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 MG |
37 | static gfp_t htlb_alloc_mask = GFP_HIGHUSER; |
38 | unsigned long hugepages_treat_as_movable; | |
a5516438 | 39 | |
c3f38a38 | 40 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
41 | unsigned int default_hstate_idx; |
42 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
43 | ||
53ba51d2 JT |
44 | __initdata LIST_HEAD(huge_boot_pages); |
45 | ||
e5ff2159 AK |
46 | /* for command line parsing */ |
47 | static struct hstate * __initdata parsed_hstate; | |
48 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 49 | static unsigned long __initdata default_hstate_size; |
e5ff2159 | 50 | |
3935baa9 | 51 | /* |
31caf665 NH |
52 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
53 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 54 | */ |
c3f38a38 | 55 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 56 | |
90481622 DG |
57 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
58 | { | |
59 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
60 | ||
61 | spin_unlock(&spool->lock); | |
62 | ||
63 | /* If no pages are used, and no other handles to the subpool | |
64 | * remain, free the subpool the subpool remain */ | |
65 | if (free) | |
66 | kfree(spool); | |
67 | } | |
68 | ||
69 | struct hugepage_subpool *hugepage_new_subpool(long nr_blocks) | |
70 | { | |
71 | struct hugepage_subpool *spool; | |
72 | ||
73 | spool = kmalloc(sizeof(*spool), GFP_KERNEL); | |
74 | if (!spool) | |
75 | return NULL; | |
76 | ||
77 | spin_lock_init(&spool->lock); | |
78 | spool->count = 1; | |
79 | spool->max_hpages = nr_blocks; | |
80 | spool->used_hpages = 0; | |
81 | ||
82 | return spool; | |
83 | } | |
84 | ||
85 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
86 | { | |
87 | spin_lock(&spool->lock); | |
88 | BUG_ON(!spool->count); | |
89 | spool->count--; | |
90 | unlock_or_release_subpool(spool); | |
91 | } | |
92 | ||
93 | static int hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
94 | long delta) | |
95 | { | |
96 | int ret = 0; | |
97 | ||
98 | if (!spool) | |
99 | return 0; | |
100 | ||
101 | spin_lock(&spool->lock); | |
102 | if ((spool->used_hpages + delta) <= spool->max_hpages) { | |
103 | spool->used_hpages += delta; | |
104 | } else { | |
105 | ret = -ENOMEM; | |
106 | } | |
107 | spin_unlock(&spool->lock); | |
108 | ||
109 | return ret; | |
110 | } | |
111 | ||
112 | static void hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
113 | long delta) | |
114 | { | |
115 | if (!spool) | |
116 | return; | |
117 | ||
118 | spin_lock(&spool->lock); | |
119 | spool->used_hpages -= delta; | |
120 | /* If hugetlbfs_put_super couldn't free spool due to | |
121 | * an outstanding quota reference, free it now. */ | |
122 | unlock_or_release_subpool(spool); | |
123 | } | |
124 | ||
125 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
126 | { | |
127 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
128 | } | |
129 | ||
130 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
131 | { | |
496ad9aa | 132 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
133 | } |
134 | ||
96822904 AW |
135 | /* |
136 | * Region tracking -- allows tracking of reservations and instantiated pages | |
137 | * across the pages in a mapping. | |
84afd99b AW |
138 | * |
139 | * The region data structures are protected by a combination of the mmap_sem | |
c748c262 | 140 | * and the hugetlb_instantiation_mutex. To access or modify a region the caller |
84afd99b | 141 | * must either hold the mmap_sem for write, or the mmap_sem for read and |
c748c262 | 142 | * the hugetlb_instantiation_mutex: |
84afd99b | 143 | * |
32f84528 | 144 | * down_write(&mm->mmap_sem); |
84afd99b | 145 | * or |
32f84528 CF |
146 | * down_read(&mm->mmap_sem); |
147 | * mutex_lock(&hugetlb_instantiation_mutex); | |
96822904 AW |
148 | */ |
149 | struct file_region { | |
150 | struct list_head link; | |
151 | long from; | |
152 | long to; | |
153 | }; | |
154 | ||
155 | static long region_add(struct list_head *head, long f, long t) | |
156 | { | |
157 | struct file_region *rg, *nrg, *trg; | |
158 | ||
159 | /* Locate the region we are either in or before. */ | |
160 | list_for_each_entry(rg, head, link) | |
161 | if (f <= rg->to) | |
162 | break; | |
163 | ||
164 | /* Round our left edge to the current segment if it encloses us. */ | |
165 | if (f > rg->from) | |
166 | f = rg->from; | |
167 | ||
168 | /* Check for and consume any regions we now overlap with. */ | |
169 | nrg = rg; | |
170 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
171 | if (&rg->link == head) | |
172 | break; | |
173 | if (rg->from > t) | |
174 | break; | |
175 | ||
176 | /* If this area reaches higher then extend our area to | |
177 | * include it completely. If this is not the first area | |
178 | * which we intend to reuse, free it. */ | |
179 | if (rg->to > t) | |
180 | t = rg->to; | |
181 | if (rg != nrg) { | |
182 | list_del(&rg->link); | |
183 | kfree(rg); | |
184 | } | |
185 | } | |
186 | nrg->from = f; | |
187 | nrg->to = t; | |
188 | return 0; | |
189 | } | |
190 | ||
191 | static long region_chg(struct list_head *head, long f, long t) | |
192 | { | |
193 | struct file_region *rg, *nrg; | |
194 | long chg = 0; | |
195 | ||
196 | /* Locate the region we are before or in. */ | |
197 | list_for_each_entry(rg, head, link) | |
198 | if (f <= rg->to) | |
199 | break; | |
200 | ||
201 | /* If we are below the current region then a new region is required. | |
202 | * Subtle, allocate a new region at the position but make it zero | |
203 | * size such that we can guarantee to record the reservation. */ | |
204 | if (&rg->link == head || t < rg->from) { | |
205 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
206 | if (!nrg) | |
207 | return -ENOMEM; | |
208 | nrg->from = f; | |
209 | nrg->to = f; | |
210 | INIT_LIST_HEAD(&nrg->link); | |
211 | list_add(&nrg->link, rg->link.prev); | |
212 | ||
213 | return t - f; | |
214 | } | |
215 | ||
216 | /* Round our left edge to the current segment if it encloses us. */ | |
217 | if (f > rg->from) | |
218 | f = rg->from; | |
219 | chg = t - f; | |
220 | ||
221 | /* Check for and consume any regions we now overlap with. */ | |
222 | list_for_each_entry(rg, rg->link.prev, link) { | |
223 | if (&rg->link == head) | |
224 | break; | |
225 | if (rg->from > t) | |
226 | return chg; | |
227 | ||
25985edc | 228 | /* We overlap with this area, if it extends further than |
96822904 AW |
229 | * us then we must extend ourselves. Account for its |
230 | * existing reservation. */ | |
231 | if (rg->to > t) { | |
232 | chg += rg->to - t; | |
233 | t = rg->to; | |
234 | } | |
235 | chg -= rg->to - rg->from; | |
236 | } | |
237 | return chg; | |
238 | } | |
239 | ||
240 | static long region_truncate(struct list_head *head, long end) | |
241 | { | |
242 | struct file_region *rg, *trg; | |
243 | long chg = 0; | |
244 | ||
245 | /* Locate the region we are either in or before. */ | |
246 | list_for_each_entry(rg, head, link) | |
247 | if (end <= rg->to) | |
248 | break; | |
249 | if (&rg->link == head) | |
250 | return 0; | |
251 | ||
252 | /* If we are in the middle of a region then adjust it. */ | |
253 | if (end > rg->from) { | |
254 | chg = rg->to - end; | |
255 | rg->to = end; | |
256 | rg = list_entry(rg->link.next, typeof(*rg), link); | |
257 | } | |
258 | ||
259 | /* Drop any remaining regions. */ | |
260 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
261 | if (&rg->link == head) | |
262 | break; | |
263 | chg += rg->to - rg->from; | |
264 | list_del(&rg->link); | |
265 | kfree(rg); | |
266 | } | |
267 | return chg; | |
268 | } | |
269 | ||
84afd99b AW |
270 | static long region_count(struct list_head *head, long f, long t) |
271 | { | |
272 | struct file_region *rg; | |
273 | long chg = 0; | |
274 | ||
275 | /* Locate each segment we overlap with, and count that overlap. */ | |
276 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
277 | long seg_from; |
278 | long seg_to; | |
84afd99b AW |
279 | |
280 | if (rg->to <= f) | |
281 | continue; | |
282 | if (rg->from >= t) | |
283 | break; | |
284 | ||
285 | seg_from = max(rg->from, f); | |
286 | seg_to = min(rg->to, t); | |
287 | ||
288 | chg += seg_to - seg_from; | |
289 | } | |
290 | ||
291 | return chg; | |
292 | } | |
293 | ||
e7c4b0bf AW |
294 | /* |
295 | * Convert the address within this vma to the page offset within | |
296 | * the mapping, in pagecache page units; huge pages here. | |
297 | */ | |
a5516438 AK |
298 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
299 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 300 | { |
a5516438 AK |
301 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
302 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
303 | } |
304 | ||
0fe6e20b NH |
305 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
306 | unsigned long address) | |
307 | { | |
308 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
309 | } | |
310 | ||
08fba699 MG |
311 | /* |
312 | * Return the size of the pages allocated when backing a VMA. In the majority | |
313 | * cases this will be same size as used by the page table entries. | |
314 | */ | |
315 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
316 | { | |
317 | struct hstate *hstate; | |
318 | ||
319 | if (!is_vm_hugetlb_page(vma)) | |
320 | return PAGE_SIZE; | |
321 | ||
322 | hstate = hstate_vma(vma); | |
323 | ||
2415cf12 | 324 | return 1UL << huge_page_shift(hstate); |
08fba699 | 325 | } |
f340ca0f | 326 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 327 | |
3340289d MG |
328 | /* |
329 | * Return the page size being used by the MMU to back a VMA. In the majority | |
330 | * of cases, the page size used by the kernel matches the MMU size. On | |
331 | * architectures where it differs, an architecture-specific version of this | |
332 | * function is required. | |
333 | */ | |
334 | #ifndef vma_mmu_pagesize | |
335 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) | |
336 | { | |
337 | return vma_kernel_pagesize(vma); | |
338 | } | |
339 | #endif | |
340 | ||
84afd99b AW |
341 | /* |
342 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
343 | * bits of the reservation map pointer, which are always clear due to | |
344 | * alignment. | |
345 | */ | |
346 | #define HPAGE_RESV_OWNER (1UL << 0) | |
347 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 348 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 349 | |
a1e78772 MG |
350 | /* |
351 | * These helpers are used to track how many pages are reserved for | |
352 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
353 | * is guaranteed to have their future faults succeed. | |
354 | * | |
355 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
356 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
357 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
358 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
359 | * |
360 | * The private mapping reservation is represented in a subtly different | |
361 | * manner to a shared mapping. A shared mapping has a region map associated | |
362 | * with the underlying file, this region map represents the backing file | |
363 | * pages which have ever had a reservation assigned which this persists even | |
364 | * after the page is instantiated. A private mapping has a region map | |
365 | * associated with the original mmap which is attached to all VMAs which | |
366 | * reference it, this region map represents those offsets which have consumed | |
367 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 368 | */ |
e7c4b0bf AW |
369 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
370 | { | |
371 | return (unsigned long)vma->vm_private_data; | |
372 | } | |
373 | ||
374 | static void set_vma_private_data(struct vm_area_struct *vma, | |
375 | unsigned long value) | |
376 | { | |
377 | vma->vm_private_data = (void *)value; | |
378 | } | |
379 | ||
84afd99b AW |
380 | struct resv_map { |
381 | struct kref refs; | |
382 | struct list_head regions; | |
383 | }; | |
384 | ||
2a4b3ded | 385 | static struct resv_map *resv_map_alloc(void) |
84afd99b AW |
386 | { |
387 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
388 | if (!resv_map) | |
389 | return NULL; | |
390 | ||
391 | kref_init(&resv_map->refs); | |
392 | INIT_LIST_HEAD(&resv_map->regions); | |
393 | ||
394 | return resv_map; | |
395 | } | |
396 | ||
2a4b3ded | 397 | static void resv_map_release(struct kref *ref) |
84afd99b AW |
398 | { |
399 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
400 | ||
401 | /* Clear out any active regions before we release the map. */ | |
402 | region_truncate(&resv_map->regions, 0); | |
403 | kfree(resv_map); | |
404 | } | |
405 | ||
406 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) | |
a1e78772 MG |
407 | { |
408 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 409 | if (!(vma->vm_flags & VM_MAYSHARE)) |
84afd99b AW |
410 | return (struct resv_map *)(get_vma_private_data(vma) & |
411 | ~HPAGE_RESV_MASK); | |
2a4b3ded | 412 | return NULL; |
a1e78772 MG |
413 | } |
414 | ||
84afd99b | 415 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 MG |
416 | { |
417 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 418 | VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); |
a1e78772 | 419 | |
84afd99b AW |
420 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
421 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
422 | } |
423 | ||
424 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
425 | { | |
04f2cbe3 | 426 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); |
f83a275d | 427 | VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); |
e7c4b0bf AW |
428 | |
429 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
430 | } |
431 | ||
432 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
433 | { | |
434 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
e7c4b0bf AW |
435 | |
436 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
437 | } |
438 | ||
04f2cbe3 | 439 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
440 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
441 | { | |
442 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 443 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
444 | vma->vm_private_data = (void *)0; |
445 | } | |
446 | ||
447 | /* Returns true if the VMA has associated reserve pages */ | |
af0ed73e | 448 | static int vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 449 | { |
af0ed73e JK |
450 | if (vma->vm_flags & VM_NORESERVE) { |
451 | /* | |
452 | * This address is already reserved by other process(chg == 0), | |
453 | * so, we should decrement reserved count. Without decrementing, | |
454 | * reserve count remains after releasing inode, because this | |
455 | * allocated page will go into page cache and is regarded as | |
456 | * coming from reserved pool in releasing step. Currently, we | |
457 | * don't have any other solution to deal with this situation | |
458 | * properly, so add work-around here. | |
459 | */ | |
460 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
461 | return 1; | |
462 | else | |
463 | return 0; | |
464 | } | |
a63884e9 JK |
465 | |
466 | /* Shared mappings always use reserves */ | |
f83a275d | 467 | if (vma->vm_flags & VM_MAYSHARE) |
7f09ca51 | 468 | return 1; |
a63884e9 JK |
469 | |
470 | /* | |
471 | * Only the process that called mmap() has reserves for | |
472 | * private mappings. | |
473 | */ | |
7f09ca51 MG |
474 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
475 | return 1; | |
a63884e9 | 476 | |
7f09ca51 | 477 | return 0; |
a1e78772 MG |
478 | } |
479 | ||
0ebabb41 NH |
480 | static void copy_gigantic_page(struct page *dst, struct page *src) |
481 | { | |
482 | int i; | |
483 | struct hstate *h = page_hstate(src); | |
484 | struct page *dst_base = dst; | |
485 | struct page *src_base = src; | |
486 | ||
487 | for (i = 0; i < pages_per_huge_page(h); ) { | |
488 | cond_resched(); | |
489 | copy_highpage(dst, src); | |
490 | ||
491 | i++; | |
492 | dst = mem_map_next(dst, dst_base, i); | |
493 | src = mem_map_next(src, src_base, i); | |
494 | } | |
495 | } | |
496 | ||
497 | void copy_huge_page(struct page *dst, struct page *src) | |
498 | { | |
499 | int i; | |
500 | struct hstate *h = page_hstate(src); | |
501 | ||
502 | if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) { | |
503 | copy_gigantic_page(dst, src); | |
504 | return; | |
505 | } | |
506 | ||
507 | might_sleep(); | |
508 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
509 | cond_resched(); | |
510 | copy_highpage(dst + i, src + i); | |
511 | } | |
512 | } | |
