Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * Generic hugetlb support. | |
6d49e352 | 3 | * (C) Nadia Yvette Chambers, April 2004 |
1da177e4 | 4 | */ |
1da177e4 LT |
5 | #include <linux/list.h> |
6 | #include <linux/init.h> | |
7 | #include <linux/module.h> | |
8 | #include <linux/mm.h> | |
e1759c21 | 9 | #include <linux/seq_file.h> |
1da177e4 LT |
10 | #include <linux/sysctl.h> |
11 | #include <linux/highmem.h> | |
cddb8a5c | 12 | #include <linux/mmu_notifier.h> |
1da177e4 | 13 | #include <linux/nodemask.h> |
63551ae0 | 14 | #include <linux/pagemap.h> |
5da7ca86 | 15 | #include <linux/mempolicy.h> |
3b32123d | 16 | #include <linux/compiler.h> |
aea47ff3 | 17 | #include <linux/cpuset.h> |
3935baa9 | 18 | #include <linux/mutex.h> |
aa888a74 | 19 | #include <linux/bootmem.h> |
a3437870 | 20 | #include <linux/sysfs.h> |
5a0e3ad6 | 21 | #include <linux/slab.h> |
0fe6e20b | 22 | #include <linux/rmap.h> |
fd6a03ed NH |
23 | #include <linux/swap.h> |
24 | #include <linux/swapops.h> | |
c8721bbb | 25 | #include <linux/page-isolation.h> |
8382d914 | 26 | #include <linux/jhash.h> |
d6606683 | 27 | |
63551ae0 DG |
28 | #include <asm/page.h> |
29 | #include <asm/pgtable.h> | |
24669e58 | 30 | #include <asm/tlb.h> |
63551ae0 | 31 | |
24669e58 | 32 | #include <linux/io.h> |
63551ae0 | 33 | #include <linux/hugetlb.h> |
9dd540e2 | 34 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 35 | #include <linux/node.h> |
7835e98b | 36 | #include "internal.h" |
1da177e4 | 37 | |
753162cd | 38 | int hugepages_treat_as_movable; |
a5516438 | 39 | |
c3f38a38 | 40 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
41 | unsigned int default_hstate_idx; |
42 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
43 | ||
53ba51d2 JT |
44 | __initdata LIST_HEAD(huge_boot_pages); |
45 | ||
e5ff2159 AK |
46 | /* for command line parsing */ |
47 | static struct hstate * __initdata parsed_hstate; | |
48 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 49 | static unsigned long __initdata default_hstate_size; |
e5ff2159 | 50 | |
3935baa9 | 51 | /* |
31caf665 NH |
52 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
53 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 54 | */ |
c3f38a38 | 55 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 56 | |
8382d914 DB |
57 | /* |
58 | * Serializes faults on the same logical page. This is used to | |
59 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
60 | */ | |
61 | static int num_fault_mutexes; | |
62 | static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp; | |
63 | ||
90481622 DG |
64 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
65 | { | |
66 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
67 | ||
68 | spin_unlock(&spool->lock); | |
69 | ||
70 | /* If no pages are used, and no other handles to the subpool | |
71 | * remain, free the subpool the subpool remain */ | |
72 | if (free) | |
73 | kfree(spool); | |
74 | } | |
75 | ||
76 | struct hugepage_subpool *hugepage_new_subpool(long nr_blocks) | |
77 | { | |
78 | struct hugepage_subpool *spool; | |
79 | ||
80 | spool = kmalloc(sizeof(*spool), GFP_KERNEL); | |
81 | if (!spool) | |
82 | return NULL; | |
83 | ||
84 | spin_lock_init(&spool->lock); | |
85 | spool->count = 1; | |
86 | spool->max_hpages = nr_blocks; | |
87 | spool->used_hpages = 0; | |
88 | ||
89 | return spool; | |
90 | } | |
91 | ||
92 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
93 | { | |
94 | spin_lock(&spool->lock); | |
95 | BUG_ON(!spool->count); | |
96 | spool->count--; | |
97 | unlock_or_release_subpool(spool); | |
98 | } | |
99 | ||
100 | static int hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
101 | long delta) | |
102 | { | |
103 | int ret = 0; | |
104 | ||
105 | if (!spool) | |
106 | return 0; | |
107 | ||
108 | spin_lock(&spool->lock); | |
109 | if ((spool->used_hpages + delta) <= spool->max_hpages) { | |
110 | spool->used_hpages += delta; | |
111 | } else { | |
112 | ret = -ENOMEM; | |
113 | } | |
114 | spin_unlock(&spool->lock); | |
115 | ||
116 | return ret; | |
117 | } | |
118 | ||
119 | static void hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
120 | long delta) | |
121 | { | |
122 | if (!spool) | |
123 | return; | |
124 | ||
125 | spin_lock(&spool->lock); | |
126 | spool->used_hpages -= delta; | |
127 | /* If hugetlbfs_put_super couldn't free spool due to | |
128 | * an outstanding quota reference, free it now. */ | |
129 | unlock_or_release_subpool(spool); | |
130 | } | |
131 | ||
132 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
133 | { | |
134 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
135 | } | |
136 | ||
137 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
138 | { | |
496ad9aa | 139 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
140 | } |
141 | ||
96822904 AW |
142 | /* |
143 | * Region tracking -- allows tracking of reservations and instantiated pages | |
144 | * across the pages in a mapping. | |
84afd99b | 145 | * |
7b24d861 DB |
146 | * The region data structures are embedded into a resv_map and |
147 | * protected by a resv_map's lock | |
96822904 AW |
148 | */ |
149 | struct file_region { | |
150 | struct list_head link; | |
151 | long from; | |
152 | long to; | |
153 | }; | |
154 | ||
1406ec9b | 155 | static long region_add(struct resv_map *resv, long f, long t) |
96822904 | 156 | { |
1406ec9b | 157 | struct list_head *head = &resv->regions; |
96822904 AW |
158 | struct file_region *rg, *nrg, *trg; |
159 | ||
7b24d861 | 160 | spin_lock(&resv->lock); |
96822904 AW |
161 | /* Locate the region we are either in or before. */ |
162 | list_for_each_entry(rg, head, link) | |
163 | if (f <= rg->to) | |
164 | break; | |
165 | ||
166 | /* Round our left edge to the current segment if it encloses us. */ | |
167 | if (f > rg->from) | |
168 | f = rg->from; | |
169 | ||
170 | /* Check for and consume any regions we now overlap with. */ | |
171 | nrg = rg; | |
172 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
173 | if (&rg->link == head) | |
174 | break; | |
175 | if (rg->from > t) | |
176 | break; | |
177 | ||
178 | /* If this area reaches higher then extend our area to | |
179 | * include it completely. If this is not the first area | |
180 | * which we intend to reuse, free it. */ | |
181 | if (rg->to > t) | |
182 | t = rg->to; | |
183 | if (rg != nrg) { | |
184 | list_del(&rg->link); | |
185 | kfree(rg); | |
186 | } | |
187 | } | |
188 | nrg->from = f; | |
189 | nrg->to = t; | |
7b24d861 | 190 | spin_unlock(&resv->lock); |
96822904 AW |
191 | return 0; |
192 | } | |
193 | ||
1406ec9b | 194 | static long region_chg(struct resv_map *resv, long f, long t) |
96822904 | 195 | { |
1406ec9b | 196 | struct list_head *head = &resv->regions; |
7b24d861 | 197 | struct file_region *rg, *nrg = NULL; |
96822904 AW |
198 | long chg = 0; |
199 | ||
7b24d861 DB |
200 | retry: |
201 | spin_lock(&resv->lock); | |
96822904 AW |
202 | /* Locate the region we are before or in. */ |
203 | list_for_each_entry(rg, head, link) | |
204 | if (f <= rg->to) | |
205 | break; | |
206 | ||
207 | /* If we are below the current region then a new region is required. | |
208 | * Subtle, allocate a new region at the position but make it zero | |
209 | * size such that we can guarantee to record the reservation. */ | |
210 | if (&rg->link == head || t < rg->from) { | |
7b24d861 DB |
211 | if (!nrg) { |
212 | spin_unlock(&resv->lock); | |
213 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
214 | if (!nrg) | |
215 | return -ENOMEM; | |
216 | ||
217 | nrg->from = f; | |
218 | nrg->to = f; | |
219 | INIT_LIST_HEAD(&nrg->link); | |
220 | goto retry; | |
221 | } | |
96822904 | 222 | |
7b24d861 DB |
223 | list_add(&nrg->link, rg->link.prev); |
224 | chg = t - f; | |
225 | goto out_nrg; | |
96822904 AW |
226 | } |
227 | ||
228 | /* Round our left edge to the current segment if it encloses us. */ | |
229 | if (f > rg->from) | |
230 | f = rg->from; | |
231 | chg = t - f; | |
232 | ||
233 | /* Check for and consume any regions we now overlap with. */ | |
234 | list_for_each_entry(rg, rg->link.prev, link) { | |
235 | if (&rg->link == head) | |
236 | break; | |
237 | if (rg->from > t) | |
7b24d861 | 238 | goto out; |
96822904 | 239 | |
25985edc | 240 | /* We overlap with this area, if it extends further than |
96822904 AW |
241 | * us then we must extend ourselves. Account for its |
242 | * existing reservation. */ | |
243 | if (rg->to > t) { | |
244 | chg += rg->to - t; | |
245 | t = rg->to; | |
246 | } | |
247 | chg -= rg->to - rg->from; | |
248 | } | |
7b24d861 DB |
249 | |
250 | out: | |
251 | spin_unlock(&resv->lock); | |
252 | /* We already know we raced and no longer need the new region */ | |
253 | kfree(nrg); | |
254 | return chg; | |
255 | out_nrg: | |
256 | spin_unlock(&resv->lock); | |
96822904 AW |
257 | return chg; |
258 | } | |
259 | ||
1406ec9b | 260 | static long region_truncate(struct resv_map *resv, long end) |
96822904 | 261 | { |
1406ec9b | 262 | struct list_head *head = &resv->regions; |
96822904 AW |
263 | struct file_region *rg, *trg; |
264 | long chg = 0; | |
265 | ||
7b24d861 | 266 | spin_lock(&resv->lock); |
96822904 AW |
267 | /* Locate the region we are either in or before. */ |
268 | list_for_each_entry(rg, head, link) | |
269 | if (end <= rg->to) | |
270 | break; | |
271 | if (&rg->link == head) | |
7b24d861 | 272 | goto out; |
96822904 AW |
273 | |
274 | /* If we are in the middle of a region then adjust it. */ | |
275 | if (end > rg->from) { | |
276 | chg = rg->to - end; | |
277 | rg->to = end; | |
278 | rg = list_entry(rg->link.next, typeof(*rg), link); | |
279 | } | |
280 | ||
281 | /* Drop any remaining regions. */ | |
282 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
283 | if (&rg->link == head) | |
284 | break; | |
285 | chg += rg->to - rg->from; | |
286 | list_del(&rg->link); | |
287 | kfree(rg); | |
288 | } | |
7b24d861 DB |
289 | |
290 | out: | |
291 | spin_unlock(&resv->lock); | |
96822904 AW |
292 | return chg; |
293 | } | |
294 | ||
1406ec9b | 295 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 296 | { |
1406ec9b | 297 | struct list_head *head = &resv->regions; |
84afd99b AW |
298 | struct file_region *rg; |
299 | long chg = 0; | |
300 | ||
7b24d861 | 301 | spin_lock(&resv->lock); |
84afd99b AW |
302 | /* Locate each segment we overlap with, and count that overlap. */ |
303 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
304 | long seg_from; |
305 | long seg_to; | |
84afd99b AW |
306 | |
307 | if (rg->to <= f) | |
308 | continue; | |
309 | if (rg->from >= t) | |
310 | break; | |
311 | ||
312 | seg_from = max(rg->from, f); | |
313 | seg_to = min(rg->to, t); | |
314 | ||
315 | chg += seg_to - seg_from; | |
316 | } | |
7b24d861 | 317 | spin_unlock(&resv->lock); |
84afd99b AW |
318 | |
319 | return chg; | |
320 | } | |
321 | ||
e7c4b0bf AW |
322 | /* |
323 | * Convert the address within this vma to the page offset within | |
324 | * the mapping, in pagecache page units; huge pages here. | |
325 | */ | |
a5516438 AK |
326 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
327 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 328 | { |
a5516438 AK |
329 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
330 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
331 | } |
332 | ||
0fe6e20b NH |
333 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
334 | unsigned long address) | |
335 | { | |
336 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
337 | } | |
338 | ||
08fba699 MG |
339 | /* |
340 | * Return the size of the pages allocated when backing a VMA. In the majority | |
341 | * cases this will be same size as used by the page table entries. | |
342 | */ | |
343 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
344 | { | |
345 | struct hstate *hstate; | |
346 | ||
347 | if (!is_vm_hugetlb_page(vma)) | |
348 | return PAGE_SIZE; | |
349 | ||
350 | hstate = hstate_vma(vma); | |
351 | ||
2415cf12 | 352 | return 1UL << huge_page_shift(hstate); |
08fba699 | 353 | } |
f340ca0f | 354 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 355 | |
3340289d MG |
356 | /* |
357 | * Return the page size being used by the MMU to back a VMA. In the majority | |
358 | * of cases, the page size used by the kernel matches the MMU size. On | |
359 | * architectures where it differs, an architecture-specific version of this | |
360 | * function is required. | |
361 | */ | |
362 | #ifndef vma_mmu_pagesize | |
363 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) | |
364 | { | |
365 | return vma_kernel_pagesize(vma); | |
366 | } | |
367 | #endif | |
368 | ||
84afd99b AW |
369 | /* |
370 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
371 | * bits of the reservation map pointer, which are always clear due to | |
372 | * alignment. | |
373 | */ | |
374 | #define HPAGE_RESV_OWNER (1UL << 0) | |
375 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 376 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 377 | |
a1e78772 MG |
378 | /* |
379 | * These helpers are used to track how many pages are reserved for | |
380 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
381 | * is guaranteed to have their future faults succeed. | |
382 | * | |
383 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
384 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
385 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
386 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
387 | * |
388 | * The private mapping reservation is represented in a subtly different | |
389 | * manner to a shared mapping. A shared mapping has a region map associated | |
390 | * with the underlying file, this region map represents the backing file | |
391 | * pages which have ever had a reservation assigned which this persists even | |
392 | * after the page is instantiated. A private mapping has a region map | |
393 | * associated with the original mmap which is attached to all VMAs which | |
394 | * reference it, this region map represents those offsets which have consumed | |
395 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 396 | */ |
e7c4b0bf AW |
397 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
398 | { | |
399 | return (unsigned long)vma->vm_private_data; | |
400 | } | |
401 | ||
402 | static void set_vma_private_data(struct vm_area_struct *vma, | |
403 | unsigned long value) | |
404 | { | |
405 | vma->vm_private_data = (void *)value; | |
406 | } | |
407 | ||
9119a41e | 408 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
409 | { |
410 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
411 | if (!resv_map) | |
412 | return NULL; | |
413 | ||
414 | kref_init(&resv_map->refs); | |
7b24d861 | 415 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
416 | INIT_LIST_HEAD(&resv_map->regions); |
417 | ||
418 | return resv_map; | |
419 | } | |
420 | ||
9119a41e | 421 | void resv_map_release(struct kref *ref) |
84afd99b AW |
422 | { |
423 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
424 | ||
425 | /* Clear out any active regions before we release the map. */ | |
1406ec9b | 426 | region_truncate(resv_map, 0); |
84afd99b AW |
427 | kfree(resv_map); |
428 | } | |
429 | ||
4e35f483 JK |
430 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
431 | { | |
432 | return inode->i_mapping->private_data; | |
433 | } | |
434 | ||
84afd99b | 435 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 436 | { |
81d1b09c | 437 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
438 | if (vma->vm_flags & VM_MAYSHARE) { |
439 | struct address_space *mapping = vma->vm_file->f_mapping; | |
440 | struct inode *inode = mapping->host; | |
441 | ||
442 | return inode_resv_map(inode); | |
443 | ||
444 | } else { | |
84afd99b AW |
445 | return (struct resv_map *)(get_vma_private_data(vma) & |
446 | ~HPAGE_RESV_MASK); | |
4e35f483 | 447 | } |
a1e78772 MG |
448 | } |
449 | ||
84afd99b | 450 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 451 | { |
81d1b09c SL |
452 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
453 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 454 | |
84afd99b AW |
455 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
456 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
457 | } |
458 | ||
459 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
460 | { | |
81d1b09c SL |
461 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
462 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
463 | |
464 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
465 | } |
466 | ||
467 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
468 | { | |
81d1b09c | 469 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
470 | |
471 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
472 | } |
473 | ||
04f2cbe3 | 474 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
475 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
476 | { | |
81d1b09c | 477 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 478 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
479 | vma->vm_private_data = (void *)0; |
480 | } | |
481 | ||
482 | /* Returns true if the VMA has associated reserve pages */ | |
af0ed73e | 483 | static int vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 484 | { |
af0ed73e JK |
485 | if (vma->vm_flags & VM_NORESERVE) { |
486 | /* | |
487 | * This address is already reserved by other process(chg == 0), | |
488 | * so, we should decrement reserved count. Without decrementing, | |
489 | * reserve count remains after releasing inode, because this | |
490 | * allocated page will go into page cache and is regarded as | |
491 | * coming from reserved pool in releasing step. Currently, we | |
492 | * don't have any other solution to deal with this situation | |
493 | * properly, so add work-around here. | |
494 | */ | |
495 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
496 | return 1; | |
497 | else | |
498 | return 0; | |
499 | } | |
a63884e9 JK |
500 | |
501 | /* Shared mappings always use reserves */ | |
f83a275d | 502 | if (vma->vm_flags & VM_MAYSHARE) |
7f09ca51 | 503 | return 1; |
a63884e9 JK |
504 | |
505 | /* | |
506 | * Only the process that called mmap() has reserves for | |
507 | * private mappings. | |
508 | */ | |
7f09ca51 MG |
509 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
510 | return 1; | |
