Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * Generic hugetlb support. | |
6d49e352 | 3 | * (C) Nadia Yvette Chambers, April 2004 |
1da177e4 | 4 | */ |
1da177e4 LT |
5 | #include <linux/list.h> |
6 | #include <linux/init.h> | |
1da177e4 | 7 | #include <linux/mm.h> |
e1759c21 | 8 | #include <linux/seq_file.h> |
1da177e4 LT |
9 | #include <linux/sysctl.h> |
10 | #include <linux/highmem.h> | |
cddb8a5c | 11 | #include <linux/mmu_notifier.h> |
1da177e4 | 12 | #include <linux/nodemask.h> |
63551ae0 | 13 | #include <linux/pagemap.h> |
5da7ca86 | 14 | #include <linux/mempolicy.h> |
3b32123d | 15 | #include <linux/compiler.h> |
aea47ff3 | 16 | #include <linux/cpuset.h> |
3935baa9 | 17 | #include <linux/mutex.h> |
aa888a74 | 18 | #include <linux/bootmem.h> |
a3437870 | 19 | #include <linux/sysfs.h> |
5a0e3ad6 | 20 | #include <linux/slab.h> |
0fe6e20b | 21 | #include <linux/rmap.h> |
fd6a03ed NH |
22 | #include <linux/swap.h> |
23 | #include <linux/swapops.h> | |
c8721bbb | 24 | #include <linux/page-isolation.h> |
8382d914 | 25 | #include <linux/jhash.h> |
d6606683 | 26 | |
63551ae0 DG |
27 | #include <asm/page.h> |
28 | #include <asm/pgtable.h> | |
24669e58 | 29 | #include <asm/tlb.h> |
63551ae0 | 30 | |
24669e58 | 31 | #include <linux/io.h> |
63551ae0 | 32 | #include <linux/hugetlb.h> |
9dd540e2 | 33 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 34 | #include <linux/node.h> |
7835e98b | 35 | #include "internal.h" |
1da177e4 | 36 | |
753162cd | 37 | int hugepages_treat_as_movable; |
a5516438 | 38 | |
c3f38a38 | 39 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
40 | unsigned int default_hstate_idx; |
41 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
641844f5 NH |
42 | /* |
43 | * Minimum page order among possible hugepage sizes, set to a proper value | |
44 | * at boot time. | |
45 | */ | |
46 | static unsigned int minimum_order __read_mostly = UINT_MAX; | |
e5ff2159 | 47 | |
53ba51d2 JT |
48 | __initdata LIST_HEAD(huge_boot_pages); |
49 | ||
e5ff2159 AK |
50 | /* for command line parsing */ |
51 | static struct hstate * __initdata parsed_hstate; | |
52 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 53 | static unsigned long __initdata default_hstate_size; |
9fee021d | 54 | static bool __initdata parsed_valid_hugepagesz = true; |
e5ff2159 | 55 | |
3935baa9 | 56 | /* |
31caf665 NH |
57 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
58 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 59 | */ |
c3f38a38 | 60 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 61 | |
8382d914 DB |
62 | /* |
63 | * Serializes faults on the same logical page. This is used to | |
64 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
65 | */ | |
66 | static int num_fault_mutexes; | |
c672c7f2 | 67 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
8382d914 | 68 | |
7ca02d0a MK |
69 | /* Forward declaration */ |
70 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
71 | ||
90481622 DG |
72 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
73 | { | |
74 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
75 | ||
76 | spin_unlock(&spool->lock); | |
77 | ||
78 | /* If no pages are used, and no other handles to the subpool | |
7ca02d0a MK |
79 | * remain, give up any reservations mased on minimum size and |
80 | * free the subpool */ | |
81 | if (free) { | |
82 | if (spool->min_hpages != -1) | |
83 | hugetlb_acct_memory(spool->hstate, | |
84 | -spool->min_hpages); | |
90481622 | 85 | kfree(spool); |
7ca02d0a | 86 | } |
90481622 DG |
87 | } |
88 | ||
7ca02d0a MK |
89 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
90 | long min_hpages) | |
90481622 DG |
91 | { |
92 | struct hugepage_subpool *spool; | |
93 | ||
c6a91820 | 94 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
95 | if (!spool) |
96 | return NULL; | |
97 | ||
98 | spin_lock_init(&spool->lock); | |
99 | spool->count = 1; | |
7ca02d0a MK |
100 | spool->max_hpages = max_hpages; |
101 | spool->hstate = h; | |
102 | spool->min_hpages = min_hpages; | |
103 | ||
104 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
105 | kfree(spool); | |
106 | return NULL; | |
107 | } | |
108 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
109 | |
110 | return spool; | |
111 | } | |
112 | ||
113 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
114 | { | |
115 | spin_lock(&spool->lock); | |
116 | BUG_ON(!spool->count); | |
117 | spool->count--; | |
118 | unlock_or_release_subpool(spool); | |
119 | } | |
120 | ||
1c5ecae3 MK |
121 | /* |
122 | * Subpool accounting for allocating and reserving pages. | |
123 | * Return -ENOMEM if there are not enough resources to satisfy the | |
124 | * the request. Otherwise, return the number of pages by which the | |
125 | * global pools must be adjusted (upward). The returned value may | |
126 | * only be different than the passed value (delta) in the case where | |
127 | * a subpool minimum size must be manitained. | |
128 | */ | |
129 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
130 | long delta) |
131 | { | |
1c5ecae3 | 132 | long ret = delta; |
90481622 DG |
133 | |
134 | if (!spool) | |
1c5ecae3 | 135 | return ret; |
90481622 DG |
136 | |
137 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
138 | |
139 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
140 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
141 | spool->used_hpages += delta; | |
142 | else { | |
143 | ret = -ENOMEM; | |
144 | goto unlock_ret; | |
145 | } | |
90481622 | 146 | } |
90481622 | 147 | |
09a95e29 MK |
148 | /* minimum size accounting */ |
149 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
150 | if (delta > spool->rsv_hpages) { |
151 | /* | |
152 | * Asking for more reserves than those already taken on | |
153 | * behalf of subpool. Return difference. | |
154 | */ | |
155 | ret = delta - spool->rsv_hpages; | |
156 | spool->rsv_hpages = 0; | |
157 | } else { | |
158 | ret = 0; /* reserves already accounted for */ | |
159 | spool->rsv_hpages -= delta; | |
160 | } | |
161 | } | |
162 | ||
163 | unlock_ret: | |
164 | spin_unlock(&spool->lock); | |
90481622 DG |
165 | return ret; |
166 | } | |
167 | ||
1c5ecae3 MK |
168 | /* |
169 | * Subpool accounting for freeing and unreserving pages. | |
170 | * Return the number of global page reservations that must be dropped. | |
171 | * The return value may only be different than the passed value (delta) | |
172 | * in the case where a subpool minimum size must be maintained. | |
173 | */ | |
174 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
175 | long delta) |
176 | { | |
1c5ecae3 MK |
177 | long ret = delta; |
178 | ||
90481622 | 179 | if (!spool) |
1c5ecae3 | 180 | return delta; |
90481622 DG |
181 | |
182 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
183 | |
184 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
185 | spool->used_hpages -= delta; | |
186 | ||
09a95e29 MK |
187 | /* minimum size accounting */ |
188 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
189 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
190 | ret = 0; | |
191 | else | |
192 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
193 | ||
194 | spool->rsv_hpages += delta; | |
195 | if (spool->rsv_hpages > spool->min_hpages) | |
196 | spool->rsv_hpages = spool->min_hpages; | |
197 | } | |
198 | ||
199 | /* | |
200 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
201 | * quota reference, free it now. | |
202 | */ | |
90481622 | 203 | unlock_or_release_subpool(spool); |
1c5ecae3 MK |
204 | |
205 | return ret; | |
90481622 DG |
206 | } |
207 | ||
208 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
209 | { | |
210 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
211 | } | |
212 | ||
213 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
214 | { | |
496ad9aa | 215 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
216 | } |
217 | ||
96822904 AW |
218 | /* |
219 | * Region tracking -- allows tracking of reservations and instantiated pages | |
220 | * across the pages in a mapping. | |
84afd99b | 221 | * |
1dd308a7 MK |
222 | * The region data structures are embedded into a resv_map and protected |
223 | * by a resv_map's lock. The set of regions within the resv_map represent | |
224 | * reservations for huge pages, or huge pages that have already been | |
225 | * instantiated within the map. The from and to elements are huge page | |
226 | * indicies into the associated mapping. from indicates the starting index | |
227 | * of the region. to represents the first index past the end of the region. | |
228 | * | |
229 | * For example, a file region structure with from == 0 and to == 4 represents | |
230 | * four huge pages in a mapping. It is important to note that the to element | |
231 | * represents the first element past the end of the region. This is used in | |
232 | * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. | |
233 | * | |
234 | * Interval notation of the form [from, to) will be used to indicate that | |
235 | * the endpoint from is inclusive and to is exclusive. | |
96822904 AW |
236 | */ |
237 | struct file_region { | |
238 | struct list_head link; | |
239 | long from; | |
240 | long to; | |
241 | }; | |
242 | ||
1dd308a7 MK |
243 | /* |
244 | * Add the huge page range represented by [f, t) to the reserve | |
5e911373 MK |
245 | * map. In the normal case, existing regions will be expanded |
246 | * to accommodate the specified range. Sufficient regions should | |
247 | * exist for expansion due to the previous call to region_chg | |
248 | * with the same range. However, it is possible that region_del | |
249 | * could have been called after region_chg and modifed the map | |
250 | * in such a way that no region exists to be expanded. In this | |
251 | * case, pull a region descriptor from the cache associated with | |
252 | * the map and use that for the new range. | |
cf3ad20b MK |
253 | * |
254 | * Return the number of new huge pages added to the map. This | |
255 | * number is greater than or equal to zero. | |
1dd308a7 | 256 | */ |
1406ec9b | 257 | static long region_add(struct resv_map *resv, long f, long t) |
96822904 | 258 | { |
1406ec9b | 259 | struct list_head *head = &resv->regions; |
96822904 | 260 | struct file_region *rg, *nrg, *trg; |
cf3ad20b | 261 | long add = 0; |
96822904 | 262 | |
7b24d861 | 263 | spin_lock(&resv->lock); |
96822904 AW |
264 | /* Locate the region we are either in or before. */ |
265 | list_for_each_entry(rg, head, link) | |
266 | if (f <= rg->to) | |
267 | break; | |
268 | ||
5e911373 MK |
269 | /* |
270 | * If no region exists which can be expanded to include the | |
271 | * specified range, the list must have been modified by an | |
272 | * interleving call to region_del(). Pull a region descriptor | |
273 | * from the cache and use it for this range. | |
274 | */ | |
275 | if (&rg->link == head || t < rg->from) { | |
276 | VM_BUG_ON(resv->region_cache_count <= 0); | |
277 | ||
278 | resv->region_cache_count--; | |
279 | nrg = list_first_entry(&resv->region_cache, struct file_region, | |
280 | link); | |
281 | list_del(&nrg->link); | |
282 | ||
283 | nrg->from = f; | |
284 | nrg->to = t; | |
285 | list_add(&nrg->link, rg->link.prev); | |
286 | ||
287 | add += t - f; | |
288 | goto out_locked; | |
289 | } | |
290 | ||
96822904 AW |
291 | /* Round our left edge to the current segment if it encloses us. */ |
292 | if (f > rg->from) | |
293 | f = rg->from; | |
294 | ||
295 | /* Check for and consume any regions we now overlap with. */ | |
296 | nrg = rg; | |
297 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
298 | if (&rg->link == head) | |
299 | break; | |
300 | if (rg->from > t) | |
301 | break; | |
302 | ||
303 | /* If this area reaches higher then extend our area to | |
304 | * include it completely. If this is not the first area | |
305 | * which we intend to reuse, free it. */ | |
306 | if (rg->to > t) | |
307 | t = rg->to; | |
308 | if (rg != nrg) { | |
cf3ad20b MK |
309 | /* Decrement return value by the deleted range. |
310 | * Another range will span this area so that by | |
311 | * end of routine add will be >= zero | |
312 | */ | |
313 | add -= (rg->to - rg->from); | |
96822904 AW |
314 | list_del(&rg->link); |
315 | kfree(rg); | |
316 | } | |
317 | } | |
cf3ad20b MK |
318 | |
319 | add += (nrg->from - f); /* Added to beginning of region */ | |
96822904 | 320 | nrg->from = f; |
cf3ad20b | 321 | add += t - nrg->to; /* Added to end of region */ |
96822904 | 322 | nrg->to = t; |
cf3ad20b | 323 | |
5e911373 MK |
324 | out_locked: |
325 | resv->adds_in_progress--; | |
7b24d861 | 326 | spin_unlock(&resv->lock); |
cf3ad20b MK |
327 | VM_BUG_ON(add < 0); |
328 | return add; | |
96822904 AW |
329 | } |
330 | ||
1dd308a7 MK |
331 | /* |
332 | * Examine the existing reserve map and determine how many | |
333 | * huge pages in the specified range [f, t) are NOT currently | |
334 | * represented. This routine is called before a subsequent | |
335 | * call to region_add that will actually modify the reserve | |
336 | * map to add the specified range [f, t). region_chg does | |
337 | * not change the number of huge pages represented by the | |
338 | * map. However, if the existing regions in the map can not | |
339 | * be expanded to represent the new range, a new file_region | |
340 | * structure is added to the map as a placeholder. This is | |
341 | * so that the subsequent region_add call will have all the | |
342 | * regions it needs and will not fail. | |
343 | * | |
5e911373 MK |
344 | * Upon entry, region_chg will also examine the cache of region descriptors |
345 | * associated with the map. If there are not enough descriptors cached, one | |
346 | * will be allocated for the in progress add operation. | |
347 | * | |
348 | * Returns the number of huge pages that need to be added to the existing | |
349 | * reservation map for the range [f, t). This number is greater or equal to | |
350 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
351 | * is needed and can not be allocated. | |
1dd308a7 | 352 | */ |
1406ec9b | 353 | static long region_chg(struct resv_map *resv, long f, long t) |
96822904 | 354 | { |
1406ec9b | 355 | struct list_head *head = &resv->regions; |
7b24d861 | 356 | struct file_region *rg, *nrg = NULL; |
96822904 AW |
357 | long chg = 0; |
358 | ||
7b24d861 DB |
359 | retry: |
360 | spin_lock(&resv->lock); | |
5e911373 MK |
361 | retry_locked: |
362 | resv->adds_in_progress++; | |
363 | ||
364 | /* | |
365 | * Check for sufficient descriptors in the cache to accommodate | |
366 | * the number of in progress add operations. | |
367 | */ | |
368 | if (resv->adds_in_progress > resv->region_cache_count) { | |
369 | struct file_region *trg; | |
370 | ||
371 | VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1); | |
372 | /* Must drop lock to allocate a new descriptor. */ | |
373 | resv->adds_in_progress--; | |
374 | spin_unlock(&resv->lock); | |
375 | ||
376 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
dbe409e4 MK |
377 | if (!trg) { |
378 | kfree(nrg); | |
5e911373 | 379 | return -ENOMEM; |
dbe409e4 | 380 | } |
5e911373 MK |
381 | |
382 | spin_lock(&resv->lock); | |
383 | list_add(&trg->link, &resv->region_cache); | |
384 | resv->region_cache_count++; | |
385 | goto retry_locked; | |
386 | } | |
387 | ||
96822904 AW |
388 | /* Locate the region we are before or in. */ |
389 | list_for_each_entry(rg, head, link) | |
390 | if (f <= rg->to) | |
391 | break; | |
392 | ||
393 | /* If we are below the current region then a new region is required. | |
394 | * Subtle, allocate a new region at the position but make it zero | |
395 | * size such that we can guarantee to record the reservation. */ | |
396 | if (&rg->link == head || t < rg->from) { | |
7b24d861 | 397 | if (!nrg) { |
5e911373 | 398 | resv->adds_in_progress--; |
7b24d861 DB |
399 | spin_unlock(&resv->lock); |
400 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
401 | if (!nrg) | |
402 | return -ENOMEM; | |
403 | ||
404 | nrg->from = f; | |
405 | nrg->to = f; | |
406 | INIT_LIST_HEAD(&nrg->link); | |
407 | goto retry; | |
408 | } | |
96822904 | 409 | |
7b24d861 DB |
410 | list_add(&nrg->link, rg->link.prev); |
411 | chg = t - f; | |
412 | goto out_nrg; | |
96822904 AW |
413 | } |
414 | ||
415 | /* Round our left edge to the current segment if it encloses us. */ | |
416 | if (f > rg->from) | |
417 | f = rg->from; | |
418 | chg = t - f; | |
419 | ||
420 | /* Check for and consume any regions we now overlap with. */ | |
421 | list_for_each_entry(rg, rg->link.prev, link) { | |
422 | if (&rg->link == head) | |
423 | break; | |
424 | if (rg->from > t) | |
7b24d861 | 425 | goto out; |
96822904 | 426 | |
25985edc | 427 | /* We overlap with this area, if it extends further than |
96822904 AW |
428 | * us then we must extend ourselves. Account for its |
429 | * existing reservation. */ | |
430 | if (rg->to > t) { | |
431 | chg += rg->to - t; | |
432 | t = rg->to; | |
433 | } | |
434 | chg -= rg->to - rg->from; | |
435 | } | |
7b24d861 DB |
436 | |
437 | out: | |
438 | spin_unlock(&resv->lock); | |
439 | /* We already know we raced and no longer need the new region */ | |
440 | kfree(nrg); | |
441 | return chg; | |
442 | out_nrg: | |
443 | spin_unlock(&resv->lock); | |
96822904 AW |
444 | return chg; |
445 | } | |
446 | ||
5e911373 MK |
447 | /* |
448 | * Abort the in progress add operation. The adds_in_progress field | |
449 | * of the resv_map keeps track of the operations in progress between | |
450 | * calls to region_chg and region_add. Operations are sometimes | |
451 | * aborted after the call to region_chg. In such cases, region_abort | |
452 | * is called to decrement the adds_in_progress counter. | |
453 | * | |
454 | * NOTE: The range arguments [f, t) are not needed or used in this | |
455 | * routine. They are kept to make reading the calling code easier as | |
456 | * arguments will match the associated region_chg call. | |
457 | */ | |
458 | static void region_abort(struct resv_map *resv, long f, long t) | |
459 | { | |
460 | spin_lock(&resv->lock); | |
461 | VM_BUG_ON(!resv->region_cache_count); | |
462 | resv->adds_in_progress--; | |
463 | spin_unlock(&resv->lock); | |
464 | } | |
465 | ||
1dd308a7 | 466 | /* |
feba16e2 MK |
467 | * Delete the specified range [f, t) from the reserve map. If the |
468 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
469 | * should be deleted. Locate the regions which intersect [f, t) | |
470 | * and either trim, delete or split the existing regions. | |
471 | * | |
472 | * Returns the number of huge pages deleted from the reserve map. | |
473 | * In the normal case, the return value is zero or more. In the | |
474 | * case where a region must be split, a new region descriptor must | |
475 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
476 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
477 | * a region and possibly return -ENOMEM. Callers specifying | |
478 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 479 | */ |
feba16e2 | 480 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 481 | { |
1406ec9b | 482 | struct list_head *head = &resv->regions; |
96822904 | 483 | struct file_region *rg, *trg; |
feba16e2 MK |
484 | struct file_region *nrg = NULL; |
485 | long del = 0; | |
96822904 | 486 | |
feba16e2 | 487 | retry: |
7b24d861 | 488 | spin_lock(&resv->lock); |
feba16e2 | 489 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
490 | /* |
491 | * Skip regions before the range to be deleted. file_region | |
492 | * ranges are normally of the form [from, to). However, there | |
493 | * may be a "placeholder" entry in the map which is of the form | |
494 | * (from, to) with from == to. Check for placeholder entries | |
495 | * at the beginning of the range to be deleted. | |
496 | */ | |
497 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 498 | continue; |
dbe409e4 | 499 | |
feba16e2 | 500 | if (rg->from >= t) |
96822904 | 501 | break; |
96822904 | 502 | |
feba16e2 MK |
503 | if (f > rg->from && t < rg->to) { /* Must split region */ |
504 | /* | |
505 | * Check for an entry in the cache before dropping | |
506 | * lock and attempting allocation. | |
507 | */ | |
508 | if (!nrg && | |
509 | resv->region_cache_count > resv->adds_in_progress) { | |
510 | nrg = list_first_entry(&resv->region_cache, | |
511 | struct file_region, | |
512 | link); | |
513 | list_del(&nrg->link); | |
514 | resv->region_cache_count--; | |
515 | } | |
96822904 | 516 | |
feba16e2 MK |
517 | if (!nrg) { |
518 | spin_unlock(&resv->lock); | |
519 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
520 | if (!nrg) | |
521 | return -ENOMEM; | |
522 | goto retry; | |
523 | } | |
524 | ||
525 | del += t - f; | |
526 | ||
527 | /* New entry for end of split region */ | |
528 | nrg->from = t; | |
529 | nrg->to = rg->to; | |
530 | INIT_LIST_HEAD(&nrg->link); | |
531 | ||
532 | /* Original entry is trimmed */ | |
533 | rg->to = f; | |
534 | ||
535 | list_add(&nrg->link, &rg->link); | |
536 | nrg = NULL; | |
96822904 | 537 | break; |
feba16e2 MK |
538 | } |
539 | ||
540 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
541 | del += rg->to - rg->from; | |
542 | list_del(&rg->link); | |
543 | kfree(rg); | |
544 | continue; | |
545 | } | |
546 | ||
547 | if (f <= rg->from) { /* Trim beginning of region */ | |
548 | del += t - rg->from; | |
549 | rg->from = t; | |
550 | } else { /* Trim end of region */ | |
551 | del += rg->to - f; | |
552 | rg->to = f; | |
553 | } | |
96822904 | 554 | } |
7b24d861 | 555 | |
7b24d861 | 556 | spin_unlock(&resv->lock); |
feba16e2 MK |
557 | kfree(nrg); |
