]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * Generic hugetlb support. | |
6d49e352 | 3 | * (C) Nadia Yvette Chambers, April 2004 |
1da177e4 | 4 | */ |
1da177e4 LT |
5 | #include <linux/list.h> |
6 | #include <linux/init.h> | |
7 | #include <linux/module.h> | |
8 | #include <linux/mm.h> | |
e1759c21 | 9 | #include <linux/seq_file.h> |
1da177e4 LT |
10 | #include <linux/sysctl.h> |
11 | #include <linux/highmem.h> | |
cddb8a5c | 12 | #include <linux/mmu_notifier.h> |
1da177e4 | 13 | #include <linux/nodemask.h> |
63551ae0 | 14 | #include <linux/pagemap.h> |
5da7ca86 | 15 | #include <linux/mempolicy.h> |
aea47ff3 | 16 | #include <linux/cpuset.h> |
3935baa9 | 17 | #include <linux/mutex.h> |
aa888a74 | 18 | #include <linux/bootmem.h> |
a3437870 | 19 | #include <linux/sysfs.h> |
5a0e3ad6 | 20 | #include <linux/slab.h> |
0fe6e20b | 21 | #include <linux/rmap.h> |
fd6a03ed NH |
22 | #include <linux/swap.h> |
23 | #include <linux/swapops.h> | |
c8721bbb | 24 | #include <linux/page-isolation.h> |
d6606683 | 25 | |
63551ae0 DG |
26 | #include <asm/page.h> |
27 | #include <asm/pgtable.h> | |
24669e58 | 28 | #include <asm/tlb.h> |
63551ae0 | 29 | |
24669e58 | 30 | #include <linux/io.h> |
63551ae0 | 31 | #include <linux/hugetlb.h> |
9dd540e2 | 32 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 33 | #include <linux/node.h> |
7835e98b | 34 | #include "internal.h" |
1da177e4 LT |
35 | |
36 | const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL; | |
396faf03 | 37 | unsigned long hugepages_treat_as_movable; |
a5516438 | 38 | |
c3f38a38 | 39 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
40 | unsigned int default_hstate_idx; |
41 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
42 | ||
53ba51d2 JT |
43 | __initdata LIST_HEAD(huge_boot_pages); |
44 | ||
e5ff2159 AK |
45 | /* for command line parsing */ |
46 | static struct hstate * __initdata parsed_hstate; | |
47 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 48 | static unsigned long __initdata default_hstate_size; |
e5ff2159 | 49 | |
3935baa9 | 50 | /* |
31caf665 NH |
51 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
52 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 53 | */ |
c3f38a38 | 54 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 55 | |
90481622 DG |
56 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
57 | { | |
58 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
59 | ||
60 | spin_unlock(&spool->lock); | |
61 | ||
62 | /* If no pages are used, and no other handles to the subpool | |
63 | * remain, free the subpool the subpool remain */ | |
64 | if (free) | |
65 | kfree(spool); | |
66 | } | |
67 | ||
68 | struct hugepage_subpool *hugepage_new_subpool(long nr_blocks) | |
69 | { | |
70 | struct hugepage_subpool *spool; | |
71 | ||
72 | spool = kmalloc(sizeof(*spool), GFP_KERNEL); | |
73 | if (!spool) | |
74 | return NULL; | |
75 | ||
76 | spin_lock_init(&spool->lock); | |
77 | spool->count = 1; | |
78 | spool->max_hpages = nr_blocks; | |
79 | spool->used_hpages = 0; | |
80 | ||
81 | return spool; | |
82 | } | |
83 | ||
84 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
85 | { | |
86 | spin_lock(&spool->lock); | |
87 | BUG_ON(!spool->count); | |
88 | spool->count--; | |
89 | unlock_or_release_subpool(spool); | |
90 | } | |
91 | ||
92 | static int hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
93 | long delta) | |
94 | { | |
95 | int ret = 0; | |
96 | ||
97 | if (!spool) | |
98 | return 0; | |
99 | ||
100 | spin_lock(&spool->lock); | |
101 | if ((spool->used_hpages + delta) <= spool->max_hpages) { | |
102 | spool->used_hpages += delta; | |
103 | } else { | |
104 | ret = -ENOMEM; | |
105 | } | |
106 | spin_unlock(&spool->lock); | |
107 | ||
108 | return ret; | |
109 | } | |
110 | ||
111 | static void hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
112 | long delta) | |
113 | { | |
114 | if (!spool) | |
115 | return; | |
116 | ||
117 | spin_lock(&spool->lock); | |
118 | spool->used_hpages -= delta; | |
119 | /* If hugetlbfs_put_super couldn't free spool due to | |
120 | * an outstanding quota reference, free it now. */ | |
121 | unlock_or_release_subpool(spool); | |
122 | } | |
123 | ||
124 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
125 | { | |
126 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
127 | } | |
128 | ||
129 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
130 | { | |
496ad9aa | 131 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
132 | } |
133 | ||
96822904 AW |
134 | /* |
135 | * Region tracking -- allows tracking of reservations and instantiated pages | |
136 | * across the pages in a mapping. | |
84afd99b | 137 | * |
7b24d861 DB |
138 | * The region data structures are embedded into a resv_map and |
139 | * protected by a resv_map's lock | |
96822904 AW |
140 | */ |
141 | struct file_region { | |
142 | struct list_head link; | |
143 | long from; | |
144 | long to; | |
145 | }; | |
146 | ||
1406ec9b | 147 | static long region_add(struct resv_map *resv, long f, long t) |
96822904 | 148 | { |
1406ec9b | 149 | struct list_head *head = &resv->regions; |
96822904 AW |
150 | struct file_region *rg, *nrg, *trg; |
151 | ||
7b24d861 | 152 | spin_lock(&resv->lock); |
96822904 AW |
153 | /* Locate the region we are either in or before. */ |
154 | list_for_each_entry(rg, head, link) | |
155 | if (f <= rg->to) | |
156 | break; | |
157 | ||
158 | /* Round our left edge to the current segment if it encloses us. */ | |
159 | if (f > rg->from) | |
160 | f = rg->from; | |
161 | ||
162 | /* Check for and consume any regions we now overlap with. */ | |
163 | nrg = rg; | |
164 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
165 | if (&rg->link == head) | |
166 | break; | |
167 | if (rg->from > t) | |
168 | break; | |
169 | ||
170 | /* If this area reaches higher then extend our area to | |
171 | * include it completely. If this is not the first area | |
172 | * which we intend to reuse, free it. */ | |
173 | if (rg->to > t) | |
174 | t = rg->to; | |
175 | if (rg != nrg) { | |
176 | list_del(&rg->link); | |
177 | kfree(rg); | |
178 | } | |
179 | } | |
180 | nrg->from = f; | |
181 | nrg->to = t; | |
7b24d861 | 182 | spin_unlock(&resv->lock); |
96822904 AW |
183 | return 0; |
184 | } | |
185 | ||
1406ec9b | 186 | static long region_chg(struct resv_map *resv, long f, long t) |
96822904 | 187 | { |
1406ec9b | 188 | struct list_head *head = &resv->regions; |
7b24d861 | 189 | struct file_region *rg, *nrg = NULL; |
96822904 AW |
190 | long chg = 0; |
191 | ||
7b24d861 DB |
192 | retry: |
193 | spin_lock(&resv->lock); | |
96822904 AW |
194 | /* Locate the region we are before or in. */ |
195 | list_for_each_entry(rg, head, link) | |
196 | if (f <= rg->to) | |
197 | break; | |
198 | ||
199 | /* If we are below the current region then a new region is required. | |
200 | * Subtle, allocate a new region at the position but make it zero | |
201 | * size such that we can guarantee to record the reservation. */ | |
202 | if (&rg->link == head || t < rg->from) { | |
7b24d861 DB |
203 | if (!nrg) { |
204 | spin_unlock(&resv->lock); | |
205 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
206 | if (!nrg) | |
207 | return -ENOMEM; | |
208 | ||
209 | nrg->from = f; | |
210 | nrg->to = f; | |
211 | INIT_LIST_HEAD(&nrg->link); | |
212 | goto retry; | |
213 | } | |
96822904 | 214 | |
7b24d861 DB |
215 | list_add(&nrg->link, rg->link.prev); |
216 | chg = t - f; | |
217 | goto out_nrg; | |
96822904 AW |
218 | } |
219 | ||
220 | /* Round our left edge to the current segment if it encloses us. */ | |
221 | if (f > rg->from) | |
222 | f = rg->from; | |
223 | chg = t - f; | |
224 | ||
225 | /* Check for and consume any regions we now overlap with. */ | |
226 | list_for_each_entry(rg, rg->link.prev, link) { | |
227 | if (&rg->link == head) | |
228 | break; | |
229 | if (rg->from > t) | |
7b24d861 | 230 | goto out; |
96822904 | 231 | |
25985edc | 232 | /* We overlap with this area, if it extends further than |
96822904 AW |
233 | * us then we must extend ourselves. Account for its |
234 | * existing reservation. */ | |
235 | if (rg->to > t) { | |
236 | chg += rg->to - t; | |
237 | t = rg->to; | |
238 | } | |
239 | chg -= rg->to - rg->from; | |
240 | } | |
7b24d861 DB |
241 | |
242 | out: | |
243 | spin_unlock(&resv->lock); | |
244 | /* We already know we raced and no longer need the new region */ | |
245 | kfree(nrg); | |
246 | return chg; | |
247 | out_nrg: | |
248 | spin_unlock(&resv->lock); | |
96822904 AW |
249 | return chg; |
250 | } | |
251 | ||
1406ec9b | 252 | static long region_truncate(struct resv_map *resv, long end) |
96822904 | 253 | { |
1406ec9b | 254 | struct list_head *head = &resv->regions; |
96822904 AW |
255 | struct file_region *rg, *trg; |
256 | long chg = 0; | |
257 | ||
7b24d861 | 258 | spin_lock(&resv->lock); |
96822904 AW |
259 | /* Locate the region we are either in or before. */ |
260 | list_for_each_entry(rg, head, link) | |
261 | if (end <= rg->to) | |
262 | break; | |
263 | if (&rg->link == head) | |
7b24d861 | 264 | goto out; |
96822904 AW |
265 | |
266 | /* If we are in the middle of a region then adjust it. */ | |
267 | if (end > rg->from) { | |
268 | chg = rg->to - end; | |
269 | rg->to = end; | |
270 | rg = list_entry(rg->link.next, typeof(*rg), link); | |
271 | } | |
272 | ||
273 | /* Drop any remaining regions. */ | |
274 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
275 | if (&rg->link == head) | |
276 | break; | |
277 | chg += rg->to - rg->from; | |
278 | list_del(&rg->link); | |
279 | kfree(rg); | |
280 | } | |
7b24d861 DB |
281 | |
282 | out: | |
283 | spin_unlock(&resv->lock); | |
96822904 AW |
284 | return chg; |
285 | } | |
286 | ||
1406ec9b | 287 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 288 | { |
1406ec9b | 289 | struct list_head *head = &resv->regions; |
84afd99b AW |
290 | struct file_region *rg; |
291 | long chg = 0; | |
292 | ||
7b24d861 | 293 | spin_lock(&resv->lock); |
84afd99b AW |
294 | /* Locate each segment we overlap with, and count that overlap. */ |
295 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
296 | long seg_from; |
297 | long seg_to; | |
84afd99b AW |
298 | |
299 | if (rg->to <= f) | |
300 | continue; | |
301 | if (rg->from >= t) | |
302 | break; | |
303 | ||
304 | seg_from = max(rg->from, f); | |
305 | seg_to = min(rg->to, t); | |
306 | ||
307 | chg += seg_to - seg_from; | |
308 | } | |
7b24d861 | 309 | spin_unlock(&resv->lock); |
84afd99b AW |
310 | |
311 | return chg; | |
312 | } | |
313 | ||
e7c4b0bf AW |
314 | /* |
315 | * Convert the address within this vma to the page offset within | |
316 | * the mapping, in pagecache page units; huge pages here. | |
317 | */ | |
a5516438 AK |
318 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
319 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 320 | { |
a5516438 AK |
321 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
322 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
323 | } |
324 | ||
0fe6e20b NH |
325 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
326 | unsigned long address) | |
327 | { | |
328 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
329 | } | |
330 | ||
08fba699 MG |
331 | /* |
332 | * Return the size of the pages allocated when backing a VMA. In the majority | |
333 | * cases this will be same size as used by the page table entries. | |
334 | */ | |
335 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
336 | { | |
337 | struct hstate *hstate; | |
338 | ||
339 | if (!is_vm_hugetlb_page(vma)) | |
340 | return PAGE_SIZE; | |
341 | ||
342 | hstate = hstate_vma(vma); | |
343 | ||
2415cf12 | 344 | return 1UL << huge_page_shift(hstate); |
08fba699 | 345 | } |
f340ca0f | 346 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 347 | |
3340289d MG |
348 | /* |
349 | * Return the page size being used by the MMU to back a VMA. In the majority | |
350 | * of cases, the page size used by the kernel matches the MMU size. On | |
351 | * architectures where it differs, an architecture-specific version of this | |
352 | * function is required. | |
353 | */ | |
354 | #ifndef vma_mmu_pagesize | |
355 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) | |
356 | { | |
357 | return vma_kernel_pagesize(vma); | |
358 | } | |
359 | #endif | |
360 | ||
84afd99b AW |
361 | /* |
362 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
363 | * bits of the reservation map pointer, which are always clear due to | |
364 | * alignment. | |
365 | */ | |
366 | #define HPAGE_RESV_OWNER (1UL << 0) | |
367 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 368 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 369 | |
a1e78772 MG |
370 | /* |
371 | * These helpers are used to track how many pages are reserved for | |
372 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
373 | * is guaranteed to have their future faults succeed. | |
374 | * | |
375 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
376 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
377 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
378 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
379 | * |
380 | * The private mapping reservation is represented in a subtly different | |
381 | * manner to a shared mapping. A shared mapping has a region map associated | |
382 | * with the underlying file, this region map represents the backing file | |
383 | * pages which have ever had a reservation assigned which this persists even | |
384 | * after the page is instantiated. A private mapping has a region map | |
385 | * associated with the original mmap which is attached to all VMAs which | |
386 | * reference it, this region map represents those offsets which have consumed | |
387 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 388 | */ |
e7c4b0bf AW |
389 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
390 | { | |
391 | return (unsigned long)vma->vm_private_data; | |
392 | } | |
393 | ||
394 | static void set_vma_private_data(struct vm_area_struct *vma, | |
395 | unsigned long value) | |
396 | { | |
397 | vma->vm_private_data = (void *)value; | |
398 | } | |
399 | ||
9119a41e | 400 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
401 | { |
402 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
403 | if (!resv_map) | |
404 | return NULL; | |
405 | ||
406 | kref_init(&resv_map->refs); | |
7b24d861 | 407 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
408 | INIT_LIST_HEAD(&resv_map->regions); |
409 | ||
410 | return resv_map; | |
411 | } | |
412 | ||
9119a41e | 413 | void resv_map_release(struct kref *ref) |
84afd99b AW |
414 | { |
415 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
416 | ||
417 | /* Clear out any active regions before we release the map. */ | |
1406ec9b | 418 | region_truncate(resv_map, 0); |
84afd99b AW |
419 | kfree(resv_map); |
420 | } | |
421 | ||
422 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) | |
a1e78772 MG |
423 | { |
424 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 425 | if (!(vma->vm_flags & VM_MAYSHARE)) |
84afd99b AW |
426 | return (struct resv_map *)(get_vma_private_data(vma) & |
427 | ~HPAGE_RESV_MASK); | |
2a4b3ded | 428 | return NULL; |
a1e78772 MG |
429 | } |
430 | ||
84afd99b | 431 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 MG |
432 | { |
433 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 434 | VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); |
a1e78772 | 435 | |
84afd99b AW |
436 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
437 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
438 | } |
439 | ||
440 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
441 | { | |
04f2cbe3 | 442 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); |
f83a275d | 443 | VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); |
e7c4b0bf AW |
444 | |
445 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
446 | } |
447 | ||
448 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
449 | { | |
450 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
e7c4b0bf AW |
451 | |
452 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
453 | } |
454 | ||
04f2cbe3 | 455 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
456 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
457 | { | |
458 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); | |
f83a275d | 459 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
460 | vma->vm_private_data = (void *)0; |
461 | } | |
462 | ||
463 | /* Returns true if the VMA has associated reserve pages */ | |
af0ed73e | 464 | static int vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 465 | { |
af0ed73e JK |
466 | if (vma->vm_flags & VM_NORESERVE) { |
467 | /* | |
468 | * This address is already reserved by other process(chg == 0), | |
469 | * so, we should decrement reserved count. Without decrementing, | |
470 | * reserve count remains after releasing inode, because this | |
471 | * allocated page will go into page cache and is regarded as | |
472 | * coming from reserved pool in releasing step. Currently, we | |
