Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * Generic hugetlb support. | |
6d49e352 | 3 | * (C) Nadia Yvette Chambers, April 2004 |
1da177e4 | 4 | */ |
1da177e4 LT |
5 | #include <linux/list.h> |
6 | #include <linux/init.h> | |
1da177e4 | 7 | #include <linux/mm.h> |
e1759c21 | 8 | #include <linux/seq_file.h> |
1da177e4 LT |
9 | #include <linux/sysctl.h> |
10 | #include <linux/highmem.h> | |
cddb8a5c | 11 | #include <linux/mmu_notifier.h> |
1da177e4 | 12 | #include <linux/nodemask.h> |
63551ae0 | 13 | #include <linux/pagemap.h> |
5da7ca86 | 14 | #include <linux/mempolicy.h> |
3b32123d | 15 | #include <linux/compiler.h> |
aea47ff3 | 16 | #include <linux/cpuset.h> |
3935baa9 | 17 | #include <linux/mutex.h> |
aa888a74 | 18 | #include <linux/bootmem.h> |
a3437870 | 19 | #include <linux/sysfs.h> |
5a0e3ad6 | 20 | #include <linux/slab.h> |
174cd4b1 | 21 | #include <linux/sched/signal.h> |
0fe6e20b | 22 | #include <linux/rmap.h> |
fd6a03ed NH |
23 | #include <linux/swap.h> |
24 | #include <linux/swapops.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> |
1a1aad8a | 35 | #include <linux/userfaultfd_k.h> |
7835e98b | 36 | #include "internal.h" |
1da177e4 | 37 | |
753162cd | 38 | int hugepages_treat_as_movable; |
a5516438 | 39 | |
c3f38a38 | 40 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
41 | unsigned int default_hstate_idx; |
42 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
641844f5 NH |
43 | /* |
44 | * Minimum page order among possible hugepage sizes, set to a proper value | |
45 | * at boot time. | |
46 | */ | |
47 | static unsigned int minimum_order __read_mostly = UINT_MAX; | |
e5ff2159 | 48 | |
53ba51d2 JT |
49 | __initdata LIST_HEAD(huge_boot_pages); |
50 | ||
e5ff2159 AK |
51 | /* for command line parsing */ |
52 | static struct hstate * __initdata parsed_hstate; | |
53 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 54 | static unsigned long __initdata default_hstate_size; |
9fee021d | 55 | static bool __initdata parsed_valid_hugepagesz = true; |
e5ff2159 | 56 | |
3935baa9 | 57 | /* |
31caf665 NH |
58 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
59 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 60 | */ |
c3f38a38 | 61 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 62 | |
8382d914 DB |
63 | /* |
64 | * Serializes faults on the same logical page. This is used to | |
65 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
66 | */ | |
67 | static int num_fault_mutexes; | |
c672c7f2 | 68 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
8382d914 | 69 | |
7ca02d0a MK |
70 | /* Forward declaration */ |
71 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
72 | ||
90481622 DG |
73 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
74 | { | |
75 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
76 | ||
77 | spin_unlock(&spool->lock); | |
78 | ||
79 | /* If no pages are used, and no other handles to the subpool | |
7ca02d0a MK |
80 | * remain, give up any reservations mased on minimum size and |
81 | * free the subpool */ | |
82 | if (free) { | |
83 | if (spool->min_hpages != -1) | |
84 | hugetlb_acct_memory(spool->hstate, | |
85 | -spool->min_hpages); | |
90481622 | 86 | kfree(spool); |
7ca02d0a | 87 | } |
90481622 DG |
88 | } |
89 | ||
7ca02d0a MK |
90 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
91 | long min_hpages) | |
90481622 DG |
92 | { |
93 | struct hugepage_subpool *spool; | |
94 | ||
c6a91820 | 95 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
96 | if (!spool) |
97 | return NULL; | |
98 | ||
99 | spin_lock_init(&spool->lock); | |
100 | spool->count = 1; | |
7ca02d0a MK |
101 | spool->max_hpages = max_hpages; |
102 | spool->hstate = h; | |
103 | spool->min_hpages = min_hpages; | |
104 | ||
105 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
106 | kfree(spool); | |
107 | return NULL; | |
108 | } | |
109 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
110 | |
111 | return spool; | |
112 | } | |
113 | ||
114 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
115 | { | |
116 | spin_lock(&spool->lock); | |
117 | BUG_ON(!spool->count); | |
118 | spool->count--; | |
119 | unlock_or_release_subpool(spool); | |
120 | } | |
121 | ||
1c5ecae3 MK |
122 | /* |
123 | * Subpool accounting for allocating and reserving pages. | |
124 | * Return -ENOMEM if there are not enough resources to satisfy the | |
125 | * the request. Otherwise, return the number of pages by which the | |
126 | * global pools must be adjusted (upward). The returned value may | |
127 | * only be different than the passed value (delta) in the case where | |
128 | * a subpool minimum size must be manitained. | |
129 | */ | |
130 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
131 | long delta) |
132 | { | |
1c5ecae3 | 133 | long ret = delta; |
90481622 DG |
134 | |
135 | if (!spool) | |
1c5ecae3 | 136 | return ret; |
90481622 DG |
137 | |
138 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
139 | |
140 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
141 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
142 | spool->used_hpages += delta; | |
143 | else { | |
144 | ret = -ENOMEM; | |
145 | goto unlock_ret; | |
146 | } | |
90481622 | 147 | } |
90481622 | 148 | |
09a95e29 MK |
149 | /* minimum size accounting */ |
150 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
151 | if (delta > spool->rsv_hpages) { |
152 | /* | |
153 | * Asking for more reserves than those already taken on | |
154 | * behalf of subpool. Return difference. | |
155 | */ | |
156 | ret = delta - spool->rsv_hpages; | |
157 | spool->rsv_hpages = 0; | |
158 | } else { | |
159 | ret = 0; /* reserves already accounted for */ | |
160 | spool->rsv_hpages -= delta; | |
161 | } | |
162 | } | |
163 | ||
164 | unlock_ret: | |
165 | spin_unlock(&spool->lock); | |
90481622 DG |
166 | return ret; |
167 | } | |
168 | ||
1c5ecae3 MK |
169 | /* |
170 | * Subpool accounting for freeing and unreserving pages. | |
171 | * Return the number of global page reservations that must be dropped. | |
172 | * The return value may only be different than the passed value (delta) | |
173 | * in the case where a subpool minimum size must be maintained. | |
174 | */ | |
175 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
176 | long delta) |
177 | { | |
1c5ecae3 MK |
178 | long ret = delta; |
179 | ||
90481622 | 180 | if (!spool) |
1c5ecae3 | 181 | return delta; |
90481622 DG |
182 | |
183 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
184 | |
185 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
186 | spool->used_hpages -= delta; | |
187 | ||
09a95e29 MK |
188 | /* minimum size accounting */ |
189 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
190 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
191 | ret = 0; | |
192 | else | |
193 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
194 | ||
195 | spool->rsv_hpages += delta; | |
196 | if (spool->rsv_hpages > spool->min_hpages) | |
197 | spool->rsv_hpages = spool->min_hpages; | |
198 | } | |
199 | ||
200 | /* | |
201 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
202 | * quota reference, free it now. | |
203 | */ | |
90481622 | 204 | unlock_or_release_subpool(spool); |
1c5ecae3 MK |
205 | |
206 | return ret; | |
90481622 DG |
207 | } |
208 | ||
209 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
210 | { | |
211 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
212 | } | |
213 | ||
214 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
215 | { | |
496ad9aa | 216 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
217 | } |
218 | ||
96822904 AW |
219 | /* |
220 | * Region tracking -- allows tracking of reservations and instantiated pages | |
221 | * across the pages in a mapping. | |
84afd99b | 222 | * |
1dd308a7 MK |
223 | * The region data structures are embedded into a resv_map and protected |
224 | * by a resv_map's lock. The set of regions within the resv_map represent | |
225 | * reservations for huge pages, or huge pages that have already been | |
226 | * instantiated within the map. The from and to elements are huge page | |
227 | * indicies into the associated mapping. from indicates the starting index | |
228 | * of the region. to represents the first index past the end of the region. | |
229 | * | |
230 | * For example, a file region structure with from == 0 and to == 4 represents | |
231 | * four huge pages in a mapping. It is important to note that the to element | |
232 | * represents the first element past the end of the region. This is used in | |
233 | * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. | |
234 | * | |
235 | * Interval notation of the form [from, to) will be used to indicate that | |
236 | * the endpoint from is inclusive and to is exclusive. | |
96822904 AW |
237 | */ |
238 | struct file_region { | |
239 | struct list_head link; | |
240 | long from; | |
241 | long to; | |
242 | }; | |
243 | ||
1dd308a7 MK |
244 | /* |
245 | * Add the huge page range represented by [f, t) to the reserve | |
5e911373 MK |
246 | * map. In the normal case, existing regions will be expanded |
247 | * to accommodate the specified range. Sufficient regions should | |
248 | * exist for expansion due to the previous call to region_chg | |
249 | * with the same range. However, it is possible that region_del | |
250 | * could have been called after region_chg and modifed the map | |
251 | * in such a way that no region exists to be expanded. In this | |
252 | * case, pull a region descriptor from the cache associated with | |
253 | * the map and use that for the new range. | |
cf3ad20b MK |
254 | * |
255 | * Return the number of new huge pages added to the map. This | |
256 | * number is greater than or equal to zero. | |
1dd308a7 | 257 | */ |
1406ec9b | 258 | static long region_add(struct resv_map *resv, long f, long t) |
96822904 | 259 | { |
1406ec9b | 260 | struct list_head *head = &resv->regions; |
96822904 | 261 | struct file_region *rg, *nrg, *trg; |
cf3ad20b | 262 | long add = 0; |
96822904 | 263 | |
7b24d861 | 264 | spin_lock(&resv->lock); |
96822904 AW |
265 | /* Locate the region we are either in or before. */ |
266 | list_for_each_entry(rg, head, link) | |
267 | if (f <= rg->to) | |
268 | break; | |
269 | ||
5e911373 MK |
270 | /* |
271 | * If no region exists which can be expanded to include the | |
272 | * specified range, the list must have been modified by an | |
273 | * interleving call to region_del(). Pull a region descriptor | |
274 | * from the cache and use it for this range. | |
275 | */ | |
276 | if (&rg->link == head || t < rg->from) { | |
277 | VM_BUG_ON(resv->region_cache_count <= 0); | |
278 | ||
279 | resv->region_cache_count--; | |
280 | nrg = list_first_entry(&resv->region_cache, struct file_region, | |
281 | link); | |
282 | list_del(&nrg->link); | |
283 | ||
284 | nrg->from = f; | |
285 | nrg->to = t; | |
286 | list_add(&nrg->link, rg->link.prev); | |
287 | ||
288 | add += t - f; | |
289 | goto out_locked; | |
290 | } | |
291 | ||
96822904 AW |
292 | /* Round our left edge to the current segment if it encloses us. */ |
293 | if (f > rg->from) | |
294 | f = rg->from; | |
295 | ||
296 | /* Check for and consume any regions we now overlap with. */ | |
297 | nrg = rg; | |
298 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
299 | if (&rg->link == head) | |
300 | break; | |
301 | if (rg->from > t) | |
302 | break; | |
303 | ||
304 | /* If this area reaches higher then extend our area to | |
305 | * include it completely. If this is not the first area | |
306 | * which we intend to reuse, free it. */ | |
307 | if (rg->to > t) | |
308 | t = rg->to; | |
309 | if (rg != nrg) { | |
cf3ad20b MK |
310 | /* Decrement return value by the deleted range. |
311 | * Another range will span this area so that by | |
312 | * end of routine add will be >= zero | |
313 | */ | |
314 | add -= (rg->to - rg->from); | |
96822904 AW |
315 | list_del(&rg->link); |
316 | kfree(rg); | |
317 | } | |
318 | } | |
cf3ad20b MK |
319 | |
320 | add += (nrg->from - f); /* Added to beginning of region */ | |
96822904 | 321 | nrg->from = f; |
cf3ad20b | 322 | add += t - nrg->to; /* Added to end of region */ |
96822904 | 323 | nrg->to = t; |
cf3ad20b | 324 | |
5e911373 MK |
325 | out_locked: |
326 | resv->adds_in_progress--; | |
7b24d861 | 327 | spin_unlock(&resv->lock); |
cf3ad20b MK |
328 | VM_BUG_ON(add < 0); |
329 | return add; | |
96822904 AW |
330 | } |
331 | ||
1dd308a7 MK |
332 | /* |
333 | * Examine the existing reserve map and determine how many | |
334 | * huge pages in the specified range [f, t) are NOT currently | |
335 | * represented. This routine is called before a subsequent | |
336 | * call to region_add that will actually modify the reserve | |
337 | * map to add the specified range [f, t). region_chg does | |
338 | * not change the number of huge pages represented by the | |
339 | * map. However, if the existing regions in the map can not | |
340 | * be expanded to represent the new range, a new file_region | |
341 | * structure is added to the map as a placeholder. This is | |
342 | * so that the subsequent region_add call will have all the | |
343 | * regions it needs and will not fail. | |
344 | * | |
5e911373 MK |
345 | * Upon entry, region_chg will also examine the cache of region descriptors |
346 | * associated with the map. If there are not enough descriptors cached, one | |
347 | * will be allocated for the in progress add operation. | |
348 | * | |
349 | * Returns the number of huge pages that need to be added to the existing | |
350 | * reservation map for the range [f, t). This number is greater or equal to | |
351 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
352 | * is needed and can not be allocated. | |
1dd308a7 | 353 | */ |
1406ec9b | 354 | static long region_chg(struct resv_map *resv, long f, long t) |
96822904 | 355 | { |
1406ec9b | 356 | struct list_head *head = &resv->regions; |
7b24d861 | 357 | struct file_region *rg, *nrg = NULL; |
96822904 AW |
358 | long chg = 0; |
359 | ||
7b24d861 DB |
360 | retry: |
361 | spin_lock(&resv->lock); | |
5e911373 MK |
362 | retry_locked: |
363 | resv->adds_in_progress++; | |
364 | ||
365 | /* | |
366 | * Check for sufficient descriptors in the cache to accommodate | |
367 | * the number of in progress add operations. | |
368 | */ | |
369 | if (resv->adds_in_progress > resv->region_cache_count) { | |
370 | struct file_region *trg; | |
371 | ||
372 | VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1); | |
373 | /* Must drop lock to allocate a new descriptor. */ | |
374 | resv->adds_in_progress--; | |
375 | spin_unlock(&resv->lock); | |
376 | ||
377 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
dbe409e4 MK |
378 | if (!trg) { |
379 | kfree(nrg); | |
5e911373 | 380 | return -ENOMEM; |
dbe409e4 | 381 | } |
5e911373 MK |
382 | |
383 | spin_lock(&resv->lock); | |
384 | list_add(&trg->link, &resv->region_cache); | |
385 | resv->region_cache_count++; | |
386 | goto retry_locked; | |
387 | } | |
388 | ||
96822904 AW |
389 | /* Locate the region we are before or in. */ |
390 | list_for_each_entry(rg, head, link) | |
391 | if (f <= rg->to) | |
392 | break; | |
393 | ||
394 | /* If we are below the current region then a new region is required. | |
395 | * Subtle, allocate a new region at the position but make it zero | |
396 | * size such that we can guarantee to record the reservation. */ | |
397 | if (&rg->link == head || t < rg->from) { | |
7b24d861 | 398 | if (!nrg) { |
5e911373 | 399 | resv->adds_in_progress--; |
7b24d861 DB |
400 | spin_unlock(&resv->lock); |
401 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
402 | if (!nrg) | |
403 | return -ENOMEM; | |
404 | ||
405 | nrg->from = f; | |
406 | nrg->to = f; | |
407 | INIT_LIST_HEAD(&nrg->link); | |
408 | goto retry; | |
409 | } | |
96822904 | 410 | |
7b24d861 DB |
411 | list_add(&nrg->link, rg->link.prev); |
412 | chg = t - f; | |
413 | goto out_nrg; | |
96822904 AW |
414 | } |
415 | ||
416 | /* Round our left edge to the current segment if it encloses us. */ | |
417 | if (f > rg->from) | |
418 | f = rg->from; | |
419 | chg = t - f; | |
420 | ||
421 | /* Check for and consume any regions we now overlap with. */ | |
422 | list_for_each_entry(rg, rg->link.prev, link) { | |
423 | if (&rg->link == head) | |
424 | break; | |
425 | if (rg->from > t) | |
7b24d861 | 426 | goto out; |
96822904 | 427 | |
25985edc | 428 | /* We overlap with this area, if it extends further than |
96822904 AW |
429 | * us then we must extend ourselves. Account for its |
430 | * existing reservation. */ | |
431 | if (rg->to > t) { | |
432 | chg += rg->to - t; | |
433 | t = rg->to; | |
434 | } | |
435 | chg -= rg->to - rg->from; | |
436 | } | |
7b24d861 DB |
437 | |
438 | out: | |
439 | spin_unlock(&resv->lock); | |
440 | /* We already know we raced and no longer need the new region */ | |
441 | kfree(nrg); | |
442 | return chg; | |
443 | out_nrg: | |
444 | spin_unlock(&resv->lock); | |
96822904 AW |
445 | return chg; |
446 | } | |
447 | ||
5e911373 MK |
448 | /* |
449 | * Abort the in progress add operation. The adds_in_progress field | |
450 | * of the resv_map keeps track of the operations in progress between | |
451 | * calls to region_chg and region_add. Operations are sometimes | |
452 | * aborted after the call to region_chg. In such cases, region_abort | |
453 | * is called to decrement the adds_in_progress counter. | |
454 | * | |
455 | * NOTE: The range arguments [f, t) are not needed or used in this | |
456 | * routine. They are kept to make reading the calling code easier as | |
457 | * arguments will match the associated region_chg call. | |
458 | */ | |
459 | static void region_abort(struct resv_map *resv, long f, long t) | |
460 | { | |
461 | spin_lock(&resv->lock); | |
462 | VM_BUG_ON(!resv->region_cache_count); | |
463 | resv->adds_in_progress--; | |
464 | spin_unlock(&resv->lock); | |
465 | } | |
466 | ||
1dd308a7 | 467 | /* |
feba16e2 MK |
468 | * Delete the specified range [f, t) from the reserve map. If the |
469 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
470 | * should be deleted. Locate the regions which intersect [f, t) | |
471 | * and either trim, delete or split the existing regions. | |
472 | * | |
473 | * Returns the number of huge pages deleted from the reserve map. | |
474 | * In the normal case, the return value is zero or more. In the | |
475 | * case where a region must be split, a new region descriptor must | |
476 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
477 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
478 | * a region and possibly return -ENOMEM. Callers specifying | |
479 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 480 | */ |
feba16e2 | 481 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 482 | { |
1406ec9b | 483 | struct list_head *head = &resv->regions; |
96822904 | 484 | struct file_region *rg, *trg; |
feba16e2 MK |
485 | struct file_region *nrg = NULL; |
486 | long del = 0; | |
96822904 | 487 | |
feba16e2 | 488 | retry: |
7b24d861 | 489 | spin_lock(&resv->lock); |
feba16e2 | 490 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
491 | /* |
492 | * Skip regions before the range to be deleted. file_region | |
493 | * ranges are normally of the form [from, to). However, there | |
494 | * may be a "placeholder" entry in the map which is of the form | |
495 | * (from, to) with from == to. Check for placeholder entries | |
496 | * at the beginning of the range to be deleted. | |
497 | */ | |
498 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 499 | continue; |
dbe409e4 | 500 | |
feba16e2 | 501 | if (rg->from >= t) |
96822904 | 502 | break; |
96822904 | 503 | |
feba16e2 MK |
504 | if (f > rg->from && t < rg->to) { /* Must split region */ |
505 | /* | |
506 | * Check for an entry in the cache before dropping | |
507 | * lock and attempting allocation. | |
508 | */ | |
509 | if (!nrg && | |
510 | resv->region_cache_count > resv->adds_in_progress) { | |
511 | nrg = list_first_entry(&resv->region_cache, | |
512 | struct file_region, | |
513 | link); | |
514 | list_del(&nrg->link); | |
515 | resv->region_cache_count--; | |
516 | } | |
96822904 | 517 | |
feba16e2 MK |
518 | if (!nrg) { |
519 | spin_unlock(&resv->lock); | |
520 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
521 | if (!nrg) | |
522 | return -ENOMEM; | |
523 | goto retry; | |
524 | } | |
525 | ||
526 | del += t - f; | |
527 | ||
528 | /* New entry for end of split region */ | |
529 | nrg->from = t; | |
530 | nrg->to = rg->to; | |
531 | INIT_LIST_HEAD(&nrg->link); | |
532 | ||
533 | /* Original entry is trimmed */ | |
534 | rg->to = f; | |
535 | ||
536 | list_add(&nrg->link, &rg->link); | |
537 | nrg = NULL; | |
96822904 | 538 | break; |
feba16e2 MK |
539 | } |
540 | ||
541 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
542 | del += rg->to - rg->from; | |
543 | list_del(&rg->link); | |
544 | kfree(rg); | |
545 | continue; | |
546 | } | |
547 | ||
548 | if (f <= rg->from) { /* Trim beginning of region */ | |
549 | del += t - rg->from; | |
550 | rg->from = t; | |
551 | } else { /* Trim end of region */ | |
552 | del += rg->to - f; | |
553 | rg->to = f; | |
554 | } | |
96822904 | 555 | } |
7b24d861 | 556 | |
7b24d861 | 557 | spin_unlock(&resv->lock); |
feba16e2 MK |
558 | kfree(nrg); |
559 | return del; | |
96822904 AW |
560 | } |
561 | ||
b5cec28d MK |
562 | /* |
563 | * A rare out of memory error was encountered which prevented removal of | |
564 | * the reserve map region for a page. The huge page itself was free'ed | |
565 | * and removed from the page cache. This routine will adjust the subpool | |
566 | * usage count, and the global reserve count if needed. By incrementing | |
567 | * these counts, the reserve map entry which could not be deleted will | |
568 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
569 | * counts. | |
570 | */ | |
72e2936c | 571 | void hugetlb_fix_reserve_counts(struct inode *inode) |
b5cec28d MK |
572 | { |
573 | struct hugepage_subpool *spool = subpool_inode(inode); | |
574 | long rsv_adjust; | |
575 | ||
576 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
72e2936c | 577 | if (rsv_adjust) { |
b5cec28d MK |
578 | struct hstate *h = hstate_inode(inode); |
579 | ||
580 | hugetlb_acct_memory(h, 1); | |
581 | } | |
582 | } | |
583 | ||
1dd308a7 MK |
584 | /* |
585 | * Count and return the number of huge pages in the reserve map | |
586 | * that intersect with the range [f, t). | |
587 | */ | |
1406ec9b | 588 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 589 | { |
1406ec9b | 590 | struct list_head *head = &resv->regions; |
84afd99b AW |
591 | struct file_region *rg; |
592 | long chg = 0; | |
593 | ||
7b24d861 | 594 | spin_lock(&resv->lock); |
84afd99b AW |
595 | /* Locate each segment we overlap with, and count that overlap. */ |
596 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
597 | long seg_from; |
598 | long seg_to; | |
84afd99b AW |
599 | |
600 | if (rg->to <= f) | |
601 | continue; | |
602 | if (rg->from >= t) | |
603 | break; | |
604 | ||
605 | seg_from = max(rg->from, f); | |
606 | seg_to = min(rg->to, t); | |
607 | ||
608 | chg += seg_to - seg_from; | |
