]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * Generic hugetlb support. | |
6d49e352 | 3 | * (C) Nadia Yvette Chambers, April 2004 |
1da177e4 | 4 | */ |
1da177e4 LT |
5 | #include <linux/list.h> |
6 | #include <linux/init.h> | |
7 | #include <linux/module.h> | |
8 | #include <linux/mm.h> | |
e1759c21 | 9 | #include <linux/seq_file.h> |
1da177e4 LT |
10 | #include <linux/sysctl.h> |
11 | #include <linux/highmem.h> | |
cddb8a5c | 12 | #include <linux/mmu_notifier.h> |
1da177e4 | 13 | #include <linux/nodemask.h> |
63551ae0 | 14 | #include <linux/pagemap.h> |
5da7ca86 | 15 | #include <linux/mempolicy.h> |
3b32123d | 16 | #include <linux/compiler.h> |
aea47ff3 | 17 | #include <linux/cpuset.h> |
3935baa9 | 18 | #include <linux/mutex.h> |
aa888a74 | 19 | #include <linux/bootmem.h> |
a3437870 | 20 | #include <linux/sysfs.h> |
5a0e3ad6 | 21 | #include <linux/slab.h> |
0fe6e20b | 22 | #include <linux/rmap.h> |
fd6a03ed NH |
23 | #include <linux/swap.h> |
24 | #include <linux/swapops.h> | |
c8721bbb | 25 | #include <linux/page-isolation.h> |
8382d914 | 26 | #include <linux/jhash.h> |
d6606683 | 27 | |
63551ae0 DG |
28 | #include <asm/page.h> |
29 | #include <asm/pgtable.h> | |
24669e58 | 30 | #include <asm/tlb.h> |
63551ae0 | 31 | |
24669e58 | 32 | #include <linux/io.h> |
63551ae0 | 33 | #include <linux/hugetlb.h> |
9dd540e2 | 34 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 35 | #include <linux/node.h> |
7835e98b | 36 | #include "internal.h" |
1da177e4 | 37 | |
753162cd | 38 | int hugepages_treat_as_movable; |
a5516438 | 39 | |
c3f38a38 | 40 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
41 | unsigned int default_hstate_idx; |
42 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
641844f5 NH |
43 | /* |
44 | * Minimum page order among possible hugepage sizes, set to a proper value | |
45 | * at boot time. | |
46 | */ | |
47 | static unsigned int minimum_order __read_mostly = UINT_MAX; | |
e5ff2159 | 48 | |
53ba51d2 JT |
49 | __initdata LIST_HEAD(huge_boot_pages); |
50 | ||
e5ff2159 AK |
51 | /* for command line parsing */ |
52 | static struct hstate * __initdata parsed_hstate; | |
53 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 54 | static unsigned long __initdata default_hstate_size; |
e5ff2159 | 55 | |
3935baa9 | 56 | /* |
31caf665 NH |
57 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
58 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 59 | */ |
c3f38a38 | 60 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 61 | |
8382d914 DB |
62 | /* |
63 | * Serializes faults on the same logical page. This is used to | |
64 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
65 | */ | |
66 | static int num_fault_mutexes; | |
c672c7f2 | 67 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
8382d914 | 68 | |
7ca02d0a MK |
69 | /* Forward declaration */ |
70 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
71 | ||
90481622 DG |
72 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
73 | { | |
74 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
75 | ||
76 | spin_unlock(&spool->lock); | |
77 | ||
78 | /* If no pages are used, and no other handles to the subpool | |
7ca02d0a MK |
79 | * remain, give up any reservations mased on minimum size and |
80 | * free the subpool */ | |
81 | if (free) { | |
82 | if (spool->min_hpages != -1) | |
83 | hugetlb_acct_memory(spool->hstate, | |
84 | -spool->min_hpages); | |
90481622 | 85 | kfree(spool); |
7ca02d0a | 86 | } |
90481622 DG |
87 | } |
88 | ||
7ca02d0a MK |
89 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
90 | long min_hpages) | |
90481622 DG |
91 | { |
92 | struct hugepage_subpool *spool; | |
93 | ||
c6a91820 | 94 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
95 | if (!spool) |
96 | return NULL; | |
97 | ||
98 | spin_lock_init(&spool->lock); | |
99 | spool->count = 1; | |
7ca02d0a MK |
100 | spool->max_hpages = max_hpages; |
101 | spool->hstate = h; | |
102 | spool->min_hpages = min_hpages; | |
103 | ||
104 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
105 | kfree(spool); | |
106 | return NULL; | |
107 | } | |
108 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
109 | |
110 | return spool; | |
111 | } | |
112 | ||
113 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
114 | { | |
115 | spin_lock(&spool->lock); | |
116 | BUG_ON(!spool->count); | |
117 | spool->count--; | |
118 | unlock_or_release_subpool(spool); | |
119 | } | |
120 | ||
1c5ecae3 MK |
121 | /* |
122 | * Subpool accounting for allocating and reserving pages. | |
123 | * Return -ENOMEM if there are not enough resources to satisfy the | |
124 | * the request. Otherwise, return the number of pages by which the | |
125 | * global pools must be adjusted (upward). The returned value may | |
126 | * only be different than the passed value (delta) in the case where | |
127 | * a subpool minimum size must be manitained. | |
128 | */ | |
129 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
130 | long delta) |
131 | { | |
1c5ecae3 | 132 | long ret = delta; |
90481622 DG |
133 | |
134 | if (!spool) | |
1c5ecae3 | 135 | return ret; |
90481622 DG |
136 | |
137 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
138 | |
139 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
140 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
141 | spool->used_hpages += delta; | |
142 | else { | |
143 | ret = -ENOMEM; | |
144 | goto unlock_ret; | |
145 | } | |
90481622 | 146 | } |
90481622 | 147 | |
1c5ecae3 MK |
148 | if (spool->min_hpages != -1) { /* minimum size accounting */ |
149 | if (delta > spool->rsv_hpages) { | |
150 | /* | |
151 | * Asking for more reserves than those already taken on | |
152 | * behalf of subpool. Return difference. | |
153 | */ | |
154 | ret = delta - spool->rsv_hpages; | |
155 | spool->rsv_hpages = 0; | |
156 | } else { | |
157 | ret = 0; /* reserves already accounted for */ | |
158 | spool->rsv_hpages -= delta; | |
159 | } | |
160 | } | |
161 | ||
162 | unlock_ret: | |
163 | spin_unlock(&spool->lock); | |
90481622 DG |
164 | return ret; |
165 | } | |
166 | ||
1c5ecae3 MK |
167 | /* |
168 | * Subpool accounting for freeing and unreserving pages. | |
169 | * Return the number of global page reservations that must be dropped. | |
170 | * The return value may only be different than the passed value (delta) | |
171 | * in the case where a subpool minimum size must be maintained. | |
172 | */ | |
173 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
174 | long delta) |
175 | { | |
1c5ecae3 MK |
176 | long ret = delta; |
177 | ||
90481622 | 178 | if (!spool) |
1c5ecae3 | 179 | return delta; |
90481622 DG |
180 | |
181 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
182 | |
183 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
184 | spool->used_hpages -= delta; | |
185 | ||
186 | if (spool->min_hpages != -1) { /* minimum size accounting */ | |
187 | if (spool->rsv_hpages + delta <= spool->min_hpages) | |
188 | ret = 0; | |
189 | else | |
190 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
191 | ||
192 | spool->rsv_hpages += delta; | |
193 | if (spool->rsv_hpages > spool->min_hpages) | |
194 | spool->rsv_hpages = spool->min_hpages; | |
195 | } | |
196 | ||
197 | /* | |
198 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
199 | * quota reference, free it now. | |
200 | */ | |
90481622 | 201 | unlock_or_release_subpool(spool); |
1c5ecae3 MK |
202 | |
203 | return ret; | |
90481622 DG |
204 | } |
205 | ||
206 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
207 | { | |
208 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
209 | } | |
210 | ||
211 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
212 | { | |
496ad9aa | 213 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
214 | } |
215 | ||
96822904 AW |
216 | /* |
217 | * Region tracking -- allows tracking of reservations and instantiated pages | |
218 | * across the pages in a mapping. | |
84afd99b | 219 | * |
1dd308a7 MK |
220 | * The region data structures are embedded into a resv_map and protected |
221 | * by a resv_map's lock. The set of regions within the resv_map represent | |
222 | * reservations for huge pages, or huge pages that have already been | |
223 | * instantiated within the map. The from and to elements are huge page | |
224 | * indicies into the associated mapping. from indicates the starting index | |
225 | * of the region. to represents the first index past the end of the region. | |
226 | * | |
227 | * For example, a file region structure with from == 0 and to == 4 represents | |
228 | * four huge pages in a mapping. It is important to note that the to element | |
229 | * represents the first element past the end of the region. This is used in | |
230 | * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. | |
231 | * | |
232 | * Interval notation of the form [from, to) will be used to indicate that | |
233 | * the endpoint from is inclusive and to is exclusive. | |
96822904 AW |
234 | */ |
235 | struct file_region { | |
236 | struct list_head link; | |
237 | long from; | |
238 | long to; | |
239 | }; | |
240 | ||
1dd308a7 MK |
241 | /* |
242 | * Add the huge page range represented by [f, t) to the reserve | |
5e911373 MK |
243 | * map. In the normal case, existing regions will be expanded |
244 | * to accommodate the specified range. Sufficient regions should | |
245 | * exist for expansion due to the previous call to region_chg | |
246 | * with the same range. However, it is possible that region_del | |
247 | * could have been called after region_chg and modifed the map | |
248 | * in such a way that no region exists to be expanded. In this | |
249 | * case, pull a region descriptor from the cache associated with | |
250 | * the map and use that for the new range. | |
cf3ad20b MK |
251 | * |
252 | * Return the number of new huge pages added to the map. This | |
253 | * number is greater than or equal to zero. | |
1dd308a7 | 254 | */ |
1406ec9b | 255 | static long region_add(struct resv_map *resv, long f, long t) |
96822904 | 256 | { |
1406ec9b | 257 | struct list_head *head = &resv->regions; |
96822904 | 258 | struct file_region *rg, *nrg, *trg; |
cf3ad20b | 259 | long add = 0; |
96822904 | 260 | |
7b24d861 | 261 | spin_lock(&resv->lock); |
96822904 AW |
262 | /* Locate the region we are either in or before. */ |
263 | list_for_each_entry(rg, head, link) | |
264 | if (f <= rg->to) | |
265 | break; | |
266 | ||
5e911373 MK |
267 | /* |
268 | * If no region exists which can be expanded to include the | |
269 | * specified range, the list must have been modified by an | |
270 | * interleving call to region_del(). Pull a region descriptor | |
271 | * from the cache and use it for this range. | |
272 | */ | |
273 | if (&rg->link == head || t < rg->from) { | |
274 | VM_BUG_ON(resv->region_cache_count <= 0); | |
275 | ||
276 | resv->region_cache_count--; | |
277 | nrg = list_first_entry(&resv->region_cache, struct file_region, | |
278 | link); | |
279 | list_del(&nrg->link); | |
280 | ||
281 | nrg->from = f; | |
282 | nrg->to = t; | |
283 | list_add(&nrg->link, rg->link.prev); | |
284 | ||
285 | add += t - f; | |
286 | goto out_locked; | |
287 | } | |
288 | ||
96822904 AW |
289 | /* Round our left edge to the current segment if it encloses us. */ |
290 | if (f > rg->from) | |
291 | f = rg->from; | |
292 | ||
293 | /* Check for and consume any regions we now overlap with. */ | |
294 | nrg = rg; | |
295 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
296 | if (&rg->link == head) | |
297 | break; | |
298 | if (rg->from > t) | |
299 | break; | |
300 | ||
301 | /* If this area reaches higher then extend our area to | |
302 | * include it completely. If this is not the first area | |
303 | * which we intend to reuse, free it. */ | |
304 | if (rg->to > t) | |
305 | t = rg->to; | |
306 | if (rg != nrg) { | |
cf3ad20b MK |
307 | /* Decrement return value by the deleted range. |
308 | * Another range will span this area so that by | |
309 | * end of routine add will be >= zero | |
310 | */ | |
311 | add -= (rg->to - rg->from); | |
96822904 AW |
312 | list_del(&rg->link); |
313 | kfree(rg); | |
314 | } | |
315 | } | |
cf3ad20b MK |
316 | |
317 | add += (nrg->from - f); /* Added to beginning of region */ | |
96822904 | 318 | nrg->from = f; |
cf3ad20b | 319 | add += t - nrg->to; /* Added to end of region */ |
96822904 | 320 | nrg->to = t; |
cf3ad20b | 321 | |
5e911373 MK |
322 | out_locked: |
323 | resv->adds_in_progress--; | |
7b24d861 | 324 | spin_unlock(&resv->lock); |
cf3ad20b MK |
325 | VM_BUG_ON(add < 0); |
326 | return add; | |
96822904 AW |
327 | } |
328 | ||
1dd308a7 MK |
329 | /* |
330 | * Examine the existing reserve map and determine how many | |
331 | * huge pages in the specified range [f, t) are NOT currently | |
332 | * represented. This routine is called before a subsequent | |
333 | * call to region_add that will actually modify the reserve | |
334 | * map to add the specified range [f, t). region_chg does | |
335 | * not change the number of huge pages represented by the | |
336 | * map. However, if the existing regions in the map can not | |
337 | * be expanded to represent the new range, a new file_region | |
338 | * structure is added to the map as a placeholder. This is | |
339 | * so that the subsequent region_add call will have all the | |
340 | * regions it needs and will not fail. | |
341 | * | |
5e911373 MK |
342 | * Upon entry, region_chg will also examine the cache of region descriptors |
343 | * associated with the map. If there are not enough descriptors cached, one | |
344 | * will be allocated for the in progress add operation. | |
345 | * | |
346 | * Returns the number of huge pages that need to be added to the existing | |
347 | * reservation map for the range [f, t). This number is greater or equal to | |
348 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
349 | * is needed and can not be allocated. | |
1dd308a7 | 350 | */ |
1406ec9b | 351 | static long region_chg(struct resv_map *resv, long f, long t) |
96822904 | 352 | { |
1406ec9b | 353 | struct list_head *head = &resv->regions; |
7b24d861 | 354 | struct file_region *rg, *nrg = NULL; |
96822904 AW |
355 | long chg = 0; |
356 | ||
7b24d861 DB |
357 | retry: |
358 | spin_lock(&resv->lock); | |
5e911373 MK |
359 | retry_locked: |
360 | resv->adds_in_progress++; | |
361 | ||
362 | /* | |
363 | * Check for sufficient descriptors in the cache to accommodate | |
364 | * the number of in progress add operations. | |
365 | */ | |
366 | if (resv->adds_in_progress > resv->region_cache_count) { | |
367 | struct file_region *trg; | |
368 | ||
369 | VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1); | |
370 | /* Must drop lock to allocate a new descriptor. */ | |
371 | resv->adds_in_progress--; | |
372 | spin_unlock(&resv->lock); | |
373 | ||
374 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
375 | if (!trg) | |
376 | return -ENOMEM; | |
377 | ||
378 | spin_lock(&resv->lock); | |
379 | list_add(&trg->link, &resv->region_cache); | |
380 | resv->region_cache_count++; | |
381 | goto retry_locked; | |
382 | } | |
383 | ||
96822904 AW |
384 | /* Locate the region we are before or in. */ |
385 | list_for_each_entry(rg, head, link) | |
386 | if (f <= rg->to) | |
387 | break; | |
388 | ||
389 | /* If we are below the current region then a new region is required. | |
390 | * Subtle, allocate a new region at the position but make it zero | |
391 | * size such that we can guarantee to record the reservation. */ | |
392 | if (&rg->link == head || t < rg->from) { | |
7b24d861 | 393 | if (!nrg) { |
5e911373 | 394 | resv->adds_in_progress--; |
7b24d861 DB |
395 | spin_unlock(&resv->lock); |
396 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
397 | if (!nrg) | |
398 | return -ENOMEM; | |
399 | ||
400 | nrg->from = f; | |
401 | nrg->to = f; | |
402 | INIT_LIST_HEAD(&nrg->link); | |
403 | goto retry; | |
404 | } | |
96822904 | 405 | |
7b24d861 DB |
406 | list_add(&nrg->link, rg->link.prev); |
407 | chg = t - f; | |
408 | goto out_nrg; | |
96822904 AW |
409 | } |
410 | ||
411 | /* Round our left edge to the current segment if it encloses us. */ | |
412 | if (f > rg->from) | |
413 | f = rg->from; | |
414 | chg = t - f; | |
415 | ||
416 | /* Check for and consume any regions we now overlap with. */ | |
417 | list_for_each_entry(rg, rg->link.prev, link) { | |
418 | if (&rg->link == head) | |
419 | break; | |
420 | if (rg->from > t) | |
7b24d861 | 421 | goto out; |
96822904 | 422 | |
25985edc | 423 | /* We overlap with this area, if it extends further than |
96822904 AW |
424 | * us then we must extend ourselves. Account for its |
425 | * existing reservation. */ | |
426 | if (rg->to > t) { | |
427 | chg += rg->to - t; | |
428 | t = rg->to; | |
429 | } | |
430 | chg -= rg->to - rg->from; | |
431 | } | |
7b24d861 DB |
432 | |
433 | out: | |
434 | spin_unlock(&resv->lock); | |
435 | /* We already know we raced and no longer need the new region */ | |
436 | kfree(nrg); | |
437 | return chg; | |
438 | out_nrg: | |
439 | spin_unlock(&resv->lock); | |
96822904 AW |
440 | return chg; |
441 | } | |
442 | ||
5e911373 MK |
443 | /* |
444 | * Abort the in progress add operation. The adds_in_progress field | |
445 | * of the resv_map keeps track of the operations in progress between | |
446 | * calls to region_chg and region_add. Operations are sometimes | |
447 | * aborted after the call to region_chg. In such cases, region_abort | |
448 | * is called to decrement the adds_in_progress counter. | |
449 | * | |
450 | * NOTE: The range arguments [f, t) are not needed or used in this | |
451 | * routine. They are kept to make reading the calling code easier as | |
452 | * arguments will match the associated region_chg call. | |
453 | */ | |
454 | static void region_abort(struct resv_map *resv, long f, long t) | |
455 | { | |
456 | spin_lock(&resv->lock); | |
457 | VM_BUG_ON(!resv->region_cache_count); | |
458 | resv->adds_in_progress--; | |
459 | spin_unlock(&resv->lock); | |
460 | } | |
461 | ||
1dd308a7 | 462 | /* |
feba16e2 MK |
463 | * Delete the specified range [f, t) from the reserve map. If the |
464 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
465 | * should be deleted. Locate the regions which intersect [f, t) | |
466 | * and either trim, delete or split the existing regions. | |
467 | * | |
468 | * Returns the number of huge pages deleted from the reserve map. | |
469 | * In the normal case, the return value is zero or more. In the | |
470 | * case where a region must be split, a new region descriptor must | |
471 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
472 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
473 | * a region and possibly return -ENOMEM. Callers specifying | |
474 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 475 | */ |
feba16e2 | 476 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 477 | { |
1406ec9b | 478 | struct list_head *head = &resv->regions; |
96822904 | 479 | struct file_region *rg, *trg; |
feba16e2 MK |
480 | struct file_region *nrg = NULL; |
481 | long del = 0; | |
96822904 | 482 | |
feba16e2 | 483 | retry: |
7b24d861 | 484 | spin_lock(&resv->lock); |
feba16e2 MK |
485 | list_for_each_entry_safe(rg, trg, head, link) { |
486 | if (rg->to <= f) | |
487 | continue; | |
488 | if (rg->from >= t) | |
96822904 | 489 | break; |
96822904 | 490 | |
feba16e2 MK |
491 | if (f > rg->from && t < rg->to) { /* Must split region */ |
492 | /* | |
493 | * Check for an entry in the cache before dropping | |
494 | * lock and attempting allocation. | |
495 | */ | |
496 | if (!nrg && | |
497 | resv->region_cache_count > resv->adds_in_progress) { | |
498 | nrg = list_first_entry(&resv->region_cache, | |
499 | struct file_region, | |
500 | link); | |
501 | list_del(&nrg->link); | |
502 | resv->region_cache_count--; | |
503 | } | |
96822904 | 504 | |
feba16e2 MK |
505 | if (!nrg) { |
506 | spin_unlock(&resv->lock); | |
507 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
508 | if (!nrg) | |
509 | return -ENOMEM; | |
510 | goto retry; | |
511 | } | |
512 | ||
513 | del += t - f; | |
514 | ||
515 | /* New entry for end of split region */ | |
516 | nrg->from = t; | |
517 | nrg->to = rg->to; | |
518 | INIT_LIST_HEAD(&nrg->link); | |
519 | ||
520 | /* Original entry is trimmed */ | |
521 | rg->to = f; | |
522 | ||
523 | list_add(&nrg->link, &rg->link); | |
524 | nrg = NULL; | |
96822904 | 525 | break; |
feba16e2 MK |
526 | } |
527 | ||
528 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
