]>
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 | */ |
c3114a84 | 1462 | 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 | } | |
c3114a84 AK |
1475 | /* |
1476 | * Move PageHWPoison flag from head page to the raw error page, | |
1477 | * which makes any subpages rather than the error page reusable. | |
1478 | */ | |
1479 | if (PageHWPoison(head) && page != head) { | |
1480 | SetPageHWPoison(page); | |
1481 | ClearPageHWPoison(head); | |
1482 | } | |
2247bb33 | 1483 | list_del(&head->lru); |
c8721bbb NH |
1484 | h->free_huge_pages--; |
1485 | h->free_huge_pages_node[nid]--; | |
c1470b33 | 1486 | h->max_huge_pages--; |
2247bb33 | 1487 | update_and_free_page(h, head); |
c8721bbb | 1488 | } |
082d5b6b | 1489 | out: |
c8721bbb | 1490 | spin_unlock(&hugetlb_lock); |
082d5b6b | 1491 | return rc; |
c8721bbb NH |
1492 | } |
1493 | ||
1494 | /* | |
1495 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1496 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
1497 | * Note that this will dissolve a free gigantic hugepage completely, if any |
1498 | * part of it lies within the given range. | |
082d5b6b GS |
1499 | * Also note that if dissolve_free_huge_page() returns with an error, all |
1500 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 1501 | */ |
082d5b6b | 1502 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 1503 | { |
c8721bbb | 1504 | unsigned long pfn; |
eb03aa00 | 1505 | struct page *page; |
082d5b6b | 1506 | int rc = 0; |
c8721bbb | 1507 | |
d0177639 | 1508 | if (!hugepages_supported()) |
082d5b6b | 1509 | return rc; |
d0177639 | 1510 | |
eb03aa00 GS |
1511 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { |
1512 | page = pfn_to_page(pfn); | |
1513 | if (PageHuge(page) && !page_count(page)) { | |
1514 | rc = dissolve_free_huge_page(page); | |
1515 | if (rc) | |
1516 | break; | |
1517 | } | |
1518 | } | |
082d5b6b GS |
1519 | |
1520 | return rc; | |
c8721bbb NH |
1521 | } |
1522 | ||
099730d6 DH |
1523 | /* |
1524 | * There are 3 ways this can get called: | |
1525 | * 1. With vma+addr: we use the VMA's memory policy | |
1526 | * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge | |
1527 | * page from any node, and let the buddy allocator itself figure | |
1528 | * it out. | |
1529 | * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page | |
1530 | * strictly from 'nid' | |
1531 | */ | |
1532 | static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h, | |
1533 | struct vm_area_struct *vma, unsigned long addr, int nid) | |
1534 | { | |
1535 | int order = huge_page_order(h); | |
1536 | gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN; | |
1537 | unsigned int cpuset_mems_cookie; | |
1538 | ||
1539 | /* | |
1540 | * We need a VMA to get a memory policy. If we do not | |
e0ec90ee DH |
1541 | * have one, we use the 'nid' argument. |
1542 | * | |
1543 | * The mempolicy stuff below has some non-inlined bits | |
1544 | * and calls ->vm_ops. That makes it hard to optimize at | |
1545 | * compile-time, even when NUMA is off and it does | |
1546 | * nothing. This helps the compiler optimize it out. | |
099730d6 | 1547 | */ |
e0ec90ee | 1548 | if (!IS_ENABLED(CONFIG_NUMA) || !vma) { |
099730d6 DH |
1549 | /* |
1550 | * If a specific node is requested, make sure to | |
1551 | * get memory from there, but only when a node | |
1552 | * is explicitly specified. | |
1553 | */ | |
1554 | if (nid != NUMA_NO_NODE) | |
1555 | gfp |= __GFP_THISNODE; | |
1556 | /* | |
1557 | * Make sure to call something that can handle | |
1558 | * nid=NUMA_NO_NODE | |
1559 | */ | |
1560 | return alloc_pages_node(nid, gfp, order); | |
1561 | } | |
1562 | ||
1563 | /* | |
1564 | * OK, so we have a VMA. Fetch the mempolicy and try to | |
e0ec90ee DH |
1565 | * allocate a huge page with it. We will only reach this |
1566 | * when CONFIG_NUMA=y. | |
099730d6 DH |
1567 | */ |
1568 | do { | |
1569 | struct page *page; | |
1570 | struct mempolicy *mpol; | |
04ec6264 | 1571 | int nid; |
099730d6 DH |
1572 | nodemask_t *nodemask; |
1573 | ||
1574 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
04ec6264 | 1575 | nid = huge_node(vma, addr, gfp, &mpol, &nodemask); |
099730d6 | 1576 | mpol_cond_put(mpol); |
04ec6264 | 1577 | page = __alloc_pages_nodemask(gfp, order, nid, nodemask); |
099730d6 DH |
1578 | if (page) |
1579 | return page; | |
1580 | } while (read_mems_allowed_retry(cpuset_mems_cookie)); | |
1581 | ||
1582 | return NULL; | |
1583 | } | |
1584 | ||
1585 | /* | |
1586 | * There are two ways to allocate a huge page: | |
1587 | * 1. When you have a VMA and an address (like a fault) | |
1588 | * 2. When you have no VMA (like when setting /proc/.../nr_hugepages) | |
1589 | * | |
1590 | * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in | |
1591 | * this case which signifies that the allocation should be done with | |
1592 | * respect for the VMA's memory policy. | |
1593 | * | |
1594 | * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This | |
1595 | * implies that memory policies will not be taken in to account. | |
1596 | */ | |
1597 | static struct page *__alloc_buddy_huge_page(struct hstate *h, | |
1598 | struct vm_area_struct *vma, unsigned long addr, int nid) | |
7893d1d5 AL |
1599 | { |
1600 | struct page *page; | |
bf50bab2 | 1601 | unsigned int r_nid; |
7893d1d5 | 1602 | |
bae7f4ae | 1603 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1604 | return NULL; |
1605 | ||
099730d6 DH |
1606 | /* |
1607 | * Make sure that anyone specifying 'nid' is not also specifying a VMA. | |
1608 | * This makes sure the caller is picking _one_ of the modes with which | |
1609 | * we can call this function, not both. | |
1610 | */ | |
1611 | if (vma || (addr != -1)) { | |
e0ec90ee DH |
1612 | VM_WARN_ON_ONCE(addr == -1); |
1613 | VM_WARN_ON_ONCE(nid != NUMA_NO_NODE); | |
099730d6 | 1614 | } |
d1c3fb1f NA |
1615 | /* |
1616 | * Assume we will successfully allocate the surplus page to | |
1617 | * prevent racing processes from causing the surplus to exceed | |
1618 | * overcommit | |
1619 | * | |
1620 | * This however introduces a different race, where a process B | |
1621 | * tries to grow the static hugepage pool while alloc_pages() is | |
1622 | * called by process A. B will only examine the per-node | |
1623 | * counters in determining if surplus huge pages can be | |
1624 | * converted to normal huge pages in adjust_pool_surplus(). A | |
1625 | * won't be able to increment the per-node counter, until the | |
1626 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
1627 | * no more huge pages can be converted from surplus to normal | |
1628 | * state (and doesn't try to convert again). Thus, we have a | |
1629 | * case where a surplus huge page exists, the pool is grown, and | |
1630 | * the surplus huge page still exists after, even though it | |
1631 | * should just have been converted to a normal huge page. This | |
1632 | * does not leak memory, though, as the hugepage will be freed | |
1633 | * once it is out of use. It also does not allow the counters to | |
1634 | * go out of whack in adjust_pool_surplus() as we don't modify | |
1635 | * the node values until we've gotten the hugepage and only the | |
1636 | * per-node value is checked there. | |
1637 | */ | |
1638 | spin_lock(&hugetlb_lock); | |
a5516438 | 1639 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
1640 | spin_unlock(&hugetlb_lock); |
1641 | return NULL; | |
1642 | } else { | |
a5516438 AK |
1643 | h->nr_huge_pages++; |
1644 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
1645 | } |
1646 | spin_unlock(&hugetlb_lock); | |
1647 | ||
099730d6 | 1648 | page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid); |
d1c3fb1f NA |
1649 | |
1650 | spin_lock(&hugetlb_lock); | |
7893d1d5 | 1651 | if (page) { |
0edaecfa | 1652 | INIT_LIST_HEAD(&page->lru); |
bf50bab2 | 1653 | r_nid = page_to_nid(page); |
f1e61557 | 1654 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
9dd540e2 | 1655 | set_hugetlb_cgroup(page, NULL); |
d1c3fb1f NA |
1656 | /* |
1657 | * We incremented the global counters already | |
1658 | */ | |
bf50bab2 NH |
1659 | h->nr_huge_pages_node[r_nid]++; |
1660 | h->surplus_huge_pages_node[r_nid]++; | |
3b116300 | 1661 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 1662 | } else { |
a5516438 AK |
1663 | h->nr_huge_pages--; |
1664 | h->surplus_huge_pages--; | |
3b116300 | 1665 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 1666 | } |
d1c3fb1f | 1667 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1668 | |
1669 | return page; | |
1670 | } | |
1671 | ||
099730d6 DH |
1672 | /* |
1673 | * Allocate a huge page from 'nid'. Note, 'nid' may be | |
1674 | * NUMA_NO_NODE, which means that it may be allocated | |
1675 | * anywhere. | |
1676 | */ | |
e0ec90ee | 1677 | static |
099730d6 DH |
1678 | struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid) |
1679 | { | |
1680 | unsigned long addr = -1; | |
1681 | ||
1682 | return __alloc_buddy_huge_page(h, NULL, addr, nid); | |
1683 | } | |
1684 | ||
1685 | /* | |
1686 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
1687 | */ | |
e0ec90ee | 1688 | static |
099730d6 DH |
1689 | struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h, |
1690 | struct vm_area_struct *vma, unsigned long addr) | |
1691 | { | |
1692 | return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE); | |
1693 | } | |
1694 | ||
bf50bab2 NH |
1695 | /* |
1696 | * This allocation function is useful in the context where vma is irrelevant. | |
1697 | * E.g. soft-offlining uses this function because it only cares physical | |
1698 | * address of error page. | |
1699 | */ | |
1700 | struct page *alloc_huge_page_node(struct hstate *h, int nid) | |
1701 | { | |
4ef91848 | 1702 | struct page *page = NULL; |
bf50bab2 NH |
1703 | |
1704 | spin_lock(&hugetlb_lock); | |
4ef91848 JK |
1705 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
1706 | page = dequeue_huge_page_node(h, nid); | |
bf50bab2 NH |
1707 | spin_unlock(&hugetlb_lock); |
1708 | ||
94ae8ba7 | 1709 | if (!page) |
099730d6 | 1710 | page = __alloc_buddy_huge_page_no_mpol(h, nid); |
bf50bab2 NH |
1711 | |
1712 | return page; | |
1713 | } | |
1714 | ||
e4e574b7 | 1715 | /* |
25985edc | 1716 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1717 | * of size 'delta'. |
1718 | */ | |
a5516438 | 1719 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1720 | { |
1721 | struct list_head surplus_list; | |
1722 | struct page *page, *tmp; | |
1723 | int ret, i; | |
1724 | int needed, allocated; | |
28073b02 | 1725 | bool alloc_ok = true; |
e4e574b7 | 1726 | |
a5516438 | 1727 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1728 | if (needed <= 0) { |
a5516438 | 1729 | h->resv_huge_pages += delta; |
e4e574b7 | 1730 | return 0; |
ac09b3a1 | 1731 | } |
e4e574b7 AL |
1732 | |
1733 | allocated = 0; | |
1734 | INIT_LIST_HEAD(&surplus_list); | |
1735 | ||
1736 | ret = -ENOMEM; | |
1737 | retry: | |
1738 | spin_unlock(&hugetlb_lock); | |
1739 | for (i = 0; i < needed; i++) { | |
099730d6 | 1740 | page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE); |
28073b02 HD |
1741 | if (!page) { |
1742 | alloc_ok = false; | |
1743 | break; | |
1744 | } | |
e4e574b7 AL |
1745 | list_add(&page->lru, &surplus_list); |
1746 | } | |
28073b02 | 1747 | allocated += i; |
e4e574b7 AL |
1748 | |
1749 | /* | |
1750 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1751 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1752 | */ | |
1753 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1754 | needed = (h->resv_huge_pages + delta) - |
1755 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1756 | if (needed > 0) { |
1757 | if (alloc_ok) | |
1758 | goto retry; | |
1759 | /* | |
1760 | * We were not able to allocate enough pages to | |
1761 | * satisfy the entire reservation so we free what | |
1762 | * we've allocated so far. | |
1763 | */ | |
1764 | goto free; | |
1765 | } | |
e4e574b7 AL |
1766 | /* |
1767 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1768 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1769 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1770 | * allocator. Commit the entire reservation here to prevent another |
1771 | * process from stealing the pages as they are added to the pool but | |
1772 | * before they are reserved. | |
e4e574b7 AL |
1773 | */ |
1774 | needed += allocated; | |
a5516438 | 1775 | h->resv_huge_pages += delta; |
e4e574b7 | 1776 | ret = 0; |
a9869b83 | 1777 | |
19fc3f0a | 1778 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1779 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1780 | if ((--needed) < 0) |
1781 | break; | |
a9869b83 NH |
1782 | /* |
1783 | * This page is now managed by the hugetlb allocator and has | |
1784 | * no users -- drop the buddy allocator's reference. | |
1785 | */ | |
1786 | put_page_testzero(page); | |
309381fe | 1787 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 1788 | enqueue_huge_page(h, page); |
19fc3f0a | 1789 | } |
28073b02 | 1790 | free: |
b0365c8d | 1791 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1792 | |
1793 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1794 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1795 | put_page(page); | |
a9869b83 | 1796 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1797 | |
1798 | return ret; | |
1799 | } | |
1800 | ||
1801 | /* | |
e5bbc8a6 MK |
1802 | * This routine has two main purposes: |
1803 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
1804 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
1805 | * to the associated reservation map. | |
1806 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
1807 | * the reservation. As many as unused_resv_pages may be freed. | |
1808 | * | |
1809 | * Called with hugetlb_lock held. However, the lock could be dropped (and | |
1810 | * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, | |
1811 | * we must make sure nobody else can claim pages we are in the process of | |
1812 | * freeing. Do this by ensuring resv_huge_page always is greater than the | |
1813 | * number of huge pages we plan to free when dropping the lock. | |
e4e574b7 | 1814 | */ |
a5516438 AK |
1815 | static void return_unused_surplus_pages(struct hstate *h, |
1816 | unsigned long unused_resv_pages) | |
e4e574b7 | 1817 | { |
e4e574b7 AL |
1818 | unsigned long nr_pages; |
1819 | ||
aa888a74 | 1820 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 1821 | if (hstate_is_gigantic(h)) |
e5bbc8a6 | 1822 | goto out; |
aa888a74 | 1823 | |
e5bbc8a6 MK |
1824 | /* |
1825 | * Part (or even all) of the reservation could have been backed | |
1826 | * by pre-allocated pages. Only free surplus pages. | |
1827 | */ | |
a5516438 | 1828 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1829 | |
685f3457 LS |
1830 | /* |
1831 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1832 | * evenly across all nodes with memory. Iterate across these nodes |
1833 | * until we can no longer free unreserved surplus pages. This occurs | |
1834 | * when the nodes with surplus pages have no free pages. | |
1835 | * free_pool_huge_page() will balance the the freed pages across the | |
1836 | * on-line nodes with memory and will handle the hstate accounting. | |
e5bbc8a6 MK |
1837 | * |
1838 | * Note that we decrement resv_huge_pages as we free the pages. If | |
1839 | * we drop the lock, resv_huge_pages will still be sufficiently large | |
1840 | * to cover subsequent pages we may free. | |
685f3457 LS |
1841 | */ |
1842 | while (nr_pages--) { | |
e5bbc8a6 MK |
1843 | h->resv_huge_pages--; |
1844 | unused_resv_pages--; | |
8cebfcd0 | 1845 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
e5bbc8a6 | 1846 | goto out; |
7848a4bf | 1847 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 | 1848 | } |
e5bbc8a6 MK |
1849 | |
1850 | out: | |
1851 | /* Fully uncommit the reservation */ | |
1852 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 AL |
1853 | } |
1854 | ||
5e911373 | 1855 | |
c37f9fb1 | 1856 | /* |
feba16e2 | 1857 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 1858 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
