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 */ | |
1fb1b0e9 MK |
804 | if (vma->vm_flags & VM_MAYSHARE) { |
805 | /* | |
806 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
807 | * be a region map for all pages. The only situation where | |
808 | * there is no region map is if a hole was punched via | |
809 | * fallocate. In this case, there really are no reverves to | |
810 | * use. This situation is indicated if chg != 0. | |
811 | */ | |
812 | if (chg) | |
813 | return false; | |
814 | else | |
815 | return true; | |
816 | } | |
a63884e9 JK |
817 | |
818 | /* | |
819 | * Only the process that called mmap() has reserves for | |
820 | * private mappings. | |
821 | */ | |
7f09ca51 | 822 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
559ec2f8 | 823 | return true; |
a63884e9 | 824 | |
559ec2f8 | 825 | return false; |
a1e78772 MG |
826 | } |
827 | ||
a5516438 | 828 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
829 | { |
830 | int nid = page_to_nid(page); | |
0edaecfa | 831 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
832 | h->free_huge_pages++; |
833 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
834 | } |
835 | ||
bf50bab2 NH |
836 | static struct page *dequeue_huge_page_node(struct hstate *h, int nid) |
837 | { | |
838 | struct page *page; | |
839 | ||
c8721bbb NH |
840 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
841 | if (!is_migrate_isolate_page(page)) | |
842 | break; | |
843 | /* | |
844 | * if 'non-isolated free hugepage' not found on the list, | |
845 | * the allocation fails. | |
846 | */ | |
847 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 848 | return NULL; |
0edaecfa | 849 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 850 | set_page_refcounted(page); |
bf50bab2 NH |
851 | h->free_huge_pages--; |
852 | h->free_huge_pages_node[nid]--; | |
853 | return page; | |
854 | } | |
855 | ||
86cdb465 NH |
856 | /* Movability of hugepages depends on migration support. */ |
857 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
858 | { | |
100873d7 | 859 | if (hugepages_treat_as_movable || hugepage_migration_supported(h)) |
86cdb465 NH |
860 | return GFP_HIGHUSER_MOVABLE; |
861 | else | |
862 | return GFP_HIGHUSER; | |
863 | } | |
864 | ||
a5516438 AK |
865 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
866 | struct vm_area_struct *vma, | |
af0ed73e JK |
867 | unsigned long address, int avoid_reserve, |
868 | long chg) | |
1da177e4 | 869 | { |
b1c12cbc | 870 | struct page *page = NULL; |
480eccf9 | 871 | struct mempolicy *mpol; |
19770b32 | 872 | nodemask_t *nodemask; |
c0ff7453 | 873 | struct zonelist *zonelist; |
dd1a239f MG |
874 | struct zone *zone; |
875 | struct zoneref *z; | |
cc9a6c87 | 876 | unsigned int cpuset_mems_cookie; |
1da177e4 | 877 | |
a1e78772 MG |
878 | /* |
879 | * A child process with MAP_PRIVATE mappings created by their parent | |
880 | * have no page reserves. This check ensures that reservations are | |
881 | * not "stolen". The child may still get SIGKILLed | |
882 | */ | |
af0ed73e | 883 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 884 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 885 | goto err; |
a1e78772 | 886 | |
04f2cbe3 | 887 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 888 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 889 | goto err; |
04f2cbe3 | 890 | |
9966c4bb | 891 | retry_cpuset: |
d26914d1 | 892 | cpuset_mems_cookie = read_mems_allowed_begin(); |
9966c4bb | 893 | zonelist = huge_zonelist(vma, address, |
86cdb465 | 894 | htlb_alloc_mask(h), &mpol, &nodemask); |
9966c4bb | 895 | |
19770b32 MG |
896 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
897 | MAX_NR_ZONES - 1, nodemask) { | |
344736f2 | 898 | if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) { |
bf50bab2 NH |
899 | page = dequeue_huge_page_node(h, zone_to_nid(zone)); |
900 | if (page) { | |
af0ed73e JK |
901 | if (avoid_reserve) |
902 | break; | |
903 | if (!vma_has_reserves(vma, chg)) | |
904 | break; | |
905 | ||
07443a85 | 906 | SetPagePrivate(page); |
af0ed73e | 907 | h->resv_huge_pages--; |
bf50bab2 NH |
908 | break; |
909 | } | |
3abf7afd | 910 | } |
1da177e4 | 911 | } |
cc9a6c87 | 912 | |
52cd3b07 | 913 | mpol_cond_put(mpol); |
d26914d1 | 914 | if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) |
cc9a6c87 | 915 | goto retry_cpuset; |
1da177e4 | 916 | return page; |
cc9a6c87 MG |
917 | |
918 | err: | |
cc9a6c87 | 919 | return NULL; |
1da177e4 LT |
920 | } |
921 | ||
1cac6f2c LC |
922 | /* |
923 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
924 | * We may have allocated or freed a huge page based on a different | |
925 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
926 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
927 | * node for alloc or free. | |
928 | */ | |
929 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
930 | { | |
931 | nid = next_node(nid, *nodes_allowed); | |
932 | if (nid == MAX_NUMNODES) | |
933 | nid = first_node(*nodes_allowed); | |
934 | VM_BUG_ON(nid >= MAX_NUMNODES); | |
935 | ||
936 | return nid; | |
937 | } | |
938 | ||
939 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
940 | { | |
941 | if (!node_isset(nid, *nodes_allowed)) | |
942 | nid = next_node_allowed(nid, nodes_allowed); | |
943 | return nid; | |
944 | } | |
945 | ||
946 | /* | |
947 | * returns the previously saved node ["this node"] from which to | |
948 | * allocate a persistent huge page for the pool and advance the | |
949 | * next node from which to allocate, handling wrap at end of node | |
950 | * mask. | |
951 | */ | |
952 | static int hstate_next_node_to_alloc(struct hstate *h, | |
953 | nodemask_t *nodes_allowed) | |
954 | { | |
955 | int nid; | |
956 | ||
957 | VM_BUG_ON(!nodes_allowed); | |
958 | ||
959 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
960 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
961 | ||
962 | return nid; | |
963 | } | |
964 | ||
965 | /* | |
966 | * helper for free_pool_huge_page() - return the previously saved | |
967 | * node ["this node"] from which to free a huge page. Advance the | |
968 | * next node id whether or not we find a free huge page to free so | |
969 | * that the next attempt to free addresses the next node. | |
970 | */ | |
971 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
972 | { | |
973 | int nid; | |
974 | ||
975 | VM_BUG_ON(!nodes_allowed); | |
976 | ||
977 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
978 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
979 | ||
980 | return nid; | |
981 | } | |
982 | ||
983 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
984 | for (nr_nodes = nodes_weight(*mask); \ | |
985 | nr_nodes > 0 && \ | |
986 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
987 | nr_nodes--) | |
988 | ||
989 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
990 | for (nr_nodes = nodes_weight(*mask); \ | |
991 | nr_nodes > 0 && \ | |
992 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
993 | nr_nodes--) | |
994 | ||
944d9fec LC |
995 | #if defined(CONFIG_CMA) && defined(CONFIG_X86_64) |
996 | static void destroy_compound_gigantic_page(struct page *page, | |
997 | unsigned long order) | |
998 | { | |
999 | int i; | |
1000 | int nr_pages = 1 << order; | |
1001 | struct page *p = page + 1; | |
1002 | ||
1003 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { | |
1004 | __ClearPageTail(p); | |
1005 | set_page_refcounted(p); | |
1006 | p->first_page = NULL; | |
1007 | } | |
1008 | ||
1009 | set_compound_order(page, 0); | |
1010 | __ClearPageHead(page); | |
1011 | } | |
1012 | ||
1013 | static void free_gigantic_page(struct page *page, unsigned order) | |
1014 | { | |
1015 | free_contig_range(page_to_pfn(page), 1 << order); | |
1016 | } | |
1017 | ||
1018 | static int __alloc_gigantic_page(unsigned long start_pfn, | |
1019 | unsigned long nr_pages) | |
1020 | { | |
1021 | unsigned long end_pfn = start_pfn + nr_pages; | |
1022 | return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE); | |
1023 | } | |
1024 | ||
1025 | static bool pfn_range_valid_gigantic(unsigned long start_pfn, | |
1026 | unsigned long nr_pages) | |
1027 | { | |
1028 | unsigned long i, end_pfn = start_pfn + nr_pages; | |
1029 | struct page *page; | |
1030 | ||
1031 | for (i = start_pfn; i < end_pfn; i++) { | |
1032 | if (!pfn_valid(i)) | |
1033 | return false; | |
1034 | ||
1035 | page = pfn_to_page(i); | |
1036 | ||
1037 | if (PageReserved(page)) | |
1038 | return false; | |
1039 | ||
1040 | if (page_count(page) > 0) | |
1041 | return false; | |
1042 | ||
1043 | if (PageHuge(page)) | |
1044 | return false; | |
1045 | } | |
1046 | ||
1047 | return true; | |
1048 | } | |
1049 | ||
1050 | static bool zone_spans_last_pfn(const struct zone *zone, | |
1051 | unsigned long start_pfn, unsigned long nr_pages) | |
1052 | { | |
1053 | unsigned long last_pfn = start_pfn + nr_pages - 1; | |
1054 | return zone_spans_pfn(zone, last_pfn); | |
1055 | } | |
1056 | ||
1057 | static struct page *alloc_gigantic_page(int nid, unsigned order) | |
1058 | { | |
1059 | unsigned long nr_pages = 1 << order; | |
1060 | unsigned long ret, pfn, flags; | |
1061 | struct zone *z; | |
1062 | ||
1063 | z = NODE_DATA(nid)->node_zones; | |
1064 | for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) { | |
1065 | spin_lock_irqsave(&z->lock, flags); | |
1066 | ||
1067 | pfn = ALIGN(z->zone_start_pfn, nr_pages); | |
1068 | while (zone_spans_last_pfn(z, pfn, nr_pages)) { | |
1069 | if (pfn_range_valid_gigantic(pfn, nr_pages)) { | |
1070 | /* | |
1071 | * We release the zone lock here because | |
1072 | * alloc_contig_range() will also lock the zone | |
1073 | * at some point. If there's an allocation | |
1074 | * spinning on this lock, it may win the race | |
1075 | * and cause alloc_contig_range() to fail... | |
1076 | */ | |
1077 | spin_unlock_irqrestore(&z->lock, flags); | |
1078 | ret = __alloc_gigantic_page(pfn, nr_pages); | |
1079 | if (!ret) | |
1080 | return pfn_to_page(pfn); | |
1081 | spin_lock_irqsave(&z->lock, flags); | |
1082 | } | |
1083 | pfn += nr_pages; | |
1084 | } | |
1085 | ||
1086 | spin_unlock_irqrestore(&z->lock, flags); | |
1087 | } | |
1088 | ||
1089 | return NULL; | |
1090 | } | |
1091 | ||
1092 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
1093 | static void prep_compound_gigantic_page(struct page *page, unsigned long order); | |
1094 | ||
1095 | static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) | |
1096 | { | |
1097 | struct page *page; | |
1098 | ||
1099 | page = alloc_gigantic_page(nid, huge_page_order(h)); | |
1100 | if (page) { | |
1101 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
1102 | prep_new_huge_page(h, page, nid); | |
1103 | } | |
1104 | ||
1105 | return page; | |
1106 | } | |
1107 | ||
1108 | static int alloc_fresh_gigantic_page(struct hstate *h, | |
1109 | nodemask_t *nodes_allowed) | |
1110 | { | |
1111 | struct page *page = NULL; | |
1112 | int nr_nodes, node; | |
1113 | ||
1114 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1115 | page = alloc_fresh_gigantic_page_node(h, node); | |
1116 | if (page) | |
1117 | return 1; | |
1118 | } | |
1119 | ||
1120 | return 0; | |
1121 | } | |
1122 | ||
1123 | static inline bool gigantic_page_supported(void) { return true; } | |
1124 | #else | |
1125 | static inline bool gigantic_page_supported(void) { return false; } | |
1126 | static inline void free_gigantic_page(struct page *page, unsigned order) { } | |
1127 | static inline void destroy_compound_gigantic_page(struct page *page, | |
1128 | unsigned long order) { } | |
1129 | static inline int alloc_fresh_gigantic_page(struct hstate *h, | |
1130 | nodemask_t *nodes_allowed) { return 0; } | |
1131 | #endif | |
1132 | ||
a5516438 | 1133 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1134 | { |
1135 | int i; | |
a5516438 | 1136 | |
944d9fec LC |
1137 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
1138 | return; | |
18229df5 | 1139 | |
a5516438 AK |
1140 | h->nr_huge_pages--; |
1141 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
1142 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
1143 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
1144 | 1 << PG_referenced | 1 << PG_dirty | | |
a7407a27 LC |
1145 | 1 << PG_active | 1 << PG_private | |
1146 | 1 << PG_writeback); | |
6af2acb6 | 1147 | } |
309381fe | 1148 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
6af2acb6 AL |
1149 | set_compound_page_dtor(page, NULL); |
1150 | set_page_refcounted(page); | |
944d9fec LC |
1151 | if (hstate_is_gigantic(h)) { |
1152 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
1153 | free_gigantic_page(page, huge_page_order(h)); | |
1154 | } else { | |
944d9fec LC |
1155 | __free_pages(page, huge_page_order(h)); |
1156 | } | |
6af2acb6 AL |
1157 | } |
1158 | ||
e5ff2159 AK |
1159 | struct hstate *size_to_hstate(unsigned long size) |
1160 | { | |
1161 | struct hstate *h; | |
1162 | ||
1163 | for_each_hstate(h) { | |
1164 | if (huge_page_size(h) == size) | |
1165 | return h; | |
1166 | } | |
1167 | return NULL; | |
1168 | } | |
1169 | ||
bcc54222 NH |
1170 | /* |
1171 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
1172 | * to hstate->hugepage_activelist.) | |
1173 | * | |
1174 | * This function can be called for tail pages, but never returns true for them. | |
1175 | */ | |
1176 | bool page_huge_active(struct page *page) | |
1177 | { | |
1178 | VM_BUG_ON_PAGE(!PageHuge(page), page); | |
1179 | return PageHead(page) && PagePrivate(&page[1]); | |
1180 | } | |
1181 | ||
1182 | /* never called for tail page */ | |
1183 | static void set_page_huge_active(struct page *page) | |
1184 | { | |
1185 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1186 | SetPagePrivate(&page[1]); | |
1187 | } | |
1188 | ||
1189 | static void clear_page_huge_active(struct page *page) | |
1190 | { | |
1191 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1192 | ClearPagePrivate(&page[1]); | |
1193 | } | |
1194 | ||
8f1d26d0 | 1195 | void free_huge_page(struct page *page) |
27a85ef1 | 1196 | { |
a5516438 AK |
1197 | /* |
1198 | * Can't pass hstate in here because it is called from the | |
1199 | * compound page destructor. | |
1200 | */ | |
e5ff2159 | 1201 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1202 | int nid = page_to_nid(page); |
90481622 DG |
1203 | struct hugepage_subpool *spool = |
1204 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 1205 | bool restore_reserve; |
27a85ef1 | 1206 | |
e5df70ab | 1207 | set_page_private(page, 0); |
23be7468 | 1208 | page->mapping = NULL; |
7893d1d5 | 1209 | BUG_ON(page_count(page)); |
0fe6e20b | 1210 | BUG_ON(page_mapcount(page)); |
07443a85 | 1211 | restore_reserve = PagePrivate(page); |
16c794b4 | 1212 | ClearPagePrivate(page); |
27a85ef1 | 1213 | |
1c5ecae3 MK |
1214 | /* |
1215 | * A return code of zero implies that the subpool will be under its | |
1216 | * minimum size if the reservation is not restored after page is free. | |
1217 | * Therefore, force restore_reserve operation. | |
1218 | */ | |
1219 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1220 | restore_reserve = true; | |
1221 | ||
27a85ef1 | 1222 | spin_lock(&hugetlb_lock); |
bcc54222 | 1223 | clear_page_huge_active(page); |
6d76dcf4 AK |
1224 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1225 | pages_per_huge_page(h), page); | |
07443a85 JK |
1226 | if (restore_reserve) |
1227 | h->resv_huge_pages++; | |
1228 | ||
944d9fec | 1229 | if (h->surplus_huge_pages_node[nid]) { |
0edaecfa AK |
1230 | /* remove the page from active list */ |
1231 | list_del(&page->lru); | |
a5516438 AK |
1232 | update_and_free_page(h, page); |
1233 | h->surplus_huge_pages--; | |
1234 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 1235 | } else { |
5d3a551c | 1236 | arch_clear_hugepage_flags(page); |
a5516438 | 1237 | enqueue_huge_page(h, page); |
7893d1d5 | 1238 | } |
27a85ef1 DG |
1239 | spin_unlock(&hugetlb_lock); |
1240 | } | |
1241 | ||
a5516438 | 1242 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1243 | { |
0edaecfa | 1244 | INIT_LIST_HEAD(&page->lru); |
b7ba30c6 AK |
1245 | set_compound_page_dtor(page, free_huge_page); |
1246 | spin_lock(&hugetlb_lock); | |
9dd540e2 | 1247 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
1248 | h->nr_huge_pages++; |
1249 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 AK |
1250 | spin_unlock(&hugetlb_lock); |
1251 | put_page(page); /* free it into the hugepage allocator */ | |
1252 | } | |
1253 | ||
2906dd52 | 1254 | static void prep_compound_gigantic_page(struct page *page, unsigned long order) |
20a0307c WF |
1255 | { |
1256 | int i; | |
1257 | int nr_pages = 1 << order; | |
1258 | struct page *p = page + 1; | |
1259 | ||
1260 | /* we rely on prep_new_huge_page to set the destructor */ | |
1261 | set_compound_order(page, order); | |
1262 | __SetPageHead(page); | |
ef5a22be | 1263 | __ClearPageReserved(page); |
