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> |
9dd540e2 | 31 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 32 | #include <linux/node.h> |
7835e98b | 33 | #include "internal.h" |
1da177e4 LT |
34 | |
35 | const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL; | |
396faf03 MG |
36 | static gfp_t htlb_alloc_mask = GFP_HIGHUSER; |
37 | unsigned long hugepages_treat_as_movable; | |
a5516438 | 38 | |
c3f38a38 | 39 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
40 | unsigned int default_hstate_idx; |
41 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
42 | ||
53ba51d2 JT |
43 | __initdata LIST_HEAD(huge_boot_pages); |
44 | ||
e5ff2159 AK |
45 | /* for command line parsing */ |
46 | static struct hstate * __initdata parsed_hstate; | |
47 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 48 | static unsigned long __initdata default_hstate_size; |
e5ff2159 | 49 | |
3935baa9 DG |
50 | /* |
51 | * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages | |
52 | */ | |
c3f38a38 | 53 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 54 | |
90481622 DG |
55 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
56 | { | |
57 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
58 | ||
59 | spin_unlock(&spool->lock); | |
60 | ||
61 | /* If no pages are used, and no other handles to the subpool | |
62 | * remain, free the subpool the subpool remain */ | |
63 | if (free) | |
64 | kfree(spool); | |
65 | } | |
66 | ||
67 | struct hugepage_subpool *hugepage_new_subpool(long nr_blocks) | |
68 | { | |
69 | struct hugepage_subpool *spool; | |
70 | ||
71 | spool = kmalloc(sizeof(*spool), GFP_KERNEL); | |
72 | if (!spool) | |
73 | return NULL; | |
74 | ||
75 | spin_lock_init(&spool->lock); | |
76 | spool->count = 1; | |
77 | spool->max_hpages = nr_blocks; | |
78 | spool->used_hpages = 0; | |
79 | ||
80 | return spool; | |
81 | } | |
82 | ||
83 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
84 | { | |
85 | spin_lock(&spool->lock); | |
86 | BUG_ON(!spool->count); | |
87 | spool->count--; | |
88 | unlock_or_release_subpool(spool); | |
89 | } | |
90 | ||
91 | static int hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
92 | long delta) | |
93 | { | |
94 | int ret = 0; | |
95 | ||
96 | if (!spool) | |
97 | return 0; | |
98 | ||
99 | spin_lock(&spool->lock); | |
100 | if ((spool->used_hpages + delta) <= spool->max_hpages) { | |
101 | spool->used_hpages += delta; | |
102 | } else { | |
103 | ret = -ENOMEM; | |
104 | } | |
105 | spin_unlock(&spool->lock); | |
106 | ||
107 | return ret; | |
108 | } | |
109 | ||
110 | static void hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
111 | long delta) | |
112 | { | |
113 | if (!spool) | |
114 | return; | |
115 | ||
116 | spin_lock(&spool->lock); | |
117 | spool->used_hpages -= delta; | |
118 | /* If hugetlbfs_put_super couldn't free spool due to | |
119 | * an outstanding quota reference, free it now. */ | |
120 | unlock_or_release_subpool(spool); | |
121 | } | |
122 | ||
123 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
124 | { | |
125 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
126 | } | |
127 | ||
128 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
129 | { | |
130 | return subpool_inode(vma->vm_file->f_dentry->d_inode); | |
131 | } | |
132 | ||
96822904 AW |
133 | /* |
134 | * Region tracking -- allows tracking of reservations and instantiated pages | |
135 | * across the pages in a mapping. | |
84afd99b AW |
136 | * |
137 | * The region data structures are protected by a combination of the mmap_sem | |
138 | * and the hugetlb_instantion_mutex. To access or modify a region the caller | |
139 | * must either hold the mmap_sem for write, or the mmap_sem for read and | |
140 | * the hugetlb_instantiation mutex: | |
141 | * | |
32f84528 | 142 | * down_write(&mm->mmap_sem); |
84afd99b | 143 | * or |
32f84528 CF |
144 | * down_read(&mm->mmap_sem); |
145 | * mutex_lock(&hugetlb_instantiation_mutex); | |
96822904 AW |
146 | */ |
147 | struct file_region { | |
148 | struct list_head link; | |
149 | long from; | |
150 | long to; | |
151 | }; | |
152 | ||
153 | static long region_add(struct list_head *head, long f, long t) | |
154 | { | |
155 | struct file_region *rg, *nrg, *trg; | |
156 | ||
157 | /* Locate the region we are either in or before. */ | |
158 | list_for_each_entry(rg, head, link) | |
159 | if (f <= rg->to) | |
160 | break; | |
161 | ||
162 | /* Round our left edge to the current segment if it encloses us. */ | |
163 | if (f > rg->from) | |
164 | f = rg->from; | |
165 | ||
166 | /* Check for and consume any regions we now overlap with. */ | |
167 | nrg = rg; | |
168 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
169 | if (&rg->link == head) | |
170 | break; | |
171 | if (rg->from > t) | |
172 | break; | |
173 | ||
174 | /* If this area reaches higher then extend our area to | |
175 | * include it completely. If this is not the first area | |
176 | * which we intend to reuse, free it. */ | |
177 | if (rg->to > t) | |
178 | t = rg->to; | |
179 | if (rg != nrg) { | |
180 | list_del(&rg->link); | |
181 | kfree(rg); | |
182 | } | |
183 | } | |
184 | nrg->from = f; | |
185 | nrg->to = t; | |
186 | return 0; | |
187 | } | |
188 | ||
189 | static long region_chg(struct list_head *head, long f, long t) | |
190 | { | |
191 | struct file_region *rg, *nrg; | |
192 | long chg = 0; | |
193 | ||
194 | /* Locate the region we are before or in. */ | |
195 | list_for_each_entry(rg, head, link) | |
196 | if (f <= rg->to) | |
197 | break; | |
198 | ||
199 | /* If we are below the current region then a new region is required. | |
200 | * Subtle, allocate a new region at the position but make it zero | |
201 | * size such that we can guarantee to record the reservation. */ | |
202 | if (&rg->link == head || t < rg->from) { | |
203 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
204 | if (!nrg) | |
205 | return -ENOMEM; | |
206 | nrg->from = f; | |
207 | nrg->to = f; | |
208 | INIT_LIST_HEAD(&nrg->link); | |
209 | list_add(&nrg->link, rg->link.prev); | |
210 | ||
211 | return t - f; | |
212 | } | |
213 | ||
214 | /* Round our left edge to the current segment if it encloses us. */ | |
215 | if (f > rg->from) | |
216 | f = rg->from; | |
217 | chg = t - f; | |
218 | ||
219 | /* Check for and consume any regions we now overlap with. */ | |
220 | list_for_each_entry(rg, rg->link.prev, link) { | |
221 | if (&rg->link == head) | |
222 | break; | |
223 | if (rg->from > t) | |
224 | return chg; | |
225 | ||
25985edc | 226 | /* We overlap with this area, if it extends further than |
96822904 AW |
227 | * us then we must extend ourselves. Account for its |
228 | * existing reservation. */ | |
229 | if (rg->to > t) { | |
230 | chg += rg->to - t; | |
231 | t = rg->to; | |
232 | } | |
233 | chg -= rg->to - rg->from; | |
234 | } | |
235 | return chg; | |
236 | } | |
237 | ||
238 | static long region_truncate(struct list_head *head, long end) | |
239 | { | |
240 | struct file_region *rg, *trg; | |
241 | long chg = 0; | |
242 | ||
243 | /* Locate the region we are either in or before. */ | |
244 | list_for_each_entry(rg, head, link) | |
245 | if (end <= rg->to) | |
246 | break; | |
247 | if (&rg->link == head) | |
248 | return 0; | |
249 | ||
250 | /* If we are in the middle of a region then adjust it. */ | |
251 | if (end > rg->from) { | |
252 | chg = rg->to - end; | |
253 | rg->to = end; | |
254 | rg = list_entry(rg->link.next, typeof(*rg), link); | |
255 | } | |
256 | ||
257 | /* Drop any remaining regions. */ | |
258 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
259 | if (&rg->link == head) | |
260 | break; | |
261 | chg += rg->to - rg->from; | |
262 | list_del(&rg->link); | |
263 | kfree(rg); | |
264 | } | |
265 | return chg; | |
266 | } | |
267 | ||
84afd99b AW |
268 | static long region_count(struct list_head *head, long f, long t) |
269 | { | |
270 | struct file_region *rg; | |
271 | long chg = 0; | |
272 | ||
273 | /* Locate each segment we overlap with, and count that overlap. */ | |
274 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
275 | long seg_from; |
276 | long seg_to; | |
84afd99b AW |
277 | |
278 | if (rg->to <= f) | |
279 | continue; | |
280 | if (rg->from >= t) | |
281 | break; | |
282 | ||
283 | seg_from = max(rg->from, f); | |
284 | seg_to = min(rg->to, t); | |
285 | ||
286 | chg += seg_to - seg_from; | |
287 | } | |
288 | ||
289 | return chg; | |
290 | } | |
291 | ||
e7c4b0bf AW |
292 | /* |
293 | * Convert the address within this vma to the page offset within | |
294 | * the mapping, in pagecache page units; huge pages here. | |
295 | */ | |
a5516438 AK |
296 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
297 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 298 | { |
a5516438 AK |
299 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
300 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
301 | } |
302 | ||
0fe6e20b NH |
303 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
304 | unsigned long address) | |
305 | { | |
306 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
307 | } | |
308 | ||
08fba699 MG |
309 | /* |
310 | * Return the size of the pages allocated when backing a VMA. In the majority | |
311 | * cases this will be same size as used by the page table entries. | |
312 | */ | |
313 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
314 | { | |
315 | struct hstate *hstate; | |
316 | ||
317 | if (!is_vm_hugetlb_page(vma)) | |
318 | return PAGE_SIZE; | |
319 | ||
320 | hstate = hstate_vma(vma); | |
321 | ||
322 | return 1UL << (hstate->order + PAGE_SHIFT); | |
323 | } | |
f340ca0f | 324 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 325 | |
3340289d MG |
326 | /* |
327 | * Return the page size being used by the MMU to back a VMA. In the majority | |
328 | * of cases, the page size used by the kernel matches the MMU size. On | |
329 | * architectures where it differs, an architecture-specific version of this | |
330 | * function is required. | |
331 | */ | |
332 | #ifndef vma_mmu_pagesize | |
333 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) | |
334 | { | |
335 | return vma_kernel_pagesize(vma); | |
336 | } | |
337 | #endif | |
338 | ||
84afd99b AW |
339 | /* |
340 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
341 | * bits of the reservation map pointer, which are always clear due to | |
342 | * alignment. | |
343 | */ | |
344 | #define HPAGE_RESV_OWNER (1UL << 0) | |
345 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 346 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 347 | |
a1e78772 MG |
348 | /* |
349 | * These helpers are used to track how many pages are reserved for | |
350 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
351 | * is guaranteed to have their future faults succeed. | |
352 | * | |
353 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
354 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
355 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
356 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
357 | * |
358 | * The private mapping reservation is represented in a subtly different | |
359 | * manner to a shared mapping. A shared mapping has a region map associated | |
360 | * with the underlying file, this region map represents the backing file | |
361 | * pages which have ever had a reservation assigned which this persists even | |
362 | * after the page is instantiated. A private mapping has a region map | |
363 | * associated with the original mmap which is attached to all VMAs which | |
364 | * reference it, this region map represents those offsets which have consumed | |
365 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 366 | */ |
e7c4b0bf AW |
367 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
368 | { | |
369 | return (unsigned long)vma->vm_private_data; | |
370 | } | |
371 | ||
372 | static void set_vma_private_data(struct vm_area_struct *vma, | |
373 | unsigned long value) | |
374 | { | |
375 | vma->vm_private_data = (void *)value; | |
376 | } | |
377 | ||
84afd99b AW |
378 | struct resv_map { |
379 | struct kref refs; | |
380 | struct list_head regions; | |
381 | }; | |
382 | ||
2a4b3ded | 383 | static struct resv_map *resv_map_alloc(void) |
84afd99b AW |
384 | { |
385 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
386 | if (!resv_map) | |
387 | return NULL; | |
388 | ||
389 | kref_init(&resv_map->refs); | |
390 | INIT_LIST_HEAD(&resv_map->regions); | |
391 | ||
392 | return resv_map; | |
393 | } | |
394 | ||
2a4b3ded | 395 | static void resv_map_release(struct kref *ref) |
84afd99b AW |
396 | { |
397 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
398 | ||
399 | /* Clear out any active regions before we release the map. */ | |
400 | region_truncate(&resv_map->regions, 0); | |
401 | kfree(resv_map); | |
402 | } | |
403 | ||
404 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) | |
a1e78772 MG |
405 | { |
406 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 407 | if (!(vma->vm_flags & VM_MAYSHARE)) |
84afd99b AW |
408 | return (struct resv_map *)(get_vma_private_data(vma) & |
409 | ~HPAGE_RESV_MASK); | |
2a4b3ded | 410 | return NULL; |
a1e78772 MG |
411 | } |
412 | ||
84afd99b | 413 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 MG |
414 | { |
415 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 416 | VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); |
a1e78772 | 417 | |
84afd99b AW |
418 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
419 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
420 | } |
421 | ||
422 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
423 | { | |
04f2cbe3 | 424 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); |
f83a275d | 425 | VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); |
e7c4b0bf AW |
426 | |
427 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
428 | } |
429 | ||
430 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
431 | { | |
432 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
e7c4b0bf AW |
433 | |
434 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
435 | } |
436 | ||
437 | /* Decrement the reserved pages in the hugepage pool by one */ | |
a5516438 AK |
438 | static void decrement_hugepage_resv_vma(struct hstate *h, |
439 | struct vm_area_struct *vma) | |
a1e78772 | 440 | { |
c37f9fb1 AW |
441 | if (vma->vm_flags & VM_NORESERVE) |
442 | return; | |
443 | ||
f83a275d | 444 | if (vma->vm_flags & VM_MAYSHARE) { |
a1e78772 | 445 | /* Shared mappings always use reserves */ |
a5516438 | 446 | h->resv_huge_pages--; |
84afd99b | 447 | } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
a1e78772 MG |
448 | /* |
449 | * Only the process that called mmap() has reserves for | |
450 | * private mappings. | |
451 | */ | |
a5516438 | 452 | h->resv_huge_pages--; |
a1e78772 MG |
453 | } |
454 | } | |
455 | ||
04f2cbe3 | 456 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
457 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
458 | { | |