1859 | * |
1860 | * vma_needs_reservation is called to determine if the huge page at addr | |
1861 | * within the vma has an associated reservation. If a reservation is | |
1862 | * needed, the value 1 is returned. The caller is then responsible for | |
1863 | * managing the global reservation and subpool usage counts. After | |
1864 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
1865 | * to add the page to the reservation map. If the page allocation fails, |
1866 | * the reservation must be ended instead of committed. vma_end_reservation | |
1867 | * is called in such cases. | |
cf3ad20b MK |
1868 | * |
1869 | * In the normal case, vma_commit_reservation returns the same value | |
1870 | * as the preceding vma_needs_reservation call. The only time this | |
1871 | * is not the case is if a reserve map was changed between calls. It | |
1872 | * is the responsibility of the caller to notice the difference and | |
1873 | * take appropriate action. | |
96b96a96 MK |
1874 | * |
1875 | * vma_add_reservation is used in error paths where a reservation must | |
1876 | * be restored when a newly allocated huge page must be freed. It is | |
1877 | * to be called after calling vma_needs_reservation to determine if a | |
1878 | * reservation exists. | |
c37f9fb1 | 1879 | */ |
5e911373 MK |
1880 | enum vma_resv_mode { |
1881 | VMA_NEEDS_RESV, | |
1882 | VMA_COMMIT_RESV, | |
feba16e2 | 1883 | VMA_END_RESV, |
96b96a96 | 1884 | VMA_ADD_RESV, |
5e911373 | 1885 | }; |
cf3ad20b MK |
1886 | static long __vma_reservation_common(struct hstate *h, |
1887 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 1888 | enum vma_resv_mode mode) |
c37f9fb1 | 1889 | { |
4e35f483 JK |
1890 | struct resv_map *resv; |
1891 | pgoff_t idx; | |
cf3ad20b | 1892 | long ret; |
c37f9fb1 | 1893 | |
4e35f483 JK |
1894 | resv = vma_resv_map(vma); |
1895 | if (!resv) | |
84afd99b | 1896 | return 1; |
c37f9fb1 | 1897 | |
4e35f483 | 1898 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
1899 | switch (mode) { |
1900 | case VMA_NEEDS_RESV: | |
cf3ad20b | 1901 | ret = region_chg(resv, idx, idx + 1); |
5e911373 MK |
1902 | break; |
1903 | case VMA_COMMIT_RESV: | |
1904 | ret = region_add(resv, idx, idx + 1); | |
1905 | break; | |
feba16e2 | 1906 | case VMA_END_RESV: |
5e911373 MK |
1907 | region_abort(resv, idx, idx + 1); |
1908 | ret = 0; | |
1909 | break; | |
96b96a96 MK |
1910 | case VMA_ADD_RESV: |
1911 | if (vma->vm_flags & VM_MAYSHARE) | |
1912 | ret = region_add(resv, idx, idx + 1); | |
1913 | else { | |
1914 | region_abort(resv, idx, idx + 1); | |
1915 | ret = region_del(resv, idx, idx + 1); | |
1916 | } | |
1917 | break; | |
5e911373 MK |
1918 | default: |
1919 | BUG(); | |
1920 | } | |
84afd99b | 1921 | |
4e35f483 | 1922 | if (vma->vm_flags & VM_MAYSHARE) |
cf3ad20b | 1923 | return ret; |
67961f9d MK |
1924 | else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { |
1925 | /* | |
1926 | * In most cases, reserves always exist for private mappings. | |
1927 | * However, a file associated with mapping could have been | |
1928 | * hole punched or truncated after reserves were consumed. | |
1929 | * As subsequent fault on such a range will not use reserves. | |
1930 | * Subtle - The reserve map for private mappings has the | |
1931 | * opposite meaning than that of shared mappings. If NO | |
1932 | * entry is in the reserve map, it means a reservation exists. | |
1933 | * If an entry exists in the reserve map, it means the | |
1934 | * reservation has already been consumed. As a result, the | |
1935 | * return value of this routine is the opposite of the | |
1936 | * value returned from reserve map manipulation routines above. | |
1937 | */ | |
1938 | if (ret) | |
1939 | return 0; | |
1940 | else | |
1941 | return 1; | |
1942 | } | |
4e35f483 | 1943 | else |
cf3ad20b | 1944 | return ret < 0 ? ret : 0; |
c37f9fb1 | 1945 | } |
cf3ad20b MK |
1946 | |
1947 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 1948 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 1949 | { |
5e911373 | 1950 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 1951 | } |
84afd99b | 1952 | |
cf3ad20b MK |
1953 | static long vma_commit_reservation(struct hstate *h, |
1954 | struct vm_area_struct *vma, unsigned long addr) | |
1955 | { | |
5e911373 MK |
1956 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
1957 | } | |
1958 | ||
feba16e2 | 1959 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
1960 | struct vm_area_struct *vma, unsigned long addr) |
1961 | { | |
feba16e2 | 1962 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
1963 | } |
1964 | ||
96b96a96 MK |
1965 | static long vma_add_reservation(struct hstate *h, |
1966 | struct vm_area_struct *vma, unsigned long addr) | |
1967 | { | |
1968 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
1969 | } | |
1970 | ||
1971 | /* | |
1972 | * This routine is called to restore a reservation on error paths. In the | |
1973 | * specific error paths, a huge page was allocated (via alloc_huge_page) | |
1974 | * and is about to be freed. If a reservation for the page existed, | |
1975 | * alloc_huge_page would have consumed the reservation and set PagePrivate | |
1976 | * in the newly allocated page. When the page is freed via free_huge_page, | |
1977 | * the global reservation count will be incremented if PagePrivate is set. | |
1978 | * However, free_huge_page can not adjust the reserve map. Adjust the | |
1979 | * reserve map here to be consistent with global reserve count adjustments | |
1980 | * to be made by free_huge_page. | |
1981 | */ | |
1982 | static void restore_reserve_on_error(struct hstate *h, | |
1983 | struct vm_area_struct *vma, unsigned long address, | |
1984 | struct page *page) | |
1985 | { | |
1986 | if (unlikely(PagePrivate(page))) { | |
1987 | long rc = vma_needs_reservation(h, vma, address); | |
1988 | ||
1989 | if (unlikely(rc < 0)) { | |
1990 | /* | |
1991 | * Rare out of memory condition in reserve map | |
1992 | * manipulation. Clear PagePrivate so that | |
1993 | * global reserve count will not be incremented | |
1994 | * by free_huge_page. This will make it appear | |
1995 | * as though the reservation for this page was | |
1996 | * consumed. This may prevent the task from | |
1997 | * faulting in the page at a later time. This | |
1998 | * is better than inconsistent global huge page | |
1999 | * accounting of reserve counts. | |
2000 | */ | |
2001 | ClearPagePrivate(page); | |
2002 | } else if (rc) { | |
2003 | rc = vma_add_reservation(h, vma, address); | |
2004 | if (unlikely(rc < 0)) | |
2005 | /* | |
2006 | * See above comment about rare out of | |
2007 | * memory condition. | |
2008 | */ | |
2009 | ClearPagePrivate(page); | |
2010 | } else | |
2011 | vma_end_reservation(h, vma, address); | |
2012 | } | |
2013 | } | |
2014 | ||
70c3547e | 2015 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 2016 | unsigned long addr, int avoid_reserve) |
1da177e4 | 2017 | { |
90481622 | 2018 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 2019 | struct hstate *h = hstate_vma(vma); |
348ea204 | 2020 | struct page *page; |
d85f69b0 MK |
2021 | long map_chg, map_commit; |
2022 | long gbl_chg; | |
6d76dcf4 AK |
2023 | int ret, idx; |
2024 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 2025 | |
6d76dcf4 | 2026 | idx = hstate_index(h); |
a1e78772 | 2027 | /* |
d85f69b0 MK |
2028 | * Examine the region/reserve map to determine if the process |
2029 | * has a reservation for the page to be allocated. A return | |
2030 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 2031 | */ |
d85f69b0 MK |
2032 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
2033 | if (map_chg < 0) | |
76dcee75 | 2034 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
2035 | |
2036 | /* | |
2037 | * Processes that did not create the mapping will have no | |
2038 | * reserves as indicated by the region/reserve map. Check | |
2039 | * that the allocation will not exceed the subpool limit. | |
2040 | * Allocations for MAP_NORESERVE mappings also need to be | |
2041 | * checked against any subpool limit. | |
2042 | */ | |
2043 | if (map_chg || avoid_reserve) { | |
2044 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2045 | if (gbl_chg < 0) { | |
feba16e2 | 2046 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2047 | return ERR_PTR(-ENOSPC); |
5e911373 | 2048 | } |
1da177e4 | 2049 | |
d85f69b0 MK |
2050 | /* |
2051 | * Even though there was no reservation in the region/reserve | |
2052 | * map, there could be reservations associated with the | |
2053 | * subpool that can be used. This would be indicated if the | |
2054 | * return value of hugepage_subpool_get_pages() is zero. | |
2055 | * However, if avoid_reserve is specified we still avoid even | |
2056 | * the subpool reservations. | |
2057 | */ | |
2058 | if (avoid_reserve) | |
2059 | gbl_chg = 1; | |
2060 | } | |
2061 | ||
6d76dcf4 | 2062 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f JZ |
2063 | if (ret) |
2064 | goto out_subpool_put; | |
2065 | ||
1da177e4 | 2066 | spin_lock(&hugetlb_lock); |
d85f69b0 MK |
2067 | /* |
2068 | * glb_chg is passed to indicate whether or not a page must be taken | |
2069 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2070 | * a reservation exists for the allocation. | |
2071 | */ | |
2072 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2073 | if (!page) { |
94ae8ba7 | 2074 | spin_unlock(&hugetlb_lock); |
099730d6 | 2075 | page = __alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2076 | if (!page) |
2077 | goto out_uncharge_cgroup; | |
a88c7695 NH |
2078 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
2079 | SetPagePrivate(page); | |
2080 | h->resv_huge_pages--; | |
2081 | } | |
79dbb236 AK |
2082 | spin_lock(&hugetlb_lock); |
2083 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 2084 | /* Fall through */ |
68842c9b | 2085 | } |
81a6fcae JK |
2086 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
2087 | spin_unlock(&hugetlb_lock); | |
348ea204 | 2088 | |
90481622 | 2089 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 2090 | |
d85f69b0 MK |
2091 | map_commit = vma_commit_reservation(h, vma, addr); |
2092 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2093 | /* |
2094 | * The page was added to the reservation map between | |
2095 | * vma_needs_reservation and vma_commit_reservation. | |
2096 | * This indicates a race with hugetlb_reserve_pages. | |
2097 | * Adjust for the subpool count incremented above AND | |
2098 | * in hugetlb_reserve_pages for the same page. Also, | |
2099 | * the reservation count added in hugetlb_reserve_pages | |
2100 | * no longer applies. | |
2101 | */ | |
2102 | long rsv_adjust; | |
2103 | ||
2104 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2105 | hugetlb_acct_memory(h, -rsv_adjust); | |
2106 | } | |
90d8b7e6 | 2107 | return page; |
8f34af6f JZ |
2108 | |
2109 | out_uncharge_cgroup: | |
2110 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
2111 | out_subpool_put: | |
d85f69b0 | 2112 | if (map_chg || avoid_reserve) |
8f34af6f | 2113 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2114 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2115 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2116 | } |
2117 | ||
74060e4d NH |
2118 | /* |
2119 | * alloc_huge_page()'s wrapper which simply returns the page if allocation | |
2120 | * succeeds, otherwise NULL. This function is called from new_vma_page(), | |
2121 | * where no ERR_VALUE is expected to be returned. | |
2122 | */ | |
2123 | struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, | |
2124 | unsigned long addr, int avoid_reserve) | |
2125 | { | |
2126 | struct page *page = alloc_huge_page(vma, addr, avoid_reserve); | |
2127 | if (IS_ERR(page)) | |
2128 | page = NULL; | |
2129 | return page; | |
2130 | } | |
2131 | ||
91f47662 | 2132 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
aa888a74 AK |
2133 | { |
2134 | struct huge_bootmem_page *m; | |
b2261026 | 2135 | int nr_nodes, node; |
aa888a74 | 2136 | |
b2261026 | 2137 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
2138 | void *addr; |
2139 | ||
8b89a116 GS |
2140 | addr = memblock_virt_alloc_try_nid_nopanic( |
2141 | huge_page_size(h), huge_page_size(h), | |
2142 | 0, BOOTMEM_ALLOC_ACCESSIBLE, node); | |
aa888a74 AK |
2143 | if (addr) { |
2144 | /* | |
2145 | * Use the beginning of the huge page to store the | |
2146 | * huge_bootmem_page struct (until gather_bootmem | |
2147 | * puts them into the mem_map). | |
2148 | */ | |
2149 | m = addr; | |
91f47662 | 2150 | goto found; |
aa888a74 | 2151 | } |
aa888a74 AK |
2152 | } |
2153 | return 0; | |
2154 | ||
2155 | found: | |
df994ead | 2156 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 AK |
2157 | /* Put them into a private list first because mem_map is not up yet */ |
2158 | list_add(&m->list, &huge_boot_pages); | |
2159 | m->hstate = h; | |
2160 | return 1; | |
2161 | } | |
2162 | ||
d00181b9 KS |
2163 | static void __init prep_compound_huge_page(struct page *page, |
2164 | unsigned int order) | |
18229df5 AW |
2165 | { |
2166 | if (unlikely(order > (MAX_ORDER - 1))) | |
2167 | prep_compound_gigantic_page(page, order); | |
2168 | else | |
2169 | prep_compound_page(page, order); | |
2170 | } | |
2171 | ||
aa888a74 AK |
2172 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
2173 | static void __init gather_bootmem_prealloc(void) | |
2174 | { | |
2175 | struct huge_bootmem_page *m; | |
2176 | ||
2177 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 2178 | struct hstate *h = m->hstate; |
ee8f248d BB |
2179 | struct page *page; |
2180 | ||
2181 | #ifdef CONFIG_HIGHMEM | |
2182 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
8b89a116 GS |
2183 | memblock_free_late(__pa(m), |
2184 | sizeof(struct huge_bootmem_page)); | |
ee8f248d BB |
2185 | #else |
2186 | page = virt_to_page(m); | |
2187 | #endif | |
aa888a74 | 2188 | WARN_ON(page_count(page) != 1); |
18229df5 | 2189 | prep_compound_huge_page(page, h->order); |
ef5a22be | 2190 | WARN_ON(PageReserved(page)); |
aa888a74 | 2191 | prep_new_huge_page(h, page, page_to_nid(page)); |
b0320c7b RA |
2192 | /* |
2193 | * If we had gigantic hugepages allocated at boot time, we need | |
2194 | * to restore the 'stolen' pages to totalram_pages in order to | |
2195 | * fix confusing memory reports from free(1) and another | |
2196 | * side-effects, like CommitLimit going negative. | |
2197 | */ | |
bae7f4ae | 2198 | if (hstate_is_gigantic(h)) |
3dcc0571 | 2199 | adjust_managed_page_count(page, 1 << h->order); |
aa888a74 AK |
2200 | } |
2201 | } | |
2202 | ||
8faa8b07 | 2203 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
2204 | { |
2205 | unsigned long i; | |
a5516438 | 2206 | |
e5ff2159 | 2207 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 2208 | if (hstate_is_gigantic(h)) { |
aa888a74 AK |
2209 | if (!alloc_bootmem_huge_page(h)) |
2210 | break; | |
9b5e5d0f | 2211 | } else if (!alloc_fresh_huge_page(h, |
8cebfcd0 | 2212 | &node_states[N_MEMORY])) |
1da177e4 | 2213 | break; |
1da177e4 | 2214 | } |
8faa8b07 | 2215 | h->max_huge_pages = i; |
e5ff2159 AK |
2216 | } |
2217 | ||
2218 | static void __init hugetlb_init_hstates(void) | |
2219 | { | |
2220 | struct hstate *h; | |
2221 | ||
2222 | for_each_hstate(h) { | |
641844f5 NH |
2223 | if (minimum_order > huge_page_order(h)) |
2224 | minimum_order = huge_page_order(h); | |
2225 | ||
8faa8b07 | 2226 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 2227 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 2228 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 2229 | } |
641844f5 | 2230 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
2231 | } |
2232 | ||
4abd32db AK |
2233 | static char * __init memfmt(char *buf, unsigned long n) |
2234 | { | |
2235 | if (n >= (1UL << 30)) | |
2236 | sprintf(buf, "%lu GB", n >> 30); | |
2237 | else if (n >= (1UL << 20)) | |
2238 | sprintf(buf, "%lu MB", n >> 20); | |
2239 | else | |
2240 | sprintf(buf, "%lu KB", n >> 10); | |
2241 | return buf; | |
2242 | } | |
2243 | ||
e5ff2159 AK |
