]>
Commit | Line | Data |
---|---|---|
1 | // SPDX-License-Identifier: GPL-2.0-only | |
2 | /* | |
3 | * Generic hugetlb support. | |
4 | * (C) Nadia Yvette Chambers, April 2004 | |
5 | */ | |
6 | #include <linux/list.h> | |
7 | #include <linux/init.h> | |
8 | #include <linux/mm.h> | |
9 | #include <linux/seq_file.h> | |
10 | #include <linux/sysctl.h> | |
11 | #include <linux/highmem.h> | |
12 | #include <linux/mmu_notifier.h> | |
13 | #include <linux/nodemask.h> | |
14 | #include <linux/pagemap.h> | |
15 | #include <linux/mempolicy.h> | |
16 | #include <linux/compiler.h> | |
17 | #include <linux/cpuset.h> | |
18 | #include <linux/mutex.h> | |
19 | #include <linux/memblock.h> | |
20 | #include <linux/sysfs.h> | |
21 | #include <linux/slab.h> | |
22 | #include <linux/sched/mm.h> | |
23 | #include <linux/mmdebug.h> | |
24 | #include <linux/sched/signal.h> | |
25 | #include <linux/rmap.h> | |
26 | #include <linux/string_helpers.h> | |
27 | #include <linux/swap.h> | |
28 | #include <linux/swapops.h> | |
29 | #include <linux/jhash.h> | |
30 | #include <linux/numa.h> | |
31 | #include <linux/llist.h> | |
32 | #include <linux/cma.h> | |
33 | ||
34 | #include <asm/page.h> | |
35 | #include <asm/pgalloc.h> | |
36 | #include <asm/tlb.h> | |
37 | ||
38 | #include <linux/io.h> | |
39 | #include <linux/hugetlb.h> | |
40 | #include <linux/hugetlb_cgroup.h> | |
41 | #include <linux/node.h> | |
42 | #include <linux/userfaultfd_k.h> | |
43 | #include <linux/page_owner.h> | |
44 | #include "internal.h" | |
45 | ||
46 | int hugetlb_max_hstate __read_mostly; | |
47 | unsigned int default_hstate_idx; | |
48 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
49 | ||
50 | #ifdef CONFIG_CMA | |
51 | static struct cma *hugetlb_cma[MAX_NUMNODES]; | |
52 | #endif | |
53 | static unsigned long hugetlb_cma_size __initdata; | |
54 | ||
55 | /* | |
56 | * Minimum page order among possible hugepage sizes, set to a proper value | |
57 | * at boot time. | |
58 | */ | |
59 | static unsigned int minimum_order __read_mostly = UINT_MAX; | |
60 | ||
61 | __initdata LIST_HEAD(huge_boot_pages); | |
62 | ||
63 | /* for command line parsing */ | |
64 | static struct hstate * __initdata parsed_hstate; | |
65 | static unsigned long __initdata default_hstate_max_huge_pages; | |
66 | static bool __initdata parsed_valid_hugepagesz = true; | |
67 | static bool __initdata parsed_default_hugepagesz; | |
68 | ||
69 | /* | |
70 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, | |
71 | * free_huge_pages, and surplus_huge_pages. | |
72 | */ | |
73 | DEFINE_SPINLOCK(hugetlb_lock); | |
74 | ||
75 | /* | |
76 | * Serializes faults on the same logical page. This is used to | |
77 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
78 | */ | |
79 | static int num_fault_mutexes; | |
80 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; | |
81 | ||
82 | static inline bool PageHugeFreed(struct page *head) | |
83 | { | |
84 | return page_private(head + 4) == -1UL; | |
85 | } | |
86 | ||
87 | static inline void SetPageHugeFreed(struct page *head) | |
88 | { | |
89 | set_page_private(head + 4, -1UL); | |
90 | } | |
91 | ||
92 | static inline void ClearPageHugeFreed(struct page *head) | |
93 | { | |
94 | set_page_private(head + 4, 0); | |
95 | } | |
96 | ||
97 | /* Forward declaration */ | |
98 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
99 | ||
100 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) | |
101 | { | |
102 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
103 | ||
104 | spin_unlock(&spool->lock); | |
105 | ||
106 | /* If no pages are used, and no other handles to the subpool | |
107 | * remain, give up any reservations based on minimum size and | |
108 | * free the subpool */ | |
109 | if (free) { | |
110 | if (spool->min_hpages != -1) | |
111 | hugetlb_acct_memory(spool->hstate, | |
112 | -spool->min_hpages); | |
113 | kfree(spool); | |
114 | } | |
115 | } | |
116 | ||
117 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, | |
118 | long min_hpages) | |
119 | { | |
120 | struct hugepage_subpool *spool; | |
121 | ||
122 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); | |
123 | if (!spool) | |
124 | return NULL; | |
125 | ||
126 | spin_lock_init(&spool->lock); | |
127 | spool->count = 1; | |
128 | spool->max_hpages = max_hpages; | |
129 | spool->hstate = h; | |
130 | spool->min_hpages = min_hpages; | |
131 | ||
132 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
133 | kfree(spool); | |
134 | return NULL; | |
135 | } | |
136 | spool->rsv_hpages = min_hpages; | |
137 | ||
138 | return spool; | |
139 | } | |
140 | ||
141 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
142 | { | |
143 | spin_lock(&spool->lock); | |
144 | BUG_ON(!spool->count); | |
145 | spool->count--; | |
146 | unlock_or_release_subpool(spool); | |
147 | } | |
148 | ||
149 | /* | |
150 | * Subpool accounting for allocating and reserving pages. | |
151 | * Return -ENOMEM if there are not enough resources to satisfy the | |
152 | * request. Otherwise, return the number of pages by which the | |
153 | * global pools must be adjusted (upward). The returned value may | |
154 | * only be different than the passed value (delta) in the case where | |
155 | * a subpool minimum size must be maintained. | |
156 | */ | |
157 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
158 | long delta) | |
159 | { | |
160 | long ret = delta; | |
161 | ||
162 | if (!spool) | |
163 | return ret; | |
164 | ||
165 | spin_lock(&spool->lock); | |
166 | ||
167 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
168 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
169 | spool->used_hpages += delta; | |
170 | else { | |
171 | ret = -ENOMEM; | |
172 | goto unlock_ret; | |
173 | } | |
174 | } | |
175 | ||
176 | /* minimum size accounting */ | |
177 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
178 | if (delta > spool->rsv_hpages) { | |
179 | /* | |
180 | * Asking for more reserves than those already taken on | |
181 | * behalf of subpool. Return difference. | |
182 | */ | |
183 | ret = delta - spool->rsv_hpages; | |
184 | spool->rsv_hpages = 0; | |
185 | } else { | |
186 | ret = 0; /* reserves already accounted for */ | |
187 | spool->rsv_hpages -= delta; | |
188 | } | |
189 | } | |
190 | ||
191 | unlock_ret: | |
192 | spin_unlock(&spool->lock); | |
193 | return ret; | |
194 | } | |
195 | ||
196 | /* | |
197 | * Subpool accounting for freeing and unreserving pages. | |
198 | * Return the number of global page reservations that must be dropped. | |
199 | * The return value may only be different than the passed value (delta) | |
200 | * in the case where a subpool minimum size must be maintained. | |
201 | */ | |
202 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
203 | long delta) | |
204 | { | |
205 | long ret = delta; | |
206 | ||
207 | if (!spool) | |
208 | return delta; | |
209 | ||
210 | spin_lock(&spool->lock); | |
211 | ||
212 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
213 | spool->used_hpages -= delta; | |
214 | ||
215 | /* minimum size accounting */ | |
216 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
217 | if (spool->rsv_hpages + delta <= spool->min_hpages) | |
218 | ret = 0; | |
219 | else | |
220 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
221 | ||
222 | spool->rsv_hpages += delta; | |
223 | if (spool->rsv_hpages > spool->min_hpages) | |
224 | spool->rsv_hpages = spool->min_hpages; | |
225 | } | |
226 | ||
227 | /* | |
228 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
229 | * quota reference, free it now. | |
230 | */ | |
231 | unlock_or_release_subpool(spool); | |
232 | ||
233 | return ret; | |
234 | } | |
235 | ||
236 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
237 | { | |
238 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
239 | } | |
240 | ||
241 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
242 | { | |
243 | return subpool_inode(file_inode(vma->vm_file)); | |
244 | } | |
245 | ||
246 | /* Helper that removes a struct file_region from the resv_map cache and returns | |
247 | * it for use. | |
248 | */ | |
249 | static struct file_region * | |
250 | get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) | |
251 | { | |
252 | struct file_region *nrg = NULL; | |
253 | ||
254 | VM_BUG_ON(resv->region_cache_count <= 0); | |
255 | ||
256 | resv->region_cache_count--; | |
257 | nrg = list_first_entry(&resv->region_cache, struct file_region, link); | |
258 | list_del(&nrg->link); | |
259 | ||
260 | nrg->from = from; | |
261 | nrg->to = to; | |
262 | ||
263 | return nrg; | |
264 | } | |
265 | ||
266 | static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, | |
267 | struct file_region *rg) | |
268 | { | |
269 | #ifdef CONFIG_CGROUP_HUGETLB | |
270 | nrg->reservation_counter = rg->reservation_counter; | |
271 | nrg->css = rg->css; | |
272 | if (rg->css) | |
273 | css_get(rg->css); | |
274 | #endif | |
275 | } | |
276 | ||
277 | /* Helper that records hugetlb_cgroup uncharge info. */ | |
278 | static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, | |
279 | struct hstate *h, | |
280 | struct resv_map *resv, | |
281 | struct file_region *nrg) | |
282 | { | |
283 | #ifdef CONFIG_CGROUP_HUGETLB | |
284 | if (h_cg) { | |
285 | nrg->reservation_counter = | |
286 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
287 | nrg->css = &h_cg->css; | |
288 | if (!resv->pages_per_hpage) | |
289 | resv->pages_per_hpage = pages_per_huge_page(h); | |
290 | /* pages_per_hpage should be the same for all entries in | |
291 | * a resv_map. | |
292 | */ | |
293 | VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); | |
294 | } else { | |
295 | nrg->reservation_counter = NULL; | |
296 | nrg->css = NULL; | |
297 | } | |
298 | #endif | |
299 | } | |
300 | ||
301 | static bool has_same_uncharge_info(struct file_region *rg, | |
302 | struct file_region *org) | |
303 | { | |
304 | #ifdef CONFIG_CGROUP_HUGETLB | |
305 | return rg && org && | |
306 | rg->reservation_counter == org->reservation_counter && | |
307 | rg->css == org->css; | |
308 | ||
309 | #else | |
310 | return true; | |
311 | #endif | |
312 | } | |
313 | ||
314 | static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) | |
315 | { | |
316 | struct file_region *nrg = NULL, *prg = NULL; | |
317 | ||
318 | prg = list_prev_entry(rg, link); | |
319 | if (&prg->link != &resv->regions && prg->to == rg->from && | |
320 | has_same_uncharge_info(prg, rg)) { | |
321 | prg->to = rg->to; | |
322 | ||
323 | list_del(&rg->link); | |
324 | kfree(rg); | |
325 | ||
326 | rg = prg; | |
327 | } | |
328 | ||
329 | nrg = list_next_entry(rg, link); | |
330 | if (&nrg->link != &resv->regions && nrg->from == rg->to && | |
331 | has_same_uncharge_info(nrg, rg)) { | |
332 | nrg->from = rg->from; | |
333 | ||
334 | list_del(&rg->link); | |
335 | kfree(rg); | |
336 | } | |
337 | } | |
338 | ||
339 | /* | |
340 | * Must be called with resv->lock held. | |
341 | * | |
342 | * Calling this with regions_needed != NULL will count the number of pages | |
343 | * to be added but will not modify the linked list. And regions_needed will | |
344 | * indicate the number of file_regions needed in the cache to carry out to add | |
345 | * the regions for this range. | |
346 | */ | |
347 | static long add_reservation_in_range(struct resv_map *resv, long f, long t, | |
348 | struct hugetlb_cgroup *h_cg, | |
349 | struct hstate *h, long *regions_needed) | |
350 | { | |
351 | long add = 0; | |
352 | struct list_head *head = &resv->regions; | |
353 | long last_accounted_offset = f; | |
354 | struct file_region *rg = NULL, *trg = NULL, *nrg = NULL; | |
355 | ||
356 | if (regions_needed) | |
357 | *regions_needed = 0; | |
358 | ||
359 | /* In this loop, we essentially handle an entry for the range | |
360 | * [last_accounted_offset, rg->from), at every iteration, with some | |
361 | * bounds checking. | |
362 | */ | |
363 | list_for_each_entry_safe(rg, trg, head, link) { | |
364 | /* Skip irrelevant regions that start before our range. */ | |
365 | if (rg->from < f) { | |
366 | /* If this region ends after the last accounted offset, | |
367 | * then we need to update last_accounted_offset. | |
368 | */ | |
369 | if (rg->to > last_accounted_offset) | |
370 | last_accounted_offset = rg->to; | |
371 | continue; | |
372 | } | |
373 | ||
374 | /* When we find a region that starts beyond our range, we've | |
375 | * finished. | |
376 | */ | |
377 | if (rg->from > t) | |
378 | break; | |
379 | ||
380 | /* Add an entry for last_accounted_offset -> rg->from, and | |
381 | * update last_accounted_offset. | |
382 | */ | |
383 | if (rg->from > last_accounted_offset) { | |
384 | add += rg->from - last_accounted_offset; | |
385 | if (!regions_needed) { | |
386 | nrg = get_file_region_entry_from_cache( | |
387 | resv, last_accounted_offset, rg->from); | |
388 | record_hugetlb_cgroup_uncharge_info(h_cg, h, | |
389 | resv, nrg); | |
390 | list_add(&nrg->link, rg->link.prev); | |
391 | coalesce_file_region(resv, nrg); | |
392 | } else | |
393 | *regions_needed += 1; | |
394 | } | |
395 | ||
396 | last_accounted_offset = rg->to; | |
397 | } | |
398 | ||
399 | /* Handle the case where our range extends beyond | |
400 | * last_accounted_offset. | |
401 | */ | |
402 | if (last_accounted_offset < t) { | |
403 | add += t - last_accounted_offset; | |
404 | if (!regions_needed) { | |
405 | nrg = get_file_region_entry_from_cache( | |
406 | resv, last_accounted_offset, t); | |
407 | record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg); | |
408 | list_add(&nrg->link, rg->link.prev); | |
409 | coalesce_file_region(resv, nrg); | |
410 | } else | |
411 | *regions_needed += 1; | |
412 | } | |
413 | ||
414 | VM_BUG_ON(add < 0); | |
415 | return add; | |
416 | } | |
417 | ||
418 | /* Must be called with resv->lock acquired. Will drop lock to allocate entries. | |
419 | */ | |
420 | static int allocate_file_region_entries(struct resv_map *resv, | |
421 | int regions_needed) | |
422 | __must_hold(&resv->lock) | |
423 | { | |
424 | struct list_head allocated_regions; | |
425 | int to_allocate = 0, i = 0; | |
426 | struct file_region *trg = NULL, *rg = NULL; | |
427 | ||
428 | VM_BUG_ON(regions_needed < 0); | |
429 | ||
430 | INIT_LIST_HEAD(&allocated_regions); | |
431 | ||
432 | /* | |
433 | * Check for sufficient descriptors in the cache to accommodate | |
434 | * the number of in progress add operations plus regions_needed. | |
435 | * | |
436 | * This is a while loop because when we drop the lock, some other call | |
437 | * to region_add or region_del may have consumed some region_entries, | |
438 | * so we keep looping here until we finally have enough entries for | |
439 | * (adds_in_progress + regions_needed). | |
440 | */ | |
441 | while (resv->region_cache_count < | |
442 | (resv->adds_in_progress + regions_needed)) { | |
443 | to_allocate = resv->adds_in_progress + regions_needed - | |
444 | resv->region_cache_count; | |
445 | ||
446 | /* At this point, we should have enough entries in the cache | |
447 | * for all the existings adds_in_progress. We should only be | |
448 | * needing to allocate for regions_needed. | |
449 | */ | |
450 | VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); | |
451 | ||
452 | spin_unlock(&resv->lock); | |
453 | for (i = 0; i < to_allocate; i++) { | |
454 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
455 | if (!trg) | |
456 | goto out_of_memory; | |
457 | list_add(&trg->link, &allocated_regions); | |
458 | } | |
459 | ||
460 | spin_lock(&resv->lock); | |
461 | ||
462 | list_splice(&allocated_regions, &resv->region_cache); | |
463 | resv->region_cache_count += to_allocate; | |
464 | } | |
465 | ||
466 | return 0; | |
467 | ||
468 | out_of_memory: | |
469 | list_for_each_entry_safe(rg, trg, &allocated_regions, link) { | |
470 | list_del(&rg->link); | |
471 | kfree(rg); | |
472 | } | |
473 | return -ENOMEM; | |
474 | } | |
475 | ||
476 | /* | |
477 | * Add the huge page range represented by [f, t) to the reserve | |
478 | * map. Regions will be taken from the cache to fill in this range. | |
479 | * Sufficient regions should exist in the cache due to the previous | |
480 | * call to region_chg with the same range, but in some cases the cache will not | |
481 | * have sufficient entries due to races with other code doing region_add or | |
482 | * region_del. The extra needed entries will be allocated. | |
483 | * | |
484 | * regions_needed is the out value provided by a previous call to region_chg. | |
485 | * | |
486 | * Return the number of new huge pages added to the map. This number is greater | |
487 | * than or equal to zero. If file_region entries needed to be allocated for | |
488 | * this operation and we were not able to allocate, it returns -ENOMEM. | |
489 | * region_add of regions of length 1 never allocate file_regions and cannot | |
490 | * fail; region_chg will always allocate at least 1 entry and a region_add for | |
491 | * 1 page will only require at most 1 entry. | |
492 | */ | |
493 | static long region_add(struct resv_map *resv, long f, long t, | |
494 | long in_regions_needed, struct hstate *h, | |
495 | struct hugetlb_cgroup *h_cg) | |
496 | { | |
497 | long add = 0, actual_regions_needed = 0; | |
498 | ||
499 | spin_lock(&resv->lock); | |
500 | retry: | |
501 | ||
502 | /* Count how many regions are actually needed to execute this add. */ | |
503 | add_reservation_in_range(resv, f, t, NULL, NULL, | |
504 | &actual_regions_needed); | |
505 | ||
506 | /* | |
507 | * Check for sufficient descriptors in the cache to accommodate | |
508 | * this add operation. Note that actual_regions_needed may be greater | |
509 | * than in_regions_needed, as the resv_map may have been modified since | |
510 | * the region_chg call. In this case, we need to make sure that we | |
511 | * allocate extra entries, such that we have enough for all the | |
512 | * existing adds_in_progress, plus the excess needed for this | |
513 | * operation. | |
514 | */ | |
515 | if (actual_regions_needed > in_regions_needed && | |
516 | resv->region_cache_count < | |
517 | resv->adds_in_progress + | |
518 | (actual_regions_needed - in_regions_needed)) { | |
519 | /* region_add operation of range 1 should never need to | |
520 | * allocate file_region entries. | |
521 | */ | |
522 | VM_BUG_ON(t - f <= 1); | |
523 | ||
524 | if (allocate_file_region_entries( | |
525 | resv, actual_regions_needed - in_regions_needed)) { | |
526 | return -ENOMEM; | |
527 | } | |
528 | ||
529 | goto retry; | |
530 | } | |
531 | ||
532 | add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); | |
533 | ||
534 | resv->adds_in_progress -= in_regions_needed; | |
535 | ||
536 | spin_unlock(&resv->lock); | |
537 | VM_BUG_ON(add < 0); | |
538 | return add; | |
539 | } | |
540 | ||
541 | /* | |
542 | * Examine the existing reserve map and determine how many | |
543 | * huge pages in the specified range [f, t) are NOT currently | |
544 | * represented. This routine is called before a subsequent | |
545 | * call to region_add that will actually modify the reserve | |
546 | * map to add the specified range [f, t). region_chg does | |
547 | * not change the number of huge pages represented by the | |
548 | * map. A number of new file_region structures is added to the cache as a | |
549 | * placeholder, for the subsequent region_add call to use. At least 1 | |
550 | * file_region structure is added. | |
551 | * | |
552 | * out_regions_needed is the number of regions added to the | |
553 | * resv->adds_in_progress. This value needs to be provided to a follow up call | |
554 | * to region_add or region_abort for proper accounting. | |
555 | * | |
556 | * Returns the number of huge pages that need to be added to the existing | |
557 | * reservation map for the range [f, t). This number is greater or equal to | |
558 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
559 | * is needed and can not be allocated. | |
560 | */ | |
561 | static long region_chg(struct resv_map *resv, long f, long t, | |
562 | long *out_regions_needed) | |
563 | { | |
564 | long chg = 0; | |
565 | ||
566 | spin_lock(&resv->lock); | |
567 | ||
568 | /* Count how many hugepages in this range are NOT represented. */ | |
569 | chg = add_reservation_in_range(resv, f, t, NULL, NULL, | |
570 | out_regions_needed); | |
571 | ||
572 | if (*out_regions_needed == 0) | |
573 | *out_regions_needed = 1; | |
574 | ||
575 | if (allocate_file_region_entries(resv, *out_regions_needed)) | |
576 | return -ENOMEM; | |
577 | ||
578 | resv->adds_in_progress += *out_regions_needed; | |
579 | ||
580 | spin_unlock(&resv->lock); | |
581 | return chg; | |
582 | } | |
583 | ||
584 | /* | |
585 | * Abort the in progress add operation. The adds_in_progress field | |
586 | * of the resv_map keeps track of the operations in progress between | |
587 | * calls to region_chg and region_add. Operations are sometimes | |
588 | * aborted after the call to region_chg. In such cases, region_abort | |
589 | * is called to decrement the adds_in_progress counter. regions_needed | |
590 | * is the value returned by the region_chg call, it is used to decrement | |
591 | * the adds_in_progress counter. | |
592 | * | |
593 | * NOTE: The range arguments [f, t) are not needed or used in this | |
594 | * routine. They are kept to make reading the calling code easier as | |
595 | * arguments will match the associated region_chg call. | |
596 | */ | |
597 | static void region_abort(struct resv_map *resv, long f, long t, | |
598 | long regions_needed) | |
599 | { | |
600 | spin_lock(&resv->lock); | |
601 | VM_BUG_ON(!resv->region_cache_count); | |
602 | resv->adds_in_progress -= regions_needed; | |
603 | spin_unlock(&resv->lock); | |
604 | } | |
605 | ||
606 | /* | |
607 | * Delete the specified range [f, t) from the reserve map. If the | |
608 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
609 | * should be deleted. Locate the regions which intersect [f, t) | |
610 | * and either trim, delete or split the existing regions. | |
611 | * | |
612 | * Returns the number of huge pages deleted from the reserve map. | |
613 | * In the normal case, the return value is zero or more. In the | |
614 | * case where a region must be split, a new region descriptor must | |
615 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
616 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
617 | * a region and possibly return -ENOMEM. Callers specifying | |
618 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
619 | */ | |
620 | static long region_del(struct resv_map *resv, long f, long t) | |
621 | { | |
622 | struct list_head *head = &resv->regions; | |
623 | struct file_region *rg, *trg; | |
624 | struct file_region *nrg = NULL; | |
625 | long del = 0; | |
626 | ||
627 | retry: | |
628 | spin_lock(&resv->lock); | |
629 | list_for_each_entry_safe(rg, trg, head, link) { | |
630 | /* | |
631 | * Skip regions before the range to be deleted. file_region | |
632 | * ranges are normally of the form [from, to). However, there | |
633 | * may be a "placeholder" entry in the map which is of the form | |
634 | * (from, to) with from == to. Check for placeholder entries | |
635 | * at the beginning of the range to be deleted. | |
636 | */ | |
637 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
638 | continue; | |
639 | ||
640 | if (rg->from >= t) | |
641 | break; | |
642 | ||
643 | if (f > rg->from && t < rg->to) { /* Must split region */ | |
644 | /* | |
645 | * Check for an entry in the cache before dropping | |
646 | * lock and attempting allocation. | |
647 | */ | |
648 | if (!nrg && | |
649 | resv->region_cache_count > resv->adds_in_progress) { | |
650 | nrg = list_first_entry(&resv->region_cache, | |
651 | struct file_region, | |
652 | link); | |
653 | list_del(&nrg->link); | |
654 | resv->region_cache_count--; | |
655 | } | |
656 | ||
657 | if (!nrg) { | |
658 | spin_unlock(&resv->lock); | |
659 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
660 | if (!nrg) | |
661 | return -ENOMEM; | |
662 | goto retry; | |
663 | } | |
664 | ||
665 | del += t - f; | |
666 | hugetlb_cgroup_uncharge_file_region( | |
667 | resv, rg, t - f); | |
668 | ||
669 | /* New entry for end of split region */ | |
670 | nrg->from = t; | |
671 | nrg->to = rg->to; | |
672 | ||
673 | copy_hugetlb_cgroup_uncharge_info(nrg, rg); | |
674 | ||
675 | INIT_LIST_HEAD(&nrg->link); | |
676 | ||
677 | /* Original entry is trimmed */ | |
678 | rg->to = f; | |
679 | ||
680 | list_add(&nrg->link, &rg->link); | |
681 | nrg = NULL; | |
682 | break; | |
683 | } | |
684 | ||
685 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
686 | del += rg->to - rg->from; | |
687 | hugetlb_cgroup_uncharge_file_region(resv, rg, | |
688 | rg->to - rg->from); | |
689 | list_del(&rg->link); | |
690 | kfree(rg); | |
691 | continue; | |
692 | } | |
693 | ||
694 | if (f <= rg->from) { /* Trim beginning of region */ | |
695 | hugetlb_cgroup_uncharge_file_region(resv, rg, | |
696 | t - rg->from); | |
697 | ||
698 | del += t - rg->from; | |
699 | rg->from = t; | |
700 | } else { /* Trim end of region */ | |
701 | hugetlb_cgroup_uncharge_file_region(resv, rg, | |
702 | rg->to - f); | |
703 | ||
704 | del += rg->to - f; | |
705 | rg->to = f; | |
706 | } | |
707 | } | |
708 | ||
709 | spin_unlock(&resv->lock); | |
710 | kfree(nrg); | |
711 | return del; | |
712 | } | |
713 | ||
714 | /* | |
715 | * A rare out of memory error was encountered which prevented removal of | |
716 | * the reserve map region for a page. The huge page itself was free'ed | |
717 | * and removed from the page cache. This routine will adjust the subpool | |
718 | * usage count, and the global reserve count if needed. By incrementing | |
719 | * these counts, the reserve map entry which could not be deleted will | |
720 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