459 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 460 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
461 | vma->vm_private_data = (void *)0; |
462 | } | |
463 | ||
464 | /* Returns true if the VMA has associated reserve pages */ | |
7f09ca51 | 465 | static int vma_has_reserves(struct vm_area_struct *vma) |
a1e78772 | 466 | { |
f83a275d | 467 | if (vma->vm_flags & VM_MAYSHARE) |
7f09ca51 MG |
468 | return 1; |
469 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) | |
470 | return 1; | |
471 | return 0; | |
a1e78772 MG |
472 | } |
473 | ||
0ebabb41 NH |
474 | static void copy_gigantic_page(struct page *dst, struct page *src) |
475 | { | |
476 | int i; | |
477 | struct hstate *h = page_hstate(src); | |
478 | struct page *dst_base = dst; | |
479 | struct page *src_base = src; | |
480 | ||
481 | for (i = 0; i < pages_per_huge_page(h); ) { | |
482 | cond_resched(); | |
483 | copy_highpage(dst, src); | |
484 | ||
485 | i++; | |
486 | dst = mem_map_next(dst, dst_base, i); | |
487 | src = mem_map_next(src, src_base, i); | |
488 | } | |
489 | } | |
490 | ||
491 | void copy_huge_page(struct page *dst, struct page *src) | |
492 | { | |
493 | int i; | |
494 | struct hstate *h = page_hstate(src); | |
495 | ||
496 | if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) { | |
497 | copy_gigantic_page(dst, src); | |
498 | return; | |
499 | } | |
500 | ||
501 | might_sleep(); | |
502 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
503 | cond_resched(); | |
504 | copy_highpage(dst + i, src + i); | |
505 | } | |
506 | } | |
507 | ||
a5516438 | 508 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
509 | { |
510 | int nid = page_to_nid(page); | |
0edaecfa | 511 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
512 | h->free_huge_pages++; |
513 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
514 | } |
515 | ||
bf50bab2 NH |
516 | static struct page *dequeue_huge_page_node(struct hstate *h, int nid) |
517 | { | |
518 | struct page *page; | |
519 | ||
520 | if (list_empty(&h->hugepage_freelists[nid])) | |
521 | return NULL; | |
522 | page = list_entry(h->hugepage_freelists[nid].next, struct page, lru); | |
0edaecfa | 523 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 524 | set_page_refcounted(page); |
bf50bab2 NH |
525 | h->free_huge_pages--; |
526 | h->free_huge_pages_node[nid]--; | |
527 | return page; | |
528 | } | |
529 | ||
a5516438 AK |
530 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
531 | struct vm_area_struct *vma, | |
04f2cbe3 | 532 | unsigned long address, int avoid_reserve) |
1da177e4 | 533 | { |
b1c12cbc | 534 | struct page *page = NULL; |
480eccf9 | 535 | struct mempolicy *mpol; |
19770b32 | 536 | nodemask_t *nodemask; |
c0ff7453 | 537 | struct zonelist *zonelist; |
dd1a239f MG |
538 | struct zone *zone; |
539 | struct zoneref *z; | |
cc9a6c87 | 540 | unsigned int cpuset_mems_cookie; |
1da177e4 | 541 | |
cc9a6c87 MG |
542 | retry_cpuset: |
543 | cpuset_mems_cookie = get_mems_allowed(); | |
c0ff7453 MX |
544 | zonelist = huge_zonelist(vma, address, |
545 | htlb_alloc_mask, &mpol, &nodemask); | |
a1e78772 MG |
546 | /* |
547 | * A child process with MAP_PRIVATE mappings created by their parent | |
548 | * have no page reserves. This check ensures that reservations are | |
549 | * not "stolen". The child may still get SIGKILLed | |
550 | */ | |
7f09ca51 | 551 | if (!vma_has_reserves(vma) && |
a5516438 | 552 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 553 | goto err; |
a1e78772 | 554 | |
04f2cbe3 | 555 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 556 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 557 | goto err; |
04f2cbe3 | 558 | |
19770b32 MG |
559 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
560 | MAX_NR_ZONES - 1, nodemask) { | |
bf50bab2 NH |
561 | if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) { |
562 | page = dequeue_huge_page_node(h, zone_to_nid(zone)); | |
563 | if (page) { | |
564 | if (!avoid_reserve) | |
565 | decrement_hugepage_resv_vma(h, vma); | |
566 | break; | |
567 | } | |
3abf7afd | 568 | } |
1da177e4 | 569 | } |
cc9a6c87 | 570 | |
52cd3b07 | 571 | mpol_cond_put(mpol); |
cc9a6c87 MG |
572 | if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page)) |
573 | goto retry_cpuset; | |
1da177e4 | 574 | return page; |
cc9a6c87 MG |
575 | |
576 | err: | |
577 | mpol_cond_put(mpol); | |
578 | return NULL; | |
1da177e4 LT |
579 | } |
580 | ||
a5516438 | 581 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
582 | { |
583 | int i; | |
a5516438 | 584 | |
18229df5 AW |
585 | VM_BUG_ON(h->order >= MAX_ORDER); |
586 | ||
a5516438 AK |
587 | h->nr_huge_pages--; |
588 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
589 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
590 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
591 | 1 << PG_referenced | 1 << PG_dirty | | |
592 | 1 << PG_active | 1 << PG_reserved | | |
593 | 1 << PG_private | 1 << PG_writeback); | |
6af2acb6 | 594 | } |
9dd540e2 | 595 | VM_BUG_ON(hugetlb_cgroup_from_page(page)); |
6af2acb6 AL |
596 | set_compound_page_dtor(page, NULL); |
597 | set_page_refcounted(page); | |
7f2e9525 | 598 | arch_release_hugepage(page); |
a5516438 | 599 | __free_pages(page, huge_page_order(h)); |
6af2acb6 AL |
600 | } |
601 | ||
e5ff2159 AK |
602 | struct hstate *size_to_hstate(unsigned long size) |
603 | { | |
604 | struct hstate *h; | |
605 | ||
606 | for_each_hstate(h) { | |
607 | if (huge_page_size(h) == size) | |
608 | return h; | |
609 | } | |
610 | return NULL; | |
611 | } | |
612 | ||
27a85ef1 DG |
613 | static void free_huge_page(struct page *page) |
614 | { | |
a5516438 AK |
615 | /* |
616 | * Can't pass hstate in here because it is called from the | |
617 | * compound page destructor. | |
618 | */ | |
e5ff2159 | 619 | struct hstate *h = page_hstate(page); |
7893d1d5 | 620 | int nid = page_to_nid(page); |
90481622 DG |
621 | struct hugepage_subpool *spool = |
622 | (struct hugepage_subpool *)page_private(page); | |
27a85ef1 | 623 | |
e5df70ab | 624 | set_page_private(page, 0); |
23be7468 | 625 | page->mapping = NULL; |
7893d1d5 | 626 | BUG_ON(page_count(page)); |
0fe6e20b | 627 | BUG_ON(page_mapcount(page)); |
27a85ef1 DG |
628 | |
629 | spin_lock(&hugetlb_lock); | |
6d76dcf4 AK |
630 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
631 | pages_per_huge_page(h), page); | |
aa888a74 | 632 | if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { |
0edaecfa AK |
633 | /* remove the page from active list */ |
634 | list_del(&page->lru); | |
a5516438 AK |
635 | update_and_free_page(h, page); |
636 | h->surplus_huge_pages--; | |
637 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 638 | } else { |
5d3a551c | 639 | arch_clear_hugepage_flags(page); |
a5516438 | 640 | enqueue_huge_page(h, page); |
7893d1d5 | 641 | } |
27a85ef1 | 642 | spin_unlock(&hugetlb_lock); |
90481622 | 643 | hugepage_subpool_put_pages(spool, 1); |
27a85ef1 DG |
644 | } |
645 | ||
a5516438 | 646 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 647 | { |
0edaecfa | 648 | INIT_LIST_HEAD(&page->lru); |
b7ba30c6 AK |
649 | set_compound_page_dtor(page, free_huge_page); |
650 | spin_lock(&hugetlb_lock); | |
9dd540e2 | 651 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
652 | h->nr_huge_pages++; |
653 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 AK |
654 | spin_unlock(&hugetlb_lock); |
655 | put_page(page); /* free it into the hugepage allocator */ | |
656 | } | |
657 | ||
20a0307c WF |
658 | static void prep_compound_gigantic_page(struct page *page, unsigned long order) |
659 | { | |
660 | int i; | |
661 | int nr_pages = 1 << order; | |
662 | struct page *p = page + 1; | |
663 | ||
664 | /* we rely on prep_new_huge_page to set the destructor */ | |
665 | set_compound_order(page, order); | |
666 | __SetPageHead(page); | |
667 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { | |
668 | __SetPageTail(p); | |
58a84aa9 | 669 | set_page_count(p, 0); |
20a0307c WF |
670 | p->first_page = page; |
671 | } | |
672 | } | |
673 | ||
674 | int PageHuge(struct page *page) | |
675 | { | |
676 | compound_page_dtor *dtor; | |
677 | ||
678 | if (!PageCompound(page)) | |
679 | return 0; | |
680 | ||
681 | page = compound_head(page); | |
682 | dtor = get_compound_page_dtor(page); | |
683 | ||
684 | return dtor == free_huge_page; | |
685 | } | |
43131e14 NH |
686 | EXPORT_SYMBOL_GPL(PageHuge); |
687 | ||
a5516438 | 688 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
1da177e4 | 689 | { |
1da177e4 | 690 | struct page *page; |
f96efd58 | 691 | |
aa888a74 AK |
692 | if (h->order >= MAX_ORDER) |
693 | return NULL; | |
694 | ||
6484eb3e | 695 | page = alloc_pages_exact_node(nid, |
551883ae NA |
696 | htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| |
697 | __GFP_REPEAT|__GFP_NOWARN, | |
a5516438 | 698 | huge_page_order(h)); |
1da177e4 | 699 | if (page) { |
7f2e9525 | 700 | if (arch_prepare_hugepage(page)) { |
caff3a2c | 701 | __free_pages(page, huge_page_order(h)); |
7b8ee84d | 702 | return NULL; |
7f2e9525 | 703 | } |
a5516438 | 704 | prep_new_huge_page(h, page, nid); |
1da177e4 | 705 | } |
63b4613c NA |
706 | |
707 | return page; | |
708 | } | |
709 | ||
9a76db09 | 710 | /* |
6ae11b27 LS |
711 | * common helper functions for hstate_next_node_to_{alloc|free}. |
712 | * We may have allocated or freed a huge page based on a different | |
713 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
714 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
715 | * node for alloc or free. | |
9a76db09 | 716 | */ |
6ae11b27 | 717 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) |
9a76db09 | 718 | { |
6ae11b27 | 719 | nid = next_node(nid, *nodes_allowed); |
9a76db09 | 720 | if (nid == MAX_NUMNODES) |
6ae11b27 | 721 | nid = first_node(*nodes_allowed); |
9a76db09 LS |
722 | VM_BUG_ON(nid >= MAX_NUMNODES); |
723 | ||
724 | return nid; | |
725 | } | |
726 | ||
6ae11b27 LS |
727 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) |
728 | { | |
729 | if (!node_isset(nid, *nodes_allowed)) | |
730 | nid = next_node_allowed(nid, nodes_allowed); | |
731 | return nid; | |
732 | } | |
733 | ||
5ced66c9 | 734 | /* |
6ae11b27 LS |
735 | * returns the previously saved node ["this node"] from which to |
736 | * allocate a persistent huge page for the pool and advance the | |
737 | * next node from which to allocate, handling wrap at end of node | |
738 | * mask. | |
5ced66c9 | 739 | */ |
6ae11b27 LS |
740 | static int hstate_next_node_to_alloc(struct hstate *h, |
741 | nodemask_t *nodes_allowed) | |
5ced66c9 | 742 | { |
6ae11b27 LS |
743 | int nid; |
744 | ||
745 | VM_BUG_ON(!nodes_allowed); | |
746 | ||
747 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
748 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
9a76db09 | 749 | |
9a76db09 | 750 | return nid; |
5ced66c9 AK |
751 | } |
752 | ||
6ae11b27 | 753 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
63b4613c NA |
754 | { |
755 | struct page *page; | |
756 | int start_nid; | |
757 | int next_nid; | |
758 | int ret = 0; | |
759 | ||
6ae11b27 | 760 | start_nid = hstate_next_node_to_alloc(h, nodes_allowed); |
e8c5c824 | 761 | next_nid = start_nid; |
63b4613c NA |
762 | |
763 | do { | |
e8c5c824 | 764 | page = alloc_fresh_huge_page_node(h, next_nid); |
9a76db09 | 765 | if (page) { |
63b4613c | 766 | ret = 1; |
9a76db09 LS |
767 | break; |
768 | } | |
6ae11b27 | 769 | next_nid = hstate_next_node_to_alloc(h, nodes_allowed); |
9a76db09 | 770 | } while (next_nid != start_nid); |
63b4613c | 771 | |
3b116300 AL |
772 | if (ret) |
773 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
774 | else | |
775 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
776 | ||
63b4613c | 777 | return ret; |
1da177e4 LT |
778 | } |
779 | ||
e8c5c824 | 780 | /* |
6ae11b27 LS |
781 | * helper for free_pool_huge_page() - return the previously saved |
782 | * node ["this node"] from which to free a huge page. Advance the | |
783 | * next node id whether or not we find a free huge page to free so | |
784 | * that the next attempt to free addresses the next node. | |
e8c5c824 | 785 | */ |
6ae11b27 | 786 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) |
e8c5c824 | 787 | { |
6ae11b27 LS |
788 | int nid; |
789 | ||
790 | VM_BUG_ON(!nodes_allowed); | |
791 | ||
792 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
793 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
9a76db09 | 794 | |
9a76db09 | 795 | return nid; |
e8c5c824 LS |
796 | } |
797 | ||
798 | /* | |
799 | * Free huge page from pool from next node to free. | |
800 | * Attempt to keep persistent huge pages more or less | |
801 | * balanced over allowed nodes. | |
802 | * Called with hugetlb_lock locked. | |
803 | */ | |
6ae11b27 LS |
804 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
805 | bool acct_surplus) | |
e8c5c824 LS |
806 | { |
807 | int start_nid; | |
808 | int next_nid; | |
809 | int ret = 0; | |
810 | ||
6ae11b27 | 811 | start_nid = hstate_next_node_to_free(h, nodes_allowed); |
e8c5c824 LS |
812 | next_nid = start_nid; |
813 | ||
814 | do { | |
685f3457 LS |
815 | /* |
816 | * If we're returning unused surplus pages, only examine | |
817 | * nodes with surplus pages. | |
818 | */ | |
819 | if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) && | |
820 | !list_empty(&h->hugepage_freelists[next_nid])) { | |
e8c5c824 LS |
821 | struct page *page = |
822 | list_entry(h->hugepage_freelists[next_nid].next, | |
823 | struct page, lru); | |
824 | list_del(&page->lru); | |
825 | h->free_huge_pages--; | |
826 | h->free_huge_pages_node[next_nid]--; | |
685f3457 LS |
827 | if (acct_surplus) { |
828 | h->surplus_huge_pages--; | |
829 | h->surplus_huge_pages_node[next_nid]--; | |
830 | } | |
e8c5c824 LS |
831 | update_and_free_page(h, page); |
832 | ret = 1; | |
9a76db09 | 833 | break; |
e8c5c824 | 834 | } |
6ae11b27 | 835 | next_nid = hstate_next_node_to_free(h, nodes_allowed); |
9a76db09 | 836 | } while (next_nid != start_nid); |
e8c5c824 LS |
837 | |
838 | return ret; | |
839 | } | |
840 | ||
bf50bab2 | 841 | static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) |
7893d1d5 AL |
842 | { |
843 | struct page *page; | |
bf50bab2 | 844 | unsigned int r_nid; |
7893d1d5 | 845 | |
aa888a74 AK |
846 | if (h->order >= MAX_ORDER) |
847 | return NULL; | |
848 | ||
d1c3fb1f NA |
849 | /* |
850 | * Assume we will successfully allocate the surplus page to | |
851 | * prevent racing processes from causing the surplus to exceed | |
852 | * overcommit | |
853 | * | |
854 | * This however introduces a different race, where a process B | |
855 | * tries to grow the static hugepage pool while alloc_pages() is | |
856 | * called by process A. B will only examine the per-node | |
857 | * counters in determining if surplus huge pages can be | |
858 | * converted to normal huge pages in adjust_pool_surplus(). A | |
859 | * won't be able to increment the per-node counter, until the | |
860 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
861 | * no more huge pages can be converted from surplus to normal | |
