]>
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/mm.h> | |
8 | #include <linux/seq_file.h> | |
9 | #include <linux/sysctl.h> | |
10 | #include <linux/highmem.h> | |
11 | #include <linux/mmu_notifier.h> | |
12 | #include <linux/nodemask.h> | |
13 | #include <linux/pagemap.h> | |
14 | #include <linux/mempolicy.h> | |
15 | #include <linux/compiler.h> | |
16 | #include <linux/cpuset.h> | |
17 | #include <linux/mutex.h> | |
18 | #include <linux/memblock.h> | |
19 | #include <linux/sysfs.h> | |
20 | #include <linux/slab.h> | |
21 | #include <linux/mmdebug.h> | |
22 | #include <linux/sched/signal.h> | |
23 | #include <linux/rmap.h> | |
24 | #include <linux/string_helpers.h> | |
25 | #include <linux/swap.h> | |
26 | #include <linux/swapops.h> | |
27 | #include <linux/jhash.h> | |
28 | #include <linux/numa.h> | |
29 | ||
30 | #include <asm/page.h> | |
31 | #include <asm/pgtable.h> | |
32 | #include <asm/tlb.h> | |
33 | ||
34 | #include <linux/io.h> | |
35 | #include <linux/hugetlb.h> | |
36 | #include <linux/hugetlb_cgroup.h> | |
37 | #include <linux/node.h> | |
38 | #include <linux/userfaultfd_k.h> | |
39 | #include <linux/page_owner.h> | |
40 | #include "internal.h" | |
41 | ||
42 | int hugetlb_max_hstate __read_mostly; | |
43 | unsigned int default_hstate_idx; | |
44 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
45 | /* | |
46 | * Minimum page order among possible hugepage sizes, set to a proper value | |
47 | * at boot time. | |
48 | */ | |
49 | static unsigned int minimum_order __read_mostly = UINT_MAX; | |
50 | ||
51 | __initdata LIST_HEAD(huge_boot_pages); | |
52 | ||
53 | /* for command line parsing */ | |
54 | static struct hstate * __initdata parsed_hstate; | |
55 | static unsigned long __initdata default_hstate_max_huge_pages; | |
56 | static unsigned long __initdata default_hstate_size; | |
57 | static bool __initdata parsed_valid_hugepagesz = true; | |
58 | ||
59 | /* | |
60 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, | |
61 | * free_huge_pages, and surplus_huge_pages. | |
62 | */ | |
63 | DEFINE_SPINLOCK(hugetlb_lock); | |
64 | ||
65 | /* | |
66 | * Serializes faults on the same logical page. This is used to | |
67 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
68 | */ | |
69 | static int num_fault_mutexes; | |
70 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; | |
71 | ||
72 | /* Forward declaration */ | |
73 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
74 | ||
75 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) | |
76 | { | |
77 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
78 | ||
79 | spin_unlock(&spool->lock); | |
80 | ||
81 | /* If no pages are used, and no other handles to the subpool | |
82 | * remain, give up any reservations mased on minimum size and | |
83 | * free the subpool */ | |
84 | if (free) { | |
85 | if (spool->min_hpages != -1) | |
86 | hugetlb_acct_memory(spool->hstate, | |
87 | -spool->min_hpages); | |
88 | kfree(spool); | |
89 | } | |
90 | } | |
91 | ||
92 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, | |
93 | long min_hpages) | |
94 | { | |
95 | struct hugepage_subpool *spool; | |
96 | ||
97 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); | |
98 | if (!spool) | |
99 | return NULL; | |
100 | ||
101 | spin_lock_init(&spool->lock); | |
102 | spool->count = 1; | |
103 | spool->max_hpages = max_hpages; | |
104 | spool->hstate = h; | |
105 | spool->min_hpages = min_hpages; | |
106 | ||
107 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
108 | kfree(spool); | |
109 | return NULL; | |
110 | } | |
111 | spool->rsv_hpages = min_hpages; | |
112 | ||
113 | return spool; | |
114 | } | |
115 | ||
116 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
117 | { | |
118 | spin_lock(&spool->lock); | |
119 | BUG_ON(!spool->count); | |
120 | spool->count--; | |
121 | unlock_or_release_subpool(spool); | |
122 | } | |
123 | ||
124 | /* | |
125 | * Subpool accounting for allocating and reserving pages. | |
126 | * Return -ENOMEM if there are not enough resources to satisfy the | |
127 | * the request. Otherwise, return the number of pages by which the | |
128 | * global pools must be adjusted (upward). The returned value may | |
129 | * only be different than the passed value (delta) in the case where | |
130 | * a subpool minimum size must be manitained. | |
131 | */ | |
132 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
133 | long delta) | |
134 | { | |
135 | long ret = delta; | |
136 | ||
137 | if (!spool) | |
138 | return ret; | |
139 | ||
140 | spin_lock(&spool->lock); | |
141 | ||
142 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
143 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
144 | spool->used_hpages += delta; | |
145 | else { | |
146 | ret = -ENOMEM; | |
147 | goto unlock_ret; | |
148 | } | |
149 | } | |
150 | ||
151 | /* minimum size accounting */ | |
152 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
153 | if (delta > spool->rsv_hpages) { | |
154 | /* | |
155 | * Asking for more reserves than those already taken on | |
156 | * behalf of subpool. Return difference. | |
157 | */ | |
158 | ret = delta - spool->rsv_hpages; | |
159 | spool->rsv_hpages = 0; | |
160 | } else { | |
161 | ret = 0; /* reserves already accounted for */ | |
162 | spool->rsv_hpages -= delta; | |
163 | } | |
164 | } | |
165 | ||
166 | unlock_ret: | |
167 | spin_unlock(&spool->lock); | |
168 | return ret; | |
169 | } | |
170 | ||
171 | /* | |
172 | * Subpool accounting for freeing and unreserving pages. | |
173 | * Return the number of global page reservations that must be dropped. | |
174 | * The return value may only be different than the passed value (delta) | |
175 | * in the case where a subpool minimum size must be maintained. | |
176 | */ | |
177 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
178 | long delta) | |
179 | { | |
180 | long ret = delta; | |
181 | ||
182 | if (!spool) | |
183 | return delta; | |
184 | ||
185 | spin_lock(&spool->lock); | |
186 | ||
187 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
188 | spool->used_hpages -= delta; | |
189 | ||
190 | /* minimum size accounting */ | |
191 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
192 | if (spool->rsv_hpages + delta <= spool->min_hpages) | |
193 | ret = 0; | |
194 | else | |
195 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
196 | ||
197 | spool->rsv_hpages += delta; | |
198 | if (spool->rsv_hpages > spool->min_hpages) | |
199 | spool->rsv_hpages = spool->min_hpages; | |
200 | } | |
201 | ||
202 | /* | |
203 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
204 | * quota reference, free it now. | |
205 | */ | |
206 | unlock_or_release_subpool(spool); | |
207 | ||
208 | return ret; | |
209 | } | |
210 | ||
211 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
212 | { | |
213 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
214 | } | |
215 | ||
216 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
217 | { | |
218 | return subpool_inode(file_inode(vma->vm_file)); | |
219 | } | |
220 | ||
221 | /* | |
222 | * Region tracking -- allows tracking of reservations and instantiated pages | |
223 | * across the pages in a mapping. | |
224 | * | |
225 | * The region data structures are embedded into a resv_map and protected | |
226 | * by a resv_map's lock. The set of regions within the resv_map represent | |
227 | * reservations for huge pages, or huge pages that have already been | |
228 | * instantiated within the map. The from and to elements are huge page | |
229 | * indicies into the associated mapping. from indicates the starting index | |
230 | * of the region. to represents the first index past the end of the region. | |
231 | * | |
232 | * For example, a file region structure with from == 0 and to == 4 represents | |
233 | * four huge pages in a mapping. It is important to note that the to element | |
234 | * represents the first element past the end of the region. This is used in | |
235 | * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. | |
236 | * | |
237 | * Interval notation of the form [from, to) will be used to indicate that | |
238 | * the endpoint from is inclusive and to is exclusive. | |
239 | */ | |
240 | struct file_region { | |
241 | struct list_head link; | |
242 | long from; | |
243 | long to; | |
244 | }; | |
245 | ||
246 | /* | |
247 | * Add the huge page range represented by [f, t) to the reserve | |
248 | * map. In the normal case, existing regions will be expanded | |
249 | * to accommodate the specified range. Sufficient regions should | |
250 | * exist for expansion due to the previous call to region_chg | |
251 | * with the same range. However, it is possible that region_del | |
252 | * could have been called after region_chg and modifed the map | |
253 | * in such a way that no region exists to be expanded. In this | |
254 | * case, pull a region descriptor from the cache associated with | |
255 | * the map and use that for the new range. | |
256 | * | |
257 | * Return the number of new huge pages added to the map. This | |
258 | * number is greater than or equal to zero. | |
259 | */ | |
260 | static long region_add(struct resv_map *resv, long f, long t) | |
261 | { | |
262 | struct list_head *head = &resv->regions; | |
263 | struct file_region *rg, *nrg, *trg; | |
264 | long add = 0; | |
265 | ||
266 | spin_lock(&resv->lock); | |
267 | /* Locate the region we are either in or before. */ | |
268 | list_for_each_entry(rg, head, link) | |
269 | if (f <= rg->to) | |
270 | break; | |
271 | ||
272 | /* | |
273 | * If no region exists which can be expanded to include the | |
274 | * specified range, the list must have been modified by an | |
275 | * interleving call to region_del(). Pull a region descriptor | |
276 | * from the cache and use it for this range. | |
277 | */ | |
278 | if (&rg->link == head || t < rg->from) { | |
279 | VM_BUG_ON(resv->region_cache_count <= 0); | |
280 | ||
281 | resv->region_cache_count--; | |
282 | nrg = list_first_entry(&resv->region_cache, struct file_region, | |
283 | link); | |
284 | list_del(&nrg->link); | |
285 | ||
286 | nrg->from = f; | |
287 | nrg->to = t; | |
288 | list_add(&nrg->link, rg->link.prev); | |
289 | ||
290 | add += t - f; | |
291 | goto out_locked; | |
292 | } | |
293 | ||
294 | /* Round our left edge to the current segment if it encloses us. */ | |
295 | if (f > rg->from) | |
296 | f = rg->from; | |
297 | ||
298 | /* Check for and consume any regions we now overlap with. */ | |
299 | nrg = rg; | |
300 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
301 | if (&rg->link == head) | |
302 | break; | |
303 | if (rg->from > t) | |
304 | break; | |
305 | ||
306 | /* If this area reaches higher then extend our area to | |
307 | * include it completely. If this is not the first area | |
308 | * which we intend to reuse, free it. */ | |
309 | if (rg->to > t) | |
310 | t = rg->to; | |
311 | if (rg != nrg) { | |
312 | /* Decrement return value by the deleted range. | |
313 | * Another range will span this area so that by | |
314 | * end of routine add will be >= zero | |
315 | */ | |
316 | add -= (rg->to - rg->from); | |
317 | list_del(&rg->link); | |
318 | kfree(rg); | |
319 | } | |
320 | } | |
321 | ||
322 | add += (nrg->from - f); /* Added to beginning of region */ | |
323 | nrg->from = f; | |
324 | add += t - nrg->to; /* Added to end of region */ | |
325 | nrg->to = t; | |
326 | ||
327 | out_locked: | |
328 | resv->adds_in_progress--; | |
329 | spin_unlock(&resv->lock); | |
330 | VM_BUG_ON(add < 0); | |
331 | return add; | |
332 | } | |
333 | ||
334 | /* | |
335 | * Examine the existing reserve map and determine how many | |
336 | * huge pages in the specified range [f, t) are NOT currently | |
337 | * represented. This routine is called before a subsequent | |
338 | * call to region_add that will actually modify the reserve | |
339 | * map to add the specified range [f, t). region_chg does | |
340 | * not change the number of huge pages represented by the | |
341 | * map. However, if the existing regions in the map can not | |
342 | * be expanded to represent the new range, a new file_region | |
343 | * structure is added to the map as a placeholder. This is | |
344 | * so that the subsequent region_add call will have all the | |
345 | * regions it needs and will not fail. | |
346 | * | |
347 | * Upon entry, region_chg will also examine the cache of region descriptors | |
348 | * associated with the map. If there are not enough descriptors cached, one | |
349 | * will be allocated for the in progress add operation. | |
350 | * | |
351 | * Returns the number of huge pages that need to be added to the existing | |
352 | * reservation map for the range [f, t). This number is greater or equal to | |
353 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
354 | * is needed and can not be allocated. | |
355 | */ | |
356 | static long region_chg(struct resv_map *resv, long f, long t) | |
357 | { | |
358 | struct list_head *head = &resv->regions; | |
359 | struct file_region *rg, *nrg = NULL; | |
360 | long chg = 0; | |
361 | ||
362 | retry: | |
363 | spin_lock(&resv->lock); | |
364 | retry_locked: | |
365 | resv->adds_in_progress++; | |
366 | ||
367 | /* | |
368 | * Check for sufficient descriptors in the cache to accommodate | |
369 | * the number of in progress add operations. | |
370 | */ | |
371 | if (resv->adds_in_progress > resv->region_cache_count) { | |
372 | struct file_region *trg; | |
373 | ||
374 | VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1); | |
375 | /* Must drop lock to allocate a new descriptor. */ | |
376 | resv->adds_in_progress--; | |
377 | spin_unlock(&resv->lock); | |
378 | ||
379 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
380 | if (!trg) { | |
381 | kfree(nrg); | |
382 | return -ENOMEM; | |
383 | } | |
384 | ||
385 | spin_lock(&resv->lock); | |
386 | list_add(&trg->link, &resv->region_cache); | |
387 | resv->region_cache_count++; | |
388 | goto retry_locked; | |
389 | } | |
390 | ||
391 | /* Locate the region we are before or in. */ | |
392 | list_for_each_entry(rg, head, link) | |
393 | if (f <= rg->to) | |
394 | break; | |
395 | ||
396 | /* If we are below the current region then a new region is required. | |
397 | * Subtle, allocate a new region at the position but make it zero | |
398 | * size such that we can guarantee to record the reservation. */ | |
399 | if (&rg->link == head || t < rg->from) { | |
400 | if (!nrg) { | |
401 | resv->adds_in_progress--; | |
402 | spin_unlock(&resv->lock); | |
403 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
404 | if (!nrg) | |
405 | return -ENOMEM; | |
406 | ||
407 | nrg->from = f; | |
408 | nrg->to = f; | |
409 | INIT_LIST_HEAD(&nrg->link); | |
410 | goto retry; | |
411 | } | |
412 | ||
413 | list_add(&nrg->link, rg->link.prev); | |
414 | chg = t - f; | |
415 | goto out_nrg; | |
416 | } | |
417 | ||
418 | /* Round our left edge to the current segment if it encloses us. */ | |
419 | if (f > rg->from) | |
420 | f = rg->from; | |
421 | chg = t - f; | |
422 | ||
423 | /* Check for and consume any regions we now overlap with. */ | |
424 | list_for_each_entry(rg, rg->link.prev, link) { | |
425 | if (&rg->link == head) | |
426 | break; | |
427 | if (rg->from > t) | |
428 | goto out; | |
429 | ||
430 | /* We overlap with this area, if it extends further than | |
431 | * us then we must extend ourselves. Account for its | |
432 | * existing reservation. */ | |
433 | if (rg->to > t) { | |
434 | chg += rg->to - t; | |
435 | t = rg->to; | |
436 | } | |
437 | chg -= rg->to - rg->from; | |
438 | } | |
439 | ||
440 | out: | |
441 | spin_unlock(&resv->lock); | |
442 | /* We already know we raced and no longer need the new region */ | |
443 | kfree(nrg); | |
444 | return chg; | |
445 | out_nrg: | |
446 | spin_unlock(&resv->lock); | |
447 | return chg; | |
448 | } | |
449 | ||
450 | /* | |
451 | * Abort the in progress add operation. The adds_in_progress field | |
452 | * of the resv_map keeps track of the operations in progress between | |
453 | * calls to region_chg and region_add. Operations are sometimes | |
454 | * aborted after the call to region_chg. In such cases, region_abort | |
455 | * is called to decrement the adds_in_progress counter. | |
456 | * | |
457 | * NOTE: The range arguments [f, t) are not needed or used in this | |
458 | * routine. They are kept to make reading the calling code easier as | |
459 | * arguments will match the associated region_chg call. | |
460 | */ | |
461 | static void region_abort(struct resv_map *resv, long f, long t) | |
462 | { | |
463 | spin_lock(&resv->lock); | |
464 | VM_BUG_ON(!resv->region_cache_count); | |
465 | resv->adds_in_progress--; | |
466 | spin_unlock(&resv->lock); | |
467 | } | |
468 | ||
469 | /* | |
470 | * Delete the specified range [f, t) from the reserve map. If the | |
471 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
472 | * should be deleted. Locate the regions which intersect [f, t) | |
473 | * and either trim, delete or split the existing regions. | |
474 | * | |
475 | * Returns the number of huge pages deleted from the reserve map. | |
476 | * In the normal case, the return value is zero or more. In the | |
477 | * case where a region must be split, a new region descriptor must | |
478 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
479 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
480 | * a region and possibly return -ENOMEM. Callers specifying | |
481 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
482 | */ | |
483 | static long region_del(struct resv_map *resv, long f, long t) | |
484 | { | |
485 | struct list_head *head = &resv->regions; | |
486 | struct file_region *rg, *trg; | |
487 | struct file_region *nrg = NULL; | |
488 | long del = 0; | |
489 | ||
490 | retry: | |
491 | spin_lock(&resv->lock); | |
492 | list_for_each_entry_safe(rg, trg, head, link) { | |
493 | /* | |
494 | * Skip regions before the range to be deleted. file_region | |
495 | * ranges are normally of the form [from, to). However, there | |
496 | * may be a "placeholder" entry in the map which is of the form | |
497 | * (from, to) with from == to. Check for placeholder entries | |
498 | * at the beginning of the range to be deleted. | |
499 | */ | |
500 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
501 | continue; | |
502 | ||
503 | if (rg->from >= t) | |
504 | break; | |
505 | ||
506 | if (f > rg->from && t < rg->to) { /* Must split region */ | |
507 | /* | |
508 | * Check for an entry in the cache before dropping | |
509 | * lock and attempting allocation. | |
510 | */ | |
511 | if (!nrg && | |
512 | resv->region_cache_count > resv->adds_in_progress) { | |
513 | nrg = list_first_entry(&resv->region_cache, | |
514 | struct file_region, | |
515 | link); | |
516 | list_del(&nrg->link); | |
517 | resv->region_cache_count--; | |
518 | } | |
519 | ||
520 | if (!nrg) { | |
521 | spin_unlock(&resv->lock); | |
522 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
523 | if (!nrg) | |
524 | return -ENOMEM; | |
525 | goto retry; | |
526 | } | |
527 | ||
528 | del += t - f; | |
529 | ||
530 | /* New entry for end of split region */ | |
531 | nrg->from = t; | |
532 | nrg->to = rg->to; | |
533 | INIT_LIST_HEAD(&nrg->link); | |
534 | ||
535 | /* Original entry is trimmed */ | |
536 | rg->to = f; | |
537 | ||
538 | list_add(&nrg->link, &rg->link); | |
539 | nrg = NULL; | |
540 | break; | |
541 | } | |
542 | ||
543 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
544 | del += rg->to - rg->from; | |
545 | list_del(&rg->link); | |
546 | kfree(rg); | |
547 | continue; | |
548 | } | |
549 | ||
550 | if (f <= rg->from) { /* Trim beginning of region */ | |
551 | del += t - rg->from; | |
552 | rg->from = t; | |
553 | } else { /* Trim end of region */ | |
554 | del += rg->to - f; | |
555 | rg->to = f; | |
556 | } | |
557 | } | |
558 | ||
559 | spin_unlock(&resv->lock); | |
560 | kfree(nrg); | |
561 | return del; | |
562 | } | |
563 | ||
564 | /* | |
565 | * A rare out of memory error was encountered which prevented removal of | |
566 | * the reserve map region for a page. The huge page itself was free'ed | |
567 | * and removed from the page cache. This routine will adjust the subpool | |
568 | * usage count, and the global reserve count if needed. By incrementing | |
569 | * these counts, the reserve map entry which could not be deleted will | |
570 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
571 | * counts. | |
572 | */ | |
573 | void hugetlb_fix_reserve_counts(struct inode *inode) | |
574 | { | |
575 | struct hugepage_subpool *spool = subpool_inode(inode); | |
576 | long rsv_adjust; | |
577 | ||
578 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
579 | if (rsv_adjust) { | |
580 | struct hstate *h = hstate_inode(inode); | |
581 | ||
582 | hugetlb_acct_memory(h, 1); | |
583 | } | |
584 | } | |
585 | ||
586 | /* | |
587 | * Count and return the number of huge pages in the reserve map | |
588 | * that intersect with the range [f, t). | |
589 | */ | |
590 | static long region_count(struct resv_map *resv, long f, long t) | |
591 | { | |
592 | struct list_head *head = &resv->regions; | |
593 | struct file_region *rg; | |
594 | long chg = 0; | |
595 | ||
596 | spin_lock(&resv->lock); | |
597 | /* Locate each segment we overlap with, and count that overlap. */ | |
598 | list_for_each_entry(rg, head, link) { | |
599 | long seg_from; | |
600 | long seg_to; | |
601 | ||
602 | if (rg->to <= f) | |
603 | continue; | |
604 | if (rg->from >= t) | |
605 | break; | |
606 | ||
607 | seg_from = max(rg->from, f); | |
608 | seg_to = min(rg->to, t); | |
609 | ||
610 | chg += seg_to - seg_from; | |
611 | } | |
612 | spin_unlock(&resv->lock); | |
613 | ||
614 | return chg; | |
615 | } | |
616 | ||
617 | /* | |
618 | * Convert the address within this vma to the page offset within | |
619 | * the mapping, in pagecache page units; huge pages here. | |
620 | */ | |
621 | static pgoff_t vma_hugecache_offset(struct hstate *h, | |
622 | struct vm_area_struct *vma, unsigned long address) | |
623 | { | |
624 | return ((address - vma->vm_start) >> huge_page_shift(h)) + | |
625 | (vma->vm_pgoff >> huge_page_order(h)); | |
626 | } | |
627 | ||
628 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, | |
629 | unsigned long address) | |