609 | } | |
7b24d861 | 610 | spin_unlock(&resv->lock); |
84afd99b AW |
611 | |
612 | return chg; | |
613 | } | |
614 | ||
e7c4b0bf AW |
615 | /* |
616 | * Convert the address within this vma to the page offset within | |
617 | * the mapping, in pagecache page units; huge pages here. | |
618 | */ | |
a5516438 AK |
619 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
620 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 621 | { |
a5516438 AK |
622 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
623 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
624 | } |
625 | ||
0fe6e20b NH |
626 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
627 | unsigned long address) | |
628 | { | |
629 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
630 | } | |
dee41079 | 631 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 632 | |
08fba699 MG |
633 | /* |
634 | * Return the size of the pages allocated when backing a VMA. In the majority | |
635 | * cases this will be same size as used by the page table entries. | |
636 | */ | |
637 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
638 | { | |
639 | struct hstate *hstate; | |
640 | ||
641 | if (!is_vm_hugetlb_page(vma)) | |
642 | return PAGE_SIZE; | |
643 | ||
644 | hstate = hstate_vma(vma); | |
645 | ||
2415cf12 | 646 | return 1UL << huge_page_shift(hstate); |
08fba699 | 647 | } |
f340ca0f | 648 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 649 | |
3340289d MG |
650 | /* |
651 | * Return the page size being used by the MMU to back a VMA. In the majority | |
652 | * of cases, the page size used by the kernel matches the MMU size. On | |
653 | * architectures where it differs, an architecture-specific version of this | |
654 | * function is required. | |
655 | */ | |
656 | #ifndef vma_mmu_pagesize | |
657 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) | |
658 | { | |
659 | return vma_kernel_pagesize(vma); | |
660 | } | |
661 | #endif | |
662 | ||
84afd99b AW |
663 | /* |
664 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
665 | * bits of the reservation map pointer, which are always clear due to | |
666 | * alignment. | |
667 | */ | |
668 | #define HPAGE_RESV_OWNER (1UL << 0) | |
669 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 670 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 671 | |
a1e78772 MG |
672 | /* |
673 | * These helpers are used to track how many pages are reserved for | |
674 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
675 | * is guaranteed to have their future faults succeed. | |
676 | * | |
677 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
678 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
679 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
680 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
681 | * |
682 | * The private mapping reservation is represented in a subtly different | |
683 | * manner to a shared mapping. A shared mapping has a region map associated | |
684 | * with the underlying file, this region map represents the backing file | |
685 | * pages which have ever had a reservation assigned which this persists even | |
686 | * after the page is instantiated. A private mapping has a region map | |
687 | * associated with the original mmap which is attached to all VMAs which | |
688 | * reference it, this region map represents those offsets which have consumed | |
689 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 690 | */ |
e7c4b0bf AW |
691 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
692 | { | |
693 | return (unsigned long)vma->vm_private_data; | |
694 | } | |
695 | ||
696 | static void set_vma_private_data(struct vm_area_struct *vma, | |
697 | unsigned long value) | |
698 | { | |
699 | vma->vm_private_data = (void *)value; | |
700 | } | |
701 | ||
9119a41e | 702 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
703 | { |
704 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
705 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
706 | ||
707 | if (!resv_map || !rg) { | |
708 | kfree(resv_map); | |
709 | kfree(rg); | |
84afd99b | 710 | return NULL; |
5e911373 | 711 | } |
84afd99b AW |
712 | |
713 | kref_init(&resv_map->refs); | |
7b24d861 | 714 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
715 | INIT_LIST_HEAD(&resv_map->regions); |
716 | ||
5e911373 MK |
717 | resv_map->adds_in_progress = 0; |
718 | ||
719 | INIT_LIST_HEAD(&resv_map->region_cache); | |
720 | list_add(&rg->link, &resv_map->region_cache); | |
721 | resv_map->region_cache_count = 1; | |
722 | ||
84afd99b AW |
723 | return resv_map; |
724 | } | |
725 | ||
9119a41e | 726 | void resv_map_release(struct kref *ref) |
84afd99b AW |
727 | { |
728 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
729 | struct list_head *head = &resv_map->region_cache; |
730 | struct file_region *rg, *trg; | |
84afd99b AW |
731 | |
732 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 733 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
734 | |
735 | /* ... and any entries left in the cache */ | |
736 | list_for_each_entry_safe(rg, trg, head, link) { | |
737 | list_del(&rg->link); | |
738 | kfree(rg); | |
739 | } | |
740 | ||
741 | VM_BUG_ON(resv_map->adds_in_progress); | |
742 | ||
84afd99b AW |
743 | kfree(resv_map); |
744 | } | |
745 | ||
4e35f483 JK |
746 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
747 | { | |
748 | return inode->i_mapping->private_data; | |
749 | } | |
750 | ||
84afd99b | 751 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 752 | { |
81d1b09c | 753 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
754 | if (vma->vm_flags & VM_MAYSHARE) { |
755 | struct address_space *mapping = vma->vm_file->f_mapping; | |
756 | struct inode *inode = mapping->host; | |
757 | ||
758 | return inode_resv_map(inode); | |
759 | ||
760 | } else { | |
84afd99b AW |
761 | return (struct resv_map *)(get_vma_private_data(vma) & |
762 | ~HPAGE_RESV_MASK); | |
4e35f483 | 763 | } |
a1e78772 MG |
764 | } |
765 | ||
84afd99b | 766 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 767 | { |
81d1b09c SL |
768 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
769 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 770 | |
84afd99b AW |
771 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
772 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
773 | } |
774 | ||
775 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
776 | { | |
81d1b09c SL |
777 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
778 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
779 | |
780 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
781 | } |
782 | ||
783 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
784 | { | |
81d1b09c | 785 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
786 | |
787 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
788 | } |
789 | ||
04f2cbe3 | 790 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
791 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
792 | { | |
81d1b09c | 793 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 794 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
795 | vma->vm_private_data = (void *)0; |
796 | } | |
797 | ||
798 | /* Returns true if the VMA has associated reserve pages */ | |
559ec2f8 | 799 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 800 | { |
af0ed73e JK |
801 | if (vma->vm_flags & VM_NORESERVE) { |
802 | /* | |
803 | * This address is already reserved by other process(chg == 0), | |
804 | * so, we should decrement reserved count. Without decrementing, | |
805 | * reserve count remains after releasing inode, because this | |
806 | * allocated page will go into page cache and is regarded as | |
807 | * coming from reserved pool in releasing step. Currently, we | |
808 | * don't have any other solution to deal with this situation | |
809 | * properly, so add work-around here. | |
810 | */ | |
811 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 812 | return true; |
af0ed73e | 813 | else |
559ec2f8 | 814 | return false; |
af0ed73e | 815 | } |
a63884e9 JK |
816 | |
817 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
818 | if (vma->vm_flags & VM_MAYSHARE) { |
819 | /* | |
820 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
821 | * be a region map for all pages. The only situation where | |
822 | * there is no region map is if a hole was punched via | |
823 | * fallocate. In this case, there really are no reverves to | |
824 | * use. This situation is indicated if chg != 0. | |
825 | */ | |
826 | if (chg) | |
827 | return false; | |
828 | else | |
829 | return true; | |
830 | } | |
a63884e9 JK |
831 | |
832 | /* | |
833 | * Only the process that called mmap() has reserves for | |
834 | * private mappings. | |
835 | */ | |
67961f9d MK |
836 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
837 | /* | |
838 | * Like the shared case above, a hole punch or truncate | |
839 | * could have been performed on the private mapping. | |
840 | * Examine the value of chg to determine if reserves | |
841 | * actually exist or were previously consumed. | |
842 | * Very Subtle - The value of chg comes from a previous | |
843 | * call to vma_needs_reserves(). The reserve map for | |
844 | * private mappings has different (opposite) semantics | |
845 | * than that of shared mappings. vma_needs_reserves() | |
846 | * has already taken this difference in semantics into | |
847 | * account. Therefore, the meaning of chg is the same | |
848 | * as in the shared case above. Code could easily be | |
849 | * combined, but keeping it separate draws attention to | |
850 | * subtle differences. | |
851 | */ | |
852 | if (chg) | |
853 | return false; | |
854 | else | |
855 | return true; | |
856 | } | |
a63884e9 | 857 | |
559ec2f8 | 858 | return false; |
a1e78772 MG |
859 | } |
860 | ||
a5516438 | 861 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
862 | { |
863 | int nid = page_to_nid(page); | |
0edaecfa | 864 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
865 | h->free_huge_pages++; |
866 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
867 | } |
868 | ||
94310cbc | 869 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
bf50bab2 NH |
870 | { |
871 | struct page *page; | |
872 | ||
c8721bbb | 873 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
243abd5b | 874 | if (!PageHWPoison(page)) |
c8721bbb NH |
875 | break; |
876 | /* | |
877 | * if 'non-isolated free hugepage' not found on the list, | |
878 | * the allocation fails. | |
879 | */ | |
880 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 881 | return NULL; |
0edaecfa | 882 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 883 | set_page_refcounted(page); |
bf50bab2 NH |
884 | h->free_huge_pages--; |
885 | h->free_huge_pages_node[nid]--; | |
886 | return page; | |
887 | } | |
888 | ||
94310cbc AK |
889 | static struct page *dequeue_huge_page_node(struct hstate *h, int nid) |
890 | { | |
891 | struct page *page; | |
892 | int node; | |
893 | ||
894 | if (nid != NUMA_NO_NODE) | |
895 | return dequeue_huge_page_node_exact(h, nid); | |
896 | ||
897 | for_each_online_node(node) { | |
898 | page = dequeue_huge_page_node_exact(h, node); | |
899 | if (page) | |
900 | return page; | |
901 | } | |
902 | return NULL; | |
903 | } | |
904 | ||
86cdb465 NH |
905 | /* Movability of hugepages depends on migration support. */ |
906 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
907 | { | |
100873d7 | 908 | if (hugepages_treat_as_movable || hugepage_migration_supported(h)) |
86cdb465 NH |
909 | return GFP_HIGHUSER_MOVABLE; |
910 | else | |
911 | return GFP_HIGHUSER; | |
912 | } | |
913 | ||
a5516438 AK |
914 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
915 | struct vm_area_struct *vma, | |
af0ed73e JK |
916 | unsigned long address, int avoid_reserve, |
917 | long chg) | |
1da177e4 | 918 | { |
b1c12cbc | 919 | struct page *page = NULL; |
480eccf9 | 920 | struct mempolicy *mpol; |
19770b32 | 921 | nodemask_t *nodemask; |
04ec6264 VB |
922 | gfp_t gfp_mask; |
923 | int nid; | |
c0ff7453 | 924 | struct zonelist *zonelist; |
dd1a239f MG |
925 | struct zone *zone; |
926 | struct zoneref *z; | |
cc9a6c87 | 927 | unsigned int cpuset_mems_cookie; |
1da177e4 | 928 | |
a1e78772 MG |
929 | /* |
930 | * A child process with MAP_PRIVATE mappings created by their parent | |
931 | * have no page reserves. This check ensures that reservations are | |
932 | * not "stolen". The child may still get SIGKILLed | |
933 | */ | |
af0ed73e | 934 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 935 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 936 | goto err; |
a1e78772 | 937 | |
04f2cbe3 | 938 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 939 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 940 | goto err; |
04f2cbe3 | 941 | |
9966c4bb | 942 | retry_cpuset: |
d26914d1 | 943 | cpuset_mems_cookie = read_mems_allowed_begin(); |
04ec6264 VB |
944 | gfp_mask = htlb_alloc_mask(h); |
945 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
946 | zonelist = node_zonelist(nid, gfp_mask); | |
9966c4bb | 947 | |
19770b32 MG |
948 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
949 | MAX_NR_ZONES - 1, nodemask) { | |
04ec6264 | 950 | if (cpuset_zone_allowed(zone, gfp_mask)) { |
bf50bab2 NH |
951 | page = dequeue_huge_page_node(h, zone_to_nid(zone)); |
952 | if (page) { | |
af0ed73e JK |
953 | if (avoid_reserve) |
954 | break; | |
955 | if (!vma_has_reserves(vma, chg)) | |
956 | break; | |
957 | ||
07443a85 | 958 | SetPagePrivate(page); |
af0ed73e | 959 | h->resv_huge_pages--; |
bf50bab2 NH |
960 | break; |
961 | } | |
3abf7afd | 962 | } |
1da177e4 | 963 | } |
cc9a6c87 | 964 | |
52cd3b07 | 965 | mpol_cond_put(mpol); |
d26914d1 | 966 | if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) |
cc9a6c87 | 967 | goto retry_cpuset; |
1da177e4 | 968 | return page; |
cc9a6c87 MG |
969 | |
970 | err: | |
cc9a6c87 | 971 | return NULL; |
1da177e4 LT |
972 | } |
973 | ||
1cac6f2c LC |
974 | /* |
975 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
976 | * We may have allocated or freed a huge page based on a different | |
977 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
978 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
979 | * node for alloc or free. | |
980 | */ | |
981 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
982 | { | |
0edaf86c | 983 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
984 | VM_BUG_ON(nid >= MAX_NUMNODES); |
985 | ||
986 | return nid; | |
987 | } | |
988 | ||
989 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
990 | { | |
991 | if (!node_isset(nid, *nodes_allowed)) | |
992 | nid = next_node_allowed(nid, nodes_allowed); | |
993 | return nid; | |
994 | } | |
995 | ||
996 | /* | |
997 | * returns the previously saved node ["this node"] from which to | |
998 | * allocate a persistent huge page for the pool and advance the | |
999 | * next node from which to allocate, handling wrap at end of node | |
1000 | * mask. | |
1001 | */ | |
1002 | static int hstate_next_node_to_alloc(struct hstate *h, | |
1003 | nodemask_t *nodes_allowed) | |
1004 | { | |
1005 | int nid; | |
1006 | ||
1007 | VM_BUG_ON(!nodes_allowed); | |
1008 | ||
1009 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1010 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1011 | ||
1012 | return nid; | |
1013 | } | |
1014 | ||
1015 | /* | |
1016 | * helper for free_pool_huge_page() - return the previously saved | |
1017 | * node ["this node"] from which to free a huge page. Advance the | |
1018 | * next node id whether or not we find a free huge page to free so | |
1019 | * that the next attempt to free addresses the next node. | |
1020 | */ | |
1021 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1022 | { | |
1023 | int nid; | |
1024 | ||
1025 | VM_BUG_ON(!nodes_allowed); | |
1026 | ||
1027 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1028 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1029 | ||
1030 | return nid; | |
1031 | } | |
1032 | ||
1033 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1034 | for (nr_nodes = nodes_weight(*mask); \ | |
1035 | nr_nodes > 0 && \ | |
1036 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1037 | nr_nodes--) | |
1038 | ||
1039 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1040 | for (nr_nodes = nodes_weight(*mask); \ | |
1041 | nr_nodes > 0 && \ | |
1042 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1043 | nr_nodes--) | |
1044 | ||
e1073d1e | 1045 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE |
944d9fec | 1046 | static void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1047 | unsigned int order) |
944d9fec LC |
1048 | { |
1049 | int i; | |
1050 | int nr_pages = 1 << order; | |
1051 | struct page *p = page + 1; | |
1052 | ||
c8cc708a | 1053 | atomic_set(compound_mapcount_ptr(page), 0); |
944d9fec | 1054 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
1d798ca3 | 1055 | clear_compound_head(p); |
944d9fec | 1056 | set_page_refcounted(p); |
944d9fec LC |
1057 | } |
1058 | ||
1059 | set_compound_order(page, 0); | |
1060 | __ClearPageHead(page); | |
1061 | } | |
1062 | ||
d00181b9 | 1063 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec LC |
1064 | { |
1065 | free_contig_range(page_to_pfn(page), 1 << order); | |
1066 | } | |
1067 | ||
1068 | static int __alloc_gigantic_page(unsigned long start_pfn, | |
1069 | unsigned long nr_pages) | |
1070 | { | |
1071 | unsigned long end_pfn = start_pfn + nr_pages; | |
ca96b625 LS |
1072 | return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE, |
1073 | GFP_KERNEL); | |
944d9fec LC |
1074 | } |
1075 | ||
f44b2dda JK |
1076 | static bool pfn_range_valid_gigantic(struct zone *z, |
1077 | unsigned long start_pfn, unsigned long nr_pages) | |
944d9fec LC |
1078 | { |
1079 | unsigned long i, end_pfn = start_pfn + nr_pages; | |
1080 | struct page *page; | |
1081 | ||
1082 | for (i = start_pfn; i < end_pfn; i++) { | |
1083 | if (!pfn_valid(i)) | |
1084 | return false; | |
1085 | ||
1086 | page = pfn_to_page(i); | |
1087 | ||
f44b2dda JK |
1088 | if (page_zone(page) != z) |
1089 | return false; | |
1090 | ||
944d9fec LC |
1091 | if (PageReserved(page)) |
1092 | return false; | |
1093 | ||
1094 | if (page_count(page) > 0) | |
1095 | return false; | |
1096 | ||
1097 | if (PageHuge(page)) | |
1098 | return false; | |
1099 | } | |
1100 | ||
1101 | return true; | |
1102 | } | |
1103 | ||
1104 | static bool zone_spans_last_pfn(const struct zone *zone, | |
1105 | unsigned long start_pfn, unsigned long nr_pages) | |
1106 | { | |
1107 | unsigned long last_pfn = start_pfn + nr_pages - 1; | |
1108 | return zone_spans_pfn(zone, last_pfn); | |
1109 | } | |
1110 | ||
d00181b9 | 1111 | static struct page *alloc_gigantic_page(int nid, unsigned int order) |
944d9fec LC |
1112 | { |
1113 | unsigned long nr_pages = 1 << order; | |
1114 | unsigned long ret, pfn, flags; | |
1115 | struct zone *z; | |
1116 | ||
1117 | z = NODE_DATA(nid)->node_zones; | |
1118 | for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) { | |
1119 | spin_lock_irqsave(&z->lock, flags); | |
1120 | ||
1121 | pfn = ALIGN(z->zone_start_pfn, nr_pages); | |
1122 | while (zone_spans_last_pfn(z, pfn, nr_pages)) { | |
f44b2dda | 1123 | if (pfn_range_valid_gigantic(z, pfn, nr_pages)) { |
944d9fec LC |
1124 | /* |
1125 | * We release the zone lock here because | |
1126 | * alloc_contig_range() will also lock the zone | |
1127 | * at some point. If there's an allocation | |
1128 | * spinning on this lock, it may win the race | |
1129 | * and cause alloc_contig_range() to fail... | |
1130 | */ | |
1131 | spin_unlock_irqrestore(&z->lock, flags); | |
1132 | ret = __alloc_gigantic_page(pfn, nr_pages); | |
1133 | if (!ret) | |
1134 | return pfn_to_page(pfn); | |
1135 | spin_lock_irqsave(&z->lock, flags); | |
1136 | } | |
1137 | pfn += nr_pages; | |
1138 | } | |
1139 | ||
1140 | spin_unlock_irqrestore(&z->lock, flags); | |
1141 | } | |
1142 | ||
1143 | return NULL; | |
1144 | } | |
1145 | ||
1146 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
d00181b9 | 1147 | static void prep_compound_gigantic_page(struct page *page, unsigned int order); |
944d9fec LC |
1148 | |
1149 | static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) | |
1150 | { | |
1151 | struct page *page; | |
1152 | ||
1153 | page = alloc_gigantic_page(nid, huge_page_order(h)); | |
1154 | if (page) { | |
1155 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
1156 | prep_new_huge_page(h, page, nid); | |
1157 | } | |
1158 | ||
1159 | return page; | |
1160 | } | |
1161 | ||
1162 | static int alloc_fresh_gigantic_page(struct hstate *h, | |
1163 | nodemask_t *nodes_allowed) | |
1164 | { | |
1165 | struct page *page = NULL; | |
1166 | int nr_nodes, node; | |
1167 | ||
1168 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1169 | page = alloc_fresh_gigantic_page_node(h, node); | |
1170 | if (page) | |
1171 | return 1; | |
1172 | } | |
1173 | ||
1174 | return 0; | |
1175 | } | |
1176 | ||
e1073d1e | 1177 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
944d9fec | 1178 | static inline bool gigantic_page_supported(void) { return false; } |
d00181b9 | 1179 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1180 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1181 | unsigned int order) { } |
944d9fec LC |
1182 | static inline int alloc_fresh_gigantic_page(struct hstate *h, |
1183 | nodemask_t *nodes_allowed) { return 0; } | |
1184 | #endif | |
1185 | ||
a5516438 | 1186 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1187 | { |
1188 | int i; | |
a5516438 | 1189 | |
944d9fec LC |
1190 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
1191 | return; | |
18229df5 | 1192 | |
a5516438 AK |
1193 | h->nr_huge_pages--; |
1194 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
1195 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
1196 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
1197 | 1 << PG_referenced | 1 << PG_dirty | | |
a7407a27 LC |
1198 | 1 << PG_active | 1 << PG_private | |
1199 | 1 << PG_writeback); | |
6af2acb6 | 1200 | } |
309381fe | 1201 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
f1e61557 | 1202 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); |
6af2acb6 | 1203 | set_page_refcounted(page); |
944d9fec LC |
1204 | if (hstate_is_gigantic(h)) { |
1205 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
1206 | free_gigantic_page(page, huge_page_order(h)); | |
1207 | } else { | |
944d9fec LC |
1208 | __free_pages(page, huge_page_order(h)); |
1209 | } | |
6af2acb6 AL |
1210 | } |
1211 | ||
e5ff2159 AK |
1212 | struct hstate *size_to_hstate(unsigned long size) |
1213 | { | |
1214 | struct hstate *h; | |
1215 | ||
1216 | for_each_hstate(h) { | |
1217 | if (huge_page_size(h) == size) | |
1218 | return h; | |
1219 | } | |
1220 | return NULL; | |
1221 | } | |
1222 | ||
bcc54222 NH |
1223 | /* |
1224 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
1225 | * to hstate->hugepage_activelist.) | |
1226 | * | |
1227 | * This function can be called for tail pages, but never returns true for them. | |
1228 | */ | |
1229 | bool page_huge_active(struct page *page) | |
1230 | { | |