862 | * state (and doesn't try to convert again). Thus, we have a | |
863 | * case where a surplus huge page exists, the pool is grown, and | |
864 | * the surplus huge page still exists after, even though it | |
865 | * should just have been converted to a normal huge page. This | |
866 | * does not leak memory, though, as the hugepage will be freed | |
867 | * once it is out of use. It also does not allow the counters to | |
868 | * go out of whack in adjust_pool_surplus() as we don't modify | |
869 | * the node values until we've gotten the hugepage and only the | |
870 | * per-node value is checked there. | |
871 | */ | |
872 | spin_lock(&hugetlb_lock); | |
a5516438 | 873 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
874 | spin_unlock(&hugetlb_lock); |
875 | return NULL; | |
876 | } else { | |
a5516438 AK |
877 | h->nr_huge_pages++; |
878 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
879 | } |
880 | spin_unlock(&hugetlb_lock); | |
881 | ||
bf50bab2 NH |
882 | if (nid == NUMA_NO_NODE) |
883 | page = alloc_pages(htlb_alloc_mask|__GFP_COMP| | |
884 | __GFP_REPEAT|__GFP_NOWARN, | |
885 | huge_page_order(h)); | |
886 | else | |
887 | page = alloc_pages_exact_node(nid, | |
888 | htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| | |
889 | __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); | |
d1c3fb1f | 890 | |
caff3a2c GS |
891 | if (page && arch_prepare_hugepage(page)) { |
892 | __free_pages(page, huge_page_order(h)); | |
ea5768c7 | 893 | page = NULL; |
caff3a2c GS |
894 | } |
895 | ||
d1c3fb1f | 896 | spin_lock(&hugetlb_lock); |
7893d1d5 | 897 | if (page) { |
0edaecfa | 898 | INIT_LIST_HEAD(&page->lru); |
bf50bab2 | 899 | r_nid = page_to_nid(page); |
7893d1d5 | 900 | set_compound_page_dtor(page, free_huge_page); |
9dd540e2 | 901 | set_hugetlb_cgroup(page, NULL); |
d1c3fb1f NA |
902 | /* |
903 | * We incremented the global counters already | |
904 | */ | |
bf50bab2 NH |
905 | h->nr_huge_pages_node[r_nid]++; |
906 | h->surplus_huge_pages_node[r_nid]++; | |
3b116300 | 907 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 908 | } else { |
a5516438 AK |
909 | h->nr_huge_pages--; |
910 | h->surplus_huge_pages--; | |
3b116300 | 911 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 912 | } |
d1c3fb1f | 913 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
914 | |
915 | return page; | |
916 | } | |
917 | ||
bf50bab2 NH |
918 | /* |
919 | * This allocation function is useful in the context where vma is irrelevant. | |
920 | * E.g. soft-offlining uses this function because it only cares physical | |
921 | * address of error page. | |
922 | */ | |
923 | struct page *alloc_huge_page_node(struct hstate *h, int nid) | |
924 | { | |
925 | struct page *page; | |
926 | ||
927 | spin_lock(&hugetlb_lock); | |
928 | page = dequeue_huge_page_node(h, nid); | |
929 | spin_unlock(&hugetlb_lock); | |
930 | ||
94ae8ba7 | 931 | if (!page) |
bf50bab2 NH |
932 | page = alloc_buddy_huge_page(h, nid); |
933 | ||
934 | return page; | |
935 | } | |
936 | ||
e4e574b7 | 937 | /* |
25985edc | 938 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
939 | * of size 'delta'. |
940 | */ | |
a5516438 | 941 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
942 | { |
943 | struct list_head surplus_list; | |
944 | struct page *page, *tmp; | |
945 | int ret, i; | |
946 | int needed, allocated; | |
28073b02 | 947 | bool alloc_ok = true; |
e4e574b7 | 948 | |
a5516438 | 949 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 950 | if (needed <= 0) { |
a5516438 | 951 | h->resv_huge_pages += delta; |
e4e574b7 | 952 | return 0; |
ac09b3a1 | 953 | } |
e4e574b7 AL |
954 | |
955 | allocated = 0; | |
956 | INIT_LIST_HEAD(&surplus_list); | |
957 | ||
958 | ret = -ENOMEM; | |
959 | retry: | |
960 | spin_unlock(&hugetlb_lock); | |
961 | for (i = 0; i < needed; i++) { | |
bf50bab2 | 962 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
28073b02 HD |
963 | if (!page) { |
964 | alloc_ok = false; | |
965 | break; | |
966 | } | |
e4e574b7 AL |
967 | list_add(&page->lru, &surplus_list); |
968 | } | |
28073b02 | 969 | allocated += i; |
e4e574b7 AL |
970 | |
971 | /* | |
972 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
973 | * because either resv_huge_pages or free_huge_pages may have changed. | |
974 | */ | |
975 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
976 | needed = (h->resv_huge_pages + delta) - |
977 | (h->free_huge_pages + allocated); | |
28073b02 HD |
978 | if (needed > 0) { |
979 | if (alloc_ok) | |
980 | goto retry; | |
981 | /* | |
982 | * We were not able to allocate enough pages to | |
983 | * satisfy the entire reservation so we free what | |
984 | * we've allocated so far. | |
985 | */ | |
986 | goto free; | |
987 | } | |
e4e574b7 AL |
988 | /* |
989 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 990 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 991 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
992 | * allocator. Commit the entire reservation here to prevent another |
993 | * process from stealing the pages as they are added to the pool but | |
994 | * before they are reserved. | |
e4e574b7 AL |
995 | */ |
996 | needed += allocated; | |
a5516438 | 997 | h->resv_huge_pages += delta; |
e4e574b7 | 998 | ret = 0; |
a9869b83 | 999 | |
19fc3f0a | 1000 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1001 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1002 | if ((--needed) < 0) |
1003 | break; | |
a9869b83 NH |
1004 | /* |
1005 | * This page is now managed by the hugetlb allocator and has | |
1006 | * no users -- drop the buddy allocator's reference. | |
1007 | */ | |
1008 | put_page_testzero(page); | |
1009 | VM_BUG_ON(page_count(page)); | |
a5516438 | 1010 | enqueue_huge_page(h, page); |
19fc3f0a | 1011 | } |
28073b02 | 1012 | free: |
b0365c8d | 1013 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1014 | |
1015 | /* Free unnecessary surplus pages to the buddy allocator */ | |
1016 | if (!list_empty(&surplus_list)) { | |
19fc3f0a | 1017 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
a9869b83 | 1018 | put_page(page); |
af767cbd | 1019 | } |
e4e574b7 | 1020 | } |
a9869b83 | 1021 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1022 | |
1023 | return ret; | |
1024 | } | |
1025 | ||
1026 | /* | |
1027 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
1028 | * allocated to satisfy the reservation must be explicitly freed if they were | |
1029 | * never used. | |
685f3457 | 1030 | * Called with hugetlb_lock held. |
e4e574b7 | 1031 | */ |
a5516438 AK |
1032 | static void return_unused_surplus_pages(struct hstate *h, |
1033 | unsigned long unused_resv_pages) | |
e4e574b7 | 1034 | { |
e4e574b7 AL |
1035 | unsigned long nr_pages; |
1036 | ||
ac09b3a1 | 1037 | /* Uncommit the reservation */ |
a5516438 | 1038 | h->resv_huge_pages -= unused_resv_pages; |
ac09b3a1 | 1039 | |
aa888a74 AK |
1040 | /* Cannot return gigantic pages currently */ |
1041 | if (h->order >= MAX_ORDER) | |
1042 | return; | |
1043 | ||
a5516438 | 1044 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1045 | |
685f3457 LS |
1046 | /* |
1047 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1048 | * evenly across all nodes with memory. Iterate across these nodes |
1049 | * until we can no longer free unreserved surplus pages. This occurs | |
1050 | * when the nodes with surplus pages have no free pages. | |
1051 | * free_pool_huge_page() will balance the the freed pages across the | |
1052 | * on-line nodes with memory and will handle the hstate accounting. | |
685f3457 LS |
1053 | */ |
1054 | while (nr_pages--) { | |
9b5e5d0f | 1055 | if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1)) |
685f3457 | 1056 | break; |
e4e574b7 AL |
1057 | } |
1058 | } | |
1059 | ||
c37f9fb1 AW |
1060 | /* |
1061 | * Determine if the huge page at addr within the vma has an associated | |
1062 | * reservation. Where it does not we will need to logically increase | |
90481622 DG |
1063 | * reservation and actually increase subpool usage before an allocation |
1064 | * can occur. Where any new reservation would be required the | |
1065 | * reservation change is prepared, but not committed. Once the page | |
1066 | * has been allocated from the subpool and instantiated the change should | |
1067 | * be committed via vma_commit_reservation. No action is required on | |
1068 | * failure. | |
c37f9fb1 | 1069 | */ |
e2f17d94 | 1070 | static long vma_needs_reservation(struct hstate *h, |
a5516438 | 1071 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 AW |
1072 | { |
1073 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1074 | struct inode *inode = mapping->host; | |
1075 | ||
f83a275d | 1076 | if (vma->vm_flags & VM_MAYSHARE) { |
a5516438 | 1077 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
c37f9fb1 AW |
1078 | return region_chg(&inode->i_mapping->private_list, |
1079 | idx, idx + 1); | |
1080 | ||
84afd99b AW |
1081 | } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1082 | return 1; | |
c37f9fb1 | 1083 | |
84afd99b | 1084 | } else { |
e2f17d94 | 1085 | long err; |
a5516438 | 1086 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
84afd99b AW |
1087 | struct resv_map *reservations = vma_resv_map(vma); |
1088 | ||
1089 | err = region_chg(&reservations->regions, idx, idx + 1); | |
1090 | if (err < 0) | |
1091 | return err; | |
1092 | return 0; | |
1093 | } | |
c37f9fb1 | 1094 | } |
a5516438 AK |
1095 | static void vma_commit_reservation(struct hstate *h, |
1096 | struct vm_area_struct *vma, unsigned long addr) | |
c37f9fb1 AW |
1097 | { |
1098 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1099 | struct inode *inode = mapping->host; | |
1100 | ||
f83a275d | 1101 | if (vma->vm_flags & VM_MAYSHARE) { |
a5516438 | 1102 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
c37f9fb1 | 1103 | region_add(&inode->i_mapping->private_list, idx, idx + 1); |
84afd99b AW |
1104 | |
1105 | } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { | |
a5516438 | 1106 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
84afd99b AW |
1107 | struct resv_map *reservations = vma_resv_map(vma); |
1108 | ||
1109 | /* Mark this page used in the map. */ | |
1110 | region_add(&reservations->regions, idx, idx + 1); | |
c37f9fb1 AW |
1111 | } |
1112 | } | |
1113 | ||
a1e78772 | 1114 | static struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 1115 | unsigned long addr, int avoid_reserve) |
1da177e4 | 1116 | { |
90481622 | 1117 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 1118 | struct hstate *h = hstate_vma(vma); |
348ea204 | 1119 | struct page *page; |
e2f17d94 | 1120 | long chg; |
6d76dcf4 AK |
1121 | int ret, idx; |
1122 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 1123 | |
6d76dcf4 | 1124 | idx = hstate_index(h); |
a1e78772 | 1125 | /* |
90481622 DG |
1126 | * Processes that did not create the mapping will have no |
1127 | * reserves and will not have accounted against subpool | |
1128 | * limit. Check that the subpool limit can be made before | |
1129 | * satisfying the allocation MAP_NORESERVE mappings may also | |
1130 | * need pages and subpool limit allocated allocated if no reserve | |
1131 | * mapping overlaps. | |
a1e78772 | 1132 | */ |
a5516438 | 1133 | chg = vma_needs_reservation(h, vma, addr); |
c37f9fb1 | 1134 | if (chg < 0) |
76dcee75 | 1135 | return ERR_PTR(-ENOMEM); |
c37f9fb1 | 1136 | if (chg) |
90481622 | 1137 | if (hugepage_subpool_get_pages(spool, chg)) |
76dcee75 | 1138 | return ERR_PTR(-ENOSPC); |
1da177e4 | 1139 | |
6d76dcf4 AK |
1140 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
1141 | if (ret) { | |
1142 | hugepage_subpool_put_pages(spool, chg); | |
1143 | return ERR_PTR(-ENOSPC); | |
1144 | } | |
1da177e4 | 1145 | spin_lock(&hugetlb_lock); |
a5516438 | 1146 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve); |
94ae8ba7 AK |
1147 | if (page) { |
1148 | /* update page cgroup details */ | |
1149 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), | |
1150 | h_cg, page); | |
1151 | spin_unlock(&hugetlb_lock); | |
1152 | } else { | |
1153 | spin_unlock(&hugetlb_lock); | |
bf50bab2 | 1154 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
68842c9b | 1155 | if (!page) { |
6d76dcf4 AK |
1156 | hugetlb_cgroup_uncharge_cgroup(idx, |
1157 | pages_per_huge_page(h), | |
1158 | h_cg); | |
90481622 | 1159 | hugepage_subpool_put_pages(spool, chg); |
76dcee75 | 1160 | return ERR_PTR(-ENOSPC); |
68842c9b | 1161 | } |
79dbb236 | 1162 | spin_lock(&hugetlb_lock); |
94ae8ba7 AK |
1163 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), |
1164 | h_cg, page); | |
79dbb236 AK |
1165 | list_move(&page->lru, &h->hugepage_activelist); |
1166 | spin_unlock(&hugetlb_lock); | |
68842c9b | 1167 | } |
348ea204 | 1168 | |
90481622 | 1169 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 1170 | |
a5516438 | 1171 | vma_commit_reservation(h, vma, addr); |
90d8b7e6 | 1172 | return page; |
b45b5bd6 DG |
1173 | } |
1174 | ||
91f47662 | 1175 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
aa888a74 AK |
1176 | { |
1177 | struct huge_bootmem_page *m; | |
9b5e5d0f | 1178 | int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); |
aa888a74 AK |
1179 | |
1180 | while (nr_nodes) { | |
1181 | void *addr; | |
1182 | ||
1183 | addr = __alloc_bootmem_node_nopanic( | |
6ae11b27 | 1184 | NODE_DATA(hstate_next_node_to_alloc(h, |
9b5e5d0f | 1185 | &node_states[N_HIGH_MEMORY])), |
aa888a74 AK |
1186 | huge_page_size(h), huge_page_size(h), 0); |
1187 | ||
1188 | if (addr) { | |
1189 | /* | |
1190 | * Use the beginning of the huge page to store the | |
1191 | * huge_bootmem_page struct (until gather_bootmem | |
1192 | * puts them into the mem_map). | |
1193 | */ | |
1194 | m = addr; | |
91f47662 | 1195 | goto found; |
aa888a74 | 1196 | } |
aa888a74 AK |
1197 | nr_nodes--; |
1198 | } | |
1199 | return 0; | |
1200 | ||
1201 | found: | |
1202 | BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); | |
1203 | /* Put them into a private list first because mem_map is not up yet */ | |
1204 | list_add(&m->list, &huge_boot_pages); | |
1205 | m->hstate = h; | |
1206 | return 1; | |
1207 | } | |
1208 | ||
18229df5 AW |
1209 | static void prep_compound_huge_page(struct page *page, int order) |
1210 | { | |
1211 | if (unlikely(order > (MAX_ORDER - 1))) | |
1212 | prep_compound_gigantic_page(page, order); | |
1213 | else | |
1214 | prep_compound_page(page, order); | |
1215 | } | |
1216 | ||
aa888a74 AK |
1217 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
1218 | static void __init gather_bootmem_prealloc(void) | |
1219 | { | |
1220 | struct huge_bootmem_page *m; | |
1221 | ||
1222 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 1223 | struct hstate *h = m->hstate; |
ee8f248d BB |
1224 | struct page *page; |
1225 | ||
1226 | #ifdef CONFIG_HIGHMEM | |
1227 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
1228 | free_bootmem_late((unsigned long)m, | |
1229 | sizeof(struct huge_bootmem_page)); | |
1230 | #else | |
1231 | page = virt_to_page(m); | |