630 | { | |
631 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
632 | } | |
633 | EXPORT_SYMBOL_GPL(linear_hugepage_index); | |
634 | ||
635 | /* | |
636 | * Return the size of the pages allocated when backing a VMA. In the majority | |
637 | * cases this will be same size as used by the page table entries. | |
638 | */ | |
639 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
640 | { | |
641 | if (vma->vm_ops && vma->vm_ops->pagesize) | |
642 | return vma->vm_ops->pagesize(vma); | |
643 | return PAGE_SIZE; | |
644 | } | |
645 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); | |
646 | ||
647 | /* | |
648 | * Return the page size being used by the MMU to back a VMA. In the majority | |
649 | * of cases, the page size used by the kernel matches the MMU size. On | |
650 | * architectures where it differs, an architecture-specific 'strong' | |
651 | * version of this symbol is required. | |
652 | */ | |
653 | __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) | |
654 | { | |
655 | return vma_kernel_pagesize(vma); | |
656 | } | |
657 | ||
658 | /* | |
659 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
660 | * bits of the reservation map pointer, which are always clear due to | |
661 | * alignment. | |
662 | */ | |
663 | #define HPAGE_RESV_OWNER (1UL << 0) | |
664 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
665 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) | |
666 | ||
667 | /* | |
668 | * These helpers are used to track how many pages are reserved for | |
669 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
670 | * is guaranteed to have their future faults succeed. | |
671 | * | |
672 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
673 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
674 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
675 | * chance of the global counters getting corrupted as a result of the values. | |
676 | * | |
677 | * The private mapping reservation is represented in a subtly different | |
678 | * manner to a shared mapping. A shared mapping has a region map associated | |
679 | * with the underlying file, this region map represents the backing file | |
680 | * pages which have ever had a reservation assigned which this persists even | |
681 | * after the page is instantiated. A private mapping has a region map | |
682 | * associated with the original mmap which is attached to all VMAs which | |
683 | * reference it, this region map represents those offsets which have consumed | |
684 | * reservation ie. where pages have been instantiated. | |
685 | */ | |
686 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) | |
687 | { | |
688 | return (unsigned long)vma->vm_private_data; | |
689 | } | |
690 | ||
691 | static void set_vma_private_data(struct vm_area_struct *vma, | |
692 | unsigned long value) | |
693 | { | |
694 | vma->vm_private_data = (void *)value; | |
695 | } | |
696 | ||
697 | struct resv_map *resv_map_alloc(void) | |
698 | { | |
699 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
700 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); | |
701 | ||
702 | if (!resv_map || !rg) { | |
703 | kfree(resv_map); | |
704 | kfree(rg); | |
705 | return NULL; | |
706 | } | |
707 | ||
708 | kref_init(&resv_map->refs); | |
709 | spin_lock_init(&resv_map->lock); | |
710 | INIT_LIST_HEAD(&resv_map->regions); | |
711 | ||
712 | resv_map->adds_in_progress = 0; | |
713 | ||
714 | INIT_LIST_HEAD(&resv_map->region_cache); | |
715 | list_add(&rg->link, &resv_map->region_cache); | |
716 | resv_map->region_cache_count = 1; | |
717 | ||
718 | return resv_map; | |
719 | } | |
720 | ||
721 | void resv_map_release(struct kref *ref) | |
722 | { | |
723 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
724 | struct list_head *head = &resv_map->region_cache; | |
725 | struct file_region *rg, *trg; | |
726 | ||
727 | /* Clear out any active regions before we release the map. */ | |
728 | region_del(resv_map, 0, LONG_MAX); | |
729 | ||
730 | /* ... and any entries left in the cache */ | |
731 | list_for_each_entry_safe(rg, trg, head, link) { | |
732 | list_del(&rg->link); | |
733 | kfree(rg); | |
734 | } | |
735 | ||
736 | VM_BUG_ON(resv_map->adds_in_progress); | |
737 | ||
738 | kfree(resv_map); | |
739 | } | |
740 | ||
741 | static inline struct resv_map *inode_resv_map(struct inode *inode) | |
742 | { | |
743 | return inode->i_mapping->private_data; | |
744 | } | |
745 | ||
746 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) | |
747 | { | |
748 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); | |
749 | if (vma->vm_flags & VM_MAYSHARE) { | |
750 | struct address_space *mapping = vma->vm_file->f_mapping; | |
751 | struct inode *inode = mapping->host; | |
752 | ||
753 | return inode_resv_map(inode); | |
754 | ||
755 | } else { | |
756 | return (struct resv_map *)(get_vma_private_data(vma) & | |
757 | ~HPAGE_RESV_MASK); | |
758 | } | |
759 | } | |
760 | ||
761 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) | |
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) & | |
767 | HPAGE_RESV_MASK) | (unsigned long)map); | |
768 | } | |
769 | ||
770 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
771 | { | |
772 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); | |
773 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
774 | ||
775 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
776 | } | |
777 | ||
778 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
779 | { | |
780 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); | |
781 | ||
782 | return (get_vma_private_data(vma) & flag) != 0; | |
783 | } | |
784 | ||
785 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ | |
786 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) | |
787 | { | |
788 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); | |
789 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
790 | vma->vm_private_data = (void *)0; | |
791 | } | |
792 | ||
793 | /* Returns true if the VMA has associated reserve pages */ | |
794 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) | |
795 | { | |
796 | if (vma->vm_flags & VM_NORESERVE) { | |
797 | /* | |
798 | * This address is already reserved by other process(chg == 0), | |
799 | * so, we should decrement reserved count. Without decrementing, | |
800 | * reserve count remains after releasing inode, because this | |
801 | * allocated page will go into page cache and is regarded as | |
802 | * coming from reserved pool in releasing step. Currently, we | |
803 | * don't have any other solution to deal with this situation | |
804 | * properly, so add work-around here. | |
805 | */ | |
806 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
807 | return true; | |
808 | else | |
809 | return false; | |
810 | } | |
811 | ||
812 | /* Shared mappings always use reserves */ | |
813 | if (vma->vm_flags & VM_MAYSHARE) { | |
814 | /* | |
815 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
816 | * be a region map for all pages. The only situation where | |
817 | * there is no region map is if a hole was punched via | |
818 | * fallocate. In this case, there really are no reverves to | |
819 | * use. This situation is indicated if chg != 0. | |
820 | */ | |
821 | if (chg) | |
822 | return false; | |
823 | else | |
824 | return true; | |
825 | } | |
826 | ||
827 | /* | |
828 | * Only the process that called mmap() has reserves for | |
829 | * private mappings. | |
830 | */ | |
831 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { | |
832 | /* | |
833 | * Like the shared case above, a hole punch or truncate | |
834 | * could have been performed on the private mapping. | |
835 | * Examine the value of chg to determine if reserves | |
836 | * actually exist or were previously consumed. | |
837 | * Very Subtle - The value of chg comes from a previous | |
838 | * call to vma_needs_reserves(). The reserve map for | |
839 | * private mappings has different (opposite) semantics | |
840 | * than that of shared mappings. vma_needs_reserves() | |
841 | * has already taken this difference in semantics into | |
842 | * account. Therefore, the meaning of chg is the same | |
843 | * as in the shared case above. Code could easily be | |
844 | * combined, but keeping it separate draws attention to | |
845 | * subtle differences. | |
846 | */ | |
847 | if (chg) | |
848 | return false; | |
849 | else | |
850 | return true; | |
851 | } | |
852 | ||
853 | return false; | |
854 | } | |
855 | ||
856 | static void enqueue_huge_page(struct hstate *h, struct page *page) | |
857 | { | |
858 | int nid = page_to_nid(page); | |
859 | list_move(&page->lru, &h->hugepage_freelists[nid]); | |
860 | h->free_huge_pages++; | |
861 | h->free_huge_pages_node[nid]++; | |
862 | } | |
863 | ||
864 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) | |
865 | { | |
866 | struct page *page; | |
867 | ||
868 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) | |
869 | if (!PageHWPoison(page)) | |
870 | break; | |
871 | /* | |
872 | * if 'non-isolated free hugepage' not found on the list, | |
873 | * the allocation fails. | |
874 | */ | |
875 | if (&h->hugepage_freelists[nid] == &page->lru) | |
876 | return NULL; | |
877 | list_move(&page->lru, &h->hugepage_activelist); | |
878 | set_page_refcounted(page); | |
879 | h->free_huge_pages--; | |
880 | h->free_huge_pages_node[nid]--; | |
881 | return page; | |
882 | } | |
883 | ||
884 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, | |
885 | nodemask_t *nmask) | |
886 | { | |
887 | unsigned int cpuset_mems_cookie; | |
888 | struct zonelist *zonelist; | |
889 | struct zone *zone; | |
890 | struct zoneref *z; | |
891 | int node = NUMA_NO_NODE; | |
892 | ||
893 | zonelist = node_zonelist(nid, gfp_mask); | |
894 | ||
895 | retry_cpuset: | |
896 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
897 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
898 | struct page *page; | |
899 | ||
900 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
901 | continue; | |
902 | /* | |
903 | * no need to ask again on the same node. Pool is node rather than | |
904 | * zone aware | |
905 | */ | |
906 | if (zone_to_nid(zone) == node) | |
907 | continue; | |
908 | node = zone_to_nid(zone); | |
909 | ||
910 | page = dequeue_huge_page_node_exact(h, node); | |
911 | if (page) | |
912 | return page; | |
913 | } | |
914 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) | |
915 | goto retry_cpuset; | |
916 | ||
917 | return NULL; | |
918 | } | |
919 | ||
920 | /* Movability of hugepages depends on migration support. */ | |
921 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
922 | { | |
923 | if (hugepage_movable_supported(h)) | |
924 | return GFP_HIGHUSER_MOVABLE; | |
925 | else | |
926 | return GFP_HIGHUSER; | |
927 | } | |
928 | ||
929 | static struct page *dequeue_huge_page_vma(struct hstate *h, | |
930 | struct vm_area_struct *vma, | |
931 | unsigned long address, int avoid_reserve, | |
932 | long chg) | |
933 | { | |
934 | struct page *page; | |
935 | struct mempolicy *mpol; | |
936 | gfp_t gfp_mask; | |
937 | nodemask_t *nodemask; | |
938 | int nid; | |
939 | ||
940 | /* | |
941 | * A child process with MAP_PRIVATE mappings created by their parent | |
942 | * have no page reserves. This check ensures that reservations are | |
943 | * not "stolen". The child may still get SIGKILLed | |
944 | */ | |
945 | if (!vma_has_reserves(vma, chg) && | |
946 | h->free_huge_pages - h->resv_huge_pages == 0) | |
947 | goto err; | |
948 | ||
949 | /* If reserves cannot be used, ensure enough pages are in the pool */ | |
950 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) | |
951 | goto err; | |
952 | ||
953 | gfp_mask = htlb_alloc_mask(h); | |
954 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
955 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); | |
956 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { | |
957 | SetPagePrivate(page); | |
958 | h->resv_huge_pages--; | |
959 | } | |
960 | ||
961 | mpol_cond_put(mpol); | |
962 | return page; | |
963 | ||
964 | err: | |
965 | return NULL; | |
966 | } | |
967 | ||
968 | /* | |
969 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
970 | * We may have allocated or freed a huge page based on a different | |
971 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
972 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
973 | * node for alloc or free. | |
974 | */ | |
975 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
976 | { | |
977 | nid = next_node_in(nid, *nodes_allowed); | |
978 | VM_BUG_ON(nid >= MAX_NUMNODES); | |
979 | ||
980 | return nid; | |
981 | } | |
982 | ||
983 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
984 | { | |
985 | if (!node_isset(nid, *nodes_allowed)) | |
986 | nid = next_node_allowed(nid, nodes_allowed); | |
987 | return nid; | |
988 | } | |
989 | ||
990 | /* | |
991 | * returns the previously saved node ["this node"] from which to | |
992 | * allocate a persistent huge page for the pool and advance the | |
993 | * next node from which to allocate, handling wrap at end of node | |
994 | * mask. | |
995 | */ | |
996 | static int hstate_next_node_to_alloc(struct hstate *h, | |
997 | nodemask_t *nodes_allowed) | |
998 | { | |
999 | int nid; | |
1000 | ||
1001 | VM_BUG_ON(!nodes_allowed); | |
1002 | ||
1003 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1004 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1005 | ||
1006 | return nid; | |
1007 | } | |
1008 | ||
1009 | /* | |
1010 | * helper for free_pool_huge_page() - return the previously saved | |
1011 | * node ["this node"] from which to free a huge page. Advance the | |
1012 | * next node id whether or not we find a free huge page to free so | |
1013 | * that the next attempt to free addresses the next node. | |
1014 | */ | |
1015 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1016 | { | |
1017 | int nid; | |
1018 | ||
1019 | VM_BUG_ON(!nodes_allowed); | |
1020 | ||
1021 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1022 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1023 | ||
1024 | return nid; | |
1025 | } | |
1026 | ||
1027 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1028 | for (nr_nodes = nodes_weight(*mask); \ | |
1029 | nr_nodes > 0 && \ | |
1030 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1031 | nr_nodes--) | |
1032 | ||
1033 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1034 | for (nr_nodes = nodes_weight(*mask); \ | |
1035 | nr_nodes > 0 && \ | |
1036 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1037 | nr_nodes--) | |
1038 | ||
1039 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE | |
1040 | static void destroy_compound_gigantic_page(struct page *page, | |
1041 | unsigned int order) | |
1042 | { | |
1043 | int i; | |
1044 | int nr_pages = 1 << order; | |
1045 | struct page *p = page + 1; | |
1046 | ||
1047 | atomic_set(compound_mapcount_ptr(page), 0); | |
1048 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { | |
1049 | clear_compound_head(p); | |
1050 | set_page_refcounted(p); | |
1051 | } | |
1052 | ||
1053 | set_compound_order(page, 0); | |
1054 | __ClearPageHead(page); | |
1055 | } | |
1056 | ||
1057 | static void free_gigantic_page(struct page *page, unsigned int order) | |
1058 | { | |
1059 | free_contig_range(page_to_pfn(page), 1 << order); | |
1060 | } | |
1061 | ||
1062 | #ifdef CONFIG_CONTIG_ALLOC | |
1063 | static int __alloc_gigantic_page(unsigned long start_pfn, | |
1064 | unsigned long nr_pages, gfp_t gfp_mask) | |
1065 | { | |
1066 | unsigned long end_pfn = start_pfn + nr_pages; | |
1067 | return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE, | |
1068 | gfp_mask); | |
1069 | } | |
1070 | ||
1071 | static bool pfn_range_valid_gigantic(struct zone *z, | |
1072 | unsigned long start_pfn, unsigned long nr_pages) | |
1073 | { | |
1074 | unsigned long i, end_pfn = start_pfn + nr_pages; | |
1075 | struct page *page; | |
1076 | ||
1077 | for (i = start_pfn; i < end_pfn; i++) { | |
1078 | if (!pfn_valid(i)) | |
1079 | return false; | |
1080 | ||
1081 | page = pfn_to_page(i); | |
1082 | ||
1083 | if (page_zone(page) != z) | |
1084 | return false; | |
1085 | ||
1086 | if (PageReserved(page)) | |
1087 | return false; | |
1088 | ||
1089 | if (page_count(page) > 0) | |
1090 | return false; | |
1091 | ||
1092 | if (PageHuge(page)) | |
1093 | return false; | |
1094 | } | |
1095 | ||
1096 | return true; | |
1097 | } | |
1098 | ||
1099 | static bool zone_spans_last_pfn(const struct zone *zone, | |
1100 | unsigned long start_pfn, unsigned long nr_pages) | |
1101 | { | |
1102 | unsigned long last_pfn = start_pfn + nr_pages - 1; | |
1103 | return zone_spans_pfn(zone, last_pfn); | |
1104 | } | |
1105 | ||
1106 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1107 | int nid, nodemask_t *nodemask) | |
1108 | { | |
1109 | unsigned int order = huge_page_order(h); | |
1110 | unsigned long nr_pages = 1 << order; | |
1111 | unsigned long ret, pfn, flags; | |
1112 | struct zonelist *zonelist; | |
1113 | struct zone *zone; | |
1114 | struct zoneref *z; | |
1115 | ||
1116 | zonelist = node_zonelist(nid, gfp_mask); | |
1117 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) { | |
1118 | spin_lock_irqsave(&zone->lock, flags); | |
1119 | ||
1120 | pfn = ALIGN(zone->zone_start_pfn, nr_pages); | |
1121 | while (zone_spans_last_pfn(zone, pfn, nr_pages)) { | |
1122 | if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) { | |
1123 | /* | |
1124 | * We release the zone lock here because | |
1125 | * alloc_contig_range() will also lock the zone | |
1126 | * at some point. If there's an allocation | |
1127 | * spinning on this lock, it may win the race | |
1128 | * and cause alloc_contig_range() to fail... | |
1129 | */ | |
1130 | spin_unlock_irqrestore(&zone->lock, flags); | |
1131 | ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask); | |
1132 | if (!ret) | |
1133 | return pfn_to_page(pfn); | |
1134 | spin_lock_irqsave(&zone->lock, flags); | |
1135 | } | |
1136 | pfn += nr_pages; | |
1137 | } | |
1138 | ||
1139 | spin_unlock_irqrestore(&zone->lock, flags); | |
1140 | } | |
1141 | ||
1142 | return NULL; | |
1143 | } | |
1144 | ||
1145 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
1146 | static void prep_compound_gigantic_page(struct page *page, unsigned int order); | |
1147 | #else /* !CONFIG_CONTIG_ALLOC */ | |
1148 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1149 | int nid, nodemask_t *nodemask) | |
1150 | { | |
1151 | return NULL; | |
1152 | } | |
1153 | #endif /* CONFIG_CONTIG_ALLOC */ | |
1154 | ||
1155 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ | |
1156 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1157 | int nid, nodemask_t *nodemask) | |
1158 | { | |
1159 | return NULL; | |
1160 | } | |
1161 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } | |
1162 | static inline void destroy_compound_gigantic_page(struct page *page, | |
1163 | unsigned int order) { } | |
1164 | #endif | |
1165 | ||
1166 | static void update_and_free_page(struct hstate *h, struct page *page) | |
1167 | { | |
1168 | int i; | |
1169 | ||
1170 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) | |
1171 | return; | |
1172 | ||
1173 | h->nr_huge_pages--; | |
1174 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
1175 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
1176 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | | |
1177 | 1 << PG_referenced | 1 << PG_dirty | | |
1178 | 1 << PG_active | 1 << PG_private | | |
1179 | 1 << PG_writeback); | |
1180 | } | |
1181 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); | |
1182 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); | |
1183 | set_page_refcounted(page); | |
1184 | if (hstate_is_gigantic(h)) { | |
1185 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
1186 | free_gigantic_page(page, huge_page_order(h)); | |
1187 | } else { | |
1188 | __free_pages(page, huge_page_order(h)); | |
1189 | } | |
1190 | } | |
1191 | ||
1192 | struct hstate *size_to_hstate(unsigned long size) | |
1193 | { | |
1194 | struct hstate *h; | |
1195 | ||
1196 | for_each_hstate(h) { | |
1197 | if (huge_page_size(h) == size) | |
1198 | return h; | |
1199 | } | |
1200 | return NULL; | |
1201 | } | |
1202 | ||
1203 | /* | |
1204 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
1205 | * to hstate->hugepage_activelist.) | |
1206 | * | |
1207 | * This function can be called for tail pages, but never returns true for them. | |
1208 | */ | |
1209 | bool page_huge_active(struct page *page) | |
1210 | { | |
1211 | VM_BUG_ON_PAGE(!PageHuge(page), page); | |
1212 | return PageHead(page) && PagePrivate(&page[1]); | |
1213 | } | |
1214 | ||
1215 | /* never called for tail page */ | |
1216 | static void set_page_huge_active(struct page *page) | |
1217 | { | |
1218 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1219 | SetPagePrivate(&page[1]); | |
1220 | } | |
1221 | ||
1222 | static void clear_page_huge_active(struct page *page) | |
1223 | { | |
1224 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1225 | ClearPagePrivate(&page[1]); | |
1226 | } | |
1227 | ||
1228 | /* | |
1229 | * Internal hugetlb specific page flag. Do not use outside of the hugetlb | |
1230 | * code | |
1231 | */ | |
1232 | static inline bool PageHugeTemporary(struct page *page) | |
1233 | { | |
1234 | if (!PageHuge(page)) | |
1235 | return false; | |
1236 | ||
1237 | return (unsigned long)page[2].mapping == -1U; | |
1238 | } | |
1239 | ||
1240 | static inline void SetPageHugeTemporary(struct page *page) | |
1241 | { | |
1242 | page[2].mapping = (void *)-1U; | |
1243 | } | |
1244 | ||
1245 | static inline void ClearPageHugeTemporary(struct page *page) | |
1246 | { | |
1247 | page[2].mapping = NULL; | |
1248 | } | |
1249 | ||
1250 | void free_huge_page(struct page *page) | |
1251 | { | |
1252 | /* | |
1253 | * Can't pass hstate in here because it is called from the | |
1254 | * compound page destructor. | |
1255 | */ | |
1256 | struct hstate *h = page_hstate(page); | |
1257 | int nid = page_to_nid(page); | |
1258 | struct hugepage_subpool *spool = | |
1259 | (struct hugepage_subpool *)page_private(page); | |
1260 | bool restore_reserve; | |
1261 | ||
1262 | VM_BUG_ON_PAGE(page_count(page), page); | |
1263 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