721 | * counts. | |
722 | */ | |
723 | void hugetlb_fix_reserve_counts(struct inode *inode) | |
724 | { | |
725 | struct hugepage_subpool *spool = subpool_inode(inode); | |
726 | long rsv_adjust; | |
727 | ||
728 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
729 | if (rsv_adjust) { | |
730 | struct hstate *h = hstate_inode(inode); | |
731 | ||
732 | hugetlb_acct_memory(h, 1); | |
733 | } | |
734 | } | |
735 | ||
736 | /* | |
737 | * Count and return the number of huge pages in the reserve map | |
738 | * that intersect with the range [f, t). | |
739 | */ | |
740 | static long region_count(struct resv_map *resv, long f, long t) | |
741 | { | |
742 | struct list_head *head = &resv->regions; | |
743 | struct file_region *rg; | |
744 | long chg = 0; | |
745 | ||
746 | spin_lock(&resv->lock); | |
747 | /* Locate each segment we overlap with, and count that overlap. */ | |
748 | list_for_each_entry(rg, head, link) { | |
749 | long seg_from; | |
750 | long seg_to; | |
751 | ||
752 | if (rg->to <= f) | |
753 | continue; | |
754 | if (rg->from >= t) | |
755 | break; | |
756 | ||
757 | seg_from = max(rg->from, f); | |
758 | seg_to = min(rg->to, t); | |
759 | ||
760 | chg += seg_to - seg_from; | |
761 | } | |
762 | spin_unlock(&resv->lock); | |
763 | ||
764 | return chg; | |
765 | } | |
766 | ||
767 | /* | |
768 | * Convert the address within this vma to the page offset within | |
769 | * the mapping, in pagecache page units; huge pages here. | |
770 | */ | |
771 | static pgoff_t vma_hugecache_offset(struct hstate *h, | |
772 | struct vm_area_struct *vma, unsigned long address) | |
773 | { | |
774 | return ((address - vma->vm_start) >> huge_page_shift(h)) + | |
775 | (vma->vm_pgoff >> huge_page_order(h)); | |
776 | } | |
777 | ||
778 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, | |
779 | unsigned long address) | |
780 | { | |
781 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
782 | } | |
783 | EXPORT_SYMBOL_GPL(linear_hugepage_index); | |
784 | ||
785 | /* | |
786 | * Return the size of the pages allocated when backing a VMA. In the majority | |
787 | * cases this will be same size as used by the page table entries. | |
788 | */ | |
789 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
790 | { | |
791 | if (vma->vm_ops && vma->vm_ops->pagesize) | |
792 | return vma->vm_ops->pagesize(vma); | |
793 | return PAGE_SIZE; | |
794 | } | |
795 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); | |
796 | ||
797 | /* | |
798 | * Return the page size being used by the MMU to back a VMA. In the majority | |
799 | * of cases, the page size used by the kernel matches the MMU size. On | |
800 | * architectures where it differs, an architecture-specific 'strong' | |
801 | * version of this symbol is required. | |
802 | */ | |
803 | __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) | |
804 | { | |
805 | return vma_kernel_pagesize(vma); | |
806 | } | |
807 | ||
808 | /* | |
809 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
810 | * bits of the reservation map pointer, which are always clear due to | |
811 | * alignment. | |
812 | */ | |
813 | #define HPAGE_RESV_OWNER (1UL << 0) | |
814 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
815 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) | |
816 | ||
817 | /* | |
818 | * These helpers are used to track how many pages are reserved for | |
819 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
820 | * is guaranteed to have their future faults succeed. | |
821 | * | |
822 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
823 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
824 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
825 | * chance of the global counters getting corrupted as a result of the values. | |
826 | * | |
827 | * The private mapping reservation is represented in a subtly different | |
828 | * manner to a shared mapping. A shared mapping has a region map associated | |
829 | * with the underlying file, this region map represents the backing file | |
830 | * pages which have ever had a reservation assigned which this persists even | |
831 | * after the page is instantiated. A private mapping has a region map | |
832 | * associated with the original mmap which is attached to all VMAs which | |
833 | * reference it, this region map represents those offsets which have consumed | |
834 | * reservation ie. where pages have been instantiated. | |
835 | */ | |
836 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) | |
837 | { | |
838 | return (unsigned long)vma->vm_private_data; | |
839 | } | |
840 | ||
841 | static void set_vma_private_data(struct vm_area_struct *vma, | |
842 | unsigned long value) | |
843 | { | |
844 | vma->vm_private_data = (void *)value; | |
845 | } | |
846 | ||
847 | static void | |
848 | resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, | |
849 | struct hugetlb_cgroup *h_cg, | |
850 | struct hstate *h) | |
851 | { | |
852 | #ifdef CONFIG_CGROUP_HUGETLB | |
853 | if (!h_cg || !h) { | |
854 | resv_map->reservation_counter = NULL; | |
855 | resv_map->pages_per_hpage = 0; | |
856 | resv_map->css = NULL; | |
857 | } else { | |
858 | resv_map->reservation_counter = | |
859 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
860 | resv_map->pages_per_hpage = pages_per_huge_page(h); | |
861 | resv_map->css = &h_cg->css; | |
862 | } | |
863 | #endif | |
864 | } | |
865 | ||
866 | struct resv_map *resv_map_alloc(void) | |
867 | { | |
868 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
869 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); | |
870 | ||
871 | if (!resv_map || !rg) { | |
872 | kfree(resv_map); | |
873 | kfree(rg); | |
874 | return NULL; | |
875 | } | |
876 | ||
877 | kref_init(&resv_map->refs); | |
878 | spin_lock_init(&resv_map->lock); | |
879 | INIT_LIST_HEAD(&resv_map->regions); | |
880 | ||
881 | resv_map->adds_in_progress = 0; | |
882 | /* | |
883 | * Initialize these to 0. On shared mappings, 0's here indicate these | |
884 | * fields don't do cgroup accounting. On private mappings, these will be | |
885 | * re-initialized to the proper values, to indicate that hugetlb cgroup | |
886 | * reservations are to be un-charged from here. | |
887 | */ | |
888 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); | |
889 | ||
890 | INIT_LIST_HEAD(&resv_map->region_cache); | |
891 | list_add(&rg->link, &resv_map->region_cache); | |
892 | resv_map->region_cache_count = 1; | |
893 | ||
894 | return resv_map; | |
895 | } | |
896 | ||
897 | void resv_map_release(struct kref *ref) | |
898 | { | |
899 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
900 | struct list_head *head = &resv_map->region_cache; | |
901 | struct file_region *rg, *trg; | |
902 | ||
903 | /* Clear out any active regions before we release the map. */ | |
904 | region_del(resv_map, 0, LONG_MAX); | |
905 | ||
906 | /* ... and any entries left in the cache */ | |
907 | list_for_each_entry_safe(rg, trg, head, link) { | |
908 | list_del(&rg->link); | |
909 | kfree(rg); | |
910 | } | |
911 | ||
912 | VM_BUG_ON(resv_map->adds_in_progress); | |
913 | ||
914 | kfree(resv_map); | |
915 | } | |
916 | ||
917 | static inline struct resv_map *inode_resv_map(struct inode *inode) | |
918 | { | |
919 | /* | |
920 | * At inode evict time, i_mapping may not point to the original | |
921 | * address space within the inode. This original address space | |
922 | * contains the pointer to the resv_map. So, always use the | |
923 | * address space embedded within the inode. | |
924 | * The VERY common case is inode->mapping == &inode->i_data but, | |
925 | * this may not be true for device special inodes. | |
926 | */ | |
927 | return (struct resv_map *)(&inode->i_data)->private_data; | |
928 | } | |
929 | ||
930 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) | |
931 | { | |
932 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); | |
933 | if (vma->vm_flags & VM_MAYSHARE) { | |
934 | struct address_space *mapping = vma->vm_file->f_mapping; | |
935 | struct inode *inode = mapping->host; | |
936 | ||
937 | return inode_resv_map(inode); | |
938 | ||
939 | } else { | |
940 | return (struct resv_map *)(get_vma_private_data(vma) & | |
941 | ~HPAGE_RESV_MASK); | |
942 | } | |
943 | } | |
944 | ||
945 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) | |
946 | { | |
947 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); | |
948 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
949 | ||
950 | set_vma_private_data(vma, (get_vma_private_data(vma) & | |
951 | HPAGE_RESV_MASK) | (unsigned long)map); | |
952 | } | |
953 | ||
954 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
955 | { | |
956 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); | |
957 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
958 | ||
959 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
960 | } | |
961 | ||
962 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
963 | { | |
964 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); | |
965 | ||
966 | return (get_vma_private_data(vma) & flag) != 0; | |
967 | } | |
968 | ||
969 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ | |
970 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) | |
971 | { | |
972 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); | |
973 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
974 | vma->vm_private_data = (void *)0; | |
975 | } | |
976 | ||
977 | /* Returns true if the VMA has associated reserve pages */ | |
978 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) | |
979 | { | |
980 | if (vma->vm_flags & VM_NORESERVE) { | |
981 | /* | |
982 | * This address is already reserved by other process(chg == 0), | |
983 | * so, we should decrement reserved count. Without decrementing, | |
984 | * reserve count remains after releasing inode, because this | |
985 | * allocated page will go into page cache and is regarded as | |
986 | * coming from reserved pool in releasing step. Currently, we | |
987 | * don't have any other solution to deal with this situation | |
988 | * properly, so add work-around here. | |
989 | */ | |
990 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
991 | return true; | |
992 | else | |
993 | return false; | |
994 | } | |
995 | ||
996 | /* Shared mappings always use reserves */ | |
997 | if (vma->vm_flags & VM_MAYSHARE) { | |
998 | /* | |
999 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
1000 | * be a region map for all pages. The only situation where | |
1001 | * there is no region map is if a hole was punched via | |
1002 | * fallocate. In this case, there really are no reserves to | |
1003 | * use. This situation is indicated if chg != 0. | |
1004 | */ | |
1005 | if (chg) | |
1006 | return false; | |
1007 | else | |
1008 | return true; | |
1009 | } | |
1010 | ||
1011 | /* | |
1012 | * Only the process that called mmap() has reserves for | |
1013 | * private mappings. | |
1014 | */ | |
1015 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { | |
1016 | /* | |
1017 | * Like the shared case above, a hole punch or truncate | |
1018 | * could have been performed on the private mapping. | |
1019 | * Examine the value of chg to determine if reserves | |
1020 | * actually exist or were previously consumed. | |
1021 | * Very Subtle - The value of chg comes from a previous | |
1022 | * call to vma_needs_reserves(). The reserve map for | |
1023 | * private mappings has different (opposite) semantics | |
1024 | * than that of shared mappings. vma_needs_reserves() | |
1025 | * has already taken this difference in semantics into | |
1026 | * account. Therefore, the meaning of chg is the same | |
1027 | * as in the shared case above. Code could easily be | |
1028 | * combined, but keeping it separate draws attention to | |
1029 | * subtle differences. | |
1030 | */ | |
1031 | if (chg) | |
1032 | return false; | |
1033 | else | |
1034 | return true; | |
1035 | } | |
1036 | ||
1037 | return false; | |
1038 | } | |
1039 | ||
1040 | static void enqueue_huge_page(struct hstate *h, struct page *page) | |
1041 | { | |
1042 | int nid = page_to_nid(page); | |
1043 | list_move(&page->lru, &h->hugepage_freelists[nid]); | |
1044 | h->free_huge_pages++; | |
1045 | h->free_huge_pages_node[nid]++; | |
1046 | SetPageHugeFreed(page); | |
1047 | } | |
1048 | ||
1049 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) | |
1050 | { | |
1051 | struct page *page; | |
1052 | bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA); | |
1053 | ||
1054 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { | |
1055 | if (nocma && is_migrate_cma_page(page)) | |
1056 | continue; | |
1057 | ||
1058 | if (PageHWPoison(page)) | |
1059 | continue; | |
1060 | ||
1061 | list_move(&page->lru, &h->hugepage_activelist); | |
1062 | set_page_refcounted(page); | |
1063 | ClearPageHugeFreed(page); | |
1064 | h->free_huge_pages--; | |
1065 | h->free_huge_pages_node[nid]--; | |
1066 | return page; | |
1067 | } | |
1068 | ||
1069 | return NULL; | |
1070 | } | |
1071 | ||
1072 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, | |
1073 | nodemask_t *nmask) | |
1074 | { | |
1075 | unsigned int cpuset_mems_cookie; | |
1076 | struct zonelist *zonelist; | |
1077 | struct zone *zone; | |
1078 | struct zoneref *z; | |
1079 | int node = NUMA_NO_NODE; | |
1080 | ||
1081 | zonelist = node_zonelist(nid, gfp_mask); | |
1082 | ||
1083 | retry_cpuset: | |
1084 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
1085 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
1086 | struct page *page; | |
1087 | ||
1088 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
1089 | continue; | |
1090 | /* | |
1091 | * no need to ask again on the same node. Pool is node rather than | |
1092 | * zone aware | |
1093 | */ | |
1094 | if (zone_to_nid(zone) == node) | |
1095 | continue; | |
1096 | node = zone_to_nid(zone); | |
1097 | ||
1098 | page = dequeue_huge_page_node_exact(h, node); | |
1099 | if (page) | |
1100 | return page; | |
1101 | } | |
1102 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) | |
1103 | goto retry_cpuset; | |
1104 | ||
1105 | return NULL; | |
1106 | } | |
1107 | ||
1108 | static struct page *dequeue_huge_page_vma(struct hstate *h, | |
1109 | struct vm_area_struct *vma, | |
1110 | unsigned long address, int avoid_reserve, | |
1111 | long chg) | |
1112 | { | |
1113 | struct page *page; | |
1114 | struct mempolicy *mpol; | |
1115 | gfp_t gfp_mask; | |
1116 | nodemask_t *nodemask; | |
1117 | int nid; | |
1118 | ||
1119 | /* | |
1120 | * A child process with MAP_PRIVATE mappings created by their parent | |
1121 | * have no page reserves. This check ensures that reservations are | |
1122 | * not "stolen". The child may still get SIGKILLed | |
1123 | */ | |
1124 | if (!vma_has_reserves(vma, chg) && | |
1125 | h->free_huge_pages - h->resv_huge_pages == 0) | |
1126 | goto err; | |
1127 | ||
1128 | /* If reserves cannot be used, ensure enough pages are in the pool */ | |
1129 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) | |
1130 | goto err; | |
1131 | ||
1132 | gfp_mask = htlb_alloc_mask(h); | |
1133 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
1134 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); | |
1135 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { | |
1136 | SetPagePrivate(page); | |
1137 | h->resv_huge_pages--; | |
1138 | } | |
1139 | ||
1140 | mpol_cond_put(mpol); | |
1141 | return page; | |
1142 | ||
1143 | err: | |
1144 | return NULL; | |
1145 | } | |
1146 | ||
1147 | /* | |
1148 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
1149 | * We may have allocated or freed a huge page based on a different | |
1150 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
1151 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
1152 | * node for alloc or free. | |
1153 | */ | |
1154 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1155 | { | |
1156 | nid = next_node_in(nid, *nodes_allowed); | |
1157 | VM_BUG_ON(nid >= MAX_NUMNODES); | |
1158 | ||
1159 | return nid; | |
1160 | } | |
1161 | ||
1162 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1163 | { | |
1164 | if (!node_isset(nid, *nodes_allowed)) | |
1165 | nid = next_node_allowed(nid, nodes_allowed); | |
1166 | return nid; | |
1167 | } | |
1168 | ||
1169 | /* | |
1170 | * returns the previously saved node ["this node"] from which to | |
1171 | * allocate a persistent huge page for the pool and advance the | |
1172 | * next node from which to allocate, handling wrap at end of node | |
1173 | * mask. | |
1174 | */ | |
1175 | static int hstate_next_node_to_alloc(struct hstate *h, | |
1176 | nodemask_t *nodes_allowed) | |
1177 | { | |
1178 | int nid; | |
1179 | ||
1180 | VM_BUG_ON(!nodes_allowed); | |
1181 | ||
1182 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1183 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1184 | ||
1185 | return nid; | |
1186 | } | |
1187 | ||
1188 | /* | |
1189 | * helper for free_pool_huge_page() - return the previously saved | |
1190 | * node ["this node"] from which to free a huge page. Advance the | |
1191 | * next node id whether or not we find a free huge page to free so | |
1192 | * that the next attempt to free addresses the next node. | |
1193 | */ | |
1194 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1195 | { | |
1196 | int nid; | |
1197 | ||
1198 | VM_BUG_ON(!nodes_allowed); | |
1199 | ||
1200 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1201 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1202 | ||
1203 | return nid; | |
1204 | } | |
1205 | ||
1206 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1207 | for (nr_nodes = nodes_weight(*mask); \ | |
1208 | nr_nodes > 0 && \ | |
1209 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1210 | nr_nodes--) | |
1211 | ||
1212 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1213 | for (nr_nodes = nodes_weight(*mask); \ | |
1214 | nr_nodes > 0 && \ | |
1215 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1216 | nr_nodes--) | |
1217 | ||
1218 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE | |
1219 | static void destroy_compound_gigantic_page(struct page *page, | |
1220 | unsigned int order) | |
1221 | { | |
1222 | int i; | |
1223 | int nr_pages = 1 << order; | |
1224 | struct page *p = page + 1; | |
1225 | ||
1226 | atomic_set(compound_mapcount_ptr(page), 0); | |
1227 | if (hpage_pincount_available(page)) | |
1228 | atomic_set(compound_pincount_ptr(page), 0); | |
1229 | ||
1230 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { | |
1231 | clear_compound_head(p); | |
1232 | set_page_refcounted(p); | |
1233 | } | |
1234 | ||
1235 | set_compound_order(page, 0); | |
1236 | page[1].compound_nr = 0; | |
1237 | __ClearPageHead(page); | |
1238 | } | |
1239 | ||
1240 | static void free_gigantic_page(struct page *page, unsigned int order) | |
1241 | { | |
1242 | /* | |
1243 | * If the page isn't allocated using the cma allocator, | |
1244 | * cma_release() returns false. | |
1245 | */ | |
1246 | #ifdef CONFIG_CMA | |
1247 | if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) | |
1248 | return; | |
1249 | #endif | |
1250 | ||
1251 | free_contig_range(page_to_pfn(page), 1 << order); | |
1252 | } | |
1253 | ||
1254 | #ifdef CONFIG_CONTIG_ALLOC | |
1255 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1256 | int nid, nodemask_t *nodemask) | |
1257 | { | |
1258 | unsigned long nr_pages = 1UL << huge_page_order(h); | |
1259 | if (nid == NUMA_NO_NODE) | |
1260 | nid = numa_mem_id(); | |
1261 | ||
1262 | #ifdef CONFIG_CMA | |
1263 | { | |
1264 | struct page *page; | |
1265 | int node; | |
1266 | ||
1267 | if (hugetlb_cma[nid]) { | |
1268 | page = cma_alloc(hugetlb_cma[nid], nr_pages, | |
1269 | huge_page_order(h), true); | |
1270 | if (page) | |
1271 | return page; | |
1272 | } | |
1273 | ||
1274 | if (!(gfp_mask & __GFP_THISNODE)) { | |
1275 | for_each_node_mask(node, *nodemask) { | |
1276 | if (node == nid || !hugetlb_cma[node]) | |
1277 | continue; | |
1278 | ||
1279 | page = cma_alloc(hugetlb_cma[node], nr_pages, | |
1280 | huge_page_order(h), true); | |
1281 | if (page) | |
1282 | return page; | |
1283 | } | |
1284 | } | |
1285 | } | |
1286 | #endif | |
1287 | ||
1288 | return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); | |
1289 | } | |
1290 | ||
1291 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
1292 | static void prep_compound_gigantic_page(struct page *page, unsigned int order); | |
1293 | #else /* !CONFIG_CONTIG_ALLOC */ | |
1294 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1295 | int nid, nodemask_t *nodemask) | |
1296 | { | |
1297 | return NULL; | |
1298 | } | |
1299 | #endif /* CONFIG_CONTIG_ALLOC */ | |
1300 | ||
1301 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ | |
1302 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1303 | int nid, nodemask_t *nodemask) | |
1304 | { | |
1305 | return NULL; | |
1306 | } | |
1307 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } | |
1308 | static inline void destroy_compound_gigantic_page(struct page *page, | |
1309 | unsigned int order) { } | |
1310 | #endif | |
1311 | ||
1312 | static void update_and_free_page(struct hstate *h, struct page *page) | |
1313 | { | |
1314 | int i; | |
1315 | ||
1316 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) | |
1317 | return; | |
1318 | ||
1319 | h->nr_huge_pages--; | |
1320 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
1321 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
1322 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | | |
1323 | 1 << PG_referenced | 1 << PG_dirty | | |
1324 | 1 << PG_active | 1 << PG_private | | |
1325 | 1 << PG_writeback); | |
1326 | } | |
1327 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); | |
1328 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); | |
1329 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); | |
1330 | set_page_refcounted(page); | |
1331 | if (hstate_is_gigantic(h)) { | |
1332 | /* | |
1333 | * Temporarily drop the hugetlb_lock, because | |
1334 | * we might block in free_gigantic_page(). | |
1335 | */ | |
1336 | spin_unlock(&hugetlb_lock); | |
1337 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
1338 | free_gigantic_page(page, huge_page_order(h)); | |
1339 | spin_lock(&hugetlb_lock); | |
1340 | } else { | |
1341 | __free_pages(page, huge_page_order(h)); | |
1342 | } | |
1343 | } | |
1344 | ||
1345 | struct hstate *size_to_hstate(unsigned long size) | |
1346 | { | |
1347 | struct hstate *h; | |
1348 | ||
1349 | for_each_hstate(h) { | |
1350 | if (huge_page_size(h) == size) | |
1351 | return h; | |
1352 | } | |
1353 | return NULL; | |
1354 | } | |
1355 | ||
1356 | /* | |
1357 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
1358 | * to hstate->hugepage_activelist.) | |
1359 | * | |
1360 | * This function can be called for tail pages, but never returns true for them. | |
1361 | */ | |
1362 | bool page_huge_active(struct page *page) | |
1363 | { | |
1364 | VM_BUG_ON_PAGE(!PageHuge(page), page); | |
1365 | return PageHead(page) && PagePrivate(&page[1]); | |
1366 | } | |
1367 | ||
1368 | /* never called for tail page */ | |
1369 | void set_page_huge_active(struct page *page) | |
1370 | { | |
1371 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1372 | SetPagePrivate(&page[1]); | |
1373 | } | |
1374 | ||
1375 | static void clear_page_huge_active(struct page *page) | |
1376 | { | |
1377 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1378 | ClearPagePrivate(&page[1]); | |
1379 | } | |
1380 | ||
1381 | /* | |
1382 | * Internal hugetlb specific page flag. Do not use outside of the hugetlb | |
1383 | * code | |
1384 | */ | |
1385 | static inline bool PageHugeTemporary(struct page *page) | |
1386 | { | |
1387 | if (!PageHuge(page)) | |
1388 | return false; | |
1389 | ||
1390 | return (unsigned long)page[2].mapping == -1U; | |
1391 | } | |
1392 | ||
1393 | static inline void SetPageHugeTemporary(struct page *page) | |
1394 | { | |
1395 | page[2].mapping = (void *)-1U; | |
1396 | } | |
1397 | ||
1398 | static inline void ClearPageHugeTemporary(struct page *page) | |
1399 | { | |
1400 | page[2].mapping = NULL; | |
1401 | } | |
1402 | ||
1403 | static void __free_huge_page(struct page *page) | |
1404 | { | |
1405 | /* | |
1406 | * Can't pass hstate in here because it is called from the | |
1407 | * compound page destructor. | |
1408 | */ | |
1409 | struct hstate *h = page_hstate(page); | |
1410 | int nid = page_to_nid(page); | |
1411 | struct hugepage_subpool *spool = | |
1412 | (struct hugepage_subpool *)page_private(page); | |
1413 | bool restore_reserve; | |
1414 | ||
1415 | VM_BUG_ON_PAGE(page_count(page), page); | |
1416 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
1417 | ||
1418 | set_page_private(page, 0); | |
1419 | page->mapping = NULL; | |
1420 | restore_reserve = PagePrivate(page); | |
1421 | ClearPagePrivate(page); | |
1422 | ||
1423 | /* | |
1424 | * If PagePrivate() was set on page, page allocation consumed a | |
1425 | * reservation. If the page was associated with a subpool, there | |
1426 | * would have been a page reserved in the subpool before allocation | |
1427 | * via hugepage_subpool_get_pages(). Since we are 'restoring' the | |
1428 | * reservtion, do not call hugepage_subpool_put_pages() as this will | |
1429 | * remove the reserved page from the subpool. | |
1430 | */ | |
1431 | if (!restore_reserve) { | |
1432 | /* | |
1433 | * A return code of zero implies that the subpool will be | |
1434 | * under its minimum size if the reservation is not restored | |
1435 | * after page is free. Therefore, force restore_reserve | |
1436 | * operation. | |
1437 | */ | |
1438 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1439 | restore_reserve = true; | |
1440 | } | |
1441 | ||
1442 | spin_lock(&hugetlb_lock); | |
