]>
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]; | |
641844f5 NH |
43 | /* |
44 | * Minimum page order among possible hugepage sizes, set to a proper value | |
45 | * at boot time. | |
46 | */ | |
47 | static unsigned int minimum_order __read_mostly = UINT_MAX; | |
e5ff2159 | 48 | |
53ba51d2 JT |
49 | __initdata LIST_HEAD(huge_boot_pages); |
50 | ||
e5ff2159 AK |
51 | /* for command line parsing */ |
52 | static struct hstate * __initdata parsed_hstate; | |
53 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 54 | static unsigned long __initdata default_hstate_size; |
e5ff2159 | 55 | |
3935baa9 | 56 | /* |
31caf665 NH |
57 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
58 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 59 | */ |
c3f38a38 | 60 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 61 | |
8382d914 DB |
62 | /* |
63 | * Serializes faults on the same logical page. This is used to | |
64 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
65 | */ | |
66 | static int num_fault_mutexes; | |
67 | static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp; | |
68 | ||
7ca02d0a MK |
69 | /* Forward declaration */ |
70 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
71 | ||
90481622 DG |
72 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
73 | { | |
74 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
75 | ||
76 | spin_unlock(&spool->lock); | |
77 | ||
78 | /* If no pages are used, and no other handles to the subpool | |
7ca02d0a MK |
79 | * remain, give up any reservations mased on minimum size and |
80 | * free the subpool */ | |
81 | if (free) { | |
82 | if (spool->min_hpages != -1) | |
83 | hugetlb_acct_memory(spool->hstate, | |
84 | -spool->min_hpages); | |
90481622 | 85 | kfree(spool); |
7ca02d0a | 86 | } |
90481622 DG |
87 | } |
88 | ||
7ca02d0a MK |
89 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
90 | long min_hpages) | |
90481622 DG |
91 | { |
92 | struct hugepage_subpool *spool; | |
93 | ||
c6a91820 | 94 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
95 | if (!spool) |
96 | return NULL; | |
97 | ||
98 | spin_lock_init(&spool->lock); | |
99 | spool->count = 1; | |
7ca02d0a MK |
100 | spool->max_hpages = max_hpages; |
101 | spool->hstate = h; | |
102 | spool->min_hpages = min_hpages; | |
103 | ||
104 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
105 | kfree(spool); | |
106 | return NULL; | |
107 | } | |
108 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
109 | |
110 | return spool; | |
111 | } | |
112 | ||
113 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
114 | { | |
115 | spin_lock(&spool->lock); | |
116 | BUG_ON(!spool->count); | |
117 | spool->count--; | |
118 | unlock_or_release_subpool(spool); | |
119 | } | |
120 | ||
1c5ecae3 MK |
121 | /* |
122 | * Subpool accounting for allocating and reserving pages. | |
123 | * Return -ENOMEM if there are not enough resources to satisfy the | |
124 | * the request. Otherwise, return the number of pages by which the | |
125 | * global pools must be adjusted (upward). The returned value may | |
126 | * only be different than the passed value (delta) in the case where | |
127 | * a subpool minimum size must be manitained. | |
128 | */ | |
129 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
130 | long delta) |
131 | { | |
1c5ecae3 | 132 | long ret = delta; |
90481622 DG |
133 | |
134 | if (!spool) | |
1c5ecae3 | 135 | return ret; |
90481622 DG |
136 | |
137 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
138 | |
139 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
140 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
141 | spool->used_hpages += delta; | |
142 | else { | |
143 | ret = -ENOMEM; | |
144 | goto unlock_ret; | |
145 | } | |
90481622 | 146 | } |
90481622 | 147 | |
1c5ecae3 MK |
148 | if (spool->min_hpages != -1) { /* minimum size accounting */ |
149 | if (delta > spool->rsv_hpages) { | |
150 | /* | |
151 | * Asking for more reserves than those already taken on | |
152 | * behalf of subpool. Return difference. | |
153 | */ | |
154 | ret = delta - spool->rsv_hpages; | |
155 | spool->rsv_hpages = 0; | |
156 | } else { | |
157 | ret = 0; /* reserves already accounted for */ | |
158 | spool->rsv_hpages -= delta; | |
159 | } | |
160 | } | |
161 | ||
162 | unlock_ret: | |
163 | spin_unlock(&spool->lock); | |
90481622 DG |
164 | return ret; |
165 | } | |
166 | ||
1c5ecae3 MK |
167 | /* |
168 | * Subpool accounting for freeing and unreserving pages. | |
169 | * Return the number of global page reservations that must be dropped. | |
170 | * The return value may only be different than the passed value (delta) | |
171 | * in the case where a subpool minimum size must be maintained. | |
172 | */ | |
173 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
174 | long delta) |
175 | { | |
1c5ecae3 MK |
176 | long ret = delta; |
177 | ||
90481622 | 178 | if (!spool) |
1c5ecae3 | 179 | return delta; |
90481622 DG |
180 | |
181 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
182 | |
183 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
184 | spool->used_hpages -= delta; | |
185 | ||
186 | if (spool->min_hpages != -1) { /* minimum size accounting */ | |
187 | if (spool->rsv_hpages + delta <= spool->min_hpages) | |
188 | ret = 0; | |
189 | else | |
190 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
191 | ||
192 | spool->rsv_hpages += delta; | |
193 | if (spool->rsv_hpages > spool->min_hpages) | |
194 | spool->rsv_hpages = spool->min_hpages; | |
195 | } | |
196 | ||
197 | /* | |
198 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
199 | * quota reference, free it now. | |
200 | */ | |
90481622 | 201 | unlock_or_release_subpool(spool); |
1c5ecae3 MK |
202 | |
203 | return ret; | |
90481622 DG |
204 | } |
205 | ||
206 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
207 | { | |
208 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
209 | } | |
210 | ||
211 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
212 | { | |
496ad9aa | 213 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
214 | } |
215 | ||
96822904 AW |
216 | /* |
217 | * Region tracking -- allows tracking of reservations and instantiated pages | |
218 | * across the pages in a mapping. | |
84afd99b | 219 | * |
7b24d861 DB |
220 | * The region data structures are embedded into a resv_map and |
221 | * protected by a resv_map's lock | |
96822904 AW |
222 | */ |
223 | struct file_region { | |
224 | struct list_head link; | |
225 | long from; | |
226 | long to; | |
227 | }; | |
228 | ||
1406ec9b | 229 | static long region_add(struct resv_map *resv, long f, long t) |
96822904 | 230 | { |
1406ec9b | 231 | struct list_head *head = &resv->regions; |
96822904 AW |
232 | struct file_region *rg, *nrg, *trg; |
233 | ||
7b24d861 | 234 | spin_lock(&resv->lock); |
96822904 AW |
235 | /* Locate the region we are either in or before. */ |
236 | list_for_each_entry(rg, head, link) | |
237 | if (f <= rg->to) | |
238 | break; | |
239 | ||
240 | /* Round our left edge to the current segment if it encloses us. */ | |
241 | if (f > rg->from) | |
242 | f = rg->from; | |
243 | ||
244 | /* Check for and consume any regions we now overlap with. */ | |
245 | nrg = rg; | |
246 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
247 | if (&rg->link == head) | |
248 | break; | |
249 | if (rg->from > t) | |
250 | break; | |
251 | ||
252 | /* If this area reaches higher then extend our area to | |
253 | * include it completely. If this is not the first area | |
254 | * which we intend to reuse, free it. */ | |
255 | if (rg->to > t) | |
256 | t = rg->to; | |
257 | if (rg != nrg) { | |
258 | list_del(&rg->link); | |
259 | kfree(rg); | |
260 | } | |
261 | } | |
262 | nrg->from = f; | |
263 | nrg->to = t; | |
7b24d861 | 264 | spin_unlock(&resv->lock); |
96822904 AW |
265 | return 0; |
266 | } | |
267 | ||
1406ec9b | 268 | static long region_chg(struct resv_map *resv, long f, long t) |
96822904 | 269 | { |
1406ec9b | 270 | struct list_head *head = &resv->regions; |
7b24d861 | 271 | struct file_region *rg, *nrg = NULL; |
96822904 AW |
272 | long chg = 0; |
273 | ||
7b24d861 DB |
274 | retry: |
275 | spin_lock(&resv->lock); | |
96822904 AW |
276 | /* Locate the region we are before or in. */ |
277 | list_for_each_entry(rg, head, link) | |
278 | if (f <= rg->to) | |
279 | break; | |
280 | ||
281 | /* If we are below the current region then a new region is required. | |
282 | * Subtle, allocate a new region at the position but make it zero | |
283 | * size such that we can guarantee to record the reservation. */ | |
284 | if (&rg->link == head || t < rg->from) { | |
7b24d861 DB |
285 | if (!nrg) { |
286 | spin_unlock(&resv->lock); | |
287 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
288 | if (!nrg) | |
289 | return -ENOMEM; | |
290 | ||
291 | nrg->from = f; | |
292 | nrg->to = f; | |
293 | INIT_LIST_HEAD(&nrg->link); | |
294 | goto retry; | |
295 | } | |
96822904 | 296 | |
7b24d861 DB |
297 | list_add(&nrg->link, rg->link.prev); |
298 | chg = t - f; | |
299 | goto out_nrg; | |
96822904 AW |
300 | } |
301 | ||
302 | /* Round our left edge to the current segment if it encloses us. */ | |
303 | if (f > rg->from) | |
304 | f = rg->from; | |
305 | chg = t - f; | |
306 | ||
307 | /* Check for and consume any regions we now overlap with. */ | |
308 | list_for_each_entry(rg, rg->link.prev, link) { | |
309 | if (&rg->link == head) | |
310 | break; | |
311 | if (rg->from > t) | |
7b24d861 | 312 | goto out; |
96822904 | 313 | |
25985edc | 314 | /* We overlap with this area, if it extends further than |
96822904 AW |
315 | * us then we must extend ourselves. Account for its |
316 | * existing reservation. */ | |
317 | if (rg->to > t) { | |
318 | chg += rg->to - t; | |
319 | t = rg->to; | |
320 | } | |
321 | chg -= rg->to - rg->from; | |
322 | } | |
7b24d861 DB |
323 | |
324 | out: | |
325 | spin_unlock(&resv->lock); | |
326 | /* We already know we raced and no longer need the new region */ | |
327 | kfree(nrg); | |
328 | return chg; | |
329 | out_nrg: | |
330 | spin_unlock(&resv->lock); | |
96822904 AW |
331 | return chg; |
332 | } | |
333 | ||
1406ec9b | 334 | static long region_truncate(struct resv_map *resv, long end) |
96822904 | 335 | { |
1406ec9b | 336 | struct list_head *head = &resv->regions; |
96822904 AW |
337 | struct file_region *rg, *trg; |
338 | long chg = 0; | |
339 | ||
7b24d861 | 340 | spin_lock(&resv->lock); |
96822904 AW |
341 | /* Locate the region we are either in or before. */ |
342 | list_for_each_entry(rg, head, link) | |
343 | if (end <= rg->to) | |
344 | break; | |
345 | if (&rg->link == head) | |
7b24d861 | 346 | goto out; |
96822904 AW |
347 | |
348 | /* If we are in the middle of a region then adjust it. */ | |
349 | if (end > rg->from) { | |
350 | chg = rg->to - end; | |
351 | rg->to = end; | |
352 | rg = list_entry(rg->link.next, typeof(*rg), link); | |
353 | } | |
354 | ||
355 | /* Drop any remaining regions. */ | |
356 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
357 | if (&rg->link == head) | |
358 | break; | |
359 | chg += rg->to - rg->from; | |
360 | list_del(&rg->link); | |
361 | kfree(rg); | |
362 | } | |
7b24d861 DB |
363 | |
364 | out: | |
365 | spin_unlock(&resv->lock); | |
96822904 AW |
366 | return chg; |
367 | } | |
368 | ||
1406ec9b | 369 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 370 | { |
1406ec9b | 371 | struct list_head *head = &resv->regions; |
84afd99b AW |
372 | struct file_region *rg; |
373 | long chg = 0; | |
374 | ||
7b24d861 | 375 | spin_lock(&resv->lock); |
84afd99b AW |
376 | /* Locate each segment we overlap with, and count that overlap. */ |
377 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
378 | long seg_from; |
379 | long seg_to; | |
84afd99b AW |
380 | |
381 | if (rg->to <= f) | |
382 | continue; | |
383 | if (rg->from >= t) | |
384 | break; | |
385 | ||
386 | seg_from = max(rg->from, f); | |
387 | seg_to = min(rg->to, t); | |
388 | ||
389 | chg += seg_to - seg_from; | |
390 | } | |
7b24d861 | 391 | spin_unlock(&resv->lock); |
84afd99b AW |
392 | |
393 | return chg; | |
394 | } | |
395 | ||
e7c4b0bf AW |
396 | /* |
397 | * Convert the address within this vma to the page offset within | |
398 | * the mapping, in pagecache page units; huge pages here. | |
399 | */ | |
a5516438 AK |
400 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
401 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 402 | { |
a5516438 AK |
403 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
404 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
405 | } |
406 | ||
0fe6e20b NH |
407 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
408 | unsigned long address) | |
409 | { | |
410 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
411 | } | |
412 | ||
08fba699 MG |
413 | /* |
414 | * Return the size of the pages allocated when backing a VMA. In the majority | |
415 | * cases this will be same size as used by the page table entries. | |
416 | */ | |
417 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
418 | { | |
419 | struct hstate *hstate; | |
420 | ||
421 | if (!is_vm_hugetlb_page(vma)) | |
422 | return PAGE_SIZE; | |
423 | ||
424 | hstate = hstate_vma(vma); | |
425 | ||
2415cf12 | 426 | return 1UL << huge_page_shift(hstate); |
08fba699 | 427 | } |
f340ca0f | 428 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 429 | |
3340289d MG |
430 | /* |
431 | * Return the page size being used by the MMU to back a VMA. In the majority | |
432 | * of cases, the page size used by the kernel matches the MMU size. On | |
433 | * architectures where it differs, an architecture-specific version of this | |
434 | * function is required. | |
435 | */ | |
436 | #ifndef vma_mmu_pagesize | |
437 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) | |
438 | { | |
439 | return vma_kernel_pagesize(vma); | |
440 | } | |
441 | #endif | |
442 | ||
84afd99b AW |
443 | /* |
444 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
445 | * bits of the reservation map pointer, which are always clear due to | |
446 | * alignment. | |
447 | */ | |
448 | #define HPAGE_RESV_OWNER (1UL << 0) | |
449 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 450 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 451 | |
a1e78772 MG |
452 | /* |
453 | * These helpers are used to track how many pages are reserved for | |
454 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
455 | * is guaranteed to have their future faults succeed. | |
456 | * | |
457 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
458 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
459 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
460 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
461 | * |
462 | * The private mapping reservation is represented in a subtly different | |
463 | * manner to a shared mapping. A shared mapping has a region map associated | |
464 | * with the underlying file, this region map represents the backing file | |
465 | * pages which have ever had a reservation assigned which this persists even | |
466 | * after the page is instantiated. A private mapping has a region map | |
467 | * associated with the original mmap which is attached to all VMAs which | |
468 | * reference it, this region map represents those offsets which have consumed | |
469 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 470 | */ |
e7c4b0bf AW |
471 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
472 | { | |
473 | return (unsigned long)vma->vm_private_data; | |
474 | } | |
475 | ||
476 | static void set_vma_private_data(struct vm_area_struct *vma, | |
477 | unsigned long value) | |
478 | { | |
479 | vma->vm_private_data = (void *)value; | |
480 | } | |
481 | ||
9119a41e | 482 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
483 | { |
484 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
485 | if (!resv_map) | |
486 | return NULL; | |
487 | ||
488 | kref_init(&resv_map->refs); | |
7b24d861 | 489 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
490 | INIT_LIST_HEAD(&resv_map->regions); |
491 | ||
492 | return resv_map; | |
493 | } | |
494 | ||
9119a41e | 495 | void resv_map_release(struct kref *ref) |
84afd99b AW |
496 | { |
497 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
498 | ||
499 | /* Clear out any active regions before we release the map. */ | |
1406ec9b | 500 | region_truncate(resv_map, 0); |
84afd99b AW |
501 | kfree(resv_map); |
502 | } | |
503 | ||
4e35f483 JK |
504 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
505 | { | |
506 | return inode->i_mapping->private_data; | |
507 | } | |
508 | ||
84afd99b | 509 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 510 | { |
81d1b09c | 511 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
512 | if (vma->vm_flags & VM_MAYSHARE) { |
513 | struct address_space *mapping = vma->vm_file->f_mapping; | |
514 | struct inode *inode = mapping->host; | |
515 | ||
516 | return inode_resv_map(inode); | |
517 | ||
518 | } else { | |
84afd99b AW |
519 | return (struct resv_map *)(get_vma_private_data(vma) & |
520 | ~HPAGE_RESV_MASK); | |
4e35f483 | 521 | } |
a1e78772 MG |
522 | } |
523 | ||
84afd99b | 524 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 525 | { |
81d1b09c SL |
526 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
527 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 528 | |
84afd99b AW |
529 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
530 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
531 | } |
532 | ||
533 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
534 | { | |
81d1b09c SL |
535 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
536 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
537 | |
538 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
539 | } |
540 | ||
541 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
542 | { | |
81d1b09c | 543 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
544 | |
545 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
546 | } |
547 | ||
04f2cbe3 | 548 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
549 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
550 | { | |
81d1b09c | 551 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 552 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
553 | vma->vm_private_data = (void *)0; |
554 | } | |
555 | ||
556 | /* Returns true if the VMA has associated reserve pages */ | |
af0ed73e | 557 | static int vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 558 | { |
af0ed73e JK |
559 | if (vma->vm_flags & VM_NORESERVE) { |
560 | /* | |
561 | * This address is already reserved by other process(chg == 0), | |