1232 | #endif | |
aa888a74 AK |
1233 | __ClearPageReserved(page); |
1234 | WARN_ON(page_count(page) != 1); | |
18229df5 | 1235 | prep_compound_huge_page(page, h->order); |
aa888a74 | 1236 | prep_new_huge_page(h, page, page_to_nid(page)); |
b0320c7b RA |
1237 | /* |
1238 | * If we had gigantic hugepages allocated at boot time, we need | |
1239 | * to restore the 'stolen' pages to totalram_pages in order to | |
1240 | * fix confusing memory reports from free(1) and another | |
1241 | * side-effects, like CommitLimit going negative. | |
1242 | */ | |
1243 | if (h->order > (MAX_ORDER - 1)) | |
1244 | totalram_pages += 1 << h->order; | |
aa888a74 AK |
1245 | } |
1246 | } | |
1247 | ||
8faa8b07 | 1248 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
1249 | { |
1250 | unsigned long i; | |
a5516438 | 1251 | |
e5ff2159 | 1252 | for (i = 0; i < h->max_huge_pages; ++i) { |
aa888a74 AK |
1253 | if (h->order >= MAX_ORDER) { |
1254 | if (!alloc_bootmem_huge_page(h)) | |
1255 | break; | |
9b5e5d0f LS |
1256 | } else if (!alloc_fresh_huge_page(h, |
1257 | &node_states[N_HIGH_MEMORY])) | |
1da177e4 | 1258 | break; |
1da177e4 | 1259 | } |
8faa8b07 | 1260 | h->max_huge_pages = i; |
e5ff2159 AK |
1261 | } |
1262 | ||
1263 | static void __init hugetlb_init_hstates(void) | |
1264 | { | |
1265 | struct hstate *h; | |
1266 | ||
1267 | for_each_hstate(h) { | |
8faa8b07 AK |
1268 | /* oversize hugepages were init'ed in early boot */ |
1269 | if (h->order < MAX_ORDER) | |
1270 | hugetlb_hstate_alloc_pages(h); | |
e5ff2159 AK |
1271 | } |
1272 | } | |
1273 | ||
4abd32db AK |
1274 | static char * __init memfmt(char *buf, unsigned long n) |
1275 | { | |
1276 | if (n >= (1UL << 30)) | |
1277 | sprintf(buf, "%lu GB", n >> 30); | |
1278 | else if (n >= (1UL << 20)) | |
1279 | sprintf(buf, "%lu MB", n >> 20); | |
1280 | else | |
1281 | sprintf(buf, "%lu KB", n >> 10); | |
1282 | return buf; | |
1283 | } | |
1284 | ||
e5ff2159 AK |
1285 | static void __init report_hugepages(void) |
1286 | { | |
1287 | struct hstate *h; | |
1288 | ||
1289 | for_each_hstate(h) { | |
4abd32db AK |
1290 | char buf[32]; |
1291 | printk(KERN_INFO "HugeTLB registered %s page size, " | |
1292 | "pre-allocated %ld pages\n", | |
1293 | memfmt(buf, huge_page_size(h)), | |
1294 | h->free_huge_pages); | |
e5ff2159 AK |
1295 | } |
1296 | } | |
1297 | ||
1da177e4 | 1298 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
1299 | static void try_to_free_low(struct hstate *h, unsigned long count, |
1300 | nodemask_t *nodes_allowed) | |
1da177e4 | 1301 | { |
4415cc8d CL |
1302 | int i; |
1303 | ||
aa888a74 AK |
1304 | if (h->order >= MAX_ORDER) |
1305 | return; | |
1306 | ||
6ae11b27 | 1307 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 1308 | struct page *page, *next; |
a5516438 AK |
1309 | struct list_head *freel = &h->hugepage_freelists[i]; |
1310 | list_for_each_entry_safe(page, next, freel, lru) { | |
1311 | if (count >= h->nr_huge_pages) | |
6b0c880d | 1312 | return; |
1da177e4 LT |
1313 | if (PageHighMem(page)) |
1314 | continue; | |
1315 | list_del(&page->lru); | |
e5ff2159 | 1316 | update_and_free_page(h, page); |
a5516438 AK |
1317 | h->free_huge_pages--; |
1318 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
1319 | } |
1320 | } | |
1321 | } | |
1322 | #else | |
6ae11b27 LS |
1323 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
1324 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
1325 | { |
1326 | } | |
1327 | #endif | |
1328 | ||
20a0307c WF |
1329 | /* |
1330 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
1331 | * balanced by operating on them in a round-robin fashion. | |
1332 | * Returns 1 if an adjustment was made. | |
1333 | */ | |
6ae11b27 LS |
1334 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
1335 | int delta) | |
20a0307c | 1336 | { |
e8c5c824 | 1337 | int start_nid, next_nid; |
20a0307c WF |
1338 | int ret = 0; |
1339 | ||
1340 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 1341 | |
e8c5c824 | 1342 | if (delta < 0) |
6ae11b27 | 1343 | start_nid = hstate_next_node_to_alloc(h, nodes_allowed); |
e8c5c824 | 1344 | else |
6ae11b27 | 1345 | start_nid = hstate_next_node_to_free(h, nodes_allowed); |
e8c5c824 LS |
1346 | next_nid = start_nid; |
1347 | ||
1348 | do { | |
1349 | int nid = next_nid; | |
1350 | if (delta < 0) { | |
e8c5c824 LS |
1351 | /* |
1352 | * To shrink on this node, there must be a surplus page | |
1353 | */ | |
9a76db09 | 1354 | if (!h->surplus_huge_pages_node[nid]) { |
6ae11b27 LS |
1355 | next_nid = hstate_next_node_to_alloc(h, |
1356 | nodes_allowed); | |
e8c5c824 | 1357 | continue; |
9a76db09 | 1358 | } |
e8c5c824 LS |
1359 | } |
1360 | if (delta > 0) { | |
e8c5c824 LS |
1361 | /* |
1362 | * Surplus cannot exceed the total number of pages | |
1363 | */ | |
1364 | if (h->surplus_huge_pages_node[nid] >= | |
9a76db09 | 1365 | h->nr_huge_pages_node[nid]) { |
6ae11b27 LS |
1366 | next_nid = hstate_next_node_to_free(h, |
1367 | nodes_allowed); | |
e8c5c824 | 1368 | continue; |
9a76db09 | 1369 | } |
e8c5c824 | 1370 | } |
20a0307c WF |
1371 | |
1372 | h->surplus_huge_pages += delta; | |
1373 | h->surplus_huge_pages_node[nid] += delta; | |
1374 | ret = 1; | |
1375 | break; | |
e8c5c824 | 1376 | } while (next_nid != start_nid); |
20a0307c | 1377 | |
20a0307c WF |
1378 | return ret; |
1379 | } | |
1380 | ||
a5516438 | 1381 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
1382 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
1383 | nodemask_t *nodes_allowed) | |
1da177e4 | 1384 | { |
7893d1d5 | 1385 | unsigned long min_count, ret; |
1da177e4 | 1386 | |
aa888a74 AK |
1387 | if (h->order >= MAX_ORDER) |
1388 | return h->max_huge_pages; | |
1389 | ||
7893d1d5 AL |
1390 | /* |
1391 | * Increase the pool size | |
1392 | * First take pages out of surplus state. Then make up the | |
1393 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f NA |
1394 | * |
1395 | * We might race with alloc_buddy_huge_page() here and be unable | |
1396 | * to convert a surplus huge page to a normal huge page. That is | |
1397 | * not critical, though, it just means the overall size of the | |
1398 | * pool might be one hugepage larger than it needs to be, but | |
1399 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 1400 | */ |
1da177e4 | 1401 | spin_lock(&hugetlb_lock); |
a5516438 | 1402 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 1403 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
1404 | break; |
1405 | } | |
1406 | ||
a5516438 | 1407 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
1408 | /* |
1409 | * If this allocation races such that we no longer need the | |
1410 | * page, free_huge_page will handle it by freeing the page | |
1411 | * and reducing the surplus. | |
1412 | */ | |
1413 | spin_unlock(&hugetlb_lock); | |
6ae11b27 | 1414 | ret = alloc_fresh_huge_page(h, nodes_allowed); |
7893d1d5 AL |
1415 | spin_lock(&hugetlb_lock); |
1416 | if (!ret) | |
1417 | goto out; | |
1418 | ||
536240f2 MG |
1419 | /* Bail for signals. Probably ctrl-c from user */ |
1420 | if (signal_pending(current)) | |
1421 | goto out; | |
7893d1d5 | 1422 | } |
7893d1d5 AL |
1423 | |
1424 | /* | |
1425 | * Decrease the pool size | |
1426 | * First return free pages to the buddy allocator (being careful | |
1427 | * to keep enough around to satisfy reservations). Then place | |
1428 | * pages into surplus state as needed so the pool will shrink | |
1429 | * to the desired size as pages become free. | |
d1c3fb1f NA |
1430 | * |
1431 | * By placing pages into the surplus state independent of the | |
1432 | * overcommit value, we are allowing the surplus pool size to | |
1433 | * exceed overcommit. There are few sane options here. Since | |
1434 | * alloc_buddy_huge_page() is checking the global counter, | |
1435 | * though, we'll note that we're not allowed to exceed surplus | |
1436 | * and won't grow the pool anywhere else. Not until one of the | |
1437 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 1438 | */ |
a5516438 | 1439 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 1440 | min_count = max(count, min_count); |
6ae11b27 | 1441 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 1442 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 1443 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 1444 | break; |
1da177e4 | 1445 | } |
a5516438 | 1446 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 1447 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
1448 | break; |
1449 | } | |
1450 | out: | |
a5516438 | 1451 | ret = persistent_huge_pages(h); |
1da177e4 | 1452 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 1453 | return ret; |
1da177e4 LT |
1454 | } |
1455 | ||
a3437870 NA |
1456 | #define HSTATE_ATTR_RO(_name) \ |
1457 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
1458 | ||
1459 | #define HSTATE_ATTR(_name) \ | |
1460 | static struct kobj_attribute _name##_attr = \ | |
1461 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
1462 | ||
1463 | static struct kobject *hugepages_kobj; | |
1464 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1465 | ||
9a305230 LS |
1466 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
1467 | ||
1468 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
1469 | { |
1470 | int i; | |
9a305230 | 1471 | |
a3437870 | 1472 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
1473 | if (hstate_kobjs[i] == kobj) { |
1474 | if (nidp) | |
1475 | *nidp = NUMA_NO_NODE; | |
a3437870 | 1476 | return &hstates[i]; |
9a305230 LS |
1477 | } |
1478 | ||
1479 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
1480 | } |
1481 | ||
06808b08 | 1482 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
1483 | struct kobj_attribute *attr, char *buf) |
1484 | { | |
9a305230 LS |
1485 | struct hstate *h; |
1486 | unsigned long nr_huge_pages; | |
1487 | int nid; | |
1488 | ||
1489 | h = kobj_to_hstate(kobj, &nid); | |
1490 | if (nid == NUMA_NO_NODE) | |
1491 | nr_huge_pages = h->nr_huge_pages; | |
1492 | else | |
1493 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
1494 | ||
1495 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 1496 | } |
adbe8726 | 1497 | |
06808b08 LS |
1498 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
1499 | struct kobject *kobj, struct kobj_attribute *attr, | |
1500 | const char *buf, size_t len) | |
a3437870 NA |
1501 | { |
1502 | int err; | |
9a305230 | 1503 | int nid; |
06808b08 | 1504 | unsigned long count; |
9a305230 | 1505 | struct hstate *h; |
bad44b5b | 1506 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 1507 | |
06808b08 | 1508 | err = strict_strtoul(buf, 10, &count); |
73ae31e5 | 1509 | if (err) |
adbe8726 | 1510 | goto out; |
a3437870 | 1511 | |
9a305230 | 1512 | h = kobj_to_hstate(kobj, &nid); |
adbe8726 EM |
1513 | if (h->order >= MAX_ORDER) { |
1514 | err = -EINVAL; | |
1515 | goto out; | |
1516 | } | |
1517 | ||
9a305230 LS |
1518 | if (nid == NUMA_NO_NODE) { |
1519 | /* | |
1520 | * global hstate attribute | |
1521 | */ | |
1522 | if (!(obey_mempolicy && | |
1523 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
1524 | NODEMASK_FREE(nodes_allowed); | |
1525 | nodes_allowed = &node_states[N_HIGH_MEMORY]; | |
1526 | } | |
1527 | } else if (nodes_allowed) { | |
1528 | /* | |
1529 | * per node hstate attribute: adjust count to global, | |
1530 | * but restrict alloc/free to the specified node. | |
1531 | */ | |
1532 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
1533 | init_nodemask_of_node(nodes_allowed, nid); | |
1534 | } else | |
1535 | nodes_allowed = &node_states[N_HIGH_MEMORY]; | |
1536 | ||
06808b08 | 1537 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 1538 | |
9b5e5d0f | 1539 | if (nodes_allowed != &node_states[N_HIGH_MEMORY]) |
06808b08 LS |
1540 | NODEMASK_FREE(nodes_allowed); |
1541 | ||
1542 | return len; | |
adbe8726 EM |
1543 | out: |
1544 | NODEMASK_FREE(nodes_allowed); | |
1545 | return err; | |
06808b08 LS |
1546 | } |
1547 | ||
1548 | static ssize_t nr_hugepages_show(struct kobject *kobj, | |
1549 | struct kobj_attribute *attr, char *buf) | |
1550 | { | |
1551 | return nr_hugepages_show_common(kobj, attr, buf); | |
1552 | } | |
1553 | ||
1554 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
1555 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1556 | { | |
1557 | return nr_hugepages_store_common(false, kobj, attr, buf, len); | |
a3437870 NA |
1558 | } |
1559 | HSTATE_ATTR(nr_hugepages); | |
1560 | ||
06808b08 LS |
1561 | #ifdef CONFIG_NUMA |
1562 | ||
1563 | /* | |
1564 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
1565 | * huge page alloc/free. | |
1566 | */ | |
1567 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
1568 | struct kobj_attribute *attr, char *buf) | |
1569 | { | |
1570 | return nr_hugepages_show_common(kobj, attr, buf); | |
1571 | } | |
1572 | ||
1573 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
1574 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1575 | { | |
1576 | return nr_hugepages_store_common(true, kobj, attr, buf, len); | |
1577 | } | |
1578 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
1579 | #endif | |
1580 | ||
1581 | ||
a3437870 NA |
1582 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
1583 | struct kobj_attribute *attr, char *buf) | |
1584 | { | |
9a305230 | 1585 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1586 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
1587 | } | |
adbe8726 | 1588 | |
a3437870 NA |
1589 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
1590 | struct kobj_attribute *attr, const char *buf, size_t count) | |
1591 | { | |
1592 | int err; | |
1593 | unsigned long input; | |
9a305230 | 1594 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 1595 | |
adbe8726 EM |
1596 | if (h->order >= MAX_ORDER) |
1597 | return -EINVAL; | |
1598 | ||
a3437870 NA |
1599 | err = strict_strtoul(buf, 10, &input); |
1600 | if (err) | |
73ae31e5 | 1601 | return err; |
a3437870 NA |
1602 | |
1603 | spin_lock(&hugetlb_lock); | |
1604 | h->nr_overcommit_huge_pages = input; | |
1605 | spin_unlock(&hugetlb_lock); | |
1606 | ||
1607 | return count; | |
1608 | } | |
1609 | HSTATE_ATTR(nr_overcommit_hugepages); | |
1610 | ||
1611 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
1612 | struct kobj_attribute *attr, char *buf) | |
1613 | { | |
9a305230 LS |
1614 | struct hstate *h; |
1615 | unsigned long free_huge_pages; | |
1616 | int nid; | |
1617 | ||
1618 | h = kobj_to_hstate(kobj, &nid); | |
1619 | if (nid == NUMA_NO_NODE) | |
1620 | free_huge_pages = h->free_huge_pages; | |
1621 | else | |
1622 | free_huge_pages = h->free_huge_pages_node[nid]; | |
1623 | ||
1624 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
1625 | } |
1626 | HSTATE_ATTR_RO(free_hugepages); | |
1627 | ||
1628 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
1629 | struct kobj_attribute *attr, char *buf) | |
1630 | { | |
9a305230 | 1631 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1632 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
1633 | } | |
1634 | HSTATE_ATTR_RO(resv_hugepages); | |
