]>
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 | |
c3f38a38 | 38 | int hugetlb_max_hstate __read_mostly; |
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 | 48 | |
3935baa9 DG |
49 | /* |
50 | * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages | |
51 | */ | |
c3f38a38 | 52 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 53 | |
90481622 DG |
54 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
55 | { | |
56 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
57 | ||
58 | spin_unlock(&spool->lock); | |
59 | ||
60 | /* If no pages are used, and no other handles to the subpool | |
61 | * remain, free the subpool the subpool remain */ | |
62 | if (free) | |
63 | kfree(spool); | |
64 | } | |
65 | ||
66 | struct hugepage_subpool *hugepage_new_subpool(long nr_blocks) | |
67 | { | |
68 | struct hugepage_subpool *spool; | |
69 | ||
70 | spool = kmalloc(sizeof(*spool), GFP_KERNEL); | |
71 | if (!spool) | |
72 | return NULL; | |
73 | ||
74 | spin_lock_init(&spool->lock); | |
75 | spool->count = 1; | |
76 | spool->max_hpages = nr_blocks; | |
77 | spool->used_hpages = 0; | |
78 | ||
79 | return spool; | |
80 | } | |
81 | ||
82 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
83 | { | |
84 | spin_lock(&spool->lock); | |
85 | BUG_ON(!spool->count); | |
86 | spool->count--; | |
87 | unlock_or_release_subpool(spool); | |
88 | } | |
89 | ||
90 | static int hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
91 | long delta) | |
92 | { | |
93 | int ret = 0; | |
94 | ||
95 | if (!spool) | |
96 | return 0; | |
97 | ||
98 | spin_lock(&spool->lock); | |
99 | if ((spool->used_hpages + delta) <= spool->max_hpages) { | |
100 | spool->used_hpages += delta; | |
101 | } else { | |
102 | ret = -ENOMEM; | |
103 | } | |
104 | spin_unlock(&spool->lock); | |
105 | ||
106 | return ret; | |
107 | } | |
108 | ||
109 | static void hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
110 | long delta) | |
111 | { | |
112 | if (!spool) | |
113 | return; | |
114 | ||
115 | spin_lock(&spool->lock); | |
116 | spool->used_hpages -= delta; | |
117 | /* If hugetlbfs_put_super couldn't free spool due to | |
118 | * an outstanding quota reference, free it now. */ | |
119 | unlock_or_release_subpool(spool); | |
120 | } | |
121 | ||
122 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
123 | { | |
124 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
125 | } | |
126 | ||
127 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
128 | { | |
129 | return subpool_inode(vma->vm_file->f_dentry->d_inode); | |
130 | } | |
131 | ||
96822904 AW |
132 | /* |
133 | * Region tracking -- allows tracking of reservations and instantiated pages | |
134 | * across the pages in a mapping. | |
84afd99b AW |
135 | * |
136 | * The region data structures are protected by a combination of the mmap_sem | |
137 | * and the hugetlb_instantion_mutex. To access or modify a region the caller | |
138 | * must either hold the mmap_sem for write, or the mmap_sem for read and | |
139 | * the hugetlb_instantiation mutex: | |
140 | * | |
32f84528 | 141 | * down_write(&mm->mmap_sem); |
84afd99b | 142 | * or |
32f84528 CF |
143 | * down_read(&mm->mmap_sem); |
144 | * mutex_lock(&hugetlb_instantiation_mutex); | |
96822904 AW |
145 | */ |
146 | struct file_region { | |
147 | struct list_head link; | |
148 | long from; | |
149 | long to; | |
150 | }; | |
151 | ||
152 | static long region_add(struct list_head *head, long f, long t) | |
153 | { | |
154 | struct file_region *rg, *nrg, *trg; | |
155 | ||
156 | /* Locate the region we are either in or before. */ | |
157 | list_for_each_entry(rg, head, link) | |
158 | if (f <= rg->to) | |
159 | break; | |
160 | ||
161 | /* Round our left edge to the current segment if it encloses us. */ | |
162 | if (f > rg->from) | |
163 | f = rg->from; | |
164 | ||
165 | /* Check for and consume any regions we now overlap with. */ | |
166 | nrg = rg; | |
167 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
168 | if (&rg->link == head) | |
169 | break; | |
170 | if (rg->from > t) | |
171 | break; | |
172 | ||
173 | /* If this area reaches higher then extend our area to | |
174 | * include it completely. If this is not the first area | |
175 | * which we intend to reuse, free it. */ | |
176 | if (rg->to > t) | |
177 | t = rg->to; | |
178 | if (rg != nrg) { | |
179 | list_del(&rg->link); | |
180 | kfree(rg); | |
181 | } | |
182 | } | |
183 | nrg->from = f; | |
184 | nrg->to = t; | |
185 | return 0; | |
186 | } | |
187 | ||
188 | static long region_chg(struct list_head *head, long f, long t) | |
189 | { | |
190 | struct file_region *rg, *nrg; | |
191 | long chg = 0; | |
192 | ||
193 | /* Locate the region we are before or in. */ | |
194 | list_for_each_entry(rg, head, link) | |
195 | if (f <= rg->to) | |
196 | break; | |
197 | ||
198 | /* If we are below the current region then a new region is required. | |
199 | * Subtle, allocate a new region at the position but make it zero | |
200 | * size such that we can guarantee to record the reservation. */ | |
201 | if (&rg->link == head || t < rg->from) { | |
202 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
203 | if (!nrg) | |
204 | return -ENOMEM; | |
205 | nrg->from = f; | |
206 | nrg->to = f; | |
207 | INIT_LIST_HEAD(&nrg->link); | |
208 | list_add(&nrg->link, rg->link.prev); | |
209 | ||
210 | return t - f; | |
211 | } | |
212 | ||
213 | /* Round our left edge to the current segment if it encloses us. */ | |
214 | if (f > rg->from) | |
215 | f = rg->from; | |
216 | chg = t - f; | |
217 | ||
218 | /* Check for and consume any regions we now overlap with. */ | |
219 | list_for_each_entry(rg, rg->link.prev, link) { | |
220 | if (&rg->link == head) | |
221 | break; | |
222 | if (rg->from > t) | |
223 | return chg; | |
224 | ||
25985edc | 225 | /* We overlap with this area, if it extends further than |
96822904 AW |
226 | * us then we must extend ourselves. Account for its |
227 | * existing reservation. */ | |
228 | if (rg->to > t) { | |
229 | chg += rg->to - t; | |
230 | t = rg->to; | |
231 | } | |
232 | chg -= rg->to - rg->from; | |
233 | } | |
234 | return chg; | |
235 | } | |
236 | ||
237 | static long region_truncate(struct list_head *head, long end) | |
238 | { | |
239 | struct file_region *rg, *trg; | |
240 | long chg = 0; | |
241 | ||
242 | /* Locate the region we are either in or before. */ | |
243 | list_for_each_entry(rg, head, link) | |
244 | if (end <= rg->to) | |
245 | break; | |
246 | if (&rg->link == head) | |
247 | return 0; | |
248 | ||
249 | /* If we are in the middle of a region then adjust it. */ | |
250 | if (end > rg->from) { | |
251 | chg = rg->to - end; | |
252 | rg->to = end; | |
253 | rg = list_entry(rg->link.next, typeof(*rg), link); | |
254 | } | |
255 | ||
256 | /* Drop any remaining regions. */ | |
257 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
258 | if (&rg->link == head) | |
259 | break; | |
260 | chg += rg->to - rg->from; | |
261 | list_del(&rg->link); | |
262 | kfree(rg); | |
263 | } | |
264 | return chg; | |
265 | } | |
266 | ||
84afd99b AW |
267 | static long region_count(struct list_head *head, long f, long t) |
268 | { | |
269 | struct file_region *rg; | |
270 | long chg = 0; | |
271 | ||
272 | /* Locate each segment we overlap with, and count that overlap. */ | |
273 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
274 | long seg_from; |
275 | long seg_to; | |
84afd99b AW |
276 | |
277 | if (rg->to <= f) | |
278 | continue; | |
279 | if (rg->from >= t) | |
280 | break; | |
281 | ||
282 | seg_from = max(rg->from, f); | |
283 | seg_to = min(rg->to, t); | |
284 | ||
285 | chg += seg_to - seg_from; | |
286 | } | |
287 | ||
288 | return chg; | |
289 | } | |
290 | ||
e7c4b0bf AW |
291 | /* |
292 | * Convert the address within this vma to the page offset within | |
293 | * the mapping, in pagecache page units; huge pages here. | |
294 | */ | |
a5516438 AK |
295 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
296 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 297 | { |
a5516438 AK |
298 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
299 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
300 | } |
301 | ||
0fe6e20b NH |
302 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
303 | unsigned long address) | |
304 | { | |
305 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
306 | } | |
307 | ||
08fba699 MG |
308 | /* |
309 | * Return the size of the pages allocated when backing a VMA. In the majority | |
310 | * cases this will be same size as used by the page table entries. | |
311 | */ | |
312 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
313 | { | |
314 | struct hstate *hstate; | |
315 | ||
316 | if (!is_vm_hugetlb_page(vma)) | |
317 | return PAGE_SIZE; | |
318 | ||
319 | hstate = hstate_vma(vma); | |
320 | ||
321 | return 1UL << (hstate->order + PAGE_SHIFT); | |
322 | } | |
f340ca0f | 323 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 324 | |
3340289d MG |
325 | /* |
326 | * Return the page size being used by the MMU to back a VMA. In the majority | |
327 | * of cases, the page size used by the kernel matches the MMU size. On | |
328 | * architectures where it differs, an architecture-specific version of this | |
329 | * function is required. | |
330 | */ | |
331 | #ifndef vma_mmu_pagesize | |
332 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) | |
333 | { | |
334 | return vma_kernel_pagesize(vma); | |
335 | } | |
336 | #endif | |
337 | ||
84afd99b AW |
338 | /* |
339 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
340 | * bits of the reservation map pointer, which are always clear due to | |
341 | * alignment. | |
342 | */ | |
343 | #define HPAGE_RESV_OWNER (1UL << 0) | |
344 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 345 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 346 | |
a1e78772 MG |
347 | /* |
348 | * These helpers are used to track how many pages are reserved for | |
349 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
350 | * is guaranteed to have their future faults succeed. | |
351 | * | |
352 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
353 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
354 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
355 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
356 | * |
357 | * The private mapping reservation is represented in a subtly different | |
358 | * manner to a shared mapping. A shared mapping has a region map associated | |
359 | * with the underlying file, this region map represents the backing file | |
360 | * pages which have ever had a reservation assigned which this persists even | |
361 | * after the page is instantiated. A private mapping has a region map | |
362 | * associated with the original mmap which is attached to all VMAs which | |
363 | * reference it, this region map represents those offsets which have consumed | |
364 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 365 | */ |
e7c4b0bf AW |
366 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
367 | { | |
368 | return (unsigned long)vma->vm_private_data; | |
369 | } | |
370 | ||
371 | static void set_vma_private_data(struct vm_area_struct *vma, | |
372 | unsigned long value) | |
373 | { | |
374 | vma->vm_private_data = (void *)value; | |
375 | } | |
376 | ||
84afd99b AW |
377 | struct resv_map { |
378 | struct kref refs; | |
379 | struct list_head regions; | |
380 | }; | |
381 | ||
2a4b3ded | 382 | static struct resv_map *resv_map_alloc(void) |
84afd99b AW |
383 | { |
384 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
385 | if (!resv_map) | |
386 | return NULL; | |
387 | ||
388 | kref_init(&resv_map->refs); | |
389 | INIT_LIST_HEAD(&resv_map->regions); | |
390 | ||
391 | return resv_map; | |
392 | } | |
393 | ||
2a4b3ded | 394 | static void resv_map_release(struct kref *ref) |
84afd99b AW |
395 | { |
396 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
397 | ||
398 | /* Clear out any active regions before we release the map. */ | |
399 | region_truncate(&resv_map->regions, 0); | |
400 | kfree(resv_map); | |
401 | } | |
402 | ||
403 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) | |
a1e78772 MG |
404 | { |
405 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 406 | if (!(vma->vm_flags & VM_MAYSHARE)) |
84afd99b AW |
407 | return (struct resv_map *)(get_vma_private_data(vma) & |
408 | ~HPAGE_RESV_MASK); | |
2a4b3ded | 409 | return NULL; |
a1e78772 MG |
410 | } |
411 | ||
84afd99b | 412 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 MG |
413 | { |
414 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 415 | VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); |
a1e78772 | 416 | |
84afd99b AW |
417 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
418 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
419 | } |
420 | ||
421 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
422 | { | |
04f2cbe3 | 423 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); |
f83a275d | 424 | VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); |
e7c4b0bf AW |
425 | |
426 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
427 | } |
428 | ||
429 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
430 | { | |
431 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
e7c4b0bf AW |
432 | |
433 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
434 | } |
435 | ||
436 | /* Decrement the reserved pages in the hugepage pool by one */ | |
a5516438 AK |
437 | static void decrement_hugepage_resv_vma(struct hstate *h, |
438 | struct vm_area_struct *vma) | |
a1e78772 | 439 | { |
c37f9fb1 AW |
440 | if (vma->vm_flags & VM_NORESERVE) |
441 | return; | |
442 | ||
f83a275d | 443 | if (vma->vm_flags & VM_MAYSHARE) { |
a1e78772 | 444 | /* Shared mappings always use reserves */ |
a5516438 | 445 | h->resv_huge_pages--; |
84afd99b | 446 | } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
a1e78772 MG |
447 | /* |
448 | * Only the process that called mmap() has reserves for | |
449 | * private mappings. | |
450 | */ | |
a5516438 | 451 | h->resv_huge_pages--; |
a1e78772 MG |
452 | } |
453 | } | |
454 | ||
04f2cbe3 | 455 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
456 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
457 | { | |
458 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 459 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
460 | vma->vm_private_data = (void *)0; |
461 | } | |
462 | ||
463 | /* Returns true if the VMA has associated reserve pages */ | |
