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> |
3b32123d | 16 | #include <linux/compiler.h> |
aea47ff3 | 17 | #include <linux/cpuset.h> |
3935baa9 | 18 | #include <linux/mutex.h> |
aa888a74 | 19 | #include <linux/bootmem.h> |
a3437870 | 20 | #include <linux/sysfs.h> |
5a0e3ad6 | 21 | #include <linux/slab.h> |
0fe6e20b | 22 | #include <linux/rmap.h> |
fd6a03ed NH |
23 | #include <linux/swap.h> |
24 | #include <linux/swapops.h> | |
c8721bbb | 25 | #include <linux/page-isolation.h> |
8382d914 | 26 | #include <linux/jhash.h> |
d6606683 | 27 | |
63551ae0 DG |
28 | #include <asm/page.h> |
29 | #include <asm/pgtable.h> | |
24669e58 | 30 | #include <asm/tlb.h> |
63551ae0 | 31 | |
24669e58 | 32 | #include <linux/io.h> |
63551ae0 | 33 | #include <linux/hugetlb.h> |
9dd540e2 | 34 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 35 | #include <linux/node.h> |
7835e98b | 36 | #include "internal.h" |
1da177e4 LT |
37 | |
38 | const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL; | |
396faf03 | 39 | unsigned long hugepages_treat_as_movable; |
a5516438 | 40 | |
c3f38a38 | 41 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
42 | unsigned int default_hstate_idx; |
43 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
44 | ||
53ba51d2 JT |
45 | __initdata LIST_HEAD(huge_boot_pages); |
46 | ||
e5ff2159 AK |
47 | /* for command line parsing */ |
48 | static struct hstate * __initdata parsed_hstate; | |
49 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 50 | static unsigned long __initdata default_hstate_size; |
e5ff2159 | 51 | |
3935baa9 | 52 | /* |
31caf665 NH |
53 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
54 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 55 | */ |
c3f38a38 | 56 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 57 | |
8382d914 DB |
58 | /* |
59 | * Serializes faults on the same logical page. This is used to | |
60 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
61 | */ | |
62 | static int num_fault_mutexes; | |
63 | static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp; | |
64 | ||
90481622 DG |
65 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
66 | { | |
67 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
68 | ||
69 | spin_unlock(&spool->lock); | |
70 | ||
71 | /* If no pages are used, and no other handles to the subpool | |
72 | * remain, free the subpool the subpool remain */ | |
73 | if (free) | |
74 | kfree(spool); | |
75 | } | |
76 | ||
77 | struct hugepage_subpool *hugepage_new_subpool(long nr_blocks) | |
78 | { | |
79 | struct hugepage_subpool *spool; | |
80 | ||
81 | spool = kmalloc(sizeof(*spool), GFP_KERNEL); | |
82 | if (!spool) | |
83 | return NULL; | |
84 | ||
85 | spin_lock_init(&spool->lock); | |
86 | spool->count = 1; | |
87 | spool->max_hpages = nr_blocks; | |
88 | spool->used_hpages = 0; | |
89 | ||
90 | return spool; | |
91 | } | |
92 | ||
93 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
94 | { | |
95 | spin_lock(&spool->lock); | |
96 | BUG_ON(!spool->count); | |
97 | spool->count--; | |
98 | unlock_or_release_subpool(spool); | |
99 | } | |
100 | ||
101 | static int hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
102 | long delta) | |
103 | { | |
104 | int ret = 0; | |
105 | ||
106 | if (!spool) | |
107 | return 0; | |
108 | ||
109 | spin_lock(&spool->lock); | |
110 | if ((spool->used_hpages + delta) <= spool->max_hpages) { | |
111 | spool->used_hpages += delta; | |
112 | } else { | |
113 | ret = -ENOMEM; | |
114 | } | |
115 | spin_unlock(&spool->lock); | |
116 | ||
117 | return ret; | |
118 | } | |
119 | ||
120 | static void hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
121 | long delta) | |
122 | { | |
123 | if (!spool) | |
124 | return; | |
125 | ||
126 | spin_lock(&spool->lock); | |
127 | spool->used_hpages -= delta; | |
128 | /* If hugetlbfs_put_super couldn't free spool due to | |
129 | * an outstanding quota reference, free it now. */ | |
130 | unlock_or_release_subpool(spool); | |
131 | } | |
132 | ||
133 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
134 | { | |
135 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
136 | } | |
137 | ||
138 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
139 | { | |
496ad9aa | 140 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
141 | } |
142 | ||
96822904 AW |
143 | /* |
144 | * Region tracking -- allows tracking of reservations and instantiated pages | |
145 | * across the pages in a mapping. | |
84afd99b | 146 | * |
7b24d861 DB |
147 | * The region data structures are embedded into a resv_map and |
148 | * protected by a resv_map's lock | |
96822904 AW |
149 | */ |
150 | struct file_region { | |
151 | struct list_head link; | |
152 | long from; | |
153 | long to; | |
154 | }; | |
155 | ||
1406ec9b | 156 | static long region_add(struct resv_map *resv, long f, long t) |
96822904 | 157 | { |
1406ec9b | 158 | struct list_head *head = &resv->regions; |
96822904 AW |
159 | struct file_region *rg, *nrg, *trg; |
160 | ||
7b24d861 | 161 | spin_lock(&resv->lock); |
96822904 AW |
162 | /* Locate the region we are either in or before. */ |
163 | list_for_each_entry(rg, head, link) | |
164 | if (f <= rg->to) | |
165 | break; | |
166 | ||
167 | /* Round our left edge to the current segment if it encloses us. */ | |
168 | if (f > rg->from) | |
169 | f = rg->from; | |
170 | ||
171 | /* Check for and consume any regions we now overlap with. */ | |
172 | nrg = rg; | |
173 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
174 | if (&rg->link == head) | |
175 | break; | |
176 | if (rg->from > t) | |
177 | break; | |
178 | ||
179 | /* If this area reaches higher then extend our area to | |
180 | * include it completely. If this is not the first area | |
181 | * which we intend to reuse, free it. */ | |
182 | if (rg->to > t) | |
183 | t = rg->to; | |
184 | if (rg != nrg) { | |
185 | list_del(&rg->link); | |
186 | kfree(rg); | |
187 | } | |
188 | } | |
189 | nrg->from = f; | |
190 | nrg->to = t; | |
7b24d861 | 191 | spin_unlock(&resv->lock); |
96822904 AW |
192 | return 0; |
193 | } | |
194 | ||
1406ec9b | 195 | static long region_chg(struct resv_map *resv, long f, long t) |
96822904 | 196 | { |
1406ec9b | 197 | struct list_head *head = &resv->regions; |
7b24d861 | 198 | struct file_region *rg, *nrg = NULL; |
96822904 AW |
199 | long chg = 0; |
200 | ||
7b24d861 DB |
201 | retry: |
202 | spin_lock(&resv->lock); | |
96822904 AW |
203 | /* Locate the region we are before or in. */ |
204 | list_for_each_entry(rg, head, link) | |
205 | if (f <= rg->to) | |
206 | break; | |
207 | ||
208 | /* If we are below the current region then a new region is required. | |
209 | * Subtle, allocate a new region at the position but make it zero | |
210 | * size such that we can guarantee to record the reservation. */ | |
211 | if (&rg->link == head || t < rg->from) { | |
7b24d861 DB |
212 | if (!nrg) { |
213 | spin_unlock(&resv->lock); | |
214 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
215 | if (!nrg) | |
216 | return -ENOMEM; | |
217 | ||
218 | nrg->from = f; | |
219 | nrg->to = f; | |
220 | INIT_LIST_HEAD(&nrg->link); | |
221 | goto retry; | |
222 | } | |
96822904 | 223 | |
7b24d861 DB |
224 | list_add(&nrg->link, rg->link.prev); |
225 | chg = t - f; | |
226 | goto out_nrg; | |
96822904 AW |
227 | } |
228 | ||
229 | /* Round our left edge to the current segment if it encloses us. */ | |
230 | if (f > rg->from) | |
231 | f = rg->from; | |
232 | chg = t - f; | |
233 | ||
234 | /* Check for and consume any regions we now overlap with. */ | |
235 | list_for_each_entry(rg, rg->link.prev, link) { | |
236 | if (&rg->link == head) | |
237 | break; | |
238 | if (rg->from > t) | |
7b24d861 | 239 | goto out; |
96822904 | 240 | |
25985edc | 241 | /* We overlap with this area, if it extends further than |
96822904 AW |
242 | * us then we must extend ourselves. Account for its |
243 | * existing reservation. */ | |
244 | if (rg->to > t) { | |
245 | chg += rg->to - t; | |
246 | t = rg->to; | |
247 | } | |
248 | chg -= rg->to - rg->from; | |
249 | } | |
7b24d861 DB |
250 | |
251 | out: | |
252 | spin_unlock(&resv->lock); | |
253 | /* We already know we raced and no longer need the new region */ | |
254 | kfree(nrg); | |
255 | return chg; | |
256 | out_nrg: | |
257 | spin_unlock(&resv->lock); | |
96822904 AW |
258 | return chg; |
259 | } | |
260 | ||
1406ec9b | 261 | static long region_truncate(struct resv_map *resv, long end) |
96822904 | 262 | { |
1406ec9b | 263 | struct list_head *head = &resv->regions; |
96822904 AW |
264 | struct file_region *rg, *trg; |
265 | long chg = 0; | |
266 | ||
7b24d861 | 267 | spin_lock(&resv->lock); |
96822904 AW |
268 | /* Locate the region we are either in or before. */ |
269 | list_for_each_entry(rg, head, link) | |
270 | if (end <= rg->to) | |
271 | break; | |
272 | if (&rg->link == head) | |
7b24d861 | 273 | goto out; |
96822904 AW |
274 | |
275 | /* If we are in the middle of a region then adjust it. */ | |
276 | if (end > rg->from) { | |
277 | chg = rg->to - end; | |
278 | rg->to = end; | |
279 | rg = list_entry(rg->link.next, typeof(*rg), link); | |
280 | } | |
281 | ||
282 | /* Drop any remaining regions. */ | |
283 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
284 | if (&rg->link == head) | |
285 | break; | |
286 | chg += rg->to - rg->from; | |
287 | list_del(&rg->link); | |
288 | kfree(rg); | |
289 | } | |
7b24d861 DB |
290 | |
291 | out: | |
292 | spin_unlock(&resv->lock); | |
96822904 AW |
293 | return chg; |
294 | } | |
295 | ||
1406ec9b | 296 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 297 | { |
1406ec9b | 298 | struct list_head *head = &resv->regions; |
84afd99b AW |
299 | struct file_region *rg; |
300 | long chg = 0; | |
301 | ||
7b24d861 | 302 | spin_lock(&resv->lock); |
84afd99b AW |
303 | /* Locate each segment we overlap with, and count that overlap. */ |
304 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
305 | long seg_from; |
306 | long seg_to; | |
84afd99b AW |
307 | |
308 | if (rg->to <= f) | |
309 | continue; | |
310 | if (rg->from >= t) | |
311 | break; | |
312 | ||
313 | seg_from = max(rg->from, f); | |
314 | seg_to = min(rg->to, t); | |
315 | ||
316 | chg += seg_to - seg_from; | |
317 | } | |
7b24d861 | 318 | spin_unlock(&resv->lock); |
84afd99b AW |
319 | |
320 | return chg; | |
321 | } | |
322 | ||
e7c4b0bf AW |
323 | /* |
324 | * Convert the address within this vma to the page offset within | |
325 | * the mapping, in pagecache page units; huge pages here. | |
326 | */ | |
a5516438 AK |
327 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
328 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 329 | { |
a5516438 AK |
330 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
331 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
332 | } |
333 | ||
0fe6e20b NH |
334 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
335 | unsigned long address) | |
336 | { | |
337 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
338 | } | |
339 | ||
08fba699 MG |
340 | /* |
341 | * Return the size of the pages allocated when backing a VMA. In the majority | |
342 | * cases this will be same size as used by the page table entries. | |
343 | */ | |
344 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
345 | { | |
346 | struct hstate *hstate; | |
347 | ||
348 | if (!is_vm_hugetlb_page(vma)) | |
349 | return PAGE_SIZE; | |
350 | ||
351 | hstate = hstate_vma(vma); | |
352 | ||
2415cf12 | 353 | return 1UL << huge_page_shift(hstate); |
08fba699 | 354 | } |
f340ca0f | 355 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 356 | |
3340289d MG |
357 | /* |
358 | * Return the page size being used by the MMU to back a VMA. In the majority | |
359 | * of cases, the page size used by the kernel matches the MMU size. On | |
360 | * architectures where it differs, an architecture-specific version of this | |
361 | * function is required. | |
362 | */ | |
363 | #ifndef vma_mmu_pagesize | |
364 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) | |
365 | { | |
366 | return vma_kernel_pagesize(vma); | |
367 | } | |
368 | #endif | |
369 | ||
84afd99b AW |
370 | /* |
371 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
372 | * bits of the reservation map pointer, which are always clear due to | |
373 | * alignment. | |
374 | */ | |
375 | #define HPAGE_RESV_OWNER (1UL << 0) | |
376 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 377 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 378 | |
a1e78772 MG |
379 | /* |
380 | * These helpers are used to track how many pages are reserved for | |
381 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
382 | * is guaranteed to have their future faults succeed. | |
383 | * | |
384 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
385 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
386 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
387 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
388 | * |
389 | * The private mapping reservation is represented in a subtly different | |
390 | * manner to a shared mapping. A shared mapping has a region map associated | |
391 | * with the underlying file, this region map represents the backing file | |
392 | * pages which have ever had a reservation assigned which this persists even | |
393 | * after the page is instantiated. A private mapping has a region map | |
394 | * associated with the original mmap which is attached to all VMAs which | |
395 | * reference it, this region map represents those offsets which have consumed | |
396 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 397 | */ |
e7c4b0bf AW |
398 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
399 | { | |
400 | return (unsigned long)vma->vm_private_data; | |
401 | } | |
402 | ||
403 | static void set_vma_private_data(struct vm_area_struct *vma, | |
404 | unsigned long value) | |
405 | { | |
406 | vma->vm_private_data = (void *)value; | |
407 | } | |
408 | ||
9119a41e | 409 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
410 | { |
411 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
412 | if (!resv_map) | |
413 | return NULL; | |
414 | ||
415 | kref_init(&resv_map->refs); | |
7b24d861 | 416 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
417 | INIT_LIST_HEAD(&resv_map->regions); |
418 | ||
419 | return resv_map; | |
420 | } | |
421 | ||
9119a41e | 422 | void resv_map_release(struct kref *ref) |
84afd99b AW |
423 | { |
424 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
425 | ||
426 | /* Clear out any active regions before we release the map. */ | |
1406ec9b | 427 | region_truncate(resv_map, 0); |
84afd99b AW |
428 | kfree(resv_map); |
429 | } | |
430 | ||
4e35f483 JK |
431 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
432 | { | |
433 | return inode->i_mapping->private_data; | |
434 | } | |
435 | ||
84afd99b | 436 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 MG |
437 | { |
438 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
4e35f483 JK |
439 | if (vma->vm_flags & VM_MAYSHARE) { |
440 | struct address_space *mapping = vma->vm_file->f_mapping; | |
441 | struct inode *inode = mapping->host; | |
442 | ||
443 | return inode_resv_map(inode); | |
444 | ||
445 | } else { | |
84afd99b AW |
446 | return (struct resv_map *)(get_vma_private_data(vma) & |
447 | ~HPAGE_RESV_MASK); | |
4e35f483 | 448 | } |
a1e78772 MG |
449 | } |
450 | ||
84afd99b | 451 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 MG |
452 | { |
453 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 454 | VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); |
a1e78772 | 455 | |
84afd99b AW |
456 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
457 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
458 | } |
459 | ||
460 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
461 | { | |
04f2cbe3 | 462 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); |
f83a275d | 463 | VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); |
e7c4b0bf AW |
464 | |
465 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
466 | } |
467 | ||
468 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
469 | { | |
470 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
e7c4b0bf AW |
471 | |
472 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
473 | } |
474 | ||
04f2cbe3 | 475 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
476 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
477 | { | |
478 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 479 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
480 | vma->vm_private_data = (void *)0; |
481 | } | |
482 | ||
483 | /* Returns true if the VMA has associated reserve pages */ | |
af0ed73e | 484 | static int vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 485 | { |
af0ed73e JK |
486 | if (vma->vm_flags & VM_NORESERVE) { |
487 | /* | |
488 | * This address is already reserved by other process(chg == 0), | |
489 | * so, we should decrement reserved count. Without decrementing, | |
490 | * reserve count remains after releasing inode, because this | |
491 | * allocated page will go into page cache and is regarded as | |
492 | * coming from reserved pool in releasing step. Currently, we | |
493 | * don't have any other solution to deal with this situation | |
494 | * properly, so add work-around here. | |
495 | */ | |
496 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
497 | return 1; | |
498 | else | |
499 | return 0; | |
500 | } | |
a63884e9 JK |
501 | |
502 | /* Shared mappings always use reserves */ | |
f83a275d | 503 | if (vma->vm_flags & VM_MAYSHARE) |
7f09ca51 | 504 | return 1; |
a63884e9 JK |
505 | |
506 | /* | |
507 | * Only the process that called mmap() has reserves for | |
508 | * private mappings. | |
509 | */ | |
7f09ca51 MG |
510 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
511 | return 1; | |
a63884e9 | 512 | |