1264 | ||
1265 | set_page_private(page, 0); | |
1266 | page->mapping = NULL; | |
1267 | restore_reserve = PagePrivate(page); | |
1268 | ClearPagePrivate(page); | |
1269 | ||
1270 | /* | |
1271 | * A return code of zero implies that the subpool will be under its | |
1272 | * minimum size if the reservation is not restored after page is free. | |
1273 | * Therefore, force restore_reserve operation. | |
1274 | */ | |
1275 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1276 | restore_reserve = true; | |
1277 | ||
1278 | spin_lock(&hugetlb_lock); | |
1279 | clear_page_huge_active(page); | |
1280 | hugetlb_cgroup_uncharge_page(hstate_index(h), | |
1281 | pages_per_huge_page(h), page); | |
1282 | if (restore_reserve) | |
1283 | h->resv_huge_pages++; | |
1284 | ||
1285 | if (PageHugeTemporary(page)) { | |
1286 | list_del(&page->lru); | |
1287 | ClearPageHugeTemporary(page); | |
1288 | update_and_free_page(h, page); | |
1289 | } else if (h->surplus_huge_pages_node[nid]) { | |
1290 | /* remove the page from active list */ | |
1291 | list_del(&page->lru); | |
1292 | update_and_free_page(h, page); | |
1293 | h->surplus_huge_pages--; | |
1294 | h->surplus_huge_pages_node[nid]--; | |
1295 | } else { | |
1296 | arch_clear_hugepage_flags(page); | |
1297 | enqueue_huge_page(h, page); | |
1298 | } | |
1299 | spin_unlock(&hugetlb_lock); | |
1300 | } | |
1301 | ||
1302 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) | |
1303 | { | |
1304 | INIT_LIST_HEAD(&page->lru); | |
1305 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); | |
1306 | spin_lock(&hugetlb_lock); | |
1307 | set_hugetlb_cgroup(page, NULL); | |
1308 | h->nr_huge_pages++; | |
1309 | h->nr_huge_pages_node[nid]++; | |
1310 | spin_unlock(&hugetlb_lock); | |
1311 | } | |
1312 | ||
1313 | static void prep_compound_gigantic_page(struct page *page, unsigned int order) | |
1314 | { | |
1315 | int i; | |
1316 | int nr_pages = 1 << order; | |
1317 | struct page *p = page + 1; | |
1318 | ||
1319 | /* we rely on prep_new_huge_page to set the destructor */ | |
1320 | set_compound_order(page, order); | |
1321 | __ClearPageReserved(page); | |
1322 | __SetPageHead(page); | |
1323 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { | |
1324 | /* | |
1325 | * For gigantic hugepages allocated through bootmem at | |
1326 | * boot, it's safer to be consistent with the not-gigantic | |
1327 | * hugepages and clear the PG_reserved bit from all tail pages | |
1328 | * too. Otherwse drivers using get_user_pages() to access tail | |
1329 | * pages may get the reference counting wrong if they see | |
1330 | * PG_reserved set on a tail page (despite the head page not | |
1331 | * having PG_reserved set). Enforcing this consistency between | |
1332 | * head and tail pages allows drivers to optimize away a check | |
1333 | * on the head page when they need know if put_page() is needed | |
1334 | * after get_user_pages(). | |
1335 | */ | |
1336 | __ClearPageReserved(p); | |
1337 | set_page_count(p, 0); | |
1338 | set_compound_head(p, page); | |
1339 | } | |
1340 | atomic_set(compound_mapcount_ptr(page), -1); | |
1341 | } | |
1342 | ||
1343 | /* | |
1344 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1345 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1346 | * details. | |
1347 | */ | |
1348 | int PageHuge(struct page *page) | |
1349 | { | |
1350 | if (!PageCompound(page)) | |
1351 | return 0; | |
1352 | ||
1353 | page = compound_head(page); | |
1354 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; | |
1355 | } | |
1356 | EXPORT_SYMBOL_GPL(PageHuge); | |
1357 | ||
1358 | /* | |
1359 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1360 | * normal or transparent huge pages. | |
1361 | */ | |
1362 | int PageHeadHuge(struct page *page_head) | |
1363 | { | |
1364 | if (!PageHead(page_head)) | |
1365 | return 0; | |
1366 | ||
1367 | return get_compound_page_dtor(page_head) == free_huge_page; | |
1368 | } | |
1369 | ||
1370 | pgoff_t __basepage_index(struct page *page) | |
1371 | { | |
1372 | struct page *page_head = compound_head(page); | |
1373 | pgoff_t index = page_index(page_head); | |
1374 | unsigned long compound_idx; | |
1375 | ||
1376 | if (!PageHuge(page_head)) | |
1377 | return page_index(page); | |
1378 | ||
1379 | if (compound_order(page_head) >= MAX_ORDER) | |
1380 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1381 | else | |
1382 | compound_idx = page - page_head; | |
1383 | ||
1384 | return (index << compound_order(page_head)) + compound_idx; | |
1385 | } | |
1386 | ||
1387 | static struct page *alloc_buddy_huge_page(struct hstate *h, | |
1388 | gfp_t gfp_mask, int nid, nodemask_t *nmask) | |
1389 | { | |
1390 | int order = huge_page_order(h); | |
1391 | struct page *page; | |
1392 | ||
1393 | gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN; | |
1394 | if (nid == NUMA_NO_NODE) | |
1395 | nid = numa_mem_id(); | |
1396 | page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); | |
1397 | if (page) | |
1398 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
1399 | else | |
1400 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
1401 | ||
1402 | return page; | |
1403 | } | |
1404 | ||
1405 | /* | |
1406 | * Common helper to allocate a fresh hugetlb page. All specific allocators | |
1407 | * should use this function to get new hugetlb pages | |
1408 | */ | |
1409 | static struct page *alloc_fresh_huge_page(struct hstate *h, | |
1410 | gfp_t gfp_mask, int nid, nodemask_t *nmask) | |
1411 | { | |
1412 | struct page *page; | |
1413 | ||
1414 | if (hstate_is_gigantic(h)) | |
1415 | page = alloc_gigantic_page(h, gfp_mask, nid, nmask); | |
1416 | else | |
1417 | page = alloc_buddy_huge_page(h, gfp_mask, | |
1418 | nid, nmask); | |
1419 | if (!page) | |
1420 | return NULL; | |
1421 | ||
1422 | if (hstate_is_gigantic(h)) | |
1423 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
1424 | prep_new_huge_page(h, page, page_to_nid(page)); | |
1425 | ||
1426 | return page; | |
1427 | } | |
1428 | ||
1429 | /* | |
1430 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
1431 | * manner. | |
1432 | */ | |
1433 | static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) | |
1434 | { | |
1435 | struct page *page; | |
1436 | int nr_nodes, node; | |
1437 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; | |
1438 | ||
1439 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1440 | page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed); | |
1441 | if (page) | |
1442 | break; | |
1443 | } | |
1444 | ||
1445 | if (!page) | |
1446 | return 0; | |
1447 | ||
1448 | put_page(page); /* free it into the hugepage allocator */ | |
1449 | ||
1450 | return 1; | |
1451 | } | |
1452 | ||
1453 | /* | |
1454 | * Free huge page from pool from next node to free. | |
1455 | * Attempt to keep persistent huge pages more or less | |
1456 | * balanced over allowed nodes. | |
1457 | * Called with hugetlb_lock locked. | |
1458 | */ | |
1459 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, | |
1460 | bool acct_surplus) | |
1461 | { | |
1462 | int nr_nodes, node; | |
1463 | int ret = 0; | |
1464 | ||
1465 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
1466 | /* | |
1467 | * If we're returning unused surplus pages, only examine | |
1468 | * nodes with surplus pages. | |
1469 | */ | |
1470 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && | |
1471 | !list_empty(&h->hugepage_freelists[node])) { | |
1472 | struct page *page = | |
1473 | list_entry(h->hugepage_freelists[node].next, | |
1474 | struct page, lru); | |
1475 | list_del(&page->lru); | |
1476 | h->free_huge_pages--; | |
1477 | h->free_huge_pages_node[node]--; | |
1478 | if (acct_surplus) { | |
1479 | h->surplus_huge_pages--; | |
1480 | h->surplus_huge_pages_node[node]--; | |
1481 | } | |
1482 | update_and_free_page(h, page); | |
1483 | ret = 1; | |
1484 | break; | |
1485 | } | |
1486 | } | |
1487 | ||
1488 | return ret; | |
1489 | } | |
1490 | ||
1491 | /* | |
1492 | * Dissolve a given free hugepage into free buddy pages. This function does | |
1493 | * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the | |
1494 | * dissolution fails because a give page is not a free hugepage, or because | |
1495 | * free hugepages are fully reserved. | |
1496 | */ | |
1497 | int dissolve_free_huge_page(struct page *page) | |
1498 | { | |
1499 | int rc = -EBUSY; | |
1500 | ||
1501 | spin_lock(&hugetlb_lock); | |
1502 | if (PageHuge(page) && !page_count(page)) { | |
1503 | struct page *head = compound_head(page); | |
1504 | struct hstate *h = page_hstate(head); | |
1505 | int nid = page_to_nid(head); | |
1506 | if (h->free_huge_pages - h->resv_huge_pages == 0) | |
1507 | goto out; | |
1508 | /* | |
1509 | * Move PageHWPoison flag from head page to the raw error page, | |
1510 | * which makes any subpages rather than the error page reusable. | |
1511 | */ | |
1512 | if (PageHWPoison(head) && page != head) { | |
1513 | SetPageHWPoison(page); | |
1514 | ClearPageHWPoison(head); | |
1515 | } | |
1516 | list_del(&head->lru); | |
1517 | h->free_huge_pages--; | |
1518 | h->free_huge_pages_node[nid]--; | |
1519 | h->max_huge_pages--; | |
1520 | update_and_free_page(h, head); | |
1521 | rc = 0; | |
1522 | } | |
1523 | out: | |
1524 | spin_unlock(&hugetlb_lock); | |
1525 | return rc; | |
1526 | } | |
1527 | ||
1528 | /* | |
1529 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1530 | * make specified memory blocks removable from the system. | |
1531 | * Note that this will dissolve a free gigantic hugepage completely, if any | |
1532 | * part of it lies within the given range. | |
1533 | * Also note that if dissolve_free_huge_page() returns with an error, all | |
1534 | * free hugepages that were dissolved before that error are lost. | |
1535 | */ | |
1536 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) | |
1537 | { | |
1538 | unsigned long pfn; | |
1539 | struct page *page; | |
1540 | int rc = 0; | |
1541 | ||
1542 | if (!hugepages_supported()) | |
1543 | return rc; | |
1544 | ||
1545 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { | |
1546 | page = pfn_to_page(pfn); | |
1547 | if (PageHuge(page) && !page_count(page)) { | |
1548 | rc = dissolve_free_huge_page(page); | |
1549 | if (rc) | |
1550 | break; | |
1551 | } | |
1552 | } | |
1553 | ||
1554 | return rc; | |
1555 | } | |
1556 | ||
1557 | /* | |
1558 | * Allocates a fresh surplus page from the page allocator. | |
1559 | */ | |
1560 | static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, | |
1561 | int nid, nodemask_t *nmask) | |
1562 | { | |
1563 | struct page *page = NULL; | |
1564 | ||
1565 | if (hstate_is_gigantic(h)) | |
1566 | return NULL; | |
1567 | ||
1568 | spin_lock(&hugetlb_lock); | |
1569 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) | |
1570 | goto out_unlock; | |
1571 | spin_unlock(&hugetlb_lock); | |
1572 | ||
1573 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask); | |
1574 | if (!page) | |
1575 | return NULL; | |
1576 | ||
1577 | spin_lock(&hugetlb_lock); | |
1578 | /* | |
1579 | * We could have raced with the pool size change. | |
1580 | * Double check that and simply deallocate the new page | |
1581 | * if we would end up overcommiting the surpluses. Abuse | |
1582 | * temporary page to workaround the nasty free_huge_page | |
1583 | * codeflow | |
1584 | */ | |
1585 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
1586 | SetPageHugeTemporary(page); | |
1587 | spin_unlock(&hugetlb_lock); | |
1588 | put_page(page); | |
1589 | return NULL; | |
1590 | } else { | |
1591 | h->surplus_huge_pages++; | |
1592 | h->surplus_huge_pages_node[page_to_nid(page)]++; | |
1593 | } | |
1594 | ||
1595 | out_unlock: | |
1596 | spin_unlock(&hugetlb_lock); | |
1597 | ||
1598 | return page; | |
1599 | } | |
1600 | ||
1601 | struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, | |
1602 | int nid, nodemask_t *nmask) | |
1603 | { | |
1604 | struct page *page; | |
1605 | ||
1606 | if (hstate_is_gigantic(h)) | |
1607 | return NULL; | |
1608 | ||
1609 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask); | |
1610 | if (!page) | |
1611 | return NULL; | |
1612 | ||
1613 | /* | |
1614 | * We do not account these pages as surplus because they are only | |
1615 | * temporary and will be released properly on the last reference | |
1616 | */ | |
1617 | SetPageHugeTemporary(page); | |
1618 | ||
1619 | return page; | |
1620 | } | |
1621 | ||
1622 | /* | |
1623 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
1624 | */ | |
1625 | static | |
1626 | struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, | |
1627 | struct vm_area_struct *vma, unsigned long addr) | |
1628 | { | |
1629 | struct page *page; | |
1630 | struct mempolicy *mpol; | |
1631 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
1632 | int nid; | |
1633 | nodemask_t *nodemask; | |
1634 | ||
1635 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
1636 | page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); | |
1637 | mpol_cond_put(mpol); | |
1638 | ||
1639 | return page; | |
1640 | } | |
1641 | ||
1642 | /* page migration callback function */ | |
1643 | struct page *alloc_huge_page_node(struct hstate *h, int nid) | |
1644 | { | |
1645 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
1646 | struct page *page = NULL; | |
1647 | ||
1648 | if (nid != NUMA_NO_NODE) | |
1649 | gfp_mask |= __GFP_THISNODE; | |
1650 | ||
1651 | spin_lock(&hugetlb_lock); | |
1652 | if (h->free_huge_pages - h->resv_huge_pages > 0) | |
1653 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL); | |
1654 | spin_unlock(&hugetlb_lock); | |
1655 | ||
1656 | if (!page) | |
1657 | page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL); | |
1658 | ||
1659 | return page; | |
1660 | } | |
1661 | ||
1662 | /* page migration callback function */ | |
1663 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, | |
1664 | nodemask_t *nmask) | |
1665 | { | |
1666 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
1667 | ||
1668 | spin_lock(&hugetlb_lock); | |
1669 | if (h->free_huge_pages - h->resv_huge_pages > 0) { | |
1670 | struct page *page; | |
1671 | ||
1672 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
1673 | if (page) { | |
1674 | spin_unlock(&hugetlb_lock); | |
1675 | return page; | |
1676 | } | |
1677 | } | |
1678 | spin_unlock(&hugetlb_lock); | |
1679 | ||
1680 | return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); | |
1681 | } | |
1682 | ||
1683 | /* mempolicy aware migration callback */ | |
1684 | struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, | |
1685 | unsigned long address) | |
1686 | { | |
1687 | struct mempolicy *mpol; | |
1688 | nodemask_t *nodemask; | |
1689 | struct page *page; | |
1690 | gfp_t gfp_mask; | |
1691 | int node; | |
1692 | ||
1693 | gfp_mask = htlb_alloc_mask(h); | |
1694 | node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
1695 | page = alloc_huge_page_nodemask(h, node, nodemask); | |
1696 | mpol_cond_put(mpol); | |
1697 | ||
1698 | return page; | |
1699 | } | |
1700 | ||
1701 | /* | |
1702 | * Increase the hugetlb pool such that it can accommodate a reservation | |
1703 | * of size 'delta'. | |
1704 | */ | |
1705 | static int gather_surplus_pages(struct hstate *h, int delta) | |
1706 | { | |
1707 | struct list_head surplus_list; | |
1708 | struct page *page, *tmp; | |
1709 | int ret, i; | |
1710 | int needed, allocated; | |
1711 | bool alloc_ok = true; | |
1712 | ||
1713 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; | |
1714 | if (needed <= 0) { | |
1715 | h->resv_huge_pages += delta; | |
1716 | return 0; | |
1717 | } | |
1718 | ||
1719 | allocated = 0; | |
1720 | INIT_LIST_HEAD(&surplus_list); | |
1721 | ||
1722 | ret = -ENOMEM; | |
1723 | retry: | |
1724 | spin_unlock(&hugetlb_lock); | |
1725 | for (i = 0; i < needed; i++) { | |
1726 | page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), | |
1727 | NUMA_NO_NODE, NULL); | |
1728 | if (!page) { | |
1729 | alloc_ok = false; | |
1730 | break; | |
1731 | } | |
1732 | list_add(&page->lru, &surplus_list); | |
1733 | cond_resched(); | |
1734 | } | |
1735 | allocated += i; | |
1736 | ||
1737 | /* | |
1738 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1739 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1740 | */ | |
1741 | spin_lock(&hugetlb_lock); | |
1742 | needed = (h->resv_huge_pages + delta) - | |
1743 | (h->free_huge_pages + allocated); | |
1744 | if (needed > 0) { | |
1745 | if (alloc_ok) | |
1746 | goto retry; | |
1747 | /* | |
1748 | * We were not able to allocate enough pages to | |
1749 | * satisfy the entire reservation so we free what | |
1750 | * we've allocated so far. | |
1751 | */ | |
1752 | goto free; | |
1753 | } | |
1754 | /* | |
1755 | * The surplus_list now contains _at_least_ the number of extra pages | |
1756 | * needed to accommodate the reservation. Add the appropriate number | |
1757 | * of pages to the hugetlb pool and free the extras back to the buddy | |
1758 | * allocator. Commit the entire reservation here to prevent another | |
1759 | * process from stealing the pages as they are added to the pool but | |
1760 | * before they are reserved. | |
1761 | */ | |
1762 | needed += allocated; | |
1763 | h->resv_huge_pages += delta; | |
1764 | ret = 0; | |
1765 | ||
1766 | /* Free the needed pages to the hugetlb pool */ | |
1767 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { | |
1768 | if ((--needed) < 0) | |
1769 | break; | |
1770 | /* | |
1771 | * This page is now managed by the hugetlb allocator and has | |
1772 | * no users -- drop the buddy allocator's reference. | |
1773 | */ | |
1774 | put_page_testzero(page); | |
1775 | VM_BUG_ON_PAGE(page_count(page), page); | |
1776 | enqueue_huge_page(h, page); | |
1777 | } | |
1778 | free: | |
1779 | spin_unlock(&hugetlb_lock); | |
1780 | ||
1781 | /* Free unnecessary surplus pages to the buddy allocator */ | |
1782 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) | |
1783 | put_page(page); | |
1784 | spin_lock(&hugetlb_lock); | |
1785 | ||
1786 | return ret; | |
1787 | } | |
1788 | ||
1789 | /* | |
1790 | * This routine has two main purposes: | |
1791 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
1792 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
1793 | * to the associated reservation map. | |
1794 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
1795 | * the reservation. As many as unused_resv_pages may be freed. | |
1796 | * | |
1797 | * Called with hugetlb_lock held. However, the lock could be dropped (and | |
1798 | * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, | |
1799 | * we must make sure nobody else can claim pages we are in the process of | |
1800 | * freeing. Do this by ensuring resv_huge_page always is greater than the | |
1801 | * number of huge pages we plan to free when dropping the lock. | |
1802 | */ | |
1803 | static void return_unused_surplus_pages(struct hstate *h, | |
1804 | unsigned long unused_resv_pages) | |
1805 | { | |
1806 | unsigned long nr_pages; | |
1807 | ||
1808 | /* Cannot return gigantic pages currently */ | |
1809 | if (hstate_is_gigantic(h)) | |
1810 | goto out; | |
1811 | ||
1812 | /* | |
1813 | * Part (or even all) of the reservation could have been backed | |
1814 | * by pre-allocated pages. Only free surplus pages. | |
1815 | */ | |
1816 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); | |
1817 | ||
1818 | /* | |
1819 | * We want to release as many surplus pages as possible, spread | |
1820 | * evenly across all nodes with memory. Iterate across these nodes | |
1821 | * until we can no longer free unreserved surplus pages. This occurs | |
1822 | * when the nodes with surplus pages have no free pages. | |
1823 | * free_pool_huge_page() will balance the the freed pages across the | |
1824 | * on-line nodes with memory and will handle the hstate accounting. | |
1825 | * | |
1826 | * Note that we decrement resv_huge_pages as we free the pages. If | |
1827 | * we drop the lock, resv_huge_pages will still be sufficiently large | |
1828 | * to cover subsequent pages we may free. | |
1829 | */ | |
1830 | while (nr_pages--) { | |
1831 | h->resv_huge_pages--; | |
1832 | unused_resv_pages--; | |
1833 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) | |
1834 | goto out; | |
1835 | cond_resched_lock(&hugetlb_lock); | |
1836 | } | |
1837 | ||
1838 | out: | |
1839 | /* Fully uncommit the reservation */ | |
1840 | h->resv_huge_pages -= unused_resv_pages; | |
1841 | } | |
1842 | ||
1843 | ||
1844 | /* | |
1845 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation | |
1846 | * are used by the huge page allocation routines to manage reservations. | |
1847 | * | |
1848 | * vma_needs_reservation is called to determine if the huge page at addr | |
1849 | * within the vma has an associated reservation. If a reservation is | |
1850 | * needed, the value 1 is returned. The caller is then responsible for | |
1851 | * managing the global reservation and subpool usage counts. After | |
1852 | * the huge page has been allocated, vma_commit_reservation is called | |
1853 | * to add the page to the reservation map. If the page allocation fails, | |
1854 | * the reservation must be ended instead of committed. vma_end_reservation | |
1855 | * is called in such cases. | |
1856 | * | |
1857 | * In the normal case, vma_commit_reservation returns the same value | |
1858 | * as the preceding vma_needs_reservation call. The only time this | |
1859 | * is not the case is if a reserve map was changed between calls. It | |
1860 | * is the responsibility of the caller to notice the difference and | |
1861 | * take appropriate action. | |
1862 | * | |
1863 | * vma_add_reservation is used in error paths where a reservation must | |
1864 | * be restored when a newly allocated huge page must be freed. It is | |
1865 | * to be called after calling vma_needs_reservation to determine if a | |
1866 | * reservation exists. | |
1867 | */ | |
1868 | enum vma_resv_mode { | |
1869 | VMA_NEEDS_RESV, | |
1870 | VMA_COMMIT_RESV, | |
1871 | VMA_END_RESV, | |
1872 | VMA_ADD_RESV, | |
1873 | }; | |
1874 | static long __vma_reservation_common(struct hstate *h, | |
1875 | struct vm_area_struct *vma, unsigned long addr, | |
1876 | enum vma_resv_mode mode) | |
1877 | { | |
1878 | struct resv_map *resv; | |
1879 | pgoff_t idx; | |
1880 | long ret; | |