562 | * so, we should decrement reserved count. Without decrementing, | |
563 | * reserve count remains after releasing inode, because this | |
564 | * allocated page will go into page cache and is regarded as | |
565 | * coming from reserved pool in releasing step. Currently, we | |
566 | * don't have any other solution to deal with this situation | |
567 | * properly, so add work-around here. | |
568 | */ | |
569 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
570 | return 1; | |
571 | else | |
572 | return 0; | |
573 | } | |
a63884e9 JK |
574 | |
575 | /* Shared mappings always use reserves */ | |
f83a275d | 576 | if (vma->vm_flags & VM_MAYSHARE) |
7f09ca51 | 577 | return 1; |
a63884e9 JK |
578 | |
579 | /* | |
580 | * Only the process that called mmap() has reserves for | |
581 | * private mappings. | |
582 | */ | |
7f09ca51 MG |
583 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
584 | return 1; | |
a63884e9 | 585 | |
7f09ca51 | 586 | return 0; |
a1e78772 MG |
587 | } |
588 | ||
a5516438 | 589 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
590 | { |
591 | int nid = page_to_nid(page); | |
0edaecfa | 592 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
593 | h->free_huge_pages++; |
594 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
595 | } |
596 | ||
bf50bab2 NH |
597 | static struct page *dequeue_huge_page_node(struct hstate *h, int nid) |
598 | { | |
599 | struct page *page; | |
600 | ||
c8721bbb NH |
601 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
602 | if (!is_migrate_isolate_page(page)) | |
603 | break; | |
604 | /* | |
605 | * if 'non-isolated free hugepage' not found on the list, | |
606 | * the allocation fails. | |
607 | */ | |
608 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 609 | return NULL; |
0edaecfa | 610 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 611 | set_page_refcounted(page); |
bf50bab2 NH |
612 | h->free_huge_pages--; |
613 | h->free_huge_pages_node[nid]--; | |
614 | return page; | |
615 | } | |
616 | ||
86cdb465 NH |
617 | /* Movability of hugepages depends on migration support. */ |
618 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
619 | { | |
100873d7 | 620 | if (hugepages_treat_as_movable || hugepage_migration_supported(h)) |
86cdb465 NH |
621 | return GFP_HIGHUSER_MOVABLE; |
622 | else | |
623 | return GFP_HIGHUSER; | |
624 | } | |
625 | ||
a5516438 AK |
626 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
627 | struct vm_area_struct *vma, | |
af0ed73e JK |
628 | unsigned long address, int avoid_reserve, |
629 | long chg) | |
1da177e4 | 630 | { |
b1c12cbc | 631 | struct page *page = NULL; |
480eccf9 | 632 | struct mempolicy *mpol; |
19770b32 | 633 | nodemask_t *nodemask; |
c0ff7453 | 634 | struct zonelist *zonelist; |
dd1a239f MG |
635 | struct zone *zone; |
636 | struct zoneref *z; | |
cc9a6c87 | 637 | unsigned int cpuset_mems_cookie; |
1da177e4 | 638 | |
a1e78772 MG |
639 | /* |
640 | * A child process with MAP_PRIVATE mappings created by their parent | |
641 | * have no page reserves. This check ensures that reservations are | |
642 | * not "stolen". The child may still get SIGKILLed | |
643 | */ | |
af0ed73e | 644 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 645 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 646 | goto err; |
a1e78772 | 647 | |
04f2cbe3 | 648 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 649 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 650 | goto err; |
04f2cbe3 | 651 | |
9966c4bb | 652 | retry_cpuset: |
d26914d1 | 653 | cpuset_mems_cookie = read_mems_allowed_begin(); |
9966c4bb | 654 | zonelist = huge_zonelist(vma, address, |
86cdb465 | 655 | htlb_alloc_mask(h), &mpol, &nodemask); |
9966c4bb | 656 | |
19770b32 MG |
657 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
658 | MAX_NR_ZONES - 1, nodemask) { | |
344736f2 | 659 | if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) { |
bf50bab2 NH |
660 | page = dequeue_huge_page_node(h, zone_to_nid(zone)); |
661 | if (page) { | |
af0ed73e JK |
662 | if (avoid_reserve) |
663 | break; | |
664 | if (!vma_has_reserves(vma, chg)) | |
665 | break; | |
666 | ||
07443a85 | 667 | SetPagePrivate(page); |
af0ed73e | 668 | h->resv_huge_pages--; |
bf50bab2 NH |
669 | break; |
670 | } | |
3abf7afd | 671 | } |
1da177e4 | 672 | } |
cc9a6c87 | 673 | |
52cd3b07 | 674 | mpol_cond_put(mpol); |
d26914d1 | 675 | if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) |
cc9a6c87 | 676 | goto retry_cpuset; |
1da177e4 | 677 | return page; |
cc9a6c87 MG |
678 | |
679 | err: | |
cc9a6c87 | 680 | return NULL; |
1da177e4 LT |
681 | } |
682 | ||
1cac6f2c LC |
683 | /* |
684 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
685 | * We may have allocated or freed a huge page based on a different | |
686 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
687 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
688 | * node for alloc or free. | |
689 | */ | |
690 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
691 | { | |
692 | nid = next_node(nid, *nodes_allowed); | |
693 | if (nid == MAX_NUMNODES) | |
694 | nid = first_node(*nodes_allowed); | |
695 | VM_BUG_ON(nid >= MAX_NUMNODES); | |
696 | ||
697 | return nid; | |
698 | } | |
699 | ||
700 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
701 | { | |
702 | if (!node_isset(nid, *nodes_allowed)) | |
703 | nid = next_node_allowed(nid, nodes_allowed); | |
704 | return nid; | |
705 | } | |
706 | ||
707 | /* | |
708 | * returns the previously saved node ["this node"] from which to | |
709 | * allocate a persistent huge page for the pool and advance the | |
710 | * next node from which to allocate, handling wrap at end of node | |
711 | * mask. | |
712 | */ | |
713 | static int hstate_next_node_to_alloc(struct hstate *h, | |
714 | nodemask_t *nodes_allowed) | |
715 | { | |
716 | int nid; | |
717 | ||
718 | VM_BUG_ON(!nodes_allowed); | |
719 | ||
720 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
721 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
722 | ||
723 | return nid; | |
724 | } | |
725 | ||
726 | /* | |
727 | * helper for free_pool_huge_page() - return the previously saved | |
728 | * node ["this node"] from which to free a huge page. Advance the | |
729 | * next node id whether or not we find a free huge page to free so | |
730 | * that the next attempt to free addresses the next node. | |
731 | */ | |
732 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
733 | { | |
734 | int nid; | |
735 | ||
736 | VM_BUG_ON(!nodes_allowed); | |
737 | ||
738 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
739 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
740 | ||
741 | return nid; | |
742 | } | |
743 | ||
744 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
745 | for (nr_nodes = nodes_weight(*mask); \ | |
746 | nr_nodes > 0 && \ | |
747 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
748 | nr_nodes--) | |
749 | ||
750 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
751 | for (nr_nodes = nodes_weight(*mask); \ | |
752 | nr_nodes > 0 && \ | |
753 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
754 | nr_nodes--) | |
755 | ||
944d9fec LC |
756 | #if defined(CONFIG_CMA) && defined(CONFIG_X86_64) |
757 | static void destroy_compound_gigantic_page(struct page *page, | |
758 | unsigned long order) | |
759 | { | |
760 | int i; | |
761 | int nr_pages = 1 << order; | |
762 | struct page *p = page + 1; | |
763 | ||
764 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { | |
765 | __ClearPageTail(p); | |
766 | set_page_refcounted(p); | |
767 | p->first_page = NULL; | |
768 | } | |
769 | ||
770 | set_compound_order(page, 0); | |
771 | __ClearPageHead(page); | |
772 | } | |
773 | ||
774 | static void free_gigantic_page(struct page *page, unsigned order) | |
775 | { | |
776 | free_contig_range(page_to_pfn(page), 1 << order); | |
777 | } | |
778 | ||
779 | static int __alloc_gigantic_page(unsigned long start_pfn, | |
780 | unsigned long nr_pages) | |
781 | { | |
782 | unsigned long end_pfn = start_pfn + nr_pages; | |
783 | return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE); | |
784 | } | |
785 | ||
786 | static bool pfn_range_valid_gigantic(unsigned long start_pfn, | |
787 | unsigned long nr_pages) | |
788 | { | |
789 | unsigned long i, end_pfn = start_pfn + nr_pages; | |
790 | struct page *page; | |
791 | ||
792 | for (i = start_pfn; i < end_pfn; i++) { | |
793 | if (!pfn_valid(i)) | |
794 | return false; | |
795 | ||
796 | page = pfn_to_page(i); | |
797 | ||
798 | if (PageReserved(page)) | |
799 | return false; | |
800 | ||
801 | if (page_count(page) > 0) | |
802 | return false; | |
803 | ||
804 | if (PageHuge(page)) | |
805 | return false; | |
806 | } | |
807 | ||
808 | return true; | |
809 | } | |
810 | ||
811 | static bool zone_spans_last_pfn(const struct zone *zone, | |
812 | unsigned long start_pfn, unsigned long nr_pages) | |
813 | { | |
814 | unsigned long last_pfn = start_pfn + nr_pages - 1; | |
815 | return zone_spans_pfn(zone, last_pfn); | |
816 | } | |
817 | ||
818 | static struct page *alloc_gigantic_page(int nid, unsigned order) | |
819 | { | |
820 | unsigned long nr_pages = 1 << order; | |
821 | unsigned long ret, pfn, flags; | |
822 | struct zone *z; | |
823 | ||
824 | z = NODE_DATA(nid)->node_zones; | |
825 | for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) { | |
826 | spin_lock_irqsave(&z->lock, flags); | |
827 | ||
828 | pfn = ALIGN(z->zone_start_pfn, nr_pages); | |
829 | while (zone_spans_last_pfn(z, pfn, nr_pages)) { | |
830 | if (pfn_range_valid_gigantic(pfn, nr_pages)) { | |
831 | /* | |
832 | * We release the zone lock here because | |
833 | * alloc_contig_range() will also lock the zone | |
834 | * at some point. If there's an allocation | |
835 | * spinning on this lock, it may win the race | |
836 | * and cause alloc_contig_range() to fail... | |
837 | */ | |
838 | spin_unlock_irqrestore(&z->lock, flags); | |
839 | ret = __alloc_gigantic_page(pfn, nr_pages); | |
840 | if (!ret) | |
841 | return pfn_to_page(pfn); | |
842 | spin_lock_irqsave(&z->lock, flags); | |
843 | } | |
844 | pfn += nr_pages; | |
845 | } | |
846 | ||
847 | spin_unlock_irqrestore(&z->lock, flags); | |
848 | } | |
849 | ||
850 | return NULL; | |
851 | } | |
852 | ||
853 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
854 | static void prep_compound_gigantic_page(struct page *page, unsigned long order); | |
855 | ||
856 | static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) | |
857 | { | |
858 | struct page *page; | |
859 | ||
860 | page = alloc_gigantic_page(nid, huge_page_order(h)); | |
861 | if (page) { | |
862 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
863 | prep_new_huge_page(h, page, nid); | |
864 | } | |
865 | ||
866 | return page; | |
867 | } | |
868 | ||
869 | static int alloc_fresh_gigantic_page(struct hstate *h, | |
870 | nodemask_t *nodes_allowed) | |
871 | { | |
872 | struct page *page = NULL; | |
873 | int nr_nodes, node; | |
874 | ||
875 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
876 | page = alloc_fresh_gigantic_page_node(h, node); | |
877 | if (page) | |
878 | return 1; | |
879 | } | |
880 | ||
881 | return 0; | |
882 | } | |
883 | ||
884 | static inline bool gigantic_page_supported(void) { return true; } | |
885 | #else | |
886 | static inline bool gigantic_page_supported(void) { return false; } | |
887 | static inline void free_gigantic_page(struct page *page, unsigned order) { } | |
888 | static inline void destroy_compound_gigantic_page(struct page *page, | |
889 | unsigned long order) { } | |
890 | static inline int alloc_fresh_gigantic_page(struct hstate *h, | |
891 | nodemask_t *nodes_allowed) { return 0; } | |
892 | #endif | |
893 | ||
a5516438 | 894 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
895 | { |
896 | int i; | |
a5516438 | 897 | |
944d9fec LC |
898 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
899 | return; | |
18229df5 | 900 | |
a5516438 AK |
901 | h->nr_huge_pages--; |
902 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
903 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
904 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
905 | 1 << PG_referenced | 1 << PG_dirty | | |
a7407a27 LC |
906 | 1 << PG_active | 1 << PG_private | |
907 | 1 << PG_writeback); | |
6af2acb6 | 908 | } |
309381fe | 909 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
6af2acb6 AL |
910 | set_compound_page_dtor(page, NULL); |
911 | set_page_refcounted(page); | |
944d9fec LC |
912 | if (hstate_is_gigantic(h)) { |
913 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
914 | free_gigantic_page(page, huge_page_order(h)); | |
915 | } else { | |
916 | arch_release_hugepage(page); | |
917 | __free_pages(page, huge_page_order(h)); | |
918 | } | |
6af2acb6 AL |
919 | } |
920 | ||
e5ff2159 AK |
921 | struct hstate *size_to_hstate(unsigned long size) |
922 | { | |
923 | struct hstate *h; | |
924 | ||
925 | for_each_hstate(h) { | |
926 | if (huge_page_size(h) == size) | |
927 | return h; | |
928 | } | |
929 | return NULL; | |
930 | } | |
931 | ||
bcc54222 NH |
932 | /* |
933 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
934 | * to hstate->hugepage_activelist.) | |
935 | * | |
936 | * This function can be called for tail pages, but never returns true for them. | |
937 | */ | |
938 | bool page_huge_active(struct page *page) | |
939 | { | |
940 | VM_BUG_ON_PAGE(!PageHuge(page), page); | |
941 | return PageHead(page) && PagePrivate(&page[1]); | |
942 | } | |
943 | ||
944 | /* never called for tail page */ | |
945 | static void set_page_huge_active(struct page *page) | |
946 | { | |
947 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
948 | SetPagePrivate(&page[1]); | |
949 | } | |
950 | ||
951 | static void clear_page_huge_active(struct page *page) | |
952 | { | |
953 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
954 | ClearPagePrivate(&page[1]); | |
955 | } | |
956 | ||
8f1d26d0 | 957 | void free_huge_page(struct page *page) |
27a85ef1 | 958 | { |
a5516438 AK |
959 | /* |
960 | * Can't pass hstate in here because it is called from the | |
961 | * compound page destructor. | |
962 | */ | |
e5ff2159 | 963 | struct hstate *h = page_hstate(page); |
7893d1d5 | 964 | int nid = page_to_nid(page); |
90481622 DG |
965 | struct hugepage_subpool *spool = |
966 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 967 | bool restore_reserve; |
27a85ef1 | 968 | |
e5df70ab | 969 | set_page_private(page, 0); |
23be7468 | 970 | page->mapping = NULL; |
7893d1d5 | 971 | BUG_ON(page_count(page)); |
0fe6e20b | 972 | BUG_ON(page_mapcount(page)); |
07443a85 | 973 | restore_reserve = PagePrivate(page); |
16c794b4 | 974 | ClearPagePrivate(page); |
27a85ef1 | 975 | |
1c5ecae3 MK |
976 | /* |
977 | * A return code of zero implies that the subpool will be under its | |
978 | * minimum size if the reservation is not restored after page is free. | |
979 | * Therefore, force restore_reserve operation. | |
980 | */ | |
981 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
982 | restore_reserve = true; | |
983 | ||
27a85ef1 | 984 | spin_lock(&hugetlb_lock); |
bcc54222 | 985 | clear_page_huge_active(page); |
6d76dcf4 AK |
986 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
987 | pages_per_huge_page(h), page); | |
07443a85 JK |
988 | if (restore_reserve) |
989 | h->resv_huge_pages++; | |
990 | ||
944d9fec | 991 | if (h->surplus_huge_pages_node[nid]) { |
0edaecfa AK |
992 | /* remove the page from active list */ |
993 | list_del(&page->lru); | |
a5516438 AK |
994 | update_and_free_page(h, page); |
995 | h->surplus_huge_pages--; | |
996 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 997 | } else { |
5d3a551c | 998 | arch_clear_hugepage_flags(page); |
a5516438 | 999 | enqueue_huge_page(h, page); |
7893d1d5 | 1000 | } |
27a85ef1 DG |
1001 | spin_unlock(&hugetlb_lock); |
1002 | } | |
1003 | ||
a5516438 | 1004 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1005 | { |
0edaecfa | 1006 | INIT_LIST_HEAD(&page->lru); |
b7ba30c6 AK |
1007 | set_compound_page_dtor(page, free_huge_page); |
1008 | spin_lock(&hugetlb_lock); | |
9dd540e2 | 1009 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
1010 | h->nr_huge_pages++; |
1011 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 AK |
1012 | spin_unlock(&hugetlb_lock); |
1013 | put_page(page); /* free it into the hugepage allocator */ | |
1014 | } | |
1015 | ||
2906dd52 | 1016 | static void prep_compound_gigantic_page(struct page *page, unsigned long order) |
20a0307c WF |
1017 | { |
1018 | int i; | |
1019 | int nr_pages = 1 << order; | |
1020 | struct page *p = page + 1; | |
1021 | ||
1022 | /* we rely on prep_new_huge_page to set the destructor */ | |
1023 | set_compound_order(page, order); | |
1024 | __SetPageHead(page); | |
ef5a22be | 1025 | __ClearPageReserved(page); |
20a0307c | 1026 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1027 | /* |
1028 | * For gigantic hugepages allocated through bootmem at | |
1029 | * boot, it's safer to be consistent with the not-gigantic | |
1030 | * hugepages and clear the PG_reserved bit from all tail pages | |
1031 | * too. Otherwse drivers using get_user_pages() to access tail | |
1032 | * pages may get the reference counting wrong if they see | |
1033 | * PG_reserved set on a tail page (despite the head page not | |
1034 | * having PG_reserved set). Enforcing this consistency between | |
1035 | * head and tail pages allows drivers to optimize away a check | |
1036 | * on the head page when they need know if put_page() is needed | |
1037 | * after get_user_pages(). | |
1038 | */ | |
1039 | __ClearPageReserved(p); | |
58a84aa9 | 1040 | set_page_count(p, 0); |
20a0307c | 1041 | p->first_page = page; |
44fc8057 DR |
1042 | /* Make sure p->first_page is always valid for PageTail() */ |
1043 | smp_wmb(); | |
1044 | __SetPageTail(p); | |
20a0307c WF |
1045 | } |
1046 | } | |
1047 | ||
7795912c AM |
1048 | /* |
1049 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1050 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1051 | * details. | |
1052 | */ | |
20a0307c WF |
1053 | int PageHuge(struct page *page) |
1054 | { | |
20a0307c WF |
1055 | if (!PageCompound(page)) |
1056 | return 0; | |
1057 | ||
1058 | page = compound_head(page); | |
758f66a2 | 1059 | return get_compound_page_dtor(page) == free_huge_page; |
20a0307c | 1060 | } |
43131e14 NH |
1061 | EXPORT_SYMBOL_GPL(PageHuge); |
1062 | ||
27c73ae7 AA |
1063 | /* |
1064 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1065 | * normal or transparent huge pages. | |
1066 | */ | |
1067 | int PageHeadHuge(struct page *page_head) | |