513 | ||
a5516438 | 514 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
515 | { |
516 | int nid = page_to_nid(page); | |
0edaecfa | 517 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
518 | h->free_huge_pages++; |
519 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
520 | } |
521 | ||
bf50bab2 NH |
522 | static struct page *dequeue_huge_page_node(struct hstate *h, int nid) |
523 | { | |
524 | struct page *page; | |
525 | ||
c8721bbb NH |
526 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
527 | if (!is_migrate_isolate_page(page)) | |
528 | break; | |
529 | /* | |
530 | * if 'non-isolated free hugepage' not found on the list, | |
531 | * the allocation fails. | |
532 | */ | |
533 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 534 | return NULL; |
0edaecfa | 535 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 536 | set_page_refcounted(page); |
bf50bab2 NH |
537 | h->free_huge_pages--; |
538 | h->free_huge_pages_node[nid]--; | |
539 | return page; | |
540 | } | |
541 | ||
a5516438 AK |
542 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
543 | struct vm_area_struct *vma, | |
af0ed73e JK |
544 | unsigned long address, int avoid_reserve, |
545 | long chg) | |
1da177e4 | 546 | { |
b1c12cbc | 547 | struct page *page = NULL; |
480eccf9 | 548 | struct mempolicy *mpol; |
19770b32 | 549 | nodemask_t *nodemask; |
c0ff7453 | 550 | struct zonelist *zonelist; |
dd1a239f MG |
551 | struct zone *zone; |
552 | struct zoneref *z; | |
cc9a6c87 | 553 | unsigned int cpuset_mems_cookie; |
1da177e4 | 554 | |
a1e78772 MG |
555 | /* |
556 | * A child process with MAP_PRIVATE mappings created by their parent | |
557 | * have no page reserves. This check ensures that reservations are | |
558 | * not "stolen". The child may still get SIGKILLed | |
559 | */ | |
af0ed73e | 560 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 561 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 562 | goto err; |
a1e78772 | 563 | |
04f2cbe3 | 564 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 565 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 566 | goto err; |
04f2cbe3 | 567 | |
9966c4bb JK |
568 | retry_cpuset: |
569 | cpuset_mems_cookie = get_mems_allowed(); | |
570 | zonelist = huge_zonelist(vma, address, | |
571 | htlb_alloc_mask, &mpol, &nodemask); | |
572 | ||
19770b32 MG |
573 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
574 | MAX_NR_ZONES - 1, nodemask) { | |
bf50bab2 NH |
575 | if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) { |
576 | page = dequeue_huge_page_node(h, zone_to_nid(zone)); | |
577 | if (page) { | |
af0ed73e JK |
578 | if (avoid_reserve) |
579 | break; | |
580 | if (!vma_has_reserves(vma, chg)) | |
581 | break; | |
582 | ||
07443a85 | 583 | SetPagePrivate(page); |
af0ed73e | 584 | h->resv_huge_pages--; |
bf50bab2 NH |
585 | break; |
586 | } | |
3abf7afd | 587 | } |
1da177e4 | 588 | } |
cc9a6c87 | 589 | |
52cd3b07 | 590 | mpol_cond_put(mpol); |
cc9a6c87 MG |
591 | if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page)) |
592 | goto retry_cpuset; | |
1da177e4 | 593 | return page; |
cc9a6c87 MG |
594 | |
595 | err: | |
cc9a6c87 | 596 | return NULL; |
1da177e4 LT |
597 | } |
598 | ||
a5516438 | 599 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
600 | { |
601 | int i; | |
a5516438 | 602 | |
18229df5 AW |
603 | VM_BUG_ON(h->order >= MAX_ORDER); |
604 | ||
a5516438 AK |
605 | h->nr_huge_pages--; |
606 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
607 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
608 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
609 | 1 << PG_referenced | 1 << PG_dirty | | |
610 | 1 << PG_active | 1 << PG_reserved | | |
611 | 1 << PG_private | 1 << PG_writeback); | |
6af2acb6 | 612 | } |
9dd540e2 | 613 | VM_BUG_ON(hugetlb_cgroup_from_page(page)); |
6af2acb6 AL |
614 | set_compound_page_dtor(page, NULL); |
615 | set_page_refcounted(page); | |
7f2e9525 | 616 | arch_release_hugepage(page); |
a5516438 | 617 | __free_pages(page, huge_page_order(h)); |
6af2acb6 AL |
618 | } |
619 | ||
e5ff2159 AK |
620 | struct hstate *size_to_hstate(unsigned long size) |
621 | { | |
622 | struct hstate *h; | |
623 | ||
624 | for_each_hstate(h) { | |
625 | if (huge_page_size(h) == size) | |
626 | return h; | |
627 | } | |
628 | return NULL; | |
629 | } | |
630 | ||
27a85ef1 DG |
631 | static void free_huge_page(struct page *page) |
632 | { | |
a5516438 AK |
633 | /* |
634 | * Can't pass hstate in here because it is called from the | |
635 | * compound page destructor. | |
636 | */ | |
e5ff2159 | 637 | struct hstate *h = page_hstate(page); |
7893d1d5 | 638 | int nid = page_to_nid(page); |
90481622 DG |
639 | struct hugepage_subpool *spool = |
640 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 641 | bool restore_reserve; |
27a85ef1 | 642 | |
e5df70ab | 643 | set_page_private(page, 0); |
23be7468 | 644 | page->mapping = NULL; |
7893d1d5 | 645 | BUG_ON(page_count(page)); |
0fe6e20b | 646 | BUG_ON(page_mapcount(page)); |
07443a85 | 647 | restore_reserve = PagePrivate(page); |
27a85ef1 DG |
648 | |
649 | spin_lock(&hugetlb_lock); | |
6d76dcf4 AK |
650 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
651 | pages_per_huge_page(h), page); | |
07443a85 JK |
652 | if (restore_reserve) |
653 | h->resv_huge_pages++; | |
654 | ||
aa888a74 | 655 | if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { |
0edaecfa AK |
656 | /* remove the page from active list */ |
657 | list_del(&page->lru); | |
a5516438 AK |
658 | update_and_free_page(h, page); |
659 | h->surplus_huge_pages--; | |
660 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 661 | } else { |
5d3a551c | 662 | arch_clear_hugepage_flags(page); |
a5516438 | 663 | enqueue_huge_page(h, page); |
7893d1d5 | 664 | } |
27a85ef1 | 665 | spin_unlock(&hugetlb_lock); |
90481622 | 666 | hugepage_subpool_put_pages(spool, 1); |
27a85ef1 DG |
667 | } |
668 | ||
a5516438 | 669 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 670 | { |
0edaecfa | 671 | INIT_LIST_HEAD(&page->lru); |
b7ba30c6 AK |
672 | set_compound_page_dtor(page, free_huge_page); |
673 | spin_lock(&hugetlb_lock); | |
9dd540e2 | 674 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
675 | h->nr_huge_pages++; |
676 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 AK |
677 | spin_unlock(&hugetlb_lock); |
678 | put_page(page); /* free it into the hugepage allocator */ | |
679 | } | |
680 | ||
20a0307c WF |
681 | static void prep_compound_gigantic_page(struct page *page, unsigned long order) |
682 | { | |
683 | int i; | |
684 | int nr_pages = 1 << order; | |
685 | struct page *p = page + 1; | |
686 | ||
687 | /* we rely on prep_new_huge_page to set the destructor */ | |
688 | set_compound_order(page, order); | |
689 | __SetPageHead(page); | |
690 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { | |
691 | __SetPageTail(p); | |
58a84aa9 | 692 | set_page_count(p, 0); |
20a0307c WF |
693 | p->first_page = page; |
694 | } | |
695 | } | |
696 | ||
7795912c AM |
697 | /* |
698 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
699 | * transparent huge pages. See the PageTransHuge() documentation for more | |
700 | * details. | |
701 | */ | |
20a0307c WF |
702 | int PageHuge(struct page *page) |
703 | { | |
704 | compound_page_dtor *dtor; | |
705 | ||
706 | if (!PageCompound(page)) | |
707 | return 0; | |
708 | ||
709 | page = compound_head(page); | |
710 | dtor = get_compound_page_dtor(page); | |
711 | ||
712 | return dtor == free_huge_page; | |
713 | } | |
43131e14 NH |
714 | EXPORT_SYMBOL_GPL(PageHuge); |
715 | ||
13d60f4b ZY |
716 | pgoff_t __basepage_index(struct page *page) |
717 | { | |
718 | struct page *page_head = compound_head(page); | |
719 | pgoff_t index = page_index(page_head); | |
720 | unsigned long compound_idx; | |
721 | ||
722 | if (!PageHuge(page_head)) | |
723 | return page_index(page); | |
724 | ||
725 | if (compound_order(page_head) >= MAX_ORDER) | |
726 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
727 | else | |
728 | compound_idx = page - page_head; | |
729 | ||
730 | return (index << compound_order(page_head)) + compound_idx; | |
731 | } | |
732 | ||
a5516438 | 733 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
1da177e4 | 734 | { |
1da177e4 | 735 | struct page *page; |
f96efd58 | 736 | |
aa888a74 AK |
737 | if (h->order >= MAX_ORDER) |
738 | return NULL; | |
739 | ||
6484eb3e | 740 | page = alloc_pages_exact_node(nid, |
551883ae NA |
741 | htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| |
742 | __GFP_REPEAT|__GFP_NOWARN, | |
a5516438 | 743 | huge_page_order(h)); |
1da177e4 | 744 | if (page) { |
7f2e9525 | 745 | if (arch_prepare_hugepage(page)) { |
caff3a2c | 746 | __free_pages(page, huge_page_order(h)); |
7b8ee84d | 747 | return NULL; |
7f2e9525 | 748 | } |
a5516438 | 749 | prep_new_huge_page(h, page, nid); |
1da177e4 | 750 | } |
63b4613c NA |
751 | |
752 | return page; | |
753 | } | |
754 | ||
9a76db09 | 755 | /* |
6ae11b27 LS |
756 | * common helper functions for hstate_next_node_to_{alloc|free}. |
757 | * We may have allocated or freed a huge page based on a different | |
758 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
759 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
760 | * node for alloc or free. | |
9a76db09 | 761 | */ |
6ae11b27 | 762 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) |
9a76db09 | 763 | { |
6ae11b27 | 764 | nid = next_node(nid, *nodes_allowed); |
9a76db09 | 765 | if (nid == MAX_NUMNODES) |
6ae11b27 | 766 | nid = first_node(*nodes_allowed); |
9a76db09 LS |
767 | VM_BUG_ON(nid >= MAX_NUMNODES); |
768 | ||
769 | return nid; | |
770 | } | |
771 | ||
6ae11b27 LS |
772 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) |
773 | { | |
774 | if (!node_isset(nid, *nodes_allowed)) | |
775 | nid = next_node_allowed(nid, nodes_allowed); | |
776 | return nid; | |
777 | } | |
778 | ||
5ced66c9 | 779 | /* |
6ae11b27 LS |
780 | * returns the previously saved node ["this node"] from which to |
781 | * allocate a persistent huge page for the pool and advance the | |
782 | * next node from which to allocate, handling wrap at end of node | |
783 | * mask. | |
5ced66c9 | 784 | */ |
6ae11b27 LS |
785 | static int hstate_next_node_to_alloc(struct hstate *h, |
786 | nodemask_t *nodes_allowed) | |
5ced66c9 | 787 | { |
6ae11b27 LS |
788 | int nid; |
789 | ||
790 | VM_BUG_ON(!nodes_allowed); | |
791 | ||
792 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
793 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
9a76db09 | 794 | |
9a76db09 | 795 | return nid; |
5ced66c9 AK |
796 | } |
797 | ||
e8c5c824 | 798 | /* |
6ae11b27 LS |
799 | * helper for free_pool_huge_page() - return the previously saved |
800 | * node ["this node"] from which to free a huge page. Advance the | |
801 | * next node id whether or not we find a free huge page to free so | |
802 | * that the next attempt to free addresses the next node. | |
e8c5c824 | 803 | */ |
6ae11b27 | 804 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) |
e8c5c824 | 805 | { |
6ae11b27 LS |
806 | int nid; |
807 | ||
808 | VM_BUG_ON(!nodes_allowed); | |
809 | ||
810 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
811 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
9a76db09 | 812 | |
9a76db09 | 813 | return nid; |
e8c5c824 LS |
814 | } |
815 | ||
b2261026 JK |
816 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ |
817 | for (nr_nodes = nodes_weight(*mask); \ | |
818 | nr_nodes > 0 && \ | |
819 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
820 | nr_nodes--) | |
821 | ||
822 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
823 | for (nr_nodes = nodes_weight(*mask); \ | |
824 | nr_nodes > 0 && \ | |
825 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
826 | nr_nodes--) | |
827 | ||
828 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) | |
829 | { | |
830 | struct page *page; | |
831 | int nr_nodes, node; | |
832 | int ret = 0; | |
833 | ||
834 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
835 | page = alloc_fresh_huge_page_node(h, node); | |
836 | if (page) { | |
837 | ret = 1; | |
838 | break; | |
839 | } | |
840 | } | |
841 | ||
842 | if (ret) | |
843 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
844 | else | |
845 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
846 | ||
847 | return ret; | |
848 | } | |
849 | ||
e8c5c824 LS |
850 | /* |
851 | * Free huge page from pool from next node to free. | |
852 | * Attempt to keep persistent huge pages more or less | |
853 | * balanced over allowed nodes. | |
854 | * Called with hugetlb_lock locked. | |
855 | */ | |
6ae11b27 LS |
856 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
857 | bool acct_surplus) | |
e8c5c824 | 858 | { |
b2261026 | 859 | int nr_nodes, node; |
e8c5c824 LS |
860 | int ret = 0; |
861 | ||
b2261026 | 862 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
863 | /* |
864 | * If we're returning unused surplus pages, only examine | |
865 | * nodes with surplus pages. | |
866 | */ | |
b2261026 JK |
867 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
868 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 869 | struct page *page = |
b2261026 | 870 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
871 | struct page, lru); |
872 | list_del(&page->lru); | |
873 | h->free_huge_pages--; | |
b2261026 | 874 | h->free_huge_pages_node[node]--; |
685f3457 LS |
875 | if (acct_surplus) { |
876 | h->surplus_huge_pages--; | |
b2261026 | 877 | h->surplus_huge_pages_node[node]--; |
685f3457 | 878 | } |
e8c5c824 LS |
879 | update_and_free_page(h, page); |
880 | ret = 1; | |
9a76db09 | 881 | break; |
e8c5c824 | 882 | } |
b2261026 | 883 | } |
e8c5c824 LS |
884 | |
885 | return ret; | |
886 | } | |
887 | ||
c8721bbb NH |
888 | /* |
889 | * Dissolve a given free hugepage into free buddy pages. This function does | |
890 | * nothing for in-use (including surplus) hugepages. | |
891 | */ | |
892 | static void dissolve_free_huge_page(struct page *page) | |
893 | { | |
894 | spin_lock(&hugetlb_lock); | |
895 | if (PageHuge(page) && !page_count(page)) { | |
896 | struct hstate *h = page_hstate(page); | |
897 | int nid = page_to_nid(page); | |
898 | list_del(&page->lru); | |
899 | h->free_huge_pages--; | |
900 | h->free_huge_pages_node[nid]--; | |
901 | update_and_free_page(h, page); | |
902 | } | |
903 | spin_unlock(&hugetlb_lock); | |
904 | } | |
905 | ||
906 | /* | |
907 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
908 | * make specified memory blocks removable from the system. | |
909 | * Note that start_pfn should aligned with (minimum) hugepage size. | |
910 | */ | |
911 | void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) | |
912 | { | |
913 | unsigned int order = 8 * sizeof(void *); | |
914 | unsigned long pfn; | |
915 | struct hstate *h; | |
916 | ||
917 | /* Set scan step to minimum hugepage size */ | |
918 | for_each_hstate(h) | |
919 | if (order > huge_page_order(h)) | |
920 | order = huge_page_order(h); | |
921 | VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order)); | |
922 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) | |
923 | dissolve_free_huge_page(pfn_to_page(pfn)); | |
924 | } | |
925 | ||
bf50bab2 | 926 | static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) |
7893d1d5 AL |
927 | { |
928 | struct page *page; | |
bf50bab2 | 929 | unsigned int r_nid; |
7893d1d5 | 930 | |
aa888a74 AK |
931 | if (h->order >= MAX_ORDER) |
932 | return NULL; | |
933 | ||
d1c3fb1f NA |
934 | /* |
935 | * Assume we will successfully allocate the surplus page to | |
936 | * prevent racing processes from causing the surplus to exceed | |
937 | * overcommit | |
938 | * | |
939 | * This however introduces a different race, where a process B | |
940 | * tries to grow the static hugepage pool while alloc_pages() is | |
941 | * called by process A. B will only examine the per-node | |
942 | * counters in determining if surplus huge pages can be | |
943 | * converted to normal huge pages in adjust_pool_surplus(). A | |
944 | * won't be able to increment the per-node counter, until the | |
945 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
946 | * no more huge pages can be converted from surplus to normal | |
947 | * state (and doesn't try to convert again). Thus, we have a | |
948 | * case where a surplus huge page exists, the pool is grown, and | |
949 | * the surplus huge page still exists after, even though it | |
950 | * should just have been converted to a normal huge page. This | |
951 | * does not leak memory, though, as the hugepage will be freed | |
952 | * once it is out of use. It also does not allow the counters to | |