7f09ca51 | 464 | static int vma_has_reserves(struct vm_area_struct *vma) |
a1e78772 | 465 | { |
f83a275d | 466 | if (vma->vm_flags & VM_MAYSHARE) |
7f09ca51 MG |
467 | return 1; |
468 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) | |
469 | return 1; | |
470 | return 0; | |
a1e78772 MG |
471 | } |
472 | ||
0ebabb41 NH |
473 | static void copy_gigantic_page(struct page *dst, struct page *src) |
474 | { | |
475 | int i; | |
476 | struct hstate *h = page_hstate(src); | |
477 | struct page *dst_base = dst; | |
478 | struct page *src_base = src; | |
479 | ||
480 | for (i = 0; i < pages_per_huge_page(h); ) { | |
481 | cond_resched(); | |
482 | copy_highpage(dst, src); | |
483 | ||
484 | i++; | |
485 | dst = mem_map_next(dst, dst_base, i); | |
486 | src = mem_map_next(src, src_base, i); | |
487 | } | |
488 | } | |
489 | ||
490 | void copy_huge_page(struct page *dst, struct page *src) | |
491 | { | |
492 | int i; | |
493 | struct hstate *h = page_hstate(src); | |
494 | ||
495 | if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) { | |
496 | copy_gigantic_page(dst, src); | |
497 | return; | |
498 | } | |
499 | ||
500 | might_sleep(); | |
501 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
502 | cond_resched(); | |
503 | copy_highpage(dst + i, src + i); | |
504 | } | |
505 | } | |
506 | ||
a5516438 | 507 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
508 | { |
509 | int nid = page_to_nid(page); | |
0edaecfa | 510 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
511 | h->free_huge_pages++; |
512 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
513 | } |
514 | ||
bf50bab2 NH |
515 | static struct page *dequeue_huge_page_node(struct hstate *h, int nid) |
516 | { | |
517 | struct page *page; | |
518 | ||
519 | if (list_empty(&h->hugepage_freelists[nid])) | |
520 | return NULL; | |
521 | page = list_entry(h->hugepage_freelists[nid].next, struct page, lru); | |
0edaecfa | 522 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 523 | set_page_refcounted(page); |
bf50bab2 NH |
524 | h->free_huge_pages--; |
525 | h->free_huge_pages_node[nid]--; | |
526 | return page; | |
527 | } | |
528 | ||
a5516438 AK |
529 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
530 | struct vm_area_struct *vma, | |
04f2cbe3 | 531 | unsigned long address, int avoid_reserve) |
1da177e4 | 532 | { |
b1c12cbc | 533 | struct page *page = NULL; |
480eccf9 | 534 | struct mempolicy *mpol; |
19770b32 | 535 | nodemask_t *nodemask; |
c0ff7453 | 536 | struct zonelist *zonelist; |
dd1a239f MG |
537 | struct zone *zone; |
538 | struct zoneref *z; | |
cc9a6c87 | 539 | unsigned int cpuset_mems_cookie; |
1da177e4 | 540 | |
cc9a6c87 MG |
541 | retry_cpuset: |
542 | cpuset_mems_cookie = get_mems_allowed(); | |
c0ff7453 MX |
543 | zonelist = huge_zonelist(vma, address, |
544 | htlb_alloc_mask, &mpol, &nodemask); | |
a1e78772 MG |
545 | /* |
546 | * A child process with MAP_PRIVATE mappings created by their parent | |
547 | * have no page reserves. This check ensures that reservations are | |
548 | * not "stolen". The child may still get SIGKILLed | |
549 | */ | |
7f09ca51 | 550 | if (!vma_has_reserves(vma) && |
a5516438 | 551 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 552 | goto err; |
a1e78772 | 553 | |
04f2cbe3 | 554 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 555 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 556 | goto err; |
04f2cbe3 | 557 | |
19770b32 MG |
558 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
559 | MAX_NR_ZONES - 1, nodemask) { | |
bf50bab2 NH |
560 | if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) { |
561 | page = dequeue_huge_page_node(h, zone_to_nid(zone)); | |
562 | if (page) { | |
563 | if (!avoid_reserve) | |
564 | decrement_hugepage_resv_vma(h, vma); | |
565 | break; | |
566 | } | |
3abf7afd | 567 | } |
1da177e4 | 568 | } |
cc9a6c87 | 569 | |
52cd3b07 | 570 | mpol_cond_put(mpol); |
cc9a6c87 MG |
571 | if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page)) |
572 | goto retry_cpuset; | |
1da177e4 | 573 | return page; |
cc9a6c87 MG |
574 | |
575 | err: | |
576 | mpol_cond_put(mpol); | |
577 | return NULL; | |
1da177e4 LT |
578 | } |
579 | ||
a5516438 | 580 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
581 | { |
582 | int i; | |
a5516438 | 583 | |
18229df5 AW |
584 | VM_BUG_ON(h->order >= MAX_ORDER); |
585 | ||
a5516438 AK |
586 | h->nr_huge_pages--; |
587 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
588 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
589 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
590 | 1 << PG_referenced | 1 << PG_dirty | | |
591 | 1 << PG_active | 1 << PG_reserved | | |
592 | 1 << PG_private | 1 << PG_writeback); | |
6af2acb6 AL |
593 | } |
594 | set_compound_page_dtor(page, NULL); | |
595 | set_page_refcounted(page); | |
7f2e9525 | 596 | arch_release_hugepage(page); |
a5516438 | 597 | __free_pages(page, huge_page_order(h)); |
6af2acb6 AL |
598 | } |
599 | ||
e5ff2159 AK |
600 | struct hstate *size_to_hstate(unsigned long size) |
601 | { | |
602 | struct hstate *h; | |
603 | ||
604 | for_each_hstate(h) { | |
605 | if (huge_page_size(h) == size) | |
606 | return h; | |
607 | } | |
608 | return NULL; | |
609 | } | |
610 | ||
27a85ef1 DG |
611 | static void free_huge_page(struct page *page) |
612 | { | |
a5516438 AK |
613 | /* |
614 | * Can't pass hstate in here because it is called from the | |
615 | * compound page destructor. | |
616 | */ | |
e5ff2159 | 617 | struct hstate *h = page_hstate(page); |
7893d1d5 | 618 | int nid = page_to_nid(page); |
90481622 DG |
619 | struct hugepage_subpool *spool = |
620 | (struct hugepage_subpool *)page_private(page); | |
27a85ef1 | 621 | |
e5df70ab | 622 | set_page_private(page, 0); |
23be7468 | 623 | page->mapping = NULL; |
7893d1d5 | 624 | BUG_ON(page_count(page)); |
0fe6e20b | 625 | BUG_ON(page_mapcount(page)); |
27a85ef1 DG |
626 | |
627 | spin_lock(&hugetlb_lock); | |
aa888a74 | 628 | if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { |
0edaecfa AK |
629 | /* remove the page from active list */ |
630 | list_del(&page->lru); | |
a5516438 AK |
631 | update_and_free_page(h, page); |
632 | h->surplus_huge_pages--; | |
633 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 634 | } else { |
a5516438 | 635 | enqueue_huge_page(h, page); |
7893d1d5 | 636 | } |
27a85ef1 | 637 | spin_unlock(&hugetlb_lock); |
90481622 | 638 | hugepage_subpool_put_pages(spool, 1); |
27a85ef1 DG |
639 | } |
640 | ||
a5516438 | 641 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 642 | { |
0edaecfa | 643 | INIT_LIST_HEAD(&page->lru); |
b7ba30c6 AK |
644 | set_compound_page_dtor(page, free_huge_page); |
645 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
646 | h->nr_huge_pages++; |
647 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 AK |
648 | spin_unlock(&hugetlb_lock); |
649 | put_page(page); /* free it into the hugepage allocator */ | |
650 | } | |
651 | ||
20a0307c WF |
652 | static void prep_compound_gigantic_page(struct page *page, unsigned long order) |
653 | { | |
654 | int i; | |
655 | int nr_pages = 1 << order; | |
656 | struct page *p = page + 1; | |
657 | ||
658 | /* we rely on prep_new_huge_page to set the destructor */ | |
659 | set_compound_order(page, order); | |
660 | __SetPageHead(page); | |
661 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { | |
662 | __SetPageTail(p); | |
58a84aa9 | 663 | set_page_count(p, 0); |
20a0307c WF |
664 | p->first_page = page; |
665 | } | |
666 | } | |
667 | ||
668 | int PageHuge(struct page *page) | |
669 | { | |
670 | compound_page_dtor *dtor; | |
671 | ||
672 | if (!PageCompound(page)) | |
673 | return 0; | |
674 | ||
675 | page = compound_head(page); | |
676 | dtor = get_compound_page_dtor(page); | |
677 | ||
678 | return dtor == free_huge_page; | |
679 | } | |
43131e14 NH |
680 | EXPORT_SYMBOL_GPL(PageHuge); |
681 | ||
a5516438 | 682 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
1da177e4 | 683 | { |
1da177e4 | 684 | struct page *page; |
f96efd58 | 685 | |
aa888a74 AK |
686 | if (h->order >= MAX_ORDER) |
687 | return NULL; | |
688 | ||
6484eb3e | 689 | page = alloc_pages_exact_node(nid, |
551883ae NA |
690 | htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| |
691 | __GFP_REPEAT|__GFP_NOWARN, | |
a5516438 | 692 | huge_page_order(h)); |
1da177e4 | 693 | if (page) { |
7f2e9525 | 694 | if (arch_prepare_hugepage(page)) { |
caff3a2c | 695 | __free_pages(page, huge_page_order(h)); |
7b8ee84d | 696 | return NULL; |
7f2e9525 | 697 | } |
a5516438 | 698 | prep_new_huge_page(h, page, nid); |
1da177e4 | 699 | } |
63b4613c NA |
700 | |
701 | return page; | |
702 | } | |
703 | ||
9a76db09 | 704 | /* |
6ae11b27 LS |
705 | * common helper functions for hstate_next_node_to_{alloc|free}. |
706 | * We may have allocated or freed a huge page based on a different | |
707 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
708 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
709 | * node for alloc or free. | |
9a76db09 | 710 | */ |
6ae11b27 | 711 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) |
9a76db09 | 712 | { |
6ae11b27 | 713 | nid = next_node(nid, *nodes_allowed); |
9a76db09 | 714 | if (nid == MAX_NUMNODES) |
6ae11b27 | 715 | nid = first_node(*nodes_allowed); |
9a76db09 LS |
716 | VM_BUG_ON(nid >= MAX_NUMNODES); |
717 | ||
718 | return nid; | |
719 | } | |
720 | ||
6ae11b27 LS |
721 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) |
722 | { | |
723 | if (!node_isset(nid, *nodes_allowed)) | |
724 | nid = next_node_allowed(nid, nodes_allowed); | |
725 | return nid; | |
726 | } | |
727 | ||
5ced66c9 | 728 | /* |
6ae11b27 LS |
729 | * returns the previously saved node ["this node"] from which to |
730 | * allocate a persistent huge page for the pool and advance the | |
731 | * next node from which to allocate, handling wrap at end of node | |
732 | * mask. | |
5ced66c9 | 733 | */ |
6ae11b27 LS |
734 | static int hstate_next_node_to_alloc(struct hstate *h, |
735 | nodemask_t *nodes_allowed) | |
5ced66c9 | 736 | { |
6ae11b27 LS |
737 | int nid; |
738 | ||
739 | VM_BUG_ON(!nodes_allowed); | |
740 | ||
741 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
742 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
9a76db09 | 743 | |
9a76db09 | 744 | return nid; |
5ced66c9 AK |
745 | } |
746 | ||
6ae11b27 | 747 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
63b4613c NA |
748 | { |
749 | struct page *page; | |
750 | int start_nid; | |
751 | int next_nid; | |
752 | int ret = 0; | |
753 | ||
6ae11b27 | 754 | start_nid = hstate_next_node_to_alloc(h, nodes_allowed); |
e8c5c824 | 755 | next_nid = start_nid; |
63b4613c NA |
756 | |
757 | do { | |
e8c5c824 | 758 | page = alloc_fresh_huge_page_node(h, next_nid); |
9a76db09 | 759 | if (page) { |
63b4613c | 760 | ret = 1; |
9a76db09 LS |
761 | break; |
762 | } | |
6ae11b27 | 763 | next_nid = hstate_next_node_to_alloc(h, nodes_allowed); |
9a76db09 | 764 | } while (next_nid != start_nid); |
63b4613c | 765 | |
3b116300 AL |
766 | if (ret) |
767 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
768 | else | |
769 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
770 | ||
63b4613c | 771 | return ret; |
1da177e4 LT |
772 | } |
773 | ||
e8c5c824 | 774 | /* |
6ae11b27 LS |
775 | * helper for free_pool_huge_page() - return the previously saved |
776 | * node ["this node"] from which to free a huge page. Advance the | |
777 | * next node id whether or not we find a free huge page to free so | |
778 | * that the next attempt to free addresses the next node. | |
e8c5c824 | 779 | */ |
6ae11b27 | 780 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) |
e8c5c824 | 781 | { |
6ae11b27 LS |
782 | int nid; |
783 | ||
784 | VM_BUG_ON(!nodes_allowed); | |
785 | ||
786 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
787 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
9a76db09 | 788 | |
9a76db09 | 789 | return nid; |
e8c5c824 LS |
790 | } |
791 | ||
792 | /* | |
793 | * Free huge page from pool from next node to free. | |
794 | * Attempt to keep persistent huge pages more or less | |
795 | * balanced over allowed nodes. | |
796 | * Called with hugetlb_lock locked. | |
797 | */ | |
6ae11b27 LS |
798 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
799 | bool acct_surplus) | |
e8c5c824 LS |
800 | { |
801 | int start_nid; | |
802 | int next_nid; | |
803 | int ret = 0; | |
804 | ||
6ae11b27 | 805 | start_nid = hstate_next_node_to_free(h, nodes_allowed); |
e8c5c824 LS |
806 | next_nid = start_nid; |
807 | ||
808 | do { | |
685f3457 LS |
809 | /* |
810 | * If we're returning unused surplus pages, only examine | |
811 | * nodes with surplus pages. | |
812 | */ | |
813 | if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) && | |
814 | !list_empty(&h->hugepage_freelists[next_nid])) { | |
e8c5c824 LS |
815 | struct page *page = |
816 | list_entry(h->hugepage_freelists[next_nid].next, | |
817 | struct page, lru); | |
818 | list_del(&page->lru); | |
819 | h->free_huge_pages--; | |
820 | h->free_huge_pages_node[next_nid]--; | |
685f3457 LS |
821 | if (acct_surplus) { |
822 | h->surplus_huge_pages--; | |
823 | h->surplus_huge_pages_node[next_nid]--; | |
824 | } | |
e8c5c824 LS |
825 | update_and_free_page(h, page); |
826 | ret = 1; | |
9a76db09 | 827 | break; |
e8c5c824 | 828 | } |
6ae11b27 | 829 | next_nid = hstate_next_node_to_free(h, nodes_allowed); |
9a76db09 | 830 | } while (next_nid != start_nid); |
e8c5c824 LS |
831 | |
832 | return ret; | |
833 | } | |
834 | ||
bf50bab2 | 835 | static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) |
7893d1d5 AL |
836 | { |
837 | struct page *page; | |
bf50bab2 | 838 | unsigned int r_nid; |
7893d1d5 | 839 | |
aa888a74 AK |
840 | if (h->order >= MAX_ORDER) |
841 | return NULL; | |
842 | ||
d1c3fb1f NA |
843 | /* |
844 | * Assume we will successfully allocate the surplus page to | |
845 | * prevent racing processes from causing the surplus to exceed | |
846 | * overcommit | |
847 | * | |
848 | * This however introduces a different race, where a process B | |
849 | * tries to grow the static hugepage pool while alloc_pages() is | |
850 | * called by process A. B will only examine the per-node | |
851 | * counters in determining if surplus huge pages can be | |
852 | * converted to normal huge pages in adjust_pool_surplus(). A | |
853 | * won't be able to increment the per-node counter, until the | |
854 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
855 | * no more huge pages can be converted from surplus to normal | |