473 | * don't have any other solution to deal with this situation | |
474 | * properly, so add work-around here. | |
475 | */ | |
476 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
477 | return 1; | |
478 | else | |
479 | return 0; | |
480 | } | |
a63884e9 JK |
481 | |
482 | /* Shared mappings always use reserves */ | |
f83a275d | 483 | if (vma->vm_flags & VM_MAYSHARE) |
7f09ca51 | 484 | return 1; |
a63884e9 JK |
485 | |
486 | /* | |
487 | * Only the process that called mmap() has reserves for | |
488 | * private mappings. | |
489 | */ | |
7f09ca51 MG |
490 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
491 | return 1; | |
a63884e9 | 492 | |
7f09ca51 | 493 | return 0; |
a1e78772 MG |
494 | } |
495 | ||
a5516438 | 496 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
497 | { |
498 | int nid = page_to_nid(page); | |
0edaecfa | 499 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
500 | h->free_huge_pages++; |
501 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
502 | } |
503 | ||
bf50bab2 NH |
504 | static struct page *dequeue_huge_page_node(struct hstate *h, int nid) |
505 | { | |
506 | struct page *page; | |
507 | ||
c8721bbb NH |
508 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
509 | if (!is_migrate_isolate_page(page)) | |
510 | break; | |
511 | /* | |
512 | * if 'non-isolated free hugepage' not found on the list, | |
513 | * the allocation fails. | |
514 | */ | |
515 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 516 | return NULL; |
0edaecfa | 517 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 518 | set_page_refcounted(page); |
bf50bab2 NH |
519 | h->free_huge_pages--; |
520 | h->free_huge_pages_node[nid]--; | |
521 | return page; | |
522 | } | |
523 | ||
86cdb465 NH |
524 | /* Movability of hugepages depends on migration support. */ |
525 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
526 | { | |
527 | if (hugepages_treat_as_movable || hugepage_migration_support(h)) | |
528 | return GFP_HIGHUSER_MOVABLE; | |
529 | else | |
530 | return GFP_HIGHUSER; | |
531 | } | |
532 | ||
a5516438 AK |
533 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
534 | struct vm_area_struct *vma, | |
af0ed73e JK |
535 | unsigned long address, int avoid_reserve, |
536 | long chg) | |
1da177e4 | 537 | { |
b1c12cbc | 538 | struct page *page = NULL; |
480eccf9 | 539 | struct mempolicy *mpol; |
19770b32 | 540 | nodemask_t *nodemask; |
c0ff7453 | 541 | struct zonelist *zonelist; |
dd1a239f MG |
542 | struct zone *zone; |
543 | struct zoneref *z; | |
cc9a6c87 | 544 | unsigned int cpuset_mems_cookie; |
1da177e4 | 545 | |
a1e78772 MG |
546 | /* |
547 | * A child process with MAP_PRIVATE mappings created by their parent | |
548 | * have no page reserves. This check ensures that reservations are | |
549 | * not "stolen". The child may still get SIGKILLed | |
550 | */ | |
af0ed73e | 551 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 552 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 553 | goto err; |
a1e78772 | 554 | |
04f2cbe3 | 555 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 556 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 557 | goto err; |
04f2cbe3 | 558 | |
9966c4bb | 559 | retry_cpuset: |
d26914d1 | 560 | cpuset_mems_cookie = read_mems_allowed_begin(); |
9966c4bb | 561 | zonelist = huge_zonelist(vma, address, |
86cdb465 | 562 | htlb_alloc_mask(h), &mpol, &nodemask); |
9966c4bb | 563 | |
19770b32 MG |
564 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
565 | MAX_NR_ZONES - 1, nodemask) { | |
86cdb465 | 566 | if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask(h))) { |
bf50bab2 NH |
567 | page = dequeue_huge_page_node(h, zone_to_nid(zone)); |
568 | if (page) { | |
af0ed73e JK |
569 | if (avoid_reserve) |
570 | break; | |
571 | if (!vma_has_reserves(vma, chg)) | |
572 | break; | |
573 | ||
07443a85 | 574 | SetPagePrivate(page); |
af0ed73e | 575 | h->resv_huge_pages--; |
bf50bab2 NH |
576 | break; |
577 | } | |
3abf7afd | 578 | } |
1da177e4 | 579 | } |
cc9a6c87 | 580 | |
52cd3b07 | 581 | mpol_cond_put(mpol); |
d26914d1 | 582 | if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) |
cc9a6c87 | 583 | goto retry_cpuset; |
1da177e4 | 584 | return page; |
cc9a6c87 MG |
585 | |
586 | err: | |
cc9a6c87 | 587 | return NULL; |
1da177e4 LT |
588 | } |
589 | ||
a5516438 | 590 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
591 | { |
592 | int i; | |
a5516438 | 593 | |
18229df5 AW |
594 | VM_BUG_ON(h->order >= MAX_ORDER); |
595 | ||
a5516438 AK |
596 | h->nr_huge_pages--; |
597 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
598 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
599 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
600 | 1 << PG_referenced | 1 << PG_dirty | | |
601 | 1 << PG_active | 1 << PG_reserved | | |
602 | 1 << PG_private | 1 << PG_writeback); | |
6af2acb6 | 603 | } |
309381fe | 604 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
6af2acb6 AL |
605 | set_compound_page_dtor(page, NULL); |
606 | set_page_refcounted(page); | |
7f2e9525 | 607 | arch_release_hugepage(page); |
a5516438 | 608 | __free_pages(page, huge_page_order(h)); |
6af2acb6 AL |
609 | } |
610 | ||
e5ff2159 AK |
611 | struct hstate *size_to_hstate(unsigned long size) |
612 | { | |
613 | struct hstate *h; | |
614 | ||
615 | for_each_hstate(h) { | |
616 | if (huge_page_size(h) == size) | |
617 | return h; | |
618 | } | |
619 | return NULL; | |
620 | } | |
621 | ||
27a85ef1 DG |
622 | static void free_huge_page(struct page *page) |
623 | { | |
a5516438 AK |
624 | /* |
625 | * Can't pass hstate in here because it is called from the | |
626 | * compound page destructor. | |
627 | */ | |
e5ff2159 | 628 | struct hstate *h = page_hstate(page); |
7893d1d5 | 629 | int nid = page_to_nid(page); |
90481622 DG |
630 | struct hugepage_subpool *spool = |
631 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 632 | bool restore_reserve; |
27a85ef1 | 633 | |
e5df70ab | 634 | set_page_private(page, 0); |
23be7468 | 635 | page->mapping = NULL; |
7893d1d5 | 636 | BUG_ON(page_count(page)); |
0fe6e20b | 637 | BUG_ON(page_mapcount(page)); |
07443a85 | 638 | restore_reserve = PagePrivate(page); |
16c794b4 | 639 | ClearPagePrivate(page); |
27a85ef1 DG |
640 | |
641 | spin_lock(&hugetlb_lock); | |
6d76dcf4 AK |
642 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
643 | pages_per_huge_page(h), page); | |
07443a85 JK |
644 | if (restore_reserve) |
645 | h->resv_huge_pages++; | |
646 | ||
aa888a74 | 647 | if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { |
0edaecfa AK |
648 | /* remove the page from active list */ |
649 | list_del(&page->lru); | |
a5516438 AK |
650 | update_and_free_page(h, page); |
651 | h->surplus_huge_pages--; | |
652 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 653 | } else { |
5d3a551c | 654 | arch_clear_hugepage_flags(page); |
a5516438 | 655 | enqueue_huge_page(h, page); |
7893d1d5 | 656 | } |
27a85ef1 | 657 | spin_unlock(&hugetlb_lock); |
90481622 | 658 | hugepage_subpool_put_pages(spool, 1); |
27a85ef1 DG |
659 | } |
660 | ||
a5516438 | 661 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 662 | { |
0edaecfa | 663 | INIT_LIST_HEAD(&page->lru); |
b7ba30c6 AK |
664 | set_compound_page_dtor(page, free_huge_page); |
665 | spin_lock(&hugetlb_lock); | |
9dd540e2 | 666 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
667 | h->nr_huge_pages++; |
668 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 AK |
669 | spin_unlock(&hugetlb_lock); |
670 | put_page(page); /* free it into the hugepage allocator */ | |
671 | } | |
672 | ||
20a0307c WF |
673 | static void prep_compound_gigantic_page(struct page *page, unsigned long order) |
674 | { | |
675 | int i; | |
676 | int nr_pages = 1 << order; | |
677 | struct page *p = page + 1; | |
678 | ||
679 | /* we rely on prep_new_huge_page to set the destructor */ | |
680 | set_compound_order(page, order); | |
681 | __SetPageHead(page); | |
ef5a22be | 682 | __ClearPageReserved(page); |
20a0307c WF |
683 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
684 | __SetPageTail(p); | |
ef5a22be AA |
685 | /* |
686 | * For gigantic hugepages allocated through bootmem at | |
687 | * boot, it's safer to be consistent with the not-gigantic | |
688 | * hugepages and clear the PG_reserved bit from all tail pages | |
689 | * too. Otherwse drivers using get_user_pages() to access tail | |
690 | * pages may get the reference counting wrong if they see | |
691 | * PG_reserved set on a tail page (despite the head page not | |
692 | * having PG_reserved set). Enforcing this consistency between | |
693 | * head and tail pages allows drivers to optimize away a check | |
694 | * on the head page when they need know if put_page() is needed | |
695 | * after get_user_pages(). | |
696 | */ | |
697 | __ClearPageReserved(p); | |
58a84aa9 | 698 | set_page_count(p, 0); |
20a0307c WF |
699 | p->first_page = page; |
700 | } | |
701 | } | |
702 | ||
7795912c AM |
703 | /* |
704 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
705 | * transparent huge pages. See the PageTransHuge() documentation for more | |
706 | * details. | |
707 | */ | |
20a0307c WF |
708 | int PageHuge(struct page *page) |
709 | { | |
20a0307c WF |
710 | if (!PageCompound(page)) |
711 | return 0; | |
712 | ||
713 | page = compound_head(page); | |
758f66a2 | 714 | return get_compound_page_dtor(page) == free_huge_page; |
20a0307c | 715 | } |
43131e14 NH |
716 | EXPORT_SYMBOL_GPL(PageHuge); |
717 | ||
27c73ae7 AA |
718 | /* |
719 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
720 | * normal or transparent huge pages. | |
721 | */ | |
722 | int PageHeadHuge(struct page *page_head) | |
723 | { | |
27c73ae7 AA |
724 | if (!PageHead(page_head)) |
725 | return 0; | |
726 | ||
758f66a2 | 727 | return get_compound_page_dtor(page_head) == free_huge_page; |
27c73ae7 | 728 | } |
27c73ae7 | 729 | |
13d60f4b ZY |
730 | pgoff_t __basepage_index(struct page *page) |
731 | { | |
732 | struct page *page_head = compound_head(page); | |
733 | pgoff_t index = page_index(page_head); | |
734 | unsigned long compound_idx; | |
735 | ||
736 | if (!PageHuge(page_head)) | |
737 | return page_index(page); | |
738 | ||
739 | if (compound_order(page_head) >= MAX_ORDER) | |
740 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
741 | else | |
742 | compound_idx = page - page_head; | |
743 | ||
744 | return (index << compound_order(page_head)) + compound_idx; | |
745 | } | |
746 | ||
a5516438 | 747 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
1da177e4 | 748 | { |
1da177e4 | 749 | struct page *page; |
f96efd58 | 750 | |
aa888a74 AK |
751 | if (h->order >= MAX_ORDER) |
752 | return NULL; | |
753 | ||
6484eb3e | 754 | page = alloc_pages_exact_node(nid, |
86cdb465 | 755 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
551883ae | 756 | __GFP_REPEAT|__GFP_NOWARN, |
a5516438 | 757 | huge_page_order(h)); |
1da177e4 | 758 | if (page) { |
7f2e9525 | 759 | if (arch_prepare_hugepage(page)) { |
caff3a2c | 760 | __free_pages(page, huge_page_order(h)); |
7b8ee84d | 761 | return NULL; |
7f2e9525 | 762 | } |
a5516438 | 763 | prep_new_huge_page(h, page, nid); |
1da177e4 | 764 | } |
63b4613c NA |
765 | |
766 | return page; | |
767 | } | |
768 | ||
9a76db09 | 769 | /* |
6ae11b27 LS |
770 | * common helper functions for hstate_next_node_to_{alloc|free}. |
771 | * We may have allocated or freed a huge page based on a different | |
772 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
773 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
774 | * node for alloc or free. | |
9a76db09 | 775 | */ |
6ae11b27 | 776 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) |
9a76db09 | 777 | { |
6ae11b27 | 778 | nid = next_node(nid, *nodes_allowed); |
9a76db09 | 779 | if (nid == MAX_NUMNODES) |
6ae11b27 | 780 | nid = first_node(*nodes_allowed); |
9a76db09 LS |
781 | VM_BUG_ON(nid >= MAX_NUMNODES); |
782 | ||
783 | return nid; | |
784 | } | |
785 | ||
6ae11b27 LS |
786 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) |
787 | { | |
788 | if (!node_isset(nid, *nodes_allowed)) | |
789 | nid = next_node_allowed(nid, nodes_allowed); | |
790 | return nid; | |
791 | } | |
792 | ||
5ced66c9 | 793 | /* |
6ae11b27 LS |
794 | * returns the previously saved node ["this node"] from which to |
795 | * allocate a persistent huge page for the pool and advance the | |
796 | * next node from which to allocate, handling wrap at end of node | |
797 | * mask. | |
5ced66c9 | 798 | */ |
6ae11b27 LS |
799 | static int hstate_next_node_to_alloc(struct hstate *h, |
800 | nodemask_t *nodes_allowed) | |
5ced66c9 | 801 | { |
6ae11b27 LS |
802 | int nid; |
803 | ||
804 | VM_BUG_ON(!nodes_allowed); | |
805 | ||
806 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
807 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
9a76db09 | 808 | |
9a76db09 | 809 | return nid; |
5ced66c9 AK |
810 | } |
811 | ||
e8c5c824 | 812 | /* |
6ae11b27 LS |
813 | * helper for free_pool_huge_page() - return the previously saved |
814 | * node ["this node"] from which to free a huge page. Advance the | |
815 | * next node id whether or not we find a free huge page to free so | |
816 | * that the next attempt to free addresses the next node. | |
e8c5c824 | 817 | */ |
6ae11b27 | 818 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) |
e8c5c824 | 819 | { |
6ae11b27 LS |
820 | int nid; |
821 | ||
822 | VM_BUG_ON(!nodes_allowed); | |
823 | ||
824 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
825 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
9a76db09 | 826 | |
9a76db09 | 827 | return nid; |
e8c5c824 LS |
828 | } |
829 | ||
b2261026 JK |
830 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ |
831 | for (nr_nodes = nodes_weight(*mask); \ | |
832 | nr_nodes > 0 && \ | |
833 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
834 | nr_nodes--) | |
835 | ||
836 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
837 | for (nr_nodes = nodes_weight(*mask); \ | |
838 | nr_nodes > 0 && \ | |
839 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
840 | nr_nodes--) | |
841 | ||
842 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) | |
843 | { | |
844 | struct page *page; | |
845 | int nr_nodes, node; | |
846 | int ret = 0; | |
847 | ||
848 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
849 | page = alloc_fresh_huge_page_node(h, node); | |
850 | if (page) { | |
851 | ret = 1; | |
852 | break; | |
853 | } | |
854 | } | |
855 | ||
856 | if (ret) | |
857 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
858 | else | |
859 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
860 | ||
861 | return ret; | |
862 | } | |
863 | ||
e8c5c824 LS |
864 | /* |
865 | * Free huge page from pool from next node to free. | |
866 | * Attempt to keep persistent huge pages more or less | |
867 | * balanced over allowed nodes. | |
868 | * Called with hugetlb_lock locked. | |
869 | */ | |
6ae11b27 LS |
870 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
871 | bool acct_surplus) | |
e8c5c824 | 872 | { |
b2261026 | 873 | int nr_nodes, node; |
e8c5c824 LS |
874 | int ret = 0; |
875 | ||
b2261026 | 876 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
877 | /* |
878 | * If we're returning unused surplus pages, only examine | |
879 | * nodes with surplus pages. | |
880 | */ | |
b2261026 JK |
881 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
882 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 883 | struct page *page = |
b2261026 | 884 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
885 | struct page, lru); |
886 | list_del(&page->lru); | |
887 | h->free_huge_pages--; | |
b2261026 | 888 | h->free_huge_pages_node[node]--; |
685f3457 LS |
889 | if (acct_surplus) { |
890 | h->surplus_huge_pages--; | |
b2261026 | 891 | h->surplus_huge_pages_node[node]--; |
685f3457 | 892 | } |
e8c5c824 LS |
893 | update_and_free_page(h, page); |
894 | ret = 1; | |
9a76db09 | 895 | break; |
e8c5c824 | 896 | } |
b2261026 | 897 | } |
e8c5c824 LS |
898 | |
899 | return ret; | |
900 | } | |
901 | ||
c8721bbb NH |
902 | /* |
903 | * Dissolve a given free hugepage into free buddy pages. This function does | |
904 | * nothing for in-use (including surplus) hugepages. | |
905 | */ | |
906 | static void dissolve_free_huge_page(struct page *page) | |
907 | { | |
908 | spin_lock(&hugetlb_lock); | |
909 | if (PageHuge(page) && !page_count(page)) { | |
910 | struct hstate *h = page_hstate(page); | |
911 | int nid = page_to_nid(page); | |
912 | list_del(&page->lru); | |
913 | h->free_huge_pages--; | |
914 | h->free_huge_pages_node[nid]--; | |
915 | update_and_free_page(h, page); | |
916 | } | |
917 | spin_unlock(&hugetlb_lock); | |
918 | } | |
919 | ||
920 | /* | |
921 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
922 | * make specified memory blocks removable from the system. | |
923 | * Note that start_pfn should aligned with (minimum) hugepage size. | |
924 | */ | |
925 | void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) | |
926 | { | |
927 | unsigned int order = 8 * sizeof(void *); | |
928 | unsigned long pfn; | |
929 | struct hstate *h; | |
