]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * Generic hugetlb support. | |
6d49e352 | 3 | * (C) Nadia Yvette Chambers, April 2004 |
1da177e4 | 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 | { | |
496ad9aa | 130 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
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 | |
c748c262 | 138 | * and the hugetlb_instantiation_mutex. To access or modify a region the caller |
84afd99b | 139 | * must either hold the mmap_sem for write, or the mmap_sem for read and |
c748c262 | 140 | * the hugetlb_instantiation_mutex: |
84afd99b | 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 | ||
2415cf12 | 322 | return 1UL << huge_page_shift(hstate); |
08fba699 | 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 | ||
04f2cbe3 | 437 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
438 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
439 | { | |
440 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 441 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
442 | vma->vm_private_data = (void *)0; |
443 | } | |
444 | ||
445 | /* Returns true if the VMA has associated reserve pages */ | |
7f09ca51 | 446 | static int vma_has_reserves(struct vm_area_struct *vma) |
a1e78772 | 447 | { |
72231b03 JK |
448 | if (vma->vm_flags & VM_NORESERVE) |
449 | return 0; | |
a63884e9 JK |
450 | |
451 | /* Shared mappings always use reserves */ | |
f83a275d | 452 | if (vma->vm_flags & VM_MAYSHARE) |
7f09ca51 | 453 | return 1; |
a63884e9 JK |
454 | |
455 | /* | |
456 | * Only the process that called mmap() has reserves for | |
457 | * private mappings. | |
458 | */ | |
7f09ca51 MG |
459 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
460 | return 1; | |
a63884e9 | 461 | |
7f09ca51 | 462 | return 0; |
a1e78772 MG |
463 | } |
464 | ||
0ebabb41 NH |
465 | static void copy_gigantic_page(struct page *dst, struct page *src) |
466 | { | |
467 | int i; | |
468 | struct hstate *h = page_hstate(src); | |
469 | struct page *dst_base = dst; | |
470 | struct page *src_base = src; | |
471 | ||
472 | for (i = 0; i < pages_per_huge_page(h); ) { | |
473 | cond_resched(); | |
474 | copy_highpage(dst, src); | |
475 | ||
476 | i++; | |
477 | dst = mem_map_next(dst, dst_base, i); | |
478 | src = mem_map_next(src, src_base, i); | |
479 | } | |
480 | } | |
481 | ||
482 | void copy_huge_page(struct page *dst, struct page *src) | |
483 | { | |
484 | int i; | |
485 | struct hstate *h = page_hstate(src); | |
486 | ||
487 | if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) { | |
488 | copy_gigantic_page(dst, src); | |
489 | return; | |
490 | } | |
491 | ||
492 | might_sleep(); | |
493 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
494 | cond_resched(); | |
495 | copy_highpage(dst + i, src + i); | |
496 | } | |
497 | } | |
498 | ||
a5516438 | 499 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
500 | { |
501 | int nid = page_to_nid(page); | |
0edaecfa | 502 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
503 | h->free_huge_pages++; |
504 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
505 | } |
506 | ||
bf50bab2 NH |
507 | static struct page *dequeue_huge_page_node(struct hstate *h, int nid) |
508 | { | |
509 | struct page *page; | |
510 | ||
511 | if (list_empty(&h->hugepage_freelists[nid])) | |
512 | return NULL; | |
513 | page = list_entry(h->hugepage_freelists[nid].next, struct page, lru); | |
0edaecfa | 514 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 515 | set_page_refcounted(page); |
bf50bab2 NH |
516 | h->free_huge_pages--; |
517 | h->free_huge_pages_node[nid]--; | |
518 | return page; | |
519 | } | |
520 | ||
a5516438 AK |
521 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
522 | struct vm_area_struct *vma, | |
04f2cbe3 | 523 | unsigned long address, int avoid_reserve) |
1da177e4 | 524 | { |
b1c12cbc | 525 | struct page *page = NULL; |
480eccf9 | 526 | struct mempolicy *mpol; |
19770b32 | 527 | nodemask_t *nodemask; |
c0ff7453 | 528 | struct zonelist *zonelist; |
dd1a239f MG |
529 | struct zone *zone; |
530 | struct zoneref *z; | |
cc9a6c87 | 531 | unsigned int cpuset_mems_cookie; |
1da177e4 | 532 | |
a1e78772 MG |
533 | /* |
534 | * A child process with MAP_PRIVATE mappings created by their parent | |
535 | * have no page reserves. This check ensures that reservations are | |
536 | * not "stolen". The child may still get SIGKILLed | |
537 | */ | |
7f09ca51 | 538 | if (!vma_has_reserves(vma) && |
a5516438 | 539 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 540 | goto err; |
a1e78772 | 541 | |
04f2cbe3 | 542 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 543 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 544 | goto err; |
04f2cbe3 | 545 | |
9966c4bb JK |
546 | retry_cpuset: |
547 | cpuset_mems_cookie = get_mems_allowed(); | |
548 | zonelist = huge_zonelist(vma, address, | |
549 | htlb_alloc_mask, &mpol, &nodemask); | |
550 | ||
19770b32 MG |
551 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
552 | MAX_NR_ZONES - 1, nodemask) { | |
bf50bab2 NH |
553 | if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) { |
554 | page = dequeue_huge_page_node(h, zone_to_nid(zone)); | |
555 | if (page) { | |
a63884e9 JK |
556 | if (!avoid_reserve && vma_has_reserves(vma)) |
557 | h->resv_huge_pages--; | |
bf50bab2 NH |
558 | break; |
559 | } | |
3abf7afd | 560 | } |
1da177e4 | 561 | } |
cc9a6c87 | 562 | |
52cd3b07 | 563 | mpol_cond_put(mpol); |
cc9a6c87 MG |
564 | if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page)) |
565 | goto retry_cpuset; | |
1da177e4 | 566 | return page; |
cc9a6c87 MG |
567 | |
568 | err: | |
cc9a6c87 | 569 | return NULL; |
1da177e4 LT |
570 | } |
571 | ||
a5516438 | 572 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
573 | { |
574 | int i; | |
a5516438 | 575 | |
18229df5 AW |
576 | VM_BUG_ON(h->order >= MAX_ORDER); |
577 | ||
a5516438 AK |
578 | h->nr_huge_pages--; |
579 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
580 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
581 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
582 | 1 << PG_referenced | 1 << PG_dirty | | |
583 | 1 << PG_active | 1 << PG_reserved | | |
584 | 1 << PG_private | 1 << PG_writeback); | |
6af2acb6 | 585 | } |
9dd540e2 | 586 | VM_BUG_ON(hugetlb_cgroup_from_page(page)); |
6af2acb6 AL |
587 | set_compound_page_dtor(page, NULL); |
588 | set_page_refcounted(page); | |
7f2e9525 | 589 | arch_release_hugepage(page); |
a5516438 | 590 | __free_pages(page, huge_page_order(h)); |
6af2acb6 AL |
591 | } |
592 | ||
e5ff2159 AK |
593 | struct hstate *size_to_hstate(unsigned long size) |
594 | { | |
595 | struct hstate *h; | |
596 | ||
597 | for_each_hstate(h) { | |
598 | if (huge_page_size(h) == size) | |
599 | return h; | |
600 | } | |
601 | return NULL; | |
602 | } | |
603 | ||
27a85ef1 DG |
604 | static void free_huge_page(struct page *page) |
605 | { | |
a5516438 AK |
606 | /* |
607 | * Can't pass hstate in here because it is called from the | |
608 | * compound page destructor. | |
609 | */ | |
e5ff2159 | 610 | struct hstate *h = page_hstate(page); |
7893d1d5 | 611 | int nid = page_to_nid(page); |
90481622 DG |
612 | struct hugepage_subpool *spool = |
613 | (struct hugepage_subpool *)page_private(page); | |
27a85ef1 | 614 | |
e5df70ab | 615 | set_page_private(page, 0); |
23be7468 | 616 | page->mapping = NULL; |
7893d1d5 | 617 | BUG_ON(page_count(page)); |
0fe6e20b | 618 | BUG_ON(page_mapcount(page)); |
27a85ef1 DG |
619 | |
620 | spin_lock(&hugetlb_lock); | |
6d76dcf4 AK |
621 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
622 | pages_per_huge_page(h), page); | |
aa888a74 | 623 | if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { |
0edaecfa AK |
624 | /* remove the page from active list */ |
625 | list_del(&page->lru); | |
a5516438 AK |
626 | update_and_free_page(h, page); |
627 | h->surplus_huge_pages--; | |
628 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 629 | } else { |
5d3a551c | 630 | arch_clear_hugepage_flags(page); |
a5516438 | 631 | enqueue_huge_page(h, page); |
7893d1d5 | 632 | } |
27a85ef1 | 633 | spin_unlock(&hugetlb_lock); |
90481622 | 634 | hugepage_subpool_put_pages(spool, 1); |
27a85ef1 DG |
635 | } |
636 | ||
a5516438 | 637 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 638 | { |
0edaecfa | 639 | INIT_LIST_HEAD(&page->lru); |
b7ba30c6 AK |
640 | set_compound_page_dtor(page, free_huge_page); |
641 | spin_lock(&hugetlb_lock); | |
9dd540e2 | 642 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
643 | h->nr_huge_pages++; |
644 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 AK |
645 | spin_unlock(&hugetlb_lock); |
646 | put_page(page); /* free it into the hugepage allocator */ | |
647 | } | |
648 | ||
20a0307c WF |
649 | static void prep_compound_gigantic_page(struct page *page, unsigned long order) |
650 | { | |
651 | int i; | |
652 | int nr_pages = 1 << order; | |
653 | struct page *p = page + 1; | |
654 | ||
655 | /* we rely on prep_new_huge_page to set the destructor */ | |
656 | set_compound_order(page, order); | |
657 | __SetPageHead(page); | |
658 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { | |
659 | __SetPageTail(p); | |
58a84aa9 | 660 | set_page_count(p, 0); |
20a0307c WF |
661 | p->first_page = page; |
662 | } | |
663 | } | |
664 | ||
7795912c AM |
665 | /* |
666 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
667 | * transparent huge pages. See the PageTransHuge() documentation for more | |
668 | * details. | |
669 | */ | |
20a0307c WF |
670 | int PageHuge(struct page *page) |
671 | { | |
672 | compound_page_dtor *dtor; | |
673 | ||
674 | if (!PageCompound(page)) | |
675 | return 0; | |
676 | ||
677 | page = compound_head(page); | |
678 | dtor = get_compound_page_dtor(page); | |
679 | ||
680 | return dtor == free_huge_page; | |
681 | } | |
43131e14 NH |
682 | EXPORT_SYMBOL_GPL(PageHuge); |
683 | ||
13d60f4b ZY |
684 | pgoff_t __basepage_index(struct page *page) |
685 | { | |
686 | struct page *page_head = compound_head(page); | |
687 | pgoff_t index = page_index(page_head); | |
688 | unsigned long compound_idx; | |
689 | ||
690 | if (!PageHuge(page_head)) | |
691 | return page_index(page); | |
692 | ||
693 | if (compound_order(page_head) >= MAX_ORDER) | |
694 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
695 | else | |
696 | compound_idx = page - page_head; | |
697 | ||
698 | return (index << compound_order(page_head)) + compound_idx; | |
699 | } | |
700 | ||
a5516438 | 701 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
1da177e4 | 702 | { |
1da177e4 | 703 | struct page *page; |
f96efd58 | 704 | |
aa888a74 AK |
705 | if (h->order >= MAX_ORDER) |
706 | return NULL; | |
707 | ||
6484eb3e | 708 | page = alloc_pages_exact_node(nid, |
551883ae NA |
709 | htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| |
710 | __GFP_REPEAT|__GFP_NOWARN, | |
a5516438 | 711 | huge_page_order(h)); |
1da177e4 | 712 | if (page) { |
7f2e9525 | 713 | if (arch_prepare_hugepage(page)) { |
caff3a2c | 714 | __free_pages(page, huge_page_order(h)); |
7b8ee84d | 715 | return NULL; |
7f2e9525 | 716 | } |
a5516438 | 717 | prep_new_huge_page(h, page, nid); |
1da177e4 | 718 | } |
63b4613c NA |
719 | |
720 | return page; | |
721 | } | |
722 | ||
9a76db09 | 723 | /* |
6ae11b27 LS |
724 | * common helper functions for hstate_next_node_to_{alloc|free}. |
725 | * We may have allocated or freed a huge page based on a different | |
726 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
727 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
728 | * node for alloc or free. | |
9a76db09 | 729 | */ |
6ae11b27 | 730 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) |
9a76db09 | 731 | { |
6ae11b27 | 732 | nid = next_node(nid, *nodes_allowed); |
9a76db09 | 733 | if (nid == MAX_NUMNODES) |
6ae11b27 | 734 | nid = first_node(*nodes_allowed); |
9a76db09 LS |
735 | VM_BUG_ON(nid >= MAX_NUMNODES); |
736 | ||
737 | return nid; | |
738 | } | |
739 | ||
6ae11b27 LS |
740 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) |
741 | { | |
742 | if (!node_isset(nid, *nodes_allowed)) | |
743 | nid = next_node_allowed(nid, nodes_allowed); | |
744 | return nid; | |
745 | } | |
746 | ||
5ced66c9 | 747 | /* |
6ae11b27 LS |
748 | * returns the previously saved node ["this node"] from which to |
749 | * allocate a persistent huge page for the pool and advance the | |
750 | * next node from which to allocate, handling wrap at end of node | |
751 | * mask. | |
5ced66c9 | 752 | */ |
6ae11b27 LS |
753 | static int hstate_next_node_to_alloc(struct hstate *h, |
754 | nodemask_t *nodes_allowed) | |
5ced66c9 | 755 | { |
6ae11b27 LS |
756 | int nid; |
757 | ||
758 | VM_BUG_ON(!nodes_allowed); | |
759 | ||
760 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
761 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
9a76db09 | 762 | |
9a76db09 | 763 | return nid; |
5ced66c9 AK |
764 | } |
765 | ||
e8c5c824 | 766 | /* |
6ae11b27 LS |
767 | * helper for free_pool_huge_page() - return the previously saved |
768 | * node ["this node"] from which to free a huge page. Advance the | |
769 | * next node id whether or not we find a free huge page to free so | |
770 | * that the next attempt to free addresses the next node. | |
e8c5c824 | 771 | */ |
6ae11b27 | 772 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) |
e8c5c824 | 773 | { |
6ae11b27 LS |
774 | int nid; |
775 | ||
776 | VM_BUG_ON(!nodes_allowed); | |
777 | ||
778 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
779 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
9a76db09 | 780 | |
9a76db09 | 781 | return nid; |
e8c5c824 LS |
782 | } |
783 | ||
b2261026 JK |
784 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ |
785 | for (nr_nodes = nodes_weight(*mask); \ | |
786 | nr_nodes > 0 && \ | |
787 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
788 | nr_nodes--) | |
789 | ||
790 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
791 | for (nr_nodes = nodes_weight(*mask); \ | |
792 | nr_nodes > 0 && \ | |
793 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
794 | nr_nodes--) | |
795 | ||
796 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) | |
797 | { | |
798 | struct page *page; | |
799 | int nr_nodes, node; | |
800 | int ret = 0; | |
801 | ||
802 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
