]>
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> | |
1da177e4 | 7 | #include <linux/mm.h> |
e1759c21 | 8 | #include <linux/seq_file.h> |
1da177e4 LT |
9 | #include <linux/sysctl.h> |
10 | #include <linux/highmem.h> | |
cddb8a5c | 11 | #include <linux/mmu_notifier.h> |
1da177e4 | 12 | #include <linux/nodemask.h> |
63551ae0 | 13 | #include <linux/pagemap.h> |
5da7ca86 | 14 | #include <linux/mempolicy.h> |
3b32123d | 15 | #include <linux/compiler.h> |
aea47ff3 | 16 | #include <linux/cpuset.h> |
3935baa9 | 17 | #include <linux/mutex.h> |
aa888a74 | 18 | #include <linux/bootmem.h> |
a3437870 | 19 | #include <linux/sysfs.h> |
5a0e3ad6 | 20 | #include <linux/slab.h> |
174cd4b1 | 21 | #include <linux/sched/signal.h> |
0fe6e20b | 22 | #include <linux/rmap.h> |
c6247f72 | 23 | #include <linux/string_helpers.h> |
fd6a03ed NH |
24 | #include <linux/swap.h> |
25 | #include <linux/swapops.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> |
1a1aad8a | 36 | #include <linux/userfaultfd_k.h> |
ab5ac90a | 37 | #include <linux/page_owner.h> |
7835e98b | 38 | #include "internal.h" |
1da177e4 | 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; |
9fee021d | 55 | static bool __initdata parsed_valid_hugepagesz = true; |
e5ff2159 | 56 | |
3935baa9 | 57 | /* |
31caf665 NH |
58 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
59 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 60 | */ |
c3f38a38 | 61 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 62 | |
8382d914 DB |
63 | /* |
64 | * Serializes faults on the same logical page. This is used to | |
65 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
66 | */ | |
67 | static int num_fault_mutexes; | |
c672c7f2 | 68 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
8382d914 | 69 | |
7ca02d0a MK |
70 | /* Forward declaration */ |
71 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
72 | ||
90481622 DG |
73 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
74 | { | |
75 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
76 | ||
77 | spin_unlock(&spool->lock); | |
78 | ||
79 | /* If no pages are used, and no other handles to the subpool | |
7ca02d0a MK |
80 | * remain, give up any reservations mased on minimum size and |
81 | * free the subpool */ | |
82 | if (free) { | |
83 | if (spool->min_hpages != -1) | |
84 | hugetlb_acct_memory(spool->hstate, | |
85 | -spool->min_hpages); | |
90481622 | 86 | kfree(spool); |
7ca02d0a | 87 | } |
90481622 DG |
88 | } |
89 | ||
7ca02d0a MK |
90 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
91 | long min_hpages) | |
90481622 DG |
92 | { |
93 | struct hugepage_subpool *spool; | |
94 | ||
c6a91820 | 95 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
96 | if (!spool) |
97 | return NULL; | |
98 | ||
99 | spin_lock_init(&spool->lock); | |
100 | spool->count = 1; | |
7ca02d0a MK |
101 | spool->max_hpages = max_hpages; |
102 | spool->hstate = h; | |
103 | spool->min_hpages = min_hpages; | |
104 | ||
105 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
106 | kfree(spool); | |
107 | return NULL; | |
108 | } | |
109 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
110 | |
111 | return spool; | |
112 | } | |
113 | ||
114 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
115 | { | |
116 | spin_lock(&spool->lock); | |
117 | BUG_ON(!spool->count); | |
118 | spool->count--; | |
119 | unlock_or_release_subpool(spool); | |
120 | } | |
121 | ||
1c5ecae3 MK |
122 | /* |
123 | * Subpool accounting for allocating and reserving pages. | |
124 | * Return -ENOMEM if there are not enough resources to satisfy the | |
125 | * the request. Otherwise, return the number of pages by which the | |
126 | * global pools must be adjusted (upward). The returned value may | |
127 | * only be different than the passed value (delta) in the case where | |
128 | * a subpool minimum size must be manitained. | |
129 | */ | |
130 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
131 | long delta) |
132 | { | |
1c5ecae3 | 133 | long ret = delta; |
90481622 DG |
134 | |
135 | if (!spool) | |
1c5ecae3 | 136 | return ret; |
90481622 DG |
137 | |
138 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
139 | |
140 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
141 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
142 | spool->used_hpages += delta; | |
143 | else { | |
144 | ret = -ENOMEM; | |
145 | goto unlock_ret; | |
146 | } | |
90481622 | 147 | } |
90481622 | 148 | |
09a95e29 MK |
149 | /* minimum size accounting */ |
150 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
151 | if (delta > spool->rsv_hpages) { |
152 | /* | |
153 | * Asking for more reserves than those already taken on | |
154 | * behalf of subpool. Return difference. | |
155 | */ | |
156 | ret = delta - spool->rsv_hpages; | |
157 | spool->rsv_hpages = 0; | |
158 | } else { | |
159 | ret = 0; /* reserves already accounted for */ | |
160 | spool->rsv_hpages -= delta; | |
161 | } | |
162 | } | |
163 | ||
164 | unlock_ret: | |
165 | spin_unlock(&spool->lock); | |
90481622 DG |
166 | return ret; |
167 | } | |
168 | ||
1c5ecae3 MK |
169 | /* |
170 | * Subpool accounting for freeing and unreserving pages. | |
171 | * Return the number of global page reservations that must be dropped. | |
172 | * The return value may only be different than the passed value (delta) | |
173 | * in the case where a subpool minimum size must be maintained. | |
174 | */ | |
175 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
176 | long delta) |
177 | { | |
1c5ecae3 MK |
178 | long ret = delta; |
179 | ||
90481622 | 180 | if (!spool) |
1c5ecae3 | 181 | return delta; |
90481622 DG |
182 | |
183 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
184 | |
185 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
186 | spool->used_hpages -= delta; | |
187 | ||
09a95e29 MK |
188 | /* minimum size accounting */ |
189 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
190 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
191 | ret = 0; | |
192 | else | |
193 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
194 | ||
195 | spool->rsv_hpages += delta; | |
196 | if (spool->rsv_hpages > spool->min_hpages) | |
197 | spool->rsv_hpages = spool->min_hpages; | |
198 | } | |
199 | ||
200 | /* | |
201 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
202 | * quota reference, free it now. | |
203 | */ | |
90481622 | 204 | unlock_or_release_subpool(spool); |
1c5ecae3 MK |
205 | |
206 | return ret; | |
90481622 DG |
207 | } |
208 | ||
209 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
210 | { | |
211 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
212 | } | |
213 | ||
214 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
215 | { | |
496ad9aa | 216 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
217 | } |
218 | ||
96822904 AW |
219 | /* |
220 | * Region tracking -- allows tracking of reservations and instantiated pages | |
221 | * across the pages in a mapping. | |
84afd99b | 222 | * |
1dd308a7 MK |
223 | * The region data structures are embedded into a resv_map and protected |
224 | * by a resv_map's lock. The set of regions within the resv_map represent | |
225 | * reservations for huge pages, or huge pages that have already been | |
226 | * instantiated within the map. The from and to elements are huge page | |
227 | * indicies into the associated mapping. from indicates the starting index | |
228 | * of the region. to represents the first index past the end of the region. | |
229 | * | |
230 | * For example, a file region structure with from == 0 and to == 4 represents | |
231 | * four huge pages in a mapping. It is important to note that the to element | |
232 | * represents the first element past the end of the region. This is used in | |
233 | * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. | |
234 | * | |
235 | * Interval notation of the form [from, to) will be used to indicate that | |
236 | * the endpoint from is inclusive and to is exclusive. | |
96822904 AW |
237 | */ |
238 | struct file_region { | |
239 | struct list_head link; | |
240 | long from; | |
241 | long to; | |
242 | }; | |
243 | ||
1dd308a7 MK |
244 | /* |
245 | * Add the huge page range represented by [f, t) to the reserve | |
5e911373 MK |
246 | * map. In the normal case, existing regions will be expanded |
247 | * to accommodate the specified range. Sufficient regions should | |
248 | * exist for expansion due to the previous call to region_chg | |
249 | * with the same range. However, it is possible that region_del | |
250 | * could have been called after region_chg and modifed the map | |
251 | * in such a way that no region exists to be expanded. In this | |
252 | * case, pull a region descriptor from the cache associated with | |
253 | * the map and use that for the new range. | |
cf3ad20b MK |
254 | * |
255 | * Return the number of new huge pages added to the map. This | |
256 | * number is greater than or equal to zero. | |
1dd308a7 | 257 | */ |
1406ec9b | 258 | static long region_add(struct resv_map *resv, long f, long t) |
96822904 | 259 | { |
1406ec9b | 260 | struct list_head *head = &resv->regions; |
96822904 | 261 | struct file_region *rg, *nrg, *trg; |
cf3ad20b | 262 | long add = 0; |
96822904 | 263 | |
7b24d861 | 264 | spin_lock(&resv->lock); |
96822904 AW |
265 | /* Locate the region we are either in or before. */ |
266 | list_for_each_entry(rg, head, link) | |
267 | if (f <= rg->to) | |
268 | break; | |
269 | ||
5e911373 MK |
270 | /* |
271 | * If no region exists which can be expanded to include the | |
272 | * specified range, the list must have been modified by an | |
273 | * interleving call to region_del(). Pull a region descriptor | |
274 | * from the cache and use it for this range. | |
275 | */ | |
276 | if (&rg->link == head || t < rg->from) { | |
277 | VM_BUG_ON(resv->region_cache_count <= 0); | |
278 | ||
279 | resv->region_cache_count--; | |
280 | nrg = list_first_entry(&resv->region_cache, struct file_region, | |
281 | link); | |
282 | list_del(&nrg->link); | |
283 | ||
284 | nrg->from = f; | |
285 | nrg->to = t; | |
286 | list_add(&nrg->link, rg->link.prev); | |
287 | ||
288 | add += t - f; | |
289 | goto out_locked; | |
290 | } | |
291 | ||
96822904 AW |
292 | /* Round our left edge to the current segment if it encloses us. */ |
293 | if (f > rg->from) | |
294 | f = rg->from; | |
295 | ||
296 | /* Check for and consume any regions we now overlap with. */ | |
297 | nrg = rg; | |
298 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
299 | if (&rg->link == head) | |
300 | break; | |
301 | if (rg->from > t) | |
302 | break; | |
303 | ||
304 | /* If this area reaches higher then extend our area to | |
305 | * include it completely. If this is not the first area | |
306 | * which we intend to reuse, free it. */ | |
307 | if (rg->to > t) | |
308 | t = rg->to; | |
309 | if (rg != nrg) { | |
cf3ad20b MK |
310 | /* Decrement return value by the deleted range. |
311 | * Another range will span this area so that by | |
312 | * end of routine add will be >= zero | |
313 | */ | |
314 | add -= (rg->to - rg->from); | |
96822904 AW |
315 | list_del(&rg->link); |
316 | kfree(rg); | |
317 | } | |
318 | } | |
cf3ad20b MK |
319 | |
320 | add += (nrg->from - f); /* Added to beginning of region */ | |
96822904 | 321 | nrg->from = f; |
cf3ad20b | 322 | add += t - nrg->to; /* Added to end of region */ |
96822904 | 323 | nrg->to = t; |
cf3ad20b | 324 | |
5e911373 MK |
325 | out_locked: |
326 | resv->adds_in_progress--; | |
7b24d861 | 327 | spin_unlock(&resv->lock); |
cf3ad20b MK |
328 | VM_BUG_ON(add < 0); |
329 | return add; | |
96822904 AW |
330 | } |
331 | ||
1dd308a7 MK |
332 | /* |
333 | * Examine the existing reserve map and determine how many | |
334 | * huge pages in the specified range [f, t) are NOT currently | |
335 | * represented. This routine is called before a subsequent | |
336 | * call to region_add that will actually modify the reserve | |
337 | * map to add the specified range [f, t). region_chg does | |
338 | * not change the number of huge pages represented by the | |
339 | * map. However, if the existing regions in the map can not | |
340 | * be expanded to represent the new range, a new file_region | |
341 | * structure is added to the map as a placeholder. This is | |
342 | * so that the subsequent region_add call will have all the | |
343 | * regions it needs and will not fail. | |
344 | * | |
5e911373 MK |
345 | * Upon entry, region_chg will also examine the cache of region descriptors |
346 | * associated with the map. If there are not enough descriptors cached, one | |
347 | * will be allocated for the in progress add operation. | |
348 | * | |
349 | * Returns the number of huge pages that need to be added to the existing | |
350 | * reservation map for the range [f, t). This number is greater or equal to | |
351 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
352 | * is needed and can not be allocated. | |
1dd308a7 | 353 | */ |
1406ec9b | 354 | static long region_chg(struct resv_map *resv, long f, long t) |
96822904 | 355 | { |
1406ec9b | 356 | struct list_head *head = &resv->regions; |
7b24d861 | 357 | struct file_region *rg, *nrg = NULL; |
96822904 AW |
358 | long chg = 0; |
359 | ||
7b24d861 DB |
360 | retry: |
361 | spin_lock(&resv->lock); | |
5e911373 MK |
362 | retry_locked: |
363 | resv->adds_in_progress++; | |
364 | ||
365 | /* | |
366 | * Check for sufficient descriptors in the cache to accommodate | |
367 | * the number of in progress add operations. | |
368 | */ | |
369 | if (resv->adds_in_progress > resv->region_cache_count) { | |
370 | struct file_region *trg; | |
371 | ||
372 | VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1); | |
373 | /* Must drop lock to allocate a new descriptor. */ | |
374 | resv->adds_in_progress--; | |
375 | spin_unlock(&resv->lock); | |
376 | ||
377 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
dbe409e4 MK |
378 | if (!trg) { |
379 | kfree(nrg); | |
5e911373 | 380 | return -ENOMEM; |
dbe409e4 | 381 | } |
5e911373 MK |
382 | |
383 | spin_lock(&resv->lock); | |
384 | list_add(&trg->link, &resv->region_cache); | |
385 | resv->region_cache_count++; | |
386 | goto retry_locked; | |
387 | } | |
388 | ||
96822904 AW |
389 | /* Locate the region we are before or in. */ |
390 | list_for_each_entry(rg, head, link) | |
391 | if (f <= rg->to) | |
392 | break; | |
393 | ||
394 | /* If we are below the current region then a new region is required. | |
395 | * Subtle, allocate a new region at the position but make it zero | |
396 | * size such that we can guarantee to record the reservation. */ | |
397 | if (&rg->link == head || t < rg->from) { | |
7b24d861 | 398 | if (!nrg) { |
5e911373 | 399 | resv->adds_in_progress--; |
7b24d861 DB |
400 | spin_unlock(&resv->lock); |
401 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
402 | if (!nrg) | |
403 | return -ENOMEM; | |
404 | ||
405 | nrg->from = f; | |
406 | nrg->to = f; | |
407 | INIT_LIST_HEAD(&nrg->link); | |
408 | goto retry; | |
409 | } | |
96822904 | 410 | |
7b24d861 DB |
411 | list_add(&nrg->link, rg->link.prev); |
412 | chg = t - f; | |
413 | goto out_nrg; | |
96822904 AW |
414 | } |
415 | ||
416 | /* Round our left edge to the current segment if it encloses us. */ | |
417 | if (f > rg->from) | |
418 | f = rg->from; | |
419 | chg = t - f; | |
420 | ||
421 | /* Check for and consume any regions we now overlap with. */ | |
422 | list_for_each_entry(rg, rg->link.prev, link) { | |
423 | if (&rg->link == head) | |
424 | break; | |
425 | if (rg->from > t) | |
7b24d861 | 426 | goto out; |
96822904 | 427 | |
25985edc | 428 | /* We overlap with this area, if it extends further than |
96822904 AW |
429 | * us then we must extend ourselves. Account for its |
430 | * existing reservation. */ | |
431 | if (rg->to > t) { | |
432 | chg += rg->to - t; | |
433 | t = rg->to; | |
434 | } | |
435 | chg -= rg->to - rg->from; | |
436 | } | |
7b24d861 DB |
437 | |
438 | out: | |
439 | spin_unlock(&resv->lock); | |
440 | /* We already know we raced and no longer need the new region */ | |
441 | kfree(nrg); | |
442 | return chg; | |
443 | out_nrg: | |
444 | spin_unlock(&resv->lock); | |
96822904 AW |
445 | return chg; |
446 | } | |
447 | ||
5e911373 MK |
448 | /* |
449 | * Abort the in progress add operation. The adds_in_progress field | |
450 | * of the resv_map keeps track of the operations in progress between | |
451 | * calls to region_chg and region_add. Operations are sometimes | |
452 | * aborted after the call to region_chg. In such cases, region_abort | |
453 | * is called to decrement the adds_in_progress counter. | |
454 | * | |
455 | * NOTE: The range arguments [f, t) are not needed or used in this | |
456 | * routine. They are kept to make reading the calling code easier as | |
457 | * arguments will match the associated region_chg call. | |
458 | */ | |
459 | static void region_abort(struct resv_map *resv, long f, long t) | |
460 | { | |
461 | spin_lock(&resv->lock); | |
462 | VM_BUG_ON(!resv->region_cache_count); | |
463 | resv->adds_in_progress--; | |
464 | spin_unlock(&resv->lock); | |
465 | } | |
466 | ||
1dd308a7 | 467 | /* |
feba16e2 MK |
468 | * Delete the specified range [f, t) from the reserve map. If the |
469 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
470 | * should be deleted. Locate the regions which intersect [f, t) | |
471 | * and either trim, delete or split the existing regions. | |
472 | * | |
473 | * Returns the number of huge pages deleted from the reserve map. | |
474 | * In the normal case, the return value is zero or more. In the | |
475 | * case where a region must be split, a new region descriptor must | |
476 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
477 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
478 | * a region and possibly return -ENOMEM. Callers specifying | |
479 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 480 | */ |
feba16e2 | 481 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 482 | { |
1406ec9b | 483 | struct list_head *head = &resv->regions; |
96822904 | 484 | struct file_region *rg, *trg; |
feba16e2 MK |
485 | struct file_region *nrg = NULL; |
486 | long del = 0; | |
96822904 | 487 | |
feba16e2 | 488 | retry: |
7b24d861 | 489 | spin_lock(&resv->lock); |
feba16e2 | 490 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
491 | /* |
492 | * Skip regions before the range to be deleted. file_region | |
493 | * ranges are normally of the form [from, to). However, there | |
494 | * may be a "placeholder" entry in the map which is of the form | |
495 | * (from, to) with from == to. Check for placeholder entries | |
496 | * at the beginning of the range to be deleted. | |
497 | */ | |
498 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 499 | continue; |
dbe409e4 | 500 | |
feba16e2 | 501 | if (rg->from >= t) |
96822904 | 502 | break; |
96822904 | 503 | |
feba16e2 MK |
504 | if (f > rg->from && t < rg->to) { /* Must split region */ |
505 | /* | |
506 | * Check for an entry in the cache before dropping | |
507 | * lock and attempting allocation. | |
508 | */ | |
509 | if (!nrg && | |
510 | resv->region_cache_count > resv->adds_in_progress) { | |
511 | nrg = list_first_entry(&resv->region_cache, | |
512 | struct file_region, | |
513 | link); | |
514 | list_del(&nrg->link); | |
515 | resv->region_cache_count--; | |
516 | } | |
96822904 | 517 | |
feba16e2 MK |
518 | if (!nrg) { |
519 | spin_unlock(&resv->lock); | |
520 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
521 | if (!nrg) | |
522 | return -ENOMEM; | |
523 | goto retry; | |
524 | } | |
525 | ||
526 | del += t - f; | |
527 | ||
528 | /* New entry for end of split region */ | |
529 | nrg->from = t; | |
530 | nrg->to = rg->to; | |
531 | INIT_LIST_HEAD(&nrg->link); | |
532 | ||
533 | /* Original entry is trimmed */ | |
534 | rg->to = f; | |
535 | ||
536 | list_add(&nrg->link, &rg->link); | |
537 | nrg = NULL; | |
96822904 | 538 | break; |
feba16e2 MK |
539 | } |
540 | ||
541 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
542 | del += rg->to - rg->from; | |
543 | list_del(&rg->link); | |
544 | kfree(rg); | |
545 | continue; | |
546 | } | |
547 | ||
548 | if (f <= rg->from) { /* Trim beginning of region */ | |
549 | del += t - rg->from; | |
550 | rg->from = t; | |
551 | } else { /* Trim end of region */ | |
552 | del += rg->to - f; | |
553 | rg->to = f; | |
554 | } | |
96822904 | 555 | } |
7b24d861 | 556 | |
7b24d861 | 557 | spin_unlock(&resv->lock); |
feba16e2 MK |
558 | kfree(nrg); |
559 | return del; | |
96822904 AW |
560 | } |
561 | ||
b5cec28d MK |
562 | /* |
563 | * A rare out of memory error was encountered which prevented removal of | |
564 | * the reserve map region for a page. The huge page itself was free'ed | |
565 | * and removed from the page cache. This routine will adjust the subpool | |
566 | * usage count, and the global reserve count if needed. By incrementing | |
567 | * these counts, the reserve map entry which could not be deleted will | |
568 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
569 | * counts. | |
570 | */ | |
72e2936c | 571 | void hugetlb_fix_reserve_counts(struct inode *inode) |
b5cec28d MK |
572 | { |
573 | struct hugepage_subpool *spool = subpool_inode(inode); | |
574 | long rsv_adjust; | |
575 | ||
576 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
72e2936c | 577 | if (rsv_adjust) { |
b5cec28d MK |
578 | struct hstate *h = hstate_inode(inode); |
579 | ||
580 | hugetlb_acct_memory(h, 1); | |
581 | } | |
582 | } | |
583 | ||
1dd308a7 MK |
584 | /* |
585 | * Count and return the number of huge pages in the reserve map | |
586 | * that intersect with the range [f, t). | |
587 | */ | |
1406ec9b | 588 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 589 | { |
1406ec9b | 590 | struct list_head *head = &resv->regions; |
84afd99b AW |
591 | struct file_region *rg; |
592 | long chg = 0; | |
593 | ||
7b24d861 | 594 | spin_lock(&resv->lock); |
84afd99b AW |
595 | /* Locate each segment we overlap with, and count that overlap. */ |
596 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
597 | long seg_from; |
598 | long seg_to; | |
84afd99b AW |
599 | |
600 | if (rg->to <= f) | |
601 | continue; | |
602 | if (rg->from >= t) | |
603 | break; | |
604 | ||
605 | seg_from = max(rg->from, f); | |
606 | seg_to = min(rg->to, t); | |
607 | ||
608 | chg += seg_to - seg_from; | |
609 | } | |
7b24d861 | 610 | spin_unlock(&resv->lock); |
84afd99b AW |
611 | |
612 | return chg; | |
613 | } | |
614 | ||
e7c4b0bf AW |
615 | /* |
616 | * Convert the address within this vma to the page offset within | |
617 | * the mapping, in pagecache page units; huge pages here. | |
618 | */ | |
a5516438 AK |
619 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
620 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 621 | { |
a5516438 AK |
622 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
623 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
624 | } |
625 | ||
0fe6e20b NH |
626 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
627 | unsigned long address) | |
628 | { | |
629 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
630 | } | |
dee41079 | 631 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 632 | |
08fba699 MG |
