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> | |
c8721bbb | 24 | #include <linux/page-isolation.h> |
d6606683 | 25 | |
63551ae0 DG |
26 | #include <asm/page.h> |
27 | #include <asm/pgtable.h> | |
24669e58 | 28 | #include <asm/tlb.h> |
63551ae0 | 29 | |
24669e58 | 30 | #include <linux/io.h> |
63551ae0 | 31 | #include <linux/hugetlb.h> |
9dd540e2 | 32 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 33 | #include <linux/node.h> |
7835e98b | 34 | #include "internal.h" |
1da177e4 LT |
35 | |
36 | const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL; | |
396faf03 | 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 | 50 | /* |
31caf665 NH |
51 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
52 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 53 | */ |
c3f38a38 | 54 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 55 | |
90481622 DG |
56 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
57 | { | |
58 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
59 | ||
60 | spin_unlock(&spool->lock); | |
61 | ||
62 | /* If no pages are used, and no other handles to the subpool | |
63 | * remain, free the subpool the subpool remain */ | |
64 | if (free) | |
65 | kfree(spool); | |
66 | } | |
67 | ||
68 | struct hugepage_subpool *hugepage_new_subpool(long nr_blocks) | |
69 | { | |
70 | struct hugepage_subpool *spool; | |
71 | ||
72 | spool = kmalloc(sizeof(*spool), GFP_KERNEL); | |
73 | if (!spool) | |
74 | return NULL; | |
75 | ||
76 | spin_lock_init(&spool->lock); | |
77 | spool->count = 1; | |
78 | spool->max_hpages = nr_blocks; | |
79 | spool->used_hpages = 0; | |
80 | ||
81 | return spool; | |
82 | } | |
83 | ||
84 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
85 | { | |
86 | spin_lock(&spool->lock); | |
87 | BUG_ON(!spool->count); | |
88 | spool->count--; | |
89 | unlock_or_release_subpool(spool); | |
90 | } | |
91 | ||
92 | static int hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
93 | long delta) | |
94 | { | |
95 | int ret = 0; | |
96 | ||
97 | if (!spool) | |
98 | return 0; | |
99 | ||
100 | spin_lock(&spool->lock); | |
101 | if ((spool->used_hpages + delta) <= spool->max_hpages) { | |
102 | spool->used_hpages += delta; | |
103 | } else { | |
104 | ret = -ENOMEM; | |
105 | } | |
106 | spin_unlock(&spool->lock); | |
107 | ||
108 | return ret; | |
109 | } | |
110 | ||
111 | static void hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
112 | long delta) | |
113 | { | |
114 | if (!spool) | |
115 | return; | |
116 | ||
117 | spin_lock(&spool->lock); | |
118 | spool->used_hpages -= delta; | |
119 | /* If hugetlbfs_put_super couldn't free spool due to | |
120 | * an outstanding quota reference, free it now. */ | |
121 | unlock_or_release_subpool(spool); | |
122 | } | |
123 | ||
124 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
125 | { | |
126 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
127 | } | |
128 | ||
129 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
130 | { | |
496ad9aa | 131 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
132 | } |
133 | ||
96822904 AW |
134 | /* |
135 | * Region tracking -- allows tracking of reservations and instantiated pages | |
136 | * across the pages in a mapping. | |
84afd99b AW |
137 | * |
138 | * The region data structures are protected by a combination of the mmap_sem | |
c748c262 | 139 | * and the hugetlb_instantiation_mutex. To access or modify a region the caller |
84afd99b | 140 | * must either hold the mmap_sem for write, or the mmap_sem for read and |
c748c262 | 141 | * the hugetlb_instantiation_mutex: |
84afd99b | 142 | * |
32f84528 | 143 | * down_write(&mm->mmap_sem); |
84afd99b | 144 | * or |
32f84528 CF |
145 | * down_read(&mm->mmap_sem); |
146 | * mutex_lock(&hugetlb_instantiation_mutex); | |
96822904 AW |
147 | */ |
148 | struct file_region { | |
149 | struct list_head link; | |
150 | long from; | |
151 | long to; | |
152 | }; | |
153 | ||
154 | static long region_add(struct list_head *head, long f, long t) | |
155 | { | |
156 | struct file_region *rg, *nrg, *trg; | |
157 | ||
158 | /* Locate the region we are either in or before. */ | |
159 | list_for_each_entry(rg, head, link) | |
160 | if (f <= rg->to) | |
161 | break; | |
162 | ||
163 | /* Round our left edge to the current segment if it encloses us. */ | |
164 | if (f > rg->from) | |
165 | f = rg->from; | |
166 | ||
167 | /* Check for and consume any regions we now overlap with. */ | |
168 | nrg = rg; | |
169 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
170 | if (&rg->link == head) | |
171 | break; | |
172 | if (rg->from > t) | |
173 | break; | |
174 | ||
175 | /* If this area reaches higher then extend our area to | |
176 | * include it completely. If this is not the first area | |
177 | * which we intend to reuse, free it. */ | |
178 | if (rg->to > t) | |
179 | t = rg->to; | |
180 | if (rg != nrg) { | |
181 | list_del(&rg->link); | |
182 | kfree(rg); | |
183 | } | |
184 | } | |
185 | nrg->from = f; | |
186 | nrg->to = t; | |
187 | return 0; | |
188 | } | |
189 | ||
190 | static long region_chg(struct list_head *head, long f, long t) | |
191 | { | |
192 | struct file_region *rg, *nrg; | |
193 | long chg = 0; | |
194 | ||
195 | /* Locate the region we are before or in. */ | |
196 | list_for_each_entry(rg, head, link) | |
197 | if (f <= rg->to) | |
198 | break; | |
199 | ||
200 | /* If we are below the current region then a new region is required. | |
201 | * Subtle, allocate a new region at the position but make it zero | |
202 | * size such that we can guarantee to record the reservation. */ | |
203 | if (&rg->link == head || t < rg->from) { | |
204 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
205 | if (!nrg) | |
206 | return -ENOMEM; | |
207 | nrg->from = f; | |
208 | nrg->to = f; | |
209 | INIT_LIST_HEAD(&nrg->link); | |
210 | list_add(&nrg->link, rg->link.prev); | |
211 | ||
212 | return t - f; | |
213 | } | |
214 | ||
215 | /* Round our left edge to the current segment if it encloses us. */ | |
216 | if (f > rg->from) | |
217 | f = rg->from; | |
218 | chg = t - f; | |
219 | ||
220 | /* Check for and consume any regions we now overlap with. */ | |
221 | list_for_each_entry(rg, rg->link.prev, link) { | |
222 | if (&rg->link == head) | |
223 | break; | |
224 | if (rg->from > t) | |
225 | return chg; | |
226 | ||
25985edc | 227 | /* We overlap with this area, if it extends further than |
96822904 AW |
228 | * us then we must extend ourselves. Account for its |
229 | * existing reservation. */ | |
230 | if (rg->to > t) { | |
231 | chg += rg->to - t; | |
232 | t = rg->to; | |
233 | } | |
234 | chg -= rg->to - rg->from; | |
235 | } | |
236 | return chg; | |
237 | } | |
238 | ||
239 | static long region_truncate(struct list_head *head, long end) | |
240 | { | |
241 | struct file_region *rg, *trg; | |
242 | long chg = 0; | |
243 | ||
244 | /* Locate the region we are either in or before. */ | |
245 | list_for_each_entry(rg, head, link) | |
246 | if (end <= rg->to) | |
247 | break; | |
248 | if (&rg->link == head) | |
249 | return 0; | |
250 | ||
251 | /* If we are in the middle of a region then adjust it. */ | |
252 | if (end > rg->from) { | |
253 | chg = rg->to - end; | |
254 | rg->to = end; | |
255 | rg = list_entry(rg->link.next, typeof(*rg), link); | |
256 | } | |
257 | ||
258 | /* Drop any remaining regions. */ | |
259 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
260 | if (&rg->link == head) | |
261 | break; | |
262 | chg += rg->to - rg->from; | |
263 | list_del(&rg->link); | |
264 | kfree(rg); | |
265 | } | |
266 | return chg; | |
267 | } | |
268 | ||
84afd99b AW |
269 | static long region_count(struct list_head *head, long f, long t) |
270 | { | |
271 | struct file_region *rg; | |
272 | long chg = 0; | |
273 | ||
274 | /* Locate each segment we overlap with, and count that overlap. */ | |
275 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
276 | long seg_from; |
277 | long seg_to; | |
84afd99b AW |
278 | |
279 | if (rg->to <= f) | |
280 | continue; | |
281 | if (rg->from >= t) | |
282 | break; | |
283 | ||
284 | seg_from = max(rg->from, f); | |
285 | seg_to = min(rg->to, t); | |
286 | ||
287 | chg += seg_to - seg_from; | |
288 | } | |
289 | ||
290 | return chg; | |
291 | } | |
292 | ||
e7c4b0bf AW |
293 | /* |
294 | * Convert the address within this vma to the page offset within | |
295 | * the mapping, in pagecache page units; huge pages here. | |
296 | */ | |
a5516438 AK |
297 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
298 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 299 | { |
a5516438 AK |
300 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
301 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
302 | } |
303 | ||
0fe6e20b NH |
304 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
305 | unsigned long address) | |
306 | { | |
307 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
308 | } | |
309 | ||
08fba699 MG |
310 | /* |
311 | * Return the size of the pages allocated when backing a VMA. In the majority | |
312 | * cases this will be same size as used by the page table entries. | |
313 | */ | |
314 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
315 | { | |
316 | struct hstate *hstate; | |
317 | ||
318 | if (!is_vm_hugetlb_page(vma)) | |
319 | return PAGE_SIZE; | |
320 | ||
321 | hstate = hstate_vma(vma); | |
322 | ||
2415cf12 | 323 | return 1UL << huge_page_shift(hstate); |
08fba699 | 324 | } |
f340ca0f | 325 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 326 | |
3340289d MG |
327 | /* |
328 | * Return the page size being used by the MMU to back a VMA. In the majority | |
329 | * of cases, the page size used by the kernel matches the MMU size. On | |
330 | * architectures where it differs, an architecture-specific version of this | |
331 | * function is required. | |
332 | */ | |
333 | #ifndef vma_mmu_pagesize | |
334 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) | |
335 | { | |
336 | return vma_kernel_pagesize(vma); | |
337 | } | |
338 | #endif | |
339 | ||
84afd99b AW |
340 | /* |
341 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
342 | * bits of the reservation map pointer, which are always clear due to | |
343 | * alignment. | |
344 | */ | |
345 | #define HPAGE_RESV_OWNER (1UL << 0) | |
346 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 347 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 348 | |
a1e78772 MG |
349 | /* |
350 | * These helpers are used to track how many pages are reserved for | |
351 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
352 | * is guaranteed to have their future faults succeed. | |
353 | * | |
354 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
355 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
356 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
357 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
358 | * |
359 | * The private mapping reservation is represented in a subtly different | |
360 | * manner to a shared mapping. A shared mapping has a region map associated | |
361 | * with the underlying file, this region map represents the backing file | |
362 | * pages which have ever had a reservation assigned which this persists even | |
363 | * after the page is instantiated. A private mapping has a region map | |
364 | * associated with the original mmap which is attached to all VMAs which | |
365 | * reference it, this region map represents those offsets which have consumed | |
366 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 367 | */ |
e7c4b0bf AW |
368 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
369 | { | |
370 | return (unsigned long)vma->vm_private_data; | |
371 | } | |
372 | ||
373 | static void set_vma_private_data(struct vm_area_struct *vma, | |
374 | unsigned long value) | |
375 | { | |
376 | vma->vm_private_data = (void *)value; | |
377 | } | |
378 | ||
9119a41e | 379 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
380 | { |
381 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
382 | if (!resv_map) | |
383 | return NULL; | |
384 | ||
385 | kref_init(&resv_map->refs); | |
386 | INIT_LIST_HEAD(&resv_map->regions); | |
387 | ||
388 | return resv_map; | |
389 | } | |
390 | ||
9119a41e | 391 | void resv_map_release(struct kref *ref) |
84afd99b AW |
392 | { |
393 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
394 | ||
395 | /* Clear out any active regions before we release the map. */ | |
396 | region_truncate(&resv_map->regions, 0); | |
397 | kfree(resv_map); | |
398 | } | |
399 | ||
400 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) | |
a1e78772 MG |
401 | { |
402 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 403 | if (!(vma->vm_flags & VM_MAYSHARE)) |
84afd99b AW |
404 | return (struct resv_map *)(get_vma_private_data(vma) & |
405 | ~HPAGE_RESV_MASK); | |
2a4b3ded | 406 | return NULL; |
a1e78772 MG |
407 | } |
408 | ||
84afd99b | 409 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 MG |
410 | { |
411 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 412 | VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); |
a1e78772 | 413 | |
84afd99b AW |
414 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
415 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
416 | } |
417 | ||
418 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
419 | { | |
04f2cbe3 | 420 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); |
f83a275d | 421 | VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); |
e7c4b0bf AW |
422 | |
423 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
424 | } |
425 | ||
426 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
427 | { | |
428 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
e7c4b0bf AW |
429 | |
430 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
431 | } |
432 | ||
04f2cbe3 | 433 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
434 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
435 | { | |
436 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 437 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
438 | vma->vm_private_data = (void *)0; |
439 | } | |
440 | ||
441 | /* Returns true if the VMA has associated reserve pages */ | |
af0ed73e | 442 | static int vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 443 | { |
af0ed73e JK |
444 | if (vma->vm_flags & VM_NORESERVE) { |
445 | /* | |
446 | * This address is already reserved by other process(chg == 0), | |
447 | * so, we should decrement reserved count. Without decrementing, | |
448 | * reserve count remains after releasing inode, because this | |
449 | * allocated page will go into page cache and is regarded as | |
450 | * coming from reserved pool in releasing step. Currently, we | |
451 | * don't have any other solution to deal with this situation | |
452 | * properly, so add work-around here. | |
453 | */ | |
454 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
455 | return 1; | |
456 | else | |
457 | return 0; | |
458 | } | |
a63884e9 JK |
459 | |
460 | /* Shared mappings always use reserves */ | |
f83a275d | 461 | if (vma->vm_flags & VM_MAYSHARE) |
7f09ca51 | 462 | return 1; |
a63884e9 JK |
463 | |
464 | /* | |
465 | * Only the process that called mmap() has reserves for | |
466 | * private mappings. | |
467 | */ | |
7f09ca51 MG |
468 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
469 | return 1; | |
a63884e9 | 470 | |
7f09ca51 | 471 | return 0; |
a1e78772 MG |
472 | } |
473 | ||
a5516438 | 474 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
475 | { |
476 | int nid = page_to_nid(page); | |
0edaecfa | 477 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
478 | h->free_huge_pages++; |
479 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
480 | } |
481 | ||
bf50bab2 NH |
482 | static struct page *dequeue_huge_page_node(struct hstate *h, int nid) |
483 | { | |
484 | struct page *page; | |
485 | ||
c8721bbb NH |
486 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
487 | if (!is_migrate_isolate_page(page)) | |
488 | break; | |
489 | /* | |
490 | * if 'non-isolated free hugepage' not found on the list, | |
491 | * the allocation fails. | |
492 | */ | |
493 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 494 | return NULL; |
0edaecfa | 495 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 496 | set_page_refcounted(page); |
bf50bab2 NH |
497 | h->free_huge_pages--; |
498 | h->free_huge_pages_node[nid]--; | |
499 | return page; | |
500 | } | |
501 | ||
86cdb465 NH |
502 | /* Movability of hugepages depends on migration support. */ |
503 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
504 | { | |
505 | if (hugepages_treat_as_movable || hugepage_migration_support(h)) | |
506 | return GFP_HIGHUSER_MOVABLE; | |
507 | else | |
508 | return GFP_HIGHUSER; | |
509 | } | |
510 | ||
a5516438 AK |