a63884e9 | 511 | |
7f09ca51 | 512 | return 0; |
a1e78772 MG |
513 | } |
514 | ||
a5516438 | 515 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
516 | { |
517 | int nid = page_to_nid(page); | |
0edaecfa | 518 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
519 | h->free_huge_pages++; |
520 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
521 | } |
522 | ||
bf50bab2 NH |
523 | static struct page *dequeue_huge_page_node(struct hstate *h, int nid) |
524 | { | |
525 | struct page *page; | |
526 | ||
c8721bbb NH |
527 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
528 | if (!is_migrate_isolate_page(page)) | |
529 | break; | |
530 | /* | |
531 | * if 'non-isolated free hugepage' not found on the list, | |
532 | * the allocation fails. | |
533 | */ | |
534 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 535 | return NULL; |
0edaecfa | 536 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 537 | set_page_refcounted(page); |
bf50bab2 NH |
538 | h->free_huge_pages--; |
539 | h->free_huge_pages_node[nid]--; | |
540 | return page; | |
541 | } | |
542 | ||
86cdb465 NH |
543 | /* Movability of hugepages depends on migration support. */ |
544 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
545 | { | |
100873d7 | 546 | if (hugepages_treat_as_movable || hugepage_migration_supported(h)) |
86cdb465 NH |
547 | return GFP_HIGHUSER_MOVABLE; |
548 | else | |
549 | return GFP_HIGHUSER; | |
550 | } | |
551 | ||
a5516438 AK |
552 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
553 | struct vm_area_struct *vma, | |
af0ed73e JK |
554 | unsigned long address, int avoid_reserve, |
555 | long chg) | |
1da177e4 | 556 | { |
b1c12cbc | 557 | struct page *page = NULL; |
480eccf9 | 558 | struct mempolicy *mpol; |
19770b32 | 559 | nodemask_t *nodemask; |
c0ff7453 | 560 | struct zonelist *zonelist; |
dd1a239f MG |
561 | struct zone *zone; |
562 | struct zoneref *z; | |
cc9a6c87 | 563 | unsigned int cpuset_mems_cookie; |
1da177e4 | 564 | |
a1e78772 MG |
565 | /* |
566 | * A child process with MAP_PRIVATE mappings created by their parent | |
567 | * have no page reserves. This check ensures that reservations are | |
568 | * not "stolen". The child may still get SIGKILLed | |
569 | */ | |
af0ed73e | 570 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 571 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 572 | goto err; |
a1e78772 | 573 | |
04f2cbe3 | 574 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 575 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 576 | goto err; |
04f2cbe3 | 577 | |
9966c4bb | 578 | retry_cpuset: |
d26914d1 | 579 | cpuset_mems_cookie = read_mems_allowed_begin(); |
9966c4bb | 580 | zonelist = huge_zonelist(vma, address, |
86cdb465 | 581 | htlb_alloc_mask(h), &mpol, &nodemask); |
9966c4bb | 582 | |
19770b32 MG |
583 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
584 | MAX_NR_ZONES - 1, nodemask) { | |
344736f2 | 585 | if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) { |
bf50bab2 NH |
586 | page = dequeue_huge_page_node(h, zone_to_nid(zone)); |
587 | if (page) { | |
af0ed73e JK |
588 | if (avoid_reserve) |
589 | break; | |
590 | if (!vma_has_reserves(vma, chg)) | |
591 | break; | |
592 | ||
07443a85 | 593 | SetPagePrivate(page); |
af0ed73e | 594 | h->resv_huge_pages--; |
bf50bab2 NH |
595 | break; |
596 | } | |
3abf7afd | 597 | } |
1da177e4 | 598 | } |
cc9a6c87 | 599 | |
52cd3b07 | 600 | mpol_cond_put(mpol); |
d26914d1 | 601 | if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) |
cc9a6c87 | 602 | goto retry_cpuset; |
1da177e4 | 603 | return page; |
cc9a6c87 MG |
604 | |
605 | err: | |
cc9a6c87 | 606 | return NULL; |
1da177e4 LT |
607 | } |
608 | ||
1cac6f2c LC |
609 | /* |
610 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
611 | * We may have allocated or freed a huge page based on a different | |
612 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
613 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
614 | * node for alloc or free. | |
615 | */ | |
616 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
617 | { | |
618 | nid = next_node(nid, *nodes_allowed); | |
619 | if (nid == MAX_NUMNODES) | |
620 | nid = first_node(*nodes_allowed); | |
621 | VM_BUG_ON(nid >= MAX_NUMNODES); | |
622 | ||
623 | return nid; | |
624 | } | |
625 | ||
626 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
627 | { | |
628 | if (!node_isset(nid, *nodes_allowed)) | |
629 | nid = next_node_allowed(nid, nodes_allowed); | |
630 | return nid; | |
631 | } | |
632 | ||
633 | /* | |
634 | * returns the previously saved node ["this node"] from which to | |
635 | * allocate a persistent huge page for the pool and advance the | |
636 | * next node from which to allocate, handling wrap at end of node | |
637 | * mask. | |
638 | */ | |
639 | static int hstate_next_node_to_alloc(struct hstate *h, | |
640 | nodemask_t *nodes_allowed) | |
641 | { | |
642 | int nid; | |
643 | ||
644 | VM_BUG_ON(!nodes_allowed); | |
645 | ||
646 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
647 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
648 | ||
649 | return nid; | |
650 | } | |
651 | ||
652 | /* | |
653 | * helper for free_pool_huge_page() - return the previously saved | |
654 | * node ["this node"] from which to free a huge page. Advance the | |
655 | * next node id whether or not we find a free huge page to free so | |
656 | * that the next attempt to free addresses the next node. | |
657 | */ | |
658 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
659 | { | |
660 | int nid; | |
661 | ||
662 | VM_BUG_ON(!nodes_allowed); | |
663 | ||
664 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
665 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
666 | ||
667 | return nid; | |
668 | } | |
669 | ||
670 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
671 | for (nr_nodes = nodes_weight(*mask); \ | |
672 | nr_nodes > 0 && \ | |
673 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
674 | nr_nodes--) | |
675 | ||
676 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
677 | for (nr_nodes = nodes_weight(*mask); \ | |
678 | nr_nodes > 0 && \ | |
679 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
680 | nr_nodes--) | |
681 | ||
944d9fec LC |
682 | #if defined(CONFIG_CMA) && defined(CONFIG_X86_64) |
683 | static void destroy_compound_gigantic_page(struct page *page, | |
684 | unsigned long order) | |
685 | { | |
686 | int i; | |
687 | int nr_pages = 1 << order; | |
688 | struct page *p = page + 1; | |
689 | ||
690 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { | |
691 | __ClearPageTail(p); | |
692 | set_page_refcounted(p); | |
693 | p->first_page = NULL; | |
694 | } | |
695 | ||
696 | set_compound_order(page, 0); | |
697 | __ClearPageHead(page); | |
698 | } | |
699 | ||
700 | static void free_gigantic_page(struct page *page, unsigned order) | |
701 | { | |
702 | free_contig_range(page_to_pfn(page), 1 << order); | |
703 | } | |
704 | ||
705 | static int __alloc_gigantic_page(unsigned long start_pfn, | |
706 | unsigned long nr_pages) | |
707 | { | |
708 | unsigned long end_pfn = start_pfn + nr_pages; | |
709 | return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE); | |
710 | } | |
711 | ||
712 | static bool pfn_range_valid_gigantic(unsigned long start_pfn, | |
713 | unsigned long nr_pages) | |
714 | { | |
715 | unsigned long i, end_pfn = start_pfn + nr_pages; | |
716 | struct page *page; | |
717 | ||
718 | for (i = start_pfn; i < end_pfn; i++) { | |
719 | if (!pfn_valid(i)) | |
720 | return false; | |
721 | ||
722 | page = pfn_to_page(i); | |
723 | ||
724 | if (PageReserved(page)) | |
725 | return false; | |
726 | ||
727 | if (page_count(page) > 0) | |
728 | return false; | |
729 | ||
730 | if (PageHuge(page)) | |
731 | return false; | |
732 | } | |
733 | ||
734 | return true; | |
735 | } | |
736 | ||
737 | static bool zone_spans_last_pfn(const struct zone *zone, | |
738 | unsigned long start_pfn, unsigned long nr_pages) | |
739 | { | |
740 | unsigned long last_pfn = start_pfn + nr_pages - 1; | |
741 | return zone_spans_pfn(zone, last_pfn); | |
742 | } | |
743 | ||
744 | static struct page *alloc_gigantic_page(int nid, unsigned order) | |
745 | { | |
746 | unsigned long nr_pages = 1 << order; | |
747 | unsigned long ret, pfn, flags; | |
748 | struct zone *z; | |
749 | ||
750 | z = NODE_DATA(nid)->node_zones; | |
751 | for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) { | |
752 | spin_lock_irqsave(&z->lock, flags); | |
753 | ||
754 | pfn = ALIGN(z->zone_start_pfn, nr_pages); | |
755 | while (zone_spans_last_pfn(z, pfn, nr_pages)) { | |
756 | if (pfn_range_valid_gigantic(pfn, nr_pages)) { | |
757 | /* | |
758 | * We release the zone lock here because | |
759 | * alloc_contig_range() will also lock the zone | |
760 | * at some point. If there's an allocation | |
761 | * spinning on this lock, it may win the race | |
762 | * and cause alloc_contig_range() to fail... | |
763 | */ | |
764 | spin_unlock_irqrestore(&z->lock, flags); | |
765 | ret = __alloc_gigantic_page(pfn, nr_pages); | |
766 | if (!ret) | |
767 | return pfn_to_page(pfn); | |
768 | spin_lock_irqsave(&z->lock, flags); | |
769 | } | |
770 | pfn += nr_pages; | |
771 | } | |
772 | ||
773 | spin_unlock_irqrestore(&z->lock, flags); | |
774 | } | |
775 | ||
776 | return NULL; | |
777 | } | |
778 | ||
779 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
780 | static void prep_compound_gigantic_page(struct page *page, unsigned long order); | |
781 | ||
782 | static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) | |
783 | { | |
784 | struct page *page; | |
785 | ||
786 | page = alloc_gigantic_page(nid, huge_page_order(h)); | |
787 | if (page) { | |
788 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
789 | prep_new_huge_page(h, page, nid); | |
790 | } | |
791 | ||
792 | return page; | |
793 | } | |
794 | ||
795 | static int alloc_fresh_gigantic_page(struct hstate *h, | |
796 | nodemask_t *nodes_allowed) | |
797 | { | |
798 | struct page *page = NULL; | |
799 | int nr_nodes, node; | |
800 | ||
801 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
802 | page = alloc_fresh_gigantic_page_node(h, node); | |
803 | if (page) | |
804 | return 1; | |
805 | } | |
806 | ||
807 | return 0; | |
808 | } | |
809 | ||
810 | static inline bool gigantic_page_supported(void) { return true; } | |
811 | #else | |
812 | static inline bool gigantic_page_supported(void) { return false; } | |
813 | static inline void free_gigantic_page(struct page *page, unsigned order) { } | |
814 | static inline void destroy_compound_gigantic_page(struct page *page, | |
815 | unsigned long order) { } | |
816 | static inline int alloc_fresh_gigantic_page(struct hstate *h, | |
817 | nodemask_t *nodes_allowed) { return 0; } | |
818 | #endif | |
819 | ||
a5516438 | 820 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
821 | { |
822 | int i; | |
a5516438 | 823 | |
944d9fec LC |
824 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
825 | return; | |
18229df5 | 826 | |
a5516438 AK |
827 | h->nr_huge_pages--; |
828 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
829 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
830 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
831 | 1 << PG_referenced | 1 << PG_dirty | | |
a7407a27 LC |
832 | 1 << PG_active | 1 << PG_private | |
833 | 1 << PG_writeback); | |
6af2acb6 | 834 | } |
309381fe | 835 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
6af2acb6 AL |
836 | set_compound_page_dtor(page, NULL); |
837 | set_page_refcounted(page); | |
944d9fec LC |
838 | if (hstate_is_gigantic(h)) { |
839 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
840 | free_gigantic_page(page, huge_page_order(h)); | |
841 | } else { | |
842 | arch_release_hugepage(page); | |
843 | __free_pages(page, huge_page_order(h)); | |
844 | } | |
6af2acb6 AL |
845 | } |
846 | ||
e5ff2159 AK |
847 | struct hstate *size_to_hstate(unsigned long size) |
848 | { | |
849 | struct hstate *h; | |
850 | ||
851 | for_each_hstate(h) { | |
852 | if (huge_page_size(h) == size) | |
853 | return h; | |
854 | } | |
855 | return NULL; | |
856 | } | |
857 | ||
8f1d26d0 | 858 | void free_huge_page(struct page *page) |
27a85ef1 | 859 | { |
a5516438 AK |
860 | /* |
861 | * Can't pass hstate in here because it is called from the | |
862 | * compound page destructor. | |
863 | */ | |
e5ff2159 | 864 | struct hstate *h = page_hstate(page); |
7893d1d5 | 865 | int nid = page_to_nid(page); |
90481622 DG |
866 | struct hugepage_subpool *spool = |
867 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 868 | bool restore_reserve; |
27a85ef1 | 869 | |
e5df70ab | 870 | set_page_private(page, 0); |
23be7468 | 871 | page->mapping = NULL; |
7893d1d5 | 872 | BUG_ON(page_count(page)); |
0fe6e20b | 873 | BUG_ON(page_mapcount(page)); |
07443a85 | 874 | restore_reserve = PagePrivate(page); |
16c794b4 | 875 | ClearPagePrivate(page); |
27a85ef1 DG |
876 | |
877 | spin_lock(&hugetlb_lock); | |
6d76dcf4 AK |
878 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
879 | pages_per_huge_page(h), page); | |
07443a85 JK |
880 | if (restore_reserve) |
881 | h->resv_huge_pages++; | |
882 | ||
944d9fec | 883 | if (h->surplus_huge_pages_node[nid]) { |
0edaecfa AK |
884 | /* remove the page from active list */ |
885 | list_del(&page->lru); | |
a5516438 AK |
886 | update_and_free_page(h, page); |
887 | h->surplus_huge_pages--; | |
888 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 889 | } else { |
5d3a551c | 890 | arch_clear_hugepage_flags(page); |
a5516438 | 891 | enqueue_huge_page(h, page); |
7893d1d5 | 892 | } |
27a85ef1 | 893 | spin_unlock(&hugetlb_lock); |
90481622 | 894 | hugepage_subpool_put_pages(spool, 1); |
27a85ef1 DG |
895 | } |
896 | ||
a5516438 | 897 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 898 | { |
0edaecfa | 899 | INIT_LIST_HEAD(&page->lru); |
b7ba30c6 AK |
900 | set_compound_page_dtor(page, free_huge_page); |
901 | spin_lock(&hugetlb_lock); | |
9dd540e2 | 902 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
903 | h->nr_huge_pages++; |
904 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 AK |
905 | spin_unlock(&hugetlb_lock); |
906 | put_page(page); /* free it into the hugepage allocator */ | |
907 | } | |
908 | ||
2906dd52 | 909 | static void prep_compound_gigantic_page(struct page *page, unsigned long order) |
20a0307c WF |
910 | { |
911 | int i; | |
912 | int nr_pages = 1 << order; | |
913 | struct page *p = page + 1; | |
914 | ||
915 | /* we rely on prep_new_huge_page to set the destructor */ | |
916 | set_compound_order(page, order); | |
917 | __SetPageHead(page); | |
ef5a22be | 918 | __ClearPageReserved(page); |
20a0307c | 919 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
920 | /* |
921 | * For gigantic hugepages allocated through bootmem at | |
922 | * boot, it's safer to be consistent with the not-gigantic | |
923 | * hugepages and clear the PG_reserved bit from all tail pages | |
924 | * too. Otherwse drivers using get_user_pages() to access tail | |
925 | * pages may get the reference counting wrong if they see | |
926 | * PG_reserved set on a tail page (despite the head page not | |
927 | * having PG_reserved set). Enforcing this consistency between | |
928 | * head and tail pages allows drivers to optimize away a check | |
929 | * on the head page when they need know if put_page() is needed | |
930 | * after get_user_pages(). | |
931 | */ | |
932 | __ClearPageReserved(p); | |
58a84aa9 | 933 | set_page_count(p, 0); |
20a0307c | 934 | p->first_page = page; |
44fc8057 DR |
935 | /* Make sure p->first_page is always valid for PageTail() */ |
936 | smp_wmb(); | |
937 | __SetPageTail(p); | |
20a0307c WF |
938 | } |
939 | } | |
940 | ||
7795912c AM |
941 | /* |
942 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
943 | * transparent huge pages. See the PageTransHuge() documentation for more | |
944 | * details. | |
945 | */ | |
20a0307c WF |
946 | int PageHuge(struct page *page) |
947 | { | |
20a0307c WF |
948 | if (!PageCompound(page)) |
949 | return 0; | |
950 | ||
951 | page = compound_head(page); | |
758f66a2 | 952 | return get_compound_page_dtor(page) == free_huge_page; |
20a0307c | 953 | } |
43131e14 NH |
954 | EXPORT_SYMBOL_GPL(PageHuge); |
955 | ||
27c73ae7 AA |
956 | /* |
957 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
958 | * normal or transparent huge pages. | |
959 | */ | |
960 | int PageHeadHuge(struct page *page_head) | |
961 | { | |
27c73ae7 AA |
962 | if (!PageHead(page_head)) |
963 | return 0; | |
964 | ||
758f66a2 | 965 | return get_compound_page_dtor(page_head) == free_huge_page; |
27c73ae7 | 966 | } |
27c73ae7 | 967 | |
13d60f4b ZY |
968 | pgoff_t __basepage_index(struct page *page) |
969 | { | |
970 | struct page *page_head = compound_head(page); | |
971 | pgoff_t index = page_index(page_head); | |
972 | unsigned long compound_idx; | |
973 | ||
974 | if (!PageHuge(page_head)) | |
975 | return page_index(page); | |
976 | ||
977 | if (compound_order(page_head) >= MAX_ORDER) | |
978 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
979 | else | |
980 | compound_idx = page - page_head; | |
981 | ||
982 | return (index << compound_order(page_head)) + compound_idx; | |
983 | } | |
984 | ||
a5516438 | 985 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
1da177e4 | 986 | { |
1da177e4 | 987 | struct page *page; |
f96efd58 | 988 | |
6484eb3e | 989 | page = alloc_pages_exact_node(nid, |
86cdb465 | 990 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
551883ae | 991 | __GFP_REPEAT|__GFP_NOWARN, |
a5516438 | 992 | huge_page_order(h)); |
1da177e4 | 993 | if (page) { |
7f2e9525 | 994 | if (arch_prepare_hugepage(page)) { |
caff3a2c | 995 | __free_pages(page, huge_page_order(h)); |
7b8ee84d | 996 | return NULL; |
7f2e9525 | 997 | } |
a5516438 | 998 | prep_new_huge_page(h, page, nid); |
1da177e4 | 999 | } |
63b4613c NA |
1000 | |
1001 | return page; | |
1002 | } | |
1003 | ||
b2261026 JK |
1004 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
1005 | { | |
1006 | struct page *page; | |
1007 | int nr_nodes, node; | |
1008 | int ret = 0; | |
1009 | ||
1010 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1011 | page = alloc_fresh_huge_page_node(h, node); | |
1012 | if (page) { | |
1013 | ret = 1; | |
1014 | break; | |
1015 | } | |
