]>
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> |
aa888a74 | 17 | #include <linux/bootmem.h> |
a3437870 | 18 | #include <linux/sysfs.h> |
5da7ca86 | 19 | |
63551ae0 DG |
20 | #include <asm/page.h> |
21 | #include <asm/pgtable.h> | |
22 | ||
23 | #include <linux/hugetlb.h> | |
7835e98b | 24 | #include "internal.h" |
1da177e4 LT |
25 | |
26 | const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL; | |
396faf03 MG |
27 | static gfp_t htlb_alloc_mask = GFP_HIGHUSER; |
28 | unsigned long hugepages_treat_as_movable; | |
a5516438 | 29 | |
e5ff2159 AK |
30 | static int max_hstate; |
31 | unsigned int default_hstate_idx; | |
32 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
33 | ||
34 | /* for command line parsing */ | |
35 | static struct hstate * __initdata parsed_hstate; | |
36 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 37 | static unsigned long __initdata default_hstate_size; |
e5ff2159 AK |
38 | |
39 | #define for_each_hstate(h) \ | |
40 | for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++) | |
396faf03 | 41 | |
3935baa9 DG |
42 | /* |
43 | * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages | |
44 | */ | |
45 | static DEFINE_SPINLOCK(hugetlb_lock); | |
0bd0f9fb | 46 | |
96822904 AW |
47 | /* |
48 | * Region tracking -- allows tracking of reservations and instantiated pages | |
49 | * across the pages in a mapping. | |
84afd99b AW |
50 | * |
51 | * The region data structures are protected by a combination of the mmap_sem | |
52 | * and the hugetlb_instantion_mutex. To access or modify a region the caller | |
53 | * must either hold the mmap_sem for write, or the mmap_sem for read and | |
54 | * the hugetlb_instantiation mutex: | |
55 | * | |
56 | * down_write(&mm->mmap_sem); | |
57 | * or | |
58 | * down_read(&mm->mmap_sem); | |
59 | * mutex_lock(&hugetlb_instantiation_mutex); | |
96822904 AW |
60 | */ |
61 | struct file_region { | |
62 | struct list_head link; | |
63 | long from; | |
64 | long to; | |
65 | }; | |
66 | ||
67 | static long region_add(struct list_head *head, long f, long t) | |
68 | { | |
69 | struct file_region *rg, *nrg, *trg; | |
70 | ||
71 | /* Locate the region we are either in or before. */ | |
72 | list_for_each_entry(rg, head, link) | |
73 | if (f <= rg->to) | |
74 | break; | |
75 | ||
76 | /* Round our left edge to the current segment if it encloses us. */ | |
77 | if (f > rg->from) | |
78 | f = rg->from; | |
79 | ||
80 | /* Check for and consume any regions we now overlap with. */ | |
81 | nrg = rg; | |
82 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
83 | if (&rg->link == head) | |
84 | break; | |
85 | if (rg->from > t) | |
86 | break; | |
87 | ||
88 | /* If this area reaches higher then extend our area to | |
89 | * include it completely. If this is not the first area | |
90 | * which we intend to reuse, free it. */ | |
91 | if (rg->to > t) | |
92 | t = rg->to; | |
93 | if (rg != nrg) { | |
94 | list_del(&rg->link); | |
95 | kfree(rg); | |
96 | } | |
97 | } | |
98 | nrg->from = f; | |
99 | nrg->to = t; | |
100 | return 0; | |
101 | } | |
102 | ||
103 | static long region_chg(struct list_head *head, long f, long t) | |
104 | { | |
105 | struct file_region *rg, *nrg; | |
106 | long chg = 0; | |
107 | ||
108 | /* Locate the region we are before or in. */ | |
109 | list_for_each_entry(rg, head, link) | |
110 | if (f <= rg->to) | |
111 | break; | |
112 | ||
113 | /* If we are below the current region then a new region is required. | |
114 | * Subtle, allocate a new region at the position but make it zero | |
115 | * size such that we can guarantee to record the reservation. */ | |
116 | if (&rg->link == head || t < rg->from) { | |
117 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
118 | if (!nrg) | |
119 | return -ENOMEM; | |
120 | nrg->from = f; | |
121 | nrg->to = f; | |
122 | INIT_LIST_HEAD(&nrg->link); | |
123 | list_add(&nrg->link, rg->link.prev); | |
124 | ||
125 | return t - f; | |
126 | } | |
127 | ||
128 | /* Round our left edge to the current segment if it encloses us. */ | |
129 | if (f > rg->from) | |
130 | f = rg->from; | |
131 | chg = t - f; | |
132 | ||
133 | /* Check for and consume any regions we now overlap with. */ | |
134 | list_for_each_entry(rg, rg->link.prev, link) { | |
135 | if (&rg->link == head) | |
136 | break; | |
137 | if (rg->from > t) | |
138 | return chg; | |
139 | ||
140 | /* We overlap with this area, if it extends futher than | |
141 | * us then we must extend ourselves. Account for its | |
142 | * existing reservation. */ | |
143 | if (rg->to > t) { | |
144 | chg += rg->to - t; | |
145 | t = rg->to; | |
146 | } | |
147 | chg -= rg->to - rg->from; | |
148 | } | |
149 | return chg; | |
150 | } | |
151 | ||
152 | static long region_truncate(struct list_head *head, long end) | |
153 | { | |
154 | struct file_region *rg, *trg; | |
155 | long chg = 0; | |
156 | ||
157 | /* Locate the region we are either in or before. */ | |
158 | list_for_each_entry(rg, head, link) | |
159 | if (end <= rg->to) | |
160 | break; | |
161 | if (&rg->link == head) | |
162 | return 0; | |
163 | ||
164 | /* If we are in the middle of a region then adjust it. */ | |
165 | if (end > rg->from) { | |
166 | chg = rg->to - end; | |
167 | rg->to = end; | |
168 | rg = list_entry(rg->link.next, typeof(*rg), link); | |
169 | } | |
170 | ||
171 | /* Drop any remaining regions. */ | |
172 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
173 | if (&rg->link == head) | |
174 | break; | |
175 | chg += rg->to - rg->from; | |
176 | list_del(&rg->link); | |
177 | kfree(rg); | |
178 | } | |
179 | return chg; | |
180 | } | |
181 | ||
84afd99b AW |
182 | static long region_count(struct list_head *head, long f, long t) |
183 | { | |
184 | struct file_region *rg; | |
185 | long chg = 0; | |
186 | ||
187 | /* Locate each segment we overlap with, and count that overlap. */ | |
188 | list_for_each_entry(rg, head, link) { | |
189 | int seg_from; | |
190 | int seg_to; | |
191 | ||
192 | if (rg->to <= f) | |
193 | continue; | |
194 | if (rg->from >= t) | |
195 | break; | |
196 | ||
197 | seg_from = max(rg->from, f); | |
198 | seg_to = min(rg->to, t); | |
199 | ||
200 | chg += seg_to - seg_from; | |
201 | } | |
202 | ||
203 | return chg; | |
204 | } | |
205 | ||
e7c4b0bf AW |
206 | /* |
207 | * Convert the address within this vma to the page offset within | |
208 | * the mapping, in pagecache page units; huge pages here. | |
209 | */ | |
a5516438 AK |
210 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
211 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 212 | { |
a5516438 AK |
213 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
214 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
215 | } |
216 | ||
84afd99b AW |
217 | /* |
218 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
219 | * bits of the reservation map pointer, which are always clear due to | |
220 | * alignment. | |
221 | */ | |
222 | #define HPAGE_RESV_OWNER (1UL << 0) | |
223 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 224 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 225 | |
a1e78772 MG |
226 | /* |
227 | * These helpers are used to track how many pages are reserved for | |
228 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
229 | * is guaranteed to have their future faults succeed. | |
230 | * | |
231 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
232 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
233 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
234 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
235 | * |
236 | * The private mapping reservation is represented in a subtly different | |
237 | * manner to a shared mapping. A shared mapping has a region map associated | |
238 | * with the underlying file, this region map represents the backing file | |
239 | * pages which have ever had a reservation assigned which this persists even | |
240 | * after the page is instantiated. A private mapping has a region map | |
241 | * associated with the original mmap which is attached to all VMAs which | |
242 | * reference it, this region map represents those offsets which have consumed | |
243 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 244 | */ |
e7c4b0bf AW |
245 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
246 | { | |
247 | return (unsigned long)vma->vm_private_data; | |
248 | } | |
249 | ||
250 | static void set_vma_private_data(struct vm_area_struct *vma, | |
251 | unsigned long value) | |
252 | { | |
253 | vma->vm_private_data = (void *)value; | |
254 | } | |
255 | ||
84afd99b AW |
256 | struct resv_map { |
257 | struct kref refs; | |
258 | struct list_head regions; | |
259 | }; | |
260 | ||
261 | struct resv_map *resv_map_alloc(void) | |
262 | { | |
263 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
264 | if (!resv_map) | |
265 | return NULL; | |
266 | ||
267 | kref_init(&resv_map->refs); | |
268 | INIT_LIST_HEAD(&resv_map->regions); | |
269 | ||
270 | return resv_map; | |
271 | } | |
272 | ||
273 | void resv_map_release(struct kref *ref) | |
274 | { | |
275 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
276 | ||
277 | /* Clear out any active regions before we release the map. */ | |
278 | region_truncate(&resv_map->regions, 0); | |
279 | kfree(resv_map); | |
280 | } | |
281 | ||
282 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) | |
a1e78772 MG |
283 | { |
284 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
285 | if (!(vma->vm_flags & VM_SHARED)) | |
84afd99b AW |
286 | return (struct resv_map *)(get_vma_private_data(vma) & |
287 | ~HPAGE_RESV_MASK); | |
a1e78772 MG |
288 | return 0; |
289 | } | |
290 | ||
84afd99b | 291 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 MG |
292 | { |
293 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
294 | VM_BUG_ON(vma->vm_flags & VM_SHARED); | |
295 | ||
84afd99b AW |
296 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
297 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
298 | } |
299 | ||
300 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
301 | { | |
04f2cbe3 | 302 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); |
e7c4b0bf AW |
303 | VM_BUG_ON(vma->vm_flags & VM_SHARED); |
304 | ||
305 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