511 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
512 | struct vm_area_struct *vma, | |
af0ed73e JK |
513 | unsigned long address, int avoid_reserve, |
514 | long chg) | |
1da177e4 | 515 | { |
b1c12cbc | 516 | struct page *page = NULL; |
480eccf9 | 517 | struct mempolicy *mpol; |
19770b32 | 518 | nodemask_t *nodemask; |
c0ff7453 | 519 | struct zonelist *zonelist; |
dd1a239f MG |
520 | struct zone *zone; |
521 | struct zoneref *z; | |
cc9a6c87 | 522 | unsigned int cpuset_mems_cookie; |
1da177e4 | 523 | |
a1e78772 MG |
524 | /* |
525 | * A child process with MAP_PRIVATE mappings created by their parent | |
526 | * have no page reserves. This check ensures that reservations are | |
527 | * not "stolen". The child may still get SIGKILLed | |
528 | */ | |
af0ed73e | 529 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 530 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 531 | goto err; |
a1e78772 | 532 | |
04f2cbe3 | 533 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 534 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 535 | goto err; |
04f2cbe3 | 536 | |
9966c4bb | 537 | retry_cpuset: |
d26914d1 | 538 | cpuset_mems_cookie = read_mems_allowed_begin(); |
9966c4bb | 539 | zonelist = huge_zonelist(vma, address, |
86cdb465 | 540 | htlb_alloc_mask(h), &mpol, &nodemask); |
9966c4bb | 541 | |
19770b32 MG |
542 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
543 | MAX_NR_ZONES - 1, nodemask) { | |
86cdb465 | 544 | if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask(h))) { |
bf50bab2 NH |
545 | page = dequeue_huge_page_node(h, zone_to_nid(zone)); |
546 | if (page) { | |
af0ed73e JK |
547 | if (avoid_reserve) |
548 | break; | |
549 | if (!vma_has_reserves(vma, chg)) | |
550 | break; | |
551 | ||
07443a85 | 552 | SetPagePrivate(page); |
af0ed73e | 553 | h->resv_huge_pages--; |
bf50bab2 NH |
554 | break; |
555 | } | |
3abf7afd | 556 | } |
1da177e4 | 557 | } |
cc9a6c87 | 558 | |
52cd3b07 | 559 | mpol_cond_put(mpol); |
d26914d1 | 560 | if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) |
cc9a6c87 | 561 | goto retry_cpuset; |
1da177e4 | 562 | return page; |
cc9a6c87 MG |
563 | |
564 | err: | |
cc9a6c87 | 565 | return NULL; |
1da177e4 LT |
566 | } |
567 | ||
a5516438 | 568 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
569 | { |
570 | int i; | |
a5516438 | 571 | |
18229df5 AW |
572 | VM_BUG_ON(h->order >= MAX_ORDER); |
573 | ||
a5516438 AK |
574 | h->nr_huge_pages--; |
575 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
576 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
577 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
578 | 1 << PG_referenced | 1 << PG_dirty | | |
579 | 1 << PG_active | 1 << PG_reserved | | |
580 | 1 << PG_private | 1 << PG_writeback); | |
6af2acb6 | 581 | } |
309381fe | 582 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
6af2acb6 AL |
583 | set_compound_page_dtor(page, NULL); |
584 | set_page_refcounted(page); | |
7f2e9525 | 585 | arch_release_hugepage(page); |
a5516438 | 586 | __free_pages(page, huge_page_order(h)); |
6af2acb6 AL |
587 | } |
588 | ||
e5ff2159 AK |
589 | struct hstate *size_to_hstate(unsigned long size) |
590 | { | |
591 | struct hstate *h; | |
592 | ||
593 | for_each_hstate(h) { | |
594 | if (huge_page_size(h) == size) | |
595 | return h; | |
596 | } | |
597 | return NULL; | |
598 | } | |
599 | ||
27a85ef1 DG |
600 | static void free_huge_page(struct page *page) |
601 | { | |
a5516438 AK |
602 | /* |
603 | * Can't pass hstate in here because it is called from the | |
604 | * compound page destructor. | |
605 | */ | |
e5ff2159 | 606 | struct hstate *h = page_hstate(page); |
7893d1d5 | 607 | int nid = page_to_nid(page); |
90481622 DG |
608 | struct hugepage_subpool *spool = |
609 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 610 | bool restore_reserve; |
27a85ef1 | 611 | |
e5df70ab | 612 | set_page_private(page, 0); |
23be7468 | 613 | page->mapping = NULL; |
7893d1d5 | 614 | BUG_ON(page_count(page)); |
0fe6e20b | 615 | BUG_ON(page_mapcount(page)); |
07443a85 | 616 | restore_reserve = PagePrivate(page); |
16c794b4 | 617 | ClearPagePrivate(page); |
27a85ef1 DG |
618 | |
619 | spin_lock(&hugetlb_lock); | |
6d76dcf4 AK |
620 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
621 | pages_per_huge_page(h), page); | |
07443a85 JK |
622 | if (restore_reserve) |
623 | h->resv_huge_pages++; | |
624 | ||
aa888a74 | 625 | if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { |
0edaecfa AK |
626 | /* remove the page from active list */ |
627 | list_del(&page->lru); | |
a5516438 AK |
628 | update_and_free_page(h, page); |
629 | h->surplus_huge_pages--; | |
630 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 631 | } else { |
5d3a551c | 632 | arch_clear_hugepage_flags(page); |
a5516438 | 633 | enqueue_huge_page(h, page); |
7893d1d5 | 634 | } |
27a85ef1 | 635 | spin_unlock(&hugetlb_lock); |
90481622 | 636 | hugepage_subpool_put_pages(spool, 1); |
27a85ef1 DG |
637 | } |
638 | ||
a5516438 | 639 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 640 | { |
0edaecfa | 641 | INIT_LIST_HEAD(&page->lru); |
b7ba30c6 AK |
642 | set_compound_page_dtor(page, free_huge_page); |
643 | spin_lock(&hugetlb_lock); | |
9dd540e2 | 644 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
645 | h->nr_huge_pages++; |
646 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 AK |
647 | spin_unlock(&hugetlb_lock); |
648 | put_page(page); /* free it into the hugepage allocator */ | |
649 | } | |
650 | ||
20a0307c WF |
651 | static void prep_compound_gigantic_page(struct page *page, unsigned long order) |
652 | { | |
653 | int i; | |
654 | int nr_pages = 1 << order; | |
655 | struct page *p = page + 1; | |
656 | ||
657 | /* we rely on prep_new_huge_page to set the destructor */ | |
658 | set_compound_order(page, order); | |
659 | __SetPageHead(page); | |
ef5a22be | 660 | __ClearPageReserved(page); |
20a0307c WF |
661 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
662 | __SetPageTail(p); | |
ef5a22be AA |
663 | /* |
664 | * For gigantic hugepages allocated through bootmem at | |
665 | * boot, it's safer to be consistent with the not-gigantic | |
666 | * hugepages and clear the PG_reserved bit from all tail pages | |
667 | * too. Otherwse drivers using get_user_pages() to access tail | |
668 | * pages may get the reference counting wrong if they see | |
669 | * PG_reserved set on a tail page (despite the head page not | |
670 | * having PG_reserved set). Enforcing this consistency between | |
671 | * head and tail pages allows drivers to optimize away a check | |
672 | * on the head page when they need know if put_page() is needed | |
673 | * after get_user_pages(). | |
674 | */ | |
675 | __ClearPageReserved(p); | |
58a84aa9 | 676 | set_page_count(p, 0); |
20a0307c WF |
677 | p->first_page = page; |
678 | } | |
679 | } | |
680 | ||
7795912c AM |
681 | /* |
682 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
683 | * transparent huge pages. See the PageTransHuge() documentation for more | |
684 | * details. | |
685 | */ | |
20a0307c WF |
686 | int PageHuge(struct page *page) |
687 | { | |
20a0307c WF |
688 | if (!PageCompound(page)) |
689 | return 0; | |
690 | ||
691 | page = compound_head(page); | |
758f66a2 | 692 | return get_compound_page_dtor(page) == free_huge_page; |
20a0307c | 693 | } |
43131e14 NH |
694 | EXPORT_SYMBOL_GPL(PageHuge); |
695 | ||
27c73ae7 AA |
696 | /* |
697 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
698 | * normal or transparent huge pages. | |
699 | */ | |
700 | int PageHeadHuge(struct page *page_head) | |
701 | { | |
27c73ae7 AA |
702 | if (!PageHead(page_head)) |
703 | return 0; | |
704 | ||
758f66a2 | 705 | return get_compound_page_dtor(page_head) == free_huge_page; |
27c73ae7 | 706 | } |
27c73ae7 | 707 | |
13d60f4b ZY |
708 | pgoff_t __basepage_index(struct page *page) |
709 | { | |
710 | struct page *page_head = compound_head(page); | |
711 | pgoff_t index = page_index(page_head); | |
712 | unsigned long compound_idx; | |
713 | ||
714 | if (!PageHuge(page_head)) | |
715 | return page_index(page); | |
716 | ||
717 | if (compound_order(page_head) >= MAX_ORDER) | |
718 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
719 | else | |
720 | compound_idx = page - page_head; | |
721 | ||
722 | return (index << compound_order(page_head)) + compound_idx; | |
723 | } | |
724 | ||
a5516438 | 725 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
1da177e4 | 726 | { |
1da177e4 | 727 | struct page *page; |
f96efd58 | 728 | |
aa888a74 AK |
729 | if (h->order >= MAX_ORDER) |
730 | return NULL; | |
731 | ||
6484eb3e | 732 | page = alloc_pages_exact_node(nid, |
86cdb465 | 733 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
551883ae | 734 | __GFP_REPEAT|__GFP_NOWARN, |
a5516438 | 735 | huge_page_order(h)); |
1da177e4 | 736 | if (page) { |
7f2e9525 | 737 | if (arch_prepare_hugepage(page)) { |
caff3a2c | 738 | __free_pages(page, huge_page_order(h)); |
7b8ee84d | 739 | return NULL; |
7f2e9525 | 740 | } |
a5516438 | 741 | prep_new_huge_page(h, page, nid); |
1da177e4 | 742 | } |
63b4613c NA |
743 | |
744 | return page; | |
745 | } | |
746 | ||
9a76db09 | 747 | /* |
6ae11b27 LS |
748 | * common helper functions for hstate_next_node_to_{alloc|free}. |
749 | * We may have allocated or freed a huge page based on a different | |
750 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
751 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
752 | * node for alloc or free. | |
9a76db09 | 753 | */ |
6ae11b27 | 754 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) |
9a76db09 | 755 | { |
6ae11b27 | 756 | nid = next_node(nid, *nodes_allowed); |
9a76db09 | 757 | if (nid == MAX_NUMNODES) |
6ae11b27 | 758 | nid = first_node(*nodes_allowed); |
9a76db09 LS |
759 | VM_BUG_ON(nid >= MAX_NUMNODES); |
760 | ||
761 | return nid; | |
762 | } | |
763 | ||
6ae11b27 LS |
764 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) |
765 | { | |
766 | if (!node_isset(nid, *nodes_allowed)) | |
767 | nid = next_node_allowed(nid, nodes_allowed); | |
768 | return nid; | |
769 | } | |
770 | ||
5ced66c9 | 771 | /* |
6ae11b27 LS |
772 | * returns the previously saved node ["this node"] from which to |
773 | * allocate a persistent huge page for the pool and advance the | |
774 | * next node from which to allocate, handling wrap at end of node | |
775 | * mask. | |
5ced66c9 | 776 | */ |
6ae11b27 LS |
777 | static int hstate_next_node_to_alloc(struct hstate *h, |
778 | nodemask_t *nodes_allowed) | |
5ced66c9 | 779 | { |
6ae11b27 LS |
780 | int nid; |
781 | ||
782 | VM_BUG_ON(!nodes_allowed); | |
783 | ||
784 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
785 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
9a76db09 | 786 | |
9a76db09 | 787 | return nid; |
5ced66c9 AK |
788 | } |
789 | ||
e8c5c824 | 790 | /* |
6ae11b27 LS |
791 | * helper for free_pool_huge_page() - return the previously saved |
792 | * node ["this node"] from which to free a huge page. Advance the | |
793 | * next node id whether or not we find a free huge page to free so | |
794 | * that the next attempt to free addresses the next node. | |
e8c5c824 | 795 | */ |
6ae11b27 | 796 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) |
e8c5c824 | 797 | { |
6ae11b27 LS |
798 | int nid; |
799 | ||
800 | VM_BUG_ON(!nodes_allowed); | |
801 | ||
802 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
803 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
9a76db09 | 804 | |
9a76db09 | 805 | return nid; |
e8c5c824 LS |
806 | } |
807 | ||
b2261026 JK |
808 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ |
809 | for (nr_nodes = nodes_weight(*mask); \ | |
810 | nr_nodes > 0 && \ | |
811 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
812 | nr_nodes--) | |
813 | ||
814 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
815 | for (nr_nodes = nodes_weight(*mask); \ | |
816 | nr_nodes > 0 && \ | |
817 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
818 | nr_nodes--) | |
819 | ||
820 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) | |
821 | { | |
822 | struct page *page; | |
823 | int nr_nodes, node; | |
824 | int ret = 0; | |
825 | ||
826 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
827 | page = alloc_fresh_huge_page_node(h, node); | |
828 | if (page) { | |
829 | ret = 1; | |
830 | break; | |
831 | } | |
832 | } | |
833 | ||
834 | if (ret) | |
835 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
836 | else | |
837 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
838 | ||
839 | return ret; | |
840 | } | |
841 | ||
e8c5c824 LS |
842 | /* |
843 | * Free huge page from pool from next node to free. | |
844 | * Attempt to keep persistent huge pages more or less | |
845 | * balanced over allowed nodes. | |
846 | * Called with hugetlb_lock locked. | |
847 | */ | |
6ae11b27 LS |
848 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
849 | bool acct_surplus) | |
e8c5c824 | 850 | { |
b2261026 | 851 | int nr_nodes, node; |
e8c5c824 LS |
852 | int ret = 0; |
853 | ||
b2261026 | 854 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
855 | /* |
856 | * If we're returning unused surplus pages, only examine | |
857 | * nodes with surplus pages. | |
858 | */ | |
b2261026 JK |
859 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
860 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 861 | struct page *page = |
b2261026 | 862 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
863 | struct page, lru); |
864 | list_del(&page->lru); | |
865 | h->free_huge_pages--; | |
b2261026 | 866 | h->free_huge_pages_node[node]--; |
685f3457 LS |
867 | if (acct_surplus) { |
868 | h->surplus_huge_pages--; | |
b2261026 | 869 | h->surplus_huge_pages_node[node]--; |
685f3457 | 870 | } |
e8c5c824 LS |
871 | update_and_free_page(h, page); |
872 | ret = 1; | |
9a76db09 | 873 | break; |
e8c5c824 | 874 | } |
b2261026 | 875 | } |
e8c5c824 LS |
876 | |
877 | return ret; | |
878 | } | |
879 | ||
c8721bbb NH |
880 | /* |
881 | * Dissolve a given free hugepage into free buddy pages. This function does | |
882 | * nothing for in-use (including surplus) hugepages. | |
883 | */ | |
884 | static void dissolve_free_huge_page(struct page *page) | |
885 | { | |
886 | spin_lock(&hugetlb_lock); | |
887 | if (PageHuge(page) && !page_count(page)) { | |
888 | struct hstate *h = page_hstate(page); | |
889 | int nid = page_to_nid(page); | |
890 | list_del(&page->lru); | |
891 | h->free_huge_pages--; | |
892 | h->free_huge_pages_node[nid]--; | |
893 | update_and_free_page(h, page); | |
894 | } | |
895 | spin_unlock(&hugetlb_lock); | |
896 | } | |
897 | ||
898 | /* | |
899 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
900 | * make specified memory blocks removable from the system. | |
901 | * Note that start_pfn should aligned with (minimum) hugepage size. | |
902 | */ | |
903 | void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) | |
904 | { | |
905 | unsigned int order = 8 * sizeof(void *); | |
906 | unsigned long pfn; | |
907 | struct hstate *h; | |
908 | ||
909 | /* Set scan step to minimum hugepage size */ | |
910 | for_each_hstate(h) | |
911 | if (order > huge_page_order(h)) | |
912 | order = huge_page_order(h); | |
913 | VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order)); | |
914 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) | |
915 | dissolve_free_huge_page(pfn_to_page(pfn)); | |
916 | } | |
917 | ||
bf50bab2 | 918 | static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) |
7893d1d5 AL |
919 | { |
920 | struct page *page; | |
bf50bab2 | 921 | unsigned int r_nid; |
7893d1d5 | 922 | |
aa888a74 AK |
923 | if (h->order >= MAX_ORDER) |
924 | return NULL; | |
925 | ||
d1c3fb1f NA |
926 | /* |
927 | * Assume we will successfully allocate the surplus page to | |
928 | * prevent racing processes from causing the surplus to exceed | |
929 | * overcommit | |
930 | * | |
931 | * This however introduces a different race, where a process B | |
932 | * tries to grow the static hugepage pool while alloc_pages() is | |
933 | * called by process A. B will only examine the per-node | |
934 | * counters in determining if surplus huge pages can be | |
935 | * converted to normal huge pages in adjust_pool_surplus(). A | |
936 | * won't be able to increment the per-node counter, until the | |
937 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
938 | * no more huge pages can be converted from surplus to normal | |
939 | * state (and doesn't try to convert again). Thus, we have a | |
940 | * case where a surplus huge page exists, the pool is grown, and | |
941 | * the surplus huge page still exists after, even though it | |
942 | * should just have been converted to a normal huge page. This | |
943 | * does not leak memory, though, as the hugepage will be freed | |
944 | * once it is out of use. It also does not allow the counters to | |