1016 | } | |
1017 | ||
1018 | if (ret) | |
1019 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
1020 | else | |
1021 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
1022 | ||
1023 | return ret; | |
1024 | } | |
1025 | ||
e8c5c824 LS |
1026 | /* |
1027 | * Free huge page from pool from next node to free. | |
1028 | * Attempt to keep persistent huge pages more or less | |
1029 | * balanced over allowed nodes. | |
1030 | * Called with hugetlb_lock locked. | |
1031 | */ | |
6ae11b27 LS |
1032 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1033 | bool acct_surplus) | |
e8c5c824 | 1034 | { |
b2261026 | 1035 | int nr_nodes, node; |
e8c5c824 LS |
1036 | int ret = 0; |
1037 | ||
b2261026 | 1038 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1039 | /* |
1040 | * If we're returning unused surplus pages, only examine | |
1041 | * nodes with surplus pages. | |
1042 | */ | |
b2261026 JK |
1043 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1044 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1045 | struct page *page = |
b2261026 | 1046 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1047 | struct page, lru); |
1048 | list_del(&page->lru); | |
1049 | h->free_huge_pages--; | |
b2261026 | 1050 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1051 | if (acct_surplus) { |
1052 | h->surplus_huge_pages--; | |
b2261026 | 1053 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1054 | } |
e8c5c824 LS |
1055 | update_and_free_page(h, page); |
1056 | ret = 1; | |
9a76db09 | 1057 | break; |
e8c5c824 | 1058 | } |
b2261026 | 1059 | } |
e8c5c824 LS |
1060 | |
1061 | return ret; | |
1062 | } | |
1063 | ||
c8721bbb NH |
1064 | /* |
1065 | * Dissolve a given free hugepage into free buddy pages. This function does | |
1066 | * nothing for in-use (including surplus) hugepages. | |
1067 | */ | |
1068 | static void dissolve_free_huge_page(struct page *page) | |
1069 | { | |
1070 | spin_lock(&hugetlb_lock); | |
1071 | if (PageHuge(page) && !page_count(page)) { | |
1072 | struct hstate *h = page_hstate(page); | |
1073 | int nid = page_to_nid(page); | |
1074 | list_del(&page->lru); | |
1075 | h->free_huge_pages--; | |
1076 | h->free_huge_pages_node[nid]--; | |
1077 | update_and_free_page(h, page); | |
1078 | } | |
1079 | spin_unlock(&hugetlb_lock); | |
1080 | } | |
1081 | ||
1082 | /* | |
1083 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1084 | * make specified memory blocks removable from the system. | |
1085 | * Note that start_pfn should aligned with (minimum) hugepage size. | |
1086 | */ | |
1087 | void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) | |
1088 | { | |
1089 | unsigned int order = 8 * sizeof(void *); | |
1090 | unsigned long pfn; | |
1091 | struct hstate *h; | |
1092 | ||
d0177639 LZ |
1093 | if (!hugepages_supported()) |
1094 | return; | |
1095 | ||
c8721bbb NH |
1096 | /* Set scan step to minimum hugepage size */ |
1097 | for_each_hstate(h) | |
1098 | if (order > huge_page_order(h)) | |
1099 | order = huge_page_order(h); | |
1100 | VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order)); | |
1101 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) | |
1102 | dissolve_free_huge_page(pfn_to_page(pfn)); | |
1103 | } | |
1104 | ||
bf50bab2 | 1105 | static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) |
7893d1d5 AL |
1106 | { |
1107 | struct page *page; | |
bf50bab2 | 1108 | unsigned int r_nid; |
7893d1d5 | 1109 | |
bae7f4ae | 1110 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1111 | return NULL; |
1112 | ||
d1c3fb1f NA |
1113 | /* |
1114 | * Assume we will successfully allocate the surplus page to | |
1115 | * prevent racing processes from causing the surplus to exceed | |
1116 | * overcommit | |
1117 | * | |
1118 | * This however introduces a different race, where a process B | |
1119 | * tries to grow the static hugepage pool while alloc_pages() is | |
1120 | * called by process A. B will only examine the per-node | |
1121 | * counters in determining if surplus huge pages can be | |
1122 | * converted to normal huge pages in adjust_pool_surplus(). A | |
1123 | * won't be able to increment the per-node counter, until the | |
1124 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
1125 | * no more huge pages can be converted from surplus to normal | |
1126 | * state (and doesn't try to convert again). Thus, we have a | |
1127 | * case where a surplus huge page exists, the pool is grown, and | |
1128 | * the surplus huge page still exists after, even though it | |
1129 | * should just have been converted to a normal huge page. This | |
1130 | * does not leak memory, though, as the hugepage will be freed | |
1131 | * once it is out of use. It also does not allow the counters to | |
1132 | * go out of whack in adjust_pool_surplus() as we don't modify | |
1133 | * the node values until we've gotten the hugepage and only the | |
1134 | * per-node value is checked there. | |
1135 | */ | |
1136 | spin_lock(&hugetlb_lock); | |
a5516438 | 1137 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
1138 | spin_unlock(&hugetlb_lock); |
1139 | return NULL; | |
1140 | } else { | |
a5516438 AK |
1141 | h->nr_huge_pages++; |
1142 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
1143 | } |
1144 | spin_unlock(&hugetlb_lock); | |
1145 | ||
bf50bab2 | 1146 | if (nid == NUMA_NO_NODE) |
86cdb465 | 1147 | page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP| |
bf50bab2 NH |
1148 | __GFP_REPEAT|__GFP_NOWARN, |
1149 | huge_page_order(h)); | |
1150 | else | |
1151 | page = alloc_pages_exact_node(nid, | |
86cdb465 | 1152 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
bf50bab2 | 1153 | __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); |
d1c3fb1f | 1154 | |
caff3a2c GS |
1155 | if (page && arch_prepare_hugepage(page)) { |
1156 | __free_pages(page, huge_page_order(h)); | |
ea5768c7 | 1157 | page = NULL; |
caff3a2c GS |
1158 | } |
1159 | ||
d1c3fb1f | 1160 | spin_lock(&hugetlb_lock); |
7893d1d5 | 1161 | if (page) { |
0edaecfa | 1162 | INIT_LIST_HEAD(&page->lru); |
bf50bab2 | 1163 | r_nid = page_to_nid(page); |
7893d1d5 | 1164 | set_compound_page_dtor(page, free_huge_page); |
9dd540e2 | 1165 | set_hugetlb_cgroup(page, NULL); |
d1c3fb1f NA |
1166 | /* |
1167 | * We incremented the global counters already | |
1168 | */ | |
bf50bab2 NH |
1169 | h->nr_huge_pages_node[r_nid]++; |
1170 | h->surplus_huge_pages_node[r_nid]++; | |
3b116300 | 1171 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 1172 | } else { |
a5516438 AK |
1173 | h->nr_huge_pages--; |
1174 | h->surplus_huge_pages--; | |
3b116300 | 1175 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 1176 | } |
d1c3fb1f | 1177 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1178 | |
1179 | return page; | |
1180 | } | |
1181 | ||
bf50bab2 NH |
1182 | /* |
1183 | * This allocation function is useful in the context where vma is irrelevant. | |
1184 | * E.g. soft-offlining uses this function because it only cares physical | |
1185 | * address of error page. | |
1186 | */ | |
1187 | struct page *alloc_huge_page_node(struct hstate *h, int nid) | |
1188 | { | |
4ef91848 | 1189 | struct page *page = NULL; |
bf50bab2 NH |
1190 | |
1191 | spin_lock(&hugetlb_lock); | |
4ef91848 JK |
1192 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
1193 | page = dequeue_huge_page_node(h, nid); | |
bf50bab2 NH |
1194 | spin_unlock(&hugetlb_lock); |
1195 | ||
94ae8ba7 | 1196 | if (!page) |
bf50bab2 NH |
1197 | page = alloc_buddy_huge_page(h, nid); |
1198 | ||
1199 | return page; | |
1200 | } | |
1201 | ||
e4e574b7 | 1202 | /* |
25985edc | 1203 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1204 | * of size 'delta'. |
1205 | */ | |
a5516438 | 1206 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1207 | { |
1208 | struct list_head surplus_list; | |
1209 | struct page *page, *tmp; | |
1210 | int ret, i; | |
1211 | int needed, allocated; | |
28073b02 | 1212 | bool alloc_ok = true; |
e4e574b7 | 1213 | |
a5516438 | 1214 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1215 | if (needed <= 0) { |
a5516438 | 1216 | h->resv_huge_pages += delta; |
e4e574b7 | 1217 | return 0; |
ac09b3a1 | 1218 | } |
e4e574b7 AL |
1219 | |
1220 | allocated = 0; | |
1221 | INIT_LIST_HEAD(&surplus_list); | |
1222 | ||
1223 | ret = -ENOMEM; | |
1224 | retry: | |
1225 | spin_unlock(&hugetlb_lock); | |
1226 | for (i = 0; i < needed; i++) { | |
bf50bab2 | 1227 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
28073b02 HD |
1228 | if (!page) { |
1229 | alloc_ok = false; | |
1230 | break; | |
1231 | } | |
e4e574b7 AL |
1232 | list_add(&page->lru, &surplus_list); |
1233 | } | |
28073b02 | 1234 | allocated += i; |
e4e574b7 AL |
1235 | |
1236 | /* | |
1237 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1238 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1239 | */ | |
1240 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1241 | needed = (h->resv_huge_pages + delta) - |
1242 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1243 | if (needed > 0) { |
1244 | if (alloc_ok) | |
1245 | goto retry; | |
1246 | /* | |
1247 | * We were not able to allocate enough pages to | |
1248 | * satisfy the entire reservation so we free what | |
1249 | * we've allocated so far. | |
1250 | */ | |
1251 | goto free; | |
1252 | } | |
e4e574b7 AL |
1253 | /* |
1254 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1255 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1256 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1257 | * allocator. Commit the entire reservation here to prevent another |
1258 | * process from stealing the pages as they are added to the pool but | |
1259 | * before they are reserved. | |
e4e574b7 AL |
1260 | */ |
1261 | needed += allocated; | |
a5516438 | 1262 | h->resv_huge_pages += delta; |
e4e574b7 | 1263 | ret = 0; |
a9869b83 | 1264 | |
19fc3f0a | 1265 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1266 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1267 | if ((--needed) < 0) |
1268 | break; | |
a9869b83 NH |
1269 | /* |
1270 | * This page is now managed by the hugetlb allocator and has | |
1271 | * no users -- drop the buddy allocator's reference. | |
1272 | */ | |
1273 | put_page_testzero(page); | |
309381fe | 1274 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 1275 | enqueue_huge_page(h, page); |
19fc3f0a | 1276 | } |
28073b02 | 1277 | free: |
b0365c8d | 1278 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1279 | |
1280 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1281 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1282 | put_page(page); | |
a9869b83 | 1283 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1284 | |
1285 | return ret; | |
1286 | } | |
1287 | ||
1288 | /* | |
1289 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
1290 | * allocated to satisfy the reservation must be explicitly freed if they were | |
1291 | * never used. | |
685f3457 | 1292 | * Called with hugetlb_lock held. |
e4e574b7 | 1293 | */ |
a5516438 AK |
1294 | static void return_unused_surplus_pages(struct hstate *h, |
1295 | unsigned long unused_resv_pages) | |
e4e574b7 | 1296 | { |
e4e574b7 AL |
1297 | unsigned long nr_pages; |
1298 | ||
ac09b3a1 | 1299 | /* Uncommit the reservation */ |
a5516438 | 1300 | h->resv_huge_pages -= unused_resv_pages; |
ac09b3a1 | 1301 | |
aa888a74 | 1302 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 1303 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1304 | return; |
1305 | ||
a5516438 | 1306 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1307 | |
685f3457 LS |
1308 | /* |
1309 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1310 | * evenly across all nodes with memory. Iterate across these nodes |
1311 | * until we can no longer free unreserved surplus pages. This occurs | |
1312 | * when the nodes with surplus pages have no free pages. | |
1313 | * free_pool_huge_page() will balance the the freed pages across the | |
1314 | * on-line nodes with memory and will handle the hstate accounting. | |
685f3457 LS |
1315 | */ |
1316 | while (nr_pages--) { | |
8cebfcd0 | 1317 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
685f3457 | 1318 | break; |
7848a4bf | 1319 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 AL |
1320 | } |
1321 | } | |
1322 | ||
c37f9fb1 AW |
1323 | /* |
1324 | * Determine if the huge page at addr within the vma has an associated | |
1325 | * reservation. Where it does not we will need to logically increase | |
90481622 DG |
1326 | * reservation and actually increase subpool usage before an allocation |
1327 | * can occur. Where any new reservation would be required the | |
1328 | * reservation change is prepared, but not committed. Once the page | |
1329 | * has been allocated from the subpool and instantiated the change should | |
1330 | * be committed via vma_commit_reservation. No action is required on | |
1331 | * failure. | |
c37f9fb1 | 1332 | */ |
e2f17d94 | 1333 | static long vma_needs_reservation(struct hstate *h, |
a5516438 | 1334 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 1335 | { |
4e35f483 JK |
1336 | struct resv_map *resv; |
1337 | pgoff_t idx; | |
1338 | long chg; | |
c37f9fb1 | 1339 | |
4e35f483 JK |
1340 | resv = vma_resv_map(vma); |
1341 | if (!resv) | |
84afd99b | 1342 | return 1; |
c37f9fb1 | 1343 | |
4e35f483 JK |
1344 | idx = vma_hugecache_offset(h, vma, addr); |
1345 | chg = region_chg(resv, idx, idx + 1); | |
84afd99b | 1346 | |
4e35f483 JK |
1347 | if (vma->vm_flags & VM_MAYSHARE) |
1348 | return chg; | |
1349 | else | |
1350 | return chg < 0 ? chg : 0; | |
c37f9fb1 | 1351 | } |
a5516438 AK |
1352 | static void vma_commit_reservation(struct hstate *h, |
1353 | struct vm_area_struct *vma, unsigned long addr) | |
c37f9fb1 | 1354 | { |
4e35f483 JK |
1355 | struct resv_map *resv; |
1356 | pgoff_t idx; | |
84afd99b | 1357 | |
4e35f483 JK |
1358 | resv = vma_resv_map(vma); |
1359 | if (!resv) | |
1360 | return; | |
84afd99b | 1361 | |
4e35f483 JK |
1362 | idx = vma_hugecache_offset(h, vma, addr); |
1363 | region_add(resv, idx, idx + 1); | |
c37f9fb1 AW |
1364 | } |
1365 | ||
a1e78772 | 1366 | static struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 1367 | unsigned long addr, int avoid_reserve) |
1da177e4 | 1368 | { |
90481622 | 1369 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 1370 | struct hstate *h = hstate_vma(vma); |
348ea204 | 1371 | struct page *page; |
e2f17d94 | 1372 | long chg; |
6d76dcf4 AK |
1373 | int ret, idx; |
1374 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 1375 | |
6d76dcf4 | 1376 | idx = hstate_index(h); |
a1e78772 | 1377 | /* |
90481622 DG |
1378 | * Processes that did not create the mapping will have no |
1379 | * reserves and will not have accounted against subpool | |
1380 | * limit. Check that the subpool limit can be made before | |
1381 | * satisfying the allocation MAP_NORESERVE mappings may also | |
1382 | * need pages and subpool limit allocated allocated if no reserve | |
1383 | * mapping overlaps. | |
a1e78772 | 1384 | */ |
a5516438 | 1385 | chg = vma_needs_reservation(h, vma, addr); |
c37f9fb1 | 1386 | if (chg < 0) |
76dcee75 | 1387 | return ERR_PTR(-ENOMEM); |
8bb3f12e JK |
1388 | if (chg || avoid_reserve) |
1389 | if (hugepage_subpool_get_pages(spool, 1)) | |
76dcee75 | 1390 | return ERR_PTR(-ENOSPC); |
1da177e4 | 1391 | |
6d76dcf4 | 1392 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f JZ |
1393 | if (ret) |
1394 | goto out_subpool_put; | |
1395 | ||
1da177e4 | 1396 | spin_lock(&hugetlb_lock); |
af0ed73e | 1397 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg); |
81a6fcae | 1398 | if (!page) { |
94ae8ba7 | 1399 | spin_unlock(&hugetlb_lock); |
bf50bab2 | 1400 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
8f34af6f JZ |
1401 | if (!page) |
1402 | goto out_uncharge_cgroup; | |
1403 | ||
79dbb236 AK |
1404 | spin_lock(&hugetlb_lock); |
1405 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 1406 | /* Fall through */ |
68842c9b | 1407 | } |
81a6fcae JK |
1408 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
1409 | spin_unlock(&hugetlb_lock); | |
348ea204 | 1410 | |
90481622 | 1411 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 1412 | |
a5516438 | 1413 | vma_commit_reservation(h, vma, addr); |
90d8b7e6 | 1414 | return page; |
8f34af6f JZ |
1415 | |
1416 | out_uncharge_cgroup: | |
1417 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
1418 | out_subpool_put: | |
1419 | if (chg || avoid_reserve) | |
1420 | hugepage_subpool_put_pages(spool, 1); | |
1421 | return ERR_PTR(-ENOSPC); | |
b45b5bd6 DG |
1422 | } |
1423 | ||
74060e4d NH |
1424 | /* |
1425 | * alloc_huge_page()'s wrapper which simply returns the page if allocation | |
1426 | * succeeds, otherwise NULL. This function is called from new_vma_page(), | |
1427 | * where no ERR_VALUE is expected to be returned. | |
1428 | */ | |
1429 | struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, | |
1430 | unsigned long addr, int avoid_reserve) | |
1431 | { | |
1432 | struct page *page = alloc_huge_page(vma, addr, avoid_reserve); | |
1433 | if (IS_ERR(page)) | |
1434 | page = NULL; | |
1435 | return page; | |
1436 | } | |
1437 | ||
91f47662 | 1438 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
aa888a74 AK |
1439 | { |
1440 | struct huge_bootmem_page *m; | |
b2261026 | 1441 | int nr_nodes, node; |
aa888a74 | 1442 | |
b2261026 | 1443 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
1444 | void *addr; |
1445 | ||
8b89a116 GS |
1446 | addr = memblock_virt_alloc_try_nid_nopanic( |
1447 | huge_page_size(h), huge_page_size(h), | |
1448 | 0, BOOTMEM_ALLOC_ACCESSIBLE, node); | |
aa888a74 AK |
1449 | if (addr) { |
1450 | /* | |
1451 | * Use the beginning of the huge page to store the | |
1452 | * huge_bootmem_page struct (until gather_bootmem | |
1453 | * puts them into the mem_map). | |
1454 | */ | |
1455 | m = addr; | |
91f47662 | 1456 | goto found; |
aa888a74 | 1457 | } |
aa888a74 AK |
1458 | } |
1459 | return 0; | |
1460 | ||
1461 | found: | |
df994ead | 1462 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 AK |
1463 | /* Put them into a private list first because mem_map is not up yet */ |
1464 | list_add(&m->list, &huge_boot_pages); | |
1465 | m->hstate = h; | |
1466 | return 1; | |
1467 | } | |
1468 | ||