306 | } |
307 | ||
308 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
309 | { | |
310 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
e7c4b0bf AW |
311 | |
312 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
313 | } |
314 | ||
315 | /* Decrement the reserved pages in the hugepage pool by one */ | |
a5516438 AK |
316 | static void decrement_hugepage_resv_vma(struct hstate *h, |
317 | struct vm_area_struct *vma) | |
a1e78772 | 318 | { |
c37f9fb1 AW |
319 | if (vma->vm_flags & VM_NORESERVE) |
320 | return; | |
321 | ||
a1e78772 MG |
322 | if (vma->vm_flags & VM_SHARED) { |
323 | /* Shared mappings always use reserves */ | |
a5516438 | 324 | h->resv_huge_pages--; |
84afd99b | 325 | } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
a1e78772 MG |
326 | /* |
327 | * Only the process that called mmap() has reserves for | |
328 | * private mappings. | |
329 | */ | |
a5516438 | 330 | h->resv_huge_pages--; |
a1e78772 MG |
331 | } |
332 | } | |
333 | ||
04f2cbe3 | 334 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
335 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
336 | { | |
337 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
338 | if (!(vma->vm_flags & VM_SHARED)) | |
339 | vma->vm_private_data = (void *)0; | |
340 | } | |
341 | ||
342 | /* Returns true if the VMA has associated reserve pages */ | |
343 | static int vma_has_private_reserves(struct vm_area_struct *vma) | |
344 | { | |
345 | if (vma->vm_flags & VM_SHARED) | |
346 | return 0; | |
84afd99b | 347 | if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
a1e78772 MG |
348 | return 0; |
349 | return 1; | |
350 | } | |
351 | ||
a5516438 AK |
352 | static void clear_huge_page(struct page *page, |
353 | unsigned long addr, unsigned long sz) | |
79ac6ba4 DG |
354 | { |
355 | int i; | |
356 | ||
357 | might_sleep(); | |
a5516438 | 358 | for (i = 0; i < sz/PAGE_SIZE; i++) { |
79ac6ba4 | 359 | cond_resched(); |
281e0e3b | 360 | clear_user_highpage(page + i, addr + i * PAGE_SIZE); |
79ac6ba4 DG |
361 | } |
362 | } | |
363 | ||
364 | static void copy_huge_page(struct page *dst, struct page *src, | |
9de455b2 | 365 | unsigned long addr, struct vm_area_struct *vma) |
79ac6ba4 DG |
366 | { |
367 | int i; | |
a5516438 | 368 | struct hstate *h = hstate_vma(vma); |
79ac6ba4 DG |
369 | |
370 | might_sleep(); | |
a5516438 | 371 | for (i = 0; i < pages_per_huge_page(h); i++) { |
79ac6ba4 | 372 | cond_resched(); |
9de455b2 | 373 | copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma); |
79ac6ba4 DG |
374 | } |
375 | } | |
376 | ||
a5516438 | 377 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
378 | { |
379 | int nid = page_to_nid(page); | |
a5516438 AK |
380 | list_add(&page->lru, &h->hugepage_freelists[nid]); |
381 | h->free_huge_pages++; | |
382 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
383 | } |
384 | ||
a5516438 | 385 | static struct page *dequeue_huge_page(struct hstate *h) |
348e1e04 NA |
386 | { |
387 | int nid; | |
388 | struct page *page = NULL; | |
389 | ||
390 | for (nid = 0; nid < MAX_NUMNODES; ++nid) { | |
a5516438 AK |
391 | if (!list_empty(&h->hugepage_freelists[nid])) { |
392 | page = list_entry(h->hugepage_freelists[nid].next, | |
348e1e04 NA |
393 | struct page, lru); |
394 | list_del(&page->lru); | |
a5516438 AK |
395 | h->free_huge_pages--; |
396 | h->free_huge_pages_node[nid]--; | |
348e1e04 NA |
397 | break; |
398 | } | |
399 | } | |
400 | return page; | |
401 | } | |
402 | ||
a5516438 AK |
403 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
404 | struct vm_area_struct *vma, | |
04f2cbe3 | 405 | unsigned long address, int avoid_reserve) |
1da177e4 | 406 | { |
31a5c6e4 | 407 | int nid; |
1da177e4 | 408 | struct page *page = NULL; |
480eccf9 | 409 | struct mempolicy *mpol; |
19770b32 | 410 | nodemask_t *nodemask; |
396faf03 | 411 | struct zonelist *zonelist = huge_zonelist(vma, address, |
19770b32 | 412 | htlb_alloc_mask, &mpol, &nodemask); |
dd1a239f MG |
413 | struct zone *zone; |
414 | struct zoneref *z; | |
1da177e4 | 415 | |
a1e78772 MG |
416 | /* |
417 | * A child process with MAP_PRIVATE mappings created by their parent | |
418 | * have no page reserves. This check ensures that reservations are | |
419 | * not "stolen". The child may still get SIGKILLed | |
420 | */ | |
421 | if (!vma_has_private_reserves(vma) && | |
a5516438 | 422 | h->free_huge_pages - h->resv_huge_pages == 0) |
a1e78772 MG |
423 | return NULL; |
424 | ||
04f2cbe3 | 425 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 426 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
04f2cbe3 MG |
427 | return NULL; |
428 | ||
19770b32 MG |
429 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
430 | MAX_NR_ZONES - 1, nodemask) { | |
54a6eb5c MG |
431 | nid = zone_to_nid(zone); |
432 | if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) && | |
a5516438 AK |
433 | !list_empty(&h->hugepage_freelists[nid])) { |
434 | page = list_entry(h->hugepage_freelists[nid].next, | |
3abf7afd AM |
435 | struct page, lru); |
436 | list_del(&page->lru); | |
a5516438 AK |
437 | h->free_huge_pages--; |
438 | h->free_huge_pages_node[nid]--; | |
04f2cbe3 MG |
439 | |
440 | if (!avoid_reserve) | |
a5516438 | 441 | decrement_hugepage_resv_vma(h, vma); |
a1e78772 | 442 | |
5ab3ee7b | 443 | break; |
3abf7afd | 444 | } |
1da177e4 | 445 | } |
52cd3b07 | 446 | mpol_cond_put(mpol); |
1da177e4 LT |
447 | return page; |
448 | } | |
449 | ||
a5516438 | 450 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
451 | { |
452 | int i; | |
a5516438 AK |
453 | |
454 | h->nr_huge_pages--; | |
455 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
456 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
6af2acb6 AL |
457 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced | |
458 | 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved | | |
459 | 1 << PG_private | 1<< PG_writeback); | |
460 | } | |
461 | set_compound_page_dtor(page, NULL); | |
462 | set_page_refcounted(page); | |
7f2e9525 | 463 | arch_release_hugepage(page); |
a5516438 | 464 | __free_pages(page, huge_page_order(h)); |
6af2acb6 AL |
465 | } |
466 | ||
e5ff2159 AK |
467 | struct hstate *size_to_hstate(unsigned long size) |
468 | { | |
469 | struct hstate *h; | |
470 | ||
471 | for_each_hstate(h) { | |
472 | if (huge_page_size(h) == size) | |
473 | return h; | |
474 | } | |
475 | return NULL; | |
476 | } | |
477 | ||
27a85ef1 DG |
478 | static void free_huge_page(struct page *page) |
479 | { | |
a5516438 AK |
480 | /* |
481 | * Can't pass hstate in here because it is called from the | |
482 | * compound page destructor. | |
483 | */ | |
e5ff2159 | 484 | struct hstate *h = page_hstate(page); |
7893d1d5 | 485 | int nid = page_to_nid(page); |
c79fb75e | 486 | struct address_space *mapping; |
27a85ef1 | 487 | |
c79fb75e | 488 | mapping = (struct address_space *) page_private(page); |
e5df70ab | 489 | set_page_private(page, 0); |
7893d1d5 | 490 | BUG_ON(page_count(page)); |
27a85ef1 DG |
491 | INIT_LIST_HEAD(&page->lru); |
492 | ||
493 | spin_lock(&hugetlb_lock); | |
aa888a74 | 494 | if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { |
a5516438 AK |
495 | update_and_free_page(h, page); |
496 | h->surplus_huge_pages--; | |
497 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 498 | } else { |
a5516438 | 499 | enqueue_huge_page(h, page); |
7893d1d5 | 500 | } |
27a85ef1 | 501 | spin_unlock(&hugetlb_lock); |
c79fb75e | 502 | if (mapping) |
9a119c05 | 503 | hugetlb_put_quota(mapping, 1); |
27a85ef1 DG |
504 | } |
505 | ||
7893d1d5 AL |
506 | /* |
507 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
508 | * balanced by operating on them in a round-robin fashion. | |
509 | * Returns 1 if an adjustment was made. | |
510 | */ | |
a5516438 | 511 | static int adjust_pool_surplus(struct hstate *h, int delta) |
7893d1d5 AL |
512 | { |
513 | static int prev_nid; | |
514 | int nid = prev_nid; | |
515 | int ret = 0; | |
516 | ||
517 | VM_BUG_ON(delta != -1 && delta != 1); | |
518 | do { | |
519 | nid = next_node(nid, node_online_map); | |
520 | if (nid == MAX_NUMNODES) | |
521 | nid = first_node(node_online_map); | |
522 | ||
523 | /* To shrink on this node, there must be a surplus page */ | |
a5516438 | 524 | if (delta < 0 && !h->surplus_huge_pages_node[nid]) |
7893d1d5 AL |
525 | continue; |
526 | /* Surplus cannot exceed the total number of pages */ | |
a5516438 AK |
527 | if (delta > 0 && h->surplus_huge_pages_node[nid] >= |
528 | h->nr_huge_pages_node[nid]) | |
7893d1d5 AL |
529 | continue; |
530 | ||
a5516438 AK |
531 | h->surplus_huge_pages += delta; |
532 | h->surplus_huge_pages_node[nid] += delta; | |
7893d1d5 AL |
533 | ret = 1; |
534 | break; | |
535 | } while (nid != prev_nid); | |
536 | ||
537 | prev_nid = nid; | |
538 | return ret; | |
539 | } | |
540 | ||
a5516438 | 541 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 AK |
542 | { |
543 | set_compound_page_dtor(page, free_huge_page); | |
544 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
545 | h->nr_huge_pages++; |
546 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 AK |
547 | spin_unlock(&hugetlb_lock); |
548 | put_page(page); /* free it into the hugepage allocator */ | |
549 | } | |
550 | ||
a5516438 | 551 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
1da177e4 | 552 | { |
1da177e4 | 553 | struct page *page; |
f96efd58 | 554 | |
aa888a74 AK |
555 | if (h->order >= MAX_ORDER) |
556 | return NULL; | |
557 | ||
63b4613c | 558 | page = alloc_pages_node(nid, |
551883ae NA |
559 | htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| |
560 | __GFP_REPEAT|__GFP_NOWARN, | |
a5516438 | 561 | huge_page_order(h)); |
1da177e4 | 562 | if (page) { |
7f2e9525 GS |
563 | if (arch_prepare_hugepage(page)) { |
564 | __free_pages(page, HUGETLB_PAGE_ORDER); | |
7b8ee84d | 565 | return NULL; |
7f2e9525 | 566 | } |
a5516438 | 567 | prep_new_huge_page(h, page, nid); |
1da177e4 | 568 | } |
63b4613c NA |
569 | |
570 | return page; | |
571 | } | |
572 | ||
5ced66c9 AK |
573 | /* |
574 | * Use a helper variable to find the next node and then | |
575 | * copy it back to hugetlb_next_nid afterwards: | |
576 | * otherwise there's a window in which a racer might | |
577 | * pass invalid nid MAX_NUMNODES to alloc_pages_node. | |
