]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * Generic hugetlb support. | |
3 | * (C) William Irwin, April 2004 | |
4 | */ | |
1da177e4 LT |
5 | #include <linux/list.h> |
6 | #include <linux/init.h> | |
7 | #include <linux/module.h> | |
8 | #include <linux/mm.h> | |
e1759c21 | 9 | #include <linux/seq_file.h> |
1da177e4 LT |
10 | #include <linux/sysctl.h> |
11 | #include <linux/highmem.h> | |
cddb8a5c | 12 | #include <linux/mmu_notifier.h> |
1da177e4 | 13 | #include <linux/nodemask.h> |
63551ae0 | 14 | #include <linux/pagemap.h> |
5da7ca86 | 15 | #include <linux/mempolicy.h> |
aea47ff3 | 16 | #include <linux/cpuset.h> |
3935baa9 | 17 | #include <linux/mutex.h> |
aa888a74 | 18 | #include <linux/bootmem.h> |
a3437870 | 19 | #include <linux/sysfs.h> |
5a0e3ad6 | 20 | #include <linux/slab.h> |
0fe6e20b | 21 | #include <linux/rmap.h> |
fd6a03ed NH |
22 | #include <linux/swap.h> |
23 | #include <linux/swapops.h> | |
d6606683 | 24 | |
63551ae0 DG |
25 | #include <asm/page.h> |
26 | #include <asm/pgtable.h> | |
24669e58 | 27 | #include <asm/tlb.h> |
63551ae0 | 28 | |
24669e58 | 29 | #include <linux/io.h> |
63551ae0 | 30 | #include <linux/hugetlb.h> |
9a305230 | 31 | #include <linux/node.h> |
7835e98b | 32 | #include "internal.h" |
1da177e4 LT |
33 | |
34 | const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL; | |
396faf03 MG |
35 | static gfp_t htlb_alloc_mask = GFP_HIGHUSER; |
36 | unsigned long hugepages_treat_as_movable; | |
a5516438 | 37 | |
47d38344 | 38 | static int hugetlb_max_hstate; |
e5ff2159 AK |
39 | unsigned int default_hstate_idx; |
40 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
41 | ||
53ba51d2 JT |
42 | __initdata LIST_HEAD(huge_boot_pages); |
43 | ||
e5ff2159 AK |
44 | /* for command line parsing */ |
45 | static struct hstate * __initdata parsed_hstate; | |
46 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 47 | static unsigned long __initdata default_hstate_size; |
e5ff2159 AK |
48 | |
49 | #define for_each_hstate(h) \ | |
47d38344 | 50 | for ((h) = hstates; (h) < &hstates[hugetlb_max_hstate]; (h)++) |
396faf03 | 51 | |
3935baa9 DG |
52 | /* |
53 | * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages | |
54 | */ | |
55 | static DEFINE_SPINLOCK(hugetlb_lock); | |
0bd0f9fb | 56 | |
90481622 DG |
57 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
58 | { | |
59 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
60 | ||
61 | spin_unlock(&spool->lock); | |
62 | ||
63 | /* If no pages are used, and no other handles to the subpool | |
64 | * remain, free the subpool the subpool remain */ | |
65 | if (free) | |
66 | kfree(spool); | |
67 | } | |
68 | ||
69 | struct hugepage_subpool *hugepage_new_subpool(long nr_blocks) | |
70 | { | |
71 | struct hugepage_subpool *spool; | |
72 | ||
73 | spool = kmalloc(sizeof(*spool), GFP_KERNEL); | |
74 | if (!spool) | |
75 | return NULL; | |
76 | ||
77 | spin_lock_init(&spool->lock); | |
78 | spool->count = 1; | |
79 | spool->max_hpages = nr_blocks; | |
80 | spool->used_hpages = 0; | |
81 | ||
82 | return spool; | |
83 | } | |
84 | ||
85 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
86 | { | |
87 | spin_lock(&spool->lock); | |
88 | BUG_ON(!spool->count); | |
89 | spool->count--; | |
90 | unlock_or_release_subpool(spool); | |
91 | } | |
92 | ||
93 | static int hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
94 | long delta) | |
95 | { | |
96 | int ret = 0; | |
97 | ||
98 | if (!spool) | |
99 | return 0; | |
100 | ||
101 | spin_lock(&spool->lock); | |
102 | if ((spool->used_hpages + delta) <= spool->max_hpages) { | |
103 | spool->used_hpages += delta; | |
104 | } else { | |
105 | ret = -ENOMEM; | |
106 | } | |
107 | spin_unlock(&spool->lock); | |
108 | ||
109 | return ret; | |
110 | } | |
111 | ||
112 | static void hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
113 | long delta) | |
114 | { | |
115 | if (!spool) | |
116 | return; | |
117 | ||
118 | spin_lock(&spool->lock); | |
119 | spool->used_hpages -= delta; | |
120 | /* If hugetlbfs_put_super couldn't free spool due to | |
121 | * an outstanding quota reference, free it now. */ | |
122 | unlock_or_release_subpool(spool); | |
123 | } | |
124 | ||
125 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
126 | { | |
127 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
128 | } | |
129 | ||
130 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
131 | { | |
132 | return subpool_inode(vma->vm_file->f_dentry->d_inode); | |
133 | } | |
134 | ||
96822904 AW |
135 | /* |
136 | * Region tracking -- allows tracking of reservations and instantiated pages | |
137 | * across the pages in a mapping. | |
84afd99b AW |
138 | * |
139 | * The region data structures are protected by a combination of the mmap_sem | |
140 | * and the hugetlb_instantion_mutex. To access or modify a region the caller | |
141 | * must either hold the mmap_sem for write, or the mmap_sem for read and | |
142 | * the hugetlb_instantiation mutex: | |
143 | * | |
32f84528 | 144 | * down_write(&mm->mmap_sem); |
84afd99b | 145 | * or |
32f84528 CF |
146 | * down_read(&mm->mmap_sem); |
147 | * mutex_lock(&hugetlb_instantiation_mutex); | |
96822904 AW |
148 | */ |
149 | struct file_region { | |
150 | struct list_head link; | |
151 | long from; | |
152 | long to; | |
153 | }; | |
154 | ||
155 | static long region_add(struct list_head *head, long f, long t) | |
156 | { | |
157 | struct file_region *rg, *nrg, *trg; | |
158 | ||
159 | /* Locate the region we are either in or before. */ | |
160 | list_for_each_entry(rg, head, link) | |
161 | if (f <= rg->to) | |
162 | break; | |
163 | ||
164 | /* Round our left edge to the current segment if it encloses us. */ | |
165 | if (f > rg->from) | |
166 | f = rg->from; | |
167 | ||
168 | /* Check for and consume any regions we now overlap with. */ | |
169 | nrg = rg; | |
170 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
171 | if (&rg->link == head) | |
172 | break; | |
173 | if (rg->from > t) | |
174 | break; | |
175 | ||
176 | /* If this area reaches higher then extend our area to | |
177 | * include it completely. If this is not the first area | |
178 | * which we intend to reuse, free it. */ | |
179 | if (rg->to > t) | |
180 | t = rg->to; | |
181 | if (rg != nrg) { | |
182 | list_del(&rg->link); | |
183 | kfree(rg); | |
184 | } | |
185 | } | |
186 | nrg->from = f; | |
187 | nrg->to = t; | |
188 | return 0; | |
189 | } | |
190 | ||
191 | static long region_chg(struct list_head *head, long f, long t) | |
192 | { | |
193 | struct file_region *rg, *nrg; | |
194 | long chg = 0; | |
195 | ||
196 | /* Locate the region we are before or in. */ | |
197 | list_for_each_entry(rg, head, link) | |
198 | if (f <= rg->to) | |
199 | break; | |
200 | ||
201 | /* If we are below the current region then a new region is required. | |
202 | * Subtle, allocate a new region at the position but make it zero | |
203 | * size such that we can guarantee to record the reservation. */ | |
204 | if (&rg->link == head || t < rg->from) { | |
205 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
206 | if (!nrg) | |
207 | return -ENOMEM; | |
208 | nrg->from = f; | |
209 | nrg->to = f; | |
210 | INIT_LIST_HEAD(&nrg->link); | |
211 | list_add(&nrg->link, rg->link.prev); | |
212 | ||
213 | return t - f; | |
214 | } | |
215 | ||
216 | /* Round our left edge to the current segment if it encloses us. */ | |
217 | if (f > rg->from) | |
218 | f = rg->from; | |
219 | chg = t - f; | |
220 | ||
221 | /* Check for and consume any regions we now overlap with. */ | |
222 | list_for_each_entry(rg, rg->link.prev, link) { | |
223 | if (&rg->link == head) | |
224 | break; | |
225 | if (rg->from > t) | |
226 | return chg; | |
227 | ||
25985edc | 228 | /* We overlap with this area, if it extends further than |
96822904 AW |
229 | * us then we must extend ourselves. Account for its |
230 | * existing reservation. */ | |
231 | if (rg->to > t) { | |
232 | chg += rg->to - t; | |
233 | t = rg->to; | |
234 | } | |
235 | chg -= rg->to - rg->from; | |
236 | } | |
237 | return chg; | |
238 | } | |
239 | ||
240 | static long region_truncate(struct list_head *head, long end) | |
241 | { | |
242 | struct file_region *rg, *trg; | |
243 | long chg = 0; | |
244 | ||
245 | /* Locate the region we are either in or before. */ | |
246 | list_for_each_entry(rg, head, link) | |
247 | if (end <= rg->to) | |
248 | break; | |
249 | if (&rg->link == head) | |
250 | return 0; | |
251 | ||
252 | /* If we are in the middle of a region then adjust it. */ | |
253 | if (end > rg->from) { | |
254 | chg = rg->to - end; | |
255 | rg->to = end; | |
256 | rg = list_entry(rg->link.next, typeof(*rg), link); | |
257 | } | |
258 | ||
259 | /* Drop any remaining regions. */ | |
260 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
261 | if (&rg->link == head) | |
262 | break; | |
263 | chg += rg->to - rg->from; | |
264 | list_del(&rg->link); | |
265 | kfree(rg); | |
266 | } | |
267 | return chg; | |
268 | } | |
269 | ||
84afd99b AW |
270 | static long region_count(struct list_head *head, long f, long t) |
271 | { | |
272 | struct file_region *rg; | |
273 | long chg = 0; | |
274 | ||
275 | /* Locate each segment we overlap with, and count that overlap. */ | |
276 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
277 | long seg_from; |
278 | long seg_to; | |
84afd99b AW |
279 | |
280 | if (rg->to <= f) | |
281 | continue; | |
282 | if (rg->from >= t) | |
283 | break; | |
284 | ||
285 | seg_from = max(rg->from, f); | |
286 | seg_to = min(rg->to, t); | |
287 | ||
288 | chg += seg_to - seg_from; | |
289 | } | |
290 | ||
291 | return chg; | |
292 | } | |
293 | ||
e7c4b0bf AW |
294 | /* |
295 | * Convert the address within this vma to the page offset within | |
296 | * the mapping, in pagecache page units; huge pages here. | |
297 | */ | |
a5516438 AK |
298 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
299 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 300 | { |
a5516438 AK |
301 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
302 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
303 | } |
304 | ||
0fe6e20b NH |
305 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
306 | unsigned long address) | |
307 | { | |
308 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
309 | } | |
310 | ||
08fba699 MG |
311 | /* |
312 | * Return the size of the pages allocated when backing a VMA. In the majority | |
313 | * cases this will be same size as used by the page table entries. | |
314 | */ | |
315 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
316 | { | |
317 | struct hstate *hstate; | |
318 | ||
319 | if (!is_vm_hugetlb_page(vma)) | |
320 | return PAGE_SIZE; | |
321 | ||
322 | hstate = hstate_vma(vma); | |
323 | ||
324 | return 1UL << (hstate->order + PAGE_SHIFT); | |
325 | } | |
f340ca0f | 326 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 327 | |
3340289d MG |
328 | /* |
329 | * Return the page size being used by the MMU to back a VMA. In the majority | |
330 | * of cases, the page size used by the kernel matches the MMU size. On | |
331 | * architectures where it differs, an architecture-specific version of this | |
332 | * function is required. | |
333 | */ | |
334 | #ifndef vma_mmu_pagesize | |
335 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) | |
336 | { | |
337 | return vma_kernel_pagesize(vma); | |
338 | } | |
339 | #endif | |
340 | ||
84afd99b AW |
341 | /* |
342 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
343 | * bits of the reservation map pointer, which are always clear due to | |
344 | * alignment. | |
345 | */ | |
346 | #define HPAGE_RESV_OWNER (1UL << 0) | |
347 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 348 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 349 | |
a1e78772 MG |
350 | /* |
351 | * These helpers are used to track how many pages are reserved for | |
352 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
353 | * is guaranteed to have their future faults succeed. | |
354 | * | |
355 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
356 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
357 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
358 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
359 | * |
360 | * The private mapping reservation is represented in a subtly different | |
361 | * manner to a shared mapping. A shared mapping has a region map associated | |
362 | * with the underlying file, this region map represents the backing file | |
363 | * pages which have ever had a reservation assigned which this persists even | |
364 | * after the page is instantiated. A private mapping has a region map | |
365 | * associated with the original mmap which is attached to all VMAs which | |
366 | * reference it, this region map represents those offsets which have consumed | |
367 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 368 | */ |
e7c4b0bf AW |
369 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
370 | { | |
371 | return (unsigned long)vma->vm_private_data; | |
372 | } | |
373 | ||
374 | static void set_vma_private_data(struct vm_area_struct *vma, | |
375 | unsigned long value) | |
376 | { | |
377 | vma->vm_private_data = (void *)value; | |
378 | } | |
379 | ||
84afd99b AW |
380 | struct resv_map { |
381 | struct kref refs; | |
382 | struct list_head regions; | |
383 | }; | |
384 | ||
2a4b3ded | 385 | static struct resv_map *resv_map_alloc(void) |
84afd99b AW |
386 | { |
387 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
388 | if (!resv_map) | |
389 | return NULL; | |
390 | ||
391 | kref_init(&resv_map->refs); | |
392 | INIT_LIST_HEAD(&resv_map->regions); | |
393 | ||
394 | return resv_map; | |
395 | } | |
396 | ||
2a4b3ded | 397 | static void resv_map_release(struct kref *ref) |
84afd99b AW |
398 | { |
399 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
400 | ||
401 | /* Clear out any active regions before we release the map. */ | |
402 | region_truncate(&resv_map->regions, 0); | |
403 | kfree(resv_map); | |
404 | } | |
405 | ||
406 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) | |
a1e78772 MG |
407 | { |
408 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 409 | if (!(vma->vm_flags & VM_MAYSHARE)) |
84afd99b AW |
410 | return (struct resv_map *)(get_vma_private_data(vma) & |
411 | ~HPAGE_RESV_MASK); | |
2a4b3ded | 412 | return NULL; |
a1e78772 MG |
413 | } |
414 | ||
84afd99b | 415 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 MG |
416 | { |
417 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 418 | VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); |
a1e78772 | 419 | |
84afd99b AW |
420 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
421 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
422 | } |
423 | ||