930 | ||
931 | /* Set scan step to minimum hugepage size */ | |
932 | for_each_hstate(h) | |
933 | if (order > huge_page_order(h)) | |
934 | order = huge_page_order(h); | |
935 | VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order)); | |
936 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) | |
937 | dissolve_free_huge_page(pfn_to_page(pfn)); | |
938 | } | |
939 | ||
bf50bab2 | 940 | static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) |
7893d1d5 AL |
941 | { |
942 | struct page *page; | |
bf50bab2 | 943 | unsigned int r_nid; |
7893d1d5 | 944 | |
aa888a74 AK |
945 | if (h->order >= MAX_ORDER) |
946 | return NULL; | |
947 | ||
d1c3fb1f NA |
948 | /* |
949 | * Assume we will successfully allocate the surplus page to | |
950 | * prevent racing processes from causing the surplus to exceed | |
951 | * overcommit | |
952 | * | |
953 | * This however introduces a different race, where a process B | |
954 | * tries to grow the static hugepage pool while alloc_pages() is | |
955 | * called by process A. B will only examine the per-node | |
956 | * counters in determining if surplus huge pages can be | |
957 | * converted to normal huge pages in adjust_pool_surplus(). A | |
958 | * won't be able to increment the per-node counter, until the | |
959 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
960 | * no more huge pages can be converted from surplus to normal | |
961 | * state (and doesn't try to convert again). Thus, we have a | |
962 | * case where a surplus huge page exists, the pool is grown, and | |
963 | * the surplus huge page still exists after, even though it | |
964 | * should just have been converted to a normal huge page. This | |
965 | * does not leak memory, though, as the hugepage will be freed | |
966 | * once it is out of use. It also does not allow the counters to | |
967 | * go out of whack in adjust_pool_surplus() as we don't modify | |
968 | * the node values until we've gotten the hugepage and only the | |
969 | * per-node value is checked there. | |
970 | */ | |
971 | spin_lock(&hugetlb_lock); | |
a5516438 | 972 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
973 | spin_unlock(&hugetlb_lock); |
974 | return NULL; | |
975 | } else { | |
a5516438 AK |
976 | h->nr_huge_pages++; |
977 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
978 | } |
979 | spin_unlock(&hugetlb_lock); | |
980 | ||
bf50bab2 | 981 | if (nid == NUMA_NO_NODE) |
86cdb465 | 982 | page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP| |
bf50bab2 NH |
983 | __GFP_REPEAT|__GFP_NOWARN, |
984 | huge_page_order(h)); | |
985 | else | |
986 | page = alloc_pages_exact_node(nid, | |
86cdb465 | 987 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
bf50bab2 | 988 | __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); |
d1c3fb1f | 989 | |
caff3a2c GS |
990 | if (page && arch_prepare_hugepage(page)) { |
991 | __free_pages(page, huge_page_order(h)); | |
ea5768c7 | 992 | page = NULL; |
caff3a2c GS |
993 | } |
994 | ||
d1c3fb1f | 995 | spin_lock(&hugetlb_lock); |
7893d1d5 | 996 | if (page) { |
0edaecfa | 997 | INIT_LIST_HEAD(&page->lru); |
bf50bab2 | 998 | r_nid = page_to_nid(page); |
7893d1d5 | 999 | set_compound_page_dtor(page, free_huge_page); |
9dd540e2 | 1000 | set_hugetlb_cgroup(page, NULL); |
d1c3fb1f NA |
1001 | /* |
1002 | * We incremented the global counters already | |
1003 | */ | |
bf50bab2 NH |
1004 | h->nr_huge_pages_node[r_nid]++; |
1005 | h->surplus_huge_pages_node[r_nid]++; | |
3b116300 | 1006 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 1007 | } else { |
a5516438 AK |
1008 | h->nr_huge_pages--; |
1009 | h->surplus_huge_pages--; | |
3b116300 | 1010 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 1011 | } |
d1c3fb1f | 1012 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1013 | |
1014 | return page; | |
1015 | } | |
1016 | ||
bf50bab2 NH |
1017 | /* |
1018 | * This allocation function is useful in the context where vma is irrelevant. | |
1019 | * E.g. soft-offlining uses this function because it only cares physical | |
1020 | * address of error page. | |
1021 | */ | |
1022 | struct page *alloc_huge_page_node(struct hstate *h, int nid) | |
1023 | { | |
4ef91848 | 1024 | struct page *page = NULL; |
bf50bab2 NH |
1025 | |
1026 | spin_lock(&hugetlb_lock); | |
4ef91848 JK |
1027 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
1028 | page = dequeue_huge_page_node(h, nid); | |
bf50bab2 NH |
1029 | spin_unlock(&hugetlb_lock); |
1030 | ||
94ae8ba7 | 1031 | if (!page) |
bf50bab2 NH |
1032 | page = alloc_buddy_huge_page(h, nid); |
1033 | ||
1034 | return page; | |
1035 | } | |
1036 | ||
e4e574b7 | 1037 | /* |
25985edc | 1038 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1039 | * of size 'delta'. |
1040 | */ | |
a5516438 | 1041 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1042 | { |
1043 | struct list_head surplus_list; | |
1044 | struct page *page, *tmp; | |
1045 | int ret, i; | |
1046 | int needed, allocated; | |
28073b02 | 1047 | bool alloc_ok = true; |
e4e574b7 | 1048 | |
a5516438 | 1049 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1050 | if (needed <= 0) { |
a5516438 | 1051 | h->resv_huge_pages += delta; |
e4e574b7 | 1052 | return 0; |
ac09b3a1 | 1053 | } |
e4e574b7 AL |
1054 | |
1055 | allocated = 0; | |
1056 | INIT_LIST_HEAD(&surplus_list); | |
1057 | ||
1058 | ret = -ENOMEM; | |
1059 | retry: | |
1060 | spin_unlock(&hugetlb_lock); | |
1061 | for (i = 0; i < needed; i++) { | |
bf50bab2 | 1062 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
28073b02 HD |
1063 | if (!page) { |
1064 | alloc_ok = false; | |
1065 | break; | |
1066 | } | |
e4e574b7 AL |
1067 | list_add(&page->lru, &surplus_list); |
1068 | } | |
28073b02 | 1069 | allocated += i; |
e4e574b7 AL |
1070 | |
1071 | /* | |
1072 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1073 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1074 | */ | |
1075 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1076 | needed = (h->resv_huge_pages + delta) - |
1077 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1078 | if (needed > 0) { |
1079 | if (alloc_ok) | |
1080 | goto retry; | |
1081 | /* | |
1082 | * We were not able to allocate enough pages to | |
1083 | * satisfy the entire reservation so we free what | |
1084 | * we've allocated so far. | |
1085 | */ | |
1086 | goto free; | |
1087 | } | |
e4e574b7 AL |
1088 | /* |
1089 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1090 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1091 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1092 | * allocator. Commit the entire reservation here to prevent another |
1093 | * process from stealing the pages as they are added to the pool but | |
1094 | * before they are reserved. | |
e4e574b7 AL |
1095 | */ |
1096 | needed += allocated; | |
a5516438 | 1097 | h->resv_huge_pages += delta; |
e4e574b7 | 1098 | ret = 0; |
a9869b83 | 1099 | |
19fc3f0a | 1100 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1101 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1102 | if ((--needed) < 0) |
1103 | break; | |
a9869b83 NH |
1104 | /* |
1105 | * This page is now managed by the hugetlb allocator and has | |
1106 | * no users -- drop the buddy allocator's reference. | |
1107 | */ | |
1108 | put_page_testzero(page); | |
309381fe | 1109 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 1110 | enqueue_huge_page(h, page); |
19fc3f0a | 1111 | } |
28073b02 | 1112 | free: |
b0365c8d | 1113 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1114 | |
1115 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1116 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1117 | put_page(page); | |
a9869b83 | 1118 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1119 | |
1120 | return ret; | |
1121 | } | |
1122 | ||
1123 | /* | |
1124 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
1125 | * allocated to satisfy the reservation must be explicitly freed if they were | |
1126 | * never used. | |
685f3457 | 1127 | * Called with hugetlb_lock held. |
e4e574b7 | 1128 | */ |
a5516438 AK |
1129 | static void return_unused_surplus_pages(struct hstate *h, |
1130 | unsigned long unused_resv_pages) | |
e4e574b7 | 1131 | { |
e4e574b7 AL |
1132 | unsigned long nr_pages; |
1133 | ||
ac09b3a1 | 1134 | /* Uncommit the reservation */ |
a5516438 | 1135 | h->resv_huge_pages -= unused_resv_pages; |
ac09b3a1 | 1136 | |
aa888a74 AK |
1137 | /* Cannot return gigantic pages currently */ |
1138 | if (h->order >= MAX_ORDER) | |
1139 | return; | |
1140 | ||
a5516438 | 1141 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1142 | |
685f3457 LS |
1143 | /* |
1144 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1145 | * evenly across all nodes with memory. Iterate across these nodes |
1146 | * until we can no longer free unreserved surplus pages. This occurs | |
1147 | * when the nodes with surplus pages have no free pages. | |
1148 | * free_pool_huge_page() will balance the the freed pages across the | |
1149 | * on-line nodes with memory and will handle the hstate accounting. | |
685f3457 LS |
1150 | */ |
1151 | while (nr_pages--) { | |
8cebfcd0 | 1152 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
685f3457 | 1153 | break; |
e4e574b7 AL |
1154 | } |
1155 | } | |
1156 | ||
c37f9fb1 AW |
1157 | /* |
1158 | * Determine if the huge page at addr within the vma has an associated | |
1159 | * reservation. Where it does not we will need to logically increase | |
90481622 DG |
1160 | * reservation and actually increase subpool usage before an allocation |
1161 | * can occur. Where any new reservation would be required the | |
1162 | * reservation change is prepared, but not committed. Once the page | |
1163 | * has been allocated from the subpool and instantiated the change should | |
1164 | * be committed via vma_commit_reservation. No action is required on | |
1165 | * failure. | |
c37f9fb1 | 1166 | */ |
e2f17d94 | 1167 | static long vma_needs_reservation(struct hstate *h, |
a5516438 | 1168 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 AW |
1169 | { |
1170 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1171 | struct inode *inode = mapping->host; | |
1172 | ||
f83a275d | 1173 | if (vma->vm_flags & VM_MAYSHARE) { |
a5516438 | 1174 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
9119a41e JK |
1175 | struct resv_map *resv = inode->i_mapping->private_data; |
1176 | ||
1406ec9b | 1177 | return region_chg(resv, idx, idx + 1); |
c37f9fb1 | 1178 | |
84afd99b AW |
1179 | } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1180 | return 1; | |
c37f9fb1 | 1181 | |
84afd99b | 1182 | } else { |
e2f17d94 | 1183 | long err; |
a5516438 | 1184 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
f522c3ac | 1185 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b | 1186 | |
1406ec9b | 1187 | err = region_chg(resv, idx, idx + 1); |
84afd99b AW |
1188 | if (err < 0) |
1189 | return err; | |
1190 | return 0; | |
1191 | } | |
c37f9fb1 | 1192 | } |
a5516438 AK |
1193 | static void vma_commit_reservation(struct hstate *h, |
1194 | struct vm_area_struct *vma, unsigned long addr) | |
c37f9fb1 AW |
1195 | { |
1196 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1197 | struct inode *inode = mapping->host; | |
1198 | ||
f83a275d | 1199 | if (vma->vm_flags & VM_MAYSHARE) { |
a5516438 | 1200 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
9119a41e JK |
1201 | struct resv_map *resv = inode->i_mapping->private_data; |
1202 | ||
1406ec9b | 1203 | region_add(resv, idx, idx + 1); |
84afd99b AW |
1204 | |
1205 | } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { | |
a5516438 | 1206 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
f522c3ac | 1207 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
1208 | |
1209 | /* Mark this page used in the map. */ | |
1406ec9b | 1210 | region_add(resv, idx, idx + 1); |
c37f9fb1 AW |
1211 | } |
1212 | } | |
1213 | ||
a1e78772 | 1214 | static struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 1215 | unsigned long addr, int avoid_reserve) |
1da177e4 | 1216 | { |
90481622 | 1217 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 1218 | struct hstate *h = hstate_vma(vma); |
348ea204 | 1219 | struct page *page; |
e2f17d94 | 1220 | long chg; |
6d76dcf4 AK |
1221 | int ret, idx; |
1222 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 1223 | |
6d76dcf4 | 1224 | idx = hstate_index(h); |
a1e78772 | 1225 | /* |
90481622 DG |
1226 | * Processes that did not create the mapping will have no |
1227 | * reserves and will not have accounted against subpool | |
1228 | * limit. Check that the subpool limit can be made before | |
1229 | * satisfying the allocation MAP_NORESERVE mappings may also | |
1230 | * need pages and subpool limit allocated allocated if no reserve | |
1231 | * mapping overlaps. | |
a1e78772 | 1232 | */ |
a5516438 | 1233 | chg = vma_needs_reservation(h, vma, addr); |
c37f9fb1 | 1234 | if (chg < 0) |
76dcee75 | 1235 | return ERR_PTR(-ENOMEM); |
8bb3f12e JK |
1236 | if (chg || avoid_reserve) |
1237 | if (hugepage_subpool_get_pages(spool, 1)) | |
76dcee75 | 1238 | return ERR_PTR(-ENOSPC); |
1da177e4 | 1239 | |
6d76dcf4 AK |
1240 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
1241 | if (ret) { | |
8bb3f12e JK |
1242 | if (chg || avoid_reserve) |
1243 | hugepage_subpool_put_pages(spool, 1); | |
6d76dcf4 AK |
1244 | return ERR_PTR(-ENOSPC); |
1245 | } | |
1da177e4 | 1246 | spin_lock(&hugetlb_lock); |
af0ed73e | 1247 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg); |
81a6fcae | 1248 | if (!page) { |
94ae8ba7 | 1249 | spin_unlock(&hugetlb_lock); |
bf50bab2 | 1250 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
68842c9b | 1251 | if (!page) { |
6d76dcf4 AK |
1252 | hugetlb_cgroup_uncharge_cgroup(idx, |
1253 | pages_per_huge_page(h), | |
1254 | h_cg); | |
8bb3f12e JK |
1255 | if (chg || avoid_reserve) |
1256 | hugepage_subpool_put_pages(spool, 1); | |
76dcee75 | 1257 | return ERR_PTR(-ENOSPC); |
68842c9b | 1258 | } |
79dbb236 AK |
1259 | spin_lock(&hugetlb_lock); |
1260 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 1261 | /* Fall through */ |
68842c9b | 1262 | } |
81a6fcae JK |
1263 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
1264 | spin_unlock(&hugetlb_lock); | |
348ea204 | 1265 | |
90481622 | 1266 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 1267 | |
a5516438 | 1268 | vma_commit_reservation(h, vma, addr); |
90d8b7e6 | 1269 | return page; |
b45b5bd6 DG |
1270 | } |
1271 | ||
74060e4d NH |
1272 | /* |
1273 | * alloc_huge_page()'s wrapper which simply returns the page if allocation | |
1274 | * succeeds, otherwise NULL. This function is called from new_vma_page(), | |
1275 | * where no ERR_VALUE is expected to be returned. | |
1276 | */ | |
1277 | struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, | |
1278 | unsigned long addr, int avoid_reserve) | |
1279 | { | |
1280 | struct page *page = alloc_huge_page(vma, addr, avoid_reserve); | |
1281 | if (IS_ERR(page)) | |
1282 | page = NULL; | |
1283 | return page; | |
1284 | } | |
1285 | ||
91f47662 | 1286 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
aa888a74 AK |
1287 | { |
1288 | struct huge_bootmem_page *m; | |
b2261026 | 1289 | int nr_nodes, node; |
aa888a74 | 1290 | |
b2261026 | 1291 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
1292 | void *addr; |
1293 | ||
8b89a116 GS |
1294 | addr = memblock_virt_alloc_try_nid_nopanic( |
1295 | huge_page_size(h), huge_page_size(h), | |
1296 | 0, BOOTMEM_ALLOC_ACCESSIBLE, node); | |
aa888a74 AK |
1297 | if (addr) { |
1298 | /* | |
1299 | * Use the beginning of the huge page to store the | |
1300 | * huge_bootmem_page struct (until gather_bootmem | |
1301 | * puts them into the mem_map). | |
1302 | */ | |
1303 | m = addr; | |
91f47662 | 1304 | goto found; |
aa888a74 | 1305 | } |
aa888a74 AK |
1306 | } |
1307 | return 0; | |
1308 | ||
1309 | found: | |
1310 | BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); | |
1311 | /* Put them into a private list first because mem_map is not up yet */ | |
1312 | list_add(&m->list, &huge_boot_pages); | |
1313 | m->hstate = h; | |
1314 | return 1; | |
1315 | } | |
1316 | ||
18229df5 AW |
1317 | static void prep_compound_huge_page(struct page *page, int order) |
1318 | { | |
1319 | if (unlikely(order > (MAX_ORDER - 1))) | |
1320 | prep_compound_gigantic_page(page, order); | |
1321 | else | |
1322 | prep_compound_page(page, order); | |
1323 | } | |
1324 | ||
aa888a74 AK |
1325 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
1326 | static void __init gather_bootmem_prealloc(void) | |
1327 | { | |
1328 | struct huge_bootmem_page *m; | |
1329 | ||
1330 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 1331 | struct hstate *h = m->hstate; |
ee8f248d BB |
1332 | struct page *page; |
1333 | ||
1334 | #ifdef CONFIG_HIGHMEM | |
1335 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
8b89a116 GS |
1336 | memblock_free_late(__pa(m), |
1337 | sizeof(struct huge_bootmem_page)); | |
ee8f248d BB |
1338 | #else |
1339 | page = virt_to_page(m); | |
1340 | #endif | |