1231 | VM_BUG_ON_PAGE(!PageHuge(page), page); | |
1232 | return PageHead(page) && PagePrivate(&page[1]); | |
1233 | } | |
1234 | ||
1235 | /* never called for tail page */ | |
1236 | static void set_page_huge_active(struct page *page) | |
1237 | { | |
1238 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1239 | SetPagePrivate(&page[1]); | |
1240 | } | |
1241 | ||
1242 | static void clear_page_huge_active(struct page *page) | |
1243 | { | |
1244 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1245 | ClearPagePrivate(&page[1]); | |
1246 | } | |
1247 | ||
8f1d26d0 | 1248 | void free_huge_page(struct page *page) |
27a85ef1 | 1249 | { |
a5516438 AK |
1250 | /* |
1251 | * Can't pass hstate in here because it is called from the | |
1252 | * compound page destructor. | |
1253 | */ | |
e5ff2159 | 1254 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1255 | int nid = page_to_nid(page); |
90481622 DG |
1256 | struct hugepage_subpool *spool = |
1257 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 1258 | bool restore_reserve; |
27a85ef1 | 1259 | |
e5df70ab | 1260 | set_page_private(page, 0); |
23be7468 | 1261 | page->mapping = NULL; |
b4330afb MK |
1262 | VM_BUG_ON_PAGE(page_count(page), page); |
1263 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
07443a85 | 1264 | restore_reserve = PagePrivate(page); |
16c794b4 | 1265 | ClearPagePrivate(page); |
27a85ef1 | 1266 | |
1c5ecae3 MK |
1267 | /* |
1268 | * A return code of zero implies that the subpool will be under its | |
1269 | * minimum size if the reservation is not restored after page is free. | |
1270 | * Therefore, force restore_reserve operation. | |
1271 | */ | |
1272 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1273 | restore_reserve = true; | |
1274 | ||
27a85ef1 | 1275 | spin_lock(&hugetlb_lock); |
bcc54222 | 1276 | clear_page_huge_active(page); |
6d76dcf4 AK |
1277 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1278 | pages_per_huge_page(h), page); | |
07443a85 JK |
1279 | if (restore_reserve) |
1280 | h->resv_huge_pages++; | |
1281 | ||
944d9fec | 1282 | if (h->surplus_huge_pages_node[nid]) { |
0edaecfa AK |
1283 | /* remove the page from active list */ |
1284 | list_del(&page->lru); | |
a5516438 AK |
1285 | update_and_free_page(h, page); |
1286 | h->surplus_huge_pages--; | |
1287 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 1288 | } else { |
5d3a551c | 1289 | arch_clear_hugepage_flags(page); |
a5516438 | 1290 | enqueue_huge_page(h, page); |
7893d1d5 | 1291 | } |
27a85ef1 DG |
1292 | spin_unlock(&hugetlb_lock); |
1293 | } | |
1294 | ||
a5516438 | 1295 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1296 | { |
0edaecfa | 1297 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1298 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
b7ba30c6 | 1299 | spin_lock(&hugetlb_lock); |
9dd540e2 | 1300 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
1301 | h->nr_huge_pages++; |
1302 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 AK |
1303 | spin_unlock(&hugetlb_lock); |
1304 | put_page(page); /* free it into the hugepage allocator */ | |
1305 | } | |
1306 | ||
d00181b9 | 1307 | static void prep_compound_gigantic_page(struct page *page, unsigned int order) |
20a0307c WF |
1308 | { |
1309 | int i; | |
1310 | int nr_pages = 1 << order; | |
1311 | struct page *p = page + 1; | |
1312 | ||
1313 | /* we rely on prep_new_huge_page to set the destructor */ | |
1314 | set_compound_order(page, order); | |
ef5a22be | 1315 | __ClearPageReserved(page); |
de09d31d | 1316 | __SetPageHead(page); |
20a0307c | 1317 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1318 | /* |
1319 | * For gigantic hugepages allocated through bootmem at | |
1320 | * boot, it's safer to be consistent with the not-gigantic | |
1321 | * hugepages and clear the PG_reserved bit from all tail pages | |
1322 | * too. Otherwse drivers using get_user_pages() to access tail | |
1323 | * pages may get the reference counting wrong if they see | |
1324 | * PG_reserved set on a tail page (despite the head page not | |
1325 | * having PG_reserved set). Enforcing this consistency between | |
1326 | * head and tail pages allows drivers to optimize away a check | |
1327 | * on the head page when they need know if put_page() is needed | |
1328 | * after get_user_pages(). | |
1329 | */ | |
1330 | __ClearPageReserved(p); | |
58a84aa9 | 1331 | set_page_count(p, 0); |
1d798ca3 | 1332 | set_compound_head(p, page); |
20a0307c | 1333 | } |
b4330afb | 1334 | atomic_set(compound_mapcount_ptr(page), -1); |
20a0307c WF |
1335 | } |
1336 | ||
7795912c AM |
1337 | /* |
1338 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1339 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1340 | * details. | |
1341 | */ | |
20a0307c WF |
1342 | int PageHuge(struct page *page) |
1343 | { | |
20a0307c WF |
1344 | if (!PageCompound(page)) |
1345 | return 0; | |
1346 | ||
1347 | page = compound_head(page); | |
f1e61557 | 1348 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1349 | } |
43131e14 NH |
1350 | EXPORT_SYMBOL_GPL(PageHuge); |
1351 | ||
27c73ae7 AA |
1352 | /* |
1353 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1354 | * normal or transparent huge pages. | |
1355 | */ | |
1356 | int PageHeadHuge(struct page *page_head) | |
1357 | { | |
27c73ae7 AA |
1358 | if (!PageHead(page_head)) |
1359 | return 0; | |
1360 | ||
758f66a2 | 1361 | return get_compound_page_dtor(page_head) == free_huge_page; |
27c73ae7 | 1362 | } |
27c73ae7 | 1363 | |
13d60f4b ZY |
1364 | pgoff_t __basepage_index(struct page *page) |
1365 | { | |
1366 | struct page *page_head = compound_head(page); | |
1367 | pgoff_t index = page_index(page_head); | |
1368 | unsigned long compound_idx; | |
1369 | ||
1370 | if (!PageHuge(page_head)) | |
1371 | return page_index(page); | |
1372 | ||
1373 | if (compound_order(page_head) >= MAX_ORDER) | |
1374 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1375 | else | |
1376 | compound_idx = page - page_head; | |
1377 | ||
1378 | return (index << compound_order(page_head)) + compound_idx; | |
1379 | } | |
1380 | ||
a5516438 | 1381 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
1da177e4 | 1382 | { |
1da177e4 | 1383 | struct page *page; |
f96efd58 | 1384 | |
96db800f | 1385 | page = __alloc_pages_node(nid, |
86cdb465 | 1386 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
551883ae | 1387 | __GFP_REPEAT|__GFP_NOWARN, |
a5516438 | 1388 | huge_page_order(h)); |
1da177e4 | 1389 | if (page) { |
a5516438 | 1390 | prep_new_huge_page(h, page, nid); |
1da177e4 | 1391 | } |
63b4613c NA |
1392 | |
1393 | return page; | |
1394 | } | |
1395 | ||
b2261026 JK |
1396 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
1397 | { | |
1398 | struct page *page; | |
1399 | int nr_nodes, node; | |
1400 | int ret = 0; | |
1401 | ||
1402 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1403 | page = alloc_fresh_huge_page_node(h, node); | |
1404 | if (page) { | |
1405 | ret = 1; | |
1406 | break; | |
1407 | } | |
1408 | } | |
1409 | ||
1410 | if (ret) | |
1411 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
1412 | else | |
1413 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
1414 | ||
1415 | return ret; | |
1416 | } | |
1417 | ||
e8c5c824 LS |
1418 | /* |
1419 | * Free huge page from pool from next node to free. | |
1420 | * Attempt to keep persistent huge pages more or less | |
1421 | * balanced over allowed nodes. | |
1422 | * Called with hugetlb_lock locked. | |
1423 | */ | |
6ae11b27 LS |
1424 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1425 | bool acct_surplus) | |
e8c5c824 | 1426 | { |
b2261026 | 1427 | int nr_nodes, node; |
e8c5c824 LS |
1428 | int ret = 0; |
1429 | ||
b2261026 | 1430 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1431 | /* |
1432 | * If we're returning unused surplus pages, only examine | |
1433 | * nodes with surplus pages. | |
1434 | */ | |
b2261026 JK |
1435 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1436 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1437 | struct page *page = |
b2261026 | 1438 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1439 | struct page, lru); |
1440 | list_del(&page->lru); | |
1441 | h->free_huge_pages--; | |
b2261026 | 1442 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1443 | if (acct_surplus) { |
1444 | h->surplus_huge_pages--; | |
b2261026 | 1445 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1446 | } |
e8c5c824 LS |
1447 | update_and_free_page(h, page); |
1448 | ret = 1; | |
9a76db09 | 1449 | break; |
e8c5c824 | 1450 | } |
b2261026 | 1451 | } |
e8c5c824 LS |
1452 | |
1453 | return ret; | |
1454 | } | |
1455 | ||
c8721bbb NH |
1456 | /* |
1457 | * Dissolve a given free hugepage into free buddy pages. This function does | |
082d5b6b GS |
1458 | * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the |
1459 | * number of free hugepages would be reduced below the number of reserved | |
1460 | * hugepages. | |
c8721bbb | 1461 | */ |
082d5b6b | 1462 | static int dissolve_free_huge_page(struct page *page) |
c8721bbb | 1463 | { |
082d5b6b GS |
1464 | int rc = 0; |
1465 | ||
c8721bbb NH |
1466 | spin_lock(&hugetlb_lock); |
1467 | if (PageHuge(page) && !page_count(page)) { | |
2247bb33 GS |
1468 | struct page *head = compound_head(page); |
1469 | struct hstate *h = page_hstate(head); | |
1470 | int nid = page_to_nid(head); | |
082d5b6b GS |
1471 | if (h->free_huge_pages - h->resv_huge_pages == 0) { |
1472 | rc = -EBUSY; | |
1473 | goto out; | |
1474 | } | |
2247bb33 | 1475 | list_del(&head->lru); |
c8721bbb NH |
1476 | h->free_huge_pages--; |
1477 | h->free_huge_pages_node[nid]--; | |
c1470b33 | 1478 | h->max_huge_pages--; |
2247bb33 | 1479 | update_and_free_page(h, head); |
c8721bbb | 1480 | } |
082d5b6b | 1481 | out: |
c8721bbb | 1482 | spin_unlock(&hugetlb_lock); |
082d5b6b | 1483 | return rc; |
c8721bbb NH |
1484 | } |
1485 | ||
1486 | /* | |
1487 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1488 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
1489 | * Note that this will dissolve a free gigantic hugepage completely, if any |
1490 | * part of it lies within the given range. | |
082d5b6b GS |
1491 | * Also note that if dissolve_free_huge_page() returns with an error, all |
1492 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 1493 | */ |
082d5b6b | 1494 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 1495 | { |
c8721bbb | 1496 | unsigned long pfn; |
eb03aa00 | 1497 | struct page *page; |
082d5b6b | 1498 | int rc = 0; |
c8721bbb | 1499 | |
d0177639 | 1500 | if (!hugepages_supported()) |
082d5b6b | 1501 | return rc; |
d0177639 | 1502 | |
eb03aa00 GS |
1503 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { |
1504 | page = pfn_to_page(pfn); | |
1505 | if (PageHuge(page) && !page_count(page)) { | |
1506 | rc = dissolve_free_huge_page(page); | |
1507 | if (rc) | |
1508 | break; | |
1509 | } | |
1510 | } | |
082d5b6b GS |
1511 | |
1512 | return rc; | |
c8721bbb NH |
1513 | } |
1514 | ||
099730d6 DH |
1515 | /* |
1516 | * There are 3 ways this can get called: | |
1517 | * 1. With vma+addr: we use the VMA's memory policy | |
1518 | * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge | |
1519 | * page from any node, and let the buddy allocator itself figure | |
1520 | * it out. | |
1521 | * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page | |
1522 | * strictly from 'nid' | |
1523 | */ | |
1524 | static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h, | |
1525 | struct vm_area_struct *vma, unsigned long addr, int nid) | |
1526 | { | |
1527 | int order = huge_page_order(h); | |
1528 | gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN; | |
1529 | unsigned int cpuset_mems_cookie; | |
1530 | ||
1531 | /* | |
1532 | * We need a VMA to get a memory policy. If we do not | |
e0ec90ee DH |
1533 | * have one, we use the 'nid' argument. |
1534 | * | |
1535 | * The mempolicy stuff below has some non-inlined bits | |
1536 | * and calls ->vm_ops. That makes it hard to optimize at | |
1537 | * compile-time, even when NUMA is off and it does | |
1538 | * nothing. This helps the compiler optimize it out. | |
099730d6 | 1539 | */ |
e0ec90ee | 1540 | if (!IS_ENABLED(CONFIG_NUMA) || !vma) { |
099730d6 DH |
1541 | /* |
1542 | * If a specific node is requested, make sure to | |
1543 | * get memory from there, but only when a node | |
1544 | * is explicitly specified. | |
1545 | */ | |
1546 | if (nid != NUMA_NO_NODE) | |
1547 | gfp |= __GFP_THISNODE; | |
1548 | /* | |
1549 | * Make sure to call something that can handle | |
1550 | * nid=NUMA_NO_NODE | |
1551 | */ | |
1552 | return alloc_pages_node(nid, gfp, order); | |
1553 | } | |
1554 | ||
1555 | /* | |
1556 | * OK, so we have a VMA. Fetch the mempolicy and try to | |
e0ec90ee DH |
1557 | * allocate a huge page with it. We will only reach this |
1558 | * when CONFIG_NUMA=y. | |
099730d6 DH |
1559 | */ |
1560 | do { | |
1561 | struct page *page; | |
1562 | struct mempolicy *mpol; | |
04ec6264 | 1563 | int nid; |
099730d6 DH |
1564 | nodemask_t *nodemask; |
1565 | ||
1566 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
04ec6264 | 1567 | nid = huge_node(vma, addr, gfp, &mpol, &nodemask); |
099730d6 | 1568 | mpol_cond_put(mpol); |
04ec6264 | 1569 | page = __alloc_pages_nodemask(gfp, order, nid, nodemask); |
099730d6 DH |
1570 | if (page) |
1571 | return page; | |
1572 | } while (read_mems_allowed_retry(cpuset_mems_cookie)); | |
1573 | ||
1574 | return NULL; | |
1575 | } | |
1576 | ||
1577 | /* | |
1578 | * There are two ways to allocate a huge page: | |
1579 | * 1. When you have a VMA and an address (like a fault) | |
1580 | * 2. When you have no VMA (like when setting /proc/.../nr_hugepages) | |
1581 | * | |
1582 | * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in | |
1583 | * this case which signifies that the allocation should be done with | |
1584 | * respect for the VMA's memory policy. | |
1585 | * | |
1586 | * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This | |
1587 | * implies that memory policies will not be taken in to account. | |
1588 | */ | |
1589 | static struct page *__alloc_buddy_huge_page(struct hstate *h, | |
1590 | struct vm_area_struct *vma, unsigned long addr, int nid) | |
7893d1d5 AL |
1591 | { |
1592 | struct page *page; | |
bf50bab2 | 1593 | unsigned int r_nid; |
7893d1d5 | 1594 | |
bae7f4ae | 1595 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1596 | return NULL; |
1597 | ||
099730d6 DH |
1598 | /* |
1599 | * Make sure that anyone specifying 'nid' is not also specifying a VMA. | |
1600 | * This makes sure the caller is picking _one_ of the modes with which | |
1601 | * we can call this function, not both. | |
1602 | */ | |
1603 | if (vma || (addr != -1)) { | |
e0ec90ee DH |
1604 | VM_WARN_ON_ONCE(addr == -1); |
1605 | VM_WARN_ON_ONCE(nid != NUMA_NO_NODE); | |
099730d6 | 1606 | } |
d1c3fb1f NA |
1607 | /* |
1608 | * Assume we will successfully allocate the surplus page to | |
1609 | * prevent racing processes from causing the surplus to exceed | |
1610 | * overcommit | |
1611 | * | |
1612 | * This however introduces a different race, where a process B | |
1613 | * tries to grow the static hugepage pool while alloc_pages() is | |
1614 | * called by process A. B will only examine the per-node | |
1615 | * counters in determining if surplus huge pages can be | |
1616 | * converted to normal huge pages in adjust_pool_surplus(). A | |
1617 | * won't be able to increment the per-node counter, until the | |
1618 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
1619 | * no more huge pages can be converted from surplus to normal | |
1620 | * state (and doesn't try to convert again). Thus, we have a | |
1621 | * case where a surplus huge page exists, the pool is grown, and | |
1622 | * the surplus huge page still exists after, even though it | |
1623 | * should just have been converted to a normal huge page. This | |
1624 | * does not leak memory, though, as the hugepage will be freed | |
1625 | * once it is out of use. It also does not allow the counters to | |
1626 | * go out of whack in adjust_pool_surplus() as we don't modify | |
1627 | * the node values until we've gotten the hugepage and only the | |
1628 | * per-node value is checked there. | |
1629 | */ | |
1630 | spin_lock(&hugetlb_lock); | |
a5516438 | 1631 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
1632 | spin_unlock(&hugetlb_lock); |
1633 | return NULL; | |
1634 | } else { | |
a5516438 AK |
1635 | h->nr_huge_pages++; |
1636 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
1637 | } |
1638 | spin_unlock(&hugetlb_lock); | |
1639 | ||
099730d6 | 1640 | page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid); |
d1c3fb1f NA |
1641 | |
1642 | spin_lock(&hugetlb_lock); | |
7893d1d5 | 1643 | if (page) { |
0edaecfa | 1644 | INIT_LIST_HEAD(&page->lru); |
bf50bab2 | 1645 | r_nid = page_to_nid(page); |
f1e61557 | 1646 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
9dd540e2 | 1647 | set_hugetlb_cgroup(page, NULL); |
d1c3fb1f NA |
1648 | /* |
1649 | * We incremented the global counters already | |
1650 | */ | |
bf50bab2 NH |
1651 | h->nr_huge_pages_node[r_nid]++; |
1652 | h->surplus_huge_pages_node[r_nid]++; | |
3b116300 | 1653 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 1654 | } else { |
a5516438 AK |
1655 | h->nr_huge_pages--; |
1656 | h->surplus_huge_pages--; | |
3b116300 | 1657 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 1658 | } |
d1c3fb1f | 1659 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1660 | |
1661 | return page; | |
1662 | } | |
1663 | ||
099730d6 DH |
1664 | /* |
1665 | * Allocate a huge page from 'nid'. Note, 'nid' may be | |
1666 | * NUMA_NO_NODE, which means that it may be allocated | |
1667 | * anywhere. | |
1668 | */ | |
e0ec90ee | 1669 | static |
099730d6 DH |
1670 | struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid) |
1671 | { | |
1672 | unsigned long addr = -1; | |
1673 | ||
1674 | return __alloc_buddy_huge_page(h, NULL, addr, nid); | |
1675 | } | |
1676 | ||
1677 | /* | |
1678 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
1679 | */ | |
e0ec90ee | 1680 | static |
099730d6 DH |
1681 | struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h, |
1682 | struct vm_area_struct *vma, unsigned long addr) | |
1683 | { | |
1684 | return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE); | |
1685 | } | |
1686 | ||
bf50bab2 NH |
1687 | /* |
1688 | * This allocation function is useful in the context where vma is irrelevant. | |
1689 | * E.g. soft-offlining uses this function because it only cares physical | |
1690 | * address of error page. | |
1691 | */ | |
1692 | struct page *alloc_huge_page_node(struct hstate *h, int nid) | |
1693 | { | |
4ef91848 | 1694 | struct page *page = NULL; |
bf50bab2 NH |
1695 | |
1696 | spin_lock(&hugetlb_lock); | |
4ef91848 JK |
1697 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
1698 | page = dequeue_huge_page_node(h, nid); | |
bf50bab2 NH |
1699 | spin_unlock(&hugetlb_lock); |
1700 | ||
94ae8ba7 | 1701 | if (!page) |
099730d6 | 1702 | page = __alloc_buddy_huge_page_no_mpol(h, nid); |
bf50bab2 NH |
1703 | |
1704 | return page; | |
1705 | } | |
1706 | ||
e4e574b7 | 1707 | /* |
25985edc | 1708 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1709 | * of size 'delta'. |
1710 | */ | |
a5516438 | 1711 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1712 | { |
1713 | struct list_head surplus_list; | |
1714 | struct page *page, *tmp; | |
1715 | int ret, i; | |
1716 | int needed, allocated; | |
28073b02 | 1717 | bool alloc_ok = true; |
e4e574b7 | 1718 | |
a5516438 | 1719 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1720 | if (needed <= 0) { |
a5516438 | 1721 | h->resv_huge_pages += delta; |
e4e574b7 | 1722 | return 0; |
ac09b3a1 | 1723 | } |
e4e574b7 AL |
1724 | |
1725 | allocated = 0; | |
1726 | INIT_LIST_HEAD(&surplus_list); | |
1727 | ||
1728 | ret = -ENOMEM; | |
1729 | retry: | |
1730 | spin_unlock(&hugetlb_lock); | |
1731 | for (i = 0; i < needed; i++) { | |
099730d6 | 1732 | page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE); |
28073b02 HD |
1733 | if (!page) { |
1734 | alloc_ok = false; | |
1735 | break; | |
1736 | } | |
e4e574b7 AL |
1737 | list_add(&page->lru, &surplus_list); |
1738 | } | |
28073b02 | 1739 | allocated += i; |
e4e574b7 AL |
1740 | |
1741 | /* | |
1742 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1743 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1744 | */ | |
1745 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1746 | needed = (h->resv_huge_pages + delta) - |
1747 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1748 | if (needed > 0) { |
1749 | if (alloc_ok) | |
1750 | goto retry; | |
1751 | /* | |
1752 | * We were not able to allocate enough pages to | |
1753 | * satisfy the entire reservation so we free what | |
1754 | * we've allocated so far. | |
1755 | */ | |
1756 | goto free; | |
1757 | } | |
e4e574b7 AL |
1758 | /* |
1759 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1760 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1761 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1762 | * allocator. Commit the entire reservation here to prevent another |
1763 | * process from stealing the pages as they are added to the pool but | |
1764 | * before they are reserved. | |
e4e574b7 AL |
1765 | */ |
1766 | needed += allocated; | |
a5516438 | 1767 | h->resv_huge_pages += delta; |
e4e574b7 | 1768 | ret = 0; |
a9869b83 | 1769 | |
19fc3f0a | 1770 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1771 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1772 | if ((--needed) < 0) |
1773 | break; | |
a9869b83 NH |
1774 | /* |
1775 | * This page is now managed by the hugetlb allocator and has | |
1776 | * no users -- drop the buddy allocator's reference. | |
1777 | */ | |
1778 | put_page_testzero(page); | |
309381fe | 1779 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 1780 | enqueue_huge_page(h, page); |
19fc3f0a | 1781 | } |
28073b02 | 1782 | free: |
b0365c8d | 1783 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1784 | |
1785 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1786 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1787 | put_page(page); | |
a9869b83 | 1788 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1789 | |
1790 | return ret; | |
1791 | } | |
1792 | ||
1793 | /* | |
e5bbc8a6 MK |
1794 | * This routine has two main purposes: |
1795 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
1796 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
1797 | * to the associated reservation map. | |
1798 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
1799 | * the reservation. As many as unused_resv_pages may be freed. | |
1800 | * | |
1801 | * Called with hugetlb_lock held. However, the lock could be dropped (and | |
1802 | * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, | |
1803 | * we must make sure nobody else can claim pages we are in the process of | |
1804 | * freeing. Do this by ensuring resv_huge_page always is greater than the | |
1805 | * number of huge pages we plan to free when dropping the lock. | |
e4e574b7 | 1806 | */ |
a5516438 AK |