2244 | static void __init report_hugepages(void) |
2245 | { | |
2246 | struct hstate *h; | |
2247 | ||
2248 | for_each_hstate(h) { | |
4abd32db | 2249 | char buf[32]; |
ffb22af5 | 2250 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
4abd32db AK |
2251 | memfmt(buf, huge_page_size(h)), |
2252 | h->free_huge_pages); | |
e5ff2159 AK |
2253 | } |
2254 | } | |
2255 | ||
1da177e4 | 2256 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
2257 | static void try_to_free_low(struct hstate *h, unsigned long count, |
2258 | nodemask_t *nodes_allowed) | |
1da177e4 | 2259 | { |
4415cc8d CL |
2260 | int i; |
2261 | ||
bae7f4ae | 2262 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2263 | return; |
2264 | ||
6ae11b27 | 2265 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 2266 | struct page *page, *next; |
a5516438 AK |
2267 | struct list_head *freel = &h->hugepage_freelists[i]; |
2268 | list_for_each_entry_safe(page, next, freel, lru) { | |
2269 | if (count >= h->nr_huge_pages) | |
6b0c880d | 2270 | return; |
1da177e4 LT |
2271 | if (PageHighMem(page)) |
2272 | continue; | |
2273 | list_del(&page->lru); | |
e5ff2159 | 2274 | update_and_free_page(h, page); |
a5516438 AK |
2275 | h->free_huge_pages--; |
2276 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
2277 | } |
2278 | } | |
2279 | } | |
2280 | #else | |
6ae11b27 LS |
2281 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
2282 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
2283 | { |
2284 | } | |
2285 | #endif | |
2286 | ||
20a0307c WF |
2287 | /* |
2288 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2289 | * balanced by operating on them in a round-robin fashion. | |
2290 | * Returns 1 if an adjustment was made. | |
2291 | */ | |
6ae11b27 LS |
2292 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
2293 | int delta) | |
20a0307c | 2294 | { |
b2261026 | 2295 | int nr_nodes, node; |
20a0307c WF |
2296 | |
2297 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 2298 | |
b2261026 JK |
2299 | if (delta < 0) { |
2300 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2301 | if (h->surplus_huge_pages_node[node]) | |
2302 | goto found; | |
e8c5c824 | 2303 | } |
b2261026 JK |
2304 | } else { |
2305 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2306 | if (h->surplus_huge_pages_node[node] < | |
2307 | h->nr_huge_pages_node[node]) | |
2308 | goto found; | |
e8c5c824 | 2309 | } |
b2261026 JK |
2310 | } |
2311 | return 0; | |
20a0307c | 2312 | |
b2261026 JK |
2313 | found: |
2314 | h->surplus_huge_pages += delta; | |
2315 | h->surplus_huge_pages_node[node] += delta; | |
2316 | return 1; | |
20a0307c WF |
2317 | } |
2318 | ||
a5516438 | 2319 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
2320 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
2321 | nodemask_t *nodes_allowed) | |
1da177e4 | 2322 | { |
7893d1d5 | 2323 | unsigned long min_count, ret; |
1da177e4 | 2324 | |
944d9fec | 2325 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
aa888a74 AK |
2326 | return h->max_huge_pages; |
2327 | ||
7893d1d5 AL |
2328 | /* |
2329 | * Increase the pool size | |
2330 | * First take pages out of surplus state. Then make up the | |
2331 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 2332 | * |
d15c7c09 | 2333 | * We might race with __alloc_buddy_huge_page() here and be unable |
d1c3fb1f NA |
2334 | * to convert a surplus huge page to a normal huge page. That is |
2335 | * not critical, though, it just means the overall size of the | |
2336 | * pool might be one hugepage larger than it needs to be, but | |
2337 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 2338 | */ |
1da177e4 | 2339 | spin_lock(&hugetlb_lock); |
a5516438 | 2340 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 2341 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
2342 | break; |
2343 | } | |
2344 | ||
a5516438 | 2345 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
2346 | /* |
2347 | * If this allocation races such that we no longer need the | |
2348 | * page, free_huge_page will handle it by freeing the page | |
2349 | * and reducing the surplus. | |
2350 | */ | |
2351 | spin_unlock(&hugetlb_lock); | |
649920c6 JH |
2352 | |
2353 | /* yield cpu to avoid soft lockup */ | |
2354 | cond_resched(); | |
2355 | ||
944d9fec LC |
2356 | if (hstate_is_gigantic(h)) |
2357 | ret = alloc_fresh_gigantic_page(h, nodes_allowed); | |
2358 | else | |
2359 | ret = alloc_fresh_huge_page(h, nodes_allowed); | |
7893d1d5 AL |
2360 | spin_lock(&hugetlb_lock); |
2361 | if (!ret) | |
2362 | goto out; | |
2363 | ||
536240f2 MG |
2364 | /* Bail for signals. Probably ctrl-c from user */ |
2365 | if (signal_pending(current)) | |
2366 | goto out; | |
7893d1d5 | 2367 | } |
7893d1d5 AL |
2368 | |
2369 | /* | |
2370 | * Decrease the pool size | |
2371 | * First return free pages to the buddy allocator (being careful | |
2372 | * to keep enough around to satisfy reservations). Then place | |
2373 | * pages into surplus state as needed so the pool will shrink | |
2374 | * to the desired size as pages become free. | |
d1c3fb1f NA |
2375 | * |
2376 | * By placing pages into the surplus state independent of the | |
2377 | * overcommit value, we are allowing the surplus pool size to | |
2378 | * exceed overcommit. There are few sane options here. Since | |
d15c7c09 | 2379 | * __alloc_buddy_huge_page() is checking the global counter, |
d1c3fb1f NA |
2380 | * though, we'll note that we're not allowed to exceed surplus |
2381 | * and won't grow the pool anywhere else. Not until one of the | |
2382 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 2383 | */ |
a5516438 | 2384 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 2385 | min_count = max(count, min_count); |
6ae11b27 | 2386 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 2387 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 2388 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 2389 | break; |
55f67141 | 2390 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 2391 | } |
a5516438 | 2392 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 2393 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
2394 | break; |
2395 | } | |
2396 | out: | |
a5516438 | 2397 | ret = persistent_huge_pages(h); |
1da177e4 | 2398 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 2399 | return ret; |
1da177e4 LT |
2400 | } |
2401 | ||
a3437870 NA |
2402 | #define HSTATE_ATTR_RO(_name) \ |
2403 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2404 | ||
2405 | #define HSTATE_ATTR(_name) \ | |
2406 | static struct kobj_attribute _name##_attr = \ | |
2407 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2408 | ||
2409 | static struct kobject *hugepages_kobj; | |
2410 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2411 | ||
9a305230 LS |
2412 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
2413 | ||
2414 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
2415 | { |
2416 | int i; | |
9a305230 | 2417 | |
a3437870 | 2418 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
2419 | if (hstate_kobjs[i] == kobj) { |
2420 | if (nidp) | |
2421 | *nidp = NUMA_NO_NODE; | |
a3437870 | 2422 | return &hstates[i]; |
9a305230 LS |
2423 | } |
2424 | ||
2425 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
2426 | } |
2427 | ||
06808b08 | 2428 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
2429 | struct kobj_attribute *attr, char *buf) |
2430 | { | |
9a305230 LS |
2431 | struct hstate *h; |
2432 | unsigned long nr_huge_pages; | |
2433 | int nid; | |
2434 | ||
2435 | h = kobj_to_hstate(kobj, &nid); | |
2436 | if (nid == NUMA_NO_NODE) | |
2437 | nr_huge_pages = h->nr_huge_pages; | |
2438 | else | |
2439 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2440 | ||
2441 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 2442 | } |
adbe8726 | 2443 | |
238d3c13 DR |
2444 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
2445 | struct hstate *h, int nid, | |
2446 | unsigned long count, size_t len) | |
a3437870 NA |
2447 | { |
2448 | int err; | |
bad44b5b | 2449 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 2450 | |
944d9fec | 2451 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) { |
adbe8726 EM |
2452 | err = -EINVAL; |
2453 | goto out; | |
2454 | } | |
2455 | ||
9a305230 LS |
2456 | if (nid == NUMA_NO_NODE) { |
2457 | /* | |
2458 | * global hstate attribute | |
2459 | */ | |
2460 | if (!(obey_mempolicy && | |
2461 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
2462 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 2463 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 LS |
2464 | } |
2465 | } else if (nodes_allowed) { | |
2466 | /* | |
2467 | * per node hstate attribute: adjust count to global, | |
2468 | * but restrict alloc/free to the specified node. | |
2469 | */ | |
2470 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2471 | init_nodemask_of_node(nodes_allowed, nid); | |
2472 | } else | |
8cebfcd0 | 2473 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 | 2474 | |
06808b08 | 2475 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 2476 | |
8cebfcd0 | 2477 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
2478 | NODEMASK_FREE(nodes_allowed); |
2479 | ||
2480 | return len; | |
adbe8726 EM |
2481 | out: |
2482 | NODEMASK_FREE(nodes_allowed); | |
2483 | return err; | |
06808b08 LS |
2484 | } |
2485 | ||
238d3c13 DR |
2486 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
2487 | struct kobject *kobj, const char *buf, | |
2488 | size_t len) | |
2489 | { | |
2490 | struct hstate *h; | |
2491 | unsigned long count; | |
2492 | int nid; | |
2493 | int err; | |
2494 | ||
2495 | err = kstrtoul(buf, 10, &count); | |
2496 | if (err) | |
2497 | return err; | |
2498 | ||
2499 | h = kobj_to_hstate(kobj, &nid); | |
2500 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2501 | } | |
2502 | ||
06808b08 LS |
2503 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
2504 | struct kobj_attribute *attr, char *buf) | |
2505 | { | |
2506 | return nr_hugepages_show_common(kobj, attr, buf); | |
2507 | } | |
2508 | ||
2509 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2510 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2511 | { | |
238d3c13 | 2512 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
2513 | } |
2514 | HSTATE_ATTR(nr_hugepages); | |
2515 | ||
06808b08 LS |
2516 | #ifdef CONFIG_NUMA |
2517 | ||
2518 | /* | |
2519 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2520 | * huge page alloc/free. | |
2521 | */ | |
2522 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
2523 | struct kobj_attribute *attr, char *buf) | |
2524 | { | |
2525 | return nr_hugepages_show_common(kobj, attr, buf); | |
2526 | } | |
2527 | ||
2528 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2529 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2530 | { | |
238d3c13 | 2531 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
2532 | } |
2533 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2534 | #endif | |
2535 | ||
2536 | ||
a3437870 NA |
2537 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
2538 | struct kobj_attribute *attr, char *buf) | |
2539 | { | |
9a305230 | 2540 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2541 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
2542 | } | |
adbe8726 | 2543 | |
a3437870 NA |
2544 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
2545 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2546 | { | |
2547 | int err; | |
2548 | unsigned long input; | |
9a305230 | 2549 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2550 | |
bae7f4ae | 2551 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2552 | return -EINVAL; |
2553 | ||
3dbb95f7 | 2554 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2555 | if (err) |
73ae31e5 | 2556 | return err; |
a3437870 NA |
2557 | |
2558 | spin_lock(&hugetlb_lock); | |
2559 | h->nr_overcommit_huge_pages = input; | |
2560 | spin_unlock(&hugetlb_lock); | |
2561 | ||
2562 | return count; | |
2563 | } | |
2564 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2565 | ||
2566 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2567 | struct kobj_attribute *attr, char *buf) | |
2568 | { | |
9a305230 LS |
2569 | struct hstate *h; |
2570 | unsigned long free_huge_pages; | |
2571 | int nid; | |
2572 | ||
2573 | h = kobj_to_hstate(kobj, &nid); | |
2574 | if (nid == NUMA_NO_NODE) | |
2575 | free_huge_pages = h->free_huge_pages; | |
2576 | else | |
2577 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2578 | ||
2579 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
2580 | } |
2581 | HSTATE_ATTR_RO(free_hugepages); | |
2582 | ||
2583 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2584 | struct kobj_attribute *attr, char *buf) | |
2585 | { | |
9a305230 | 2586 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2587 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
2588 | } | |
2589 | HSTATE_ATTR_RO(resv_hugepages); | |
2590 | ||
2591 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2592 | struct kobj_attribute *attr, char *buf) | |
2593 | { | |
9a305230 LS |
2594 | struct hstate *h; |
2595 | unsigned long surplus_huge_pages; | |
2596 | int nid; | |
2597 | ||
2598 | h = kobj_to_hstate(kobj, &nid); | |
2599 | if (nid == NUMA_NO_NODE) | |
2600 | surplus_huge_pages = h->surplus_huge_pages; | |
2601 | else | |
2602 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2603 | ||
2604 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
2605 | } |
2606 | HSTATE_ATTR_RO(surplus_hugepages); | |
2607 | ||
2608 | static struct attribute *hstate_attrs[] = { | |
2609 | &nr_hugepages_attr.attr, | |
2610 | &nr_overcommit_hugepages_attr.attr, | |
2611 | &free_hugepages_attr.attr, | |
2612 | &resv_hugepages_attr.attr, | |
2613 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
2614 | #ifdef CONFIG_NUMA |
2615 | &nr_hugepages_mempolicy_attr.attr, | |
2616 | #endif | |
a3437870 NA |
2617 | NULL, |
2618 | }; | |
2619 | ||
2620 | static struct attribute_group hstate_attr_group = { | |
2621 | .attrs = hstate_attrs, | |
2622 | }; | |
2623 | ||
094e9539 JM |
2624 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
2625 | struct kobject **hstate_kobjs, | |
2626 | struct attribute_group *hstate_attr_group) | |
a3437870 NA |
2627 | { |
2628 | int retval; | |
972dc4de | 2629 | int hi = hstate_index(h); |
a3437870 | 2630 | |
9a305230 LS |
2631 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
2632 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
2633 | return -ENOMEM; |
2634 | ||
9a305230 | 2635 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 2636 | if (retval) |
9a305230 | 2637 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
2638 | |
2639 | return retval; | |
2640 | } | |
2641 | ||
2642 | static void __init hugetlb_sysfs_init(void) | |
2643 | { | |
2644 | struct hstate *h; | |
2645 | int err; | |
2646 | ||
2647 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
2648 | if (!hugepages_kobj) | |
2649 | return; | |
2650 | ||
2651 | for_each_hstate(h) { | |
9a305230 LS |
2652 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
2653 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 2654 | if (err) |
ffb22af5 | 2655 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
2656 | } |
2657 | } | |
2658 | ||
9a305230 LS |
2659 | #ifdef CONFIG_NUMA |
2660 | ||
2661 | /* | |
2662 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
2663 | * with node devices in node_devices[] using a parallel array. The array |
2664 | * index of a node device or _hstate == node id. | |
2665 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
2666 | * the base kernel, on the hugetlb module. |
2667 | */ | |
2668 | struct node_hstate { | |
2669 | struct kobject *hugepages_kobj; | |
2670 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2671 | }; | |
b4e289a6 | 2672 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
2673 | |
2674 | /* | |
10fbcf4c | 2675 | * A subset of global hstate attributes for node devices |
9a305230 LS |
2676 | */ |
2677 | static struct attribute *per_node_hstate_attrs[] = { | |
2678 | &nr_hugepages_attr.attr, | |