803 | page = alloc_fresh_huge_page_node(h, node); | |
804 | if (page) { | |
805 | ret = 1; | |
806 | break; | |
807 | } | |
808 | } | |
809 | ||
810 | if (ret) | |
811 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
812 | else | |
813 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
814 | ||
815 | return ret; | |
816 | } | |
817 | ||
e8c5c824 LS |
818 | /* |
819 | * Free huge page from pool from next node to free. | |
820 | * Attempt to keep persistent huge pages more or less | |
821 | * balanced over allowed nodes. | |
822 | * Called with hugetlb_lock locked. | |
823 | */ | |
6ae11b27 LS |
824 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
825 | bool acct_surplus) | |
e8c5c824 | 826 | { |
b2261026 | 827 | int nr_nodes, node; |
e8c5c824 LS |
828 | int ret = 0; |
829 | ||
b2261026 | 830 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
831 | /* |
832 | * If we're returning unused surplus pages, only examine | |
833 | * nodes with surplus pages. | |
834 | */ | |
b2261026 JK |
835 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
836 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 837 | struct page *page = |
b2261026 | 838 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
839 | struct page, lru); |
840 | list_del(&page->lru); | |
841 | h->free_huge_pages--; | |
b2261026 | 842 | h->free_huge_pages_node[node]--; |
685f3457 LS |
843 | if (acct_surplus) { |
844 | h->surplus_huge_pages--; | |
b2261026 | 845 | h->surplus_huge_pages_node[node]--; |
685f3457 | 846 | } |
e8c5c824 LS |
847 | update_and_free_page(h, page); |
848 | ret = 1; | |
9a76db09 | 849 | break; |
e8c5c824 | 850 | } |
b2261026 | 851 | } |
e8c5c824 LS |
852 | |
853 | return ret; | |
854 | } | |
855 | ||
bf50bab2 | 856 | static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) |
7893d1d5 AL |
857 | { |
858 | struct page *page; | |
bf50bab2 | 859 | unsigned int r_nid; |
7893d1d5 | 860 | |
aa888a74 AK |
861 | if (h->order >= MAX_ORDER) |
862 | return NULL; | |
863 | ||
d1c3fb1f NA |
864 | /* |
865 | * Assume we will successfully allocate the surplus page to | |
866 | * prevent racing processes from causing the surplus to exceed | |
867 | * overcommit | |
868 | * | |
869 | * This however introduces a different race, where a process B | |
870 | * tries to grow the static hugepage pool while alloc_pages() is | |
871 | * called by process A. B will only examine the per-node | |
872 | * counters in determining if surplus huge pages can be | |
873 | * converted to normal huge pages in adjust_pool_surplus(). A | |
874 | * won't be able to increment the per-node counter, until the | |
875 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
876 | * no more huge pages can be converted from surplus to normal | |
877 | * state (and doesn't try to convert again). Thus, we have a | |
878 | * case where a surplus huge page exists, the pool is grown, and | |
879 | * the surplus huge page still exists after, even though it | |
880 | * should just have been converted to a normal huge page. This | |
881 | * does not leak memory, though, as the hugepage will be freed | |
882 | * once it is out of use. It also does not allow the counters to | |
883 | * go out of whack in adjust_pool_surplus() as we don't modify | |
884 | * the node values until we've gotten the hugepage and only the | |
885 | * per-node value is checked there. | |
886 | */ | |
887 | spin_lock(&hugetlb_lock); | |
a5516438 | 888 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
889 | spin_unlock(&hugetlb_lock); |
890 | return NULL; | |
891 | } else { | |
a5516438 AK |
892 | h->nr_huge_pages++; |
893 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
894 | } |
895 | spin_unlock(&hugetlb_lock); | |
896 | ||
bf50bab2 NH |
897 | if (nid == NUMA_NO_NODE) |
898 | page = alloc_pages(htlb_alloc_mask|__GFP_COMP| | |
899 | __GFP_REPEAT|__GFP_NOWARN, | |
900 | huge_page_order(h)); | |
901 | else | |
902 | page = alloc_pages_exact_node(nid, | |
903 | htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| | |
904 | __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); | |
d1c3fb1f | 905 | |
caff3a2c GS |
906 | if (page && arch_prepare_hugepage(page)) { |
907 | __free_pages(page, huge_page_order(h)); | |
ea5768c7 | 908 | page = NULL; |
caff3a2c GS |
909 | } |
910 | ||
d1c3fb1f | 911 | spin_lock(&hugetlb_lock); |
7893d1d5 | 912 | if (page) { |
0edaecfa | 913 | INIT_LIST_HEAD(&page->lru); |
bf50bab2 | 914 | r_nid = page_to_nid(page); |
7893d1d5 | 915 | set_compound_page_dtor(page, free_huge_page); |
9dd540e2 | 916 | set_hugetlb_cgroup(page, NULL); |
d1c3fb1f NA |
917 | /* |
918 | * We incremented the global counters already | |
919 | */ | |
bf50bab2 NH |
920 | h->nr_huge_pages_node[r_nid]++; |
921 | h->surplus_huge_pages_node[r_nid]++; | |
3b116300 | 922 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 923 | } else { |
a5516438 AK |
924 | h->nr_huge_pages--; |
925 | h->surplus_huge_pages--; | |
3b116300 | 926 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 927 | } |
d1c3fb1f | 928 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
929 | |
930 | return page; | |
931 | } | |
932 | ||
bf50bab2 NH |
933 | /* |
934 | * This allocation function is useful in the context where vma is irrelevant. | |
935 | * E.g. soft-offlining uses this function because it only cares physical | |
936 | * address of error page. | |
937 | */ | |
938 | struct page *alloc_huge_page_node(struct hstate *h, int nid) | |
939 | { | |
940 | struct page *page; | |
941 | ||
942 | spin_lock(&hugetlb_lock); | |
943 | page = dequeue_huge_page_node(h, nid); | |
944 | spin_unlock(&hugetlb_lock); | |
945 | ||
94ae8ba7 | 946 | if (!page) |
bf50bab2 NH |
947 | page = alloc_buddy_huge_page(h, nid); |
948 | ||
949 | return page; | |
950 | } | |
951 | ||
e4e574b7 | 952 | /* |
25985edc | 953 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
954 | * of size 'delta'. |
955 | */ | |
a5516438 | 956 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
957 | { |
958 | struct list_head surplus_list; | |
959 | struct page *page, *tmp; | |
960 | int ret, i; | |
961 | int needed, allocated; | |
28073b02 | 962 | bool alloc_ok = true; |
e4e574b7 | 963 | |
a5516438 | 964 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 965 | if (needed <= 0) { |
a5516438 | 966 | h->resv_huge_pages += delta; |
e4e574b7 | 967 | return 0; |
ac09b3a1 | 968 | } |
e4e574b7 AL |
969 | |
970 | allocated = 0; | |
971 | INIT_LIST_HEAD(&surplus_list); | |
972 | ||
973 | ret = -ENOMEM; | |
974 | retry: | |
975 | spin_unlock(&hugetlb_lock); | |
976 | for (i = 0; i < needed; i++) { | |
bf50bab2 | 977 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
28073b02 HD |
978 | if (!page) { |
979 | alloc_ok = false; | |
980 | break; | |
981 | } | |
e4e574b7 AL |
982 | list_add(&page->lru, &surplus_list); |
983 | } | |
28073b02 | 984 | allocated += i; |
e4e574b7 AL |
985 | |
986 | /* | |
987 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
988 | * because either resv_huge_pages or free_huge_pages may have changed. | |
989 | */ | |
990 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
991 | needed = (h->resv_huge_pages + delta) - |
992 | (h->free_huge_pages + allocated); | |
28073b02 HD |
993 | if (needed > 0) { |
994 | if (alloc_ok) | |
995 | goto retry; | |
996 | /* | |
997 | * We were not able to allocate enough pages to | |
998 | * satisfy the entire reservation so we free what | |
999 | * we've allocated so far. | |
1000 | */ | |
1001 | goto free; | |
1002 | } | |
e4e574b7 AL |
1003 | /* |
1004 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1005 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1006 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1007 | * allocator. Commit the entire reservation here to prevent another |
1008 | * process from stealing the pages as they are added to the pool but | |
1009 | * before they are reserved. | |
e4e574b7 AL |
1010 | */ |
1011 | needed += allocated; | |
a5516438 | 1012 | h->resv_huge_pages += delta; |
e4e574b7 | 1013 | ret = 0; |
a9869b83 | 1014 | |
19fc3f0a | 1015 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1016 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1017 | if ((--needed) < 0) |
1018 | break; | |
a9869b83 NH |
1019 | /* |
1020 | * This page is now managed by the hugetlb allocator and has | |
1021 | * no users -- drop the buddy allocator's reference. | |
1022 | */ | |
1023 | put_page_testzero(page); | |
1024 | VM_BUG_ON(page_count(page)); | |
a5516438 | 1025 | enqueue_huge_page(h, page); |
19fc3f0a | 1026 | } |
28073b02 | 1027 | free: |
b0365c8d | 1028 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1029 | |
1030 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1031 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1032 | put_page(page); | |
a9869b83 | 1033 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1034 | |
1035 | return ret; | |
1036 | } | |
1037 | ||
1038 | /* | |
1039 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
1040 | * allocated to satisfy the reservation must be explicitly freed if they were | |
1041 | * never used. | |
685f3457 | 1042 | * Called with hugetlb_lock held. |
e4e574b7 | 1043 | */ |
a5516438 AK |
1044 | static void return_unused_surplus_pages(struct hstate *h, |
1045 | unsigned long unused_resv_pages) | |
e4e574b7 | 1046 | { |
e4e574b7 AL |
1047 | unsigned long nr_pages; |
1048 | ||
ac09b3a1 | 1049 | /* Uncommit the reservation */ |
a5516438 | 1050 | h->resv_huge_pages -= unused_resv_pages; |
ac09b3a1 | 1051 | |
aa888a74 AK |
1052 | /* Cannot return gigantic pages currently */ |
1053 | if (h->order >= MAX_ORDER) | |
1054 | return; | |
1055 | ||
a5516438 | 1056 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1057 | |
685f3457 LS |
1058 | /* |
1059 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1060 | * evenly across all nodes with memory. Iterate across these nodes |
1061 | * until we can no longer free unreserved surplus pages. This occurs | |
1062 | * when the nodes with surplus pages have no free pages. | |
1063 | * free_pool_huge_page() will balance the the freed pages across the | |
1064 | * on-line nodes with memory and will handle the hstate accounting. | |
685f3457 LS |
1065 | */ |
1066 | while (nr_pages--) { | |
8cebfcd0 | 1067 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
685f3457 | 1068 | break; |
e4e574b7 AL |
1069 | } |
1070 | } | |
1071 | ||
c37f9fb1 AW |
1072 | /* |
1073 | * Determine if the huge page at addr within the vma has an associated | |
1074 | * reservation. Where it does not we will need to logically increase | |
90481622 DG |
1075 | * reservation and actually increase subpool usage before an allocation |
1076 | * can occur. Where any new reservation would be required the | |
1077 | * reservation change is prepared, but not committed. Once the page | |
1078 | * has been allocated from the subpool and instantiated the change should | |
1079 | * be committed via vma_commit_reservation. No action is required on | |
1080 | * failure. | |
c37f9fb1 | 1081 | */ |
e2f17d94 | 1082 | static long vma_needs_reservation(struct hstate *h, |
a5516438 | 1083 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 AW |
1084 | { |
1085 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1086 | struct inode *inode = mapping->host; | |
1087 | ||
f83a275d | 1088 | if (vma->vm_flags & VM_MAYSHARE) { |
a5516438 | 1089 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
c37f9fb1 AW |
1090 | return region_chg(&inode->i_mapping->private_list, |
1091 | idx, idx + 1); | |
1092 | ||
84afd99b AW |
1093 | } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1094 | return 1; | |
c37f9fb1 | 1095 | |
84afd99b | 1096 | } else { |
e2f17d94 | 1097 | long err; |
a5516438 | 1098 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
84afd99b AW |
1099 | struct resv_map *reservations = vma_resv_map(vma); |
1100 | ||
1101 | err = region_chg(&reservations->regions, idx, idx + 1); | |
1102 | if (err < 0) | |
1103 | return err; | |
1104 | return 0; | |
1105 | } | |
c37f9fb1 | 1106 | } |
a5516438 AK |
1107 | static void vma_commit_reservation(struct hstate *h, |
1108 | struct vm_area_struct *vma, unsigned long addr) | |
c37f9fb1 AW |
1109 | { |
1110 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1111 | struct inode *inode = mapping->host; | |
1112 | ||
f83a275d | 1113 | if (vma->vm_flags & VM_MAYSHARE) { |
a5516438 | 1114 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
c37f9fb1 | 1115 | region_add(&inode->i_mapping->private_list, idx, idx + 1); |
84afd99b AW |
1116 | |
1117 | } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { | |
a5516438 | 1118 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
84afd99b AW |
1119 | struct resv_map *reservations = vma_resv_map(vma); |
1120 | ||
1121 | /* Mark this page used in the map. */ | |
1122 | region_add(&reservations->regions, idx, idx + 1); | |
c37f9fb1 AW |
1123 | } |
1124 | } | |
1125 | ||
a1e78772 | 1126 | static struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 1127 | unsigned long addr, int avoid_reserve) |
1da177e4 | 1128 | { |
90481622 | 1129 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 1130 | struct hstate *h = hstate_vma(vma); |
348ea204 | 1131 | struct page *page; |
e2f17d94 | 1132 | long chg; |
6d76dcf4 AK |
1133 | int ret, idx; |
1134 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 1135 | |
6d76dcf4 | 1136 | idx = hstate_index(h); |
a1e78772 | 1137 | /* |
90481622 DG |
1138 | * Processes that did not create the mapping will have no |
1139 | * reserves and will not have accounted against subpool | |
1140 | * limit. Check that the subpool limit can be made before | |
1141 | * satisfying the allocation MAP_NORESERVE mappings may also | |
1142 | * need pages and subpool limit allocated allocated if no reserve | |
1143 | * mapping overlaps. | |
a1e78772 | 1144 | */ |
a5516438 | 1145 | chg = vma_needs_reservation(h, vma, addr); |
c37f9fb1 | 1146 | if (chg < 0) |
76dcee75 | 1147 | return ERR_PTR(-ENOMEM); |
c37f9fb1 | 1148 | if (chg) |
90481622 | 1149 | if (hugepage_subpool_get_pages(spool, chg)) |
76dcee75 | 1150 | return ERR_PTR(-ENOSPC); |
1da177e4 | 1151 | |
6d76dcf4 AK |
1152 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
1153 | if (ret) { | |
1154 | hugepage_subpool_put_pages(spool, chg); | |
1155 | return ERR_PTR(-ENOSPC); | |
1156 | } | |