856 | * state (and doesn't try to convert again). Thus, we have a | |
857 | * case where a surplus huge page exists, the pool is grown, and | |
858 | * the surplus huge page still exists after, even though it | |
859 | * should just have been converted to a normal huge page. This | |
860 | * does not leak memory, though, as the hugepage will be freed | |
861 | * once it is out of use. It also does not allow the counters to | |
862 | * go out of whack in adjust_pool_surplus() as we don't modify | |
863 | * the node values until we've gotten the hugepage and only the | |
864 | * per-node value is checked there. | |
865 | */ | |
866 | spin_lock(&hugetlb_lock); | |
a5516438 | 867 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
868 | spin_unlock(&hugetlb_lock); |
869 | return NULL; | |
870 | } else { | |
a5516438 AK |
871 | h->nr_huge_pages++; |
872 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
873 | } |
874 | spin_unlock(&hugetlb_lock); | |
875 | ||
bf50bab2 NH |
876 | if (nid == NUMA_NO_NODE) |
877 | page = alloc_pages(htlb_alloc_mask|__GFP_COMP| | |
878 | __GFP_REPEAT|__GFP_NOWARN, | |
879 | huge_page_order(h)); | |
880 | else | |
881 | page = alloc_pages_exact_node(nid, | |
882 | htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| | |
883 | __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); | |
d1c3fb1f | 884 | |
caff3a2c GS |
885 | if (page && arch_prepare_hugepage(page)) { |
886 | __free_pages(page, huge_page_order(h)); | |
ea5768c7 | 887 | page = NULL; |
caff3a2c GS |
888 | } |
889 | ||
d1c3fb1f | 890 | spin_lock(&hugetlb_lock); |
7893d1d5 | 891 | if (page) { |
0edaecfa | 892 | INIT_LIST_HEAD(&page->lru); |
bf50bab2 | 893 | r_nid = page_to_nid(page); |
7893d1d5 | 894 | set_compound_page_dtor(page, free_huge_page); |
d1c3fb1f NA |
895 | /* |
896 | * We incremented the global counters already | |
897 | */ | |
bf50bab2 NH |
898 | h->nr_huge_pages_node[r_nid]++; |
899 | h->surplus_huge_pages_node[r_nid]++; | |
3b116300 | 900 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 901 | } else { |
a5516438 AK |
902 | h->nr_huge_pages--; |
903 | h->surplus_huge_pages--; | |
3b116300 | 904 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 905 | } |
d1c3fb1f | 906 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
907 | |
908 | return page; | |
909 | } | |
910 | ||
bf50bab2 NH |
911 | /* |
912 | * This allocation function is useful in the context where vma is irrelevant. | |
913 | * E.g. soft-offlining uses this function because it only cares physical | |
914 | * address of error page. | |
915 | */ | |
916 | struct page *alloc_huge_page_node(struct hstate *h, int nid) | |
917 | { | |
918 | struct page *page; | |
919 | ||
920 | spin_lock(&hugetlb_lock); | |
921 | page = dequeue_huge_page_node(h, nid); | |
922 | spin_unlock(&hugetlb_lock); | |
923 | ||
924 | if (!page) | |
925 | page = alloc_buddy_huge_page(h, nid); | |
926 | ||
927 | return page; | |
928 | } | |
929 | ||
e4e574b7 | 930 | /* |
25985edc | 931 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
932 | * of size 'delta'. |
933 | */ | |
a5516438 | 934 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
935 | { |
936 | struct list_head surplus_list; | |
937 | struct page *page, *tmp; | |
938 | int ret, i; | |
939 | int needed, allocated; | |
28073b02 | 940 | bool alloc_ok = true; |
e4e574b7 | 941 | |
a5516438 | 942 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 943 | if (needed <= 0) { |
a5516438 | 944 | h->resv_huge_pages += delta; |
e4e574b7 | 945 | return 0; |
ac09b3a1 | 946 | } |
e4e574b7 AL |
947 | |
948 | allocated = 0; | |
949 | INIT_LIST_HEAD(&surplus_list); | |
950 | ||
951 | ret = -ENOMEM; | |
952 | retry: | |
953 | spin_unlock(&hugetlb_lock); | |
954 | for (i = 0; i < needed; i++) { | |
bf50bab2 | 955 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
28073b02 HD |
956 | if (!page) { |
957 | alloc_ok = false; | |
958 | break; | |
959 | } | |
e4e574b7 AL |
960 | list_add(&page->lru, &surplus_list); |
961 | } | |
28073b02 | 962 | allocated += i; |
e4e574b7 AL |
963 | |
964 | /* | |
965 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
966 | * because either resv_huge_pages or free_huge_pages may have changed. | |
967 | */ | |
968 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
969 | needed = (h->resv_huge_pages + delta) - |
970 | (h->free_huge_pages + allocated); | |
28073b02 HD |
971 | if (needed > 0) { |
972 | if (alloc_ok) | |
973 | goto retry; | |
974 | /* | |
975 | * We were not able to allocate enough pages to | |
976 | * satisfy the entire reservation so we free what | |
977 | * we've allocated so far. | |
978 | */ | |
979 | goto free; | |
980 | } | |
e4e574b7 AL |
981 | /* |
982 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 983 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 984 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
985 | * allocator. Commit the entire reservation here to prevent another |
986 | * process from stealing the pages as they are added to the pool but | |
987 | * before they are reserved. | |
e4e574b7 AL |
988 | */ |
989 | needed += allocated; | |
a5516438 | 990 | h->resv_huge_pages += delta; |
e4e574b7 | 991 | ret = 0; |
a9869b83 | 992 | |
19fc3f0a | 993 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 994 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
995 | if ((--needed) < 0) |
996 | break; | |
a9869b83 NH |
997 | /* |
998 | * This page is now managed by the hugetlb allocator and has | |
999 | * no users -- drop the buddy allocator's reference. | |
1000 | */ | |
1001 | put_page_testzero(page); | |
1002 | VM_BUG_ON(page_count(page)); | |
a5516438 | 1003 | enqueue_huge_page(h, page); |
19fc3f0a | 1004 | } |
28073b02 | 1005 | free: |
b0365c8d | 1006 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1007 | |
1008 | /* Free unnecessary surplus pages to the buddy allocator */ | |
1009 | if (!list_empty(&surplus_list)) { | |
19fc3f0a | 1010 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
a9869b83 | 1011 | put_page(page); |
af767cbd | 1012 | } |
e4e574b7 | 1013 | } |
a9869b83 | 1014 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1015 | |
1016 | return ret; | |
1017 | } | |
1018 | ||
1019 | /* | |
1020 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
1021 | * allocated to satisfy the reservation must be explicitly freed if they were | |
1022 | * never used. | |
685f3457 | 1023 | * Called with hugetlb_lock held. |
e4e574b7 | 1024 | */ |
a5516438 AK |
1025 | static void return_unused_surplus_pages(struct hstate *h, |
1026 | unsigned long unused_resv_pages) | |
e4e574b7 | 1027 | { |
e4e574b7 AL |
1028 | unsigned long nr_pages; |
1029 | ||
ac09b3a1 | 1030 | /* Uncommit the reservation */ |
a5516438 | 1031 | h->resv_huge_pages -= unused_resv_pages; |
ac09b3a1 | 1032 | |
aa888a74 AK |
1033 | /* Cannot return gigantic pages currently */ |
1034 | if (h->order >= MAX_ORDER) | |
1035 | return; | |
1036 | ||
a5516438 | 1037 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1038 | |
685f3457 LS |
1039 | /* |
1040 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1041 | * evenly across all nodes with memory. Iterate across these nodes |
1042 | * until we can no longer free unreserved surplus pages. This occurs | |
1043 | * when the nodes with surplus pages have no free pages. | |
1044 | * free_pool_huge_page() will balance the the freed pages across the | |
1045 | * on-line nodes with memory and will handle the hstate accounting. | |
685f3457 LS |
1046 | */ |
1047 | while (nr_pages--) { | |
9b5e5d0f | 1048 | if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1)) |
685f3457 | 1049 | break; |
e4e574b7 AL |
1050 | } |
1051 | } | |
1052 | ||
c37f9fb1 AW |
1053 | /* |
1054 | * Determine if the huge page at addr within the vma has an associated | |
1055 | * reservation. Where it does not we will need to logically increase | |
90481622 DG |
1056 | * reservation and actually increase subpool usage before an allocation |
1057 | * can occur. Where any new reservation would be required the | |
1058 | * reservation change is prepared, but not committed. Once the page | |
1059 | * has been allocated from the subpool and instantiated the change should | |
1060 | * be committed via vma_commit_reservation. No action is required on | |
1061 | * failure. | |
c37f9fb1 | 1062 | */ |
e2f17d94 | 1063 | static long vma_needs_reservation(struct hstate *h, |
a5516438 | 1064 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 AW |
1065 | { |
1066 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1067 | struct inode *inode = mapping->host; | |
1068 | ||
f83a275d | 1069 | if (vma->vm_flags & VM_MAYSHARE) { |
a5516438 | 1070 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
c37f9fb1 AW |
1071 | return region_chg(&inode->i_mapping->private_list, |
1072 | idx, idx + 1); | |
1073 | ||
84afd99b AW |
1074 | } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1075 | return 1; | |
c37f9fb1 | 1076 | |
84afd99b | 1077 | } else { |
e2f17d94 | 1078 | long err; |
a5516438 | 1079 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
84afd99b AW |
1080 | struct resv_map *reservations = vma_resv_map(vma); |
1081 | ||
1082 | err = region_chg(&reservations->regions, idx, idx + 1); | |
1083 | if (err < 0) | |
1084 | return err; | |
1085 | return 0; | |
1086 | } | |
c37f9fb1 | 1087 | } |
a5516438 AK |
1088 | static void vma_commit_reservation(struct hstate *h, |
1089 | struct vm_area_struct *vma, unsigned long addr) | |
c37f9fb1 AW |
1090 | { |
1091 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1092 | struct inode *inode = mapping->host; | |
1093 | ||
f83a275d | 1094 | if (vma->vm_flags & VM_MAYSHARE) { |
a5516438 | 1095 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
c37f9fb1 | 1096 | region_add(&inode->i_mapping->private_list, idx, idx + 1); |
84afd99b AW |
1097 | |
1098 | } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { | |
a5516438 | 1099 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
84afd99b AW |
1100 | struct resv_map *reservations = vma_resv_map(vma); |
1101 | ||
1102 | /* Mark this page used in the map. */ | |
1103 | region_add(&reservations->regions, idx, idx + 1); | |
c37f9fb1 AW |
1104 | } |
1105 | } | |
1106 | ||
a1e78772 | 1107 | static struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 1108 | unsigned long addr, int avoid_reserve) |
1da177e4 | 1109 | { |
90481622 | 1110 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 1111 | struct hstate *h = hstate_vma(vma); |
348ea204 | 1112 | struct page *page; |
e2f17d94 | 1113 | long chg; |
a1e78772 MG |
1114 | |
1115 | /* | |
90481622 DG |
1116 | * Processes that did not create the mapping will have no |
1117 | * reserves and will not have accounted against subpool | |
1118 | * limit. Check that the subpool limit can be made before | |
1119 | * satisfying the allocation MAP_NORESERVE mappings may also | |
1120 | * need pages and subpool limit allocated allocated if no reserve | |
1121 | * mapping overlaps. | |
a1e78772 | 1122 | */ |
a5516438 | 1123 | chg = vma_needs_reservation(h, vma, addr); |
c37f9fb1 | 1124 | if (chg < 0) |
76dcee75 | 1125 | return ERR_PTR(-ENOMEM); |
c37f9fb1 | 1126 | if (chg) |
90481622 | 1127 | if (hugepage_subpool_get_pages(spool, chg)) |
76dcee75 | 1128 | return ERR_PTR(-ENOSPC); |
1da177e4 LT |
1129 | |
1130 | spin_lock(&hugetlb_lock); | |
a5516438 | 1131 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve); |
1da177e4 | 1132 | spin_unlock(&hugetlb_lock); |
b45b5bd6 | 1133 | |
68842c9b | 1134 | if (!page) { |
bf50bab2 | 1135 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
68842c9b | 1136 | if (!page) { |
90481622 | 1137 | hugepage_subpool_put_pages(spool, chg); |
76dcee75 | 1138 | return ERR_PTR(-ENOSPC); |
68842c9b KC |
1139 | } |
1140 | } | |
348ea204 | 1141 | |
90481622 | 1142 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 1143 | |
a5516438 | 1144 | vma_commit_reservation(h, vma, addr); |
c37f9fb1 | 1145 | |
90d8b7e6 | 1146 | return page; |
b45b5bd6 DG |
1147 | } |
1148 | ||
91f47662 | 1149 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
aa888a74 AK |
1150 | { |
1151 | struct huge_bootmem_page *m; | |
9b5e5d0f | 1152 | int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); |
aa888a74 AK |
1153 | |
1154 | while (nr_nodes) { | |
1155 | void *addr; | |
1156 | ||
1157 | addr = __alloc_bootmem_node_nopanic( | |
6ae11b27 | 1158 | NODE_DATA(hstate_next_node_to_alloc(h, |
9b5e5d0f | 1159 | &node_states[N_HIGH_MEMORY])), |
aa888a74 AK |
1160 | huge_page_size(h), huge_page_size(h), 0); |
1161 | ||
1162 | if (addr) { | |
1163 | /* | |
1164 | * Use the beginning of the huge page to store the | |
1165 | * huge_bootmem_page struct (until gather_bootmem | |
1166 | * puts them into the mem_map). | |
1167 | */ | |
1168 | m = addr; | |
91f47662 | 1169 | goto found; |
aa888a74 | 1170 | } |
aa888a74 AK |
1171 | nr_nodes--; |
1172 | } | |
1173 | return 0; | |
1174 | ||
1175 | found: | |
1176 | BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); | |
1177 | /* Put them into a private list first because mem_map is not up yet */ | |
1178 | list_add(&m->list, &huge_boot_pages); | |
1179 | m->hstate = h; | |
1180 | return 1; | |
1181 | } | |
1182 | ||
18229df5 AW |
1183 | static void prep_compound_huge_page(struct page *page, int order) |
1184 | { | |
1185 | if (unlikely(order > (MAX_ORDER - 1))) | |
1186 | prep_compound_gigantic_page(page, order); | |
1187 | else | |
1188 | prep_compound_page(page, order); | |
1189 | } | |
1190 | ||
aa888a74 AK |
1191 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
1192 | static void __init gather_bootmem_prealloc(void) | |
1193 | { | |
1194 | struct huge_bootmem_page *m; | |
1195 | ||
1196 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 1197 | struct hstate *h = m->hstate; |
ee8f248d BB |
1198 | struct page *page; |
1199 | ||
1200 | #ifdef CONFIG_HIGHMEM | |
1201 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
1202 | free_bootmem_late((unsigned long)m, | |
1203 | sizeof(struct huge_bootmem_page)); | |
1204 | #else | |
1205 | page = virt_to_page(m); | |
1206 | #endif | |
aa888a74 AK |
1207 | __ClearPageReserved(page); |
1208 | WARN_ON(page_count(page) != 1); | |
18229df5 | 1209 | prep_compound_huge_page(page, h->order); |
aa888a74 | 1210 | prep_new_huge_page(h, page, page_to_nid(page)); |
b0320c7b RA |
1211 | /* |
1212 | * If we had gigantic hugepages allocated at boot time, we need | |
1213 | * to restore the 'stolen' pages to totalram_pages in order to | |
1214 | * fix confusing memory reports from free(1) and another | |