953 | * go out of whack in adjust_pool_surplus() as we don't modify | |
954 | * the node values until we've gotten the hugepage and only the | |
955 | * per-node value is checked there. | |
956 | */ | |
957 | spin_lock(&hugetlb_lock); | |
a5516438 | 958 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
959 | spin_unlock(&hugetlb_lock); |
960 | return NULL; | |
961 | } else { | |
a5516438 AK |
962 | h->nr_huge_pages++; |
963 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
964 | } |
965 | spin_unlock(&hugetlb_lock); | |
966 | ||
bf50bab2 NH |
967 | if (nid == NUMA_NO_NODE) |
968 | page = alloc_pages(htlb_alloc_mask|__GFP_COMP| | |
969 | __GFP_REPEAT|__GFP_NOWARN, | |
970 | huge_page_order(h)); | |
971 | else | |
972 | page = alloc_pages_exact_node(nid, | |
973 | htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE| | |
974 | __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); | |
d1c3fb1f | 975 | |
caff3a2c GS |
976 | if (page && arch_prepare_hugepage(page)) { |
977 | __free_pages(page, huge_page_order(h)); | |
ea5768c7 | 978 | page = NULL; |
caff3a2c GS |
979 | } |
980 | ||
d1c3fb1f | 981 | spin_lock(&hugetlb_lock); |
7893d1d5 | 982 | if (page) { |
0edaecfa | 983 | INIT_LIST_HEAD(&page->lru); |
bf50bab2 | 984 | r_nid = page_to_nid(page); |
7893d1d5 | 985 | set_compound_page_dtor(page, free_huge_page); |
9dd540e2 | 986 | set_hugetlb_cgroup(page, NULL); |
d1c3fb1f NA |
987 | /* |
988 | * We incremented the global counters already | |
989 | */ | |
bf50bab2 NH |
990 | h->nr_huge_pages_node[r_nid]++; |
991 | h->surplus_huge_pages_node[r_nid]++; | |
3b116300 | 992 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 993 | } else { |
a5516438 AK |
994 | h->nr_huge_pages--; |
995 | h->surplus_huge_pages--; | |
3b116300 | 996 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 997 | } |
d1c3fb1f | 998 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
999 | |
1000 | return page; | |
1001 | } | |
1002 | ||
bf50bab2 NH |
1003 | /* |
1004 | * This allocation function is useful in the context where vma is irrelevant. | |
1005 | * E.g. soft-offlining uses this function because it only cares physical | |
1006 | * address of error page. | |
1007 | */ | |
1008 | struct page *alloc_huge_page_node(struct hstate *h, int nid) | |
1009 | { | |
4ef91848 | 1010 | struct page *page = NULL; |
bf50bab2 NH |
1011 | |
1012 | spin_lock(&hugetlb_lock); | |
4ef91848 JK |
1013 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
1014 | page = dequeue_huge_page_node(h, nid); | |
bf50bab2 NH |
1015 | spin_unlock(&hugetlb_lock); |
1016 | ||
94ae8ba7 | 1017 | if (!page) |
bf50bab2 NH |
1018 | page = alloc_buddy_huge_page(h, nid); |
1019 | ||
1020 | return page; | |
1021 | } | |
1022 | ||
e4e574b7 | 1023 | /* |
25985edc | 1024 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1025 | * of size 'delta'. |
1026 | */ | |
a5516438 | 1027 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1028 | { |
1029 | struct list_head surplus_list; | |
1030 | struct page *page, *tmp; | |
1031 | int ret, i; | |
1032 | int needed, allocated; | |
28073b02 | 1033 | bool alloc_ok = true; |
e4e574b7 | 1034 | |
a5516438 | 1035 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1036 | if (needed <= 0) { |
a5516438 | 1037 | h->resv_huge_pages += delta; |
e4e574b7 | 1038 | return 0; |
ac09b3a1 | 1039 | } |
e4e574b7 AL |
1040 | |
1041 | allocated = 0; | |
1042 | INIT_LIST_HEAD(&surplus_list); | |
1043 | ||
1044 | ret = -ENOMEM; | |
1045 | retry: | |
1046 | spin_unlock(&hugetlb_lock); | |
1047 | for (i = 0; i < needed; i++) { | |
bf50bab2 | 1048 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
28073b02 HD |
1049 | if (!page) { |
1050 | alloc_ok = false; | |
1051 | break; | |
1052 | } | |
e4e574b7 AL |
1053 | list_add(&page->lru, &surplus_list); |
1054 | } | |
28073b02 | 1055 | allocated += i; |
e4e574b7 AL |
1056 | |
1057 | /* | |
1058 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1059 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1060 | */ | |
1061 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1062 | needed = (h->resv_huge_pages + delta) - |
1063 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1064 | if (needed > 0) { |
1065 | if (alloc_ok) | |
1066 | goto retry; | |
1067 | /* | |
1068 | * We were not able to allocate enough pages to | |
1069 | * satisfy the entire reservation so we free what | |
1070 | * we've allocated so far. | |
1071 | */ | |
1072 | goto free; | |
1073 | } | |
e4e574b7 AL |
1074 | /* |
1075 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1076 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1077 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1078 | * allocator. Commit the entire reservation here to prevent another |
1079 | * process from stealing the pages as they are added to the pool but | |
1080 | * before they are reserved. | |
e4e574b7 AL |
1081 | */ |
1082 | needed += allocated; | |
a5516438 | 1083 | h->resv_huge_pages += delta; |
e4e574b7 | 1084 | ret = 0; |
a9869b83 | 1085 | |
19fc3f0a | 1086 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1087 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1088 | if ((--needed) < 0) |
1089 | break; | |
a9869b83 NH |
1090 | /* |
1091 | * This page is now managed by the hugetlb allocator and has | |
1092 | * no users -- drop the buddy allocator's reference. | |
1093 | */ | |
1094 | put_page_testzero(page); | |
1095 | VM_BUG_ON(page_count(page)); | |
a5516438 | 1096 | enqueue_huge_page(h, page); |
19fc3f0a | 1097 | } |
28073b02 | 1098 | free: |
b0365c8d | 1099 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1100 | |
1101 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1102 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1103 | put_page(page); | |
a9869b83 | 1104 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1105 | |
1106 | return ret; | |
1107 | } | |
1108 | ||
1109 | /* | |
1110 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
1111 | * allocated to satisfy the reservation must be explicitly freed if they were | |
1112 | * never used. | |
685f3457 | 1113 | * Called with hugetlb_lock held. |
e4e574b7 | 1114 | */ |
a5516438 AK |
1115 | static void return_unused_surplus_pages(struct hstate *h, |
1116 | unsigned long unused_resv_pages) | |
e4e574b7 | 1117 | { |
e4e574b7 AL |
1118 | unsigned long nr_pages; |
1119 | ||
ac09b3a1 | 1120 | /* Uncommit the reservation */ |
a5516438 | 1121 | h->resv_huge_pages -= unused_resv_pages; |
ac09b3a1 | 1122 | |
aa888a74 AK |
1123 | /* Cannot return gigantic pages currently */ |
1124 | if (h->order >= MAX_ORDER) | |
1125 | return; | |
1126 | ||
a5516438 | 1127 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1128 | |
685f3457 LS |
1129 | /* |
1130 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1131 | * evenly across all nodes with memory. Iterate across these nodes |
1132 | * until we can no longer free unreserved surplus pages. This occurs | |
1133 | * when the nodes with surplus pages have no free pages. | |
1134 | * free_pool_huge_page() will balance the the freed pages across the | |
1135 | * on-line nodes with memory and will handle the hstate accounting. | |
685f3457 LS |
1136 | */ |
1137 | while (nr_pages--) { | |
8cebfcd0 | 1138 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
685f3457 | 1139 | break; |
e4e574b7 AL |
1140 | } |
1141 | } | |
1142 | ||
c37f9fb1 AW |
1143 | /* |
1144 | * Determine if the huge page at addr within the vma has an associated | |
1145 | * reservation. Where it does not we will need to logically increase | |
90481622 DG |
1146 | * reservation and actually increase subpool usage before an allocation |
1147 | * can occur. Where any new reservation would be required the | |
1148 | * reservation change is prepared, but not committed. Once the page | |
1149 | * has been allocated from the subpool and instantiated the change should | |
1150 | * be committed via vma_commit_reservation. No action is required on | |
1151 | * failure. | |
c37f9fb1 | 1152 | */ |
e2f17d94 | 1153 | static long vma_needs_reservation(struct hstate *h, |
a5516438 | 1154 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 AW |
1155 | { |
1156 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1157 | struct inode *inode = mapping->host; | |
1158 | ||
f83a275d | 1159 | if (vma->vm_flags & VM_MAYSHARE) { |
a5516438 | 1160 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
c37f9fb1 AW |
1161 | return region_chg(&inode->i_mapping->private_list, |
1162 | idx, idx + 1); | |
1163 | ||
84afd99b AW |
1164 | } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1165 | return 1; | |
c37f9fb1 | 1166 | |
84afd99b | 1167 | } else { |
e2f17d94 | 1168 | long err; |
a5516438 | 1169 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
f522c3ac | 1170 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b | 1171 | |
f522c3ac | 1172 | err = region_chg(&resv->regions, idx, idx + 1); |
84afd99b AW |
1173 | if (err < 0) |
1174 | return err; | |
1175 | return 0; | |
1176 | } | |
c37f9fb1 | 1177 | } |
a5516438 AK |
1178 | static void vma_commit_reservation(struct hstate *h, |
1179 | struct vm_area_struct *vma, unsigned long addr) | |
c37f9fb1 AW |
1180 | { |
1181 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1182 | struct inode *inode = mapping->host; | |
1183 | ||
f83a275d | 1184 | if (vma->vm_flags & VM_MAYSHARE) { |
a5516438 | 1185 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
c37f9fb1 | 1186 | region_add(&inode->i_mapping->private_list, idx, idx + 1); |
84afd99b AW |
1187 | |
1188 | } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { | |
a5516438 | 1189 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
f522c3ac | 1190 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
1191 | |
1192 | /* Mark this page used in the map. */ | |
f522c3ac | 1193 | region_add(&resv->regions, idx, idx + 1); |
c37f9fb1 AW |
1194 | } |
1195 | } | |
1196 | ||
a1e78772 | 1197 | static struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 1198 | unsigned long addr, int avoid_reserve) |
1da177e4 | 1199 | { |
90481622 | 1200 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 1201 | struct hstate *h = hstate_vma(vma); |
348ea204 | 1202 | struct page *page; |
e2f17d94 | 1203 | long chg; |
6d76dcf4 AK |
1204 | int ret, idx; |
1205 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 1206 | |
6d76dcf4 | 1207 | idx = hstate_index(h); |
a1e78772 | 1208 | /* |
90481622 DG |
1209 | * Processes that did not create the mapping will have no |
1210 | * reserves and will not have accounted against subpool | |
1211 | * limit. Check that the subpool limit can be made before | |
1212 | * satisfying the allocation MAP_NORESERVE mappings may also | |
1213 | * need pages and subpool limit allocated allocated if no reserve | |
1214 | * mapping overlaps. | |
a1e78772 | 1215 | */ |
a5516438 | 1216 | chg = vma_needs_reservation(h, vma, addr); |
c37f9fb1 | 1217 | if (chg < 0) |
76dcee75 | 1218 | return ERR_PTR(-ENOMEM); |
8bb3f12e JK |
1219 | if (chg || avoid_reserve) |
1220 | if (hugepage_subpool_get_pages(spool, 1)) | |
76dcee75 | 1221 | return ERR_PTR(-ENOSPC); |
1da177e4 | 1222 | |
6d76dcf4 AK |
1223 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
1224 | if (ret) { | |
8bb3f12e JK |
1225 | if (chg || avoid_reserve) |
1226 | hugepage_subpool_put_pages(spool, 1); | |
6d76dcf4 AK |
1227 | return ERR_PTR(-ENOSPC); |
1228 | } | |
1da177e4 | 1229 | spin_lock(&hugetlb_lock); |
af0ed73e | 1230 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg); |
81a6fcae | 1231 | if (!page) { |
94ae8ba7 | 1232 | spin_unlock(&hugetlb_lock); |
bf50bab2 | 1233 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
68842c9b | 1234 | if (!page) { |
6d76dcf4 AK |
1235 | hugetlb_cgroup_uncharge_cgroup(idx, |
1236 | pages_per_huge_page(h), | |
1237 | h_cg); | |
8bb3f12e JK |
1238 | if (chg || avoid_reserve) |
1239 | hugepage_subpool_put_pages(spool, 1); | |
76dcee75 | 1240 | return ERR_PTR(-ENOSPC); |
68842c9b | 1241 | } |
79dbb236 AK |
1242 | spin_lock(&hugetlb_lock); |
1243 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 1244 | /* Fall through */ |
68842c9b | 1245 | } |
81a6fcae JK |
1246 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
1247 | spin_unlock(&hugetlb_lock); | |
348ea204 | 1248 | |
90481622 | 1249 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 1250 | |
a5516438 | 1251 | vma_commit_reservation(h, vma, addr); |
90d8b7e6 | 1252 | return page; |
b45b5bd6 DG |
1253 | } |
1254 | ||
74060e4d NH |
1255 | /* |
1256 | * alloc_huge_page()'s wrapper which simply returns the page if allocation | |
1257 | * succeeds, otherwise NULL. This function is called from new_vma_page(), | |
1258 | * where no ERR_VALUE is expected to be returned. | |
1259 | */ | |
1260 | struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, | |
1261 | unsigned long addr, int avoid_reserve) | |
1262 | { | |
1263 | struct page *page = alloc_huge_page(vma, addr, avoid_reserve); | |
1264 | if (IS_ERR(page)) | |
1265 | page = NULL; | |
1266 | return page; | |
1267 | } | |
1268 | ||
91f47662 | 1269 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
aa888a74 AK |
1270 | { |
1271 | struct huge_bootmem_page *m; | |
b2261026 | 1272 | int nr_nodes, node; |
aa888a74 | 1273 | |
b2261026 | 1274 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
1275 | void *addr; |
1276 | ||
b2261026 | 1277 | addr = __alloc_bootmem_node_nopanic(NODE_DATA(node), |
aa888a74 AK |
1278 | huge_page_size(h), huge_page_size(h), 0); |
1279 | ||
1280 | if (addr) { | |
1281 | /* | |
1282 | * Use the beginning of the huge page to store the | |
1283 | * huge_bootmem_page struct (until gather_bootmem | |
1284 | * puts them into the mem_map). | |
1285 | */ | |
1286 | m = addr; | |
91f47662 | 1287 | goto found; |
aa888a74 | 1288 | } |
aa888a74 AK |
1289 | } |
1290 | return 0; | |
1291 | ||
1292 | found: | |
1293 | BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); | |
1294 | /* Put them into a private list first because mem_map is not up yet */ | |
1295 | list_add(&m->list, &huge_boot_pages); | |
1296 | m->hstate = h; | |
1297 | return 1; | |
1298 | } | |
1299 | ||
18229df5 AW |
1300 | static void prep_compound_huge_page(struct page *page, int order) |
1301 | { | |
1302 | if (unlikely(order > (MAX_ORDER - 1))) | |
1303 | prep_compound_gigantic_page(page, order); | |
1304 | else | |
1305 | prep_compound_page(page, order); | |
1306 | } | |
1307 | ||
aa888a74 AK |
1308 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
1309 | static void __init gather_bootmem_prealloc(void) | |
1310 | { | |
1311 | struct huge_bootmem_page *m; | |
1312 | ||
1313 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 1314 | struct hstate *h = m->hstate; |
ee8f248d BB |
1315 | struct page *page; |
1316 | ||
1317 | #ifdef CONFIG_HIGHMEM | |
1318 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
1319 | free_bootmem_late((unsigned long)m, | |
1320 | sizeof(struct huge_bootmem_page)); | |
1321 | #else | |
1322 | page = virt_to_page(m); | |
1323 | #endif | |
aa888a74 AK |
1324 | __ClearPageReserved(page); |
1325 | WARN_ON(page_count(page) != 1); | |
18229df5 | 1326 | prep_compound_huge_page(page, h->order); |
aa888a74 | 1327 | prep_new_huge_page(h, page, page_to_nid(page)); |
b0320c7b RA |
1328 | /* |
1329 | * If we had gigantic hugepages allocated at boot time, we need | |
1330 | * to restore the 'stolen' pages to totalram_pages in order to | |
1331 | * fix confusing memory reports from free(1) and another | |
1332 | * side-effects, like CommitLimit going negative. | |
1333 | */ | |
1334 | if (h->order > (MAX_ORDER - 1)) | |
3dcc0571 | 1335 | adjust_managed_page_count(page, 1 << h->order); |
aa888a74 AK |
1336 | } |
1337 | } | |
1338 | ||
8faa8b07 | 1339 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
1340 | { |
1341 | unsigned long i; | |
a5516438 | 1342 | |
e5ff2159 | 1343 | for (i = 0; i < h->max_huge_pages; ++i) { |
aa888a74 AK |
1344 | if (h->order >= MAX_ORDER) { |
1345 | if (!alloc_bootmem_huge_page(h)) | |
1346 | break; | |
9b5e5d0f | 1347 | } else if (!alloc_fresh_huge_page(h, |
8cebfcd0 | 1348 | &node_states[N_MEMORY])) |
1da177e4 | 1349 | break; |
1da177e4 | 1350 | } |
8faa8b07 | 1351 | h->max_huge_pages = i; |
e5ff2159 AK |
1352 | } |
1353 | ||
1354 | static void __init hugetlb_init_hstates(void) | |
1355 | { | |
1356 | struct hstate *h; | |
1357 | ||
1358 | for_each_hstate(h) { | |
8faa8b07 AK |
1359 | /* oversize hugepages were init'ed in early boot */ |
1360 | if (h->order < MAX_ORDER) | |
1361 | hugetlb_hstate_alloc_pages(h); | |
e5ff2159 AK |
1362 | } |
1363 | } | |
1364 | ||
4abd32db AK |