20a0307c | 1264 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1265 | /* |
1266 | * For gigantic hugepages allocated through bootmem at | |
1267 | * boot, it's safer to be consistent with the not-gigantic | |
1268 | * hugepages and clear the PG_reserved bit from all tail pages | |
1269 | * too. Otherwse drivers using get_user_pages() to access tail | |
1270 | * pages may get the reference counting wrong if they see | |
1271 | * PG_reserved set on a tail page (despite the head page not | |
1272 | * having PG_reserved set). Enforcing this consistency between | |
1273 | * head and tail pages allows drivers to optimize away a check | |
1274 | * on the head page when they need know if put_page() is needed | |
1275 | * after get_user_pages(). | |
1276 | */ | |
1277 | __ClearPageReserved(p); | |
58a84aa9 | 1278 | set_page_count(p, 0); |
20a0307c | 1279 | p->first_page = page; |
44fc8057 DR |
1280 | /* Make sure p->first_page is always valid for PageTail() */ |
1281 | smp_wmb(); | |
1282 | __SetPageTail(p); | |
20a0307c WF |
1283 | } |
1284 | } | |
1285 | ||
7795912c AM |
1286 | /* |
1287 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1288 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1289 | * details. | |
1290 | */ | |
20a0307c WF |
1291 | int PageHuge(struct page *page) |
1292 | { | |
20a0307c WF |
1293 | if (!PageCompound(page)) |
1294 | return 0; | |
1295 | ||
1296 | page = compound_head(page); | |
758f66a2 | 1297 | return get_compound_page_dtor(page) == free_huge_page; |
20a0307c | 1298 | } |
43131e14 NH |
1299 | EXPORT_SYMBOL_GPL(PageHuge); |
1300 | ||
27c73ae7 AA |
1301 | /* |
1302 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1303 | * normal or transparent huge pages. | |
1304 | */ | |
1305 | int PageHeadHuge(struct page *page_head) | |
1306 | { | |
27c73ae7 AA |
1307 | if (!PageHead(page_head)) |
1308 | return 0; | |
1309 | ||
758f66a2 | 1310 | return get_compound_page_dtor(page_head) == free_huge_page; |
27c73ae7 | 1311 | } |
27c73ae7 | 1312 | |
13d60f4b ZY |
1313 | pgoff_t __basepage_index(struct page *page) |
1314 | { | |
1315 | struct page *page_head = compound_head(page); | |
1316 | pgoff_t index = page_index(page_head); | |
1317 | unsigned long compound_idx; | |
1318 | ||
1319 | if (!PageHuge(page_head)) | |
1320 | return page_index(page); | |
1321 | ||
1322 | if (compound_order(page_head) >= MAX_ORDER) | |
1323 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1324 | else | |
1325 | compound_idx = page - page_head; | |
1326 | ||
1327 | return (index << compound_order(page_head)) + compound_idx; | |
1328 | } | |
1329 | ||
a5516438 | 1330 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
1da177e4 | 1331 | { |
1da177e4 | 1332 | struct page *page; |
f96efd58 | 1333 | |
96db800f | 1334 | page = __alloc_pages_node(nid, |
86cdb465 | 1335 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
551883ae | 1336 | __GFP_REPEAT|__GFP_NOWARN, |
a5516438 | 1337 | huge_page_order(h)); |
1da177e4 | 1338 | if (page) { |
a5516438 | 1339 | prep_new_huge_page(h, page, nid); |
1da177e4 | 1340 | } |
63b4613c NA |
1341 | |
1342 | return page; | |
1343 | } | |
1344 | ||
b2261026 JK |
1345 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
1346 | { | |
1347 | struct page *page; | |
1348 | int nr_nodes, node; | |
1349 | int ret = 0; | |
1350 | ||
1351 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1352 | page = alloc_fresh_huge_page_node(h, node); | |
1353 | if (page) { | |
1354 | ret = 1; | |
1355 | break; | |
1356 | } | |
1357 | } | |
1358 | ||
1359 | if (ret) | |
1360 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
1361 | else | |
1362 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
1363 | ||
1364 | return ret; | |
1365 | } | |
1366 | ||
e8c5c824 LS |
1367 | /* |
1368 | * Free huge page from pool from next node to free. | |
1369 | * Attempt to keep persistent huge pages more or less | |
1370 | * balanced over allowed nodes. | |
1371 | * Called with hugetlb_lock locked. | |
1372 | */ | |
6ae11b27 LS |
1373 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1374 | bool acct_surplus) | |
e8c5c824 | 1375 | { |
b2261026 | 1376 | int nr_nodes, node; |
e8c5c824 LS |
1377 | int ret = 0; |
1378 | ||
b2261026 | 1379 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1380 | /* |
1381 | * If we're returning unused surplus pages, only examine | |
1382 | * nodes with surplus pages. | |
1383 | */ | |
b2261026 JK |
1384 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1385 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1386 | struct page *page = |
b2261026 | 1387 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1388 | struct page, lru); |
1389 | list_del(&page->lru); | |
1390 | h->free_huge_pages--; | |
b2261026 | 1391 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1392 | if (acct_surplus) { |
1393 | h->surplus_huge_pages--; | |
b2261026 | 1394 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1395 | } |
e8c5c824 LS |
1396 | update_and_free_page(h, page); |
1397 | ret = 1; | |
9a76db09 | 1398 | break; |
e8c5c824 | 1399 | } |
b2261026 | 1400 | } |
e8c5c824 LS |
1401 | |
1402 | return ret; | |
1403 | } | |
1404 | ||
c8721bbb NH |
1405 | /* |
1406 | * Dissolve a given free hugepage into free buddy pages. This function does | |
1407 | * nothing for in-use (including surplus) hugepages. | |
1408 | */ | |
1409 | static void dissolve_free_huge_page(struct page *page) | |
1410 | { | |
1411 | spin_lock(&hugetlb_lock); | |
1412 | if (PageHuge(page) && !page_count(page)) { | |
1413 | struct hstate *h = page_hstate(page); | |
1414 | int nid = page_to_nid(page); | |
1415 | list_del(&page->lru); | |
1416 | h->free_huge_pages--; | |
1417 | h->free_huge_pages_node[nid]--; | |
1418 | update_and_free_page(h, page); | |
1419 | } | |
1420 | spin_unlock(&hugetlb_lock); | |
1421 | } | |
1422 | ||
1423 | /* | |
1424 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1425 | * make specified memory blocks removable from the system. | |
1426 | * Note that start_pfn should aligned with (minimum) hugepage size. | |
1427 | */ | |
1428 | void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) | |
1429 | { | |
c8721bbb | 1430 | unsigned long pfn; |
c8721bbb | 1431 | |
d0177639 LZ |
1432 | if (!hugepages_supported()) |
1433 | return; | |
1434 | ||
641844f5 NH |
1435 | VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order)); |
1436 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) | |
c8721bbb NH |
1437 | dissolve_free_huge_page(pfn_to_page(pfn)); |
1438 | } | |
1439 | ||
bf50bab2 | 1440 | static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) |
7893d1d5 AL |
1441 | { |
1442 | struct page *page; | |
bf50bab2 | 1443 | unsigned int r_nid; |
7893d1d5 | 1444 | |
bae7f4ae | 1445 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1446 | return NULL; |
1447 | ||
d1c3fb1f NA |
1448 | /* |
1449 | * Assume we will successfully allocate the surplus page to | |
1450 | * prevent racing processes from causing the surplus to exceed | |
1451 | * overcommit | |
1452 | * | |
1453 | * This however introduces a different race, where a process B | |
1454 | * tries to grow the static hugepage pool while alloc_pages() is | |
1455 | * called by process A. B will only examine the per-node | |
1456 | * counters in determining if surplus huge pages can be | |
1457 | * converted to normal huge pages in adjust_pool_surplus(). A | |
1458 | * won't be able to increment the per-node counter, until the | |
1459 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
1460 | * no more huge pages can be converted from surplus to normal | |
1461 | * state (and doesn't try to convert again). Thus, we have a | |
1462 | * case where a surplus huge page exists, the pool is grown, and | |
1463 | * the surplus huge page still exists after, even though it | |
1464 | * should just have been converted to a normal huge page. This | |
1465 | * does not leak memory, though, as the hugepage will be freed | |
1466 | * once it is out of use. It also does not allow the counters to | |
1467 | * go out of whack in adjust_pool_surplus() as we don't modify | |
1468 | * the node values until we've gotten the hugepage and only the | |
1469 | * per-node value is checked there. | |
1470 | */ | |
1471 | spin_lock(&hugetlb_lock); | |
a5516438 | 1472 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
1473 | spin_unlock(&hugetlb_lock); |
1474 | return NULL; | |
1475 | } else { | |
a5516438 AK |
1476 | h->nr_huge_pages++; |
1477 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
1478 | } |
1479 | spin_unlock(&hugetlb_lock); | |
1480 | ||
bf50bab2 | 1481 | if (nid == NUMA_NO_NODE) |
86cdb465 | 1482 | page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP| |
bf50bab2 NH |
1483 | __GFP_REPEAT|__GFP_NOWARN, |
1484 | huge_page_order(h)); | |
1485 | else | |
96db800f | 1486 | page = __alloc_pages_node(nid, |
86cdb465 | 1487 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
bf50bab2 | 1488 | __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); |
d1c3fb1f NA |
1489 | |
1490 | spin_lock(&hugetlb_lock); | |
7893d1d5 | 1491 | if (page) { |
0edaecfa | 1492 | INIT_LIST_HEAD(&page->lru); |
bf50bab2 | 1493 | r_nid = page_to_nid(page); |
7893d1d5 | 1494 | set_compound_page_dtor(page, free_huge_page); |
9dd540e2 | 1495 | set_hugetlb_cgroup(page, NULL); |
d1c3fb1f NA |
1496 | /* |
1497 | * We incremented the global counters already | |
1498 | */ | |
bf50bab2 NH |
1499 | h->nr_huge_pages_node[r_nid]++; |
1500 | h->surplus_huge_pages_node[r_nid]++; | |
3b116300 | 1501 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 1502 | } else { |
a5516438 AK |
1503 | h->nr_huge_pages--; |
1504 | h->surplus_huge_pages--; | |
3b116300 | 1505 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 1506 | } |
d1c3fb1f | 1507 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1508 | |
1509 | return page; | |
1510 | } | |
1511 | ||
bf50bab2 NH |
1512 | /* |
1513 | * This allocation function is useful in the context where vma is irrelevant. | |
1514 | * E.g. soft-offlining uses this function because it only cares physical | |
1515 | * address of error page. | |
1516 | */ | |
1517 | struct page *alloc_huge_page_node(struct hstate *h, int nid) | |
1518 | { | |
4ef91848 | 1519 | struct page *page = NULL; |
bf50bab2 NH |
1520 | |
1521 | spin_lock(&hugetlb_lock); | |
4ef91848 JK |
1522 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
1523 | page = dequeue_huge_page_node(h, nid); | |
bf50bab2 NH |
1524 | spin_unlock(&hugetlb_lock); |
1525 | ||
94ae8ba7 | 1526 | if (!page) |
bf50bab2 NH |
1527 | page = alloc_buddy_huge_page(h, nid); |
1528 | ||
1529 | return page; | |
1530 | } | |
1531 | ||
e4e574b7 | 1532 | /* |
25985edc | 1533 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1534 | * of size 'delta'. |
1535 | */ | |
a5516438 | 1536 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1537 | { |
1538 | struct list_head surplus_list; | |
1539 | struct page *page, *tmp; | |
1540 | int ret, i; | |
1541 | int needed, allocated; | |
28073b02 | 1542 | bool alloc_ok = true; |
e4e574b7 | 1543 | |
a5516438 | 1544 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1545 | if (needed <= 0) { |
a5516438 | 1546 | h->resv_huge_pages += delta; |
e4e574b7 | 1547 | return 0; |
ac09b3a1 | 1548 | } |
e4e574b7 AL |
1549 | |
1550 | allocated = 0; | |
1551 | INIT_LIST_HEAD(&surplus_list); | |
1552 | ||
1553 | ret = -ENOMEM; | |
1554 | retry: | |
1555 | spin_unlock(&hugetlb_lock); | |
1556 | for (i = 0; i < needed; i++) { | |
bf50bab2 | 1557 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
28073b02 HD |
1558 | if (!page) { |
1559 | alloc_ok = false; | |
1560 | break; | |
1561 | } | |
e4e574b7 AL |
1562 | list_add(&page->lru, &surplus_list); |
1563 | } | |
28073b02 | 1564 | allocated += i; |
e4e574b7 AL |
1565 | |
1566 | /* | |
1567 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1568 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1569 | */ | |
1570 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1571 | needed = (h->resv_huge_pages + delta) - |
1572 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1573 | if (needed > 0) { |
1574 | if (alloc_ok) | |
1575 | goto retry; | |
1576 | /* | |
1577 | * We were not able to allocate enough pages to | |
1578 | * satisfy the entire reservation so we free what | |
1579 | * we've allocated so far. | |
1580 | */ | |
1581 | goto free; | |
1582 | } | |
e4e574b7 AL |
1583 | /* |
1584 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1585 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1586 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1587 | * allocator. Commit the entire reservation here to prevent another |
1588 | * process from stealing the pages as they are added to the pool but | |
1589 | * before they are reserved. | |
e4e574b7 AL |
1590 | */ |
1591 | needed += allocated; | |
a5516438 | 1592 | h->resv_huge_pages += delta; |
e4e574b7 | 1593 | ret = 0; |
a9869b83 | 1594 | |
19fc3f0a | 1595 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1596 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1597 | if ((--needed) < 0) |
1598 | break; | |
a9869b83 NH |
1599 | /* |
1600 | * This page is now managed by the hugetlb allocator and has | |
1601 | * no users -- drop the buddy allocator's reference. | |
1602 | */ | |
1603 | put_page_testzero(page); | |
309381fe | 1604 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 1605 | enqueue_huge_page(h, page); |
19fc3f0a | 1606 | } |
28073b02 | 1607 | free: |
b0365c8d | 1608 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1609 | |
1610 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1611 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1612 | put_page(page); | |
a9869b83 | 1613 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1614 | |
1615 | return ret; | |
1616 | } | |
1617 | ||
1618 | /* | |
1619 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
1620 | * allocated to satisfy the reservation must be explicitly freed if they were | |
1621 | * never used. | |
685f3457 | 1622 | * Called with hugetlb_lock held. |
e4e574b7 | 1623 | */ |
a5516438 AK |
1624 | static void return_unused_surplus_pages(struct hstate *h, |
1625 | unsigned long unused_resv_pages) | |
e4e574b7 | 1626 | { |
e4e574b7 AL |
1627 | unsigned long nr_pages; |
1628 | ||
ac09b3a1 | 1629 | /* Uncommit the reservation */ |
a5516438 | 1630 | h->resv_huge_pages -= unused_resv_pages; |
ac09b3a1 | 1631 | |
aa888a74 | 1632 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 1633 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1634 | return; |
1635 | ||
a5516438 | 1636 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1637 | |
685f3457 LS |
1638 | /* |
1639 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1640 | * evenly across all nodes with memory. Iterate across these nodes |
1641 | * until we can no longer free unreserved surplus pages. This occurs | |
1642 | * when the nodes with surplus pages have no free pages. | |
1643 | * free_pool_huge_page() will balance the the freed pages across the | |
1644 | * on-line nodes with memory and will handle the hstate accounting. | |
685f3457 LS |
1645 | */ |
1646 | while (nr_pages--) { | |
8cebfcd0 | 1647 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
685f3457 | 1648 | break; |
7848a4bf | 1649 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 AL |
1650 | } |
1651 | } | |
1652 | ||
5e911373 | 1653 | |
c37f9fb1 | 1654 | /* |
feba16e2 | 1655 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 1656 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
1657 | * |
1658 | * vma_needs_reservation is called to determine if the huge page at addr | |
1659 | * within the vma has an associated reservation. If a reservation is | |
1660 | * needed, the value 1 is returned. The caller is then responsible for | |
1661 | * managing the global reservation and subpool usage counts. After | |
1662 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
1663 | * to add the page to the reservation map. If the page allocation fails, |
1664 | * the reservation must be ended instead of committed. vma_end_reservation | |
1665 | * is called in such cases. | |
cf3ad20b MK |
1666 | * |
1667 | * In the normal case, vma_commit_reservation returns the same value | |
1668 | * as the preceding vma_needs_reservation call. The only time this | |
1669 | * is not the case is if a reserve map was changed between calls. It | |
1670 | * is the responsibility of the caller to notice the difference and | |
1671 | * take appropriate action. | |
c37f9fb1 | 1672 | */ |
5e911373 MK |
1673 | enum vma_resv_mode { |
1674 | VMA_NEEDS_RESV, | |
1675 | VMA_COMMIT_RESV, | |
feba16e2 | 1676 | VMA_END_RESV, |
5e911373 | 1677 | }; |
cf3ad20b MK |
1678 | static long __vma_reservation_common(struct hstate *h, |
1679 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 1680 | enum vma_resv_mode mode) |