1443 | clear_page_huge_active(page); | |
1444 | hugetlb_cgroup_uncharge_page(hstate_index(h), | |
1445 | pages_per_huge_page(h), page); | |
1446 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), | |
1447 | pages_per_huge_page(h), page); | |
1448 | if (restore_reserve) | |
1449 | h->resv_huge_pages++; | |
1450 | ||
1451 | if (PageHugeTemporary(page)) { | |
1452 | list_del(&page->lru); | |
1453 | ClearPageHugeTemporary(page); | |
1454 | update_and_free_page(h, page); | |
1455 | } else if (h->surplus_huge_pages_node[nid]) { | |
1456 | /* remove the page from active list */ | |
1457 | list_del(&page->lru); | |
1458 | update_and_free_page(h, page); | |
1459 | h->surplus_huge_pages--; | |
1460 | h->surplus_huge_pages_node[nid]--; | |
1461 | } else { | |
1462 | arch_clear_hugepage_flags(page); | |
1463 | enqueue_huge_page(h, page); | |
1464 | } | |
1465 | spin_unlock(&hugetlb_lock); | |
1466 | } | |
1467 | ||
1468 | /* | |
1469 | * As free_huge_page() can be called from a non-task context, we have | |
1470 | * to defer the actual freeing in a workqueue to prevent potential | |
1471 | * hugetlb_lock deadlock. | |
1472 | * | |
1473 | * free_hpage_workfn() locklessly retrieves the linked list of pages to | |
1474 | * be freed and frees them one-by-one. As the page->mapping pointer is | |
1475 | * going to be cleared in __free_huge_page() anyway, it is reused as the | |
1476 | * llist_node structure of a lockless linked list of huge pages to be freed. | |
1477 | */ | |
1478 | static LLIST_HEAD(hpage_freelist); | |
1479 | ||
1480 | static void free_hpage_workfn(struct work_struct *work) | |
1481 | { | |
1482 | struct llist_node *node; | |
1483 | struct page *page; | |
1484 | ||
1485 | node = llist_del_all(&hpage_freelist); | |
1486 | ||
1487 | while (node) { | |
1488 | page = container_of((struct address_space **)node, | |
1489 | struct page, mapping); | |
1490 | node = node->next; | |
1491 | __free_huge_page(page); | |
1492 | } | |
1493 | } | |
1494 | static DECLARE_WORK(free_hpage_work, free_hpage_workfn); | |
1495 | ||
1496 | void free_huge_page(struct page *page) | |
1497 | { | |
1498 | /* | |
1499 | * Defer freeing if in non-task context to avoid hugetlb_lock deadlock. | |
1500 | */ | |
1501 | if (!in_task()) { | |
1502 | /* | |
1503 | * Only call schedule_work() if hpage_freelist is previously | |
1504 | * empty. Otherwise, schedule_work() had been called but the | |
1505 | * workfn hasn't retrieved the list yet. | |
1506 | */ | |
1507 | if (llist_add((struct llist_node *)&page->mapping, | |
1508 | &hpage_freelist)) | |
1509 | schedule_work(&free_hpage_work); | |
1510 | return; | |
1511 | } | |
1512 | ||
1513 | __free_huge_page(page); | |
1514 | } | |
1515 | ||
1516 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) | |
1517 | { | |
1518 | INIT_LIST_HEAD(&page->lru); | |
1519 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); | |
1520 | set_hugetlb_cgroup(page, NULL); | |
1521 | set_hugetlb_cgroup_rsvd(page, NULL); | |
1522 | spin_lock(&hugetlb_lock); | |
1523 | h->nr_huge_pages++; | |
1524 | h->nr_huge_pages_node[nid]++; | |
1525 | ClearPageHugeFreed(page); | |
1526 | spin_unlock(&hugetlb_lock); | |
1527 | } | |
1528 | ||
1529 | static void prep_compound_gigantic_page(struct page *page, unsigned int order) | |
1530 | { | |
1531 | int i; | |
1532 | int nr_pages = 1 << order; | |
1533 | struct page *p = page + 1; | |
1534 | ||
1535 | /* we rely on prep_new_huge_page to set the destructor */ | |
1536 | set_compound_order(page, order); | |
1537 | __ClearPageReserved(page); | |
1538 | __SetPageHead(page); | |
1539 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { | |
1540 | /* | |
1541 | * For gigantic hugepages allocated through bootmem at | |
1542 | * boot, it's safer to be consistent with the not-gigantic | |
1543 | * hugepages and clear the PG_reserved bit from all tail pages | |
1544 | * too. Otherwise drivers using get_user_pages() to access tail | |
1545 | * pages may get the reference counting wrong if they see | |
1546 | * PG_reserved set on a tail page (despite the head page not | |
1547 | * having PG_reserved set). Enforcing this consistency between | |
1548 | * head and tail pages allows drivers to optimize away a check | |
1549 | * on the head page when they need know if put_page() is needed | |
1550 | * after get_user_pages(). | |
1551 | */ | |
1552 | __ClearPageReserved(p); | |
1553 | set_page_count(p, 0); | |
1554 | set_compound_head(p, page); | |
1555 | } | |
1556 | atomic_set(compound_mapcount_ptr(page), -1); | |
1557 | ||
1558 | if (hpage_pincount_available(page)) | |
1559 | atomic_set(compound_pincount_ptr(page), 0); | |
1560 | } | |
1561 | ||
1562 | /* | |
1563 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1564 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1565 | * details. | |
1566 | */ | |
1567 | int PageHuge(struct page *page) | |
1568 | { | |
1569 | if (!PageCompound(page)) | |
1570 | return 0; | |
1571 | ||
1572 | page = compound_head(page); | |
1573 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; | |
1574 | } | |
1575 | EXPORT_SYMBOL_GPL(PageHuge); | |
1576 | ||
1577 | /* | |
1578 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1579 | * normal or transparent huge pages. | |
1580 | */ | |
1581 | int PageHeadHuge(struct page *page_head) | |
1582 | { | |
1583 | if (!PageHead(page_head)) | |
1584 | return 0; | |
1585 | ||
1586 | return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; | |
1587 | } | |
1588 | ||
1589 | /* | |
1590 | * Find and lock address space (mapping) in write mode. | |
1591 | * | |
1592 | * Upon entry, the page is locked which means that page_mapping() is | |
1593 | * stable. Due to locking order, we can only trylock_write. If we can | |
1594 | * not get the lock, simply return NULL to caller. | |
1595 | */ | |
1596 | struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) | |
1597 | { | |
1598 | struct address_space *mapping = page_mapping(hpage); | |
1599 | ||
1600 | if (!mapping) | |
1601 | return mapping; | |
1602 | ||
1603 | if (i_mmap_trylock_write(mapping)) | |
1604 | return mapping; | |
1605 | ||
1606 | return NULL; | |
1607 | } | |
1608 | ||
1609 | pgoff_t __basepage_index(struct page *page) | |
1610 | { | |
1611 | struct page *page_head = compound_head(page); | |
1612 | pgoff_t index = page_index(page_head); | |
1613 | unsigned long compound_idx; | |
1614 | ||
1615 | if (!PageHuge(page_head)) | |
1616 | return page_index(page); | |
1617 | ||
1618 | if (compound_order(page_head) >= MAX_ORDER) | |
1619 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1620 | else | |
1621 | compound_idx = page - page_head; | |
1622 | ||
1623 | return (index << compound_order(page_head)) + compound_idx; | |
1624 | } | |
1625 | ||
1626 | static struct page *alloc_buddy_huge_page(struct hstate *h, | |
1627 | gfp_t gfp_mask, int nid, nodemask_t *nmask, | |
1628 | nodemask_t *node_alloc_noretry) | |
1629 | { | |
1630 | int order = huge_page_order(h); | |
1631 | struct page *page; | |
1632 | bool alloc_try_hard = true; | |
1633 | ||
1634 | /* | |
1635 | * By default we always try hard to allocate the page with | |
1636 | * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in | |
1637 | * a loop (to adjust global huge page counts) and previous allocation | |
1638 | * failed, do not continue to try hard on the same node. Use the | |
1639 | * node_alloc_noretry bitmap to manage this state information. | |
1640 | */ | |
1641 | if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) | |
1642 | alloc_try_hard = false; | |
1643 | gfp_mask |= __GFP_COMP|__GFP_NOWARN; | |
1644 | if (alloc_try_hard) | |
1645 | gfp_mask |= __GFP_RETRY_MAYFAIL; | |
1646 | if (nid == NUMA_NO_NODE) | |
1647 | nid = numa_mem_id(); | |
1648 | page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); | |
1649 | if (page) | |
1650 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
1651 | else | |
1652 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
1653 | ||
1654 | /* | |
1655 | * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this | |
1656 | * indicates an overall state change. Clear bit so that we resume | |
1657 | * normal 'try hard' allocations. | |
1658 | */ | |
1659 | if (node_alloc_noretry && page && !alloc_try_hard) | |
1660 | node_clear(nid, *node_alloc_noretry); | |
1661 | ||
1662 | /* | |
1663 | * If we tried hard to get a page but failed, set bit so that | |
1664 | * subsequent attempts will not try as hard until there is an | |
1665 | * overall state change. | |
1666 | */ | |
1667 | if (node_alloc_noretry && !page && alloc_try_hard) | |
1668 | node_set(nid, *node_alloc_noretry); | |
1669 | ||
1670 | return page; | |
1671 | } | |
1672 | ||
1673 | /* | |
1674 | * Common helper to allocate a fresh hugetlb page. All specific allocators | |
1675 | * should use this function to get new hugetlb pages | |
1676 | */ | |
1677 | static struct page *alloc_fresh_huge_page(struct hstate *h, | |
1678 | gfp_t gfp_mask, int nid, nodemask_t *nmask, | |
1679 | nodemask_t *node_alloc_noretry) | |
1680 | { | |
1681 | struct page *page; | |
1682 | ||
1683 | if (hstate_is_gigantic(h)) | |
1684 | page = alloc_gigantic_page(h, gfp_mask, nid, nmask); | |
1685 | else | |
1686 | page = alloc_buddy_huge_page(h, gfp_mask, | |
1687 | nid, nmask, node_alloc_noretry); | |
1688 | if (!page) | |
1689 | return NULL; | |
1690 | ||
1691 | if (hstate_is_gigantic(h)) | |
1692 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
1693 | prep_new_huge_page(h, page, page_to_nid(page)); | |
1694 | ||
1695 | return page; | |
1696 | } | |
1697 | ||
1698 | /* | |
1699 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
1700 | * manner. | |
1701 | */ | |
1702 | static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, | |
1703 | nodemask_t *node_alloc_noretry) | |
1704 | { | |
1705 | struct page *page; | |
1706 | int nr_nodes, node; | |
1707 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; | |
1708 | ||
1709 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1710 | page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, | |
1711 | node_alloc_noretry); | |
1712 | if (page) | |
1713 | break; | |
1714 | } | |
1715 | ||
1716 | if (!page) | |
1717 | return 0; | |
1718 | ||
1719 | put_page(page); /* free it into the hugepage allocator */ | |
1720 | ||
1721 | return 1; | |
1722 | } | |
1723 | ||
1724 | /* | |
1725 | * Free huge page from pool from next node to free. | |
1726 | * Attempt to keep persistent huge pages more or less | |
1727 | * balanced over allowed nodes. | |
1728 | * Called with hugetlb_lock locked. | |
1729 | */ | |
1730 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, | |
1731 | bool acct_surplus) | |
1732 | { | |
1733 | int nr_nodes, node; | |
1734 | int ret = 0; | |
1735 | ||
1736 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
1737 | /* | |
1738 | * If we're returning unused surplus pages, only examine | |
1739 | * nodes with surplus pages. | |
1740 | */ | |
1741 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && | |
1742 | !list_empty(&h->hugepage_freelists[node])) { | |
1743 | struct page *page = | |
1744 | list_entry(h->hugepage_freelists[node].next, | |
1745 | struct page, lru); | |
1746 | list_del(&page->lru); | |
1747 | h->free_huge_pages--; | |
1748 | h->free_huge_pages_node[node]--; | |
1749 | if (acct_surplus) { | |
1750 | h->surplus_huge_pages--; | |
1751 | h->surplus_huge_pages_node[node]--; | |
1752 | } | |
1753 | update_and_free_page(h, page); | |
1754 | ret = 1; | |
1755 | break; | |
1756 | } | |
1757 | } | |
1758 | ||
1759 | return ret; | |
1760 | } | |
1761 | ||
1762 | /* | |
1763 | * Dissolve a given free hugepage into free buddy pages. This function does | |
1764 | * nothing for in-use hugepages and non-hugepages. | |
1765 | * This function returns values like below: | |
1766 | * | |
1767 | * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use | |
1768 | * (allocated or reserved.) | |
1769 | * 0: successfully dissolved free hugepages or the page is not a | |
1770 | * hugepage (considered as already dissolved) | |
1771 | */ | |
1772 | int dissolve_free_huge_page(struct page *page) | |
1773 | { | |
1774 | int rc = -EBUSY; | |
1775 | ||
1776 | retry: | |
1777 | /* Not to disrupt normal path by vainly holding hugetlb_lock */ | |
1778 | if (!PageHuge(page)) | |
1779 | return 0; | |
1780 | ||
1781 | spin_lock(&hugetlb_lock); | |
1782 | if (!PageHuge(page)) { | |
1783 | rc = 0; | |
1784 | goto out; | |
1785 | } | |
1786 | ||
1787 | if (!page_count(page)) { | |
1788 | struct page *head = compound_head(page); | |
1789 | struct hstate *h = page_hstate(head); | |
1790 | int nid = page_to_nid(head); | |
1791 | if (h->free_huge_pages - h->resv_huge_pages == 0) | |
1792 | goto out; | |
1793 | ||
1794 | /* | |
1795 | * We should make sure that the page is already on the free list | |
1796 | * when it is dissolved. | |
1797 | */ | |
1798 | if (unlikely(!PageHugeFreed(head))) { | |
1799 | spin_unlock(&hugetlb_lock); | |
1800 | cond_resched(); | |
1801 | ||
1802 | /* | |
1803 | * Theoretically, we should return -EBUSY when we | |
1804 | * encounter this race. In fact, we have a chance | |
1805 | * to successfully dissolve the page if we do a | |
1806 | * retry. Because the race window is quite small. | |
1807 | * If we seize this opportunity, it is an optimization | |
1808 | * for increasing the success rate of dissolving page. | |
1809 | */ | |
1810 | goto retry; | |
1811 | } | |
1812 | ||
1813 | /* | |
1814 | * Move PageHWPoison flag from head page to the raw error page, | |
1815 | * which makes any subpages rather than the error page reusable. | |
1816 | */ | |
1817 | if (PageHWPoison(head) && page != head) { | |
1818 | SetPageHWPoison(page); | |
1819 | ClearPageHWPoison(head); | |
1820 | } | |
1821 | list_del(&head->lru); | |
1822 | h->free_huge_pages--; | |
1823 | h->free_huge_pages_node[nid]--; | |
1824 | h->max_huge_pages--; | |
1825 | update_and_free_page(h, head); | |
1826 | rc = 0; | |
1827 | } | |
1828 | out: | |
1829 | spin_unlock(&hugetlb_lock); | |
1830 | return rc; | |
1831 | } | |
1832 | ||
1833 | /* | |
1834 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1835 | * make specified memory blocks removable from the system. | |
1836 | * Note that this will dissolve a free gigantic hugepage completely, if any | |
1837 | * part of it lies within the given range. | |
1838 | * Also note that if dissolve_free_huge_page() returns with an error, all | |
1839 | * free hugepages that were dissolved before that error are lost. | |
1840 | */ | |
1841 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) | |
1842 | { | |
1843 | unsigned long pfn; | |
1844 | struct page *page; | |
1845 | int rc = 0; | |
1846 | ||
1847 | if (!hugepages_supported()) | |
1848 | return rc; | |
1849 | ||
1850 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { | |
1851 | page = pfn_to_page(pfn); | |
1852 | rc = dissolve_free_huge_page(page); | |
1853 | if (rc) | |
1854 | break; | |
1855 | } | |
1856 | ||
1857 | return rc; | |
1858 | } | |
1859 | ||
1860 | /* | |
1861 | * Allocates a fresh surplus page from the page allocator. | |
1862 | */ | |
1863 | static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, | |
1864 | int nid, nodemask_t *nmask) | |
1865 | { | |
1866 | struct page *page = NULL; | |
1867 | ||
1868 | if (hstate_is_gigantic(h)) | |
1869 | return NULL; | |
1870 | ||
1871 | spin_lock(&hugetlb_lock); | |
1872 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) | |
1873 | goto out_unlock; | |
1874 | spin_unlock(&hugetlb_lock); | |
1875 | ||
1876 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); | |
1877 | if (!page) | |
1878 | return NULL; | |
1879 | ||
1880 | spin_lock(&hugetlb_lock); | |
1881 | /* | |
1882 | * We could have raced with the pool size change. | |
1883 | * Double check that and simply deallocate the new page | |
1884 | * if we would end up overcommiting the surpluses. Abuse | |
1885 | * temporary page to workaround the nasty free_huge_page | |
1886 | * codeflow | |
1887 | */ | |
1888 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
1889 | SetPageHugeTemporary(page); | |
1890 | spin_unlock(&hugetlb_lock); | |
1891 | put_page(page); | |
1892 | return NULL; | |
1893 | } else { | |
1894 | h->surplus_huge_pages++; | |
1895 | h->surplus_huge_pages_node[page_to_nid(page)]++; | |
1896 | } | |
1897 | ||
1898 | out_unlock: | |
1899 | spin_unlock(&hugetlb_lock); | |
1900 | ||
1901 | return page; | |
1902 | } | |
1903 | ||
1904 | static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, | |
1905 | int nid, nodemask_t *nmask) | |
1906 | { | |
1907 | struct page *page; | |
1908 | ||
1909 | if (hstate_is_gigantic(h)) | |
1910 | return NULL; | |
1911 | ||
1912 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); | |
1913 | if (!page) | |
1914 | return NULL; | |
1915 | ||
1916 | /* | |
1917 | * We do not account these pages as surplus because they are only | |
1918 | * temporary and will be released properly on the last reference | |
1919 | */ | |
1920 | SetPageHugeTemporary(page); | |
1921 | ||
1922 | return page; | |
1923 | } | |
1924 | ||
1925 | /* | |
1926 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
1927 | */ | |
1928 | static | |
1929 | struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, | |
1930 | struct vm_area_struct *vma, unsigned long addr) | |
1931 | { | |
1932 | struct page *page; | |
1933 | struct mempolicy *mpol; | |
1934 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
1935 | int nid; | |
1936 | nodemask_t *nodemask; | |
1937 | ||
1938 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
1939 | page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); | |
1940 | mpol_cond_put(mpol); | |
1941 | ||
1942 | return page; | |
1943 | } | |
1944 | ||
1945 | /* page migration callback function */ | |
1946 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, | |
1947 | nodemask_t *nmask, gfp_t gfp_mask) | |
1948 | { | |
1949 | spin_lock(&hugetlb_lock); | |
1950 | if (h->free_huge_pages - h->resv_huge_pages > 0) { | |
1951 | struct page *page; | |
1952 | ||
1953 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
1954 | if (page) { | |
1955 | spin_unlock(&hugetlb_lock); | |
1956 | return page; | |
1957 | } | |
1958 | } | |
1959 | spin_unlock(&hugetlb_lock); | |
1960 | ||
1961 | return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); | |
1962 | } | |
1963 | ||
1964 | /* mempolicy aware migration callback */ | |
1965 | struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, | |
1966 | unsigned long address) | |
1967 | { | |
1968 | struct mempolicy *mpol; | |
1969 | nodemask_t *nodemask; | |
1970 | struct page *page; | |
1971 | gfp_t gfp_mask; | |
1972 | int node; | |
1973 | ||
1974 | gfp_mask = htlb_alloc_mask(h); | |
1975 | node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
1976 | page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); | |
1977 | mpol_cond_put(mpol); | |
1978 | ||
1979 | return page; | |
1980 | } | |
1981 | ||
1982 | /* | |
1983 | * Increase the hugetlb pool such that it can accommodate a reservation | |
1984 | * of size 'delta'. | |
1985 | */ | |
1986 | static int gather_surplus_pages(struct hstate *h, long delta) | |
1987 | __must_hold(&hugetlb_lock) | |
1988 | { | |
1989 | struct list_head surplus_list; | |
1990 | struct page *page, *tmp; | |
1991 | int ret; | |
1992 | long i; | |
1993 | long needed, allocated; | |
1994 | bool alloc_ok = true; | |
1995 | ||
1996 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; | |
1997 | if (needed <= 0) { | |
1998 | h->resv_huge_pages += delta; | |
1999 | return 0; | |
2000 | } | |
2001 | ||
2002 | allocated = 0; | |
2003 | INIT_LIST_HEAD(&surplus_list); | |
2004 | ||
2005 | ret = -ENOMEM; | |
2006 | retry: | |
2007 | spin_unlock(&hugetlb_lock); | |
2008 | for (i = 0; i < needed; i++) { | |
2009 | page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), | |
2010 | NUMA_NO_NODE, NULL); | |
2011 | if (!page) { | |
2012 | alloc_ok = false; | |
2013 | break; | |
2014 | } | |
2015 | list_add(&page->lru, &surplus_list); | |
2016 | cond_resched(); | |
2017 | } | |
2018 | allocated += i; | |
2019 | ||
2020 | /* | |
2021 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
2022 | * because either resv_huge_pages or free_huge_pages may have changed. | |
2023 | */ | |
2024 | spin_lock(&hugetlb_lock); | |
2025 | needed = (h->resv_huge_pages + delta) - | |
2026 | (h->free_huge_pages + allocated); | |
2027 | if (needed > 0) { | |
2028 | if (alloc_ok) | |
2029 | goto retry; | |
2030 | /* | |
2031 | * We were not able to allocate enough pages to | |
2032 | * satisfy the entire reservation so we free what | |
2033 | * we've allocated so far. | |
2034 | */ | |
2035 | goto free; | |
2036 | } | |
2037 | /* | |
2038 | * The surplus_list now contains _at_least_ the number of extra pages | |
2039 | * needed to accommodate the reservation. Add the appropriate number | |
2040 | * of pages to the hugetlb pool and free the extras back to the buddy | |
2041 | * allocator. Commit the entire reservation here to prevent another | |
2042 | * process from stealing the pages as they are added to the pool but | |
2043 | * before they are reserved. | |
2044 | */ | |
2045 | needed += allocated; | |
2046 | h->resv_huge_pages += delta; | |
2047 | ret = 0; | |
2048 | ||
2049 | /* Free the needed pages to the hugetlb pool */ | |
2050 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { | |
2051 | if ((--needed) < 0) | |
2052 | break; | |
2053 | /* | |
2054 | * This page is now managed by the hugetlb allocator and has | |
2055 | * no users -- drop the buddy allocator's reference. | |
2056 | */ | |
2057 | VM_BUG_ON_PAGE(!put_page_testzero(page), page); | |
2058 | enqueue_huge_page(h, page); | |
2059 | } | |
2060 | free: | |
2061 | spin_unlock(&hugetlb_lock); | |
2062 | ||
2063 | /* Free unnecessary surplus pages to the buddy allocator */ | |
2064 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) | |
2065 | put_page(page); | |
2066 | spin_lock(&hugetlb_lock); | |
2067 | ||
2068 | return ret; | |
2069 | } | |
2070 | ||
2071 | /* | |
2072 | * This routine has two main purposes: | |
2073 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
2074 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
2075 | * to the associated reservation map. | |
2076 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
2077 | * the reservation. As many as unused_resv_pages may be freed. | |
2078 | * | |
2079 | * Called with hugetlb_lock held. However, the lock could be dropped (and | |
2080 | * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, | |
2081 | * we must make sure nobody else can claim pages we are in the process of | |
2082 | * freeing. Do this by ensuring resv_huge_page always is greater than the | |
2083 | * number of huge pages we plan to free when dropping the lock. | |
2084 | */ | |
2085 | static void return_unused_surplus_pages(struct hstate *h, | |
2086 | unsigned long unused_resv_pages) | |
2087 | { | |
2088 | unsigned long nr_pages; | |
2089 | ||
2090 | /* Cannot return gigantic pages currently */ | |
2091 | if (hstate_is_gigantic(h)) | |
2092 | goto out; | |
2093 | ||
2094 | /* | |
2095 | * Part (or even all) of the reservation could have been backed | |
2096 | * by pre-allocated pages. Only free surplus pages. | |
2097 | */ | |
2098 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); | |
2099 | ||
2100 | /* | |
2101 | * We want to release as many surplus pages as possible, spread | |
2102 | * evenly across all nodes with memory. Iterate across these nodes | |
2103 | * until we can no longer free unreserved surplus pages. This occurs | |
2104 | * when the nodes with surplus pages have no free pages. | |
2105 | * free_pool_huge_page() will balance the freed pages across the | |
2106 | * on-line nodes with memory and will handle the hstate accounting. | |
2107 | * | |
2108 | * Note that we decrement resv_huge_pages as we free the pages. If | |
2109 | * we drop the lock, resv_huge_pages will still be sufficiently large | |
2110 | * to cover subsequent pages we may free. | |
2111 | */ | |
2112 | while (nr_pages--) { | |
2113 | h->resv_huge_pages--; | |
2114 | unused_resv_pages--; | |
2115 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) | |
2116 | goto out; | |
2117 | cond_resched_lock(&hugetlb_lock); | |
2118 | } | |
2119 | ||
2120 | out: | |
2121 | /* Fully uncommit the reservation */ | |
2122 | h->resv_huge_pages -= unused_resv_pages; | |
2123 | } | |
2124 | ||
2125 | ||
2126 | /* | |
2127 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation | |
2128 | * are used by the huge page allocation routines to manage reservations. | |
2129 | * | |
2130 | * vma_needs_reservation is called to determine if the huge page at addr | |
2131 | * within the vma has an associated reservation. If a reservation is | |
2132 | * needed, the value 1 is returned. The caller is then responsible for | |
2133 | * managing the global reservation and subpool usage counts. After | |
2134 | * the huge page has been allocated, vma_commit_reservation is called | |
2135 | * to add the page to the reservation map. If the page allocation fails, | |
2136 | * the reservation must be ended instead of committed. vma_end_reservation | |
2137 | * is called in such cases. | |
2138 | * | |
2139 | * In the normal case, vma_commit_reservation returns the same value | |
2140 | * as the preceding vma_needs_reservation call. The only time this | |
2141 | * is not the case is if a reserve map was changed between calls. It | |
2142 | * is the responsibility of the caller to notice the difference and | |
2143 | * take appropriate action. | |
2144 | * | |
2145 | * vma_add_reservation is used in error paths where a reservation must | |
2146 | * be restored when a newly allocated huge page must be freed. It is | |
2147 | * to be called after calling vma_needs_reservation to determine if a | |
