Loading...
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
46int hugetlb_max_hstate __read_mostly;
47unsigned int default_hstate_idx;
48struct hstate hstates[HUGE_MAX_HSTATE];
49
50#ifdef CONFIG_CMA
51static struct cma *hugetlb_cma[MAX_NUMNODES];
52#endif
53static 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 */
59static unsigned int minimum_order __read_mostly = UINT_MAX;
60
61__initdata LIST_HEAD(huge_boot_pages);
62
63/* for command line parsing */
64static struct hstate * __initdata parsed_hstate;
65static unsigned long __initdata default_hstate_max_huge_pages;
66static bool __initdata parsed_valid_hugepagesz = true;
67static 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 */
73DEFINE_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 */
79static int num_fault_mutexes;
80struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
81
82/* Forward declaration */
83static int hugetlb_acct_memory(struct hstate *h, long delta);
84
85static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
86{
87 bool free = (spool->count == 0) && (spool->used_hpages == 0);
88
89 spin_unlock(&spool->lock);
90
91 /* If no pages are used, and no other handles to the subpool
92 * remain, give up any reservations based on minimum size and
93 * free the subpool */
94 if (free) {
95 if (spool->min_hpages != -1)
96 hugetlb_acct_memory(spool->hstate,
97 -spool->min_hpages);
98 kfree(spool);
99 }
100}
101
102struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
103 long min_hpages)
104{
105 struct hugepage_subpool *spool;
106
107 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
108 if (!spool)
109 return NULL;
110
111 spin_lock_init(&spool->lock);
112 spool->count = 1;
113 spool->max_hpages = max_hpages;
114 spool->hstate = h;
115 spool->min_hpages = min_hpages;
116
117 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
118 kfree(spool);
119 return NULL;
120 }
121 spool->rsv_hpages = min_hpages;
122
123 return spool;
124}
125
126void hugepage_put_subpool(struct hugepage_subpool *spool)
127{
128 spin_lock(&spool->lock);
129 BUG_ON(!spool->count);
130 spool->count--;
131 unlock_or_release_subpool(spool);
132}
133
134/*
135 * Subpool accounting for allocating and reserving pages.
136 * Return -ENOMEM if there are not enough resources to satisfy the
137 * request. Otherwise, return the number of pages by which the
138 * global pools must be adjusted (upward). The returned value may
139 * only be different than the passed value (delta) in the case where
140 * a subpool minimum size must be maintained.
141 */
142static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
143 long delta)
144{
145 long ret = delta;
146
147 if (!spool)
148 return ret;
149
150 spin_lock(&spool->lock);
151
152 if (spool->max_hpages != -1) { /* maximum size accounting */
153 if ((spool->used_hpages + delta) <= spool->max_hpages)
154 spool->used_hpages += delta;
155 else {
156 ret = -ENOMEM;
157 goto unlock_ret;
158 }
159 }
160
161 /* minimum size accounting */
162 if (spool->min_hpages != -1 && spool->rsv_hpages) {
163 if (delta > spool->rsv_hpages) {
164 /*
165 * Asking for more reserves than those already taken on
166 * behalf of subpool. Return difference.
167 */
168 ret = delta - spool->rsv_hpages;
169 spool->rsv_hpages = 0;
170 } else {
171 ret = 0; /* reserves already accounted for */
172 spool->rsv_hpages -= delta;
173 }
174 }
175
176unlock_ret:
177 spin_unlock(&spool->lock);
178 return ret;
179}
180
181/*
182 * Subpool accounting for freeing and unreserving pages.
183 * Return the number of global page reservations that must be dropped.
184 * The return value may only be different than the passed value (delta)
185 * in the case where a subpool minimum size must be maintained.
186 */
187static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
188 long delta)
189{
190 long ret = delta;
191
192 if (!spool)
193 return delta;
194
195 spin_lock(&spool->lock);
196
197 if (spool->max_hpages != -1) /* maximum size accounting */
198 spool->used_hpages -= delta;
199
200 /* minimum size accounting */
201 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
202 if (spool->rsv_hpages + delta <= spool->min_hpages)
203 ret = 0;
204 else
205 ret = spool->rsv_hpages + delta - spool->min_hpages;
206
207 spool->rsv_hpages += delta;
208 if (spool->rsv_hpages > spool->min_hpages)
209 spool->rsv_hpages = spool->min_hpages;
210 }
211
212 /*
213 * If hugetlbfs_put_super couldn't free spool due to an outstanding
214 * quota reference, free it now.
215 */
216 unlock_or_release_subpool(spool);
217
218 return ret;
219}
220
221static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
222{
223 return HUGETLBFS_SB(inode->i_sb)->spool;
224}
225
226static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
227{
228 return subpool_inode(file_inode(vma->vm_file));
229}
230
231/* Helper that removes a struct file_region from the resv_map cache and returns
232 * it for use.
233 */
234static struct file_region *
235get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
236{
237 struct file_region *nrg = NULL;
238
239 VM_BUG_ON(resv->region_cache_count <= 0);
240
241 resv->region_cache_count--;
242 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
243 VM_BUG_ON(!nrg);
244 list_del(&nrg->link);
245
246 nrg->from = from;
247 nrg->to = to;
248
249 return nrg;
250}
251
252static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
253 struct file_region *rg)
254{
255#ifdef CONFIG_CGROUP_HUGETLB
256 nrg->reservation_counter = rg->reservation_counter;
257 nrg->css = rg->css;
258 if (rg->css)
259 css_get(rg->css);
260#endif
261}
262
263/* Helper that records hugetlb_cgroup uncharge info. */
264static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
265 struct hstate *h,
266 struct resv_map *resv,
267 struct file_region *nrg)
268{
269#ifdef CONFIG_CGROUP_HUGETLB
270 if (h_cg) {
271 nrg->reservation_counter =
272 &h_cg->rsvd_hugepage[hstate_index(h)];
273 nrg->css = &h_cg->css;
274 if (!resv->pages_per_hpage)
275 resv->pages_per_hpage = pages_per_huge_page(h);
276 /* pages_per_hpage should be the same for all entries in
277 * a resv_map.
278 */
279 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
280 } else {
281 nrg->reservation_counter = NULL;
282 nrg->css = NULL;
283 }
284#endif
285}
286
287static bool has_same_uncharge_info(struct file_region *rg,
288 struct file_region *org)
289{
290#ifdef CONFIG_CGROUP_HUGETLB
291 return rg && org &&
292 rg->reservation_counter == org->reservation_counter &&
293 rg->css == org->css;
294
295#else
296 return true;
297#endif
298}
299
300static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
301{
302 struct file_region *nrg = NULL, *prg = NULL;
303
304 prg = list_prev_entry(rg, link);
305 if (&prg->link != &resv->regions && prg->to == rg->from &&
306 has_same_uncharge_info(prg, rg)) {
307 prg->to = rg->to;
308
309 list_del(&rg->link);
310 kfree(rg);
311
312 coalesce_file_region(resv, prg);
313 return;
314 }
315
316 nrg = list_next_entry(rg, link);
317 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
318 has_same_uncharge_info(nrg, rg)) {
319 nrg->from = rg->from;
320
321 list_del(&rg->link);
322 kfree(rg);
323
324 coalesce_file_region(resv, nrg);
325 return;
326 }
327}
328
329/* Must be called with resv->lock held. Calling this with count_only == true
330 * will count the number of pages to be added but will not modify the linked
331 * list. If regions_needed != NULL and count_only == true, then regions_needed
332 * will indicate the number of file_regions needed in the cache to carry out to
333 * add the regions for this range.
334 */
335static long add_reservation_in_range(struct resv_map *resv, long f, long t,
336 struct hugetlb_cgroup *h_cg,
337 struct hstate *h, long *regions_needed,
338 bool count_only)
339{
340 long add = 0;
341 struct list_head *head = &resv->regions;
342 long last_accounted_offset = f;
343 struct file_region *rg = NULL, *trg = NULL, *nrg = NULL;
344
345 if (regions_needed)
346 *regions_needed = 0;
347
348 /* In this loop, we essentially handle an entry for the range
349 * [last_accounted_offset, rg->from), at every iteration, with some
350 * bounds checking.
351 */
352 list_for_each_entry_safe(rg, trg, head, link) {
353 /* Skip irrelevant regions that start before our range. */
354 if (rg->from < f) {
355 /* If this region ends after the last accounted offset,
356 * then we need to update last_accounted_offset.
357 */
358 if (rg->to > last_accounted_offset)
359 last_accounted_offset = rg->to;
360 continue;
361 }
362
363 /* When we find a region that starts beyond our range, we've
364 * finished.
365 */
366 if (rg->from > t)
367 break;
368
369 /* Add an entry for last_accounted_offset -> rg->from, and
370 * update last_accounted_offset.
371 */
372 if (rg->from > last_accounted_offset) {
373 add += rg->from - last_accounted_offset;
374 if (!count_only) {
375 nrg = get_file_region_entry_from_cache(
376 resv, last_accounted_offset, rg->from);
377 record_hugetlb_cgroup_uncharge_info(h_cg, h,
378 resv, nrg);
379 list_add(&nrg->link, rg->link.prev);
380 coalesce_file_region(resv, nrg);
381 } else if (regions_needed)
382 *regions_needed += 1;
383 }
384
385 last_accounted_offset = rg->to;
386 }
387
388 /* Handle the case where our range extends beyond
389 * last_accounted_offset.
390 */
391 if (last_accounted_offset < t) {
392 add += t - last_accounted_offset;
393 if (!count_only) {
394 nrg = get_file_region_entry_from_cache(
395 resv, last_accounted_offset, t);
396 record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg);
397 list_add(&nrg->link, rg->link.prev);
398 coalesce_file_region(resv, nrg);
399 } else if (regions_needed)
400 *regions_needed += 1;
401 }
402
403 VM_BUG_ON(add < 0);
404 return add;
405}
406
407/* Must be called with resv->lock acquired. Will drop lock to allocate entries.
408 */
409static int allocate_file_region_entries(struct resv_map *resv,
410 int regions_needed)
411 __must_hold(&resv->lock)
412{
413 struct list_head allocated_regions;
414 int to_allocate = 0, i = 0;
415 struct file_region *trg = NULL, *rg = NULL;
416
417 VM_BUG_ON(regions_needed < 0);
418
419 INIT_LIST_HEAD(&allocated_regions);
420
421 /*
422 * Check for sufficient descriptors in the cache to accommodate
423 * the number of in progress add operations plus regions_needed.
424 *
425 * This is a while loop because when we drop the lock, some other call
426 * to region_add or region_del may have consumed some region_entries,
427 * so we keep looping here until we finally have enough entries for
428 * (adds_in_progress + regions_needed).
429 */
430 while (resv->region_cache_count <
431 (resv->adds_in_progress + regions_needed)) {
432 to_allocate = resv->adds_in_progress + regions_needed -
433 resv->region_cache_count;
434
435 /* At this point, we should have enough entries in the cache
436 * for all the existings adds_in_progress. We should only be
437 * needing to allocate for regions_needed.
438 */
439 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
440
441 spin_unlock(&resv->lock);
442 for (i = 0; i < to_allocate; i++) {
443 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
444 if (!trg)
445 goto out_of_memory;
446 list_add(&trg->link, &allocated_regions);
447 }
448
449 spin_lock(&resv->lock);
450
451 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
452 list_del(&rg->link);
453 list_add(&rg->link, &resv->region_cache);
454 resv->region_cache_count++;
455 }
456 }
457
458 return 0;
459
460out_of_memory:
461 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
462 list_del(&rg->link);
463 kfree(rg);
464 }
465 return -ENOMEM;
466}
467
468/*
469 * Add the huge page range represented by [f, t) to the reserve
470 * map. Regions will be taken from the cache to fill in this range.
471 * Sufficient regions should exist in the cache due to the previous
472 * call to region_chg with the same range, but in some cases the cache will not
473 * have sufficient entries due to races with other code doing region_add or
474 * region_del. The extra needed entries will be allocated.
475 *
476 * regions_needed is the out value provided by a previous call to region_chg.
477 *
478 * Return the number of new huge pages added to the map. This number is greater
479 * than or equal to zero. If file_region entries needed to be allocated for
480 * this operation and we were not able to allocate, it returns -ENOMEM.
481 * region_add of regions of length 1 never allocate file_regions and cannot
482 * fail; region_chg will always allocate at least 1 entry and a region_add for
483 * 1 page will only require at most 1 entry.
484 */
485static long region_add(struct resv_map *resv, long f, long t,
486 long in_regions_needed, struct hstate *h,
487 struct hugetlb_cgroup *h_cg)
488{
489 long add = 0, actual_regions_needed = 0;
490
491 spin_lock(&resv->lock);
492retry:
493
494 /* Count how many regions are actually needed to execute this add. */
495 add_reservation_in_range(resv, f, t, NULL, NULL, &actual_regions_needed,
496 true);
497
498 /*
499 * Check for sufficient descriptors in the cache to accommodate
500 * this add operation. Note that actual_regions_needed may be greater
501 * than in_regions_needed, as the resv_map may have been modified since
502 * the region_chg call. In this case, we need to make sure that we
503 * allocate extra entries, such that we have enough for all the
504 * existing adds_in_progress, plus the excess needed for this
505 * operation.
506 */
507 if (actual_regions_needed > in_regions_needed &&
508 resv->region_cache_count <
509 resv->adds_in_progress +
510 (actual_regions_needed - in_regions_needed)) {
511 /* region_add operation of range 1 should never need to
512 * allocate file_region entries.
513 */
514 VM_BUG_ON(t - f <= 1);
515
516 if (allocate_file_region_entries(
517 resv, actual_regions_needed - in_regions_needed)) {
518 return -ENOMEM;
519 }
520
521 goto retry;
522 }
523
524 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL, false);
525
526 resv->adds_in_progress -= in_regions_needed;
527
528 spin_unlock(&resv->lock);
529 VM_BUG_ON(add < 0);
530 return add;
531}
532
533/*
534 * Examine the existing reserve map and determine how many
535 * huge pages in the specified range [f, t) are NOT currently
536 * represented. This routine is called before a subsequent
537 * call to region_add that will actually modify the reserve
538 * map to add the specified range [f, t). region_chg does
539 * not change the number of huge pages represented by the
540 * map. A number of new file_region structures is added to the cache as a
541 * placeholder, for the subsequent region_add call to use. At least 1
542 * file_region structure is added.
543 *
544 * out_regions_needed is the number of regions added to the
545 * resv->adds_in_progress. This value needs to be provided to a follow up call
546 * to region_add or region_abort for proper accounting.
547 *
548 * Returns the number of huge pages that need to be added to the existing
549 * reservation map for the range [f, t). This number is greater or equal to
550 * zero. -ENOMEM is returned if a new file_region structure or cache entry
551 * is needed and can not be allocated.
552 */
553static long region_chg(struct resv_map *resv, long f, long t,
554 long *out_regions_needed)
555{
556 long chg = 0;
557
558 spin_lock(&resv->lock);
559
560 /* Count how many hugepages in this range are NOT respresented. */
561 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
562 out_regions_needed, true);
563
564 if (*out_regions_needed == 0)
565 *out_regions_needed = 1;
566
567 if (allocate_file_region_entries(resv, *out_regions_needed))
568 return -ENOMEM;
569
570 resv->adds_in_progress += *out_regions_needed;
571
572 spin_unlock(&resv->lock);
573 return chg;
574}
575
576/*
577 * Abort the in progress add operation. The adds_in_progress field
578 * of the resv_map keeps track of the operations in progress between
579 * calls to region_chg and region_add. Operations are sometimes
580 * aborted after the call to region_chg. In such cases, region_abort
581 * is called to decrement the adds_in_progress counter. regions_needed
582 * is the value returned by the region_chg call, it is used to decrement
583 * the adds_in_progress counter.
584 *
585 * NOTE: The range arguments [f, t) are not needed or used in this
586 * routine. They are kept to make reading the calling code easier as
587 * arguments will match the associated region_chg call.
588 */
589static void region_abort(struct resv_map *resv, long f, long t,
590 long regions_needed)
591{
592 spin_lock(&resv->lock);
593 VM_BUG_ON(!resv->region_cache_count);
594 resv->adds_in_progress -= regions_needed;
595 spin_unlock(&resv->lock);
596}
597
598/*
599 * Delete the specified range [f, t) from the reserve map. If the
600 * t parameter is LONG_MAX, this indicates that ALL regions after f
601 * should be deleted. Locate the regions which intersect [f, t)
602 * and either trim, delete or split the existing regions.
603 *
604 * Returns the number of huge pages deleted from the reserve map.
605 * In the normal case, the return value is zero or more. In the
606 * case where a region must be split, a new region descriptor must
607 * be allocated. If the allocation fails, -ENOMEM will be returned.
608 * NOTE: If the parameter t == LONG_MAX, then we will never split
609 * a region and possibly return -ENOMEM. Callers specifying
610 * t == LONG_MAX do not need to check for -ENOMEM error.
611 */
612static long region_del(struct resv_map *resv, long f, long t)
613{
614 struct list_head *head = &resv->regions;
615 struct file_region *rg, *trg;
616 struct file_region *nrg = NULL;
617 long del = 0;
618
619retry:
620 spin_lock(&resv->lock);
621 list_for_each_entry_safe(rg, trg, head, link) {
622 /*
623 * Skip regions before the range to be deleted. file_region
624 * ranges are normally of the form [from, to). However, there
625 * may be a "placeholder" entry in the map which is of the form
626 * (from, to) with from == to. Check for placeholder entries
627 * at the beginning of the range to be deleted.
628 */
629 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
630 continue;
631
632 if (rg->from >= t)
633 break;
634
635 if (f > rg->from && t < rg->to) { /* Must split region */
636 /*
637 * Check for an entry in the cache before dropping
638 * lock and attempting allocation.
639 */
640 if (!nrg &&
641 resv->region_cache_count > resv->adds_in_progress) {
642 nrg = list_first_entry(&resv->region_cache,
643 struct file_region,
644 link);
645 list_del(&nrg->link);
646 resv->region_cache_count--;
647 }
648
649 if (!nrg) {
650 spin_unlock(&resv->lock);
651 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
652 if (!nrg)
653 return -ENOMEM;
654 goto retry;
655 }
656
657 del += t - f;
658
659 /* New entry for end of split region */
660 nrg->from = t;
661 nrg->to = rg->to;
662
663 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
664
665 INIT_LIST_HEAD(&nrg->link);
666
667 /* Original entry is trimmed */
668 rg->to = f;
669
670 hugetlb_cgroup_uncharge_file_region(
671 resv, rg, nrg->to - nrg->from);
672
673 list_add(&nrg->link, &rg->link);
674 nrg = NULL;
675 break;
676 }
677
678 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
679 del += rg->to - rg->from;
680 hugetlb_cgroup_uncharge_file_region(resv, rg,
681 rg->to - rg->from);
682 list_del(&rg->link);
683 kfree(rg);
684 continue;
685 }
686
687 if (f <= rg->from) { /* Trim beginning of region */
688 del += t - rg->from;
689 rg->from = t;
690
691 hugetlb_cgroup_uncharge_file_region(resv, rg,
692 t - rg->from);
693 } else { /* Trim end of region */
694 del += rg->to - f;
695 rg->to = f;
696
697 hugetlb_cgroup_uncharge_file_region(resv, rg,
698 rg->to - f);
699 }
700 }
701
702 spin_unlock(&resv->lock);
703 kfree(nrg);
704 return del;
705}
706
707/*
708 * A rare out of memory error was encountered which prevented removal of
709 * the reserve map region for a page. The huge page itself was free'ed
710 * and removed from the page cache. This routine will adjust the subpool
711 * usage count, and the global reserve count if needed. By incrementing
712 * these counts, the reserve map entry which could not be deleted will
713 * appear as a "reserved" entry instead of simply dangling with incorrect
714 * counts.
715 */
716void hugetlb_fix_reserve_counts(struct inode *inode)
717{
718 struct hugepage_subpool *spool = subpool_inode(inode);
719 long rsv_adjust;
720
721 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
722 if (rsv_adjust) {
723 struct hstate *h = hstate_inode(inode);
724
725 hugetlb_acct_memory(h, 1);
726 }
727}
728
729/*
730 * Count and return the number of huge pages in the reserve map
731 * that intersect with the range [f, t).
732 */
733static long region_count(struct resv_map *resv, long f, long t)
734{
735 struct list_head *head = &resv->regions;
736 struct file_region *rg;
737 long chg = 0;
738
739 spin_lock(&resv->lock);
740 /* Locate each segment we overlap with, and count that overlap. */
741 list_for_each_entry(rg, head, link) {
742 long seg_from;
743 long seg_to;
744
745 if (rg->to <= f)
746 continue;
747 if (rg->from >= t)
748 break;
749
750 seg_from = max(rg->from, f);
751 seg_to = min(rg->to, t);
752
753 chg += seg_to - seg_from;
754 }
755 spin_unlock(&resv->lock);
756
757 return chg;
758}
759
760/*
761 * Convert the address within this vma to the page offset within
762 * the mapping, in pagecache page units; huge pages here.
763 */
764static pgoff_t vma_hugecache_offset(struct hstate *h,
765 struct vm_area_struct *vma, unsigned long address)
766{
767 return ((address - vma->vm_start) >> huge_page_shift(h)) +
768 (vma->vm_pgoff >> huge_page_order(h));
769}
770
771pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
772 unsigned long address)
773{
774 return vma_hugecache_offset(hstate_vma(vma), vma, address);
775}
776EXPORT_SYMBOL_GPL(linear_hugepage_index);
777
778/*
779 * Return the size of the pages allocated when backing a VMA. In the majority
780 * cases this will be same size as used by the page table entries.
781 */
782unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
783{
784 if (vma->vm_ops && vma->vm_ops->pagesize)
785 return vma->vm_ops->pagesize(vma);
786 return PAGE_SIZE;
787}
788EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
789
790/*
791 * Return the page size being used by the MMU to back a VMA. In the majority
792 * of cases, the page size used by the kernel matches the MMU size. On
793 * architectures where it differs, an architecture-specific 'strong'
794 * version of this symbol is required.
795 */
796__weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
797{
798 return vma_kernel_pagesize(vma);
799}
800
801/*
802 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
803 * bits of the reservation map pointer, which are always clear due to
804 * alignment.
805 */
806#define HPAGE_RESV_OWNER (1UL << 0)
807#define HPAGE_RESV_UNMAPPED (1UL << 1)
808#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
809
810/*
811 * These helpers are used to track how many pages are reserved for
812 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
813 * is guaranteed to have their future faults succeed.
814 *
815 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
816 * the reserve counters are updated with the hugetlb_lock held. It is safe
817 * to reset the VMA at fork() time as it is not in use yet and there is no
818 * chance of the global counters getting corrupted as a result of the values.
819 *
820 * The private mapping reservation is represented in a subtly different
821 * manner to a shared mapping. A shared mapping has a region map associated
822 * with the underlying file, this region map represents the backing file
823 * pages which have ever had a reservation assigned which this persists even
824 * after the page is instantiated. A private mapping has a region map
825 * associated with the original mmap which is attached to all VMAs which
826 * reference it, this region map represents those offsets which have consumed
827 * reservation ie. where pages have been instantiated.
828 */
829static unsigned long get_vma_private_data(struct vm_area_struct *vma)
830{
831 return (unsigned long)vma->vm_private_data;
832}
833
834static void set_vma_private_data(struct vm_area_struct *vma,
835 unsigned long value)
836{
837 vma->vm_private_data = (void *)value;
838}
839
840static void
841resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
842 struct hugetlb_cgroup *h_cg,
843 struct hstate *h)
844{
845#ifdef CONFIG_CGROUP_HUGETLB
846 if (!h_cg || !h) {
847 resv_map->reservation_counter = NULL;
848 resv_map->pages_per_hpage = 0;
849 resv_map->css = NULL;
850 } else {
851 resv_map->reservation_counter =
852 &h_cg->rsvd_hugepage[hstate_index(h)];
853 resv_map->pages_per_hpage = pages_per_huge_page(h);
854 resv_map->css = &h_cg->css;
855 }
856#endif
857}
858
859struct resv_map *resv_map_alloc(void)
860{
861 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
862 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
863
864 if (!resv_map || !rg) {
865 kfree(resv_map);
866 kfree(rg);
867 return NULL;
868 }
869
870 kref_init(&resv_map->refs);
871 spin_lock_init(&resv_map->lock);
872 INIT_LIST_HEAD(&resv_map->regions);
873
874 resv_map->adds_in_progress = 0;
875 /*
876 * Initialize these to 0. On shared mappings, 0's here indicate these
877 * fields don't do cgroup accounting. On private mappings, these will be
878 * re-initialized to the proper values, to indicate that hugetlb cgroup
879 * reservations are to be un-charged from here.
880 */
881 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
882
883 INIT_LIST_HEAD(&resv_map->region_cache);
884 list_add(&rg->link, &resv_map->region_cache);
885 resv_map->region_cache_count = 1;
886
887 return resv_map;
888}
889
890void resv_map_release(struct kref *ref)
891{
892 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
893 struct list_head *head = &resv_map->region_cache;
894 struct file_region *rg, *trg;
895
896 /* Clear out any active regions before we release the map. */
897 region_del(resv_map, 0, LONG_MAX);
898
899 /* ... and any entries left in the cache */
900 list_for_each_entry_safe(rg, trg, head, link) {
901 list_del(&rg->link);
902 kfree(rg);
903 }
904
905 VM_BUG_ON(resv_map->adds_in_progress);
906
907 kfree(resv_map);
908}
909
910static inline struct resv_map *inode_resv_map(struct inode *inode)
911{
912 /*
913 * At inode evict time, i_mapping may not point to the original
914 * address space within the inode. This original address space
915 * contains the pointer to the resv_map. So, always use the
916 * address space embedded within the inode.
917 * The VERY common case is inode->mapping == &inode->i_data but,
918 * this may not be true for device special inodes.