633 | /* |
634 | * Return the size of the pages allocated when backing a VMA. In the majority | |
635 | * cases this will be same size as used by the page table entries. | |
636 | */ | |
637 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
638 | { | |
639 | struct hstate *hstate; | |
640 | ||
641 | if (!is_vm_hugetlb_page(vma)) | |
642 | return PAGE_SIZE; | |
643 | ||
644 | hstate = hstate_vma(vma); | |
645 | ||
2415cf12 | 646 | return 1UL << huge_page_shift(hstate); |
08fba699 | 647 | } |
f340ca0f | 648 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 649 | |
3340289d MG |
650 | /* |
651 | * Return the page size being used by the MMU to back a VMA. In the majority | |
652 | * of cases, the page size used by the kernel matches the MMU size. On | |
653 | * architectures where it differs, an architecture-specific version of this | |
654 | * function is required. | |
655 | */ | |
656 | #ifndef vma_mmu_pagesize | |
657 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) | |
658 | { | |
659 | return vma_kernel_pagesize(vma); | |
660 | } | |
661 | #endif | |
662 | ||
84afd99b AW |
663 | /* |
664 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
665 | * bits of the reservation map pointer, which are always clear due to | |
666 | * alignment. | |
667 | */ | |
668 | #define HPAGE_RESV_OWNER (1UL << 0) | |
669 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 670 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 671 | |
a1e78772 MG |
672 | /* |
673 | * These helpers are used to track how many pages are reserved for | |
674 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
675 | * is guaranteed to have their future faults succeed. | |
676 | * | |
677 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
678 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
679 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
680 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
681 | * |
682 | * The private mapping reservation is represented in a subtly different | |
683 | * manner to a shared mapping. A shared mapping has a region map associated | |
684 | * with the underlying file, this region map represents the backing file | |
685 | * pages which have ever had a reservation assigned which this persists even | |
686 | * after the page is instantiated. A private mapping has a region map | |
687 | * associated with the original mmap which is attached to all VMAs which | |
688 | * reference it, this region map represents those offsets which have consumed | |
689 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 690 | */ |
e7c4b0bf AW |
691 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
692 | { | |
693 | return (unsigned long)vma->vm_private_data; | |
694 | } | |
695 | ||
696 | static void set_vma_private_data(struct vm_area_struct *vma, | |
697 | unsigned long value) | |
698 | { | |
699 | vma->vm_private_data = (void *)value; | |
700 | } | |
701 | ||
9119a41e | 702 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
703 | { |
704 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
705 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
706 | ||
707 | if (!resv_map || !rg) { | |
708 | kfree(resv_map); | |
709 | kfree(rg); | |
84afd99b | 710 | return NULL; |
5e911373 | 711 | } |
84afd99b AW |
712 | |
713 | kref_init(&resv_map->refs); | |
7b24d861 | 714 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
715 | INIT_LIST_HEAD(&resv_map->regions); |
716 | ||
5e911373 MK |
717 | resv_map->adds_in_progress = 0; |
718 | ||
719 | INIT_LIST_HEAD(&resv_map->region_cache); | |
720 | list_add(&rg->link, &resv_map->region_cache); | |
721 | resv_map->region_cache_count = 1; | |
722 | ||
84afd99b AW |
723 | return resv_map; |
724 | } | |
725 | ||
9119a41e | 726 | void resv_map_release(struct kref *ref) |
84afd99b AW |
727 | { |
728 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
729 | struct list_head *head = &resv_map->region_cache; |
730 | struct file_region *rg, *trg; | |
84afd99b AW |
731 | |
732 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 733 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
734 | |
735 | /* ... and any entries left in the cache */ | |
736 | list_for_each_entry_safe(rg, trg, head, link) { | |
737 | list_del(&rg->link); | |
738 | kfree(rg); | |
739 | } | |
740 | ||
741 | VM_BUG_ON(resv_map->adds_in_progress); | |
742 | ||
84afd99b AW |
743 | kfree(resv_map); |
744 | } | |
745 | ||
4e35f483 JK |
746 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
747 | { | |
748 | return inode->i_mapping->private_data; | |
749 | } | |
750 | ||
84afd99b | 751 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 752 | { |
81d1b09c | 753 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
754 | if (vma->vm_flags & VM_MAYSHARE) { |
755 | struct address_space *mapping = vma->vm_file->f_mapping; | |
756 | struct inode *inode = mapping->host; | |
757 | ||
758 | return inode_resv_map(inode); | |
759 | ||
760 | } else { | |
84afd99b AW |
761 | return (struct resv_map *)(get_vma_private_data(vma) & |
762 | ~HPAGE_RESV_MASK); | |
4e35f483 | 763 | } |
a1e78772 MG |
764 | } |
765 | ||
84afd99b | 766 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 767 | { |
81d1b09c SL |
768 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
769 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 770 | |
84afd99b AW |
771 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
772 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
773 | } |
774 | ||
775 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
776 | { | |
81d1b09c SL |
777 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
778 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
779 | |
780 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
781 | } |
782 | ||
783 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
784 | { | |
81d1b09c | 785 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
786 | |
787 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
788 | } |
789 | ||
04f2cbe3 | 790 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
791 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
792 | { | |
81d1b09c | 793 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 794 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
795 | vma->vm_private_data = (void *)0; |
796 | } | |
797 | ||
798 | /* Returns true if the VMA has associated reserve pages */ | |
559ec2f8 | 799 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 800 | { |
af0ed73e JK |
801 | if (vma->vm_flags & VM_NORESERVE) { |
802 | /* | |
803 | * This address is already reserved by other process(chg == 0), | |
804 | * so, we should decrement reserved count. Without decrementing, | |
805 | * reserve count remains after releasing inode, because this | |
806 | * allocated page will go into page cache and is regarded as | |
807 | * coming from reserved pool in releasing step. Currently, we | |
808 | * don't have any other solution to deal with this situation | |
809 | * properly, so add work-around here. | |
810 | */ | |
811 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 812 | return true; |
af0ed73e | 813 | else |
559ec2f8 | 814 | return false; |
af0ed73e | 815 | } |
a63884e9 JK |
816 | |
817 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
818 | if (vma->vm_flags & VM_MAYSHARE) { |
819 | /* | |
820 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
821 | * be a region map for all pages. The only situation where | |
822 | * there is no region map is if a hole was punched via | |
823 | * fallocate. In this case, there really are no reverves to | |
824 | * use. This situation is indicated if chg != 0. | |
825 | */ | |
826 | if (chg) | |
827 | return false; | |
828 | else | |
829 | return true; | |
830 | } | |
a63884e9 JK |
831 | |
832 | /* | |
833 | * Only the process that called mmap() has reserves for | |
834 | * private mappings. | |
835 | */ | |
67961f9d MK |
836 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
837 | /* | |
838 | * Like the shared case above, a hole punch or truncate | |
839 | * could have been performed on the private mapping. | |
840 | * Examine the value of chg to determine if reserves | |
841 | * actually exist or were previously consumed. | |
842 | * Very Subtle - The value of chg comes from a previous | |
843 | * call to vma_needs_reserves(). The reserve map for | |
844 | * private mappings has different (opposite) semantics | |
845 | * than that of shared mappings. vma_needs_reserves() | |
846 | * has already taken this difference in semantics into | |
847 | * account. Therefore, the meaning of chg is the same | |
848 | * as in the shared case above. Code could easily be | |
849 | * combined, but keeping it separate draws attention to | |
850 | * subtle differences. | |
851 | */ | |
852 | if (chg) | |
853 | return false; | |
854 | else | |
855 | return true; | |
856 | } | |
a63884e9 | 857 | |
559ec2f8 | 858 | return false; |
a1e78772 MG |
859 | } |
860 | ||
a5516438 | 861 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
862 | { |
863 | int nid = page_to_nid(page); | |
0edaecfa | 864 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
865 | h->free_huge_pages++; |
866 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
867 | } |
868 | ||
94310cbc | 869 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
bf50bab2 NH |
870 | { |
871 | struct page *page; | |
872 | ||
c8721bbb | 873 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
243abd5b | 874 | if (!PageHWPoison(page)) |
c8721bbb NH |
875 | break; |
876 | /* | |
877 | * if 'non-isolated free hugepage' not found on the list, | |
878 | * the allocation fails. | |
879 | */ | |
880 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 881 | return NULL; |
0edaecfa | 882 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 883 | set_page_refcounted(page); |
bf50bab2 NH |
884 | h->free_huge_pages--; |
885 | h->free_huge_pages_node[nid]--; | |
886 | return page; | |
887 | } | |
888 | ||
3e59fcb0 MH |
889 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, |
890 | nodemask_t *nmask) | |
94310cbc | 891 | { |
3e59fcb0 MH |
892 | unsigned int cpuset_mems_cookie; |
893 | struct zonelist *zonelist; | |
894 | struct zone *zone; | |
895 | struct zoneref *z; | |
896 | int node = -1; | |
94310cbc | 897 | |
3e59fcb0 MH |
898 | zonelist = node_zonelist(nid, gfp_mask); |
899 | ||
900 | retry_cpuset: | |
901 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
902 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
903 | struct page *page; | |
904 | ||
905 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
906 | continue; | |
907 | /* | |
908 | * no need to ask again on the same node. Pool is node rather than | |
909 | * zone aware | |
910 | */ | |
911 | if (zone_to_nid(zone) == node) | |
912 | continue; | |
913 | node = zone_to_nid(zone); | |
94310cbc | 914 | |
94310cbc AK |
915 | page = dequeue_huge_page_node_exact(h, node); |
916 | if (page) | |
917 | return page; | |
918 | } | |
3e59fcb0 MH |
919 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) |
920 | goto retry_cpuset; | |
921 | ||
94310cbc AK |
922 | return NULL; |
923 | } | |
924 | ||
86cdb465 NH |
925 | /* Movability of hugepages depends on migration support. */ |
926 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
927 | { | |
d6cb41cc | 928 | if (hugepage_migration_supported(h)) |
86cdb465 NH |
929 | return GFP_HIGHUSER_MOVABLE; |
930 | else | |
931 | return GFP_HIGHUSER; | |
932 | } | |
933 | ||
a5516438 AK |
934 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
935 | struct vm_area_struct *vma, | |
af0ed73e JK |
936 | unsigned long address, int avoid_reserve, |
937 | long chg) | |
1da177e4 | 938 | { |
3e59fcb0 | 939 | struct page *page; |
480eccf9 | 940 | struct mempolicy *mpol; |
04ec6264 | 941 | gfp_t gfp_mask; |
3e59fcb0 | 942 | nodemask_t *nodemask; |
04ec6264 | 943 | int nid; |
1da177e4 | 944 | |
a1e78772 MG |
945 | /* |
946 | * A child process with MAP_PRIVATE mappings created by their parent | |
947 | * have no page reserves. This check ensures that reservations are | |
948 | * not "stolen". The child may still get SIGKILLed | |
949 | */ | |
af0ed73e | 950 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 951 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 952 | goto err; |
a1e78772 | 953 | |
04f2cbe3 | 954 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 955 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 956 | goto err; |
04f2cbe3 | 957 | |
04ec6264 VB |
958 | gfp_mask = htlb_alloc_mask(h); |
959 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
3e59fcb0 MH |
960 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); |
961 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { | |
962 | SetPagePrivate(page); | |
963 | h->resv_huge_pages--; | |
1da177e4 | 964 | } |
cc9a6c87 | 965 | |
52cd3b07 | 966 | mpol_cond_put(mpol); |
1da177e4 | 967 | return page; |
cc9a6c87 MG |
968 | |
969 | err: | |
cc9a6c87 | 970 | return NULL; |
1da177e4 LT |
971 | } |
972 | ||
1cac6f2c LC |
973 | /* |
974 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
975 | * We may have allocated or freed a huge page based on a different | |
976 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
977 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
978 | * node for alloc or free. | |
979 | */ | |
980 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
981 | { | |
0edaf86c | 982 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
983 | VM_BUG_ON(nid >= MAX_NUMNODES); |
984 | ||
985 | return nid; | |
986 | } | |
987 | ||
988 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
989 | { | |
990 | if (!node_isset(nid, *nodes_allowed)) | |
991 | nid = next_node_allowed(nid, nodes_allowed); | |
992 | return nid; | |
993 | } | |
994 | ||
995 | /* | |
996 | * returns the previously saved node ["this node"] from which to | |
997 | * allocate a persistent huge page for the pool and advance the | |
998 | * next node from which to allocate, handling wrap at end of node | |
999 | * mask. | |
1000 | */ | |
1001 | static int hstate_next_node_to_alloc(struct hstate *h, | |
1002 | nodemask_t *nodes_allowed) | |
1003 | { | |
1004 | int nid; | |
1005 | ||
1006 | VM_BUG_ON(!nodes_allowed); | |
1007 | ||
1008 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1009 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1010 | ||
1011 | return nid; | |
1012 | } | |
1013 | ||
1014 | /* | |
1015 | * helper for free_pool_huge_page() - return the previously saved | |
1016 | * node ["this node"] from which to free a huge page. Advance the | |
1017 | * next node id whether or not we find a free huge page to free so | |
1018 | * that the next attempt to free addresses the next node. | |
1019 | */ | |
1020 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1021 | { | |
1022 | int nid; | |
1023 | ||
1024 | VM_BUG_ON(!nodes_allowed); | |
1025 | ||
1026 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1027 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1028 | ||
1029 | return nid; | |
1030 | } | |
1031 | ||
1032 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1033 | for (nr_nodes = nodes_weight(*mask); \ | |
1034 | nr_nodes > 0 && \ | |
1035 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1036 | nr_nodes--) | |
1037 | ||
1038 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1039 | for (nr_nodes = nodes_weight(*mask); \ | |
1040 | nr_nodes > 0 && \ | |
1041 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1042 | nr_nodes--) | |
1043 | ||
e1073d1e | 1044 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE |
944d9fec | 1045 | static void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1046 | unsigned int order) |
944d9fec LC |
1047 | { |
1048 | int i; | |
1049 | int nr_pages = 1 << order; | |
1050 | struct page *p = page + 1; | |
1051 | ||
c8cc708a | 1052 | atomic_set(compound_mapcount_ptr(page), 0); |
944d9fec | 1053 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
1d798ca3 | 1054 | clear_compound_head(p); |
944d9fec | 1055 | set_page_refcounted(p); |
944d9fec LC |
1056 | } |
1057 | ||
1058 | set_compound_order(page, 0); | |
1059 | __ClearPageHead(page); | |
1060 | } | |
1061 | ||
d00181b9 | 1062 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec LC |
1063 | { |
1064 | free_contig_range(page_to_pfn(page), 1 << order); | |
1065 | } | |
1066 | ||
1067 | static int __alloc_gigantic_page(unsigned long start_pfn, | |
79b63f12 | 1068 | unsigned long nr_pages, gfp_t gfp_mask) |
944d9fec LC |
1069 | { |
1070 | unsigned long end_pfn = start_pfn + nr_pages; | |
ca96b625 | 1071 | return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE, |
79b63f12 | 1072 | gfp_mask); |
944d9fec LC |
1073 | } |
1074 | ||
f44b2dda JK |
1075 | static bool pfn_range_valid_gigantic(struct zone *z, |
1076 | unsigned long start_pfn, unsigned long nr_pages) | |
944d9fec LC |
1077 | { |
1078 | unsigned long i, end_pfn = start_pfn + nr_pages; | |
1079 | struct page *page; | |
1080 | ||
1081 | for (i = start_pfn; i < end_pfn; i++) { | |
1082 | if (!pfn_valid(i)) | |
1083 | return false; | |
1084 | ||
1085 | page = pfn_to_page(i); | |
1086 | ||
f44b2dda JK |
1087 | if (page_zone(page) != z) |
1088 | return false; | |
1089 | ||
944d9fec LC |
1090 | if (PageReserved(page)) |
1091 | return false; | |
1092 | ||
1093 | if (page_count(page) > 0) | |
1094 | return false; | |
1095 | ||
1096 | if (PageHuge(page)) | |
1097 | return false; | |
1098 | } | |
1099 | ||
1100 | return true; | |
1101 | } | |
1102 | ||
1103 | static bool zone_spans_last_pfn(const struct zone *zone, | |
1104 | unsigned long start_pfn, unsigned long nr_pages) | |
1105 | { | |
1106 | unsigned long last_pfn = start_pfn + nr_pages - 1; | |
1107 | return zone_spans_pfn(zone, last_pfn); | |
1108 | } | |
1109 | ||
d9cc948f MH |
1110 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
1111 | int nid, nodemask_t *nodemask) | |
944d9fec | 1112 | { |
79b63f12 | 1113 | unsigned int order = huge_page_order(h); |
944d9fec LC |
1114 | unsigned long nr_pages = 1 << order; |
1115 | unsigned long ret, pfn, flags; | |
79b63f12 MH |
1116 | struct zonelist *zonelist; |
1117 | struct zone *zone; | |
1118 | struct zoneref *z; | |
944d9fec | 1119 | |
79b63f12 | 1120 | zonelist = node_zonelist(nid, gfp_mask); |
d9cc948f | 1121 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) { |
79b63f12 | 1122 | spin_lock_irqsave(&zone->lock, flags); |
944d9fec | 1123 | |
79b63f12 MH |
1124 | pfn = ALIGN(zone->zone_start_pfn, nr_pages); |
1125 | while (zone_spans_last_pfn(zone, pfn, nr_pages)) { | |
1126 | if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) { | |
944d9fec LC |
1127 | /* |
1128 | * We release the zone lock here because | |
1129 | * alloc_contig_range() will also lock the zone | |
1130 | * at some point. If there's an allocation | |
1131 | * spinning on this lock, it may win the race | |
1132 | * and cause alloc_contig_range() to fail... | |
1133 | */ | |
79b63f12 MH |
1134 | spin_unlock_irqrestore(&zone->lock, flags); |
1135 | ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask); | |
944d9fec LC |
1136 | if (!ret) |
1137 | return pfn_to_page(pfn); | |
79b63f12 | 1138 | spin_lock_irqsave(&zone->lock, flags); |
944d9fec LC |
1139 | } |
1140 | pfn += nr_pages; | |
1141 | } | |
1142 | ||
79b63f12 | 1143 | spin_unlock_irqrestore(&zone->lock, flags); |
944d9fec LC |
1144 | } |
1145 | ||
1146 | return NULL; | |
1147 | } | |
1148 | ||
1149 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
d00181b9 | 1150 | static void prep_compound_gigantic_page(struct page *page, unsigned int order); |
944d9fec | 1151 | |
e1073d1e | 1152 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
944d9fec | 1153 | static inline bool gigantic_page_supported(void) { return false; } |
d9cc948f MH |
1154 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
1155 | int nid, nodemask_t *nodemask) { return NULL; } | |
d00181b9 | 1156 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1157 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1158 | unsigned int order) { } |
944d9fec LC |
1159 | #endif |
1160 | ||
a5516438 | 1161 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1162 | { |
1163 | int i; | |
a5516438 | 1164 | |
944d9fec LC |
1165 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
1166 | return; | |
18229df5 | 1167 | |
a5516438 AK |
1168 | h->nr_huge_pages--; |
1169 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
1170 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
1171 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
1172 | 1 << PG_referenced | 1 << PG_dirty | | |
a7407a27 LC |
1173 | 1 << PG_active | 1 << PG_private | |
1174 | 1 << PG_writeback); | |
6af2acb6 | 1175 | } |
309381fe | 1176 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
f1e61557 | 1177 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); |
6af2acb6 | 1178 | set_page_refcounted(page); |
944d9fec LC |
1179 | if (hstate_is_gigantic(h)) { |
1180 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
1181 | free_gigantic_page(page, huge_page_order(h)); | |
1182 | } else { | |
944d9fec LC |
1183 | __free_pages(page, huge_page_order(h)); |
1184 | } | |
6af2acb6 AL |
1185 | } |
1186 | ||
e5ff2159 AK |
1187 | struct hstate *size_to_hstate(unsigned long size) |
1188 | { | |
1189 | struct hstate *h; | |
1190 | ||
1191 | for_each_hstate(h) { | |
1192 | if (huge_page_size(h) == size) | |
1193 | return h; | |
1194 | } | |
1195 | return NULL; | |
1196 | } | |
1197 | ||
bcc54222 NH |
1198 | /* |
1199 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
1200 | * to hstate->hugepage_activelist.) | |
1201 | * | |
1202 | * This function can be called for tail pages, but never returns true for them. | |
1203 | */ | |
1204 | bool page_huge_active(struct page *page) | |
1205 | { | |
1206 | VM_BUG_ON_PAGE(!PageHuge(page), page); | |
1207 | return PageHead(page) && PagePrivate(&page[1]); | |
1208 | } | |
1209 | ||
1210 | /* never called for tail page */ | |
1211 | static void set_page_huge_active(struct page *page) | |
1212 | { | |
1213 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1214 | SetPagePrivate(&page[1]); | |
1215 | } | |
1216 | ||
1217 | static void clear_page_huge_active(struct page *page) | |
1218 | { | |
1219 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1220 | ClearPagePrivate(&page[1]); | |
1221 | } | |
1222 | ||
ab5ac90a MH |
1223 | /* |
1224 | * Internal hugetlb specific page flag. Do not use outside of the hugetlb | |
1225 | * code | |
1226 | */ | |
1227 | static inline bool PageHugeTemporary(struct page *page) | |
1228 | { | |
1229 | if (!PageHuge(page)) | |
1230 | return false; | |
1231 | ||
1232 | return (unsigned long)page[2].mapping == -1U; | |
1233 | } | |
1234 | ||
1235 | static inline void SetPageHugeTemporary(struct page *page) | |
1236 | { | |
1237 | page[2].mapping = (void *)-1U; | |
1238 | } | |
1239 | ||
1240 | static inline void ClearPageHugeTemporary(struct page *page) | |
1241 | { | |
1242 | page[2].mapping = NULL; | |
1243 | } | |
1244 | ||
8f1d26d0 | 1245 | void free_huge_page(struct page *page) |
27a85ef1 | 1246 | { |
a5516438 AK |
1247 | /* |
1248 | * Can't pass hstate in here because it is called from the | |
1249 | * compound page destructor. | |
1250 | */ | |