1881 | ||
1882 | resv = vma_resv_map(vma); | |
1883 | if (!resv) | |
1884 | return 1; | |
1885 | ||
1886 | idx = vma_hugecache_offset(h, vma, addr); | |
1887 | switch (mode) { | |
1888 | case VMA_NEEDS_RESV: | |
1889 | ret = region_chg(resv, idx, idx + 1); | |
1890 | break; | |
1891 | case VMA_COMMIT_RESV: | |
1892 | ret = region_add(resv, idx, idx + 1); | |
1893 | break; | |
1894 | case VMA_END_RESV: | |
1895 | region_abort(resv, idx, idx + 1); | |
1896 | ret = 0; | |
1897 | break; | |
1898 | case VMA_ADD_RESV: | |
1899 | if (vma->vm_flags & VM_MAYSHARE) | |
1900 | ret = region_add(resv, idx, idx + 1); | |
1901 | else { | |
1902 | region_abort(resv, idx, idx + 1); | |
1903 | ret = region_del(resv, idx, idx + 1); | |
1904 | } | |
1905 | break; | |
1906 | default: | |
1907 | BUG(); | |
1908 | } | |
1909 | ||
1910 | if (vma->vm_flags & VM_MAYSHARE) | |
1911 | return ret; | |
1912 | else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { | |
1913 | /* | |
1914 | * In most cases, reserves always exist for private mappings. | |
1915 | * However, a file associated with mapping could have been | |
1916 | * hole punched or truncated after reserves were consumed. | |
1917 | * As subsequent fault on such a range will not use reserves. | |
1918 | * Subtle - The reserve map for private mappings has the | |
1919 | * opposite meaning than that of shared mappings. If NO | |
1920 | * entry is in the reserve map, it means a reservation exists. | |
1921 | * If an entry exists in the reserve map, it means the | |
1922 | * reservation has already been consumed. As a result, the | |
1923 | * return value of this routine is the opposite of the | |
1924 | * value returned from reserve map manipulation routines above. | |
1925 | */ | |
1926 | if (ret) | |
1927 | return 0; | |
1928 | else | |
1929 | return 1; | |
1930 | } | |
1931 | else | |
1932 | return ret < 0 ? ret : 0; | |
1933 | } | |
1934 | ||
1935 | static long vma_needs_reservation(struct hstate *h, | |
1936 | struct vm_area_struct *vma, unsigned long addr) | |
1937 | { | |
1938 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); | |
1939 | } | |
1940 | ||
1941 | static long vma_commit_reservation(struct hstate *h, | |
1942 | struct vm_area_struct *vma, unsigned long addr) | |
1943 | { | |
1944 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); | |
1945 | } | |
1946 | ||
1947 | static void vma_end_reservation(struct hstate *h, | |
1948 | struct vm_area_struct *vma, unsigned long addr) | |
1949 | { | |
1950 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); | |
1951 | } | |
1952 | ||
1953 | static long vma_add_reservation(struct hstate *h, | |
1954 | struct vm_area_struct *vma, unsigned long addr) | |
1955 | { | |
1956 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
1957 | } | |
1958 | ||
1959 | /* | |
1960 | * This routine is called to restore a reservation on error paths. In the | |
1961 | * specific error paths, a huge page was allocated (via alloc_huge_page) | |
1962 | * and is about to be freed. If a reservation for the page existed, | |
1963 | * alloc_huge_page would have consumed the reservation and set PagePrivate | |
1964 | * in the newly allocated page. When the page is freed via free_huge_page, | |
1965 | * the global reservation count will be incremented if PagePrivate is set. | |
1966 | * However, free_huge_page can not adjust the reserve map. Adjust the | |
1967 | * reserve map here to be consistent with global reserve count adjustments | |
1968 | * to be made by free_huge_page. | |
1969 | */ | |
1970 | static void restore_reserve_on_error(struct hstate *h, | |
1971 | struct vm_area_struct *vma, unsigned long address, | |
1972 | struct page *page) | |
1973 | { | |
1974 | if (unlikely(PagePrivate(page))) { | |
1975 | long rc = vma_needs_reservation(h, vma, address); | |
1976 | ||
1977 | if (unlikely(rc < 0)) { | |
1978 | /* | |
1979 | * Rare out of memory condition in reserve map | |
1980 | * manipulation. Clear PagePrivate so that | |
1981 | * global reserve count will not be incremented | |
1982 | * by free_huge_page. This will make it appear | |
1983 | * as though the reservation for this page was | |
1984 | * consumed. This may prevent the task from | |
1985 | * faulting in the page at a later time. This | |
1986 | * is better than inconsistent global huge page | |
1987 | * accounting of reserve counts. | |
1988 | */ | |
1989 | ClearPagePrivate(page); | |
1990 | } else if (rc) { | |
1991 | rc = vma_add_reservation(h, vma, address); | |
1992 | if (unlikely(rc < 0)) | |
1993 | /* | |
1994 | * See above comment about rare out of | |
1995 | * memory condition. | |
1996 | */ | |
1997 | ClearPagePrivate(page); | |
1998 | } else | |
1999 | vma_end_reservation(h, vma, address); | |
2000 | } | |
2001 | } | |
2002 | ||
2003 | struct page *alloc_huge_page(struct vm_area_struct *vma, | |
2004 | unsigned long addr, int avoid_reserve) | |
2005 | { | |
2006 | struct hugepage_subpool *spool = subpool_vma(vma); | |
2007 | struct hstate *h = hstate_vma(vma); | |
2008 | struct page *page; | |
2009 | long map_chg, map_commit; | |
2010 | long gbl_chg; | |
2011 | int ret, idx; | |
2012 | struct hugetlb_cgroup *h_cg; | |
2013 | ||
2014 | idx = hstate_index(h); | |
2015 | /* | |
2016 | * Examine the region/reserve map to determine if the process | |
2017 | * has a reservation for the page to be allocated. A return | |
2018 | * code of zero indicates a reservation exists (no change). | |
2019 | */ | |
2020 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); | |
2021 | if (map_chg < 0) | |
2022 | return ERR_PTR(-ENOMEM); | |
2023 | ||
2024 | /* | |
2025 | * Processes that did not create the mapping will have no | |
2026 | * reserves as indicated by the region/reserve map. Check | |
2027 | * that the allocation will not exceed the subpool limit. | |
2028 | * Allocations for MAP_NORESERVE mappings also need to be | |
2029 | * checked against any subpool limit. | |
2030 | */ | |
2031 | if (map_chg || avoid_reserve) { | |
2032 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2033 | if (gbl_chg < 0) { | |
2034 | vma_end_reservation(h, vma, addr); | |
2035 | return ERR_PTR(-ENOSPC); | |
2036 | } | |
2037 | ||
2038 | /* | |
2039 | * Even though there was no reservation in the region/reserve | |
2040 | * map, there could be reservations associated with the | |
2041 | * subpool that can be used. This would be indicated if the | |
2042 | * return value of hugepage_subpool_get_pages() is zero. | |
2043 | * However, if avoid_reserve is specified we still avoid even | |
2044 | * the subpool reservations. | |
2045 | */ | |
2046 | if (avoid_reserve) | |
2047 | gbl_chg = 1; | |
2048 | } | |
2049 | ||
2050 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); | |
2051 | if (ret) | |
2052 | goto out_subpool_put; | |
2053 | ||
2054 | spin_lock(&hugetlb_lock); | |
2055 | /* | |
2056 | * glb_chg is passed to indicate whether or not a page must be taken | |
2057 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2058 | * a reservation exists for the allocation. | |
2059 | */ | |
2060 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
2061 | if (!page) { | |
2062 | spin_unlock(&hugetlb_lock); | |
2063 | page = alloc_buddy_huge_page_with_mpol(h, vma, addr); | |
2064 | if (!page) | |
2065 | goto out_uncharge_cgroup; | |
2066 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { | |
2067 | SetPagePrivate(page); | |
2068 | h->resv_huge_pages--; | |
2069 | } | |
2070 | spin_lock(&hugetlb_lock); | |
2071 | list_move(&page->lru, &h->hugepage_activelist); | |
2072 | /* Fall through */ | |
2073 | } | |
2074 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); | |
2075 | spin_unlock(&hugetlb_lock); | |
2076 | ||
2077 | set_page_private(page, (unsigned long)spool); | |
2078 | ||
2079 | map_commit = vma_commit_reservation(h, vma, addr); | |
2080 | if (unlikely(map_chg > map_commit)) { | |
2081 | /* | |
2082 | * The page was added to the reservation map between | |
2083 | * vma_needs_reservation and vma_commit_reservation. | |
2084 | * This indicates a race with hugetlb_reserve_pages. | |
2085 | * Adjust for the subpool count incremented above AND | |
2086 | * in hugetlb_reserve_pages for the same page. Also, | |
2087 | * the reservation count added in hugetlb_reserve_pages | |
2088 | * no longer applies. | |
2089 | */ | |
2090 | long rsv_adjust; | |
2091 | ||
2092 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2093 | hugetlb_acct_memory(h, -rsv_adjust); | |
2094 | } | |
2095 | return page; | |
2096 | ||
2097 | out_uncharge_cgroup: | |
2098 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
2099 | out_subpool_put: | |
2100 | if (map_chg || avoid_reserve) | |
2101 | hugepage_subpool_put_pages(spool, 1); | |
2102 | vma_end_reservation(h, vma, addr); | |
2103 | return ERR_PTR(-ENOSPC); | |
2104 | } | |
2105 | ||
2106 | int alloc_bootmem_huge_page(struct hstate *h) | |
2107 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); | |
2108 | int __alloc_bootmem_huge_page(struct hstate *h) | |
2109 | { | |
2110 | struct huge_bootmem_page *m; | |
2111 | int nr_nodes, node; | |
2112 | ||
2113 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { | |
2114 | void *addr; | |
2115 | ||
2116 | addr = memblock_alloc_try_nid_raw( | |
2117 | huge_page_size(h), huge_page_size(h), | |
2118 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); | |
2119 | if (addr) { | |
2120 | /* | |
2121 | * Use the beginning of the huge page to store the | |
2122 | * huge_bootmem_page struct (until gather_bootmem | |
2123 | * puts them into the mem_map). | |
2124 | */ | |
2125 | m = addr; | |
2126 | goto found; | |
2127 | } | |
2128 | } | |
2129 | return 0; | |
2130 | ||
2131 | found: | |
2132 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); | |
2133 | /* Put them into a private list first because mem_map is not up yet */ | |
2134 | INIT_LIST_HEAD(&m->list); | |
2135 | list_add(&m->list, &huge_boot_pages); | |
2136 | m->hstate = h; | |
2137 | return 1; | |
2138 | } | |
2139 | ||
2140 | static void __init prep_compound_huge_page(struct page *page, | |
2141 | unsigned int order) | |
2142 | { | |
2143 | if (unlikely(order > (MAX_ORDER - 1))) | |
2144 | prep_compound_gigantic_page(page, order); | |
2145 | else | |
2146 | prep_compound_page(page, order); | |
2147 | } | |
2148 | ||
2149 | /* Put bootmem huge pages into the standard lists after mem_map is up */ | |
2150 | static void __init gather_bootmem_prealloc(void) | |
2151 | { | |
2152 | struct huge_bootmem_page *m; | |
2153 | ||
2154 | list_for_each_entry(m, &huge_boot_pages, list) { | |
2155 | struct page *page = virt_to_page(m); | |
2156 | struct hstate *h = m->hstate; | |
2157 | ||
2158 | WARN_ON(page_count(page) != 1); | |
2159 | prep_compound_huge_page(page, h->order); | |
2160 | WARN_ON(PageReserved(page)); | |
2161 | prep_new_huge_page(h, page, page_to_nid(page)); | |
2162 | put_page(page); /* free it into the hugepage allocator */ | |
2163 | ||
2164 | /* | |
2165 | * If we had gigantic hugepages allocated at boot time, we need | |
2166 | * to restore the 'stolen' pages to totalram_pages in order to | |
2167 | * fix confusing memory reports from free(1) and another | |
2168 | * side-effects, like CommitLimit going negative. | |
2169 | */ | |
2170 | if (hstate_is_gigantic(h)) | |
2171 | adjust_managed_page_count(page, 1 << h->order); | |
2172 | cond_resched(); | |
2173 | } | |
2174 | } | |
2175 | ||
2176 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) | |
2177 | { | |
2178 | unsigned long i; | |
2179 | ||
2180 | for (i = 0; i < h->max_huge_pages; ++i) { | |
2181 | if (hstate_is_gigantic(h)) { | |
2182 | if (!alloc_bootmem_huge_page(h)) | |
2183 | break; | |
2184 | } else if (!alloc_pool_huge_page(h, | |
2185 | &node_states[N_MEMORY])) | |
2186 | break; | |
2187 | cond_resched(); | |
2188 | } | |
2189 | if (i < h->max_huge_pages) { | |
2190 | char buf[32]; | |
2191 | ||
2192 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
2193 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", | |
2194 | h->max_huge_pages, buf, i); | |
2195 | h->max_huge_pages = i; | |
2196 | } | |
2197 | } | |
2198 | ||
2199 | static void __init hugetlb_init_hstates(void) | |
2200 | { | |
2201 | struct hstate *h; | |
2202 | ||
2203 | for_each_hstate(h) { | |
2204 | if (minimum_order > huge_page_order(h)) | |
2205 | minimum_order = huge_page_order(h); | |
2206 | ||
2207 | /* oversize hugepages were init'ed in early boot */ | |
2208 | if (!hstate_is_gigantic(h)) | |
2209 | hugetlb_hstate_alloc_pages(h); | |
2210 | } | |
2211 | VM_BUG_ON(minimum_order == UINT_MAX); | |
2212 | } | |
2213 | ||
2214 | static void __init report_hugepages(void) | |
2215 | { | |
2216 | struct hstate *h; | |
2217 | ||
2218 | for_each_hstate(h) { | |
2219 | char buf[32]; | |
2220 | ||
2221 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
2222 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", | |
2223 | buf, h->free_huge_pages); | |
2224 | } | |
2225 | } | |
2226 | ||
2227 | #ifdef CONFIG_HIGHMEM | |
2228 | static void try_to_free_low(struct hstate *h, unsigned long count, | |
2229 | nodemask_t *nodes_allowed) | |
2230 | { | |
2231 | int i; | |
2232 | ||
2233 | if (hstate_is_gigantic(h)) | |
2234 | return; | |
2235 | ||
2236 | for_each_node_mask(i, *nodes_allowed) { | |
2237 | struct page *page, *next; | |
2238 | struct list_head *freel = &h->hugepage_freelists[i]; | |
2239 | list_for_each_entry_safe(page, next, freel, lru) { | |
2240 | if (count >= h->nr_huge_pages) | |
2241 | return; | |
2242 | if (PageHighMem(page)) | |
2243 | continue; | |
2244 | list_del(&page->lru); | |
2245 | update_and_free_page(h, page); | |
2246 | h->free_huge_pages--; | |
2247 | h->free_huge_pages_node[page_to_nid(page)]--; | |
2248 | } | |
2249 | } | |
2250 | } | |
2251 | #else | |
2252 | static inline void try_to_free_low(struct hstate *h, unsigned long count, | |
2253 | nodemask_t *nodes_allowed) | |
2254 | { | |
2255 | } | |
2256 | #endif | |
2257 | ||
2258 | /* | |
2259 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2260 | * balanced by operating on them in a round-robin fashion. | |
2261 | * Returns 1 if an adjustment was made. | |
2262 | */ | |
2263 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, | |
2264 | int delta) | |
2265 | { | |
2266 | int nr_nodes, node; | |
2267 | ||
2268 | VM_BUG_ON(delta != -1 && delta != 1); | |
2269 | ||
2270 | if (delta < 0) { | |
2271 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2272 | if (h->surplus_huge_pages_node[node]) | |
2273 | goto found; | |
2274 | } | |
2275 | } else { | |
2276 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2277 | if (h->surplus_huge_pages_node[node] < | |
2278 | h->nr_huge_pages_node[node]) | |
2279 | goto found; | |
2280 | } | |
2281 | } | |
2282 | return 0; | |
2283 | ||
2284 | found: | |
2285 | h->surplus_huge_pages += delta; | |
2286 | h->surplus_huge_pages_node[node] += delta; | |
2287 | return 1; | |
2288 | } | |
2289 | ||
2290 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) | |
2291 | static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, | |
2292 | nodemask_t *nodes_allowed) | |
2293 | { | |
2294 | unsigned long min_count, ret; | |
2295 | ||
2296 | spin_lock(&hugetlb_lock); | |
2297 | ||
2298 | /* | |
2299 | * Check for a node specific request. | |
2300 | * Changing node specific huge page count may require a corresponding | |
2301 | * change to the global count. In any case, the passed node mask | |
2302 | * (nodes_allowed) will restrict alloc/free to the specified node. | |
2303 | */ | |
2304 | if (nid != NUMA_NO_NODE) { | |
2305 | unsigned long old_count = count; | |
2306 | ||
2307 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2308 | /* | |
2309 | * User may have specified a large count value which caused the | |
2310 | * above calculation to overflow. In this case, they wanted | |
2311 | * to allocate as many huge pages as possible. Set count to | |
2312 | * largest possible value to align with their intention. | |
2313 | */ | |
2314 | if (count < old_count) | |
2315 | count = ULONG_MAX; | |
2316 | } | |
2317 | ||
2318 | /* | |
2319 | * Gigantic pages runtime allocation depend on the capability for large | |
2320 | * page range allocation. | |
2321 | * If the system does not provide this feature, return an error when | |
2322 | * the user tries to allocate gigantic pages but let the user free the | |
2323 | * boottime allocated gigantic pages. | |
2324 | */ | |
2325 | if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { | |
2326 | if (count > persistent_huge_pages(h)) { | |
2327 | spin_unlock(&hugetlb_lock); | |
2328 | return -EINVAL; | |
2329 | } | |
2330 | /* Fall through to decrease pool */ | |
2331 | } | |
2332 | ||
2333 | /* | |
2334 | * Increase the pool size | |
2335 | * First take pages out of surplus state. Then make up the | |
2336 | * remaining difference by allocating fresh huge pages. | |
2337 | * | |
2338 | * We might race with alloc_surplus_huge_page() here and be unable | |
2339 | * to convert a surplus huge page to a normal huge page. That is | |
2340 | * not critical, though, it just means the overall size of the | |
2341 | * pool might be one hugepage larger than it needs to be, but | |
2342 | * within all the constraints specified by the sysctls. | |
2343 | */ | |
2344 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { | |
2345 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) | |
2346 | break; | |
2347 | } | |
2348 | ||
2349 | while (count > persistent_huge_pages(h)) { | |
2350 | /* | |
2351 | * If this allocation races such that we no longer need the | |
2352 | * page, free_huge_page will handle it by freeing the page | |
2353 | * and reducing the surplus. | |
2354 | */ | |
2355 | spin_unlock(&hugetlb_lock); | |
2356 | ||
2357 | /* yield cpu to avoid soft lockup */ | |
2358 | cond_resched(); | |
2359 | ||
2360 | ret = alloc_pool_huge_page(h, nodes_allowed); | |
2361 | spin_lock(&hugetlb_lock); | |
2362 | if (!ret) | |
2363 | goto out; | |
2364 | ||
2365 | /* Bail for signals. Probably ctrl-c from user */ | |
2366 | if (signal_pending(current)) | |
2367 | goto out; | |
2368 | } | |
2369 | ||
2370 | /* | |
2371 | * Decrease the pool size | |
2372 | * First return free pages to the buddy allocator (being careful | |
2373 | * to keep enough around to satisfy reservations). Then place | |
2374 | * pages into surplus state as needed so the pool will shrink | |
2375 | * to the desired size as pages become free. | |
2376 | * | |
2377 | * By placing pages into the surplus state independent of the | |
2378 | * overcommit value, we are allowing the surplus pool size to | |
2379 | * exceed overcommit. There are few sane options here. Since | |
2380 | * alloc_surplus_huge_page() is checking the global counter, | |
2381 | * though, we'll note that we're not allowed to exceed surplus | |
2382 | * and won't grow the pool anywhere else. Not until one of the | |
2383 | * sysctls are changed, or the surplus pages go out of use. | |
2384 | */ | |
2385 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; | |
2386 | min_count = max(count, min_count); | |
2387 | try_to_free_low(h, min_count, nodes_allowed); | |
2388 | while (min_count < persistent_huge_pages(h)) { | |
2389 | if (!free_pool_huge_page(h, nodes_allowed, 0)) | |
2390 | break; | |
2391 | cond_resched_lock(&hugetlb_lock); | |
2392 | } | |
2393 | while (count < persistent_huge_pages(h)) { | |
2394 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) | |
2395 | break; | |
2396 | } | |
2397 | out: | |
2398 | h->max_huge_pages = persistent_huge_pages(h); | |
2399 | spin_unlock(&hugetlb_lock); | |
2400 | ||
2401 | return 0; | |
2402 | } | |
2403 | ||
2404 | #define HSTATE_ATTR_RO(_name) \ | |
2405 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2406 | ||
2407 | #define HSTATE_ATTR(_name) \ | |
2408 | static struct kobj_attribute _name##_attr = \ | |
2409 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2410 | ||
2411 | static struct kobject *hugepages_kobj; | |
2412 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2413 | ||
2414 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); | |
2415 | ||
2416 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
2417 | { | |
2418 | int i; | |
2419 | ||
2420 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
2421 | if (hstate_kobjs[i] == kobj) { | |
2422 | if (nidp) | |
2423 | *nidp = NUMA_NO_NODE; | |
2424 | return &hstates[i]; | |
2425 | } | |
2426 | ||
2427 | return kobj_to_node_hstate(kobj, nidp); | |
2428 | } | |
2429 | ||
2430 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, | |
2431 | struct kobj_attribute *attr, char *buf) | |
2432 | { | |
2433 | struct hstate *h; | |
2434 | unsigned long nr_huge_pages; | |
2435 | int nid; | |
2436 | ||
2437 | h = kobj_to_hstate(kobj, &nid); | |
2438 | if (nid == NUMA_NO_NODE) | |
2439 | nr_huge_pages = h->nr_huge_pages; | |
2440 | else | |
2441 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2442 | ||
2443 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
2444 | } | |
2445 | ||
2446 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, | |
2447 | struct hstate *h, int nid, | |
2448 | unsigned long count, size_t len) | |
2449 | { | |
2450 | int err; | |
2451 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); | |
2452 | ||
2453 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) { | |
2454 | err = -EINVAL; | |
2455 | goto out; | |
2456 | } | |
2457 | ||
2458 | if (nid == NUMA_NO_NODE) { | |
2459 | /* | |
2460 | * global hstate attribute | |
2461 | */ | |
2462 | if (!(obey_mempolicy && | |
2463 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
2464 | NODEMASK_FREE(nodes_allowed); | |
2465 | nodes_allowed = &node_states[N_MEMORY]; | |
2466 | } | |
2467 | } else if (nodes_allowed) { | |
2468 | /* | |
2469 | * Node specific request. count adjustment happens in | |
2470 | * set_max_huge_pages() after acquiring hugetlb_lock. | |
2471 | */ | |
2472 | init_nodemask_of_node(nodes_allowed, nid); | |
2473 | } else { | |
2474 | /* | |
2475 | * Node specific request, but we could not allocate the few | |
2476 | * words required for a node mask. We are unlikely to hit | |
2477 | * this condition. Since we can not pass down the appropriate | |
2478 | * node mask, just return ENOMEM. | |
2479 | */ | |
2480 | err = -ENOMEM; | |