1068 | { | |
27c73ae7 AA |
1069 | if (!PageHead(page_head)) |
1070 | return 0; | |
1071 | ||
758f66a2 | 1072 | return get_compound_page_dtor(page_head) == free_huge_page; |
27c73ae7 | 1073 | } |
27c73ae7 | 1074 | |
13d60f4b ZY |
1075 | pgoff_t __basepage_index(struct page *page) |
1076 | { | |
1077 | struct page *page_head = compound_head(page); | |
1078 | pgoff_t index = page_index(page_head); | |
1079 | unsigned long compound_idx; | |
1080 | ||
1081 | if (!PageHuge(page_head)) | |
1082 | return page_index(page); | |
1083 | ||
1084 | if (compound_order(page_head) >= MAX_ORDER) | |
1085 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1086 | else | |
1087 | compound_idx = page - page_head; | |
1088 | ||
1089 | return (index << compound_order(page_head)) + compound_idx; | |
1090 | } | |
1091 | ||
a5516438 | 1092 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
1da177e4 | 1093 | { |
1da177e4 | 1094 | struct page *page; |
f96efd58 | 1095 | |
6484eb3e | 1096 | page = alloc_pages_exact_node(nid, |
86cdb465 | 1097 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
551883ae | 1098 | __GFP_REPEAT|__GFP_NOWARN, |
a5516438 | 1099 | huge_page_order(h)); |
1da177e4 | 1100 | if (page) { |
7f2e9525 | 1101 | if (arch_prepare_hugepage(page)) { |
caff3a2c | 1102 | __free_pages(page, huge_page_order(h)); |
7b8ee84d | 1103 | return NULL; |
7f2e9525 | 1104 | } |
a5516438 | 1105 | prep_new_huge_page(h, page, nid); |
1da177e4 | 1106 | } |
63b4613c NA |
1107 | |
1108 | return page; | |
1109 | } | |
1110 | ||
b2261026 JK |
1111 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
1112 | { | |
1113 | struct page *page; | |
1114 | int nr_nodes, node; | |
1115 | int ret = 0; | |
1116 | ||
1117 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1118 | page = alloc_fresh_huge_page_node(h, node); | |
1119 | if (page) { | |
1120 | ret = 1; | |
1121 | break; | |
1122 | } | |
1123 | } | |
1124 | ||
1125 | if (ret) | |
1126 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
1127 | else | |
1128 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
1129 | ||
1130 | return ret; | |
1131 | } | |
1132 | ||
e8c5c824 LS |
1133 | /* |
1134 | * Free huge page from pool from next node to free. | |
1135 | * Attempt to keep persistent huge pages more or less | |
1136 | * balanced over allowed nodes. | |
1137 | * Called with hugetlb_lock locked. | |
1138 | */ | |
6ae11b27 LS |
1139 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1140 | bool acct_surplus) | |
e8c5c824 | 1141 | { |
b2261026 | 1142 | int nr_nodes, node; |
e8c5c824 LS |
1143 | int ret = 0; |
1144 | ||
b2261026 | 1145 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1146 | /* |
1147 | * If we're returning unused surplus pages, only examine | |
1148 | * nodes with surplus pages. | |
1149 | */ | |
b2261026 JK |
1150 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1151 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1152 | struct page *page = |
b2261026 | 1153 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1154 | struct page, lru); |
1155 | list_del(&page->lru); | |
1156 | h->free_huge_pages--; | |
b2261026 | 1157 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1158 | if (acct_surplus) { |
1159 | h->surplus_huge_pages--; | |
b2261026 | 1160 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1161 | } |
e8c5c824 LS |
1162 | update_and_free_page(h, page); |
1163 | ret = 1; | |
9a76db09 | 1164 | break; |
e8c5c824 | 1165 | } |
b2261026 | 1166 | } |
e8c5c824 LS |
1167 | |
1168 | return ret; | |
1169 | } | |
1170 | ||
c8721bbb NH |
1171 | /* |
1172 | * Dissolve a given free hugepage into free buddy pages. This function does | |
1173 | * nothing for in-use (including surplus) hugepages. | |
1174 | */ | |
1175 | static void dissolve_free_huge_page(struct page *page) | |
1176 | { | |
1177 | spin_lock(&hugetlb_lock); | |
1178 | if (PageHuge(page) && !page_count(page)) { | |
1179 | struct hstate *h = page_hstate(page); | |
1180 | int nid = page_to_nid(page); | |
1181 | list_del(&page->lru); | |
1182 | h->free_huge_pages--; | |
1183 | h->free_huge_pages_node[nid]--; | |
1184 | update_and_free_page(h, page); | |
1185 | } | |
1186 | spin_unlock(&hugetlb_lock); | |
1187 | } | |
1188 | ||
1189 | /* | |
1190 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1191 | * make specified memory blocks removable from the system. | |
1192 | * Note that start_pfn should aligned with (minimum) hugepage size. | |
1193 | */ | |
1194 | void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) | |
1195 | { | |
c8721bbb | 1196 | unsigned long pfn; |
c8721bbb | 1197 | |
d0177639 LZ |
1198 | if (!hugepages_supported()) |
1199 | return; | |
1200 | ||
641844f5 NH |
1201 | VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order)); |
1202 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) | |
c8721bbb NH |
1203 | dissolve_free_huge_page(pfn_to_page(pfn)); |
1204 | } | |
1205 | ||
bf50bab2 | 1206 | static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) |
7893d1d5 AL |
1207 | { |
1208 | struct page *page; | |
bf50bab2 | 1209 | unsigned int r_nid; |
7893d1d5 | 1210 | |
bae7f4ae | 1211 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1212 | return NULL; |
1213 | ||
d1c3fb1f NA |
1214 | /* |
1215 | * Assume we will successfully allocate the surplus page to | |
1216 | * prevent racing processes from causing the surplus to exceed | |
1217 | * overcommit | |
1218 | * | |
1219 | * This however introduces a different race, where a process B | |
1220 | * tries to grow the static hugepage pool while alloc_pages() is | |
1221 | * called by process A. B will only examine the per-node | |
1222 | * counters in determining if surplus huge pages can be | |
1223 | * converted to normal huge pages in adjust_pool_surplus(). A | |
1224 | * won't be able to increment the per-node counter, until the | |
1225 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
1226 | * no more huge pages can be converted from surplus to normal | |
1227 | * state (and doesn't try to convert again). Thus, we have a | |
1228 | * case where a surplus huge page exists, the pool is grown, and | |
1229 | * the surplus huge page still exists after, even though it | |
1230 | * should just have been converted to a normal huge page. This | |
1231 | * does not leak memory, though, as the hugepage will be freed | |
1232 | * once it is out of use. It also does not allow the counters to | |
1233 | * go out of whack in adjust_pool_surplus() as we don't modify | |
1234 | * the node values until we've gotten the hugepage and only the | |
1235 | * per-node value is checked there. | |
1236 | */ | |
1237 | spin_lock(&hugetlb_lock); | |
a5516438 | 1238 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
1239 | spin_unlock(&hugetlb_lock); |
1240 | return NULL; | |
1241 | } else { | |
a5516438 AK |
1242 | h->nr_huge_pages++; |
1243 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
1244 | } |
1245 | spin_unlock(&hugetlb_lock); | |
1246 | ||
bf50bab2 | 1247 | if (nid == NUMA_NO_NODE) |
86cdb465 | 1248 | page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP| |
bf50bab2 NH |
1249 | __GFP_REPEAT|__GFP_NOWARN, |
1250 | huge_page_order(h)); | |
1251 | else | |
1252 | page = alloc_pages_exact_node(nid, | |
86cdb465 | 1253 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
bf50bab2 | 1254 | __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); |
d1c3fb1f | 1255 | |
caff3a2c GS |
1256 | if (page && arch_prepare_hugepage(page)) { |
1257 | __free_pages(page, huge_page_order(h)); | |
ea5768c7 | 1258 | page = NULL; |
caff3a2c GS |
1259 | } |
1260 | ||
d1c3fb1f | 1261 | spin_lock(&hugetlb_lock); |
7893d1d5 | 1262 | if (page) { |
0edaecfa | 1263 | INIT_LIST_HEAD(&page->lru); |
bf50bab2 | 1264 | r_nid = page_to_nid(page); |
7893d1d5 | 1265 | set_compound_page_dtor(page, free_huge_page); |
9dd540e2 | 1266 | set_hugetlb_cgroup(page, NULL); |
d1c3fb1f NA |
1267 | /* |
1268 | * We incremented the global counters already | |
1269 | */ | |
bf50bab2 NH |
1270 | h->nr_huge_pages_node[r_nid]++; |
1271 | h->surplus_huge_pages_node[r_nid]++; | |
3b116300 | 1272 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 1273 | } else { |
a5516438 AK |
1274 | h->nr_huge_pages--; |
1275 | h->surplus_huge_pages--; | |
3b116300 | 1276 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 1277 | } |
d1c3fb1f | 1278 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1279 | |
1280 | return page; | |
1281 | } | |
1282 | ||
bf50bab2 NH |
1283 | /* |
1284 | * This allocation function is useful in the context where vma is irrelevant. | |
1285 | * E.g. soft-offlining uses this function because it only cares physical | |
1286 | * address of error page. | |
1287 | */ | |
1288 | struct page *alloc_huge_page_node(struct hstate *h, int nid) | |
1289 | { | |
4ef91848 | 1290 | struct page *page = NULL; |
bf50bab2 NH |
1291 | |
1292 | spin_lock(&hugetlb_lock); | |
4ef91848 JK |
1293 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
1294 | page = dequeue_huge_page_node(h, nid); | |
bf50bab2 NH |
1295 | spin_unlock(&hugetlb_lock); |
1296 | ||
94ae8ba7 | 1297 | if (!page) |
bf50bab2 NH |
1298 | page = alloc_buddy_huge_page(h, nid); |
1299 | ||
1300 | return page; | |
1301 | } | |
1302 | ||
e4e574b7 | 1303 | /* |
25985edc | 1304 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1305 | * of size 'delta'. |
1306 | */ | |
a5516438 | 1307 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1308 | { |
1309 | struct list_head surplus_list; | |
1310 | struct page *page, *tmp; | |
1311 | int ret, i; | |
1312 | int needed, allocated; | |
28073b02 | 1313 | bool alloc_ok = true; |
e4e574b7 | 1314 | |
a5516438 | 1315 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1316 | if (needed <= 0) { |
a5516438 | 1317 | h->resv_huge_pages += delta; |
e4e574b7 | 1318 | return 0; |
ac09b3a1 | 1319 | } |
e4e574b7 AL |
1320 | |
1321 | allocated = 0; | |
1322 | INIT_LIST_HEAD(&surplus_list); | |
1323 | ||
1324 | ret = -ENOMEM; | |
1325 | retry: | |
1326 | spin_unlock(&hugetlb_lock); | |
1327 | for (i = 0; i < needed; i++) { | |
bf50bab2 | 1328 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
28073b02 HD |
1329 | if (!page) { |
1330 | alloc_ok = false; | |
1331 | break; | |
1332 | } | |
e4e574b7 AL |
1333 | list_add(&page->lru, &surplus_list); |
1334 | } | |
28073b02 | 1335 | allocated += i; |
e4e574b7 AL |
1336 | |
1337 | /* | |
1338 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1339 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1340 | */ | |
1341 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1342 | needed = (h->resv_huge_pages + delta) - |
1343 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1344 | if (needed > 0) { |
1345 | if (alloc_ok) | |
1346 | goto retry; | |
1347 | /* | |
1348 | * We were not able to allocate enough pages to | |
1349 | * satisfy the entire reservation so we free what | |
1350 | * we've allocated so far. | |
1351 | */ | |
1352 | goto free; | |
1353 | } | |
e4e574b7 AL |
1354 | /* |
1355 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1356 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1357 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1358 | * allocator. Commit the entire reservation here to prevent another |
1359 | * process from stealing the pages as they are added to the pool but | |
1360 | * before they are reserved. | |
e4e574b7 AL |
1361 | */ |
1362 | needed += allocated; | |
a5516438 | 1363 | h->resv_huge_pages += delta; |
e4e574b7 | 1364 | ret = 0; |
a9869b83 | 1365 | |
19fc3f0a | 1366 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1367 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1368 | if ((--needed) < 0) |
1369 | break; | |
a9869b83 NH |
1370 | /* |
1371 | * This page is now managed by the hugetlb allocator and has | |
1372 | * no users -- drop the buddy allocator's reference. | |
1373 | */ | |
1374 | put_page_testzero(page); | |
309381fe | 1375 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 1376 | enqueue_huge_page(h, page); |
19fc3f0a | 1377 | } |
28073b02 | 1378 | free: |
b0365c8d | 1379 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1380 | |
1381 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1382 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1383 | put_page(page); | |
a9869b83 | 1384 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1385 | |
1386 | return ret; | |
1387 | } | |
1388 | ||
1389 | /* | |
1390 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
1391 | * allocated to satisfy the reservation must be explicitly freed if they were | |
1392 | * never used. | |
685f3457 | 1393 | * Called with hugetlb_lock held. |
e4e574b7 | 1394 | */ |
a5516438 AK |
1395 | static void return_unused_surplus_pages(struct hstate *h, |
1396 | unsigned long unused_resv_pages) | |
e4e574b7 | 1397 | { |
e4e574b7 AL |
1398 | unsigned long nr_pages; |
1399 | ||
ac09b3a1 | 1400 | /* Uncommit the reservation */ |
a5516438 | 1401 | h->resv_huge_pages -= unused_resv_pages; |
ac09b3a1 | 1402 | |
aa888a74 | 1403 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 1404 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1405 | return; |
1406 | ||
a5516438 | 1407 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1408 | |
685f3457 LS |
1409 | /* |
1410 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1411 | * evenly across all nodes with memory. Iterate across these nodes |
1412 | * until we can no longer free unreserved surplus pages. This occurs | |
1413 | * when the nodes with surplus pages have no free pages. | |
1414 | * free_pool_huge_page() will balance the the freed pages across the | |
1415 | * on-line nodes with memory and will handle the hstate accounting. | |
685f3457 LS |
1416 | */ |
1417 | while (nr_pages--) { | |
8cebfcd0 | 1418 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
685f3457 | 1419 | break; |
7848a4bf | 1420 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 AL |
1421 | } |
1422 | } | |
1423 | ||
c37f9fb1 AW |
1424 | /* |
1425 | * Determine if the huge page at addr within the vma has an associated | |
1426 | * reservation. Where it does not we will need to logically increase | |
90481622 DG |
1427 | * reservation and actually increase subpool usage before an allocation |
1428 | * can occur. Where any new reservation would be required the | |
1429 | * reservation change is prepared, but not committed. Once the page | |
1430 | * has been allocated from the subpool and instantiated the change should | |
1431 | * be committed via vma_commit_reservation. No action is required on | |
1432 | * failure. | |
c37f9fb1 | 1433 | */ |
e2f17d94 | 1434 | static long vma_needs_reservation(struct hstate *h, |
a5516438 | 1435 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 1436 | { |
4e35f483 JK |
1437 | struct resv_map *resv; |
1438 | pgoff_t idx; | |
1439 | long chg; | |
c37f9fb1 | 1440 | |
4e35f483 JK |
1441 | resv = vma_resv_map(vma); |
1442 | if (!resv) | |
84afd99b | 1443 | return 1; |
c37f9fb1 | 1444 | |
4e35f483 JK |
1445 | idx = vma_hugecache_offset(h, vma, addr); |
1446 | chg = region_chg(resv, idx, idx + 1); | |
84afd99b | 1447 | |
4e35f483 JK |
1448 | if (vma->vm_flags & VM_MAYSHARE) |
1449 | return chg; | |
1450 | else | |
1451 | return chg < 0 ? chg : 0; | |
c37f9fb1 | 1452 | } |
a5516438 AK |
1453 | static void vma_commit_reservation(struct hstate *h, |
1454 | struct vm_area_struct *vma, unsigned long addr) | |
c37f9fb1 | 1455 | { |
4e35f483 JK |
1456 | struct resv_map *resv; |
1457 | pgoff_t idx; | |
84afd99b | 1458 | |
4e35f483 JK |
1459 | resv = vma_resv_map(vma); |
1460 | if (!resv) | |
1461 | return; | |
84afd99b | 1462 | |
4e35f483 JK |
1463 | idx = vma_hugecache_offset(h, vma, addr); |
1464 | region_add(resv, idx, idx + 1); | |
c37f9fb1 AW |
1465 | } |
1466 | ||
a1e78772 | 1467 | static struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 1468 | unsigned long addr, int avoid_reserve) |
1da177e4 | 1469 | { |
90481622 | 1470 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 1471 | struct hstate *h = hstate_vma(vma); |
348ea204 | 1472 | struct page *page; |
e2f17d94 | 1473 | long chg; |
6d76dcf4 AK |
1474 | int ret, idx; |
1475 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 1476 | |
6d76dcf4 | 1477 | idx = hstate_index(h); |
a1e78772 | 1478 | /* |
90481622 DG |
1479 | * Processes that did not create the mapping will have no |
1480 | * reserves and will not have accounted against subpool | |
1481 | * limit. Check that the subpool limit can be made before | |
1482 | * satisfying the allocation MAP_NORESERVE mappings may also | |
1483 | * need pages and subpool limit allocated allocated if no reserve | |
1484 | * mapping overlaps. | |
a1e78772 | 1485 | */ |
a5516438 | 1486 | chg = vma_needs_reservation(h, vma, addr); |
c37f9fb1 | 1487 | if (chg < 0) |
76dcee75 | 1488 | return ERR_PTR(-ENOMEM); |
8bb3f12e | 1489 | if (chg || avoid_reserve) |
1c5ecae3 | 1490 | if (hugepage_subpool_get_pages(spool, 1) < 0) |
76dcee75 | 1491 | return ERR_PTR(-ENOSPC); |
1da177e4 | 1492 | |
6d76dcf4 | 1493 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f JZ |
1494 | if (ret) |
1495 | goto out_subpool_put; | |
1496 | ||
1da177e4 | 1497 | spin_lock(&hugetlb_lock); |
af0ed73e | 1498 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg); |
81a6fcae | 1499 | if (!page) { |
94ae8ba7 | 1500 | spin_unlock(&hugetlb_lock); |
bf50bab2 | 1501 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
8f34af6f JZ |
1502 | if (!page) |
1503 | goto out_uncharge_cgroup; | |
1504 | ||
79dbb236 AK |
1505 | spin_lock(&hugetlb_lock); |
1506 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 1507 | /* Fall through */ |
68842c9b | 1508 | } |
81a6fcae JK |
1509 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
1510 | spin_unlock(&hugetlb_lock); | |
348ea204 | 1511 | |
90481622 | 1512 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 1513 | |
a5516438 | 1514 | vma_commit_reservation(h, vma, addr); |
90d8b7e6 | 1515 | return page; |
8f34af6f JZ |
1516 | |
1517 | out_uncharge_cgroup: | |
1518 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
1519 | out_subpool_put: | |
1520 | if (chg || avoid_reserve) | |
1521 | hugepage_subpool_put_pages(spool, 1); | |
1522 | return ERR_PTR(-ENOSPC); | |
b45b5bd6 DG |
1523 | } |
1524 | ||
74060e4d NH |
1525 | /* |
1526 | * alloc_huge_page()'s wrapper which simply returns the page if allocation | |
1527 | * succeeds, otherwise NULL. This function is called from new_vma_page(), | |
1528 | * where no ERR_VALUE is expected to be returned. | |
1529 | */ | |
1530 | struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, | |
1531 | unsigned long addr, int avoid_reserve) | |