558 | return del; | |
96822904 AW |
559 | } |
560 | ||
b5cec28d MK |
561 | /* |
562 | * A rare out of memory error was encountered which prevented removal of | |
563 | * the reserve map region for a page. The huge page itself was free'ed | |
564 | * and removed from the page cache. This routine will adjust the subpool | |
565 | * usage count, and the global reserve count if needed. By incrementing | |
566 | * these counts, the reserve map entry which could not be deleted will | |
567 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
568 | * counts. | |
569 | */ | |
570 | void hugetlb_fix_reserve_counts(struct inode *inode, bool restore_reserve) | |
571 | { | |
572 | struct hugepage_subpool *spool = subpool_inode(inode); | |
573 | long rsv_adjust; | |
574 | ||
575 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
576 | if (restore_reserve && rsv_adjust) { | |
577 | struct hstate *h = hstate_inode(inode); | |
578 | ||
579 | hugetlb_acct_memory(h, 1); | |
580 | } | |
581 | } | |
582 | ||
1dd308a7 MK |
583 | /* |
584 | * Count and return the number of huge pages in the reserve map | |
585 | * that intersect with the range [f, t). | |
586 | */ | |
1406ec9b | 587 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 588 | { |
1406ec9b | 589 | struct list_head *head = &resv->regions; |
84afd99b AW |
590 | struct file_region *rg; |
591 | long chg = 0; | |
592 | ||
7b24d861 | 593 | spin_lock(&resv->lock); |
84afd99b AW |
594 | /* Locate each segment we overlap with, and count that overlap. */ |
595 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
596 | long seg_from; |
597 | long seg_to; | |
84afd99b AW |
598 | |
599 | if (rg->to <= f) | |
600 | continue; | |
601 | if (rg->from >= t) | |
602 | break; | |
603 | ||
604 | seg_from = max(rg->from, f); | |
605 | seg_to = min(rg->to, t); | |
606 | ||
607 | chg += seg_to - seg_from; | |
608 | } | |
7b24d861 | 609 | spin_unlock(&resv->lock); |
84afd99b AW |
610 | |
611 | return chg; | |
612 | } | |
613 | ||
e7c4b0bf AW |
614 | /* |
615 | * Convert the address within this vma to the page offset within | |
616 | * the mapping, in pagecache page units; huge pages here. | |
617 | */ | |
a5516438 AK |
618 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
619 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 620 | { |
a5516438 AK |
621 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
622 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
623 | } |
624 | ||
0fe6e20b NH |
625 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
626 | unsigned long address) | |
627 | { | |
628 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
629 | } | |
dee41079 | 630 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 631 | |
08fba699 MG |
632 | /* |
633 | * Return the size of the pages allocated when backing a VMA. In the majority | |
634 | * cases this will be same size as used by the page table entries. | |
635 | */ | |
636 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
637 | { | |
638 | struct hstate *hstate; | |
639 | ||
640 | if (!is_vm_hugetlb_page(vma)) | |
641 | return PAGE_SIZE; | |
642 | ||
643 | hstate = hstate_vma(vma); | |
644 | ||
2415cf12 | 645 | return 1UL << huge_page_shift(hstate); |
08fba699 | 646 | } |
f340ca0f | 647 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 648 | |
3340289d MG |
649 | /* |
650 | * Return the page size being used by the MMU to back a VMA. In the majority | |
651 | * of cases, the page size used by the kernel matches the MMU size. On | |
652 | * architectures where it differs, an architecture-specific version of this | |
653 | * function is required. | |
654 | */ | |
655 | #ifndef vma_mmu_pagesize | |
656 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) | |
657 | { | |
658 | return vma_kernel_pagesize(vma); | |
659 | } | |
660 | #endif | |
661 | ||
84afd99b AW |
662 | /* |
663 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
664 | * bits of the reservation map pointer, which are always clear due to | |
665 | * alignment. | |
666 | */ | |
667 | #define HPAGE_RESV_OWNER (1UL << 0) | |
668 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 669 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 670 | |
a1e78772 MG |
671 | /* |
672 | * These helpers are used to track how many pages are reserved for | |
673 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
674 | * is guaranteed to have their future faults succeed. | |
675 | * | |
676 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
677 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
678 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
679 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
680 | * |
681 | * The private mapping reservation is represented in a subtly different | |
682 | * manner to a shared mapping. A shared mapping has a region map associated | |
683 | * with the underlying file, this region map represents the backing file | |
684 | * pages which have ever had a reservation assigned which this persists even | |
685 | * after the page is instantiated. A private mapping has a region map | |
686 | * associated with the original mmap which is attached to all VMAs which | |
687 | * reference it, this region map represents those offsets which have consumed | |
688 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 689 | */ |
e7c4b0bf AW |
690 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
691 | { | |
692 | return (unsigned long)vma->vm_private_data; | |
693 | } | |
694 | ||
695 | static void set_vma_private_data(struct vm_area_struct *vma, | |
696 | unsigned long value) | |
697 | { | |
698 | vma->vm_private_data = (void *)value; | |
699 | } | |
700 | ||
9119a41e | 701 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
702 | { |
703 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
704 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
705 | ||
706 | if (!resv_map || !rg) { | |
707 | kfree(resv_map); | |
708 | kfree(rg); | |
84afd99b | 709 | return NULL; |
5e911373 | 710 | } |
84afd99b AW |
711 | |
712 | kref_init(&resv_map->refs); | |
7b24d861 | 713 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
714 | INIT_LIST_HEAD(&resv_map->regions); |
715 | ||
5e911373 MK |
716 | resv_map->adds_in_progress = 0; |
717 | ||
718 | INIT_LIST_HEAD(&resv_map->region_cache); | |
719 | list_add(&rg->link, &resv_map->region_cache); | |
720 | resv_map->region_cache_count = 1; | |
721 | ||
84afd99b AW |
722 | return resv_map; |
723 | } | |
724 | ||
9119a41e | 725 | void resv_map_release(struct kref *ref) |
84afd99b AW |
726 | { |
727 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
728 | struct list_head *head = &resv_map->region_cache; |
729 | struct file_region *rg, *trg; | |
84afd99b AW |
730 | |
731 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 732 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
733 | |
734 | /* ... and any entries left in the cache */ | |
735 | list_for_each_entry_safe(rg, trg, head, link) { | |
736 | list_del(&rg->link); | |
737 | kfree(rg); | |
738 | } | |
739 | ||
740 | VM_BUG_ON(resv_map->adds_in_progress); | |
741 | ||
84afd99b AW |
742 | kfree(resv_map); |
743 | } | |
744 | ||
4e35f483 JK |
745 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
746 | { | |
747 | return inode->i_mapping->private_data; | |
748 | } | |
749 | ||
84afd99b | 750 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 751 | { |
81d1b09c | 752 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
753 | if (vma->vm_flags & VM_MAYSHARE) { |
754 | struct address_space *mapping = vma->vm_file->f_mapping; | |
755 | struct inode *inode = mapping->host; | |
756 | ||
757 | return inode_resv_map(inode); | |
758 | ||
759 | } else { | |
84afd99b AW |
760 | return (struct resv_map *)(get_vma_private_data(vma) & |
761 | ~HPAGE_RESV_MASK); | |
4e35f483 | 762 | } |
a1e78772 MG |
763 | } |
764 | ||
84afd99b | 765 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 766 | { |
81d1b09c SL |
767 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
768 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 769 | |
84afd99b AW |
770 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
771 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
772 | } |
773 | ||
774 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
775 | { | |
81d1b09c SL |
776 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
777 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
778 | |
779 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
780 | } |
781 | ||
782 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
783 | { | |
81d1b09c | 784 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
785 | |
786 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
787 | } |
788 | ||
04f2cbe3 | 789 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
790 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
791 | { | |
81d1b09c | 792 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 793 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
794 | vma->vm_private_data = (void *)0; |
795 | } | |
796 | ||
797 | /* Returns true if the VMA has associated reserve pages */ | |
559ec2f8 | 798 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 799 | { |
af0ed73e JK |
800 | if (vma->vm_flags & VM_NORESERVE) { |
801 | /* | |
802 | * This address is already reserved by other process(chg == 0), | |
803 | * so, we should decrement reserved count. Without decrementing, | |
804 | * reserve count remains after releasing inode, because this | |
805 | * allocated page will go into page cache and is regarded as | |
806 | * coming from reserved pool in releasing step. Currently, we | |
807 | * don't have any other solution to deal with this situation | |
808 | * properly, so add work-around here. | |
809 | */ | |
810 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 811 | return true; |
af0ed73e | 812 | else |
559ec2f8 | 813 | return false; |
af0ed73e | 814 | } |
a63884e9 JK |
815 | |
816 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
817 | if (vma->vm_flags & VM_MAYSHARE) { |
818 | /* | |
819 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
820 | * be a region map for all pages. The only situation where | |
821 | * there is no region map is if a hole was punched via | |
822 | * fallocate. In this case, there really are no reverves to | |
823 | * use. This situation is indicated if chg != 0. | |
824 | */ | |
825 | if (chg) | |
826 | return false; | |
827 | else | |
828 | return true; | |
829 | } | |
a63884e9 JK |
830 | |
831 | /* | |
832 | * Only the process that called mmap() has reserves for | |
833 | * private mappings. | |
834 | */ | |
67961f9d MK |
835 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
836 | /* | |
837 | * Like the shared case above, a hole punch or truncate | |
838 | * could have been performed on the private mapping. | |
839 | * Examine the value of chg to determine if reserves | |
840 | * actually exist or were previously consumed. | |
841 | * Very Subtle - The value of chg comes from a previous | |
842 | * call to vma_needs_reserves(). The reserve map for | |
843 | * private mappings has different (opposite) semantics | |
844 | * than that of shared mappings. vma_needs_reserves() | |
845 | * has already taken this difference in semantics into | |
846 | * account. Therefore, the meaning of chg is the same | |
847 | * as in the shared case above. Code could easily be | |
848 | * combined, but keeping it separate draws attention to | |
849 | * subtle differences. | |
850 | */ | |
851 | if (chg) | |
852 | return false; | |
853 | else | |
854 | return true; | |
855 | } | |
a63884e9 | 856 | |
559ec2f8 | 857 | return false; |
a1e78772 MG |
858 | } |
859 | ||
a5516438 | 860 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
861 | { |
862 | int nid = page_to_nid(page); | |
0edaecfa | 863 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
864 | h->free_huge_pages++; |
865 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
866 | } |
867 | ||
bf50bab2 NH |
868 | static struct page *dequeue_huge_page_node(struct hstate *h, int nid) |
869 | { | |
870 | struct page *page; | |
871 | ||
c8721bbb NH |
872 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
873 | if (!is_migrate_isolate_page(page)) | |
874 | break; | |
875 | /* | |
876 | * if 'non-isolated free hugepage' not found on the list, | |
877 | * the allocation fails. | |
878 | */ | |
879 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 880 | return NULL; |
0edaecfa | 881 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 882 | set_page_refcounted(page); |
bf50bab2 NH |
883 | h->free_huge_pages--; |
884 | h->free_huge_pages_node[nid]--; | |
885 | return page; | |
886 | } | |
887 | ||
86cdb465 NH |
888 | /* Movability of hugepages depends on migration support. */ |
889 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
890 | { | |
100873d7 | 891 | if (hugepages_treat_as_movable || hugepage_migration_supported(h)) |
86cdb465 NH |
892 | return GFP_HIGHUSER_MOVABLE; |
893 | else | |
894 | return GFP_HIGHUSER; | |
895 | } | |
896 | ||
a5516438 AK |
897 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
898 | struct vm_area_struct *vma, | |
af0ed73e JK |
899 | unsigned long address, int avoid_reserve, |
900 | long chg) | |
1da177e4 | 901 | { |
b1c12cbc | 902 | struct page *page = NULL; |
480eccf9 | 903 | struct mempolicy *mpol; |
19770b32 | 904 | nodemask_t *nodemask; |
c0ff7453 | 905 | struct zonelist *zonelist; |
dd1a239f MG |
906 | struct zone *zone; |
907 | struct zoneref *z; | |
cc9a6c87 | 908 | unsigned int cpuset_mems_cookie; |
1da177e4 | 909 | |
a1e78772 MG |
910 | /* |
911 | * A child process with MAP_PRIVATE mappings created by their parent | |
912 | * have no page reserves. This check ensures that reservations are | |
913 | * not "stolen". The child may still get SIGKILLed | |
914 | */ | |
af0ed73e | 915 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 916 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 917 | goto err; |
a1e78772 | 918 | |
04f2cbe3 | 919 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 920 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 921 | goto err; |
04f2cbe3 | 922 | |
9966c4bb | 923 | retry_cpuset: |
d26914d1 | 924 | cpuset_mems_cookie = read_mems_allowed_begin(); |
9966c4bb | 925 | zonelist = huge_zonelist(vma, address, |
86cdb465 | 926 | htlb_alloc_mask(h), &mpol, &nodemask); |
9966c4bb | 927 | |
19770b32 MG |
928 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
929 | MAX_NR_ZONES - 1, nodemask) { | |
344736f2 | 930 | if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) { |
bf50bab2 NH |
931 | page = dequeue_huge_page_node(h, zone_to_nid(zone)); |
932 | if (page) { | |
af0ed73e JK |
933 | if (avoid_reserve) |
934 | break; | |
935 | if (!vma_has_reserves(vma, chg)) | |
936 | break; | |
937 | ||
07443a85 | 938 | SetPagePrivate(page); |
af0ed73e | 939 | h->resv_huge_pages--; |
bf50bab2 NH |
940 | break; |
941 | } | |
3abf7afd | 942 | } |
1da177e4 | 943 | } |
cc9a6c87 | 944 | |
52cd3b07 | 945 | mpol_cond_put(mpol); |
d26914d1 | 946 | if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) |
cc9a6c87 | 947 | goto retry_cpuset; |
1da177e4 | 948 | return page; |
cc9a6c87 MG |
949 | |
950 | err: | |
cc9a6c87 | 951 | return NULL; |
1da177e4 LT |
952 | } |
953 | ||
1cac6f2c LC |
954 | /* |
955 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
956 | * We may have allocated or freed a huge page based on a different | |
957 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
958 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
959 | * node for alloc or free. | |
960 | */ | |
961 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
962 | { | |
0edaf86c | 963 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
964 | VM_BUG_ON(nid >= MAX_NUMNODES); |
965 | ||
966 | return nid; | |
967 | } | |
968 | ||
969 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
970 | { | |
971 | if (!node_isset(nid, *nodes_allowed)) | |
972 | nid = next_node_allowed(nid, nodes_allowed); | |
973 | return nid; | |
974 | } | |
975 | ||
976 | /* | |
977 | * returns the previously saved node ["this node"] from which to | |
978 | * allocate a persistent huge page for the pool and advance the | |
979 | * next node from which to allocate, handling wrap at end of node | |
980 | * mask. | |
981 | */ | |
982 | static int hstate_next_node_to_alloc(struct hstate *h, | |
983 | nodemask_t *nodes_allowed) | |
984 | { | |
985 | int nid; | |
986 | ||
987 | VM_BUG_ON(!nodes_allowed); | |
988 | ||
989 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
990 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
991 | ||
992 | return nid; | |
993 | } | |
994 | ||
995 | /* | |
996 | * helper for free_pool_huge_page() - return the previously saved | |
997 | * node ["this node"] from which to free a huge page. Advance the | |
998 | * next node id whether or not we find a free huge page to free so | |
999 | * that the next attempt to free addresses the next node. | |
1000 | */ | |
1001 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1002 | { | |
1003 | int nid; | |
1004 | ||
1005 | VM_BUG_ON(!nodes_allowed); | |
1006 | ||
1007 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1008 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1009 | ||
1010 | return nid; | |
1011 | } | |
1012 | ||
1013 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1014 | for (nr_nodes = nodes_weight(*mask); \ | |
1015 | nr_nodes > 0 && \ | |
1016 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1017 | nr_nodes--) | |
1018 | ||
1019 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1020 | for (nr_nodes = nodes_weight(*mask); \ | |
1021 | nr_nodes > 0 && \ | |
1022 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1023 | nr_nodes--) | |
1024 | ||
080fe206 | 1025 | #if defined(CONFIG_X86_64) && ((defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)) |
944d9fec | 1026 | static void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1027 | unsigned int order) |
944d9fec LC |
1028 | { |
1029 | int i; | |
1030 | int nr_pages = 1 << order; | |
1031 | struct page *p = page + 1; | |
1032 | ||
1033 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { | |
1d798ca3 | 1034 | clear_compound_head(p); |
944d9fec | 1035 | set_page_refcounted(p); |
944d9fec LC |
1036 | } |
1037 | ||
1038 | set_compound_order(page, 0); | |
1039 | __ClearPageHead(page); | |
1040 | } | |
1041 | ||
d00181b9 | 1042 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec LC |
1043 | { |
1044 | free_contig_range(page_to_pfn(page), 1 << order); | |
1045 | } | |
1046 | ||
1047 | static int __alloc_gigantic_page(unsigned long start_pfn, | |
1048 | unsigned long nr_pages) | |
1049 | { | |
1050 | unsigned long end_pfn = start_pfn + nr_pages; | |
1051 | return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE); | |
1052 | } | |
1053 | ||
f44b2dda JK |
1054 | static bool pfn_range_valid_gigantic(struct zone *z, |
1055 | unsigned long start_pfn, unsigned long nr_pages) | |
944d9fec LC |
1056 | { |
1057 | unsigned long i, end_pfn = start_pfn + nr_pages; | |
1058 | struct page *page; | |
1059 | ||
1060 | for (i = start_pfn; i < end_pfn; i++) { | |
1061 | if (!pfn_valid(i)) | |
1062 | return false; | |
1063 | ||
1064 | page = pfn_to_page(i); | |
1065 | ||
f44b2dda JK |
1066 | if (page_zone(page) != z) |
1067 | return false; | |
1068 | ||
944d9fec LC |
1069 | if (PageReserved(page)) |
1070 | return false; | |
1071 | ||
1072 | if (page_count(page) > 0) | |
1073 | return false; | |
1074 | ||
1075 | if (PageHuge(page)) | |
1076 | return false; | |
1077 | } | |
1078 | ||
1079 | return true; | |
1080 | } | |
1081 | ||
1082 | static bool zone_spans_last_pfn(const struct zone *zone, | |
1083 | unsigned long start_pfn, unsigned long nr_pages) | |
1084 | { | |
1085 | unsigned long last_pfn = start_pfn + nr_pages - 1; | |
1086 | return zone_spans_pfn(zone, last_pfn); | |
1087 | } | |
1088 | ||
d00181b9 | 1089 | static struct page *alloc_gigantic_page(int nid, unsigned int order) |
944d9fec LC |
1090 | { |
1091 | unsigned long nr_pages = 1 << order; | |
1092 | unsigned long ret, pfn, flags; | |
1093 | struct zone *z; | |
1094 | ||
1095 | z = NODE_DATA(nid)->node_zones; | |
1096 | for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) { | |
1097 | spin_lock_irqsave(&z->lock, flags); | |
1098 | ||
1099 | pfn = ALIGN(z->zone_start_pfn, nr_pages); | |
1100 | while (zone_spans_last_pfn(z, pfn, nr_pages)) { | |
f44b2dda | 1101 | if (pfn_range_valid_gigantic(z, pfn, nr_pages)) { |
944d9fec LC |
1102 | /* |
1103 | * We release the zone lock here because | |
1104 | * alloc_contig_range() will also lock the zone | |
1105 | * at some point. If there's an allocation | |
1106 | * spinning on this lock, it may win the race | |
1107 | * and cause alloc_contig_range() to fail... | |
1108 | */ | |
1109 | spin_unlock_irqrestore(&z->lock, flags); | |
1110 | ret = __alloc_gigantic_page(pfn, nr_pages); | |
1111 | if (!ret) | |
1112 | return pfn_to_page(pfn); | |
1113 | spin_lock_irqsave(&z->lock, flags); | |
1114 | } | |
1115 | pfn += nr_pages; | |
1116 | } | |
1117 | ||
1118 | spin_unlock_irqrestore(&z->lock, flags); | |
1119 | } | |
1120 | ||
1121 | return NULL; | |
1122 | } | |
1123 | ||
1124 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
d00181b9 | 1125 | static void prep_compound_gigantic_page(struct page *page, unsigned int order); |
944d9fec LC |
1126 | |
1127 | static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) | |
1128 | { | |
1129 | struct page *page; | |
1130 | ||
1131 | page = alloc_gigantic_page(nid, huge_page_order(h)); | |