2679 | &free_hugepages_attr.attr, | |
2680 | &surplus_hugepages_attr.attr, | |
2681 | NULL, | |
2682 | }; | |
2683 | ||
2684 | static struct attribute_group per_node_hstate_attr_group = { | |
2685 | .attrs = per_node_hstate_attrs, | |
2686 | }; | |
2687 | ||
2688 | /* | |
10fbcf4c | 2689 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
2690 | * Returns node id via non-NULL nidp. |
2691 | */ | |
2692 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2693 | { | |
2694 | int nid; | |
2695 | ||
2696 | for (nid = 0; nid < nr_node_ids; nid++) { | |
2697 | struct node_hstate *nhs = &node_hstates[nid]; | |
2698 | int i; | |
2699 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
2700 | if (nhs->hstate_kobjs[i] == kobj) { | |
2701 | if (nidp) | |
2702 | *nidp = nid; | |
2703 | return &hstates[i]; | |
2704 | } | |
2705 | } | |
2706 | ||
2707 | BUG(); | |
2708 | return NULL; | |
2709 | } | |
2710 | ||
2711 | /* | |
10fbcf4c | 2712 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
2713 | * No-op if no hstate attributes attached. |
2714 | */ | |
3cd8b44f | 2715 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
2716 | { |
2717 | struct hstate *h; | |
10fbcf4c | 2718 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2719 | |
2720 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 2721 | return; /* no hstate attributes */ |
9a305230 | 2722 | |
972dc4de AK |
2723 | for_each_hstate(h) { |
2724 | int idx = hstate_index(h); | |
2725 | if (nhs->hstate_kobjs[idx]) { | |
2726 | kobject_put(nhs->hstate_kobjs[idx]); | |
2727 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 2728 | } |
972dc4de | 2729 | } |
9a305230 LS |
2730 | |
2731 | kobject_put(nhs->hugepages_kobj); | |
2732 | nhs->hugepages_kobj = NULL; | |
2733 | } | |
2734 | ||
9a305230 LS |
2735 | |
2736 | /* | |
10fbcf4c | 2737 | * Register hstate attributes for a single node device. |
9a305230 LS |
2738 | * No-op if attributes already registered. |
2739 | */ | |
3cd8b44f | 2740 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
2741 | { |
2742 | struct hstate *h; | |
10fbcf4c | 2743 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2744 | int err; |
2745 | ||
2746 | if (nhs->hugepages_kobj) | |
2747 | return; /* already allocated */ | |
2748 | ||
2749 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 2750 | &node->dev.kobj); |
9a305230 LS |
2751 | if (!nhs->hugepages_kobj) |
2752 | return; | |
2753 | ||
2754 | for_each_hstate(h) { | |
2755 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
2756 | nhs->hstate_kobjs, | |
2757 | &per_node_hstate_attr_group); | |
2758 | if (err) { | |
ffb22af5 AM |
2759 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
2760 | h->name, node->dev.id); | |
9a305230 LS |
2761 | hugetlb_unregister_node(node); |
2762 | break; | |
2763 | } | |
2764 | } | |
2765 | } | |
2766 | ||
2767 | /* | |
9b5e5d0f | 2768 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
2769 | * devices of nodes that have memory. All on-line nodes should have |
2770 | * registered their associated device by this time. | |
9a305230 | 2771 | */ |
7d9ca000 | 2772 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
2773 | { |
2774 | int nid; | |
2775 | ||
8cebfcd0 | 2776 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 2777 | struct node *node = node_devices[nid]; |
10fbcf4c | 2778 | if (node->dev.id == nid) |
9a305230 LS |
2779 | hugetlb_register_node(node); |
2780 | } | |
2781 | ||
2782 | /* | |
10fbcf4c | 2783 | * Let the node device driver know we're here so it can |
9a305230 LS |
2784 | * [un]register hstate attributes on node hotplug. |
2785 | */ | |
2786 | register_hugetlbfs_with_node(hugetlb_register_node, | |
2787 | hugetlb_unregister_node); | |
2788 | } | |
2789 | #else /* !CONFIG_NUMA */ | |
2790 | ||
2791 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2792 | { | |
2793 | BUG(); | |
2794 | if (nidp) | |
2795 | *nidp = -1; | |
2796 | return NULL; | |
2797 | } | |
2798 | ||
9a305230 LS |
2799 | static void hugetlb_register_all_nodes(void) { } |
2800 | ||
2801 | #endif | |
2802 | ||
a3437870 NA |
2803 | static int __init hugetlb_init(void) |
2804 | { | |
8382d914 DB |
2805 | int i; |
2806 | ||
457c1b27 | 2807 | if (!hugepages_supported()) |
0ef89d25 | 2808 | return 0; |
a3437870 | 2809 | |
e11bfbfc NP |
2810 | if (!size_to_hstate(default_hstate_size)) { |
2811 | default_hstate_size = HPAGE_SIZE; | |
2812 | if (!size_to_hstate(default_hstate_size)) | |
2813 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 2814 | } |
972dc4de | 2815 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
f8b74815 VT |
2816 | if (default_hstate_max_huge_pages) { |
2817 | if (!default_hstate.max_huge_pages) | |
2818 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
2819 | } | |
a3437870 NA |
2820 | |
2821 | hugetlb_init_hstates(); | |
aa888a74 | 2822 | gather_bootmem_prealloc(); |
a3437870 NA |
2823 | report_hugepages(); |
2824 | ||
2825 | hugetlb_sysfs_init(); | |
9a305230 | 2826 | hugetlb_register_all_nodes(); |
7179e7bf | 2827 | hugetlb_cgroup_file_init(); |
9a305230 | 2828 | |
8382d914 DB |
2829 | #ifdef CONFIG_SMP |
2830 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
2831 | #else | |
2832 | num_fault_mutexes = 1; | |
2833 | #endif | |
c672c7f2 | 2834 | hugetlb_fault_mutex_table = |
8382d914 | 2835 | kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); |
c672c7f2 | 2836 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
2837 | |
2838 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 2839 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
2840 | return 0; |
2841 | } | |
3e89e1c5 | 2842 | subsys_initcall(hugetlb_init); |
a3437870 NA |
2843 | |
2844 | /* Should be called on processing a hugepagesz=... option */ | |
9fee021d VT |
2845 | void __init hugetlb_bad_size(void) |
2846 | { | |
2847 | parsed_valid_hugepagesz = false; | |
2848 | } | |
2849 | ||
d00181b9 | 2850 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
2851 | { |
2852 | struct hstate *h; | |
8faa8b07 AK |
2853 | unsigned long i; |
2854 | ||
a3437870 | 2855 | if (size_to_hstate(PAGE_SIZE << order)) { |
598d8091 | 2856 | pr_warn("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
2857 | return; |
2858 | } | |
47d38344 | 2859 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 2860 | BUG_ON(order == 0); |
47d38344 | 2861 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
2862 | h->order = order; |
2863 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
2864 | h->nr_huge_pages = 0; |
2865 | h->free_huge_pages = 0; | |
2866 | for (i = 0; i < MAX_NUMNODES; ++i) | |
2867 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 2868 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
2869 | h->next_nid_to_alloc = first_memory_node; |
2870 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
2871 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
2872 | huge_page_size(h)/1024); | |
8faa8b07 | 2873 | |
a3437870 NA |
2874 | parsed_hstate = h; |
2875 | } | |
2876 | ||
e11bfbfc | 2877 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
2878 | { |
2879 | unsigned long *mhp; | |
8faa8b07 | 2880 | static unsigned long *last_mhp; |
a3437870 | 2881 | |
9fee021d VT |
2882 | if (!parsed_valid_hugepagesz) { |
2883 | pr_warn("hugepages = %s preceded by " | |
2884 | "an unsupported hugepagesz, ignoring\n", s); | |
2885 | parsed_valid_hugepagesz = true; | |
2886 | return 1; | |
2887 | } | |
a3437870 | 2888 | /* |
47d38344 | 2889 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
2890 | * so this hugepages= parameter goes to the "default hstate". |
2891 | */ | |
9fee021d | 2892 | else if (!hugetlb_max_hstate) |
a3437870 NA |
2893 | mhp = &default_hstate_max_huge_pages; |
2894 | else | |
2895 | mhp = &parsed_hstate->max_huge_pages; | |
2896 | ||
8faa8b07 | 2897 | if (mhp == last_mhp) { |
598d8091 | 2898 | pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); |
8faa8b07 AK |
2899 | return 1; |
2900 | } | |
2901 | ||
a3437870 NA |
2902 | if (sscanf(s, "%lu", mhp) <= 0) |
2903 | *mhp = 0; | |
2904 | ||
8faa8b07 AK |
2905 | /* |
2906 | * Global state is always initialized later in hugetlb_init. | |
2907 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
2908 | * use the bootmem allocator. | |
2909 | */ | |
47d38344 | 2910 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
2911 | hugetlb_hstate_alloc_pages(parsed_hstate); |
2912 | ||
2913 | last_mhp = mhp; | |
2914 | ||
a3437870 NA |
2915 | return 1; |
2916 | } | |
e11bfbfc NP |
2917 | __setup("hugepages=", hugetlb_nrpages_setup); |
2918 | ||
2919 | static int __init hugetlb_default_setup(char *s) | |
2920 | { | |
2921 | default_hstate_size = memparse(s, &s); | |
2922 | return 1; | |
2923 | } | |
2924 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 2925 | |
8a213460 NA |
2926 | static unsigned int cpuset_mems_nr(unsigned int *array) |
2927 | { | |
2928 | int node; | |
2929 | unsigned int nr = 0; | |
2930 | ||
2931 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
2932 | nr += array[node]; | |
2933 | ||
2934 | return nr; | |
2935 | } | |
2936 | ||
2937 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
2938 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
2939 | struct ctl_table *table, int write, | |
2940 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 2941 | { |
e5ff2159 | 2942 | struct hstate *h = &default_hstate; |
238d3c13 | 2943 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 2944 | int ret; |
e5ff2159 | 2945 | |
457c1b27 | 2946 | if (!hugepages_supported()) |
86613628 | 2947 | return -EOPNOTSUPP; |
457c1b27 | 2948 | |
e5ff2159 AK |
2949 | table->data = &tmp; |
2950 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2951 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2952 | if (ret) | |
2953 | goto out; | |
e5ff2159 | 2954 | |
238d3c13 DR |
2955 | if (write) |
2956 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
2957 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
2958 | out: |
2959 | return ret; | |
1da177e4 | 2960 | } |
396faf03 | 2961 | |
06808b08 LS |
2962 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2963 | void __user *buffer, size_t *length, loff_t *ppos) | |
2964 | { | |
2965 | ||
2966 | return hugetlb_sysctl_handler_common(false, table, write, | |
2967 | buffer, length, ppos); | |
2968 | } | |
2969 | ||
2970 | #ifdef CONFIG_NUMA | |
2971 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2972 | void __user *buffer, size_t *length, loff_t *ppos) | |
2973 | { | |
2974 | return hugetlb_sysctl_handler_common(true, table, write, | |
2975 | buffer, length, ppos); | |
2976 | } | |
2977 | #endif /* CONFIG_NUMA */ | |
2978 | ||
a3d0c6aa | 2979 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 2980 | void __user *buffer, |
a3d0c6aa NA |
2981 | size_t *length, loff_t *ppos) |
2982 | { | |
a5516438 | 2983 | struct hstate *h = &default_hstate; |
e5ff2159 | 2984 | unsigned long tmp; |
08d4a246 | 2985 | int ret; |
e5ff2159 | 2986 | |
457c1b27 | 2987 | if (!hugepages_supported()) |
86613628 | 2988 | return -EOPNOTSUPP; |
457c1b27 | 2989 | |
c033a93c | 2990 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 2991 | |
bae7f4ae | 2992 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
2993 | return -EINVAL; |
2994 | ||
e5ff2159 AK |
2995 | table->data = &tmp; |
2996 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2997 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2998 | if (ret) | |
2999 | goto out; | |
e5ff2159 AK |
3000 | |
3001 | if (write) { | |
3002 | spin_lock(&hugetlb_lock); | |
3003 | h->nr_overcommit_huge_pages = tmp; | |
3004 | spin_unlock(&hugetlb_lock); | |
3005 | } | |
08d4a246 MH |
3006 | out: |
3007 | return ret; | |
a3d0c6aa NA |
3008 | } |
3009 | ||
1da177e4 LT |
3010 | #endif /* CONFIG_SYSCTL */ |
3011 | ||
e1759c21 | 3012 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 3013 | { |
a5516438 | 3014 | struct hstate *h = &default_hstate; |
457c1b27 NA |
3015 | if (!hugepages_supported()) |
3016 | return; | |
e1759c21 | 3017 | seq_printf(m, |
4f98a2fe RR |
3018 | "HugePages_Total: %5lu\n" |
3019 | "HugePages_Free: %5lu\n" | |
3020 | "HugePages_Rsvd: %5lu\n" | |
3021 | "HugePages_Surp: %5lu\n" | |
3022 | "Hugepagesize: %8lu kB\n", | |
a5516438 AK |
3023 | h->nr_huge_pages, |
3024 | h->free_huge_pages, | |
3025 | h->resv_huge_pages, | |
3026 | h->surplus_huge_pages, | |
3027 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
3028 | } |
3029 | ||
3030 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
3031 | { | |
a5516438 | 3032 | struct hstate *h = &default_hstate; |
457c1b27 NA |
3033 | if (!hugepages_supported()) |
3034 | return 0; | |
1da177e4 LT |
3035 | return sprintf(buf, |
3036 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
3037 | "Node %d HugePages_Free: %5u\n" |
3038 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
3039 | nid, h->nr_huge_pages_node[nid], |
3040 | nid, h->free_huge_pages_node[nid], | |
3041 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
3042 | } |
3043 | ||
949f7ec5 DR |
3044 | void hugetlb_show_meminfo(void) |
3045 | { | |
3046 | struct hstate *h; | |
3047 | int nid; | |
3048 | ||
457c1b27 NA |
3049 | if (!hugepages_supported()) |
3050 | return; | |
3051 | ||
949f7ec5 DR |
3052 | for_each_node_state(nid, N_MEMORY) |
3053 | for_each_hstate(h) | |
3054 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
3055 | nid, | |
3056 | h->nr_huge_pages_node[nid], | |
3057 | h->free_huge_pages_node[nid], | |
3058 | h->surplus_huge_pages_node[nid], | |
3059 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
3060 | } | |
3061 | ||
5d317b2b NH |
3062 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
3063 | { | |
3064 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
3065 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
3066 | } | |
3067 | ||
1da177e4 LT |
3068 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
3069 | unsigned long hugetlb_total_pages(void) | |
3070 | { | |
d0028588 WL |
3071 | struct hstate *h; |
3072 | unsigned long nr_total_pages = 0; | |
3073 | ||
3074 | for_each_hstate(h) | |
3075 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
3076 | return nr_total_pages; | |
1da177e4 | 3077 | } |
1da177e4 | 3078 | |
a5516438 | 3079 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
3080 | { |
3081 | int ret = -ENOMEM; | |
3082 | ||
3083 | spin_lock(&hugetlb_lock); | |
3084 | /* | |
3085 | * When cpuset is configured, it breaks the strict hugetlb page | |
3086 | * reservation as the accounting is done on a global variable. Such | |
3087 | * reservation is completely rubbish in the presence of cpuset because | |
3088 | * the reservation is not checked against page availability for the | |
3089 | * current cpuset. Application can still potentially OOM'ed by kernel | |
3090 | * with lack of free htlb page in cpuset that the task is in. | |
3091 | * Attempt to enforce strict accounting with cpuset is almost | |
3092 | * impossible (or too ugly) because cpuset is too fluid that | |
3093 | * task or memory node can be dynamically moved between cpusets. | |
3094 | * | |
3095 | * The change of semantics for shared hugetlb mapping with cpuset is | |
3096 | * undesirable. However, in order to preserve some of the semantics, | |
3097 | * we fall back to check against current free page availability as | |
3098 | * a best attempt and hopefully to minimize the impact of changing | |
3099 | * semantics that cpuset has. | |
3100 | */ | |
3101 | if (delta > 0) { | |
a5516438 | 3102 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
3103 | goto out; |
3104 | ||
a5516438 AK |
3105 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
3106 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
3107 | goto out; |
3108 | } | |
3109 | } | |
3110 | ||
3111 | ret = 0; | |
3112 | if (delta < 0) | |
a5516438 | 3113 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
3114 | |
3115 | out: | |
3116 | spin_unlock(&hugetlb_lock); | |
3117 | return ret; | |
3118 | } | |
3119 | ||
84afd99b AW |
3120 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
3121 | { | |
f522c3ac | 3122 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