945 | * go out of whack in adjust_pool_surplus() as we don't modify | |
946 | * the node values until we've gotten the hugepage and only the | |
947 | * per-node value is checked there. | |
948 | */ | |
949 | spin_lock(&hugetlb_lock); | |
a5516438 | 950 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
951 | spin_unlock(&hugetlb_lock); |
952 | return NULL; | |
953 | } else { | |
a5516438 AK |
954 | h->nr_huge_pages++; |
955 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
956 | } |
957 | spin_unlock(&hugetlb_lock); | |
958 | ||
bf50bab2 | 959 | if (nid == NUMA_NO_NODE) |
86cdb465 | 960 | page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP| |
bf50bab2 NH |
961 | __GFP_REPEAT|__GFP_NOWARN, |
962 | huge_page_order(h)); | |
963 | else | |
964 | page = alloc_pages_exact_node(nid, | |
86cdb465 | 965 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
bf50bab2 | 966 | __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); |
d1c3fb1f | 967 | |
caff3a2c GS |
968 | if (page && arch_prepare_hugepage(page)) { |
969 | __free_pages(page, huge_page_order(h)); | |
ea5768c7 | 970 | page = NULL; |
caff3a2c GS |
971 | } |
972 | ||
d1c3fb1f | 973 | spin_lock(&hugetlb_lock); |
7893d1d5 | 974 | if (page) { |
0edaecfa | 975 | INIT_LIST_HEAD(&page->lru); |
bf50bab2 | 976 | r_nid = page_to_nid(page); |
7893d1d5 | 977 | set_compound_page_dtor(page, free_huge_page); |
9dd540e2 | 978 | set_hugetlb_cgroup(page, NULL); |
d1c3fb1f NA |
979 | /* |
980 | * We incremented the global counters already | |
981 | */ | |
bf50bab2 NH |
982 | h->nr_huge_pages_node[r_nid]++; |
983 | h->surplus_huge_pages_node[r_nid]++; | |
3b116300 | 984 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 985 | } else { |
a5516438 AK |
986 | h->nr_huge_pages--; |
987 | h->surplus_huge_pages--; | |
3b116300 | 988 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 989 | } |
d1c3fb1f | 990 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
991 | |
992 | return page; | |
993 | } | |
994 | ||
bf50bab2 NH |
995 | /* |
996 | * This allocation function is useful in the context where vma is irrelevant. | |
997 | * E.g. soft-offlining uses this function because it only cares physical | |
998 | * address of error page. | |
999 | */ | |
1000 | struct page *alloc_huge_page_node(struct hstate *h, int nid) | |
1001 | { | |
4ef91848 | 1002 | struct page *page = NULL; |
bf50bab2 NH |
1003 | |
1004 | spin_lock(&hugetlb_lock); | |
4ef91848 JK |
1005 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
1006 | page = dequeue_huge_page_node(h, nid); | |
bf50bab2 NH |
1007 | spin_unlock(&hugetlb_lock); |
1008 | ||
94ae8ba7 | 1009 | if (!page) |
bf50bab2 NH |
1010 | page = alloc_buddy_huge_page(h, nid); |
1011 | ||
1012 | return page; | |
1013 | } | |
1014 | ||
e4e574b7 | 1015 | /* |
25985edc | 1016 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1017 | * of size 'delta'. |
1018 | */ | |
a5516438 | 1019 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1020 | { |
1021 | struct list_head surplus_list; | |
1022 | struct page *page, *tmp; | |
1023 | int ret, i; | |
1024 | int needed, allocated; | |
28073b02 | 1025 | bool alloc_ok = true; |
e4e574b7 | 1026 | |
a5516438 | 1027 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1028 | if (needed <= 0) { |
a5516438 | 1029 | h->resv_huge_pages += delta; |
e4e574b7 | 1030 | return 0; |
ac09b3a1 | 1031 | } |
e4e574b7 AL |
1032 | |
1033 | allocated = 0; | |
1034 | INIT_LIST_HEAD(&surplus_list); | |
1035 | ||
1036 | ret = -ENOMEM; | |
1037 | retry: | |
1038 | spin_unlock(&hugetlb_lock); | |
1039 | for (i = 0; i < needed; i++) { | |
bf50bab2 | 1040 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
28073b02 HD |
1041 | if (!page) { |
1042 | alloc_ok = false; | |
1043 | break; | |
1044 | } | |
e4e574b7 AL |
1045 | list_add(&page->lru, &surplus_list); |
1046 | } | |
28073b02 | 1047 | allocated += i; |
e4e574b7 AL |
1048 | |
1049 | /* | |
1050 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1051 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1052 | */ | |
1053 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1054 | needed = (h->resv_huge_pages + delta) - |
1055 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1056 | if (needed > 0) { |
1057 | if (alloc_ok) | |
1058 | goto retry; | |
1059 | /* | |
1060 | * We were not able to allocate enough pages to | |
1061 | * satisfy the entire reservation so we free what | |
1062 | * we've allocated so far. | |
1063 | */ | |
1064 | goto free; | |
1065 | } | |
e4e574b7 AL |
1066 | /* |
1067 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1068 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1069 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1070 | * allocator. Commit the entire reservation here to prevent another |
1071 | * process from stealing the pages as they are added to the pool but | |
1072 | * before they are reserved. | |
e4e574b7 AL |
1073 | */ |
1074 | needed += allocated; | |
a5516438 | 1075 | h->resv_huge_pages += delta; |
e4e574b7 | 1076 | ret = 0; |
a9869b83 | 1077 | |
19fc3f0a | 1078 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1079 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1080 | if ((--needed) < 0) |
1081 | break; | |
a9869b83 NH |
1082 | /* |
1083 | * This page is now managed by the hugetlb allocator and has | |
1084 | * no users -- drop the buddy allocator's reference. | |
1085 | */ | |
1086 | put_page_testzero(page); | |
309381fe | 1087 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 1088 | enqueue_huge_page(h, page); |
19fc3f0a | 1089 | } |
28073b02 | 1090 | free: |
b0365c8d | 1091 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1092 | |
1093 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1094 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1095 | put_page(page); | |
a9869b83 | 1096 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1097 | |
1098 | return ret; | |
1099 | } | |
1100 | ||
1101 | /* | |
1102 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
1103 | * allocated to satisfy the reservation must be explicitly freed if they were | |
1104 | * never used. | |
685f3457 | 1105 | * Called with hugetlb_lock held. |
e4e574b7 | 1106 | */ |
a5516438 AK |
1107 | static void return_unused_surplus_pages(struct hstate *h, |
1108 | unsigned long unused_resv_pages) | |
e4e574b7 | 1109 | { |
e4e574b7 AL |
1110 | unsigned long nr_pages; |
1111 | ||
ac09b3a1 | 1112 | /* Uncommit the reservation */ |
a5516438 | 1113 | h->resv_huge_pages -= unused_resv_pages; |
ac09b3a1 | 1114 | |
aa888a74 AK |
1115 | /* Cannot return gigantic pages currently */ |
1116 | if (h->order >= MAX_ORDER) | |
1117 | return; | |
1118 | ||
a5516438 | 1119 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1120 | |
685f3457 LS |
1121 | /* |
1122 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1123 | * evenly across all nodes with memory. Iterate across these nodes |
1124 | * until we can no longer free unreserved surplus pages. This occurs | |
1125 | * when the nodes with surplus pages have no free pages. | |
1126 | * free_pool_huge_page() will balance the the freed pages across the | |
1127 | * on-line nodes with memory and will handle the hstate accounting. | |
685f3457 LS |
1128 | */ |
1129 | while (nr_pages--) { | |
8cebfcd0 | 1130 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
685f3457 | 1131 | break; |
e4e574b7 AL |
1132 | } |
1133 | } | |
1134 | ||
c37f9fb1 AW |
1135 | /* |
1136 | * Determine if the huge page at addr within the vma has an associated | |
1137 | * reservation. Where it does not we will need to logically increase | |
90481622 DG |
1138 | * reservation and actually increase subpool usage before an allocation |
1139 | * can occur. Where any new reservation would be required the | |
1140 | * reservation change is prepared, but not committed. Once the page | |
1141 | * has been allocated from the subpool and instantiated the change should | |
1142 | * be committed via vma_commit_reservation. No action is required on | |
1143 | * failure. | |
c37f9fb1 | 1144 | */ |
e2f17d94 | 1145 | static long vma_needs_reservation(struct hstate *h, |
a5516438 | 1146 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 AW |
1147 | { |
1148 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1149 | struct inode *inode = mapping->host; | |
1150 | ||
f83a275d | 1151 | if (vma->vm_flags & VM_MAYSHARE) { |
a5516438 | 1152 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
9119a41e JK |
1153 | struct resv_map *resv = inode->i_mapping->private_data; |
1154 | ||
1155 | return region_chg(&resv->regions, idx, idx + 1); | |
c37f9fb1 | 1156 | |
84afd99b AW |
1157 | } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1158 | return 1; | |
c37f9fb1 | 1159 | |
84afd99b | 1160 | } else { |
e2f17d94 | 1161 | long err; |
a5516438 | 1162 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
f522c3ac | 1163 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b | 1164 | |
f522c3ac | 1165 | err = region_chg(&resv->regions, idx, idx + 1); |
84afd99b AW |
1166 | if (err < 0) |
1167 | return err; | |
1168 | return 0; | |
1169 | } | |
c37f9fb1 | 1170 | } |
a5516438 AK |
1171 | static void vma_commit_reservation(struct hstate *h, |
1172 | struct vm_area_struct *vma, unsigned long addr) | |
c37f9fb1 AW |
1173 | { |
1174 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1175 | struct inode *inode = mapping->host; | |
1176 | ||
f83a275d | 1177 | if (vma->vm_flags & VM_MAYSHARE) { |
a5516438 | 1178 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
9119a41e JK |
1179 | struct resv_map *resv = inode->i_mapping->private_data; |
1180 | ||
1181 | region_add(&resv->regions, idx, idx + 1); | |
84afd99b AW |
1182 | |
1183 | } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { | |
a5516438 | 1184 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
f522c3ac | 1185 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
1186 | |
1187 | /* Mark this page used in the map. */ | |
f522c3ac | 1188 | region_add(&resv->regions, idx, idx + 1); |
c37f9fb1 AW |
1189 | } |
1190 | } | |
1191 | ||
a1e78772 | 1192 | static struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 1193 | unsigned long addr, int avoid_reserve) |
1da177e4 | 1194 | { |
90481622 | 1195 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 1196 | struct hstate *h = hstate_vma(vma); |
348ea204 | 1197 | struct page *page; |
e2f17d94 | 1198 | long chg; |
6d76dcf4 AK |
1199 | int ret, idx; |
1200 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 1201 | |
6d76dcf4 | 1202 | idx = hstate_index(h); |
a1e78772 | 1203 | /* |
90481622 DG |
1204 | * Processes that did not create the mapping will have no |
1205 | * reserves and will not have accounted against subpool | |
1206 | * limit. Check that the subpool limit can be made before | |
1207 | * satisfying the allocation MAP_NORESERVE mappings may also | |
1208 | * need pages and subpool limit allocated allocated if no reserve | |
1209 | * mapping overlaps. | |
a1e78772 | 1210 | */ |
a5516438 | 1211 | chg = vma_needs_reservation(h, vma, addr); |
c37f9fb1 | 1212 | if (chg < 0) |
76dcee75 | 1213 | return ERR_PTR(-ENOMEM); |
8bb3f12e JK |
1214 | if (chg || avoid_reserve) |
1215 | if (hugepage_subpool_get_pages(spool, 1)) | |
76dcee75 | 1216 | return ERR_PTR(-ENOSPC); |
1da177e4 | 1217 | |
6d76dcf4 AK |
1218 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
1219 | if (ret) { | |
8bb3f12e JK |
1220 | if (chg || avoid_reserve) |
1221 | hugepage_subpool_put_pages(spool, 1); | |
6d76dcf4 AK |
1222 | return ERR_PTR(-ENOSPC); |
1223 | } | |
1da177e4 | 1224 | spin_lock(&hugetlb_lock); |
af0ed73e | 1225 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg); |
81a6fcae | 1226 | if (!page) { |
94ae8ba7 | 1227 | spin_unlock(&hugetlb_lock); |
bf50bab2 | 1228 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
68842c9b | 1229 | if (!page) { |
6d76dcf4 AK |
1230 | hugetlb_cgroup_uncharge_cgroup(idx, |
1231 | pages_per_huge_page(h), | |
1232 | h_cg); | |
8bb3f12e JK |
1233 | if (chg || avoid_reserve) |
1234 | hugepage_subpool_put_pages(spool, 1); | |
76dcee75 | 1235 | return ERR_PTR(-ENOSPC); |
68842c9b | 1236 | } |
79dbb236 AK |
1237 | spin_lock(&hugetlb_lock); |
1238 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 1239 | /* Fall through */ |
68842c9b | 1240 | } |
81a6fcae JK |
1241 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
1242 | spin_unlock(&hugetlb_lock); | |
348ea204 | 1243 | |
90481622 | 1244 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 1245 | |
a5516438 | 1246 | vma_commit_reservation(h, vma, addr); |
90d8b7e6 | 1247 | return page; |
b45b5bd6 DG |
1248 | } |
1249 | ||
74060e4d NH |
1250 | /* |
1251 | * alloc_huge_page()'s wrapper which simply returns the page if allocation | |
1252 | * succeeds, otherwise NULL. This function is called from new_vma_page(), | |
1253 | * where no ERR_VALUE is expected to be returned. | |
1254 | */ | |
1255 | struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, | |
1256 | unsigned long addr, int avoid_reserve) | |
1257 | { | |
1258 | struct page *page = alloc_huge_page(vma, addr, avoid_reserve); | |
1259 | if (IS_ERR(page)) | |
1260 | page = NULL; | |
1261 | return page; | |
1262 | } | |
1263 | ||
91f47662 | 1264 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
aa888a74 AK |
1265 | { |
1266 | struct huge_bootmem_page *m; | |
b2261026 | 1267 | int nr_nodes, node; |
aa888a74 | 1268 | |
b2261026 | 1269 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
1270 | void *addr; |
1271 | ||
8b89a116 GS |
1272 | addr = memblock_virt_alloc_try_nid_nopanic( |
1273 | huge_page_size(h), huge_page_size(h), | |
1274 | 0, BOOTMEM_ALLOC_ACCESSIBLE, node); | |
aa888a74 AK |
1275 | if (addr) { |
1276 | /* | |
1277 | * Use the beginning of the huge page to store the | |
1278 | * huge_bootmem_page struct (until gather_bootmem | |
1279 | * puts them into the mem_map). | |
1280 | */ | |
1281 | m = addr; | |
91f47662 | 1282 | goto found; |
aa888a74 | 1283 | } |
aa888a74 AK |
1284 | } |
1285 | return 0; | |
1286 | ||
1287 | found: | |
1288 | BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); | |
1289 | /* Put them into a private list first because mem_map is not up yet */ | |
1290 | list_add(&m->list, &huge_boot_pages); | |
1291 | m->hstate = h; | |
1292 | return 1; | |
1293 | } | |
1294 | ||
18229df5 AW |
1295 | static void prep_compound_huge_page(struct page *page, int order) |
1296 | { | |
1297 | if (unlikely(order > (MAX_ORDER - 1))) | |
1298 | prep_compound_gigantic_page(page, order); | |
1299 | else | |
1300 | prep_compound_page(page, order); | |
1301 | } | |
1302 | ||
aa888a74 AK |
1303 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
1304 | static void __init gather_bootmem_prealloc(void) | |
1305 | { | |
1306 | struct huge_bootmem_page *m; | |
1307 | ||
1308 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 1309 | struct hstate *h = m->hstate; |
ee8f248d BB |
1310 | struct page *page; |
1311 | ||
1312 | #ifdef CONFIG_HIGHMEM | |
1313 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
8b89a116 GS |
1314 | memblock_free_late(__pa(m), |
1315 | sizeof(struct huge_bootmem_page)); | |
ee8f248d BB |
1316 | #else |
1317 | page = virt_to_page(m); | |
1318 | #endif | |
aa888a74 | 1319 | WARN_ON(page_count(page) != 1); |
18229df5 | 1320 | prep_compound_huge_page(page, h->order); |
ef5a22be | 1321 | WARN_ON(PageReserved(page)); |
aa888a74 | 1322 | prep_new_huge_page(h, page, page_to_nid(page)); |
b0320c7b RA |
1323 | /* |
1324 | * If we had gigantic hugepages allocated at boot time, we need | |
1325 | * to restore the 'stolen' pages to totalram_pages in order to | |
1326 | * fix confusing memory reports from free(1) and another | |
1327 | * side-effects, like CommitLimit going negative. | |
1328 | */ | |
1329 | if (h->order > (MAX_ORDER - 1)) | |
3dcc0571 | 1330 | adjust_managed_page_count(page, 1 << h->order); |
aa888a74 AK |
1331 | } |
1332 | } | |
1333 | ||
8faa8b07 | 1334 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
1335 | { |
1336 | unsigned long i; | |
a5516438 | 1337 | |
e5ff2159 | 1338 | for (i = 0; i < h->max_huge_pages; ++i) { |
aa888a74 AK |
1339 | if (h->order >= MAX_ORDER) { |
1340 | if (!alloc_bootmem_huge_page(h)) | |
1341 | break; | |
9b5e5d0f | 1342 | } else if (!alloc_fresh_huge_page(h, |
8cebfcd0 | 1343 | &node_states[N_MEMORY])) |
1da177e4 | 1344 | break; |
1da177e4 | 1345 | } |
8faa8b07 | 1346 | h->max_huge_pages = i; |
e5ff2159 AK |
1347 | } |
1348 | ||
1349 | static void __init hugetlb_init_hstates(void) | |
1350 | { | |
1351 | struct hstate *h; | |
1352 | ||
1353 | for_each_hstate(h) { | |
8faa8b07 AK |
1354 | /* oversize hugepages were init'ed in early boot */ |
1355 | if (h->order < MAX_ORDER) | |
1356 | hugetlb_hstate_alloc_pages(h); | |
e5ff2159 AK |
1357 | } |
1358 | } | |
1359 | ||