1132 | if (page) { | |
1133 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
1134 | prep_new_huge_page(h, page, nid); | |
1135 | } | |
1136 | ||
1137 | return page; | |
1138 | } | |
1139 | ||
1140 | static int alloc_fresh_gigantic_page(struct hstate *h, | |
1141 | nodemask_t *nodes_allowed) | |
1142 | { | |
1143 | struct page *page = NULL; | |
1144 | int nr_nodes, node; | |
1145 | ||
1146 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1147 | page = alloc_fresh_gigantic_page_node(h, node); | |
1148 | if (page) | |
1149 | return 1; | |
1150 | } | |
1151 | ||
1152 | return 0; | |
1153 | } | |
1154 | ||
1155 | static inline bool gigantic_page_supported(void) { return true; } | |
1156 | #else | |
1157 | static inline bool gigantic_page_supported(void) { return false; } | |
d00181b9 | 1158 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1159 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1160 | unsigned int order) { } |
944d9fec LC |
1161 | static inline int alloc_fresh_gigantic_page(struct hstate *h, |
1162 | nodemask_t *nodes_allowed) { return 0; } | |
1163 | #endif | |
1164 | ||
a5516438 | 1165 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1166 | { |
1167 | int i; | |
a5516438 | 1168 | |
944d9fec LC |
1169 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
1170 | return; | |
18229df5 | 1171 | |
a5516438 AK |
1172 | h->nr_huge_pages--; |
1173 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
1174 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
1175 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
1176 | 1 << PG_referenced | 1 << PG_dirty | | |
a7407a27 LC |
1177 | 1 << PG_active | 1 << PG_private | |
1178 | 1 << PG_writeback); | |
6af2acb6 | 1179 | } |
309381fe | 1180 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
f1e61557 | 1181 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); |
6af2acb6 | 1182 | set_page_refcounted(page); |
944d9fec LC |
1183 | if (hstate_is_gigantic(h)) { |
1184 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
1185 | free_gigantic_page(page, huge_page_order(h)); | |
1186 | } else { | |
944d9fec LC |
1187 | __free_pages(page, huge_page_order(h)); |
1188 | } | |
6af2acb6 AL |
1189 | } |
1190 | ||
e5ff2159 AK |
1191 | struct hstate *size_to_hstate(unsigned long size) |
1192 | { | |
1193 | struct hstate *h; | |
1194 | ||
1195 | for_each_hstate(h) { | |
1196 | if (huge_page_size(h) == size) | |
1197 | return h; | |
1198 | } | |
1199 | return NULL; | |
1200 | } | |
1201 | ||
bcc54222 NH |
1202 | /* |
1203 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
1204 | * to hstate->hugepage_activelist.) | |
1205 | * | |
1206 | * This function can be called for tail pages, but never returns true for them. | |
1207 | */ | |
1208 | bool page_huge_active(struct page *page) | |
1209 | { | |
1210 | VM_BUG_ON_PAGE(!PageHuge(page), page); | |
1211 | return PageHead(page) && PagePrivate(&page[1]); | |
1212 | } | |
1213 | ||
1214 | /* never called for tail page */ | |
1215 | static void set_page_huge_active(struct page *page) | |
1216 | { | |
1217 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1218 | SetPagePrivate(&page[1]); | |
1219 | } | |
1220 | ||
1221 | static void clear_page_huge_active(struct page *page) | |
1222 | { | |
1223 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1224 | ClearPagePrivate(&page[1]); | |
1225 | } | |
1226 | ||
8f1d26d0 | 1227 | void free_huge_page(struct page *page) |
27a85ef1 | 1228 | { |
a5516438 AK |
1229 | /* |
1230 | * Can't pass hstate in here because it is called from the | |
1231 | * compound page destructor. | |
1232 | */ | |
e5ff2159 | 1233 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1234 | int nid = page_to_nid(page); |
90481622 DG |
1235 | struct hugepage_subpool *spool = |
1236 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 1237 | bool restore_reserve; |
27a85ef1 | 1238 | |
e5df70ab | 1239 | set_page_private(page, 0); |
23be7468 | 1240 | page->mapping = NULL; |
b4330afb MK |
1241 | VM_BUG_ON_PAGE(page_count(page), page); |
1242 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
07443a85 | 1243 | restore_reserve = PagePrivate(page); |
16c794b4 | 1244 | ClearPagePrivate(page); |
27a85ef1 | 1245 | |
1c5ecae3 MK |
1246 | /* |
1247 | * A return code of zero implies that the subpool will be under its | |
1248 | * minimum size if the reservation is not restored after page is free. | |
1249 | * Therefore, force restore_reserve operation. | |
1250 | */ | |
1251 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1252 | restore_reserve = true; | |
1253 | ||
27a85ef1 | 1254 | spin_lock(&hugetlb_lock); |
bcc54222 | 1255 | clear_page_huge_active(page); |
6d76dcf4 AK |
1256 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1257 | pages_per_huge_page(h), page); | |
07443a85 JK |
1258 | if (restore_reserve) |
1259 | h->resv_huge_pages++; | |
1260 | ||
944d9fec | 1261 | if (h->surplus_huge_pages_node[nid]) { |
0edaecfa AK |
1262 | /* remove the page from active list */ |
1263 | list_del(&page->lru); | |
a5516438 AK |
1264 | update_and_free_page(h, page); |
1265 | h->surplus_huge_pages--; | |
1266 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 1267 | } else { |
5d3a551c | 1268 | arch_clear_hugepage_flags(page); |
a5516438 | 1269 | enqueue_huge_page(h, page); |
7893d1d5 | 1270 | } |
27a85ef1 DG |
1271 | spin_unlock(&hugetlb_lock); |
1272 | } | |
1273 | ||
a5516438 | 1274 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1275 | { |
0edaecfa | 1276 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1277 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
b7ba30c6 | 1278 | spin_lock(&hugetlb_lock); |
9dd540e2 | 1279 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
1280 | h->nr_huge_pages++; |
1281 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 AK |
1282 | spin_unlock(&hugetlb_lock); |
1283 | put_page(page); /* free it into the hugepage allocator */ | |
1284 | } | |
1285 | ||
d00181b9 | 1286 | static void prep_compound_gigantic_page(struct page *page, unsigned int order) |
20a0307c WF |
1287 | { |
1288 | int i; | |
1289 | int nr_pages = 1 << order; | |
1290 | struct page *p = page + 1; | |
1291 | ||
1292 | /* we rely on prep_new_huge_page to set the destructor */ | |
1293 | set_compound_order(page, order); | |
ef5a22be | 1294 | __ClearPageReserved(page); |
de09d31d | 1295 | __SetPageHead(page); |
20a0307c | 1296 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1297 | /* |
1298 | * For gigantic hugepages allocated through bootmem at | |
1299 | * boot, it's safer to be consistent with the not-gigantic | |
1300 | * hugepages and clear the PG_reserved bit from all tail pages | |
1301 | * too. Otherwse drivers using get_user_pages() to access tail | |
1302 | * pages may get the reference counting wrong if they see | |
1303 | * PG_reserved set on a tail page (despite the head page not | |
1304 | * having PG_reserved set). Enforcing this consistency between | |
1305 | * head and tail pages allows drivers to optimize away a check | |
1306 | * on the head page when they need know if put_page() is needed | |
1307 | * after get_user_pages(). | |
1308 | */ | |
1309 | __ClearPageReserved(p); | |
58a84aa9 | 1310 | set_page_count(p, 0); |
1d798ca3 | 1311 | set_compound_head(p, page); |
20a0307c | 1312 | } |
b4330afb | 1313 | atomic_set(compound_mapcount_ptr(page), -1); |
20a0307c WF |
1314 | } |
1315 | ||
7795912c AM |
1316 | /* |
1317 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1318 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1319 | * details. | |
1320 | */ | |
20a0307c WF |
1321 | int PageHuge(struct page *page) |
1322 | { | |
20a0307c WF |
1323 | if (!PageCompound(page)) |
1324 | return 0; | |
1325 | ||
1326 | page = compound_head(page); | |
f1e61557 | 1327 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1328 | } |
43131e14 NH |
1329 | EXPORT_SYMBOL_GPL(PageHuge); |
1330 | ||
27c73ae7 AA |
1331 | /* |
1332 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1333 | * normal or transparent huge pages. | |
1334 | */ | |
1335 | int PageHeadHuge(struct page *page_head) | |
1336 | { | |
27c73ae7 AA |
1337 | if (!PageHead(page_head)) |
1338 | return 0; | |
1339 | ||
758f66a2 | 1340 | return get_compound_page_dtor(page_head) == free_huge_page; |
27c73ae7 | 1341 | } |
27c73ae7 | 1342 | |
13d60f4b ZY |
1343 | pgoff_t __basepage_index(struct page *page) |
1344 | { | |
1345 | struct page *page_head = compound_head(page); | |
1346 | pgoff_t index = page_index(page_head); | |
1347 | unsigned long compound_idx; | |
1348 | ||
1349 | if (!PageHuge(page_head)) | |
1350 | return page_index(page); | |
1351 | ||
1352 | if (compound_order(page_head) >= MAX_ORDER) | |
1353 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1354 | else | |
1355 | compound_idx = page - page_head; | |
1356 | ||
1357 | return (index << compound_order(page_head)) + compound_idx; | |
1358 | } | |
1359 | ||
a5516438 | 1360 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
1da177e4 | 1361 | { |
1da177e4 | 1362 | struct page *page; |
f96efd58 | 1363 | |
96db800f | 1364 | page = __alloc_pages_node(nid, |
86cdb465 | 1365 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
551883ae | 1366 | __GFP_REPEAT|__GFP_NOWARN, |
a5516438 | 1367 | huge_page_order(h)); |
1da177e4 | 1368 | if (page) { |
a5516438 | 1369 | prep_new_huge_page(h, page, nid); |
1da177e4 | 1370 | } |
63b4613c NA |
1371 | |
1372 | return page; | |
1373 | } | |
1374 | ||
b2261026 JK |
1375 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
1376 | { | |
1377 | struct page *page; | |
1378 | int nr_nodes, node; | |
1379 | int ret = 0; | |
1380 | ||
1381 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1382 | page = alloc_fresh_huge_page_node(h, node); | |
1383 | if (page) { | |
1384 | ret = 1; | |
1385 | break; | |
1386 | } | |
1387 | } | |
1388 | ||
1389 | if (ret) | |
1390 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
1391 | else | |
1392 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
1393 | ||
1394 | return ret; | |
1395 | } | |
1396 | ||
e8c5c824 LS |
1397 | /* |
1398 | * Free huge page from pool from next node to free. | |
1399 | * Attempt to keep persistent huge pages more or less | |
1400 | * balanced over allowed nodes. | |
1401 | * Called with hugetlb_lock locked. | |
1402 | */ | |
6ae11b27 LS |
1403 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1404 | bool acct_surplus) | |
e8c5c824 | 1405 | { |
b2261026 | 1406 | int nr_nodes, node; |
e8c5c824 LS |
1407 | int ret = 0; |
1408 | ||
b2261026 | 1409 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1410 | /* |
1411 | * If we're returning unused surplus pages, only examine | |
1412 | * nodes with surplus pages. | |
1413 | */ | |
b2261026 JK |
1414 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1415 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1416 | struct page *page = |
b2261026 | 1417 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1418 | struct page, lru); |
1419 | list_del(&page->lru); | |
1420 | h->free_huge_pages--; | |
b2261026 | 1421 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1422 | if (acct_surplus) { |
1423 | h->surplus_huge_pages--; | |
b2261026 | 1424 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1425 | } |
e8c5c824 LS |
1426 | update_and_free_page(h, page); |
1427 | ret = 1; | |
9a76db09 | 1428 | break; |
e8c5c824 | 1429 | } |
b2261026 | 1430 | } |
e8c5c824 LS |
1431 | |
1432 | return ret; | |
1433 | } | |
1434 | ||
c8721bbb NH |
1435 | /* |
1436 | * Dissolve a given free hugepage into free buddy pages. This function does | |
1437 | * nothing for in-use (including surplus) hugepages. | |
1438 | */ | |
1439 | static void dissolve_free_huge_page(struct page *page) | |
1440 | { | |
1441 | spin_lock(&hugetlb_lock); | |
1442 | if (PageHuge(page) && !page_count(page)) { | |
1443 | struct hstate *h = page_hstate(page); | |
1444 | int nid = page_to_nid(page); | |
1445 | list_del(&page->lru); | |
1446 | h->free_huge_pages--; | |
1447 | h->free_huge_pages_node[nid]--; | |
1448 | update_and_free_page(h, page); | |
1449 | } | |
1450 | spin_unlock(&hugetlb_lock); | |
1451 | } | |
1452 | ||
1453 | /* | |
1454 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1455 | * make specified memory blocks removable from the system. | |
1456 | * Note that start_pfn should aligned with (minimum) hugepage size. | |
1457 | */ | |
1458 | void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) | |
1459 | { | |
c8721bbb | 1460 | unsigned long pfn; |
c8721bbb | 1461 | |
d0177639 LZ |
1462 | if (!hugepages_supported()) |
1463 | return; | |
1464 | ||
641844f5 NH |
1465 | VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order)); |
1466 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) | |
c8721bbb NH |
1467 | dissolve_free_huge_page(pfn_to_page(pfn)); |
1468 | } | |
1469 | ||
099730d6 DH |
1470 | /* |
1471 | * There are 3 ways this can get called: | |
1472 | * 1. With vma+addr: we use the VMA's memory policy | |
1473 | * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge | |
1474 | * page from any node, and let the buddy allocator itself figure | |
1475 | * it out. | |
1476 | * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page | |
1477 | * strictly from 'nid' | |
1478 | */ | |
1479 | static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h, | |
1480 | struct vm_area_struct *vma, unsigned long addr, int nid) | |
1481 | { | |
1482 | int order = huge_page_order(h); | |
1483 | gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN; | |
1484 | unsigned int cpuset_mems_cookie; | |
1485 | ||
1486 | /* | |
1487 | * We need a VMA to get a memory policy. If we do not | |
e0ec90ee DH |
1488 | * have one, we use the 'nid' argument. |
1489 | * | |
1490 | * The mempolicy stuff below has some non-inlined bits | |
1491 | * and calls ->vm_ops. That makes it hard to optimize at | |
1492 | * compile-time, even when NUMA is off and it does | |
1493 | * nothing. This helps the compiler optimize it out. | |
099730d6 | 1494 | */ |
e0ec90ee | 1495 | if (!IS_ENABLED(CONFIG_NUMA) || !vma) { |
099730d6 DH |
1496 | /* |
1497 | * If a specific node is requested, make sure to | |
1498 | * get memory from there, but only when a node | |
1499 | * is explicitly specified. | |
1500 | */ | |
1501 | if (nid != NUMA_NO_NODE) | |
1502 | gfp |= __GFP_THISNODE; | |
1503 | /* | |
1504 | * Make sure to call something that can handle | |
1505 | * nid=NUMA_NO_NODE | |
1506 | */ | |
1507 | return alloc_pages_node(nid, gfp, order); | |
1508 | } | |
1509 | ||
1510 | /* | |
1511 | * OK, so we have a VMA. Fetch the mempolicy and try to | |
e0ec90ee DH |
1512 | * allocate a huge page with it. We will only reach this |
1513 | * when CONFIG_NUMA=y. | |
099730d6 DH |
1514 | */ |
1515 | do { | |
1516 | struct page *page; | |
1517 | struct mempolicy *mpol; | |
1518 | struct zonelist *zl; | |
1519 | nodemask_t *nodemask; | |
1520 | ||
1521 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
1522 | zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask); | |
1523 | mpol_cond_put(mpol); | |
1524 | page = __alloc_pages_nodemask(gfp, order, zl, nodemask); | |
1525 | if (page) | |
1526 | return page; | |
1527 | } while (read_mems_allowed_retry(cpuset_mems_cookie)); | |
1528 | ||
1529 | return NULL; | |
1530 | } | |
1531 | ||
1532 | /* | |
1533 | * There are two ways to allocate a huge page: | |
1534 | * 1. When you have a VMA and an address (like a fault) | |
1535 | * 2. When you have no VMA (like when setting /proc/.../nr_hugepages) | |
1536 | * | |
1537 | * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in | |
1538 | * this case which signifies that the allocation should be done with | |
1539 | * respect for the VMA's memory policy. | |
1540 | * | |
1541 | * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This | |
1542 | * implies that memory policies will not be taken in to account. | |
1543 | */ | |
1544 | static struct page *__alloc_buddy_huge_page(struct hstate *h, | |
1545 | struct vm_area_struct *vma, unsigned long addr, int nid) | |
7893d1d5 AL |
1546 | { |
1547 | struct page *page; | |
bf50bab2 | 1548 | unsigned int r_nid; |
7893d1d5 | 1549 | |
bae7f4ae | 1550 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1551 | return NULL; |
1552 | ||
099730d6 DH |
1553 | /* |
1554 | * Make sure that anyone specifying 'nid' is not also specifying a VMA. | |
1555 | * This makes sure the caller is picking _one_ of the modes with which | |
1556 | * we can call this function, not both. | |
1557 | */ | |
1558 | if (vma || (addr != -1)) { | |
e0ec90ee DH |
1559 | VM_WARN_ON_ONCE(addr == -1); |
1560 | VM_WARN_ON_ONCE(nid != NUMA_NO_NODE); | |
099730d6 | 1561 | } |
d1c3fb1f NA |
1562 | /* |
1563 | * Assume we will successfully allocate the surplus page to | |
1564 | * prevent racing processes from causing the surplus to exceed | |
1565 | * overcommit | |
1566 | * | |
1567 | * This however introduces a different race, where a process B | |
1568 | * tries to grow the static hugepage pool while alloc_pages() is | |
1569 | * called by process A. B will only examine the per-node | |
1570 | * counters in determining if surplus huge pages can be | |
1571 | * converted to normal huge pages in adjust_pool_surplus(). A | |
1572 | * won't be able to increment the per-node counter, until the | |
1573 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
1574 | * no more huge pages can be converted from surplus to normal | |
1575 | * state (and doesn't try to convert again). Thus, we have a | |
1576 | * case where a surplus huge page exists, the pool is grown, and | |
1577 | * the surplus huge page still exists after, even though it | |
1578 | * should just have been converted to a normal huge page. This | |
1579 | * does not leak memory, though, as the hugepage will be freed | |
1580 | * once it is out of use. It also does not allow the counters to | |
1581 | * go out of whack in adjust_pool_surplus() as we don't modify | |
1582 | * the node values until we've gotten the hugepage and only the | |
1583 | * per-node value is checked there. | |
1584 | */ | |
1585 | spin_lock(&hugetlb_lock); | |
a5516438 | 1586 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
1587 | spin_unlock(&hugetlb_lock); |
1588 | return NULL; | |
1589 | } else { | |
a5516438 AK |
1590 | h->nr_huge_pages++; |
1591 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
1592 | } |
1593 | spin_unlock(&hugetlb_lock); | |
1594 | ||
099730d6 | 1595 | page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid); |
d1c3fb1f NA |
1596 | |
1597 | spin_lock(&hugetlb_lock); | |
7893d1d5 | 1598 | if (page) { |
0edaecfa | 1599 | INIT_LIST_HEAD(&page->lru); |
bf50bab2 | 1600 | r_nid = page_to_nid(page); |
f1e61557 | 1601 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
9dd540e2 | 1602 | set_hugetlb_cgroup(page, NULL); |
d1c3fb1f NA |
1603 | /* |
1604 | * We incremented the global counters already | |
1605 | */ | |
bf50bab2 NH |
1606 | h->nr_huge_pages_node[r_nid]++; |
1607 | h->surplus_huge_pages_node[r_nid]++; | |
3b116300 | 1608 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 1609 | } else { |
a5516438 AK |
1610 | h->nr_huge_pages--; |
1611 | h->surplus_huge_pages--; | |
3b116300 | 1612 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 1613 | } |
d1c3fb1f | 1614 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1615 | |
1616 | return page; | |
1617 | } | |
1618 | ||
099730d6 DH |
1619 | /* |
1620 | * Allocate a huge page from 'nid'. Note, 'nid' may be | |
1621 | * NUMA_NO_NODE, which means that it may be allocated | |
1622 | * anywhere. | |
1623 | */ | |
e0ec90ee | 1624 | static |
099730d6 DH |
1625 | struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid) |
1626 | { | |
1627 | unsigned long addr = -1; | |
1628 | ||
1629 | return __alloc_buddy_huge_page(h, NULL, addr, nid); | |
1630 | } | |
1631 | ||
1632 | /* | |
1633 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
1634 | */ | |
e0ec90ee | 1635 | static |
099730d6 DH |
1636 | struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h, |
1637 | struct vm_area_struct *vma, unsigned long addr) | |
1638 | { | |
1639 | return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE); | |
1640 | } | |
1641 | ||
bf50bab2 NH |
1642 | /* |
1643 | * This allocation function is useful in the context where vma is irrelevant. | |
1644 | * E.g. soft-offlining uses this function because it only cares physical | |
1645 | * address of error page. | |
1646 | */ | |
1647 | struct page *alloc_huge_page_node(struct hstate *h, int nid) | |
1648 | { | |
4ef91848 | 1649 | struct page *page = NULL; |
bf50bab2 NH |
1650 | |
1651 | spin_lock(&hugetlb_lock); | |
4ef91848 JK |
1652 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
1653 | page = dequeue_huge_page_node(h, nid); | |
bf50bab2 NH |
1654 | spin_unlock(&hugetlb_lock); |
1655 | ||
94ae8ba7 | 1656 | if (!page) |
099730d6 | 1657 | page = __alloc_buddy_huge_page_no_mpol(h, nid); |
bf50bab2 NH |
1658 | |
1659 | return page; | |
1660 | } | |
1661 | ||
e4e574b7 | 1662 | /* |
25985edc | 1663 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1664 | * of size 'delta'. |
1665 | */ | |
a5516438 | 1666 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1667 | { |
1668 | struct list_head surplus_list; | |
1669 | struct page *page, *tmp; | |
1670 | int ret, i; | |
1671 | int needed, allocated; | |
28073b02 | 1672 | bool alloc_ok = true; |
e4e574b7 | 1673 | |
a5516438 | 1674 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1675 | if (needed <= 0) { |
a5516438 | 1676 | h->resv_huge_pages += delta; |
e4e574b7 | 1677 | return 0; |
ac09b3a1 | 1678 | } |