1215 | * side-effects, like CommitLimit going negative. | |
1216 | */ | |
1217 | if (h->order > (MAX_ORDER - 1)) | |
1218 | totalram_pages += 1 << h->order; | |
aa888a74 AK |
1219 | } |
1220 | } | |
1221 | ||
8faa8b07 | 1222 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
1223 | { |
1224 | unsigned long i; | |
a5516438 | 1225 | |
e5ff2159 | 1226 | for (i = 0; i < h->max_huge_pages; ++i) { |
aa888a74 AK |
1227 | if (h->order >= MAX_ORDER) { |
1228 | if (!alloc_bootmem_huge_page(h)) | |
1229 | break; | |
9b5e5d0f LS |
1230 | } else if (!alloc_fresh_huge_page(h, |
1231 | &node_states[N_HIGH_MEMORY])) | |
1da177e4 | 1232 | break; |
1da177e4 | 1233 | } |
8faa8b07 | 1234 | h->max_huge_pages = i; |
e5ff2159 AK |
1235 | } |
1236 | ||
1237 | static void __init hugetlb_init_hstates(void) | |
1238 | { | |
1239 | struct hstate *h; | |
1240 | ||
1241 | for_each_hstate(h) { | |
8faa8b07 AK |
1242 | /* oversize hugepages were init'ed in early boot */ |
1243 | if (h->order < MAX_ORDER) | |
1244 | hugetlb_hstate_alloc_pages(h); | |
e5ff2159 AK |
1245 | } |
1246 | } | |
1247 | ||
4abd32db AK |
1248 | static char * __init memfmt(char *buf, unsigned long n) |
1249 | { | |
1250 | if (n >= (1UL << 30)) | |
1251 | sprintf(buf, "%lu GB", n >> 30); | |
1252 | else if (n >= (1UL << 20)) | |
1253 | sprintf(buf, "%lu MB", n >> 20); | |
1254 | else | |
1255 | sprintf(buf, "%lu KB", n >> 10); | |
1256 | return buf; | |
1257 | } | |
1258 | ||
e5ff2159 AK |
1259 | static void __init report_hugepages(void) |
1260 | { | |
1261 | struct hstate *h; | |
1262 | ||
1263 | for_each_hstate(h) { | |
4abd32db AK |
1264 | char buf[32]; |
1265 | printk(KERN_INFO "HugeTLB registered %s page size, " | |
1266 | "pre-allocated %ld pages\n", | |
1267 | memfmt(buf, huge_page_size(h)), | |
1268 | h->free_huge_pages); | |
e5ff2159 AK |
1269 | } |
1270 | } | |
1271 | ||
1da177e4 | 1272 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
1273 | static void try_to_free_low(struct hstate *h, unsigned long count, |
1274 | nodemask_t *nodes_allowed) | |
1da177e4 | 1275 | { |
4415cc8d CL |
1276 | int i; |
1277 | ||
aa888a74 AK |
1278 | if (h->order >= MAX_ORDER) |
1279 | return; | |
1280 | ||
6ae11b27 | 1281 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 1282 | struct page *page, *next; |
a5516438 AK |
1283 | struct list_head *freel = &h->hugepage_freelists[i]; |
1284 | list_for_each_entry_safe(page, next, freel, lru) { | |
1285 | if (count >= h->nr_huge_pages) | |
6b0c880d | 1286 | return; |
1da177e4 LT |
1287 | if (PageHighMem(page)) |
1288 | continue; | |
1289 | list_del(&page->lru); | |
e5ff2159 | 1290 | update_and_free_page(h, page); |
a5516438 AK |
1291 | h->free_huge_pages--; |
1292 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
1293 | } |
1294 | } | |
1295 | } | |
1296 | #else | |
6ae11b27 LS |
1297 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
1298 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
1299 | { |
1300 | } | |
1301 | #endif | |
1302 | ||
20a0307c WF |
1303 | /* |
1304 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
1305 | * balanced by operating on them in a round-robin fashion. | |
1306 | * Returns 1 if an adjustment was made. | |
1307 | */ | |
6ae11b27 LS |
1308 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
1309 | int delta) | |
20a0307c | 1310 | { |
e8c5c824 | 1311 | int start_nid, next_nid; |
20a0307c WF |
1312 | int ret = 0; |
1313 | ||
1314 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 1315 | |
e8c5c824 | 1316 | if (delta < 0) |
6ae11b27 | 1317 | start_nid = hstate_next_node_to_alloc(h, nodes_allowed); |
e8c5c824 | 1318 | else |
6ae11b27 | 1319 | start_nid = hstate_next_node_to_free(h, nodes_allowed); |
e8c5c824 LS |
1320 | next_nid = start_nid; |
1321 | ||
1322 | do { | |
1323 | int nid = next_nid; | |
1324 | if (delta < 0) { | |
e8c5c824 LS |
1325 | /* |
1326 | * To shrink on this node, there must be a surplus page | |
1327 | */ | |
9a76db09 | 1328 | if (!h->surplus_huge_pages_node[nid]) { |
6ae11b27 LS |
1329 | next_nid = hstate_next_node_to_alloc(h, |
1330 | nodes_allowed); | |
e8c5c824 | 1331 | continue; |
9a76db09 | 1332 | } |
e8c5c824 LS |
1333 | } |
1334 | if (delta > 0) { | |
e8c5c824 LS |
1335 | /* |
1336 | * Surplus cannot exceed the total number of pages | |
1337 | */ | |
1338 | if (h->surplus_huge_pages_node[nid] >= | |
9a76db09 | 1339 | h->nr_huge_pages_node[nid]) { |
6ae11b27 LS |
1340 | next_nid = hstate_next_node_to_free(h, |
1341 | nodes_allowed); | |
e8c5c824 | 1342 | continue; |
9a76db09 | 1343 | } |
e8c5c824 | 1344 | } |
20a0307c WF |
1345 | |
1346 | h->surplus_huge_pages += delta; | |
1347 | h->surplus_huge_pages_node[nid] += delta; | |
1348 | ret = 1; | |
1349 | break; | |
e8c5c824 | 1350 | } while (next_nid != start_nid); |
20a0307c | 1351 | |
20a0307c WF |
1352 | return ret; |
1353 | } | |
1354 | ||
a5516438 | 1355 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
1356 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
1357 | nodemask_t *nodes_allowed) | |
1da177e4 | 1358 | { |
7893d1d5 | 1359 | unsigned long min_count, ret; |
1da177e4 | 1360 | |
aa888a74 AK |
1361 | if (h->order >= MAX_ORDER) |
1362 | return h->max_huge_pages; | |
1363 | ||
7893d1d5 AL |
1364 | /* |
1365 | * Increase the pool size | |
1366 | * First take pages out of surplus state. Then make up the | |
1367 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f NA |
1368 | * |
1369 | * We might race with alloc_buddy_huge_page() here and be unable | |
1370 | * to convert a surplus huge page to a normal huge page. That is | |
1371 | * not critical, though, it just means the overall size of the | |
1372 | * pool might be one hugepage larger than it needs to be, but | |
1373 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 1374 | */ |
1da177e4 | 1375 | spin_lock(&hugetlb_lock); |
a5516438 | 1376 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 1377 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
1378 | break; |
1379 | } | |
1380 | ||
a5516438 | 1381 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
1382 | /* |
1383 | * If this allocation races such that we no longer need the | |
1384 | * page, free_huge_page will handle it by freeing the page | |
1385 | * and reducing the surplus. | |
1386 | */ | |
1387 | spin_unlock(&hugetlb_lock); | |
6ae11b27 | 1388 | ret = alloc_fresh_huge_page(h, nodes_allowed); |
7893d1d5 AL |
1389 | spin_lock(&hugetlb_lock); |
1390 | if (!ret) | |
1391 | goto out; | |
1392 | ||
536240f2 MG |
1393 | /* Bail for signals. Probably ctrl-c from user */ |
1394 | if (signal_pending(current)) | |
1395 | goto out; | |
7893d1d5 | 1396 | } |
7893d1d5 AL |
1397 | |
1398 | /* | |
1399 | * Decrease the pool size | |
1400 | * First return free pages to the buddy allocator (being careful | |
1401 | * to keep enough around to satisfy reservations). Then place | |
1402 | * pages into surplus state as needed so the pool will shrink | |
1403 | * to the desired size as pages become free. | |
d1c3fb1f NA |
1404 | * |
1405 | * By placing pages into the surplus state independent of the | |
1406 | * overcommit value, we are allowing the surplus pool size to | |
1407 | * exceed overcommit. There are few sane options here. Since | |
1408 | * alloc_buddy_huge_page() is checking the global counter, | |
1409 | * though, we'll note that we're not allowed to exceed surplus | |
1410 | * and won't grow the pool anywhere else. Not until one of the | |
1411 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 1412 | */ |
a5516438 | 1413 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 1414 | min_count = max(count, min_count); |
6ae11b27 | 1415 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 1416 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 1417 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 1418 | break; |
1da177e4 | 1419 | } |
a5516438 | 1420 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 1421 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
1422 | break; |
1423 | } | |
1424 | out: | |
a5516438 | 1425 | ret = persistent_huge_pages(h); |
1da177e4 | 1426 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 1427 | return ret; |
1da177e4 LT |
1428 | } |
1429 | ||
a3437870 NA |
1430 | #define HSTATE_ATTR_RO(_name) \ |
1431 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
1432 | ||
1433 | #define HSTATE_ATTR(_name) \ | |
1434 | static struct kobj_attribute _name##_attr = \ | |
1435 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
1436 | ||
1437 | static struct kobject *hugepages_kobj; | |
1438 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1439 | ||
9a305230 LS |
1440 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
1441 | ||
1442 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
1443 | { |
1444 | int i; | |
9a305230 | 1445 | |
a3437870 | 1446 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
1447 | if (hstate_kobjs[i] == kobj) { |
1448 | if (nidp) | |
1449 | *nidp = NUMA_NO_NODE; | |
a3437870 | 1450 | return &hstates[i]; |
9a305230 LS |
1451 | } |
1452 | ||
1453 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
1454 | } |
1455 | ||
06808b08 | 1456 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
1457 | struct kobj_attribute *attr, char *buf) |
1458 | { | |
9a305230 LS |
1459 | struct hstate *h; |
1460 | unsigned long nr_huge_pages; | |
1461 | int nid; | |
1462 | ||
1463 | h = kobj_to_hstate(kobj, &nid); | |
1464 | if (nid == NUMA_NO_NODE) | |
1465 | nr_huge_pages = h->nr_huge_pages; | |
1466 | else | |
1467 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
1468 | ||
1469 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 1470 | } |
adbe8726 | 1471 | |
06808b08 LS |
1472 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
1473 | struct kobject *kobj, struct kobj_attribute *attr, | |
1474 | const char *buf, size_t len) | |
a3437870 NA |
1475 | { |
1476 | int err; | |
9a305230 | 1477 | int nid; |
06808b08 | 1478 | unsigned long count; |
9a305230 | 1479 | struct hstate *h; |
bad44b5b | 1480 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 1481 | |
06808b08 | 1482 | err = strict_strtoul(buf, 10, &count); |
73ae31e5 | 1483 | if (err) |
adbe8726 | 1484 | goto out; |
a3437870 | 1485 | |
9a305230 | 1486 | h = kobj_to_hstate(kobj, &nid); |
adbe8726 EM |
1487 | if (h->order >= MAX_ORDER) { |
1488 | err = -EINVAL; | |
1489 | goto out; | |
1490 | } | |
1491 | ||
9a305230 LS |
1492 | if (nid == NUMA_NO_NODE) { |
1493 | /* | |
1494 | * global hstate attribute | |
1495 | */ | |
1496 | if (!(obey_mempolicy && | |
1497 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
1498 | NODEMASK_FREE(nodes_allowed); | |
1499 | nodes_allowed = &node_states[N_HIGH_MEMORY]; | |
1500 | } | |
1501 | } else if (nodes_allowed) { | |
1502 | /* | |
1503 | * per node hstate attribute: adjust count to global, | |
1504 | * but restrict alloc/free to the specified node. | |
1505 | */ | |
1506 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
1507 | init_nodemask_of_node(nodes_allowed, nid); | |
1508 | } else | |
1509 | nodes_allowed = &node_states[N_HIGH_MEMORY]; | |
1510 | ||
06808b08 | 1511 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 1512 | |
9b5e5d0f | 1513 | if (nodes_allowed != &node_states[N_HIGH_MEMORY]) |
06808b08 LS |
1514 | NODEMASK_FREE(nodes_allowed); |
1515 | ||
1516 | return len; | |
adbe8726 EM |
1517 | out: |
1518 | NODEMASK_FREE(nodes_allowed); | |
1519 | return err; | |
06808b08 LS |
1520 | } |
1521 | ||
1522 | static ssize_t nr_hugepages_show(struct kobject *kobj, | |
1523 | struct kobj_attribute *attr, char *buf) | |
1524 | { | |
1525 | return nr_hugepages_show_common(kobj, attr, buf); | |
1526 | } | |
1527 | ||
1528 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
1529 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1530 | { | |
1531 | return nr_hugepages_store_common(false, kobj, attr, buf, len); | |
a3437870 NA |
1532 | } |
1533 | HSTATE_ATTR(nr_hugepages); | |
1534 | ||
06808b08 LS |
1535 | #ifdef CONFIG_NUMA |
1536 | ||
1537 | /* | |
1538 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
1539 | * huge page alloc/free. | |
1540 | */ | |
1541 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
1542 | struct kobj_attribute *attr, char *buf) | |
1543 | { | |
1544 | return nr_hugepages_show_common(kobj, attr, buf); | |
1545 | } | |
1546 | ||
1547 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
1548 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1549 | { | |
1550 | return nr_hugepages_store_common(true, kobj, attr, buf, len); | |
1551 | } | |
1552 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
1553 | #endif | |
1554 | ||
1555 | ||
a3437870 NA |
1556 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
1557 | struct kobj_attribute *attr, char *buf) | |
1558 | { | |
9a305230 | 1559 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1560 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
1561 | } | |
adbe8726 | 1562 | |
a3437870 NA |
1563 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
1564 | struct kobj_attribute *attr, const char *buf, size_t count) | |
1565 | { | |
1566 | int err; | |
1567 | unsigned long input; | |
9a305230 | 1568 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 1569 | |
adbe8726 EM |
1570 | if (h->order >= MAX_ORDER) |
1571 | return -EINVAL; | |
1572 | ||
a3437870 NA |
1573 | err = strict_strtoul(buf, 10, &input); |
1574 | if (err) | |
73ae31e5 | 1575 | return err; |
a3437870 NA |
1576 | |
1577 | spin_lock(&hugetlb_lock); | |
1578 | h->nr_overcommit_huge_pages = input; | |
1579 | spin_unlock(&hugetlb_lock); | |
1580 | ||
1581 | return count; | |
1582 | } | |
1583 | HSTATE_ATTR(nr_overcommit_hugepages); | |
1584 | ||
1585 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
1586 | struct kobj_attribute *attr, char *buf) | |
1587 | { | |
9a305230 LS |
1588 | struct hstate *h; |
1589 | unsigned long free_huge_pages; | |
1590 | int nid; | |
1591 | ||
1592 | h = kobj_to_hstate(kobj, &nid); | |
1593 | if (nid == NUMA_NO_NODE) | |
1594 | free_huge_pages = h->free_huge_pages; | |
1595 | else | |
1596 | free_huge_pages = h->free_huge_pages_node[nid]; | |
1597 | ||
1598 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
1599 | } |
1600 | HSTATE_ATTR_RO(free_hugepages); | |
1601 | ||
1602 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
1603 | struct kobj_attribute *attr, char *buf) | |
1604 | { | |
9a305230 | 1605 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1606 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
1607 | } | |
1608 | HSTATE_ATTR_RO(resv_hugepages); | |