1da177e4 | 1157 | spin_lock(&hugetlb_lock); |
a5516438 | 1158 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve); |
81a6fcae | 1159 | if (!page) { |
94ae8ba7 | 1160 | spin_unlock(&hugetlb_lock); |
bf50bab2 | 1161 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
68842c9b | 1162 | if (!page) { |
6d76dcf4 AK |
1163 | hugetlb_cgroup_uncharge_cgroup(idx, |
1164 | pages_per_huge_page(h), | |
1165 | h_cg); | |
90481622 | 1166 | hugepage_subpool_put_pages(spool, chg); |
76dcee75 | 1167 | return ERR_PTR(-ENOSPC); |
68842c9b | 1168 | } |
79dbb236 AK |
1169 | spin_lock(&hugetlb_lock); |
1170 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 1171 | /* Fall through */ |
68842c9b | 1172 | } |
81a6fcae JK |
1173 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
1174 | spin_unlock(&hugetlb_lock); | |
348ea204 | 1175 | |
90481622 | 1176 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 1177 | |
a5516438 | 1178 | vma_commit_reservation(h, vma, addr); |
90d8b7e6 | 1179 | return page; |
b45b5bd6 DG |
1180 | } |
1181 | ||
91f47662 | 1182 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
aa888a74 AK |
1183 | { |
1184 | struct huge_bootmem_page *m; | |
b2261026 | 1185 | int nr_nodes, node; |
aa888a74 | 1186 | |
b2261026 | 1187 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
1188 | void *addr; |
1189 | ||
b2261026 | 1190 | addr = __alloc_bootmem_node_nopanic(NODE_DATA(node), |
aa888a74 AK |
1191 | huge_page_size(h), huge_page_size(h), 0); |
1192 | ||
1193 | if (addr) { | |
1194 | /* | |
1195 | * Use the beginning of the huge page to store the | |
1196 | * huge_bootmem_page struct (until gather_bootmem | |
1197 | * puts them into the mem_map). | |
1198 | */ | |
1199 | m = addr; | |
91f47662 | 1200 | goto found; |
aa888a74 | 1201 | } |
aa888a74 AK |
1202 | } |
1203 | return 0; | |
1204 | ||
1205 | found: | |
1206 | BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); | |
1207 | /* Put them into a private list first because mem_map is not up yet */ | |
1208 | list_add(&m->list, &huge_boot_pages); | |
1209 | m->hstate = h; | |
1210 | return 1; | |
1211 | } | |
1212 | ||
18229df5 AW |
1213 | static void prep_compound_huge_page(struct page *page, int order) |
1214 | { | |
1215 | if (unlikely(order > (MAX_ORDER - 1))) | |
1216 | prep_compound_gigantic_page(page, order); | |
1217 | else | |
1218 | prep_compound_page(page, order); | |
1219 | } | |
1220 | ||
aa888a74 AK |
1221 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
1222 | static void __init gather_bootmem_prealloc(void) | |
1223 | { | |
1224 | struct huge_bootmem_page *m; | |
1225 | ||
1226 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 1227 | struct hstate *h = m->hstate; |
ee8f248d BB |
1228 | struct page *page; |
1229 | ||
1230 | #ifdef CONFIG_HIGHMEM | |
1231 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
1232 | free_bootmem_late((unsigned long)m, | |
1233 | sizeof(struct huge_bootmem_page)); | |
1234 | #else | |
1235 | page = virt_to_page(m); | |
1236 | #endif | |
aa888a74 AK |
1237 | __ClearPageReserved(page); |
1238 | WARN_ON(page_count(page) != 1); | |
18229df5 | 1239 | prep_compound_huge_page(page, h->order); |
aa888a74 | 1240 | prep_new_huge_page(h, page, page_to_nid(page)); |
b0320c7b RA |
1241 | /* |
1242 | * If we had gigantic hugepages allocated at boot time, we need | |
1243 | * to restore the 'stolen' pages to totalram_pages in order to | |
1244 | * fix confusing memory reports from free(1) and another | |
1245 | * side-effects, like CommitLimit going negative. | |
1246 | */ | |
1247 | if (h->order > (MAX_ORDER - 1)) | |
3dcc0571 | 1248 | adjust_managed_page_count(page, 1 << h->order); |
aa888a74 AK |
1249 | } |
1250 | } | |
1251 | ||
8faa8b07 | 1252 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
1253 | { |
1254 | unsigned long i; | |
a5516438 | 1255 | |
e5ff2159 | 1256 | for (i = 0; i < h->max_huge_pages; ++i) { |
aa888a74 AK |
1257 | if (h->order >= MAX_ORDER) { |
1258 | if (!alloc_bootmem_huge_page(h)) | |
1259 | break; | |
9b5e5d0f | 1260 | } else if (!alloc_fresh_huge_page(h, |
8cebfcd0 | 1261 | &node_states[N_MEMORY])) |
1da177e4 | 1262 | break; |
1da177e4 | 1263 | } |
8faa8b07 | 1264 | h->max_huge_pages = i; |
e5ff2159 AK |
1265 | } |
1266 | ||
1267 | static void __init hugetlb_init_hstates(void) | |
1268 | { | |
1269 | struct hstate *h; | |
1270 | ||
1271 | for_each_hstate(h) { | |
8faa8b07 AK |
1272 | /* oversize hugepages were init'ed in early boot */ |
1273 | if (h->order < MAX_ORDER) | |
1274 | hugetlb_hstate_alloc_pages(h); | |
e5ff2159 AK |
1275 | } |
1276 | } | |
1277 | ||
4abd32db AK |
1278 | static char * __init memfmt(char *buf, unsigned long n) |
1279 | { | |
1280 | if (n >= (1UL << 30)) | |
1281 | sprintf(buf, "%lu GB", n >> 30); | |
1282 | else if (n >= (1UL << 20)) | |
1283 | sprintf(buf, "%lu MB", n >> 20); | |
1284 | else | |
1285 | sprintf(buf, "%lu KB", n >> 10); | |
1286 | return buf; | |
1287 | } | |
1288 | ||
e5ff2159 AK |
1289 | static void __init report_hugepages(void) |
1290 | { | |
1291 | struct hstate *h; | |
1292 | ||
1293 | for_each_hstate(h) { | |
4abd32db | 1294 | char buf[32]; |
ffb22af5 | 1295 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
4abd32db AK |
1296 | memfmt(buf, huge_page_size(h)), |
1297 | h->free_huge_pages); | |
e5ff2159 AK |
1298 | } |
1299 | } | |
1300 | ||
1da177e4 | 1301 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
1302 | static void try_to_free_low(struct hstate *h, unsigned long count, |
1303 | nodemask_t *nodes_allowed) | |
1da177e4 | 1304 | { |
4415cc8d CL |
1305 | int i; |
1306 | ||
aa888a74 AK |
1307 | if (h->order >= MAX_ORDER) |
1308 | return; | |
1309 | ||
6ae11b27 | 1310 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 1311 | struct page *page, *next; |
a5516438 AK |
1312 | struct list_head *freel = &h->hugepage_freelists[i]; |
1313 | list_for_each_entry_safe(page, next, freel, lru) { | |
1314 | if (count >= h->nr_huge_pages) | |
6b0c880d | 1315 | return; |
1da177e4 LT |
1316 | if (PageHighMem(page)) |
1317 | continue; | |
1318 | list_del(&page->lru); | |
e5ff2159 | 1319 | update_and_free_page(h, page); |
a5516438 AK |
1320 | h->free_huge_pages--; |
1321 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
1322 | } |
1323 | } | |
1324 | } | |
1325 | #else | |
6ae11b27 LS |
1326 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
1327 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
1328 | { |
1329 | } | |
1330 | #endif | |
1331 | ||
20a0307c WF |
1332 | /* |
1333 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
1334 | * balanced by operating on them in a round-robin fashion. | |
1335 | * Returns 1 if an adjustment was made. | |
1336 | */ | |
6ae11b27 LS |
1337 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
1338 | int delta) | |
20a0307c | 1339 | { |
b2261026 | 1340 | int nr_nodes, node; |
20a0307c WF |
1341 | |
1342 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 1343 | |
b2261026 JK |
1344 | if (delta < 0) { |
1345 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1346 | if (h->surplus_huge_pages_node[node]) | |
1347 | goto found; | |
e8c5c824 | 1348 | } |
b2261026 JK |
1349 | } else { |
1350 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
1351 | if (h->surplus_huge_pages_node[node] < | |
1352 | h->nr_huge_pages_node[node]) | |
1353 | goto found; | |
e8c5c824 | 1354 | } |
b2261026 JK |
1355 | } |
1356 | return 0; | |
20a0307c | 1357 | |
b2261026 JK |
1358 | found: |
1359 | h->surplus_huge_pages += delta; | |
1360 | h->surplus_huge_pages_node[node] += delta; | |
1361 | return 1; | |
20a0307c WF |
1362 | } |
1363 | ||
a5516438 | 1364 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
1365 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
1366 | nodemask_t *nodes_allowed) | |
1da177e4 | 1367 | { |
7893d1d5 | 1368 | unsigned long min_count, ret; |
1da177e4 | 1369 | |
aa888a74 AK |
1370 | if (h->order >= MAX_ORDER) |
1371 | return h->max_huge_pages; | |
1372 | ||
7893d1d5 AL |
1373 | /* |
1374 | * Increase the pool size | |
1375 | * First take pages out of surplus state. Then make up the | |
1376 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f NA |
1377 | * |
1378 | * We might race with alloc_buddy_huge_page() here and be unable | |
1379 | * to convert a surplus huge page to a normal huge page. That is | |
1380 | * not critical, though, it just means the overall size of the | |
1381 | * pool might be one hugepage larger than it needs to be, but | |
1382 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 1383 | */ |
1da177e4 | 1384 | spin_lock(&hugetlb_lock); |
a5516438 | 1385 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 1386 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
1387 | break; |
1388 | } | |
1389 | ||
a5516438 | 1390 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
1391 | /* |
1392 | * If this allocation races such that we no longer need the | |
1393 | * page, free_huge_page will handle it by freeing the page | |
1394 | * and reducing the surplus. | |
1395 | */ | |
1396 | spin_unlock(&hugetlb_lock); | |
6ae11b27 | 1397 | ret = alloc_fresh_huge_page(h, nodes_allowed); |
7893d1d5 AL |
1398 | spin_lock(&hugetlb_lock); |
1399 | if (!ret) | |
1400 | goto out; | |
1401 | ||
536240f2 MG |
1402 | /* Bail for signals. Probably ctrl-c from user */ |
1403 | if (signal_pending(current)) | |
1404 | goto out; | |
7893d1d5 | 1405 | } |
7893d1d5 AL |
1406 | |
1407 | /* | |
1408 | * Decrease the pool size | |
1409 | * First return free pages to the buddy allocator (being careful | |
1410 | * to keep enough around to satisfy reservations). Then place | |
1411 | * pages into surplus state as needed so the pool will shrink | |
1412 | * to the desired size as pages become free. | |
d1c3fb1f NA |
1413 | * |
1414 | * By placing pages into the surplus state independent of the | |
1415 | * overcommit value, we are allowing the surplus pool size to | |
1416 | * exceed overcommit. There are few sane options here. Since | |
1417 | * alloc_buddy_huge_page() is checking the global counter, | |
1418 | * though, we'll note that we're not allowed to exceed surplus | |
1419 | * and won't grow the pool anywhere else. Not until one of the | |
1420 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 1421 | */ |
a5516438 | 1422 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 1423 | min_count = max(count, min_count); |
6ae11b27 | 1424 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 1425 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 1426 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 1427 | break; |
1da177e4 | 1428 | } |
a5516438 | 1429 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 1430 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
1431 | break; |
1432 | } | |
1433 | out: | |
a5516438 | 1434 | ret = persistent_huge_pages(h); |
1da177e4 | 1435 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 1436 | return ret; |
1da177e4 LT |
1437 | } |
1438 | ||
a3437870 NA |
1439 | #define HSTATE_ATTR_RO(_name) \ |
1440 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
1441 | ||
1442 | #define HSTATE_ATTR(_name) \ | |
1443 | static struct kobj_attribute _name##_attr = \ | |
1444 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
1445 | ||
1446 | static struct kobject *hugepages_kobj; | |
1447 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1448 | ||
9a305230 LS |
1449 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
1450 | ||
1451 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
1452 | { |
1453 | int i; | |
9a305230 | 1454 | |
a3437870 | 1455 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
1456 | if (hstate_kobjs[i] == kobj) { |
1457 | if (nidp) | |
1458 | *nidp = NUMA_NO_NODE; | |
a3437870 | 1459 | return &hstates[i]; |
9a305230 LS |
1460 | } |
1461 | ||
1462 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
1463 | } |
1464 | ||
06808b08 | 1465 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
1466 | struct kobj_attribute *attr, char *buf) |
1467 | { | |
9a305230 LS |
1468 | struct hstate *h; |
1469 | unsigned long nr_huge_pages; | |
1470 | int nid; | |
1471 | ||
1472 | h = kobj_to_hstate(kobj, &nid); | |
1473 | if (nid == NUMA_NO_NODE) | |
1474 | nr_huge_pages = h->nr_huge_pages; | |
1475 | else | |
1476 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
1477 | ||
1478 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 1479 | } |
adbe8726 | 1480 | |
06808b08 LS |
1481 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
1482 | struct kobject *kobj, struct kobj_attribute *attr, | |
1483 | const char *buf, size_t len) | |
a3437870 NA |
1484 | { |
1485 | int err; | |
9a305230 | 1486 | int nid; |
06808b08 | 1487 | unsigned long count; |
9a305230 | 1488 | struct hstate *h; |
bad44b5b | 1489 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 1490 | |
3dbb95f7 | 1491 | err = kstrtoul(buf, 10, &count); |
73ae31e5 | 1492 | if (err) |
adbe8726 | 1493 | goto out; |
a3437870 | 1494 | |
9a305230 | 1495 | h = kobj_to_hstate(kobj, &nid); |
adbe8726 EM |
1496 | if (h->order >= MAX_ORDER) { |
1497 | err = -EINVAL; | |
1498 | goto out; | |
1499 | } | |
1500 | ||
9a305230 LS |
1501 | if (nid == NUMA_NO_NODE) { |
1502 | /* | |
1503 | * global hstate attribute | |
1504 | */ | |
1505 | if (!(obey_mempolicy && | |
1506 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
1507 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 1508 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 LS |
1509 | } |
1510 | } else if (nodes_allowed) { | |
1511 | /* | |
1512 | * per node hstate attribute: adjust count to global, | |
1513 | * but restrict alloc/free to the specified node. | |
1514 | */ | |
1515 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
1516 | init_nodemask_of_node(nodes_allowed, nid); | |
1517 | } else | |
8cebfcd0 | 1518 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 | 1519 | |
06808b08 | 1520 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 1521 | |
8cebfcd0 | 1522 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
1523 | NODEMASK_FREE(nodes_allowed); |
1524 | ||
1525 | return len; | |
adbe8726 EM |
1526 | out: |
1527 | NODEMASK_FREE(nodes_allowed); | |
1528 | return err; | |
06808b08 LS |
1529 | } |
1530 | ||
1531 | static ssize_t nr_hugepages_show(struct kobject *kobj, | |
1532 | struct kobj_attribute *attr, char *buf) | |