424 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
425 | { | |
04f2cbe3 | 426 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); |
f83a275d | 427 | VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); |
e7c4b0bf AW |
428 | |
429 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
430 | } |
431 | ||
432 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
433 | { | |
434 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
e7c4b0bf AW |
435 | |
436 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
437 | } |
438 | ||
439 | /* Decrement the reserved pages in the hugepage pool by one */ | |
a5516438 AK |
440 | static void decrement_hugepage_resv_vma(struct hstate *h, |
441 | struct vm_area_struct *vma) | |
a1e78772 | 442 | { |
c37f9fb1 AW |
443 | if (vma->vm_flags & VM_NORESERVE) |
444 | return; | |
445 | ||
f83a275d | 446 | if (vma->vm_flags & VM_MAYSHARE) { |
a1e78772 | 447 | /* Shared mappings always use reserves */ |
a5516438 | 448 | h->resv_huge_pages--; |
84afd99b | 449 | } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
a1e78772 MG |
450 | /* |
451 | * Only the process that called mmap() has reserves for | |
452 | * private mappings. | |
453 | */ | |
a5516438 | 454 | h->resv_huge_pages--; |
a1e78772 MG |
455 | } |
456 | } | |
457 | ||
04f2cbe3 | 458 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
459 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
460 | { | |
461 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 462 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
463 | vma->vm_private_data = (void *)0; |
464 | } | |
465 | ||
466 | /* Returns true if the VMA has associated reserve pages */ | |
7f09ca51 | 467 | static int vma_has_reserves(struct vm_area_struct *vma) |
a1e78772 | 468 | { |
f83a275d | 469 | if (vma->vm_flags & VM_MAYSHARE) |
7f09ca51 MG |
470 | return 1; |
471 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) | |
472 | return 1; | |
473 | return 0; | |
a1e78772 MG |
474 | } |
475 | ||
0ebabb41 NH |
476 | static void copy_gigantic_page(struct page *dst, struct page *src) |
477 | { | |
478 | int i; | |
479 | struct hstate *h = page_hstate(src); | |
480 | struct page *dst_base = dst; | |
481 | struct page *src_base = src; | |
482 | ||
483 | for (i = 0; i < pages_per_huge_page(h); ) { | |
484 | cond_resched(); | |
485 | copy_highpage(dst, src); | |
486 | ||
487 | i++; | |
488 | dst = mem_map_next(dst, dst_base, i); | |
489 | src = mem_map_next(src, src_base, i); | |
490 | } | |
491 | } | |
492 | ||
493 | void copy_huge_page(struct page *dst, struct page *src) | |
494 | { | |
495 | int i; | |
496 | struct hstate *h = page_hstate(src); | |
497 | ||
498 | if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) { | |
499 | copy_gigantic_page(dst, src); | |
500 | return; | |
501 | } | |
502 | ||
503 | might_sleep(); | |
504 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
505 | cond_resched(); | |
506 | copy_highpage(dst + i, src + i); | |
507 | } | |
508 | } | |
509 | ||
a5516438 | 510 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
511 | { |
512 | int nid = page_to_nid(page); | |
0edaecfa | 513 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
514 | h->free_huge_pages++; |
515 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
516 | } |
517 | ||
bf50bab2 NH |
518 | static struct page *dequeue_huge_page_node(struct hstate *h, int nid) |
519 | { | |
520 | struct page *page; | |
521 | ||
522 | if (list_empty(&h->hugepage_freelists[nid])) | |
523 | return NULL; | |
524 | page = list_entry(h->hugepage_freelists[nid].next, struct page, lru); | |
0edaecfa | 525 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 526 | set_page_refcounted(page); |
bf50bab2 NH |
527 | h->free_huge_pages--; |
528 | h->free_huge_pages_node[nid]--; | |
529 | return page; | |
530 | } | |
531 | ||
a5516438 AK |
532 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
533 | struct vm_area_struct *vma, | |
04f2cbe3 | 534 | unsigned long address, int avoid_reserve) |
1da177e4 | 535 | { |
b1c12cbc | 536 | struct page *page = NULL; |
480eccf9 | 537 | struct mempolicy *mpol; |
19770b32 | 538 | nodemask_t *nodemask; |
c0ff7453 | 539 | struct zonelist *zonelist; |
dd1a239f MG |
540 | struct zone *zone; |
541 | struct zoneref *z; | |
cc9a6c87 | 542 | unsigned int cpuset_mems_cookie; |
1da177e4 | 543 | |
cc9a6c87 MG |
544 | retry_cpuset: |
545 | cpuset_mems_cookie = get_mems_allowed(); | |
c0ff7453 MX |
546 | zonelist = huge_zonelist(vma, address, |
547 | htlb_alloc_mask, &mpol, &nodemask); | |
a1e78772 MG |
548 | /* |
549 | * A child process with MAP_PRIVATE mappings created by their parent | |
550 | * have no page reserves. This check ensures that reservations are | |
551 | * not "stolen". The child may still get SIGKILLed | |
552 | */ | |
7f09ca51 | 553 | if (!vma_has_reserves(vma) && |
a5516438 | 554 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 555 | goto err; |
a1e78772 | 556 | |
04f2cbe3 | 557 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 558 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 559 | goto err; |
04f2cbe3 | 560 | |
19770b32 MG |
561 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
562 | MAX_NR_ZONES - 1, nodemask) { | |
bf50bab2 NH |
563 | if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) { |
564 | page = dequeue_huge_page_node(h, zone_to_nid(zone)); | |
565 | if (page) { | |
566 | if (!avoid_reserve) | |
567 | decrement_hugepage_resv_vma(h, vma); | |
568 | break; | |
569 | } | |
3abf7afd | 570 | } |
1da177e4 | 571 | } |
cc9a6c87 | 572 | |
52cd3b07 | 573 | mpol_cond_put(mpol); |
cc9a6c87 MG |
574 | if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page)) |
575 | goto retry_cpuset; | |
1da177e4 | 576 | return page; |
cc9a6c87 MG |
577 | |
578 | err: | |
579 | mpol_cond_put(mpol); | |
580 | return NULL; | |
1da177e4 LT |
581 | } |
582 | ||
a5516438 | 583 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
584 | { |
585 | int i; | |
a5516438 | 586 | |
18229df5 AW |
587 | VM_BUG_ON(h->order >= MAX_ORDER); |
588 | ||
a5516438 AK |
589 | h->nr_huge_pages--; |
590 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
591 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
592 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
593 | 1 << PG_referenced | 1 << PG_dirty | | |
594 | 1 << PG_active | 1 << PG_reserved | | |
595 | 1 << PG_private | 1 << PG_writeback); | |
6af2acb6 AL |
596 | } |
597 | set_compound_page_dtor(page, NULL); | |
598 | set_page_refcounted(page); | |
7f2e9525 | 599 | arch_release_hugepage(page); |
a5516438 | 600 | __free_pages(page, huge_page_order(h)); |
6af2acb6 AL |
601 | } |
602 | ||
e5ff2159 AK |
603 | struct hstate *size_to_hstate(unsigned long size) |
604 | { | |
605 | struct hstate *h; | |
606 | ||
607 | for_each_hstate(h) { | |
608 | if (huge_page_size(h) == size) | |
609 | return h; | |
610 | } | |
611 | return NULL; | |
612 | } | |
613 | ||
27a85ef1 DG |
614 | static void free_huge_page(struct page *page) |
615 | { | |
a5516438 AK |
616 | /* |
617 | * Can't pass hstate in here because it is called from the | |
618 | * compound page destructor. | |
619 | */ | |
e5ff2159 | 620 | struct hstate *h = page_hstate(page); |
7893d1d5 | 621 | int nid = page_to_nid(page); |
90481622 DG |
622 | struct hugepage_subpool *spool = |
623 | (struct hugepage_subpool *)page_private(page); | |
27a85ef1 | 624 | |
e5df70ab | 625 | set_page_private(page, 0); |
23be7468 | 626 | page->mapping = NULL; |
7893d1d5 | 627 | BUG_ON(page_count(page)); |
0fe6e20b | 628 | BUG_ON(page_mapcount(page)); |
27a85ef1 DG |
629 | |
630 | spin_lock(&hugetlb_lock); | |
aa888a74 | 631 | if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { |
0edaecfa AK |
632 | /* remove the page from active list */ |
633 | list_del(&page->lru); | |
a5516438 AK |
634 | update_and_free_page(h, page); |
635 | h->surplus_huge_pages--; | |
636 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 637 | } else { |
a5516438 | 638 | enqueue_huge_page(h, page); |
7893d1d5 | 639 | } |
27a85ef1 | 640 | spin_unlock(&hugetlb_lock); |
90481622 | 641 | hugepage_subpool_put_pages(spool, 1); |
27a85ef1 DG |
642 | } |
643 | ||
a5516438 | 644 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 645 | { |
0edaecfa | 646 | INIT_LIST_HEAD(&page->lru); |
b7ba30c6 AK |
647 | set_compound_page_dtor(page, free_huge_page); |
648 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
649 | h->nr_huge_pages++; |
650 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 AK |
651 | spin_unlock(&hugetlb_lock); |
652 | put_page(page); /* free it into the hugepage allocator */ | |
653 | } | |
654 | ||
20a0307c WF |
655 | static void prep_compound_gigantic_page(struct page *page, unsigned long order) |
656 | { | |
657 | int i; | |
658 | int nr_pages = 1 << order; | |
659 | struct page *p = page + 1; | |
660 | ||
661 | /* we rely on prep_new_huge_page to set the destructor */ | |
662 | set_compound_order(page, order); | |
663 | __SetPageHead(page); | |
664 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { | |
665 | __SetPageTail(p); | |
58a84aa9 | 666 | set_page_count(p, 0); |
20a0307c WF |
667 | p->first_page = page; |
668 | } | |
669 | } | |
670 | ||
671 | int PageHuge(struct page *page) | |
672 | { | |
673 | compound_page_dtor *dtor; | |
674 | ||
675 | if (!PageCompound(page)) | |
676 | return 0; | |
677 | ||
678 | page = compound_head(page); | |
679 | dtor = get_compound_page_dtor(page); | |
680 | ||
681 | return dtor == free_huge_page; | |
682 | } | |
43131e14 NH |
683 | EXPORT_SYMBOL_GPL(PageHuge); |
684 | ||
a5516438 | 685 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
1da177e4 | 686 | { |
1da177e4 | 687 | struct page *page; |
f96efd58 | 688 | |
aa888a74 AK |
689 | if (h->order >= MAX_ORDER) |
690 | return NULL; | |
691 | ||
6484eb3e | 692 | page = alloc_pages_exact_node(nid, |
551883ae NA |
693 | htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| |
694 | __GFP_REPEAT|__GFP_NOWARN, | |
a5516438 | 695 | huge_page_order(h)); |
1da177e4 | 696 | if (page) { |
7f2e9525 | 697 | if (arch_prepare_hugepage(page)) { |
caff3a2c | 698 | __free_pages(page, huge_page_order(h)); |
7b8ee84d | 699 | return NULL; |
7f2e9525 | 700 | } |
a5516438 | 701 | prep_new_huge_page(h, page, nid); |
1da177e4 | 702 | } |
63b4613c NA |
703 | |
704 | return page; | |
705 | } | |
706 | ||
9a76db09 | 707 | /* |
6ae11b27 LS |
708 | * common helper functions for hstate_next_node_to_{alloc|free}. |
709 | * We may have allocated or freed a huge page based on a different | |
710 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
711 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
712 | * node for alloc or free. | |
9a76db09 | 713 | */ |
6ae11b27 | 714 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) |
9a76db09 | 715 | { |
6ae11b27 | 716 | nid = next_node(nid, *nodes_allowed); |
9a76db09 | 717 | if (nid == MAX_NUMNODES) |
6ae11b27 | 718 | nid = first_node(*nodes_allowed); |
9a76db09 LS |
719 | VM_BUG_ON(nid >= MAX_NUMNODES); |
720 | ||
721 | return nid; | |
722 | } | |
723 | ||
6ae11b27 LS |
724 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) |
725 | { | |
726 | if (!node_isset(nid, *nodes_allowed)) | |
727 | nid = next_node_allowed(nid, nodes_allowed); | |
728 | return nid; | |
729 | } | |
730 | ||
5ced66c9 | 731 | /* |
6ae11b27 LS |
732 | * returns the previously saved node ["this node"] from which to |
733 | * allocate a persistent huge page for the pool and advance the | |
734 | * next node from which to allocate, handling wrap at end of node | |
735 | * mask. | |
5ced66c9 | 736 | */ |
6ae11b27 LS |
737 | static int hstate_next_node_to_alloc(struct hstate *h, |
738 | nodemask_t *nodes_allowed) | |
5ced66c9 | 739 | { |
6ae11b27 LS |
740 | int nid; |
741 | ||
742 | VM_BUG_ON(!nodes_allowed); | |
743 | ||
744 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
745 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
9a76db09 | 746 | |
9a76db09 | 747 | return nid; |
5ced66c9 AK |
748 | } |
749 | ||
6ae11b27 | 750 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
63b4613c NA |
751 | { |
752 | struct page *page; | |
753 | int start_nid; | |
754 | int next_nid; | |
755 | int ret = 0; | |
756 | ||
6ae11b27 | 757 | start_nid = hstate_next_node_to_alloc(h, nodes_allowed); |
e8c5c824 | 758 | next_nid = start_nid; |
63b4613c NA |
759 | |
760 | do { | |
e8c5c824 | 761 | page = alloc_fresh_huge_page_node(h, next_nid); |
9a76db09 | 762 | if (page) { |
63b4613c | 763 | ret = 1; |
9a76db09 LS |
764 | break; |
765 | } | |
6ae11b27 | 766 | next_nid = hstate_next_node_to_alloc(h, nodes_allowed); |
9a76db09 | 767 | } while (next_nid != start_nid); |
63b4613c | 768 | |
3b116300 AL |
769 | if (ret) |
770 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
771 | else | |
772 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
773 | ||
63b4613c | 774 | return ret; |
1da177e4 LT |
775 | } |
776 | ||
e8c5c824 | 777 | /* |
6ae11b27 LS |
778 | * helper for free_pool_huge_page() - return the previously saved |
779 | * node ["this node"] from which to free a huge page. Advance the | |
780 | * next node id whether or not we find a free huge page to free so | |
781 | * that the next attempt to free addresses the next node. | |
e8c5c824 | 782 | */ |
6ae11b27 | 783 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) |
e8c5c824 | 784 | { |
6ae11b27 LS |
785 | int nid; |
786 | ||
787 | VM_BUG_ON(!nodes_allowed); | |
788 | ||
789 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
790 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
9a76db09 | 791 | |
9a76db09 | 792 | return nid; |
e8c5c824 LS |
793 | } |
794 | ||
795 | /* | |
796 | * Free huge page from pool from next node to free. | |
797 | * Attempt to keep persistent huge pages more or less | |
798 | * balanced over allowed nodes. | |
799 | * Called with hugetlb_lock locked. | |
800 | */ | |
6ae11b27 LS |
801 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
802 | bool acct_surplus) | |
e8c5c824 LS |
803 | { |
804 | int start_nid; | |
805 | int next_nid; | |
806 | int ret = 0; | |
807 | ||
6ae11b27 | 808 | start_nid = hstate_next_node_to_free(h, nodes_allowed); |
e8c5c824 LS |
809 | next_nid = start_nid; |
810 | ||
811 | do { | |
685f3457 LS |
812 | /* |
813 | * If we're returning unused surplus pages, only examine | |
814 | * nodes with surplus pages. | |
815 | */ | |
816 | if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) && | |
817 | !list_empty(&h->hugepage_freelists[next_nid])) { | |
e8c5c824 LS |
818 | struct page *page = |
819 | list_entry(h->hugepage_freelists[next_nid].next, | |
820 | struct page, lru); | |
821 | list_del(&page->lru); | |
822 | h->free_huge_pages--; | |
823 | h->free_huge_pages_node[next_nid]--; | |
685f3457 LS |
824 | if (acct_surplus) { |