e5ff2159 | 1251 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1252 | int nid = page_to_nid(page); |
90481622 DG |
1253 | struct hugepage_subpool *spool = |
1254 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 1255 | bool restore_reserve; |
27a85ef1 | 1256 | |
e5df70ab | 1257 | set_page_private(page, 0); |
23be7468 | 1258 | page->mapping = NULL; |
b4330afb MK |
1259 | VM_BUG_ON_PAGE(page_count(page), page); |
1260 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
07443a85 | 1261 | restore_reserve = PagePrivate(page); |
16c794b4 | 1262 | ClearPagePrivate(page); |
27a85ef1 | 1263 | |
1c5ecae3 MK |
1264 | /* |
1265 | * A return code of zero implies that the subpool will be under its | |
1266 | * minimum size if the reservation is not restored after page is free. | |
1267 | * Therefore, force restore_reserve operation. | |
1268 | */ | |
1269 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1270 | restore_reserve = true; | |
1271 | ||
27a85ef1 | 1272 | spin_lock(&hugetlb_lock); |
bcc54222 | 1273 | clear_page_huge_active(page); |
6d76dcf4 AK |
1274 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1275 | pages_per_huge_page(h), page); | |
07443a85 JK |
1276 | if (restore_reserve) |
1277 | h->resv_huge_pages++; | |
1278 | ||
ab5ac90a MH |
1279 | if (PageHugeTemporary(page)) { |
1280 | list_del(&page->lru); | |
1281 | ClearPageHugeTemporary(page); | |
1282 | update_and_free_page(h, page); | |
1283 | } else if (h->surplus_huge_pages_node[nid]) { | |
0edaecfa AK |
1284 | /* remove the page from active list */ |
1285 | list_del(&page->lru); | |
a5516438 AK |
1286 | update_and_free_page(h, page); |
1287 | h->surplus_huge_pages--; | |
1288 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 1289 | } else { |
5d3a551c | 1290 | arch_clear_hugepage_flags(page); |
a5516438 | 1291 | enqueue_huge_page(h, page); |
7893d1d5 | 1292 | } |
27a85ef1 DG |
1293 | spin_unlock(&hugetlb_lock); |
1294 | } | |
1295 | ||
a5516438 | 1296 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1297 | { |
0edaecfa | 1298 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1299 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
b7ba30c6 | 1300 | spin_lock(&hugetlb_lock); |
9dd540e2 | 1301 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
1302 | h->nr_huge_pages++; |
1303 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 | 1304 | spin_unlock(&hugetlb_lock); |
b7ba30c6 AK |
1305 | } |
1306 | ||
d00181b9 | 1307 | static void prep_compound_gigantic_page(struct page *page, unsigned int order) |
20a0307c WF |
1308 | { |
1309 | int i; | |
1310 | int nr_pages = 1 << order; | |
1311 | struct page *p = page + 1; | |
1312 | ||
1313 | /* we rely on prep_new_huge_page to set the destructor */ | |
1314 | set_compound_order(page, order); | |
ef5a22be | 1315 | __ClearPageReserved(page); |
de09d31d | 1316 | __SetPageHead(page); |
20a0307c | 1317 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1318 | /* |
1319 | * For gigantic hugepages allocated through bootmem at | |
1320 | * boot, it's safer to be consistent with the not-gigantic | |
1321 | * hugepages and clear the PG_reserved bit from all tail pages | |
1322 | * too. Otherwse drivers using get_user_pages() to access tail | |
1323 | * pages may get the reference counting wrong if they see | |
1324 | * PG_reserved set on a tail page (despite the head page not | |
1325 | * having PG_reserved set). Enforcing this consistency between | |
1326 | * head and tail pages allows drivers to optimize away a check | |
1327 | * on the head page when they need know if put_page() is needed | |
1328 | * after get_user_pages(). | |
1329 | */ | |
1330 | __ClearPageReserved(p); | |
58a84aa9 | 1331 | set_page_count(p, 0); |
1d798ca3 | 1332 | set_compound_head(p, page); |
20a0307c | 1333 | } |
b4330afb | 1334 | atomic_set(compound_mapcount_ptr(page), -1); |
20a0307c WF |
1335 | } |
1336 | ||
7795912c AM |
1337 | /* |
1338 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1339 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1340 | * details. | |
1341 | */ | |
20a0307c WF |
1342 | int PageHuge(struct page *page) |
1343 | { | |
20a0307c WF |
1344 | if (!PageCompound(page)) |
1345 | return 0; | |
1346 | ||
1347 | page = compound_head(page); | |
f1e61557 | 1348 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1349 | } |
43131e14 NH |
1350 | EXPORT_SYMBOL_GPL(PageHuge); |
1351 | ||
27c73ae7 AA |
1352 | /* |
1353 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1354 | * normal or transparent huge pages. | |
1355 | */ | |
1356 | int PageHeadHuge(struct page *page_head) | |
1357 | { | |
27c73ae7 AA |
1358 | if (!PageHead(page_head)) |
1359 | return 0; | |
1360 | ||
758f66a2 | 1361 | return get_compound_page_dtor(page_head) == free_huge_page; |
27c73ae7 | 1362 | } |
27c73ae7 | 1363 | |
13d60f4b ZY |
1364 | pgoff_t __basepage_index(struct page *page) |
1365 | { | |
1366 | struct page *page_head = compound_head(page); | |
1367 | pgoff_t index = page_index(page_head); | |
1368 | unsigned long compound_idx; | |
1369 | ||
1370 | if (!PageHuge(page_head)) | |
1371 | return page_index(page); | |
1372 | ||
1373 | if (compound_order(page_head) >= MAX_ORDER) | |
1374 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1375 | else | |
1376 | compound_idx = page - page_head; | |
1377 | ||
1378 | return (index << compound_order(page_head)) + compound_idx; | |
1379 | } | |
1380 | ||
af0fb9df MH |
1381 | static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h, |
1382 | gfp_t gfp_mask, int nid, nodemask_t *nmask) | |
1da177e4 | 1383 | { |
af0fb9df | 1384 | int order = huge_page_order(h); |
1da177e4 | 1385 | struct page *page; |
f96efd58 | 1386 | |
af0fb9df MH |
1387 | gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN; |
1388 | if (nid == NUMA_NO_NODE) | |
1389 | nid = numa_mem_id(); | |
1390 | page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); | |
1391 | if (page) | |
1392 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
1393 | else | |
1394 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
63b4613c NA |
1395 | |
1396 | return page; | |
1397 | } | |
1398 | ||
af0fb9df MH |
1399 | /* |
1400 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
1401 | * manner. | |
1402 | */ | |
b2261026 JK |
1403 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
1404 | { | |
1405 | struct page *page; | |
1406 | int nr_nodes, node; | |
af0fb9df | 1407 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
b2261026 JK |
1408 | |
1409 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
d9cc948f MH |
1410 | if (hstate_is_gigantic(h)) |
1411 | page = alloc_gigantic_page(h, gfp_mask, | |
1412 | node, nodes_allowed); | |
1413 | else | |
1414 | page = __hugetlb_alloc_buddy_huge_page(h, gfp_mask, | |
1415 | node, nodes_allowed); | |
af0fb9df | 1416 | if (page) |
b2261026 | 1417 | break; |
af0fb9df | 1418 | |
b2261026 JK |
1419 | } |
1420 | ||
af0fb9df MH |
1421 | if (!page) |
1422 | return 0; | |
b2261026 | 1423 | |
d9cc948f MH |
1424 | if (hstate_is_gigantic(h)) |
1425 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
af0fb9df MH |
1426 | prep_new_huge_page(h, page, page_to_nid(page)); |
1427 | put_page(page); /* free it into the hugepage allocator */ | |
1428 | ||
1429 | return 1; | |
b2261026 JK |
1430 | } |
1431 | ||
e8c5c824 LS |
1432 | /* |
1433 | * Free huge page from pool from next node to free. | |
1434 | * Attempt to keep persistent huge pages more or less | |
1435 | * balanced over allowed nodes. | |
1436 | * Called with hugetlb_lock locked. | |
1437 | */ | |
6ae11b27 LS |
1438 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1439 | bool acct_surplus) | |
e8c5c824 | 1440 | { |
b2261026 | 1441 | int nr_nodes, node; |
e8c5c824 LS |
1442 | int ret = 0; |
1443 | ||
b2261026 | 1444 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1445 | /* |
1446 | * If we're returning unused surplus pages, only examine | |
1447 | * nodes with surplus pages. | |
1448 | */ | |
b2261026 JK |
1449 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1450 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1451 | struct page *page = |
b2261026 | 1452 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1453 | struct page, lru); |
1454 | list_del(&page->lru); | |
1455 | h->free_huge_pages--; | |
b2261026 | 1456 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1457 | if (acct_surplus) { |
1458 | h->surplus_huge_pages--; | |
b2261026 | 1459 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1460 | } |
e8c5c824 LS |
1461 | update_and_free_page(h, page); |
1462 | ret = 1; | |
9a76db09 | 1463 | break; |
e8c5c824 | 1464 | } |
b2261026 | 1465 | } |
e8c5c824 LS |
1466 | |
1467 | return ret; | |
1468 | } | |
1469 | ||
c8721bbb NH |
1470 | /* |
1471 | * Dissolve a given free hugepage into free buddy pages. This function does | |
082d5b6b GS |
1472 | * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the |
1473 | * number of free hugepages would be reduced below the number of reserved | |
1474 | * hugepages. | |
c8721bbb | 1475 | */ |
c3114a84 | 1476 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 1477 | { |
082d5b6b GS |
1478 | int rc = 0; |
1479 | ||
c8721bbb NH |
1480 | spin_lock(&hugetlb_lock); |
1481 | if (PageHuge(page) && !page_count(page)) { | |
2247bb33 GS |
1482 | struct page *head = compound_head(page); |
1483 | struct hstate *h = page_hstate(head); | |
1484 | int nid = page_to_nid(head); | |
082d5b6b GS |
1485 | if (h->free_huge_pages - h->resv_huge_pages == 0) { |
1486 | rc = -EBUSY; | |
1487 | goto out; | |
1488 | } | |
c3114a84 AK |
1489 | /* |
1490 | * Move PageHWPoison flag from head page to the raw error page, | |
1491 | * which makes any subpages rather than the error page reusable. | |
1492 | */ | |
1493 | if (PageHWPoison(head) && page != head) { | |
1494 | SetPageHWPoison(page); | |
1495 | ClearPageHWPoison(head); | |
1496 | } | |
2247bb33 | 1497 | list_del(&head->lru); |
c8721bbb NH |
1498 | h->free_huge_pages--; |
1499 | h->free_huge_pages_node[nid]--; | |
c1470b33 | 1500 | h->max_huge_pages--; |
2247bb33 | 1501 | update_and_free_page(h, head); |
c8721bbb | 1502 | } |
082d5b6b | 1503 | out: |
c8721bbb | 1504 | spin_unlock(&hugetlb_lock); |
082d5b6b | 1505 | return rc; |
c8721bbb NH |
1506 | } |
1507 | ||
1508 | /* | |
1509 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1510 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
1511 | * Note that this will dissolve a free gigantic hugepage completely, if any |
1512 | * part of it lies within the given range. | |
082d5b6b GS |
1513 | * Also note that if dissolve_free_huge_page() returns with an error, all |
1514 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 1515 | */ |
082d5b6b | 1516 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 1517 | { |
c8721bbb | 1518 | unsigned long pfn; |
eb03aa00 | 1519 | struct page *page; |
082d5b6b | 1520 | int rc = 0; |
c8721bbb | 1521 | |
d0177639 | 1522 | if (!hugepages_supported()) |
082d5b6b | 1523 | return rc; |
d0177639 | 1524 | |
eb03aa00 GS |
1525 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { |
1526 | page = pfn_to_page(pfn); | |
1527 | if (PageHuge(page) && !page_count(page)) { | |
1528 | rc = dissolve_free_huge_page(page); | |
1529 | if (rc) | |
1530 | break; | |
1531 | } | |
1532 | } | |
082d5b6b GS |
1533 | |
1534 | return rc; | |
c8721bbb NH |
1535 | } |
1536 | ||
ab5ac90a MH |
1537 | /* |
1538 | * Allocates a fresh surplus page from the page allocator. | |
1539 | */ | |
1540 | static struct page *__alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, | |
aaf14e40 | 1541 | int nid, nodemask_t *nmask) |
7893d1d5 | 1542 | { |
9980d744 | 1543 | struct page *page = NULL; |
7893d1d5 | 1544 | |
bae7f4ae | 1545 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1546 | return NULL; |
1547 | ||
d1c3fb1f | 1548 | spin_lock(&hugetlb_lock); |
9980d744 MH |
1549 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) |
1550 | goto out_unlock; | |
d1c3fb1f NA |
1551 | spin_unlock(&hugetlb_lock); |
1552 | ||
aaf14e40 | 1553 | page = __hugetlb_alloc_buddy_huge_page(h, gfp_mask, nid, nmask); |
9980d744 MH |
1554 | if (!page) |
1555 | goto out_unlock; | |
d1c3fb1f NA |
1556 | |
1557 | spin_lock(&hugetlb_lock); | |
9980d744 MH |
1558 | /* |
1559 | * We could have raced with the pool size change. | |
1560 | * Double check that and simply deallocate the new page | |
1561 | * if we would end up overcommiting the surpluses. Abuse | |
1562 | * temporary page to workaround the nasty free_huge_page | |
1563 | * codeflow | |
1564 | */ | |
1565 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
1566 | SetPageHugeTemporary(page); | |
1567 | put_page(page); | |
1568 | page = NULL; | |
1569 | } else { | |
1570 | int r_nid; | |
1571 | ||
1572 | h->surplus_huge_pages++; | |
1573 | h->nr_huge_pages++; | |
0edaecfa | 1574 | INIT_LIST_HEAD(&page->lru); |
bf50bab2 | 1575 | r_nid = page_to_nid(page); |
f1e61557 | 1576 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
9dd540e2 | 1577 | set_hugetlb_cgroup(page, NULL); |
bf50bab2 NH |
1578 | h->nr_huge_pages_node[r_nid]++; |
1579 | h->surplus_huge_pages_node[r_nid]++; | |
7893d1d5 | 1580 | } |
9980d744 MH |
1581 | |
1582 | out_unlock: | |
d1c3fb1f | 1583 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1584 | |
1585 | return page; | |
1586 | } | |
1587 | ||
ab5ac90a MH |
1588 | static struct page *__alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, |
1589 | int nid, nodemask_t *nmask) | |
1590 | { | |
1591 | struct page *page; | |
1592 | ||
1593 | if (hstate_is_gigantic(h)) | |
1594 | return NULL; | |
1595 | ||
1596 | page = __hugetlb_alloc_buddy_huge_page(h, gfp_mask, nid, nmask); | |
1597 | if (!page) | |
1598 | return NULL; | |
1599 | ||
1600 | /* | |
1601 | * We do not account these pages as surplus because they are only | |
1602 | * temporary and will be released properly on the last reference | |
1603 | */ | |
1604 | prep_new_huge_page(h, page, page_to_nid(page)); | |
1605 | SetPageHugeTemporary(page); | |
1606 | ||
1607 | return page; | |
1608 | } | |
1609 | ||
099730d6 DH |
1610 | /* |
1611 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
1612 | */ | |
e0ec90ee | 1613 | static |
099730d6 DH |
1614 | struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h, |
1615 | struct vm_area_struct *vma, unsigned long addr) | |
1616 | { | |
aaf14e40 MH |
1617 | struct page *page; |
1618 | struct mempolicy *mpol; | |
1619 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
1620 | int nid; | |
1621 | nodemask_t *nodemask; | |
1622 | ||
1623 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
ab5ac90a | 1624 | page = __alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); |
aaf14e40 MH |
1625 | mpol_cond_put(mpol); |
1626 | ||
1627 | return page; | |
099730d6 DH |
1628 | } |
1629 | ||
ab5ac90a | 1630 | /* page migration callback function */ |
bf50bab2 NH |
1631 | struct page *alloc_huge_page_node(struct hstate *h, int nid) |
1632 | { | |
aaf14e40 | 1633 | gfp_t gfp_mask = htlb_alloc_mask(h); |
4ef91848 | 1634 | struct page *page = NULL; |
bf50bab2 | 1635 | |
aaf14e40 MH |
1636 | if (nid != NUMA_NO_NODE) |
1637 | gfp_mask |= __GFP_THISNODE; | |
1638 | ||
bf50bab2 | 1639 | spin_lock(&hugetlb_lock); |
4ef91848 | 1640 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
3e59fcb0 | 1641 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL); |
bf50bab2 NH |
1642 | spin_unlock(&hugetlb_lock); |
1643 | ||
94ae8ba7 | 1644 | if (!page) |
ab5ac90a | 1645 | page = __alloc_migrate_huge_page(h, gfp_mask, nid, NULL); |
bf50bab2 NH |
1646 | |
1647 | return page; | |
1648 | } | |
1649 | ||
ab5ac90a | 1650 | /* page migration callback function */ |
3e59fcb0 MH |
1651 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, |
1652 | nodemask_t *nmask) | |
4db9b2ef | 1653 | { |
aaf14e40 | 1654 | gfp_t gfp_mask = htlb_alloc_mask(h); |
4db9b2ef MH |
1655 | |
1656 | spin_lock(&hugetlb_lock); | |
1657 | if (h->free_huge_pages - h->resv_huge_pages > 0) { | |
3e59fcb0 MH |
1658 | struct page *page; |
1659 | ||
1660 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
1661 | if (page) { | |
1662 | spin_unlock(&hugetlb_lock); | |
1663 | return page; | |
4db9b2ef MH |
1664 | } |
1665 | } | |
1666 | spin_unlock(&hugetlb_lock); | |
4db9b2ef | 1667 | |
ab5ac90a | 1668 | return __alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); |
4db9b2ef MH |
1669 | } |
1670 | ||
e4e574b7 | 1671 | /* |
25985edc | 1672 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1673 | * of size 'delta'. |
1674 | */ | |
a5516438 | 1675 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1676 | { |
1677 | struct list_head surplus_list; | |
1678 | struct page *page, *tmp; | |
1679 | int ret, i; | |
1680 | int needed, allocated; | |
28073b02 | 1681 | bool alloc_ok = true; |
e4e574b7 | 1682 | |
a5516438 | 1683 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1684 | if (needed <= 0) { |
a5516438 | 1685 | h->resv_huge_pages += delta; |
e4e574b7 | 1686 | return 0; |
ac09b3a1 | 1687 | } |
e4e574b7 AL |
1688 | |
1689 | allocated = 0; | |
1690 | INIT_LIST_HEAD(&surplus_list); | |
1691 | ||
1692 | ret = -ENOMEM; | |
1693 | retry: | |
1694 | spin_unlock(&hugetlb_lock); | |
1695 | for (i = 0; i < needed; i++) { | |
ab5ac90a | 1696 | page = __alloc_surplus_huge_page(h, htlb_alloc_mask(h), |
aaf14e40 | 1697 | NUMA_NO_NODE, NULL); |
28073b02 HD |
1698 | if (!page) { |
1699 | alloc_ok = false; | |
1700 | break; | |
1701 | } | |
e4e574b7 | 1702 | list_add(&page->lru, &surplus_list); |
69ed779a | 1703 | cond_resched(); |
e4e574b7 | 1704 | } |
28073b02 | 1705 | allocated += i; |
e4e574b7 AL |
1706 | |
1707 | /* | |
1708 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1709 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1710 | */ | |
1711 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1712 | needed = (h->resv_huge_pages + delta) - |
1713 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1714 | if (needed > 0) { |
1715 | if (alloc_ok) | |
1716 | goto retry; | |
1717 | /* | |
1718 | * We were not able to allocate enough pages to | |
1719 | * satisfy the entire reservation so we free what | |
1720 | * we've allocated so far. | |
1721 | */ | |
1722 | goto free; | |
1723 | } | |
e4e574b7 AL |
1724 | /* |
1725 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1726 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1727 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1728 | * allocator. Commit the entire reservation here to prevent another |
1729 | * process from stealing the pages as they are added to the pool but | |
1730 | * before they are reserved. | |
e4e574b7 AL |
1731 | */ |
1732 | needed += allocated; | |
a5516438 | 1733 | h->resv_huge_pages += delta; |
e4e574b7 | 1734 | ret = 0; |
a9869b83 | 1735 | |
19fc3f0a | 1736 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1737 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1738 | if ((--needed) < 0) |
1739 | break; | |
a9869b83 NH |
1740 | /* |
1741 | * This page is now managed by the hugetlb allocator and has | |
1742 | * no users -- drop the buddy allocator's reference. | |
1743 | */ | |
1744 | put_page_testzero(page); | |
309381fe | 1745 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 1746 | enqueue_huge_page(h, page); |
19fc3f0a | 1747 | } |
28073b02 | 1748 | free: |
b0365c8d | 1749 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1750 | |
1751 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1752 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1753 | put_page(page); | |
a9869b83 | 1754 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1755 | |
1756 | return ret; | |
1757 | } | |
1758 | ||
1759 | /* | |
e5bbc8a6 MK |
1760 | * This routine has two main purposes: |
1761 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
1762 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
1763 | * to the associated reservation map. | |
1764 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
1765 | * the reservation. As many as unused_resv_pages may be freed. | |
1766 | * | |
1767 | * Called with hugetlb_lock held. However, the lock could be dropped (and | |
1768 | * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, | |
1769 | * we must make sure nobody else can claim pages we are in the process of | |
1770 | * freeing. Do this by ensuring resv_huge_page always is greater than the | |
1771 | * number of huge pages we plan to free when dropping the lock. | |
e4e574b7 | 1772 | */ |
a5516438 AK |
1773 | static void return_unused_surplus_pages(struct hstate *h, |
1774 | unsigned long unused_resv_pages) | |
e4e574b7 | 1775 | { |
e4e574b7 AL |
1776 | unsigned long nr_pages; |
1777 | ||
aa888a74 | 1778 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 1779 | if (hstate_is_gigantic(h)) |
e5bbc8a6 | 1780 | goto out; |
aa888a74 | 1781 | |
e5bbc8a6 MK |
1782 | /* |
1783 | * Part (or even all) of the reservation could have been backed | |
1784 | * by pre-allocated pages. Only free surplus pages. | |
1785 | */ | |
a5516438 | 1786 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1787 | |
685f3457 LS |
1788 | /* |
1789 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1790 | * evenly across all nodes with memory. Iterate across these nodes |
1791 | * until we can no longer free unreserved surplus pages. This occurs | |
1792 | * when the nodes with surplus pages have no free pages. | |
1793 | * free_pool_huge_page() will balance the the freed pages across the | |
1794 | * on-line nodes with memory and will handle the hstate accounting. | |
e5bbc8a6 MK |
1795 | * |
1796 | * Note that we decrement resv_huge_pages as we free the pages. If | |
1797 | * we drop the lock, resv_huge_pages will still be sufficiently large | |
1798 | * to cover subsequent pages we may free. | |
685f3457 LS |
1799 | */ |
1800 | while (nr_pages--) { | |
e5bbc8a6 MK |
1801 | h->resv_huge_pages--; |
1802 | unused_resv_pages--; | |
8cebfcd0 | 1803 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
e5bbc8a6 | 1804 | goto out; |
7848a4bf | 1805 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 | 1806 | } |
e5bbc8a6 MK |
1807 | |
1808 | out: | |
1809 | /* Fully uncommit the reservation */ | |
1810 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 AL |
1811 | } |
1812 | ||
5e911373 | 1813 | |
c37f9fb1 | 1814 | /* |
feba16e2 | 1815 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 1816 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
1817 | * |
1818 | * vma_needs_reservation is called to determine if the huge page at addr | |
1819 | * within the vma has an associated reservation. If a reservation is | |
1820 | * needed, the value 1 is returned. The caller is then responsible for | |
1821 | * managing the global reservation and subpool usage counts. After | |
1822 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
1823 | * to add the page to the reservation map. If the page allocation fails, |
1824 | * the reservation must be ended instead of committed. vma_end_reservation | |
1825 | * is called in such cases. | |
cf3ad20b MK |
1826 | * |
1827 | * In the normal case, vma_commit_reservation returns the same value | |
1828 | * as the preceding vma_needs_reservation call. The only time this | |
1829 | * is not the case is if a reserve map was changed between calls. It | |