3123 | |
3124 | /* | |
3125 | * This new VMA should share its siblings reservation map if present. | |
3126 | * The VMA will only ever have a valid reservation map pointer where | |
3127 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 3128 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
3129 | * after this open call completes. It is therefore safe to take a |
3130 | * new reference here without additional locking. | |
3131 | */ | |
4e35f483 | 3132 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 3133 | kref_get(&resv->refs); |
84afd99b AW |
3134 | } |
3135 | ||
a1e78772 MG |
3136 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
3137 | { | |
a5516438 | 3138 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 3139 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 3140 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 3141 | unsigned long reserve, start, end; |
1c5ecae3 | 3142 | long gbl_reserve; |
84afd99b | 3143 | |
4e35f483 JK |
3144 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3145 | return; | |
84afd99b | 3146 | |
4e35f483 JK |
3147 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
3148 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 3149 | |
4e35f483 | 3150 | reserve = (end - start) - region_count(resv, start, end); |
84afd99b | 3151 | |
4e35f483 JK |
3152 | kref_put(&resv->refs, resv_map_release); |
3153 | ||
3154 | if (reserve) { | |
1c5ecae3 MK |
3155 | /* |
3156 | * Decrement reserve counts. The global reserve count may be | |
3157 | * adjusted if the subpool has a minimum size. | |
3158 | */ | |
3159 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
3160 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 3161 | } |
a1e78772 MG |
3162 | } |
3163 | ||
1da177e4 LT |
3164 | /* |
3165 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
3166 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
3167 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
3168 | * this far. | |
3169 | */ | |
11bac800 | 3170 | static int hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
3171 | { |
3172 | BUG(); | |
d0217ac0 | 3173 | return 0; |
1da177e4 LT |
3174 | } |
3175 | ||
f0f37e2f | 3176 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 3177 | .fault = hugetlb_vm_op_fault, |
84afd99b | 3178 | .open = hugetlb_vm_op_open, |
a1e78772 | 3179 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
3180 | }; |
3181 | ||
1e8f889b DG |
3182 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
3183 | int writable) | |
63551ae0 DG |
3184 | { |
3185 | pte_t entry; | |
3186 | ||
1e8f889b | 3187 | if (writable) { |
106c992a GS |
3188 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
3189 | vma->vm_page_prot))); | |
63551ae0 | 3190 | } else { |
106c992a GS |
3191 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
3192 | vma->vm_page_prot)); | |
63551ae0 DG |
3193 | } |
3194 | entry = pte_mkyoung(entry); | |
3195 | entry = pte_mkhuge(entry); | |
d9ed9faa | 3196 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
3197 | |
3198 | return entry; | |
3199 | } | |
3200 | ||
1e8f889b DG |
3201 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
3202 | unsigned long address, pte_t *ptep) | |
3203 | { | |
3204 | pte_t entry; | |
3205 | ||
106c992a | 3206 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 3207 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 3208 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
3209 | } |
3210 | ||
d5ed7444 | 3211 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
3212 | { |
3213 | swp_entry_t swp; | |
3214 | ||
3215 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 3216 | return false; |
4a705fef NH |
3217 | swp = pte_to_swp_entry(pte); |
3218 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
d5ed7444 | 3219 | return true; |
4a705fef | 3220 | else |
d5ed7444 | 3221 | return false; |
4a705fef NH |
3222 | } |
3223 | ||
3224 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
3225 | { | |
3226 | swp_entry_t swp; | |
3227 | ||
3228 | if (huge_pte_none(pte) || pte_present(pte)) | |
3229 | return 0; | |
3230 | swp = pte_to_swp_entry(pte); | |
3231 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
3232 | return 1; | |
3233 | else | |
3234 | return 0; | |
3235 | } | |
1e8f889b | 3236 | |
63551ae0 DG |
3237 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
3238 | struct vm_area_struct *vma) | |
3239 | { | |
3240 | pte_t *src_pte, *dst_pte, entry; | |
3241 | struct page *ptepage; | |
1c59827d | 3242 | unsigned long addr; |
1e8f889b | 3243 | int cow; |
a5516438 AK |
3244 | struct hstate *h = hstate_vma(vma); |
3245 | unsigned long sz = huge_page_size(h); | |
e8569dd2 AS |
3246 | unsigned long mmun_start; /* For mmu_notifiers */ |
3247 | unsigned long mmun_end; /* For mmu_notifiers */ | |
3248 | int ret = 0; | |
1e8f889b DG |
3249 | |
3250 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 3251 | |
e8569dd2 AS |
3252 | mmun_start = vma->vm_start; |
3253 | mmun_end = vma->vm_end; | |
3254 | if (cow) | |
3255 | mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); | |
3256 | ||
a5516438 | 3257 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 3258 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 3259 | src_pte = huge_pte_offset(src, addr, sz); |
c74df32c HD |
3260 | if (!src_pte) |
3261 | continue; | |
a5516438 | 3262 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
3263 | if (!dst_pte) { |
3264 | ret = -ENOMEM; | |
3265 | break; | |
3266 | } | |
c5c99429 LW |
3267 | |
3268 | /* If the pagetables are shared don't copy or take references */ | |
3269 | if (dst_pte == src_pte) | |
3270 | continue; | |
3271 | ||
cb900f41 KS |
3272 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
3273 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3274 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef NH |
3275 | entry = huge_ptep_get(src_pte); |
3276 | if (huge_pte_none(entry)) { /* skip none entry */ | |
3277 | ; | |
3278 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
3279 | is_hugetlb_entry_hwpoisoned(entry))) { | |
3280 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
3281 | ||
3282 | if (is_write_migration_entry(swp_entry) && cow) { | |
3283 | /* | |
3284 | * COW mappings require pages in both | |
3285 | * parent and child to be set to read. | |
3286 | */ | |
3287 | make_migration_entry_read(&swp_entry); | |
3288 | entry = swp_entry_to_pte(swp_entry); | |
e5251fd4 PA |
3289 | set_huge_swap_pte_at(src, addr, src_pte, |
3290 | entry, sz); | |
4a705fef | 3291 | } |
e5251fd4 | 3292 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); |
4a705fef | 3293 | } else { |
34ee645e | 3294 | if (cow) { |
7f2e9525 | 3295 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e JR |
3296 | mmu_notifier_invalidate_range(src, mmun_start, |
3297 | mmun_end); | |
3298 | } | |
0253d634 | 3299 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
3300 | ptepage = pte_page(entry); |
3301 | get_page(ptepage); | |
53f9263b | 3302 | page_dup_rmap(ptepage, true); |
1c59827d | 3303 | set_huge_pte_at(dst, addr, dst_pte, entry); |
5d317b2b | 3304 | hugetlb_count_add(pages_per_huge_page(h), dst); |
1c59827d | 3305 | } |
cb900f41 KS |
3306 | spin_unlock(src_ptl); |
3307 | spin_unlock(dst_ptl); | |
63551ae0 | 3308 | } |
63551ae0 | 3309 | |
e8569dd2 AS |
3310 | if (cow) |
3311 | mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); | |
3312 | ||
3313 | return ret; | |
63551ae0 DG |
3314 | } |
3315 | ||
24669e58 AK |
3316 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
3317 | unsigned long start, unsigned long end, | |
3318 | struct page *ref_page) | |
63551ae0 DG |
3319 | { |
3320 | struct mm_struct *mm = vma->vm_mm; | |
3321 | unsigned long address; | |
c7546f8f | 3322 | pte_t *ptep; |
63551ae0 | 3323 | pte_t pte; |
cb900f41 | 3324 | spinlock_t *ptl; |
63551ae0 | 3325 | struct page *page; |
a5516438 AK |
3326 | struct hstate *h = hstate_vma(vma); |
3327 | unsigned long sz = huge_page_size(h); | |
2ec74c3e SG |
3328 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
3329 | const unsigned long mmun_end = end; /* For mmu_notifiers */ | |
a5516438 | 3330 | |
63551ae0 | 3331 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
3332 | BUG_ON(start & ~huge_page_mask(h)); |
3333 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 3334 | |
07e32661 AK |
3335 | /* |
3336 | * This is a hugetlb vma, all the pte entries should point | |
3337 | * to huge page. | |
3338 | */ | |
3339 | tlb_remove_check_page_size_change(tlb, sz); | |
24669e58 | 3340 | tlb_start_vma(tlb, vma); |
2ec74c3e | 3341 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
569f48b8 | 3342 | address = start; |
569f48b8 | 3343 | for (; address < end; address += sz) { |
7868a208 | 3344 | ptep = huge_pte_offset(mm, address, sz); |
4c887265 | 3345 | if (!ptep) |
c7546f8f DG |
3346 | continue; |
3347 | ||
cb900f41 | 3348 | ptl = huge_pte_lock(h, mm, ptep); |
31d49da5 AK |
3349 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3350 | spin_unlock(ptl); | |
3351 | continue; | |
3352 | } | |
39dde65c | 3353 | |
6629326b | 3354 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
3355 | if (huge_pte_none(pte)) { |
3356 | spin_unlock(ptl); | |
3357 | continue; | |
3358 | } | |
6629326b HD |
3359 | |
3360 | /* | |
9fbc1f63 NH |
3361 | * Migrating hugepage or HWPoisoned hugepage is already |
3362 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 3363 | */ |
9fbc1f63 | 3364 | if (unlikely(!pte_present(pte))) { |
9386fac3 | 3365 | huge_pte_clear(mm, address, ptep, sz); |
31d49da5 AK |
3366 | spin_unlock(ptl); |
3367 | continue; | |
8c4894c6 | 3368 | } |
6629326b HD |
3369 | |
3370 | page = pte_page(pte); | |
04f2cbe3 MG |
3371 | /* |
3372 | * If a reference page is supplied, it is because a specific | |
3373 | * page is being unmapped, not a range. Ensure the page we | |
3374 | * are about to unmap is the actual page of interest. | |
3375 | */ | |
3376 | if (ref_page) { | |
31d49da5 AK |
3377 | if (page != ref_page) { |
3378 | spin_unlock(ptl); | |
3379 | continue; | |
3380 | } | |
04f2cbe3 MG |
3381 | /* |
3382 | * Mark the VMA as having unmapped its page so that | |
3383 | * future faults in this VMA will fail rather than | |
3384 | * looking like data was lost | |
3385 | */ | |
3386 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
3387 | } | |
3388 | ||
c7546f8f | 3389 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 3390 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 3391 | if (huge_pte_dirty(pte)) |
6649a386 | 3392 | set_page_dirty(page); |
9e81130b | 3393 | |
5d317b2b | 3394 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
d281ee61 | 3395 | page_remove_rmap(page, true); |
31d49da5 | 3396 | |
cb900f41 | 3397 | spin_unlock(ptl); |
e77b0852 | 3398 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
3399 | /* |
3400 | * Bail out after unmapping reference page if supplied | |
3401 | */ | |
3402 | if (ref_page) | |
3403 | break; | |
fe1668ae | 3404 | } |
2ec74c3e | 3405 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
24669e58 | 3406 | tlb_end_vma(tlb, vma); |
1da177e4 | 3407 | } |
63551ae0 | 3408 | |
d833352a MG |
3409 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
3410 | struct vm_area_struct *vma, unsigned long start, | |
3411 | unsigned long end, struct page *ref_page) | |
3412 | { | |
3413 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
3414 | ||
3415 | /* | |
3416 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
3417 | * test will fail on a vma being torn down, and not grab a page table | |
3418 | * on its way out. We're lucky that the flag has such an appropriate | |
3419 | * name, and can in fact be safely cleared here. We could clear it | |
3420 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 3421 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 3422 | * because in the context this is called, the VMA is about to be |
c8c06efa | 3423 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
3424 | */ |
3425 | vma->vm_flags &= ~VM_MAYSHARE; | |
3426 | } | |
3427 | ||
502717f4 | 3428 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 3429 | unsigned long end, struct page *ref_page) |
502717f4 | 3430 | { |
24669e58 AK |
3431 | struct mm_struct *mm; |
3432 | struct mmu_gather tlb; | |
3433 | ||
3434 | mm = vma->vm_mm; | |
3435 | ||
2b047252 | 3436 | tlb_gather_mmu(&tlb, mm, start, end); |
24669e58 AK |
3437 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
3438 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 KC |
3439 | } |
3440 | ||
04f2cbe3 MG |
3441 | /* |
3442 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
3443 | * mappping it owns the reserve page for. The intention is to unmap the page | |
3444 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
3445 | * same region. | |
3446 | */ | |
2f4612af DB |
3447 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
3448 | struct page *page, unsigned long address) | |
04f2cbe3 | 3449 | { |
7526674d | 3450 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
3451 | struct vm_area_struct *iter_vma; |
3452 | struct address_space *mapping; | |
04f2cbe3 MG |
3453 | pgoff_t pgoff; |
3454 | ||
3455 | /* | |
3456 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
3457 | * from page cache lookup which is in HPAGE_SIZE units. | |
3458 | */ | |
7526674d | 3459 | address = address & huge_page_mask(h); |
36e4f20a MH |
3460 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
3461 | vma->vm_pgoff; | |
93c76a3d | 3462 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 3463 | |
4eb2b1dc MG |
3464 | /* |
3465 | * Take the mapping lock for the duration of the table walk. As | |
3466 | * this mapping should be shared between all the VMAs, | |
3467 | * __unmap_hugepage_range() is called as the lock is already held | |
3468 | */ | |
83cde9e8 | 3469 | i_mmap_lock_write(mapping); |
6b2dbba8 | 3470 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
3471 | /* Do not unmap the current VMA */ |
3472 | if (iter_vma == vma) | |
3473 | continue; | |
3474 | ||
2f84a899 MG |
3475 | /* |
3476 | * Shared VMAs have their own reserves and do not affect | |
3477 | * MAP_PRIVATE accounting but it is possible that a shared | |
3478 | * VMA is using the same page so check and skip such VMAs. | |
3479 | */ | |
3480 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
3481 | continue; | |
3482 | ||
04f2cbe3 MG |
3483 | /* |
3484 | * Unmap the page from other VMAs without their own reserves. | |
3485 | * They get marked to be SIGKILLed if they fault in these | |
3486 | * areas. This is because a future no-page fault on this VMA | |
3487 | * could insert a zeroed page instead of the data existing | |
3488 | * from the time of fork. This would look like data corruption | |
3489 | */ | |
3490 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
3491 | unmap_hugepage_range(iter_vma, address, |
3492 | address + huge_page_size(h), page); | |
04f2cbe3 | 3493 | } |
83cde9e8 | 3494 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
3495 | } |
3496 | ||
0fe6e20b NH |
3497 | /* |
3498 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
3499 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
3500 | * cannot race with other handlers or page migration. | |
3501 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 3502 | */ |
1e8f889b | 3503 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
3999f52e AK |
3504 | unsigned long address, pte_t *ptep, |
3505 | struct page *pagecache_page, spinlock_t *ptl) | |
1e8f889b | 3506 | { |
3999f52e | 3507 | pte_t pte; |
a5516438 | 3508 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 3509 | struct page *old_page, *new_page; |
ad4404a2 | 3510 | int ret = 0, outside_reserve = 0; |
2ec74c3e SG |
3511 | unsigned long mmun_start; /* For mmu_notifiers */ |
3512 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1e8f889b | 3513 | |
3999f52e | 3514 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
3515 | old_page = pte_page(pte); |
3516 | ||
04f2cbe3 | 3517 | retry_avoidcopy: |
1e8f889b DG |
3518 | /* If no-one else is actually using this page, avoid the copy |