578 | * But we don't need to use a spin_lock here: it really | |
579 | * doesn't matter if occasionally a racer chooses the | |
580 | * same nid as we do. Move nid forward in the mask even | |
581 | * if we just successfully allocated a hugepage so that | |
582 | * the next caller gets hugepages on the next node. | |
583 | */ | |
584 | static int hstate_next_node(struct hstate *h) | |
585 | { | |
586 | int next_nid; | |
587 | next_nid = next_node(h->hugetlb_next_nid, node_online_map); | |
588 | if (next_nid == MAX_NUMNODES) | |
589 | next_nid = first_node(node_online_map); | |
590 | h->hugetlb_next_nid = next_nid; | |
591 | return next_nid; | |
592 | } | |
593 | ||
a5516438 | 594 | static int alloc_fresh_huge_page(struct hstate *h) |
63b4613c NA |
595 | { |
596 | struct page *page; | |
597 | int start_nid; | |
598 | int next_nid; | |
599 | int ret = 0; | |
600 | ||
a5516438 | 601 | start_nid = h->hugetlb_next_nid; |
63b4613c NA |
602 | |
603 | do { | |
a5516438 | 604 | page = alloc_fresh_huge_page_node(h, h->hugetlb_next_nid); |
63b4613c NA |
605 | if (page) |
606 | ret = 1; | |
5ced66c9 | 607 | next_nid = hstate_next_node(h); |
a5516438 | 608 | } while (!page && h->hugetlb_next_nid != start_nid); |
63b4613c | 609 | |
3b116300 AL |
610 | if (ret) |
611 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
612 | else | |
613 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
614 | ||
63b4613c | 615 | return ret; |
1da177e4 LT |
616 | } |
617 | ||
a5516438 AK |
618 | static struct page *alloc_buddy_huge_page(struct hstate *h, |
619 | struct vm_area_struct *vma, unsigned long address) | |
7893d1d5 AL |
620 | { |
621 | struct page *page; | |
d1c3fb1f | 622 | unsigned int nid; |
7893d1d5 | 623 | |
aa888a74 AK |
624 | if (h->order >= MAX_ORDER) |
625 | return NULL; | |
626 | ||
d1c3fb1f NA |
627 | /* |
628 | * Assume we will successfully allocate the surplus page to | |
629 | * prevent racing processes from causing the surplus to exceed | |
630 | * overcommit | |
631 | * | |
632 | * This however introduces a different race, where a process B | |
633 | * tries to grow the static hugepage pool while alloc_pages() is | |
634 | * called by process A. B will only examine the per-node | |
635 | * counters in determining if surplus huge pages can be | |
636 | * converted to normal huge pages in adjust_pool_surplus(). A | |
637 | * won't be able to increment the per-node counter, until the | |
638 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
639 | * no more huge pages can be converted from surplus to normal | |
640 | * state (and doesn't try to convert again). Thus, we have a | |
641 | * case where a surplus huge page exists, the pool is grown, and | |
642 | * the surplus huge page still exists after, even though it | |
643 | * should just have been converted to a normal huge page. This | |
644 | * does not leak memory, though, as the hugepage will be freed | |
645 | * once it is out of use. It also does not allow the counters to | |
646 | * go out of whack in adjust_pool_surplus() as we don't modify | |
647 | * the node values until we've gotten the hugepage and only the | |
648 | * per-node value is checked there. | |
649 | */ | |
650 | spin_lock(&hugetlb_lock); | |
a5516438 | 651 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
652 | spin_unlock(&hugetlb_lock); |
653 | return NULL; | |
654 | } else { | |
a5516438 AK |
655 | h->nr_huge_pages++; |
656 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
657 | } |
658 | spin_unlock(&hugetlb_lock); | |
659 | ||
551883ae NA |
660 | page = alloc_pages(htlb_alloc_mask|__GFP_COMP| |
661 | __GFP_REPEAT|__GFP_NOWARN, | |
a5516438 | 662 | huge_page_order(h)); |
d1c3fb1f NA |
663 | |
664 | spin_lock(&hugetlb_lock); | |
7893d1d5 | 665 | if (page) { |
2668db91 AL |
666 | /* |
667 | * This page is now managed by the hugetlb allocator and has | |
668 | * no users -- drop the buddy allocator's reference. | |
669 | */ | |
670 | put_page_testzero(page); | |
671 | VM_BUG_ON(page_count(page)); | |
d1c3fb1f | 672 | nid = page_to_nid(page); |
7893d1d5 | 673 | set_compound_page_dtor(page, free_huge_page); |
d1c3fb1f NA |
674 | /* |
675 | * We incremented the global counters already | |
676 | */ | |
a5516438 AK |
677 | h->nr_huge_pages_node[nid]++; |
678 | h->surplus_huge_pages_node[nid]++; | |
3b116300 | 679 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 680 | } else { |
a5516438 AK |
681 | h->nr_huge_pages--; |
682 | h->surplus_huge_pages--; | |
3b116300 | 683 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 684 | } |
d1c3fb1f | 685 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
686 | |
687 | return page; | |
688 | } | |
689 | ||
e4e574b7 AL |
690 | /* |
691 | * Increase the hugetlb pool such that it can accomodate a reservation | |
692 | * of size 'delta'. | |
693 | */ | |
a5516438 | 694 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
695 | { |
696 | struct list_head surplus_list; | |
697 | struct page *page, *tmp; | |
698 | int ret, i; | |
699 | int needed, allocated; | |
700 | ||
a5516438 | 701 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 702 | if (needed <= 0) { |
a5516438 | 703 | h->resv_huge_pages += delta; |
e4e574b7 | 704 | return 0; |
ac09b3a1 | 705 | } |
e4e574b7 AL |
706 | |
707 | allocated = 0; | |
708 | INIT_LIST_HEAD(&surplus_list); | |
709 | ||
710 | ret = -ENOMEM; | |
711 | retry: | |
712 | spin_unlock(&hugetlb_lock); | |
713 | for (i = 0; i < needed; i++) { | |
a5516438 | 714 | page = alloc_buddy_huge_page(h, NULL, 0); |
e4e574b7 AL |
715 | if (!page) { |
716 | /* | |
717 | * We were not able to allocate enough pages to | |
718 | * satisfy the entire reservation so we free what | |
719 | * we've allocated so far. | |
720 | */ | |
721 | spin_lock(&hugetlb_lock); | |
722 | needed = 0; | |
723 | goto free; | |
724 | } | |
725 | ||
726 | list_add(&page->lru, &surplus_list); | |
727 | } | |
728 | allocated += needed; | |
729 | ||
730 | /* | |
731 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
732 | * because either resv_huge_pages or free_huge_pages may have changed. | |
733 | */ | |
734 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
735 | needed = (h->resv_huge_pages + delta) - |
736 | (h->free_huge_pages + allocated); | |
e4e574b7 AL |
737 | if (needed > 0) |
738 | goto retry; | |
739 | ||
740 | /* | |
741 | * The surplus_list now contains _at_least_ the number of extra pages | |
742 | * needed to accomodate the reservation. Add the appropriate number | |
743 | * of pages to the hugetlb pool and free the extras back to the buddy | |
ac09b3a1 AL |
744 | * allocator. Commit the entire reservation here to prevent another |
745 | * process from stealing the pages as they are added to the pool but | |
746 | * before they are reserved. | |
e4e574b7 AL |
747 | */ |
748 | needed += allocated; | |
a5516438 | 749 | h->resv_huge_pages += delta; |
e4e574b7 AL |
750 | ret = 0; |
751 | free: | |
19fc3f0a | 752 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 753 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
754 | if ((--needed) < 0) |
755 | break; | |
e4e574b7 | 756 | list_del(&page->lru); |
a5516438 | 757 | enqueue_huge_page(h, page); |
19fc3f0a AL |
758 | } |
759 | ||
760 | /* Free unnecessary surplus pages to the buddy allocator */ | |
761 | if (!list_empty(&surplus_list)) { | |
762 | spin_unlock(&hugetlb_lock); | |
763 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { | |
764 | list_del(&page->lru); | |
af767cbd | 765 | /* |
2668db91 AL |
766 | * The page has a reference count of zero already, so |
767 | * call free_huge_page directly instead of using | |
768 | * put_page. This must be done with hugetlb_lock | |
af767cbd AL |
769 | * unlocked which is safe because free_huge_page takes |
770 | * hugetlb_lock before deciding how to free the page. | |
771 | */ | |
2668db91 | 772 | free_huge_page(page); |
af767cbd | 773 | } |
19fc3f0a | 774 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
775 | } |
776 | ||
777 | return ret; | |
778 | } | |
779 | ||
780 | /* | |
781 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
782 | * allocated to satisfy the reservation must be explicitly freed if they were | |
783 | * never used. | |
784 | */ | |
a5516438 AK |
785 | static void return_unused_surplus_pages(struct hstate *h, |
786 | unsigned long unused_resv_pages) | |
e4e574b7 AL |
787 | { |
788 | static int nid = -1; | |
789 | struct page *page; | |
790 | unsigned long nr_pages; | |
791 | ||
11320d17 NA |
792 | /* |
793 | * We want to release as many surplus pages as possible, spread | |
794 | * evenly across all nodes. Iterate across all nodes until we | |
795 | * can no longer free unreserved surplus pages. This occurs when | |
796 | * the nodes with surplus pages have no free pages. | |
797 | */ | |
798 | unsigned long remaining_iterations = num_online_nodes(); | |
799 | ||
ac09b3a1 | 800 | /* Uncommit the reservation */ |
a5516438 | 801 | h->resv_huge_pages -= unused_resv_pages; |
ac09b3a1 | 802 | |
aa888a74 AK |
803 | /* Cannot return gigantic pages currently */ |
804 | if (h->order >= MAX_ORDER) | |
805 | return; | |
806 | ||
a5516438 | 807 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 808 | |
11320d17 | 809 | while (remaining_iterations-- && nr_pages) { |
e4e574b7 AL |
810 | nid = next_node(nid, node_online_map); |
811 | if (nid == MAX_NUMNODES) | |
812 | nid = first_node(node_online_map); | |
813 | ||
a5516438 | 814 | if (!h->surplus_huge_pages_node[nid]) |
e4e574b7 AL |
815 | continue; |
816 | ||
a5516438 AK |
817 | if (!list_empty(&h->hugepage_freelists[nid])) { |
818 | page = list_entry(h->hugepage_freelists[nid].next, | |
e4e574b7 AL |
819 | struct page, lru); |
820 | list_del(&page->lru); | |
a5516438 AK |
821 | update_and_free_page(h, page); |
822 | h->free_huge_pages--; | |
823 | h->free_huge_pages_node[nid]--; | |
824 | h->surplus_huge_pages--; | |
825 | h->surplus_huge_pages_node[nid]--; | |
e4e574b7 | 826 | nr_pages--; |
11320d17 | 827 | remaining_iterations = num_online_nodes(); |
e4e574b7 AL |
828 | } |
829 | } | |
830 | } | |
831 | ||
c37f9fb1 AW |
832 | /* |
833 | * Determine if the huge page at addr within the vma has an associated | |
834 | * reservation. Where it does not we will need to logically increase | |