1830 | * is the responsibility of the caller to notice the difference and | |
1831 | * take appropriate action. | |
96b96a96 MK |
1832 | * |
1833 | * vma_add_reservation is used in error paths where a reservation must | |
1834 | * be restored when a newly allocated huge page must be freed. It is | |
1835 | * to be called after calling vma_needs_reservation to determine if a | |
1836 | * reservation exists. | |
c37f9fb1 | 1837 | */ |
5e911373 MK |
1838 | enum vma_resv_mode { |
1839 | VMA_NEEDS_RESV, | |
1840 | VMA_COMMIT_RESV, | |
feba16e2 | 1841 | VMA_END_RESV, |
96b96a96 | 1842 | VMA_ADD_RESV, |
5e911373 | 1843 | }; |
cf3ad20b MK |
1844 | static long __vma_reservation_common(struct hstate *h, |
1845 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 1846 | enum vma_resv_mode mode) |
c37f9fb1 | 1847 | { |
4e35f483 JK |
1848 | struct resv_map *resv; |
1849 | pgoff_t idx; | |
cf3ad20b | 1850 | long ret; |
c37f9fb1 | 1851 | |
4e35f483 JK |
1852 | resv = vma_resv_map(vma); |
1853 | if (!resv) | |
84afd99b | 1854 | return 1; |
c37f9fb1 | 1855 | |
4e35f483 | 1856 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
1857 | switch (mode) { |
1858 | case VMA_NEEDS_RESV: | |
cf3ad20b | 1859 | ret = region_chg(resv, idx, idx + 1); |
5e911373 MK |
1860 | break; |
1861 | case VMA_COMMIT_RESV: | |
1862 | ret = region_add(resv, idx, idx + 1); | |
1863 | break; | |
feba16e2 | 1864 | case VMA_END_RESV: |
5e911373 MK |
1865 | region_abort(resv, idx, idx + 1); |
1866 | ret = 0; | |
1867 | break; | |
96b96a96 MK |
1868 | case VMA_ADD_RESV: |
1869 | if (vma->vm_flags & VM_MAYSHARE) | |
1870 | ret = region_add(resv, idx, idx + 1); | |
1871 | else { | |
1872 | region_abort(resv, idx, idx + 1); | |
1873 | ret = region_del(resv, idx, idx + 1); | |
1874 | } | |
1875 | break; | |
5e911373 MK |
1876 | default: |
1877 | BUG(); | |
1878 | } | |
84afd99b | 1879 | |
4e35f483 | 1880 | if (vma->vm_flags & VM_MAYSHARE) |
cf3ad20b | 1881 | return ret; |
67961f9d MK |
1882 | else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { |
1883 | /* | |
1884 | * In most cases, reserves always exist for private mappings. | |
1885 | * However, a file associated with mapping could have been | |
1886 | * hole punched or truncated after reserves were consumed. | |
1887 | * As subsequent fault on such a range will not use reserves. | |
1888 | * Subtle - The reserve map for private mappings has the | |
1889 | * opposite meaning than that of shared mappings. If NO | |
1890 | * entry is in the reserve map, it means a reservation exists. | |
1891 | * If an entry exists in the reserve map, it means the | |
1892 | * reservation has already been consumed. As a result, the | |
1893 | * return value of this routine is the opposite of the | |
1894 | * value returned from reserve map manipulation routines above. | |
1895 | */ | |
1896 | if (ret) | |
1897 | return 0; | |
1898 | else | |
1899 | return 1; | |
1900 | } | |
4e35f483 | 1901 | else |
cf3ad20b | 1902 | return ret < 0 ? ret : 0; |
c37f9fb1 | 1903 | } |
cf3ad20b MK |
1904 | |
1905 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 1906 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 1907 | { |
5e911373 | 1908 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 1909 | } |
84afd99b | 1910 | |
cf3ad20b MK |
1911 | static long vma_commit_reservation(struct hstate *h, |
1912 | struct vm_area_struct *vma, unsigned long addr) | |
1913 | { | |
5e911373 MK |
1914 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
1915 | } | |
1916 | ||
feba16e2 | 1917 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
1918 | struct vm_area_struct *vma, unsigned long addr) |
1919 | { | |
feba16e2 | 1920 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
1921 | } |
1922 | ||
96b96a96 MK |
1923 | static long vma_add_reservation(struct hstate *h, |
1924 | struct vm_area_struct *vma, unsigned long addr) | |
1925 | { | |
1926 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
1927 | } | |
1928 | ||
1929 | /* | |
1930 | * This routine is called to restore a reservation on error paths. In the | |
1931 | * specific error paths, a huge page was allocated (via alloc_huge_page) | |
1932 | * and is about to be freed. If a reservation for the page existed, | |
1933 | * alloc_huge_page would have consumed the reservation and set PagePrivate | |
1934 | * in the newly allocated page. When the page is freed via free_huge_page, | |
1935 | * the global reservation count will be incremented if PagePrivate is set. | |
1936 | * However, free_huge_page can not adjust the reserve map. Adjust the | |
1937 | * reserve map here to be consistent with global reserve count adjustments | |
1938 | * to be made by free_huge_page. | |
1939 | */ | |
1940 | static void restore_reserve_on_error(struct hstate *h, | |
1941 | struct vm_area_struct *vma, unsigned long address, | |
1942 | struct page *page) | |
1943 | { | |
1944 | if (unlikely(PagePrivate(page))) { | |
1945 | long rc = vma_needs_reservation(h, vma, address); | |
1946 | ||
1947 | if (unlikely(rc < 0)) { | |
1948 | /* | |
1949 | * Rare out of memory condition in reserve map | |
1950 | * manipulation. Clear PagePrivate so that | |
1951 | * global reserve count will not be incremented | |
1952 | * by free_huge_page. This will make it appear | |
1953 | * as though the reservation for this page was | |
1954 | * consumed. This may prevent the task from | |
1955 | * faulting in the page at a later time. This | |
1956 | * is better than inconsistent global huge page | |
1957 | * accounting of reserve counts. | |
1958 | */ | |
1959 | ClearPagePrivate(page); | |
1960 | } else if (rc) { | |
1961 | rc = vma_add_reservation(h, vma, address); | |
1962 | if (unlikely(rc < 0)) | |
1963 | /* | |
1964 | * See above comment about rare out of | |
1965 | * memory condition. | |
1966 | */ | |
1967 | ClearPagePrivate(page); | |
1968 | } else | |
1969 | vma_end_reservation(h, vma, address); | |
1970 | } | |
1971 | } | |
1972 | ||
70c3547e | 1973 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 1974 | unsigned long addr, int avoid_reserve) |
1da177e4 | 1975 | { |
90481622 | 1976 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 1977 | struct hstate *h = hstate_vma(vma); |
348ea204 | 1978 | struct page *page; |
d85f69b0 MK |
1979 | long map_chg, map_commit; |
1980 | long gbl_chg; | |
6d76dcf4 AK |
1981 | int ret, idx; |
1982 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 1983 | |
6d76dcf4 | 1984 | idx = hstate_index(h); |
a1e78772 | 1985 | /* |
d85f69b0 MK |
1986 | * Examine the region/reserve map to determine if the process |
1987 | * has a reservation for the page to be allocated. A return | |
1988 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 1989 | */ |
d85f69b0 MK |
1990 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
1991 | if (map_chg < 0) | |
76dcee75 | 1992 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
1993 | |
1994 | /* | |
1995 | * Processes that did not create the mapping will have no | |
1996 | * reserves as indicated by the region/reserve map. Check | |
1997 | * that the allocation will not exceed the subpool limit. | |
1998 | * Allocations for MAP_NORESERVE mappings also need to be | |
1999 | * checked against any subpool limit. | |
2000 | */ | |
2001 | if (map_chg || avoid_reserve) { | |
2002 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2003 | if (gbl_chg < 0) { | |
feba16e2 | 2004 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2005 | return ERR_PTR(-ENOSPC); |
5e911373 | 2006 | } |
1da177e4 | 2007 | |
d85f69b0 MK |
2008 | /* |
2009 | * Even though there was no reservation in the region/reserve | |
2010 | * map, there could be reservations associated with the | |
2011 | * subpool that can be used. This would be indicated if the | |
2012 | * return value of hugepage_subpool_get_pages() is zero. | |
2013 | * However, if avoid_reserve is specified we still avoid even | |
2014 | * the subpool reservations. | |
2015 | */ | |
2016 | if (avoid_reserve) | |
2017 | gbl_chg = 1; | |
2018 | } | |
2019 | ||
6d76dcf4 | 2020 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f JZ |
2021 | if (ret) |
2022 | goto out_subpool_put; | |
2023 | ||
1da177e4 | 2024 | spin_lock(&hugetlb_lock); |
d85f69b0 MK |
2025 | /* |
2026 | * glb_chg is passed to indicate whether or not a page must be taken | |
2027 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2028 | * a reservation exists for the allocation. | |
2029 | */ | |
2030 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2031 | if (!page) { |
94ae8ba7 | 2032 | spin_unlock(&hugetlb_lock); |
099730d6 | 2033 | page = __alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2034 | if (!page) |
2035 | goto out_uncharge_cgroup; | |
a88c7695 NH |
2036 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
2037 | SetPagePrivate(page); | |
2038 | h->resv_huge_pages--; | |
2039 | } | |
79dbb236 AK |
2040 | spin_lock(&hugetlb_lock); |
2041 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 2042 | /* Fall through */ |
68842c9b | 2043 | } |
81a6fcae JK |
2044 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
2045 | spin_unlock(&hugetlb_lock); | |
348ea204 | 2046 | |
90481622 | 2047 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 2048 | |
d85f69b0 MK |
2049 | map_commit = vma_commit_reservation(h, vma, addr); |
2050 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2051 | /* |
2052 | * The page was added to the reservation map between | |
2053 | * vma_needs_reservation and vma_commit_reservation. | |
2054 | * This indicates a race with hugetlb_reserve_pages. | |
2055 | * Adjust for the subpool count incremented above AND | |
2056 | * in hugetlb_reserve_pages for the same page. Also, | |
2057 | * the reservation count added in hugetlb_reserve_pages | |
2058 | * no longer applies. | |
2059 | */ | |
2060 | long rsv_adjust; | |
2061 | ||
2062 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2063 | hugetlb_acct_memory(h, -rsv_adjust); | |
2064 | } | |
90d8b7e6 | 2065 | return page; |
8f34af6f JZ |
2066 | |
2067 | out_uncharge_cgroup: | |
2068 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
2069 | out_subpool_put: | |
d85f69b0 | 2070 | if (map_chg || avoid_reserve) |
8f34af6f | 2071 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2072 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2073 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2074 | } |
2075 | ||
74060e4d NH |
2076 | /* |
2077 | * alloc_huge_page()'s wrapper which simply returns the page if allocation | |
2078 | * succeeds, otherwise NULL. This function is called from new_vma_page(), | |
2079 | * where no ERR_VALUE is expected to be returned. | |
2080 | */ | |
2081 | struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, | |
2082 | unsigned long addr, int avoid_reserve) | |
2083 | { | |
2084 | struct page *page = alloc_huge_page(vma, addr, avoid_reserve); | |
2085 | if (IS_ERR(page)) | |
2086 | page = NULL; | |
2087 | return page; | |
2088 | } | |
2089 | ||
e24a1307 AK |
2090 | int alloc_bootmem_huge_page(struct hstate *h) |
2091 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); | |
2092 | int __alloc_bootmem_huge_page(struct hstate *h) | |
aa888a74 AK |
2093 | { |
2094 | struct huge_bootmem_page *m; | |
b2261026 | 2095 | int nr_nodes, node; |
aa888a74 | 2096 | |
b2261026 | 2097 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
2098 | void *addr; |
2099 | ||
8b89a116 GS |
2100 | addr = memblock_virt_alloc_try_nid_nopanic( |
2101 | huge_page_size(h), huge_page_size(h), | |
2102 | 0, BOOTMEM_ALLOC_ACCESSIBLE, node); | |
aa888a74 AK |
2103 | if (addr) { |
2104 | /* | |
2105 | * Use the beginning of the huge page to store the | |
2106 | * huge_bootmem_page struct (until gather_bootmem | |
2107 | * puts them into the mem_map). | |
2108 | */ | |
2109 | m = addr; | |
91f47662 | 2110 | goto found; |
aa888a74 | 2111 | } |
aa888a74 AK |
2112 | } |
2113 | return 0; | |
2114 | ||
2115 | found: | |
df994ead | 2116 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 AK |
2117 | /* Put them into a private list first because mem_map is not up yet */ |
2118 | list_add(&m->list, &huge_boot_pages); | |
2119 | m->hstate = h; | |
2120 | return 1; | |
2121 | } | |
2122 | ||
d00181b9 KS |
2123 | static void __init prep_compound_huge_page(struct page *page, |
2124 | unsigned int order) | |
18229df5 AW |
2125 | { |
2126 | if (unlikely(order > (MAX_ORDER - 1))) | |
2127 | prep_compound_gigantic_page(page, order); | |
2128 | else | |
2129 | prep_compound_page(page, order); | |
2130 | } | |
2131 | ||
aa888a74 AK |
2132 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
2133 | static void __init gather_bootmem_prealloc(void) | |
2134 | { | |
2135 | struct huge_bootmem_page *m; | |
2136 | ||
2137 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 2138 | struct hstate *h = m->hstate; |
ee8f248d BB |
2139 | struct page *page; |
2140 | ||
2141 | #ifdef CONFIG_HIGHMEM | |
2142 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
8b89a116 GS |
2143 | memblock_free_late(__pa(m), |
2144 | sizeof(struct huge_bootmem_page)); | |
ee8f248d BB |
2145 | #else |
2146 | page = virt_to_page(m); | |
2147 | #endif | |
aa888a74 | 2148 | WARN_ON(page_count(page) != 1); |
18229df5 | 2149 | prep_compound_huge_page(page, h->order); |
ef5a22be | 2150 | WARN_ON(PageReserved(page)); |
aa888a74 | 2151 | prep_new_huge_page(h, page, page_to_nid(page)); |
af0fb9df MH |
2152 | put_page(page); /* free it into the hugepage allocator */ |
2153 | ||
b0320c7b RA |
2154 | /* |
2155 | * If we had gigantic hugepages allocated at boot time, we need | |
2156 | * to restore the 'stolen' pages to totalram_pages in order to | |
2157 | * fix confusing memory reports from free(1) and another | |
2158 | * side-effects, like CommitLimit going negative. | |
2159 | */ | |
bae7f4ae | 2160 | if (hstate_is_gigantic(h)) |
3dcc0571 | 2161 | adjust_managed_page_count(page, 1 << h->order); |
aa888a74 AK |
2162 | } |
2163 | } | |
2164 | ||
8faa8b07 | 2165 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
2166 | { |
2167 | unsigned long i; | |
a5516438 | 2168 | |
e5ff2159 | 2169 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 2170 | if (hstate_is_gigantic(h)) { |
aa888a74 AK |
2171 | if (!alloc_bootmem_huge_page(h)) |
2172 | break; | |
9b5e5d0f | 2173 | } else if (!alloc_fresh_huge_page(h, |
8cebfcd0 | 2174 | &node_states[N_MEMORY])) |
1da177e4 | 2175 | break; |
69ed779a | 2176 | cond_resched(); |
1da177e4 | 2177 | } |
d715cf80 LH |
2178 | if (i < h->max_huge_pages) { |
2179 | char buf[32]; | |
2180 | ||
c6247f72 | 2181 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
2182 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
2183 | h->max_huge_pages, buf, i); | |
2184 | h->max_huge_pages = i; | |
2185 | } | |
e5ff2159 AK |
2186 | } |
2187 | ||
2188 | static void __init hugetlb_init_hstates(void) | |
2189 | { | |
2190 | struct hstate *h; | |
2191 | ||
2192 | for_each_hstate(h) { | |
641844f5 NH |
2193 | if (minimum_order > huge_page_order(h)) |
2194 | minimum_order = huge_page_order(h); | |
2195 | ||
8faa8b07 | 2196 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 2197 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 2198 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 2199 | } |
641844f5 | 2200 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
2201 | } |
2202 | ||
2203 | static void __init report_hugepages(void) | |
2204 | { | |
2205 | struct hstate *h; | |
2206 | ||
2207 | for_each_hstate(h) { | |
4abd32db | 2208 | char buf[32]; |
c6247f72 MW |
2209 | |
2210 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
ffb22af5 | 2211 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 2212 | buf, h->free_huge_pages); |
e5ff2159 AK |
2213 | } |
2214 | } | |
2215 | ||
1da177e4 | 2216 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
2217 | static void try_to_free_low(struct hstate *h, unsigned long count, |
2218 | nodemask_t *nodes_allowed) | |
1da177e4 | 2219 | { |
4415cc8d CL |
2220 | int i; |
2221 | ||
bae7f4ae | 2222 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2223 | return; |
2224 | ||
6ae11b27 | 2225 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 2226 | struct page *page, *next; |
a5516438 AK |
2227 | struct list_head *freel = &h->hugepage_freelists[i]; |
2228 | list_for_each_entry_safe(page, next, freel, lru) { | |
2229 | if (count >= h->nr_huge_pages) | |
6b0c880d | 2230 | return; |
1da177e4 LT |
2231 | if (PageHighMem(page)) |
2232 | continue; | |
2233 | list_del(&page->lru); | |
e5ff2159 | 2234 | update_and_free_page(h, page); |
a5516438 AK |
2235 | h->free_huge_pages--; |
2236 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
2237 | } |
2238 | } | |
2239 | } | |
2240 | #else | |
6ae11b27 LS |
2241 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
2242 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
2243 | { |
2244 | } | |
2245 | #endif | |
2246 | ||
20a0307c WF |
2247 | /* |
2248 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2249 | * balanced by operating on them in a round-robin fashion. | |
2250 | * Returns 1 if an adjustment was made. | |
2251 | */ | |
6ae11b27 LS |
2252 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
2253 | int delta) | |
20a0307c | 2254 | { |
b2261026 | 2255 | int nr_nodes, node; |
20a0307c WF |
2256 | |
2257 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 2258 | |
b2261026 JK |
2259 | if (delta < 0) { |
2260 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2261 | if (h->surplus_huge_pages_node[node]) | |
2262 | goto found; | |
e8c5c824 | 2263 | } |
b2261026 JK |
2264 | } else { |
2265 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2266 | if (h->surplus_huge_pages_node[node] < | |
2267 | h->nr_huge_pages_node[node]) | |
2268 | goto found; | |
e8c5c824 | 2269 | } |
b2261026 JK |
2270 | } |
2271 | return 0; | |
20a0307c | 2272 | |
b2261026 JK |
2273 | found: |
2274 | h->surplus_huge_pages += delta; | |
2275 | h->surplus_huge_pages_node[node] += delta; | |
2276 | return 1; | |
20a0307c WF |
2277 | } |
2278 | ||
a5516438 | 2279 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
2280 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
2281 | nodemask_t *nodes_allowed) | |
1da177e4 | 2282 | { |
7893d1d5 | 2283 | unsigned long min_count, ret; |
1da177e4 | 2284 | |
944d9fec | 2285 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
aa888a74 AK |
2286 | return h->max_huge_pages; |
2287 | ||
7893d1d5 AL |
2288 | /* |
2289 | * Increase the pool size | |
2290 | * First take pages out of surplus state. Then make up the | |
2291 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 2292 | * |
ab5ac90a | 2293 | * We might race with __alloc_surplus_huge_page() here and be unable |
d1c3fb1f NA |
2294 | * to convert a surplus huge page to a normal huge page. That is |
2295 | * not critical, though, it just means the overall size of the | |
2296 | * pool might be one hugepage larger than it needs to be, but | |
2297 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 2298 | */ |
1da177e4 | 2299 | spin_lock(&hugetlb_lock); |
a5516438 | 2300 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 2301 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
2302 | break; |
2303 | } | |
2304 | ||
a5516438 | 2305 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
2306 | /* |
2307 | * If this allocation races such that we no longer need the | |
2308 | * page, free_huge_page will handle it by freeing the page | |
2309 | * and reducing the surplus. | |
2310 | */ | |
2311 | spin_unlock(&hugetlb_lock); | |
649920c6 JH |
2312 | |
2313 | /* yield cpu to avoid soft lockup */ | |
2314 | cond_resched(); | |
2315 | ||
d9cc948f | 2316 | ret = alloc_fresh_huge_page(h, nodes_allowed); |
7893d1d5 AL |
2317 | spin_lock(&hugetlb_lock); |
2318 | if (!ret) | |
2319 | goto out; | |
2320 | ||
536240f2 MG |
2321 | /* Bail for signals. Probably ctrl-c from user */ |
2322 | if (signal_pending(current)) | |
2323 | goto out; | |
7893d1d5 | 2324 | } |
7893d1d5 AL |
2325 | |
2326 | /* | |
2327 | * Decrease the pool size | |
2328 | * First return free pages to the buddy allocator (being careful | |
2329 | * to keep enough around to satisfy reservations). Then place | |
2330 | * pages into surplus state as needed so the pool will shrink | |
2331 | * to the desired size as pages become free. | |
d1c3fb1f NA |
2332 | * |
2333 | * By placing pages into the surplus state independent of the | |
2334 | * overcommit value, we are allowing the surplus pool size to | |
2335 | * exceed overcommit. There are few sane options here. Since | |
ab5ac90a | 2336 | * __alloc_surplus_huge_page() is checking the global counter, |
d1c3fb1f NA |
2337 | * though, we'll note that we're not allowed to exceed surplus |
2338 | * and won't grow the pool anywhere else. Not until one of the | |
2339 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 2340 | */ |
a5516438 | 2341 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 2342 | min_count = max(count, min_count); |
6ae11b27 | 2343 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 2344 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 2345 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 2346 | break; |
55f67141 | 2347 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 2348 | } |
a5516438 | 2349 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 2350 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
2351 | break; |
2352 | } | |
2353 | out: | |
a5516438 | 2354 | ret = persistent_huge_pages(h); |
1da177e4 | 2355 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 2356 | return ret; |
1da177e4 LT |
2357 | } |
2358 | ||
a3437870 NA |
2359 | #define HSTATE_ATTR_RO(_name) \ |
2360 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2361 | ||
2362 | #define HSTATE_ATTR(_name) \ | |
2363 | static struct kobj_attribute _name##_attr = \ | |
2364 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2365 | ||
2366 | static struct kobject *hugepages_kobj; | |
2367 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2368 | ||
9a305230 LS |
2369 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
2370 | ||
2371 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
2372 | { |
2373 | int i; | |
9a305230 | 2374 | |
a3437870 | 2375 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
2376 | if (hstate_kobjs[i] == kobj) { |
2377 | if (nidp) | |
2378 | *nidp = NUMA_NO_NODE; | |
a3437870 | 2379 | return &hstates[i]; |
9a305230 LS |
2380 | } |
2381 | ||
2382 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
2383 | } |
2384 | ||
06808b08 | 2385 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
2386 | struct kobj_attribute *attr, char *buf) |
2387 | { | |
9a305230 LS |
2388 | struct hstate *h; |
2389 | unsigned long nr_huge_pages; | |
2390 | int nid; | |
2391 | ||
2392 | h = kobj_to_hstate(kobj, &nid); | |
2393 | if (nid == NUMA_NO_NODE) | |
2394 | nr_huge_pages = h->nr_huge_pages; | |
2395 | else | |
2396 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2397 | ||
2398 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 2399 | } |
adbe8726 | 2400 | |
238d3c13 DR |
2401 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
2402 | struct hstate *h, int nid, | |
2403 | unsigned long count, size_t len) | |
a3437870 NA |
2404 | { |
2405 | int err; | |