1365 | static char * __init memfmt(char *buf, unsigned long n) |
1366 | { | |
1367 | if (n >= (1UL << 30)) | |
1368 | sprintf(buf, "%lu GB", n >> 30); | |
1369 | else if (n >= (1UL << 20)) | |
1370 | sprintf(buf, "%lu MB", n >> 20); | |
1371 | else | |
1372 | sprintf(buf, "%lu KB", n >> 10); | |
1373 | return buf; | |
1374 | } | |
1375 | ||
e5ff2159 AK |
1376 | static void __init report_hugepages(void) |
1377 | { | |
1378 | struct hstate *h; | |
1379 | ||
1380 | for_each_hstate(h) { | |
4abd32db | 1381 | char buf[32]; |
ffb22af5 | 1382 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
4abd32db AK |
1383 | memfmt(buf, huge_page_size(h)), |
1384 | h->free_huge_pages); | |
e5ff2159 AK |
1385 | } |
1386 | } | |
1387 | ||
1da177e4 | 1388 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
1389 | static void try_to_free_low(struct hstate *h, unsigned long count, |
1390 | nodemask_t *nodes_allowed) | |
1da177e4 | 1391 | { |
4415cc8d CL |
1392 | int i; |
1393 | ||
aa888a74 AK |
1394 | if (h->order >= MAX_ORDER) |
1395 | return; | |
1396 | ||
6ae11b27 | 1397 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 1398 | struct page *page, *next; |
a5516438 AK |
1399 | struct list_head *freel = &h->hugepage_freelists[i]; |
1400 | list_for_each_entry_safe(page, next, freel, lru) { | |
1401 | if (count >= h->nr_huge_pages) | |
6b0c880d | 1402 | return; |
1da177e4 LT |
1403 | if (PageHighMem(page)) |
1404 | continue; | |
1405 | list_del(&page->lru); | |
e5ff2159 | 1406 | update_and_free_page(h, page); |
a5516438 AK |
1407 | h->free_huge_pages--; |
1408 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
1409 | } |
1410 | } | |
1411 | } | |
1412 | #else | |
6ae11b27 LS |
1413 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
1414 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
1415 | { |
1416 | } | |
1417 | #endif | |
1418 | ||
20a0307c WF |
1419 | /* |
1420 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
1421 | * balanced by operating on them in a round-robin fashion. | |
1422 | * Returns 1 if an adjustment was made. | |
1423 | */ | |
6ae11b27 LS |
1424 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
1425 | int delta) | |
20a0307c | 1426 | { |
b2261026 | 1427 | int nr_nodes, node; |
20a0307c WF |
1428 | |
1429 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 1430 | |
b2261026 JK |
1431 | if (delta < 0) { |
1432 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1433 | if (h->surplus_huge_pages_node[node]) | |
1434 | goto found; | |
e8c5c824 | 1435 | } |
b2261026 JK |
1436 | } else { |
1437 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
1438 | if (h->surplus_huge_pages_node[node] < | |
1439 | h->nr_huge_pages_node[node]) | |
1440 | goto found; | |
e8c5c824 | 1441 | } |
b2261026 JK |
1442 | } |
1443 | return 0; | |
20a0307c | 1444 | |
b2261026 JK |
1445 | found: |
1446 | h->surplus_huge_pages += delta; | |
1447 | h->surplus_huge_pages_node[node] += delta; | |
1448 | return 1; | |
20a0307c WF |
1449 | } |
1450 | ||
a5516438 | 1451 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
1452 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
1453 | nodemask_t *nodes_allowed) | |
1da177e4 | 1454 | { |
7893d1d5 | 1455 | unsigned long min_count, ret; |
1da177e4 | 1456 | |
aa888a74 AK |
1457 | if (h->order >= MAX_ORDER) |
1458 | return h->max_huge_pages; | |
1459 | ||
7893d1d5 AL |
1460 | /* |
1461 | * Increase the pool size | |
1462 | * First take pages out of surplus state. Then make up the | |
1463 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f NA |
1464 | * |
1465 | * We might race with alloc_buddy_huge_page() here and be unable | |
1466 | * to convert a surplus huge page to a normal huge page. That is | |
1467 | * not critical, though, it just means the overall size of the | |
1468 | * pool might be one hugepage larger than it needs to be, but | |
1469 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 1470 | */ |
1da177e4 | 1471 | spin_lock(&hugetlb_lock); |
a5516438 | 1472 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 1473 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
1474 | break; |
1475 | } | |
1476 | ||
a5516438 | 1477 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
1478 | /* |
1479 | * If this allocation races such that we no longer need the | |
1480 | * page, free_huge_page will handle it by freeing the page | |
1481 | * and reducing the surplus. | |
1482 | */ | |
1483 | spin_unlock(&hugetlb_lock); | |
6ae11b27 | 1484 | ret = alloc_fresh_huge_page(h, nodes_allowed); |
7893d1d5 AL |
1485 | spin_lock(&hugetlb_lock); |
1486 | if (!ret) | |
1487 | goto out; | |
1488 | ||
536240f2 MG |
1489 | /* Bail for signals. Probably ctrl-c from user */ |
1490 | if (signal_pending(current)) | |
1491 | goto out; | |
7893d1d5 | 1492 | } |
7893d1d5 AL |
1493 | |
1494 | /* | |
1495 | * Decrease the pool size | |
1496 | * First return free pages to the buddy allocator (being careful | |
1497 | * to keep enough around to satisfy reservations). Then place | |
1498 | * pages into surplus state as needed so the pool will shrink | |
1499 | * to the desired size as pages become free. | |
d1c3fb1f NA |
1500 | * |
1501 | * By placing pages into the surplus state independent of the | |
1502 | * overcommit value, we are allowing the surplus pool size to | |
1503 | * exceed overcommit. There are few sane options here. Since | |
1504 | * alloc_buddy_huge_page() is checking the global counter, | |
1505 | * though, we'll note that we're not allowed to exceed surplus | |
1506 | * and won't grow the pool anywhere else. Not until one of the | |
1507 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 1508 | */ |
a5516438 | 1509 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 1510 | min_count = max(count, min_count); |
6ae11b27 | 1511 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 1512 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 1513 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 1514 | break; |
1da177e4 | 1515 | } |
a5516438 | 1516 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 1517 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
1518 | break; |
1519 | } | |
1520 | out: | |
a5516438 | 1521 | ret = persistent_huge_pages(h); |
1da177e4 | 1522 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 1523 | return ret; |
1da177e4 LT |
1524 | } |
1525 | ||
a3437870 NA |
1526 | #define HSTATE_ATTR_RO(_name) \ |
1527 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
1528 | ||
1529 | #define HSTATE_ATTR(_name) \ | |
1530 | static struct kobj_attribute _name##_attr = \ | |
1531 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
1532 | ||
1533 | static struct kobject *hugepages_kobj; | |
1534 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1535 | ||
9a305230 LS |
1536 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
1537 | ||
1538 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
1539 | { |
1540 | int i; | |
9a305230 | 1541 | |
a3437870 | 1542 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
1543 | if (hstate_kobjs[i] == kobj) { |
1544 | if (nidp) | |
1545 | *nidp = NUMA_NO_NODE; | |
a3437870 | 1546 | return &hstates[i]; |
9a305230 LS |
1547 | } |
1548 | ||
1549 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
1550 | } |
1551 | ||
06808b08 | 1552 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
1553 | struct kobj_attribute *attr, char *buf) |
1554 | { | |
9a305230 LS |
1555 | struct hstate *h; |
1556 | unsigned long nr_huge_pages; | |
1557 | int nid; | |
1558 | ||
1559 | h = kobj_to_hstate(kobj, &nid); | |
1560 | if (nid == NUMA_NO_NODE) | |
1561 | nr_huge_pages = h->nr_huge_pages; | |
1562 | else | |
1563 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
1564 | ||
1565 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 1566 | } |
adbe8726 | 1567 | |
06808b08 LS |
1568 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
1569 | struct kobject *kobj, struct kobj_attribute *attr, | |
1570 | const char *buf, size_t len) | |
a3437870 NA |
1571 | { |
1572 | int err; | |
9a305230 | 1573 | int nid; |
06808b08 | 1574 | unsigned long count; |
9a305230 | 1575 | struct hstate *h; |
bad44b5b | 1576 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 1577 | |
3dbb95f7 | 1578 | err = kstrtoul(buf, 10, &count); |
73ae31e5 | 1579 | if (err) |
adbe8726 | 1580 | goto out; |
a3437870 | 1581 | |
9a305230 | 1582 | h = kobj_to_hstate(kobj, &nid); |
adbe8726 EM |
1583 | if (h->order >= MAX_ORDER) { |
1584 | err = -EINVAL; | |
1585 | goto out; | |
1586 | } | |
1587 | ||
9a305230 LS |
1588 | if (nid == NUMA_NO_NODE) { |
1589 | /* | |
1590 | * global hstate attribute | |
1591 | */ | |
1592 | if (!(obey_mempolicy && | |
1593 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
1594 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 1595 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 LS |
1596 | } |
1597 | } else if (nodes_allowed) { | |
1598 | /* | |
1599 | * per node hstate attribute: adjust count to global, | |
1600 | * but restrict alloc/free to the specified node. | |
1601 | */ | |
1602 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
1603 | init_nodemask_of_node(nodes_allowed, nid); | |
1604 | } else | |
8cebfcd0 | 1605 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 | 1606 | |
06808b08 | 1607 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 1608 | |
8cebfcd0 | 1609 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
1610 | NODEMASK_FREE(nodes_allowed); |
1611 | ||
1612 | return len; | |
adbe8726 EM |
1613 | out: |
1614 | NODEMASK_FREE(nodes_allowed); | |
1615 | return err; | |
06808b08 LS |
1616 | } |
1617 | ||
1618 | static ssize_t nr_hugepages_show(struct kobject *kobj, | |
1619 | struct kobj_attribute *attr, char *buf) | |
1620 | { | |
1621 | return nr_hugepages_show_common(kobj, attr, buf); | |
1622 | } | |
1623 | ||
1624 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
1625 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1626 | { | |
1627 | return nr_hugepages_store_common(false, kobj, attr, buf, len); | |
a3437870 NA |
1628 | } |
1629 | HSTATE_ATTR(nr_hugepages); | |
1630 | ||
06808b08 LS |
1631 | #ifdef CONFIG_NUMA |
1632 | ||
1633 | /* | |
1634 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
1635 | * huge page alloc/free. | |
1636 | */ | |
1637 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
1638 | struct kobj_attribute *attr, char *buf) | |
1639 | { | |
1640 | return nr_hugepages_show_common(kobj, attr, buf); | |
1641 | } | |
1642 | ||
1643 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
1644 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1645 | { | |
1646 | return nr_hugepages_store_common(true, kobj, attr, buf, len); | |
1647 | } | |
1648 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
1649 | #endif | |
1650 | ||
1651 | ||
a3437870 NA |
1652 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
1653 | struct kobj_attribute *attr, char *buf) | |
1654 | { | |
9a305230 | 1655 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1656 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
1657 | } | |
adbe8726 | 1658 | |
a3437870 NA |
1659 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
1660 | struct kobj_attribute *attr, const char *buf, size_t count) | |
1661 | { | |
1662 | int err; | |
1663 | unsigned long input; | |
9a305230 | 1664 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 1665 | |
adbe8726 EM |
1666 | if (h->order >= MAX_ORDER) |
1667 | return -EINVAL; | |
1668 | ||
3dbb95f7 | 1669 | err = kstrtoul(buf, 10, &input); |
a3437870 | 1670 | if (err) |
73ae31e5 | 1671 | return err; |
a3437870 NA |
1672 | |
1673 | spin_lock(&hugetlb_lock); | |
1674 | h->nr_overcommit_huge_pages = input; | |
1675 | spin_unlock(&hugetlb_lock); | |
1676 | ||
1677 | return count; | |
1678 | } | |
1679 | HSTATE_ATTR(nr_overcommit_hugepages); | |
1680 | ||
1681 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
1682 | struct kobj_attribute *attr, char *buf) | |
1683 | { | |
9a305230 LS |
1684 | struct hstate *h; |
1685 | unsigned long free_huge_pages; | |
1686 | int nid; | |
1687 | ||
1688 | h = kobj_to_hstate(kobj, &nid); | |
1689 | if (nid == NUMA_NO_NODE) | |
1690 | free_huge_pages = h->free_huge_pages; | |
1691 | else | |
1692 | free_huge_pages = h->free_huge_pages_node[nid]; | |
1693 | ||
1694 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
1695 | } |
1696 | HSTATE_ATTR_RO(free_hugepages); | |
1697 | ||
1698 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
1699 | struct kobj_attribute *attr, char *buf) | |
1700 | { | |
9a305230 | 1701 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1702 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
1703 | } | |
1704 | HSTATE_ATTR_RO(resv_hugepages); | |
1705 | ||
1706 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
1707 | struct kobj_attribute *attr, char *buf) | |
1708 | { | |
9a305230 LS |
1709 | struct hstate *h; |
1710 | unsigned long surplus_huge_pages; | |
1711 | int nid; | |
1712 | ||
1713 | h = kobj_to_hstate(kobj, &nid); | |
1714 | if (nid == NUMA_NO_NODE) | |
1715 | surplus_huge_pages = h->surplus_huge_pages; | |
1716 | else | |
1717 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
1718 | ||
1719 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
1720 | } |
1721 | HSTATE_ATTR_RO(surplus_hugepages); | |
1722 | ||
1723 | static struct attribute *hstate_attrs[] = { | |
1724 | &nr_hugepages_attr.attr, | |
1725 | &nr_overcommit_hugepages_attr.attr, | |
1726 | &free_hugepages_attr.attr, | |
1727 | &resv_hugepages_attr.attr, | |
1728 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
1729 | #ifdef CONFIG_NUMA |
1730 | &nr_hugepages_mempolicy_attr.attr, | |
1731 | #endif | |
a3437870 NA |
1732 | NULL, |
1733 | }; | |
1734 | ||
1735 | static struct attribute_group hstate_attr_group = { | |
1736 | .attrs = hstate_attrs, | |
1737 | }; | |
1738 | ||
094e9539 JM |
1739 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
1740 | struct kobject **hstate_kobjs, | |
1741 | struct attribute_group *hstate_attr_group) | |
a3437870 NA |
1742 | { |
1743 | int retval; | |
972dc4de | 1744 | int hi = hstate_index(h); |
a3437870 | 1745 | |
9a305230 LS |
1746 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
1747 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
1748 | return -ENOMEM; |
1749 | ||
9a305230 | 1750 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 1751 | if (retval) |
9a305230 | 1752 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
1753 | |
1754 | return retval; | |
1755 | } | |
1756 | ||
1757 | static void __init hugetlb_sysfs_init(void) | |
1758 | { | |
1759 | struct hstate *h; | |
1760 | int err; | |
1761 | ||
1762 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
1763 | if (!hugepages_kobj) | |
1764 | return; | |
1765 | ||
1766 | for_each_hstate(h) { | |
9a305230 LS |
1767 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
1768 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 1769 | if (err) |
ffb22af5 | 1770 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
1771 | } |
1772 | } | |
1773 | ||
9a305230 LS |
1774 | #ifdef CONFIG_NUMA |
1775 | ||
1776 | /* | |
1777 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
1778 | * with node devices in node_devices[] using a parallel array. The array |
1779 | * index of a node device or _hstate == node id. | |
1780 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
1781 | * the base kernel, on the hugetlb module. |
1782 | */ | |
1783 | struct node_hstate { | |
1784 | struct kobject *hugepages_kobj; | |
1785 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1786 | }; | |
1787 | struct node_hstate node_hstates[MAX_NUMNODES]; | |
1788 | ||
1789 | /* | |
10fbcf4c | 1790 | * A subset of global hstate attributes for node devices |
9a305230 LS |
1791 | */ |
1792 | static struct attribute *per_node_hstate_attrs[] = { | |
1793 | &nr_hugepages_attr.attr, | |
1794 | &free_hugepages_attr.attr, | |
1795 | &surplus_hugepages_attr.attr, | |
1796 | NULL, | |
1797 | }; | |
1798 | ||
1799 | static struct attribute_group per_node_hstate_attr_group = { | |
1800 | .attrs = per_node_hstate_attrs, | |
1801 | }; | |
1802 | ||
1803 | /* | |
10fbcf4c | 1804 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
1805 | * Returns node id via non-NULL nidp. |
1806 | */ | |
1807 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
1808 | { | |
1809 | int nid; | |
1810 | ||
1811 | for (nid = 0; nid < nr_node_ids; nid++) { | |
1812 | struct node_hstate *nhs = &node_hstates[nid]; | |
1813 | int i; | |
1814 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
1815 | if (nhs->hstate_kobjs[i] == kobj) { | |
1816 | if (nidp) | |