aa888a74 | 1341 | WARN_ON(page_count(page) != 1); |
18229df5 | 1342 | prep_compound_huge_page(page, h->order); |
ef5a22be | 1343 | WARN_ON(PageReserved(page)); |
aa888a74 | 1344 | prep_new_huge_page(h, page, page_to_nid(page)); |
b0320c7b RA |
1345 | /* |
1346 | * If we had gigantic hugepages allocated at boot time, we need | |
1347 | * to restore the 'stolen' pages to totalram_pages in order to | |
1348 | * fix confusing memory reports from free(1) and another | |
1349 | * side-effects, like CommitLimit going negative. | |
1350 | */ | |
1351 | if (h->order > (MAX_ORDER - 1)) | |
3dcc0571 | 1352 | adjust_managed_page_count(page, 1 << h->order); |
aa888a74 AK |
1353 | } |
1354 | } | |
1355 | ||
8faa8b07 | 1356 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
1357 | { |
1358 | unsigned long i; | |
a5516438 | 1359 | |
e5ff2159 | 1360 | for (i = 0; i < h->max_huge_pages; ++i) { |
aa888a74 AK |
1361 | if (h->order >= MAX_ORDER) { |
1362 | if (!alloc_bootmem_huge_page(h)) | |
1363 | break; | |
9b5e5d0f | 1364 | } else if (!alloc_fresh_huge_page(h, |
8cebfcd0 | 1365 | &node_states[N_MEMORY])) |
1da177e4 | 1366 | break; |
1da177e4 | 1367 | } |
8faa8b07 | 1368 | h->max_huge_pages = i; |
e5ff2159 AK |
1369 | } |
1370 | ||
1371 | static void __init hugetlb_init_hstates(void) | |
1372 | { | |
1373 | struct hstate *h; | |
1374 | ||
1375 | for_each_hstate(h) { | |
8faa8b07 AK |
1376 | /* oversize hugepages were init'ed in early boot */ |
1377 | if (h->order < MAX_ORDER) | |
1378 | hugetlb_hstate_alloc_pages(h); | |
e5ff2159 AK |
1379 | } |
1380 | } | |
1381 | ||
4abd32db AK |
1382 | static char * __init memfmt(char *buf, unsigned long n) |
1383 | { | |
1384 | if (n >= (1UL << 30)) | |
1385 | sprintf(buf, "%lu GB", n >> 30); | |
1386 | else if (n >= (1UL << 20)) | |
1387 | sprintf(buf, "%lu MB", n >> 20); | |
1388 | else | |
1389 | sprintf(buf, "%lu KB", n >> 10); | |
1390 | return buf; | |
1391 | } | |
1392 | ||
e5ff2159 AK |
1393 | static void __init report_hugepages(void) |
1394 | { | |
1395 | struct hstate *h; | |
1396 | ||
1397 | for_each_hstate(h) { | |
4abd32db | 1398 | char buf[32]; |
ffb22af5 | 1399 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
4abd32db AK |
1400 | memfmt(buf, huge_page_size(h)), |
1401 | h->free_huge_pages); | |
e5ff2159 AK |
1402 | } |
1403 | } | |
1404 | ||
1da177e4 | 1405 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
1406 | static void try_to_free_low(struct hstate *h, unsigned long count, |
1407 | nodemask_t *nodes_allowed) | |
1da177e4 | 1408 | { |
4415cc8d CL |
1409 | int i; |
1410 | ||
aa888a74 AK |
1411 | if (h->order >= MAX_ORDER) |
1412 | return; | |
1413 | ||
6ae11b27 | 1414 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 1415 | struct page *page, *next; |
a5516438 AK |
1416 | struct list_head *freel = &h->hugepage_freelists[i]; |
1417 | list_for_each_entry_safe(page, next, freel, lru) { | |
1418 | if (count >= h->nr_huge_pages) | |
6b0c880d | 1419 | return; |
1da177e4 LT |
1420 | if (PageHighMem(page)) |
1421 | continue; | |
1422 | list_del(&page->lru); | |
e5ff2159 | 1423 | update_and_free_page(h, page); |
a5516438 AK |
1424 | h->free_huge_pages--; |
1425 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
1426 | } |
1427 | } | |
1428 | } | |
1429 | #else | |
6ae11b27 LS |
1430 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
1431 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
1432 | { |
1433 | } | |
1434 | #endif | |
1435 | ||
20a0307c WF |
1436 | /* |
1437 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
1438 | * balanced by operating on them in a round-robin fashion. | |
1439 | * Returns 1 if an adjustment was made. | |
1440 | */ | |
6ae11b27 LS |
1441 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
1442 | int delta) | |
20a0307c | 1443 | { |
b2261026 | 1444 | int nr_nodes, node; |
20a0307c WF |
1445 | |
1446 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 1447 | |
b2261026 JK |
1448 | if (delta < 0) { |
1449 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1450 | if (h->surplus_huge_pages_node[node]) | |
1451 | goto found; | |
e8c5c824 | 1452 | } |
b2261026 JK |
1453 | } else { |
1454 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
1455 | if (h->surplus_huge_pages_node[node] < | |
1456 | h->nr_huge_pages_node[node]) | |
1457 | goto found; | |
e8c5c824 | 1458 | } |
b2261026 JK |
1459 | } |
1460 | return 0; | |
20a0307c | 1461 | |
b2261026 JK |
1462 | found: |
1463 | h->surplus_huge_pages += delta; | |
1464 | h->surplus_huge_pages_node[node] += delta; | |
1465 | return 1; | |
20a0307c WF |
1466 | } |
1467 | ||
a5516438 | 1468 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
1469 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
1470 | nodemask_t *nodes_allowed) | |
1da177e4 | 1471 | { |
7893d1d5 | 1472 | unsigned long min_count, ret; |
1da177e4 | 1473 | |
aa888a74 AK |
1474 | if (h->order >= MAX_ORDER) |
1475 | return h->max_huge_pages; | |
1476 | ||
7893d1d5 AL |
1477 | /* |
1478 | * Increase the pool size | |
1479 | * First take pages out of surplus state. Then make up the | |
1480 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f NA |
1481 | * |
1482 | * We might race with alloc_buddy_huge_page() here and be unable | |
1483 | * to convert a surplus huge page to a normal huge page. That is | |
1484 | * not critical, though, it just means the overall size of the | |
1485 | * pool might be one hugepage larger than it needs to be, but | |
1486 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 1487 | */ |
1da177e4 | 1488 | spin_lock(&hugetlb_lock); |
a5516438 | 1489 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 1490 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
1491 | break; |
1492 | } | |
1493 | ||
a5516438 | 1494 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
1495 | /* |
1496 | * If this allocation races such that we no longer need the | |
1497 | * page, free_huge_page will handle it by freeing the page | |
1498 | * and reducing the surplus. | |
1499 | */ | |
1500 | spin_unlock(&hugetlb_lock); | |
6ae11b27 | 1501 | ret = alloc_fresh_huge_page(h, nodes_allowed); |
7893d1d5 AL |
1502 | spin_lock(&hugetlb_lock); |
1503 | if (!ret) | |
1504 | goto out; | |
1505 | ||
536240f2 MG |
1506 | /* Bail for signals. Probably ctrl-c from user */ |
1507 | if (signal_pending(current)) | |
1508 | goto out; | |
7893d1d5 | 1509 | } |
7893d1d5 AL |
1510 | |
1511 | /* | |
1512 | * Decrease the pool size | |
1513 | * First return free pages to the buddy allocator (being careful | |
1514 | * to keep enough around to satisfy reservations). Then place | |
1515 | * pages into surplus state as needed so the pool will shrink | |
1516 | * to the desired size as pages become free. | |
d1c3fb1f NA |
1517 | * |
1518 | * By placing pages into the surplus state independent of the | |
1519 | * overcommit value, we are allowing the surplus pool size to | |
1520 | * exceed overcommit. There are few sane options here. Since | |
1521 | * alloc_buddy_huge_page() is checking the global counter, | |
1522 | * though, we'll note that we're not allowed to exceed surplus | |
1523 | * and won't grow the pool anywhere else. Not until one of the | |
1524 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 1525 | */ |
a5516438 | 1526 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 1527 | min_count = max(count, min_count); |
6ae11b27 | 1528 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 1529 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 1530 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 1531 | break; |
1da177e4 | 1532 | } |
a5516438 | 1533 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 1534 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
1535 | break; |
1536 | } | |
1537 | out: | |
a5516438 | 1538 | ret = persistent_huge_pages(h); |
1da177e4 | 1539 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 1540 | return ret; |
1da177e4 LT |
1541 | } |
1542 | ||
a3437870 NA |
1543 | #define HSTATE_ATTR_RO(_name) \ |
1544 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
1545 | ||
1546 | #define HSTATE_ATTR(_name) \ | |
1547 | static struct kobj_attribute _name##_attr = \ | |
1548 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
1549 | ||
1550 | static struct kobject *hugepages_kobj; | |
1551 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1552 | ||
9a305230 LS |
1553 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
1554 | ||
1555 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
1556 | { |
1557 | int i; | |
9a305230 | 1558 | |
a3437870 | 1559 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
1560 | if (hstate_kobjs[i] == kobj) { |
1561 | if (nidp) | |
1562 | *nidp = NUMA_NO_NODE; | |
a3437870 | 1563 | return &hstates[i]; |
9a305230 LS |
1564 | } |
1565 | ||
1566 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
1567 | } |
1568 | ||
06808b08 | 1569 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
1570 | struct kobj_attribute *attr, char *buf) |
1571 | { | |
9a305230 LS |
1572 | struct hstate *h; |
1573 | unsigned long nr_huge_pages; | |
1574 | int nid; | |
1575 | ||
1576 | h = kobj_to_hstate(kobj, &nid); | |
1577 | if (nid == NUMA_NO_NODE) | |
1578 | nr_huge_pages = h->nr_huge_pages; | |
1579 | else | |
1580 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
1581 | ||
1582 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 1583 | } |
adbe8726 | 1584 | |
06808b08 LS |
1585 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
1586 | struct kobject *kobj, struct kobj_attribute *attr, | |
1587 | const char *buf, size_t len) | |
a3437870 NA |
1588 | { |
1589 | int err; | |
9a305230 | 1590 | int nid; |
06808b08 | 1591 | unsigned long count; |
9a305230 | 1592 | struct hstate *h; |
bad44b5b | 1593 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 1594 | |
3dbb95f7 | 1595 | err = kstrtoul(buf, 10, &count); |
73ae31e5 | 1596 | if (err) |
adbe8726 | 1597 | goto out; |
a3437870 | 1598 | |
9a305230 | 1599 | h = kobj_to_hstate(kobj, &nid); |
adbe8726 EM |
1600 | if (h->order >= MAX_ORDER) { |
1601 | err = -EINVAL; | |
1602 | goto out; | |
1603 | } | |
1604 | ||
9a305230 LS |
1605 | if (nid == NUMA_NO_NODE) { |
1606 | /* | |
1607 | * global hstate attribute | |
1608 | */ | |
1609 | if (!(obey_mempolicy && | |
1610 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
1611 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 1612 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 LS |
1613 | } |
1614 | } else if (nodes_allowed) { | |
1615 | /* | |
1616 | * per node hstate attribute: adjust count to global, | |
1617 | * but restrict alloc/free to the specified node. | |
1618 | */ | |
1619 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
1620 | init_nodemask_of_node(nodes_allowed, nid); | |
1621 | } else | |
8cebfcd0 | 1622 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 | 1623 | |
06808b08 | 1624 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 1625 | |
8cebfcd0 | 1626 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
1627 | NODEMASK_FREE(nodes_allowed); |
1628 | ||
1629 | return len; | |
adbe8726 EM |
1630 | out: |
1631 | NODEMASK_FREE(nodes_allowed); | |
1632 | return err; | |
06808b08 LS |
1633 | } |
1634 | ||
1635 | static ssize_t nr_hugepages_show(struct kobject *kobj, | |
1636 | struct kobj_attribute *attr, char *buf) | |
1637 | { | |
1638 | return nr_hugepages_show_common(kobj, attr, buf); | |
1639 | } | |
1640 | ||
1641 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
1642 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1643 | { | |
1644 | return nr_hugepages_store_common(false, kobj, attr, buf, len); | |
a3437870 NA |
1645 | } |
1646 | HSTATE_ATTR(nr_hugepages); | |
1647 | ||
06808b08 LS |
1648 | #ifdef CONFIG_NUMA |
1649 | ||
1650 | /* | |
1651 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
1652 | * huge page alloc/free. | |
1653 | */ | |
1654 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
1655 | struct kobj_attribute *attr, char *buf) | |
1656 | { | |
1657 | return nr_hugepages_show_common(kobj, attr, buf); | |
1658 | } | |
1659 | ||
1660 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
1661 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1662 | { | |
1663 | return nr_hugepages_store_common(true, kobj, attr, buf, len); | |
1664 | } | |
1665 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
1666 | #endif | |
1667 | ||
1668 | ||
a3437870 NA |
1669 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
1670 | struct kobj_attribute *attr, char *buf) | |
1671 | { | |
9a305230 | 1672 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1673 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
1674 | } | |
adbe8726 | 1675 | |
a3437870 NA |
1676 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
1677 | struct kobj_attribute *attr, const char *buf, size_t count) | |
1678 | { | |
1679 | int err; | |
1680 | unsigned long input; | |
9a305230 | 1681 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 1682 | |
adbe8726 EM |
1683 | if (h->order >= MAX_ORDER) |
1684 | return -EINVAL; | |
1685 | ||
3dbb95f7 | 1686 | err = kstrtoul(buf, 10, &input); |
a3437870 | 1687 | if (err) |
73ae31e5 | 1688 | return err; |
a3437870 NA |
1689 | |
1690 | spin_lock(&hugetlb_lock); | |
1691 | h->nr_overcommit_huge_pages = input; | |
1692 | spin_unlock(&hugetlb_lock); | |
1693 | ||
1694 | return count; | |
1695 | } | |
1696 | HSTATE_ATTR(nr_overcommit_hugepages); | |
1697 | ||
1698 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
1699 | struct kobj_attribute *attr, char *buf) | |
1700 | { | |
9a305230 LS |
1701 | struct hstate *h; |
1702 | unsigned long free_huge_pages; | |
1703 | int nid; | |
1704 | ||
1705 | h = kobj_to_hstate(kobj, &nid); | |
1706 | if (nid == NUMA_NO_NODE) | |
1707 | free_huge_pages = h->free_huge_pages; | |
1708 | else | |
1709 | free_huge_pages = h->free_huge_pages_node[nid]; | |
1710 | ||
1711 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
1712 | } |
1713 | HSTATE_ATTR_RO(free_hugepages); | |
1714 | ||
1715 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
1716 | struct kobj_attribute *attr, char *buf) | |
1717 | { | |
9a305230 | 1718 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1719 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
1720 | } | |
1721 | HSTATE_ATTR_RO(resv_hugepages); | |
1722 | ||
1723 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
1724 | struct kobj_attribute *attr, char *buf) | |
1725 | { | |
9a305230 LS |
1726 | struct hstate *h; |
1727 | unsigned long surplus_huge_pages; | |
1728 | int nid; | |
1729 | ||
1730 | h = kobj_to_hstate(kobj, &nid); | |
1731 | if (nid == NUMA_NO_NODE) | |
1732 | surplus_huge_pages = h->surplus_huge_pages; | |
1733 | else | |
1734 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
1735 | ||
1736 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
1737 | } |
1738 | HSTATE_ATTR_RO(surplus_hugepages); | |
1739 | ||
1740 | static struct attribute *hstate_attrs[] = { | |
1741 | &nr_hugepages_attr.attr, | |
1742 | &nr_overcommit_hugepages_attr.attr, | |
1743 | &free_hugepages_attr.attr, | |
1744 | &resv_hugepages_attr.attr, | |
1745 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
1746 | #ifdef CONFIG_NUMA |
1747 | &nr_hugepages_mempolicy_attr.attr, | |
1748 | #endif | |
a3437870 NA |
1749 | NULL, |
1750 | }; | |
1751 | ||
1752 | static struct attribute_group hstate_attr_group = { | |
1753 | .attrs = hstate_attrs, | |
1754 | }; | |
1755 | ||
094e9539 JM |
1756 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
1757 | struct kobject **hstate_kobjs, | |
1758 | struct attribute_group *hstate_attr_group) | |
a3437870 NA |
1759 | { |
1760 | int retval; | |
972dc4de | 1761 | int hi = hstate_index(h); |
a3437870 | 1762 | |
9a305230 LS |
1763 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
1764 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
1765 | return -ENOMEM; |
1766 | ||
9a305230 | 1767 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 1768 | if (retval) |
9a305230 | 1769 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
1770 | |
1771 | return retval; | |
1772 | } | |
1773 | ||
1774 | static void __init hugetlb_sysfs_init(void) | |
1775 | { | |
1776 | struct hstate *h; | |
1777 | int err; | |
1778 | ||
1779 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
1780 | if (!hugepages_kobj) | |
1781 | return; | |
1782 | ||
1783 | for_each_hstate(h) { | |
9a305230 LS |
1784 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
1785 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 1786 | if (err) |
ffb22af5 | 1787 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
1788 | } |
1789 | } | |
1790 | ||
9a305230 LS |
1791 | #ifdef CONFIG_NUMA |
1792 | ||
1793 | /* | |
1794 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
1795 | * with node devices in node_devices[] using a parallel array. The array |
1796 | * index of a node device or _hstate == node id. | |
1797 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
1798 | * the base kernel, on the hugetlb module. |
1799 | */ | |
1800 | struct node_hstate { | |
1801 | struct kobject *hugepages_kobj; | |