c37f9fb1 | 1681 | { |
4e35f483 JK |
1682 | struct resv_map *resv; |
1683 | pgoff_t idx; | |
cf3ad20b | 1684 | long ret; |
c37f9fb1 | 1685 | |
4e35f483 JK |
1686 | resv = vma_resv_map(vma); |
1687 | if (!resv) | |
84afd99b | 1688 | return 1; |
c37f9fb1 | 1689 | |
4e35f483 | 1690 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
1691 | switch (mode) { |
1692 | case VMA_NEEDS_RESV: | |
cf3ad20b | 1693 | ret = region_chg(resv, idx, idx + 1); |
5e911373 MK |
1694 | break; |
1695 | case VMA_COMMIT_RESV: | |
1696 | ret = region_add(resv, idx, idx + 1); | |
1697 | break; | |
feba16e2 | 1698 | case VMA_END_RESV: |
5e911373 MK |
1699 | region_abort(resv, idx, idx + 1); |
1700 | ret = 0; | |
1701 | break; | |
1702 | default: | |
1703 | BUG(); | |
1704 | } | |
84afd99b | 1705 | |
4e35f483 | 1706 | if (vma->vm_flags & VM_MAYSHARE) |
cf3ad20b | 1707 | return ret; |
4e35f483 | 1708 | else |
cf3ad20b | 1709 | return ret < 0 ? ret : 0; |
c37f9fb1 | 1710 | } |
cf3ad20b MK |
1711 | |
1712 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 1713 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 1714 | { |
5e911373 | 1715 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 1716 | } |
84afd99b | 1717 | |
cf3ad20b MK |
1718 | static long vma_commit_reservation(struct hstate *h, |
1719 | struct vm_area_struct *vma, unsigned long addr) | |
1720 | { | |
5e911373 MK |
1721 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
1722 | } | |
1723 | ||
feba16e2 | 1724 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
1725 | struct vm_area_struct *vma, unsigned long addr) |
1726 | { | |
feba16e2 | 1727 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
1728 | } |
1729 | ||
70c3547e | 1730 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 1731 | unsigned long addr, int avoid_reserve) |
1da177e4 | 1732 | { |
90481622 | 1733 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 1734 | struct hstate *h = hstate_vma(vma); |
348ea204 | 1735 | struct page *page; |
d85f69b0 MK |
1736 | long map_chg, map_commit; |
1737 | long gbl_chg; | |
6d76dcf4 AK |
1738 | int ret, idx; |
1739 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 1740 | |
6d76dcf4 | 1741 | idx = hstate_index(h); |
a1e78772 | 1742 | /* |
d85f69b0 MK |
1743 | * Examine the region/reserve map to determine if the process |
1744 | * has a reservation for the page to be allocated. A return | |
1745 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 1746 | */ |
d85f69b0 MK |
1747 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
1748 | if (map_chg < 0) | |
76dcee75 | 1749 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
1750 | |
1751 | /* | |
1752 | * Processes that did not create the mapping will have no | |
1753 | * reserves as indicated by the region/reserve map. Check | |
1754 | * that the allocation will not exceed the subpool limit. | |
1755 | * Allocations for MAP_NORESERVE mappings also need to be | |
1756 | * checked against any subpool limit. | |
1757 | */ | |
1758 | if (map_chg || avoid_reserve) { | |
1759 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
1760 | if (gbl_chg < 0) { | |
feba16e2 | 1761 | vma_end_reservation(h, vma, addr); |
76dcee75 | 1762 | return ERR_PTR(-ENOSPC); |
5e911373 | 1763 | } |
1da177e4 | 1764 | |
d85f69b0 MK |
1765 | /* |
1766 | * Even though there was no reservation in the region/reserve | |
1767 | * map, there could be reservations associated with the | |
1768 | * subpool that can be used. This would be indicated if the | |
1769 | * return value of hugepage_subpool_get_pages() is zero. | |
1770 | * However, if avoid_reserve is specified we still avoid even | |
1771 | * the subpool reservations. | |
1772 | */ | |
1773 | if (avoid_reserve) | |
1774 | gbl_chg = 1; | |
1775 | } | |
1776 | ||
6d76dcf4 | 1777 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f JZ |
1778 | if (ret) |
1779 | goto out_subpool_put; | |
1780 | ||
1da177e4 | 1781 | spin_lock(&hugetlb_lock); |
d85f69b0 MK |
1782 | /* |
1783 | * glb_chg is passed to indicate whether or not a page must be taken | |
1784 | * from the global free pool (global change). gbl_chg == 0 indicates | |
1785 | * a reservation exists for the allocation. | |
1786 | */ | |
1787 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 1788 | if (!page) { |
94ae8ba7 | 1789 | spin_unlock(&hugetlb_lock); |
bf50bab2 | 1790 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
8f34af6f JZ |
1791 | if (!page) |
1792 | goto out_uncharge_cgroup; | |
1793 | ||
79dbb236 AK |
1794 | spin_lock(&hugetlb_lock); |
1795 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 1796 | /* Fall through */ |
68842c9b | 1797 | } |
81a6fcae JK |
1798 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
1799 | spin_unlock(&hugetlb_lock); | |
348ea204 | 1800 | |
90481622 | 1801 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 1802 | |
d85f69b0 MK |
1803 | map_commit = vma_commit_reservation(h, vma, addr); |
1804 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
1805 | /* |
1806 | * The page was added to the reservation map between | |
1807 | * vma_needs_reservation and vma_commit_reservation. | |
1808 | * This indicates a race with hugetlb_reserve_pages. | |
1809 | * Adjust for the subpool count incremented above AND | |
1810 | * in hugetlb_reserve_pages for the same page. Also, | |
1811 | * the reservation count added in hugetlb_reserve_pages | |
1812 | * no longer applies. | |
1813 | */ | |
1814 | long rsv_adjust; | |
1815 | ||
1816 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
1817 | hugetlb_acct_memory(h, -rsv_adjust); | |
1818 | } | |
90d8b7e6 | 1819 | return page; |
8f34af6f JZ |
1820 | |
1821 | out_uncharge_cgroup: | |
1822 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
1823 | out_subpool_put: | |
d85f69b0 | 1824 | if (map_chg || avoid_reserve) |
8f34af6f | 1825 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 1826 | vma_end_reservation(h, vma, addr); |
8f34af6f | 1827 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
1828 | } |
1829 | ||
74060e4d NH |
1830 | /* |
1831 | * alloc_huge_page()'s wrapper which simply returns the page if allocation | |
1832 | * succeeds, otherwise NULL. This function is called from new_vma_page(), | |
1833 | * where no ERR_VALUE is expected to be returned. | |
1834 | */ | |
1835 | struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, | |
1836 | unsigned long addr, int avoid_reserve) | |
1837 | { | |
1838 | struct page *page = alloc_huge_page(vma, addr, avoid_reserve); | |
1839 | if (IS_ERR(page)) | |
1840 | page = NULL; | |
1841 | return page; | |
1842 | } | |
1843 | ||
91f47662 | 1844 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
aa888a74 AK |
1845 | { |
1846 | struct huge_bootmem_page *m; | |
b2261026 | 1847 | int nr_nodes, node; |
aa888a74 | 1848 | |
b2261026 | 1849 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
1850 | void *addr; |
1851 | ||
8b89a116 GS |
1852 | addr = memblock_virt_alloc_try_nid_nopanic( |
1853 | huge_page_size(h), huge_page_size(h), | |
1854 | 0, BOOTMEM_ALLOC_ACCESSIBLE, node); | |
aa888a74 AK |
1855 | if (addr) { |
1856 | /* | |
1857 | * Use the beginning of the huge page to store the | |
1858 | * huge_bootmem_page struct (until gather_bootmem | |
1859 | * puts them into the mem_map). | |
1860 | */ | |
1861 | m = addr; | |
91f47662 | 1862 | goto found; |
aa888a74 | 1863 | } |
aa888a74 AK |
1864 | } |
1865 | return 0; | |
1866 | ||
1867 | found: | |
df994ead | 1868 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 AK |
1869 | /* Put them into a private list first because mem_map is not up yet */ |
1870 | list_add(&m->list, &huge_boot_pages); | |
1871 | m->hstate = h; | |
1872 | return 1; | |
1873 | } | |
1874 | ||
f412c97a | 1875 | static void __init prep_compound_huge_page(struct page *page, int order) |
18229df5 AW |
1876 | { |
1877 | if (unlikely(order > (MAX_ORDER - 1))) | |
1878 | prep_compound_gigantic_page(page, order); | |
1879 | else | |
1880 | prep_compound_page(page, order); | |
1881 | } | |
1882 | ||
aa888a74 AK |
1883 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
1884 | static void __init gather_bootmem_prealloc(void) | |
1885 | { | |
1886 | struct huge_bootmem_page *m; | |
1887 | ||
1888 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 1889 | struct hstate *h = m->hstate; |
ee8f248d BB |
1890 | struct page *page; |
1891 | ||
1892 | #ifdef CONFIG_HIGHMEM | |
1893 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
8b89a116 GS |
1894 | memblock_free_late(__pa(m), |
1895 | sizeof(struct huge_bootmem_page)); | |
ee8f248d BB |
1896 | #else |
1897 | page = virt_to_page(m); | |
1898 | #endif | |
aa888a74 | 1899 | WARN_ON(page_count(page) != 1); |
18229df5 | 1900 | prep_compound_huge_page(page, h->order); |
ef5a22be | 1901 | WARN_ON(PageReserved(page)); |
aa888a74 | 1902 | prep_new_huge_page(h, page, page_to_nid(page)); |
b0320c7b RA |
1903 | /* |
1904 | * If we had gigantic hugepages allocated at boot time, we need | |
1905 | * to restore the 'stolen' pages to totalram_pages in order to | |
1906 | * fix confusing memory reports from free(1) and another | |
1907 | * side-effects, like CommitLimit going negative. | |
1908 | */ | |
bae7f4ae | 1909 | if (hstate_is_gigantic(h)) |
3dcc0571 | 1910 | adjust_managed_page_count(page, 1 << h->order); |
aa888a74 AK |
1911 | } |
1912 | } | |
1913 | ||
8faa8b07 | 1914 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
1915 | { |
1916 | unsigned long i; | |
a5516438 | 1917 | |
e5ff2159 | 1918 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 1919 | if (hstate_is_gigantic(h)) { |
aa888a74 AK |
1920 | if (!alloc_bootmem_huge_page(h)) |
1921 | break; | |
9b5e5d0f | 1922 | } else if (!alloc_fresh_huge_page(h, |
8cebfcd0 | 1923 | &node_states[N_MEMORY])) |
1da177e4 | 1924 | break; |
1da177e4 | 1925 | } |
8faa8b07 | 1926 | h->max_huge_pages = i; |
e5ff2159 AK |
1927 | } |
1928 | ||
1929 | static void __init hugetlb_init_hstates(void) | |
1930 | { | |
1931 | struct hstate *h; | |
1932 | ||
1933 | for_each_hstate(h) { | |
641844f5 NH |
1934 | if (minimum_order > huge_page_order(h)) |
1935 | minimum_order = huge_page_order(h); | |
1936 | ||
8faa8b07 | 1937 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 1938 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 1939 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 1940 | } |
641844f5 | 1941 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
1942 | } |
1943 | ||
4abd32db AK |
1944 | static char * __init memfmt(char *buf, unsigned long n) |
1945 | { | |
1946 | if (n >= (1UL << 30)) | |
1947 | sprintf(buf, "%lu GB", n >> 30); | |
1948 | else if (n >= (1UL << 20)) | |
1949 | sprintf(buf, "%lu MB", n >> 20); | |
1950 | else | |
1951 | sprintf(buf, "%lu KB", n >> 10); | |
1952 | return buf; | |
1953 | } | |
1954 | ||
e5ff2159 AK |
1955 | static void __init report_hugepages(void) |
1956 | { | |
1957 | struct hstate *h; | |
1958 | ||
1959 | for_each_hstate(h) { | |
4abd32db | 1960 | char buf[32]; |
ffb22af5 | 1961 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
4abd32db AK |
1962 | memfmt(buf, huge_page_size(h)), |
1963 | h->free_huge_pages); | |
e5ff2159 AK |
1964 | } |
1965 | } | |
1966 | ||
1da177e4 | 1967 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
1968 | static void try_to_free_low(struct hstate *h, unsigned long count, |
1969 | nodemask_t *nodes_allowed) | |
1da177e4 | 1970 | { |
4415cc8d CL |
1971 | int i; |
1972 | ||
bae7f4ae | 1973 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1974 | return; |
1975 | ||
6ae11b27 | 1976 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 1977 | struct page *page, *next; |
a5516438 AK |
1978 | struct list_head *freel = &h->hugepage_freelists[i]; |
1979 | list_for_each_entry_safe(page, next, freel, lru) { | |
1980 | if (count >= h->nr_huge_pages) | |
6b0c880d | 1981 | return; |
1da177e4 LT |
1982 | if (PageHighMem(page)) |
1983 | continue; | |
1984 | list_del(&page->lru); | |
e5ff2159 | 1985 | update_and_free_page(h, page); |
a5516438 AK |
1986 | h->free_huge_pages--; |
1987 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
1988 | } |
1989 | } | |
1990 | } | |
1991 | #else | |
6ae11b27 LS |
1992 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
1993 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
1994 | { |
1995 | } | |
1996 | #endif | |
1997 | ||
20a0307c WF |
1998 | /* |
1999 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2000 | * balanced by operating on them in a round-robin fashion. | |
2001 | * Returns 1 if an adjustment was made. | |
2002 | */ | |
6ae11b27 LS |
2003 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
2004 | int delta) | |
20a0307c | 2005 | { |
b2261026 | 2006 | int nr_nodes, node; |
20a0307c WF |
2007 | |
2008 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 2009 | |
b2261026 JK |
2010 | if (delta < 0) { |
2011 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2012 | if (h->surplus_huge_pages_node[node]) | |
2013 | goto found; | |
e8c5c824 | 2014 | } |
b2261026 JK |
2015 | } else { |
2016 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2017 | if (h->surplus_huge_pages_node[node] < | |
2018 | h->nr_huge_pages_node[node]) | |
2019 | goto found; | |
e8c5c824 | 2020 | } |
b2261026 JK |
2021 | } |
2022 | return 0; | |
20a0307c | 2023 | |
b2261026 JK |
2024 | found: |
2025 | h->surplus_huge_pages += delta; | |
2026 | h->surplus_huge_pages_node[node] += delta; | |
2027 | return 1; | |
20a0307c WF |
2028 | } |
2029 | ||
a5516438 | 2030 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
2031 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
2032 | nodemask_t *nodes_allowed) | |
1da177e4 | 2033 | { |
7893d1d5 | 2034 | unsigned long min_count, ret; |
1da177e4 | 2035 | |
944d9fec | 2036 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
aa888a74 AK |
2037 | return h->max_huge_pages; |
2038 | ||
7893d1d5 AL |
2039 | /* |
2040 | * Increase the pool size | |
2041 | * First take pages out of surplus state. Then make up the | |
2042 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f NA |
2043 | * |
2044 | * We might race with alloc_buddy_huge_page() here and be unable | |
2045 | * to convert a surplus huge page to a normal huge page. That is | |
2046 | * not critical, though, it just means the overall size of the | |
2047 | * pool might be one hugepage larger than it needs to be, but | |
2048 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 2049 | */ |
1da177e4 | 2050 | spin_lock(&hugetlb_lock); |
a5516438 | 2051 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 2052 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
2053 | break; |
2054 | } | |
2055 | ||
a5516438 | 2056 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
2057 | /* |
2058 | * If this allocation races such that we no longer need the | |
2059 | * page, free_huge_page will handle it by freeing the page | |
2060 | * and reducing the surplus. | |
2061 | */ | |
2062 | spin_unlock(&hugetlb_lock); | |
944d9fec LC |
2063 | if (hstate_is_gigantic(h)) |
2064 | ret = alloc_fresh_gigantic_page(h, nodes_allowed); | |
2065 | else | |
2066 | ret = alloc_fresh_huge_page(h, nodes_allowed); | |
7893d1d5 AL |
2067 | spin_lock(&hugetlb_lock); |
2068 | if (!ret) | |
2069 | goto out; | |
2070 | ||
536240f2 MG |
2071 | /* Bail for signals. Probably ctrl-c from user */ |
2072 | if (signal_pending(current)) | |
2073 | goto out; | |
7893d1d5 | 2074 | } |
7893d1d5 AL |
2075 | |
2076 | /* | |
2077 | * Decrease the pool size | |
2078 | * First return free pages to the buddy allocator (being careful | |
2079 | * to keep enough around to satisfy reservations). Then place | |
2080 | * pages into surplus state as needed so the pool will shrink | |
2081 | * to the desired size as pages become free. | |
d1c3fb1f NA |
2082 | * |
2083 | * By placing pages into the surplus state independent of the | |
2084 | * overcommit value, we are allowing the surplus pool size to | |
2085 | * exceed overcommit. There are few sane options here. Since | |
2086 | * alloc_buddy_huge_page() is checking the global counter, | |
2087 | * though, we'll note that we're not allowed to exceed surplus | |
2088 | * and won't grow the pool anywhere else. Not until one of the | |
2089 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 2090 | */ |
a5516438 | 2091 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 2092 | min_count = max(count, min_count); |
6ae11b27 | 2093 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 2094 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 2095 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 2096 | break; |
55f67141 | 2097 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 2098 | } |
a5516438 | 2099 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 2100 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
2101 | break; |
2102 | } | |
2103 | out: | |
a5516438 | 2104 | ret = persistent_huge_pages(h); |