1533 | { | |
1534 | return nr_hugepages_show_common(kobj, attr, buf); | |
1535 | } | |
1536 | ||
1537 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
1538 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1539 | { | |
1540 | return nr_hugepages_store_common(false, kobj, attr, buf, len); | |
a3437870 NA |
1541 | } |
1542 | HSTATE_ATTR(nr_hugepages); | |
1543 | ||
06808b08 LS |
1544 | #ifdef CONFIG_NUMA |
1545 | ||
1546 | /* | |
1547 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
1548 | * huge page alloc/free. | |
1549 | */ | |
1550 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
1551 | struct kobj_attribute *attr, char *buf) | |
1552 | { | |
1553 | return nr_hugepages_show_common(kobj, attr, buf); | |
1554 | } | |
1555 | ||
1556 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
1557 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1558 | { | |
1559 | return nr_hugepages_store_common(true, kobj, attr, buf, len); | |
1560 | } | |
1561 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
1562 | #endif | |
1563 | ||
1564 | ||
a3437870 NA |
1565 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
1566 | struct kobj_attribute *attr, char *buf) | |
1567 | { | |
9a305230 | 1568 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1569 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
1570 | } | |
adbe8726 | 1571 | |
a3437870 NA |
1572 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
1573 | struct kobj_attribute *attr, const char *buf, size_t count) | |
1574 | { | |
1575 | int err; | |
1576 | unsigned long input; | |
9a305230 | 1577 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 1578 | |
adbe8726 EM |
1579 | if (h->order >= MAX_ORDER) |
1580 | return -EINVAL; | |
1581 | ||
3dbb95f7 | 1582 | err = kstrtoul(buf, 10, &input); |
a3437870 | 1583 | if (err) |
73ae31e5 | 1584 | return err; |
a3437870 NA |
1585 | |
1586 | spin_lock(&hugetlb_lock); | |
1587 | h->nr_overcommit_huge_pages = input; | |
1588 | spin_unlock(&hugetlb_lock); | |
1589 | ||
1590 | return count; | |
1591 | } | |
1592 | HSTATE_ATTR(nr_overcommit_hugepages); | |
1593 | ||
1594 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
1595 | struct kobj_attribute *attr, char *buf) | |
1596 | { | |
9a305230 LS |
1597 | struct hstate *h; |
1598 | unsigned long free_huge_pages; | |
1599 | int nid; | |
1600 | ||
1601 | h = kobj_to_hstate(kobj, &nid); | |
1602 | if (nid == NUMA_NO_NODE) | |
1603 | free_huge_pages = h->free_huge_pages; | |
1604 | else | |
1605 | free_huge_pages = h->free_huge_pages_node[nid]; | |
1606 | ||
1607 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
1608 | } |
1609 | HSTATE_ATTR_RO(free_hugepages); | |
1610 | ||
1611 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
1612 | struct kobj_attribute *attr, char *buf) | |
1613 | { | |
9a305230 | 1614 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1615 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
1616 | } | |
1617 | HSTATE_ATTR_RO(resv_hugepages); | |
1618 | ||
1619 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
1620 | struct kobj_attribute *attr, char *buf) | |
1621 | { | |
9a305230 LS |
1622 | struct hstate *h; |
1623 | unsigned long surplus_huge_pages; | |
1624 | int nid; | |
1625 | ||
1626 | h = kobj_to_hstate(kobj, &nid); | |
1627 | if (nid == NUMA_NO_NODE) | |
1628 | surplus_huge_pages = h->surplus_huge_pages; | |
1629 | else | |
1630 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
1631 | ||
1632 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
1633 | } |
1634 | HSTATE_ATTR_RO(surplus_hugepages); | |
1635 | ||
1636 | static struct attribute *hstate_attrs[] = { | |
1637 | &nr_hugepages_attr.attr, | |
1638 | &nr_overcommit_hugepages_attr.attr, | |
1639 | &free_hugepages_attr.attr, | |
1640 | &resv_hugepages_attr.attr, | |
1641 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
1642 | #ifdef CONFIG_NUMA |
1643 | &nr_hugepages_mempolicy_attr.attr, | |
1644 | #endif | |
a3437870 NA |
1645 | NULL, |
1646 | }; | |
1647 | ||
1648 | static struct attribute_group hstate_attr_group = { | |
1649 | .attrs = hstate_attrs, | |
1650 | }; | |
1651 | ||
094e9539 JM |
1652 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
1653 | struct kobject **hstate_kobjs, | |
1654 | struct attribute_group *hstate_attr_group) | |
a3437870 NA |
1655 | { |
1656 | int retval; | |
972dc4de | 1657 | int hi = hstate_index(h); |
a3437870 | 1658 | |
9a305230 LS |
1659 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
1660 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
1661 | return -ENOMEM; |
1662 | ||
9a305230 | 1663 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 1664 | if (retval) |
9a305230 | 1665 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
1666 | |
1667 | return retval; | |
1668 | } | |
1669 | ||
1670 | static void __init hugetlb_sysfs_init(void) | |
1671 | { | |
1672 | struct hstate *h; | |
1673 | int err; | |
1674 | ||
1675 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
1676 | if (!hugepages_kobj) | |
1677 | return; | |
1678 | ||
1679 | for_each_hstate(h) { | |
9a305230 LS |
1680 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
1681 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 1682 | if (err) |
ffb22af5 | 1683 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
1684 | } |
1685 | } | |
1686 | ||
9a305230 LS |
1687 | #ifdef CONFIG_NUMA |
1688 | ||
1689 | /* | |
1690 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
1691 | * with node devices in node_devices[] using a parallel array. The array |
1692 | * index of a node device or _hstate == node id. | |
1693 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
1694 | * the base kernel, on the hugetlb module. |
1695 | */ | |
1696 | struct node_hstate { | |
1697 | struct kobject *hugepages_kobj; | |
1698 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1699 | }; | |
1700 | struct node_hstate node_hstates[MAX_NUMNODES]; | |
1701 | ||
1702 | /* | |
10fbcf4c | 1703 | * A subset of global hstate attributes for node devices |
9a305230 LS |
1704 | */ |
1705 | static struct attribute *per_node_hstate_attrs[] = { | |
1706 | &nr_hugepages_attr.attr, | |
1707 | &free_hugepages_attr.attr, | |
1708 | &surplus_hugepages_attr.attr, | |
1709 | NULL, | |
1710 | }; | |
1711 | ||
1712 | static struct attribute_group per_node_hstate_attr_group = { | |
1713 | .attrs = per_node_hstate_attrs, | |
1714 | }; | |
1715 | ||
1716 | /* | |
10fbcf4c | 1717 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
1718 | * Returns node id via non-NULL nidp. |
1719 | */ | |
1720 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
1721 | { | |
1722 | int nid; | |
1723 | ||
1724 | for (nid = 0; nid < nr_node_ids; nid++) { | |
1725 | struct node_hstate *nhs = &node_hstates[nid]; | |
1726 | int i; | |
1727 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
1728 | if (nhs->hstate_kobjs[i] == kobj) { | |
1729 | if (nidp) | |
1730 | *nidp = nid; | |
1731 | return &hstates[i]; | |
1732 | } | |
1733 | } | |
1734 | ||
1735 | BUG(); | |
1736 | return NULL; | |
1737 | } | |
1738 | ||
1739 | /* | |
10fbcf4c | 1740 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
1741 | * No-op if no hstate attributes attached. |
1742 | */ | |
3cd8b44f | 1743 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
1744 | { |
1745 | struct hstate *h; | |
10fbcf4c | 1746 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
1747 | |
1748 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 1749 | return; /* no hstate attributes */ |
9a305230 | 1750 | |
972dc4de AK |
1751 | for_each_hstate(h) { |
1752 | int idx = hstate_index(h); | |
1753 | if (nhs->hstate_kobjs[idx]) { | |
1754 | kobject_put(nhs->hstate_kobjs[idx]); | |
1755 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 1756 | } |
972dc4de | 1757 | } |
9a305230 LS |
1758 | |
1759 | kobject_put(nhs->hugepages_kobj); | |
1760 | nhs->hugepages_kobj = NULL; | |
1761 | } | |
1762 | ||
1763 | /* | |
10fbcf4c | 1764 | * hugetlb module exit: unregister hstate attributes from node devices |
9a305230 LS |
1765 | * that have them. |
1766 | */ | |
1767 | static void hugetlb_unregister_all_nodes(void) | |
1768 | { | |
1769 | int nid; | |
1770 | ||
1771 | /* | |
10fbcf4c | 1772 | * disable node device registrations. |
9a305230 LS |
1773 | */ |
1774 | register_hugetlbfs_with_node(NULL, NULL); | |
1775 | ||
1776 | /* | |
1777 | * remove hstate attributes from any nodes that have them. | |
1778 | */ | |
1779 | for (nid = 0; nid < nr_node_ids; nid++) | |
8732794b | 1780 | hugetlb_unregister_node(node_devices[nid]); |
9a305230 LS |
1781 | } |
1782 | ||
1783 | /* | |
10fbcf4c | 1784 | * Register hstate attributes for a single node device. |
9a305230 LS |
1785 | * No-op if attributes already registered. |
1786 | */ | |
3cd8b44f | 1787 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
1788 | { |
1789 | struct hstate *h; | |
10fbcf4c | 1790 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
1791 | int err; |
1792 | ||
1793 | if (nhs->hugepages_kobj) | |
1794 | return; /* already allocated */ | |
1795 | ||
1796 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 1797 | &node->dev.kobj); |
9a305230 LS |
1798 | if (!nhs->hugepages_kobj) |
1799 | return; | |
1800 | ||
1801 | for_each_hstate(h) { | |
1802 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
1803 | nhs->hstate_kobjs, | |
1804 | &per_node_hstate_attr_group); | |
1805 | if (err) { | |
ffb22af5 AM |
1806 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
1807 | h->name, node->dev.id); | |
9a305230 LS |
1808 | hugetlb_unregister_node(node); |
1809 | break; | |
1810 | } | |
1811 | } | |
1812 | } | |
1813 | ||
1814 | /* | |
9b5e5d0f | 1815 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
1816 | * devices of nodes that have memory. All on-line nodes should have |
1817 | * registered their associated device by this time. | |
9a305230 LS |
1818 | */ |
1819 | static void hugetlb_register_all_nodes(void) | |
1820 | { | |
1821 | int nid; | |
1822 | ||
8cebfcd0 | 1823 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 1824 | struct node *node = node_devices[nid]; |
10fbcf4c | 1825 | if (node->dev.id == nid) |
9a305230 LS |
1826 | hugetlb_register_node(node); |
1827 | } | |
1828 | ||
1829 | /* | |
10fbcf4c | 1830 | * Let the node device driver know we're here so it can |
9a305230 LS |
1831 | * [un]register hstate attributes on node hotplug. |
1832 | */ | |
1833 | register_hugetlbfs_with_node(hugetlb_register_node, | |
1834 | hugetlb_unregister_node); | |
1835 | } | |
1836 | #else /* !CONFIG_NUMA */ | |
1837 | ||
1838 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
1839 | { | |
1840 | BUG(); | |
1841 | if (nidp) | |
1842 | *nidp = -1; | |
1843 | return NULL; | |
1844 | } | |
1845 | ||
1846 | static void hugetlb_unregister_all_nodes(void) { } | |
1847 | ||
1848 | static void hugetlb_register_all_nodes(void) { } | |
1849 | ||
1850 | #endif | |
1851 | ||
a3437870 NA |
1852 | static void __exit hugetlb_exit(void) |
1853 | { | |
1854 | struct hstate *h; | |
1855 | ||
9a305230 LS |
1856 | hugetlb_unregister_all_nodes(); |
1857 | ||
a3437870 | 1858 | for_each_hstate(h) { |
972dc4de | 1859 | kobject_put(hstate_kobjs[hstate_index(h)]); |
a3437870 NA |
1860 | } |
1861 | ||
1862 | kobject_put(hugepages_kobj); | |
1863 | } | |
1864 | module_exit(hugetlb_exit); | |
1865 | ||
1866 | static int __init hugetlb_init(void) | |
1867 | { | |
0ef89d25 BH |
1868 | /* Some platform decide whether they support huge pages at boot |
1869 | * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when | |
1870 | * there is no such support | |
1871 | */ | |
1872 | if (HPAGE_SHIFT == 0) | |
1873 | return 0; | |
a3437870 | 1874 | |
e11bfbfc NP |
1875 | if (!size_to_hstate(default_hstate_size)) { |
1876 | default_hstate_size = HPAGE_SIZE; | |
1877 | if (!size_to_hstate(default_hstate_size)) | |
1878 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 1879 | } |
972dc4de | 1880 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
e11bfbfc NP |
1881 | if (default_hstate_max_huge_pages) |
1882 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
a3437870 NA |
1883 | |
1884 | hugetlb_init_hstates(); | |
aa888a74 | 1885 | gather_bootmem_prealloc(); |
a3437870 NA |
1886 | report_hugepages(); |
1887 | ||
1888 | hugetlb_sysfs_init(); | |
9a305230 | 1889 | hugetlb_register_all_nodes(); |
7179e7bf | 1890 | hugetlb_cgroup_file_init(); |
9a305230 | 1891 | |
a3437870 NA |
1892 | return 0; |
1893 | } | |
1894 | module_init(hugetlb_init); | |
1895 | ||
1896 | /* Should be called on processing a hugepagesz=... option */ | |
1897 | void __init hugetlb_add_hstate(unsigned order) | |
1898 | { | |
1899 | struct hstate *h; | |
8faa8b07 AK |
1900 | unsigned long i; |
1901 | ||
a3437870 | 1902 | if (size_to_hstate(PAGE_SIZE << order)) { |
ffb22af5 | 1903 | pr_warning("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
1904 | return; |
1905 | } | |
47d38344 | 1906 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 1907 | BUG_ON(order == 0); |
47d38344 | 1908 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
1909 | h->order = order; |
1910 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
1911 | h->nr_huge_pages = 0; |
1912 | h->free_huge_pages = 0; | |
1913 | for (i = 0; i < MAX_NUMNODES; ++i) | |
1914 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 1915 | INIT_LIST_HEAD(&h->hugepage_activelist); |
8cebfcd0 LJ |
1916 | h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); |
1917 | h->next_nid_to_free = first_node(node_states[N_MEMORY]); | |
a3437870 NA |
1918 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
1919 | huge_page_size(h)/1024); | |
8faa8b07 | 1920 | |
a3437870 NA |
1921 | parsed_hstate = h; |
1922 | } | |
1923 | ||
e11bfbfc | 1924 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
1925 | { |
1926 | unsigned long *mhp; | |
8faa8b07 | 1927 | static unsigned long *last_mhp; |
a3437870 NA |
1928 | |
1929 | /* | |
47d38344 | 1930 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
1931 | * so this hugepages= parameter goes to the "default hstate". |
1932 | */ | |
47d38344 | 1933 | if (!hugetlb_max_hstate) |
a3437870 NA |
1934 | mhp = &default_hstate_max_huge_pages; |
1935 | else | |
1936 | mhp = &parsed_hstate->max_huge_pages; | |
1937 | ||
8faa8b07 | 1938 | if (mhp == last_mhp) { |
ffb22af5 AM |