bad44b5b | 2406 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 2407 | |
944d9fec | 2408 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) { |
adbe8726 EM |
2409 | err = -EINVAL; |
2410 | goto out; | |
2411 | } | |
2412 | ||
9a305230 LS |
2413 | if (nid == NUMA_NO_NODE) { |
2414 | /* | |
2415 | * global hstate attribute | |
2416 | */ | |
2417 | if (!(obey_mempolicy && | |
2418 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
2419 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 2420 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 LS |
2421 | } |
2422 | } else if (nodes_allowed) { | |
2423 | /* | |
2424 | * per node hstate attribute: adjust count to global, | |
2425 | * but restrict alloc/free to the specified node. | |
2426 | */ | |
2427 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2428 | init_nodemask_of_node(nodes_allowed, nid); | |
2429 | } else | |
8cebfcd0 | 2430 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 | 2431 | |
06808b08 | 2432 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 2433 | |
8cebfcd0 | 2434 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
2435 | NODEMASK_FREE(nodes_allowed); |
2436 | ||
2437 | return len; | |
adbe8726 EM |
2438 | out: |
2439 | NODEMASK_FREE(nodes_allowed); | |
2440 | return err; | |
06808b08 LS |
2441 | } |
2442 | ||
238d3c13 DR |
2443 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
2444 | struct kobject *kobj, const char *buf, | |
2445 | size_t len) | |
2446 | { | |
2447 | struct hstate *h; | |
2448 | unsigned long count; | |
2449 | int nid; | |
2450 | int err; | |
2451 | ||
2452 | err = kstrtoul(buf, 10, &count); | |
2453 | if (err) | |
2454 | return err; | |
2455 | ||
2456 | h = kobj_to_hstate(kobj, &nid); | |
2457 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2458 | } | |
2459 | ||
06808b08 LS |
2460 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
2461 | struct kobj_attribute *attr, char *buf) | |
2462 | { | |
2463 | return nr_hugepages_show_common(kobj, attr, buf); | |
2464 | } | |
2465 | ||
2466 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2467 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2468 | { | |
238d3c13 | 2469 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
2470 | } |
2471 | HSTATE_ATTR(nr_hugepages); | |
2472 | ||
06808b08 LS |
2473 | #ifdef CONFIG_NUMA |
2474 | ||
2475 | /* | |
2476 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2477 | * huge page alloc/free. | |
2478 | */ | |
2479 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
2480 | struct kobj_attribute *attr, char *buf) | |
2481 | { | |
2482 | return nr_hugepages_show_common(kobj, attr, buf); | |
2483 | } | |
2484 | ||
2485 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2486 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2487 | { | |
238d3c13 | 2488 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
2489 | } |
2490 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2491 | #endif | |
2492 | ||
2493 | ||
a3437870 NA |
2494 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
2495 | struct kobj_attribute *attr, char *buf) | |
2496 | { | |
9a305230 | 2497 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2498 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
2499 | } | |
adbe8726 | 2500 | |
a3437870 NA |
2501 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
2502 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2503 | { | |
2504 | int err; | |
2505 | unsigned long input; | |
9a305230 | 2506 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2507 | |
bae7f4ae | 2508 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2509 | return -EINVAL; |
2510 | ||
3dbb95f7 | 2511 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2512 | if (err) |
73ae31e5 | 2513 | return err; |
a3437870 NA |
2514 | |
2515 | spin_lock(&hugetlb_lock); | |
2516 | h->nr_overcommit_huge_pages = input; | |
2517 | spin_unlock(&hugetlb_lock); | |
2518 | ||
2519 | return count; | |
2520 | } | |
2521 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2522 | ||
2523 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2524 | struct kobj_attribute *attr, char *buf) | |
2525 | { | |
9a305230 LS |
2526 | struct hstate *h; |
2527 | unsigned long free_huge_pages; | |
2528 | int nid; | |
2529 | ||
2530 | h = kobj_to_hstate(kobj, &nid); | |
2531 | if (nid == NUMA_NO_NODE) | |
2532 | free_huge_pages = h->free_huge_pages; | |
2533 | else | |
2534 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2535 | ||
2536 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
2537 | } |
2538 | HSTATE_ATTR_RO(free_hugepages); | |
2539 | ||
2540 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2541 | struct kobj_attribute *attr, char *buf) | |
2542 | { | |
9a305230 | 2543 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2544 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
2545 | } | |
2546 | HSTATE_ATTR_RO(resv_hugepages); | |
2547 | ||
2548 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2549 | struct kobj_attribute *attr, char *buf) | |
2550 | { | |
9a305230 LS |
2551 | struct hstate *h; |
2552 | unsigned long surplus_huge_pages; | |
2553 | int nid; | |
2554 | ||
2555 | h = kobj_to_hstate(kobj, &nid); | |
2556 | if (nid == NUMA_NO_NODE) | |
2557 | surplus_huge_pages = h->surplus_huge_pages; | |
2558 | else | |
2559 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2560 | ||
2561 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
2562 | } |
2563 | HSTATE_ATTR_RO(surplus_hugepages); | |
2564 | ||
2565 | static struct attribute *hstate_attrs[] = { | |
2566 | &nr_hugepages_attr.attr, | |
2567 | &nr_overcommit_hugepages_attr.attr, | |
2568 | &free_hugepages_attr.attr, | |
2569 | &resv_hugepages_attr.attr, | |
2570 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
2571 | #ifdef CONFIG_NUMA |
2572 | &nr_hugepages_mempolicy_attr.attr, | |
2573 | #endif | |
a3437870 NA |
2574 | NULL, |
2575 | }; | |
2576 | ||
67e5ed96 | 2577 | static const struct attribute_group hstate_attr_group = { |
a3437870 NA |
2578 | .attrs = hstate_attrs, |
2579 | }; | |
2580 | ||
094e9539 JM |
2581 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
2582 | struct kobject **hstate_kobjs, | |
67e5ed96 | 2583 | const struct attribute_group *hstate_attr_group) |
a3437870 NA |
2584 | { |
2585 | int retval; | |
972dc4de | 2586 | int hi = hstate_index(h); |
a3437870 | 2587 | |
9a305230 LS |
2588 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
2589 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
2590 | return -ENOMEM; |
2591 | ||
9a305230 | 2592 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 2593 | if (retval) |
9a305230 | 2594 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
2595 | |
2596 | return retval; | |
2597 | } | |
2598 | ||
2599 | static void __init hugetlb_sysfs_init(void) | |
2600 | { | |
2601 | struct hstate *h; | |
2602 | int err; | |
2603 | ||
2604 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
2605 | if (!hugepages_kobj) | |
2606 | return; | |
2607 | ||
2608 | for_each_hstate(h) { | |
9a305230 LS |
2609 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
2610 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 2611 | if (err) |
ffb22af5 | 2612 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
2613 | } |
2614 | } | |
2615 | ||
9a305230 LS |
2616 | #ifdef CONFIG_NUMA |
2617 | ||
2618 | /* | |
2619 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
2620 | * with node devices in node_devices[] using a parallel array. The array |
2621 | * index of a node device or _hstate == node id. | |
2622 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
2623 | * the base kernel, on the hugetlb module. |
2624 | */ | |
2625 | struct node_hstate { | |
2626 | struct kobject *hugepages_kobj; | |
2627 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2628 | }; | |
b4e289a6 | 2629 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
2630 | |
2631 | /* | |
10fbcf4c | 2632 | * A subset of global hstate attributes for node devices |
9a305230 LS |
2633 | */ |
2634 | static struct attribute *per_node_hstate_attrs[] = { | |
2635 | &nr_hugepages_attr.attr, | |
2636 | &free_hugepages_attr.attr, | |
2637 | &surplus_hugepages_attr.attr, | |
2638 | NULL, | |
2639 | }; | |
2640 | ||
67e5ed96 | 2641 | static const struct attribute_group per_node_hstate_attr_group = { |
9a305230 LS |
2642 | .attrs = per_node_hstate_attrs, |
2643 | }; | |
2644 | ||
2645 | /* | |
10fbcf4c | 2646 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
2647 | * Returns node id via non-NULL nidp. |
2648 | */ | |
2649 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2650 | { | |
2651 | int nid; | |
2652 | ||
2653 | for (nid = 0; nid < nr_node_ids; nid++) { | |
2654 | struct node_hstate *nhs = &node_hstates[nid]; | |
2655 | int i; | |
2656 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
2657 | if (nhs->hstate_kobjs[i] == kobj) { | |
2658 | if (nidp) | |
2659 | *nidp = nid; | |
2660 | return &hstates[i]; | |
2661 | } | |
2662 | } | |
2663 | ||
2664 | BUG(); | |
2665 | return NULL; | |
2666 | } | |
2667 | ||
2668 | /* | |
10fbcf4c | 2669 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
2670 | * No-op if no hstate attributes attached. |
2671 | */ | |
3cd8b44f | 2672 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
2673 | { |
2674 | struct hstate *h; | |
10fbcf4c | 2675 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2676 | |
2677 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 2678 | return; /* no hstate attributes */ |
9a305230 | 2679 | |
972dc4de AK |
2680 | for_each_hstate(h) { |
2681 | int idx = hstate_index(h); | |
2682 | if (nhs->hstate_kobjs[idx]) { | |
2683 | kobject_put(nhs->hstate_kobjs[idx]); | |
2684 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 2685 | } |
972dc4de | 2686 | } |
9a305230 LS |
2687 | |
2688 | kobject_put(nhs->hugepages_kobj); | |
2689 | nhs->hugepages_kobj = NULL; | |
2690 | } | |
2691 | ||
9a305230 LS |
2692 | |
2693 | /* | |
10fbcf4c | 2694 | * Register hstate attributes for a single node device. |
9a305230 LS |
2695 | * No-op if attributes already registered. |
2696 | */ | |
3cd8b44f | 2697 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
2698 | { |
2699 | struct hstate *h; | |
10fbcf4c | 2700 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2701 | int err; |
2702 | ||
2703 | if (nhs->hugepages_kobj) | |
2704 | return; /* already allocated */ | |
2705 | ||
2706 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 2707 | &node->dev.kobj); |
9a305230 LS |
2708 | if (!nhs->hugepages_kobj) |
2709 | return; | |
2710 | ||
2711 | for_each_hstate(h) { | |
2712 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
2713 | nhs->hstate_kobjs, | |
2714 | &per_node_hstate_attr_group); | |
2715 | if (err) { | |
ffb22af5 AM |
2716 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
2717 | h->name, node->dev.id); | |
9a305230 LS |
2718 | hugetlb_unregister_node(node); |
2719 | break; | |
2720 | } | |
2721 | } | |
2722 | } | |
2723 | ||
2724 | /* | |
9b5e5d0f | 2725 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
2726 | * devices of nodes that have memory. All on-line nodes should have |
2727 | * registered their associated device by this time. | |
9a305230 | 2728 | */ |
7d9ca000 | 2729 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
2730 | { |
2731 | int nid; | |
2732 | ||
8cebfcd0 | 2733 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 2734 | struct node *node = node_devices[nid]; |
10fbcf4c | 2735 | if (node->dev.id == nid) |
9a305230 LS |
2736 | hugetlb_register_node(node); |
2737 | } | |
2738 | ||
2739 | /* | |
10fbcf4c | 2740 | * Let the node device driver know we're here so it can |
9a305230 LS |
2741 | * [un]register hstate attributes on node hotplug. |
2742 | */ | |
2743 | register_hugetlbfs_with_node(hugetlb_register_node, | |
2744 | hugetlb_unregister_node); | |
2745 | } | |
2746 | #else /* !CONFIG_NUMA */ | |
2747 | ||
2748 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2749 | { | |
2750 | BUG(); | |
2751 | if (nidp) | |
2752 | *nidp = -1; | |
2753 | return NULL; | |
2754 | } | |
2755 | ||
9a305230 LS |
2756 | static void hugetlb_register_all_nodes(void) { } |
2757 | ||
2758 | #endif | |
2759 | ||
a3437870 NA |
2760 | static int __init hugetlb_init(void) |
2761 | { | |
8382d914 DB |
2762 | int i; |
2763 | ||
457c1b27 | 2764 | if (!hugepages_supported()) |
0ef89d25 | 2765 | return 0; |
a3437870 | 2766 | |
e11bfbfc | 2767 | if (!size_to_hstate(default_hstate_size)) { |
d715cf80 LH |
2768 | if (default_hstate_size != 0) { |
2769 | pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n", | |
2770 | default_hstate_size, HPAGE_SIZE); | |
2771 | } | |
2772 | ||
e11bfbfc NP |
2773 | default_hstate_size = HPAGE_SIZE; |
2774 | if (!size_to_hstate(default_hstate_size)) | |
2775 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 2776 | } |
972dc4de | 2777 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
f8b74815 VT |
2778 | if (default_hstate_max_huge_pages) { |
2779 | if (!default_hstate.max_huge_pages) | |
2780 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
2781 | } | |
a3437870 NA |
2782 | |
2783 | hugetlb_init_hstates(); | |
aa888a74 | 2784 | gather_bootmem_prealloc(); |
a3437870 NA |
2785 | report_hugepages(); |
2786 | ||
2787 | hugetlb_sysfs_init(); | |
9a305230 | 2788 | hugetlb_register_all_nodes(); |
7179e7bf | 2789 | hugetlb_cgroup_file_init(); |
9a305230 | 2790 | |
8382d914 DB |
2791 | #ifdef CONFIG_SMP |
2792 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
2793 | #else | |
2794 | num_fault_mutexes = 1; | |
2795 | #endif | |
c672c7f2 | 2796 | hugetlb_fault_mutex_table = |
8382d914 | 2797 | kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); |
c672c7f2 | 2798 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
2799 | |
2800 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 2801 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
2802 | return 0; |
2803 | } | |
3e89e1c5 | 2804 | subsys_initcall(hugetlb_init); |
a3437870 NA |
2805 | |
2806 | /* Should be called on processing a hugepagesz=... option */ | |
9fee021d VT |
2807 | void __init hugetlb_bad_size(void) |
2808 | { | |
2809 | parsed_valid_hugepagesz = false; | |
2810 | } | |
2811 | ||
d00181b9 | 2812 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
2813 | { |
2814 | struct hstate *h; | |
8faa8b07 AK |
2815 | unsigned long i; |
2816 | ||
a3437870 | 2817 | if (size_to_hstate(PAGE_SIZE << order)) { |
598d8091 | 2818 | pr_warn("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
2819 | return; |
2820 | } | |
47d38344 | 2821 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 2822 | BUG_ON(order == 0); |
47d38344 | 2823 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
2824 | h->order = order; |
2825 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
2826 | h->nr_huge_pages = 0; |
2827 | h->free_huge_pages = 0; | |
2828 | for (i = 0; i < MAX_NUMNODES; ++i) | |
2829 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 2830 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
2831 | h->next_nid_to_alloc = first_memory_node; |
2832 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
2833 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
2834 | huge_page_size(h)/1024); | |
8faa8b07 | 2835 | |
a3437870 NA |
2836 | parsed_hstate = h; |
2837 | } | |
2838 | ||
e11bfbfc | 2839 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
2840 | { |
2841 | unsigned long *mhp; | |
8faa8b07 | 2842 | static unsigned long *last_mhp; |
a3437870 | 2843 | |
9fee021d VT |
2844 | if (!parsed_valid_hugepagesz) { |
2845 | pr_warn("hugepages = %s preceded by " | |
2846 | "an unsupported hugepagesz, ignoring\n", s); | |
2847 | parsed_valid_hugepagesz = true; | |
2848 | return 1; | |
2849 | } | |
a3437870 | 2850 | /* |
47d38344 | 2851 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
2852 | * so this hugepages= parameter goes to the "default hstate". |
2853 | */ | |
9fee021d | 2854 | else if (!hugetlb_max_hstate) |
a3437870 NA |
2855 | mhp = &default_hstate_max_huge_pages; |
2856 | else | |
2857 | mhp = &parsed_hstate->max_huge_pages; | |
2858 | ||
8faa8b07 | 2859 | if (mhp == last_mhp) { |
598d8091 | 2860 | pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); |
8faa8b07 AK |
2861 | return 1; |
2862 | } | |
2863 | ||
a3437870 NA |
2864 | if (sscanf(s, "%lu", mhp) <= 0) |
2865 | *mhp = 0; | |
2866 | ||
8faa8b07 AK |
2867 | /* |
2868 | * Global state is always initialized later in hugetlb_init. | |
2869 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
2870 | * use the bootmem allocator. | |
2871 | */ | |
47d38344 | 2872 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
2873 | hugetlb_hstate_alloc_pages(parsed_hstate); |
2874 | ||
2875 | last_mhp = mhp; | |
2876 | ||
a3437870 NA |
2877 | return 1; |
2878 | } | |
e11bfbfc NP |
2879 | __setup("hugepages=", hugetlb_nrpages_setup); |
2880 | ||
2881 | static int __init hugetlb_default_setup(char *s) | |
2882 | { | |
2883 | default_hstate_size = memparse(s, &s); | |
2884 | return 1; | |
2885 | } | |
2886 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 2887 | |
8a213460 NA |
2888 | static unsigned int cpuset_mems_nr(unsigned int *array) |
2889 | { | |
2890 | int node; | |
2891 | unsigned int nr = 0; | |
2892 | ||
2893 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
2894 | nr += array[node]; | |
2895 | ||
2896 | return nr; | |
2897 | } | |
2898 | ||
2899 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
2900 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
2901 | struct ctl_table *table, int write, | |
2902 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 2903 | { |
e5ff2159 | 2904 | struct hstate *h = &default_hstate; |
238d3c13 | 2905 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 2906 | int ret; |
e5ff2159 | 2907 | |
457c1b27 | 2908 | if (!hugepages_supported()) |
86613628 | 2909 | return -EOPNOTSUPP; |
457c1b27 | 2910 | |
e5ff2159 AK |
2911 | table->data = &tmp; |
2912 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2913 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2914 | if (ret) | |
2915 | goto out; | |
e5ff2159 | 2916 | |
238d3c13 DR |
2917 | if (write) |
2918 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
2919 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
2920 | out: |
2921 | return ret; | |
1da177e4 | 2922 | } |
396faf03 | 2923 | |
06808b08 LS |
2924 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2925 | void __user *buffer, size_t *length, loff_t *ppos) | |
2926 | { | |
2927 | ||
2928 | return hugetlb_sysctl_handler_common(false, table, write, | |
2929 | buffer, length, ppos); | |
2930 | } | |
2931 | ||
2932 | #ifdef CONFIG_NUMA | |
2933 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2934 | void __user *buffer, size_t *length, loff_t *ppos) | |
2935 | { | |
2936 | return hugetlb_sysctl_handler_common(true, table, write, | |
2937 | buffer, length, ppos); | |
2938 | } | |
2939 | #endif /* CONFIG_NUMA */ | |
2940 | ||
a3d0c6aa | 2941 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 2942 | void __user *buffer, |
a3d0c6aa NA |
2943 | size_t *length, loff_t *ppos) |
2944 | { | |
a5516438 | 2945 | struct hstate *h = &default_hstate; |
e5ff2159 | 2946 | unsigned long tmp; |
08d4a246 | 2947 | int ret; |
e5ff2159 | 2948 | |
457c1b27 | 2949 | if (!hugepages_supported()) |
86613628 | 2950 | return -EOPNOTSUPP; |
457c1b27 | 2951 | |
c033a93c | 2952 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 2953 | |
bae7f4ae | 2954 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
2955 | return -EINVAL; |
2956 | ||
e5ff2159 AK |
2957 | table->data = &tmp; |
2958 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2959 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2960 | if (ret) | |
2961 | goto out; | |
e5ff2159 AK |
2962 | |
2963 | if (write) { | |
2964 | spin_lock(&hugetlb_lock); | |
2965 | h->nr_overcommit_huge_pages = tmp; | |
2966 | spin_unlock(&hugetlb_lock); | |
2967 | } | |
08d4a246 MH |
2968 | out: |
2969 | return ret; | |
a3d0c6aa NA |
2970 | } |
2971 | ||
1da177e4 LT |
2972 | #endif /* CONFIG_SYSCTL */ |
2973 | ||
e1759c21 | 2974 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 2975 | { |
fcb2b0c5 RG |
2976 | struct hstate *h; |
2977 | unsigned long total = 0; | |
2978 | ||
457c1b27 NA |
2979 | if (!hugepages_supported()) |
2980 | return; | |
fcb2b0c5 RG |
2981 | |
2982 | for_each_hstate(h) { | |
2983 | unsigned long count = h->nr_huge_pages; | |
2984 | ||
2985 | total += (PAGE_SIZE << huge_page_order(h)) * count; | |
2986 | ||
2987 | if (h == &default_hstate) | |
2988 | seq_printf(m, | |
2989 | "HugePages_Total: %5lu\n" | |
2990 | "HugePages_Free: %5lu\n" | |
2991 | "HugePages_Rsvd: %5lu\n" | |
2992 | "HugePages_Surp: %5lu\n" | |
2993 | "Hugepagesize: %8lu kB\n", | |
2994 | count, | |
2995 | h->free_huge_pages, | |
2996 | h->resv_huge_pages, | |
2997 | h->surplus_huge_pages, | |
2998 | (PAGE_SIZE << huge_page_order(h)) / 1024); | |
2999 | } | |
3000 | ||
3001 | seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); | |
1da177e4 LT |
3002 | } |
3003 | ||
3004 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
3005 | { | |
a5516438 | 3006 | struct hstate *h = &default_hstate; |
457c1b27 NA |
3007 | if (!hugepages_supported()) |
3008 | return 0; | |
1da177e4 LT |
3009 | return sprintf(buf, |
3010 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
3011 | "Node %d HugePages_Free: %5u\n" |
3012 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
3013 | nid, h->nr_huge_pages_node[nid], |
3014 | nid, h->free_huge_pages_node[nid], | |
3015 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
3016 | } |
3017 | ||
949f7ec5 DR |
3018 | void hugetlb_show_meminfo(void) |
3019 | { | |
3020 | struct hstate *h; | |
3021 | int nid; | |
3022 | ||
457c1b27 NA |
3023 | if (!hugepages_supported()) |
3024 | return; | |
3025 | ||
949f7ec5 DR |
3026 | for_each_node_state(nid, N_MEMORY) |
3027 | for_each_hstate(h) | |
3028 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
3029 | nid, | |
3030 | h->nr_huge_pages_node[nid], | |
3031 | h->free_huge_pages_node[nid], | |
3032 | h->surplus_huge_pages_node[nid], | |
3033 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
3034 | } | |
3035 | ||
5d317b2b NH |
3036 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
3037 | { | |
3038 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
3039 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
3040 | } | |
3041 | ||
1da177e4 LT |
3042 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
3043 | unsigned long hugetlb_total_pages(void) | |
3044 | { | |
d0028588 WL |
3045 | struct hstate *h; |
3046 | unsigned long nr_total_pages = 0; | |
3047 | ||
3048 | for_each_hstate(h) | |
3049 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
3050 | return nr_total_pages; | |
1da177e4 | 3051 | } |
1da177e4 | 3052 | |
a5516438 | 3053 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
3054 | { |
3055 | int ret = -ENOMEM; | |
3056 | ||
3057 | spin_lock(&hugetlb_lock); | |
3058 | /* | |
3059 | * When cpuset is configured, it breaks the strict hugetlb page | |
3060 | * reservation as the accounting is done on a global variable. Such | |
3061 | * reservation is completely rubbish in the presence of cpuset because | |