825 | h->surplus_huge_pages--; | |
826 | h->surplus_huge_pages_node[next_nid]--; | |
827 | } | |
e8c5c824 LS |
828 | update_and_free_page(h, page); |
829 | ret = 1; | |
9a76db09 | 830 | break; |
e8c5c824 | 831 | } |
6ae11b27 | 832 | next_nid = hstate_next_node_to_free(h, nodes_allowed); |
9a76db09 | 833 | } while (next_nid != start_nid); |
e8c5c824 LS |
834 | |
835 | return ret; | |
836 | } | |
837 | ||
bf50bab2 | 838 | static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) |
7893d1d5 AL |
839 | { |
840 | struct page *page; | |
bf50bab2 | 841 | unsigned int r_nid; |
7893d1d5 | 842 | |
aa888a74 AK |
843 | if (h->order >= MAX_ORDER) |
844 | return NULL; | |
845 | ||
d1c3fb1f NA |
846 | /* |
847 | * Assume we will successfully allocate the surplus page to | |
848 | * prevent racing processes from causing the surplus to exceed | |
849 | * overcommit | |
850 | * | |
851 | * This however introduces a different race, where a process B | |
852 | * tries to grow the static hugepage pool while alloc_pages() is | |
853 | * called by process A. B will only examine the per-node | |
854 | * counters in determining if surplus huge pages can be | |
855 | * converted to normal huge pages in adjust_pool_surplus(). A | |
856 | * won't be able to increment the per-node counter, until the | |
857 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
858 | * no more huge pages can be converted from surplus to normal | |
859 | * state (and doesn't try to convert again). Thus, we have a | |
860 | * case where a surplus huge page exists, the pool is grown, and | |
861 | * the surplus huge page still exists after, even though it | |
862 | * should just have been converted to a normal huge page. This | |
863 | * does not leak memory, though, as the hugepage will be freed | |
864 | * once it is out of use. It also does not allow the counters to | |
865 | * go out of whack in adjust_pool_surplus() as we don't modify | |
866 | * the node values until we've gotten the hugepage and only the | |
867 | * per-node value is checked there. | |
868 | */ | |
869 | spin_lock(&hugetlb_lock); | |
a5516438 | 870 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
871 | spin_unlock(&hugetlb_lock); |
872 | return NULL; | |
873 | } else { | |
a5516438 AK |
874 | h->nr_huge_pages++; |
875 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
876 | } |
877 | spin_unlock(&hugetlb_lock); | |
878 | ||
bf50bab2 NH |
879 | if (nid == NUMA_NO_NODE) |
880 | page = alloc_pages(htlb_alloc_mask|__GFP_COMP| | |
881 | __GFP_REPEAT|__GFP_NOWARN, | |
882 | huge_page_order(h)); | |
883 | else | |
884 | page = alloc_pages_exact_node(nid, | |
885 | htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| | |
886 | __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); | |
d1c3fb1f | 887 | |
caff3a2c GS |
888 | if (page && arch_prepare_hugepage(page)) { |
889 | __free_pages(page, huge_page_order(h)); | |
ea5768c7 | 890 | page = NULL; |
caff3a2c GS |
891 | } |
892 | ||
d1c3fb1f | 893 | spin_lock(&hugetlb_lock); |
7893d1d5 | 894 | if (page) { |
0edaecfa | 895 | INIT_LIST_HEAD(&page->lru); |
bf50bab2 | 896 | r_nid = page_to_nid(page); |
7893d1d5 | 897 | set_compound_page_dtor(page, free_huge_page); |
d1c3fb1f NA |
898 | /* |
899 | * We incremented the global counters already | |
900 | */ | |
bf50bab2 NH |
901 | h->nr_huge_pages_node[r_nid]++; |
902 | h->surplus_huge_pages_node[r_nid]++; | |
3b116300 | 903 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 904 | } else { |
a5516438 AK |
905 | h->nr_huge_pages--; |
906 | h->surplus_huge_pages--; | |
3b116300 | 907 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 908 | } |
d1c3fb1f | 909 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
910 | |
911 | return page; | |
912 | } | |
913 | ||
bf50bab2 NH |
914 | /* |
915 | * This allocation function is useful in the context where vma is irrelevant. | |
916 | * E.g. soft-offlining uses this function because it only cares physical | |
917 | * address of error page. | |
918 | */ | |
919 | struct page *alloc_huge_page_node(struct hstate *h, int nid) | |
920 | { | |
921 | struct page *page; | |
922 | ||
923 | spin_lock(&hugetlb_lock); | |
924 | page = dequeue_huge_page_node(h, nid); | |
925 | spin_unlock(&hugetlb_lock); | |
926 | ||
927 | if (!page) | |
928 | page = alloc_buddy_huge_page(h, nid); | |
929 | ||
930 | return page; | |
931 | } | |
932 | ||
e4e574b7 | 933 | /* |
25985edc | 934 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
935 | * of size 'delta'. |
936 | */ | |
a5516438 | 937 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
938 | { |
939 | struct list_head surplus_list; | |
940 | struct page *page, *tmp; | |
941 | int ret, i; | |
942 | int needed, allocated; | |
28073b02 | 943 | bool alloc_ok = true; |
e4e574b7 | 944 | |
a5516438 | 945 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 946 | if (needed <= 0) { |
a5516438 | 947 | h->resv_huge_pages += delta; |
e4e574b7 | 948 | return 0; |
ac09b3a1 | 949 | } |
e4e574b7 AL |
950 | |
951 | allocated = 0; | |
952 | INIT_LIST_HEAD(&surplus_list); | |
953 | ||
954 | ret = -ENOMEM; | |
955 | retry: | |
956 | spin_unlock(&hugetlb_lock); | |
957 | for (i = 0; i < needed; i++) { | |
bf50bab2 | 958 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
28073b02 HD |
959 | if (!page) { |
960 | alloc_ok = false; | |
961 | break; | |
962 | } | |
e4e574b7 AL |
963 | list_add(&page->lru, &surplus_list); |
964 | } | |
28073b02 | 965 | allocated += i; |
e4e574b7 AL |
966 | |
967 | /* | |
968 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
969 | * because either resv_huge_pages or free_huge_pages may have changed. | |
970 | */ | |
971 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
972 | needed = (h->resv_huge_pages + delta) - |
973 | (h->free_huge_pages + allocated); | |
28073b02 HD |
974 | if (needed > 0) { |
975 | if (alloc_ok) | |
976 | goto retry; | |
977 | /* | |
978 | * We were not able to allocate enough pages to | |
979 | * satisfy the entire reservation so we free what | |
980 | * we've allocated so far. | |
981 | */ | |
982 | goto free; | |
983 | } | |
e4e574b7 AL |
984 | /* |
985 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 986 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 987 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
988 | * allocator. Commit the entire reservation here to prevent another |
989 | * process from stealing the pages as they are added to the pool but | |
990 | * before they are reserved. | |
e4e574b7 AL |
991 | */ |
992 | needed += allocated; | |
a5516438 | 993 | h->resv_huge_pages += delta; |
e4e574b7 | 994 | ret = 0; |
a9869b83 | 995 | |
19fc3f0a | 996 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 997 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
998 | if ((--needed) < 0) |
999 | break; | |
a9869b83 NH |
1000 | /* |
1001 | * This page is now managed by the hugetlb allocator and has | |
1002 | * no users -- drop the buddy allocator's reference. | |
1003 | */ | |
1004 | put_page_testzero(page); | |
1005 | VM_BUG_ON(page_count(page)); | |
a5516438 | 1006 | enqueue_huge_page(h, page); |
19fc3f0a | 1007 | } |
28073b02 | 1008 | free: |
b0365c8d | 1009 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1010 | |
1011 | /* Free unnecessary surplus pages to the buddy allocator */ | |
1012 | if (!list_empty(&surplus_list)) { | |
19fc3f0a | 1013 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
a9869b83 | 1014 | put_page(page); |
af767cbd | 1015 | } |
e4e574b7 | 1016 | } |
a9869b83 | 1017 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1018 | |
1019 | return ret; | |
1020 | } | |
1021 | ||
1022 | /* | |
1023 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
1024 | * allocated to satisfy the reservation must be explicitly freed if they were | |
1025 | * never used. | |
685f3457 | 1026 | * Called with hugetlb_lock held. |
e4e574b7 | 1027 | */ |
a5516438 AK |
1028 | static void return_unused_surplus_pages(struct hstate *h, |
1029 | unsigned long unused_resv_pages) | |
e4e574b7 | 1030 | { |
e4e574b7 AL |
1031 | unsigned long nr_pages; |
1032 | ||
ac09b3a1 | 1033 | /* Uncommit the reservation */ |
a5516438 | 1034 | h->resv_huge_pages -= unused_resv_pages; |
ac09b3a1 | 1035 | |
aa888a74 AK |
1036 | /* Cannot return gigantic pages currently */ |
1037 | if (h->order >= MAX_ORDER) | |
1038 | return; | |
1039 | ||
a5516438 | 1040 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1041 | |
685f3457 LS |
1042 | /* |
1043 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1044 | * evenly across all nodes with memory. Iterate across these nodes |
1045 | * until we can no longer free unreserved surplus pages. This occurs | |
1046 | * when the nodes with surplus pages have no free pages. | |
1047 | * free_pool_huge_page() will balance the the freed pages across the | |
1048 | * on-line nodes with memory and will handle the hstate accounting. | |
685f3457 LS |
1049 | */ |
1050 | while (nr_pages--) { | |
9b5e5d0f | 1051 | if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1)) |
685f3457 | 1052 | break; |
e4e574b7 AL |
1053 | } |
1054 | } | |
1055 | ||
c37f9fb1 AW |
1056 | /* |
1057 | * Determine if the huge page at addr within the vma has an associated | |
1058 | * reservation. Where it does not we will need to logically increase | |
90481622 DG |
1059 | * reservation and actually increase subpool usage before an allocation |
1060 | * can occur. Where any new reservation would be required the | |
1061 | * reservation change is prepared, but not committed. Once the page | |
1062 | * has been allocated from the subpool and instantiated the change should | |
1063 | * be committed via vma_commit_reservation. No action is required on | |
1064 | * failure. | |
c37f9fb1 | 1065 | */ |
e2f17d94 | 1066 | static long vma_needs_reservation(struct hstate *h, |
a5516438 | 1067 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 AW |
1068 | { |
1069 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1070 | struct inode *inode = mapping->host; | |
1071 | ||
f83a275d | 1072 | if (vma->vm_flags & VM_MAYSHARE) { |
a5516438 | 1073 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
c37f9fb1 AW |
1074 | return region_chg(&inode->i_mapping->private_list, |
1075 | idx, idx + 1); | |
1076 | ||
84afd99b AW |
1077 | } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1078 | return 1; | |
c37f9fb1 | 1079 | |
84afd99b | 1080 | } else { |
e2f17d94 | 1081 | long err; |
a5516438 | 1082 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
84afd99b AW |
1083 | struct resv_map *reservations = vma_resv_map(vma); |
1084 | ||
1085 | err = region_chg(&reservations->regions, idx, idx + 1); | |
1086 | if (err < 0) | |
1087 | return err; | |
1088 | return 0; | |
1089 | } | |
c37f9fb1 | 1090 | } |
a5516438 AK |
1091 | static void vma_commit_reservation(struct hstate *h, |
1092 | struct vm_area_struct *vma, unsigned long addr) | |
c37f9fb1 AW |
1093 | { |
1094 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1095 | struct inode *inode = mapping->host; | |
1096 | ||
f83a275d | 1097 | if (vma->vm_flags & VM_MAYSHARE) { |
a5516438 | 1098 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
c37f9fb1 | 1099 | region_add(&inode->i_mapping->private_list, idx, idx + 1); |
84afd99b AW |
1100 | |
1101 | } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { | |
a5516438 | 1102 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
84afd99b AW |
1103 | struct resv_map *reservations = vma_resv_map(vma); |
1104 | ||
1105 | /* Mark this page used in the map. */ | |
1106 | region_add(&reservations->regions, idx, idx + 1); | |
c37f9fb1 AW |
1107 | } |
1108 | } | |
1109 | ||
a1e78772 | 1110 | static struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 1111 | unsigned long addr, int avoid_reserve) |
1da177e4 | 1112 | { |
90481622 | 1113 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 1114 | struct hstate *h = hstate_vma(vma); |
348ea204 | 1115 | struct page *page; |
e2f17d94 | 1116 | long chg; |
a1e78772 MG |
1117 | |
1118 | /* | |
90481622 DG |
1119 | * Processes that did not create the mapping will have no |
1120 | * reserves and will not have accounted against subpool | |
1121 | * limit. Check that the subpool limit can be made before | |
1122 | * satisfying the allocation MAP_NORESERVE mappings may also | |
1123 | * need pages and subpool limit allocated allocated if no reserve | |
1124 | * mapping overlaps. | |
a1e78772 | 1125 | */ |
a5516438 | 1126 | chg = vma_needs_reservation(h, vma, addr); |
c37f9fb1 | 1127 | if (chg < 0) |
76dcee75 | 1128 | return ERR_PTR(-ENOMEM); |
c37f9fb1 | 1129 | if (chg) |
90481622 | 1130 | if (hugepage_subpool_get_pages(spool, chg)) |
76dcee75 | 1131 | return ERR_PTR(-ENOSPC); |
1da177e4 LT |
1132 | |
1133 | spin_lock(&hugetlb_lock); | |
a5516438 | 1134 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve); |
1da177e4 | 1135 | spin_unlock(&hugetlb_lock); |
b45b5bd6 | 1136 | |
68842c9b | 1137 | if (!page) { |
bf50bab2 | 1138 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
68842c9b | 1139 | if (!page) { |
90481622 | 1140 | hugepage_subpool_put_pages(spool, chg); |
76dcee75 | 1141 | return ERR_PTR(-ENOSPC); |
68842c9b KC |
1142 | } |
1143 | } | |
348ea204 | 1144 | |
90481622 | 1145 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 1146 | |
a5516438 | 1147 | vma_commit_reservation(h, vma, addr); |
c37f9fb1 | 1148 | |
90d8b7e6 | 1149 | return page; |
b45b5bd6 DG |
1150 | } |
1151 | ||
91f47662 | 1152 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
aa888a74 AK |
1153 | { |
1154 | struct huge_bootmem_page *m; | |
9b5e5d0f | 1155 | int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); |
aa888a74 AK |
1156 | |
1157 | while (nr_nodes) { | |
1158 | void *addr; | |
1159 | ||
1160 | addr = __alloc_bootmem_node_nopanic( | |
6ae11b27 | 1161 | NODE_DATA(hstate_next_node_to_alloc(h, |
9b5e5d0f | 1162 | &node_states[N_HIGH_MEMORY])), |
aa888a74 AK |
1163 | huge_page_size(h), huge_page_size(h), 0); |
1164 | ||
1165 | if (addr) { | |
1166 | /* | |
1167 | * Use the beginning of the huge page to store the | |
1168 | * huge_bootmem_page struct (until gather_bootmem | |
1169 | * puts them into the mem_map). | |
1170 | */ | |
1171 | m = addr; | |
91f47662 | 1172 | goto found; |
aa888a74 | 1173 | } |
aa888a74 AK |
1174 | nr_nodes--; |
1175 | } | |
1176 | return 0; | |
1177 | ||
1178 | found: | |
1179 | BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); | |
1180 | /* Put them into a private list first because mem_map is not up yet */ | |
1181 | list_add(&m->list, &huge_boot_pages); | |
1182 | m->hstate = h; | |
1183 | return 1; | |
1184 | } | |
1185 | ||
18229df5 AW |
1186 | static void prep_compound_huge_page(struct page *page, int order) |
1187 | { | |
1188 | if (unlikely(order > (MAX_ORDER - 1))) | |