7f09ca51 | 513 | return 0; |
a1e78772 MG |
514 | } |
515 | ||
a5516438 | 516 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
517 | { |
518 | int nid = page_to_nid(page); | |
0edaecfa | 519 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
520 | h->free_huge_pages++; |
521 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
522 | } |
523 | ||
bf50bab2 NH |
524 | static struct page *dequeue_huge_page_node(struct hstate *h, int nid) |
525 | { | |
526 | struct page *page; | |
527 | ||
c8721bbb NH |
528 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
529 | if (!is_migrate_isolate_page(page)) | |
530 | break; | |
531 | /* | |
532 | * if 'non-isolated free hugepage' not found on the list, | |
533 | * the allocation fails. | |
534 | */ | |
535 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 536 | return NULL; |
0edaecfa | 537 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 538 | set_page_refcounted(page); |
bf50bab2 NH |
539 | h->free_huge_pages--; |
540 | h->free_huge_pages_node[nid]--; | |
541 | return page; | |
542 | } | |
543 | ||
86cdb465 NH |
544 | /* Movability of hugepages depends on migration support. */ |
545 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
546 | { | |
547 | if (hugepages_treat_as_movable || hugepage_migration_support(h)) | |
548 | return GFP_HIGHUSER_MOVABLE; | |
549 | else | |
550 | return GFP_HIGHUSER; | |
551 | } | |
552 | ||
a5516438 AK |
553 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
554 | struct vm_area_struct *vma, | |
af0ed73e JK |
555 | unsigned long address, int avoid_reserve, |
556 | long chg) | |
1da177e4 | 557 | { |
b1c12cbc | 558 | struct page *page = NULL; |
480eccf9 | 559 | struct mempolicy *mpol; |
19770b32 | 560 | nodemask_t *nodemask; |
c0ff7453 | 561 | struct zonelist *zonelist; |
dd1a239f MG |
562 | struct zone *zone; |
563 | struct zoneref *z; | |
cc9a6c87 | 564 | unsigned int cpuset_mems_cookie; |
1da177e4 | 565 | |
a1e78772 MG |
566 | /* |
567 | * A child process with MAP_PRIVATE mappings created by their parent | |
568 | * have no page reserves. This check ensures that reservations are | |
569 | * not "stolen". The child may still get SIGKILLed | |
570 | */ | |
af0ed73e | 571 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 572 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 573 | goto err; |
a1e78772 | 574 | |
04f2cbe3 | 575 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 576 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 577 | goto err; |
04f2cbe3 | 578 | |
9966c4bb | 579 | retry_cpuset: |
d26914d1 | 580 | cpuset_mems_cookie = read_mems_allowed_begin(); |
9966c4bb | 581 | zonelist = huge_zonelist(vma, address, |
86cdb465 | 582 | htlb_alloc_mask(h), &mpol, &nodemask); |
9966c4bb | 583 | |
19770b32 MG |
584 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
585 | MAX_NR_ZONES - 1, nodemask) { | |
86cdb465 | 586 | if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask(h))) { |
bf50bab2 NH |
587 | page = dequeue_huge_page_node(h, zone_to_nid(zone)); |
588 | if (page) { | |
af0ed73e JK |
589 | if (avoid_reserve) |
590 | break; | |
591 | if (!vma_has_reserves(vma, chg)) | |
592 | break; | |
593 | ||
07443a85 | 594 | SetPagePrivate(page); |
af0ed73e | 595 | h->resv_huge_pages--; |
bf50bab2 NH |
596 | break; |
597 | } | |
3abf7afd | 598 | } |
1da177e4 | 599 | } |
cc9a6c87 | 600 | |
52cd3b07 | 601 | mpol_cond_put(mpol); |
d26914d1 | 602 | if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) |
cc9a6c87 | 603 | goto retry_cpuset; |
1da177e4 | 604 | return page; |
cc9a6c87 MG |
605 | |
606 | err: | |
cc9a6c87 | 607 | return NULL; |
1da177e4 LT |
608 | } |
609 | ||
1cac6f2c LC |
610 | /* |
611 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
612 | * We may have allocated or freed a huge page based on a different | |
613 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
614 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
615 | * node for alloc or free. | |
616 | */ | |
617 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
618 | { | |
619 | nid = next_node(nid, *nodes_allowed); | |
620 | if (nid == MAX_NUMNODES) | |
621 | nid = first_node(*nodes_allowed); | |
622 | VM_BUG_ON(nid >= MAX_NUMNODES); | |
623 | ||
624 | return nid; | |
625 | } | |
626 | ||
627 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
628 | { | |
629 | if (!node_isset(nid, *nodes_allowed)) | |
630 | nid = next_node_allowed(nid, nodes_allowed); | |
631 | return nid; | |
632 | } | |
633 | ||
634 | /* | |
635 | * returns the previously saved node ["this node"] from which to | |
636 | * allocate a persistent huge page for the pool and advance the | |
637 | * next node from which to allocate, handling wrap at end of node | |
638 | * mask. | |
639 | */ | |
640 | static int hstate_next_node_to_alloc(struct hstate *h, | |
641 | nodemask_t *nodes_allowed) | |
642 | { | |
643 | int nid; | |
644 | ||
645 | VM_BUG_ON(!nodes_allowed); | |
646 | ||
647 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
648 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
649 | ||
650 | return nid; | |
651 | } | |
652 | ||
653 | /* | |
654 | * helper for free_pool_huge_page() - return the previously saved | |
655 | * node ["this node"] from which to free a huge page. Advance the | |
656 | * next node id whether or not we find a free huge page to free so | |
657 | * that the next attempt to free addresses the next node. | |
658 | */ | |
659 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
660 | { | |
661 | int nid; | |
662 | ||
663 | VM_BUG_ON(!nodes_allowed); | |
664 | ||
665 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
666 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
667 | ||
668 | return nid; | |
669 | } | |
670 | ||
671 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
672 | for (nr_nodes = nodes_weight(*mask); \ | |
673 | nr_nodes > 0 && \ | |
674 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
675 | nr_nodes--) | |
676 | ||
677 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
678 | for (nr_nodes = nodes_weight(*mask); \ | |
679 | nr_nodes > 0 && \ | |
680 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
681 | nr_nodes--) | |
682 | ||
944d9fec LC |
683 | #if defined(CONFIG_CMA) && defined(CONFIG_X86_64) |
684 | static void destroy_compound_gigantic_page(struct page *page, | |
685 | unsigned long order) | |
686 | { | |
687 | int i; | |
688 | int nr_pages = 1 << order; | |
689 | struct page *p = page + 1; | |
690 | ||
691 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { | |
692 | __ClearPageTail(p); | |
693 | set_page_refcounted(p); | |
694 | p->first_page = NULL; | |
695 | } | |
696 | ||
697 | set_compound_order(page, 0); | |
698 | __ClearPageHead(page); | |
699 | } | |
700 | ||
701 | static void free_gigantic_page(struct page *page, unsigned order) | |
702 | { | |
703 | free_contig_range(page_to_pfn(page), 1 << order); | |
704 | } | |
705 | ||
706 | static int __alloc_gigantic_page(unsigned long start_pfn, | |
707 | unsigned long nr_pages) | |
708 | { | |
709 | unsigned long end_pfn = start_pfn + nr_pages; | |
710 | return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE); | |
711 | } | |
712 | ||
713 | static bool pfn_range_valid_gigantic(unsigned long start_pfn, | |
714 | unsigned long nr_pages) | |
715 | { | |
716 | unsigned long i, end_pfn = start_pfn + nr_pages; | |
717 | struct page *page; | |
718 | ||
719 | for (i = start_pfn; i < end_pfn; i++) { | |
720 | if (!pfn_valid(i)) | |
721 | return false; | |
722 | ||
723 | page = pfn_to_page(i); | |
724 | ||
725 | if (PageReserved(page)) | |
726 | return false; | |
727 | ||
728 | if (page_count(page) > 0) | |
729 | return false; | |
730 | ||
731 | if (PageHuge(page)) | |
732 | return false; | |
733 | } | |
734 | ||
735 | return true; | |
736 | } | |
737 | ||
738 | static bool zone_spans_last_pfn(const struct zone *zone, | |
739 | unsigned long start_pfn, unsigned long nr_pages) | |
740 | { | |
741 | unsigned long last_pfn = start_pfn + nr_pages - 1; | |
742 | return zone_spans_pfn(zone, last_pfn); | |
743 | } | |
744 | ||
745 | static struct page *alloc_gigantic_page(int nid, unsigned order) | |
746 | { | |
747 | unsigned long nr_pages = 1 << order; | |
748 | unsigned long ret, pfn, flags; | |
749 | struct zone *z; | |
750 | ||
751 | z = NODE_DATA(nid)->node_zones; | |
752 | for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) { | |
753 | spin_lock_irqsave(&z->lock, flags); | |
754 | ||
755 | pfn = ALIGN(z->zone_start_pfn, nr_pages); | |
756 | while (zone_spans_last_pfn(z, pfn, nr_pages)) { | |
757 | if (pfn_range_valid_gigantic(pfn, nr_pages)) { | |
758 | /* | |
759 | * We release the zone lock here because | |
760 | * alloc_contig_range() will also lock the zone | |
761 | * at some point. If there's an allocation | |
762 | * spinning on this lock, it may win the race | |
763 | * and cause alloc_contig_range() to fail... | |
764 | */ | |
765 | spin_unlock_irqrestore(&z->lock, flags); | |
766 | ret = __alloc_gigantic_page(pfn, nr_pages); | |
767 | if (!ret) | |
768 | return pfn_to_page(pfn); | |
769 | spin_lock_irqsave(&z->lock, flags); | |
770 | } | |
771 | pfn += nr_pages; | |
772 | } | |
773 | ||
774 | spin_unlock_irqrestore(&z->lock, flags); | |
775 | } | |
776 | ||
777 | return NULL; | |
778 | } | |
779 | ||
780 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
781 | static void prep_compound_gigantic_page(struct page *page, unsigned long order); | |
782 | ||
783 | static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) | |
784 | { | |
785 | struct page *page; | |
786 | ||
787 | page = alloc_gigantic_page(nid, huge_page_order(h)); | |
788 | if (page) { | |
789 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
790 | prep_new_huge_page(h, page, nid); | |
791 | } | |
792 | ||
793 | return page; | |
794 | } | |
795 | ||
796 | static int alloc_fresh_gigantic_page(struct hstate *h, | |
797 | nodemask_t *nodes_allowed) | |
798 | { | |
799 | struct page *page = NULL; | |
800 | int nr_nodes, node; | |
801 | ||
802 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
803 | page = alloc_fresh_gigantic_page_node(h, node); | |
804 | if (page) | |
805 | return 1; | |
806 | } | |
807 | ||
808 | return 0; | |
809 | } | |
810 | ||
811 | static inline bool gigantic_page_supported(void) { return true; } | |
812 | #else | |
813 | static inline bool gigantic_page_supported(void) { return false; } | |
814 | static inline void free_gigantic_page(struct page *page, unsigned order) { } | |
815 | static inline void destroy_compound_gigantic_page(struct page *page, | |
816 | unsigned long order) { } | |
817 | static inline int alloc_fresh_gigantic_page(struct hstate *h, | |
818 | nodemask_t *nodes_allowed) { return 0; } | |
819 | #endif | |
820 | ||
a5516438 | 821 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
822 | { |
823 | int i; | |
a5516438 | 824 | |
944d9fec LC |
825 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
826 | return; | |
18229df5 | 827 | |
a5516438 AK |
828 | h->nr_huge_pages--; |
829 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
830 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
831 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
832 | 1 << PG_referenced | 1 << PG_dirty | | |
a7407a27 LC |
833 | 1 << PG_active | 1 << PG_private | |
834 | 1 << PG_writeback); | |
6af2acb6 | 835 | } |
309381fe | 836 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
6af2acb6 AL |
837 | set_compound_page_dtor(page, NULL); |
838 | set_page_refcounted(page); | |
944d9fec LC |
839 | if (hstate_is_gigantic(h)) { |
840 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
841 | free_gigantic_page(page, huge_page_order(h)); | |
842 | } else { | |
843 | arch_release_hugepage(page); | |
844 | __free_pages(page, huge_page_order(h)); | |
845 | } | |
6af2acb6 AL |
846 | } |
847 | ||
e5ff2159 AK |
848 | struct hstate *size_to_hstate(unsigned long size) |
849 | { | |
850 | struct hstate *h; | |
851 | ||
852 | for_each_hstate(h) { | |
853 | if (huge_page_size(h) == size) | |
854 | return h; | |
855 | } | |
856 | return NULL; | |
857 | } | |
858 | ||
27a85ef1 DG |
859 | static void free_huge_page(struct page *page) |
860 | { | |
a5516438 AK |
861 | /* |
862 | * Can't pass hstate in here because it is called from the | |
863 | * compound page destructor. | |
864 | */ | |
e5ff2159 | 865 | struct hstate *h = page_hstate(page); |
7893d1d5 | 866 | int nid = page_to_nid(page); |
90481622 DG |
867 | struct hugepage_subpool *spool = |
868 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 869 | bool restore_reserve; |
27a85ef1 | 870 | |
e5df70ab | 871 | set_page_private(page, 0); |
23be7468 | 872 | page->mapping = NULL; |
7893d1d5 | 873 | BUG_ON(page_count(page)); |
0fe6e20b | 874 | BUG_ON(page_mapcount(page)); |
07443a85 | 875 | restore_reserve = PagePrivate(page); |
16c794b4 | 876 | ClearPagePrivate(page); |
27a85ef1 DG |
877 | |
878 | spin_lock(&hugetlb_lock); | |
6d76dcf4 AK |
879 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
880 | pages_per_huge_page(h), page); | |
07443a85 JK |
881 | if (restore_reserve) |
882 | h->resv_huge_pages++; | |
883 | ||
944d9fec | 884 | if (h->surplus_huge_pages_node[nid]) { |
0edaecfa AK |
885 | /* remove the page from active list */ |
886 | list_del(&page->lru); | |
a5516438 AK |
887 | update_and_free_page(h, page); |
888 | h->surplus_huge_pages--; | |
889 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 890 | } else { |
5d3a551c | 891 | arch_clear_hugepage_flags(page); |
a5516438 | 892 | enqueue_huge_page(h, page); |
7893d1d5 | 893 | } |
27a85ef1 | 894 | spin_unlock(&hugetlb_lock); |
90481622 | 895 | hugepage_subpool_put_pages(spool, 1); |
27a85ef1 DG |
896 | } |
897 | ||
a5516438 | 898 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 899 | { |
0edaecfa | 900 | INIT_LIST_HEAD(&page->lru); |
b7ba30c6 AK |
901 | set_compound_page_dtor(page, free_huge_page); |
902 | spin_lock(&hugetlb_lock); | |
9dd540e2 | 903 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
904 | h->nr_huge_pages++; |
905 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 AK |
906 | spin_unlock(&hugetlb_lock); |
907 | put_page(page); /* free it into the hugepage allocator */ | |
908 | } | |
909 | ||
2906dd52 | 910 | static void prep_compound_gigantic_page(struct page *page, unsigned long order) |
20a0307c WF |
911 | { |
912 | int i; | |
913 | int nr_pages = 1 << order; | |
914 | struct page *p = page + 1; | |
915 | ||
916 | /* we rely on prep_new_huge_page to set the destructor */ | |
917 | set_compound_order(page, order); | |
918 | __SetPageHead(page); | |
ef5a22be | 919 | __ClearPageReserved(page); |
20a0307c WF |
920 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
921 | __SetPageTail(p); | |
ef5a22be AA |
922 | /* |
923 | * For gigantic hugepages allocated through bootmem at | |
924 | * boot, it's safer to be consistent with the not-gigantic | |
925 | * hugepages and clear the PG_reserved bit from all tail pages | |
926 | * too. Otherwse drivers using get_user_pages() to access tail | |
927 | * pages may get the reference counting wrong if they see | |
928 | * PG_reserved set on a tail page (despite the head page not | |
929 | * having PG_reserved set). Enforcing this consistency between | |
930 | * head and tail pages allows drivers to optimize away a check | |
931 | * on the head page when they need know if put_page() is needed | |
932 | * after get_user_pages(). | |
933 | */ | |
934 | __ClearPageReserved(p); | |
58a84aa9 | 935 | set_page_count(p, 0); |
20a0307c WF |
936 | p->first_page = page; |
937 | } | |
938 | } | |
939 | ||
7795912c AM |
940 | /* |
941 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
942 | * transparent huge pages. See the PageTransHuge() documentation for more | |
943 | * details. | |
944 | */ | |
20a0307c WF |
945 | int PageHuge(struct page *page) |
946 | { | |
20a0307c WF |
947 | if (!PageCompound(page)) |
948 | return 0; | |
949 | ||
950 | page = compound_head(page); | |
758f66a2 | 951 | return get_compound_page_dtor(page) == free_huge_page; |
20a0307c | 952 | } |
43131e14 NH |
953 | EXPORT_SYMBOL_GPL(PageHuge); |
954 | ||
27c73ae7 AA |
955 | /* |
956 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
957 | * normal or transparent huge pages. | |
958 | */ | |
959 | int PageHeadHuge(struct page *page_head) | |
960 | { | |
27c73ae7 AA |
961 | if (!PageHead(page_head)) |
962 | return 0; | |
963 | ||
758f66a2 | 964 | return get_compound_page_dtor(page_head) == free_huge_page; |
27c73ae7 | 965 | } |
27c73ae7 | 966 | |
13d60f4b ZY |
967 | pgoff_t __basepage_index(struct page *page) |
968 | { | |
969 | struct page *page_head = compound_head(page); | |
970 | pgoff_t index = page_index(page_head); | |
971 | unsigned long compound_idx; | |
972 | ||
973 | if (!PageHuge(page_head)) | |
974 | return page_index(page); | |
975 | ||
976 | if (compound_order(page_head) >= MAX_ORDER) | |
977 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
978 | else | |
979 | compound_idx = page - page_head; | |
980 | ||
981 | return (index << compound_order(page_head)) + compound_idx; | |
982 | } | |
983 | ||
a5516438 | 984 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
1da177e4 | 985 | { |
1da177e4 | 986 | struct page *page; |
f96efd58 | 987 | |
6484eb3e | 988 | page = alloc_pages_exact_node(nid, |
86cdb465 | 989 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
551883ae | 990 | __GFP_REPEAT|__GFP_NOWARN, |
a5516438 | 991 | huge_page_order(h)); |
1da177e4 | 992 | if (page) { |
7f2e9525 | 993 | if (arch_prepare_hugepage(page)) { |
caff3a2c | 994 | __free_pages(page, huge_page_order(h)); |
7b8ee84d | 995 | return NULL; |
7f2e9525 | 996 | } |
a5516438 | 997 | prep_new_huge_page(h, page, nid); |
1da177e4 | 998 | } |
63b4613c NA |
999 | |
1000 | return page; | |
1001 | } | |
1002 | ||
b2261026 JK |
1003 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
1004 | { | |
1005 | struct page *page; | |