1802 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1803 | }; | |
1804 | struct node_hstate node_hstates[MAX_NUMNODES]; | |
1805 | ||
1806 | /* | |
10fbcf4c | 1807 | * A subset of global hstate attributes for node devices |
9a305230 LS |
1808 | */ |
1809 | static struct attribute *per_node_hstate_attrs[] = { | |
1810 | &nr_hugepages_attr.attr, | |
1811 | &free_hugepages_attr.attr, | |
1812 | &surplus_hugepages_attr.attr, | |
1813 | NULL, | |
1814 | }; | |
1815 | ||
1816 | static struct attribute_group per_node_hstate_attr_group = { | |
1817 | .attrs = per_node_hstate_attrs, | |
1818 | }; | |
1819 | ||
1820 | /* | |
10fbcf4c | 1821 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
1822 | * Returns node id via non-NULL nidp. |
1823 | */ | |
1824 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
1825 | { | |
1826 | int nid; | |
1827 | ||
1828 | for (nid = 0; nid < nr_node_ids; nid++) { | |
1829 | struct node_hstate *nhs = &node_hstates[nid]; | |
1830 | int i; | |
1831 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
1832 | if (nhs->hstate_kobjs[i] == kobj) { | |
1833 | if (nidp) | |
1834 | *nidp = nid; | |
1835 | return &hstates[i]; | |
1836 | } | |
1837 | } | |
1838 | ||
1839 | BUG(); | |
1840 | return NULL; | |
1841 | } | |
1842 | ||
1843 | /* | |
10fbcf4c | 1844 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
1845 | * No-op if no hstate attributes attached. |
1846 | */ | |
3cd8b44f | 1847 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
1848 | { |
1849 | struct hstate *h; | |
10fbcf4c | 1850 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
1851 | |
1852 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 1853 | return; /* no hstate attributes */ |
9a305230 | 1854 | |
972dc4de AK |
1855 | for_each_hstate(h) { |
1856 | int idx = hstate_index(h); | |
1857 | if (nhs->hstate_kobjs[idx]) { | |
1858 | kobject_put(nhs->hstate_kobjs[idx]); | |
1859 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 1860 | } |
972dc4de | 1861 | } |
9a305230 LS |
1862 | |
1863 | kobject_put(nhs->hugepages_kobj); | |
1864 | nhs->hugepages_kobj = NULL; | |
1865 | } | |
1866 | ||
1867 | /* | |
10fbcf4c | 1868 | * hugetlb module exit: unregister hstate attributes from node devices |
9a305230 LS |
1869 | * that have them. |
1870 | */ | |
1871 | static void hugetlb_unregister_all_nodes(void) | |
1872 | { | |
1873 | int nid; | |
1874 | ||
1875 | /* | |
10fbcf4c | 1876 | * disable node device registrations. |
9a305230 LS |
1877 | */ |
1878 | register_hugetlbfs_with_node(NULL, NULL); | |
1879 | ||
1880 | /* | |
1881 | * remove hstate attributes from any nodes that have them. | |
1882 | */ | |
1883 | for (nid = 0; nid < nr_node_ids; nid++) | |
8732794b | 1884 | hugetlb_unregister_node(node_devices[nid]); |
9a305230 LS |
1885 | } |
1886 | ||
1887 | /* | |
10fbcf4c | 1888 | * Register hstate attributes for a single node device. |
9a305230 LS |
1889 | * No-op if attributes already registered. |
1890 | */ | |
3cd8b44f | 1891 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
1892 | { |
1893 | struct hstate *h; | |
10fbcf4c | 1894 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
1895 | int err; |
1896 | ||
1897 | if (nhs->hugepages_kobj) | |
1898 | return; /* already allocated */ | |
1899 | ||
1900 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 1901 | &node->dev.kobj); |
9a305230 LS |
1902 | if (!nhs->hugepages_kobj) |
1903 | return; | |
1904 | ||
1905 | for_each_hstate(h) { | |
1906 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
1907 | nhs->hstate_kobjs, | |
1908 | &per_node_hstate_attr_group); | |
1909 | if (err) { | |
ffb22af5 AM |
1910 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
1911 | h->name, node->dev.id); | |
9a305230 LS |
1912 | hugetlb_unregister_node(node); |
1913 | break; | |
1914 | } | |
1915 | } | |
1916 | } | |
1917 | ||
1918 | /* | |
9b5e5d0f | 1919 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
1920 | * devices of nodes that have memory. All on-line nodes should have |
1921 | * registered their associated device by this time. | |
9a305230 LS |
1922 | */ |
1923 | static void hugetlb_register_all_nodes(void) | |
1924 | { | |
1925 | int nid; | |
1926 | ||
8cebfcd0 | 1927 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 1928 | struct node *node = node_devices[nid]; |
10fbcf4c | 1929 | if (node->dev.id == nid) |
9a305230 LS |
1930 | hugetlb_register_node(node); |
1931 | } | |
1932 | ||
1933 | /* | |
10fbcf4c | 1934 | * Let the node device driver know we're here so it can |
9a305230 LS |
1935 | * [un]register hstate attributes on node hotplug. |
1936 | */ | |
1937 | register_hugetlbfs_with_node(hugetlb_register_node, | |
1938 | hugetlb_unregister_node); | |
1939 | } | |
1940 | #else /* !CONFIG_NUMA */ | |
1941 | ||
1942 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
1943 | { | |
1944 | BUG(); | |
1945 | if (nidp) | |
1946 | *nidp = -1; | |
1947 | return NULL; | |
1948 | } | |
1949 | ||
1950 | static void hugetlb_unregister_all_nodes(void) { } | |
1951 | ||
1952 | static void hugetlb_register_all_nodes(void) { } | |
1953 | ||
1954 | #endif | |
1955 | ||
a3437870 NA |
1956 | static void __exit hugetlb_exit(void) |
1957 | { | |
1958 | struct hstate *h; | |
1959 | ||
9a305230 LS |
1960 | hugetlb_unregister_all_nodes(); |
1961 | ||
a3437870 | 1962 | for_each_hstate(h) { |
972dc4de | 1963 | kobject_put(hstate_kobjs[hstate_index(h)]); |
a3437870 NA |
1964 | } |
1965 | ||
1966 | kobject_put(hugepages_kobj); | |
1967 | } | |
1968 | module_exit(hugetlb_exit); | |
1969 | ||
1970 | static int __init hugetlb_init(void) | |
1971 | { | |
0ef89d25 BH |
1972 | /* Some platform decide whether they support huge pages at boot |
1973 | * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when | |
1974 | * there is no such support | |
1975 | */ | |
1976 | if (HPAGE_SHIFT == 0) | |
1977 | return 0; | |
a3437870 | 1978 | |
e11bfbfc NP |
1979 | if (!size_to_hstate(default_hstate_size)) { |
1980 | default_hstate_size = HPAGE_SIZE; | |
1981 | if (!size_to_hstate(default_hstate_size)) | |
1982 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 1983 | } |
972dc4de | 1984 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
e11bfbfc NP |
1985 | if (default_hstate_max_huge_pages) |
1986 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
a3437870 NA |
1987 | |
1988 | hugetlb_init_hstates(); | |
aa888a74 | 1989 | gather_bootmem_prealloc(); |
a3437870 NA |
1990 | report_hugepages(); |
1991 | ||
1992 | hugetlb_sysfs_init(); | |
9a305230 | 1993 | hugetlb_register_all_nodes(); |
7179e7bf | 1994 | hugetlb_cgroup_file_init(); |
9a305230 | 1995 | |
a3437870 NA |
1996 | return 0; |
1997 | } | |
1998 | module_init(hugetlb_init); | |
1999 | ||
2000 | /* Should be called on processing a hugepagesz=... option */ | |
2001 | void __init hugetlb_add_hstate(unsigned order) | |
2002 | { | |
2003 | struct hstate *h; | |
8faa8b07 AK |
2004 | unsigned long i; |
2005 | ||
a3437870 | 2006 | if (size_to_hstate(PAGE_SIZE << order)) { |
ffb22af5 | 2007 | pr_warning("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
2008 | return; |
2009 | } | |
47d38344 | 2010 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 2011 | BUG_ON(order == 0); |
47d38344 | 2012 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
2013 | h->order = order; |
2014 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
2015 | h->nr_huge_pages = 0; |
2016 | h->free_huge_pages = 0; | |
2017 | for (i = 0; i < MAX_NUMNODES; ++i) | |
2018 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 2019 | INIT_LIST_HEAD(&h->hugepage_activelist); |
8cebfcd0 LJ |
2020 | h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); |
2021 | h->next_nid_to_free = first_node(node_states[N_MEMORY]); | |
a3437870 NA |
2022 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
2023 | huge_page_size(h)/1024); | |
8faa8b07 | 2024 | |
a3437870 NA |
2025 | parsed_hstate = h; |
2026 | } | |
2027 | ||
e11bfbfc | 2028 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
2029 | { |
2030 | unsigned long *mhp; | |
8faa8b07 | 2031 | static unsigned long *last_mhp; |
a3437870 NA |
2032 | |
2033 | /* | |
47d38344 | 2034 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
2035 | * so this hugepages= parameter goes to the "default hstate". |
2036 | */ | |
47d38344 | 2037 | if (!hugetlb_max_hstate) |
a3437870 NA |
2038 | mhp = &default_hstate_max_huge_pages; |
2039 | else | |
2040 | mhp = &parsed_hstate->max_huge_pages; | |
2041 | ||
8faa8b07 | 2042 | if (mhp == last_mhp) { |
ffb22af5 AM |
2043 | pr_warning("hugepages= specified twice without " |
2044 | "interleaving hugepagesz=, ignoring\n"); | |
8faa8b07 AK |
2045 | return 1; |
2046 | } | |
2047 | ||
a3437870 NA |
2048 | if (sscanf(s, "%lu", mhp) <= 0) |
2049 | *mhp = 0; | |
2050 | ||
8faa8b07 AK |
2051 | /* |
2052 | * Global state is always initialized later in hugetlb_init. | |
2053 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
2054 | * use the bootmem allocator. | |
2055 | */ | |
47d38344 | 2056 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
2057 | hugetlb_hstate_alloc_pages(parsed_hstate); |
2058 | ||
2059 | last_mhp = mhp; | |
2060 | ||
a3437870 NA |
2061 | return 1; |
2062 | } | |
e11bfbfc NP |
2063 | __setup("hugepages=", hugetlb_nrpages_setup); |
2064 | ||
2065 | static int __init hugetlb_default_setup(char *s) | |
2066 | { | |
2067 | default_hstate_size = memparse(s, &s); | |
2068 | return 1; | |
2069 | } | |
2070 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 2071 | |
8a213460 NA |
2072 | static unsigned int cpuset_mems_nr(unsigned int *array) |
2073 | { | |
2074 | int node; | |
2075 | unsigned int nr = 0; | |
2076 | ||
2077 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
2078 | nr += array[node]; | |
2079 | ||
2080 | return nr; | |
2081 | } | |
2082 | ||
2083 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
2084 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
2085 | struct ctl_table *table, int write, | |
2086 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 2087 | { |
e5ff2159 AK |
2088 | struct hstate *h = &default_hstate; |
2089 | unsigned long tmp; | |
08d4a246 | 2090 | int ret; |
e5ff2159 | 2091 | |
c033a93c | 2092 | tmp = h->max_huge_pages; |
e5ff2159 | 2093 | |
adbe8726 EM |
2094 | if (write && h->order >= MAX_ORDER) |
2095 | return -EINVAL; | |
2096 | ||
e5ff2159 AK |
2097 | table->data = &tmp; |
2098 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2099 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2100 | if (ret) | |
2101 | goto out; | |
e5ff2159 | 2102 | |
06808b08 | 2103 | if (write) { |
bad44b5b DR |
2104 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, |
2105 | GFP_KERNEL | __GFP_NORETRY); | |
06808b08 LS |
2106 | if (!(obey_mempolicy && |
2107 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
2108 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 2109 | nodes_allowed = &node_states[N_MEMORY]; |
06808b08 LS |
2110 | } |
2111 | h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); | |
2112 | ||
8cebfcd0 | 2113 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
2114 | NODEMASK_FREE(nodes_allowed); |
2115 | } | |
08d4a246 MH |
2116 | out: |
2117 | return ret; | |
1da177e4 | 2118 | } |
396faf03 | 2119 | |
06808b08 LS |
2120 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2121 | void __user *buffer, size_t *length, loff_t *ppos) | |
2122 | { | |
2123 | ||
2124 | return hugetlb_sysctl_handler_common(false, table, write, | |
2125 | buffer, length, ppos); | |
2126 | } | |
2127 | ||
2128 | #ifdef CONFIG_NUMA | |
2129 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2130 | void __user *buffer, size_t *length, loff_t *ppos) | |
2131 | { | |
2132 | return hugetlb_sysctl_handler_common(true, table, write, | |
2133 | buffer, length, ppos); | |
2134 | } | |
2135 | #endif /* CONFIG_NUMA */ | |
2136 | ||
a3d0c6aa | 2137 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 2138 | void __user *buffer, |
a3d0c6aa NA |
2139 | size_t *length, loff_t *ppos) |
2140 | { | |
a5516438 | 2141 | struct hstate *h = &default_hstate; |
e5ff2159 | 2142 | unsigned long tmp; |
08d4a246 | 2143 | int ret; |
e5ff2159 | 2144 | |
c033a93c | 2145 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 2146 | |
adbe8726 EM |
2147 | if (write && h->order >= MAX_ORDER) |
2148 | return -EINVAL; | |
2149 | ||
e5ff2159 AK |
2150 | table->data = &tmp; |
2151 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2152 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2153 | if (ret) | |
2154 | goto out; | |
e5ff2159 AK |
2155 | |
2156 | if (write) { | |
2157 | spin_lock(&hugetlb_lock); | |
2158 | h->nr_overcommit_huge_pages = tmp; | |
2159 | spin_unlock(&hugetlb_lock); | |
2160 | } | |
08d4a246 MH |
2161 | out: |
2162 | return ret; | |
a3d0c6aa NA |
2163 | } |
2164 | ||
1da177e4 LT |
2165 | #endif /* CONFIG_SYSCTL */ |
2166 | ||
e1759c21 | 2167 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 2168 | { |
a5516438 | 2169 | struct hstate *h = &default_hstate; |
e1759c21 | 2170 | seq_printf(m, |
4f98a2fe RR |
2171 | "HugePages_Total: %5lu\n" |
2172 | "HugePages_Free: %5lu\n" | |
2173 | "HugePages_Rsvd: %5lu\n" | |
2174 | "HugePages_Surp: %5lu\n" | |
2175 | "Hugepagesize: %8lu kB\n", | |
a5516438 AK |
2176 | h->nr_huge_pages, |
2177 | h->free_huge_pages, | |
2178 | h->resv_huge_pages, | |
2179 | h->surplus_huge_pages, | |
2180 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
2181 | } |
2182 | ||
2183 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
2184 | { | |
a5516438 | 2185 | struct hstate *h = &default_hstate; |
1da177e4 LT |
2186 | return sprintf(buf, |
2187 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
2188 | "Node %d HugePages_Free: %5u\n" |
2189 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
2190 | nid, h->nr_huge_pages_node[nid], |
2191 | nid, h->free_huge_pages_node[nid], | |
2192 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
2193 | } |
2194 | ||
949f7ec5 DR |
2195 | void hugetlb_show_meminfo(void) |
2196 | { | |
2197 | struct hstate *h; | |
2198 | int nid; | |
2199 | ||
2200 | for_each_node_state(nid, N_MEMORY) | |
2201 | for_each_hstate(h) | |
2202 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
2203 | nid, | |
2204 | h->nr_huge_pages_node[nid], | |
2205 | h->free_huge_pages_node[nid], | |
2206 | h->surplus_huge_pages_node[nid], | |
2207 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
2208 | } | |
2209 | ||
1da177e4 LT |
2210 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
2211 | unsigned long hugetlb_total_pages(void) | |
2212 | { | |
d0028588 WL |
2213 | struct hstate *h; |
2214 | unsigned long nr_total_pages = 0; | |
2215 | ||
2216 | for_each_hstate(h) | |
2217 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
2218 | return nr_total_pages; | |
1da177e4 | 2219 | } |
1da177e4 | 2220 | |
a5516438 | 2221 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
2222 | { |
2223 | int ret = -ENOMEM; | |
2224 | ||
2225 | spin_lock(&hugetlb_lock); | |
2226 | /* | |
2227 | * When cpuset is configured, it breaks the strict hugetlb page | |
2228 | * reservation as the accounting is done on a global variable. Such | |
2229 | * reservation is completely rubbish in the presence of cpuset because | |
2230 | * the reservation is not checked against page availability for the | |
2231 | * current cpuset. Application can still potentially OOM'ed by kernel | |
2232 | * with lack of free htlb page in cpuset that the task is in. | |
2233 | * Attempt to enforce strict accounting with cpuset is almost | |
2234 | * impossible (or too ugly) because cpuset is too fluid that | |
2235 | * task or memory node can be dynamically moved between cpusets. | |
2236 | * | |
2237 | * The change of semantics for shared hugetlb mapping with cpuset is | |
2238 | * undesirable. However, in order to preserve some of the semantics, | |
2239 | * we fall back to check against current free page availability as | |
2240 | * a best attempt and hopefully to minimize the impact of changing | |
2241 | * semantics that cpuset has. | |
2242 | */ | |
2243 | if (delta > 0) { | |
a5516438 | 2244 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
2245 | goto out; |
2246 | ||
a5516438 AK |
2247 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
2248 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
2249 | goto out; |
2250 | } | |
2251 | } | |
2252 | ||
2253 | ret = 0; | |
2254 | if (delta < 0) | |
a5516438 | 2255 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
2256 | |
2257 | out: | |
2258 | spin_unlock(&hugetlb_lock); | |
2259 | return ret; | |
2260 | } | |
2261 | ||
84afd99b AW |
2262 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
2263 | { | |
f522c3ac | 2264 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
2265 | |
2266 | /* | |
2267 | * This new VMA should share its siblings reservation map if present. | |
2268 | * The VMA will only ever have a valid reservation map pointer where | |
2269 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 2270 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