1807 | static void return_unused_surplus_pages(struct hstate *h, |
1808 | unsigned long unused_resv_pages) | |
e4e574b7 | 1809 | { |
e4e574b7 AL |
1810 | unsigned long nr_pages; |
1811 | ||
aa888a74 | 1812 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 1813 | if (hstate_is_gigantic(h)) |
e5bbc8a6 | 1814 | goto out; |
aa888a74 | 1815 | |
e5bbc8a6 MK |
1816 | /* |
1817 | * Part (or even all) of the reservation could have been backed | |
1818 | * by pre-allocated pages. Only free surplus pages. | |
1819 | */ | |
a5516438 | 1820 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1821 | |
685f3457 LS |
1822 | /* |
1823 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1824 | * evenly across all nodes with memory. Iterate across these nodes |
1825 | * until we can no longer free unreserved surplus pages. This occurs | |
1826 | * when the nodes with surplus pages have no free pages. | |
1827 | * free_pool_huge_page() will balance the the freed pages across the | |
1828 | * on-line nodes with memory and will handle the hstate accounting. | |
e5bbc8a6 MK |
1829 | * |
1830 | * Note that we decrement resv_huge_pages as we free the pages. If | |
1831 | * we drop the lock, resv_huge_pages will still be sufficiently large | |
1832 | * to cover subsequent pages we may free. | |
685f3457 LS |
1833 | */ |
1834 | while (nr_pages--) { | |
e5bbc8a6 MK |
1835 | h->resv_huge_pages--; |
1836 | unused_resv_pages--; | |
8cebfcd0 | 1837 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
e5bbc8a6 | 1838 | goto out; |
7848a4bf | 1839 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 | 1840 | } |
e5bbc8a6 MK |
1841 | |
1842 | out: | |
1843 | /* Fully uncommit the reservation */ | |
1844 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 AL |
1845 | } |
1846 | ||
5e911373 | 1847 | |
c37f9fb1 | 1848 | /* |
feba16e2 | 1849 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 1850 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
1851 | * |
1852 | * vma_needs_reservation is called to determine if the huge page at addr | |
1853 | * within the vma has an associated reservation. If a reservation is | |
1854 | * needed, the value 1 is returned. The caller is then responsible for | |
1855 | * managing the global reservation and subpool usage counts. After | |
1856 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
1857 | * to add the page to the reservation map. If the page allocation fails, |
1858 | * the reservation must be ended instead of committed. vma_end_reservation | |
1859 | * is called in such cases. | |
cf3ad20b MK |
1860 | * |
1861 | * In the normal case, vma_commit_reservation returns the same value | |
1862 | * as the preceding vma_needs_reservation call. The only time this | |
1863 | * is not the case is if a reserve map was changed between calls. It | |
1864 | * is the responsibility of the caller to notice the difference and | |
1865 | * take appropriate action. | |
96b96a96 MK |
1866 | * |
1867 | * vma_add_reservation is used in error paths where a reservation must | |
1868 | * be restored when a newly allocated huge page must be freed. It is | |
1869 | * to be called after calling vma_needs_reservation to determine if a | |
1870 | * reservation exists. | |
c37f9fb1 | 1871 | */ |
5e911373 MK |
1872 | enum vma_resv_mode { |
1873 | VMA_NEEDS_RESV, | |
1874 | VMA_COMMIT_RESV, | |
feba16e2 | 1875 | VMA_END_RESV, |
96b96a96 | 1876 | VMA_ADD_RESV, |
5e911373 | 1877 | }; |
cf3ad20b MK |
1878 | static long __vma_reservation_common(struct hstate *h, |
1879 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 1880 | enum vma_resv_mode mode) |
c37f9fb1 | 1881 | { |
4e35f483 JK |
1882 | struct resv_map *resv; |
1883 | pgoff_t idx; | |
cf3ad20b | 1884 | long ret; |
c37f9fb1 | 1885 | |
4e35f483 JK |
1886 | resv = vma_resv_map(vma); |
1887 | if (!resv) | |
84afd99b | 1888 | return 1; |
c37f9fb1 | 1889 | |
4e35f483 | 1890 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
1891 | switch (mode) { |
1892 | case VMA_NEEDS_RESV: | |
cf3ad20b | 1893 | ret = region_chg(resv, idx, idx + 1); |
5e911373 MK |
1894 | break; |
1895 | case VMA_COMMIT_RESV: | |
1896 | ret = region_add(resv, idx, idx + 1); | |
1897 | break; | |
feba16e2 | 1898 | case VMA_END_RESV: |
5e911373 MK |
1899 | region_abort(resv, idx, idx + 1); |
1900 | ret = 0; | |
1901 | break; | |
96b96a96 MK |
1902 | case VMA_ADD_RESV: |
1903 | if (vma->vm_flags & VM_MAYSHARE) | |
1904 | ret = region_add(resv, idx, idx + 1); | |
1905 | else { | |
1906 | region_abort(resv, idx, idx + 1); | |
1907 | ret = region_del(resv, idx, idx + 1); | |
1908 | } | |
1909 | break; | |
5e911373 MK |
1910 | default: |
1911 | BUG(); | |
1912 | } | |
84afd99b | 1913 | |
4e35f483 | 1914 | if (vma->vm_flags & VM_MAYSHARE) |
cf3ad20b | 1915 | return ret; |
67961f9d MK |
1916 | else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { |
1917 | /* | |
1918 | * In most cases, reserves always exist for private mappings. | |
1919 | * However, a file associated with mapping could have been | |
1920 | * hole punched or truncated after reserves were consumed. | |
1921 | * As subsequent fault on such a range will not use reserves. | |
1922 | * Subtle - The reserve map for private mappings has the | |
1923 | * opposite meaning than that of shared mappings. If NO | |
1924 | * entry is in the reserve map, it means a reservation exists. | |
1925 | * If an entry exists in the reserve map, it means the | |
1926 | * reservation has already been consumed. As a result, the | |
1927 | * return value of this routine is the opposite of the | |
1928 | * value returned from reserve map manipulation routines above. | |
1929 | */ | |
1930 | if (ret) | |
1931 | return 0; | |
1932 | else | |
1933 | return 1; | |
1934 | } | |
4e35f483 | 1935 | else |
cf3ad20b | 1936 | return ret < 0 ? ret : 0; |
c37f9fb1 | 1937 | } |
cf3ad20b MK |
1938 | |
1939 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 1940 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 1941 | { |
5e911373 | 1942 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 1943 | } |
84afd99b | 1944 | |
cf3ad20b MK |
1945 | static long vma_commit_reservation(struct hstate *h, |
1946 | struct vm_area_struct *vma, unsigned long addr) | |
1947 | { | |
5e911373 MK |
1948 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
1949 | } | |
1950 | ||
feba16e2 | 1951 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
1952 | struct vm_area_struct *vma, unsigned long addr) |
1953 | { | |
feba16e2 | 1954 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
1955 | } |
1956 | ||
96b96a96 MK |
1957 | static long vma_add_reservation(struct hstate *h, |
1958 | struct vm_area_struct *vma, unsigned long addr) | |
1959 | { | |
1960 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
1961 | } | |
1962 | ||
1963 | /* | |
1964 | * This routine is called to restore a reservation on error paths. In the | |
1965 | * specific error paths, a huge page was allocated (via alloc_huge_page) | |
1966 | * and is about to be freed. If a reservation for the page existed, | |
1967 | * alloc_huge_page would have consumed the reservation and set PagePrivate | |
1968 | * in the newly allocated page. When the page is freed via free_huge_page, | |
1969 | * the global reservation count will be incremented if PagePrivate is set. | |
1970 | * However, free_huge_page can not adjust the reserve map. Adjust the | |
1971 | * reserve map here to be consistent with global reserve count adjustments | |
1972 | * to be made by free_huge_page. | |
1973 | */ | |
1974 | static void restore_reserve_on_error(struct hstate *h, | |
1975 | struct vm_area_struct *vma, unsigned long address, | |
1976 | struct page *page) | |
1977 | { | |
1978 | if (unlikely(PagePrivate(page))) { | |
1979 | long rc = vma_needs_reservation(h, vma, address); | |
1980 | ||
1981 | if (unlikely(rc < 0)) { | |
1982 | /* | |
1983 | * Rare out of memory condition in reserve map | |
1984 | * manipulation. Clear PagePrivate so that | |
1985 | * global reserve count will not be incremented | |
1986 | * by free_huge_page. This will make it appear | |
1987 | * as though the reservation for this page was | |
1988 | * consumed. This may prevent the task from | |
1989 | * faulting in the page at a later time. This | |
1990 | * is better than inconsistent global huge page | |
1991 | * accounting of reserve counts. | |
1992 | */ | |
1993 | ClearPagePrivate(page); | |
1994 | } else if (rc) { | |
1995 | rc = vma_add_reservation(h, vma, address); | |
1996 | if (unlikely(rc < 0)) | |
1997 | /* | |
1998 | * See above comment about rare out of | |
1999 | * memory condition. | |
2000 | */ | |
2001 | ClearPagePrivate(page); | |
2002 | } else | |
2003 | vma_end_reservation(h, vma, address); | |
2004 | } | |
2005 | } | |
2006 | ||
70c3547e | 2007 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 2008 | unsigned long addr, int avoid_reserve) |
1da177e4 | 2009 | { |
90481622 | 2010 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 2011 | struct hstate *h = hstate_vma(vma); |
348ea204 | 2012 | struct page *page; |
d85f69b0 MK |
2013 | long map_chg, map_commit; |
2014 | long gbl_chg; | |
6d76dcf4 AK |
2015 | int ret, idx; |
2016 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 2017 | |
6d76dcf4 | 2018 | idx = hstate_index(h); |
a1e78772 | 2019 | /* |
d85f69b0 MK |
2020 | * Examine the region/reserve map to determine if the process |
2021 | * has a reservation for the page to be allocated. A return | |
2022 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 2023 | */ |
d85f69b0 MK |
2024 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
2025 | if (map_chg < 0) | |
76dcee75 | 2026 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
2027 | |
2028 | /* | |
2029 | * Processes that did not create the mapping will have no | |
2030 | * reserves as indicated by the region/reserve map. Check | |
2031 | * that the allocation will not exceed the subpool limit. | |
2032 | * Allocations for MAP_NORESERVE mappings also need to be | |
2033 | * checked against any subpool limit. | |
2034 | */ | |
2035 | if (map_chg || avoid_reserve) { | |
2036 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2037 | if (gbl_chg < 0) { | |
feba16e2 | 2038 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2039 | return ERR_PTR(-ENOSPC); |
5e911373 | 2040 | } |
1da177e4 | 2041 | |
d85f69b0 MK |
2042 | /* |
2043 | * Even though there was no reservation in the region/reserve | |
2044 | * map, there could be reservations associated with the | |
2045 | * subpool that can be used. This would be indicated if the | |
2046 | * return value of hugepage_subpool_get_pages() is zero. | |
2047 | * However, if avoid_reserve is specified we still avoid even | |
2048 | * the subpool reservations. | |
2049 | */ | |
2050 | if (avoid_reserve) | |
2051 | gbl_chg = 1; | |
2052 | } | |
2053 | ||
6d76dcf4 | 2054 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f JZ |
2055 | if (ret) |
2056 | goto out_subpool_put; | |
2057 | ||
1da177e4 | 2058 | spin_lock(&hugetlb_lock); |
d85f69b0 MK |
2059 | /* |
2060 | * glb_chg is passed to indicate whether or not a page must be taken | |
2061 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2062 | * a reservation exists for the allocation. | |
2063 | */ | |
2064 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2065 | if (!page) { |
94ae8ba7 | 2066 | spin_unlock(&hugetlb_lock); |
099730d6 | 2067 | page = __alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2068 | if (!page) |
2069 | goto out_uncharge_cgroup; | |
a88c7695 NH |
2070 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
2071 | SetPagePrivate(page); | |
2072 | h->resv_huge_pages--; | |
2073 | } | |
79dbb236 AK |
2074 | spin_lock(&hugetlb_lock); |
2075 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 2076 | /* Fall through */ |
68842c9b | 2077 | } |
81a6fcae JK |
2078 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
2079 | spin_unlock(&hugetlb_lock); | |
348ea204 | 2080 | |
90481622 | 2081 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 2082 | |
d85f69b0 MK |
2083 | map_commit = vma_commit_reservation(h, vma, addr); |
2084 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2085 | /* |
2086 | * The page was added to the reservation map between | |
2087 | * vma_needs_reservation and vma_commit_reservation. | |
2088 | * This indicates a race with hugetlb_reserve_pages. | |
2089 | * Adjust for the subpool count incremented above AND | |
2090 | * in hugetlb_reserve_pages for the same page. Also, | |
2091 | * the reservation count added in hugetlb_reserve_pages | |
2092 | * no longer applies. | |
2093 | */ | |
2094 | long rsv_adjust; | |
2095 | ||
2096 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2097 | hugetlb_acct_memory(h, -rsv_adjust); | |
2098 | } | |
90d8b7e6 | 2099 | return page; |
8f34af6f JZ |
2100 | |
2101 | out_uncharge_cgroup: | |
2102 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
2103 | out_subpool_put: | |
d85f69b0 | 2104 | if (map_chg || avoid_reserve) |
8f34af6f | 2105 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2106 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2107 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2108 | } |
2109 | ||
74060e4d NH |
2110 | /* |
2111 | * alloc_huge_page()'s wrapper which simply returns the page if allocation | |
2112 | * succeeds, otherwise NULL. This function is called from new_vma_page(), | |
2113 | * where no ERR_VALUE is expected to be returned. | |
2114 | */ | |
2115 | struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, | |
2116 | unsigned long addr, int avoid_reserve) | |
2117 | { | |
2118 | struct page *page = alloc_huge_page(vma, addr, avoid_reserve); | |
2119 | if (IS_ERR(page)) | |
2120 | page = NULL; | |
2121 | return page; | |
2122 | } | |
2123 | ||
91f47662 | 2124 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
aa888a74 AK |
2125 | { |
2126 | struct huge_bootmem_page *m; | |
b2261026 | 2127 | int nr_nodes, node; |
aa888a74 | 2128 | |
b2261026 | 2129 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
2130 | void *addr; |
2131 | ||
8b89a116 GS |
2132 | addr = memblock_virt_alloc_try_nid_nopanic( |
2133 | huge_page_size(h), huge_page_size(h), | |
2134 | 0, BOOTMEM_ALLOC_ACCESSIBLE, node); | |
aa888a74 AK |
2135 | if (addr) { |
2136 | /* | |
2137 | * Use the beginning of the huge page to store the | |
2138 | * huge_bootmem_page struct (until gather_bootmem | |
2139 | * puts them into the mem_map). | |
2140 | */ | |
2141 | m = addr; | |
91f47662 | 2142 | goto found; |
aa888a74 | 2143 | } |
aa888a74 AK |
2144 | } |
2145 | return 0; | |
2146 | ||
2147 | found: | |
df994ead | 2148 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 AK |
2149 | /* Put them into a private list first because mem_map is not up yet */ |
2150 | list_add(&m->list, &huge_boot_pages); | |
2151 | m->hstate = h; | |
2152 | return 1; | |
2153 | } | |
2154 | ||
d00181b9 KS |
2155 | static void __init prep_compound_huge_page(struct page *page, |
2156 | unsigned int order) | |
18229df5 AW |
2157 | { |
2158 | if (unlikely(order > (MAX_ORDER - 1))) | |
2159 | prep_compound_gigantic_page(page, order); | |
2160 | else | |
2161 | prep_compound_page(page, order); | |
2162 | } | |
2163 | ||
aa888a74 AK |
2164 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
2165 | static void __init gather_bootmem_prealloc(void) | |
2166 | { | |
2167 | struct huge_bootmem_page *m; | |
2168 | ||
2169 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 2170 | struct hstate *h = m->hstate; |
ee8f248d BB |
2171 | struct page *page; |
2172 | ||
2173 | #ifdef CONFIG_HIGHMEM | |
2174 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
8b89a116 GS |
2175 | memblock_free_late(__pa(m), |
2176 | sizeof(struct huge_bootmem_page)); | |
ee8f248d BB |
2177 | #else |
2178 | page = virt_to_page(m); | |
2179 | #endif | |
aa888a74 | 2180 | WARN_ON(page_count(page) != 1); |
18229df5 | 2181 | prep_compound_huge_page(page, h->order); |
ef5a22be | 2182 | WARN_ON(PageReserved(page)); |
aa888a74 | 2183 | prep_new_huge_page(h, page, page_to_nid(page)); |
b0320c7b RA |
2184 | /* |
2185 | * If we had gigantic hugepages allocated at boot time, we need | |
2186 | * to restore the 'stolen' pages to totalram_pages in order to | |
2187 | * fix confusing memory reports from free(1) and another | |
2188 | * side-effects, like CommitLimit going negative. | |
2189 | */ | |
bae7f4ae | 2190 | if (hstate_is_gigantic(h)) |
3dcc0571 | 2191 | adjust_managed_page_count(page, 1 << h->order); |
aa888a74 AK |
2192 | } |
2193 | } | |
2194 | ||
8faa8b07 | 2195 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
2196 | { |
2197 | unsigned long i; | |
a5516438 | 2198 | |
e5ff2159 | 2199 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 2200 | if (hstate_is_gigantic(h)) { |
aa888a74 AK |
2201 | if (!alloc_bootmem_huge_page(h)) |
2202 | break; | |
9b5e5d0f | 2203 | } else if (!alloc_fresh_huge_page(h, |
8cebfcd0 | 2204 | &node_states[N_MEMORY])) |
1da177e4 | 2205 | break; |
1da177e4 | 2206 | } |
8faa8b07 | 2207 | h->max_huge_pages = i; |
e5ff2159 AK |
2208 | } |
2209 | ||
2210 | static void __init hugetlb_init_hstates(void) | |
2211 | { | |
2212 | struct hstate *h; | |
2213 | ||
2214 | for_each_hstate(h) { | |
641844f5 NH |
2215 | if (minimum_order > huge_page_order(h)) |
2216 | minimum_order = huge_page_order(h); | |
2217 | ||
8faa8b07 | 2218 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 2219 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 2220 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 2221 | } |
641844f5 | 2222 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
2223 | } |
2224 | ||
4abd32db AK |
2225 | static char * __init memfmt(char *buf, unsigned long n) |
2226 | { | |
2227 | if (n >= (1UL << 30)) | |
2228 | sprintf(buf, "%lu GB", n >> 30); | |
2229 | else if (n >= (1UL << 20)) | |
2230 | sprintf(buf, "%lu MB", n >> 20); | |
2231 | else | |
2232 | sprintf(buf, "%lu KB", n >> 10); | |
2233 | return buf; | |
2234 | } | |
2235 | ||
e5ff2159 AK |
2236 | static void __init report_hugepages(void) |
2237 | { | |
2238 | struct hstate *h; | |
2239 | ||
2240 | for_each_hstate(h) { | |
4abd32db | 2241 | char buf[32]; |
ffb22af5 | 2242 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
4abd32db AK |
2243 | memfmt(buf, huge_page_size(h)), |
2244 | h->free_huge_pages); | |
e5ff2159 AK |
2245 | } |
2246 | } | |
2247 | ||
1da177e4 | 2248 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
2249 | static void try_to_free_low(struct hstate *h, unsigned long count, |
2250 | nodemask_t *nodes_allowed) | |
1da177e4 | 2251 | { |
4415cc8d CL |
2252 | int i; |
2253 | ||
bae7f4ae | 2254 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2255 | return; |
2256 | ||
6ae11b27 | 2257 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 2258 | struct page *page, *next; |
a5516438 AK |
2259 | struct list_head *freel = &h->hugepage_freelists[i]; |
2260 | list_for_each_entry_safe(page, next, freel, lru) { | |
2261 | if (count >= h->nr_huge_pages) | |
6b0c880d | 2262 | return; |
1da177e4 LT |
2263 | if (PageHighMem(page)) |
2264 | continue; | |
2265 | list_del(&page->lru); | |
e5ff2159 | 2266 | update_and_free_page(h, page); |
a5516438 AK |
2267 | h->free_huge_pages--; |
2268 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
2269 | } |
2270 | } | |
2271 | } | |
2272 | #else | |
6ae11b27 LS |
2273 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
2274 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
2275 | { |
2276 | } | |
2277 | #endif | |
2278 | ||
20a0307c WF |
2279 | /* |
2280 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2281 | * balanced by operating on them in a round-robin fashion. | |
2282 | * Returns 1 if an adjustment was made. | |
2283 | */ | |
6ae11b27 LS |
2284 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
2285 | int delta) | |
20a0307c | 2286 | { |
b2261026 | 2287 | int nr_nodes, node; |
20a0307c WF |
2288 | |
2289 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 2290 | |
b2261026 JK |
2291 | if (delta < 0) { |
2292 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2293 | if (h->surplus_huge_pages_node[node]) | |
2294 | goto found; | |
e8c5c824 | 2295 | } |
b2261026 JK |
2296 | } else { |
2297 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2298 | if (h->surplus_huge_pages_node[node] < | |
2299 | h->nr_huge_pages_node[node]) | |
2300 | goto found; | |
e8c5c824 | 2301 | } |
b2261026 JK |
2302 | } |
2303 | return 0; | |
20a0307c | 2304 | |
b2261026 JK |
2305 | found: |
2306 | h->surplus_huge_pages += delta; | |
2307 | h->surplus_huge_pages_node[node] += delta; | |
2308 | return 1; | |
20a0307c WF |
2309 | } |
2310 | ||
a5516438 | 2311 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
2312 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
2313 | nodemask_t *nodes_allowed) | |
1da177e4 | 2314 | { |
7893d1d5 | 2315 | unsigned long min_count, ret; |
1da177e4 | 2316 | |
944d9fec | 2317 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
aa888a74 AK |
2318 | return h->max_huge_pages; |
2319 | ||
7893d1d5 AL |
2320 | /* |
2321 | * Increase the pool size | |
2322 | * First take pages out of surplus state. Then make up the | |
2323 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 2324 | * |
d15c7c09 | 2325 | * We might race with __alloc_buddy_huge_page() here and be unable |
d1c3fb1f NA |
2326 | * to convert a surplus huge page to a normal huge page. That is |
2327 | * not critical, though, it just means the overall size of the | |
2328 | * pool might be one hugepage larger than it needs to be, but | |
2329 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 2330 | */ |
1da177e4 | 2331 | spin_lock(&hugetlb_lock); |
a5516438 | 2332 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 2333 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
2334 | break; |
2335 | } | |
2336 | ||
a5516438 | 2337 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
2338 | /* |
2339 | * If this allocation races such that we no longer need the | |
2340 | * page, free_huge_page will handle it by freeing the page | |
2341 | * and reducing the surplus. | |
2342 | */ | |
2343 | spin_unlock(&hugetlb_lock); | |
649920c6 JH |
2344 | |
2345 | /* yield cpu to avoid soft lockup */ | |
2346 | cond_resched(); | |
2347 | ||
944d9fec LC |
2348 | if (hstate_is_gigantic(h)) |
2349 | ret = alloc_fresh_gigantic_page(h, nodes_allowed); | |
2350 | else | |
2351 | ret = alloc_fresh_huge_page(h, nodes_allowed); | |
7893d1d5 AL |
2352 | spin_lock(&hugetlb_lock); |
2353 | if (!ret) | |
2354 | goto out; | |
2355 | ||
536240f2 MG |
2356 | /* Bail for signals. Probably ctrl-c from user */ |
2357 | if (signal_pending(current)) | |
2358 | goto out; | |
7893d1d5 | 2359 | } |
7893d1d5 AL |
2360 | |
2361 | /* | |
2362 | * Decrease the pool size | |
2363 | * First return free pages to the buddy allocator (being careful | |
2364 | * to keep enough around to satisfy reservations). Then place | |
2365 | * pages into surplus state as needed so the pool will shrink | |
2366 | * to the desired size as pages become free. | |
d1c3fb1f NA |
2367 | * |
2368 | * By placing pages into the surplus state independent of the | |
2369 | * overcommit value, we are allowing the surplus pool size to | |