e4e574b7 AL |
1679 | |
1680 | allocated = 0; | |
1681 | INIT_LIST_HEAD(&surplus_list); | |
1682 | ||
1683 | ret = -ENOMEM; | |
1684 | retry: | |
1685 | spin_unlock(&hugetlb_lock); | |
1686 | for (i = 0; i < needed; i++) { | |
099730d6 | 1687 | page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE); |
28073b02 HD |
1688 | if (!page) { |
1689 | alloc_ok = false; | |
1690 | break; | |
1691 | } | |
e4e574b7 AL |
1692 | list_add(&page->lru, &surplus_list); |
1693 | } | |
28073b02 | 1694 | allocated += i; |
e4e574b7 AL |
1695 | |
1696 | /* | |
1697 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1698 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1699 | */ | |
1700 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1701 | needed = (h->resv_huge_pages + delta) - |
1702 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1703 | if (needed > 0) { |
1704 | if (alloc_ok) | |
1705 | goto retry; | |
1706 | /* | |
1707 | * We were not able to allocate enough pages to | |
1708 | * satisfy the entire reservation so we free what | |
1709 | * we've allocated so far. | |
1710 | */ | |
1711 | goto free; | |
1712 | } | |
e4e574b7 AL |
1713 | /* |
1714 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1715 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1716 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1717 | * allocator. Commit the entire reservation here to prevent another |
1718 | * process from stealing the pages as they are added to the pool but | |
1719 | * before they are reserved. | |
e4e574b7 AL |
1720 | */ |
1721 | needed += allocated; | |
a5516438 | 1722 | h->resv_huge_pages += delta; |
e4e574b7 | 1723 | ret = 0; |
a9869b83 | 1724 | |
19fc3f0a | 1725 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1726 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1727 | if ((--needed) < 0) |
1728 | break; | |
a9869b83 NH |
1729 | /* |
1730 | * This page is now managed by the hugetlb allocator and has | |
1731 | * no users -- drop the buddy allocator's reference. | |
1732 | */ | |
1733 | put_page_testzero(page); | |
309381fe | 1734 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 1735 | enqueue_huge_page(h, page); |
19fc3f0a | 1736 | } |
28073b02 | 1737 | free: |
b0365c8d | 1738 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1739 | |
1740 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1741 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1742 | put_page(page); | |
a9869b83 | 1743 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1744 | |
1745 | return ret; | |
1746 | } | |
1747 | ||
1748 | /* | |
1749 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
1750 | * allocated to satisfy the reservation must be explicitly freed if they were | |
1751 | * never used. | |
685f3457 | 1752 | * Called with hugetlb_lock held. |
e4e574b7 | 1753 | */ |
a5516438 AK |
1754 | static void return_unused_surplus_pages(struct hstate *h, |
1755 | unsigned long unused_resv_pages) | |
e4e574b7 | 1756 | { |
e4e574b7 AL |
1757 | unsigned long nr_pages; |
1758 | ||
ac09b3a1 | 1759 | /* Uncommit the reservation */ |
a5516438 | 1760 | h->resv_huge_pages -= unused_resv_pages; |
ac09b3a1 | 1761 | |
aa888a74 | 1762 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 1763 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1764 | return; |
1765 | ||
a5516438 | 1766 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1767 | |
685f3457 LS |
1768 | /* |
1769 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1770 | * evenly across all nodes with memory. Iterate across these nodes |
1771 | * until we can no longer free unreserved surplus pages. This occurs | |
1772 | * when the nodes with surplus pages have no free pages. | |
1773 | * free_pool_huge_page() will balance the the freed pages across the | |
1774 | * on-line nodes with memory and will handle the hstate accounting. | |
685f3457 LS |
1775 | */ |
1776 | while (nr_pages--) { | |
8cebfcd0 | 1777 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
685f3457 | 1778 | break; |
7848a4bf | 1779 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 AL |
1780 | } |
1781 | } | |
1782 | ||
5e911373 | 1783 | |
c37f9fb1 | 1784 | /* |
feba16e2 | 1785 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 1786 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
1787 | * |
1788 | * vma_needs_reservation is called to determine if the huge page at addr | |
1789 | * within the vma has an associated reservation. If a reservation is | |
1790 | * needed, the value 1 is returned. The caller is then responsible for | |
1791 | * managing the global reservation and subpool usage counts. After | |
1792 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
1793 | * to add the page to the reservation map. If the page allocation fails, |
1794 | * the reservation must be ended instead of committed. vma_end_reservation | |
1795 | * is called in such cases. | |
cf3ad20b MK |
1796 | * |
1797 | * In the normal case, vma_commit_reservation returns the same value | |
1798 | * as the preceding vma_needs_reservation call. The only time this | |
1799 | * is not the case is if a reserve map was changed between calls. It | |
1800 | * is the responsibility of the caller to notice the difference and | |
1801 | * take appropriate action. | |
c37f9fb1 | 1802 | */ |
5e911373 MK |
1803 | enum vma_resv_mode { |
1804 | VMA_NEEDS_RESV, | |
1805 | VMA_COMMIT_RESV, | |
feba16e2 | 1806 | VMA_END_RESV, |
5e911373 | 1807 | }; |
cf3ad20b MK |
1808 | static long __vma_reservation_common(struct hstate *h, |
1809 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 1810 | enum vma_resv_mode mode) |
c37f9fb1 | 1811 | { |
4e35f483 JK |
1812 | struct resv_map *resv; |
1813 | pgoff_t idx; | |
cf3ad20b | 1814 | long ret; |
c37f9fb1 | 1815 | |
4e35f483 JK |
1816 | resv = vma_resv_map(vma); |
1817 | if (!resv) | |
84afd99b | 1818 | return 1; |
c37f9fb1 | 1819 | |
4e35f483 | 1820 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
1821 | switch (mode) { |
1822 | case VMA_NEEDS_RESV: | |
cf3ad20b | 1823 | ret = region_chg(resv, idx, idx + 1); |
5e911373 MK |
1824 | break; |
1825 | case VMA_COMMIT_RESV: | |
1826 | ret = region_add(resv, idx, idx + 1); | |
1827 | break; | |
feba16e2 | 1828 | case VMA_END_RESV: |
5e911373 MK |
1829 | region_abort(resv, idx, idx + 1); |
1830 | ret = 0; | |
1831 | break; | |
1832 | default: | |
1833 | BUG(); | |
1834 | } | |
84afd99b | 1835 | |
4e35f483 | 1836 | if (vma->vm_flags & VM_MAYSHARE) |
cf3ad20b | 1837 | return ret; |
67961f9d MK |
1838 | else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { |
1839 | /* | |
1840 | * In most cases, reserves always exist for private mappings. | |
1841 | * However, a file associated with mapping could have been | |
1842 | * hole punched or truncated after reserves were consumed. | |
1843 | * As subsequent fault on such a range will not use reserves. | |
1844 | * Subtle - The reserve map for private mappings has the | |
1845 | * opposite meaning than that of shared mappings. If NO | |
1846 | * entry is in the reserve map, it means a reservation exists. | |
1847 | * If an entry exists in the reserve map, it means the | |
1848 | * reservation has already been consumed. As a result, the | |
1849 | * return value of this routine is the opposite of the | |
1850 | * value returned from reserve map manipulation routines above. | |
1851 | */ | |
1852 | if (ret) | |
1853 | return 0; | |
1854 | else | |
1855 | return 1; | |
1856 | } | |
4e35f483 | 1857 | else |
cf3ad20b | 1858 | return ret < 0 ? ret : 0; |
c37f9fb1 | 1859 | } |
cf3ad20b MK |
1860 | |
1861 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 1862 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 1863 | { |
5e911373 | 1864 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 1865 | } |
84afd99b | 1866 | |
cf3ad20b MK |
1867 | static long vma_commit_reservation(struct hstate *h, |
1868 | struct vm_area_struct *vma, unsigned long addr) | |
1869 | { | |
5e911373 MK |
1870 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
1871 | } | |
1872 | ||
feba16e2 | 1873 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
1874 | struct vm_area_struct *vma, unsigned long addr) |
1875 | { | |
feba16e2 | 1876 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
1877 | } |
1878 | ||
70c3547e | 1879 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 1880 | unsigned long addr, int avoid_reserve) |
1da177e4 | 1881 | { |
90481622 | 1882 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 1883 | struct hstate *h = hstate_vma(vma); |
348ea204 | 1884 | struct page *page; |
d85f69b0 MK |
1885 | long map_chg, map_commit; |
1886 | long gbl_chg; | |
6d76dcf4 AK |
1887 | int ret, idx; |
1888 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 1889 | |
6d76dcf4 | 1890 | idx = hstate_index(h); |
a1e78772 | 1891 | /* |
d85f69b0 MK |
1892 | * Examine the region/reserve map to determine if the process |
1893 | * has a reservation for the page to be allocated. A return | |
1894 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 1895 | */ |
d85f69b0 MK |
1896 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
1897 | if (map_chg < 0) | |
76dcee75 | 1898 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
1899 | |
1900 | /* | |
1901 | * Processes that did not create the mapping will have no | |
1902 | * reserves as indicated by the region/reserve map. Check | |
1903 | * that the allocation will not exceed the subpool limit. | |
1904 | * Allocations for MAP_NORESERVE mappings also need to be | |
1905 | * checked against any subpool limit. | |
1906 | */ | |
1907 | if (map_chg || avoid_reserve) { | |
1908 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
1909 | if (gbl_chg < 0) { | |
feba16e2 | 1910 | vma_end_reservation(h, vma, addr); |
76dcee75 | 1911 | return ERR_PTR(-ENOSPC); |
5e911373 | 1912 | } |
1da177e4 | 1913 | |
d85f69b0 MK |
1914 | /* |
1915 | * Even though there was no reservation in the region/reserve | |
1916 | * map, there could be reservations associated with the | |
1917 | * subpool that can be used. This would be indicated if the | |
1918 | * return value of hugepage_subpool_get_pages() is zero. | |
1919 | * However, if avoid_reserve is specified we still avoid even | |
1920 | * the subpool reservations. | |
1921 | */ | |
1922 | if (avoid_reserve) | |
1923 | gbl_chg = 1; | |
1924 | } | |
1925 | ||
6d76dcf4 | 1926 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f JZ |
1927 | if (ret) |
1928 | goto out_subpool_put; | |
1929 | ||
1da177e4 | 1930 | spin_lock(&hugetlb_lock); |
d85f69b0 MK |
1931 | /* |
1932 | * glb_chg is passed to indicate whether or not a page must be taken | |
1933 | * from the global free pool (global change). gbl_chg == 0 indicates | |
1934 | * a reservation exists for the allocation. | |
1935 | */ | |
1936 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 1937 | if (!page) { |
94ae8ba7 | 1938 | spin_unlock(&hugetlb_lock); |
099730d6 | 1939 | page = __alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
1940 | if (!page) |
1941 | goto out_uncharge_cgroup; | |
a88c7695 NH |
1942 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
1943 | SetPagePrivate(page); | |
1944 | h->resv_huge_pages--; | |
1945 | } | |
79dbb236 AK |
1946 | spin_lock(&hugetlb_lock); |
1947 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 1948 | /* Fall through */ |
68842c9b | 1949 | } |
81a6fcae JK |
1950 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
1951 | spin_unlock(&hugetlb_lock); | |
348ea204 | 1952 | |
90481622 | 1953 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 1954 | |
d85f69b0 MK |
1955 | map_commit = vma_commit_reservation(h, vma, addr); |
1956 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
1957 | /* |
1958 | * The page was added to the reservation map between | |
1959 | * vma_needs_reservation and vma_commit_reservation. | |
1960 | * This indicates a race with hugetlb_reserve_pages. | |
1961 | * Adjust for the subpool count incremented above AND | |
1962 | * in hugetlb_reserve_pages for the same page. Also, | |
1963 | * the reservation count added in hugetlb_reserve_pages | |
1964 | * no longer applies. | |
1965 | */ | |
1966 | long rsv_adjust; | |
1967 | ||
1968 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
1969 | hugetlb_acct_memory(h, -rsv_adjust); | |
1970 | } | |
90d8b7e6 | 1971 | return page; |
8f34af6f JZ |
1972 | |
1973 | out_uncharge_cgroup: | |
1974 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
1975 | out_subpool_put: | |
d85f69b0 | 1976 | if (map_chg || avoid_reserve) |
8f34af6f | 1977 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 1978 | vma_end_reservation(h, vma, addr); |
8f34af6f | 1979 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
1980 | } |
1981 | ||
74060e4d NH |
1982 | /* |
1983 | * alloc_huge_page()'s wrapper which simply returns the page if allocation | |
1984 | * succeeds, otherwise NULL. This function is called from new_vma_page(), | |
1985 | * where no ERR_VALUE is expected to be returned. | |
1986 | */ | |
1987 | struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, | |
1988 | unsigned long addr, int avoid_reserve) | |
1989 | { | |
1990 | struct page *page = alloc_huge_page(vma, addr, avoid_reserve); | |
1991 | if (IS_ERR(page)) | |
1992 | page = NULL; | |
1993 | return page; | |
1994 | } | |
1995 | ||
91f47662 | 1996 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
aa888a74 AK |
1997 | { |
1998 | struct huge_bootmem_page *m; | |
b2261026 | 1999 | int nr_nodes, node; |
aa888a74 | 2000 | |
b2261026 | 2001 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
2002 | void *addr; |
2003 | ||
8b89a116 GS |
2004 | addr = memblock_virt_alloc_try_nid_nopanic( |
2005 | huge_page_size(h), huge_page_size(h), | |
2006 | 0, BOOTMEM_ALLOC_ACCESSIBLE, node); | |
aa888a74 AK |
2007 | if (addr) { |
2008 | /* | |
2009 | * Use the beginning of the huge page to store the | |
2010 | * huge_bootmem_page struct (until gather_bootmem | |
2011 | * puts them into the mem_map). | |
2012 | */ | |
2013 | m = addr; | |
91f47662 | 2014 | goto found; |
aa888a74 | 2015 | } |
aa888a74 AK |
2016 | } |
2017 | return 0; | |
2018 | ||
2019 | found: | |
df994ead | 2020 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 AK |
2021 | /* Put them into a private list first because mem_map is not up yet */ |
2022 | list_add(&m->list, &huge_boot_pages); | |
2023 | m->hstate = h; | |
2024 | return 1; | |
2025 | } | |
2026 | ||
d00181b9 KS |
2027 | static void __init prep_compound_huge_page(struct page *page, |
2028 | unsigned int order) | |
18229df5 AW |
2029 | { |
2030 | if (unlikely(order > (MAX_ORDER - 1))) | |
2031 | prep_compound_gigantic_page(page, order); | |
2032 | else | |
2033 | prep_compound_page(page, order); | |
2034 | } | |
2035 | ||
aa888a74 AK |
2036 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
2037 | static void __init gather_bootmem_prealloc(void) | |
2038 | { | |
2039 | struct huge_bootmem_page *m; | |
2040 | ||
2041 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 2042 | struct hstate *h = m->hstate; |
ee8f248d BB |
2043 | struct page *page; |
2044 | ||
2045 | #ifdef CONFIG_HIGHMEM | |
2046 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
8b89a116 GS |
2047 | memblock_free_late(__pa(m), |
2048 | sizeof(struct huge_bootmem_page)); | |
ee8f248d BB |
2049 | #else |
2050 | page = virt_to_page(m); | |
2051 | #endif | |
aa888a74 | 2052 | WARN_ON(page_count(page) != 1); |
18229df5 | 2053 | prep_compound_huge_page(page, h->order); |
ef5a22be | 2054 | WARN_ON(PageReserved(page)); |
aa888a74 | 2055 | prep_new_huge_page(h, page, page_to_nid(page)); |
b0320c7b RA |
2056 | /* |
2057 | * If we had gigantic hugepages allocated at boot time, we need | |
2058 | * to restore the 'stolen' pages to totalram_pages in order to | |
2059 | * fix confusing memory reports from free(1) and another | |
2060 | * side-effects, like CommitLimit going negative. | |
2061 | */ | |
bae7f4ae | 2062 | if (hstate_is_gigantic(h)) |
3dcc0571 | 2063 | adjust_managed_page_count(page, 1 << h->order); |
aa888a74 AK |
2064 | } |
2065 | } | |
2066 | ||
8faa8b07 | 2067 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
2068 | { |
2069 | unsigned long i; | |
a5516438 | 2070 | |
e5ff2159 | 2071 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 2072 | if (hstate_is_gigantic(h)) { |
aa888a74 AK |
2073 | if (!alloc_bootmem_huge_page(h)) |
2074 | break; | |
9b5e5d0f | 2075 | } else if (!alloc_fresh_huge_page(h, |
8cebfcd0 | 2076 | &node_states[N_MEMORY])) |
1da177e4 | 2077 | break; |
1da177e4 | 2078 | } |
8faa8b07 | 2079 | h->max_huge_pages = i; |
e5ff2159 AK |
2080 | } |
2081 | ||
2082 | static void __init hugetlb_init_hstates(void) | |
2083 | { | |
2084 | struct hstate *h; | |
2085 | ||
2086 | for_each_hstate(h) { | |
641844f5 NH |
2087 | if (minimum_order > huge_page_order(h)) |
2088 | minimum_order = huge_page_order(h); | |
2089 | ||
8faa8b07 | 2090 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 2091 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 2092 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 2093 | } |
641844f5 | 2094 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
2095 | } |
2096 | ||
4abd32db AK |
2097 | static char * __init memfmt(char *buf, unsigned long n) |
2098 | { | |
2099 | if (n >= (1UL << 30)) | |
2100 | sprintf(buf, "%lu GB", n >> 30); | |
2101 | else if (n >= (1UL << 20)) | |
2102 | sprintf(buf, "%lu MB", n >> 20); | |
2103 | else | |
2104 | sprintf(buf, "%lu KB", n >> 10); | |
2105 | return buf; | |
2106 | } | |
2107 | ||
e5ff2159 AK |
2108 | static void __init report_hugepages(void) |
2109 | { | |
2110 | struct hstate *h; | |
2111 | ||
2112 | for_each_hstate(h) { | |
4abd32db | 2113 | char buf[32]; |
ffb22af5 | 2114 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
4abd32db AK |
2115 | memfmt(buf, huge_page_size(h)), |
2116 | h->free_huge_pages); | |
e5ff2159 AK |
2117 | } |
2118 | } | |
2119 | ||
1da177e4 | 2120 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
2121 | static void try_to_free_low(struct hstate *h, unsigned long count, |
2122 | nodemask_t *nodes_allowed) | |
1da177e4 | 2123 | { |
4415cc8d CL |
2124 | int i; |
2125 | ||
bae7f4ae | 2126 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2127 | return; |
2128 | ||
6ae11b27 | 2129 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 2130 | struct page *page, *next; |
a5516438 AK |
2131 | struct list_head *freel = &h->hugepage_freelists[i]; |
2132 | list_for_each_entry_safe(page, next, freel, lru) { | |
2133 | if (count >= h->nr_huge_pages) | |
6b0c880d | 2134 | return; |
1da177e4 LT |
2135 | if (PageHighMem(page)) |
2136 | continue; | |
2137 | list_del(&page->lru); | |
e5ff2159 | 2138 | update_and_free_page(h, page); |
a5516438 AK |
2139 | h->free_huge_pages--; |
2140 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
2141 | } |
2142 | } | |
2143 | } | |
2144 | #else | |
6ae11b27 LS |
2145 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
2146 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
2147 | { |
2148 | } | |
2149 | #endif | |
2150 | ||
20a0307c WF |
2151 | /* |
2152 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2153 | * balanced by operating on them in a round-robin fashion. | |
2154 | * Returns 1 if an adjustment was made. | |
2155 | */ | |
6ae11b27 LS |
2156 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
2157 | int delta) | |
20a0307c | 2158 | { |
b2261026 | 2159 | int nr_nodes, node; |
20a0307c WF |
2160 | |
2161 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 2162 | |
b2261026 JK |
2163 | if (delta < 0) { |
2164 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2165 | if (h->surplus_huge_pages_node[node]) | |
2166 | goto found; | |
e8c5c824 | 2167 | } |
b2261026 JK |
2168 | } else { |
2169 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2170 | if (h->surplus_huge_pages_node[node] < | |
2171 | h->nr_huge_pages_node[node]) | |
2172 | goto found; | |
e8c5c824 | 2173 | } |
b2261026 JK |
2174 | } |
2175 | return 0; | |
20a0307c | 2176 | |
b2261026 JK |
2177 | found: |
2178 | h->surplus_huge_pages += delta; | |
2179 | h->surplus_huge_pages_node[node] += delta; | |
2180 | return 1; | |
20a0307c WF |
2181 | } |
2182 | ||
a5516438 | 2183 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
2184 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
2185 | nodemask_t *nodes_allowed) | |
1da177e4 | 2186 | { |
7893d1d5 | 2187 | unsigned long min_count, ret; |
1da177e4 | 2188 | |
944d9fec | 2189 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
aa888a74 AK |
2190 | return h->max_huge_pages; |
2191 | ||
7893d1d5 AL |
2192 | /* |
2193 | * Increase the pool size | |
2194 | * First take pages out of surplus state. Then make up the | |
2195 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 2196 | * |
d15c7c09 | 2197 | * We might race with __alloc_buddy_huge_page() here and be unable |
d1c3fb1f NA |
2198 | * to convert a surplus huge page to a normal huge page. That is |
2199 | * not critical, though, it just means the overall size of the | |
2200 | * pool might be one hugepage larger than it needs to be, but | |
2201 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 2202 | */ |
1da177e4 | 2203 | spin_lock(&hugetlb_lock); |
a5516438 | 2204 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 2205 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
2206 | break; |
2207 | } | |
2208 | ||
a5516438 | 2209 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
2210 | /* |
2211 | * If this allocation races such that we no longer need the | |
2212 | * page, free_huge_page will handle it by freeing the page | |