529 | del += rg->to - rg->from; | |
530 | list_del(&rg->link); | |
531 | kfree(rg); | |
532 | continue; | |
533 | } | |
534 | ||
535 | if (f <= rg->from) { /* Trim beginning of region */ | |
536 | del += t - rg->from; | |
537 | rg->from = t; | |
538 | } else { /* Trim end of region */ | |
539 | del += rg->to - f; | |
540 | rg->to = f; | |
541 | } | |
96822904 | 542 | } |
7b24d861 | 543 | |
7b24d861 | 544 | spin_unlock(&resv->lock); |
feba16e2 MK |
545 | kfree(nrg); |
546 | return del; | |
96822904 AW |
547 | } |
548 | ||
b5cec28d MK |
549 | /* |
550 | * A rare out of memory error was encountered which prevented removal of | |
551 | * the reserve map region for a page. The huge page itself was free'ed | |
552 | * and removed from the page cache. This routine will adjust the subpool | |
553 | * usage count, and the global reserve count if needed. By incrementing | |
554 | * these counts, the reserve map entry which could not be deleted will | |
555 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
556 | * counts. | |
557 | */ | |
558 | void hugetlb_fix_reserve_counts(struct inode *inode, bool restore_reserve) | |
559 | { | |
560 | struct hugepage_subpool *spool = subpool_inode(inode); | |
561 | long rsv_adjust; | |
562 | ||
563 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
564 | if (restore_reserve && rsv_adjust) { | |
565 | struct hstate *h = hstate_inode(inode); | |
566 | ||
567 | hugetlb_acct_memory(h, 1); | |
568 | } | |
569 | } | |
570 | ||
1dd308a7 MK |
571 | /* |
572 | * Count and return the number of huge pages in the reserve map | |
573 | * that intersect with the range [f, t). | |
574 | */ | |
1406ec9b | 575 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 576 | { |
1406ec9b | 577 | struct list_head *head = &resv->regions; |
84afd99b AW |
578 | struct file_region *rg; |
579 | long chg = 0; | |
580 | ||
7b24d861 | 581 | spin_lock(&resv->lock); |
84afd99b AW |
582 | /* Locate each segment we overlap with, and count that overlap. */ |
583 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
584 | long seg_from; |
585 | long seg_to; | |
84afd99b AW |
586 | |
587 | if (rg->to <= f) | |
588 | continue; | |
589 | if (rg->from >= t) | |
590 | break; | |
591 | ||
592 | seg_from = max(rg->from, f); | |
593 | seg_to = min(rg->to, t); | |
594 | ||
595 | chg += seg_to - seg_from; | |
596 | } | |
7b24d861 | 597 | spin_unlock(&resv->lock); |
84afd99b AW |
598 | |
599 | return chg; | |
600 | } | |
601 | ||
e7c4b0bf AW |
602 | /* |
603 | * Convert the address within this vma to the page offset within | |
604 | * the mapping, in pagecache page units; huge pages here. | |
605 | */ | |
a5516438 AK |
606 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
607 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 608 | { |
a5516438 AK |
609 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
610 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
611 | } |
612 | ||
0fe6e20b NH |
613 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
614 | unsigned long address) | |
615 | { | |
616 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
617 | } | |
618 | ||
08fba699 MG |
619 | /* |
620 | * Return the size of the pages allocated when backing a VMA. In the majority | |
621 | * cases this will be same size as used by the page table entries. | |
622 | */ | |
623 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
624 | { | |
625 | struct hstate *hstate; | |
626 | ||
627 | if (!is_vm_hugetlb_page(vma)) | |
628 | return PAGE_SIZE; | |
629 | ||
630 | hstate = hstate_vma(vma); | |
631 | ||
2415cf12 | 632 | return 1UL << huge_page_shift(hstate); |
08fba699 | 633 | } |
f340ca0f | 634 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 635 | |
3340289d MG |
636 | /* |
637 | * Return the page size being used by the MMU to back a VMA. In the majority | |
638 | * of cases, the page size used by the kernel matches the MMU size. On | |
639 | * architectures where it differs, an architecture-specific version of this | |
640 | * function is required. | |
641 | */ | |
642 | #ifndef vma_mmu_pagesize | |
643 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) | |
644 | { | |
645 | return vma_kernel_pagesize(vma); | |
646 | } | |
647 | #endif | |
648 | ||
84afd99b AW |
649 | /* |
650 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
651 | * bits of the reservation map pointer, which are always clear due to | |
652 | * alignment. | |
653 | */ | |
654 | #define HPAGE_RESV_OWNER (1UL << 0) | |
655 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 656 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 657 | |
a1e78772 MG |
658 | /* |
659 | * These helpers are used to track how many pages are reserved for | |
660 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
661 | * is guaranteed to have their future faults succeed. | |
662 | * | |
663 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
664 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
665 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
666 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
667 | * |
668 | * The private mapping reservation is represented in a subtly different | |
669 | * manner to a shared mapping. A shared mapping has a region map associated | |
670 | * with the underlying file, this region map represents the backing file | |
671 | * pages which have ever had a reservation assigned which this persists even | |
672 | * after the page is instantiated. A private mapping has a region map | |
673 | * associated with the original mmap which is attached to all VMAs which | |
674 | * reference it, this region map represents those offsets which have consumed | |
675 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 676 | */ |
e7c4b0bf AW |
677 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
678 | { | |
679 | return (unsigned long)vma->vm_private_data; | |
680 | } | |
681 | ||
682 | static void set_vma_private_data(struct vm_area_struct *vma, | |
683 | unsigned long value) | |
684 | { | |
685 | vma->vm_private_data = (void *)value; | |
686 | } | |
687 | ||
9119a41e | 688 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
689 | { |
690 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
691 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
692 | ||
693 | if (!resv_map || !rg) { | |
694 | kfree(resv_map); | |
695 | kfree(rg); | |
84afd99b | 696 | return NULL; |
5e911373 | 697 | } |
84afd99b AW |
698 | |
699 | kref_init(&resv_map->refs); | |
7b24d861 | 700 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
701 | INIT_LIST_HEAD(&resv_map->regions); |
702 | ||
5e911373 MK |
703 | resv_map->adds_in_progress = 0; |
704 | ||
705 | INIT_LIST_HEAD(&resv_map->region_cache); | |
706 | list_add(&rg->link, &resv_map->region_cache); | |
707 | resv_map->region_cache_count = 1; | |
708 | ||
84afd99b AW |
709 | return resv_map; |
710 | } | |
711 | ||
9119a41e | 712 | void resv_map_release(struct kref *ref) |
84afd99b AW |
713 | { |
714 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
715 | struct list_head *head = &resv_map->region_cache; |
716 | struct file_region *rg, *trg; | |
84afd99b AW |
717 | |
718 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 719 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
720 | |
721 | /* ... and any entries left in the cache */ | |
722 | list_for_each_entry_safe(rg, trg, head, link) { | |
723 | list_del(&rg->link); | |
724 | kfree(rg); | |
725 | } | |
726 | ||
727 | VM_BUG_ON(resv_map->adds_in_progress); | |
728 | ||
84afd99b AW |
729 | kfree(resv_map); |
730 | } | |
731 | ||
4e35f483 JK |
732 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
733 | { | |
734 | return inode->i_mapping->private_data; | |
735 | } | |
736 | ||
84afd99b | 737 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 738 | { |
81d1b09c | 739 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
740 | if (vma->vm_flags & VM_MAYSHARE) { |
741 | struct address_space *mapping = vma->vm_file->f_mapping; | |
742 | struct inode *inode = mapping->host; | |
743 | ||
744 | return inode_resv_map(inode); | |
745 | ||
746 | } else { | |
84afd99b AW |
747 | return (struct resv_map *)(get_vma_private_data(vma) & |
748 | ~HPAGE_RESV_MASK); | |
4e35f483 | 749 | } |
a1e78772 MG |
750 | } |
751 | ||
84afd99b | 752 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 753 | { |
81d1b09c SL |
754 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
755 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 756 | |
84afd99b AW |
757 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
758 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
759 | } |
760 | ||
761 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
762 | { | |
81d1b09c SL |
763 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
764 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
765 | |
766 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
767 | } |
768 | ||
769 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
770 | { | |
81d1b09c | 771 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
772 | |
773 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
774 | } |
775 | ||
04f2cbe3 | 776 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
777 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
778 | { | |
81d1b09c | 779 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 780 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
781 | vma->vm_private_data = (void *)0; |
782 | } | |
783 | ||
784 | /* Returns true if the VMA has associated reserve pages */ | |
559ec2f8 | 785 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 786 | { |
af0ed73e JK |
787 | if (vma->vm_flags & VM_NORESERVE) { |
788 | /* | |
789 | * This address is already reserved by other process(chg == 0), | |
790 | * so, we should decrement reserved count. Without decrementing, | |
791 | * reserve count remains after releasing inode, because this | |
792 | * allocated page will go into page cache and is regarded as | |
793 | * coming from reserved pool in releasing step. Currently, we | |
794 | * don't have any other solution to deal with this situation | |
795 | * properly, so add work-around here. | |
796 | */ | |
797 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 798 | return true; |
af0ed73e | 799 | else |
559ec2f8 | 800 | return false; |
af0ed73e | 801 | } |
a63884e9 JK |
802 | |
803 | /* Shared mappings always use reserves */ | |
f83a275d | 804 | if (vma->vm_flags & VM_MAYSHARE) |
559ec2f8 | 805 | return true; |
a63884e9 JK |
806 | |
807 | /* | |
808 | * Only the process that called mmap() has reserves for | |
809 | * private mappings. | |
810 | */ | |
7f09ca51 | 811 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
559ec2f8 | 812 | return true; |
a63884e9 | 813 | |
559ec2f8 | 814 | return false; |
a1e78772 MG |
815 | } |
816 | ||
a5516438 | 817 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
818 | { |
819 | int nid = page_to_nid(page); | |
0edaecfa | 820 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
821 | h->free_huge_pages++; |
822 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
823 | } |
824 | ||
bf50bab2 NH |
825 | static struct page *dequeue_huge_page_node(struct hstate *h, int nid) |
826 | { | |
827 | struct page *page; | |
828 | ||
c8721bbb NH |
829 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
830 | if (!is_migrate_isolate_page(page)) | |
831 | break; | |
832 | /* | |
833 | * if 'non-isolated free hugepage' not found on the list, | |
834 | * the allocation fails. | |
835 | */ | |
836 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 837 | return NULL; |
0edaecfa | 838 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 839 | set_page_refcounted(page); |
bf50bab2 NH |
840 | h->free_huge_pages--; |
841 | h->free_huge_pages_node[nid]--; | |
842 | return page; | |
843 | } | |
844 | ||
86cdb465 NH |
845 | /* Movability of hugepages depends on migration support. */ |
846 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
847 | { | |
100873d7 | 848 | if (hugepages_treat_as_movable || hugepage_migration_supported(h)) |
86cdb465 NH |
849 | return GFP_HIGHUSER_MOVABLE; |
850 | else | |
851 | return GFP_HIGHUSER; | |
852 | } | |
853 | ||
a5516438 AK |
854 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
855 | struct vm_area_struct *vma, | |
af0ed73e JK |
856 | unsigned long address, int avoid_reserve, |
857 | long chg) | |
1da177e4 | 858 | { |
b1c12cbc | 859 | struct page *page = NULL; |
480eccf9 | 860 | struct mempolicy *mpol; |
19770b32 | 861 | nodemask_t *nodemask; |
c0ff7453 | 862 | struct zonelist *zonelist; |
dd1a239f MG |
863 | struct zone *zone; |
864 | struct zoneref *z; | |
cc9a6c87 | 865 | unsigned int cpuset_mems_cookie; |
1da177e4 | 866 | |
a1e78772 MG |
867 | /* |
868 | * A child process with MAP_PRIVATE mappings created by their parent | |
869 | * have no page reserves. This check ensures that reservations are | |
870 | * not "stolen". The child may still get SIGKILLed | |
871 | */ | |
af0ed73e | 872 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 873 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 874 | goto err; |
a1e78772 | 875 | |
04f2cbe3 | 876 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 877 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 878 | goto err; |
04f2cbe3 | 879 | |
9966c4bb | 880 | retry_cpuset: |
d26914d1 | 881 | cpuset_mems_cookie = read_mems_allowed_begin(); |
9966c4bb | 882 | zonelist = huge_zonelist(vma, address, |
86cdb465 | 883 | htlb_alloc_mask(h), &mpol, &nodemask); |
9966c4bb | 884 | |
19770b32 MG |
885 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
886 | MAX_NR_ZONES - 1, nodemask) { | |
344736f2 | 887 | if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) { |
bf50bab2 NH |
888 | page = dequeue_huge_page_node(h, zone_to_nid(zone)); |
889 | if (page) { | |
af0ed73e JK |
890 | if (avoid_reserve) |
891 | break; | |
892 | if (!vma_has_reserves(vma, chg)) | |
893 | break; | |
894 | ||
07443a85 | 895 | SetPagePrivate(page); |
af0ed73e | 896 | h->resv_huge_pages--; |
bf50bab2 NH |
897 | break; |
898 | } | |
3abf7afd | 899 | } |
1da177e4 | 900 | } |
cc9a6c87 | 901 | |
52cd3b07 | 902 | mpol_cond_put(mpol); |
d26914d1 | 903 | if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) |
cc9a6c87 | 904 | goto retry_cpuset; |
1da177e4 | 905 | return page; |
cc9a6c87 MG |
906 | |
907 | err: | |
cc9a6c87 | 908 | return NULL; |
1da177e4 LT |
909 | } |
910 | ||
1cac6f2c LC |
911 | /* |
912 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
913 | * We may have allocated or freed a huge page based on a different | |
914 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
915 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
916 | * node for alloc or free. | |
917 | */ | |
918 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
919 | { | |
920 | nid = next_node(nid, *nodes_allowed); | |
921 | if (nid == MAX_NUMNODES) | |
922 | nid = first_node(*nodes_allowed); | |
923 | VM_BUG_ON(nid >= MAX_NUMNODES); | |
924 | ||
925 | return nid; | |
926 | } | |
927 | ||
928 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
929 | { | |
930 | if (!node_isset(nid, *nodes_allowed)) | |
931 | nid = next_node_allowed(nid, nodes_allowed); | |
932 | return nid; | |
933 | } | |
934 | ||
935 | /* | |
936 | * returns the previously saved node ["this node"] from which to | |
937 | * allocate a persistent huge page for the pool and advance the | |
938 | * next node from which to allocate, handling wrap at end of node | |
939 | * mask. | |
940 | */ | |
941 | static int hstate_next_node_to_alloc(struct hstate *h, | |
942 | nodemask_t *nodes_allowed) | |
943 | { | |
944 | int nid; | |
945 | ||
946 | VM_BUG_ON(!nodes_allowed); | |
947 | ||
948 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
949 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
950 | ||
951 | return nid; | |
952 | } | |
953 | ||
954 | /* | |
955 | * helper for free_pool_huge_page() - return the previously saved | |
956 | * node ["this node"] from which to free a huge page. Advance the | |
957 | * next node id whether or not we find a free huge page to free so | |
958 | * that the next attempt to free addresses the next node. | |
959 | */ | |
960 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
961 | { | |
962 | int nid; | |
963 | ||
964 | VM_BUG_ON(!nodes_allowed); | |
965 | ||
966 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
967 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
968 | ||
969 | return nid; | |
970 | } | |
971 | ||
972 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
973 | for (nr_nodes = nodes_weight(*mask); \ | |
974 | nr_nodes > 0 && \ | |
975 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
976 | nr_nodes--) | |
977 | ||
978 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
979 | for (nr_nodes = nodes_weight(*mask); \ | |
980 | nr_nodes > 0 && \ | |
981 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
982 | nr_nodes--) | |
983 | ||
944d9fec LC |
984 | #if defined(CONFIG_CMA) && defined(CONFIG_X86_64) |
985 | static void destroy_compound_gigantic_page(struct page *page, | |
986 | unsigned long order) | |
987 | { | |
988 | int i; | |
989 | int nr_pages = 1 << order; | |
990 | struct page *p = page + 1; | |
991 | ||
992 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { | |
993 | __ClearPageTail(p); | |
994 | set_page_refcounted(p); | |
995 | p->first_page = NULL; | |
996 | } | |
997 | ||
998 | set_compound_order(page, 0); | |
999 | __ClearPageHead(page); | |
1000 | } | |
1001 | ||
1002 | static void free_gigantic_page(struct page *page, unsigned order) | |
1003 | { | |
1004 | free_contig_range(page_to_pfn(page), 1 << order); | |
1005 | } | |
1006 | ||
1007 | static int __alloc_gigantic_page(unsigned long start_pfn, | |
1008 | unsigned long nr_pages) | |
1009 | { | |
1010 | unsigned long end_pfn = start_pfn + nr_pages; | |
1011 | return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE); | |
1012 | } | |
1013 | ||
1014 | static bool pfn_range_valid_gigantic(unsigned long start_pfn, | |
1015 | unsigned long nr_pages) | |
1016 | { | |
1017 | unsigned long i, end_pfn = start_pfn + nr_pages; | |
1018 | struct page *page; | |
1019 | ||
1020 | for (i = start_pfn; i < end_pfn; i++) { | |
1021 | if (!pfn_valid(i)) | |
1022 | return false; | |
1023 | ||
1024 | page = pfn_to_page(i); | |
1025 | ||
1026 | if (PageReserved(page)) | |
1027 | return false; | |
1028 | ||
1029 | if (page_count(page) > 0) | |
1030 | return false; | |
1031 | ||
1032 | if (PageHuge(page)) | |
1033 | return false; | |
1034 | } | |
1035 | ||
1036 | return true; | |
1037 | } | |
1038 | ||
1039 | static bool zone_spans_last_pfn(const struct zone *zone, | |
1040 | unsigned long start_pfn, unsigned long nr_pages) | |
1041 | { | |
1042 | unsigned long last_pfn = start_pfn + nr_pages - 1; | |
1043 | return zone_spans_pfn(zone, last_pfn); | |
1044 | } | |
1045 | ||
1046 | static struct page *alloc_gigantic_page(int nid, unsigned order) | |
1047 | { | |
1048 | unsigned long nr_pages = 1 << order; | |
1049 | unsigned long ret, pfn, flags; | |
1050 | struct zone *z; | |
1051 | ||
1052 | z = NODE_DATA(nid)->node_zones; | |
1053 | for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) { | |
1054 | spin_lock_irqsave(&z->lock, flags); | |
1055 | ||
1056 | pfn = ALIGN(z->zone_start_pfn, nr_pages); | |
1057 | while (zone_spans_last_pfn(z, pfn, nr_pages)) { | |
1058 | if (pfn_range_valid_gigantic(pfn, nr_pages)) { | |
1059 | /* | |
1060 | * We release the zone lock here because | |
1061 | * alloc_contig_range() will also lock the zone | |
1062 | * at some point. If there's an allocation | |
1063 | * spinning on this lock, it may win the race | |
1064 | * and cause alloc_contig_range() to fail... | |
1065 | */ | |
1066 | spin_unlock_irqrestore(&z->lock, flags); | |
1067 | ret = __alloc_gigantic_page(pfn, nr_pages); | |
1068 | if (!ret) | |
1069 | return pfn_to_page(pfn); | |
1070 | spin_lock_irqsave(&z->lock, flags); | |
1071 | } | |
1072 | pfn += nr_pages; | |
1073 | } | |
1074 | ||
1075 | spin_unlock_irqrestore(&z->lock, flags); | |
1076 | } | |
1077 | ||
1078 | return NULL; | |
1079 | } | |
1080 | ||
1081 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
1082 | static void prep_compound_gigantic_page(struct page *page, unsigned long order); | |
1083 | ||
1084 | static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) | |
1085 | { | |
1086 | struct page *page; | |
1087 | ||
1088 | page = alloc_gigantic_page(nid, huge_page_order(h)); | |
1089 | if (page) { | |
