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