1635 | ||
1636 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
1637 | struct kobj_attribute *attr, char *buf) | |
1638 | { | |
9a305230 LS |
1639 | struct hstate *h; |
1640 | unsigned long surplus_huge_pages; | |
1641 | int nid; | |
1642 | ||
1643 | h = kobj_to_hstate(kobj, &nid); | |
1644 | if (nid == NUMA_NO_NODE) | |
1645 | surplus_huge_pages = h->surplus_huge_pages; | |
1646 | else | |
1647 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
1648 | ||
1649 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
1650 | } |
1651 | HSTATE_ATTR_RO(surplus_hugepages); | |
1652 | ||
1653 | static struct attribute *hstate_attrs[] = { | |
1654 | &nr_hugepages_attr.attr, | |
1655 | &nr_overcommit_hugepages_attr.attr, | |
1656 | &free_hugepages_attr.attr, | |
1657 | &resv_hugepages_attr.attr, | |
1658 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
1659 | #ifdef CONFIG_NUMA |
1660 | &nr_hugepages_mempolicy_attr.attr, | |
1661 | #endif | |
a3437870 NA |
1662 | NULL, |
1663 | }; | |
1664 | ||
1665 | static struct attribute_group hstate_attr_group = { | |
1666 | .attrs = hstate_attrs, | |
1667 | }; | |
1668 | ||
094e9539 JM |
1669 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
1670 | struct kobject **hstate_kobjs, | |
1671 | struct attribute_group *hstate_attr_group) | |
a3437870 NA |
1672 | { |
1673 | int retval; | |
972dc4de | 1674 | int hi = hstate_index(h); |
a3437870 | 1675 | |
9a305230 LS |
1676 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
1677 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
1678 | return -ENOMEM; |
1679 | ||
9a305230 | 1680 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 1681 | if (retval) |
9a305230 | 1682 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
1683 | |
1684 | return retval; | |
1685 | } | |
1686 | ||
1687 | static void __init hugetlb_sysfs_init(void) | |
1688 | { | |
1689 | struct hstate *h; | |
1690 | int err; | |
1691 | ||
1692 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
1693 | if (!hugepages_kobj) | |
1694 | return; | |
1695 | ||
1696 | for_each_hstate(h) { | |
9a305230 LS |
1697 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
1698 | hstate_kobjs, &hstate_attr_group); | |
a3437870 NA |
1699 | if (err) |
1700 | printk(KERN_ERR "Hugetlb: Unable to add hstate %s", | |
1701 | h->name); | |
1702 | } | |
1703 | } | |
1704 | ||
9a305230 LS |
1705 | #ifdef CONFIG_NUMA |
1706 | ||
1707 | /* | |
1708 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
1709 | * with node devices in node_devices[] using a parallel array. The array |
1710 | * index of a node device or _hstate == node id. | |
1711 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
1712 | * the base kernel, on the hugetlb module. |
1713 | */ | |
1714 | struct node_hstate { | |
1715 | struct kobject *hugepages_kobj; | |
1716 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1717 | }; | |
1718 | struct node_hstate node_hstates[MAX_NUMNODES]; | |
1719 | ||
1720 | /* | |
10fbcf4c | 1721 | * A subset of global hstate attributes for node devices |
9a305230 LS |
1722 | */ |
1723 | static struct attribute *per_node_hstate_attrs[] = { | |
1724 | &nr_hugepages_attr.attr, | |
1725 | &free_hugepages_attr.attr, | |
1726 | &surplus_hugepages_attr.attr, | |
1727 | NULL, | |
1728 | }; | |
1729 | ||
1730 | static struct attribute_group per_node_hstate_attr_group = { | |
1731 | .attrs = per_node_hstate_attrs, | |
1732 | }; | |
1733 | ||
1734 | /* | |
10fbcf4c | 1735 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
1736 | * Returns node id via non-NULL nidp. |
1737 | */ | |
1738 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
1739 | { | |
1740 | int nid; | |
1741 | ||
1742 | for (nid = 0; nid < nr_node_ids; nid++) { | |
1743 | struct node_hstate *nhs = &node_hstates[nid]; | |
1744 | int i; | |
1745 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
1746 | if (nhs->hstate_kobjs[i] == kobj) { | |
1747 | if (nidp) | |
1748 | *nidp = nid; | |
1749 | return &hstates[i]; | |
1750 | } | |
1751 | } | |
1752 | ||
1753 | BUG(); | |
1754 | return NULL; | |
1755 | } | |
1756 | ||
1757 | /* | |
10fbcf4c | 1758 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
1759 | * No-op if no hstate attributes attached. |
1760 | */ | |
1761 | void hugetlb_unregister_node(struct node *node) | |
1762 | { | |
1763 | struct hstate *h; | |
10fbcf4c | 1764 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
1765 | |
1766 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 1767 | return; /* no hstate attributes */ |
9a305230 | 1768 | |
972dc4de AK |
1769 | for_each_hstate(h) { |
1770 | int idx = hstate_index(h); | |
1771 | if (nhs->hstate_kobjs[idx]) { | |
1772 | kobject_put(nhs->hstate_kobjs[idx]); | |
1773 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 1774 | } |
972dc4de | 1775 | } |
9a305230 LS |
1776 | |
1777 | kobject_put(nhs->hugepages_kobj); | |
1778 | nhs->hugepages_kobj = NULL; | |
1779 | } | |
1780 | ||
1781 | /* | |
10fbcf4c | 1782 | * hugetlb module exit: unregister hstate attributes from node devices |
9a305230 LS |
1783 | * that have them. |
1784 | */ | |
1785 | static void hugetlb_unregister_all_nodes(void) | |
1786 | { | |
1787 | int nid; | |
1788 | ||
1789 | /* | |
10fbcf4c | 1790 | * disable node device registrations. |
9a305230 LS |
1791 | */ |
1792 | register_hugetlbfs_with_node(NULL, NULL); | |
1793 | ||
1794 | /* | |
1795 | * remove hstate attributes from any nodes that have them. | |
1796 | */ | |
1797 | for (nid = 0; nid < nr_node_ids; nid++) | |
1798 | hugetlb_unregister_node(&node_devices[nid]); | |
1799 | } | |
1800 | ||
1801 | /* | |
10fbcf4c | 1802 | * Register hstate attributes for a single node device. |
9a305230 LS |
1803 | * No-op if attributes already registered. |
1804 | */ | |
1805 | void hugetlb_register_node(struct node *node) | |
1806 | { | |
1807 | struct hstate *h; | |
10fbcf4c | 1808 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
1809 | int err; |
1810 | ||
1811 | if (nhs->hugepages_kobj) | |
1812 | return; /* already allocated */ | |
1813 | ||
1814 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 1815 | &node->dev.kobj); |
9a305230 LS |
1816 | if (!nhs->hugepages_kobj) |
1817 | return; | |
1818 | ||
1819 | for_each_hstate(h) { | |
1820 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
1821 | nhs->hstate_kobjs, | |
1822 | &per_node_hstate_attr_group); | |
1823 | if (err) { | |
1824 | printk(KERN_ERR "Hugetlb: Unable to add hstate %s" | |
1825 | " for node %d\n", | |
10fbcf4c | 1826 | h->name, node->dev.id); |
9a305230 LS |
1827 | hugetlb_unregister_node(node); |
1828 | break; | |
1829 | } | |
1830 | } | |
1831 | } | |
1832 | ||
1833 | /* | |
9b5e5d0f | 1834 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
1835 | * devices of nodes that have memory. All on-line nodes should have |
1836 | * registered their associated device by this time. | |
9a305230 LS |
1837 | */ |
1838 | static void hugetlb_register_all_nodes(void) | |
1839 | { | |
1840 | int nid; | |
1841 | ||
9b5e5d0f | 1842 | for_each_node_state(nid, N_HIGH_MEMORY) { |
9a305230 | 1843 | struct node *node = &node_devices[nid]; |
10fbcf4c | 1844 | if (node->dev.id == nid) |
9a305230 LS |
1845 | hugetlb_register_node(node); |
1846 | } | |
1847 | ||
1848 | /* | |
10fbcf4c | 1849 | * Let the node device driver know we're here so it can |
9a305230 LS |
1850 | * [un]register hstate attributes on node hotplug. |
1851 | */ | |
1852 | register_hugetlbfs_with_node(hugetlb_register_node, | |
1853 | hugetlb_unregister_node); | |
1854 | } | |
1855 | #else /* !CONFIG_NUMA */ | |
1856 | ||
1857 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
1858 | { | |
1859 | BUG(); | |
1860 | if (nidp) | |
1861 | *nidp = -1; | |
1862 | return NULL; | |
1863 | } | |
1864 | ||
1865 | static void hugetlb_unregister_all_nodes(void) { } | |
1866 | ||
1867 | static void hugetlb_register_all_nodes(void) { } | |
1868 | ||
1869 | #endif | |
1870 | ||
a3437870 NA |
1871 | static void __exit hugetlb_exit(void) |
1872 | { | |
1873 | struct hstate *h; | |
1874 | ||
9a305230 LS |
1875 | hugetlb_unregister_all_nodes(); |
1876 | ||
a3437870 | 1877 | for_each_hstate(h) { |
972dc4de | 1878 | kobject_put(hstate_kobjs[hstate_index(h)]); |
a3437870 NA |
1879 | } |
1880 | ||
1881 | kobject_put(hugepages_kobj); | |
1882 | } | |
1883 | module_exit(hugetlb_exit); | |
1884 | ||
1885 | static int __init hugetlb_init(void) | |
1886 | { | |
0ef89d25 BH |
1887 | /* Some platform decide whether they support huge pages at boot |
1888 | * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when | |
1889 | * there is no such support | |
1890 | */ | |
1891 | if (HPAGE_SHIFT == 0) | |
1892 | return 0; | |
a3437870 | 1893 | |
e11bfbfc NP |
1894 | if (!size_to_hstate(default_hstate_size)) { |
1895 | default_hstate_size = HPAGE_SIZE; | |
1896 | if (!size_to_hstate(default_hstate_size)) | |
1897 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 1898 | } |
972dc4de | 1899 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
e11bfbfc NP |
1900 | if (default_hstate_max_huge_pages) |
1901 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
a3437870 NA |
1902 | |
1903 | hugetlb_init_hstates(); | |
1904 | ||
aa888a74 AK |
1905 | gather_bootmem_prealloc(); |
1906 | ||
a3437870 NA |
1907 | report_hugepages(); |
1908 | ||
1909 | hugetlb_sysfs_init(); | |
1910 | ||
9a305230 LS |
1911 | hugetlb_register_all_nodes(); |
1912 | ||
a3437870 NA |
1913 | return 0; |
1914 | } | |
1915 | module_init(hugetlb_init); | |
1916 | ||
1917 | /* Should be called on processing a hugepagesz=... option */ | |
1918 | void __init hugetlb_add_hstate(unsigned order) | |
1919 | { | |
1920 | struct hstate *h; | |
8faa8b07 AK |
1921 | unsigned long i; |
1922 | ||
a3437870 NA |
1923 | if (size_to_hstate(PAGE_SIZE << order)) { |
1924 | printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n"); | |
1925 | return; | |
1926 | } | |
47d38344 | 1927 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 1928 | BUG_ON(order == 0); |
47d38344 | 1929 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
1930 | h->order = order; |
1931 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
1932 | h->nr_huge_pages = 0; |
1933 | h->free_huge_pages = 0; | |
1934 | for (i = 0; i < MAX_NUMNODES; ++i) | |
1935 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 1936 | INIT_LIST_HEAD(&h->hugepage_activelist); |
9b5e5d0f LS |
1937 | h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]); |
1938 | h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]); | |
a3437870 NA |
1939 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
1940 | huge_page_size(h)/1024); | |
abb8206c AK |
1941 | /* |
1942 | * Add cgroup control files only if the huge page consists | |
1943 | * of more than two normal pages. This is because we use | |
1944 | * page[2].lru.next for storing cgoup details. | |
1945 | */ | |
1946 | if (order >= HUGETLB_CGROUP_MIN_ORDER) | |
1947 | hugetlb_cgroup_file_init(hugetlb_max_hstate - 1); | |
8faa8b07 | 1948 | |
a3437870 NA |
1949 | parsed_hstate = h; |
1950 | } | |
1951 | ||
e11bfbfc | 1952 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
1953 | { |
1954 | unsigned long *mhp; | |
8faa8b07 | 1955 | static unsigned long *last_mhp; |
a3437870 NA |
1956 | |
1957 | /* | |
47d38344 | 1958 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
1959 | * so this hugepages= parameter goes to the "default hstate". |
1960 | */ | |
47d38344 | 1961 | if (!hugetlb_max_hstate) |
a3437870 NA |
1962 | mhp = &default_hstate_max_huge_pages; |
1963 | else | |
1964 | mhp = &parsed_hstate->max_huge_pages; | |
1965 | ||
8faa8b07 AK |
1966 | if (mhp == last_mhp) { |
1967 | printk(KERN_WARNING "hugepages= specified twice without " | |
1968 | "interleaving hugepagesz=, ignoring\n"); | |
1969 | return 1; | |
1970 | } | |
1971 | ||
a3437870 NA |
1972 | if (sscanf(s, "%lu", mhp) <= 0) |
1973 | *mhp = 0; | |
1974 | ||
8faa8b07 AK |
1975 | /* |
1976 | * Global state is always initialized later in hugetlb_init. | |
1977 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
1978 | * use the bootmem allocator. | |
1979 | */ | |
47d38344 | 1980 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
1981 | hugetlb_hstate_alloc_pages(parsed_hstate); |
1982 | ||
1983 | last_mhp = mhp; | |
1984 | ||
a3437870 NA |
1985 | return 1; |
1986 | } | |
e11bfbfc NP |
1987 | __setup("hugepages=", hugetlb_nrpages_setup); |
1988 | ||
1989 | static int __init hugetlb_default_setup(char *s) | |
1990 | { | |
1991 | default_hstate_size = memparse(s, &s); | |
1992 | return 1; | |
1993 | } | |
1994 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 1995 | |
8a213460 NA |
1996 | static unsigned int cpuset_mems_nr(unsigned int *array) |
1997 | { | |
1998 | int node; | |
1999 | unsigned int nr = 0; | |
2000 | ||
2001 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
2002 | nr += array[node]; | |
2003 | ||
2004 | return nr; | |
2005 | } | |
2006 | ||
2007 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
2008 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
2009 | struct ctl_table *table, int write, | |
2010 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 2011 | { |
e5ff2159 AK |
2012 | struct hstate *h = &default_hstate; |
2013 | unsigned long tmp; | |
08d4a246 | 2014 | int ret; |
e5ff2159 | 2015 | |
c033a93c | 2016 | tmp = h->max_huge_pages; |
e5ff2159 | 2017 | |
adbe8726 EM |
2018 | if (write && h->order >= MAX_ORDER) |
2019 | return -EINVAL; | |
2020 | ||
e5ff2159 AK |
2021 | table->data = &tmp; |
2022 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2023 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2024 | if (ret) | |
2025 | goto out; | |
e5ff2159 | 2026 | |
06808b08 | 2027 | if (write) { |
bad44b5b DR |
2028 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, |
2029 | GFP_KERNEL | __GFP_NORETRY); | |
06808b08 LS |
2030 | if (!(obey_mempolicy && |
2031 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
2032 | NODEMASK_FREE(nodes_allowed); | |
2033 | nodes_allowed = &node_states[N_HIGH_MEMORY]; | |
2034 | } | |
2035 | h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); | |
2036 | ||
2037 | if (nodes_allowed != &node_states[N_HIGH_MEMORY]) | |
2038 | NODEMASK_FREE(nodes_allowed); | |
2039 | } | |
08d4a246 MH |
2040 | out: |
2041 | return ret; | |
1da177e4 | 2042 | } |
396faf03 | 2043 | |
06808b08 LS |
2044 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2045 | void __user *buffer, size_t *length, loff_t *ppos) | |
2046 | { | |
2047 | ||
2048 | return hugetlb_sysctl_handler_common(false, table, write, | |
2049 | buffer, length, ppos); | |
2050 | } | |
2051 | ||
2052 | #ifdef CONFIG_NUMA | |
2053 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2054 | void __user *buffer, size_t *length, loff_t *ppos) | |
2055 | { | |
2056 | return hugetlb_sysctl_handler_common(true, table, write, | |
2057 | buffer, length, ppos); | |
2058 | } | |
2059 | #endif /* CONFIG_NUMA */ | |
2060 | ||
396faf03 | 2061 | int hugetlb_treat_movable_handler(struct ctl_table *table, int write, |