1817 | *nidp = nid; | |
1818 | return &hstates[i]; | |
1819 | } | |
1820 | } | |
1821 | ||
1822 | BUG(); | |
1823 | return NULL; | |
1824 | } | |
1825 | ||
1826 | /* | |
10fbcf4c | 1827 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
1828 | * No-op if no hstate attributes attached. |
1829 | */ | |
3cd8b44f | 1830 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
1831 | { |
1832 | struct hstate *h; | |
10fbcf4c | 1833 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
1834 | |
1835 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 1836 | return; /* no hstate attributes */ |
9a305230 | 1837 | |
972dc4de AK |
1838 | for_each_hstate(h) { |
1839 | int idx = hstate_index(h); | |
1840 | if (nhs->hstate_kobjs[idx]) { | |
1841 | kobject_put(nhs->hstate_kobjs[idx]); | |
1842 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 1843 | } |
972dc4de | 1844 | } |
9a305230 LS |
1845 | |
1846 | kobject_put(nhs->hugepages_kobj); | |
1847 | nhs->hugepages_kobj = NULL; | |
1848 | } | |
1849 | ||
1850 | /* | |
10fbcf4c | 1851 | * hugetlb module exit: unregister hstate attributes from node devices |
9a305230 LS |
1852 | * that have them. |
1853 | */ | |
1854 | static void hugetlb_unregister_all_nodes(void) | |
1855 | { | |
1856 | int nid; | |
1857 | ||
1858 | /* | |
10fbcf4c | 1859 | * disable node device registrations. |
9a305230 LS |
1860 | */ |
1861 | register_hugetlbfs_with_node(NULL, NULL); | |
1862 | ||
1863 | /* | |
1864 | * remove hstate attributes from any nodes that have them. | |
1865 | */ | |
1866 | for (nid = 0; nid < nr_node_ids; nid++) | |
8732794b | 1867 | hugetlb_unregister_node(node_devices[nid]); |
9a305230 LS |
1868 | } |
1869 | ||
1870 | /* | |
10fbcf4c | 1871 | * Register hstate attributes for a single node device. |
9a305230 LS |
1872 | * No-op if attributes already registered. |
1873 | */ | |
3cd8b44f | 1874 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
1875 | { |
1876 | struct hstate *h; | |
10fbcf4c | 1877 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
1878 | int err; |
1879 | ||
1880 | if (nhs->hugepages_kobj) | |
1881 | return; /* already allocated */ | |
1882 | ||
1883 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 1884 | &node->dev.kobj); |
9a305230 LS |
1885 | if (!nhs->hugepages_kobj) |
1886 | return; | |
1887 | ||
1888 | for_each_hstate(h) { | |
1889 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
1890 | nhs->hstate_kobjs, | |
1891 | &per_node_hstate_attr_group); | |
1892 | if (err) { | |
ffb22af5 AM |
1893 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
1894 | h->name, node->dev.id); | |
9a305230 LS |
1895 | hugetlb_unregister_node(node); |
1896 | break; | |
1897 | } | |
1898 | } | |
1899 | } | |
1900 | ||
1901 | /* | |
9b5e5d0f | 1902 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
1903 | * devices of nodes that have memory. All on-line nodes should have |
1904 | * registered their associated device by this time. | |
9a305230 LS |
1905 | */ |
1906 | static void hugetlb_register_all_nodes(void) | |
1907 | { | |
1908 | int nid; | |
1909 | ||
8cebfcd0 | 1910 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 1911 | struct node *node = node_devices[nid]; |
10fbcf4c | 1912 | if (node->dev.id == nid) |
9a305230 LS |
1913 | hugetlb_register_node(node); |
1914 | } | |
1915 | ||
1916 | /* | |
10fbcf4c | 1917 | * Let the node device driver know we're here so it can |
9a305230 LS |
1918 | * [un]register hstate attributes on node hotplug. |
1919 | */ | |
1920 | register_hugetlbfs_with_node(hugetlb_register_node, | |
1921 | hugetlb_unregister_node); | |
1922 | } | |
1923 | #else /* !CONFIG_NUMA */ | |
1924 | ||
1925 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
1926 | { | |
1927 | BUG(); | |
1928 | if (nidp) | |
1929 | *nidp = -1; | |
1930 | return NULL; | |
1931 | } | |
1932 | ||
1933 | static void hugetlb_unregister_all_nodes(void) { } | |
1934 | ||
1935 | static void hugetlb_register_all_nodes(void) { } | |
1936 | ||
1937 | #endif | |
1938 | ||
a3437870 NA |
1939 | static void __exit hugetlb_exit(void) |
1940 | { | |
1941 | struct hstate *h; | |
1942 | ||
9a305230 LS |
1943 | hugetlb_unregister_all_nodes(); |
1944 | ||
a3437870 | 1945 | for_each_hstate(h) { |
972dc4de | 1946 | kobject_put(hstate_kobjs[hstate_index(h)]); |
a3437870 NA |
1947 | } |
1948 | ||
1949 | kobject_put(hugepages_kobj); | |
1950 | } | |
1951 | module_exit(hugetlb_exit); | |
1952 | ||
1953 | static int __init hugetlb_init(void) | |
1954 | { | |
0ef89d25 BH |
1955 | /* Some platform decide whether they support huge pages at boot |
1956 | * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when | |
1957 | * there is no such support | |
1958 | */ | |
1959 | if (HPAGE_SHIFT == 0) | |
1960 | return 0; | |
a3437870 | 1961 | |
e11bfbfc NP |
1962 | if (!size_to_hstate(default_hstate_size)) { |
1963 | default_hstate_size = HPAGE_SIZE; | |
1964 | if (!size_to_hstate(default_hstate_size)) | |
1965 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 1966 | } |
972dc4de | 1967 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
e11bfbfc NP |
1968 | if (default_hstate_max_huge_pages) |
1969 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
a3437870 NA |
1970 | |
1971 | hugetlb_init_hstates(); | |
aa888a74 | 1972 | gather_bootmem_prealloc(); |
a3437870 NA |
1973 | report_hugepages(); |
1974 | ||
1975 | hugetlb_sysfs_init(); | |
9a305230 | 1976 | hugetlb_register_all_nodes(); |
7179e7bf | 1977 | hugetlb_cgroup_file_init(); |
9a305230 | 1978 | |
a3437870 NA |
1979 | return 0; |
1980 | } | |
1981 | module_init(hugetlb_init); | |
1982 | ||
1983 | /* Should be called on processing a hugepagesz=... option */ | |
1984 | void __init hugetlb_add_hstate(unsigned order) | |
1985 | { | |
1986 | struct hstate *h; | |
8faa8b07 AK |
1987 | unsigned long i; |
1988 | ||
a3437870 | 1989 | if (size_to_hstate(PAGE_SIZE << order)) { |
ffb22af5 | 1990 | pr_warning("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
1991 | return; |
1992 | } | |
47d38344 | 1993 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 1994 | BUG_ON(order == 0); |
47d38344 | 1995 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
1996 | h->order = order; |
1997 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
1998 | h->nr_huge_pages = 0; |
1999 | h->free_huge_pages = 0; | |
2000 | for (i = 0; i < MAX_NUMNODES; ++i) | |
2001 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 2002 | INIT_LIST_HEAD(&h->hugepage_activelist); |
8cebfcd0 LJ |
2003 | h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); |
2004 | h->next_nid_to_free = first_node(node_states[N_MEMORY]); | |
a3437870 NA |
2005 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
2006 | huge_page_size(h)/1024); | |
8faa8b07 | 2007 | |
a3437870 NA |
2008 | parsed_hstate = h; |
2009 | } | |
2010 | ||
e11bfbfc | 2011 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
2012 | { |
2013 | unsigned long *mhp; | |
8faa8b07 | 2014 | static unsigned long *last_mhp; |
a3437870 NA |
2015 | |
2016 | /* | |
47d38344 | 2017 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
2018 | * so this hugepages= parameter goes to the "default hstate". |
2019 | */ | |
47d38344 | 2020 | if (!hugetlb_max_hstate) |
a3437870 NA |
2021 | mhp = &default_hstate_max_huge_pages; |
2022 | else | |
2023 | mhp = &parsed_hstate->max_huge_pages; | |
2024 | ||
8faa8b07 | 2025 | if (mhp == last_mhp) { |
ffb22af5 AM |
2026 | pr_warning("hugepages= specified twice without " |
2027 | "interleaving hugepagesz=, ignoring\n"); | |
8faa8b07 AK |
2028 | return 1; |
2029 | } | |
2030 | ||
a3437870 NA |
2031 | if (sscanf(s, "%lu", mhp) <= 0) |
2032 | *mhp = 0; | |
2033 | ||
8faa8b07 AK |
2034 | /* |
2035 | * Global state is always initialized later in hugetlb_init. | |
2036 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
2037 | * use the bootmem allocator. | |
2038 | */ | |
47d38344 | 2039 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
2040 | hugetlb_hstate_alloc_pages(parsed_hstate); |
2041 | ||
2042 | last_mhp = mhp; | |
2043 | ||
a3437870 NA |
2044 | return 1; |
2045 | } | |
e11bfbfc NP |
2046 | __setup("hugepages=", hugetlb_nrpages_setup); |
2047 | ||
2048 | static int __init hugetlb_default_setup(char *s) | |
2049 | { | |
2050 | default_hstate_size = memparse(s, &s); | |
2051 | return 1; | |
2052 | } | |
2053 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 2054 | |
8a213460 NA |
2055 | static unsigned int cpuset_mems_nr(unsigned int *array) |
2056 | { | |
2057 | int node; | |
2058 | unsigned int nr = 0; | |
2059 | ||
2060 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
2061 | nr += array[node]; | |
2062 | ||
2063 | return nr; | |
2064 | } | |
2065 | ||
2066 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
2067 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
2068 | struct ctl_table *table, int write, | |
2069 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 2070 | { |
e5ff2159 AK |
2071 | struct hstate *h = &default_hstate; |
2072 | unsigned long tmp; | |
08d4a246 | 2073 | int ret; |
e5ff2159 | 2074 | |
c033a93c | 2075 | tmp = h->max_huge_pages; |
e5ff2159 | 2076 | |
adbe8726 EM |
2077 | if (write && h->order >= MAX_ORDER) |
2078 | return -EINVAL; | |
2079 | ||
e5ff2159 AK |
2080 | table->data = &tmp; |
2081 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2082 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2083 | if (ret) | |
2084 | goto out; | |
e5ff2159 | 2085 | |
06808b08 | 2086 | if (write) { |
bad44b5b DR |
2087 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, |
2088 | GFP_KERNEL | __GFP_NORETRY); | |
06808b08 LS |
2089 | if (!(obey_mempolicy && |
2090 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
2091 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 2092 | nodes_allowed = &node_states[N_MEMORY]; |
06808b08 LS |
2093 | } |
2094 | h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); | |
2095 | ||
8cebfcd0 | 2096 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
2097 | NODEMASK_FREE(nodes_allowed); |
2098 | } | |
08d4a246 MH |
2099 | out: |
2100 | return ret; | |
1da177e4 | 2101 | } |
396faf03 | 2102 | |
06808b08 LS |
2103 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2104 | void __user *buffer, size_t *length, loff_t *ppos) | |
2105 | { | |
2106 | ||
2107 | return hugetlb_sysctl_handler_common(false, table, write, | |
2108 | buffer, length, ppos); | |
2109 | } | |
2110 | ||
2111 | #ifdef CONFIG_NUMA | |
2112 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2113 | void __user *buffer, size_t *length, loff_t *ppos) | |
2114 | { | |
2115 | return hugetlb_sysctl_handler_common(true, table, write, | |
2116 | buffer, length, ppos); | |
2117 | } | |
2118 | #endif /* CONFIG_NUMA */ | |
2119 | ||
396faf03 | 2120 | int hugetlb_treat_movable_handler(struct ctl_table *table, int write, |
8d65af78 | 2121 | void __user *buffer, |
396faf03 MG |
2122 | size_t *length, loff_t *ppos) |
2123 | { | |
8d65af78 | 2124 | proc_dointvec(table, write, buffer, length, ppos); |
396faf03 MG |
2125 | if (hugepages_treat_as_movable) |
2126 | htlb_alloc_mask = GFP_HIGHUSER_MOVABLE; | |
2127 | else | |
2128 | htlb_alloc_mask = GFP_HIGHUSER; | |
2129 | return 0; | |
2130 | } | |
2131 | ||
a3d0c6aa | 2132 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 2133 | void __user *buffer, |
a3d0c6aa NA |
2134 | size_t *length, loff_t *ppos) |
2135 | { | |
a5516438 | 2136 | struct hstate *h = &default_hstate; |
e5ff2159 | 2137 | unsigned long tmp; |
08d4a246 | 2138 | int ret; |
e5ff2159 | 2139 | |
c033a93c | 2140 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 2141 | |
adbe8726 EM |
2142 | if (write && h->order >= MAX_ORDER) |
2143 | return -EINVAL; | |
2144 | ||
e5ff2159 AK |
2145 | table->data = &tmp; |
2146 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2147 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2148 | if (ret) | |
2149 | goto out; | |
e5ff2159 AK |
2150 | |
2151 | if (write) { | |
2152 | spin_lock(&hugetlb_lock); | |
2153 | h->nr_overcommit_huge_pages = tmp; | |
2154 | spin_unlock(&hugetlb_lock); | |
2155 | } | |
08d4a246 MH |
2156 | out: |
2157 | return ret; | |
a3d0c6aa NA |
2158 | } |
2159 | ||
1da177e4 LT |
2160 | #endif /* CONFIG_SYSCTL */ |
2161 | ||
e1759c21 | 2162 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 2163 | { |
a5516438 | 2164 | struct hstate *h = &default_hstate; |
e1759c21 | 2165 | seq_printf(m, |
4f98a2fe RR |
2166 | "HugePages_Total: %5lu\n" |
2167 | "HugePages_Free: %5lu\n" | |
2168 | "HugePages_Rsvd: %5lu\n" | |
2169 | "HugePages_Surp: %5lu\n" | |
2170 | "Hugepagesize: %8lu kB\n", | |
a5516438 AK |
2171 | h->nr_huge_pages, |
2172 | h->free_huge_pages, | |
2173 | h->resv_huge_pages, | |
2174 | h->surplus_huge_pages, | |
2175 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
2176 | } |
2177 | ||
2178 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
2179 | { | |
a5516438 | 2180 | struct hstate *h = &default_hstate; |
1da177e4 LT |
2181 | return sprintf(buf, |
2182 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
2183 | "Node %d HugePages_Free: %5u\n" |
2184 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
2185 | nid, h->nr_huge_pages_node[nid], |
2186 | nid, h->free_huge_pages_node[nid], | |
2187 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
2188 | } |
2189 | ||
949f7ec5 DR |
2190 | void hugetlb_show_meminfo(void) |
2191 | { | |
2192 | struct hstate *h; | |
2193 | int nid; | |
2194 | ||
2195 | for_each_node_state(nid, N_MEMORY) | |
2196 | for_each_hstate(h) | |
2197 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
2198 | nid, | |
2199 | h->nr_huge_pages_node[nid], | |
2200 | h->free_huge_pages_node[nid], | |
2201 | h->surplus_huge_pages_node[nid], | |
2202 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
2203 | } | |
2204 | ||
1da177e4 LT |
2205 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
2206 | unsigned long hugetlb_total_pages(void) | |
2207 | { | |
d0028588 WL |
2208 | struct hstate *h; |
2209 | unsigned long nr_total_pages = 0; | |
2210 | ||
2211 | for_each_hstate(h) | |
2212 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
2213 | return nr_total_pages; | |
1da177e4 | 2214 | } |
1da177e4 | 2215 | |
a5516438 | 2216 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
2217 | { |
2218 | int ret = -ENOMEM; | |
2219 | ||
2220 | spin_lock(&hugetlb_lock); | |
2221 | /* | |
2222 | * When cpuset is configured, it breaks the strict hugetlb page | |
2223 | * reservation as the accounting is done on a global variable. Such | |
2224 | * reservation is completely rubbish in the presence of cpuset because | |
2225 | * the reservation is not checked against page availability for the | |
2226 | * current cpuset. Application can still potentially OOM'ed by kernel | |
2227 | * with lack of free htlb page in cpuset that the task is in. | |
2228 | * Attempt to enforce strict accounting with cpuset is almost | |
2229 | * impossible (or too ugly) because cpuset is too fluid that | |
2230 | * task or memory node can be dynamically moved between cpusets. | |
2231 | * | |
2232 | * The change of semantics for shared hugetlb mapping with cpuset is | |
2233 | * undesirable. However, in order to preserve some of the semantics, | |
2234 | * we fall back to check against current free page availability as | |
2235 | * a best attempt and hopefully to minimize the impact of changing | |
2236 | * semantics that cpuset has. | |
2237 | */ | |
2238 | if (delta > 0) { | |
a5516438 | 2239 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
2240 | goto out; |
2241 | ||
a5516438 AK |
2242 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
2243 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
2244 | goto out; |
2245 | } | |
2246 | } | |
2247 | ||
2248 | ret = 0; | |
2249 | if (delta < 0) | |
a5516438 | 2250 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
2251 | |
2252 | out: | |
2253 | spin_unlock(&hugetlb_lock); | |
2254 | return ret; | |
2255 | } | |
2256 | ||
84afd99b AW |
2257 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
2258 | { | |
f522c3ac | 2259 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
2260 | |
2261 | /* | |
2262 | * This new VMA should share its siblings reservation map if present. | |
2263 | * The VMA will only ever have a valid reservation map pointer where | |
2264 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 2265 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
2266 | * after this open call completes. It is therefore safe to take a |
2267 | * new reference here without additional locking. | |
2268 | */ | |
f522c3ac JK |
2269 | if (resv) |
2270 | kref_get(&resv->refs); | |