1532 | { | |
1533 | struct page *page = alloc_huge_page(vma, addr, avoid_reserve); | |
1534 | if (IS_ERR(page)) | |
1535 | page = NULL; | |
1536 | return page; | |
1537 | } | |
1538 | ||
91f47662 | 1539 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
aa888a74 AK |
1540 | { |
1541 | struct huge_bootmem_page *m; | |
b2261026 | 1542 | int nr_nodes, node; |
aa888a74 | 1543 | |
b2261026 | 1544 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
1545 | void *addr; |
1546 | ||
8b89a116 GS |
1547 | addr = memblock_virt_alloc_try_nid_nopanic( |
1548 | huge_page_size(h), huge_page_size(h), | |
1549 | 0, BOOTMEM_ALLOC_ACCESSIBLE, node); | |
aa888a74 AK |
1550 | if (addr) { |
1551 | /* | |
1552 | * Use the beginning of the huge page to store the | |
1553 | * huge_bootmem_page struct (until gather_bootmem | |
1554 | * puts them into the mem_map). | |
1555 | */ | |
1556 | m = addr; | |
91f47662 | 1557 | goto found; |
aa888a74 | 1558 | } |
aa888a74 AK |
1559 | } |
1560 | return 0; | |
1561 | ||
1562 | found: | |
df994ead | 1563 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 AK |
1564 | /* Put them into a private list first because mem_map is not up yet */ |
1565 | list_add(&m->list, &huge_boot_pages); | |
1566 | m->hstate = h; | |
1567 | return 1; | |
1568 | } | |
1569 | ||
f412c97a | 1570 | static void __init prep_compound_huge_page(struct page *page, int order) |
18229df5 AW |
1571 | { |
1572 | if (unlikely(order > (MAX_ORDER - 1))) | |
1573 | prep_compound_gigantic_page(page, order); | |
1574 | else | |
1575 | prep_compound_page(page, order); | |
1576 | } | |
1577 | ||
aa888a74 AK |
1578 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
1579 | static void __init gather_bootmem_prealloc(void) | |
1580 | { | |
1581 | struct huge_bootmem_page *m; | |
1582 | ||
1583 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 1584 | struct hstate *h = m->hstate; |
ee8f248d BB |
1585 | struct page *page; |
1586 | ||
1587 | #ifdef CONFIG_HIGHMEM | |
1588 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
8b89a116 GS |
1589 | memblock_free_late(__pa(m), |
1590 | sizeof(struct huge_bootmem_page)); | |
ee8f248d BB |
1591 | #else |
1592 | page = virt_to_page(m); | |
1593 | #endif | |
aa888a74 | 1594 | WARN_ON(page_count(page) != 1); |
18229df5 | 1595 | prep_compound_huge_page(page, h->order); |
ef5a22be | 1596 | WARN_ON(PageReserved(page)); |
aa888a74 | 1597 | prep_new_huge_page(h, page, page_to_nid(page)); |
b0320c7b RA |
1598 | /* |
1599 | * If we had gigantic hugepages allocated at boot time, we need | |
1600 | * to restore the 'stolen' pages to totalram_pages in order to | |
1601 | * fix confusing memory reports from free(1) and another | |
1602 | * side-effects, like CommitLimit going negative. | |
1603 | */ | |
bae7f4ae | 1604 | if (hstate_is_gigantic(h)) |
3dcc0571 | 1605 | adjust_managed_page_count(page, 1 << h->order); |
aa888a74 AK |
1606 | } |
1607 | } | |
1608 | ||
8faa8b07 | 1609 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
1610 | { |
1611 | unsigned long i; | |
a5516438 | 1612 | |
e5ff2159 | 1613 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 1614 | if (hstate_is_gigantic(h)) { |
aa888a74 AK |
1615 | if (!alloc_bootmem_huge_page(h)) |
1616 | break; | |
9b5e5d0f | 1617 | } else if (!alloc_fresh_huge_page(h, |
8cebfcd0 | 1618 | &node_states[N_MEMORY])) |
1da177e4 | 1619 | break; |
1da177e4 | 1620 | } |
8faa8b07 | 1621 | h->max_huge_pages = i; |
e5ff2159 AK |
1622 | } |
1623 | ||
1624 | static void __init hugetlb_init_hstates(void) | |
1625 | { | |
1626 | struct hstate *h; | |
1627 | ||
1628 | for_each_hstate(h) { | |
641844f5 NH |
1629 | if (minimum_order > huge_page_order(h)) |
1630 | minimum_order = huge_page_order(h); | |
1631 | ||
8faa8b07 | 1632 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 1633 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 1634 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 1635 | } |
641844f5 | 1636 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
1637 | } |
1638 | ||
4abd32db AK |
1639 | static char * __init memfmt(char *buf, unsigned long n) |
1640 | { | |
1641 | if (n >= (1UL << 30)) | |
1642 | sprintf(buf, "%lu GB", n >> 30); | |
1643 | else if (n >= (1UL << 20)) | |
1644 | sprintf(buf, "%lu MB", n >> 20); | |
1645 | else | |
1646 | sprintf(buf, "%lu KB", n >> 10); | |
1647 | return buf; | |
1648 | } | |
1649 | ||
e5ff2159 AK |
1650 | static void __init report_hugepages(void) |
1651 | { | |
1652 | struct hstate *h; | |
1653 | ||
1654 | for_each_hstate(h) { | |
4abd32db | 1655 | char buf[32]; |
ffb22af5 | 1656 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
4abd32db AK |
1657 | memfmt(buf, huge_page_size(h)), |
1658 | h->free_huge_pages); | |
e5ff2159 AK |
1659 | } |
1660 | } | |
1661 | ||
1da177e4 | 1662 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
1663 | static void try_to_free_low(struct hstate *h, unsigned long count, |
1664 | nodemask_t *nodes_allowed) | |
1da177e4 | 1665 | { |
4415cc8d CL |
1666 | int i; |
1667 | ||
bae7f4ae | 1668 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1669 | return; |
1670 | ||
6ae11b27 | 1671 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 1672 | struct page *page, *next; |
a5516438 AK |
1673 | struct list_head *freel = &h->hugepage_freelists[i]; |
1674 | list_for_each_entry_safe(page, next, freel, lru) { | |
1675 | if (count >= h->nr_huge_pages) | |
6b0c880d | 1676 | return; |
1da177e4 LT |
1677 | if (PageHighMem(page)) |
1678 | continue; | |
1679 | list_del(&page->lru); | |
e5ff2159 | 1680 | update_and_free_page(h, page); |
a5516438 AK |
1681 | h->free_huge_pages--; |
1682 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
1683 | } |
1684 | } | |
1685 | } | |
1686 | #else | |
6ae11b27 LS |
1687 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
1688 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
1689 | { |
1690 | } | |
1691 | #endif | |
1692 | ||
20a0307c WF |
1693 | /* |
1694 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
1695 | * balanced by operating on them in a round-robin fashion. | |
1696 | * Returns 1 if an adjustment was made. | |
1697 | */ | |
6ae11b27 LS |
1698 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
1699 | int delta) | |
20a0307c | 1700 | { |
b2261026 | 1701 | int nr_nodes, node; |
20a0307c WF |
1702 | |
1703 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 1704 | |
b2261026 JK |
1705 | if (delta < 0) { |
1706 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1707 | if (h->surplus_huge_pages_node[node]) | |
1708 | goto found; | |
e8c5c824 | 1709 | } |
b2261026 JK |
1710 | } else { |
1711 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
1712 | if (h->surplus_huge_pages_node[node] < | |
1713 | h->nr_huge_pages_node[node]) | |
1714 | goto found; | |
e8c5c824 | 1715 | } |
b2261026 JK |
1716 | } |
1717 | return 0; | |
20a0307c | 1718 | |
b2261026 JK |
1719 | found: |
1720 | h->surplus_huge_pages += delta; | |
1721 | h->surplus_huge_pages_node[node] += delta; | |
1722 | return 1; | |
20a0307c WF |
1723 | } |
1724 | ||
a5516438 | 1725 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
1726 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
1727 | nodemask_t *nodes_allowed) | |
1da177e4 | 1728 | { |
7893d1d5 | 1729 | unsigned long min_count, ret; |
1da177e4 | 1730 | |
944d9fec | 1731 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
aa888a74 AK |
1732 | return h->max_huge_pages; |
1733 | ||
7893d1d5 AL |
1734 | /* |
1735 | * Increase the pool size | |
1736 | * First take pages out of surplus state. Then make up the | |
1737 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f NA |
1738 | * |
1739 | * We might race with alloc_buddy_huge_page() here and be unable | |
1740 | * to convert a surplus huge page to a normal huge page. That is | |
1741 | * not critical, though, it just means the overall size of the | |
1742 | * pool might be one hugepage larger than it needs to be, but | |
1743 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 1744 | */ |
1da177e4 | 1745 | spin_lock(&hugetlb_lock); |
a5516438 | 1746 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 1747 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
1748 | break; |
1749 | } | |
1750 | ||
a5516438 | 1751 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
1752 | /* |
1753 | * If this allocation races such that we no longer need the | |
1754 | * page, free_huge_page will handle it by freeing the page | |
1755 | * and reducing the surplus. | |
1756 | */ | |
1757 | spin_unlock(&hugetlb_lock); | |
944d9fec LC |
1758 | if (hstate_is_gigantic(h)) |
1759 | ret = alloc_fresh_gigantic_page(h, nodes_allowed); | |
1760 | else | |
1761 | ret = alloc_fresh_huge_page(h, nodes_allowed); | |
7893d1d5 AL |
1762 | spin_lock(&hugetlb_lock); |
1763 | if (!ret) | |
1764 | goto out; | |
1765 | ||
536240f2 MG |
1766 | /* Bail for signals. Probably ctrl-c from user */ |
1767 | if (signal_pending(current)) | |
1768 | goto out; | |
7893d1d5 | 1769 | } |
7893d1d5 AL |
1770 | |
1771 | /* | |
1772 | * Decrease the pool size | |
1773 | * First return free pages to the buddy allocator (being careful | |
1774 | * to keep enough around to satisfy reservations). Then place | |
1775 | * pages into surplus state as needed so the pool will shrink | |
1776 | * to the desired size as pages become free. | |
d1c3fb1f NA |
1777 | * |
1778 | * By placing pages into the surplus state independent of the | |
1779 | * overcommit value, we are allowing the surplus pool size to | |
1780 | * exceed overcommit. There are few sane options here. Since | |
1781 | * alloc_buddy_huge_page() is checking the global counter, | |
1782 | * though, we'll note that we're not allowed to exceed surplus | |
1783 | * and won't grow the pool anywhere else. Not until one of the | |
1784 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 1785 | */ |
a5516438 | 1786 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 1787 | min_count = max(count, min_count); |
6ae11b27 | 1788 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 1789 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 1790 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 1791 | break; |
55f67141 | 1792 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 1793 | } |
a5516438 | 1794 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 1795 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
1796 | break; |
1797 | } | |
1798 | out: | |
a5516438 | 1799 | ret = persistent_huge_pages(h); |
1da177e4 | 1800 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 1801 | return ret; |
1da177e4 LT |
1802 | } |
1803 | ||
a3437870 NA |
1804 | #define HSTATE_ATTR_RO(_name) \ |
1805 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
1806 | ||
1807 | #define HSTATE_ATTR(_name) \ | |
1808 | static struct kobj_attribute _name##_attr = \ | |
1809 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
1810 | ||
1811 | static struct kobject *hugepages_kobj; | |
1812 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1813 | ||
9a305230 LS |
1814 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
1815 | ||
1816 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
1817 | { |
1818 | int i; | |
9a305230 | 1819 | |
a3437870 | 1820 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
1821 | if (hstate_kobjs[i] == kobj) { |
1822 | if (nidp) | |
1823 | *nidp = NUMA_NO_NODE; | |
a3437870 | 1824 | return &hstates[i]; |
9a305230 LS |
1825 | } |
1826 | ||
1827 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
1828 | } |
1829 | ||
06808b08 | 1830 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
1831 | struct kobj_attribute *attr, char *buf) |
1832 | { | |
9a305230 LS |
1833 | struct hstate *h; |
1834 | unsigned long nr_huge_pages; | |
1835 | int nid; | |
1836 | ||
1837 | h = kobj_to_hstate(kobj, &nid); | |
1838 | if (nid == NUMA_NO_NODE) | |
1839 | nr_huge_pages = h->nr_huge_pages; | |
1840 | else | |
1841 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
1842 | ||
1843 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 1844 | } |
adbe8726 | 1845 | |
238d3c13 DR |
1846 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
1847 | struct hstate *h, int nid, | |
1848 | unsigned long count, size_t len) | |
a3437870 NA |
1849 | { |
1850 | int err; | |
bad44b5b | 1851 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 1852 | |
944d9fec | 1853 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) { |
adbe8726 EM |
1854 | err = -EINVAL; |
1855 | goto out; | |
1856 | } | |
1857 | ||
9a305230 LS |
1858 | if (nid == NUMA_NO_NODE) { |
1859 | /* | |
1860 | * global hstate attribute | |
1861 | */ | |
1862 | if (!(obey_mempolicy && | |
1863 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
1864 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 1865 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 LS |
1866 | } |
1867 | } else if (nodes_allowed) { | |
1868 | /* | |
1869 | * per node hstate attribute: adjust count to global, | |
1870 | * but restrict alloc/free to the specified node. | |
1871 | */ | |
1872 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
1873 | init_nodemask_of_node(nodes_allowed, nid); | |
1874 | } else | |
8cebfcd0 | 1875 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 | 1876 | |
06808b08 | 1877 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 1878 | |
8cebfcd0 | 1879 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
1880 | NODEMASK_FREE(nodes_allowed); |
1881 | ||
1882 | return len; | |
adbe8726 EM |
1883 | out: |
1884 | NODEMASK_FREE(nodes_allowed); | |
1885 | return err; | |
06808b08 LS |
1886 | } |
1887 | ||
238d3c13 DR |
1888 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
1889 | struct kobject *kobj, const char *buf, | |
1890 | size_t len) | |
1891 | { | |
1892 | struct hstate *h; | |
1893 | unsigned long count; | |
1894 | int nid; | |
1895 | int err; | |
1896 | ||
1897 | err = kstrtoul(buf, 10, &count); | |
1898 | if (err) | |
1899 | return err; | |
1900 | ||
1901 | h = kobj_to_hstate(kobj, &nid); | |
1902 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
1903 | } | |
1904 | ||
06808b08 LS |
1905 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
1906 | struct kobj_attribute *attr, char *buf) | |
1907 | { | |
1908 | return nr_hugepages_show_common(kobj, attr, buf); | |
1909 | } | |
1910 | ||
1911 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
1912 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1913 | { | |
238d3c13 | 1914 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
1915 | } |
1916 | HSTATE_ATTR(nr_hugepages); | |
1917 | ||
06808b08 LS |
1918 | #ifdef CONFIG_NUMA |
1919 | ||
1920 | /* | |
1921 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
1922 | * huge page alloc/free. | |
1923 | */ | |
1924 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
1925 | struct kobj_attribute *attr, char *buf) | |
1926 | { | |
1927 | return nr_hugepages_show_common(kobj, attr, buf); | |
1928 | } | |
1929 | ||
1930 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
1931 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1932 | { | |
238d3c13 | 1933 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
1934 | } |
1935 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
1936 | #endif | |
1937 | ||
1938 | ||
a3437870 NA |
1939 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
1940 | struct kobj_attribute *attr, char *buf) | |
1941 | { | |
9a305230 | 1942 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1943 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
1944 | } | |
adbe8726 | 1945 | |
a3437870 NA |
1946 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
1947 | struct kobj_attribute *attr, const char *buf, size_t count) | |
1948 | { | |
1949 | int err; | |
1950 | unsigned long input; | |
9a305230 | 1951 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 1952 | |
bae7f4ae | 1953 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
1954 | return -EINVAL; |
1955 | ||
3dbb95f7 | 1956 | err = kstrtoul(buf, 10, &input); |
a3437870 | 1957 | if (err) |
73ae31e5 | 1958 | return err; |
a3437870 NA |
1959 | |
1960 | spin_lock(&hugetlb_lock); | |
1961 | h->nr_overcommit_huge_pages = input; | |
1962 | spin_unlock(&hugetlb_lock); | |
1963 | ||
1964 | return count; | |
1965 | } | |
1966 | HSTATE_ATTR(nr_overcommit_hugepages); | |
1967 | ||
1968 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
1969 | struct kobj_attribute *attr, char *buf) | |
1970 | { | |
9a305230 LS |
1971 | struct hstate *h; |
1972 | unsigned long free_huge_pages; | |
1973 | int nid; | |
1974 | ||
1975 | h = kobj_to_hstate(kobj, &nid); | |
1976 | if (nid == NUMA_NO_NODE) | |
1977 | free_huge_pages = h->free_huge_pages; | |
1978 | else | |
1979 | free_huge_pages = h->free_huge_pages_node[nid]; | |
1980 | ||
1981 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
1982 | } |
1983 | HSTATE_ATTR_RO(free_hugepages); | |
1984 | ||
1985 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
1986 | struct kobj_attribute *attr, char *buf) | |
1987 | { | |
9a305230 | 1988 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
1989 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
1990 | } | |
1991 | HSTATE_ATTR_RO(resv_hugepages); | |
1992 | ||
1993 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
1994 | struct kobj_attribute *attr, char *buf) | |
1995 | { | |
9a305230 LS |
1996 | struct hstate *h; |
1997 | unsigned long surplus_huge_pages; | |
1998 | int nid; | |
1999 | ||
2000 | h = kobj_to_hstate(kobj, &nid); | |
2001 | if (nid == NUMA_NO_NODE) | |
2002 | surplus_huge_pages = h->surplus_huge_pages; | |
2003 | else | |
2004 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2005 | ||
2006 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
2007 | } |
2008 | HSTATE_ATTR_RO(surplus_hugepages); | |
2009 | ||
2010 | static struct attribute *hstate_attrs[] = { | |
2011 | &nr_hugepages_attr.attr, | |
2012 | &nr_overcommit_hugepages_attr.attr, | |
2013 | &free_hugepages_attr.attr, | |
2014 | &resv_hugepages_attr.attr, | |
2015 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
2016 | #ifdef CONFIG_NUMA |
2017 | &nr_hugepages_mempolicy_attr.attr, | |
2018 | #endif | |
a3437870 NA |
2019 | NULL, |
2020 | }; | |
2021 | ||
2022 | static struct attribute_group hstate_attr_group = { | |
2023 | .attrs = hstate_attrs, | |
2024 | }; | |
2025 | ||
094e9539 JM |
2026 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
2027 | struct kobject **hstate_kobjs, | |
2028 | struct attribute_group *hstate_attr_group) | |
a3437870 NA |
2029 | { |
2030 | int retval; | |
972dc4de | 2031 | int hi = hstate_index(h); |
a3437870 | 2032 | |
9a305230 LS |