1939 | pr_warning("hugepages= specified twice without " |
1940 | "interleaving hugepagesz=, ignoring\n"); | |
8faa8b07 AK |
1941 | return 1; |
1942 | } | |
1943 | ||
a3437870 NA |
1944 | if (sscanf(s, "%lu", mhp) <= 0) |
1945 | *mhp = 0; | |
1946 | ||
8faa8b07 AK |
1947 | /* |
1948 | * Global state is always initialized later in hugetlb_init. | |
1949 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
1950 | * use the bootmem allocator. | |
1951 | */ | |
47d38344 | 1952 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
1953 | hugetlb_hstate_alloc_pages(parsed_hstate); |
1954 | ||
1955 | last_mhp = mhp; | |
1956 | ||
a3437870 NA |
1957 | return 1; |
1958 | } | |
e11bfbfc NP |
1959 | __setup("hugepages=", hugetlb_nrpages_setup); |
1960 | ||
1961 | static int __init hugetlb_default_setup(char *s) | |
1962 | { | |
1963 | default_hstate_size = memparse(s, &s); | |
1964 | return 1; | |
1965 | } | |
1966 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 1967 | |
8a213460 NA |
1968 | static unsigned int cpuset_mems_nr(unsigned int *array) |
1969 | { | |
1970 | int node; | |
1971 | unsigned int nr = 0; | |
1972 | ||
1973 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
1974 | nr += array[node]; | |
1975 | ||
1976 | return nr; | |
1977 | } | |
1978 | ||
1979 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
1980 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
1981 | struct ctl_table *table, int write, | |
1982 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 1983 | { |
e5ff2159 AK |
1984 | struct hstate *h = &default_hstate; |
1985 | unsigned long tmp; | |
08d4a246 | 1986 | int ret; |
e5ff2159 | 1987 | |
c033a93c | 1988 | tmp = h->max_huge_pages; |
e5ff2159 | 1989 | |
adbe8726 EM |
1990 | if (write && h->order >= MAX_ORDER) |
1991 | return -EINVAL; | |
1992 | ||
e5ff2159 AK |
1993 | table->data = &tmp; |
1994 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
1995 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
1996 | if (ret) | |
1997 | goto out; | |
e5ff2159 | 1998 | |
06808b08 | 1999 | if (write) { |
bad44b5b DR |
2000 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, |
2001 | GFP_KERNEL | __GFP_NORETRY); | |
06808b08 LS |
2002 | if (!(obey_mempolicy && |
2003 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
2004 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 2005 | nodes_allowed = &node_states[N_MEMORY]; |
06808b08 LS |
2006 | } |
2007 | h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); | |
2008 | ||
8cebfcd0 | 2009 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
2010 | NODEMASK_FREE(nodes_allowed); |
2011 | } | |
08d4a246 MH |
2012 | out: |
2013 | return ret; | |
1da177e4 | 2014 | } |
396faf03 | 2015 | |
06808b08 LS |
2016 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2017 | void __user *buffer, size_t *length, loff_t *ppos) | |
2018 | { | |
2019 | ||
2020 | return hugetlb_sysctl_handler_common(false, table, write, | |
2021 | buffer, length, ppos); | |
2022 | } | |
2023 | ||
2024 | #ifdef CONFIG_NUMA | |
2025 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2026 | void __user *buffer, size_t *length, loff_t *ppos) | |
2027 | { | |
2028 | return hugetlb_sysctl_handler_common(true, table, write, | |
2029 | buffer, length, ppos); | |
2030 | } | |
2031 | #endif /* CONFIG_NUMA */ | |
2032 | ||
396faf03 | 2033 | int hugetlb_treat_movable_handler(struct ctl_table *table, int write, |
8d65af78 | 2034 | void __user *buffer, |
396faf03 MG |
2035 | size_t *length, loff_t *ppos) |
2036 | { | |
8d65af78 | 2037 | proc_dointvec(table, write, buffer, length, ppos); |
396faf03 MG |
2038 | if (hugepages_treat_as_movable) |
2039 | htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; | |
2040 | else | |
2041 | htlb_alloc_mask = GFP_HIGHUSER; | |
2042 | return 0; | |
2043 | } | |
2044 | ||
a3d0c6aa | 2045 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 2046 | void __user *buffer, |
a3d0c6aa NA |
2047 | size_t *length, loff_t *ppos) |
2048 | { | |
a5516438 | 2049 | struct hstate *h = &default_hstate; |
e5ff2159 | 2050 | unsigned long tmp; |
08d4a246 | 2051 | int ret; |
e5ff2159 | 2052 | |
c033a93c | 2053 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 2054 | |
adbe8726 EM |
2055 | if (write && h->order >= MAX_ORDER) |
2056 | return -EINVAL; | |
2057 | ||
e5ff2159 AK |
2058 | table->data = &tmp; |
2059 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2060 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2061 | if (ret) | |
2062 | goto out; | |
e5ff2159 AK |
2063 | |
2064 | if (write) { | |
2065 | spin_lock(&hugetlb_lock); | |
2066 | h->nr_overcommit_huge_pages = tmp; | |
2067 | spin_unlock(&hugetlb_lock); | |
2068 | } | |
08d4a246 MH |
2069 | out: |
2070 | return ret; | |
a3d0c6aa NA |
2071 | } |
2072 | ||
1da177e4 LT |
2073 | #endif /* CONFIG_SYSCTL */ |
2074 | ||
e1759c21 | 2075 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 2076 | { |
a5516438 | 2077 | struct hstate *h = &default_hstate; |
e1759c21 | 2078 | seq_printf(m, |
4f98a2fe RR |
2079 | "HugePages_Total: %5lu\n" |
2080 | "HugePages_Free: %5lu\n" | |
2081 | "HugePages_Rsvd: %5lu\n" | |
2082 | "HugePages_Surp: %5lu\n" | |
2083 | "Hugepagesize: %8lu kB\n", | |
a5516438 AK |
2084 | h->nr_huge_pages, |
2085 | h->free_huge_pages, | |
2086 | h->resv_huge_pages, | |
2087 | h->surplus_huge_pages, | |
2088 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
2089 | } |
2090 | ||
2091 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
2092 | { | |
a5516438 | 2093 | struct hstate *h = &default_hstate; |
1da177e4 LT |
2094 | return sprintf(buf, |
2095 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
2096 | "Node %d HugePages_Free: %5u\n" |
2097 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
2098 | nid, h->nr_huge_pages_node[nid], |
2099 | nid, h->free_huge_pages_node[nid], | |
2100 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
2101 | } |
2102 | ||
949f7ec5 DR |
2103 | void hugetlb_show_meminfo(void) |
2104 | { | |
2105 | struct hstate *h; | |
2106 | int nid; | |
2107 | ||
2108 | for_each_node_state(nid, N_MEMORY) | |
2109 | for_each_hstate(h) | |
2110 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
2111 | nid, | |
2112 | h->nr_huge_pages_node[nid], | |
2113 | h->free_huge_pages_node[nid], | |
2114 | h->surplus_huge_pages_node[nid], | |
2115 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
2116 | } | |
2117 | ||
1da177e4 LT |
2118 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
2119 | unsigned long hugetlb_total_pages(void) | |
2120 | { | |
d0028588 WL |
2121 | struct hstate *h; |
2122 | unsigned long nr_total_pages = 0; | |
2123 | ||
2124 | for_each_hstate(h) | |
2125 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
2126 | return nr_total_pages; | |
1da177e4 | 2127 | } |
1da177e4 | 2128 | |
a5516438 | 2129 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
2130 | { |
2131 | int ret = -ENOMEM; | |
2132 | ||
2133 | spin_lock(&hugetlb_lock); | |
2134 | /* | |
2135 | * When cpuset is configured, it breaks the strict hugetlb page | |
2136 | * reservation as the accounting is done on a global variable. Such | |
2137 | * reservation is completely rubbish in the presence of cpuset because | |
2138 | * the reservation is not checked against page availability for the | |
2139 | * current cpuset. Application can still potentially OOM'ed by kernel | |
2140 | * with lack of free htlb page in cpuset that the task is in. | |
2141 | * Attempt to enforce strict accounting with cpuset is almost | |
2142 | * impossible (or too ugly) because cpuset is too fluid that | |
2143 | * task or memory node can be dynamically moved between cpusets. | |
2144 | * | |
2145 | * The change of semantics for shared hugetlb mapping with cpuset is | |
2146 | * undesirable. However, in order to preserve some of the semantics, | |
2147 | * we fall back to check against current free page availability as | |
2148 | * a best attempt and hopefully to minimize the impact of changing | |
2149 | * semantics that cpuset has. | |
2150 | */ | |
2151 | if (delta > 0) { | |
a5516438 | 2152 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
2153 | goto out; |
2154 | ||
a5516438 AK |
2155 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
2156 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
2157 | goto out; |
2158 | } | |
2159 | } | |
2160 | ||
2161 | ret = 0; | |
2162 | if (delta < 0) | |
a5516438 | 2163 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
2164 | |
2165 | out: | |
2166 | spin_unlock(&hugetlb_lock); | |
2167 | return ret; | |
2168 | } | |
2169 | ||
84afd99b AW |
2170 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
2171 | { | |
2172 | struct resv_map *reservations = vma_resv_map(vma); | |
2173 | ||
2174 | /* | |
2175 | * This new VMA should share its siblings reservation map if present. | |
2176 | * The VMA will only ever have a valid reservation map pointer where | |
2177 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 2178 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
2179 | * after this open call completes. It is therefore safe to take a |
2180 | * new reference here without additional locking. | |
2181 | */ | |
2182 | if (reservations) | |
2183 | kref_get(&reservations->refs); | |
2184 | } | |
2185 | ||
c50ac050 DH |
2186 | static void resv_map_put(struct vm_area_struct *vma) |
2187 | { | |
2188 | struct resv_map *reservations = vma_resv_map(vma); | |
2189 | ||
2190 | if (!reservations) | |
2191 | return; | |
2192 | kref_put(&reservations->refs, resv_map_release); | |
2193 | } | |
2194 | ||
a1e78772 MG |
2195 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
2196 | { | |
a5516438 | 2197 | struct hstate *h = hstate_vma(vma); |
84afd99b | 2198 | struct resv_map *reservations = vma_resv_map(vma); |
90481622 | 2199 | struct hugepage_subpool *spool = subpool_vma(vma); |
84afd99b AW |
2200 | unsigned long reserve; |
2201 | unsigned long start; | |
2202 | unsigned long end; | |
2203 | ||
2204 | if (reservations) { | |
a5516438 AK |
2205 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
2206 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b AW |
2207 | |
2208 | reserve = (end - start) - | |
2209 | region_count(&reservations->regions, start, end); | |
2210 | ||
c50ac050 | 2211 | resv_map_put(vma); |
84afd99b | 2212 | |
7251ff78 | 2213 | if (reserve) { |
a5516438 | 2214 | hugetlb_acct_memory(h, -reserve); |
90481622 | 2215 | hugepage_subpool_put_pages(spool, reserve); |
7251ff78 | 2216 | } |
84afd99b | 2217 | } |
a1e78772 MG |
2218 | } |
2219 | ||
1da177e4 LT |
2220 | /* |
2221 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
2222 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
2223 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
2224 | * this far. | |
2225 | */ | |
d0217ac0 | 2226 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
2227 | { |
2228 | BUG(); | |
d0217ac0 | 2229 | return 0; |
1da177e4 LT |
2230 | } |
2231 | ||
f0f37e2f | 2232 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 2233 | .fault = hugetlb_vm_op_fault, |
84afd99b | 2234 | .open = hugetlb_vm_op_open, |
a1e78772 | 2235 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
2236 | }; |
2237 | ||
1e8f889b DG |
2238 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
2239 | int writable) | |
63551ae0 DG |
2240 | { |
2241 | pte_t entry; | |
2242 | ||
1e8f889b | 2243 | if (writable) { |
106c992a GS |
2244 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
2245 | vma->vm_page_prot))); | |
63551ae0 | 2246 | } else { |
106c992a GS |
2247 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
2248 | vma->vm_page_prot)); | |
63551ae0 DG |
2249 | } |
2250 | entry = pte_mkyoung(entry); | |
2251 | entry = pte_mkhuge(entry); | |
d9ed9faa | 2252 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
2253 | |
2254 | return entry; | |
2255 | } | |
2256 | ||
1e8f889b DG |
2257 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
2258 | unsigned long address, pte_t *ptep) | |
2259 | { | |
2260 | pte_t entry; | |
2261 | ||
106c992a | 2262 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 2263 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 2264 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
2265 | } |
2266 | ||
2267 | ||
63551ae0 DG |
2268 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
2269 | struct vm_area_struct *vma) | |
2270 | { | |
2271 | pte_t *src_pte, *dst_pte, entry; | |
2272 | struct page *ptepage; | |
1c59827d | 2273 | unsigned long addr; |
1e8f889b | 2274 | int cow; |
a5516438 AK |
2275 | struct hstate *h = hstate_vma(vma); |
2276 | unsigned long sz = huge_page_size(h); | |
1e8f889b DG |
2277 | |
2278 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 2279 | |
a5516438 | 2280 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
c74df32c HD |
2281 | src_pte = huge_pte_offset(src, addr); |
2282 | if (!src_pte) | |
2283 | continue; | |
a5516438 | 2284 | dst_pte = huge_pte_alloc(dst, addr, sz); |
63551ae0 DG |
2285 | if (!dst_pte) |
2286 | goto nomem; | |
c5c99429 LW |
2287 | |
2288 | /* If the pagetables are shared don't copy or take references */ | |
2289 | if (dst_pte == src_pte) | |
2290 | continue; | |
2291 | ||
c74df32c | 2292 | spin_lock(&dst->page_table_lock); |
46478758 | 2293 | spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); |
7f2e9525 | 2294 | if (!huge_pte_none(huge_ptep_get(src_pte))) { |
1e8f889b | 2295 | if (cow) |
7f2e9525 GS |
2296 | huge_ptep_set_wrprotect(src, addr, src_pte); |
2297 | entry = huge_ptep_get(src_pte); | |
1c59827d HD |
2298 | ptepage = pte_page(entry); |
2299 | get_page(ptepage); | |
0fe6e20b | 2300 | page_dup_rmap(ptepage); |
1c59827d HD |
2301 | set_huge_pte_at(dst, addr, dst_pte, entry); |
2302 | } | |
2303 | spin_unlock(&src->page_table_lock); | |
c74df32c | 2304 | spin_unlock(&dst->page_table_lock); |
63551ae0 DG |
2305 | } |
2306 | return 0; | |
2307 | ||
2308 | nomem: | |
2309 | return -ENOMEM; | |
2310 | } | |
2311 | ||
290408d4 NH |
2312 | static int is_hugetlb_entry_migration(pte_t pte) |
2313 | { | |
2314 | swp_entry_t swp; | |
2315 | ||
2316 | if (huge_pte_none(pte) || pte_present(pte)) | |
2317 | return 0; | |
2318 | swp = pte_to_swp_entry(pte); | |
32f84528 | 2319 | if (non_swap_entry(swp) && is_migration_entry(swp)) |
290408d4 | 2320 | return 1; |
32f84528 | 2321 | else |
290408d4 NH |
2322 | return 0; |