1189 | prep_compound_gigantic_page(page, order); | |
1190 | else | |
1191 | prep_compound_page(page, order); | |
1192 | } | |
1193 | ||
aa888a74 AK |
1194 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
1195 | static void __init gather_bootmem_prealloc(void) | |
1196 | { | |
1197 | struct huge_bootmem_page *m; | |
1198 | ||
1199 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 1200 | struct hstate *h = m->hstate; |
ee8f248d BB |
1201 | struct page *page; |
1202 | ||
1203 | #ifdef CONFIG_HIGHMEM | |
1204 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
1205 | free_bootmem_late((unsigned long)m, | |
1206 | sizeof(struct huge_bootmem_page)); | |
1207 | #else | |
1208 | page = virt_to_page(m); | |
1209 | #endif | |
aa888a74 AK |
1210 | __ClearPageReserved(page); |
1211 | WARN_ON(page_count(page) != 1); | |
18229df5 | 1212 | prep_compound_huge_page(page, h->order); |
aa888a74 | 1213 | prep_new_huge_page(h, page, page_to_nid(page)); |
b0320c7b RA |
1214 | /* |
1215 | * If we had gigantic hugepages allocated at boot time, we need | |
1216 | * to restore the 'stolen' pages to totalram_pages in order to | |
1217 | * fix confusing memory reports from free(1) and another | |
1218 | * side-effects, like CommitLimit going negative. | |
1219 | */ | |
1220 | if (h->order > (MAX_ORDER - 1)) | |
1221 | totalram_pages += 1 << h->order; | |
aa888a74 AK |
1222 | } |
1223 | } | |
1224 | ||
8faa8b07 | 1225 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
1226 | { |
1227 | unsigned long i; | |
a5516438 | 1228 | |
e5ff2159 | 1229 | for (i = 0; i < h->max_huge_pages; ++i) { |
aa888a74 AK |
1230 | if (h->order >= MAX_ORDER) { |
1231 | if (!alloc_bootmem_huge_page(h)) | |
1232 | break; | |
9b5e5d0f LS |
1233 | } else if (!alloc_fresh_huge_page(h, |
1234 | &node_states[N_HIGH_MEMORY])) | |
1da177e4 | 1235 | break; |
1da177e4 | 1236 | } |
8faa8b07 | 1237 | h->max_huge_pages = i; |
e5ff2159 AK |
1238 | } |
1239 | ||
1240 | static void __init hugetlb_init_hstates(void) | |
1241 | { | |
1242 | struct hstate *h; | |
1243 | ||
1244 | for_each_hstate(h) { | |
8faa8b07 AK |
1245 | /* oversize hugepages were init'ed in early boot */ |
1246 | if (h->order < MAX_ORDER) | |
1247 | hugetlb_hstate_alloc_pages(h); | |
e5ff2159 AK |
1248 | } |
1249 | } | |
1250 | ||
4abd32db AK |
1251 | static char * __init memfmt(char *buf, unsigned long n) |
1252 | { | |
1253 | if (n >= (1UL << 30)) | |
1254 | sprintf(buf, "%lu GB", n >> 30); | |
1255 | else if (n >= (1UL << 20)) | |
1256 | sprintf(buf, "%lu MB", n >> 20); | |
1257 | else | |
1258 | sprintf(buf, "%lu KB", n >> 10); | |
1259 | return buf; | |
1260 | } | |
1261 | ||
e5ff2159 AK |
1262 | static void __init report_hugepages(void) |
1263 | { | |
1264 | struct hstate *h; | |
1265 | ||
1266 | for_each_hstate(h) { | |
4abd32db AK |
1267 | char buf[32]; |
1268 | printk(KERN_INFO "HugeTLB registered %s page size, " | |
1269 | "pre-allocated %ld pages\n", | |
1270 | memfmt(buf, huge_page_size(h)), | |
1271 | h->free_huge_pages); | |
e5ff2159 AK |
1272 | } |
1273 | } | |
1274 | ||
1da177e4 | 1275 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
1276 | static void try_to_free_low(struct hstate *h, unsigned long count, |
1277 | nodemask_t *nodes_allowed) | |
1da177e4 | 1278 | { |
4415cc8d CL |
1279 | int i; |
1280 | ||
aa888a74 AK |
1281 | if (h->order >= MAX_ORDER) |
1282 | return; | |
1283 | ||
6ae11b27 | 1284 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 1285 | struct page *page, *next; |
a5516438 AK |
1286 | struct list_head *freel = &h->hugepage_freelists[i]; |
1287 | list_for_each_entry_safe(page, next, freel, lru) { | |
1288 | if (count >= h->nr_huge_pages) | |
6b0c880d | 1289 | return; |
1da177e4 LT |
1290 | if (PageHighMem(page)) |
1291 | continue; | |
1292 | list_del(&page->lru); | |
e5ff2159 | 1293 | update_and_free_page(h, page); |
a5516438 AK |
1294 | h->free_huge_pages--; |
1295 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
1296 | } |
1297 | } | |
1298 | } | |
1299 | #else | |
6ae11b27 LS |
1300 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
1301 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
1302 | { |
1303 | } | |
1304 | #endif | |
1305 | ||
20a0307c WF |
1306 | /* |
1307 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
1308 | * balanced by operating on them in a round-robin fashion. | |
1309 | * Returns 1 if an adjustment was made. | |
1310 | */ | |
6ae11b27 LS |
1311 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
1312 | int delta) | |
20a0307c | 1313 | { |
e8c5c824 | 1314 | int start_nid, next_nid; |
20a0307c WF |
1315 | int ret = 0; |
1316 | ||
1317 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 1318 | |
e8c5c824 | 1319 | if (delta < 0) |
6ae11b27 | 1320 | start_nid = hstate_next_node_to_alloc(h, nodes_allowed); |
e8c5c824 | 1321 | else |
6ae11b27 | 1322 | start_nid = hstate_next_node_to_free(h, nodes_allowed); |
e8c5c824 LS |
1323 | next_nid = start_nid; |
1324 | ||
1325 | do { | |
1326 | int nid = next_nid; | |
1327 | if (delta < 0) { | |
e8c5c824 LS |
1328 | /* |
1329 | * To shrink on this node, there must be a surplus page | |
1330 | */ | |
9a76db09 | 1331 | if (!h->surplus_huge_pages_node[nid]) { |
6ae11b27 LS |
1332 | next_nid = hstate_next_node_to_alloc(h, |
1333 | nodes_allowed); | |
e8c5c824 | 1334 | continue; |
9a76db09 | 1335 | } |
e8c5c824 LS |
1336 | } |
1337 | if (delta > 0) { | |
e8c5c824 LS |
1338 | /* |
1339 | * Surplus cannot exceed the total number of pages | |
1340 | */ | |
1341 | if (h->surplus_huge_pages_node[nid] >= | |
9a76db09 | 1342 | h->nr_huge_pages_node[nid]) { |
6ae11b27 LS |
1343 | next_nid = hstate_next_node_to_free(h, |
1344 | nodes_allowed); | |
e8c5c824 | 1345 | continue; |
9a76db09 | 1346 | } |
e8c5c824 | 1347 | } |
20a0307c WF |
1348 | |
1349 | h->surplus_huge_pages += delta; | |
1350 | h->surplus_huge_pages_node[nid] += delta; | |
1351 | ret = 1; | |
1352 | break; | |
e8c5c824 | 1353 | } while (next_nid != start_nid); |
20a0307c | 1354 | |
20a0307c WF |
1355 | return ret; |
1356 | } | |
1357 | ||
a5516438 | 1358 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
1359 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
1360 | nodemask_t *nodes_allowed) | |
1da177e4 | 1361 | { |
7893d1d5 | 1362 | unsigned long min_count, ret; |
1da177e4 | 1363 | |
aa888a74 AK |
1364 | if (h->order >= MAX_ORDER) |
1365 | return h->max_huge_pages; | |
1366 | ||
7893d1d5 AL |
1367 | /* |
1368 | * Increase the pool size | |
1369 | * First take pages out of surplus state. Then make up the | |
1370 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f NA |
1371 | * |
1372 | * We might race with alloc_buddy_huge_page() here and be unable | |
1373 | * to convert a surplus huge page to a normal huge page. That is | |
1374 | * not critical, though, it just means the overall size of the | |
1375 | * pool might be one hugepage larger than it needs to be, but | |
1376 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 1377 | */ |
1da177e4 | 1378 | spin_lock(&hugetlb_lock); |
a5516438 | 1379 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 1380 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
1381 | break; |
1382 | } | |
1383 | ||
a5516438 | 1384 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
1385 | /* |
1386 | * If this allocation races such that we no longer need the | |
1387 | * page, free_huge_page will handle it by freeing the page | |
1388 | * and reducing the surplus. | |
1389 | */ | |
1390 | spin_unlock(&hugetlb_lock); | |
6ae11b27 | 1391 | ret = alloc_fresh_huge_page(h, nodes_allowed); |
7893d1d5 AL |
1392 | spin_lock(&hugetlb_lock); |
1393 | if (!ret) | |
1394 | goto out; | |
1395 | ||
536240f2 MG |
1396 | /* Bail for signals. Probably ctrl-c from user */ |
1397 | if (signal_pending(current)) | |
1398 | goto out; | |
7893d1d5 | 1399 | } |
7893d1d5 AL |
1400 | |
1401 | /* | |
1402 | * Decrease the pool size | |
1403 | * First return free pages to the buddy allocator (being careful | |
1404 | * to keep enough around to satisfy reservations). Then place | |
1405 | * pages into surplus state as needed so the pool will shrink | |
1406 | * to the desired size as pages become free. | |
d1c3fb1f NA |
1407 | * |
1408 | * By placing pages into the surplus state independent of the | |
1409 | * overcommit value, we are allowing the surplus pool size to | |
1410 | * exceed overcommit. There are few sane options here. Since | |
1411 | * alloc_buddy_huge_page() is checking the global counter, | |
1412 | * though, we'll note that we're not allowed to exceed surplus | |
1413 | * and won't grow the pool anywhere else. Not until one of the | |
1414 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 1415 | */ |
a5516438 | 1416 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 1417 | min_count = max(count, min_count); |
6ae11b27 | 1418 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 1419 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 1420 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 1421 | break; |
1da177e4 | 1422 | } |
a5516438 | 1423 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 1424 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
1425 | break; |
1426 | } | |
1427 | out: | |
a5516438 | 1428 | ret = persistent_huge_pages(h); |
1da177e4 | 1429 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 1430 | return ret; |
1da177e4 LT |
1431 | } |
1432 | ||
a3437870 NA |
1433 | #define HSTATE_ATTR_RO(_name) \ |
1434 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
1435 | ||
1436 | #define HSTATE_ATTR(_name) \ | |
1437 | static struct kobj_attribute _name##_attr = \ | |
1438 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
1439 | ||
1440 | static struct kobject *hugepages_kobj; | |
1441 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1442 | ||
9a305230 LS |
1443 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
1444 | ||
1445 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
1446 | { |
1447 | int i; | |
9a305230 | 1448 | |
a3437870 | 1449 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
1450 | if (hstate_kobjs[i] == kobj) { |
1451 | if (nidp) | |
1452 | *nidp = NUMA_NO_NODE; | |
a3437870 | 1453 | return &hstates[i]; |
9a305230 LS |
1454 | } |
1455 | ||
1456 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
1457 | } |
1458 | ||
06808b08 | 1459 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
1460 | struct kobj_attribute *attr, char *buf) |
1461 | { | |
9a305230 LS |
1462 | struct hstate *h; |
1463 | unsigned long nr_huge_pages; | |
1464 | int nid; | |
1465 | ||
1466 | h = kobj_to_hstate(kobj, &nid); | |
1467 | if (nid == NUMA_NO_NODE) | |
1468 | nr_huge_pages = h->nr_huge_pages; | |
1469 | else | |
1470 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
1471 | ||
1472 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 1473 | } |
adbe8726 | 1474 | |
06808b08 LS |
1475 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
1476 | struct kobject *kobj, struct kobj_attribute *attr, | |
1477 | const char *buf, size_t len) | |
a3437870 NA |
1478 | { |
1479 | int err; | |
9a305230 | 1480 | int nid; |
06808b08 | 1481 | unsigned long count; |
9a305230 | 1482 | struct hstate *h; |
bad44b5b | 1483 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 1484 | |
06808b08 | 1485 | err = strict_strtoul(buf, 10, &count); |
73ae31e5 | 1486 | if (err) |
adbe8726 | 1487 | goto out; |
a3437870 | 1488 | |
9a305230 | 1489 | h = kobj_to_hstate(kobj, &nid); |
adbe8726 EM |
1490 | if (h->order >= MAX_ORDER) { |
1491 | err = -EINVAL; | |
1492 | goto out; | |
1493 | } | |
1494 | ||
9a305230 LS |
1495 | if (nid == NUMA_NO_NODE) { |
1496 | /* | |
1497 | * global hstate attribute | |
1498 | */ | |
1499 | if (!(obey_mempolicy && | |
1500 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
1501 | NODEMASK_FREE(nodes_allowed); | |
1502 | nodes_allowed = &node_states[N_HIGH_MEMORY]; | |
1503 | } | |
1504 | } else if (nodes_allowed) { | |
1505 | /* | |
1506 | * per node hstate attribute: adjust count to global, | |
1507 | * but restrict alloc/free to the specified node. | |
1508 | */ | |
1509 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
1510 | init_nodemask_of_node(nodes_allowed, nid); | |
1511 | } else | |
1512 | nodes_allowed = &node_states[N_HIGH_MEMORY]; | |
1513 | ||
06808b08 | 1514 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 1515 | |
9b5e5d0f | 1516 | if (nodes_allowed != &node_states[N_HIGH_MEMORY]) |
06808b08 LS |
1517 | NODEMASK_FREE(nodes_allowed); |
1518 | ||
1519 | return len; | |
adbe8726 EM |
1520 | out: |
1521 | NODEMASK_FREE(nodes_allowed); | |
1522 | return err; | |
06808b08 LS |
1523 | } |
1524 | ||
1525 | static ssize_t nr_hugepages_show(struct kobject *kobj, | |
1526 | struct kobj_attribute *attr, char *buf) | |
1527 | { | |
1528 | return nr_hugepages_show_common(kobj, attr, buf); | |
1529 | } | |
1530 | ||
1531 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
1532 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1533 | { | |
1534 | return nr_hugepages_store_common(false, kobj, attr, buf, len); | |
a3437870 NA |
1535 | } |
1536 | HSTATE_ATTR(nr_hugepages); | |
1537 | ||
06808b08 LS |
1538 | #ifdef CONFIG_NUMA |
1539 | ||
1540 | /* | |
1541 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
1542 | * huge page alloc/free. | |
1543 | */ | |
1544 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
1545 | struct kobj_attribute *attr, char *buf) | |
1546 | { | |
1547 | return nr_hugepages_show_common(kobj, attr, buf); | |
1548 | } | |
1549 | ||
1550 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
1551 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1552 | { | |
1553 | return nr_hugepages_store_common(true, kobj, attr, buf, len); | |
1554 | } | |
1555 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
1556 | #endif | |
1557 | ||
1558 | ||
a3437870 NA |
1559 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
1560 | struct kobj_attribute *attr, char *buf) | |
1561 | { | |
9a305230 | 1562 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1563 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
1564 | } | |
adbe8726 | 1565 | |
a3437870 NA |
1566 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
1567 | struct kobj_attribute *attr, const char *buf, size_t count) | |
1568 | { | |
1569 | int err; | |
1570 | unsigned long input; | |
9a305230 | 1571 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 1572 | |
adbe8726 EM |
1573 | if (h->order >= MAX_ORDER) |
1574 | return -EINVAL; | |
1575 | ||
a3437870 NA |
1576 | err = strict_strtoul(buf, 10, &input); |
1577 | if (err) | |
73ae31e5 | 1578 | return err; |
a3437870 NA |
1579 | |
1580 | spin_lock(&hugetlb_lock); | |
1581 | h->nr_overcommit_huge_pages = input; | |
1582 | spin_unlock(&hugetlb_lock); | |
1583 | ||