2213 | * and reducing the surplus. | |
2214 | */ | |
2215 | spin_unlock(&hugetlb_lock); | |
944d9fec LC |
2216 | if (hstate_is_gigantic(h)) |
2217 | ret = alloc_fresh_gigantic_page(h, nodes_allowed); | |
2218 | else | |
2219 | ret = alloc_fresh_huge_page(h, nodes_allowed); | |
7893d1d5 AL |
2220 | spin_lock(&hugetlb_lock); |
2221 | if (!ret) | |
2222 | goto out; | |
2223 | ||
536240f2 MG |
2224 | /* Bail for signals. Probably ctrl-c from user */ |
2225 | if (signal_pending(current)) | |
2226 | goto out; | |
7893d1d5 | 2227 | } |
7893d1d5 AL |
2228 | |
2229 | /* | |
2230 | * Decrease the pool size | |
2231 | * First return free pages to the buddy allocator (being careful | |
2232 | * to keep enough around to satisfy reservations). Then place | |
2233 | * pages into surplus state as needed so the pool will shrink | |
2234 | * to the desired size as pages become free. | |
d1c3fb1f NA |
2235 | * |
2236 | * By placing pages into the surplus state independent of the | |
2237 | * overcommit value, we are allowing the surplus pool size to | |
2238 | * exceed overcommit. There are few sane options here. Since | |
d15c7c09 | 2239 | * __alloc_buddy_huge_page() is checking the global counter, |
d1c3fb1f NA |
2240 | * though, we'll note that we're not allowed to exceed surplus |
2241 | * and won't grow the pool anywhere else. Not until one of the | |
2242 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 2243 | */ |
a5516438 | 2244 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 2245 | min_count = max(count, min_count); |
6ae11b27 | 2246 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 2247 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 2248 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 2249 | break; |
55f67141 | 2250 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 2251 | } |
a5516438 | 2252 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 2253 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
2254 | break; |
2255 | } | |
2256 | out: | |
a5516438 | 2257 | ret = persistent_huge_pages(h); |
1da177e4 | 2258 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 2259 | return ret; |
1da177e4 LT |
2260 | } |
2261 | ||
a3437870 NA |
2262 | #define HSTATE_ATTR_RO(_name) \ |
2263 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2264 | ||
2265 | #define HSTATE_ATTR(_name) \ | |
2266 | static struct kobj_attribute _name##_attr = \ | |
2267 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2268 | ||
2269 | static struct kobject *hugepages_kobj; | |
2270 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2271 | ||
9a305230 LS |
2272 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
2273 | ||
2274 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
2275 | { |
2276 | int i; | |
9a305230 | 2277 | |
a3437870 | 2278 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
2279 | if (hstate_kobjs[i] == kobj) { |
2280 | if (nidp) | |
2281 | *nidp = NUMA_NO_NODE; | |
a3437870 | 2282 | return &hstates[i]; |
9a305230 LS |
2283 | } |
2284 | ||
2285 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
2286 | } |
2287 | ||
06808b08 | 2288 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
2289 | struct kobj_attribute *attr, char *buf) |
2290 | { | |
9a305230 LS |
2291 | struct hstate *h; |
2292 | unsigned long nr_huge_pages; | |
2293 | int nid; | |
2294 | ||
2295 | h = kobj_to_hstate(kobj, &nid); | |
2296 | if (nid == NUMA_NO_NODE) | |
2297 | nr_huge_pages = h->nr_huge_pages; | |
2298 | else | |
2299 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2300 | ||
2301 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 2302 | } |
adbe8726 | 2303 | |
238d3c13 DR |
2304 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
2305 | struct hstate *h, int nid, | |
2306 | unsigned long count, size_t len) | |
a3437870 NA |
2307 | { |
2308 | int err; | |
bad44b5b | 2309 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 2310 | |
944d9fec | 2311 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) { |
adbe8726 EM |
2312 | err = -EINVAL; |
2313 | goto out; | |
2314 | } | |
2315 | ||
9a305230 LS |
2316 | if (nid == NUMA_NO_NODE) { |
2317 | /* | |
2318 | * global hstate attribute | |
2319 | */ | |
2320 | if (!(obey_mempolicy && | |
2321 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
2322 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 2323 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 LS |
2324 | } |
2325 | } else if (nodes_allowed) { | |
2326 | /* | |
2327 | * per node hstate attribute: adjust count to global, | |
2328 | * but restrict alloc/free to the specified node. | |
2329 | */ | |
2330 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2331 | init_nodemask_of_node(nodes_allowed, nid); | |
2332 | } else | |
8cebfcd0 | 2333 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 | 2334 | |
06808b08 | 2335 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 2336 | |
8cebfcd0 | 2337 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
2338 | NODEMASK_FREE(nodes_allowed); |
2339 | ||
2340 | return len; | |
adbe8726 EM |
2341 | out: |
2342 | NODEMASK_FREE(nodes_allowed); | |
2343 | return err; | |
06808b08 LS |
2344 | } |
2345 | ||
238d3c13 DR |
2346 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
2347 | struct kobject *kobj, const char *buf, | |
2348 | size_t len) | |
2349 | { | |
2350 | struct hstate *h; | |
2351 | unsigned long count; | |
2352 | int nid; | |
2353 | int err; | |
2354 | ||
2355 | err = kstrtoul(buf, 10, &count); | |
2356 | if (err) | |
2357 | return err; | |
2358 | ||
2359 | h = kobj_to_hstate(kobj, &nid); | |
2360 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2361 | } | |
2362 | ||
06808b08 LS |
2363 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
2364 | struct kobj_attribute *attr, char *buf) | |
2365 | { | |
2366 | return nr_hugepages_show_common(kobj, attr, buf); | |
2367 | } | |
2368 | ||
2369 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2370 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2371 | { | |
238d3c13 | 2372 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
2373 | } |
2374 | HSTATE_ATTR(nr_hugepages); | |
2375 | ||
06808b08 LS |
2376 | #ifdef CONFIG_NUMA |
2377 | ||
2378 | /* | |
2379 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2380 | * huge page alloc/free. | |
2381 | */ | |
2382 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
2383 | struct kobj_attribute *attr, char *buf) | |
2384 | { | |
2385 | return nr_hugepages_show_common(kobj, attr, buf); | |
2386 | } | |
2387 | ||
2388 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2389 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2390 | { | |
238d3c13 | 2391 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
2392 | } |
2393 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2394 | #endif | |
2395 | ||
2396 | ||
a3437870 NA |
2397 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
2398 | struct kobj_attribute *attr, char *buf) | |
2399 | { | |
9a305230 | 2400 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2401 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
2402 | } | |
adbe8726 | 2403 | |
a3437870 NA |
2404 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
2405 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2406 | { | |
2407 | int err; | |
2408 | unsigned long input; | |
9a305230 | 2409 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2410 | |
bae7f4ae | 2411 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2412 | return -EINVAL; |
2413 | ||
3dbb95f7 | 2414 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2415 | if (err) |
73ae31e5 | 2416 | return err; |
a3437870 NA |
2417 | |
2418 | spin_lock(&hugetlb_lock); | |
2419 | h->nr_overcommit_huge_pages = input; | |
2420 | spin_unlock(&hugetlb_lock); | |
2421 | ||
2422 | return count; | |
2423 | } | |
2424 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2425 | ||
2426 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2427 | struct kobj_attribute *attr, char *buf) | |
2428 | { | |
9a305230 LS |
2429 | struct hstate *h; |
2430 | unsigned long free_huge_pages; | |
2431 | int nid; | |
2432 | ||
2433 | h = kobj_to_hstate(kobj, &nid); | |
2434 | if (nid == NUMA_NO_NODE) | |
2435 | free_huge_pages = h->free_huge_pages; | |
2436 | else | |
2437 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2438 | ||
2439 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
2440 | } |
2441 | HSTATE_ATTR_RO(free_hugepages); | |
2442 | ||
2443 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2444 | struct kobj_attribute *attr, char *buf) | |
2445 | { | |
9a305230 | 2446 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2447 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
2448 | } | |
2449 | HSTATE_ATTR_RO(resv_hugepages); | |
2450 | ||
2451 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2452 | struct kobj_attribute *attr, char *buf) | |
2453 | { | |
9a305230 LS |
2454 | struct hstate *h; |
2455 | unsigned long surplus_huge_pages; | |
2456 | int nid; | |
2457 | ||
2458 | h = kobj_to_hstate(kobj, &nid); | |
2459 | if (nid == NUMA_NO_NODE) | |
2460 | surplus_huge_pages = h->surplus_huge_pages; | |
2461 | else | |
2462 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2463 | ||
2464 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
2465 | } |
2466 | HSTATE_ATTR_RO(surplus_hugepages); | |
2467 | ||
2468 | static struct attribute *hstate_attrs[] = { | |
2469 | &nr_hugepages_attr.attr, | |
2470 | &nr_overcommit_hugepages_attr.attr, | |
2471 | &free_hugepages_attr.attr, | |
2472 | &resv_hugepages_attr.attr, | |
2473 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
2474 | #ifdef CONFIG_NUMA |
2475 | &nr_hugepages_mempolicy_attr.attr, | |
2476 | #endif | |
a3437870 NA |
2477 | NULL, |
2478 | }; | |
2479 | ||
2480 | static struct attribute_group hstate_attr_group = { | |
2481 | .attrs = hstate_attrs, | |
2482 | }; | |
2483 | ||
094e9539 JM |
2484 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
2485 | struct kobject **hstate_kobjs, | |
2486 | struct attribute_group *hstate_attr_group) | |
a3437870 NA |
2487 | { |
2488 | int retval; | |
972dc4de | 2489 | int hi = hstate_index(h); |
a3437870 | 2490 | |
9a305230 LS |
2491 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
2492 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
2493 | return -ENOMEM; |
2494 | ||
9a305230 | 2495 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 2496 | if (retval) |
9a305230 | 2497 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
2498 | |
2499 | return retval; | |
2500 | } | |
2501 | ||
2502 | static void __init hugetlb_sysfs_init(void) | |
2503 | { | |
2504 | struct hstate *h; | |
2505 | int err; | |
2506 | ||
2507 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
2508 | if (!hugepages_kobj) | |
2509 | return; | |
2510 | ||
2511 | for_each_hstate(h) { | |
9a305230 LS |
2512 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
2513 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 2514 | if (err) |
ffb22af5 | 2515 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
2516 | } |
2517 | } | |
2518 | ||
9a305230 LS |
2519 | #ifdef CONFIG_NUMA |
2520 | ||
2521 | /* | |
2522 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
2523 | * with node devices in node_devices[] using a parallel array. The array |
2524 | * index of a node device or _hstate == node id. | |
2525 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
2526 | * the base kernel, on the hugetlb module. |
2527 | */ | |
2528 | struct node_hstate { | |
2529 | struct kobject *hugepages_kobj; | |
2530 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2531 | }; | |
b4e289a6 | 2532 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
2533 | |
2534 | /* | |
10fbcf4c | 2535 | * A subset of global hstate attributes for node devices |
9a305230 LS |
2536 | */ |
2537 | static struct attribute *per_node_hstate_attrs[] = { | |
2538 | &nr_hugepages_attr.attr, | |
2539 | &free_hugepages_attr.attr, | |
2540 | &surplus_hugepages_attr.attr, | |
2541 | NULL, | |
2542 | }; | |
2543 | ||
2544 | static struct attribute_group per_node_hstate_attr_group = { | |
2545 | .attrs = per_node_hstate_attrs, | |
2546 | }; | |
2547 | ||
2548 | /* | |
10fbcf4c | 2549 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
2550 | * Returns node id via non-NULL nidp. |
2551 | */ | |
2552 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2553 | { | |
2554 | int nid; | |
2555 | ||
2556 | for (nid = 0; nid < nr_node_ids; nid++) { | |
2557 | struct node_hstate *nhs = &node_hstates[nid]; | |
2558 | int i; | |
2559 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
2560 | if (nhs->hstate_kobjs[i] == kobj) { | |
2561 | if (nidp) | |
2562 | *nidp = nid; | |
2563 | return &hstates[i]; | |
2564 | } | |
2565 | } | |
2566 | ||
2567 | BUG(); | |
2568 | return NULL; | |
2569 | } | |
2570 | ||
2571 | /* | |
10fbcf4c | 2572 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
2573 | * No-op if no hstate attributes attached. |
2574 | */ | |
3cd8b44f | 2575 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
2576 | { |
2577 | struct hstate *h; | |
10fbcf4c | 2578 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2579 | |
2580 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 2581 | return; /* no hstate attributes */ |
9a305230 | 2582 | |
972dc4de AK |
2583 | for_each_hstate(h) { |
2584 | int idx = hstate_index(h); | |
2585 | if (nhs->hstate_kobjs[idx]) { | |
2586 | kobject_put(nhs->hstate_kobjs[idx]); | |
2587 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 2588 | } |
972dc4de | 2589 | } |
9a305230 LS |
2590 | |
2591 | kobject_put(nhs->hugepages_kobj); | |
2592 | nhs->hugepages_kobj = NULL; | |
2593 | } | |
2594 | ||
9a305230 LS |
2595 | |
2596 | /* | |
10fbcf4c | 2597 | * Register hstate attributes for a single node device. |
9a305230 LS |
2598 | * No-op if attributes already registered. |
2599 | */ | |
3cd8b44f | 2600 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
2601 | { |
2602 | struct hstate *h; | |
10fbcf4c | 2603 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2604 | int err; |
2605 | ||
2606 | if (nhs->hugepages_kobj) | |
2607 | return; /* already allocated */ | |
2608 | ||
2609 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 2610 | &node->dev.kobj); |
9a305230 LS |
2611 | if (!nhs->hugepages_kobj) |
2612 | return; | |
2613 | ||
2614 | for_each_hstate(h) { | |
2615 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
2616 | nhs->hstate_kobjs, | |
2617 | &per_node_hstate_attr_group); | |
2618 | if (err) { | |
ffb22af5 AM |
2619 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
2620 | h->name, node->dev.id); | |
9a305230 LS |
2621 | hugetlb_unregister_node(node); |
2622 | break; | |
2623 | } | |
2624 | } | |
2625 | } | |
2626 | ||
2627 | /* | |
9b5e5d0f | 2628 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
2629 | * devices of nodes that have memory. All on-line nodes should have |
2630 | * registered their associated device by this time. | |
9a305230 | 2631 | */ |
7d9ca000 | 2632 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
2633 | { |
2634 | int nid; | |
2635 | ||
8cebfcd0 | 2636 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 2637 | struct node *node = node_devices[nid]; |
10fbcf4c | 2638 | if (node->dev.id == nid) |
9a305230 LS |
2639 | hugetlb_register_node(node); |
2640 | } | |
2641 | ||
2642 | /* | |
10fbcf4c | 2643 | * Let the node device driver know we're here so it can |
9a305230 LS |
2644 | * [un]register hstate attributes on node hotplug. |
2645 | */ | |
2646 | register_hugetlbfs_with_node(hugetlb_register_node, | |
2647 | hugetlb_unregister_node); | |
2648 | } | |
2649 | #else /* !CONFIG_NUMA */ | |
2650 | ||
2651 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2652 | { | |
2653 | BUG(); | |
2654 | if (nidp) | |
2655 | *nidp = -1; | |
2656 | return NULL; | |
2657 | } | |
2658 | ||
9a305230 LS |
2659 | static void hugetlb_register_all_nodes(void) { } |
2660 | ||
2661 | #endif | |
2662 | ||
a3437870 NA |
2663 | static int __init hugetlb_init(void) |
2664 | { | |
8382d914 DB |
2665 | int i; |
2666 | ||
457c1b27 | 2667 | if (!hugepages_supported()) |
0ef89d25 | 2668 | return 0; |
a3437870 | 2669 | |
e11bfbfc NP |
2670 | if (!size_to_hstate(default_hstate_size)) { |
2671 | default_hstate_size = HPAGE_SIZE; | |
2672 | if (!size_to_hstate(default_hstate_size)) | |
2673 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 2674 | } |
972dc4de | 2675 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
f8b74815 VT |
2676 | if (default_hstate_max_huge_pages) { |
2677 | if (!default_hstate.max_huge_pages) | |
2678 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
2679 | } | |
a3437870 NA |
2680 | |
2681 | hugetlb_init_hstates(); | |
aa888a74 | 2682 | gather_bootmem_prealloc(); |
a3437870 NA |
2683 | report_hugepages(); |
2684 | ||
2685 | hugetlb_sysfs_init(); | |
9a305230 | 2686 | hugetlb_register_all_nodes(); |
7179e7bf | 2687 | hugetlb_cgroup_file_init(); |
9a305230 | 2688 | |
8382d914 DB |
2689 | #ifdef CONFIG_SMP |
2690 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
2691 | #else | |
2692 | num_fault_mutexes = 1; | |
2693 | #endif | |
c672c7f2 | 2694 | hugetlb_fault_mutex_table = |
8382d914 | 2695 | kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); |
c672c7f2 | 2696 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
2697 | |
2698 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 2699 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
2700 | return 0; |
2701 | } | |
3e89e1c5 | 2702 | subsys_initcall(hugetlb_init); |
a3437870 NA |
2703 | |
2704 | /* Should be called on processing a hugepagesz=... option */ | |
9fee021d VT |
2705 | void __init hugetlb_bad_size(void) |
2706 | { | |
2707 | parsed_valid_hugepagesz = false; | |
2708 | } | |
2709 | ||
d00181b9 | 2710 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
2711 | { |
2712 | struct hstate *h; | |
8faa8b07 AK |
2713 | unsigned long i; |
2714 | ||
a3437870 | 2715 | if (size_to_hstate(PAGE_SIZE << order)) { |
598d8091 | 2716 | pr_warn("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
2717 | return; |
2718 | } | |
47d38344 | 2719 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 2720 | BUG_ON(order == 0); |
47d38344 | 2721 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
2722 | h->order = order; |
2723 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
2724 | h->nr_huge_pages = 0; |
2725 | h->free_huge_pages = 0; | |
2726 | for (i = 0; i < MAX_NUMNODES; ++i) | |
2727 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 2728 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
2729 | h->next_nid_to_alloc = first_memory_node; |
2730 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
2731 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
2732 | huge_page_size(h)/1024); | |
8faa8b07 | 2733 | |
a3437870 NA |
2734 | parsed_hstate = h; |
2735 | } | |
2736 | ||
e11bfbfc | 2737 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
2738 | { |
2739 | unsigned long *mhp; | |
8faa8b07 | 2740 | static unsigned long *last_mhp; |
a3437870 | 2741 | |
9fee021d VT |
2742 | if (!parsed_valid_hugepagesz) { |
2743 | pr_warn("hugepages = %s preceded by " | |
2744 | "an unsupported hugepagesz, ignoring\n", s); | |
2745 | parsed_valid_hugepagesz = true; | |
2746 | return 1; | |
2747 | } | |
a3437870 | 2748 | /* |
47d38344 | 2749 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
2750 | * so this hugepages= parameter goes to the "default hstate". |
2751 | */ | |
9fee021d | 2752 | else if (!hugetlb_max_hstate) |
a3437870 NA |
2753 | mhp = &default_hstate_max_huge_pages; |
2754 | else | |
2755 | mhp = &parsed_hstate->max_huge_pages; | |
2756 | ||
8faa8b07 | 2757 | if (mhp == last_mhp) { |
598d8091 | 2758 | pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); |
8faa8b07 AK |
2759 | return 1; |
2760 | } | |
2761 | ||
a3437870 NA |
2762 | if (sscanf(s, "%lu", mhp) <= 0) |
2763 | *mhp = 0; | |
2764 | ||
8faa8b07 AK |
2765 | /* |
2766 | * Global state is always initialized later in hugetlb_init. | |
2767 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
2768 | * use the bootmem allocator. | |
2769 | */ | |
47d38344 | 2770 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
2771 | hugetlb_hstate_alloc_pages(parsed_hstate); |
2772 | ||
2773 | last_mhp = mhp; | |
2774 | ||
a3437870 NA |
2775 | return 1; |
2776 | } | |
e11bfbfc NP |
2777 | __setup("hugepages=", hugetlb_nrpages_setup); |
2778 | ||
2779 | static int __init hugetlb_default_setup(char *s) | |
2780 | { | |
2781 | default_hstate_size = memparse(s, &s); | |
2782 | return 1; | |
2783 | } | |
2784 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 2785 | |
8a213460 NA |
2786 | static unsigned int cpuset_mems_nr(unsigned int *array) |
2787 | { | |
2788 | int node; | |
2789 | unsigned int nr = 0; | |
2790 | ||
2791 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
2792 | nr += array[node]; | |
2793 | ||
2794 | return nr; | |
2795 | } | |
2796 | ||
2797 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
2798 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
2799 | struct ctl_table *table, int write, | |
2800 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 2801 | { |
e5ff2159 | 2802 | struct hstate *h = &default_hstate; |
238d3c13 | 2803 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 2804 | int ret; |
e5ff2159 | 2805 | |
457c1b27 | 2806 | if (!hugepages_supported()) |
86613628 | 2807 | return -EOPNOTSUPP; |
457c1b27 | 2808 | |
e5ff2159 AK |
2809 | table->data = &tmp; |