2370 | * exceed overcommit. There are few sane options here. Since | |
d15c7c09 | 2371 | * __alloc_buddy_huge_page() is checking the global counter, |
d1c3fb1f NA |
2372 | * though, we'll note that we're not allowed to exceed surplus |
2373 | * and won't grow the pool anywhere else. Not until one of the | |
2374 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 2375 | */ |
a5516438 | 2376 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 2377 | min_count = max(count, min_count); |
6ae11b27 | 2378 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 2379 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 2380 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 2381 | break; |
55f67141 | 2382 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 2383 | } |
a5516438 | 2384 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 2385 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
2386 | break; |
2387 | } | |
2388 | out: | |
a5516438 | 2389 | ret = persistent_huge_pages(h); |
1da177e4 | 2390 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 2391 | return ret; |
1da177e4 LT |
2392 | } |
2393 | ||
a3437870 NA |
2394 | #define HSTATE_ATTR_RO(_name) \ |
2395 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2396 | ||
2397 | #define HSTATE_ATTR(_name) \ | |
2398 | static struct kobj_attribute _name##_attr = \ | |
2399 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2400 | ||
2401 | static struct kobject *hugepages_kobj; | |
2402 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2403 | ||
9a305230 LS |
2404 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
2405 | ||
2406 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
2407 | { |
2408 | int i; | |
9a305230 | 2409 | |
a3437870 | 2410 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
2411 | if (hstate_kobjs[i] == kobj) { |
2412 | if (nidp) | |
2413 | *nidp = NUMA_NO_NODE; | |
a3437870 | 2414 | return &hstates[i]; |
9a305230 LS |
2415 | } |
2416 | ||
2417 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
2418 | } |
2419 | ||
06808b08 | 2420 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
2421 | struct kobj_attribute *attr, char *buf) |
2422 | { | |
9a305230 LS |
2423 | struct hstate *h; |
2424 | unsigned long nr_huge_pages; | |
2425 | int nid; | |
2426 | ||
2427 | h = kobj_to_hstate(kobj, &nid); | |
2428 | if (nid == NUMA_NO_NODE) | |
2429 | nr_huge_pages = h->nr_huge_pages; | |
2430 | else | |
2431 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2432 | ||
2433 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 2434 | } |
adbe8726 | 2435 | |
238d3c13 DR |
2436 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
2437 | struct hstate *h, int nid, | |
2438 | unsigned long count, size_t len) | |
a3437870 NA |
2439 | { |
2440 | int err; | |
bad44b5b | 2441 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 2442 | |
944d9fec | 2443 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) { |
adbe8726 EM |
2444 | err = -EINVAL; |
2445 | goto out; | |
2446 | } | |
2447 | ||
9a305230 LS |
2448 | if (nid == NUMA_NO_NODE) { |
2449 | /* | |
2450 | * global hstate attribute | |
2451 | */ | |
2452 | if (!(obey_mempolicy && | |
2453 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
2454 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 2455 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 LS |
2456 | } |
2457 | } else if (nodes_allowed) { | |
2458 | /* | |
2459 | * per node hstate attribute: adjust count to global, | |
2460 | * but restrict alloc/free to the specified node. | |
2461 | */ | |
2462 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2463 | init_nodemask_of_node(nodes_allowed, nid); | |
2464 | } else | |
8cebfcd0 | 2465 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 | 2466 | |
06808b08 | 2467 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 2468 | |
8cebfcd0 | 2469 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
2470 | NODEMASK_FREE(nodes_allowed); |
2471 | ||
2472 | return len; | |
adbe8726 EM |
2473 | out: |
2474 | NODEMASK_FREE(nodes_allowed); | |
2475 | return err; | |
06808b08 LS |
2476 | } |
2477 | ||
238d3c13 DR |
2478 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
2479 | struct kobject *kobj, const char *buf, | |
2480 | size_t len) | |
2481 | { | |
2482 | struct hstate *h; | |
2483 | unsigned long count; | |
2484 | int nid; | |
2485 | int err; | |
2486 | ||
2487 | err = kstrtoul(buf, 10, &count); | |
2488 | if (err) | |
2489 | return err; | |
2490 | ||
2491 | h = kobj_to_hstate(kobj, &nid); | |
2492 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2493 | } | |
2494 | ||
06808b08 LS |
2495 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
2496 | struct kobj_attribute *attr, char *buf) | |
2497 | { | |
2498 | return nr_hugepages_show_common(kobj, attr, buf); | |
2499 | } | |
2500 | ||
2501 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2502 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2503 | { | |
238d3c13 | 2504 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
2505 | } |
2506 | HSTATE_ATTR(nr_hugepages); | |
2507 | ||
06808b08 LS |
2508 | #ifdef CONFIG_NUMA |
2509 | ||
2510 | /* | |
2511 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2512 | * huge page alloc/free. | |
2513 | */ | |
2514 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
2515 | struct kobj_attribute *attr, char *buf) | |
2516 | { | |
2517 | return nr_hugepages_show_common(kobj, attr, buf); | |
2518 | } | |
2519 | ||
2520 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2521 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2522 | { | |
238d3c13 | 2523 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
2524 | } |
2525 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2526 | #endif | |
2527 | ||
2528 | ||
a3437870 NA |
2529 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
2530 | struct kobj_attribute *attr, char *buf) | |
2531 | { | |
9a305230 | 2532 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2533 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
2534 | } | |
adbe8726 | 2535 | |
a3437870 NA |
2536 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
2537 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2538 | { | |
2539 | int err; | |
2540 | unsigned long input; | |
9a305230 | 2541 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2542 | |
bae7f4ae | 2543 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2544 | return -EINVAL; |
2545 | ||
3dbb95f7 | 2546 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2547 | if (err) |
73ae31e5 | 2548 | return err; |
a3437870 NA |
2549 | |
2550 | spin_lock(&hugetlb_lock); | |
2551 | h->nr_overcommit_huge_pages = input; | |
2552 | spin_unlock(&hugetlb_lock); | |
2553 | ||
2554 | return count; | |
2555 | } | |
2556 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2557 | ||
2558 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2559 | struct kobj_attribute *attr, char *buf) | |
2560 | { | |
9a305230 LS |
2561 | struct hstate *h; |
2562 | unsigned long free_huge_pages; | |
2563 | int nid; | |
2564 | ||
2565 | h = kobj_to_hstate(kobj, &nid); | |
2566 | if (nid == NUMA_NO_NODE) | |
2567 | free_huge_pages = h->free_huge_pages; | |
2568 | else | |
2569 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2570 | ||
2571 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
2572 | } |
2573 | HSTATE_ATTR_RO(free_hugepages); | |
2574 | ||
2575 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2576 | struct kobj_attribute *attr, char *buf) | |
2577 | { | |
9a305230 | 2578 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2579 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
2580 | } | |
2581 | HSTATE_ATTR_RO(resv_hugepages); | |
2582 | ||
2583 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2584 | struct kobj_attribute *attr, char *buf) | |
2585 | { | |
9a305230 LS |
2586 | struct hstate *h; |
2587 | unsigned long surplus_huge_pages; | |
2588 | int nid; | |
2589 | ||
2590 | h = kobj_to_hstate(kobj, &nid); | |
2591 | if (nid == NUMA_NO_NODE) | |
2592 | surplus_huge_pages = h->surplus_huge_pages; | |
2593 | else | |
2594 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2595 | ||
2596 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
2597 | } |
2598 | HSTATE_ATTR_RO(surplus_hugepages); | |
2599 | ||
2600 | static struct attribute *hstate_attrs[] = { | |
2601 | &nr_hugepages_attr.attr, | |
2602 | &nr_overcommit_hugepages_attr.attr, | |
2603 | &free_hugepages_attr.attr, | |
2604 | &resv_hugepages_attr.attr, | |
2605 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
2606 | #ifdef CONFIG_NUMA |
2607 | &nr_hugepages_mempolicy_attr.attr, | |
2608 | #endif | |
a3437870 NA |
2609 | NULL, |
2610 | }; | |
2611 | ||
2612 | static struct attribute_group hstate_attr_group = { | |
2613 | .attrs = hstate_attrs, | |
2614 | }; | |
2615 | ||
094e9539 JM |
2616 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
2617 | struct kobject **hstate_kobjs, | |
2618 | struct attribute_group *hstate_attr_group) | |
a3437870 NA |
2619 | { |
2620 | int retval; | |
972dc4de | 2621 | int hi = hstate_index(h); |
a3437870 | 2622 | |
9a305230 LS |
2623 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
2624 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
2625 | return -ENOMEM; |
2626 | ||
9a305230 | 2627 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 2628 | if (retval) |
9a305230 | 2629 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
2630 | |
2631 | return retval; | |
2632 | } | |
2633 | ||
2634 | static void __init hugetlb_sysfs_init(void) | |
2635 | { | |
2636 | struct hstate *h; | |
2637 | int err; | |
2638 | ||
2639 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
2640 | if (!hugepages_kobj) | |
2641 | return; | |
2642 | ||
2643 | for_each_hstate(h) { | |
9a305230 LS |
2644 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
2645 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 2646 | if (err) |
ffb22af5 | 2647 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
2648 | } |
2649 | } | |
2650 | ||
9a305230 LS |
2651 | #ifdef CONFIG_NUMA |
2652 | ||
2653 | /* | |
2654 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
2655 | * with node devices in node_devices[] using a parallel array. The array |
2656 | * index of a node device or _hstate == node id. | |
2657 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
2658 | * the base kernel, on the hugetlb module. |
2659 | */ | |
2660 | struct node_hstate { | |
2661 | struct kobject *hugepages_kobj; | |
2662 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2663 | }; | |
b4e289a6 | 2664 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
2665 | |
2666 | /* | |
10fbcf4c | 2667 | * A subset of global hstate attributes for node devices |
9a305230 LS |
2668 | */ |
2669 | static struct attribute *per_node_hstate_attrs[] = { | |
2670 | &nr_hugepages_attr.attr, | |
2671 | &free_hugepages_attr.attr, | |
2672 | &surplus_hugepages_attr.attr, | |
2673 | NULL, | |
2674 | }; | |
2675 | ||
2676 | static struct attribute_group per_node_hstate_attr_group = { | |
2677 | .attrs = per_node_hstate_attrs, | |
2678 | }; | |
2679 | ||
2680 | /* | |
10fbcf4c | 2681 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
2682 | * Returns node id via non-NULL nidp. |
2683 | */ | |
2684 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2685 | { | |
2686 | int nid; | |
2687 | ||
2688 | for (nid = 0; nid < nr_node_ids; nid++) { | |
2689 | struct node_hstate *nhs = &node_hstates[nid]; | |
2690 | int i; | |
2691 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
2692 | if (nhs->hstate_kobjs[i] == kobj) { | |
2693 | if (nidp) | |
2694 | *nidp = nid; | |
2695 | return &hstates[i]; | |
2696 | } | |
2697 | } | |
2698 | ||
2699 | BUG(); | |
2700 | return NULL; | |
2701 | } | |
2702 | ||
2703 | /* | |
10fbcf4c | 2704 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
2705 | * No-op if no hstate attributes attached. |
2706 | */ | |
3cd8b44f | 2707 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
2708 | { |
2709 | struct hstate *h; | |
10fbcf4c | 2710 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2711 | |
2712 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 2713 | return; /* no hstate attributes */ |
9a305230 | 2714 | |
972dc4de AK |
2715 | for_each_hstate(h) { |
2716 | int idx = hstate_index(h); | |
2717 | if (nhs->hstate_kobjs[idx]) { | |
2718 | kobject_put(nhs->hstate_kobjs[idx]); | |
2719 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 2720 | } |
972dc4de | 2721 | } |
9a305230 LS |
2722 | |
2723 | kobject_put(nhs->hugepages_kobj); | |
2724 | nhs->hugepages_kobj = NULL; | |
2725 | } | |
2726 | ||
9a305230 LS |
2727 | |
2728 | /* | |
10fbcf4c | 2729 | * Register hstate attributes for a single node device. |
9a305230 LS |
2730 | * No-op if attributes already registered. |
2731 | */ | |
3cd8b44f | 2732 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
2733 | { |
2734 | struct hstate *h; | |
10fbcf4c | 2735 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2736 | int err; |
2737 | ||
2738 | if (nhs->hugepages_kobj) | |
2739 | return; /* already allocated */ | |
2740 | ||
2741 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 2742 | &node->dev.kobj); |
9a305230 LS |
2743 | if (!nhs->hugepages_kobj) |
2744 | return; | |
2745 | ||
2746 | for_each_hstate(h) { | |
2747 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
2748 | nhs->hstate_kobjs, | |
2749 | &per_node_hstate_attr_group); | |
2750 | if (err) { | |
ffb22af5 AM |
2751 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
2752 | h->name, node->dev.id); | |
9a305230 LS |
2753 | hugetlb_unregister_node(node); |
2754 | break; | |
2755 | } | |
2756 | } | |
2757 | } | |
2758 | ||
2759 | /* | |
9b5e5d0f | 2760 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
2761 | * devices of nodes that have memory. All on-line nodes should have |
2762 | * registered their associated device by this time. | |
9a305230 | 2763 | */ |
7d9ca000 | 2764 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
2765 | { |
2766 | int nid; | |
2767 | ||
8cebfcd0 | 2768 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 2769 | struct node *node = node_devices[nid]; |
10fbcf4c | 2770 | if (node->dev.id == nid) |
9a305230 LS |
2771 | hugetlb_register_node(node); |
2772 | } | |
2773 | ||
2774 | /* | |
10fbcf4c | 2775 | * Let the node device driver know we're here so it can |
9a305230 LS |
2776 | * [un]register hstate attributes on node hotplug. |
2777 | */ | |
2778 | register_hugetlbfs_with_node(hugetlb_register_node, | |
2779 | hugetlb_unregister_node); | |
2780 | } | |
2781 | #else /* !CONFIG_NUMA */ | |
2782 | ||
2783 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2784 | { | |
2785 | BUG(); | |
2786 | if (nidp) | |
2787 | *nidp = -1; | |
2788 | return NULL; | |
2789 | } | |
2790 | ||
9a305230 LS |
2791 | static void hugetlb_register_all_nodes(void) { } |
2792 | ||
2793 | #endif | |
2794 | ||
a3437870 NA |
2795 | static int __init hugetlb_init(void) |
2796 | { | |
8382d914 DB |
2797 | int i; |
2798 | ||
457c1b27 | 2799 | if (!hugepages_supported()) |
0ef89d25 | 2800 | return 0; |
a3437870 | 2801 | |
e11bfbfc NP |
2802 | if (!size_to_hstate(default_hstate_size)) { |
2803 | default_hstate_size = HPAGE_SIZE; | |
2804 | if (!size_to_hstate(default_hstate_size)) | |
2805 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 2806 | } |
972dc4de | 2807 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
f8b74815 VT |
2808 | if (default_hstate_max_huge_pages) { |
2809 | if (!default_hstate.max_huge_pages) | |
2810 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
2811 | } | |
a3437870 NA |
2812 | |
2813 | hugetlb_init_hstates(); | |
aa888a74 | 2814 | gather_bootmem_prealloc(); |
a3437870 NA |
2815 | report_hugepages(); |
2816 | ||
2817 | hugetlb_sysfs_init(); | |
9a305230 | 2818 | hugetlb_register_all_nodes(); |
7179e7bf | 2819 | hugetlb_cgroup_file_init(); |
9a305230 | 2820 | |
8382d914 DB |
2821 | #ifdef CONFIG_SMP |
2822 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
2823 | #else | |
2824 | num_fault_mutexes = 1; | |
2825 | #endif | |
c672c7f2 | 2826 | hugetlb_fault_mutex_table = |
8382d914 | 2827 | kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); |
c672c7f2 | 2828 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
2829 | |
2830 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 2831 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
2832 | return 0; |
2833 | } | |
3e89e1c5 | 2834 | subsys_initcall(hugetlb_init); |
a3437870 NA |
2835 | |
2836 | /* Should be called on processing a hugepagesz=... option */ | |
9fee021d VT |
2837 | void __init hugetlb_bad_size(void) |
2838 | { | |
2839 | parsed_valid_hugepagesz = false; | |
2840 | } | |
2841 | ||
d00181b9 | 2842 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
2843 | { |
2844 | struct hstate *h; | |
8faa8b07 AK |
2845 | unsigned long i; |
2846 | ||
a3437870 | 2847 | if (size_to_hstate(PAGE_SIZE << order)) { |
598d8091 | 2848 | pr_warn("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
2849 | return; |
2850 | } | |
47d38344 | 2851 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 2852 | BUG_ON(order == 0); |
47d38344 | 2853 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
2854 | h->order = order; |
2855 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
2856 | h->nr_huge_pages = 0; |
2857 | h->free_huge_pages = 0; | |
2858 | for (i = 0; i < MAX_NUMNODES; ++i) | |
2859 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 2860 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
2861 | h->next_nid_to_alloc = first_memory_node; |
2862 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
2863 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
2864 | huge_page_size(h)/1024); | |
8faa8b07 | 2865 | |
a3437870 NA |
2866 | parsed_hstate = h; |
2867 | } | |
2868 | ||
e11bfbfc | 2869 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
2870 | { |
2871 | unsigned long *mhp; | |
8faa8b07 | 2872 | static unsigned long *last_mhp; |
a3437870 | 2873 | |
9fee021d VT |
2874 | if (!parsed_valid_hugepagesz) { |
2875 | pr_warn("hugepages = %s preceded by " | |
2876 | "an unsupported hugepagesz, ignoring\n", s); | |
2877 | parsed_valid_hugepagesz = true; | |
2878 | return 1; | |
2879 | } | |
a3437870 | 2880 | /* |
47d38344 | 2881 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
2882 | * so this hugepages= parameter goes to the "default hstate". |
2883 | */ | |
9fee021d | 2884 | else if (!hugetlb_max_hstate) |
a3437870 NA |
2885 | mhp = &default_hstate_max_huge_pages; |
2886 | else | |
2887 | mhp = &parsed_hstate->max_huge_pages; | |
2888 | ||
8faa8b07 | 2889 | if (mhp == last_mhp) { |
598d8091 | 2890 | pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); |
8faa8b07 AK |
2891 | return 1; |
2892 | } | |
2893 | ||
a3437870 NA |
2894 | if (sscanf(s, "%lu", mhp) <= 0) |
2895 | *mhp = 0; | |
2896 | ||
8faa8b07 AK |
2897 | /* |
2898 | * Global state is always initialized later in hugetlb_init. | |
2899 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
2900 | * use the bootmem allocator. | |
2901 | */ | |
47d38344 | 2902 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
2903 | hugetlb_hstate_alloc_pages(parsed_hstate); |
2904 | ||
2905 | last_mhp = mhp; | |
2906 | ||
a3437870 NA |
2907 | return 1; |
2908 | } | |
e11bfbfc NP |
2909 | __setup("hugepages=", hugetlb_nrpages_setup); |
2910 | ||
2911 | static int __init hugetlb_default_setup(char *s) | |
2912 | { | |
2913 | default_hstate_size = memparse(s, &s); | |
2914 | return 1; | |
2915 | } | |
2916 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 2917 | |
8a213460 NA |
2918 | static unsigned int cpuset_mems_nr(unsigned int *array) |
2919 | { | |
2920 | int node; | |
2921 | unsigned int nr = 0; | |
2922 | ||
2923 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
2924 | nr += array[node]; | |
2925 | ||
2926 | return nr; | |
2927 | } | |
2928 | ||
2929 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
2930 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
2931 | struct ctl_table *table, int write, | |
2932 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 2933 | { |
e5ff2159 | 2934 | struct hstate *h = &default_hstate; |
238d3c13 | 2935 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 2936 | int ret; |
e5ff2159 | 2937 | |
457c1b27 | 2938 | if (!hugepages_supported()) |
86613628 | 2939 | return -EOPNOTSUPP; |
457c1b27 | 2940 | |
e5ff2159 AK |
2941 | table->data = &tmp; |
2942 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2943 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2944 | if (ret) | |
2945 | goto out; | |
e5ff2159 | 2946 | |
238d3c13 DR |
2947 | if (write) |
2948 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
2949 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
2950 | out: |
2951 | return ret; | |
1da177e4 | 2952 | } |
396faf03 | 2953 | |
06808b08 LS |
2954 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2955 | void __user *buffer, size_t *length, loff_t *ppos) | |
2956 | { | |
2957 | ||
2958 | return hugetlb_sysctl_handler_common(false, table, write, | |
2959 | buffer, length, ppos); | |
2960 | } | |
2961 | ||
2962 | #ifdef CONFIG_NUMA | |
2963 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2964 | void __user *buffer, size_t *length, loff_t *ppos) | |
2965 | { | |
2966 | return hugetlb_sysctl_handler_common(true, table, write, | |
2967 | buffer, length, ppos); | |
2968 | } | |
2969 | #endif /* CONFIG_NUMA */ | |
2970 | ||
a3d0c6aa | 2971 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 2972 | void __user *buffer, |
a3d0c6aa NA |
2973 | size_t *length, loff_t *ppos) |
2974 | { | |
a5516438 | 2975 | struct hstate *h = &default_hstate; |
e5ff2159 | 2976 | unsigned long tmp; |
08d4a246 | 2977 | int ret; |
e5ff2159 | 2978 | |
457c1b27 | 2979 | if (!hugepages_supported()) |
86613628 | 2980 | return -EOPNOTSUPP; |
457c1b27 | 2981 | |
c033a93c | 2982 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 2983 | |
bae7f4ae | 2984 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
2985 | return -EINVAL; |
2986 | ||
e5ff2159 AK |
2987 | table->data = &tmp; |
2988 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2989 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2990 | if (ret) | |
2991 | goto out; | |
e5ff2159 AK |
2992 | |
2993 | if (write) { | |
2994 | spin_lock(&hugetlb_lock); | |
2995 | h->nr_overcommit_huge_pages = tmp; | |
2996 | spin_unlock(&hugetlb_lock); | |
2997 | } | |
08d4a246 MH |
2998 | out: |
2999 | return ret; | |
a3d0c6aa NA |
3000 | } |
3001 | ||
1da177e4 LT |
3002 | #endif /* CONFIG_SYSCTL */ |
3003 | ||
e1759c21 | 3004 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 3005 | { |
a5516438 | 3006 | struct hstate *h = &default_hstate; |
457c1b27 NA |
3007 | if (!hugepages_supported()) |
3008 | return; | |
e1759c21 | 3009 | seq_printf(m, |
4f98a2fe RR |
3010 | "HugePages_Total: %5lu\n" |
3011 | "HugePages_Free: %5lu\n" | |
3012 | "HugePages_Rsvd: %5lu\n" | |
3013 | "HugePages_Surp: %5lu\n" | |
3014 | "Hugepagesize: %8lu kB\n", | |
a5516438 AK |
3015 | h->nr_huge_pages, |
3016 | h->free_huge_pages, | |