1090 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
1091 | prep_new_huge_page(h, page, nid); | |
1092 | } | |
1093 | ||
1094 | return page; | |
1095 | } | |
1096 | ||
1097 | static int alloc_fresh_gigantic_page(struct hstate *h, | |
1098 | nodemask_t *nodes_allowed) | |
1099 | { | |
1100 | struct page *page = NULL; | |
1101 | int nr_nodes, node; | |
1102 | ||
1103 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1104 | page = alloc_fresh_gigantic_page_node(h, node); | |
1105 | if (page) | |
1106 | return 1; | |
1107 | } | |
1108 | ||
1109 | return 0; | |
1110 | } | |
1111 | ||
1112 | static inline bool gigantic_page_supported(void) { return true; } | |
1113 | #else | |
1114 | static inline bool gigantic_page_supported(void) { return false; } | |
1115 | static inline void free_gigantic_page(struct page *page, unsigned order) { } | |
1116 | static inline void destroy_compound_gigantic_page(struct page *page, | |
1117 | unsigned long order) { } | |
1118 | static inline int alloc_fresh_gigantic_page(struct hstate *h, | |
1119 | nodemask_t *nodes_allowed) { return 0; } | |
1120 | #endif | |
1121 | ||
a5516438 | 1122 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1123 | { |
1124 | int i; | |
a5516438 | 1125 | |
944d9fec LC |
1126 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
1127 | return; | |
18229df5 | 1128 | |
a5516438 AK |
1129 | h->nr_huge_pages--; |
1130 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
1131 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
1132 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
1133 | 1 << PG_referenced | 1 << PG_dirty | | |
a7407a27 LC |
1134 | 1 << PG_active | 1 << PG_private | |
1135 | 1 << PG_writeback); | |
6af2acb6 | 1136 | } |
309381fe | 1137 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
6af2acb6 AL |
1138 | set_compound_page_dtor(page, NULL); |
1139 | set_page_refcounted(page); | |
944d9fec LC |
1140 | if (hstate_is_gigantic(h)) { |
1141 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
1142 | free_gigantic_page(page, huge_page_order(h)); | |
1143 | } else { | |
944d9fec LC |
1144 | __free_pages(page, huge_page_order(h)); |
1145 | } | |
6af2acb6 AL |
1146 | } |
1147 | ||
e5ff2159 AK |
1148 | struct hstate *size_to_hstate(unsigned long size) |
1149 | { | |
1150 | struct hstate *h; | |
1151 | ||
1152 | for_each_hstate(h) { | |
1153 | if (huge_page_size(h) == size) | |
1154 | return h; | |
1155 | } | |
1156 | return NULL; | |
1157 | } | |
1158 | ||
bcc54222 NH |
1159 | /* |
1160 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
1161 | * to hstate->hugepage_activelist.) | |
1162 | * | |
1163 | * This function can be called for tail pages, but never returns true for them. | |
1164 | */ | |
1165 | bool page_huge_active(struct page *page) | |
1166 | { | |
1167 | VM_BUG_ON_PAGE(!PageHuge(page), page); | |
1168 | return PageHead(page) && PagePrivate(&page[1]); | |
1169 | } | |
1170 | ||
1171 | /* never called for tail page */ | |
1172 | static void set_page_huge_active(struct page *page) | |
1173 | { | |
1174 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1175 | SetPagePrivate(&page[1]); | |
1176 | } | |
1177 | ||
1178 | static void clear_page_huge_active(struct page *page) | |
1179 | { | |
1180 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1181 | ClearPagePrivate(&page[1]); | |
1182 | } | |
1183 | ||
8f1d26d0 | 1184 | void free_huge_page(struct page *page) |
27a85ef1 | 1185 | { |
a5516438 AK |
1186 | /* |
1187 | * Can't pass hstate in here because it is called from the | |
1188 | * compound page destructor. | |
1189 | */ | |
e5ff2159 | 1190 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1191 | int nid = page_to_nid(page); |
90481622 DG |
1192 | struct hugepage_subpool *spool = |
1193 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 1194 | bool restore_reserve; |
27a85ef1 | 1195 | |
e5df70ab | 1196 | set_page_private(page, 0); |
23be7468 | 1197 | page->mapping = NULL; |
7893d1d5 | 1198 | BUG_ON(page_count(page)); |
0fe6e20b | 1199 | BUG_ON(page_mapcount(page)); |
07443a85 | 1200 | restore_reserve = PagePrivate(page); |
16c794b4 | 1201 | ClearPagePrivate(page); |
27a85ef1 | 1202 | |
1c5ecae3 MK |
1203 | /* |
1204 | * A return code of zero implies that the subpool will be under its | |
1205 | * minimum size if the reservation is not restored after page is free. | |
1206 | * Therefore, force restore_reserve operation. | |
1207 | */ | |
1208 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1209 | restore_reserve = true; | |
1210 | ||
27a85ef1 | 1211 | spin_lock(&hugetlb_lock); |
bcc54222 | 1212 | clear_page_huge_active(page); |
6d76dcf4 AK |
1213 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1214 | pages_per_huge_page(h), page); | |
07443a85 JK |
1215 | if (restore_reserve) |
1216 | h->resv_huge_pages++; | |
1217 | ||
944d9fec | 1218 | if (h->surplus_huge_pages_node[nid]) { |
0edaecfa AK |
1219 | /* remove the page from active list */ |
1220 | list_del(&page->lru); | |
a5516438 AK |
1221 | update_and_free_page(h, page); |
1222 | h->surplus_huge_pages--; | |
1223 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 1224 | } else { |
5d3a551c | 1225 | arch_clear_hugepage_flags(page); |
a5516438 | 1226 | enqueue_huge_page(h, page); |
7893d1d5 | 1227 | } |
27a85ef1 DG |
1228 | spin_unlock(&hugetlb_lock); |
1229 | } | |
1230 | ||
a5516438 | 1231 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1232 | { |
0edaecfa | 1233 | INIT_LIST_HEAD(&page->lru); |
b7ba30c6 AK |
1234 | set_compound_page_dtor(page, free_huge_page); |
1235 | spin_lock(&hugetlb_lock); | |
9dd540e2 | 1236 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
1237 | h->nr_huge_pages++; |
1238 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 AK |
1239 | spin_unlock(&hugetlb_lock); |
1240 | put_page(page); /* free it into the hugepage allocator */ | |
1241 | } | |
1242 | ||
2906dd52 | 1243 | static void prep_compound_gigantic_page(struct page *page, unsigned long order) |
20a0307c WF |
1244 | { |
1245 | int i; | |
1246 | int nr_pages = 1 << order; | |
1247 | struct page *p = page + 1; | |
1248 | ||
1249 | /* we rely on prep_new_huge_page to set the destructor */ | |
1250 | set_compound_order(page, order); | |
1251 | __SetPageHead(page); | |
ef5a22be | 1252 | __ClearPageReserved(page); |
20a0307c | 1253 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1254 | /* |
1255 | * For gigantic hugepages allocated through bootmem at | |
1256 | * boot, it's safer to be consistent with the not-gigantic | |
1257 | * hugepages and clear the PG_reserved bit from all tail pages | |
1258 | * too. Otherwse drivers using get_user_pages() to access tail | |
1259 | * pages may get the reference counting wrong if they see | |
1260 | * PG_reserved set on a tail page (despite the head page not | |
1261 | * having PG_reserved set). Enforcing this consistency between | |
1262 | * head and tail pages allows drivers to optimize away a check | |
1263 | * on the head page when they need know if put_page() is needed | |
1264 | * after get_user_pages(). | |
1265 | */ | |
1266 | __ClearPageReserved(p); | |
58a84aa9 | 1267 | set_page_count(p, 0); |
20a0307c | 1268 | p->first_page = page; |
44fc8057 DR |
1269 | /* Make sure p->first_page is always valid for PageTail() */ |
1270 | smp_wmb(); | |
1271 | __SetPageTail(p); | |
20a0307c WF |
1272 | } |
1273 | } | |
1274 | ||
7795912c AM |
1275 | /* |
1276 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1277 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1278 | * details. | |
1279 | */ | |
20a0307c WF |
1280 | int PageHuge(struct page *page) |
1281 | { | |
20a0307c WF |
1282 | if (!PageCompound(page)) |
1283 | return 0; | |
1284 | ||
1285 | page = compound_head(page); | |
758f66a2 | 1286 | return get_compound_page_dtor(page) == free_huge_page; |
20a0307c | 1287 | } |
43131e14 NH |
1288 | EXPORT_SYMBOL_GPL(PageHuge); |
1289 | ||
27c73ae7 AA |
1290 | /* |
1291 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1292 | * normal or transparent huge pages. | |
1293 | */ | |
1294 | int PageHeadHuge(struct page *page_head) | |
1295 | { | |
27c73ae7 AA |
1296 | if (!PageHead(page_head)) |
1297 | return 0; | |
1298 | ||
758f66a2 | 1299 | return get_compound_page_dtor(page_head) == free_huge_page; |
27c73ae7 | 1300 | } |
27c73ae7 | 1301 | |
13d60f4b ZY |
1302 | pgoff_t __basepage_index(struct page *page) |
1303 | { | |
1304 | struct page *page_head = compound_head(page); | |
1305 | pgoff_t index = page_index(page_head); | |
1306 | unsigned long compound_idx; | |
1307 | ||
1308 | if (!PageHuge(page_head)) | |
1309 | return page_index(page); | |
1310 | ||
1311 | if (compound_order(page_head) >= MAX_ORDER) | |
1312 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1313 | else | |
1314 | compound_idx = page - page_head; | |
1315 | ||
1316 | return (index << compound_order(page_head)) + compound_idx; | |
1317 | } | |
1318 | ||
a5516438 | 1319 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
1da177e4 | 1320 | { |
1da177e4 | 1321 | struct page *page; |
f96efd58 | 1322 | |
6484eb3e | 1323 | page = alloc_pages_exact_node(nid, |
86cdb465 | 1324 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
551883ae | 1325 | __GFP_REPEAT|__GFP_NOWARN, |
a5516438 | 1326 | huge_page_order(h)); |
1da177e4 | 1327 | if (page) { |
a5516438 | 1328 | prep_new_huge_page(h, page, nid); |
1da177e4 | 1329 | } |
63b4613c NA |
1330 | |
1331 | return page; | |
1332 | } | |
1333 | ||
b2261026 JK |
1334 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
1335 | { | |
1336 | struct page *page; | |
1337 | int nr_nodes, node; | |
1338 | int ret = 0; | |
1339 | ||
1340 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1341 | page = alloc_fresh_huge_page_node(h, node); | |
1342 | if (page) { | |
1343 | ret = 1; | |
1344 | break; | |
1345 | } | |
1346 | } | |
1347 | ||
1348 | if (ret) | |
1349 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
1350 | else | |
1351 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
1352 | ||
1353 | return ret; | |
1354 | } | |
1355 | ||
e8c5c824 LS |
1356 | /* |
1357 | * Free huge page from pool from next node to free. | |
1358 | * Attempt to keep persistent huge pages more or less | |
1359 | * balanced over allowed nodes. | |
1360 | * Called with hugetlb_lock locked. | |
1361 | */ | |
6ae11b27 LS |
1362 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1363 | bool acct_surplus) | |
e8c5c824 | 1364 | { |
b2261026 | 1365 | int nr_nodes, node; |
e8c5c824 LS |
1366 | int ret = 0; |
1367 | ||
b2261026 | 1368 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1369 | /* |
1370 | * If we're returning unused surplus pages, only examine | |
1371 | * nodes with surplus pages. | |
1372 | */ | |
b2261026 JK |
1373 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1374 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1375 | struct page *page = |
b2261026 | 1376 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1377 | struct page, lru); |
1378 | list_del(&page->lru); | |
1379 | h->free_huge_pages--; | |
b2261026 | 1380 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1381 | if (acct_surplus) { |
1382 | h->surplus_huge_pages--; | |
b2261026 | 1383 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1384 | } |
e8c5c824 LS |
1385 | update_and_free_page(h, page); |
1386 | ret = 1; | |
9a76db09 | 1387 | break; |
e8c5c824 | 1388 | } |
b2261026 | 1389 | } |
e8c5c824 LS |
1390 | |
1391 | return ret; | |
1392 | } | |
1393 | ||
c8721bbb NH |
1394 | /* |
1395 | * Dissolve a given free hugepage into free buddy pages. This function does | |
1396 | * nothing for in-use (including surplus) hugepages. | |
1397 | */ | |
1398 | static void dissolve_free_huge_page(struct page *page) | |
1399 | { | |
1400 | spin_lock(&hugetlb_lock); | |
1401 | if (PageHuge(page) && !page_count(page)) { | |
1402 | struct hstate *h = page_hstate(page); | |
1403 | int nid = page_to_nid(page); | |
1404 | list_del(&page->lru); | |
1405 | h->free_huge_pages--; | |
1406 | h->free_huge_pages_node[nid]--; | |
1407 | update_and_free_page(h, page); | |
1408 | } | |
1409 | spin_unlock(&hugetlb_lock); | |
1410 | } | |
1411 | ||
1412 | /* | |
1413 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1414 | * make specified memory blocks removable from the system. | |
1415 | * Note that start_pfn should aligned with (minimum) hugepage size. | |
1416 | */ | |
1417 | void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) | |
1418 | { | |
c8721bbb | 1419 | unsigned long pfn; |
c8721bbb | 1420 | |
d0177639 LZ |
1421 | if (!hugepages_supported()) |
1422 | return; | |
1423 | ||
641844f5 NH |
1424 | VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order)); |
1425 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) | |
c8721bbb NH |
1426 | dissolve_free_huge_page(pfn_to_page(pfn)); |
1427 | } | |
1428 | ||
bf50bab2 | 1429 | static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) |
7893d1d5 AL |
1430 | { |
1431 | struct page *page; | |
bf50bab2 | 1432 | unsigned int r_nid; |
7893d1d5 | 1433 | |
bae7f4ae | 1434 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1435 | return NULL; |
1436 | ||
d1c3fb1f NA |
1437 | /* |
1438 | * Assume we will successfully allocate the surplus page to | |
1439 | * prevent racing processes from causing the surplus to exceed | |
1440 | * overcommit | |
1441 | * | |
1442 | * This however introduces a different race, where a process B | |
1443 | * tries to grow the static hugepage pool while alloc_pages() is | |
1444 | * called by process A. B will only examine the per-node | |
1445 | * counters in determining if surplus huge pages can be | |
1446 | * converted to normal huge pages in adjust_pool_surplus(). A | |
1447 | * won't be able to increment the per-node counter, until the | |
1448 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
1449 | * no more huge pages can be converted from surplus to normal | |
1450 | * state (and doesn't try to convert again). Thus, we have a | |
1451 | * case where a surplus huge page exists, the pool is grown, and | |
1452 | * the surplus huge page still exists after, even though it | |
1453 | * should just have been converted to a normal huge page. This | |
1454 | * does not leak memory, though, as the hugepage will be freed | |
1455 | * once it is out of use. It also does not allow the counters to | |
1456 | * go out of whack in adjust_pool_surplus() as we don't modify | |
1457 | * the node values until we've gotten the hugepage and only the | |
1458 | * per-node value is checked there. | |
1459 | */ | |
1460 | spin_lock(&hugetlb_lock); | |
a5516438 | 1461 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
1462 | spin_unlock(&hugetlb_lock); |
1463 | return NULL; | |
1464 | } else { | |
a5516438 AK |
1465 | h->nr_huge_pages++; |
1466 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
1467 | } |
1468 | spin_unlock(&hugetlb_lock); | |
1469 | ||
bf50bab2 | 1470 | if (nid == NUMA_NO_NODE) |
86cdb465 | 1471 | page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP| |
bf50bab2 NH |
1472 | __GFP_REPEAT|__GFP_NOWARN, |
1473 | huge_page_order(h)); | |
1474 | else | |
1475 | page = alloc_pages_exact_node(nid, | |
86cdb465 | 1476 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
bf50bab2 | 1477 | __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); |
d1c3fb1f NA |
1478 | |
1479 | spin_lock(&hugetlb_lock); | |
7893d1d5 | 1480 | if (page) { |
0edaecfa | 1481 | INIT_LIST_HEAD(&page->lru); |
bf50bab2 | 1482 | r_nid = page_to_nid(page); |
7893d1d5 | 1483 | set_compound_page_dtor(page, free_huge_page); |
9dd540e2 | 1484 | set_hugetlb_cgroup(page, NULL); |
d1c3fb1f NA |
1485 | /* |
1486 | * We incremented the global counters already | |
1487 | */ | |
bf50bab2 NH |
1488 | h->nr_huge_pages_node[r_nid]++; |
1489 | h->surplus_huge_pages_node[r_nid]++; | |
3b116300 | 1490 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 1491 | } else { |
a5516438 AK |
1492 | h->nr_huge_pages--; |
1493 | h->surplus_huge_pages--; | |
3b116300 | 1494 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 1495 | } |
d1c3fb1f | 1496 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1497 | |
1498 | return page; | |
1499 | } | |
1500 | ||
bf50bab2 NH |
1501 | /* |
1502 | * This allocation function is useful in the context where vma is irrelevant. | |
1503 | * E.g. soft-offlining uses this function because it only cares physical | |
1504 | * address of error page. | |
1505 | */ | |
1506 | struct page *alloc_huge_page_node(struct hstate *h, int nid) | |
1507 | { | |
4ef91848 | 1508 | struct page *page = NULL; |
bf50bab2 NH |
1509 | |
1510 | spin_lock(&hugetlb_lock); | |
4ef91848 JK |
1511 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
1512 | page = dequeue_huge_page_node(h, nid); | |
bf50bab2 NH |
1513 | spin_unlock(&hugetlb_lock); |
1514 | ||
94ae8ba7 | 1515 | if (!page) |
bf50bab2 NH |
1516 | page = alloc_buddy_huge_page(h, nid); |
1517 | ||
1518 | return page; | |
1519 | } | |
1520 | ||
e4e574b7 | 1521 | /* |
25985edc | 1522 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1523 | * of size 'delta'. |
1524 | */ | |
a5516438 | 1525 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1526 | { |
1527 | struct list_head surplus_list; | |
1528 | struct page *page, *tmp; | |
1529 | int ret, i; | |
1530 | int needed, allocated; | |
28073b02 | 1531 | bool alloc_ok = true; |
e4e574b7 | 1532 | |
a5516438 | 1533 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1534 | if (needed <= 0) { |
a5516438 | 1535 | h->resv_huge_pages += delta; |
e4e574b7 | 1536 | return 0; |
ac09b3a1 | 1537 | } |
e4e574b7 AL |
1538 | |
1539 | allocated = 0; | |
1540 | INIT_LIST_HEAD(&surplus_list); | |
1541 | ||
1542 | ret = -ENOMEM; | |
1543 | retry: | |
1544 | spin_unlock(&hugetlb_lock); | |
1545 | for (i = 0; i < needed; i++) { | |
bf50bab2 | 1546 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
28073b02 HD |
1547 | if (!page) { |
1548 | alloc_ok = false; | |
1549 | break; | |
1550 | } | |
e4e574b7 AL |
1551 | list_add(&page->lru, &surplus_list); |
1552 | } | |
28073b02 | 1553 | allocated += i; |
e4e574b7 AL |
1554 | |
1555 | /* | |
1556 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1557 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1558 | */ | |
1559 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1560 | needed = (h->resv_huge_pages + delta) - |
1561 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1562 | if (needed > 0) { |
1563 | if (alloc_ok) | |
1564 | goto retry; | |
1565 | /* | |
1566 | * We were not able to allocate enough pages to | |
1567 | * satisfy the entire reservation so we free what | |
1568 | * we've allocated so far. | |
1569 | */ | |
1570 | goto free; | |
1571 | } | |
e4e574b7 AL |
1572 | /* |
1573 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1574 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1575 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1576 | * allocator. Commit the entire reservation here to prevent another |
1577 | * process from stealing the pages as they are added to the pool but | |
1578 | * before they are reserved. | |
e4e574b7 AL |
1579 | */ |
1580 | needed += allocated; | |
a5516438 | 1581 | h->resv_huge_pages += delta; |
e4e574b7 | 1582 | ret = 0; |
a9869b83 | 1583 | |
19fc3f0a | 1584 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1585 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1586 | if ((--needed) < 0) |
1587 | break; | |
a9869b83 NH |
1588 | /* |
1589 | * This page is now managed by the hugetlb allocator and has | |
1590 | * no users -- drop the buddy allocator's reference. | |
1591 | */ | |
1592 | put_page_testzero(page); | |
309381fe | 1593 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 1594 | enqueue_huge_page(h, page); |
19fc3f0a | 1595 | } |
28073b02 | 1596 | free: |
b0365c8d | 1597 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1598 | |
1599 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1600 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1601 | put_page(page); | |
a9869b83 | 1602 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1603 | |
1604 | return ret; | |
1605 | } | |
1606 | ||
1607 | /* | |
1608 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
1609 | * allocated to satisfy the reservation must be explicitly freed if they were | |
1610 | * never used. | |
685f3457 | 1611 | * Called with hugetlb_lock held. |
e4e574b7 | 1612 | */ |
a5516438 AK |
1613 | static void return_unused_surplus_pages(struct hstate *h, |
1614 | unsigned long unused_resv_pages) | |
e4e574b7 | 1615 | { |