835 | * reservation and actually increase quota before an allocation can occur. | |
836 | * Where any new reservation would be required the reservation change is | |
837 | * prepared, but not committed. Once the page has been quota'd allocated | |
838 | * an instantiated the change should be committed via vma_commit_reservation. | |
839 | * No action is required on failure. | |
840 | */ | |
a5516438 AK |
841 | static int vma_needs_reservation(struct hstate *h, |
842 | struct vm_area_struct *vma, unsigned long addr) | |
c37f9fb1 AW |
843 | { |
844 | struct address_space *mapping = vma->vm_file->f_mapping; | |
845 | struct inode *inode = mapping->host; | |
846 | ||
847 | if (vma->vm_flags & VM_SHARED) { | |
a5516438 | 848 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
c37f9fb1 AW |
849 | return region_chg(&inode->i_mapping->private_list, |
850 | idx, idx + 1); | |
851 | ||
84afd99b AW |
852 | } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
853 | return 1; | |
c37f9fb1 | 854 | |
84afd99b AW |
855 | } else { |
856 | int err; | |
a5516438 | 857 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
84afd99b AW |
858 | struct resv_map *reservations = vma_resv_map(vma); |
859 | ||
860 | err = region_chg(&reservations->regions, idx, idx + 1); | |
861 | if (err < 0) | |
862 | return err; | |
863 | return 0; | |
864 | } | |
c37f9fb1 | 865 | } |
a5516438 AK |
866 | static void vma_commit_reservation(struct hstate *h, |
867 | struct vm_area_struct *vma, unsigned long addr) | |
c37f9fb1 AW |
868 | { |
869 | struct address_space *mapping = vma->vm_file->f_mapping; | |
870 | struct inode *inode = mapping->host; | |
871 | ||
872 | if (vma->vm_flags & VM_SHARED) { | |
a5516438 | 873 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
c37f9fb1 | 874 | region_add(&inode->i_mapping->private_list, idx, idx + 1); |
84afd99b AW |
875 | |
876 | } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { | |
a5516438 | 877 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
84afd99b AW |
878 | struct resv_map *reservations = vma_resv_map(vma); |
879 | ||
880 | /* Mark this page used in the map. */ | |
881 | region_add(&reservations->regions, idx, idx + 1); | |
c37f9fb1 AW |
882 | } |
883 | } | |
884 | ||
a1e78772 | 885 | static struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 886 | unsigned long addr, int avoid_reserve) |
1da177e4 | 887 | { |
a5516438 | 888 | struct hstate *h = hstate_vma(vma); |
348ea204 | 889 | struct page *page; |
a1e78772 MG |
890 | struct address_space *mapping = vma->vm_file->f_mapping; |
891 | struct inode *inode = mapping->host; | |
c37f9fb1 | 892 | unsigned int chg; |
a1e78772 MG |
893 | |
894 | /* | |
895 | * Processes that did not create the mapping will have no reserves and | |
896 | * will not have accounted against quota. Check that the quota can be | |
897 | * made before satisfying the allocation | |
c37f9fb1 AW |
898 | * MAP_NORESERVE mappings may also need pages and quota allocated |
899 | * if no reserve mapping overlaps. | |
a1e78772 | 900 | */ |
a5516438 | 901 | chg = vma_needs_reservation(h, vma, addr); |
c37f9fb1 AW |
902 | if (chg < 0) |
903 | return ERR_PTR(chg); | |
904 | if (chg) | |
a1e78772 MG |
905 | if (hugetlb_get_quota(inode->i_mapping, chg)) |
906 | return ERR_PTR(-ENOSPC); | |
1da177e4 LT |
907 | |
908 | spin_lock(&hugetlb_lock); | |
a5516438 | 909 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve); |
1da177e4 | 910 | spin_unlock(&hugetlb_lock); |
b45b5bd6 | 911 | |
68842c9b | 912 | if (!page) { |
a5516438 | 913 | page = alloc_buddy_huge_page(h, vma, addr); |
68842c9b | 914 | if (!page) { |
a1e78772 | 915 | hugetlb_put_quota(inode->i_mapping, chg); |
68842c9b KC |
916 | return ERR_PTR(-VM_FAULT_OOM); |
917 | } | |
918 | } | |
348ea204 | 919 | |
a1e78772 MG |
920 | set_page_refcounted(page); |
921 | set_page_private(page, (unsigned long) mapping); | |
90d8b7e6 | 922 | |
a5516438 | 923 | vma_commit_reservation(h, vma, addr); |
c37f9fb1 | 924 | |
90d8b7e6 | 925 | return page; |
b45b5bd6 DG |
926 | } |
927 | ||
aa888a74 AK |
928 | static __initdata LIST_HEAD(huge_boot_pages); |
929 | ||
930 | struct huge_bootmem_page { | |
931 | struct list_head list; | |
932 | struct hstate *hstate; | |
933 | }; | |
934 | ||
935 | static int __init alloc_bootmem_huge_page(struct hstate *h) | |
936 | { | |
937 | struct huge_bootmem_page *m; | |
938 | int nr_nodes = nodes_weight(node_online_map); | |
939 | ||
940 | while (nr_nodes) { | |
941 | void *addr; | |
942 | ||
943 | addr = __alloc_bootmem_node_nopanic( | |
944 | NODE_DATA(h->hugetlb_next_nid), | |
945 | huge_page_size(h), huge_page_size(h), 0); | |
946 | ||
947 | if (addr) { | |
948 | /* | |
949 | * Use the beginning of the huge page to store the | |
950 | * huge_bootmem_page struct (until gather_bootmem | |
951 | * puts them into the mem_map). | |
952 | */ | |
953 | m = addr; | |
954 | if (m) | |
955 | goto found; | |
956 | } | |
957 | hstate_next_node(h); | |
958 | nr_nodes--; | |
959 | } | |
960 | return 0; | |
961 | ||
962 | found: | |
963 | BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); | |
964 | /* Put them into a private list first because mem_map is not up yet */ | |
965 | list_add(&m->list, &huge_boot_pages); | |
966 | m->hstate = h; | |
967 | return 1; | |
968 | } | |
969 | ||
970 | /* Put bootmem huge pages into the standard lists after mem_map is up */ | |
971 | static void __init gather_bootmem_prealloc(void) | |
972 | { | |
973 | struct huge_bootmem_page *m; | |
974 | ||
975 | list_for_each_entry(m, &huge_boot_pages, list) { | |
976 | struct page *page = virt_to_page(m); | |
977 | struct hstate *h = m->hstate; | |
978 | __ClearPageReserved(page); | |
979 | WARN_ON(page_count(page) != 1); | |
980 | prep_compound_page(page, h->order); | |
981 | prep_new_huge_page(h, page, page_to_nid(page)); | |
982 | } | |
983 | } | |
984 | ||
8faa8b07 | 985 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
986 | { |
987 | unsigned long i; | |
a5516438 | 988 | |
e5ff2159 | 989 | for (i = 0; i < h->max_huge_pages; ++i) { |
aa888a74 AK |
990 | if (h->order >= MAX_ORDER) { |
991 | if (!alloc_bootmem_huge_page(h)) | |
992 | break; | |
993 | } else if (!alloc_fresh_huge_page(h)) | |
1da177e4 | 994 | break; |
1da177e4 | 995 | } |
8faa8b07 | 996 | h->max_huge_pages = i; |
e5ff2159 AK |
997 | } |
998 | ||
999 | static void __init hugetlb_init_hstates(void) | |
1000 | { | |
1001 | struct hstate *h; | |
1002 | ||
1003 | for_each_hstate(h) { | |
8faa8b07 AK |
1004 | /* oversize hugepages were init'ed in early boot */ |
1005 | if (h->order < MAX_ORDER) | |
1006 | hugetlb_hstate_alloc_pages(h); | |
e5ff2159 AK |
1007 | } |
1008 | } | |
1009 | ||
4abd32db AK |
1010 | static char * __init memfmt(char *buf, unsigned long n) |
1011 | { | |
1012 | if (n >= (1UL << 30)) | |
1013 | sprintf(buf, "%lu GB", n >> 30); | |
1014 | else if (n >= (1UL << 20)) | |
1015 | sprintf(buf, "%lu MB", n >> 20); | |
1016 | else | |
1017 | sprintf(buf, "%lu KB", n >> 10); | |
1018 | return buf; | |
1019 | } | |
1020 | ||
e5ff2159 AK |
1021 | static void __init report_hugepages(void) |
1022 | { | |
1023 | struct hstate *h; | |
1024 | ||
1025 | for_each_hstate(h) { | |
4abd32db AK |
1026 | char buf[32]; |
1027 | printk(KERN_INFO "HugeTLB registered %s page size, " | |
1028 | "pre-allocated %ld pages\n", | |
1029 | memfmt(buf, huge_page_size(h)), | |
1030 | h->free_huge_pages); | |
e5ff2159 AK |
1031 | } |
1032 | } | |
1033 | ||
1da177e4 | 1034 | #ifdef CONFIG_SYSCTL |
1da177e4 | 1035 | #ifdef CONFIG_HIGHMEM |
a5516438 | 1036 | static void try_to_free_low(struct hstate *h, unsigned long count) |
1da177e4 | 1037 | { |
4415cc8d CL |
1038 | int i; |
1039 | ||
aa888a74 AK |
1040 | if (h->order >= MAX_ORDER) |
1041 | return; | |
1042 | ||
1da177e4 LT |
1043 | for (i = 0; i < MAX_NUMNODES; ++i) { |
1044 | struct page *page, *next; | |
a5516438 AK |
1045 | struct list_head *freel = &h->hugepage_freelists[i]; |
1046 | list_for_each_entry_safe(page, next, freel, lru) { | |
1047 | if (count >= h->nr_huge_pages) | |
6b0c880d | 1048 | return; |
1da177e4 LT |
1049 | if (PageHighMem(page)) |
1050 | continue; | |
1051 | list_del(&page->lru); | |
e5ff2159 | 1052 | update_and_free_page(h, page); |
a5516438 AK |
1053 | h->free_huge_pages--; |
1054 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
1055 | } |
1056 | } | |
1057 | } | |
1058 | #else | |
a5516438 | 1059 | static inline void try_to_free_low(struct hstate *h, unsigned long count) |
1da177e4 LT |
1060 | { |
1061 | } | |
1062 | #endif | |
1063 | ||
a5516438 | 1064 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
e5ff2159 | 1065 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count) |
1da177e4 | 1066 | { |
7893d1d5 | 1067 | unsigned long min_count, ret; |
1da177e4 | 1068 | |
aa888a74 AK |
1069 | if (h->order >= MAX_ORDER) |
1070 | return h->max_huge_pages; | |
1071 | ||
7893d1d5 AL |
1072 | /* |
1073 | * Increase the pool size | |
1074 | * First take pages out of surplus state. Then make up the | |
1075 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f NA |
1076 | * |
1077 | * We might race with alloc_buddy_huge_page() here and be unable | |
1078 | * to convert a surplus huge page to a normal huge page. That is | |
1079 | * not critical, though, it just means the overall size of the | |
1080 | * pool might be one hugepage larger than it needs to be, but | |
1081 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 1082 | */ |
1da177e4 | 1083 | spin_lock(&hugetlb_lock); |
a5516438 AK |
1084 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
1085 | if (!adjust_pool_surplus(h, -1)) | |
7893d1d5 AL |
1086 | break; |
1087 | } | |
1088 | ||
a5516438 | 1089 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
1090 | /* |
1091 | * If this allocation races such that we no longer need the | |
1092 | * page, free_huge_page will handle it by freeing the page | |
1093 | * and reducing the surplus. | |
1094 | */ | |
1095 | spin_unlock(&hugetlb_lock); | |
a5516438 | 1096 | ret = alloc_fresh_huge_page(h); |
7893d1d5 AL |
1097 | spin_lock(&hugetlb_lock); |
1098 | if (!ret) | |
1099 | goto out; | |
1100 | ||
1101 | } | |
7893d1d5 AL |
1102 | |
1103 | /* | |
1104 | * Decrease the pool size | |