919 */
920 return (struct resv_map *)(&inode->i_data)->private_data;
921}
922
923static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
924{
925 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
926 if (vma->vm_flags & VM_MAYSHARE) {
927 struct address_space *mapping = vma->vm_file->f_mapping;
928 struct inode *inode = mapping->host;
929
930 return inode_resv_map(inode);
931
932 } else {
933 return (struct resv_map *)(get_vma_private_data(vma) &
934 ~HPAGE_RESV_MASK);
935 }
936}
937
938static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
939{
940 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
941 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
942
943 set_vma_private_data(vma, (get_vma_private_data(vma) &
944 HPAGE_RESV_MASK) | (unsigned long)map);
945}
946
947static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
948{
949 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
950 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
951
952 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
953}
954
955static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
956{
957 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
958
959 return (get_vma_private_data(vma) & flag) != 0;
960}
961
962/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
963void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
964{
965 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
966 if (!(vma->vm_flags & VM_MAYSHARE))
967 vma->vm_private_data = (void *)0;
968}
969
970/* Returns true if the VMA has associated reserve pages */
971static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
972{
973 if (vma->vm_flags & VM_NORESERVE) {
974 /*
975 * This address is already reserved by other process(chg == 0),
976 * so, we should decrement reserved count. Without decrementing,
977 * reserve count remains after releasing inode, because this
978 * allocated page will go into page cache and is regarded as
979 * coming from reserved pool in releasing step. Currently, we
980 * don't have any other solution to deal with this situation
981 * properly, so add work-around here.
982 */
983 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
984 return true;
985 else
986 return false;
987 }
988
989 /* Shared mappings always use reserves */
990 if (vma->vm_flags & VM_MAYSHARE) {
991 /*
992 * We know VM_NORESERVE is not set. Therefore, there SHOULD
993 * be a region map for all pages. The only situation where
994 * there is no region map is if a hole was punched via
995 * fallocate. In this case, there really are no reserves to
996 * use. This situation is indicated if chg != 0.
997 */
998 if (chg)
999 return false;
1000 else
1001 return true;
1002 }
1003
1004 /*
1005 * Only the process that called mmap() has reserves for
1006 * private mappings.
1007 */
1008 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1009 /*
1010 * Like the shared case above, a hole punch or truncate
1011 * could have been performed on the private mapping.
1012 * Examine the value of chg to determine if reserves
1013 * actually exist or were previously consumed.
1014 * Very Subtle - The value of chg comes from a previous
1015 * call to vma_needs_reserves(). The reserve map for
1016 * private mappings has different (opposite) semantics
1017 * than that of shared mappings. vma_needs_reserves()
1018 * has already taken this difference in semantics into
1019 * account. Therefore, the meaning of chg is the same
1020 * as in the shared case above. Code could easily be
1021 * combined, but keeping it separate draws attention to
1022 * subtle differences.
1023 */
1024 if (chg)
1025 return false;
1026 else
1027 return true;
1028 }
1029
1030 return false;
1031}
1032
1033static void enqueue_huge_page(struct hstate *h, struct page *page)
1034{
1035 int nid = page_to_nid(page);
1036 list_move(&page->lru, &h->hugepage_freelists[nid]);
1037 h->free_huge_pages++;
1038 h->free_huge_pages_node[nid]++;
1039}
1040
1041static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1042{
1043 struct page *page;
1044 bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA);
1045
1046 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1047 if (nocma && is_migrate_cma_page(page))
1048 continue;
1049
1050 if (!PageHWPoison(page))
1051 break;
1052 }
1053
1054 /*
1055 * if 'non-isolated free hugepage' not found on the list,
1056 * the allocation fails.
1057 */
1058 if (&h->hugepage_freelists[nid] == &page->lru)
1059 return NULL;
1060 list_move(&page->lru, &h->hugepage_activelist);
1061 set_page_refcounted(page);
1062 h->free_huge_pages--;
1063 h->free_huge_pages_node[nid]--;
1064 return page;
1065}
1066
1067static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1068 nodemask_t *nmask)
1069{
1070 unsigned int cpuset_mems_cookie;
1071 struct zonelist *zonelist;
1072 struct zone *zone;
1073 struct zoneref *z;
1074 int node = NUMA_NO_NODE;
1075
1076 zonelist = node_zonelist(nid, gfp_mask);
1077
1078retry_cpuset:
1079 cpuset_mems_cookie = read_mems_allowed_begin();
1080 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1081 struct page *page;
1082
1083 if (!cpuset_zone_allowed(zone, gfp_mask))
1084 continue;
1085 /*
1086 * no need to ask again on the same node. Pool is node rather than
1087 * zone aware
1088 */
1089 if (zone_to_nid(zone) == node)
1090 continue;
1091 node = zone_to_nid(zone);
1092
1093 page = dequeue_huge_page_node_exact(h, node);
1094 if (page)
1095 return page;
1096 }
1097 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1098 goto retry_cpuset;
1099
1100 return NULL;
1101}
1102
1103static struct page *dequeue_huge_page_vma(struct hstate *h,
1104 struct vm_area_struct *vma,
1105 unsigned long address, int avoid_reserve,
1106 long chg)
1107{
1108 struct page *page;
1109 struct mempolicy *mpol;
1110 gfp_t gfp_mask;
1111 nodemask_t *nodemask;
1112 int nid;
1113
1114 /*
1115 * A child process with MAP_PRIVATE mappings created by their parent
1116 * have no page reserves. This check ensures that reservations are
1117 * not "stolen". The child may still get SIGKILLed
1118 */
1119 if (!vma_has_reserves(vma, chg) &&
1120 h->free_huge_pages - h->resv_huge_pages == 0)
1121 goto err;
1122
1123 /* If reserves cannot be used, ensure enough pages are in the pool */
1124 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1125 goto err;
1126
1127 gfp_mask = htlb_alloc_mask(h);
1128 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1129 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1130 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1131 SetPagePrivate(page);
1132 h->resv_huge_pages--;
1133 }
1134
1135 mpol_cond_put(mpol);
1136 return page;
1137
1138err:
1139 return NULL;
1140}
1141
1142/*
1143 * common helper functions for hstate_next_node_to_{alloc|free}.
1144 * We may have allocated or freed a huge page based on a different
1145 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1146 * be outside of *nodes_allowed. Ensure that we use an allowed
1147 * node for alloc or free.
1148 */
1149static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1150{
1151 nid = next_node_in(nid, *nodes_allowed);
1152 VM_BUG_ON(nid >= MAX_NUMNODES);
1153
1154 return nid;
1155}
1156
1157static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1158{
1159 if (!node_isset(nid, *nodes_allowed))
1160 nid = next_node_allowed(nid, nodes_allowed);
1161 return nid;
1162}
1163
1164/*
1165 * returns the previously saved node ["this node"] from which to
1166 * allocate a persistent huge page for the pool and advance the
1167 * next node from which to allocate, handling wrap at end of node
1168 * mask.
1169 */
1170static int hstate_next_node_to_alloc(struct hstate *h,
1171 nodemask_t *nodes_allowed)
1172{
1173 int nid;
1174
1175 VM_BUG_ON(!nodes_allowed);
1176
1177 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1178 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1179
1180 return nid;
1181}
1182
1183/*
1184 * helper for free_pool_huge_page() - return the previously saved
1185 * node ["this node"] from which to free a huge page. Advance the
1186 * next node id whether or not we find a free huge page to free so
1187 * that the next attempt to free addresses the next node.
1188 */
1189static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1190{
1191 int nid;
1192
1193 VM_BUG_ON(!nodes_allowed);
1194
1195 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1196 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1197
1198 return nid;
1199}
1200
1201#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1202 for (nr_nodes = nodes_weight(*mask); \
1203 nr_nodes > 0 && \
1204 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1205 nr_nodes--)
1206
1207#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1208 for (nr_nodes = nodes_weight(*mask); \
1209 nr_nodes > 0 && \
1210 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1211 nr_nodes--)
1212
1213#ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1214static void destroy_compound_gigantic_page(struct page *page,
1215 unsigned int order)
1216{
1217 int i;
1218 int nr_pages = 1 << order;
1219 struct page *p = page + 1;
1220
1221 atomic_set(compound_mapcount_ptr(page), 0);
1222 if (hpage_pincount_available(page))
1223 atomic_set(compound_pincount_ptr(page), 0);
1224
1225 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1226 clear_compound_head(p);
1227 set_page_refcounted(p);
1228 }
1229
1230 set_compound_order(page, 0);
1231 __ClearPageHead(page);
1232}
1233
1234static void free_gigantic_page(struct page *page, unsigned int order)
1235{
1236 /*
1237 * If the page isn't allocated using the cma allocator,
1238 * cma_release() returns false.
1239 */
1240#ifdef CONFIG_CMA
1241 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1242 return;
1243#endif
1244
1245 free_contig_range(page_to_pfn(page), 1 << order);
1246}
1247
1248#ifdef CONFIG_CONTIG_ALLOC
1249static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1250 int nid, nodemask_t *nodemask)
1251{
1252 unsigned long nr_pages = 1UL << huge_page_order(h);
1253 if (nid == NUMA_NO_NODE)
1254 nid = numa_mem_id();
1255
1256#ifdef CONFIG_CMA
1257 {
1258 struct page *page;
1259 int node;
1260
1261 if (hugetlb_cma[nid]) {
1262 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1263 huge_page_order(h), true);
1264 if (page)
1265 return page;
1266 }
1267
1268 if (!(gfp_mask & __GFP_THISNODE)) {
1269 for_each_node_mask(node, *nodemask) {
1270 if (node == nid || !hugetlb_cma[node])
1271 continue;
1272
1273 page = cma_alloc(hugetlb_cma[node], nr_pages,
1274 huge_page_order(h), true);
1275 if (page)
1276 return page;
1277 }
1278 }
1279 }
1280#endif
1281
1282 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1283}
1284
1285static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1286static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1287#else /* !CONFIG_CONTIG_ALLOC */
1288static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1289 int nid, nodemask_t *nodemask)
1290{
1291 return NULL;
1292}
1293#endif /* CONFIG_CONTIG_ALLOC */
1294
1295#else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1296static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1297 int nid, nodemask_t *nodemask)
1298{
1299 return NULL;
1300}
1301static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1302static inline void destroy_compound_gigantic_page(struct page *page,
1303 unsigned int order) { }
1304#endif
1305
1306static void update_and_free_page(struct hstate *h, struct page *page)
1307{
1308 int i;
1309
1310 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1311 return;
1312
1313 h->nr_huge_pages--;
1314 h->nr_huge_pages_node[page_to_nid(page)]--;
1315 for (i = 0; i < pages_per_huge_page(h); i++) {
1316 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1317 1 << PG_referenced | 1 << PG_dirty |
1318 1 << PG_active | 1 << PG_private |
1319 1 << PG_writeback);
1320 }
1321 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1322 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1323 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1324 set_page_refcounted(page);
1325 if (hstate_is_gigantic(h)) {
1326 /*
1327 * Temporarily drop the hugetlb_lock, because
1328 * we might block in free_gigantic_page().
1329 */
1330 spin_unlock(&hugetlb_lock);
1331 destroy_compound_gigantic_page(page, huge_page_order(h));
1332 free_gigantic_page(page, huge_page_order(h));
1333 spin_lock(&hugetlb_lock);
1334 } else {
1335 __free_pages(page, huge_page_order(h));
1336 }
1337}
1338
1339struct hstate *size_to_hstate(unsigned long size)
1340{
1341 struct hstate *h;
1342
1343 for_each_hstate(h) {
1344 if (huge_page_size(h) == size)
1345 return h;
1346 }
1347 return NULL;
1348}
1349
1350/*
1351 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1352 * to hstate->hugepage_activelist.)
1353 *
1354 * This function can be called for tail pages, but never returns true for them.
1355 */
1356bool page_huge_active(struct page *page)
1357{
1358 VM_BUG_ON_PAGE(!PageHuge(page), page);
1359 return PageHead(page) && PagePrivate(&page[1]);
1360}
1361
1362/* never called for tail page */
1363static void set_page_huge_active(struct page *page)
1364{
1365 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1366 SetPagePrivate(&page[1]);
1367}
1368
1369static void clear_page_huge_active(struct page *page)
1370{
1371 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1372 ClearPagePrivate(&page[1]);
1373}
1374
1375/*
1376 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1377 * code
1378 */
1379static inline bool PageHugeTemporary(struct page *page)
1380{
1381 if (!PageHuge(page))
1382 return false;
1383
1384 return (unsigned long)page[2].mapping == -1U;
1385}
1386
1387static inline void SetPageHugeTemporary(struct page *page)
1388{
1389 page[2].mapping = (void *)-1U;
1390}
1391
1392static inline void ClearPageHugeTemporary(struct page *page)
1393{
1394 page[2].mapping = NULL;
1395}
1396
1397static void __free_huge_page(struct page *page)
1398{
1399 /*
1400 * Can't pass hstate in here because it is called from the
1401 * compound page destructor.
1402 */
1403 struct hstate *h = page_hstate(page);
1404 int nid = page_to_nid(page);
1405 struct hugepage_subpool *spool =
1406 (struct hugepage_subpool *)page_private(page);
1407 bool restore_reserve;
1408
1409 VM_BUG_ON_PAGE(page_count(page), page);
1410 VM_BUG_ON_PAGE(page_mapcount(page), page);
1411
1412 set_page_private(page, 0);
1413 page->mapping = NULL;
1414 restore_reserve = PagePrivate(page);
1415 ClearPagePrivate(page);
1416
1417 /*
1418 * If PagePrivate() was set on page, page allocation consumed a
1419 * reservation. If the page was associated with a subpool, there
1420 * would have been a page reserved in the subpool before allocation
1421 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1422 * reservtion, do not call hugepage_subpool_put_pages() as this will
1423 * remove the reserved page from the subpool.
1424 */
1425 if (!restore_reserve) {
1426 /*
1427 * A return code of zero implies that the subpool will be
1428 * under its minimum size if the reservation is not restored
1429 * after page is free. Therefore, force restore_reserve
1430 * operation.
1431 */
1432 if (hugepage_subpool_put_pages(spool, 1) == 0)
1433 restore_reserve = true;
1434 }
1435
1436 spin_lock(&hugetlb_lock);
1437 clear_page_huge_active(page);
1438 hugetlb_cgroup_uncharge_page(hstate_index(h),
1439 pages_per_huge_page(h), page);
1440 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1441 pages_per_huge_page(h), page);
1442 if (restore_reserve)
1443 h->resv_huge_pages++;
1444
1445 if (PageHugeTemporary(page)) {
1446 list_del(&page->lru);
1447 ClearPageHugeTemporary(page);
1448 update_and_free_page(h, page);
1449 } else if (h->surplus_huge_pages_node[nid]) {
1450 /* remove the page from active list */
1451 list_del(&page->lru);
1452 update_and_free_page(h, page);
1453 h->surplus_huge_pages--;
1454 h->surplus_huge_pages_node[nid]--;
1455 } else {
1456 arch_clear_hugepage_flags(page);
1457 enqueue_huge_page(h, page);
1458 }
1459 spin_unlock(&hugetlb_lock);
1460}
1461
1462/*
1463 * As free_huge_page() can be called from a non-task context, we have
1464 * to defer the actual freeing in a workqueue to prevent potential
1465 * hugetlb_lock deadlock.
1466 *
1467 * free_hpage_workfn() locklessly retrieves the linked list of pages to
1468 * be freed and frees them one-by-one. As the page->mapping pointer is
1469 * going to be cleared in __free_huge_page() anyway, it is reused as the
1470 * llist_node structure of a lockless linked list of huge pages to be freed.
1471 */
1472static LLIST_HEAD(hpage_freelist);
1473
1474static void free_hpage_workfn(struct work_struct *work)
1475{
1476 struct llist_node *node;
1477 struct page *page;
1478
1479 node = llist_del_all(&hpage_freelist);
1480
1481 while (node) {
1482 page = container_of((struct address_space **)node,
1483 struct page, mapping);
1484 node = node->next;
1485 __free_huge_page(page);
1486 }
1487}
1488static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1489
1490void free_huge_page(struct page *page)
1491{
1492 /*
1493 * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1494 */
1495 if (!in_task()) {
1496 /*
1497 * Only call schedule_work() if hpage_freelist is previously
1498 * empty. Otherwise, schedule_work() had been called but the
1499 * workfn hasn't retrieved the list yet.
1500 */
1501 if (llist_add((struct llist_node *)&page->mapping,
1502 &hpage_freelist))
1503 schedule_work(&free_hpage_work);
1504 return;
1505 }
1506
1507 __free_huge_page(page);
1508}
1509
1510static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1511{
1512 INIT_LIST_HEAD(&page->lru);
1513 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1514 spin_lock(&hugetlb_lock);
1515 set_hugetlb_cgroup(page, NULL);
1516 set_hugetlb_cgroup_rsvd(page, NULL);
1517 h->nr_huge_pages++;
1518 h->nr_huge_pages_node[nid]++;
1519 spin_unlock(&hugetlb_lock);
1520}
1521
1522static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1523{
1524 int i;
1525 int nr_pages = 1 << order;
1526 struct page *p = page + 1;
1527
1528 /* we rely on prep_new_huge_page to set the destructor */
1529 set_compound_order(page, order);
1530 __ClearPageReserved(page);
1531 __SetPageHead(page);
1532 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1533 /*
1534 * For gigantic hugepages allocated through bootmem at
1535 * boot, it's safer to be consistent with the not-gigantic
1536 * hugepages and clear the PG_reserved bit from all tail pages
1537 * too. Otherwise drivers using get_user_pages() to access tail
1538 * pages may get the reference counting wrong if they see
1539 * PG_reserved set on a tail page (despite the head page not
1540 * having PG_reserved set). Enforcing this consistency between
1541 * head and tail pages allows drivers to optimize away a check
1542 * on the head page when they need know if put_page() is needed
1543 * after get_user_pages().
1544 */
1545 __ClearPageReserved(p);
1546 set_page_count(p, 0);
1547 set_compound_head(p, page);
1548 }
1549 atomic_set(compound_mapcount_ptr(page), -1);
1550
1551 if (hpage_pincount_available(page))
1552 atomic_set(compound_pincount_ptr(page), 0);
1553}
1554
1555/*
1556 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1557 * transparent huge pages. See the PageTransHuge() documentation for more
1558 * details.
1559 */
1560int PageHuge(struct page *page)
1561{
1562 if (!PageCompound(page))
1563 return 0;
1564
1565 page = compound_head(page);
1566 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1567}
1568EXPORT_SYMBOL_GPL(PageHuge);
1569
1570/*
1571 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1572 * normal or transparent huge pages.
1573 */
1574int PageHeadHuge(struct page *page_head)
1575{
1576 if (!PageHead(page_head))
1577 return 0;
1578
1579 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1580}
1581
1582/*
1583 * Find address_space associated with hugetlbfs page.
1584 * Upon entry page is locked and page 'was' mapped although mapped state
1585 * could change. If necessary, use anon_vma to find vma and associated
1586 * address space. The returned mapping may be stale, but it can not be
1587 * invalid as page lock (which is held) is required to destroy mapping.
1588 */
1589static struct address_space *_get_hugetlb_page_mapping(struct page *hpage)
1590{
1591 struct anon_vma *anon_vma;
1592 pgoff_t pgoff_start, pgoff_end;
1593 struct anon_vma_chain *avc;
1594 struct address_space *mapping = page_mapping(hpage);
1595
1596 /* Simple file based mapping */
1597 if (mapping)
1598 return mapping;
1599
1600 /*
1601 * Even anonymous hugetlbfs mappings are associated with an
1602 * underlying hugetlbfs file (see hugetlb_file_setup in mmap
1603 * code). Find a vma associated with the anonymous vma, and
1604 * use the file pointer to get address_space.
1605 */
1606 anon_vma = page_lock_anon_vma_read(hpage);
1607 if (!anon_vma)
1608 return mapping; /* NULL */
1609
1610 /* Use first found vma */
1611 pgoff_start = page_to_pgoff(hpage);
1612 pgoff_end = pgoff_start + pages_per_huge_page(page_hstate(hpage)) - 1;
1613 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1614 pgoff_start, pgoff_end) {
1615 struct vm_area_struct *vma = avc->vma;
1616
1617 mapping = vma->vm_file->f_mapping;
1618 break;
1619 }
1620
1621 anon_vma_unlock_read(anon_vma);
1622 return mapping;
1623}
1624
1625/*
1626 * Find and lock address space (mapping) in write mode.
1627 *
1628 * Upon entry, the page is locked which allows us to find the mapping
1629 * even in the case of an anon page. However, locking order dictates
1630 * the i_mmap_rwsem be acquired BEFORE the page lock. This is hugetlbfs
1631 * specific. So, we first try to lock the sema while still holding the
1632 * page lock. If this works, great! If not, then we need to drop the
1633 * page lock and then acquire i_mmap_rwsem and reacquire page lock. Of
1634 * course, need to revalidate state along the way.
1635 */
1636struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1637{
1638 struct address_space *mapping, *mapping2;
1639
1640 mapping = _get_hugetlb_page_mapping(hpage);
1641retry:
1642 if (!mapping)
1643 return mapping;
1644
1645 /*
1646 * If no contention, take lock and return
1647 */
1648 if (i_mmap_trylock_write(mapping))
1649 return mapping;
1650
1651 /*
1652 * Must drop page lock and wait on mapping sema.
1653 * Note: Once page lock is dropped, mapping could become invalid.
1654 * As a hack, increase map count until we lock page again.
1655 */
1656 atomic_inc(&hpage->_mapcount);
1657 unlock_page(hpage);
1658 i_mmap_lock_write(mapping);
1659 lock_page(hpage);
1660 atomic_add_negative(-1, &hpage->_mapcount);
1661
1662 /* verify page is still mapped */
1663 if (!page_mapped(hpage)) {
1664 i_mmap_unlock_write(mapping);
1665 return NULL;
1666 }
1667
1668 /*
1669 * Get address space again and verify it is the same one
1670 * we locked. If not, drop lock and retry.
1671 */
1672 mapping2 = _get_hugetlb_page_mapping(hpage);
1673 if (mapping2 != mapping) {
1674 i_mmap_unlock_write(mapping);
1675 mapping = mapping2;
1676 goto retry;
1677 }
1678
1679 return mapping;
1680}
1681
1682pgoff_t __basepage_index(struct page *page)
1683{
1684 struct page *page_head = compound_head(page);
1685 pgoff_t index = page_index(page_head);
1686 unsigned long compound_idx;
1687
1688 if (!PageHuge(page_head))
1689 return page_index(page);
1690
1691 if (compound_order(page_head) >= MAX_ORDER)
1692 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1693 else
1694 compound_idx = page - page_head;
1695
1696 return (index << compound_order(page_head)) + compound_idx;
1697}
1698
1699static struct page *alloc_buddy_huge_page(struct hstate *h,
1700 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1701 nodemask_t *node_alloc_noretry)
1702{
1703 int order = huge_page_order(h);
1704 struct page *page;
1705 bool alloc_try_hard = true;
1706
1707 /*
1708 * By default we always try hard to allocate the page with
1709 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1710 * a loop (to adjust global huge page counts) and previous allocation
1711 * failed, do not continue to try hard on the same node. Use the
1712 * node_alloc_noretry bitmap to manage this state information.
1713 */
1714 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1715 alloc_try_hard = false;
1716 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1717 if (alloc_try_hard)
1718 gfp_mask |= __GFP_RETRY_MAYFAIL;
1719 if (nid == NUMA_NO_NODE)
1720 nid = numa_mem_id();
1721 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1722 if (page)
1723 __count_vm_event(HTLB_BUDDY_PGALLOC);
1724 else
1725 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1726
1727 /*
1728 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1729 * indicates an overall state change. Clear bit so that we resume
1730 * normal 'try hard' allocations.
1731 */
1732 if (node_alloc_noretry && page && !alloc_try_hard)
1733 node_clear(nid, *node_alloc_noretry);
1734
1735 /*
1736 * If we tried hard to get a page but failed, set bit so that
1737 * subsequent attempts will not try as hard until there is an
1738 * overall state change.
1739 */
1740 if (node_alloc_noretry && !page && alloc_try_hard)
1741 node_set(nid, *node_alloc_noretry);
1742
1743 return page;
1744}
1745
1746/*
1747 * Common helper to allocate a fresh hugetlb page. All specific allocators
1748 * should use this function to get new hugetlb pages
1749 */
1750static struct page *alloc_fresh_huge_page(struct hstate *h,
1751 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1752 nodemask_t *node_alloc_noretry)
1753{
1754 struct page *page;
1755
1756 if (hstate_is_gigantic(h))
1757 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1758 else
1759 page = alloc_buddy_huge_page(h, gfp_mask,
1760 nid, nmask, node_alloc_noretry);
1761 if (!page)
1762 return NULL;
1763
1764 if (hstate_is_gigantic(h))
1765 prep_compound_gigantic_page(page, huge_page_order(h));
1766 prep_new_huge_page(h, page, page_to_nid(page));
1767
1768 return page;
1769}
1770
1771/*
1772 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1773 * manner.
1774 */
1775static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1776 nodemask_t *node_alloc_noretry)
1777{
1778 struct page *page;
1779 int nr_nodes, node;
1780 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1781
1782 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1783 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1784 node_alloc_noretry);
1785 if (page)
1786 break;
1787 }
1788
1789 if (!page)
1790 return 0;
1791
1792 put_page(page); /* free it into the hugepage allocator */
1793
1794 return 1;
1795}
1796
1797/*
1798 * Free huge page from pool from next node to free.
1799 * Attempt to keep persistent huge pages more or less
1800 * balanced over allowed nodes.
1801 * Called with hugetlb_lock locked.
1802 */
1803static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1804 bool acct_surplus)
1805{
1806 int nr_nodes, node;
1807 int ret = 0;
1808
1809 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1810 /*
1811 * If we're returning unused surplus pages, only examine
1812 * nodes with surplus pages.
1813 */
1814 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1815 !list_empty(&h->hugepage_freelists[node])) {
1816 struct page *page =
1817 list_entry(h->hugepage_freelists[node].next,
1818 struct page, lru);
1819 list_del(&page->lru);
1820 h->free_huge_pages--;
1821 h->free_huge_pages_node[node]--;
1822 if (acct_surplus) {
1823 h->surplus_huge_pages--;
1824 h->surplus_huge_pages_node[node]--;
1825 }
1826 update_and_free_page(h, page);
1827 ret = 1;
1828 break;
1829 }
1830 }
1831
1832 return ret;
1833}
1834
1835/*
1836 * Dissolve a given free hugepage into free buddy pages. This function does
1837 * nothing for in-use hugepages and non-hugepages.
1838 * This function returns values like below:
1839 *
1840 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1841 * (allocated or reserved.)
1842 * 0: successfully dissolved free hugepages or the page is not a
1843 * hugepage (considered as already dissolved)
1844 */
1845int dissolve_free_huge_page(struct page *page)
1846{
1847 int rc = -EBUSY;
1848
1849 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1850 if (!PageHuge(page))
1851 return 0;
1852
1853 spin_lock(&hugetlb_lock);
1854 if (!PageHuge(page)) {
1855 rc = 0;
1856 goto out;
1857 }
1858
1859 if (!page_count(page)) {
1860 struct page *head = compound_head(page);
1861 struct hstate *h = page_hstate(head);
1862 int nid = page_to_nid(head);
1863 if (h->free_huge_pages - h->resv_huge_pages == 0)
1864 goto out;
1865 /*
1866 * Move PageHWPoison flag from head page to the raw error page,
1867 * which makes any subpages rather than the error page reusable.