f412c97a | 1469 | static void __init prep_compound_huge_page(struct page *page, int order) |
18229df5 AW |
1470 | { |
1471 | if (unlikely(order > (MAX_ORDER - 1))) | |
1472 | prep_compound_gigantic_page(page, order); | |
1473 | else | |
1474 | prep_compound_page(page, order); | |
1475 | } | |
1476 | ||
aa888a74 AK |
1477 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
1478 | static void __init gather_bootmem_prealloc(void) | |
1479 | { | |
1480 | struct huge_bootmem_page *m; | |
1481 | ||
1482 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 1483 | struct hstate *h = m->hstate; |
ee8f248d BB |
1484 | struct page *page; |
1485 | ||
1486 | #ifdef CONFIG_HIGHMEM | |
1487 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
8b89a116 GS |
1488 | memblock_free_late(__pa(m), |
1489 | sizeof(struct huge_bootmem_page)); | |
ee8f248d BB |
1490 | #else |
1491 | page = virt_to_page(m); | |
1492 | #endif | |
aa888a74 | 1493 | WARN_ON(page_count(page) != 1); |
18229df5 | 1494 | prep_compound_huge_page(page, h->order); |
ef5a22be | 1495 | WARN_ON(PageReserved(page)); |
aa888a74 | 1496 | prep_new_huge_page(h, page, page_to_nid(page)); |
b0320c7b RA |
1497 | /* |
1498 | * If we had gigantic hugepages allocated at boot time, we need | |
1499 | * to restore the 'stolen' pages to totalram_pages in order to | |
1500 | * fix confusing memory reports from free(1) and another | |
1501 | * side-effects, like CommitLimit going negative. | |
1502 | */ | |
bae7f4ae | 1503 | if (hstate_is_gigantic(h)) |
3dcc0571 | 1504 | adjust_managed_page_count(page, 1 << h->order); |
aa888a74 AK |
1505 | } |
1506 | } | |
1507 | ||
8faa8b07 | 1508 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
1509 | { |
1510 | unsigned long i; | |
a5516438 | 1511 | |
e5ff2159 | 1512 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 1513 | if (hstate_is_gigantic(h)) { |
aa888a74 AK |
1514 | if (!alloc_bootmem_huge_page(h)) |
1515 | break; | |
9b5e5d0f | 1516 | } else if (!alloc_fresh_huge_page(h, |
8cebfcd0 | 1517 | &node_states[N_MEMORY])) |
1da177e4 | 1518 | break; |
1da177e4 | 1519 | } |
8faa8b07 | 1520 | h->max_huge_pages = i; |
e5ff2159 AK |
1521 | } |
1522 | ||
1523 | static void __init hugetlb_init_hstates(void) | |
1524 | { | |
1525 | struct hstate *h; | |
1526 | ||
1527 | for_each_hstate(h) { | |
8faa8b07 | 1528 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 1529 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 1530 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 AK |
1531 | } |
1532 | } | |
1533 | ||
4abd32db AK |
1534 | static char * __init memfmt(char *buf, unsigned long n) |
1535 | { | |
1536 | if (n >= (1UL << 30)) | |
1537 | sprintf(buf, "%lu GB", n >> 30); | |
1538 | else if (n >= (1UL << 20)) | |
1539 | sprintf(buf, "%lu MB", n >> 20); | |
1540 | else | |
1541 | sprintf(buf, "%lu KB", n >> 10); | |
1542 | return buf; | |
1543 | } | |
1544 | ||
e5ff2159 AK |
1545 | static void __init report_hugepages(void) |
1546 | { | |
1547 | struct hstate *h; | |
1548 | ||
1549 | for_each_hstate(h) { | |
4abd32db | 1550 | char buf[32]; |
ffb22af5 | 1551 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
4abd32db AK |
1552 | memfmt(buf, huge_page_size(h)), |
1553 | h->free_huge_pages); | |
e5ff2159 AK |
1554 | } |
1555 | } | |
1556 | ||
1da177e4 | 1557 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
1558 | static void try_to_free_low(struct hstate *h, unsigned long count, |
1559 | nodemask_t *nodes_allowed) | |
1da177e4 | 1560 | { |
4415cc8d CL |
1561 | int i; |
1562 | ||
bae7f4ae | 1563 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1564 | return; |
1565 | ||
6ae11b27 | 1566 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 1567 | struct page *page, *next; |
a5516438 AK |
1568 | struct list_head *freel = &h->hugepage_freelists[i]; |
1569 | list_for_each_entry_safe(page, next, freel, lru) { | |
1570 | if (count >= h->nr_huge_pages) | |
6b0c880d | 1571 | return; |
1da177e4 LT |
1572 | if (PageHighMem(page)) |
1573 | continue; | |
1574 | list_del(&page->lru); | |
e5ff2159 | 1575 | update_and_free_page(h, page); |
a5516438 AK |
1576 | h->free_huge_pages--; |
1577 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
1578 | } |
1579 | } | |
1580 | } | |
1581 | #else | |
6ae11b27 LS |
1582 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
1583 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
1584 | { |
1585 | } | |
1586 | #endif | |
1587 | ||
20a0307c WF |
1588 | /* |
1589 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
1590 | * balanced by operating on them in a round-robin fashion. | |
1591 | * Returns 1 if an adjustment was made. | |
1592 | */ | |
6ae11b27 LS |
1593 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
1594 | int delta) | |
20a0307c | 1595 | { |
b2261026 | 1596 | int nr_nodes, node; |
20a0307c WF |
1597 | |
1598 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 1599 | |
b2261026 JK |
1600 | if (delta < 0) { |
1601 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1602 | if (h->surplus_huge_pages_node[node]) | |
1603 | goto found; | |
e8c5c824 | 1604 | } |
b2261026 JK |
1605 | } else { |
1606 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
1607 | if (h->surplus_huge_pages_node[node] < | |
1608 | h->nr_huge_pages_node[node]) | |
1609 | goto found; | |
e8c5c824 | 1610 | } |
b2261026 JK |
1611 | } |
1612 | return 0; | |
20a0307c | 1613 | |
b2261026 JK |
1614 | found: |
1615 | h->surplus_huge_pages += delta; | |
1616 | h->surplus_huge_pages_node[node] += delta; | |
1617 | return 1; | |
20a0307c WF |
1618 | } |
1619 | ||
a5516438 | 1620 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
1621 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
1622 | nodemask_t *nodes_allowed) | |
1da177e4 | 1623 | { |
7893d1d5 | 1624 | unsigned long min_count, ret; |
1da177e4 | 1625 | |
944d9fec | 1626 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
aa888a74 AK |
1627 | return h->max_huge_pages; |
1628 | ||
7893d1d5 AL |
1629 | /* |
1630 | * Increase the pool size | |
1631 | * First take pages out of surplus state. Then make up the | |
1632 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f NA |
1633 | * |
1634 | * We might race with alloc_buddy_huge_page() here and be unable | |
1635 | * to convert a surplus huge page to a normal huge page. That is | |
1636 | * not critical, though, it just means the overall size of the | |
1637 | * pool might be one hugepage larger than it needs to be, but | |
1638 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 1639 | */ |
1da177e4 | 1640 | spin_lock(&hugetlb_lock); |
a5516438 | 1641 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 1642 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
1643 | break; |
1644 | } | |
1645 | ||
a5516438 | 1646 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
1647 | /* |
1648 | * If this allocation races such that we no longer need the | |
1649 | * page, free_huge_page will handle it by freeing the page | |
1650 | * and reducing the surplus. | |
1651 | */ | |
1652 | spin_unlock(&hugetlb_lock); | |
944d9fec LC |
1653 | if (hstate_is_gigantic(h)) |
1654 | ret = alloc_fresh_gigantic_page(h, nodes_allowed); | |
1655 | else | |
1656 | ret = alloc_fresh_huge_page(h, nodes_allowed); | |
7893d1d5 AL |
1657 | spin_lock(&hugetlb_lock); |
1658 | if (!ret) | |
1659 | goto out; | |
1660 | ||
536240f2 MG |
1661 | /* Bail for signals. Probably ctrl-c from user */ |
1662 | if (signal_pending(current)) | |
1663 | goto out; | |
7893d1d5 | 1664 | } |
7893d1d5 AL |
1665 | |
1666 | /* | |
1667 | * Decrease the pool size | |
1668 | * First return free pages to the buddy allocator (being careful | |
1669 | * to keep enough around to satisfy reservations). Then place | |
1670 | * pages into surplus state as needed so the pool will shrink | |
1671 | * to the desired size as pages become free. | |
d1c3fb1f NA |
1672 | * |
1673 | * By placing pages into the surplus state independent of the | |
1674 | * overcommit value, we are allowing the surplus pool size to | |
1675 | * exceed overcommit. There are few sane options here. Since | |
1676 | * alloc_buddy_huge_page() is checking the global counter, | |
1677 | * though, we'll note that we're not allowed to exceed surplus | |
1678 | * and won't grow the pool anywhere else. Not until one of the | |
1679 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 1680 | */ |
a5516438 | 1681 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 1682 | min_count = max(count, min_count); |
6ae11b27 | 1683 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 1684 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 1685 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 1686 | break; |
55f67141 | 1687 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 1688 | } |
a5516438 | 1689 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 1690 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
1691 | break; |
1692 | } | |
1693 | out: | |
a5516438 | 1694 | ret = persistent_huge_pages(h); |
1da177e4 | 1695 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 1696 | return ret; |
1da177e4 LT |
1697 | } |
1698 | ||
a3437870 NA |
1699 | #define HSTATE_ATTR_RO(_name) \ |
1700 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
1701 | ||
1702 | #define HSTATE_ATTR(_name) \ | |
1703 | static struct kobj_attribute _name##_attr = \ | |
1704 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
1705 | ||
1706 | static struct kobject *hugepages_kobj; | |
1707 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1708 | ||
9a305230 LS |
1709 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
1710 | ||
1711 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
1712 | { |
1713 | int i; | |
9a305230 | 1714 | |
a3437870 | 1715 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
1716 | if (hstate_kobjs[i] == kobj) { |
1717 | if (nidp) | |
1718 | *nidp = NUMA_NO_NODE; | |
a3437870 | 1719 | return &hstates[i]; |
9a305230 LS |
1720 | } |
1721 | ||
1722 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
1723 | } |
1724 | ||
06808b08 | 1725 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
1726 | struct kobj_attribute *attr, char *buf) |
1727 | { | |
9a305230 LS |
1728 | struct hstate *h; |
1729 | unsigned long nr_huge_pages; | |
1730 | int nid; | |
1731 | ||
1732 | h = kobj_to_hstate(kobj, &nid); | |
1733 | if (nid == NUMA_NO_NODE) | |
1734 | nr_huge_pages = h->nr_huge_pages; | |
1735 | else | |
1736 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
1737 | ||
1738 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 1739 | } |
adbe8726 | 1740 | |
238d3c13 DR |
1741 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
1742 | struct hstate *h, int nid, | |
1743 | unsigned long count, size_t len) | |
a3437870 NA |
1744 | { |
1745 | int err; | |
bad44b5b | 1746 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 1747 | |
944d9fec | 1748 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) { |
adbe8726 EM |
1749 | err = -EINVAL; |
1750 | goto out; | |
1751 | } | |
1752 | ||
9a305230 LS |
1753 | if (nid == NUMA_NO_NODE) { |
1754 | /* | |
1755 | * global hstate attribute | |
1756 | */ | |
1757 | if (!(obey_mempolicy && | |
1758 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
1759 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 1760 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 LS |
1761 | } |
1762 | } else if (nodes_allowed) { | |
1763 | /* | |
1764 | * per node hstate attribute: adjust count to global, | |
1765 | * but restrict alloc/free to the specified node. | |
1766 | */ | |
1767 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
1768 | init_nodemask_of_node(nodes_allowed, nid); | |
1769 | } else | |
8cebfcd0 | 1770 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 | 1771 | |
06808b08 | 1772 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 1773 | |
8cebfcd0 | 1774 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
1775 | NODEMASK_FREE(nodes_allowed); |
1776 | ||
1777 | return len; | |
adbe8726 EM |
1778 | out: |
1779 | NODEMASK_FREE(nodes_allowed); | |
1780 | return err; | |
06808b08 LS |
1781 | } |
1782 | ||
238d3c13 DR |
1783 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
1784 | struct kobject *kobj, const char *buf, | |
1785 | size_t len) | |
1786 | { | |
1787 | struct hstate *h; | |
1788 | unsigned long count; | |
1789 | int nid; | |
1790 | int err; | |
1791 | ||
1792 | err = kstrtoul(buf, 10, &count); | |
1793 | if (err) | |
1794 | return err; | |
1795 | ||
1796 | h = kobj_to_hstate(kobj, &nid); | |
1797 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
1798 | } | |
1799 | ||
06808b08 LS |
1800 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
1801 | struct kobj_attribute *attr, char *buf) | |
1802 | { | |
1803 | return nr_hugepages_show_common(kobj, attr, buf); | |
1804 | } | |
1805 | ||
1806 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
1807 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1808 | { | |
238d3c13 | 1809 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
1810 | } |
1811 | HSTATE_ATTR(nr_hugepages); | |
1812 | ||
06808b08 LS |
1813 | #ifdef CONFIG_NUMA |
1814 | ||
1815 | /* | |
1816 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
1817 | * huge page alloc/free. | |
1818 | */ | |
1819 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
1820 | struct kobj_attribute *attr, char *buf) | |
1821 | { | |
1822 | return nr_hugepages_show_common(kobj, attr, buf); | |
1823 | } | |
1824 | ||
1825 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
1826 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1827 | { | |
238d3c13 | 1828 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
1829 | } |
1830 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
1831 | #endif | |
1832 | ||
1833 | ||
a3437870 NA |
1834 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
1835 | struct kobj_attribute *attr, char *buf) | |
1836 | { | |
9a305230 | 1837 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1838 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
1839 | } | |
adbe8726 | 1840 | |
a3437870 NA |
1841 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
1842 | struct kobj_attribute *attr, const char *buf, size_t count) | |
1843 | { | |
1844 | int err; | |
1845 | unsigned long input; | |
9a305230 | 1846 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 1847 | |
bae7f4ae | 1848 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
1849 | return -EINVAL; |
1850 | ||
3dbb95f7 | 1851 | err = kstrtoul(buf, 10, &input); |
a3437870 | 1852 | if (err) |
73ae31e5 | 1853 | return err; |
a3437870 NA |
1854 | |
1855 | spin_lock(&hugetlb_lock); | |
1856 | h->nr_overcommit_huge_pages = input; | |
1857 | spin_unlock(&hugetlb_lock); | |
1858 | ||
1859 | return count; | |
1860 | } | |
1861 | HSTATE_ATTR(nr_overcommit_hugepages); | |
1862 | ||
1863 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
1864 | struct kobj_attribute *attr, char *buf) | |
1865 | { | |
9a305230 LS |
1866 | struct hstate *h; |
1867 | unsigned long free_huge_pages; | |
1868 | int nid; | |
1869 | ||
1870 | h = kobj_to_hstate(kobj, &nid); | |
1871 | if (nid == NUMA_NO_NODE) | |
1872 | free_huge_pages = h->free_huge_pages; | |
1873 | else | |
1874 | free_huge_pages = h->free_huge_pages_node[nid]; | |
1875 | ||
1876 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
1877 | } |
1878 | HSTATE_ATTR_RO(free_hugepages); | |
1879 | ||
1880 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
1881 | struct kobj_attribute *attr, char *buf) | |
1882 | { | |
9a305230 | 1883 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1884 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
1885 | } | |
1886 | HSTATE_ATTR_RO(resv_hugepages); | |
1887 | ||
1888 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
1889 | struct kobj_attribute *attr, char *buf) | |
1890 | { | |
9a305230 LS |
1891 | struct hstate *h; |
1892 | unsigned long surplus_huge_pages; | |
1893 | int nid; | |
1894 | ||
1895 | h = kobj_to_hstate(kobj, &nid); | |
1896 | if (nid == NUMA_NO_NODE) | |
1897 | surplus_huge_pages = h->surplus_huge_pages; | |
1898 | else | |
1899 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
1900 | ||
1901 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
1902 | } |
1903 | HSTATE_ATTR_RO(surplus_hugepages); | |
1904 | ||
1905 | static struct attribute *hstate_attrs[] = { | |
1906 | &nr_hugepages_attr.attr, | |
1907 | &nr_overcommit_hugepages_attr.attr, | |
1908 | &free_hugepages_attr.attr, | |
1909 | &resv_hugepages_attr.attr, | |
1910 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
1911 | #ifdef CONFIG_NUMA |
1912 | &nr_hugepages_mempolicy_attr.attr, | |
1913 | #endif | |
a3437870 NA |
1914 | NULL, |
1915 | }; | |
1916 | ||
1917 | static struct attribute_group hstate_attr_group = { | |
1918 | .attrs = hstate_attrs, | |
1919 | }; | |
1920 | ||
094e9539 JM |
1921 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
1922 | struct kobject **hstate_kobjs, | |
1923 | struct attribute_group *hstate_attr_group) | |
a3437870 NA |
1924 | { |
1925 | int retval; | |
972dc4de | 1926 | int hi = hstate_index(h); |
a3437870 | 1927 | |
9a305230 LS |
1928 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
1929 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
1930 | return -ENOMEM; |
1931 | ||
9a305230 | 1932 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 1933 | if (retval) |
9a305230 | 1934 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
1935 | |
1936 | return retval; | |
1937 | } | |
1938 | ||
1939 | static void __init hugetlb_sysfs_init(void) | |
1940 | { | |
1941 | struct hstate *h; | |
1942 | int err; | |
1943 | ||
1944 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
1945 | if (!hugepages_kobj) | |
1946 | return; | |
1947 | ||
1948 | for_each_hstate(h) { | |
9a305230 LS |
1949 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
1950 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 1951 | if (err) |
ffb22af5 | 1952 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
1953 | } |
1954 | } | |
1955 | ||
9a305230 LS |
1956 | #ifdef CONFIG_NUMA |
1957 | ||
1958 | /* | |
1959 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
1960 | * with node devices in node_devices[] using a parallel array. The array |
1961 | * index of a node device or _hstate == node id. | |