2148 | * reservation exists. | |
2149 | */ | |
2150 | enum vma_resv_mode { | |
2151 | VMA_NEEDS_RESV, | |
2152 | VMA_COMMIT_RESV, | |
2153 | VMA_END_RESV, | |
2154 | VMA_ADD_RESV, | |
2155 | }; | |
2156 | static long __vma_reservation_common(struct hstate *h, | |
2157 | struct vm_area_struct *vma, unsigned long addr, | |
2158 | enum vma_resv_mode mode) | |
2159 | { | |
2160 | struct resv_map *resv; | |
2161 | pgoff_t idx; | |
2162 | long ret; | |
2163 | long dummy_out_regions_needed; | |
2164 | ||
2165 | resv = vma_resv_map(vma); | |
2166 | if (!resv) | |
2167 | return 1; | |
2168 | ||
2169 | idx = vma_hugecache_offset(h, vma, addr); | |
2170 | switch (mode) { | |
2171 | case VMA_NEEDS_RESV: | |
2172 | ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); | |
2173 | /* We assume that vma_reservation_* routines always operate on | |
2174 | * 1 page, and that adding to resv map a 1 page entry can only | |
2175 | * ever require 1 region. | |
2176 | */ | |
2177 | VM_BUG_ON(dummy_out_regions_needed != 1); | |
2178 | break; | |
2179 | case VMA_COMMIT_RESV: | |
2180 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); | |
2181 | /* region_add calls of range 1 should never fail. */ | |
2182 | VM_BUG_ON(ret < 0); | |
2183 | break; | |
2184 | case VMA_END_RESV: | |
2185 | region_abort(resv, idx, idx + 1, 1); | |
2186 | ret = 0; | |
2187 | break; | |
2188 | case VMA_ADD_RESV: | |
2189 | if (vma->vm_flags & VM_MAYSHARE) { | |
2190 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); | |
2191 | /* region_add calls of range 1 should never fail. */ | |
2192 | VM_BUG_ON(ret < 0); | |
2193 | } else { | |
2194 | region_abort(resv, idx, idx + 1, 1); | |
2195 | ret = region_del(resv, idx, idx + 1); | |
2196 | } | |
2197 | break; | |
2198 | default: | |
2199 | BUG(); | |
2200 | } | |
2201 | ||
2202 | if (vma->vm_flags & VM_MAYSHARE) | |
2203 | return ret; | |
2204 | else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { | |
2205 | /* | |
2206 | * In most cases, reserves always exist for private mappings. | |
2207 | * However, a file associated with mapping could have been | |
2208 | * hole punched or truncated after reserves were consumed. | |
2209 | * As subsequent fault on such a range will not use reserves. | |
2210 | * Subtle - The reserve map for private mappings has the | |
2211 | * opposite meaning than that of shared mappings. If NO | |
2212 | * entry is in the reserve map, it means a reservation exists. | |
2213 | * If an entry exists in the reserve map, it means the | |
2214 | * reservation has already been consumed. As a result, the | |
2215 | * return value of this routine is the opposite of the | |
2216 | * value returned from reserve map manipulation routines above. | |
2217 | */ | |
2218 | if (ret) | |
2219 | return 0; | |
2220 | else | |
2221 | return 1; | |
2222 | } | |
2223 | else | |
2224 | return ret < 0 ? ret : 0; | |
2225 | } | |
2226 | ||
2227 | static long vma_needs_reservation(struct hstate *h, | |
2228 | struct vm_area_struct *vma, unsigned long addr) | |
2229 | { | |
2230 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); | |
2231 | } | |
2232 | ||
2233 | static long vma_commit_reservation(struct hstate *h, | |
2234 | struct vm_area_struct *vma, unsigned long addr) | |
2235 | { | |
2236 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); | |
2237 | } | |
2238 | ||
2239 | static void vma_end_reservation(struct hstate *h, | |
2240 | struct vm_area_struct *vma, unsigned long addr) | |
2241 | { | |
2242 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); | |
2243 | } | |
2244 | ||
2245 | static long vma_add_reservation(struct hstate *h, | |
2246 | struct vm_area_struct *vma, unsigned long addr) | |
2247 | { | |
2248 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
2249 | } | |
2250 | ||
2251 | /* | |
2252 | * This routine is called to restore a reservation on error paths. In the | |
2253 | * specific error paths, a huge page was allocated (via alloc_huge_page) | |
2254 | * and is about to be freed. If a reservation for the page existed, | |
2255 | * alloc_huge_page would have consumed the reservation and set PagePrivate | |
2256 | * in the newly allocated page. When the page is freed via free_huge_page, | |
2257 | * the global reservation count will be incremented if PagePrivate is set. | |
2258 | * However, free_huge_page can not adjust the reserve map. Adjust the | |
2259 | * reserve map here to be consistent with global reserve count adjustments | |
2260 | * to be made by free_huge_page. | |
2261 | */ | |
2262 | static void restore_reserve_on_error(struct hstate *h, | |
2263 | struct vm_area_struct *vma, unsigned long address, | |
2264 | struct page *page) | |
2265 | { | |
2266 | if (unlikely(PagePrivate(page))) { | |
2267 | long rc = vma_needs_reservation(h, vma, address); | |
2268 | ||
2269 | if (unlikely(rc < 0)) { | |
2270 | /* | |
2271 | * Rare out of memory condition in reserve map | |
2272 | * manipulation. Clear PagePrivate so that | |
2273 | * global reserve count will not be incremented | |
2274 | * by free_huge_page. This will make it appear | |
2275 | * as though the reservation for this page was | |
2276 | * consumed. This may prevent the task from | |
2277 | * faulting in the page at a later time. This | |
2278 | * is better than inconsistent global huge page | |
2279 | * accounting of reserve counts. | |
2280 | */ | |
2281 | ClearPagePrivate(page); | |
2282 | } else if (rc) { | |
2283 | rc = vma_add_reservation(h, vma, address); | |
2284 | if (unlikely(rc < 0)) | |
2285 | /* | |
2286 | * See above comment about rare out of | |
2287 | * memory condition. | |
2288 | */ | |
2289 | ClearPagePrivate(page); | |
2290 | } else | |
2291 | vma_end_reservation(h, vma, address); | |
2292 | } | |
2293 | } | |
2294 | ||
2295 | struct page *alloc_huge_page(struct vm_area_struct *vma, | |
2296 | unsigned long addr, int avoid_reserve) | |
2297 | { | |
2298 | struct hugepage_subpool *spool = subpool_vma(vma); | |
2299 | struct hstate *h = hstate_vma(vma); | |
2300 | struct page *page; | |
2301 | long map_chg, map_commit; | |
2302 | long gbl_chg; | |
2303 | int ret, idx; | |
2304 | struct hugetlb_cgroup *h_cg; | |
2305 | bool deferred_reserve; | |
2306 | ||
2307 | idx = hstate_index(h); | |
2308 | /* | |
2309 | * Examine the region/reserve map to determine if the process | |
2310 | * has a reservation for the page to be allocated. A return | |
2311 | * code of zero indicates a reservation exists (no change). | |
2312 | */ | |
2313 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); | |
2314 | if (map_chg < 0) | |
2315 | return ERR_PTR(-ENOMEM); | |
2316 | ||
2317 | /* | |
2318 | * Processes that did not create the mapping will have no | |
2319 | * reserves as indicated by the region/reserve map. Check | |
2320 | * that the allocation will not exceed the subpool limit. | |
2321 | * Allocations for MAP_NORESERVE mappings also need to be | |
2322 | * checked against any subpool limit. | |
2323 | */ | |
2324 | if (map_chg || avoid_reserve) { | |
2325 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2326 | if (gbl_chg < 0) { | |
2327 | vma_end_reservation(h, vma, addr); | |
2328 | return ERR_PTR(-ENOSPC); | |
2329 | } | |
2330 | ||
2331 | /* | |
2332 | * Even though there was no reservation in the region/reserve | |
2333 | * map, there could be reservations associated with the | |
2334 | * subpool that can be used. This would be indicated if the | |
2335 | * return value of hugepage_subpool_get_pages() is zero. | |
2336 | * However, if avoid_reserve is specified we still avoid even | |
2337 | * the subpool reservations. | |
2338 | */ | |
2339 | if (avoid_reserve) | |
2340 | gbl_chg = 1; | |
2341 | } | |
2342 | ||
2343 | /* If this allocation is not consuming a reservation, charge it now. | |
2344 | */ | |
2345 | deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma); | |
2346 | if (deferred_reserve) { | |
2347 | ret = hugetlb_cgroup_charge_cgroup_rsvd( | |
2348 | idx, pages_per_huge_page(h), &h_cg); | |
2349 | if (ret) | |
2350 | goto out_subpool_put; | |
2351 | } | |
2352 | ||
2353 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); | |
2354 | if (ret) | |
2355 | goto out_uncharge_cgroup_reservation; | |
2356 | ||
2357 | spin_lock(&hugetlb_lock); | |
2358 | /* | |
2359 | * glb_chg is passed to indicate whether or not a page must be taken | |
2360 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2361 | * a reservation exists for the allocation. | |
2362 | */ | |
2363 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
2364 | if (!page) { | |
2365 | spin_unlock(&hugetlb_lock); | |
2366 | page = alloc_buddy_huge_page_with_mpol(h, vma, addr); | |
2367 | if (!page) | |
2368 | goto out_uncharge_cgroup; | |
2369 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { | |
2370 | SetPagePrivate(page); | |
2371 | h->resv_huge_pages--; | |
2372 | } | |
2373 | spin_lock(&hugetlb_lock); | |
2374 | list_add(&page->lru, &h->hugepage_activelist); | |
2375 | /* Fall through */ | |
2376 | } | |
2377 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); | |
2378 | /* If allocation is not consuming a reservation, also store the | |
2379 | * hugetlb_cgroup pointer on the page. | |
2380 | */ | |
2381 | if (deferred_reserve) { | |
2382 | hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), | |
2383 | h_cg, page); | |
2384 | } | |
2385 | ||
2386 | spin_unlock(&hugetlb_lock); | |
2387 | ||
2388 | set_page_private(page, (unsigned long)spool); | |
2389 | ||
2390 | map_commit = vma_commit_reservation(h, vma, addr); | |
2391 | if (unlikely(map_chg > map_commit)) { | |
2392 | /* | |
2393 | * The page was added to the reservation map between | |
2394 | * vma_needs_reservation and vma_commit_reservation. | |
2395 | * This indicates a race with hugetlb_reserve_pages. | |
2396 | * Adjust for the subpool count incremented above AND | |
2397 | * in hugetlb_reserve_pages for the same page. Also, | |
2398 | * the reservation count added in hugetlb_reserve_pages | |
2399 | * no longer applies. | |
2400 | */ | |
2401 | long rsv_adjust; | |
2402 | ||
2403 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2404 | hugetlb_acct_memory(h, -rsv_adjust); | |
2405 | if (deferred_reserve) | |
2406 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), | |
2407 | pages_per_huge_page(h), page); | |
2408 | } | |
2409 | return page; | |
2410 | ||
2411 | out_uncharge_cgroup: | |
2412 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
2413 | out_uncharge_cgroup_reservation: | |
2414 | if (deferred_reserve) | |
2415 | hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), | |
2416 | h_cg); | |
2417 | out_subpool_put: | |
2418 | if (map_chg || avoid_reserve) | |
2419 | hugepage_subpool_put_pages(spool, 1); | |
2420 | vma_end_reservation(h, vma, addr); | |
2421 | return ERR_PTR(-ENOSPC); | |
2422 | } | |
2423 | ||
2424 | int alloc_bootmem_huge_page(struct hstate *h) | |
2425 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); | |
2426 | int __alloc_bootmem_huge_page(struct hstate *h) | |
2427 | { | |
2428 | struct huge_bootmem_page *m; | |
2429 | int nr_nodes, node; | |
2430 | ||
2431 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { | |
2432 | void *addr; | |
2433 | ||
2434 | addr = memblock_alloc_try_nid_raw( | |
2435 | huge_page_size(h), huge_page_size(h), | |
2436 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); | |
2437 | if (addr) { | |
2438 | /* | |
2439 | * Use the beginning of the huge page to store the | |
2440 | * huge_bootmem_page struct (until gather_bootmem | |
2441 | * puts them into the mem_map). | |
2442 | */ | |
2443 | m = addr; | |
2444 | goto found; | |
2445 | } | |
2446 | } | |
2447 | return 0; | |
2448 | ||
2449 | found: | |
2450 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); | |
2451 | /* Put them into a private list first because mem_map is not up yet */ | |
2452 | INIT_LIST_HEAD(&m->list); | |
2453 | list_add(&m->list, &huge_boot_pages); | |
2454 | m->hstate = h; | |
2455 | return 1; | |
2456 | } | |
2457 | ||
2458 | static void __init prep_compound_huge_page(struct page *page, | |
2459 | unsigned int order) | |
2460 | { | |
2461 | if (unlikely(order > (MAX_ORDER - 1))) | |
2462 | prep_compound_gigantic_page(page, order); | |
2463 | else | |
2464 | prep_compound_page(page, order); | |
2465 | } | |
2466 | ||
2467 | /* Put bootmem huge pages into the standard lists after mem_map is up */ | |
2468 | static void __init gather_bootmem_prealloc(void) | |
2469 | { | |
2470 | struct huge_bootmem_page *m; | |
2471 | ||
2472 | list_for_each_entry(m, &huge_boot_pages, list) { | |
2473 | struct page *page = virt_to_page(m); | |
2474 | struct hstate *h = m->hstate; | |
2475 | ||
2476 | WARN_ON(page_count(page) != 1); | |
2477 | prep_compound_huge_page(page, h->order); | |
2478 | WARN_ON(PageReserved(page)); | |
2479 | prep_new_huge_page(h, page, page_to_nid(page)); | |
2480 | put_page(page); /* free it into the hugepage allocator */ | |
2481 | ||
2482 | /* | |
2483 | * If we had gigantic hugepages allocated at boot time, we need | |
2484 | * to restore the 'stolen' pages to totalram_pages in order to | |
2485 | * fix confusing memory reports from free(1) and another | |
2486 | * side-effects, like CommitLimit going negative. | |
2487 | */ | |
2488 | if (hstate_is_gigantic(h)) | |
2489 | adjust_managed_page_count(page, 1 << h->order); | |
2490 | cond_resched(); | |
2491 | } | |
2492 | } | |
2493 | ||
2494 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) | |
2495 | { | |
2496 | unsigned long i; | |
2497 | nodemask_t *node_alloc_noretry; | |
2498 | ||
2499 | if (!hstate_is_gigantic(h)) { | |
2500 | /* | |
2501 | * Bit mask controlling how hard we retry per-node allocations. | |
2502 | * Ignore errors as lower level routines can deal with | |
2503 | * node_alloc_noretry == NULL. If this kmalloc fails at boot | |
2504 | * time, we are likely in bigger trouble. | |
2505 | */ | |
2506 | node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), | |
2507 | GFP_KERNEL); | |
2508 | } else { | |
2509 | /* allocations done at boot time */ | |
2510 | node_alloc_noretry = NULL; | |
2511 | } | |
2512 | ||
2513 | /* bit mask controlling how hard we retry per-node allocations */ | |
2514 | if (node_alloc_noretry) | |
2515 | nodes_clear(*node_alloc_noretry); | |
2516 | ||
2517 | for (i = 0; i < h->max_huge_pages; ++i) { | |
2518 | if (hstate_is_gigantic(h)) { | |
2519 | if (hugetlb_cma_size) { | |
2520 | pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); | |
2521 | break; | |
2522 | } | |
2523 | if (!alloc_bootmem_huge_page(h)) | |
2524 | break; | |
2525 | } else if (!alloc_pool_huge_page(h, | |
2526 | &node_states[N_MEMORY], | |
2527 | node_alloc_noretry)) | |
2528 | break; | |
2529 | cond_resched(); | |
2530 | } | |
2531 | if (i < h->max_huge_pages) { | |
2532 | char buf[32]; | |
2533 | ||
2534 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
2535 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", | |
2536 | h->max_huge_pages, buf, i); | |
2537 | h->max_huge_pages = i; | |
2538 | } | |
2539 | ||
2540 | kfree(node_alloc_noretry); | |
2541 | } | |
2542 | ||
2543 | static void __init hugetlb_init_hstates(void) | |
2544 | { | |
2545 | struct hstate *h; | |
2546 | ||
2547 | for_each_hstate(h) { | |
2548 | if (minimum_order > huge_page_order(h)) | |
2549 | minimum_order = huge_page_order(h); | |
2550 | ||
2551 | /* oversize hugepages were init'ed in early boot */ | |
2552 | if (!hstate_is_gigantic(h)) | |
2553 | hugetlb_hstate_alloc_pages(h); | |
2554 | } | |
2555 | VM_BUG_ON(minimum_order == UINT_MAX); | |
2556 | } | |
2557 | ||
2558 | static void __init report_hugepages(void) | |
2559 | { | |
2560 | struct hstate *h; | |
2561 | ||
2562 | for_each_hstate(h) { | |
2563 | char buf[32]; | |
2564 | ||
2565 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
2566 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", | |
2567 | buf, h->free_huge_pages); | |
2568 | } | |
2569 | } | |
2570 | ||
2571 | #ifdef CONFIG_HIGHMEM | |
2572 | static void try_to_free_low(struct hstate *h, unsigned long count, | |
2573 | nodemask_t *nodes_allowed) | |
2574 | { | |
2575 | int i; | |
2576 | ||
2577 | if (hstate_is_gigantic(h)) | |
2578 | return; | |
2579 | ||
2580 | for_each_node_mask(i, *nodes_allowed) { | |
2581 | struct page *page, *next; | |
2582 | struct list_head *freel = &h->hugepage_freelists[i]; | |
2583 | list_for_each_entry_safe(page, next, freel, lru) { | |
2584 | if (count >= h->nr_huge_pages) | |
2585 | return; | |
2586 | if (PageHighMem(page)) | |
2587 | continue; | |
2588 | list_del(&page->lru); | |
2589 | update_and_free_page(h, page); | |
2590 | h->free_huge_pages--; | |
2591 | h->free_huge_pages_node[page_to_nid(page)]--; | |
2592 | } | |
2593 | } | |
2594 | } | |
2595 | #else | |
2596 | static inline void try_to_free_low(struct hstate *h, unsigned long count, | |
2597 | nodemask_t *nodes_allowed) | |
2598 | { | |
2599 | } | |
2600 | #endif | |
2601 | ||
2602 | /* | |
2603 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2604 | * balanced by operating on them in a round-robin fashion. | |
2605 | * Returns 1 if an adjustment was made. | |
2606 | */ | |
2607 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, | |
2608 | int delta) | |
2609 | { | |
2610 | int nr_nodes, node; | |
2611 | ||
2612 | VM_BUG_ON(delta != -1 && delta != 1); | |
2613 | ||
2614 | if (delta < 0) { | |
2615 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2616 | if (h->surplus_huge_pages_node[node]) | |
2617 | goto found; | |
2618 | } | |
2619 | } else { | |
2620 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2621 | if (h->surplus_huge_pages_node[node] < | |
2622 | h->nr_huge_pages_node[node]) | |
2623 | goto found; | |
2624 | } | |
2625 | } | |
2626 | return 0; | |
2627 | ||
2628 | found: | |
2629 | h->surplus_huge_pages += delta; | |
2630 | h->surplus_huge_pages_node[node] += delta; | |
2631 | return 1; | |
2632 | } | |
2633 | ||
2634 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) | |
2635 | static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, | |
2636 | nodemask_t *nodes_allowed) | |
2637 | { | |
2638 | unsigned long min_count, ret; | |
2639 | NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); | |
2640 | ||
2641 | /* | |
2642 | * Bit mask controlling how hard we retry per-node allocations. | |
2643 | * If we can not allocate the bit mask, do not attempt to allocate | |
2644 | * the requested huge pages. | |
2645 | */ | |
2646 | if (node_alloc_noretry) | |
2647 | nodes_clear(*node_alloc_noretry); | |
2648 | else | |
2649 | return -ENOMEM; | |
2650 | ||
2651 | spin_lock(&hugetlb_lock); | |
2652 | ||
2653 | /* | |
2654 | * Check for a node specific request. | |
2655 | * Changing node specific huge page count may require a corresponding | |
2656 | * change to the global count. In any case, the passed node mask | |
2657 | * (nodes_allowed) will restrict alloc/free to the specified node. | |
2658 | */ | |
2659 | if (nid != NUMA_NO_NODE) { | |
2660 | unsigned long old_count = count; | |
2661 | ||
2662 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2663 | /* | |
2664 | * User may have specified a large count value which caused the | |
2665 | * above calculation to overflow. In this case, they wanted | |
2666 | * to allocate as many huge pages as possible. Set count to | |
2667 | * largest possible value to align with their intention. | |
2668 | */ | |
2669 | if (count < old_count) | |
2670 | count = ULONG_MAX; | |
2671 | } | |
2672 | ||
2673 | /* | |
2674 | * Gigantic pages runtime allocation depend on the capability for large | |
2675 | * page range allocation. | |
2676 | * If the system does not provide this feature, return an error when | |
2677 | * the user tries to allocate gigantic pages but let the user free the | |
2678 | * boottime allocated gigantic pages. | |
2679 | */ | |
2680 | if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { | |
2681 | if (count > persistent_huge_pages(h)) { | |
2682 | spin_unlock(&hugetlb_lock); | |
2683 | NODEMASK_FREE(node_alloc_noretry); | |
2684 | return -EINVAL; | |
2685 | } | |
2686 | /* Fall through to decrease pool */ | |
2687 | } | |
2688 | ||
2689 | /* | |
2690 | * Increase the pool size | |
2691 | * First take pages out of surplus state. Then make up the | |
2692 | * remaining difference by allocating fresh huge pages. | |
2693 | * | |
2694 | * We might race with alloc_surplus_huge_page() here and be unable | |
2695 | * to convert a surplus huge page to a normal huge page. That is | |
2696 | * not critical, though, it just means the overall size of the | |
2697 | * pool might be one hugepage larger than it needs to be, but | |
2698 | * within all the constraints specified by the sysctls. | |
2699 | */ | |
2700 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { | |
2701 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) | |
2702 | break; | |
2703 | } | |
2704 | ||
2705 | while (count > persistent_huge_pages(h)) { | |
2706 | /* | |
2707 | * If this allocation races such that we no longer need the | |
2708 | * page, free_huge_page will handle it by freeing the page | |
2709 | * and reducing the surplus. | |
2710 | */ | |
2711 | spin_unlock(&hugetlb_lock); | |
2712 | ||
2713 | /* yield cpu to avoid soft lockup */ | |
2714 | cond_resched(); | |
2715 | ||
2716 | ret = alloc_pool_huge_page(h, nodes_allowed, | |
2717 | node_alloc_noretry); | |
2718 | spin_lock(&hugetlb_lock); | |
2719 | if (!ret) | |
2720 | goto out; | |
2721 | ||
2722 | /* Bail for signals. Probably ctrl-c from user */ | |
2723 | if (signal_pending(current)) | |
2724 | goto out; | |
2725 | } | |
2726 | ||
2727 | /* | |
2728 | * Decrease the pool size | |
2729 | * First return free pages to the buddy allocator (being careful | |
2730 | * to keep enough around to satisfy reservations). Then place | |
2731 | * pages into surplus state as needed so the pool will shrink | |
2732 | * to the desired size as pages become free. | |
2733 | * | |
2734 | * By placing pages into the surplus state independent of the | |
2735 | * overcommit value, we are allowing the surplus pool size to | |
2736 | * exceed overcommit. There are few sane options here. Since | |
2737 | * alloc_surplus_huge_page() is checking the global counter, | |
2738 | * though, we'll note that we're not allowed to exceed surplus | |
2739 | * and won't grow the pool anywhere else. Not until one of the | |
2740 | * sysctls are changed, or the surplus pages go out of use. | |
2741 | */ | |
2742 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; | |
2743 | min_count = max(count, min_count); | |
2744 | try_to_free_low(h, min_count, nodes_allowed); | |
2745 | while (min_count < persistent_huge_pages(h)) { | |
2746 | if (!free_pool_huge_page(h, nodes_allowed, 0)) | |
2747 | break; | |
2748 | cond_resched_lock(&hugetlb_lock); | |
2749 | } | |
2750 | while (count < persistent_huge_pages(h)) { | |
2751 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) | |
2752 | break; | |
2753 | } | |
2754 | out: | |
2755 | h->max_huge_pages = persistent_huge_pages(h); | |
2756 | spin_unlock(&hugetlb_lock); | |
2757 | ||
2758 | NODEMASK_FREE(node_alloc_noretry); | |
2759 | ||
2760 | return 0; | |
2761 | } | |
2762 | ||
2763 | #define HSTATE_ATTR_RO(_name) \ | |
2764 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2765 | ||
2766 | #define HSTATE_ATTR(_name) \ | |
2767 | static struct kobj_attribute _name##_attr = \ | |
2768 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2769 | ||
2770 | static struct kobject *hugepages_kobj; | |
2771 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2772 | ||
2773 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); | |
2774 | ||
2775 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
2776 | { | |
2777 | int i; | |
2778 | ||
2779 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
2780 | if (hstate_kobjs[i] == kobj) { | |
2781 | if (nidp) | |
2782 | *nidp = NUMA_NO_NODE; | |
2783 | return &hstates[i]; | |
2784 | } | |
2785 | ||
2786 | return kobj_to_node_hstate(kobj, nidp); | |
2787 | } | |
2788 | ||
2789 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, | |
2790 | struct kobj_attribute *attr, char *buf) | |
2791 | { | |
2792 | struct hstate *h; | |
2793 | unsigned long nr_huge_pages; | |
2794 | int nid; | |
2795 | ||
2796 | h = kobj_to_hstate(kobj, &nid); | |
2797 | if (nid == NUMA_NO_NODE) | |
2798 | nr_huge_pages = h->nr_huge_pages; | |
2799 | else | |
2800 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2801 | ||
2802 | return sysfs_emit(buf, "%lu\n", nr_huge_pages); | |
2803 | } | |
2804 | ||
2805 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, | |
2806 | struct hstate *h, int nid, | |
2807 | unsigned long count, size_t len) | |
2808 | { | |
2809 | int err; | |
2810 | nodemask_t nodes_allowed, *n_mask; | |
2811 | ||
2812 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) | |
2813 | return -EINVAL; | |
2814 | ||
2815 | if (nid == NUMA_NO_NODE) { | |
2816 | /* | |
2817 | * global hstate attribute | |
2818 | */ | |
2819 | if (!(obey_mempolicy && | |
2820 | init_nodemask_of_mempolicy(&nodes_allowed))) | |
2821 | n_mask = &node_states[N_MEMORY]; | |
2822 | else | |
2823 | n_mask = &nodes_allowed; | |
2824 | } else { | |
2825 | /* | |
2826 | * Node specific request. count adjustment happens in | |
2827 | * set_max_huge_pages() after acquiring hugetlb_lock. | |
2828 | */ | |
2829 | init_nodemask_of_node(&nodes_allowed, nid); | |
2830 | n_mask = &nodes_allowed; | |
2831 | } | |
2832 | ||
2833 | err = set_max_huge_pages(h, count, nid, n_mask); | |
2834 | ||
2835 | return err ? err : len; | |
2836 | } | |
2837 | ||
2838 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, | |
2839 | struct kobject *kobj, const char *buf, | |
2840 | size_t len) | |
2841 | { | |
2842 | struct hstate *h; | |