1868 */
1869 if (PageHWPoison(head) && page != head) {
1870 SetPageHWPoison(page);
1871 ClearPageHWPoison(head);
1872 }
1873 list_del(&head->lru);
1874 h->free_huge_pages--;
1875 h->free_huge_pages_node[nid]--;
1876 h->max_huge_pages--;
1877 update_and_free_page(h, head);
1878 rc = 0;
1879 }
1880out:
1881 spin_unlock(&hugetlb_lock);
1882 return rc;
1883}
1884
1885/*
1886 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1887 * make specified memory blocks removable from the system.
1888 * Note that this will dissolve a free gigantic hugepage completely, if any
1889 * part of it lies within the given range.
1890 * Also note that if dissolve_free_huge_page() returns with an error, all
1891 * free hugepages that were dissolved before that error are lost.
1892 */
1893int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1894{
1895 unsigned long pfn;
1896 struct page *page;
1897 int rc = 0;
1898
1899 if (!hugepages_supported())
1900 return rc;
1901
1902 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1903 page = pfn_to_page(pfn);
1904 rc = dissolve_free_huge_page(page);
1905 if (rc)
1906 break;
1907 }
1908
1909 return rc;
1910}
1911
1912/*
1913 * Allocates a fresh surplus page from the page allocator.
1914 */
1915static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1916 int nid, nodemask_t *nmask)
1917{
1918 struct page *page = NULL;
1919
1920 if (hstate_is_gigantic(h))
1921 return NULL;
1922
1923 spin_lock(&hugetlb_lock);
1924 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1925 goto out_unlock;
1926 spin_unlock(&hugetlb_lock);
1927
1928 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1929 if (!page)
1930 return NULL;
1931
1932 spin_lock(&hugetlb_lock);
1933 /*
1934 * We could have raced with the pool size change.
1935 * Double check that and simply deallocate the new page
1936 * if we would end up overcommiting the surpluses. Abuse
1937 * temporary page to workaround the nasty free_huge_page
1938 * codeflow
1939 */
1940 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1941 SetPageHugeTemporary(page);
1942 spin_unlock(&hugetlb_lock);
1943 put_page(page);
1944 return NULL;
1945 } else {
1946 h->surplus_huge_pages++;
1947 h->surplus_huge_pages_node[page_to_nid(page)]++;
1948 }
1949
1950out_unlock:
1951 spin_unlock(&hugetlb_lock);
1952
1953 return page;
1954}
1955
1956static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1957 int nid, nodemask_t *nmask)
1958{
1959 struct page *page;
1960
1961 if (hstate_is_gigantic(h))
1962 return NULL;
1963
1964 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1965 if (!page)
1966 return NULL;
1967
1968 /*
1969 * We do not account these pages as surplus because they are only
1970 * temporary and will be released properly on the last reference
1971 */
1972 SetPageHugeTemporary(page);
1973
1974 return page;
1975}
1976
1977/*
1978 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1979 */
1980static
1981struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1982 struct vm_area_struct *vma, unsigned long addr)
1983{
1984 struct page *page;
1985 struct mempolicy *mpol;
1986 gfp_t gfp_mask = htlb_alloc_mask(h);
1987 int nid;
1988 nodemask_t *nodemask;
1989
1990 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1991 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1992 mpol_cond_put(mpol);
1993
1994 return page;
1995}
1996
1997/* page migration callback function */
1998struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1999 nodemask_t *nmask, gfp_t gfp_mask)
2000{
2001 spin_lock(&hugetlb_lock);
2002 if (h->free_huge_pages - h->resv_huge_pages > 0) {
2003 struct page *page;
2004
2005 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
2006 if (page) {
2007 spin_unlock(&hugetlb_lock);
2008 return page;
2009 }
2010 }
2011 spin_unlock(&hugetlb_lock);
2012
2013 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2014}
2015
2016/* mempolicy aware migration callback */
2017struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2018 unsigned long address)
2019{
2020 struct mempolicy *mpol;
2021 nodemask_t *nodemask;
2022 struct page *page;
2023 gfp_t gfp_mask;
2024 int node;
2025
2026 gfp_mask = htlb_alloc_mask(h);
2027 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2028 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
2029 mpol_cond_put(mpol);
2030
2031 return page;
2032}
2033
2034/*
2035 * Increase the hugetlb pool such that it can accommodate a reservation
2036 * of size 'delta'.
2037 */
2038static int gather_surplus_pages(struct hstate *h, int delta)
2039 __must_hold(&hugetlb_lock)
2040{
2041 struct list_head surplus_list;
2042 struct page *page, *tmp;
2043 int ret, i;
2044 int needed, allocated;
2045 bool alloc_ok = true;
2046
2047 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2048 if (needed <= 0) {
2049 h->resv_huge_pages += delta;
2050 return 0;
2051 }
2052
2053 allocated = 0;
2054 INIT_LIST_HEAD(&surplus_list);
2055
2056 ret = -ENOMEM;
2057retry:
2058 spin_unlock(&hugetlb_lock);
2059 for (i = 0; i < needed; i++) {
2060 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2061 NUMA_NO_NODE, NULL);
2062 if (!page) {
2063 alloc_ok = false;
2064 break;
2065 }
2066 list_add(&page->lru, &surplus_list);
2067 cond_resched();
2068 }
2069 allocated += i;
2070
2071 /*
2072 * After retaking hugetlb_lock, we need to recalculate 'needed'
2073 * because either resv_huge_pages or free_huge_pages may have changed.
2074 */
2075 spin_lock(&hugetlb_lock);
2076 needed = (h->resv_huge_pages + delta) -
2077 (h->free_huge_pages + allocated);
2078 if (needed > 0) {
2079 if (alloc_ok)
2080 goto retry;
2081 /*
2082 * We were not able to allocate enough pages to
2083 * satisfy the entire reservation so we free what
2084 * we've allocated so far.
2085 */
2086 goto free;
2087 }
2088 /*
2089 * The surplus_list now contains _at_least_ the number of extra pages
2090 * needed to accommodate the reservation. Add the appropriate number
2091 * of pages to the hugetlb pool and free the extras back to the buddy
2092 * allocator. Commit the entire reservation here to prevent another
2093 * process from stealing the pages as they are added to the pool but
2094 * before they are reserved.
2095 */
2096 needed += allocated;
2097 h->resv_huge_pages += delta;
2098 ret = 0;
2099
2100 /* Free the needed pages to the hugetlb pool */
2101 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2102 if ((--needed) < 0)
2103 break;
2104 /*
2105 * This page is now managed by the hugetlb allocator and has
2106 * no users -- drop the buddy allocator's reference.
2107 */
2108 put_page_testzero(page);
2109 VM_BUG_ON_PAGE(page_count(page), page);
2110 enqueue_huge_page(h, page);
2111 }
2112free:
2113 spin_unlock(&hugetlb_lock);
2114
2115 /* Free unnecessary surplus pages to the buddy allocator */
2116 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2117 put_page(page);
2118 spin_lock(&hugetlb_lock);
2119
2120 return ret;
2121}
2122
2123/*
2124 * This routine has two main purposes:
2125 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2126 * in unused_resv_pages. This corresponds to the prior adjustments made
2127 * to the associated reservation map.
2128 * 2) Free any unused surplus pages that may have been allocated to satisfy
2129 * the reservation. As many as unused_resv_pages may be freed.
2130 *
2131 * Called with hugetlb_lock held. However, the lock could be dropped (and
2132 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
2133 * we must make sure nobody else can claim pages we are in the process of
2134 * freeing. Do this by ensuring resv_huge_page always is greater than the
2135 * number of huge pages we plan to free when dropping the lock.
2136 */
2137static void return_unused_surplus_pages(struct hstate *h,
2138 unsigned long unused_resv_pages)
2139{
2140 unsigned long nr_pages;
2141
2142 /* Cannot return gigantic pages currently */
2143 if (hstate_is_gigantic(h))
2144 goto out;
2145
2146 /*
2147 * Part (or even all) of the reservation could have been backed
2148 * by pre-allocated pages. Only free surplus pages.
2149 */
2150 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2151
2152 /*
2153 * We want to release as many surplus pages as possible, spread
2154 * evenly across all nodes with memory. Iterate across these nodes
2155 * until we can no longer free unreserved surplus pages. This occurs
2156 * when the nodes with surplus pages have no free pages.
2157 * free_pool_huge_page() will balance the freed pages across the
2158 * on-line nodes with memory and will handle the hstate accounting.
2159 *
2160 * Note that we decrement resv_huge_pages as we free the pages. If
2161 * we drop the lock, resv_huge_pages will still be sufficiently large
2162 * to cover subsequent pages we may free.
2163 */
2164 while (nr_pages--) {
2165 h->resv_huge_pages--;
2166 unused_resv_pages--;
2167 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
2168 goto out;
2169 cond_resched_lock(&hugetlb_lock);
2170 }
2171
2172out:
2173 /* Fully uncommit the reservation */
2174 h->resv_huge_pages -= unused_resv_pages;
2175}
2176
2177
2178/*
2179 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2180 * are used by the huge page allocation routines to manage reservations.
2181 *
2182 * vma_needs_reservation is called to determine if the huge page at addr
2183 * within the vma has an associated reservation. If a reservation is
2184 * needed, the value 1 is returned. The caller is then responsible for
2185 * managing the global reservation and subpool usage counts. After
2186 * the huge page has been allocated, vma_commit_reservation is called
2187 * to add the page to the reservation map. If the page allocation fails,
2188 * the reservation must be ended instead of committed. vma_end_reservation
2189 * is called in such cases.
2190 *
2191 * In the normal case, vma_commit_reservation returns the same value
2192 * as the preceding vma_needs_reservation call. The only time this
2193 * is not the case is if a reserve map was changed between calls. It
2194 * is the responsibility of the caller to notice the difference and
2195 * take appropriate action.
2196 *
2197 * vma_add_reservation is used in error paths where a reservation must
2198 * be restored when a newly allocated huge page must be freed. It is
2199 * to be called after calling vma_needs_reservation to determine if a
2200 * reservation exists.
2201 */
2202enum vma_resv_mode {
2203 VMA_NEEDS_RESV,
2204 VMA_COMMIT_RESV,
2205 VMA_END_RESV,
2206 VMA_ADD_RESV,
2207};
2208static long __vma_reservation_common(struct hstate *h,
2209 struct vm_area_struct *vma, unsigned long addr,
2210 enum vma_resv_mode mode)
2211{
2212 struct resv_map *resv;
2213 pgoff_t idx;
2214 long ret;
2215 long dummy_out_regions_needed;
2216
2217 resv = vma_resv_map(vma);
2218 if (!resv)
2219 return 1;
2220
2221 idx = vma_hugecache_offset(h, vma, addr);
2222 switch (mode) {
2223 case VMA_NEEDS_RESV:
2224 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2225 /* We assume that vma_reservation_* routines always operate on
2226 * 1 page, and that adding to resv map a 1 page entry can only
2227 * ever require 1 region.
2228 */
2229 VM_BUG_ON(dummy_out_regions_needed != 1);
2230 break;
2231 case VMA_COMMIT_RESV:
2232 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2233 /* region_add calls of range 1 should never fail. */
2234 VM_BUG_ON(ret < 0);
2235 break;
2236 case VMA_END_RESV:
2237 region_abort(resv, idx, idx + 1, 1);
2238 ret = 0;
2239 break;
2240 case VMA_ADD_RESV:
2241 if (vma->vm_flags & VM_MAYSHARE) {
2242 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2243 /* region_add calls of range 1 should never fail. */
2244 VM_BUG_ON(ret < 0);
2245 } else {
2246 region_abort(resv, idx, idx + 1, 1);
2247 ret = region_del(resv, idx, idx + 1);
2248 }
2249 break;
2250 default:
2251 BUG();
2252 }
2253
2254 if (vma->vm_flags & VM_MAYSHARE)
2255 return ret;
2256 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
2257 /*
2258 * In most cases, reserves always exist for private mappings.
2259 * However, a file associated with mapping could have been
2260 * hole punched or truncated after reserves were consumed.
2261 * As subsequent fault on such a range will not use reserves.
2262 * Subtle - The reserve map for private mappings has the
2263 * opposite meaning than that of shared mappings. If NO
2264 * entry is in the reserve map, it means a reservation exists.
2265 * If an entry exists in the reserve map, it means the
2266 * reservation has already been consumed. As a result, the
2267 * return value of this routine is the opposite of the
2268 * value returned from reserve map manipulation routines above.
2269 */
2270 if (ret)
2271 return 0;
2272 else
2273 return 1;
2274 }
2275 else
2276 return ret < 0 ? ret : 0;
2277}
2278
2279static long vma_needs_reservation(struct hstate *h,
2280 struct vm_area_struct *vma, unsigned long addr)
2281{
2282 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2283}
2284
2285static long vma_commit_reservation(struct hstate *h,
2286 struct vm_area_struct *vma, unsigned long addr)
2287{
2288 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2289}
2290
2291static void vma_end_reservation(struct hstate *h,
2292 struct vm_area_struct *vma, unsigned long addr)
2293{
2294 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2295}
2296
2297static long vma_add_reservation(struct hstate *h,
2298 struct vm_area_struct *vma, unsigned long addr)
2299{
2300 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2301}
2302
2303/*
2304 * This routine is called to restore a reservation on error paths. In the
2305 * specific error paths, a huge page was allocated (via alloc_huge_page)
2306 * and is about to be freed. If a reservation for the page existed,
2307 * alloc_huge_page would have consumed the reservation and set PagePrivate
2308 * in the newly allocated page. When the page is freed via free_huge_page,
2309 * the global reservation count will be incremented if PagePrivate is set.
2310 * However, free_huge_page can not adjust the reserve map. Adjust the
2311 * reserve map here to be consistent with global reserve count adjustments
2312 * to be made by free_huge_page.
2313 */
2314static void restore_reserve_on_error(struct hstate *h,
2315 struct vm_area_struct *vma, unsigned long address,
2316 struct page *page)
2317{
2318 if (unlikely(PagePrivate(page))) {
2319 long rc = vma_needs_reservation(h, vma, address);
2320
2321 if (unlikely(rc < 0)) {
2322 /*
2323 * Rare out of memory condition in reserve map
2324 * manipulation. Clear PagePrivate so that
2325 * global reserve count will not be incremented
2326 * by free_huge_page. This will make it appear
2327 * as though the reservation for this page was
2328 * consumed. This may prevent the task from
2329 * faulting in the page at a later time. This
2330 * is better than inconsistent global huge page
2331 * accounting of reserve counts.
2332 */
2333 ClearPagePrivate(page);
2334 } else if (rc) {
2335 rc = vma_add_reservation(h, vma, address);
2336 if (unlikely(rc < 0))
2337 /*
2338 * See above comment about rare out of
2339 * memory condition.
2340 */
2341 ClearPagePrivate(page);
2342 } else
2343 vma_end_reservation(h, vma, address);
2344 }
2345}
2346
2347struct page *alloc_huge_page(struct vm_area_struct *vma,
2348 unsigned long addr, int avoid_reserve)
2349{
2350 struct hugepage_subpool *spool = subpool_vma(vma);
2351 struct hstate *h = hstate_vma(vma);
2352 struct page *page;
2353 long map_chg, map_commit;
2354 long gbl_chg;
2355 int ret, idx;
2356 struct hugetlb_cgroup *h_cg;
2357 bool deferred_reserve;
2358
2359 idx = hstate_index(h);
2360 /*
2361 * Examine the region/reserve map to determine if the process
2362 * has a reservation for the page to be allocated. A return
2363 * code of zero indicates a reservation exists (no change).
2364 */
2365 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2366 if (map_chg < 0)
2367 return ERR_PTR(-ENOMEM);
2368
2369 /*
2370 * Processes that did not create the mapping will have no
2371 * reserves as indicated by the region/reserve map. Check
2372 * that the allocation will not exceed the subpool limit.
2373 * Allocations for MAP_NORESERVE mappings also need to be
2374 * checked against any subpool limit.
2375 */
2376 if (map_chg || avoid_reserve) {
2377 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2378 if (gbl_chg < 0) {
2379 vma_end_reservation(h, vma, addr);
2380 return ERR_PTR(-ENOSPC);
2381 }
2382
2383 /*
2384 * Even though there was no reservation in the region/reserve
2385 * map, there could be reservations associated with the
2386 * subpool that can be used. This would be indicated if the
2387 * return value of hugepage_subpool_get_pages() is zero.
2388 * However, if avoid_reserve is specified we still avoid even
2389 * the subpool reservations.
2390 */
2391 if (avoid_reserve)
2392 gbl_chg = 1;
2393 }
2394
2395 /* If this allocation is not consuming a reservation, charge it now.
2396 */
2397 deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma);
2398 if (deferred_reserve) {
2399 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2400 idx, pages_per_huge_page(h), &h_cg);
2401 if (ret)
2402 goto out_subpool_put;
2403 }
2404
2405 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2406 if (ret)
2407 goto out_uncharge_cgroup_reservation;
2408
2409 spin_lock(&hugetlb_lock);
2410 /*
2411 * glb_chg is passed to indicate whether or not a page must be taken
2412 * from the global free pool (global change). gbl_chg == 0 indicates
2413 * a reservation exists for the allocation.
2414 */
2415 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2416 if (!page) {
2417 spin_unlock(&hugetlb_lock);
2418 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2419 if (!page)
2420 goto out_uncharge_cgroup;
2421 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2422 SetPagePrivate(page);
2423 h->resv_huge_pages--;
2424 }
2425 spin_lock(&hugetlb_lock);
2426 list_move(&page->lru, &h->hugepage_activelist);
2427 /* Fall through */
2428 }
2429 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2430 /* If allocation is not consuming a reservation, also store the
2431 * hugetlb_cgroup pointer on the page.
2432 */
2433 if (deferred_reserve) {
2434 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2435 h_cg, page);
2436 }
2437
2438 spin_unlock(&hugetlb_lock);
2439
2440 set_page_private(page, (unsigned long)spool);
2441
2442 map_commit = vma_commit_reservation(h, vma, addr);
2443 if (unlikely(map_chg > map_commit)) {
2444 /*
2445 * The page was added to the reservation map between
2446 * vma_needs_reservation and vma_commit_reservation.
2447 * This indicates a race with hugetlb_reserve_pages.
2448 * Adjust for the subpool count incremented above AND
2449 * in hugetlb_reserve_pages for the same page. Also,
2450 * the reservation count added in hugetlb_reserve_pages
2451 * no longer applies.
2452 */
2453 long rsv_adjust;
2454
2455 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2456 hugetlb_acct_memory(h, -rsv_adjust);
2457 }
2458 return page;
2459
2460out_uncharge_cgroup:
2461 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2462out_uncharge_cgroup_reservation:
2463 if (deferred_reserve)
2464 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2465 h_cg);
2466out_subpool_put:
2467 if (map_chg || avoid_reserve)
2468 hugepage_subpool_put_pages(spool, 1);
2469 vma_end_reservation(h, vma, addr);
2470 return ERR_PTR(-ENOSPC);
2471}
2472
2473int alloc_bootmem_huge_page(struct hstate *h)
2474 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2475int __alloc_bootmem_huge_page(struct hstate *h)
2476{
2477 struct huge_bootmem_page *m;
2478 int nr_nodes, node;
2479
2480 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2481 void *addr;
2482
2483 addr = memblock_alloc_try_nid_raw(
2484 huge_page_size(h), huge_page_size(h),
2485 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2486 if (addr) {
2487 /*
2488 * Use the beginning of the huge page to store the
2489 * huge_bootmem_page struct (until gather_bootmem
2490 * puts them into the mem_map).
2491 */
2492 m = addr;
2493 goto found;
2494 }
2495 }
2496 return 0;
2497
2498found:
2499 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2500 /* Put them into a private list first because mem_map is not up yet */
2501 INIT_LIST_HEAD(&m->list);
2502 list_add(&m->list, &huge_boot_pages);
2503 m->hstate = h;
2504 return 1;
2505}
2506
2507static void __init prep_compound_huge_page(struct page *page,
2508 unsigned int order)
2509{
2510 if (unlikely(order > (MAX_ORDER - 1)))
2511 prep_compound_gigantic_page(page, order);
2512 else
2513 prep_compound_page(page, order);
2514}
2515
2516/* Put bootmem huge pages into the standard lists after mem_map is up */
2517static void __init gather_bootmem_prealloc(void)
2518{
2519 struct huge_bootmem_page *m;
2520
2521 list_for_each_entry(m, &huge_boot_pages, list) {
2522 struct page *page = virt_to_page(m);
2523 struct hstate *h = m->hstate;
2524
2525 WARN_ON(page_count(page) != 1);
2526 prep_compound_huge_page(page, h->order);
2527 WARN_ON(PageReserved(page));
2528 prep_new_huge_page(h, page, page_to_nid(page));
2529 put_page(page); /* free it into the hugepage allocator */
2530
2531 /*
2532 * If we had gigantic hugepages allocated at boot time, we need
2533 * to restore the 'stolen' pages to totalram_pages in order to
2534 * fix confusing memory reports from free(1) and another
2535 * side-effects, like CommitLimit going negative.
2536 */
2537 if (hstate_is_gigantic(h))
2538 adjust_managed_page_count(page, 1 << h->order);
2539 cond_resched();
2540 }
2541}
2542
2543static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2544{
2545 unsigned long i;
2546 nodemask_t *node_alloc_noretry;
2547
2548 if (!hstate_is_gigantic(h)) {
2549 /*
2550 * Bit mask controlling how hard we retry per-node allocations.
2551 * Ignore errors as lower level routines can deal with
2552 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2553 * time, we are likely in bigger trouble.
2554 */
2555 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2556 GFP_KERNEL);
2557 } else {
2558 /* allocations done at boot time */
2559 node_alloc_noretry = NULL;
2560 }
2561
2562 /* bit mask controlling how hard we retry per-node allocations */
2563 if (node_alloc_noretry)
2564 nodes_clear(*node_alloc_noretry);
2565
2566 for (i = 0; i < h->max_huge_pages; ++i) {
2567 if (hstate_is_gigantic(h)) {
2568 if (hugetlb_cma_size) {
2569 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2570 break;
2571 }
2572 if (!alloc_bootmem_huge_page(h))
2573 break;
2574 } else if (!alloc_pool_huge_page(h,
2575 &node_states[N_MEMORY],
2576 node_alloc_noretry))
2577 break;
2578 cond_resched();
2579 }
2580 if (i < h->max_huge_pages) {
2581 char buf[32];
2582
2583 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2584 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2585 h->max_huge_pages, buf, i);
2586 h->max_huge_pages = i;
2587 }
2588
2589 kfree(node_alloc_noretry);
2590}
2591
2592static void __init hugetlb_init_hstates(void)
2593{
2594 struct hstate *h;
2595
2596 for_each_hstate(h) {
2597 if (minimum_order > huge_page_order(h))
2598 minimum_order = huge_page_order(h);
2599
2600 /* oversize hugepages were init'ed in early boot */
2601 if (!hstate_is_gigantic(h))
2602 hugetlb_hstate_alloc_pages(h);
2603 }
2604 VM_BUG_ON(minimum_order == UINT_MAX);
2605}
2606
2607static void __init report_hugepages(void)
2608{
2609 struct hstate *h;
2610
2611 for_each_hstate(h) {
2612 char buf[32];
2613
2614 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2615 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2616 buf, h->free_huge_pages);
2617 }
2618}
2619
2620#ifdef CONFIG_HIGHMEM
2621static void try_to_free_low(struct hstate *h, unsigned long count,
2622 nodemask_t *nodes_allowed)
2623{
2624 int i;
2625
2626 if (hstate_is_gigantic(h))
2627 return;
2628
2629 for_each_node_mask(i, *nodes_allowed) {
2630 struct page *page, *next;
2631 struct list_head *freel = &h->hugepage_freelists[i];
2632 list_for_each_entry_safe(page, next, freel, lru) {
2633 if (count >= h->nr_huge_pages)
2634 return;
2635 if (PageHighMem(page))
2636 continue;
2637 list_del(&page->lru);
2638 update_and_free_page(h, page);
2639 h->free_huge_pages--;
2640 h->free_huge_pages_node[page_to_nid(page)]--;
2641 }
2642 }
2643}
2644#else
2645static inline void try_to_free_low(struct hstate *h, unsigned long count,
2646 nodemask_t *nodes_allowed)
2647{
2648}
2649#endif
2650
2651/*
2652 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2653 * balanced by operating on them in a round-robin fashion.
2654 * Returns 1 if an adjustment was made.
2655 */
2656static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2657 int delta)
2658{
2659 int nr_nodes, node;
2660
2661 VM_BUG_ON(delta != -1 && delta != 1);
2662
2663 if (delta < 0) {
2664 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2665 if (h->surplus_huge_pages_node[node])
2666 goto found;
2667 }
2668 } else {
2669 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2670 if (h->surplus_huge_pages_node[node] <
2671 h->nr_huge_pages_node[node])
2672 goto found;
2673 }
2674 }
2675 return 0;
2676
2677found:
2678 h->surplus_huge_pages += delta;
2679 h->surplus_huge_pages_node[node] += delta;
2680 return 1;
2681}
2682
2683#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2684static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2685 nodemask_t *nodes_allowed)
2686{
2687 unsigned long min_count, ret;
2688 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2689
2690 /*
2691 * Bit mask controlling how hard we retry per-node allocations.
2692 * If we can not allocate the bit mask, do not attempt to allocate
2693 * the requested huge pages.
2694 */
2695 if (node_alloc_noretry)
2696 nodes_clear(*node_alloc_noretry);
2697 else
2698 return -ENOMEM;
2699
2700 spin_lock(&hugetlb_lock);
2701
2702 /*
2703 * Check for a node specific request.
2704 * Changing node specific huge page count may require a corresponding
2705 * change to the global count. In any case, the passed node mask
2706 * (nodes_allowed) will restrict alloc/free to the specified node.
2707 */
2708 if (nid != NUMA_NO_NODE) {
2709 unsigned long old_count = count;
2710
2711 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2712 /*
2713 * User may have specified a large count value which caused the
2714 * above calculation to overflow. In this case, they wanted
2715 * to allocate as many huge pages as possible. Set count to
2716 * largest possible value to align with their intention.