2271 | * after this open call completes. It is therefore safe to take a |
2272 | * new reference here without additional locking. | |
2273 | */ | |
f522c3ac JK |
2274 | if (resv) |
2275 | kref_get(&resv->refs); | |
84afd99b AW |
2276 | } |
2277 | ||
a1e78772 MG |
2278 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
2279 | { | |
a5516438 | 2280 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 2281 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 2282 | struct hugepage_subpool *spool = subpool_vma(vma); |
84afd99b AW |
2283 | unsigned long reserve; |
2284 | unsigned long start; | |
2285 | unsigned long end; | |
2286 | ||
f522c3ac | 2287 | if (resv) { |
a5516438 AK |
2288 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
2289 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b AW |
2290 | |
2291 | reserve = (end - start) - | |
1406ec9b | 2292 | region_count(resv, start, end); |
84afd99b | 2293 | |
f031dd27 | 2294 | kref_put(&resv->refs, resv_map_release); |
84afd99b | 2295 | |
7251ff78 | 2296 | if (reserve) { |
a5516438 | 2297 | hugetlb_acct_memory(h, -reserve); |
90481622 | 2298 | hugepage_subpool_put_pages(spool, reserve); |
7251ff78 | 2299 | } |
84afd99b | 2300 | } |
a1e78772 MG |
2301 | } |
2302 | ||
1da177e4 LT |
2303 | /* |
2304 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
2305 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
2306 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
2307 | * this far. | |
2308 | */ | |
d0217ac0 | 2309 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
2310 | { |
2311 | BUG(); | |
d0217ac0 | 2312 | return 0; |
1da177e4 LT |
2313 | } |
2314 | ||
f0f37e2f | 2315 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 2316 | .fault = hugetlb_vm_op_fault, |
84afd99b | 2317 | .open = hugetlb_vm_op_open, |
a1e78772 | 2318 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
2319 | }; |
2320 | ||
1e8f889b DG |
2321 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
2322 | int writable) | |
63551ae0 DG |
2323 | { |
2324 | pte_t entry; | |
2325 | ||
1e8f889b | 2326 | if (writable) { |
106c992a GS |
2327 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
2328 | vma->vm_page_prot))); | |
63551ae0 | 2329 | } else { |
106c992a GS |
2330 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
2331 | vma->vm_page_prot)); | |
63551ae0 DG |
2332 | } |
2333 | entry = pte_mkyoung(entry); | |
2334 | entry = pte_mkhuge(entry); | |
d9ed9faa | 2335 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
2336 | |
2337 | return entry; | |
2338 | } | |
2339 | ||
1e8f889b DG |
2340 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
2341 | unsigned long address, pte_t *ptep) | |
2342 | { | |
2343 | pte_t entry; | |
2344 | ||
106c992a | 2345 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 2346 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 2347 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
2348 | } |
2349 | ||
2350 | ||
63551ae0 DG |
2351 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
2352 | struct vm_area_struct *vma) | |
2353 | { | |
2354 | pte_t *src_pte, *dst_pte, entry; | |
2355 | struct page *ptepage; | |
1c59827d | 2356 | unsigned long addr; |
1e8f889b | 2357 | int cow; |
a5516438 AK |
2358 | struct hstate *h = hstate_vma(vma); |
2359 | unsigned long sz = huge_page_size(h); | |
e8569dd2 AS |
2360 | unsigned long mmun_start; /* For mmu_notifiers */ |
2361 | unsigned long mmun_end; /* For mmu_notifiers */ | |
2362 | int ret = 0; | |
1e8f889b DG |
2363 | |
2364 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 2365 | |
e8569dd2 AS |
2366 | mmun_start = vma->vm_start; |
2367 | mmun_end = vma->vm_end; | |
2368 | if (cow) | |
2369 | mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); | |
2370 | ||
a5516438 | 2371 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 2372 | spinlock_t *src_ptl, *dst_ptl; |
c74df32c HD |
2373 | src_pte = huge_pte_offset(src, addr); |
2374 | if (!src_pte) | |
2375 | continue; | |
a5516438 | 2376 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
2377 | if (!dst_pte) { |
2378 | ret = -ENOMEM; | |
2379 | break; | |
2380 | } | |
c5c99429 LW |
2381 | |
2382 | /* If the pagetables are shared don't copy or take references */ | |
2383 | if (dst_pte == src_pte) | |
2384 | continue; | |
2385 | ||
cb900f41 KS |
2386 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
2387 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
2388 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
7f2e9525 | 2389 | if (!huge_pte_none(huge_ptep_get(src_pte))) { |
1e8f889b | 2390 | if (cow) |
7f2e9525 GS |
2391 | huge_ptep_set_wrprotect(src, addr, src_pte); |
2392 | entry = huge_ptep_get(src_pte); | |
1c59827d HD |
2393 | ptepage = pte_page(entry); |
2394 | get_page(ptepage); | |
0fe6e20b | 2395 | page_dup_rmap(ptepage); |
1c59827d HD |
2396 | set_huge_pte_at(dst, addr, dst_pte, entry); |
2397 | } | |
cb900f41 KS |
2398 | spin_unlock(src_ptl); |
2399 | spin_unlock(dst_ptl); | |
63551ae0 | 2400 | } |
63551ae0 | 2401 | |
e8569dd2 AS |
2402 | if (cow) |
2403 | mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); | |
2404 | ||
2405 | return ret; | |
63551ae0 DG |
2406 | } |
2407 | ||
290408d4 NH |
2408 | static int is_hugetlb_entry_migration(pte_t pte) |
2409 | { | |
2410 | swp_entry_t swp; | |
2411 | ||
2412 | if (huge_pte_none(pte) || pte_present(pte)) | |
2413 | return 0; | |
2414 | swp = pte_to_swp_entry(pte); | |
32f84528 | 2415 | if (non_swap_entry(swp) && is_migration_entry(swp)) |
290408d4 | 2416 | return 1; |
32f84528 | 2417 | else |
290408d4 NH |
2418 | return 0; |
2419 | } | |
2420 | ||
fd6a03ed NH |
2421 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) |
2422 | { | |
2423 | swp_entry_t swp; | |
2424 | ||
2425 | if (huge_pte_none(pte) || pte_present(pte)) | |
2426 | return 0; | |
2427 | swp = pte_to_swp_entry(pte); | |
32f84528 | 2428 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) |
fd6a03ed | 2429 | return 1; |
32f84528 | 2430 | else |
fd6a03ed NH |
2431 | return 0; |
2432 | } | |
2433 | ||
24669e58 AK |
2434 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
2435 | unsigned long start, unsigned long end, | |
2436 | struct page *ref_page) | |
63551ae0 | 2437 | { |
24669e58 | 2438 | int force_flush = 0; |
63551ae0 DG |
2439 | struct mm_struct *mm = vma->vm_mm; |
2440 | unsigned long address; | |
c7546f8f | 2441 | pte_t *ptep; |
63551ae0 | 2442 | pte_t pte; |
cb900f41 | 2443 | spinlock_t *ptl; |
63551ae0 | 2444 | struct page *page; |
a5516438 AK |
2445 | struct hstate *h = hstate_vma(vma); |
2446 | unsigned long sz = huge_page_size(h); | |
2ec74c3e SG |
2447 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
2448 | const unsigned long mmun_end = end; /* For mmu_notifiers */ | |
a5516438 | 2449 | |
63551ae0 | 2450 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
2451 | BUG_ON(start & ~huge_page_mask(h)); |
2452 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 2453 | |
24669e58 | 2454 | tlb_start_vma(tlb, vma); |
2ec74c3e | 2455 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
24669e58 | 2456 | again: |
a5516438 | 2457 | for (address = start; address < end; address += sz) { |
c7546f8f | 2458 | ptep = huge_pte_offset(mm, address); |
4c887265 | 2459 | if (!ptep) |
c7546f8f DG |
2460 | continue; |
2461 | ||
cb900f41 | 2462 | ptl = huge_pte_lock(h, mm, ptep); |
39dde65c | 2463 | if (huge_pmd_unshare(mm, &address, ptep)) |
cb900f41 | 2464 | goto unlock; |
39dde65c | 2465 | |
6629326b HD |
2466 | pte = huge_ptep_get(ptep); |
2467 | if (huge_pte_none(pte)) | |
cb900f41 | 2468 | goto unlock; |
6629326b HD |
2469 | |
2470 | /* | |
2471 | * HWPoisoned hugepage is already unmapped and dropped reference | |
2472 | */ | |
8c4894c6 | 2473 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { |
106c992a | 2474 | huge_pte_clear(mm, address, ptep); |
cb900f41 | 2475 | goto unlock; |
8c4894c6 | 2476 | } |
6629326b HD |
2477 | |
2478 | page = pte_page(pte); | |
04f2cbe3 MG |
2479 | /* |
2480 | * If a reference page is supplied, it is because a specific | |
2481 | * page is being unmapped, not a range. Ensure the page we | |
2482 | * are about to unmap is the actual page of interest. | |
2483 | */ | |
2484 | if (ref_page) { | |
04f2cbe3 | 2485 | if (page != ref_page) |
cb900f41 | 2486 | goto unlock; |
04f2cbe3 MG |
2487 | |
2488 | /* | |
2489 | * Mark the VMA as having unmapped its page so that | |
2490 | * future faults in this VMA will fail rather than | |
2491 | * looking like data was lost | |
2492 | */ | |
2493 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
2494 | } | |
2495 | ||
c7546f8f | 2496 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
24669e58 | 2497 | tlb_remove_tlb_entry(tlb, ptep, address); |
106c992a | 2498 | if (huge_pte_dirty(pte)) |
6649a386 | 2499 | set_page_dirty(page); |
9e81130b | 2500 | |
24669e58 AK |
2501 | page_remove_rmap(page); |
2502 | force_flush = !__tlb_remove_page(tlb, page); | |
cb900f41 KS |
2503 | if (force_flush) { |
2504 | spin_unlock(ptl); | |
24669e58 | 2505 | break; |
cb900f41 | 2506 | } |
9e81130b | 2507 | /* Bail out after unmapping reference page if supplied */ |
cb900f41 KS |
2508 | if (ref_page) { |
2509 | spin_unlock(ptl); | |
9e81130b | 2510 | break; |
cb900f41 KS |
2511 | } |
2512 | unlock: | |
2513 | spin_unlock(ptl); | |
63551ae0 | 2514 | } |
24669e58 AK |
2515 | /* |
2516 | * mmu_gather ran out of room to batch pages, we break out of | |
2517 | * the PTE lock to avoid doing the potential expensive TLB invalidate | |
2518 | * and page-free while holding it. | |
2519 | */ | |
2520 | if (force_flush) { | |
2521 | force_flush = 0; | |
2522 | tlb_flush_mmu(tlb); | |
2523 | if (address < end && !ref_page) | |
2524 | goto again; | |
fe1668ae | 2525 | } |
2ec74c3e | 2526 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
24669e58 | 2527 | tlb_end_vma(tlb, vma); |
1da177e4 | 2528 | } |
63551ae0 | 2529 | |
d833352a MG |
2530 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
2531 | struct vm_area_struct *vma, unsigned long start, | |
2532 | unsigned long end, struct page *ref_page) | |
2533 | { | |
2534 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
2535 | ||
2536 | /* | |
2537 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
2538 | * test will fail on a vma being torn down, and not grab a page table | |
2539 | * on its way out. We're lucky that the flag has such an appropriate | |
2540 | * name, and can in fact be safely cleared here. We could clear it | |
2541 | * before the __unmap_hugepage_range above, but all that's necessary | |
2542 | * is to clear it before releasing the i_mmap_mutex. This works | |
2543 | * because in the context this is called, the VMA is about to be | |
2544 | * destroyed and the i_mmap_mutex is held. | |
2545 | */ | |
2546 | vma->vm_flags &= ~VM_MAYSHARE; | |
2547 | } | |
2548 | ||
502717f4 | 2549 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 2550 | unsigned long end, struct page *ref_page) |
502717f4 | 2551 | { |
24669e58 AK |
2552 | struct mm_struct *mm; |
2553 | struct mmu_gather tlb; | |
2554 | ||
2555 | mm = vma->vm_mm; | |
2556 | ||
2b047252 | 2557 | tlb_gather_mmu(&tlb, mm, start, end); |
24669e58 AK |
2558 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
2559 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 KC |
2560 | } |
2561 | ||
04f2cbe3 MG |
2562 | /* |
2563 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
2564 | * mappping it owns the reserve page for. The intention is to unmap the page | |
2565 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
2566 | * same region. | |
2567 | */ | |
2a4b3ded HH |
2568 | static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
2569 | struct page *page, unsigned long address) | |
04f2cbe3 | 2570 | { |
7526674d | 2571 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
2572 | struct vm_area_struct *iter_vma; |
2573 | struct address_space *mapping; | |
04f2cbe3 MG |
2574 | pgoff_t pgoff; |
2575 | ||
2576 | /* | |
2577 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
2578 | * from page cache lookup which is in HPAGE_SIZE units. | |
2579 | */ | |
7526674d | 2580 | address = address & huge_page_mask(h); |
36e4f20a MH |
2581 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
2582 | vma->vm_pgoff; | |
496ad9aa | 2583 | mapping = file_inode(vma->vm_file)->i_mapping; |
04f2cbe3 | 2584 | |
4eb2b1dc MG |
2585 | /* |
2586 | * Take the mapping lock for the duration of the table walk. As | |
2587 | * this mapping should be shared between all the VMAs, | |
2588 | * __unmap_hugepage_range() is called as the lock is already held | |
2589 | */ | |
3d48ae45 | 2590 | mutex_lock(&mapping->i_mmap_mutex); |
6b2dbba8 | 2591 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
2592 | /* Do not unmap the current VMA */ |
2593 | if (iter_vma == vma) | |
2594 | continue; | |
2595 | ||
2596 | /* | |
2597 | * Unmap the page from other VMAs without their own reserves. | |
2598 | * They get marked to be SIGKILLed if they fault in these | |
2599 | * areas. This is because a future no-page fault on this VMA | |
2600 | * could insert a zeroed page instead of the data existing | |
2601 | * from the time of fork. This would look like data corruption | |
2602 | */ | |
2603 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
2604 | unmap_hugepage_range(iter_vma, address, |
2605 | address + huge_page_size(h), page); | |
04f2cbe3 | 2606 | } |
3d48ae45 | 2607 | mutex_unlock(&mapping->i_mmap_mutex); |
04f2cbe3 MG |
2608 | |
2609 | return 1; | |
2610 | } | |
2611 | ||
0fe6e20b NH |
2612 | /* |
2613 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
2614 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
2615 | * cannot race with other handlers or page migration. | |
2616 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 2617 | */ |
1e8f889b | 2618 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
04f2cbe3 | 2619 | unsigned long address, pte_t *ptep, pte_t pte, |
cb900f41 | 2620 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 2621 | { |
a5516438 | 2622 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 2623 | struct page *old_page, *new_page; |
04f2cbe3 | 2624 | int outside_reserve = 0; |
2ec74c3e SG |
2625 | unsigned long mmun_start; /* For mmu_notifiers */ |
2626 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1e8f889b DG |
2627 | |
2628 | old_page = pte_page(pte); | |
2629 | ||
04f2cbe3 | 2630 | retry_avoidcopy: |
1e8f889b DG |
2631 | /* If no-one else is actually using this page, avoid the copy |
2632 | * and just make the page writable */ | |
37a2140d JK |
2633 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
2634 | page_move_anon_rmap(old_page, vma, address); | |
1e8f889b | 2635 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 2636 | return 0; |
1e8f889b DG |
2637 | } |
2638 | ||
04f2cbe3 MG |
2639 | /* |
2640 | * If the process that created a MAP_PRIVATE mapping is about to | |
2641 | * perform a COW due to a shared page count, attempt to satisfy | |
2642 | * the allocation without using the existing reserves. The pagecache | |
2643 | * page is used to determine if the reserve at this address was | |
2644 | * consumed or not. If reserves were used, a partial faulted mapping | |
2645 | * at the time of fork() could consume its reserves on COW instead | |
2646 | * of the full address range. | |
2647 | */ | |
5944d011 | 2648 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
2649 | old_page != pagecache_page) |
2650 | outside_reserve = 1; | |
2651 | ||
1e8f889b | 2652 | page_cache_get(old_page); |
b76c8cfb | 2653 | |
cb900f41 KS |
2654 | /* Drop page table lock as buddy allocator may be called */ |
2655 | spin_unlock(ptl); | |
04f2cbe3 | 2656 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 2657 | |
2fc39cec | 2658 | if (IS_ERR(new_page)) { |
76dcee75 | 2659 | long err = PTR_ERR(new_page); |
1e8f889b | 2660 | page_cache_release(old_page); |
04f2cbe3 MG |
2661 | |
2662 | /* | |
2663 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
2664 | * it is due to references held by a child and an insufficient | |
2665 | * huge page pool. To guarantee the original mappers | |
2666 | * reliability, unmap the page from child processes. The child | |
2667 | * may get SIGKILLed if it later faults. | |
2668 | */ | |
2669 | if (outside_reserve) { | |
2670 | BUG_ON(huge_pte_none(pte)); | |
2671 | if (unmap_ref_private(mm, vma, old_page, address)) { | |
04f2cbe3 | 2672 | BUG_ON(huge_pte_none(pte)); |
cb900f41 | 2673 | spin_lock(ptl); |
a734bcc8 HD |
2674 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
2675 | if (likely(pte_same(huge_ptep_get(ptep), pte))) | |
2676 | goto retry_avoidcopy; | |
2677 | /* | |
cb900f41 KS |
2678 | * race occurs while re-acquiring page table |
2679 | * lock, and our job is done. | |
a734bcc8 HD |
2680 | */ |
2681 | return 0; | |
04f2cbe3 MG |
2682 | } |
2683 | WARN_ON_ONCE(1); | |
2684 | } | |
2685 | ||
b76c8cfb | 2686 | /* Caller expects lock to be held */ |
cb900f41 | 2687 | spin_lock(ptl); |
76dcee75 AK |
2688 | if (err == -ENOMEM) |
2689 | return VM_FAULT_OOM; | |
2690 | else | |
2691 | return VM_FAULT_SIGBUS; | |
1e8f889b DG |