2481 | goto out; | |
2482 | } | |
2483 | ||
2484 | err = set_max_huge_pages(h, count, nid, nodes_allowed); | |
2485 | ||
2486 | out: | |
2487 | if (nodes_allowed != &node_states[N_MEMORY]) | |
2488 | NODEMASK_FREE(nodes_allowed); | |
2489 | ||
2490 | return err ? err : len; | |
2491 | } | |
2492 | ||
2493 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, | |
2494 | struct kobject *kobj, const char *buf, | |
2495 | size_t len) | |
2496 | { | |
2497 | struct hstate *h; | |
2498 | unsigned long count; | |
2499 | int nid; | |
2500 | int err; | |
2501 | ||
2502 | err = kstrtoul(buf, 10, &count); | |
2503 | if (err) | |
2504 | return err; | |
2505 | ||
2506 | h = kobj_to_hstate(kobj, &nid); | |
2507 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2508 | } | |
2509 | ||
2510 | static ssize_t nr_hugepages_show(struct kobject *kobj, | |
2511 | struct kobj_attribute *attr, char *buf) | |
2512 | { | |
2513 | return nr_hugepages_show_common(kobj, attr, buf); | |
2514 | } | |
2515 | ||
2516 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2517 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2518 | { | |
2519 | return nr_hugepages_store_common(false, kobj, buf, len); | |
2520 | } | |
2521 | HSTATE_ATTR(nr_hugepages); | |
2522 | ||
2523 | #ifdef CONFIG_NUMA | |
2524 | ||
2525 | /* | |
2526 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2527 | * huge page alloc/free. | |
2528 | */ | |
2529 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
2530 | struct kobj_attribute *attr, char *buf) | |
2531 | { | |
2532 | return nr_hugepages_show_common(kobj, attr, buf); | |
2533 | } | |
2534 | ||
2535 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2536 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2537 | { | |
2538 | return nr_hugepages_store_common(true, kobj, buf, len); | |
2539 | } | |
2540 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2541 | #endif | |
2542 | ||
2543 | ||
2544 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, | |
2545 | struct kobj_attribute *attr, char *buf) | |
2546 | { | |
2547 | struct hstate *h = kobj_to_hstate(kobj, NULL); | |
2548 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); | |
2549 | } | |
2550 | ||
2551 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, | |
2552 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2553 | { | |
2554 | int err; | |
2555 | unsigned long input; | |
2556 | struct hstate *h = kobj_to_hstate(kobj, NULL); | |
2557 | ||
2558 | if (hstate_is_gigantic(h)) | |
2559 | return -EINVAL; | |
2560 | ||
2561 | err = kstrtoul(buf, 10, &input); | |
2562 | if (err) | |
2563 | return err; | |
2564 | ||
2565 | spin_lock(&hugetlb_lock); | |
2566 | h->nr_overcommit_huge_pages = input; | |
2567 | spin_unlock(&hugetlb_lock); | |
2568 | ||
2569 | return count; | |
2570 | } | |
2571 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2572 | ||
2573 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2574 | struct kobj_attribute *attr, char *buf) | |
2575 | { | |
2576 | struct hstate *h; | |
2577 | unsigned long free_huge_pages; | |
2578 | int nid; | |
2579 | ||
2580 | h = kobj_to_hstate(kobj, &nid); | |
2581 | if (nid == NUMA_NO_NODE) | |
2582 | free_huge_pages = h->free_huge_pages; | |
2583 | else | |
2584 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2585 | ||
2586 | return sprintf(buf, "%lu\n", free_huge_pages); | |
2587 | } | |
2588 | HSTATE_ATTR_RO(free_hugepages); | |
2589 | ||
2590 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2591 | struct kobj_attribute *attr, char *buf) | |
2592 | { | |
2593 | struct hstate *h = kobj_to_hstate(kobj, NULL); | |
2594 | return sprintf(buf, "%lu\n", h->resv_huge_pages); | |
2595 | } | |
2596 | HSTATE_ATTR_RO(resv_hugepages); | |
2597 | ||
2598 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2599 | struct kobj_attribute *attr, char *buf) | |
2600 | { | |
2601 | struct hstate *h; | |
2602 | unsigned long surplus_huge_pages; | |
2603 | int nid; | |
2604 | ||
2605 | h = kobj_to_hstate(kobj, &nid); | |
2606 | if (nid == NUMA_NO_NODE) | |
2607 | surplus_huge_pages = h->surplus_huge_pages; | |
2608 | else | |
2609 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2610 | ||
2611 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
2612 | } | |
2613 | HSTATE_ATTR_RO(surplus_hugepages); | |
2614 | ||
2615 | static struct attribute *hstate_attrs[] = { | |
2616 | &nr_hugepages_attr.attr, | |
2617 | &nr_overcommit_hugepages_attr.attr, | |
2618 | &free_hugepages_attr.attr, | |
2619 | &resv_hugepages_attr.attr, | |
2620 | &surplus_hugepages_attr.attr, | |
2621 | #ifdef CONFIG_NUMA | |
2622 | &nr_hugepages_mempolicy_attr.attr, | |
2623 | #endif | |
2624 | NULL, | |
2625 | }; | |
2626 | ||
2627 | static const struct attribute_group hstate_attr_group = { | |
2628 | .attrs = hstate_attrs, | |
2629 | }; | |
2630 | ||
2631 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, | |
2632 | struct kobject **hstate_kobjs, | |
2633 | const struct attribute_group *hstate_attr_group) | |
2634 | { | |
2635 | int retval; | |
2636 | int hi = hstate_index(h); | |
2637 | ||
2638 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); | |
2639 | if (!hstate_kobjs[hi]) | |
2640 | return -ENOMEM; | |
2641 | ||
2642 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); | |
2643 | if (retval) | |
2644 | kobject_put(hstate_kobjs[hi]); | |
2645 | ||
2646 | return retval; | |
2647 | } | |
2648 | ||
2649 | static void __init hugetlb_sysfs_init(void) | |
2650 | { | |
2651 | struct hstate *h; | |
2652 | int err; | |
2653 | ||
2654 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
2655 | if (!hugepages_kobj) | |
2656 | return; | |
2657 | ||
2658 | for_each_hstate(h) { | |
2659 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, | |
2660 | hstate_kobjs, &hstate_attr_group); | |
2661 | if (err) | |
2662 | pr_err("Hugetlb: Unable to add hstate %s", h->name); | |
2663 | } | |
2664 | } | |
2665 | ||
2666 | #ifdef CONFIG_NUMA | |
2667 | ||
2668 | /* | |
2669 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
2670 | * with node devices in node_devices[] using a parallel array. The array | |
2671 | * index of a node device or _hstate == node id. | |
2672 | * This is here to avoid any static dependency of the node device driver, in | |
2673 | * the base kernel, on the hugetlb module. | |
2674 | */ | |
2675 | struct node_hstate { | |
2676 | struct kobject *hugepages_kobj; | |
2677 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2678 | }; | |
2679 | static struct node_hstate node_hstates[MAX_NUMNODES]; | |
2680 | ||
2681 | /* | |
2682 | * A subset of global hstate attributes for node devices | |
2683 | */ | |
2684 | static struct attribute *per_node_hstate_attrs[] = { | |
2685 | &nr_hugepages_attr.attr, | |
2686 | &free_hugepages_attr.attr, | |
2687 | &surplus_hugepages_attr.attr, | |
2688 | NULL, | |
2689 | }; | |
2690 | ||
2691 | static const struct attribute_group per_node_hstate_attr_group = { | |
2692 | .attrs = per_node_hstate_attrs, | |
2693 | }; | |
2694 | ||
2695 | /* | |
2696 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. | |
2697 | * Returns node id via non-NULL nidp. | |
2698 | */ | |
2699 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2700 | { | |
2701 | int nid; | |
2702 | ||
2703 | for (nid = 0; nid < nr_node_ids; nid++) { | |
2704 | struct node_hstate *nhs = &node_hstates[nid]; | |
2705 | int i; | |
2706 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
2707 | if (nhs->hstate_kobjs[i] == kobj) { | |
2708 | if (nidp) | |
2709 | *nidp = nid; | |
2710 | return &hstates[i]; | |
2711 | } | |
2712 | } | |
2713 | ||
2714 | BUG(); | |
2715 | return NULL; | |
2716 | } | |
2717 | ||
2718 | /* | |
2719 | * Unregister hstate attributes from a single node device. | |
2720 | * No-op if no hstate attributes attached. | |
2721 | */ | |
2722 | static void hugetlb_unregister_node(struct node *node) | |
2723 | { | |
2724 | struct hstate *h; | |
2725 | struct node_hstate *nhs = &node_hstates[node->dev.id]; | |
2726 | ||
2727 | if (!nhs->hugepages_kobj) | |
2728 | return; /* no hstate attributes */ | |
2729 | ||
2730 | for_each_hstate(h) { | |
2731 | int idx = hstate_index(h); | |
2732 | if (nhs->hstate_kobjs[idx]) { | |
2733 | kobject_put(nhs->hstate_kobjs[idx]); | |
2734 | nhs->hstate_kobjs[idx] = NULL; | |
2735 | } | |
2736 | } | |
2737 | ||
2738 | kobject_put(nhs->hugepages_kobj); | |
2739 | nhs->hugepages_kobj = NULL; | |
2740 | } | |
2741 | ||
2742 | ||
2743 | /* | |
2744 | * Register hstate attributes for a single node device. | |
2745 | * No-op if attributes already registered. | |
2746 | */ | |
2747 | static void hugetlb_register_node(struct node *node) | |
2748 | { | |
2749 | struct hstate *h; | |
2750 | struct node_hstate *nhs = &node_hstates[node->dev.id]; | |
2751 | int err; | |
2752 | ||
2753 | if (nhs->hugepages_kobj) | |
2754 | return; /* already allocated */ | |
2755 | ||
2756 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
2757 | &node->dev.kobj); | |
2758 | if (!nhs->hugepages_kobj) | |
2759 | return; | |
2760 | ||
2761 | for_each_hstate(h) { | |
2762 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
2763 | nhs->hstate_kobjs, | |
2764 | &per_node_hstate_attr_group); | |
2765 | if (err) { | |
2766 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", | |
2767 | h->name, node->dev.id); | |
2768 | hugetlb_unregister_node(node); | |
2769 | break; | |
2770 | } | |
2771 | } | |
2772 | } | |
2773 | ||
2774 | /* | |
2775 | * hugetlb init time: register hstate attributes for all registered node | |
2776 | * devices of nodes that have memory. All on-line nodes should have | |
2777 | * registered their associated device by this time. | |
2778 | */ | |
2779 | static void __init hugetlb_register_all_nodes(void) | |
2780 | { | |
2781 | int nid; | |
2782 | ||
2783 | for_each_node_state(nid, N_MEMORY) { | |
2784 | struct node *node = node_devices[nid]; | |
2785 | if (node->dev.id == nid) | |
2786 | hugetlb_register_node(node); | |
2787 | } | |
2788 | ||
2789 | /* | |
2790 | * Let the node device driver know we're here so it can | |
2791 | * [un]register hstate attributes on node hotplug. | |
2792 | */ | |
2793 | register_hugetlbfs_with_node(hugetlb_register_node, | |
2794 | hugetlb_unregister_node); | |
2795 | } | |
2796 | #else /* !CONFIG_NUMA */ | |
2797 | ||
2798 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2799 | { | |
2800 | BUG(); | |
2801 | if (nidp) | |
2802 | *nidp = -1; | |
2803 | return NULL; | |
2804 | } | |
2805 | ||
2806 | static void hugetlb_register_all_nodes(void) { } | |
2807 | ||
2808 | #endif | |
2809 | ||
2810 | static int __init hugetlb_init(void) | |
2811 | { | |
2812 | int i; | |
2813 | ||
2814 | if (!hugepages_supported()) | |
2815 | return 0; | |
2816 | ||
2817 | if (!size_to_hstate(default_hstate_size)) { | |
2818 | if (default_hstate_size != 0) { | |
2819 | pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n", | |
2820 | default_hstate_size, HPAGE_SIZE); | |
2821 | } | |
2822 | ||
2823 | default_hstate_size = HPAGE_SIZE; | |
2824 | if (!size_to_hstate(default_hstate_size)) | |
2825 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
2826 | } | |
2827 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); | |
2828 | if (default_hstate_max_huge_pages) { | |
2829 | if (!default_hstate.max_huge_pages) | |
2830 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
2831 | } | |
2832 | ||
2833 | hugetlb_init_hstates(); | |
2834 | gather_bootmem_prealloc(); | |
2835 | report_hugepages(); | |
2836 | ||
2837 | hugetlb_sysfs_init(); | |
2838 | hugetlb_register_all_nodes(); | |
2839 | hugetlb_cgroup_file_init(); | |
2840 | ||
2841 | #ifdef CONFIG_SMP | |
2842 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
2843 | #else | |
2844 | num_fault_mutexes = 1; | |
2845 | #endif | |
2846 | hugetlb_fault_mutex_table = | |
2847 | kmalloc_array(num_fault_mutexes, sizeof(struct mutex), | |
2848 | GFP_KERNEL); | |
2849 | BUG_ON(!hugetlb_fault_mutex_table); | |
2850 | ||
2851 | for (i = 0; i < num_fault_mutexes; i++) | |
2852 | mutex_init(&hugetlb_fault_mutex_table[i]); | |
2853 | return 0; | |
2854 | } | |
2855 | subsys_initcall(hugetlb_init); | |
2856 | ||
2857 | /* Should be called on processing a hugepagesz=... option */ | |
2858 | void __init hugetlb_bad_size(void) | |
2859 | { | |
2860 | parsed_valid_hugepagesz = false; | |
2861 | } | |
2862 | ||
2863 | void __init hugetlb_add_hstate(unsigned int order) | |
2864 | { | |
2865 | struct hstate *h; | |
2866 | unsigned long i; | |
2867 | ||
2868 | if (size_to_hstate(PAGE_SIZE << order)) { | |
2869 | pr_warn("hugepagesz= specified twice, ignoring\n"); | |
2870 | return; | |
2871 | } | |
2872 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); | |
2873 | BUG_ON(order == 0); | |
2874 | h = &hstates[hugetlb_max_hstate++]; | |
2875 | h->order = order; | |
2876 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
2877 | h->nr_huge_pages = 0; | |
2878 | h->free_huge_pages = 0; | |
2879 | for (i = 0; i < MAX_NUMNODES; ++i) | |
2880 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
2881 | INIT_LIST_HEAD(&h->hugepage_activelist); | |
2882 | h->next_nid_to_alloc = first_memory_node; | |
2883 | h->next_nid_to_free = first_memory_node; | |
2884 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", | |
2885 | huge_page_size(h)/1024); | |
2886 | ||
2887 | parsed_hstate = h; | |
2888 | } | |
2889 | ||
2890 | static int __init hugetlb_nrpages_setup(char *s) | |
2891 | { | |
2892 | unsigned long *mhp; | |
2893 | static unsigned long *last_mhp; | |
2894 | ||
2895 | if (!parsed_valid_hugepagesz) { | |
2896 | pr_warn("hugepages = %s preceded by " | |
2897 | "an unsupported hugepagesz, ignoring\n", s); | |
2898 | parsed_valid_hugepagesz = true; | |
2899 | return 1; | |
2900 | } | |
2901 | /* | |
2902 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, | |
2903 | * so this hugepages= parameter goes to the "default hstate". | |
2904 | */ | |
2905 | else if (!hugetlb_max_hstate) | |
2906 | mhp = &default_hstate_max_huge_pages; | |
2907 | else | |
2908 | mhp = &parsed_hstate->max_huge_pages; | |
2909 | ||
2910 | if (mhp == last_mhp) { | |
2911 | pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); | |
2912 | return 1; | |
2913 | } | |
2914 | ||
2915 | if (sscanf(s, "%lu", mhp) <= 0) | |
2916 | *mhp = 0; | |
2917 | ||
2918 | /* | |
2919 | * Global state is always initialized later in hugetlb_init. | |
2920 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
2921 | * use the bootmem allocator. | |
2922 | */ | |
2923 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) | |
2924 | hugetlb_hstate_alloc_pages(parsed_hstate); | |
2925 | ||
2926 | last_mhp = mhp; | |
2927 | ||
2928 | return 1; | |
2929 | } | |
2930 | __setup("hugepages=", hugetlb_nrpages_setup); | |
2931 | ||
2932 | static int __init hugetlb_default_setup(char *s) | |
2933 | { | |
2934 | default_hstate_size = memparse(s, &s); | |
2935 | return 1; | |
2936 | } | |
2937 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
2938 | ||
2939 | static unsigned int cpuset_mems_nr(unsigned int *array) | |
2940 | { | |
2941 | int node; | |
2942 | unsigned int nr = 0; | |
2943 | ||
2944 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
2945 | nr += array[node]; | |
2946 | ||
2947 | return nr; | |
2948 | } | |
2949 | ||
2950 | #ifdef CONFIG_SYSCTL | |
2951 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, | |
2952 | struct ctl_table *table, int write, | |
2953 | void __user *buffer, size_t *length, loff_t *ppos) | |
2954 | { | |
2955 | struct hstate *h = &default_hstate; | |
2956 | unsigned long tmp = h->max_huge_pages; | |
2957 | int ret; | |
2958 | ||
2959 | if (!hugepages_supported()) | |
2960 | return -EOPNOTSUPP; | |
2961 | ||
2962 | table->data = &tmp; | |
2963 | table->maxlen = sizeof(unsigned long); | |
2964 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); | |
2965 | if (ret) | |
2966 | goto out; | |
2967 | ||
2968 | if (write) | |
2969 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
2970 | NUMA_NO_NODE, tmp, *length); | |
2971 | out: | |
2972 | return ret; | |
2973 | } | |
2974 | ||
2975 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, | |
2976 | void __user *buffer, size_t *length, loff_t *ppos) | |
2977 | { | |
2978 | ||
2979 | return hugetlb_sysctl_handler_common(false, table, write, | |
2980 | buffer, length, ppos); | |
2981 | } | |
2982 | ||
2983 | #ifdef CONFIG_NUMA | |
2984 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2985 | void __user *buffer, size_t *length, loff_t *ppos) | |
2986 | { | |
2987 | return hugetlb_sysctl_handler_common(true, table, write, | |
2988 | buffer, length, ppos); | |
2989 | } | |
2990 | #endif /* CONFIG_NUMA */ | |
2991 | ||
2992 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, | |
2993 | void __user *buffer, | |
2994 | size_t *length, loff_t *ppos) | |
2995 | { | |
2996 | struct hstate *h = &default_hstate; | |
2997 | unsigned long tmp; | |
2998 | int ret; | |
2999 | ||
3000 | if (!hugepages_supported()) | |
3001 | return -EOPNOTSUPP; | |
3002 | ||
3003 | tmp = h->nr_overcommit_huge_pages; | |
3004 | ||
3005 | if (write && hstate_is_gigantic(h)) | |
3006 | return -EINVAL; | |
3007 | ||
3008 | table->data = &tmp; | |
3009 | table->maxlen = sizeof(unsigned long); | |
3010 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); | |
3011 | if (ret) | |
3012 | goto out; | |
3013 | ||
3014 | if (write) { | |
3015 | spin_lock(&hugetlb_lock); | |
3016 | h->nr_overcommit_huge_pages = tmp; | |
3017 | spin_unlock(&hugetlb_lock); | |
3018 | } | |
3019 | out: | |
3020 | return ret; | |
3021 | } | |
3022 | ||
3023 | #endif /* CONFIG_SYSCTL */ | |
3024 | ||
3025 | void hugetlb_report_meminfo(struct seq_file *m) | |
3026 | { | |
3027 | struct hstate *h; | |
3028 | unsigned long total = 0; | |
3029 | ||
3030 | if (!hugepages_supported()) | |
3031 | return; | |
3032 | ||
3033 | for_each_hstate(h) { | |
3034 | unsigned long count = h->nr_huge_pages; | |
3035 | ||
3036 | total += (PAGE_SIZE << huge_page_order(h)) * count; | |
3037 | ||
3038 | if (h == &default_hstate) | |
3039 | seq_printf(m, | |
3040 | "HugePages_Total: %5lu\n" | |
3041 | "HugePages_Free: %5lu\n" | |
3042 | "HugePages_Rsvd: %5lu\n" | |
3043 | "HugePages_Surp: %5lu\n" | |
3044 | "Hugepagesize: %8lu kB\n", | |
3045 | count, | |
3046 | h->free_huge_pages, | |
3047 | h->resv_huge_pages, | |
3048 | h->surplus_huge_pages, | |
3049 | (PAGE_SIZE << huge_page_order(h)) / 1024); | |
3050 | } | |
3051 | ||
3052 | seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); | |
3053 | } | |
3054 | ||
3055 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
3056 | { | |
3057 | struct hstate *h = &default_hstate; | |
3058 | if (!hugepages_supported()) | |
3059 | return 0; | |
3060 | return sprintf(buf, | |
3061 | "Node %d HugePages_Total: %5u\n" | |
3062 | "Node %d HugePages_Free: %5u\n" | |
3063 | "Node %d HugePages_Surp: %5u\n", | |
3064 | nid, h->nr_huge_pages_node[nid], | |
3065 | nid, h->free_huge_pages_node[nid], | |
3066 | nid, h->surplus_huge_pages_node[nid]); | |
3067 | } | |
3068 | ||
3069 | void hugetlb_show_meminfo(void) | |
3070 | { | |
3071 | struct hstate *h; | |
3072 | int nid; | |
3073 | ||
3074 | if (!hugepages_supported()) | |
3075 | return; | |
3076 | ||
3077 | for_each_node_state(nid, N_MEMORY) | |
3078 | for_each_hstate(h) | |
3079 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
3080 | nid, | |
3081 | h->nr_huge_pages_node[nid], | |
3082 | h->free_huge_pages_node[nid], | |
3083 | h->surplus_huge_pages_node[nid], | |
3084 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
3085 | } | |
3086 | ||
3087 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) | |
3088 | { | |
3089 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
3090 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
3091 | } | |
3092 | ||
3093 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ | |
3094 | unsigned long hugetlb_total_pages(void) | |
3095 | { | |
3096 | struct hstate *h; | |
3097 | unsigned long nr_total_pages = 0; | |
3098 | ||
3099 | for_each_hstate(h) | |
3100 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
3101 | return nr_total_pages; | |
3102 | } | |
3103 | ||
3104 | static int hugetlb_acct_memory(struct hstate *h, long delta) | |
3105 | { | |
3106 | int ret = -ENOMEM; | |
3107 | ||
3108 | spin_lock(&hugetlb_lock); | |
3109 | /* | |
3110 | * When cpuset is configured, it breaks the strict hugetlb page | |
3111 | * reservation as the accounting is done on a global variable. Such | |
3112 | * reservation is completely rubbish in the presence of cpuset because | |
3113 | * the reservation is not checked against page availability for the | |
3114 | * current cpuset. Application can still potentially OOM'ed by kernel | |
3115 | * with lack of free htlb page in cpuset that the task is in. | |
3116 | * Attempt to enforce strict accounting with cpuset is almost | |
3117 | * impossible (or too ugly) because cpuset is too fluid that | |
3118 | * task or memory node can be dynamically moved between cpusets. | |
3119 | * | |
3120 | * The change of semantics for shared hugetlb mapping with cpuset is | |
3121 | * undesirable. However, in order to preserve some of the semantics, | |
3122 | * we fall back to check against current free page availability as | |
3123 | * a best attempt and hopefully to minimize the impact of changing | |
3124 | * semantics that cpuset has. | |
3125 | */ | |
3126 | if (delta > 0) { | |
3127 | if (gather_surplus_pages(h, delta) < 0) | |
3128 | goto out; | |
3129 | ||
3130 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { | |
3131 | return_unused_surplus_pages(h, delta); | |
3132 | goto out; | |
3133 | } | |
3134 | } | |
3135 | ||
3136 | ret = 0; | |
3137 | if (delta < 0) | |
3138 | return_unused_surplus_pages(h, (unsigned long) -delta); | |
3139 | ||
3140 | out: | |
3141 | spin_unlock(&hugetlb_lock); | |
3142 | return ret; | |
3143 | } | |
3144 | ||
3145 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) | |
3146 | { | |
3147 | struct resv_map *resv = vma_resv_map(vma); | |
3148 | ||
3149 | /* | |
3150 | * This new VMA should share its siblings reservation map if present. | |
3151 | * The VMA will only ever have a valid reservation map pointer where | |
3152 | * it is being copied for another still existing VMA. As that VMA | |