2717 */
2718 if (count < old_count)
2719 count = ULONG_MAX;
2720 }
2721
2722 /*
2723 * Gigantic pages runtime allocation depend on the capability for large
2724 * page range allocation.
2725 * If the system does not provide this feature, return an error when
2726 * the user tries to allocate gigantic pages but let the user free the
2727 * boottime allocated gigantic pages.
2728 */
2729 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2730 if (count > persistent_huge_pages(h)) {
2731 spin_unlock(&hugetlb_lock);
2732 NODEMASK_FREE(node_alloc_noretry);
2733 return -EINVAL;
2734 }
2735 /* Fall through to decrease pool */
2736 }
2737
2738 /*
2739 * Increase the pool size
2740 * First take pages out of surplus state. Then make up the
2741 * remaining difference by allocating fresh huge pages.
2742 *
2743 * We might race with alloc_surplus_huge_page() here and be unable
2744 * to convert a surplus huge page to a normal huge page. That is
2745 * not critical, though, it just means the overall size of the
2746 * pool might be one hugepage larger than it needs to be, but
2747 * within all the constraints specified by the sysctls.
2748 */
2749 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2750 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2751 break;
2752 }
2753
2754 while (count > persistent_huge_pages(h)) {
2755 /*
2756 * If this allocation races such that we no longer need the
2757 * page, free_huge_page will handle it by freeing the page
2758 * and reducing the surplus.
2759 */
2760 spin_unlock(&hugetlb_lock);
2761
2762 /* yield cpu to avoid soft lockup */
2763 cond_resched();
2764
2765 ret = alloc_pool_huge_page(h, nodes_allowed,
2766 node_alloc_noretry);
2767 spin_lock(&hugetlb_lock);
2768 if (!ret)
2769 goto out;
2770
2771 /* Bail for signals. Probably ctrl-c from user */
2772 if (signal_pending(current))
2773 goto out;
2774 }
2775
2776 /*
2777 * Decrease the pool size
2778 * First return free pages to the buddy allocator (being careful
2779 * to keep enough around to satisfy reservations). Then place
2780 * pages into surplus state as needed so the pool will shrink
2781 * to the desired size as pages become free.
2782 *
2783 * By placing pages into the surplus state independent of the
2784 * overcommit value, we are allowing the surplus pool size to
2785 * exceed overcommit. There are few sane options here. Since
2786 * alloc_surplus_huge_page() is checking the global counter,
2787 * though, we'll note that we're not allowed to exceed surplus
2788 * and won't grow the pool anywhere else. Not until one of the
2789 * sysctls are changed, or the surplus pages go out of use.
2790 */
2791 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2792 min_count = max(count, min_count);
2793 try_to_free_low(h, min_count, nodes_allowed);
2794 while (min_count < persistent_huge_pages(h)) {
2795 if (!free_pool_huge_page(h, nodes_allowed, 0))
2796 break;
2797 cond_resched_lock(&hugetlb_lock);
2798 }
2799 while (count < persistent_huge_pages(h)) {
2800 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2801 break;
2802 }
2803out:
2804 h->max_huge_pages = persistent_huge_pages(h);
2805 spin_unlock(&hugetlb_lock);
2806
2807 NODEMASK_FREE(node_alloc_noretry);
2808
2809 return 0;
2810}
2811
2812#define HSTATE_ATTR_RO(_name) \
2813 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2814
2815#define HSTATE_ATTR(_name) \
2816 static struct kobj_attribute _name##_attr = \
2817 __ATTR(_name, 0644, _name##_show, _name##_store)
2818
2819static struct kobject *hugepages_kobj;
2820static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2821
2822static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2823
2824static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2825{
2826 int i;
2827
2828 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2829 if (hstate_kobjs[i] == kobj) {
2830 if (nidp)
2831 *nidp = NUMA_NO_NODE;
2832 return &hstates[i];
2833 }
2834
2835 return kobj_to_node_hstate(kobj, nidp);
2836}
2837
2838static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2839 struct kobj_attribute *attr, char *buf)
2840{
2841 struct hstate *h;
2842 unsigned long nr_huge_pages;
2843 int nid;
2844
2845 h = kobj_to_hstate(kobj, &nid);
2846 if (nid == NUMA_NO_NODE)
2847 nr_huge_pages = h->nr_huge_pages;
2848 else
2849 nr_huge_pages = h->nr_huge_pages_node[nid];
2850
2851 return sprintf(buf, "%lu\n", nr_huge_pages);
2852}
2853
2854static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2855 struct hstate *h, int nid,
2856 unsigned long count, size_t len)
2857{
2858 int err;
2859 nodemask_t nodes_allowed, *n_mask;
2860
2861 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2862 return -EINVAL;
2863
2864 if (nid == NUMA_NO_NODE) {
2865 /*
2866 * global hstate attribute
2867 */
2868 if (!(obey_mempolicy &&
2869 init_nodemask_of_mempolicy(&nodes_allowed)))
2870 n_mask = &node_states[N_MEMORY];
2871 else
2872 n_mask = &nodes_allowed;
2873 } else {
2874 /*
2875 * Node specific request. count adjustment happens in
2876 * set_max_huge_pages() after acquiring hugetlb_lock.
2877 */
2878 init_nodemask_of_node(&nodes_allowed, nid);
2879 n_mask = &nodes_allowed;
2880 }
2881
2882 err = set_max_huge_pages(h, count, nid, n_mask);
2883
2884 return err ? err : len;
2885}
2886
2887static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2888 struct kobject *kobj, const char *buf,
2889 size_t len)
2890{
2891 struct hstate *h;
2892 unsigned long count;
2893 int nid;
2894 int err;
2895
2896 err = kstrtoul(buf, 10, &count);
2897 if (err)
2898 return err;
2899
2900 h = kobj_to_hstate(kobj, &nid);
2901 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2902}
2903
2904static ssize_t nr_hugepages_show(struct kobject *kobj,
2905 struct kobj_attribute *attr, char *buf)
2906{
2907 return nr_hugepages_show_common(kobj, attr, buf);
2908}
2909
2910static ssize_t nr_hugepages_store(struct kobject *kobj,
2911 struct kobj_attribute *attr, const char *buf, size_t len)
2912{
2913 return nr_hugepages_store_common(false, kobj, buf, len);
2914}
2915HSTATE_ATTR(nr_hugepages);
2916
2917#ifdef CONFIG_NUMA
2918
2919/*
2920 * hstate attribute for optionally mempolicy-based constraint on persistent
2921 * huge page alloc/free.
2922 */
2923static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2924 struct kobj_attribute *attr, char *buf)
2925{
2926 return nr_hugepages_show_common(kobj, attr, buf);
2927}
2928
2929static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2930 struct kobj_attribute *attr, const char *buf, size_t len)
2931{
2932 return nr_hugepages_store_common(true, kobj, buf, len);
2933}
2934HSTATE_ATTR(nr_hugepages_mempolicy);
2935#endif
2936
2937
2938static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2939 struct kobj_attribute *attr, char *buf)
2940{
2941 struct hstate *h = kobj_to_hstate(kobj, NULL);
2942 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2943}
2944
2945static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2946 struct kobj_attribute *attr, const char *buf, size_t count)
2947{
2948 int err;
2949 unsigned long input;
2950 struct hstate *h = kobj_to_hstate(kobj, NULL);
2951
2952 if (hstate_is_gigantic(h))
2953 return -EINVAL;
2954
2955 err = kstrtoul(buf, 10, &input);
2956 if (err)
2957 return err;
2958
2959 spin_lock(&hugetlb_lock);
2960 h->nr_overcommit_huge_pages = input;
2961 spin_unlock(&hugetlb_lock);
2962
2963 return count;
2964}
2965HSTATE_ATTR(nr_overcommit_hugepages);
2966
2967static ssize_t free_hugepages_show(struct kobject *kobj,
2968 struct kobj_attribute *attr, char *buf)
2969{
2970 struct hstate *h;
2971 unsigned long free_huge_pages;
2972 int nid;
2973
2974 h = kobj_to_hstate(kobj, &nid);
2975 if (nid == NUMA_NO_NODE)
2976 free_huge_pages = h->free_huge_pages;
2977 else
2978 free_huge_pages = h->free_huge_pages_node[nid];
2979
2980 return sprintf(buf, "%lu\n", free_huge_pages);
2981}
2982HSTATE_ATTR_RO(free_hugepages);
2983
2984static ssize_t resv_hugepages_show(struct kobject *kobj,
2985 struct kobj_attribute *attr, char *buf)
2986{
2987 struct hstate *h = kobj_to_hstate(kobj, NULL);
2988 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2989}
2990HSTATE_ATTR_RO(resv_hugepages);
2991
2992static ssize_t surplus_hugepages_show(struct kobject *kobj,
2993 struct kobj_attribute *attr, char *buf)
2994{
2995 struct hstate *h;
2996 unsigned long surplus_huge_pages;
2997 int nid;
2998
2999 h = kobj_to_hstate(kobj, &nid);
3000 if (nid == NUMA_NO_NODE)
3001 surplus_huge_pages = h->surplus_huge_pages;
3002 else
3003 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3004
3005 return sprintf(buf, "%lu\n", surplus_huge_pages);
3006}
3007HSTATE_ATTR_RO(surplus_hugepages);
3008
3009static struct attribute *hstate_attrs[] = {
3010 &nr_hugepages_attr.attr,
3011 &nr_overcommit_hugepages_attr.attr,
3012 &free_hugepages_attr.attr,
3013 &resv_hugepages_attr.attr,
3014 &surplus_hugepages_attr.attr,
3015#ifdef CONFIG_NUMA
3016 &nr_hugepages_mempolicy_attr.attr,
3017#endif
3018 NULL,
3019};
3020
3021static const struct attribute_group hstate_attr_group = {
3022 .attrs = hstate_attrs,
3023};
3024
3025static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3026 struct kobject **hstate_kobjs,
3027 const struct attribute_group *hstate_attr_group)
3028{
3029 int retval;
3030 int hi = hstate_index(h);
3031
3032 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3033 if (!hstate_kobjs[hi])
3034 return -ENOMEM;
3035
3036 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3037 if (retval)
3038 kobject_put(hstate_kobjs[hi]);
3039
3040 return retval;
3041}
3042
3043static void __init hugetlb_sysfs_init(void)
3044{
3045 struct hstate *h;
3046 int err;
3047
3048 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3049 if (!hugepages_kobj)
3050 return;
3051
3052 for_each_hstate(h) {
3053 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3054 hstate_kobjs, &hstate_attr_group);
3055 if (err)
3056 pr_err("HugeTLB: Unable to add hstate %s", h->name);
3057 }
3058}
3059
3060#ifdef CONFIG_NUMA
3061
3062/*
3063 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3064 * with node devices in node_devices[] using a parallel array. The array
3065 * index of a node device or _hstate == node id.
3066 * This is here to avoid any static dependency of the node device driver, in
3067 * the base kernel, on the hugetlb module.
3068 */
3069struct node_hstate {
3070 struct kobject *hugepages_kobj;
3071 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3072};
3073static struct node_hstate node_hstates[MAX_NUMNODES];
3074
3075/*
3076 * A subset of global hstate attributes for node devices
3077 */
3078static struct attribute *per_node_hstate_attrs[] = {
3079 &nr_hugepages_attr.attr,
3080 &free_hugepages_attr.attr,
3081 &surplus_hugepages_attr.attr,
3082 NULL,
3083};
3084
3085static const struct attribute_group per_node_hstate_attr_group = {
3086 .attrs = per_node_hstate_attrs,
3087};
3088
3089/*
3090 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3091 * Returns node id via non-NULL nidp.
3092 */
3093static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3094{
3095 int nid;
3096
3097 for (nid = 0; nid < nr_node_ids; nid++) {
3098 struct node_hstate *nhs = &node_hstates[nid];
3099 int i;
3100 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3101 if (nhs->hstate_kobjs[i] == kobj) {
3102 if (nidp)
3103 *nidp = nid;
3104 return &hstates[i];
3105 }
3106 }
3107
3108 BUG();
3109 return NULL;
3110}
3111
3112/*
3113 * Unregister hstate attributes from a single node device.
3114 * No-op if no hstate attributes attached.
3115 */
3116static void hugetlb_unregister_node(struct node *node)
3117{
3118 struct hstate *h;
3119 struct node_hstate *nhs = &node_hstates[node->dev.id];
3120
3121 if (!nhs->hugepages_kobj)
3122 return; /* no hstate attributes */
3123
3124 for_each_hstate(h) {
3125 int idx = hstate_index(h);
3126 if (nhs->hstate_kobjs[idx]) {
3127 kobject_put(nhs->hstate_kobjs[idx]);
3128 nhs->hstate_kobjs[idx] = NULL;
3129 }
3130 }
3131
3132 kobject_put(nhs->hugepages_kobj);
3133 nhs->hugepages_kobj = NULL;
3134}
3135
3136
3137/*
3138 * Register hstate attributes for a single node device.
3139 * No-op if attributes already registered.
3140 */
3141static void hugetlb_register_node(struct node *node)
3142{
3143 struct hstate *h;
3144 struct node_hstate *nhs = &node_hstates[node->dev.id];
3145 int err;
3146
3147 if (nhs->hugepages_kobj)
3148 return; /* already allocated */
3149
3150 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3151 &node->dev.kobj);
3152 if (!nhs->hugepages_kobj)
3153 return;
3154
3155 for_each_hstate(h) {
3156 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3157 nhs->hstate_kobjs,
3158 &per_node_hstate_attr_group);
3159 if (err) {
3160 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3161 h->name, node->dev.id);
3162 hugetlb_unregister_node(node);
3163 break;
3164 }
3165 }
3166}
3167
3168/*
3169 * hugetlb init time: register hstate attributes for all registered node
3170 * devices of nodes that have memory. All on-line nodes should have
3171 * registered their associated device by this time.
3172 */
3173static void __init hugetlb_register_all_nodes(void)
3174{
3175 int nid;
3176
3177 for_each_node_state(nid, N_MEMORY) {
3178 struct node *node = node_devices[nid];
3179 if (node->dev.id == nid)
3180 hugetlb_register_node(node);
3181 }
3182
3183 /*
3184 * Let the node device driver know we're here so it can
3185 * [un]register hstate attributes on node hotplug.
3186 */
3187 register_hugetlbfs_with_node(hugetlb_register_node,
3188 hugetlb_unregister_node);
3189}
3190#else /* !CONFIG_NUMA */
3191
3192static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3193{
3194 BUG();
3195 if (nidp)
3196 *nidp = -1;
3197 return NULL;
3198}
3199
3200static void hugetlb_register_all_nodes(void) { }
3201
3202#endif
3203
3204static int __init hugetlb_init(void)
3205{
3206 int i;
3207
3208 if (!hugepages_supported()) {
3209 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3210 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3211 return 0;
3212 }
3213
3214 /*
3215 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3216 * architectures depend on setup being done here.
3217 */
3218 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3219 if (!parsed_default_hugepagesz) {
3220 /*
3221 * If we did not parse a default huge page size, set
3222 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3223 * number of huge pages for this default size was implicitly
3224 * specified, set that here as well.
3225 * Note that the implicit setting will overwrite an explicit
3226 * setting. A warning will be printed in this case.
3227 */
3228 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3229 if (default_hstate_max_huge_pages) {
3230 if (default_hstate.max_huge_pages) {
3231 char buf[32];
3232
3233 string_get_size(huge_page_size(&default_hstate),
3234 1, STRING_UNITS_2, buf, 32);
3235 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3236 default_hstate.max_huge_pages, buf);
3237 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3238 default_hstate_max_huge_pages);
3239 }
3240 default_hstate.max_huge_pages =
3241 default_hstate_max_huge_pages;
3242 }
3243 }
3244
3245 hugetlb_cma_check();
3246 hugetlb_init_hstates();
3247 gather_bootmem_prealloc();
3248 report_hugepages();
3249
3250 hugetlb_sysfs_init();
3251 hugetlb_register_all_nodes();
3252 hugetlb_cgroup_file_init();
3253
3254#ifdef CONFIG_SMP
3255 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3256#else
3257 num_fault_mutexes = 1;
3258#endif
3259 hugetlb_fault_mutex_table =
3260 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3261 GFP_KERNEL);
3262 BUG_ON(!hugetlb_fault_mutex_table);
3263
3264 for (i = 0; i < num_fault_mutexes; i++)
3265 mutex_init(&hugetlb_fault_mutex_table[i]);
3266 return 0;
3267}
3268subsys_initcall(hugetlb_init);
3269
3270/* Overwritten by architectures with more huge page sizes */
3271bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3272{
3273 return size == HPAGE_SIZE;
3274}
3275
3276void __init hugetlb_add_hstate(unsigned int order)
3277{
3278 struct hstate *h;
3279 unsigned long i;
3280
3281 if (size_to_hstate(PAGE_SIZE << order)) {
3282 return;
3283 }
3284 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3285 BUG_ON(order == 0);
3286 h = &hstates[hugetlb_max_hstate++];
3287 h->order = order;
3288 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
3289 h->nr_huge_pages = 0;
3290 h->free_huge_pages = 0;
3291 for (i = 0; i < MAX_NUMNODES; ++i)
3292 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3293 INIT_LIST_HEAD(&h->hugepage_activelist);
3294 h->next_nid_to_alloc = first_memory_node;
3295 h->next_nid_to_free = first_memory_node;
3296 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3297 huge_page_size(h)/1024);
3298
3299 parsed_hstate = h;
3300}
3301
3302/*
3303 * hugepages command line processing
3304 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3305 * specification. If not, ignore the hugepages value. hugepages can also
3306 * be the first huge page command line option in which case it implicitly
3307 * specifies the number of huge pages for the default size.
3308 */
3309static int __init hugepages_setup(char *s)
3310{
3311 unsigned long *mhp;
3312 static unsigned long *last_mhp;
3313
3314 if (!parsed_valid_hugepagesz) {
3315 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3316 parsed_valid_hugepagesz = true;
3317 return 0;
3318 }
3319
3320 /*
3321 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3322 * yet, so this hugepages= parameter goes to the "default hstate".
3323 * Otherwise, it goes with the previously parsed hugepagesz or
3324 * default_hugepagesz.
3325 */
3326 else if (!hugetlb_max_hstate)
3327 mhp = &default_hstate_max_huge_pages;
3328 else
3329 mhp = &parsed_hstate->max_huge_pages;
3330
3331 if (mhp == last_mhp) {
3332 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3333 return 0;
3334 }
3335
3336 if (sscanf(s, "%lu", mhp) <= 0)
3337 *mhp = 0;
3338
3339 /*
3340 * Global state is always initialized later in hugetlb_init.
3341 * But we need to allocate >= MAX_ORDER hstates here early to still
3342 * use the bootmem allocator.
3343 */
3344 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
3345 hugetlb_hstate_alloc_pages(parsed_hstate);
3346
3347 last_mhp = mhp;
3348
3349 return 1;
3350}
3351__setup("hugepages=", hugepages_setup);
3352
3353/*
3354 * hugepagesz command line processing
3355 * A specific huge page size can only be specified once with hugepagesz.
3356 * hugepagesz is followed by hugepages on the command line. The global
3357 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3358 * hugepagesz argument was valid.
3359 */
3360static int __init hugepagesz_setup(char *s)
3361{
3362 unsigned long size;
3363 struct hstate *h;
3364
3365 parsed_valid_hugepagesz = false;
3366 size = (unsigned long)memparse(s, NULL);
3367
3368 if (!arch_hugetlb_valid_size(size)) {
3369 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3370 return 0;
3371 }
3372
3373 h = size_to_hstate(size);
3374 if (h) {
3375 /*
3376 * hstate for this size already exists. This is normally
3377 * an error, but is allowed if the existing hstate is the
3378 * default hstate. More specifically, it is only allowed if
3379 * the number of huge pages for the default hstate was not
3380 * previously specified.
3381 */
3382 if (!parsed_default_hugepagesz || h != &default_hstate ||
3383 default_hstate.max_huge_pages) {
3384 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3385 return 0;
3386 }
3387
3388 /*
3389 * No need to call hugetlb_add_hstate() as hstate already
3390 * exists. But, do set parsed_hstate so that a following
3391 * hugepages= parameter will be applied to this hstate.
3392 */
3393 parsed_hstate = h;
3394 parsed_valid_hugepagesz = true;
3395 return 1;
3396 }
3397
3398 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3399 parsed_valid_hugepagesz = true;
3400 return 1;
3401}
3402__setup("hugepagesz=", hugepagesz_setup);
3403
3404/*
3405 * default_hugepagesz command line input
3406 * Only one instance of default_hugepagesz allowed on command line.
3407 */
3408static int __init default_hugepagesz_setup(char *s)
3409{
3410 unsigned long size;
3411
3412 parsed_valid_hugepagesz = false;
3413 if (parsed_default_hugepagesz) {
3414 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3415 return 0;
3416 }
3417
3418 size = (unsigned long)memparse(s, NULL);
3419
3420 if (!arch_hugetlb_valid_size(size)) {
3421 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3422 return 0;
3423 }
3424
3425 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3426 parsed_valid_hugepagesz = true;
3427 parsed_default_hugepagesz = true;
3428 default_hstate_idx = hstate_index(size_to_hstate(size));
3429
3430 /*
3431 * The number of default huge pages (for this size) could have been
3432 * specified as the first hugetlb parameter: hugepages=X. If so,
3433 * then default_hstate_max_huge_pages is set. If the default huge
3434 * page size is gigantic (>= MAX_ORDER), then the pages must be
3435 * allocated here from bootmem allocator.
3436 */
3437 if (default_hstate_max_huge_pages) {
3438 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3439 if (hstate_is_gigantic(&default_hstate))
3440 hugetlb_hstate_alloc_pages(&default_hstate);
3441 default_hstate_max_huge_pages = 0;
3442 }
3443
3444 return 1;
3445}
3446__setup("default_hugepagesz=", default_hugepagesz_setup);
3447
3448static unsigned int allowed_mems_nr(struct hstate *h)
3449{
3450 int node;
3451 unsigned int nr = 0;
3452 nodemask_t *mpol_allowed;
3453 unsigned int *array = h->free_huge_pages_node;
3454 gfp_t gfp_mask = htlb_alloc_mask(h);
3455
3456 mpol_allowed = policy_nodemask_current(gfp_mask);
3457
3458 for_each_node_mask(node, cpuset_current_mems_allowed) {
3459 if (!mpol_allowed ||
3460 (mpol_allowed && node_isset(node, *mpol_allowed)))
3461 nr += array[node];
3462 }
3463
3464 return nr;
3465}
3466
3467#ifdef CONFIG_SYSCTL
3468static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3469 void *buffer, size_t *length,
3470 loff_t *ppos, unsigned long *out)
3471{
3472 struct ctl_table dup_table;
3473
3474 /*
3475 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3476 * can duplicate the @table and alter the duplicate of it.
3477 */
3478 dup_table = *table;
3479 dup_table.data = out;
3480
3481 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3482}
3483
3484static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3485 struct ctl_table *table, int write,
3486 void *buffer, size_t *length, loff_t *ppos)
3487{
3488 struct hstate *h = &default_hstate;
3489 unsigned long tmp = h->max_huge_pages;
3490 int ret;
3491
3492 if (!hugepages_supported())
3493 return -EOPNOTSUPP;
3494
3495 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3496 &tmp);
3497 if (ret)
3498 goto out;
3499
3500 if (write)
3501 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3502 NUMA_NO_NODE, tmp, *length);
3503out:
3504 return ret;
3505}
3506
3507int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3508 void *buffer, size_t *length, loff_t *ppos)
3509{
3510
3511 return hugetlb_sysctl_handler_common(false, table, write,
3512 buffer, length, ppos);
3513}
3514
3515#ifdef CONFIG_NUMA
3516int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3517 void *buffer, size_t *length, loff_t *ppos)
3518{
3519 return hugetlb_sysctl_handler_common(true, table, write,
3520 buffer, length, ppos);
3521}
3522#endif /* CONFIG_NUMA */
3523
3524int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3525 void *buffer, size_t *length, loff_t *ppos)
3526{
3527 struct hstate *h = &default_hstate;
3528 unsigned long tmp;
3529 int ret;
3530
3531 if (!hugepages_supported())
3532 return -EOPNOTSUPP;
3533
3534 tmp = h->nr_overcommit_huge_pages;
3535
3536 if (write && hstate_is_gigantic(h))
3537 return -EINVAL;
3538
3539 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3540 &tmp);
3541 if (ret)
3542 goto out;
3543
3544 if (write) {
3545 spin_lock(&hugetlb_lock);
3546 h->nr_overcommit_huge_pages = tmp;
3547 spin_unlock(&hugetlb_lock);
3548 }
3549out:
3550 return ret;
3551}
3552
3553#endif /* CONFIG_SYSCTL */
3554
3555void hugetlb_report_meminfo(struct seq_file *m)
3556{
3557 struct hstate *h;
3558 unsigned long total = 0;
3559
3560 if (!hugepages_supported())
3561 return;
3562
3563 for_each_hstate(h) {
3564 unsigned long count = h->nr_huge_pages;
3565
3566 total += (PAGE_SIZE << huge_page_order(h)) * count;
3567
3568 if (h == &default_hstate)
3569 seq_printf(m,
3570 "HugePages_Total: %5lu\n"
3571 "HugePages_Free: %5lu\n"
3572 "HugePages_Rsvd: %5lu\n"
3573 "HugePages_Surp: %5lu\n"
3574 "Hugepagesize: %8lu kB\n",
3575 count,
3576 h->free_huge_pages,
3577 h->resv_huge_pages,
3578 h->surplus_huge_pages,
3579 (PAGE_SIZE << huge_page_order(h)) / 1024);
3580 }
3581
3582 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3583}
3584
3585int hugetlb_report_node_meminfo(int nid, char *buf)
3586{
3587 struct hstate *h = &default_hstate;
3588 if (!hugepages_supported())
3589 return 0;
3590 return sprintf(buf,
3591 "Node %d HugePages_Total: %5u\n"
3592 "Node %d HugePages_Free: %5u\n"
3593 "Node %d HugePages_Surp: %5u\n",
3594 nid, h->nr_huge_pages_node[nid],
3595 nid, h->free_huge_pages_node[nid],
3596 nid, h->surplus_huge_pages_node[nid]);
3597}
3598
3599void hugetlb_show_meminfo(void)
3600{
3601 struct hstate *h;
3602 int nid;
3603
3604 if (!hugepages_supported())
3605 return;
3606
3607 for_each_node_state(nid, N_MEMORY)
3608 for_each_hstate(h)
3609 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3610 nid,
3611 h->nr_huge_pages_node[nid],
3612 h->free_huge_pages_node[nid],
3613 h->surplus_huge_pages_node[nid],
3614 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3615}
3616
3617void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3618{
3619 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3620 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3621}
3622
3623/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3624unsigned long hugetlb_total_pages(void)
3625{
3626 struct hstate *h;
3627 unsigned long nr_total_pages = 0;
3628
3629 for_each_hstate(h)
3630 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3631 return nr_total_pages;
3632}
3633
3634static int hugetlb_acct_memory(struct hstate *h, long delta)
3635{
3636 int ret = -ENOMEM;
3637
3638 spin_lock(&hugetlb_lock);
3639 /*
3640 * When cpuset is configured, it breaks the strict hugetlb page
3641 * reservation as the accounting is done on a global variable. Such
3642 * reservation is completely rubbish in the presence of cpuset because
3643 * the reservation is not checked against page availability for the
3644 * current cpuset. Application can still potentially OOM'ed by kernel
3645 * with lack of free htlb page in cpuset that the task is in.