2810 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2811 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2812 | if (ret) | |
2813 | goto out; | |
e5ff2159 | 2814 | |
238d3c13 DR |
2815 | if (write) |
2816 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
2817 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
2818 | out: |
2819 | return ret; | |
1da177e4 | 2820 | } |
396faf03 | 2821 | |
06808b08 LS |
2822 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2823 | void __user *buffer, size_t *length, loff_t *ppos) | |
2824 | { | |
2825 | ||
2826 | return hugetlb_sysctl_handler_common(false, table, write, | |
2827 | buffer, length, ppos); | |
2828 | } | |
2829 | ||
2830 | #ifdef CONFIG_NUMA | |
2831 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2832 | void __user *buffer, size_t *length, loff_t *ppos) | |
2833 | { | |
2834 | return hugetlb_sysctl_handler_common(true, table, write, | |
2835 | buffer, length, ppos); | |
2836 | } | |
2837 | #endif /* CONFIG_NUMA */ | |
2838 | ||
a3d0c6aa | 2839 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 2840 | void __user *buffer, |
a3d0c6aa NA |
2841 | size_t *length, loff_t *ppos) |
2842 | { | |
a5516438 | 2843 | struct hstate *h = &default_hstate; |
e5ff2159 | 2844 | unsigned long tmp; |
08d4a246 | 2845 | int ret; |
e5ff2159 | 2846 | |
457c1b27 | 2847 | if (!hugepages_supported()) |
86613628 | 2848 | return -EOPNOTSUPP; |
457c1b27 | 2849 | |
c033a93c | 2850 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 2851 | |
bae7f4ae | 2852 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
2853 | return -EINVAL; |
2854 | ||
e5ff2159 AK |
2855 | table->data = &tmp; |
2856 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2857 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2858 | if (ret) | |
2859 | goto out; | |
e5ff2159 AK |
2860 | |
2861 | if (write) { | |
2862 | spin_lock(&hugetlb_lock); | |
2863 | h->nr_overcommit_huge_pages = tmp; | |
2864 | spin_unlock(&hugetlb_lock); | |
2865 | } | |
08d4a246 MH |
2866 | out: |
2867 | return ret; | |
a3d0c6aa NA |
2868 | } |
2869 | ||
1da177e4 LT |
2870 | #endif /* CONFIG_SYSCTL */ |
2871 | ||
e1759c21 | 2872 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 2873 | { |
a5516438 | 2874 | struct hstate *h = &default_hstate; |
457c1b27 NA |
2875 | if (!hugepages_supported()) |
2876 | return; | |
e1759c21 | 2877 | seq_printf(m, |
4f98a2fe RR |
2878 | "HugePages_Total: %5lu\n" |
2879 | "HugePages_Free: %5lu\n" | |
2880 | "HugePages_Rsvd: %5lu\n" | |
2881 | "HugePages_Surp: %5lu\n" | |
2882 | "Hugepagesize: %8lu kB\n", | |
a5516438 AK |
2883 | h->nr_huge_pages, |
2884 | h->free_huge_pages, | |
2885 | h->resv_huge_pages, | |
2886 | h->surplus_huge_pages, | |
2887 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
2888 | } |
2889 | ||
2890 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
2891 | { | |
a5516438 | 2892 | struct hstate *h = &default_hstate; |
457c1b27 NA |
2893 | if (!hugepages_supported()) |
2894 | return 0; | |
1da177e4 LT |
2895 | return sprintf(buf, |
2896 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
2897 | "Node %d HugePages_Free: %5u\n" |
2898 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
2899 | nid, h->nr_huge_pages_node[nid], |
2900 | nid, h->free_huge_pages_node[nid], | |
2901 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
2902 | } |
2903 | ||
949f7ec5 DR |
2904 | void hugetlb_show_meminfo(void) |
2905 | { | |
2906 | struct hstate *h; | |
2907 | int nid; | |
2908 | ||
457c1b27 NA |
2909 | if (!hugepages_supported()) |
2910 | return; | |
2911 | ||
949f7ec5 DR |
2912 | for_each_node_state(nid, N_MEMORY) |
2913 | for_each_hstate(h) | |
2914 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
2915 | nid, | |
2916 | h->nr_huge_pages_node[nid], | |
2917 | h->free_huge_pages_node[nid], | |
2918 | h->surplus_huge_pages_node[nid], | |
2919 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
2920 | } | |
2921 | ||
5d317b2b NH |
2922 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
2923 | { | |
2924 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
2925 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
2926 | } | |
2927 | ||
1da177e4 LT |
2928 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
2929 | unsigned long hugetlb_total_pages(void) | |
2930 | { | |
d0028588 WL |
2931 | struct hstate *h; |
2932 | unsigned long nr_total_pages = 0; | |
2933 | ||
2934 | for_each_hstate(h) | |
2935 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
2936 | return nr_total_pages; | |
1da177e4 | 2937 | } |
1da177e4 | 2938 | |
a5516438 | 2939 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
2940 | { |
2941 | int ret = -ENOMEM; | |
2942 | ||
2943 | spin_lock(&hugetlb_lock); | |
2944 | /* | |
2945 | * When cpuset is configured, it breaks the strict hugetlb page | |
2946 | * reservation as the accounting is done on a global variable. Such | |
2947 | * reservation is completely rubbish in the presence of cpuset because | |
2948 | * the reservation is not checked against page availability for the | |
2949 | * current cpuset. Application can still potentially OOM'ed by kernel | |
2950 | * with lack of free htlb page in cpuset that the task is in. | |
2951 | * Attempt to enforce strict accounting with cpuset is almost | |
2952 | * impossible (or too ugly) because cpuset is too fluid that | |
2953 | * task or memory node can be dynamically moved between cpusets. | |
2954 | * | |
2955 | * The change of semantics for shared hugetlb mapping with cpuset is | |
2956 | * undesirable. However, in order to preserve some of the semantics, | |
2957 | * we fall back to check against current free page availability as | |
2958 | * a best attempt and hopefully to minimize the impact of changing | |
2959 | * semantics that cpuset has. | |
2960 | */ | |
2961 | if (delta > 0) { | |
a5516438 | 2962 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
2963 | goto out; |
2964 | ||
a5516438 AK |
2965 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
2966 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
2967 | goto out; |
2968 | } | |
2969 | } | |
2970 | ||
2971 | ret = 0; | |
2972 | if (delta < 0) | |
a5516438 | 2973 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
2974 | |
2975 | out: | |
2976 | spin_unlock(&hugetlb_lock); | |
2977 | return ret; | |
2978 | } | |
2979 | ||
84afd99b AW |
2980 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
2981 | { | |
f522c3ac | 2982 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
2983 | |
2984 | /* | |
2985 | * This new VMA should share its siblings reservation map if present. | |
2986 | * The VMA will only ever have a valid reservation map pointer where | |
2987 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 2988 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
2989 | * after this open call completes. It is therefore safe to take a |
2990 | * new reference here without additional locking. | |
2991 | */ | |
4e35f483 | 2992 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 2993 | kref_get(&resv->refs); |
84afd99b AW |
2994 | } |
2995 | ||
a1e78772 MG |
2996 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
2997 | { | |
a5516438 | 2998 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 2999 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 3000 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 3001 | unsigned long reserve, start, end; |
1c5ecae3 | 3002 | long gbl_reserve; |
84afd99b | 3003 | |
4e35f483 JK |
3004 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3005 | return; | |
84afd99b | 3006 | |
4e35f483 JK |
3007 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
3008 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 3009 | |
4e35f483 | 3010 | reserve = (end - start) - region_count(resv, start, end); |
84afd99b | 3011 | |
4e35f483 JK |
3012 | kref_put(&resv->refs, resv_map_release); |
3013 | ||
3014 | if (reserve) { | |
1c5ecae3 MK |
3015 | /* |
3016 | * Decrement reserve counts. The global reserve count may be | |
3017 | * adjusted if the subpool has a minimum size. | |
3018 | */ | |
3019 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
3020 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 3021 | } |
a1e78772 MG |
3022 | } |
3023 | ||
1da177e4 LT |
3024 | /* |
3025 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
3026 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
3027 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
3028 | * this far. | |
3029 | */ | |
d0217ac0 | 3030 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
3031 | { |
3032 | BUG(); | |
d0217ac0 | 3033 | return 0; |
1da177e4 LT |
3034 | } |
3035 | ||
f0f37e2f | 3036 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 3037 | .fault = hugetlb_vm_op_fault, |
84afd99b | 3038 | .open = hugetlb_vm_op_open, |
a1e78772 | 3039 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
3040 | }; |
3041 | ||
1e8f889b DG |
3042 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
3043 | int writable) | |
63551ae0 DG |
3044 | { |
3045 | pte_t entry; | |
3046 | ||
1e8f889b | 3047 | if (writable) { |
106c992a GS |
3048 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
3049 | vma->vm_page_prot))); | |
63551ae0 | 3050 | } else { |
106c992a GS |
3051 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
3052 | vma->vm_page_prot)); | |
63551ae0 DG |
3053 | } |
3054 | entry = pte_mkyoung(entry); | |
3055 | entry = pte_mkhuge(entry); | |
d9ed9faa | 3056 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
3057 | |
3058 | return entry; | |
3059 | } | |
3060 | ||
1e8f889b DG |
3061 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
3062 | unsigned long address, pte_t *ptep) | |
3063 | { | |
3064 | pte_t entry; | |
3065 | ||
106c992a | 3066 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 3067 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 3068 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
3069 | } |
3070 | ||
4a705fef NH |
3071 | static int is_hugetlb_entry_migration(pte_t pte) |
3072 | { | |
3073 | swp_entry_t swp; | |
3074 | ||
3075 | if (huge_pte_none(pte) || pte_present(pte)) | |
3076 | return 0; | |
3077 | swp = pte_to_swp_entry(pte); | |
3078 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
3079 | return 1; | |
3080 | else | |
3081 | return 0; | |
3082 | } | |
3083 | ||
3084 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
3085 | { | |
3086 | swp_entry_t swp; | |
3087 | ||
3088 | if (huge_pte_none(pte) || pte_present(pte)) | |
3089 | return 0; | |
3090 | swp = pte_to_swp_entry(pte); | |
3091 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
3092 | return 1; | |
3093 | else | |
3094 | return 0; | |
3095 | } | |
1e8f889b | 3096 | |
63551ae0 DG |
3097 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
3098 | struct vm_area_struct *vma) | |
3099 | { | |
3100 | pte_t *src_pte, *dst_pte, entry; | |
3101 | struct page *ptepage; | |
1c59827d | 3102 | unsigned long addr; |
1e8f889b | 3103 | int cow; |
a5516438 AK |
3104 | struct hstate *h = hstate_vma(vma); |
3105 | unsigned long sz = huge_page_size(h); | |
e8569dd2 AS |
3106 | unsigned long mmun_start; /* For mmu_notifiers */ |
3107 | unsigned long mmun_end; /* For mmu_notifiers */ | |
3108 | int ret = 0; | |
1e8f889b DG |
3109 | |
3110 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 3111 | |
e8569dd2 AS |
3112 | mmun_start = vma->vm_start; |
3113 | mmun_end = vma->vm_end; | |
3114 | if (cow) | |
3115 | mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); | |
3116 | ||
a5516438 | 3117 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 3118 | spinlock_t *src_ptl, *dst_ptl; |
c74df32c HD |
3119 | src_pte = huge_pte_offset(src, addr); |
3120 | if (!src_pte) | |
3121 | continue; | |
a5516438 | 3122 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
3123 | if (!dst_pte) { |
3124 | ret = -ENOMEM; | |
3125 | break; | |
3126 | } | |
c5c99429 LW |
3127 | |
3128 | /* If the pagetables are shared don't copy or take references */ | |
3129 | if (dst_pte == src_pte) | |
3130 | continue; | |
3131 | ||
cb900f41 KS |
3132 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
3133 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3134 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef NH |
3135 | entry = huge_ptep_get(src_pte); |
3136 | if (huge_pte_none(entry)) { /* skip none entry */ | |
3137 | ; | |
3138 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
3139 | is_hugetlb_entry_hwpoisoned(entry))) { | |
3140 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
3141 | ||
3142 | if (is_write_migration_entry(swp_entry) && cow) { | |
3143 | /* | |
3144 | * COW mappings require pages in both | |
3145 | * parent and child to be set to read. | |
3146 | */ | |
3147 | make_migration_entry_read(&swp_entry); | |
3148 | entry = swp_entry_to_pte(swp_entry); | |
3149 | set_huge_pte_at(src, addr, src_pte, entry); | |
3150 | } | |
3151 | set_huge_pte_at(dst, addr, dst_pte, entry); | |
3152 | } else { | |
34ee645e | 3153 | if (cow) { |
7f2e9525 | 3154 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e JR |
3155 | mmu_notifier_invalidate_range(src, mmun_start, |
3156 | mmun_end); | |
3157 | } | |
0253d634 | 3158 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
3159 | ptepage = pte_page(entry); |
3160 | get_page(ptepage); | |
53f9263b | 3161 | page_dup_rmap(ptepage, true); |
1c59827d | 3162 | set_huge_pte_at(dst, addr, dst_pte, entry); |
5d317b2b | 3163 | hugetlb_count_add(pages_per_huge_page(h), dst); |
1c59827d | 3164 | } |
cb900f41 KS |
3165 | spin_unlock(src_ptl); |
3166 | spin_unlock(dst_ptl); | |
63551ae0 | 3167 | } |
63551ae0 | 3168 | |
e8569dd2 AS |
3169 | if (cow) |
3170 | mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); | |
3171 | ||
3172 | return ret; | |
63551ae0 DG |
3173 | } |
3174 | ||
24669e58 AK |
3175 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
3176 | unsigned long start, unsigned long end, | |
3177 | struct page *ref_page) | |
63551ae0 | 3178 | { |
24669e58 | 3179 | int force_flush = 0; |
63551ae0 DG |
3180 | struct mm_struct *mm = vma->vm_mm; |
3181 | unsigned long address; | |
c7546f8f | 3182 | pte_t *ptep; |
63551ae0 | 3183 | pte_t pte; |
cb900f41 | 3184 | spinlock_t *ptl; |
63551ae0 | 3185 | struct page *page; |
a5516438 AK |
3186 | struct hstate *h = hstate_vma(vma); |
3187 | unsigned long sz = huge_page_size(h); | |
2ec74c3e SG |
3188 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
3189 | const unsigned long mmun_end = end; /* For mmu_notifiers */ | |
a5516438 | 3190 | |
63551ae0 | 3191 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
3192 | BUG_ON(start & ~huge_page_mask(h)); |
3193 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 3194 | |
24669e58 | 3195 | tlb_start_vma(tlb, vma); |
2ec74c3e | 3196 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
569f48b8 | 3197 | address = start; |
24669e58 | 3198 | again: |
569f48b8 | 3199 | for (; address < end; address += sz) { |
c7546f8f | 3200 | ptep = huge_pte_offset(mm, address); |
4c887265 | 3201 | if (!ptep) |
c7546f8f DG |
3202 | continue; |
3203 | ||
cb900f41 | 3204 | ptl = huge_pte_lock(h, mm, ptep); |
39dde65c | 3205 | if (huge_pmd_unshare(mm, &address, ptep)) |
cb900f41 | 3206 | goto unlock; |
39dde65c | 3207 | |
6629326b HD |
3208 | pte = huge_ptep_get(ptep); |
3209 | if (huge_pte_none(pte)) | |
cb900f41 | 3210 | goto unlock; |
6629326b HD |
3211 | |
3212 | /* | |
9fbc1f63 NH |
3213 | * Migrating hugepage or HWPoisoned hugepage is already |
3214 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 3215 | */ |
9fbc1f63 | 3216 | if (unlikely(!pte_present(pte))) { |
106c992a | 3217 | huge_pte_clear(mm, address, ptep); |
cb900f41 | 3218 | goto unlock; |
8c4894c6 | 3219 | } |
6629326b HD |
3220 | |
3221 | page = pte_page(pte); | |
04f2cbe3 MG |
3222 | /* |
3223 | * If a reference page is supplied, it is because a specific | |
3224 | * page is being unmapped, not a range. Ensure the page we | |
3225 | * are about to unmap is the actual page of interest. | |
3226 | */ | |
3227 | if (ref_page) { | |
04f2cbe3 | 3228 | if (page != ref_page) |
cb900f41 | 3229 | goto unlock; |
04f2cbe3 MG |
3230 | |
3231 | /* | |
3232 | * Mark the VMA as having unmapped its page so that | |
3233 | * future faults in this VMA will fail rather than | |
3234 | * looking like data was lost | |
3235 | */ | |
3236 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
3237 | } | |
3238 | ||
c7546f8f | 3239 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
24669e58 | 3240 | tlb_remove_tlb_entry(tlb, ptep, address); |
106c992a | 3241 | if (huge_pte_dirty(pte)) |
6649a386 | 3242 | set_page_dirty(page); |
9e81130b | 3243 | |
5d317b2b | 3244 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
d281ee61 | 3245 | page_remove_rmap(page, true); |
24669e58 | 3246 | force_flush = !__tlb_remove_page(tlb, page); |
cb900f41 | 3247 | if (force_flush) { |
569f48b8 | 3248 | address += sz; |
cb900f41 | 3249 | spin_unlock(ptl); |
24669e58 | 3250 | break; |
cb900f41 | 3251 | } |
9e81130b | 3252 | /* Bail out after unmapping reference page if supplied */ |
cb900f41 KS |
3253 | if (ref_page) { |
3254 | spin_unlock(ptl); | |
9e81130b | 3255 | break; |
cb900f41 KS |
3256 | } |
3257 | unlock: | |
3258 | spin_unlock(ptl); | |
63551ae0 | 3259 | } |
24669e58 AK |
3260 | /* |
3261 | * mmu_gather ran out of room to batch pages, we break out of | |
3262 | * the PTE lock to avoid doing the potential expensive TLB invalidate | |
3263 | * and page-free while holding it. | |
3264 | */ | |
3265 | if (force_flush) { | |
3266 | force_flush = 0; | |
3267 | tlb_flush_mmu(tlb); | |
3268 | if (address < end && !ref_page) | |
3269 | goto again; | |
fe1668ae | 3270 | } |
2ec74c3e | 3271 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
24669e58 | 3272 | tlb_end_vma(tlb, vma); |
1da177e4 | 3273 | } |
63551ae0 | 3274 | |
d833352a MG |
3275 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
3276 | struct vm_area_struct *vma, unsigned long start, | |
3277 | unsigned long end, struct page *ref_page) | |
3278 | { | |
3279 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
3280 | ||
3281 | /* | |
3282 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
3283 | * test will fail on a vma being torn down, and not grab a page table | |
3284 | * on its way out. We're lucky that the flag has such an appropriate | |
3285 | * name, and can in fact be safely cleared here. We could clear it | |
3286 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 3287 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 3288 | * because in the context this is called, the VMA is about to be |
c8c06efa | 3289 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
3290 | */ |
3291 | vma->vm_flags &= ~VM_MAYSHARE; | |
3292 | } | |
3293 | ||
502717f4 | 3294 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 3295 | unsigned long end, struct page *ref_page) |
502717f4 | 3296 | { |
24669e58 AK |
3297 | struct mm_struct *mm; |
3298 | struct mmu_gather tlb; | |
3299 | ||
3300 | mm = vma->vm_mm; | |
3301 | ||
2b047252 | 3302 | tlb_gather_mmu(&tlb, mm, start, end); |
24669e58 AK |
3303 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
3304 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 CK |
3305 | } |
3306 | ||
04f2cbe3 MG |
3307 | /* |
3308 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
3309 | * mappping it owns the reserve page for. The intention is to unmap the page | |
3310 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
3311 | * same region. | |
3312 | */ | |
2f4612af DB |
3313 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
3314 | struct page *page, unsigned long address) | |
04f2cbe3 | 3315 | { |
7526674d | 3316 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
3317 | struct vm_area_struct *iter_vma; |
3318 | struct address_space *mapping; | |
04f2cbe3 MG |
3319 | pgoff_t pgoff; |
3320 | ||
3321 | /* | |
3322 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
3323 | * from page cache lookup which is in HPAGE_SIZE units. | |
3324 | */ | |
7526674d | 3325 | address = address & huge_page_mask(h); |
36e4f20a MH |
3326 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
3327 | vma->vm_pgoff; | |
496ad9aa | 3328 | mapping = file_inode(vma->vm_file)->i_mapping; |
04f2cbe3 | 3329 | |
4eb2b1dc MG |
3330 | /* |
3331 | * Take the mapping lock for the duration of the table walk. As | |
3332 | * this mapping should be shared between all the VMAs, | |
3333 | * __unmap_hugepage_range() is called as the lock is already held | |
3334 | */ | |
83cde9e8 | 3335 | i_mmap_lock_write(mapping); |
6b2dbba8 | 3336 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
3337 | /* Do not unmap the current VMA */ |
3338 | if (iter_vma == vma) | |
3339 | continue; | |
3340 | ||
2f84a899 MG |
3341 | /* |
3342 | * Shared VMAs have their own reserves and do not affect | |
3343 | * MAP_PRIVATE accounting but it is possible that a shared | |
3344 | * VMA is using the same page so check and skip such VMAs. | |
3345 | */ | |
3346 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
3347 | continue; | |
3348 | ||
04f2cbe3 MG |
3349 | /* |
3350 | * Unmap the page from other VMAs without their own reserves. | |
3351 | * They get marked to be SIGKILLed if they fault in these | |
3352 | * areas. This is because a future no-page fault on this VMA | |
3353 | * could insert a zeroed page instead of the data existing | |
3354 | * from the time of fork. This would look like data corruption | |
3355 | */ | |
3356 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
3357 | unmap_hugepage_range(iter_vma, address, |
3358 | address + huge_page_size(h), page); | |
04f2cbe3 | 3359 | } |