2033 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
2034 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
2035 | return -ENOMEM; |
2036 | ||
9a305230 | 2037 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 2038 | if (retval) |
9a305230 | 2039 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
2040 | |
2041 | return retval; | |
2042 | } | |
2043 | ||
2044 | static void __init hugetlb_sysfs_init(void) | |
2045 | { | |
2046 | struct hstate *h; | |
2047 | int err; | |
2048 | ||
2049 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
2050 | if (!hugepages_kobj) | |
2051 | return; | |
2052 | ||
2053 | for_each_hstate(h) { | |
9a305230 LS |
2054 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
2055 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 2056 | if (err) |
ffb22af5 | 2057 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
2058 | } |
2059 | } | |
2060 | ||
9a305230 LS |
2061 | #ifdef CONFIG_NUMA |
2062 | ||
2063 | /* | |
2064 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
2065 | * with node devices in node_devices[] using a parallel array. The array |
2066 | * index of a node device or _hstate == node id. | |
2067 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
2068 | * the base kernel, on the hugetlb module. |
2069 | */ | |
2070 | struct node_hstate { | |
2071 | struct kobject *hugepages_kobj; | |
2072 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2073 | }; | |
2074 | struct node_hstate node_hstates[MAX_NUMNODES]; | |
2075 | ||
2076 | /* | |
10fbcf4c | 2077 | * A subset of global hstate attributes for node devices |
9a305230 LS |
2078 | */ |
2079 | static struct attribute *per_node_hstate_attrs[] = { | |
2080 | &nr_hugepages_attr.attr, | |
2081 | &free_hugepages_attr.attr, | |
2082 | &surplus_hugepages_attr.attr, | |
2083 | NULL, | |
2084 | }; | |
2085 | ||
2086 | static struct attribute_group per_node_hstate_attr_group = { | |
2087 | .attrs = per_node_hstate_attrs, | |
2088 | }; | |
2089 | ||
2090 | /* | |
10fbcf4c | 2091 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
2092 | * Returns node id via non-NULL nidp. |
2093 | */ | |
2094 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2095 | { | |
2096 | int nid; | |
2097 | ||
2098 | for (nid = 0; nid < nr_node_ids; nid++) { | |
2099 | struct node_hstate *nhs = &node_hstates[nid]; | |
2100 | int i; | |
2101 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
2102 | if (nhs->hstate_kobjs[i] == kobj) { | |
2103 | if (nidp) | |
2104 | *nidp = nid; | |
2105 | return &hstates[i]; | |
2106 | } | |
2107 | } | |
2108 | ||
2109 | BUG(); | |
2110 | return NULL; | |
2111 | } | |
2112 | ||
2113 | /* | |
10fbcf4c | 2114 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
2115 | * No-op if no hstate attributes attached. |
2116 | */ | |
3cd8b44f | 2117 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
2118 | { |
2119 | struct hstate *h; | |
10fbcf4c | 2120 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2121 | |
2122 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 2123 | return; /* no hstate attributes */ |
9a305230 | 2124 | |
972dc4de AK |
2125 | for_each_hstate(h) { |
2126 | int idx = hstate_index(h); | |
2127 | if (nhs->hstate_kobjs[idx]) { | |
2128 | kobject_put(nhs->hstate_kobjs[idx]); | |
2129 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 2130 | } |
972dc4de | 2131 | } |
9a305230 LS |
2132 | |
2133 | kobject_put(nhs->hugepages_kobj); | |
2134 | nhs->hugepages_kobj = NULL; | |
2135 | } | |
2136 | ||
2137 | /* | |
10fbcf4c | 2138 | * hugetlb module exit: unregister hstate attributes from node devices |
9a305230 LS |
2139 | * that have them. |
2140 | */ | |
2141 | static void hugetlb_unregister_all_nodes(void) | |
2142 | { | |
2143 | int nid; | |
2144 | ||
2145 | /* | |
10fbcf4c | 2146 | * disable node device registrations. |
9a305230 LS |
2147 | */ |
2148 | register_hugetlbfs_with_node(NULL, NULL); | |
2149 | ||
2150 | /* | |
2151 | * remove hstate attributes from any nodes that have them. | |
2152 | */ | |
2153 | for (nid = 0; nid < nr_node_ids; nid++) | |
8732794b | 2154 | hugetlb_unregister_node(node_devices[nid]); |
9a305230 LS |
2155 | } |
2156 | ||
2157 | /* | |
10fbcf4c | 2158 | * Register hstate attributes for a single node device. |
9a305230 LS |
2159 | * No-op if attributes already registered. |
2160 | */ | |
3cd8b44f | 2161 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
2162 | { |
2163 | struct hstate *h; | |
10fbcf4c | 2164 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2165 | int err; |
2166 | ||
2167 | if (nhs->hugepages_kobj) | |
2168 | return; /* already allocated */ | |
2169 | ||
2170 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 2171 | &node->dev.kobj); |
9a305230 LS |
2172 | if (!nhs->hugepages_kobj) |
2173 | return; | |
2174 | ||
2175 | for_each_hstate(h) { | |
2176 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
2177 | nhs->hstate_kobjs, | |
2178 | &per_node_hstate_attr_group); | |
2179 | if (err) { | |
ffb22af5 AM |
2180 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
2181 | h->name, node->dev.id); | |
9a305230 LS |
2182 | hugetlb_unregister_node(node); |
2183 | break; | |
2184 | } | |
2185 | } | |
2186 | } | |
2187 | ||
2188 | /* | |
9b5e5d0f | 2189 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
2190 | * devices of nodes that have memory. All on-line nodes should have |
2191 | * registered their associated device by this time. | |
9a305230 | 2192 | */ |
7d9ca000 | 2193 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
2194 | { |
2195 | int nid; | |
2196 | ||
8cebfcd0 | 2197 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 2198 | struct node *node = node_devices[nid]; |
10fbcf4c | 2199 | if (node->dev.id == nid) |
9a305230 LS |
2200 | hugetlb_register_node(node); |
2201 | } | |
2202 | ||
2203 | /* | |
10fbcf4c | 2204 | * Let the node device driver know we're here so it can |
9a305230 LS |
2205 | * [un]register hstate attributes on node hotplug. |
2206 | */ | |
2207 | register_hugetlbfs_with_node(hugetlb_register_node, | |
2208 | hugetlb_unregister_node); | |
2209 | } | |
2210 | #else /* !CONFIG_NUMA */ | |
2211 | ||
2212 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2213 | { | |
2214 | BUG(); | |
2215 | if (nidp) | |
2216 | *nidp = -1; | |
2217 | return NULL; | |
2218 | } | |
2219 | ||
2220 | static void hugetlb_unregister_all_nodes(void) { } | |
2221 | ||
2222 | static void hugetlb_register_all_nodes(void) { } | |
2223 | ||
2224 | #endif | |
2225 | ||
a3437870 NA |
2226 | static void __exit hugetlb_exit(void) |
2227 | { | |
2228 | struct hstate *h; | |
2229 | ||
9a305230 LS |
2230 | hugetlb_unregister_all_nodes(); |
2231 | ||
a3437870 | 2232 | for_each_hstate(h) { |
972dc4de | 2233 | kobject_put(hstate_kobjs[hstate_index(h)]); |
a3437870 NA |
2234 | } |
2235 | ||
2236 | kobject_put(hugepages_kobj); | |
8382d914 | 2237 | kfree(htlb_fault_mutex_table); |
a3437870 NA |
2238 | } |
2239 | module_exit(hugetlb_exit); | |
2240 | ||
2241 | static int __init hugetlb_init(void) | |
2242 | { | |
8382d914 DB |
2243 | int i; |
2244 | ||
457c1b27 | 2245 | if (!hugepages_supported()) |
0ef89d25 | 2246 | return 0; |
a3437870 | 2247 | |
e11bfbfc NP |
2248 | if (!size_to_hstate(default_hstate_size)) { |
2249 | default_hstate_size = HPAGE_SIZE; | |
2250 | if (!size_to_hstate(default_hstate_size)) | |
2251 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 2252 | } |
972dc4de | 2253 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
e11bfbfc NP |
2254 | if (default_hstate_max_huge_pages) |
2255 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
a3437870 NA |
2256 | |
2257 | hugetlb_init_hstates(); | |
aa888a74 | 2258 | gather_bootmem_prealloc(); |
a3437870 NA |
2259 | report_hugepages(); |
2260 | ||
2261 | hugetlb_sysfs_init(); | |
9a305230 | 2262 | hugetlb_register_all_nodes(); |
7179e7bf | 2263 | hugetlb_cgroup_file_init(); |
9a305230 | 2264 | |
8382d914 DB |
2265 | #ifdef CONFIG_SMP |
2266 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
2267 | #else | |
2268 | num_fault_mutexes = 1; | |
2269 | #endif | |
2270 | htlb_fault_mutex_table = | |
2271 | kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); | |
2272 | BUG_ON(!htlb_fault_mutex_table); | |
2273 | ||
2274 | for (i = 0; i < num_fault_mutexes; i++) | |
2275 | mutex_init(&htlb_fault_mutex_table[i]); | |
a3437870 NA |
2276 | return 0; |
2277 | } | |
2278 | module_init(hugetlb_init); | |
2279 | ||
2280 | /* Should be called on processing a hugepagesz=... option */ | |
2281 | void __init hugetlb_add_hstate(unsigned order) | |
2282 | { | |
2283 | struct hstate *h; | |
8faa8b07 AK |
2284 | unsigned long i; |
2285 | ||
a3437870 | 2286 | if (size_to_hstate(PAGE_SIZE << order)) { |
ffb22af5 | 2287 | pr_warning("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
2288 | return; |
2289 | } | |
47d38344 | 2290 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 2291 | BUG_ON(order == 0); |
47d38344 | 2292 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
2293 | h->order = order; |
2294 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
2295 | h->nr_huge_pages = 0; |
2296 | h->free_huge_pages = 0; | |
2297 | for (i = 0; i < MAX_NUMNODES; ++i) | |
2298 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 2299 | INIT_LIST_HEAD(&h->hugepage_activelist); |
8cebfcd0 LJ |
2300 | h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); |
2301 | h->next_nid_to_free = first_node(node_states[N_MEMORY]); | |
a3437870 NA |
2302 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
2303 | huge_page_size(h)/1024); | |
8faa8b07 | 2304 | |
a3437870 NA |
2305 | parsed_hstate = h; |
2306 | } | |
2307 | ||
e11bfbfc | 2308 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
2309 | { |
2310 | unsigned long *mhp; | |
8faa8b07 | 2311 | static unsigned long *last_mhp; |
a3437870 NA |
2312 | |
2313 | /* | |
47d38344 | 2314 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
2315 | * so this hugepages= parameter goes to the "default hstate". |
2316 | */ | |
47d38344 | 2317 | if (!hugetlb_max_hstate) |
a3437870 NA |
2318 | mhp = &default_hstate_max_huge_pages; |
2319 | else | |
2320 | mhp = &parsed_hstate->max_huge_pages; | |
2321 | ||
8faa8b07 | 2322 | if (mhp == last_mhp) { |
ffb22af5 AM |
2323 | pr_warning("hugepages= specified twice without " |
2324 | "interleaving hugepagesz=, ignoring\n"); | |
8faa8b07 AK |
2325 | return 1; |
2326 | } | |
2327 | ||
a3437870 NA |
2328 | if (sscanf(s, "%lu", mhp) <= 0) |
2329 | *mhp = 0; | |
2330 | ||
8faa8b07 AK |
2331 | /* |
2332 | * Global state is always initialized later in hugetlb_init. | |
2333 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
2334 | * use the bootmem allocator. | |
2335 | */ | |
47d38344 | 2336 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
2337 | hugetlb_hstate_alloc_pages(parsed_hstate); |
2338 | ||
2339 | last_mhp = mhp; | |
2340 | ||
a3437870 NA |
2341 | return 1; |
2342 | } | |
e11bfbfc NP |
2343 | __setup("hugepages=", hugetlb_nrpages_setup); |
2344 | ||
2345 | static int __init hugetlb_default_setup(char *s) | |
2346 | { | |
2347 | default_hstate_size = memparse(s, &s); | |
2348 | return 1; | |
2349 | } | |
2350 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 2351 | |
8a213460 NA |
2352 | static unsigned int cpuset_mems_nr(unsigned int *array) |
2353 | { | |
2354 | int node; | |
2355 | unsigned int nr = 0; | |
2356 | ||
2357 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
2358 | nr += array[node]; | |
2359 | ||
2360 | return nr; | |
2361 | } | |
2362 | ||
2363 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
2364 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
2365 | struct ctl_table *table, int write, | |
2366 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 2367 | { |
e5ff2159 | 2368 | struct hstate *h = &default_hstate; |
238d3c13 | 2369 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 2370 | int ret; |
e5ff2159 | 2371 | |
457c1b27 NA |
2372 | if (!hugepages_supported()) |
2373 | return -ENOTSUPP; | |
2374 | ||
e5ff2159 AK |
2375 | table->data = &tmp; |
2376 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2377 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2378 | if (ret) | |
2379 | goto out; | |
e5ff2159 | 2380 | |
238d3c13 DR |
2381 | if (write) |
2382 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
2383 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
2384 | out: |
2385 | return ret; | |
1da177e4 | 2386 | } |
396faf03 | 2387 | |
06808b08 LS |
2388 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2389 | void __user *buffer, size_t *length, loff_t *ppos) | |
2390 | { | |
2391 | ||
2392 | return hugetlb_sysctl_handler_common(false, table, write, | |
2393 | buffer, length, ppos); | |
2394 | } | |
2395 | ||
2396 | #ifdef CONFIG_NUMA | |
2397 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2398 | void __user *buffer, size_t *length, loff_t *ppos) | |
2399 | { | |
2400 | return hugetlb_sysctl_handler_common(true, table, write, | |
2401 | buffer, length, ppos); | |
2402 | } | |
2403 | #endif /* CONFIG_NUMA */ | |
2404 | ||
a3d0c6aa | 2405 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 2406 | void __user *buffer, |
a3d0c6aa NA |
2407 | size_t *length, loff_t *ppos) |
2408 | { | |
a5516438 | 2409 | struct hstate *h = &default_hstate; |
e5ff2159 | 2410 | unsigned long tmp; |
08d4a246 | 2411 | int ret; |
e5ff2159 | 2412 | |
457c1b27 NA |
2413 | if (!hugepages_supported()) |
2414 | return -ENOTSUPP; | |
2415 | ||
c033a93c | 2416 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 2417 | |
bae7f4ae | 2418 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
2419 | return -EINVAL; |
2420 | ||
e5ff2159 AK |
2421 | table->data = &tmp; |
2422 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2423 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2424 | if (ret) | |
2425 | goto out; | |
e5ff2159 AK |
2426 | |
2427 | if (write) { | |
2428 | spin_lock(&hugetlb_lock); | |
2429 | h->nr_overcommit_huge_pages = tmp; | |
2430 | spin_unlock(&hugetlb_lock); | |
2431 | } | |
08d4a246 MH |
2432 | out: |
2433 | return ret; | |
a3d0c6aa NA |
2434 | } |
2435 | ||
1da177e4 LT |
2436 | #endif /* CONFIG_SYSCTL */ |
2437 | ||
e1759c21 | 2438 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 2439 | { |
a5516438 | 2440 | struct hstate *h = &default_hstate; |
457c1b27 NA |
2441 | if (!hugepages_supported()) |
2442 | return; | |
e1759c21 | 2443 | seq_printf(m, |
4f98a2fe RR |
2444 | "HugePages_Total: %5lu\n" |
2445 | "HugePages_Free: %5lu\n" | |
2446 | "HugePages_Rsvd: %5lu\n" | |
2447 | "HugePages_Surp: %5lu\n" | |
2448 | "Hugepagesize: %8lu kB\n", | |
a5516438 AK |
2449 | h->nr_huge_pages, |
2450 | h->free_huge_pages, | |
2451 | h->resv_huge_pages, | |
2452 | h->surplus_huge_pages, | |
2453 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
2454 | } |
2455 | ||
2456 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
2457 | { | |
a5516438 | 2458 | struct hstate *h = &default_hstate; |
457c1b27 NA |
2459 | if (!hugepages_supported()) |
2460 | return 0; | |
1da177e4 LT |
2461 | return sprintf(buf, |
2462 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
2463 | "Node %d HugePages_Free: %5u\n" |
2464 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
2465 | nid, h->nr_huge_pages_node[nid], |
2466 | nid, h->free_huge_pages_node[nid], | |
2467 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
2468 | } |
2469 | ||
949f7ec5 DR |
2470 | void hugetlb_show_meminfo(void) |
2471 | { | |
2472 | struct hstate *h; | |
2473 | int nid; | |
2474 | ||
457c1b27 NA |
2475 | if (!hugepages_supported()) |
2476 | return; | |
2477 | ||
949f7ec5 DR |
2478 | for_each_node_state(nid, N_MEMORY) |
2479 | for_each_hstate(h) | |
2480 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
2481 | nid, | |
2482 | h->nr_huge_pages_node[nid], | |
2483 | h->free_huge_pages_node[nid], | |
2484 | h->surplus_huge_pages_node[nid], | |
2485 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
2486 | } | |
2487 | ||
1da177e4 LT |
2488 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
2489 | unsigned long hugetlb_total_pages(void) | |
2490 | { | |
d0028588 WL |
2491 | struct hstate *h; |
2492 | unsigned long nr_total_pages = 0; | |
2493 | ||
2494 | for_each_hstate(h) | |
2495 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
2496 | return nr_total_pages; | |
1da177e4 | 2497 | } |
1da177e4 | 2498 | |
a5516438 | 2499 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
2500 | { |
2501 | int ret = -ENOMEM; | |
2502 | ||
2503 | spin_lock(&hugetlb_lock); | |
2504 | /* | |
2505 | * When cpuset is configured, it breaks the strict hugetlb page | |
2506 | * reservation as the accounting is done on a global variable. Such | |
2507 | * reservation is completely rubbish in the presence of cpuset because | |
2508 | * the reservation is not checked against page availability for the | |
2509 | * current cpuset. Application can still potentially OOM'ed by kernel | |
2510 | * with lack of free htlb page in cpuset that the task is in. | |
2511 | * Attempt to enforce strict accounting with cpuset is almost | |
2512 | * impossible (or too ugly) because cpuset is too fluid that | |
2513 | * task or memory node can be dynamically moved between cpusets. | |
2514 | * | |
2515 | * The change of semantics for shared hugetlb mapping with cpuset is | |
2516 | * undesirable. However, in order to preserve some of the semantics, | |
2517 | * we fall back to check against current free page availability as | |
2518 | * a best attempt and hopefully to minimize the impact of changing | |
2519 | * semantics that cpuset has. | |
2520 | */ | |
2521 | if (delta > 0) { | |
a5516438 | 2522 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
2523 | goto out; |
2524 | ||
a5516438 AK |
2525 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
2526 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
2527 | goto out; |
2528 | } | |
2529 | } | |
2530 | ||
2531 | ret = 0; | |
2532 | if (delta < 0) | |
a5516438 | 2533 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
2534 | |
2535 | out: | |
2536 | spin_unlock(&hugetlb_lock); | |
2537 | return ret; | |
2538 | } | |
2539 | ||
84afd99b AW |
2540 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
2541 | { | |
f522c3ac | 2542 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
2543 | |
2544 | /* | |
2545 | * This new VMA should share its siblings reservation map if present. | |
2546 | * The VMA will only ever have a valid reservation map pointer where | |
2547 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 2548 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
2549 | * after this open call completes. It is therefore safe to take a |