1006 | int nr_nodes, node; | |
1007 | int ret = 0; | |
1008 | ||
1009 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1010 | page = alloc_fresh_huge_page_node(h, node); | |
1011 | if (page) { | |
1012 | ret = 1; | |
1013 | break; | |
1014 | } | |
1015 | } | |
1016 | ||
1017 | if (ret) | |
1018 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
1019 | else | |
1020 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
1021 | ||
1022 | return ret; | |
1023 | } | |
1024 | ||
e8c5c824 LS |
1025 | /* |
1026 | * Free huge page from pool from next node to free. | |
1027 | * Attempt to keep persistent huge pages more or less | |
1028 | * balanced over allowed nodes. | |
1029 | * Called with hugetlb_lock locked. | |
1030 | */ | |
6ae11b27 LS |
1031 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1032 | bool acct_surplus) | |
e8c5c824 | 1033 | { |
b2261026 | 1034 | int nr_nodes, node; |
e8c5c824 LS |
1035 | int ret = 0; |
1036 | ||
b2261026 | 1037 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1038 | /* |
1039 | * If we're returning unused surplus pages, only examine | |
1040 | * nodes with surplus pages. | |
1041 | */ | |
b2261026 JK |
1042 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1043 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1044 | struct page *page = |
b2261026 | 1045 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1046 | struct page, lru); |
1047 | list_del(&page->lru); | |
1048 | h->free_huge_pages--; | |
b2261026 | 1049 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1050 | if (acct_surplus) { |
1051 | h->surplus_huge_pages--; | |
b2261026 | 1052 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1053 | } |
e8c5c824 LS |
1054 | update_and_free_page(h, page); |
1055 | ret = 1; | |
9a76db09 | 1056 | break; |
e8c5c824 | 1057 | } |
b2261026 | 1058 | } |
e8c5c824 LS |
1059 | |
1060 | return ret; | |
1061 | } | |
1062 | ||
c8721bbb NH |
1063 | /* |
1064 | * Dissolve a given free hugepage into free buddy pages. This function does | |
1065 | * nothing for in-use (including surplus) hugepages. | |
1066 | */ | |
1067 | static void dissolve_free_huge_page(struct page *page) | |
1068 | { | |
1069 | spin_lock(&hugetlb_lock); | |
1070 | if (PageHuge(page) && !page_count(page)) { | |
1071 | struct hstate *h = page_hstate(page); | |
1072 | int nid = page_to_nid(page); | |
1073 | list_del(&page->lru); | |
1074 | h->free_huge_pages--; | |
1075 | h->free_huge_pages_node[nid]--; | |
1076 | update_and_free_page(h, page); | |
1077 | } | |
1078 | spin_unlock(&hugetlb_lock); | |
1079 | } | |
1080 | ||
1081 | /* | |
1082 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1083 | * make specified memory blocks removable from the system. | |
1084 | * Note that start_pfn should aligned with (minimum) hugepage size. | |
1085 | */ | |
1086 | void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) | |
1087 | { | |
1088 | unsigned int order = 8 * sizeof(void *); | |
1089 | unsigned long pfn; | |
1090 | struct hstate *h; | |
1091 | ||
1092 | /* Set scan step to minimum hugepage size */ | |
1093 | for_each_hstate(h) | |
1094 | if (order > huge_page_order(h)) | |
1095 | order = huge_page_order(h); | |
1096 | VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order)); | |
1097 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) | |
1098 | dissolve_free_huge_page(pfn_to_page(pfn)); | |
1099 | } | |
1100 | ||
bf50bab2 | 1101 | static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) |
7893d1d5 AL |
1102 | { |
1103 | struct page *page; | |
bf50bab2 | 1104 | unsigned int r_nid; |
7893d1d5 | 1105 | |
bae7f4ae | 1106 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1107 | return NULL; |
1108 | ||
d1c3fb1f NA |
1109 | /* |
1110 | * Assume we will successfully allocate the surplus page to | |
1111 | * prevent racing processes from causing the surplus to exceed | |
1112 | * overcommit | |
1113 | * | |
1114 | * This however introduces a different race, where a process B | |
1115 | * tries to grow the static hugepage pool while alloc_pages() is | |
1116 | * called by process A. B will only examine the per-node | |
1117 | * counters in determining if surplus huge pages can be | |
1118 | * converted to normal huge pages in adjust_pool_surplus(). A | |
1119 | * won't be able to increment the per-node counter, until the | |
1120 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
1121 | * no more huge pages can be converted from surplus to normal | |
1122 | * state (and doesn't try to convert again). Thus, we have a | |
1123 | * case where a surplus huge page exists, the pool is grown, and | |
1124 | * the surplus huge page still exists after, even though it | |
1125 | * should just have been converted to a normal huge page. This | |
1126 | * does not leak memory, though, as the hugepage will be freed | |
1127 | * once it is out of use. It also does not allow the counters to | |
1128 | * go out of whack in adjust_pool_surplus() as we don't modify | |
1129 | * the node values until we've gotten the hugepage and only the | |
1130 | * per-node value is checked there. | |
1131 | */ | |
1132 | spin_lock(&hugetlb_lock); | |
a5516438 | 1133 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
1134 | spin_unlock(&hugetlb_lock); |
1135 | return NULL; | |
1136 | } else { | |
a5516438 AK |
1137 | h->nr_huge_pages++; |
1138 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
1139 | } |
1140 | spin_unlock(&hugetlb_lock); | |
1141 | ||
bf50bab2 | 1142 | if (nid == NUMA_NO_NODE) |
86cdb465 | 1143 | page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP| |
bf50bab2 NH |
1144 | __GFP_REPEAT|__GFP_NOWARN, |
1145 | huge_page_order(h)); | |
1146 | else | |
1147 | page = alloc_pages_exact_node(nid, | |
86cdb465 | 1148 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
bf50bab2 | 1149 | __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); |
d1c3fb1f | 1150 | |
caff3a2c GS |
1151 | if (page && arch_prepare_hugepage(page)) { |
1152 | __free_pages(page, huge_page_order(h)); | |
ea5768c7 | 1153 | page = NULL; |
caff3a2c GS |
1154 | } |
1155 | ||
d1c3fb1f | 1156 | spin_lock(&hugetlb_lock); |
7893d1d5 | 1157 | if (page) { |
0edaecfa | 1158 | INIT_LIST_HEAD(&page->lru); |
bf50bab2 | 1159 | r_nid = page_to_nid(page); |
7893d1d5 | 1160 | set_compound_page_dtor(page, free_huge_page); |
9dd540e2 | 1161 | set_hugetlb_cgroup(page, NULL); |
d1c3fb1f NA |
1162 | /* |
1163 | * We incremented the global counters already | |
1164 | */ | |
bf50bab2 NH |
1165 | h->nr_huge_pages_node[r_nid]++; |
1166 | h->surplus_huge_pages_node[r_nid]++; | |
3b116300 | 1167 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 1168 | } else { |
a5516438 AK |
1169 | h->nr_huge_pages--; |
1170 | h->surplus_huge_pages--; | |
3b116300 | 1171 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 1172 | } |
d1c3fb1f | 1173 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1174 | |
1175 | return page; | |
1176 | } | |
1177 | ||
bf50bab2 NH |
1178 | /* |
1179 | * This allocation function is useful in the context where vma is irrelevant. | |
1180 | * E.g. soft-offlining uses this function because it only cares physical | |
1181 | * address of error page. | |
1182 | */ | |
1183 | struct page *alloc_huge_page_node(struct hstate *h, int nid) | |
1184 | { | |
4ef91848 | 1185 | struct page *page = NULL; |
bf50bab2 NH |
1186 | |
1187 | spin_lock(&hugetlb_lock); | |
4ef91848 JK |
1188 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
1189 | page = dequeue_huge_page_node(h, nid); | |
bf50bab2 NH |
1190 | spin_unlock(&hugetlb_lock); |
1191 | ||
94ae8ba7 | 1192 | if (!page) |
bf50bab2 NH |
1193 | page = alloc_buddy_huge_page(h, nid); |
1194 | ||
1195 | return page; | |
1196 | } | |
1197 | ||
e4e574b7 | 1198 | /* |
25985edc | 1199 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1200 | * of size 'delta'. |
1201 | */ | |
a5516438 | 1202 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1203 | { |
1204 | struct list_head surplus_list; | |
1205 | struct page *page, *tmp; | |
1206 | int ret, i; | |
1207 | int needed, allocated; | |
28073b02 | 1208 | bool alloc_ok = true; |
e4e574b7 | 1209 | |
a5516438 | 1210 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1211 | if (needed <= 0) { |
a5516438 | 1212 | h->resv_huge_pages += delta; |
e4e574b7 | 1213 | return 0; |
ac09b3a1 | 1214 | } |
e4e574b7 AL |
1215 | |
1216 | allocated = 0; | |
1217 | INIT_LIST_HEAD(&surplus_list); | |
1218 | ||
1219 | ret = -ENOMEM; | |
1220 | retry: | |
1221 | spin_unlock(&hugetlb_lock); | |
1222 | for (i = 0; i < needed; i++) { | |
bf50bab2 | 1223 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
28073b02 HD |
1224 | if (!page) { |
1225 | alloc_ok = false; | |
1226 | break; | |
1227 | } | |
e4e574b7 AL |
1228 | list_add(&page->lru, &surplus_list); |
1229 | } | |
28073b02 | 1230 | allocated += i; |
e4e574b7 AL |
1231 | |
1232 | /* | |
1233 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1234 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1235 | */ | |
1236 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1237 | needed = (h->resv_huge_pages + delta) - |
1238 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1239 | if (needed > 0) { |
1240 | if (alloc_ok) | |
1241 | goto retry; | |
1242 | /* | |
1243 | * We were not able to allocate enough pages to | |
1244 | * satisfy the entire reservation so we free what | |
1245 | * we've allocated so far. | |
1246 | */ | |
1247 | goto free; | |
1248 | } | |
e4e574b7 AL |
1249 | /* |
1250 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1251 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1252 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1253 | * allocator. Commit the entire reservation here to prevent another |
1254 | * process from stealing the pages as they are added to the pool but | |
1255 | * before they are reserved. | |
e4e574b7 AL |
1256 | */ |
1257 | needed += allocated; | |
a5516438 | 1258 | h->resv_huge_pages += delta; |
e4e574b7 | 1259 | ret = 0; |
a9869b83 | 1260 | |
19fc3f0a | 1261 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1262 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1263 | if ((--needed) < 0) |
1264 | break; | |
a9869b83 NH |
1265 | /* |
1266 | * This page is now managed by the hugetlb allocator and has | |
1267 | * no users -- drop the buddy allocator's reference. | |
1268 | */ | |
1269 | put_page_testzero(page); | |
309381fe | 1270 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 1271 | enqueue_huge_page(h, page); |
19fc3f0a | 1272 | } |
28073b02 | 1273 | free: |
b0365c8d | 1274 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1275 | |
1276 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1277 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1278 | put_page(page); | |
a9869b83 | 1279 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1280 | |
1281 | return ret; | |
1282 | } | |
1283 | ||
1284 | /* | |
1285 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
1286 | * allocated to satisfy the reservation must be explicitly freed if they were | |
1287 | * never used. | |
685f3457 | 1288 | * Called with hugetlb_lock held. |
e4e574b7 | 1289 | */ |
a5516438 AK |
1290 | static void return_unused_surplus_pages(struct hstate *h, |
1291 | unsigned long unused_resv_pages) | |
e4e574b7 | 1292 | { |
e4e574b7 AL |
1293 | unsigned long nr_pages; |
1294 | ||
ac09b3a1 | 1295 | /* Uncommit the reservation */ |
a5516438 | 1296 | h->resv_huge_pages -= unused_resv_pages; |
ac09b3a1 | 1297 | |
aa888a74 | 1298 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 1299 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1300 | return; |
1301 | ||
a5516438 | 1302 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1303 | |
685f3457 LS |
1304 | /* |
1305 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1306 | * evenly across all nodes with memory. Iterate across these nodes |
1307 | * until we can no longer free unreserved surplus pages. This occurs | |
1308 | * when the nodes with surplus pages have no free pages. | |
1309 | * free_pool_huge_page() will balance the the freed pages across the | |
1310 | * on-line nodes with memory and will handle the hstate accounting. | |
685f3457 LS |
1311 | */ |
1312 | while (nr_pages--) { | |
8cebfcd0 | 1313 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
685f3457 | 1314 | break; |
7848a4bf | 1315 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 AL |
1316 | } |
1317 | } | |
1318 | ||
c37f9fb1 AW |
1319 | /* |
1320 | * Determine if the huge page at addr within the vma has an associated | |
1321 | * reservation. Where it does not we will need to logically increase | |
90481622 DG |
1322 | * reservation and actually increase subpool usage before an allocation |
1323 | * can occur. Where any new reservation would be required the | |
1324 | * reservation change is prepared, but not committed. Once the page | |
1325 | * has been allocated from the subpool and instantiated the change should | |
1326 | * be committed via vma_commit_reservation. No action is required on | |
1327 | * failure. | |
c37f9fb1 | 1328 | */ |
e2f17d94 | 1329 | static long vma_needs_reservation(struct hstate *h, |
a5516438 | 1330 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 1331 | { |
4e35f483 JK |
1332 | struct resv_map *resv; |
1333 | pgoff_t idx; | |
1334 | long chg; | |
c37f9fb1 | 1335 | |
4e35f483 JK |
1336 | resv = vma_resv_map(vma); |
1337 | if (!resv) | |
84afd99b | 1338 | return 1; |
c37f9fb1 | 1339 | |
4e35f483 JK |
1340 | idx = vma_hugecache_offset(h, vma, addr); |
1341 | chg = region_chg(resv, idx, idx + 1); | |
84afd99b | 1342 | |
4e35f483 JK |
1343 | if (vma->vm_flags & VM_MAYSHARE) |
1344 | return chg; | |
1345 | else | |
1346 | return chg < 0 ? chg : 0; | |
c37f9fb1 | 1347 | } |
a5516438 AK |
1348 | static void vma_commit_reservation(struct hstate *h, |
1349 | struct vm_area_struct *vma, unsigned long addr) | |
c37f9fb1 | 1350 | { |
4e35f483 JK |
1351 | struct resv_map *resv; |
1352 | pgoff_t idx; | |
84afd99b | 1353 | |
4e35f483 JK |
1354 | resv = vma_resv_map(vma); |
1355 | if (!resv) | |
1356 | return; | |
84afd99b | 1357 | |
4e35f483 JK |
1358 | idx = vma_hugecache_offset(h, vma, addr); |
1359 | region_add(resv, idx, idx + 1); | |
c37f9fb1 AW |
1360 | } |
1361 | ||
a1e78772 | 1362 | static struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 1363 | unsigned long addr, int avoid_reserve) |
1da177e4 | 1364 | { |
90481622 | 1365 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 1366 | struct hstate *h = hstate_vma(vma); |
348ea204 | 1367 | struct page *page; |
e2f17d94 | 1368 | long chg; |
6d76dcf4 AK |
1369 | int ret, idx; |
1370 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 1371 | |
6d76dcf4 | 1372 | idx = hstate_index(h); |
a1e78772 | 1373 | /* |
90481622 DG |
1374 | * Processes that did not create the mapping will have no |
1375 | * reserves and will not have accounted against subpool | |
1376 | * limit. Check that the subpool limit can be made before | |
1377 | * satisfying the allocation MAP_NORESERVE mappings may also | |
1378 | * need pages and subpool limit allocated allocated if no reserve | |
1379 | * mapping overlaps. | |
a1e78772 | 1380 | */ |
a5516438 | 1381 | chg = vma_needs_reservation(h, vma, addr); |
c37f9fb1 | 1382 | if (chg < 0) |
76dcee75 | 1383 | return ERR_PTR(-ENOMEM); |
8bb3f12e JK |
1384 | if (chg || avoid_reserve) |
1385 | if (hugepage_subpool_get_pages(spool, 1)) | |
76dcee75 | 1386 | return ERR_PTR(-ENOSPC); |
1da177e4 | 1387 | |
6d76dcf4 AK |
1388 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
1389 | if (ret) { | |
8bb3f12e JK |
1390 | if (chg || avoid_reserve) |
1391 | hugepage_subpool_put_pages(spool, 1); | |
6d76dcf4 AK |
1392 | return ERR_PTR(-ENOSPC); |
1393 | } | |
1da177e4 | 1394 | spin_lock(&hugetlb_lock); |
af0ed73e | 1395 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg); |
81a6fcae | 1396 | if (!page) { |
94ae8ba7 | 1397 | spin_unlock(&hugetlb_lock); |
bf50bab2 | 1398 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
68842c9b | 1399 | if (!page) { |
6d76dcf4 AK |
1400 | hugetlb_cgroup_uncharge_cgroup(idx, |
1401 | pages_per_huge_page(h), | |
1402 | h_cg); | |
8bb3f12e JK |
1403 | if (chg || avoid_reserve) |
1404 | hugepage_subpool_put_pages(spool, 1); | |
76dcee75 | 1405 | return ERR_PTR(-ENOSPC); |
68842c9b | 1406 | } |
79dbb236 AK |
1407 | spin_lock(&hugetlb_lock); |
1408 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 1409 | /* Fall through */ |
68842c9b | 1410 | } |
81a6fcae JK |
1411 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
1412 | spin_unlock(&hugetlb_lock); | |
348ea204 | 1413 | |
90481622 | 1414 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 1415 | |
a5516438 | 1416 | vma_commit_reservation(h, vma, addr); |
90d8b7e6 | 1417 | return page; |
b45b5bd6 DG |
1418 | } |
1419 | ||
74060e4d NH |
1420 | /* |
1421 | * alloc_huge_page()'s wrapper which simply returns the page if allocation | |
1422 | * succeeds, otherwise NULL. This function is called from new_vma_page(), | |
1423 | * where no ERR_VALUE is expected to be returned. | |
1424 | */ | |
1425 | struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, | |
1426 | unsigned long addr, int avoid_reserve) | |
1427 | { | |
1428 | struct page *page = alloc_huge_page(vma, addr, avoid_reserve); | |
1429 | if (IS_ERR(page)) | |
1430 | page = NULL; | |
1431 | return page; | |
1432 | } | |
1433 | ||
91f47662 | 1434 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
aa888a74 AK |
1435 | { |
1436 | struct huge_bootmem_page *m; | |
b2261026 | 1437 | int nr_nodes, node; |
aa888a74 | 1438 | |
b2261026 | 1439 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
1440 | void *addr; |
1441 | ||
8b89a116 GS |
1442 | addr = memblock_virt_alloc_try_nid_nopanic( |
1443 | huge_page_size(h), huge_page_size(h), | |
1444 | 0, BOOTMEM_ALLOC_ACCESSIBLE, node); | |
aa888a74 AK |
1445 | if (addr) { |