e4e574b7 AL |
1616 | unsigned long nr_pages; |
1617 | ||
ac09b3a1 | 1618 | /* Uncommit the reservation */ |
a5516438 | 1619 | h->resv_huge_pages -= unused_resv_pages; |
ac09b3a1 | 1620 | |
aa888a74 | 1621 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 1622 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1623 | return; |
1624 | ||
a5516438 | 1625 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1626 | |
685f3457 LS |
1627 | /* |
1628 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1629 | * evenly across all nodes with memory. Iterate across these nodes |
1630 | * until we can no longer free unreserved surplus pages. This occurs | |
1631 | * when the nodes with surplus pages have no free pages. | |
1632 | * free_pool_huge_page() will balance the the freed pages across the | |
1633 | * on-line nodes with memory and will handle the hstate accounting. | |
685f3457 LS |
1634 | */ |
1635 | while (nr_pages--) { | |
8cebfcd0 | 1636 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
685f3457 | 1637 | break; |
7848a4bf | 1638 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 AL |
1639 | } |
1640 | } | |
1641 | ||
5e911373 | 1642 | |
c37f9fb1 | 1643 | /* |
feba16e2 | 1644 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 1645 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
1646 | * |
1647 | * vma_needs_reservation is called to determine if the huge page at addr | |
1648 | * within the vma has an associated reservation. If a reservation is | |
1649 | * needed, the value 1 is returned. The caller is then responsible for | |
1650 | * managing the global reservation and subpool usage counts. After | |
1651 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
1652 | * to add the page to the reservation map. If the page allocation fails, |
1653 | * the reservation must be ended instead of committed. vma_end_reservation | |
1654 | * is called in such cases. | |
cf3ad20b MK |
1655 | * |
1656 | * In the normal case, vma_commit_reservation returns the same value | |
1657 | * as the preceding vma_needs_reservation call. The only time this | |
1658 | * is not the case is if a reserve map was changed between calls. It | |
1659 | * is the responsibility of the caller to notice the difference and | |
1660 | * take appropriate action. | |
c37f9fb1 | 1661 | */ |
5e911373 MK |
1662 | enum vma_resv_mode { |
1663 | VMA_NEEDS_RESV, | |
1664 | VMA_COMMIT_RESV, | |
feba16e2 | 1665 | VMA_END_RESV, |
5e911373 | 1666 | }; |
cf3ad20b MK |
1667 | static long __vma_reservation_common(struct hstate *h, |
1668 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 1669 | enum vma_resv_mode mode) |
c37f9fb1 | 1670 | { |
4e35f483 JK |
1671 | struct resv_map *resv; |
1672 | pgoff_t idx; | |
cf3ad20b | 1673 | long ret; |
c37f9fb1 | 1674 | |
4e35f483 JK |
1675 | resv = vma_resv_map(vma); |
1676 | if (!resv) | |
84afd99b | 1677 | return 1; |
c37f9fb1 | 1678 | |
4e35f483 | 1679 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
1680 | switch (mode) { |
1681 | case VMA_NEEDS_RESV: | |
cf3ad20b | 1682 | ret = region_chg(resv, idx, idx + 1); |
5e911373 MK |
1683 | break; |
1684 | case VMA_COMMIT_RESV: | |
1685 | ret = region_add(resv, idx, idx + 1); | |
1686 | break; | |
feba16e2 | 1687 | case VMA_END_RESV: |
5e911373 MK |
1688 | region_abort(resv, idx, idx + 1); |
1689 | ret = 0; | |
1690 | break; | |
1691 | default: | |
1692 | BUG(); | |
1693 | } | |
84afd99b | 1694 | |
4e35f483 | 1695 | if (vma->vm_flags & VM_MAYSHARE) |
cf3ad20b | 1696 | return ret; |
4e35f483 | 1697 | else |
cf3ad20b | 1698 | return ret < 0 ? ret : 0; |
c37f9fb1 | 1699 | } |
cf3ad20b MK |
1700 | |
1701 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 1702 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 1703 | { |
5e911373 | 1704 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 1705 | } |
84afd99b | 1706 | |
cf3ad20b MK |
1707 | static long vma_commit_reservation(struct hstate *h, |
1708 | struct vm_area_struct *vma, unsigned long addr) | |
1709 | { | |
5e911373 MK |
1710 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
1711 | } | |
1712 | ||
feba16e2 | 1713 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
1714 | struct vm_area_struct *vma, unsigned long addr) |
1715 | { | |
feba16e2 | 1716 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
1717 | } |
1718 | ||
a1e78772 | 1719 | static struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 1720 | unsigned long addr, int avoid_reserve) |
1da177e4 | 1721 | { |
90481622 | 1722 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 1723 | struct hstate *h = hstate_vma(vma); |
348ea204 | 1724 | struct page *page; |
33039678 | 1725 | long chg, commit; |
6d76dcf4 AK |
1726 | int ret, idx; |
1727 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 1728 | |
6d76dcf4 | 1729 | idx = hstate_index(h); |
a1e78772 | 1730 | /* |
90481622 DG |
1731 | * Processes that did not create the mapping will have no |
1732 | * reserves and will not have accounted against subpool | |
1733 | * limit. Check that the subpool limit can be made before | |
1734 | * satisfying the allocation MAP_NORESERVE mappings may also | |
1735 | * need pages and subpool limit allocated allocated if no reserve | |
1736 | * mapping overlaps. | |
a1e78772 | 1737 | */ |
a5516438 | 1738 | chg = vma_needs_reservation(h, vma, addr); |
c37f9fb1 | 1739 | if (chg < 0) |
76dcee75 | 1740 | return ERR_PTR(-ENOMEM); |
8bb3f12e | 1741 | if (chg || avoid_reserve) |
5e911373 | 1742 | if (hugepage_subpool_get_pages(spool, 1) < 0) { |
feba16e2 | 1743 | vma_end_reservation(h, vma, addr); |
76dcee75 | 1744 | return ERR_PTR(-ENOSPC); |
5e911373 | 1745 | } |
1da177e4 | 1746 | |
6d76dcf4 | 1747 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f JZ |
1748 | if (ret) |
1749 | goto out_subpool_put; | |
1750 | ||
1da177e4 | 1751 | spin_lock(&hugetlb_lock); |
af0ed73e | 1752 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg); |
81a6fcae | 1753 | if (!page) { |
94ae8ba7 | 1754 | spin_unlock(&hugetlb_lock); |
bf50bab2 | 1755 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
8f34af6f JZ |
1756 | if (!page) |
1757 | goto out_uncharge_cgroup; | |
1758 | ||
79dbb236 AK |
1759 | spin_lock(&hugetlb_lock); |
1760 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 1761 | /* Fall through */ |
68842c9b | 1762 | } |
81a6fcae JK |
1763 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
1764 | spin_unlock(&hugetlb_lock); | |
348ea204 | 1765 | |
90481622 | 1766 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 1767 | |
33039678 MK |
1768 | commit = vma_commit_reservation(h, vma, addr); |
1769 | if (unlikely(chg > commit)) { | |
1770 | /* | |
1771 | * The page was added to the reservation map between | |
1772 | * vma_needs_reservation and vma_commit_reservation. | |
1773 | * This indicates a race with hugetlb_reserve_pages. | |
1774 | * Adjust for the subpool count incremented above AND | |
1775 | * in hugetlb_reserve_pages for the same page. Also, | |
1776 | * the reservation count added in hugetlb_reserve_pages | |
1777 | * no longer applies. | |
1778 | */ | |
1779 | long rsv_adjust; | |
1780 | ||
1781 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
1782 | hugetlb_acct_memory(h, -rsv_adjust); | |
1783 | } | |
90d8b7e6 | 1784 | return page; |
8f34af6f JZ |
1785 | |
1786 | out_uncharge_cgroup: | |
1787 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
1788 | out_subpool_put: | |
1789 | if (chg || avoid_reserve) | |
1790 | hugepage_subpool_put_pages(spool, 1); | |
feba16e2 | 1791 | vma_end_reservation(h, vma, addr); |
8f34af6f | 1792 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
1793 | } |
1794 | ||
74060e4d NH |
1795 | /* |
1796 | * alloc_huge_page()'s wrapper which simply returns the page if allocation | |
1797 | * succeeds, otherwise NULL. This function is called from new_vma_page(), | |
1798 | * where no ERR_VALUE is expected to be returned. | |
1799 | */ | |
1800 | struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, | |
1801 | unsigned long addr, int avoid_reserve) | |
1802 | { | |
1803 | struct page *page = alloc_huge_page(vma, addr, avoid_reserve); | |
1804 | if (IS_ERR(page)) | |
1805 | page = NULL; | |
1806 | return page; | |
1807 | } | |
1808 | ||
91f47662 | 1809 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
aa888a74 AK |
1810 | { |
1811 | struct huge_bootmem_page *m; | |
b2261026 | 1812 | int nr_nodes, node; |
aa888a74 | 1813 | |
b2261026 | 1814 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
1815 | void *addr; |
1816 | ||
8b89a116 GS |
1817 | addr = memblock_virt_alloc_try_nid_nopanic( |
1818 | huge_page_size(h), huge_page_size(h), | |
1819 | 0, BOOTMEM_ALLOC_ACCESSIBLE, node); | |
aa888a74 AK |
1820 | if (addr) { |
1821 | /* | |
1822 | * Use the beginning of the huge page to store the | |
1823 | * huge_bootmem_page struct (until gather_bootmem | |
1824 | * puts them into the mem_map). | |
1825 | */ | |
1826 | m = addr; | |
91f47662 | 1827 | goto found; |
aa888a74 | 1828 | } |
aa888a74 AK |
1829 | } |
1830 | return 0; | |
1831 | ||
1832 | found: | |
df994ead | 1833 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 AK |
1834 | /* Put them into a private list first because mem_map is not up yet */ |
1835 | list_add(&m->list, &huge_boot_pages); | |
1836 | m->hstate = h; | |
1837 | return 1; | |
1838 | } | |
1839 | ||
f412c97a | 1840 | static void __init prep_compound_huge_page(struct page *page, int order) |
18229df5 AW |
1841 | { |
1842 | if (unlikely(order > (MAX_ORDER - 1))) | |
1843 | prep_compound_gigantic_page(page, order); | |
1844 | else | |
1845 | prep_compound_page(page, order); | |
1846 | } | |
1847 | ||
aa888a74 AK |
1848 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
1849 | static void __init gather_bootmem_prealloc(void) | |
1850 | { | |
1851 | struct huge_bootmem_page *m; | |
1852 | ||
1853 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 1854 | struct hstate *h = m->hstate; |
ee8f248d BB |
1855 | struct page *page; |
1856 | ||
1857 | #ifdef CONFIG_HIGHMEM | |
1858 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
8b89a116 GS |
1859 | memblock_free_late(__pa(m), |
1860 | sizeof(struct huge_bootmem_page)); | |
ee8f248d BB |
1861 | #else |
1862 | page = virt_to_page(m); | |
1863 | #endif | |
aa888a74 | 1864 | WARN_ON(page_count(page) != 1); |
18229df5 | 1865 | prep_compound_huge_page(page, h->order); |
ef5a22be | 1866 | WARN_ON(PageReserved(page)); |
aa888a74 | 1867 | prep_new_huge_page(h, page, page_to_nid(page)); |
b0320c7b RA |
1868 | /* |
1869 | * If we had gigantic hugepages allocated at boot time, we need | |
1870 | * to restore the 'stolen' pages to totalram_pages in order to | |
1871 | * fix confusing memory reports from free(1) and another | |
1872 | * side-effects, like CommitLimit going negative. | |
1873 | */ | |
bae7f4ae | 1874 | if (hstate_is_gigantic(h)) |
3dcc0571 | 1875 | adjust_managed_page_count(page, 1 << h->order); |
aa888a74 AK |
1876 | } |
1877 | } | |
1878 | ||
8faa8b07 | 1879 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
1880 | { |
1881 | unsigned long i; | |
a5516438 | 1882 | |
e5ff2159 | 1883 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 1884 | if (hstate_is_gigantic(h)) { |
aa888a74 AK |
1885 | if (!alloc_bootmem_huge_page(h)) |
1886 | break; | |
9b5e5d0f | 1887 | } else if (!alloc_fresh_huge_page(h, |
8cebfcd0 | 1888 | &node_states[N_MEMORY])) |
1da177e4 | 1889 | break; |
1da177e4 | 1890 | } |
8faa8b07 | 1891 | h->max_huge_pages = i; |
e5ff2159 AK |
1892 | } |
1893 | ||
1894 | static void __init hugetlb_init_hstates(void) | |
1895 | { | |
1896 | struct hstate *h; | |
1897 | ||
1898 | for_each_hstate(h) { | |
641844f5 NH |
1899 | if (minimum_order > huge_page_order(h)) |
1900 | minimum_order = huge_page_order(h); | |
1901 | ||
8faa8b07 | 1902 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 1903 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 1904 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 1905 | } |
641844f5 | 1906 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
1907 | } |
1908 | ||
4abd32db AK |
1909 | static char * __init memfmt(char *buf, unsigned long n) |
1910 | { | |
1911 | if (n >= (1UL << 30)) | |
1912 | sprintf(buf, "%lu GB", n >> 30); | |
1913 | else if (n >= (1UL << 20)) | |
1914 | sprintf(buf, "%lu MB", n >> 20); | |
1915 | else | |
1916 | sprintf(buf, "%lu KB", n >> 10); | |
1917 | return buf; | |
1918 | } | |
1919 | ||
e5ff2159 AK |
1920 | static void __init report_hugepages(void) |
1921 | { | |
1922 | struct hstate *h; | |
1923 | ||
1924 | for_each_hstate(h) { | |
4abd32db | 1925 | char buf[32]; |
ffb22af5 | 1926 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
4abd32db AK |
1927 | memfmt(buf, huge_page_size(h)), |
1928 | h->free_huge_pages); | |
e5ff2159 AK |
1929 | } |
1930 | } | |
1931 | ||
1da177e4 | 1932 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
1933 | static void try_to_free_low(struct hstate *h, unsigned long count, |
1934 | nodemask_t *nodes_allowed) | |
1da177e4 | 1935 | { |
4415cc8d CL |
1936 | int i; |
1937 | ||
bae7f4ae | 1938 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1939 | return; |
1940 | ||
6ae11b27 | 1941 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 1942 | struct page *page, *next; |
a5516438 AK |
1943 | struct list_head *freel = &h->hugepage_freelists[i]; |
1944 | list_for_each_entry_safe(page, next, freel, lru) { | |
1945 | if (count >= h->nr_huge_pages) | |
6b0c880d | 1946 | return; |
1da177e4 LT |
1947 | if (PageHighMem(page)) |
1948 | continue; | |
1949 | list_del(&page->lru); | |
e5ff2159 | 1950 | update_and_free_page(h, page); |
a5516438 AK |
1951 | h->free_huge_pages--; |
1952 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
1953 | } |
1954 | } | |
1955 | } | |
1956 | #else | |
6ae11b27 LS |
1957 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
1958 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
1959 | { |
1960 | } | |
1961 | #endif | |
1962 | ||
20a0307c WF |
1963 | /* |
1964 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
1965 | * balanced by operating on them in a round-robin fashion. | |
1966 | * Returns 1 if an adjustment was made. | |
1967 | */ | |
6ae11b27 LS |
1968 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
1969 | int delta) | |
20a0307c | 1970 | { |
b2261026 | 1971 | int nr_nodes, node; |
20a0307c WF |
1972 | |
1973 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 1974 | |
b2261026 JK |
1975 | if (delta < 0) { |
1976 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1977 | if (h->surplus_huge_pages_node[node]) | |
1978 | goto found; | |
e8c5c824 | 1979 | } |
b2261026 JK |
1980 | } else { |
1981 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
1982 | if (h->surplus_huge_pages_node[node] < | |
1983 | h->nr_huge_pages_node[node]) | |
1984 | goto found; | |
e8c5c824 | 1985 | } |
b2261026 JK |
1986 | } |
1987 | return 0; | |
20a0307c | 1988 | |
b2261026 JK |
1989 | found: |
1990 | h->surplus_huge_pages += delta; | |
1991 | h->surplus_huge_pages_node[node] += delta; | |
1992 | return 1; | |
20a0307c WF |
1993 | } |
1994 | ||
a5516438 | 1995 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
1996 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
1997 | nodemask_t *nodes_allowed) | |
1da177e4 | 1998 | { |
7893d1d5 | 1999 | unsigned long min_count, ret; |
1da177e4 | 2000 | |
944d9fec | 2001 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
aa888a74 AK |
2002 | return h->max_huge_pages; |
2003 | ||
7893d1d5 AL |
2004 | /* |
2005 | * Increase the pool size | |
2006 | * First take pages out of surplus state. Then make up the | |
2007 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f NA |
2008 | * |
2009 | * We might race with alloc_buddy_huge_page() here and be unable | |
2010 | * to convert a surplus huge page to a normal huge page. That is | |
2011 | * not critical, though, it just means the overall size of the | |
2012 | * pool might be one hugepage larger than it needs to be, but | |
2013 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 2014 | */ |
1da177e4 | 2015 | spin_lock(&hugetlb_lock); |
a5516438 | 2016 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 2017 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
2018 | break; |
2019 | } | |
2020 | ||
a5516438 | 2021 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
2022 | /* |
2023 | * If this allocation races such that we no longer need the | |
2024 | * page, free_huge_page will handle it by freeing the page | |
2025 | * and reducing the surplus. | |
2026 | */ | |
2027 | spin_unlock(&hugetlb_lock); | |
944d9fec LC |
2028 | if (hstate_is_gigantic(h)) |
2029 | ret = alloc_fresh_gigantic_page(h, nodes_allowed); | |
2030 | else | |
2031 | ret = alloc_fresh_huge_page(h, nodes_allowed); | |
7893d1d5 AL |
2032 | spin_lock(&hugetlb_lock); |
2033 | if (!ret) | |
2034 | goto out; | |
2035 | ||
536240f2 MG |
2036 | /* Bail for signals. Probably ctrl-c from user */ |
2037 | if (signal_pending(current)) | |
2038 | goto out; | |
7893d1d5 | 2039 | } |
7893d1d5 AL |
2040 | |
2041 | /* | |
2042 | * Decrease the pool size | |
2043 | * First return free pages to the buddy allocator (being careful | |
2044 | * to keep enough around to satisfy reservations). Then place | |
2045 | * pages into surplus state as needed so the pool will shrink | |
2046 | * to the desired size as pages become free. | |
d1c3fb1f NA |
2047 | * |
2048 | * By placing pages into the surplus state independent of the | |
2049 | * overcommit value, we are allowing the surplus pool size to | |
2050 | * exceed overcommit. There are few sane options here. Since | |
2051 | * alloc_buddy_huge_page() is checking the global counter, | |
2052 | * though, we'll note that we're not allowed to exceed surplus | |
2053 | * and won't grow the pool anywhere else. Not until one of the | |
2054 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 2055 | */ |
a5516438 | 2056 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 2057 | min_count = max(count, min_count); |
6ae11b27 | 2058 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 2059 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 2060 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 2061 | break; |
55f67141 | 2062 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 2063 | } |
a5516438 | 2064 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 2065 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
2066 | break; |
2067 | } | |
2068 | out: | |
a5516438 | 2069 | ret = persistent_huge_pages(h); |
1da177e4 | 2070 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 2071 | return ret; |
1da177e4 LT |
2072 | } |
2073 | ||
a3437870 NA |
2074 | #define HSTATE_ATTR_RO(_name) \ |
2075 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2076 | ||
2077 | #define HSTATE_ATTR(_name) \ | |
2078 | static struct kobj_attribute _name##_attr = \ | |
2079 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2080 | ||
2081 | static struct kobject *hugepages_kobj; | |
2082 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2083 | ||
9a305230 LS |
2084 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
2085 | ||
2086 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
2087 | { |
2088 | int i; | |
9a305230 | 2089 | |
a3437870 | 2090 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
2091 | if (hstate_kobjs[i] == kobj) { |
2092 | if (nidp) | |
2093 | *nidp = NUMA_NO_NODE; | |
a3437870 | 2094 | return &hstates[i]; |
9a305230 LS |
2095 | } |
2096 | ||
2097 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
2098 | } |
2099 | ||
06808b08 | 2100 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
2101 | struct kobj_attribute *attr, char *buf) |
2102 | { | |
9a305230 LS |
2103 | struct hstate *h; |
2104 | unsigned long nr_huge_pages; | |
2105 | int nid; | |
2106 | ||
2107 | h = kobj_to_hstate(kobj, &nid); | |
2108 | if (nid == NUMA_NO_NODE) | |
2109 | nr_huge_pages = h->nr_huge_pages; | |
2110 | else | |
2111 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2112 | ||
2113 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 2114 | } |
adbe8726 | 2115 | |
238d3c13 DR |
2116 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
2117 | struct hstate *h, int nid, | |
2118 | unsigned long count, size_t len) | |
a3437870 NA |
2119 | { |
2120 | int err; | |
bad44b5b | 2121 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 2122 | |
944d9fec | 2123 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) { |
adbe8726 EM |
2124 | err = -EINVAL; |
2125 | goto out; | |
2126 | } | |
2127 | ||