1da177e4 | 2105 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 2106 | return ret; |
1da177e4 LT |
2107 | } |
2108 | ||
a3437870 NA |
2109 | #define HSTATE_ATTR_RO(_name) \ |
2110 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2111 | ||
2112 | #define HSTATE_ATTR(_name) \ | |
2113 | static struct kobj_attribute _name##_attr = \ | |
2114 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2115 | ||
2116 | static struct kobject *hugepages_kobj; | |
2117 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2118 | ||
9a305230 LS |
2119 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
2120 | ||
2121 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
2122 | { |
2123 | int i; | |
9a305230 | 2124 | |
a3437870 | 2125 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
2126 | if (hstate_kobjs[i] == kobj) { |
2127 | if (nidp) | |
2128 | *nidp = NUMA_NO_NODE; | |
a3437870 | 2129 | return &hstates[i]; |
9a305230 LS |
2130 | } |
2131 | ||
2132 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
2133 | } |
2134 | ||
06808b08 | 2135 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
2136 | struct kobj_attribute *attr, char *buf) |
2137 | { | |
9a305230 LS |
2138 | struct hstate *h; |
2139 | unsigned long nr_huge_pages; | |
2140 | int nid; | |
2141 | ||
2142 | h = kobj_to_hstate(kobj, &nid); | |
2143 | if (nid == NUMA_NO_NODE) | |
2144 | nr_huge_pages = h->nr_huge_pages; | |
2145 | else | |
2146 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2147 | ||
2148 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 2149 | } |
adbe8726 | 2150 | |
238d3c13 DR |
2151 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
2152 | struct hstate *h, int nid, | |
2153 | unsigned long count, size_t len) | |
a3437870 NA |
2154 | { |
2155 | int err; | |
bad44b5b | 2156 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 2157 | |
944d9fec | 2158 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) { |
adbe8726 EM |
2159 | err = -EINVAL; |
2160 | goto out; | |
2161 | } | |
2162 | ||
9a305230 LS |
2163 | if (nid == NUMA_NO_NODE) { |
2164 | /* | |
2165 | * global hstate attribute | |
2166 | */ | |
2167 | if (!(obey_mempolicy && | |
2168 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
2169 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 2170 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 LS |
2171 | } |
2172 | } else if (nodes_allowed) { | |
2173 | /* | |
2174 | * per node hstate attribute: adjust count to global, | |
2175 | * but restrict alloc/free to the specified node. | |
2176 | */ | |
2177 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2178 | init_nodemask_of_node(nodes_allowed, nid); | |
2179 | } else | |
8cebfcd0 | 2180 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 | 2181 | |
06808b08 | 2182 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 2183 | |
8cebfcd0 | 2184 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
2185 | NODEMASK_FREE(nodes_allowed); |
2186 | ||
2187 | return len; | |
adbe8726 EM |
2188 | out: |
2189 | NODEMASK_FREE(nodes_allowed); | |
2190 | return err; | |
06808b08 LS |
2191 | } |
2192 | ||
238d3c13 DR |
2193 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
2194 | struct kobject *kobj, const char *buf, | |
2195 | size_t len) | |
2196 | { | |
2197 | struct hstate *h; | |
2198 | unsigned long count; | |
2199 | int nid; | |
2200 | int err; | |
2201 | ||
2202 | err = kstrtoul(buf, 10, &count); | |
2203 | if (err) | |
2204 | return err; | |
2205 | ||
2206 | h = kobj_to_hstate(kobj, &nid); | |
2207 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2208 | } | |
2209 | ||
06808b08 LS |
2210 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
2211 | struct kobj_attribute *attr, char *buf) | |
2212 | { | |
2213 | return nr_hugepages_show_common(kobj, attr, buf); | |
2214 | } | |
2215 | ||
2216 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2217 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2218 | { | |
238d3c13 | 2219 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
2220 | } |
2221 | HSTATE_ATTR(nr_hugepages); | |
2222 | ||
06808b08 LS |
2223 | #ifdef CONFIG_NUMA |
2224 | ||
2225 | /* | |
2226 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2227 | * huge page alloc/free. | |
2228 | */ | |
2229 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
2230 | struct kobj_attribute *attr, char *buf) | |
2231 | { | |
2232 | return nr_hugepages_show_common(kobj, attr, buf); | |
2233 | } | |
2234 | ||
2235 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2236 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2237 | { | |
238d3c13 | 2238 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
2239 | } |
2240 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2241 | #endif | |
2242 | ||
2243 | ||
a3437870 NA |
2244 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
2245 | struct kobj_attribute *attr, char *buf) | |
2246 | { | |
9a305230 | 2247 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2248 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
2249 | } | |
adbe8726 | 2250 | |
a3437870 NA |
2251 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
2252 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2253 | { | |
2254 | int err; | |
2255 | unsigned long input; | |
9a305230 | 2256 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2257 | |
bae7f4ae | 2258 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2259 | return -EINVAL; |
2260 | ||
3dbb95f7 | 2261 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2262 | if (err) |
73ae31e5 | 2263 | return err; |
a3437870 NA |
2264 | |
2265 | spin_lock(&hugetlb_lock); | |
2266 | h->nr_overcommit_huge_pages = input; | |
2267 | spin_unlock(&hugetlb_lock); | |
2268 | ||
2269 | return count; | |
2270 | } | |
2271 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2272 | ||
2273 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2274 | struct kobj_attribute *attr, char *buf) | |
2275 | { | |
9a305230 LS |
2276 | struct hstate *h; |
2277 | unsigned long free_huge_pages; | |
2278 | int nid; | |
2279 | ||
2280 | h = kobj_to_hstate(kobj, &nid); | |
2281 | if (nid == NUMA_NO_NODE) | |
2282 | free_huge_pages = h->free_huge_pages; | |
2283 | else | |
2284 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2285 | ||
2286 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
2287 | } |
2288 | HSTATE_ATTR_RO(free_hugepages); | |
2289 | ||
2290 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2291 | struct kobj_attribute *attr, char *buf) | |
2292 | { | |
9a305230 | 2293 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2294 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
2295 | } | |
2296 | HSTATE_ATTR_RO(resv_hugepages); | |
2297 | ||
2298 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2299 | struct kobj_attribute *attr, char *buf) | |
2300 | { | |
9a305230 LS |
2301 | struct hstate *h; |
2302 | unsigned long surplus_huge_pages; | |
2303 | int nid; | |
2304 | ||
2305 | h = kobj_to_hstate(kobj, &nid); | |
2306 | if (nid == NUMA_NO_NODE) | |
2307 | surplus_huge_pages = h->surplus_huge_pages; | |
2308 | else | |
2309 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2310 | ||
2311 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
2312 | } |
2313 | HSTATE_ATTR_RO(surplus_hugepages); | |
2314 | ||
2315 | static struct attribute *hstate_attrs[] = { | |
2316 | &nr_hugepages_attr.attr, | |
2317 | &nr_overcommit_hugepages_attr.attr, | |
2318 | &free_hugepages_attr.attr, | |
2319 | &resv_hugepages_attr.attr, | |
2320 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
2321 | #ifdef CONFIG_NUMA |
2322 | &nr_hugepages_mempolicy_attr.attr, | |
2323 | #endif | |
a3437870 NA |
2324 | NULL, |
2325 | }; | |
2326 | ||
2327 | static struct attribute_group hstate_attr_group = { | |
2328 | .attrs = hstate_attrs, | |
2329 | }; | |
2330 | ||
094e9539 JM |
2331 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
2332 | struct kobject **hstate_kobjs, | |
2333 | struct attribute_group *hstate_attr_group) | |
a3437870 NA |
2334 | { |
2335 | int retval; | |
972dc4de | 2336 | int hi = hstate_index(h); |
a3437870 | 2337 | |
9a305230 LS |
2338 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
2339 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
2340 | return -ENOMEM; |
2341 | ||
9a305230 | 2342 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 2343 | if (retval) |
9a305230 | 2344 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
2345 | |
2346 | return retval; | |
2347 | } | |
2348 | ||
2349 | static void __init hugetlb_sysfs_init(void) | |
2350 | { | |
2351 | struct hstate *h; | |
2352 | int err; | |
2353 | ||
2354 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
2355 | if (!hugepages_kobj) | |
2356 | return; | |
2357 | ||
2358 | for_each_hstate(h) { | |
9a305230 LS |
2359 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
2360 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 2361 | if (err) |
ffb22af5 | 2362 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
2363 | } |
2364 | } | |
2365 | ||
9a305230 LS |
2366 | #ifdef CONFIG_NUMA |
2367 | ||
2368 | /* | |
2369 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
2370 | * with node devices in node_devices[] using a parallel array. The array |
2371 | * index of a node device or _hstate == node id. | |
2372 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
2373 | * the base kernel, on the hugetlb module. |
2374 | */ | |
2375 | struct node_hstate { | |
2376 | struct kobject *hugepages_kobj; | |
2377 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2378 | }; | |
2379 | struct node_hstate node_hstates[MAX_NUMNODES]; | |
2380 | ||
2381 | /* | |
10fbcf4c | 2382 | * A subset of global hstate attributes for node devices |
9a305230 LS |
2383 | */ |
2384 | static struct attribute *per_node_hstate_attrs[] = { | |
2385 | &nr_hugepages_attr.attr, | |
2386 | &free_hugepages_attr.attr, | |
2387 | &surplus_hugepages_attr.attr, | |
2388 | NULL, | |
2389 | }; | |
2390 | ||
2391 | static struct attribute_group per_node_hstate_attr_group = { | |
2392 | .attrs = per_node_hstate_attrs, | |
2393 | }; | |
2394 | ||
2395 | /* | |
10fbcf4c | 2396 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
2397 | * Returns node id via non-NULL nidp. |
2398 | */ | |
2399 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2400 | { | |
2401 | int nid; | |
2402 | ||
2403 | for (nid = 0; nid < nr_node_ids; nid++) { | |
2404 | struct node_hstate *nhs = &node_hstates[nid]; | |
2405 | int i; | |
2406 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
2407 | if (nhs->hstate_kobjs[i] == kobj) { | |
2408 | if (nidp) | |
2409 | *nidp = nid; | |
2410 | return &hstates[i]; | |
2411 | } | |
2412 | } | |
2413 | ||
2414 | BUG(); | |
2415 | return NULL; | |
2416 | } | |
2417 | ||
2418 | /* | |
10fbcf4c | 2419 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
2420 | * No-op if no hstate attributes attached. |
2421 | */ | |
3cd8b44f | 2422 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
2423 | { |
2424 | struct hstate *h; | |
10fbcf4c | 2425 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2426 | |
2427 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 2428 | return; /* no hstate attributes */ |
9a305230 | 2429 | |
972dc4de AK |
2430 | for_each_hstate(h) { |
2431 | int idx = hstate_index(h); | |
2432 | if (nhs->hstate_kobjs[idx]) { | |
2433 | kobject_put(nhs->hstate_kobjs[idx]); | |
2434 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 2435 | } |
972dc4de | 2436 | } |
9a305230 LS |
2437 | |
2438 | kobject_put(nhs->hugepages_kobj); | |
2439 | nhs->hugepages_kobj = NULL; | |
2440 | } | |
2441 | ||
2442 | /* | |
10fbcf4c | 2443 | * hugetlb module exit: unregister hstate attributes from node devices |
9a305230 LS |
2444 | * that have them. |
2445 | */ | |
2446 | static void hugetlb_unregister_all_nodes(void) | |
2447 | { | |
2448 | int nid; | |
2449 | ||
2450 | /* | |
10fbcf4c | 2451 | * disable node device registrations. |
9a305230 LS |
2452 | */ |
2453 | register_hugetlbfs_with_node(NULL, NULL); | |
2454 | ||
2455 | /* | |
2456 | * remove hstate attributes from any nodes that have them. | |
2457 | */ | |
2458 | for (nid = 0; nid < nr_node_ids; nid++) | |
8732794b | 2459 | hugetlb_unregister_node(node_devices[nid]); |
9a305230 LS |
2460 | } |
2461 | ||
2462 | /* | |
10fbcf4c | 2463 | * Register hstate attributes for a single node device. |
9a305230 LS |
2464 | * No-op if attributes already registered. |
2465 | */ | |
3cd8b44f | 2466 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
2467 | { |
2468 | struct hstate *h; | |
10fbcf4c | 2469 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2470 | int err; |
2471 | ||
2472 | if (nhs->hugepages_kobj) | |
2473 | return; /* already allocated */ | |
2474 | ||
2475 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 2476 | &node->dev.kobj); |
9a305230 LS |
2477 | if (!nhs->hugepages_kobj) |
2478 | return; | |
2479 | ||
2480 | for_each_hstate(h) { | |
2481 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
2482 | nhs->hstate_kobjs, | |
2483 | &per_node_hstate_attr_group); | |
2484 | if (err) { | |
ffb22af5 AM |
2485 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
2486 | h->name, node->dev.id); | |
9a305230 LS |
2487 | hugetlb_unregister_node(node); |
2488 | break; | |
2489 | } | |
2490 | } | |
2491 | } | |
2492 | ||
2493 | /* | |
9b5e5d0f | 2494 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
2495 | * devices of nodes that have memory. All on-line nodes should have |
2496 | * registered their associated device by this time. | |
9a305230 | 2497 | */ |
7d9ca000 | 2498 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
2499 | { |
2500 | int nid; | |
2501 | ||
8cebfcd0 | 2502 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 2503 | struct node *node = node_devices[nid]; |
10fbcf4c | 2504 | if (node->dev.id == nid) |
9a305230 LS |
2505 | hugetlb_register_node(node); |
2506 | } | |
2507 | ||
2508 | /* | |
10fbcf4c | 2509 | * Let the node device driver know we're here so it can |
9a305230 LS |
2510 | * [un]register hstate attributes on node hotplug. |
2511 | */ | |
2512 | register_hugetlbfs_with_node(hugetlb_register_node, | |
2513 | hugetlb_unregister_node); | |
2514 | } | |
2515 | #else /* !CONFIG_NUMA */ | |
2516 | ||
2517 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2518 | { | |
2519 | BUG(); | |
2520 | if (nidp) | |
2521 | *nidp = -1; | |
2522 | return NULL; | |
2523 | } | |
2524 | ||
2525 | static void hugetlb_unregister_all_nodes(void) { } | |
2526 | ||
2527 | static void hugetlb_register_all_nodes(void) { } | |
2528 | ||
2529 | #endif | |
2530 | ||
a3437870 NA |
2531 | static void __exit hugetlb_exit(void) |
2532 | { | |
2533 | struct hstate *h; | |
2534 | ||
9a305230 LS |
2535 | hugetlb_unregister_all_nodes(); |
2536 | ||
a3437870 | 2537 | for_each_hstate(h) { |
972dc4de | 2538 | kobject_put(hstate_kobjs[hstate_index(h)]); |
a3437870 NA |
2539 | } |
2540 | ||
2541 | kobject_put(hugepages_kobj); | |
c672c7f2 | 2542 | kfree(hugetlb_fault_mutex_table); |
a3437870 NA |
2543 | } |
2544 | module_exit(hugetlb_exit); | |
2545 | ||
2546 | static int __init hugetlb_init(void) | |
2547 | { | |
8382d914 DB |
2548 | int i; |
2549 | ||
457c1b27 | 2550 | if (!hugepages_supported()) |
0ef89d25 | 2551 | return 0; |
a3437870 | 2552 | |
e11bfbfc NP |
2553 | if (!size_to_hstate(default_hstate_size)) { |
2554 | default_hstate_size = HPAGE_SIZE; | |
2555 | if (!size_to_hstate(default_hstate_size)) | |
2556 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 2557 | } |
972dc4de | 2558 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
e11bfbfc NP |
2559 | if (default_hstate_max_huge_pages) |
2560 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
a3437870 NA |
2561 | |
2562 | hugetlb_init_hstates(); | |
aa888a74 | 2563 | gather_bootmem_prealloc(); |
a3437870 NA |
2564 | report_hugepages(); |
2565 | ||
2566 | hugetlb_sysfs_init(); | |
9a305230 | 2567 | hugetlb_register_all_nodes(); |
7179e7bf | 2568 | hugetlb_cgroup_file_init(); |
9a305230 | 2569 | |
8382d914 DB |
2570 | #ifdef CONFIG_SMP |
2571 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
2572 | #else | |
2573 | num_fault_mutexes = 1; | |
2574 | #endif | |
c672c7f2 | 2575 | hugetlb_fault_mutex_table = |
8382d914 | 2576 | kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); |
c672c7f2 | 2577 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
2578 | |
2579 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 2580 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
2581 | return 0; |
2582 | } | |
2583 | module_init(hugetlb_init); | |
2584 | ||
2585 | /* Should be called on processing a hugepagesz=... option */ | |
2586 | void __init hugetlb_add_hstate(unsigned order) | |
2587 | { | |
2588 | struct hstate *h; | |
8faa8b07 AK |
2589 | unsigned long i; |
2590 | ||