1105 | * First return free pages to the buddy allocator (being careful | |
1106 | * to keep enough around to satisfy reservations). Then place | |
1107 | * pages into surplus state as needed so the pool will shrink | |
1108 | * to the desired size as pages become free. | |
d1c3fb1f NA |
1109 | * |
1110 | * By placing pages into the surplus state independent of the | |
1111 | * overcommit value, we are allowing the surplus pool size to | |
1112 | * exceed overcommit. There are few sane options here. Since | |
1113 | * alloc_buddy_huge_page() is checking the global counter, | |
1114 | * though, we'll note that we're not allowed to exceed surplus | |
1115 | * and won't grow the pool anywhere else. Not until one of the | |
1116 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 1117 | */ |
a5516438 | 1118 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 1119 | min_count = max(count, min_count); |
a5516438 AK |
1120 | try_to_free_low(h, min_count); |
1121 | while (min_count < persistent_huge_pages(h)) { | |
1122 | struct page *page = dequeue_huge_page(h); | |
1da177e4 LT |
1123 | if (!page) |
1124 | break; | |
a5516438 | 1125 | update_and_free_page(h, page); |
1da177e4 | 1126 | } |
a5516438 AK |
1127 | while (count < persistent_huge_pages(h)) { |
1128 | if (!adjust_pool_surplus(h, 1)) | |
7893d1d5 AL |
1129 | break; |
1130 | } | |
1131 | out: | |
a5516438 | 1132 | ret = persistent_huge_pages(h); |
1da177e4 | 1133 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 1134 | return ret; |
1da177e4 LT |
1135 | } |
1136 | ||
a3437870 NA |
1137 | #define HSTATE_ATTR_RO(_name) \ |
1138 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
1139 | ||
1140 | #define HSTATE_ATTR(_name) \ | |
1141 | static struct kobj_attribute _name##_attr = \ | |
1142 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
1143 | ||
1144 | static struct kobject *hugepages_kobj; | |
1145 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1146 | ||
1147 | static struct hstate *kobj_to_hstate(struct kobject *kobj) | |
1148 | { | |
1149 | int i; | |
1150 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
1151 | if (hstate_kobjs[i] == kobj) | |
1152 | return &hstates[i]; | |
1153 | BUG(); | |
1154 | return NULL; | |
1155 | } | |
1156 | ||
1157 | static ssize_t nr_hugepages_show(struct kobject *kobj, | |
1158 | struct kobj_attribute *attr, char *buf) | |
1159 | { | |
1160 | struct hstate *h = kobj_to_hstate(kobj); | |
1161 | return sprintf(buf, "%lu\n", h->nr_huge_pages); | |
1162 | } | |
1163 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
1164 | struct kobj_attribute *attr, const char *buf, size_t count) | |
1165 | { | |
1166 | int err; | |
1167 | unsigned long input; | |
1168 | struct hstate *h = kobj_to_hstate(kobj); | |
1169 | ||
1170 | err = strict_strtoul(buf, 10, &input); | |
1171 | if (err) | |
1172 | return 0; | |
1173 | ||
1174 | h->max_huge_pages = set_max_huge_pages(h, input); | |
1175 | ||
1176 | return count; | |
1177 | } | |
1178 | HSTATE_ATTR(nr_hugepages); | |
1179 | ||
1180 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, | |
1181 | struct kobj_attribute *attr, char *buf) | |
1182 | { | |
1183 | struct hstate *h = kobj_to_hstate(kobj); | |
1184 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); | |
1185 | } | |
1186 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, | |
1187 | struct kobj_attribute *attr, const char *buf, size_t count) | |
1188 | { | |
1189 | int err; | |
1190 | unsigned long input; | |
1191 | struct hstate *h = kobj_to_hstate(kobj); | |
1192 | ||
1193 | err = strict_strtoul(buf, 10, &input); | |
1194 | if (err) | |
1195 | return 0; | |
1196 | ||
1197 | spin_lock(&hugetlb_lock); | |
1198 | h->nr_overcommit_huge_pages = input; | |
1199 | spin_unlock(&hugetlb_lock); | |
1200 | ||
1201 | return count; | |
1202 | } | |
1203 | HSTATE_ATTR(nr_overcommit_hugepages); | |
1204 | ||
1205 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
1206 | struct kobj_attribute *attr, char *buf) | |
1207 | { | |
1208 | struct hstate *h = kobj_to_hstate(kobj); | |
1209 | return sprintf(buf, "%lu\n", h->free_huge_pages); | |
1210 | } | |
1211 | HSTATE_ATTR_RO(free_hugepages); | |
1212 | ||
1213 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
1214 | struct kobj_attribute *attr, char *buf) | |
1215 | { | |
1216 | struct hstate *h = kobj_to_hstate(kobj); | |
1217 | return sprintf(buf, "%lu\n", h->resv_huge_pages); | |
1218 | } | |
1219 | HSTATE_ATTR_RO(resv_hugepages); | |
1220 | ||
1221 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
1222 | struct kobj_attribute *attr, char *buf) | |
1223 | { | |
1224 | struct hstate *h = kobj_to_hstate(kobj); | |
1225 | return sprintf(buf, "%lu\n", h->surplus_huge_pages); | |
1226 | } | |
1227 | HSTATE_ATTR_RO(surplus_hugepages); | |
1228 | ||
1229 | static struct attribute *hstate_attrs[] = { | |
1230 | &nr_hugepages_attr.attr, | |
1231 | &nr_overcommit_hugepages_attr.attr, | |
1232 | &free_hugepages_attr.attr, | |
1233 | &resv_hugepages_attr.attr, | |
1234 | &surplus_hugepages_attr.attr, | |
1235 | NULL, | |
1236 | }; | |
1237 | ||
1238 | static struct attribute_group hstate_attr_group = { | |
1239 | .attrs = hstate_attrs, | |
1240 | }; | |
1241 | ||
1242 | static int __init hugetlb_sysfs_add_hstate(struct hstate *h) | |
1243 | { | |
1244 | int retval; | |
1245 | ||
1246 | hstate_kobjs[h - hstates] = kobject_create_and_add(h->name, | |
1247 | hugepages_kobj); | |
1248 | if (!hstate_kobjs[h - hstates]) | |
1249 | return -ENOMEM; | |
1250 | ||
1251 | retval = sysfs_create_group(hstate_kobjs[h - hstates], | |
1252 | &hstate_attr_group); | |
1253 | if (retval) | |
1254 | kobject_put(hstate_kobjs[h - hstates]); | |
1255 | ||
1256 | return retval; | |
1257 | } | |
1258 | ||
1259 | static void __init hugetlb_sysfs_init(void) | |
1260 | { | |
1261 | struct hstate *h; | |
1262 | int err; | |
1263 | ||
1264 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
1265 | if (!hugepages_kobj) | |
1266 | return; | |
1267 | ||
1268 | for_each_hstate(h) { | |
1269 | err = hugetlb_sysfs_add_hstate(h); | |
1270 | if (err) | |
1271 | printk(KERN_ERR "Hugetlb: Unable to add hstate %s", | |
1272 | h->name); | |
1273 | } | |
1274 | } | |
1275 | ||
1276 | static void __exit hugetlb_exit(void) | |
1277 | { | |
1278 | struct hstate *h; | |
1279 | ||
1280 | for_each_hstate(h) { | |
1281 | kobject_put(hstate_kobjs[h - hstates]); | |
1282 | } | |
1283 | ||
1284 | kobject_put(hugepages_kobj); | |
1285 | } | |
1286 | module_exit(hugetlb_exit); | |
1287 | ||
1288 | static int __init hugetlb_init(void) | |
1289 | { | |
1290 | BUILD_BUG_ON(HPAGE_SHIFT == 0); | |
1291 | ||
e11bfbfc NP |
1292 | if (!size_to_hstate(default_hstate_size)) { |
1293 | default_hstate_size = HPAGE_SIZE; | |
1294 | if (!size_to_hstate(default_hstate_size)) | |
1295 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 1296 | } |
e11bfbfc NP |
1297 | default_hstate_idx = size_to_hstate(default_hstate_size) - hstates; |
1298 | if (default_hstate_max_huge_pages) | |
1299 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
a3437870 NA |
1300 | |
1301 | hugetlb_init_hstates(); | |
1302 | ||
aa888a74 AK |
1303 | gather_bootmem_prealloc(); |
1304 | ||
a3437870 NA |
1305 | report_hugepages(); |
1306 | ||
1307 | hugetlb_sysfs_init(); | |
1308 | ||
1309 | return 0; | |
1310 | } | |
1311 | module_init(hugetlb_init); | |
1312 | ||
1313 | /* Should be called on processing a hugepagesz=... option */ | |
1314 | void __init hugetlb_add_hstate(unsigned order) | |
1315 | { | |
1316 | struct hstate *h; | |
8faa8b07 AK |
1317 | unsigned long i; |
1318 | ||
a3437870 NA |
1319 | if (size_to_hstate(PAGE_SIZE << order)) { |
1320 | printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n"); | |
1321 | return; | |
1322 | } | |
1323 | BUG_ON(max_hstate >= HUGE_MAX_HSTATE); | |
1324 | BUG_ON(order == 0); | |
1325 | h = &hstates[max_hstate++]; | |
1326 | h->order = order; | |
1327 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
1328 | h->nr_huge_pages = 0; |
1329 | h->free_huge_pages = 0; | |
1330 | for (i = 0; i < MAX_NUMNODES; ++i) | |
1331 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
1332 | h->hugetlb_next_nid = first_node(node_online_map); | |
a3437870 NA |
1333 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
1334 | huge_page_size(h)/1024); | |
8faa8b07 | 1335 | |
a3437870 NA |
1336 | parsed_hstate = h; |
1337 | } | |
1338 | ||
e11bfbfc | 1339 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
1340 | { |
1341 | unsigned long *mhp; | |
8faa8b07 | 1342 | static unsigned long *last_mhp; |
a3437870 NA |
1343 | |
1344 | /* | |
1345 | * !max_hstate means we haven't parsed a hugepagesz= parameter yet, | |
1346 | * so this hugepages= parameter goes to the "default hstate". | |
1347 | */ | |
1348 | if (!max_hstate) | |
1349 | mhp = &default_hstate_max_huge_pages; | |
1350 | else | |
1351 | mhp = &parsed_hstate->max_huge_pages; | |
1352 | ||
8faa8b07 AK |
1353 | if (mhp == last_mhp) { |
1354 | printk(KERN_WARNING "hugepages= specified twice without " | |
1355 | "interleaving hugepagesz=, ignoring\n"); | |
1356 | return 1; | |
1357 | } | |
1358 | ||
a3437870 NA |
1359 | if (sscanf(s, "%lu", mhp) <= 0) |
1360 | *mhp = 0; | |
1361 | ||
8faa8b07 AK |
1362 | /* |
1363 | * Global state is always initialized later in hugetlb_init. | |
1364 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
1365 | * use the bootmem allocator. | |
1366 | */ | |
1367 | if (max_hstate && parsed_hstate->order >= MAX_ORDER) | |
1368 | hugetlb_hstate_alloc_pages(parsed_hstate); | |
1369 | ||
1370 | last_mhp = mhp; | |
1371 | ||
a3437870 NA |
1372 | return 1; |
1373 | } | |
e11bfbfc NP |
1374 | __setup("hugepages=", hugetlb_nrpages_setup); |
1375 | ||
1376 | static int __init hugetlb_default_setup(char *s) | |
1377 | { | |
1378 | default_hstate_size = memparse(s, &s); | |
1379 | return 1; | |
1380 | } | |
1381 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 NA |
1382 | |
1383 | static unsigned int cpuset_mems_nr(unsigned int *array) | |
1384 | { | |
1385 | int node; | |
1386 | unsigned int nr = 0; | |
1387 | ||
1388 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
1389 | nr += array[node]; | |
1390 | ||
1391 | return nr; | |
1392 | } | |
1393 | ||
1da177e4 LT |