8d65af78 | 2062 | void __user *buffer, |
396faf03 MG |
2063 | size_t *length, loff_t *ppos) |
2064 | { | |
8d65af78 | 2065 | proc_dointvec(table, write, buffer, length, ppos); |
396faf03 MG |
2066 | if (hugepages_treat_as_movable) |
2067 | htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; | |
2068 | else | |
2069 | htlb_alloc_mask = GFP_HIGHUSER; | |
2070 | return 0; | |
2071 | } | |
2072 | ||
a3d0c6aa | 2073 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 2074 | void __user *buffer, |
a3d0c6aa NA |
2075 | size_t *length, loff_t *ppos) |
2076 | { | |
a5516438 | 2077 | struct hstate *h = &default_hstate; |
e5ff2159 | 2078 | unsigned long tmp; |
08d4a246 | 2079 | int ret; |
e5ff2159 | 2080 | |
c033a93c | 2081 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 2082 | |
adbe8726 EM |
2083 | if (write && h->order >= MAX_ORDER) |
2084 | return -EINVAL; | |
2085 | ||
e5ff2159 AK |
2086 | table->data = &tmp; |
2087 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2088 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2089 | if (ret) | |
2090 | goto out; | |
e5ff2159 AK |
2091 | |
2092 | if (write) { | |
2093 | spin_lock(&hugetlb_lock); | |
2094 | h->nr_overcommit_huge_pages = tmp; | |
2095 | spin_unlock(&hugetlb_lock); | |
2096 | } | |
08d4a246 MH |
2097 | out: |
2098 | return ret; | |
a3d0c6aa NA |
2099 | } |
2100 | ||
1da177e4 LT |
2101 | #endif /* CONFIG_SYSCTL */ |
2102 | ||
e1759c21 | 2103 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 2104 | { |
a5516438 | 2105 | struct hstate *h = &default_hstate; |
e1759c21 | 2106 | seq_printf(m, |
4f98a2fe RR |
2107 | "HugePages_Total: %5lu\n" |
2108 | "HugePages_Free: %5lu\n" | |
2109 | "HugePages_Rsvd: %5lu\n" | |
2110 | "HugePages_Surp: %5lu\n" | |
2111 | "Hugepagesize: %8lu kB\n", | |
a5516438 AK |
2112 | h->nr_huge_pages, |
2113 | h->free_huge_pages, | |
2114 | h->resv_huge_pages, | |
2115 | h->surplus_huge_pages, | |
2116 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
2117 | } |
2118 | ||
2119 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
2120 | { | |
a5516438 | 2121 | struct hstate *h = &default_hstate; |
1da177e4 LT |
2122 | return sprintf(buf, |
2123 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
2124 | "Node %d HugePages_Free: %5u\n" |
2125 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
2126 | nid, h->nr_huge_pages_node[nid], |
2127 | nid, h->free_huge_pages_node[nid], | |
2128 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
2129 | } |
2130 | ||
1da177e4 LT |
2131 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
2132 | unsigned long hugetlb_total_pages(void) | |
2133 | { | |
a5516438 AK |
2134 | struct hstate *h = &default_hstate; |
2135 | return h->nr_huge_pages * pages_per_huge_page(h); | |
1da177e4 | 2136 | } |
1da177e4 | 2137 | |
a5516438 | 2138 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
2139 | { |
2140 | int ret = -ENOMEM; | |
2141 | ||
2142 | spin_lock(&hugetlb_lock); | |
2143 | /* | |
2144 | * When cpuset is configured, it breaks the strict hugetlb page | |
2145 | * reservation as the accounting is done on a global variable. Such | |
2146 | * reservation is completely rubbish in the presence of cpuset because | |
2147 | * the reservation is not checked against page availability for the | |
2148 | * current cpuset. Application can still potentially OOM'ed by kernel | |
2149 | * with lack of free htlb page in cpuset that the task is in. | |
2150 | * Attempt to enforce strict accounting with cpuset is almost | |
2151 | * impossible (or too ugly) because cpuset is too fluid that | |
2152 | * task or memory node can be dynamically moved between cpusets. | |
2153 | * | |
2154 | * The change of semantics for shared hugetlb mapping with cpuset is | |
2155 | * undesirable. However, in order to preserve some of the semantics, | |
2156 | * we fall back to check against current free page availability as | |
2157 | * a best attempt and hopefully to minimize the impact of changing | |
2158 | * semantics that cpuset has. | |
2159 | */ | |
2160 | if (delta > 0) { | |
a5516438 | 2161 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
2162 | goto out; |
2163 | ||
a5516438 AK |
2164 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
2165 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
2166 | goto out; |
2167 | } | |
2168 | } | |
2169 | ||
2170 | ret = 0; | |
2171 | if (delta < 0) | |
a5516438 | 2172 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
2173 | |
2174 | out: | |
2175 | spin_unlock(&hugetlb_lock); | |
2176 | return ret; | |
2177 | } | |
2178 | ||
84afd99b AW |
2179 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
2180 | { | |
2181 | struct resv_map *reservations = vma_resv_map(vma); | |
2182 | ||
2183 | /* | |
2184 | * This new VMA should share its siblings reservation map if present. | |
2185 | * The VMA will only ever have a valid reservation map pointer where | |
2186 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 2187 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
2188 | * after this open call completes. It is therefore safe to take a |
2189 | * new reference here without additional locking. | |
2190 | */ | |
2191 | if (reservations) | |
2192 | kref_get(&reservations->refs); | |
2193 | } | |
2194 | ||
c50ac050 DH |
2195 | static void resv_map_put(struct vm_area_struct *vma) |
2196 | { | |
2197 | struct resv_map *reservations = vma_resv_map(vma); | |
2198 | ||
2199 | if (!reservations) | |
2200 | return; | |
2201 | kref_put(&reservations->refs, resv_map_release); | |
2202 | } | |
2203 | ||
a1e78772 MG |
2204 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
2205 | { | |
a5516438 | 2206 | struct hstate *h = hstate_vma(vma); |
84afd99b | 2207 | struct resv_map *reservations = vma_resv_map(vma); |
90481622 | 2208 | struct hugepage_subpool *spool = subpool_vma(vma); |
84afd99b AW |
2209 | unsigned long reserve; |
2210 | unsigned long start; | |
2211 | unsigned long end; | |
2212 | ||
2213 | if (reservations) { | |
a5516438 AK |
2214 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
2215 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b AW |
2216 | |
2217 | reserve = (end - start) - | |
2218 | region_count(&reservations->regions, start, end); | |
2219 | ||
c50ac050 | 2220 | resv_map_put(vma); |
84afd99b | 2221 | |
7251ff78 | 2222 | if (reserve) { |
a5516438 | 2223 | hugetlb_acct_memory(h, -reserve); |
90481622 | 2224 | hugepage_subpool_put_pages(spool, reserve); |
7251ff78 | 2225 | } |
84afd99b | 2226 | } |
a1e78772 MG |
2227 | } |
2228 | ||
1da177e4 LT |
2229 | /* |
2230 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
2231 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
2232 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
2233 | * this far. | |
2234 | */ | |
d0217ac0 | 2235 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
2236 | { |
2237 | BUG(); | |
d0217ac0 | 2238 | return 0; |
1da177e4 LT |
2239 | } |
2240 | ||
f0f37e2f | 2241 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 2242 | .fault = hugetlb_vm_op_fault, |
84afd99b | 2243 | .open = hugetlb_vm_op_open, |
a1e78772 | 2244 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
2245 | }; |
2246 | ||
1e8f889b DG |
2247 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
2248 | int writable) | |
63551ae0 DG |
2249 | { |
2250 | pte_t entry; | |
2251 | ||
1e8f889b | 2252 | if (writable) { |
63551ae0 DG |
2253 | entry = |
2254 | pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); | |
2255 | } else { | |
7f2e9525 | 2256 | entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot)); |
63551ae0 DG |
2257 | } |
2258 | entry = pte_mkyoung(entry); | |
2259 | entry = pte_mkhuge(entry); | |
d9ed9faa | 2260 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
2261 | |
2262 | return entry; | |
2263 | } | |
2264 | ||
1e8f889b DG |
2265 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
2266 | unsigned long address, pte_t *ptep) | |
2267 | { | |
2268 | pte_t entry; | |
2269 | ||
7f2e9525 | 2270 | entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 2271 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 2272 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
2273 | } |
2274 | ||
2275 | ||
63551ae0 DG |
2276 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
2277 | struct vm_area_struct *vma) | |
2278 | { | |
2279 | pte_t *src_pte, *dst_pte, entry; | |
2280 | struct page *ptepage; | |
1c59827d | 2281 | unsigned long addr; |
1e8f889b | 2282 | int cow; |
a5516438 AK |
2283 | struct hstate *h = hstate_vma(vma); |
2284 | unsigned long sz = huge_page_size(h); | |
1e8f889b DG |
2285 | |
2286 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 2287 | |
a5516438 | 2288 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
c74df32c HD |
2289 | src_pte = huge_pte_offset(src, addr); |
2290 | if (!src_pte) | |
2291 | continue; | |
a5516438 | 2292 | dst_pte = huge_pte_alloc(dst, addr, sz); |
63551ae0 DG |
2293 | if (!dst_pte) |
2294 | goto nomem; | |
c5c99429 LW |
2295 | |
2296 | /* If the pagetables are shared don't copy or take references */ | |
2297 | if (dst_pte == src_pte) | |
2298 | continue; | |
2299 | ||
c74df32c | 2300 | spin_lock(&dst->page_table_lock); |
46478758 | 2301 | spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); |
7f2e9525 | 2302 | if (!huge_pte_none(huge_ptep_get(src_pte))) { |
1e8f889b | 2303 | if (cow) |
7f2e9525 GS |
2304 | huge_ptep_set_wrprotect(src, addr, src_pte); |
2305 | entry = huge_ptep_get(src_pte); | |
1c59827d HD |
2306 | ptepage = pte_page(entry); |
2307 | get_page(ptepage); | |
0fe6e20b | 2308 | page_dup_rmap(ptepage); |
1c59827d HD |
2309 | set_huge_pte_at(dst, addr, dst_pte, entry); |
2310 | } | |
2311 | spin_unlock(&src->page_table_lock); | |
c74df32c | 2312 | spin_unlock(&dst->page_table_lock); |
63551ae0 DG |
2313 | } |
2314 | return 0; | |
2315 | ||
2316 | nomem: | |
2317 | return -ENOMEM; | |
2318 | } | |
2319 | ||
290408d4 NH |
2320 | static int is_hugetlb_entry_migration(pte_t pte) |
2321 | { | |
2322 | swp_entry_t swp; | |
2323 | ||
2324 | if (huge_pte_none(pte) || pte_present(pte)) | |
2325 | return 0; | |
2326 | swp = pte_to_swp_entry(pte); | |
32f84528 | 2327 | if (non_swap_entry(swp) && is_migration_entry(swp)) |
290408d4 | 2328 | return 1; |
32f84528 | 2329 | else |
290408d4 NH |
2330 | return 0; |
2331 | } | |
2332 | ||
fd6a03ed NH |
2333 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) |
2334 | { | |
2335 | swp_entry_t swp; | |
2336 | ||
2337 | if (huge_pte_none(pte) || pte_present(pte)) | |
2338 | return 0; | |
2339 | swp = pte_to_swp_entry(pte); | |
32f84528 | 2340 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) |
fd6a03ed | 2341 | return 1; |
32f84528 | 2342 | else |
fd6a03ed NH |
2343 | return 0; |
2344 | } | |
2345 | ||
24669e58 AK |
2346 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
2347 | unsigned long start, unsigned long end, | |
2348 | struct page *ref_page) | |
63551ae0 | 2349 | { |
24669e58 | 2350 | int force_flush = 0; |
63551ae0 DG |
2351 | struct mm_struct *mm = vma->vm_mm; |
2352 | unsigned long address; | |
c7546f8f | 2353 | pte_t *ptep; |
63551ae0 DG |
2354 | pte_t pte; |
2355 | struct page *page; | |
a5516438 AK |
2356 | struct hstate *h = hstate_vma(vma); |
2357 | unsigned long sz = huge_page_size(h); | |
2ec74c3e SG |
2358 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
2359 | const unsigned long mmun_end = end; /* For mmu_notifiers */ | |
a5516438 | 2360 | |
63551ae0 | 2361 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
2362 | BUG_ON(start & ~huge_page_mask(h)); |
2363 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 2364 | |
24669e58 | 2365 | tlb_start_vma(tlb, vma); |
2ec74c3e | 2366 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
24669e58 | 2367 | again: |
508034a3 | 2368 | spin_lock(&mm->page_table_lock); |
a5516438 | 2369 | for (address = start; address < end; address += sz) { |
c7546f8f | 2370 | ptep = huge_pte_offset(mm, address); |
4c887265 | 2371 | if (!ptep) |
c7546f8f DG |
2372 | continue; |
2373 | ||
39dde65c CK |
2374 | if (huge_pmd_unshare(mm, &address, ptep)) |
2375 | continue; | |
2376 | ||
6629326b HD |
2377 | pte = huge_ptep_get(ptep); |
2378 | if (huge_pte_none(pte)) | |
2379 | continue; | |
2380 | ||
2381 | /* | |
2382 | * HWPoisoned hugepage is already unmapped and dropped reference | |
2383 | */ | |
2384 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) | |
2385 | continue; | |
2386 | ||
2387 | page = pte_page(pte); | |
04f2cbe3 MG |
2388 | /* |
2389 | * If a reference page is supplied, it is because a specific | |
2390 | * page is being unmapped, not a range. Ensure the page we | |
2391 | * are about to unmap is the actual page of interest. | |
2392 | */ | |
2393 | if (ref_page) { | |
04f2cbe3 MG |
2394 | if (page != ref_page) |
2395 | continue; | |
2396 | ||
2397 | /* | |
2398 | * Mark the VMA as having unmapped its page so that | |
2399 | * future faults in this VMA will fail rather than | |
2400 | * looking like data was lost | |
2401 | */ | |
2402 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
2403 | } | |
2404 | ||
c7546f8f | 2405 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
24669e58 | 2406 | tlb_remove_tlb_entry(tlb, ptep, address); |
6649a386 KC |
2407 | if (pte_dirty(pte)) |
2408 | set_page_dirty(page); | |
9e81130b | 2409 | |
24669e58 AK |
2410 | page_remove_rmap(page); |
2411 | force_flush = !__tlb_remove_page(tlb, page); | |
2412 | if (force_flush) | |
2413 | break; | |
9e81130b HD |
2414 | /* Bail out after unmapping reference page if supplied */ |
2415 | if (ref_page) | |
2416 | break; | |
63551ae0 | 2417 | } |
cd2934a3 | 2418 | spin_unlock(&mm->page_table_lock); |
24669e58 AK |
2419 | /* |
2420 | * mmu_gather ran out of room to batch pages, we break out of | |
2421 | * the PTE lock to avoid doing the potential expensive TLB invalidate | |
2422 | * and page-free while holding it. | |
2423 | */ | |
2424 | if (force_flush) { | |
2425 | force_flush = 0; | |
2426 | tlb_flush_mmu(tlb); | |
2427 | if (address < end && !ref_page) | |
2428 | goto again; | |
fe1668ae | 2429 | } |
2ec74c3e | 2430 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
24669e58 | 2431 | tlb_end_vma(tlb, vma); |
1da177e4 | 2432 | } |
63551ae0 | 2433 | |
d833352a MG |
2434 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
2435 | struct vm_area_struct *vma, unsigned long start, | |
2436 | unsigned long end, struct page *ref_page) | |
2437 | { | |
2438 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
2439 | ||
2440 | /* | |
2441 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
2442 | * test will fail on a vma being torn down, and not grab a page table | |
2443 | * on its way out. We're lucky that the flag has such an appropriate | |
2444 | * name, and can in fact be safely cleared here. We could clear it | |
2445 | * before the __unmap_hugepage_range above, but all that's necessary | |
2446 | * is to clear it before releasing the i_mmap_mutex. This works | |