3646 * Attempt to enforce strict accounting with cpuset is almost
3647 * impossible (or too ugly) because cpuset is too fluid that
3648 * task or memory node can be dynamically moved between cpusets.
3649 *
3650 * The change of semantics for shared hugetlb mapping with cpuset is
3651 * undesirable. However, in order to preserve some of the semantics,
3652 * we fall back to check against current free page availability as
3653 * a best attempt and hopefully to minimize the impact of changing
3654 * semantics that cpuset has.
3655 *
3656 * Apart from cpuset, we also have memory policy mechanism that
3657 * also determines from which node the kernel will allocate memory
3658 * in a NUMA system. So similar to cpuset, we also should consider
3659 * the memory policy of the current task. Similar to the description
3660 * above.
3661 */
3662 if (delta > 0) {
3663 if (gather_surplus_pages(h, delta) < 0)
3664 goto out;
3665
3666 if (delta > allowed_mems_nr(h)) {
3667 return_unused_surplus_pages(h, delta);
3668 goto out;
3669 }
3670 }
3671
3672 ret = 0;
3673 if (delta < 0)
3674 return_unused_surplus_pages(h, (unsigned long) -delta);
3675
3676out:
3677 spin_unlock(&hugetlb_lock);
3678 return ret;
3679}
3680
3681static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3682{
3683 struct resv_map *resv = vma_resv_map(vma);
3684
3685 /*
3686 * This new VMA should share its siblings reservation map if present.
3687 * The VMA will only ever have a valid reservation map pointer where
3688 * it is being copied for another still existing VMA. As that VMA
3689 * has a reference to the reservation map it cannot disappear until
3690 * after this open call completes. It is therefore safe to take a
3691 * new reference here without additional locking.
3692 */
3693 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3694 kref_get(&resv->refs);
3695}
3696
3697static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3698{
3699 struct hstate *h = hstate_vma(vma);
3700 struct resv_map *resv = vma_resv_map(vma);
3701 struct hugepage_subpool *spool = subpool_vma(vma);
3702 unsigned long reserve, start, end;
3703 long gbl_reserve;
3704
3705 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3706 return;
3707
3708 start = vma_hugecache_offset(h, vma, vma->vm_start);
3709 end = vma_hugecache_offset(h, vma, vma->vm_end);
3710
3711 reserve = (end - start) - region_count(resv, start, end);
3712 hugetlb_cgroup_uncharge_counter(resv, start, end);
3713 if (reserve) {
3714 /*
3715 * Decrement reserve counts. The global reserve count may be
3716 * adjusted if the subpool has a minimum size.
3717 */
3718 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3719 hugetlb_acct_memory(h, -gbl_reserve);
3720 }
3721
3722 kref_put(&resv->refs, resv_map_release);
3723}
3724
3725static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3726{
3727 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3728 return -EINVAL;
3729 return 0;
3730}
3731
3732static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3733{
3734 struct hstate *hstate = hstate_vma(vma);
3735
3736 return 1UL << huge_page_shift(hstate);
3737}
3738
3739/*
3740 * We cannot handle pagefaults against hugetlb pages at all. They cause
3741 * handle_mm_fault() to try to instantiate regular-sized pages in the
3742 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3743 * this far.
3744 */
3745static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3746{
3747 BUG();
3748 return 0;
3749}
3750
3751/*
3752 * When a new function is introduced to vm_operations_struct and added
3753 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3754 * This is because under System V memory model, mappings created via
3755 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3756 * their original vm_ops are overwritten with shm_vm_ops.
3757 */
3758const struct vm_operations_struct hugetlb_vm_ops = {
3759 .fault = hugetlb_vm_op_fault,
3760 .open = hugetlb_vm_op_open,
3761 .close = hugetlb_vm_op_close,
3762 .split = hugetlb_vm_op_split,
3763 .pagesize = hugetlb_vm_op_pagesize,
3764};
3765
3766static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3767 int writable)
3768{
3769 pte_t entry;
3770
3771 if (writable) {
3772 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3773 vma->vm_page_prot)));
3774 } else {
3775 entry = huge_pte_wrprotect(mk_huge_pte(page,
3776 vma->vm_page_prot));
3777 }
3778 entry = pte_mkyoung(entry);
3779 entry = pte_mkhuge(entry);
3780 entry = arch_make_huge_pte(entry, vma, page, writable);
3781
3782 return entry;
3783}
3784
3785static void set_huge_ptep_writable(struct vm_area_struct *vma,
3786 unsigned long address, pte_t *ptep)
3787{
3788 pte_t entry;
3789
3790 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3791 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3792 update_mmu_cache(vma, address, ptep);
3793}
3794
3795bool is_hugetlb_entry_migration(pte_t pte)
3796{
3797 swp_entry_t swp;
3798
3799 if (huge_pte_none(pte) || pte_present(pte))
3800 return false;
3801 swp = pte_to_swp_entry(pte);
3802 if (non_swap_entry(swp) && is_migration_entry(swp))
3803 return true;
3804 else
3805 return false;
3806}
3807
3808static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3809{
3810 swp_entry_t swp;
3811
3812 if (huge_pte_none(pte) || pte_present(pte))
3813 return 0;
3814 swp = pte_to_swp_entry(pte);
3815 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3816 return 1;
3817 else
3818 return 0;
3819}
3820
3821int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3822 struct vm_area_struct *vma)
3823{
3824 pte_t *src_pte, *dst_pte, entry, dst_entry;
3825 struct page *ptepage;
3826 unsigned long addr;
3827 int cow;
3828 struct hstate *h = hstate_vma(vma);
3829 unsigned long sz = huge_page_size(h);
3830 struct address_space *mapping = vma->vm_file->f_mapping;
3831 struct mmu_notifier_range range;
3832 int ret = 0;
3833
3834 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3835
3836 if (cow) {
3837 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3838 vma->vm_start,
3839 vma->vm_end);
3840 mmu_notifier_invalidate_range_start(&range);
3841 } else {
3842 /*
3843 * For shared mappings i_mmap_rwsem must be held to call
3844 * huge_pte_alloc, otherwise the returned ptep could go
3845 * away if part of a shared pmd and another thread calls
3846 * huge_pmd_unshare.
3847 */
3848 i_mmap_lock_read(mapping);
3849 }
3850
3851 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3852 spinlock_t *src_ptl, *dst_ptl;
3853 src_pte = huge_pte_offset(src, addr, sz);
3854 if (!src_pte)
3855 continue;
3856 dst_pte = huge_pte_alloc(dst, addr, sz);
3857 if (!dst_pte) {
3858 ret = -ENOMEM;
3859 break;
3860 }
3861
3862 /*
3863 * If the pagetables are shared don't copy or take references.
3864 * dst_pte == src_pte is the common case of src/dest sharing.
3865 *
3866 * However, src could have 'unshared' and dst shares with
3867 * another vma. If dst_pte !none, this implies sharing.
3868 * Check here before taking page table lock, and once again
3869 * after taking the lock below.
3870 */
3871 dst_entry = huge_ptep_get(dst_pte);
3872 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3873 continue;
3874
3875 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3876 src_ptl = huge_pte_lockptr(h, src, src_pte);
3877 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3878 entry = huge_ptep_get(src_pte);
3879 dst_entry = huge_ptep_get(dst_pte);
3880 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3881 /*
3882 * Skip if src entry none. Also, skip in the
3883 * unlikely case dst entry !none as this implies
3884 * sharing with another vma.
3885 */
3886 ;
3887 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3888 is_hugetlb_entry_hwpoisoned(entry))) {
3889 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3890
3891 if (is_write_migration_entry(swp_entry) && cow) {
3892 /*
3893 * COW mappings require pages in both
3894 * parent and child to be set to read.
3895 */
3896 make_migration_entry_read(&swp_entry);
3897 entry = swp_entry_to_pte(swp_entry);
3898 set_huge_swap_pte_at(src, addr, src_pte,
3899 entry, sz);
3900 }
3901 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3902 } else {
3903 if (cow) {
3904 /*
3905 * No need to notify as we are downgrading page
3906 * table protection not changing it to point
3907 * to a new page.
3908 *
3909 * See Documentation/vm/mmu_notifier.rst
3910 */
3911 huge_ptep_set_wrprotect(src, addr, src_pte);
3912 }
3913 entry = huge_ptep_get(src_pte);
3914 ptepage = pte_page(entry);
3915 get_page(ptepage);
3916 page_dup_rmap(ptepage, true);
3917 set_huge_pte_at(dst, addr, dst_pte, entry);
3918 hugetlb_count_add(pages_per_huge_page(h), dst);
3919 }
3920 spin_unlock(src_ptl);
3921 spin_unlock(dst_ptl);
3922 }
3923
3924 if (cow)
3925 mmu_notifier_invalidate_range_end(&range);
3926 else
3927 i_mmap_unlock_read(mapping);
3928
3929 return ret;
3930}
3931
3932void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3933 unsigned long start, unsigned long end,
3934 struct page *ref_page)
3935{
3936 struct mm_struct *mm = vma->vm_mm;
3937 unsigned long address;
3938 pte_t *ptep;
3939 pte_t pte;
3940 spinlock_t *ptl;
3941 struct page *page;
3942 struct hstate *h = hstate_vma(vma);
3943 unsigned long sz = huge_page_size(h);
3944 struct mmu_notifier_range range;
3945
3946 WARN_ON(!is_vm_hugetlb_page(vma));
3947 BUG_ON(start & ~huge_page_mask(h));
3948 BUG_ON(end & ~huge_page_mask(h));
3949
3950 /*
3951 * This is a hugetlb vma, all the pte entries should point
3952 * to huge page.
3953 */
3954 tlb_change_page_size(tlb, sz);
3955 tlb_start_vma(tlb, vma);
3956
3957 /*
3958 * If sharing possible, alert mmu notifiers of worst case.
3959 */
3960 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3961 end);
3962 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3963 mmu_notifier_invalidate_range_start(&range);
3964 address = start;
3965 for (; address < end; address += sz) {
3966 ptep = huge_pte_offset(mm, address, sz);
3967 if (!ptep)
3968 continue;
3969
3970 ptl = huge_pte_lock(h, mm, ptep);
3971 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
3972 spin_unlock(ptl);
3973 /*
3974 * We just unmapped a page of PMDs by clearing a PUD.
3975 * The caller's TLB flush range should cover this area.
3976 */
3977 continue;
3978 }
3979
3980 pte = huge_ptep_get(ptep);
3981 if (huge_pte_none(pte)) {
3982 spin_unlock(ptl);
3983 continue;
3984 }
3985
3986 /*
3987 * Migrating hugepage or HWPoisoned hugepage is already
3988 * unmapped and its refcount is dropped, so just clear pte here.
3989 */
3990 if (unlikely(!pte_present(pte))) {
3991 huge_pte_clear(mm, address, ptep, sz);
3992 spin_unlock(ptl);
3993 continue;
3994 }
3995
3996 page = pte_page(pte);
3997 /*
3998 * If a reference page is supplied, it is because a specific
3999 * page is being unmapped, not a range. Ensure the page we
4000 * are about to unmap is the actual page of interest.
4001 */
4002 if (ref_page) {
4003 if (page != ref_page) {
4004 spin_unlock(ptl);
4005 continue;
4006 }
4007 /*
4008 * Mark the VMA as having unmapped its page so that
4009 * future faults in this VMA will fail rather than
4010 * looking like data was lost
4011 */
4012 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
4013 }
4014
4015 pte = huge_ptep_get_and_clear(mm, address, ptep);
4016 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
4017 if (huge_pte_dirty(pte))
4018 set_page_dirty(page);
4019
4020 hugetlb_count_sub(pages_per_huge_page(h), mm);
4021 page_remove_rmap(page, true);
4022
4023 spin_unlock(ptl);
4024 tlb_remove_page_size(tlb, page, huge_page_size(h));
4025 /*
4026 * Bail out after unmapping reference page if supplied
4027 */
4028 if (ref_page)
4029 break;
4030 }
4031 mmu_notifier_invalidate_range_end(&range);
4032 tlb_end_vma(tlb, vma);
4033}
4034
4035void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4036 struct vm_area_struct *vma, unsigned long start,
4037 unsigned long end, struct page *ref_page)
4038{
4039 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4040
4041 /*
4042 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4043 * test will fail on a vma being torn down, and not grab a page table
4044 * on its way out. We're lucky that the flag has such an appropriate
4045 * name, and can in fact be safely cleared here. We could clear it
4046 * before the __unmap_hugepage_range above, but all that's necessary
4047 * is to clear it before releasing the i_mmap_rwsem. This works
4048 * because in the context this is called, the VMA is about to be
4049 * destroyed and the i_mmap_rwsem is held.
4050 */
4051 vma->vm_flags &= ~VM_MAYSHARE;
4052}
4053
4054void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
4055 unsigned long end, struct page *ref_page)
4056{
4057 struct mm_struct *mm;
4058 struct mmu_gather tlb;
4059 unsigned long tlb_start = start;
4060 unsigned long tlb_end = end;
4061
4062 /*
4063 * If shared PMDs were possibly used within this vma range, adjust
4064 * start/end for worst case tlb flushing.
4065 * Note that we can not be sure if PMDs are shared until we try to
4066 * unmap pages. However, we want to make sure TLB flushing covers
4067 * the largest possible range.
4068 */
4069 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
4070
4071 mm = vma->vm_mm;
4072
4073 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
4074 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
4075 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
4076}
4077
4078/*
4079 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4080 * mappping it owns the reserve page for. The intention is to unmap the page
4081 * from other VMAs and let the children be SIGKILLed if they are faulting the
4082 * same region.
4083 */
4084static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4085 struct page *page, unsigned long address)
4086{
4087 struct hstate *h = hstate_vma(vma);
4088 struct vm_area_struct *iter_vma;
4089 struct address_space *mapping;
4090 pgoff_t pgoff;
4091
4092 /*
4093 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4094 * from page cache lookup which is in HPAGE_SIZE units.
4095 */
4096 address = address & huge_page_mask(h);
4097 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4098 vma->vm_pgoff;
4099 mapping = vma->vm_file->f_mapping;
4100
4101 /*
4102 * Take the mapping lock for the duration of the table walk. As
4103 * this mapping should be shared between all the VMAs,
4104 * __unmap_hugepage_range() is called as the lock is already held
4105 */
4106 i_mmap_lock_write(mapping);
4107 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
4108 /* Do not unmap the current VMA */
4109 if (iter_vma == vma)
4110 continue;
4111
4112 /*
4113 * Shared VMAs have their own reserves and do not affect
4114 * MAP_PRIVATE accounting but it is possible that a shared
4115 * VMA is using the same page so check and skip such VMAs.
4116 */
4117 if (iter_vma->vm_flags & VM_MAYSHARE)
4118 continue;
4119
4120 /*
4121 * Unmap the page from other VMAs without their own reserves.
4122 * They get marked to be SIGKILLed if they fault in these
4123 * areas. This is because a future no-page fault on this VMA
4124 * could insert a zeroed page instead of the data existing
4125 * from the time of fork. This would look like data corruption
4126 */
4127 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
4128 unmap_hugepage_range(iter_vma, address,
4129 address + huge_page_size(h), page);
4130 }
4131 i_mmap_unlock_write(mapping);
4132}
4133
4134/*
4135 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4136 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4137 * cannot race with other handlers or page migration.
4138 * Keep the pte_same checks anyway to make transition from the mutex easier.
4139 */
4140static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
4141 unsigned long address, pte_t *ptep,
4142 struct page *pagecache_page, spinlock_t *ptl)
4143{
4144 pte_t pte;
4145 struct hstate *h = hstate_vma(vma);
4146 struct page *old_page, *new_page;
4147 int outside_reserve = 0;
4148 vm_fault_t ret = 0;
4149 unsigned long haddr = address & huge_page_mask(h);
4150 struct mmu_notifier_range range;
4151
4152 pte = huge_ptep_get(ptep);
4153 old_page = pte_page(pte);
4154
4155retry_avoidcopy:
4156 /* If no-one else is actually using this page, avoid the copy
4157 * and just make the page writable */
4158 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
4159 page_move_anon_rmap(old_page, vma);
4160 set_huge_ptep_writable(vma, haddr, ptep);
4161 return 0;
4162 }
4163
4164 /*
4165 * If the process that created a MAP_PRIVATE mapping is about to
4166 * perform a COW due to a shared page count, attempt to satisfy
4167 * the allocation without using the existing reserves. The pagecache
4168 * page is used to determine if the reserve at this address was
4169 * consumed or not. If reserves were used, a partial faulted mapping
4170 * at the time of fork() could consume its reserves on COW instead
4171 * of the full address range.
4172 */
4173 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4174 old_page != pagecache_page)
4175 outside_reserve = 1;
4176
4177 get_page(old_page);
4178
4179 /*
4180 * Drop page table lock as buddy allocator may be called. It will
4181 * be acquired again before returning to the caller, as expected.
4182 */
4183 spin_unlock(ptl);
4184 new_page = alloc_huge_page(vma, haddr, outside_reserve);
4185
4186 if (IS_ERR(new_page)) {
4187 /*
4188 * If a process owning a MAP_PRIVATE mapping fails to COW,
4189 * it is due to references held by a child and an insufficient
4190 * huge page pool. To guarantee the original mappers
4191 * reliability, unmap the page from child processes. The child
4192 * may get SIGKILLed if it later faults.
4193 */
4194 if (outside_reserve) {
4195 put_page(old_page);
4196 BUG_ON(huge_pte_none(pte));
4197 unmap_ref_private(mm, vma, old_page, haddr);
4198 BUG_ON(huge_pte_none(pte));
4199 spin_lock(ptl);
4200 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4201 if (likely(ptep &&
4202 pte_same(huge_ptep_get(ptep), pte)))
4203 goto retry_avoidcopy;
4204 /*
4205 * race occurs while re-acquiring page table
4206 * lock, and our job is done.
4207 */
4208 return 0;
4209 }
4210
4211 ret = vmf_error(PTR_ERR(new_page));
4212 goto out_release_old;
4213 }
4214
4215 /*
4216 * When the original hugepage is shared one, it does not have
4217 * anon_vma prepared.
4218 */
4219 if (unlikely(anon_vma_prepare(vma))) {
4220 ret = VM_FAULT_OOM;
4221 goto out_release_all;
4222 }
4223
4224 copy_user_huge_page(new_page, old_page, address, vma,
4225 pages_per_huge_page(h));
4226 __SetPageUptodate(new_page);
4227
4228 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4229 haddr + huge_page_size(h));
4230 mmu_notifier_invalidate_range_start(&range);
4231
4232 /*
4233 * Retake the page table lock to check for racing updates
4234 * before the page tables are altered
4235 */
4236 spin_lock(ptl);
4237 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4238 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4239 ClearPagePrivate(new_page);
4240
4241 /* Break COW */
4242 huge_ptep_clear_flush(vma, haddr, ptep);
4243 mmu_notifier_invalidate_range(mm, range.start, range.end);
4244 set_huge_pte_at(mm, haddr, ptep,
4245 make_huge_pte(vma, new_page, 1));
4246 page_remove_rmap(old_page, true);
4247 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4248 set_page_huge_active(new_page);
4249 /* Make the old page be freed below */
4250 new_page = old_page;
4251 }
4252 spin_unlock(ptl);
4253 mmu_notifier_invalidate_range_end(&range);
4254out_release_all:
4255 restore_reserve_on_error(h, vma, haddr, new_page);
4256 put_page(new_page);
4257out_release_old:
4258 put_page(old_page);
4259
4260 spin_lock(ptl); /* Caller expects lock to be held */
4261 return ret;
4262}
4263
4264/* Return the pagecache page at a given address within a VMA */
4265static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4266 struct vm_area_struct *vma, unsigned long address)
4267{
4268 struct address_space *mapping;
4269 pgoff_t idx;
4270
4271 mapping = vma->vm_file->f_mapping;
4272 idx = vma_hugecache_offset(h, vma, address);
4273
4274 return find_lock_page(mapping, idx);
4275}
4276
4277/*
4278 * Return whether there is a pagecache page to back given address within VMA.
4279 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4280 */
4281static bool hugetlbfs_pagecache_present(struct hstate *h,
4282 struct vm_area_struct *vma, unsigned long address)
4283{
4284 struct address_space *mapping;
4285 pgoff_t idx;
4286 struct page *page;
4287
4288 mapping = vma->vm_file->f_mapping;
4289 idx = vma_hugecache_offset(h, vma, address);
4290
4291 page = find_get_page(mapping, idx);
4292 if (page)
4293 put_page(page);
4294 return page != NULL;
4295}
4296
4297int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4298 pgoff_t idx)
4299{
4300 struct inode *inode = mapping->host;
4301 struct hstate *h = hstate_inode(inode);
4302 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4303
4304 if (err)
4305 return err;
4306 ClearPagePrivate(page);
4307
4308 /*
4309 * set page dirty so that it will not be removed from cache/file
4310 * by non-hugetlbfs specific code paths.
4311 */
4312 set_page_dirty(page);
4313
4314 spin_lock(&inode->i_lock);
4315 inode->i_blocks += blocks_per_huge_page(h);
4316 spin_unlock(&inode->i_lock);
4317 return 0;
4318}
4319
4320static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4321 struct vm_area_struct *vma,
4322 struct address_space *mapping, pgoff_t idx,
4323 unsigned long address, pte_t *ptep, unsigned int flags)
4324{
4325 struct hstate *h = hstate_vma(vma);
4326 vm_fault_t ret = VM_FAULT_SIGBUS;
4327 int anon_rmap = 0;
4328 unsigned long size;
4329 struct page *page;
4330 pte_t new_pte;
4331 spinlock_t *ptl;
4332 unsigned long haddr = address & huge_page_mask(h);
4333 bool new_page = false;
4334
4335 /*
4336 * Currently, we are forced to kill the process in the event the
4337 * original mapper has unmapped pages from the child due to a failed
4338 * COW. Warn that such a situation has occurred as it may not be obvious
4339 */
4340 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4341 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4342 current->pid);
4343 return ret;
4344 }
4345
4346 /*
4347 * We can not race with truncation due to holding i_mmap_rwsem.
4348 * i_size is modified when holding i_mmap_rwsem, so check here
4349 * once for faults beyond end of file.
4350 */
4351 size = i_size_read(mapping->host) >> huge_page_shift(h);
4352 if (idx >= size)
4353 goto out;
4354
4355retry:
4356 page = find_lock_page(mapping, idx);
4357 if (!page) {
4358 /*
4359 * Check for page in userfault range
4360 */
4361 if (userfaultfd_missing(vma)) {
4362 u32 hash;
4363 struct vm_fault vmf = {
4364 .vma = vma,
4365 .address = haddr,
4366 .flags = flags,
4367 /*
4368 * Hard to debug if it ends up being
4369 * used by a callee that assumes
4370 * something about the other
4371 * uninitialized fields... same as in
4372 * memory.c
4373 */
4374 };
4375
4376 /*
4377 * hugetlb_fault_mutex and i_mmap_rwsem must be
4378 * dropped before handling userfault. Reacquire
4379 * after handling fault to make calling code simpler.
4380 */
4381 hash = hugetlb_fault_mutex_hash(mapping, idx);
4382 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4383 i_mmap_unlock_read(mapping);
4384 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
4385 i_mmap_lock_read(mapping);
4386 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4387 goto out;
4388 }
4389
4390 page = alloc_huge_page(vma, haddr, 0);
4391 if (IS_ERR(page)) {
4392 /*
4393 * Returning error will result in faulting task being
4394 * sent SIGBUS. The hugetlb fault mutex prevents two
4395 * tasks from racing to fault in the same page which
4396 * could result in false unable to allocate errors.
4397 * Page migration does not take the fault mutex, but
4398 * does a clear then write of pte's under page table
4399 * lock. Page fault code could race with migration,
4400 * notice the clear pte and try to allocate a page
4401 * here. Before returning error, get ptl and make
4402 * sure there really is no pte entry.
4403 */
4404 ptl = huge_pte_lock(h, mm, ptep);
4405 if (!huge_pte_none(huge_ptep_get(ptep))) {
4406 ret = 0;
4407 spin_unlock(ptl);
4408 goto out;
4409 }
4410 spin_unlock(ptl);
4411 ret = vmf_error(PTR_ERR(page));
4412 goto out;
4413 }
4414 clear_huge_page(page, address, pages_per_huge_page(h));
4415 __SetPageUptodate(page);
4416 new_page = true;
4417
4418 if (vma->vm_flags & VM_MAYSHARE) {
4419 int err = huge_add_to_page_cache(page, mapping, idx);
4420 if (err) {
4421 put_page(page);
4422 if (err == -EEXIST)
4423 goto retry;
4424 goto out;
4425 }
4426 } else {
4427 lock_page(page);
4428 if (unlikely(anon_vma_prepare(vma))) {
4429 ret = VM_FAULT_OOM;
4430 goto backout_unlocked;
4431 }
4432 anon_rmap = 1;
4433 }
4434 } else {
4435 /*
4436 * If memory error occurs between mmap() and fault, some process
4437 * don't have hwpoisoned swap entry for errored virtual address.
4438 * So we need to block hugepage fault by PG_hwpoison bit check.
4439 */
4440 if (unlikely(PageHWPoison(page))) {
4441 ret = VM_FAULT_HWPOISON |
4442 VM_FAULT_SET_HINDEX(hstate_index(h));
4443 goto backout_unlocked;
4444 }
4445 }
4446
4447 /*
4448 * If we are going to COW a private mapping later, we examine the
4449 * pending reservations for this page now. This will ensure that
4450 * any allocations necessary to record that reservation occur outside
4451 * the spinlock.