1446 | /* | |
1447 | * Use the beginning of the huge page to store the | |
1448 | * huge_bootmem_page struct (until gather_bootmem | |
1449 | * puts them into the mem_map). | |
1450 | */ | |
1451 | m = addr; | |
91f47662 | 1452 | goto found; |
aa888a74 | 1453 | } |
aa888a74 AK |
1454 | } |
1455 | return 0; | |
1456 | ||
1457 | found: | |
1458 | BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); | |
1459 | /* Put them into a private list first because mem_map is not up yet */ | |
1460 | list_add(&m->list, &huge_boot_pages); | |
1461 | m->hstate = h; | |
1462 | return 1; | |
1463 | } | |
1464 | ||
f412c97a | 1465 | static void __init prep_compound_huge_page(struct page *page, int order) |
18229df5 AW |
1466 | { |
1467 | if (unlikely(order > (MAX_ORDER - 1))) | |
1468 | prep_compound_gigantic_page(page, order); | |
1469 | else | |
1470 | prep_compound_page(page, order); | |
1471 | } | |
1472 | ||
aa888a74 AK |
1473 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
1474 | static void __init gather_bootmem_prealloc(void) | |
1475 | { | |
1476 | struct huge_bootmem_page *m; | |
1477 | ||
1478 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 1479 | struct hstate *h = m->hstate; |
ee8f248d BB |
1480 | struct page *page; |
1481 | ||
1482 | #ifdef CONFIG_HIGHMEM | |
1483 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
8b89a116 GS |
1484 | memblock_free_late(__pa(m), |
1485 | sizeof(struct huge_bootmem_page)); | |
ee8f248d BB |
1486 | #else |
1487 | page = virt_to_page(m); | |
1488 | #endif | |
aa888a74 | 1489 | WARN_ON(page_count(page) != 1); |
18229df5 | 1490 | prep_compound_huge_page(page, h->order); |
ef5a22be | 1491 | WARN_ON(PageReserved(page)); |
aa888a74 | 1492 | prep_new_huge_page(h, page, page_to_nid(page)); |
b0320c7b RA |
1493 | /* |
1494 | * If we had gigantic hugepages allocated at boot time, we need | |
1495 | * to restore the 'stolen' pages to totalram_pages in order to | |
1496 | * fix confusing memory reports from free(1) and another | |
1497 | * side-effects, like CommitLimit going negative. | |
1498 | */ | |
bae7f4ae | 1499 | if (hstate_is_gigantic(h)) |
3dcc0571 | 1500 | adjust_managed_page_count(page, 1 << h->order); |
aa888a74 AK |
1501 | } |
1502 | } | |
1503 | ||
8faa8b07 | 1504 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
1505 | { |
1506 | unsigned long i; | |
a5516438 | 1507 | |
e5ff2159 | 1508 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 1509 | if (hstate_is_gigantic(h)) { |
aa888a74 AK |
1510 | if (!alloc_bootmem_huge_page(h)) |
1511 | break; | |
9b5e5d0f | 1512 | } else if (!alloc_fresh_huge_page(h, |
8cebfcd0 | 1513 | &node_states[N_MEMORY])) |
1da177e4 | 1514 | break; |
1da177e4 | 1515 | } |
8faa8b07 | 1516 | h->max_huge_pages = i; |
e5ff2159 AK |
1517 | } |
1518 | ||
1519 | static void __init hugetlb_init_hstates(void) | |
1520 | { | |
1521 | struct hstate *h; | |
1522 | ||
1523 | for_each_hstate(h) { | |
8faa8b07 | 1524 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 1525 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 1526 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 AK |
1527 | } |
1528 | } | |
1529 | ||
4abd32db AK |
1530 | static char * __init memfmt(char *buf, unsigned long n) |
1531 | { | |
1532 | if (n >= (1UL << 30)) | |
1533 | sprintf(buf, "%lu GB", n >> 30); | |
1534 | else if (n >= (1UL << 20)) | |
1535 | sprintf(buf, "%lu MB", n >> 20); | |
1536 | else | |
1537 | sprintf(buf, "%lu KB", n >> 10); | |
1538 | return buf; | |
1539 | } | |
1540 | ||
e5ff2159 AK |
1541 | static void __init report_hugepages(void) |
1542 | { | |
1543 | struct hstate *h; | |
1544 | ||
1545 | for_each_hstate(h) { | |
4abd32db | 1546 | char buf[32]; |
ffb22af5 | 1547 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
4abd32db AK |
1548 | memfmt(buf, huge_page_size(h)), |
1549 | h->free_huge_pages); | |
e5ff2159 AK |
1550 | } |
1551 | } | |
1552 | ||
1da177e4 | 1553 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
1554 | static void try_to_free_low(struct hstate *h, unsigned long count, |
1555 | nodemask_t *nodes_allowed) | |
1da177e4 | 1556 | { |
4415cc8d CL |
1557 | int i; |
1558 | ||
bae7f4ae | 1559 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1560 | return; |
1561 | ||
6ae11b27 | 1562 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 1563 | struct page *page, *next; |
a5516438 AK |
1564 | struct list_head *freel = &h->hugepage_freelists[i]; |
1565 | list_for_each_entry_safe(page, next, freel, lru) { | |
1566 | if (count >= h->nr_huge_pages) | |
6b0c880d | 1567 | return; |
1da177e4 LT |
1568 | if (PageHighMem(page)) |
1569 | continue; | |
1570 | list_del(&page->lru); | |
e5ff2159 | 1571 | update_and_free_page(h, page); |
a5516438 AK |
1572 | h->free_huge_pages--; |
1573 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
1574 | } |
1575 | } | |
1576 | } | |
1577 | #else | |
6ae11b27 LS |
1578 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
1579 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
1580 | { |
1581 | } | |
1582 | #endif | |
1583 | ||
20a0307c WF |
1584 | /* |
1585 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
1586 | * balanced by operating on them in a round-robin fashion. | |
1587 | * Returns 1 if an adjustment was made. | |
1588 | */ | |
6ae11b27 LS |
1589 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
1590 | int delta) | |
20a0307c | 1591 | { |
b2261026 | 1592 | int nr_nodes, node; |
20a0307c WF |
1593 | |
1594 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 1595 | |
b2261026 JK |
1596 | if (delta < 0) { |
1597 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1598 | if (h->surplus_huge_pages_node[node]) | |
1599 | goto found; | |
e8c5c824 | 1600 | } |
b2261026 JK |
1601 | } else { |
1602 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
1603 | if (h->surplus_huge_pages_node[node] < | |
1604 | h->nr_huge_pages_node[node]) | |
1605 | goto found; | |
e8c5c824 | 1606 | } |
b2261026 JK |
1607 | } |
1608 | return 0; | |
20a0307c | 1609 | |
b2261026 JK |
1610 | found: |
1611 | h->surplus_huge_pages += delta; | |
1612 | h->surplus_huge_pages_node[node] += delta; | |
1613 | return 1; | |
20a0307c WF |
1614 | } |
1615 | ||
a5516438 | 1616 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
1617 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
1618 | nodemask_t *nodes_allowed) | |
1da177e4 | 1619 | { |
7893d1d5 | 1620 | unsigned long min_count, ret; |
1da177e4 | 1621 | |
944d9fec | 1622 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
aa888a74 AK |
1623 | return h->max_huge_pages; |
1624 | ||
7893d1d5 AL |
1625 | /* |
1626 | * Increase the pool size | |
1627 | * First take pages out of surplus state. Then make up the | |
1628 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f NA |
1629 | * |
1630 | * We might race with alloc_buddy_huge_page() here and be unable | |
1631 | * to convert a surplus huge page to a normal huge page. That is | |
1632 | * not critical, though, it just means the overall size of the | |
1633 | * pool might be one hugepage larger than it needs to be, but | |
1634 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 1635 | */ |
1da177e4 | 1636 | spin_lock(&hugetlb_lock); |
a5516438 | 1637 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 1638 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
1639 | break; |
1640 | } | |
1641 | ||
a5516438 | 1642 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
1643 | /* |
1644 | * If this allocation races such that we no longer need the | |
1645 | * page, free_huge_page will handle it by freeing the page | |
1646 | * and reducing the surplus. | |
1647 | */ | |
1648 | spin_unlock(&hugetlb_lock); | |
944d9fec LC |
1649 | if (hstate_is_gigantic(h)) |
1650 | ret = alloc_fresh_gigantic_page(h, nodes_allowed); | |
1651 | else | |
1652 | ret = alloc_fresh_huge_page(h, nodes_allowed); | |
7893d1d5 AL |
1653 | spin_lock(&hugetlb_lock); |
1654 | if (!ret) | |
1655 | goto out; | |
1656 | ||
536240f2 MG |
1657 | /* Bail for signals. Probably ctrl-c from user */ |
1658 | if (signal_pending(current)) | |
1659 | goto out; | |
7893d1d5 | 1660 | } |
7893d1d5 AL |
1661 | |
1662 | /* | |
1663 | * Decrease the pool size | |
1664 | * First return free pages to the buddy allocator (being careful | |
1665 | * to keep enough around to satisfy reservations). Then place | |
1666 | * pages into surplus state as needed so the pool will shrink | |
1667 | * to the desired size as pages become free. | |
d1c3fb1f NA |
1668 | * |
1669 | * By placing pages into the surplus state independent of the | |
1670 | * overcommit value, we are allowing the surplus pool size to | |
1671 | * exceed overcommit. There are few sane options here. Since | |
1672 | * alloc_buddy_huge_page() is checking the global counter, | |
1673 | * though, we'll note that we're not allowed to exceed surplus | |
1674 | * and won't grow the pool anywhere else. Not until one of the | |
1675 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 1676 | */ |
a5516438 | 1677 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 1678 | min_count = max(count, min_count); |
6ae11b27 | 1679 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 1680 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 1681 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 1682 | break; |
55f67141 | 1683 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 1684 | } |
a5516438 | 1685 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 1686 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
1687 | break; |
1688 | } | |
1689 | out: | |
a5516438 | 1690 | ret = persistent_huge_pages(h); |
1da177e4 | 1691 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 1692 | return ret; |
1da177e4 LT |
1693 | } |
1694 | ||
a3437870 NA |
1695 | #define HSTATE_ATTR_RO(_name) \ |
1696 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
1697 | ||
1698 | #define HSTATE_ATTR(_name) \ | |
1699 | static struct kobj_attribute _name##_attr = \ | |
1700 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
1701 | ||
1702 | static struct kobject *hugepages_kobj; | |
1703 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1704 | ||
9a305230 LS |
1705 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
1706 | ||
1707 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
1708 | { |
1709 | int i; | |
9a305230 | 1710 | |
a3437870 | 1711 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
1712 | if (hstate_kobjs[i] == kobj) { |
1713 | if (nidp) | |
1714 | *nidp = NUMA_NO_NODE; | |
a3437870 | 1715 | return &hstates[i]; |
9a305230 LS |
1716 | } |
1717 | ||
1718 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
1719 | } |
1720 | ||
06808b08 | 1721 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
1722 | struct kobj_attribute *attr, char *buf) |
1723 | { | |
9a305230 LS |
1724 | struct hstate *h; |
1725 | unsigned long nr_huge_pages; | |
1726 | int nid; | |
1727 | ||
1728 | h = kobj_to_hstate(kobj, &nid); | |
1729 | if (nid == NUMA_NO_NODE) | |
1730 | nr_huge_pages = h->nr_huge_pages; | |
1731 | else | |
1732 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
1733 | ||
1734 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 1735 | } |
adbe8726 | 1736 | |
06808b08 LS |
1737 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
1738 | struct kobject *kobj, struct kobj_attribute *attr, | |
1739 | const char *buf, size_t len) | |
a3437870 NA |
1740 | { |
1741 | int err; | |
9a305230 | 1742 | int nid; |
06808b08 | 1743 | unsigned long count; |
9a305230 | 1744 | struct hstate *h; |
bad44b5b | 1745 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 1746 | |
3dbb95f7 | 1747 | err = kstrtoul(buf, 10, &count); |
73ae31e5 | 1748 | if (err) |
adbe8726 | 1749 | goto out; |
a3437870 | 1750 | |
9a305230 | 1751 | h = kobj_to_hstate(kobj, &nid); |
944d9fec | 1752 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) { |
adbe8726 EM |
1753 | err = -EINVAL; |
1754 | goto out; | |
1755 | } | |
1756 | ||
9a305230 LS |
1757 | if (nid == NUMA_NO_NODE) { |
1758 | /* | |
1759 | * global hstate attribute | |
1760 | */ | |
1761 | if (!(obey_mempolicy && | |
1762 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
1763 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 1764 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 LS |
1765 | } |
1766 | } else if (nodes_allowed) { | |
1767 | /* | |
1768 | * per node hstate attribute: adjust count to global, | |
1769 | * but restrict alloc/free to the specified node. | |
1770 | */ | |
1771 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
1772 | init_nodemask_of_node(nodes_allowed, nid); | |
1773 | } else | |
8cebfcd0 | 1774 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 | 1775 | |
06808b08 | 1776 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 1777 | |
8cebfcd0 | 1778 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
1779 | NODEMASK_FREE(nodes_allowed); |
1780 | ||
1781 | return len; | |
adbe8726 EM |
1782 | out: |
1783 | NODEMASK_FREE(nodes_allowed); | |
1784 | return err; | |
06808b08 LS |
1785 | } |
1786 | ||
1787 | static ssize_t nr_hugepages_show(struct kobject *kobj, | |
1788 | struct kobj_attribute *attr, char *buf) | |
1789 | { | |
1790 | return nr_hugepages_show_common(kobj, attr, buf); | |
1791 | } | |
1792 | ||
1793 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
1794 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1795 | { | |
1796 | return nr_hugepages_store_common(false, kobj, attr, buf, len); | |
a3437870 NA |
1797 | } |
1798 | HSTATE_ATTR(nr_hugepages); | |
1799 | ||
06808b08 LS |
1800 | #ifdef CONFIG_NUMA |
1801 | ||
1802 | /* | |
1803 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
1804 | * huge page alloc/free. | |
1805 | */ | |
1806 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
1807 | struct kobj_attribute *attr, char *buf) | |
1808 | { | |
1809 | return nr_hugepages_show_common(kobj, attr, buf); | |
1810 | } | |
1811 | ||
1812 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
1813 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1814 | { | |
1815 | return nr_hugepages_store_common(true, kobj, attr, buf, len); | |
1816 | } | |
1817 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
1818 | #endif | |
1819 | ||
1820 | ||
a3437870 NA |
1821 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
1822 | struct kobj_attribute *attr, char *buf) | |
1823 | { | |
9a305230 | 1824 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1825 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
1826 | } | |
adbe8726 | 1827 | |
a3437870 NA |
1828 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
1829 | struct kobj_attribute *attr, const char *buf, size_t count) | |
1830 | { | |
1831 | int err; | |
1832 | unsigned long input; | |
9a305230 | 1833 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 1834 | |
bae7f4ae | 1835 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
1836 | return -EINVAL; |
1837 | ||
3dbb95f7 | 1838 | err = kstrtoul(buf, 10, &input); |
a3437870 | 1839 | if (err) |
73ae31e5 | 1840 | return err; |
a3437870 NA |
1841 | |
1842 | spin_lock(&hugetlb_lock); | |
1843 | h->nr_overcommit_huge_pages = input; | |
1844 | spin_unlock(&hugetlb_lock); | |
1845 | ||
1846 | return count; | |
1847 | } | |
1848 | HSTATE_ATTR(nr_overcommit_hugepages); | |
1849 | ||
1850 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
1851 | struct kobj_attribute *attr, char *buf) | |
1852 | { | |
9a305230 LS |
1853 | struct hstate *h; |
1854 | unsigned long free_huge_pages; | |
1855 | int nid; | |
1856 | ||
1857 | h = kobj_to_hstate(kobj, &nid); | |
1858 | if (nid == NUMA_NO_NODE) | |
1859 | free_huge_pages = h->free_huge_pages; | |
1860 | else | |
1861 | free_huge_pages = h->free_huge_pages_node[nid]; | |
1862 | ||
1863 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
1864 | } |
1865 | HSTATE_ATTR_RO(free_hugepages); | |
1866 | ||
1867 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
1868 | struct kobj_attribute *attr, char *buf) | |
1869 | { | |
9a305230 | 1870 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1871 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
1872 | } | |
1873 | HSTATE_ATTR_RO(resv_hugepages); | |
1874 | ||
1875 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
1876 | struct kobj_attribute *attr, char *buf) | |
1877 | { | |
9a305230 LS |
1878 | struct hstate *h; |
1879 | unsigned long surplus_huge_pages; | |
1880 | int nid; | |
1881 | ||
1882 | h = kobj_to_hstate(kobj, &nid); | |
1883 | if (nid == NUMA_NO_NODE) | |
1884 | surplus_huge_pages = h->surplus_huge_pages; | |
1885 | else | |
1886 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
1887 | ||
1888 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
1889 | } |
1890 | HSTATE_ATTR_RO(surplus_hugepages); | |
1891 | ||
1892 | static struct attribute *hstate_attrs[] = { | |
1893 | &nr_hugepages_attr.attr, | |
1894 | &nr_overcommit_hugepages_attr.attr, | |
1895 | &free_hugepages_attr.attr, | |
1896 | &resv_hugepages_attr.attr, | |
1897 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
1898 | #ifdef CONFIG_NUMA |
1899 | &nr_hugepages_mempolicy_attr.attr, | |
1900 | #endif | |
a3437870 NA |
1901 | NULL, |
1902 | }; | |
1903 | ||
1904 | static struct attribute_group hstate_attr_group = { | |
1905 | .attrs = hstate_attrs, | |
1906 | }; | |
1907 | ||
094e9539 JM |
1908 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
1909 | struct kobject **hstate_kobjs, | |
1910 | struct attribute_group *hstate_attr_group) | |
a3437870 NA |
1911 | { |
1912 | int retval; | |
972dc4de | 1913 | int hi = hstate_index(h); |
a3437870 | 1914 | |
9a305230 LS |
1915 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
1916 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
1917 | return -ENOMEM; |
1918 | ||
9a305230 | 1919 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 1920 | if (retval) |
9a305230 | 1921 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