84afd99b AW |
2271 | } |
2272 | ||
c50ac050 DH |
2273 | static void resv_map_put(struct vm_area_struct *vma) |
2274 | { | |
f522c3ac | 2275 | struct resv_map *resv = vma_resv_map(vma); |
c50ac050 | 2276 | |
f522c3ac | 2277 | if (!resv) |
c50ac050 | 2278 | return; |
f522c3ac | 2279 | kref_put(&resv->refs, resv_map_release); |
c50ac050 DH |
2280 | } |
2281 | ||
a1e78772 MG |
2282 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
2283 | { | |
a5516438 | 2284 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 2285 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 2286 | struct hugepage_subpool *spool = subpool_vma(vma); |
84afd99b AW |
2287 | unsigned long reserve; |
2288 | unsigned long start; | |
2289 | unsigned long end; | |
2290 | ||
f522c3ac | 2291 | if (resv) { |
a5516438 AK |
2292 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
2293 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b AW |
2294 | |
2295 | reserve = (end - start) - | |
f522c3ac | 2296 | region_count(&resv->regions, start, end); |
84afd99b | 2297 | |
c50ac050 | 2298 | resv_map_put(vma); |
84afd99b | 2299 | |
7251ff78 | 2300 | if (reserve) { |
a5516438 | 2301 | hugetlb_acct_memory(h, -reserve); |
90481622 | 2302 | hugepage_subpool_put_pages(spool, reserve); |
7251ff78 | 2303 | } |
84afd99b | 2304 | } |
a1e78772 MG |
2305 | } |
2306 | ||
1da177e4 LT |
2307 | /* |
2308 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
2309 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
2310 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
2311 | * this far. | |
2312 | */ | |
d0217ac0 | 2313 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
2314 | { |
2315 | BUG(); | |
d0217ac0 | 2316 | return 0; |
1da177e4 LT |
2317 | } |
2318 | ||
f0f37e2f | 2319 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 2320 | .fault = hugetlb_vm_op_fault, |
84afd99b | 2321 | .open = hugetlb_vm_op_open, |
a1e78772 | 2322 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
2323 | }; |
2324 | ||
1e8f889b DG |
2325 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
2326 | int writable) | |
63551ae0 DG |
2327 | { |
2328 | pte_t entry; | |
2329 | ||
1e8f889b | 2330 | if (writable) { |
106c992a GS |
2331 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
2332 | vma->vm_page_prot))); | |
63551ae0 | 2333 | } else { |
106c992a GS |
2334 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
2335 | vma->vm_page_prot)); | |
63551ae0 DG |
2336 | } |
2337 | entry = pte_mkyoung(entry); | |
2338 | entry = pte_mkhuge(entry); | |
d9ed9faa | 2339 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
2340 | |
2341 | return entry; | |
2342 | } | |
2343 | ||
1e8f889b DG |
2344 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
2345 | unsigned long address, pte_t *ptep) | |
2346 | { | |
2347 | pte_t entry; | |
2348 | ||
106c992a | 2349 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 2350 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 2351 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
2352 | } |
2353 | ||
2354 | ||
63551ae0 DG |
2355 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
2356 | struct vm_area_struct *vma) | |
2357 | { | |
2358 | pte_t *src_pte, *dst_pte, entry; | |
2359 | struct page *ptepage; | |
1c59827d | 2360 | unsigned long addr; |
1e8f889b | 2361 | int cow; |
a5516438 AK |
2362 | struct hstate *h = hstate_vma(vma); |
2363 | unsigned long sz = huge_page_size(h); | |
1e8f889b DG |
2364 | |
2365 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 2366 | |
a5516438 | 2367 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
c74df32c HD |
2368 | src_pte = huge_pte_offset(src, addr); |
2369 | if (!src_pte) | |
2370 | continue; | |
a5516438 | 2371 | dst_pte = huge_pte_alloc(dst, addr, sz); |
63551ae0 DG |
2372 | if (!dst_pte) |
2373 | goto nomem; | |
c5c99429 LW |
2374 | |
2375 | /* If the pagetables are shared don't copy or take references */ | |
2376 | if (dst_pte == src_pte) | |
2377 | continue; | |
2378 | ||
c74df32c | 2379 | spin_lock(&dst->page_table_lock); |
46478758 | 2380 | spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING); |
7f2e9525 | 2381 | if (!huge_pte_none(huge_ptep_get(src_pte))) { |
1e8f889b | 2382 | if (cow) |
7f2e9525 GS |
2383 | huge_ptep_set_wrprotect(src, addr, src_pte); |
2384 | entry = huge_ptep_get(src_pte); | |
1c59827d HD |
2385 | ptepage = pte_page(entry); |
2386 | get_page(ptepage); | |
0fe6e20b | 2387 | page_dup_rmap(ptepage); |
1c59827d HD |
2388 | set_huge_pte_at(dst, addr, dst_pte, entry); |
2389 | } | |
2390 | spin_unlock(&src->page_table_lock); | |
c74df32c | 2391 | spin_unlock(&dst->page_table_lock); |
63551ae0 DG |
2392 | } |
2393 | return 0; | |
2394 | ||
2395 | nomem: | |
2396 | return -ENOMEM; | |
2397 | } | |
2398 | ||
290408d4 NH |
2399 | static int is_hugetlb_entry_migration(pte_t pte) |
2400 | { | |
2401 | swp_entry_t swp; | |
2402 | ||
2403 | if (huge_pte_none(pte) || pte_present(pte)) | |
2404 | return 0; | |
2405 | swp = pte_to_swp_entry(pte); | |
32f84528 | 2406 | if (non_swap_entry(swp) && is_migration_entry(swp)) |
290408d4 | 2407 | return 1; |
32f84528 | 2408 | else |
290408d4 NH |
2409 | return 0; |
2410 | } | |
2411 | ||
fd6a03ed NH |
2412 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) |
2413 | { | |
2414 | swp_entry_t swp; | |
2415 | ||
2416 | if (huge_pte_none(pte) || pte_present(pte)) | |
2417 | return 0; | |
2418 | swp = pte_to_swp_entry(pte); | |
32f84528 | 2419 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) |
fd6a03ed | 2420 | return 1; |
32f84528 | 2421 | else |
fd6a03ed NH |
2422 | return 0; |
2423 | } | |
2424 | ||
24669e58 AK |
2425 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
2426 | unsigned long start, unsigned long end, | |
2427 | struct page *ref_page) | |
63551ae0 | 2428 | { |
24669e58 | 2429 | int force_flush = 0; |
63551ae0 DG |
2430 | struct mm_struct *mm = vma->vm_mm; |
2431 | unsigned long address; | |
c7546f8f | 2432 | pte_t *ptep; |
63551ae0 DG |
2433 | pte_t pte; |
2434 | struct page *page; | |
a5516438 AK |
2435 | struct hstate *h = hstate_vma(vma); |
2436 | unsigned long sz = huge_page_size(h); | |
2ec74c3e SG |
2437 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
2438 | const unsigned long mmun_end = end; /* For mmu_notifiers */ | |
a5516438 | 2439 | |
63551ae0 | 2440 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
2441 | BUG_ON(start & ~huge_page_mask(h)); |
2442 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 2443 | |
24669e58 | 2444 | tlb_start_vma(tlb, vma); |
2ec74c3e | 2445 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
24669e58 | 2446 | again: |
508034a3 | 2447 | spin_lock(&mm->page_table_lock); |
a5516438 | 2448 | for (address = start; address < end; address += sz) { |
c7546f8f | 2449 | ptep = huge_pte_offset(mm, address); |
4c887265 | 2450 | if (!ptep) |
c7546f8f DG |
2451 | continue; |
2452 | ||
39dde65c CK |
2453 | if (huge_pmd_unshare(mm, &address, ptep)) |
2454 | continue; | |
2455 | ||
6629326b HD |
2456 | pte = huge_ptep_get(ptep); |
2457 | if (huge_pte_none(pte)) | |
2458 | continue; | |
2459 | ||
2460 | /* | |
2461 | * HWPoisoned hugepage is already unmapped and dropped reference | |
2462 | */ | |
8c4894c6 | 2463 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { |
106c992a | 2464 | huge_pte_clear(mm, address, ptep); |
6629326b | 2465 | continue; |
8c4894c6 | 2466 | } |
6629326b HD |
2467 | |
2468 | page = pte_page(pte); | |
04f2cbe3 MG |
2469 | /* |
2470 | * If a reference page is supplied, it is because a specific | |
2471 | * page is being unmapped, not a range. Ensure the page we | |
2472 | * are about to unmap is the actual page of interest. | |
2473 | */ | |
2474 | if (ref_page) { | |
04f2cbe3 MG |
2475 | if (page != ref_page) |
2476 | continue; | |
2477 | ||
2478 | /* | |
2479 | * Mark the VMA as having unmapped its page so that | |
2480 | * future faults in this VMA will fail rather than | |
2481 | * looking like data was lost | |
2482 | */ | |
2483 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
2484 | } | |
2485 | ||
c7546f8f | 2486 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
24669e58 | 2487 | tlb_remove_tlb_entry(tlb, ptep, address); |
106c992a | 2488 | if (huge_pte_dirty(pte)) |
6649a386 | 2489 | set_page_dirty(page); |
9e81130b | 2490 | |
24669e58 AK |
2491 | page_remove_rmap(page); |
2492 | force_flush = !__tlb_remove_page(tlb, page); | |
2493 | if (force_flush) | |
2494 | break; | |
9e81130b HD |
2495 | /* Bail out after unmapping reference page if supplied */ |
2496 | if (ref_page) | |
2497 | break; | |
63551ae0 | 2498 | } |
cd2934a3 | 2499 | spin_unlock(&mm->page_table_lock); |
24669e58 AK |
2500 | /* |
2501 | * mmu_gather ran out of room to batch pages, we break out of | |
2502 | * the PTE lock to avoid doing the potential expensive TLB invalidate | |
2503 | * and page-free while holding it. | |
2504 | */ | |
2505 | if (force_flush) { | |
2506 | force_flush = 0; | |
2507 | tlb_flush_mmu(tlb); | |
2508 | if (address < end && !ref_page) | |
2509 | goto again; | |
fe1668ae | 2510 | } |
2ec74c3e | 2511 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
24669e58 | 2512 | tlb_end_vma(tlb, vma); |
1da177e4 | 2513 | } |
63551ae0 | 2514 | |
d833352a MG |
2515 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
2516 | struct vm_area_struct *vma, unsigned long start, | |
2517 | unsigned long end, struct page *ref_page) | |
2518 | { | |
2519 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
2520 | ||
2521 | /* | |
2522 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
2523 | * test will fail on a vma being torn down, and not grab a page table | |
2524 | * on its way out. We're lucky that the flag has such an appropriate | |
2525 | * name, and can in fact be safely cleared here. We could clear it | |
2526 | * before the __unmap_hugepage_range above, but all that's necessary | |
2527 | * is to clear it before releasing the i_mmap_mutex. This works | |
2528 | * because in the context this is called, the VMA is about to be | |
2529 | * destroyed and the i_mmap_mutex is held. | |
2530 | */ | |
2531 | vma->vm_flags &= ~VM_MAYSHARE; | |
2532 | } | |
2533 | ||
502717f4 | 2534 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 2535 | unsigned long end, struct page *ref_page) |
502717f4 | 2536 | { |
24669e58 AK |
2537 | struct mm_struct *mm; |
2538 | struct mmu_gather tlb; | |
2539 | ||
2540 | mm = vma->vm_mm; | |
2541 | ||
2b047252 | 2542 | tlb_gather_mmu(&tlb, mm, start, end); |
24669e58 AK |
2543 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
2544 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 CK |
2545 | } |
2546 | ||
04f2cbe3 MG |
2547 | /* |
2548 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
2549 | * mappping it owns the reserve page for. The intention is to unmap the page | |
2550 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
2551 | * same region. | |
2552 | */ | |
2a4b3ded HH |
2553 | static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
2554 | struct page *page, unsigned long address) | |
04f2cbe3 | 2555 | { |
7526674d | 2556 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
2557 | struct vm_area_struct *iter_vma; |
2558 | struct address_space *mapping; | |
04f2cbe3 MG |
2559 | pgoff_t pgoff; |
2560 | ||
2561 | /* | |
2562 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
2563 | * from page cache lookup which is in HPAGE_SIZE units. | |
2564 | */ | |
7526674d | 2565 | address = address & huge_page_mask(h); |
36e4f20a MH |
2566 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
2567 | vma->vm_pgoff; | |
496ad9aa | 2568 | mapping = file_inode(vma->vm_file)->i_mapping; |
04f2cbe3 | 2569 | |
4eb2b1dc MG |
2570 | /* |
2571 | * Take the mapping lock for the duration of the table walk. As | |
2572 | * this mapping should be shared between all the VMAs, | |
2573 | * __unmap_hugepage_range() is called as the lock is already held | |
2574 | */ | |
3d48ae45 | 2575 | mutex_lock(&mapping->i_mmap_mutex); |
6b2dbba8 | 2576 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
2577 | /* Do not unmap the current VMA */ |
2578 | if (iter_vma == vma) | |
2579 | continue; | |
2580 | ||
2581 | /* | |
2582 | * Unmap the page from other VMAs without their own reserves. | |
2583 | * They get marked to be SIGKILLed if they fault in these | |
2584 | * areas. This is because a future no-page fault on this VMA | |
2585 | * could insert a zeroed page instead of the data existing | |
2586 | * from the time of fork. This would look like data corruption | |
2587 | */ | |
2588 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
2589 | unmap_hugepage_range(iter_vma, address, |
2590 | address + huge_page_size(h), page); | |
04f2cbe3 | 2591 | } |
3d48ae45 | 2592 | mutex_unlock(&mapping->i_mmap_mutex); |
04f2cbe3 MG |
2593 | |
2594 | return 1; | |
2595 | } | |
2596 | ||
0fe6e20b NH |
2597 | /* |
2598 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
2599 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
2600 | * cannot race with other handlers or page migration. | |
2601 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 2602 | */ |
1e8f889b | 2603 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
04f2cbe3 MG |
2604 | unsigned long address, pte_t *ptep, pte_t pte, |
2605 | struct page *pagecache_page) | |
1e8f889b | 2606 | { |
a5516438 | 2607 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 2608 | struct page *old_page, *new_page; |
04f2cbe3 | 2609 | int outside_reserve = 0; |
2ec74c3e SG |
2610 | unsigned long mmun_start; /* For mmu_notifiers */ |
2611 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1e8f889b DG |
2612 | |
2613 | old_page = pte_page(pte); | |
2614 | ||
04f2cbe3 | 2615 | retry_avoidcopy: |
1e8f889b DG |
2616 | /* If no-one else is actually using this page, avoid the copy |
2617 | * and just make the page writable */ | |
37a2140d JK |
2618 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
2619 | page_move_anon_rmap(old_page, vma, address); | |
1e8f889b | 2620 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 2621 | return 0; |
1e8f889b DG |
2622 | } |
2623 | ||
04f2cbe3 MG |
2624 | /* |
2625 | * If the process that created a MAP_PRIVATE mapping is about to | |
2626 | * perform a COW due to a shared page count, attempt to satisfy | |
2627 | * the allocation without using the existing reserves. The pagecache | |
2628 | * page is used to determine if the reserve at this address was | |
2629 | * consumed or not. If reserves were used, a partial faulted mapping | |
2630 | * at the time of fork() could consume its reserves on COW instead | |
2631 | * of the full address range. | |
2632 | */ | |
5944d011 | 2633 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
2634 | old_page != pagecache_page) |
2635 | outside_reserve = 1; | |
2636 | ||
1e8f889b | 2637 | page_cache_get(old_page); |
b76c8cfb LW |
2638 | |
2639 | /* Drop page_table_lock as buddy allocator may be called */ | |
2640 | spin_unlock(&mm->page_table_lock); | |
04f2cbe3 | 2641 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 2642 | |
2fc39cec | 2643 | if (IS_ERR(new_page)) { |
76dcee75 | 2644 | long err = PTR_ERR(new_page); |
1e8f889b | 2645 | page_cache_release(old_page); |
04f2cbe3 MG |
2646 | |
2647 | /* | |
2648 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
2649 | * it is due to references held by a child and an insufficient | |
2650 | * huge page pool. To guarantee the original mappers | |
2651 | * reliability, unmap the page from child processes. The child | |
2652 | * may get SIGKILLed if it later faults. | |
2653 | */ | |
2654 | if (outside_reserve) { | |
2655 | BUG_ON(huge_pte_none(pte)); | |
2656 | if (unmap_ref_private(mm, vma, old_page, address)) { | |
04f2cbe3 | 2657 | BUG_ON(huge_pte_none(pte)); |
b76c8cfb | 2658 | spin_lock(&mm->page_table_lock); |
a734bcc8 HD |
2659 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
2660 | if (likely(pte_same(huge_ptep_get(ptep), pte))) | |
2661 | goto retry_avoidcopy; | |
2662 | /* | |
2663 | * race occurs while re-acquiring page_table_lock, and | |
2664 | * our job is done. | |
2665 | */ | |
2666 | return 0; | |
04f2cbe3 MG |
2667 | } |
2668 | WARN_ON_ONCE(1); | |
2669 | } | |
2670 | ||
b76c8cfb LW |
2671 | /* Caller expects lock to be held */ |
2672 | spin_lock(&mm->page_table_lock); | |
76dcee75 AK |
2673 | if (err == -ENOMEM) |
2674 | return VM_FAULT_OOM; | |
2675 | else | |
2676 | return VM_FAULT_SIGBUS; | |
1e8f889b DG |
2677 | } |
2678 | ||
0fe6e20b NH |
2679 | /* |
2680 | * When the original hugepage is shared one, it does not have | |
2681 | * anon_vma prepared. | |
2682 | */ | |
44e2aa93 | 2683 | if (unlikely(anon_vma_prepare(vma))) { |