4452 */
4453 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4454 if (vma_needs_reservation(h, vma, haddr) < 0) {
4455 ret = VM_FAULT_OOM;
4456 goto backout_unlocked;
4457 }
4458 /* Just decrements count, does not deallocate */
4459 vma_end_reservation(h, vma, haddr);
4460 }
4461
4462 ptl = huge_pte_lock(h, mm, ptep);
4463 ret = 0;
4464 if (!huge_pte_none(huge_ptep_get(ptep)))
4465 goto backout;
4466
4467 if (anon_rmap) {
4468 ClearPagePrivate(page);
4469 hugepage_add_new_anon_rmap(page, vma, haddr);
4470 } else
4471 page_dup_rmap(page, true);
4472 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4473 && (vma->vm_flags & VM_SHARED)));
4474 set_huge_pte_at(mm, haddr, ptep, new_pte);
4475
4476 hugetlb_count_add(pages_per_huge_page(h), mm);
4477 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4478 /* Optimization, do the COW without a second fault */
4479 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4480 }
4481
4482 spin_unlock(ptl);
4483
4484 /*
4485 * Only make newly allocated pages active. Existing pages found
4486 * in the pagecache could be !page_huge_active() if they have been
4487 * isolated for migration.
4488 */
4489 if (new_page)
4490 set_page_huge_active(page);
4491
4492 unlock_page(page);
4493out:
4494 return ret;
4495
4496backout:
4497 spin_unlock(ptl);
4498backout_unlocked:
4499 unlock_page(page);
4500 restore_reserve_on_error(h, vma, haddr, page);
4501 put_page(page);
4502 goto out;
4503}
4504
4505#ifdef CONFIG_SMP
4506u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4507{
4508 unsigned long key[2];
4509 u32 hash;
4510
4511 key[0] = (unsigned long) mapping;
4512 key[1] = idx;
4513
4514 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4515
4516 return hash & (num_fault_mutexes - 1);
4517}
4518#else
4519/*
4520 * For uniprocesor systems we always use a single mutex, so just
4521 * return 0 and avoid the hashing overhead.
4522 */
4523u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4524{
4525 return 0;
4526}
4527#endif
4528
4529vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4530 unsigned long address, unsigned int flags)
4531{
4532 pte_t *ptep, entry;
4533 spinlock_t *ptl;
4534 vm_fault_t ret;
4535 u32 hash;
4536 pgoff_t idx;
4537 struct page *page = NULL;
4538 struct page *pagecache_page = NULL;
4539 struct hstate *h = hstate_vma(vma);
4540 struct address_space *mapping;
4541 int need_wait_lock = 0;
4542 unsigned long haddr = address & huge_page_mask(h);
4543
4544 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4545 if (ptep) {
4546 /*
4547 * Since we hold no locks, ptep could be stale. That is
4548 * OK as we are only making decisions based on content and
4549 * not actually modifying content here.
4550 */
4551 entry = huge_ptep_get(ptep);
4552 if (unlikely(is_hugetlb_entry_migration(entry))) {
4553 migration_entry_wait_huge(vma, mm, ptep);
4554 return 0;
4555 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4556 return VM_FAULT_HWPOISON_LARGE |
4557 VM_FAULT_SET_HINDEX(hstate_index(h));
4558 }
4559
4560 /*
4561 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4562 * until finished with ptep. This serves two purposes:
4563 * 1) It prevents huge_pmd_unshare from being called elsewhere
4564 * and making the ptep no longer valid.
4565 * 2) It synchronizes us with i_size modifications during truncation.
4566 *
4567 * ptep could have already be assigned via huge_pte_offset. That
4568 * is OK, as huge_pte_alloc will return the same value unless
4569 * something has changed.
4570 */
4571 mapping = vma->vm_file->f_mapping;
4572 i_mmap_lock_read(mapping);
4573 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4574 if (!ptep) {
4575 i_mmap_unlock_read(mapping);
4576 return VM_FAULT_OOM;
4577 }
4578
4579 /*
4580 * Serialize hugepage allocation and instantiation, so that we don't
4581 * get spurious allocation failures if two CPUs race to instantiate
4582 * the same page in the page cache.
4583 */
4584 idx = vma_hugecache_offset(h, vma, haddr);
4585 hash = hugetlb_fault_mutex_hash(mapping, idx);
4586 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4587
4588 entry = huge_ptep_get(ptep);
4589 if (huge_pte_none(entry)) {
4590 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4591 goto out_mutex;
4592 }
4593
4594 ret = 0;
4595
4596 /*
4597 * entry could be a migration/hwpoison entry at this point, so this
4598 * check prevents the kernel from going below assuming that we have
4599 * an active hugepage in pagecache. This goto expects the 2nd page
4600 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4601 * properly handle it.
4602 */
4603 if (!pte_present(entry))
4604 goto out_mutex;
4605
4606 /*
4607 * If we are going to COW the mapping later, we examine the pending
4608 * reservations for this page now. This will ensure that any
4609 * allocations necessary to record that reservation occur outside the
4610 * spinlock. For private mappings, we also lookup the pagecache
4611 * page now as it is used to determine if a reservation has been
4612 * consumed.
4613 */
4614 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4615 if (vma_needs_reservation(h, vma, haddr) < 0) {
4616 ret = VM_FAULT_OOM;
4617 goto out_mutex;
4618 }
4619 /* Just decrements count, does not deallocate */
4620 vma_end_reservation(h, vma, haddr);
4621
4622 if (!(vma->vm_flags & VM_MAYSHARE))
4623 pagecache_page = hugetlbfs_pagecache_page(h,
4624 vma, haddr);
4625 }
4626
4627 ptl = huge_pte_lock(h, mm, ptep);
4628
4629 /* Check for a racing update before calling hugetlb_cow */
4630 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4631 goto out_ptl;
4632
4633 /*
4634 * hugetlb_cow() requires page locks of pte_page(entry) and
4635 * pagecache_page, so here we need take the former one
4636 * when page != pagecache_page or !pagecache_page.
4637 */
4638 page = pte_page(entry);
4639 if (page != pagecache_page)
4640 if (!trylock_page(page)) {
4641 need_wait_lock = 1;
4642 goto out_ptl;
4643 }
4644
4645 get_page(page);
4646
4647 if (flags & FAULT_FLAG_WRITE) {
4648 if (!huge_pte_write(entry)) {
4649 ret = hugetlb_cow(mm, vma, address, ptep,
4650 pagecache_page, ptl);
4651 goto out_put_page;
4652 }
4653 entry = huge_pte_mkdirty(entry);
4654 }
4655 entry = pte_mkyoung(entry);
4656 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4657 flags & FAULT_FLAG_WRITE))
4658 update_mmu_cache(vma, haddr, ptep);
4659out_put_page:
4660 if (page != pagecache_page)
4661 unlock_page(page);
4662 put_page(page);
4663out_ptl:
4664 spin_unlock(ptl);
4665
4666 if (pagecache_page) {
4667 unlock_page(pagecache_page);
4668 put_page(pagecache_page);
4669 }
4670out_mutex:
4671 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4672 i_mmap_unlock_read(mapping);
4673 /*
4674 * Generally it's safe to hold refcount during waiting page lock. But
4675 * here we just wait to defer the next page fault to avoid busy loop and
4676 * the page is not used after unlocked before returning from the current
4677 * page fault. So we are safe from accessing freed page, even if we wait
4678 * here without taking refcount.
4679 */
4680 if (need_wait_lock)
4681 wait_on_page_locked(page);
4682 return ret;
4683}
4684
4685/*
4686 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4687 * modifications for huge pages.
4688 */
4689int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4690 pte_t *dst_pte,
4691 struct vm_area_struct *dst_vma,
4692 unsigned long dst_addr,
4693 unsigned long src_addr,
4694 struct page **pagep)
4695{
4696 struct address_space *mapping;
4697 pgoff_t idx;
4698 unsigned long size;
4699 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4700 struct hstate *h = hstate_vma(dst_vma);
4701 pte_t _dst_pte;
4702 spinlock_t *ptl;
4703 int ret;
4704 struct page *page;
4705
4706 if (!*pagep) {
4707 ret = -ENOMEM;
4708 page = alloc_huge_page(dst_vma, dst_addr, 0);
4709 if (IS_ERR(page))
4710 goto out;
4711
4712 ret = copy_huge_page_from_user(page,
4713 (const void __user *) src_addr,
4714 pages_per_huge_page(h), false);
4715
4716 /* fallback to copy_from_user outside mmap_lock */
4717 if (unlikely(ret)) {
4718 ret = -ENOENT;
4719 *pagep = page;
4720 /* don't free the page */
4721 goto out;
4722 }
4723 } else {
4724 page = *pagep;
4725 *pagep = NULL;
4726 }
4727
4728 /*
4729 * The memory barrier inside __SetPageUptodate makes sure that
4730 * preceding stores to the page contents become visible before
4731 * the set_pte_at() write.
4732 */
4733 __SetPageUptodate(page);
4734
4735 mapping = dst_vma->vm_file->f_mapping;
4736 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4737
4738 /*
4739 * If shared, add to page cache
4740 */
4741 if (vm_shared) {
4742 size = i_size_read(mapping->host) >> huge_page_shift(h);
4743 ret = -EFAULT;
4744 if (idx >= size)
4745 goto out_release_nounlock;
4746
4747 /*
4748 * Serialization between remove_inode_hugepages() and
4749 * huge_add_to_page_cache() below happens through the
4750 * hugetlb_fault_mutex_table that here must be hold by
4751 * the caller.
4752 */
4753 ret = huge_add_to_page_cache(page, mapping, idx);
4754 if (ret)
4755 goto out_release_nounlock;
4756 }
4757
4758 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4759 spin_lock(ptl);
4760
4761 /*
4762 * Recheck the i_size after holding PT lock to make sure not
4763 * to leave any page mapped (as page_mapped()) beyond the end
4764 * of the i_size (remove_inode_hugepages() is strict about
4765 * enforcing that). If we bail out here, we'll also leave a
4766 * page in the radix tree in the vm_shared case beyond the end
4767 * of the i_size, but remove_inode_hugepages() will take care
4768 * of it as soon as we drop the hugetlb_fault_mutex_table.
4769 */
4770 size = i_size_read(mapping->host) >> huge_page_shift(h);
4771 ret = -EFAULT;
4772 if (idx >= size)
4773 goto out_release_unlock;
4774
4775 ret = -EEXIST;
4776 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4777 goto out_release_unlock;
4778
4779 if (vm_shared) {
4780 page_dup_rmap(page, true);
4781 } else {
4782 ClearPagePrivate(page);
4783 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4784 }
4785
4786 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4787 if (dst_vma->vm_flags & VM_WRITE)
4788 _dst_pte = huge_pte_mkdirty(_dst_pte);
4789 _dst_pte = pte_mkyoung(_dst_pte);
4790
4791 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4792
4793 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4794 dst_vma->vm_flags & VM_WRITE);
4795 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4796
4797 /* No need to invalidate - it was non-present before */
4798 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4799
4800 spin_unlock(ptl);
4801 set_page_huge_active(page);
4802 if (vm_shared)
4803 unlock_page(page);
4804 ret = 0;
4805out:
4806 return ret;
4807out_release_unlock:
4808 spin_unlock(ptl);
4809 if (vm_shared)
4810 unlock_page(page);
4811out_release_nounlock:
4812 put_page(page);
4813 goto out;
4814}
4815
4816long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4817 struct page **pages, struct vm_area_struct **vmas,
4818 unsigned long *position, unsigned long *nr_pages,
4819 long i, unsigned int flags, int *locked)
4820{
4821 unsigned long pfn_offset;
4822 unsigned long vaddr = *position;
4823 unsigned long remainder = *nr_pages;
4824 struct hstate *h = hstate_vma(vma);
4825 int err = -EFAULT;
4826
4827 while (vaddr < vma->vm_end && remainder) {
4828 pte_t *pte;
4829 spinlock_t *ptl = NULL;
4830 int absent;
4831 struct page *page;
4832
4833 /*
4834 * If we have a pending SIGKILL, don't keep faulting pages and
4835 * potentially allocating memory.
4836 */
4837 if (fatal_signal_pending(current)) {
4838 remainder = 0;
4839 break;
4840 }
4841
4842 /*
4843 * Some archs (sparc64, sh*) have multiple pte_ts to
4844 * each hugepage. We have to make sure we get the
4845 * first, for the page indexing below to work.
4846 *
4847 * Note that page table lock is not held when pte is null.
4848 */
4849 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4850 huge_page_size(h));
4851 if (pte)
4852 ptl = huge_pte_lock(h, mm, pte);
4853 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4854
4855 /*
4856 * When coredumping, it suits get_dump_page if we just return
4857 * an error where there's an empty slot with no huge pagecache
4858 * to back it. This way, we avoid allocating a hugepage, and
4859 * the sparse dumpfile avoids allocating disk blocks, but its
4860 * huge holes still show up with zeroes where they need to be.
4861 */
4862 if (absent && (flags & FOLL_DUMP) &&
4863 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4864 if (pte)
4865 spin_unlock(ptl);
4866 remainder = 0;
4867 break;
4868 }
4869
4870 /*
4871 * We need call hugetlb_fault for both hugepages under migration
4872 * (in which case hugetlb_fault waits for the migration,) and
4873 * hwpoisoned hugepages (in which case we need to prevent the
4874 * caller from accessing to them.) In order to do this, we use
4875 * here is_swap_pte instead of is_hugetlb_entry_migration and
4876 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4877 * both cases, and because we can't follow correct pages
4878 * directly from any kind of swap entries.
4879 */
4880 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4881 ((flags & FOLL_WRITE) &&
4882 !huge_pte_write(huge_ptep_get(pte)))) {
4883 vm_fault_t ret;
4884 unsigned int fault_flags = 0;
4885
4886 if (pte)
4887 spin_unlock(ptl);
4888 if (flags & FOLL_WRITE)
4889 fault_flags |= FAULT_FLAG_WRITE;
4890 if (locked)
4891 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4892 FAULT_FLAG_KILLABLE;
4893 if (flags & FOLL_NOWAIT)
4894 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4895 FAULT_FLAG_RETRY_NOWAIT;
4896 if (flags & FOLL_TRIED) {
4897 /*
4898 * Note: FAULT_FLAG_ALLOW_RETRY and
4899 * FAULT_FLAG_TRIED can co-exist
4900 */
4901 fault_flags |= FAULT_FLAG_TRIED;
4902 }
4903 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4904 if (ret & VM_FAULT_ERROR) {
4905 err = vm_fault_to_errno(ret, flags);
4906 remainder = 0;
4907 break;
4908 }
4909 if (ret & VM_FAULT_RETRY) {
4910 if (locked &&
4911 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4912 *locked = 0;
4913 *nr_pages = 0;
4914 /*
4915 * VM_FAULT_RETRY must not return an
4916 * error, it will return zero
4917 * instead.
4918 *
4919 * No need to update "position" as the
4920 * caller will not check it after
4921 * *nr_pages is set to 0.
4922 */
4923 return i;
4924 }
4925 continue;
4926 }
4927
4928 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4929 page = pte_page(huge_ptep_get(pte));
4930
4931 /*
4932 * If subpage information not requested, update counters
4933 * and skip the same_page loop below.
4934 */
4935 if (!pages && !vmas && !pfn_offset &&
4936 (vaddr + huge_page_size(h) < vma->vm_end) &&
4937 (remainder >= pages_per_huge_page(h))) {
4938 vaddr += huge_page_size(h);
4939 remainder -= pages_per_huge_page(h);
4940 i += pages_per_huge_page(h);
4941 spin_unlock(ptl);
4942 continue;
4943 }
4944
4945same_page:
4946 if (pages) {
4947 pages[i] = mem_map_offset(page, pfn_offset);
4948 /*
4949 * try_grab_page() should always succeed here, because:
4950 * a) we hold the ptl lock, and b) we've just checked
4951 * that the huge page is present in the page tables. If
4952 * the huge page is present, then the tail pages must
4953 * also be present. The ptl prevents the head page and
4954 * tail pages from being rearranged in any way. So this
4955 * page must be available at this point, unless the page
4956 * refcount overflowed:
4957 */
4958 if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) {
4959 spin_unlock(ptl);
4960 remainder = 0;
4961 err = -ENOMEM;
4962 break;
4963 }
4964 }
4965
4966 if (vmas)
4967 vmas[i] = vma;
4968
4969 vaddr += PAGE_SIZE;
4970 ++pfn_offset;
4971 --remainder;
4972 ++i;
4973 if (vaddr < vma->vm_end && remainder &&
4974 pfn_offset < pages_per_huge_page(h)) {
4975 /*
4976 * We use pfn_offset to avoid touching the pageframes
4977 * of this compound page.
4978 */
4979 goto same_page;
4980 }
4981 spin_unlock(ptl);
4982 }
4983 *nr_pages = remainder;
4984 /*
4985 * setting position is actually required only if remainder is
4986 * not zero but it's faster not to add a "if (remainder)"
4987 * branch.
4988 */
4989 *position = vaddr;
4990
4991 return i ? i : err;
4992}
4993
4994#ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4995/*
4996 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4997 * implement this.
4998 */
4999#define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
5000#endif
5001
5002unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
5003 unsigned long address, unsigned long end, pgprot_t newprot)
5004{
5005 struct mm_struct *mm = vma->vm_mm;
5006 unsigned long start = address;
5007 pte_t *ptep;
5008 pte_t pte;
5009 struct hstate *h = hstate_vma(vma);
5010 unsigned long pages = 0;
5011 bool shared_pmd = false;
5012 struct mmu_notifier_range range;
5013
5014 /*
5015 * In the case of shared PMDs, the area to flush could be beyond
5016 * start/end. Set range.start/range.end to cover the maximum possible
5017 * range if PMD sharing is possible.
5018 */
5019 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5020 0, vma, mm, start, end);
5021 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5022
5023 BUG_ON(address >= end);
5024 flush_cache_range(vma, range.start, range.end);
5025
5026 mmu_notifier_invalidate_range_start(&range);
5027 i_mmap_lock_write(vma->vm_file->f_mapping);
5028 for (; address < end; address += huge_page_size(h)) {
5029 spinlock_t *ptl;
5030 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5031 if (!ptep)
5032 continue;
5033 ptl = huge_pte_lock(h, mm, ptep);
5034 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
5035 pages++;
5036 spin_unlock(ptl);
5037 shared_pmd = true;
5038 continue;
5039 }
5040 pte = huge_ptep_get(ptep);
5041 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5042 spin_unlock(ptl);
5043 continue;
5044 }
5045 if (unlikely(is_hugetlb_entry_migration(pte))) {
5046 swp_entry_t entry = pte_to_swp_entry(pte);
5047
5048 if (is_write_migration_entry(entry)) {
5049 pte_t newpte;
5050
5051 make_migration_entry_read(&entry);
5052 newpte = swp_entry_to_pte(entry);
5053 set_huge_swap_pte_at(mm, address, ptep,
5054 newpte, huge_page_size(h));
5055 pages++;
5056 }
5057 spin_unlock(ptl);
5058 continue;
5059 }
5060 if (!huge_pte_none(pte)) {
5061 pte_t old_pte;
5062
5063 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5064 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
5065 pte = arch_make_huge_pte(pte, vma, NULL, 0);
5066 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
5067 pages++;
5068 }
5069 spin_unlock(ptl);
5070 }
5071 /*
5072 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5073 * may have cleared our pud entry and done put_page on the page table:
5074 * once we release i_mmap_rwsem, another task can do the final put_page
5075 * and that page table be reused and filled with junk. If we actually
5076 * did unshare a page of pmds, flush the range corresponding to the pud.
5077 */
5078 if (shared_pmd)
5079 flush_hugetlb_tlb_range(vma, range.start, range.end);
5080 else
5081 flush_hugetlb_tlb_range(vma, start, end);
5082 /*
5083 * No need to call mmu_notifier_invalidate_range() we are downgrading
5084 * page table protection not changing it to point to a new page.
5085 *
5086 * See Documentation/vm/mmu_notifier.rst
5087 */
5088 i_mmap_unlock_write(vma->vm_file->f_mapping);
5089 mmu_notifier_invalidate_range_end(&range);
5090
5091 return pages << h->order;
5092}
5093
5094int hugetlb_reserve_pages(struct inode *inode,
5095 long from, long to,
5096 struct vm_area_struct *vma,
5097 vm_flags_t vm_flags)
5098{
5099 long ret, chg, add = -1;
5100 struct hstate *h = hstate_inode(inode);
5101 struct hugepage_subpool *spool = subpool_inode(inode);
5102 struct resv_map *resv_map;
5103 struct hugetlb_cgroup *h_cg = NULL;
5104 long gbl_reserve, regions_needed = 0;
5105
5106 /* This should never happen */
5107 if (from > to) {
5108 VM_WARN(1, "%s called with a negative range\n", __func__);
5109 return -EINVAL;
5110 }
5111
5112 /*
5113 * Only apply hugepage reservation if asked. At fault time, an
5114 * attempt will be made for VM_NORESERVE to allocate a page
5115 * without using reserves
5116 */
5117 if (vm_flags & VM_NORESERVE)
5118 return 0;
5119
5120 /*
5121 * Shared mappings base their reservation on the number of pages that
5122 * are already allocated on behalf of the file. Private mappings need
5123 * to reserve the full area even if read-only as mprotect() may be
5124 * called to make the mapping read-write. Assume !vma is a shm mapping
5125 */
5126 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5127 /*
5128 * resv_map can not be NULL as hugetlb_reserve_pages is only
5129 * called for inodes for which resv_maps were created (see
5130 * hugetlbfs_get_inode).
5131 */
5132 resv_map = inode_resv_map(inode);
5133
5134 chg = region_chg(resv_map, from, to, ®ions_needed);
5135
5136 } else {
5137 /* Private mapping. */
5138 resv_map = resv_map_alloc();
5139 if (!resv_map)
5140 return -ENOMEM;
5141
5142 chg = to - from;
5143
5144 set_vma_resv_map(vma, resv_map);
5145 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5146 }
5147
5148 if (chg < 0) {
5149 ret = chg;
5150 goto out_err;
5151 }
5152
5153 ret = hugetlb_cgroup_charge_cgroup_rsvd(
5154 hstate_index(h), chg * pages_per_huge_page(h), &h_cg);
5155
5156 if (ret < 0) {
5157 ret = -ENOMEM;
5158 goto out_err;
5159 }
5160
5161 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5162 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5163 * of the resv_map.
5164 */
5165 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5166 }
5167
5168 /*
5169 * There must be enough pages in the subpool for the mapping. If
5170 * the subpool has a minimum size, there may be some global
5171 * reservations already in place (gbl_reserve).
5172 */
5173 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5174 if (gbl_reserve < 0) {
5175 ret = -ENOSPC;
5176 goto out_uncharge_cgroup;
5177 }
5178
5179 /*
5180 * Check enough hugepages are available for the reservation.
5181 * Hand the pages back to the subpool if there are not
5182 */
5183 ret = hugetlb_acct_memory(h, gbl_reserve);
5184 if (ret < 0) {
5185 goto out_put_pages;
5186 }
5187
5188 /*
5189 * Account for the reservations made. Shared mappings record regions
5190 * that have reservations as they are shared by multiple VMAs.
5191 * When the last VMA disappears, the region map says how much
5192 * the reservation was and the page cache tells how much of
5193 * the reservation was consumed. Private mappings are per-VMA and
5194 * only the consumed reservations are tracked. When the VMA
5195 * disappears, the original reservation is the VMA size and the
5196 * consumed reservations are stored in the map. Hence, nothing
5197 * else has to be done for private mappings here
5198 */
5199 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5200 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5201
5202 if (unlikely(add < 0)) {
5203 hugetlb_acct_memory(h, -gbl_reserve);
5204 goto out_put_pages;
5205 } else if (unlikely(chg > add)) {
5206 /*
5207 * pages in this range were added to the reserve
5208 * map between region_chg and region_add. This
5209 * indicates a race with alloc_huge_page. Adjust
5210 * the subpool and reserve counts modified above
5211 * based on the difference.
5212 */
5213 long rsv_adjust;
5214
5215 hugetlb_cgroup_uncharge_cgroup_rsvd(
5216 hstate_index(h),
5217 (chg - add) * pages_per_huge_page(h), h_cg);
5218
5219 rsv_adjust = hugepage_subpool_put_pages(spool,
5220 chg - add);
5221 hugetlb_acct_memory(h, -rsv_adjust);
5222 }
5223 }
5224 return 0;
5225out_put_pages:
5226 /* put back original number of pages, chg */
5227 (void)hugepage_subpool_put_pages(spool, chg);
5228out_uncharge_cgroup:
5229 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5230 chg * pages_per_huge_page(h), h_cg);
5231out_err:
5232 if (!vma || vma->vm_flags & VM_MAYSHARE)
5233 /* Only call region_abort if the region_chg succeeded but the
5234 * region_add failed or didn't run.
5235 */
5236 if (chg >= 0 && add < 0)
5237 region_abort(resv_map, from, to, regions_needed);
5238 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5239 kref_put(&resv_map->refs, resv_map_release);
5240 return ret;
5241}
5242
5243long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5244 long freed)
5245{
5246 struct hstate *h = hstate_inode(inode);
5247 struct resv_map *resv_map = inode_resv_map(inode);
5248 long chg = 0;
5249 struct hugepage_subpool *spool = subpool_inode(inode);
5250 long gbl_reserve;
5251
5252 /*
5253 * Since this routine can be called in the evict inode path for all
5254 * hugetlbfs inodes, resv_map could be NULL.
5255 */
5256 if (resv_map) {
5257 chg = region_del(resv_map, start, end);
5258 /*
5259 * region_del() can fail in the rare case where a region
5260 * must be split and another region descriptor can not be
5261 * allocated. If end == LONG_MAX, it will not fail.
5262 */
5263 if (chg < 0)
5264 return chg;
5265 }
5266
5267 spin_lock(&inode->i_lock);
5268 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5269 spin_unlock(&inode->i_lock);
5270
5271 /*
5272 * If the subpool has a minimum size, the number of global
5273 * reservations to be released may be adjusted.
5274 */
5275 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5276 hugetlb_acct_memory(h, -gbl_reserve);
5277
5278 return 0;
5279}
5280
5281#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5282static unsigned long page_table_shareable(struct vm_area_struct *svma,
5283 struct vm_area_struct *vma,
5284 unsigned long addr, pgoff_t idx)
5285{
5286 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5287 svma->vm_start;
5288 unsigned long sbase = saddr & PUD_MASK;
5289 unsigned long s_end = sbase + PUD_SIZE;
5290
5291 /* Allow segments to share if only one is marked locked */
5292 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5293 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5294
5295 /*
5296 * match the virtual addresses, permission and the alignment of the
5297 * page table page.