1962 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
1963 | * the base kernel, on the hugetlb module. |
1964 | */ | |
1965 | struct node_hstate { | |
1966 | struct kobject *hugepages_kobj; | |
1967 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1968 | }; | |
1969 | struct node_hstate node_hstates[MAX_NUMNODES]; | |
1970 | ||
1971 | /* | |
10fbcf4c | 1972 | * A subset of global hstate attributes for node devices |
9a305230 LS |
1973 | */ |
1974 | static struct attribute *per_node_hstate_attrs[] = { | |
1975 | &nr_hugepages_attr.attr, | |
1976 | &free_hugepages_attr.attr, | |
1977 | &surplus_hugepages_attr.attr, | |
1978 | NULL, | |
1979 | }; | |
1980 | ||
1981 | static struct attribute_group per_node_hstate_attr_group = { | |
1982 | .attrs = per_node_hstate_attrs, | |
1983 | }; | |
1984 | ||
1985 | /* | |
10fbcf4c | 1986 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
1987 | * Returns node id via non-NULL nidp. |
1988 | */ | |
1989 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
1990 | { | |
1991 | int nid; | |
1992 | ||
1993 | for (nid = 0; nid < nr_node_ids; nid++) { | |
1994 | struct node_hstate *nhs = &node_hstates[nid]; | |
1995 | int i; | |
1996 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
1997 | if (nhs->hstate_kobjs[i] == kobj) { | |
1998 | if (nidp) | |
1999 | *nidp = nid; | |
2000 | return &hstates[i]; | |
2001 | } | |
2002 | } | |
2003 | ||
2004 | BUG(); | |
2005 | return NULL; | |
2006 | } | |
2007 | ||
2008 | /* | |
10fbcf4c | 2009 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
2010 | * No-op if no hstate attributes attached. |
2011 | */ | |
3cd8b44f | 2012 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
2013 | { |
2014 | struct hstate *h; | |
10fbcf4c | 2015 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2016 | |
2017 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 2018 | return; /* no hstate attributes */ |
9a305230 | 2019 | |
972dc4de AK |
2020 | for_each_hstate(h) { |
2021 | int idx = hstate_index(h); | |
2022 | if (nhs->hstate_kobjs[idx]) { | |
2023 | kobject_put(nhs->hstate_kobjs[idx]); | |
2024 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 2025 | } |
972dc4de | 2026 | } |
9a305230 LS |
2027 | |
2028 | kobject_put(nhs->hugepages_kobj); | |
2029 | nhs->hugepages_kobj = NULL; | |
2030 | } | |
2031 | ||
2032 | /* | |
10fbcf4c | 2033 | * hugetlb module exit: unregister hstate attributes from node devices |
9a305230 LS |
2034 | * that have them. |
2035 | */ | |
2036 | static void hugetlb_unregister_all_nodes(void) | |
2037 | { | |
2038 | int nid; | |
2039 | ||
2040 | /* | |
10fbcf4c | 2041 | * disable node device registrations. |
9a305230 LS |
2042 | */ |
2043 | register_hugetlbfs_with_node(NULL, NULL); | |
2044 | ||
2045 | /* | |
2046 | * remove hstate attributes from any nodes that have them. | |
2047 | */ | |
2048 | for (nid = 0; nid < nr_node_ids; nid++) | |
8732794b | 2049 | hugetlb_unregister_node(node_devices[nid]); |
9a305230 LS |
2050 | } |
2051 | ||
2052 | /* | |
10fbcf4c | 2053 | * Register hstate attributes for a single node device. |
9a305230 LS |
2054 | * No-op if attributes already registered. |
2055 | */ | |
3cd8b44f | 2056 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
2057 | { |
2058 | struct hstate *h; | |
10fbcf4c | 2059 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2060 | int err; |
2061 | ||
2062 | if (nhs->hugepages_kobj) | |
2063 | return; /* already allocated */ | |
2064 | ||
2065 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 2066 | &node->dev.kobj); |
9a305230 LS |
2067 | if (!nhs->hugepages_kobj) |
2068 | return; | |
2069 | ||
2070 | for_each_hstate(h) { | |
2071 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
2072 | nhs->hstate_kobjs, | |
2073 | &per_node_hstate_attr_group); | |
2074 | if (err) { | |
ffb22af5 AM |
2075 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
2076 | h->name, node->dev.id); | |
9a305230 LS |
2077 | hugetlb_unregister_node(node); |
2078 | break; | |
2079 | } | |
2080 | } | |
2081 | } | |
2082 | ||
2083 | /* | |
9b5e5d0f | 2084 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
2085 | * devices of nodes that have memory. All on-line nodes should have |
2086 | * registered their associated device by this time. | |
9a305230 | 2087 | */ |
7d9ca000 | 2088 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
2089 | { |
2090 | int nid; | |
2091 | ||
8cebfcd0 | 2092 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 2093 | struct node *node = node_devices[nid]; |
10fbcf4c | 2094 | if (node->dev.id == nid) |
9a305230 LS |
2095 | hugetlb_register_node(node); |
2096 | } | |
2097 | ||
2098 | /* | |
10fbcf4c | 2099 | * Let the node device driver know we're here so it can |
9a305230 LS |
2100 | * [un]register hstate attributes on node hotplug. |
2101 | */ | |
2102 | register_hugetlbfs_with_node(hugetlb_register_node, | |
2103 | hugetlb_unregister_node); | |
2104 | } | |
2105 | #else /* !CONFIG_NUMA */ | |
2106 | ||
2107 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2108 | { | |
2109 | BUG(); | |
2110 | if (nidp) | |
2111 | *nidp = -1; | |
2112 | return NULL; | |
2113 | } | |
2114 | ||
2115 | static void hugetlb_unregister_all_nodes(void) { } | |
2116 | ||
2117 | static void hugetlb_register_all_nodes(void) { } | |
2118 | ||
2119 | #endif | |
2120 | ||
a3437870 NA |
2121 | static void __exit hugetlb_exit(void) |
2122 | { | |
2123 | struct hstate *h; | |
2124 | ||
9a305230 LS |
2125 | hugetlb_unregister_all_nodes(); |
2126 | ||
a3437870 | 2127 | for_each_hstate(h) { |
972dc4de | 2128 | kobject_put(hstate_kobjs[hstate_index(h)]); |
a3437870 NA |
2129 | } |
2130 | ||
2131 | kobject_put(hugepages_kobj); | |
8382d914 | 2132 | kfree(htlb_fault_mutex_table); |
a3437870 NA |
2133 | } |
2134 | module_exit(hugetlb_exit); | |
2135 | ||
2136 | static int __init hugetlb_init(void) | |
2137 | { | |
8382d914 DB |
2138 | int i; |
2139 | ||
457c1b27 | 2140 | if (!hugepages_supported()) |
0ef89d25 | 2141 | return 0; |
a3437870 | 2142 | |
e11bfbfc NP |
2143 | if (!size_to_hstate(default_hstate_size)) { |
2144 | default_hstate_size = HPAGE_SIZE; | |
2145 | if (!size_to_hstate(default_hstate_size)) | |
2146 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 2147 | } |
972dc4de | 2148 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
e11bfbfc NP |
2149 | if (default_hstate_max_huge_pages) |
2150 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
a3437870 NA |
2151 | |
2152 | hugetlb_init_hstates(); | |
aa888a74 | 2153 | gather_bootmem_prealloc(); |
a3437870 NA |
2154 | report_hugepages(); |
2155 | ||
2156 | hugetlb_sysfs_init(); | |
9a305230 | 2157 | hugetlb_register_all_nodes(); |
7179e7bf | 2158 | hugetlb_cgroup_file_init(); |
9a305230 | 2159 | |
8382d914 DB |
2160 | #ifdef CONFIG_SMP |
2161 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
2162 | #else | |
2163 | num_fault_mutexes = 1; | |
2164 | #endif | |
2165 | htlb_fault_mutex_table = | |
2166 | kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); | |
2167 | BUG_ON(!htlb_fault_mutex_table); | |
2168 | ||
2169 | for (i = 0; i < num_fault_mutexes; i++) | |
2170 | mutex_init(&htlb_fault_mutex_table[i]); | |
a3437870 NA |
2171 | return 0; |
2172 | } | |
2173 | module_init(hugetlb_init); | |
2174 | ||
2175 | /* Should be called on processing a hugepagesz=... option */ | |
2176 | void __init hugetlb_add_hstate(unsigned order) | |
2177 | { | |
2178 | struct hstate *h; | |
8faa8b07 AK |
2179 | unsigned long i; |
2180 | ||
a3437870 | 2181 | if (size_to_hstate(PAGE_SIZE << order)) { |
ffb22af5 | 2182 | pr_warning("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
2183 | return; |
2184 | } | |
47d38344 | 2185 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 2186 | BUG_ON(order == 0); |
47d38344 | 2187 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
2188 | h->order = order; |
2189 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
2190 | h->nr_huge_pages = 0; |
2191 | h->free_huge_pages = 0; | |
2192 | for (i = 0; i < MAX_NUMNODES; ++i) | |
2193 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 2194 | INIT_LIST_HEAD(&h->hugepage_activelist); |
8cebfcd0 LJ |
2195 | h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); |
2196 | h->next_nid_to_free = first_node(node_states[N_MEMORY]); | |
a3437870 NA |
2197 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
2198 | huge_page_size(h)/1024); | |
8faa8b07 | 2199 | |
a3437870 NA |
2200 | parsed_hstate = h; |
2201 | } | |
2202 | ||
e11bfbfc | 2203 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
2204 | { |
2205 | unsigned long *mhp; | |
8faa8b07 | 2206 | static unsigned long *last_mhp; |
a3437870 NA |
2207 | |
2208 | /* | |
47d38344 | 2209 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
2210 | * so this hugepages= parameter goes to the "default hstate". |
2211 | */ | |
47d38344 | 2212 | if (!hugetlb_max_hstate) |
a3437870 NA |
2213 | mhp = &default_hstate_max_huge_pages; |
2214 | else | |
2215 | mhp = &parsed_hstate->max_huge_pages; | |
2216 | ||
8faa8b07 | 2217 | if (mhp == last_mhp) { |
ffb22af5 AM |
2218 | pr_warning("hugepages= specified twice without " |
2219 | "interleaving hugepagesz=, ignoring\n"); | |
8faa8b07 AK |
2220 | return 1; |
2221 | } | |
2222 | ||
a3437870 NA |
2223 | if (sscanf(s, "%lu", mhp) <= 0) |
2224 | *mhp = 0; | |
2225 | ||
8faa8b07 AK |
2226 | /* |
2227 | * Global state is always initialized later in hugetlb_init. | |
2228 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
2229 | * use the bootmem allocator. | |
2230 | */ | |
47d38344 | 2231 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
2232 | hugetlb_hstate_alloc_pages(parsed_hstate); |
2233 | ||
2234 | last_mhp = mhp; | |
2235 | ||
a3437870 NA |
2236 | return 1; |
2237 | } | |
e11bfbfc NP |
2238 | __setup("hugepages=", hugetlb_nrpages_setup); |
2239 | ||
2240 | static int __init hugetlb_default_setup(char *s) | |
2241 | { | |
2242 | default_hstate_size = memparse(s, &s); | |
2243 | return 1; | |
2244 | } | |
2245 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 2246 | |
8a213460 NA |
2247 | static unsigned int cpuset_mems_nr(unsigned int *array) |
2248 | { | |
2249 | int node; | |
2250 | unsigned int nr = 0; | |
2251 | ||
2252 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
2253 | nr += array[node]; | |
2254 | ||
2255 | return nr; | |
2256 | } | |
2257 | ||
2258 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
2259 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
2260 | struct ctl_table *table, int write, | |
2261 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 2262 | { |
e5ff2159 | 2263 | struct hstate *h = &default_hstate; |
238d3c13 | 2264 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 2265 | int ret; |
e5ff2159 | 2266 | |
457c1b27 NA |
2267 | if (!hugepages_supported()) |
2268 | return -ENOTSUPP; | |
2269 | ||
e5ff2159 AK |
2270 | table->data = &tmp; |
2271 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2272 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2273 | if (ret) | |
2274 | goto out; | |
e5ff2159 | 2275 | |
238d3c13 DR |
2276 | if (write) |
2277 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
2278 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
2279 | out: |
2280 | return ret; | |
1da177e4 | 2281 | } |
396faf03 | 2282 | |
06808b08 LS |
2283 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2284 | void __user *buffer, size_t *length, loff_t *ppos) | |
2285 | { | |
2286 | ||
2287 | return hugetlb_sysctl_handler_common(false, table, write, | |
2288 | buffer, length, ppos); | |
2289 | } | |
2290 | ||
2291 | #ifdef CONFIG_NUMA | |
2292 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2293 | void __user *buffer, size_t *length, loff_t *ppos) | |
2294 | { | |
2295 | return hugetlb_sysctl_handler_common(true, table, write, | |
2296 | buffer, length, ppos); | |
2297 | } | |
2298 | #endif /* CONFIG_NUMA */ | |
2299 | ||
a3d0c6aa | 2300 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 2301 | void __user *buffer, |
a3d0c6aa NA |
2302 | size_t *length, loff_t *ppos) |
2303 | { | |
a5516438 | 2304 | struct hstate *h = &default_hstate; |
e5ff2159 | 2305 | unsigned long tmp; |
08d4a246 | 2306 | int ret; |
e5ff2159 | 2307 | |
457c1b27 NA |
2308 | if (!hugepages_supported()) |
2309 | return -ENOTSUPP; | |
2310 | ||
c033a93c | 2311 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 2312 | |
bae7f4ae | 2313 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
2314 | return -EINVAL; |
2315 | ||
e5ff2159 AK |
2316 | table->data = &tmp; |
2317 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2318 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2319 | if (ret) | |
2320 | goto out; | |
e5ff2159 AK |
2321 | |
2322 | if (write) { | |
2323 | spin_lock(&hugetlb_lock); | |
2324 | h->nr_overcommit_huge_pages = tmp; | |
2325 | spin_unlock(&hugetlb_lock); | |
2326 | } | |
08d4a246 MH |
2327 | out: |
2328 | return ret; | |
a3d0c6aa NA |
2329 | } |
2330 | ||
1da177e4 LT |
2331 | #endif /* CONFIG_SYSCTL */ |
2332 | ||
e1759c21 | 2333 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 2334 | { |
a5516438 | 2335 | struct hstate *h = &default_hstate; |
457c1b27 NA |
2336 | if (!hugepages_supported()) |
2337 | return; | |
e1759c21 | 2338 | seq_printf(m, |
4f98a2fe RR |
2339 | "HugePages_Total: %5lu\n" |
2340 | "HugePages_Free: %5lu\n" | |
2341 | "HugePages_Rsvd: %5lu\n" | |
2342 | "HugePages_Surp: %5lu\n" | |
2343 | "Hugepagesize: %8lu kB\n", | |
a5516438 AK |
2344 | h->nr_huge_pages, |
2345 | h->free_huge_pages, | |
2346 | h->resv_huge_pages, | |
2347 | h->surplus_huge_pages, | |
2348 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
2349 | } |
2350 | ||
2351 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
2352 | { | |
a5516438 | 2353 | struct hstate *h = &default_hstate; |
457c1b27 NA |
2354 | if (!hugepages_supported()) |
2355 | return 0; | |
1da177e4 LT |
2356 | return sprintf(buf, |
2357 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
2358 | "Node %d HugePages_Free: %5u\n" |
2359 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
2360 | nid, h->nr_huge_pages_node[nid], |
2361 | nid, h->free_huge_pages_node[nid], | |
2362 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
2363 | } |
2364 | ||
949f7ec5 DR |
2365 | void hugetlb_show_meminfo(void) |
2366 | { | |
2367 | struct hstate *h; | |
2368 | int nid; | |
2369 | ||
457c1b27 NA |
2370 | if (!hugepages_supported()) |
2371 | return; | |
2372 | ||
949f7ec5 DR |
2373 | for_each_node_state(nid, N_MEMORY) |
2374 | for_each_hstate(h) | |
2375 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
2376 | nid, | |
2377 | h->nr_huge_pages_node[nid], | |
2378 | h->free_huge_pages_node[nid], | |
2379 | h->surplus_huge_pages_node[nid], | |
2380 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
2381 | } | |
2382 | ||
1da177e4 LT |
2383 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
2384 | unsigned long hugetlb_total_pages(void) | |
2385 | { | |
d0028588 WL |
2386 | struct hstate *h; |
2387 | unsigned long nr_total_pages = 0; | |
2388 | ||
2389 | for_each_hstate(h) | |
2390 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
2391 | return nr_total_pages; | |
1da177e4 | 2392 | } |
1da177e4 | 2393 | |
a5516438 | 2394 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
2395 | { |
2396 | int ret = -ENOMEM; | |
2397 | ||
2398 | spin_lock(&hugetlb_lock); | |
2399 | /* | |
2400 | * When cpuset is configured, it breaks the strict hugetlb page | |
2401 | * reservation as the accounting is done on a global variable. Such | |
2402 | * reservation is completely rubbish in the presence of cpuset because | |
2403 | * the reservation is not checked against page availability for the | |
2404 | * current cpuset. Application can still potentially OOM'ed by kernel | |
2405 | * with lack of free htlb page in cpuset that the task is in. | |
2406 | * Attempt to enforce strict accounting with cpuset is almost | |
2407 | * impossible (or too ugly) because cpuset is too fluid that | |
2408 | * task or memory node can be dynamically moved between cpusets. | |
2409 | * | |
2410 | * The change of semantics for shared hugetlb mapping with cpuset is | |
2411 | * undesirable. However, in order to preserve some of the semantics, | |
2412 | * we fall back to check against current free page availability as | |
2413 | * a best attempt and hopefully to minimize the impact of changing | |
2414 | * semantics that cpuset has. | |
2415 | */ | |
2416 | if (delta > 0) { | |
a5516438 | 2417 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
2418 | goto out; |
2419 | ||
a5516438 AK |
2420 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
2421 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
2422 | goto out; |
2423 | } | |
2424 | } | |
2425 | ||
2426 | ret = 0; | |
2427 | if (delta < 0) | |
a5516438 | 2428 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
2429 | |
2430 | out: | |
2431 | spin_unlock(&hugetlb_lock); | |
2432 | return ret; | |
2433 | } | |
2434 | ||
84afd99b AW |
2435 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
2436 | { | |
f522c3ac | 2437 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
2438 | |
2439 | /* | |
2440 | * This new VMA should share its siblings reservation map if present. | |
2441 | * The VMA will only ever have a valid reservation map pointer where | |
2442 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 2443 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
2444 | * after this open call completes. It is therefore safe to take a |
2445 | * new reference here without additional locking. | |
2446 | */ | |
4e35f483 | 2447 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 2448 | kref_get(&resv->refs); |
84afd99b AW |
2449 | } |
2450 | ||
a1e78772 MG |
2451 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
2452 | { | |
a5516438 | 2453 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 2454 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 2455 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 2456 | unsigned long reserve, start, end; |
84afd99b | 2457 | |
4e35f483 JK |
2458 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
2459 | return; | |
84afd99b | 2460 | |
4e35f483 JK |
2461 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
2462 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 2463 | |