1609 | ||
1610 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
1611 | struct kobj_attribute *attr, char *buf) | |
1612 | { | |
9a305230 LS |
1613 | struct hstate *h; |
1614 | unsigned long surplus_huge_pages; | |
1615 | int nid; | |
1616 | ||
1617 | h = kobj_to_hstate(kobj, &nid); | |
1618 | if (nid == NUMA_NO_NODE) | |
1619 | surplus_huge_pages = h->surplus_huge_pages; | |
1620 | else | |
1621 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
1622 | ||
1623 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
1624 | } |
1625 | HSTATE_ATTR_RO(surplus_hugepages); | |
1626 | ||
1627 | static struct attribute *hstate_attrs[] = { | |
1628 | &nr_hugepages_attr.attr, | |
1629 | &nr_overcommit_hugepages_attr.attr, | |
1630 | &free_hugepages_attr.attr, | |
1631 | &resv_hugepages_attr.attr, | |
1632 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
1633 | #ifdef CONFIG_NUMA |
1634 | &nr_hugepages_mempolicy_attr.attr, | |
1635 | #endif | |
a3437870 NA |
1636 | NULL, |
1637 | }; | |
1638 | ||
1639 | static struct attribute_group hstate_attr_group = { | |
1640 | .attrs = hstate_attrs, | |
1641 | }; | |
1642 | ||
094e9539 JM |
1643 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
1644 | struct kobject **hstate_kobjs, | |
1645 | struct attribute_group *hstate_attr_group) | |
a3437870 NA |
1646 | { |
1647 | int retval; | |
972dc4de | 1648 | int hi = hstate_index(h); |
a3437870 | 1649 | |
9a305230 LS |
1650 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
1651 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
1652 | return -ENOMEM; |
1653 | ||
9a305230 | 1654 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 1655 | if (retval) |
9a305230 | 1656 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
1657 | |
1658 | return retval; | |
1659 | } | |
1660 | ||
1661 | static void __init hugetlb_sysfs_init(void) | |
1662 | { | |
1663 | struct hstate *h; | |
1664 | int err; | |
1665 | ||
1666 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
1667 | if (!hugepages_kobj) | |
1668 | return; | |
1669 | ||
1670 | for_each_hstate(h) { | |
9a305230 LS |
1671 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
1672 | hstate_kobjs, &hstate_attr_group); | |
a3437870 NA |
1673 | if (err) |
1674 | printk(KERN_ERR "Hugetlb: Unable to add hstate %s", | |
1675 | h->name); | |
1676 | } | |
1677 | } | |
1678 | ||
9a305230 LS |
1679 | #ifdef CONFIG_NUMA |
1680 | ||
1681 | /* | |
1682 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
1683 | * with node devices in node_devices[] using a parallel array. The array |
1684 | * index of a node device or _hstate == node id. | |
1685 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
1686 | * the base kernel, on the hugetlb module. |
1687 | */ | |
1688 | struct node_hstate { | |
1689 | struct kobject *hugepages_kobj; | |
1690 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1691 | }; | |
1692 | struct node_hstate node_hstates[MAX_NUMNODES]; | |
1693 | ||
1694 | /* | |
10fbcf4c | 1695 | * A subset of global hstate attributes for node devices |
9a305230 LS |
1696 | */ |
1697 | static struct attribute *per_node_hstate_attrs[] = { | |
1698 | &nr_hugepages_attr.attr, | |
1699 | &free_hugepages_attr.attr, | |
1700 | &surplus_hugepages_attr.attr, | |
1701 | NULL, | |
1702 | }; | |
1703 | ||
1704 | static struct attribute_group per_node_hstate_attr_group = { | |
1705 | .attrs = per_node_hstate_attrs, | |
1706 | }; | |
1707 | ||
1708 | /* | |
10fbcf4c | 1709 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
1710 | * Returns node id via non-NULL nidp. |
1711 | */ | |
1712 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
1713 | { | |
1714 | int nid; | |
1715 | ||
1716 | for (nid = 0; nid < nr_node_ids; nid++) { | |
1717 | struct node_hstate *nhs = &node_hstates[nid]; | |
1718 | int i; | |
1719 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
1720 | if (nhs->hstate_kobjs[i] == kobj) { | |
1721 | if (nidp) | |
1722 | *nidp = nid; | |
1723 | return &hstates[i]; | |
1724 | } | |
1725 | } | |
1726 | ||
1727 | BUG(); | |
1728 | return NULL; | |
1729 | } | |
1730 | ||
1731 | /* | |
10fbcf4c | 1732 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
1733 | * No-op if no hstate attributes attached. |
1734 | */ | |
1735 | void hugetlb_unregister_node(struct node *node) | |
1736 | { | |
1737 | struct hstate *h; | |
10fbcf4c | 1738 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
1739 | |
1740 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 1741 | return; /* no hstate attributes */ |
9a305230 | 1742 | |
972dc4de AK |
1743 | for_each_hstate(h) { |
1744 | int idx = hstate_index(h); | |
1745 | if (nhs->hstate_kobjs[idx]) { | |
1746 | kobject_put(nhs->hstate_kobjs[idx]); | |
1747 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 1748 | } |
972dc4de | 1749 | } |
9a305230 LS |
1750 | |
1751 | kobject_put(nhs->hugepages_kobj); | |
1752 | nhs->hugepages_kobj = NULL; | |
1753 | } | |
1754 | ||
1755 | /* | |
10fbcf4c | 1756 | * hugetlb module exit: unregister hstate attributes from node devices |
9a305230 LS |
1757 | * that have them. |
1758 | */ | |
1759 | static void hugetlb_unregister_all_nodes(void) | |
1760 | { | |
1761 | int nid; | |
1762 | ||
1763 | /* | |
10fbcf4c | 1764 | * disable node device registrations. |
9a305230 LS |
1765 | */ |
1766 | register_hugetlbfs_with_node(NULL, NULL); | |
1767 | ||
1768 | /* | |
1769 | * remove hstate attributes from any nodes that have them. | |
1770 | */ | |
1771 | for (nid = 0; nid < nr_node_ids; nid++) | |
1772 | hugetlb_unregister_node(&node_devices[nid]); | |
1773 | } | |
1774 | ||
1775 | /* | |
10fbcf4c | 1776 | * Register hstate attributes for a single node device. |
9a305230 LS |
1777 | * No-op if attributes already registered. |
1778 | */ | |
1779 | void hugetlb_register_node(struct node *node) | |
1780 | { | |
1781 | struct hstate *h; | |
10fbcf4c | 1782 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
1783 | int err; |
1784 | ||
1785 | if (nhs->hugepages_kobj) | |
1786 | return; /* already allocated */ | |
1787 | ||
1788 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 1789 | &node->dev.kobj); |
9a305230 LS |
1790 | if (!nhs->hugepages_kobj) |
1791 | return; | |
1792 | ||
1793 | for_each_hstate(h) { | |
1794 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
1795 | nhs->hstate_kobjs, | |
1796 | &per_node_hstate_attr_group); | |
1797 | if (err) { | |
1798 | printk(KERN_ERR "Hugetlb: Unable to add hstate %s" | |
1799 | " for node %d\n", | |
10fbcf4c | 1800 | h->name, node->dev.id); |
9a305230 LS |
1801 | hugetlb_unregister_node(node); |
1802 | break; | |
1803 | } | |
1804 | } | |
1805 | } | |
1806 | ||
1807 | /* | |
9b5e5d0f | 1808 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
1809 | * devices of nodes that have memory. All on-line nodes should have |
1810 | * registered their associated device by this time. | |
9a305230 LS |
1811 | */ |
1812 | static void hugetlb_register_all_nodes(void) | |
1813 | { | |
1814 | int nid; | |
1815 | ||
9b5e5d0f | 1816 | for_each_node_state(nid, N_HIGH_MEMORY) { |
9a305230 | 1817 | struct node *node = &node_devices[nid]; |
10fbcf4c | 1818 | if (node->dev.id == nid) |
9a305230 LS |
1819 | hugetlb_register_node(node); |
1820 | } | |
1821 | ||
1822 | /* | |
10fbcf4c | 1823 | * Let the node device driver know we're here so it can |
9a305230 LS |
1824 | * [un]register hstate attributes on node hotplug. |
1825 | */ | |
1826 | register_hugetlbfs_with_node(hugetlb_register_node, | |
1827 | hugetlb_unregister_node); | |
1828 | } | |
1829 | #else /* !CONFIG_NUMA */ | |
1830 | ||
1831 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
1832 | { | |
1833 | BUG(); | |
1834 | if (nidp) | |
1835 | *nidp = -1; | |
1836 | return NULL; | |
1837 | } | |
1838 | ||
1839 | static void hugetlb_unregister_all_nodes(void) { } | |
1840 | ||
1841 | static void hugetlb_register_all_nodes(void) { } | |
1842 | ||
1843 | #endif | |
1844 | ||
a3437870 NA |
1845 | static void __exit hugetlb_exit(void) |
1846 | { | |
1847 | struct hstate *h; | |
1848 | ||
9a305230 LS |
1849 | hugetlb_unregister_all_nodes(); |
1850 | ||
a3437870 | 1851 | for_each_hstate(h) { |
972dc4de | 1852 | kobject_put(hstate_kobjs[hstate_index(h)]); |
a3437870 NA |
1853 | } |
1854 | ||
1855 | kobject_put(hugepages_kobj); | |
1856 | } | |
1857 | module_exit(hugetlb_exit); | |
1858 | ||
1859 | static int __init hugetlb_init(void) | |
1860 | { | |
0ef89d25 BH |
1861 | /* Some platform decide whether they support huge pages at boot |
1862 | * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when | |
1863 | * there is no such support | |
1864 | */ | |
1865 | if (HPAGE_SHIFT == 0) | |
1866 | return 0; | |
a3437870 | 1867 | |
e11bfbfc NP |
1868 | if (!size_to_hstate(default_hstate_size)) { |
1869 | default_hstate_size = HPAGE_SIZE; | |
1870 | if (!size_to_hstate(default_hstate_size)) | |
1871 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 1872 | } |
972dc4de | 1873 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
e11bfbfc NP |
1874 | if (default_hstate_max_huge_pages) |
1875 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
a3437870 NA |
1876 | |
1877 | hugetlb_init_hstates(); | |
1878 | ||
aa888a74 AK |
1879 | gather_bootmem_prealloc(); |
1880 | ||
a3437870 NA |
1881 | report_hugepages(); |
1882 | ||
1883 | hugetlb_sysfs_init(); | |
1884 | ||
9a305230 LS |
1885 | hugetlb_register_all_nodes(); |
1886 | ||
a3437870 NA |
1887 | return 0; |
1888 | } | |
1889 | module_init(hugetlb_init); | |
1890 | ||
1891 | /* Should be called on processing a hugepagesz=... option */ | |
1892 | void __init hugetlb_add_hstate(unsigned order) | |
1893 | { | |
1894 | struct hstate *h; | |
8faa8b07 AK |
1895 | unsigned long i; |
1896 | ||
a3437870 NA |
1897 | if (size_to_hstate(PAGE_SIZE << order)) { |
1898 | printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n"); | |
1899 | return; | |
1900 | } | |
47d38344 | 1901 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 1902 | BUG_ON(order == 0); |
47d38344 | 1903 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
1904 | h->order = order; |
1905 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
1906 | h->nr_huge_pages = 0; |
1907 | h->free_huge_pages = 0; | |
1908 | for (i = 0; i < MAX_NUMNODES; ++i) | |
1909 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 1910 | INIT_LIST_HEAD(&h->hugepage_activelist); |
9b5e5d0f LS |
1911 | h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]); |
1912 | h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]); | |
a3437870 NA |
1913 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
1914 | huge_page_size(h)/1024); | |
8faa8b07 | 1915 | |
a3437870 NA |
1916 | parsed_hstate = h; |
1917 | } | |
1918 | ||
e11bfbfc | 1919 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
1920 | { |
1921 | unsigned long *mhp; | |
8faa8b07 | 1922 | static unsigned long *last_mhp; |
a3437870 NA |
1923 | |
1924 | /* | |
47d38344 | 1925 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
1926 | * so this hugepages= parameter goes to the "default hstate". |
1927 | */ | |
47d38344 | 1928 | if (!hugetlb_max_hstate) |
a3437870 NA |
1929 | mhp = &default_hstate_max_huge_pages; |
1930 | else | |
1931 | mhp = &parsed_hstate->max_huge_pages; | |
1932 | ||
8faa8b07 AK |
1933 | if (mhp == last_mhp) { |
1934 | printk(KERN_WARNING "hugepages= specified twice without " | |
1935 | "interleaving hugepagesz=, ignoring\n"); | |
1936 | return 1; | |
1937 | } | |
1938 | ||
a3437870 NA |
1939 | if (sscanf(s, "%lu", mhp) <= 0) |
1940 | *mhp = 0; | |
1941 | ||
8faa8b07 AK |
1942 | /* |
1943 | * Global state is always initialized later in hugetlb_init. | |
1944 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
1945 | * use the bootmem allocator. | |
1946 | */ | |
47d38344 | 1947 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
1948 | hugetlb_hstate_alloc_pages(parsed_hstate); |
1949 | ||
1950 | last_mhp = mhp; | |
1951 | ||
a3437870 NA |
1952 | return 1; |
1953 | } | |
e11bfbfc NP |
1954 | __setup("hugepages=", hugetlb_nrpages_setup); |
1955 | ||
1956 | static int __init hugetlb_default_setup(char *s) | |
1957 | { | |
1958 | default_hstate_size = memparse(s, &s); | |
1959 | return 1; | |
1960 | } | |
1961 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 1962 | |
8a213460 NA |
1963 | static unsigned int cpuset_mems_nr(unsigned int *array) |
1964 | { | |
1965 | int node; | |
1966 | unsigned int nr = 0; | |
1967 | ||
1968 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
1969 | nr += array[node]; | |
1970 | ||
1971 | return nr; | |
1972 | } | |
1973 | ||
1974 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
1975 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
1976 | struct ctl_table *table, int write, | |
1977 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 1978 | { |
e5ff2159 AK |
1979 | struct hstate *h = &default_hstate; |
1980 | unsigned long tmp; | |
08d4a246 | 1981 | int ret; |
e5ff2159 | 1982 | |
c033a93c | 1983 | tmp = h->max_huge_pages; |
e5ff2159 | 1984 | |
adbe8726 EM |
1985 | if (write && h->order >= MAX_ORDER) |
1986 | return -EINVAL; | |
1987 | ||
e5ff2159 AK |
1988 | table->data = &tmp; |
1989 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
1990 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
1991 | if (ret) | |
1992 | goto out; | |
e5ff2159 | 1993 | |
06808b08 | 1994 | if (write) { |
bad44b5b DR |
1995 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, |
1996 | GFP_KERNEL | __GFP_NORETRY); | |
06808b08 LS |
1997 | if (!(obey_mempolicy && |
1998 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
1999 | NODEMASK_FREE(nodes_allowed); | |
2000 | nodes_allowed = &node_states[N_HIGH_MEMORY]; | |
2001 | } | |
2002 | h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); | |
2003 | ||
2004 | if (nodes_allowed != &node_states[N_HIGH_MEMORY]) | |
2005 | NODEMASK_FREE(nodes_allowed); | |
2006 | } | |
08d4a246 MH |
2007 | out: |
2008 | return ret; | |
1da177e4 | 2009 | } |
396faf03 | 2010 | |
06808b08 LS |
2011 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2012 | void __user *buffer, size_t *length, loff_t *ppos) | |
2013 | { | |
2014 | ||
2015 | return hugetlb_sysctl_handler_common(false, table, write, | |
2016 | buffer, length, ppos); | |
2017 | } | |
2018 | ||
2019 | #ifdef CONFIG_NUMA | |
2020 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2021 | void __user *buffer, size_t *length, loff_t *ppos) | |
2022 | { | |
2023 | return hugetlb_sysctl_handler_common(true, table, write, | |
2024 | buffer, length, ppos); | |