4abd32db AK |
1360 | static char * __init memfmt(char *buf, unsigned long n) |
1361 | { | |
1362 | if (n >= (1UL << 30)) | |
1363 | sprintf(buf, "%lu GB", n >> 30); | |
1364 | else if (n >= (1UL << 20)) | |
1365 | sprintf(buf, "%lu MB", n >> 20); | |
1366 | else | |
1367 | sprintf(buf, "%lu KB", n >> 10); | |
1368 | return buf; | |
1369 | } | |
1370 | ||
e5ff2159 AK |
1371 | static void __init report_hugepages(void) |
1372 | { | |
1373 | struct hstate *h; | |
1374 | ||
1375 | for_each_hstate(h) { | |
4abd32db | 1376 | char buf[32]; |
ffb22af5 | 1377 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
4abd32db AK |
1378 | memfmt(buf, huge_page_size(h)), |
1379 | h->free_huge_pages); | |
e5ff2159 AK |
1380 | } |
1381 | } | |
1382 | ||
1da177e4 | 1383 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
1384 | static void try_to_free_low(struct hstate *h, unsigned long count, |
1385 | nodemask_t *nodes_allowed) | |
1da177e4 | 1386 | { |
4415cc8d CL |
1387 | int i; |
1388 | ||
aa888a74 AK |
1389 | if (h->order >= MAX_ORDER) |
1390 | return; | |
1391 | ||
6ae11b27 | 1392 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 1393 | struct page *page, *next; |
a5516438 AK |
1394 | struct list_head *freel = &h->hugepage_freelists[i]; |
1395 | list_for_each_entry_safe(page, next, freel, lru) { | |
1396 | if (count >= h->nr_huge_pages) | |
6b0c880d | 1397 | return; |
1da177e4 LT |
1398 | if (PageHighMem(page)) |
1399 | continue; | |
1400 | list_del(&page->lru); | |
e5ff2159 | 1401 | update_and_free_page(h, page); |
a5516438 AK |
1402 | h->free_huge_pages--; |
1403 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
1404 | } |
1405 | } | |
1406 | } | |
1407 | #else | |
6ae11b27 LS |
1408 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
1409 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
1410 | { |
1411 | } | |
1412 | #endif | |
1413 | ||
20a0307c WF |
1414 | /* |
1415 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
1416 | * balanced by operating on them in a round-robin fashion. | |
1417 | * Returns 1 if an adjustment was made. | |
1418 | */ | |
6ae11b27 LS |
1419 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
1420 | int delta) | |
20a0307c | 1421 | { |
b2261026 | 1422 | int nr_nodes, node; |
20a0307c WF |
1423 | |
1424 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 1425 | |
b2261026 JK |
1426 | if (delta < 0) { |
1427 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1428 | if (h->surplus_huge_pages_node[node]) | |
1429 | goto found; | |
e8c5c824 | 1430 | } |
b2261026 JK |
1431 | } else { |
1432 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
1433 | if (h->surplus_huge_pages_node[node] < | |
1434 | h->nr_huge_pages_node[node]) | |
1435 | goto found; | |
e8c5c824 | 1436 | } |
b2261026 JK |
1437 | } |
1438 | return 0; | |
20a0307c | 1439 | |
b2261026 JK |
1440 | found: |
1441 | h->surplus_huge_pages += delta; | |
1442 | h->surplus_huge_pages_node[node] += delta; | |
1443 | return 1; | |
20a0307c WF |
1444 | } |
1445 | ||
a5516438 | 1446 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
1447 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
1448 | nodemask_t *nodes_allowed) | |
1da177e4 | 1449 | { |
7893d1d5 | 1450 | unsigned long min_count, ret; |
1da177e4 | 1451 | |
aa888a74 AK |
1452 | if (h->order >= MAX_ORDER) |
1453 | return h->max_huge_pages; | |
1454 | ||
7893d1d5 AL |
1455 | /* |
1456 | * Increase the pool size | |
1457 | * First take pages out of surplus state. Then make up the | |
1458 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f NA |
1459 | * |
1460 | * We might race with alloc_buddy_huge_page() here and be unable | |
1461 | * to convert a surplus huge page to a normal huge page. That is | |
1462 | * not critical, though, it just means the overall size of the | |
1463 | * pool might be one hugepage larger than it needs to be, but | |
1464 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 1465 | */ |
1da177e4 | 1466 | spin_lock(&hugetlb_lock); |
a5516438 | 1467 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 1468 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
1469 | break; |
1470 | } | |
1471 | ||
a5516438 | 1472 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
1473 | /* |
1474 | * If this allocation races such that we no longer need the | |
1475 | * page, free_huge_page will handle it by freeing the page | |
1476 | * and reducing the surplus. | |
1477 | */ | |
1478 | spin_unlock(&hugetlb_lock); | |
6ae11b27 | 1479 | ret = alloc_fresh_huge_page(h, nodes_allowed); |
7893d1d5 AL |
1480 | spin_lock(&hugetlb_lock); |
1481 | if (!ret) | |
1482 | goto out; | |
1483 | ||
536240f2 MG |
1484 | /* Bail for signals. Probably ctrl-c from user */ |
1485 | if (signal_pending(current)) | |
1486 | goto out; | |
7893d1d5 | 1487 | } |
7893d1d5 AL |
1488 | |
1489 | /* | |
1490 | * Decrease the pool size | |
1491 | * First return free pages to the buddy allocator (being careful | |
1492 | * to keep enough around to satisfy reservations). Then place | |
1493 | * pages into surplus state as needed so the pool will shrink | |
1494 | * to the desired size as pages become free. | |
d1c3fb1f NA |
1495 | * |
1496 | * By placing pages into the surplus state independent of the | |
1497 | * overcommit value, we are allowing the surplus pool size to | |
1498 | * exceed overcommit. There are few sane options here. Since | |
1499 | * alloc_buddy_huge_page() is checking the global counter, | |
1500 | * though, we'll note that we're not allowed to exceed surplus | |
1501 | * and won't grow the pool anywhere else. Not until one of the | |
1502 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 1503 | */ |
a5516438 | 1504 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 1505 | min_count = max(count, min_count); |
6ae11b27 | 1506 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 1507 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 1508 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 1509 | break; |
1da177e4 | 1510 | } |
a5516438 | 1511 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 1512 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
1513 | break; |
1514 | } | |
1515 | out: | |
a5516438 | 1516 | ret = persistent_huge_pages(h); |
1da177e4 | 1517 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 1518 | return ret; |
1da177e4 LT |
1519 | } |
1520 | ||
a3437870 NA |
1521 | #define HSTATE_ATTR_RO(_name) \ |
1522 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
1523 | ||
1524 | #define HSTATE_ATTR(_name) \ | |
1525 | static struct kobj_attribute _name##_attr = \ | |
1526 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
1527 | ||
1528 | static struct kobject *hugepages_kobj; | |
1529 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1530 | ||
9a305230 LS |
1531 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
1532 | ||
1533 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
1534 | { |
1535 | int i; | |
9a305230 | 1536 | |
a3437870 | 1537 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
1538 | if (hstate_kobjs[i] == kobj) { |
1539 | if (nidp) | |
1540 | *nidp = NUMA_NO_NODE; | |
a3437870 | 1541 | return &hstates[i]; |
9a305230 LS |
1542 | } |
1543 | ||
1544 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
1545 | } |
1546 | ||
06808b08 | 1547 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
1548 | struct kobj_attribute *attr, char *buf) |
1549 | { | |
9a305230 LS |
1550 | struct hstate *h; |
1551 | unsigned long nr_huge_pages; | |
1552 | int nid; | |
1553 | ||
1554 | h = kobj_to_hstate(kobj, &nid); | |
1555 | if (nid == NUMA_NO_NODE) | |
1556 | nr_huge_pages = h->nr_huge_pages; | |
1557 | else | |
1558 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
1559 | ||
1560 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 1561 | } |
adbe8726 | 1562 | |
06808b08 LS |
1563 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
1564 | struct kobject *kobj, struct kobj_attribute *attr, | |
1565 | const char *buf, size_t len) | |
a3437870 NA |
1566 | { |
1567 | int err; | |
9a305230 | 1568 | int nid; |
06808b08 | 1569 | unsigned long count; |
9a305230 | 1570 | struct hstate *h; |
bad44b5b | 1571 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 1572 | |
3dbb95f7 | 1573 | err = kstrtoul(buf, 10, &count); |
73ae31e5 | 1574 | if (err) |
adbe8726 | 1575 | goto out; |
a3437870 | 1576 | |
9a305230 | 1577 | h = kobj_to_hstate(kobj, &nid); |
adbe8726 EM |
1578 | if (h->order >= MAX_ORDER) { |
1579 | err = -EINVAL; | |
1580 | goto out; | |
1581 | } | |
1582 | ||
9a305230 LS |
1583 | if (nid == NUMA_NO_NODE) { |
1584 | /* | |
1585 | * global hstate attribute | |
1586 | */ | |
1587 | if (!(obey_mempolicy && | |
1588 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
1589 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 1590 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 LS |
1591 | } |
1592 | } else if (nodes_allowed) { | |
1593 | /* | |
1594 | * per node hstate attribute: adjust count to global, | |
1595 | * but restrict alloc/free to the specified node. | |
1596 | */ | |
1597 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
1598 | init_nodemask_of_node(nodes_allowed, nid); | |
1599 | } else | |
8cebfcd0 | 1600 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 | 1601 | |
06808b08 | 1602 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 1603 | |
8cebfcd0 | 1604 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
1605 | NODEMASK_FREE(nodes_allowed); |
1606 | ||
1607 | return len; | |
adbe8726 EM |
1608 | out: |
1609 | NODEMASK_FREE(nodes_allowed); | |
1610 | return err; | |
06808b08 LS |
1611 | } |
1612 | ||
1613 | static ssize_t nr_hugepages_show(struct kobject *kobj, | |
1614 | struct kobj_attribute *attr, char *buf) | |
1615 | { | |
1616 | return nr_hugepages_show_common(kobj, attr, buf); | |
1617 | } | |
1618 | ||
1619 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
1620 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1621 | { | |
1622 | return nr_hugepages_store_common(false, kobj, attr, buf, len); | |
a3437870 NA |
1623 | } |
1624 | HSTATE_ATTR(nr_hugepages); | |
1625 | ||
06808b08 LS |
1626 | #ifdef CONFIG_NUMA |
1627 | ||
1628 | /* | |
1629 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
1630 | * huge page alloc/free. | |
1631 | */ | |
1632 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
1633 | struct kobj_attribute *attr, char *buf) | |
1634 | { | |
1635 | return nr_hugepages_show_common(kobj, attr, buf); | |
1636 | } | |
1637 | ||
1638 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
1639 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1640 | { | |
1641 | return nr_hugepages_store_common(true, kobj, attr, buf, len); | |
1642 | } | |
1643 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
1644 | #endif | |
1645 | ||
1646 | ||
a3437870 NA |
1647 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
1648 | struct kobj_attribute *attr, char *buf) | |
1649 | { | |
9a305230 | 1650 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1651 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
1652 | } | |
adbe8726 | 1653 | |
a3437870 NA |
1654 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
1655 | struct kobj_attribute *attr, const char *buf, size_t count) | |
1656 | { | |
1657 | int err; | |
1658 | unsigned long input; | |
9a305230 | 1659 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 1660 | |
adbe8726 EM |
1661 | if (h->order >= MAX_ORDER) |
1662 | return -EINVAL; | |
1663 | ||
3dbb95f7 | 1664 | err = kstrtoul(buf, 10, &input); |
a3437870 | 1665 | if (err) |
73ae31e5 | 1666 | return err; |
a3437870 NA |
1667 | |
1668 | spin_lock(&hugetlb_lock); | |
1669 | h->nr_overcommit_huge_pages = input; | |
1670 | spin_unlock(&hugetlb_lock); | |
1671 | ||
1672 | return count; | |
1673 | } | |
1674 | HSTATE_ATTR(nr_overcommit_hugepages); | |
1675 | ||
1676 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
1677 | struct kobj_attribute *attr, char *buf) | |
1678 | { | |
9a305230 LS |
1679 | struct hstate *h; |
1680 | unsigned long free_huge_pages; | |
1681 | int nid; | |
1682 | ||
1683 | h = kobj_to_hstate(kobj, &nid); | |
1684 | if (nid == NUMA_NO_NODE) | |
1685 | free_huge_pages = h->free_huge_pages; | |
1686 | else | |
1687 | free_huge_pages = h->free_huge_pages_node[nid]; | |
1688 | ||
1689 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
1690 | } |
1691 | HSTATE_ATTR_RO(free_hugepages); | |
1692 | ||
1693 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
1694 | struct kobj_attribute *attr, char *buf) | |
1695 | { | |
9a305230 | 1696 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1697 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
1698 | } | |
1699 | HSTATE_ATTR_RO(resv_hugepages); | |
1700 | ||
1701 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
1702 | struct kobj_attribute *attr, char *buf) | |
1703 | { | |
9a305230 LS |
1704 | struct hstate *h; |
1705 | unsigned long surplus_huge_pages; | |
1706 | int nid; | |
1707 | ||
1708 | h = kobj_to_hstate(kobj, &nid); | |
1709 | if (nid == NUMA_NO_NODE) | |
1710 | surplus_huge_pages = h->surplus_huge_pages; | |
1711 | else | |
1712 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
1713 | ||
1714 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
1715 | } |
1716 | HSTATE_ATTR_RO(surplus_hugepages); | |
1717 | ||
1718 | static struct attribute *hstate_attrs[] = { | |
1719 | &nr_hugepages_attr.attr, | |
1720 | &nr_overcommit_hugepages_attr.attr, | |
1721 | &free_hugepages_attr.attr, | |
1722 | &resv_hugepages_attr.attr, | |
1723 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
1724 | #ifdef CONFIG_NUMA |
1725 | &nr_hugepages_mempolicy_attr.attr, | |
1726 | #endif | |
a3437870 NA |
1727 | NULL, |
1728 | }; | |
1729 | ||
1730 | static struct attribute_group hstate_attr_group = { | |
1731 | .attrs = hstate_attrs, | |
1732 | }; | |
1733 | ||
094e9539 JM |
1734 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
1735 | struct kobject **hstate_kobjs, | |
1736 | struct attribute_group *hstate_attr_group) | |
a3437870 NA |
1737 | { |
1738 | int retval; | |
972dc4de | 1739 | int hi = hstate_index(h); |
a3437870 | 1740 | |
9a305230 LS |
1741 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
1742 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
1743 | return -ENOMEM; |
1744 | ||
9a305230 | 1745 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 1746 | if (retval) |
9a305230 | 1747 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
1748 | |
1749 | return retval; | |
1750 | } | |
1751 | ||
1752 | static void __init hugetlb_sysfs_init(void) | |
1753 | { | |
1754 | struct hstate *h; | |
1755 | int err; | |
1756 | ||
1757 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
1758 | if (!hugepages_kobj) | |
1759 | return; | |
1760 | ||
1761 | for_each_hstate(h) { | |
9a305230 LS |
1762 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
1763 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 1764 | if (err) |
ffb22af5 | 1765 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
1766 | } |
1767 | } | |
1768 | ||
9a305230 LS |
1769 | #ifdef CONFIG_NUMA |
1770 | ||
1771 | /* | |
1772 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
1773 | * with node devices in node_devices[] using a parallel array. The array |
1774 | * index of a node device or _hstate == node id. | |
1775 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
1776 | * the base kernel, on the hugetlb module. |
1777 | */ | |
1778 | struct node_hstate { | |
1779 | struct kobject *hugepages_kobj; | |
1780 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1781 | }; | |
1782 | struct node_hstate node_hstates[MAX_NUMNODES]; | |
1783 | ||
1784 | /* | |
10fbcf4c | 1785 | * A subset of global hstate attributes for node devices |
9a305230 LS |
1786 | */ |
1787 | static struct attribute *per_node_hstate_attrs[] = { | |
1788 | &nr_hugepages_attr.attr, | |
1789 | &free_hugepages_attr.attr, | |
1790 | &surplus_hugepages_attr.attr, | |
1791 | NULL, | |
1792 | }; | |
1793 | ||
1794 | static struct attribute_group per_node_hstate_attr_group = { | |
1795 | .attrs = per_node_hstate_attrs, | |
1796 | }; | |
1797 | ||
1798 | /* | |
10fbcf4c | 1799 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
1800 | * Returns node id via non-NULL nidp. |
1801 | */ | |
1802 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
1803 | { | |
1804 | int nid; | |
1805 | ||
1806 | for (nid = 0; nid < nr_node_ids; nid++) { | |
1807 | struct node_hstate *nhs = &node_hstates[nid]; | |
1808 | int i; | |
1809 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
1810 | if (nhs->hstate_kobjs[i] == kobj) { | |
1811 | if (nidp) | |
1812 | *nidp = nid; | |
1813 | return &hstates[i]; | |
1814 | } | |
1815 | } | |
1816 | ||
1817 | BUG(); | |
1818 | return NULL; | |
1819 | } | |
1820 | ||
1821 | /* | |