83cde9e8 | 3360 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
3361 | } |
3362 | ||
0fe6e20b NH |
3363 | /* |
3364 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
3365 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
3366 | * cannot race with other handlers or page migration. | |
3367 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 3368 | */ |
1e8f889b | 3369 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
04f2cbe3 | 3370 | unsigned long address, pte_t *ptep, pte_t pte, |
cb900f41 | 3371 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 3372 | { |
a5516438 | 3373 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 3374 | struct page *old_page, *new_page; |
ad4404a2 | 3375 | int ret = 0, outside_reserve = 0; |
2ec74c3e SG |
3376 | unsigned long mmun_start; /* For mmu_notifiers */ |
3377 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1e8f889b DG |
3378 | |
3379 | old_page = pte_page(pte); | |
3380 | ||
04f2cbe3 | 3381 | retry_avoidcopy: |
1e8f889b DG |
3382 | /* If no-one else is actually using this page, avoid the copy |
3383 | * and just make the page writable */ | |
37a2140d JK |
3384 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
3385 | page_move_anon_rmap(old_page, vma, address); | |
1e8f889b | 3386 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 3387 | return 0; |
1e8f889b DG |
3388 | } |
3389 | ||
04f2cbe3 MG |
3390 | /* |
3391 | * If the process that created a MAP_PRIVATE mapping is about to | |
3392 | * perform a COW due to a shared page count, attempt to satisfy | |
3393 | * the allocation without using the existing reserves. The pagecache | |
3394 | * page is used to determine if the reserve at this address was | |
3395 | * consumed or not. If reserves were used, a partial faulted mapping | |
3396 | * at the time of fork() could consume its reserves on COW instead | |
3397 | * of the full address range. | |
3398 | */ | |
5944d011 | 3399 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
3400 | old_page != pagecache_page) |
3401 | outside_reserve = 1; | |
3402 | ||
09cbfeaf | 3403 | get_page(old_page); |
b76c8cfb | 3404 | |
ad4404a2 DB |
3405 | /* |
3406 | * Drop page table lock as buddy allocator may be called. It will | |
3407 | * be acquired again before returning to the caller, as expected. | |
3408 | */ | |
cb900f41 | 3409 | spin_unlock(ptl); |
04f2cbe3 | 3410 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 3411 | |
2fc39cec | 3412 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
3413 | /* |
3414 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
3415 | * it is due to references held by a child and an insufficient | |
3416 | * huge page pool. To guarantee the original mappers | |
3417 | * reliability, unmap the page from child processes. The child | |
3418 | * may get SIGKILLed if it later faults. | |
3419 | */ | |
3420 | if (outside_reserve) { | |
09cbfeaf | 3421 | put_page(old_page); |
04f2cbe3 | 3422 | BUG_ON(huge_pte_none(pte)); |
2f4612af DB |
3423 | unmap_ref_private(mm, vma, old_page, address); |
3424 | BUG_ON(huge_pte_none(pte)); | |
3425 | spin_lock(ptl); | |
3426 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); | |
3427 | if (likely(ptep && | |
3428 | pte_same(huge_ptep_get(ptep), pte))) | |
3429 | goto retry_avoidcopy; | |
3430 | /* | |
3431 | * race occurs while re-acquiring page table | |
3432 | * lock, and our job is done. | |
3433 | */ | |
3434 | return 0; | |
04f2cbe3 MG |
3435 | } |
3436 | ||
ad4404a2 DB |
3437 | ret = (PTR_ERR(new_page) == -ENOMEM) ? |
3438 | VM_FAULT_OOM : VM_FAULT_SIGBUS; | |
3439 | goto out_release_old; | |
1e8f889b DG |
3440 | } |
3441 | ||
0fe6e20b NH |
3442 | /* |
3443 | * When the original hugepage is shared one, it does not have | |
3444 | * anon_vma prepared. | |
3445 | */ | |
44e2aa93 | 3446 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
3447 | ret = VM_FAULT_OOM; |
3448 | goto out_release_all; | |
44e2aa93 | 3449 | } |
0fe6e20b | 3450 | |
47ad8475 AA |
3451 | copy_user_huge_page(new_page, old_page, address, vma, |
3452 | pages_per_huge_page(h)); | |
0ed361de | 3453 | __SetPageUptodate(new_page); |
bcc54222 | 3454 | set_page_huge_active(new_page); |
1e8f889b | 3455 | |
2ec74c3e SG |
3456 | mmun_start = address & huge_page_mask(h); |
3457 | mmun_end = mmun_start + huge_page_size(h); | |
3458 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
ad4404a2 | 3459 | |
b76c8cfb | 3460 | /* |
cb900f41 | 3461 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
3462 | * before the page tables are altered |
3463 | */ | |
cb900f41 | 3464 | spin_lock(ptl); |
a5516438 | 3465 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
a9af0c5d | 3466 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
3467 | ClearPagePrivate(new_page); |
3468 | ||
1e8f889b | 3469 | /* Break COW */ |
8fe627ec | 3470 | huge_ptep_clear_flush(vma, address, ptep); |
34ee645e | 3471 | mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); |
1e8f889b DG |
3472 | set_huge_pte_at(mm, address, ptep, |
3473 | make_huge_pte(vma, new_page, 1)); | |
d281ee61 | 3474 | page_remove_rmap(old_page, true); |
cd67f0d2 | 3475 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
3476 | /* Make the old page be freed below */ |
3477 | new_page = old_page; | |
3478 | } | |
cb900f41 | 3479 | spin_unlock(ptl); |
2ec74c3e | 3480 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
ad4404a2 | 3481 | out_release_all: |
09cbfeaf | 3482 | put_page(new_page); |
ad4404a2 | 3483 | out_release_old: |
09cbfeaf | 3484 | put_page(old_page); |
8312034f | 3485 | |
ad4404a2 DB |
3486 | spin_lock(ptl); /* Caller expects lock to be held */ |
3487 | return ret; | |
1e8f889b DG |
3488 | } |
3489 | ||
04f2cbe3 | 3490 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
3491 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
3492 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
3493 | { |
3494 | struct address_space *mapping; | |
e7c4b0bf | 3495 | pgoff_t idx; |
04f2cbe3 MG |
3496 | |
3497 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 3498 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
3499 | |
3500 | return find_lock_page(mapping, idx); | |
3501 | } | |
3502 | ||
3ae77f43 HD |
3503 | /* |
3504 | * Return whether there is a pagecache page to back given address within VMA. | |
3505 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
3506 | */ | |
3507 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
3508 | struct vm_area_struct *vma, unsigned long address) |
3509 | { | |
3510 | struct address_space *mapping; | |
3511 | pgoff_t idx; | |
3512 | struct page *page; | |
3513 | ||
3514 | mapping = vma->vm_file->f_mapping; | |
3515 | idx = vma_hugecache_offset(h, vma, address); | |
3516 | ||
3517 | page = find_get_page(mapping, idx); | |
3518 | if (page) | |
3519 | put_page(page); | |
3520 | return page != NULL; | |
3521 | } | |
3522 | ||
ab76ad54 MK |
3523 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
3524 | pgoff_t idx) | |
3525 | { | |
3526 | struct inode *inode = mapping->host; | |
3527 | struct hstate *h = hstate_inode(inode); | |
3528 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
3529 | ||
3530 | if (err) | |
3531 | return err; | |
3532 | ClearPagePrivate(page); | |
3533 | ||
3534 | spin_lock(&inode->i_lock); | |
3535 | inode->i_blocks += blocks_per_huge_page(h); | |
3536 | spin_unlock(&inode->i_lock); | |
3537 | return 0; | |
3538 | } | |
3539 | ||
a1ed3dda | 3540 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
8382d914 DB |
3541 | struct address_space *mapping, pgoff_t idx, |
3542 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 3543 | { |
a5516438 | 3544 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 3545 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 3546 | int anon_rmap = 0; |
4c887265 | 3547 | unsigned long size; |
4c887265 | 3548 | struct page *page; |
1e8f889b | 3549 | pte_t new_pte; |
cb900f41 | 3550 | spinlock_t *ptl; |
4c887265 | 3551 | |
04f2cbe3 MG |
3552 | /* |
3553 | * Currently, we are forced to kill the process in the event the | |
3554 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 3555 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
3556 | */ |
3557 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 3558 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 3559 | current->pid); |
04f2cbe3 MG |
3560 | return ret; |
3561 | } | |
3562 | ||
4c887265 AL |
3563 | /* |
3564 | * Use page lock to guard against racing truncation | |
3565 | * before we get page_table_lock. | |
3566 | */ | |
6bda666a CL |
3567 | retry: |
3568 | page = find_lock_page(mapping, idx); | |
3569 | if (!page) { | |
a5516438 | 3570 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
3571 | if (idx >= size) |
3572 | goto out; | |
04f2cbe3 | 3573 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 3574 | if (IS_ERR(page)) { |
76dcee75 AK |
3575 | ret = PTR_ERR(page); |
3576 | if (ret == -ENOMEM) | |
3577 | ret = VM_FAULT_OOM; | |
3578 | else | |
3579 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
3580 | goto out; |
3581 | } | |
47ad8475 | 3582 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 3583 | __SetPageUptodate(page); |
bcc54222 | 3584 | set_page_huge_active(page); |
ac9b9c66 | 3585 | |
f83a275d | 3586 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 3587 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
3588 | if (err) { |
3589 | put_page(page); | |
6bda666a CL |
3590 | if (err == -EEXIST) |
3591 | goto retry; | |
3592 | goto out; | |
3593 | } | |
23be7468 | 3594 | } else { |
6bda666a | 3595 | lock_page(page); |
0fe6e20b NH |
3596 | if (unlikely(anon_vma_prepare(vma))) { |
3597 | ret = VM_FAULT_OOM; | |
3598 | goto backout_unlocked; | |
3599 | } | |
409eb8c2 | 3600 | anon_rmap = 1; |
23be7468 | 3601 | } |
0fe6e20b | 3602 | } else { |
998b4382 NH |
3603 | /* |
3604 | * If memory error occurs between mmap() and fault, some process | |
3605 | * don't have hwpoisoned swap entry for errored virtual address. | |
3606 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
3607 | */ | |
3608 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 3609 | ret = VM_FAULT_HWPOISON | |
972dc4de | 3610 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
3611 | goto backout_unlocked; |
3612 | } | |
6bda666a | 3613 | } |
1e8f889b | 3614 | |
57303d80 AW |
3615 | /* |
3616 | * If we are going to COW a private mapping later, we examine the | |
3617 | * pending reservations for this page now. This will ensure that | |
3618 | * any allocations necessary to record that reservation occur outside | |
3619 | * the spinlock. | |
3620 | */ | |
5e911373 | 3621 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
2b26736c AW |
3622 | if (vma_needs_reservation(h, vma, address) < 0) { |
3623 | ret = VM_FAULT_OOM; | |
3624 | goto backout_unlocked; | |
3625 | } | |
5e911373 | 3626 | /* Just decrements count, does not deallocate */ |
feba16e2 | 3627 | vma_end_reservation(h, vma, address); |
5e911373 | 3628 | } |
57303d80 | 3629 | |
cb900f41 KS |
3630 | ptl = huge_pte_lockptr(h, mm, ptep); |
3631 | spin_lock(ptl); | |
a5516438 | 3632 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
3633 | if (idx >= size) |
3634 | goto backout; | |
3635 | ||
83c54070 | 3636 | ret = 0; |
7f2e9525 | 3637 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
3638 | goto backout; |
3639 | ||
07443a85 JK |
3640 | if (anon_rmap) { |
3641 | ClearPagePrivate(page); | |
409eb8c2 | 3642 | hugepage_add_new_anon_rmap(page, vma, address); |
ac714904 | 3643 | } else |
53f9263b | 3644 | page_dup_rmap(page, true); |
1e8f889b DG |
3645 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
3646 | && (vma->vm_flags & VM_SHARED))); | |
3647 | set_huge_pte_at(mm, address, ptep, new_pte); | |
3648 | ||
5d317b2b | 3649 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 3650 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 3651 | /* Optimization, do the COW without a second fault */ |
cb900f41 | 3652 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); |
1e8f889b DG |
3653 | } |
3654 | ||
cb900f41 | 3655 | spin_unlock(ptl); |
4c887265 AL |
3656 | unlock_page(page); |
3657 | out: | |
ac9b9c66 | 3658 | return ret; |
4c887265 AL |
3659 | |
3660 | backout: | |
cb900f41 | 3661 | spin_unlock(ptl); |
2b26736c | 3662 | backout_unlocked: |
4c887265 AL |
3663 | unlock_page(page); |
3664 | put_page(page); | |
3665 | goto out; | |
ac9b9c66 HD |
3666 | } |
3667 | ||
8382d914 | 3668 | #ifdef CONFIG_SMP |
c672c7f2 | 3669 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
8382d914 DB |
3670 | struct vm_area_struct *vma, |
3671 | struct address_space *mapping, | |
3672 | pgoff_t idx, unsigned long address) | |
3673 | { | |
3674 | unsigned long key[2]; | |
3675 | u32 hash; | |
3676 | ||
3677 | if (vma->vm_flags & VM_SHARED) { | |
3678 | key[0] = (unsigned long) mapping; | |
3679 | key[1] = idx; | |
3680 | } else { | |
3681 | key[0] = (unsigned long) mm; | |
3682 | key[1] = address >> huge_page_shift(h); | |
3683 | } | |
3684 | ||
3685 | hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); | |
3686 | ||
3687 | return hash & (num_fault_mutexes - 1); | |
3688 | } | |
3689 | #else | |
3690 | /* | |
3691 | * For uniprocesor systems we always use a single mutex, so just | |
3692 | * return 0 and avoid the hashing overhead. | |
3693 | */ | |
c672c7f2 | 3694 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
8382d914 DB |
3695 | struct vm_area_struct *vma, |
3696 | struct address_space *mapping, | |
3697 | pgoff_t idx, unsigned long address) | |
3698 | { | |
3699 | return 0; | |
3700 | } | |
3701 | #endif | |
3702 | ||
86e5216f | 3703 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 3704 | unsigned long address, unsigned int flags) |
86e5216f | 3705 | { |
8382d914 | 3706 | pte_t *ptep, entry; |
cb900f41 | 3707 | spinlock_t *ptl; |
1e8f889b | 3708 | int ret; |
8382d914 DB |
3709 | u32 hash; |
3710 | pgoff_t idx; | |
0fe6e20b | 3711 | struct page *page = NULL; |
57303d80 | 3712 | struct page *pagecache_page = NULL; |
a5516438 | 3713 | struct hstate *h = hstate_vma(vma); |
8382d914 | 3714 | struct address_space *mapping; |
0f792cf9 | 3715 | int need_wait_lock = 0; |
86e5216f | 3716 | |
1e16a539 KH |
3717 | address &= huge_page_mask(h); |
3718 | ||
fd6a03ed NH |
3719 | ptep = huge_pte_offset(mm, address); |
3720 | if (ptep) { | |
3721 | entry = huge_ptep_get(ptep); | |
290408d4 | 3722 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 3723 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
3724 | return 0; |
3725 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 3726 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 3727 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
0d777df5 NH |
3728 | } else { |
3729 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); | |
3730 | if (!ptep) | |
3731 | return VM_FAULT_OOM; | |
fd6a03ed NH |
3732 | } |
3733 | ||
8382d914 DB |
3734 | mapping = vma->vm_file->f_mapping; |
3735 | idx = vma_hugecache_offset(h, vma, address); | |
3736 | ||
3935baa9 DG |
3737 | /* |
3738 | * Serialize hugepage allocation and instantiation, so that we don't | |
3739 | * get spurious allocation failures if two CPUs race to instantiate | |
3740 | * the same page in the page cache. | |
3741 | */ | |
c672c7f2 MK |
3742 | hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address); |
3743 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
8382d914 | 3744 | |
7f2e9525 GS |
3745 | entry = huge_ptep_get(ptep); |
3746 | if (huge_pte_none(entry)) { | |
8382d914 | 3747 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 3748 | goto out_mutex; |
3935baa9 | 3749 | } |
86e5216f | 3750 | |
83c54070 | 3751 | ret = 0; |
1e8f889b | 3752 | |
0f792cf9 NH |
3753 | /* |
3754 | * entry could be a migration/hwpoison entry at this point, so this | |
3755 | * check prevents the kernel from going below assuming that we have | |
3756 | * a active hugepage in pagecache. This goto expects the 2nd page fault, | |
3757 | * and is_hugetlb_entry_(migration|hwpoisoned) check will properly | |
3758 | * handle it. | |
3759 | */ | |
3760 | if (!pte_present(entry)) | |
3761 | goto out_mutex; | |
3762 | ||
57303d80 AW |
3763 | /* |
3764 | * If we are going to COW the mapping later, we examine the pending | |
3765 | * reservations for this page now. This will ensure that any | |
3766 | * allocations necessary to record that reservation occur outside the | |
3767 | * spinlock. For private mappings, we also lookup the pagecache | |
3768 | * page now as it is used to determine if a reservation has been | |
3769 | * consumed. | |
3770 | */ | |
106c992a | 3771 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
2b26736c AW |
3772 | if (vma_needs_reservation(h, vma, address) < 0) { |
3773 | ret = VM_FAULT_OOM; | |
b4d1d99f | 3774 | goto out_mutex; |
2b26736c | 3775 | } |
5e911373 | 3776 | /* Just decrements count, does not deallocate */ |
feba16e2 | 3777 | vma_end_reservation(h, vma, address); |
57303d80 | 3778 | |
f83a275d | 3779 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
3780 | pagecache_page = hugetlbfs_pagecache_page(h, |
3781 | vma, address); | |
3782 | } | |
3783 | ||
0f792cf9 NH |
3784 | ptl = huge_pte_lock(h, mm, ptep); |
3785 | ||
3786 | /* Check for a racing update before calling hugetlb_cow */ | |
3787 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
3788 | goto out_ptl; | |
3789 | ||
56c9cfb1 NH |
3790 | /* |
3791 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
3792 | * pagecache_page, so here we need take the former one | |
3793 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
3794 | */ |
3795 | page = pte_page(entry); | |
3796 | if (page != pagecache_page) | |
0f792cf9 NH |
3797 | if (!trylock_page(page)) { |
3798 | need_wait_lock = 1; | |
3799 | goto out_ptl; | |
3800 | } | |
b4d1d99f | 3801 | |
0f792cf9 | 3802 | get_page(page); |
b4d1d99f | 3803 | |
788c7df4 | 3804 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 3805 | if (!huge_pte_write(entry)) { |
57303d80 | 3806 | ret = hugetlb_cow(mm, vma, address, ptep, entry, |
cb900f41 | 3807 | pagecache_page, ptl); |
0f792cf9 | 3808 | goto out_put_page; |
b4d1d99f | 3809 | } |
106c992a | 3810 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
3811 | } |
3812 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
3813 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
3814 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 3815 | update_mmu_cache(vma, address, ptep); |
0f792cf9 NH |
3816 | out_put_page: |
3817 | if (page != pagecache_page) | |
3818 | unlock_page(page); | |
3819 | put_page(page); | |
cb900f41 KS |
3820 | out_ptl: |
3821 | spin_unlock(ptl); | |
57303d80 AW |
3822 | |
3823 | if (pagecache_page) { | |
3824 | unlock_page(pagecache_page); | |
3825 | put_page(pagecache_page); | |
3826 | } | |
b4d1d99f | 3827 | out_mutex: |
c672c7f2 | 3828 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
0f792cf9 NH |
3829 | /* |
3830 | * Generally it's safe to hold refcount during waiting page lock. But | |
3831 | * here we just wait to defer the next page fault to avoid busy loop and | |
3832 | * the page is not used after unlocked before returning from the current | |
3833 | * page fault. So we are safe from accessing freed page, even if we wait | |
3834 | * here without taking refcount. | |
3835 | */ | |
3836 | if (need_wait_lock) | |
3837 | wait_on_page_locked(page); | |
1e8f889b | 3838 | return ret; |
86e5216f AL |
3839 | } |
3840 | ||
28a35716 ML |
3841 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
3842 | struct page **pages, struct vm_area_struct **vmas, | |
3843 | unsigned long *position, unsigned long *nr_pages, | |
3844 | long i, unsigned int flags) | |
63551ae0 | 3845 | { |
d5d4b0aa CK |
3846 | unsigned long pfn_offset; |
3847 | unsigned long vaddr = *position; | |
28a35716 | 3848 | unsigned long remainder = *nr_pages; |
a5516438 | 3849 | struct hstate *h = hstate_vma(vma); |
63551ae0 | 3850 | |
63551ae0 | 3851 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 3852 | pte_t *pte; |
cb900f41 | 3853 | spinlock_t *ptl = NULL; |
2a15efc9 | 3854 | int absent; |
4c887265 | 3855 | struct page *page; |
63551ae0 | 3856 | |
02057967 DR |
3857 | /* |
3858 | * If we have a pending SIGKILL, don't keep faulting pages and | |
3859 | * potentially allocating memory. | |
3860 | */ | |
3861 | if (unlikely(fatal_signal_pending(current))) { | |
3862 | remainder = 0; | |
3863 | break; | |
3864 | } | |
3865 | ||
4c887265 AL |
3866 | /* |
3867 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 3868 | * each hugepage. We have to make sure we get the |
4c887265 | 3869 | * first, for the page indexing below to work. |
cb900f41 KS |
3870 | * |
3871 | * Note that page table lock is not held when pte is null. | |
4c887265 | 3872 | */ |
a5516438 | 3873 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); |
cb900f41 KS |
3874 | if (pte) |
3875 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
3876 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
3877 | ||
3878 | /* | |
3879 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
3880 | * an error where there's an empty slot with no huge pagecache |
3881 | * to back it. This way, we avoid allocating a hugepage, and | |
3882 | * the sparse dumpfile avoids allocating disk blocks, but its | |
3883 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 3884 | */ |
3ae77f43 HD |
3885 | if (absent && (flags & FOLL_DUMP) && |
3886 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
3887 | if (pte) |
3888 | spin_unlock(ptl); | |
2a15efc9 HD |
3889 | remainder = 0; |
3890 | break; | |
3891 | } | |
63551ae0 | 3892 | |
9cc3a5bd NH |
3893 | /* |