9a305230 LS |
2128 | if (nid == NUMA_NO_NODE) { |
2129 | /* | |
2130 | * global hstate attribute | |
2131 | */ | |
2132 | if (!(obey_mempolicy && | |
2133 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
2134 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 2135 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 LS |
2136 | } |
2137 | } else if (nodes_allowed) { | |
2138 | /* | |
2139 | * per node hstate attribute: adjust count to global, | |
2140 | * but restrict alloc/free to the specified node. | |
2141 | */ | |
2142 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2143 | init_nodemask_of_node(nodes_allowed, nid); | |
2144 | } else | |
8cebfcd0 | 2145 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 | 2146 | |
06808b08 | 2147 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 2148 | |
8cebfcd0 | 2149 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
2150 | NODEMASK_FREE(nodes_allowed); |
2151 | ||
2152 | return len; | |
adbe8726 EM |
2153 | out: |
2154 | NODEMASK_FREE(nodes_allowed); | |
2155 | return err; | |
06808b08 LS |
2156 | } |
2157 | ||
238d3c13 DR |
2158 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
2159 | struct kobject *kobj, const char *buf, | |
2160 | size_t len) | |
2161 | { | |
2162 | struct hstate *h; | |
2163 | unsigned long count; | |
2164 | int nid; | |
2165 | int err; | |
2166 | ||
2167 | err = kstrtoul(buf, 10, &count); | |
2168 | if (err) | |
2169 | return err; | |
2170 | ||
2171 | h = kobj_to_hstate(kobj, &nid); | |
2172 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2173 | } | |
2174 | ||
06808b08 LS |
2175 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
2176 | struct kobj_attribute *attr, char *buf) | |
2177 | { | |
2178 | return nr_hugepages_show_common(kobj, attr, buf); | |
2179 | } | |
2180 | ||
2181 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2182 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2183 | { | |
238d3c13 | 2184 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
2185 | } |
2186 | HSTATE_ATTR(nr_hugepages); | |
2187 | ||
06808b08 LS |
2188 | #ifdef CONFIG_NUMA |
2189 | ||
2190 | /* | |
2191 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2192 | * huge page alloc/free. | |
2193 | */ | |
2194 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
2195 | struct kobj_attribute *attr, char *buf) | |
2196 | { | |
2197 | return nr_hugepages_show_common(kobj, attr, buf); | |
2198 | } | |
2199 | ||
2200 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2201 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2202 | { | |
238d3c13 | 2203 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
2204 | } |
2205 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2206 | #endif | |
2207 | ||
2208 | ||
a3437870 NA |
2209 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
2210 | struct kobj_attribute *attr, char *buf) | |
2211 | { | |
9a305230 | 2212 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2213 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
2214 | } | |
adbe8726 | 2215 | |
a3437870 NA |
2216 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
2217 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2218 | { | |
2219 | int err; | |
2220 | unsigned long input; | |
9a305230 | 2221 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2222 | |
bae7f4ae | 2223 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2224 | return -EINVAL; |
2225 | ||
3dbb95f7 | 2226 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2227 | if (err) |
73ae31e5 | 2228 | return err; |
a3437870 NA |
2229 | |
2230 | spin_lock(&hugetlb_lock); | |
2231 | h->nr_overcommit_huge_pages = input; | |
2232 | spin_unlock(&hugetlb_lock); | |
2233 | ||
2234 | return count; | |
2235 | } | |
2236 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2237 | ||
2238 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2239 | struct kobj_attribute *attr, char *buf) | |
2240 | { | |
9a305230 LS |
2241 | struct hstate *h; |
2242 | unsigned long free_huge_pages; | |
2243 | int nid; | |
2244 | ||
2245 | h = kobj_to_hstate(kobj, &nid); | |
2246 | if (nid == NUMA_NO_NODE) | |
2247 | free_huge_pages = h->free_huge_pages; | |
2248 | else | |
2249 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2250 | ||
2251 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
2252 | } |
2253 | HSTATE_ATTR_RO(free_hugepages); | |
2254 | ||
2255 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2256 | struct kobj_attribute *attr, char *buf) | |
2257 | { | |
9a305230 | 2258 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2259 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
2260 | } | |
2261 | HSTATE_ATTR_RO(resv_hugepages); | |
2262 | ||
2263 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2264 | struct kobj_attribute *attr, char *buf) | |
2265 | { | |
9a305230 LS |
2266 | struct hstate *h; |
2267 | unsigned long surplus_huge_pages; | |
2268 | int nid; | |
2269 | ||
2270 | h = kobj_to_hstate(kobj, &nid); | |
2271 | if (nid == NUMA_NO_NODE) | |
2272 | surplus_huge_pages = h->surplus_huge_pages; | |
2273 | else | |
2274 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2275 | ||
2276 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
2277 | } |
2278 | HSTATE_ATTR_RO(surplus_hugepages); | |
2279 | ||
2280 | static struct attribute *hstate_attrs[] = { | |
2281 | &nr_hugepages_attr.attr, | |
2282 | &nr_overcommit_hugepages_attr.attr, | |
2283 | &free_hugepages_attr.attr, | |
2284 | &resv_hugepages_attr.attr, | |
2285 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
2286 | #ifdef CONFIG_NUMA |
2287 | &nr_hugepages_mempolicy_attr.attr, | |
2288 | #endif | |
a3437870 NA |
2289 | NULL, |
2290 | }; | |
2291 | ||
2292 | static struct attribute_group hstate_attr_group = { | |
2293 | .attrs = hstate_attrs, | |
2294 | }; | |
2295 | ||
094e9539 JM |
2296 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
2297 | struct kobject **hstate_kobjs, | |
2298 | struct attribute_group *hstate_attr_group) | |
a3437870 NA |
2299 | { |
2300 | int retval; | |
972dc4de | 2301 | int hi = hstate_index(h); |
a3437870 | 2302 | |
9a305230 LS |
2303 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
2304 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
2305 | return -ENOMEM; |
2306 | ||
9a305230 | 2307 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 2308 | if (retval) |
9a305230 | 2309 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
2310 | |
2311 | return retval; | |
2312 | } | |
2313 | ||
2314 | static void __init hugetlb_sysfs_init(void) | |
2315 | { | |
2316 | struct hstate *h; | |
2317 | int err; | |
2318 | ||
2319 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
2320 | if (!hugepages_kobj) | |
2321 | return; | |
2322 | ||
2323 | for_each_hstate(h) { | |
9a305230 LS |
2324 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
2325 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 2326 | if (err) |
ffb22af5 | 2327 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
2328 | } |
2329 | } | |
2330 | ||
9a305230 LS |
2331 | #ifdef CONFIG_NUMA |
2332 | ||
2333 | /* | |
2334 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
2335 | * with node devices in node_devices[] using a parallel array. The array |
2336 | * index of a node device or _hstate == node id. | |
2337 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
2338 | * the base kernel, on the hugetlb module. |
2339 | */ | |
2340 | struct node_hstate { | |
2341 | struct kobject *hugepages_kobj; | |
2342 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2343 | }; | |
2344 | struct node_hstate node_hstates[MAX_NUMNODES]; | |
2345 | ||
2346 | /* | |
10fbcf4c | 2347 | * A subset of global hstate attributes for node devices |
9a305230 LS |
2348 | */ |
2349 | static struct attribute *per_node_hstate_attrs[] = { | |
2350 | &nr_hugepages_attr.attr, | |
2351 | &free_hugepages_attr.attr, | |
2352 | &surplus_hugepages_attr.attr, | |
2353 | NULL, | |
2354 | }; | |
2355 | ||
2356 | static struct attribute_group per_node_hstate_attr_group = { | |
2357 | .attrs = per_node_hstate_attrs, | |
2358 | }; | |
2359 | ||
2360 | /* | |
10fbcf4c | 2361 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
2362 | * Returns node id via non-NULL nidp. |
2363 | */ | |
2364 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2365 | { | |
2366 | int nid; | |
2367 | ||
2368 | for (nid = 0; nid < nr_node_ids; nid++) { | |
2369 | struct node_hstate *nhs = &node_hstates[nid]; | |
2370 | int i; | |
2371 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
2372 | if (nhs->hstate_kobjs[i] == kobj) { | |
2373 | if (nidp) | |
2374 | *nidp = nid; | |
2375 | return &hstates[i]; | |
2376 | } | |
2377 | } | |
2378 | ||
2379 | BUG(); | |
2380 | return NULL; | |
2381 | } | |
2382 | ||
2383 | /* | |
10fbcf4c | 2384 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
2385 | * No-op if no hstate attributes attached. |
2386 | */ | |
3cd8b44f | 2387 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
2388 | { |
2389 | struct hstate *h; | |
10fbcf4c | 2390 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2391 | |
2392 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 2393 | return; /* no hstate attributes */ |
9a305230 | 2394 | |
972dc4de AK |
2395 | for_each_hstate(h) { |
2396 | int idx = hstate_index(h); | |
2397 | if (nhs->hstate_kobjs[idx]) { | |
2398 | kobject_put(nhs->hstate_kobjs[idx]); | |
2399 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 2400 | } |
972dc4de | 2401 | } |
9a305230 LS |
2402 | |
2403 | kobject_put(nhs->hugepages_kobj); | |
2404 | nhs->hugepages_kobj = NULL; | |
2405 | } | |
2406 | ||
2407 | /* | |
10fbcf4c | 2408 | * hugetlb module exit: unregister hstate attributes from node devices |
9a305230 LS |
2409 | * that have them. |
2410 | */ | |
2411 | static void hugetlb_unregister_all_nodes(void) | |
2412 | { | |
2413 | int nid; | |
2414 | ||
2415 | /* | |
10fbcf4c | 2416 | * disable node device registrations. |
9a305230 LS |
2417 | */ |
2418 | register_hugetlbfs_with_node(NULL, NULL); | |
2419 | ||
2420 | /* | |
2421 | * remove hstate attributes from any nodes that have them. | |
2422 | */ | |
2423 | for (nid = 0; nid < nr_node_ids; nid++) | |
8732794b | 2424 | hugetlb_unregister_node(node_devices[nid]); |
9a305230 LS |
2425 | } |
2426 | ||
2427 | /* | |
10fbcf4c | 2428 | * Register hstate attributes for a single node device. |
9a305230 LS |
2429 | * No-op if attributes already registered. |
2430 | */ | |
3cd8b44f | 2431 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
2432 | { |
2433 | struct hstate *h; | |
10fbcf4c | 2434 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2435 | int err; |
2436 | ||
2437 | if (nhs->hugepages_kobj) | |
2438 | return; /* already allocated */ | |
2439 | ||
2440 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 2441 | &node->dev.kobj); |
9a305230 LS |
2442 | if (!nhs->hugepages_kobj) |
2443 | return; | |
2444 | ||
2445 | for_each_hstate(h) { | |
2446 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
2447 | nhs->hstate_kobjs, | |
2448 | &per_node_hstate_attr_group); | |
2449 | if (err) { | |
ffb22af5 AM |
2450 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
2451 | h->name, node->dev.id); | |
9a305230 LS |
2452 | hugetlb_unregister_node(node); |
2453 | break; | |
2454 | } | |
2455 | } | |
2456 | } | |
2457 | ||
2458 | /* | |
9b5e5d0f | 2459 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
2460 | * devices of nodes that have memory. All on-line nodes should have |
2461 | * registered their associated device by this time. | |
9a305230 | 2462 | */ |
7d9ca000 | 2463 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
2464 | { |
2465 | int nid; | |
2466 | ||
8cebfcd0 | 2467 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 2468 | struct node *node = node_devices[nid]; |
10fbcf4c | 2469 | if (node->dev.id == nid) |
9a305230 LS |
2470 | hugetlb_register_node(node); |
2471 | } | |
2472 | ||
2473 | /* | |
10fbcf4c | 2474 | * Let the node device driver know we're here so it can |
9a305230 LS |
2475 | * [un]register hstate attributes on node hotplug. |
2476 | */ | |
2477 | register_hugetlbfs_with_node(hugetlb_register_node, | |
2478 | hugetlb_unregister_node); | |
2479 | } | |
2480 | #else /* !CONFIG_NUMA */ | |
2481 | ||
2482 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2483 | { | |
2484 | BUG(); | |
2485 | if (nidp) | |
2486 | *nidp = -1; | |
2487 | return NULL; | |
2488 | } | |
2489 | ||
2490 | static void hugetlb_unregister_all_nodes(void) { } | |
2491 | ||
2492 | static void hugetlb_register_all_nodes(void) { } | |
2493 | ||
2494 | #endif | |
2495 | ||
a3437870 NA |
2496 | static void __exit hugetlb_exit(void) |
2497 | { | |
2498 | struct hstate *h; | |
2499 | ||
9a305230 LS |
2500 | hugetlb_unregister_all_nodes(); |
2501 | ||
a3437870 | 2502 | for_each_hstate(h) { |
972dc4de | 2503 | kobject_put(hstate_kobjs[hstate_index(h)]); |
a3437870 NA |
2504 | } |
2505 | ||
2506 | kobject_put(hugepages_kobj); | |
c672c7f2 | 2507 | kfree(hugetlb_fault_mutex_table); |
a3437870 NA |
2508 | } |
2509 | module_exit(hugetlb_exit); | |
2510 | ||
2511 | static int __init hugetlb_init(void) | |
2512 | { | |
8382d914 DB |
2513 | int i; |
2514 | ||
457c1b27 | 2515 | if (!hugepages_supported()) |
0ef89d25 | 2516 | return 0; |
a3437870 | 2517 | |
e11bfbfc NP |
2518 | if (!size_to_hstate(default_hstate_size)) { |
2519 | default_hstate_size = HPAGE_SIZE; | |
2520 | if (!size_to_hstate(default_hstate_size)) | |
2521 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 2522 | } |
972dc4de | 2523 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
e11bfbfc NP |
2524 | if (default_hstate_max_huge_pages) |
2525 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
a3437870 NA |
2526 | |
2527 | hugetlb_init_hstates(); | |
aa888a74 | 2528 | gather_bootmem_prealloc(); |
a3437870 NA |
2529 | report_hugepages(); |
2530 | ||
2531 | hugetlb_sysfs_init(); | |
9a305230 | 2532 | hugetlb_register_all_nodes(); |
7179e7bf | 2533 | hugetlb_cgroup_file_init(); |
9a305230 | 2534 | |
8382d914 DB |
2535 | #ifdef CONFIG_SMP |
2536 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
2537 | #else | |
2538 | num_fault_mutexes = 1; | |
2539 | #endif | |
c672c7f2 | 2540 | hugetlb_fault_mutex_table = |
8382d914 | 2541 | kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); |
c672c7f2 | 2542 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
2543 | |
2544 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 2545 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
2546 | return 0; |
2547 | } | |
2548 | module_init(hugetlb_init); | |
2549 | ||
2550 | /* Should be called on processing a hugepagesz=... option */ | |
2551 | void __init hugetlb_add_hstate(unsigned order) | |
2552 | { | |
2553 | struct hstate *h; | |
8faa8b07 AK |
2554 | unsigned long i; |
2555 | ||
a3437870 | 2556 | if (size_to_hstate(PAGE_SIZE << order)) { |
ffb22af5 | 2557 | pr_warning("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
2558 | return; |
2559 | } | |
47d38344 | 2560 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 2561 | BUG_ON(order == 0); |
47d38344 | 2562 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
2563 | h->order = order; |
2564 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
2565 | h->nr_huge_pages = 0; |
2566 | h->free_huge_pages = 0; | |
2567 | for (i = 0; i < MAX_NUMNODES; ++i) | |
2568 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 2569 | INIT_LIST_HEAD(&h->hugepage_activelist); |
8cebfcd0 LJ |
2570 | h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); |
2571 | h->next_nid_to_free = first_node(node_states[N_MEMORY]); | |
a3437870 NA |
2572 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
2573 | huge_page_size(h)/1024); | |
8faa8b07 | 2574 | |
a3437870 NA |
2575 | parsed_hstate = h; |
2576 | } | |
2577 | ||
e11bfbfc | 2578 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
2579 | { |
2580 | unsigned long *mhp; | |
8faa8b07 | 2581 | static unsigned long *last_mhp; |
a3437870 NA |
2582 | |
2583 | /* | |
47d38344 | 2584 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
2585 | * so this hugepages= parameter goes to the "default hstate". |
2586 | */ | |
47d38344 | 2587 | if (!hugetlb_max_hstate) |
a3437870 NA |
2588 | mhp = &default_hstate_max_huge_pages; |
2589 | else | |
2590 | mhp = &parsed_hstate->max_huge_pages; | |
2591 | ||
8faa8b07 | 2592 | if (mhp == last_mhp) { |
ffb22af5 AM |
2593 | pr_warning("hugepages= specified twice without " |
2594 | "interleaving hugepagesz=, ignoring\n"); | |
8faa8b07 AK |
2595 | return 1; |
2596 | } | |
2597 | ||
a3437870 NA |
2598 | if (sscanf(s, "%lu", mhp) <= 0) |
2599 | *mhp = 0; | |
2600 | ||
8faa8b07 AK |
2601 | /* |
2602 | * Global state is always initialized later in hugetlb_init. | |
2603 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
2604 | * use the bootmem allocator. | |
2605 | */ | |
47d38344 | 2606 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
2607 | hugetlb_hstate_alloc_pages(parsed_hstate); |
2608 | ||
2609 | last_mhp = mhp; | |
2610 | ||
a3437870 NA |
2611 | return 1; |
2612 | } | |
e11bfbfc NP |
2613 | __setup("hugepages=", hugetlb_nrpages_setup); |
2614 | ||
2615 | static int __init hugetlb_default_setup(char *s) | |
2616 | { | |
2617 | default_hstate_size = memparse(s, &s); | |
2618 | return 1; | |
2619 | } | |
2620 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 2621 | |
8a213460 NA |
2622 | static unsigned int cpuset_mems_nr(unsigned int *array) |
2623 | { | |
2624 | int node; | |
2625 | unsigned int nr = 0; | |
2626 | ||
2627 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
2628 | nr += array[node]; | |
2629 | ||
2630 | return nr; | |
2631 | } | |
2632 | ||
2633 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
2634 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
2635 | struct ctl_table *table, int write, | |
2636 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 2637 | { |
e5ff2159 | 2638 | struct hstate *h = &default_hstate; |
238d3c13 | 2639 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 2640 | int ret; |
e5ff2159 | 2641 | |
457c1b27 NA |
2642 | if (!hugepages_supported()) |
2643 | return -ENOTSUPP; | |
2644 | ||
e5ff2159 AK |
2645 | table->data = &tmp; |
2646 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2647 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2648 | if (ret) | |
2649 | goto out; | |
e5ff2159 | 2650 | |
238d3c13 DR |
2651 | if (write) |
2652 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
2653 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
2654 | out: |
2655 | return ret; | |
1da177e4 | 2656 | } |
396faf03 | 2657 | |
06808b08 LS |
2658 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2659 | void __user *buffer, size_t *length, loff_t *ppos) | |
2660 | { | |
2661 | ||
2662 | return hugetlb_sysctl_handler_common(false, table, write, | |
2663 | buffer, length, ppos); | |
2664 | } | |
2665 | ||
2666 | #ifdef CONFIG_NUMA | |
2667 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2668 | void __user *buffer, size_t *length, loff_t *ppos) | |
2669 | { | |
2670 | return hugetlb_sysctl_handler_common(true, table, write, | |
2671 | buffer, length, ppos); | |
2672 | } | |
2673 | #endif /* CONFIG_NUMA */ | |
2674 | ||
a3d0c6aa | 2675 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 2676 | void __user *buffer, |
a3d0c6aa NA |
2677 | size_t *length, loff_t *ppos) |
2678 | { | |
a5516438 | 2679 | struct hstate *h = &default_hstate; |
e5ff2159 | 2680 | unsigned long tmp; |
08d4a246 | 2681 | int ret; |
e5ff2159 | 2682 | |
457c1b27 NA |
2683 | if (!hugepages_supported()) |
2684 | return -ENOTSUPP; | |
2685 | ||
c033a93c | 2686 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 2687 | |
bae7f4ae | 2688 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
2689 | return -EINVAL; |
2690 | ||
e5ff2159 AK |
2691 | table->data = &tmp; |
2692 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2693 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2694 | if (ret) | |
2695 | goto out; | |
e5ff2159 AK |
2696 | |
2697 | if (write) { | |
2698 | spin_lock(&hugetlb_lock); | |
2699 | h->nr_overcommit_huge_pages = tmp; | |
2700 | spin_unlock(&hugetlb_lock); | |
2701 | } | |
08d4a246 MH |
2702 | out: |
2703 | return ret; | |
a3d0c6aa NA |
2704 | } |
2705 | ||
1da177e4 LT |
2706 | #endif /* CONFIG_SYSCTL */ |
2707 | ||
e1759c21 | 2708 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 2709 | { |
a5516438 | 2710 | struct hstate *h = &default_hstate; |
457c1b27 NA |
2711 | if (!hugepages_supported()) |