1922 | |
1923 | return retval; | |
1924 | } | |
1925 | ||
1926 | static void __init hugetlb_sysfs_init(void) | |
1927 | { | |
1928 | struct hstate *h; | |
1929 | int err; | |
1930 | ||
1931 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
1932 | if (!hugepages_kobj) | |
1933 | return; | |
1934 | ||
1935 | for_each_hstate(h) { | |
9a305230 LS |
1936 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
1937 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 1938 | if (err) |
ffb22af5 | 1939 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
1940 | } |
1941 | } | |
1942 | ||
9a305230 LS |
1943 | #ifdef CONFIG_NUMA |
1944 | ||
1945 | /* | |
1946 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
1947 | * with node devices in node_devices[] using a parallel array. The array |
1948 | * index of a node device or _hstate == node id. | |
1949 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
1950 | * the base kernel, on the hugetlb module. |
1951 | */ | |
1952 | struct node_hstate { | |
1953 | struct kobject *hugepages_kobj; | |
1954 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1955 | }; | |
1956 | struct node_hstate node_hstates[MAX_NUMNODES]; | |
1957 | ||
1958 | /* | |
10fbcf4c | 1959 | * A subset of global hstate attributes for node devices |
9a305230 LS |
1960 | */ |
1961 | static struct attribute *per_node_hstate_attrs[] = { | |
1962 | &nr_hugepages_attr.attr, | |
1963 | &free_hugepages_attr.attr, | |
1964 | &surplus_hugepages_attr.attr, | |
1965 | NULL, | |
1966 | }; | |
1967 | ||
1968 | static struct attribute_group per_node_hstate_attr_group = { | |
1969 | .attrs = per_node_hstate_attrs, | |
1970 | }; | |
1971 | ||
1972 | /* | |
10fbcf4c | 1973 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
1974 | * Returns node id via non-NULL nidp. |
1975 | */ | |
1976 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
1977 | { | |
1978 | int nid; | |
1979 | ||
1980 | for (nid = 0; nid < nr_node_ids; nid++) { | |
1981 | struct node_hstate *nhs = &node_hstates[nid]; | |
1982 | int i; | |
1983 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
1984 | if (nhs->hstate_kobjs[i] == kobj) { | |
1985 | if (nidp) | |
1986 | *nidp = nid; | |
1987 | return &hstates[i]; | |
1988 | } | |
1989 | } | |
1990 | ||
1991 | BUG(); | |
1992 | return NULL; | |
1993 | } | |
1994 | ||
1995 | /* | |
10fbcf4c | 1996 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
1997 | * No-op if no hstate attributes attached. |
1998 | */ | |
3cd8b44f | 1999 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
2000 | { |
2001 | struct hstate *h; | |
10fbcf4c | 2002 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2003 | |
2004 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 2005 | return; /* no hstate attributes */ |
9a305230 | 2006 | |
972dc4de AK |
2007 | for_each_hstate(h) { |
2008 | int idx = hstate_index(h); | |
2009 | if (nhs->hstate_kobjs[idx]) { | |
2010 | kobject_put(nhs->hstate_kobjs[idx]); | |
2011 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 2012 | } |
972dc4de | 2013 | } |
9a305230 LS |
2014 | |
2015 | kobject_put(nhs->hugepages_kobj); | |
2016 | nhs->hugepages_kobj = NULL; | |
2017 | } | |
2018 | ||
2019 | /* | |
10fbcf4c | 2020 | * hugetlb module exit: unregister hstate attributes from node devices |
9a305230 LS |
2021 | * that have them. |
2022 | */ | |
2023 | static void hugetlb_unregister_all_nodes(void) | |
2024 | { | |
2025 | int nid; | |
2026 | ||
2027 | /* | |
10fbcf4c | 2028 | * disable node device registrations. |
9a305230 LS |
2029 | */ |
2030 | register_hugetlbfs_with_node(NULL, NULL); | |
2031 | ||
2032 | /* | |
2033 | * remove hstate attributes from any nodes that have them. | |
2034 | */ | |
2035 | for (nid = 0; nid < nr_node_ids; nid++) | |
8732794b | 2036 | hugetlb_unregister_node(node_devices[nid]); |
9a305230 LS |
2037 | } |
2038 | ||
2039 | /* | |
10fbcf4c | 2040 | * Register hstate attributes for a single node device. |
9a305230 LS |
2041 | * No-op if attributes already registered. |
2042 | */ | |
3cd8b44f | 2043 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
2044 | { |
2045 | struct hstate *h; | |
10fbcf4c | 2046 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2047 | int err; |
2048 | ||
2049 | if (nhs->hugepages_kobj) | |
2050 | return; /* already allocated */ | |
2051 | ||
2052 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 2053 | &node->dev.kobj); |
9a305230 LS |
2054 | if (!nhs->hugepages_kobj) |
2055 | return; | |
2056 | ||
2057 | for_each_hstate(h) { | |
2058 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
2059 | nhs->hstate_kobjs, | |
2060 | &per_node_hstate_attr_group); | |
2061 | if (err) { | |
ffb22af5 AM |
2062 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
2063 | h->name, node->dev.id); | |
9a305230 LS |
2064 | hugetlb_unregister_node(node); |
2065 | break; | |
2066 | } | |
2067 | } | |
2068 | } | |
2069 | ||
2070 | /* | |
9b5e5d0f | 2071 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
2072 | * devices of nodes that have memory. All on-line nodes should have |
2073 | * registered their associated device by this time. | |
9a305230 LS |
2074 | */ |
2075 | static void hugetlb_register_all_nodes(void) | |
2076 | { | |
2077 | int nid; | |
2078 | ||
8cebfcd0 | 2079 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 2080 | struct node *node = node_devices[nid]; |
10fbcf4c | 2081 | if (node->dev.id == nid) |
9a305230 LS |
2082 | hugetlb_register_node(node); |
2083 | } | |
2084 | ||
2085 | /* | |
10fbcf4c | 2086 | * Let the node device driver know we're here so it can |
9a305230 LS |
2087 | * [un]register hstate attributes on node hotplug. |
2088 | */ | |
2089 | register_hugetlbfs_with_node(hugetlb_register_node, | |
2090 | hugetlb_unregister_node); | |
2091 | } | |
2092 | #else /* !CONFIG_NUMA */ | |
2093 | ||
2094 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2095 | { | |
2096 | BUG(); | |
2097 | if (nidp) | |
2098 | *nidp = -1; | |
2099 | return NULL; | |
2100 | } | |
2101 | ||
2102 | static void hugetlb_unregister_all_nodes(void) { } | |
2103 | ||
2104 | static void hugetlb_register_all_nodes(void) { } | |
2105 | ||
2106 | #endif | |
2107 | ||
a3437870 NA |
2108 | static void __exit hugetlb_exit(void) |
2109 | { | |
2110 | struct hstate *h; | |
2111 | ||
9a305230 LS |
2112 | hugetlb_unregister_all_nodes(); |
2113 | ||
a3437870 | 2114 | for_each_hstate(h) { |
972dc4de | 2115 | kobject_put(hstate_kobjs[hstate_index(h)]); |
a3437870 NA |
2116 | } |
2117 | ||
2118 | kobject_put(hugepages_kobj); | |
8382d914 | 2119 | kfree(htlb_fault_mutex_table); |
a3437870 NA |
2120 | } |
2121 | module_exit(hugetlb_exit); | |
2122 | ||
2123 | static int __init hugetlb_init(void) | |
2124 | { | |
8382d914 DB |
2125 | int i; |
2126 | ||
457c1b27 | 2127 | if (!hugepages_supported()) |
0ef89d25 | 2128 | return 0; |
a3437870 | 2129 | |
e11bfbfc NP |
2130 | if (!size_to_hstate(default_hstate_size)) { |
2131 | default_hstate_size = HPAGE_SIZE; | |
2132 | if (!size_to_hstate(default_hstate_size)) | |
2133 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 2134 | } |
972dc4de | 2135 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
e11bfbfc NP |
2136 | if (default_hstate_max_huge_pages) |
2137 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
a3437870 NA |
2138 | |
2139 | hugetlb_init_hstates(); | |
aa888a74 | 2140 | gather_bootmem_prealloc(); |
a3437870 NA |
2141 | report_hugepages(); |
2142 | ||
2143 | hugetlb_sysfs_init(); | |
9a305230 | 2144 | hugetlb_register_all_nodes(); |
7179e7bf | 2145 | hugetlb_cgroup_file_init(); |
9a305230 | 2146 | |
8382d914 DB |
2147 | #ifdef CONFIG_SMP |
2148 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
2149 | #else | |
2150 | num_fault_mutexes = 1; | |
2151 | #endif | |
2152 | htlb_fault_mutex_table = | |
2153 | kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); | |
2154 | BUG_ON(!htlb_fault_mutex_table); | |
2155 | ||
2156 | for (i = 0; i < num_fault_mutexes; i++) | |
2157 | mutex_init(&htlb_fault_mutex_table[i]); | |
a3437870 NA |
2158 | return 0; |
2159 | } | |
2160 | module_init(hugetlb_init); | |
2161 | ||
2162 | /* Should be called on processing a hugepagesz=... option */ | |
2163 | void __init hugetlb_add_hstate(unsigned order) | |
2164 | { | |
2165 | struct hstate *h; | |
8faa8b07 AK |
2166 | unsigned long i; |
2167 | ||
a3437870 | 2168 | if (size_to_hstate(PAGE_SIZE << order)) { |
ffb22af5 | 2169 | pr_warning("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
2170 | return; |
2171 | } | |
47d38344 | 2172 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 2173 | BUG_ON(order == 0); |
47d38344 | 2174 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
2175 | h->order = order; |
2176 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
2177 | h->nr_huge_pages = 0; |
2178 | h->free_huge_pages = 0; | |
2179 | for (i = 0; i < MAX_NUMNODES; ++i) | |
2180 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 2181 | INIT_LIST_HEAD(&h->hugepage_activelist); |
8cebfcd0 LJ |
2182 | h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); |
2183 | h->next_nid_to_free = first_node(node_states[N_MEMORY]); | |
a3437870 NA |
2184 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
2185 | huge_page_size(h)/1024); | |
8faa8b07 | 2186 | |
a3437870 NA |
2187 | parsed_hstate = h; |
2188 | } | |
2189 | ||
e11bfbfc | 2190 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
2191 | { |
2192 | unsigned long *mhp; | |
8faa8b07 | 2193 | static unsigned long *last_mhp; |
a3437870 NA |
2194 | |
2195 | /* | |
47d38344 | 2196 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
2197 | * so this hugepages= parameter goes to the "default hstate". |
2198 | */ | |
47d38344 | 2199 | if (!hugetlb_max_hstate) |
a3437870 NA |
2200 | mhp = &default_hstate_max_huge_pages; |
2201 | else | |
2202 | mhp = &parsed_hstate->max_huge_pages; | |
2203 | ||
8faa8b07 | 2204 | if (mhp == last_mhp) { |
ffb22af5 AM |
2205 | pr_warning("hugepages= specified twice without " |
2206 | "interleaving hugepagesz=, ignoring\n"); | |
8faa8b07 AK |
2207 | return 1; |
2208 | } | |
2209 | ||
a3437870 NA |
2210 | if (sscanf(s, "%lu", mhp) <= 0) |
2211 | *mhp = 0; | |
2212 | ||
8faa8b07 AK |
2213 | /* |
2214 | * Global state is always initialized later in hugetlb_init. | |
2215 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
2216 | * use the bootmem allocator. | |
2217 | */ | |
47d38344 | 2218 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
2219 | hugetlb_hstate_alloc_pages(parsed_hstate); |
2220 | ||
2221 | last_mhp = mhp; | |
2222 | ||
a3437870 NA |
2223 | return 1; |
2224 | } | |
e11bfbfc NP |
2225 | __setup("hugepages=", hugetlb_nrpages_setup); |
2226 | ||
2227 | static int __init hugetlb_default_setup(char *s) | |
2228 | { | |
2229 | default_hstate_size = memparse(s, &s); | |
2230 | return 1; | |
2231 | } | |
2232 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 2233 | |
8a213460 NA |
2234 | static unsigned int cpuset_mems_nr(unsigned int *array) |
2235 | { | |
2236 | int node; | |
2237 | unsigned int nr = 0; | |
2238 | ||
2239 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
2240 | nr += array[node]; | |
2241 | ||
2242 | return nr; | |
2243 | } | |
2244 | ||
2245 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
2246 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
2247 | struct ctl_table *table, int write, | |
2248 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 2249 | { |
e5ff2159 AK |
2250 | struct hstate *h = &default_hstate; |
2251 | unsigned long tmp; | |
08d4a246 | 2252 | int ret; |
e5ff2159 | 2253 | |
457c1b27 NA |
2254 | if (!hugepages_supported()) |
2255 | return -ENOTSUPP; | |
2256 | ||
c033a93c | 2257 | tmp = h->max_huge_pages; |
e5ff2159 | 2258 | |
944d9fec | 2259 | if (write && hstate_is_gigantic(h) && !gigantic_page_supported()) |
adbe8726 EM |
2260 | return -EINVAL; |
2261 | ||
e5ff2159 AK |
2262 | table->data = &tmp; |
2263 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2264 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2265 | if (ret) | |
2266 | goto out; | |
e5ff2159 | 2267 | |
06808b08 | 2268 | if (write) { |
bad44b5b DR |
2269 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, |
2270 | GFP_KERNEL | __GFP_NORETRY); | |
06808b08 LS |
2271 | if (!(obey_mempolicy && |
2272 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
2273 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 2274 | nodes_allowed = &node_states[N_MEMORY]; |
06808b08 LS |
2275 | } |
2276 | h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); | |
2277 | ||
8cebfcd0 | 2278 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
2279 | NODEMASK_FREE(nodes_allowed); |
2280 | } | |
08d4a246 MH |
2281 | out: |
2282 | return ret; | |
1da177e4 | 2283 | } |
396faf03 | 2284 | |
06808b08 LS |
2285 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2286 | void __user *buffer, size_t *length, loff_t *ppos) | |
2287 | { | |
2288 | ||
2289 | return hugetlb_sysctl_handler_common(false, table, write, | |
2290 | buffer, length, ppos); | |
2291 | } | |
2292 | ||
2293 | #ifdef CONFIG_NUMA | |
2294 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2295 | void __user *buffer, size_t *length, loff_t *ppos) | |
2296 | { | |
2297 | return hugetlb_sysctl_handler_common(true, table, write, | |
2298 | buffer, length, ppos); | |
2299 | } | |
2300 | #endif /* CONFIG_NUMA */ | |
2301 | ||
a3d0c6aa | 2302 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 2303 | void __user *buffer, |
a3d0c6aa NA |
2304 | size_t *length, loff_t *ppos) |
2305 | { | |
a5516438 | 2306 | struct hstate *h = &default_hstate; |
e5ff2159 | 2307 | unsigned long tmp; |
08d4a246 | 2308 | int ret; |
e5ff2159 | 2309 | |
457c1b27 NA |
2310 | if (!hugepages_supported()) |
2311 | return -ENOTSUPP; | |
2312 | ||
c033a93c | 2313 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 2314 | |
bae7f4ae | 2315 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
2316 | return -EINVAL; |
2317 | ||
e5ff2159 AK |
2318 | table->data = &tmp; |
2319 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2320 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2321 | if (ret) | |
2322 | goto out; | |
e5ff2159 AK |
2323 | |
2324 | if (write) { | |
2325 | spin_lock(&hugetlb_lock); | |
2326 | h->nr_overcommit_huge_pages = tmp; | |
2327 | spin_unlock(&hugetlb_lock); | |
2328 | } | |
08d4a246 MH |
2329 | out: |
2330 | return ret; | |
a3d0c6aa NA |
2331 | } |
2332 | ||
1da177e4 LT |
2333 | #endif /* CONFIG_SYSCTL */ |
2334 | ||
e1759c21 | 2335 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 2336 | { |
a5516438 | 2337 | struct hstate *h = &default_hstate; |
457c1b27 NA |
2338 | if (!hugepages_supported()) |
2339 | return; | |
e1759c21 | 2340 | seq_printf(m, |
4f98a2fe RR |
2341 | "HugePages_Total: %5lu\n" |
2342 | "HugePages_Free: %5lu\n" | |
2343 | "HugePages_Rsvd: %5lu\n" | |
2344 | "HugePages_Surp: %5lu\n" | |
2345 | "Hugepagesize: %8lu kB\n", | |
a5516438 AK |
2346 | h->nr_huge_pages, |
2347 | h->free_huge_pages, | |
2348 | h->resv_huge_pages, | |
2349 | h->surplus_huge_pages, | |
2350 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
2351 | } |
2352 | ||
2353 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
2354 | { | |
a5516438 | 2355 | struct hstate *h = &default_hstate; |
457c1b27 NA |
2356 | if (!hugepages_supported()) |
2357 | return 0; | |
1da177e4 LT |
2358 | return sprintf(buf, |
2359 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
2360 | "Node %d HugePages_Free: %5u\n" |
2361 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
2362 | nid, h->nr_huge_pages_node[nid], |
2363 | nid, h->free_huge_pages_node[nid], | |
2364 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
2365 | } |
2366 | ||
949f7ec5 DR |
2367 | void hugetlb_show_meminfo(void) |
2368 | { | |
2369 | struct hstate *h; | |
2370 | int nid; | |
2371 | ||
457c1b27 NA |
2372 | if (!hugepages_supported()) |
2373 | return; | |
2374 | ||
949f7ec5 DR |
2375 | for_each_node_state(nid, N_MEMORY) |
2376 | for_each_hstate(h) | |
2377 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
2378 | nid, | |
2379 | h->nr_huge_pages_node[nid], | |
2380 | h->free_huge_pages_node[nid], | |
2381 | h->surplus_huge_pages_node[nid], | |
2382 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
2383 | } | |
2384 | ||
1da177e4 LT |
2385 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
2386 | unsigned long hugetlb_total_pages(void) | |
2387 | { | |
d0028588 WL |
2388 | struct hstate *h; |
2389 | unsigned long nr_total_pages = 0; | |
2390 | ||
2391 | for_each_hstate(h) | |
2392 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
2393 | return nr_total_pages; | |
1da177e4 | 2394 | } |
1da177e4 | 2395 | |
a5516438 | 2396 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
2397 | { |
2398 | int ret = -ENOMEM; | |
2399 | ||
2400 | spin_lock(&hugetlb_lock); | |
2401 | /* | |
2402 | * When cpuset is configured, it breaks the strict hugetlb page | |
2403 | * reservation as the accounting is done on a global variable. Such | |
2404 | * reservation is completely rubbish in the presence of cpuset because | |
2405 | * the reservation is not checked against page availability for the | |
2406 | * current cpuset. Application can still potentially OOM'ed by kernel | |
2407 | * with lack of free htlb page in cpuset that the task is in. | |
2408 | * Attempt to enforce strict accounting with cpuset is almost | |
2409 | * impossible (or too ugly) because cpuset is too fluid that | |
2410 | * task or memory node can be dynamically moved between cpusets. | |
2411 | * | |
2412 | * The change of semantics for shared hugetlb mapping with cpuset is | |
2413 | * undesirable. However, in order to preserve some of the semantics, | |
2414 | * we fall back to check against current free page availability as | |
2415 | * a best attempt and hopefully to minimize the impact of changing | |
