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