2025 | } | |
2026 | #endif /* CONFIG_NUMA */ | |
2027 | ||
396faf03 | 2028 | int hugetlb_treat_movable_handler(struct ctl_table *table, int write, |
8d65af78 | 2029 | void __user *buffer, |
396faf03 MG |
2030 | size_t *length, loff_t *ppos) |
2031 | { | |
8d65af78 | 2032 | proc_dointvec(table, write, buffer, length, ppos); |
396faf03 MG |
2033 | if (hugepages_treat_as_movable) |
2034 | htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; | |
2035 | else | |
2036 | htlb_alloc_mask = GFP_HIGHUSER; | |
2037 | return 0; | |
2038 | } | |
2039 | ||
a3d0c6aa | 2040 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 2041 | void __user *buffer, |
a3d0c6aa NA |
2042 | size_t *length, loff_t *ppos) |
2043 | { | |
a5516438 | 2044 | struct hstate *h = &default_hstate; |
e5ff2159 | 2045 | unsigned long tmp; |
08d4a246 | 2046 | int ret; |
e5ff2159 | 2047 | |
c033a93c | 2048 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 2049 | |
adbe8726 EM |
2050 | if (write && h->order >= MAX_ORDER) |
2051 | return -EINVAL; | |
2052 | ||
e5ff2159 AK |
2053 | table->data = &tmp; |
2054 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2055 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2056 | if (ret) | |
2057 | goto out; | |
e5ff2159 AK |
2058 | |
2059 | if (write) { | |
2060 | spin_lock(&hugetlb_lock); | |
2061 | h->nr_overcommit_huge_pages = tmp; | |
2062 | spin_unlock(&hugetlb_lock); | |
2063 | } | |
08d4a246 MH |
2064 | out: |
2065 | return ret; | |
a3d0c6aa NA |
2066 | } |
2067 | ||
1da177e4 LT |
2068 | #endif /* CONFIG_SYSCTL */ |
2069 | ||
e1759c21 | 2070 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 2071 | { |
a5516438 | 2072 | struct hstate *h = &default_hstate; |
e1759c21 | 2073 | seq_printf(m, |
4f98a2fe RR |
2074 | "HugePages_Total: %5lu\n" |
2075 | "HugePages_Free: %5lu\n" | |
2076 | "HugePages_Rsvd: %5lu\n" | |
2077 | "HugePages_Surp: %5lu\n" | |
2078 | "Hugepagesize: %8lu kB\n", | |
a5516438 AK |
2079 | h->nr_huge_pages, |
2080 | h->free_huge_pages, | |
2081 | h->resv_huge_pages, | |
2082 | h->surplus_huge_pages, | |
2083 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
2084 | } |
2085 | ||
2086 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
2087 | { | |
a5516438 | 2088 | struct hstate *h = &default_hstate; |
1da177e4 LT |
2089 | return sprintf(buf, |
2090 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
2091 | "Node %d HugePages_Free: %5u\n" |
2092 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
2093 | nid, h->nr_huge_pages_node[nid], |
2094 | nid, h->free_huge_pages_node[nid], | |
2095 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
2096 | } |
2097 | ||
1da177e4 LT |
2098 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
2099 | unsigned long hugetlb_total_pages(void) | |
2100 | { | |
a5516438 AK |
2101 | struct hstate *h = &default_hstate; |
2102 | return h->nr_huge_pages * pages_per_huge_page(h); | |
1da177e4 | 2103 | } |
1da177e4 | 2104 | |
a5516438 | 2105 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
2106 | { |
2107 | int ret = -ENOMEM; | |
2108 | ||
2109 | spin_lock(&hugetlb_lock); | |
2110 | /* | |
2111 | * When cpuset is configured, it breaks the strict hugetlb page | |
2112 | * reservation as the accounting is done on a global variable. Such | |
2113 | * reservation is completely rubbish in the presence of cpuset because | |
2114 | * the reservation is not checked against page availability for the | |
2115 | * current cpuset. Application can still potentially OOM'ed by kernel | |
2116 | * with lack of free htlb page in cpuset that the task is in. | |
2117 | * Attempt to enforce strict accounting with cpuset is almost | |
2118 | * impossible (or too ugly) because cpuset is too fluid that | |
2119 | * task or memory node can be dynamically moved between cpusets. | |
2120 | * | |
2121 | * The change of semantics for shared hugetlb mapping with cpuset is | |
2122 | * undesirable. However, in order to preserve some of the semantics, | |
2123 | * we fall back to check against current free page availability as | |
2124 | * a best attempt and hopefully to minimize the impact of changing | |
2125 | * semantics that cpuset has. | |
2126 | */ | |
2127 | if (delta > 0) { | |
a5516438 | 2128 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
2129 | goto out; |
2130 | ||
a5516438 AK |
2131 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
2132 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
2133 | goto out; |
2134 | } | |
2135 | } | |
2136 | ||
2137 | ret = 0; | |
2138 | if (delta < 0) | |
a5516438 | 2139 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
2140 | |
2141 | out: | |
2142 | spin_unlock(&hugetlb_lock); | |
2143 | return ret; | |
2144 | } | |
2145 | ||
84afd99b AW |
2146 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
2147 | { | |
2148 | struct resv_map *reservations = vma_resv_map(vma); | |
2149 | ||
2150 | /* | |
2151 | * This new VMA should share its siblings reservation map if present. | |
2152 | * The VMA will only ever have a valid reservation map pointer where | |
2153 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 2154 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
2155 | * after this open call completes. It is therefore safe to take a |
2156 | * new reference here without additional locking. | |
2157 | */ | |
2158 | if (reservations) | |
2159 | kref_get(&reservations->refs); | |
2160 | } | |
2161 | ||
c50ac050 DH |
2162 | static void resv_map_put(struct vm_area_struct *vma) |
2163 | { | |
2164 | struct resv_map *reservations = vma_resv_map(vma); | |
2165 | ||
2166 | if (!reservations) | |
2167 | return; | |
2168 | kref_put(&reservations->refs, resv_map_release); | |
2169 | } | |
2170 | ||
a1e78772 MG |
2171 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
2172 | { | |
a5516438 | 2173 | struct hstate *h = hstate_vma(vma); |
84afd99b | 2174 | struct resv_map *reservations = vma_resv_map(vma); |
90481622 | 2175 | struct hugepage_subpool *spool = subpool_vma(vma); |
84afd99b AW |
2176 | unsigned long reserve; |
2177 | unsigned long start; | |
2178 | unsigned long end; | |
2179 | ||
2180 | if (reservations) { | |
a5516438 AK |
2181 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
2182 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b AW |
2183 | |
2184 | reserve = (end - start) - | |
2185 | region_count(&reservations->regions, start, end); | |
2186 | ||
c50ac050 | 2187 | resv_map_put(vma); |
84afd99b | 2188 | |
7251ff78 | 2189 | if (reserve) { |
a5516438 | 2190 | hugetlb_acct_memory(h, -reserve); |
90481622 | 2191 | hugepage_subpool_put_pages(spool, reserve); |
7251ff78 | 2192 | } |
84afd99b | 2193 | } |
a1e78772 MG |
2194 | } |
2195 | ||
1da177e4 LT |
2196 | /* |
2197 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
2198 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
2199 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
2200 | * this far. | |
2201 | */ | |
d0217ac0 | 2202 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
2203 | { |
2204 | BUG(); | |
d0217ac0 | 2205 | return 0; |
1da177e4 LT |
2206 | } |
2207 | ||
f0f37e2f | 2208 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 2209 | .fault = hugetlb_vm_op_fault, |
84afd99b | 2210 | .open = hugetlb_vm_op_open, |
a1e78772 | 2211 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
2212 | }; |
2213 | ||
1e8f889b DG |
2214 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
2215 | int writable) | |
63551ae0 DG |
2216 | { |
2217 | pte_t entry; | |
2218 | ||
1e8f889b | 2219 | if (writable) { |
63551ae0 DG |
2220 | entry = |
2221 | pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); | |
2222 | } else { | |
7f2e9525 | 2223 | entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot)); |
63551ae0 DG |
2224 | } |
2225 | entry = pte_mkyoung(entry); | |
2226 | entry = pte_mkhuge(entry); | |
d9ed9faa | 2227 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
2228 | |
2229 | return entry; | |
2230 | } | |
2231 | ||
1e8f889b DG |
2232 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
2233 | unsigned long address, pte_t *ptep) | |
2234 | { | |
2235 | pte_t entry; | |
2236 | ||
7f2e9525 | 2237 | entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 2238 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 2239 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
2240 | } |
2241 | ||
2242 | ||
63551ae0 DG |
2243 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
2244 | struct vm_area_struct *vma) | |
2245 | { | |
2246 | pte_t *src_pte, *dst_pte, entry; | |
2247 | struct page *ptepage; | |
1c59827d | 2248 | unsigned long addr; |
1e8f889b | 2249 | int cow; |
a5516438 AK |
2250 | struct hstate *h = hstate_vma(vma); |
2251 | unsigned long sz = huge_page_size(h); | |
1e8f889b DG |
2252 | |
2253 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 2254 | |
a5516438 | 2255 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
c74df32c HD |
2256 | src_pte = huge_pte_offset(src, addr); |
2257 | if (!src_pte) | |
2258 | continue; | |
a5516438 | 2259 | dst_pte = huge_pte_alloc(dst, addr, sz); |
63551ae0 DG |
2260 | if (!dst_pte) |
2261 | goto nomem; | |
c5c99429 LW |
2262 | |
2263 | /* If the pagetables are shared don't copy or take references */ | |
2264 | if (dst_pte == src_pte) | |
2265 | continue; | |
2266 | ||
c74df32c | 2267 | spin_lock(&dst->page_table_lock); |
46478758 | 2268 | spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); |
7f2e9525 | 2269 | if (!huge_pte_none(huge_ptep_get(src_pte))) { |
1e8f889b | 2270 | if (cow) |
7f2e9525 GS |
2271 | huge_ptep_set_wrprotect(src, addr, src_pte); |
2272 | entry = huge_ptep_get(src_pte); | |
1c59827d HD |
2273 | ptepage = pte_page(entry); |
2274 | get_page(ptepage); | |
0fe6e20b | 2275 | page_dup_rmap(ptepage); |
1c59827d HD |
2276 | set_huge_pte_at(dst, addr, dst_pte, entry); |
2277 | } | |
2278 | spin_unlock(&src->page_table_lock); | |
c74df32c | 2279 | spin_unlock(&dst->page_table_lock); |
63551ae0 DG |
2280 | } |
2281 | return 0; | |
2282 | ||
2283 | nomem: | |
2284 | return -ENOMEM; | |
2285 | } | |
2286 | ||
290408d4 NH |
2287 | static int is_hugetlb_entry_migration(pte_t pte) |
2288 | { | |
2289 | swp_entry_t swp; | |
2290 | ||
2291 | if (huge_pte_none(pte) || pte_present(pte)) | |
2292 | return 0; | |
2293 | swp = pte_to_swp_entry(pte); | |
32f84528 | 2294 | if (non_swap_entry(swp) && is_migration_entry(swp)) |
290408d4 | 2295 | return 1; |
32f84528 | 2296 | else |
290408d4 NH |
2297 | return 0; |
2298 | } | |
2299 | ||
fd6a03ed NH |
2300 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) |
2301 | { | |
2302 | swp_entry_t swp; | |
2303 | ||
2304 | if (huge_pte_none(pte) || pte_present(pte)) | |
2305 | return 0; | |
2306 | swp = pte_to_swp_entry(pte); | |
32f84528 | 2307 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) |
fd6a03ed | 2308 | return 1; |
32f84528 | 2309 | else |
fd6a03ed NH |
2310 | return 0; |
2311 | } | |
2312 | ||
24669e58 AK |
2313 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
2314 | unsigned long start, unsigned long end, | |
2315 | struct page *ref_page) | |
63551ae0 | 2316 | { |
24669e58 | 2317 | int force_flush = 0; |
63551ae0 DG |
2318 | struct mm_struct *mm = vma->vm_mm; |
2319 | unsigned long address; | |
c7546f8f | 2320 | pte_t *ptep; |
63551ae0 DG |
2321 | pte_t pte; |
2322 | struct page *page; | |
a5516438 AK |
2323 | struct hstate *h = hstate_vma(vma); |
2324 | unsigned long sz = huge_page_size(h); | |
2325 | ||
63551ae0 | 2326 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
2327 | BUG_ON(start & ~huge_page_mask(h)); |
2328 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 2329 | |
24669e58 | 2330 | tlb_start_vma(tlb, vma); |
cddb8a5c | 2331 | mmu_notifier_invalidate_range_start(mm, start, end); |
24669e58 | 2332 | again: |
508034a3 | 2333 | spin_lock(&mm->page_table_lock); |
a5516438 | 2334 | for (address = start; address < end; address += sz) { |
c7546f8f | 2335 | ptep = huge_pte_offset(mm, address); |
4c887265 | 2336 | if (!ptep) |
c7546f8f DG |
2337 | continue; |
2338 | ||
39dde65c KC |
2339 | if (huge_pmd_unshare(mm, &address, ptep)) |
2340 | continue; | |
2341 | ||
6629326b HD |
2342 | pte = huge_ptep_get(ptep); |
2343 | if (huge_pte_none(pte)) | |
2344 | continue; | |
2345 | ||
2346 | /* | |
2347 | * HWPoisoned hugepage is already unmapped and dropped reference | |
2348 | */ | |
2349 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) | |
2350 | continue; | |
2351 | ||
2352 | page = pte_page(pte); | |
04f2cbe3 MG |
2353 | /* |
2354 | * If a reference page is supplied, it is because a specific | |
2355 | * page is being unmapped, not a range. Ensure the page we | |
2356 | * are about to unmap is the actual page of interest. | |
2357 | */ | |
2358 | if (ref_page) { | |
04f2cbe3 MG |
2359 | if (page != ref_page) |
2360 | continue; | |
2361 | ||
2362 | /* | |
2363 | * Mark the VMA as having unmapped its page so that | |
2364 | * future faults in this VMA will fail rather than | |
2365 | * looking like data was lost | |
2366 | */ | |
2367 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
2368 | } | |
2369 | ||
c7546f8f | 2370 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
24669e58 | 2371 | tlb_remove_tlb_entry(tlb, ptep, address); |
6649a386 KC |
2372 | if (pte_dirty(pte)) |
2373 | set_page_dirty(page); | |
9e81130b | 2374 | |
24669e58 AK |
2375 | page_remove_rmap(page); |
2376 | force_flush = !__tlb_remove_page(tlb, page); | |
2377 | if (force_flush) | |
2378 | break; | |
9e81130b HD |
2379 | /* Bail out after unmapping reference page if supplied */ |
2380 | if (ref_page) | |
2381 | break; | |
63551ae0 | 2382 | } |
cd2934a3 | 2383 | spin_unlock(&mm->page_table_lock); |
24669e58 AK |
2384 | /* |
2385 | * mmu_gather ran out of room to batch pages, we break out of | |
2386 | * the PTE lock to avoid doing the potential expensive TLB invalidate | |
2387 | * and page-free while holding it. | |
2388 | */ | |
2389 | if (force_flush) { | |
2390 | force_flush = 0; | |
2391 | tlb_flush_mmu(tlb); | |
2392 | if (address < end && !ref_page) | |
2393 | goto again; | |
fe1668ae | 2394 | } |
24669e58 AK |
2395 | mmu_notifier_invalidate_range_end(mm, start, end); |
2396 | tlb_end_vma(tlb, vma); | |
1da177e4 | 2397 | } |
63551ae0 | 2398 | |
502717f4 | 2399 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 2400 | unsigned long end, struct page *ref_page) |
502717f4 | 2401 | { |
24669e58 AK |
2402 | struct mm_struct *mm; |
2403 | struct mmu_gather tlb; | |
2404 | ||
2405 | mm = vma->vm_mm; | |
2406 | ||