5298 */
5299 if (pmd_index(addr) != pmd_index(saddr) ||
5300 vm_flags != svm_flags ||
5301 sbase < svma->vm_start || svma->vm_end < s_end)
5302 return 0;
5303
5304 return saddr;
5305}
5306
5307static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5308{
5309 unsigned long base = addr & PUD_MASK;
5310 unsigned long end = base + PUD_SIZE;
5311
5312 /*
5313 * check on proper vm_flags and page table alignment
5314 */
5315 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5316 return true;
5317 return false;
5318}
5319
5320/*
5321 * Determine if start,end range within vma could be mapped by shared pmd.
5322 * If yes, adjust start and end to cover range associated with possible
5323 * shared pmd mappings.
5324 */
5325void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5326 unsigned long *start, unsigned long *end)
5327{
5328 unsigned long a_start, a_end;
5329
5330 if (!(vma->vm_flags & VM_MAYSHARE))
5331 return;
5332
5333 /* Extend the range to be PUD aligned for a worst case scenario */
5334 a_start = ALIGN_DOWN(*start, PUD_SIZE);
5335 a_end = ALIGN(*end, PUD_SIZE);
5336
5337 /*
5338 * Intersect the range with the vma range, since pmd sharing won't be
5339 * across vma after all
5340 */
5341 *start = max(vma->vm_start, a_start);
5342 *end = min(vma->vm_end, a_end);
5343}
5344
5345/*
5346 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5347 * and returns the corresponding pte. While this is not necessary for the
5348 * !shared pmd case because we can allocate the pmd later as well, it makes the
5349 * code much cleaner.
5350 *
5351 * This routine must be called with i_mmap_rwsem held in at least read mode.
5352 * For hugetlbfs, this prevents removal of any page table entries associated
5353 * with the address space. This is important as we are setting up sharing
5354 * based on existing page table entries (mappings).
5355 */
5356pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5357{
5358 struct vm_area_struct *vma = find_vma(mm, addr);
5359 struct address_space *mapping = vma->vm_file->f_mapping;
5360 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5361 vma->vm_pgoff;
5362 struct vm_area_struct *svma;
5363 unsigned long saddr;
5364 pte_t *spte = NULL;
5365 pte_t *pte;
5366 spinlock_t *ptl;
5367
5368 if (!vma_shareable(vma, addr))
5369 return (pte_t *)pmd_alloc(mm, pud, addr);
5370
5371 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5372 if (svma == vma)
5373 continue;
5374
5375 saddr = page_table_shareable(svma, vma, addr, idx);
5376 if (saddr) {
5377 spte = huge_pte_offset(svma->vm_mm, saddr,
5378 vma_mmu_pagesize(svma));
5379 if (spte) {
5380 get_page(virt_to_page(spte));
5381 break;
5382 }
5383 }
5384 }
5385
5386 if (!spte)
5387 goto out;
5388
5389 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
5390 if (pud_none(*pud)) {
5391 pud_populate(mm, pud,
5392 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5393 mm_inc_nr_pmds(mm);
5394 } else {
5395 put_page(virt_to_page(spte));
5396 }
5397 spin_unlock(ptl);
5398out:
5399 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5400 return pte;
5401}
5402
5403/*
5404 * unmap huge page backed by shared pte.
5405 *
5406 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5407 * indicated by page_count > 1, unmap is achieved by clearing pud and
5408 * decrementing the ref count. If count == 1, the pte page is not shared.
5409 *
5410 * Called with page table lock held and i_mmap_rwsem held in write mode.
5411 *
5412 * returns: 1 successfully unmapped a shared pte page
5413 * 0 the underlying pte page is not shared, or it is the last user
5414 */
5415int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5416 unsigned long *addr, pte_t *ptep)
5417{
5418 pgd_t *pgd = pgd_offset(mm, *addr);
5419 p4d_t *p4d = p4d_offset(pgd, *addr);
5420 pud_t *pud = pud_offset(p4d, *addr);
5421
5422 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
5423 BUG_ON(page_count(virt_to_page(ptep)) == 0);
5424 if (page_count(virt_to_page(ptep)) == 1)
5425 return 0;
5426
5427 pud_clear(pud);
5428 put_page(virt_to_page(ptep));
5429 mm_dec_nr_pmds(mm);
5430 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5431 return 1;
5432}
5433#define want_pmd_share() (1)
5434#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5435pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5436{
5437 return NULL;
5438}
5439
5440int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5441 unsigned long *addr, pte_t *ptep)
5442{
5443 return 0;
5444}
5445
5446void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5447 unsigned long *start, unsigned long *end)
5448{
5449}
5450#define want_pmd_share() (0)
5451#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5452
5453#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5454pte_t *huge_pte_alloc(struct mm_struct *mm,
5455 unsigned long addr, unsigned long sz)
5456{
5457 pgd_t *pgd;
5458 p4d_t *p4d;
5459 pud_t *pud;
5460 pte_t *pte = NULL;
5461
5462 pgd = pgd_offset(mm, addr);
5463 p4d = p4d_alloc(mm, pgd, addr);
5464 if (!p4d)
5465 return NULL;
5466 pud = pud_alloc(mm, p4d, addr);
5467 if (pud) {
5468 if (sz == PUD_SIZE) {
5469 pte = (pte_t *)pud;
5470 } else {
5471 BUG_ON(sz != PMD_SIZE);
5472 if (want_pmd_share() && pud_none(*pud))
5473 pte = huge_pmd_share(mm, addr, pud);
5474 else
5475 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5476 }
5477 }
5478 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
5479
5480 return pte;
5481}
5482
5483/*
5484 * huge_pte_offset() - Walk the page table to resolve the hugepage
5485 * entry at address @addr
5486 *
5487 * Return: Pointer to page table entry (PUD or PMD) for
5488 * address @addr, or NULL if a !p*d_present() entry is encountered and the
5489 * size @sz doesn't match the hugepage size at this level of the page
5490 * table.
5491 */
5492pte_t *huge_pte_offset(struct mm_struct *mm,
5493 unsigned long addr, unsigned long sz)
5494{
5495 pgd_t *pgd;
5496 p4d_t *p4d;
5497 pud_t *pud;
5498 pmd_t *pmd;
5499
5500 pgd = pgd_offset(mm, addr);
5501 if (!pgd_present(*pgd))
5502 return NULL;
5503 p4d = p4d_offset(pgd, addr);
5504 if (!p4d_present(*p4d))
5505 return NULL;
5506
5507 pud = pud_offset(p4d, addr);
5508 if (sz == PUD_SIZE)
5509 /* must be pud huge, non-present or none */
5510 return (pte_t *)pud;
5511 if (!pud_present(*pud))
5512 return NULL;
5513 /* must have a valid entry and size to go further */
5514
5515 pmd = pmd_offset(pud, addr);
5516 /* must be pmd huge, non-present or none */
5517 return (pte_t *)pmd;
5518}
5519
5520#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5521
5522/*
5523 * These functions are overwritable if your architecture needs its own
5524 * behavior.
5525 */
5526struct page * __weak
5527follow_huge_addr(struct mm_struct *mm, unsigned long address,
5528 int write)
5529{
5530 return ERR_PTR(-EINVAL);
5531}
5532
5533struct page * __weak
5534follow_huge_pd(struct vm_area_struct *vma,
5535 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5536{
5537 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5538 return NULL;
5539}
5540
5541struct page * __weak
5542follow_huge_pmd(struct mm_struct *mm, unsigned long address,
5543 pmd_t *pmd, int flags)
5544{
5545 struct page *page = NULL;
5546 spinlock_t *ptl;
5547 pte_t pte;
5548
5549 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5550 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
5551 (FOLL_PIN | FOLL_GET)))
5552 return NULL;
5553
5554retry:
5555 ptl = pmd_lockptr(mm, pmd);
5556 spin_lock(ptl);
5557 /*
5558 * make sure that the address range covered by this pmd is not
5559 * unmapped from other threads.
5560 */
5561 if (!pmd_huge(*pmd))
5562 goto out;
5563 pte = huge_ptep_get((pte_t *)pmd);
5564 if (pte_present(pte)) {
5565 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
5566 /*
5567 * try_grab_page() should always succeed here, because: a) we
5568 * hold the pmd (ptl) lock, and b) we've just checked that the
5569 * huge pmd (head) page is present in the page tables. The ptl
5570 * prevents the head page and tail pages from being rearranged
5571 * in any way. So this page must be available at this point,
5572 * unless the page refcount overflowed:
5573 */
5574 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5575 page = NULL;
5576 goto out;
5577 }
5578 } else {
5579 if (is_hugetlb_entry_migration(pte)) {
5580 spin_unlock(ptl);
5581 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5582 goto retry;
5583 }
5584 /*
5585 * hwpoisoned entry is treated as no_page_table in
5586 * follow_page_mask().
5587 */
5588 }
5589out:
5590 spin_unlock(ptl);
5591 return page;
5592}
5593
5594struct page * __weak
5595follow_huge_pud(struct mm_struct *mm, unsigned long address,
5596 pud_t *pud, int flags)
5597{
5598 if (flags & (FOLL_GET | FOLL_PIN))
5599 return NULL;
5600
5601 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5602}
5603
5604struct page * __weak
5605follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5606{
5607 if (flags & (FOLL_GET | FOLL_PIN))
5608 return NULL;
5609
5610 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5611}
5612
5613bool isolate_huge_page(struct page *page, struct list_head *list)
5614{
5615 bool ret = true;
5616
5617 VM_BUG_ON_PAGE(!PageHead(page), page);
5618 spin_lock(&hugetlb_lock);
5619 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
5620 ret = false;
5621 goto unlock;
5622 }
5623 clear_page_huge_active(page);
5624 list_move_tail(&page->lru, list);
5625unlock:
5626 spin_unlock(&hugetlb_lock);
5627 return ret;
5628}
5629
5630void putback_active_hugepage(struct page *page)
5631{
5632 VM_BUG_ON_PAGE(!PageHead(page), page);
5633 spin_lock(&hugetlb_lock);
5634 set_page_huge_active(page);
5635 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5636 spin_unlock(&hugetlb_lock);
5637 put_page(page);
5638}
5639
5640void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5641{
5642 struct hstate *h = page_hstate(oldpage);
5643
5644 hugetlb_cgroup_migrate(oldpage, newpage);
5645 set_page_owner_migrate_reason(newpage, reason);
5646
5647 /*
5648 * transfer temporary state of the new huge page. This is
5649 * reverse to other transitions because the newpage is going to
5650 * be final while the old one will be freed so it takes over
5651 * the temporary status.
5652 *
5653 * Also note that we have to transfer the per-node surplus state
5654 * here as well otherwise the global surplus count will not match
5655 * the per-node's.
5656 */
5657 if (PageHugeTemporary(newpage)) {
5658 int old_nid = page_to_nid(oldpage);
5659 int new_nid = page_to_nid(newpage);
5660
5661 SetPageHugeTemporary(oldpage);
5662 ClearPageHugeTemporary(newpage);
5663
5664 spin_lock(&hugetlb_lock);
5665 if (h->surplus_huge_pages_node[old_nid]) {
5666 h->surplus_huge_pages_node[old_nid]--;
5667 h->surplus_huge_pages_node[new_nid]++;
5668 }
5669 spin_unlock(&hugetlb_lock);
5670 }
5671}
5672
5673#ifdef CONFIG_CMA
5674static bool cma_reserve_called __initdata;
5675
5676static int __init cmdline_parse_hugetlb_cma(char *p)
5677{
5678 hugetlb_cma_size = memparse(p, &p);
5679 return 0;
5680}
5681
5682early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
5683
5684void __init hugetlb_cma_reserve(int order)
5685{
5686 unsigned long size, reserved, per_node;
5687 int nid;
5688
5689 cma_reserve_called = true;
5690
5691 if (!hugetlb_cma_size)
5692 return;
5693
5694 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
5695 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
5696 (PAGE_SIZE << order) / SZ_1M);
5697 return;
5698 }
5699
5700 /*
5701 * If 3 GB area is requested on a machine with 4 numa nodes,
5702 * let's allocate 1 GB on first three nodes and ignore the last one.
5703 */
5704 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
5705 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
5706 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
5707
5708 reserved = 0;
5709 for_each_node_state(nid, N_ONLINE) {
5710 int res;
5711 char name[20];
5712
5713 size = min(per_node, hugetlb_cma_size - reserved);
5714 size = round_up(size, PAGE_SIZE << order);
5715
5716 snprintf(name, 20, "hugetlb%d", nid);
5717 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
5718 0, false, name,
5719 &hugetlb_cma[nid], nid);
5720 if (res) {
5721 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
5722 res, nid);
5723 continue;
5724 }
5725
5726 reserved += size;
5727 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
5728 size / SZ_1M, nid);
5729
5730 if (reserved >= hugetlb_cma_size)
5731 break;
5732 }
5733}
5734
5735void __init hugetlb_cma_check(void)
5736{
5737 if (!hugetlb_cma_size || cma_reserve_called)
5738 return;
5739
5740 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
5741}
5742
5743#endif /* CONFIG_CMA */
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/mmdebug.h>
23#include <linux/sched/signal.h>
24#include <linux/rmap.h>
25#include <linux/string_helpers.h>
26#include <linux/swap.h>
27#include <linux/swapops.h>
28#include <linux/jhash.h>
29#include <linux/numa.h>
30
31#include <asm/page.h>
32#include <asm/pgtable.h>
33#include <asm/tlb.h>
34
35#include <linux/io.h>
36#include <linux/hugetlb.h>
37#include <linux/hugetlb_cgroup.h>
38#include <linux/node.h>
39#include <linux/userfaultfd_k.h>
40#include <linux/page_owner.h>
41#include "internal.h"
42
43int hugetlb_max_hstate __read_mostly;
44unsigned int default_hstate_idx;
45struct hstate hstates[HUGE_MAX_HSTATE];
46/*
47 * Minimum page order among possible hugepage sizes, set to a proper value
48 * at boot time.
49 */
50static unsigned int minimum_order __read_mostly = UINT_MAX;
51
52__initdata LIST_HEAD(huge_boot_pages);
53
54/* for command line parsing */
55static struct hstate * __initdata parsed_hstate;
56static unsigned long __initdata default_hstate_max_huge_pages;
57static unsigned long __initdata default_hstate_size;
58static bool __initdata parsed_valid_hugepagesz = true;
59
60/*
61 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
62 * free_huge_pages, and surplus_huge_pages.
63 */
64DEFINE_SPINLOCK(hugetlb_lock);
65
66/*
67 * Serializes faults on the same logical page. This is used to
68 * prevent spurious OOMs when the hugepage pool is fully utilized.
69 */
70static int num_fault_mutexes;
71struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
72
73/* Forward declaration */
74static int hugetlb_acct_memory(struct hstate *h, long delta);
75
76static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
77{
78 bool free = (spool->count == 0) && (spool->used_hpages == 0);
79
80 spin_unlock(&spool->lock);
81
82 /* If no pages are used, and no other handles to the subpool
83 * remain, give up any reservations mased on minimum size and
84 * free the subpool */
85 if (free) {
86 if (spool->min_hpages != -1)
87 hugetlb_acct_memory(spool->hstate,
88 -spool->min_hpages);
89 kfree(spool);
90 }
91}
92
93struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
94 long min_hpages)
95{
96 struct hugepage_subpool *spool;
97
98 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
99 if (!spool)
100 return NULL;
101
102 spin_lock_init(&spool->lock);
103 spool->count = 1;
104 spool->max_hpages = max_hpages;
105 spool->hstate = h;
106 spool->min_hpages = min_hpages;
107
108 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
109 kfree(spool);
110 return NULL;
111 }
112 spool->rsv_hpages = min_hpages;
113
114 return spool;
115}
116
117void hugepage_put_subpool(struct hugepage_subpool *spool)
118{
119 spin_lock(&spool->lock);
120 BUG_ON(!spool->count);
121 spool->count--;
122 unlock_or_release_subpool(spool);
123}
124
125/*
126 * Subpool accounting for allocating and reserving pages.
127 * Return -ENOMEM if there are not enough resources to satisfy the
128 * the request. Otherwise, return the number of pages by which the
129 * global pools must be adjusted (upward). The returned value may
130 * only be different than the passed value (delta) in the case where
131 * a subpool minimum size must be manitained.
132 */
133static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
134 long delta)
135{
136 long ret = delta;
137
138 if (!spool)
139 return ret;
140
141 spin_lock(&spool->lock);
142
143 if (spool->max_hpages != -1) { /* maximum size accounting */
144 if ((spool->used_hpages + delta) <= spool->max_hpages)
145 spool->used_hpages += delta;
146 else {
147 ret = -ENOMEM;
148 goto unlock_ret;
149 }
150 }
151
152 /* minimum size accounting */
153 if (spool->min_hpages != -1 && spool->rsv_hpages) {
154 if (delta > spool->rsv_hpages) {
155 /*
156 * Asking for more reserves than those already taken on
157 * behalf of subpool. Return difference.
158 */
159 ret = delta - spool->rsv_hpages;
160 spool->rsv_hpages = 0;
161 } else {
162 ret = 0; /* reserves already accounted for */
163 spool->rsv_hpages -= delta;
164 }
165 }
166
167unlock_ret:
168 spin_unlock(&spool->lock);
169 return ret;
170}
171
172/*
173 * Subpool accounting for freeing and unreserving pages.
174 * Return the number of global page reservations that must be dropped.
175 * The return value may only be different than the passed value (delta)
176 * in the case where a subpool minimum size must be maintained.
177 */
178static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
179 long delta)
180{
181 long ret = delta;
182
183 if (!spool)
184 return delta;
185
186 spin_lock(&spool->lock);
187
188 if (spool->max_hpages != -1) /* maximum size accounting */
189 spool->used_hpages -= delta;
190
191 /* minimum size accounting */
192 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
193 if (spool->rsv_hpages + delta <= spool->min_hpages)
194 ret = 0;
195 else
196 ret = spool->rsv_hpages + delta - spool->min_hpages;
197
198 spool->rsv_hpages += delta;
199 if (spool->rsv_hpages > spool->min_hpages)
200 spool->rsv_hpages = spool->min_hpages;
201 }
202
203 /*
204 * If hugetlbfs_put_super couldn't free spool due to an outstanding
205 * quota reference, free it now.
206 */
207 unlock_or_release_subpool(spool);
208
209 return ret;
210}
211
212static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
213{
214 return HUGETLBFS_SB(inode->i_sb)->spool;
215}
216
217static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
218{
219 return subpool_inode(file_inode(vma->vm_file));
220}
221
222/*
223 * Region tracking -- allows tracking of reservations and instantiated pages
224 * across the pages in a mapping.
225 *
226 * The region data structures are embedded into a resv_map and protected
227 * by a resv_map's lock. The set of regions within the resv_map represent
228 * reservations for huge pages, or huge pages that have already been
229 * instantiated within the map. The from and to elements are huge page
230 * indicies into the associated mapping. from indicates the starting index
231 * of the region. to represents the first index past the end of the region.
232 *
233 * For example, a file region structure with from == 0 and to == 4 represents
234 * four huge pages in a mapping. It is important to note that the to element
235 * represents the first element past the end of the region. This is used in
236 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
237 *
238 * Interval notation of the form [from, to) will be used to indicate that
239 * the endpoint from is inclusive and to is exclusive.
240 */
241struct file_region {
242 struct list_head link;
243 long from;
244 long to;
245};
246
247/*
248 * Add the huge page range represented by [f, t) to the reserve
249 * map. In the normal case, existing regions will be expanded
250 * to accommodate the specified range. Sufficient regions should
251 * exist for expansion due to the previous call to region_chg
252 * with the same range. However, it is possible that region_del
253 * could have been called after region_chg and modifed the map
254 * in such a way that no region exists to be expanded. In this
255 * case, pull a region descriptor from the cache associated with
256 * the map and use that for the new range.
257 *
258 * Return the number of new huge pages added to the map. This
259 * number is greater than or equal to zero.
260 */
261static long region_add(struct resv_map *resv, long f, long t)
262{
263 struct list_head *head = &resv->regions;
264 struct file_region *rg, *nrg, *trg;
265 long add = 0;
266
267 spin_lock(&resv->lock);
268 /* Locate the region we are either in or before. */
269 list_for_each_entry(rg, head, link)
270 if (f <= rg->to)
271 break;
272
273 /*
274 * If no region exists which can be expanded to include the
275 * specified range, the list must have been modified by an
276 * interleving call to region_del(). Pull a region descriptor
277 * from the cache and use it for this range.
278 */
279 if (&rg->link == head || t < rg->from) {
280 VM_BUG_ON(resv->region_cache_count <= 0);
281
282 resv->region_cache_count--;
283 nrg = list_first_entry(&resv->region_cache, struct file_region,
284 link);
285 list_del(&nrg->link);
286
287 nrg->from = f;
288 nrg->to = t;
289 list_add(&nrg->link, rg->link.prev);
290
291 add += t - f;
292 goto out_locked;
293 }
294
295 /* Round our left edge to the current segment if it encloses us. */
296 if (f > rg->from)
297 f = rg->from;
298
299 /* Check for and consume any regions we now overlap with. */
300 nrg = rg;
301 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
302 if (&rg->link == head)
303 break;
304 if (rg->from > t)
305 break;
306
307 /* If this area reaches higher then extend our area to
308 * include it completely. If this is not the first area
309 * which we intend to reuse, free it. */
310 if (rg->to > t)
311 t = rg->to;
312 if (rg != nrg) {
313 /* Decrement return value by the deleted range.
314 * Another range will span this area so that by
315 * end of routine add will be >= zero
316 */
317 add -= (rg->to - rg->from);
318 list_del(&rg->link);
319 kfree(rg);
320 }
321 }
322
323 add += (nrg->from - f); /* Added to beginning of region */
324 nrg->from = f;
325 add += t - nrg->to; /* Added to end of region */
326 nrg->to = t;
327
328out_locked:
329 resv->adds_in_progress--;
330 spin_unlock(&resv->lock);
331 VM_BUG_ON(add < 0);
332 return add;
333}
334
335/*
336 * Examine the existing reserve map and determine how many
337 * huge pages in the specified range [f, t) are NOT currently
338 * represented. This routine is called before a subsequent
339 * call to region_add that will actually modify the reserve
340 * map to add the specified range [f, t). region_chg does
341 * not change the number of huge pages represented by the
342 * map. However, if the existing regions in the map can not
343 * be expanded to represent the new range, a new file_region
344 * structure is added to the map as a placeholder. This is
345 * so that the subsequent region_add call will have all the
346 * regions it needs and will not fail.
347 *
348 * Upon entry, region_chg will also examine the cache of region descriptors
349 * associated with the map. If there are not enough descriptors cached, one
350 * will be allocated for the in progress add operation.
351 *
352 * Returns the number of huge pages that need to be added to the existing
353 * reservation map for the range [f, t). This number is greater or equal to
354 * zero. -ENOMEM is returned if a new file_region structure or cache entry
355 * is needed and can not be allocated.
356 */
357static long region_chg(struct resv_map *resv, long f, long t)
358{
359 struct list_head *head = &resv->regions;
360 struct file_region *rg, *nrg = NULL;
361 long chg = 0;
362
363retry:
364 spin_lock(&resv->lock);
365retry_locked:
366 resv->adds_in_progress++;
367
368 /*
369 * Check for sufficient descriptors in the cache to accommodate
370 * the number of in progress add operations.
371 */
372 if (resv->adds_in_progress > resv->region_cache_count) {
373 struct file_region *trg;
374
375 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
376 /* Must drop lock to allocate a new descriptor. */
377 resv->adds_in_progress--;
378 spin_unlock(&resv->lock);
379
380 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
381 if (!trg) {
382 kfree(nrg);
383 return -ENOMEM;
384 }
385
386 spin_lock(&resv->lock);
387 list_add(&trg->link, &resv->region_cache);
388 resv->region_cache_count++;
389 goto retry_locked;
390 }
391
392 /* Locate the region we are before or in. */
393 list_for_each_entry(rg, head, link)
394 if (f <= rg->to)
395 break;
396
397 /* If we are below the current region then a new region is required.
398 * Subtle, allocate a new region at the position but make it zero
399 * size such that we can guarantee to record the reservation. */
400 if (&rg->link == head || t < rg->from) {
401 if (!nrg) {
402 resv->adds_in_progress--;
403 spin_unlock(&resv->lock);
404 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
405 if (!nrg)
406 return -ENOMEM;
407
408 nrg->from = f;
409 nrg->to = f;
410 INIT_LIST_HEAD(&nrg->link);
411 goto retry;
412 }
413
414 list_add(&nrg->link, rg->link.prev);
415 chg = t - f;
416 goto out_nrg;
417 }
418
419 /* Round our left edge to the current segment if it encloses us. */
420 if (f > rg->from)
421 f = rg->from;
422 chg = t - f;
423
424 /* Check for and consume any regions we now overlap with. */
425 list_for_each_entry(rg, rg->link.prev, link) {
426 if (&rg->link == head)
427 break;
428 if (rg->from > t)
429 goto out;
430
431 /* We overlap with this area, if it extends further than
432 * us then we must extend ourselves. Account for its
433 * existing reservation. */
434 if (rg->to > t) {
435 chg += rg->to - t;
436 t = rg->to;
437 }
438 chg -= rg->to - rg->from;
439 }
440
441out:
442 spin_unlock(&resv->lock);
443 /* We already know we raced and no longer need the new region */
444 kfree(nrg);
445 return chg;
446out_nrg:
447 spin_unlock(&resv->lock);
448 return chg;
449}
450
451/*
452 * Abort the in progress add operation. The adds_in_progress field
453 * of the resv_map keeps track of the operations in progress between
454 * calls to region_chg and region_add. Operations are sometimes
455 * aborted after the call to region_chg. In such cases, region_abort
456 * is called to decrement the adds_in_progress counter.
457 *
458 * NOTE: The range arguments [f, t) are not needed or used in this
459 * routine. They are kept to make reading the calling code easier as
460 * arguments will match the associated region_chg call.
461 */
462static void region_abort(struct resv_map *resv, long f, long t)
463{
464 spin_lock(&resv->lock);
465 VM_BUG_ON(!resv->region_cache_count);
466 resv->adds_in_progress--;
467 spin_unlock(&resv->lock);
468}
469
470/*
471 * Delete the specified range [f, t) from the reserve map. If the
472 * t parameter is LONG_MAX, this indicates that ALL regions after f
473 * should be deleted. Locate the regions which intersect [f, t)
474 * and either trim, delete or split the existing regions.
475 *
476 * Returns the number of huge pages deleted from the reserve map.