10fbcf4c | 1822 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
1823 | * No-op if no hstate attributes attached. |
1824 | */ | |
3cd8b44f | 1825 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
1826 | { |
1827 | struct hstate *h; | |
10fbcf4c | 1828 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
1829 | |
1830 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 1831 | return; /* no hstate attributes */ |
9a305230 | 1832 | |
972dc4de AK |
1833 | for_each_hstate(h) { |
1834 | int idx = hstate_index(h); | |
1835 | if (nhs->hstate_kobjs[idx]) { | |
1836 | kobject_put(nhs->hstate_kobjs[idx]); | |
1837 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 1838 | } |
972dc4de | 1839 | } |
9a305230 LS |
1840 | |
1841 | kobject_put(nhs->hugepages_kobj); | |
1842 | nhs->hugepages_kobj = NULL; | |
1843 | } | |
1844 | ||
1845 | /* | |
10fbcf4c | 1846 | * hugetlb module exit: unregister hstate attributes from node devices |
9a305230 LS |
1847 | * that have them. |
1848 | */ | |
1849 | static void hugetlb_unregister_all_nodes(void) | |
1850 | { | |
1851 | int nid; | |
1852 | ||
1853 | /* | |
10fbcf4c | 1854 | * disable node device registrations. |
9a305230 LS |
1855 | */ |
1856 | register_hugetlbfs_with_node(NULL, NULL); | |
1857 | ||
1858 | /* | |
1859 | * remove hstate attributes from any nodes that have them. | |
1860 | */ | |
1861 | for (nid = 0; nid < nr_node_ids; nid++) | |
8732794b | 1862 | hugetlb_unregister_node(node_devices[nid]); |
9a305230 LS |
1863 | } |
1864 | ||
1865 | /* | |
10fbcf4c | 1866 | * Register hstate attributes for a single node device. |
9a305230 LS |
1867 | * No-op if attributes already registered. |
1868 | */ | |
3cd8b44f | 1869 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
1870 | { |
1871 | struct hstate *h; | |
10fbcf4c | 1872 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
1873 | int err; |
1874 | ||
1875 | if (nhs->hugepages_kobj) | |
1876 | return; /* already allocated */ | |
1877 | ||
1878 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 1879 | &node->dev.kobj); |
9a305230 LS |
1880 | if (!nhs->hugepages_kobj) |
1881 | return; | |
1882 | ||
1883 | for_each_hstate(h) { | |
1884 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
1885 | nhs->hstate_kobjs, | |
1886 | &per_node_hstate_attr_group); | |
1887 | if (err) { | |
ffb22af5 AM |
1888 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
1889 | h->name, node->dev.id); | |
9a305230 LS |
1890 | hugetlb_unregister_node(node); |
1891 | break; | |
1892 | } | |
1893 | } | |
1894 | } | |
1895 | ||
1896 | /* | |
9b5e5d0f | 1897 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
1898 | * devices of nodes that have memory. All on-line nodes should have |
1899 | * registered their associated device by this time. | |
9a305230 LS |
1900 | */ |
1901 | static void hugetlb_register_all_nodes(void) | |
1902 | { | |
1903 | int nid; | |
1904 | ||
8cebfcd0 | 1905 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 1906 | struct node *node = node_devices[nid]; |
10fbcf4c | 1907 | if (node->dev.id == nid) |
9a305230 LS |
1908 | hugetlb_register_node(node); |
1909 | } | |
1910 | ||
1911 | /* | |
10fbcf4c | 1912 | * Let the node device driver know we're here so it can |
9a305230 LS |
1913 | * [un]register hstate attributes on node hotplug. |
1914 | */ | |
1915 | register_hugetlbfs_with_node(hugetlb_register_node, | |
1916 | hugetlb_unregister_node); | |
1917 | } | |
1918 | #else /* !CONFIG_NUMA */ | |
1919 | ||
1920 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
1921 | { | |
1922 | BUG(); | |
1923 | if (nidp) | |
1924 | *nidp = -1; | |
1925 | return NULL; | |
1926 | } | |
1927 | ||
1928 | static void hugetlb_unregister_all_nodes(void) { } | |
1929 | ||
1930 | static void hugetlb_register_all_nodes(void) { } | |
1931 | ||
1932 | #endif | |
1933 | ||
a3437870 NA |
1934 | static void __exit hugetlb_exit(void) |
1935 | { | |
1936 | struct hstate *h; | |
1937 | ||
9a305230 LS |
1938 | hugetlb_unregister_all_nodes(); |
1939 | ||
a3437870 | 1940 | for_each_hstate(h) { |
972dc4de | 1941 | kobject_put(hstate_kobjs[hstate_index(h)]); |
a3437870 NA |
1942 | } |
1943 | ||
1944 | kobject_put(hugepages_kobj); | |
1945 | } | |
1946 | module_exit(hugetlb_exit); | |
1947 | ||
1948 | static int __init hugetlb_init(void) | |
1949 | { | |
0ef89d25 BH |
1950 | /* Some platform decide whether they support huge pages at boot |
1951 | * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when | |
1952 | * there is no such support | |
1953 | */ | |
1954 | if (HPAGE_SHIFT == 0) | |
1955 | return 0; | |
a3437870 | 1956 | |
e11bfbfc NP |
1957 | if (!size_to_hstate(default_hstate_size)) { |
1958 | default_hstate_size = HPAGE_SIZE; | |
1959 | if (!size_to_hstate(default_hstate_size)) | |
1960 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 1961 | } |
972dc4de | 1962 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
e11bfbfc NP |
1963 | if (default_hstate_max_huge_pages) |
1964 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
a3437870 NA |
1965 | |
1966 | hugetlb_init_hstates(); | |
aa888a74 | 1967 | gather_bootmem_prealloc(); |
a3437870 NA |
1968 | report_hugepages(); |
1969 | ||
1970 | hugetlb_sysfs_init(); | |
9a305230 | 1971 | hugetlb_register_all_nodes(); |
7179e7bf | 1972 | hugetlb_cgroup_file_init(); |
9a305230 | 1973 | |
a3437870 NA |
1974 | return 0; |
1975 | } | |
1976 | module_init(hugetlb_init); | |
1977 | ||
1978 | /* Should be called on processing a hugepagesz=... option */ | |
1979 | void __init hugetlb_add_hstate(unsigned order) | |
1980 | { | |
1981 | struct hstate *h; | |
8faa8b07 AK |
1982 | unsigned long i; |
1983 | ||
a3437870 | 1984 | if (size_to_hstate(PAGE_SIZE << order)) { |
ffb22af5 | 1985 | pr_warning("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
1986 | return; |
1987 | } | |
47d38344 | 1988 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 1989 | BUG_ON(order == 0); |
47d38344 | 1990 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
1991 | h->order = order; |
1992 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
1993 | h->nr_huge_pages = 0; |
1994 | h->free_huge_pages = 0; | |
1995 | for (i = 0; i < MAX_NUMNODES; ++i) | |
1996 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 1997 | INIT_LIST_HEAD(&h->hugepage_activelist); |
8cebfcd0 LJ |
1998 | h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); |
1999 | h->next_nid_to_free = first_node(node_states[N_MEMORY]); | |
a3437870 NA |
2000 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
2001 | huge_page_size(h)/1024); | |
8faa8b07 | 2002 | |
a3437870 NA |
2003 | parsed_hstate = h; |
2004 | } | |
2005 | ||
e11bfbfc | 2006 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
2007 | { |
2008 | unsigned long *mhp; | |
8faa8b07 | 2009 | static unsigned long *last_mhp; |
a3437870 NA |
2010 | |
2011 | /* | |
47d38344 | 2012 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
2013 | * so this hugepages= parameter goes to the "default hstate". |
2014 | */ | |
47d38344 | 2015 | if (!hugetlb_max_hstate) |
a3437870 NA |
2016 | mhp = &default_hstate_max_huge_pages; |
2017 | else | |
2018 | mhp = &parsed_hstate->max_huge_pages; | |
2019 | ||
8faa8b07 | 2020 | if (mhp == last_mhp) { |
ffb22af5 AM |
2021 | pr_warning("hugepages= specified twice without " |
2022 | "interleaving hugepagesz=, ignoring\n"); | |
8faa8b07 AK |
2023 | return 1; |
2024 | } | |
2025 | ||
a3437870 NA |
2026 | if (sscanf(s, "%lu", mhp) <= 0) |
2027 | *mhp = 0; | |
2028 | ||
8faa8b07 AK |
2029 | /* |
2030 | * Global state is always initialized later in hugetlb_init. | |
2031 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
2032 | * use the bootmem allocator. | |
2033 | */ | |
47d38344 | 2034 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
2035 | hugetlb_hstate_alloc_pages(parsed_hstate); |
2036 | ||
2037 | last_mhp = mhp; | |
2038 | ||
a3437870 NA |
2039 | return 1; |
2040 | } | |
e11bfbfc NP |
2041 | __setup("hugepages=", hugetlb_nrpages_setup); |
2042 | ||
2043 | static int __init hugetlb_default_setup(char *s) | |
2044 | { | |
2045 | default_hstate_size = memparse(s, &s); | |
2046 | return 1; | |
2047 | } | |
2048 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 2049 | |
8a213460 NA |
2050 | static unsigned int cpuset_mems_nr(unsigned int *array) |
2051 | { | |
2052 | int node; | |
2053 | unsigned int nr = 0; | |
2054 | ||
2055 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
2056 | nr += array[node]; | |
2057 | ||
2058 | return nr; | |
2059 | } | |
2060 | ||
2061 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
2062 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
2063 | struct ctl_table *table, int write, | |
2064 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 2065 | { |
e5ff2159 AK |
2066 | struct hstate *h = &default_hstate; |
2067 | unsigned long tmp; | |
08d4a246 | 2068 | int ret; |
e5ff2159 | 2069 | |
c033a93c | 2070 | tmp = h->max_huge_pages; |
e5ff2159 | 2071 | |
adbe8726 EM |
2072 | if (write && h->order >= MAX_ORDER) |
2073 | return -EINVAL; | |
2074 | ||
e5ff2159 AK |
2075 | table->data = &tmp; |
2076 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2077 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2078 | if (ret) | |
2079 | goto out; | |
e5ff2159 | 2080 | |
06808b08 | 2081 | if (write) { |
bad44b5b DR |
2082 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, |
2083 | GFP_KERNEL | __GFP_NORETRY); | |
06808b08 LS |
2084 | if (!(obey_mempolicy && |
2085 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
2086 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 2087 | nodes_allowed = &node_states[N_MEMORY]; |
06808b08 LS |
2088 | } |
2089 | h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); | |
2090 | ||
8cebfcd0 | 2091 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
2092 | NODEMASK_FREE(nodes_allowed); |
2093 | } | |
08d4a246 MH |
2094 | out: |
2095 | return ret; | |
1da177e4 | 2096 | } |
396faf03 | 2097 | |
06808b08 LS |
2098 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2099 | void __user *buffer, size_t *length, loff_t *ppos) | |
2100 | { | |
2101 | ||
2102 | return hugetlb_sysctl_handler_common(false, table, write, | |
2103 | buffer, length, ppos); | |
2104 | } | |
2105 | ||
2106 | #ifdef CONFIG_NUMA | |
2107 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2108 | void __user *buffer, size_t *length, loff_t *ppos) | |
2109 | { | |
2110 | return hugetlb_sysctl_handler_common(true, table, write, | |
2111 | buffer, length, ppos); | |
2112 | } | |
2113 | #endif /* CONFIG_NUMA */ | |
2114 | ||
a3d0c6aa | 2115 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 2116 | void __user *buffer, |
a3d0c6aa NA |
2117 | size_t *length, loff_t *ppos) |
2118 | { | |
a5516438 | 2119 | struct hstate *h = &default_hstate; |
e5ff2159 | 2120 | unsigned long tmp; |
08d4a246 | 2121 | int ret; |
e5ff2159 | 2122 | |
c033a93c | 2123 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 2124 | |
adbe8726 EM |
2125 | if (write && h->order >= MAX_ORDER) |
2126 | return -EINVAL; | |
2127 | ||
e5ff2159 AK |
2128 | table->data = &tmp; |
2129 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2130 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2131 | if (ret) | |
2132 | goto out; | |
e5ff2159 AK |
2133 | |
2134 | if (write) { | |
2135 | spin_lock(&hugetlb_lock); | |
2136 | h->nr_overcommit_huge_pages = tmp; | |
2137 | spin_unlock(&hugetlb_lock); | |
2138 | } | |
08d4a246 MH |
2139 | out: |
2140 | return ret; | |
a3d0c6aa NA |
2141 | } |
2142 | ||
1da177e4 LT |
2143 | #endif /* CONFIG_SYSCTL */ |
2144 | ||
e1759c21 | 2145 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 2146 | { |
a5516438 | 2147 | struct hstate *h = &default_hstate; |
e1759c21 | 2148 | seq_printf(m, |
4f98a2fe RR |
2149 | "HugePages_Total: %5lu\n" |
2150 | "HugePages_Free: %5lu\n" | |
2151 | "HugePages_Rsvd: %5lu\n" | |
2152 | "HugePages_Surp: %5lu\n" | |
2153 | "Hugepagesize: %8lu kB\n", | |
a5516438 AK |
2154 | h->nr_huge_pages, |
2155 | h->free_huge_pages, | |
2156 | h->resv_huge_pages, | |
2157 | h->surplus_huge_pages, | |
2158 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
2159 | } |
2160 | ||
2161 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
2162 | { | |
a5516438 | 2163 | struct hstate *h = &default_hstate; |
1da177e4 LT |
2164 | return sprintf(buf, |
2165 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
2166 | "Node %d HugePages_Free: %5u\n" |
2167 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
2168 | nid, h->nr_huge_pages_node[nid], |
2169 | nid, h->free_huge_pages_node[nid], | |
2170 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
2171 | } |
2172 | ||
949f7ec5 DR |
2173 | void hugetlb_show_meminfo(void) |
2174 | { | |
2175 | struct hstate *h; | |
2176 | int nid; | |
2177 | ||
2178 | for_each_node_state(nid, N_MEMORY) | |
2179 | for_each_hstate(h) | |
2180 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
2181 | nid, | |
2182 | h->nr_huge_pages_node[nid], | |
2183 | h->free_huge_pages_node[nid], | |
2184 | h->surplus_huge_pages_node[nid], | |
2185 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
2186 | } | |
2187 | ||
1da177e4 LT |
2188 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
2189 | unsigned long hugetlb_total_pages(void) | |
2190 | { | |
d0028588 WL |
2191 | struct hstate *h; |
2192 | unsigned long nr_total_pages = 0; | |
2193 | ||
2194 | for_each_hstate(h) | |
2195 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
2196 | return nr_total_pages; | |
1da177e4 | 2197 | } |
1da177e4 | 2198 | |
a5516438 | 2199 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
2200 | { |
2201 | int ret = -ENOMEM; | |
2202 | ||
2203 | spin_lock(&hugetlb_lock); | |
2204 | /* | |
2205 | * When cpuset is configured, it breaks the strict hugetlb page | |
2206 | * reservation as the accounting is done on a global variable. Such | |
2207 | * reservation is completely rubbish in the presence of cpuset because | |
2208 | * the reservation is not checked against page availability for the | |
2209 | * current cpuset. Application can still potentially OOM'ed by kernel | |
2210 | * with lack of free htlb page in cpuset that the task is in. | |
2211 | * Attempt to enforce strict accounting with cpuset is almost | |
2212 | * impossible (or too ugly) because cpuset is too fluid that | |
2213 | * task or memory node can be dynamically moved between cpusets. | |
2214 | * | |
2215 | * The change of semantics for shared hugetlb mapping with cpuset is | |
2216 | * undesirable. However, in order to preserve some of the semantics, | |
2217 | * we fall back to check against current free page availability as | |
2218 | * a best attempt and hopefully to minimize the impact of changing | |
2219 | * semantics that cpuset has. | |
2220 | */ | |
2221 | if (delta > 0) { | |
a5516438 | 2222 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
2223 | goto out; |
2224 | ||
a5516438 AK |
2225 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
2226 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
2227 | goto out; |
2228 | } | |
2229 | } | |
2230 | ||
2231 | ret = 0; | |
2232 | if (delta < 0) | |
a5516438 | 2233 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
2234 | |
2235 | out: | |
2236 | spin_unlock(&hugetlb_lock); | |
2237 | return ret; | |
2238 | } | |
2239 | ||
84afd99b AW |
2240 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
2241 | { | |
f522c3ac | 2242 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
2243 | |
2244 | /* | |
2245 | * This new VMA should share its siblings reservation map if present. | |
2246 | * The VMA will only ever have a valid reservation map pointer where | |
2247 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 2248 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
2249 | * after this open call completes. It is therefore safe to take a |
2250 | * new reference here without additional locking. | |
2251 | */ | |
f522c3ac JK |
2252 | if (resv) |
2253 | kref_get(&resv->refs); | |
84afd99b AW |
2254 | } |
2255 | ||
c50ac050 DH |
2256 | static void resv_map_put(struct vm_area_struct *vma) |
2257 | { | |
f522c3ac | 2258 | struct resv_map *resv = vma_resv_map(vma); |
c50ac050 | 2259 | |
f522c3ac | 2260 | if (!resv) |
c50ac050 | 2261 | return; |
f522c3ac | 2262 | kref_put(&resv->refs, resv_map_release); |
c50ac050 DH |
2263 | } |
2264 | ||
a1e78772 MG |
2265 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
2266 | { | |
a5516438 | 2267 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 2268 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 2269 | struct hugepage_subpool *spool = subpool_vma(vma); |
84afd99b AW |
2270 | unsigned long reserve; |
2271 | unsigned long start; | |
2272 | unsigned long end; | |
2273 | ||
f522c3ac | 2274 | if (resv) { |
a5516438 AK |
2275 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