1394 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
1395 | struct file *file, void __user *buffer, | |
1396 | size_t *length, loff_t *ppos) | |
1397 | { | |
e5ff2159 AK |
1398 | struct hstate *h = &default_hstate; |
1399 | unsigned long tmp; | |
1400 | ||
1401 | if (!write) | |
1402 | tmp = h->max_huge_pages; | |
1403 | ||
1404 | table->data = &tmp; | |
1405 | table->maxlen = sizeof(unsigned long); | |
1da177e4 | 1406 | proc_doulongvec_minmax(table, write, file, buffer, length, ppos); |
e5ff2159 AK |
1407 | |
1408 | if (write) | |
1409 | h->max_huge_pages = set_max_huge_pages(h, tmp); | |
1410 | ||
1da177e4 LT |
1411 | return 0; |
1412 | } | |
396faf03 MG |
1413 | |
1414 | int hugetlb_treat_movable_handler(struct ctl_table *table, int write, | |
1415 | struct file *file, void __user *buffer, | |
1416 | size_t *length, loff_t *ppos) | |
1417 | { | |
1418 | proc_dointvec(table, write, file, buffer, length, ppos); | |
1419 | if (hugepages_treat_as_movable) | |
1420 | htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; | |
1421 | else | |
1422 | htlb_alloc_mask = GFP_HIGHUSER; | |
1423 | return 0; | |
1424 | } | |
1425 | ||
a3d0c6aa NA |
1426 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
1427 | struct file *file, void __user *buffer, | |
1428 | size_t *length, loff_t *ppos) | |
1429 | { | |
a5516438 | 1430 | struct hstate *h = &default_hstate; |
e5ff2159 AK |
1431 | unsigned long tmp; |
1432 | ||
1433 | if (!write) | |
1434 | tmp = h->nr_overcommit_huge_pages; | |
1435 | ||
1436 | table->data = &tmp; | |
1437 | table->maxlen = sizeof(unsigned long); | |
a3d0c6aa | 1438 | proc_doulongvec_minmax(table, write, file, buffer, length, ppos); |
e5ff2159 AK |
1439 | |
1440 | if (write) { | |
1441 | spin_lock(&hugetlb_lock); | |
1442 | h->nr_overcommit_huge_pages = tmp; | |
1443 | spin_unlock(&hugetlb_lock); | |
1444 | } | |
1445 | ||
a3d0c6aa NA |
1446 | return 0; |
1447 | } | |
1448 | ||
1da177e4 LT |
1449 | #endif /* CONFIG_SYSCTL */ |
1450 | ||
1451 | int hugetlb_report_meminfo(char *buf) | |
1452 | { | |
a5516438 | 1453 | struct hstate *h = &default_hstate; |
1da177e4 LT |
1454 | return sprintf(buf, |
1455 | "HugePages_Total: %5lu\n" | |
1456 | "HugePages_Free: %5lu\n" | |
a43a8c39 | 1457 | "HugePages_Rsvd: %5lu\n" |
7893d1d5 | 1458 | "HugePages_Surp: %5lu\n" |
1da177e4 | 1459 | "Hugepagesize: %5lu kB\n", |
a5516438 AK |
1460 | h->nr_huge_pages, |
1461 | h->free_huge_pages, | |
1462 | h->resv_huge_pages, | |
1463 | h->surplus_huge_pages, | |
1464 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
1465 | } |
1466 | ||
1467 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
1468 | { | |
a5516438 | 1469 | struct hstate *h = &default_hstate; |
1da177e4 LT |
1470 | return sprintf(buf, |
1471 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
1472 | "Node %d HugePages_Free: %5u\n" |
1473 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
1474 | nid, h->nr_huge_pages_node[nid], |
1475 | nid, h->free_huge_pages_node[nid], | |
1476 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
1477 | } |
1478 | ||
1da177e4 LT |
1479 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
1480 | unsigned long hugetlb_total_pages(void) | |
1481 | { | |
a5516438 AK |
1482 | struct hstate *h = &default_hstate; |
1483 | return h->nr_huge_pages * pages_per_huge_page(h); | |
1da177e4 | 1484 | } |
1da177e4 | 1485 | |
a5516438 | 1486 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
1487 | { |
1488 | int ret = -ENOMEM; | |
1489 | ||
1490 | spin_lock(&hugetlb_lock); | |
1491 | /* | |
1492 | * When cpuset is configured, it breaks the strict hugetlb page | |
1493 | * reservation as the accounting is done on a global variable. Such | |
1494 | * reservation is completely rubbish in the presence of cpuset because | |
1495 | * the reservation is not checked against page availability for the | |
1496 | * current cpuset. Application can still potentially OOM'ed by kernel | |
1497 | * with lack of free htlb page in cpuset that the task is in. | |
1498 | * Attempt to enforce strict accounting with cpuset is almost | |
1499 | * impossible (or too ugly) because cpuset is too fluid that | |
1500 | * task or memory node can be dynamically moved between cpusets. | |
1501 | * | |
1502 | * The change of semantics for shared hugetlb mapping with cpuset is | |
1503 | * undesirable. However, in order to preserve some of the semantics, | |
1504 | * we fall back to check against current free page availability as | |
1505 | * a best attempt and hopefully to minimize the impact of changing | |
1506 | * semantics that cpuset has. | |
1507 | */ | |
1508 | if (delta > 0) { | |
a5516438 | 1509 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
1510 | goto out; |
1511 | ||
a5516438 AK |
1512 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
1513 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
1514 | goto out; |
1515 | } | |
1516 | } | |
1517 | ||
1518 | ret = 0; | |
1519 | if (delta < 0) | |
a5516438 | 1520 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
1521 | |
1522 | out: | |
1523 | spin_unlock(&hugetlb_lock); | |
1524 | return ret; | |
1525 | } | |
1526 | ||
84afd99b AW |
1527 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
1528 | { | |
1529 | struct resv_map *reservations = vma_resv_map(vma); | |
1530 | ||
1531 | /* | |
1532 | * This new VMA should share its siblings reservation map if present. | |
1533 | * The VMA will only ever have a valid reservation map pointer where | |
1534 | * it is being copied for another still existing VMA. As that VMA | |
1535 | * has a reference to the reservation map it cannot dissappear until | |
1536 | * after this open call completes. It is therefore safe to take a | |
1537 | * new reference here without additional locking. | |
1538 | */ | |
1539 | if (reservations) | |
1540 | kref_get(&reservations->refs); | |
1541 | } | |
1542 | ||
a1e78772 MG |
1543 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
1544 | { | |
a5516438 | 1545 | struct hstate *h = hstate_vma(vma); |
84afd99b AW |
1546 | struct resv_map *reservations = vma_resv_map(vma); |
1547 | unsigned long reserve; | |
1548 | unsigned long start; | |
1549 | unsigned long end; | |
1550 | ||
1551 | if (reservations) { | |
a5516438 AK |
1552 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
1553 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b AW |
1554 | |
1555 | reserve = (end - start) - | |
1556 | region_count(&reservations->regions, start, end); | |
1557 | ||
1558 | kref_put(&reservations->refs, resv_map_release); | |
1559 | ||
1560 | if (reserve) | |
a5516438 | 1561 | hugetlb_acct_memory(h, -reserve); |
84afd99b | 1562 | } |
a1e78772 MG |
1563 | } |
1564 | ||
1da177e4 LT |
1565 | /* |
1566 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
1567 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
1568 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
1569 | * this far. | |
1570 | */ | |
d0217ac0 | 1571 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
1572 | { |
1573 | BUG(); | |
d0217ac0 | 1574 | return 0; |
1da177e4 LT |
1575 | } |
1576 | ||
1577 | struct vm_operations_struct hugetlb_vm_ops = { | |
d0217ac0 | 1578 | .fault = hugetlb_vm_op_fault, |
84afd99b | 1579 | .open = hugetlb_vm_op_open, |
a1e78772 | 1580 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
1581 | }; |
1582 | ||
1e8f889b DG |
1583 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
1584 | int writable) | |
63551ae0 DG |
1585 | { |
1586 | pte_t entry; | |
1587 | ||
1e8f889b | 1588 | if (writable) { |
63551ae0 DG |
1589 | entry = |
1590 | pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); | |
1591 | } else { | |
7f2e9525 | 1592 | entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot)); |
63551ae0 DG |
1593 | } |
1594 | entry = pte_mkyoung(entry); | |
1595 | entry = pte_mkhuge(entry); | |
1596 | ||
1597 | return entry; | |
1598 | } | |
1599 | ||
1e8f889b DG |
1600 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
1601 | unsigned long address, pte_t *ptep) | |
1602 | { | |
1603 | pte_t entry; | |
1604 | ||
7f2e9525 GS |
1605 | entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep))); |
1606 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) { | |
8dab5241 | 1607 | update_mmu_cache(vma, address, entry); |
8dab5241 | 1608 | } |
1e8f889b DG |
1609 | } |
1610 | ||
1611 | ||
63551ae0 DG |
1612 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
1613 | struct vm_area_struct *vma) | |
1614 | { | |
1615 | pte_t *src_pte, *dst_pte, entry; | |
1616 | struct page *ptepage; | |
1c59827d | 1617 | unsigned long addr; |
1e8f889b | 1618 | int cow; |
a5516438 AK |
1619 | struct hstate *h = hstate_vma(vma); |
1620 | unsigned long sz = huge_page_size(h); | |
1e8f889b DG |
1621 | |
1622 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 1623 | |
a5516438 | 1624 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
c74df32c HD |
1625 | src_pte = huge_pte_offset(src, addr); |
1626 | if (!src_pte) | |
1627 | continue; | |
a5516438 | 1628 | dst_pte = huge_pte_alloc(dst, addr, sz); |
63551ae0 DG |
1629 | if (!dst_pte) |
1630 | goto nomem; | |
c5c99429 LW |
1631 | |
1632 | /* If the pagetables are shared don't copy or take references */ | |
1633 | if (dst_pte == src_pte) | |
1634 | continue; | |
1635 | ||
c74df32c | 1636 | spin_lock(&dst->page_table_lock); |
46478758 | 1637 | spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); |
7f2e9525 | 1638 | if (!huge_pte_none(huge_ptep_get(src_pte))) { |
1e8f889b | 1639 | if (cow) |
7f2e9525 GS |
1640 | huge_ptep_set_wrprotect(src, addr, src_pte); |
1641 | entry = huge_ptep_get(src_pte); | |
1c59827d HD |
1642 | ptepage = pte_page(entry); |
1643 | get_page(ptepage); | |
1c59827d HD |
1644 | set_huge_pte_at(dst, addr, dst_pte, entry); |
1645 | } | |
1646 | spin_unlock(&src->page_table_lock); | |
c74df32c | 1647 | spin_unlock(&dst->page_table_lock); |
63551ae0 DG |
1648 | } |
1649 | return 0; | |
1650 | ||
1651 | nomem: | |
1652 | return -ENOMEM; | |
1653 | } | |
1654 | ||
502717f4 | 1655 | void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 1656 | unsigned long end, struct page *ref_page) |
63551ae0 DG |
1657 | { |
1658 | struct mm_struct *mm = vma->vm_mm; | |
1659 | unsigned long address; | |