2692 | } |
2693 | ||
0fe6e20b NH |
2694 | /* |
2695 | * When the original hugepage is shared one, it does not have | |
2696 | * anon_vma prepared. | |
2697 | */ | |
44e2aa93 | 2698 | if (unlikely(anon_vma_prepare(vma))) { |
ea4039a3 HD |
2699 | page_cache_release(new_page); |
2700 | page_cache_release(old_page); | |
44e2aa93 | 2701 | /* Caller expects lock to be held */ |
cb900f41 | 2702 | spin_lock(ptl); |
0fe6e20b | 2703 | return VM_FAULT_OOM; |
44e2aa93 | 2704 | } |
0fe6e20b | 2705 | |
47ad8475 AA |
2706 | copy_user_huge_page(new_page, old_page, address, vma, |
2707 | pages_per_huge_page(h)); | |
0ed361de | 2708 | __SetPageUptodate(new_page); |
1e8f889b | 2709 | |
2ec74c3e SG |
2710 | mmun_start = address & huge_page_mask(h); |
2711 | mmun_end = mmun_start + huge_page_size(h); | |
2712 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
b76c8cfb | 2713 | /* |
cb900f41 | 2714 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
2715 | * before the page tables are altered |
2716 | */ | |
cb900f41 | 2717 | spin_lock(ptl); |
a5516438 | 2718 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
7f2e9525 | 2719 | if (likely(pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
2720 | ClearPagePrivate(new_page); |
2721 | ||
1e8f889b | 2722 | /* Break COW */ |
8fe627ec | 2723 | huge_ptep_clear_flush(vma, address, ptep); |
1e8f889b DG |
2724 | set_huge_pte_at(mm, address, ptep, |
2725 | make_huge_pte(vma, new_page, 1)); | |
0fe6e20b | 2726 | page_remove_rmap(old_page); |
cd67f0d2 | 2727 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
2728 | /* Make the old page be freed below */ |
2729 | new_page = old_page; | |
2730 | } | |
cb900f41 | 2731 | spin_unlock(ptl); |
2ec74c3e | 2732 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
1e8f889b DG |
2733 | page_cache_release(new_page); |
2734 | page_cache_release(old_page); | |
8312034f JK |
2735 | |
2736 | /* Caller expects lock to be held */ | |
cb900f41 | 2737 | spin_lock(ptl); |
83c54070 | 2738 | return 0; |
1e8f889b DG |
2739 | } |
2740 | ||
04f2cbe3 | 2741 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
2742 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
2743 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
2744 | { |
2745 | struct address_space *mapping; | |
e7c4b0bf | 2746 | pgoff_t idx; |
04f2cbe3 MG |
2747 | |
2748 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 2749 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
2750 | |
2751 | return find_lock_page(mapping, idx); | |
2752 | } | |
2753 | ||
3ae77f43 HD |
2754 | /* |
2755 | * Return whether there is a pagecache page to back given address within VMA. | |
2756 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
2757 | */ | |
2758 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
2759 | struct vm_area_struct *vma, unsigned long address) |
2760 | { | |
2761 | struct address_space *mapping; | |
2762 | pgoff_t idx; | |
2763 | struct page *page; | |
2764 | ||
2765 | mapping = vma->vm_file->f_mapping; | |
2766 | idx = vma_hugecache_offset(h, vma, address); | |
2767 | ||
2768 | page = find_get_page(mapping, idx); | |
2769 | if (page) | |
2770 | put_page(page); | |
2771 | return page != NULL; | |
2772 | } | |
2773 | ||
a1ed3dda | 2774 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 2775 | unsigned long address, pte_t *ptep, unsigned int flags) |
ac9b9c66 | 2776 | { |
a5516438 | 2777 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 2778 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 2779 | int anon_rmap = 0; |
e7c4b0bf | 2780 | pgoff_t idx; |
4c887265 | 2781 | unsigned long size; |
4c887265 AL |
2782 | struct page *page; |
2783 | struct address_space *mapping; | |
1e8f889b | 2784 | pte_t new_pte; |
cb900f41 | 2785 | spinlock_t *ptl; |
4c887265 | 2786 | |
04f2cbe3 MG |
2787 | /* |
2788 | * Currently, we are forced to kill the process in the event the | |
2789 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 2790 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
2791 | */ |
2792 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
ffb22af5 AM |
2793 | pr_warning("PID %d killed due to inadequate hugepage pool\n", |
2794 | current->pid); | |
04f2cbe3 MG |
2795 | return ret; |
2796 | } | |
2797 | ||
4c887265 | 2798 | mapping = vma->vm_file->f_mapping; |
a5516438 | 2799 | idx = vma_hugecache_offset(h, vma, address); |
4c887265 AL |
2800 | |
2801 | /* | |
2802 | * Use page lock to guard against racing truncation | |
2803 | * before we get page_table_lock. | |
2804 | */ | |
6bda666a CL |
2805 | retry: |
2806 | page = find_lock_page(mapping, idx); | |
2807 | if (!page) { | |
a5516438 | 2808 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
2809 | if (idx >= size) |
2810 | goto out; | |
04f2cbe3 | 2811 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 2812 | if (IS_ERR(page)) { |
76dcee75 AK |
2813 | ret = PTR_ERR(page); |
2814 | if (ret == -ENOMEM) | |
2815 | ret = VM_FAULT_OOM; | |
2816 | else | |
2817 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
2818 | goto out; |
2819 | } | |
47ad8475 | 2820 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 2821 | __SetPageUptodate(page); |
ac9b9c66 | 2822 | |
f83a275d | 2823 | if (vma->vm_flags & VM_MAYSHARE) { |
6bda666a | 2824 | int err; |
45c682a6 | 2825 | struct inode *inode = mapping->host; |
6bda666a CL |
2826 | |
2827 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
2828 | if (err) { | |
2829 | put_page(page); | |
6bda666a CL |
2830 | if (err == -EEXIST) |
2831 | goto retry; | |
2832 | goto out; | |
2833 | } | |
07443a85 | 2834 | ClearPagePrivate(page); |
45c682a6 KC |
2835 | |
2836 | spin_lock(&inode->i_lock); | |
a5516438 | 2837 | inode->i_blocks += blocks_per_huge_page(h); |
45c682a6 | 2838 | spin_unlock(&inode->i_lock); |
23be7468 | 2839 | } else { |
6bda666a | 2840 | lock_page(page); |
0fe6e20b NH |
2841 | if (unlikely(anon_vma_prepare(vma))) { |
2842 | ret = VM_FAULT_OOM; | |
2843 | goto backout_unlocked; | |
2844 | } | |
409eb8c2 | 2845 | anon_rmap = 1; |
23be7468 | 2846 | } |
0fe6e20b | 2847 | } else { |
998b4382 NH |
2848 | /* |
2849 | * If memory error occurs between mmap() and fault, some process | |
2850 | * don't have hwpoisoned swap entry for errored virtual address. | |
2851 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
2852 | */ | |
2853 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 2854 | ret = VM_FAULT_HWPOISON | |
972dc4de | 2855 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
2856 | goto backout_unlocked; |
2857 | } | |
6bda666a | 2858 | } |
1e8f889b | 2859 | |
57303d80 AW |
2860 | /* |
2861 | * If we are going to COW a private mapping later, we examine the | |
2862 | * pending reservations for this page now. This will ensure that | |
2863 | * any allocations necessary to record that reservation occur outside | |
2864 | * the spinlock. | |
2865 | */ | |
788c7df4 | 2866 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) |
2b26736c AW |
2867 | if (vma_needs_reservation(h, vma, address) < 0) { |
2868 | ret = VM_FAULT_OOM; | |
2869 | goto backout_unlocked; | |
2870 | } | |
57303d80 | 2871 | |
cb900f41 KS |
2872 | ptl = huge_pte_lockptr(h, mm, ptep); |
2873 | spin_lock(ptl); | |
a5516438 | 2874 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
2875 | if (idx >= size) |
2876 | goto backout; | |
2877 | ||
83c54070 | 2878 | ret = 0; |
7f2e9525 | 2879 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
2880 | goto backout; |
2881 | ||
07443a85 JK |
2882 | if (anon_rmap) { |
2883 | ClearPagePrivate(page); | |
409eb8c2 | 2884 | hugepage_add_new_anon_rmap(page, vma, address); |
07443a85 | 2885 | } |
409eb8c2 HD |
2886 | else |
2887 | page_dup_rmap(page); | |
1e8f889b DG |
2888 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
2889 | && (vma->vm_flags & VM_SHARED))); | |
2890 | set_huge_pte_at(mm, address, ptep, new_pte); | |
2891 | ||
788c7df4 | 2892 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 2893 | /* Optimization, do the COW without a second fault */ |
cb900f41 | 2894 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); |
1e8f889b DG |
2895 | } |
2896 | ||
cb900f41 | 2897 | spin_unlock(ptl); |
4c887265 AL |
2898 | unlock_page(page); |
2899 | out: | |
ac9b9c66 | 2900 | return ret; |
4c887265 AL |
2901 | |
2902 | backout: | |
cb900f41 | 2903 | spin_unlock(ptl); |
2b26736c | 2904 | backout_unlocked: |
4c887265 AL |
2905 | unlock_page(page); |
2906 | put_page(page); | |
2907 | goto out; | |
ac9b9c66 HD |
2908 | } |
2909 | ||
86e5216f | 2910 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 2911 | unsigned long address, unsigned int flags) |
86e5216f AL |
2912 | { |
2913 | pte_t *ptep; | |
2914 | pte_t entry; | |
cb900f41 | 2915 | spinlock_t *ptl; |
1e8f889b | 2916 | int ret; |
0fe6e20b | 2917 | struct page *page = NULL; |
57303d80 | 2918 | struct page *pagecache_page = NULL; |
3935baa9 | 2919 | static DEFINE_MUTEX(hugetlb_instantiation_mutex); |
a5516438 | 2920 | struct hstate *h = hstate_vma(vma); |
86e5216f | 2921 | |
1e16a539 KH |
2922 | address &= huge_page_mask(h); |
2923 | ||
fd6a03ed NH |
2924 | ptep = huge_pte_offset(mm, address); |
2925 | if (ptep) { | |
2926 | entry = huge_ptep_get(ptep); | |
290408d4 | 2927 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 2928 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
2929 | return 0; |
2930 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 2931 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 2932 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
2933 | } |
2934 | ||
a5516438 | 2935 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); |
86e5216f AL |
2936 | if (!ptep) |
2937 | return VM_FAULT_OOM; | |
2938 | ||
3935baa9 DG |
2939 | /* |
2940 | * Serialize hugepage allocation and instantiation, so that we don't | |
2941 | * get spurious allocation failures if two CPUs race to instantiate | |
2942 | * the same page in the page cache. | |
2943 | */ | |
2944 | mutex_lock(&hugetlb_instantiation_mutex); | |
7f2e9525 GS |
2945 | entry = huge_ptep_get(ptep); |
2946 | if (huge_pte_none(entry)) { | |
788c7df4 | 2947 | ret = hugetlb_no_page(mm, vma, address, ptep, flags); |
b4d1d99f | 2948 | goto out_mutex; |
3935baa9 | 2949 | } |
86e5216f | 2950 | |
83c54070 | 2951 | ret = 0; |
1e8f889b | 2952 | |
57303d80 AW |
2953 | /* |
2954 | * If we are going to COW the mapping later, we examine the pending | |
2955 | * reservations for this page now. This will ensure that any | |
2956 | * allocations necessary to record that reservation occur outside the | |
2957 | * spinlock. For private mappings, we also lookup the pagecache | |
2958 | * page now as it is used to determine if a reservation has been | |
2959 | * consumed. | |
2960 | */ | |
106c992a | 2961 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
2b26736c AW |
2962 | if (vma_needs_reservation(h, vma, address) < 0) { |
2963 | ret = VM_FAULT_OOM; | |
b4d1d99f | 2964 | goto out_mutex; |
2b26736c | 2965 | } |
57303d80 | 2966 | |
f83a275d | 2967 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
2968 | pagecache_page = hugetlbfs_pagecache_page(h, |
2969 | vma, address); | |
2970 | } | |
2971 | ||
56c9cfb1 NH |
2972 | /* |
2973 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
2974 | * pagecache_page, so here we need take the former one | |
2975 | * when page != pagecache_page or !pagecache_page. | |
2976 | * Note that locking order is always pagecache_page -> page, | |
2977 | * so no worry about deadlock. | |
2978 | */ | |
2979 | page = pte_page(entry); | |
66aebce7 | 2980 | get_page(page); |
56c9cfb1 | 2981 | if (page != pagecache_page) |
0fe6e20b | 2982 | lock_page(page); |
0fe6e20b | 2983 | |
cb900f41 KS |
2984 | ptl = huge_pte_lockptr(h, mm, ptep); |
2985 | spin_lock(ptl); | |
1e8f889b | 2986 | /* Check for a racing update before calling hugetlb_cow */ |
b4d1d99f | 2987 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) |
cb900f41 | 2988 | goto out_ptl; |
b4d1d99f DG |
2989 | |
2990 | ||
788c7df4 | 2991 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 2992 | if (!huge_pte_write(entry)) { |
57303d80 | 2993 | ret = hugetlb_cow(mm, vma, address, ptep, entry, |
cb900f41 KS |
2994 | pagecache_page, ptl); |
2995 | goto out_ptl; | |
b4d1d99f | 2996 | } |
106c992a | 2997 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
2998 | } |
2999 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
3000 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
3001 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 3002 | update_mmu_cache(vma, address, ptep); |
b4d1d99f | 3003 | |
cb900f41 KS |
3004 | out_ptl: |
3005 | spin_unlock(ptl); | |
57303d80 AW |
3006 | |
3007 | if (pagecache_page) { | |
3008 | unlock_page(pagecache_page); | |
3009 | put_page(pagecache_page); | |
3010 | } | |
1f64d69c DN |
3011 | if (page != pagecache_page) |
3012 | unlock_page(page); | |
66aebce7 | 3013 | put_page(page); |
57303d80 | 3014 | |
b4d1d99f | 3015 | out_mutex: |
3935baa9 | 3016 | mutex_unlock(&hugetlb_instantiation_mutex); |
1e8f889b DG |
3017 | |
3018 | return ret; | |
86e5216f AL |
3019 | } |
3020 | ||
28a35716 ML |
3021 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
3022 | struct page **pages, struct vm_area_struct **vmas, | |
3023 | unsigned long *position, unsigned long *nr_pages, | |
3024 | long i, unsigned int flags) | |
63551ae0 | 3025 | { |
d5d4b0aa KC |
3026 | unsigned long pfn_offset; |
3027 | unsigned long vaddr = *position; | |
28a35716 | 3028 | unsigned long remainder = *nr_pages; |
a5516438 | 3029 | struct hstate *h = hstate_vma(vma); |
63551ae0 | 3030 | |
63551ae0 | 3031 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 3032 | pte_t *pte; |
cb900f41 | 3033 | spinlock_t *ptl = NULL; |
2a15efc9 | 3034 | int absent; |
4c887265 | 3035 | struct page *page; |
63551ae0 | 3036 | |
4c887265 AL |
3037 | /* |
3038 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 3039 | * each hugepage. We have to make sure we get the |
4c887265 | 3040 | * first, for the page indexing below to work. |
cb900f41 KS |
3041 | * |
3042 | * Note that page table lock is not held when pte is null. | |
4c887265 | 3043 | */ |
a5516438 | 3044 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); |
cb900f41 KS |
3045 | if (pte) |
3046 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
3047 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
3048 | ||
3049 | /* | |
3050 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
3051 | * an error where there's an empty slot with no huge pagecache |
3052 | * to back it. This way, we avoid allocating a hugepage, and | |
3053 | * the sparse dumpfile avoids allocating disk blocks, but its | |
3054 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 3055 | */ |
3ae77f43 HD |
3056 | if (absent && (flags & FOLL_DUMP) && |
3057 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
3058 | if (pte) |
3059 | spin_unlock(ptl); | |
2a15efc9 HD |
3060 | remainder = 0; |
3061 | break; | |
3062 | } | |
63551ae0 | 3063 | |
9cc3a5bd NH |
3064 | /* |
3065 | * We need call hugetlb_fault for both hugepages under migration | |
3066 | * (in which case hugetlb_fault waits for the migration,) and | |
3067 | * hwpoisoned hugepages (in which case we need to prevent the | |
3068 | * caller from accessing to them.) In order to do this, we use | |
3069 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
3070 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
3071 | * both cases, and because we can't follow correct pages | |
3072 | * directly from any kind of swap entries. | |
3073 | */ | |
3074 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
3075 | ((flags & FOLL_WRITE) && |
3076 | !huge_pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 3077 | int ret; |
63551ae0 | 3078 | |
cb900f41 KS |
3079 | if (pte) |
3080 | spin_unlock(ptl); | |
2a15efc9 HD |
3081 | ret = hugetlb_fault(mm, vma, vaddr, |
3082 | (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); | |
a89182c7 | 3083 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 3084 | continue; |
63551ae0 | 3085 | |
4c887265 | 3086 | remainder = 0; |
4c887265 AL |
3087 | break; |
3088 | } | |
3089 | ||
a5516438 | 3090 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 3091 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 3092 | same_page: |
d6692183 | 3093 | if (pages) { |
2a15efc9 | 3094 | pages[i] = mem_map_offset(page, pfn_offset); |
a0368d4e | 3095 | get_page_foll(pages[i]); |
d6692183 | 3096 | } |
63551ae0 DG |
3097 | |
3098 | if (vmas) | |
3099 | vmas[i] = vma; | |
3100 | ||
3101 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 3102 | ++pfn_offset; |
63551ae0 DG |
3103 | --remainder; |
3104 | ++i; | |
d5d4b0aa | 3105 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 3106 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa KC |
3107 | /* |
3108 | * We use pfn_offset to avoid touching the pageframes | |
3109 | * of this compound page. | |
3110 | */ | |
3111 | goto same_page; | |
3112 | } | |
cb900f41 | 3113 | spin_unlock(ptl); |
63551ae0 | 3114 | } |
28a35716 | 3115 | *nr_pages = remainder; |
63551ae0 DG |
3116 | *position = vaddr; |
3117 | ||