ea4039a3 HD |
2684 | page_cache_release(new_page); |
2685 | page_cache_release(old_page); | |
44e2aa93 DN |
2686 | /* Caller expects lock to be held */ |
2687 | spin_lock(&mm->page_table_lock); | |
0fe6e20b | 2688 | return VM_FAULT_OOM; |
44e2aa93 | 2689 | } |
0fe6e20b | 2690 | |
47ad8475 AA |
2691 | copy_user_huge_page(new_page, old_page, address, vma, |
2692 | pages_per_huge_page(h)); | |
0ed361de | 2693 | __SetPageUptodate(new_page); |
1e8f889b | 2694 | |
2ec74c3e SG |
2695 | mmun_start = address & huge_page_mask(h); |
2696 | mmun_end = mmun_start + huge_page_size(h); | |
2697 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
b76c8cfb LW |
2698 | /* |
2699 | * Retake the page_table_lock to check for racing updates | |
2700 | * before the page tables are altered | |
2701 | */ | |
2702 | spin_lock(&mm->page_table_lock); | |
a5516438 | 2703 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
7f2e9525 | 2704 | if (likely(pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
2705 | ClearPagePrivate(new_page); |
2706 | ||
1e8f889b | 2707 | /* Break COW */ |
8fe627ec | 2708 | huge_ptep_clear_flush(vma, address, ptep); |
1e8f889b DG |
2709 | set_huge_pte_at(mm, address, ptep, |
2710 | make_huge_pte(vma, new_page, 1)); | |
0fe6e20b | 2711 | page_remove_rmap(old_page); |
cd67f0d2 | 2712 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
2713 | /* Make the old page be freed below */ |
2714 | new_page = old_page; | |
2715 | } | |
2ec74c3e SG |
2716 | spin_unlock(&mm->page_table_lock); |
2717 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); | |
1e8f889b DG |
2718 | page_cache_release(new_page); |
2719 | page_cache_release(old_page); | |
8312034f JK |
2720 | |
2721 | /* Caller expects lock to be held */ | |
2722 | spin_lock(&mm->page_table_lock); | |
83c54070 | 2723 | return 0; |
1e8f889b DG |
2724 | } |
2725 | ||
04f2cbe3 | 2726 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
2727 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
2728 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
2729 | { |
2730 | struct address_space *mapping; | |
e7c4b0bf | 2731 | pgoff_t idx; |
04f2cbe3 MG |
2732 | |
2733 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 2734 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
2735 | |
2736 | return find_lock_page(mapping, idx); | |
2737 | } | |
2738 | ||
3ae77f43 HD |
2739 | /* |
2740 | * Return whether there is a pagecache page to back given address within VMA. | |
2741 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
2742 | */ | |
2743 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
2744 | struct vm_area_struct *vma, unsigned long address) |
2745 | { | |
2746 | struct address_space *mapping; | |
2747 | pgoff_t idx; | |
2748 | struct page *page; | |
2749 | ||
2750 | mapping = vma->vm_file->f_mapping; | |
2751 | idx = vma_hugecache_offset(h, vma, address); | |
2752 | ||
2753 | page = find_get_page(mapping, idx); | |
2754 | if (page) | |
2755 | put_page(page); | |
2756 | return page != NULL; | |
2757 | } | |
2758 | ||
a1ed3dda | 2759 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 2760 | unsigned long address, pte_t *ptep, unsigned int flags) |
ac9b9c66 | 2761 | { |
a5516438 | 2762 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 2763 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 2764 | int anon_rmap = 0; |
e7c4b0bf | 2765 | pgoff_t idx; |
4c887265 | 2766 | unsigned long size; |
4c887265 AL |
2767 | struct page *page; |
2768 | struct address_space *mapping; | |
1e8f889b | 2769 | pte_t new_pte; |
4c887265 | 2770 | |
04f2cbe3 MG |
2771 | /* |
2772 | * Currently, we are forced to kill the process in the event the | |
2773 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 2774 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
2775 | */ |
2776 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
ffb22af5 AM |
2777 | pr_warning("PID %d killed due to inadequate hugepage pool\n", |
2778 | current->pid); | |
04f2cbe3 MG |
2779 | return ret; |
2780 | } | |
2781 | ||
4c887265 | 2782 | mapping = vma->vm_file->f_mapping; |
a5516438 | 2783 | idx = vma_hugecache_offset(h, vma, address); |
4c887265 AL |
2784 | |
2785 | /* | |
2786 | * Use page lock to guard against racing truncation | |
2787 | * before we get page_table_lock. | |
2788 | */ | |
6bda666a CL |
2789 | retry: |
2790 | page = find_lock_page(mapping, idx); | |
2791 | if (!page) { | |
a5516438 | 2792 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
2793 | if (idx >= size) |
2794 | goto out; | |
04f2cbe3 | 2795 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 2796 | if (IS_ERR(page)) { |
76dcee75 AK |
2797 | ret = PTR_ERR(page); |
2798 | if (ret == -ENOMEM) | |
2799 | ret = VM_FAULT_OOM; | |
2800 | else | |
2801 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
2802 | goto out; |
2803 | } | |
47ad8475 | 2804 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 2805 | __SetPageUptodate(page); |
ac9b9c66 | 2806 | |
f83a275d | 2807 | if (vma->vm_flags & VM_MAYSHARE) { |
6bda666a | 2808 | int err; |
45c682a6 | 2809 | struct inode *inode = mapping->host; |
6bda666a CL |
2810 | |
2811 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
2812 | if (err) { | |
2813 | put_page(page); | |
6bda666a CL |
2814 | if (err == -EEXIST) |
2815 | goto retry; | |
2816 | goto out; | |
2817 | } | |
07443a85 | 2818 | ClearPagePrivate(page); |
45c682a6 KC |
2819 | |
2820 | spin_lock(&inode->i_lock); | |
a5516438 | 2821 | inode->i_blocks += blocks_per_huge_page(h); |
45c682a6 | 2822 | spin_unlock(&inode->i_lock); |
23be7468 | 2823 | } else { |
6bda666a | 2824 | lock_page(page); |
0fe6e20b NH |
2825 | if (unlikely(anon_vma_prepare(vma))) { |
2826 | ret = VM_FAULT_OOM; | |
2827 | goto backout_unlocked; | |
2828 | } | |
409eb8c2 | 2829 | anon_rmap = 1; |
23be7468 | 2830 | } |
0fe6e20b | 2831 | } else { |
998b4382 NH |
2832 | /* |
2833 | * If memory error occurs between mmap() and fault, some process | |
2834 | * don't have hwpoisoned swap entry for errored virtual address. | |
2835 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
2836 | */ | |
2837 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 2838 | ret = VM_FAULT_HWPOISON | |
972dc4de | 2839 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
2840 | goto backout_unlocked; |
2841 | } | |
6bda666a | 2842 | } |
1e8f889b | 2843 | |
57303d80 AW |
2844 | /* |
2845 | * If we are going to COW a private mapping later, we examine the | |
2846 | * pending reservations for this page now. This will ensure that | |
2847 | * any allocations necessary to record that reservation occur outside | |
2848 | * the spinlock. | |
2849 | */ | |
788c7df4 | 2850 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) |
2b26736c AW |
2851 | if (vma_needs_reservation(h, vma, address) < 0) { |
2852 | ret = VM_FAULT_OOM; | |
2853 | goto backout_unlocked; | |
2854 | } | |
57303d80 | 2855 | |
ac9b9c66 | 2856 | spin_lock(&mm->page_table_lock); |
a5516438 | 2857 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
2858 | if (idx >= size) |
2859 | goto backout; | |
2860 | ||
83c54070 | 2861 | ret = 0; |
7f2e9525 | 2862 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
2863 | goto backout; |
2864 | ||
07443a85 JK |
2865 | if (anon_rmap) { |
2866 | ClearPagePrivate(page); | |
409eb8c2 | 2867 | hugepage_add_new_anon_rmap(page, vma, address); |
07443a85 | 2868 | } |
409eb8c2 HD |
2869 | else |
2870 | page_dup_rmap(page); | |
1e8f889b DG |
2871 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
2872 | && (vma->vm_flags & VM_SHARED))); | |
2873 | set_huge_pte_at(mm, address, ptep, new_pte); | |
2874 | ||
788c7df4 | 2875 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 2876 | /* Optimization, do the COW without a second fault */ |
04f2cbe3 | 2877 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page); |
1e8f889b DG |
2878 | } |
2879 | ||
ac9b9c66 | 2880 | spin_unlock(&mm->page_table_lock); |
4c887265 AL |
2881 | unlock_page(page); |
2882 | out: | |
ac9b9c66 | 2883 | return ret; |
4c887265 AL |
2884 | |
2885 | backout: | |
2886 | spin_unlock(&mm->page_table_lock); | |
2b26736c | 2887 | backout_unlocked: |
4c887265 AL |
2888 | unlock_page(page); |
2889 | put_page(page); | |
2890 | goto out; | |
ac9b9c66 HD |
2891 | } |
2892 | ||
86e5216f | 2893 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 2894 | unsigned long address, unsigned int flags) |
86e5216f AL |
2895 | { |
2896 | pte_t *ptep; | |
2897 | pte_t entry; | |
1e8f889b | 2898 | int ret; |
0fe6e20b | 2899 | struct page *page = NULL; |
57303d80 | 2900 | struct page *pagecache_page = NULL; |
3935baa9 | 2901 | static DEFINE_MUTEX(hugetlb_instantiation_mutex); |
a5516438 | 2902 | struct hstate *h = hstate_vma(vma); |
86e5216f | 2903 | |
1e16a539 KH |
2904 | address &= huge_page_mask(h); |
2905 | ||
fd6a03ed NH |
2906 | ptep = huge_pte_offset(mm, address); |
2907 | if (ptep) { | |
2908 | entry = huge_ptep_get(ptep); | |
290408d4 | 2909 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
30dad309 | 2910 | migration_entry_wait_huge(mm, ptep); |
290408d4 NH |
2911 | return 0; |
2912 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 2913 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 2914 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
2915 | } |
2916 | ||
a5516438 | 2917 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); |
86e5216f AL |
2918 | if (!ptep) |
2919 | return VM_FAULT_OOM; | |
2920 | ||
3935baa9 DG |
2921 | /* |
2922 | * Serialize hugepage allocation and instantiation, so that we don't | |
2923 | * get spurious allocation failures if two CPUs race to instantiate | |
2924 | * the same page in the page cache. | |
2925 | */ | |
2926 | mutex_lock(&hugetlb_instantiation_mutex); | |
7f2e9525 GS |
2927 | entry = huge_ptep_get(ptep); |
2928 | if (huge_pte_none(entry)) { | |
788c7df4 | 2929 | ret = hugetlb_no_page(mm, vma, address, ptep, flags); |
b4d1d99f | 2930 | goto out_mutex; |
3935baa9 | 2931 | } |
86e5216f | 2932 | |
83c54070 | 2933 | ret = 0; |
1e8f889b | 2934 | |
57303d80 AW |
2935 | /* |
2936 | * If we are going to COW the mapping later, we examine the pending | |
2937 | * reservations for this page now. This will ensure that any | |
2938 | * allocations necessary to record that reservation occur outside the | |
2939 | * spinlock. For private mappings, we also lookup the pagecache | |
2940 | * page now as it is used to determine if a reservation has been | |
2941 | * consumed. | |
2942 | */ | |
106c992a | 2943 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
2b26736c AW |
2944 | if (vma_needs_reservation(h, vma, address) < 0) { |
2945 | ret = VM_FAULT_OOM; | |
b4d1d99f | 2946 | goto out_mutex; |
2b26736c | 2947 | } |
57303d80 | 2948 | |
f83a275d | 2949 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
2950 | pagecache_page = hugetlbfs_pagecache_page(h, |
2951 | vma, address); | |
2952 | } | |
2953 | ||
56c9cfb1 NH |
2954 | /* |
2955 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
2956 | * pagecache_page, so here we need take the former one | |
2957 | * when page != pagecache_page or !pagecache_page. | |
2958 | * Note that locking order is always pagecache_page -> page, | |
2959 | * so no worry about deadlock. | |
2960 | */ | |
2961 | page = pte_page(entry); | |
66aebce7 | 2962 | get_page(page); |
56c9cfb1 | 2963 | if (page != pagecache_page) |
0fe6e20b | 2964 | lock_page(page); |
0fe6e20b | 2965 | |
1e8f889b DG |
2966 | spin_lock(&mm->page_table_lock); |
2967 | /* Check for a racing update before calling hugetlb_cow */ | |
b4d1d99f DG |
2968 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) |
2969 | goto out_page_table_lock; | |
2970 | ||
2971 | ||
788c7df4 | 2972 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 2973 | if (!huge_pte_write(entry)) { |
57303d80 AW |
2974 | ret = hugetlb_cow(mm, vma, address, ptep, entry, |
2975 | pagecache_page); | |
b4d1d99f DG |
2976 | goto out_page_table_lock; |
2977 | } | |
106c992a | 2978 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
2979 | } |
2980 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
2981 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
2982 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 2983 | update_mmu_cache(vma, address, ptep); |
b4d1d99f DG |
2984 | |
2985 | out_page_table_lock: | |
1e8f889b | 2986 | spin_unlock(&mm->page_table_lock); |
57303d80 AW |
2987 | |
2988 | if (pagecache_page) { | |
2989 | unlock_page(pagecache_page); | |
2990 | put_page(pagecache_page); | |
2991 | } | |
1f64d69c DN |
2992 | if (page != pagecache_page) |
2993 | unlock_page(page); | |
66aebce7 | 2994 | put_page(page); |
57303d80 | 2995 | |
b4d1d99f | 2996 | out_mutex: |
3935baa9 | 2997 | mutex_unlock(&hugetlb_instantiation_mutex); |
1e8f889b DG |
2998 | |
2999 | return ret; | |
86e5216f AL |
3000 | } |
3001 | ||
28a35716 ML |
3002 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
3003 | struct page **pages, struct vm_area_struct **vmas, | |
3004 | unsigned long *position, unsigned long *nr_pages, | |
3005 | long i, unsigned int flags) | |
63551ae0 | 3006 | { |
d5d4b0aa CK |
3007 | unsigned long pfn_offset; |
3008 | unsigned long vaddr = *position; | |
28a35716 | 3009 | unsigned long remainder = *nr_pages; |
a5516438 | 3010 | struct hstate *h = hstate_vma(vma); |
63551ae0 | 3011 | |
1c59827d | 3012 | spin_lock(&mm->page_table_lock); |
63551ae0 | 3013 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 3014 | pte_t *pte; |
2a15efc9 | 3015 | int absent; |
4c887265 | 3016 | struct page *page; |
63551ae0 | 3017 | |
4c887265 AL |
3018 | /* |
3019 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 3020 | * each hugepage. We have to make sure we get the |
4c887265 AL |
3021 | * first, for the page indexing below to work. |
3022 | */ | |
a5516438 | 3023 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); |
2a15efc9 HD |
3024 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
3025 | ||
3026 | /* | |
3027 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
3028 | * an error where there's an empty slot with no huge pagecache |
3029 | * to back it. This way, we avoid allocating a hugepage, and | |
3030 | * the sparse dumpfile avoids allocating disk blocks, but its | |
3031 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 3032 | */ |
3ae77f43 HD |
3033 | if (absent && (flags & FOLL_DUMP) && |
3034 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
2a15efc9 HD |
3035 | remainder = 0; |
3036 | break; | |
3037 | } | |
63551ae0 | 3038 | |
9cc3a5bd NH |
3039 | /* |
3040 | * We need call hugetlb_fault for both hugepages under migration | |
3041 | * (in which case hugetlb_fault waits for the migration,) and | |
3042 | * hwpoisoned hugepages (in which case we need to prevent the | |
3043 | * caller from accessing to them.) In order to do this, we use | |
3044 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
3045 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
3046 | * both cases, and because we can't follow correct pages | |
3047 | * directly from any kind of swap entries. | |
3048 | */ | |
3049 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
3050 | ((flags & FOLL_WRITE) && |
3051 | !huge_pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 3052 | int ret; |
63551ae0 | 3053 | |
4c887265 | 3054 | spin_unlock(&mm->page_table_lock); |
2a15efc9 HD |
3055 | ret = hugetlb_fault(mm, vma, vaddr, |
3056 | (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); | |
4c887265 | 3057 | spin_lock(&mm->page_table_lock); |
a89182c7 | 3058 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 3059 | continue; |
63551ae0 | 3060 | |
4c887265 | 3061 | remainder = 0; |
4c887265 AL |
3062 | break; |
3063 | } | |
3064 | ||
a5516438 | 3065 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 3066 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 3067 | same_page: |
d6692183 | 3068 | if (pages) { |
2a15efc9 | 3069 | pages[i] = mem_map_offset(page, pfn_offset); |
4b2e38ad | 3070 | get_page(pages[i]); |
d6692183 | 3071 | } |
63551ae0 DG |
3072 | |
3073 | if (vmas) | |
3074 | vmas[i] = vma; | |
3075 | ||
3076 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 3077 | ++pfn_offset; |
63551ae0 DG |
3078 | --remainder; |
3079 | ++i; | |
d5d4b0aa | 3080 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 3081 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa CK |
3082 | /* |
3083 | * We use pfn_offset to avoid touching the pageframes | |
3084 | * of this compound page. | |
3085 | */ | |
3086 | goto same_page; | |
3087 | } | |
63551ae0 | 3088 | } |
1c59827d | 3089 | spin_unlock(&mm->page_table_lock); |
28a35716 | 3090 | *nr_pages = remainder; |