2407 | tlb_gather_mmu(&tlb, mm, 0); | |
2408 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); | |
2409 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 KC |
2410 | } |
2411 | ||
04f2cbe3 MG |
2412 | /* |
2413 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
2414 | * mappping it owns the reserve page for. The intention is to unmap the page | |
2415 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
2416 | * same region. | |
2417 | */ | |
2a4b3ded HH |
2418 | static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
2419 | struct page *page, unsigned long address) | |
04f2cbe3 | 2420 | { |
7526674d | 2421 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
2422 | struct vm_area_struct *iter_vma; |
2423 | struct address_space *mapping; | |
2424 | struct prio_tree_iter iter; | |
2425 | pgoff_t pgoff; | |
2426 | ||
2427 | /* | |
2428 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
2429 | * from page cache lookup which is in HPAGE_SIZE units. | |
2430 | */ | |
7526674d | 2431 | address = address & huge_page_mask(h); |
0c176d52 | 2432 | pgoff = vma_hugecache_offset(h, vma, address); |
90481622 | 2433 | mapping = vma->vm_file->f_dentry->d_inode->i_mapping; |
04f2cbe3 | 2434 | |
4eb2b1dc MG |
2435 | /* |
2436 | * Take the mapping lock for the duration of the table walk. As | |
2437 | * this mapping should be shared between all the VMAs, | |
2438 | * __unmap_hugepage_range() is called as the lock is already held | |
2439 | */ | |
3d48ae45 | 2440 | mutex_lock(&mapping->i_mmap_mutex); |
04f2cbe3 MG |
2441 | vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) { |
2442 | /* Do not unmap the current VMA */ | |
2443 | if (iter_vma == vma) | |
2444 | continue; | |
2445 | ||
2446 | /* | |
2447 | * Unmap the page from other VMAs without their own reserves. | |
2448 | * They get marked to be SIGKILLed if they fault in these | |
2449 | * areas. This is because a future no-page fault on this VMA | |
2450 | * could insert a zeroed page instead of the data existing | |
2451 | * from the time of fork. This would look like data corruption | |
2452 | */ | |
2453 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
2454 | unmap_hugepage_range(iter_vma, address, |
2455 | address + huge_page_size(h), page); | |
04f2cbe3 | 2456 | } |
3d48ae45 | 2457 | mutex_unlock(&mapping->i_mmap_mutex); |
04f2cbe3 MG |
2458 | |
2459 | return 1; | |
2460 | } | |
2461 | ||
0fe6e20b NH |
2462 | /* |
2463 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
2464 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
2465 | * cannot race with other handlers or page migration. | |
2466 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 2467 | */ |
1e8f889b | 2468 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
04f2cbe3 MG |
2469 | unsigned long address, pte_t *ptep, pte_t pte, |
2470 | struct page *pagecache_page) | |
1e8f889b | 2471 | { |
a5516438 | 2472 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 2473 | struct page *old_page, *new_page; |
79ac6ba4 | 2474 | int avoidcopy; |
04f2cbe3 | 2475 | int outside_reserve = 0; |
1e8f889b DG |
2476 | |
2477 | old_page = pte_page(pte); | |
2478 | ||
04f2cbe3 | 2479 | retry_avoidcopy: |
1e8f889b DG |
2480 | /* If no-one else is actually using this page, avoid the copy |
2481 | * and just make the page writable */ | |
0fe6e20b | 2482 | avoidcopy = (page_mapcount(old_page) == 1); |
1e8f889b | 2483 | if (avoidcopy) { |
56c9cfb1 NH |
2484 | if (PageAnon(old_page)) |
2485 | page_move_anon_rmap(old_page, vma, address); | |
1e8f889b | 2486 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 2487 | return 0; |
1e8f889b DG |
2488 | } |
2489 | ||
04f2cbe3 MG |
2490 | /* |
2491 | * If the process that created a MAP_PRIVATE mapping is about to | |
2492 | * perform a COW due to a shared page count, attempt to satisfy | |
2493 | * the allocation without using the existing reserves. The pagecache | |
2494 | * page is used to determine if the reserve at this address was | |
2495 | * consumed or not. If reserves were used, a partial faulted mapping | |
2496 | * at the time of fork() could consume its reserves on COW instead | |
2497 | * of the full address range. | |
2498 | */ | |
f83a275d | 2499 | if (!(vma->vm_flags & VM_MAYSHARE) && |
04f2cbe3 MG |
2500 | is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
2501 | old_page != pagecache_page) | |
2502 | outside_reserve = 1; | |
2503 | ||
1e8f889b | 2504 | page_cache_get(old_page); |
b76c8cfb LW |
2505 | |
2506 | /* Drop page_table_lock as buddy allocator may be called */ | |
2507 | spin_unlock(&mm->page_table_lock); | |
04f2cbe3 | 2508 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 2509 | |
2fc39cec | 2510 | if (IS_ERR(new_page)) { |
76dcee75 | 2511 | long err = PTR_ERR(new_page); |
1e8f889b | 2512 | page_cache_release(old_page); |
04f2cbe3 MG |
2513 | |
2514 | /* | |
2515 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
2516 | * it is due to references held by a child and an insufficient | |
2517 | * huge page pool. To guarantee the original mappers | |
2518 | * reliability, unmap the page from child processes. The child | |
2519 | * may get SIGKILLed if it later faults. | |
2520 | */ | |
2521 | if (outside_reserve) { | |
2522 | BUG_ON(huge_pte_none(pte)); | |
2523 | if (unmap_ref_private(mm, vma, old_page, address)) { | |
04f2cbe3 | 2524 | BUG_ON(huge_pte_none(pte)); |
b76c8cfb | 2525 | spin_lock(&mm->page_table_lock); |
a734bcc8 HD |
2526 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
2527 | if (likely(pte_same(huge_ptep_get(ptep), pte))) | |
2528 | goto retry_avoidcopy; | |
2529 | /* | |
2530 | * race occurs while re-acquiring page_table_lock, and | |
2531 | * our job is done. | |
2532 | */ | |
2533 | return 0; | |
04f2cbe3 MG |
2534 | } |
2535 | WARN_ON_ONCE(1); | |
2536 | } | |
2537 | ||
b76c8cfb LW |
2538 | /* Caller expects lock to be held */ |
2539 | spin_lock(&mm->page_table_lock); | |
76dcee75 AK |
2540 | if (err == -ENOMEM) |
2541 | return VM_FAULT_OOM; | |
2542 | else | |
2543 | return VM_FAULT_SIGBUS; | |
1e8f889b DG |
2544 | } |
2545 | ||
0fe6e20b NH |
2546 | /* |
2547 | * When the original hugepage is shared one, it does not have | |
2548 | * anon_vma prepared. | |
2549 | */ | |
44e2aa93 | 2550 | if (unlikely(anon_vma_prepare(vma))) { |
ea4039a3 HD |
2551 | page_cache_release(new_page); |
2552 | page_cache_release(old_page); | |
44e2aa93 DN |
2553 | /* Caller expects lock to be held */ |
2554 | spin_lock(&mm->page_table_lock); | |
0fe6e20b | 2555 | return VM_FAULT_OOM; |
44e2aa93 | 2556 | } |
0fe6e20b | 2557 | |
47ad8475 AA |
2558 | copy_user_huge_page(new_page, old_page, address, vma, |
2559 | pages_per_huge_page(h)); | |
0ed361de | 2560 | __SetPageUptodate(new_page); |
1e8f889b | 2561 | |
b76c8cfb LW |
2562 | /* |
2563 | * Retake the page_table_lock to check for racing updates | |
2564 | * before the page tables are altered | |
2565 | */ | |
2566 | spin_lock(&mm->page_table_lock); | |
a5516438 | 2567 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
7f2e9525 | 2568 | if (likely(pte_same(huge_ptep_get(ptep), pte))) { |
1e8f889b | 2569 | /* Break COW */ |
3edd4fc9 DD |
2570 | mmu_notifier_invalidate_range_start(mm, |
2571 | address & huge_page_mask(h), | |
2572 | (address & huge_page_mask(h)) + huge_page_size(h)); | |
8fe627ec | 2573 | huge_ptep_clear_flush(vma, address, ptep); |
1e8f889b DG |
2574 | set_huge_pte_at(mm, address, ptep, |
2575 | make_huge_pte(vma, new_page, 1)); | |
0fe6e20b | 2576 | page_remove_rmap(old_page); |
cd67f0d2 | 2577 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
2578 | /* Make the old page be freed below */ |
2579 | new_page = old_page; | |
3edd4fc9 DD |
2580 | mmu_notifier_invalidate_range_end(mm, |
2581 | address & huge_page_mask(h), | |
2582 | (address & huge_page_mask(h)) + huge_page_size(h)); | |
1e8f889b DG |
2583 | } |
2584 | page_cache_release(new_page); | |
2585 | page_cache_release(old_page); | |
83c54070 | 2586 | return 0; |
1e8f889b DG |
2587 | } |
2588 | ||
04f2cbe3 | 2589 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
2590 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
2591 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
2592 | { |
2593 | struct address_space *mapping; | |
e7c4b0bf | 2594 | pgoff_t idx; |
04f2cbe3 MG |
2595 | |
2596 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 2597 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
2598 | |
2599 | return find_lock_page(mapping, idx); | |
2600 | } | |
2601 | ||
3ae77f43 HD |
2602 | /* |
2603 | * Return whether there is a pagecache page to back given address within VMA. | |
2604 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
2605 | */ | |
2606 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
2607 | struct vm_area_struct *vma, unsigned long address) |
2608 | { | |
2609 | struct address_space *mapping; | |
2610 | pgoff_t idx; | |
2611 | struct page *page; | |
2612 | ||
2613 | mapping = vma->vm_file->f_mapping; | |
2614 | idx = vma_hugecache_offset(h, vma, address); | |
2615 | ||
2616 | page = find_get_page(mapping, idx); | |
2617 | if (page) | |
2618 | put_page(page); | |
2619 | return page != NULL; | |
2620 | } | |
2621 | ||
a1ed3dda | 2622 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 2623 | unsigned long address, pte_t *ptep, unsigned int flags) |
ac9b9c66 | 2624 | { |
a5516438 | 2625 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 2626 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 2627 | int anon_rmap = 0; |
e7c4b0bf | 2628 | pgoff_t idx; |
4c887265 | 2629 | unsigned long size; |
4c887265 AL |
2630 | struct page *page; |
2631 | struct address_space *mapping; | |
1e8f889b | 2632 | pte_t new_pte; |
4c887265 | 2633 | |
04f2cbe3 MG |
2634 | /* |
2635 | * Currently, we are forced to kill the process in the event the | |
2636 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 2637 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
2638 | */ |
2639 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
2640 | printk(KERN_WARNING | |
2641 | "PID %d killed due to inadequate hugepage pool\n", | |
2642 | current->pid); | |
2643 | return ret; | |
2644 | } | |
2645 | ||
4c887265 | 2646 | mapping = vma->vm_file->f_mapping; |
a5516438 | 2647 | idx = vma_hugecache_offset(h, vma, address); |
4c887265 AL |
2648 | |
2649 | /* | |
2650 | * Use page lock to guard against racing truncation | |
2651 | * before we get page_table_lock. | |
2652 | */ | |
6bda666a CL |
2653 | retry: |
2654 | page = find_lock_page(mapping, idx); | |
2655 | if (!page) { | |
a5516438 | 2656 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
2657 | if (idx >= size) |
2658 | goto out; | |
04f2cbe3 | 2659 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 2660 | if (IS_ERR(page)) { |
76dcee75 AK |
2661 | ret = PTR_ERR(page); |
2662 | if (ret == -ENOMEM) | |
2663 | ret = VM_FAULT_OOM; | |
2664 | else | |
2665 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
2666 | goto out; |
2667 | } | |
47ad8475 | 2668 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 2669 | __SetPageUptodate(page); |
ac9b9c66 | 2670 | |
f83a275d | 2671 | if (vma->vm_flags & VM_MAYSHARE) { |
6bda666a | 2672 | int err; |
45c682a6 | 2673 | struct inode *inode = mapping->host; |
6bda666a CL |
2674 | |
2675 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
2676 | if (err) { | |
2677 | put_page(page); | |
6bda666a CL |
2678 | if (err == -EEXIST) |
2679 | goto retry; | |
2680 | goto out; | |
2681 | } | |
45c682a6 KC |
2682 | |
2683 | spin_lock(&inode->i_lock); | |
a5516438 | 2684 | inode->i_blocks += blocks_per_huge_page(h); |
45c682a6 | 2685 | spin_unlock(&inode->i_lock); |
23be7468 | 2686 | } else { |
6bda666a | 2687 | lock_page(page); |
0fe6e20b NH |
2688 | if (unlikely(anon_vma_prepare(vma))) { |
2689 | ret = VM_FAULT_OOM; | |
2690 | goto backout_unlocked; | |
2691 | } | |
409eb8c2 | 2692 | anon_rmap = 1; |
23be7468 | 2693 | } |
0fe6e20b | 2694 | } else { |
998b4382 NH |
2695 | /* |
2696 | * If memory error occurs between mmap() and fault, some process | |
2697 | * don't have hwpoisoned swap entry for errored virtual address. | |
2698 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
2699 | */ | |
2700 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 2701 | ret = VM_FAULT_HWPOISON | |
972dc4de | 2702 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
2703 | goto backout_unlocked; |
2704 | } | |
6bda666a | 2705 | } |
1e8f889b | 2706 | |
57303d80 AW |
2707 | /* |
2708 | * If we are going to COW a private mapping later, we examine the | |
2709 | * pending reservations for this page now. This will ensure that | |
2710 | * any allocations necessary to record that reservation occur outside | |
2711 | * the spinlock. | |
2712 | */ | |
788c7df4 | 2713 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) |
2b26736c AW |
2714 | if (vma_needs_reservation(h, vma, address) < 0) { |
2715 | ret = VM_FAULT_OOM; | |
2716 | goto backout_unlocked; | |
2717 | } | |
57303d80 | 2718 | |
ac9b9c66 | 2719 | spin_lock(&mm->page_table_lock); |
a5516438 | 2720 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
2721 | if (idx >= size) |
2722 | goto backout; | |
2723 | ||
83c54070 | 2724 | ret = 0; |
7f2e9525 | 2725 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
2726 | goto backout; |
2727 | ||
409eb8c2 HD |
2728 | if (anon_rmap) |
2729 | hugepage_add_new_anon_rmap(page, vma, address); | |
2730 | else | |
2731 | page_dup_rmap(page); | |
1e8f889b DG |
2732 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
2733 | && (vma->vm_flags & VM_SHARED))); | |
2734 | set_huge_pte_at(mm, address, ptep, new_pte); | |
2735 | ||
788c7df4 | 2736 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 2737 | /* Optimization, do the COW without a second fault */ |
04f2cbe3 | 2738 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); |
1e8f889b DG |
2739 | } |
2740 | ||
ac9b9c66 | 2741 | spin_unlock(&mm->page_table_lock); |
4c887265 AL |
2742 | unlock_page(page); |
2743 | out: | |
ac9b9c66 | 2744 | return ret; |
4c887265 AL |
2745 | |
2746 | backout: | |
2747 | spin_unlock(&mm->page_table_lock); | |
2b26736c | 2748 | backout_unlocked: |
4c887265 AL |
2749 | unlock_page(page); |
2750 | put_page(page); | |
2751 | goto out; | |
ac9b9c66 HD |
2752 | } |
2753 | ||
86e5216f | 2754 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 2755 | unsigned long address, unsigned int flags) |
86e5216f AL |
2756 | { |
2757 | pte_t *ptep; | |
2758 | pte_t entry; | |
1e8f889b | 2759 | int ret; |
0fe6e20b | 2760 | struct page *page = NULL; |