2843 | unsigned long count; | |
2844 | int nid; | |
2845 | int err; | |
2846 | ||
2847 | err = kstrtoul(buf, 10, &count); | |
2848 | if (err) | |
2849 | return err; | |
2850 | ||
2851 | h = kobj_to_hstate(kobj, &nid); | |
2852 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2853 | } | |
2854 | ||
2855 | static ssize_t nr_hugepages_show(struct kobject *kobj, | |
2856 | struct kobj_attribute *attr, char *buf) | |
2857 | { | |
2858 | return nr_hugepages_show_common(kobj, attr, buf); | |
2859 | } | |
2860 | ||
2861 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2862 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2863 | { | |
2864 | return nr_hugepages_store_common(false, kobj, buf, len); | |
2865 | } | |
2866 | HSTATE_ATTR(nr_hugepages); | |
2867 | ||
2868 | #ifdef CONFIG_NUMA | |
2869 | ||
2870 | /* | |
2871 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2872 | * huge page alloc/free. | |
2873 | */ | |
2874 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
2875 | struct kobj_attribute *attr, | |
2876 | char *buf) | |
2877 | { | |
2878 | return nr_hugepages_show_common(kobj, attr, buf); | |
2879 | } | |
2880 | ||
2881 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2882 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2883 | { | |
2884 | return nr_hugepages_store_common(true, kobj, buf, len); | |
2885 | } | |
2886 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2887 | #endif | |
2888 | ||
2889 | ||
2890 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, | |
2891 | struct kobj_attribute *attr, char *buf) | |
2892 | { | |
2893 | struct hstate *h = kobj_to_hstate(kobj, NULL); | |
2894 | return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); | |
2895 | } | |
2896 | ||
2897 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, | |
2898 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2899 | { | |
2900 | int err; | |
2901 | unsigned long input; | |
2902 | struct hstate *h = kobj_to_hstate(kobj, NULL); | |
2903 | ||
2904 | if (hstate_is_gigantic(h)) | |
2905 | return -EINVAL; | |
2906 | ||
2907 | err = kstrtoul(buf, 10, &input); | |
2908 | if (err) | |
2909 | return err; | |
2910 | ||
2911 | spin_lock(&hugetlb_lock); | |
2912 | h->nr_overcommit_huge_pages = input; | |
2913 | spin_unlock(&hugetlb_lock); | |
2914 | ||
2915 | return count; | |
2916 | } | |
2917 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2918 | ||
2919 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2920 | struct kobj_attribute *attr, char *buf) | |
2921 | { | |
2922 | struct hstate *h; | |
2923 | unsigned long free_huge_pages; | |
2924 | int nid; | |
2925 | ||
2926 | h = kobj_to_hstate(kobj, &nid); | |
2927 | if (nid == NUMA_NO_NODE) | |
2928 | free_huge_pages = h->free_huge_pages; | |
2929 | else | |
2930 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2931 | ||
2932 | return sysfs_emit(buf, "%lu\n", free_huge_pages); | |
2933 | } | |
2934 | HSTATE_ATTR_RO(free_hugepages); | |
2935 | ||
2936 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2937 | struct kobj_attribute *attr, char *buf) | |
2938 | { | |
2939 | struct hstate *h = kobj_to_hstate(kobj, NULL); | |
2940 | return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); | |
2941 | } | |
2942 | HSTATE_ATTR_RO(resv_hugepages); | |
2943 | ||
2944 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2945 | struct kobj_attribute *attr, char *buf) | |
2946 | { | |
2947 | struct hstate *h; | |
2948 | unsigned long surplus_huge_pages; | |
2949 | int nid; | |
2950 | ||
2951 | h = kobj_to_hstate(kobj, &nid); | |
2952 | if (nid == NUMA_NO_NODE) | |
2953 | surplus_huge_pages = h->surplus_huge_pages; | |
2954 | else | |
2955 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2956 | ||
2957 | return sysfs_emit(buf, "%lu\n", surplus_huge_pages); | |
2958 | } | |
2959 | HSTATE_ATTR_RO(surplus_hugepages); | |
2960 | ||
2961 | static struct attribute *hstate_attrs[] = { | |
2962 | &nr_hugepages_attr.attr, | |
2963 | &nr_overcommit_hugepages_attr.attr, | |
2964 | &free_hugepages_attr.attr, | |
2965 | &resv_hugepages_attr.attr, | |
2966 | &surplus_hugepages_attr.attr, | |
2967 | #ifdef CONFIG_NUMA | |
2968 | &nr_hugepages_mempolicy_attr.attr, | |
2969 | #endif | |
2970 | NULL, | |
2971 | }; | |
2972 | ||
2973 | static const struct attribute_group hstate_attr_group = { | |
2974 | .attrs = hstate_attrs, | |
2975 | }; | |
2976 | ||
2977 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, | |
2978 | struct kobject **hstate_kobjs, | |
2979 | const struct attribute_group *hstate_attr_group) | |
2980 | { | |
2981 | int retval; | |
2982 | int hi = hstate_index(h); | |
2983 | ||
2984 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); | |
2985 | if (!hstate_kobjs[hi]) | |
2986 | return -ENOMEM; | |
2987 | ||
2988 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); | |
2989 | if (retval) | |
2990 | kobject_put(hstate_kobjs[hi]); | |
2991 | ||
2992 | return retval; | |
2993 | } | |
2994 | ||
2995 | static void __init hugetlb_sysfs_init(void) | |
2996 | { | |
2997 | struct hstate *h; | |
2998 | int err; | |
2999 | ||
3000 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
3001 | if (!hugepages_kobj) | |
3002 | return; | |
3003 | ||
3004 | for_each_hstate(h) { | |
3005 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, | |
3006 | hstate_kobjs, &hstate_attr_group); | |
3007 | if (err) | |
3008 | pr_err("HugeTLB: Unable to add hstate %s", h->name); | |
3009 | } | |
3010 | } | |
3011 | ||
3012 | #ifdef CONFIG_NUMA | |
3013 | ||
3014 | /* | |
3015 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
3016 | * with node devices in node_devices[] using a parallel array. The array | |
3017 | * index of a node device or _hstate == node id. | |
3018 | * This is here to avoid any static dependency of the node device driver, in | |
3019 | * the base kernel, on the hugetlb module. | |
3020 | */ | |
3021 | struct node_hstate { | |
3022 | struct kobject *hugepages_kobj; | |
3023 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
3024 | }; | |
3025 | static struct node_hstate node_hstates[MAX_NUMNODES]; | |
3026 | ||
3027 | /* | |
3028 | * A subset of global hstate attributes for node devices | |
3029 | */ | |
3030 | static struct attribute *per_node_hstate_attrs[] = { | |
3031 | &nr_hugepages_attr.attr, | |
3032 | &free_hugepages_attr.attr, | |
3033 | &surplus_hugepages_attr.attr, | |
3034 | NULL, | |
3035 | }; | |
3036 | ||
3037 | static const struct attribute_group per_node_hstate_attr_group = { | |
3038 | .attrs = per_node_hstate_attrs, | |
3039 | }; | |
3040 | ||
3041 | /* | |
3042 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. | |
3043 | * Returns node id via non-NULL nidp. | |
3044 | */ | |
3045 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3046 | { | |
3047 | int nid; | |
3048 | ||
3049 | for (nid = 0; nid < nr_node_ids; nid++) { | |
3050 | struct node_hstate *nhs = &node_hstates[nid]; | |
3051 | int i; | |
3052 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
3053 | if (nhs->hstate_kobjs[i] == kobj) { | |
3054 | if (nidp) | |
3055 | *nidp = nid; | |
3056 | return &hstates[i]; | |
3057 | } | |
3058 | } | |
3059 | ||
3060 | BUG(); | |
3061 | return NULL; | |
3062 | } | |
3063 | ||
3064 | /* | |
3065 | * Unregister hstate attributes from a single node device. | |
3066 | * No-op if no hstate attributes attached. | |
3067 | */ | |
3068 | static void hugetlb_unregister_node(struct node *node) | |
3069 | { | |
3070 | struct hstate *h; | |
3071 | struct node_hstate *nhs = &node_hstates[node->dev.id]; | |
3072 | ||
3073 | if (!nhs->hugepages_kobj) | |
3074 | return; /* no hstate attributes */ | |
3075 | ||
3076 | for_each_hstate(h) { | |
3077 | int idx = hstate_index(h); | |
3078 | if (nhs->hstate_kobjs[idx]) { | |
3079 | kobject_put(nhs->hstate_kobjs[idx]); | |
3080 | nhs->hstate_kobjs[idx] = NULL; | |
3081 | } | |
3082 | } | |
3083 | ||
3084 | kobject_put(nhs->hugepages_kobj); | |
3085 | nhs->hugepages_kobj = NULL; | |
3086 | } | |
3087 | ||
3088 | ||
3089 | /* | |
3090 | * Register hstate attributes for a single node device. | |
3091 | * No-op if attributes already registered. | |
3092 | */ | |
3093 | static void hugetlb_register_node(struct node *node) | |
3094 | { | |
3095 | struct hstate *h; | |
3096 | struct node_hstate *nhs = &node_hstates[node->dev.id]; | |
3097 | int err; | |
3098 | ||
3099 | if (nhs->hugepages_kobj) | |
3100 | return; /* already allocated */ | |
3101 | ||
3102 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
3103 | &node->dev.kobj); | |
3104 | if (!nhs->hugepages_kobj) | |
3105 | return; | |
3106 | ||
3107 | for_each_hstate(h) { | |
3108 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
3109 | nhs->hstate_kobjs, | |
3110 | &per_node_hstate_attr_group); | |
3111 | if (err) { | |
3112 | pr_err("HugeTLB: Unable to add hstate %s for node %d\n", | |
3113 | h->name, node->dev.id); | |
3114 | hugetlb_unregister_node(node); | |
3115 | break; | |
3116 | } | |
3117 | } | |
3118 | } | |
3119 | ||
3120 | /* | |
3121 | * hugetlb init time: register hstate attributes for all registered node | |
3122 | * devices of nodes that have memory. All on-line nodes should have | |
3123 | * registered their associated device by this time. | |
3124 | */ | |
3125 | static void __init hugetlb_register_all_nodes(void) | |
3126 | { | |
3127 | int nid; | |
3128 | ||
3129 | for_each_node_state(nid, N_MEMORY) { | |
3130 | struct node *node = node_devices[nid]; | |
3131 | if (node->dev.id == nid) | |
3132 | hugetlb_register_node(node); | |
3133 | } | |
3134 | ||
3135 | /* | |
3136 | * Let the node device driver know we're here so it can | |
3137 | * [un]register hstate attributes on node hotplug. | |
3138 | */ | |
3139 | register_hugetlbfs_with_node(hugetlb_register_node, | |
3140 | hugetlb_unregister_node); | |
3141 | } | |
3142 | #else /* !CONFIG_NUMA */ | |
3143 | ||
3144 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3145 | { | |
3146 | BUG(); | |
3147 | if (nidp) | |
3148 | *nidp = -1; | |
3149 | return NULL; | |
3150 | } | |
3151 | ||
3152 | static void hugetlb_register_all_nodes(void) { } | |
3153 | ||
3154 | #endif | |
3155 | ||
3156 | static int __init hugetlb_init(void) | |
3157 | { | |
3158 | int i; | |
3159 | ||
3160 | if (!hugepages_supported()) { | |
3161 | if (hugetlb_max_hstate || default_hstate_max_huge_pages) | |
3162 | pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); | |
3163 | return 0; | |
3164 | } | |
3165 | ||
3166 | /* | |
3167 | * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some | |
3168 | * architectures depend on setup being done here. | |
3169 | */ | |
3170 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
3171 | if (!parsed_default_hugepagesz) { | |
3172 | /* | |
3173 | * If we did not parse a default huge page size, set | |
3174 | * default_hstate_idx to HPAGE_SIZE hstate. And, if the | |
3175 | * number of huge pages for this default size was implicitly | |
3176 | * specified, set that here as well. | |
3177 | * Note that the implicit setting will overwrite an explicit | |
3178 | * setting. A warning will be printed in this case. | |
3179 | */ | |
3180 | default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); | |
3181 | if (default_hstate_max_huge_pages) { | |
3182 | if (default_hstate.max_huge_pages) { | |
3183 | char buf[32]; | |
3184 | ||
3185 | string_get_size(huge_page_size(&default_hstate), | |
3186 | 1, STRING_UNITS_2, buf, 32); | |
3187 | pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", | |
3188 | default_hstate.max_huge_pages, buf); | |
3189 | pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", | |
3190 | default_hstate_max_huge_pages); | |
3191 | } | |
3192 | default_hstate.max_huge_pages = | |
3193 | default_hstate_max_huge_pages; | |
3194 | } | |
3195 | } | |
3196 | ||
3197 | hugetlb_cma_check(); | |
3198 | hugetlb_init_hstates(); | |
3199 | gather_bootmem_prealloc(); | |
3200 | report_hugepages(); | |
3201 | ||
3202 | hugetlb_sysfs_init(); | |
3203 | hugetlb_register_all_nodes(); | |
3204 | hugetlb_cgroup_file_init(); | |
3205 | ||
3206 | #ifdef CONFIG_SMP | |
3207 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
3208 | #else | |
3209 | num_fault_mutexes = 1; | |
3210 | #endif | |
3211 | hugetlb_fault_mutex_table = | |
3212 | kmalloc_array(num_fault_mutexes, sizeof(struct mutex), | |
3213 | GFP_KERNEL); | |
3214 | BUG_ON(!hugetlb_fault_mutex_table); | |
3215 | ||
3216 | for (i = 0; i < num_fault_mutexes; i++) | |
3217 | mutex_init(&hugetlb_fault_mutex_table[i]); | |
3218 | return 0; | |
3219 | } | |
3220 | subsys_initcall(hugetlb_init); | |
3221 | ||
3222 | /* Overwritten by architectures with more huge page sizes */ | |
3223 | bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) | |
3224 | { | |
3225 | return size == HPAGE_SIZE; | |
3226 | } | |
3227 | ||
3228 | void __init hugetlb_add_hstate(unsigned int order) | |
3229 | { | |
3230 | struct hstate *h; | |
3231 | unsigned long i; | |
3232 | ||
3233 | if (size_to_hstate(PAGE_SIZE << order)) { | |
3234 | return; | |
3235 | } | |
3236 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); | |
3237 | BUG_ON(order == 0); | |
3238 | h = &hstates[hugetlb_max_hstate++]; | |
3239 | h->order = order; | |
3240 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
3241 | for (i = 0; i < MAX_NUMNODES; ++i) | |
3242 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
3243 | INIT_LIST_HEAD(&h->hugepage_activelist); | |
3244 | h->next_nid_to_alloc = first_memory_node; | |
3245 | h->next_nid_to_free = first_memory_node; | |
3246 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", | |
3247 | huge_page_size(h)/1024); | |
3248 | ||
3249 | parsed_hstate = h; | |
3250 | } | |
3251 | ||
3252 | /* | |
3253 | * hugepages command line processing | |
3254 | * hugepages normally follows a valid hugepagsz or default_hugepagsz | |
3255 | * specification. If not, ignore the hugepages value. hugepages can also | |
3256 | * be the first huge page command line option in which case it implicitly | |
3257 | * specifies the number of huge pages for the default size. | |
3258 | */ | |
3259 | static int __init hugepages_setup(char *s) | |
3260 | { | |
3261 | unsigned long *mhp; | |
3262 | static unsigned long *last_mhp; | |
3263 | ||
3264 | if (!parsed_valid_hugepagesz) { | |
3265 | pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); | |
3266 | parsed_valid_hugepagesz = true; | |
3267 | return 0; | |
3268 | } | |
3269 | ||
3270 | /* | |
3271 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter | |
3272 | * yet, so this hugepages= parameter goes to the "default hstate". | |
3273 | * Otherwise, it goes with the previously parsed hugepagesz or | |
3274 | * default_hugepagesz. | |
3275 | */ | |
3276 | else if (!hugetlb_max_hstate) | |
3277 | mhp = &default_hstate_max_huge_pages; | |
3278 | else | |
3279 | mhp = &parsed_hstate->max_huge_pages; | |
3280 | ||
3281 | if (mhp == last_mhp) { | |
3282 | pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); | |
3283 | return 0; | |
3284 | } | |
3285 | ||
3286 | if (sscanf(s, "%lu", mhp) <= 0) | |
3287 | *mhp = 0; | |
3288 | ||
3289 | /* | |
3290 | * Global state is always initialized later in hugetlb_init. | |
3291 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
3292 | * use the bootmem allocator. | |
3293 | */ | |
3294 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) | |
3295 | hugetlb_hstate_alloc_pages(parsed_hstate); | |
3296 | ||
3297 | last_mhp = mhp; | |
3298 | ||
3299 | return 1; | |
3300 | } | |
3301 | __setup("hugepages=", hugepages_setup); | |
3302 | ||
3303 | /* | |
3304 | * hugepagesz command line processing | |
3305 | * A specific huge page size can only be specified once with hugepagesz. | |
3306 | * hugepagesz is followed by hugepages on the command line. The global | |
3307 | * variable 'parsed_valid_hugepagesz' is used to determine if prior | |
3308 | * hugepagesz argument was valid. | |
3309 | */ | |
3310 | static int __init hugepagesz_setup(char *s) | |
3311 | { | |
3312 | unsigned long size; | |
3313 | struct hstate *h; | |
3314 | ||
3315 | parsed_valid_hugepagesz = false; | |
3316 | size = (unsigned long)memparse(s, NULL); | |
3317 | ||
3318 | if (!arch_hugetlb_valid_size(size)) { | |
3319 | pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); | |
3320 | return 0; | |
3321 | } | |
3322 | ||
3323 | h = size_to_hstate(size); | |
3324 | if (h) { | |
3325 | /* | |
3326 | * hstate for this size already exists. This is normally | |
3327 | * an error, but is allowed if the existing hstate is the | |
3328 | * default hstate. More specifically, it is only allowed if | |
3329 | * the number of huge pages for the default hstate was not | |
3330 | * previously specified. | |
3331 | */ | |
3332 | if (!parsed_default_hugepagesz || h != &default_hstate || | |
3333 | default_hstate.max_huge_pages) { | |
3334 | pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); | |
3335 | return 0; | |
3336 | } | |
3337 | ||
3338 | /* | |
3339 | * No need to call hugetlb_add_hstate() as hstate already | |
3340 | * exists. But, do set parsed_hstate so that a following | |
3341 | * hugepages= parameter will be applied to this hstate. | |
3342 | */ | |
3343 | parsed_hstate = h; | |
3344 | parsed_valid_hugepagesz = true; | |
3345 | return 1; | |
3346 | } | |
3347 | ||
3348 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); | |
3349 | parsed_valid_hugepagesz = true; | |
3350 | return 1; | |
3351 | } | |
3352 | __setup("hugepagesz=", hugepagesz_setup); | |
3353 | ||
3354 | /* | |
3355 | * default_hugepagesz command line input | |
3356 | * Only one instance of default_hugepagesz allowed on command line. | |
3357 | */ | |
3358 | static int __init default_hugepagesz_setup(char *s) | |
3359 | { | |
3360 | unsigned long size; | |
3361 | ||
3362 | parsed_valid_hugepagesz = false; | |
3363 | if (parsed_default_hugepagesz) { | |
3364 | pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); | |
3365 | return 0; | |
3366 | } | |
3367 | ||
3368 | size = (unsigned long)memparse(s, NULL); | |
3369 | ||
3370 | if (!arch_hugetlb_valid_size(size)) { | |
3371 | pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); | |
3372 | return 0; | |
3373 | } | |
3374 | ||
3375 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); | |
3376 | parsed_valid_hugepagesz = true; | |
3377 | parsed_default_hugepagesz = true; | |
3378 | default_hstate_idx = hstate_index(size_to_hstate(size)); | |
3379 | ||
3380 | /* | |
3381 | * The number of default huge pages (for this size) could have been | |
3382 | * specified as the first hugetlb parameter: hugepages=X. If so, | |
3383 | * then default_hstate_max_huge_pages is set. If the default huge | |
3384 | * page size is gigantic (>= MAX_ORDER), then the pages must be | |
3385 | * allocated here from bootmem allocator. | |
3386 | */ | |
3387 | if (default_hstate_max_huge_pages) { | |
3388 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
3389 | if (hstate_is_gigantic(&default_hstate)) | |
3390 | hugetlb_hstate_alloc_pages(&default_hstate); | |
3391 | default_hstate_max_huge_pages = 0; | |
3392 | } | |
3393 | ||
3394 | return 1; | |
3395 | } | |
3396 | __setup("default_hugepagesz=", default_hugepagesz_setup); | |
3397 | ||
3398 | static unsigned int allowed_mems_nr(struct hstate *h) | |
3399 | { | |
3400 | int node; | |
3401 | unsigned int nr = 0; | |
3402 | nodemask_t *mpol_allowed; | |
3403 | unsigned int *array = h->free_huge_pages_node; | |
3404 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
3405 | ||
3406 | mpol_allowed = policy_nodemask_current(gfp_mask); | |
3407 | ||
3408 | for_each_node_mask(node, cpuset_current_mems_allowed) { | |
3409 | if (!mpol_allowed || | |
3410 | (mpol_allowed && node_isset(node, *mpol_allowed))) | |
3411 | nr += array[node]; | |
3412 | } | |
3413 | ||
3414 | return nr; | |
3415 | } | |
3416 | ||
3417 | #ifdef CONFIG_SYSCTL | |
3418 | static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, | |
3419 | void *buffer, size_t *length, | |
3420 | loff_t *ppos, unsigned long *out) | |
3421 | { | |
3422 | struct ctl_table dup_table; | |
3423 | ||
3424 | /* | |
3425 | * In order to avoid races with __do_proc_doulongvec_minmax(), we | |
3426 | * can duplicate the @table and alter the duplicate of it. | |
3427 | */ | |
3428 | dup_table = *table; | |
3429 | dup_table.data = out; | |
3430 | ||
3431 | return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); | |
3432 | } | |
3433 | ||
3434 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, | |
3435 | struct ctl_table *table, int write, | |
3436 | void *buffer, size_t *length, loff_t *ppos) | |
3437 | { | |
3438 | struct hstate *h = &default_hstate; | |
3439 | unsigned long tmp = h->max_huge_pages; | |
3440 | int ret; | |
3441 | ||
3442 | if (!hugepages_supported()) | |
3443 | return -EOPNOTSUPP; | |
3444 | ||
3445 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, | |
3446 | &tmp); | |
3447 | if (ret) | |
3448 | goto out; | |
3449 | ||
3450 | if (write) | |
3451 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
3452 | NUMA_NO_NODE, tmp, *length); | |
3453 | out: | |
3454 | return ret; | |
3455 | } | |
3456 | ||
3457 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, | |
3458 | void *buffer, size_t *length, loff_t *ppos) | |
3459 | { | |
3460 | ||
3461 | return hugetlb_sysctl_handler_common(false, table, write, | |
3462 | buffer, length, ppos); | |
3463 | } | |
3464 | ||
3465 | #ifdef CONFIG_NUMA | |
3466 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
3467 | void *buffer, size_t *length, loff_t *ppos) | |
3468 | { | |
3469 | return hugetlb_sysctl_handler_common(true, table, write, | |
3470 | buffer, length, ppos); | |
3471 | } | |
3472 | #endif /* CONFIG_NUMA */ | |
3473 | ||
3474 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, | |
3475 | void *buffer, size_t *length, loff_t *ppos) | |
3476 | { | |
3477 | struct hstate *h = &default_hstate; | |
3478 | unsigned long tmp; | |
3479 | int ret; | |
3480 | ||
3481 | if (!hugepages_supported()) | |
3482 | return -EOPNOTSUPP; | |
3483 | ||
3484 | tmp = h->nr_overcommit_huge_pages; | |
3485 | ||
3486 | if (write && hstate_is_gigantic(h)) | |
3487 | return -EINVAL; | |
3488 | ||
3489 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, | |
3490 | &tmp); | |
3491 | if (ret) | |
3492 | goto out; | |
3493 | ||
3494 | if (write) { | |
3495 | spin_lock(&hugetlb_lock); | |
3496 | h->nr_overcommit_huge_pages = tmp; | |
3497 | spin_unlock(&hugetlb_lock); | |
3498 | } | |
3499 | out: | |
3500 | return ret; | |
3501 | } | |
3502 | ||
3503 | #endif /* CONFIG_SYSCTL */ | |
3504 | ||
3505 | void hugetlb_report_meminfo(struct seq_file *m) | |
3506 | { | |
3507 | struct hstate *h; | |
3508 | unsigned long total = 0; | |
3509 | ||
3510 | if (!hugepages_supported()) | |
3511 | return; | |
3512 | ||
3513 | for_each_hstate(h) { | |
3514 | unsigned long count = h->nr_huge_pages; | |
3515 | ||
3516 | total += (PAGE_SIZE << huge_page_order(h)) * count; | |
3517 | ||
3518 | if (h == &default_hstate) | |
3519 | seq_printf(m, | |
3520 | "HugePages_Total: %5lu\n" | |
3521 | "HugePages_Free: %5lu\n" | |
3522 | "HugePages_Rsvd: %5lu\n" | |
3523 | "HugePages_Surp: %5lu\n" | |
3524 | "Hugepagesize: %8lu kB\n", | |
3525 | count, | |
3526 | h->free_huge_pages, | |
3527 | h->resv_huge_pages, | |
3528 | h->surplus_huge_pages, | |
3529 | (PAGE_SIZE << huge_page_order(h)) / 1024); | |
3530 | } | |
3531 | ||
3532 | seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); | |
3533 | } | |
3534 | ||
3535 | int hugetlb_report_node_meminfo(char *buf, int len, int nid) | |
3536 | { | |
3537 | struct hstate *h = &default_hstate; | |
3538 | ||
3539 | if (!hugepages_supported()) | |
3540 | return 0; | |
3541 | ||
3542 | return sysfs_emit_at(buf, len, | |
3543 | "Node %d HugePages_Total: %5u\n" | |
3544 | "Node %d HugePages_Free: %5u\n" | |
3545 | "Node %d HugePages_Surp: %5u\n", | |
3546 | nid, h->nr_huge_pages_node[nid], | |
3547 | nid, h->free_huge_pages_node[nid], | |
3548 | nid, h->surplus_huge_pages_node[nid]); | |
3549 | } | |
3550 | ||
3551 | void hugetlb_show_meminfo(void) | |
3552 | { | |
3553 | struct hstate *h; | |
3554 | int nid; | |
3555 | ||
3556 | if (!hugepages_supported()) | |
3557 | return; | |
3558 | ||
3559 | for_each_node_state(nid, N_MEMORY) | |
3560 | for_each_hstate(h) | |
3561 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
3562 | nid, | |
3563 | h->nr_huge_pages_node[nid], | |
3564 | h->free_huge_pages_node[nid], | |
3565 | h->surplus_huge_pages_node[nid], | |
3566 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
3567 | } | |
3568 | ||
3569 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) | |
3570 | { | |
3571 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
3572 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
3573 | } | |
3574 | ||
3575 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ | |
3576 | unsigned long hugetlb_total_pages(void) | |
3577 | { | |
3578 | struct hstate *h; | |
3579 | unsigned long nr_total_pages = 0; | |
3580 | ||
3581 | for_each_hstate(h) | |
3582 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
3583 | return nr_total_pages; | |
3584 | } | |
3585 | ||
3586 | static int hugetlb_acct_memory(struct hstate *h, long delta) | |
3587 | { | |
3588 | int ret = -ENOMEM; | |
3589 | ||
3590 | spin_lock(&hugetlb_lock); | |
3591 | /* | |
3592 | * When cpuset is configured, it breaks the strict hugetlb page | |
3593 | * reservation as the accounting is done on a global variable. Such | |
3594 | * reservation is completely rubbish in the presence of cpuset because | |
3595 | * the reservation is not checked against page availability for the | |
3596 | * current cpuset. Application can still potentially OOM'ed by kernel | |
3597 | * with lack of free htlb page in cpuset that the task is in. | |
3598 | * Attempt to enforce strict accounting with cpuset is almost | |