63551ae0 DG |
3091 | *position = vaddr; |
3092 | ||
2a15efc9 | 3093 | return i ? i : -EFAULT; |
63551ae0 | 3094 | } |
8f860591 | 3095 | |
7da4d641 | 3096 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
3097 | unsigned long address, unsigned long end, pgprot_t newprot) |
3098 | { | |
3099 | struct mm_struct *mm = vma->vm_mm; | |
3100 | unsigned long start = address; | |
3101 | pte_t *ptep; | |
3102 | pte_t pte; | |
a5516438 | 3103 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 3104 | unsigned long pages = 0; |
8f860591 ZY |
3105 | |
3106 | BUG_ON(address >= end); | |
3107 | flush_cache_range(vma, address, end); | |
3108 | ||
3d48ae45 | 3109 | mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex); |
8f860591 | 3110 | spin_lock(&mm->page_table_lock); |
a5516438 | 3111 | for (; address < end; address += huge_page_size(h)) { |
8f860591 ZY |
3112 | ptep = huge_pte_offset(mm, address); |
3113 | if (!ptep) | |
3114 | continue; | |
7da4d641 PZ |
3115 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3116 | pages++; | |
39dde65c | 3117 | continue; |
7da4d641 | 3118 | } |
7f2e9525 | 3119 | if (!huge_pte_none(huge_ptep_get(ptep))) { |
8f860591 | 3120 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
106c992a | 3121 | pte = pte_mkhuge(huge_pte_modify(pte, newprot)); |
be7517d6 | 3122 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
8f860591 | 3123 | set_huge_pte_at(mm, address, ptep, pte); |
7da4d641 | 3124 | pages++; |
8f860591 ZY |
3125 | } |
3126 | } | |
3127 | spin_unlock(&mm->page_table_lock); | |
d833352a MG |
3128 | /* |
3129 | * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare | |
3130 | * may have cleared our pud entry and done put_page on the page table: | |
3131 | * once we release i_mmap_mutex, another task can do the final put_page | |
3132 | * and that page table be reused and filled with junk. | |
3133 | */ | |
8f860591 | 3134 | flush_tlb_range(vma, start, end); |
d833352a | 3135 | mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex); |
7da4d641 PZ |
3136 | |
3137 | return pages << h->order; | |
8f860591 ZY |
3138 | } |
3139 | ||
a1e78772 MG |
3140 | int hugetlb_reserve_pages(struct inode *inode, |
3141 | long from, long to, | |
5a6fe125 | 3142 | struct vm_area_struct *vma, |
ca16d140 | 3143 | vm_flags_t vm_flags) |
e4e574b7 | 3144 | { |
17c9d12e | 3145 | long ret, chg; |
a5516438 | 3146 | struct hstate *h = hstate_inode(inode); |
90481622 | 3147 | struct hugepage_subpool *spool = subpool_inode(inode); |
e4e574b7 | 3148 | |
17c9d12e MG |
3149 | /* |
3150 | * Only apply hugepage reservation if asked. At fault time, an | |
3151 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 3152 | * without using reserves |
17c9d12e | 3153 | */ |
ca16d140 | 3154 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
3155 | return 0; |
3156 | ||
a1e78772 MG |
3157 | /* |
3158 | * Shared mappings base their reservation on the number of pages that | |
3159 | * are already allocated on behalf of the file. Private mappings need | |
3160 | * to reserve the full area even if read-only as mprotect() may be | |
3161 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
3162 | */ | |
f83a275d | 3163 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
a1e78772 | 3164 | chg = region_chg(&inode->i_mapping->private_list, from, to); |
17c9d12e MG |
3165 | else { |
3166 | struct resv_map *resv_map = resv_map_alloc(); | |
3167 | if (!resv_map) | |
3168 | return -ENOMEM; | |
3169 | ||
a1e78772 | 3170 | chg = to - from; |
84afd99b | 3171 | |
17c9d12e MG |
3172 | set_vma_resv_map(vma, resv_map); |
3173 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
3174 | } | |
3175 | ||
c50ac050 DH |
3176 | if (chg < 0) { |
3177 | ret = chg; | |
3178 | goto out_err; | |
3179 | } | |
8a630112 | 3180 | |
90481622 | 3181 | /* There must be enough pages in the subpool for the mapping */ |
c50ac050 DH |
3182 | if (hugepage_subpool_get_pages(spool, chg)) { |
3183 | ret = -ENOSPC; | |
3184 | goto out_err; | |
3185 | } | |
5a6fe125 MG |
3186 | |
3187 | /* | |
17c9d12e | 3188 | * Check enough hugepages are available for the reservation. |
90481622 | 3189 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 3190 | */ |
a5516438 | 3191 | ret = hugetlb_acct_memory(h, chg); |
68842c9b | 3192 | if (ret < 0) { |
90481622 | 3193 | hugepage_subpool_put_pages(spool, chg); |
c50ac050 | 3194 | goto out_err; |
68842c9b | 3195 | } |
17c9d12e MG |
3196 | |
3197 | /* | |
3198 | * Account for the reservations made. Shared mappings record regions | |
3199 | * that have reservations as they are shared by multiple VMAs. | |
3200 | * When the last VMA disappears, the region map says how much | |
3201 | * the reservation was and the page cache tells how much of | |
3202 | * the reservation was consumed. Private mappings are per-VMA and | |
3203 | * only the consumed reservations are tracked. When the VMA | |
3204 | * disappears, the original reservation is the VMA size and the | |
3205 | * consumed reservations are stored in the map. Hence, nothing | |
3206 | * else has to be done for private mappings here | |
3207 | */ | |
f83a275d | 3208 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
a1e78772 | 3209 | region_add(&inode->i_mapping->private_list, from, to); |
a43a8c39 | 3210 | return 0; |
c50ac050 | 3211 | out_err: |
4523e145 DH |
3212 | if (vma) |
3213 | resv_map_put(vma); | |
c50ac050 | 3214 | return ret; |
a43a8c39 CK |
3215 | } |
3216 | ||
3217 | void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) | |
3218 | { | |
a5516438 | 3219 | struct hstate *h = hstate_inode(inode); |
a43a8c39 | 3220 | long chg = region_truncate(&inode->i_mapping->private_list, offset); |
90481622 | 3221 | struct hugepage_subpool *spool = subpool_inode(inode); |
45c682a6 KC |
3222 | |
3223 | spin_lock(&inode->i_lock); | |
e4c6f8be | 3224 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
3225 | spin_unlock(&inode->i_lock); |
3226 | ||
90481622 | 3227 | hugepage_subpool_put_pages(spool, (chg - freed)); |
a5516438 | 3228 | hugetlb_acct_memory(h, -(chg - freed)); |
a43a8c39 | 3229 | } |
93f70f90 | 3230 | |
3212b535 SC |
3231 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
3232 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
3233 | struct vm_area_struct *vma, | |
3234 | unsigned long addr, pgoff_t idx) | |
3235 | { | |
3236 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
3237 | svma->vm_start; | |
3238 | unsigned long sbase = saddr & PUD_MASK; | |
3239 | unsigned long s_end = sbase + PUD_SIZE; | |
3240 | ||
3241 | /* Allow segments to share if only one is marked locked */ | |
3242 | unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED; | |
3243 | unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED; | |
3244 | ||
3245 | /* | |
3246 | * match the virtual addresses, permission and the alignment of the | |
3247 | * page table page. | |
3248 | */ | |
3249 | if (pmd_index(addr) != pmd_index(saddr) || | |
3250 | vm_flags != svm_flags || | |
3251 | sbase < svma->vm_start || svma->vm_end < s_end) | |
3252 | return 0; | |
3253 | ||
3254 | return saddr; | |
3255 | } | |
3256 | ||
3257 | static int vma_shareable(struct vm_area_struct *vma, unsigned long addr) | |
3258 | { | |
3259 | unsigned long base = addr & PUD_MASK; | |
3260 | unsigned long end = base + PUD_SIZE; | |
3261 | ||
3262 | /* | |
3263 | * check on proper vm_flags and page table alignment | |
3264 | */ | |
3265 | if (vma->vm_flags & VM_MAYSHARE && | |
3266 | vma->vm_start <= base && end <= vma->vm_end) | |
3267 | return 1; | |
3268 | return 0; | |
3269 | } | |
3270 | ||
3271 | /* | |
3272 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
3273 | * and returns the corresponding pte. While this is not necessary for the | |
3274 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
3275 | * code much cleaner. pmd allocation is essential for the shared case because | |
3276 | * pud has to be populated inside the same i_mmap_mutex section - otherwise | |
3277 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a | |
3278 | * bad pmd for sharing. | |
3279 | */ | |
3280 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
3281 | { | |
3282 | struct vm_area_struct *vma = find_vma(mm, addr); | |
3283 | struct address_space *mapping = vma->vm_file->f_mapping; | |
3284 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
3285 | vma->vm_pgoff; | |
3286 | struct vm_area_struct *svma; | |
3287 | unsigned long saddr; | |
3288 | pte_t *spte = NULL; | |
3289 | pte_t *pte; | |
3290 | ||
3291 | if (!vma_shareable(vma, addr)) | |
3292 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
3293 | ||
3294 | mutex_lock(&mapping->i_mmap_mutex); | |
3295 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { | |
3296 | if (svma == vma) | |
3297 | continue; | |
3298 | ||
3299 | saddr = page_table_shareable(svma, vma, addr, idx); | |
3300 | if (saddr) { | |
3301 | spte = huge_pte_offset(svma->vm_mm, saddr); | |
3302 | if (spte) { | |
3303 | get_page(virt_to_page(spte)); | |
3304 | break; | |
3305 | } | |
3306 | } | |
3307 | } | |
3308 | ||
3309 | if (!spte) | |
3310 | goto out; | |
3311 | ||
3312 | spin_lock(&mm->page_table_lock); | |
3313 | if (pud_none(*pud)) | |
3314 | pud_populate(mm, pud, | |
3315 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
3316 | else | |
3317 | put_page(virt_to_page(spte)); | |
3318 | spin_unlock(&mm->page_table_lock); | |
3319 | out: | |
3320 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3321 | mutex_unlock(&mapping->i_mmap_mutex); | |
3322 | return pte; | |
3323 | } | |
3324 | ||
3325 | /* | |
3326 | * unmap huge page backed by shared pte. | |
3327 | * | |
3328 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
3329 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
3330 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
3331 | * | |
3332 | * called with vma->vm_mm->page_table_lock held. | |
3333 | * | |
3334 | * returns: 1 successfully unmapped a shared pte page | |
3335 | * 0 the underlying pte page is not shared, or it is the last user | |
3336 | */ | |
3337 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
3338 | { | |
3339 | pgd_t *pgd = pgd_offset(mm, *addr); | |
3340 | pud_t *pud = pud_offset(pgd, *addr); | |
3341 | ||
3342 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
3343 | if (page_count(virt_to_page(ptep)) == 1) | |
3344 | return 0; | |
3345 | ||
3346 | pud_clear(pud); | |
3347 | put_page(virt_to_page(ptep)); | |
3348 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; | |
3349 | return 1; | |
3350 | } | |
9e5fc74c SC |
3351 | #define want_pmd_share() (1) |
3352 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
3353 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
3354 | { | |
3355 | return NULL; | |
3356 | } | |
3357 | #define want_pmd_share() (0) | |
3212b535 SC |
3358 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
3359 | ||
9e5fc74c SC |
3360 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
3361 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
3362 | unsigned long addr, unsigned long sz) | |
3363 | { | |
3364 | pgd_t *pgd; | |
3365 | pud_t *pud; | |
3366 | pte_t *pte = NULL; | |
3367 | ||
3368 | pgd = pgd_offset(mm, addr); | |
3369 | pud = pud_alloc(mm, pgd, addr); | |
3370 | if (pud) { | |
3371 | if (sz == PUD_SIZE) { | |
3372 | pte = (pte_t *)pud; | |
3373 | } else { | |
3374 | BUG_ON(sz != PMD_SIZE); | |
3375 | if (want_pmd_share() && pud_none(*pud)) | |
3376 | pte = huge_pmd_share(mm, addr, pud); | |
3377 | else | |
3378 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3379 | } | |
3380 | } | |
3381 | BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); | |
3382 | ||
3383 | return pte; | |
3384 | } | |
3385 | ||
3386 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) | |
3387 | { | |
3388 | pgd_t *pgd; | |
3389 | pud_t *pud; | |
3390 | pmd_t *pmd = NULL; | |
3391 | ||
3392 | pgd = pgd_offset(mm, addr); | |
3393 | if (pgd_present(*pgd)) { | |
3394 | pud = pud_offset(pgd, addr); | |
3395 | if (pud_present(*pud)) { | |
3396 | if (pud_huge(*pud)) | |
3397 | return (pte_t *)pud; | |
3398 | pmd = pmd_offset(pud, addr); | |
3399 | } | |
3400 | } | |
3401 | return (pte_t *) pmd; | |
3402 | } | |
3403 | ||
3404 | struct page * | |
3405 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, | |
3406 | pmd_t *pmd, int write) | |
3407 | { | |
3408 | struct page *page; | |
3409 | ||
3410 | page = pte_page(*(pte_t *)pmd); | |
3411 | if (page) | |
3412 | page += ((address & ~PMD_MASK) >> PAGE_SHIFT); | |
3413 | return page; | |
3414 | } | |
3415 | ||
3416 | struct page * | |
3417 | follow_huge_pud(struct mm_struct *mm, unsigned long address, | |
3418 | pud_t *pud, int write) | |
3419 | { | |
3420 | struct page *page; | |
3421 | ||
3422 | page = pte_page(*(pte_t *)pud); | |
3423 | if (page) | |
3424 | page += ((address & ~PUD_MASK) >> PAGE_SHIFT); | |
3425 | return page; | |
3426 | } | |
3427 | ||
3428 | #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */ | |
3429 | ||
3430 | /* Can be overriden by architectures */ | |
3431 | __attribute__((weak)) struct page * | |
3432 | follow_huge_pud(struct mm_struct *mm, unsigned long address, | |
3433 | pud_t *pud, int write) | |
3434 | { | |
3435 | BUG(); | |
3436 | return NULL; | |
3437 | } | |
3438 | ||
3439 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ | |
3440 | ||
d5bd9106 AK |
3441 | #ifdef CONFIG_MEMORY_FAILURE |
3442 | ||
6de2b1aa NH |
3443 | /* Should be called in hugetlb_lock */ |
3444 | static int is_hugepage_on_freelist(struct page *hpage) | |
3445 | { | |
3446 | struct page *page; | |
3447 | struct page *tmp; | |
3448 | struct hstate *h = page_hstate(hpage); | |
3449 | int nid = page_to_nid(hpage); | |
3450 | ||
3451 | list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru) | |
3452 | if (page == hpage) | |
3453 | return 1; | |
3454 | return 0; | |
3455 | } | |
3456 | ||
93f70f90 NH |
3457 | /* |
3458 | * This function is called from memory failure code. | |
3459 | * Assume the caller holds page lock of the head page. | |
3460 | */ | |
6de2b1aa | 3461 | int dequeue_hwpoisoned_huge_page(struct page *hpage) |
93f70f90 NH |
3462 | { |
3463 | struct hstate *h = page_hstate(hpage); | |
3464 | int nid = page_to_nid(hpage); | |
6de2b1aa | 3465 | int ret = -EBUSY; |
93f70f90 NH |
3466 | |
3467 | spin_lock(&hugetlb_lock); | |
6de2b1aa | 3468 | if (is_hugepage_on_freelist(hpage)) { |
56f2fb14 NH |
3469 | /* |
3470 | * Hwpoisoned hugepage isn't linked to activelist or freelist, | |
3471 | * but dangling hpage->lru can trigger list-debug warnings | |
3472 | * (this happens when we call unpoison_memory() on it), | |
3473 | * so let it point to itself with list_del_init(). | |
3474 | */ | |
3475 | list_del_init(&hpage->lru); | |
8c6c2ecb | 3476 | set_page_refcounted(hpage); |
6de2b1aa NH |
3477 | h->free_huge_pages--; |
3478 | h->free_huge_pages_node[nid]--; | |
3479 | ret = 0; | |
3480 | } | |
93f70f90 | 3481 | spin_unlock(&hugetlb_lock); |
6de2b1aa | 3482 | return ret; |
93f70f90 | 3483 | } |
6de2b1aa | 3484 | #endif |
31caf665 NH |
3485 | |
3486 | bool isolate_huge_page(struct page *page, struct list_head *list) | |
3487 | { | |
3488 | VM_BUG_ON(!PageHead(page)); | |
3489 | if (!get_page_unless_zero(page)) | |
3490 | return false; | |
3491 | spin_lock(&hugetlb_lock); | |
3492 | list_move_tail(&page->lru, list); | |
3493 | spin_unlock(&hugetlb_lock); | |
3494 | return true; | |
3495 | } | |
3496 | ||
3497 | void putback_active_hugepage(struct page *page) | |
3498 | { | |
3499 | VM_BUG_ON(!PageHead(page)); | |
3500 | spin_lock(&hugetlb_lock); | |
3501 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); | |
3502 | spin_unlock(&hugetlb_lock); | |
3503 | put_page(page); | |
3504 | } | |
c8721bbb NH |
3505 | |
3506 | bool is_hugepage_active(struct page *page) | |
3507 | { | |
3508 | VM_BUG_ON(!PageHuge(page)); | |
3509 | /* | |
3510 | * This function can be called for a tail page because the caller, | |
3511 | * scan_movable_pages, scans through a given pfn-range which typically | |
3512 | * covers one memory block. In systems using gigantic hugepage (1GB | |
3513 | * for x86_64,) a hugepage is larger than a memory block, and we don't | |
3514 | * support migrating such large hugepages for now, so return false | |
3515 | * when called for tail pages. | |
3516 | */ | |
3517 | if (PageTail(page)) | |
3518 | return false; | |
3519 | /* | |
3520 | * Refcount of a hwpoisoned hugepages is 1, but they are not active, | |
3521 | * so we should return false for them. | |
3522 | */ | |
3523 | if (unlikely(PageHWPoison(page))) | |
3524 | return false; | |
3525 | return page_count(page) > 0; | |
3526 | } |