a3437870 | 2591 | if (size_to_hstate(PAGE_SIZE << order)) { |
ffb22af5 | 2592 | pr_warning("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
2593 | return; |
2594 | } | |
47d38344 | 2595 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 2596 | BUG_ON(order == 0); |
47d38344 | 2597 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
2598 | h->order = order; |
2599 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
2600 | h->nr_huge_pages = 0; |
2601 | h->free_huge_pages = 0; | |
2602 | for (i = 0; i < MAX_NUMNODES; ++i) | |
2603 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 2604 | INIT_LIST_HEAD(&h->hugepage_activelist); |
8cebfcd0 LJ |
2605 | h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); |
2606 | h->next_nid_to_free = first_node(node_states[N_MEMORY]); | |
a3437870 NA |
2607 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
2608 | huge_page_size(h)/1024); | |
8faa8b07 | 2609 | |
a3437870 NA |
2610 | parsed_hstate = h; |
2611 | } | |
2612 | ||
e11bfbfc | 2613 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
2614 | { |
2615 | unsigned long *mhp; | |
8faa8b07 | 2616 | static unsigned long *last_mhp; |
a3437870 NA |
2617 | |
2618 | /* | |
47d38344 | 2619 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
2620 | * so this hugepages= parameter goes to the "default hstate". |
2621 | */ | |
47d38344 | 2622 | if (!hugetlb_max_hstate) |
a3437870 NA |
2623 | mhp = &default_hstate_max_huge_pages; |
2624 | else | |
2625 | mhp = &parsed_hstate->max_huge_pages; | |
2626 | ||
8faa8b07 | 2627 | if (mhp == last_mhp) { |
ffb22af5 AM |
2628 | pr_warning("hugepages= specified twice without " |
2629 | "interleaving hugepagesz=, ignoring\n"); | |
8faa8b07 AK |
2630 | return 1; |
2631 | } | |
2632 | ||
a3437870 NA |
2633 | if (sscanf(s, "%lu", mhp) <= 0) |
2634 | *mhp = 0; | |
2635 | ||
8faa8b07 AK |
2636 | /* |
2637 | * Global state is always initialized later in hugetlb_init. | |
2638 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
2639 | * use the bootmem allocator. | |
2640 | */ | |
47d38344 | 2641 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
2642 | hugetlb_hstate_alloc_pages(parsed_hstate); |
2643 | ||
2644 | last_mhp = mhp; | |
2645 | ||
a3437870 NA |
2646 | return 1; |
2647 | } | |
e11bfbfc NP |
2648 | __setup("hugepages=", hugetlb_nrpages_setup); |
2649 | ||
2650 | static int __init hugetlb_default_setup(char *s) | |
2651 | { | |
2652 | default_hstate_size = memparse(s, &s); | |
2653 | return 1; | |
2654 | } | |
2655 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 2656 | |
8a213460 NA |
2657 | static unsigned int cpuset_mems_nr(unsigned int *array) |
2658 | { | |
2659 | int node; | |
2660 | unsigned int nr = 0; | |
2661 | ||
2662 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
2663 | nr += array[node]; | |
2664 | ||
2665 | return nr; | |
2666 | } | |
2667 | ||
2668 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
2669 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
2670 | struct ctl_table *table, int write, | |
2671 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 2672 | { |
e5ff2159 | 2673 | struct hstate *h = &default_hstate; |
238d3c13 | 2674 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 2675 | int ret; |
e5ff2159 | 2676 | |
457c1b27 NA |
2677 | if (!hugepages_supported()) |
2678 | return -ENOTSUPP; | |
2679 | ||
e5ff2159 AK |
2680 | table->data = &tmp; |
2681 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2682 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2683 | if (ret) | |
2684 | goto out; | |
e5ff2159 | 2685 | |
238d3c13 DR |
2686 | if (write) |
2687 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
2688 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
2689 | out: |
2690 | return ret; | |
1da177e4 | 2691 | } |
396faf03 | 2692 | |
06808b08 LS |
2693 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2694 | void __user *buffer, size_t *length, loff_t *ppos) | |
2695 | { | |
2696 | ||
2697 | return hugetlb_sysctl_handler_common(false, table, write, | |
2698 | buffer, length, ppos); | |
2699 | } | |
2700 | ||
2701 | #ifdef CONFIG_NUMA | |
2702 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2703 | void __user *buffer, size_t *length, loff_t *ppos) | |
2704 | { | |
2705 | return hugetlb_sysctl_handler_common(true, table, write, | |
2706 | buffer, length, ppos); | |
2707 | } | |
2708 | #endif /* CONFIG_NUMA */ | |
2709 | ||
a3d0c6aa | 2710 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 2711 | void __user *buffer, |
a3d0c6aa NA |
2712 | size_t *length, loff_t *ppos) |
2713 | { | |
a5516438 | 2714 | struct hstate *h = &default_hstate; |
e5ff2159 | 2715 | unsigned long tmp; |
08d4a246 | 2716 | int ret; |
e5ff2159 | 2717 | |
457c1b27 NA |
2718 | if (!hugepages_supported()) |
2719 | return -ENOTSUPP; | |
2720 | ||
c033a93c | 2721 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 2722 | |
bae7f4ae | 2723 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
2724 | return -EINVAL; |
2725 | ||
e5ff2159 AK |
2726 | table->data = &tmp; |
2727 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2728 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2729 | if (ret) | |
2730 | goto out; | |
e5ff2159 AK |
2731 | |
2732 | if (write) { | |
2733 | spin_lock(&hugetlb_lock); | |
2734 | h->nr_overcommit_huge_pages = tmp; | |
2735 | spin_unlock(&hugetlb_lock); | |
2736 | } | |
08d4a246 MH |
2737 | out: |
2738 | return ret; | |
a3d0c6aa NA |
2739 | } |
2740 | ||
1da177e4 LT |
2741 | #endif /* CONFIG_SYSCTL */ |
2742 | ||
e1759c21 | 2743 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 2744 | { |
a5516438 | 2745 | struct hstate *h = &default_hstate; |
457c1b27 NA |
2746 | if (!hugepages_supported()) |
2747 | return; | |
e1759c21 | 2748 | seq_printf(m, |
4f98a2fe RR |
2749 | "HugePages_Total: %5lu\n" |
2750 | "HugePages_Free: %5lu\n" | |
2751 | "HugePages_Rsvd: %5lu\n" | |
2752 | "HugePages_Surp: %5lu\n" | |
2753 | "Hugepagesize: %8lu kB\n", | |
a5516438 AK |
2754 | h->nr_huge_pages, |
2755 | h->free_huge_pages, | |
2756 | h->resv_huge_pages, | |
2757 | h->surplus_huge_pages, | |
2758 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
2759 | } |
2760 | ||
2761 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
2762 | { | |
a5516438 | 2763 | struct hstate *h = &default_hstate; |
457c1b27 NA |
2764 | if (!hugepages_supported()) |
2765 | return 0; | |
1da177e4 LT |
2766 | return sprintf(buf, |
2767 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
2768 | "Node %d HugePages_Free: %5u\n" |
2769 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
2770 | nid, h->nr_huge_pages_node[nid], |
2771 | nid, h->free_huge_pages_node[nid], | |
2772 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
2773 | } |
2774 | ||
949f7ec5 DR |
2775 | void hugetlb_show_meminfo(void) |
2776 | { | |
2777 | struct hstate *h; | |
2778 | int nid; | |
2779 | ||
457c1b27 NA |
2780 | if (!hugepages_supported()) |
2781 | return; | |
2782 | ||
949f7ec5 DR |
2783 | for_each_node_state(nid, N_MEMORY) |
2784 | for_each_hstate(h) | |
2785 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
2786 | nid, | |
2787 | h->nr_huge_pages_node[nid], | |
2788 | h->free_huge_pages_node[nid], | |
2789 | h->surplus_huge_pages_node[nid], | |
2790 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
2791 | } | |
2792 | ||
1da177e4 LT |
2793 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
2794 | unsigned long hugetlb_total_pages(void) | |
2795 | { | |
d0028588 WL |
2796 | struct hstate *h; |
2797 | unsigned long nr_total_pages = 0; | |
2798 | ||
2799 | for_each_hstate(h) | |
2800 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
2801 | return nr_total_pages; | |
1da177e4 | 2802 | } |
1da177e4 | 2803 | |
a5516438 | 2804 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
2805 | { |
2806 | int ret = -ENOMEM; | |
2807 | ||
2808 | spin_lock(&hugetlb_lock); | |
2809 | /* | |
2810 | * When cpuset is configured, it breaks the strict hugetlb page | |
2811 | * reservation as the accounting is done on a global variable. Such | |
2812 | * reservation is completely rubbish in the presence of cpuset because | |
2813 | * the reservation is not checked against page availability for the | |
2814 | * current cpuset. Application can still potentially OOM'ed by kernel | |
2815 | * with lack of free htlb page in cpuset that the task is in. | |
2816 | * Attempt to enforce strict accounting with cpuset is almost | |
2817 | * impossible (or too ugly) because cpuset is too fluid that | |
2818 | * task or memory node can be dynamically moved between cpusets. | |
2819 | * | |
2820 | * The change of semantics for shared hugetlb mapping with cpuset is | |
2821 | * undesirable. However, in order to preserve some of the semantics, | |
2822 | * we fall back to check against current free page availability as | |
2823 | * a best attempt and hopefully to minimize the impact of changing | |
2824 | * semantics that cpuset has. | |
2825 | */ | |
2826 | if (delta > 0) { | |
a5516438 | 2827 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
2828 | goto out; |
2829 | ||
a5516438 AK |
2830 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
2831 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
2832 | goto out; |
2833 | } | |
2834 | } | |
2835 | ||
2836 | ret = 0; | |
2837 | if (delta < 0) | |
a5516438 | 2838 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
2839 | |
2840 | out: | |
2841 | spin_unlock(&hugetlb_lock); | |
2842 | return ret; | |
2843 | } | |
2844 | ||
84afd99b AW |
2845 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
2846 | { | |
f522c3ac | 2847 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
2848 | |
2849 | /* | |
2850 | * This new VMA should share its siblings reservation map if present. | |
2851 | * The VMA will only ever have a valid reservation map pointer where | |
2852 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 2853 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
2854 | * after this open call completes. It is therefore safe to take a |
2855 | * new reference here without additional locking. | |
2856 | */ | |
4e35f483 | 2857 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 2858 | kref_get(&resv->refs); |
84afd99b AW |
2859 | } |
2860 | ||
a1e78772 MG |
2861 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
2862 | { | |
a5516438 | 2863 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 2864 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 2865 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 2866 | unsigned long reserve, start, end; |
1c5ecae3 | 2867 | long gbl_reserve; |
84afd99b | 2868 | |
4e35f483 JK |
2869 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
2870 | return; | |
84afd99b | 2871 | |
4e35f483 JK |
2872 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
2873 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 2874 | |
4e35f483 | 2875 | reserve = (end - start) - region_count(resv, start, end); |
84afd99b | 2876 | |
4e35f483 JK |
2877 | kref_put(&resv->refs, resv_map_release); |
2878 | ||
2879 | if (reserve) { | |
1c5ecae3 MK |
2880 | /* |
2881 | * Decrement reserve counts. The global reserve count may be | |
2882 | * adjusted if the subpool has a minimum size. | |
2883 | */ | |
2884 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
2885 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 2886 | } |
a1e78772 MG |
2887 | } |
2888 | ||
1da177e4 LT |
2889 | /* |
2890 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
2891 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
2892 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
2893 | * this far. | |
2894 | */ | |
d0217ac0 | 2895 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
2896 | { |
2897 | BUG(); | |
d0217ac0 | 2898 | return 0; |
1da177e4 LT |
2899 | } |
2900 | ||
f0f37e2f | 2901 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 2902 | .fault = hugetlb_vm_op_fault, |
84afd99b | 2903 | .open = hugetlb_vm_op_open, |
a1e78772 | 2904 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
2905 | }; |
2906 | ||
1e8f889b DG |
2907 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
2908 | int writable) | |
63551ae0 DG |
2909 | { |
2910 | pte_t entry; | |
2911 | ||
1e8f889b | 2912 | if (writable) { |
106c992a GS |
2913 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
2914 | vma->vm_page_prot))); | |
63551ae0 | 2915 | } else { |
106c992a GS |
2916 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
2917 | vma->vm_page_prot)); | |
63551ae0 DG |
2918 | } |
2919 | entry = pte_mkyoung(entry); | |
2920 | entry = pte_mkhuge(entry); | |
d9ed9faa | 2921 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
2922 | |
2923 | return entry; | |
2924 | } | |
2925 | ||
1e8f889b DG |
2926 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
2927 | unsigned long address, pte_t *ptep) | |
2928 | { | |
2929 | pte_t entry; | |
2930 | ||
106c992a | 2931 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 2932 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 2933 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
2934 | } |
2935 | ||
4a705fef NH |
2936 | static int is_hugetlb_entry_migration(pte_t pte) |
2937 | { | |
2938 | swp_entry_t swp; | |
2939 | ||
2940 | if (huge_pte_none(pte) || pte_present(pte)) | |
2941 | return 0; | |
2942 | swp = pte_to_swp_entry(pte); | |
2943 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
2944 | return 1; | |
2945 | else | |
2946 | return 0; | |
2947 | } | |
2948 | ||
2949 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
2950 | { | |
2951 | swp_entry_t swp; | |
2952 | ||
2953 | if (huge_pte_none(pte) || pte_present(pte)) | |
2954 | return 0; | |
2955 | swp = pte_to_swp_entry(pte); | |
2956 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
2957 | return 1; | |
2958 | else | |
2959 | return 0; | |
2960 | } | |
1e8f889b | 2961 | |
63551ae0 DG |
2962 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
2963 | struct vm_area_struct *vma) | |
2964 | { | |
2965 | pte_t *src_pte, *dst_pte, entry; | |
2966 | struct page *ptepage; | |
1c59827d | 2967 | unsigned long addr; |
1e8f889b | 2968 | int cow; |
a5516438 AK |
2969 | struct hstate *h = hstate_vma(vma); |
2970 | unsigned long sz = huge_page_size(h); | |
e8569dd2 AS |
2971 | unsigned long mmun_start; /* For mmu_notifiers */ |
2972 | unsigned long mmun_end; /* For mmu_notifiers */ | |
2973 | int ret = 0; | |
1e8f889b DG |
2974 | |
2975 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 2976 | |
e8569dd2 AS |
2977 | mmun_start = vma->vm_start; |
2978 | mmun_end = vma->vm_end; | |
2979 | if (cow) | |
2980 | mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); | |
2981 | ||
a5516438 | 2982 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 2983 | spinlock_t *src_ptl, *dst_ptl; |
c74df32c HD |
2984 | src_pte = huge_pte_offset(src, addr); |
2985 | if (!src_pte) | |
2986 | continue; | |
a5516438 | 2987 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
2988 | if (!dst_pte) { |
2989 | ret = -ENOMEM; | |
2990 | break; | |
2991 | } | |
c5c99429 LW |
2992 | |
2993 | /* If the pagetables are shared don't copy or take references */ | |
2994 | if (dst_pte == src_pte) | |
2995 | continue; | |
2996 | ||
cb900f41 KS |
2997 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
2998 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
2999 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef NH |
3000 | entry = huge_ptep_get(src_pte); |
3001 | if (huge_pte_none(entry)) { /* skip none entry */ | |
3002 | ; | |
3003 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
3004 | is_hugetlb_entry_hwpoisoned(entry))) { | |
3005 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
3006 | ||
3007 | if (is_write_migration_entry(swp_entry) && cow) { | |
3008 | /* | |
3009 | * COW mappings require pages in both | |
3010 | * parent and child to be set to read. | |
3011 | */ | |
3012 | make_migration_entry_read(&swp_entry); | |
3013 | entry = swp_entry_to_pte(swp_entry); | |
3014 | set_huge_pte_at(src, addr, src_pte, entry); | |
3015 | } | |
3016 | set_huge_pte_at(dst, addr, dst_pte, entry); | |
3017 | } else { | |
34ee645e | 3018 | if (cow) { |
7f2e9525 | 3019 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e JR |
3020 | mmu_notifier_invalidate_range(src, mmun_start, |
3021 | mmun_end); | |
3022 | } | |
0253d634 | 3023 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
3024 | ptepage = pte_page(entry); |
3025 | get_page(ptepage); | |
0fe6e20b | 3026 | page_dup_rmap(ptepage); |
1c59827d HD |
3027 | set_huge_pte_at(dst, addr, dst_pte, entry); |
3028 | } | |
cb900f41 KS |
3029 | spin_unlock(src_ptl); |
3030 | spin_unlock(dst_ptl); | |
63551ae0 | 3031 | } |
63551ae0 | 3032 | |
e8569dd2 AS |
3033 | if (cow) |
3034 | mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); | |
3035 | ||
3036 | return ret; | |
63551ae0 DG |
3037 | } |
3038 | ||
24669e58 AK |