2276 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b AW |
2277 | |
2278 | reserve = (end - start) - | |
f522c3ac | 2279 | region_count(&resv->regions, start, end); |
84afd99b | 2280 | |
c50ac050 | 2281 | resv_map_put(vma); |
84afd99b | 2282 | |
7251ff78 | 2283 | if (reserve) { |
a5516438 | 2284 | hugetlb_acct_memory(h, -reserve); |
90481622 | 2285 | hugepage_subpool_put_pages(spool, reserve); |
7251ff78 | 2286 | } |
84afd99b | 2287 | } |
a1e78772 MG |
2288 | } |
2289 | ||
1da177e4 LT |
2290 | /* |
2291 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
2292 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
2293 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
2294 | * this far. | |
2295 | */ | |
d0217ac0 | 2296 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
2297 | { |
2298 | BUG(); | |
d0217ac0 | 2299 | return 0; |
1da177e4 LT |
2300 | } |
2301 | ||
f0f37e2f | 2302 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 2303 | .fault = hugetlb_vm_op_fault, |
84afd99b | 2304 | .open = hugetlb_vm_op_open, |
a1e78772 | 2305 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
2306 | }; |
2307 | ||
1e8f889b DG |
2308 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
2309 | int writable) | |
63551ae0 DG |
2310 | { |
2311 | pte_t entry; | |
2312 | ||
1e8f889b | 2313 | if (writable) { |
106c992a GS |
2314 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
2315 | vma->vm_page_prot))); | |
63551ae0 | 2316 | } else { |
106c992a GS |
2317 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
2318 | vma->vm_page_prot)); | |
63551ae0 DG |
2319 | } |
2320 | entry = pte_mkyoung(entry); | |
2321 | entry = pte_mkhuge(entry); | |
d9ed9faa | 2322 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
2323 | |
2324 | return entry; | |
2325 | } | |
2326 | ||
1e8f889b DG |
2327 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
2328 | unsigned long address, pte_t *ptep) | |
2329 | { | |
2330 | pte_t entry; | |
2331 | ||
106c992a | 2332 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 2333 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 2334 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
2335 | } |
2336 | ||
2337 | ||
63551ae0 DG |
2338 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
2339 | struct vm_area_struct *vma) | |
2340 | { | |
2341 | pte_t *src_pte, *dst_pte, entry; | |
2342 | struct page *ptepage; | |
1c59827d | 2343 | unsigned long addr; |
1e8f889b | 2344 | int cow; |
a5516438 AK |
2345 | struct hstate *h = hstate_vma(vma); |
2346 | unsigned long sz = huge_page_size(h); | |
e8569dd2 AS |
2347 | unsigned long mmun_start; /* For mmu_notifiers */ |
2348 | unsigned long mmun_end; /* For mmu_notifiers */ | |
2349 | int ret = 0; | |
1e8f889b DG |
2350 | |
2351 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 2352 | |
e8569dd2 AS |
2353 | mmun_start = vma->vm_start; |
2354 | mmun_end = vma->vm_end; | |
2355 | if (cow) | |
2356 | mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); | |
2357 | ||
a5516438 | 2358 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 2359 | spinlock_t *src_ptl, *dst_ptl; |
c74df32c HD |
2360 | src_pte = huge_pte_offset(src, addr); |
2361 | if (!src_pte) | |
2362 | continue; | |
a5516438 | 2363 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
2364 | if (!dst_pte) { |
2365 | ret = -ENOMEM; | |
2366 | break; | |
2367 | } | |
c5c99429 LW |
2368 | |
2369 | /* If the pagetables are shared don't copy or take references */ | |
2370 | if (dst_pte == src_pte) | |
2371 | continue; | |
2372 | ||
cb900f41 KS |
2373 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
2374 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
2375 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
7f2e9525 | 2376 | if (!huge_pte_none(huge_ptep_get(src_pte))) { |
1e8f889b | 2377 | if (cow) |
7f2e9525 GS |
2378 | huge_ptep_set_wrprotect(src, addr, src_pte); |
2379 | entry = huge_ptep_get(src_pte); | |
1c59827d HD |
2380 | ptepage = pte_page(entry); |
2381 | get_page(ptepage); | |
0fe6e20b | 2382 | page_dup_rmap(ptepage); |
1c59827d HD |
2383 | set_huge_pte_at(dst, addr, dst_pte, entry); |
2384 | } | |
cb900f41 KS |
2385 | spin_unlock(src_ptl); |
2386 | spin_unlock(dst_ptl); | |
63551ae0 | 2387 | } |
63551ae0 | 2388 | |
e8569dd2 AS |
2389 | if (cow) |
2390 | mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); | |
2391 | ||
2392 | return ret; | |
63551ae0 DG |
2393 | } |
2394 | ||
290408d4 NH |
2395 | static int is_hugetlb_entry_migration(pte_t pte) |
2396 | { | |
2397 | swp_entry_t swp; | |
2398 | ||
2399 | if (huge_pte_none(pte) || pte_present(pte)) | |
2400 | return 0; | |
2401 | swp = pte_to_swp_entry(pte); | |
32f84528 | 2402 | if (non_swap_entry(swp) && is_migration_entry(swp)) |
290408d4 | 2403 | return 1; |
32f84528 | 2404 | else |
290408d4 NH |
2405 | return 0; |
2406 | } | |
2407 | ||
fd6a03ed NH |
2408 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) |
2409 | { | |
2410 | swp_entry_t swp; | |
2411 | ||
2412 | if (huge_pte_none(pte) || pte_present(pte)) | |
2413 | return 0; | |
2414 | swp = pte_to_swp_entry(pte); | |
32f84528 | 2415 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) |
fd6a03ed | 2416 | return 1; |
32f84528 | 2417 | else |
fd6a03ed NH |
2418 | return 0; |
2419 | } | |
2420 | ||
24669e58 AK |
2421 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
2422 | unsigned long start, unsigned long end, | |
2423 | struct page *ref_page) | |
63551ae0 | 2424 | { |
24669e58 | 2425 | int force_flush = 0; |
63551ae0 DG |
2426 | struct mm_struct *mm = vma->vm_mm; |
2427 | unsigned long address; | |
c7546f8f | 2428 | pte_t *ptep; |
63551ae0 | 2429 | pte_t pte; |
cb900f41 | 2430 | spinlock_t *ptl; |
63551ae0 | 2431 | struct page *page; |
a5516438 AK |
2432 | struct hstate *h = hstate_vma(vma); |
2433 | unsigned long sz = huge_page_size(h); | |
2ec74c3e SG |
2434 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
2435 | const unsigned long mmun_end = end; /* For mmu_notifiers */ | |
a5516438 | 2436 | |
63551ae0 | 2437 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
2438 | BUG_ON(start & ~huge_page_mask(h)); |
2439 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 2440 | |
24669e58 | 2441 | tlb_start_vma(tlb, vma); |
2ec74c3e | 2442 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
24669e58 | 2443 | again: |
a5516438 | 2444 | for (address = start; address < end; address += sz) { |
c7546f8f | 2445 | ptep = huge_pte_offset(mm, address); |
4c887265 | 2446 | if (!ptep) |
c7546f8f DG |
2447 | continue; |
2448 | ||
cb900f41 | 2449 | ptl = huge_pte_lock(h, mm, ptep); |
39dde65c | 2450 | if (huge_pmd_unshare(mm, &address, ptep)) |
cb900f41 | 2451 | goto unlock; |
39dde65c | 2452 | |
6629326b HD |
2453 | pte = huge_ptep_get(ptep); |
2454 | if (huge_pte_none(pte)) | |
cb900f41 | 2455 | goto unlock; |
6629326b HD |
2456 | |
2457 | /* | |
2458 | * HWPoisoned hugepage is already unmapped and dropped reference | |
2459 | */ | |
8c4894c6 | 2460 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { |
106c992a | 2461 | huge_pte_clear(mm, address, ptep); |
cb900f41 | 2462 | goto unlock; |
8c4894c6 | 2463 | } |
6629326b HD |
2464 | |
2465 | page = pte_page(pte); | |
04f2cbe3 MG |
2466 | /* |
2467 | * If a reference page is supplied, it is because a specific | |
2468 | * page is being unmapped, not a range. Ensure the page we | |
2469 | * are about to unmap is the actual page of interest. | |
2470 | */ | |
2471 | if (ref_page) { | |
04f2cbe3 | 2472 | if (page != ref_page) |
cb900f41 | 2473 | goto unlock; |
04f2cbe3 MG |
2474 | |
2475 | /* | |
2476 | * Mark the VMA as having unmapped its page so that | |
2477 | * future faults in this VMA will fail rather than | |
2478 | * looking like data was lost | |
2479 | */ | |
2480 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
2481 | } | |
2482 | ||
c7546f8f | 2483 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
24669e58 | 2484 | tlb_remove_tlb_entry(tlb, ptep, address); |
106c992a | 2485 | if (huge_pte_dirty(pte)) |
6649a386 | 2486 | set_page_dirty(page); |
9e81130b | 2487 | |
24669e58 AK |
2488 | page_remove_rmap(page); |
2489 | force_flush = !__tlb_remove_page(tlb, page); | |
cb900f41 KS |
2490 | if (force_flush) { |
2491 | spin_unlock(ptl); | |
24669e58 | 2492 | break; |
cb900f41 | 2493 | } |
9e81130b | 2494 | /* Bail out after unmapping reference page if supplied */ |
cb900f41 KS |
2495 | if (ref_page) { |
2496 | spin_unlock(ptl); | |
9e81130b | 2497 | break; |
cb900f41 KS |
2498 | } |
2499 | unlock: | |
2500 | spin_unlock(ptl); | |
63551ae0 | 2501 | } |
24669e58 AK |
2502 | /* |
2503 | * mmu_gather ran out of room to batch pages, we break out of | |
2504 | * the PTE lock to avoid doing the potential expensive TLB invalidate | |
2505 | * and page-free while holding it. | |
2506 | */ | |
2507 | if (force_flush) { | |
2508 | force_flush = 0; | |
2509 | tlb_flush_mmu(tlb); | |
2510 | if (address < end && !ref_page) | |
2511 | goto again; | |
fe1668ae | 2512 | } |
2ec74c3e | 2513 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
24669e58 | 2514 | tlb_end_vma(tlb, vma); |
1da177e4 | 2515 | } |
63551ae0 | 2516 | |
d833352a MG |
2517 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
2518 | struct vm_area_struct *vma, unsigned long start, | |
2519 | unsigned long end, struct page *ref_page) | |
2520 | { | |
2521 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
2522 | ||
2523 | /* | |
2524 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
2525 | * test will fail on a vma being torn down, and not grab a page table | |
2526 | * on its way out. We're lucky that the flag has such an appropriate | |
2527 | * name, and can in fact be safely cleared here. We could clear it | |
2528 | * before the __unmap_hugepage_range above, but all that's necessary | |
2529 | * is to clear it before releasing the i_mmap_mutex. This works | |
2530 | * because in the context this is called, the VMA is about to be | |
2531 | * destroyed and the i_mmap_mutex is held. | |
2532 | */ | |
2533 | vma->vm_flags &= ~VM_MAYSHARE; | |
2534 | } | |
2535 | ||
502717f4 | 2536 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 2537 | unsigned long end, struct page *ref_page) |
502717f4 | 2538 | { |
24669e58 AK |
2539 | struct mm_struct *mm; |
2540 | struct mmu_gather tlb; | |
2541 | ||
2542 | mm = vma->vm_mm; | |
2543 | ||
2b047252 | 2544 | tlb_gather_mmu(&tlb, mm, start, end); |
24669e58 AK |
2545 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
2546 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 CK |
2547 | } |
2548 | ||
04f2cbe3 MG |
2549 | /* |
2550 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
2551 | * mappping it owns the reserve page for. The intention is to unmap the page | |
2552 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
2553 | * same region. | |
2554 | */ | |
2a4b3ded HH |
2555 | static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
2556 | struct page *page, unsigned long address) | |
04f2cbe3 | 2557 | { |
7526674d | 2558 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
2559 | struct vm_area_struct *iter_vma; |
2560 | struct address_space *mapping; | |
04f2cbe3 MG |
2561 | pgoff_t pgoff; |
2562 | ||
2563 | /* | |
2564 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
2565 | * from page cache lookup which is in HPAGE_SIZE units. | |
2566 | */ | |
7526674d | 2567 | address = address & huge_page_mask(h); |
36e4f20a MH |
2568 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
2569 | vma->vm_pgoff; | |
496ad9aa | 2570 | mapping = file_inode(vma->vm_file)->i_mapping; |
04f2cbe3 | 2571 | |
4eb2b1dc MG |
2572 | /* |
2573 | * Take the mapping lock for the duration of the table walk. As | |
2574 | * this mapping should be shared between all the VMAs, | |
2575 | * __unmap_hugepage_range() is called as the lock is already held | |
2576 | */ | |
3d48ae45 | 2577 | mutex_lock(&mapping->i_mmap_mutex); |
6b2dbba8 | 2578 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
2579 | /* Do not unmap the current VMA */ |
2580 | if (iter_vma == vma) | |
2581 | continue; | |
2582 | ||
2583 | /* | |
2584 | * Unmap the page from other VMAs without their own reserves. | |
2585 | * They get marked to be SIGKILLed if they fault in these | |
2586 | * areas. This is because a future no-page fault on this VMA | |
2587 | * could insert a zeroed page instead of the data existing | |
2588 | * from the time of fork. This would look like data corruption | |
2589 | */ | |
2590 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
2591 | unmap_hugepage_range(iter_vma, address, |
2592 | address + huge_page_size(h), page); | |
04f2cbe3 | 2593 | } |
3d48ae45 | 2594 | mutex_unlock(&mapping->i_mmap_mutex); |
04f2cbe3 MG |
2595 | |
2596 | return 1; | |
2597 | } | |
2598 | ||
0fe6e20b NH |
2599 | /* |
2600 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
2601 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
2602 | * cannot race with other handlers or page migration. | |
2603 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 2604 | */ |
1e8f889b | 2605 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
04f2cbe3 | 2606 | unsigned long address, pte_t *ptep, pte_t pte, |
cb900f41 | 2607 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 2608 | { |
a5516438 | 2609 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 2610 | struct page *old_page, *new_page; |
04f2cbe3 | 2611 | int outside_reserve = 0; |
2ec74c3e SG |
2612 | unsigned long mmun_start; /* For mmu_notifiers */ |
2613 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1e8f889b DG |
2614 | |
2615 | old_page = pte_page(pte); | |
2616 | ||
04f2cbe3 | 2617 | retry_avoidcopy: |
1e8f889b DG |
2618 | /* If no-one else is actually using this page, avoid the copy |
2619 | * and just make the page writable */ | |
37a2140d JK |
2620 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
2621 | page_move_anon_rmap(old_page, vma, address); | |
1e8f889b | 2622 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 2623 | return 0; |
1e8f889b DG |
2624 | } |
2625 | ||
04f2cbe3 MG |
2626 | /* |
2627 | * If the process that created a MAP_PRIVATE mapping is about to | |
2628 | * perform a COW due to a shared page count, attempt to satisfy | |
2629 | * the allocation without using the existing reserves. The pagecache | |
2630 | * page is used to determine if the reserve at this address was | |
2631 | * consumed or not. If reserves were used, a partial faulted mapping | |
2632 | * at the time of fork() could consume its reserves on COW instead | |
2633 | * of the full address range. | |
2634 | */ | |
5944d011 | 2635 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
2636 | old_page != pagecache_page) |
2637 | outside_reserve = 1; | |
2638 | ||
1e8f889b | 2639 | page_cache_get(old_page); |
b76c8cfb | 2640 | |
cb900f41 KS |
2641 | /* Drop page table lock as buddy allocator may be called */ |
2642 | spin_unlock(ptl); | |
04f2cbe3 | 2643 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 2644 | |
2fc39cec | 2645 | if (IS_ERR(new_page)) { |
76dcee75 | 2646 | long err = PTR_ERR(new_page); |
1e8f889b | 2647 | page_cache_release(old_page); |
04f2cbe3 MG |
2648 | |
2649 | /* | |
2650 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
2651 | * it is due to references held by a child and an insufficient | |
2652 | * huge page pool. To guarantee the original mappers | |
2653 | * reliability, unmap the page from child processes. The child | |
2654 | * may get SIGKILLed if it later faults. | |
2655 | */ | |
2656 | if (outside_reserve) { | |
2657 | BUG_ON(huge_pte_none(pte)); | |
2658 | if (unmap_ref_private(mm, vma, old_page, address)) { | |
04f2cbe3 | 2659 | BUG_ON(huge_pte_none(pte)); |
cb900f41 | 2660 | spin_lock(ptl); |
a734bcc8 HD |
2661 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
2662 | if (likely(pte_same(huge_ptep_get(ptep), pte))) | |
2663 | goto retry_avoidcopy; | |
2664 | /* | |
cb900f41 KS |
2665 | * race occurs while re-acquiring page table |
2666 | * lock, and our job is done. | |
a734bcc8 HD |
2667 | */ |
2668 | return 0; | |
04f2cbe3 MG |
2669 | } |
2670 | WARN_ON_ONCE(1); | |
2671 | } | |
2672 | ||
b76c8cfb | 2673 | /* Caller expects lock to be held */ |
cb900f41 | 2674 | spin_lock(ptl); |
76dcee75 AK |
2675 | if (err == -ENOMEM) |
2676 | return VM_FAULT_OOM; | |
2677 | else | |
2678 | return VM_FAULT_SIGBUS; | |
1e8f889b DG |
2679 | } |
2680 | ||
0fe6e20b NH |
2681 | /* |
2682 | * When the original hugepage is shared one, it does not have | |
2683 | * anon_vma prepared. | |
2684 | */ | |
44e2aa93 | 2685 | if (unlikely(anon_vma_prepare(vma))) { |
ea4039a3 HD |
2686 | page_cache_release(new_page); |
2687 | page_cache_release(old_page); | |
44e2aa93 | 2688 | /* Caller expects lock to be held */ |
cb900f41 | 2689 | spin_lock(ptl); |
0fe6e20b | 2690 | return VM_FAULT_OOM; |
44e2aa93 | 2691 | } |