3519 | * and just make the page writable */ | |
37a2140d | 3520 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
5a49973d | 3521 | page_move_anon_rmap(old_page, vma); |
1e8f889b | 3522 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 3523 | return 0; |
1e8f889b DG |
3524 | } |
3525 | ||
04f2cbe3 MG |
3526 | /* |
3527 | * If the process that created a MAP_PRIVATE mapping is about to | |
3528 | * perform a COW due to a shared page count, attempt to satisfy | |
3529 | * the allocation without using the existing reserves. The pagecache | |
3530 | * page is used to determine if the reserve at this address was | |
3531 | * consumed or not. If reserves were used, a partial faulted mapping | |
3532 | * at the time of fork() could consume its reserves on COW instead | |
3533 | * of the full address range. | |
3534 | */ | |
5944d011 | 3535 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
3536 | old_page != pagecache_page) |
3537 | outside_reserve = 1; | |
3538 | ||
09cbfeaf | 3539 | get_page(old_page); |
b76c8cfb | 3540 | |
ad4404a2 DB |
3541 | /* |
3542 | * Drop page table lock as buddy allocator may be called. It will | |
3543 | * be acquired again before returning to the caller, as expected. | |
3544 | */ | |
cb900f41 | 3545 | spin_unlock(ptl); |
04f2cbe3 | 3546 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 3547 | |
2fc39cec | 3548 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
3549 | /* |
3550 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
3551 | * it is due to references held by a child and an insufficient | |
3552 | * huge page pool. To guarantee the original mappers | |
3553 | * reliability, unmap the page from child processes. The child | |
3554 | * may get SIGKILLed if it later faults. | |
3555 | */ | |
3556 | if (outside_reserve) { | |
09cbfeaf | 3557 | put_page(old_page); |
04f2cbe3 | 3558 | BUG_ON(huge_pte_none(pte)); |
2f4612af DB |
3559 | unmap_ref_private(mm, vma, old_page, address); |
3560 | BUG_ON(huge_pte_none(pte)); | |
3561 | spin_lock(ptl); | |
7868a208 PA |
3562 | ptep = huge_pte_offset(mm, address & huge_page_mask(h), |
3563 | huge_page_size(h)); | |
2f4612af DB |
3564 | if (likely(ptep && |
3565 | pte_same(huge_ptep_get(ptep), pte))) | |
3566 | goto retry_avoidcopy; | |
3567 | /* | |
3568 | * race occurs while re-acquiring page table | |
3569 | * lock, and our job is done. | |
3570 | */ | |
3571 | return 0; | |
04f2cbe3 MG |
3572 | } |
3573 | ||
ad4404a2 DB |
3574 | ret = (PTR_ERR(new_page) == -ENOMEM) ? |
3575 | VM_FAULT_OOM : VM_FAULT_SIGBUS; | |
3576 | goto out_release_old; | |
1e8f889b DG |
3577 | } |
3578 | ||
0fe6e20b NH |
3579 | /* |
3580 | * When the original hugepage is shared one, it does not have | |
3581 | * anon_vma prepared. | |
3582 | */ | |
44e2aa93 | 3583 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
3584 | ret = VM_FAULT_OOM; |
3585 | goto out_release_all; | |
44e2aa93 | 3586 | } |
0fe6e20b | 3587 | |
47ad8475 AA |
3588 | copy_user_huge_page(new_page, old_page, address, vma, |
3589 | pages_per_huge_page(h)); | |
0ed361de | 3590 | __SetPageUptodate(new_page); |
bcc54222 | 3591 | set_page_huge_active(new_page); |
1e8f889b | 3592 | |
2ec74c3e SG |
3593 | mmun_start = address & huge_page_mask(h); |
3594 | mmun_end = mmun_start + huge_page_size(h); | |
3595 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
ad4404a2 | 3596 | |
b76c8cfb | 3597 | /* |
cb900f41 | 3598 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
3599 | * before the page tables are altered |
3600 | */ | |
cb900f41 | 3601 | spin_lock(ptl); |
7868a208 PA |
3602 | ptep = huge_pte_offset(mm, address & huge_page_mask(h), |
3603 | huge_page_size(h)); | |
a9af0c5d | 3604 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
3605 | ClearPagePrivate(new_page); |
3606 | ||
1e8f889b | 3607 | /* Break COW */ |
8fe627ec | 3608 | huge_ptep_clear_flush(vma, address, ptep); |
34ee645e | 3609 | mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); |
1e8f889b DG |
3610 | set_huge_pte_at(mm, address, ptep, |
3611 | make_huge_pte(vma, new_page, 1)); | |
d281ee61 | 3612 | page_remove_rmap(old_page, true); |
cd67f0d2 | 3613 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
3614 | /* Make the old page be freed below */ |
3615 | new_page = old_page; | |
3616 | } | |
cb900f41 | 3617 | spin_unlock(ptl); |
2ec74c3e | 3618 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
ad4404a2 | 3619 | out_release_all: |
96b96a96 | 3620 | restore_reserve_on_error(h, vma, address, new_page); |
09cbfeaf | 3621 | put_page(new_page); |
ad4404a2 | 3622 | out_release_old: |
09cbfeaf | 3623 | put_page(old_page); |
8312034f | 3624 | |
ad4404a2 DB |
3625 | spin_lock(ptl); /* Caller expects lock to be held */ |
3626 | return ret; | |
1e8f889b DG |
3627 | } |
3628 | ||
04f2cbe3 | 3629 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
3630 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
3631 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
3632 | { |
3633 | struct address_space *mapping; | |
e7c4b0bf | 3634 | pgoff_t idx; |
04f2cbe3 MG |
3635 | |
3636 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 3637 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
3638 | |
3639 | return find_lock_page(mapping, idx); | |
3640 | } | |
3641 | ||
3ae77f43 HD |
3642 | /* |
3643 | * Return whether there is a pagecache page to back given address within VMA. | |
3644 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
3645 | */ | |
3646 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
3647 | struct vm_area_struct *vma, unsigned long address) |
3648 | { | |
3649 | struct address_space *mapping; | |
3650 | pgoff_t idx; | |
3651 | struct page *page; | |
3652 | ||
3653 | mapping = vma->vm_file->f_mapping; | |
3654 | idx = vma_hugecache_offset(h, vma, address); | |
3655 | ||
3656 | page = find_get_page(mapping, idx); | |
3657 | if (page) | |
3658 | put_page(page); | |
3659 | return page != NULL; | |
3660 | } | |
3661 | ||
ab76ad54 MK |
3662 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
3663 | pgoff_t idx) | |
3664 | { | |
3665 | struct inode *inode = mapping->host; | |
3666 | struct hstate *h = hstate_inode(inode); | |
3667 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
3668 | ||
3669 | if (err) | |
3670 | return err; | |
3671 | ClearPagePrivate(page); | |
3672 | ||
3673 | spin_lock(&inode->i_lock); | |
3674 | inode->i_blocks += blocks_per_huge_page(h); | |
3675 | spin_unlock(&inode->i_lock); | |
3676 | return 0; | |
3677 | } | |
3678 | ||
a1ed3dda | 3679 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
8382d914 DB |
3680 | struct address_space *mapping, pgoff_t idx, |
3681 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 3682 | { |
a5516438 | 3683 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 3684 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 3685 | int anon_rmap = 0; |
4c887265 | 3686 | unsigned long size; |
4c887265 | 3687 | struct page *page; |
1e8f889b | 3688 | pte_t new_pte; |
cb900f41 | 3689 | spinlock_t *ptl; |
4c887265 | 3690 | |
04f2cbe3 MG |
3691 | /* |
3692 | * Currently, we are forced to kill the process in the event the | |
3693 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 3694 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
3695 | */ |
3696 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 3697 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 3698 | current->pid); |
04f2cbe3 MG |
3699 | return ret; |
3700 | } | |
3701 | ||
4c887265 AL |
3702 | /* |
3703 | * Use page lock to guard against racing truncation | |
3704 | * before we get page_table_lock. | |
3705 | */ | |
6bda666a CL |
3706 | retry: |
3707 | page = find_lock_page(mapping, idx); | |
3708 | if (!page) { | |
a5516438 | 3709 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
3710 | if (idx >= size) |
3711 | goto out; | |
1a1aad8a MK |
3712 | |
3713 | /* | |
3714 | * Check for page in userfault range | |
3715 | */ | |
3716 | if (userfaultfd_missing(vma)) { | |
3717 | u32 hash; | |
3718 | struct vm_fault vmf = { | |
3719 | .vma = vma, | |
3720 | .address = address, | |
3721 | .flags = flags, | |
3722 | /* | |
3723 | * Hard to debug if it ends up being | |
3724 | * used by a callee that assumes | |
3725 | * something about the other | |
3726 | * uninitialized fields... same as in | |
3727 | * memory.c | |
3728 | */ | |
3729 | }; | |
3730 | ||
3731 | /* | |
3732 | * hugetlb_fault_mutex must be dropped before | |
3733 | * handling userfault. Reacquire after handling | |
3734 | * fault to make calling code simpler. | |
3735 | */ | |
3736 | hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, | |
3737 | idx, address); | |
3738 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
3739 | ret = handle_userfault(&vmf, VM_UFFD_MISSING); | |
3740 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
3741 | goto out; | |
3742 | } | |
3743 | ||
04f2cbe3 | 3744 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 3745 | if (IS_ERR(page)) { |
76dcee75 AK |
3746 | ret = PTR_ERR(page); |
3747 | if (ret == -ENOMEM) | |
3748 | ret = VM_FAULT_OOM; | |
3749 | else | |
3750 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
3751 | goto out; |
3752 | } | |
47ad8475 | 3753 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 3754 | __SetPageUptodate(page); |
bcc54222 | 3755 | set_page_huge_active(page); |
ac9b9c66 | 3756 | |
f83a275d | 3757 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 3758 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
3759 | if (err) { |
3760 | put_page(page); | |
6bda666a CL |
3761 | if (err == -EEXIST) |
3762 | goto retry; | |
3763 | goto out; | |
3764 | } | |
23be7468 | 3765 | } else { |
6bda666a | 3766 | lock_page(page); |
0fe6e20b NH |
3767 | if (unlikely(anon_vma_prepare(vma))) { |
3768 | ret = VM_FAULT_OOM; | |
3769 | goto backout_unlocked; | |
3770 | } | |
409eb8c2 | 3771 | anon_rmap = 1; |
23be7468 | 3772 | } |
0fe6e20b | 3773 | } else { |
998b4382 NH |
3774 | /* |
3775 | * If memory error occurs between mmap() and fault, some process | |
3776 | * don't have hwpoisoned swap entry for errored virtual address. | |
3777 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
3778 | */ | |
3779 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 3780 | ret = VM_FAULT_HWPOISON | |
972dc4de | 3781 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
3782 | goto backout_unlocked; |
3783 | } | |
6bda666a | 3784 | } |
1e8f889b | 3785 | |
57303d80 AW |
3786 | /* |
3787 | * If we are going to COW a private mapping later, we examine the | |
3788 | * pending reservations for this page now. This will ensure that | |
3789 | * any allocations necessary to record that reservation occur outside | |
3790 | * the spinlock. | |
3791 | */ | |
5e911373 | 3792 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
2b26736c AW |
3793 | if (vma_needs_reservation(h, vma, address) < 0) { |
3794 | ret = VM_FAULT_OOM; | |
3795 | goto backout_unlocked; | |
3796 | } | |
5e911373 | 3797 | /* Just decrements count, does not deallocate */ |
feba16e2 | 3798 | vma_end_reservation(h, vma, address); |
5e911373 | 3799 | } |
57303d80 | 3800 | |
8bea8052 | 3801 | ptl = huge_pte_lock(h, mm, ptep); |
a5516438 | 3802 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
3803 | if (idx >= size) |
3804 | goto backout; | |
3805 | ||
83c54070 | 3806 | ret = 0; |
7f2e9525 | 3807 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
3808 | goto backout; |
3809 | ||
07443a85 JK |
3810 | if (anon_rmap) { |
3811 | ClearPagePrivate(page); | |
409eb8c2 | 3812 | hugepage_add_new_anon_rmap(page, vma, address); |
ac714904 | 3813 | } else |
53f9263b | 3814 | page_dup_rmap(page, true); |
1e8f889b DG |
3815 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
3816 | && (vma->vm_flags & VM_SHARED))); | |
3817 | set_huge_pte_at(mm, address, ptep, new_pte); | |
3818 | ||
5d317b2b | 3819 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 3820 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 3821 | /* Optimization, do the COW without a second fault */ |
3999f52e | 3822 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); |
1e8f889b DG |
3823 | } |
3824 | ||
cb900f41 | 3825 | spin_unlock(ptl); |
4c887265 AL |
3826 | unlock_page(page); |
3827 | out: | |
ac9b9c66 | 3828 | return ret; |
4c887265 AL |
3829 | |
3830 | backout: | |
cb900f41 | 3831 | spin_unlock(ptl); |
2b26736c | 3832 | backout_unlocked: |
4c887265 | 3833 | unlock_page(page); |
96b96a96 | 3834 | restore_reserve_on_error(h, vma, address, page); |
4c887265 AL |
3835 | put_page(page); |
3836 | goto out; | |
ac9b9c66 HD |
3837 | } |
3838 | ||
8382d914 | 3839 | #ifdef CONFIG_SMP |
c672c7f2 | 3840 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
8382d914 DB |
3841 | struct vm_area_struct *vma, |
3842 | struct address_space *mapping, | |
3843 | pgoff_t idx, unsigned long address) | |
3844 | { | |
3845 | unsigned long key[2]; | |
3846 | u32 hash; | |
3847 | ||
3848 | if (vma->vm_flags & VM_SHARED) { | |
3849 | key[0] = (unsigned long) mapping; | |
3850 | key[1] = idx; | |
3851 | } else { | |
3852 | key[0] = (unsigned long) mm; | |
3853 | key[1] = address >> huge_page_shift(h); | |
3854 | } | |
3855 | ||
3856 | hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); | |
3857 | ||
3858 | return hash & (num_fault_mutexes - 1); | |
3859 | } | |
3860 | #else | |
3861 | /* | |
3862 | * For uniprocesor systems we always use a single mutex, so just | |
3863 | * return 0 and avoid the hashing overhead. | |
3864 | */ | |
c672c7f2 | 3865 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
8382d914 DB |
3866 | struct vm_area_struct *vma, |
3867 | struct address_space *mapping, | |
3868 | pgoff_t idx, unsigned long address) | |
3869 | { | |
3870 | return 0; | |
3871 | } | |
3872 | #endif | |
3873 | ||
86e5216f | 3874 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 3875 | unsigned long address, unsigned int flags) |
86e5216f | 3876 | { |
8382d914 | 3877 | pte_t *ptep, entry; |
cb900f41 | 3878 | spinlock_t *ptl; |
1e8f889b | 3879 | int ret; |
8382d914 DB |
3880 | u32 hash; |
3881 | pgoff_t idx; | |
0fe6e20b | 3882 | struct page *page = NULL; |
57303d80 | 3883 | struct page *pagecache_page = NULL; |
a5516438 | 3884 | struct hstate *h = hstate_vma(vma); |
8382d914 | 3885 | struct address_space *mapping; |
0f792cf9 | 3886 | int need_wait_lock = 0; |
86e5216f | 3887 | |
1e16a539 KH |
3888 | address &= huge_page_mask(h); |
3889 | ||
7868a208 | 3890 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
fd6a03ed NH |
3891 | if (ptep) { |
3892 | entry = huge_ptep_get(ptep); | |
290408d4 | 3893 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 3894 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
3895 | return 0; |
3896 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 3897 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 3898 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
0d777df5 NH |
3899 | } else { |
3900 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); | |
3901 | if (!ptep) | |
3902 | return VM_FAULT_OOM; | |
fd6a03ed NH |
3903 | } |
3904 | ||
8382d914 DB |
3905 | mapping = vma->vm_file->f_mapping; |
3906 | idx = vma_hugecache_offset(h, vma, address); | |
3907 | ||
3935baa9 DG |
3908 | /* |
3909 | * Serialize hugepage allocation and instantiation, so that we don't | |
3910 | * get spurious allocation failures if two CPUs race to instantiate | |
3911 | * the same page in the page cache. | |
3912 | */ | |
c672c7f2 MK |
3913 | hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address); |
3914 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
8382d914 | 3915 | |
7f2e9525 GS |
3916 | entry = huge_ptep_get(ptep); |
3917 | if (huge_pte_none(entry)) { | |
8382d914 | 3918 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 3919 | goto out_mutex; |
3935baa9 | 3920 | } |
86e5216f | 3921 | |
83c54070 | 3922 | ret = 0; |
1e8f889b | 3923 | |
0f792cf9 NH |
3924 | /* |
3925 | * entry could be a migration/hwpoison entry at this point, so this | |
3926 | * check prevents the kernel from going below assuming that we have | |