4e35f483 | 2464 | reserve = (end - start) - region_count(resv, start, end); |
84afd99b | 2465 | |
4e35f483 JK |
2466 | kref_put(&resv->refs, resv_map_release); |
2467 | ||
2468 | if (reserve) { | |
2469 | hugetlb_acct_memory(h, -reserve); | |
2470 | hugepage_subpool_put_pages(spool, reserve); | |
84afd99b | 2471 | } |
a1e78772 MG |
2472 | } |
2473 | ||
1da177e4 LT |
2474 | /* |
2475 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
2476 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
2477 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
2478 | * this far. | |
2479 | */ | |
d0217ac0 | 2480 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
2481 | { |
2482 | BUG(); | |
d0217ac0 | 2483 | return 0; |
1da177e4 LT |
2484 | } |
2485 | ||
f0f37e2f | 2486 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 2487 | .fault = hugetlb_vm_op_fault, |
84afd99b | 2488 | .open = hugetlb_vm_op_open, |
a1e78772 | 2489 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
2490 | }; |
2491 | ||
1e8f889b DG |
2492 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
2493 | int writable) | |
63551ae0 DG |
2494 | { |
2495 | pte_t entry; | |
2496 | ||
1e8f889b | 2497 | if (writable) { |
106c992a GS |
2498 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
2499 | vma->vm_page_prot))); | |
63551ae0 | 2500 | } else { |
106c992a GS |
2501 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
2502 | vma->vm_page_prot)); | |
63551ae0 DG |
2503 | } |
2504 | entry = pte_mkyoung(entry); | |
2505 | entry = pte_mkhuge(entry); | |
d9ed9faa | 2506 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
2507 | |
2508 | return entry; | |
2509 | } | |
2510 | ||
1e8f889b DG |
2511 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
2512 | unsigned long address, pte_t *ptep) | |
2513 | { | |
2514 | pte_t entry; | |
2515 | ||
106c992a | 2516 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 2517 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 2518 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
2519 | } |
2520 | ||
4a705fef NH |
2521 | static int is_hugetlb_entry_migration(pte_t pte) |
2522 | { | |
2523 | swp_entry_t swp; | |
2524 | ||
2525 | if (huge_pte_none(pte) || pte_present(pte)) | |
2526 | return 0; | |
2527 | swp = pte_to_swp_entry(pte); | |
2528 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
2529 | return 1; | |
2530 | else | |
2531 | return 0; | |
2532 | } | |
2533 | ||
2534 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
2535 | { | |
2536 | swp_entry_t swp; | |
2537 | ||
2538 | if (huge_pte_none(pte) || pte_present(pte)) | |
2539 | return 0; | |
2540 | swp = pte_to_swp_entry(pte); | |
2541 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
2542 | return 1; | |
2543 | else | |
2544 | return 0; | |
2545 | } | |
1e8f889b | 2546 | |
63551ae0 DG |
2547 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
2548 | struct vm_area_struct *vma) | |
2549 | { | |
2550 | pte_t *src_pte, *dst_pte, entry; | |
2551 | struct page *ptepage; | |
1c59827d | 2552 | unsigned long addr; |
1e8f889b | 2553 | int cow; |
a5516438 AK |
2554 | struct hstate *h = hstate_vma(vma); |
2555 | unsigned long sz = huge_page_size(h); | |
e8569dd2 AS |
2556 | unsigned long mmun_start; /* For mmu_notifiers */ |
2557 | unsigned long mmun_end; /* For mmu_notifiers */ | |
2558 | int ret = 0; | |
1e8f889b DG |
2559 | |
2560 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 2561 | |
e8569dd2 AS |
2562 | mmun_start = vma->vm_start; |
2563 | mmun_end = vma->vm_end; | |
2564 | if (cow) | |
2565 | mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); | |
2566 | ||
a5516438 | 2567 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 2568 | spinlock_t *src_ptl, *dst_ptl; |
c74df32c HD |
2569 | src_pte = huge_pte_offset(src, addr); |
2570 | if (!src_pte) | |
2571 | continue; | |
a5516438 | 2572 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
2573 | if (!dst_pte) { |
2574 | ret = -ENOMEM; | |
2575 | break; | |
2576 | } | |
c5c99429 LW |
2577 | |
2578 | /* If the pagetables are shared don't copy or take references */ | |
2579 | if (dst_pte == src_pte) | |
2580 | continue; | |
2581 | ||
cb900f41 KS |
2582 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
2583 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
2584 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef NH |
2585 | entry = huge_ptep_get(src_pte); |
2586 | if (huge_pte_none(entry)) { /* skip none entry */ | |
2587 | ; | |
2588 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
2589 | is_hugetlb_entry_hwpoisoned(entry))) { | |
2590 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
2591 | ||
2592 | if (is_write_migration_entry(swp_entry) && cow) { | |
2593 | /* | |
2594 | * COW mappings require pages in both | |
2595 | * parent and child to be set to read. | |
2596 | */ | |
2597 | make_migration_entry_read(&swp_entry); | |
2598 | entry = swp_entry_to_pte(swp_entry); | |
2599 | set_huge_pte_at(src, addr, src_pte, entry); | |
2600 | } | |
2601 | set_huge_pte_at(dst, addr, dst_pte, entry); | |
2602 | } else { | |
34ee645e | 2603 | if (cow) { |
7f2e9525 | 2604 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e JR |
2605 | mmu_notifier_invalidate_range(src, mmun_start, |
2606 | mmun_end); | |
2607 | } | |
0253d634 | 2608 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
2609 | ptepage = pte_page(entry); |
2610 | get_page(ptepage); | |
0fe6e20b | 2611 | page_dup_rmap(ptepage); |
1c59827d HD |
2612 | set_huge_pte_at(dst, addr, dst_pte, entry); |
2613 | } | |
cb900f41 KS |
2614 | spin_unlock(src_ptl); |
2615 | spin_unlock(dst_ptl); | |
63551ae0 | 2616 | } |
63551ae0 | 2617 | |
e8569dd2 AS |
2618 | if (cow) |
2619 | mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); | |
2620 | ||
2621 | return ret; | |
63551ae0 DG |
2622 | } |
2623 | ||
24669e58 AK |
2624 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
2625 | unsigned long start, unsigned long end, | |
2626 | struct page *ref_page) | |
63551ae0 | 2627 | { |
24669e58 | 2628 | int force_flush = 0; |
63551ae0 DG |
2629 | struct mm_struct *mm = vma->vm_mm; |
2630 | unsigned long address; | |
c7546f8f | 2631 | pte_t *ptep; |
63551ae0 | 2632 | pte_t pte; |
cb900f41 | 2633 | spinlock_t *ptl; |
63551ae0 | 2634 | struct page *page; |
a5516438 AK |
2635 | struct hstate *h = hstate_vma(vma); |
2636 | unsigned long sz = huge_page_size(h); | |
2ec74c3e SG |
2637 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
2638 | const unsigned long mmun_end = end; /* For mmu_notifiers */ | |
a5516438 | 2639 | |
63551ae0 | 2640 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
2641 | BUG_ON(start & ~huge_page_mask(h)); |
2642 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 2643 | |
24669e58 | 2644 | tlb_start_vma(tlb, vma); |
2ec74c3e | 2645 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
569f48b8 | 2646 | address = start; |
24669e58 | 2647 | again: |
569f48b8 | 2648 | for (; address < end; address += sz) { |
c7546f8f | 2649 | ptep = huge_pte_offset(mm, address); |
4c887265 | 2650 | if (!ptep) |
c7546f8f DG |
2651 | continue; |
2652 | ||
cb900f41 | 2653 | ptl = huge_pte_lock(h, mm, ptep); |
39dde65c | 2654 | if (huge_pmd_unshare(mm, &address, ptep)) |
cb900f41 | 2655 | goto unlock; |
39dde65c | 2656 | |
6629326b HD |
2657 | pte = huge_ptep_get(ptep); |
2658 | if (huge_pte_none(pte)) | |
cb900f41 | 2659 | goto unlock; |
6629326b HD |
2660 | |
2661 | /* | |
9fbc1f63 NH |
2662 | * Migrating hugepage or HWPoisoned hugepage is already |
2663 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 2664 | */ |
9fbc1f63 | 2665 | if (unlikely(!pte_present(pte))) { |
106c992a | 2666 | huge_pte_clear(mm, address, ptep); |
cb900f41 | 2667 | goto unlock; |
8c4894c6 | 2668 | } |
6629326b HD |
2669 | |
2670 | page = pte_page(pte); | |
04f2cbe3 MG |
2671 | /* |
2672 | * If a reference page is supplied, it is because a specific | |
2673 | * page is being unmapped, not a range. Ensure the page we | |
2674 | * are about to unmap is the actual page of interest. | |
2675 | */ | |
2676 | if (ref_page) { | |
04f2cbe3 | 2677 | if (page != ref_page) |
cb900f41 | 2678 | goto unlock; |
04f2cbe3 MG |
2679 | |
2680 | /* | |
2681 | * Mark the VMA as having unmapped its page so that | |
2682 | * future faults in this VMA will fail rather than | |
2683 | * looking like data was lost | |
2684 | */ | |
2685 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
2686 | } | |
2687 | ||
c7546f8f | 2688 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
24669e58 | 2689 | tlb_remove_tlb_entry(tlb, ptep, address); |
106c992a | 2690 | if (huge_pte_dirty(pte)) |
6649a386 | 2691 | set_page_dirty(page); |
9e81130b | 2692 | |
24669e58 AK |
2693 | page_remove_rmap(page); |
2694 | force_flush = !__tlb_remove_page(tlb, page); | |
cb900f41 | 2695 | if (force_flush) { |
569f48b8 | 2696 | address += sz; |
cb900f41 | 2697 | spin_unlock(ptl); |
24669e58 | 2698 | break; |
cb900f41 | 2699 | } |
9e81130b | 2700 | /* Bail out after unmapping reference page if supplied */ |
cb900f41 KS |
2701 | if (ref_page) { |
2702 | spin_unlock(ptl); | |
9e81130b | 2703 | break; |
cb900f41 KS |
2704 | } |
2705 | unlock: | |
2706 | spin_unlock(ptl); | |
63551ae0 | 2707 | } |
24669e58 AK |
2708 | /* |
2709 | * mmu_gather ran out of room to batch pages, we break out of | |
2710 | * the PTE lock to avoid doing the potential expensive TLB invalidate | |
2711 | * and page-free while holding it. | |
2712 | */ | |
2713 | if (force_flush) { | |
2714 | force_flush = 0; | |
2715 | tlb_flush_mmu(tlb); | |
2716 | if (address < end && !ref_page) | |
2717 | goto again; | |
fe1668ae | 2718 | } |
2ec74c3e | 2719 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
24669e58 | 2720 | tlb_end_vma(tlb, vma); |
1da177e4 | 2721 | } |
63551ae0 | 2722 | |
d833352a MG |
2723 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
2724 | struct vm_area_struct *vma, unsigned long start, | |
2725 | unsigned long end, struct page *ref_page) | |
2726 | { | |
2727 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
2728 | ||
2729 | /* | |
2730 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
2731 | * test will fail on a vma being torn down, and not grab a page table | |
2732 | * on its way out. We're lucky that the flag has such an appropriate | |
2733 | * name, and can in fact be safely cleared here. We could clear it | |
2734 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 2735 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 2736 | * because in the context this is called, the VMA is about to be |
c8c06efa | 2737 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
2738 | */ |
2739 | vma->vm_flags &= ~VM_MAYSHARE; | |
2740 | } | |
2741 | ||
502717f4 | 2742 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 2743 | unsigned long end, struct page *ref_page) |
502717f4 | 2744 | { |
24669e58 AK |
2745 | struct mm_struct *mm; |
2746 | struct mmu_gather tlb; | |
2747 | ||
2748 | mm = vma->vm_mm; | |
2749 | ||
2b047252 | 2750 | tlb_gather_mmu(&tlb, mm, start, end); |
24669e58 AK |
2751 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
2752 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 CK |
2753 | } |
2754 | ||
04f2cbe3 MG |
2755 | /* |
2756 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
2757 | * mappping it owns the reserve page for. The intention is to unmap the page | |
2758 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
2759 | * same region. | |
2760 | */ | |
2f4612af DB |
2761 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
2762 | struct page *page, unsigned long address) | |
04f2cbe3 | 2763 | { |
7526674d | 2764 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
2765 | struct vm_area_struct *iter_vma; |
2766 | struct address_space *mapping; | |
04f2cbe3 MG |
2767 | pgoff_t pgoff; |
2768 | ||
2769 | /* | |
2770 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
2771 | * from page cache lookup which is in HPAGE_SIZE units. | |
2772 | */ | |
7526674d | 2773 | address = address & huge_page_mask(h); |
36e4f20a MH |
2774 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
2775 | vma->vm_pgoff; | |
496ad9aa | 2776 | mapping = file_inode(vma->vm_file)->i_mapping; |
04f2cbe3 | 2777 | |
4eb2b1dc MG |
2778 | /* |
2779 | * Take the mapping lock for the duration of the table walk. As | |
2780 | * this mapping should be shared between all the VMAs, | |
2781 | * __unmap_hugepage_range() is called as the lock is already held | |
2782 | */ | |
83cde9e8 | 2783 | i_mmap_lock_write(mapping); |
6b2dbba8 | 2784 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
2785 | /* Do not unmap the current VMA */ |
2786 | if (iter_vma == vma) | |
2787 | continue; | |
2788 | ||
2789 | /* | |
2790 | * Unmap the page from other VMAs without their own reserves. | |
2791 | * They get marked to be SIGKILLed if they fault in these | |
2792 | * areas. This is because a future no-page fault on this VMA | |
2793 | * could insert a zeroed page instead of the data existing | |
2794 | * from the time of fork. This would look like data corruption | |
2795 | */ | |
2796 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
2797 | unmap_hugepage_range(iter_vma, address, |
2798 | address + huge_page_size(h), page); | |
04f2cbe3 | 2799 | } |
83cde9e8 | 2800 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
2801 | } |
2802 | ||
0fe6e20b NH |
2803 | /* |
2804 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
2805 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
2806 | * cannot race with other handlers or page migration. | |
2807 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 2808 | */ |
1e8f889b | 2809 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
04f2cbe3 | 2810 | unsigned long address, pte_t *ptep, pte_t pte, |
cb900f41 | 2811 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 2812 | { |
a5516438 | 2813 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 2814 | struct page *old_page, *new_page; |
ad4404a2 | 2815 | int ret = 0, outside_reserve = 0; |
2ec74c3e SG |
2816 | unsigned long mmun_start; /* For mmu_notifiers */ |
2817 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1e8f889b DG |
2818 | |
2819 | old_page = pte_page(pte); | |
2820 | ||
04f2cbe3 | 2821 | retry_avoidcopy: |
1e8f889b DG |
2822 | /* If no-one else is actually using this page, avoid the copy |
2823 | * and just make the page writable */ | |
37a2140d JK |
2824 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
2825 | page_move_anon_rmap(old_page, vma, address); | |
1e8f889b | 2826 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 2827 | return 0; |
1e8f889b DG |
2828 | } |
2829 | ||
04f2cbe3 MG |
2830 | /* |
2831 | * If the process that created a MAP_PRIVATE mapping is about to | |
2832 | * perform a COW due to a shared page count, attempt to satisfy | |
2833 | * the allocation without using the existing reserves. The pagecache | |
2834 | * page is used to determine if the reserve at this address was | |
2835 | * consumed or not. If reserves were used, a partial faulted mapping | |
2836 | * at the time of fork() could consume its reserves on COW instead | |
2837 | * of the full address range. | |
2838 | */ | |
5944d011 | 2839 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
2840 | old_page != pagecache_page) |
2841 | outside_reserve = 1; | |
2842 | ||
1e8f889b | 2843 | page_cache_get(old_page); |
b76c8cfb | 2844 | |
ad4404a2 DB |
2845 | /* |
2846 | * Drop page table lock as buddy allocator may be called. It will | |
2847 | * be acquired again before returning to the caller, as expected. | |
2848 | */ | |
cb900f41 | 2849 | spin_unlock(ptl); |
04f2cbe3 | 2850 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 2851 | |
2fc39cec | 2852 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
2853 | /* |
2854 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
2855 | * it is due to references held by a child and an insufficient | |
2856 | * huge page pool. To guarantee the original mappers | |
2857 | * reliability, unmap the page from child processes. The child | |
2858 | * may get SIGKILLed if it later faults. | |
2859 | */ | |
2860 | if (outside_reserve) { | |
ad4404a2 | 2861 | page_cache_release(old_page); |
04f2cbe3 | 2862 | BUG_ON(huge_pte_none(pte)); |
2f4612af DB |
2863 | unmap_ref_private(mm, vma, old_page, address); |
2864 | BUG_ON(huge_pte_none(pte)); | |
2865 | spin_lock(ptl); | |
2866 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); | |
2867 | if (likely(ptep && | |
2868 | pte_same(huge_ptep_get(ptep), pte))) | |
2869 | goto retry_avoidcopy; | |
2870 | /* | |
2871 | * race occurs while re-acquiring page table | |
2872 | * lock, and our job is done. | |
2873 | */ | |
2874 | return 0; | |
04f2cbe3 MG |
2875 | } |
2876 | ||
ad4404a2 DB |
2877 | ret = (PTR_ERR(new_page) == -ENOMEM) ? |
2878 | VM_FAULT_OOM : VM_FAULT_SIGBUS; | |
2879 | goto out_release_old; | |
1e8f889b DG |
2880 | } |
2881 | ||
0fe6e20b NH |
2882 | /* |
2883 | * When the original hugepage is shared one, it does not have | |
2884 | * anon_vma prepared. | |
2885 | */ | |
44e2aa93 | 2886 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
2887 | ret = VM_FAULT_OOM; |
2888 | goto out_release_all; | |
44e2aa93 | 2889 | } |
0fe6e20b | 2890 | |
47ad8475 AA |
2891 | copy_user_huge_page(new_page, old_page, address, vma, |
2892 | pages_per_huge_page(h)); | |
0ed361de | 2893 | __SetPageUptodate(new_page); |
1e8f889b | 2894 | |
2ec74c3e SG |
2895 | mmun_start = address & huge_page_mask(h); |
2896 | mmun_end = mmun_start + huge_page_size(h); | |
2897 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
ad4404a2 | 2898 | |
b76c8cfb | 2899 | /* |
cb900f41 | 2900 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
2901 | * before the page tables are altered |
2902 | */ | |
cb900f41 | 2903 | spin_lock(ptl); |
a5516438 | 2904 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
a9af0c5d | 2905 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
2906 | ClearPagePrivate(new_page); |
2907 | ||
1e8f889b | 2908 | /* Break COW */ |
8fe627ec | 2909 | huge_ptep_clear_flush(vma, address, ptep); |
34ee645e | 2910 | mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); |
1e8f889b DG |
2911 | set_huge_pte_at(mm, address, ptep, |
2912 | make_huge_pte(vma, new_page, 1)); | |
0fe6e20b | 2913 | page_remove_rmap(old_page); |