3062 | * the reservation is not checked against page availability for the | |
3063 | * current cpuset. Application can still potentially OOM'ed by kernel | |
3064 | * with lack of free htlb page in cpuset that the task is in. | |
3065 | * Attempt to enforce strict accounting with cpuset is almost | |
3066 | * impossible (or too ugly) because cpuset is too fluid that | |
3067 | * task or memory node can be dynamically moved between cpusets. | |
3068 | * | |
3069 | * The change of semantics for shared hugetlb mapping with cpuset is | |
3070 | * undesirable. However, in order to preserve some of the semantics, | |
3071 | * we fall back to check against current free page availability as | |
3072 | * a best attempt and hopefully to minimize the impact of changing | |
3073 | * semantics that cpuset has. | |
3074 | */ | |
3075 | if (delta > 0) { | |
a5516438 | 3076 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
3077 | goto out; |
3078 | ||
a5516438 AK |
3079 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
3080 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
3081 | goto out; |
3082 | } | |
3083 | } | |
3084 | ||
3085 | ret = 0; | |
3086 | if (delta < 0) | |
a5516438 | 3087 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
3088 | |
3089 | out: | |
3090 | spin_unlock(&hugetlb_lock); | |
3091 | return ret; | |
3092 | } | |
3093 | ||
84afd99b AW |
3094 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
3095 | { | |
f522c3ac | 3096 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
3097 | |
3098 | /* | |
3099 | * This new VMA should share its siblings reservation map if present. | |
3100 | * The VMA will only ever have a valid reservation map pointer where | |
3101 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 3102 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
3103 | * after this open call completes. It is therefore safe to take a |
3104 | * new reference here without additional locking. | |
3105 | */ | |
4e35f483 | 3106 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 3107 | kref_get(&resv->refs); |
84afd99b AW |
3108 | } |
3109 | ||
a1e78772 MG |
3110 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
3111 | { | |
a5516438 | 3112 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 3113 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 3114 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 3115 | unsigned long reserve, start, end; |
1c5ecae3 | 3116 | long gbl_reserve; |
84afd99b | 3117 | |
4e35f483 JK |
3118 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3119 | return; | |
84afd99b | 3120 | |
4e35f483 JK |
3121 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
3122 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 3123 | |
4e35f483 | 3124 | reserve = (end - start) - region_count(resv, start, end); |
84afd99b | 3125 | |
4e35f483 JK |
3126 | kref_put(&resv->refs, resv_map_release); |
3127 | ||
3128 | if (reserve) { | |
1c5ecae3 MK |
3129 | /* |
3130 | * Decrement reserve counts. The global reserve count may be | |
3131 | * adjusted if the subpool has a minimum size. | |
3132 | */ | |
3133 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
3134 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 3135 | } |
a1e78772 MG |
3136 | } |
3137 | ||
31383c68 DW |
3138 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) |
3139 | { | |
3140 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
3141 | return -EINVAL; | |
3142 | return 0; | |
3143 | } | |
3144 | ||
1da177e4 LT |
3145 | /* |
3146 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
3147 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
3148 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
3149 | * this far. | |
3150 | */ | |
11bac800 | 3151 | static int hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
3152 | { |
3153 | BUG(); | |
d0217ac0 | 3154 | return 0; |
1da177e4 LT |
3155 | } |
3156 | ||
f0f37e2f | 3157 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 3158 | .fault = hugetlb_vm_op_fault, |
84afd99b | 3159 | .open = hugetlb_vm_op_open, |
a1e78772 | 3160 | .close = hugetlb_vm_op_close, |
31383c68 | 3161 | .split = hugetlb_vm_op_split, |
1da177e4 LT |
3162 | }; |
3163 | ||
1e8f889b DG |
3164 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
3165 | int writable) | |
63551ae0 DG |
3166 | { |
3167 | pte_t entry; | |
3168 | ||
1e8f889b | 3169 | if (writable) { |
106c992a GS |
3170 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
3171 | vma->vm_page_prot))); | |
63551ae0 | 3172 | } else { |
106c992a GS |
3173 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
3174 | vma->vm_page_prot)); | |
63551ae0 DG |
3175 | } |
3176 | entry = pte_mkyoung(entry); | |
3177 | entry = pte_mkhuge(entry); | |
d9ed9faa | 3178 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
3179 | |
3180 | return entry; | |
3181 | } | |
3182 | ||
1e8f889b DG |
3183 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
3184 | unsigned long address, pte_t *ptep) | |
3185 | { | |
3186 | pte_t entry; | |
3187 | ||
106c992a | 3188 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 3189 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 3190 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
3191 | } |
3192 | ||
d5ed7444 | 3193 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
3194 | { |
3195 | swp_entry_t swp; | |
3196 | ||
3197 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 3198 | return false; |
4a705fef NH |
3199 | swp = pte_to_swp_entry(pte); |
3200 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
d5ed7444 | 3201 | return true; |
4a705fef | 3202 | else |
d5ed7444 | 3203 | return false; |
4a705fef NH |
3204 | } |
3205 | ||
3206 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
3207 | { | |
3208 | swp_entry_t swp; | |
3209 | ||
3210 | if (huge_pte_none(pte) || pte_present(pte)) | |
3211 | return 0; | |
3212 | swp = pte_to_swp_entry(pte); | |
3213 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
3214 | return 1; | |
3215 | else | |
3216 | return 0; | |
3217 | } | |
1e8f889b | 3218 | |
63551ae0 DG |
3219 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
3220 | struct vm_area_struct *vma) | |
3221 | { | |
3222 | pte_t *src_pte, *dst_pte, entry; | |
3223 | struct page *ptepage; | |
1c59827d | 3224 | unsigned long addr; |
1e8f889b | 3225 | int cow; |
a5516438 AK |
3226 | struct hstate *h = hstate_vma(vma); |
3227 | unsigned long sz = huge_page_size(h); | |
e8569dd2 AS |
3228 | unsigned long mmun_start; /* For mmu_notifiers */ |
3229 | unsigned long mmun_end; /* For mmu_notifiers */ | |
3230 | int ret = 0; | |
1e8f889b DG |
3231 | |
3232 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 3233 | |
e8569dd2 AS |
3234 | mmun_start = vma->vm_start; |
3235 | mmun_end = vma->vm_end; | |
3236 | if (cow) | |
3237 | mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); | |
3238 | ||
a5516438 | 3239 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 3240 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 3241 | src_pte = huge_pte_offset(src, addr, sz); |
c74df32c HD |
3242 | if (!src_pte) |
3243 | continue; | |
a5516438 | 3244 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
3245 | if (!dst_pte) { |
3246 | ret = -ENOMEM; | |
3247 | break; | |
3248 | } | |
c5c99429 LW |
3249 | |
3250 | /* If the pagetables are shared don't copy or take references */ | |
3251 | if (dst_pte == src_pte) | |
3252 | continue; | |
3253 | ||
cb900f41 KS |
3254 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
3255 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3256 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef NH |
3257 | entry = huge_ptep_get(src_pte); |
3258 | if (huge_pte_none(entry)) { /* skip none entry */ | |
3259 | ; | |
3260 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
3261 | is_hugetlb_entry_hwpoisoned(entry))) { | |
3262 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
3263 | ||
3264 | if (is_write_migration_entry(swp_entry) && cow) { | |
3265 | /* | |
3266 | * COW mappings require pages in both | |
3267 | * parent and child to be set to read. | |
3268 | */ | |
3269 | make_migration_entry_read(&swp_entry); | |
3270 | entry = swp_entry_to_pte(swp_entry); | |
e5251fd4 PA |
3271 | set_huge_swap_pte_at(src, addr, src_pte, |
3272 | entry, sz); | |
4a705fef | 3273 | } |
e5251fd4 | 3274 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); |
4a705fef | 3275 | } else { |
34ee645e | 3276 | if (cow) { |
0f10851e JG |
3277 | /* |
3278 | * No need to notify as we are downgrading page | |
3279 | * table protection not changing it to point | |
3280 | * to a new page. | |
3281 | * | |
3282 | * See Documentation/vm/mmu_notifier.txt | |
3283 | */ | |
7f2e9525 | 3284 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e | 3285 | } |
0253d634 | 3286 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
3287 | ptepage = pte_page(entry); |
3288 | get_page(ptepage); | |
53f9263b | 3289 | page_dup_rmap(ptepage, true); |
1c59827d | 3290 | set_huge_pte_at(dst, addr, dst_pte, entry); |
5d317b2b | 3291 | hugetlb_count_add(pages_per_huge_page(h), dst); |
1c59827d | 3292 | } |
cb900f41 KS |
3293 | spin_unlock(src_ptl); |
3294 | spin_unlock(dst_ptl); | |
63551ae0 | 3295 | } |
63551ae0 | 3296 | |
e8569dd2 AS |
3297 | if (cow) |
3298 | mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); | |
3299 | ||
3300 | return ret; | |
63551ae0 DG |
3301 | } |
3302 | ||
24669e58 AK |
3303 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
3304 | unsigned long start, unsigned long end, | |
3305 | struct page *ref_page) | |
63551ae0 DG |
3306 | { |
3307 | struct mm_struct *mm = vma->vm_mm; | |
3308 | unsigned long address; | |
c7546f8f | 3309 | pte_t *ptep; |
63551ae0 | 3310 | pte_t pte; |
cb900f41 | 3311 | spinlock_t *ptl; |
63551ae0 | 3312 | struct page *page; |
a5516438 AK |
3313 | struct hstate *h = hstate_vma(vma); |
3314 | unsigned long sz = huge_page_size(h); | |
2ec74c3e SG |
3315 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
3316 | const unsigned long mmun_end = end; /* For mmu_notifiers */ | |
a5516438 | 3317 | |
63551ae0 | 3318 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
3319 | BUG_ON(start & ~huge_page_mask(h)); |
3320 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 3321 | |
07e32661 AK |
3322 | /* |
3323 | * This is a hugetlb vma, all the pte entries should point | |
3324 | * to huge page. | |
3325 | */ | |
3326 | tlb_remove_check_page_size_change(tlb, sz); | |
24669e58 | 3327 | tlb_start_vma(tlb, vma); |
2ec74c3e | 3328 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
569f48b8 | 3329 | address = start; |
569f48b8 | 3330 | for (; address < end; address += sz) { |
7868a208 | 3331 | ptep = huge_pte_offset(mm, address, sz); |
4c887265 | 3332 | if (!ptep) |
c7546f8f DG |
3333 | continue; |
3334 | ||
cb900f41 | 3335 | ptl = huge_pte_lock(h, mm, ptep); |
31d49da5 AK |
3336 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3337 | spin_unlock(ptl); | |
3338 | continue; | |
3339 | } | |
39dde65c | 3340 | |
6629326b | 3341 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
3342 | if (huge_pte_none(pte)) { |
3343 | spin_unlock(ptl); | |
3344 | continue; | |
3345 | } | |
6629326b HD |
3346 | |
3347 | /* | |
9fbc1f63 NH |
3348 | * Migrating hugepage or HWPoisoned hugepage is already |
3349 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 3350 | */ |
9fbc1f63 | 3351 | if (unlikely(!pte_present(pte))) { |
9386fac3 | 3352 | huge_pte_clear(mm, address, ptep, sz); |
31d49da5 AK |
3353 | spin_unlock(ptl); |
3354 | continue; | |
8c4894c6 | 3355 | } |
6629326b HD |
3356 | |
3357 | page = pte_page(pte); | |
04f2cbe3 MG |
3358 | /* |
3359 | * If a reference page is supplied, it is because a specific | |
3360 | * page is being unmapped, not a range. Ensure the page we | |
3361 | * are about to unmap is the actual page of interest. | |
3362 | */ | |
3363 | if (ref_page) { | |
31d49da5 AK |
3364 | if (page != ref_page) { |
3365 | spin_unlock(ptl); | |
3366 | continue; | |
3367 | } | |
04f2cbe3 MG |
3368 | /* |
3369 | * Mark the VMA as having unmapped its page so that | |
3370 | * future faults in this VMA will fail rather than | |
3371 | * looking like data was lost | |
3372 | */ | |
3373 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
3374 | } | |
3375 | ||
c7546f8f | 3376 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 3377 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 3378 | if (huge_pte_dirty(pte)) |
6649a386 | 3379 | set_page_dirty(page); |
9e81130b | 3380 | |
5d317b2b | 3381 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
d281ee61 | 3382 | page_remove_rmap(page, true); |
31d49da5 | 3383 | |
cb900f41 | 3384 | spin_unlock(ptl); |
e77b0852 | 3385 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
3386 | /* |
3387 | * Bail out after unmapping reference page if supplied | |
3388 | */ | |
3389 | if (ref_page) | |
3390 | break; | |
fe1668ae | 3391 | } |
2ec74c3e | 3392 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
24669e58 | 3393 | tlb_end_vma(tlb, vma); |
1da177e4 | 3394 | } |
63551ae0 | 3395 | |
d833352a MG |
3396 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
3397 | struct vm_area_struct *vma, unsigned long start, | |
3398 | unsigned long end, struct page *ref_page) | |
3399 | { | |
3400 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
3401 | ||
3402 | /* | |
3403 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
3404 | * test will fail on a vma being torn down, and not grab a page table | |
3405 | * on its way out. We're lucky that the flag has such an appropriate | |
3406 | * name, and can in fact be safely cleared here. We could clear it | |
3407 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 3408 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 3409 | * because in the context this is called, the VMA is about to be |
c8c06efa | 3410 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
3411 | */ |
3412 | vma->vm_flags &= ~VM_MAYSHARE; | |
3413 | } | |
3414 | ||
502717f4 | 3415 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 3416 | unsigned long end, struct page *ref_page) |
502717f4 | 3417 | { |
24669e58 AK |
3418 | struct mm_struct *mm; |
3419 | struct mmu_gather tlb; | |
3420 | ||
3421 | mm = vma->vm_mm; | |
3422 | ||
2b047252 | 3423 | tlb_gather_mmu(&tlb, mm, start, end); |
24669e58 AK |
3424 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
3425 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 KC |
3426 | } |
3427 | ||
04f2cbe3 MG |
3428 | /* |
3429 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
3430 | * mappping it owns the reserve page for. The intention is to unmap the page | |
3431 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
3432 | * same region. | |
3433 | */ | |
2f4612af DB |
3434 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
3435 | struct page *page, unsigned long address) | |
04f2cbe3 | 3436 | { |
7526674d | 3437 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
3438 | struct vm_area_struct *iter_vma; |
3439 | struct address_space *mapping; | |
04f2cbe3 MG |
3440 | pgoff_t pgoff; |
3441 | ||
3442 | /* | |
3443 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
3444 | * from page cache lookup which is in HPAGE_SIZE units. | |
3445 | */ | |
7526674d | 3446 | address = address & huge_page_mask(h); |
36e4f20a MH |
3447 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
3448 | vma->vm_pgoff; | |
93c76a3d | 3449 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 3450 | |
4eb2b1dc MG |
3451 | /* |
3452 | * Take the mapping lock for the duration of the table walk. As | |
3453 | * this mapping should be shared between all the VMAs, | |
3454 | * __unmap_hugepage_range() is called as the lock is already held | |
3455 | */ | |
83cde9e8 | 3456 | i_mmap_lock_write(mapping); |
6b2dbba8 | 3457 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
3458 | /* Do not unmap the current VMA */ |
3459 | if (iter_vma == vma) | |
3460 | continue; | |
3461 | ||
2f84a899 MG |
3462 | /* |
3463 | * Shared VMAs have their own reserves and do not affect | |
3464 | * MAP_PRIVATE accounting but it is possible that a shared | |
3465 | * VMA is using the same page so check and skip such VMAs. | |
3466 | */ | |
3467 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
3468 | continue; | |
3469 | ||
04f2cbe3 MG |
3470 | /* |
3471 | * Unmap the page from other VMAs without their own reserves. | |
3472 | * They get marked to be SIGKILLed if they fault in these | |
3473 | * areas. This is because a future no-page fault on this VMA | |
3474 | * could insert a zeroed page instead of the data existing | |
3475 | * from the time of fork. This would look like data corruption | |
3476 | */ | |
3477 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
3478 | unmap_hugepage_range(iter_vma, address, |
3479 | address + huge_page_size(h), page); | |
04f2cbe3 | 3480 | } |
83cde9e8 | 3481 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
3482 | } |
3483 | ||
0fe6e20b NH |
3484 | /* |
3485 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
3486 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
3487 | * cannot race with other handlers or page migration. | |
3488 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 3489 | */ |
1e8f889b | 3490 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
3999f52e AK |
3491 | unsigned long address, pte_t *ptep, |
3492 | struct page *pagecache_page, spinlock_t *ptl) | |
1e8f889b | 3493 | { |
3999f52e | 3494 | pte_t pte; |
a5516438 | 3495 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 3496 | struct page *old_page, *new_page; |
ad4404a2 | 3497 | int ret = 0, outside_reserve = 0; |
2ec74c3e SG |
3498 | unsigned long mmun_start; /* For mmu_notifiers */ |
3499 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1e8f889b | 3500 | |
3999f52e | 3501 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
3502 | old_page = pte_page(pte); |
3503 | ||
04f2cbe3 | 3504 | retry_avoidcopy: |
1e8f889b DG |
3505 | /* If no-one else is actually using this page, avoid the copy |
3506 | * and just make the page writable */ | |
37a2140d | 3507 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
5a49973d | 3508 | page_move_anon_rmap(old_page, vma); |
1e8f889b | 3509 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 3510 | return 0; |
1e8f889b DG |
3511 | } |
3512 | ||
04f2cbe3 MG |
3513 | /* |
3514 | * If the process that created a MAP_PRIVATE mapping is about to | |
3515 | * perform a COW due to a shared page count, attempt to satisfy | |
3516 | * the allocation without using the existing reserves. The pagecache | |
3517 | * page is used to determine if the reserve at this address was | |
3518 | * consumed or not. If reserves were used, a partial faulted mapping | |
3519 | * at the time of fork() could consume its reserves on COW instead | |
3520 | * of the full address range. | |
3521 | */ | |
5944d011 | 3522 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
3523 | old_page != pagecache_page) |
3524 | outside_reserve = 1; | |
3525 | ||
09cbfeaf | 3526 | get_page(old_page); |
b76c8cfb | 3527 | |
ad4404a2 DB |
3528 | /* |
3529 | * Drop page table lock as buddy allocator may be called. It will | |
3530 | * be acquired again before returning to the caller, as expected. | |
3531 | */ | |
cb900f41 | 3532 | spin_unlock(ptl); |
04f2cbe3 | 3533 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 3534 | |
2fc39cec | 3535 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
3536 | /* |
3537 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
3538 | * it is due to references held by a child and an insufficient | |
3539 | * huge page pool. To guarantee the original mappers | |
3540 | * reliability, unmap the page from child processes. The child | |
3541 | * may get SIGKILLed if it later faults. | |
3542 | */ | |
3543 | if (outside_reserve) { | |
09cbfeaf | 3544 | put_page(old_page); |
04f2cbe3 | 3545 | BUG_ON(huge_pte_none(pte)); |
2f4612af DB |
3546 | unmap_ref_private(mm, vma, old_page, address); |
3547 | BUG_ON(huge_pte_none(pte)); | |
3548 | spin_lock(ptl); | |
7868a208 PA |
3549 | ptep = huge_pte_offset(mm, address & huge_page_mask(h), |
3550 | huge_page_size(h)); | |
2f4612af DB |
3551 | if (likely(ptep && |
3552 | pte_same(huge_ptep_get(ptep), pte))) | |
3553 | goto retry_avoidcopy; | |
3554 | /* | |
3555 | * race occurs while re-acquiring page table | |
3556 | * lock, and our job is done. | |
3557 | */ | |
3558 | return 0; | |
04f2cbe3 MG |
3559 | } |
3560 | ||
ad4404a2 DB |
3561 | ret = (PTR_ERR(new_page) == -ENOMEM) ? |
3562 | VM_FAULT_OOM : VM_FAULT_SIGBUS; | |
3563 | goto out_release_old; | |
1e8f889b DG |
3564 | } |
3565 | ||
0fe6e20b NH |
3566 | /* |
3567 | * When the original hugepage is shared one, it does not have | |
3568 | * anon_vma prepared. | |
3569 | */ | |
44e2aa93 | 3570 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
3571 | ret = VM_FAULT_OOM; |
3572 | goto out_release_all; | |
44e2aa93 | 3573 | } |
0fe6e20b | 3574 | |
47ad8475 AA |
3575 | copy_user_huge_page(new_page, old_page, address, vma, |
3576 | pages_per_huge_page(h)); | |
0ed361de | 3577 | __SetPageUptodate(new_page); |
bcc54222 | 3578 | set_page_huge_active(new_page); |
1e8f889b | 3579 | |
2ec74c3e SG |
3580 | mmun_start = address & huge_page_mask(h); |
3581 | mmun_end = mmun_start + huge_page_size(h); | |
3582 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
ad4404a2 | 3583 | |
b76c8cfb | 3584 | /* |
cb900f41 | 3585 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
3586 | * before the page tables are altered |
3587 | */ | |
cb900f41 | 3588 | spin_lock(ptl); |
7868a208 PA |
3589 | ptep = huge_pte_offset(mm, address & huge_page_mask(h), |
3590 | huge_page_size(h)); | |
a9af0c5d | 3591 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
3592 | ClearPagePrivate(new_page); |
3593 | ||
1e8f889b | 3594 | /* Break COW */ |
8fe627ec | 3595 | huge_ptep_clear_flush(vma, address, ptep); |
34ee645e | 3596 | mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); |
1e8f889b DG |
3597 | set_huge_pte_at(mm, address, ptep, |
3598 | make_huge_pte(vma, new_page, 1)); | |
d281ee61 | 3599 | page_remove_rmap(old_page, true); |
cd67f0d2 | 3600 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
3601 | /* Make the old page be freed below */ |
3602 | new_page = old_page; | |
3603 | } | |
cb900f41 | 3604 | spin_unlock(ptl); |
2ec74c3e | 3605 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
ad4404a2 | 3606 | out_release_all: |
96b96a96 | 3607 | restore_reserve_on_error(h, vma, address, new_page); |
09cbfeaf | 3608 | put_page(new_page); |
ad4404a2 | 3609 | out_release_old: |
09cbfeaf | 3610 | put_page(old_page); |
8312034f | 3611 | |
ad4404a2 DB |
3612 | spin_lock(ptl); /* Caller expects lock to be held */ |
3613 | return ret; | |
1e8f889b DG |
3614 | } |
3615 | ||
04f2cbe3 | 3616 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
3617 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
3618 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
3619 | { |
3620 | struct address_space *mapping; | |
e7c4b0bf | 3621 | pgoff_t idx; |
04f2cbe3 MG |
3622 | |
3623 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 3624 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
3625 | |
3626 | return find_lock_page(mapping, idx); | |