2416 | * semantics that cpuset has. | |
2417 | */ | |
2418 | if (delta > 0) { | |
a5516438 | 2419 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
2420 | goto out; |
2421 | ||
a5516438 AK |
2422 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
2423 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
2424 | goto out; |
2425 | } | |
2426 | } | |
2427 | ||
2428 | ret = 0; | |
2429 | if (delta < 0) | |
a5516438 | 2430 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
2431 | |
2432 | out: | |
2433 | spin_unlock(&hugetlb_lock); | |
2434 | return ret; | |
2435 | } | |
2436 | ||
84afd99b AW |
2437 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
2438 | { | |
f522c3ac | 2439 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
2440 | |
2441 | /* | |
2442 | * This new VMA should share its siblings reservation map if present. | |
2443 | * The VMA will only ever have a valid reservation map pointer where | |
2444 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 2445 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
2446 | * after this open call completes. It is therefore safe to take a |
2447 | * new reference here without additional locking. | |
2448 | */ | |
4e35f483 | 2449 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 2450 | kref_get(&resv->refs); |
84afd99b AW |
2451 | } |
2452 | ||
a1e78772 MG |
2453 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
2454 | { | |
a5516438 | 2455 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 2456 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 2457 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 2458 | unsigned long reserve, start, end; |
84afd99b | 2459 | |
4e35f483 JK |
2460 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
2461 | return; | |
84afd99b | 2462 | |
4e35f483 JK |
2463 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
2464 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 2465 | |
4e35f483 | 2466 | reserve = (end - start) - region_count(resv, start, end); |
84afd99b | 2467 | |
4e35f483 JK |
2468 | kref_put(&resv->refs, resv_map_release); |
2469 | ||
2470 | if (reserve) { | |
2471 | hugetlb_acct_memory(h, -reserve); | |
2472 | hugepage_subpool_put_pages(spool, reserve); | |
84afd99b | 2473 | } |
a1e78772 MG |
2474 | } |
2475 | ||
1da177e4 LT |
2476 | /* |
2477 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
2478 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
2479 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
2480 | * this far. | |
2481 | */ | |
d0217ac0 | 2482 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
2483 | { |
2484 | BUG(); | |
d0217ac0 | 2485 | return 0; |
1da177e4 LT |
2486 | } |
2487 | ||
f0f37e2f | 2488 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 2489 | .fault = hugetlb_vm_op_fault, |
84afd99b | 2490 | .open = hugetlb_vm_op_open, |
a1e78772 | 2491 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
2492 | }; |
2493 | ||
1e8f889b DG |
2494 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
2495 | int writable) | |
63551ae0 DG |
2496 | { |
2497 | pte_t entry; | |
2498 | ||
1e8f889b | 2499 | if (writable) { |
106c992a GS |
2500 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
2501 | vma->vm_page_prot))); | |
63551ae0 | 2502 | } else { |
106c992a GS |
2503 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
2504 | vma->vm_page_prot)); | |
63551ae0 DG |
2505 | } |
2506 | entry = pte_mkyoung(entry); | |
2507 | entry = pte_mkhuge(entry); | |
d9ed9faa | 2508 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
2509 | |
2510 | return entry; | |
2511 | } | |
2512 | ||
1e8f889b DG |
2513 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
2514 | unsigned long address, pte_t *ptep) | |
2515 | { | |
2516 | pte_t entry; | |
2517 | ||
106c992a | 2518 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 2519 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 2520 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
2521 | } |
2522 | ||
2523 | ||
63551ae0 DG |
2524 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
2525 | struct vm_area_struct *vma) | |
2526 | { | |
2527 | pte_t *src_pte, *dst_pte, entry; | |
2528 | struct page *ptepage; | |
1c59827d | 2529 | unsigned long addr; |
1e8f889b | 2530 | int cow; |
a5516438 AK |
2531 | struct hstate *h = hstate_vma(vma); |
2532 | unsigned long sz = huge_page_size(h); | |
e8569dd2 AS |
2533 | unsigned long mmun_start; /* For mmu_notifiers */ |
2534 | unsigned long mmun_end; /* For mmu_notifiers */ | |
2535 | int ret = 0; | |
1e8f889b DG |
2536 | |
2537 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 2538 | |
e8569dd2 AS |
2539 | mmun_start = vma->vm_start; |
2540 | mmun_end = vma->vm_end; | |
2541 | if (cow) | |
2542 | mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); | |
2543 | ||
a5516438 | 2544 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 2545 | spinlock_t *src_ptl, *dst_ptl; |
c74df32c HD |
2546 | src_pte = huge_pte_offset(src, addr); |
2547 | if (!src_pte) | |
2548 | continue; | |
a5516438 | 2549 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
2550 | if (!dst_pte) { |
2551 | ret = -ENOMEM; | |
2552 | break; | |
2553 | } | |
c5c99429 LW |
2554 | |
2555 | /* If the pagetables are shared don't copy or take references */ | |
2556 | if (dst_pte == src_pte) | |
2557 | continue; | |
2558 | ||
cb900f41 KS |
2559 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
2560 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
2561 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
7f2e9525 | 2562 | if (!huge_pte_none(huge_ptep_get(src_pte))) { |
1e8f889b | 2563 | if (cow) |
7f2e9525 GS |
2564 | huge_ptep_set_wrprotect(src, addr, src_pte); |
2565 | entry = huge_ptep_get(src_pte); | |
1c59827d HD |
2566 | ptepage = pte_page(entry); |
2567 | get_page(ptepage); | |
0fe6e20b | 2568 | page_dup_rmap(ptepage); |
1c59827d HD |
2569 | set_huge_pte_at(dst, addr, dst_pte, entry); |
2570 | } | |
cb900f41 KS |
2571 | spin_unlock(src_ptl); |
2572 | spin_unlock(dst_ptl); | |
63551ae0 | 2573 | } |
63551ae0 | 2574 | |
e8569dd2 AS |
2575 | if (cow) |
2576 | mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); | |
2577 | ||
2578 | return ret; | |
63551ae0 DG |
2579 | } |
2580 | ||
290408d4 NH |
2581 | static int is_hugetlb_entry_migration(pte_t pte) |
2582 | { | |
2583 | swp_entry_t swp; | |
2584 | ||
2585 | if (huge_pte_none(pte) || pte_present(pte)) | |
2586 | return 0; | |
2587 | swp = pte_to_swp_entry(pte); | |
32f84528 | 2588 | if (non_swap_entry(swp) && is_migration_entry(swp)) |
290408d4 | 2589 | return 1; |
32f84528 | 2590 | else |
290408d4 NH |
2591 | return 0; |
2592 | } | |
2593 | ||
fd6a03ed NH |
2594 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) |
2595 | { | |
2596 | swp_entry_t swp; | |
2597 | ||
2598 | if (huge_pte_none(pte) || pte_present(pte)) | |
2599 | return 0; | |
2600 | swp = pte_to_swp_entry(pte); | |
32f84528 | 2601 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) |
fd6a03ed | 2602 | return 1; |
32f84528 | 2603 | else |
fd6a03ed NH |
2604 | return 0; |
2605 | } | |
2606 | ||
24669e58 AK |
2607 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
2608 | unsigned long start, unsigned long end, | |
2609 | struct page *ref_page) | |
63551ae0 | 2610 | { |
24669e58 | 2611 | int force_flush = 0; |
63551ae0 DG |
2612 | struct mm_struct *mm = vma->vm_mm; |
2613 | unsigned long address; | |
c7546f8f | 2614 | pte_t *ptep; |
63551ae0 | 2615 | pte_t pte; |
cb900f41 | 2616 | spinlock_t *ptl; |
63551ae0 | 2617 | struct page *page; |
a5516438 AK |
2618 | struct hstate *h = hstate_vma(vma); |
2619 | unsigned long sz = huge_page_size(h); | |
2ec74c3e SG |
2620 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
2621 | const unsigned long mmun_end = end; /* For mmu_notifiers */ | |
a5516438 | 2622 | |
63551ae0 | 2623 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
2624 | BUG_ON(start & ~huge_page_mask(h)); |
2625 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 2626 | |
24669e58 | 2627 | tlb_start_vma(tlb, vma); |
2ec74c3e | 2628 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
24669e58 | 2629 | again: |
a5516438 | 2630 | for (address = start; address < end; address += sz) { |
c7546f8f | 2631 | ptep = huge_pte_offset(mm, address); |
4c887265 | 2632 | if (!ptep) |
c7546f8f DG |
2633 | continue; |
2634 | ||
cb900f41 | 2635 | ptl = huge_pte_lock(h, mm, ptep); |
39dde65c | 2636 | if (huge_pmd_unshare(mm, &address, ptep)) |
cb900f41 | 2637 | goto unlock; |
39dde65c | 2638 | |
6629326b HD |
2639 | pte = huge_ptep_get(ptep); |
2640 | if (huge_pte_none(pte)) | |
cb900f41 | 2641 | goto unlock; |
6629326b HD |
2642 | |
2643 | /* | |
2644 | * HWPoisoned hugepage is already unmapped and dropped reference | |
2645 | */ | |
8c4894c6 | 2646 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { |
106c992a | 2647 | huge_pte_clear(mm, address, ptep); |
cb900f41 | 2648 | goto unlock; |
8c4894c6 | 2649 | } |
6629326b HD |
2650 | |
2651 | page = pte_page(pte); | |
04f2cbe3 MG |
2652 | /* |
2653 | * If a reference page is supplied, it is because a specific | |
2654 | * page is being unmapped, not a range. Ensure the page we | |
2655 | * are about to unmap is the actual page of interest. | |
2656 | */ | |
2657 | if (ref_page) { | |
04f2cbe3 | 2658 | if (page != ref_page) |
cb900f41 | 2659 | goto unlock; |
04f2cbe3 MG |
2660 | |
2661 | /* | |
2662 | * Mark the VMA as having unmapped its page so that | |
2663 | * future faults in this VMA will fail rather than | |
2664 | * looking like data was lost | |
2665 | */ | |
2666 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
2667 | } | |
2668 | ||
c7546f8f | 2669 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
24669e58 | 2670 | tlb_remove_tlb_entry(tlb, ptep, address); |
106c992a | 2671 | if (huge_pte_dirty(pte)) |
6649a386 | 2672 | set_page_dirty(page); |
9e81130b | 2673 | |
24669e58 AK |
2674 | page_remove_rmap(page); |
2675 | force_flush = !__tlb_remove_page(tlb, page); | |
cb900f41 KS |
2676 | if (force_flush) { |
2677 | spin_unlock(ptl); | |
24669e58 | 2678 | break; |
cb900f41 | 2679 | } |
9e81130b | 2680 | /* Bail out after unmapping reference page if supplied */ |
cb900f41 KS |
2681 | if (ref_page) { |
2682 | spin_unlock(ptl); | |
9e81130b | 2683 | break; |
cb900f41 KS |
2684 | } |
2685 | unlock: | |
2686 | spin_unlock(ptl); | |
63551ae0 | 2687 | } |
24669e58 AK |
2688 | /* |
2689 | * mmu_gather ran out of room to batch pages, we break out of | |
2690 | * the PTE lock to avoid doing the potential expensive TLB invalidate | |
2691 | * and page-free while holding it. | |
2692 | */ | |
2693 | if (force_flush) { | |
2694 | force_flush = 0; | |
2695 | tlb_flush_mmu(tlb); | |
2696 | if (address < end && !ref_page) | |
2697 | goto again; | |
fe1668ae | 2698 | } |
2ec74c3e | 2699 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
24669e58 | 2700 | tlb_end_vma(tlb, vma); |
1da177e4 | 2701 | } |
63551ae0 | 2702 | |
d833352a MG |
2703 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
2704 | struct vm_area_struct *vma, unsigned long start, | |
2705 | unsigned long end, struct page *ref_page) | |
2706 | { | |
2707 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
2708 | ||
2709 | /* | |
2710 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
2711 | * test will fail on a vma being torn down, and not grab a page table | |
2712 | * on its way out. We're lucky that the flag has such an appropriate | |
2713 | * name, and can in fact be safely cleared here. We could clear it | |
2714 | * before the __unmap_hugepage_range above, but all that's necessary | |
2715 | * is to clear it before releasing the i_mmap_mutex. This works | |
2716 | * because in the context this is called, the VMA is about to be | |
2717 | * destroyed and the i_mmap_mutex is held. | |
2718 | */ | |
2719 | vma->vm_flags &= ~VM_MAYSHARE; | |
2720 | } | |
2721 | ||
502717f4 | 2722 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 2723 | unsigned long end, struct page *ref_page) |
502717f4 | 2724 | { |
24669e58 AK |
2725 | struct mm_struct *mm; |
2726 | struct mmu_gather tlb; | |
2727 | ||
2728 | mm = vma->vm_mm; | |
2729 | ||
2b047252 | 2730 | tlb_gather_mmu(&tlb, mm, start, end); |
24669e58 AK |
2731 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
2732 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 CK |
2733 | } |
2734 | ||
04f2cbe3 MG |
2735 | /* |
2736 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
2737 | * mappping it owns the reserve page for. The intention is to unmap the page | |
2738 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
2739 | * same region. | |
2740 | */ | |
2a4b3ded HH |
2741 | static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
2742 | struct page *page, unsigned long address) | |
04f2cbe3 | 2743 | { |
7526674d | 2744 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
2745 | struct vm_area_struct *iter_vma; |
2746 | struct address_space *mapping; | |
04f2cbe3 MG |
2747 | pgoff_t pgoff; |
2748 | ||
2749 | /* | |
2750 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
2751 | * from page cache lookup which is in HPAGE_SIZE units. | |
2752 | */ | |
7526674d | 2753 | address = address & huge_page_mask(h); |
36e4f20a MH |
2754 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
2755 | vma->vm_pgoff; | |
496ad9aa | 2756 | mapping = file_inode(vma->vm_file)->i_mapping; |
04f2cbe3 | 2757 | |
4eb2b1dc MG |
2758 | /* |
2759 | * Take the mapping lock for the duration of the table walk. As | |
2760 | * this mapping should be shared between all the VMAs, | |
2761 | * __unmap_hugepage_range() is called as the lock is already held | |
2762 | */ | |
3d48ae45 | 2763 | mutex_lock(&mapping->i_mmap_mutex); |
6b2dbba8 | 2764 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
2765 | /* Do not unmap the current VMA */ |
2766 | if (iter_vma == vma) | |
2767 | continue; | |
2768 | ||
2769 | /* | |
2770 | * Unmap the page from other VMAs without their own reserves. | |
2771 | * They get marked to be SIGKILLed if they fault in these | |
2772 | * areas. This is because a future no-page fault on this VMA | |
2773 | * could insert a zeroed page instead of the data existing | |
2774 | * from the time of fork. This would look like data corruption | |
2775 | */ | |
2776 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
2777 | unmap_hugepage_range(iter_vma, address, |
2778 | address + huge_page_size(h), page); | |
04f2cbe3 | 2779 | } |
3d48ae45 | 2780 | mutex_unlock(&mapping->i_mmap_mutex); |
04f2cbe3 MG |
2781 | |
2782 | return 1; | |
2783 | } | |
2784 | ||
0fe6e20b NH |
2785 | /* |
2786 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
2787 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
2788 | * cannot race with other handlers or page migration. | |
2789 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 2790 | */ |
1e8f889b | 2791 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
04f2cbe3 | 2792 | unsigned long address, pte_t *ptep, pte_t pte, |
cb900f41 | 2793 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 2794 | { |
a5516438 | 2795 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 2796 | struct page *old_page, *new_page; |
04f2cbe3 | 2797 | int outside_reserve = 0; |
2ec74c3e SG |
2798 | unsigned long mmun_start; /* For mmu_notifiers */ |
2799 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1e8f889b DG |
2800 | |
2801 | old_page = pte_page(pte); | |
2802 | ||
04f2cbe3 | 2803 | retry_avoidcopy: |
1e8f889b DG |
2804 | /* If no-one else is actually using this page, avoid the copy |
2805 | * and just make the page writable */ | |
37a2140d JK |
2806 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
2807 | page_move_anon_rmap(old_page, vma, address); | |
1e8f889b | 2808 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 2809 | return 0; |
1e8f889b DG |
2810 | } |
2811 | ||
04f2cbe3 MG |
2812 | /* |
2813 | * If the process that created a MAP_PRIVATE mapping is about to | |
2814 | * perform a COW due to a shared page count, attempt to satisfy | |
2815 | * the allocation without using the existing reserves. The pagecache | |
2816 | * page is used to determine if the reserve at this address was | |
2817 | * consumed or not. If reserves were used, a partial faulted mapping | |
2818 | * at the time of fork() could consume its reserves on COW instead | |
2819 | * of the full address range. | |
2820 | */ | |
5944d011 | 2821 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
2822 | old_page != pagecache_page) |
2823 | outside_reserve = 1; | |
2824 | ||
1e8f889b | 2825 | page_cache_get(old_page); |
b76c8cfb | 2826 | |
cb900f41 KS |
2827 | /* Drop page table lock as buddy allocator may be called */ |
2828 | spin_unlock(ptl); | |
04f2cbe3 | 2829 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 2830 | |
2fc39cec | 2831 | if (IS_ERR(new_page)) { |
76dcee75 | 2832 | long err = PTR_ERR(new_page); |
1e8f889b | 2833 | page_cache_release(old_page); |
04f2cbe3 MG |
2834 | |
2835 | /* | |
2836 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
2837 | * it is due to references held by a child and an insufficient | |
2838 | * huge page pool. To guarantee the original mappers | |
2839 | * reliability, unmap the page from child processes. The child | |
2840 | * may get SIGKILLed if it later faults. | |
2841 | */ | |
2842 | if (outside_reserve) { | |
2843 | BUG_ON(huge_pte_none(pte)); | |
2844 | if (unmap_ref_private(mm, vma, old_page, address)) { | |
04f2cbe3 | 2845 | BUG_ON(huge_pte_none(pte)); |
cb900f41 | 2846 | spin_lock(ptl); |
a734bcc8 | 2847 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
a9af0c5d NH |
2848 | if (likely(ptep && |
2849 | pte_same(huge_ptep_get(ptep), pte))) | |
a734bcc8 HD |
2850 | goto retry_avoidcopy; |
2851 | /* | |
cb900f41 KS |
2852 | * race occurs while re-acquiring page table |
2853 | * lock, and our job is done. | |
a734bcc8 HD |
2854 | */ |
2855 | return 0; | |
04f2cbe3 MG |
2856 | } |