2323 | } | |
2324 | ||
fd6a03ed NH |
2325 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) |
2326 | { | |
2327 | swp_entry_t swp; | |
2328 | ||
2329 | if (huge_pte_none(pte) || pte_present(pte)) | |
2330 | return 0; | |
2331 | swp = pte_to_swp_entry(pte); | |
32f84528 | 2332 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) |
fd6a03ed | 2333 | return 1; |
32f84528 | 2334 | else |
fd6a03ed NH |
2335 | return 0; |
2336 | } | |
2337 | ||
24669e58 AK |
2338 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
2339 | unsigned long start, unsigned long end, | |
2340 | struct page *ref_page) | |
63551ae0 | 2341 | { |
24669e58 | 2342 | int force_flush = 0; |
63551ae0 DG |
2343 | struct mm_struct *mm = vma->vm_mm; |
2344 | unsigned long address; | |
c7546f8f | 2345 | pte_t *ptep; |
63551ae0 DG |
2346 | pte_t pte; |
2347 | struct page *page; | |
a5516438 AK |
2348 | struct hstate *h = hstate_vma(vma); |
2349 | unsigned long sz = huge_page_size(h); | |
2ec74c3e SG |
2350 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
2351 | const unsigned long mmun_end = end; /* For mmu_notifiers */ | |
a5516438 | 2352 | |
63551ae0 | 2353 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
2354 | BUG_ON(start & ~huge_page_mask(h)); |
2355 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 2356 | |
24669e58 | 2357 | tlb_start_vma(tlb, vma); |
2ec74c3e | 2358 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
24669e58 | 2359 | again: |
508034a3 | 2360 | spin_lock(&mm->page_table_lock); |
a5516438 | 2361 | for (address = start; address < end; address += sz) { |
c7546f8f | 2362 | ptep = huge_pte_offset(mm, address); |
4c887265 | 2363 | if (!ptep) |
c7546f8f DG |
2364 | continue; |
2365 | ||
39dde65c KC |
2366 | if (huge_pmd_unshare(mm, &address, ptep)) |
2367 | continue; | |
2368 | ||
6629326b HD |
2369 | pte = huge_ptep_get(ptep); |
2370 | if (huge_pte_none(pte)) | |
2371 | continue; | |
2372 | ||
2373 | /* | |
2374 | * HWPoisoned hugepage is already unmapped and dropped reference | |
2375 | */ | |
8c4894c6 | 2376 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { |
106c992a | 2377 | huge_pte_clear(mm, address, ptep); |
6629326b | 2378 | continue; |
8c4894c6 | 2379 | } |
6629326b HD |
2380 | |
2381 | page = pte_page(pte); | |
04f2cbe3 MG |
2382 | /* |
2383 | * If a reference page is supplied, it is because a specific | |
2384 | * page is being unmapped, not a range. Ensure the page we | |
2385 | * are about to unmap is the actual page of interest. | |
2386 | */ | |
2387 | if (ref_page) { | |
04f2cbe3 MG |
2388 | if (page != ref_page) |
2389 | continue; | |
2390 | ||
2391 | /* | |
2392 | * Mark the VMA as having unmapped its page so that | |
2393 | * future faults in this VMA will fail rather than | |
2394 | * looking like data was lost | |
2395 | */ | |
2396 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
2397 | } | |
2398 | ||
c7546f8f | 2399 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
24669e58 | 2400 | tlb_remove_tlb_entry(tlb, ptep, address); |
106c992a | 2401 | if (huge_pte_dirty(pte)) |
6649a386 | 2402 | set_page_dirty(page); |
9e81130b | 2403 | |
24669e58 AK |
2404 | page_remove_rmap(page); |
2405 | force_flush = !__tlb_remove_page(tlb, page); | |
2406 | if (force_flush) | |
2407 | break; | |
9e81130b HD |
2408 | /* Bail out after unmapping reference page if supplied */ |
2409 | if (ref_page) | |
2410 | break; | |
63551ae0 | 2411 | } |
cd2934a3 | 2412 | spin_unlock(&mm->page_table_lock); |
24669e58 AK |
2413 | /* |
2414 | * mmu_gather ran out of room to batch pages, we break out of | |
2415 | * the PTE lock to avoid doing the potential expensive TLB invalidate | |
2416 | * and page-free while holding it. | |
2417 | */ | |
2418 | if (force_flush) { | |
2419 | force_flush = 0; | |
2420 | tlb_flush_mmu(tlb); | |
2421 | if (address < end && !ref_page) | |
2422 | goto again; | |
fe1668ae | 2423 | } |
2ec74c3e | 2424 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
24669e58 | 2425 | tlb_end_vma(tlb, vma); |
1da177e4 | 2426 | } |
63551ae0 | 2427 | |
d833352a MG |
2428 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
2429 | struct vm_area_struct *vma, unsigned long start, | |
2430 | unsigned long end, struct page *ref_page) | |
2431 | { | |
2432 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
2433 | ||
2434 | /* | |
2435 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
2436 | * test will fail on a vma being torn down, and not grab a page table | |
2437 | * on its way out. We're lucky that the flag has such an appropriate | |
2438 | * name, and can in fact be safely cleared here. We could clear it | |
2439 | * before the __unmap_hugepage_range above, but all that's necessary | |
2440 | * is to clear it before releasing the i_mmap_mutex. This works | |
2441 | * because in the context this is called, the VMA is about to be | |
2442 | * destroyed and the i_mmap_mutex is held. | |
2443 | */ | |
2444 | vma->vm_flags &= ~VM_MAYSHARE; | |
2445 | } | |
2446 | ||
502717f4 | 2447 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 2448 | unsigned long end, struct page *ref_page) |
502717f4 | 2449 | { |
24669e58 AK |
2450 | struct mm_struct *mm; |
2451 | struct mmu_gather tlb; | |
2452 | ||
2453 | mm = vma->vm_mm; | |
2454 | ||
2b047252 | 2455 | tlb_gather_mmu(&tlb, mm, start, end); |
24669e58 AK |
2456 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
2457 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 KC |
2458 | } |
2459 | ||
04f2cbe3 MG |
2460 | /* |
2461 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
2462 | * mappping it owns the reserve page for. The intention is to unmap the page | |
2463 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
2464 | * same region. | |
2465 | */ | |
2a4b3ded HH |
2466 | static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
2467 | struct page *page, unsigned long address) | |
04f2cbe3 | 2468 | { |
7526674d | 2469 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
2470 | struct vm_area_struct *iter_vma; |
2471 | struct address_space *mapping; | |
04f2cbe3 MG |
2472 | pgoff_t pgoff; |
2473 | ||
2474 | /* | |
2475 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
2476 | * from page cache lookup which is in HPAGE_SIZE units. | |
2477 | */ | |
7526674d | 2478 | address = address & huge_page_mask(h); |
36e4f20a MH |
2479 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
2480 | vma->vm_pgoff; | |
496ad9aa | 2481 | mapping = file_inode(vma->vm_file)->i_mapping; |
04f2cbe3 | 2482 | |
4eb2b1dc MG |
2483 | /* |
2484 | * Take the mapping lock for the duration of the table walk. As | |
2485 | * this mapping should be shared between all the VMAs, | |
2486 | * __unmap_hugepage_range() is called as the lock is already held | |
2487 | */ | |
3d48ae45 | 2488 | mutex_lock(&mapping->i_mmap_mutex); |
6b2dbba8 | 2489 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
2490 | /* Do not unmap the current VMA */ |
2491 | if (iter_vma == vma) | |
2492 | continue; | |
2493 | ||
2494 | /* | |
2495 | * Unmap the page from other VMAs without their own reserves. | |
2496 | * They get marked to be SIGKILLed if they fault in these | |
2497 | * areas. This is because a future no-page fault on this VMA | |
2498 | * could insert a zeroed page instead of the data existing | |
2499 | * from the time of fork. This would look like data corruption | |
2500 | */ | |
2501 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
2502 | unmap_hugepage_range(iter_vma, address, |
2503 | address + huge_page_size(h), page); | |
04f2cbe3 | 2504 | } |
3d48ae45 | 2505 | mutex_unlock(&mapping->i_mmap_mutex); |
04f2cbe3 MG |
2506 | |
2507 | return 1; | |
2508 | } | |
2509 | ||
0fe6e20b NH |
2510 | /* |
2511 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
2512 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
2513 | * cannot race with other handlers or page migration. | |
2514 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 2515 | */ |
1e8f889b | 2516 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
04f2cbe3 MG |
2517 | unsigned long address, pte_t *ptep, pte_t pte, |
2518 | struct page *pagecache_page) | |
1e8f889b | 2519 | { |
a5516438 | 2520 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 2521 | struct page *old_page, *new_page; |
04f2cbe3 | 2522 | int outside_reserve = 0; |
2ec74c3e SG |
2523 | unsigned long mmun_start; /* For mmu_notifiers */ |
2524 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1e8f889b DG |
2525 | |
2526 | old_page = pte_page(pte); | |
2527 | ||
04f2cbe3 | 2528 | retry_avoidcopy: |
1e8f889b DG |
2529 | /* If no-one else is actually using this page, avoid the copy |
2530 | * and just make the page writable */ | |
37a2140d JK |
2531 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
2532 | page_move_anon_rmap(old_page, vma, address); | |
1e8f889b | 2533 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 2534 | return 0; |
1e8f889b DG |
2535 | } |
2536 | ||
04f2cbe3 MG |
2537 | /* |
2538 | * If the process that created a MAP_PRIVATE mapping is about to | |
2539 | * perform a COW due to a shared page count, attempt to satisfy | |
2540 | * the allocation without using the existing reserves. The pagecache | |
2541 | * page is used to determine if the reserve at this address was | |
2542 | * consumed or not. If reserves were used, a partial faulted mapping | |
2543 | * at the time of fork() could consume its reserves on COW instead | |
2544 | * of the full address range. | |
2545 | */ | |
f83a275d | 2546 | if (!(vma->vm_flags & VM_MAYSHARE) && |
04f2cbe3 MG |
2547 | is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
2548 | old_page != pagecache_page) | |
2549 | outside_reserve = 1; | |
2550 | ||
1e8f889b | 2551 | page_cache_get(old_page); |
b76c8cfb LW |
2552 | |
2553 | /* Drop page_table_lock as buddy allocator may be called */ | |
2554 | spin_unlock(&mm->page_table_lock); | |
04f2cbe3 | 2555 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 2556 | |
2fc39cec | 2557 | if (IS_ERR(new_page)) { |
76dcee75 | 2558 | long err = PTR_ERR(new_page); |
1e8f889b | 2559 | page_cache_release(old_page); |
04f2cbe3 MG |
2560 | |
2561 | /* | |
2562 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
2563 | * it is due to references held by a child and an insufficient | |
2564 | * huge page pool. To guarantee the original mappers | |
2565 | * reliability, unmap the page from child processes. The child | |
2566 | * may get SIGKILLed if it later faults. | |
2567 | */ | |
2568 | if (outside_reserve) { | |
2569 | BUG_ON(huge_pte_none(pte)); | |
2570 | if (unmap_ref_private(mm, vma, old_page, address)) { | |
04f2cbe3 | 2571 | BUG_ON(huge_pte_none(pte)); |
b76c8cfb | 2572 | spin_lock(&mm->page_table_lock); |
a734bcc8 HD |
2573 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
2574 | if (likely(pte_same(huge_ptep_get(ptep), pte))) | |
2575 | goto retry_avoidcopy; | |
2576 | /* | |
2577 | * race occurs while re-acquiring page_table_lock, and | |
2578 | * our job is done. | |
2579 | */ | |
2580 | return 0; | |
04f2cbe3 MG |
2581 | } |
2582 | WARN_ON_ONCE(1); | |
2583 | } | |
2584 | ||
b76c8cfb LW |
2585 | /* Caller expects lock to be held */ |
2586 | spin_lock(&mm->page_table_lock); | |
76dcee75 AK |
2587 | if (err == -ENOMEM) |
2588 | return VM_FAULT_OOM; | |
2589 | else | |
2590 | return VM_FAULT_SIGBUS; | |
1e8f889b DG |
2591 | } |
2592 | ||
0fe6e20b NH |
2593 | /* |
2594 | * When the original hugepage is shared one, it does not have | |
2595 | * anon_vma prepared. | |
2596 | */ | |
44e2aa93 | 2597 | if (unlikely(anon_vma_prepare(vma))) { |
ea4039a3 HD |
2598 | page_cache_release(new_page); |
2599 | page_cache_release(old_page); | |
44e2aa93 DN |
2600 | /* Caller expects lock to be held */ |
2601 | spin_lock(&mm->page_table_lock); | |
0fe6e20b | 2602 | return VM_FAULT_OOM; |
44e2aa93 | 2603 | } |
0fe6e20b | 2604 | |
47ad8475 AA |
2605 | copy_user_huge_page(new_page, old_page, address, vma, |
2606 | pages_per_huge_page(h)); | |
0ed361de | 2607 | __SetPageUptodate(new_page); |
1e8f889b | 2608 | |
2ec74c3e SG |
2609 | mmun_start = address & huge_page_mask(h); |
2610 | mmun_end = mmun_start + huge_page_size(h); | |
2611 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
b76c8cfb LW |
2612 | /* |
2613 | * Retake the page_table_lock to check for racing updates | |
2614 | * before the page tables are altered | |
2615 | */ | |
2616 | spin_lock(&mm->page_table_lock); | |
a5516438 | 2617 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
7f2e9525 | 2618 | if (likely(pte_same(huge_ptep_get(ptep), pte))) { |
1e8f889b | 2619 | /* Break COW */ |
8fe627ec | 2620 | huge_ptep_clear_flush(vma, address, ptep); |
1e8f889b DG |
2621 | set_huge_pte_at(mm, address, ptep, |
2622 | make_huge_pte(vma, new_page, 1)); | |
0fe6e20b | 2623 | page_remove_rmap(old_page); |
cd67f0d2 | 2624 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
2625 | /* Make the old page be freed below */ |
2626 | new_page = old_page; | |
2627 | } | |
2ec74c3e SG |
2628 | spin_unlock(&mm->page_table_lock); |
2629 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); | |
2630 | /* Caller expects lock to be held */ | |
2631 | spin_lock(&mm->page_table_lock); | |
1e8f889b DG |
2632 | page_cache_release(new_page); |
2633 | page_cache_release(old_page); | |
83c54070 | 2634 | return 0; |
1e8f889b DG |
2635 | } |
2636 | ||
04f2cbe3 | 2637 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
2638 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
2639 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
2640 | { |
2641 | struct address_space *mapping; | |
e7c4b0bf | 2642 | pgoff_t idx; |
04f2cbe3 MG |
2643 | |
2644 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 2645 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
2646 | |
2647 | return find_lock_page(mapping, idx); | |
2648 | } | |
2649 | ||
3ae77f43 HD |
2650 | /* |
2651 | * Return whether there is a pagecache page to back given address within VMA. | |
2652 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
2653 | */ | |
2654 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
2655 | struct vm_area_struct *vma, unsigned long address) |
2656 | { | |
2657 | struct address_space *mapping; | |
2658 | pgoff_t idx; | |
2659 | struct page *page; | |
2660 | ||
2661 | mapping = vma->vm_file->f_mapping; | |