3894 | * We need call hugetlb_fault for both hugepages under migration | |
3895 | * (in which case hugetlb_fault waits for the migration,) and | |
3896 | * hwpoisoned hugepages (in which case we need to prevent the | |
3897 | * caller from accessing to them.) In order to do this, we use | |
3898 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
3899 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
3900 | * both cases, and because we can't follow correct pages | |
3901 | * directly from any kind of swap entries. | |
3902 | */ | |
3903 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
3904 | ((flags & FOLL_WRITE) && |
3905 | !huge_pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 3906 | int ret; |
63551ae0 | 3907 | |
cb900f41 KS |
3908 | if (pte) |
3909 | spin_unlock(ptl); | |
2a15efc9 HD |
3910 | ret = hugetlb_fault(mm, vma, vaddr, |
3911 | (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); | |
a89182c7 | 3912 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 3913 | continue; |
63551ae0 | 3914 | |
4c887265 | 3915 | remainder = 0; |
4c887265 AL |
3916 | break; |
3917 | } | |
3918 | ||
a5516438 | 3919 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 3920 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 3921 | same_page: |
d6692183 | 3922 | if (pages) { |
2a15efc9 | 3923 | pages[i] = mem_map_offset(page, pfn_offset); |
ddc58f27 | 3924 | get_page(pages[i]); |
d6692183 | 3925 | } |
63551ae0 DG |
3926 | |
3927 | if (vmas) | |
3928 | vmas[i] = vma; | |
3929 | ||
3930 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 3931 | ++pfn_offset; |
63551ae0 DG |
3932 | --remainder; |
3933 | ++i; | |
d5d4b0aa | 3934 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 3935 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa CK |
3936 | /* |
3937 | * We use pfn_offset to avoid touching the pageframes | |
3938 | * of this compound page. | |
3939 | */ | |
3940 | goto same_page; | |
3941 | } | |
cb900f41 | 3942 | spin_unlock(ptl); |
63551ae0 | 3943 | } |
28a35716 | 3944 | *nr_pages = remainder; |
63551ae0 DG |
3945 | *position = vaddr; |
3946 | ||
2a15efc9 | 3947 | return i ? i : -EFAULT; |
63551ae0 | 3948 | } |
8f860591 | 3949 | |
7da4d641 | 3950 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
3951 | unsigned long address, unsigned long end, pgprot_t newprot) |
3952 | { | |
3953 | struct mm_struct *mm = vma->vm_mm; | |
3954 | unsigned long start = address; | |
3955 | pte_t *ptep; | |
3956 | pte_t pte; | |
a5516438 | 3957 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 3958 | unsigned long pages = 0; |
8f860591 ZY |
3959 | |
3960 | BUG_ON(address >= end); | |
3961 | flush_cache_range(vma, address, end); | |
3962 | ||
a5338093 | 3963 | mmu_notifier_invalidate_range_start(mm, start, end); |
83cde9e8 | 3964 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 3965 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 3966 | spinlock_t *ptl; |
8f860591 ZY |
3967 | ptep = huge_pte_offset(mm, address); |
3968 | if (!ptep) | |
3969 | continue; | |
cb900f41 | 3970 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
3971 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3972 | pages++; | |
cb900f41 | 3973 | spin_unlock(ptl); |
39dde65c | 3974 | continue; |
7da4d641 | 3975 | } |
a8bda28d NH |
3976 | pte = huge_ptep_get(ptep); |
3977 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
3978 | spin_unlock(ptl); | |
3979 | continue; | |
3980 | } | |
3981 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
3982 | swp_entry_t entry = pte_to_swp_entry(pte); | |
3983 | ||
3984 | if (is_write_migration_entry(entry)) { | |
3985 | pte_t newpte; | |
3986 | ||
3987 | make_migration_entry_read(&entry); | |
3988 | newpte = swp_entry_to_pte(entry); | |
3989 | set_huge_pte_at(mm, address, ptep, newpte); | |
3990 | pages++; | |
3991 | } | |
3992 | spin_unlock(ptl); | |
3993 | continue; | |
3994 | } | |
3995 | if (!huge_pte_none(pte)) { | |
8f860591 | 3996 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
106c992a | 3997 | pte = pte_mkhuge(huge_pte_modify(pte, newprot)); |
be7517d6 | 3998 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
8f860591 | 3999 | set_huge_pte_at(mm, address, ptep, pte); |
7da4d641 | 4000 | pages++; |
8f860591 | 4001 | } |
cb900f41 | 4002 | spin_unlock(ptl); |
8f860591 | 4003 | } |
d833352a | 4004 | /* |
c8c06efa | 4005 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 4006 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 4007 | * once we release i_mmap_rwsem, another task can do the final put_page |
d833352a MG |
4008 | * and that page table be reused and filled with junk. |
4009 | */ | |
8f860591 | 4010 | flush_tlb_range(vma, start, end); |
34ee645e | 4011 | mmu_notifier_invalidate_range(mm, start, end); |
83cde9e8 | 4012 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
a5338093 | 4013 | mmu_notifier_invalidate_range_end(mm, start, end); |
7da4d641 PZ |
4014 | |
4015 | return pages << h->order; | |
8f860591 ZY |
4016 | } |
4017 | ||
a1e78772 MG |
4018 | int hugetlb_reserve_pages(struct inode *inode, |
4019 | long from, long to, | |
5a6fe125 | 4020 | struct vm_area_struct *vma, |
ca16d140 | 4021 | vm_flags_t vm_flags) |
e4e574b7 | 4022 | { |
17c9d12e | 4023 | long ret, chg; |
a5516438 | 4024 | struct hstate *h = hstate_inode(inode); |
90481622 | 4025 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 4026 | struct resv_map *resv_map; |
1c5ecae3 | 4027 | long gbl_reserve; |
e4e574b7 | 4028 | |
17c9d12e MG |
4029 | /* |
4030 | * Only apply hugepage reservation if asked. At fault time, an | |
4031 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 4032 | * without using reserves |
17c9d12e | 4033 | */ |
ca16d140 | 4034 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
4035 | return 0; |
4036 | ||
a1e78772 MG |
4037 | /* |
4038 | * Shared mappings base their reservation on the number of pages that | |
4039 | * are already allocated on behalf of the file. Private mappings need | |
4040 | * to reserve the full area even if read-only as mprotect() may be | |
4041 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
4042 | */ | |
9119a41e | 4043 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4e35f483 | 4044 | resv_map = inode_resv_map(inode); |
9119a41e | 4045 | |
1406ec9b | 4046 | chg = region_chg(resv_map, from, to); |
9119a41e JK |
4047 | |
4048 | } else { | |
4049 | resv_map = resv_map_alloc(); | |
17c9d12e MG |
4050 | if (!resv_map) |
4051 | return -ENOMEM; | |
4052 | ||
a1e78772 | 4053 | chg = to - from; |
84afd99b | 4054 | |
17c9d12e MG |
4055 | set_vma_resv_map(vma, resv_map); |
4056 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
4057 | } | |
4058 | ||
c50ac050 DH |
4059 | if (chg < 0) { |
4060 | ret = chg; | |
4061 | goto out_err; | |
4062 | } | |
8a630112 | 4063 | |
1c5ecae3 MK |
4064 | /* |
4065 | * There must be enough pages in the subpool for the mapping. If | |
4066 | * the subpool has a minimum size, there may be some global | |
4067 | * reservations already in place (gbl_reserve). | |
4068 | */ | |
4069 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
4070 | if (gbl_reserve < 0) { | |
c50ac050 DH |
4071 | ret = -ENOSPC; |
4072 | goto out_err; | |
4073 | } | |
5a6fe125 MG |
4074 | |
4075 | /* | |
17c9d12e | 4076 | * Check enough hugepages are available for the reservation. |
90481622 | 4077 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 4078 | */ |
1c5ecae3 | 4079 | ret = hugetlb_acct_memory(h, gbl_reserve); |
68842c9b | 4080 | if (ret < 0) { |
1c5ecae3 MK |
4081 | /* put back original number of pages, chg */ |
4082 | (void)hugepage_subpool_put_pages(spool, chg); | |
c50ac050 | 4083 | goto out_err; |
68842c9b | 4084 | } |
17c9d12e MG |
4085 | |
4086 | /* | |
4087 | * Account for the reservations made. Shared mappings record regions | |
4088 | * that have reservations as they are shared by multiple VMAs. | |
4089 | * When the last VMA disappears, the region map says how much | |
4090 | * the reservation was and the page cache tells how much of | |
4091 | * the reservation was consumed. Private mappings are per-VMA and | |
4092 | * only the consumed reservations are tracked. When the VMA | |
4093 | * disappears, the original reservation is the VMA size and the | |
4094 | * consumed reservations are stored in the map. Hence, nothing | |
4095 | * else has to be done for private mappings here | |
4096 | */ | |
33039678 MK |
4097 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4098 | long add = region_add(resv_map, from, to); | |
4099 | ||
4100 | if (unlikely(chg > add)) { | |
4101 | /* | |
4102 | * pages in this range were added to the reserve | |
4103 | * map between region_chg and region_add. This | |
4104 | * indicates a race with alloc_huge_page. Adjust | |
4105 | * the subpool and reserve counts modified above | |
4106 | * based on the difference. | |
4107 | */ | |
4108 | long rsv_adjust; | |
4109 | ||
4110 | rsv_adjust = hugepage_subpool_put_pages(spool, | |
4111 | chg - add); | |
4112 | hugetlb_acct_memory(h, -rsv_adjust); | |
4113 | } | |
4114 | } | |
a43a8c39 | 4115 | return 0; |
c50ac050 | 4116 | out_err: |
5e911373 MK |
4117 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
4118 | region_abort(resv_map, from, to); | |
f031dd27 JK |
4119 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
4120 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 4121 | return ret; |
a43a8c39 CK |
4122 | } |
4123 | ||
b5cec28d MK |
4124 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
4125 | long freed) | |
a43a8c39 | 4126 | { |
a5516438 | 4127 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 4128 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 4129 | long chg = 0; |
90481622 | 4130 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 4131 | long gbl_reserve; |
45c682a6 | 4132 | |
b5cec28d MK |
4133 | if (resv_map) { |
4134 | chg = region_del(resv_map, start, end); | |
4135 | /* | |
4136 | * region_del() can fail in the rare case where a region | |
4137 | * must be split and another region descriptor can not be | |
4138 | * allocated. If end == LONG_MAX, it will not fail. | |
4139 | */ | |
4140 | if (chg < 0) | |
4141 | return chg; | |
4142 | } | |
4143 | ||
45c682a6 | 4144 | spin_lock(&inode->i_lock); |
e4c6f8be | 4145 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
4146 | spin_unlock(&inode->i_lock); |
4147 | ||
1c5ecae3 MK |
4148 | /* |
4149 | * If the subpool has a minimum size, the number of global | |
4150 | * reservations to be released may be adjusted. | |
4151 | */ | |
4152 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
4153 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
4154 | |
4155 | return 0; | |
a43a8c39 | 4156 | } |
93f70f90 | 4157 | |
3212b535 SC |
4158 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
4159 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
4160 | struct vm_area_struct *vma, | |
4161 | unsigned long addr, pgoff_t idx) | |
4162 | { | |
4163 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
4164 | svma->vm_start; | |
4165 | unsigned long sbase = saddr & PUD_MASK; | |
4166 | unsigned long s_end = sbase + PUD_SIZE; | |
4167 | ||
4168 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
4169 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
4170 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
4171 | |
4172 | /* | |
4173 | * match the virtual addresses, permission and the alignment of the | |
4174 | * page table page. | |
4175 | */ | |
4176 | if (pmd_index(addr) != pmd_index(saddr) || | |
4177 | vm_flags != svm_flags || | |
4178 | sbase < svma->vm_start || svma->vm_end < s_end) | |
4179 | return 0; | |
4180 | ||
4181 | return saddr; | |
4182 | } | |
4183 | ||
31aafb45 | 4184 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
4185 | { |
4186 | unsigned long base = addr & PUD_MASK; | |
4187 | unsigned long end = base + PUD_SIZE; | |
4188 | ||
4189 | /* | |
4190 | * check on proper vm_flags and page table alignment | |
4191 | */ | |
4192 | if (vma->vm_flags & VM_MAYSHARE && | |
4193 | vma->vm_start <= base && end <= vma->vm_end) | |
31aafb45 NK |
4194 | return true; |
4195 | return false; | |
3212b535 SC |
4196 | } |
4197 | ||
4198 | /* | |
4199 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
4200 | * and returns the corresponding pte. While this is not necessary for the | |
4201 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
4202 | * code much cleaner. pmd allocation is essential for the shared case because | |
c8c06efa | 4203 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise |
3212b535 SC |
4204 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a |
4205 | * bad pmd for sharing. | |
4206 | */ | |
4207 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4208 | { | |
4209 | struct vm_area_struct *vma = find_vma(mm, addr); | |
4210 | struct address_space *mapping = vma->vm_file->f_mapping; | |
4211 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
4212 | vma->vm_pgoff; | |
4213 | struct vm_area_struct *svma; | |
4214 | unsigned long saddr; | |
4215 | pte_t *spte = NULL; | |
4216 | pte_t *pte; | |
cb900f41 | 4217 | spinlock_t *ptl; |
3212b535 SC |
4218 | |
4219 | if (!vma_shareable(vma, addr)) | |
4220 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
4221 | ||
83cde9e8 | 4222 | i_mmap_lock_write(mapping); |
3212b535 SC |
4223 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
4224 | if (svma == vma) | |
4225 | continue; | |
4226 | ||
4227 | saddr = page_table_shareable(svma, vma, addr, idx); | |
4228 | if (saddr) { | |
4229 | spte = huge_pte_offset(svma->vm_mm, saddr); | |
4230 | if (spte) { | |
dc6c9a35 | 4231 | mm_inc_nr_pmds(mm); |
3212b535 SC |
4232 | get_page(virt_to_page(spte)); |
4233 | break; | |
4234 | } | |
4235 | } | |
4236 | } | |
4237 | ||
4238 | if (!spte) | |
4239 | goto out; | |
4240 | ||
cb900f41 KS |
4241 | ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); |
4242 | spin_lock(ptl); | |
dc6c9a35 | 4243 | if (pud_none(*pud)) { |
3212b535 SC |
4244 | pud_populate(mm, pud, |
4245 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
dc6c9a35 | 4246 | } else { |
3212b535 | 4247 | put_page(virt_to_page(spte)); |
dc6c9a35 KS |
4248 | mm_inc_nr_pmds(mm); |
4249 | } | |
cb900f41 | 4250 | spin_unlock(ptl); |
3212b535 SC |
4251 | out: |
4252 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
83cde9e8 | 4253 | i_mmap_unlock_write(mapping); |
3212b535 SC |
4254 | return pte; |
4255 | } | |
4256 | ||
4257 | /* | |
4258 | * unmap huge page backed by shared pte. | |
4259 | * | |
4260 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
4261 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
4262 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
4263 | * | |
cb900f41 | 4264 | * called with page table lock held. |
3212b535 SC |
4265 | * |
4266 | * returns: 1 successfully unmapped a shared pte page | |
4267 | * 0 the underlying pte page is not shared, or it is the last user | |
4268 | */ | |
4269 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4270 | { | |
4271 | pgd_t *pgd = pgd_offset(mm, *addr); | |
4272 | pud_t *pud = pud_offset(pgd, *addr); | |
4273 | ||
4274 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
4275 | if (page_count(virt_to_page(ptep)) == 1) | |
4276 | return 0; | |
4277 | ||
4278 | pud_clear(pud); | |
4279 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 4280 | mm_dec_nr_pmds(mm); |
3212b535 SC |
4281 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
4282 | return 1; | |
4283 | } | |
9e5fc74c SC |
4284 | #define want_pmd_share() (1) |
4285 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
4286 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4287 | { | |
4288 | return NULL; | |
4289 | } | |
e81f2d22 ZZ |
4290 | |
4291 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4292 | { | |
4293 | return 0; | |
4294 | } | |
9e5fc74c | 4295 | #define want_pmd_share() (0) |
3212b535 SC |
4296 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
4297 | ||
9e5fc74c SC |
4298 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
4299 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
4300 | unsigned long addr, unsigned long sz) | |
4301 | { | |
4302 | pgd_t *pgd; | |
4303 | pud_t *pud; | |
4304 | pte_t *pte = NULL; | |
4305 | ||
4306 | pgd = pgd_offset(mm, addr); | |
4307 | pud = pud_alloc(mm, pgd, addr); | |
4308 | if (pud) { | |
4309 | if (sz == PUD_SIZE) { | |
4310 | pte = (pte_t *)pud; | |
4311 | } else { | |
4312 | BUG_ON(sz != PMD_SIZE); | |
4313 | if (want_pmd_share() && pud_none(*pud)) | |
4314 | pte = huge_pmd_share(mm, addr, pud); | |
4315 | else | |
4316 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
4317 | } | |
4318 | } | |
4319 | BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); | |
4320 | ||
4321 | return pte; | |
4322 | } | |
4323 | ||
4324 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) | |
4325 | { | |
4326 | pgd_t *pgd; | |
4327 | pud_t *pud; | |
4328 | pmd_t *pmd = NULL; | |
4329 | ||
4330 | pgd = pgd_offset(mm, addr); | |
4331 | if (pgd_present(*pgd)) { | |
4332 | pud = pud_offset(pgd, addr); | |
4333 | if (pud_present(*pud)) { | |
4334 | if (pud_huge(*pud)) | |
4335 | return (pte_t *)pud; | |
4336 | pmd = pmd_offset(pud, addr); | |
4337 | } | |
4338 | } | |
4339 | return (pte_t *) pmd; | |
4340 | } | |
4341 | ||
61f77eda NH |
4342 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
4343 | ||
4344 | /* | |
4345 | * These functions are overwritable if your architecture needs its own | |
4346 | * behavior. | |
4347 | */ | |
4348 | struct page * __weak | |
4349 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
4350 | int write) | |
4351 | { | |
4352 | return ERR_PTR(-EINVAL); | |
4353 | } | |
4354 | ||
4355 | struct page * __weak | |
9e5fc74c | 4356 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 4357 | pmd_t *pmd, int flags) |
9e5fc74c | 4358 | { |
e66f17ff NH |
4359 | struct page *page = NULL; |
4360 | spinlock_t *ptl; | |
4361 | retry: | |
4362 | ptl = pmd_lockptr(mm, pmd); | |
4363 | spin_lock(ptl); | |
4364 | /* | |
4365 | * make sure that the address range covered by this pmd is not | |
4366 | * unmapped from other threads. | |
4367 | */ | |
4368 | if (!pmd_huge(*pmd)) | |
4369 | goto out; | |
4370 | if (pmd_present(*pmd)) { | |
97534127 | 4371 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
e66f17ff NH |
4372 | if (flags & FOLL_GET) |
4373 | get_page(page); | |
4374 | } else { | |
4375 | if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) { | |
4376 | spin_unlock(ptl); | |
4377 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
4378 | goto retry; | |
4379 | } | |
4380 | /* | |
4381 | * hwpoisoned entry is treated as no_page_table in | |
4382 | * follow_page_mask(). | |
4383 | */ | |
4384 | } | |
4385 | out: | |
4386 | spin_unlock(ptl); | |
9e5fc74c SC |
4387 | return page; |
4388 | } | |
4389 | ||
61f77eda | 4390 | struct page * __weak |
9e5fc74c | 4391 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 4392 | pud_t *pud, int flags) |
9e5fc74c | 4393 | { |
e66f17ff NH |
4394 | if (flags & FOLL_GET) |
4395 | return NULL; | |
9e5fc74c | 4396 | |
e66f17ff | 4397 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
4398 | } |
4399 | ||
d5bd9106 AK |
4400 | #ifdef CONFIG_MEMORY_FAILURE |
4401 | ||
93f70f90 NH |
4402 | /* |
4403 | * This function is called from memory failure code. | |
4404 | * Assume the caller holds page lock of the head page. | |
4405 | */ | |
6de2b1aa | 4406 | int dequeue_hwpoisoned_huge_page(struct page *hpage) |
93f70f90 NH |
4407 | { |
4408 | struct hstate *h = page_hstate(hpage); | |
4409 | int nid = page_to_nid(hpage); | |
6de2b1aa | 4410 | int ret = -EBUSY; |
93f70f90 NH |
4411 | |
4412 | spin_lock(&hugetlb_lock); | |
7e1f049e NH |
4413 | /* |
4414 | * Just checking !page_huge_active is not enough, because that could be | |
4415 | * an isolated/hwpoisoned hugepage (which have >0 refcount). | |
4416 | */ | |
4417 | if (!page_huge_active(hpage) && !page_count(hpage)) { | |
56f2fb14 NH |
4418 | /* |
4419 | * Hwpoisoned hugepage isn't linked to activelist or freelist, | |
4420 | * but dangling hpage->lru can trigger list-debug warnings | |
4421 | * (this happens when we call unpoison_memory() on it), | |
4422 | * so let it point to itself with list_del_init(). | |
4423 | */ | |
4424 | list_del_init(&hpage->lru); | |
8c6c2ecb | 4425 | set_page_refcounted(hpage); |
6de2b1aa NH |
4426 | h->free_huge_pages--; |
4427 | h->free_huge_pages_node[nid]--; | |
4428 | ret = 0; | |
4429 | } | |
93f70f90 | 4430 | spin_unlock(&hugetlb_lock); |
6de2b1aa | 4431 | return ret; |
93f70f90 | 4432 | } |
6de2b1aa | 4433 | #endif |
31caf665 NH |
4434 | |
4435 | bool isolate_huge_page(struct page *page, struct list_head *list) | |
4436 | { | |
bcc54222 NH |
4437 | bool ret = true; |
4438 | ||
309381fe | 4439 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4440 | spin_lock(&hugetlb_lock); |
bcc54222 NH |
4441 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { |
4442 | ret = false; | |
4443 | goto unlock; | |
4444 | } | |
4445 | clear_page_huge_active(page); | |
31caf665 | 4446 | list_move_tail(&page->lru, list); |
bcc54222 | 4447 | unlock: |
31caf665 | 4448 | spin_unlock(&hugetlb_lock); |
bcc54222 | 4449 | return ret; |
31caf665 NH |
4450 | } |
4451 | ||
4452 | void putback_active_hugepage(struct page *page) | |
4453 | { | |
309381fe | 4454 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4455 | spin_lock(&hugetlb_lock); |
bcc54222 | 4456 | set_page_huge_active(page); |
31caf665 NH |
4457 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
4458 | spin_unlock(&hugetlb_lock); | |
4459 | put_page(page); | |
4460 | } |