1584 | return count; | |
1585 | } | |
1586 | HSTATE_ATTR(nr_overcommit_hugepages); | |
1587 | ||
1588 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
1589 | struct kobj_attribute *attr, char *buf) | |
1590 | { | |
9a305230 LS |
1591 | struct hstate *h; |
1592 | unsigned long free_huge_pages; | |
1593 | int nid; | |
1594 | ||
1595 | h = kobj_to_hstate(kobj, &nid); | |
1596 | if (nid == NUMA_NO_NODE) | |
1597 | free_huge_pages = h->free_huge_pages; | |
1598 | else | |
1599 | free_huge_pages = h->free_huge_pages_node[nid]; | |
1600 | ||
1601 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
1602 | } |
1603 | HSTATE_ATTR_RO(free_hugepages); | |
1604 | ||
1605 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
1606 | struct kobj_attribute *attr, char *buf) | |
1607 | { | |
9a305230 | 1608 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1609 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
1610 | } | |
1611 | HSTATE_ATTR_RO(resv_hugepages); | |
1612 | ||
1613 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
1614 | struct kobj_attribute *attr, char *buf) | |
1615 | { | |
9a305230 LS |
1616 | struct hstate *h; |
1617 | unsigned long surplus_huge_pages; | |
1618 | int nid; | |
1619 | ||
1620 | h = kobj_to_hstate(kobj, &nid); | |
1621 | if (nid == NUMA_NO_NODE) | |
1622 | surplus_huge_pages = h->surplus_huge_pages; | |
1623 | else | |
1624 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
1625 | ||
1626 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
1627 | } |
1628 | HSTATE_ATTR_RO(surplus_hugepages); | |
1629 | ||
1630 | static struct attribute *hstate_attrs[] = { | |
1631 | &nr_hugepages_attr.attr, | |
1632 | &nr_overcommit_hugepages_attr.attr, | |
1633 | &free_hugepages_attr.attr, | |
1634 | &resv_hugepages_attr.attr, | |
1635 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
1636 | #ifdef CONFIG_NUMA |
1637 | &nr_hugepages_mempolicy_attr.attr, | |
1638 | #endif | |
a3437870 NA |
1639 | NULL, |
1640 | }; | |
1641 | ||
1642 | static struct attribute_group hstate_attr_group = { | |
1643 | .attrs = hstate_attrs, | |
1644 | }; | |
1645 | ||
094e9539 JM |
1646 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
1647 | struct kobject **hstate_kobjs, | |
1648 | struct attribute_group *hstate_attr_group) | |
a3437870 NA |
1649 | { |
1650 | int retval; | |
972dc4de | 1651 | int hi = hstate_index(h); |
a3437870 | 1652 | |
9a305230 LS |
1653 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
1654 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
1655 | return -ENOMEM; |
1656 | ||
9a305230 | 1657 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 1658 | if (retval) |
9a305230 | 1659 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
1660 | |
1661 | return retval; | |
1662 | } | |
1663 | ||
1664 | static void __init hugetlb_sysfs_init(void) | |
1665 | { | |
1666 | struct hstate *h; | |
1667 | int err; | |
1668 | ||
1669 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
1670 | if (!hugepages_kobj) | |
1671 | return; | |
1672 | ||
1673 | for_each_hstate(h) { | |
9a305230 LS |
1674 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
1675 | hstate_kobjs, &hstate_attr_group); | |
a3437870 NA |
1676 | if (err) |
1677 | printk(KERN_ERR "Hugetlb: Unable to add hstate %s", | |
1678 | h->name); | |
1679 | } | |
1680 | } | |
1681 | ||
9a305230 LS |
1682 | #ifdef CONFIG_NUMA |
1683 | ||
1684 | /* | |
1685 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
1686 | * with node devices in node_devices[] using a parallel array. The array |
1687 | * index of a node device or _hstate == node id. | |
1688 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
1689 | * the base kernel, on the hugetlb module. |
1690 | */ | |
1691 | struct node_hstate { | |
1692 | struct kobject *hugepages_kobj; | |
1693 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1694 | }; | |
1695 | struct node_hstate node_hstates[MAX_NUMNODES]; | |
1696 | ||
1697 | /* | |
10fbcf4c | 1698 | * A subset of global hstate attributes for node devices |
9a305230 LS |
1699 | */ |
1700 | static struct attribute *per_node_hstate_attrs[] = { | |
1701 | &nr_hugepages_attr.attr, | |
1702 | &free_hugepages_attr.attr, | |
1703 | &surplus_hugepages_attr.attr, | |
1704 | NULL, | |
1705 | }; | |
1706 | ||
1707 | static struct attribute_group per_node_hstate_attr_group = { | |
1708 | .attrs = per_node_hstate_attrs, | |
1709 | }; | |
1710 | ||
1711 | /* | |
10fbcf4c | 1712 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
1713 | * Returns node id via non-NULL nidp. |
1714 | */ | |
1715 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
1716 | { | |
1717 | int nid; | |
1718 | ||
1719 | for (nid = 0; nid < nr_node_ids; nid++) { | |
1720 | struct node_hstate *nhs = &node_hstates[nid]; | |
1721 | int i; | |
1722 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
1723 | if (nhs->hstate_kobjs[i] == kobj) { | |
1724 | if (nidp) | |
1725 | *nidp = nid; | |
1726 | return &hstates[i]; | |
1727 | } | |
1728 | } | |
1729 | ||
1730 | BUG(); | |
1731 | return NULL; | |
1732 | } | |
1733 | ||
1734 | /* | |
10fbcf4c | 1735 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
1736 | * No-op if no hstate attributes attached. |
1737 | */ | |
1738 | void hugetlb_unregister_node(struct node *node) | |
1739 | { | |
1740 | struct hstate *h; | |
10fbcf4c | 1741 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
1742 | |
1743 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 1744 | return; /* no hstate attributes */ |
9a305230 | 1745 | |
972dc4de AK |
1746 | for_each_hstate(h) { |
1747 | int idx = hstate_index(h); | |
1748 | if (nhs->hstate_kobjs[idx]) { | |
1749 | kobject_put(nhs->hstate_kobjs[idx]); | |
1750 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 1751 | } |
972dc4de | 1752 | } |
9a305230 LS |
1753 | |
1754 | kobject_put(nhs->hugepages_kobj); | |
1755 | nhs->hugepages_kobj = NULL; | |
1756 | } | |
1757 | ||
1758 | /* | |
10fbcf4c | 1759 | * hugetlb module exit: unregister hstate attributes from node devices |
9a305230 LS |
1760 | * that have them. |
1761 | */ | |
1762 | static void hugetlb_unregister_all_nodes(void) | |
1763 | { | |
1764 | int nid; | |
1765 | ||
1766 | /* | |
10fbcf4c | 1767 | * disable node device registrations. |
9a305230 LS |
1768 | */ |
1769 | register_hugetlbfs_with_node(NULL, NULL); | |
1770 | ||
1771 | /* | |
1772 | * remove hstate attributes from any nodes that have them. | |
1773 | */ | |
1774 | for (nid = 0; nid < nr_node_ids; nid++) | |
1775 | hugetlb_unregister_node(&node_devices[nid]); | |
1776 | } | |
1777 | ||
1778 | /* | |
10fbcf4c | 1779 | * Register hstate attributes for a single node device. |
9a305230 LS |
1780 | * No-op if attributes already registered. |
1781 | */ | |
1782 | void hugetlb_register_node(struct node *node) | |
1783 | { | |
1784 | struct hstate *h; | |
10fbcf4c | 1785 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
1786 | int err; |
1787 | ||
1788 | if (nhs->hugepages_kobj) | |
1789 | return; /* already allocated */ | |
1790 | ||
1791 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 1792 | &node->dev.kobj); |
9a305230 LS |
1793 | if (!nhs->hugepages_kobj) |
1794 | return; | |
1795 | ||
1796 | for_each_hstate(h) { | |
1797 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
1798 | nhs->hstate_kobjs, | |
1799 | &per_node_hstate_attr_group); | |
1800 | if (err) { | |
1801 | printk(KERN_ERR "Hugetlb: Unable to add hstate %s" | |
1802 | " for node %d\n", | |
10fbcf4c | 1803 | h->name, node->dev.id); |
9a305230 LS |
1804 | hugetlb_unregister_node(node); |
1805 | break; | |
1806 | } | |
1807 | } | |
1808 | } | |
1809 | ||
1810 | /* | |
9b5e5d0f | 1811 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
1812 | * devices of nodes that have memory. All on-line nodes should have |
1813 | * registered their associated device by this time. | |
9a305230 LS |
1814 | */ |
1815 | static void hugetlb_register_all_nodes(void) | |
1816 | { | |
1817 | int nid; | |
1818 | ||
9b5e5d0f | 1819 | for_each_node_state(nid, N_HIGH_MEMORY) { |
9a305230 | 1820 | struct node *node = &node_devices[nid]; |
10fbcf4c | 1821 | if (node->dev.id == nid) |
9a305230 LS |
1822 | hugetlb_register_node(node); |
1823 | } | |
1824 | ||
1825 | /* | |
10fbcf4c | 1826 | * Let the node device driver know we're here so it can |
9a305230 LS |
1827 | * [un]register hstate attributes on node hotplug. |
1828 | */ | |
1829 | register_hugetlbfs_with_node(hugetlb_register_node, | |
1830 | hugetlb_unregister_node); | |
1831 | } | |
1832 | #else /* !CONFIG_NUMA */ | |
1833 | ||
1834 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
1835 | { | |
1836 | BUG(); | |
1837 | if (nidp) | |
1838 | *nidp = -1; | |
1839 | return NULL; | |
1840 | } | |
1841 | ||
1842 | static void hugetlb_unregister_all_nodes(void) { } | |
1843 | ||
1844 | static void hugetlb_register_all_nodes(void) { } | |
1845 | ||
1846 | #endif | |
1847 | ||
a3437870 NA |
1848 | static void __exit hugetlb_exit(void) |
1849 | { | |
1850 | struct hstate *h; | |
1851 | ||
9a305230 LS |
1852 | hugetlb_unregister_all_nodes(); |
1853 | ||
a3437870 | 1854 | for_each_hstate(h) { |
972dc4de | 1855 | kobject_put(hstate_kobjs[hstate_index(h)]); |
a3437870 NA |
1856 | } |
1857 | ||
1858 | kobject_put(hugepages_kobj); | |
1859 | } | |
1860 | module_exit(hugetlb_exit); | |
1861 | ||
1862 | static int __init hugetlb_init(void) | |
1863 | { | |
0ef89d25 BH |
1864 | /* Some platform decide whether they support huge pages at boot |
1865 | * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when | |
1866 | * there is no such support | |
1867 | */ | |
1868 | if (HPAGE_SHIFT == 0) | |
1869 | return 0; | |
a3437870 | 1870 | |
e11bfbfc NP |
1871 | if (!size_to_hstate(default_hstate_size)) { |
1872 | default_hstate_size = HPAGE_SIZE; | |
1873 | if (!size_to_hstate(default_hstate_size)) | |
1874 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 1875 | } |
972dc4de | 1876 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
e11bfbfc NP |
1877 | if (default_hstate_max_huge_pages) |
1878 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
a3437870 NA |
1879 | |
1880 | hugetlb_init_hstates(); | |
1881 | ||
aa888a74 AK |
1882 | gather_bootmem_prealloc(); |
1883 | ||
a3437870 NA |
1884 | report_hugepages(); |
1885 | ||
1886 | hugetlb_sysfs_init(); | |
1887 | ||
9a305230 LS |
1888 | hugetlb_register_all_nodes(); |
1889 | ||
a3437870 NA |
1890 | return 0; |
1891 | } | |
1892 | module_init(hugetlb_init); | |
1893 | ||
1894 | /* Should be called on processing a hugepagesz=... option */ | |
1895 | void __init hugetlb_add_hstate(unsigned order) | |
1896 | { | |
1897 | struct hstate *h; | |
8faa8b07 AK |
1898 | unsigned long i; |
1899 | ||
a3437870 NA |
1900 | if (size_to_hstate(PAGE_SIZE << order)) { |
1901 | printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n"); | |
1902 | return; | |
1903 | } | |
47d38344 | 1904 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 1905 | BUG_ON(order == 0); |
47d38344 | 1906 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
1907 | h->order = order; |
1908 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
1909 | h->nr_huge_pages = 0; |
1910 | h->free_huge_pages = 0; | |
1911 | for (i = 0; i < MAX_NUMNODES; ++i) | |
1912 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 1913 | INIT_LIST_HEAD(&h->hugepage_activelist); |
9b5e5d0f LS |
1914 | h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]); |
1915 | h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]); | |
a3437870 NA |
1916 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
1917 | huge_page_size(h)/1024); | |
8faa8b07 | 1918 | |
a3437870 NA |
1919 | parsed_hstate = h; |
1920 | } | |
1921 | ||
e11bfbfc | 1922 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
1923 | { |
1924 | unsigned long *mhp; | |
8faa8b07 | 1925 | static unsigned long *last_mhp; |
a3437870 NA |
1926 | |
1927 | /* | |
47d38344 | 1928 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
1929 | * so this hugepages= parameter goes to the "default hstate". |
1930 | */ | |
47d38344 | 1931 | if (!hugetlb_max_hstate) |
a3437870 NA |
1932 | mhp = &default_hstate_max_huge_pages; |
1933 | else | |
1934 | mhp = &parsed_hstate->max_huge_pages; | |
1935 | ||
8faa8b07 AK |
1936 | if (mhp == last_mhp) { |
1937 | printk(KERN_WARNING "hugepages= specified twice without " | |
1938 | "interleaving hugepagesz=, ignoring\n"); | |
1939 | return 1; | |
1940 | } | |
1941 | ||
a3437870 NA |
1942 | if (sscanf(s, "%lu", mhp) <= 0) |
1943 | *mhp = 0; | |
1944 | ||
8faa8b07 AK |
1945 | /* |
1946 | * Global state is always initialized later in hugetlb_init. | |
1947 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
1948 | * use the bootmem allocator. | |
1949 | */ | |
47d38344 | 1950 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
1951 | hugetlb_hstate_alloc_pages(parsed_hstate); |
1952 | ||
1953 | last_mhp = mhp; | |
1954 | ||
a3437870 NA |
1955 | return 1; |
1956 | } | |
e11bfbfc NP |
1957 | __setup("hugepages=", hugetlb_nrpages_setup); |
1958 | ||
1959 | static int __init hugetlb_default_setup(char *s) | |
1960 | { | |
1961 | default_hstate_size = memparse(s, &s); | |
1962 | return 1; | |
1963 | } | |
1964 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 1965 | |
8a213460 NA |
1966 | static unsigned int cpuset_mems_nr(unsigned int *array) |
1967 | { | |
1968 | int node; | |
1969 | unsigned int nr = 0; | |
1970 | ||
1971 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
1972 | nr += array[node]; | |
1973 | ||
1974 | return nr; | |
1975 | } | |
1976 | ||
1977 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
1978 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
1979 | struct ctl_table *table, int write, | |
1980 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 1981 | { |
e5ff2159 AK |
1982 | struct hstate *h = &default_hstate; |
1983 | unsigned long tmp; | |
08d4a246 | 1984 | int ret; |
e5ff2159 | 1985 | |
c033a93c | 1986 | tmp = h->max_huge_pages; |
e5ff2159 | 1987 | |
adbe8726 EM |
1988 | if (write && h->order >= MAX_ORDER) |
1989 | return -EINVAL; | |
1990 | ||
e5ff2159 AK |
1991 | table->data = &tmp; |
1992 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
1993 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
1994 | if (ret) | |
1995 | goto out; | |
e5ff2159 | 1996 | |
06808b08 | 1997 | if (write) { |
bad44b5b DR |
1998 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, |
1999 | GFP_KERNEL | __GFP_NORETRY); | |
06808b08 LS |
2000 | if (!(obey_mempolicy && |
2001 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
2002 | NODEMASK_FREE(nodes_allowed); | |
2003 | nodes_allowed = &node_states[N_HIGH_MEMORY]; | |
2004 | } | |
2005 | h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); | |
2006 | ||