3627 | } | |
3628 | ||
3ae77f43 HD |
3629 | /* |
3630 | * Return whether there is a pagecache page to back given address within VMA. | |
3631 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
3632 | */ | |
3633 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
3634 | struct vm_area_struct *vma, unsigned long address) |
3635 | { | |
3636 | struct address_space *mapping; | |
3637 | pgoff_t idx; | |
3638 | struct page *page; | |
3639 | ||
3640 | mapping = vma->vm_file->f_mapping; | |
3641 | idx = vma_hugecache_offset(h, vma, address); | |
3642 | ||
3643 | page = find_get_page(mapping, idx); | |
3644 | if (page) | |
3645 | put_page(page); | |
3646 | return page != NULL; | |
3647 | } | |
3648 | ||
ab76ad54 MK |
3649 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
3650 | pgoff_t idx) | |
3651 | { | |
3652 | struct inode *inode = mapping->host; | |
3653 | struct hstate *h = hstate_inode(inode); | |
3654 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
3655 | ||
3656 | if (err) | |
3657 | return err; | |
3658 | ClearPagePrivate(page); | |
3659 | ||
3660 | spin_lock(&inode->i_lock); | |
3661 | inode->i_blocks += blocks_per_huge_page(h); | |
3662 | spin_unlock(&inode->i_lock); | |
3663 | return 0; | |
3664 | } | |
3665 | ||
a1ed3dda | 3666 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
8382d914 DB |
3667 | struct address_space *mapping, pgoff_t idx, |
3668 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 3669 | { |
a5516438 | 3670 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 3671 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 3672 | int anon_rmap = 0; |
4c887265 | 3673 | unsigned long size; |
4c887265 | 3674 | struct page *page; |
1e8f889b | 3675 | pte_t new_pte; |
cb900f41 | 3676 | spinlock_t *ptl; |
4c887265 | 3677 | |
04f2cbe3 MG |
3678 | /* |
3679 | * Currently, we are forced to kill the process in the event the | |
3680 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 3681 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
3682 | */ |
3683 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 3684 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 3685 | current->pid); |
04f2cbe3 MG |
3686 | return ret; |
3687 | } | |
3688 | ||
4c887265 AL |
3689 | /* |
3690 | * Use page lock to guard against racing truncation | |
3691 | * before we get page_table_lock. | |
3692 | */ | |
6bda666a CL |
3693 | retry: |
3694 | page = find_lock_page(mapping, idx); | |
3695 | if (!page) { | |
a5516438 | 3696 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
3697 | if (idx >= size) |
3698 | goto out; | |
1a1aad8a MK |
3699 | |
3700 | /* | |
3701 | * Check for page in userfault range | |
3702 | */ | |
3703 | if (userfaultfd_missing(vma)) { | |
3704 | u32 hash; | |
3705 | struct vm_fault vmf = { | |
3706 | .vma = vma, | |
3707 | .address = address, | |
3708 | .flags = flags, | |
3709 | /* | |
3710 | * Hard to debug if it ends up being | |
3711 | * used by a callee that assumes | |
3712 | * something about the other | |
3713 | * uninitialized fields... same as in | |
3714 | * memory.c | |
3715 | */ | |
3716 | }; | |
3717 | ||
3718 | /* | |
3719 | * hugetlb_fault_mutex must be dropped before | |
3720 | * handling userfault. Reacquire after handling | |
3721 | * fault to make calling code simpler. | |
3722 | */ | |
3723 | hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, | |
3724 | idx, address); | |
3725 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
3726 | ret = handle_userfault(&vmf, VM_UFFD_MISSING); | |
3727 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
3728 | goto out; | |
3729 | } | |
3730 | ||
04f2cbe3 | 3731 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 3732 | if (IS_ERR(page)) { |
76dcee75 AK |
3733 | ret = PTR_ERR(page); |
3734 | if (ret == -ENOMEM) | |
3735 | ret = VM_FAULT_OOM; | |
3736 | else | |
3737 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
3738 | goto out; |
3739 | } | |
47ad8475 | 3740 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 3741 | __SetPageUptodate(page); |
bcc54222 | 3742 | set_page_huge_active(page); |
ac9b9c66 | 3743 | |
f83a275d | 3744 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 3745 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
3746 | if (err) { |
3747 | put_page(page); | |
6bda666a CL |
3748 | if (err == -EEXIST) |
3749 | goto retry; | |
3750 | goto out; | |
3751 | } | |
23be7468 | 3752 | } else { |
6bda666a | 3753 | lock_page(page); |
0fe6e20b NH |
3754 | if (unlikely(anon_vma_prepare(vma))) { |
3755 | ret = VM_FAULT_OOM; | |
3756 | goto backout_unlocked; | |
3757 | } | |
409eb8c2 | 3758 | anon_rmap = 1; |
23be7468 | 3759 | } |
0fe6e20b | 3760 | } else { |
998b4382 NH |
3761 | /* |
3762 | * If memory error occurs between mmap() and fault, some process | |
3763 | * don't have hwpoisoned swap entry for errored virtual address. | |
3764 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
3765 | */ | |
3766 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 3767 | ret = VM_FAULT_HWPOISON | |
972dc4de | 3768 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
3769 | goto backout_unlocked; |
3770 | } | |
6bda666a | 3771 | } |
1e8f889b | 3772 | |
57303d80 AW |
3773 | /* |
3774 | * If we are going to COW a private mapping later, we examine the | |
3775 | * pending reservations for this page now. This will ensure that | |
3776 | * any allocations necessary to record that reservation occur outside | |
3777 | * the spinlock. | |
3778 | */ | |
5e911373 | 3779 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
2b26736c AW |
3780 | if (vma_needs_reservation(h, vma, address) < 0) { |
3781 | ret = VM_FAULT_OOM; | |
3782 | goto backout_unlocked; | |
3783 | } | |
5e911373 | 3784 | /* Just decrements count, does not deallocate */ |
feba16e2 | 3785 | vma_end_reservation(h, vma, address); |
5e911373 | 3786 | } |
57303d80 | 3787 | |
8bea8052 | 3788 | ptl = huge_pte_lock(h, mm, ptep); |
a5516438 | 3789 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
3790 | if (idx >= size) |
3791 | goto backout; | |
3792 | ||
83c54070 | 3793 | ret = 0; |
7f2e9525 | 3794 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
3795 | goto backout; |
3796 | ||
07443a85 JK |
3797 | if (anon_rmap) { |
3798 | ClearPagePrivate(page); | |
409eb8c2 | 3799 | hugepage_add_new_anon_rmap(page, vma, address); |
ac714904 | 3800 | } else |
53f9263b | 3801 | page_dup_rmap(page, true); |
1e8f889b DG |
3802 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
3803 | && (vma->vm_flags & VM_SHARED))); | |
3804 | set_huge_pte_at(mm, address, ptep, new_pte); | |
3805 | ||
5d317b2b | 3806 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 3807 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 3808 | /* Optimization, do the COW without a second fault */ |
3999f52e | 3809 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); |
1e8f889b DG |
3810 | } |
3811 | ||
cb900f41 | 3812 | spin_unlock(ptl); |
4c887265 AL |
3813 | unlock_page(page); |
3814 | out: | |
ac9b9c66 | 3815 | return ret; |
4c887265 AL |
3816 | |
3817 | backout: | |
cb900f41 | 3818 | spin_unlock(ptl); |
2b26736c | 3819 | backout_unlocked: |
4c887265 | 3820 | unlock_page(page); |
96b96a96 | 3821 | restore_reserve_on_error(h, vma, address, page); |
4c887265 AL |
3822 | put_page(page); |
3823 | goto out; | |
ac9b9c66 HD |
3824 | } |
3825 | ||
8382d914 | 3826 | #ifdef CONFIG_SMP |
c672c7f2 | 3827 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
8382d914 DB |
3828 | struct vm_area_struct *vma, |
3829 | struct address_space *mapping, | |
3830 | pgoff_t idx, unsigned long address) | |
3831 | { | |
3832 | unsigned long key[2]; | |
3833 | u32 hash; | |
3834 | ||
3835 | if (vma->vm_flags & VM_SHARED) { | |
3836 | key[0] = (unsigned long) mapping; | |
3837 | key[1] = idx; | |
3838 | } else { | |
3839 | key[0] = (unsigned long) mm; | |
3840 | key[1] = address >> huge_page_shift(h); | |
3841 | } | |
3842 | ||
3843 | hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); | |
3844 | ||
3845 | return hash & (num_fault_mutexes - 1); | |
3846 | } | |
3847 | #else | |
3848 | /* | |
3849 | * For uniprocesor systems we always use a single mutex, so just | |
3850 | * return 0 and avoid the hashing overhead. | |
3851 | */ | |
c672c7f2 | 3852 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
8382d914 DB |
3853 | struct vm_area_struct *vma, |
3854 | struct address_space *mapping, | |
3855 | pgoff_t idx, unsigned long address) | |
3856 | { | |
3857 | return 0; | |
3858 | } | |
3859 | #endif | |
3860 | ||
86e5216f | 3861 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 3862 | unsigned long address, unsigned int flags) |
86e5216f | 3863 | { |
8382d914 | 3864 | pte_t *ptep, entry; |
cb900f41 | 3865 | spinlock_t *ptl; |
1e8f889b | 3866 | int ret; |
8382d914 DB |
3867 | u32 hash; |
3868 | pgoff_t idx; | |
0fe6e20b | 3869 | struct page *page = NULL; |
57303d80 | 3870 | struct page *pagecache_page = NULL; |
a5516438 | 3871 | struct hstate *h = hstate_vma(vma); |
8382d914 | 3872 | struct address_space *mapping; |
0f792cf9 | 3873 | int need_wait_lock = 0; |
86e5216f | 3874 | |
1e16a539 KH |
3875 | address &= huge_page_mask(h); |
3876 | ||
7868a208 | 3877 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
fd6a03ed NH |
3878 | if (ptep) { |
3879 | entry = huge_ptep_get(ptep); | |
290408d4 | 3880 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 3881 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
3882 | return 0; |
3883 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 3884 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 3885 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
0d777df5 NH |
3886 | } else { |
3887 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); | |
3888 | if (!ptep) | |
3889 | return VM_FAULT_OOM; | |
fd6a03ed NH |
3890 | } |
3891 | ||
8382d914 DB |
3892 | mapping = vma->vm_file->f_mapping; |
3893 | idx = vma_hugecache_offset(h, vma, address); | |
3894 | ||
3935baa9 DG |
3895 | /* |
3896 | * Serialize hugepage allocation and instantiation, so that we don't | |
3897 | * get spurious allocation failures if two CPUs race to instantiate | |
3898 | * the same page in the page cache. | |
3899 | */ | |
c672c7f2 MK |
3900 | hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address); |
3901 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
8382d914 | 3902 | |
7f2e9525 GS |
3903 | entry = huge_ptep_get(ptep); |
3904 | if (huge_pte_none(entry)) { | |
8382d914 | 3905 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 3906 | goto out_mutex; |
3935baa9 | 3907 | } |
86e5216f | 3908 | |
83c54070 | 3909 | ret = 0; |
1e8f889b | 3910 | |
0f792cf9 NH |
3911 | /* |
3912 | * entry could be a migration/hwpoison entry at this point, so this | |
3913 | * check prevents the kernel from going below assuming that we have | |
3914 | * a active hugepage in pagecache. This goto expects the 2nd page fault, | |
3915 | * and is_hugetlb_entry_(migration|hwpoisoned) check will properly | |
3916 | * handle it. | |
3917 | */ | |
3918 | if (!pte_present(entry)) | |
3919 | goto out_mutex; | |
3920 | ||
57303d80 AW |
3921 | /* |
3922 | * If we are going to COW the mapping later, we examine the pending | |
3923 | * reservations for this page now. This will ensure that any | |
3924 | * allocations necessary to record that reservation occur outside the | |
3925 | * spinlock. For private mappings, we also lookup the pagecache | |
3926 | * page now as it is used to determine if a reservation has been | |
3927 | * consumed. | |
3928 | */ | |
106c992a | 3929 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
2b26736c AW |
3930 | if (vma_needs_reservation(h, vma, address) < 0) { |
3931 | ret = VM_FAULT_OOM; | |
b4d1d99f | 3932 | goto out_mutex; |
2b26736c | 3933 | } |
5e911373 | 3934 | /* Just decrements count, does not deallocate */ |
feba16e2 | 3935 | vma_end_reservation(h, vma, address); |
57303d80 | 3936 | |
f83a275d | 3937 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
3938 | pagecache_page = hugetlbfs_pagecache_page(h, |
3939 | vma, address); | |
3940 | } | |
3941 | ||
0f792cf9 NH |
3942 | ptl = huge_pte_lock(h, mm, ptep); |
3943 | ||
3944 | /* Check for a racing update before calling hugetlb_cow */ | |
3945 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
3946 | goto out_ptl; | |
3947 | ||
56c9cfb1 NH |
3948 | /* |
3949 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
3950 | * pagecache_page, so here we need take the former one | |
3951 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
3952 | */ |
3953 | page = pte_page(entry); | |
3954 | if (page != pagecache_page) | |
0f792cf9 NH |
3955 | if (!trylock_page(page)) { |
3956 | need_wait_lock = 1; | |
3957 | goto out_ptl; | |
3958 | } | |
b4d1d99f | 3959 | |
0f792cf9 | 3960 | get_page(page); |
b4d1d99f | 3961 | |
788c7df4 | 3962 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 3963 | if (!huge_pte_write(entry)) { |
3999f52e AK |
3964 | ret = hugetlb_cow(mm, vma, address, ptep, |
3965 | pagecache_page, ptl); | |
0f792cf9 | 3966 | goto out_put_page; |
b4d1d99f | 3967 | } |
106c992a | 3968 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
3969 | } |
3970 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
3971 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
3972 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 3973 | update_mmu_cache(vma, address, ptep); |
0f792cf9 NH |
3974 | out_put_page: |
3975 | if (page != pagecache_page) | |
3976 | unlock_page(page); | |
3977 | put_page(page); | |
cb900f41 KS |
3978 | out_ptl: |
3979 | spin_unlock(ptl); | |
57303d80 AW |
3980 | |
3981 | if (pagecache_page) { | |
3982 | unlock_page(pagecache_page); | |
3983 | put_page(pagecache_page); | |
3984 | } | |
b4d1d99f | 3985 | out_mutex: |
c672c7f2 | 3986 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
0f792cf9 NH |
3987 | /* |
3988 | * Generally it's safe to hold refcount during waiting page lock. But | |
3989 | * here we just wait to defer the next page fault to avoid busy loop and | |
3990 | * the page is not used after unlocked before returning from the current | |
3991 | * page fault. So we are safe from accessing freed page, even if we wait | |
3992 | * here without taking refcount. | |
3993 | */ | |
3994 | if (need_wait_lock) | |
3995 | wait_on_page_locked(page); | |
1e8f889b | 3996 | return ret; |
86e5216f AL |
3997 | } |
3998 | ||
8fb5debc MK |
3999 | /* |
4000 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
4001 | * modifications for huge pages. | |
4002 | */ | |
4003 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
4004 | pte_t *dst_pte, | |
4005 | struct vm_area_struct *dst_vma, | |
4006 | unsigned long dst_addr, | |
4007 | unsigned long src_addr, | |
4008 | struct page **pagep) | |
4009 | { | |
1e392147 AA |
4010 | struct address_space *mapping; |
4011 | pgoff_t idx; | |
4012 | unsigned long size; | |
1c9e8def | 4013 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
4014 | struct hstate *h = hstate_vma(dst_vma); |
4015 | pte_t _dst_pte; | |
4016 | spinlock_t *ptl; | |
4017 | int ret; | |
4018 | struct page *page; | |
4019 | ||
4020 | if (!*pagep) { | |
4021 | ret = -ENOMEM; | |
4022 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
4023 | if (IS_ERR(page)) | |
4024 | goto out; | |
4025 | ||
4026 | ret = copy_huge_page_from_user(page, | |
4027 | (const void __user *) src_addr, | |
810a56b9 | 4028 | pages_per_huge_page(h), false); |
8fb5debc MK |
4029 | |
4030 | /* fallback to copy_from_user outside mmap_sem */ | |
4031 | if (unlikely(ret)) { | |
4032 | ret = -EFAULT; | |
4033 | *pagep = page; | |
4034 | /* don't free the page */ | |
4035 | goto out; | |
4036 | } | |
4037 | } else { | |
4038 | page = *pagep; | |
4039 | *pagep = NULL; | |
4040 | } | |
4041 | ||
4042 | /* | |
4043 | * The memory barrier inside __SetPageUptodate makes sure that | |
4044 | * preceding stores to the page contents become visible before | |
4045 | * the set_pte_at() write. | |
4046 | */ | |
4047 | __SetPageUptodate(page); | |
4048 | set_page_huge_active(page); | |
4049 | ||
1e392147 AA |
4050 | mapping = dst_vma->vm_file->f_mapping; |
4051 | idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
4052 | ||
1c9e8def MK |
4053 | /* |
4054 | * If shared, add to page cache | |
4055 | */ | |
4056 | if (vm_shared) { | |
1e392147 AA |
4057 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4058 | ret = -EFAULT; | |
4059 | if (idx >= size) | |
4060 | goto out_release_nounlock; | |
1c9e8def | 4061 | |
1e392147 AA |
4062 | /* |
4063 | * Serialization between remove_inode_hugepages() and | |
4064 | * huge_add_to_page_cache() below happens through the | |
4065 | * hugetlb_fault_mutex_table that here must be hold by | |
4066 | * the caller. | |
4067 | */ | |
1c9e8def MK |
4068 | ret = huge_add_to_page_cache(page, mapping, idx); |
4069 | if (ret) | |
4070 | goto out_release_nounlock; | |
4071 | } | |
4072 | ||
8fb5debc MK |
4073 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); |
4074 | spin_lock(ptl); | |
4075 | ||
1e392147 AA |
4076 | /* |
4077 | * Recheck the i_size after holding PT lock to make sure not | |
4078 | * to leave any page mapped (as page_mapped()) beyond the end | |
4079 | * of the i_size (remove_inode_hugepages() is strict about | |
4080 | * enforcing that). If we bail out here, we'll also leave a | |
4081 | * page in the radix tree in the vm_shared case beyond the end | |
4082 | * of the i_size, but remove_inode_hugepages() will take care | |
4083 | * of it as soon as we drop the hugetlb_fault_mutex_table. | |
4084 | */ | |
4085 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
4086 | ret = -EFAULT; | |
4087 | if (idx >= size) | |
4088 | goto out_release_unlock; | |
4089 | ||
8fb5debc MK |
4090 | ret = -EEXIST; |
4091 | if (!huge_pte_none(huge_ptep_get(dst_pte))) | |
4092 | goto out_release_unlock; | |
4093 | ||
1c9e8def MK |
4094 | if (vm_shared) { |
4095 | page_dup_rmap(page, true); | |
4096 | } else { | |
4097 | ClearPagePrivate(page); | |
4098 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); | |
4099 | } | |
8fb5debc MK |
4100 | |
4101 | _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); | |
4102 | if (dst_vma->vm_flags & VM_WRITE) | |
4103 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
4104 | _dst_pte = pte_mkyoung(_dst_pte); | |
4105 | ||
4106 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); | |
4107 | ||
4108 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
4109 | dst_vma->vm_flags & VM_WRITE); | |
4110 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
4111 | ||
4112 | /* No need to invalidate - it was non-present before */ | |
4113 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
4114 | ||
4115 | spin_unlock(ptl); | |
1c9e8def MK |
4116 | if (vm_shared) |
4117 | unlock_page(page); | |
8fb5debc MK |
4118 | ret = 0; |
4119 | out: | |
4120 | return ret; | |
4121 | out_release_unlock: | |
4122 | spin_unlock(ptl); | |
1c9e8def MK |
4123 | if (vm_shared) |
4124 | unlock_page(page); | |
5af10dfd | 4125 | out_release_nounlock: |
8fb5debc MK |
4126 | put_page(page); |
4127 | goto out; | |
4128 | } | |
4129 | ||
28a35716 ML |
4130 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
4131 | struct page **pages, struct vm_area_struct **vmas, | |
4132 | unsigned long *position, unsigned long *nr_pages, | |
87ffc118 | 4133 | long i, unsigned int flags, int *nonblocking) |
63551ae0 | 4134 | { |
d5d4b0aa KC |
4135 | unsigned long pfn_offset; |
4136 | unsigned long vaddr = *position; | |
28a35716 | 4137 | unsigned long remainder = *nr_pages; |
a5516438 | 4138 | struct hstate *h = hstate_vma(vma); |
2be7cfed | 4139 | int err = -EFAULT; |
63551ae0 | 4140 | |
63551ae0 | 4141 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 4142 | pte_t *pte; |
cb900f41 | 4143 | spinlock_t *ptl = NULL; |
2a15efc9 | 4144 | int absent; |
4c887265 | 4145 | struct page *page; |
63551ae0 | 4146 | |
02057967 DR |
4147 | /* |
4148 | * If we have a pending SIGKILL, don't keep faulting pages and | |
4149 | * potentially allocating memory. | |
4150 | */ | |
4151 | if (unlikely(fatal_signal_pending(current))) { | |
4152 | remainder = 0; | |
4153 | break; | |
4154 | } | |
4155 | ||
4c887265 AL |
4156 | /* |
4157 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 4158 | * each hugepage. We have to make sure we get the |
4c887265 | 4159 | * first, for the page indexing below to work. |
cb900f41 KS |
4160 | * |
4161 | * Note that page table lock is not held when pte is null. | |
4c887265 | 4162 | */ |
7868a208 PA |
4163 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
4164 | huge_page_size(h)); | |
cb900f41 KS |
4165 | if (pte) |
4166 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
4167 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
4168 | ||
4169 | /* | |
4170 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
4171 | * an error where there's an empty slot with no huge pagecache |
4172 | * to back it. This way, we avoid allocating a hugepage, and | |
4173 | * the sparse dumpfile avoids allocating disk blocks, but its | |
4174 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 4175 | */ |
3ae77f43 HD |
4176 | if (absent && (flags & FOLL_DUMP) && |
4177 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
4178 | if (pte) |
4179 | spin_unlock(ptl); | |
2a15efc9 HD |
4180 | remainder = 0; |
4181 | break; | |
4182 | } | |
63551ae0 | 4183 | |
9cc3a5bd NH |
4184 | /* |
4185 | * We need call hugetlb_fault for both hugepages under migration | |
4186 | * (in which case hugetlb_fault waits for the migration,) and | |
4187 | * hwpoisoned hugepages (in which case we need to prevent the | |
4188 | * caller from accessing to them.) In order to do this, we use | |
4189 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
4190 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
4191 | * both cases, and because we can't follow correct pages | |
4192 | * directly from any kind of swap entries. | |
4193 | */ | |
4194 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
4195 | ((flags & FOLL_WRITE) && |
4196 | !huge_pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 4197 | int ret; |
87ffc118 | 4198 | unsigned int fault_flags = 0; |
63551ae0 | 4199 | |
cb900f41 KS |
4200 | if (pte) |
4201 | spin_unlock(ptl); | |
87ffc118 AA |
4202 | if (flags & FOLL_WRITE) |
4203 | fault_flags |= FAULT_FLAG_WRITE; | |
4204 | if (nonblocking) | |
4205 | fault_flags |= FAULT_FLAG_ALLOW_RETRY; | |
4206 | if (flags & FOLL_NOWAIT) | |
4207 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
4208 | FAULT_FLAG_RETRY_NOWAIT; | |
4209 | if (flags & FOLL_TRIED) { | |
4210 | VM_WARN_ON_ONCE(fault_flags & | |
4211 | FAULT_FLAG_ALLOW_RETRY); | |
4212 | fault_flags |= FAULT_FLAG_TRIED; | |
4213 | } | |
4214 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
4215 | if (ret & VM_FAULT_ERROR) { | |
2be7cfed | 4216 | err = vm_fault_to_errno(ret, flags); |
87ffc118 AA |
4217 | remainder = 0; |
4218 | break; | |
4219 | } | |
4220 | if (ret & VM_FAULT_RETRY) { | |
4221 | if (nonblocking) | |
4222 | *nonblocking = 0; | |
4223 | *nr_pages = 0; | |
4224 | /* | |
4225 | * VM_FAULT_RETRY must not return an | |
4226 | * error, it will return zero | |
4227 | * instead. | |
4228 | * | |
4229 | * No need to update "position" as the | |
4230 | * caller will not check it after | |