3017 | h->resv_huge_pages, | |
3018 | h->surplus_huge_pages, | |
3019 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
3020 | } |
3021 | ||
3022 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
3023 | { | |
a5516438 | 3024 | struct hstate *h = &default_hstate; |
457c1b27 NA |
3025 | if (!hugepages_supported()) |
3026 | return 0; | |
1da177e4 LT |
3027 | return sprintf(buf, |
3028 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
3029 | "Node %d HugePages_Free: %5u\n" |
3030 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
3031 | nid, h->nr_huge_pages_node[nid], |
3032 | nid, h->free_huge_pages_node[nid], | |
3033 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
3034 | } |
3035 | ||
949f7ec5 DR |
3036 | void hugetlb_show_meminfo(void) |
3037 | { | |
3038 | struct hstate *h; | |
3039 | int nid; | |
3040 | ||
457c1b27 NA |
3041 | if (!hugepages_supported()) |
3042 | return; | |
3043 | ||
949f7ec5 DR |
3044 | for_each_node_state(nid, N_MEMORY) |
3045 | for_each_hstate(h) | |
3046 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
3047 | nid, | |
3048 | h->nr_huge_pages_node[nid], | |
3049 | h->free_huge_pages_node[nid], | |
3050 | h->surplus_huge_pages_node[nid], | |
3051 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
3052 | } | |
3053 | ||
5d317b2b NH |
3054 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
3055 | { | |
3056 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
3057 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
3058 | } | |
3059 | ||
1da177e4 LT |
3060 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
3061 | unsigned long hugetlb_total_pages(void) | |
3062 | { | |
d0028588 WL |
3063 | struct hstate *h; |
3064 | unsigned long nr_total_pages = 0; | |
3065 | ||
3066 | for_each_hstate(h) | |
3067 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
3068 | return nr_total_pages; | |
1da177e4 | 3069 | } |
1da177e4 | 3070 | |
a5516438 | 3071 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
3072 | { |
3073 | int ret = -ENOMEM; | |
3074 | ||
3075 | spin_lock(&hugetlb_lock); | |
3076 | /* | |
3077 | * When cpuset is configured, it breaks the strict hugetlb page | |
3078 | * reservation as the accounting is done on a global variable. Such | |
3079 | * reservation is completely rubbish in the presence of cpuset because | |
3080 | * the reservation is not checked against page availability for the | |
3081 | * current cpuset. Application can still potentially OOM'ed by kernel | |
3082 | * with lack of free htlb page in cpuset that the task is in. | |
3083 | * Attempt to enforce strict accounting with cpuset is almost | |
3084 | * impossible (or too ugly) because cpuset is too fluid that | |
3085 | * task or memory node can be dynamically moved between cpusets. | |
3086 | * | |
3087 | * The change of semantics for shared hugetlb mapping with cpuset is | |
3088 | * undesirable. However, in order to preserve some of the semantics, | |
3089 | * we fall back to check against current free page availability as | |
3090 | * a best attempt and hopefully to minimize the impact of changing | |
3091 | * semantics that cpuset has. | |
3092 | */ | |
3093 | if (delta > 0) { | |
a5516438 | 3094 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
3095 | goto out; |
3096 | ||
a5516438 AK |
3097 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
3098 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
3099 | goto out; |
3100 | } | |
3101 | } | |
3102 | ||
3103 | ret = 0; | |
3104 | if (delta < 0) | |
a5516438 | 3105 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
3106 | |
3107 | out: | |
3108 | spin_unlock(&hugetlb_lock); | |
3109 | return ret; | |
3110 | } | |
3111 | ||
84afd99b AW |
3112 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
3113 | { | |
f522c3ac | 3114 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
3115 | |
3116 | /* | |
3117 | * This new VMA should share its siblings reservation map if present. | |
3118 | * The VMA will only ever have a valid reservation map pointer where | |
3119 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 3120 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
3121 | * after this open call completes. It is therefore safe to take a |
3122 | * new reference here without additional locking. | |
3123 | */ | |
4e35f483 | 3124 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 3125 | kref_get(&resv->refs); |
84afd99b AW |
3126 | } |
3127 | ||
a1e78772 MG |
3128 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
3129 | { | |
a5516438 | 3130 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 3131 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 3132 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 3133 | unsigned long reserve, start, end; |
1c5ecae3 | 3134 | long gbl_reserve; |
84afd99b | 3135 | |
4e35f483 JK |
3136 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3137 | return; | |
84afd99b | 3138 | |
4e35f483 JK |
3139 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
3140 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 3141 | |
4e35f483 | 3142 | reserve = (end - start) - region_count(resv, start, end); |
84afd99b | 3143 | |
4e35f483 JK |
3144 | kref_put(&resv->refs, resv_map_release); |
3145 | ||
3146 | if (reserve) { | |
1c5ecae3 MK |
3147 | /* |
3148 | * Decrement reserve counts. The global reserve count may be | |
3149 | * adjusted if the subpool has a minimum size. | |
3150 | */ | |
3151 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
3152 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 3153 | } |
a1e78772 MG |
3154 | } |
3155 | ||
1da177e4 LT |
3156 | /* |
3157 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
3158 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
3159 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
3160 | * this far. | |
3161 | */ | |
11bac800 | 3162 | static int hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
3163 | { |
3164 | BUG(); | |
d0217ac0 | 3165 | return 0; |
1da177e4 LT |
3166 | } |
3167 | ||
f0f37e2f | 3168 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 3169 | .fault = hugetlb_vm_op_fault, |
84afd99b | 3170 | .open = hugetlb_vm_op_open, |
a1e78772 | 3171 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
3172 | }; |
3173 | ||
1e8f889b DG |
3174 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
3175 | int writable) | |
63551ae0 DG |
3176 | { |
3177 | pte_t entry; | |
3178 | ||
1e8f889b | 3179 | if (writable) { |
106c992a GS |
3180 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
3181 | vma->vm_page_prot))); | |
63551ae0 | 3182 | } else { |
106c992a GS |
3183 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
3184 | vma->vm_page_prot)); | |
63551ae0 DG |
3185 | } |
3186 | entry = pte_mkyoung(entry); | |
3187 | entry = pte_mkhuge(entry); | |
d9ed9faa | 3188 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
3189 | |
3190 | return entry; | |
3191 | } | |
3192 | ||
1e8f889b DG |
3193 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
3194 | unsigned long address, pte_t *ptep) | |
3195 | { | |
3196 | pte_t entry; | |
3197 | ||
106c992a | 3198 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 3199 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 3200 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
3201 | } |
3202 | ||
d5ed7444 | 3203 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
3204 | { |
3205 | swp_entry_t swp; | |
3206 | ||
3207 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 3208 | return false; |
4a705fef NH |
3209 | swp = pte_to_swp_entry(pte); |
3210 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
d5ed7444 | 3211 | return true; |
4a705fef | 3212 | else |
d5ed7444 | 3213 | return false; |
4a705fef NH |
3214 | } |
3215 | ||
3216 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
3217 | { | |
3218 | swp_entry_t swp; | |
3219 | ||
3220 | if (huge_pte_none(pte) || pte_present(pte)) | |
3221 | return 0; | |
3222 | swp = pte_to_swp_entry(pte); | |
3223 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
3224 | return 1; | |
3225 | else | |
3226 | return 0; | |
3227 | } | |
1e8f889b | 3228 | |
63551ae0 DG |
3229 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
3230 | struct vm_area_struct *vma) | |
3231 | { | |
3232 | pte_t *src_pte, *dst_pte, entry; | |
3233 | struct page *ptepage; | |
1c59827d | 3234 | unsigned long addr; |
1e8f889b | 3235 | int cow; |
a5516438 AK |
3236 | struct hstate *h = hstate_vma(vma); |
3237 | unsigned long sz = huge_page_size(h); | |
e8569dd2 AS |
3238 | unsigned long mmun_start; /* For mmu_notifiers */ |
3239 | unsigned long mmun_end; /* For mmu_notifiers */ | |
3240 | int ret = 0; | |
1e8f889b DG |
3241 | |
3242 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 3243 | |
e8569dd2 AS |
3244 | mmun_start = vma->vm_start; |
3245 | mmun_end = vma->vm_end; | |
3246 | if (cow) | |
3247 | mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); | |
3248 | ||
a5516438 | 3249 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 3250 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 3251 | src_pte = huge_pte_offset(src, addr, sz); |
c74df32c HD |
3252 | if (!src_pte) |
3253 | continue; | |
a5516438 | 3254 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
3255 | if (!dst_pte) { |
3256 | ret = -ENOMEM; | |
3257 | break; | |
3258 | } | |
c5c99429 LW |
3259 | |
3260 | /* If the pagetables are shared don't copy or take references */ | |
3261 | if (dst_pte == src_pte) | |
3262 | continue; | |
3263 | ||
cb900f41 KS |
3264 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
3265 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3266 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef NH |
3267 | entry = huge_ptep_get(src_pte); |
3268 | if (huge_pte_none(entry)) { /* skip none entry */ | |
3269 | ; | |
3270 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
3271 | is_hugetlb_entry_hwpoisoned(entry))) { | |
3272 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
3273 | ||
3274 | if (is_write_migration_entry(swp_entry) && cow) { | |
3275 | /* | |
3276 | * COW mappings require pages in both | |
3277 | * parent and child to be set to read. | |
3278 | */ | |
3279 | make_migration_entry_read(&swp_entry); | |
3280 | entry = swp_entry_to_pte(swp_entry); | |
e5251fd4 PA |
3281 | set_huge_swap_pte_at(src, addr, src_pte, |
3282 | entry, sz); | |
4a705fef | 3283 | } |
e5251fd4 | 3284 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); |
4a705fef | 3285 | } else { |
34ee645e | 3286 | if (cow) { |
7f2e9525 | 3287 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e JR |
3288 | mmu_notifier_invalidate_range(src, mmun_start, |
3289 | mmun_end); | |
3290 | } | |
0253d634 | 3291 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
3292 | ptepage = pte_page(entry); |
3293 | get_page(ptepage); | |
53f9263b | 3294 | page_dup_rmap(ptepage, true); |
1c59827d | 3295 | set_huge_pte_at(dst, addr, dst_pte, entry); |
5d317b2b | 3296 | hugetlb_count_add(pages_per_huge_page(h), dst); |
1c59827d | 3297 | } |
cb900f41 KS |
3298 | spin_unlock(src_ptl); |
3299 | spin_unlock(dst_ptl); | |
63551ae0 | 3300 | } |
63551ae0 | 3301 | |
e8569dd2 AS |
3302 | if (cow) |
3303 | mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); | |
3304 | ||
3305 | return ret; | |
63551ae0 DG |
3306 | } |
3307 | ||
24669e58 AK |
3308 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
3309 | unsigned long start, unsigned long end, | |
3310 | struct page *ref_page) | |
63551ae0 DG |
3311 | { |
3312 | struct mm_struct *mm = vma->vm_mm; | |
3313 | unsigned long address; | |
c7546f8f | 3314 | pte_t *ptep; |
63551ae0 | 3315 | pte_t pte; |
cb900f41 | 3316 | spinlock_t *ptl; |
63551ae0 | 3317 | struct page *page; |
a5516438 AK |
3318 | struct hstate *h = hstate_vma(vma); |
3319 | unsigned long sz = huge_page_size(h); | |
2ec74c3e SG |
3320 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
3321 | const unsigned long mmun_end = end; /* For mmu_notifiers */ | |
a5516438 | 3322 | |
63551ae0 | 3323 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
3324 | BUG_ON(start & ~huge_page_mask(h)); |
3325 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 3326 | |
07e32661 AK |
3327 | /* |
3328 | * This is a hugetlb vma, all the pte entries should point | |
3329 | * to huge page. | |
3330 | */ | |
3331 | tlb_remove_check_page_size_change(tlb, sz); | |
24669e58 | 3332 | tlb_start_vma(tlb, vma); |
2ec74c3e | 3333 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
569f48b8 | 3334 | address = start; |
569f48b8 | 3335 | for (; address < end; address += sz) { |
7868a208 | 3336 | ptep = huge_pte_offset(mm, address, sz); |
4c887265 | 3337 | if (!ptep) |
c7546f8f DG |
3338 | continue; |
3339 | ||
cb900f41 | 3340 | ptl = huge_pte_lock(h, mm, ptep); |
31d49da5 AK |
3341 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3342 | spin_unlock(ptl); | |
3343 | continue; | |
3344 | } | |
39dde65c | 3345 | |
6629326b | 3346 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
3347 | if (huge_pte_none(pte)) { |
3348 | spin_unlock(ptl); | |
3349 | continue; | |
3350 | } | |
6629326b HD |
3351 | |
3352 | /* | |
9fbc1f63 NH |
3353 | * Migrating hugepage or HWPoisoned hugepage is already |
3354 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 3355 | */ |
9fbc1f63 | 3356 | if (unlikely(!pte_present(pte))) { |
9386fac3 | 3357 | huge_pte_clear(mm, address, ptep, sz); |
31d49da5 AK |
3358 | spin_unlock(ptl); |
3359 | continue; | |
8c4894c6 | 3360 | } |
6629326b HD |
3361 | |
3362 | page = pte_page(pte); | |
04f2cbe3 MG |
3363 | /* |
3364 | * If a reference page is supplied, it is because a specific | |
3365 | * page is being unmapped, not a range. Ensure the page we | |
3366 | * are about to unmap is the actual page of interest. | |
3367 | */ | |
3368 | if (ref_page) { | |
31d49da5 AK |
3369 | if (page != ref_page) { |
3370 | spin_unlock(ptl); | |
3371 | continue; | |
3372 | } | |
04f2cbe3 MG |
3373 | /* |
3374 | * Mark the VMA as having unmapped its page so that | |
3375 | * future faults in this VMA will fail rather than | |
3376 | * looking like data was lost | |
3377 | */ | |
3378 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
3379 | } | |
3380 | ||
c7546f8f | 3381 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 3382 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 3383 | if (huge_pte_dirty(pte)) |
6649a386 | 3384 | set_page_dirty(page); |
9e81130b | 3385 | |
5d317b2b | 3386 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
d281ee61 | 3387 | page_remove_rmap(page, true); |
31d49da5 | 3388 | |
cb900f41 | 3389 | spin_unlock(ptl); |
e77b0852 | 3390 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
3391 | /* |
3392 | * Bail out after unmapping reference page if supplied | |
3393 | */ | |
3394 | if (ref_page) | |
3395 | break; | |
fe1668ae | 3396 | } |
2ec74c3e | 3397 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
24669e58 | 3398 | tlb_end_vma(tlb, vma); |
1da177e4 | 3399 | } |
63551ae0 | 3400 | |
d833352a MG |
3401 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
3402 | struct vm_area_struct *vma, unsigned long start, | |
3403 | unsigned long end, struct page *ref_page) | |
3404 | { | |
3405 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
3406 | ||
3407 | /* | |
3408 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
3409 | * test will fail on a vma being torn down, and not grab a page table | |
3410 | * on its way out. We're lucky that the flag has such an appropriate | |
3411 | * name, and can in fact be safely cleared here. We could clear it | |
3412 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 3413 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 3414 | * because in the context this is called, the VMA is about to be |
c8c06efa | 3415 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
3416 | */ |
3417 | vma->vm_flags &= ~VM_MAYSHARE; | |
3418 | } | |
3419 | ||
502717f4 | 3420 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 3421 | unsigned long end, struct page *ref_page) |
502717f4 | 3422 | { |
24669e58 AK |
3423 | struct mm_struct *mm; |
3424 | struct mmu_gather tlb; | |
3425 | ||
3426 | mm = vma->vm_mm; | |
3427 | ||
2b047252 | 3428 | tlb_gather_mmu(&tlb, mm, start, end); |
24669e58 AK |
3429 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
3430 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 CK |
3431 | } |
3432 | ||
04f2cbe3 MG |
3433 | /* |
3434 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
3435 | * mappping it owns the reserve page for. The intention is to unmap the page | |
3436 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
3437 | * same region. | |
3438 | */ | |
2f4612af DB |
3439 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
3440 | struct page *page, unsigned long address) | |
04f2cbe3 | 3441 | { |
7526674d | 3442 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
3443 | struct vm_area_struct *iter_vma; |
3444 | struct address_space *mapping; | |
04f2cbe3 MG |
3445 | pgoff_t pgoff; |
3446 | ||
3447 | /* | |
3448 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
3449 | * from page cache lookup which is in HPAGE_SIZE units. | |
3450 | */ | |
7526674d | 3451 | address = address & huge_page_mask(h); |
36e4f20a MH |
3452 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
3453 | vma->vm_pgoff; | |
93c76a3d | 3454 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 3455 | |
4eb2b1dc MG |
3456 | /* |
3457 | * Take the mapping lock for the duration of the table walk. As | |
3458 | * this mapping should be shared between all the VMAs, | |
3459 | * __unmap_hugepage_range() is called as the lock is already held | |
3460 | */ | |
83cde9e8 | 3461 | i_mmap_lock_write(mapping); |
6b2dbba8 | 3462 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
3463 | /* Do not unmap the current VMA */ |
3464 | if (iter_vma == vma) | |
3465 | continue; | |
3466 | ||
2f84a899 MG |
3467 | /* |
3468 | * Shared VMAs have their own reserves and do not affect | |
3469 | * MAP_PRIVATE accounting but it is possible that a shared | |
3470 | * VMA is using the same page so check and skip such VMAs. | |
3471 | */ | |
3472 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
3473 | continue; | |
3474 | ||
04f2cbe3 MG |
3475 | /* |
3476 | * Unmap the page from other VMAs without their own reserves. | |
3477 | * They get marked to be SIGKILLed if they fault in these | |
3478 | * areas. This is because a future no-page fault on this VMA | |
3479 | * could insert a zeroed page instead of the data existing | |
3480 | * from the time of fork. This would look like data corruption | |
3481 | */ | |
3482 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
3483 | unmap_hugepage_range(iter_vma, address, |
3484 | address + huge_page_size(h), page); | |
04f2cbe3 | 3485 | } |
83cde9e8 | 3486 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
3487 | } |
3488 | ||
0fe6e20b NH |
3489 | /* |
3490 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
3491 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
3492 | * cannot race with other handlers or page migration. | |
3493 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 3494 | */ |
1e8f889b | 3495 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
3999f52e AK |
3496 | unsigned long address, pte_t *ptep, |
3497 | struct page *pagecache_page, spinlock_t *ptl) | |
1e8f889b | 3498 | { |
3999f52e | 3499 | pte_t pte; |
a5516438 | 3500 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 3501 | struct page *old_page, *new_page; |
ad4404a2 | 3502 | int ret = 0, outside_reserve = 0; |
2ec74c3e SG |
3503 | unsigned long mmun_start; /* For mmu_notifiers */ |
3504 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1e8f889b | 3505 | |
3999f52e | 3506 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
3507 | old_page = pte_page(pte); |
3508 | ||
04f2cbe3 | 3509 | retry_avoidcopy: |
1e8f889b DG |
3510 | /* If no-one else is actually using this page, avoid the copy |
3511 | * and just make the page writable */ | |
37a2140d | 3512 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
5a49973d | 3513 | page_move_anon_rmap(old_page, vma); |
1e8f889b | 3514 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 3515 | return 0; |
1e8f889b DG |
3516 | } |
3517 | ||
04f2cbe3 MG |
3518 | /* |
3519 | * If the process that created a MAP_PRIVATE mapping is about to | |
3520 | * perform a COW due to a shared page count, attempt to satisfy | |
3521 | * the allocation without using the existing reserves. The pagecache | |
3522 | * page is used to determine if the reserve at this address was | |
3523 | * consumed or not. If reserves were used, a partial faulted mapping | |
3524 | * at the time of fork() could consume its reserves on COW instead | |
3525 | * of the full address range. | |
3526 | */ | |
5944d011 | 3527 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
3528 | old_page != pagecache_page) |
3529 | outside_reserve = 1; | |
3530 | ||
09cbfeaf | 3531 | get_page(old_page); |
b76c8cfb | 3532 | |
ad4404a2 DB |
3533 | /* |
3534 | * Drop page table lock as buddy allocator may be called. It will | |
3535 | * be acquired again before returning to the caller, as expected. | |
3536 | */ | |
cb900f41 | 3537 | spin_unlock(ptl); |
04f2cbe3 | 3538 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 3539 | |
2fc39cec | 3540 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
3541 | /* |
3542 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
3543 | * it is due to references held by a child and an insufficient | |
3544 | * huge page pool. To guarantee the original mappers | |
3545 | * reliability, unmap the page from child processes. The child | |
3546 | * may get SIGKILLed if it later faults. | |
3547 | */ | |
3548 | if (outside_reserve) { | |
09cbfeaf | 3549 | put_page(old_page); |
04f2cbe3 | 3550 | BUG_ON(huge_pte_none(pte)); |
2f4612af DB |
3551 | unmap_ref_private(mm, vma, old_page, address); |
3552 | BUG_ON(huge_pte_none(pte)); | |
3553 | spin_lock(ptl); | |
7868a208 PA |
3554 | ptep = huge_pte_offset(mm, address & huge_page_mask(h), |
3555 | huge_page_size(h)); | |
2f4612af DB |
3556 | if (likely(ptep && |
3557 | pte_same(huge_ptep_get(ptep), pte))) | |
3558 | goto retry_avoidcopy; | |
3559 | /* | |
3560 | * race occurs while re-acquiring page table | |
3561 | * lock, and our job is done. | |
3562 | */ | |
3563 | return 0; | |
04f2cbe3 MG |
3564 | } |
3565 | ||
ad4404a2 DB |
3566 | ret = (PTR_ERR(new_page) == -ENOMEM) ? |
3567 | VM_FAULT_OOM : VM_FAULT_SIGBUS; | |
3568 | goto out_release_old; | |
1e8f889b DG |
3569 | } |
3570 | ||
0fe6e20b NH |
3571 | /* |
3572 | * When the original hugepage is shared one, it does not have | |
3573 | * anon_vma prepared. | |
3574 | */ | |
44e2aa93 | 3575 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
3576 | ret = VM_FAULT_OOM; |
3577 | goto out_release_all; | |
44e2aa93 | 3578 | } |
0fe6e20b | 3579 | |
47ad8475 AA |
3580 | copy_user_huge_page(new_page, old_page, address, vma, |
3581 | pages_per_huge_page(h)); | |
0ed361de | 3582 | __SetPageUptodate(new_page); |
bcc54222 | 3583 | set_page_huge_active(new_page); |
1e8f889b | 3584 | |
2ec74c3e SG |
3585 | mmun_start = address & huge_page_mask(h); |
3586 | mmun_end = mmun_start + huge_page_size(h); | |