3599 | * impossible (or too ugly) because cpuset is too fluid that | |
3600 | * task or memory node can be dynamically moved between cpusets. | |
3601 | * | |
3602 | * The change of semantics for shared hugetlb mapping with cpuset is | |
3603 | * undesirable. However, in order to preserve some of the semantics, | |
3604 | * we fall back to check against current free page availability as | |
3605 | * a best attempt and hopefully to minimize the impact of changing | |
3606 | * semantics that cpuset has. | |
3607 | * | |
3608 | * Apart from cpuset, we also have memory policy mechanism that | |
3609 | * also determines from which node the kernel will allocate memory | |
3610 | * in a NUMA system. So similar to cpuset, we also should consider | |
3611 | * the memory policy of the current task. Similar to the description | |
3612 | * above. | |
3613 | */ | |
3614 | if (delta > 0) { | |
3615 | if (gather_surplus_pages(h, delta) < 0) | |
3616 | goto out; | |
3617 | ||
3618 | if (delta > allowed_mems_nr(h)) { | |
3619 | return_unused_surplus_pages(h, delta); | |
3620 | goto out; | |
3621 | } | |
3622 | } | |
3623 | ||
3624 | ret = 0; | |
3625 | if (delta < 0) | |
3626 | return_unused_surplus_pages(h, (unsigned long) -delta); | |
3627 | ||
3628 | out: | |
3629 | spin_unlock(&hugetlb_lock); | |
3630 | return ret; | |
3631 | } | |
3632 | ||
3633 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) | |
3634 | { | |
3635 | struct resv_map *resv = vma_resv_map(vma); | |
3636 | ||
3637 | /* | |
3638 | * This new VMA should share its siblings reservation map if present. | |
3639 | * The VMA will only ever have a valid reservation map pointer where | |
3640 | * it is being copied for another still existing VMA. As that VMA | |
3641 | * has a reference to the reservation map it cannot disappear until | |
3642 | * after this open call completes. It is therefore safe to take a | |
3643 | * new reference here without additional locking. | |
3644 | */ | |
3645 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) | |
3646 | kref_get(&resv->refs); | |
3647 | } | |
3648 | ||
3649 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) | |
3650 | { | |
3651 | struct hstate *h = hstate_vma(vma); | |
3652 | struct resv_map *resv = vma_resv_map(vma); | |
3653 | struct hugepage_subpool *spool = subpool_vma(vma); | |
3654 | unsigned long reserve, start, end; | |
3655 | long gbl_reserve; | |
3656 | ||
3657 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) | |
3658 | return; | |
3659 | ||
3660 | start = vma_hugecache_offset(h, vma, vma->vm_start); | |
3661 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
3662 | ||
3663 | reserve = (end - start) - region_count(resv, start, end); | |
3664 | hugetlb_cgroup_uncharge_counter(resv, start, end); | |
3665 | if (reserve) { | |
3666 | /* | |
3667 | * Decrement reserve counts. The global reserve count may be | |
3668 | * adjusted if the subpool has a minimum size. | |
3669 | */ | |
3670 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
3671 | hugetlb_acct_memory(h, -gbl_reserve); | |
3672 | } | |
3673 | ||
3674 | kref_put(&resv->refs, resv_map_release); | |
3675 | } | |
3676 | ||
3677 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) | |
3678 | { | |
3679 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
3680 | return -EINVAL; | |
3681 | return 0; | |
3682 | } | |
3683 | ||
3684 | static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) | |
3685 | { | |
3686 | struct hstate *hstate = hstate_vma(vma); | |
3687 | ||
3688 | return 1UL << huge_page_shift(hstate); | |
3689 | } | |
3690 | ||
3691 | /* | |
3692 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
3693 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
3694 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
3695 | * this far. | |
3696 | */ | |
3697 | static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) | |
3698 | { | |
3699 | BUG(); | |
3700 | return 0; | |
3701 | } | |
3702 | ||
3703 | /* | |
3704 | * When a new function is introduced to vm_operations_struct and added | |
3705 | * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. | |
3706 | * This is because under System V memory model, mappings created via | |
3707 | * shmget/shmat with "huge page" specified are backed by hugetlbfs files, | |
3708 | * their original vm_ops are overwritten with shm_vm_ops. | |
3709 | */ | |
3710 | const struct vm_operations_struct hugetlb_vm_ops = { | |
3711 | .fault = hugetlb_vm_op_fault, | |
3712 | .open = hugetlb_vm_op_open, | |
3713 | .close = hugetlb_vm_op_close, | |
3714 | .may_split = hugetlb_vm_op_split, | |
3715 | .pagesize = hugetlb_vm_op_pagesize, | |
3716 | }; | |
3717 | ||
3718 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, | |
3719 | int writable) | |
3720 | { | |
3721 | pte_t entry; | |
3722 | ||
3723 | if (writable) { | |
3724 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, | |
3725 | vma->vm_page_prot))); | |
3726 | } else { | |
3727 | entry = huge_pte_wrprotect(mk_huge_pte(page, | |
3728 | vma->vm_page_prot)); | |
3729 | } | |
3730 | entry = pte_mkyoung(entry); | |
3731 | entry = pte_mkhuge(entry); | |
3732 | entry = arch_make_huge_pte(entry, vma, page, writable); | |
3733 | ||
3734 | return entry; | |
3735 | } | |
3736 | ||
3737 | static void set_huge_ptep_writable(struct vm_area_struct *vma, | |
3738 | unsigned long address, pte_t *ptep) | |
3739 | { | |
3740 | pte_t entry; | |
3741 | ||
3742 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); | |
3743 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) | |
3744 | update_mmu_cache(vma, address, ptep); | |
3745 | } | |
3746 | ||
3747 | bool is_hugetlb_entry_migration(pte_t pte) | |
3748 | { | |
3749 | swp_entry_t swp; | |
3750 | ||
3751 | if (huge_pte_none(pte) || pte_present(pte)) | |
3752 | return false; | |
3753 | swp = pte_to_swp_entry(pte); | |
3754 | if (is_migration_entry(swp)) | |
3755 | return true; | |
3756 | else | |
3757 | return false; | |
3758 | } | |
3759 | ||
3760 | static bool is_hugetlb_entry_hwpoisoned(pte_t pte) | |
3761 | { | |
3762 | swp_entry_t swp; | |
3763 | ||
3764 | if (huge_pte_none(pte) || pte_present(pte)) | |
3765 | return false; | |
3766 | swp = pte_to_swp_entry(pte); | |
3767 | if (is_hwpoison_entry(swp)) | |
3768 | return true; | |
3769 | else | |
3770 | return false; | |
3771 | } | |
3772 | ||
3773 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, | |
3774 | struct vm_area_struct *vma) | |
3775 | { | |
3776 | pte_t *src_pte, *dst_pte, entry, dst_entry; | |
3777 | struct page *ptepage; | |
3778 | unsigned long addr; | |
3779 | int cow; | |
3780 | struct hstate *h = hstate_vma(vma); | |
3781 | unsigned long sz = huge_page_size(h); | |
3782 | struct address_space *mapping = vma->vm_file->f_mapping; | |
3783 | struct mmu_notifier_range range; | |
3784 | int ret = 0; | |
3785 | ||
3786 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
3787 | ||
3788 | if (cow) { | |
3789 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, | |
3790 | vma->vm_start, | |
3791 | vma->vm_end); | |
3792 | mmu_notifier_invalidate_range_start(&range); | |
3793 | } else { | |
3794 | /* | |
3795 | * For shared mappings i_mmap_rwsem must be held to call | |
3796 | * huge_pte_alloc, otherwise the returned ptep could go | |
3797 | * away if part of a shared pmd and another thread calls | |
3798 | * huge_pmd_unshare. | |
3799 | */ | |
3800 | i_mmap_lock_read(mapping); | |
3801 | } | |
3802 | ||
3803 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { | |
3804 | spinlock_t *src_ptl, *dst_ptl; | |
3805 | src_pte = huge_pte_offset(src, addr, sz); | |
3806 | if (!src_pte) | |
3807 | continue; | |
3808 | dst_pte = huge_pte_alloc(dst, addr, sz); | |
3809 | if (!dst_pte) { | |
3810 | ret = -ENOMEM; | |
3811 | break; | |
3812 | } | |
3813 | ||
3814 | /* | |
3815 | * If the pagetables are shared don't copy or take references. | |
3816 | * dst_pte == src_pte is the common case of src/dest sharing. | |
3817 | * | |
3818 | * However, src could have 'unshared' and dst shares with | |
3819 | * another vma. If dst_pte !none, this implies sharing. | |
3820 | * Check here before taking page table lock, and once again | |
3821 | * after taking the lock below. | |
3822 | */ | |
3823 | dst_entry = huge_ptep_get(dst_pte); | |
3824 | if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) | |
3825 | continue; | |
3826 | ||
3827 | dst_ptl = huge_pte_lock(h, dst, dst_pte); | |
3828 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3829 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
3830 | entry = huge_ptep_get(src_pte); | |
3831 | dst_entry = huge_ptep_get(dst_pte); | |
3832 | if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { | |
3833 | /* | |
3834 | * Skip if src entry none. Also, skip in the | |
3835 | * unlikely case dst entry !none as this implies | |
3836 | * sharing with another vma. | |
3837 | */ | |
3838 | ; | |
3839 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
3840 | is_hugetlb_entry_hwpoisoned(entry))) { | |
3841 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
3842 | ||
3843 | if (is_write_migration_entry(swp_entry) && cow) { | |
3844 | /* | |
3845 | * COW mappings require pages in both | |
3846 | * parent and child to be set to read. | |
3847 | */ | |
3848 | make_migration_entry_read(&swp_entry); | |
3849 | entry = swp_entry_to_pte(swp_entry); | |
3850 | set_huge_swap_pte_at(src, addr, src_pte, | |
3851 | entry, sz); | |
3852 | } | |
3853 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); | |
3854 | } else { | |
3855 | if (cow) { | |
3856 | /* | |
3857 | * No need to notify as we are downgrading page | |
3858 | * table protection not changing it to point | |
3859 | * to a new page. | |
3860 | * | |
3861 | * See Documentation/vm/mmu_notifier.rst | |
3862 | */ | |
3863 | huge_ptep_set_wrprotect(src, addr, src_pte); | |
3864 | } | |
3865 | entry = huge_ptep_get(src_pte); | |
3866 | ptepage = pte_page(entry); | |
3867 | get_page(ptepage); | |
3868 | page_dup_rmap(ptepage, true); | |
3869 | set_huge_pte_at(dst, addr, dst_pte, entry); | |
3870 | hugetlb_count_add(pages_per_huge_page(h), dst); | |
3871 | } | |
3872 | spin_unlock(src_ptl); | |
3873 | spin_unlock(dst_ptl); | |
3874 | } | |
3875 | ||
3876 | if (cow) | |
3877 | mmu_notifier_invalidate_range_end(&range); | |
3878 | else | |
3879 | i_mmap_unlock_read(mapping); | |
3880 | ||
3881 | return ret; | |
3882 | } | |
3883 | ||
3884 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, | |
3885 | unsigned long start, unsigned long end, | |
3886 | struct page *ref_page) | |
3887 | { | |
3888 | struct mm_struct *mm = vma->vm_mm; | |
3889 | unsigned long address; | |
3890 | pte_t *ptep; | |
3891 | pte_t pte; | |
3892 | spinlock_t *ptl; | |
3893 | struct page *page; | |
3894 | struct hstate *h = hstate_vma(vma); | |
3895 | unsigned long sz = huge_page_size(h); | |
3896 | struct mmu_notifier_range range; | |
3897 | ||
3898 | WARN_ON(!is_vm_hugetlb_page(vma)); | |
3899 | BUG_ON(start & ~huge_page_mask(h)); | |
3900 | BUG_ON(end & ~huge_page_mask(h)); | |
3901 | ||
3902 | /* | |
3903 | * This is a hugetlb vma, all the pte entries should point | |
3904 | * to huge page. | |
3905 | */ | |
3906 | tlb_change_page_size(tlb, sz); | |
3907 | tlb_start_vma(tlb, vma); | |
3908 | ||
3909 | /* | |
3910 | * If sharing possible, alert mmu notifiers of worst case. | |
3911 | */ | |
3912 | mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, | |
3913 | end); | |
3914 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); | |
3915 | mmu_notifier_invalidate_range_start(&range); | |
3916 | address = start; | |
3917 | for (; address < end; address += sz) { | |
3918 | ptep = huge_pte_offset(mm, address, sz); | |
3919 | if (!ptep) | |
3920 | continue; | |
3921 | ||
3922 | ptl = huge_pte_lock(h, mm, ptep); | |
3923 | if (huge_pmd_unshare(mm, vma, &address, ptep)) { | |
3924 | spin_unlock(ptl); | |
3925 | /* | |
3926 | * We just unmapped a page of PMDs by clearing a PUD. | |
3927 | * The caller's TLB flush range should cover this area. | |
3928 | */ | |
3929 | continue; | |
3930 | } | |
3931 | ||
3932 | pte = huge_ptep_get(ptep); | |
3933 | if (huge_pte_none(pte)) { | |
3934 | spin_unlock(ptl); | |
3935 | continue; | |
3936 | } | |
3937 | ||
3938 | /* | |
3939 | * Migrating hugepage or HWPoisoned hugepage is already | |
3940 | * unmapped and its refcount is dropped, so just clear pte here. | |
3941 | */ | |
3942 | if (unlikely(!pte_present(pte))) { | |
3943 | huge_pte_clear(mm, address, ptep, sz); | |
3944 | spin_unlock(ptl); | |
3945 | continue; | |
3946 | } | |
3947 | ||
3948 | page = pte_page(pte); | |
3949 | /* | |
3950 | * If a reference page is supplied, it is because a specific | |
3951 | * page is being unmapped, not a range. Ensure the page we | |
3952 | * are about to unmap is the actual page of interest. | |
3953 | */ | |
3954 | if (ref_page) { | |
3955 | if (page != ref_page) { | |
3956 | spin_unlock(ptl); | |
3957 | continue; | |
3958 | } | |
3959 | /* | |
3960 | * Mark the VMA as having unmapped its page so that | |
3961 | * future faults in this VMA will fail rather than | |
3962 | * looking like data was lost | |
3963 | */ | |
3964 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
3965 | } | |
3966 | ||
3967 | pte = huge_ptep_get_and_clear(mm, address, ptep); | |
3968 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); | |
3969 | if (huge_pte_dirty(pte)) | |
3970 | set_page_dirty(page); | |
3971 | ||
3972 | hugetlb_count_sub(pages_per_huge_page(h), mm); | |
3973 | page_remove_rmap(page, true); | |
3974 | ||
3975 | spin_unlock(ptl); | |
3976 | tlb_remove_page_size(tlb, page, huge_page_size(h)); | |
3977 | /* | |
3978 | * Bail out after unmapping reference page if supplied | |
3979 | */ | |
3980 | if (ref_page) | |
3981 | break; | |
3982 | } | |
3983 | mmu_notifier_invalidate_range_end(&range); | |
3984 | tlb_end_vma(tlb, vma); | |
3985 | } | |
3986 | ||
3987 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, | |
3988 | struct vm_area_struct *vma, unsigned long start, | |
3989 | unsigned long end, struct page *ref_page) | |
3990 | { | |
3991 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
3992 | ||
3993 | /* | |
3994 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
3995 | * test will fail on a vma being torn down, and not grab a page table | |
3996 | * on its way out. We're lucky that the flag has such an appropriate | |
3997 | * name, and can in fact be safely cleared here. We could clear it | |
3998 | * before the __unmap_hugepage_range above, but all that's necessary | |
3999 | * is to clear it before releasing the i_mmap_rwsem. This works | |
4000 | * because in the context this is called, the VMA is about to be | |
4001 | * destroyed and the i_mmap_rwsem is held. | |
4002 | */ | |
4003 | vma->vm_flags &= ~VM_MAYSHARE; | |
4004 | } | |
4005 | ||
4006 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, | |
4007 | unsigned long end, struct page *ref_page) | |
4008 | { | |
4009 | struct mm_struct *mm; | |
4010 | struct mmu_gather tlb; | |
4011 | unsigned long tlb_start = start; | |
4012 | unsigned long tlb_end = end; | |
4013 | ||
4014 | /* | |
4015 | * If shared PMDs were possibly used within this vma range, adjust | |
4016 | * start/end for worst case tlb flushing. | |
4017 | * Note that we can not be sure if PMDs are shared until we try to | |
4018 | * unmap pages. However, we want to make sure TLB flushing covers | |
4019 | * the largest possible range. | |
4020 | */ | |
4021 | adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end); | |
4022 | ||
4023 | mm = vma->vm_mm; | |
4024 | ||
4025 | tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end); | |
4026 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); | |
4027 | tlb_finish_mmu(&tlb, tlb_start, tlb_end); | |
4028 | } | |
4029 | ||
4030 | /* | |
4031 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
4032 | * mappping it owns the reserve page for. The intention is to unmap the page | |
4033 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
4034 | * same region. | |
4035 | */ | |
4036 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, | |
4037 | struct page *page, unsigned long address) | |
4038 | { | |
4039 | struct hstate *h = hstate_vma(vma); | |
4040 | struct vm_area_struct *iter_vma; | |
4041 | struct address_space *mapping; | |
4042 | pgoff_t pgoff; | |
4043 | ||
4044 | /* | |
4045 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
4046 | * from page cache lookup which is in HPAGE_SIZE units. | |
4047 | */ | |
4048 | address = address & huge_page_mask(h); | |
4049 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + | |
4050 | vma->vm_pgoff; | |
4051 | mapping = vma->vm_file->f_mapping; | |
4052 | ||
4053 | /* | |
4054 | * Take the mapping lock for the duration of the table walk. As | |
4055 | * this mapping should be shared between all the VMAs, | |
4056 | * __unmap_hugepage_range() is called as the lock is already held | |
4057 | */ | |
4058 | i_mmap_lock_write(mapping); | |
4059 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { | |
4060 | /* Do not unmap the current VMA */ | |
4061 | if (iter_vma == vma) | |
4062 | continue; | |
4063 | ||
4064 | /* | |
4065 | * Shared VMAs have their own reserves and do not affect | |
4066 | * MAP_PRIVATE accounting but it is possible that a shared | |
4067 | * VMA is using the same page so check and skip such VMAs. | |
4068 | */ | |
4069 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
4070 | continue; | |
4071 | ||
4072 | /* | |
4073 | * Unmap the page from other VMAs without their own reserves. | |
4074 | * They get marked to be SIGKILLed if they fault in these | |
4075 | * areas. This is because a future no-page fault on this VMA | |
4076 | * could insert a zeroed page instead of the data existing | |
4077 | * from the time of fork. This would look like data corruption | |
4078 | */ | |
4079 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
4080 | unmap_hugepage_range(iter_vma, address, | |
4081 | address + huge_page_size(h), page); | |
4082 | } | |
4083 | i_mmap_unlock_write(mapping); | |
4084 | } | |
4085 | ||
4086 | /* | |
4087 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
4088 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we | |
4089 | * cannot race with other handlers or page migration. | |
4090 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
4091 | */ | |
4092 | static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, | |
4093 | unsigned long address, pte_t *ptep, | |
4094 | struct page *pagecache_page, spinlock_t *ptl) | |
4095 | { | |
4096 | pte_t pte; | |
4097 | struct hstate *h = hstate_vma(vma); | |
4098 | struct page *old_page, *new_page; | |
4099 | int outside_reserve = 0; | |
4100 | vm_fault_t ret = 0; | |
4101 | unsigned long haddr = address & huge_page_mask(h); | |
4102 | struct mmu_notifier_range range; | |
4103 | ||
4104 | pte = huge_ptep_get(ptep); | |
4105 | old_page = pte_page(pte); | |
4106 | ||
4107 | retry_avoidcopy: | |
4108 | /* If no-one else is actually using this page, avoid the copy | |
4109 | * and just make the page writable */ | |
4110 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { | |
4111 | page_move_anon_rmap(old_page, vma); | |
4112 | set_huge_ptep_writable(vma, haddr, ptep); | |
4113 | return 0; | |
4114 | } | |
4115 | ||
4116 | /* | |
4117 | * If the process that created a MAP_PRIVATE mapping is about to | |
4118 | * perform a COW due to a shared page count, attempt to satisfy | |
4119 | * the allocation without using the existing reserves. The pagecache | |
4120 | * page is used to determine if the reserve at this address was | |
4121 | * consumed or not. If reserves were used, a partial faulted mapping | |
4122 | * at the time of fork() could consume its reserves on COW instead | |
4123 | * of the full address range. | |
4124 | */ | |
4125 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && | |
4126 | old_page != pagecache_page) | |
4127 | outside_reserve = 1; | |
4128 | ||
4129 | get_page(old_page); | |
4130 | ||
4131 | /* | |
4132 | * Drop page table lock as buddy allocator may be called. It will | |
4133 | * be acquired again before returning to the caller, as expected. | |
4134 | */ | |
4135 | spin_unlock(ptl); | |
4136 | new_page = alloc_huge_page(vma, haddr, outside_reserve); | |
4137 | ||
4138 | if (IS_ERR(new_page)) { | |
4139 | /* | |
4140 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
4141 | * it is due to references held by a child and an insufficient | |
4142 | * huge page pool. To guarantee the original mappers | |
4143 | * reliability, unmap the page from child processes. The child | |
4144 | * may get SIGKILLed if it later faults. | |
4145 | */ | |
4146 | if (outside_reserve) { | |
4147 | struct address_space *mapping = vma->vm_file->f_mapping; | |
4148 | pgoff_t idx; | |
4149 | u32 hash; | |
4150 | ||
4151 | put_page(old_page); | |
4152 | BUG_ON(huge_pte_none(pte)); | |
4153 | /* | |
4154 | * Drop hugetlb_fault_mutex and i_mmap_rwsem before | |
4155 | * unmapping. unmapping needs to hold i_mmap_rwsem | |
4156 | * in write mode. Dropping i_mmap_rwsem in read mode | |
4157 | * here is OK as COW mappings do not interact with | |
4158 | * PMD sharing. | |
4159 | * | |
4160 | * Reacquire both after unmap operation. | |
4161 | */ | |
4162 | idx = vma_hugecache_offset(h, vma, haddr); | |
4163 | hash = hugetlb_fault_mutex_hash(mapping, idx); | |
4164 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
4165 | i_mmap_unlock_read(mapping); | |
4166 | ||
4167 | unmap_ref_private(mm, vma, old_page, haddr); | |
4168 | ||
4169 | i_mmap_lock_read(mapping); | |
4170 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
4171 | spin_lock(ptl); | |
4172 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); | |
4173 | if (likely(ptep && | |
4174 | pte_same(huge_ptep_get(ptep), pte))) | |
4175 | goto retry_avoidcopy; | |
4176 | /* | |
4177 | * race occurs while re-acquiring page table | |
4178 | * lock, and our job is done. | |
4179 | */ | |
4180 | return 0; | |
4181 | } | |
4182 | ||
4183 | ret = vmf_error(PTR_ERR(new_page)); | |
4184 | goto out_release_old; | |
4185 | } | |
4186 | ||
4187 | /* | |
4188 | * When the original hugepage is shared one, it does not have | |
4189 | * anon_vma prepared. | |
4190 | */ | |
4191 | if (unlikely(anon_vma_prepare(vma))) { | |
4192 | ret = VM_FAULT_OOM; | |
4193 | goto out_release_all; | |
4194 | } | |
4195 | ||
4196 | copy_user_huge_page(new_page, old_page, address, vma, | |
4197 | pages_per_huge_page(h)); | |
4198 | __SetPageUptodate(new_page); | |
4199 | ||
4200 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, | |
4201 | haddr + huge_page_size(h)); | |
4202 | mmu_notifier_invalidate_range_start(&range); | |
4203 | ||
4204 | /* | |
4205 | * Retake the page table lock to check for racing updates | |
4206 | * before the page tables are altered | |
4207 | */ | |
4208 | spin_lock(ptl); | |
4209 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); | |
4210 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { | |
4211 | ClearPagePrivate(new_page); | |
4212 | ||
4213 | /* Break COW */ | |
4214 | huge_ptep_clear_flush(vma, haddr, ptep); | |
4215 | mmu_notifier_invalidate_range(mm, range.start, range.end); | |
4216 | set_huge_pte_at(mm, haddr, ptep, | |
4217 | make_huge_pte(vma, new_page, 1)); | |
4218 | page_remove_rmap(old_page, true); | |
4219 | hugepage_add_new_anon_rmap(new_page, vma, haddr); | |
4220 | set_page_huge_active(new_page); | |
4221 | /* Make the old page be freed below */ | |
4222 | new_page = old_page; | |
4223 | } | |
4224 | spin_unlock(ptl); | |
4225 | mmu_notifier_invalidate_range_end(&range); | |
4226 | out_release_all: | |
4227 | restore_reserve_on_error(h, vma, haddr, new_page); | |
4228 | put_page(new_page); | |
4229 | out_release_old: | |
4230 | put_page(old_page); | |
4231 | ||
4232 | spin_lock(ptl); /* Caller expects lock to be held */ | |
4233 | return ret; | |
4234 | } | |
4235 | ||
4236 | /* Return the pagecache page at a given address within a VMA */ | |
4237 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, | |
4238 | struct vm_area_struct *vma, unsigned long address) | |
4239 | { | |
4240 | struct address_space *mapping; | |
4241 | pgoff_t idx; | |
4242 | ||
4243 | mapping = vma->vm_file->f_mapping; | |
4244 | idx = vma_hugecache_offset(h, vma, address); | |
4245 | ||
4246 | return find_lock_page(mapping, idx); | |
4247 | } | |
4248 | ||
4249 | /* | |
4250 | * Return whether there is a pagecache page to back given address within VMA. | |
4251 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
4252 | */ | |
4253 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
4254 | struct vm_area_struct *vma, unsigned long address) | |
4255 | { | |
4256 | struct address_space *mapping; | |
4257 | pgoff_t idx; | |
4258 | struct page *page; | |
4259 | ||
4260 | mapping = vma->vm_file->f_mapping; | |
4261 | idx = vma_hugecache_offset(h, vma, address); | |
4262 | ||
4263 | page = find_get_page(mapping, idx); | |
4264 | if (page) | |
4265 | put_page(page); | |
4266 | return page != NULL; | |
4267 | } | |
4268 | ||
4269 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, | |
4270 | pgoff_t idx) | |
4271 | { | |
4272 | struct inode *inode = mapping->host; | |
4273 | struct hstate *h = hstate_inode(inode); | |
4274 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
4275 | ||
4276 | if (err) | |
4277 | return err; | |
4278 | ClearPagePrivate(page); | |
4279 | ||
4280 | /* | |
4281 | * set page dirty so that it will not be removed from cache/file | |
4282 | * by non-hugetlbfs specific code paths. | |
4283 | */ | |
4284 | set_page_dirty(page); | |
4285 | ||
4286 | spin_lock(&inode->i_lock); | |
4287 | inode->i_blocks += blocks_per_huge_page(h); | |
4288 | spin_unlock(&inode->i_lock); | |
4289 | return 0; | |