2550 | * new reference here without additional locking. | |
2551 | */ | |
4e35f483 | 2552 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 2553 | kref_get(&resv->refs); |
84afd99b AW |
2554 | } |
2555 | ||
a1e78772 MG |
2556 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
2557 | { | |
a5516438 | 2558 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 2559 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 2560 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 2561 | unsigned long reserve, start, end; |
1c5ecae3 | 2562 | long gbl_reserve; |
84afd99b | 2563 | |
4e35f483 JK |
2564 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
2565 | return; | |
84afd99b | 2566 | |
4e35f483 JK |
2567 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
2568 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 2569 | |
4e35f483 | 2570 | reserve = (end - start) - region_count(resv, start, end); |
84afd99b | 2571 | |
4e35f483 JK |
2572 | kref_put(&resv->refs, resv_map_release); |
2573 | ||
2574 | if (reserve) { | |
1c5ecae3 MK |
2575 | /* |
2576 | * Decrement reserve counts. The global reserve count may be | |
2577 | * adjusted if the subpool has a minimum size. | |
2578 | */ | |
2579 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
2580 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 2581 | } |
a1e78772 MG |
2582 | } |
2583 | ||
1da177e4 LT |
2584 | /* |
2585 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
2586 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
2587 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
2588 | * this far. | |
2589 | */ | |
d0217ac0 | 2590 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
2591 | { |
2592 | BUG(); | |
d0217ac0 | 2593 | return 0; |
1da177e4 LT |
2594 | } |
2595 | ||
f0f37e2f | 2596 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 2597 | .fault = hugetlb_vm_op_fault, |
84afd99b | 2598 | .open = hugetlb_vm_op_open, |
a1e78772 | 2599 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
2600 | }; |
2601 | ||
1e8f889b DG |
2602 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
2603 | int writable) | |
63551ae0 DG |
2604 | { |
2605 | pte_t entry; | |
2606 | ||
1e8f889b | 2607 | if (writable) { |
106c992a GS |
2608 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
2609 | vma->vm_page_prot))); | |
63551ae0 | 2610 | } else { |
106c992a GS |
2611 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
2612 | vma->vm_page_prot)); | |
63551ae0 DG |
2613 | } |
2614 | entry = pte_mkyoung(entry); | |
2615 | entry = pte_mkhuge(entry); | |
d9ed9faa | 2616 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
2617 | |
2618 | return entry; | |
2619 | } | |
2620 | ||
1e8f889b DG |
2621 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
2622 | unsigned long address, pte_t *ptep) | |
2623 | { | |
2624 | pte_t entry; | |
2625 | ||
106c992a | 2626 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 2627 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 2628 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
2629 | } |
2630 | ||
4a705fef NH |
2631 | static int is_hugetlb_entry_migration(pte_t pte) |
2632 | { | |
2633 | swp_entry_t swp; | |
2634 | ||
2635 | if (huge_pte_none(pte) || pte_present(pte)) | |
2636 | return 0; | |
2637 | swp = pte_to_swp_entry(pte); | |
2638 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
2639 | return 1; | |
2640 | else | |
2641 | return 0; | |
2642 | } | |
2643 | ||
2644 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
2645 | { | |
2646 | swp_entry_t swp; | |
2647 | ||
2648 | if (huge_pte_none(pte) || pte_present(pte)) | |
2649 | return 0; | |
2650 | swp = pte_to_swp_entry(pte); | |
2651 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
2652 | return 1; | |
2653 | else | |
2654 | return 0; | |
2655 | } | |
1e8f889b | 2656 | |
63551ae0 DG |
2657 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
2658 | struct vm_area_struct *vma) | |
2659 | { | |
2660 | pte_t *src_pte, *dst_pte, entry; | |
2661 | struct page *ptepage; | |
1c59827d | 2662 | unsigned long addr; |
1e8f889b | 2663 | int cow; |
a5516438 AK |
2664 | struct hstate *h = hstate_vma(vma); |
2665 | unsigned long sz = huge_page_size(h); | |
e8569dd2 AS |
2666 | unsigned long mmun_start; /* For mmu_notifiers */ |
2667 | unsigned long mmun_end; /* For mmu_notifiers */ | |
2668 | int ret = 0; | |
1e8f889b DG |
2669 | |
2670 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 2671 | |
e8569dd2 AS |
2672 | mmun_start = vma->vm_start; |
2673 | mmun_end = vma->vm_end; | |
2674 | if (cow) | |
2675 | mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); | |
2676 | ||
a5516438 | 2677 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 2678 | spinlock_t *src_ptl, *dst_ptl; |
c74df32c HD |
2679 | src_pte = huge_pte_offset(src, addr); |
2680 | if (!src_pte) | |
2681 | continue; | |
a5516438 | 2682 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
2683 | if (!dst_pte) { |
2684 | ret = -ENOMEM; | |
2685 | break; | |
2686 | } | |
c5c99429 LW |
2687 | |
2688 | /* If the pagetables are shared don't copy or take references */ | |
2689 | if (dst_pte == src_pte) | |
2690 | continue; | |
2691 | ||
cb900f41 KS |
2692 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
2693 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
2694 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef NH |
2695 | entry = huge_ptep_get(src_pte); |
2696 | if (huge_pte_none(entry)) { /* skip none entry */ | |
2697 | ; | |
2698 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
2699 | is_hugetlb_entry_hwpoisoned(entry))) { | |
2700 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
2701 | ||
2702 | if (is_write_migration_entry(swp_entry) && cow) { | |
2703 | /* | |
2704 | * COW mappings require pages in both | |
2705 | * parent and child to be set to read. | |
2706 | */ | |
2707 | make_migration_entry_read(&swp_entry); | |
2708 | entry = swp_entry_to_pte(swp_entry); | |
2709 | set_huge_pte_at(src, addr, src_pte, entry); | |
2710 | } | |
2711 | set_huge_pte_at(dst, addr, dst_pte, entry); | |
2712 | } else { | |
34ee645e | 2713 | if (cow) { |
7f2e9525 | 2714 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e JR |
2715 | mmu_notifier_invalidate_range(src, mmun_start, |
2716 | mmun_end); | |
2717 | } | |
0253d634 | 2718 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
2719 | ptepage = pte_page(entry); |
2720 | get_page(ptepage); | |
0fe6e20b | 2721 | page_dup_rmap(ptepage); |
1c59827d HD |
2722 | set_huge_pte_at(dst, addr, dst_pte, entry); |
2723 | } | |
cb900f41 KS |
2724 | spin_unlock(src_ptl); |
2725 | spin_unlock(dst_ptl); | |
63551ae0 | 2726 | } |
63551ae0 | 2727 | |
e8569dd2 AS |
2728 | if (cow) |
2729 | mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); | |
2730 | ||
2731 | return ret; | |
63551ae0 DG |
2732 | } |
2733 | ||
24669e58 AK |
2734 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
2735 | unsigned long start, unsigned long end, | |
2736 | struct page *ref_page) | |
63551ae0 | 2737 | { |
24669e58 | 2738 | int force_flush = 0; |
63551ae0 DG |
2739 | struct mm_struct *mm = vma->vm_mm; |
2740 | unsigned long address; | |
c7546f8f | 2741 | pte_t *ptep; |
63551ae0 | 2742 | pte_t pte; |
cb900f41 | 2743 | spinlock_t *ptl; |
63551ae0 | 2744 | struct page *page; |
a5516438 AK |
2745 | struct hstate *h = hstate_vma(vma); |
2746 | unsigned long sz = huge_page_size(h); | |
2ec74c3e SG |
2747 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
2748 | const unsigned long mmun_end = end; /* For mmu_notifiers */ | |
a5516438 | 2749 | |
63551ae0 | 2750 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
2751 | BUG_ON(start & ~huge_page_mask(h)); |
2752 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 2753 | |
24669e58 | 2754 | tlb_start_vma(tlb, vma); |
2ec74c3e | 2755 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
569f48b8 | 2756 | address = start; |
24669e58 | 2757 | again: |
569f48b8 | 2758 | for (; address < end; address += sz) { |
c7546f8f | 2759 | ptep = huge_pte_offset(mm, address); |
4c887265 | 2760 | if (!ptep) |
c7546f8f DG |
2761 | continue; |
2762 | ||
cb900f41 | 2763 | ptl = huge_pte_lock(h, mm, ptep); |
39dde65c | 2764 | if (huge_pmd_unshare(mm, &address, ptep)) |
cb900f41 | 2765 | goto unlock; |
39dde65c | 2766 | |
6629326b HD |
2767 | pte = huge_ptep_get(ptep); |
2768 | if (huge_pte_none(pte)) | |
cb900f41 | 2769 | goto unlock; |
6629326b HD |
2770 | |
2771 | /* | |
9fbc1f63 NH |
2772 | * Migrating hugepage or HWPoisoned hugepage is already |
2773 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 2774 | */ |
9fbc1f63 | 2775 | if (unlikely(!pte_present(pte))) { |
106c992a | 2776 | huge_pte_clear(mm, address, ptep); |
cb900f41 | 2777 | goto unlock; |
8c4894c6 | 2778 | } |
6629326b HD |
2779 | |
2780 | page = pte_page(pte); | |
04f2cbe3 MG |
2781 | /* |
2782 | * If a reference page is supplied, it is because a specific | |
2783 | * page is being unmapped, not a range. Ensure the page we | |
2784 | * are about to unmap is the actual page of interest. | |
2785 | */ | |
2786 | if (ref_page) { | |
04f2cbe3 | 2787 | if (page != ref_page) |
cb900f41 | 2788 | goto unlock; |
04f2cbe3 MG |
2789 | |
2790 | /* | |
2791 | * Mark the VMA as having unmapped its page so that | |
2792 | * future faults in this VMA will fail rather than | |
2793 | * looking like data was lost | |
2794 | */ | |
2795 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
2796 | } | |
2797 | ||
c7546f8f | 2798 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
24669e58 | 2799 | tlb_remove_tlb_entry(tlb, ptep, address); |
106c992a | 2800 | if (huge_pte_dirty(pte)) |
6649a386 | 2801 | set_page_dirty(page); |
9e81130b | 2802 | |
24669e58 AK |
2803 | page_remove_rmap(page); |
2804 | force_flush = !__tlb_remove_page(tlb, page); | |
cb900f41 | 2805 | if (force_flush) { |
569f48b8 | 2806 | address += sz; |
cb900f41 | 2807 | spin_unlock(ptl); |
24669e58 | 2808 | break; |
cb900f41 | 2809 | } |
9e81130b | 2810 | /* Bail out after unmapping reference page if supplied */ |
cb900f41 KS |
2811 | if (ref_page) { |
2812 | spin_unlock(ptl); | |
9e81130b | 2813 | break; |
cb900f41 KS |
2814 | } |
2815 | unlock: | |
2816 | spin_unlock(ptl); | |
63551ae0 | 2817 | } |
24669e58 AK |
2818 | /* |
2819 | * mmu_gather ran out of room to batch pages, we break out of | |
2820 | * the PTE lock to avoid doing the potential expensive TLB invalidate | |
2821 | * and page-free while holding it. | |
2822 | */ | |
2823 | if (force_flush) { | |
2824 | force_flush = 0; | |
2825 | tlb_flush_mmu(tlb); | |
2826 | if (address < end && !ref_page) | |
2827 | goto again; | |
fe1668ae | 2828 | } |
2ec74c3e | 2829 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
24669e58 | 2830 | tlb_end_vma(tlb, vma); |
1da177e4 | 2831 | } |
63551ae0 | 2832 | |
d833352a MG |
2833 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
2834 | struct vm_area_struct *vma, unsigned long start, | |
2835 | unsigned long end, struct page *ref_page) | |
2836 | { | |
2837 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
2838 | ||
2839 | /* | |
2840 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
2841 | * test will fail on a vma being torn down, and not grab a page table | |
2842 | * on its way out. We're lucky that the flag has such an appropriate | |
2843 | * name, and can in fact be safely cleared here. We could clear it | |
2844 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 2845 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 2846 | * because in the context this is called, the VMA is about to be |
c8c06efa | 2847 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
2848 | */ |
2849 | vma->vm_flags &= ~VM_MAYSHARE; | |
2850 | } | |
2851 | ||
502717f4 | 2852 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 2853 | unsigned long end, struct page *ref_page) |
502717f4 | 2854 | { |
24669e58 AK |
2855 | struct mm_struct *mm; |
2856 | struct mmu_gather tlb; | |
2857 | ||
2858 | mm = vma->vm_mm; | |
2859 | ||
2b047252 | 2860 | tlb_gather_mmu(&tlb, mm, start, end); |
24669e58 AK |
2861 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
2862 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 KC |
2863 | } |
2864 | ||
04f2cbe3 MG |
2865 | /* |
2866 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
2867 | * mappping it owns the reserve page for. The intention is to unmap the page | |
2868 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
2869 | * same region. | |
2870 | */ | |
2f4612af DB |
2871 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
2872 | struct page *page, unsigned long address) | |
04f2cbe3 | 2873 | { |
7526674d | 2874 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
2875 | struct vm_area_struct *iter_vma; |
2876 | struct address_space *mapping; | |
04f2cbe3 MG |
2877 | pgoff_t pgoff; |
2878 | ||
2879 | /* | |
2880 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
2881 | * from page cache lookup which is in HPAGE_SIZE units. | |
2882 | */ | |
7526674d | 2883 | address = address & huge_page_mask(h); |
36e4f20a MH |
2884 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
2885 | vma->vm_pgoff; | |
496ad9aa | 2886 | mapping = file_inode(vma->vm_file)->i_mapping; |
04f2cbe3 | 2887 | |
4eb2b1dc MG |
2888 | /* |
2889 | * Take the mapping lock for the duration of the table walk. As | |
2890 | * this mapping should be shared between all the VMAs, | |
2891 | * __unmap_hugepage_range() is called as the lock is already held | |
2892 | */ | |
83cde9e8 | 2893 | i_mmap_lock_write(mapping); |
6b2dbba8 | 2894 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
2895 | /* Do not unmap the current VMA */ |
2896 | if (iter_vma == vma) | |
2897 | continue; | |
2898 | ||
2899 | /* | |
2900 | * Unmap the page from other VMAs without their own reserves. | |
2901 | * They get marked to be SIGKILLed if they fault in these | |
2902 | * areas. This is because a future no-page fault on this VMA | |
2903 | * could insert a zeroed page instead of the data existing | |
2904 | * from the time of fork. This would look like data corruption | |
2905 | */ | |
2906 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
2907 | unmap_hugepage_range(iter_vma, address, |
2908 | address + huge_page_size(h), page); | |
04f2cbe3 | 2909 | } |
83cde9e8 | 2910 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
2911 | } |
2912 | ||
0fe6e20b NH |
2913 | /* |
2914 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
2915 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
2916 | * cannot race with other handlers or page migration. | |
2917 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 2918 | */ |
1e8f889b | 2919 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
04f2cbe3 | 2920 | unsigned long address, pte_t *ptep, pte_t pte, |
cb900f41 | 2921 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 2922 | { |
a5516438 | 2923 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 2924 | struct page *old_page, *new_page; |
ad4404a2 | 2925 | int ret = 0, outside_reserve = 0; |
2ec74c3e SG |
2926 | unsigned long mmun_start; /* For mmu_notifiers */ |
2927 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1e8f889b DG |
2928 | |
2929 | old_page = pte_page(pte); | |
2930 | ||
04f2cbe3 | 2931 | retry_avoidcopy: |
1e8f889b DG |
2932 | /* If no-one else is actually using this page, avoid the copy |
2933 | * and just make the page writable */ | |
37a2140d JK |
2934 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
2935 | page_move_anon_rmap(old_page, vma, address); | |
1e8f889b | 2936 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 2937 | return 0; |
1e8f889b DG |
2938 | } |
2939 | ||
04f2cbe3 MG |
2940 | /* |
2941 | * If the process that created a MAP_PRIVATE mapping is about to | |
2942 | * perform a COW due to a shared page count, attempt to satisfy | |
2943 | * the allocation without using the existing reserves. The pagecache | |
2944 | * page is used to determine if the reserve at this address was | |
2945 | * consumed or not. If reserves were used, a partial faulted mapping | |
2946 | * at the time of fork() could consume its reserves on COW instead | |
2947 | * of the full address range. | |
2948 | */ | |
5944d011 | 2949 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
2950 | old_page != pagecache_page) |
2951 | outside_reserve = 1; | |
2952 | ||
1e8f889b | 2953 | page_cache_get(old_page); |
b76c8cfb | 2954 | |
ad4404a2 DB |
2955 | /* |
2956 | * Drop page table lock as buddy allocator may be called. It will | |
2957 | * be acquired again before returning to the caller, as expected. | |
2958 | */ | |
cb900f41 | 2959 | spin_unlock(ptl); |
04f2cbe3 | 2960 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 2961 | |
2fc39cec | 2962 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
2963 | /* |
2964 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
2965 | * it is due to references held by a child and an insufficient | |
2966 | * huge page pool. To guarantee the original mappers | |
2967 | * reliability, unmap the page from child processes. The child | |
2968 | * may get SIGKILLed if it later faults. | |
2969 | */ | |
2970 | if (outside_reserve) { | |
ad4404a2 | 2971 | page_cache_release(old_page); |
04f2cbe3 | 2972 | BUG_ON(huge_pte_none(pte)); |
2f4612af DB |
2973 | unmap_ref_private(mm, vma, old_page, address); |
2974 | BUG_ON(huge_pte_none(pte)); | |
2975 | spin_lock(ptl); | |
2976 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); | |
2977 | if (likely(ptep && | |
2978 | pte_same(huge_ptep_get(ptep), pte))) | |
2979 | goto retry_avoidcopy; | |
2980 | /* | |
2981 | * race occurs while re-acquiring page table | |
2982 | * lock, and our job is done. | |
2983 | */ | |
2984 | return 0; | |
04f2cbe3 MG |
2985 | } |
2986 | ||
ad4404a2 DB |
2987 | ret = (PTR_ERR(new_page) == -ENOMEM) ? |
2988 | VM_FAULT_OOM : VM_FAULT_SIGBUS; | |
2989 | goto out_release_old; | |
1e8f889b DG |
2990 | } |
2991 | ||
0fe6e20b NH |
2992 | /* |
2993 | * When the original hugepage is shared one, it does not have | |
2994 | * anon_vma prepared. | |
2995 | */ | |
44e2aa93 | 2996 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
2997 | ret = VM_FAULT_OOM; |
2998 | goto out_release_all; | |
44e2aa93 | 2999 | } |
0fe6e20b | 3000 | |
47ad8475 AA |
3001 | copy_user_huge_page(new_page, old_page, address, vma, |
3002 | pages_per_huge_page(h)); | |
0ed361de | 3003 | __SetPageUptodate(new_page); |
bcc54222 | 3004 | set_page_huge_active(new_page); |
1e8f889b | 3005 | |
2ec74c3e SG |
3006 | mmun_start = address & huge_page_mask(h); |
3007 | mmun_end = mmun_start + huge_page_size(h); | |
3008 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