2857 | WARN_ON_ONCE(1); | |
2858 | } | |
2859 | ||
b76c8cfb | 2860 | /* Caller expects lock to be held */ |
cb900f41 | 2861 | spin_lock(ptl); |
76dcee75 AK |
2862 | if (err == -ENOMEM) |
2863 | return VM_FAULT_OOM; | |
2864 | else | |
2865 | return VM_FAULT_SIGBUS; | |
1e8f889b DG |
2866 | } |
2867 | ||
0fe6e20b NH |
2868 | /* |
2869 | * When the original hugepage is shared one, it does not have | |
2870 | * anon_vma prepared. | |
2871 | */ | |
44e2aa93 | 2872 | if (unlikely(anon_vma_prepare(vma))) { |
ea4039a3 HD |
2873 | page_cache_release(new_page); |
2874 | page_cache_release(old_page); | |
44e2aa93 | 2875 | /* Caller expects lock to be held */ |
cb900f41 | 2876 | spin_lock(ptl); |
0fe6e20b | 2877 | return VM_FAULT_OOM; |
44e2aa93 | 2878 | } |
0fe6e20b | 2879 | |
47ad8475 AA |
2880 | copy_user_huge_page(new_page, old_page, address, vma, |
2881 | pages_per_huge_page(h)); | |
0ed361de | 2882 | __SetPageUptodate(new_page); |
1e8f889b | 2883 | |
2ec74c3e SG |
2884 | mmun_start = address & huge_page_mask(h); |
2885 | mmun_end = mmun_start + huge_page_size(h); | |
2886 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
b76c8cfb | 2887 | /* |
cb900f41 | 2888 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
2889 | * before the page tables are altered |
2890 | */ | |
cb900f41 | 2891 | spin_lock(ptl); |
a5516438 | 2892 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
a9af0c5d | 2893 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
2894 | ClearPagePrivate(new_page); |
2895 | ||
1e8f889b | 2896 | /* Break COW */ |
8fe627ec | 2897 | huge_ptep_clear_flush(vma, address, ptep); |
1e8f889b DG |
2898 | set_huge_pte_at(mm, address, ptep, |
2899 | make_huge_pte(vma, new_page, 1)); | |
0fe6e20b | 2900 | page_remove_rmap(old_page); |
cd67f0d2 | 2901 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
2902 | /* Make the old page be freed below */ |
2903 | new_page = old_page; | |
2904 | } | |
cb900f41 | 2905 | spin_unlock(ptl); |
2ec74c3e | 2906 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
1e8f889b DG |
2907 | page_cache_release(new_page); |
2908 | page_cache_release(old_page); | |
8312034f JK |
2909 | |
2910 | /* Caller expects lock to be held */ | |
cb900f41 | 2911 | spin_lock(ptl); |
83c54070 | 2912 | return 0; |
1e8f889b DG |
2913 | } |
2914 | ||
04f2cbe3 | 2915 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
2916 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
2917 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
2918 | { |
2919 | struct address_space *mapping; | |
e7c4b0bf | 2920 | pgoff_t idx; |
04f2cbe3 MG |
2921 | |
2922 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 2923 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
2924 | |
2925 | return find_lock_page(mapping, idx); | |
2926 | } | |
2927 | ||
3ae77f43 HD |
2928 | /* |
2929 | * Return whether there is a pagecache page to back given address within VMA. | |
2930 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
2931 | */ | |
2932 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
2933 | struct vm_area_struct *vma, unsigned long address) |
2934 | { | |
2935 | struct address_space *mapping; | |
2936 | pgoff_t idx; | |
2937 | struct page *page; | |
2938 | ||
2939 | mapping = vma->vm_file->f_mapping; | |
2940 | idx = vma_hugecache_offset(h, vma, address); | |
2941 | ||
2942 | page = find_get_page(mapping, idx); | |
2943 | if (page) | |
2944 | put_page(page); | |
2945 | return page != NULL; | |
2946 | } | |
2947 | ||
a1ed3dda | 2948 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
8382d914 DB |
2949 | struct address_space *mapping, pgoff_t idx, |
2950 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 2951 | { |
a5516438 | 2952 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 2953 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 2954 | int anon_rmap = 0; |
4c887265 | 2955 | unsigned long size; |
4c887265 | 2956 | struct page *page; |
1e8f889b | 2957 | pte_t new_pte; |
cb900f41 | 2958 | spinlock_t *ptl; |
4c887265 | 2959 | |
04f2cbe3 MG |
2960 | /* |
2961 | * Currently, we are forced to kill the process in the event the | |
2962 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 2963 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
2964 | */ |
2965 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
ffb22af5 AM |
2966 | pr_warning("PID %d killed due to inadequate hugepage pool\n", |
2967 | current->pid); | |
04f2cbe3 MG |
2968 | return ret; |
2969 | } | |
2970 | ||
4c887265 AL |
2971 | /* |
2972 | * Use page lock to guard against racing truncation | |
2973 | * before we get page_table_lock. | |
2974 | */ | |
6bda666a CL |
2975 | retry: |
2976 | page = find_lock_page(mapping, idx); | |
2977 | if (!page) { | |
a5516438 | 2978 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
2979 | if (idx >= size) |
2980 | goto out; | |
04f2cbe3 | 2981 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 2982 | if (IS_ERR(page)) { |
76dcee75 AK |
2983 | ret = PTR_ERR(page); |
2984 | if (ret == -ENOMEM) | |
2985 | ret = VM_FAULT_OOM; | |
2986 | else | |
2987 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
2988 | goto out; |
2989 | } | |
47ad8475 | 2990 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 2991 | __SetPageUptodate(page); |
ac9b9c66 | 2992 | |
f83a275d | 2993 | if (vma->vm_flags & VM_MAYSHARE) { |
6bda666a | 2994 | int err; |
45c682a6 | 2995 | struct inode *inode = mapping->host; |
6bda666a CL |
2996 | |
2997 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
2998 | if (err) { | |
2999 | put_page(page); | |
6bda666a CL |
3000 | if (err == -EEXIST) |
3001 | goto retry; | |
3002 | goto out; | |
3003 | } | |
07443a85 | 3004 | ClearPagePrivate(page); |
45c682a6 KC |
3005 | |
3006 | spin_lock(&inode->i_lock); | |
a5516438 | 3007 | inode->i_blocks += blocks_per_huge_page(h); |
45c682a6 | 3008 | spin_unlock(&inode->i_lock); |
23be7468 | 3009 | } else { |
6bda666a | 3010 | lock_page(page); |
0fe6e20b NH |
3011 | if (unlikely(anon_vma_prepare(vma))) { |
3012 | ret = VM_FAULT_OOM; | |
3013 | goto backout_unlocked; | |
3014 | } | |
409eb8c2 | 3015 | anon_rmap = 1; |
23be7468 | 3016 | } |
0fe6e20b | 3017 | } else { |
998b4382 NH |
3018 | /* |
3019 | * If memory error occurs between mmap() and fault, some process | |
3020 | * don't have hwpoisoned swap entry for errored virtual address. | |
3021 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
3022 | */ | |
3023 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 3024 | ret = VM_FAULT_HWPOISON | |
972dc4de | 3025 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
3026 | goto backout_unlocked; |
3027 | } | |
6bda666a | 3028 | } |
1e8f889b | 3029 | |
57303d80 AW |
3030 | /* |
3031 | * If we are going to COW a private mapping later, we examine the | |
3032 | * pending reservations for this page now. This will ensure that | |
3033 | * any allocations necessary to record that reservation occur outside | |
3034 | * the spinlock. | |
3035 | */ | |
788c7df4 | 3036 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) |
2b26736c AW |
3037 | if (vma_needs_reservation(h, vma, address) < 0) { |
3038 | ret = VM_FAULT_OOM; | |
3039 | goto backout_unlocked; | |
3040 | } | |
57303d80 | 3041 | |
cb900f41 KS |
3042 | ptl = huge_pte_lockptr(h, mm, ptep); |
3043 | spin_lock(ptl); | |
a5516438 | 3044 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
3045 | if (idx >= size) |
3046 | goto backout; | |
3047 | ||
83c54070 | 3048 | ret = 0; |
7f2e9525 | 3049 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
3050 | goto backout; |
3051 | ||
07443a85 JK |
3052 | if (anon_rmap) { |
3053 | ClearPagePrivate(page); | |
409eb8c2 | 3054 | hugepage_add_new_anon_rmap(page, vma, address); |
ac714904 | 3055 | } else |
409eb8c2 | 3056 | page_dup_rmap(page); |
1e8f889b DG |
3057 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
3058 | && (vma->vm_flags & VM_SHARED))); | |
3059 | set_huge_pte_at(mm, address, ptep, new_pte); | |
3060 | ||
788c7df4 | 3061 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 3062 | /* Optimization, do the COW without a second fault */ |
cb900f41 | 3063 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); |
1e8f889b DG |
3064 | } |
3065 | ||
cb900f41 | 3066 | spin_unlock(ptl); |
4c887265 AL |
3067 | unlock_page(page); |
3068 | out: | |
ac9b9c66 | 3069 | return ret; |
4c887265 AL |
3070 | |
3071 | backout: | |
cb900f41 | 3072 | spin_unlock(ptl); |
2b26736c | 3073 | backout_unlocked: |
4c887265 AL |
3074 | unlock_page(page); |
3075 | put_page(page); | |
3076 | goto out; | |
ac9b9c66 HD |
3077 | } |
3078 | ||
8382d914 DB |
3079 | #ifdef CONFIG_SMP |
3080 | static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm, | |
3081 | struct vm_area_struct *vma, | |
3082 | struct address_space *mapping, | |
3083 | pgoff_t idx, unsigned long address) | |
3084 | { | |
3085 | unsigned long key[2]; | |
3086 | u32 hash; | |
3087 | ||
3088 | if (vma->vm_flags & VM_SHARED) { | |
3089 | key[0] = (unsigned long) mapping; | |
3090 | key[1] = idx; | |
3091 | } else { | |
3092 | key[0] = (unsigned long) mm; | |
3093 | key[1] = address >> huge_page_shift(h); | |
3094 | } | |
3095 | ||
3096 | hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); | |
3097 | ||
3098 | return hash & (num_fault_mutexes - 1); | |
3099 | } | |
3100 | #else | |
3101 | /* | |
3102 | * For uniprocesor systems we always use a single mutex, so just | |
3103 | * return 0 and avoid the hashing overhead. | |
3104 | */ | |
3105 | static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm, | |
3106 | struct vm_area_struct *vma, | |
3107 | struct address_space *mapping, | |
3108 | pgoff_t idx, unsigned long address) | |
3109 | { | |
3110 | return 0; | |
3111 | } | |
3112 | #endif | |
3113 | ||
86e5216f | 3114 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 3115 | unsigned long address, unsigned int flags) |
86e5216f | 3116 | { |
8382d914 | 3117 | pte_t *ptep, entry; |
cb900f41 | 3118 | spinlock_t *ptl; |
1e8f889b | 3119 | int ret; |
8382d914 DB |
3120 | u32 hash; |
3121 | pgoff_t idx; | |
0fe6e20b | 3122 | struct page *page = NULL; |
57303d80 | 3123 | struct page *pagecache_page = NULL; |
a5516438 | 3124 | struct hstate *h = hstate_vma(vma); |
8382d914 | 3125 | struct address_space *mapping; |
86e5216f | 3126 | |
1e16a539 KH |
3127 | address &= huge_page_mask(h); |
3128 | ||
fd6a03ed NH |
3129 | ptep = huge_pte_offset(mm, address); |
3130 | if (ptep) { | |
3131 | entry = huge_ptep_get(ptep); | |
290408d4 | 3132 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 3133 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
3134 | return 0; |
3135 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 3136 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 3137 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
3138 | } |
3139 | ||
a5516438 | 3140 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); |
86e5216f AL |
3141 | if (!ptep) |
3142 | return VM_FAULT_OOM; | |
3143 | ||
8382d914 DB |
3144 | mapping = vma->vm_file->f_mapping; |
3145 | idx = vma_hugecache_offset(h, vma, address); | |
3146 | ||
3935baa9 DG |
3147 | /* |
3148 | * Serialize hugepage allocation and instantiation, so that we don't | |
3149 | * get spurious allocation failures if two CPUs race to instantiate | |
3150 | * the same page in the page cache. | |
3151 | */ | |
8382d914 DB |
3152 | hash = fault_mutex_hash(h, mm, vma, mapping, idx, address); |
3153 | mutex_lock(&htlb_fault_mutex_table[hash]); | |
3154 | ||
7f2e9525 GS |
3155 | entry = huge_ptep_get(ptep); |
3156 | if (huge_pte_none(entry)) { | |
8382d914 | 3157 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 3158 | goto out_mutex; |
3935baa9 | 3159 | } |
86e5216f | 3160 | |
83c54070 | 3161 | ret = 0; |
1e8f889b | 3162 | |
57303d80 AW |
3163 | /* |
3164 | * If we are going to COW the mapping later, we examine the pending | |
3165 | * reservations for this page now. This will ensure that any | |
3166 | * allocations necessary to record that reservation occur outside the | |
3167 | * spinlock. For private mappings, we also lookup the pagecache | |
3168 | * page now as it is used to determine if a reservation has been | |
3169 | * consumed. | |
3170 | */ | |
106c992a | 3171 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
2b26736c AW |
3172 | if (vma_needs_reservation(h, vma, address) < 0) { |
3173 | ret = VM_FAULT_OOM; | |
b4d1d99f | 3174 | goto out_mutex; |
2b26736c | 3175 | } |
57303d80 | 3176 | |
f83a275d | 3177 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
3178 | pagecache_page = hugetlbfs_pagecache_page(h, |
3179 | vma, address); | |
3180 | } | |
3181 | ||
56c9cfb1 NH |
3182 | /* |
3183 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
3184 | * pagecache_page, so here we need take the former one | |
3185 | * when page != pagecache_page or !pagecache_page. | |
3186 | * Note that locking order is always pagecache_page -> page, | |
3187 | * so no worry about deadlock. | |
3188 | */ | |
3189 | page = pte_page(entry); | |
66aebce7 | 3190 | get_page(page); |
56c9cfb1 | 3191 | if (page != pagecache_page) |
0fe6e20b | 3192 | lock_page(page); |
0fe6e20b | 3193 | |
cb900f41 KS |
3194 | ptl = huge_pte_lockptr(h, mm, ptep); |
3195 | spin_lock(ptl); | |
1e8f889b | 3196 | /* Check for a racing update before calling hugetlb_cow */ |
b4d1d99f | 3197 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) |
cb900f41 | 3198 | goto out_ptl; |
b4d1d99f DG |
3199 | |
3200 | ||
788c7df4 | 3201 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 3202 | if (!huge_pte_write(entry)) { |
57303d80 | 3203 | ret = hugetlb_cow(mm, vma, address, ptep, entry, |
cb900f41 KS |
3204 | pagecache_page, ptl); |
3205 | goto out_ptl; | |
b4d1d99f | 3206 | } |
106c992a | 3207 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
3208 | } |
3209 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
3210 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
3211 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 3212 | update_mmu_cache(vma, address, ptep); |
b4d1d99f | 3213 | |
cb900f41 KS |
3214 | out_ptl: |
3215 | spin_unlock(ptl); | |
57303d80 AW |
3216 | |
3217 | if (pagecache_page) { | |
3218 | unlock_page(pagecache_page); | |
3219 | put_page(pagecache_page); | |
3220 | } | |
1f64d69c DN |
3221 | if (page != pagecache_page) |
3222 | unlock_page(page); | |
66aebce7 | 3223 | put_page(page); |
57303d80 | 3224 | |
b4d1d99f | 3225 | out_mutex: |
8382d914 | 3226 | mutex_unlock(&htlb_fault_mutex_table[hash]); |
1e8f889b | 3227 | return ret; |
86e5216f AL |
3228 | } |
3229 | ||
28a35716 ML |
3230 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
3231 | struct page **pages, struct vm_area_struct **vmas, | |
3232 | unsigned long *position, unsigned long *nr_pages, | |
3233 | long i, unsigned int flags) | |
63551ae0 | 3234 | { |
d5d4b0aa CK |
3235 | unsigned long pfn_offset; |
3236 | unsigned long vaddr = *position; | |
28a35716 | 3237 | unsigned long remainder = *nr_pages; |
a5516438 | 3238 | struct hstate *h = hstate_vma(vma); |
63551ae0 | 3239 | |
63551ae0 | 3240 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 3241 | pte_t *pte; |
cb900f41 | 3242 | spinlock_t *ptl = NULL; |
2a15efc9 | 3243 | int absent; |
4c887265 | 3244 | struct page *page; |
63551ae0 | 3245 | |
4c887265 AL |
3246 | /* |
3247 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 3248 | * each hugepage. We have to make sure we get the |
4c887265 | 3249 | * first, for the page indexing below to work. |
cb900f41 KS |
3250 | * |
3251 | * Note that page table lock is not held when pte is null. | |
4c887265 | 3252 | */ |
a5516438 | 3253 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); |
cb900f41 KS |
3254 | if (pte) |
3255 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
3256 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
3257 | ||
3258 | /* | |
3259 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
3260 | * an error where there's an empty slot with no huge pagecache |
3261 | * to back it. This way, we avoid allocating a hugepage, and | |
3262 | * the sparse dumpfile avoids allocating disk blocks, but its | |
3263 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 3264 | */ |
3ae77f43 HD |
3265 | if (absent && (flags & FOLL_DUMP) && |
3266 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
3267 | if (pte) |
3268 | spin_unlock(ptl); | |
2a15efc9 HD |
3269 | remainder = 0; |
3270 | break; | |
3271 | } | |
63551ae0 | 3272 | |
9cc3a5bd NH |
3273 | /* |
3274 | * We need call hugetlb_fault for both hugepages under migration | |
3275 | * (in which case hugetlb_fault waits for the migration,) and | |
3276 | * hwpoisoned hugepages (in which case we need to prevent the | |
3277 | * caller from accessing to them.) In order to do this, we use | |
3278 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
3279 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
3280 | * both cases, and because we can't follow correct pages | |
3281 | * directly from any kind of swap entries. | |
3282 | */ | |
3283 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
3284 | ((flags & FOLL_WRITE) && |
3285 | !huge_pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 3286 | int ret; |
63551ae0 | 3287 | |
cb900f41 KS |
3288 | if (pte) |
3289 | spin_unlock(ptl); | |
2a15efc9 HD |
3290 | ret = hugetlb_fault(mm, vma, vaddr, |
3291 | (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); | |
a89182c7 | 3292 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 3293 | continue; |
63551ae0 | 3294 | |
4c887265 | 3295 | remainder = 0; |
4c887265 AL |
3296 | break; |
3297 | } | |
3298 | ||
a5516438 | 3299 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 3300 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 3301 | same_page: |
d6692183 | 3302 | if (pages) { |
2a15efc9 | 3303 | pages[i] = mem_map_offset(page, pfn_offset); |
a0368d4e | 3304 | get_page_foll(pages[i]); |
d6692183 | 3305 | } |