2447 | * because in the context this is called, the VMA is about to be | |
2448 | * destroyed and the i_mmap_mutex is held. | |
2449 | */ | |
2450 | vma->vm_flags &= ~VM_MAYSHARE; | |
2451 | } | |
2452 | ||
502717f4 | 2453 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 2454 | unsigned long end, struct page *ref_page) |
502717f4 | 2455 | { |
24669e58 AK |
2456 | struct mm_struct *mm; |
2457 | struct mmu_gather tlb; | |
2458 | ||
2459 | mm = vma->vm_mm; | |
2460 | ||
2461 | tlb_gather_mmu(&tlb, mm, 0); | |
2462 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); | |
2463 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 CK |
2464 | } |
2465 | ||
04f2cbe3 MG |
2466 | /* |
2467 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
2468 | * mappping it owns the reserve page for. The intention is to unmap the page | |
2469 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
2470 | * same region. | |
2471 | */ | |
2a4b3ded HH |
2472 | static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
2473 | struct page *page, unsigned long address) | |
04f2cbe3 | 2474 | { |
7526674d | 2475 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
2476 | struct vm_area_struct *iter_vma; |
2477 | struct address_space *mapping; | |
04f2cbe3 MG |
2478 | pgoff_t pgoff; |
2479 | ||
2480 | /* | |
2481 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
2482 | * from page cache lookup which is in HPAGE_SIZE units. | |
2483 | */ | |
7526674d | 2484 | address = address & huge_page_mask(h); |
36e4f20a MH |
2485 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
2486 | vma->vm_pgoff; | |
90481622 | 2487 | mapping = vma->vm_file->f_dentry->d_inode->i_mapping; |
04f2cbe3 | 2488 | |
4eb2b1dc MG |
2489 | /* |
2490 | * Take the mapping lock for the duration of the table walk. As | |
2491 | * this mapping should be shared between all the VMAs, | |
2492 | * __unmap_hugepage_range() is called as the lock is already held | |
2493 | */ | |
3d48ae45 | 2494 | mutex_lock(&mapping->i_mmap_mutex); |
6b2dbba8 | 2495 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
2496 | /* Do not unmap the current VMA */ |
2497 | if (iter_vma == vma) | |
2498 | continue; | |
2499 | ||
2500 | /* | |
2501 | * Unmap the page from other VMAs without their own reserves. | |
2502 | * They get marked to be SIGKILLed if they fault in these | |
2503 | * areas. This is because a future no-page fault on this VMA | |
2504 | * could insert a zeroed page instead of the data existing | |
2505 | * from the time of fork. This would look like data corruption | |
2506 | */ | |
2507 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
2508 | unmap_hugepage_range(iter_vma, address, |
2509 | address + huge_page_size(h), page); | |
04f2cbe3 | 2510 | } |
3d48ae45 | 2511 | mutex_unlock(&mapping->i_mmap_mutex); |
04f2cbe3 MG |
2512 | |
2513 | return 1; | |
2514 | } | |
2515 | ||
0fe6e20b NH |
2516 | /* |
2517 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
2518 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
2519 | * cannot race with other handlers or page migration. | |
2520 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 2521 | */ |
1e8f889b | 2522 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
04f2cbe3 MG |
2523 | unsigned long address, pte_t *ptep, pte_t pte, |
2524 | struct page *pagecache_page) | |
1e8f889b | 2525 | { |
a5516438 | 2526 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 2527 | struct page *old_page, *new_page; |
79ac6ba4 | 2528 | int avoidcopy; |
04f2cbe3 | 2529 | int outside_reserve = 0; |
2ec74c3e SG |
2530 | unsigned long mmun_start; /* For mmu_notifiers */ |
2531 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1e8f889b DG |
2532 | |
2533 | old_page = pte_page(pte); | |
2534 | ||
04f2cbe3 | 2535 | retry_avoidcopy: |
1e8f889b DG |
2536 | /* If no-one else is actually using this page, avoid the copy |
2537 | * and just make the page writable */ | |
0fe6e20b | 2538 | avoidcopy = (page_mapcount(old_page) == 1); |
1e8f889b | 2539 | if (avoidcopy) { |
56c9cfb1 NH |
2540 | if (PageAnon(old_page)) |
2541 | page_move_anon_rmap(old_page, vma, address); | |
1e8f889b | 2542 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 2543 | return 0; |
1e8f889b DG |
2544 | } |
2545 | ||
04f2cbe3 MG |
2546 | /* |
2547 | * If the process that created a MAP_PRIVATE mapping is about to | |
2548 | * perform a COW due to a shared page count, attempt to satisfy | |
2549 | * the allocation without using the existing reserves. The pagecache | |
2550 | * page is used to determine if the reserve at this address was | |
2551 | * consumed or not. If reserves were used, a partial faulted mapping | |
2552 | * at the time of fork() could consume its reserves on COW instead | |
2553 | * of the full address range. | |
2554 | */ | |
f83a275d | 2555 | if (!(vma->vm_flags & VM_MAYSHARE) && |
04f2cbe3 MG |
2556 | is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
2557 | old_page != pagecache_page) | |
2558 | outside_reserve = 1; | |
2559 | ||
1e8f889b | 2560 | page_cache_get(old_page); |
b76c8cfb LW |
2561 | |
2562 | /* Drop page_table_lock as buddy allocator may be called */ | |
2563 | spin_unlock(&mm->page_table_lock); | |
04f2cbe3 | 2564 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 2565 | |
2fc39cec | 2566 | if (IS_ERR(new_page)) { |
76dcee75 | 2567 | long err = PTR_ERR(new_page); |
1e8f889b | 2568 | page_cache_release(old_page); |
04f2cbe3 MG |
2569 | |
2570 | /* | |
2571 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
2572 | * it is due to references held by a child and an insufficient | |
2573 | * huge page pool. To guarantee the original mappers | |
2574 | * reliability, unmap the page from child processes. The child | |
2575 | * may get SIGKILLed if it later faults. | |
2576 | */ | |
2577 | if (outside_reserve) { | |
2578 | BUG_ON(huge_pte_none(pte)); | |
2579 | if (unmap_ref_private(mm, vma, old_page, address)) { | |
04f2cbe3 | 2580 | BUG_ON(huge_pte_none(pte)); |
b76c8cfb | 2581 | spin_lock(&mm->page_table_lock); |
a734bcc8 HD |
2582 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
2583 | if (likely(pte_same(huge_ptep_get(ptep), pte))) | |
2584 | goto retry_avoidcopy; | |
2585 | /* | |
2586 | * race occurs while re-acquiring page_table_lock, and | |
2587 | * our job is done. | |
2588 | */ | |
2589 | return 0; | |
04f2cbe3 MG |
2590 | } |
2591 | WARN_ON_ONCE(1); | |
2592 | } | |
2593 | ||
b76c8cfb LW |
2594 | /* Caller expects lock to be held */ |
2595 | spin_lock(&mm->page_table_lock); | |
76dcee75 AK |
2596 | if (err == -ENOMEM) |
2597 | return VM_FAULT_OOM; | |
2598 | else | |
2599 | return VM_FAULT_SIGBUS; | |
1e8f889b DG |
2600 | } |
2601 | ||
0fe6e20b NH |
2602 | /* |
2603 | * When the original hugepage is shared one, it does not have | |
2604 | * anon_vma prepared. | |
2605 | */ | |
44e2aa93 | 2606 | if (unlikely(anon_vma_prepare(vma))) { |
ea4039a3 HD |
2607 | page_cache_release(new_page); |
2608 | page_cache_release(old_page); | |
44e2aa93 DN |
2609 | /* Caller expects lock to be held */ |
2610 | spin_lock(&mm->page_table_lock); | |
0fe6e20b | 2611 | return VM_FAULT_OOM; |
44e2aa93 | 2612 | } |
0fe6e20b | 2613 | |
47ad8475 AA |
2614 | copy_user_huge_page(new_page, old_page, address, vma, |
2615 | pages_per_huge_page(h)); | |
0ed361de | 2616 | __SetPageUptodate(new_page); |
1e8f889b | 2617 | |
2ec74c3e SG |
2618 | mmun_start = address & huge_page_mask(h); |
2619 | mmun_end = mmun_start + huge_page_size(h); | |
2620 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
b76c8cfb LW |
2621 | /* |
2622 | * Retake the page_table_lock to check for racing updates | |
2623 | * before the page tables are altered | |
2624 | */ | |
2625 | spin_lock(&mm->page_table_lock); | |
a5516438 | 2626 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
7f2e9525 | 2627 | if (likely(pte_same(huge_ptep_get(ptep), pte))) { |
1e8f889b | 2628 | /* Break COW */ |
8fe627ec | 2629 | huge_ptep_clear_flush(vma, address, ptep); |
1e8f889b DG |
2630 | set_huge_pte_at(mm, address, ptep, |
2631 | make_huge_pte(vma, new_page, 1)); | |
0fe6e20b | 2632 | page_remove_rmap(old_page); |
cd67f0d2 | 2633 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
2634 | /* Make the old page be freed below */ |
2635 | new_page = old_page; | |
2636 | } | |
2ec74c3e SG |
2637 | spin_unlock(&mm->page_table_lock); |
2638 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); | |
2639 | /* Caller expects lock to be held */ | |
2640 | spin_lock(&mm->page_table_lock); | |
1e8f889b DG |
2641 | page_cache_release(new_page); |
2642 | page_cache_release(old_page); | |
83c54070 | 2643 | return 0; |
1e8f889b DG |
2644 | } |
2645 | ||
04f2cbe3 | 2646 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
2647 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
2648 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
2649 | { |
2650 | struct address_space *mapping; | |
e7c4b0bf | 2651 | pgoff_t idx; |
04f2cbe3 MG |
2652 | |
2653 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 2654 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
2655 | |
2656 | return find_lock_page(mapping, idx); | |
2657 | } | |
2658 | ||
3ae77f43 HD |
2659 | /* |
2660 | * Return whether there is a pagecache page to back given address within VMA. | |
2661 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
2662 | */ | |
2663 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
2664 | struct vm_area_struct *vma, unsigned long address) |
2665 | { | |
2666 | struct address_space *mapping; | |
2667 | pgoff_t idx; | |
2668 | struct page *page; | |
2669 | ||
2670 | mapping = vma->vm_file->f_mapping; | |
2671 | idx = vma_hugecache_offset(h, vma, address); | |
2672 | ||
2673 | page = find_get_page(mapping, idx); | |
2674 | if (page) | |
2675 | put_page(page); | |
2676 | return page != NULL; | |
2677 | } | |
2678 | ||
a1ed3dda | 2679 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 2680 | unsigned long address, pte_t *ptep, unsigned int flags) |
ac9b9c66 | 2681 | { |
a5516438 | 2682 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 2683 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 2684 | int anon_rmap = 0; |
e7c4b0bf | 2685 | pgoff_t idx; |
4c887265 | 2686 | unsigned long size; |
4c887265 AL |
2687 | struct page *page; |
2688 | struct address_space *mapping; | |
1e8f889b | 2689 | pte_t new_pte; |
4c887265 | 2690 | |
04f2cbe3 MG |
2691 | /* |
2692 | * Currently, we are forced to kill the process in the event the | |
2693 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 2694 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
2695 | */ |
2696 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
2697 | printk(KERN_WARNING | |
2698 | "PID %d killed due to inadequate hugepage pool\n", | |
2699 | current->pid); | |
2700 | return ret; | |
2701 | } | |
2702 | ||
4c887265 | 2703 | mapping = vma->vm_file->f_mapping; |
a5516438 | 2704 | idx = vma_hugecache_offset(h, vma, address); |
4c887265 AL |
2705 | |
2706 | /* | |
2707 | * Use page lock to guard against racing truncation | |
2708 | * before we get page_table_lock. | |
2709 | */ | |
6bda666a CL |
2710 | retry: |
2711 | page = find_lock_page(mapping, idx); | |
2712 | if (!page) { | |
a5516438 | 2713 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
2714 | if (idx >= size) |
2715 | goto out; | |
04f2cbe3 | 2716 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 2717 | if (IS_ERR(page)) { |
76dcee75 AK |
2718 | ret = PTR_ERR(page); |
2719 | if (ret == -ENOMEM) | |
2720 | ret = VM_FAULT_OOM; | |
2721 | else | |
2722 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
2723 | goto out; |
2724 | } | |
47ad8475 | 2725 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 2726 | __SetPageUptodate(page); |
ac9b9c66 | 2727 | |
f83a275d | 2728 | if (vma->vm_flags & VM_MAYSHARE) { |
6bda666a | 2729 | int err; |
45c682a6 | 2730 | struct inode *inode = mapping->host; |
6bda666a CL |
2731 | |
2732 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
2733 | if (err) { | |
2734 | put_page(page); | |
6bda666a CL |
2735 | if (err == -EEXIST) |
2736 | goto retry; | |
2737 | goto out; | |
2738 | } | |
45c682a6 KC |
2739 | |
2740 | spin_lock(&inode->i_lock); | |
a5516438 | 2741 | inode->i_blocks += blocks_per_huge_page(h); |
45c682a6 | 2742 | spin_unlock(&inode->i_lock); |
23be7468 | 2743 | } else { |
6bda666a | 2744 | lock_page(page); |
0fe6e20b NH |
2745 | if (unlikely(anon_vma_prepare(vma))) { |
2746 | ret = VM_FAULT_OOM; | |
2747 | goto backout_unlocked; | |
2748 | } | |
409eb8c2 | 2749 | anon_rmap = 1; |
23be7468 | 2750 | } |
0fe6e20b | 2751 | } else { |
998b4382 NH |
2752 | /* |
2753 | * If memory error occurs between mmap() and fault, some process | |
2754 | * don't have hwpoisoned swap entry for errored virtual address. | |
2755 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
2756 | */ | |
2757 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 2758 | ret = VM_FAULT_HWPOISON | |
972dc4de | 2759 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
2760 | goto backout_unlocked; |
2761 | } | |
6bda666a | 2762 | } |
1e8f889b | 2763 | |
57303d80 AW |
2764 | /* |
2765 | * If we are going to COW a private mapping later, we examine the | |
2766 | * pending reservations for this page now. This will ensure that | |
2767 | * any allocations necessary to record that reservation occur outside | |
2768 | * the spinlock. | |
2769 | */ | |
788c7df4 | 2770 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) |
2b26736c AW |
2771 | if (vma_needs_reservation(h, vma, address) < 0) { |
2772 | ret = VM_FAULT_OOM; | |
2773 | goto backout_unlocked; | |
2774 | } | |
57303d80 | 2775 | |
ac9b9c66 | 2776 | spin_lock(&mm->page_table_lock); |
a5516438 | 2777 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
2778 | if (idx >= size) |
2779 | goto backout; | |
2780 | ||
83c54070 | 2781 | ret = 0; |
7f2e9525 | 2782 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
2783 | goto backout; |
2784 | ||
409eb8c2 HD |
2785 | if (anon_rmap) |
2786 | hugepage_add_new_anon_rmap(page, vma, address); | |
2787 | else | |
2788 | page_dup_rmap(page); | |
1e8f889b DG |
2789 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
2790 | && (vma->vm_flags & VM_SHARED))); | |
2791 | set_huge_pte_at(mm, address, ptep, new_pte); | |
2792 | ||
788c7df4 | 2793 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 2794 | /* Optimization, do the COW without a second fault */ |
04f2cbe3 | 2795 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); |
1e8f889b DG |
2796 | } |
2797 | ||
ac9b9c66 | 2798 | spin_unlock(&mm->page_table_lock); |
4c887265 AL |
2799 | unlock_page(page); |
2800 | out: | |
ac9b9c66 | 2801 | return ret; |
4c887265 AL |
2802 | |
2803 | backout: | |
2804 | spin_unlock(&mm->page_table_lock); | |
2b26736c | 2805 | backout_unlocked: |
4c887265 AL |
2806 | unlock_page(page); |
2807 | put_page(page); | |
2808 | goto out; | |
ac9b9c66 HD |
2809 | } |
2810 | ||
86e5216f | 2811 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 2812 | unsigned long address, unsigned int flags) |
86e5216f AL |