4290 | } | |
4291 | ||
4292 | static vm_fault_t hugetlb_no_page(struct mm_struct *mm, | |
4293 | struct vm_area_struct *vma, | |
4294 | struct address_space *mapping, pgoff_t idx, | |
4295 | unsigned long address, pte_t *ptep, unsigned int flags) | |
4296 | { | |
4297 | struct hstate *h = hstate_vma(vma); | |
4298 | vm_fault_t ret = VM_FAULT_SIGBUS; | |
4299 | int anon_rmap = 0; | |
4300 | unsigned long size; | |
4301 | struct page *page; | |
4302 | pte_t new_pte; | |
4303 | spinlock_t *ptl; | |
4304 | unsigned long haddr = address & huge_page_mask(h); | |
4305 | bool new_page = false; | |
4306 | ||
4307 | /* | |
4308 | * Currently, we are forced to kill the process in the event the | |
4309 | * original mapper has unmapped pages from the child due to a failed | |
4310 | * COW. Warn that such a situation has occurred as it may not be obvious | |
4311 | */ | |
4312 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
4313 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", | |
4314 | current->pid); | |
4315 | return ret; | |
4316 | } | |
4317 | ||
4318 | /* | |
4319 | * We can not race with truncation due to holding i_mmap_rwsem. | |
4320 | * i_size is modified when holding i_mmap_rwsem, so check here | |
4321 | * once for faults beyond end of file. | |
4322 | */ | |
4323 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
4324 | if (idx >= size) | |
4325 | goto out; | |
4326 | ||
4327 | retry: | |
4328 | page = find_lock_page(mapping, idx); | |
4329 | if (!page) { | |
4330 | /* | |
4331 | * Check for page in userfault range | |
4332 | */ | |
4333 | if (userfaultfd_missing(vma)) { | |
4334 | u32 hash; | |
4335 | struct vm_fault vmf = { | |
4336 | .vma = vma, | |
4337 | .address = haddr, | |
4338 | .flags = flags, | |
4339 | /* | |
4340 | * Hard to debug if it ends up being | |
4341 | * used by a callee that assumes | |
4342 | * something about the other | |
4343 | * uninitialized fields... same as in | |
4344 | * memory.c | |
4345 | */ | |
4346 | }; | |
4347 | ||
4348 | /* | |
4349 | * hugetlb_fault_mutex and i_mmap_rwsem must be | |
4350 | * dropped before handling userfault. Reacquire | |
4351 | * after handling fault to make calling code simpler. | |
4352 | */ | |
4353 | hash = hugetlb_fault_mutex_hash(mapping, idx); | |
4354 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
4355 | i_mmap_unlock_read(mapping); | |
4356 | ret = handle_userfault(&vmf, VM_UFFD_MISSING); | |
4357 | i_mmap_lock_read(mapping); | |
4358 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
4359 | goto out; | |
4360 | } | |
4361 | ||
4362 | page = alloc_huge_page(vma, haddr, 0); | |
4363 | if (IS_ERR(page)) { | |
4364 | /* | |
4365 | * Returning error will result in faulting task being | |
4366 | * sent SIGBUS. The hugetlb fault mutex prevents two | |
4367 | * tasks from racing to fault in the same page which | |
4368 | * could result in false unable to allocate errors. | |
4369 | * Page migration does not take the fault mutex, but | |
4370 | * does a clear then write of pte's under page table | |
4371 | * lock. Page fault code could race with migration, | |
4372 | * notice the clear pte and try to allocate a page | |
4373 | * here. Before returning error, get ptl and make | |
4374 | * sure there really is no pte entry. | |
4375 | */ | |
4376 | ptl = huge_pte_lock(h, mm, ptep); | |
4377 | if (!huge_pte_none(huge_ptep_get(ptep))) { | |
4378 | ret = 0; | |
4379 | spin_unlock(ptl); | |
4380 | goto out; | |
4381 | } | |
4382 | spin_unlock(ptl); | |
4383 | ret = vmf_error(PTR_ERR(page)); | |
4384 | goto out; | |
4385 | } | |
4386 | clear_huge_page(page, address, pages_per_huge_page(h)); | |
4387 | __SetPageUptodate(page); | |
4388 | new_page = true; | |
4389 | ||
4390 | if (vma->vm_flags & VM_MAYSHARE) { | |
4391 | int err = huge_add_to_page_cache(page, mapping, idx); | |
4392 | if (err) { | |
4393 | put_page(page); | |
4394 | if (err == -EEXIST) | |
4395 | goto retry; | |
4396 | goto out; | |
4397 | } | |
4398 | } else { | |
4399 | lock_page(page); | |
4400 | if (unlikely(anon_vma_prepare(vma))) { | |
4401 | ret = VM_FAULT_OOM; | |
4402 | goto backout_unlocked; | |
4403 | } | |
4404 | anon_rmap = 1; | |
4405 | } | |
4406 | } else { | |
4407 | /* | |
4408 | * If memory error occurs between mmap() and fault, some process | |
4409 | * don't have hwpoisoned swap entry for errored virtual address. | |
4410 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
4411 | */ | |
4412 | if (unlikely(PageHWPoison(page))) { | |
4413 | ret = VM_FAULT_HWPOISON_LARGE | | |
4414 | VM_FAULT_SET_HINDEX(hstate_index(h)); | |
4415 | goto backout_unlocked; | |
4416 | } | |
4417 | } | |
4418 | ||
4419 | /* | |
4420 | * If we are going to COW a private mapping later, we examine the | |
4421 | * pending reservations for this page now. This will ensure that | |
4422 | * any allocations necessary to record that reservation occur outside | |
4423 | * the spinlock. | |
4424 | */ | |
4425 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { | |
4426 | if (vma_needs_reservation(h, vma, haddr) < 0) { | |
4427 | ret = VM_FAULT_OOM; | |
4428 | goto backout_unlocked; | |
4429 | } | |
4430 | /* Just decrements count, does not deallocate */ | |
4431 | vma_end_reservation(h, vma, haddr); | |
4432 | } | |
4433 | ||
4434 | ptl = huge_pte_lock(h, mm, ptep); | |
4435 | ret = 0; | |
4436 | if (!huge_pte_none(huge_ptep_get(ptep))) | |
4437 | goto backout; | |
4438 | ||
4439 | if (anon_rmap) { | |
4440 | ClearPagePrivate(page); | |
4441 | hugepage_add_new_anon_rmap(page, vma, haddr); | |
4442 | } else | |
4443 | page_dup_rmap(page, true); | |
4444 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) | |
4445 | && (vma->vm_flags & VM_SHARED))); | |
4446 | set_huge_pte_at(mm, haddr, ptep, new_pte); | |
4447 | ||
4448 | hugetlb_count_add(pages_per_huge_page(h), mm); | |
4449 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { | |
4450 | /* Optimization, do the COW without a second fault */ | |
4451 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); | |
4452 | } | |
4453 | ||
4454 | spin_unlock(ptl); | |
4455 | ||
4456 | /* | |
4457 | * Only make newly allocated pages active. Existing pages found | |
4458 | * in the pagecache could be !page_huge_active() if they have been | |
4459 | * isolated for migration. | |
4460 | */ | |
4461 | if (new_page) | |
4462 | set_page_huge_active(page); | |
4463 | ||
4464 | unlock_page(page); | |
4465 | out: | |
4466 | return ret; | |
4467 | ||
4468 | backout: | |
4469 | spin_unlock(ptl); | |
4470 | backout_unlocked: | |
4471 | unlock_page(page); | |
4472 | restore_reserve_on_error(h, vma, haddr, page); | |
4473 | put_page(page); | |
4474 | goto out; | |
4475 | } | |
4476 | ||
4477 | #ifdef CONFIG_SMP | |
4478 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) | |
4479 | { | |
4480 | unsigned long key[2]; | |
4481 | u32 hash; | |
4482 | ||
4483 | key[0] = (unsigned long) mapping; | |
4484 | key[1] = idx; | |
4485 | ||
4486 | hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); | |
4487 | ||
4488 | return hash & (num_fault_mutexes - 1); | |
4489 | } | |
4490 | #else | |
4491 | /* | |
4492 | * For uniprocesor systems we always use a single mutex, so just | |
4493 | * return 0 and avoid the hashing overhead. | |
4494 | */ | |
4495 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) | |
4496 | { | |
4497 | return 0; | |
4498 | } | |
4499 | #endif | |
4500 | ||
4501 | vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, | |
4502 | unsigned long address, unsigned int flags) | |
4503 | { | |
4504 | pte_t *ptep, entry; | |
4505 | spinlock_t *ptl; | |
4506 | vm_fault_t ret; | |
4507 | u32 hash; | |
4508 | pgoff_t idx; | |
4509 | struct page *page = NULL; | |
4510 | struct page *pagecache_page = NULL; | |
4511 | struct hstate *h = hstate_vma(vma); | |
4512 | struct address_space *mapping; | |
4513 | int need_wait_lock = 0; | |
4514 | unsigned long haddr = address & huge_page_mask(h); | |
4515 | ||
4516 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); | |
4517 | if (ptep) { | |
4518 | /* | |
4519 | * Since we hold no locks, ptep could be stale. That is | |
4520 | * OK as we are only making decisions based on content and | |
4521 | * not actually modifying content here. | |
4522 | */ | |
4523 | entry = huge_ptep_get(ptep); | |
4524 | if (unlikely(is_hugetlb_entry_migration(entry))) { | |
4525 | migration_entry_wait_huge(vma, mm, ptep); | |
4526 | return 0; | |
4527 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
4528 | return VM_FAULT_HWPOISON_LARGE | | |
4529 | VM_FAULT_SET_HINDEX(hstate_index(h)); | |
4530 | } | |
4531 | ||
4532 | /* | |
4533 | * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold | |
4534 | * until finished with ptep. This serves two purposes: | |
4535 | * 1) It prevents huge_pmd_unshare from being called elsewhere | |
4536 | * and making the ptep no longer valid. | |
4537 | * 2) It synchronizes us with i_size modifications during truncation. | |
4538 | * | |
4539 | * ptep could have already be assigned via huge_pte_offset. That | |
4540 | * is OK, as huge_pte_alloc will return the same value unless | |
4541 | * something has changed. | |
4542 | */ | |
4543 | mapping = vma->vm_file->f_mapping; | |
4544 | i_mmap_lock_read(mapping); | |
4545 | ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); | |
4546 | if (!ptep) { | |
4547 | i_mmap_unlock_read(mapping); | |
4548 | return VM_FAULT_OOM; | |
4549 | } | |
4550 | ||
4551 | /* | |
4552 | * Serialize hugepage allocation and instantiation, so that we don't | |
4553 | * get spurious allocation failures if two CPUs race to instantiate | |
4554 | * the same page in the page cache. | |
4555 | */ | |
4556 | idx = vma_hugecache_offset(h, vma, haddr); | |
4557 | hash = hugetlb_fault_mutex_hash(mapping, idx); | |
4558 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
4559 | ||
4560 | entry = huge_ptep_get(ptep); | |
4561 | if (huge_pte_none(entry)) { | |
4562 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); | |
4563 | goto out_mutex; | |
4564 | } | |
4565 | ||
4566 | ret = 0; | |
4567 | ||
4568 | /* | |
4569 | * entry could be a migration/hwpoison entry at this point, so this | |
4570 | * check prevents the kernel from going below assuming that we have | |
4571 | * an active hugepage in pagecache. This goto expects the 2nd page | |
4572 | * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will | |
4573 | * properly handle it. | |
4574 | */ | |
4575 | if (!pte_present(entry)) | |
4576 | goto out_mutex; | |
4577 | ||
4578 | /* | |
4579 | * If we are going to COW the mapping later, we examine the pending | |
4580 | * reservations for this page now. This will ensure that any | |
4581 | * allocations necessary to record that reservation occur outside the | |
4582 | * spinlock. For private mappings, we also lookup the pagecache | |
4583 | * page now as it is used to determine if a reservation has been | |
4584 | * consumed. | |
4585 | */ | |
4586 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { | |
4587 | if (vma_needs_reservation(h, vma, haddr) < 0) { | |
4588 | ret = VM_FAULT_OOM; | |
4589 | goto out_mutex; | |
4590 | } | |
4591 | /* Just decrements count, does not deallocate */ | |
4592 | vma_end_reservation(h, vma, haddr); | |
4593 | ||
4594 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
4595 | pagecache_page = hugetlbfs_pagecache_page(h, | |
4596 | vma, haddr); | |
4597 | } | |
4598 | ||
4599 | ptl = huge_pte_lock(h, mm, ptep); | |
4600 | ||
4601 | /* Check for a racing update before calling hugetlb_cow */ | |
4602 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
4603 | goto out_ptl; | |
4604 | ||
4605 | /* | |
4606 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
4607 | * pagecache_page, so here we need take the former one | |
4608 | * when page != pagecache_page or !pagecache_page. | |
4609 | */ | |
4610 | page = pte_page(entry); | |
4611 | if (page != pagecache_page) | |
4612 | if (!trylock_page(page)) { | |
4613 | need_wait_lock = 1; | |
4614 | goto out_ptl; | |
4615 | } | |
4616 | ||
4617 | get_page(page); | |
4618 | ||
4619 | if (flags & FAULT_FLAG_WRITE) { | |
4620 | if (!huge_pte_write(entry)) { | |
4621 | ret = hugetlb_cow(mm, vma, address, ptep, | |
4622 | pagecache_page, ptl); | |
4623 | goto out_put_page; | |
4624 | } | |
4625 | entry = huge_pte_mkdirty(entry); | |
4626 | } | |
4627 | entry = pte_mkyoung(entry); | |
4628 | if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, | |
4629 | flags & FAULT_FLAG_WRITE)) | |
4630 | update_mmu_cache(vma, haddr, ptep); | |
4631 | out_put_page: | |
4632 | if (page != pagecache_page) | |
4633 | unlock_page(page); | |
4634 | put_page(page); | |
4635 | out_ptl: | |
4636 | spin_unlock(ptl); | |
4637 | ||
4638 | if (pagecache_page) { | |
4639 | unlock_page(pagecache_page); | |
4640 | put_page(pagecache_page); | |
4641 | } | |
4642 | out_mutex: | |
4643 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
4644 | i_mmap_unlock_read(mapping); | |
4645 | /* | |
4646 | * Generally it's safe to hold refcount during waiting page lock. But | |
4647 | * here we just wait to defer the next page fault to avoid busy loop and | |
4648 | * the page is not used after unlocked before returning from the current | |
4649 | * page fault. So we are safe from accessing freed page, even if we wait | |
4650 | * here without taking refcount. | |
4651 | */ | |
4652 | if (need_wait_lock) | |
4653 | wait_on_page_locked(page); | |
4654 | return ret; | |
4655 | } | |
4656 | ||
4657 | /* | |
4658 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
4659 | * modifications for huge pages. | |
4660 | */ | |
4661 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
4662 | pte_t *dst_pte, | |
4663 | struct vm_area_struct *dst_vma, | |
4664 | unsigned long dst_addr, | |
4665 | unsigned long src_addr, | |
4666 | struct page **pagep) | |
4667 | { | |
4668 | struct address_space *mapping; | |
4669 | pgoff_t idx; | |
4670 | unsigned long size; | |
4671 | int vm_shared = dst_vma->vm_flags & VM_SHARED; | |
4672 | struct hstate *h = hstate_vma(dst_vma); | |
4673 | pte_t _dst_pte; | |
4674 | spinlock_t *ptl; | |
4675 | int ret; | |
4676 | struct page *page; | |
4677 | ||
4678 | if (!*pagep) { | |
4679 | ret = -ENOMEM; | |
4680 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
4681 | if (IS_ERR(page)) | |
4682 | goto out; | |
4683 | ||
4684 | ret = copy_huge_page_from_user(page, | |
4685 | (const void __user *) src_addr, | |
4686 | pages_per_huge_page(h), false); | |
4687 | ||
4688 | /* fallback to copy_from_user outside mmap_lock */ | |
4689 | if (unlikely(ret)) { | |
4690 | ret = -ENOENT; | |
4691 | *pagep = page; | |
4692 | /* don't free the page */ | |
4693 | goto out; | |
4694 | } | |
4695 | } else { | |
4696 | page = *pagep; | |
4697 | *pagep = NULL; | |
4698 | } | |
4699 | ||
4700 | /* | |
4701 | * The memory barrier inside __SetPageUptodate makes sure that | |
4702 | * preceding stores to the page contents become visible before | |
4703 | * the set_pte_at() write. | |
4704 | */ | |
4705 | __SetPageUptodate(page); | |
4706 | ||
4707 | mapping = dst_vma->vm_file->f_mapping; | |
4708 | idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
4709 | ||
4710 | /* | |
4711 | * If shared, add to page cache | |
4712 | */ | |
4713 | if (vm_shared) { | |
4714 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
4715 | ret = -EFAULT; | |
4716 | if (idx >= size) | |
4717 | goto out_release_nounlock; | |
4718 | ||
4719 | /* | |
4720 | * Serialization between remove_inode_hugepages() and | |
4721 | * huge_add_to_page_cache() below happens through the | |
4722 | * hugetlb_fault_mutex_table that here must be hold by | |
4723 | * the caller. | |
4724 | */ | |
4725 | ret = huge_add_to_page_cache(page, mapping, idx); | |
4726 | if (ret) | |
4727 | goto out_release_nounlock; | |
4728 | } | |
4729 | ||
4730 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); | |
4731 | spin_lock(ptl); | |
4732 | ||
4733 | /* | |
4734 | * Recheck the i_size after holding PT lock to make sure not | |
4735 | * to leave any page mapped (as page_mapped()) beyond the end | |
4736 | * of the i_size (remove_inode_hugepages() is strict about | |
4737 | * enforcing that). If we bail out here, we'll also leave a | |
4738 | * page in the radix tree in the vm_shared case beyond the end | |
4739 | * of the i_size, but remove_inode_hugepages() will take care | |
4740 | * of it as soon as we drop the hugetlb_fault_mutex_table. | |
4741 | */ | |
4742 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
4743 | ret = -EFAULT; | |
4744 | if (idx >= size) | |
4745 | goto out_release_unlock; | |
4746 | ||
4747 | ret = -EEXIST; | |
4748 | if (!huge_pte_none(huge_ptep_get(dst_pte))) | |
4749 | goto out_release_unlock; | |
4750 | ||
4751 | if (vm_shared) { | |
4752 | page_dup_rmap(page, true); | |
4753 | } else { | |
4754 | ClearPagePrivate(page); | |
4755 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); | |
4756 | } | |
4757 | ||
4758 | _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); | |
4759 | if (dst_vma->vm_flags & VM_WRITE) | |
4760 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
4761 | _dst_pte = pte_mkyoung(_dst_pte); | |
4762 | ||
4763 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); | |
4764 | ||
4765 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
4766 | dst_vma->vm_flags & VM_WRITE); | |
4767 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
4768 | ||
4769 | /* No need to invalidate - it was non-present before */ | |
4770 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
4771 | ||
4772 | spin_unlock(ptl); | |
4773 | set_page_huge_active(page); | |
4774 | if (vm_shared) | |
4775 | unlock_page(page); | |
4776 | ret = 0; | |
4777 | out: | |
4778 | return ret; | |
4779 | out_release_unlock: | |
4780 | spin_unlock(ptl); | |
4781 | if (vm_shared) | |
4782 | unlock_page(page); | |
4783 | out_release_nounlock: | |
4784 | put_page(page); | |
4785 | goto out; | |
4786 | } | |
4787 | ||
4788 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, | |
4789 | struct page **pages, struct vm_area_struct **vmas, | |
4790 | unsigned long *position, unsigned long *nr_pages, | |
4791 | long i, unsigned int flags, int *locked) | |
4792 | { | |
4793 | unsigned long pfn_offset; | |
4794 | unsigned long vaddr = *position; | |
4795 | unsigned long remainder = *nr_pages; | |
4796 | struct hstate *h = hstate_vma(vma); | |
4797 | int err = -EFAULT; | |
4798 | ||
4799 | while (vaddr < vma->vm_end && remainder) { | |
4800 | pte_t *pte; | |
4801 | spinlock_t *ptl = NULL; | |
4802 | int absent; | |
4803 | struct page *page; | |
4804 | ||
4805 | /* | |
4806 | * If we have a pending SIGKILL, don't keep faulting pages and | |
4807 | * potentially allocating memory. | |
4808 | */ | |
4809 | if (fatal_signal_pending(current)) { | |
4810 | remainder = 0; | |
4811 | break; | |
4812 | } | |
4813 | ||
4814 | /* | |
4815 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
4816 | * each hugepage. We have to make sure we get the | |
4817 | * first, for the page indexing below to work. | |
4818 | * | |
4819 | * Note that page table lock is not held when pte is null. | |
4820 | */ | |
4821 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), | |
4822 | huge_page_size(h)); | |
4823 | if (pte) | |
4824 | ptl = huge_pte_lock(h, mm, pte); | |
4825 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); | |
4826 | ||
4827 | /* | |
4828 | * When coredumping, it suits get_dump_page if we just return | |
4829 | * an error where there's an empty slot with no huge pagecache | |
4830 | * to back it. This way, we avoid allocating a hugepage, and | |
4831 | * the sparse dumpfile avoids allocating disk blocks, but its | |
4832 | * huge holes still show up with zeroes where they need to be. | |
4833 | */ | |
4834 | if (absent && (flags & FOLL_DUMP) && | |
4835 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
4836 | if (pte) | |
4837 | spin_unlock(ptl); | |
4838 | remainder = 0; | |
4839 | break; | |
4840 | } | |
4841 | ||
4842 | /* | |
4843 | * We need call hugetlb_fault for both hugepages under migration | |
4844 | * (in which case hugetlb_fault waits for the migration,) and | |
4845 | * hwpoisoned hugepages (in which case we need to prevent the | |
4846 | * caller from accessing to them.) In order to do this, we use | |
4847 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
4848 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
4849 | * both cases, and because we can't follow correct pages | |
4850 | * directly from any kind of swap entries. | |
4851 | */ | |
4852 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
4853 | ((flags & FOLL_WRITE) && | |
4854 | !huge_pte_write(huge_ptep_get(pte)))) { | |
4855 | vm_fault_t ret; | |
4856 | unsigned int fault_flags = 0; | |
4857 | ||
4858 | if (pte) | |
4859 | spin_unlock(ptl); | |
4860 | if (flags & FOLL_WRITE) | |
4861 | fault_flags |= FAULT_FLAG_WRITE; | |
4862 | if (locked) | |
4863 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
4864 | FAULT_FLAG_KILLABLE; | |
4865 | if (flags & FOLL_NOWAIT) | |
4866 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
4867 | FAULT_FLAG_RETRY_NOWAIT; | |
4868 | if (flags & FOLL_TRIED) { | |
4869 | /* | |
4870 | * Note: FAULT_FLAG_ALLOW_RETRY and | |
4871 | * FAULT_FLAG_TRIED can co-exist | |
4872 | */ | |
4873 | fault_flags |= FAULT_FLAG_TRIED; | |
4874 | } | |
4875 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
4876 | if (ret & VM_FAULT_ERROR) { | |
4877 | err = vm_fault_to_errno(ret, flags); | |
4878 | remainder = 0; | |
4879 | break; | |
4880 | } | |
4881 | if (ret & VM_FAULT_RETRY) { | |
4882 | if (locked && | |
4883 | !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) | |
4884 | *locked = 0; | |
4885 | *nr_pages = 0; | |
4886 | /* | |
4887 | * VM_FAULT_RETRY must not return an | |
4888 | * error, it will return zero | |
4889 | * instead. | |
4890 | * | |
4891 | * No need to update "position" as the | |
4892 | * caller will not check it after | |
4893 | * *nr_pages is set to 0. | |
4894 | */ | |
4895 | return i; | |
4896 | } | |
4897 | continue; | |
4898 | } | |
4899 | ||
4900 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; | |
4901 | page = pte_page(huge_ptep_get(pte)); | |
4902 | ||
4903 | /* | |
4904 | * If subpage information not requested, update counters | |
4905 | * and skip the same_page loop below. | |
4906 | */ | |
4907 | if (!pages && !vmas && !pfn_offset && | |
4908 | (vaddr + huge_page_size(h) < vma->vm_end) && | |
4909 | (remainder >= pages_per_huge_page(h))) { | |
4910 | vaddr += huge_page_size(h); | |
4911 | remainder -= pages_per_huge_page(h); | |
4912 | i += pages_per_huge_page(h); | |
4913 | spin_unlock(ptl); | |
4914 | continue; | |
4915 | } | |
4916 | ||
4917 | same_page: | |
4918 | if (pages) { | |
4919 | pages[i] = mem_map_offset(page, pfn_offset); | |
4920 | /* | |
4921 | * try_grab_page() should always succeed here, because: | |
4922 | * a) we hold the ptl lock, and b) we've just checked | |
4923 | * that the huge page is present in the page tables. If | |
4924 | * the huge page is present, then the tail pages must | |
4925 | * also be present. The ptl prevents the head page and | |
4926 | * tail pages from being rearranged in any way. So this | |
4927 | * page must be available at this point, unless the page | |
4928 | * refcount overflowed: | |
4929 | */ | |
4930 | if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) { | |
4931 | spin_unlock(ptl); | |
4932 | remainder = 0; | |
4933 | err = -ENOMEM; | |
4934 | break; | |
4935 | } | |
4936 | } | |
4937 | ||
4938 | if (vmas) | |
4939 | vmas[i] = vma; | |
4940 | ||
4941 | vaddr += PAGE_SIZE; | |
4942 | ++pfn_offset; | |
4943 | --remainder; | |
4944 | ++i; | |
4945 | if (vaddr < vma->vm_end && remainder && | |
4946 | pfn_offset < pages_per_huge_page(h)) { | |
4947 | /* | |
4948 | * We use pfn_offset to avoid touching the pageframes | |
4949 | * of this compound page. | |
4950 | */ | |
4951 | goto same_page; | |
4952 | } | |
4953 | spin_unlock(ptl); | |
4954 | } | |
4955 | *nr_pages = remainder; | |
4956 | /* | |
4957 | * setting position is actually required only if remainder is | |
4958 | * not zero but it's faster not to add a "if (remainder)" | |
4959 | * branch. | |
4960 | */ | |
4961 | *position = vaddr; | |
4962 | ||
4963 | return i ? i : err; | |
4964 | } | |
4965 | ||
4966 | #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE | |
4967 | /* | |
4968 | * ARCHes with special requirements for evicting HUGETLB backing TLB entries can | |
4969 | * implement this. | |
4970 | */ | |
4971 | #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) | |
4972 | #endif | |
4973 | ||
4974 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, | |
4975 | unsigned long address, unsigned long end, pgprot_t newprot) | |
4976 | { | |
4977 | struct mm_struct *mm = vma->vm_mm; | |
4978 | unsigned long start = address; | |
4979 | pte_t *ptep; | |
4980 | pte_t pte; | |
4981 | struct hstate *h = hstate_vma(vma); | |
4982 | unsigned long pages = 0; | |
4983 | bool shared_pmd = false; | |
4984 | struct mmu_notifier_range range; | |
4985 | ||
4986 | /* | |
4987 | * In the case of shared PMDs, the area to flush could be beyond | |
4988 | * start/end. Set range.start/range.end to cover the maximum possible | |
4989 | * range if PMD sharing is possible. | |
4990 | */ | |
4991 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, | |
4992 | 0, vma, mm, start, end); | |
4993 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); | |
4994 | ||
4995 | BUG_ON(address >= end); | |
4996 | flush_cache_range(vma, range.start, range.end); | |
4997 | ||
4998 | mmu_notifier_invalidate_range_start(&range); | |
4999 | i_mmap_lock_write(vma->vm_file->f_mapping); | |
5000 | for (; address < end; address += huge_page_size(h)) { | |
5001 | spinlock_t *ptl; | |