c7546f8f | 1660 | pte_t *ptep; |
63551ae0 DG |
1661 | pte_t pte; |
1662 | struct page *page; | |
fe1668ae | 1663 | struct page *tmp; |
a5516438 AK |
1664 | struct hstate *h = hstate_vma(vma); |
1665 | unsigned long sz = huge_page_size(h); | |
1666 | ||
c0a499c2 CK |
1667 | /* |
1668 | * A page gathering list, protected by per file i_mmap_lock. The | |
1669 | * lock is used to avoid list corruption from multiple unmapping | |
1670 | * of the same page since we are using page->lru. | |
1671 | */ | |
fe1668ae | 1672 | LIST_HEAD(page_list); |
63551ae0 DG |
1673 | |
1674 | WARN_ON(!is_vm_hugetlb_page(vma)); | |
a5516438 AK |
1675 | BUG_ON(start & ~huge_page_mask(h)); |
1676 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 1677 | |
508034a3 | 1678 | spin_lock(&mm->page_table_lock); |
a5516438 | 1679 | for (address = start; address < end; address += sz) { |
c7546f8f | 1680 | ptep = huge_pte_offset(mm, address); |
4c887265 | 1681 | if (!ptep) |
c7546f8f DG |
1682 | continue; |
1683 | ||
39dde65c CK |
1684 | if (huge_pmd_unshare(mm, &address, ptep)) |
1685 | continue; | |
1686 | ||
04f2cbe3 MG |
1687 | /* |
1688 | * If a reference page is supplied, it is because a specific | |
1689 | * page is being unmapped, not a range. Ensure the page we | |
1690 | * are about to unmap is the actual page of interest. | |
1691 | */ | |
1692 | if (ref_page) { | |
1693 | pte = huge_ptep_get(ptep); | |
1694 | if (huge_pte_none(pte)) | |
1695 | continue; | |
1696 | page = pte_page(pte); | |
1697 | if (page != ref_page) | |
1698 | continue; | |
1699 | ||
1700 | /* | |
1701 | * Mark the VMA as having unmapped its page so that | |
1702 | * future faults in this VMA will fail rather than | |
1703 | * looking like data was lost | |
1704 | */ | |
1705 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
1706 | } | |
1707 | ||
c7546f8f | 1708 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
7f2e9525 | 1709 | if (huge_pte_none(pte)) |
63551ae0 | 1710 | continue; |
c7546f8f | 1711 | |
63551ae0 | 1712 | page = pte_page(pte); |
6649a386 KC |
1713 | if (pte_dirty(pte)) |
1714 | set_page_dirty(page); | |
fe1668ae | 1715 | list_add(&page->lru, &page_list); |
63551ae0 | 1716 | } |
1da177e4 | 1717 | spin_unlock(&mm->page_table_lock); |
508034a3 | 1718 | flush_tlb_range(vma, start, end); |
fe1668ae CK |
1719 | list_for_each_entry_safe(page, tmp, &page_list, lru) { |
1720 | list_del(&page->lru); | |
1721 | put_page(page); | |
1722 | } | |
1da177e4 | 1723 | } |
63551ae0 | 1724 | |
502717f4 | 1725 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 1726 | unsigned long end, struct page *ref_page) |
502717f4 | 1727 | { |
a137e1cc AK |
1728 | spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); |
1729 | __unmap_hugepage_range(vma, start, end, ref_page); | |
1730 | spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); | |
502717f4 CK |
1731 | } |
1732 | ||
04f2cbe3 MG |
1733 | /* |
1734 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
1735 | * mappping it owns the reserve page for. The intention is to unmap the page | |
1736 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
1737 | * same region. | |
1738 | */ | |
1739 | int unmap_ref_private(struct mm_struct *mm, | |
1740 | struct vm_area_struct *vma, | |
1741 | struct page *page, | |
1742 | unsigned long address) | |
1743 | { | |
1744 | struct vm_area_struct *iter_vma; | |
1745 | struct address_space *mapping; | |
1746 | struct prio_tree_iter iter; | |
1747 | pgoff_t pgoff; | |
1748 | ||
1749 | /* | |
1750 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
1751 | * from page cache lookup which is in HPAGE_SIZE units. | |
1752 | */ | |
1753 | address = address & huge_page_mask(hstate_vma(vma)); | |
1754 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) | |
1755 | + (vma->vm_pgoff >> PAGE_SHIFT); | |
1756 | mapping = (struct address_space *)page_private(page); | |
1757 | ||
1758 | vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) { | |
1759 | /* Do not unmap the current VMA */ | |
1760 | if (iter_vma == vma) | |
1761 | continue; | |
1762 | ||
1763 | /* | |
1764 | * Unmap the page from other VMAs without their own reserves. | |
1765 | * They get marked to be SIGKILLed if they fault in these | |
1766 | * areas. This is because a future no-page fault on this VMA | |
1767 | * could insert a zeroed page instead of the data existing | |
1768 | * from the time of fork. This would look like data corruption | |
1769 | */ | |
1770 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
1771 | unmap_hugepage_range(iter_vma, | |
1772 | address, address + HPAGE_SIZE, | |
1773 | page); | |
1774 | } | |
1775 | ||
1776 | return 1; | |
1777 | } | |
1778 | ||
1e8f889b | 1779 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
04f2cbe3 MG |
1780 | unsigned long address, pte_t *ptep, pte_t pte, |
1781 | struct page *pagecache_page) | |
1e8f889b | 1782 | { |
a5516438 | 1783 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 1784 | struct page *old_page, *new_page; |
79ac6ba4 | 1785 | int avoidcopy; |
04f2cbe3 | 1786 | int outside_reserve = 0; |
1e8f889b DG |
1787 | |
1788 | old_page = pte_page(pte); | |
1789 | ||
04f2cbe3 | 1790 | retry_avoidcopy: |
1e8f889b DG |
1791 | /* If no-one else is actually using this page, avoid the copy |
1792 | * and just make the page writable */ | |
1793 | avoidcopy = (page_count(old_page) == 1); | |
1794 | if (avoidcopy) { | |
1795 | set_huge_ptep_writable(vma, address, ptep); | |
83c54070 | 1796 | return 0; |
1e8f889b DG |
1797 | } |
1798 | ||
04f2cbe3 MG |
1799 | /* |
1800 | * If the process that created a MAP_PRIVATE mapping is about to | |
1801 | * perform a COW due to a shared page count, attempt to satisfy | |
1802 | * the allocation without using the existing reserves. The pagecache | |
1803 | * page is used to determine if the reserve at this address was | |
1804 | * consumed or not. If reserves were used, a partial faulted mapping | |
1805 | * at the time of fork() could consume its reserves on COW instead | |
1806 | * of the full address range. | |
1807 | */ | |
1808 | if (!(vma->vm_flags & VM_SHARED) && | |
1809 | is_vma_resv_set(vma, HPAGE_RESV_OWNER) && | |
1810 | old_page != pagecache_page) | |
1811 | outside_reserve = 1; | |
1812 | ||
1e8f889b | 1813 | page_cache_get(old_page); |
04f2cbe3 | 1814 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 1815 | |
2fc39cec | 1816 | if (IS_ERR(new_page)) { |
1e8f889b | 1817 | page_cache_release(old_page); |
04f2cbe3 MG |
1818 | |
1819 | /* | |
1820 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
1821 | * it is due to references held by a child and an insufficient | |
1822 | * huge page pool. To guarantee the original mappers | |
1823 | * reliability, unmap the page from child processes. The child | |
1824 | * may get SIGKILLed if it later faults. | |
1825 | */ | |
1826 | if (outside_reserve) { | |
1827 | BUG_ON(huge_pte_none(pte)); | |
1828 | if (unmap_ref_private(mm, vma, old_page, address)) { | |
1829 | BUG_ON(page_count(old_page) != 1); | |
1830 | BUG_ON(huge_pte_none(pte)); | |
1831 | goto retry_avoidcopy; | |
1832 | } | |
1833 | WARN_ON_ONCE(1); | |
1834 | } | |
1835 | ||
2fc39cec | 1836 | return -PTR_ERR(new_page); |
1e8f889b DG |
1837 | } |
1838 | ||
1839 | spin_unlock(&mm->page_table_lock); | |
9de455b2 | 1840 | copy_huge_page(new_page, old_page, address, vma); |
0ed361de | 1841 | __SetPageUptodate(new_page); |
1e8f889b DG |
1842 | spin_lock(&mm->page_table_lock); |
1843 | ||
a5516438 | 1844 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
7f2e9525 | 1845 | if (likely(pte_same(huge_ptep_get(ptep), pte))) { |
1e8f889b | 1846 | /* Break COW */ |
8fe627ec | 1847 | huge_ptep_clear_flush(vma, address, ptep); |
1e8f889b DG |
1848 | set_huge_pte_at(mm, address, ptep, |
1849 | make_huge_pte(vma, new_page, 1)); | |
1850 | /* Make the old page be freed below */ | |
1851 | new_page = old_page; | |
1852 | } | |
1853 | page_cache_release(new_page); | |
1854 | page_cache_release(old_page); | |
83c54070 | 1855 | return 0; |
1e8f889b DG |
1856 | } |
1857 | ||
04f2cbe3 | 1858 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
1859 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
1860 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
1861 | { |
1862 | struct address_space *mapping; | |
e7c4b0bf | 1863 | pgoff_t idx; |
04f2cbe3 MG |
1864 | |
1865 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 1866 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
1867 | |
1868 | return find_lock_page(mapping, idx); | |
1869 | } | |
1870 | ||
a1ed3dda | 1871 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
1e8f889b | 1872 | unsigned long address, pte_t *ptep, int write_access) |
ac9b9c66 | 1873 | { |
a5516438 | 1874 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 1875 | int ret = VM_FAULT_SIGBUS; |
e7c4b0bf | 1876 | pgoff_t idx; |
4c887265 | 1877 | unsigned long size; |
4c887265 AL |
1878 | struct page *page; |
1879 | struct address_space *mapping; | |
1e8f889b | 1880 | pte_t new_pte; |
4c887265 | 1881 | |
04f2cbe3 MG |
1882 | /* |
1883 | * Currently, we are forced to kill the process in the event the | |
1884 | * original mapper has unmapped pages from the child due to a failed | |
1885 | * COW. Warn that such a situation has occured as it may not be obvious | |
1886 | */ | |
1887 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
1888 | printk(KERN_WARNING | |
1889 | "PID %d killed due to inadequate hugepage pool\n", | |
1890 | current->pid); | |
1891 | return ret; | |
1892 | } | |
1893 | ||
4c887265 | 1894 | mapping = vma->vm_file->f_mapping; |
a5516438 | 1895 | idx = vma_hugecache_offset(h, vma, address); |
4c887265 AL |
1896 | |
1897 | /* | |
1898 | * Use page lock to guard against racing truncation | |
1899 | * before we get page_table_lock. | |
1900 | */ | |
6bda666a CL |
1901 | retry: |
1902 | page = find_lock_page(mapping, idx); | |
1903 | if (!page) { | |
a5516438 | 1904 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
1905 | if (idx >= size) |
1906 | goto out; | |
04f2cbe3 | 1907 | page = alloc_huge_page(vma, address, 0); |
2fc39cec AL |
1908 | if (IS_ERR(page)) { |
1909 | ret = -PTR_ERR(page); | |
6bda666a CL |
1910 | goto out; |
1911 | } | |
a5516438 | 1912 | clear_huge_page(page, address, huge_page_size(h)); |