3153 | * has a reference to the reservation map it cannot disappear until | |
3154 | * after this open call completes. It is therefore safe to take a | |
3155 | * new reference here without additional locking. | |
3156 | */ | |
3157 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) | |
3158 | kref_get(&resv->refs); | |
3159 | } | |
3160 | ||
3161 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) | |
3162 | { | |
3163 | struct hstate *h = hstate_vma(vma); | |
3164 | struct resv_map *resv = vma_resv_map(vma); | |
3165 | struct hugepage_subpool *spool = subpool_vma(vma); | |
3166 | unsigned long reserve, start, end; | |
3167 | long gbl_reserve; | |
3168 | ||
3169 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) | |
3170 | return; | |
3171 | ||
3172 | start = vma_hugecache_offset(h, vma, vma->vm_start); | |
3173 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
3174 | ||
3175 | reserve = (end - start) - region_count(resv, start, end); | |
3176 | ||
3177 | kref_put(&resv->refs, resv_map_release); | |
3178 | ||
3179 | if (reserve) { | |
3180 | /* | |
3181 | * Decrement reserve counts. The global reserve count may be | |
3182 | * adjusted if the subpool has a minimum size. | |
3183 | */ | |
3184 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
3185 | hugetlb_acct_memory(h, -gbl_reserve); | |
3186 | } | |
3187 | } | |
3188 | ||
3189 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) | |
3190 | { | |
3191 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
3192 | return -EINVAL; | |
3193 | return 0; | |
3194 | } | |
3195 | ||
3196 | static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) | |
3197 | { | |
3198 | struct hstate *hstate = hstate_vma(vma); | |
3199 | ||
3200 | return 1UL << huge_page_shift(hstate); | |
3201 | } | |
3202 | ||
3203 | /* | |
3204 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
3205 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
3206 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
3207 | * this far. | |
3208 | */ | |
3209 | static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) | |
3210 | { | |
3211 | BUG(); | |
3212 | return 0; | |
3213 | } | |
3214 | ||
3215 | /* | |
3216 | * When a new function is introduced to vm_operations_struct and added | |
3217 | * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. | |
3218 | * This is because under System V memory model, mappings created via | |
3219 | * shmget/shmat with "huge page" specified are backed by hugetlbfs files, | |
3220 | * their original vm_ops are overwritten with shm_vm_ops. | |
3221 | */ | |
3222 | const struct vm_operations_struct hugetlb_vm_ops = { | |
3223 | .fault = hugetlb_vm_op_fault, | |
3224 | .open = hugetlb_vm_op_open, | |
3225 | .close = hugetlb_vm_op_close, | |
3226 | .split = hugetlb_vm_op_split, | |
3227 | .pagesize = hugetlb_vm_op_pagesize, | |
3228 | }; | |
3229 | ||
3230 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, | |
3231 | int writable) | |
3232 | { | |
3233 | pte_t entry; | |
3234 | ||
3235 | if (writable) { | |
3236 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, | |
3237 | vma->vm_page_prot))); | |
3238 | } else { | |
3239 | entry = huge_pte_wrprotect(mk_huge_pte(page, | |
3240 | vma->vm_page_prot)); | |
3241 | } | |
3242 | entry = pte_mkyoung(entry); | |
3243 | entry = pte_mkhuge(entry); | |
3244 | entry = arch_make_huge_pte(entry, vma, page, writable); | |
3245 | ||
3246 | return entry; | |
3247 | } | |
3248 | ||
3249 | static void set_huge_ptep_writable(struct vm_area_struct *vma, | |
3250 | unsigned long address, pte_t *ptep) | |
3251 | { | |
3252 | pte_t entry; | |
3253 | ||
3254 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); | |
3255 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) | |
3256 | update_mmu_cache(vma, address, ptep); | |
3257 | } | |
3258 | ||
3259 | bool is_hugetlb_entry_migration(pte_t pte) | |
3260 | { | |
3261 | swp_entry_t swp; | |
3262 | ||
3263 | if (huge_pte_none(pte) || pte_present(pte)) | |
3264 | return false; | |
3265 | swp = pte_to_swp_entry(pte); | |
3266 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
3267 | return true; | |
3268 | else | |
3269 | return false; | |
3270 | } | |
3271 | ||
3272 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
3273 | { | |
3274 | swp_entry_t swp; | |
3275 | ||
3276 | if (huge_pte_none(pte) || pte_present(pte)) | |
3277 | return 0; | |
3278 | swp = pte_to_swp_entry(pte); | |
3279 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
3280 | return 1; | |
3281 | else | |
3282 | return 0; | |
3283 | } | |
3284 | ||
3285 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, | |
3286 | struct vm_area_struct *vma) | |
3287 | { | |
3288 | pte_t *src_pte, *dst_pte, entry, dst_entry; | |
3289 | struct page *ptepage; | |
3290 | unsigned long addr; | |
3291 | int cow; | |
3292 | struct hstate *h = hstate_vma(vma); | |
3293 | unsigned long sz = huge_page_size(h); | |
3294 | struct mmu_notifier_range range; | |
3295 | int ret = 0; | |
3296 | ||
3297 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
3298 | ||
3299 | if (cow) { | |
3300 | mmu_notifier_range_init(&range, src, vma->vm_start, | |
3301 | vma->vm_end); | |
3302 | mmu_notifier_invalidate_range_start(&range); | |
3303 | } | |
3304 | ||
3305 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { | |
3306 | spinlock_t *src_ptl, *dst_ptl; | |
3307 | src_pte = huge_pte_offset(src, addr, sz); | |
3308 | if (!src_pte) | |
3309 | continue; | |
3310 | dst_pte = huge_pte_alloc(dst, addr, sz); | |
3311 | if (!dst_pte) { | |
3312 | ret = -ENOMEM; | |
3313 | break; | |
3314 | } | |
3315 | ||
3316 | /* | |
3317 | * If the pagetables are shared don't copy or take references. | |
3318 | * dst_pte == src_pte is the common case of src/dest sharing. | |
3319 | * | |
3320 | * However, src could have 'unshared' and dst shares with | |
3321 | * another vma. If dst_pte !none, this implies sharing. | |
3322 | * Check here before taking page table lock, and once again | |
3323 | * after taking the lock below. | |
3324 | */ | |
3325 | dst_entry = huge_ptep_get(dst_pte); | |
3326 | if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) | |
3327 | continue; | |
3328 | ||
3329 | dst_ptl = huge_pte_lock(h, dst, dst_pte); | |
3330 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3331 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
3332 | entry = huge_ptep_get(src_pte); | |
3333 | dst_entry = huge_ptep_get(dst_pte); | |
3334 | if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { | |
3335 | /* | |
3336 | * Skip if src entry none. Also, skip in the | |
3337 | * unlikely case dst entry !none as this implies | |
3338 | * sharing with another vma. | |
3339 | */ | |
3340 | ; | |
3341 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
3342 | is_hugetlb_entry_hwpoisoned(entry))) { | |
3343 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
3344 | ||
3345 | if (is_write_migration_entry(swp_entry) && cow) { | |
3346 | /* | |
3347 | * COW mappings require pages in both | |
3348 | * parent and child to be set to read. | |
3349 | */ | |
3350 | make_migration_entry_read(&swp_entry); | |
3351 | entry = swp_entry_to_pte(swp_entry); | |
3352 | set_huge_swap_pte_at(src, addr, src_pte, | |
3353 | entry, sz); | |
3354 | } | |
3355 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); | |
3356 | } else { | |
3357 | if (cow) { | |
3358 | /* | |
3359 | * No need to notify as we are downgrading page | |
3360 | * table protection not changing it to point | |
3361 | * to a new page. | |
3362 | * | |
3363 | * See Documentation/vm/mmu_notifier.rst | |
3364 | */ | |
3365 | huge_ptep_set_wrprotect(src, addr, src_pte); | |
3366 | } | |
3367 | entry = huge_ptep_get(src_pte); | |
3368 | ptepage = pte_page(entry); | |
3369 | get_page(ptepage); | |
3370 | page_dup_rmap(ptepage, true); | |
3371 | set_huge_pte_at(dst, addr, dst_pte, entry); | |
3372 | hugetlb_count_add(pages_per_huge_page(h), dst); | |
3373 | } | |
3374 | spin_unlock(src_ptl); | |
3375 | spin_unlock(dst_ptl); | |
3376 | } | |
3377 | ||
3378 | if (cow) | |
3379 | mmu_notifier_invalidate_range_end(&range); | |
3380 | ||
3381 | return ret; | |
3382 | } | |
3383 | ||
3384 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, | |
3385 | unsigned long start, unsigned long end, | |
3386 | struct page *ref_page) | |
3387 | { | |
3388 | struct mm_struct *mm = vma->vm_mm; | |
3389 | unsigned long address; | |
3390 | pte_t *ptep; | |
3391 | pte_t pte; | |
3392 | spinlock_t *ptl; | |
3393 | struct page *page; | |
3394 | struct hstate *h = hstate_vma(vma); | |
3395 | unsigned long sz = huge_page_size(h); | |
3396 | struct mmu_notifier_range range; | |
3397 | ||
3398 | WARN_ON(!is_vm_hugetlb_page(vma)); | |
3399 | BUG_ON(start & ~huge_page_mask(h)); | |
3400 | BUG_ON(end & ~huge_page_mask(h)); | |
3401 | ||
3402 | /* | |
3403 | * This is a hugetlb vma, all the pte entries should point | |
3404 | * to huge page. | |
3405 | */ | |
3406 | tlb_change_page_size(tlb, sz); | |
3407 | tlb_start_vma(tlb, vma); | |
3408 | ||
3409 | /* | |
3410 | * If sharing possible, alert mmu notifiers of worst case. | |
3411 | */ | |
3412 | mmu_notifier_range_init(&range, mm, start, end); | |
3413 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); | |
3414 | mmu_notifier_invalidate_range_start(&range); | |
3415 | address = start; | |
3416 | for (; address < end; address += sz) { | |
3417 | ptep = huge_pte_offset(mm, address, sz); | |
3418 | if (!ptep) | |
3419 | continue; | |
3420 | ||
3421 | ptl = huge_pte_lock(h, mm, ptep); | |
3422 | if (huge_pmd_unshare(mm, &address, ptep)) { | |
3423 | spin_unlock(ptl); | |
3424 | /* | |
3425 | * We just unmapped a page of PMDs by clearing a PUD. | |
3426 | * The caller's TLB flush range should cover this area. | |
3427 | */ | |
3428 | continue; | |
3429 | } | |
3430 | ||
3431 | pte = huge_ptep_get(ptep); | |
3432 | if (huge_pte_none(pte)) { | |
3433 | spin_unlock(ptl); | |
3434 | continue; | |
3435 | } | |
3436 | ||
3437 | /* | |
3438 | * Migrating hugepage or HWPoisoned hugepage is already | |
3439 | * unmapped and its refcount is dropped, so just clear pte here. | |
3440 | */ | |
3441 | if (unlikely(!pte_present(pte))) { | |
3442 | huge_pte_clear(mm, address, ptep, sz); | |
3443 | spin_unlock(ptl); | |
3444 | continue; | |
3445 | } | |
3446 | ||
3447 | page = pte_page(pte); | |
3448 | /* | |
3449 | * If a reference page is supplied, it is because a specific | |
3450 | * page is being unmapped, not a range. Ensure the page we | |
3451 | * are about to unmap is the actual page of interest. | |
3452 | */ | |
3453 | if (ref_page) { | |
3454 | if (page != ref_page) { | |
3455 | spin_unlock(ptl); | |
3456 | continue; | |
3457 | } | |
3458 | /* | |
3459 | * Mark the VMA as having unmapped its page so that | |
3460 | * future faults in this VMA will fail rather than | |
3461 | * looking like data was lost | |
3462 | */ | |
3463 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
3464 | } | |
3465 | ||
3466 | pte = huge_ptep_get_and_clear(mm, address, ptep); | |
3467 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); | |
3468 | if (huge_pte_dirty(pte)) | |
3469 | set_page_dirty(page); | |
3470 | ||
3471 | hugetlb_count_sub(pages_per_huge_page(h), mm); | |
3472 | page_remove_rmap(page, true); | |
3473 | ||
3474 | spin_unlock(ptl); | |
3475 | tlb_remove_page_size(tlb, page, huge_page_size(h)); | |
3476 | /* | |
3477 | * Bail out after unmapping reference page if supplied | |
3478 | */ | |
3479 | if (ref_page) | |
3480 | break; | |
3481 | } | |
3482 | mmu_notifier_invalidate_range_end(&range); | |
3483 | tlb_end_vma(tlb, vma); | |
3484 | } | |
3485 | ||
3486 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, | |
3487 | struct vm_area_struct *vma, unsigned long start, | |
3488 | unsigned long end, struct page *ref_page) | |
3489 | { | |
3490 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
3491 | ||
3492 | /* | |
3493 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
3494 | * test will fail on a vma being torn down, and not grab a page table | |
3495 | * on its way out. We're lucky that the flag has such an appropriate | |
3496 | * name, and can in fact be safely cleared here. We could clear it | |
3497 | * before the __unmap_hugepage_range above, but all that's necessary | |
3498 | * is to clear it before releasing the i_mmap_rwsem. This works | |
3499 | * because in the context this is called, the VMA is about to be | |
3500 | * destroyed and the i_mmap_rwsem is held. | |
3501 | */ | |
3502 | vma->vm_flags &= ~VM_MAYSHARE; | |
3503 | } | |
3504 | ||
3505 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, | |
3506 | unsigned long end, struct page *ref_page) | |
3507 | { | |
3508 | struct mm_struct *mm; | |
3509 | struct mmu_gather tlb; | |
3510 | unsigned long tlb_start = start; | |
3511 | unsigned long tlb_end = end; | |
3512 | ||
3513 | /* | |
3514 | * If shared PMDs were possibly used within this vma range, adjust | |
3515 | * start/end for worst case tlb flushing. | |
3516 | * Note that we can not be sure if PMDs are shared until we try to | |
3517 | * unmap pages. However, we want to make sure TLB flushing covers | |
3518 | * the largest possible range. | |
3519 | */ | |
3520 | adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end); | |
3521 | ||
3522 | mm = vma->vm_mm; | |
3523 | ||
3524 | tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end); | |
3525 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); | |
3526 | tlb_finish_mmu(&tlb, tlb_start, tlb_end); | |
3527 | } | |
3528 | ||
3529 | /* | |
3530 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
3531 | * mappping it owns the reserve page for. The intention is to unmap the page | |
3532 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
3533 | * same region. | |
3534 | */ | |
3535 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, | |
3536 | struct page *page, unsigned long address) | |
3537 | { | |
3538 | struct hstate *h = hstate_vma(vma); | |
3539 | struct vm_area_struct *iter_vma; | |
3540 | struct address_space *mapping; | |
3541 | pgoff_t pgoff; | |
3542 | ||
3543 | /* | |
3544 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
3545 | * from page cache lookup which is in HPAGE_SIZE units. | |
3546 | */ | |
3547 | address = address & huge_page_mask(h); | |
3548 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + | |
3549 | vma->vm_pgoff; | |
3550 | mapping = vma->vm_file->f_mapping; | |
3551 | ||
3552 | /* | |
3553 | * Take the mapping lock for the duration of the table walk. As | |
3554 | * this mapping should be shared between all the VMAs, | |
3555 | * __unmap_hugepage_range() is called as the lock is already held | |
3556 | */ | |
3557 | i_mmap_lock_write(mapping); | |
3558 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { | |
3559 | /* Do not unmap the current VMA */ | |
3560 | if (iter_vma == vma) | |
3561 | continue; | |
3562 | ||
3563 | /* | |
3564 | * Shared VMAs have their own reserves and do not affect | |
3565 | * MAP_PRIVATE accounting but it is possible that a shared | |
3566 | * VMA is using the same page so check and skip such VMAs. | |
3567 | */ | |
3568 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
3569 | continue; | |
3570 | ||
3571 | /* | |
3572 | * Unmap the page from other VMAs without their own reserves. | |
3573 | * They get marked to be SIGKILLed if they fault in these | |
3574 | * areas. This is because a future no-page fault on this VMA | |
3575 | * could insert a zeroed page instead of the data existing | |
3576 | * from the time of fork. This would look like data corruption | |
3577 | */ | |
3578 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
3579 | unmap_hugepage_range(iter_vma, address, | |
3580 | address + huge_page_size(h), page); | |
3581 | } | |
3582 | i_mmap_unlock_write(mapping); | |
3583 | } | |
3584 | ||
3585 | /* | |
3586 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
3587 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we | |
3588 | * cannot race with other handlers or page migration. | |
3589 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
3590 | */ | |
3591 | static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, | |
3592 | unsigned long address, pte_t *ptep, | |
3593 | struct page *pagecache_page, spinlock_t *ptl) | |
3594 | { | |
3595 | pte_t pte; | |
3596 | struct hstate *h = hstate_vma(vma); | |
3597 | struct page *old_page, *new_page; | |
3598 | int outside_reserve = 0; | |
3599 | vm_fault_t ret = 0; | |
3600 | unsigned long haddr = address & huge_page_mask(h); | |
3601 | struct mmu_notifier_range range; | |
3602 | ||
3603 | pte = huge_ptep_get(ptep); | |
3604 | old_page = pte_page(pte); | |
3605 | ||
3606 | retry_avoidcopy: | |
3607 | /* If no-one else is actually using this page, avoid the copy | |
3608 | * and just make the page writable */ | |
3609 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { | |
3610 | page_move_anon_rmap(old_page, vma); | |
3611 | set_huge_ptep_writable(vma, haddr, ptep); | |
3612 | return 0; | |
3613 | } | |
3614 | ||
3615 | /* | |
3616 | * If the process that created a MAP_PRIVATE mapping is about to | |
3617 | * perform a COW due to a shared page count, attempt to satisfy | |
3618 | * the allocation without using the existing reserves. The pagecache | |
3619 | * page is used to determine if the reserve at this address was | |
3620 | * consumed or not. If reserves were used, a partial faulted mapping | |
3621 | * at the time of fork() could consume its reserves on COW instead | |
3622 | * of the full address range. | |
3623 | */ | |
3624 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && | |
3625 | old_page != pagecache_page) | |
3626 | outside_reserve = 1; | |
3627 | ||
3628 | get_page(old_page); | |
3629 | ||
3630 | /* | |
3631 | * Drop page table lock as buddy allocator may be called. It will | |
3632 | * be acquired again before returning to the caller, as expected. | |
3633 | */ | |
3634 | spin_unlock(ptl); | |
3635 | new_page = alloc_huge_page(vma, haddr, outside_reserve); | |
3636 | ||
3637 | if (IS_ERR(new_page)) { | |
3638 | /* | |
3639 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
3640 | * it is due to references held by a child and an insufficient | |
3641 | * huge page pool. To guarantee the original mappers | |
3642 | * reliability, unmap the page from child processes. The child | |
3643 | * may get SIGKILLed if it later faults. | |
3644 | */ | |
3645 | if (outside_reserve) { | |
3646 | put_page(old_page); | |
3647 | BUG_ON(huge_pte_none(pte)); | |
3648 | unmap_ref_private(mm, vma, old_page, haddr); | |
3649 | BUG_ON(huge_pte_none(pte)); | |
3650 | spin_lock(ptl); | |
3651 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); | |
3652 | if (likely(ptep && | |
3653 | pte_same(huge_ptep_get(ptep), pte))) | |
3654 | goto retry_avoidcopy; | |
3655 | /* | |
3656 | * race occurs while re-acquiring page table | |
3657 | * lock, and our job is done. | |
3658 | */ | |
3659 | return 0; | |
3660 | } | |
3661 | ||
3662 | ret = vmf_error(PTR_ERR(new_page)); | |
3663 | goto out_release_old; | |
3664 | } | |
3665 | ||
3666 | /* | |
3667 | * When the original hugepage is shared one, it does not have | |
3668 | * anon_vma prepared. | |
3669 | */ | |
3670 | if (unlikely(anon_vma_prepare(vma))) { | |
3671 | ret = VM_FAULT_OOM; | |
3672 | goto out_release_all; | |
3673 | } | |
3674 | ||
3675 | copy_user_huge_page(new_page, old_page, address, vma, | |
3676 | pages_per_huge_page(h)); | |
3677 | __SetPageUptodate(new_page); | |
3678 | ||
3679 | mmu_notifier_range_init(&range, mm, haddr, haddr + huge_page_size(h)); | |
3680 | mmu_notifier_invalidate_range_start(&range); | |
3681 | ||
3682 | /* | |
3683 | * Retake the page table lock to check for racing updates | |
3684 | * before the page tables are altered | |
3685 | */ | |
3686 | spin_lock(ptl); | |
3687 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); | |
3688 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { | |
3689 | ClearPagePrivate(new_page); | |
3690 | ||
3691 | /* Break COW */ | |
3692 | huge_ptep_clear_flush(vma, haddr, ptep); | |
3693 | mmu_notifier_invalidate_range(mm, range.start, range.end); | |
3694 | set_huge_pte_at(mm, haddr, ptep, | |
3695 | make_huge_pte(vma, new_page, 1)); | |
3696 | page_remove_rmap(old_page, true); | |
3697 | hugepage_add_new_anon_rmap(new_page, vma, haddr); | |
3698 | set_page_huge_active(new_page); | |
3699 | /* Make the old page be freed below */ | |
3700 | new_page = old_page; | |
3701 | } | |
3702 | spin_unlock(ptl); | |
3703 | mmu_notifier_invalidate_range_end(&range); | |
3704 | out_release_all: | |
3705 | restore_reserve_on_error(h, vma, haddr, new_page); | |
3706 | put_page(new_page); | |
3707 | out_release_old: | |
3708 | put_page(old_page); | |
3709 | ||
3710 | spin_lock(ptl); /* Caller expects lock to be held */ | |
3711 | return ret; | |
3712 | } | |
3713 | ||
3714 | /* Return the pagecache page at a given address within a VMA */ | |
3715 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, | |
3716 | struct vm_area_struct *vma, unsigned long address) | |
3717 | { | |
3718 | struct address_space *mapping; | |
3719 | pgoff_t idx; | |
3720 | ||
3721 | mapping = vma->vm_file->f_mapping; | |
3722 | idx = vma_hugecache_offset(h, vma, address); | |
3723 | ||
3724 | return find_lock_page(mapping, idx); | |
3725 | } | |
3726 | ||
3727 | /* | |
3728 | * Return whether there is a pagecache page to back given address within VMA. | |
3729 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
3730 | */ | |
3731 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
3732 | struct vm_area_struct *vma, unsigned long address) | |
3733 | { | |
3734 | struct address_space *mapping; | |
3735 | pgoff_t idx; | |
3736 | struct page *page; | |
3737 | ||
3738 | mapping = vma->vm_file->f_mapping; | |
3739 | idx = vma_hugecache_offset(h, vma, address); | |
3740 | ||
3741 | page = find_get_page(mapping, idx); | |
3742 | if (page) | |
3743 | put_page(page); | |
3744 | return page != NULL; | |
3745 | } | |
3746 | ||
3747 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, | |
3748 | pgoff_t idx) | |
3749 | { | |
3750 | struct inode *inode = mapping->host; | |
3751 | struct hstate *h = hstate_inode(inode); | |
3752 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
3753 | ||
3754 | if (err) | |
3755 | return err; | |