cd67f0d2 | 2914 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
2915 | /* Make the old page be freed below */ |
2916 | new_page = old_page; | |
2917 | } | |
cb900f41 | 2918 | spin_unlock(ptl); |
2ec74c3e | 2919 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
ad4404a2 | 2920 | out_release_all: |
1e8f889b | 2921 | page_cache_release(new_page); |
ad4404a2 | 2922 | out_release_old: |
1e8f889b | 2923 | page_cache_release(old_page); |
8312034f | 2924 | |
ad4404a2 DB |
2925 | spin_lock(ptl); /* Caller expects lock to be held */ |
2926 | return ret; | |
1e8f889b DG |
2927 | } |
2928 | ||
04f2cbe3 | 2929 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
2930 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
2931 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
2932 | { |
2933 | struct address_space *mapping; | |
e7c4b0bf | 2934 | pgoff_t idx; |
04f2cbe3 MG |
2935 | |
2936 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 2937 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
2938 | |
2939 | return find_lock_page(mapping, idx); | |
2940 | } | |
2941 | ||
3ae77f43 HD |
2942 | /* |
2943 | * Return whether there is a pagecache page to back given address within VMA. | |
2944 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
2945 | */ | |
2946 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
2947 | struct vm_area_struct *vma, unsigned long address) |
2948 | { | |
2949 | struct address_space *mapping; | |
2950 | pgoff_t idx; | |
2951 | struct page *page; | |
2952 | ||
2953 | mapping = vma->vm_file->f_mapping; | |
2954 | idx = vma_hugecache_offset(h, vma, address); | |
2955 | ||
2956 | page = find_get_page(mapping, idx); | |
2957 | if (page) | |
2958 | put_page(page); | |
2959 | return page != NULL; | |
2960 | } | |
2961 | ||
a1ed3dda | 2962 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
8382d914 DB |
2963 | struct address_space *mapping, pgoff_t idx, |
2964 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 2965 | { |
a5516438 | 2966 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 2967 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 2968 | int anon_rmap = 0; |
4c887265 | 2969 | unsigned long size; |
4c887265 | 2970 | struct page *page; |
1e8f889b | 2971 | pte_t new_pte; |
cb900f41 | 2972 | spinlock_t *ptl; |
4c887265 | 2973 | |
04f2cbe3 MG |
2974 | /* |
2975 | * Currently, we are forced to kill the process in the event the | |
2976 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 2977 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
2978 | */ |
2979 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
ffb22af5 AM |
2980 | pr_warning("PID %d killed due to inadequate hugepage pool\n", |
2981 | current->pid); | |
04f2cbe3 MG |
2982 | return ret; |
2983 | } | |
2984 | ||
4c887265 AL |
2985 | /* |
2986 | * Use page lock to guard against racing truncation | |
2987 | * before we get page_table_lock. | |
2988 | */ | |
6bda666a CL |
2989 | retry: |
2990 | page = find_lock_page(mapping, idx); | |
2991 | if (!page) { | |
a5516438 | 2992 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
2993 | if (idx >= size) |
2994 | goto out; | |
04f2cbe3 | 2995 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 2996 | if (IS_ERR(page)) { |
76dcee75 AK |
2997 | ret = PTR_ERR(page); |
2998 | if (ret == -ENOMEM) | |
2999 | ret = VM_FAULT_OOM; | |
3000 | else | |
3001 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
3002 | goto out; |
3003 | } | |
47ad8475 | 3004 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 3005 | __SetPageUptodate(page); |
ac9b9c66 | 3006 | |
f83a275d | 3007 | if (vma->vm_flags & VM_MAYSHARE) { |
6bda666a | 3008 | int err; |
45c682a6 | 3009 | struct inode *inode = mapping->host; |
6bda666a CL |
3010 | |
3011 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
3012 | if (err) { | |
3013 | put_page(page); | |
6bda666a CL |
3014 | if (err == -EEXIST) |
3015 | goto retry; | |
3016 | goto out; | |
3017 | } | |
07443a85 | 3018 | ClearPagePrivate(page); |
45c682a6 KC |
3019 | |
3020 | spin_lock(&inode->i_lock); | |
a5516438 | 3021 | inode->i_blocks += blocks_per_huge_page(h); |
45c682a6 | 3022 | spin_unlock(&inode->i_lock); |
23be7468 | 3023 | } else { |
6bda666a | 3024 | lock_page(page); |
0fe6e20b NH |
3025 | if (unlikely(anon_vma_prepare(vma))) { |
3026 | ret = VM_FAULT_OOM; | |
3027 | goto backout_unlocked; | |
3028 | } | |
409eb8c2 | 3029 | anon_rmap = 1; |
23be7468 | 3030 | } |
0fe6e20b | 3031 | } else { |
998b4382 NH |
3032 | /* |
3033 | * If memory error occurs between mmap() and fault, some process | |
3034 | * don't have hwpoisoned swap entry for errored virtual address. | |
3035 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
3036 | */ | |
3037 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 3038 | ret = VM_FAULT_HWPOISON | |
972dc4de | 3039 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
3040 | goto backout_unlocked; |
3041 | } | |
6bda666a | 3042 | } |
1e8f889b | 3043 | |
57303d80 AW |
3044 | /* |
3045 | * If we are going to COW a private mapping later, we examine the | |
3046 | * pending reservations for this page now. This will ensure that | |
3047 | * any allocations necessary to record that reservation occur outside | |
3048 | * the spinlock. | |
3049 | */ | |
788c7df4 | 3050 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) |
2b26736c AW |
3051 | if (vma_needs_reservation(h, vma, address) < 0) { |
3052 | ret = VM_FAULT_OOM; | |
3053 | goto backout_unlocked; | |
3054 | } | |
57303d80 | 3055 | |
cb900f41 KS |
3056 | ptl = huge_pte_lockptr(h, mm, ptep); |
3057 | spin_lock(ptl); | |
a5516438 | 3058 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
3059 | if (idx >= size) |
3060 | goto backout; | |
3061 | ||
83c54070 | 3062 | ret = 0; |
7f2e9525 | 3063 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
3064 | goto backout; |
3065 | ||
07443a85 JK |
3066 | if (anon_rmap) { |
3067 | ClearPagePrivate(page); | |
409eb8c2 | 3068 | hugepage_add_new_anon_rmap(page, vma, address); |
ac714904 | 3069 | } else |
409eb8c2 | 3070 | page_dup_rmap(page); |
1e8f889b DG |
3071 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
3072 | && (vma->vm_flags & VM_SHARED))); | |
3073 | set_huge_pte_at(mm, address, ptep, new_pte); | |
3074 | ||
788c7df4 | 3075 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 3076 | /* Optimization, do the COW without a second fault */ |
cb900f41 | 3077 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); |
1e8f889b DG |
3078 | } |
3079 | ||
cb900f41 | 3080 | spin_unlock(ptl); |
4c887265 AL |
3081 | unlock_page(page); |
3082 | out: | |
ac9b9c66 | 3083 | return ret; |
4c887265 AL |
3084 | |
3085 | backout: | |
cb900f41 | 3086 | spin_unlock(ptl); |
2b26736c | 3087 | backout_unlocked: |
4c887265 AL |
3088 | unlock_page(page); |
3089 | put_page(page); | |
3090 | goto out; | |
ac9b9c66 HD |
3091 | } |
3092 | ||
8382d914 DB |
3093 | #ifdef CONFIG_SMP |
3094 | static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm, | |
3095 | struct vm_area_struct *vma, | |
3096 | struct address_space *mapping, | |
3097 | pgoff_t idx, unsigned long address) | |
3098 | { | |
3099 | unsigned long key[2]; | |
3100 | u32 hash; | |
3101 | ||
3102 | if (vma->vm_flags & VM_SHARED) { | |
3103 | key[0] = (unsigned long) mapping; | |
3104 | key[1] = idx; | |
3105 | } else { | |
3106 | key[0] = (unsigned long) mm; | |
3107 | key[1] = address >> huge_page_shift(h); | |
3108 | } | |
3109 | ||
3110 | hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); | |
3111 | ||
3112 | return hash & (num_fault_mutexes - 1); | |
3113 | } | |
3114 | #else | |
3115 | /* | |
3116 | * For uniprocesor systems we always use a single mutex, so just | |
3117 | * return 0 and avoid the hashing overhead. | |
3118 | */ | |
3119 | static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm, | |
3120 | struct vm_area_struct *vma, | |
3121 | struct address_space *mapping, | |
3122 | pgoff_t idx, unsigned long address) | |
3123 | { | |
3124 | return 0; | |
3125 | } | |
3126 | #endif | |
3127 | ||
86e5216f | 3128 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 3129 | unsigned long address, unsigned int flags) |
86e5216f | 3130 | { |
8382d914 | 3131 | pte_t *ptep, entry; |
cb900f41 | 3132 | spinlock_t *ptl; |
1e8f889b | 3133 | int ret; |
8382d914 DB |
3134 | u32 hash; |
3135 | pgoff_t idx; | |
0fe6e20b | 3136 | struct page *page = NULL; |
57303d80 | 3137 | struct page *pagecache_page = NULL; |
a5516438 | 3138 | struct hstate *h = hstate_vma(vma); |
8382d914 | 3139 | struct address_space *mapping; |
0f792cf9 | 3140 | int need_wait_lock = 0; |
86e5216f | 3141 | |
1e16a539 KH |
3142 | address &= huge_page_mask(h); |
3143 | ||
fd6a03ed NH |
3144 | ptep = huge_pte_offset(mm, address); |
3145 | if (ptep) { | |
3146 | entry = huge_ptep_get(ptep); | |
290408d4 | 3147 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 3148 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
3149 | return 0; |
3150 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 3151 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 3152 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
3153 | } |
3154 | ||
a5516438 | 3155 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); |
86e5216f AL |
3156 | if (!ptep) |
3157 | return VM_FAULT_OOM; | |
3158 | ||
8382d914 DB |
3159 | mapping = vma->vm_file->f_mapping; |
3160 | idx = vma_hugecache_offset(h, vma, address); | |
3161 | ||
3935baa9 DG |
3162 | /* |
3163 | * Serialize hugepage allocation and instantiation, so that we don't | |
3164 | * get spurious allocation failures if two CPUs race to instantiate | |
3165 | * the same page in the page cache. | |
3166 | */ | |
8382d914 DB |
3167 | hash = fault_mutex_hash(h, mm, vma, mapping, idx, address); |
3168 | mutex_lock(&htlb_fault_mutex_table[hash]); | |
3169 | ||
7f2e9525 GS |
3170 | entry = huge_ptep_get(ptep); |
3171 | if (huge_pte_none(entry)) { | |
8382d914 | 3172 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 3173 | goto out_mutex; |
3935baa9 | 3174 | } |
86e5216f | 3175 | |
83c54070 | 3176 | ret = 0; |
1e8f889b | 3177 | |
0f792cf9 NH |
3178 | /* |
3179 | * entry could be a migration/hwpoison entry at this point, so this | |
3180 | * check prevents the kernel from going below assuming that we have | |
3181 | * a active hugepage in pagecache. This goto expects the 2nd page fault, | |
3182 | * and is_hugetlb_entry_(migration|hwpoisoned) check will properly | |
3183 | * handle it. | |
3184 | */ | |
3185 | if (!pte_present(entry)) | |
3186 | goto out_mutex; | |
3187 | ||
57303d80 AW |
3188 | /* |
3189 | * If we are going to COW the mapping later, we examine the pending | |
3190 | * reservations for this page now. This will ensure that any | |
3191 | * allocations necessary to record that reservation occur outside the | |
3192 | * spinlock. For private mappings, we also lookup the pagecache | |
3193 | * page now as it is used to determine if a reservation has been | |
3194 | * consumed. | |
3195 | */ | |
106c992a | 3196 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
2b26736c AW |
3197 | if (vma_needs_reservation(h, vma, address) < 0) { |
3198 | ret = VM_FAULT_OOM; | |
b4d1d99f | 3199 | goto out_mutex; |
2b26736c | 3200 | } |
57303d80 | 3201 | |
f83a275d | 3202 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
3203 | pagecache_page = hugetlbfs_pagecache_page(h, |
3204 | vma, address); | |
3205 | } | |
3206 | ||
0f792cf9 NH |
3207 | ptl = huge_pte_lock(h, mm, ptep); |
3208 | ||
3209 | /* Check for a racing update before calling hugetlb_cow */ | |
3210 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
3211 | goto out_ptl; | |
3212 | ||
56c9cfb1 NH |
3213 | /* |
3214 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
3215 | * pagecache_page, so here we need take the former one | |
3216 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
3217 | */ |
3218 | page = pte_page(entry); | |
3219 | if (page != pagecache_page) | |
0f792cf9 NH |
3220 | if (!trylock_page(page)) { |
3221 | need_wait_lock = 1; | |
3222 | goto out_ptl; | |
3223 | } | |
b4d1d99f | 3224 | |
0f792cf9 | 3225 | get_page(page); |
b4d1d99f | 3226 | |
788c7df4 | 3227 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 3228 | if (!huge_pte_write(entry)) { |
57303d80 | 3229 | ret = hugetlb_cow(mm, vma, address, ptep, entry, |
cb900f41 | 3230 | pagecache_page, ptl); |
0f792cf9 | 3231 | goto out_put_page; |
b4d1d99f | 3232 | } |
106c992a | 3233 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
3234 | } |
3235 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
3236 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
3237 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 3238 | update_mmu_cache(vma, address, ptep); |
0f792cf9 NH |
3239 | out_put_page: |
3240 | if (page != pagecache_page) | |
3241 | unlock_page(page); | |
3242 | put_page(page); | |
cb900f41 KS |
3243 | out_ptl: |
3244 | spin_unlock(ptl); | |
57303d80 AW |
3245 | |
3246 | if (pagecache_page) { | |
3247 | unlock_page(pagecache_page); | |
3248 | put_page(pagecache_page); | |
3249 | } | |
b4d1d99f | 3250 | out_mutex: |
8382d914 | 3251 | mutex_unlock(&htlb_fault_mutex_table[hash]); |
0f792cf9 NH |
3252 | /* |
3253 | * Generally it's safe to hold refcount during waiting page lock. But | |
3254 | * here we just wait to defer the next page fault to avoid busy loop and | |
3255 | * the page is not used after unlocked before returning from the current | |
3256 | * page fault. So we are safe from accessing freed page, even if we wait | |
3257 | * here without taking refcount. | |
3258 | */ | |
3259 | if (need_wait_lock) | |
3260 | wait_on_page_locked(page); | |
1e8f889b | 3261 | return ret; |
86e5216f AL |
3262 | } |
3263 | ||
28a35716 ML |
3264 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
3265 | struct page **pages, struct vm_area_struct **vmas, | |
3266 | unsigned long *position, unsigned long *nr_pages, | |
3267 | long i, unsigned int flags) | |
63551ae0 | 3268 | { |
d5d4b0aa CK |
3269 | unsigned long pfn_offset; |
3270 | unsigned long vaddr = *position; | |
28a35716 | 3271 | unsigned long remainder = *nr_pages; |
a5516438 | 3272 | struct hstate *h = hstate_vma(vma); |
63551ae0 | 3273 | |
63551ae0 | 3274 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 3275 | pte_t *pte; |
cb900f41 | 3276 | spinlock_t *ptl = NULL; |
2a15efc9 | 3277 | int absent; |
4c887265 | 3278 | struct page *page; |
63551ae0 | 3279 | |
4c887265 AL |
3280 | /* |
3281 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 3282 | * each hugepage. We have to make sure we get the |
4c887265 | 3283 | * first, for the page indexing below to work. |
cb900f41 KS |
3284 | * |
3285 | * Note that page table lock is not held when pte is null. | |
4c887265 | 3286 | */ |
a5516438 | 3287 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); |
cb900f41 KS |
3288 | if (pte) |
3289 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
3290 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
3291 | ||
3292 | /* | |
3293 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
3294 | * an error where there's an empty slot with no huge pagecache |
3295 | * to back it. This way, we avoid allocating a hugepage, and | |
3296 | * the sparse dumpfile avoids allocating disk blocks, but its | |
3297 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 3298 | */ |
3ae77f43 HD |
3299 | if (absent && (flags & FOLL_DUMP) && |
3300 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
3301 | if (pte) |
3302 | spin_unlock(ptl); | |
2a15efc9 HD |
3303 | remainder = 0; |
3304 | break; | |
3305 | } | |
63551ae0 | 3306 | |
9cc3a5bd NH |
3307 | /* |
3308 | * We need call hugetlb_fault for both hugepages under migration | |
3309 | * (in which case hugetlb_fault waits for the migration,) and | |
3310 | * hwpoisoned hugepages (in which case we need to prevent the | |
3311 | * caller from accessing to them.) In order to do this, we use | |
3312 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
3313 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
3314 | * both cases, and because we can't follow correct pages | |
3315 | * directly from any kind of swap entries. | |
3316 | */ | |
3317 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
3318 | ((flags & FOLL_WRITE) && |
3319 | !huge_pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 3320 | int ret; |
63551ae0 | 3321 | |
cb900f41 KS |
3322 | if (pte) |
3323 | spin_unlock(ptl); | |
2a15efc9 HD |
3324 | ret = hugetlb_fault(mm, vma, vaddr, |
3325 | (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); | |
a89182c7 | 3326 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 3327 | continue; |
63551ae0 | 3328 | |
4c887265 | 3329 | remainder = 0; |
4c887265 AL |
3330 | break; |
3331 | } | |
3332 | ||
a5516438 | 3333 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 3334 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 3335 | same_page: |
d6692183 | 3336 | if (pages) { |
2a15efc9 | 3337 | pages[i] = mem_map_offset(page, pfn_offset); |
a0368d4e | 3338 | get_page_foll(pages[i]); |
d6692183 | 3339 | } |
63551ae0 DG |
3340 | |
3341 | if (vmas) | |
3342 | vmas[i] = vma; | |
3343 | ||
3344 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 3345 | ++pfn_offset; |
63551ae0 DG |
3346 | --remainder; |
3347 | ++i; | |
d5d4b0aa | 3348 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 3349 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa CK |
3350 | /* |
3351 | * We use pfn_offset to avoid touching the pageframes | |
3352 | * of this compound page. | |
3353 | */ | |
3354 | goto same_page; | |
3355 | } | |
cb900f41 | 3356 | spin_unlock(ptl); |
63551ae0 | 3357 | } |
28a35716 | 3358 | *nr_pages = remainder; |
63551ae0 DG |
3359 | *position = vaddr; |
3360 | ||
2a15efc9 | 3361 | return i ? i : -EFAULT; |
63551ae0 | 3362 | } |
8f860591 | 3363 | |
7da4d641 | 3364 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
3365 | unsigned long address, unsigned long end, pgprot_t newprot) |
3366 | { | |
3367 | struct mm_struct *mm = vma->vm_mm; | |
3368 | unsigned long start = address; | |
3369 | pte_t *ptep; | |
3370 | pte_t pte; | |
a5516438 | 3371 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 3372 | unsigned long pages = 0; |
8f860591 ZY |
3373 | |
3374 | BUG_ON(address >= end); | |