0ed361de | 1913 | __SetPageUptodate(page); |
ac9b9c66 | 1914 | |
6bda666a CL |
1915 | if (vma->vm_flags & VM_SHARED) { |
1916 | int err; | |
45c682a6 | 1917 | struct inode *inode = mapping->host; |
6bda666a CL |
1918 | |
1919 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
1920 | if (err) { | |
1921 | put_page(page); | |
6bda666a CL |
1922 | if (err == -EEXIST) |
1923 | goto retry; | |
1924 | goto out; | |
1925 | } | |
45c682a6 KC |
1926 | |
1927 | spin_lock(&inode->i_lock); | |
a5516438 | 1928 | inode->i_blocks += blocks_per_huge_page(h); |
45c682a6 | 1929 | spin_unlock(&inode->i_lock); |
6bda666a CL |
1930 | } else |
1931 | lock_page(page); | |
1932 | } | |
1e8f889b | 1933 | |
ac9b9c66 | 1934 | spin_lock(&mm->page_table_lock); |
a5516438 | 1935 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
1936 | if (idx >= size) |
1937 | goto backout; | |
1938 | ||
83c54070 | 1939 | ret = 0; |
7f2e9525 | 1940 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
1941 | goto backout; |
1942 | ||
1e8f889b DG |
1943 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
1944 | && (vma->vm_flags & VM_SHARED))); | |
1945 | set_huge_pte_at(mm, address, ptep, new_pte); | |
1946 | ||
1947 | if (write_access && !(vma->vm_flags & VM_SHARED)) { | |
1948 | /* Optimization, do the COW without a second fault */ | |
04f2cbe3 | 1949 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); |
1e8f889b DG |
1950 | } |
1951 | ||
ac9b9c66 | 1952 | spin_unlock(&mm->page_table_lock); |
4c887265 AL |
1953 | unlock_page(page); |
1954 | out: | |
ac9b9c66 | 1955 | return ret; |
4c887265 AL |
1956 | |
1957 | backout: | |
1958 | spin_unlock(&mm->page_table_lock); | |
4c887265 AL |
1959 | unlock_page(page); |
1960 | put_page(page); | |
1961 | goto out; | |
ac9b9c66 HD |
1962 | } |
1963 | ||
86e5216f AL |
1964 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
1965 | unsigned long address, int write_access) | |
1966 | { | |
1967 | pte_t *ptep; | |
1968 | pte_t entry; | |
1e8f889b | 1969 | int ret; |
3935baa9 | 1970 | static DEFINE_MUTEX(hugetlb_instantiation_mutex); |
a5516438 | 1971 | struct hstate *h = hstate_vma(vma); |
86e5216f | 1972 | |
a5516438 | 1973 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); |
86e5216f AL |
1974 | if (!ptep) |
1975 | return VM_FAULT_OOM; | |
1976 | ||
3935baa9 DG |
1977 | /* |
1978 | * Serialize hugepage allocation and instantiation, so that we don't | |
1979 | * get spurious allocation failures if two CPUs race to instantiate | |
1980 | * the same page in the page cache. | |
1981 | */ | |
1982 | mutex_lock(&hugetlb_instantiation_mutex); | |
7f2e9525 GS |
1983 | entry = huge_ptep_get(ptep); |
1984 | if (huge_pte_none(entry)) { | |
3935baa9 DG |
1985 | ret = hugetlb_no_page(mm, vma, address, ptep, write_access); |
1986 | mutex_unlock(&hugetlb_instantiation_mutex); | |
1987 | return ret; | |
1988 | } | |
86e5216f | 1989 | |
83c54070 | 1990 | ret = 0; |
1e8f889b DG |
1991 | |
1992 | spin_lock(&mm->page_table_lock); | |
1993 | /* Check for a racing update before calling hugetlb_cow */ | |
7f2e9525 | 1994 | if (likely(pte_same(entry, huge_ptep_get(ptep)))) |
04f2cbe3 MG |
1995 | if (write_access && !pte_write(entry)) { |
1996 | struct page *page; | |
a5516438 | 1997 | page = hugetlbfs_pagecache_page(h, vma, address); |
04f2cbe3 MG |
1998 | ret = hugetlb_cow(mm, vma, address, ptep, entry, page); |
1999 | if (page) { | |
2000 | unlock_page(page); | |
2001 | put_page(page); | |
2002 | } | |
2003 | } | |
1e8f889b | 2004 | spin_unlock(&mm->page_table_lock); |
3935baa9 | 2005 | mutex_unlock(&hugetlb_instantiation_mutex); |
1e8f889b DG |
2006 | |
2007 | return ret; | |
86e5216f AL |
2008 | } |
2009 | ||
ceb86879 AK |
2010 | /* Can be overriden by architectures */ |
2011 | __attribute__((weak)) struct page * | |
2012 | follow_huge_pud(struct mm_struct *mm, unsigned long address, | |
2013 | pud_t *pud, int write) | |
2014 | { | |
2015 | BUG(); | |
2016 | return NULL; | |
2017 | } | |
2018 | ||
63551ae0 DG |
2019 | int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
2020 | struct page **pages, struct vm_area_struct **vmas, | |
5b23dbe8 AL |
2021 | unsigned long *position, int *length, int i, |
2022 | int write) | |
63551ae0 | 2023 | { |
d5d4b0aa CK |
2024 | unsigned long pfn_offset; |
2025 | unsigned long vaddr = *position; | |
63551ae0 | 2026 | int remainder = *length; |
a5516438 | 2027 | struct hstate *h = hstate_vma(vma); |
63551ae0 | 2028 | |
1c59827d | 2029 | spin_lock(&mm->page_table_lock); |
63551ae0 | 2030 | while (vaddr < vma->vm_end && remainder) { |
4c887265 AL |
2031 | pte_t *pte; |
2032 | struct page *page; | |
63551ae0 | 2033 | |
4c887265 AL |
2034 | /* |
2035 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2036 | * each hugepage. We have to make * sure we get the | |
2037 | * first, for the page indexing below to work. | |
2038 | */ | |
a5516438 | 2039 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); |
63551ae0 | 2040 | |
7f2e9525 GS |
2041 | if (!pte || huge_pte_none(huge_ptep_get(pte)) || |
2042 | (write && !pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 2043 | int ret; |
63551ae0 | 2044 | |
4c887265 | 2045 | spin_unlock(&mm->page_table_lock); |
5b23dbe8 | 2046 | ret = hugetlb_fault(mm, vma, vaddr, write); |
4c887265 | 2047 | spin_lock(&mm->page_table_lock); |
a89182c7 | 2048 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 2049 | continue; |
63551ae0 | 2050 | |
4c887265 AL |
2051 | remainder = 0; |
2052 | if (!i) | |
2053 | i = -EFAULT; | |
2054 | break; | |
2055 | } | |
2056 | ||
a5516438 | 2057 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 2058 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 2059 | same_page: |
d6692183 CK |
2060 | if (pages) { |
2061 | get_page(page); | |
d5d4b0aa | 2062 | pages[i] = page + pfn_offset; |
d6692183 | 2063 | } |
63551ae0 DG |
2064 | |
2065 | if (vmas) | |
2066 | vmas[i] = vma; | |
2067 | ||
2068 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 2069 | ++pfn_offset; |
63551ae0 DG |
2070 | --remainder; |
2071 | ++i; | |
d5d4b0aa | 2072 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 2073 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa CK |
2074 | /* |
2075 | * We use pfn_offset to avoid touching the pageframes | |
2076 | * of this compound page. | |
2077 | */ | |
2078 | goto same_page; | |
2079 | } | |
63551ae0 | 2080 | } |
1c59827d | 2081 | spin_unlock(&mm->page_table_lock); |
63551ae0 DG |
2082 | *length = remainder; |
2083 | *position = vaddr; | |
2084 | ||
2085 | return i; | |
2086 | } | |
8f860591 ZY |
2087 | |
2088 | void hugetlb_change_protection(struct vm_area_struct *vma, | |
2089 | unsigned long address, unsigned long end, pgprot_t newprot) | |
2090 | { | |
2091 | struct mm_struct *mm = vma->vm_mm; | |
2092 | unsigned long start = address; | |
2093 | pte_t *ptep; | |
2094 | pte_t pte; | |
a5516438 | 2095 | struct hstate *h = hstate_vma(vma); |
8f860591 ZY |
2096 | |
2097 | BUG_ON(address >= end); | |
2098 | flush_cache_range(vma, address, end); | |
2099 | ||
39dde65c | 2100 | spin_lock(&vma->vm_file->f_mapping->i_mmap_lock); |
8f860591 | 2101 | spin_lock(&mm->page_table_lock); |
a5516438 | 2102 | for (; address < end; address += huge_page_size(h)) { |
8f860591 ZY |
2103 | ptep = huge_pte_offset(mm, address); |
2104 | if (!ptep) | |
2105 | continue; | |
39dde65c CK |
2106 | if (huge_pmd_unshare(mm, &address, ptep)) |
2107 | continue; | |
7f2e9525 | 2108 | if (!huge_pte_none(huge_ptep_get(ptep))) { |
8f860591 ZY |
2109 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
2110 | pte = pte_mkhuge(pte_modify(pte, newprot)); | |
2111 | set_huge_pte_at(mm, address, ptep, pte); | |
8f860591 ZY |
2112 | } |
2113 | } | |
2114 | spin_unlock(&mm->page_table_lock); | |
39dde65c | 2115 | spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock); |
8f860591 ZY |
2116 | |
2117 | flush_tlb_range(vma, start, end); | |
2118 | } | |
2119 | ||
a1e78772 MG |
2120 | int hugetlb_reserve_pages(struct inode *inode, |
2121 | long from, long to, | |
2122 | struct vm_area_struct *vma) | |
e4e574b7 AL |
2123 | { |
2124 | long ret, chg; | |
a5516438 | 2125 | struct hstate *h = hstate_inode(inode); |
e4e574b7 | 2126 | |
c37f9fb1 AW |
2127 | if (vma && vma->vm_flags & VM_NORESERVE) |
2128 | return 0; | |
2129 | ||
a1e78772 MG |
2130 | /* |
2131 | * Shared mappings base their reservation on the number of pages that | |
2132 | * are already allocated on behalf of the file. Private mappings need | |
2133 | * to reserve the full area even if read-only as mprotect() may be | |
2134 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
2135 | */ | |
2136 | if (!vma || vma->vm_flags & VM_SHARED) | |
2137 | chg = region_chg(&inode->i_mapping->private_list, from, to); | |
2138 | else { | |
84afd99b AW |
2139 | struct resv_map *resv_map = resv_map_alloc(); |
2140 | if (!resv_map) | |
2141 | return -ENOMEM; | |
2142 | ||
a1e78772 | 2143 | chg = to - from; |
84afd99b AW |
2144 | |
2145 | set_vma_resv_map(vma, resv_map); | |
04f2cbe3 | 2146 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); |
a1e78772 MG |
2147 | } |
2148 | ||
e4e574b7 AL |
2149 | if (chg < 0) |
2150 | return chg; | |
8a630112 | 2151 | |
90d8b7e6 AL |
2152 | if (hugetlb_get_quota(inode->i_mapping, chg)) |
2153 | return -ENOSPC; | |
a5516438 | 2154 | ret = hugetlb_acct_memory(h, chg); |
68842c9b KC |
2155 | if (ret < 0) { |
2156 | hugetlb_put_quota(inode->i_mapping, chg); | |
a43a8c39 | 2157 | return ret; |
68842c9b | 2158 | } |
a1e78772 MG |
2159 | if (!vma || vma->vm_flags & VM_SHARED) |
2160 | region_add(&inode->i_mapping->private_list, from, to); | |
a43a8c39 CK |
2161 | return 0; |
2162 | } | |
2163 | ||
2164 | void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) | |
2165 | { | |
a5516438 | 2166 | struct hstate *h = hstate_inode(inode); |
a43a8c39 | 2167 | long chg = region_truncate(&inode->i_mapping->private_list, offset); |
45c682a6 KC |
2168 | |
2169 | spin_lock(&inode->i_lock); | |
a5516438 | 2170 | inode->i_blocks -= blocks_per_huge_page(h); |
45c682a6 KC |
2171 | spin_unlock(&inode->i_lock); |
2172 | ||
90d8b7e6 | 2173 | hugetlb_put_quota(inode->i_mapping, (chg - freed)); |
a5516438 | 2174 | hugetlb_acct_memory(h, -(chg - freed)); |
a43a8c39 | 2175 | } |