2813 | { |
2814 | pte_t *ptep; | |
2815 | pte_t entry; | |
1e8f889b | 2816 | int ret; |
0fe6e20b | 2817 | struct page *page = NULL; |
57303d80 | 2818 | struct page *pagecache_page = NULL; |
3935baa9 | 2819 | static DEFINE_MUTEX(hugetlb_instantiation_mutex); |
a5516438 | 2820 | struct hstate *h = hstate_vma(vma); |
86e5216f | 2821 | |
1e16a539 KH |
2822 | address &= huge_page_mask(h); |
2823 | ||
fd6a03ed NH |
2824 | ptep = huge_pte_offset(mm, address); |
2825 | if (ptep) { | |
2826 | entry = huge_ptep_get(ptep); | |
290408d4 NH |
2827 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
2828 | migration_entry_wait(mm, (pmd_t *)ptep, address); | |
2829 | return 0; | |
2830 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 2831 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 2832 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
2833 | } |
2834 | ||
a5516438 | 2835 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); |
86e5216f AL |
2836 | if (!ptep) |
2837 | return VM_FAULT_OOM; | |
2838 | ||
3935baa9 DG |
2839 | /* |
2840 | * Serialize hugepage allocation and instantiation, so that we don't | |
2841 | * get spurious allocation failures if two CPUs race to instantiate | |
2842 | * the same page in the page cache. | |
2843 | */ | |
2844 | mutex_lock(&hugetlb_instantiation_mutex); | |
7f2e9525 GS |
2845 | entry = huge_ptep_get(ptep); |
2846 | if (huge_pte_none(entry)) { | |
788c7df4 | 2847 | ret = hugetlb_no_page(mm, vma, address, ptep, flags); |
b4d1d99f | 2848 | goto out_mutex; |
3935baa9 | 2849 | } |
86e5216f | 2850 | |
83c54070 | 2851 | ret = 0; |
1e8f889b | 2852 | |
57303d80 AW |
2853 | /* |
2854 | * If we are going to COW the mapping later, we examine the pending | |
2855 | * reservations for this page now. This will ensure that any | |
2856 | * allocations necessary to record that reservation occur outside the | |
2857 | * spinlock. For private mappings, we also lookup the pagecache | |
2858 | * page now as it is used to determine if a reservation has been | |
2859 | * consumed. | |
2860 | */ | |
788c7df4 | 2861 | if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) { |
2b26736c AW |
2862 | if (vma_needs_reservation(h, vma, address) < 0) { |
2863 | ret = VM_FAULT_OOM; | |
b4d1d99f | 2864 | goto out_mutex; |
2b26736c | 2865 | } |
57303d80 | 2866 | |
f83a275d | 2867 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
2868 | pagecache_page = hugetlbfs_pagecache_page(h, |
2869 | vma, address); | |
2870 | } | |
2871 | ||
56c9cfb1 NH |
2872 | /* |
2873 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
2874 | * pagecache_page, so here we need take the former one | |
2875 | * when page != pagecache_page or !pagecache_page. | |
2876 | * Note that locking order is always pagecache_page -> page, | |
2877 | * so no worry about deadlock. | |
2878 | */ | |
2879 | page = pte_page(entry); | |
66aebce7 | 2880 | get_page(page); |
56c9cfb1 | 2881 | if (page != pagecache_page) |
0fe6e20b | 2882 | lock_page(page); |
0fe6e20b | 2883 | |
1e8f889b DG |
2884 | spin_lock(&mm->page_table_lock); |
2885 | /* Check for a racing update before calling hugetlb_cow */ | |
b4d1d99f DG |
2886 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) |
2887 | goto out_page_table_lock; | |
2888 | ||
2889 | ||
788c7df4 | 2890 | if (flags & FAULT_FLAG_WRITE) { |
b4d1d99f | 2891 | if (!pte_write(entry)) { |
57303d80 AW |
2892 | ret = hugetlb_cow(mm, vma, address, ptep, entry, |
2893 | pagecache_page); | |
b4d1d99f DG |
2894 | goto out_page_table_lock; |
2895 | } | |
2896 | entry = pte_mkdirty(entry); | |
2897 | } | |
2898 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
2899 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
2900 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 2901 | update_mmu_cache(vma, address, ptep); |
b4d1d99f DG |
2902 | |
2903 | out_page_table_lock: | |
1e8f889b | 2904 | spin_unlock(&mm->page_table_lock); |
57303d80 AW |
2905 | |
2906 | if (pagecache_page) { | |
2907 | unlock_page(pagecache_page); | |
2908 | put_page(pagecache_page); | |
2909 | } | |
1f64d69c DN |
2910 | if (page != pagecache_page) |
2911 | unlock_page(page); | |
66aebce7 | 2912 | put_page(page); |
57303d80 | 2913 | |
b4d1d99f | 2914 | out_mutex: |
3935baa9 | 2915 | mutex_unlock(&hugetlb_instantiation_mutex); |
1e8f889b DG |
2916 | |
2917 | return ret; | |
86e5216f AL |
2918 | } |
2919 | ||
ceb86879 AK |
2920 | /* Can be overriden by architectures */ |
2921 | __attribute__((weak)) struct page * | |
2922 | follow_huge_pud(struct mm_struct *mm, unsigned long address, | |
2923 | pud_t *pud, int write) | |
2924 | { | |
2925 | BUG(); | |
2926 | return NULL; | |
2927 | } | |
2928 | ||
63551ae0 DG |
2929 | int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
2930 | struct page **pages, struct vm_area_struct **vmas, | |
5b23dbe8 | 2931 | unsigned long *position, int *length, int i, |
2a15efc9 | 2932 | unsigned int flags) |
63551ae0 | 2933 | { |
d5d4b0aa CK |
2934 | unsigned long pfn_offset; |
2935 | unsigned long vaddr = *position; | |
63551ae0 | 2936 | int remainder = *length; |
a5516438 | 2937 | struct hstate *h = hstate_vma(vma); |
63551ae0 | 2938 | |
1c59827d | 2939 | spin_lock(&mm->page_table_lock); |
63551ae0 | 2940 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 2941 | pte_t *pte; |
2a15efc9 | 2942 | int absent; |
4c887265 | 2943 | struct page *page; |
63551ae0 | 2944 | |
4c887265 AL |
2945 | /* |
2946 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 2947 | * each hugepage. We have to make sure we get the |
4c887265 AL |
2948 | * first, for the page indexing below to work. |
2949 | */ | |
a5516438 | 2950 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); |
2a15efc9 HD |
2951 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
2952 | ||
2953 | /* | |
2954 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
2955 | * an error where there's an empty slot with no huge pagecache |
2956 | * to back it. This way, we avoid allocating a hugepage, and | |
2957 | * the sparse dumpfile avoids allocating disk blocks, but its | |
2958 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 2959 | */ |
3ae77f43 HD |
2960 | if (absent && (flags & FOLL_DUMP) && |
2961 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
2a15efc9 HD |
2962 | remainder = 0; |
2963 | break; | |
2964 | } | |
63551ae0 | 2965 | |
2a15efc9 HD |
2966 | if (absent || |
2967 | ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 2968 | int ret; |
63551ae0 | 2969 | |
4c887265 | 2970 | spin_unlock(&mm->page_table_lock); |
2a15efc9 HD |
2971 | ret = hugetlb_fault(mm, vma, vaddr, |
2972 | (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); | |
4c887265 | 2973 | spin_lock(&mm->page_table_lock); |
a89182c7 | 2974 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 2975 | continue; |
63551ae0 | 2976 | |
4c887265 | 2977 | remainder = 0; |
4c887265 AL |
2978 | break; |
2979 | } | |
2980 | ||
a5516438 | 2981 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 2982 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 2983 | same_page: |
d6692183 | 2984 | if (pages) { |
2a15efc9 | 2985 | pages[i] = mem_map_offset(page, pfn_offset); |
4b2e38ad | 2986 | get_page(pages[i]); |
d6692183 | 2987 | } |
63551ae0 DG |
2988 | |
2989 | if (vmas) | |
2990 | vmas[i] = vma; | |
2991 | ||
2992 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 2993 | ++pfn_offset; |
63551ae0 DG |
2994 | --remainder; |
2995 | ++i; | |
d5d4b0aa | 2996 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 2997 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa CK |
2998 | /* |
2999 | * We use pfn_offset to avoid touching the pageframes | |
3000 | * of this compound page. | |
3001 | */ | |
3002 | goto same_page; | |
3003 | } | |
63551ae0 | 3004 | } |
1c59827d | 3005 | spin_unlock(&mm->page_table_lock); |
63551ae0 DG |
3006 | *length = remainder; |
3007 | *position = vaddr; | |
3008 | ||
2a15efc9 | 3009 | return i ? i : -EFAULT; |
63551ae0 | 3010 | } |
8f860591 ZY |
3011 | |
3012 | void hugetlb_change_protection(struct vm_area_struct *vma, | |
3013 | unsigned long address, unsigned long end, pgprot_t newprot) | |
3014 | { | |
3015 | struct mm_struct *mm = vma->vm_mm; | |
3016 | unsigned long start = address; | |
3017 | pte_t *ptep; | |
3018 | pte_t pte; | |
a5516438 | 3019 | struct hstate *h = hstate_vma(vma); |
8f860591 ZY |
3020 | |
3021 | BUG_ON(address >= end); | |
3022 | flush_cache_range(vma, address, end); | |
3023 | ||
3d48ae45 | 3024 | mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex); |
8f860591 | 3025 | spin_lock(&mm->page_table_lock); |
a5516438 | 3026 | for (; address < end; address += huge_page_size(h)) { |
8f860591 ZY |
3027 | ptep = huge_pte_offset(mm, address); |
3028 | if (!ptep) | |
3029 | continue; | |
39dde65c CK |
3030 | if (huge_pmd_unshare(mm, &address, ptep)) |
3031 | continue; | |
7f2e9525 | 3032 | if (!huge_pte_none(huge_ptep_get(ptep))) { |
8f860591 ZY |
3033 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
3034 | pte = pte_mkhuge(pte_modify(pte, newprot)); | |
3035 | set_huge_pte_at(mm, address, ptep, pte); | |
8f860591 ZY |
3036 | } |
3037 | } | |
3038 | spin_unlock(&mm->page_table_lock); | |
d833352a MG |
3039 | /* |
3040 | * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare | |
3041 | * may have cleared our pud entry and done put_page on the page table: | |
3042 | * once we release i_mmap_mutex, another task can do the final put_page | |
3043 | * and that page table be reused and filled with junk. | |
3044 | */ | |
8f860591 | 3045 | flush_tlb_range(vma, start, end); |
d833352a | 3046 | mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex); |
8f860591 ZY |
3047 | } |
3048 | ||
a1e78772 MG |
3049 | int hugetlb_reserve_pages(struct inode *inode, |
3050 | long from, long to, | |
5a6fe125 | 3051 | struct vm_area_struct *vma, |
ca16d140 | 3052 | vm_flags_t vm_flags) |
e4e574b7 | 3053 | { |
17c9d12e | 3054 | long ret, chg; |
a5516438 | 3055 | struct hstate *h = hstate_inode(inode); |
90481622 | 3056 | struct hugepage_subpool *spool = subpool_inode(inode); |
e4e574b7 | 3057 | |
17c9d12e MG |
3058 | /* |
3059 | * Only apply hugepage reservation if asked. At fault time, an | |
3060 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 3061 | * without using reserves |
17c9d12e | 3062 | */ |
ca16d140 | 3063 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
3064 | return 0; |
3065 | ||
a1e78772 MG |
3066 | /* |
3067 | * Shared mappings base their reservation on the number of pages that | |
3068 | * are already allocated on behalf of the file. Private mappings need | |
3069 | * to reserve the full area even if read-only as mprotect() may be | |
3070 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
3071 | */ | |
f83a275d | 3072 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
a1e78772 | 3073 | chg = region_chg(&inode->i_mapping->private_list, from, to); |
17c9d12e MG |
3074 | else { |
3075 | struct resv_map *resv_map = resv_map_alloc(); | |
3076 | if (!resv_map) | |
3077 | return -ENOMEM; | |
3078 | ||
a1e78772 | 3079 | chg = to - from; |
84afd99b | 3080 | |
17c9d12e MG |
3081 | set_vma_resv_map(vma, resv_map); |
3082 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
3083 | } | |
3084 | ||
c50ac050 DH |
3085 | if (chg < 0) { |
3086 | ret = chg; | |
3087 | goto out_err; | |
3088 | } | |
8a630112 | 3089 | |
90481622 | 3090 | /* There must be enough pages in the subpool for the mapping */ |
c50ac050 DH |
3091 | if (hugepage_subpool_get_pages(spool, chg)) { |
3092 | ret = -ENOSPC; | |
3093 | goto out_err; | |
3094 | } | |
5a6fe125 MG |
3095 | |
3096 | /* | |
17c9d12e | 3097 | * Check enough hugepages are available for the reservation. |
90481622 | 3098 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 3099 | */ |
a5516438 | 3100 | ret = hugetlb_acct_memory(h, chg); |
68842c9b | 3101 | if (ret < 0) { |
90481622 | 3102 | hugepage_subpool_put_pages(spool, chg); |
c50ac050 | 3103 | goto out_err; |
68842c9b | 3104 | } |
17c9d12e MG |
3105 | |
3106 | /* | |
3107 | * Account for the reservations made. Shared mappings record regions | |
3108 | * that have reservations as they are shared by multiple VMAs. | |
3109 | * When the last VMA disappears, the region map says how much | |
3110 | * the reservation was and the page cache tells how much of | |
3111 | * the reservation was consumed. Private mappings are per-VMA and | |
3112 | * only the consumed reservations are tracked. When the VMA | |
3113 | * disappears, the original reservation is the VMA size and the | |
3114 | * consumed reservations are stored in the map. Hence, nothing | |
3115 | * else has to be done for private mappings here | |
3116 | */ | |
f83a275d | 3117 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
a1e78772 | 3118 | region_add(&inode->i_mapping->private_list, from, to); |
a43a8c39 | 3119 | return 0; |
c50ac050 | 3120 | out_err: |
4523e145 DH |
3121 | if (vma) |
3122 | resv_map_put(vma); | |
c50ac050 | 3123 | return ret; |
a43a8c39 CK |
3124 | } |
3125 | ||
3126 | void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) | |
3127 | { | |
a5516438 | 3128 | struct hstate *h = hstate_inode(inode); |
a43a8c39 | 3129 | long chg = region_truncate(&inode->i_mapping->private_list, offset); |
90481622 | 3130 | struct hugepage_subpool *spool = subpool_inode(inode); |
45c682a6 KC |
3131 | |
3132 | spin_lock(&inode->i_lock); | |
e4c6f8be | 3133 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
3134 | spin_unlock(&inode->i_lock); |
3135 | ||
90481622 | 3136 | hugepage_subpool_put_pages(spool, (chg - freed)); |
a5516438 | 3137 | hugetlb_acct_memory(h, -(chg - freed)); |
a43a8c39 | 3138 | } |
93f70f90 | 3139 | |
d5bd9106 AK |
3140 | #ifdef CONFIG_MEMORY_FAILURE |
3141 | ||
6de2b1aa NH |
3142 | /* Should be called in hugetlb_lock */ |
3143 | static int is_hugepage_on_freelist(struct page *hpage) | |
3144 | { | |
3145 | struct page *page; | |
3146 | struct page *tmp; | |
3147 | struct hstate *h = page_hstate(hpage); | |
3148 | int nid = page_to_nid(hpage); | |
3149 | ||
3150 | list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru) | |
3151 | if (page == hpage) | |
3152 | return 1; | |
3153 | return 0; | |
3154 | } | |
3155 | ||
93f70f90 NH |
3156 | /* |
3157 | * This function is called from memory failure code. | |
3158 | * Assume the caller holds page lock of the head page. | |
3159 | */ | |
6de2b1aa | 3160 | int dequeue_hwpoisoned_huge_page(struct page *hpage) |
93f70f90 NH |
3161 | { |
3162 | struct hstate *h = page_hstate(hpage); | |
3163 | int nid = page_to_nid(hpage); | |
6de2b1aa | 3164 | int ret = -EBUSY; |
93f70f90 NH |
3165 | |
3166 | spin_lock(&hugetlb_lock); | |
6de2b1aa NH |
3167 | if (is_hugepage_on_freelist(hpage)) { |
3168 | list_del(&hpage->lru); | |
8c6c2ecb | 3169 | set_page_refcounted(hpage); |
6de2b1aa NH |
3170 | h->free_huge_pages--; |
3171 | h->free_huge_pages_node[nid]--; | |
3172 | ret = 0; | |
3173 | } | |
93f70f90 | 3174 | spin_unlock(&hugetlb_lock); |
6de2b1aa | 3175 | return ret; |
93f70f90 | 3176 | } |
6de2b1aa | 3177 | #endif |