2007 | if (nodes_allowed != &node_states[N_HIGH_MEMORY]) | |
2008 | NODEMASK_FREE(nodes_allowed); | |
2009 | } | |
08d4a246 MH |
2010 | out: |
2011 | return ret; | |
1da177e4 | 2012 | } |
396faf03 | 2013 | |
06808b08 LS |
2014 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2015 | void __user *buffer, size_t *length, loff_t *ppos) | |
2016 | { | |
2017 | ||
2018 | return hugetlb_sysctl_handler_common(false, table, write, | |
2019 | buffer, length, ppos); | |
2020 | } | |
2021 | ||
2022 | #ifdef CONFIG_NUMA | |
2023 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2024 | void __user *buffer, size_t *length, loff_t *ppos) | |
2025 | { | |
2026 | return hugetlb_sysctl_handler_common(true, table, write, | |
2027 | buffer, length, ppos); | |
2028 | } | |
2029 | #endif /* CONFIG_NUMA */ | |
2030 | ||
396faf03 | 2031 | int hugetlb_treat_movable_handler(struct ctl_table *table, int write, |
8d65af78 | 2032 | void __user *buffer, |
396faf03 MG |
2033 | size_t *length, loff_t *ppos) |
2034 | { | |
8d65af78 | 2035 | proc_dointvec(table, write, buffer, length, ppos); |
396faf03 MG |
2036 | if (hugepages_treat_as_movable) |
2037 | htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; | |
2038 | else | |
2039 | htlb_alloc_mask = GFP_HIGHUSER; | |
2040 | return 0; | |
2041 | } | |
2042 | ||
a3d0c6aa | 2043 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 2044 | void __user *buffer, |
a3d0c6aa NA |
2045 | size_t *length, loff_t *ppos) |
2046 | { | |
a5516438 | 2047 | struct hstate *h = &default_hstate; |
e5ff2159 | 2048 | unsigned long tmp; |
08d4a246 | 2049 | int ret; |
e5ff2159 | 2050 | |
c033a93c | 2051 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 2052 | |
adbe8726 EM |
2053 | if (write && h->order >= MAX_ORDER) |
2054 | return -EINVAL; | |
2055 | ||
e5ff2159 AK |
2056 | table->data = &tmp; |
2057 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2058 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2059 | if (ret) | |
2060 | goto out; | |
e5ff2159 AK |
2061 | |
2062 | if (write) { | |
2063 | spin_lock(&hugetlb_lock); | |
2064 | h->nr_overcommit_huge_pages = tmp; | |
2065 | spin_unlock(&hugetlb_lock); | |
2066 | } | |
08d4a246 MH |
2067 | out: |
2068 | return ret; | |
a3d0c6aa NA |
2069 | } |
2070 | ||
1da177e4 LT |
2071 | #endif /* CONFIG_SYSCTL */ |
2072 | ||
e1759c21 | 2073 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 2074 | { |
a5516438 | 2075 | struct hstate *h = &default_hstate; |
e1759c21 | 2076 | seq_printf(m, |
4f98a2fe RR |
2077 | "HugePages_Total: %5lu\n" |
2078 | "HugePages_Free: %5lu\n" | |
2079 | "HugePages_Rsvd: %5lu\n" | |
2080 | "HugePages_Surp: %5lu\n" | |
2081 | "Hugepagesize: %8lu kB\n", | |
a5516438 AK |
2082 | h->nr_huge_pages, |
2083 | h->free_huge_pages, | |
2084 | h->resv_huge_pages, | |
2085 | h->surplus_huge_pages, | |
2086 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
2087 | } |
2088 | ||
2089 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
2090 | { | |
a5516438 | 2091 | struct hstate *h = &default_hstate; |
1da177e4 LT |
2092 | return sprintf(buf, |
2093 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
2094 | "Node %d HugePages_Free: %5u\n" |
2095 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
2096 | nid, h->nr_huge_pages_node[nid], |
2097 | nid, h->free_huge_pages_node[nid], | |
2098 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
2099 | } |
2100 | ||
1da177e4 LT |
2101 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
2102 | unsigned long hugetlb_total_pages(void) | |
2103 | { | |
a5516438 AK |
2104 | struct hstate *h = &default_hstate; |
2105 | return h->nr_huge_pages * pages_per_huge_page(h); | |
1da177e4 | 2106 | } |
1da177e4 | 2107 | |
a5516438 | 2108 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
2109 | { |
2110 | int ret = -ENOMEM; | |
2111 | ||
2112 | spin_lock(&hugetlb_lock); | |
2113 | /* | |
2114 | * When cpuset is configured, it breaks the strict hugetlb page | |
2115 | * reservation as the accounting is done on a global variable. Such | |
2116 | * reservation is completely rubbish in the presence of cpuset because | |
2117 | * the reservation is not checked against page availability for the | |
2118 | * current cpuset. Application can still potentially OOM'ed by kernel | |
2119 | * with lack of free htlb page in cpuset that the task is in. | |
2120 | * Attempt to enforce strict accounting with cpuset is almost | |
2121 | * impossible (or too ugly) because cpuset is too fluid that | |
2122 | * task or memory node can be dynamically moved between cpusets. | |
2123 | * | |
2124 | * The change of semantics for shared hugetlb mapping with cpuset is | |
2125 | * undesirable. However, in order to preserve some of the semantics, | |
2126 | * we fall back to check against current free page availability as | |
2127 | * a best attempt and hopefully to minimize the impact of changing | |
2128 | * semantics that cpuset has. | |
2129 | */ | |
2130 | if (delta > 0) { | |
a5516438 | 2131 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
2132 | goto out; |
2133 | ||
a5516438 AK |
2134 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
2135 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
2136 | goto out; |
2137 | } | |
2138 | } | |
2139 | ||
2140 | ret = 0; | |
2141 | if (delta < 0) | |
a5516438 | 2142 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
2143 | |
2144 | out: | |
2145 | spin_unlock(&hugetlb_lock); | |
2146 | return ret; | |
2147 | } | |
2148 | ||
84afd99b AW |
2149 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
2150 | { | |
2151 | struct resv_map *reservations = vma_resv_map(vma); | |
2152 | ||
2153 | /* | |
2154 | * This new VMA should share its siblings reservation map if present. | |
2155 | * The VMA will only ever have a valid reservation map pointer where | |
2156 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 2157 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
2158 | * after this open call completes. It is therefore safe to take a |
2159 | * new reference here without additional locking. | |
2160 | */ | |
2161 | if (reservations) | |
2162 | kref_get(&reservations->refs); | |
2163 | } | |
2164 | ||
c50ac050 DH |
2165 | static void resv_map_put(struct vm_area_struct *vma) |
2166 | { | |
2167 | struct resv_map *reservations = vma_resv_map(vma); | |
2168 | ||
2169 | if (!reservations) | |
2170 | return; | |
2171 | kref_put(&reservations->refs, resv_map_release); | |
2172 | } | |
2173 | ||
a1e78772 MG |
2174 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
2175 | { | |
a5516438 | 2176 | struct hstate *h = hstate_vma(vma); |
84afd99b | 2177 | struct resv_map *reservations = vma_resv_map(vma); |
90481622 | 2178 | struct hugepage_subpool *spool = subpool_vma(vma); |
84afd99b AW |
2179 | unsigned long reserve; |
2180 | unsigned long start; | |
2181 | unsigned long end; | |
2182 | ||
2183 | if (reservations) { | |
a5516438 AK |
2184 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
2185 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b AW |
2186 | |
2187 | reserve = (end - start) - | |
2188 | region_count(&reservations->regions, start, end); | |
2189 | ||
c50ac050 | 2190 | resv_map_put(vma); |
84afd99b | 2191 | |
7251ff78 | 2192 | if (reserve) { |
a5516438 | 2193 | hugetlb_acct_memory(h, -reserve); |
90481622 | 2194 | hugepage_subpool_put_pages(spool, reserve); |
7251ff78 | 2195 | } |
84afd99b | 2196 | } |
a1e78772 MG |
2197 | } |
2198 | ||
1da177e4 LT |
2199 | /* |
2200 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
2201 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
2202 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
2203 | * this far. | |
2204 | */ | |
d0217ac0 | 2205 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
2206 | { |
2207 | BUG(); | |
d0217ac0 | 2208 | return 0; |
1da177e4 LT |
2209 | } |
2210 | ||
f0f37e2f | 2211 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 2212 | .fault = hugetlb_vm_op_fault, |
84afd99b | 2213 | .open = hugetlb_vm_op_open, |
a1e78772 | 2214 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
2215 | }; |
2216 | ||
1e8f889b DG |
2217 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
2218 | int writable) | |
63551ae0 DG |
2219 | { |
2220 | pte_t entry; | |
2221 | ||
1e8f889b | 2222 | if (writable) { |
63551ae0 DG |
2223 | entry = |
2224 | pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); | |
2225 | } else { | |
7f2e9525 | 2226 | entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot)); |
63551ae0 DG |
2227 | } |
2228 | entry = pte_mkyoung(entry); | |
2229 | entry = pte_mkhuge(entry); | |
d9ed9faa | 2230 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
2231 | |
2232 | return entry; | |
2233 | } | |
2234 | ||
1e8f889b DG |
2235 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
2236 | unsigned long address, pte_t *ptep) | |
2237 | { | |
2238 | pte_t entry; | |
2239 | ||
7f2e9525 | 2240 | entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 2241 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 2242 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
2243 | } |
2244 | ||
2245 | ||
63551ae0 DG |
2246 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
2247 | struct vm_area_struct *vma) | |
2248 | { | |
2249 | pte_t *src_pte, *dst_pte, entry; | |
2250 | struct page *ptepage; | |
1c59827d | 2251 | unsigned long addr; |
1e8f889b | 2252 | int cow; |
a5516438 AK |
2253 | struct hstate *h = hstate_vma(vma); |
2254 | unsigned long sz = huge_page_size(h); | |
1e8f889b DG |
2255 | |
2256 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 2257 | |
a5516438 | 2258 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
c74df32c HD |
2259 | src_pte = huge_pte_offset(src, addr); |
2260 | if (!src_pte) | |
2261 | continue; | |
a5516438 | 2262 | dst_pte = huge_pte_alloc(dst, addr, sz); |
63551ae0 DG |
2263 | if (!dst_pte) |
2264 | goto nomem; | |
c5c99429 LW |
2265 | |
2266 | /* If the pagetables are shared don't copy or take references */ | |
2267 | if (dst_pte == src_pte) | |
2268 | continue; | |
2269 | ||
c74df32c | 2270 | spin_lock(&dst->page_table_lock); |
46478758 | 2271 | spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); |
7f2e9525 | 2272 | if (!huge_pte_none(huge_ptep_get(src_pte))) { |
1e8f889b | 2273 | if (cow) |
7f2e9525 GS |
2274 | huge_ptep_set_wrprotect(src, addr, src_pte); |
2275 | entry = huge_ptep_get(src_pte); | |
1c59827d HD |
2276 | ptepage = pte_page(entry); |
2277 | get_page(ptepage); | |
0fe6e20b | 2278 | page_dup_rmap(ptepage); |
1c59827d HD |
2279 | set_huge_pte_at(dst, addr, dst_pte, entry); |
2280 | } | |
2281 | spin_unlock(&src->page_table_lock); | |
c74df32c | 2282 | spin_unlock(&dst->page_table_lock); |
63551ae0 DG |
2283 | } |
2284 | return 0; | |
2285 | ||
2286 | nomem: | |
2287 | return -ENOMEM; | |
2288 | } | |
2289 | ||
290408d4 NH |
2290 | static int is_hugetlb_entry_migration(pte_t pte) |
2291 | { | |
2292 | swp_entry_t swp; | |
2293 | ||
2294 | if (huge_pte_none(pte) || pte_present(pte)) | |
2295 | return 0; | |
2296 | swp = pte_to_swp_entry(pte); | |
32f84528 | 2297 | if (non_swap_entry(swp) && is_migration_entry(swp)) |
290408d4 | 2298 | return 1; |
32f84528 | 2299 | else |
290408d4 NH |
2300 | return 0; |
2301 | } | |
2302 | ||
fd6a03ed NH |
2303 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) |
2304 | { | |
2305 | swp_entry_t swp; | |
2306 | ||
2307 | if (huge_pte_none(pte) || pte_present(pte)) | |
2308 | return 0; | |
2309 | swp = pte_to_swp_entry(pte); | |
32f84528 | 2310 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) |
fd6a03ed | 2311 | return 1; |
32f84528 | 2312 | else |
fd6a03ed NH |
2313 | return 0; |
2314 | } | |
2315 | ||
24669e58 AK |
2316 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
2317 | unsigned long start, unsigned long end, | |
2318 | struct page *ref_page) | |
63551ae0 | 2319 | { |
24669e58 | 2320 | int force_flush = 0; |
63551ae0 DG |
2321 | struct mm_struct *mm = vma->vm_mm; |
2322 | unsigned long address; | |
c7546f8f | 2323 | pte_t *ptep; |
63551ae0 DG |
2324 | pte_t pte; |
2325 | struct page *page; | |
a5516438 AK |
2326 | struct hstate *h = hstate_vma(vma); |
2327 | unsigned long sz = huge_page_size(h); | |
2328 | ||
63551ae0 | 2329 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
2330 | BUG_ON(start & ~huge_page_mask(h)); |
2331 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 2332 | |
24669e58 | 2333 | tlb_start_vma(tlb, vma); |
cddb8a5c | 2334 | mmu_notifier_invalidate_range_start(mm, start, end); |
24669e58 | 2335 | again: |
508034a3 | 2336 | spin_lock(&mm->page_table_lock); |
a5516438 | 2337 | for (address = start; address < end; address += sz) { |
c7546f8f | 2338 | ptep = huge_pte_offset(mm, address); |
4c887265 | 2339 | if (!ptep) |
c7546f8f DG |
2340 | continue; |
2341 | ||
39dde65c KC |
2342 | if (huge_pmd_unshare(mm, &address, ptep)) |
2343 | continue; | |
2344 | ||
6629326b HD |
2345 | pte = huge_ptep_get(ptep); |
2346 | if (huge_pte_none(pte)) | |
2347 | continue; | |
2348 | ||
2349 | /* | |
2350 | * HWPoisoned hugepage is already unmapped and dropped reference | |
2351 | */ | |
2352 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) | |
2353 | continue; | |
2354 | ||
2355 | page = pte_page(pte); | |
04f2cbe3 MG |
2356 | /* |
2357 | * If a reference page is supplied, it is because a specific | |
2358 | * page is being unmapped, not a range. Ensure the page we | |
2359 | * are about to unmap is the actual page of interest. | |
2360 | */ | |
2361 | if (ref_page) { | |
04f2cbe3 MG |
2362 | if (page != ref_page) |
2363 | continue; | |
2364 | ||
2365 | /* | |
2366 | * Mark the VMA as having unmapped its page so that | |
2367 | * future faults in this VMA will fail rather than | |
2368 | * looking like data was lost | |
2369 | */ | |
2370 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
2371 | } | |
2372 | ||
c7546f8f | 2373 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
24669e58 | 2374 | tlb_remove_tlb_entry(tlb, ptep, address); |
6649a386 KC |
2375 | if (pte_dirty(pte)) |
2376 | set_page_dirty(page); | |
9e81130b | 2377 | |
24669e58 AK |
2378 | page_remove_rmap(page); |
2379 | force_flush = !__tlb_remove_page(tlb, page); | |
2380 | if (force_flush) | |
2381 | break; | |
9e81130b HD |
2382 | /* Bail out after unmapping reference page if supplied */ |
2383 | if (ref_page) | |
2384 | break; | |
63551ae0 | 2385 | } |
cd2934a3 | 2386 | spin_unlock(&mm->page_table_lock); |
24669e58 AK |
2387 | /* |
2388 | * mmu_gather ran out of room to batch pages, we break out of | |
2389 | * the PTE lock to avoid doing the potential expensive TLB invalidate | |
2390 | * and page-free while holding it. | |
2391 | */ | |
2392 | if (force_flush) { | |
2393 | force_flush = 0; | |
2394 | tlb_flush_mmu(tlb); | |
2395 | if (address < end && !ref_page) | |