ad4404a2 | 3009 | |
b76c8cfb | 3010 | /* |
cb900f41 | 3011 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
3012 | * before the page tables are altered |
3013 | */ | |
cb900f41 | 3014 | spin_lock(ptl); |
a5516438 | 3015 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
a9af0c5d | 3016 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
3017 | ClearPagePrivate(new_page); |
3018 | ||
1e8f889b | 3019 | /* Break COW */ |
8fe627ec | 3020 | huge_ptep_clear_flush(vma, address, ptep); |
34ee645e | 3021 | mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); |
1e8f889b DG |
3022 | set_huge_pte_at(mm, address, ptep, |
3023 | make_huge_pte(vma, new_page, 1)); | |
0fe6e20b | 3024 | page_remove_rmap(old_page); |
cd67f0d2 | 3025 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
3026 | /* Make the old page be freed below */ |
3027 | new_page = old_page; | |
3028 | } | |
cb900f41 | 3029 | spin_unlock(ptl); |
2ec74c3e | 3030 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
ad4404a2 | 3031 | out_release_all: |
1e8f889b | 3032 | page_cache_release(new_page); |
ad4404a2 | 3033 | out_release_old: |
1e8f889b | 3034 | page_cache_release(old_page); |
8312034f | 3035 | |
ad4404a2 DB |
3036 | spin_lock(ptl); /* Caller expects lock to be held */ |
3037 | return ret; | |
1e8f889b DG |
3038 | } |
3039 | ||
04f2cbe3 | 3040 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
3041 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
3042 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
3043 | { |
3044 | struct address_space *mapping; | |
e7c4b0bf | 3045 | pgoff_t idx; |
04f2cbe3 MG |
3046 | |
3047 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 3048 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
3049 | |
3050 | return find_lock_page(mapping, idx); | |
3051 | } | |
3052 | ||
3ae77f43 HD |
3053 | /* |
3054 | * Return whether there is a pagecache page to back given address within VMA. | |
3055 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
3056 | */ | |
3057 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
3058 | struct vm_area_struct *vma, unsigned long address) |
3059 | { | |
3060 | struct address_space *mapping; | |
3061 | pgoff_t idx; | |
3062 | struct page *page; | |
3063 | ||
3064 | mapping = vma->vm_file->f_mapping; | |
3065 | idx = vma_hugecache_offset(h, vma, address); | |
3066 | ||
3067 | page = find_get_page(mapping, idx); | |
3068 | if (page) | |
3069 | put_page(page); | |
3070 | return page != NULL; | |
3071 | } | |
3072 | ||
a1ed3dda | 3073 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
8382d914 DB |
3074 | struct address_space *mapping, pgoff_t idx, |
3075 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 3076 | { |
a5516438 | 3077 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 3078 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 3079 | int anon_rmap = 0; |
4c887265 | 3080 | unsigned long size; |
4c887265 | 3081 | struct page *page; |
1e8f889b | 3082 | pte_t new_pte; |
cb900f41 | 3083 | spinlock_t *ptl; |
4c887265 | 3084 | |
04f2cbe3 MG |
3085 | /* |
3086 | * Currently, we are forced to kill the process in the event the | |
3087 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 3088 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
3089 | */ |
3090 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
ffb22af5 AM |
3091 | pr_warning("PID %d killed due to inadequate hugepage pool\n", |
3092 | current->pid); | |
04f2cbe3 MG |
3093 | return ret; |
3094 | } | |
3095 | ||
4c887265 AL |
3096 | /* |
3097 | * Use page lock to guard against racing truncation | |
3098 | * before we get page_table_lock. | |
3099 | */ | |
6bda666a CL |
3100 | retry: |
3101 | page = find_lock_page(mapping, idx); | |
3102 | if (!page) { | |
a5516438 | 3103 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
3104 | if (idx >= size) |
3105 | goto out; | |
04f2cbe3 | 3106 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 3107 | if (IS_ERR(page)) { |
76dcee75 AK |
3108 | ret = PTR_ERR(page); |
3109 | if (ret == -ENOMEM) | |
3110 | ret = VM_FAULT_OOM; | |
3111 | else | |
3112 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
3113 | goto out; |
3114 | } | |
47ad8475 | 3115 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 3116 | __SetPageUptodate(page); |
bcc54222 | 3117 | set_page_huge_active(page); |
ac9b9c66 | 3118 | |
f83a275d | 3119 | if (vma->vm_flags & VM_MAYSHARE) { |
6bda666a | 3120 | int err; |
45c682a6 | 3121 | struct inode *inode = mapping->host; |
6bda666a CL |
3122 | |
3123 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
3124 | if (err) { | |
3125 | put_page(page); | |
6bda666a CL |
3126 | if (err == -EEXIST) |
3127 | goto retry; | |
3128 | goto out; | |
3129 | } | |
07443a85 | 3130 | ClearPagePrivate(page); |
45c682a6 KC |
3131 | |
3132 | spin_lock(&inode->i_lock); | |
a5516438 | 3133 | inode->i_blocks += blocks_per_huge_page(h); |
45c682a6 | 3134 | spin_unlock(&inode->i_lock); |
23be7468 | 3135 | } else { |
6bda666a | 3136 | lock_page(page); |
0fe6e20b NH |
3137 | if (unlikely(anon_vma_prepare(vma))) { |
3138 | ret = VM_FAULT_OOM; | |
3139 | goto backout_unlocked; | |
3140 | } | |
409eb8c2 | 3141 | anon_rmap = 1; |
23be7468 | 3142 | } |
0fe6e20b | 3143 | } else { |
998b4382 NH |
3144 | /* |
3145 | * If memory error occurs between mmap() and fault, some process | |
3146 | * don't have hwpoisoned swap entry for errored virtual address. | |
3147 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
3148 | */ | |
3149 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 3150 | ret = VM_FAULT_HWPOISON | |
972dc4de | 3151 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
3152 | goto backout_unlocked; |
3153 | } | |
6bda666a | 3154 | } |
1e8f889b | 3155 | |
57303d80 AW |
3156 | /* |
3157 | * If we are going to COW a private mapping later, we examine the | |
3158 | * pending reservations for this page now. This will ensure that | |
3159 | * any allocations necessary to record that reservation occur outside | |
3160 | * the spinlock. | |
3161 | */ | |
788c7df4 | 3162 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) |
2b26736c AW |
3163 | if (vma_needs_reservation(h, vma, address) < 0) { |
3164 | ret = VM_FAULT_OOM; | |
3165 | goto backout_unlocked; | |
3166 | } | |
57303d80 | 3167 | |
cb900f41 KS |
3168 | ptl = huge_pte_lockptr(h, mm, ptep); |
3169 | spin_lock(ptl); | |
a5516438 | 3170 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
3171 | if (idx >= size) |
3172 | goto backout; | |
3173 | ||
83c54070 | 3174 | ret = 0; |
7f2e9525 | 3175 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
3176 | goto backout; |
3177 | ||
07443a85 JK |
3178 | if (anon_rmap) { |
3179 | ClearPagePrivate(page); | |
409eb8c2 | 3180 | hugepage_add_new_anon_rmap(page, vma, address); |
ac714904 | 3181 | } else |
409eb8c2 | 3182 | page_dup_rmap(page); |
1e8f889b DG |
3183 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
3184 | && (vma->vm_flags & VM_SHARED))); | |
3185 | set_huge_pte_at(mm, address, ptep, new_pte); | |
3186 | ||
788c7df4 | 3187 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 3188 | /* Optimization, do the COW without a second fault */ |
cb900f41 | 3189 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); |
1e8f889b DG |
3190 | } |
3191 | ||
cb900f41 | 3192 | spin_unlock(ptl); |
4c887265 AL |
3193 | unlock_page(page); |
3194 | out: | |
ac9b9c66 | 3195 | return ret; |
4c887265 AL |
3196 | |
3197 | backout: | |
cb900f41 | 3198 | spin_unlock(ptl); |
2b26736c | 3199 | backout_unlocked: |
4c887265 AL |
3200 | unlock_page(page); |
3201 | put_page(page); | |
3202 | goto out; | |
ac9b9c66 HD |
3203 | } |
3204 | ||
8382d914 DB |
3205 | #ifdef CONFIG_SMP |
3206 | static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm, | |
3207 | struct vm_area_struct *vma, | |
3208 | struct address_space *mapping, | |
3209 | pgoff_t idx, unsigned long address) | |
3210 | { | |
3211 | unsigned long key[2]; | |
3212 | u32 hash; | |
3213 | ||
3214 | if (vma->vm_flags & VM_SHARED) { | |
3215 | key[0] = (unsigned long) mapping; | |
3216 | key[1] = idx; | |
3217 | } else { | |
3218 | key[0] = (unsigned long) mm; | |
3219 | key[1] = address >> huge_page_shift(h); | |
3220 | } | |
3221 | ||
3222 | hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); | |
3223 | ||
3224 | return hash & (num_fault_mutexes - 1); | |
3225 | } | |
3226 | #else | |
3227 | /* | |
3228 | * For uniprocesor systems we always use a single mutex, so just | |
3229 | * return 0 and avoid the hashing overhead. | |
3230 | */ | |
3231 | static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm, | |
3232 | struct vm_area_struct *vma, | |
3233 | struct address_space *mapping, | |
3234 | pgoff_t idx, unsigned long address) | |
3235 | { | |
3236 | return 0; | |
3237 | } | |
3238 | #endif | |
3239 | ||
86e5216f | 3240 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 3241 | unsigned long address, unsigned int flags) |
86e5216f | 3242 | { |
8382d914 | 3243 | pte_t *ptep, entry; |
cb900f41 | 3244 | spinlock_t *ptl; |
1e8f889b | 3245 | int ret; |
8382d914 DB |
3246 | u32 hash; |
3247 | pgoff_t idx; | |
0fe6e20b | 3248 | struct page *page = NULL; |
57303d80 | 3249 | struct page *pagecache_page = NULL; |
a5516438 | 3250 | struct hstate *h = hstate_vma(vma); |
8382d914 | 3251 | struct address_space *mapping; |
0f792cf9 | 3252 | int need_wait_lock = 0; |
86e5216f | 3253 | |
1e16a539 KH |
3254 | address &= huge_page_mask(h); |
3255 | ||
fd6a03ed NH |
3256 | ptep = huge_pte_offset(mm, address); |
3257 | if (ptep) { | |
3258 | entry = huge_ptep_get(ptep); | |
290408d4 | 3259 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 3260 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
3261 | return 0; |
3262 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 3263 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 3264 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
3265 | } |
3266 | ||
a5516438 | 3267 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); |
86e5216f AL |
3268 | if (!ptep) |
3269 | return VM_FAULT_OOM; | |
3270 | ||
8382d914 DB |
3271 | mapping = vma->vm_file->f_mapping; |
3272 | idx = vma_hugecache_offset(h, vma, address); | |
3273 | ||
3935baa9 DG |
3274 | /* |
3275 | * Serialize hugepage allocation and instantiation, so that we don't | |
3276 | * get spurious allocation failures if two CPUs race to instantiate | |
3277 | * the same page in the page cache. | |
3278 | */ | |
8382d914 DB |
3279 | hash = fault_mutex_hash(h, mm, vma, mapping, idx, address); |
3280 | mutex_lock(&htlb_fault_mutex_table[hash]); | |
3281 | ||
7f2e9525 GS |
3282 | entry = huge_ptep_get(ptep); |
3283 | if (huge_pte_none(entry)) { | |
8382d914 | 3284 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 3285 | goto out_mutex; |
3935baa9 | 3286 | } |
86e5216f | 3287 | |
83c54070 | 3288 | ret = 0; |
1e8f889b | 3289 | |
0f792cf9 NH |
3290 | /* |
3291 | * entry could be a migration/hwpoison entry at this point, so this | |
3292 | * check prevents the kernel from going below assuming that we have | |
3293 | * a active hugepage in pagecache. This goto expects the 2nd page fault, | |
3294 | * and is_hugetlb_entry_(migration|hwpoisoned) check will properly | |
3295 | * handle it. | |
3296 | */ | |
3297 | if (!pte_present(entry)) | |
3298 | goto out_mutex; | |
3299 | ||
57303d80 AW |
3300 | /* |
3301 | * If we are going to COW the mapping later, we examine the pending | |
3302 | * reservations for this page now. This will ensure that any | |
3303 | * allocations necessary to record that reservation occur outside the | |
3304 | * spinlock. For private mappings, we also lookup the pagecache | |
3305 | * page now as it is used to determine if a reservation has been | |
3306 | * consumed. | |
3307 | */ | |
106c992a | 3308 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
2b26736c AW |
3309 | if (vma_needs_reservation(h, vma, address) < 0) { |
3310 | ret = VM_FAULT_OOM; | |
b4d1d99f | 3311 | goto out_mutex; |
2b26736c | 3312 | } |
57303d80 | 3313 | |
f83a275d | 3314 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
3315 | pagecache_page = hugetlbfs_pagecache_page(h, |
3316 | vma, address); | |
3317 | } | |
3318 | ||
0f792cf9 NH |
3319 | ptl = huge_pte_lock(h, mm, ptep); |
3320 | ||
3321 | /* Check for a racing update before calling hugetlb_cow */ | |
3322 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
3323 | goto out_ptl; | |
3324 | ||
56c9cfb1 NH |
3325 | /* |
3326 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
3327 | * pagecache_page, so here we need take the former one | |
3328 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
3329 | */ |
3330 | page = pte_page(entry); | |
3331 | if (page != pagecache_page) | |
0f792cf9 NH |
3332 | if (!trylock_page(page)) { |
3333 | need_wait_lock = 1; | |
3334 | goto out_ptl; | |
3335 | } | |
b4d1d99f | 3336 | |
0f792cf9 | 3337 | get_page(page); |
b4d1d99f | 3338 | |
788c7df4 | 3339 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 3340 | if (!huge_pte_write(entry)) { |
57303d80 | 3341 | ret = hugetlb_cow(mm, vma, address, ptep, entry, |
cb900f41 | 3342 | pagecache_page, ptl); |
0f792cf9 | 3343 | goto out_put_page; |
b4d1d99f | 3344 | } |
106c992a | 3345 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
3346 | } |
3347 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
3348 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
3349 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 3350 | update_mmu_cache(vma, address, ptep); |
0f792cf9 NH |
3351 | out_put_page: |
3352 | if (page != pagecache_page) | |
3353 | unlock_page(page); | |
3354 | put_page(page); | |
cb900f41 KS |
3355 | out_ptl: |
3356 | spin_unlock(ptl); | |
57303d80 AW |
3357 | |
3358 | if (pagecache_page) { | |
3359 | unlock_page(pagecache_page); | |
3360 | put_page(pagecache_page); | |
3361 | } | |
b4d1d99f | 3362 | out_mutex: |
8382d914 | 3363 | mutex_unlock(&htlb_fault_mutex_table[hash]); |
0f792cf9 NH |
3364 | /* |
3365 | * Generally it's safe to hold refcount during waiting page lock. But | |
3366 | * here we just wait to defer the next page fault to avoid busy loop and | |
3367 | * the page is not used after unlocked before returning from the current | |
3368 | * page fault. So we are safe from accessing freed page, even if we wait | |
3369 | * here without taking refcount. | |
3370 | */ | |
3371 | if (need_wait_lock) | |
3372 | wait_on_page_locked(page); | |
1e8f889b | 3373 | return ret; |
86e5216f AL |
3374 | } |
3375 | ||
28a35716 ML |
3376 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
3377 | struct page **pages, struct vm_area_struct **vmas, | |
3378 | unsigned long *position, unsigned long *nr_pages, | |
3379 | long i, unsigned int flags) | |
63551ae0 | 3380 | { |
d5d4b0aa KC |
3381 | unsigned long pfn_offset; |
3382 | unsigned long vaddr = *position; | |
28a35716 | 3383 | unsigned long remainder = *nr_pages; |
a5516438 | 3384 | struct hstate *h = hstate_vma(vma); |
63551ae0 | 3385 | |
63551ae0 | 3386 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 3387 | pte_t *pte; |
cb900f41 | 3388 | spinlock_t *ptl = NULL; |
2a15efc9 | 3389 | int absent; |
4c887265 | 3390 | struct page *page; |
63551ae0 | 3391 | |
02057967 DR |
3392 | /* |
3393 | * If we have a pending SIGKILL, don't keep faulting pages and | |
3394 | * potentially allocating memory. | |
3395 | */ | |
3396 | if (unlikely(fatal_signal_pending(current))) { | |
3397 | remainder = 0; | |
3398 | break; | |
3399 | } | |
3400 | ||
4c887265 AL |
3401 | /* |
3402 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 3403 | * each hugepage. We have to make sure we get the |
4c887265 | 3404 | * first, for the page indexing below to work. |
cb900f41 KS |
3405 | * |
3406 | * Note that page table lock is not held when pte is null. | |
4c887265 | 3407 | */ |
a5516438 | 3408 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); |
cb900f41 KS |
3409 | if (pte) |
3410 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
3411 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
3412 | ||
3413 | /* | |
3414 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
3415 | * an error where there's an empty slot with no huge pagecache |
3416 | * to back it. This way, we avoid allocating a hugepage, and | |
3417 | * the sparse dumpfile avoids allocating disk blocks, but its | |
3418 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 3419 | */ |
3ae77f43 HD |
3420 | if (absent && (flags & FOLL_DUMP) && |
3421 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
3422 | if (pte) |
3423 | spin_unlock(ptl); | |
2a15efc9 HD |
3424 | remainder = 0; |
3425 | break; | |
3426 | } | |
63551ae0 | 3427 | |
9cc3a5bd NH |
3428 | /* |
3429 | * We need call hugetlb_fault for both hugepages under migration | |
3430 | * (in which case hugetlb_fault waits for the migration,) and | |
3431 | * hwpoisoned hugepages (in which case we need to prevent the | |
3432 | * caller from accessing to them.) In order to do this, we use | |
3433 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
3434 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
3435 | * both cases, and because we can't follow correct pages | |
3436 | * directly from any kind of swap entries. | |
3437 | */ | |
3438 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
3439 | ((flags & FOLL_WRITE) && |
3440 | !huge_pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 3441 | int ret; |
63551ae0 | 3442 | |
cb900f41 KS |
3443 | if (pte) |
3444 | spin_unlock(ptl); | |
2a15efc9 HD |
3445 | ret = hugetlb_fault(mm, vma, vaddr, |
3446 | (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); | |
a89182c7 | 3447 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 3448 | continue; |
63551ae0 | 3449 | |
4c887265 | 3450 | remainder = 0; |
4c887265 AL |
3451 | break; |
3452 | } | |
3453 | ||
a5516438 | 3454 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 3455 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 3456 | same_page: |
d6692183 | 3457 | if (pages) { |
2a15efc9 | 3458 | pages[i] = mem_map_offset(page, pfn_offset); |
a0368d4e | 3459 | get_page_foll(pages[i]); |
d6692183 | 3460 | } |
63551ae0 DG |
3461 | |
3462 | if (vmas) | |
3463 | vmas[i] = vma; | |
3464 | ||
3465 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 3466 | ++pfn_offset; |
63551ae0 DG |
3467 | --remainder; |
3468 | ++i; | |
d5d4b0aa | 3469 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 3470 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa KC |
3471 | /* |
3472 | * We use pfn_offset to avoid touching the pageframes | |
3473 | * of this compound page. | |
3474 | */ | |
3475 | goto same_page; | |
3476 | } | |