3039 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
3040 | unsigned long start, unsigned long end, | |
3041 | struct page *ref_page) | |
63551ae0 | 3042 | { |
24669e58 | 3043 | int force_flush = 0; |
63551ae0 DG |
3044 | struct mm_struct *mm = vma->vm_mm; |
3045 | unsigned long address; | |
c7546f8f | 3046 | pte_t *ptep; |
63551ae0 | 3047 | pte_t pte; |
cb900f41 | 3048 | spinlock_t *ptl; |
63551ae0 | 3049 | struct page *page; |
a5516438 AK |
3050 | struct hstate *h = hstate_vma(vma); |
3051 | unsigned long sz = huge_page_size(h); | |
2ec74c3e SG |
3052 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
3053 | const unsigned long mmun_end = end; /* For mmu_notifiers */ | |
a5516438 | 3054 | |
63551ae0 | 3055 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
3056 | BUG_ON(start & ~huge_page_mask(h)); |
3057 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 3058 | |
24669e58 | 3059 | tlb_start_vma(tlb, vma); |
2ec74c3e | 3060 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
569f48b8 | 3061 | address = start; |
24669e58 | 3062 | again: |
569f48b8 | 3063 | for (; address < end; address += sz) { |
c7546f8f | 3064 | ptep = huge_pte_offset(mm, address); |
4c887265 | 3065 | if (!ptep) |
c7546f8f DG |
3066 | continue; |
3067 | ||
cb900f41 | 3068 | ptl = huge_pte_lock(h, mm, ptep); |
39dde65c | 3069 | if (huge_pmd_unshare(mm, &address, ptep)) |
cb900f41 | 3070 | goto unlock; |
39dde65c | 3071 | |
6629326b HD |
3072 | pte = huge_ptep_get(ptep); |
3073 | if (huge_pte_none(pte)) | |
cb900f41 | 3074 | goto unlock; |
6629326b HD |
3075 | |
3076 | /* | |
9fbc1f63 NH |
3077 | * Migrating hugepage or HWPoisoned hugepage is already |
3078 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 3079 | */ |
9fbc1f63 | 3080 | if (unlikely(!pte_present(pte))) { |
106c992a | 3081 | huge_pte_clear(mm, address, ptep); |
cb900f41 | 3082 | goto unlock; |
8c4894c6 | 3083 | } |
6629326b HD |
3084 | |
3085 | page = pte_page(pte); | |
04f2cbe3 MG |
3086 | /* |
3087 | * If a reference page is supplied, it is because a specific | |
3088 | * page is being unmapped, not a range. Ensure the page we | |
3089 | * are about to unmap is the actual page of interest. | |
3090 | */ | |
3091 | if (ref_page) { | |
04f2cbe3 | 3092 | if (page != ref_page) |
cb900f41 | 3093 | goto unlock; |
04f2cbe3 MG |
3094 | |
3095 | /* | |
3096 | * Mark the VMA as having unmapped its page so that | |
3097 | * future faults in this VMA will fail rather than | |
3098 | * looking like data was lost | |
3099 | */ | |
3100 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
3101 | } | |
3102 | ||
c7546f8f | 3103 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
24669e58 | 3104 | tlb_remove_tlb_entry(tlb, ptep, address); |
106c992a | 3105 | if (huge_pte_dirty(pte)) |
6649a386 | 3106 | set_page_dirty(page); |
9e81130b | 3107 | |
24669e58 AK |
3108 | page_remove_rmap(page); |
3109 | force_flush = !__tlb_remove_page(tlb, page); | |
cb900f41 | 3110 | if (force_flush) { |
569f48b8 | 3111 | address += sz; |
cb900f41 | 3112 | spin_unlock(ptl); |
24669e58 | 3113 | break; |
cb900f41 | 3114 | } |
9e81130b | 3115 | /* Bail out after unmapping reference page if supplied */ |
cb900f41 KS |
3116 | if (ref_page) { |
3117 | spin_unlock(ptl); | |
9e81130b | 3118 | break; |
cb900f41 KS |
3119 | } |
3120 | unlock: | |
3121 | spin_unlock(ptl); | |
63551ae0 | 3122 | } |
24669e58 AK |
3123 | /* |
3124 | * mmu_gather ran out of room to batch pages, we break out of | |
3125 | * the PTE lock to avoid doing the potential expensive TLB invalidate | |
3126 | * and page-free while holding it. | |
3127 | */ | |
3128 | if (force_flush) { | |
3129 | force_flush = 0; | |
3130 | tlb_flush_mmu(tlb); | |
3131 | if (address < end && !ref_page) | |
3132 | goto again; | |
fe1668ae | 3133 | } |
2ec74c3e | 3134 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
24669e58 | 3135 | tlb_end_vma(tlb, vma); |
1da177e4 | 3136 | } |
63551ae0 | 3137 | |
d833352a MG |
3138 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
3139 | struct vm_area_struct *vma, unsigned long start, | |
3140 | unsigned long end, struct page *ref_page) | |
3141 | { | |
3142 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
3143 | ||
3144 | /* | |
3145 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
3146 | * test will fail on a vma being torn down, and not grab a page table | |
3147 | * on its way out. We're lucky that the flag has such an appropriate | |
3148 | * name, and can in fact be safely cleared here. We could clear it | |
3149 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 3150 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 3151 | * because in the context this is called, the VMA is about to be |
c8c06efa | 3152 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
3153 | */ |
3154 | vma->vm_flags &= ~VM_MAYSHARE; | |
3155 | } | |
3156 | ||
502717f4 | 3157 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 3158 | unsigned long end, struct page *ref_page) |
502717f4 | 3159 | { |
24669e58 AK |
3160 | struct mm_struct *mm; |
3161 | struct mmu_gather tlb; | |
3162 | ||
3163 | mm = vma->vm_mm; | |
3164 | ||
2b047252 | 3165 | tlb_gather_mmu(&tlb, mm, start, end); |
24669e58 AK |
3166 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
3167 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 CK |
3168 | } |
3169 | ||
04f2cbe3 MG |
3170 | /* |
3171 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
3172 | * mappping it owns the reserve page for. The intention is to unmap the page | |
3173 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
3174 | * same region. | |
3175 | */ | |
2f4612af DB |
3176 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
3177 | struct page *page, unsigned long address) | |
04f2cbe3 | 3178 | { |
7526674d | 3179 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
3180 | struct vm_area_struct *iter_vma; |
3181 | struct address_space *mapping; | |
04f2cbe3 MG |
3182 | pgoff_t pgoff; |
3183 | ||
3184 | /* | |
3185 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
3186 | * from page cache lookup which is in HPAGE_SIZE units. | |
3187 | */ | |
7526674d | 3188 | address = address & huge_page_mask(h); |
36e4f20a MH |
3189 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
3190 | vma->vm_pgoff; | |
496ad9aa | 3191 | mapping = file_inode(vma->vm_file)->i_mapping; |
04f2cbe3 | 3192 | |
4eb2b1dc MG |
3193 | /* |
3194 | * Take the mapping lock for the duration of the table walk. As | |
3195 | * this mapping should be shared between all the VMAs, | |
3196 | * __unmap_hugepage_range() is called as the lock is already held | |
3197 | */ | |
83cde9e8 | 3198 | i_mmap_lock_write(mapping); |
6b2dbba8 | 3199 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
3200 | /* Do not unmap the current VMA */ |
3201 | if (iter_vma == vma) | |
3202 | continue; | |
3203 | ||
3204 | /* | |
3205 | * Unmap the page from other VMAs without their own reserves. | |
3206 | * They get marked to be SIGKILLed if they fault in these | |
3207 | * areas. This is because a future no-page fault on this VMA | |
3208 | * could insert a zeroed page instead of the data existing | |
3209 | * from the time of fork. This would look like data corruption | |
3210 | */ | |
3211 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
3212 | unmap_hugepage_range(iter_vma, address, |
3213 | address + huge_page_size(h), page); | |
04f2cbe3 | 3214 | } |
83cde9e8 | 3215 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
3216 | } |
3217 | ||
0fe6e20b NH |
3218 | /* |
3219 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
3220 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
3221 | * cannot race with other handlers or page migration. | |
3222 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 3223 | */ |
1e8f889b | 3224 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
04f2cbe3 | 3225 | unsigned long address, pte_t *ptep, pte_t pte, |
cb900f41 | 3226 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 3227 | { |
a5516438 | 3228 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 3229 | struct page *old_page, *new_page; |
ad4404a2 | 3230 | int ret = 0, outside_reserve = 0; |
2ec74c3e SG |
3231 | unsigned long mmun_start; /* For mmu_notifiers */ |
3232 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1e8f889b DG |
3233 | |
3234 | old_page = pte_page(pte); | |
3235 | ||
04f2cbe3 | 3236 | retry_avoidcopy: |
1e8f889b DG |
3237 | /* If no-one else is actually using this page, avoid the copy |
3238 | * and just make the page writable */ | |
37a2140d JK |
3239 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
3240 | page_move_anon_rmap(old_page, vma, address); | |
1e8f889b | 3241 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 3242 | return 0; |
1e8f889b DG |
3243 | } |
3244 | ||
04f2cbe3 MG |
3245 | /* |
3246 | * If the process that created a MAP_PRIVATE mapping is about to | |
3247 | * perform a COW due to a shared page count, attempt to satisfy | |
3248 | * the allocation without using the existing reserves. The pagecache | |
3249 | * page is used to determine if the reserve at this address was | |
3250 | * consumed or not. If reserves were used, a partial faulted mapping | |
3251 | * at the time of fork() could consume its reserves on COW instead | |
3252 | * of the full address range. | |
3253 | */ | |
5944d011 | 3254 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
3255 | old_page != pagecache_page) |
3256 | outside_reserve = 1; | |
3257 | ||
1e8f889b | 3258 | page_cache_get(old_page); |
b76c8cfb | 3259 | |
ad4404a2 DB |
3260 | /* |
3261 | * Drop page table lock as buddy allocator may be called. It will | |
3262 | * be acquired again before returning to the caller, as expected. | |
3263 | */ | |
cb900f41 | 3264 | spin_unlock(ptl); |
04f2cbe3 | 3265 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 3266 | |
2fc39cec | 3267 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
3268 | /* |
3269 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
3270 | * it is due to references held by a child and an insufficient | |
3271 | * huge page pool. To guarantee the original mappers | |
3272 | * reliability, unmap the page from child processes. The child | |
3273 | * may get SIGKILLed if it later faults. | |
3274 | */ | |
3275 | if (outside_reserve) { | |
ad4404a2 | 3276 | page_cache_release(old_page); |
04f2cbe3 | 3277 | BUG_ON(huge_pte_none(pte)); |
2f4612af DB |
3278 | unmap_ref_private(mm, vma, old_page, address); |
3279 | BUG_ON(huge_pte_none(pte)); | |
3280 | spin_lock(ptl); | |
3281 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); | |
3282 | if (likely(ptep && | |
3283 | pte_same(huge_ptep_get(ptep), pte))) | |
3284 | goto retry_avoidcopy; | |
3285 | /* | |
3286 | * race occurs while re-acquiring page table | |
3287 | * lock, and our job is done. | |
3288 | */ | |
3289 | return 0; | |
04f2cbe3 MG |
3290 | } |
3291 | ||
ad4404a2 DB |
3292 | ret = (PTR_ERR(new_page) == -ENOMEM) ? |
3293 | VM_FAULT_OOM : VM_FAULT_SIGBUS; | |
3294 | goto out_release_old; | |
1e8f889b DG |
3295 | } |
3296 | ||
0fe6e20b NH |
3297 | /* |
3298 | * When the original hugepage is shared one, it does not have | |
3299 | * anon_vma prepared. | |
3300 | */ | |
44e2aa93 | 3301 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
3302 | ret = VM_FAULT_OOM; |
3303 | goto out_release_all; | |
44e2aa93 | 3304 | } |
0fe6e20b | 3305 | |
47ad8475 AA |
3306 | copy_user_huge_page(new_page, old_page, address, vma, |
3307 | pages_per_huge_page(h)); | |
0ed361de | 3308 | __SetPageUptodate(new_page); |
bcc54222 | 3309 | set_page_huge_active(new_page); |
1e8f889b | 3310 | |
2ec74c3e SG |
3311 | mmun_start = address & huge_page_mask(h); |
3312 | mmun_end = mmun_start + huge_page_size(h); | |
3313 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
ad4404a2 | 3314 | |
b76c8cfb | 3315 | /* |
cb900f41 | 3316 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
3317 | * before the page tables are altered |
3318 | */ | |
cb900f41 | 3319 | spin_lock(ptl); |
a5516438 | 3320 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
a9af0c5d | 3321 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
3322 | ClearPagePrivate(new_page); |
3323 | ||
1e8f889b | 3324 | /* Break COW */ |
8fe627ec | 3325 | huge_ptep_clear_flush(vma, address, ptep); |
34ee645e | 3326 | mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); |
1e8f889b DG |
3327 | set_huge_pte_at(mm, address, ptep, |
3328 | make_huge_pte(vma, new_page, 1)); | |
0fe6e20b | 3329 | page_remove_rmap(old_page); |
cd67f0d2 | 3330 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
3331 | /* Make the old page be freed below */ |
3332 | new_page = old_page; | |
3333 | } | |
cb900f41 | 3334 | spin_unlock(ptl); |
2ec74c3e | 3335 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
ad4404a2 | 3336 | out_release_all: |
1e8f889b | 3337 | page_cache_release(new_page); |
ad4404a2 | 3338 | out_release_old: |
1e8f889b | 3339 | page_cache_release(old_page); |
8312034f | 3340 | |
ad4404a2 DB |
3341 | spin_lock(ptl); /* Caller expects lock to be held */ |
3342 | return ret; | |
1e8f889b DG |
3343 | } |
3344 | ||
04f2cbe3 | 3345 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
3346 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
3347 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
3348 | { |
3349 | struct address_space *mapping; | |
e7c4b0bf | 3350 | pgoff_t idx; |
04f2cbe3 MG |
3351 | |
3352 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 3353 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
3354 | |
3355 | return find_lock_page(mapping, idx); | |
3356 | } | |
3357 | ||
3ae77f43 HD |
3358 | /* |
3359 | * Return whether there is a pagecache page to back given address within VMA. | |
3360 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
3361 | */ | |
3362 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
3363 | struct vm_area_struct *vma, unsigned long address) |
3364 | { | |
3365 | struct address_space *mapping; | |
3366 | pgoff_t idx; | |
3367 | struct page *page; | |
3368 | ||
3369 | mapping = vma->vm_file->f_mapping; | |
3370 | idx = vma_hugecache_offset(h, vma, address); | |
3371 | ||
3372 | page = find_get_page(mapping, idx); | |
3373 | if (page) | |
3374 | put_page(page); | |
3375 | return page != NULL; | |
3376 | } | |
3377 | ||
ab76ad54 MK |
3378 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
3379 | pgoff_t idx) | |
3380 | { | |
3381 | struct inode *inode = mapping->host; | |
3382 | struct hstate *h = hstate_inode(inode); | |
3383 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
3384 | ||
3385 | if (err) | |
3386 | return err; | |
3387 | ClearPagePrivate(page); | |
3388 | ||
3389 | spin_lock(&inode->i_lock); | |
3390 | inode->i_blocks += blocks_per_huge_page(h); | |
3391 | spin_unlock(&inode->i_lock); | |
3392 | return 0; | |
3393 | } | |
3394 | ||
a1ed3dda | 3395 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
8382d914 DB |
3396 | struct address_space *mapping, pgoff_t idx, |
3397 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 3398 | { |
a5516438 | 3399 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 3400 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 3401 | int anon_rmap = 0; |
4c887265 | 3402 | unsigned long size; |
4c887265 | 3403 | struct page *page; |
1e8f889b | 3404 | pte_t new_pte; |
cb900f41 | 3405 | spinlock_t *ptl; |
4c887265 | 3406 | |
04f2cbe3 MG |
3407 | /* |
3408 | * Currently, we are forced to kill the process in the event the | |
3409 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 3410 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
3411 | */ |
3412 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
ffb22af5 AM |
3413 | pr_warning("PID %d killed due to inadequate hugepage pool\n", |
3414 | current->pid); | |
04f2cbe3 MG |
3415 | return ret; |
3416 | } | |
3417 | ||
4c887265 AL |
3418 | /* |
3419 | * Use page lock to guard against racing truncation | |
3420 | * before we get page_table_lock. | |
3421 | */ | |
6bda666a CL |
3422 | retry: |
3423 | page = find_lock_page(mapping, idx); | |
3424 | if (!page) { | |
a5516438 | 3425 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
3426 | if (idx >= size) |
3427 | goto out; | |
04f2cbe3 | 3428 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 3429 | if (IS_ERR(page)) { |
76dcee75 AK |
3430 | ret = PTR_ERR(page); |
3431 | if (ret == -ENOMEM) | |
3432 | ret = VM_FAULT_OOM; | |
3433 | else | |
3434 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
3435 | goto out; |
3436 | } | |
47ad8475 | 3437 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 3438 | __SetPageUptodate(page); |
bcc54222 | 3439 | set_page_huge_active(page); |
ac9b9c66 | 3440 | |
f83a275d | 3441 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 3442 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
3443 | if (err) { |
3444 | put_page(page); | |
6bda666a CL |