0fe6e20b | 2692 | |
47ad8475 AA |
2693 | copy_user_huge_page(new_page, old_page, address, vma, |
2694 | pages_per_huge_page(h)); | |
0ed361de | 2695 | __SetPageUptodate(new_page); |
1e8f889b | 2696 | |
2ec74c3e SG |
2697 | mmun_start = address & huge_page_mask(h); |
2698 | mmun_end = mmun_start + huge_page_size(h); | |
2699 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
b76c8cfb | 2700 | /* |
cb900f41 | 2701 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
2702 | * before the page tables are altered |
2703 | */ | |
cb900f41 | 2704 | spin_lock(ptl); |
a5516438 | 2705 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
7f2e9525 | 2706 | if (likely(pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
2707 | ClearPagePrivate(new_page); |
2708 | ||
1e8f889b | 2709 | /* Break COW */ |
8fe627ec | 2710 | huge_ptep_clear_flush(vma, address, ptep); |
1e8f889b DG |
2711 | set_huge_pte_at(mm, address, ptep, |
2712 | make_huge_pte(vma, new_page, 1)); | |
0fe6e20b | 2713 | page_remove_rmap(old_page); |
cd67f0d2 | 2714 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
2715 | /* Make the old page be freed below */ |
2716 | new_page = old_page; | |
2717 | } | |
cb900f41 | 2718 | spin_unlock(ptl); |
2ec74c3e | 2719 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
1e8f889b DG |
2720 | page_cache_release(new_page); |
2721 | page_cache_release(old_page); | |
8312034f JK |
2722 | |
2723 | /* Caller expects lock to be held */ | |
cb900f41 | 2724 | spin_lock(ptl); |
83c54070 | 2725 | return 0; |
1e8f889b DG |
2726 | } |
2727 | ||
04f2cbe3 | 2728 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
2729 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
2730 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
2731 | { |
2732 | struct address_space *mapping; | |
e7c4b0bf | 2733 | pgoff_t idx; |
04f2cbe3 MG |
2734 | |
2735 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 2736 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
2737 | |
2738 | return find_lock_page(mapping, idx); | |
2739 | } | |
2740 | ||
3ae77f43 HD |
2741 | /* |
2742 | * Return whether there is a pagecache page to back given address within VMA. | |
2743 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
2744 | */ | |
2745 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
2746 | struct vm_area_struct *vma, unsigned long address) |
2747 | { | |
2748 | struct address_space *mapping; | |
2749 | pgoff_t idx; | |
2750 | struct page *page; | |
2751 | ||
2752 | mapping = vma->vm_file->f_mapping; | |
2753 | idx = vma_hugecache_offset(h, vma, address); | |
2754 | ||
2755 | page = find_get_page(mapping, idx); | |
2756 | if (page) | |
2757 | put_page(page); | |
2758 | return page != NULL; | |
2759 | } | |
2760 | ||
a1ed3dda | 2761 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 2762 | unsigned long address, pte_t *ptep, unsigned int flags) |
ac9b9c66 | 2763 | { |
a5516438 | 2764 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 2765 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 2766 | int anon_rmap = 0; |
e7c4b0bf | 2767 | pgoff_t idx; |
4c887265 | 2768 | unsigned long size; |
4c887265 AL |
2769 | struct page *page; |
2770 | struct address_space *mapping; | |
1e8f889b | 2771 | pte_t new_pte; |
cb900f41 | 2772 | spinlock_t *ptl; |
4c887265 | 2773 | |
04f2cbe3 MG |
2774 | /* |
2775 | * Currently, we are forced to kill the process in the event the | |
2776 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 2777 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
2778 | */ |
2779 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
ffb22af5 AM |
2780 | pr_warning("PID %d killed due to inadequate hugepage pool\n", |
2781 | current->pid); | |
04f2cbe3 MG |
2782 | return ret; |
2783 | } | |
2784 | ||
4c887265 | 2785 | mapping = vma->vm_file->f_mapping; |
a5516438 | 2786 | idx = vma_hugecache_offset(h, vma, address); |
4c887265 AL |
2787 | |
2788 | /* | |
2789 | * Use page lock to guard against racing truncation | |
2790 | * before we get page_table_lock. | |
2791 | */ | |
6bda666a CL |
2792 | retry: |
2793 | page = find_lock_page(mapping, idx); | |
2794 | if (!page) { | |
a5516438 | 2795 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
2796 | if (idx >= size) |
2797 | goto out; | |
04f2cbe3 | 2798 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 2799 | if (IS_ERR(page)) { |
76dcee75 AK |
2800 | ret = PTR_ERR(page); |
2801 | if (ret == -ENOMEM) | |
2802 | ret = VM_FAULT_OOM; | |
2803 | else | |
2804 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
2805 | goto out; |
2806 | } | |
47ad8475 | 2807 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 2808 | __SetPageUptodate(page); |
ac9b9c66 | 2809 | |
f83a275d | 2810 | if (vma->vm_flags & VM_MAYSHARE) { |
6bda666a | 2811 | int err; |
45c682a6 | 2812 | struct inode *inode = mapping->host; |
6bda666a CL |
2813 | |
2814 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
2815 | if (err) { | |
2816 | put_page(page); | |
6bda666a CL |
2817 | if (err == -EEXIST) |
2818 | goto retry; | |
2819 | goto out; | |
2820 | } | |
07443a85 | 2821 | ClearPagePrivate(page); |
45c682a6 KC |
2822 | |
2823 | spin_lock(&inode->i_lock); | |
a5516438 | 2824 | inode->i_blocks += blocks_per_huge_page(h); |
45c682a6 | 2825 | spin_unlock(&inode->i_lock); |
23be7468 | 2826 | } else { |
6bda666a | 2827 | lock_page(page); |
0fe6e20b NH |
2828 | if (unlikely(anon_vma_prepare(vma))) { |
2829 | ret = VM_FAULT_OOM; | |
2830 | goto backout_unlocked; | |
2831 | } | |
409eb8c2 | 2832 | anon_rmap = 1; |
23be7468 | 2833 | } |
0fe6e20b | 2834 | } else { |
998b4382 NH |
2835 | /* |
2836 | * If memory error occurs between mmap() and fault, some process | |
2837 | * don't have hwpoisoned swap entry for errored virtual address. | |
2838 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
2839 | */ | |
2840 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 2841 | ret = VM_FAULT_HWPOISON | |
972dc4de | 2842 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
2843 | goto backout_unlocked; |
2844 | } | |
6bda666a | 2845 | } |
1e8f889b | 2846 | |
57303d80 AW |
2847 | /* |
2848 | * If we are going to COW a private mapping later, we examine the | |
2849 | * pending reservations for this page now. This will ensure that | |
2850 | * any allocations necessary to record that reservation occur outside | |
2851 | * the spinlock. | |
2852 | */ | |
788c7df4 | 2853 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) |
2b26736c AW |
2854 | if (vma_needs_reservation(h, vma, address) < 0) { |
2855 | ret = VM_FAULT_OOM; | |
2856 | goto backout_unlocked; | |
2857 | } | |
57303d80 | 2858 | |
cb900f41 KS |
2859 | ptl = huge_pte_lockptr(h, mm, ptep); |
2860 | spin_lock(ptl); | |
a5516438 | 2861 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
2862 | if (idx >= size) |
2863 | goto backout; | |
2864 | ||
83c54070 | 2865 | ret = 0; |
7f2e9525 | 2866 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
2867 | goto backout; |
2868 | ||
07443a85 JK |
2869 | if (anon_rmap) { |
2870 | ClearPagePrivate(page); | |
409eb8c2 | 2871 | hugepage_add_new_anon_rmap(page, vma, address); |
07443a85 | 2872 | } |
409eb8c2 HD |
2873 | else |
2874 | page_dup_rmap(page); | |
1e8f889b DG |
2875 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
2876 | && (vma->vm_flags & VM_SHARED))); | |
2877 | set_huge_pte_at(mm, address, ptep, new_pte); | |
2878 | ||
788c7df4 | 2879 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 2880 | /* Optimization, do the COW without a second fault */ |
cb900f41 | 2881 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); |
1e8f889b DG |
2882 | } |
2883 | ||
cb900f41 | 2884 | spin_unlock(ptl); |
4c887265 AL |
2885 | unlock_page(page); |
2886 | out: | |
ac9b9c66 | 2887 | return ret; |
4c887265 AL |
2888 | |
2889 | backout: | |
cb900f41 | 2890 | spin_unlock(ptl); |
2b26736c | 2891 | backout_unlocked: |
4c887265 AL |
2892 | unlock_page(page); |
2893 | put_page(page); | |
2894 | goto out; | |
ac9b9c66 HD |
2895 | } |
2896 | ||
86e5216f | 2897 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 2898 | unsigned long address, unsigned int flags) |
86e5216f AL |
2899 | { |
2900 | pte_t *ptep; | |
2901 | pte_t entry; | |
cb900f41 | 2902 | spinlock_t *ptl; |
1e8f889b | 2903 | int ret; |
0fe6e20b | 2904 | struct page *page = NULL; |
57303d80 | 2905 | struct page *pagecache_page = NULL; |
3935baa9 | 2906 | static DEFINE_MUTEX(hugetlb_instantiation_mutex); |
a5516438 | 2907 | struct hstate *h = hstate_vma(vma); |
86e5216f | 2908 | |
1e16a539 KH |
2909 | address &= huge_page_mask(h); |
2910 | ||
fd6a03ed NH |
2911 | ptep = huge_pte_offset(mm, address); |
2912 | if (ptep) { | |
2913 | entry = huge_ptep_get(ptep); | |
290408d4 | 2914 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 2915 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
2916 | return 0; |
2917 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 2918 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 2919 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
2920 | } |
2921 | ||
a5516438 | 2922 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); |
86e5216f AL |
2923 | if (!ptep) |
2924 | return VM_FAULT_OOM; | |
2925 | ||
3935baa9 DG |
2926 | /* |
2927 | * Serialize hugepage allocation and instantiation, so that we don't | |
2928 | * get spurious allocation failures if two CPUs race to instantiate | |
2929 | * the same page in the page cache. | |
2930 | */ | |
2931 | mutex_lock(&hugetlb_instantiation_mutex); | |
7f2e9525 GS |
2932 | entry = huge_ptep_get(ptep); |
2933 | if (huge_pte_none(entry)) { | |
788c7df4 | 2934 | ret = hugetlb_no_page(mm, vma, address, ptep, flags); |
b4d1d99f | 2935 | goto out_mutex; |
3935baa9 | 2936 | } |
86e5216f | 2937 | |
83c54070 | 2938 | ret = 0; |
1e8f889b | 2939 | |
57303d80 AW |
2940 | /* |
2941 | * If we are going to COW the mapping later, we examine the pending | |
2942 | * reservations for this page now. This will ensure that any | |
2943 | * allocations necessary to record that reservation occur outside the | |
2944 | * spinlock. For private mappings, we also lookup the pagecache | |
2945 | * page now as it is used to determine if a reservation has been | |
2946 | * consumed. | |
2947 | */ | |
106c992a | 2948 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
2b26736c AW |
2949 | if (vma_needs_reservation(h, vma, address) < 0) { |
2950 | ret = VM_FAULT_OOM; | |
b4d1d99f | 2951 | goto out_mutex; |
2b26736c | 2952 | } |
57303d80 | 2953 | |
f83a275d | 2954 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
2955 | pagecache_page = hugetlbfs_pagecache_page(h, |
2956 | vma, address); | |
2957 | } | |
2958 | ||
56c9cfb1 NH |
2959 | /* |
2960 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
2961 | * pagecache_page, so here we need take the former one | |
2962 | * when page != pagecache_page or !pagecache_page. | |
2963 | * Note that locking order is always pagecache_page -> page, | |
2964 | * so no worry about deadlock. | |
2965 | */ | |
2966 | page = pte_page(entry); | |
66aebce7 | 2967 | get_page(page); |
56c9cfb1 | 2968 | if (page != pagecache_page) |
0fe6e20b | 2969 | lock_page(page); |
0fe6e20b | 2970 | |
cb900f41 KS |
2971 | ptl = huge_pte_lockptr(h, mm, ptep); |
2972 | spin_lock(ptl); | |
1e8f889b | 2973 | /* Check for a racing update before calling hugetlb_cow */ |
b4d1d99f | 2974 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) |
cb900f41 | 2975 | goto out_ptl; |
b4d1d99f DG |
2976 | |
2977 | ||
788c7df4 | 2978 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 2979 | if (!huge_pte_write(entry)) { |
57303d80 | 2980 | ret = hugetlb_cow(mm, vma, address, ptep, entry, |
cb900f41 KS |
2981 | pagecache_page, ptl); |
2982 | goto out_ptl; | |
b4d1d99f | 2983 | } |
106c992a | 2984 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
2985 | } |
2986 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
2987 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
2988 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 2989 | update_mmu_cache(vma, address, ptep); |
b4d1d99f | 2990 | |
cb900f41 KS |
2991 | out_ptl: |
2992 | spin_unlock(ptl); | |
57303d80 AW |
2993 | |
2994 | if (pagecache_page) { | |
2995 | unlock_page(pagecache_page); | |
2996 | put_page(pagecache_page); | |
2997 | } | |
1f64d69c DN |
2998 | if (page != pagecache_page) |
2999 | unlock_page(page); | |
66aebce7 | 3000 | put_page(page); |
57303d80 | 3001 | |
b4d1d99f | 3002 | out_mutex: |
3935baa9 | 3003 | mutex_unlock(&hugetlb_instantiation_mutex); |
1e8f889b DG |
3004 | |
3005 | return ret; | |
86e5216f AL |
3006 | } |
3007 | ||
28a35716 ML |
3008 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
3009 | struct page **pages, struct vm_area_struct **vmas, | |
3010 | unsigned long *position, unsigned long *nr_pages, | |
3011 | long i, unsigned int flags) | |
63551ae0 | 3012 | { |
d5d4b0aa CK |
3013 | unsigned long pfn_offset; |
3014 | unsigned long vaddr = *position; | |
28a35716 | 3015 | unsigned long remainder = *nr_pages; |
a5516438 | 3016 | struct hstate *h = hstate_vma(vma); |
63551ae0 | 3017 | |
63551ae0 | 3018 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 3019 | pte_t *pte; |
cb900f41 | 3020 | spinlock_t *ptl = NULL; |
2a15efc9 | 3021 | int absent; |
4c887265 | 3022 | struct page *page; |
63551ae0 | 3023 | |
4c887265 AL |
3024 | /* |
3025 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 3026 | * each hugepage. We have to make sure we get the |
4c887265 | 3027 | * first, for the page indexing below to work. |
cb900f41 KS |
3028 | * |
3029 | * Note that page table lock is not held when pte is null. | |
4c887265 | 3030 | */ |
a5516438 | 3031 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); |
cb900f41 KS |
3032 | if (pte) |
3033 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
3034 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
3035 | ||
3036 | /* | |
3037 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
3038 | * an error where there's an empty slot with no huge pagecache |
3039 | * to back it. This way, we avoid allocating a hugepage, and | |
3040 | * the sparse dumpfile avoids allocating disk blocks, but its | |
3041 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 3042 | */ |
3ae77f43 HD |
3043 | if (absent && (flags & FOLL_DUMP) && |
3044 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
3045 | if (pte) |
3046 | spin_unlock(ptl); | |
2a15efc9 HD |
3047 | remainder = 0; |
3048 | break; | |
3049 | } | |
63551ae0 | 3050 | |
9cc3a5bd NH |
3051 | /* |
3052 | * We need call hugetlb_fault for both hugepages under migration | |
3053 | * (in which case hugetlb_fault waits for the migration,) and | |
3054 | * hwpoisoned hugepages (in which case we need to prevent the | |
3055 | * caller from accessing to them.) In order to do this, we use | |
3056 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
3057 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
3058 | * both cases, and because we can't follow correct pages | |
3059 | * directly from any kind of swap entries. | |
3060 | */ | |
3061 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
3062 | ((flags & FOLL_WRITE) && |
3063 | !huge_pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 3064 | int ret; |
63551ae0 | 3065 | |
cb900f41 KS |
3066 | if (pte) |
3067 | spin_unlock(ptl); | |
2a15efc9 HD |
3068 | ret = hugetlb_fault(mm, vma, vaddr, |
3069 | (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); | |
a89182c7 | 3070 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 3071 | continue; |
63551ae0 | 3072 | |
4c887265 | 3073 | remainder = 0; |
4c887265 AL |
3074 | break; |
3075 | } | |
3076 | ||
a5516438 | 3077 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 3078 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 3079 | same_page: |
d6692183 | 3080 | if (pages) { |
2a15efc9 | 3081 | pages[i] = mem_map_offset(page, pfn_offset); |
a0368d4e | 3082 | get_page_foll(pages[i]); |
d6692183 | 3083 | } |
63551ae0 DG |
3084 | |
3085 | if (vmas) | |
3086 | vmas[i] = vma; | |
3087 | ||
3088 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 3089 | ++pfn_offset; |
63551ae0 DG |
3090 | --remainder; |
3091 | ++i; | |
d5d4b0aa | 3092 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 3093 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa CK |
3094 | /* |
3095 | * We use pfn_offset to avoid touching the pageframes | |
3096 | * of this compound page. | |
3097 | */ | |
3098 | goto same_page; | |
3099 | } | |
cb900f41 | 3100 | spin_unlock(ptl); |
63551ae0 | 3101 | } |
28a35716 | 3102 | *nr_pages = remainder; |
63551ae0 DG |
3103 | *position = vaddr; |
3104 | ||
2a15efc9 | 3105 | return i ? i : -EFAULT; |
63551ae0 | 3106 | } |
8f860591 | 3107 | |
7da4d641 | 3108 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
3109 | unsigned long address, unsigned long end, pgprot_t newprot) |