2712 | return; | |
e1759c21 | 2713 | seq_printf(m, |
4f98a2fe RR |
2714 | "HugePages_Total: %5lu\n" |
2715 | "HugePages_Free: %5lu\n" | |
2716 | "HugePages_Rsvd: %5lu\n" | |
2717 | "HugePages_Surp: %5lu\n" | |
2718 | "Hugepagesize: %8lu kB\n", | |
a5516438 AK |
2719 | h->nr_huge_pages, |
2720 | h->free_huge_pages, | |
2721 | h->resv_huge_pages, | |
2722 | h->surplus_huge_pages, | |
2723 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
2724 | } |
2725 | ||
2726 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
2727 | { | |
a5516438 | 2728 | struct hstate *h = &default_hstate; |
457c1b27 NA |
2729 | if (!hugepages_supported()) |
2730 | return 0; | |
1da177e4 LT |
2731 | return sprintf(buf, |
2732 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
2733 | "Node %d HugePages_Free: %5u\n" |
2734 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
2735 | nid, h->nr_huge_pages_node[nid], |
2736 | nid, h->free_huge_pages_node[nid], | |
2737 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
2738 | } |
2739 | ||
949f7ec5 DR |
2740 | void hugetlb_show_meminfo(void) |
2741 | { | |
2742 | struct hstate *h; | |
2743 | int nid; | |
2744 | ||
457c1b27 NA |
2745 | if (!hugepages_supported()) |
2746 | return; | |
2747 | ||
949f7ec5 DR |
2748 | for_each_node_state(nid, N_MEMORY) |
2749 | for_each_hstate(h) | |
2750 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
2751 | nid, | |
2752 | h->nr_huge_pages_node[nid], | |
2753 | h->free_huge_pages_node[nid], | |
2754 | h->surplus_huge_pages_node[nid], | |
2755 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
2756 | } | |
2757 | ||
1da177e4 LT |
2758 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
2759 | unsigned long hugetlb_total_pages(void) | |
2760 | { | |
d0028588 WL |
2761 | struct hstate *h; |
2762 | unsigned long nr_total_pages = 0; | |
2763 | ||
2764 | for_each_hstate(h) | |
2765 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
2766 | return nr_total_pages; | |
1da177e4 | 2767 | } |
1da177e4 | 2768 | |
a5516438 | 2769 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
2770 | { |
2771 | int ret = -ENOMEM; | |
2772 | ||
2773 | spin_lock(&hugetlb_lock); | |
2774 | /* | |
2775 | * When cpuset is configured, it breaks the strict hugetlb page | |
2776 | * reservation as the accounting is done on a global variable. Such | |
2777 | * reservation is completely rubbish in the presence of cpuset because | |
2778 | * the reservation is not checked against page availability for the | |
2779 | * current cpuset. Application can still potentially OOM'ed by kernel | |
2780 | * with lack of free htlb page in cpuset that the task is in. | |
2781 | * Attempt to enforce strict accounting with cpuset is almost | |
2782 | * impossible (or too ugly) because cpuset is too fluid that | |
2783 | * task or memory node can be dynamically moved between cpusets. | |
2784 | * | |
2785 | * The change of semantics for shared hugetlb mapping with cpuset is | |
2786 | * undesirable. However, in order to preserve some of the semantics, | |
2787 | * we fall back to check against current free page availability as | |
2788 | * a best attempt and hopefully to minimize the impact of changing | |
2789 | * semantics that cpuset has. | |
2790 | */ | |
2791 | if (delta > 0) { | |
a5516438 | 2792 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
2793 | goto out; |
2794 | ||
a5516438 AK |
2795 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
2796 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
2797 | goto out; |
2798 | } | |
2799 | } | |
2800 | ||
2801 | ret = 0; | |
2802 | if (delta < 0) | |
a5516438 | 2803 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
2804 | |
2805 | out: | |
2806 | spin_unlock(&hugetlb_lock); | |
2807 | return ret; | |
2808 | } | |
2809 | ||
84afd99b AW |
2810 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
2811 | { | |
f522c3ac | 2812 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
2813 | |
2814 | /* | |
2815 | * This new VMA should share its siblings reservation map if present. | |
2816 | * The VMA will only ever have a valid reservation map pointer where | |
2817 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 2818 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
2819 | * after this open call completes. It is therefore safe to take a |
2820 | * new reference here without additional locking. | |
2821 | */ | |
4e35f483 | 2822 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 2823 | kref_get(&resv->refs); |
84afd99b AW |
2824 | } |
2825 | ||
a1e78772 MG |
2826 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
2827 | { | |
a5516438 | 2828 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 2829 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 2830 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 2831 | unsigned long reserve, start, end; |
1c5ecae3 | 2832 | long gbl_reserve; |
84afd99b | 2833 | |
4e35f483 JK |
2834 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
2835 | return; | |
84afd99b | 2836 | |
4e35f483 JK |
2837 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
2838 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 2839 | |
4e35f483 | 2840 | reserve = (end - start) - region_count(resv, start, end); |
84afd99b | 2841 | |
4e35f483 JK |
2842 | kref_put(&resv->refs, resv_map_release); |
2843 | ||
2844 | if (reserve) { | |
1c5ecae3 MK |
2845 | /* |
2846 | * Decrement reserve counts. The global reserve count may be | |
2847 | * adjusted if the subpool has a minimum size. | |
2848 | */ | |
2849 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
2850 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 2851 | } |
a1e78772 MG |
2852 | } |
2853 | ||
1da177e4 LT |
2854 | /* |
2855 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
2856 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
2857 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
2858 | * this far. | |
2859 | */ | |
d0217ac0 | 2860 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
2861 | { |
2862 | BUG(); | |
d0217ac0 | 2863 | return 0; |
1da177e4 LT |
2864 | } |
2865 | ||
f0f37e2f | 2866 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 2867 | .fault = hugetlb_vm_op_fault, |
84afd99b | 2868 | .open = hugetlb_vm_op_open, |
a1e78772 | 2869 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
2870 | }; |
2871 | ||
1e8f889b DG |
2872 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
2873 | int writable) | |
63551ae0 DG |
2874 | { |
2875 | pte_t entry; | |
2876 | ||
1e8f889b | 2877 | if (writable) { |
106c992a GS |
2878 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
2879 | vma->vm_page_prot))); | |
63551ae0 | 2880 | } else { |
106c992a GS |
2881 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
2882 | vma->vm_page_prot)); | |
63551ae0 DG |
2883 | } |
2884 | entry = pte_mkyoung(entry); | |
2885 | entry = pte_mkhuge(entry); | |
d9ed9faa | 2886 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
2887 | |
2888 | return entry; | |
2889 | } | |
2890 | ||
1e8f889b DG |
2891 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
2892 | unsigned long address, pte_t *ptep) | |
2893 | { | |
2894 | pte_t entry; | |
2895 | ||
106c992a | 2896 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 2897 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 2898 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
2899 | } |
2900 | ||
4a705fef NH |
2901 | static int is_hugetlb_entry_migration(pte_t pte) |
2902 | { | |
2903 | swp_entry_t swp; | |
2904 | ||
2905 | if (huge_pte_none(pte) || pte_present(pte)) | |
2906 | return 0; | |
2907 | swp = pte_to_swp_entry(pte); | |
2908 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
2909 | return 1; | |
2910 | else | |
2911 | return 0; | |
2912 | } | |
2913 | ||
2914 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
2915 | { | |
2916 | swp_entry_t swp; | |
2917 | ||
2918 | if (huge_pte_none(pte) || pte_present(pte)) | |
2919 | return 0; | |
2920 | swp = pte_to_swp_entry(pte); | |
2921 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
2922 | return 1; | |
2923 | else | |
2924 | return 0; | |
2925 | } | |
1e8f889b | 2926 | |
63551ae0 DG |
2927 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
2928 | struct vm_area_struct *vma) | |
2929 | { | |
2930 | pte_t *src_pte, *dst_pte, entry; | |
2931 | struct page *ptepage; | |
1c59827d | 2932 | unsigned long addr; |
1e8f889b | 2933 | int cow; |
a5516438 AK |
2934 | struct hstate *h = hstate_vma(vma); |
2935 | unsigned long sz = huge_page_size(h); | |
e8569dd2 AS |
2936 | unsigned long mmun_start; /* For mmu_notifiers */ |
2937 | unsigned long mmun_end; /* For mmu_notifiers */ | |
2938 | int ret = 0; | |
1e8f889b DG |
2939 | |
2940 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 2941 | |
e8569dd2 AS |
2942 | mmun_start = vma->vm_start; |
2943 | mmun_end = vma->vm_end; | |
2944 | if (cow) | |
2945 | mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); | |
2946 | ||
a5516438 | 2947 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 2948 | spinlock_t *src_ptl, *dst_ptl; |
c74df32c HD |
2949 | src_pte = huge_pte_offset(src, addr); |
2950 | if (!src_pte) | |
2951 | continue; | |
a5516438 | 2952 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
2953 | if (!dst_pte) { |
2954 | ret = -ENOMEM; | |
2955 | break; | |
2956 | } | |
c5c99429 LW |
2957 | |
2958 | /* If the pagetables are shared don't copy or take references */ | |
2959 | if (dst_pte == src_pte) | |
2960 | continue; | |
2961 | ||
cb900f41 KS |
2962 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
2963 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
2964 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef NH |
2965 | entry = huge_ptep_get(src_pte); |
2966 | if (huge_pte_none(entry)) { /* skip none entry */ | |
2967 | ; | |
2968 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
2969 | is_hugetlb_entry_hwpoisoned(entry))) { | |
2970 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
2971 | ||
2972 | if (is_write_migration_entry(swp_entry) && cow) { | |
2973 | /* | |
2974 | * COW mappings require pages in both | |
2975 | * parent and child to be set to read. | |
2976 | */ | |
2977 | make_migration_entry_read(&swp_entry); | |
2978 | entry = swp_entry_to_pte(swp_entry); | |
2979 | set_huge_pte_at(src, addr, src_pte, entry); | |
2980 | } | |
2981 | set_huge_pte_at(dst, addr, dst_pte, entry); | |
2982 | } else { | |
34ee645e | 2983 | if (cow) { |
7f2e9525 | 2984 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e JR |
2985 | mmu_notifier_invalidate_range(src, mmun_start, |
2986 | mmun_end); | |
2987 | } | |
0253d634 | 2988 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
2989 | ptepage = pte_page(entry); |
2990 | get_page(ptepage); | |
0fe6e20b | 2991 | page_dup_rmap(ptepage); |
1c59827d HD |
2992 | set_huge_pte_at(dst, addr, dst_pte, entry); |
2993 | } | |
cb900f41 KS |
2994 | spin_unlock(src_ptl); |
2995 | spin_unlock(dst_ptl); | |
63551ae0 | 2996 | } |
63551ae0 | 2997 | |
e8569dd2 AS |
2998 | if (cow) |
2999 | mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); | |
3000 | ||
3001 | return ret; | |
63551ae0 DG |
3002 | } |
3003 | ||
24669e58 AK |
3004 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
3005 | unsigned long start, unsigned long end, | |
3006 | struct page *ref_page) | |
63551ae0 | 3007 | { |
24669e58 | 3008 | int force_flush = 0; |
63551ae0 DG |
3009 | struct mm_struct *mm = vma->vm_mm; |
3010 | unsigned long address; | |
c7546f8f | 3011 | pte_t *ptep; |
63551ae0 | 3012 | pte_t pte; |
cb900f41 | 3013 | spinlock_t *ptl; |
63551ae0 | 3014 | struct page *page; |
a5516438 AK |
3015 | struct hstate *h = hstate_vma(vma); |
3016 | unsigned long sz = huge_page_size(h); | |
2ec74c3e SG |
3017 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
3018 | const unsigned long mmun_end = end; /* For mmu_notifiers */ | |
a5516438 | 3019 | |
63551ae0 | 3020 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
3021 | BUG_ON(start & ~huge_page_mask(h)); |
3022 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 3023 | |
24669e58 | 3024 | tlb_start_vma(tlb, vma); |
2ec74c3e | 3025 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
569f48b8 | 3026 | address = start; |
24669e58 | 3027 | again: |
569f48b8 | 3028 | for (; address < end; address += sz) { |
c7546f8f | 3029 | ptep = huge_pte_offset(mm, address); |
4c887265 | 3030 | if (!ptep) |
c7546f8f DG |
3031 | continue; |
3032 | ||
cb900f41 | 3033 | ptl = huge_pte_lock(h, mm, ptep); |
39dde65c | 3034 | if (huge_pmd_unshare(mm, &address, ptep)) |
cb900f41 | 3035 | goto unlock; |
39dde65c | 3036 | |
6629326b HD |
3037 | pte = huge_ptep_get(ptep); |
3038 | if (huge_pte_none(pte)) | |
cb900f41 | 3039 | goto unlock; |
6629326b HD |
3040 | |
3041 | /* | |
9fbc1f63 NH |
3042 | * Migrating hugepage or HWPoisoned hugepage is already |
3043 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 3044 | */ |
9fbc1f63 | 3045 | if (unlikely(!pte_present(pte))) { |
106c992a | 3046 | huge_pte_clear(mm, address, ptep); |
cb900f41 | 3047 | goto unlock; |
8c4894c6 | 3048 | } |
6629326b HD |
3049 | |
3050 | page = pte_page(pte); | |
04f2cbe3 MG |
3051 | /* |
3052 | * If a reference page is supplied, it is because a specific | |
3053 | * page is being unmapped, not a range. Ensure the page we | |
3054 | * are about to unmap is the actual page of interest. | |
3055 | */ | |
3056 | if (ref_page) { | |
04f2cbe3 | 3057 | if (page != ref_page) |
cb900f41 | 3058 | goto unlock; |
04f2cbe3 MG |
3059 | |
3060 | /* | |
3061 | * Mark the VMA as having unmapped its page so that | |
3062 | * future faults in this VMA will fail rather than | |
3063 | * looking like data was lost | |
3064 | */ | |
3065 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
3066 | } | |
3067 | ||
c7546f8f | 3068 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
24669e58 | 3069 | tlb_remove_tlb_entry(tlb, ptep, address); |
106c992a | 3070 | if (huge_pte_dirty(pte)) |
6649a386 | 3071 | set_page_dirty(page); |
9e81130b | 3072 | |
24669e58 AK |
3073 | page_remove_rmap(page); |
3074 | force_flush = !__tlb_remove_page(tlb, page); | |
cb900f41 | 3075 | if (force_flush) { |
569f48b8 | 3076 | address += sz; |
cb900f41 | 3077 | spin_unlock(ptl); |
24669e58 | 3078 | break; |
cb900f41 | 3079 | } |
9e81130b | 3080 | /* Bail out after unmapping reference page if supplied */ |
cb900f41 KS |
3081 | if (ref_page) { |
3082 | spin_unlock(ptl); | |
9e81130b | 3083 | break; |
cb900f41 KS |
3084 | } |
3085 | unlock: | |
3086 | spin_unlock(ptl); | |
63551ae0 | 3087 | } |
24669e58 AK |
3088 | /* |
3089 | * mmu_gather ran out of room to batch pages, we break out of | |
3090 | * the PTE lock to avoid doing the potential expensive TLB invalidate | |
3091 | * and page-free while holding it. | |
3092 | */ | |
3093 | if (force_flush) { | |
3094 | force_flush = 0; | |
3095 | tlb_flush_mmu(tlb); | |
3096 | if (address < end && !ref_page) | |
3097 | goto again; | |
fe1668ae | 3098 | } |
2ec74c3e | 3099 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
24669e58 | 3100 | tlb_end_vma(tlb, vma); |
1da177e4 | 3101 | } |
63551ae0 | 3102 | |
d833352a MG |
3103 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
3104 | struct vm_area_struct *vma, unsigned long start, | |
3105 | unsigned long end, struct page *ref_page) | |
3106 | { | |
3107 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
3108 | ||
3109 | /* | |
3110 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
3111 | * test will fail on a vma being torn down, and not grab a page table | |
3112 | * on its way out. We're lucky that the flag has such an appropriate | |
3113 | * name, and can in fact be safely cleared here. We could clear it | |
3114 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 3115 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 3116 | * because in the context this is called, the VMA is about to be |
c8c06efa | 3117 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
3118 | */ |
3119 | vma->vm_flags &= ~VM_MAYSHARE; | |
3120 | } | |
3121 | ||
502717f4 | 3122 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 3123 | unsigned long end, struct page *ref_page) |
502717f4 | 3124 | { |
24669e58 AK |
3125 | struct mm_struct *mm; |
3126 | struct mmu_gather tlb; | |
3127 | ||
3128 | mm = vma->vm_mm; | |
3129 | ||
2b047252 | 3130 | tlb_gather_mmu(&tlb, mm, start, end); |
24669e58 AK |
3131 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
3132 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 KC |
3133 | } |
3134 | ||
04f2cbe3 MG |
3135 | /* |
3136 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
3137 | * mappping it owns the reserve page for. The intention is to unmap the page | |
3138 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
3139 | * same region. | |
3140 | */ | |
2f4612af DB |
3141 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
3142 | struct page *page, unsigned long address) | |
04f2cbe3 | 3143 | { |
7526674d | 3144 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
3145 | struct vm_area_struct *iter_vma; |
3146 | struct address_space *mapping; | |
04f2cbe3 MG |
3147 | pgoff_t pgoff; |
3148 | ||
3149 | /* | |
3150 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
3151 | * from page cache lookup which is in HPAGE_SIZE units. | |
3152 | */ | |
7526674d | 3153 | address = address & huge_page_mask(h); |
36e4f20a MH |
3154 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
3155 | vma->vm_pgoff; | |
496ad9aa | 3156 | mapping = file_inode(vma->vm_file)->i_mapping; |
04f2cbe3 | 3157 | |
4eb2b1dc MG |
3158 | /* |
3159 | * Take the mapping lock for the duration of the table walk. As | |
3160 | * this mapping should be shared between all the VMAs, | |
3161 | * __unmap_hugepage_range() is called as the lock is already held | |
3162 | */ | |
83cde9e8 | 3163 | i_mmap_lock_write(mapping); |
6b2dbba8 | 3164 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
3165 | /* Do not unmap the current VMA */ |
3166 | if (iter_vma == vma) | |
3167 | continue; | |
3168 | ||
3169 | /* | |
3170 | * Unmap the page from other VMAs without their own reserves. | |
3171 | * They get marked to be SIGKILLed if they fault in these | |
3172 | * areas. This is because a future no-page fault on this VMA | |
3173 | * could insert a zeroed page instead of the data existing | |
3174 | * from the time of fork. This would look like data corruption | |
3175 | */ | |
3176 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
3177 | unmap_hugepage_range(iter_vma, address, |
3178 | address + huge_page_size(h), page); | |
04f2cbe3 | 3179 | } |
83cde9e8 | 3180 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
3181 | } |
3182 | ||
0fe6e20b NH |
3183 | /* |
3184 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
3185 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
3186 | * cannot race with other handlers or page migration. | |
3187 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 3188 | */ |
1e8f889b | 3189 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
04f2cbe3 | 3190 | unsigned long address, pte_t *ptep, pte_t pte, |
cb900f41 | 3191 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 3192 | { |
a5516438 | 3193 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 3194 | struct page *old_page, *new_page; |
ad4404a2 | 3195 | int ret = 0, outside_reserve = 0; |
2ec74c3e SG |
3196 | unsigned long mmun_start; /* For mmu_notifiers */ |
3197 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1e8f889b DG |
3198 | |
3199 | old_page = pte_page(pte); | |
3200 | ||
04f2cbe3 | 3201 | retry_avoidcopy: |
1e8f889b DG |
3202 | /* If no-one else is actually using this page, avoid the copy |
3203 | * and just make the page writable */ | |
37a2140d JK |
3204 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
3205 | page_move_anon_rmap(old_page, vma, address); | |
1e8f889b | 3206 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 3207 | return 0; |
1e8f889b DG |
3208 | } |
3209 | ||
04f2cbe3 MG |
3210 | /* |
3211 | * If the process that created a MAP_PRIVATE mapping is about to | |
3212 | * perform a COW due to a shared page count, attempt to satisfy | |
3213 | * the allocation without using the existing reserves. The pagecache | |
3214 | * page is used to determine if the reserve at this address was | |
3215 | * consumed or not. If reserves were used, a partial faulted mapping | |
3216 | * at the time of fork() could consume its reserves on COW instead | |
3217 | * of the full address range. | |
3218 | */ | |