3445 | if (err == -EEXIST) |
3446 | goto retry; | |
3447 | goto out; | |
3448 | } | |
23be7468 | 3449 | } else { |
6bda666a | 3450 | lock_page(page); |
0fe6e20b NH |
3451 | if (unlikely(anon_vma_prepare(vma))) { |
3452 | ret = VM_FAULT_OOM; | |
3453 | goto backout_unlocked; | |
3454 | } | |
409eb8c2 | 3455 | anon_rmap = 1; |
23be7468 | 3456 | } |
0fe6e20b | 3457 | } else { |
998b4382 NH |
3458 | /* |
3459 | * If memory error occurs between mmap() and fault, some process | |
3460 | * don't have hwpoisoned swap entry for errored virtual address. | |
3461 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
3462 | */ | |
3463 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 3464 | ret = VM_FAULT_HWPOISON | |
972dc4de | 3465 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
3466 | goto backout_unlocked; |
3467 | } | |
6bda666a | 3468 | } |
1e8f889b | 3469 | |
57303d80 AW |
3470 | /* |
3471 | * If we are going to COW a private mapping later, we examine the | |
3472 | * pending reservations for this page now. This will ensure that | |
3473 | * any allocations necessary to record that reservation occur outside | |
3474 | * the spinlock. | |
3475 | */ | |
5e911373 | 3476 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
2b26736c AW |
3477 | if (vma_needs_reservation(h, vma, address) < 0) { |
3478 | ret = VM_FAULT_OOM; | |
3479 | goto backout_unlocked; | |
3480 | } | |
5e911373 | 3481 | /* Just decrements count, does not deallocate */ |
feba16e2 | 3482 | vma_end_reservation(h, vma, address); |
5e911373 | 3483 | } |
57303d80 | 3484 | |
cb900f41 KS |
3485 | ptl = huge_pte_lockptr(h, mm, ptep); |
3486 | spin_lock(ptl); | |
a5516438 | 3487 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
3488 | if (idx >= size) |
3489 | goto backout; | |
3490 | ||
83c54070 | 3491 | ret = 0; |
7f2e9525 | 3492 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
3493 | goto backout; |
3494 | ||
07443a85 JK |
3495 | if (anon_rmap) { |
3496 | ClearPagePrivate(page); | |
409eb8c2 | 3497 | hugepage_add_new_anon_rmap(page, vma, address); |
ac714904 | 3498 | } else |
409eb8c2 | 3499 | page_dup_rmap(page); |
1e8f889b DG |
3500 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
3501 | && (vma->vm_flags & VM_SHARED))); | |
3502 | set_huge_pte_at(mm, address, ptep, new_pte); | |
3503 | ||
788c7df4 | 3504 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 3505 | /* Optimization, do the COW without a second fault */ |
cb900f41 | 3506 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); |
1e8f889b DG |
3507 | } |
3508 | ||
cb900f41 | 3509 | spin_unlock(ptl); |
4c887265 AL |
3510 | unlock_page(page); |
3511 | out: | |
ac9b9c66 | 3512 | return ret; |
4c887265 AL |
3513 | |
3514 | backout: | |
cb900f41 | 3515 | spin_unlock(ptl); |
2b26736c | 3516 | backout_unlocked: |
4c887265 AL |
3517 | unlock_page(page); |
3518 | put_page(page); | |
3519 | goto out; | |
ac9b9c66 HD |
3520 | } |
3521 | ||
8382d914 | 3522 | #ifdef CONFIG_SMP |
c672c7f2 | 3523 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
8382d914 DB |
3524 | struct vm_area_struct *vma, |
3525 | struct address_space *mapping, | |
3526 | pgoff_t idx, unsigned long address) | |
3527 | { | |
3528 | unsigned long key[2]; | |
3529 | u32 hash; | |
3530 | ||
3531 | if (vma->vm_flags & VM_SHARED) { | |
3532 | key[0] = (unsigned long) mapping; | |
3533 | key[1] = idx; | |
3534 | } else { | |
3535 | key[0] = (unsigned long) mm; | |
3536 | key[1] = address >> huge_page_shift(h); | |
3537 | } | |
3538 | ||
3539 | hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); | |
3540 | ||
3541 | return hash & (num_fault_mutexes - 1); | |
3542 | } | |
3543 | #else | |
3544 | /* | |
3545 | * For uniprocesor systems we always use a single mutex, so just | |
3546 | * return 0 and avoid the hashing overhead. | |
3547 | */ | |
c672c7f2 | 3548 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
8382d914 DB |
3549 | struct vm_area_struct *vma, |
3550 | struct address_space *mapping, | |
3551 | pgoff_t idx, unsigned long address) | |
3552 | { | |
3553 | return 0; | |
3554 | } | |
3555 | #endif | |
3556 | ||
86e5216f | 3557 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 3558 | unsigned long address, unsigned int flags) |
86e5216f | 3559 | { |
8382d914 | 3560 | pte_t *ptep, entry; |
cb900f41 | 3561 | spinlock_t *ptl; |
1e8f889b | 3562 | int ret; |
8382d914 DB |
3563 | u32 hash; |
3564 | pgoff_t idx; | |
0fe6e20b | 3565 | struct page *page = NULL; |
57303d80 | 3566 | struct page *pagecache_page = NULL; |
a5516438 | 3567 | struct hstate *h = hstate_vma(vma); |
8382d914 | 3568 | struct address_space *mapping; |
0f792cf9 | 3569 | int need_wait_lock = 0; |
86e5216f | 3570 | |
1e16a539 KH |
3571 | address &= huge_page_mask(h); |
3572 | ||
fd6a03ed NH |
3573 | ptep = huge_pte_offset(mm, address); |
3574 | if (ptep) { | |
3575 | entry = huge_ptep_get(ptep); | |
290408d4 | 3576 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 3577 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
3578 | return 0; |
3579 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 3580 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 3581 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
3582 | } |
3583 | ||
a5516438 | 3584 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); |
86e5216f AL |
3585 | if (!ptep) |
3586 | return VM_FAULT_OOM; | |
3587 | ||
8382d914 DB |
3588 | mapping = vma->vm_file->f_mapping; |
3589 | idx = vma_hugecache_offset(h, vma, address); | |
3590 | ||
3935baa9 DG |
3591 | /* |
3592 | * Serialize hugepage allocation and instantiation, so that we don't | |
3593 | * get spurious allocation failures if two CPUs race to instantiate | |
3594 | * the same page in the page cache. | |
3595 | */ | |
c672c7f2 MK |
3596 | hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address); |
3597 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
8382d914 | 3598 | |
7f2e9525 GS |
3599 | entry = huge_ptep_get(ptep); |
3600 | if (huge_pte_none(entry)) { | |
8382d914 | 3601 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 3602 | goto out_mutex; |
3935baa9 | 3603 | } |
86e5216f | 3604 | |
83c54070 | 3605 | ret = 0; |
1e8f889b | 3606 | |
0f792cf9 NH |
3607 | /* |
3608 | * entry could be a migration/hwpoison entry at this point, so this | |
3609 | * check prevents the kernel from going below assuming that we have | |
3610 | * a active hugepage in pagecache. This goto expects the 2nd page fault, | |
3611 | * and is_hugetlb_entry_(migration|hwpoisoned) check will properly | |
3612 | * handle it. | |
3613 | */ | |
3614 | if (!pte_present(entry)) | |
3615 | goto out_mutex; | |
3616 | ||
57303d80 AW |
3617 | /* |
3618 | * If we are going to COW the mapping later, we examine the pending | |
3619 | * reservations for this page now. This will ensure that any | |
3620 | * allocations necessary to record that reservation occur outside the | |
3621 | * spinlock. For private mappings, we also lookup the pagecache | |
3622 | * page now as it is used to determine if a reservation has been | |
3623 | * consumed. | |
3624 | */ | |
106c992a | 3625 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
2b26736c AW |
3626 | if (vma_needs_reservation(h, vma, address) < 0) { |
3627 | ret = VM_FAULT_OOM; | |
b4d1d99f | 3628 | goto out_mutex; |
2b26736c | 3629 | } |
5e911373 | 3630 | /* Just decrements count, does not deallocate */ |
feba16e2 | 3631 | vma_end_reservation(h, vma, address); |
57303d80 | 3632 | |
f83a275d | 3633 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
3634 | pagecache_page = hugetlbfs_pagecache_page(h, |
3635 | vma, address); | |
3636 | } | |
3637 | ||
0f792cf9 NH |
3638 | ptl = huge_pte_lock(h, mm, ptep); |
3639 | ||
3640 | /* Check for a racing update before calling hugetlb_cow */ | |
3641 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
3642 | goto out_ptl; | |
3643 | ||
56c9cfb1 NH |
3644 | /* |
3645 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
3646 | * pagecache_page, so here we need take the former one | |
3647 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
3648 | */ |
3649 | page = pte_page(entry); | |
3650 | if (page != pagecache_page) | |
0f792cf9 NH |
3651 | if (!trylock_page(page)) { |
3652 | need_wait_lock = 1; | |
3653 | goto out_ptl; | |
3654 | } | |
b4d1d99f | 3655 | |
0f792cf9 | 3656 | get_page(page); |
b4d1d99f | 3657 | |
788c7df4 | 3658 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 3659 | if (!huge_pte_write(entry)) { |
57303d80 | 3660 | ret = hugetlb_cow(mm, vma, address, ptep, entry, |
cb900f41 | 3661 | pagecache_page, ptl); |
0f792cf9 | 3662 | goto out_put_page; |
b4d1d99f | 3663 | } |
106c992a | 3664 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
3665 | } |
3666 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
3667 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
3668 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 3669 | update_mmu_cache(vma, address, ptep); |
0f792cf9 NH |
3670 | out_put_page: |
3671 | if (page != pagecache_page) | |
3672 | unlock_page(page); | |
3673 | put_page(page); | |
cb900f41 KS |
3674 | out_ptl: |
3675 | spin_unlock(ptl); | |
57303d80 AW |
3676 | |
3677 | if (pagecache_page) { | |
3678 | unlock_page(pagecache_page); | |
3679 | put_page(pagecache_page); | |
3680 | } | |
b4d1d99f | 3681 | out_mutex: |
c672c7f2 | 3682 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
0f792cf9 NH |
3683 | /* |
3684 | * Generally it's safe to hold refcount during waiting page lock. But | |
3685 | * here we just wait to defer the next page fault to avoid busy loop and | |
3686 | * the page is not used after unlocked before returning from the current | |
3687 | * page fault. So we are safe from accessing freed page, even if we wait | |
3688 | * here without taking refcount. | |
3689 | */ | |
3690 | if (need_wait_lock) | |
3691 | wait_on_page_locked(page); | |
1e8f889b | 3692 | return ret; |
86e5216f AL |
3693 | } |
3694 | ||
28a35716 ML |
3695 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
3696 | struct page **pages, struct vm_area_struct **vmas, | |
3697 | unsigned long *position, unsigned long *nr_pages, | |
3698 | long i, unsigned int flags) | |
63551ae0 | 3699 | { |
d5d4b0aa CK |
3700 | unsigned long pfn_offset; |
3701 | unsigned long vaddr = *position; | |
28a35716 | 3702 | unsigned long remainder = *nr_pages; |
a5516438 | 3703 | struct hstate *h = hstate_vma(vma); |
63551ae0 | 3704 | |
63551ae0 | 3705 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 3706 | pte_t *pte; |
cb900f41 | 3707 | spinlock_t *ptl = NULL; |
2a15efc9 | 3708 | int absent; |
4c887265 | 3709 | struct page *page; |
63551ae0 | 3710 | |
02057967 DR |
3711 | /* |
3712 | * If we have a pending SIGKILL, don't keep faulting pages and | |
3713 | * potentially allocating memory. | |
3714 | */ | |
3715 | if (unlikely(fatal_signal_pending(current))) { | |
3716 | remainder = 0; | |
3717 | break; | |
3718 | } | |
3719 | ||
4c887265 AL |
3720 | /* |
3721 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 3722 | * each hugepage. We have to make sure we get the |
4c887265 | 3723 | * first, for the page indexing below to work. |
cb900f41 KS |
3724 | * |
3725 | * Note that page table lock is not held when pte is null. | |
4c887265 | 3726 | */ |
a5516438 | 3727 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); |
cb900f41 KS |
3728 | if (pte) |
3729 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
3730 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
3731 | ||
3732 | /* | |
3733 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
3734 | * an error where there's an empty slot with no huge pagecache |
3735 | * to back it. This way, we avoid allocating a hugepage, and | |
3736 | * the sparse dumpfile avoids allocating disk blocks, but its | |
3737 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 3738 | */ |
3ae77f43 HD |
3739 | if (absent && (flags & FOLL_DUMP) && |
3740 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
3741 | if (pte) |
3742 | spin_unlock(ptl); | |
2a15efc9 HD |
3743 | remainder = 0; |
3744 | break; | |
3745 | } | |
63551ae0 | 3746 | |
9cc3a5bd NH |
3747 | /* |
3748 | * We need call hugetlb_fault for both hugepages under migration | |
3749 | * (in which case hugetlb_fault waits for the migration,) and | |
3750 | * hwpoisoned hugepages (in which case we need to prevent the | |
3751 | * caller from accessing to them.) In order to do this, we use | |
3752 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
3753 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
3754 | * both cases, and because we can't follow correct pages | |
3755 | * directly from any kind of swap entries. | |
3756 | */ | |
3757 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
3758 | ((flags & FOLL_WRITE) && |
3759 | !huge_pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 3760 | int ret; |
63551ae0 | 3761 | |
cb900f41 KS |
3762 | if (pte) |
3763 | spin_unlock(ptl); | |
2a15efc9 HD |
3764 | ret = hugetlb_fault(mm, vma, vaddr, |
3765 | (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); | |
a89182c7 | 3766 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 3767 | continue; |
63551ae0 | 3768 | |
4c887265 | 3769 | remainder = 0; |
4c887265 AL |
3770 | break; |
3771 | } | |
3772 | ||
a5516438 | 3773 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 3774 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 3775 | same_page: |
d6692183 | 3776 | if (pages) { |
2a15efc9 | 3777 | pages[i] = mem_map_offset(page, pfn_offset); |
a0368d4e | 3778 | get_page_foll(pages[i]); |
d6692183 | 3779 | } |
63551ae0 DG |
3780 | |
3781 | if (vmas) | |
3782 | vmas[i] = vma; | |
3783 | ||
3784 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 3785 | ++pfn_offset; |
63551ae0 DG |
3786 | --remainder; |
3787 | ++i; | |
d5d4b0aa | 3788 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 3789 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa CK |
3790 | /* |
3791 | * We use pfn_offset to avoid touching the pageframes | |
3792 | * of this compound page. | |
3793 | */ | |
3794 | goto same_page; | |
3795 | } | |
cb900f41 | 3796 | spin_unlock(ptl); |
63551ae0 | 3797 | } |
28a35716 | 3798 | *nr_pages = remainder; |
63551ae0 DG |
3799 | *position = vaddr; |
3800 | ||
2a15efc9 | 3801 | return i ? i : -EFAULT; |
63551ae0 | 3802 | } |
8f860591 | 3803 | |
7da4d641 | 3804 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
3805 | unsigned long address, unsigned long end, pgprot_t newprot) |
3806 | { | |
3807 | struct mm_struct *mm = vma->vm_mm; | |
3808 | unsigned long start = address; | |
3809 | pte_t *ptep; | |
3810 | pte_t pte; | |
a5516438 | 3811 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 3812 | unsigned long pages = 0; |
8f860591 ZY |
3813 | |
3814 | BUG_ON(address >= end); | |
3815 | flush_cache_range(vma, address, end); | |
3816 | ||
a5338093 | 3817 | mmu_notifier_invalidate_range_start(mm, start, end); |
83cde9e8 | 3818 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 3819 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 3820 | spinlock_t *ptl; |
8f860591 ZY |
3821 | ptep = huge_pte_offset(mm, address); |
3822 | if (!ptep) | |
3823 | continue; | |
cb900f41 | 3824 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
3825 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3826 | pages++; | |
cb900f41 | 3827 | spin_unlock(ptl); |
39dde65c | 3828 | continue; |
7da4d641 | 3829 | } |
a8bda28d NH |
3830 | pte = huge_ptep_get(ptep); |
3831 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
3832 | spin_unlock(ptl); | |
3833 | continue; | |
3834 | } | |
3835 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
3836 | swp_entry_t entry = pte_to_swp_entry(pte); | |
3837 | ||
3838 | if (is_write_migration_entry(entry)) { | |
3839 | pte_t newpte; | |
3840 | ||
3841 | make_migration_entry_read(&entry); | |
3842 | newpte = swp_entry_to_pte(entry); | |
3843 | set_huge_pte_at(mm, address, ptep, newpte); | |
3844 | pages++; | |
3845 | } | |
3846 | spin_unlock(ptl); | |
3847 | continue; | |
3848 | } | |
3849 | if (!huge_pte_none(pte)) { | |
8f860591 | 3850 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
106c992a | 3851 | pte = pte_mkhuge(huge_pte_modify(pte, newprot)); |
be7517d6 | 3852 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
8f860591 | 3853 | set_huge_pte_at(mm, address, ptep, pte); |
7da4d641 | 3854 | pages++; |
8f860591 | 3855 | } |
cb900f41 | 3856 | spin_unlock(ptl); |
8f860591 | 3857 | } |
d833352a | 3858 | /* |
c8c06efa | 3859 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 3860 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 3861 | * once we release i_mmap_rwsem, another task can do the final put_page |
d833352a MG |
3862 | * and that page table be reused and filled with junk. |
3863 | */ | |
8f860591 | 3864 | flush_tlb_range(vma, start, end); |
34ee645e | 3865 | mmu_notifier_invalidate_range(mm, start, end); |