2396 | goto again; | |
fe1668ae | 2397 | } |
24669e58 AK |
2398 | mmu_notifier_invalidate_range_end(mm, start, end); |
2399 | tlb_end_vma(tlb, vma); | |
1da177e4 | 2400 | } |
63551ae0 | 2401 | |
502717f4 | 2402 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 2403 | unsigned long end, struct page *ref_page) |
502717f4 | 2404 | { |
24669e58 AK |
2405 | struct mm_struct *mm; |
2406 | struct mmu_gather tlb; | |
2407 | ||
2408 | mm = vma->vm_mm; | |
2409 | ||
2410 | tlb_gather_mmu(&tlb, mm, 0); | |
2411 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); | |
2412 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 KC |
2413 | } |
2414 | ||
04f2cbe3 MG |
2415 | /* |
2416 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
2417 | * mappping it owns the reserve page for. The intention is to unmap the page | |
2418 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
2419 | * same region. | |
2420 | */ | |
2a4b3ded HH |
2421 | static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
2422 | struct page *page, unsigned long address) | |
04f2cbe3 | 2423 | { |
7526674d | 2424 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
2425 | struct vm_area_struct *iter_vma; |
2426 | struct address_space *mapping; | |
2427 | struct prio_tree_iter iter; | |
2428 | pgoff_t pgoff; | |
2429 | ||
2430 | /* | |
2431 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
2432 | * from page cache lookup which is in HPAGE_SIZE units. | |
2433 | */ | |
7526674d | 2434 | address = address & huge_page_mask(h); |
0c176d52 | 2435 | pgoff = vma_hugecache_offset(h, vma, address); |
90481622 | 2436 | mapping = vma->vm_file->f_dentry->d_inode->i_mapping; |
04f2cbe3 | 2437 | |
4eb2b1dc MG |
2438 | /* |
2439 | * Take the mapping lock for the duration of the table walk. As | |
2440 | * this mapping should be shared between all the VMAs, | |
2441 | * __unmap_hugepage_range() is called as the lock is already held | |
2442 | */ | |
3d48ae45 | 2443 | mutex_lock(&mapping->i_mmap_mutex); |
04f2cbe3 MG |
2444 | vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) { |
2445 | /* Do not unmap the current VMA */ | |
2446 | if (iter_vma == vma) | |
2447 | continue; | |
2448 | ||
2449 | /* | |
2450 | * Unmap the page from other VMAs without their own reserves. | |
2451 | * They get marked to be SIGKILLed if they fault in these | |
2452 | * areas. This is because a future no-page fault on this VMA | |
2453 | * could insert a zeroed page instead of the data existing | |
2454 | * from the time of fork. This would look like data corruption | |
2455 | */ | |
2456 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
2457 | unmap_hugepage_range(iter_vma, address, |
2458 | address + huge_page_size(h), page); | |
04f2cbe3 | 2459 | } |
3d48ae45 | 2460 | mutex_unlock(&mapping->i_mmap_mutex); |
04f2cbe3 MG |
2461 | |
2462 | return 1; | |
2463 | } | |
2464 | ||
0fe6e20b NH |
2465 | /* |
2466 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
2467 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
2468 | * cannot race with other handlers or page migration. | |
2469 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 2470 | */ |
1e8f889b | 2471 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
04f2cbe3 MG |
2472 | unsigned long address, pte_t *ptep, pte_t pte, |
2473 | struct page *pagecache_page) | |
1e8f889b | 2474 | { |
a5516438 | 2475 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 2476 | struct page *old_page, *new_page; |
79ac6ba4 | 2477 | int avoidcopy; |
04f2cbe3 | 2478 | int outside_reserve = 0; |
1e8f889b DG |
2479 | |
2480 | old_page = pte_page(pte); | |
2481 | ||
04f2cbe3 | 2482 | retry_avoidcopy: |
1e8f889b DG |
2483 | /* If no-one else is actually using this page, avoid the copy |
2484 | * and just make the page writable */ | |
0fe6e20b | 2485 | avoidcopy = (page_mapcount(old_page) == 1); |
1e8f889b | 2486 | if (avoidcopy) { |
56c9cfb1 NH |
2487 | if (PageAnon(old_page)) |
2488 | page_move_anon_rmap(old_page, vma, address); | |
1e8f889b | 2489 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 2490 | return 0; |
1e8f889b DG |
2491 | } |
2492 | ||
04f2cbe3 MG |
2493 | /* |
2494 | * If the process that created a MAP_PRIVATE mapping is about to | |
2495 | * perform a COW due to a shared page count, attempt to satisfy | |
2496 | * the allocation without using the existing reserves. The pagecache | |
2497 | * page is used to determine if the reserve at this address was | |
2498 | * consumed or not. If reserves were used, a partial faulted mapping | |
2499 | * at the time of fork() could consume its reserves on COW instead | |
2500 | * of the full address range. | |
2501 | */ | |
f83a275d | 2502 | if (!(vma->vm_flags & VM_MAYSHARE) && |
04f2cbe3 MG |
2503 | is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
2504 | old_page != pagecache_page) | |
2505 | outside_reserve = 1; | |
2506 | ||
1e8f889b | 2507 | page_cache_get(old_page); |
b76c8cfb LW |
2508 | |
2509 | /* Drop page_table_lock as buddy allocator may be called */ | |
2510 | spin_unlock(&mm->page_table_lock); | |
04f2cbe3 | 2511 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 2512 | |
2fc39cec | 2513 | if (IS_ERR(new_page)) { |
76dcee75 | 2514 | long err = PTR_ERR(new_page); |
1e8f889b | 2515 | page_cache_release(old_page); |
04f2cbe3 MG |
2516 | |
2517 | /* | |
2518 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
2519 | * it is due to references held by a child and an insufficient | |
2520 | * huge page pool. To guarantee the original mappers | |
2521 | * reliability, unmap the page from child processes. The child | |
2522 | * may get SIGKILLed if it later faults. | |
2523 | */ | |
2524 | if (outside_reserve) { | |
2525 | BUG_ON(huge_pte_none(pte)); | |
2526 | if (unmap_ref_private(mm, vma, old_page, address)) { | |
04f2cbe3 | 2527 | BUG_ON(huge_pte_none(pte)); |
b76c8cfb | 2528 | spin_lock(&mm->page_table_lock); |
a734bcc8 HD |
2529 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
2530 | if (likely(pte_same(huge_ptep_get(ptep), pte))) | |
2531 | goto retry_avoidcopy; | |
2532 | /* | |
2533 | * race occurs while re-acquiring page_table_lock, and | |
2534 | * our job is done. | |
2535 | */ | |
2536 | return 0; | |
04f2cbe3 MG |
2537 | } |
2538 | WARN_ON_ONCE(1); | |
2539 | } | |
2540 | ||
b76c8cfb LW |
2541 | /* Caller expects lock to be held */ |
2542 | spin_lock(&mm->page_table_lock); | |
76dcee75 AK |
2543 | if (err == -ENOMEM) |
2544 | return VM_FAULT_OOM; | |
2545 | else | |
2546 | return VM_FAULT_SIGBUS; | |
1e8f889b DG |
2547 | } |
2548 | ||
0fe6e20b NH |
2549 | /* |
2550 | * When the original hugepage is shared one, it does not have | |
2551 | * anon_vma prepared. | |
2552 | */ | |
44e2aa93 | 2553 | if (unlikely(anon_vma_prepare(vma))) { |
ea4039a3 HD |
2554 | page_cache_release(new_page); |
2555 | page_cache_release(old_page); | |
44e2aa93 DN |
2556 | /* Caller expects lock to be held */ |
2557 | spin_lock(&mm->page_table_lock); | |
0fe6e20b | 2558 | return VM_FAULT_OOM; |
44e2aa93 | 2559 | } |
0fe6e20b | 2560 | |
47ad8475 AA |
2561 | copy_user_huge_page(new_page, old_page, address, vma, |
2562 | pages_per_huge_page(h)); | |
0ed361de | 2563 | __SetPageUptodate(new_page); |
1e8f889b | 2564 | |
b76c8cfb LW |
2565 | /* |
2566 | * Retake the page_table_lock to check for racing updates | |
2567 | * before the page tables are altered | |
2568 | */ | |
2569 | spin_lock(&mm->page_table_lock); | |
a5516438 | 2570 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
7f2e9525 | 2571 | if (likely(pte_same(huge_ptep_get(ptep), pte))) { |
1e8f889b | 2572 | /* Break COW */ |
3edd4fc9 DD |
2573 | mmu_notifier_invalidate_range_start(mm, |
2574 | address & huge_page_mask(h), | |
2575 | (address & huge_page_mask(h)) + huge_page_size(h)); | |
8fe627ec | 2576 | huge_ptep_clear_flush(vma, address, ptep); |
1e8f889b DG |
2577 | set_huge_pte_at(mm, address, ptep, |
2578 | make_huge_pte(vma, new_page, 1)); | |
0fe6e20b | 2579 | page_remove_rmap(old_page); |
cd67f0d2 | 2580 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
2581 | /* Make the old page be freed below */ |
2582 | new_page = old_page; | |
3edd4fc9 DD |
2583 | mmu_notifier_invalidate_range_end(mm, |
2584 | address & huge_page_mask(h), | |
2585 | (address & huge_page_mask(h)) + huge_page_size(h)); | |
1e8f889b DG |
2586 | } |
2587 | page_cache_release(new_page); | |
2588 | page_cache_release(old_page); | |
83c54070 | 2589 | return 0; |
1e8f889b DG |
2590 | } |
2591 | ||
04f2cbe3 | 2592 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
2593 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
2594 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
2595 | { |
2596 | struct address_space *mapping; | |
e7c4b0bf | 2597 | pgoff_t idx; |
04f2cbe3 MG |
2598 | |
2599 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 2600 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
2601 | |
2602 | return find_lock_page(mapping, idx); | |
2603 | } | |
2604 | ||
3ae77f43 HD |
2605 | /* |
2606 | * Return whether there is a pagecache page to back given address within VMA. | |
2607 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
2608 | */ | |
2609 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
2610 | struct vm_area_struct *vma, unsigned long address) |
2611 | { | |
2612 | struct address_space *mapping; | |
2613 | pgoff_t idx; | |
2614 | struct page *page; | |
2615 | ||
2616 | mapping = vma->vm_file->f_mapping; | |
2617 | idx = vma_hugecache_offset(h, vma, address); | |
2618 | ||
2619 | page = find_get_page(mapping, idx); | |
2620 | if (page) | |
2621 | put_page(page); | |
2622 | return page != NULL; | |
2623 | } | |
2624 | ||
a1ed3dda | 2625 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 2626 | unsigned long address, pte_t *ptep, unsigned int flags) |
ac9b9c66 | 2627 | { |
a5516438 | 2628 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 2629 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 2630 | int anon_rmap = 0; |
e7c4b0bf | 2631 | pgoff_t idx; |
4c887265 | 2632 | unsigned long size; |
4c887265 AL |
2633 | struct page *page; |
2634 | struct address_space *mapping; | |
1e8f889b | 2635 | pte_t new_pte; |
4c887265 | 2636 | |
04f2cbe3 MG |
2637 | /* |
2638 | * Currently, we are forced to kill the process in the event the | |
2639 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 2640 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
2641 | */ |
2642 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
2643 | printk(KERN_WARNING | |
2644 | "PID %d killed due to inadequate hugepage pool\n", | |
2645 | current->pid); | |
2646 | return ret; | |
2647 | } | |
2648 | ||
4c887265 | 2649 | mapping = vma->vm_file->f_mapping; |
a5516438 | 2650 | idx = vma_hugecache_offset(h, vma, address); |
4c887265 AL |
2651 | |
2652 | /* | |
2653 | * Use page lock to guard against racing truncation | |
2654 | * before we get page_table_lock. | |
2655 | */ | |
6bda666a CL |
2656 | retry: |
2657 | page = find_lock_page(mapping, idx); | |
2658 | if (!page) { | |
a5516438 | 2659 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
2660 | if (idx >= size) |
2661 | goto out; | |
04f2cbe3 | 2662 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 2663 | if (IS_ERR(page)) { |
76dcee75 AK |
2664 | ret = PTR_ERR(page); |
2665 | if (ret == -ENOMEM) | |
2666 | ret = VM_FAULT_OOM; | |
2667 | else | |
2668 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
2669 | goto out; |
2670 | } | |
47ad8475 | 2671 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 2672 | __SetPageUptodate(page); |
ac9b9c66 | 2673 | |
f83a275d | 2674 | if (vma->vm_flags & VM_MAYSHARE) { |
6bda666a | 2675 | int err; |
45c682a6 | 2676 | struct inode *inode = mapping->host; |
6bda666a CL |
2677 | |
2678 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
2679 | if (err) { | |
2680 | put_page(page); | |
6bda666a CL |
2681 | if (err == -EEXIST) |
2682 | goto retry; | |
2683 | goto out; | |
2684 | } | |
45c682a6 KC |
2685 | |
2686 | spin_lock(&inode->i_lock); | |
a5516438 | 2687 | inode->i_blocks += blocks_per_huge_page(h); |
45c682a6 | 2688 | spin_unlock(&inode->i_lock); |
23be7468 | 2689 | } else { |
6bda666a | 2690 | lock_page(page); |
0fe6e20b NH |
2691 | if (unlikely(anon_vma_prepare(vma))) { |
2692 | ret = VM_FAULT_OOM; | |
2693 | goto backout_unlocked; | |
2694 | } | |
409eb8c2 | 2695 | anon_rmap = 1; |
23be7468 | 2696 | } |
0fe6e20b | 2697 | } else { |
998b4382 NH |
2698 | /* |
2699 | * If memory error occurs between mmap() and fault, some process | |
2700 | * don't have hwpoisoned swap entry for errored virtual address. | |
2701 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
2702 | */ | |
2703 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 2704 | ret = VM_FAULT_HWPOISON | |
972dc4de | 2705 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
2706 | goto backout_unlocked; |
2707 | } | |
6bda666a | 2708 | } |
1e8f889b | 2709 | |
57303d80 AW |
2710 | /* |
2711 | * If we are going to COW a private mapping later, we examine the | |
2712 | * pending reservations for this page now. This will ensure that | |
2713 | * any allocations necessary to record that reservation occur outside | |
2714 | * the spinlock. | |
2715 | */ | |
788c7df4 | 2716 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) |
2b26736c AW |
2717 | if (vma_needs_reservation(h, vma, address) < 0) { |
2718 | ret = VM_FAULT_OOM; | |
2719 | goto backout_unlocked; | |
2720 | } | |
57303d80 | 2721 | |
ac9b9c66 | 2722 | spin_lock(&mm->page_table_lock); |
a5516438 | 2723 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
2724 | if (idx >= size) |
2725 | goto backout; | |
2726 | ||
83c54070 | 2727 | ret = 0; |
7f2e9525 | 2728 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
2729 | goto backout; |
2730 | ||
409eb8c2 HD |
2731 | if (anon_rmap) |
2732 | hugepage_add_new_anon_rmap(page, vma, address); | |
2733 | else | |
2734 | page_dup_rmap(page); | |
1e8f889b DG |
2735 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
2736 | && (vma->vm_flags & VM_SHARED))); | |
2737 | set_huge_pte_at(mm, address, ptep, new_pte); | |
2738 | ||
788c7df4 | 2739 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 2740 | /* Optimization, do the COW without a second fault */ |
04f2cbe3 | 2741 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); |
1e8f889b DG |
2742 | } |
2743 | ||
ac9b9c66 | 2744 | spin_unlock(&mm->page_table_lock); |
4c887265 AL |
2745 | unlock_page(page); |
2746 | out: | |
ac9b9c66 | 2747 | return ret; |
4c887265 AL |
2748 | |
2749 | backout: | |
2750 | spin_unlock(&mm->page_table_lock); | |
2b26736c | 2751 | backout_unlocked: |
4c887265 AL |
2752 | unlock_page(page); |
2753 | put_page(page); | |
2754 | goto out; | |
ac9b9c66 HD |
2755 | } |
2756 | ||