5944d011 | 3219 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
3220 | old_page != pagecache_page) |
3221 | outside_reserve = 1; | |
3222 | ||
1e8f889b | 3223 | page_cache_get(old_page); |
b76c8cfb | 3224 | |
ad4404a2 DB |
3225 | /* |
3226 | * Drop page table lock as buddy allocator may be called. It will | |
3227 | * be acquired again before returning to the caller, as expected. | |
3228 | */ | |
cb900f41 | 3229 | spin_unlock(ptl); |
04f2cbe3 | 3230 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 3231 | |
2fc39cec | 3232 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
3233 | /* |
3234 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
3235 | * it is due to references held by a child and an insufficient | |
3236 | * huge page pool. To guarantee the original mappers | |
3237 | * reliability, unmap the page from child processes. The child | |
3238 | * may get SIGKILLed if it later faults. | |
3239 | */ | |
3240 | if (outside_reserve) { | |
ad4404a2 | 3241 | page_cache_release(old_page); |
04f2cbe3 | 3242 | BUG_ON(huge_pte_none(pte)); |
2f4612af DB |
3243 | unmap_ref_private(mm, vma, old_page, address); |
3244 | BUG_ON(huge_pte_none(pte)); | |
3245 | spin_lock(ptl); | |
3246 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); | |
3247 | if (likely(ptep && | |
3248 | pte_same(huge_ptep_get(ptep), pte))) | |
3249 | goto retry_avoidcopy; | |
3250 | /* | |
3251 | * race occurs while re-acquiring page table | |
3252 | * lock, and our job is done. | |
3253 | */ | |
3254 | return 0; | |
04f2cbe3 MG |
3255 | } |
3256 | ||
ad4404a2 DB |
3257 | ret = (PTR_ERR(new_page) == -ENOMEM) ? |
3258 | VM_FAULT_OOM : VM_FAULT_SIGBUS; | |
3259 | goto out_release_old; | |
1e8f889b DG |
3260 | } |
3261 | ||
0fe6e20b NH |
3262 | /* |
3263 | * When the original hugepage is shared one, it does not have | |
3264 | * anon_vma prepared. | |
3265 | */ | |
44e2aa93 | 3266 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
3267 | ret = VM_FAULT_OOM; |
3268 | goto out_release_all; | |
44e2aa93 | 3269 | } |
0fe6e20b | 3270 | |
47ad8475 AA |
3271 | copy_user_huge_page(new_page, old_page, address, vma, |
3272 | pages_per_huge_page(h)); | |
0ed361de | 3273 | __SetPageUptodate(new_page); |
bcc54222 | 3274 | set_page_huge_active(new_page); |
1e8f889b | 3275 | |
2ec74c3e SG |
3276 | mmun_start = address & huge_page_mask(h); |
3277 | mmun_end = mmun_start + huge_page_size(h); | |
3278 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
ad4404a2 | 3279 | |
b76c8cfb | 3280 | /* |
cb900f41 | 3281 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
3282 | * before the page tables are altered |
3283 | */ | |
cb900f41 | 3284 | spin_lock(ptl); |
a5516438 | 3285 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
a9af0c5d | 3286 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
3287 | ClearPagePrivate(new_page); |
3288 | ||
1e8f889b | 3289 | /* Break COW */ |
8fe627ec | 3290 | huge_ptep_clear_flush(vma, address, ptep); |
34ee645e | 3291 | mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); |
1e8f889b DG |
3292 | set_huge_pte_at(mm, address, ptep, |
3293 | make_huge_pte(vma, new_page, 1)); | |
0fe6e20b | 3294 | page_remove_rmap(old_page); |
cd67f0d2 | 3295 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
3296 | /* Make the old page be freed below */ |
3297 | new_page = old_page; | |
3298 | } | |
cb900f41 | 3299 | spin_unlock(ptl); |
2ec74c3e | 3300 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
ad4404a2 | 3301 | out_release_all: |
1e8f889b | 3302 | page_cache_release(new_page); |
ad4404a2 | 3303 | out_release_old: |
1e8f889b | 3304 | page_cache_release(old_page); |
8312034f | 3305 | |
ad4404a2 DB |
3306 | spin_lock(ptl); /* Caller expects lock to be held */ |
3307 | return ret; | |
1e8f889b DG |
3308 | } |
3309 | ||
04f2cbe3 | 3310 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
3311 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
3312 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
3313 | { |
3314 | struct address_space *mapping; | |
e7c4b0bf | 3315 | pgoff_t idx; |
04f2cbe3 MG |
3316 | |
3317 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 3318 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
3319 | |
3320 | return find_lock_page(mapping, idx); | |
3321 | } | |
3322 | ||
3ae77f43 HD |
3323 | /* |
3324 | * Return whether there is a pagecache page to back given address within VMA. | |
3325 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
3326 | */ | |
3327 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
3328 | struct vm_area_struct *vma, unsigned long address) |
3329 | { | |
3330 | struct address_space *mapping; | |
3331 | pgoff_t idx; | |
3332 | struct page *page; | |
3333 | ||
3334 | mapping = vma->vm_file->f_mapping; | |
3335 | idx = vma_hugecache_offset(h, vma, address); | |
3336 | ||
3337 | page = find_get_page(mapping, idx); | |
3338 | if (page) | |
3339 | put_page(page); | |
3340 | return page != NULL; | |
3341 | } | |
3342 | ||
a1ed3dda | 3343 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
8382d914 DB |
3344 | struct address_space *mapping, pgoff_t idx, |
3345 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 3346 | { |
a5516438 | 3347 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 3348 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 3349 | int anon_rmap = 0; |
4c887265 | 3350 | unsigned long size; |
4c887265 | 3351 | struct page *page; |
1e8f889b | 3352 | pte_t new_pte; |
cb900f41 | 3353 | spinlock_t *ptl; |
4c887265 | 3354 | |
04f2cbe3 MG |
3355 | /* |
3356 | * Currently, we are forced to kill the process in the event the | |
3357 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 3358 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
3359 | */ |
3360 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
ffb22af5 AM |
3361 | pr_warning("PID %d killed due to inadequate hugepage pool\n", |
3362 | current->pid); | |
04f2cbe3 MG |
3363 | return ret; |
3364 | } | |
3365 | ||
4c887265 AL |
3366 | /* |
3367 | * Use page lock to guard against racing truncation | |
3368 | * before we get page_table_lock. | |
3369 | */ | |
6bda666a CL |
3370 | retry: |
3371 | page = find_lock_page(mapping, idx); | |
3372 | if (!page) { | |
a5516438 | 3373 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
3374 | if (idx >= size) |
3375 | goto out; | |
04f2cbe3 | 3376 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 3377 | if (IS_ERR(page)) { |
76dcee75 AK |
3378 | ret = PTR_ERR(page); |
3379 | if (ret == -ENOMEM) | |
3380 | ret = VM_FAULT_OOM; | |
3381 | else | |
3382 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
3383 | goto out; |
3384 | } | |
47ad8475 | 3385 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 3386 | __SetPageUptodate(page); |
bcc54222 | 3387 | set_page_huge_active(page); |
ac9b9c66 | 3388 | |
f83a275d | 3389 | if (vma->vm_flags & VM_MAYSHARE) { |
6bda666a | 3390 | int err; |
45c682a6 | 3391 | struct inode *inode = mapping->host; |
6bda666a CL |
3392 | |
3393 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
3394 | if (err) { | |
3395 | put_page(page); | |
6bda666a CL |
3396 | if (err == -EEXIST) |
3397 | goto retry; | |
3398 | goto out; | |
3399 | } | |
07443a85 | 3400 | ClearPagePrivate(page); |
45c682a6 KC |
3401 | |
3402 | spin_lock(&inode->i_lock); | |
a5516438 | 3403 | inode->i_blocks += blocks_per_huge_page(h); |
45c682a6 | 3404 | spin_unlock(&inode->i_lock); |
23be7468 | 3405 | } else { |
6bda666a | 3406 | lock_page(page); |
0fe6e20b NH |
3407 | if (unlikely(anon_vma_prepare(vma))) { |
3408 | ret = VM_FAULT_OOM; | |
3409 | goto backout_unlocked; | |
3410 | } | |
409eb8c2 | 3411 | anon_rmap = 1; |
23be7468 | 3412 | } |
0fe6e20b | 3413 | } else { |
998b4382 NH |
3414 | /* |
3415 | * If memory error occurs between mmap() and fault, some process | |
3416 | * don't have hwpoisoned swap entry for errored virtual address. | |
3417 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
3418 | */ | |
3419 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 3420 | ret = VM_FAULT_HWPOISON | |
972dc4de | 3421 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
3422 | goto backout_unlocked; |
3423 | } | |
6bda666a | 3424 | } |
1e8f889b | 3425 | |
57303d80 AW |
3426 | /* |
3427 | * If we are going to COW a private mapping later, we examine the | |
3428 | * pending reservations for this page now. This will ensure that | |
3429 | * any allocations necessary to record that reservation occur outside | |
3430 | * the spinlock. | |
3431 | */ | |
5e911373 | 3432 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
2b26736c AW |
3433 | if (vma_needs_reservation(h, vma, address) < 0) { |
3434 | ret = VM_FAULT_OOM; | |
3435 | goto backout_unlocked; | |
3436 | } | |
5e911373 | 3437 | /* Just decrements count, does not deallocate */ |
feba16e2 | 3438 | vma_end_reservation(h, vma, address); |
5e911373 | 3439 | } |
57303d80 | 3440 | |
cb900f41 KS |
3441 | ptl = huge_pte_lockptr(h, mm, ptep); |
3442 | spin_lock(ptl); | |
a5516438 | 3443 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
3444 | if (idx >= size) |
3445 | goto backout; | |
3446 | ||
83c54070 | 3447 | ret = 0; |
7f2e9525 | 3448 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
3449 | goto backout; |
3450 | ||
07443a85 JK |
3451 | if (anon_rmap) { |
3452 | ClearPagePrivate(page); | |
409eb8c2 | 3453 | hugepage_add_new_anon_rmap(page, vma, address); |
ac714904 | 3454 | } else |
409eb8c2 | 3455 | page_dup_rmap(page); |
1e8f889b DG |
3456 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
3457 | && (vma->vm_flags & VM_SHARED))); | |
3458 | set_huge_pte_at(mm, address, ptep, new_pte); | |
3459 | ||
788c7df4 | 3460 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 3461 | /* Optimization, do the COW without a second fault */ |
cb900f41 | 3462 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); |
1e8f889b DG |
3463 | } |
3464 | ||
cb900f41 | 3465 | spin_unlock(ptl); |
4c887265 AL |
3466 | unlock_page(page); |
3467 | out: | |
ac9b9c66 | 3468 | return ret; |
4c887265 AL |
3469 | |
3470 | backout: | |
cb900f41 | 3471 | spin_unlock(ptl); |
2b26736c | 3472 | backout_unlocked: |
4c887265 AL |
3473 | unlock_page(page); |
3474 | put_page(page); | |
3475 | goto out; | |
ac9b9c66 HD |
3476 | } |
3477 | ||
8382d914 | 3478 | #ifdef CONFIG_SMP |
c672c7f2 | 3479 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
8382d914 DB |
3480 | struct vm_area_struct *vma, |
3481 | struct address_space *mapping, | |
3482 | pgoff_t idx, unsigned long address) | |
3483 | { | |
3484 | unsigned long key[2]; | |
3485 | u32 hash; | |
3486 | ||
3487 | if (vma->vm_flags & VM_SHARED) { | |
3488 | key[0] = (unsigned long) mapping; | |
3489 | key[1] = idx; | |
3490 | } else { | |
3491 | key[0] = (unsigned long) mm; | |
3492 | key[1] = address >> huge_page_shift(h); | |
3493 | } | |
3494 | ||
3495 | hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); | |
3496 | ||
3497 | return hash & (num_fault_mutexes - 1); | |
3498 | } | |
3499 | #else | |
3500 | /* | |
3501 | * For uniprocesor systems we always use a single mutex, so just | |
3502 | * return 0 and avoid the hashing overhead. | |
3503 | */ | |
c672c7f2 | 3504 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
8382d914 DB |
3505 | struct vm_area_struct *vma, |
3506 | struct address_space *mapping, | |
3507 | pgoff_t idx, unsigned long address) | |
3508 | { | |
3509 | return 0; | |
3510 | } | |
3511 | #endif | |
3512 | ||
86e5216f | 3513 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 3514 | unsigned long address, unsigned int flags) |
86e5216f | 3515 | { |
8382d914 | 3516 | pte_t *ptep, entry; |
cb900f41 | 3517 | spinlock_t *ptl; |
1e8f889b | 3518 | int ret; |
8382d914 DB |
3519 | u32 hash; |
3520 | pgoff_t idx; | |
0fe6e20b | 3521 | struct page *page = NULL; |
57303d80 | 3522 | struct page *pagecache_page = NULL; |
a5516438 | 3523 | struct hstate *h = hstate_vma(vma); |
8382d914 | 3524 | struct address_space *mapping; |
0f792cf9 | 3525 | int need_wait_lock = 0; |
86e5216f | 3526 | |
1e16a539 KH |
3527 | address &= huge_page_mask(h); |
3528 | ||
fd6a03ed NH |
3529 | ptep = huge_pte_offset(mm, address); |
3530 | if (ptep) { | |
3531 | entry = huge_ptep_get(ptep); | |
290408d4 | 3532 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 3533 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
3534 | return 0; |
3535 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 3536 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 3537 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
3538 | } |
3539 | ||
a5516438 | 3540 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); |
86e5216f AL |
3541 | if (!ptep) |
3542 | return VM_FAULT_OOM; | |
3543 | ||
8382d914 DB |
3544 | mapping = vma->vm_file->f_mapping; |
3545 | idx = vma_hugecache_offset(h, vma, address); | |
3546 | ||
3935baa9 DG |
3547 | /* |
3548 | * Serialize hugepage allocation and instantiation, so that we don't | |
3549 | * get spurious allocation failures if two CPUs race to instantiate | |
3550 | * the same page in the page cache. | |
3551 | */ | |
c672c7f2 MK |
3552 | hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address); |
3553 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
8382d914 | 3554 | |
7f2e9525 GS |
3555 | entry = huge_ptep_get(ptep); |
3556 | if (huge_pte_none(entry)) { | |
8382d914 | 3557 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 3558 | goto out_mutex; |
3935baa9 | 3559 | } |
86e5216f | 3560 | |
83c54070 | 3561 | ret = 0; |
1e8f889b | 3562 | |
0f792cf9 NH |
3563 | /* |
3564 | * entry could be a migration/hwpoison entry at this point, so this | |
3565 | * check prevents the kernel from going below assuming that we have | |
3566 | * a active hugepage in pagecache. This goto expects the 2nd page fault, | |
3567 | * and is_hugetlb_entry_(migration|hwpoisoned) check will properly | |
3568 | * handle it. | |
3569 | */ | |
3570 | if (!pte_present(entry)) | |
3571 | goto out_mutex; | |
3572 | ||
57303d80 AW |
3573 | /* |
3574 | * If we are going to COW the mapping later, we examine the pending | |
3575 | * reservations for this page now. This will ensure that any | |
3576 | * allocations necessary to record that reservation occur outside the | |
3577 | * spinlock. For private mappings, we also lookup the pagecache | |
3578 | * page now as it is used to determine if a reservation has been | |
3579 | * consumed. | |
3580 | */ | |
106c992a | 3581 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
2b26736c AW |
3582 | if (vma_needs_reservation(h, vma, address) < 0) { |
3583 | ret = VM_FAULT_OOM; | |
b4d1d99f | 3584 | goto out_mutex; |
2b26736c | 3585 | } |
5e911373 | 3586 | /* Just decrements count, does not deallocate */ |
feba16e2 | 3587 | vma_end_reservation(h, vma, address); |
57303d80 | 3588 | |
f83a275d | 3589 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
3590 | pagecache_page = hugetlbfs_pagecache_page(h, |
3591 | vma, address); | |
3592 | } | |
3593 | ||
0f792cf9 NH |
3594 | ptl = huge_pte_lock(h, mm, ptep); |
3595 | ||
3596 | /* Check for a racing update before calling hugetlb_cow */ | |
3597 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
3598 | goto out_ptl; | |
3599 | ||
56c9cfb1 NH |
3600 | /* |
3601 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
3602 | * pagecache_page, so here we need take the former one | |
3603 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
3604 | */ |
3605 | page = pte_page(entry); | |
3606 | if (page != pagecache_page) | |
0f792cf9 NH |
3607 | if (!trylock_page(page)) { |
3608 | need_wait_lock = 1; | |
3609 | goto out_ptl; | |
3610 | } | |
b4d1d99f | 3611 | |
0f792cf9 | 3612 | get_page(page); |
b4d1d99f | 3613 | |
788c7df4 | 3614 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 3615 | if (!huge_pte_write(entry)) { |
57303d80 | 3616 | ret = hugetlb_cow(mm, vma, address, ptep, entry, |
cb900f41 | 3617 | pagecache_page, ptl); |
0f792cf9 | 3618 | goto out_put_page; |
b4d1d99f | 3619 | } |
106c992a | 3620 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
3621 | } |
3622 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
3623 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
3624 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 3625 | update_mmu_cache(vma, address, ptep); |
0f792cf9 NH |
3626 | out_put_page: |
3627 | if (page != pagecache_page) | |
3628 | unlock_page(page); | |
3629 | put_page(page); | |
cb900f41 KS |
3630 | out_ptl: |
3631 | spin_unlock(ptl); | |
57303d80 AW |
3632 | |
3633 | if (pagecache_page) { | |
3634 | unlock_page(pagecache_page); | |
3635 | put_page(pagecache_page); | |
3636 | } | |
b4d1d99f | 3637 | out_mutex: |
c672c7f2 | 3638 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
0f792cf9 NH |
3639 | /* |
3640 | * Generally it's safe to hold refcount during waiting page lock. But | |
3641 | * here we just wait to defer the next page fault to avoid busy loop and | |
3642 | * the page is not used after unlocked before returning from the current | |
3643 | * page fault. So we are safe from accessing freed page, even if we wait | |
3644 | * here without taking refcount. | |
3645 | */ | |
3646 | if (need_wait_lock) | |
3647 | wait_on_page_locked(page); | |
1e8f889b | 3648 | return ret; |
86e5216f AL |
3649 | } |
3650 | ||
28a35716 ML |
3651 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
3652 | struct page **pages, struct vm_area_struct **vmas, | |
3653 | unsigned long *position, unsigned long *nr_pages, | |
3654 | long i, unsigned int flags) | |
63551ae0 | 3655 | { |
d5d4b0aa KC |
3656 | unsigned long pfn_offset; |
3657 | unsigned long vaddr = *position; | |
28a35716 | 3658 | unsigned long remainder = *nr_pages; |
a5516438 | 3659 | struct hstate *h = hstate_vma(vma); |
63551ae0 | 3660 | |
63551ae0 | 3661 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 3662 | pte_t *pte; |
cb900f41 | 3663 | spinlock_t *ptl = NULL; |
2a15efc9 | 3664 | int absent; |
4c887265 | 3665 | struct page *page; |
63551ae0 | 3666 | |
02057967 DR |
3667 | /* |
3668 | * If we have a pending SIGKILL, don't keep faulting pages and | |
3669 | * potentially allocating memory. | |
3670 | */ | |
3671 | if (unlikely(fatal_signal_pending(current))) { | |
3672 | remainder = 0; | |
3673 | break; | |
3674 | } | |
3675 | ||
4c887265 AL |
3676 | /* |
3677 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 3678 | * each hugepage. We have to make sure we get the |
4c887265 | 3679 | * first, for the page indexing below to work. |
cb900f41 KS |
3680 | * |
3681 | * Note that page table lock is not held when pte is null. | |
4c887265 | 3682 | */ |
a5516438 | 3683 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); |
cb900f41 KS |
3684 | if (pte) |
3685 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
3686 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
3687 | ||
3688 | /* | |
3689 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
3690 | * an error where there's an empty slot with no huge pagecache |
3691 | * to back it. This way, we avoid allocating a hugepage, and | |
3692 | * the sparse dumpfile avoids allocating disk blocks, but its | |
3693 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 3694 | */ |
3ae77f43 HD |
3695 | if (absent && (flags & FOLL_DUMP) && |
3696 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
3697 | if (pte) |
3698 | spin_unlock(ptl); | |
2a15efc9 HD |
3699 | remainder = 0; |
3700 | break; | |
3701 | } | |
63551ae0 | 3702 | |
9cc3a5bd NH |
3703 | /* |
3704 | * We need call hugetlb_fault for both hugepages under migration | |
3705 | * (in which case hugetlb_fault waits for the migration,) and | |
3706 | * hwpoisoned hugepages (in which case we need to prevent the | |
3707 | * caller from accessing to them.) In order to do this, we use | |
3708 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
3709 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
3710 | * both cases, and because we can't follow correct pages | |
3711 | * directly from any kind of swap entries. | |
3712 | */ | |
3713 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
3714 | ((flags & FOLL_WRITE) && |
3715 | !huge_pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 3716 | int ret; |
63551ae0 | 3717 | |
cb900f41 KS |
3718 | if (pte) |
3719 | spin_unlock(ptl); | |
2a15efc9 HD |
3720 | ret = hugetlb_fault(mm, vma, vaddr, |
3721 | (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); | |
a89182c7 | 3722 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 3723 | continue; |
63551ae0 | 3724 | |
4c887265 | 3725 | remainder = 0; |
4c887265 AL |
3726 | break; |
3727 | } | |