63551ae0 DG |
3306 | |
3307 | if (vmas) | |
3308 | vmas[i] = vma; | |
3309 | ||
3310 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 3311 | ++pfn_offset; |
63551ae0 DG |
3312 | --remainder; |
3313 | ++i; | |
d5d4b0aa | 3314 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 3315 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa CK |
3316 | /* |
3317 | * We use pfn_offset to avoid touching the pageframes | |
3318 | * of this compound page. | |
3319 | */ | |
3320 | goto same_page; | |
3321 | } | |
cb900f41 | 3322 | spin_unlock(ptl); |
63551ae0 | 3323 | } |
28a35716 | 3324 | *nr_pages = remainder; |
63551ae0 DG |
3325 | *position = vaddr; |
3326 | ||
2a15efc9 | 3327 | return i ? i : -EFAULT; |
63551ae0 | 3328 | } |
8f860591 | 3329 | |
7da4d641 | 3330 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
3331 | unsigned long address, unsigned long end, pgprot_t newprot) |
3332 | { | |
3333 | struct mm_struct *mm = vma->vm_mm; | |
3334 | unsigned long start = address; | |
3335 | pte_t *ptep; | |
3336 | pte_t pte; | |
a5516438 | 3337 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 3338 | unsigned long pages = 0; |
8f860591 ZY |
3339 | |
3340 | BUG_ON(address >= end); | |
3341 | flush_cache_range(vma, address, end); | |
3342 | ||
a5338093 | 3343 | mmu_notifier_invalidate_range_start(mm, start, end); |
3d48ae45 | 3344 | mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex); |
a5516438 | 3345 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 3346 | spinlock_t *ptl; |
8f860591 ZY |
3347 | ptep = huge_pte_offset(mm, address); |
3348 | if (!ptep) | |
3349 | continue; | |
cb900f41 | 3350 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
3351 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3352 | pages++; | |
cb900f41 | 3353 | spin_unlock(ptl); |
39dde65c | 3354 | continue; |
7da4d641 | 3355 | } |
7f2e9525 | 3356 | if (!huge_pte_none(huge_ptep_get(ptep))) { |
8f860591 | 3357 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
106c992a | 3358 | pte = pte_mkhuge(huge_pte_modify(pte, newprot)); |
be7517d6 | 3359 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
8f860591 | 3360 | set_huge_pte_at(mm, address, ptep, pte); |
7da4d641 | 3361 | pages++; |
8f860591 | 3362 | } |
cb900f41 | 3363 | spin_unlock(ptl); |
8f860591 | 3364 | } |
d833352a MG |
3365 | /* |
3366 | * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare | |
3367 | * may have cleared our pud entry and done put_page on the page table: | |
3368 | * once we release i_mmap_mutex, another task can do the final put_page | |
3369 | * and that page table be reused and filled with junk. | |
3370 | */ | |
8f860591 | 3371 | flush_tlb_range(vma, start, end); |
d833352a | 3372 | mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex); |
a5338093 | 3373 | mmu_notifier_invalidate_range_end(mm, start, end); |
7da4d641 PZ |
3374 | |
3375 | return pages << h->order; | |
8f860591 ZY |
3376 | } |
3377 | ||
a1e78772 MG |
3378 | int hugetlb_reserve_pages(struct inode *inode, |
3379 | long from, long to, | |
5a6fe125 | 3380 | struct vm_area_struct *vma, |
ca16d140 | 3381 | vm_flags_t vm_flags) |
e4e574b7 | 3382 | { |
17c9d12e | 3383 | long ret, chg; |
a5516438 | 3384 | struct hstate *h = hstate_inode(inode); |
90481622 | 3385 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 3386 | struct resv_map *resv_map; |
e4e574b7 | 3387 | |
17c9d12e MG |
3388 | /* |
3389 | * Only apply hugepage reservation if asked. At fault time, an | |
3390 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 3391 | * without using reserves |
17c9d12e | 3392 | */ |
ca16d140 | 3393 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
3394 | return 0; |
3395 | ||
a1e78772 MG |
3396 | /* |
3397 | * Shared mappings base their reservation on the number of pages that | |
3398 | * are already allocated on behalf of the file. Private mappings need | |
3399 | * to reserve the full area even if read-only as mprotect() may be | |
3400 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
3401 | */ | |
9119a41e | 3402 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4e35f483 | 3403 | resv_map = inode_resv_map(inode); |
9119a41e | 3404 | |
1406ec9b | 3405 | chg = region_chg(resv_map, from, to); |
9119a41e JK |
3406 | |
3407 | } else { | |
3408 | resv_map = resv_map_alloc(); | |
17c9d12e MG |
3409 | if (!resv_map) |
3410 | return -ENOMEM; | |
3411 | ||
a1e78772 | 3412 | chg = to - from; |
84afd99b | 3413 | |
17c9d12e MG |
3414 | set_vma_resv_map(vma, resv_map); |
3415 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
3416 | } | |
3417 | ||
c50ac050 DH |
3418 | if (chg < 0) { |
3419 | ret = chg; | |
3420 | goto out_err; | |
3421 | } | |
8a630112 | 3422 | |
90481622 | 3423 | /* There must be enough pages in the subpool for the mapping */ |
c50ac050 DH |
3424 | if (hugepage_subpool_get_pages(spool, chg)) { |
3425 | ret = -ENOSPC; | |
3426 | goto out_err; | |
3427 | } | |
5a6fe125 MG |
3428 | |
3429 | /* | |
17c9d12e | 3430 | * Check enough hugepages are available for the reservation. |
90481622 | 3431 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 3432 | */ |
a5516438 | 3433 | ret = hugetlb_acct_memory(h, chg); |
68842c9b | 3434 | if (ret < 0) { |
90481622 | 3435 | hugepage_subpool_put_pages(spool, chg); |
c50ac050 | 3436 | goto out_err; |
68842c9b | 3437 | } |
17c9d12e MG |
3438 | |
3439 | /* | |
3440 | * Account for the reservations made. Shared mappings record regions | |
3441 | * that have reservations as they are shared by multiple VMAs. | |
3442 | * When the last VMA disappears, the region map says how much | |
3443 | * the reservation was and the page cache tells how much of | |
3444 | * the reservation was consumed. Private mappings are per-VMA and | |
3445 | * only the consumed reservations are tracked. When the VMA | |
3446 | * disappears, the original reservation is the VMA size and the | |
3447 | * consumed reservations are stored in the map. Hence, nothing | |
3448 | * else has to be done for private mappings here | |
3449 | */ | |
f83a275d | 3450 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
1406ec9b | 3451 | region_add(resv_map, from, to); |
a43a8c39 | 3452 | return 0; |
c50ac050 | 3453 | out_err: |
f031dd27 JK |
3454 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3455 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 3456 | return ret; |
a43a8c39 CK |
3457 | } |
3458 | ||
3459 | void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) | |
3460 | { | |
a5516438 | 3461 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 3462 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 3463 | long chg = 0; |
90481622 | 3464 | struct hugepage_subpool *spool = subpool_inode(inode); |
45c682a6 | 3465 | |
9119a41e | 3466 | if (resv_map) |
1406ec9b | 3467 | chg = region_truncate(resv_map, offset); |
45c682a6 | 3468 | spin_lock(&inode->i_lock); |
e4c6f8be | 3469 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
3470 | spin_unlock(&inode->i_lock); |
3471 | ||
90481622 | 3472 | hugepage_subpool_put_pages(spool, (chg - freed)); |
a5516438 | 3473 | hugetlb_acct_memory(h, -(chg - freed)); |
a43a8c39 | 3474 | } |
93f70f90 | 3475 | |
3212b535 SC |
3476 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
3477 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
3478 | struct vm_area_struct *vma, | |
3479 | unsigned long addr, pgoff_t idx) | |
3480 | { | |
3481 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
3482 | svma->vm_start; | |
3483 | unsigned long sbase = saddr & PUD_MASK; | |
3484 | unsigned long s_end = sbase + PUD_SIZE; | |
3485 | ||
3486 | /* Allow segments to share if only one is marked locked */ | |
3487 | unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED; | |
3488 | unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED; | |
3489 | ||
3490 | /* | |
3491 | * match the virtual addresses, permission and the alignment of the | |
3492 | * page table page. | |
3493 | */ | |
3494 | if (pmd_index(addr) != pmd_index(saddr) || | |
3495 | vm_flags != svm_flags || | |
3496 | sbase < svma->vm_start || svma->vm_end < s_end) | |
3497 | return 0; | |
3498 | ||
3499 | return saddr; | |
3500 | } | |
3501 | ||
3502 | static int vma_shareable(struct vm_area_struct *vma, unsigned long addr) | |
3503 | { | |
3504 | unsigned long base = addr & PUD_MASK; | |
3505 | unsigned long end = base + PUD_SIZE; | |
3506 | ||
3507 | /* | |
3508 | * check on proper vm_flags and page table alignment | |
3509 | */ | |
3510 | if (vma->vm_flags & VM_MAYSHARE && | |
3511 | vma->vm_start <= base && end <= vma->vm_end) | |
3512 | return 1; | |
3513 | return 0; | |
3514 | } | |
3515 | ||
3516 | /* | |
3517 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
3518 | * and returns the corresponding pte. While this is not necessary for the | |
3519 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
3520 | * code much cleaner. pmd allocation is essential for the shared case because | |
3521 | * pud has to be populated inside the same i_mmap_mutex section - otherwise | |
3522 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a | |
3523 | * bad pmd for sharing. | |
3524 | */ | |
3525 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
3526 | { | |
3527 | struct vm_area_struct *vma = find_vma(mm, addr); | |
3528 | struct address_space *mapping = vma->vm_file->f_mapping; | |
3529 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
3530 | vma->vm_pgoff; | |
3531 | struct vm_area_struct *svma; | |
3532 | unsigned long saddr; | |
3533 | pte_t *spte = NULL; | |
3534 | pte_t *pte; | |
cb900f41 | 3535 | spinlock_t *ptl; |
3212b535 SC |
3536 | |
3537 | if (!vma_shareable(vma, addr)) | |
3538 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
3539 | ||
3540 | mutex_lock(&mapping->i_mmap_mutex); | |
3541 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { | |
3542 | if (svma == vma) | |
3543 | continue; | |
3544 | ||
3545 | saddr = page_table_shareable(svma, vma, addr, idx); | |
3546 | if (saddr) { | |
3547 | spte = huge_pte_offset(svma->vm_mm, saddr); | |
3548 | if (spte) { | |
3549 | get_page(virt_to_page(spte)); | |
3550 | break; | |
3551 | } | |
3552 | } | |
3553 | } | |
3554 | ||
3555 | if (!spte) | |
3556 | goto out; | |
3557 | ||
cb900f41 KS |
3558 | ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); |
3559 | spin_lock(ptl); | |
3212b535 SC |
3560 | if (pud_none(*pud)) |
3561 | pud_populate(mm, pud, | |
3562 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
3563 | else | |
3564 | put_page(virt_to_page(spte)); | |
cb900f41 | 3565 | spin_unlock(ptl); |
3212b535 SC |
3566 | out: |
3567 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3568 | mutex_unlock(&mapping->i_mmap_mutex); | |
3569 | return pte; | |
3570 | } | |
3571 | ||
3572 | /* | |
3573 | * unmap huge page backed by shared pte. | |
3574 | * | |
3575 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
3576 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
3577 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
3578 | * | |
cb900f41 | 3579 | * called with page table lock held. |
3212b535 SC |
3580 | * |
3581 | * returns: 1 successfully unmapped a shared pte page | |
3582 | * 0 the underlying pte page is not shared, or it is the last user | |
3583 | */ | |
3584 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
3585 | { | |
3586 | pgd_t *pgd = pgd_offset(mm, *addr); | |
3587 | pud_t *pud = pud_offset(pgd, *addr); | |
3588 | ||
3589 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
3590 | if (page_count(virt_to_page(ptep)) == 1) | |
3591 | return 0; | |
3592 | ||
3593 | pud_clear(pud); | |
3594 | put_page(virt_to_page(ptep)); | |
3595 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; | |
3596 | return 1; | |
3597 | } | |
9e5fc74c SC |
3598 | #define want_pmd_share() (1) |
3599 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
3600 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
3601 | { | |
3602 | return NULL; | |
3603 | } | |
3604 | #define want_pmd_share() (0) | |
3212b535 SC |
3605 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
3606 | ||
9e5fc74c SC |
3607 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
3608 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
3609 | unsigned long addr, unsigned long sz) | |
3610 | { | |
3611 | pgd_t *pgd; | |
3612 | pud_t *pud; | |
3613 | pte_t *pte = NULL; | |
3614 | ||
3615 | pgd = pgd_offset(mm, addr); | |
3616 | pud = pud_alloc(mm, pgd, addr); | |
3617 | if (pud) { | |
3618 | if (sz == PUD_SIZE) { | |
3619 | pte = (pte_t *)pud; | |
3620 | } else { | |
3621 | BUG_ON(sz != PMD_SIZE); | |
3622 | if (want_pmd_share() && pud_none(*pud)) | |
3623 | pte = huge_pmd_share(mm, addr, pud); | |
3624 | else | |
3625 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3626 | } | |
3627 | } | |
3628 | BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); | |
3629 | ||
3630 | return pte; | |
3631 | } | |
3632 | ||
3633 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) | |
3634 | { | |
3635 | pgd_t *pgd; | |
3636 | pud_t *pud; | |
3637 | pmd_t *pmd = NULL; | |
3638 | ||
3639 | pgd = pgd_offset(mm, addr); | |
3640 | if (pgd_present(*pgd)) { | |
3641 | pud = pud_offset(pgd, addr); | |
3642 | if (pud_present(*pud)) { | |
3643 | if (pud_huge(*pud)) | |
3644 | return (pte_t *)pud; | |
3645 | pmd = pmd_offset(pud, addr); | |
3646 | } | |
3647 | } | |
3648 | return (pte_t *) pmd; | |
3649 | } | |
3650 | ||
3651 | struct page * | |
3652 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, | |
3653 | pmd_t *pmd, int write) | |
3654 | { | |
3655 | struct page *page; | |
3656 | ||
3657 | page = pte_page(*(pte_t *)pmd); | |
3658 | if (page) | |
3659 | page += ((address & ~PMD_MASK) >> PAGE_SHIFT); | |
3660 | return page; | |
3661 | } | |
3662 | ||
3663 | struct page * | |
3664 | follow_huge_pud(struct mm_struct *mm, unsigned long address, | |
3665 | pud_t *pud, int write) | |
3666 | { | |
3667 | struct page *page; | |
3668 | ||
3669 | page = pte_page(*(pte_t *)pud); | |
3670 | if (page) | |
3671 | page += ((address & ~PUD_MASK) >> PAGE_SHIFT); | |
3672 | return page; | |
3673 | } | |
3674 | ||
3675 | #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */ | |
3676 | ||
3677 | /* Can be overriden by architectures */ | |
3b32123d | 3678 | struct page * __weak |
9e5fc74c SC |
3679 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
3680 | pud_t *pud, int write) | |
3681 | { | |
3682 | BUG(); | |
3683 | return NULL; | |
3684 | } | |
3685 | ||
3686 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ | |
3687 | ||
d5bd9106 AK |
3688 | #ifdef CONFIG_MEMORY_FAILURE |
3689 | ||
6de2b1aa NH |
3690 | /* Should be called in hugetlb_lock */ |
3691 | static int is_hugepage_on_freelist(struct page *hpage) | |
3692 | { | |
3693 | struct page *page; | |
3694 | struct page *tmp; | |
3695 | struct hstate *h = page_hstate(hpage); | |
3696 | int nid = page_to_nid(hpage); | |
3697 | ||
3698 | list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru) | |
3699 | if (page == hpage) | |
3700 | return 1; | |
3701 | return 0; | |
3702 | } | |
3703 | ||
93f70f90 NH |
3704 | /* |
3705 | * This function is called from memory failure code. | |
3706 | * Assume the caller holds page lock of the head page. | |
3707 | */ | |
6de2b1aa | 3708 | int dequeue_hwpoisoned_huge_page(struct page *hpage) |
93f70f90 NH |
3709 | { |
3710 | struct hstate *h = page_hstate(hpage); | |
3711 | int nid = page_to_nid(hpage); | |
6de2b1aa | 3712 | int ret = -EBUSY; |
93f70f90 NH |
3713 | |
3714 | spin_lock(&hugetlb_lock); | |
6de2b1aa | 3715 | if (is_hugepage_on_freelist(hpage)) { |
56f2fb14 NH |
3716 | /* |
3717 | * Hwpoisoned hugepage isn't linked to activelist or freelist, | |
3718 | * but dangling hpage->lru can trigger list-debug warnings | |
3719 | * (this happens when we call unpoison_memory() on it), | |
3720 | * so let it point to itself with list_del_init(). | |
3721 | */ | |
3722 | list_del_init(&hpage->lru); | |
8c6c2ecb | 3723 | set_page_refcounted(hpage); |
6de2b1aa NH |
3724 | h->free_huge_pages--; |
3725 | h->free_huge_pages_node[nid]--; | |
3726 | ret = 0; | |
3727 | } | |
93f70f90 | 3728 | spin_unlock(&hugetlb_lock); |
6de2b1aa | 3729 | return ret; |
93f70f90 | 3730 | } |
6de2b1aa | 3731 | #endif |
31caf665 NH |
3732 | |
3733 | bool isolate_huge_page(struct page *page, struct list_head *list) | |
3734 | { | |
309381fe | 3735 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 NH |
3736 | if (!get_page_unless_zero(page)) |
3737 | return false; | |
3738 | spin_lock(&hugetlb_lock); | |
3739 | list_move_tail(&page->lru, list); | |
3740 | spin_unlock(&hugetlb_lock); | |
3741 | return true; | |
3742 | } | |
3743 | ||
3744 | void putback_active_hugepage(struct page *page) | |
3745 | { | |
309381fe | 3746 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 NH |
3747 | spin_lock(&hugetlb_lock); |
3748 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); | |
3749 | spin_unlock(&hugetlb_lock); | |
3750 | put_page(page); | |
3751 | } | |
c8721bbb NH |
3752 | |
3753 | bool is_hugepage_active(struct page *page) | |
3754 | { | |
309381fe | 3755 | VM_BUG_ON_PAGE(!PageHuge(page), page); |
c8721bbb NH |
3756 | /* |
3757 | * This function can be called for a tail page because the caller, | |
3758 | * scan_movable_pages, scans through a given pfn-range which typically | |
3759 | * covers one memory block. In systems using gigantic hugepage (1GB | |
3760 | * for x86_64,) a hugepage is larger than a memory block, and we don't | |
3761 | * support migrating such large hugepages for now, so return false | |
3762 | * when called for tail pages. | |
3763 | */ | |
3764 | if (PageTail(page)) | |
3765 | return false; | |
3766 | /* | |
3767 | * Refcount of a hwpoisoned hugepages is 1, but they are not active, | |
3768 | * so we should return false for them. | |
3769 | */ | |
3770 | if (unlikely(PageHWPoison(page))) | |
3771 | return false; | |
3772 | return page_count(page) > 0; | |
3773 | } |