2a15efc9 | 3118 | return i ? i : -EFAULT; |
63551ae0 | 3119 | } |
8f860591 | 3120 | |
7da4d641 | 3121 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
3122 | unsigned long address, unsigned long end, pgprot_t newprot) |
3123 | { | |
3124 | struct mm_struct *mm = vma->vm_mm; | |
3125 | unsigned long start = address; | |
3126 | pte_t *ptep; | |
3127 | pte_t pte; | |
a5516438 | 3128 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 3129 | unsigned long pages = 0; |
8f860591 ZY |
3130 | |
3131 | BUG_ON(address >= end); | |
3132 | flush_cache_range(vma, address, end); | |
3133 | ||
3d48ae45 | 3134 | mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex); |
a5516438 | 3135 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 3136 | spinlock_t *ptl; |
8f860591 ZY |
3137 | ptep = huge_pte_offset(mm, address); |
3138 | if (!ptep) | |
3139 | continue; | |
cb900f41 | 3140 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
3141 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3142 | pages++; | |
cb900f41 | 3143 | spin_unlock(ptl); |
39dde65c | 3144 | continue; |
7da4d641 | 3145 | } |
7f2e9525 | 3146 | if (!huge_pte_none(huge_ptep_get(ptep))) { |
8f860591 | 3147 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
106c992a | 3148 | pte = pte_mkhuge(huge_pte_modify(pte, newprot)); |
be7517d6 | 3149 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
8f860591 | 3150 | set_huge_pte_at(mm, address, ptep, pte); |
7da4d641 | 3151 | pages++; |
8f860591 | 3152 | } |
cb900f41 | 3153 | spin_unlock(ptl); |
8f860591 | 3154 | } |
d833352a MG |
3155 | /* |
3156 | * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare | |
3157 | * may have cleared our pud entry and done put_page on the page table: | |
3158 | * once we release i_mmap_mutex, another task can do the final put_page | |
3159 | * and that page table be reused and filled with junk. | |
3160 | */ | |
8f860591 | 3161 | flush_tlb_range(vma, start, end); |
d833352a | 3162 | mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex); |
7da4d641 PZ |
3163 | |
3164 | return pages << h->order; | |
8f860591 ZY |
3165 | } |
3166 | ||
a1e78772 MG |
3167 | int hugetlb_reserve_pages(struct inode *inode, |
3168 | long from, long to, | |
5a6fe125 | 3169 | struct vm_area_struct *vma, |
ca16d140 | 3170 | vm_flags_t vm_flags) |
e4e574b7 | 3171 | { |
17c9d12e | 3172 | long ret, chg; |
a5516438 | 3173 | struct hstate *h = hstate_inode(inode); |
90481622 | 3174 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 3175 | struct resv_map *resv_map; |
e4e574b7 | 3176 | |
17c9d12e MG |
3177 | /* |
3178 | * Only apply hugepage reservation if asked. At fault time, an | |
3179 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 3180 | * without using reserves |
17c9d12e | 3181 | */ |
ca16d140 | 3182 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
3183 | return 0; |
3184 | ||
a1e78772 MG |
3185 | /* |
3186 | * Shared mappings base their reservation on the number of pages that | |
3187 | * are already allocated on behalf of the file. Private mappings need | |
3188 | * to reserve the full area even if read-only as mprotect() may be | |
3189 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
3190 | */ | |
9119a41e JK |
3191 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
3192 | resv_map = inode->i_mapping->private_data; | |
3193 | ||
1406ec9b | 3194 | chg = region_chg(resv_map, from, to); |
9119a41e JK |
3195 | |
3196 | } else { | |
3197 | resv_map = resv_map_alloc(); | |
17c9d12e MG |
3198 | if (!resv_map) |
3199 | return -ENOMEM; | |
3200 | ||
a1e78772 | 3201 | chg = to - from; |
84afd99b | 3202 | |
17c9d12e MG |
3203 | set_vma_resv_map(vma, resv_map); |
3204 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
3205 | } | |
3206 | ||
c50ac050 DH |
3207 | if (chg < 0) { |
3208 | ret = chg; | |
3209 | goto out_err; | |
3210 | } | |
8a630112 | 3211 | |
90481622 | 3212 | /* There must be enough pages in the subpool for the mapping */ |
c50ac050 DH |
3213 | if (hugepage_subpool_get_pages(spool, chg)) { |
3214 | ret = -ENOSPC; | |
3215 | goto out_err; | |
3216 | } | |
5a6fe125 MG |
3217 | |
3218 | /* | |
17c9d12e | 3219 | * Check enough hugepages are available for the reservation. |
90481622 | 3220 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 3221 | */ |
a5516438 | 3222 | ret = hugetlb_acct_memory(h, chg); |
68842c9b | 3223 | if (ret < 0) { |
90481622 | 3224 | hugepage_subpool_put_pages(spool, chg); |
c50ac050 | 3225 | goto out_err; |
68842c9b | 3226 | } |
17c9d12e MG |
3227 | |
3228 | /* | |
3229 | * Account for the reservations made. Shared mappings record regions | |
3230 | * that have reservations as they are shared by multiple VMAs. | |
3231 | * When the last VMA disappears, the region map says how much | |
3232 | * the reservation was and the page cache tells how much of | |
3233 | * the reservation was consumed. Private mappings are per-VMA and | |
3234 | * only the consumed reservations are tracked. When the VMA | |
3235 | * disappears, the original reservation is the VMA size and the | |
3236 | * consumed reservations are stored in the map. Hence, nothing | |
3237 | * else has to be done for private mappings here | |
3238 | */ | |
f83a275d | 3239 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
1406ec9b | 3240 | region_add(resv_map, from, to); |
a43a8c39 | 3241 | return 0; |
c50ac050 | 3242 | out_err: |
f031dd27 JK |
3243 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3244 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 3245 | return ret; |
a43a8c39 KC |
3246 | } |
3247 | ||
3248 | void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) | |
3249 | { | |
a5516438 | 3250 | struct hstate *h = hstate_inode(inode); |
9119a41e JK |
3251 | struct resv_map *resv_map = inode->i_mapping->private_data; |
3252 | long chg = 0; | |
90481622 | 3253 | struct hugepage_subpool *spool = subpool_inode(inode); |
45c682a6 | 3254 | |
9119a41e | 3255 | if (resv_map) |
1406ec9b | 3256 | chg = region_truncate(resv_map, offset); |
45c682a6 | 3257 | spin_lock(&inode->i_lock); |
e4c6f8be | 3258 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
3259 | spin_unlock(&inode->i_lock); |
3260 | ||
90481622 | 3261 | hugepage_subpool_put_pages(spool, (chg - freed)); |
a5516438 | 3262 | hugetlb_acct_memory(h, -(chg - freed)); |
a43a8c39 | 3263 | } |
93f70f90 | 3264 | |
3212b535 SC |
3265 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
3266 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
3267 | struct vm_area_struct *vma, | |
3268 | unsigned long addr, pgoff_t idx) | |
3269 | { | |
3270 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
3271 | svma->vm_start; | |
3272 | unsigned long sbase = saddr & PUD_MASK; | |
3273 | unsigned long s_end = sbase + PUD_SIZE; | |
3274 | ||
3275 | /* Allow segments to share if only one is marked locked */ | |
3276 | unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED; | |
3277 | unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED; | |
3278 | ||
3279 | /* | |
3280 | * match the virtual addresses, permission and the alignment of the | |
3281 | * page table page. | |
3282 | */ | |
3283 | if (pmd_index(addr) != pmd_index(saddr) || | |
3284 | vm_flags != svm_flags || | |
3285 | sbase < svma->vm_start || svma->vm_end < s_end) | |
3286 | return 0; | |
3287 | ||
3288 | return saddr; | |
3289 | } | |
3290 | ||
3291 | static int vma_shareable(struct vm_area_struct *vma, unsigned long addr) | |
3292 | { | |
3293 | unsigned long base = addr & PUD_MASK; | |
3294 | unsigned long end = base + PUD_SIZE; | |
3295 | ||
3296 | /* | |
3297 | * check on proper vm_flags and page table alignment | |
3298 | */ | |
3299 | if (vma->vm_flags & VM_MAYSHARE && | |
3300 | vma->vm_start <= base && end <= vma->vm_end) | |
3301 | return 1; | |
3302 | return 0; | |
3303 | } | |
3304 | ||
3305 | /* | |
3306 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
3307 | * and returns the corresponding pte. While this is not necessary for the | |
3308 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
3309 | * code much cleaner. pmd allocation is essential for the shared case because | |
3310 | * pud has to be populated inside the same i_mmap_mutex section - otherwise | |
3311 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a | |
3312 | * bad pmd for sharing. | |
3313 | */ | |
3314 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
3315 | { | |
3316 | struct vm_area_struct *vma = find_vma(mm, addr); | |
3317 | struct address_space *mapping = vma->vm_file->f_mapping; | |
3318 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
3319 | vma->vm_pgoff; | |
3320 | struct vm_area_struct *svma; | |
3321 | unsigned long saddr; | |
3322 | pte_t *spte = NULL; | |
3323 | pte_t *pte; | |
cb900f41 | 3324 | spinlock_t *ptl; |
3212b535 SC |
3325 | |
3326 | if (!vma_shareable(vma, addr)) | |
3327 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
3328 | ||
3329 | mutex_lock(&mapping->i_mmap_mutex); | |
3330 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { | |
3331 | if (svma == vma) | |
3332 | continue; | |
3333 | ||
3334 | saddr = page_table_shareable(svma, vma, addr, idx); | |
3335 | if (saddr) { | |
3336 | spte = huge_pte_offset(svma->vm_mm, saddr); | |
3337 | if (spte) { | |
3338 | get_page(virt_to_page(spte)); | |
3339 | break; | |
3340 | } | |
3341 | } | |
3342 | } | |
3343 | ||
3344 | if (!spte) | |
3345 | goto out; | |
3346 | ||
cb900f41 KS |
3347 | ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); |
3348 | spin_lock(ptl); | |
3212b535 SC |
3349 | if (pud_none(*pud)) |
3350 | pud_populate(mm, pud, | |
3351 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
3352 | else | |
3353 | put_page(virt_to_page(spte)); | |
cb900f41 | 3354 | spin_unlock(ptl); |
3212b535 SC |
3355 | out: |
3356 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3357 | mutex_unlock(&mapping->i_mmap_mutex); | |
3358 | return pte; | |
3359 | } | |
3360 | ||
3361 | /* | |
3362 | * unmap huge page backed by shared pte. | |
3363 | * | |
3364 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
3365 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
3366 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
3367 | * | |
cb900f41 | 3368 | * called with page table lock held. |
3212b535 SC |
3369 | * |
3370 | * returns: 1 successfully unmapped a shared pte page | |
3371 | * 0 the underlying pte page is not shared, or it is the last user | |
3372 | */ | |
3373 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
3374 | { | |
3375 | pgd_t *pgd = pgd_offset(mm, *addr); | |
3376 | pud_t *pud = pud_offset(pgd, *addr); | |
3377 | ||
3378 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
3379 | if (page_count(virt_to_page(ptep)) == 1) | |
3380 | return 0; | |
3381 | ||
3382 | pud_clear(pud); | |
3383 | put_page(virt_to_page(ptep)); | |
3384 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; | |
3385 | return 1; | |
3386 | } | |
9e5fc74c SC |
3387 | #define want_pmd_share() (1) |
3388 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
3389 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
3390 | { | |
3391 | return NULL; | |
3392 | } | |
3393 | #define want_pmd_share() (0) | |
3212b535 SC |
3394 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
3395 | ||
9e5fc74c SC |
3396 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
3397 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
3398 | unsigned long addr, unsigned long sz) | |
3399 | { | |
3400 | pgd_t *pgd; | |
3401 | pud_t *pud; | |
3402 | pte_t *pte = NULL; | |
3403 | ||
3404 | pgd = pgd_offset(mm, addr); | |
3405 | pud = pud_alloc(mm, pgd, addr); | |
3406 | if (pud) { | |
3407 | if (sz == PUD_SIZE) { | |
3408 | pte = (pte_t *)pud; | |
3409 | } else { | |
3410 | BUG_ON(sz != PMD_SIZE); | |
3411 | if (want_pmd_share() && pud_none(*pud)) | |
3412 | pte = huge_pmd_share(mm, addr, pud); | |
3413 | else | |
3414 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3415 | } | |
3416 | } | |
3417 | BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); | |
3418 | ||
3419 | return pte; | |
3420 | } | |
3421 | ||
3422 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) | |
3423 | { | |
3424 | pgd_t *pgd; | |
3425 | pud_t *pud; | |
3426 | pmd_t *pmd = NULL; | |
3427 | ||
3428 | pgd = pgd_offset(mm, addr); | |
3429 | if (pgd_present(*pgd)) { | |
3430 | pud = pud_offset(pgd, addr); | |
3431 | if (pud_present(*pud)) { | |
3432 | if (pud_huge(*pud)) | |
3433 | return (pte_t *)pud; | |
3434 | pmd = pmd_offset(pud, addr); | |
3435 | } | |
3436 | } | |
3437 | return (pte_t *) pmd; | |
3438 | } | |
3439 | ||
3440 | struct page * | |
3441 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, | |
3442 | pmd_t *pmd, int write) | |
3443 | { | |
3444 | struct page *page; | |
3445 | ||
3446 | page = pte_page(*(pte_t *)pmd); | |
3447 | if (page) | |
3448 | page += ((address & ~PMD_MASK) >> PAGE_SHIFT); | |
3449 | return page; | |
3450 | } | |
3451 | ||
3452 | struct page * | |
3453 | follow_huge_pud(struct mm_struct *mm, unsigned long address, | |
3454 | pud_t *pud, int write) | |
3455 | { | |
3456 | struct page *page; | |
3457 | ||
3458 | page = pte_page(*(pte_t *)pud); | |
3459 | if (page) | |
3460 | page += ((address & ~PUD_MASK) >> PAGE_SHIFT); | |
3461 | return page; | |
3462 | } | |
3463 | ||
3464 | #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */ | |
3465 | ||
3466 | /* Can be overriden by architectures */ | |
3467 | __attribute__((weak)) struct page * | |
3468 | follow_huge_pud(struct mm_struct *mm, unsigned long address, | |
3469 | pud_t *pud, int write) | |
3470 | { | |
3471 | BUG(); | |
3472 | return NULL; | |
3473 | } | |
3474 | ||
3475 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ | |
3476 | ||
d5bd9106 AK |
3477 | #ifdef CONFIG_MEMORY_FAILURE |
3478 | ||
6de2b1aa NH |
3479 | /* Should be called in hugetlb_lock */ |
3480 | static int is_hugepage_on_freelist(struct page *hpage) | |
3481 | { | |
3482 | struct page *page; | |
3483 | struct page *tmp; | |
3484 | struct hstate *h = page_hstate(hpage); | |
3485 | int nid = page_to_nid(hpage); | |
3486 | ||
3487 | list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru) | |
3488 | if (page == hpage) | |
3489 | return 1; | |
3490 | return 0; | |
3491 | } | |
3492 | ||
93f70f90 NH |
3493 | /* |
3494 | * This function is called from memory failure code. | |
3495 | * Assume the caller holds page lock of the head page. | |
3496 | */ | |
6de2b1aa | 3497 | int dequeue_hwpoisoned_huge_page(struct page *hpage) |
93f70f90 NH |
3498 | { |
3499 | struct hstate *h = page_hstate(hpage); | |
3500 | int nid = page_to_nid(hpage); | |
6de2b1aa | 3501 | int ret = -EBUSY; |
93f70f90 NH |
3502 | |
3503 | spin_lock(&hugetlb_lock); | |
6de2b1aa | 3504 | if (is_hugepage_on_freelist(hpage)) { |
56f2fb14 NH |
3505 | /* |
3506 | * Hwpoisoned hugepage isn't linked to activelist or freelist, | |
3507 | * but dangling hpage->lru can trigger list-debug warnings | |
3508 | * (this happens when we call unpoison_memory() on it), | |
3509 | * so let it point to itself with list_del_init(). | |
3510 | */ | |
3511 | list_del_init(&hpage->lru); | |
8c6c2ecb | 3512 | set_page_refcounted(hpage); |
6de2b1aa NH |
3513 | h->free_huge_pages--; |
3514 | h->free_huge_pages_node[nid]--; | |
3515 | ret = 0; | |
3516 | } | |
93f70f90 | 3517 | spin_unlock(&hugetlb_lock); |
6de2b1aa | 3518 | return ret; |
93f70f90 | 3519 | } |
6de2b1aa | 3520 | #endif |
31caf665 NH |
3521 | |
3522 | bool isolate_huge_page(struct page *page, struct list_head *list) | |
3523 | { | |
309381fe | 3524 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 NH |
3525 | if (!get_page_unless_zero(page)) |
3526 | return false; | |
3527 | spin_lock(&hugetlb_lock); | |
3528 | list_move_tail(&page->lru, list); | |
3529 | spin_unlock(&hugetlb_lock); | |
3530 | return true; | |
3531 | } | |
3532 | ||
3533 | void putback_active_hugepage(struct page *page) | |
3534 | { | |
309381fe | 3535 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 NH |
3536 | spin_lock(&hugetlb_lock); |
3537 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); | |
3538 | spin_unlock(&hugetlb_lock); | |
3539 | put_page(page); | |
3540 | } | |
c8721bbb NH |
3541 | |
3542 | bool is_hugepage_active(struct page *page) | |
3543 | { | |
309381fe | 3544 | VM_BUG_ON_PAGE(!PageHuge(page), page); |
c8721bbb NH |
3545 | /* |
3546 | * This function can be called for a tail page because the caller, | |
3547 | * scan_movable_pages, scans through a given pfn-range which typically | |
3548 | * covers one memory block. In systems using gigantic hugepage (1GB | |
3549 | * for x86_64,) a hugepage is larger than a memory block, and we don't | |
3550 | * support migrating such large hugepages for now, so return false | |
3551 | * when called for tail pages. | |
3552 | */ | |
3553 | if (PageTail(page)) | |
3554 | return false; | |
3555 | /* | |
3556 | * Refcount of a hwpoisoned hugepages is 1, but they are not active, | |
3557 | * so we should return false for them. | |
3558 | */ | |
3559 | if (unlikely(PageHWPoison(page))) | |
3560 | return false; | |
3561 | return page_count(page) > 0; | |
3562 | } |