5002 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); | |
5003 | if (!ptep) | |
5004 | continue; | |
5005 | ptl = huge_pte_lock(h, mm, ptep); | |
5006 | if (huge_pmd_unshare(mm, vma, &address, ptep)) { | |
5007 | pages++; | |
5008 | spin_unlock(ptl); | |
5009 | shared_pmd = true; | |
5010 | continue; | |
5011 | } | |
5012 | pte = huge_ptep_get(ptep); | |
5013 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
5014 | spin_unlock(ptl); | |
5015 | continue; | |
5016 | } | |
5017 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
5018 | swp_entry_t entry = pte_to_swp_entry(pte); | |
5019 | ||
5020 | if (is_write_migration_entry(entry)) { | |
5021 | pte_t newpte; | |
5022 | ||
5023 | make_migration_entry_read(&entry); | |
5024 | newpte = swp_entry_to_pte(entry); | |
5025 | set_huge_swap_pte_at(mm, address, ptep, | |
5026 | newpte, huge_page_size(h)); | |
5027 | pages++; | |
5028 | } | |
5029 | spin_unlock(ptl); | |
5030 | continue; | |
5031 | } | |
5032 | if (!huge_pte_none(pte)) { | |
5033 | pte_t old_pte; | |
5034 | ||
5035 | old_pte = huge_ptep_modify_prot_start(vma, address, ptep); | |
5036 | pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); | |
5037 | pte = arch_make_huge_pte(pte, vma, NULL, 0); | |
5038 | huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); | |
5039 | pages++; | |
5040 | } | |
5041 | spin_unlock(ptl); | |
5042 | } | |
5043 | /* | |
5044 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare | |
5045 | * may have cleared our pud entry and done put_page on the page table: | |
5046 | * once we release i_mmap_rwsem, another task can do the final put_page | |
5047 | * and that page table be reused and filled with junk. If we actually | |
5048 | * did unshare a page of pmds, flush the range corresponding to the pud. | |
5049 | */ | |
5050 | if (shared_pmd) | |
5051 | flush_hugetlb_tlb_range(vma, range.start, range.end); | |
5052 | else | |
5053 | flush_hugetlb_tlb_range(vma, start, end); | |
5054 | /* | |
5055 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
5056 | * page table protection not changing it to point to a new page. | |
5057 | * | |
5058 | * See Documentation/vm/mmu_notifier.rst | |
5059 | */ | |
5060 | i_mmap_unlock_write(vma->vm_file->f_mapping); | |
5061 | mmu_notifier_invalidate_range_end(&range); | |
5062 | ||
5063 | return pages << h->order; | |
5064 | } | |
5065 | ||
5066 | int hugetlb_reserve_pages(struct inode *inode, | |
5067 | long from, long to, | |
5068 | struct vm_area_struct *vma, | |
5069 | vm_flags_t vm_flags) | |
5070 | { | |
5071 | long ret, chg, add = -1; | |
5072 | struct hstate *h = hstate_inode(inode); | |
5073 | struct hugepage_subpool *spool = subpool_inode(inode); | |
5074 | struct resv_map *resv_map; | |
5075 | struct hugetlb_cgroup *h_cg = NULL; | |
5076 | long gbl_reserve, regions_needed = 0; | |
5077 | ||
5078 | /* This should never happen */ | |
5079 | if (from > to) { | |
5080 | VM_WARN(1, "%s called with a negative range\n", __func__); | |
5081 | return -EINVAL; | |
5082 | } | |
5083 | ||
5084 | /* | |
5085 | * Only apply hugepage reservation if asked. At fault time, an | |
5086 | * attempt will be made for VM_NORESERVE to allocate a page | |
5087 | * without using reserves | |
5088 | */ | |
5089 | if (vm_flags & VM_NORESERVE) | |
5090 | return 0; | |
5091 | ||
5092 | /* | |
5093 | * Shared mappings base their reservation on the number of pages that | |
5094 | * are already allocated on behalf of the file. Private mappings need | |
5095 | * to reserve the full area even if read-only as mprotect() may be | |
5096 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
5097 | */ | |
5098 | if (!vma || vma->vm_flags & VM_MAYSHARE) { | |
5099 | /* | |
5100 | * resv_map can not be NULL as hugetlb_reserve_pages is only | |
5101 | * called for inodes for which resv_maps were created (see | |
5102 | * hugetlbfs_get_inode). | |
5103 | */ | |
5104 | resv_map = inode_resv_map(inode); | |
5105 | ||
5106 | chg = region_chg(resv_map, from, to, ®ions_needed); | |
5107 | ||
5108 | } else { | |
5109 | /* Private mapping. */ | |
5110 | resv_map = resv_map_alloc(); | |
5111 | if (!resv_map) | |
5112 | return -ENOMEM; | |
5113 | ||
5114 | chg = to - from; | |
5115 | ||
5116 | set_vma_resv_map(vma, resv_map); | |
5117 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
5118 | } | |
5119 | ||
5120 | if (chg < 0) { | |
5121 | ret = chg; | |
5122 | goto out_err; | |
5123 | } | |
5124 | ||
5125 | ret = hugetlb_cgroup_charge_cgroup_rsvd( | |
5126 | hstate_index(h), chg * pages_per_huge_page(h), &h_cg); | |
5127 | ||
5128 | if (ret < 0) { | |
5129 | ret = -ENOMEM; | |
5130 | goto out_err; | |
5131 | } | |
5132 | ||
5133 | if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { | |
5134 | /* For private mappings, the hugetlb_cgroup uncharge info hangs | |
5135 | * of the resv_map. | |
5136 | */ | |
5137 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); | |
5138 | } | |
5139 | ||
5140 | /* | |
5141 | * There must be enough pages in the subpool for the mapping. If | |
5142 | * the subpool has a minimum size, there may be some global | |
5143 | * reservations already in place (gbl_reserve). | |
5144 | */ | |
5145 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
5146 | if (gbl_reserve < 0) { | |
5147 | ret = -ENOSPC; | |
5148 | goto out_uncharge_cgroup; | |
5149 | } | |
5150 | ||
5151 | /* | |
5152 | * Check enough hugepages are available for the reservation. | |
5153 | * Hand the pages back to the subpool if there are not | |
5154 | */ | |
5155 | ret = hugetlb_acct_memory(h, gbl_reserve); | |
5156 | if (ret < 0) { | |
5157 | goto out_put_pages; | |
5158 | } | |
5159 | ||
5160 | /* | |
5161 | * Account for the reservations made. Shared mappings record regions | |
5162 | * that have reservations as they are shared by multiple VMAs. | |
5163 | * When the last VMA disappears, the region map says how much | |
5164 | * the reservation was and the page cache tells how much of | |
5165 | * the reservation was consumed. Private mappings are per-VMA and | |
5166 | * only the consumed reservations are tracked. When the VMA | |
5167 | * disappears, the original reservation is the VMA size and the | |
5168 | * consumed reservations are stored in the map. Hence, nothing | |
5169 | * else has to be done for private mappings here | |
5170 | */ | |
5171 | if (!vma || vma->vm_flags & VM_MAYSHARE) { | |
5172 | add = region_add(resv_map, from, to, regions_needed, h, h_cg); | |
5173 | ||
5174 | if (unlikely(add < 0)) { | |
5175 | hugetlb_acct_memory(h, -gbl_reserve); | |
5176 | ret = add; | |
5177 | goto out_put_pages; | |
5178 | } else if (unlikely(chg > add)) { | |
5179 | /* | |
5180 | * pages in this range were added to the reserve | |
5181 | * map between region_chg and region_add. This | |
5182 | * indicates a race with alloc_huge_page. Adjust | |
5183 | * the subpool and reserve counts modified above | |
5184 | * based on the difference. | |
5185 | */ | |
5186 | long rsv_adjust; | |
5187 | ||
5188 | hugetlb_cgroup_uncharge_cgroup_rsvd( | |
5189 | hstate_index(h), | |
5190 | (chg - add) * pages_per_huge_page(h), h_cg); | |
5191 | ||
5192 | rsv_adjust = hugepage_subpool_put_pages(spool, | |
5193 | chg - add); | |
5194 | hugetlb_acct_memory(h, -rsv_adjust); | |
5195 | } | |
5196 | } | |
5197 | return 0; | |
5198 | out_put_pages: | |
5199 | /* put back original number of pages, chg */ | |
5200 | (void)hugepage_subpool_put_pages(spool, chg); | |
5201 | out_uncharge_cgroup: | |
5202 | hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), | |
5203 | chg * pages_per_huge_page(h), h_cg); | |
5204 | out_err: | |
5205 | if (!vma || vma->vm_flags & VM_MAYSHARE) | |
5206 | /* Only call region_abort if the region_chg succeeded but the | |
5207 | * region_add failed or didn't run. | |
5208 | */ | |
5209 | if (chg >= 0 && add < 0) | |
5210 | region_abort(resv_map, from, to, regions_needed); | |
5211 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) | |
5212 | kref_put(&resv_map->refs, resv_map_release); | |
5213 | return ret; | |
5214 | } | |
5215 | ||
5216 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, | |
5217 | long freed) | |
5218 | { | |
5219 | struct hstate *h = hstate_inode(inode); | |
5220 | struct resv_map *resv_map = inode_resv_map(inode); | |
5221 | long chg = 0; | |
5222 | struct hugepage_subpool *spool = subpool_inode(inode); | |
5223 | long gbl_reserve; | |
5224 | ||
5225 | /* | |
5226 | * Since this routine can be called in the evict inode path for all | |
5227 | * hugetlbfs inodes, resv_map could be NULL. | |
5228 | */ | |
5229 | if (resv_map) { | |
5230 | chg = region_del(resv_map, start, end); | |
5231 | /* | |
5232 | * region_del() can fail in the rare case where a region | |
5233 | * must be split and another region descriptor can not be | |
5234 | * allocated. If end == LONG_MAX, it will not fail. | |
5235 | */ | |
5236 | if (chg < 0) | |
5237 | return chg; | |
5238 | } | |
5239 | ||
5240 | spin_lock(&inode->i_lock); | |
5241 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); | |
5242 | spin_unlock(&inode->i_lock); | |
5243 | ||
5244 | /* | |
5245 | * If the subpool has a minimum size, the number of global | |
5246 | * reservations to be released may be adjusted. | |
5247 | */ | |
5248 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
5249 | hugetlb_acct_memory(h, -gbl_reserve); | |
5250 | ||
5251 | return 0; | |
5252 | } | |
5253 | ||
5254 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE | |
5255 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
5256 | struct vm_area_struct *vma, | |
5257 | unsigned long addr, pgoff_t idx) | |
5258 | { | |
5259 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
5260 | svma->vm_start; | |
5261 | unsigned long sbase = saddr & PUD_MASK; | |
5262 | unsigned long s_end = sbase + PUD_SIZE; | |
5263 | ||
5264 | /* Allow segments to share if only one is marked locked */ | |
5265 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
5266 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
5267 | ||
5268 | /* | |
5269 | * match the virtual addresses, permission and the alignment of the | |
5270 | * page table page. | |
5271 | */ | |
5272 | if (pmd_index(addr) != pmd_index(saddr) || | |
5273 | vm_flags != svm_flags || | |
5274 | sbase < svma->vm_start || svma->vm_end < s_end) | |
5275 | return 0; | |
5276 | ||
5277 | return saddr; | |
5278 | } | |
5279 | ||
5280 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) | |
5281 | { | |
5282 | unsigned long base = addr & PUD_MASK; | |
5283 | unsigned long end = base + PUD_SIZE; | |
5284 | ||
5285 | /* | |
5286 | * check on proper vm_flags and page table alignment | |
5287 | */ | |
5288 | if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) | |
5289 | return true; | |
5290 | return false; | |
5291 | } | |
5292 | ||
5293 | /* | |
5294 | * Determine if start,end range within vma could be mapped by shared pmd. | |
5295 | * If yes, adjust start and end to cover range associated with possible | |
5296 | * shared pmd mappings. | |
5297 | */ | |
5298 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
5299 | unsigned long *start, unsigned long *end) | |
5300 | { | |
5301 | unsigned long a_start, a_end; | |
5302 | ||
5303 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
5304 | return; | |
5305 | ||
5306 | /* Extend the range to be PUD aligned for a worst case scenario */ | |
5307 | a_start = ALIGN_DOWN(*start, PUD_SIZE); | |
5308 | a_end = ALIGN(*end, PUD_SIZE); | |
5309 | ||
5310 | /* | |
5311 | * Intersect the range with the vma range, since pmd sharing won't be | |
5312 | * across vma after all | |
5313 | */ | |
5314 | *start = max(vma->vm_start, a_start); | |
5315 | *end = min(vma->vm_end, a_end); | |
5316 | } | |
5317 | ||
5318 | /* | |
5319 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
5320 | * and returns the corresponding pte. While this is not necessary for the | |
5321 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
5322 | * code much cleaner. | |
5323 | * | |
5324 | * This routine must be called with i_mmap_rwsem held in at least read mode if | |
5325 | * sharing is possible. For hugetlbfs, this prevents removal of any page | |
5326 | * table entries associated with the address space. This is important as we | |
5327 | * are setting up sharing based on existing page table entries (mappings). | |
5328 | * | |
5329 | * NOTE: This routine is only called from huge_pte_alloc. Some callers of | |
5330 | * huge_pte_alloc know that sharing is not possible and do not take | |
5331 | * i_mmap_rwsem as a performance optimization. This is handled by the | |
5332 | * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is | |
5333 | * only required for subsequent processing. | |
5334 | */ | |
5335 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
5336 | { | |
5337 | struct vm_area_struct *vma = find_vma(mm, addr); | |
5338 | struct address_space *mapping = vma->vm_file->f_mapping; | |
5339 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
5340 | vma->vm_pgoff; | |
5341 | struct vm_area_struct *svma; | |
5342 | unsigned long saddr; | |
5343 | pte_t *spte = NULL; | |
5344 | pte_t *pte; | |
5345 | spinlock_t *ptl; | |
5346 | ||
5347 | if (!vma_shareable(vma, addr)) | |
5348 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
5349 | ||
5350 | i_mmap_assert_locked(mapping); | |
5351 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { | |
5352 | if (svma == vma) | |
5353 | continue; | |
5354 | ||
5355 | saddr = page_table_shareable(svma, vma, addr, idx); | |
5356 | if (saddr) { | |
5357 | spte = huge_pte_offset(svma->vm_mm, saddr, | |
5358 | vma_mmu_pagesize(svma)); | |
5359 | if (spte) { | |
5360 | get_page(virt_to_page(spte)); | |
5361 | break; | |
5362 | } | |
5363 | } | |
5364 | } | |
5365 | ||
5366 | if (!spte) | |
5367 | goto out; | |
5368 | ||
5369 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); | |
5370 | if (pud_none(*pud)) { | |
5371 | pud_populate(mm, pud, | |
5372 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
5373 | mm_inc_nr_pmds(mm); | |
5374 | } else { | |
5375 | put_page(virt_to_page(spte)); | |
5376 | } | |
5377 | spin_unlock(ptl); | |
5378 | out: | |
5379 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
5380 | return pte; | |
5381 | } | |
5382 | ||
5383 | /* | |
5384 | * unmap huge page backed by shared pte. | |
5385 | * | |
5386 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
5387 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
5388 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
5389 | * | |
5390 | * Called with page table lock held and i_mmap_rwsem held in write mode. | |
5391 | * | |
5392 | * returns: 1 successfully unmapped a shared pte page | |
5393 | * 0 the underlying pte page is not shared, or it is the last user | |
5394 | */ | |
5395 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, | |
5396 | unsigned long *addr, pte_t *ptep) | |
5397 | { | |
5398 | pgd_t *pgd = pgd_offset(mm, *addr); | |
5399 | p4d_t *p4d = p4d_offset(pgd, *addr); | |
5400 | pud_t *pud = pud_offset(p4d, *addr); | |
5401 | ||
5402 | i_mmap_assert_write_locked(vma->vm_file->f_mapping); | |
5403 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
5404 | if (page_count(virt_to_page(ptep)) == 1) | |
5405 | return 0; | |
5406 | ||
5407 | pud_clear(pud); | |
5408 | put_page(virt_to_page(ptep)); | |
5409 | mm_dec_nr_pmds(mm); | |
5410 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; | |
5411 | return 1; | |
5412 | } | |
5413 | #define want_pmd_share() (1) | |
5414 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
5415 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
5416 | { | |
5417 | return NULL; | |
5418 | } | |
5419 | ||
5420 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, | |
5421 | unsigned long *addr, pte_t *ptep) | |
5422 | { | |
5423 | return 0; | |
5424 | } | |
5425 | ||
5426 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
5427 | unsigned long *start, unsigned long *end) | |
5428 | { | |
5429 | } | |
5430 | #define want_pmd_share() (0) | |
5431 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
5432 | ||
5433 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB | |
5434 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
5435 | unsigned long addr, unsigned long sz) | |
5436 | { | |
5437 | pgd_t *pgd; | |
5438 | p4d_t *p4d; | |
5439 | pud_t *pud; | |
5440 | pte_t *pte = NULL; | |
5441 | ||
5442 | pgd = pgd_offset(mm, addr); | |
5443 | p4d = p4d_alloc(mm, pgd, addr); | |
5444 | if (!p4d) | |
5445 | return NULL; | |
5446 | pud = pud_alloc(mm, p4d, addr); | |
5447 | if (pud) { | |
5448 | if (sz == PUD_SIZE) { | |
5449 | pte = (pte_t *)pud; | |
5450 | } else { | |
5451 | BUG_ON(sz != PMD_SIZE); | |
5452 | if (want_pmd_share() && pud_none(*pud)) | |
5453 | pte = huge_pmd_share(mm, addr, pud); | |
5454 | else | |
5455 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
5456 | } | |
5457 | } | |
5458 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); | |
5459 | ||
5460 | return pte; | |
5461 | } | |
5462 | ||
5463 | /* | |
5464 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
5465 | * entry at address @addr | |
5466 | * | |
5467 | * Return: Pointer to page table entry (PUD or PMD) for | |
5468 | * address @addr, or NULL if a !p*d_present() entry is encountered and the | |
5469 | * size @sz doesn't match the hugepage size at this level of the page | |
5470 | * table. | |
5471 | */ | |
5472 | pte_t *huge_pte_offset(struct mm_struct *mm, | |
5473 | unsigned long addr, unsigned long sz) | |
5474 | { | |
5475 | pgd_t *pgd; | |
5476 | p4d_t *p4d; | |
5477 | pud_t *pud; | |
5478 | pmd_t *pmd; | |
5479 | ||
5480 | pgd = pgd_offset(mm, addr); | |
5481 | if (!pgd_present(*pgd)) | |
5482 | return NULL; | |
5483 | p4d = p4d_offset(pgd, addr); | |
5484 | if (!p4d_present(*p4d)) | |
5485 | return NULL; | |
5486 | ||
5487 | pud = pud_offset(p4d, addr); | |
5488 | if (sz == PUD_SIZE) | |
5489 | /* must be pud huge, non-present or none */ | |
5490 | return (pte_t *)pud; | |
5491 | if (!pud_present(*pud)) | |
5492 | return NULL; | |
5493 | /* must have a valid entry and size to go further */ | |
5494 | ||
5495 | pmd = pmd_offset(pud, addr); | |
5496 | /* must be pmd huge, non-present or none */ | |
5497 | return (pte_t *)pmd; | |
5498 | } | |
5499 | ||
5500 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ | |
5501 | ||
5502 | /* | |
5503 | * These functions are overwritable if your architecture needs its own | |
5504 | * behavior. | |
5505 | */ | |
5506 | struct page * __weak | |
5507 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
5508 | int write) | |
5509 | { | |
5510 | return ERR_PTR(-EINVAL); | |
5511 | } | |
5512 | ||
5513 | struct page * __weak | |
5514 | follow_huge_pd(struct vm_area_struct *vma, | |
5515 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
5516 | { | |
5517 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
5518 | return NULL; | |
5519 | } | |
5520 | ||
5521 | struct page * __weak | |
5522 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, | |
5523 | pmd_t *pmd, int flags) | |
5524 | { | |
5525 | struct page *page = NULL; | |
5526 | spinlock_t *ptl; | |
5527 | pte_t pte; | |
5528 | ||
5529 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ | |
5530 | if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == | |
5531 | (FOLL_PIN | FOLL_GET))) | |
5532 | return NULL; | |
5533 | ||
5534 | retry: | |
5535 | ptl = pmd_lockptr(mm, pmd); | |
5536 | spin_lock(ptl); | |
5537 | /* | |
5538 | * make sure that the address range covered by this pmd is not | |
5539 | * unmapped from other threads. | |
5540 | */ | |
5541 | if (!pmd_huge(*pmd)) | |
5542 | goto out; | |
5543 | pte = huge_ptep_get((pte_t *)pmd); | |
5544 | if (pte_present(pte)) { | |
5545 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); | |
5546 | /* | |
5547 | * try_grab_page() should always succeed here, because: a) we | |
5548 | * hold the pmd (ptl) lock, and b) we've just checked that the | |
5549 | * huge pmd (head) page is present in the page tables. The ptl | |
5550 | * prevents the head page and tail pages from being rearranged | |
5551 | * in any way. So this page must be available at this point, | |
5552 | * unless the page refcount overflowed: | |
5553 | */ | |
5554 | if (WARN_ON_ONCE(!try_grab_page(page, flags))) { | |
5555 | page = NULL; | |
5556 | goto out; | |
5557 | } | |
5558 | } else { | |
5559 | if (is_hugetlb_entry_migration(pte)) { | |
5560 | spin_unlock(ptl); | |
5561 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
5562 | goto retry; | |
5563 | } | |
5564 | /* | |
5565 | * hwpoisoned entry is treated as no_page_table in | |
5566 | * follow_page_mask(). | |
5567 | */ | |
5568 | } | |
5569 | out: | |
5570 | spin_unlock(ptl); | |
5571 | return page; | |
5572 | } | |
5573 | ||
5574 | struct page * __weak | |
5575 | follow_huge_pud(struct mm_struct *mm, unsigned long address, | |
5576 | pud_t *pud, int flags) | |
5577 | { | |
5578 | if (flags & (FOLL_GET | FOLL_PIN)) | |
5579 | return NULL; | |
5580 | ||
5581 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); | |
5582 | } | |
5583 | ||
5584 | struct page * __weak | |
5585 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
5586 | { | |
5587 | if (flags & (FOLL_GET | FOLL_PIN)) | |
5588 | return NULL; | |
5589 | ||
5590 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
5591 | } | |
5592 | ||
5593 | bool isolate_huge_page(struct page *page, struct list_head *list) | |
5594 | { | |
5595 | bool ret = true; | |
5596 | ||
5597 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
5598 | spin_lock(&hugetlb_lock); | |
5599 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { | |
5600 | ret = false; | |
5601 | goto unlock; | |
5602 | } | |
5603 | clear_page_huge_active(page); | |
5604 | list_move_tail(&page->lru, list); | |
5605 | unlock: | |
5606 | spin_unlock(&hugetlb_lock); | |
5607 | return ret; | |
5608 | } | |
5609 | ||
5610 | void putback_active_hugepage(struct page *page) | |
5611 | { | |
5612 | VM_BUG_ON_PAGE(!PageHead(page), page); | |
5613 | spin_lock(&hugetlb_lock); | |
5614 | set_page_huge_active(page); | |
5615 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); | |
5616 | spin_unlock(&hugetlb_lock); | |
5617 | put_page(page); | |
5618 | } | |
5619 | ||
5620 | void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) | |
5621 | { | |
5622 | struct hstate *h = page_hstate(oldpage); | |
5623 | ||
5624 | hugetlb_cgroup_migrate(oldpage, newpage); | |
5625 | set_page_owner_migrate_reason(newpage, reason); | |
5626 | ||
5627 | /* | |
5628 | * transfer temporary state of the new huge page. This is | |
5629 | * reverse to other transitions because the newpage is going to | |
5630 | * be final while the old one will be freed so it takes over | |
5631 | * the temporary status. | |
5632 | * | |
5633 | * Also note that we have to transfer the per-node surplus state | |
5634 | * here as well otherwise the global surplus count will not match | |
5635 | * the per-node's. | |
5636 | */ | |
5637 | if (PageHugeTemporary(newpage)) { | |
5638 | int old_nid = page_to_nid(oldpage); | |
5639 | int new_nid = page_to_nid(newpage); | |
5640 | ||
5641 | SetPageHugeTemporary(oldpage); | |
5642 | ClearPageHugeTemporary(newpage); | |
5643 | ||
5644 | spin_lock(&hugetlb_lock); | |
5645 | if (h->surplus_huge_pages_node[old_nid]) { | |
5646 | h->surplus_huge_pages_node[old_nid]--; | |
5647 | h->surplus_huge_pages_node[new_nid]++; | |
5648 | } | |
5649 | spin_unlock(&hugetlb_lock); | |
5650 | } | |
5651 | } | |
5652 | ||
5653 | #ifdef CONFIG_CMA | |
5654 | static bool cma_reserve_called __initdata; | |
5655 | ||
5656 | static int __init cmdline_parse_hugetlb_cma(char *p) | |
5657 | { | |
5658 | hugetlb_cma_size = memparse(p, &p); | |
5659 | return 0; | |
5660 | } | |
5661 | ||
5662 | early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); | |
5663 | ||
5664 | void __init hugetlb_cma_reserve(int order) | |
5665 | { | |
5666 | unsigned long size, reserved, per_node; | |
5667 | int nid; | |
5668 | ||
5669 | cma_reserve_called = true; | |
5670 | ||
5671 | if (!hugetlb_cma_size) | |
5672 | return; | |
5673 | ||
5674 | if (hugetlb_cma_size < (PAGE_SIZE << order)) { | |
5675 | pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", | |
5676 | (PAGE_SIZE << order) / SZ_1M); | |
5677 | return; | |
5678 | } | |
5679 | ||
5680 | /* | |
5681 | * If 3 GB area is requested on a machine with 4 numa nodes, | |
5682 | * let's allocate 1 GB on first three nodes and ignore the last one. | |
5683 | */ | |
5684 | per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); | |
5685 | pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", | |
5686 | hugetlb_cma_size / SZ_1M, per_node / SZ_1M); | |
5687 | ||
5688 | reserved = 0; | |
5689 | for_each_node_state(nid, N_ONLINE) { | |
5690 | int res; | |
5691 | char name[CMA_MAX_NAME]; | |
5692 | ||
5693 | size = min(per_node, hugetlb_cma_size - reserved); | |
5694 | size = round_up(size, PAGE_SIZE << order); | |
5695 | ||
5696 | snprintf(name, sizeof(name), "hugetlb%d", nid); | |
5697 | res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order, | |
5698 | 0, false, name, | |
5699 | &hugetlb_cma[nid], nid); | |
5700 | if (res) { | |
5701 | pr_warn("hugetlb_cma: reservation failed: err %d, node %d", | |
5702 | res, nid); | |
5703 | continue; | |
5704 | } | |
5705 | ||
5706 | reserved += size; | |
5707 | pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", | |
5708 | size / SZ_1M, nid); | |
5709 | ||
5710 | if (reserved >= hugetlb_cma_size) | |
5711 | break; | |
5712 | } | |
5713 | } | |
5714 | ||
5715 | void __init hugetlb_cma_check(void) | |
5716 | { | |
5717 | if (!hugetlb_cma_size || cma_reserve_called) | |
5718 | return; | |
5719 | ||
5720 | pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); | |
5721 | } | |
5722 | ||
5723 | #endif /* CONFIG_CMA */ |