3756 | ClearPagePrivate(page); | |
3757 | ||
3758 | /* | |
3759 | * set page dirty so that it will not be removed from cache/file | |
3760 | * by non-hugetlbfs specific code paths. | |
3761 | */ | |
3762 | set_page_dirty(page); | |
3763 | ||
3764 | spin_lock(&inode->i_lock); | |
3765 | inode->i_blocks += blocks_per_huge_page(h); | |
3766 | spin_unlock(&inode->i_lock); | |
3767 | return 0; | |
3768 | } | |
3769 | ||
3770 | static vm_fault_t hugetlb_no_page(struct mm_struct *mm, | |
3771 | struct vm_area_struct *vma, | |
3772 | struct address_space *mapping, pgoff_t idx, | |
3773 | unsigned long address, pte_t *ptep, unsigned int flags) | |
3774 | { | |
3775 | struct hstate *h = hstate_vma(vma); | |
3776 | vm_fault_t ret = VM_FAULT_SIGBUS; | |
3777 | int anon_rmap = 0; | |
3778 | unsigned long size; | |
3779 | struct page *page; | |
3780 | pte_t new_pte; | |
3781 | spinlock_t *ptl; | |
3782 | unsigned long haddr = address & huge_page_mask(h); | |
3783 | bool new_page = false; | |
3784 | ||
3785 | /* | |
3786 | * Currently, we are forced to kill the process in the event the | |
3787 | * original mapper has unmapped pages from the child due to a failed | |
3788 | * COW. Warn that such a situation has occurred as it may not be obvious | |
3789 | */ | |
3790 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
3791 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", | |
3792 | current->pid); | |
3793 | return ret; | |
3794 | } | |
3795 | ||
3796 | /* | |
3797 | * Use page lock to guard against racing truncation | |
3798 | * before we get page_table_lock. | |
3799 | */ | |
3800 | retry: | |
3801 | page = find_lock_page(mapping, idx); | |
3802 | if (!page) { | |
3803 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
3804 | if (idx >= size) | |
3805 | goto out; | |
3806 | ||
3807 | /* | |
3808 | * Check for page in userfault range | |
3809 | */ | |
3810 | if (userfaultfd_missing(vma)) { | |
3811 | u32 hash; | |
3812 | struct vm_fault vmf = { | |
3813 | .vma = vma, | |
3814 | .address = haddr, | |
3815 | .flags = flags, | |
3816 | /* | |
3817 | * Hard to debug if it ends up being | |
3818 | * used by a callee that assumes | |
3819 | * something about the other | |
3820 | * uninitialized fields... same as in | |
3821 | * memory.c | |
3822 | */ | |
3823 | }; | |
3824 | ||
3825 | /* | |
3826 | * hugetlb_fault_mutex must be dropped before | |
3827 | * handling userfault. Reacquire after handling | |
3828 | * fault to make calling code simpler. | |
3829 | */ | |
3830 | hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, | |
3831 | idx, haddr); | |
3832 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
3833 | ret = handle_userfault(&vmf, VM_UFFD_MISSING); | |
3834 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
3835 | goto out; | |
3836 | } | |
3837 | ||
3838 | page = alloc_huge_page(vma, haddr, 0); | |
3839 | if (IS_ERR(page)) { | |
3840 | ret = vmf_error(PTR_ERR(page)); | |
3841 | goto out; | |
3842 | } | |
3843 | clear_huge_page(page, address, pages_per_huge_page(h)); | |
3844 | __SetPageUptodate(page); | |
3845 | new_page = true; | |
3846 | ||
3847 | if (vma->vm_flags & VM_MAYSHARE) { | |
3848 | int err = huge_add_to_page_cache(page, mapping, idx); | |
3849 | if (err) { | |
3850 | put_page(page); | |
3851 | if (err == -EEXIST) | |
3852 | goto retry; | |
3853 | goto out; | |
3854 | } | |
3855 | } else { | |
3856 | lock_page(page); | |
3857 | if (unlikely(anon_vma_prepare(vma))) { | |
3858 | ret = VM_FAULT_OOM; | |
3859 | goto backout_unlocked; | |
3860 | } | |
3861 | anon_rmap = 1; | |
3862 | } | |
3863 | } else { | |
3864 | /* | |
3865 | * If memory error occurs between mmap() and fault, some process | |
3866 | * don't have hwpoisoned swap entry for errored virtual address. | |
3867 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
3868 | */ | |
3869 | if (unlikely(PageHWPoison(page))) { | |
3870 | ret = VM_FAULT_HWPOISON | | |
3871 | VM_FAULT_SET_HINDEX(hstate_index(h)); | |
3872 | goto backout_unlocked; | |
3873 | } | |
3874 | } | |
3875 | ||
3876 | /* | |
3877 | * If we are going to COW a private mapping later, we examine the | |
3878 | * pending reservations for this page now. This will ensure that | |
3879 | * any allocations necessary to record that reservation occur outside | |
3880 | * the spinlock. | |
3881 | */ | |
3882 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { | |
3883 | if (vma_needs_reservation(h, vma, haddr) < 0) { | |
3884 | ret = VM_FAULT_OOM; | |
3885 | goto backout_unlocked; | |
3886 | } | |
3887 | /* Just decrements count, does not deallocate */ | |
3888 | vma_end_reservation(h, vma, haddr); | |
3889 | } | |
3890 | ||
3891 | ptl = huge_pte_lock(h, mm, ptep); | |
3892 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
3893 | if (idx >= size) | |
3894 | goto backout; | |
3895 | ||
3896 | ret = 0; | |
3897 | if (!huge_pte_none(huge_ptep_get(ptep))) | |
3898 | goto backout; | |
3899 | ||
3900 | if (anon_rmap) { | |
3901 | ClearPagePrivate(page); | |
3902 | hugepage_add_new_anon_rmap(page, vma, haddr); | |
3903 | } else | |
3904 | page_dup_rmap(page, true); | |
3905 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) | |
3906 | && (vma->vm_flags & VM_SHARED))); | |
3907 | set_huge_pte_at(mm, haddr, ptep, new_pte); | |
3908 | ||
3909 | hugetlb_count_add(pages_per_huge_page(h), mm); | |
3910 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { | |
3911 | /* Optimization, do the COW without a second fault */ | |
3912 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); | |
3913 | } | |
3914 | ||
3915 | spin_unlock(ptl); | |
3916 | ||
3917 | /* | |
3918 | * Only make newly allocated pages active. Existing pages found | |
3919 | * in the pagecache could be !page_huge_active() if they have been | |
3920 | * isolated for migration. | |
3921 | */ | |
3922 | if (new_page) | |
3923 | set_page_huge_active(page); | |
3924 | ||
3925 | unlock_page(page); | |
3926 | out: | |
3927 | return ret; | |
3928 | ||
3929 | backout: | |
3930 | spin_unlock(ptl); | |
3931 | backout_unlocked: | |
3932 | unlock_page(page); | |
3933 | restore_reserve_on_error(h, vma, haddr, page); | |
3934 | put_page(page); | |
3935 | goto out; | |
3936 | } | |
3937 | ||
3938 | #ifdef CONFIG_SMP | |
3939 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, | |
3940 | struct vm_area_struct *vma, | |
3941 | struct address_space *mapping, | |
3942 | pgoff_t idx, unsigned long address) | |
3943 | { | |
3944 | unsigned long key[2]; | |
3945 | u32 hash; | |
3946 | ||
3947 | if (vma->vm_flags & VM_SHARED) { | |
3948 | key[0] = (unsigned long) mapping; | |
3949 | key[1] = idx; | |
3950 | } else { | |
3951 | key[0] = (unsigned long) mm; | |
3952 | key[1] = address >> huge_page_shift(h); | |
3953 | } | |
3954 | ||
3955 | hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); | |
3956 | ||
3957 | return hash & (num_fault_mutexes - 1); | |
3958 | } | |
3959 | #else | |
3960 | /* | |
3961 | * For uniprocesor systems we always use a single mutex, so just | |
3962 | * return 0 and avoid the hashing overhead. | |
3963 | */ | |
3964 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, | |
3965 | struct vm_area_struct *vma, | |
3966 | struct address_space *mapping, | |
3967 | pgoff_t idx, unsigned long address) | |
3968 | { | |
3969 | return 0; | |
3970 | } | |
3971 | #endif | |
3972 | ||
3973 | vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, | |
3974 | unsigned long address, unsigned int flags) | |
3975 | { | |
3976 | pte_t *ptep, entry; | |
3977 | spinlock_t *ptl; | |
3978 | vm_fault_t ret; | |
3979 | u32 hash; | |
3980 | pgoff_t idx; | |
3981 | struct page *page = NULL; | |
3982 | struct page *pagecache_page = NULL; | |
3983 | struct hstate *h = hstate_vma(vma); | |
3984 | struct address_space *mapping; | |
3985 | int need_wait_lock = 0; | |
3986 | unsigned long haddr = address & huge_page_mask(h); | |
3987 | ||
3988 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); | |
3989 | if (ptep) { | |
3990 | entry = huge_ptep_get(ptep); | |
3991 | if (unlikely(is_hugetlb_entry_migration(entry))) { | |
3992 | migration_entry_wait_huge(vma, mm, ptep); | |
3993 | return 0; | |
3994 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
3995 | return VM_FAULT_HWPOISON_LARGE | | |
3996 | VM_FAULT_SET_HINDEX(hstate_index(h)); | |
3997 | } else { | |
3998 | ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); | |
3999 | if (!ptep) | |
4000 | return VM_FAULT_OOM; | |
4001 | } | |
4002 | ||
4003 | mapping = vma->vm_file->f_mapping; | |
4004 | idx = vma_hugecache_offset(h, vma, haddr); | |
4005 | ||
4006 | /* | |
4007 | * Serialize hugepage allocation and instantiation, so that we don't | |
4008 | * get spurious allocation failures if two CPUs race to instantiate | |
4009 | * the same page in the page cache. | |
4010 | */ | |
4011 | hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, haddr); | |
4012 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
4013 | ||
4014 | entry = huge_ptep_get(ptep); | |
4015 | if (huge_pte_none(entry)) { | |
4016 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); | |
4017 | goto out_mutex; | |
4018 | } | |
4019 | ||
4020 | ret = 0; | |
4021 | ||
4022 | /* | |
4023 | * entry could be a migration/hwpoison entry at this point, so this | |
4024 | * check prevents the kernel from going below assuming that we have | |
4025 | * a active hugepage in pagecache. This goto expects the 2nd page fault, | |
4026 | * and is_hugetlb_entry_(migration|hwpoisoned) check will properly | |
4027 | * handle it. | |
4028 | */ | |
4029 | if (!pte_present(entry)) | |
4030 | goto out_mutex; | |
4031 | ||
4032 | /* | |
4033 | * If we are going to COW the mapping later, we examine the pending | |
4034 | * reservations for this page now. This will ensure that any | |
4035 | * allocations necessary to record that reservation occur outside the | |
4036 | * spinlock. For private mappings, we also lookup the pagecache | |
4037 | * page now as it is used to determine if a reservation has been | |
4038 | * consumed. | |
4039 | */ | |
4040 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { | |
4041 | if (vma_needs_reservation(h, vma, haddr) < 0) { | |
4042 | ret = VM_FAULT_OOM; | |
4043 | goto out_mutex; | |
4044 | } | |
4045 | /* Just decrements count, does not deallocate */ | |
4046 | vma_end_reservation(h, vma, haddr); | |
4047 | ||
4048 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
4049 | pagecache_page = hugetlbfs_pagecache_page(h, | |
4050 | vma, haddr); | |
4051 | } | |
4052 | ||
4053 | ptl = huge_pte_lock(h, mm, ptep); | |
4054 | ||
4055 | /* Check for a racing update before calling hugetlb_cow */ | |
4056 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
4057 | goto out_ptl; | |
4058 | ||
4059 | /* | |
4060 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
4061 | * pagecache_page, so here we need take the former one | |
4062 | * when page != pagecache_page or !pagecache_page. | |
4063 | */ | |
4064 | page = pte_page(entry); | |
4065 | if (page != pagecache_page) | |
4066 | if (!trylock_page(page)) { | |
4067 | need_wait_lock = 1; | |
4068 | goto out_ptl; | |
4069 | } | |
4070 | ||
4071 | get_page(page); | |
4072 | ||
4073 | if (flags & FAULT_FLAG_WRITE) { | |
4074 | if (!huge_pte_write(entry)) { | |
4075 | ret = hugetlb_cow(mm, vma, address, ptep, | |
4076 | pagecache_page, ptl); | |
4077 | goto out_put_page; | |
4078 | } | |
4079 | entry = huge_pte_mkdirty(entry); | |
4080 | } | |
4081 | entry = pte_mkyoung(entry); | |
4082 | if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, | |
4083 | flags & FAULT_FLAG_WRITE)) | |
4084 | update_mmu_cache(vma, haddr, ptep); | |
4085 | out_put_page: | |
4086 | if (page != pagecache_page) | |
4087 | unlock_page(page); | |
4088 | put_page(page); | |
4089 | out_ptl: | |
4090 | spin_unlock(ptl); | |
4091 | ||
4092 | if (pagecache_page) { | |
4093 | unlock_page(pagecache_page); | |
4094 | put_page(pagecache_page); | |
4095 | } | |
4096 | out_mutex: | |
4097 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
4098 | /* | |
4099 | * Generally it's safe to hold refcount during waiting page lock. But | |
4100 | * here we just wait to defer the next page fault to avoid busy loop and | |
4101 | * the page is not used after unlocked before returning from the current | |
4102 | * page fault. So we are safe from accessing freed page, even if we wait | |
4103 | * here without taking refcount. | |
4104 | */ | |
4105 | if (need_wait_lock) | |
4106 | wait_on_page_locked(page); | |
4107 | return ret; | |
4108 | } | |
4109 | ||
4110 | /* | |
4111 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
4112 | * modifications for huge pages. | |
4113 | */ | |
4114 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
4115 | pte_t *dst_pte, | |
4116 | struct vm_area_struct *dst_vma, | |
4117 | unsigned long dst_addr, | |
4118 | unsigned long src_addr, | |
4119 | struct page **pagep) | |
4120 | { | |
4121 | struct address_space *mapping; | |
4122 | pgoff_t idx; | |
4123 | unsigned long size; | |
4124 | int vm_shared = dst_vma->vm_flags & VM_SHARED; | |
4125 | struct hstate *h = hstate_vma(dst_vma); | |
4126 | pte_t _dst_pte; | |
4127 | spinlock_t *ptl; | |
4128 | int ret; | |
4129 | struct page *page; | |
4130 | ||
4131 | if (!*pagep) { | |
4132 | ret = -ENOMEM; | |
4133 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
4134 | if (IS_ERR(page)) | |
4135 | goto out; | |
4136 | ||
4137 | ret = copy_huge_page_from_user(page, | |
4138 | (const void __user *) src_addr, | |
4139 | pages_per_huge_page(h), false); | |
4140 | ||
4141 | /* fallback to copy_from_user outside mmap_sem */ | |
4142 | if (unlikely(ret)) { | |
4143 | ret = -ENOENT; | |
4144 | *pagep = page; | |
4145 | /* don't free the page */ | |
4146 | goto out; | |
4147 | } | |
4148 | } else { | |
4149 | page = *pagep; | |
4150 | *pagep = NULL; | |
4151 | } | |
4152 | ||
4153 | /* | |
4154 | * The memory barrier inside __SetPageUptodate makes sure that | |
4155 | * preceding stores to the page contents become visible before | |
4156 | * the set_pte_at() write. | |
4157 | */ | |
4158 | __SetPageUptodate(page); | |
4159 | ||
4160 | mapping = dst_vma->vm_file->f_mapping; | |
4161 | idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
4162 | ||
4163 | /* | |
4164 | * If shared, add to page cache | |
4165 | */ | |
4166 | if (vm_shared) { | |
4167 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
4168 | ret = -EFAULT; | |
4169 | if (idx >= size) | |
4170 | goto out_release_nounlock; | |
4171 | ||
4172 | /* | |
4173 | * Serialization between remove_inode_hugepages() and | |
4174 | * huge_add_to_page_cache() below happens through the | |
4175 | * hugetlb_fault_mutex_table that here must be hold by | |
4176 | * the caller. | |
4177 | */ | |
4178 | ret = huge_add_to_page_cache(page, mapping, idx); | |
4179 | if (ret) | |
4180 | goto out_release_nounlock; | |
4181 | } | |
4182 | ||
4183 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); | |
4184 | spin_lock(ptl); | |
4185 | ||
4186 | /* | |
4187 | * Recheck the i_size after holding PT lock to make sure not | |
4188 | * to leave any page mapped (as page_mapped()) beyond the end | |
4189 | * of the i_size (remove_inode_hugepages() is strict about | |
4190 | * enforcing that). If we bail out here, we'll also leave a | |
4191 | * page in the radix tree in the vm_shared case beyond the end | |
4192 | * of the i_size, but remove_inode_hugepages() will take care | |
4193 | * of it as soon as we drop the hugetlb_fault_mutex_table. | |
4194 | */ | |
4195 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
4196 | ret = -EFAULT; | |
4197 | if (idx >= size) | |
4198 | goto out_release_unlock; | |
4199 | ||
4200 | ret = -EEXIST; | |
4201 | if (!huge_pte_none(huge_ptep_get(dst_pte))) | |
4202 | goto out_release_unlock; | |
4203 | ||
4204 | if (vm_shared) { | |
4205 | page_dup_rmap(page, true); | |
4206 | } else { | |
4207 | ClearPagePrivate(page); | |
4208 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); | |
4209 | } | |
4210 | ||
4211 | _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); | |
4212 | if (dst_vma->vm_flags & VM_WRITE) | |
4213 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
4214 | _dst_pte = pte_mkyoung(_dst_pte); | |
4215 | ||
4216 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); | |
4217 | ||
4218 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
4219 | dst_vma->vm_flags & VM_WRITE); | |
4220 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
4221 | ||
4222 | /* No need to invalidate - it was non-present before */ | |
4223 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
4224 | ||
4225 | spin_unlock(ptl); | |
4226 | set_page_huge_active(page); | |
4227 | if (vm_shared) | |
4228 | unlock_page(page); | |
4229 | ret = 0; | |
4230 | out: | |
4231 | return ret; | |
4232 | out_release_unlock: | |
4233 | spin_unlock(ptl); | |
4234 | if (vm_shared) | |
4235 | unlock_page(page); | |
4236 | out_release_nounlock: | |
4237 | put_page(page); | |
4238 | goto out; | |
4239 | } | |
4240 | ||
4241 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, | |
4242 | struct page **pages, struct vm_area_struct **vmas, | |
4243 | unsigned long *position, unsigned long *nr_pages, | |
4244 | long i, unsigned int flags, int *nonblocking) | |
4245 | { | |
4246 | unsigned long pfn_offset; | |
4247 | unsigned long vaddr = *position; | |
4248 | unsigned long remainder = *nr_pages; | |
4249 | struct hstate *h = hstate_vma(vma); | |
4250 | int err = -EFAULT; | |
4251 | ||
4252 | while (vaddr < vma->vm_end && remainder) { | |
4253 | pte_t *pte; | |
4254 | spinlock_t *ptl = NULL; | |
4255 | int absent; | |
4256 | struct page *page; | |
4257 | ||
4258 | /* | |
4259 | * If we have a pending SIGKILL, don't keep faulting pages and | |
4260 | * potentially allocating memory. | |
4261 | */ | |
4262 | if (fatal_signal_pending(current)) { | |
4263 | remainder = 0; | |
4264 | break; | |
4265 | } | |
4266 | ||
4267 | /* | |
4268 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
4269 | * each hugepage. We have to make sure we get the | |
4270 | * first, for the page indexing below to work. | |
4271 | * | |
4272 | * Note that page table lock is not held when pte is null. | |
4273 | */ | |
4274 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), | |
4275 | huge_page_size(h)); | |
4276 | if (pte) | |
4277 | ptl = huge_pte_lock(h, mm, pte); | |
4278 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); | |
4279 | ||
4280 | /* | |
4281 | * When coredumping, it suits get_dump_page if we just return | |
4282 | * an error where there's an empty slot with no huge pagecache | |
4283 | * to back it. This way, we avoid allocating a hugepage, and | |
4284 | * the sparse dumpfile avoids allocating disk blocks, but its | |
4285 | * huge holes still show up with zeroes where they need to be. | |
4286 | */ | |
4287 | if (absent && (flags & FOLL_DUMP) && | |
4288 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
4289 | if (pte) | |
4290 | spin_unlock(ptl); | |
4291 | remainder = 0; | |
4292 | break; | |
4293 | } | |
4294 | ||
4295 | /* | |
4296 | * We need call hugetlb_fault for both hugepages under migration | |
4297 | * (in which case hugetlb_fault waits for the migration,) and | |
4298 | * hwpoisoned hugepages (in which case we need to prevent the | |
4299 | * caller from accessing to them.) In order to do this, we use | |
4300 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
4301 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
4302 | * both cases, and because we can't follow correct pages | |
4303 | * directly from any kind of swap entries. | |
4304 | */ | |
4305 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
4306 | ((flags & FOLL_WRITE) && | |
4307 | !huge_pte_write(huge_ptep_get(pte)))) { | |
4308 | vm_fault_t ret; | |
4309 | unsigned int fault_flags = 0; | |
4310 | ||
4311 | if (pte) | |
4312 | spin_unlock(ptl); | |
4313 | if (flags & FOLL_WRITE) | |
4314 | fault_flags |= FAULT_FLAG_WRITE; | |
4315 | if (nonblocking) | |
4316 | fault_flags |= FAULT_FLAG_ALLOW_RETRY; | |
4317 | if (flags & FOLL_NOWAIT) | |
4318 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
4319 | FAULT_FLAG_RETRY_NOWAIT; | |
4320 | if (flags & FOLL_TRIED) { | |
4321 | VM_WARN_ON_ONCE(fault_flags & | |
4322 | FAULT_FLAG_ALLOW_RETRY); | |
4323 | fault_flags |= FAULT_FLAG_TRIED; | |
4324 | } | |
4325 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
4326 | if (ret & VM_FAULT_ERROR) { | |
4327 | err = vm_fault_to_errno(ret, flags); | |
4328 | remainder = 0; | |
4329 | break; | |
4330 | } | |
4331 | if (ret & VM_FAULT_RETRY) { | |
4332 | if (nonblocking && | |
4333 | !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) | |
4334 | *nonblocking = 0; | |
4335 | *nr_pages = 0; | |
4336 | /* | |
4337 | * VM_FAULT_RETRY must not return an | |
4338 | * error, it will return zero | |
4339 | * instead. | |
4340 | * | |
4341 | * No need to update "position" as the | |
4342 | * caller will not check it after | |
4343 | * *nr_pages is set to 0. | |
4344 | */ | |
4345 | return i; | |
4346 | } | |
4347 | continue; | |
4348 | } | |
4349 | ||
4350 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; | |
4351 | page = pte_page(huge_ptep_get(pte)); | |
4352 | ||
4353 | /* | |
4354 | * Instead of doing 'try_get_page()' below in the same_page | |
4355 | * loop, just check the count once here. | |
4356 | */ | |
4357 | if (unlikely(page_count(page) <= 0)) { | |
4358 | if (pages) { | |
4359 | spin_unlock(ptl); | |
4360 | remainder = 0; | |
4361 | err = -ENOMEM; | |
4362 | break; | |
4363 | } | |
4364 | } | |
4365 | same_page: | |
4366 | if (pages) { | |
4367 | pages[i] = mem_map_offset(page, pfn_offset); | |
4368 | get_page(pages[i]); | |
4369 | } | |
4370 | ||
4371 | if (vmas) | |
4372 | vmas[i] = vma; | |
4373 | ||
4374 | vaddr += PAGE_SIZE; | |
4375 | ++pfn_offset; | |
4376 | --remainder; | |
4377 | ++i; | |
4378 | if (vaddr < vma->vm_end && remainder && | |
4379 | pfn_offset < pages_per_huge_page(h)) { | |
4380 | /* | |
4381 | * We use pfn_offset to avoid touching the pageframes | |
4382 | * of this compound page. | |
4383 | */ | |
4384 | goto same_page; | |
4385 | } | |
4386 | spin_unlock(ptl); | |
4387 | } | |
4388 | *nr_pages = remainder; | |
4389 | /* | |
4390 | * setting position is actually required only if remainder is | |