3728 | ||
a5516438 | 3729 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 3730 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 3731 | same_page: |
d6692183 | 3732 | if (pages) { |
2a15efc9 | 3733 | pages[i] = mem_map_offset(page, pfn_offset); |
a0368d4e | 3734 | get_page_foll(pages[i]); |
d6692183 | 3735 | } |
63551ae0 DG |
3736 | |
3737 | if (vmas) | |
3738 | vmas[i] = vma; | |
3739 | ||
3740 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 3741 | ++pfn_offset; |
63551ae0 DG |
3742 | --remainder; |
3743 | ++i; | |
d5d4b0aa | 3744 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 3745 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa KC |
3746 | /* |
3747 | * We use pfn_offset to avoid touching the pageframes | |
3748 | * of this compound page. | |
3749 | */ | |
3750 | goto same_page; | |
3751 | } | |
cb900f41 | 3752 | spin_unlock(ptl); |
63551ae0 | 3753 | } |
28a35716 | 3754 | *nr_pages = remainder; |
63551ae0 DG |
3755 | *position = vaddr; |
3756 | ||
2a15efc9 | 3757 | return i ? i : -EFAULT; |
63551ae0 | 3758 | } |
8f860591 | 3759 | |
7da4d641 | 3760 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
3761 | unsigned long address, unsigned long end, pgprot_t newprot) |
3762 | { | |
3763 | struct mm_struct *mm = vma->vm_mm; | |
3764 | unsigned long start = address; | |
3765 | pte_t *ptep; | |
3766 | pte_t pte; | |
a5516438 | 3767 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 3768 | unsigned long pages = 0; |
8f860591 ZY |
3769 | |
3770 | BUG_ON(address >= end); | |
3771 | flush_cache_range(vma, address, end); | |
3772 | ||
a5338093 | 3773 | mmu_notifier_invalidate_range_start(mm, start, end); |
83cde9e8 | 3774 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 3775 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 3776 | spinlock_t *ptl; |
8f860591 ZY |
3777 | ptep = huge_pte_offset(mm, address); |
3778 | if (!ptep) | |
3779 | continue; | |
cb900f41 | 3780 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
3781 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3782 | pages++; | |
cb900f41 | 3783 | spin_unlock(ptl); |
39dde65c | 3784 | continue; |
7da4d641 | 3785 | } |
a8bda28d NH |
3786 | pte = huge_ptep_get(ptep); |
3787 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
3788 | spin_unlock(ptl); | |
3789 | continue; | |
3790 | } | |
3791 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
3792 | swp_entry_t entry = pte_to_swp_entry(pte); | |
3793 | ||
3794 | if (is_write_migration_entry(entry)) { | |
3795 | pte_t newpte; | |
3796 | ||
3797 | make_migration_entry_read(&entry); | |
3798 | newpte = swp_entry_to_pte(entry); | |
3799 | set_huge_pte_at(mm, address, ptep, newpte); | |
3800 | pages++; | |
3801 | } | |
3802 | spin_unlock(ptl); | |
3803 | continue; | |
3804 | } | |
3805 | if (!huge_pte_none(pte)) { | |
8f860591 | 3806 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
106c992a | 3807 | pte = pte_mkhuge(huge_pte_modify(pte, newprot)); |
be7517d6 | 3808 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
8f860591 | 3809 | set_huge_pte_at(mm, address, ptep, pte); |
7da4d641 | 3810 | pages++; |
8f860591 | 3811 | } |
cb900f41 | 3812 | spin_unlock(ptl); |
8f860591 | 3813 | } |
d833352a | 3814 | /* |
c8c06efa | 3815 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 3816 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 3817 | * once we release i_mmap_rwsem, another task can do the final put_page |
d833352a MG |
3818 | * and that page table be reused and filled with junk. |
3819 | */ | |
8f860591 | 3820 | flush_tlb_range(vma, start, end); |
34ee645e | 3821 | mmu_notifier_invalidate_range(mm, start, end); |
83cde9e8 | 3822 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
a5338093 | 3823 | mmu_notifier_invalidate_range_end(mm, start, end); |
7da4d641 PZ |
3824 | |
3825 | return pages << h->order; | |
8f860591 ZY |
3826 | } |
3827 | ||
a1e78772 MG |
3828 | int hugetlb_reserve_pages(struct inode *inode, |
3829 | long from, long to, | |
5a6fe125 | 3830 | struct vm_area_struct *vma, |
ca16d140 | 3831 | vm_flags_t vm_flags) |
e4e574b7 | 3832 | { |
17c9d12e | 3833 | long ret, chg; |
a5516438 | 3834 | struct hstate *h = hstate_inode(inode); |
90481622 | 3835 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 3836 | struct resv_map *resv_map; |
1c5ecae3 | 3837 | long gbl_reserve; |
e4e574b7 | 3838 | |
17c9d12e MG |
3839 | /* |
3840 | * Only apply hugepage reservation if asked. At fault time, an | |
3841 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 3842 | * without using reserves |
17c9d12e | 3843 | */ |
ca16d140 | 3844 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
3845 | return 0; |
3846 | ||
a1e78772 MG |
3847 | /* |
3848 | * Shared mappings base their reservation on the number of pages that | |
3849 | * are already allocated on behalf of the file. Private mappings need | |
3850 | * to reserve the full area even if read-only as mprotect() may be | |
3851 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
3852 | */ | |
9119a41e | 3853 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4e35f483 | 3854 | resv_map = inode_resv_map(inode); |
9119a41e | 3855 | |
1406ec9b | 3856 | chg = region_chg(resv_map, from, to); |
9119a41e JK |
3857 | |
3858 | } else { | |
3859 | resv_map = resv_map_alloc(); | |
17c9d12e MG |
3860 | if (!resv_map) |
3861 | return -ENOMEM; | |
3862 | ||
a1e78772 | 3863 | chg = to - from; |
84afd99b | 3864 | |
17c9d12e MG |
3865 | set_vma_resv_map(vma, resv_map); |
3866 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
3867 | } | |
3868 | ||
c50ac050 DH |
3869 | if (chg < 0) { |
3870 | ret = chg; | |
3871 | goto out_err; | |
3872 | } | |
8a630112 | 3873 | |
1c5ecae3 MK |
3874 | /* |
3875 | * There must be enough pages in the subpool for the mapping. If | |
3876 | * the subpool has a minimum size, there may be some global | |
3877 | * reservations already in place (gbl_reserve). | |
3878 | */ | |
3879 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
3880 | if (gbl_reserve < 0) { | |
c50ac050 DH |
3881 | ret = -ENOSPC; |
3882 | goto out_err; | |
3883 | } | |
5a6fe125 MG |
3884 | |
3885 | /* | |
17c9d12e | 3886 | * Check enough hugepages are available for the reservation. |
90481622 | 3887 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 3888 | */ |
1c5ecae3 | 3889 | ret = hugetlb_acct_memory(h, gbl_reserve); |
68842c9b | 3890 | if (ret < 0) { |
1c5ecae3 MK |
3891 | /* put back original number of pages, chg */ |
3892 | (void)hugepage_subpool_put_pages(spool, chg); | |
c50ac050 | 3893 | goto out_err; |
68842c9b | 3894 | } |
17c9d12e MG |
3895 | |
3896 | /* | |
3897 | * Account for the reservations made. Shared mappings record regions | |
3898 | * that have reservations as they are shared by multiple VMAs. | |
3899 | * When the last VMA disappears, the region map says how much | |
3900 | * the reservation was and the page cache tells how much of | |
3901 | * the reservation was consumed. Private mappings are per-VMA and | |
3902 | * only the consumed reservations are tracked. When the VMA | |
3903 | * disappears, the original reservation is the VMA size and the | |
3904 | * consumed reservations are stored in the map. Hence, nothing | |
3905 | * else has to be done for private mappings here | |
3906 | */ | |
33039678 MK |
3907 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
3908 | long add = region_add(resv_map, from, to); | |
3909 | ||
3910 | if (unlikely(chg > add)) { | |
3911 | /* | |
3912 | * pages in this range were added to the reserve | |
3913 | * map between region_chg and region_add. This | |
3914 | * indicates a race with alloc_huge_page. Adjust | |
3915 | * the subpool and reserve counts modified above | |
3916 | * based on the difference. | |
3917 | */ | |
3918 | long rsv_adjust; | |
3919 | ||
3920 | rsv_adjust = hugepage_subpool_put_pages(spool, | |
3921 | chg - add); | |
3922 | hugetlb_acct_memory(h, -rsv_adjust); | |
3923 | } | |
3924 | } | |
a43a8c39 | 3925 | return 0; |
c50ac050 | 3926 | out_err: |
5e911373 MK |
3927 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
3928 | region_abort(resv_map, from, to); | |
f031dd27 JK |
3929 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3930 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 3931 | return ret; |
a43a8c39 KC |
3932 | } |
3933 | ||
b5cec28d MK |
3934 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
3935 | long freed) | |
a43a8c39 | 3936 | { |
a5516438 | 3937 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 3938 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 3939 | long chg = 0; |
90481622 | 3940 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 3941 | long gbl_reserve; |
45c682a6 | 3942 | |
b5cec28d MK |
3943 | if (resv_map) { |
3944 | chg = region_del(resv_map, start, end); | |
3945 | /* | |
3946 | * region_del() can fail in the rare case where a region | |
3947 | * must be split and another region descriptor can not be | |
3948 | * allocated. If end == LONG_MAX, it will not fail. | |
3949 | */ | |
3950 | if (chg < 0) | |
3951 | return chg; | |
3952 | } | |
3953 | ||
45c682a6 | 3954 | spin_lock(&inode->i_lock); |
e4c6f8be | 3955 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
3956 | spin_unlock(&inode->i_lock); |
3957 | ||
1c5ecae3 MK |
3958 | /* |
3959 | * If the subpool has a minimum size, the number of global | |
3960 | * reservations to be released may be adjusted. | |
3961 | */ | |
3962 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
3963 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
3964 | |
3965 | return 0; | |
a43a8c39 | 3966 | } |
93f70f90 | 3967 | |
3212b535 SC |
3968 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
3969 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
3970 | struct vm_area_struct *vma, | |
3971 | unsigned long addr, pgoff_t idx) | |
3972 | { | |
3973 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
3974 | svma->vm_start; | |
3975 | unsigned long sbase = saddr & PUD_MASK; | |
3976 | unsigned long s_end = sbase + PUD_SIZE; | |
3977 | ||
3978 | /* Allow segments to share if only one is marked locked */ | |
3979 | unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED; | |
3980 | unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED; | |
3981 | ||
3982 | /* | |
3983 | * match the virtual addresses, permission and the alignment of the | |
3984 | * page table page. | |
3985 | */ | |
3986 | if (pmd_index(addr) != pmd_index(saddr) || | |
3987 | vm_flags != svm_flags || | |
3988 | sbase < svma->vm_start || svma->vm_end < s_end) | |
3989 | return 0; | |
3990 | ||
3991 | return saddr; | |
3992 | } | |
3993 | ||
31aafb45 | 3994 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
3995 | { |
3996 | unsigned long base = addr & PUD_MASK; | |
3997 | unsigned long end = base + PUD_SIZE; | |
3998 | ||
3999 | /* | |
4000 | * check on proper vm_flags and page table alignment | |
4001 | */ | |
4002 | if (vma->vm_flags & VM_MAYSHARE && | |
4003 | vma->vm_start <= base && end <= vma->vm_end) | |
31aafb45 NK |
4004 | return true; |
4005 | return false; | |
3212b535 SC |
4006 | } |
4007 | ||
4008 | /* | |
4009 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
4010 | * and returns the corresponding pte. While this is not necessary for the | |
4011 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
4012 | * code much cleaner. pmd allocation is essential for the shared case because | |
c8c06efa | 4013 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise |
3212b535 SC |
4014 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a |
4015 | * bad pmd for sharing. | |
4016 | */ | |
4017 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4018 | { | |
4019 | struct vm_area_struct *vma = find_vma(mm, addr); | |
4020 | struct address_space *mapping = vma->vm_file->f_mapping; | |
4021 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
4022 | vma->vm_pgoff; | |
4023 | struct vm_area_struct *svma; | |
4024 | unsigned long saddr; | |
4025 | pte_t *spte = NULL; | |
4026 | pte_t *pte; | |
cb900f41 | 4027 | spinlock_t *ptl; |
3212b535 SC |
4028 | |
4029 | if (!vma_shareable(vma, addr)) | |
4030 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
4031 | ||
83cde9e8 | 4032 | i_mmap_lock_write(mapping); |
3212b535 SC |
4033 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
4034 | if (svma == vma) | |
4035 | continue; | |
4036 | ||
4037 | saddr = page_table_shareable(svma, vma, addr, idx); | |
4038 | if (saddr) { | |
4039 | spte = huge_pte_offset(svma->vm_mm, saddr); | |
4040 | if (spte) { | |
dc6c9a35 | 4041 | mm_inc_nr_pmds(mm); |
3212b535 SC |
4042 | get_page(virt_to_page(spte)); |
4043 | break; | |
4044 | } | |
4045 | } | |
4046 | } | |
4047 | ||
4048 | if (!spte) | |
4049 | goto out; | |
4050 | ||
cb900f41 KS |
4051 | ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); |
4052 | spin_lock(ptl); | |
dc6c9a35 | 4053 | if (pud_none(*pud)) { |
3212b535 SC |
4054 | pud_populate(mm, pud, |
4055 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
dc6c9a35 | 4056 | } else { |
3212b535 | 4057 | put_page(virt_to_page(spte)); |
dc6c9a35 KS |
4058 | mm_inc_nr_pmds(mm); |
4059 | } | |
cb900f41 | 4060 | spin_unlock(ptl); |
3212b535 SC |
4061 | out: |
4062 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
83cde9e8 | 4063 | i_mmap_unlock_write(mapping); |
3212b535 SC |
4064 | return pte; |
4065 | } | |
4066 | ||
4067 | /* | |
4068 | * unmap huge page backed by shared pte. | |
4069 | * | |
4070 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
4071 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
4072 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
4073 | * | |
cb900f41 | 4074 | * called with page table lock held. |
3212b535 SC |
4075 | * |
4076 | * returns: 1 successfully unmapped a shared pte page | |
4077 | * 0 the underlying pte page is not shared, or it is the last user | |
4078 | */ | |
4079 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4080 | { | |
4081 | pgd_t *pgd = pgd_offset(mm, *addr); | |
4082 | pud_t *pud = pud_offset(pgd, *addr); | |
4083 | ||
4084 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
4085 | if (page_count(virt_to_page(ptep)) == 1) | |
4086 | return 0; | |
4087 | ||
4088 | pud_clear(pud); | |
4089 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 4090 | mm_dec_nr_pmds(mm); |
3212b535 SC |
4091 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
4092 | return 1; | |
4093 | } | |
9e5fc74c SC |
4094 | #define want_pmd_share() (1) |
4095 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
4096 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4097 | { | |
4098 | return NULL; | |
4099 | } | |
e81f2d22 ZZ |
4100 | |
4101 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4102 | { | |
4103 | return 0; | |
4104 | } | |
9e5fc74c | 4105 | #define want_pmd_share() (0) |
3212b535 SC |
4106 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
4107 | ||
9e5fc74c SC |
4108 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
4109 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
4110 | unsigned long addr, unsigned long sz) | |
4111 | { | |
4112 | pgd_t *pgd; | |
4113 | pud_t *pud; | |
4114 | pte_t *pte = NULL; | |
4115 | ||
4116 | pgd = pgd_offset(mm, addr); | |
4117 | pud = pud_alloc(mm, pgd, addr); | |
4118 | if (pud) { | |
4119 | if (sz == PUD_SIZE) { | |
4120 | pte = (pte_t *)pud; | |
4121 | } else { | |
4122 | BUG_ON(sz != PMD_SIZE); | |
4123 | if (want_pmd_share() && pud_none(*pud)) | |
4124 | pte = huge_pmd_share(mm, addr, pud); | |
4125 | else | |
4126 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
4127 | } | |
4128 | } | |
4129 | BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); | |
4130 | ||
4131 | return pte; | |
4132 | } | |
4133 | ||
4134 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) | |
4135 | { | |
4136 | pgd_t *pgd; | |
4137 | pud_t *pud; | |
4138 | pmd_t *pmd = NULL; | |
4139 | ||
4140 | pgd = pgd_offset(mm, addr); | |
4141 | if (pgd_present(*pgd)) { | |
4142 | pud = pud_offset(pgd, addr); | |
4143 | if (pud_present(*pud)) { | |
4144 | if (pud_huge(*pud)) | |
4145 | return (pte_t *)pud; | |
4146 | pmd = pmd_offset(pud, addr); | |
4147 | } | |
4148 | } | |
4149 | return (pte_t *) pmd; | |
4150 | } | |
4151 | ||
61f77eda NH |
4152 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
4153 | ||
4154 | /* | |
4155 | * These functions are overwritable if your architecture needs its own | |
4156 | * behavior. | |
4157 | */ | |
4158 | struct page * __weak | |
4159 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
4160 | int write) | |
4161 | { | |
4162 | return ERR_PTR(-EINVAL); | |
4163 | } | |
4164 | ||
4165 | struct page * __weak | |
9e5fc74c | 4166 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 4167 | pmd_t *pmd, int flags) |
9e5fc74c | 4168 | { |
e66f17ff NH |
4169 | struct page *page = NULL; |
4170 | spinlock_t *ptl; | |
4171 | retry: | |
4172 | ptl = pmd_lockptr(mm, pmd); | |
4173 | spin_lock(ptl); | |
4174 | /* | |
4175 | * make sure that the address range covered by this pmd is not | |
4176 | * unmapped from other threads. | |
4177 | */ | |
4178 | if (!pmd_huge(*pmd)) | |
4179 | goto out; | |
4180 | if (pmd_present(*pmd)) { | |
97534127 | 4181 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
e66f17ff NH |
4182 | if (flags & FOLL_GET) |
4183 | get_page(page); | |
4184 | } else { | |
4185 | if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) { | |
4186 | spin_unlock(ptl); | |
4187 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
4188 | goto retry; | |
4189 | } | |
4190 | /* | |
4191 | * hwpoisoned entry is treated as no_page_table in | |
4192 | * follow_page_mask(). | |
4193 | */ | |
4194 | } | |
4195 | out: | |
4196 | spin_unlock(ptl); | |
9e5fc74c SC |
4197 | return page; |
4198 | } | |
4199 | ||
61f77eda | 4200 | struct page * __weak |
9e5fc74c | 4201 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 4202 | pud_t *pud, int flags) |
9e5fc74c | 4203 | { |
e66f17ff NH |
4204 | if (flags & FOLL_GET) |
4205 | return NULL; | |
9e5fc74c | 4206 | |
e66f17ff | 4207 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
4208 | } |
4209 | ||
d5bd9106 AK |
4210 | #ifdef CONFIG_MEMORY_FAILURE |
4211 | ||
93f70f90 NH |
4212 | /* |
4213 | * This function is called from memory failure code. | |
4214 | * Assume the caller holds page lock of the head page. | |
4215 | */ | |
6de2b1aa | 4216 | int dequeue_hwpoisoned_huge_page(struct page *hpage) |
93f70f90 NH |
4217 | { |
4218 | struct hstate *h = page_hstate(hpage); | |
4219 | int nid = page_to_nid(hpage); | |
6de2b1aa | 4220 | int ret = -EBUSY; |
93f70f90 NH |
4221 | |
4222 | spin_lock(&hugetlb_lock); | |
7e1f049e NH |
4223 | /* |
4224 | * Just checking !page_huge_active is not enough, because that could be | |
4225 | * an isolated/hwpoisoned hugepage (which have >0 refcount). | |
4226 | */ | |
4227 | if (!page_huge_active(hpage) && !page_count(hpage)) { | |
56f2fb14 NH |
4228 | /* |
4229 | * Hwpoisoned hugepage isn't linked to activelist or freelist, | |
4230 | * but dangling hpage->lru can trigger list-debug warnings | |
4231 | * (this happens when we call unpoison_memory() on it), | |
4232 | * so let it point to itself with list_del_init(). | |
4233 | */ | |
4234 | list_del_init(&hpage->lru); | |
8c6c2ecb | 4235 | set_page_refcounted(hpage); |
6de2b1aa NH |
4236 | h->free_huge_pages--; |
4237 | h->free_huge_pages_node[nid]--; | |
4238 | ret = 0; | |
4239 | } | |
93f70f90 | 4240 | spin_unlock(&hugetlb_lock); |
6de2b1aa | 4241 | return ret; |
93f70f90 | 4242 | } |
6de2b1aa | 4243 | #endif |
31caf665 NH |
4244 | |
4245 | bool isolate_huge_page(struct page *page, struct list_head *list) | |
4246 | { | |
bcc54222 NH |
4247 | bool ret = true; |
4248 | ||
309381fe | 4249 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4250 | spin_lock(&hugetlb_lock); |
bcc54222 NH |
4251 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { |
4252 | ret = false; | |
4253 | goto unlock; | |
4254 | } | |
4255 | clear_page_huge_active(page); | |
31caf665 | 4256 | list_move_tail(&page->lru, list); |
bcc54222 | 4257 | unlock: |
31caf665 | 4258 | spin_unlock(&hugetlb_lock); |
bcc54222 | 4259 | return ret; |
31caf665 NH |
4260 | } |
4261 | ||
4262 | void putback_active_hugepage(struct page *page) | |
4263 | { | |
309381fe | 4264 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4265 | spin_lock(&hugetlb_lock); |
bcc54222 | 4266 | set_page_huge_active(page); |
31caf665 NH |
4267 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
4268 | spin_unlock(&hugetlb_lock); | |
4269 | put_page(page); | |
4270 | } |