57303d80 | 2761 | struct page *pagecache_page = NULL; |
3935baa9 | 2762 | static DEFINE_MUTEX(hugetlb_instantiation_mutex); |
a5516438 | 2763 | struct hstate *h = hstate_vma(vma); |
86e5216f | 2764 | |
1e16a539 KH |
2765 | address &= huge_page_mask(h); |
2766 | ||
fd6a03ed NH |
2767 | ptep = huge_pte_offset(mm, address); |
2768 | if (ptep) { | |
2769 | entry = huge_ptep_get(ptep); | |
290408d4 NH |
2770 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
2771 | migration_entry_wait(mm, (pmd_t *)ptep, address); | |
2772 | return 0; | |
2773 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 2774 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 2775 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
2776 | } |
2777 | ||
a5516438 | 2778 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); |
86e5216f AL |
2779 | if (!ptep) |
2780 | return VM_FAULT_OOM; | |
2781 | ||
3935baa9 DG |
2782 | /* |
2783 | * Serialize hugepage allocation and instantiation, so that we don't | |
2784 | * get spurious allocation failures if two CPUs race to instantiate | |
2785 | * the same page in the page cache. | |
2786 | */ | |
2787 | mutex_lock(&hugetlb_instantiation_mutex); | |
7f2e9525 GS |
2788 | entry = huge_ptep_get(ptep); |
2789 | if (huge_pte_none(entry)) { | |
788c7df4 | 2790 | ret = hugetlb_no_page(mm, vma, address, ptep, flags); |
b4d1d99f | 2791 | goto out_mutex; |
3935baa9 | 2792 | } |
86e5216f | 2793 | |
83c54070 | 2794 | ret = 0; |
1e8f889b | 2795 | |
57303d80 AW |
2796 | /* |
2797 | * If we are going to COW the mapping later, we examine the pending | |
2798 | * reservations for this page now. This will ensure that any | |
2799 | * allocations necessary to record that reservation occur outside the | |
2800 | * spinlock. For private mappings, we also lookup the pagecache | |
2801 | * page now as it is used to determine if a reservation has been | |
2802 | * consumed. | |
2803 | */ | |
788c7df4 | 2804 | if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) { |
2b26736c AW |
2805 | if (vma_needs_reservation(h, vma, address) < 0) { |
2806 | ret = VM_FAULT_OOM; | |
b4d1d99f | 2807 | goto out_mutex; |
2b26736c | 2808 | } |
57303d80 | 2809 | |
f83a275d | 2810 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
2811 | pagecache_page = hugetlbfs_pagecache_page(h, |
2812 | vma, address); | |
2813 | } | |
2814 | ||
56c9cfb1 NH |
2815 | /* |
2816 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
2817 | * pagecache_page, so here we need take the former one | |
2818 | * when page != pagecache_page or !pagecache_page. | |
2819 | * Note that locking order is always pagecache_page -> page, | |
2820 | * so no worry about deadlock. | |
2821 | */ | |
2822 | page = pte_page(entry); | |
66aebce7 | 2823 | get_page(page); |
56c9cfb1 | 2824 | if (page != pagecache_page) |
0fe6e20b | 2825 | lock_page(page); |
0fe6e20b | 2826 | |
1e8f889b DG |
2827 | spin_lock(&mm->page_table_lock); |
2828 | /* Check for a racing update before calling hugetlb_cow */ | |
b4d1d99f DG |
2829 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) |
2830 | goto out_page_table_lock; | |
2831 | ||
2832 | ||
788c7df4 | 2833 | if (flags & FAULT_FLAG_WRITE) { |
b4d1d99f | 2834 | if (!pte_write(entry)) { |
57303d80 AW |
2835 | ret = hugetlb_cow(mm, vma, address, ptep, entry, |
2836 | pagecache_page); | |
b4d1d99f DG |
2837 | goto out_page_table_lock; |
2838 | } | |
2839 | entry = pte_mkdirty(entry); | |
2840 | } | |
2841 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
2842 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
2843 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 2844 | update_mmu_cache(vma, address, ptep); |
b4d1d99f DG |
2845 | |
2846 | out_page_table_lock: | |
1e8f889b | 2847 | spin_unlock(&mm->page_table_lock); |
57303d80 AW |
2848 | |
2849 | if (pagecache_page) { | |
2850 | unlock_page(pagecache_page); | |
2851 | put_page(pagecache_page); | |
2852 | } | |
1f64d69c DN |
2853 | if (page != pagecache_page) |
2854 | unlock_page(page); | |
66aebce7 | 2855 | put_page(page); |
57303d80 | 2856 | |
b4d1d99f | 2857 | out_mutex: |
3935baa9 | 2858 | mutex_unlock(&hugetlb_instantiation_mutex); |
1e8f889b DG |
2859 | |
2860 | return ret; | |
86e5216f AL |
2861 | } |
2862 | ||
ceb86879 AK |
2863 | /* Can be overriden by architectures */ |
2864 | __attribute__((weak)) struct page * | |
2865 | follow_huge_pud(struct mm_struct *mm, unsigned long address, | |
2866 | pud_t *pud, int write) | |
2867 | { | |
2868 | BUG(); | |
2869 | return NULL; | |
2870 | } | |
2871 | ||
63551ae0 DG |
2872 | int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
2873 | struct page **pages, struct vm_area_struct **vmas, | |
5b23dbe8 | 2874 | unsigned long *position, int *length, int i, |
2a15efc9 | 2875 | unsigned int flags) |
63551ae0 | 2876 | { |
d5d4b0aa KC |
2877 | unsigned long pfn_offset; |
2878 | unsigned long vaddr = *position; | |
63551ae0 | 2879 | int remainder = *length; |
a5516438 | 2880 | struct hstate *h = hstate_vma(vma); |
63551ae0 | 2881 | |
1c59827d | 2882 | spin_lock(&mm->page_table_lock); |
63551ae0 | 2883 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 2884 | pte_t *pte; |
2a15efc9 | 2885 | int absent; |
4c887265 | 2886 | struct page *page; |
63551ae0 | 2887 | |
4c887265 AL |
2888 | /* |
2889 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 2890 | * each hugepage. We have to make sure we get the |
4c887265 AL |
2891 | * first, for the page indexing below to work. |
2892 | */ | |
a5516438 | 2893 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); |
2a15efc9 HD |
2894 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
2895 | ||
2896 | /* | |
2897 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
2898 | * an error where there's an empty slot with no huge pagecache |
2899 | * to back it. This way, we avoid allocating a hugepage, and | |
2900 | * the sparse dumpfile avoids allocating disk blocks, but its | |
2901 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 2902 | */ |
3ae77f43 HD |
2903 | if (absent && (flags & FOLL_DUMP) && |
2904 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
2a15efc9 HD |
2905 | remainder = 0; |
2906 | break; | |
2907 | } | |
63551ae0 | 2908 | |
2a15efc9 HD |
2909 | if (absent || |
2910 | ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 2911 | int ret; |
63551ae0 | 2912 | |
4c887265 | 2913 | spin_unlock(&mm->page_table_lock); |
2a15efc9 HD |
2914 | ret = hugetlb_fault(mm, vma, vaddr, |
2915 | (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); | |
4c887265 | 2916 | spin_lock(&mm->page_table_lock); |
a89182c7 | 2917 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 2918 | continue; |
63551ae0 | 2919 | |
4c887265 | 2920 | remainder = 0; |
4c887265 AL |
2921 | break; |
2922 | } | |
2923 | ||
a5516438 | 2924 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 2925 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 2926 | same_page: |
d6692183 | 2927 | if (pages) { |
2a15efc9 | 2928 | pages[i] = mem_map_offset(page, pfn_offset); |
4b2e38ad | 2929 | get_page(pages[i]); |
d6692183 | 2930 | } |
63551ae0 DG |
2931 | |
2932 | if (vmas) | |
2933 | vmas[i] = vma; | |
2934 | ||
2935 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 2936 | ++pfn_offset; |
63551ae0 DG |
2937 | --remainder; |
2938 | ++i; | |
d5d4b0aa | 2939 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 2940 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa KC |
2941 | /* |
2942 | * We use pfn_offset to avoid touching the pageframes | |
2943 | * of this compound page. | |
2944 | */ | |
2945 | goto same_page; | |
2946 | } | |
63551ae0 | 2947 | } |
1c59827d | 2948 | spin_unlock(&mm->page_table_lock); |
63551ae0 DG |
2949 | *length = remainder; |
2950 | *position = vaddr; | |
2951 | ||
2a15efc9 | 2952 | return i ? i : -EFAULT; |
63551ae0 | 2953 | } |
8f860591 ZY |
2954 | |
2955 | void hugetlb_change_protection(struct vm_area_struct *vma, | |
2956 | unsigned long address, unsigned long end, pgprot_t newprot) | |
2957 | { | |
2958 | struct mm_struct *mm = vma->vm_mm; | |
2959 | unsigned long start = address; | |
2960 | pte_t *ptep; | |
2961 | pte_t pte; | |
a5516438 | 2962 | struct hstate *h = hstate_vma(vma); |
8f860591 ZY |
2963 | |
2964 | BUG_ON(address >= end); | |
2965 | flush_cache_range(vma, address, end); | |
2966 | ||
3d48ae45 | 2967 | mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex); |
8f860591 | 2968 | spin_lock(&mm->page_table_lock); |
a5516438 | 2969 | for (; address < end; address += huge_page_size(h)) { |
8f860591 ZY |
2970 | ptep = huge_pte_offset(mm, address); |
2971 | if (!ptep) | |
2972 | continue; | |
39dde65c KC |
2973 | if (huge_pmd_unshare(mm, &address, ptep)) |
2974 | continue; | |
7f2e9525 | 2975 | if (!huge_pte_none(huge_ptep_get(ptep))) { |
8f860591 ZY |
2976 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
2977 | pte = pte_mkhuge(pte_modify(pte, newprot)); | |
2978 | set_huge_pte_at(mm, address, ptep, pte); | |
8f860591 ZY |
2979 | } |
2980 | } | |
2981 | spin_unlock(&mm->page_table_lock); | |
3d48ae45 | 2982 | mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex); |
8f860591 ZY |
2983 | |
2984 | flush_tlb_range(vma, start, end); | |
2985 | } | |
2986 | ||
a1e78772 MG |
2987 | int hugetlb_reserve_pages(struct inode *inode, |
2988 | long from, long to, | |
5a6fe125 | 2989 | struct vm_area_struct *vma, |
ca16d140 | 2990 | vm_flags_t vm_flags) |
e4e574b7 | 2991 | { |
17c9d12e | 2992 | long ret, chg; |
a5516438 | 2993 | struct hstate *h = hstate_inode(inode); |
90481622 | 2994 | struct hugepage_subpool *spool = subpool_inode(inode); |
e4e574b7 | 2995 | |
17c9d12e MG |
2996 | /* |
2997 | * Only apply hugepage reservation if asked. At fault time, an | |
2998 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 2999 | * without using reserves |
17c9d12e | 3000 | */ |
ca16d140 | 3001 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
3002 | return 0; |
3003 | ||
a1e78772 MG |
3004 | /* |
3005 | * Shared mappings base their reservation on the number of pages that | |
3006 | * are already allocated on behalf of the file. Private mappings need | |
3007 | * to reserve the full area even if read-only as mprotect() may be | |
3008 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
3009 | */ | |
f83a275d | 3010 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
a1e78772 | 3011 | chg = region_chg(&inode->i_mapping->private_list, from, to); |
17c9d12e MG |
3012 | else { |
3013 | struct resv_map *resv_map = resv_map_alloc(); | |
3014 | if (!resv_map) | |
3015 | return -ENOMEM; | |
3016 | ||
a1e78772 | 3017 | chg = to - from; |
84afd99b | 3018 | |
17c9d12e MG |
3019 | set_vma_resv_map(vma, resv_map); |
3020 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
3021 | } | |
3022 | ||
c50ac050 DH |
3023 | if (chg < 0) { |
3024 | ret = chg; | |
3025 | goto out_err; | |
3026 | } | |
8a630112 | 3027 | |
90481622 | 3028 | /* There must be enough pages in the subpool for the mapping */ |
c50ac050 DH |
3029 | if (hugepage_subpool_get_pages(spool, chg)) { |
3030 | ret = -ENOSPC; | |
3031 | goto out_err; | |
3032 | } | |
5a6fe125 MG |
3033 | |
3034 | /* | |
17c9d12e | 3035 | * Check enough hugepages are available for the reservation. |
90481622 | 3036 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 3037 | */ |
a5516438 | 3038 | ret = hugetlb_acct_memory(h, chg); |
68842c9b | 3039 | if (ret < 0) { |
90481622 | 3040 | hugepage_subpool_put_pages(spool, chg); |
c50ac050 | 3041 | goto out_err; |
68842c9b | 3042 | } |
17c9d12e MG |
3043 | |
3044 | /* | |
3045 | * Account for the reservations made. Shared mappings record regions | |
3046 | * that have reservations as they are shared by multiple VMAs. | |
3047 | * When the last VMA disappears, the region map says how much | |
3048 | * the reservation was and the page cache tells how much of | |
3049 | * the reservation was consumed. Private mappings are per-VMA and | |
3050 | * only the consumed reservations are tracked. When the VMA | |
3051 | * disappears, the original reservation is the VMA size and the | |
3052 | * consumed reservations are stored in the map. Hence, nothing | |
3053 | * else has to be done for private mappings here | |
3054 | */ | |
f83a275d | 3055 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
a1e78772 | 3056 | region_add(&inode->i_mapping->private_list, from, to); |
a43a8c39 | 3057 | return 0; |
c50ac050 | 3058 | out_err: |
4523e145 DH |
3059 | if (vma) |
3060 | resv_map_put(vma); | |
c50ac050 | 3061 | return ret; |
a43a8c39 KC |
3062 | } |
3063 | ||
3064 | void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) | |
3065 | { | |
a5516438 | 3066 | struct hstate *h = hstate_inode(inode); |
a43a8c39 | 3067 | long chg = region_truncate(&inode->i_mapping->private_list, offset); |
90481622 | 3068 | struct hugepage_subpool *spool = subpool_inode(inode); |
45c682a6 KC |
3069 | |
3070 | spin_lock(&inode->i_lock); | |
e4c6f8be | 3071 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
3072 | spin_unlock(&inode->i_lock); |
3073 | ||
90481622 | 3074 | hugepage_subpool_put_pages(spool, (chg - freed)); |
a5516438 | 3075 | hugetlb_acct_memory(h, -(chg - freed)); |
a43a8c39 | 3076 | } |
93f70f90 | 3077 | |
d5bd9106 AK |
3078 | #ifdef CONFIG_MEMORY_FAILURE |
3079 | ||
6de2b1aa NH |
3080 | /* Should be called in hugetlb_lock */ |
3081 | static int is_hugepage_on_freelist(struct page *hpage) | |
3082 | { | |
3083 | struct page *page; | |
3084 | struct page *tmp; | |
3085 | struct hstate *h = page_hstate(hpage); | |
3086 | int nid = page_to_nid(hpage); | |
3087 | ||
3088 | list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru) | |
3089 | if (page == hpage) | |
3090 | return 1; | |
3091 | return 0; | |
3092 | } | |
3093 | ||
93f70f90 NH |
3094 | /* |
3095 | * This function is called from memory failure code. | |
3096 | * Assume the caller holds page lock of the head page. | |
3097 | */ | |
6de2b1aa | 3098 | int dequeue_hwpoisoned_huge_page(struct page *hpage) |
93f70f90 NH |
3099 | { |
3100 | struct hstate *h = page_hstate(hpage); | |
3101 | int nid = page_to_nid(hpage); | |
6de2b1aa | 3102 | int ret = -EBUSY; |
93f70f90 NH |
3103 | |
3104 | spin_lock(&hugetlb_lock); | |
6de2b1aa NH |
3105 | if (is_hugepage_on_freelist(hpage)) { |
3106 | list_del(&hpage->lru); | |
8c6c2ecb | 3107 | set_page_refcounted(hpage); |
6de2b1aa NH |
3108 | h->free_huge_pages--; |
3109 | h->free_huge_pages_node[nid]--; | |
3110 | ret = 0; | |
3111 | } | |
93f70f90 | 3112 | spin_unlock(&hugetlb_lock); |
6de2b1aa | 3113 | return ret; |
93f70f90 | 3114 | } |
6de2b1aa | 3115 | #endif |