3587 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
ad4404a2 | 3588 | |
b76c8cfb | 3589 | /* |
cb900f41 | 3590 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
3591 | * before the page tables are altered |
3592 | */ | |
cb900f41 | 3593 | spin_lock(ptl); |
7868a208 PA |
3594 | ptep = huge_pte_offset(mm, address & huge_page_mask(h), |
3595 | huge_page_size(h)); | |
a9af0c5d | 3596 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
3597 | ClearPagePrivate(new_page); |
3598 | ||
1e8f889b | 3599 | /* Break COW */ |
8fe627ec | 3600 | huge_ptep_clear_flush(vma, address, ptep); |
34ee645e | 3601 | mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); |
1e8f889b DG |
3602 | set_huge_pte_at(mm, address, ptep, |
3603 | make_huge_pte(vma, new_page, 1)); | |
d281ee61 | 3604 | page_remove_rmap(old_page, true); |
cd67f0d2 | 3605 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
3606 | /* Make the old page be freed below */ |
3607 | new_page = old_page; | |
3608 | } | |
cb900f41 | 3609 | spin_unlock(ptl); |
2ec74c3e | 3610 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
ad4404a2 | 3611 | out_release_all: |
96b96a96 | 3612 | restore_reserve_on_error(h, vma, address, new_page); |
09cbfeaf | 3613 | put_page(new_page); |
ad4404a2 | 3614 | out_release_old: |
09cbfeaf | 3615 | put_page(old_page); |
8312034f | 3616 | |
ad4404a2 DB |
3617 | spin_lock(ptl); /* Caller expects lock to be held */ |
3618 | return ret; | |
1e8f889b DG |
3619 | } |
3620 | ||
04f2cbe3 | 3621 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
3622 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
3623 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
3624 | { |
3625 | struct address_space *mapping; | |
e7c4b0bf | 3626 | pgoff_t idx; |
04f2cbe3 MG |
3627 | |
3628 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 3629 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
3630 | |
3631 | return find_lock_page(mapping, idx); | |
3632 | } | |
3633 | ||
3ae77f43 HD |
3634 | /* |
3635 | * Return whether there is a pagecache page to back given address within VMA. | |
3636 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
3637 | */ | |
3638 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
3639 | struct vm_area_struct *vma, unsigned long address) |
3640 | { | |
3641 | struct address_space *mapping; | |
3642 | pgoff_t idx; | |
3643 | struct page *page; | |
3644 | ||
3645 | mapping = vma->vm_file->f_mapping; | |
3646 | idx = vma_hugecache_offset(h, vma, address); | |
3647 | ||
3648 | page = find_get_page(mapping, idx); | |
3649 | if (page) | |
3650 | put_page(page); | |
3651 | return page != NULL; | |
3652 | } | |
3653 | ||
ab76ad54 MK |
3654 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
3655 | pgoff_t idx) | |
3656 | { | |
3657 | struct inode *inode = mapping->host; | |
3658 | struct hstate *h = hstate_inode(inode); | |
3659 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
3660 | ||
3661 | if (err) | |
3662 | return err; | |
3663 | ClearPagePrivate(page); | |
3664 | ||
3665 | spin_lock(&inode->i_lock); | |
3666 | inode->i_blocks += blocks_per_huge_page(h); | |
3667 | spin_unlock(&inode->i_lock); | |
3668 | return 0; | |
3669 | } | |
3670 | ||
a1ed3dda | 3671 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
8382d914 DB |
3672 | struct address_space *mapping, pgoff_t idx, |
3673 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 3674 | { |
a5516438 | 3675 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 3676 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 3677 | int anon_rmap = 0; |
4c887265 | 3678 | unsigned long size; |
4c887265 | 3679 | struct page *page; |
1e8f889b | 3680 | pte_t new_pte; |
cb900f41 | 3681 | spinlock_t *ptl; |
4c887265 | 3682 | |
04f2cbe3 MG |
3683 | /* |
3684 | * Currently, we are forced to kill the process in the event the | |
3685 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 3686 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
3687 | */ |
3688 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 3689 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 3690 | current->pid); |
04f2cbe3 MG |
3691 | return ret; |
3692 | } | |
3693 | ||
4c887265 AL |
3694 | /* |
3695 | * Use page lock to guard against racing truncation | |
3696 | * before we get page_table_lock. | |
3697 | */ | |
6bda666a CL |
3698 | retry: |
3699 | page = find_lock_page(mapping, idx); | |
3700 | if (!page) { | |
a5516438 | 3701 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
3702 | if (idx >= size) |
3703 | goto out; | |
1a1aad8a MK |
3704 | |
3705 | /* | |
3706 | * Check for page in userfault range | |
3707 | */ | |
3708 | if (userfaultfd_missing(vma)) { | |
3709 | u32 hash; | |
3710 | struct vm_fault vmf = { | |
3711 | .vma = vma, | |
3712 | .address = address, | |
3713 | .flags = flags, | |
3714 | /* | |
3715 | * Hard to debug if it ends up being | |
3716 | * used by a callee that assumes | |
3717 | * something about the other | |
3718 | * uninitialized fields... same as in | |
3719 | * memory.c | |
3720 | */ | |
3721 | }; | |
3722 | ||
3723 | /* | |
3724 | * hugetlb_fault_mutex must be dropped before | |
3725 | * handling userfault. Reacquire after handling | |
3726 | * fault to make calling code simpler. | |
3727 | */ | |
3728 | hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, | |
3729 | idx, address); | |
3730 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
3731 | ret = handle_userfault(&vmf, VM_UFFD_MISSING); | |
3732 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
3733 | goto out; | |
3734 | } | |
3735 | ||
04f2cbe3 | 3736 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 3737 | if (IS_ERR(page)) { |
76dcee75 AK |
3738 | ret = PTR_ERR(page); |
3739 | if (ret == -ENOMEM) | |
3740 | ret = VM_FAULT_OOM; | |
3741 | else | |
3742 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
3743 | goto out; |
3744 | } | |
47ad8475 | 3745 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 3746 | __SetPageUptodate(page); |
bcc54222 | 3747 | set_page_huge_active(page); |
ac9b9c66 | 3748 | |
f83a275d | 3749 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 3750 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
3751 | if (err) { |
3752 | put_page(page); | |
6bda666a CL |
3753 | if (err == -EEXIST) |
3754 | goto retry; | |
3755 | goto out; | |
3756 | } | |
23be7468 | 3757 | } else { |
6bda666a | 3758 | lock_page(page); |
0fe6e20b NH |
3759 | if (unlikely(anon_vma_prepare(vma))) { |
3760 | ret = VM_FAULT_OOM; | |
3761 | goto backout_unlocked; | |
3762 | } | |
409eb8c2 | 3763 | anon_rmap = 1; |
23be7468 | 3764 | } |
0fe6e20b | 3765 | } else { |
998b4382 NH |
3766 | /* |
3767 | * If memory error occurs between mmap() and fault, some process | |
3768 | * don't have hwpoisoned swap entry for errored virtual address. | |
3769 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
3770 | */ | |
3771 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 3772 | ret = VM_FAULT_HWPOISON | |
972dc4de | 3773 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
3774 | goto backout_unlocked; |
3775 | } | |
6bda666a | 3776 | } |
1e8f889b | 3777 | |
57303d80 AW |
3778 | /* |
3779 | * If we are going to COW a private mapping later, we examine the | |
3780 | * pending reservations for this page now. This will ensure that | |
3781 | * any allocations necessary to record that reservation occur outside | |
3782 | * the spinlock. | |
3783 | */ | |
5e911373 | 3784 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
2b26736c AW |
3785 | if (vma_needs_reservation(h, vma, address) < 0) { |
3786 | ret = VM_FAULT_OOM; | |
3787 | goto backout_unlocked; | |
3788 | } | |
5e911373 | 3789 | /* Just decrements count, does not deallocate */ |
feba16e2 | 3790 | vma_end_reservation(h, vma, address); |
5e911373 | 3791 | } |
57303d80 | 3792 | |
8bea8052 | 3793 | ptl = huge_pte_lock(h, mm, ptep); |
a5516438 | 3794 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
3795 | if (idx >= size) |
3796 | goto backout; | |
3797 | ||
83c54070 | 3798 | ret = 0; |
7f2e9525 | 3799 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
3800 | goto backout; |
3801 | ||
07443a85 JK |
3802 | if (anon_rmap) { |
3803 | ClearPagePrivate(page); | |
409eb8c2 | 3804 | hugepage_add_new_anon_rmap(page, vma, address); |
ac714904 | 3805 | } else |
53f9263b | 3806 | page_dup_rmap(page, true); |
1e8f889b DG |
3807 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
3808 | && (vma->vm_flags & VM_SHARED))); | |
3809 | set_huge_pte_at(mm, address, ptep, new_pte); | |
3810 | ||
5d317b2b | 3811 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 3812 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 3813 | /* Optimization, do the COW without a second fault */ |
3999f52e | 3814 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); |
1e8f889b DG |
3815 | } |
3816 | ||
cb900f41 | 3817 | spin_unlock(ptl); |
4c887265 AL |
3818 | unlock_page(page); |
3819 | out: | |
ac9b9c66 | 3820 | return ret; |
4c887265 AL |
3821 | |
3822 | backout: | |
cb900f41 | 3823 | spin_unlock(ptl); |
2b26736c | 3824 | backout_unlocked: |
4c887265 | 3825 | unlock_page(page); |
96b96a96 | 3826 | restore_reserve_on_error(h, vma, address, page); |
4c887265 AL |
3827 | put_page(page); |
3828 | goto out; | |
ac9b9c66 HD |
3829 | } |
3830 | ||
8382d914 | 3831 | #ifdef CONFIG_SMP |
c672c7f2 | 3832 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
8382d914 DB |
3833 | struct vm_area_struct *vma, |
3834 | struct address_space *mapping, | |
3835 | pgoff_t idx, unsigned long address) | |
3836 | { | |
3837 | unsigned long key[2]; | |
3838 | u32 hash; | |
3839 | ||
3840 | if (vma->vm_flags & VM_SHARED) { | |
3841 | key[0] = (unsigned long) mapping; | |
3842 | key[1] = idx; | |
3843 | } else { | |
3844 | key[0] = (unsigned long) mm; | |
3845 | key[1] = address >> huge_page_shift(h); | |
3846 | } | |
3847 | ||
3848 | hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); | |
3849 | ||
3850 | return hash & (num_fault_mutexes - 1); | |
3851 | } | |
3852 | #else | |
3853 | /* | |
3854 | * For uniprocesor systems we always use a single mutex, so just | |
3855 | * return 0 and avoid the hashing overhead. | |
3856 | */ | |
c672c7f2 | 3857 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
8382d914 DB |
3858 | struct vm_area_struct *vma, |
3859 | struct address_space *mapping, | |
3860 | pgoff_t idx, unsigned long address) | |
3861 | { | |
3862 | return 0; | |
3863 | } | |
3864 | #endif | |
3865 | ||
86e5216f | 3866 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 3867 | unsigned long address, unsigned int flags) |
86e5216f | 3868 | { |
8382d914 | 3869 | pte_t *ptep, entry; |
cb900f41 | 3870 | spinlock_t *ptl; |
1e8f889b | 3871 | int ret; |
8382d914 DB |
3872 | u32 hash; |
3873 | pgoff_t idx; | |
0fe6e20b | 3874 | struct page *page = NULL; |
57303d80 | 3875 | struct page *pagecache_page = NULL; |
a5516438 | 3876 | struct hstate *h = hstate_vma(vma); |
8382d914 | 3877 | struct address_space *mapping; |
0f792cf9 | 3878 | int need_wait_lock = 0; |
86e5216f | 3879 | |
1e16a539 KH |
3880 | address &= huge_page_mask(h); |
3881 | ||
7868a208 | 3882 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
fd6a03ed NH |
3883 | if (ptep) { |
3884 | entry = huge_ptep_get(ptep); | |
290408d4 | 3885 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 3886 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
3887 | return 0; |
3888 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 3889 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 3890 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
0d777df5 NH |
3891 | } else { |
3892 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); | |
3893 | if (!ptep) | |
3894 | return VM_FAULT_OOM; | |
fd6a03ed NH |
3895 | } |
3896 | ||
8382d914 DB |
3897 | mapping = vma->vm_file->f_mapping; |
3898 | idx = vma_hugecache_offset(h, vma, address); | |
3899 | ||
3935baa9 DG |
3900 | /* |
3901 | * Serialize hugepage allocation and instantiation, so that we don't | |
3902 | * get spurious allocation failures if two CPUs race to instantiate | |
3903 | * the same page in the page cache. | |
3904 | */ | |
c672c7f2 MK |
3905 | hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address); |
3906 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
8382d914 | 3907 | |
7f2e9525 GS |
3908 | entry = huge_ptep_get(ptep); |
3909 | if (huge_pte_none(entry)) { | |
8382d914 | 3910 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 3911 | goto out_mutex; |
3935baa9 | 3912 | } |
86e5216f | 3913 | |
83c54070 | 3914 | ret = 0; |
1e8f889b | 3915 | |
0f792cf9 NH |
3916 | /* |
3917 | * entry could be a migration/hwpoison entry at this point, so this | |
3918 | * check prevents the kernel from going below assuming that we have | |
3919 | * a active hugepage in pagecache. This goto expects the 2nd page fault, | |
3920 | * and is_hugetlb_entry_(migration|hwpoisoned) check will properly | |
3921 | * handle it. | |
3922 | */ | |
3923 | if (!pte_present(entry)) | |
3924 | goto out_mutex; | |
3925 | ||
57303d80 AW |
3926 | /* |
3927 | * If we are going to COW the mapping later, we examine the pending | |
3928 | * reservations for this page now. This will ensure that any | |
3929 | * allocations necessary to record that reservation occur outside the | |
3930 | * spinlock. For private mappings, we also lookup the pagecache | |
3931 | * page now as it is used to determine if a reservation has been | |
3932 | * consumed. | |
3933 | */ | |
106c992a | 3934 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
2b26736c AW |
3935 | if (vma_needs_reservation(h, vma, address) < 0) { |
3936 | ret = VM_FAULT_OOM; | |
b4d1d99f | 3937 | goto out_mutex; |
2b26736c | 3938 | } |
5e911373 | 3939 | /* Just decrements count, does not deallocate */ |
feba16e2 | 3940 | vma_end_reservation(h, vma, address); |
57303d80 | 3941 | |
f83a275d | 3942 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
3943 | pagecache_page = hugetlbfs_pagecache_page(h, |
3944 | vma, address); | |
3945 | } | |
3946 | ||
0f792cf9 NH |
3947 | ptl = huge_pte_lock(h, mm, ptep); |
3948 | ||
3949 | /* Check for a racing update before calling hugetlb_cow */ | |
3950 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
3951 | goto out_ptl; | |
3952 | ||
56c9cfb1 NH |
3953 | /* |
3954 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
3955 | * pagecache_page, so here we need take the former one | |
3956 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
3957 | */ |
3958 | page = pte_page(entry); | |
3959 | if (page != pagecache_page) | |
0f792cf9 NH |
3960 | if (!trylock_page(page)) { |
3961 | need_wait_lock = 1; | |
3962 | goto out_ptl; | |
3963 | } | |
b4d1d99f | 3964 | |
0f792cf9 | 3965 | get_page(page); |
b4d1d99f | 3966 | |
788c7df4 | 3967 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 3968 | if (!huge_pte_write(entry)) { |
3999f52e AK |
3969 | ret = hugetlb_cow(mm, vma, address, ptep, |
3970 | pagecache_page, ptl); | |
0f792cf9 | 3971 | goto out_put_page; |
b4d1d99f | 3972 | } |
106c992a | 3973 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
3974 | } |
3975 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
3976 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
3977 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 3978 | update_mmu_cache(vma, address, ptep); |
0f792cf9 NH |
3979 | out_put_page: |
3980 | if (page != pagecache_page) | |
3981 | unlock_page(page); | |
3982 | put_page(page); | |
cb900f41 KS |
3983 | out_ptl: |
3984 | spin_unlock(ptl); | |
57303d80 AW |
3985 | |
3986 | if (pagecache_page) { | |
3987 | unlock_page(pagecache_page); | |
3988 | put_page(pagecache_page); | |
3989 | } | |
b4d1d99f | 3990 | out_mutex: |
c672c7f2 | 3991 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
0f792cf9 NH |
3992 | /* |
3993 | * Generally it's safe to hold refcount during waiting page lock. But | |
3994 | * here we just wait to defer the next page fault to avoid busy loop and | |
3995 | * the page is not used after unlocked before returning from the current | |
3996 | * page fault. So we are safe from accessing freed page, even if we wait | |
3997 | * here without taking refcount. | |
3998 | */ | |
3999 | if (need_wait_lock) | |
4000 | wait_on_page_locked(page); | |
1e8f889b | 4001 | return ret; |
86e5216f AL |
4002 | } |
4003 | ||
8fb5debc MK |
4004 | /* |
4005 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
4006 | * modifications for huge pages. | |
4007 | */ | |
4008 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
4009 | pte_t *dst_pte, | |
4010 | struct vm_area_struct *dst_vma, | |
4011 | unsigned long dst_addr, | |
4012 | unsigned long src_addr, | |
4013 | struct page **pagep) | |
4014 | { | |
1c9e8def | 4015 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
4016 | struct hstate *h = hstate_vma(dst_vma); |
4017 | pte_t _dst_pte; | |
4018 | spinlock_t *ptl; | |
4019 | int ret; | |
4020 | struct page *page; | |
4021 | ||
4022 | if (!*pagep) { | |
4023 | ret = -ENOMEM; | |
4024 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
4025 | if (IS_ERR(page)) | |
4026 | goto out; | |
4027 | ||
4028 | ret = copy_huge_page_from_user(page, | |
4029 | (const void __user *) src_addr, | |
810a56b9 | 4030 | pages_per_huge_page(h), false); |
8fb5debc MK |
4031 | |
4032 | /* fallback to copy_from_user outside mmap_sem */ | |
4033 | if (unlikely(ret)) { | |
4034 | ret = -EFAULT; | |
4035 | *pagep = page; | |
4036 | /* don't free the page */ | |
4037 | goto out; | |
4038 | } | |
4039 | } else { | |
4040 | page = *pagep; | |
4041 | *pagep = NULL; | |
4042 | } | |
4043 | ||
4044 | /* | |
4045 | * The memory barrier inside __SetPageUptodate makes sure that | |
4046 | * preceding stores to the page contents become visible before | |
4047 | * the set_pte_at() write. | |
4048 | */ | |
4049 | __SetPageUptodate(page); | |
4050 | set_page_huge_active(page); | |
4051 | ||
1c9e8def MK |
4052 | /* |
4053 | * If shared, add to page cache | |
4054 | */ | |
4055 | if (vm_shared) { | |
4056 | struct address_space *mapping = dst_vma->vm_file->f_mapping; | |
4057 | pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
4058 | ||
4059 | ret = huge_add_to_page_cache(page, mapping, idx); | |
4060 | if (ret) | |
4061 | goto out_release_nounlock; | |
4062 | } | |
4063 | ||
8fb5debc MK |
4064 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); |
4065 | spin_lock(ptl); | |
4066 | ||
4067 | ret = -EEXIST; | |
4068 | if (!huge_pte_none(huge_ptep_get(dst_pte))) | |
4069 | goto out_release_unlock; | |
4070 | ||
1c9e8def MK |
4071 | if (vm_shared) { |
4072 | page_dup_rmap(page, true); | |
4073 | } else { | |
4074 | ClearPagePrivate(page); | |
4075 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); | |
4076 | } | |
8fb5debc MK |
4077 | |
4078 | _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); | |
4079 | if (dst_vma->vm_flags & VM_WRITE) | |
4080 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
4081 | _dst_pte = pte_mkyoung(_dst_pte); | |
4082 | ||
4083 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); | |
4084 | ||
4085 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
4086 | dst_vma->vm_flags & VM_WRITE); | |
4087 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
4088 | ||
4089 | /* No need to invalidate - it was non-present before */ | |
4090 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
4091 | ||
4092 | spin_unlock(ptl); | |
1c9e8def MK |
4093 | if (vm_shared) |
4094 | unlock_page(page); | |
8fb5debc MK |
4095 | ret = 0; |
4096 | out: | |
4097 | return ret; | |
4098 | out_release_unlock: | |
4099 | spin_unlock(ptl); | |
1c9e8def MK |
4100 | out_release_nounlock: |
4101 | if (vm_shared) | |
4102 | unlock_page(page); | |
8fb5debc MK |
4103 | put_page(page); |
4104 | goto out; | |
4105 | } | |
4106 | ||
28a35716 ML |
4107 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
4108 | struct page **pages, struct vm_area_struct **vmas, | |
4109 | unsigned long *position, unsigned long *nr_pages, | |
87ffc118 | 4110 | long i, unsigned int flags, int *nonblocking) |
63551ae0 | 4111 | { |
d5d4b0aa CK |
4112 | unsigned long pfn_offset; |
4113 | unsigned long vaddr = *position; | |
28a35716 | 4114 | unsigned long remainder = *nr_pages; |
a5516438 | 4115 | struct hstate *h = hstate_vma(vma); |
63551ae0 | 4116 | |
63551ae0 | 4117 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 4118 | pte_t *pte; |
cb900f41 | 4119 | spinlock_t *ptl = NULL; |
2a15efc9 | 4120 | int absent; |
4c887265 | 4121 | struct page *page; |
63551ae0 | 4122 | |
02057967 DR |
4123 | /* |
4124 | * If we have a pending SIGKILL, don't keep faulting pages and | |
4125 | * potentially allocating memory. | |
4126 | */ | |
4127 | if (unlikely(fatal_signal_pending(current))) { | |
4128 | remainder = 0; | |
4129 | break; | |
4130 | } | |
4131 | ||
4c887265 AL |
4132 | /* |
4133 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 4134 | * each hugepage. We have to make sure we get the |
4c887265 | 4135 | * first, for the page indexing below to work. |
cb900f41 KS |
4136 | * |
4137 | * Note that page table lock is not held when pte is null. | |
4c887265 | 4138 | */ |
7868a208 PA |
4139 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
4140 | huge_page_size(h)); | |
cb900f41 KS |
4141 | if (pte) |
4142 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
4143 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
4144 | ||
4145 | /* | |
4146 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
4147 | * an error where there's an empty slot with no huge pagecache |
4148 | * to back it. This way, we avoid allocating a hugepage, and | |
4149 | * the sparse dumpfile avoids allocating disk blocks, but its | |
4150 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 4151 | */ |
3ae77f43 HD |
4152 | if (absent && (flags & FOLL_DUMP) && |
4153 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
4154 | if (pte) |
4155 | spin_unlock(ptl); | |
2a15efc9 HD |
4156 | remainder = 0; |
4157 | break; | |
4158 | } | |
63551ae0 | 4159 | |
9cc3a5bd NH |
4160 | /* |
4161 | * We need call hugetlb_fault for both hugepages under migration | |
4162 | * (in which case hugetlb_fault waits for the migration,) and | |
4163 | * hwpoisoned hugepages (in which case we need to prevent the | |
4164 | * caller from accessing to them.) In order to do this, we use | |
4165 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
4166 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
4167 | * both cases, and because we can't follow correct pages | |
4168 | * directly from any kind of swap entries. | |
4169 | */ | |
4170 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
4171 | ((flags & FOLL_WRITE) && |
4172 | !huge_pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 4173 | int ret; |
87ffc118 | 4174 | unsigned int fault_flags = 0; |
63551ae0 | 4175 | |
cb900f41 KS |
4176 | if (pte) |
4177 | spin_unlock(ptl); | |
87ffc118 AA |
4178 | if (flags & FOLL_WRITE) |
4179 | fault_flags |= FAULT_FLAG_WRITE; | |
4180 | if (nonblocking) | |
4181 | fault_flags |= FAULT_FLAG_ALLOW_RETRY; | |
4182 | if (flags & FOLL_NOWAIT) | |
4183 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
4184 | FAULT_FLAG_RETRY_NOWAIT; | |