4391 | * not zero but it's faster not to add a "if (remainder)" | |
4392 | * branch. | |
4393 | */ | |
4394 | *position = vaddr; | |
4395 | ||
4396 | return i ? i : err; | |
4397 | } | |
4398 | ||
4399 | #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE | |
4400 | /* | |
4401 | * ARCHes with special requirements for evicting HUGETLB backing TLB entries can | |
4402 | * implement this. | |
4403 | */ | |
4404 | #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) | |
4405 | #endif | |
4406 | ||
4407 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, | |
4408 | unsigned long address, unsigned long end, pgprot_t newprot) | |
4409 | { | |
4410 | struct mm_struct *mm = vma->vm_mm; | |
4411 | unsigned long start = address; | |
4412 | pte_t *ptep; | |
4413 | pte_t pte; | |
4414 | struct hstate *h = hstate_vma(vma); | |
4415 | unsigned long pages = 0; | |
4416 | bool shared_pmd = false; | |
4417 | struct mmu_notifier_range range; | |
4418 | ||
4419 | /* | |
4420 | * In the case of shared PMDs, the area to flush could be beyond | |
4421 | * start/end. Set range.start/range.end to cover the maximum possible | |
4422 | * range if PMD sharing is possible. | |
4423 | */ | |
4424 | mmu_notifier_range_init(&range, mm, start, end); | |
4425 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); | |
4426 | ||
4427 | BUG_ON(address >= end); | |
4428 | flush_cache_range(vma, range.start, range.end); | |
4429 | ||
4430 | mmu_notifier_invalidate_range_start(&range); | |
4431 | i_mmap_lock_write(vma->vm_file->f_mapping); | |
4432 | for (; address < end; address += huge_page_size(h)) { | |
4433 | spinlock_t *ptl; | |
4434 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); | |
4435 | if (!ptep) | |
4436 | continue; | |
4437 | ptl = huge_pte_lock(h, mm, ptep); | |
4438 | if (huge_pmd_unshare(mm, &address, ptep)) { | |
4439 | pages++; | |
4440 | spin_unlock(ptl); | |
4441 | shared_pmd = true; | |
4442 | continue; | |
4443 | } | |
4444 | pte = huge_ptep_get(ptep); | |
4445 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
4446 | spin_unlock(ptl); | |
4447 | continue; | |
4448 | } | |
4449 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
4450 | swp_entry_t entry = pte_to_swp_entry(pte); | |
4451 | ||
4452 | if (is_write_migration_entry(entry)) { | |
4453 | pte_t newpte; | |
4454 | ||
4455 | make_migration_entry_read(&entry); | |
4456 | newpte = swp_entry_to_pte(entry); | |
4457 | set_huge_swap_pte_at(mm, address, ptep, | |
4458 | newpte, huge_page_size(h)); | |
4459 | pages++; | |
4460 | } | |
4461 | spin_unlock(ptl); | |
4462 | continue; | |
4463 | } | |
4464 | if (!huge_pte_none(pte)) { | |
4465 | pte_t old_pte; | |
4466 | ||
4467 | old_pte = huge_ptep_modify_prot_start(vma, address, ptep); | |
4468 | pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); | |
4469 | pte = arch_make_huge_pte(pte, vma, NULL, 0); | |
4470 | huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); | |
4471 | pages++; | |
4472 | } | |
4473 | spin_unlock(ptl); | |
4474 | } | |
4475 | /* | |
4476 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare | |
4477 | * may have cleared our pud entry and done put_page on the page table: | |
4478 | * once we release i_mmap_rwsem, another task can do the final put_page | |
4479 | * and that page table be reused and filled with junk. If we actually | |
4480 | * did unshare a page of pmds, flush the range corresponding to the pud. | |
4481 | */ | |
4482 | if (shared_pmd) | |
4483 | flush_hugetlb_tlb_range(vma, range.start, range.end); | |
4484 | else | |
4485 | flush_hugetlb_tlb_range(vma, start, end); | |
4486 | /* | |
4487 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
4488 | * page table protection not changing it to point to a new page. | |
4489 | * | |
4490 | * See Documentation/vm/mmu_notifier.rst | |
4491 | */ | |
4492 | i_mmap_unlock_write(vma->vm_file->f_mapping); | |
4493 | mmu_notifier_invalidate_range_end(&range); | |
4494 | ||
4495 | return pages << h->order; | |
4496 | } | |
4497 | ||
4498 | int hugetlb_reserve_pages(struct inode *inode, | |
4499 | long from, long to, | |
4500 | struct vm_area_struct *vma, | |
4501 | vm_flags_t vm_flags) | |
4502 | { | |
4503 | long ret, chg; | |
4504 | struct hstate *h = hstate_inode(inode); | |
4505 | struct hugepage_subpool *spool = subpool_inode(inode); | |
4506 | struct resv_map *resv_map; | |
4507 | long gbl_reserve; | |
4508 | ||
4509 | /* This should never happen */ | |
4510 | if (from > to) { | |
4511 | VM_WARN(1, "%s called with a negative range\n", __func__); | |
4512 | return -EINVAL; | |
4513 | } | |
4514 | ||
4515 | /* | |
4516 | * Only apply hugepage reservation if asked. At fault time, an | |
4517 | * attempt will be made for VM_NORESERVE to allocate a page | |
4518 | * without using reserves | |
4519 | */ | |
4520 | if (vm_flags & VM_NORESERVE) | |
4521 | return 0; | |
4522 | ||
4523 | /* | |
4524 | * Shared mappings base their reservation on the number of pages that | |
4525 | * are already allocated on behalf of the file. Private mappings need | |
4526 | * to reserve the full area even if read-only as mprotect() may be | |
4527 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
4528 | */ | |
4529 | if (!vma || vma->vm_flags & VM_MAYSHARE) { | |
4530 | resv_map = inode_resv_map(inode); | |
4531 | ||
4532 | chg = region_chg(resv_map, from, to); | |
4533 | ||
4534 | } else { | |
4535 | resv_map = resv_map_alloc(); | |
4536 | if (!resv_map) | |
4537 | return -ENOMEM; | |
4538 | ||
4539 | chg = to - from; | |
4540 | ||
4541 | set_vma_resv_map(vma, resv_map); | |
4542 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
4543 | } | |
4544 | ||
4545 | if (chg < 0) { | |
4546 | ret = chg; | |
4547 | goto out_err; | |
4548 | } | |
4549 | ||
4550 | /* | |
4551 | * There must be enough pages in the subpool for the mapping. If | |
4552 | * the subpool has a minimum size, there may be some global | |
4553 | * reservations already in place (gbl_reserve). | |
4554 | */ | |
4555 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
4556 | if (gbl_reserve < 0) { | |
4557 | ret = -ENOSPC; | |
4558 | goto out_err; | |
4559 | } | |
4560 | ||
4561 | /* | |
4562 | * Check enough hugepages are available for the reservation. | |
4563 | * Hand the pages back to the subpool if there are not | |
4564 | */ | |
4565 | ret = hugetlb_acct_memory(h, gbl_reserve); | |
4566 | if (ret < 0) { | |
4567 | /* put back original number of pages, chg */ | |
4568 | (void)hugepage_subpool_put_pages(spool, chg); | |
4569 | goto out_err; | |
4570 | } | |
4571 | ||
4572 | /* | |
4573 | * Account for the reservations made. Shared mappings record regions | |
4574 | * that have reservations as they are shared by multiple VMAs. | |
4575 | * When the last VMA disappears, the region map says how much | |
4576 | * the reservation was and the page cache tells how much of | |
4577 | * the reservation was consumed. Private mappings are per-VMA and | |
4578 | * only the consumed reservations are tracked. When the VMA | |
4579 | * disappears, the original reservation is the VMA size and the | |
4580 | * consumed reservations are stored in the map. Hence, nothing | |
4581 | * else has to be done for private mappings here | |
4582 | */ | |
4583 | if (!vma || vma->vm_flags & VM_MAYSHARE) { | |
4584 | long add = region_add(resv_map, from, to); | |
4585 | ||
4586 | if (unlikely(chg > add)) { | |
4587 | /* | |
4588 | * pages in this range were added to the reserve | |
4589 | * map between region_chg and region_add. This | |
4590 | * indicates a race with alloc_huge_page. Adjust | |
4591 | * the subpool and reserve counts modified above | |
4592 | * based on the difference. | |
4593 | */ | |
4594 | long rsv_adjust; | |
4595 | ||
4596 | rsv_adjust = hugepage_subpool_put_pages(spool, | |
4597 | chg - add); | |
4598 | hugetlb_acct_memory(h, -rsv_adjust); | |
4599 | } | |
4600 | } | |
4601 | return 0; | |
4602 | out_err: | |
4603 | if (!vma || vma->vm_flags & VM_MAYSHARE) | |
4604 | /* Don't call region_abort if region_chg failed */ | |
4605 | if (chg >= 0) | |
4606 | region_abort(resv_map, from, to); | |
4607 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) | |
4608 | kref_put(&resv_map->refs, resv_map_release); | |
4609 | return ret; | |
4610 | } | |
4611 | ||
4612 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, | |
4613 | long freed) | |
4614 | { | |
4615 | struct hstate *h = hstate_inode(inode); | |
4616 | struct resv_map *resv_map = inode_resv_map(inode); | |
4617 | long chg = 0; | |
4618 | struct hugepage_subpool *spool = subpool_inode(inode); | |
4619 | long gbl_reserve; | |
4620 | ||
4621 | if (resv_map) { | |
4622 | chg = region_del(resv_map, start, end); | |
4623 | /* | |
4624 | * region_del() can fail in the rare case where a region | |
4625 | * must be split and another region descriptor can not be | |
4626 | * allocated. If end == LONG_MAX, it will not fail. | |
4627 | */ | |
4628 | if (chg < 0) | |
4629 | return chg; | |
4630 | } | |
4631 | ||
4632 | spin_lock(&inode->i_lock); | |
4633 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); | |
4634 | spin_unlock(&inode->i_lock); | |
4635 | ||
4636 | /* | |
4637 | * If the subpool has a minimum size, the number of global | |
4638 | * reservations to be released may be adjusted. | |
4639 | */ | |
4640 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
4641 | hugetlb_acct_memory(h, -gbl_reserve); | |
4642 | ||
4643 | return 0; | |
4644 | } | |
4645 | ||
4646 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE | |
4647 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
4648 | struct vm_area_struct *vma, | |
4649 | unsigned long addr, pgoff_t idx) | |
4650 | { | |
4651 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
4652 | svma->vm_start; | |
4653 | unsigned long sbase = saddr & PUD_MASK; | |
4654 | unsigned long s_end = sbase + PUD_SIZE; | |
4655 | ||
4656 | /* Allow segments to share if only one is marked locked */ | |
4657 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
4658 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
4659 | ||
4660 | /* | |
4661 | * match the virtual addresses, permission and the alignment of the | |
4662 | * page table page. | |
4663 | */ | |
4664 | if (pmd_index(addr) != pmd_index(saddr) || | |
4665 | vm_flags != svm_flags || | |
4666 | sbase < svma->vm_start || svma->vm_end < s_end) | |
4667 | return 0; | |
4668 | ||
4669 | return saddr; | |
4670 | } | |
4671 | ||
4672 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) | |
4673 | { | |
4674 | unsigned long base = addr & PUD_MASK; | |
4675 | unsigned long end = base + PUD_SIZE; | |
4676 | ||
4677 | /* | |
4678 | * check on proper vm_flags and page table alignment | |
4679 | */ | |
4680 | if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) | |
4681 | return true; | |
4682 | return false; | |
4683 | } | |
4684 | ||
4685 | /* | |
4686 | * Determine if start,end range within vma could be mapped by shared pmd. | |
4687 | * If yes, adjust start and end to cover range associated with possible | |
4688 | * shared pmd mappings. | |
4689 | */ | |
4690 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
4691 | unsigned long *start, unsigned long *end) | |
4692 | { | |
4693 | unsigned long check_addr = *start; | |
4694 | ||
4695 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
4696 | return; | |
4697 | ||
4698 | for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) { | |
4699 | unsigned long a_start = check_addr & PUD_MASK; | |
4700 | unsigned long a_end = a_start + PUD_SIZE; | |
4701 | ||
4702 | /* | |
4703 | * If sharing is possible, adjust start/end if necessary. | |
4704 | */ | |
4705 | if (range_in_vma(vma, a_start, a_end)) { | |
4706 | if (a_start < *start) | |
4707 | *start = a_start; | |
4708 | if (a_end > *end) | |
4709 | *end = a_end; | |
4710 | } | |
4711 | } | |
4712 | } | |
4713 | ||
4714 | /* | |
4715 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
4716 | * and returns the corresponding pte. While this is not necessary for the | |
4717 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
4718 | * code much cleaner. pmd allocation is essential for the shared case because | |
4719 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise | |
4720 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a | |
4721 | * bad pmd for sharing. | |
4722 | */ | |
4723 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4724 | { | |
4725 | struct vm_area_struct *vma = find_vma(mm, addr); | |
4726 | struct address_space *mapping = vma->vm_file->f_mapping; | |
4727 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
4728 | vma->vm_pgoff; | |
4729 | struct vm_area_struct *svma; | |
4730 | unsigned long saddr; | |
4731 | pte_t *spte = NULL; | |
4732 | pte_t *pte; | |
4733 | spinlock_t *ptl; | |
4734 | ||
4735 | if (!vma_shareable(vma, addr)) | |
4736 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
4737 | ||
4738 | i_mmap_lock_write(mapping); | |
4739 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { | |
4740 | if (svma == vma) | |
4741 | continue; | |
4742 | ||
4743 | saddr = page_table_shareable(svma, vma, addr, idx); | |
4744 | if (saddr) { | |
4745 | spte = huge_pte_offset(svma->vm_mm, saddr, | |
4746 | vma_mmu_pagesize(svma)); | |
4747 | if (spte) { | |
4748 | get_page(virt_to_page(spte)); | |
4749 | break; | |
4750 | } | |
4751 | } | |
4752 | } | |
4753 | ||
4754 | if (!spte) | |
4755 | goto out; | |
4756 | ||
4757 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); | |
4758 | if (pud_none(*pud)) { | |
4759 | pud_populate(mm, pud, | |
4760 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
4761 | mm_inc_nr_pmds(mm); | |
4762 | } else { | |
4763 | put_page(virt_to_page(spte)); | |
4764 | } | |
4765 | spin_unlock(ptl); | |
4766 | out: | |
4767 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
4768 | i_mmap_unlock_write(mapping); | |
4769 | return pte; | |
4770 | } | |
4771 | ||
4772 | /* | |
4773 | * unmap huge page backed by shared pte. | |
4774 | * | |
4775 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
4776 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
4777 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
4778 | * | |
4779 | * called with page table lock held. | |
4780 | * | |
4781 | * returns: 1 successfully unmapped a shared pte page | |
4782 | * 0 the underlying pte page is not shared, or it is the last user | |
4783 | */ | |
4784 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4785 | { | |
4786 | pgd_t *pgd = pgd_offset(mm, *addr); | |
4787 | p4d_t *p4d = p4d_offset(pgd, *addr); | |
4788 | pud_t *pud = pud_offset(p4d, *addr); | |
4789 | ||
4790 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
4791 | if (page_count(virt_to_page(ptep)) == 1) | |
4792 | return 0; | |
4793 | ||
4794 | pud_clear(pud); | |
4795 | put_page(virt_to_page(ptep)); | |
4796 | mm_dec_nr_pmds(mm); | |
4797 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; | |
4798 | return 1; | |
4799 | } | |
4800 | #define want_pmd_share() (1) | |
4801 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
4802 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4803 | { | |
4804 | return NULL; | |
4805 | } | |
4806 | ||
4807 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4808 | { | |
4809 | return 0; | |
4810 | } | |
4811 | ||
4812 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
4813 | unsigned long *start, unsigned long *end) | |
4814 | { | |
4815 | } | |
4816 | #define want_pmd_share() (0) | |
4817 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
4818 | ||
4819 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB | |
4820 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
4821 | unsigned long addr, unsigned long sz) | |
4822 | { | |
4823 | pgd_t *pgd; | |
4824 | p4d_t *p4d; | |
4825 | pud_t *pud; | |
4826 | pte_t *pte = NULL; | |
4827 | ||
4828 | pgd = pgd_offset(mm, addr); | |
4829 | p4d = p4d_alloc(mm, pgd, addr); | |
4830 | if (!p4d) | |
4831 | return NULL; | |
4832 | pud = pud_alloc(mm, p4d, addr); | |
4833 | if (pud) { | |
4834 | if (sz == PUD_SIZE) { | |
4835 | pte = (pte_t *)pud; | |
4836 | } else { | |
4837 | BUG_ON(sz != PMD_SIZE); | |
4838 | if (want_pmd_share() && pud_none(*pud)) | |
4839 | pte = huge_pmd_share(mm, addr, pud); | |
4840 | else | |
4841 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
4842 | } | |
4843 | } | |
4844 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); | |
4845 | ||
4846 | return pte; | |
4847 | } | |
4848 | ||
4849 | /* | |
4850 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
4851 | * entry at address @addr | |
4852 | * | |
4853 | * Return: Pointer to page table or swap entry (PUD or PMD) for | |
4854 | * address @addr, or NULL if a p*d_none() entry is encountered and the | |
4855 | * size @sz doesn't match the hugepage size at this level of the page | |
4856 | * table. | |
4857 | */ | |
4858 | pte_t *huge_pte_offset(struct mm_struct *mm, | |
4859 | unsigned long addr, unsigned long sz) | |
4860 | { | |
4861 | pgd_t *pgd; | |
4862 | p4d_t *p4d; | |
4863 | pud_t *pud; | |
4864 | pmd_t *pmd; | |
4865 | ||
4866 | pgd = pgd_offset(mm, addr); | |
4867 | if (!pgd_present(*pgd)) | |
4868 | return NULL; | |
4869 | p4d = p4d_offset(pgd, addr); | |
4870 | if (!p4d_present(*p4d)) | |
4871 | return NULL; | |
4872 | ||
4873 | pud = pud_offset(p4d, addr); | |
4874 | if (sz != PUD_SIZE && pud_none(*pud)) | |
4875 | return NULL; | |
4876 | /* hugepage or swap? */ | |
4877 | if (pud_huge(*pud) || !pud_present(*pud)) | |
4878 | return (pte_t *)pud; | |
4879 | ||
4880 | pmd = pmd_offset(pud, addr); | |
4881 | if (sz != PMD_SIZE && pmd_none(*pmd)) | |
4882 | return NULL; | |
4883 | /* hugepage or swap? */ | |
4884 | if (pmd_huge(*pmd) || !pmd_present(*pmd)) | |
4885 | return (pte_t *)pmd; | |
4886 | ||
4887 | return NULL; | |
4888 | } | |
4889 | ||
4890 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ | |
4891 | ||
4892 | /* | |
4893 | * These functions are overwritable if your architecture needs its own | |
4894 | * behavior. | |
4895 | */ | |
4896 | struct page * __weak | |
4897 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
4898 | int write) | |
4899 | { | |
4900 | return ERR_PTR(-EINVAL); | |
4901 | } | |
4902 | ||
4903 | struct page * __weak | |
4904 | follow_huge_pd(struct vm_area_struct *vma, | |
4905 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
4906 | { | |
4907 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
4908 | return NULL; | |
4909 | } | |
4910 | ||
4911 | struct page * __weak | |
4912 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, | |
4913 | pmd_t *pmd, int flags) | |
4914 | { | |
4915 | struct page *page = NULL; | |
4916 | spinlock_t *ptl; | |
4917 | pte_t pte; | |
4918 | retry: | |
4919 | ptl = pmd_lockptr(mm, pmd); | |
4920 | spin_lock(ptl); | |
4921 | /* | |
4922 | * make sure that the address range covered by this pmd is not | |
4923 | * unmapped from other threads. | |
4924 | */ | |
4925 | if (!pmd_huge(*pmd)) | |
4926 | goto out; | |
4927 | pte = huge_ptep_get((pte_t *)pmd); | |
4928 | if (pte_present(pte)) { | |
4929 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); | |
4930 | if (flags & FOLL_GET) | |
4931 | get_page(page); | |
4932 | } else { | |
4933 | if (is_hugetlb_entry_migration(pte)) { | |
4934 | spin_unlock(ptl); | |
4935 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
4936 | goto retry; | |
4937 | } | |
4938 | /* | |
4939 | * hwpoisoned entry is treated as no_page_table in | |
4940 | * follow_page_mask(). | |
4941 | */ | |
4942 | } | |
4943 | out: | |
4944 | spin_unlock(ptl); | |
4945 | return page; | |
4946 | } | |
4947 | ||
4948 | struct page * __weak | |
4949 | follow_huge_pud(struct mm_struct *mm, unsigned long address, | |
4950 | pud_t *pud, int flags) | |
4951 | { | |
4952 | if (flags & FOLL_GET) | |
4953 | return NULL; | |
4954 | ||
4955 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); | |
4956 | } | |
4957 | ||
4958 | struct page * __weak | |
4959 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
4960 | { | |
4961 | if (flags & FOLL_GET) | |
4962 | return NULL; | |
4963 | ||
4964 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
4965 | } | |
4966 | ||
4967 | bool isolate_huge_page(struct page *page, struct list_head *list) | |
4968 | { | |
4969 | bool ret = true; | |
4970 | ||
4971 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
4972 | spin_lock(&hugetlb_lock); | |
4973 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { | |
4974 | ret = false; | |
4975 | goto unlock; | |
4976 | } | |
4977 | clear_page_huge_active(page); | |
4978 | list_move_tail(&page->lru, list); | |
4979 | unlock: | |
4980 | spin_unlock(&hugetlb_lock); | |
4981 | return ret; | |
4982 | } | |
4983 | ||
4984 | void putback_active_hugepage(struct page *page) | |
4985 | { | |
4986 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
4987 | spin_lock(&hugetlb_lock); | |
4988 | set_page_huge_active(page); | |
4989 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); | |
4990 | spin_unlock(&hugetlb_lock); | |
4991 | put_page(page); | |
4992 | } | |
4993 | ||
4994 | void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) | |
4995 | { | |
4996 | struct hstate *h = page_hstate(oldpage); | |
4997 | ||
4998 | hugetlb_cgroup_migrate(oldpage, newpage); | |
4999 | set_page_owner_migrate_reason(newpage, reason); | |
5000 | ||
5001 | /* | |
5002 | * transfer temporary state of the new huge page. This is | |
5003 | * reverse to other transitions because the newpage is going to | |
5004 | * be final while the old one will be freed so it takes over | |
5005 | * the temporary status. | |
5006 | * | |
5007 | * Also note that we have to transfer the per-node surplus state | |
5008 | * here as well otherwise the global surplus count will not match | |
5009 | * the per-node's. | |
5010 | */ | |
5011 | if (PageHugeTemporary(newpage)) { | |
5012 | int old_nid = page_to_nid(oldpage); | |
5013 | int new_nid = page_to_nid(newpage); | |
5014 | ||
5015 | SetPageHugeTemporary(oldpage); | |
5016 | ClearPageHugeTemporary(newpage); | |
5017 | ||
5018 | spin_lock(&hugetlb_lock); | |
5019 | if (h->surplus_huge_pages_node[old_nid]) { | |
5020 | h->surplus_huge_pages_node[old_nid]--; | |
5021 | h->surplus_huge_pages_node[new_nid]++; | |
5022 | } | |
5023 | spin_unlock(&hugetlb_lock); | |
5024 | } | |
5025 | } |