83cde9e8 | 3866 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
a5338093 | 3867 | mmu_notifier_invalidate_range_end(mm, start, end); |
7da4d641 PZ |
3868 | |
3869 | return pages << h->order; | |
8f860591 ZY |
3870 | } |
3871 | ||
a1e78772 MG |
3872 | int hugetlb_reserve_pages(struct inode *inode, |
3873 | long from, long to, | |
5a6fe125 | 3874 | struct vm_area_struct *vma, |
ca16d140 | 3875 | vm_flags_t vm_flags) |
e4e574b7 | 3876 | { |
17c9d12e | 3877 | long ret, chg; |
a5516438 | 3878 | struct hstate *h = hstate_inode(inode); |
90481622 | 3879 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 3880 | struct resv_map *resv_map; |
1c5ecae3 | 3881 | long gbl_reserve; |
e4e574b7 | 3882 | |
17c9d12e MG |
3883 | /* |
3884 | * Only apply hugepage reservation if asked. At fault time, an | |
3885 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 3886 | * without using reserves |
17c9d12e | 3887 | */ |
ca16d140 | 3888 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
3889 | return 0; |
3890 | ||
a1e78772 MG |
3891 | /* |
3892 | * Shared mappings base their reservation on the number of pages that | |
3893 | * are already allocated on behalf of the file. Private mappings need | |
3894 | * to reserve the full area even if read-only as mprotect() may be | |
3895 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
3896 | */ | |
9119a41e | 3897 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4e35f483 | 3898 | resv_map = inode_resv_map(inode); |
9119a41e | 3899 | |
1406ec9b | 3900 | chg = region_chg(resv_map, from, to); |
9119a41e JK |
3901 | |
3902 | } else { | |
3903 | resv_map = resv_map_alloc(); | |
17c9d12e MG |
3904 | if (!resv_map) |
3905 | return -ENOMEM; | |
3906 | ||
a1e78772 | 3907 | chg = to - from; |
84afd99b | 3908 | |
17c9d12e MG |
3909 | set_vma_resv_map(vma, resv_map); |
3910 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
3911 | } | |
3912 | ||
c50ac050 DH |
3913 | if (chg < 0) { |
3914 | ret = chg; | |
3915 | goto out_err; | |
3916 | } | |
8a630112 | 3917 | |
1c5ecae3 MK |
3918 | /* |
3919 | * There must be enough pages in the subpool for the mapping. If | |
3920 | * the subpool has a minimum size, there may be some global | |
3921 | * reservations already in place (gbl_reserve). | |
3922 | */ | |
3923 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
3924 | if (gbl_reserve < 0) { | |
c50ac050 DH |
3925 | ret = -ENOSPC; |
3926 | goto out_err; | |
3927 | } | |
5a6fe125 MG |
3928 | |
3929 | /* | |
17c9d12e | 3930 | * Check enough hugepages are available for the reservation. |
90481622 | 3931 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 3932 | */ |
1c5ecae3 | 3933 | ret = hugetlb_acct_memory(h, gbl_reserve); |
68842c9b | 3934 | if (ret < 0) { |
1c5ecae3 MK |
3935 | /* put back original number of pages, chg */ |
3936 | (void)hugepage_subpool_put_pages(spool, chg); | |
c50ac050 | 3937 | goto out_err; |
68842c9b | 3938 | } |
17c9d12e MG |
3939 | |
3940 | /* | |
3941 | * Account for the reservations made. Shared mappings record regions | |
3942 | * that have reservations as they are shared by multiple VMAs. | |
3943 | * When the last VMA disappears, the region map says how much | |
3944 | * the reservation was and the page cache tells how much of | |
3945 | * the reservation was consumed. Private mappings are per-VMA and | |
3946 | * only the consumed reservations are tracked. When the VMA | |
3947 | * disappears, the original reservation is the VMA size and the | |
3948 | * consumed reservations are stored in the map. Hence, nothing | |
3949 | * else has to be done for private mappings here | |
3950 | */ | |
33039678 MK |
3951 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
3952 | long add = region_add(resv_map, from, to); | |
3953 | ||
3954 | if (unlikely(chg > add)) { | |
3955 | /* | |
3956 | * pages in this range were added to the reserve | |
3957 | * map between region_chg and region_add. This | |
3958 | * indicates a race with alloc_huge_page. Adjust | |
3959 | * the subpool and reserve counts modified above | |
3960 | * based on the difference. | |
3961 | */ | |
3962 | long rsv_adjust; | |
3963 | ||
3964 | rsv_adjust = hugepage_subpool_put_pages(spool, | |
3965 | chg - add); | |
3966 | hugetlb_acct_memory(h, -rsv_adjust); | |
3967 | } | |
3968 | } | |
a43a8c39 | 3969 | return 0; |
c50ac050 | 3970 | out_err: |
5e911373 MK |
3971 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
3972 | region_abort(resv_map, from, to); | |
f031dd27 JK |
3973 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3974 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 3975 | return ret; |
a43a8c39 CK |
3976 | } |
3977 | ||
b5cec28d MK |
3978 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
3979 | long freed) | |
a43a8c39 | 3980 | { |
a5516438 | 3981 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 3982 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 3983 | long chg = 0; |
90481622 | 3984 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 3985 | long gbl_reserve; |
45c682a6 | 3986 | |
b5cec28d MK |
3987 | if (resv_map) { |
3988 | chg = region_del(resv_map, start, end); | |
3989 | /* | |
3990 | * region_del() can fail in the rare case where a region | |
3991 | * must be split and another region descriptor can not be | |
3992 | * allocated. If end == LONG_MAX, it will not fail. | |
3993 | */ | |
3994 | if (chg < 0) | |
3995 | return chg; | |
3996 | } | |
3997 | ||
45c682a6 | 3998 | spin_lock(&inode->i_lock); |
e4c6f8be | 3999 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
4000 | spin_unlock(&inode->i_lock); |
4001 | ||
1c5ecae3 MK |
4002 | /* |
4003 | * If the subpool has a minimum size, the number of global | |
4004 | * reservations to be released may be adjusted. | |
4005 | */ | |
4006 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
4007 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
4008 | |
4009 | return 0; | |
a43a8c39 | 4010 | } |
93f70f90 | 4011 | |
3212b535 SC |
4012 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
4013 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
4014 | struct vm_area_struct *vma, | |
4015 | unsigned long addr, pgoff_t idx) | |
4016 | { | |
4017 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
4018 | svma->vm_start; | |
4019 | unsigned long sbase = saddr & PUD_MASK; | |
4020 | unsigned long s_end = sbase + PUD_SIZE; | |
4021 | ||
4022 | /* Allow segments to share if only one is marked locked */ | |
4023 | unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED; | |
4024 | unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED; | |
4025 | ||
4026 | /* | |
4027 | * match the virtual addresses, permission and the alignment of the | |
4028 | * page table page. | |
4029 | */ | |
4030 | if (pmd_index(addr) != pmd_index(saddr) || | |
4031 | vm_flags != svm_flags || | |
4032 | sbase < svma->vm_start || svma->vm_end < s_end) | |
4033 | return 0; | |
4034 | ||
4035 | return saddr; | |
4036 | } | |
4037 | ||
31aafb45 | 4038 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
4039 | { |
4040 | unsigned long base = addr & PUD_MASK; | |
4041 | unsigned long end = base + PUD_SIZE; | |
4042 | ||
4043 | /* | |
4044 | * check on proper vm_flags and page table alignment | |
4045 | */ | |
4046 | if (vma->vm_flags & VM_MAYSHARE && | |
4047 | vma->vm_start <= base && end <= vma->vm_end) | |
31aafb45 NK |
4048 | return true; |
4049 | return false; | |
3212b535 SC |
4050 | } |
4051 | ||
4052 | /* | |
4053 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
4054 | * and returns the corresponding pte. While this is not necessary for the | |
4055 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
4056 | * code much cleaner. pmd allocation is essential for the shared case because | |
c8c06efa | 4057 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise |
3212b535 SC |
4058 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a |
4059 | * bad pmd for sharing. | |
4060 | */ | |
4061 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4062 | { | |
4063 | struct vm_area_struct *vma = find_vma(mm, addr); | |
4064 | struct address_space *mapping = vma->vm_file->f_mapping; | |
4065 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
4066 | vma->vm_pgoff; | |
4067 | struct vm_area_struct *svma; | |
4068 | unsigned long saddr; | |
4069 | pte_t *spte = NULL; | |
4070 | pte_t *pte; | |
cb900f41 | 4071 | spinlock_t *ptl; |
3212b535 SC |
4072 | |
4073 | if (!vma_shareable(vma, addr)) | |
4074 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
4075 | ||
83cde9e8 | 4076 | i_mmap_lock_write(mapping); |
3212b535 SC |
4077 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
4078 | if (svma == vma) | |
4079 | continue; | |
4080 | ||
4081 | saddr = page_table_shareable(svma, vma, addr, idx); | |
4082 | if (saddr) { | |
4083 | spte = huge_pte_offset(svma->vm_mm, saddr); | |
4084 | if (spte) { | |
dc6c9a35 | 4085 | mm_inc_nr_pmds(mm); |
3212b535 SC |
4086 | get_page(virt_to_page(spte)); |
4087 | break; | |
4088 | } | |
4089 | } | |
4090 | } | |
4091 | ||
4092 | if (!spte) | |
4093 | goto out; | |
4094 | ||
cb900f41 KS |
4095 | ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); |
4096 | spin_lock(ptl); | |
dc6c9a35 | 4097 | if (pud_none(*pud)) { |
3212b535 SC |
4098 | pud_populate(mm, pud, |
4099 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
dc6c9a35 | 4100 | } else { |
3212b535 | 4101 | put_page(virt_to_page(spte)); |
dc6c9a35 KS |
4102 | mm_inc_nr_pmds(mm); |
4103 | } | |
cb900f41 | 4104 | spin_unlock(ptl); |
3212b535 SC |
4105 | out: |
4106 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
83cde9e8 | 4107 | i_mmap_unlock_write(mapping); |
3212b535 SC |
4108 | return pte; |
4109 | } | |
4110 | ||
4111 | /* | |
4112 | * unmap huge page backed by shared pte. | |
4113 | * | |
4114 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
4115 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
4116 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
4117 | * | |
cb900f41 | 4118 | * called with page table lock held. |
3212b535 SC |
4119 | * |
4120 | * returns: 1 successfully unmapped a shared pte page | |
4121 | * 0 the underlying pte page is not shared, or it is the last user | |
4122 | */ | |
4123 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4124 | { | |
4125 | pgd_t *pgd = pgd_offset(mm, *addr); | |
4126 | pud_t *pud = pud_offset(pgd, *addr); | |
4127 | ||
4128 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
4129 | if (page_count(virt_to_page(ptep)) == 1) | |
4130 | return 0; | |
4131 | ||
4132 | pud_clear(pud); | |
4133 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 4134 | mm_dec_nr_pmds(mm); |
3212b535 SC |
4135 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
4136 | return 1; | |
4137 | } | |
9e5fc74c SC |
4138 | #define want_pmd_share() (1) |
4139 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
4140 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4141 | { | |
4142 | return NULL; | |
4143 | } | |
e81f2d22 ZZ |
4144 | |
4145 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4146 | { | |
4147 | return 0; | |
4148 | } | |
9e5fc74c | 4149 | #define want_pmd_share() (0) |
3212b535 SC |
4150 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
4151 | ||
9e5fc74c SC |
4152 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
4153 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
4154 | unsigned long addr, unsigned long sz) | |
4155 | { | |
4156 | pgd_t *pgd; | |
4157 | pud_t *pud; | |
4158 | pte_t *pte = NULL; | |
4159 | ||
4160 | pgd = pgd_offset(mm, addr); | |
4161 | pud = pud_alloc(mm, pgd, addr); | |
4162 | if (pud) { | |
4163 | if (sz == PUD_SIZE) { | |
4164 | pte = (pte_t *)pud; | |
4165 | } else { | |
4166 | BUG_ON(sz != PMD_SIZE); | |
4167 | if (want_pmd_share() && pud_none(*pud)) | |
4168 | pte = huge_pmd_share(mm, addr, pud); | |
4169 | else | |
4170 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
4171 | } | |
4172 | } | |
4173 | BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); | |
4174 | ||
4175 | return pte; | |
4176 | } | |
4177 | ||
4178 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) | |
4179 | { | |
4180 | pgd_t *pgd; | |
4181 | pud_t *pud; | |
4182 | pmd_t *pmd = NULL; | |
4183 | ||
4184 | pgd = pgd_offset(mm, addr); | |
4185 | if (pgd_present(*pgd)) { | |
4186 | pud = pud_offset(pgd, addr); | |
4187 | if (pud_present(*pud)) { | |
4188 | if (pud_huge(*pud)) | |
4189 | return (pte_t *)pud; | |
4190 | pmd = pmd_offset(pud, addr); | |
4191 | } | |
4192 | } | |
4193 | return (pte_t *) pmd; | |
4194 | } | |
4195 | ||
61f77eda NH |
4196 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
4197 | ||
4198 | /* | |
4199 | * These functions are overwritable if your architecture needs its own | |
4200 | * behavior. | |
4201 | */ | |
4202 | struct page * __weak | |
4203 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
4204 | int write) | |
4205 | { | |
4206 | return ERR_PTR(-EINVAL); | |
4207 | } | |
4208 | ||
4209 | struct page * __weak | |
9e5fc74c | 4210 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 4211 | pmd_t *pmd, int flags) |
9e5fc74c | 4212 | { |
e66f17ff NH |
4213 | struct page *page = NULL; |
4214 | spinlock_t *ptl; | |
4215 | retry: | |
4216 | ptl = pmd_lockptr(mm, pmd); | |
4217 | spin_lock(ptl); | |
4218 | /* | |
4219 | * make sure that the address range covered by this pmd is not | |
4220 | * unmapped from other threads. | |
4221 | */ | |
4222 | if (!pmd_huge(*pmd)) | |
4223 | goto out; | |
4224 | if (pmd_present(*pmd)) { | |
97534127 | 4225 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
e66f17ff NH |
4226 | if (flags & FOLL_GET) |
4227 | get_page(page); | |
4228 | } else { | |
4229 | if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) { | |
4230 | spin_unlock(ptl); | |
4231 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
4232 | goto retry; | |
4233 | } | |
4234 | /* | |
4235 | * hwpoisoned entry is treated as no_page_table in | |
4236 | * follow_page_mask(). | |
4237 | */ | |
4238 | } | |
4239 | out: | |
4240 | spin_unlock(ptl); | |
9e5fc74c SC |
4241 | return page; |
4242 | } | |
4243 | ||
61f77eda | 4244 | struct page * __weak |
9e5fc74c | 4245 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 4246 | pud_t *pud, int flags) |
9e5fc74c | 4247 | { |
e66f17ff NH |
4248 | if (flags & FOLL_GET) |
4249 | return NULL; | |
9e5fc74c | 4250 | |
e66f17ff | 4251 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
4252 | } |
4253 | ||
d5bd9106 AK |
4254 | #ifdef CONFIG_MEMORY_FAILURE |
4255 | ||
93f70f90 NH |
4256 | /* |
4257 | * This function is called from memory failure code. | |
4258 | * Assume the caller holds page lock of the head page. | |
4259 | */ | |
6de2b1aa | 4260 | int dequeue_hwpoisoned_huge_page(struct page *hpage) |
93f70f90 NH |
4261 | { |
4262 | struct hstate *h = page_hstate(hpage); | |
4263 | int nid = page_to_nid(hpage); | |
6de2b1aa | 4264 | int ret = -EBUSY; |
93f70f90 NH |
4265 | |
4266 | spin_lock(&hugetlb_lock); | |
7e1f049e NH |
4267 | /* |
4268 | * Just checking !page_huge_active is not enough, because that could be | |
4269 | * an isolated/hwpoisoned hugepage (which have >0 refcount). | |
4270 | */ | |
4271 | if (!page_huge_active(hpage) && !page_count(hpage)) { | |
56f2fb14 NH |
4272 | /* |
4273 | * Hwpoisoned hugepage isn't linked to activelist or freelist, | |
4274 | * but dangling hpage->lru can trigger list-debug warnings | |
4275 | * (this happens when we call unpoison_memory() on it), | |
4276 | * so let it point to itself with list_del_init(). | |
4277 | */ | |
4278 | list_del_init(&hpage->lru); | |
8c6c2ecb | 4279 | set_page_refcounted(hpage); |
6de2b1aa NH |
4280 | h->free_huge_pages--; |
4281 | h->free_huge_pages_node[nid]--; | |
4282 | ret = 0; | |
4283 | } | |
93f70f90 | 4284 | spin_unlock(&hugetlb_lock); |
6de2b1aa | 4285 | return ret; |
93f70f90 | 4286 | } |
6de2b1aa | 4287 | #endif |
31caf665 NH |
4288 | |
4289 | bool isolate_huge_page(struct page *page, struct list_head *list) | |
4290 | { | |
bcc54222 NH |
4291 | bool ret = true; |
4292 | ||
309381fe | 4293 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4294 | spin_lock(&hugetlb_lock); |
bcc54222 NH |
4295 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { |
4296 | ret = false; | |
4297 | goto unlock; | |
4298 | } | |
4299 | clear_page_huge_active(page); | |
31caf665 | 4300 | list_move_tail(&page->lru, list); |
bcc54222 | 4301 | unlock: |
31caf665 | 4302 | spin_unlock(&hugetlb_lock); |
bcc54222 | 4303 | return ret; |
31caf665 NH |
4304 | } |
4305 | ||
4306 | void putback_active_hugepage(struct page *page) | |
4307 | { | |
309381fe | 4308 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4309 | spin_lock(&hugetlb_lock); |
bcc54222 | 4310 | set_page_huge_active(page); |
31caf665 NH |
4311 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
4312 | spin_unlock(&hugetlb_lock); | |
4313 | put_page(page); | |
4314 | } |