2662 | idx = vma_hugecache_offset(h, vma, address); | |
2663 | ||
2664 | page = find_get_page(mapping, idx); | |
2665 | if (page) | |
2666 | put_page(page); | |
2667 | return page != NULL; | |
2668 | } | |
2669 | ||
a1ed3dda | 2670 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 2671 | unsigned long address, pte_t *ptep, unsigned int flags) |
ac9b9c66 | 2672 | { |
a5516438 | 2673 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 2674 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 2675 | int anon_rmap = 0; |
e7c4b0bf | 2676 | pgoff_t idx; |
4c887265 | 2677 | unsigned long size; |
4c887265 AL |
2678 | struct page *page; |
2679 | struct address_space *mapping; | |
1e8f889b | 2680 | pte_t new_pte; |
4c887265 | 2681 | |
04f2cbe3 MG |
2682 | /* |
2683 | * Currently, we are forced to kill the process in the event the | |
2684 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 2685 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
2686 | */ |
2687 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
ffb22af5 AM |
2688 | pr_warning("PID %d killed due to inadequate hugepage pool\n", |
2689 | current->pid); | |
04f2cbe3 MG |
2690 | return ret; |
2691 | } | |
2692 | ||
4c887265 | 2693 | mapping = vma->vm_file->f_mapping; |
a5516438 | 2694 | idx = vma_hugecache_offset(h, vma, address); |
4c887265 AL |
2695 | |
2696 | /* | |
2697 | * Use page lock to guard against racing truncation | |
2698 | * before we get page_table_lock. | |
2699 | */ | |
6bda666a CL |
2700 | retry: |
2701 | page = find_lock_page(mapping, idx); | |
2702 | if (!page) { | |
a5516438 | 2703 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
2704 | if (idx >= size) |
2705 | goto out; | |
04f2cbe3 | 2706 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 2707 | if (IS_ERR(page)) { |
76dcee75 AK |
2708 | ret = PTR_ERR(page); |
2709 | if (ret == -ENOMEM) | |
2710 | ret = VM_FAULT_OOM; | |
2711 | else | |
2712 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
2713 | goto out; |
2714 | } | |
47ad8475 | 2715 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 2716 | __SetPageUptodate(page); |
ac9b9c66 | 2717 | |
f83a275d | 2718 | if (vma->vm_flags & VM_MAYSHARE) { |
6bda666a | 2719 | int err; |
45c682a6 | 2720 | struct inode *inode = mapping->host; |
6bda666a CL |
2721 | |
2722 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
2723 | if (err) { | |
2724 | put_page(page); | |
6bda666a CL |
2725 | if (err == -EEXIST) |
2726 | goto retry; | |
2727 | goto out; | |
2728 | } | |
45c682a6 KC |
2729 | |
2730 | spin_lock(&inode->i_lock); | |
a5516438 | 2731 | inode->i_blocks += blocks_per_huge_page(h); |
45c682a6 | 2732 | spin_unlock(&inode->i_lock); |
23be7468 | 2733 | } else { |
6bda666a | 2734 | lock_page(page); |
0fe6e20b NH |
2735 | if (unlikely(anon_vma_prepare(vma))) { |
2736 | ret = VM_FAULT_OOM; | |
2737 | goto backout_unlocked; | |
2738 | } | |
409eb8c2 | 2739 | anon_rmap = 1; |
23be7468 | 2740 | } |
0fe6e20b | 2741 | } else { |
998b4382 NH |
2742 | /* |
2743 | * If memory error occurs between mmap() and fault, some process | |
2744 | * don't have hwpoisoned swap entry for errored virtual address. | |
2745 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
2746 | */ | |
2747 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 2748 | ret = VM_FAULT_HWPOISON | |
972dc4de | 2749 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
2750 | goto backout_unlocked; |
2751 | } | |
6bda666a | 2752 | } |
1e8f889b | 2753 | |
57303d80 AW |
2754 | /* |
2755 | * If we are going to COW a private mapping later, we examine the | |
2756 | * pending reservations for this page now. This will ensure that | |
2757 | * any allocations necessary to record that reservation occur outside | |
2758 | * the spinlock. | |
2759 | */ | |
788c7df4 | 2760 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) |
2b26736c AW |
2761 | if (vma_needs_reservation(h, vma, address) < 0) { |
2762 | ret = VM_FAULT_OOM; | |
2763 | goto backout_unlocked; | |
2764 | } | |
57303d80 | 2765 | |
ac9b9c66 | 2766 | spin_lock(&mm->page_table_lock); |
a5516438 | 2767 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
2768 | if (idx >= size) |
2769 | goto backout; | |
2770 | ||
83c54070 | 2771 | ret = 0; |
7f2e9525 | 2772 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
2773 | goto backout; |
2774 | ||
409eb8c2 HD |
2775 | if (anon_rmap) |
2776 | hugepage_add_new_anon_rmap(page, vma, address); | |
2777 | else | |
2778 | page_dup_rmap(page); | |
1e8f889b DG |
2779 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
2780 | && (vma->vm_flags & VM_SHARED))); | |
2781 | set_huge_pte_at(mm, address, ptep, new_pte); | |
2782 | ||
788c7df4 | 2783 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 2784 | /* Optimization, do the COW without a second fault */ |
04f2cbe3 | 2785 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); |
1e8f889b DG |
2786 | } |
2787 | ||
ac9b9c66 | 2788 | spin_unlock(&mm->page_table_lock); |
4c887265 AL |
2789 | unlock_page(page); |
2790 | out: | |
ac9b9c66 | 2791 | return ret; |
4c887265 AL |
2792 | |
2793 | backout: | |
2794 | spin_unlock(&mm->page_table_lock); | |
2b26736c | 2795 | backout_unlocked: |
4c887265 AL |
2796 | unlock_page(page); |
2797 | put_page(page); | |
2798 | goto out; | |
ac9b9c66 HD |
2799 | } |
2800 | ||
86e5216f | 2801 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 2802 | unsigned long address, unsigned int flags) |
86e5216f AL |
2803 | { |
2804 | pte_t *ptep; | |
2805 | pte_t entry; | |
1e8f889b | 2806 | int ret; |
0fe6e20b | 2807 | struct page *page = NULL; |
57303d80 | 2808 | struct page *pagecache_page = NULL; |
3935baa9 | 2809 | static DEFINE_MUTEX(hugetlb_instantiation_mutex); |
a5516438 | 2810 | struct hstate *h = hstate_vma(vma); |
86e5216f | 2811 | |
1e16a539 KH |
2812 | address &= huge_page_mask(h); |
2813 | ||
fd6a03ed NH |
2814 | ptep = huge_pte_offset(mm, address); |
2815 | if (ptep) { | |
2816 | entry = huge_ptep_get(ptep); | |
290408d4 | 2817 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
30dad309 | 2818 | migration_entry_wait_huge(mm, ptep); |
290408d4 NH |
2819 | return 0; |
2820 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 2821 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 2822 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
2823 | } |
2824 | ||
a5516438 | 2825 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); |
86e5216f AL |
2826 | if (!ptep) |
2827 | return VM_FAULT_OOM; | |
2828 | ||
3935baa9 DG |
2829 | /* |
2830 | * Serialize hugepage allocation and instantiation, so that we don't | |
2831 | * get spurious allocation failures if two CPUs race to instantiate | |
2832 | * the same page in the page cache. | |
2833 | */ | |
2834 | mutex_lock(&hugetlb_instantiation_mutex); | |
7f2e9525 GS |
2835 | entry = huge_ptep_get(ptep); |
2836 | if (huge_pte_none(entry)) { | |
788c7df4 | 2837 | ret = hugetlb_no_page(mm, vma, address, ptep, flags); |
b4d1d99f | 2838 | goto out_mutex; |
3935baa9 | 2839 | } |
86e5216f | 2840 | |
83c54070 | 2841 | ret = 0; |
1e8f889b | 2842 | |
57303d80 AW |
2843 | /* |
2844 | * If we are going to COW the mapping later, we examine the pending | |
2845 | * reservations for this page now. This will ensure that any | |
2846 | * allocations necessary to record that reservation occur outside the | |
2847 | * spinlock. For private mappings, we also lookup the pagecache | |
2848 | * page now as it is used to determine if a reservation has been | |
2849 | * consumed. | |
2850 | */ | |
106c992a | 2851 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
2b26736c AW |
2852 | if (vma_needs_reservation(h, vma, address) < 0) { |
2853 | ret = VM_FAULT_OOM; | |
b4d1d99f | 2854 | goto out_mutex; |
2b26736c | 2855 | } |
57303d80 | 2856 | |
f83a275d | 2857 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
2858 | pagecache_page = hugetlbfs_pagecache_page(h, |
2859 | vma, address); | |
2860 | } | |
2861 | ||
56c9cfb1 NH |
2862 | /* |
2863 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
2864 | * pagecache_page, so here we need take the former one | |
2865 | * when page != pagecache_page or !pagecache_page. | |
2866 | * Note that locking order is always pagecache_page -> page, | |
2867 | * so no worry about deadlock. | |
2868 | */ | |
2869 | page = pte_page(entry); | |
66aebce7 | 2870 | get_page(page); |
56c9cfb1 | 2871 | if (page != pagecache_page) |
0fe6e20b | 2872 | lock_page(page); |
0fe6e20b | 2873 | |
1e8f889b DG |
2874 | spin_lock(&mm->page_table_lock); |
2875 | /* Check for a racing update before calling hugetlb_cow */ | |
b4d1d99f DG |
2876 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) |
2877 | goto out_page_table_lock; | |
2878 | ||
2879 | ||
788c7df4 | 2880 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 2881 | if (!huge_pte_write(entry)) { |
57303d80 AW |
2882 | ret = hugetlb_cow(mm, vma, address, ptep, entry, |
2883 | pagecache_page); | |
b4d1d99f DG |
2884 | goto out_page_table_lock; |
2885 | } | |
106c992a | 2886 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
2887 | } |
2888 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
2889 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
2890 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 2891 | update_mmu_cache(vma, address, ptep); |
b4d1d99f DG |
2892 | |
2893 | out_page_table_lock: | |
1e8f889b | 2894 | spin_unlock(&mm->page_table_lock); |
57303d80 AW |
2895 | |
2896 | if (pagecache_page) { | |
2897 | unlock_page(pagecache_page); | |
2898 | put_page(pagecache_page); | |
2899 | } | |
1f64d69c DN |
2900 | if (page != pagecache_page) |
2901 | unlock_page(page); | |
66aebce7 | 2902 | put_page(page); |
57303d80 | 2903 | |
b4d1d99f | 2904 | out_mutex: |
3935baa9 | 2905 | mutex_unlock(&hugetlb_instantiation_mutex); |
1e8f889b DG |
2906 | |
2907 | return ret; | |
86e5216f AL |
2908 | } |
2909 | ||
28a35716 ML |
2910 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
2911 | struct page **pages, struct vm_area_struct **vmas, | |
2912 | unsigned long *position, unsigned long *nr_pages, | |
2913 | long i, unsigned int flags) | |
63551ae0 | 2914 | { |
d5d4b0aa KC |
2915 | unsigned long pfn_offset; |
2916 | unsigned long vaddr = *position; | |
28a35716 | 2917 | unsigned long remainder = *nr_pages; |
a5516438 | 2918 | struct hstate *h = hstate_vma(vma); |
63551ae0 | 2919 | |
1c59827d | 2920 | spin_lock(&mm->page_table_lock); |
63551ae0 | 2921 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 2922 | pte_t *pte; |
2a15efc9 | 2923 | int absent; |
4c887265 | 2924 | struct page *page; |
63551ae0 | 2925 | |
4c887265 AL |
2926 | /* |
2927 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 2928 | * each hugepage. We have to make sure we get the |
4c887265 AL |
2929 | * first, for the page indexing below to work. |
2930 | */ | |
a5516438 | 2931 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); |
2a15efc9 HD |
2932 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
2933 | ||
2934 | /* | |
2935 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
2936 | * an error where there's an empty slot with no huge pagecache |
2937 | * to back it. This way, we avoid allocating a hugepage, and | |
2938 | * the sparse dumpfile avoids allocating disk blocks, but its | |
2939 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 2940 | */ |
3ae77f43 HD |
2941 | if (absent && (flags & FOLL_DUMP) && |
2942 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
2a15efc9 HD |
2943 | remainder = 0; |
2944 | break; | |
2945 | } | |
63551ae0 | 2946 | |
9cc3a5bd NH |
2947 | /* |
2948 | * We need call hugetlb_fault for both hugepages under migration | |
2949 | * (in which case hugetlb_fault waits for the migration,) and | |
2950 | * hwpoisoned hugepages (in which case we need to prevent the | |
2951 | * caller from accessing to them.) In order to do this, we use | |
2952 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
2953 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
2954 | * both cases, and because we can't follow correct pages | |
2955 | * directly from any kind of swap entries. | |
2956 | */ | |
2957 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
2958 | ((flags & FOLL_WRITE) && |
2959 | !huge_pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 2960 | int ret; |
63551ae0 | 2961 | |
4c887265 | 2962 | spin_unlock(&mm->page_table_lock); |
2a15efc9 HD |
2963 | ret = hugetlb_fault(mm, vma, vaddr, |
2964 | (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); | |
4c887265 | 2965 | spin_lock(&mm->page_table_lock); |
a89182c7 | 2966 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 2967 | continue; |
63551ae0 | 2968 | |
4c887265 | 2969 | remainder = 0; |
4c887265 AL |
2970 | break; |
2971 | } | |
2972 | ||
a5516438 | 2973 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 2974 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 2975 | same_page: |
d6692183 | 2976 | if (pages) { |
2a15efc9 | 2977 | pages[i] = mem_map_offset(page, pfn_offset); |
4b2e38ad | 2978 | get_page(pages[i]); |
d6692183 | 2979 | } |
63551ae0 DG |
2980 | |
2981 | if (vmas) | |
2982 | vmas[i] = vma; | |
2983 | ||
2984 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 2985 | ++pfn_offset; |
63551ae0 DG |
2986 | --remainder; |
2987 | ++i; | |
d5d4b0aa | 2988 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 2989 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa KC |
2990 | /* |
2991 | * We use pfn_offset to avoid touching the pageframes | |
2992 | * of this compound page. | |
2993 | */ | |
2994 | goto same_page; | |
2995 | } | |
63551ae0 | 2996 | } |
1c59827d | 2997 | spin_unlock(&mm->page_table_lock); |
28a35716 | 2998 | *nr_pages = remainder; |
63551ae0 DG |
2999 | *position = vaddr; |
3000 | ||
2a15efc9 | 3001 | return i ? i : -EFAULT; |
63551ae0 | 3002 | } |
8f860591 | 3003 | |
7da4d641 | 3004 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
3005 | unsigned long address, unsigned long end, pgprot_t newprot) |
3006 | { | |
3007 | struct mm_struct *mm = vma->vm_mm; | |
3008 | unsigned long start = address; | |
3009 | pte_t *ptep; | |
3010 | pte_t pte; | |
a5516438 | 3011 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 3012 | unsigned long pages = 0; |
8f860591 ZY |
3013 | |
3014 | BUG_ON(address >= end); | |
3015 | flush_cache_range(vma, address, end); | |
3016 | ||