3110 | { | |
3111 | struct mm_struct *mm = vma->vm_mm; | |
3112 | unsigned long start = address; | |
3113 | pte_t *ptep; | |
3114 | pte_t pte; | |
a5516438 | 3115 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 3116 | unsigned long pages = 0; |
8f860591 ZY |
3117 | |
3118 | BUG_ON(address >= end); | |
3119 | flush_cache_range(vma, address, end); | |
3120 | ||
3d48ae45 | 3121 | mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex); |
a5516438 | 3122 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 3123 | spinlock_t *ptl; |
8f860591 ZY |
3124 | ptep = huge_pte_offset(mm, address); |
3125 | if (!ptep) | |
3126 | continue; | |
cb900f41 | 3127 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
3128 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3129 | pages++; | |
cb900f41 | 3130 | spin_unlock(ptl); |
39dde65c | 3131 | continue; |
7da4d641 | 3132 | } |
7f2e9525 | 3133 | if (!huge_pte_none(huge_ptep_get(ptep))) { |
8f860591 | 3134 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
106c992a | 3135 | pte = pte_mkhuge(huge_pte_modify(pte, newprot)); |
be7517d6 | 3136 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
8f860591 | 3137 | set_huge_pte_at(mm, address, ptep, pte); |
7da4d641 | 3138 | pages++; |
8f860591 | 3139 | } |
cb900f41 | 3140 | spin_unlock(ptl); |
8f860591 | 3141 | } |
d833352a MG |
3142 | /* |
3143 | * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare | |
3144 | * may have cleared our pud entry and done put_page on the page table: | |
3145 | * once we release i_mmap_mutex, another task can do the final put_page | |
3146 | * and that page table be reused and filled with junk. | |
3147 | */ | |
8f860591 | 3148 | flush_tlb_range(vma, start, end); |
d833352a | 3149 | mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex); |
7da4d641 PZ |
3150 | |
3151 | return pages << h->order; | |
8f860591 ZY |
3152 | } |
3153 | ||
a1e78772 MG |
3154 | int hugetlb_reserve_pages(struct inode *inode, |
3155 | long from, long to, | |
5a6fe125 | 3156 | struct vm_area_struct *vma, |
ca16d140 | 3157 | vm_flags_t vm_flags) |
e4e574b7 | 3158 | { |
17c9d12e | 3159 | long ret, chg; |
a5516438 | 3160 | struct hstate *h = hstate_inode(inode); |
90481622 | 3161 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 3162 | struct resv_map *resv_map; |
e4e574b7 | 3163 | |
17c9d12e MG |
3164 | /* |
3165 | * Only apply hugepage reservation if asked. At fault time, an | |
3166 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 3167 | * without using reserves |
17c9d12e | 3168 | */ |
ca16d140 | 3169 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
3170 | return 0; |
3171 | ||
a1e78772 MG |
3172 | /* |
3173 | * Shared mappings base their reservation on the number of pages that | |
3174 | * are already allocated on behalf of the file. Private mappings need | |
3175 | * to reserve the full area even if read-only as mprotect() may be | |
3176 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
3177 | */ | |
9119a41e JK |
3178 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
3179 | resv_map = inode->i_mapping->private_data; | |
3180 | ||
3181 | chg = region_chg(&resv_map->regions, from, to); | |
3182 | ||
3183 | } else { | |
3184 | resv_map = resv_map_alloc(); | |
17c9d12e MG |
3185 | if (!resv_map) |
3186 | return -ENOMEM; | |
3187 | ||
a1e78772 | 3188 | chg = to - from; |
84afd99b | 3189 | |
17c9d12e MG |
3190 | set_vma_resv_map(vma, resv_map); |
3191 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
3192 | } | |
3193 | ||
c50ac050 DH |
3194 | if (chg < 0) { |
3195 | ret = chg; | |
3196 | goto out_err; | |
3197 | } | |
8a630112 | 3198 | |
90481622 | 3199 | /* There must be enough pages in the subpool for the mapping */ |
c50ac050 DH |
3200 | if (hugepage_subpool_get_pages(spool, chg)) { |
3201 | ret = -ENOSPC; | |
3202 | goto out_err; | |
3203 | } | |
5a6fe125 MG |
3204 | |
3205 | /* | |
17c9d12e | 3206 | * Check enough hugepages are available for the reservation. |
90481622 | 3207 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 3208 | */ |
a5516438 | 3209 | ret = hugetlb_acct_memory(h, chg); |
68842c9b | 3210 | if (ret < 0) { |
90481622 | 3211 | hugepage_subpool_put_pages(spool, chg); |
c50ac050 | 3212 | goto out_err; |
68842c9b | 3213 | } |
17c9d12e MG |
3214 | |
3215 | /* | |
3216 | * Account for the reservations made. Shared mappings record regions | |
3217 | * that have reservations as they are shared by multiple VMAs. | |
3218 | * When the last VMA disappears, the region map says how much | |
3219 | * the reservation was and the page cache tells how much of | |
3220 | * the reservation was consumed. Private mappings are per-VMA and | |
3221 | * only the consumed reservations are tracked. When the VMA | |
3222 | * disappears, the original reservation is the VMA size and the | |
3223 | * consumed reservations are stored in the map. Hence, nothing | |
3224 | * else has to be done for private mappings here | |
3225 | */ | |
f83a275d | 3226 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
9119a41e | 3227 | region_add(&resv_map->regions, from, to); |
a43a8c39 | 3228 | return 0; |
c50ac050 | 3229 | out_err: |
4523e145 DH |
3230 | if (vma) |
3231 | resv_map_put(vma); | |
c50ac050 | 3232 | return ret; |
a43a8c39 CK |
3233 | } |
3234 | ||
3235 | void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) | |
3236 | { | |
a5516438 | 3237 | struct hstate *h = hstate_inode(inode); |
9119a41e JK |
3238 | struct resv_map *resv_map = inode->i_mapping->private_data; |
3239 | long chg = 0; | |
90481622 | 3240 | struct hugepage_subpool *spool = subpool_inode(inode); |
45c682a6 | 3241 | |
9119a41e JK |
3242 | if (resv_map) |
3243 | chg = region_truncate(&resv_map->regions, offset); | |
45c682a6 | 3244 | spin_lock(&inode->i_lock); |
e4c6f8be | 3245 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
3246 | spin_unlock(&inode->i_lock); |
3247 | ||
90481622 | 3248 | hugepage_subpool_put_pages(spool, (chg - freed)); |
a5516438 | 3249 | hugetlb_acct_memory(h, -(chg - freed)); |
a43a8c39 | 3250 | } |
93f70f90 | 3251 | |
3212b535 SC |
3252 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
3253 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
3254 | struct vm_area_struct *vma, | |
3255 | unsigned long addr, pgoff_t idx) | |
3256 | { | |
3257 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
3258 | svma->vm_start; | |
3259 | unsigned long sbase = saddr & PUD_MASK; | |
3260 | unsigned long s_end = sbase + PUD_SIZE; | |
3261 | ||
3262 | /* Allow segments to share if only one is marked locked */ | |
3263 | unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED; | |
3264 | unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED; | |
3265 | ||
3266 | /* | |
3267 | * match the virtual addresses, permission and the alignment of the | |
3268 | * page table page. | |
3269 | */ | |
3270 | if (pmd_index(addr) != pmd_index(saddr) || | |
3271 | vm_flags != svm_flags || | |
3272 | sbase < svma->vm_start || svma->vm_end < s_end) | |
3273 | return 0; | |
3274 | ||
3275 | return saddr; | |
3276 | } | |
3277 | ||
3278 | static int vma_shareable(struct vm_area_struct *vma, unsigned long addr) | |
3279 | { | |
3280 | unsigned long base = addr & PUD_MASK; | |
3281 | unsigned long end = base + PUD_SIZE; | |
3282 | ||
3283 | /* | |
3284 | * check on proper vm_flags and page table alignment | |
3285 | */ | |
3286 | if (vma->vm_flags & VM_MAYSHARE && | |
3287 | vma->vm_start <= base && end <= vma->vm_end) | |
3288 | return 1; | |
3289 | return 0; | |
3290 | } | |
3291 | ||
3292 | /* | |
3293 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
3294 | * and returns the corresponding pte. While this is not necessary for the | |
3295 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
3296 | * code much cleaner. pmd allocation is essential for the shared case because | |
3297 | * pud has to be populated inside the same i_mmap_mutex section - otherwise | |
3298 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a | |
3299 | * bad pmd for sharing. | |
3300 | */ | |
3301 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
3302 | { | |
3303 | struct vm_area_struct *vma = find_vma(mm, addr); | |
3304 | struct address_space *mapping = vma->vm_file->f_mapping; | |
3305 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
3306 | vma->vm_pgoff; | |
3307 | struct vm_area_struct *svma; | |
3308 | unsigned long saddr; | |
3309 | pte_t *spte = NULL; | |
3310 | pte_t *pte; | |
cb900f41 | 3311 | spinlock_t *ptl; |
3212b535 SC |
3312 | |
3313 | if (!vma_shareable(vma, addr)) | |
3314 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
3315 | ||
3316 | mutex_lock(&mapping->i_mmap_mutex); | |
3317 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { | |
3318 | if (svma == vma) | |
3319 | continue; | |
3320 | ||
3321 | saddr = page_table_shareable(svma, vma, addr, idx); | |
3322 | if (saddr) { | |
3323 | spte = huge_pte_offset(svma->vm_mm, saddr); | |
3324 | if (spte) { | |
3325 | get_page(virt_to_page(spte)); | |
3326 | break; | |
3327 | } | |
3328 | } | |
3329 | } | |
3330 | ||
3331 | if (!spte) | |
3332 | goto out; | |
3333 | ||
cb900f41 KS |
3334 | ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); |
3335 | spin_lock(ptl); | |
3212b535 SC |
3336 | if (pud_none(*pud)) |
3337 | pud_populate(mm, pud, | |
3338 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
3339 | else | |
3340 | put_page(virt_to_page(spte)); | |
cb900f41 | 3341 | spin_unlock(ptl); |
3212b535 SC |
3342 | out: |
3343 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3344 | mutex_unlock(&mapping->i_mmap_mutex); | |
3345 | return pte; | |
3346 | } | |
3347 | ||
3348 | /* | |
3349 | * unmap huge page backed by shared pte. | |
3350 | * | |
3351 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
3352 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
3353 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
3354 | * | |
cb900f41 | 3355 | * called with page table lock held. |
3212b535 SC |
3356 | * |
3357 | * returns: 1 successfully unmapped a shared pte page | |
3358 | * 0 the underlying pte page is not shared, or it is the last user | |
3359 | */ | |
3360 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
3361 | { | |
3362 | pgd_t *pgd = pgd_offset(mm, *addr); | |
3363 | pud_t *pud = pud_offset(pgd, *addr); | |
3364 | ||
3365 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
3366 | if (page_count(virt_to_page(ptep)) == 1) | |
3367 | return 0; | |
3368 | ||
3369 | pud_clear(pud); | |
3370 | put_page(virt_to_page(ptep)); | |
3371 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; | |
3372 | return 1; | |
3373 | } | |
9e5fc74c SC |
3374 | #define want_pmd_share() (1) |
3375 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
3376 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
3377 | { | |
3378 | return NULL; | |
3379 | } | |
3380 | #define want_pmd_share() (0) | |
3212b535 SC |
3381 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
3382 | ||
9e5fc74c SC |
3383 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
3384 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
3385 | unsigned long addr, unsigned long sz) | |
3386 | { | |
3387 | pgd_t *pgd; | |
3388 | pud_t *pud; | |
3389 | pte_t *pte = NULL; | |
3390 | ||
3391 | pgd = pgd_offset(mm, addr); | |
3392 | pud = pud_alloc(mm, pgd, addr); | |
3393 | if (pud) { | |
3394 | if (sz == PUD_SIZE) { | |
3395 | pte = (pte_t *)pud; | |
3396 | } else { | |
3397 | BUG_ON(sz != PMD_SIZE); | |
3398 | if (want_pmd_share() && pud_none(*pud)) | |
3399 | pte = huge_pmd_share(mm, addr, pud); | |
3400 | else | |
3401 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3402 | } | |
3403 | } | |
3404 | BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); | |
3405 | ||
3406 | return pte; | |
3407 | } | |
3408 | ||
3409 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) | |
3410 | { | |
3411 | pgd_t *pgd; | |
3412 | pud_t *pud; | |
3413 | pmd_t *pmd = NULL; | |
3414 | ||
3415 | pgd = pgd_offset(mm, addr); | |
3416 | if (pgd_present(*pgd)) { | |
3417 | pud = pud_offset(pgd, addr); | |
3418 | if (pud_present(*pud)) { | |
3419 | if (pud_huge(*pud)) | |
3420 | return (pte_t *)pud; | |
3421 | pmd = pmd_offset(pud, addr); | |
3422 | } | |
3423 | } | |
3424 | return (pte_t *) pmd; | |
3425 | } | |
3426 | ||
3427 | struct page * | |
3428 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, | |
3429 | pmd_t *pmd, int write) | |
3430 | { | |
3431 | struct page *page; | |
3432 | ||
3433 | page = pte_page(*(pte_t *)pmd); | |
3434 | if (page) | |
3435 | page += ((address & ~PMD_MASK) >> PAGE_SHIFT); | |
3436 | return page; | |
3437 | } | |
3438 | ||
3439 | struct page * | |
3440 | follow_huge_pud(struct mm_struct *mm, unsigned long address, | |
3441 | pud_t *pud, int write) | |
3442 | { | |
3443 | struct page *page; | |
3444 | ||
3445 | page = pte_page(*(pte_t *)pud); | |
3446 | if (page) | |
3447 | page += ((address & ~PUD_MASK) >> PAGE_SHIFT); | |
3448 | return page; | |
3449 | } | |
3450 | ||
3451 | #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */ | |
3452 | ||
3453 | /* Can be overriden by architectures */ | |
3454 | __attribute__((weak)) struct page * | |
3455 | follow_huge_pud(struct mm_struct *mm, unsigned long address, | |
3456 | pud_t *pud, int write) | |
3457 | { | |
3458 | BUG(); | |
3459 | return NULL; | |
3460 | } | |
3461 | ||
3462 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ | |
3463 | ||
d5bd9106 AK |
3464 | #ifdef CONFIG_MEMORY_FAILURE |
3465 | ||
6de2b1aa NH |
3466 | /* Should be called in hugetlb_lock */ |
3467 | static int is_hugepage_on_freelist(struct page *hpage) | |
3468 | { | |
3469 | struct page *page; | |
3470 | struct page *tmp; | |
3471 | struct hstate *h = page_hstate(hpage); | |
3472 | int nid = page_to_nid(hpage); | |
3473 | ||
3474 | list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru) | |
3475 | if (page == hpage) | |
3476 | return 1; | |
3477 | return 0; | |
3478 | } | |
3479 | ||
93f70f90 NH |
3480 | /* |
3481 | * This function is called from memory failure code. | |
3482 | * Assume the caller holds page lock of the head page. | |
3483 | */ | |
6de2b1aa | 3484 | int dequeue_hwpoisoned_huge_page(struct page *hpage) |
93f70f90 NH |
3485 | { |
3486 | struct hstate *h = page_hstate(hpage); | |
3487 | int nid = page_to_nid(hpage); | |
6de2b1aa | 3488 | int ret = -EBUSY; |
93f70f90 NH |
3489 | |
3490 | spin_lock(&hugetlb_lock); | |
6de2b1aa | 3491 | if (is_hugepage_on_freelist(hpage)) { |
56f2fb14 NH |
3492 | /* |
3493 | * Hwpoisoned hugepage isn't linked to activelist or freelist, | |
3494 | * but dangling hpage->lru can trigger list-debug warnings | |
3495 | * (this happens when we call unpoison_memory() on it), | |
3496 | * so let it point to itself with list_del_init(). | |
3497 | */ | |
3498 | list_del_init(&hpage->lru); | |
8c6c2ecb | 3499 | set_page_refcounted(hpage); |
6de2b1aa NH |
3500 | h->free_huge_pages--; |
3501 | h->free_huge_pages_node[nid]--; | |
3502 | ret = 0; | |
3503 | } | |
93f70f90 | 3504 | spin_unlock(&hugetlb_lock); |
6de2b1aa | 3505 | return ret; |
93f70f90 | 3506 | } |
6de2b1aa | 3507 | #endif |
31caf665 NH |
3508 | |
3509 | bool isolate_huge_page(struct page *page, struct list_head *list) | |
3510 | { | |
309381fe | 3511 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 NH |
3512 | if (!get_page_unless_zero(page)) |
3513 | return false; | |
3514 | spin_lock(&hugetlb_lock); | |
3515 | list_move_tail(&page->lru, list); | |
3516 | spin_unlock(&hugetlb_lock); | |
3517 | return true; | |
3518 | } | |
3519 | ||
3520 | void putback_active_hugepage(struct page *page) | |
3521 | { | |
309381fe | 3522 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 NH |
3523 | spin_lock(&hugetlb_lock); |
3524 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); | |
3525 | spin_unlock(&hugetlb_lock); | |
3526 | put_page(page); | |
3527 | } | |
c8721bbb NH |
3528 | |
3529 | bool is_hugepage_active(struct page *page) | |
3530 | { | |
309381fe | 3531 | VM_BUG_ON_PAGE(!PageHuge(page), page); |
c8721bbb NH |
3532 | /* |
3533 | * This function can be called for a tail page because the caller, | |
3534 | * scan_movable_pages, scans through a given pfn-range which typically | |
3535 | * covers one memory block. In systems using gigantic hugepage (1GB | |
3536 | * for x86_64,) a hugepage is larger than a memory block, and we don't | |
3537 | * support migrating such large hugepages for now, so return false | |
3538 | * when called for tail pages. | |
3539 | */ | |
3540 | if (PageTail(page)) | |
3541 | return false; | |
3542 | /* | |
3543 | * Refcount of a hwpoisoned hugepages is 1, but they are not active, | |
3544 | * so we should return false for them. | |
3545 | */ | |
3546 | if (unlikely(PageHWPoison(page))) | |
3547 | return false; | |
3548 | return page_count(page) > 0; | |
3549 | } |