4185 | if (flags & FOLL_TRIED) { | |
4186 | VM_WARN_ON_ONCE(fault_flags & | |
4187 | FAULT_FLAG_ALLOW_RETRY); | |
4188 | fault_flags |= FAULT_FLAG_TRIED; | |
4189 | } | |
4190 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
4191 | if (ret & VM_FAULT_ERROR) { | |
9a291a7c JM |
4192 | int err = vm_fault_to_errno(ret, flags); |
4193 | ||
4194 | if (err) | |
4195 | return err; | |
4196 | ||
87ffc118 AA |
4197 | remainder = 0; |
4198 | break; | |
4199 | } | |
4200 | if (ret & VM_FAULT_RETRY) { | |
4201 | if (nonblocking) | |
4202 | *nonblocking = 0; | |
4203 | *nr_pages = 0; | |
4204 | /* | |
4205 | * VM_FAULT_RETRY must not return an | |
4206 | * error, it will return zero | |
4207 | * instead. | |
4208 | * | |
4209 | * No need to update "position" as the | |
4210 | * caller will not check it after | |
4211 | * *nr_pages is set to 0. | |
4212 | */ | |
4213 | return i; | |
4214 | } | |
4215 | continue; | |
4c887265 AL |
4216 | } |
4217 | ||
a5516438 | 4218 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 4219 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 4220 | same_page: |
d6692183 | 4221 | if (pages) { |
2a15efc9 | 4222 | pages[i] = mem_map_offset(page, pfn_offset); |
ddc58f27 | 4223 | get_page(pages[i]); |
d6692183 | 4224 | } |
63551ae0 DG |
4225 | |
4226 | if (vmas) | |
4227 | vmas[i] = vma; | |
4228 | ||
4229 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 4230 | ++pfn_offset; |
63551ae0 DG |
4231 | --remainder; |
4232 | ++i; | |
d5d4b0aa | 4233 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 4234 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa CK |
4235 | /* |
4236 | * We use pfn_offset to avoid touching the pageframes | |
4237 | * of this compound page. | |
4238 | */ | |
4239 | goto same_page; | |
4240 | } | |
cb900f41 | 4241 | spin_unlock(ptl); |
63551ae0 | 4242 | } |
28a35716 | 4243 | *nr_pages = remainder; |
87ffc118 AA |
4244 | /* |
4245 | * setting position is actually required only if remainder is | |
4246 | * not zero but it's faster not to add a "if (remainder)" | |
4247 | * branch. | |
4248 | */ | |
63551ae0 DG |
4249 | *position = vaddr; |
4250 | ||
2a15efc9 | 4251 | return i ? i : -EFAULT; |
63551ae0 | 4252 | } |
8f860591 | 4253 | |
5491ae7b AK |
4254 | #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE |
4255 | /* | |
4256 | * ARCHes with special requirements for evicting HUGETLB backing TLB entries can | |
4257 | * implement this. | |
4258 | */ | |
4259 | #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) | |
4260 | #endif | |
4261 | ||
7da4d641 | 4262 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
4263 | unsigned long address, unsigned long end, pgprot_t newprot) |
4264 | { | |
4265 | struct mm_struct *mm = vma->vm_mm; | |
4266 | unsigned long start = address; | |
4267 | pte_t *ptep; | |
4268 | pte_t pte; | |
a5516438 | 4269 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 4270 | unsigned long pages = 0; |
8f860591 ZY |
4271 | |
4272 | BUG_ON(address >= end); | |
4273 | flush_cache_range(vma, address, end); | |
4274 | ||
a5338093 | 4275 | mmu_notifier_invalidate_range_start(mm, start, end); |
83cde9e8 | 4276 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 4277 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 4278 | spinlock_t *ptl; |
7868a208 | 4279 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
8f860591 ZY |
4280 | if (!ptep) |
4281 | continue; | |
cb900f41 | 4282 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
4283 | if (huge_pmd_unshare(mm, &address, ptep)) { |
4284 | pages++; | |
cb900f41 | 4285 | spin_unlock(ptl); |
39dde65c | 4286 | continue; |
7da4d641 | 4287 | } |
a8bda28d NH |
4288 | pte = huge_ptep_get(ptep); |
4289 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
4290 | spin_unlock(ptl); | |
4291 | continue; | |
4292 | } | |
4293 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
4294 | swp_entry_t entry = pte_to_swp_entry(pte); | |
4295 | ||
4296 | if (is_write_migration_entry(entry)) { | |
4297 | pte_t newpte; | |
4298 | ||
4299 | make_migration_entry_read(&entry); | |
4300 | newpte = swp_entry_to_pte(entry); | |
e5251fd4 PA |
4301 | set_huge_swap_pte_at(mm, address, ptep, |
4302 | newpte, huge_page_size(h)); | |
a8bda28d NH |
4303 | pages++; |
4304 | } | |
4305 | spin_unlock(ptl); | |
4306 | continue; | |
4307 | } | |
4308 | if (!huge_pte_none(pte)) { | |
8f860591 | 4309 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
106c992a | 4310 | pte = pte_mkhuge(huge_pte_modify(pte, newprot)); |
be7517d6 | 4311 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
8f860591 | 4312 | set_huge_pte_at(mm, address, ptep, pte); |
7da4d641 | 4313 | pages++; |
8f860591 | 4314 | } |
cb900f41 | 4315 | spin_unlock(ptl); |
8f860591 | 4316 | } |
d833352a | 4317 | /* |
c8c06efa | 4318 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 4319 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 4320 | * once we release i_mmap_rwsem, another task can do the final put_page |
d833352a MG |
4321 | * and that page table be reused and filled with junk. |
4322 | */ | |
5491ae7b | 4323 | flush_hugetlb_tlb_range(vma, start, end); |
34ee645e | 4324 | mmu_notifier_invalidate_range(mm, start, end); |
83cde9e8 | 4325 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
a5338093 | 4326 | mmu_notifier_invalidate_range_end(mm, start, end); |
7da4d641 PZ |
4327 | |
4328 | return pages << h->order; | |
8f860591 ZY |
4329 | } |
4330 | ||
a1e78772 MG |
4331 | int hugetlb_reserve_pages(struct inode *inode, |
4332 | long from, long to, | |
5a6fe125 | 4333 | struct vm_area_struct *vma, |
ca16d140 | 4334 | vm_flags_t vm_flags) |
e4e574b7 | 4335 | { |
17c9d12e | 4336 | long ret, chg; |
a5516438 | 4337 | struct hstate *h = hstate_inode(inode); |
90481622 | 4338 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 4339 | struct resv_map *resv_map; |
1c5ecae3 | 4340 | long gbl_reserve; |
e4e574b7 | 4341 | |
17c9d12e MG |
4342 | /* |
4343 | * Only apply hugepage reservation if asked. At fault time, an | |
4344 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 4345 | * without using reserves |
17c9d12e | 4346 | */ |
ca16d140 | 4347 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
4348 | return 0; |
4349 | ||
a1e78772 MG |
4350 | /* |
4351 | * Shared mappings base their reservation on the number of pages that | |
4352 | * are already allocated on behalf of the file. Private mappings need | |
4353 | * to reserve the full area even if read-only as mprotect() may be | |
4354 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
4355 | */ | |
9119a41e | 4356 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4e35f483 | 4357 | resv_map = inode_resv_map(inode); |
9119a41e | 4358 | |
1406ec9b | 4359 | chg = region_chg(resv_map, from, to); |
9119a41e JK |
4360 | |
4361 | } else { | |
4362 | resv_map = resv_map_alloc(); | |
17c9d12e MG |
4363 | if (!resv_map) |
4364 | return -ENOMEM; | |
4365 | ||
a1e78772 | 4366 | chg = to - from; |
84afd99b | 4367 | |
17c9d12e MG |
4368 | set_vma_resv_map(vma, resv_map); |
4369 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
4370 | } | |
4371 | ||
c50ac050 DH |
4372 | if (chg < 0) { |
4373 | ret = chg; | |
4374 | goto out_err; | |
4375 | } | |
8a630112 | 4376 | |
1c5ecae3 MK |
4377 | /* |
4378 | * There must be enough pages in the subpool for the mapping. If | |
4379 | * the subpool has a minimum size, there may be some global | |
4380 | * reservations already in place (gbl_reserve). | |
4381 | */ | |
4382 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
4383 | if (gbl_reserve < 0) { | |
c50ac050 DH |
4384 | ret = -ENOSPC; |
4385 | goto out_err; | |
4386 | } | |
5a6fe125 MG |
4387 | |
4388 | /* | |
17c9d12e | 4389 | * Check enough hugepages are available for the reservation. |
90481622 | 4390 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 4391 | */ |
1c5ecae3 | 4392 | ret = hugetlb_acct_memory(h, gbl_reserve); |
68842c9b | 4393 | if (ret < 0) { |
1c5ecae3 MK |
4394 | /* put back original number of pages, chg */ |
4395 | (void)hugepage_subpool_put_pages(spool, chg); | |
c50ac050 | 4396 | goto out_err; |
68842c9b | 4397 | } |
17c9d12e MG |
4398 | |
4399 | /* | |
4400 | * Account for the reservations made. Shared mappings record regions | |
4401 | * that have reservations as they are shared by multiple VMAs. | |
4402 | * When the last VMA disappears, the region map says how much | |
4403 | * the reservation was and the page cache tells how much of | |
4404 | * the reservation was consumed. Private mappings are per-VMA and | |
4405 | * only the consumed reservations are tracked. When the VMA | |
4406 | * disappears, the original reservation is the VMA size and the | |
4407 | * consumed reservations are stored in the map. Hence, nothing | |
4408 | * else has to be done for private mappings here | |
4409 | */ | |
33039678 MK |
4410 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4411 | long add = region_add(resv_map, from, to); | |
4412 | ||
4413 | if (unlikely(chg > add)) { | |
4414 | /* | |
4415 | * pages in this range were added to the reserve | |
4416 | * map between region_chg and region_add. This | |
4417 | * indicates a race with alloc_huge_page. Adjust | |
4418 | * the subpool and reserve counts modified above | |
4419 | * based on the difference. | |
4420 | */ | |
4421 | long rsv_adjust; | |
4422 | ||
4423 | rsv_adjust = hugepage_subpool_put_pages(spool, | |
4424 | chg - add); | |
4425 | hugetlb_acct_memory(h, -rsv_adjust); | |
4426 | } | |
4427 | } | |
a43a8c39 | 4428 | return 0; |
c50ac050 | 4429 | out_err: |
5e911373 | 4430 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
ff8c0c53 MK |
4431 | /* Don't call region_abort if region_chg failed */ |
4432 | if (chg >= 0) | |
4433 | region_abort(resv_map, from, to); | |
f031dd27 JK |
4434 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
4435 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 4436 | return ret; |
a43a8c39 CK |
4437 | } |
4438 | ||
b5cec28d MK |
4439 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
4440 | long freed) | |
a43a8c39 | 4441 | { |
a5516438 | 4442 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 4443 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 4444 | long chg = 0; |
90481622 | 4445 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 4446 | long gbl_reserve; |
45c682a6 | 4447 | |
b5cec28d MK |
4448 | if (resv_map) { |
4449 | chg = region_del(resv_map, start, end); | |
4450 | /* | |
4451 | * region_del() can fail in the rare case where a region | |
4452 | * must be split and another region descriptor can not be | |
4453 | * allocated. If end == LONG_MAX, it will not fail. | |
4454 | */ | |
4455 | if (chg < 0) | |
4456 | return chg; | |
4457 | } | |
4458 | ||
45c682a6 | 4459 | spin_lock(&inode->i_lock); |
e4c6f8be | 4460 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
4461 | spin_unlock(&inode->i_lock); |
4462 | ||
1c5ecae3 MK |
4463 | /* |
4464 | * If the subpool has a minimum size, the number of global | |
4465 | * reservations to be released may be adjusted. | |
4466 | */ | |
4467 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
4468 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
4469 | |
4470 | return 0; | |
a43a8c39 | 4471 | } |
93f70f90 | 4472 | |
3212b535 SC |
4473 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
4474 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
4475 | struct vm_area_struct *vma, | |
4476 | unsigned long addr, pgoff_t idx) | |
4477 | { | |
4478 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
4479 | svma->vm_start; | |
4480 | unsigned long sbase = saddr & PUD_MASK; | |
4481 | unsigned long s_end = sbase + PUD_SIZE; | |
4482 | ||
4483 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
4484 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
4485 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
4486 | |
4487 | /* | |
4488 | * match the virtual addresses, permission and the alignment of the | |
4489 | * page table page. | |
4490 | */ | |
4491 | if (pmd_index(addr) != pmd_index(saddr) || | |
4492 | vm_flags != svm_flags || | |
4493 | sbase < svma->vm_start || svma->vm_end < s_end) | |
4494 | return 0; | |
4495 | ||
4496 | return saddr; | |
4497 | } | |
4498 | ||
31aafb45 | 4499 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
4500 | { |
4501 | unsigned long base = addr & PUD_MASK; | |
4502 | unsigned long end = base + PUD_SIZE; | |
4503 | ||
4504 | /* | |
4505 | * check on proper vm_flags and page table alignment | |
4506 | */ | |
4507 | if (vma->vm_flags & VM_MAYSHARE && | |
4508 | vma->vm_start <= base && end <= vma->vm_end) | |
31aafb45 NK |
4509 | return true; |
4510 | return false; | |
3212b535 SC |
4511 | } |
4512 | ||
4513 | /* | |
4514 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
4515 | * and returns the corresponding pte. While this is not necessary for the | |
4516 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
4517 | * code much cleaner. pmd allocation is essential for the shared case because | |
c8c06efa | 4518 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise |
3212b535 SC |
4519 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a |
4520 | * bad pmd for sharing. | |
4521 | */ | |
4522 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4523 | { | |
4524 | struct vm_area_struct *vma = find_vma(mm, addr); | |
4525 | struct address_space *mapping = vma->vm_file->f_mapping; | |
4526 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
4527 | vma->vm_pgoff; | |
4528 | struct vm_area_struct *svma; | |
4529 | unsigned long saddr; | |
4530 | pte_t *spte = NULL; | |
4531 | pte_t *pte; | |
cb900f41 | 4532 | spinlock_t *ptl; |
3212b535 SC |
4533 | |
4534 | if (!vma_shareable(vma, addr)) | |
4535 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
4536 | ||
83cde9e8 | 4537 | i_mmap_lock_write(mapping); |
3212b535 SC |
4538 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
4539 | if (svma == vma) | |
4540 | continue; | |
4541 | ||
4542 | saddr = page_table_shareable(svma, vma, addr, idx); | |
4543 | if (saddr) { | |
7868a208 PA |
4544 | spte = huge_pte_offset(svma->vm_mm, saddr, |
4545 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
4546 | if (spte) { |
4547 | get_page(virt_to_page(spte)); | |
4548 | break; | |
4549 | } | |
4550 | } | |
4551 | } | |
4552 | ||
4553 | if (!spte) | |
4554 | goto out; | |
4555 | ||
8bea8052 | 4556 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 4557 | if (pud_none(*pud)) { |
3212b535 SC |
4558 | pud_populate(mm, pud, |
4559 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 4560 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 4561 | } else { |
3212b535 | 4562 | put_page(virt_to_page(spte)); |
dc6c9a35 | 4563 | } |
cb900f41 | 4564 | spin_unlock(ptl); |
3212b535 SC |
4565 | out: |
4566 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
83cde9e8 | 4567 | i_mmap_unlock_write(mapping); |
3212b535 SC |
4568 | return pte; |
4569 | } | |
4570 | ||
4571 | /* | |
4572 | * unmap huge page backed by shared pte. | |
4573 | * | |
4574 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
4575 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
4576 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
4577 | * | |
cb900f41 | 4578 | * called with page table lock held. |
3212b535 SC |
4579 | * |
4580 | * returns: 1 successfully unmapped a shared pte page | |
4581 | * 0 the underlying pte page is not shared, or it is the last user | |
4582 | */ | |
4583 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4584 | { | |
4585 | pgd_t *pgd = pgd_offset(mm, *addr); | |
c2febafc KS |
4586 | p4d_t *p4d = p4d_offset(pgd, *addr); |
4587 | pud_t *pud = pud_offset(p4d, *addr); | |
3212b535 SC |
4588 | |
4589 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
4590 | if (page_count(virt_to_page(ptep)) == 1) | |
4591 | return 0; | |
4592 | ||
4593 | pud_clear(pud); | |
4594 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 4595 | mm_dec_nr_pmds(mm); |
3212b535 SC |
4596 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
4597 | return 1; | |
4598 | } | |
9e5fc74c SC |
4599 | #define want_pmd_share() (1) |
4600 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
4601 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4602 | { | |
4603 | return NULL; | |
4604 | } | |
e81f2d22 ZZ |
4605 | |
4606 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4607 | { | |
4608 | return 0; | |
4609 | } | |
9e5fc74c | 4610 | #define want_pmd_share() (0) |
3212b535 SC |
4611 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
4612 | ||
9e5fc74c SC |
4613 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
4614 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
4615 | unsigned long addr, unsigned long sz) | |
4616 | { | |
4617 | pgd_t *pgd; | |
c2febafc | 4618 | p4d_t *p4d; |
9e5fc74c SC |
4619 | pud_t *pud; |
4620 | pte_t *pte = NULL; | |
4621 | ||
4622 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
4623 | p4d = p4d_offset(pgd, addr); |
4624 | pud = pud_alloc(mm, p4d, addr); | |
9e5fc74c SC |
4625 | if (pud) { |
4626 | if (sz == PUD_SIZE) { | |
4627 | pte = (pte_t *)pud; | |
4628 | } else { | |
4629 | BUG_ON(sz != PMD_SIZE); | |
4630 | if (want_pmd_share() && pud_none(*pud)) | |
4631 | pte = huge_pmd_share(mm, addr, pud); | |
4632 | else | |
4633 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
4634 | } | |
4635 | } | |
4e666314 | 4636 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
4637 | |
4638 | return pte; | |
4639 | } | |
4640 | ||
7868a208 PA |
4641 | pte_t *huge_pte_offset(struct mm_struct *mm, |
4642 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
4643 | { |
4644 | pgd_t *pgd; | |
c2febafc | 4645 | p4d_t *p4d; |
9e5fc74c | 4646 | pud_t *pud; |
c2febafc | 4647 | pmd_t *pmd; |
9e5fc74c SC |
4648 | |
4649 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
4650 | if (!pgd_present(*pgd)) |
4651 | return NULL; | |
4652 | p4d = p4d_offset(pgd, addr); | |
4653 | if (!p4d_present(*p4d)) | |
4654 | return NULL; | |
4655 | pud = pud_offset(p4d, addr); | |
4656 | if (!pud_present(*pud)) | |
4657 | return NULL; | |
4658 | if (pud_huge(*pud)) | |
4659 | return (pte_t *)pud; | |
4660 | pmd = pmd_offset(pud, addr); | |
9e5fc74c SC |
4661 | return (pte_t *) pmd; |
4662 | } | |
4663 | ||
61f77eda NH |
4664 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
4665 | ||
4666 | /* | |
4667 | * These functions are overwritable if your architecture needs its own | |
4668 | * behavior. | |
4669 | */ | |
4670 | struct page * __weak | |
4671 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
4672 | int write) | |
4673 | { | |
4674 | return ERR_PTR(-EINVAL); | |
4675 | } | |
4676 | ||
4dc71451 AK |
4677 | struct page * __weak |
4678 | follow_huge_pd(struct vm_area_struct *vma, | |
4679 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
4680 | { | |
4681 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
4682 | return NULL; | |
4683 | } | |
4684 | ||
61f77eda | 4685 | struct page * __weak |
9e5fc74c | 4686 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 4687 | pmd_t *pmd, int flags) |
9e5fc74c | 4688 | { |
e66f17ff NH |
4689 | struct page *page = NULL; |
4690 | spinlock_t *ptl; | |
c9d398fa | 4691 | pte_t pte; |
e66f17ff NH |
4692 | retry: |
4693 | ptl = pmd_lockptr(mm, pmd); | |
4694 | spin_lock(ptl); | |
4695 | /* | |
4696 | * make sure that the address range covered by this pmd is not | |
4697 | * unmapped from other threads. | |
4698 | */ | |
4699 | if (!pmd_huge(*pmd)) | |
4700 | goto out; | |
c9d398fa NH |
4701 | pte = huge_ptep_get((pte_t *)pmd); |
4702 | if (pte_present(pte)) { | |
97534127 | 4703 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
e66f17ff NH |
4704 | if (flags & FOLL_GET) |
4705 | get_page(page); | |
4706 | } else { | |
c9d398fa | 4707 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff NH |
4708 | spin_unlock(ptl); |
4709 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
4710 | goto retry; | |
4711 | } | |
4712 | /* | |
4713 | * hwpoisoned entry is treated as no_page_table in | |
4714 | * follow_page_mask(). | |
4715 | */ | |
4716 | } | |
4717 | out: | |
4718 | spin_unlock(ptl); | |
9e5fc74c SC |
4719 | return page; |
4720 | } | |
4721 | ||
61f77eda | 4722 | struct page * __weak |
9e5fc74c | 4723 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 4724 | pud_t *pud, int flags) |
9e5fc74c | 4725 | { |
e66f17ff NH |
4726 | if (flags & FOLL_GET) |
4727 | return NULL; | |
9e5fc74c | 4728 | |
e66f17ff | 4729 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
4730 | } |
4731 | ||
faaa5b62 AK |
4732 | struct page * __weak |
4733 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
4734 | { | |
4735 | if (flags & FOLL_GET) | |
4736 | return NULL; | |
4737 | ||
4738 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
4739 | } | |
4740 | ||
d5bd9106 AK |
4741 | #ifdef CONFIG_MEMORY_FAILURE |
4742 | ||
93f70f90 NH |
4743 | /* |
4744 | * This function is called from memory failure code. | |
93f70f90 | 4745 | */ |
6de2b1aa | 4746 | int dequeue_hwpoisoned_huge_page(struct page *hpage) |
93f70f90 NH |
4747 | { |
4748 | struct hstate *h = page_hstate(hpage); | |
4749 | int nid = page_to_nid(hpage); | |
6de2b1aa | 4750 | int ret = -EBUSY; |
93f70f90 NH |
4751 | |
4752 | spin_lock(&hugetlb_lock); | |
7e1f049e NH |
4753 | /* |
4754 | * Just checking !page_huge_active is not enough, because that could be | |
4755 | * an isolated/hwpoisoned hugepage (which have >0 refcount). | |
4756 | */ | |
4757 | if (!page_huge_active(hpage) && !page_count(hpage)) { | |
56f2fb14 NH |
4758 | /* |
4759 | * Hwpoisoned hugepage isn't linked to activelist or freelist, | |
4760 | * but dangling hpage->lru can trigger list-debug warnings | |
4761 | * (this happens when we call unpoison_memory() on it), | |
4762 | * so let it point to itself with list_del_init(). | |
4763 | */ | |
4764 | list_del_init(&hpage->lru); | |
8c6c2ecb | 4765 | set_page_refcounted(hpage); |
6de2b1aa NH |
4766 | h->free_huge_pages--; |
4767 | h->free_huge_pages_node[nid]--; | |
4768 | ret = 0; | |
4769 | } | |
93f70f90 | 4770 | spin_unlock(&hugetlb_lock); |
6de2b1aa | 4771 | return ret; |
93f70f90 | 4772 | } |
6de2b1aa | 4773 | #endif |
31caf665 NH |
4774 | |
4775 | bool isolate_huge_page(struct page *page, struct list_head *list) | |
4776 | { | |
bcc54222 NH |
4777 | bool ret = true; |
4778 | ||
309381fe | 4779 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4780 | spin_lock(&hugetlb_lock); |
bcc54222 NH |
4781 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { |
4782 | ret = false; | |
4783 | goto unlock; | |
4784 | } | |
4785 | clear_page_huge_active(page); | |
31caf665 | 4786 | list_move_tail(&page->lru, list); |
bcc54222 | 4787 | unlock: |
31caf665 | 4788 | spin_unlock(&hugetlb_lock); |
bcc54222 | 4789 | return ret; |
31caf665 NH |
4790 | } |
4791 | ||
4792 | void putback_active_hugepage(struct page *page) | |
4793 | { | |
309381fe | 4794 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4795 | spin_lock(&hugetlb_lock); |
bcc54222 | 4796 | set_page_huge_active(page); |
31caf665 NH |
4797 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
4798 | spin_unlock(&hugetlb_lock); | |
4799 | put_page(page); | |
4800 | } |