3d48ae45 | 3017 | mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex); |
8f860591 | 3018 | spin_lock(&mm->page_table_lock); |
a5516438 | 3019 | for (; address < end; address += huge_page_size(h)) { |
8f860591 ZY |
3020 | ptep = huge_pte_offset(mm, address); |
3021 | if (!ptep) | |
3022 | continue; | |
7da4d641 PZ |
3023 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3024 | pages++; | |
39dde65c | 3025 | continue; |
7da4d641 | 3026 | } |
7f2e9525 | 3027 | if (!huge_pte_none(huge_ptep_get(ptep))) { |
8f860591 | 3028 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
106c992a | 3029 | pte = pte_mkhuge(huge_pte_modify(pte, newprot)); |
be7517d6 | 3030 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
8f860591 | 3031 | set_huge_pte_at(mm, address, ptep, pte); |
7da4d641 | 3032 | pages++; |
8f860591 ZY |
3033 | } |
3034 | } | |
3035 | spin_unlock(&mm->page_table_lock); | |
d833352a MG |
3036 | /* |
3037 | * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare | |
3038 | * may have cleared our pud entry and done put_page on the page table: | |
3039 | * once we release i_mmap_mutex, another task can do the final put_page | |
3040 | * and that page table be reused and filled with junk. | |
3041 | */ | |
8f860591 | 3042 | flush_tlb_range(vma, start, end); |
d833352a | 3043 | mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex); |
7da4d641 PZ |
3044 | |
3045 | return pages << h->order; | |
8f860591 ZY |
3046 | } |
3047 | ||
a1e78772 MG |
3048 | int hugetlb_reserve_pages(struct inode *inode, |
3049 | long from, long to, | |
5a6fe125 | 3050 | struct vm_area_struct *vma, |
ca16d140 | 3051 | vm_flags_t vm_flags) |
e4e574b7 | 3052 | { |
17c9d12e | 3053 | long ret, chg; |
a5516438 | 3054 | struct hstate *h = hstate_inode(inode); |
90481622 | 3055 | struct hugepage_subpool *spool = subpool_inode(inode); |
e4e574b7 | 3056 | |
17c9d12e MG |
3057 | /* |
3058 | * Only apply hugepage reservation if asked. At fault time, an | |
3059 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 3060 | * without using reserves |
17c9d12e | 3061 | */ |
ca16d140 | 3062 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
3063 | return 0; |
3064 | ||
a1e78772 MG |
3065 | /* |
3066 | * Shared mappings base their reservation on the number of pages that | |
3067 | * are already allocated on behalf of the file. Private mappings need | |
3068 | * to reserve the full area even if read-only as mprotect() may be | |
3069 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
3070 | */ | |
f83a275d | 3071 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
a1e78772 | 3072 | chg = region_chg(&inode->i_mapping->private_list, from, to); |
17c9d12e MG |
3073 | else { |
3074 | struct resv_map *resv_map = resv_map_alloc(); | |
3075 | if (!resv_map) | |
3076 | return -ENOMEM; | |
3077 | ||
a1e78772 | 3078 | chg = to - from; |
84afd99b | 3079 | |
17c9d12e MG |
3080 | set_vma_resv_map(vma, resv_map); |
3081 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
3082 | } | |
3083 | ||
c50ac050 DH |
3084 | if (chg < 0) { |
3085 | ret = chg; | |
3086 | goto out_err; | |
3087 | } | |
8a630112 | 3088 | |
90481622 | 3089 | /* There must be enough pages in the subpool for the mapping */ |
c50ac050 DH |
3090 | if (hugepage_subpool_get_pages(spool, chg)) { |
3091 | ret = -ENOSPC; | |
3092 | goto out_err; | |
3093 | } | |
5a6fe125 MG |
3094 | |
3095 | /* | |
17c9d12e | 3096 | * Check enough hugepages are available for the reservation. |
90481622 | 3097 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 3098 | */ |
a5516438 | 3099 | ret = hugetlb_acct_memory(h, chg); |
68842c9b | 3100 | if (ret < 0) { |
90481622 | 3101 | hugepage_subpool_put_pages(spool, chg); |
c50ac050 | 3102 | goto out_err; |
68842c9b | 3103 | } |
17c9d12e MG |
3104 | |
3105 | /* | |
3106 | * Account for the reservations made. Shared mappings record regions | |
3107 | * that have reservations as they are shared by multiple VMAs. | |
3108 | * When the last VMA disappears, the region map says how much | |
3109 | * the reservation was and the page cache tells how much of | |
3110 | * the reservation was consumed. Private mappings are per-VMA and | |
3111 | * only the consumed reservations are tracked. When the VMA | |
3112 | * disappears, the original reservation is the VMA size and the | |
3113 | * consumed reservations are stored in the map. Hence, nothing | |
3114 | * else has to be done for private mappings here | |
3115 | */ | |
f83a275d | 3116 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
a1e78772 | 3117 | region_add(&inode->i_mapping->private_list, from, to); |
a43a8c39 | 3118 | return 0; |
c50ac050 | 3119 | out_err: |
4523e145 DH |
3120 | if (vma) |
3121 | resv_map_put(vma); | |
c50ac050 | 3122 | return ret; |
a43a8c39 KC |
3123 | } |
3124 | ||
3125 | void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) | |
3126 | { | |
a5516438 | 3127 | struct hstate *h = hstate_inode(inode); |
a43a8c39 | 3128 | long chg = region_truncate(&inode->i_mapping->private_list, offset); |
90481622 | 3129 | struct hugepage_subpool *spool = subpool_inode(inode); |
45c682a6 KC |
3130 | |
3131 | spin_lock(&inode->i_lock); | |
e4c6f8be | 3132 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
3133 | spin_unlock(&inode->i_lock); |
3134 | ||
90481622 | 3135 | hugepage_subpool_put_pages(spool, (chg - freed)); |
a5516438 | 3136 | hugetlb_acct_memory(h, -(chg - freed)); |
a43a8c39 | 3137 | } |
93f70f90 | 3138 | |
3212b535 SC |
3139 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
3140 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
3141 | struct vm_area_struct *vma, | |
3142 | unsigned long addr, pgoff_t idx) | |
3143 | { | |
3144 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
3145 | svma->vm_start; | |
3146 | unsigned long sbase = saddr & PUD_MASK; | |
3147 | unsigned long s_end = sbase + PUD_SIZE; | |
3148 | ||
3149 | /* Allow segments to share if only one is marked locked */ | |
3150 | unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED; | |
3151 | unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED; | |
3152 | ||
3153 | /* | |
3154 | * match the virtual addresses, permission and the alignment of the | |
3155 | * page table page. | |
3156 | */ | |
3157 | if (pmd_index(addr) != pmd_index(saddr) || | |
3158 | vm_flags != svm_flags || | |
3159 | sbase < svma->vm_start || svma->vm_end < s_end) | |
3160 | return 0; | |
3161 | ||
3162 | return saddr; | |
3163 | } | |
3164 | ||
3165 | static int vma_shareable(struct vm_area_struct *vma, unsigned long addr) | |
3166 | { | |
3167 | unsigned long base = addr & PUD_MASK; | |
3168 | unsigned long end = base + PUD_SIZE; | |
3169 | ||
3170 | /* | |
3171 | * check on proper vm_flags and page table alignment | |
3172 | */ | |
3173 | if (vma->vm_flags & VM_MAYSHARE && | |
3174 | vma->vm_start <= base && end <= vma->vm_end) | |
3175 | return 1; | |
3176 | return 0; | |
3177 | } | |
3178 | ||
3179 | /* | |
3180 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
3181 | * and returns the corresponding pte. While this is not necessary for the | |
3182 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
3183 | * code much cleaner. pmd allocation is essential for the shared case because | |
3184 | * pud has to be populated inside the same i_mmap_mutex section - otherwise | |
3185 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a | |
3186 | * bad pmd for sharing. | |
3187 | */ | |
3188 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
3189 | { | |
3190 | struct vm_area_struct *vma = find_vma(mm, addr); | |
3191 | struct address_space *mapping = vma->vm_file->f_mapping; | |
3192 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
3193 | vma->vm_pgoff; | |
3194 | struct vm_area_struct *svma; | |
3195 | unsigned long saddr; | |
3196 | pte_t *spte = NULL; | |
3197 | pte_t *pte; | |
3198 | ||
3199 | if (!vma_shareable(vma, addr)) | |
3200 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
3201 | ||
3202 | mutex_lock(&mapping->i_mmap_mutex); | |
3203 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { | |
3204 | if (svma == vma) | |
3205 | continue; | |
3206 | ||
3207 | saddr = page_table_shareable(svma, vma, addr, idx); | |
3208 | if (saddr) { | |
3209 | spte = huge_pte_offset(svma->vm_mm, saddr); | |
3210 | if (spte) { | |
3211 | get_page(virt_to_page(spte)); | |
3212 | break; | |
3213 | } | |
3214 | } | |
3215 | } | |
3216 | ||
3217 | if (!spte) | |
3218 | goto out; | |
3219 | ||
3220 | spin_lock(&mm->page_table_lock); | |
3221 | if (pud_none(*pud)) | |
3222 | pud_populate(mm, pud, | |
3223 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
3224 | else | |
3225 | put_page(virt_to_page(spte)); | |
3226 | spin_unlock(&mm->page_table_lock); | |
3227 | out: | |
3228 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3229 | mutex_unlock(&mapping->i_mmap_mutex); | |
3230 | return pte; | |
3231 | } | |
3232 | ||
3233 | /* | |
3234 | * unmap huge page backed by shared pte. | |
3235 | * | |
3236 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
3237 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
3238 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
3239 | * | |
3240 | * called with vma->vm_mm->page_table_lock held. | |
3241 | * | |
3242 | * returns: 1 successfully unmapped a shared pte page | |
3243 | * 0 the underlying pte page is not shared, or it is the last user | |
3244 | */ | |
3245 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
3246 | { | |
3247 | pgd_t *pgd = pgd_offset(mm, *addr); | |
3248 | pud_t *pud = pud_offset(pgd, *addr); | |
3249 | ||
3250 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
3251 | if (page_count(virt_to_page(ptep)) == 1) | |
3252 | return 0; | |
3253 | ||
3254 | pud_clear(pud); | |
3255 | put_page(virt_to_page(ptep)); | |
3256 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; | |
3257 | return 1; | |
3258 | } | |
9e5fc74c SC |
3259 | #define want_pmd_share() (1) |
3260 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
3261 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
3262 | { | |
3263 | return NULL; | |
3264 | } | |
3265 | #define want_pmd_share() (0) | |
3212b535 SC |
3266 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
3267 | ||
9e5fc74c SC |
3268 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
3269 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
3270 | unsigned long addr, unsigned long sz) | |
3271 | { | |
3272 | pgd_t *pgd; | |
3273 | pud_t *pud; | |
3274 | pte_t *pte = NULL; | |
3275 | ||
3276 | pgd = pgd_offset(mm, addr); | |
3277 | pud = pud_alloc(mm, pgd, addr); | |
3278 | if (pud) { | |
3279 | if (sz == PUD_SIZE) { | |
3280 | pte = (pte_t *)pud; | |
3281 | } else { | |
3282 | BUG_ON(sz != PMD_SIZE); | |
3283 | if (want_pmd_share() && pud_none(*pud)) | |
3284 | pte = huge_pmd_share(mm, addr, pud); | |
3285 | else | |
3286 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3287 | } | |
3288 | } | |
3289 | BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); | |
3290 | ||
3291 | return pte; | |
3292 | } | |
3293 | ||
3294 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) | |
3295 | { | |
3296 | pgd_t *pgd; | |
3297 | pud_t *pud; | |
3298 | pmd_t *pmd = NULL; | |
3299 | ||
3300 | pgd = pgd_offset(mm, addr); | |
3301 | if (pgd_present(*pgd)) { | |
3302 | pud = pud_offset(pgd, addr); | |
3303 | if (pud_present(*pud)) { | |
3304 | if (pud_huge(*pud)) | |
3305 | return (pte_t *)pud; | |
3306 | pmd = pmd_offset(pud, addr); | |
3307 | } | |
3308 | } | |
3309 | return (pte_t *) pmd; | |
3310 | } | |
3311 | ||
3312 | struct page * | |
3313 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, | |
3314 | pmd_t *pmd, int write) | |
3315 | { | |
3316 | struct page *page; | |
3317 | ||
3318 | page = pte_page(*(pte_t *)pmd); | |
3319 | if (page) | |
3320 | page += ((address & ~PMD_MASK) >> PAGE_SHIFT); | |
3321 | return page; | |
3322 | } | |
3323 | ||
3324 | struct page * | |
3325 | follow_huge_pud(struct mm_struct *mm, unsigned long address, | |
3326 | pud_t *pud, int write) | |
3327 | { | |
3328 | struct page *page; | |
3329 | ||
3330 | page = pte_page(*(pte_t *)pud); | |
3331 | if (page) | |
3332 | page += ((address & ~PUD_MASK) >> PAGE_SHIFT); | |
3333 | return page; | |
3334 | } | |
3335 | ||
3336 | #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */ | |
3337 | ||
3338 | /* Can be overriden by architectures */ | |
3339 | __attribute__((weak)) struct page * | |
3340 | follow_huge_pud(struct mm_struct *mm, unsigned long address, | |
3341 | pud_t *pud, int write) | |
3342 | { | |
3343 | BUG(); | |
3344 | return NULL; | |
3345 | } | |
3346 | ||
3347 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ | |
3348 | ||
d5bd9106 AK |
3349 | #ifdef CONFIG_MEMORY_FAILURE |
3350 | ||
6de2b1aa NH |
3351 | /* Should be called in hugetlb_lock */ |
3352 | static int is_hugepage_on_freelist(struct page *hpage) | |
3353 | { | |
3354 | struct page *page; | |
3355 | struct page *tmp; | |
3356 | struct hstate *h = page_hstate(hpage); | |
3357 | int nid = page_to_nid(hpage); | |
3358 | ||
3359 | list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru) | |
3360 | if (page == hpage) | |
3361 | return 1; | |
3362 | return 0; | |
3363 | } | |
3364 | ||
93f70f90 NH |
3365 | /* |
3366 | * This function is called from memory failure code. | |
3367 | * Assume the caller holds page lock of the head page. | |
3368 | */ | |
6de2b1aa | 3369 | int dequeue_hwpoisoned_huge_page(struct page *hpage) |
93f70f90 NH |
3370 | { |
3371 | struct hstate *h = page_hstate(hpage); | |
3372 | int nid = page_to_nid(hpage); | |
6de2b1aa | 3373 | int ret = -EBUSY; |
93f70f90 NH |
3374 | |
3375 | spin_lock(&hugetlb_lock); | |
6de2b1aa | 3376 | if (is_hugepage_on_freelist(hpage)) { |
56f2fb14 NH |
3377 | /* |
3378 | * Hwpoisoned hugepage isn't linked to activelist or freelist, | |
3379 | * but dangling hpage->lru can trigger list-debug warnings | |
3380 | * (this happens when we call unpoison_memory() on it), | |
3381 | * so let it point to itself with list_del_init(). | |
3382 | */ | |
3383 | list_del_init(&hpage->lru); | |
8c6c2ecb | 3384 | set_page_refcounted(hpage); |
6de2b1aa NH |
3385 | h->free_huge_pages--; |
3386 | h->free_huge_pages_node[nid]--; | |
3387 | ret = 0; | |
3388 | } | |
93f70f90 | 3389 | spin_unlock(&hugetlb_lock); |
6de2b1aa | 3390 | return ret; |
93f70f90 | 3391 | } |
6de2b1aa | 3392 | #endif |