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