3927 | * a active hugepage in pagecache. This goto expects the 2nd page fault, | |
3928 | * and is_hugetlb_entry_(migration|hwpoisoned) check will properly | |
3929 | * handle it. | |
3930 | */ | |
3931 | if (!pte_present(entry)) | |
3932 | goto out_mutex; | |
3933 | ||
57303d80 AW |
3934 | /* |
3935 | * If we are going to COW the mapping later, we examine the pending | |
3936 | * reservations for this page now. This will ensure that any | |
3937 | * allocations necessary to record that reservation occur outside the | |
3938 | * spinlock. For private mappings, we also lookup the pagecache | |
3939 | * page now as it is used to determine if a reservation has been | |
3940 | * consumed. | |
3941 | */ | |
106c992a | 3942 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
2b26736c AW |
3943 | if (vma_needs_reservation(h, vma, address) < 0) { |
3944 | ret = VM_FAULT_OOM; | |
b4d1d99f | 3945 | goto out_mutex; |
2b26736c | 3946 | } |
5e911373 | 3947 | /* Just decrements count, does not deallocate */ |
feba16e2 | 3948 | vma_end_reservation(h, vma, address); |
57303d80 | 3949 | |
f83a275d | 3950 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
3951 | pagecache_page = hugetlbfs_pagecache_page(h, |
3952 | vma, address); | |
3953 | } | |
3954 | ||
0f792cf9 NH |
3955 | ptl = huge_pte_lock(h, mm, ptep); |
3956 | ||
3957 | /* Check for a racing update before calling hugetlb_cow */ | |
3958 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
3959 | goto out_ptl; | |
3960 | ||
56c9cfb1 NH |
3961 | /* |
3962 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
3963 | * pagecache_page, so here we need take the former one | |
3964 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
3965 | */ |
3966 | page = pte_page(entry); | |
3967 | if (page != pagecache_page) | |
0f792cf9 NH |
3968 | if (!trylock_page(page)) { |
3969 | need_wait_lock = 1; | |
3970 | goto out_ptl; | |
3971 | } | |
b4d1d99f | 3972 | |
0f792cf9 | 3973 | get_page(page); |
b4d1d99f | 3974 | |
788c7df4 | 3975 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 3976 | if (!huge_pte_write(entry)) { |
3999f52e AK |
3977 | ret = hugetlb_cow(mm, vma, address, ptep, |
3978 | pagecache_page, ptl); | |
0f792cf9 | 3979 | goto out_put_page; |
b4d1d99f | 3980 | } |
106c992a | 3981 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
3982 | } |
3983 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
3984 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
3985 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 3986 | update_mmu_cache(vma, address, ptep); |
0f792cf9 NH |
3987 | out_put_page: |
3988 | if (page != pagecache_page) | |
3989 | unlock_page(page); | |
3990 | put_page(page); | |
cb900f41 KS |
3991 | out_ptl: |
3992 | spin_unlock(ptl); | |
57303d80 AW |
3993 | |
3994 | if (pagecache_page) { | |
3995 | unlock_page(pagecache_page); | |
3996 | put_page(pagecache_page); | |
3997 | } | |
b4d1d99f | 3998 | out_mutex: |
c672c7f2 | 3999 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
0f792cf9 NH |
4000 | /* |
4001 | * Generally it's safe to hold refcount during waiting page lock. But | |
4002 | * here we just wait to defer the next page fault to avoid busy loop and | |
4003 | * the page is not used after unlocked before returning from the current | |
4004 | * page fault. So we are safe from accessing freed page, even if we wait | |
4005 | * here without taking refcount. | |
4006 | */ | |
4007 | if (need_wait_lock) | |
4008 | wait_on_page_locked(page); | |
1e8f889b | 4009 | return ret; |
86e5216f AL |
4010 | } |
4011 | ||
8fb5debc MK |
4012 | /* |
4013 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
4014 | * modifications for huge pages. | |
4015 | */ | |
4016 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
4017 | pte_t *dst_pte, | |
4018 | struct vm_area_struct *dst_vma, | |
4019 | unsigned long dst_addr, | |
4020 | unsigned long src_addr, | |
4021 | struct page **pagep) | |
4022 | { | |
1c9e8def | 4023 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
4024 | struct hstate *h = hstate_vma(dst_vma); |
4025 | pte_t _dst_pte; | |
4026 | spinlock_t *ptl; | |
4027 | int ret; | |
4028 | struct page *page; | |
4029 | ||
4030 | if (!*pagep) { | |
4031 | ret = -ENOMEM; | |
4032 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
4033 | if (IS_ERR(page)) | |
4034 | goto out; | |
4035 | ||
4036 | ret = copy_huge_page_from_user(page, | |
4037 | (const void __user *) src_addr, | |
810a56b9 | 4038 | pages_per_huge_page(h), false); |
8fb5debc MK |
4039 | |
4040 | /* fallback to copy_from_user outside mmap_sem */ | |
4041 | if (unlikely(ret)) { | |
4042 | ret = -EFAULT; | |
4043 | *pagep = page; | |
4044 | /* don't free the page */ | |
4045 | goto out; | |
4046 | } | |
4047 | } else { | |
4048 | page = *pagep; | |
4049 | *pagep = NULL; | |
4050 | } | |
4051 | ||
4052 | /* | |
4053 | * The memory barrier inside __SetPageUptodate makes sure that | |
4054 | * preceding stores to the page contents become visible before | |
4055 | * the set_pte_at() write. | |
4056 | */ | |
4057 | __SetPageUptodate(page); | |
4058 | set_page_huge_active(page); | |
4059 | ||
1c9e8def MK |
4060 | /* |
4061 | * If shared, add to page cache | |
4062 | */ | |
4063 | if (vm_shared) { | |
4064 | struct address_space *mapping = dst_vma->vm_file->f_mapping; | |
4065 | pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
4066 | ||
4067 | ret = huge_add_to_page_cache(page, mapping, idx); | |
4068 | if (ret) | |
4069 | goto out_release_nounlock; | |
4070 | } | |
4071 | ||
8fb5debc MK |
4072 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); |
4073 | spin_lock(ptl); | |
4074 | ||
4075 | ret = -EEXIST; | |
4076 | if (!huge_pte_none(huge_ptep_get(dst_pte))) | |
4077 | goto out_release_unlock; | |
4078 | ||
1c9e8def MK |
4079 | if (vm_shared) { |
4080 | page_dup_rmap(page, true); | |
4081 | } else { | |
4082 | ClearPagePrivate(page); | |
4083 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); | |
4084 | } | |
8fb5debc MK |
4085 | |
4086 | _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); | |
4087 | if (dst_vma->vm_flags & VM_WRITE) | |
4088 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
4089 | _dst_pte = pte_mkyoung(_dst_pte); | |
4090 | ||
4091 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); | |
4092 | ||
4093 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
4094 | dst_vma->vm_flags & VM_WRITE); | |
4095 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
4096 | ||
4097 | /* No need to invalidate - it was non-present before */ | |
4098 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
4099 | ||
4100 | spin_unlock(ptl); | |
1c9e8def MK |
4101 | if (vm_shared) |
4102 | unlock_page(page); | |
8fb5debc MK |
4103 | ret = 0; |
4104 | out: | |
4105 | return ret; | |
4106 | out_release_unlock: | |
4107 | spin_unlock(ptl); | |
1c9e8def MK |
4108 | out_release_nounlock: |
4109 | if (vm_shared) | |
4110 | unlock_page(page); | |
8fb5debc MK |
4111 | put_page(page); |
4112 | goto out; | |
4113 | } | |
4114 | ||
28a35716 ML |
4115 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
4116 | struct page **pages, struct vm_area_struct **vmas, | |
4117 | unsigned long *position, unsigned long *nr_pages, | |
87ffc118 | 4118 | long i, unsigned int flags, int *nonblocking) |
63551ae0 | 4119 | { |
d5d4b0aa KC |
4120 | unsigned long pfn_offset; |
4121 | unsigned long vaddr = *position; | |
28a35716 | 4122 | unsigned long remainder = *nr_pages; |
a5516438 | 4123 | struct hstate *h = hstate_vma(vma); |
63551ae0 | 4124 | |
63551ae0 | 4125 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 4126 | pte_t *pte; |
cb900f41 | 4127 | spinlock_t *ptl = NULL; |
2a15efc9 | 4128 | int absent; |
4c887265 | 4129 | struct page *page; |
63551ae0 | 4130 | |
02057967 DR |
4131 | /* |
4132 | * If we have a pending SIGKILL, don't keep faulting pages and | |
4133 | * potentially allocating memory. | |
4134 | */ | |
4135 | if (unlikely(fatal_signal_pending(current))) { | |
4136 | remainder = 0; | |
4137 | break; | |
4138 | } | |
4139 | ||
4c887265 AL |
4140 | /* |
4141 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 4142 | * each hugepage. We have to make sure we get the |
4c887265 | 4143 | * first, for the page indexing below to work. |
cb900f41 KS |
4144 | * |
4145 | * Note that page table lock is not held when pte is null. | |
4c887265 | 4146 | */ |
7868a208 PA |
4147 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
4148 | huge_page_size(h)); | |
cb900f41 KS |
4149 | if (pte) |
4150 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
4151 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
4152 | ||
4153 | /* | |
4154 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
4155 | * an error where there's an empty slot with no huge pagecache |
4156 | * to back it. This way, we avoid allocating a hugepage, and | |
4157 | * the sparse dumpfile avoids allocating disk blocks, but its | |
4158 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 4159 | */ |
3ae77f43 HD |
4160 | if (absent && (flags & FOLL_DUMP) && |
4161 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
4162 | if (pte) |
4163 | spin_unlock(ptl); | |
2a15efc9 HD |
4164 | remainder = 0; |
4165 | break; | |
4166 | } | |
63551ae0 | 4167 | |
9cc3a5bd NH |
4168 | /* |
4169 | * We need call hugetlb_fault for both hugepages under migration | |
4170 | * (in which case hugetlb_fault waits for the migration,) and | |
4171 | * hwpoisoned hugepages (in which case we need to prevent the | |
4172 | * caller from accessing to them.) In order to do this, we use | |
4173 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
4174 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
4175 | * both cases, and because we can't follow correct pages | |
4176 | * directly from any kind of swap entries. | |
4177 | */ | |
4178 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
4179 | ((flags & FOLL_WRITE) && |
4180 | !huge_pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 4181 | int ret; |
87ffc118 | 4182 | unsigned int fault_flags = 0; |
63551ae0 | 4183 | |
cb900f41 KS |
4184 | if (pte) |
4185 | spin_unlock(ptl); | |
87ffc118 AA |
4186 | if (flags & FOLL_WRITE) |
4187 | fault_flags |= FAULT_FLAG_WRITE; | |
4188 | if (nonblocking) | |
4189 | fault_flags |= FAULT_FLAG_ALLOW_RETRY; | |
4190 | if (flags & FOLL_NOWAIT) | |
4191 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
4192 | FAULT_FLAG_RETRY_NOWAIT; | |
4193 | if (flags & FOLL_TRIED) { | |
4194 | VM_WARN_ON_ONCE(fault_flags & | |
4195 | FAULT_FLAG_ALLOW_RETRY); | |
4196 | fault_flags |= FAULT_FLAG_TRIED; | |
4197 | } | |
4198 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
4199 | if (ret & VM_FAULT_ERROR) { | |
9a291a7c JM |
4200 | int err = vm_fault_to_errno(ret, flags); |
4201 | ||
4202 | if (err) | |
4203 | return err; | |
4204 | ||
87ffc118 AA |
4205 | remainder = 0; |
4206 | break; | |
4207 | } | |
4208 | if (ret & VM_FAULT_RETRY) { | |
4209 | if (nonblocking) | |
4210 | *nonblocking = 0; | |
4211 | *nr_pages = 0; | |
4212 | /* | |
4213 | * VM_FAULT_RETRY must not return an | |
4214 | * error, it will return zero | |
4215 | * instead. | |
4216 | * | |
4217 | * No need to update "position" as the | |
4218 | * caller will not check it after | |
4219 | * *nr_pages is set to 0. | |
4220 | */ | |
4221 | return i; | |
4222 | } | |
4223 | continue; | |
4c887265 AL |
4224 | } |
4225 | ||
a5516438 | 4226 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 4227 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 4228 | same_page: |
d6692183 | 4229 | if (pages) { |
2a15efc9 | 4230 | pages[i] = mem_map_offset(page, pfn_offset); |
ddc58f27 | 4231 | get_page(pages[i]); |
d6692183 | 4232 | } |
63551ae0 DG |
4233 | |
4234 | if (vmas) | |
4235 | vmas[i] = vma; | |
4236 | ||
4237 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 4238 | ++pfn_offset; |
63551ae0 DG |
4239 | --remainder; |
4240 | ++i; | |
d5d4b0aa | 4241 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 4242 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa KC |
4243 | /* |
4244 | * We use pfn_offset to avoid touching the pageframes | |
4245 | * of this compound page. | |
4246 | */ | |
4247 | goto same_page; | |
4248 | } | |
cb900f41 | 4249 | spin_unlock(ptl); |
63551ae0 | 4250 | } |
28a35716 | 4251 | *nr_pages = remainder; |
87ffc118 AA |
4252 | /* |
4253 | * setting position is actually required only if remainder is | |
4254 | * not zero but it's faster not to add a "if (remainder)" | |
4255 | * branch. | |
4256 | */ | |
63551ae0 DG |
4257 | *position = vaddr; |
4258 | ||
2a15efc9 | 4259 | return i ? i : -EFAULT; |
63551ae0 | 4260 | } |
8f860591 | 4261 | |
5491ae7b AK |
4262 | #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE |
4263 | /* | |
4264 | * ARCHes with special requirements for evicting HUGETLB backing TLB entries can | |
4265 | * implement this. | |
4266 | */ | |
4267 | #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) | |
4268 | #endif | |
4269 | ||
7da4d641 | 4270 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
4271 | unsigned long address, unsigned long end, pgprot_t newprot) |
4272 | { | |
4273 | struct mm_struct *mm = vma->vm_mm; | |
4274 | unsigned long start = address; | |
4275 | pte_t *ptep; | |
4276 | pte_t pte; | |
a5516438 | 4277 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 4278 | unsigned long pages = 0; |
8f860591 ZY |
4279 | |
4280 | BUG_ON(address >= end); | |
4281 | flush_cache_range(vma, address, end); | |
4282 | ||
a5338093 | 4283 | mmu_notifier_invalidate_range_start(mm, start, end); |
83cde9e8 | 4284 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 4285 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 4286 | spinlock_t *ptl; |
7868a208 | 4287 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
8f860591 ZY |
4288 | if (!ptep) |
4289 | continue; | |
cb900f41 | 4290 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
4291 | if (huge_pmd_unshare(mm, &address, ptep)) { |
4292 | pages++; | |
cb900f41 | 4293 | spin_unlock(ptl); |
39dde65c | 4294 | continue; |
7da4d641 | 4295 | } |
a8bda28d NH |
4296 | pte = huge_ptep_get(ptep); |
4297 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
4298 | spin_unlock(ptl); | |
4299 | continue; | |
4300 | } | |
4301 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
4302 | swp_entry_t entry = pte_to_swp_entry(pte); | |
4303 | ||
4304 | if (is_write_migration_entry(entry)) { | |
4305 | pte_t newpte; | |
4306 | ||
4307 | make_migration_entry_read(&entry); | |
4308 | newpte = swp_entry_to_pte(entry); | |
e5251fd4 PA |
4309 | set_huge_swap_pte_at(mm, address, ptep, |
4310 | newpte, huge_page_size(h)); | |
a8bda28d NH |
4311 | pages++; |
4312 | } | |
4313 | spin_unlock(ptl); | |
4314 | continue; | |
4315 | } | |
4316 | if (!huge_pte_none(pte)) { | |
8f860591 | 4317 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
106c992a | 4318 | pte = pte_mkhuge(huge_pte_modify(pte, newprot)); |
be7517d6 | 4319 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
8f860591 | 4320 | set_huge_pte_at(mm, address, ptep, pte); |
7da4d641 | 4321 | pages++; |
8f860591 | 4322 | } |
cb900f41 | 4323 | spin_unlock(ptl); |
8f860591 | 4324 | } |
d833352a | 4325 | /* |
c8c06efa | 4326 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 4327 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 4328 | * once we release i_mmap_rwsem, another task can do the final put_page |
d833352a MG |
4329 | * and that page table be reused and filled with junk. |
4330 | */ | |
5491ae7b | 4331 | flush_hugetlb_tlb_range(vma, start, end); |
34ee645e | 4332 | mmu_notifier_invalidate_range(mm, start, end); |
83cde9e8 | 4333 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
a5338093 | 4334 | mmu_notifier_invalidate_range_end(mm, start, end); |
7da4d641 PZ |
4335 | |
4336 | return pages << h->order; | |
8f860591 ZY |
4337 | } |
4338 | ||
a1e78772 MG |
4339 | int hugetlb_reserve_pages(struct inode *inode, |
4340 | long from, long to, | |
5a6fe125 | 4341 | struct vm_area_struct *vma, |
ca16d140 | 4342 | vm_flags_t vm_flags) |
e4e574b7 | 4343 | { |
17c9d12e | 4344 | long ret, chg; |
a5516438 | 4345 | struct hstate *h = hstate_inode(inode); |