86e5216f | 2757 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 2758 | unsigned long address, unsigned int flags) |
86e5216f AL |
2759 | { |
2760 | pte_t *ptep; | |
2761 | pte_t entry; | |
1e8f889b | 2762 | int ret; |
0fe6e20b | 2763 | struct page *page = NULL; |
57303d80 | 2764 | struct page *pagecache_page = NULL; |
3935baa9 | 2765 | static DEFINE_MUTEX(hugetlb_instantiation_mutex); |
a5516438 | 2766 | struct hstate *h = hstate_vma(vma); |
86e5216f | 2767 | |
1e16a539 KH |
2768 | address &= huge_page_mask(h); |
2769 | ||
fd6a03ed NH |
2770 | ptep = huge_pte_offset(mm, address); |
2771 | if (ptep) { | |
2772 | entry = huge_ptep_get(ptep); | |
290408d4 NH |
2773 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
2774 | migration_entry_wait(mm, (pmd_t *)ptep, address); | |
2775 | return 0; | |
2776 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 2777 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 2778 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
2779 | } |
2780 | ||
a5516438 | 2781 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); |
86e5216f AL |
2782 | if (!ptep) |
2783 | return VM_FAULT_OOM; | |
2784 | ||
3935baa9 DG |
2785 | /* |
2786 | * Serialize hugepage allocation and instantiation, so that we don't | |
2787 | * get spurious allocation failures if two CPUs race to instantiate | |
2788 | * the same page in the page cache. | |
2789 | */ | |
2790 | mutex_lock(&hugetlb_instantiation_mutex); | |
7f2e9525 GS |
2791 | entry = huge_ptep_get(ptep); |
2792 | if (huge_pte_none(entry)) { | |
788c7df4 | 2793 | ret = hugetlb_no_page(mm, vma, address, ptep, flags); |
b4d1d99f | 2794 | goto out_mutex; |
3935baa9 | 2795 | } |
86e5216f | 2796 | |
83c54070 | 2797 | ret = 0; |
1e8f889b | 2798 | |
57303d80 AW |
2799 | /* |
2800 | * If we are going to COW the mapping later, we examine the pending | |
2801 | * reservations for this page now. This will ensure that any | |
2802 | * allocations necessary to record that reservation occur outside the | |
2803 | * spinlock. For private mappings, we also lookup the pagecache | |
2804 | * page now as it is used to determine if a reservation has been | |
2805 | * consumed. | |
2806 | */ | |
788c7df4 | 2807 | if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) { |
2b26736c AW |
2808 | if (vma_needs_reservation(h, vma, address) < 0) { |
2809 | ret = VM_FAULT_OOM; | |
b4d1d99f | 2810 | goto out_mutex; |
2b26736c | 2811 | } |
57303d80 | 2812 | |
f83a275d | 2813 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
2814 | pagecache_page = hugetlbfs_pagecache_page(h, |
2815 | vma, address); | |
2816 | } | |
2817 | ||
56c9cfb1 NH |
2818 | /* |
2819 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
2820 | * pagecache_page, so here we need take the former one | |
2821 | * when page != pagecache_page or !pagecache_page. | |
2822 | * Note that locking order is always pagecache_page -> page, | |
2823 | * so no worry about deadlock. | |
2824 | */ | |
2825 | page = pte_page(entry); | |
66aebce7 | 2826 | get_page(page); |
56c9cfb1 | 2827 | if (page != pagecache_page) |
0fe6e20b | 2828 | lock_page(page); |
0fe6e20b | 2829 | |
1e8f889b DG |
2830 | spin_lock(&mm->page_table_lock); |
2831 | /* Check for a racing update before calling hugetlb_cow */ | |
b4d1d99f DG |
2832 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) |
2833 | goto out_page_table_lock; | |
2834 | ||
2835 | ||
788c7df4 | 2836 | if (flags & FAULT_FLAG_WRITE) { |
b4d1d99f | 2837 | if (!pte_write(entry)) { |
57303d80 AW |
2838 | ret = hugetlb_cow(mm, vma, address, ptep, entry, |
2839 | pagecache_page); | |
b4d1d99f DG |
2840 | goto out_page_table_lock; |
2841 | } | |
2842 | entry = pte_mkdirty(entry); | |
2843 | } | |
2844 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
2845 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
2846 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 2847 | update_mmu_cache(vma, address, ptep); |
b4d1d99f DG |
2848 | |
2849 | out_page_table_lock: | |
1e8f889b | 2850 | spin_unlock(&mm->page_table_lock); |
57303d80 AW |
2851 | |
2852 | if (pagecache_page) { | |
2853 | unlock_page(pagecache_page); | |
2854 | put_page(pagecache_page); | |
2855 | } | |
1f64d69c DN |
2856 | if (page != pagecache_page) |
2857 | unlock_page(page); | |
66aebce7 | 2858 | put_page(page); |
57303d80 | 2859 | |
b4d1d99f | 2860 | out_mutex: |
3935baa9 | 2861 | mutex_unlock(&hugetlb_instantiation_mutex); |
1e8f889b DG |
2862 | |
2863 | return ret; | |
86e5216f AL |
2864 | } |
2865 | ||
ceb86879 AK |
2866 | /* Can be overriden by architectures */ |
2867 | __attribute__((weak)) struct page * | |
2868 | follow_huge_pud(struct mm_struct *mm, unsigned long address, | |
2869 | pud_t *pud, int write) | |
2870 | { | |
2871 | BUG(); | |
2872 | return NULL; | |
2873 | } | |
2874 | ||
63551ae0 DG |
2875 | int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
2876 | struct page **pages, struct vm_area_struct **vmas, | |
5b23dbe8 | 2877 | unsigned long *position, int *length, int i, |
2a15efc9 | 2878 | unsigned int flags) |
63551ae0 | 2879 | { |
d5d4b0aa KC |
2880 | unsigned long pfn_offset; |
2881 | unsigned long vaddr = *position; | |
63551ae0 | 2882 | int remainder = *length; |
a5516438 | 2883 | struct hstate *h = hstate_vma(vma); |
63551ae0 | 2884 | |
1c59827d | 2885 | spin_lock(&mm->page_table_lock); |
63551ae0 | 2886 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 2887 | pte_t *pte; |
2a15efc9 | 2888 | int absent; |
4c887265 | 2889 | struct page *page; |
63551ae0 | 2890 | |
4c887265 AL |
2891 | /* |
2892 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 2893 | * each hugepage. We have to make sure we get the |
4c887265 AL |
2894 | * first, for the page indexing below to work. |
2895 | */ | |
a5516438 | 2896 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); |
2a15efc9 HD |
2897 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
2898 | ||
2899 | /* | |
2900 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
2901 | * an error where there's an empty slot with no huge pagecache |
2902 | * to back it. This way, we avoid allocating a hugepage, and | |
2903 | * the sparse dumpfile avoids allocating disk blocks, but its | |
2904 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 2905 | */ |
3ae77f43 HD |
2906 | if (absent && (flags & FOLL_DUMP) && |
2907 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
2a15efc9 HD |
2908 | remainder = 0; |
2909 | break; | |
2910 | } | |
63551ae0 | 2911 | |
2a15efc9 HD |
2912 | if (absent || |
2913 | ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 2914 | int ret; |
63551ae0 | 2915 | |
4c887265 | 2916 | spin_unlock(&mm->page_table_lock); |
2a15efc9 HD |
2917 | ret = hugetlb_fault(mm, vma, vaddr, |
2918 | (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); | |
4c887265 | 2919 | spin_lock(&mm->page_table_lock); |
a89182c7 | 2920 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 2921 | continue; |
63551ae0 | 2922 | |
4c887265 | 2923 | remainder = 0; |
4c887265 AL |
2924 | break; |
2925 | } | |
2926 | ||
a5516438 | 2927 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 2928 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 2929 | same_page: |
d6692183 | 2930 | if (pages) { |
2a15efc9 | 2931 | pages[i] = mem_map_offset(page, pfn_offset); |
4b2e38ad | 2932 | get_page(pages[i]); |
d6692183 | 2933 | } |
63551ae0 DG |
2934 | |
2935 | if (vmas) | |
2936 | vmas[i] = vma; | |
2937 | ||
2938 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 2939 | ++pfn_offset; |
63551ae0 DG |
2940 | --remainder; |
2941 | ++i; | |
d5d4b0aa | 2942 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 2943 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa KC |
2944 | /* |
2945 | * We use pfn_offset to avoid touching the pageframes | |
2946 | * of this compound page. | |
2947 | */ | |
2948 | goto same_page; | |
2949 | } | |
63551ae0 | 2950 | } |
1c59827d | 2951 | spin_unlock(&mm->page_table_lock); |
63551ae0 DG |
2952 | *length = remainder; |
2953 | *position = vaddr; | |
2954 | ||
2a15efc9 | 2955 | return i ? i : -EFAULT; |
63551ae0 | 2956 | } |
8f860591 ZY |
2957 | |
2958 | void hugetlb_change_protection(struct vm_area_struct *vma, | |
2959 | unsigned long address, unsigned long end, pgprot_t newprot) | |
2960 | { | |
2961 | struct mm_struct *mm = vma->vm_mm; | |
2962 | unsigned long start = address; | |
2963 | pte_t *ptep; | |
2964 | pte_t pte; | |
a5516438 | 2965 | struct hstate *h = hstate_vma(vma); |
8f860591 ZY |
2966 | |
2967 | BUG_ON(address >= end); | |
2968 | flush_cache_range(vma, address, end); | |
2969 | ||
3d48ae45 | 2970 | mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex); |
8f860591 | 2971 | spin_lock(&mm->page_table_lock); |
a5516438 | 2972 | for (; address < end; address += huge_page_size(h)) { |
8f860591 ZY |
2973 | ptep = huge_pte_offset(mm, address); |
2974 | if (!ptep) | |
2975 | continue; | |
39dde65c KC |
2976 | if (huge_pmd_unshare(mm, &address, ptep)) |
2977 | continue; | |
7f2e9525 | 2978 | if (!huge_pte_none(huge_ptep_get(ptep))) { |
8f860591 ZY |
2979 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
2980 | pte = pte_mkhuge(pte_modify(pte, newprot)); | |
2981 | set_huge_pte_at(mm, address, ptep, pte); | |
8f860591 ZY |
2982 | } |
2983 | } | |
2984 | spin_unlock(&mm->page_table_lock); | |
3d48ae45 | 2985 | mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex); |
8f860591 ZY |
2986 | |
2987 | flush_tlb_range(vma, start, end); | |
2988 | } | |
2989 | ||
a1e78772 MG |
2990 | int hugetlb_reserve_pages(struct inode *inode, |
2991 | long from, long to, | |
5a6fe125 | 2992 | struct vm_area_struct *vma, |
ca16d140 | 2993 | vm_flags_t vm_flags) |
e4e574b7 | 2994 | { |
17c9d12e | 2995 | long ret, chg; |
a5516438 | 2996 | struct hstate *h = hstate_inode(inode); |
90481622 | 2997 | struct hugepage_subpool *spool = subpool_inode(inode); |
e4e574b7 | 2998 | |
17c9d12e MG |
2999 | /* |
3000 | * Only apply hugepage reservation if asked. At fault time, an | |
3001 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 3002 | * without using reserves |
17c9d12e | 3003 | */ |
ca16d140 | 3004 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
3005 | return 0; |
3006 | ||
a1e78772 MG |
3007 | /* |
3008 | * Shared mappings base their reservation on the number of pages that | |
3009 | * are already allocated on behalf of the file. Private mappings need | |
3010 | * to reserve the full area even if read-only as mprotect() may be | |
3011 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
3012 | */ | |
f83a275d | 3013 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
a1e78772 | 3014 | chg = region_chg(&inode->i_mapping->private_list, from, to); |
17c9d12e MG |
3015 | else { |
3016 | struct resv_map *resv_map = resv_map_alloc(); | |
3017 | if (!resv_map) | |
3018 | return -ENOMEM; | |
3019 | ||
a1e78772 | 3020 | chg = to - from; |
84afd99b | 3021 | |
17c9d12e MG |
3022 | set_vma_resv_map(vma, resv_map); |
3023 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
3024 | } | |
3025 | ||
c50ac050 DH |
3026 | if (chg < 0) { |
3027 | ret = chg; | |
3028 | goto out_err; | |
3029 | } | |
8a630112 | 3030 | |
90481622 | 3031 | /* There must be enough pages in the subpool for the mapping */ |
c50ac050 DH |
3032 | if (hugepage_subpool_get_pages(spool, chg)) { |
3033 | ret = -ENOSPC; | |
3034 | goto out_err; | |
3035 | } | |
5a6fe125 MG |
3036 | |
3037 | /* | |
17c9d12e | 3038 | * Check enough hugepages are available for the reservation. |
90481622 | 3039 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 3040 | */ |
a5516438 | 3041 | ret = hugetlb_acct_memory(h, chg); |
68842c9b | 3042 | if (ret < 0) { |
90481622 | 3043 | hugepage_subpool_put_pages(spool, chg); |
c50ac050 | 3044 | goto out_err; |
68842c9b | 3045 | } |
17c9d12e MG |
3046 | |
3047 | /* | |
3048 | * Account for the reservations made. Shared mappings record regions | |
3049 | * that have reservations as they are shared by multiple VMAs. | |
3050 | * When the last VMA disappears, the region map says how much | |
3051 | * the reservation was and the page cache tells how much of | |
3052 | * the reservation was consumed. Private mappings are per-VMA and | |
3053 | * only the consumed reservations are tracked. When the VMA | |
3054 | * disappears, the original reservation is the VMA size and the | |
3055 | * consumed reservations are stored in the map. Hence, nothing | |
3056 | * else has to be done for private mappings here | |
3057 | */ | |
f83a275d | 3058 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
a1e78772 | 3059 | region_add(&inode->i_mapping->private_list, from, to); |
a43a8c39 | 3060 | return 0; |
c50ac050 | 3061 | out_err: |
4523e145 DH |
3062 | if (vma) |
3063 | resv_map_put(vma); | |
c50ac050 | 3064 | return ret; |
a43a8c39 KC |
3065 | } |
3066 | ||
3067 | void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) | |
3068 | { | |
a5516438 | 3069 | struct hstate *h = hstate_inode(inode); |
a43a8c39 | 3070 | long chg = region_truncate(&inode->i_mapping->private_list, offset); |
90481622 | 3071 | struct hugepage_subpool *spool = subpool_inode(inode); |
45c682a6 KC |
3072 | |
3073 | spin_lock(&inode->i_lock); | |
e4c6f8be | 3074 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
3075 | spin_unlock(&inode->i_lock); |
3076 | ||
90481622 | 3077 | hugepage_subpool_put_pages(spool, (chg - freed)); |
a5516438 | 3078 | hugetlb_acct_memory(h, -(chg - freed)); |
a43a8c39 | 3079 | } |
93f70f90 | 3080 | |
d5bd9106 AK |
3081 | #ifdef CONFIG_MEMORY_FAILURE |
3082 | ||
6de2b1aa NH |
3083 | /* Should be called in hugetlb_lock */ |
3084 | static int is_hugepage_on_freelist(struct page *hpage) | |
3085 | { | |
3086 | struct page *page; | |
3087 | struct page *tmp; | |
3088 | struct hstate *h = page_hstate(hpage); | |
3089 | int nid = page_to_nid(hpage); | |
3090 | ||
3091 | list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru) | |
3092 | if (page == hpage) | |
3093 | return 1; | |
3094 | return 0; | |
3095 | } | |
3096 | ||
93f70f90 NH |
3097 | /* |
3098 | * This function is called from memory failure code. | |
3099 | * Assume the caller holds page lock of the head page. | |
3100 | */ | |
6de2b1aa | 3101 | int dequeue_hwpoisoned_huge_page(struct page *hpage) |
93f70f90 NH |
3102 | { |
3103 | struct hstate *h = page_hstate(hpage); | |
3104 | int nid = page_to_nid(hpage); | |
6de2b1aa | 3105 | int ret = -EBUSY; |
93f70f90 NH |
3106 | |
3107 | spin_lock(&hugetlb_lock); | |
6de2b1aa NH |
3108 | if (is_hugepage_on_freelist(hpage)) { |
3109 | list_del(&hpage->lru); | |
8c6c2ecb | 3110 | set_page_refcounted(hpage); |
6de2b1aa NH |
3111 | h->free_huge_pages--; |
3112 | h->free_huge_pages_node[nid]--; | |
3113 | ret = 0; | |
3114 | } | |
93f70f90 | 3115 | spin_unlock(&hugetlb_lock); |
6de2b1aa | 3116 | return ret; |
93f70f90 | 3117 | } |
6de2b1aa | 3118 | #endif |