cb900f41 | 3477 | spin_unlock(ptl); |
63551ae0 | 3478 | } |
28a35716 | 3479 | *nr_pages = remainder; |
63551ae0 DG |
3480 | *position = vaddr; |
3481 | ||
2a15efc9 | 3482 | return i ? i : -EFAULT; |
63551ae0 | 3483 | } |
8f860591 | 3484 | |
7da4d641 | 3485 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
3486 | unsigned long address, unsigned long end, pgprot_t newprot) |
3487 | { | |
3488 | struct mm_struct *mm = vma->vm_mm; | |
3489 | unsigned long start = address; | |
3490 | pte_t *ptep; | |
3491 | pte_t pte; | |
a5516438 | 3492 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 3493 | unsigned long pages = 0; |
8f860591 ZY |
3494 | |
3495 | BUG_ON(address >= end); | |
3496 | flush_cache_range(vma, address, end); | |
3497 | ||
a5338093 | 3498 | mmu_notifier_invalidate_range_start(mm, start, end); |
83cde9e8 | 3499 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 3500 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 3501 | spinlock_t *ptl; |
8f860591 ZY |
3502 | ptep = huge_pte_offset(mm, address); |
3503 | if (!ptep) | |
3504 | continue; | |
cb900f41 | 3505 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
3506 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3507 | pages++; | |
cb900f41 | 3508 | spin_unlock(ptl); |
39dde65c | 3509 | continue; |
7da4d641 | 3510 | } |
a8bda28d NH |
3511 | pte = huge_ptep_get(ptep); |
3512 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
3513 | spin_unlock(ptl); | |
3514 | continue; | |
3515 | } | |
3516 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
3517 | swp_entry_t entry = pte_to_swp_entry(pte); | |
3518 | ||
3519 | if (is_write_migration_entry(entry)) { | |
3520 | pte_t newpte; | |
3521 | ||
3522 | make_migration_entry_read(&entry); | |
3523 | newpte = swp_entry_to_pte(entry); | |
3524 | set_huge_pte_at(mm, address, ptep, newpte); | |
3525 | pages++; | |
3526 | } | |
3527 | spin_unlock(ptl); | |
3528 | continue; | |
3529 | } | |
3530 | if (!huge_pte_none(pte)) { | |
8f860591 | 3531 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
106c992a | 3532 | pte = pte_mkhuge(huge_pte_modify(pte, newprot)); |
be7517d6 | 3533 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
8f860591 | 3534 | set_huge_pte_at(mm, address, ptep, pte); |
7da4d641 | 3535 | pages++; |
8f860591 | 3536 | } |
cb900f41 | 3537 | spin_unlock(ptl); |
8f860591 | 3538 | } |
d833352a | 3539 | /* |
c8c06efa | 3540 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 3541 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 3542 | * once we release i_mmap_rwsem, another task can do the final put_page |
d833352a MG |
3543 | * and that page table be reused and filled with junk. |
3544 | */ | |
8f860591 | 3545 | flush_tlb_range(vma, start, end); |
34ee645e | 3546 | mmu_notifier_invalidate_range(mm, start, end); |
83cde9e8 | 3547 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
a5338093 | 3548 | mmu_notifier_invalidate_range_end(mm, start, end); |
7da4d641 PZ |
3549 | |
3550 | return pages << h->order; | |
8f860591 ZY |
3551 | } |
3552 | ||
a1e78772 MG |
3553 | int hugetlb_reserve_pages(struct inode *inode, |
3554 | long from, long to, | |
5a6fe125 | 3555 | struct vm_area_struct *vma, |
ca16d140 | 3556 | vm_flags_t vm_flags) |
e4e574b7 | 3557 | { |
17c9d12e | 3558 | long ret, chg; |
a5516438 | 3559 | struct hstate *h = hstate_inode(inode); |
90481622 | 3560 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 3561 | struct resv_map *resv_map; |
1c5ecae3 | 3562 | long gbl_reserve; |
e4e574b7 | 3563 | |
17c9d12e MG |
3564 | /* |
3565 | * Only apply hugepage reservation if asked. At fault time, an | |
3566 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 3567 | * without using reserves |
17c9d12e | 3568 | */ |
ca16d140 | 3569 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
3570 | return 0; |
3571 | ||
a1e78772 MG |
3572 | /* |
3573 | * Shared mappings base their reservation on the number of pages that | |
3574 | * are already allocated on behalf of the file. Private mappings need | |
3575 | * to reserve the full area even if read-only as mprotect() may be | |
3576 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
3577 | */ | |
9119a41e | 3578 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4e35f483 | 3579 | resv_map = inode_resv_map(inode); |
9119a41e | 3580 | |
1406ec9b | 3581 | chg = region_chg(resv_map, from, to); |
9119a41e JK |
3582 | |
3583 | } else { | |
3584 | resv_map = resv_map_alloc(); | |
17c9d12e MG |
3585 | if (!resv_map) |
3586 | return -ENOMEM; | |
3587 | ||
a1e78772 | 3588 | chg = to - from; |
84afd99b | 3589 | |
17c9d12e MG |
3590 | set_vma_resv_map(vma, resv_map); |
3591 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
3592 | } | |
3593 | ||
c50ac050 DH |
3594 | if (chg < 0) { |
3595 | ret = chg; | |
3596 | goto out_err; | |
3597 | } | |
8a630112 | 3598 | |
1c5ecae3 MK |
3599 | /* |
3600 | * There must be enough pages in the subpool for the mapping. If | |
3601 | * the subpool has a minimum size, there may be some global | |
3602 | * reservations already in place (gbl_reserve). | |
3603 | */ | |
3604 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
3605 | if (gbl_reserve < 0) { | |
c50ac050 DH |
3606 | ret = -ENOSPC; |
3607 | goto out_err; | |
3608 | } | |
5a6fe125 MG |
3609 | |
3610 | /* | |
17c9d12e | 3611 | * Check enough hugepages are available for the reservation. |
90481622 | 3612 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 3613 | */ |
1c5ecae3 | 3614 | ret = hugetlb_acct_memory(h, gbl_reserve); |
68842c9b | 3615 | if (ret < 0) { |
1c5ecae3 MK |
3616 | /* put back original number of pages, chg */ |
3617 | (void)hugepage_subpool_put_pages(spool, chg); | |
c50ac050 | 3618 | goto out_err; |
68842c9b | 3619 | } |
17c9d12e MG |
3620 | |
3621 | /* | |
3622 | * Account for the reservations made. Shared mappings record regions | |
3623 | * that have reservations as they are shared by multiple VMAs. | |
3624 | * When the last VMA disappears, the region map says how much | |
3625 | * the reservation was and the page cache tells how much of | |
3626 | * the reservation was consumed. Private mappings are per-VMA and | |
3627 | * only the consumed reservations are tracked. When the VMA | |
3628 | * disappears, the original reservation is the VMA size and the | |
3629 | * consumed reservations are stored in the map. Hence, nothing | |
3630 | * else has to be done for private mappings here | |
3631 | */ | |
f83a275d | 3632 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
1406ec9b | 3633 | region_add(resv_map, from, to); |
a43a8c39 | 3634 | return 0; |
c50ac050 | 3635 | out_err: |
f031dd27 JK |
3636 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3637 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 3638 | return ret; |
a43a8c39 KC |
3639 | } |
3640 | ||
3641 | void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) | |
3642 | { | |
a5516438 | 3643 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 3644 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 3645 | long chg = 0; |
90481622 | 3646 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 3647 | long gbl_reserve; |
45c682a6 | 3648 | |
9119a41e | 3649 | if (resv_map) |
1406ec9b | 3650 | chg = region_truncate(resv_map, offset); |
45c682a6 | 3651 | spin_lock(&inode->i_lock); |
e4c6f8be | 3652 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
3653 | spin_unlock(&inode->i_lock); |
3654 | ||
1c5ecae3 MK |
3655 | /* |
3656 | * If the subpool has a minimum size, the number of global | |
3657 | * reservations to be released may be adjusted. | |
3658 | */ | |
3659 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
3660 | hugetlb_acct_memory(h, -gbl_reserve); | |
a43a8c39 | 3661 | } |
93f70f90 | 3662 | |
3212b535 SC |
3663 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
3664 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
3665 | struct vm_area_struct *vma, | |
3666 | unsigned long addr, pgoff_t idx) | |
3667 | { | |
3668 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
3669 | svma->vm_start; | |
3670 | unsigned long sbase = saddr & PUD_MASK; | |
3671 | unsigned long s_end = sbase + PUD_SIZE; | |
3672 | ||
3673 | /* Allow segments to share if only one is marked locked */ | |
3674 | unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED; | |
3675 | unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED; | |
3676 | ||
3677 | /* | |
3678 | * match the virtual addresses, permission and the alignment of the | |
3679 | * page table page. | |
3680 | */ | |
3681 | if (pmd_index(addr) != pmd_index(saddr) || | |
3682 | vm_flags != svm_flags || | |
3683 | sbase < svma->vm_start || svma->vm_end < s_end) | |
3684 | return 0; | |
3685 | ||
3686 | return saddr; | |
3687 | } | |
3688 | ||
3689 | static int vma_shareable(struct vm_area_struct *vma, unsigned long addr) | |
3690 | { | |
3691 | unsigned long base = addr & PUD_MASK; | |
3692 | unsigned long end = base + PUD_SIZE; | |
3693 | ||
3694 | /* | |
3695 | * check on proper vm_flags and page table alignment | |
3696 | */ | |
3697 | if (vma->vm_flags & VM_MAYSHARE && | |
3698 | vma->vm_start <= base && end <= vma->vm_end) | |
3699 | return 1; | |
3700 | return 0; | |
3701 | } | |
3702 | ||
3703 | /* | |
3704 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
3705 | * and returns the corresponding pte. While this is not necessary for the | |
3706 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
3707 | * code much cleaner. pmd allocation is essential for the shared case because | |
c8c06efa | 3708 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise |
3212b535 SC |
3709 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a |
3710 | * bad pmd for sharing. | |
3711 | */ | |
3712 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
3713 | { | |
3714 | struct vm_area_struct *vma = find_vma(mm, addr); | |
3715 | struct address_space *mapping = vma->vm_file->f_mapping; | |
3716 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
3717 | vma->vm_pgoff; | |
3718 | struct vm_area_struct *svma; | |
3719 | unsigned long saddr; | |
3720 | pte_t *spte = NULL; | |
3721 | pte_t *pte; | |
cb900f41 | 3722 | spinlock_t *ptl; |
3212b535 SC |
3723 | |
3724 | if (!vma_shareable(vma, addr)) | |
3725 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
3726 | ||
83cde9e8 | 3727 | i_mmap_lock_write(mapping); |
3212b535 SC |
3728 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
3729 | if (svma == vma) | |
3730 | continue; | |
3731 | ||
3732 | saddr = page_table_shareable(svma, vma, addr, idx); | |
3733 | if (saddr) { | |
3734 | spte = huge_pte_offset(svma->vm_mm, saddr); | |
3735 | if (spte) { | |
dc6c9a35 | 3736 | mm_inc_nr_pmds(mm); |
3212b535 SC |
3737 | get_page(virt_to_page(spte)); |
3738 | break; | |
3739 | } | |
3740 | } | |
3741 | } | |
3742 | ||
3743 | if (!spte) | |
3744 | goto out; | |
3745 | ||
cb900f41 KS |
3746 | ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); |
3747 | spin_lock(ptl); | |
dc6c9a35 | 3748 | if (pud_none(*pud)) { |
3212b535 SC |
3749 | pud_populate(mm, pud, |
3750 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
dc6c9a35 | 3751 | } else { |
3212b535 | 3752 | put_page(virt_to_page(spte)); |
dc6c9a35 KS |
3753 | mm_inc_nr_pmds(mm); |
3754 | } | |
cb900f41 | 3755 | spin_unlock(ptl); |
3212b535 SC |
3756 | out: |
3757 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
83cde9e8 | 3758 | i_mmap_unlock_write(mapping); |
3212b535 SC |
3759 | return pte; |
3760 | } | |
3761 | ||
3762 | /* | |
3763 | * unmap huge page backed by shared pte. | |
3764 | * | |
3765 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
3766 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
3767 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
3768 | * | |
cb900f41 | 3769 | * called with page table lock held. |
3212b535 SC |
3770 | * |
3771 | * returns: 1 successfully unmapped a shared pte page | |
3772 | * 0 the underlying pte page is not shared, or it is the last user | |
3773 | */ | |
3774 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
3775 | { | |
3776 | pgd_t *pgd = pgd_offset(mm, *addr); | |
3777 | pud_t *pud = pud_offset(pgd, *addr); | |
3778 | ||
3779 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
3780 | if (page_count(virt_to_page(ptep)) == 1) | |
3781 | return 0; | |
3782 | ||
3783 | pud_clear(pud); | |
3784 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 3785 | mm_dec_nr_pmds(mm); |
3212b535 SC |
3786 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
3787 | return 1; | |
3788 | } | |
9e5fc74c SC |
3789 | #define want_pmd_share() (1) |
3790 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
3791 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
3792 | { | |
3793 | return NULL; | |
3794 | } | |
e81f2d22 ZZ |
3795 | |
3796 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
3797 | { | |
3798 | return 0; | |
3799 | } | |
9e5fc74c | 3800 | #define want_pmd_share() (0) |
3212b535 SC |
3801 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
3802 | ||
9e5fc74c SC |
3803 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
3804 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
3805 | unsigned long addr, unsigned long sz) | |
3806 | { | |
3807 | pgd_t *pgd; | |
3808 | pud_t *pud; | |
3809 | pte_t *pte = NULL; | |
3810 | ||
3811 | pgd = pgd_offset(mm, addr); | |
3812 | pud = pud_alloc(mm, pgd, addr); | |
3813 | if (pud) { | |
3814 | if (sz == PUD_SIZE) { | |
3815 | pte = (pte_t *)pud; | |
3816 | } else { | |
3817 | BUG_ON(sz != PMD_SIZE); | |
3818 | if (want_pmd_share() && pud_none(*pud)) | |
3819 | pte = huge_pmd_share(mm, addr, pud); | |
3820 | else | |
3821 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3822 | } | |
3823 | } | |
3824 | BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); | |
3825 | ||
3826 | return pte; | |
3827 | } | |
3828 | ||
3829 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) | |
3830 | { | |
3831 | pgd_t *pgd; | |
3832 | pud_t *pud; | |
3833 | pmd_t *pmd = NULL; | |
3834 | ||
3835 | pgd = pgd_offset(mm, addr); | |
3836 | if (pgd_present(*pgd)) { | |
3837 | pud = pud_offset(pgd, addr); | |
3838 | if (pud_present(*pud)) { | |
3839 | if (pud_huge(*pud)) | |
3840 | return (pte_t *)pud; | |
3841 | pmd = pmd_offset(pud, addr); | |
3842 | } | |
3843 | } | |
3844 | return (pte_t *) pmd; | |
3845 | } | |
3846 | ||
61f77eda NH |
3847 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
3848 | ||
3849 | /* | |
3850 | * These functions are overwritable if your architecture needs its own | |
3851 | * behavior. | |
3852 | */ | |
3853 | struct page * __weak | |
3854 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
3855 | int write) | |
3856 | { | |
3857 | return ERR_PTR(-EINVAL); | |
3858 | } | |
3859 | ||
3860 | struct page * __weak | |
9e5fc74c | 3861 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 3862 | pmd_t *pmd, int flags) |
9e5fc74c | 3863 | { |
e66f17ff NH |
3864 | struct page *page = NULL; |
3865 | spinlock_t *ptl; | |
3866 | retry: | |
3867 | ptl = pmd_lockptr(mm, pmd); | |
3868 | spin_lock(ptl); | |
3869 | /* | |
3870 | * make sure that the address range covered by this pmd is not | |
3871 | * unmapped from other threads. | |
3872 | */ | |
3873 | if (!pmd_huge(*pmd)) | |
3874 | goto out; | |
3875 | if (pmd_present(*pmd)) { | |
97534127 | 3876 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
e66f17ff NH |
3877 | if (flags & FOLL_GET) |
3878 | get_page(page); | |
3879 | } else { | |
3880 | if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) { | |
3881 | spin_unlock(ptl); | |
3882 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
3883 | goto retry; | |
3884 | } | |
3885 | /* | |
3886 | * hwpoisoned entry is treated as no_page_table in | |
3887 | * follow_page_mask(). | |
3888 | */ | |
3889 | } | |
3890 | out: | |
3891 | spin_unlock(ptl); | |
9e5fc74c SC |
3892 | return page; |
3893 | } | |
3894 | ||
61f77eda | 3895 | struct page * __weak |
9e5fc74c | 3896 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 3897 | pud_t *pud, int flags) |
9e5fc74c | 3898 | { |
e66f17ff NH |
3899 | if (flags & FOLL_GET) |
3900 | return NULL; | |
9e5fc74c | 3901 | |
e66f17ff | 3902 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
3903 | } |
3904 | ||
d5bd9106 AK |
3905 | #ifdef CONFIG_MEMORY_FAILURE |
3906 | ||
93f70f90 NH |
3907 | /* |
3908 | * This function is called from memory failure code. | |
3909 | * Assume the caller holds page lock of the head page. | |
3910 | */ | |
6de2b1aa | 3911 | int dequeue_hwpoisoned_huge_page(struct page *hpage) |
93f70f90 NH |
3912 | { |
3913 | struct hstate *h = page_hstate(hpage); | |
3914 | int nid = page_to_nid(hpage); | |
6de2b1aa | 3915 | int ret = -EBUSY; |
93f70f90 NH |
3916 | |
3917 | spin_lock(&hugetlb_lock); | |
7e1f049e NH |
3918 | /* |
3919 | * Just checking !page_huge_active is not enough, because that could be | |
3920 | * an isolated/hwpoisoned hugepage (which have >0 refcount). | |
3921 | */ | |
3922 | if (!page_huge_active(hpage) && !page_count(hpage)) { | |
56f2fb14 NH |
3923 | /* |
3924 | * Hwpoisoned hugepage isn't linked to activelist or freelist, | |
3925 | * but dangling hpage->lru can trigger list-debug warnings | |
3926 | * (this happens when we call unpoison_memory() on it), | |
3927 | * so let it point to itself with list_del_init(). | |
3928 | */ | |
3929 | list_del_init(&hpage->lru); | |
8c6c2ecb | 3930 | set_page_refcounted(hpage); |
6de2b1aa NH |
3931 | h->free_huge_pages--; |
3932 | h->free_huge_pages_node[nid]--; | |
3933 | ret = 0; | |
3934 | } | |
93f70f90 | 3935 | spin_unlock(&hugetlb_lock); |
6de2b1aa | 3936 | return ret; |
93f70f90 | 3937 | } |
6de2b1aa | 3938 | #endif |
31caf665 NH |
3939 | |
3940 | bool isolate_huge_page(struct page *page, struct list_head *list) | |
3941 | { | |
bcc54222 NH |
3942 | bool ret = true; |
3943 | ||
309381fe | 3944 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 3945 | spin_lock(&hugetlb_lock); |
bcc54222 NH |
3946 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { |
3947 | ret = false; | |
3948 | goto unlock; | |
3949 | } | |
3950 | clear_page_huge_active(page); | |
31caf665 | 3951 | list_move_tail(&page->lru, list); |
bcc54222 | 3952 | unlock: |
31caf665 | 3953 | spin_unlock(&hugetlb_lock); |
bcc54222 | 3954 | return ret; |
31caf665 NH |
3955 | } |
3956 | ||
3957 | void putback_active_hugepage(struct page *page) | |
3958 | { | |
309381fe | 3959 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 3960 | spin_lock(&hugetlb_lock); |
bcc54222 | 3961 | set_page_huge_active(page); |
31caf665 NH |
3962 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
3963 | spin_unlock(&hugetlb_lock); | |
3964 | put_page(page); | |
3965 | } |