | |
1749 | long chg = region_truncate(&inode->i_mapping->private_list, offset); | |
45c682a6 KC |
1750 | |
1751 | spin_lock(&inode->i_lock); | |
1752 | inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed; | |
1753 | spin_unlock(&inode->i_lock); | |
1754 | ||
90d8b7e6 AL |
1755 | hugetlb_put_quota(inode->i_mapping, (chg - freed)); |
1756 | hugetlb_acct_memory(-(chg - freed)); | |
a43a8c39 | 1757 | } |