4231 | * *nr_pages is set to 0. | |
4232 | */ | |
4233 | return i; | |
4234 | } | |
4235 | continue; | |
4c887265 AL |
4236 | } |
4237 | ||
a5516438 | 4238 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 4239 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 4240 | same_page: |
d6692183 | 4241 | if (pages) { |
2a15efc9 | 4242 | pages[i] = mem_map_offset(page, pfn_offset); |
ddc58f27 | 4243 | get_page(pages[i]); |
d6692183 | 4244 | } |
63551ae0 DG |
4245 | |
4246 | if (vmas) | |
4247 | vmas[i] = vma; | |
4248 | ||
4249 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 4250 | ++pfn_offset; |
63551ae0 DG |
4251 | --remainder; |
4252 | ++i; | |
d5d4b0aa | 4253 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 4254 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa KC |
4255 | /* |
4256 | * We use pfn_offset to avoid touching the pageframes | |
4257 | * of this compound page. | |
4258 | */ | |
4259 | goto same_page; | |
4260 | } | |
cb900f41 | 4261 | spin_unlock(ptl); |
63551ae0 | 4262 | } |
28a35716 | 4263 | *nr_pages = remainder; |
87ffc118 AA |
4264 | /* |
4265 | * setting position is actually required only if remainder is | |
4266 | * not zero but it's faster not to add a "if (remainder)" | |
4267 | * branch. | |
4268 | */ | |
63551ae0 DG |
4269 | *position = vaddr; |
4270 | ||
2be7cfed | 4271 | return i ? i : err; |
63551ae0 | 4272 | } |
8f860591 | 4273 | |
5491ae7b AK |
4274 | #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE |
4275 | /* | |
4276 | * ARCHes with special requirements for evicting HUGETLB backing TLB entries can | |
4277 | * implement this. | |
4278 | */ | |
4279 | #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) | |
4280 | #endif | |
4281 | ||
7da4d641 | 4282 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
4283 | unsigned long address, unsigned long end, pgprot_t newprot) |
4284 | { | |
4285 | struct mm_struct *mm = vma->vm_mm; | |
4286 | unsigned long start = address; | |
4287 | pte_t *ptep; | |
4288 | pte_t pte; | |
a5516438 | 4289 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 4290 | unsigned long pages = 0; |
8f860591 ZY |
4291 | |
4292 | BUG_ON(address >= end); | |
4293 | flush_cache_range(vma, address, end); | |
4294 | ||
a5338093 | 4295 | mmu_notifier_invalidate_range_start(mm, start, end); |
83cde9e8 | 4296 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 4297 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 4298 | spinlock_t *ptl; |
7868a208 | 4299 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
8f860591 ZY |
4300 | if (!ptep) |
4301 | continue; | |
cb900f41 | 4302 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
4303 | if (huge_pmd_unshare(mm, &address, ptep)) { |
4304 | pages++; | |
cb900f41 | 4305 | spin_unlock(ptl); |
39dde65c | 4306 | continue; |
7da4d641 | 4307 | } |
a8bda28d NH |
4308 | pte = huge_ptep_get(ptep); |
4309 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
4310 | spin_unlock(ptl); | |
4311 | continue; | |
4312 | } | |
4313 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
4314 | swp_entry_t entry = pte_to_swp_entry(pte); | |
4315 | ||
4316 | if (is_write_migration_entry(entry)) { | |
4317 | pte_t newpte; | |
4318 | ||
4319 | make_migration_entry_read(&entry); | |
4320 | newpte = swp_entry_to_pte(entry); | |
e5251fd4 PA |
4321 | set_huge_swap_pte_at(mm, address, ptep, |
4322 | newpte, huge_page_size(h)); | |
a8bda28d NH |
4323 | pages++; |
4324 | } | |
4325 | spin_unlock(ptl); | |
4326 | continue; | |
4327 | } | |
4328 | if (!huge_pte_none(pte)) { | |
8f860591 | 4329 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
106c992a | 4330 | pte = pte_mkhuge(huge_pte_modify(pte, newprot)); |
be7517d6 | 4331 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
8f860591 | 4332 | set_huge_pte_at(mm, address, ptep, pte); |
7da4d641 | 4333 | pages++; |
8f860591 | 4334 | } |
cb900f41 | 4335 | spin_unlock(ptl); |
8f860591 | 4336 | } |
d833352a | 4337 | /* |
c8c06efa | 4338 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 4339 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 4340 | * once we release i_mmap_rwsem, another task can do the final put_page |
d833352a MG |
4341 | * and that page table be reused and filled with junk. |
4342 | */ | |
5491ae7b | 4343 | flush_hugetlb_tlb_range(vma, start, end); |
0f10851e JG |
4344 | /* |
4345 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
4346 | * page table protection not changing it to point to a new page. | |
4347 | * | |
4348 | * See Documentation/vm/mmu_notifier.txt | |
4349 | */ | |
83cde9e8 | 4350 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
a5338093 | 4351 | mmu_notifier_invalidate_range_end(mm, start, end); |
7da4d641 PZ |
4352 | |
4353 | return pages << h->order; | |
8f860591 ZY |
4354 | } |
4355 | ||
a1e78772 MG |
4356 | int hugetlb_reserve_pages(struct inode *inode, |
4357 | long from, long to, | |
5a6fe125 | 4358 | struct vm_area_struct *vma, |
ca16d140 | 4359 | vm_flags_t vm_flags) |
e4e574b7 | 4360 | { |
17c9d12e | 4361 | long ret, chg; |
a5516438 | 4362 | struct hstate *h = hstate_inode(inode); |
90481622 | 4363 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 4364 | struct resv_map *resv_map; |
1c5ecae3 | 4365 | long gbl_reserve; |
e4e574b7 | 4366 | |
17c9d12e MG |
4367 | /* |
4368 | * Only apply hugepage reservation if asked. At fault time, an | |
4369 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 4370 | * without using reserves |
17c9d12e | 4371 | */ |
ca16d140 | 4372 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
4373 | return 0; |
4374 | ||
a1e78772 MG |
4375 | /* |
4376 | * Shared mappings base their reservation on the number of pages that | |
4377 | * are already allocated on behalf of the file. Private mappings need | |
4378 | * to reserve the full area even if read-only as mprotect() may be | |
4379 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
4380 | */ | |
9119a41e | 4381 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4e35f483 | 4382 | resv_map = inode_resv_map(inode); |
9119a41e | 4383 | |
1406ec9b | 4384 | chg = region_chg(resv_map, from, to); |
9119a41e JK |
4385 | |
4386 | } else { | |
4387 | resv_map = resv_map_alloc(); | |
17c9d12e MG |
4388 | if (!resv_map) |
4389 | return -ENOMEM; | |
4390 | ||
a1e78772 | 4391 | chg = to - from; |
84afd99b | 4392 | |
17c9d12e MG |
4393 | set_vma_resv_map(vma, resv_map); |
4394 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
4395 | } | |
4396 | ||
c50ac050 DH |
4397 | if (chg < 0) { |
4398 | ret = chg; | |
4399 | goto out_err; | |
4400 | } | |
8a630112 | 4401 | |
1c5ecae3 MK |
4402 | /* |
4403 | * There must be enough pages in the subpool for the mapping. If | |
4404 | * the subpool has a minimum size, there may be some global | |
4405 | * reservations already in place (gbl_reserve). | |
4406 | */ | |
4407 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
4408 | if (gbl_reserve < 0) { | |
c50ac050 DH |
4409 | ret = -ENOSPC; |
4410 | goto out_err; | |
4411 | } | |
5a6fe125 MG |
4412 | |
4413 | /* | |
17c9d12e | 4414 | * Check enough hugepages are available for the reservation. |
90481622 | 4415 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 4416 | */ |
1c5ecae3 | 4417 | ret = hugetlb_acct_memory(h, gbl_reserve); |
68842c9b | 4418 | if (ret < 0) { |
1c5ecae3 MK |
4419 | /* put back original number of pages, chg */ |
4420 | (void)hugepage_subpool_put_pages(spool, chg); | |
c50ac050 | 4421 | goto out_err; |
68842c9b | 4422 | } |
17c9d12e MG |
4423 | |
4424 | /* | |
4425 | * Account for the reservations made. Shared mappings record regions | |
4426 | * that have reservations as they are shared by multiple VMAs. | |
4427 | * When the last VMA disappears, the region map says how much | |
4428 | * the reservation was and the page cache tells how much of | |
4429 | * the reservation was consumed. Private mappings are per-VMA and | |
4430 | * only the consumed reservations are tracked. When the VMA | |
4431 | * disappears, the original reservation is the VMA size and the | |
4432 | * consumed reservations are stored in the map. Hence, nothing | |
4433 | * else has to be done for private mappings here | |
4434 | */ | |
33039678 MK |
4435 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4436 | long add = region_add(resv_map, from, to); | |
4437 | ||
4438 | if (unlikely(chg > add)) { | |
4439 | /* | |
4440 | * pages in this range were added to the reserve | |
4441 | * map between region_chg and region_add. This | |
4442 | * indicates a race with alloc_huge_page. Adjust | |
4443 | * the subpool and reserve counts modified above | |
4444 | * based on the difference. | |
4445 | */ | |
4446 | long rsv_adjust; | |
4447 | ||
4448 | rsv_adjust = hugepage_subpool_put_pages(spool, | |
4449 | chg - add); | |
4450 | hugetlb_acct_memory(h, -rsv_adjust); | |
4451 | } | |
4452 | } | |
a43a8c39 | 4453 | return 0; |
c50ac050 | 4454 | out_err: |
5e911373 | 4455 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
ff8c0c53 MK |
4456 | /* Don't call region_abort if region_chg failed */ |
4457 | if (chg >= 0) | |
4458 | region_abort(resv_map, from, to); | |
f031dd27 JK |
4459 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
4460 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 4461 | return ret; |
a43a8c39 KC |
4462 | } |
4463 | ||
b5cec28d MK |
4464 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
4465 | long freed) | |
a43a8c39 | 4466 | { |
a5516438 | 4467 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 4468 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 4469 | long chg = 0; |
90481622 | 4470 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 4471 | long gbl_reserve; |
45c682a6 | 4472 | |
b5cec28d MK |
4473 | if (resv_map) { |
4474 | chg = region_del(resv_map, start, end); | |
4475 | /* | |
4476 | * region_del() can fail in the rare case where a region | |
4477 | * must be split and another region descriptor can not be | |
4478 | * allocated. If end == LONG_MAX, it will not fail. | |
4479 | */ | |
4480 | if (chg < 0) | |
4481 | return chg; | |
4482 | } | |
4483 | ||
45c682a6 | 4484 | spin_lock(&inode->i_lock); |
e4c6f8be | 4485 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
4486 | spin_unlock(&inode->i_lock); |
4487 | ||
1c5ecae3 MK |
4488 | /* |
4489 | * If the subpool has a minimum size, the number of global | |
4490 | * reservations to be released may be adjusted. | |
4491 | */ | |
4492 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
4493 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
4494 | |
4495 | return 0; | |
a43a8c39 | 4496 | } |
93f70f90 | 4497 | |
3212b535 SC |
4498 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
4499 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
4500 | struct vm_area_struct *vma, | |
4501 | unsigned long addr, pgoff_t idx) | |
4502 | { | |
4503 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
4504 | svma->vm_start; | |
4505 | unsigned long sbase = saddr & PUD_MASK; | |
4506 | unsigned long s_end = sbase + PUD_SIZE; | |
4507 | ||
4508 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
4509 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
4510 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
4511 | |
4512 | /* | |
4513 | * match the virtual addresses, permission and the alignment of the | |
4514 | * page table page. | |
4515 | */ | |
4516 | if (pmd_index(addr) != pmd_index(saddr) || | |
4517 | vm_flags != svm_flags || | |
4518 | sbase < svma->vm_start || svma->vm_end < s_end) | |
4519 | return 0; | |
4520 | ||
4521 | return saddr; | |
4522 | } | |
4523 | ||
31aafb45 | 4524 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
4525 | { |
4526 | unsigned long base = addr & PUD_MASK; | |
4527 | unsigned long end = base + PUD_SIZE; | |
4528 | ||
4529 | /* | |
4530 | * check on proper vm_flags and page table alignment | |
4531 | */ | |
4532 | if (vma->vm_flags & VM_MAYSHARE && | |
4533 | vma->vm_start <= base && end <= vma->vm_end) | |
31aafb45 NK |
4534 | return true; |
4535 | return false; | |
3212b535 SC |
4536 | } |
4537 | ||
4538 | /* | |
4539 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
4540 | * and returns the corresponding pte. While this is not necessary for the | |
4541 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
4542 | * code much cleaner. pmd allocation is essential for the shared case because | |
c8c06efa | 4543 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise |
3212b535 SC |
4544 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a |
4545 | * bad pmd for sharing. | |
4546 | */ | |
4547 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4548 | { | |
4549 | struct vm_area_struct *vma = find_vma(mm, addr); | |
4550 | struct address_space *mapping = vma->vm_file->f_mapping; | |
4551 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
4552 | vma->vm_pgoff; | |
4553 | struct vm_area_struct *svma; | |
4554 | unsigned long saddr; | |
4555 | pte_t *spte = NULL; | |
4556 | pte_t *pte; | |
cb900f41 | 4557 | spinlock_t *ptl; |
3212b535 SC |
4558 | |
4559 | if (!vma_shareable(vma, addr)) | |
4560 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
4561 | ||
83cde9e8 | 4562 | i_mmap_lock_write(mapping); |
3212b535 SC |
4563 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
4564 | if (svma == vma) | |
4565 | continue; | |
4566 | ||
4567 | saddr = page_table_shareable(svma, vma, addr, idx); | |
4568 | if (saddr) { | |
7868a208 PA |
4569 | spte = huge_pte_offset(svma->vm_mm, saddr, |
4570 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
4571 | if (spte) { |
4572 | get_page(virt_to_page(spte)); | |
4573 | break; | |
4574 | } | |
4575 | } | |
4576 | } | |
4577 | ||
4578 | if (!spte) | |
4579 | goto out; | |
4580 | ||
8bea8052 | 4581 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 4582 | if (pud_none(*pud)) { |
3212b535 SC |
4583 | pud_populate(mm, pud, |
4584 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 4585 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 4586 | } else { |
3212b535 | 4587 | put_page(virt_to_page(spte)); |
dc6c9a35 | 4588 | } |
cb900f41 | 4589 | spin_unlock(ptl); |
3212b535 SC |
4590 | out: |
4591 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
83cde9e8 | 4592 | i_mmap_unlock_write(mapping); |
3212b535 SC |
4593 | return pte; |
4594 | } | |
4595 | ||
4596 | /* | |
4597 | * unmap huge page backed by shared pte. | |
4598 | * | |
4599 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
4600 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
4601 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
4602 | * | |
cb900f41 | 4603 | * called with page table lock held. |
3212b535 SC |
4604 | * |
4605 | * returns: 1 successfully unmapped a shared pte page | |
4606 | * 0 the underlying pte page is not shared, or it is the last user | |
4607 | */ | |
4608 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4609 | { | |
4610 | pgd_t *pgd = pgd_offset(mm, *addr); | |
c2febafc KS |
4611 | p4d_t *p4d = p4d_offset(pgd, *addr); |
4612 | pud_t *pud = pud_offset(p4d, *addr); | |
3212b535 SC |
4613 | |
4614 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
4615 | if (page_count(virt_to_page(ptep)) == 1) | |
4616 | return 0; | |
4617 | ||
4618 | pud_clear(pud); | |
4619 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 4620 | mm_dec_nr_pmds(mm); |
3212b535 SC |
4621 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
4622 | return 1; | |
4623 | } | |
9e5fc74c SC |
4624 | #define want_pmd_share() (1) |
4625 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
4626 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4627 | { | |
4628 | return NULL; | |
4629 | } | |
e81f2d22 ZZ |
4630 | |
4631 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4632 | { | |
4633 | return 0; | |
4634 | } | |
9e5fc74c | 4635 | #define want_pmd_share() (0) |
3212b535 SC |
4636 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
4637 | ||
9e5fc74c SC |
4638 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
4639 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
4640 | unsigned long addr, unsigned long sz) | |
4641 | { | |
4642 | pgd_t *pgd; | |
c2febafc | 4643 | p4d_t *p4d; |
9e5fc74c SC |
4644 | pud_t *pud; |
4645 | pte_t *pte = NULL; | |
4646 | ||
4647 | pgd = pgd_offset(mm, addr); | |
f4f0a3d8 KS |
4648 | p4d = p4d_alloc(mm, pgd, addr); |
4649 | if (!p4d) | |
4650 | return NULL; | |
c2febafc | 4651 | pud = pud_alloc(mm, p4d, addr); |
9e5fc74c SC |
4652 | if (pud) { |
4653 | if (sz == PUD_SIZE) { | |
4654 | pte = (pte_t *)pud; | |
4655 | } else { | |
4656 | BUG_ON(sz != PMD_SIZE); | |
4657 | if (want_pmd_share() && pud_none(*pud)) | |
4658 | pte = huge_pmd_share(mm, addr, pud); | |
4659 | else | |
4660 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
4661 | } | |
4662 | } | |
4e666314 | 4663 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
4664 | |
4665 | return pte; | |
4666 | } | |
4667 | ||
9b19df29 PA |
4668 | /* |
4669 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
4670 | * entry at address @addr | |
4671 | * | |
4672 | * Return: Pointer to page table or swap entry (PUD or PMD) for | |
4673 | * address @addr, or NULL if a p*d_none() entry is encountered and the | |
4674 | * size @sz doesn't match the hugepage size at this level of the page | |
4675 | * table. | |
4676 | */ | |
7868a208 PA |
4677 | pte_t *huge_pte_offset(struct mm_struct *mm, |
4678 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
4679 | { |
4680 | pgd_t *pgd; | |
c2febafc | 4681 | p4d_t *p4d; |
9e5fc74c | 4682 | pud_t *pud; |
c2febafc | 4683 | pmd_t *pmd; |
9e5fc74c SC |
4684 | |
4685 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
4686 | if (!pgd_present(*pgd)) |
4687 | return NULL; | |
4688 | p4d = p4d_offset(pgd, addr); | |
4689 | if (!p4d_present(*p4d)) | |
4690 | return NULL; | |
9b19df29 | 4691 | |
c2febafc | 4692 | pud = pud_offset(p4d, addr); |
9b19df29 | 4693 | if (sz != PUD_SIZE && pud_none(*pud)) |
c2febafc | 4694 | return NULL; |
9b19df29 PA |
4695 | /* hugepage or swap? */ |
4696 | if (pud_huge(*pud) || !pud_present(*pud)) | |
c2febafc | 4697 | return (pte_t *)pud; |
9b19df29 | 4698 | |
c2febafc | 4699 | pmd = pmd_offset(pud, addr); |
9b19df29 PA |
4700 | if (sz != PMD_SIZE && pmd_none(*pmd)) |
4701 | return NULL; | |
4702 | /* hugepage or swap? */ | |
4703 | if (pmd_huge(*pmd) || !pmd_present(*pmd)) | |
4704 | return (pte_t *)pmd; | |
4705 | ||
4706 | return NULL; | |
9e5fc74c SC |
4707 | } |
4708 | ||
61f77eda NH |
4709 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
4710 | ||
4711 | /* | |
4712 | * These functions are overwritable if your architecture needs its own | |
4713 | * behavior. | |
4714 | */ | |
4715 | struct page * __weak | |
4716 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
4717 | int write) | |
4718 | { | |
4719 | return ERR_PTR(-EINVAL); | |
4720 | } | |
4721 | ||
4dc71451 AK |
4722 | struct page * __weak |
4723 | follow_huge_pd(struct vm_area_struct *vma, | |
4724 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
4725 | { | |
4726 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
4727 | return NULL; | |
4728 | } | |
4729 | ||
61f77eda | 4730 | struct page * __weak |
9e5fc74c | 4731 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 4732 | pmd_t *pmd, int flags) |
9e5fc74c | 4733 | { |
e66f17ff NH |
4734 | struct page *page = NULL; |
4735 | spinlock_t *ptl; | |
c9d398fa | 4736 | pte_t pte; |
e66f17ff NH |
4737 | retry: |
4738 | ptl = pmd_lockptr(mm, pmd); | |
4739 | spin_lock(ptl); | |
4740 | /* | |
4741 | * make sure that the address range covered by this pmd is not | |
4742 | * unmapped from other threads. | |
4743 | */ | |
4744 | if (!pmd_huge(*pmd)) | |
4745 | goto out; | |
c9d398fa NH |
4746 | pte = huge_ptep_get((pte_t *)pmd); |
4747 | if (pte_present(pte)) { | |
97534127 | 4748 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
e66f17ff NH |
4749 | if (flags & FOLL_GET) |
4750 | get_page(page); | |
4751 | } else { | |
c9d398fa | 4752 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff NH |
4753 | spin_unlock(ptl); |
4754 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
4755 | goto retry; | |
4756 | } | |
4757 | /* | |
4758 | * hwpoisoned entry is treated as no_page_table in | |
4759 | * follow_page_mask(). | |
4760 | */ | |
4761 | } | |
4762 | out: | |
4763 | spin_unlock(ptl); | |
9e5fc74c SC |
4764 | return page; |
4765 | } | |
4766 | ||
61f77eda | 4767 | struct page * __weak |
9e5fc74c | 4768 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 4769 | pud_t *pud, int flags) |
9e5fc74c | 4770 | { |
e66f17ff NH |
4771 | if (flags & FOLL_GET) |
4772 | return NULL; | |
9e5fc74c | 4773 | |
e66f17ff | 4774 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
4775 | } |
4776 | ||
faaa5b62 AK |
4777 | struct page * __weak |
4778 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
4779 | { | |
4780 | if (flags & FOLL_GET) | |
4781 | return NULL; | |
4782 | ||
4783 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
4784 | } | |
4785 | ||
31caf665 NH |
4786 | bool isolate_huge_page(struct page *page, struct list_head *list) |
4787 | { | |
bcc54222 NH |
4788 | bool ret = true; |
4789 | ||
309381fe | 4790 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4791 | spin_lock(&hugetlb_lock); |
bcc54222 NH |
4792 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { |
4793 | ret = false; | |
4794 | goto unlock; | |
4795 | } | |
4796 | clear_page_huge_active(page); | |
31caf665 | 4797 | list_move_tail(&page->lru, list); |
bcc54222 | 4798 | unlock: |
31caf665 | 4799 | spin_unlock(&hugetlb_lock); |
bcc54222 | 4800 | return ret; |
31caf665 NH |
4801 | } |
4802 | ||
4803 | void putback_active_hugepage(struct page *page) | |
4804 | { | |
309381fe | 4805 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4806 | spin_lock(&hugetlb_lock); |
bcc54222 | 4807 | set_page_huge_active(page); |
31caf665 NH |
4808 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
4809 | spin_unlock(&hugetlb_lock); | |
4810 | put_page(page); | |
4811 | } | |
ab5ac90a MH |
4812 | |
4813 | void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) | |
4814 | { | |
4815 | struct hstate *h = page_hstate(oldpage); | |
4816 | ||
4817 | hugetlb_cgroup_migrate(oldpage, newpage); | |
4818 | set_page_owner_migrate_reason(newpage, reason); | |
4819 | ||
4820 | /* | |
4821 | * transfer temporary state of the new huge page. This is | |
4822 | * reverse to other transitions because the newpage is going to | |
4823 | * be final while the old one will be freed so it takes over | |
4824 | * the temporary status. | |
4825 | * | |
4826 | * Also note that we have to transfer the per-node surplus state | |
4827 | * here as well otherwise the global surplus count will not match | |
4828 | * the per-node's. | |
4829 | */ | |
4830 | if (PageHugeTemporary(newpage)) { | |
4831 | int old_nid = page_to_nid(oldpage); | |
4832 | int new_nid = page_to_nid(newpage); | |
4833 | ||
4834 | SetPageHugeTemporary(oldpage); | |
4835 | ClearPageHugeTemporary(newpage); | |
4836 | ||
4837 | spin_lock(&hugetlb_lock); | |
4838 | if (h->surplus_huge_pages_node[old_nid]) { | |
4839 | h->surplus_huge_pages_node[old_nid]--; | |
4840 | h->surplus_huge_pages_node[new_nid]++; | |
4841 | } | |
4842 | spin_unlock(&hugetlb_lock); | |
4843 | } | |
4844 | } |