90481622 | 4346 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 4347 | struct resv_map *resv_map; |
1c5ecae3 | 4348 | long gbl_reserve; |
e4e574b7 | 4349 | |
17c9d12e MG |
4350 | /* |
4351 | * Only apply hugepage reservation if asked. At fault time, an | |
4352 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 4353 | * without using reserves |
17c9d12e | 4354 | */ |
ca16d140 | 4355 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
4356 | return 0; |
4357 | ||
a1e78772 MG |
4358 | /* |
4359 | * Shared mappings base their reservation on the number of pages that | |
4360 | * are already allocated on behalf of the file. Private mappings need | |
4361 | * to reserve the full area even if read-only as mprotect() may be | |
4362 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
4363 | */ | |
9119a41e | 4364 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4e35f483 | 4365 | resv_map = inode_resv_map(inode); |
9119a41e | 4366 | |
1406ec9b | 4367 | chg = region_chg(resv_map, from, to); |
9119a41e JK |
4368 | |
4369 | } else { | |
4370 | resv_map = resv_map_alloc(); | |
17c9d12e MG |
4371 | if (!resv_map) |
4372 | return -ENOMEM; | |
4373 | ||
a1e78772 | 4374 | chg = to - from; |
84afd99b | 4375 | |
17c9d12e MG |
4376 | set_vma_resv_map(vma, resv_map); |
4377 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
4378 | } | |
4379 | ||
c50ac050 DH |
4380 | if (chg < 0) { |
4381 | ret = chg; | |
4382 | goto out_err; | |
4383 | } | |
8a630112 | 4384 | |
1c5ecae3 MK |
4385 | /* |
4386 | * There must be enough pages in the subpool for the mapping. If | |
4387 | * the subpool has a minimum size, there may be some global | |
4388 | * reservations already in place (gbl_reserve). | |
4389 | */ | |
4390 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
4391 | if (gbl_reserve < 0) { | |
c50ac050 DH |
4392 | ret = -ENOSPC; |
4393 | goto out_err; | |
4394 | } | |
5a6fe125 MG |
4395 | |
4396 | /* | |
17c9d12e | 4397 | * Check enough hugepages are available for the reservation. |
90481622 | 4398 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 4399 | */ |
1c5ecae3 | 4400 | ret = hugetlb_acct_memory(h, gbl_reserve); |
68842c9b | 4401 | if (ret < 0) { |
1c5ecae3 MK |
4402 | /* put back original number of pages, chg */ |
4403 | (void)hugepage_subpool_put_pages(spool, chg); | |
c50ac050 | 4404 | goto out_err; |
68842c9b | 4405 | } |
17c9d12e MG |
4406 | |
4407 | /* | |
4408 | * Account for the reservations made. Shared mappings record regions | |
4409 | * that have reservations as they are shared by multiple VMAs. | |
4410 | * When the last VMA disappears, the region map says how much | |
4411 | * the reservation was and the page cache tells how much of | |
4412 | * the reservation was consumed. Private mappings are per-VMA and | |
4413 | * only the consumed reservations are tracked. When the VMA | |
4414 | * disappears, the original reservation is the VMA size and the | |
4415 | * consumed reservations are stored in the map. Hence, nothing | |
4416 | * else has to be done for private mappings here | |
4417 | */ | |
33039678 MK |
4418 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4419 | long add = region_add(resv_map, from, to); | |
4420 | ||
4421 | if (unlikely(chg > add)) { | |
4422 | /* | |
4423 | * pages in this range were added to the reserve | |
4424 | * map between region_chg and region_add. This | |
4425 | * indicates a race with alloc_huge_page. Adjust | |
4426 | * the subpool and reserve counts modified above | |
4427 | * based on the difference. | |
4428 | */ | |
4429 | long rsv_adjust; | |
4430 | ||
4431 | rsv_adjust = hugepage_subpool_put_pages(spool, | |
4432 | chg - add); | |
4433 | hugetlb_acct_memory(h, -rsv_adjust); | |
4434 | } | |
4435 | } | |
a43a8c39 | 4436 | return 0; |
c50ac050 | 4437 | out_err: |
5e911373 | 4438 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
ff8c0c53 MK |
4439 | /* Don't call region_abort if region_chg failed */ |
4440 | if (chg >= 0) | |
4441 | region_abort(resv_map, from, to); | |
f031dd27 JK |
4442 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
4443 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 4444 | return ret; |
a43a8c39 KC |
4445 | } |
4446 | ||
b5cec28d MK |
4447 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
4448 | long freed) | |
a43a8c39 | 4449 | { |
a5516438 | 4450 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 4451 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 4452 | long chg = 0; |
90481622 | 4453 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 4454 | long gbl_reserve; |
45c682a6 | 4455 | |
b5cec28d MK |
4456 | if (resv_map) { |
4457 | chg = region_del(resv_map, start, end); | |
4458 | /* | |
4459 | * region_del() can fail in the rare case where a region | |
4460 | * must be split and another region descriptor can not be | |
4461 | * allocated. If end == LONG_MAX, it will not fail. | |
4462 | */ | |
4463 | if (chg < 0) | |
4464 | return chg; | |
4465 | } | |
4466 | ||
45c682a6 | 4467 | spin_lock(&inode->i_lock); |
e4c6f8be | 4468 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
4469 | spin_unlock(&inode->i_lock); |
4470 | ||
1c5ecae3 MK |
4471 | /* |
4472 | * If the subpool has a minimum size, the number of global | |
4473 | * reservations to be released may be adjusted. | |
4474 | */ | |
4475 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
4476 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
4477 | |
4478 | return 0; | |
a43a8c39 | 4479 | } |
93f70f90 | 4480 | |
3212b535 SC |
4481 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
4482 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
4483 | struct vm_area_struct *vma, | |
4484 | unsigned long addr, pgoff_t idx) | |
4485 | { | |
4486 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
4487 | svma->vm_start; | |
4488 | unsigned long sbase = saddr & PUD_MASK; | |
4489 | unsigned long s_end = sbase + PUD_SIZE; | |
4490 | ||
4491 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
4492 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
4493 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
4494 | |
4495 | /* | |
4496 | * match the virtual addresses, permission and the alignment of the | |
4497 | * page table page. | |
4498 | */ | |
4499 | if (pmd_index(addr) != pmd_index(saddr) || | |
4500 | vm_flags != svm_flags || | |
4501 | sbase < svma->vm_start || svma->vm_end < s_end) | |
4502 | return 0; | |
4503 | ||
4504 | return saddr; | |
4505 | } | |
4506 | ||
31aafb45 | 4507 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
4508 | { |
4509 | unsigned long base = addr & PUD_MASK; | |
4510 | unsigned long end = base + PUD_SIZE; | |
4511 | ||
4512 | /* | |
4513 | * check on proper vm_flags and page table alignment | |
4514 | */ | |
4515 | if (vma->vm_flags & VM_MAYSHARE && | |
4516 | vma->vm_start <= base && end <= vma->vm_end) | |
31aafb45 NK |
4517 | return true; |
4518 | return false; | |
3212b535 SC |
4519 | } |
4520 | ||
4521 | /* | |
4522 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
4523 | * and returns the corresponding pte. While this is not necessary for the | |
4524 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
4525 | * code much cleaner. pmd allocation is essential for the shared case because | |
c8c06efa | 4526 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise |
3212b535 SC |
4527 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a |
4528 | * bad pmd for sharing. | |
4529 | */ | |
4530 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4531 | { | |
4532 | struct vm_area_struct *vma = find_vma(mm, addr); | |
4533 | struct address_space *mapping = vma->vm_file->f_mapping; | |
4534 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
4535 | vma->vm_pgoff; | |
4536 | struct vm_area_struct *svma; | |
4537 | unsigned long saddr; | |
4538 | pte_t *spte = NULL; | |
4539 | pte_t *pte; | |
cb900f41 | 4540 | spinlock_t *ptl; |
3212b535 SC |
4541 | |
4542 | if (!vma_shareable(vma, addr)) | |
4543 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
4544 | ||
83cde9e8 | 4545 | i_mmap_lock_write(mapping); |
3212b535 SC |
4546 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
4547 | if (svma == vma) | |
4548 | continue; | |
4549 | ||
4550 | saddr = page_table_shareable(svma, vma, addr, idx); | |
4551 | if (saddr) { | |
7868a208 PA |
4552 | spte = huge_pte_offset(svma->vm_mm, saddr, |
4553 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
4554 | if (spte) { |
4555 | get_page(virt_to_page(spte)); | |
4556 | break; | |
4557 | } | |
4558 | } | |
4559 | } | |
4560 | ||
4561 | if (!spte) | |
4562 | goto out; | |
4563 | ||
8bea8052 | 4564 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 4565 | if (pud_none(*pud)) { |
3212b535 SC |
4566 | pud_populate(mm, pud, |
4567 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 4568 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 4569 | } else { |
3212b535 | 4570 | put_page(virt_to_page(spte)); |
dc6c9a35 | 4571 | } |
cb900f41 | 4572 | spin_unlock(ptl); |
3212b535 SC |
4573 | out: |
4574 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
83cde9e8 | 4575 | i_mmap_unlock_write(mapping); |
3212b535 SC |
4576 | return pte; |
4577 | } | |
4578 | ||
4579 | /* | |
4580 | * unmap huge page backed by shared pte. | |
4581 | * | |
4582 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
4583 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
4584 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
4585 | * | |
cb900f41 | 4586 | * called with page table lock held. |
3212b535 SC |
4587 | * |
4588 | * returns: 1 successfully unmapped a shared pte page | |
4589 | * 0 the underlying pte page is not shared, or it is the last user | |
4590 | */ | |
4591 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4592 | { | |
4593 | pgd_t *pgd = pgd_offset(mm, *addr); | |
c2febafc KS |
4594 | p4d_t *p4d = p4d_offset(pgd, *addr); |
4595 | pud_t *pud = pud_offset(p4d, *addr); | |
3212b535 SC |
4596 | |
4597 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
4598 | if (page_count(virt_to_page(ptep)) == 1) | |
4599 | return 0; | |
4600 | ||
4601 | pud_clear(pud); | |
4602 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 4603 | mm_dec_nr_pmds(mm); |
3212b535 SC |
4604 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
4605 | return 1; | |
4606 | } | |
9e5fc74c SC |
4607 | #define want_pmd_share() (1) |
4608 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
4609 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4610 | { | |
4611 | return NULL; | |
4612 | } | |
e81f2d22 ZZ |
4613 | |
4614 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4615 | { | |
4616 | return 0; | |
4617 | } | |
9e5fc74c | 4618 | #define want_pmd_share() (0) |
3212b535 SC |
4619 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
4620 | ||
9e5fc74c SC |
4621 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
4622 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
4623 | unsigned long addr, unsigned long sz) | |
4624 | { | |
4625 | pgd_t *pgd; | |
c2febafc | 4626 | p4d_t *p4d; |
9e5fc74c SC |
4627 | pud_t *pud; |
4628 | pte_t *pte = NULL; | |
4629 | ||
4630 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
4631 | p4d = p4d_offset(pgd, addr); |
4632 | pud = pud_alloc(mm, p4d, addr); | |
9e5fc74c SC |
4633 | if (pud) { |
4634 | if (sz == PUD_SIZE) { | |
4635 | pte = (pte_t *)pud; | |
4636 | } else { | |
4637 | BUG_ON(sz != PMD_SIZE); | |
4638 | if (want_pmd_share() && pud_none(*pud)) | |
4639 | pte = huge_pmd_share(mm, addr, pud); | |
4640 | else | |
4641 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
4642 | } | |
4643 | } | |
4e666314 | 4644 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
4645 | |
4646 | return pte; | |
4647 | } | |
4648 | ||
7868a208 PA |
4649 | pte_t *huge_pte_offset(struct mm_struct *mm, |
4650 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
4651 | { |
4652 | pgd_t *pgd; | |
c2febafc | 4653 | p4d_t *p4d; |
9e5fc74c | 4654 | pud_t *pud; |
c2febafc | 4655 | pmd_t *pmd; |
9e5fc74c SC |
4656 | |
4657 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
4658 | if (!pgd_present(*pgd)) |
4659 | return NULL; | |
4660 | p4d = p4d_offset(pgd, addr); | |
4661 | if (!p4d_present(*p4d)) | |
4662 | return NULL; | |
4663 | pud = pud_offset(p4d, addr); | |
4664 | if (!pud_present(*pud)) | |
4665 | return NULL; | |
4666 | if (pud_huge(*pud)) | |
4667 | return (pte_t *)pud; | |
4668 | pmd = pmd_offset(pud, addr); | |
9e5fc74c SC |
4669 | return (pte_t *) pmd; |
4670 | } | |
4671 | ||
61f77eda NH |
4672 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
4673 | ||
4674 | /* | |
4675 | * These functions are overwritable if your architecture needs its own | |
4676 | * behavior. | |
4677 | */ | |
4678 | struct page * __weak | |
4679 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
4680 | int write) | |
4681 | { | |
4682 | return ERR_PTR(-EINVAL); | |
4683 | } | |
4684 | ||
4dc71451 AK |
4685 | struct page * __weak |
4686 | follow_huge_pd(struct vm_area_struct *vma, | |
4687 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
4688 | { | |
4689 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
4690 | return NULL; | |
4691 | } | |
4692 | ||
61f77eda | 4693 | struct page * __weak |
9e5fc74c | 4694 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 4695 | pmd_t *pmd, int flags) |
9e5fc74c | 4696 | { |
e66f17ff NH |
4697 | struct page *page = NULL; |
4698 | spinlock_t *ptl; | |
c9d398fa | 4699 | pte_t pte; |
e66f17ff NH |
4700 | retry: |
4701 | ptl = pmd_lockptr(mm, pmd); | |
4702 | spin_lock(ptl); | |
4703 | /* | |
4704 | * make sure that the address range covered by this pmd is not | |
4705 | * unmapped from other threads. | |
4706 | */ | |
4707 | if (!pmd_huge(*pmd)) | |
4708 | goto out; | |
c9d398fa NH |
4709 | pte = huge_ptep_get((pte_t *)pmd); |
4710 | if (pte_present(pte)) { | |
97534127 | 4711 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
e66f17ff NH |
4712 | if (flags & FOLL_GET) |
4713 | get_page(page); | |
4714 | } else { | |
c9d398fa | 4715 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff NH |
4716 | spin_unlock(ptl); |
4717 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
4718 | goto retry; | |
4719 | } | |
4720 | /* | |
4721 | * hwpoisoned entry is treated as no_page_table in | |
4722 | * follow_page_mask(). | |
4723 | */ | |
4724 | } | |
4725 | out: | |
4726 | spin_unlock(ptl); | |
9e5fc74c SC |
4727 | return page; |
4728 | } | |
4729 | ||
61f77eda | 4730 | struct page * __weak |
9e5fc74c | 4731 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 4732 | pud_t *pud, int flags) |
9e5fc74c | 4733 | { |
e66f17ff NH |
4734 | if (flags & FOLL_GET) |
4735 | return NULL; | |
9e5fc74c | 4736 | |
e66f17ff | 4737 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
4738 | } |
4739 | ||
faaa5b62 AK |
4740 | struct page * __weak |
4741 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
4742 | { | |
4743 | if (flags & FOLL_GET) | |
4744 | return NULL; | |
4745 | ||
4746 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
4747 | } | |
4748 | ||
31caf665 NH |
4749 | bool isolate_huge_page(struct page *page, struct list_head *list) |
4750 | { | |
bcc54222 NH |
4751 | bool ret = true; |
4752 | ||
309381fe | 4753 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4754 | spin_lock(&hugetlb_lock); |
bcc54222 NH |
4755 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { |
4756 | ret = false; | |
4757 | goto unlock; | |
4758 | } | |
4759 | clear_page_huge_active(page); | |
31caf665 | 4760 | list_move_tail(&page->lru, list); |
bcc54222 | 4761 | unlock: |
31caf665 | 4762 | spin_unlock(&hugetlb_lock); |
bcc54222 | 4763 | return ret; |
31caf665 NH |
4764 | } |
4765 | ||
4766 | void putback_active_hugepage(struct page *page) | |
4767 | { | |
309381fe | 4768 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4769 | spin_lock(&hugetlb_lock); |
bcc54222 | 4770 | set_page_huge_active(page); |
31caf665 NH |
4771 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
4772 | spin_unlock(&hugetlb_lock); | |
4773 | put_page(page); | |
4774 | } |