3375 | flush_cache_range(vma, address, end); | |
3376 | ||
a5338093 | 3377 | mmu_notifier_invalidate_range_start(mm, start, end); |
83cde9e8 | 3378 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 3379 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 3380 | spinlock_t *ptl; |
8f860591 ZY |
3381 | ptep = huge_pte_offset(mm, address); |
3382 | if (!ptep) | |
3383 | continue; | |
cb900f41 | 3384 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
3385 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3386 | pages++; | |
cb900f41 | 3387 | spin_unlock(ptl); |
39dde65c | 3388 | continue; |
7da4d641 | 3389 | } |
a8bda28d NH |
3390 | pte = huge_ptep_get(ptep); |
3391 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
3392 | spin_unlock(ptl); | |
3393 | continue; | |
3394 | } | |
3395 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
3396 | swp_entry_t entry = pte_to_swp_entry(pte); | |
3397 | ||
3398 | if (is_write_migration_entry(entry)) { | |
3399 | pte_t newpte; | |
3400 | ||
3401 | make_migration_entry_read(&entry); | |
3402 | newpte = swp_entry_to_pte(entry); | |
3403 | set_huge_pte_at(mm, address, ptep, newpte); | |
3404 | pages++; | |
3405 | } | |
3406 | spin_unlock(ptl); | |
3407 | continue; | |
3408 | } | |
3409 | if (!huge_pte_none(pte)) { | |
8f860591 | 3410 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
106c992a | 3411 | pte = pte_mkhuge(huge_pte_modify(pte, newprot)); |
be7517d6 | 3412 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
8f860591 | 3413 | set_huge_pte_at(mm, address, ptep, pte); |
7da4d641 | 3414 | pages++; |
8f860591 | 3415 | } |
cb900f41 | 3416 | spin_unlock(ptl); |
8f860591 | 3417 | } |
d833352a | 3418 | /* |
c8c06efa | 3419 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 3420 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 3421 | * once we release i_mmap_rwsem, another task can do the final put_page |
d833352a MG |
3422 | * and that page table be reused and filled with junk. |
3423 | */ | |
8f860591 | 3424 | flush_tlb_range(vma, start, end); |
34ee645e | 3425 | mmu_notifier_invalidate_range(mm, start, end); |
83cde9e8 | 3426 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
a5338093 | 3427 | mmu_notifier_invalidate_range_end(mm, start, end); |
7da4d641 PZ |
3428 | |
3429 | return pages << h->order; | |
8f860591 ZY |
3430 | } |
3431 | ||
a1e78772 MG |
3432 | int hugetlb_reserve_pages(struct inode *inode, |
3433 | long from, long to, | |
5a6fe125 | 3434 | struct vm_area_struct *vma, |
ca16d140 | 3435 | vm_flags_t vm_flags) |
e4e574b7 | 3436 | { |
17c9d12e | 3437 | long ret, chg; |
a5516438 | 3438 | struct hstate *h = hstate_inode(inode); |
90481622 | 3439 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 3440 | struct resv_map *resv_map; |
e4e574b7 | 3441 | |
17c9d12e MG |
3442 | /* |
3443 | * Only apply hugepage reservation if asked. At fault time, an | |
3444 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 3445 | * without using reserves |
17c9d12e | 3446 | */ |
ca16d140 | 3447 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
3448 | return 0; |
3449 | ||
a1e78772 MG |
3450 | /* |
3451 | * Shared mappings base their reservation on the number of pages that | |
3452 | * are already allocated on behalf of the file. Private mappings need | |
3453 | * to reserve the full area even if read-only as mprotect() may be | |
3454 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
3455 | */ | |
9119a41e | 3456 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4e35f483 | 3457 | resv_map = inode_resv_map(inode); |
9119a41e | 3458 | |
1406ec9b | 3459 | chg = region_chg(resv_map, from, to); |
9119a41e JK |
3460 | |
3461 | } else { | |
3462 | resv_map = resv_map_alloc(); | |
17c9d12e MG |
3463 | if (!resv_map) |
3464 | return -ENOMEM; | |
3465 | ||
a1e78772 | 3466 | chg = to - from; |
84afd99b | 3467 | |
17c9d12e MG |
3468 | set_vma_resv_map(vma, resv_map); |
3469 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
3470 | } | |
3471 | ||
c50ac050 DH |
3472 | if (chg < 0) { |
3473 | ret = chg; | |
3474 | goto out_err; | |
3475 | } | |
8a630112 | 3476 | |
90481622 | 3477 | /* There must be enough pages in the subpool for the mapping */ |
c50ac050 DH |
3478 | if (hugepage_subpool_get_pages(spool, chg)) { |
3479 | ret = -ENOSPC; | |
3480 | goto out_err; | |
3481 | } | |
5a6fe125 MG |
3482 | |
3483 | /* | |
17c9d12e | 3484 | * Check enough hugepages are available for the reservation. |
90481622 | 3485 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 3486 | */ |
a5516438 | 3487 | ret = hugetlb_acct_memory(h, chg); |
68842c9b | 3488 | if (ret < 0) { |
90481622 | 3489 | hugepage_subpool_put_pages(spool, chg); |
c50ac050 | 3490 | goto out_err; |
68842c9b | 3491 | } |
17c9d12e MG |
3492 | |
3493 | /* | |
3494 | * Account for the reservations made. Shared mappings record regions | |
3495 | * that have reservations as they are shared by multiple VMAs. | |
3496 | * When the last VMA disappears, the region map says how much | |
3497 | * the reservation was and the page cache tells how much of | |
3498 | * the reservation was consumed. Private mappings are per-VMA and | |
3499 | * only the consumed reservations are tracked. When the VMA | |
3500 | * disappears, the original reservation is the VMA size and the | |
3501 | * consumed reservations are stored in the map. Hence, nothing | |
3502 | * else has to be done for private mappings here | |
3503 | */ | |
f83a275d | 3504 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
1406ec9b | 3505 | region_add(resv_map, from, to); |
a43a8c39 | 3506 | return 0; |
c50ac050 | 3507 | out_err: |
f031dd27 JK |
3508 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3509 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 3510 | return ret; |
a43a8c39 CK |
3511 | } |
3512 | ||
3513 | void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) | |
3514 | { | |
a5516438 | 3515 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 3516 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 3517 | long chg = 0; |
90481622 | 3518 | struct hugepage_subpool *spool = subpool_inode(inode); |
45c682a6 | 3519 | |
9119a41e | 3520 | if (resv_map) |
1406ec9b | 3521 | chg = region_truncate(resv_map, offset); |
45c682a6 | 3522 | spin_lock(&inode->i_lock); |
e4c6f8be | 3523 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
3524 | spin_unlock(&inode->i_lock); |
3525 | ||
90481622 | 3526 | hugepage_subpool_put_pages(spool, (chg - freed)); |
a5516438 | 3527 | hugetlb_acct_memory(h, -(chg - freed)); |
a43a8c39 | 3528 | } |
93f70f90 | 3529 | |
3212b535 SC |
3530 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
3531 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
3532 | struct vm_area_struct *vma, | |
3533 | unsigned long addr, pgoff_t idx) | |
3534 | { | |
3535 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
3536 | svma->vm_start; | |
3537 | unsigned long sbase = saddr & PUD_MASK; | |
3538 | unsigned long s_end = sbase + PUD_SIZE; | |
3539 | ||
3540 | /* Allow segments to share if only one is marked locked */ | |
3541 | unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED; | |
3542 | unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED; | |
3543 | ||
3544 | /* | |
3545 | * match the virtual addresses, permission and the alignment of the | |
3546 | * page table page. | |
3547 | */ | |
3548 | if (pmd_index(addr) != pmd_index(saddr) || | |
3549 | vm_flags != svm_flags || | |
3550 | sbase < svma->vm_start || svma->vm_end < s_end) | |
3551 | return 0; | |
3552 | ||
3553 | return saddr; | |
3554 | } | |
3555 | ||
3556 | static int vma_shareable(struct vm_area_struct *vma, unsigned long addr) | |
3557 | { | |
3558 | unsigned long base = addr & PUD_MASK; | |
3559 | unsigned long end = base + PUD_SIZE; | |
3560 | ||
3561 | /* | |
3562 | * check on proper vm_flags and page table alignment | |
3563 | */ | |
3564 | if (vma->vm_flags & VM_MAYSHARE && | |
3565 | vma->vm_start <= base && end <= vma->vm_end) | |
3566 | return 1; | |
3567 | return 0; | |
3568 | } | |
3569 | ||
3570 | /* | |
3571 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
3572 | * and returns the corresponding pte. While this is not necessary for the | |
3573 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
3574 | * code much cleaner. pmd allocation is essential for the shared case because | |
c8c06efa | 3575 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise |
3212b535 SC |
3576 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a |
3577 | * bad pmd for sharing. | |
3578 | */ | |
3579 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
3580 | { | |
3581 | struct vm_area_struct *vma = find_vma(mm, addr); | |
3582 | struct address_space *mapping = vma->vm_file->f_mapping; | |
3583 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
3584 | vma->vm_pgoff; | |
3585 | struct vm_area_struct *svma; | |
3586 | unsigned long saddr; | |
3587 | pte_t *spte = NULL; | |
3588 | pte_t *pte; | |
cb900f41 | 3589 | spinlock_t *ptl; |
3212b535 SC |
3590 | |
3591 | if (!vma_shareable(vma, addr)) | |
3592 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
3593 | ||
83cde9e8 | 3594 | i_mmap_lock_write(mapping); |
3212b535 SC |
3595 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
3596 | if (svma == vma) | |
3597 | continue; | |
3598 | ||
3599 | saddr = page_table_shareable(svma, vma, addr, idx); | |
3600 | if (saddr) { | |
3601 | spte = huge_pte_offset(svma->vm_mm, saddr); | |
3602 | if (spte) { | |
dc6c9a35 | 3603 | mm_inc_nr_pmds(mm); |
3212b535 SC |
3604 | get_page(virt_to_page(spte)); |
3605 | break; | |
3606 | } | |
3607 | } | |
3608 | } | |
3609 | ||
3610 | if (!spte) | |
3611 | goto out; | |
3612 | ||
cb900f41 KS |
3613 | ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); |
3614 | spin_lock(ptl); | |
dc6c9a35 | 3615 | if (pud_none(*pud)) { |
3212b535 SC |
3616 | pud_populate(mm, pud, |
3617 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
dc6c9a35 | 3618 | } else { |
3212b535 | 3619 | put_page(virt_to_page(spte)); |
dc6c9a35 KS |
3620 | mm_inc_nr_pmds(mm); |
3621 | } | |
cb900f41 | 3622 | spin_unlock(ptl); |
3212b535 SC |
3623 | out: |
3624 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
83cde9e8 | 3625 | i_mmap_unlock_write(mapping); |
3212b535 SC |
3626 | return pte; |
3627 | } | |
3628 | ||
3629 | /* | |
3630 | * unmap huge page backed by shared pte. | |
3631 | * | |
3632 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
3633 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
3634 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
3635 | * | |
cb900f41 | 3636 | * called with page table lock held. |
3212b535 SC |
3637 | * |
3638 | * returns: 1 successfully unmapped a shared pte page | |
3639 | * 0 the underlying pte page is not shared, or it is the last user | |
3640 | */ | |
3641 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
3642 | { | |
3643 | pgd_t *pgd = pgd_offset(mm, *addr); | |
3644 | pud_t *pud = pud_offset(pgd, *addr); | |
3645 | ||
3646 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
3647 | if (page_count(virt_to_page(ptep)) == 1) | |
3648 | return 0; | |
3649 | ||
3650 | pud_clear(pud); | |
3651 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 3652 | mm_dec_nr_pmds(mm); |
3212b535 SC |
3653 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
3654 | return 1; | |
3655 | } | |
9e5fc74c SC |
3656 | #define want_pmd_share() (1) |
3657 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
3658 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
3659 | { | |
3660 | return NULL; | |
3661 | } | |
3662 | #define want_pmd_share() (0) | |
3212b535 SC |
3663 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
3664 | ||
9e5fc74c SC |
3665 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
3666 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
3667 | unsigned long addr, unsigned long sz) | |
3668 | { | |
3669 | pgd_t *pgd; | |
3670 | pud_t *pud; | |
3671 | pte_t *pte = NULL; | |
3672 | ||
3673 | pgd = pgd_offset(mm, addr); | |
3674 | pud = pud_alloc(mm, pgd, addr); | |
3675 | if (pud) { | |
3676 | if (sz == PUD_SIZE) { | |
3677 | pte = (pte_t *)pud; | |
3678 | } else { | |
3679 | BUG_ON(sz != PMD_SIZE); | |
3680 | if (want_pmd_share() && pud_none(*pud)) | |
3681 | pte = huge_pmd_share(mm, addr, pud); | |
3682 | else | |
3683 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3684 | } | |
3685 | } | |
3686 | BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); | |
3687 | ||
3688 | return pte; | |
3689 | } | |
3690 | ||
3691 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) | |
3692 | { | |
3693 | pgd_t *pgd; | |
3694 | pud_t *pud; | |
3695 | pmd_t *pmd = NULL; | |
3696 | ||
3697 | pgd = pgd_offset(mm, addr); | |
3698 | if (pgd_present(*pgd)) { | |
3699 | pud = pud_offset(pgd, addr); | |
3700 | if (pud_present(*pud)) { | |
3701 | if (pud_huge(*pud)) | |
3702 | return (pte_t *)pud; | |
3703 | pmd = pmd_offset(pud, addr); | |
3704 | } | |
3705 | } | |
3706 | return (pte_t *) pmd; | |
3707 | } | |
3708 | ||
61f77eda NH |
3709 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
3710 | ||
3711 | /* | |
3712 | * These functions are overwritable if your architecture needs its own | |
3713 | * behavior. | |
3714 | */ | |
3715 | struct page * __weak | |
3716 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
3717 | int write) | |
3718 | { | |
3719 | return ERR_PTR(-EINVAL); | |
3720 | } | |
3721 | ||
3722 | struct page * __weak | |
9e5fc74c | 3723 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 3724 | pmd_t *pmd, int flags) |
9e5fc74c | 3725 | { |
e66f17ff NH |
3726 | struct page *page = NULL; |
3727 | spinlock_t *ptl; | |
3728 | retry: | |
3729 | ptl = pmd_lockptr(mm, pmd); | |
3730 | spin_lock(ptl); | |
3731 | /* | |
3732 | * make sure that the address range covered by this pmd is not | |
3733 | * unmapped from other threads. | |
3734 | */ | |
3735 | if (!pmd_huge(*pmd)) | |
3736 | goto out; | |
3737 | if (pmd_present(*pmd)) { | |
3738 | page = pte_page(*(pte_t *)pmd) + | |
3739 | ((address & ~PMD_MASK) >> PAGE_SHIFT); | |
3740 | if (flags & FOLL_GET) | |
3741 | get_page(page); | |
3742 | } else { | |
3743 | if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) { | |
3744 | spin_unlock(ptl); | |
3745 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
3746 | goto retry; | |
3747 | } | |
3748 | /* | |
3749 | * hwpoisoned entry is treated as no_page_table in | |
3750 | * follow_page_mask(). | |
3751 | */ | |
3752 | } | |
3753 | out: | |
3754 | spin_unlock(ptl); | |
9e5fc74c SC |
3755 | return page; |
3756 | } | |
3757 | ||
61f77eda | 3758 | struct page * __weak |
9e5fc74c | 3759 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 3760 | pud_t *pud, int flags) |
9e5fc74c | 3761 | { |
e66f17ff NH |
3762 | if (flags & FOLL_GET) |
3763 | return NULL; | |
9e5fc74c | 3764 | |
e66f17ff | 3765 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
3766 | } |
3767 | ||
d5bd9106 AK |
3768 | #ifdef CONFIG_MEMORY_FAILURE |
3769 | ||
6de2b1aa NH |
3770 | /* Should be called in hugetlb_lock */ |
3771 | static int is_hugepage_on_freelist(struct page *hpage) | |
3772 | { | |
3773 | struct page *page; | |
3774 | struct page *tmp; | |
3775 | struct hstate *h = page_hstate(hpage); | |
3776 | int nid = page_to_nid(hpage); | |
3777 | ||
3778 | list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru) | |
3779 | if (page == hpage) | |
3780 | return 1; | |
3781 | return 0; | |
3782 | } | |
3783 | ||
93f70f90 NH |
3784 | /* |
3785 | * This function is called from memory failure code. | |
3786 | * Assume the caller holds page lock of the head page. | |
3787 | */ | |
6de2b1aa | 3788 | int dequeue_hwpoisoned_huge_page(struct page *hpage) |
93f70f90 NH |
3789 | { |
3790 | struct hstate *h = page_hstate(hpage); | |
3791 | int nid = page_to_nid(hpage); | |
6de2b1aa | 3792 | int ret = -EBUSY; |
93f70f90 NH |
3793 | |
3794 | spin_lock(&hugetlb_lock); | |
6de2b1aa | 3795 | if (is_hugepage_on_freelist(hpage)) { |
56f2fb14 NH |
3796 | /* |
3797 | * Hwpoisoned hugepage isn't linked to activelist or freelist, | |
3798 | * but dangling hpage->lru can trigger list-debug warnings | |
3799 | * (this happens when we call unpoison_memory() on it), | |
3800 | * so let it point to itself with list_del_init(). | |
3801 | */ | |
3802 | list_del_init(&hpage->lru); | |
8c6c2ecb | 3803 | set_page_refcounted(hpage); |
6de2b1aa NH |
3804 | h->free_huge_pages--; |
3805 | h->free_huge_pages_node[nid]--; | |
3806 | ret = 0; | |
3807 | } | |
93f70f90 | 3808 | spin_unlock(&hugetlb_lock); |
6de2b1aa | 3809 | return ret; |
93f70f90 | 3810 | } |
6de2b1aa | 3811 | #endif |
31caf665 NH |
3812 | |
3813 | bool isolate_huge_page(struct page *page, struct list_head *list) | |
3814 | { | |
309381fe | 3815 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 NH |
3816 | if (!get_page_unless_zero(page)) |
3817 | return false; | |
3818 | spin_lock(&hugetlb_lock); | |
3819 | list_move_tail(&page->lru, list); | |
3820 | spin_unlock(&hugetlb_lock); | |
3821 | return true; | |
3822 | } | |
3823 | ||
3824 | void putback_active_hugepage(struct page *page) | |
3825 | { | |
309381fe | 3826 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 NH |
3827 | spin_lock(&hugetlb_lock); |
3828 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); | |
3829 | spin_unlock(&hugetlb_lock); | |
3830 | put_page(page); | |
3831 | } | |
c8721bbb NH |
3832 | |
3833 | bool is_hugepage_active(struct page *page) | |
3834 | { | |
309381fe | 3835 | VM_BUG_ON_PAGE(!PageHuge(page), page); |
c8721bbb NH |
3836 | /* |
3837 | * This function can be called for a tail page because the caller, | |
3838 | * scan_movable_pages, scans through a given pfn-range which typically | |
3839 | * covers one memory block. In systems using gigantic hugepage (1GB | |
3840 | * for x86_64,) a hugepage is larger than a memory block, and we don't | |
3841 | * support migrating such large hugepages for now, so return false | |
3842 | * when called for tail pages. | |
3843 | */ | |
3844 | if (PageTail(page)) | |
3845 | return false; | |
3846 | /* | |
3847 | * Refcount of a hwpoisoned hugepages is 1, but they are not active, | |
3848 | * so we should return false for them. | |
3849 | */ | |
3850 | if (unlikely(PageHWPoison(page))) | |
3851 | return false; | |
3852 | return page_count(page) > 0; | |
3853 | } |