477 * In the normal case, the return value is zero or more. In the
478 * case where a region must be split, a new region descriptor must
479 * be allocated. If the allocation fails, -ENOMEM will be returned.
480 * NOTE: If the parameter t == LONG_MAX, then we will never split
481 * a region and possibly return -ENOMEM. Callers specifying
482 * t == LONG_MAX do not need to check for -ENOMEM error.
483 */
484static long region_del(struct resv_map *resv, long f, long t)
485{
486 struct list_head *head = &resv->regions;
487 struct file_region *rg, *trg;
488 struct file_region *nrg = NULL;
489 long del = 0;
490
491retry:
492 spin_lock(&resv->lock);
493 list_for_each_entry_safe(rg, trg, head, link) {
494 /*
495 * Skip regions before the range to be deleted. file_region
496 * ranges are normally of the form [from, to). However, there
497 * may be a "placeholder" entry in the map which is of the form
498 * (from, to) with from == to. Check for placeholder entries
499 * at the beginning of the range to be deleted.
500 */
501 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
502 continue;
503
504 if (rg->from >= t)
505 break;
506
507 if (f > rg->from && t < rg->to) { /* Must split region */
508 /*
509 * Check for an entry in the cache before dropping
510 * lock and attempting allocation.
511 */
512 if (!nrg &&
513 resv->region_cache_count > resv->adds_in_progress) {
514 nrg = list_first_entry(&resv->region_cache,
515 struct file_region,
516 link);
517 list_del(&nrg->link);
518 resv->region_cache_count--;
519 }
520
521 if (!nrg) {
522 spin_unlock(&resv->lock);
523 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
524 if (!nrg)
525 return -ENOMEM;
526 goto retry;
527 }
528
529 del += t - f;
530
531 /* New entry for end of split region */
532 nrg->from = t;
533 nrg->to = rg->to;
534 INIT_LIST_HEAD(&nrg->link);
535
536 /* Original entry is trimmed */
537 rg->to = f;
538
539 list_add(&nrg->link, &rg->link);
540 nrg = NULL;
541 break;
542 }
543
544 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
545 del += rg->to - rg->from;
546 list_del(&rg->link);
547 kfree(rg);
548 continue;
549 }
550
551 if (f <= rg->from) { /* Trim beginning of region */
552 del += t - rg->from;
553 rg->from = t;
554 } else { /* Trim end of region */
555 del += rg->to - f;
556 rg->to = f;
557 }
558 }
559
560 spin_unlock(&resv->lock);
561 kfree(nrg);
562 return del;
563}
564
565/*
566 * A rare out of memory error was encountered which prevented removal of
567 * the reserve map region for a page. The huge page itself was free'ed
568 * and removed from the page cache. This routine will adjust the subpool
569 * usage count, and the global reserve count if needed. By incrementing
570 * these counts, the reserve map entry which could not be deleted will
571 * appear as a "reserved" entry instead of simply dangling with incorrect
572 * counts.
573 */
574void hugetlb_fix_reserve_counts(struct inode *inode)
575{
576 struct hugepage_subpool *spool = subpool_inode(inode);
577 long rsv_adjust;
578
579 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
580 if (rsv_adjust) {
581 struct hstate *h = hstate_inode(inode);
582
583 hugetlb_acct_memory(h, 1);
584 }
585}
586
587/*
588 * Count and return the number of huge pages in the reserve map
589 * that intersect with the range [f, t).
590 */
591static long region_count(struct resv_map *resv, long f, long t)
592{
593 struct list_head *head = &resv->regions;
594 struct file_region *rg;
595 long chg = 0;
596
597 spin_lock(&resv->lock);
598 /* Locate each segment we overlap with, and count that overlap. */
599 list_for_each_entry(rg, head, link) {
600 long seg_from;
601 long seg_to;
602
603 if (rg->to <= f)
604 continue;
605 if (rg->from >= t)
606 break;
607
608 seg_from = max(rg->from, f);
609 seg_to = min(rg->to, t);
610
611 chg += seg_to - seg_from;
612 }
613 spin_unlock(&resv->lock);
614
615 return chg;
616}
617
618/*
619 * Convert the address within this vma to the page offset within
620 * the mapping, in pagecache page units; huge pages here.
621 */
622static pgoff_t vma_hugecache_offset(struct hstate *h,
623 struct vm_area_struct *vma, unsigned long address)
624{
625 return ((address - vma->vm_start) >> huge_page_shift(h)) +
626 (vma->vm_pgoff >> huge_page_order(h));
627}
628
629pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
630 unsigned long address)
631{
632 return vma_hugecache_offset(hstate_vma(vma), vma, address);
633}
634EXPORT_SYMBOL_GPL(linear_hugepage_index);
635
636/*
637 * Return the size of the pages allocated when backing a VMA. In the majority
638 * cases this will be same size as used by the page table entries.
639 */
640unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
641{
642 if (vma->vm_ops && vma->vm_ops->pagesize)
643 return vma->vm_ops->pagesize(vma);
644 return PAGE_SIZE;
645}
646EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
647
648/*
649 * Return the page size being used by the MMU to back a VMA. In the majority
650 * of cases, the page size used by the kernel matches the MMU size. On
651 * architectures where it differs, an architecture-specific 'strong'
652 * version of this symbol is required.
653 */
654__weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
655{
656 return vma_kernel_pagesize(vma);
657}
658
659/*
660 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
661 * bits of the reservation map pointer, which are always clear due to
662 * alignment.
663 */
664#define HPAGE_RESV_OWNER (1UL << 0)
665#define HPAGE_RESV_UNMAPPED (1UL << 1)
666#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
667
668/*
669 * These helpers are used to track how many pages are reserved for
670 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
671 * is guaranteed to have their future faults succeed.
672 *
673 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
674 * the reserve counters are updated with the hugetlb_lock held. It is safe
675 * to reset the VMA at fork() time as it is not in use yet and there is no
676 * chance of the global counters getting corrupted as a result of the values.
677 *
678 * The private mapping reservation is represented in a subtly different
679 * manner to a shared mapping. A shared mapping has a region map associated
680 * with the underlying file, this region map represents the backing file
681 * pages which have ever had a reservation assigned which this persists even
682 * after the page is instantiated. A private mapping has a region map
683 * associated with the original mmap which is attached to all VMAs which
684 * reference it, this region map represents those offsets which have consumed
685 * reservation ie. where pages have been instantiated.
686 */
687static unsigned long get_vma_private_data(struct vm_area_struct *vma)
688{
689 return (unsigned long)vma->vm_private_data;
690}
691
692static void set_vma_private_data(struct vm_area_struct *vma,
693 unsigned long value)
694{
695 vma->vm_private_data = (void *)value;
696}
697
698struct resv_map *resv_map_alloc(void)
699{
700 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
701 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
702
703 if (!resv_map || !rg) {
704 kfree(resv_map);
705 kfree(rg);
706 return NULL;
707 }
708
709 kref_init(&resv_map->refs);
710 spin_lock_init(&resv_map->lock);
711 INIT_LIST_HEAD(&resv_map->regions);
712
713 resv_map->adds_in_progress = 0;
714
715 INIT_LIST_HEAD(&resv_map->region_cache);
716 list_add(&rg->link, &resv_map->region_cache);
717 resv_map->region_cache_count = 1;
718
719 return resv_map;
720}
721
722void resv_map_release(struct kref *ref)
723{
724 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
725 struct list_head *head = &resv_map->region_cache;
726 struct file_region *rg, *trg;
727
728 /* Clear out any active regions before we release the map. */
729 region_del(resv_map, 0, LONG_MAX);
730
731 /* ... and any entries left in the cache */
732 list_for_each_entry_safe(rg, trg, head, link) {
733 list_del(&rg->link);
734 kfree(rg);
735 }
736
737 VM_BUG_ON(resv_map->adds_in_progress);
738
739 kfree(resv_map);
740}
741
742static inline struct resv_map *inode_resv_map(struct inode *inode)
743{
744 /*
745 * At inode evict time, i_mapping may not point to the original
746 * address space within the inode. This original address space
747 * contains the pointer to the resv_map. So, always use the
748 * address space embedded within the inode.
749 * The VERY common case is inode->mapping == &inode->i_data but,
750 * this may not be true for device special inodes.
751 */
752 return (struct resv_map *)(&inode->i_data)->private_data;
753}
754
755static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
756{
757 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
758 if (vma->vm_flags & VM_MAYSHARE) {
759 struct address_space *mapping = vma->vm_file->f_mapping;
760 struct inode *inode = mapping->host;
761
762 return inode_resv_map(inode);
763
764 } else {
765 return (struct resv_map *)(get_vma_private_data(vma) &
766 ~HPAGE_RESV_MASK);
767 }
768}
769
770static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
771{
772 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
773 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
774
775 set_vma_private_data(vma, (get_vma_private_data(vma) &
776 HPAGE_RESV_MASK) | (unsigned long)map);
777}
778
779static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
780{
781 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
782 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
783
784 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
785}
786
787static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
788{
789 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
790
791 return (get_vma_private_data(vma) & flag) != 0;
792}
793
794/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
795void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
796{
797 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
798 if (!(vma->vm_flags & VM_MAYSHARE))
799 vma->vm_private_data = (void *)0;
800}
801
802/* Returns true if the VMA has associated reserve pages */
803static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
804{
805 if (vma->vm_flags & VM_NORESERVE) {
806 /*
807 * This address is already reserved by other process(chg == 0),
808 * so, we should decrement reserved count. Without decrementing,
809 * reserve count remains after releasing inode, because this
810 * allocated page will go into page cache and is regarded as
811 * coming from reserved pool in releasing step. Currently, we
812 * don't have any other solution to deal with this situation
813 * properly, so add work-around here.
814 */
815 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
816 return true;
817 else
818 return false;
819 }
820
821 /* Shared mappings always use reserves */
822 if (vma->vm_flags & VM_MAYSHARE) {
823 /*
824 * We know VM_NORESERVE is not set. Therefore, there SHOULD
825 * be a region map for all pages. The only situation where
826 * there is no region map is if a hole was punched via
827 * fallocate. In this case, there really are no reverves to
828 * use. This situation is indicated if chg != 0.
829 */
830 if (chg)
831 return false;
832 else
833 return true;
834 }
835
836 /*
837 * Only the process that called mmap() has reserves for
838 * private mappings.
839 */
840 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
841 /*
842 * Like the shared case above, a hole punch or truncate
843 * could have been performed on the private mapping.
844 * Examine the value of chg to determine if reserves
845 * actually exist or were previously consumed.
846 * Very Subtle - The value of chg comes from a previous
847 * call to vma_needs_reserves(). The reserve map for
848 * private mappings has different (opposite) semantics
849 * than that of shared mappings. vma_needs_reserves()
850 * has already taken this difference in semantics into
851 * account. Therefore, the meaning of chg is the same
852 * as in the shared case above. Code could easily be
853 * combined, but keeping it separate draws attention to
854 * subtle differences.
855 */
856 if (chg)
857 return false;
858 else
859 return true;
860 }
861
862 return false;
863}
864
865static void enqueue_huge_page(struct hstate *h, struct page *page)
866{
867 int nid = page_to_nid(page);
868 list_move(&page->lru, &h->hugepage_freelists[nid]);
869 h->free_huge_pages++;
870 h->free_huge_pages_node[nid]++;
871}
872
873static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
874{
875 struct page *page;
876
877 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
878 if (!PageHWPoison(page))
879 break;
880 /*
881 * if 'non-isolated free hugepage' not found on the list,
882 * the allocation fails.
883 */
884 if (&h->hugepage_freelists[nid] == &page->lru)
885 return NULL;
886 list_move(&page->lru, &h->hugepage_activelist);
887 set_page_refcounted(page);
888 h->free_huge_pages--;
889 h->free_huge_pages_node[nid]--;
890 return page;
891}
892
893static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
894 nodemask_t *nmask)
895{
896 unsigned int cpuset_mems_cookie;
897 struct zonelist *zonelist;
898 struct zone *zone;
899 struct zoneref *z;
900 int node = NUMA_NO_NODE;
901
902 zonelist = node_zonelist(nid, gfp_mask);
903
904retry_cpuset:
905 cpuset_mems_cookie = read_mems_allowed_begin();
906 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
907 struct page *page;
908
909 if (!cpuset_zone_allowed(zone, gfp_mask))
910 continue;
911 /*
912 * no need to ask again on the same node. Pool is node rather than
913 * zone aware
914 */
915 if (zone_to_nid(zone) == node)
916 continue;
917 node = zone_to_nid(zone);
918
919 page = dequeue_huge_page_node_exact(h, node);
920 if (page)
921 return page;
922 }
923 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
924 goto retry_cpuset;
925
926 return NULL;
927}
928
929/* Movability of hugepages depends on migration support. */
930static inline gfp_t htlb_alloc_mask(struct hstate *h)
931{
932 if (hugepage_movable_supported(h))
933 return GFP_HIGHUSER_MOVABLE;
934 else
935 return GFP_HIGHUSER;
936}
937
938static struct page *dequeue_huge_page_vma(struct hstate *h,
939 struct vm_area_struct *vma,
940 unsigned long address, int avoid_reserve,
941 long chg)
942{
943 struct page *page;
944 struct mempolicy *mpol;
945 gfp_t gfp_mask;
946 nodemask_t *nodemask;
947 int nid;
948
949 /*
950 * A child process with MAP_PRIVATE mappings created by their parent
951 * have no page reserves. This check ensures that reservations are
952 * not "stolen". The child may still get SIGKILLed
953 */
954 if (!vma_has_reserves(vma, chg) &&
955 h->free_huge_pages - h->resv_huge_pages == 0)
956 goto err;
957
958 /* If reserves cannot be used, ensure enough pages are in the pool */
959 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
960 goto err;
961
962 gfp_mask = htlb_alloc_mask(h);
963 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
964 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
965 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
966 SetPagePrivate(page);
967 h->resv_huge_pages--;
968 }
969
970 mpol_cond_put(mpol);
971 return page;
972
973err:
974 return NULL;
975}
976
977/*
978 * common helper functions for hstate_next_node_to_{alloc|free}.
979 * We may have allocated or freed a huge page based on a different
980 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
981 * be outside of *nodes_allowed. Ensure that we use an allowed
982 * node for alloc or free.
983 */
984static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
985{
986 nid = next_node_in(nid, *nodes_allowed);
987 VM_BUG_ON(nid >= MAX_NUMNODES);
988
989 return nid;
990}
991
992static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
993{
994 if (!node_isset(nid, *nodes_allowed))
995 nid = next_node_allowed(nid, nodes_allowed);
996 return nid;
997}
998
999/*
1000 * returns the previously saved node ["this node"] from which to
1001 * allocate a persistent huge page for the pool and advance the
1002 * next node from which to allocate, handling wrap at end of node
1003 * mask.
1004 */
1005static int hstate_next_node_to_alloc(struct hstate *h,
1006 nodemask_t *nodes_allowed)
1007{
1008 int nid;
1009
1010 VM_BUG_ON(!nodes_allowed);
1011
1012 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1013 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1014
1015 return nid;
1016}
1017
1018/*
1019 * helper for free_pool_huge_page() - return the previously saved
1020 * node ["this node"] from which to free a huge page. Advance the
1021 * next node id whether or not we find a free huge page to free so
1022 * that the next attempt to free addresses the next node.
1023 */
1024static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1025{
1026 int nid;
1027
1028 VM_BUG_ON(!nodes_allowed);
1029
1030 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1031 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1032
1033 return nid;
1034}
1035
1036#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1037 for (nr_nodes = nodes_weight(*mask); \
1038 nr_nodes > 0 && \
1039 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1040 nr_nodes--)
1041
1042#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1043 for (nr_nodes = nodes_weight(*mask); \
1044 nr_nodes > 0 && \
1045 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1046 nr_nodes--)
1047
1048#ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1049static void destroy_compound_gigantic_page(struct page *page,
1050 unsigned int order)
1051{
1052 int i;
1053 int nr_pages = 1 << order;
1054 struct page *p = page + 1;
1055
1056 atomic_set(compound_mapcount_ptr(page), 0);
1057 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1058 clear_compound_head(p);
1059 set_page_refcounted(p);
1060 }
1061
1062 set_compound_order(page, 0);
1063 __ClearPageHead(page);
1064}
1065
1066static void free_gigantic_page(struct page *page, unsigned int order)
1067{
1068 free_contig_range(page_to_pfn(page), 1 << order);
1069}
1070
1071#ifdef CONFIG_CONTIG_ALLOC
1072static int __alloc_gigantic_page(unsigned long start_pfn,
1073 unsigned long nr_pages, gfp_t gfp_mask)
1074{
1075 unsigned long end_pfn = start_pfn + nr_pages;
1076 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1077 gfp_mask);
1078}
1079
1080static bool pfn_range_valid_gigantic(struct zone *z,
1081 unsigned long start_pfn, unsigned long nr_pages)
1082{
1083 unsigned long i, end_pfn = start_pfn + nr_pages;
1084 struct page *page;
1085
1086 for (i = start_pfn; i < end_pfn; i++) {
1087 page = pfn_to_online_page(i);
1088 if (!page)
1089 return false;
1090
1091 if (page_zone(page) != z)
1092 return false;
1093
1094 if (PageReserved(page))
1095 return false;
1096
1097 if (page_count(page) > 0)
1098 return false;
1099
1100 if (PageHuge(page))
1101 return false;
1102 }
1103
1104 return true;
1105}
1106
1107static bool zone_spans_last_pfn(const struct zone *zone,
1108 unsigned long start_pfn, unsigned long nr_pages)
1109{
1110 unsigned long last_pfn = start_pfn + nr_pages - 1;
1111 return zone_spans_pfn(zone, last_pfn);
1112}
1113
1114static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1115 int nid, nodemask_t *nodemask)
1116{
1117 unsigned int order = huge_page_order(h);
1118 unsigned long nr_pages = 1 << order;
1119 unsigned long ret, pfn, flags;
1120 struct zonelist *zonelist;
1121 struct zone *zone;
1122 struct zoneref *z;
1123
1124 zonelist = node_zonelist(nid, gfp_mask);
1125 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) {
1126 spin_lock_irqsave(&zone->lock, flags);
1127
1128 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
1129 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
1130 if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) {
1131 /*
1132 * We release the zone lock here because
1133 * alloc_contig_range() will also lock the zone
1134 * at some point. If there's an allocation
1135 * spinning on this lock, it may win the race
1136 * and cause alloc_contig_range() to fail...
1137 */
1138 spin_unlock_irqrestore(&zone->lock, flags);
1139 ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask);
1140 if (!ret)
1141 return pfn_to_page(pfn);
1142 spin_lock_irqsave(&zone->lock, flags);
1143 }
1144 pfn += nr_pages;
1145 }
1146
1147 spin_unlock_irqrestore(&zone->lock, flags);
1148 }
1149
1150 return NULL;
1151}
1152
1153static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1154static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1155#else /* !CONFIG_CONTIG_ALLOC */
1156static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1157 int nid, nodemask_t *nodemask)
1158{
1159 return NULL;
1160}
1161#endif /* CONFIG_CONTIG_ALLOC */
1162
1163#else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1164static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1165 int nid, nodemask_t *nodemask)
1166{
1167 return NULL;
1168}
1169static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1170static inline void destroy_compound_gigantic_page(struct page *page,
1171 unsigned int order) { }
1172#endif
1173
1174static void update_and_free_page(struct hstate *h, struct page *page)
1175{
1176 int i;
1177
1178 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1179 return;
1180
1181 h->nr_huge_pages--;
1182 h->nr_huge_pages_node[page_to_nid(page)]--;
1183 for (i = 0; i < pages_per_huge_page(h); i++) {
1184 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1185 1 << PG_referenced | 1 << PG_dirty |
1186 1 << PG_active | 1 << PG_private |
1187 1 << PG_writeback);
1188 }
1189 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1190 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1191 set_page_refcounted(page);
1192 if (hstate_is_gigantic(h)) {
1193 destroy_compound_gigantic_page(page, huge_page_order(h));
1194 free_gigantic_page(page, huge_page_order(h));
1195 } else {
1196 __free_pages(page, huge_page_order(h));
1197 }
1198}
1199
1200struct hstate *size_to_hstate(unsigned long size)
1201{
1202 struct hstate *h;
1203
1204 for_each_hstate(h) {
1205 if (huge_page_size(h) == size)
1206 return h;
1207 }
1208 return NULL;
1209}
1210
1211/*
1212 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1213 * to hstate->hugepage_activelist.)
1214 *
1215 * This function can be called for tail pages, but never returns true for them.
1216 */
1217bool page_huge_active(struct page *page)
1218{
1219 VM_BUG_ON_PAGE(!PageHuge(page), page);
1220 return PageHead(page) && PagePrivate(&page[1]);
1221}
1222
1223/* never called for tail page */
1224static void set_page_huge_active(struct page *page)
1225{
1226 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1227 SetPagePrivate(&page[1]);
1228}
1229
1230static void clear_page_huge_active(struct page *page)
1231{
1232 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1233 ClearPagePrivate(&page[1]);
1234}
1235
1236/*
1237 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1238 * code
1239 */
1240static inline bool PageHugeTemporary(struct page *page)
1241{
1242 if (!PageHuge(page))
1243 return false;
1244
1245 return (unsigned long)page[2].mapping == -1U;
1246}
1247
1248static inline void SetPageHugeTemporary(struct page *page)
1249{
1250 page[2].mapping = (void *)-1U;
1251}
1252
1253static inline void ClearPageHugeTemporary(struct page *page)
1254{
1255 page[2].mapping = NULL;
1256}
1257
1258void free_huge_page(struct page *page)
1259{
1260 /*
1261 * Can't pass hstate in here because it is called from the
1262 * compound page destructor.
1263 */
1264 struct hstate *h = page_hstate(page);
1265 int nid = page_to_nid(page);
1266 struct hugepage_subpool *spool =
1267 (struct hugepage_subpool *)page_private(page);
1268 bool restore_reserve;
1269
1270 VM_BUG_ON_PAGE(page_count(page), page);
1271 VM_BUG_ON_PAGE(page_mapcount(page), page);
1272
1273 set_page_private(page, 0);
1274 page->mapping = NULL;
1275 restore_reserve = PagePrivate(page);
1276 ClearPagePrivate(page);
1277
1278 /*
1279 * If PagePrivate() was set on page, page allocation consumed a
1280 * reservation. If the page was associated with a subpool, there
1281 * would have been a page reserved in the subpool before allocation
1282 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1283 * reservtion, do not call hugepage_subpool_put_pages() as this will
1284 * remove the reserved page from the subpool.
1285 */
1286 if (!restore_reserve) {
1287 /*
1288 * A return code of zero implies that the subpool will be
1289 * under its minimum size if the reservation is not restored
1290 * after page is free. Therefore, force restore_reserve
1291 * operation.
1292 */
1293 if (hugepage_subpool_put_pages(spool, 1) == 0)
1294 restore_reserve = true;
1295 }
1296
1297 spin_lock(&hugetlb_lock);
1298 clear_page_huge_active(page);
1299 hugetlb_cgroup_uncharge_page(hstate_index(h),
1300 pages_per_huge_page(h), page);
1301 if (restore_reserve)
1302 h->resv_huge_pages++;
1303
1304 if (PageHugeTemporary(page)) {
1305 list_del(&page->lru);
1306 ClearPageHugeTemporary(page);
1307 update_and_free_page(h, page);
1308 } else if (h->surplus_huge_pages_node[nid]) {
1309 /* remove the page from active list */
1310 list_del(&page->lru);
1311 update_and_free_page(h, page);
1312 h->surplus_huge_pages--;
1313 h->surplus_huge_pages_node[nid]--;
1314 } else {
1315 arch_clear_hugepage_flags(page);
1316 enqueue_huge_page(h, page);
1317 }
1318 spin_unlock(&hugetlb_lock);
1319}
1320
1321static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1322{
1323 INIT_LIST_HEAD(&page->lru);
1324 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1325 spin_lock(&hugetlb_lock);
1326 set_hugetlb_cgroup(page, NULL);
1327 h->nr_huge_pages++;
1328 h->nr_huge_pages_node[nid]++;
1329 spin_unlock(&hugetlb_lock);
1330}
1331
1332static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1333{
1334 int i;
1335 int nr_pages = 1 << order;
1336 struct page *p = page + 1;
1337
1338 /* we rely on prep_new_huge_page to set the destructor */
1339 set_compound_order(page, order);
1340 __ClearPageReserved(page);
1341 __SetPageHead(page);
1342 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1343 /*
1344 * For gigantic hugepages allocated through bootmem at
1345 * boot, it's safer to be consistent with the not-gigantic
1346 * hugepages and clear the PG_reserved bit from all tail pages
1347 * too. Otherwse drivers using get_user_pages() to access tail
1348 * pages may get the reference counting wrong if they see
1349 * PG_reserved set on a tail page (despite the head page not
1350 * having PG_reserved set). Enforcing this consistency between
1351 * head and tail pages allows drivers to optimize away a check
1352 * on the head page when they need know if put_page() is needed
1353 * after get_user_pages().
1354 */
1355 __ClearPageReserved(p);
1356 set_page_count(p, 0);
1357 set_compound_head(p, page);
1358 }
1359 atomic_set(compound_mapcount_ptr(page), -1);
1360}
1361
1362/*
1363 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1364 * transparent huge pages. See the PageTransHuge() documentation for more
1365 * details.
1366 */
1367int PageHuge(struct page *page)
1368{
1369 if (!PageCompound(page))
1370 return 0;
1371
1372 page = compound_head(page);
1373 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1374}
1375EXPORT_SYMBOL_GPL(PageHuge);
1376
1377/*
1378 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1379 * normal or transparent huge pages.
1380 */
1381int PageHeadHuge(struct page *page_head)
1382{
1383 if (!PageHead(page_head))
1384 return 0;
1385
1386 return get_compound_page_dtor(page_head) == free_huge_page;
1387}
1388
1389pgoff_t __basepage_index(struct page *page)
1390{
1391 struct page *page_head = compound_head(page);
1392 pgoff_t index = page_index(page_head);
1393 unsigned long compound_idx;
1394
1395 if (!PageHuge(page_head))
1396 return page_index(page);
1397
1398 if (compound_order(page_head) >= MAX_ORDER)
1399 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1400 else
1401 compound_idx = page - page_head;
1402
1403 return (index << compound_order(page_head)) + compound_idx;
1404}
1405
1406static struct page *alloc_buddy_huge_page(struct hstate *h,
1407 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1408 nodemask_t *node_alloc_noretry)
1409{
1410 int order = huge_page_order(h);
1411 struct page *page;
1412 bool alloc_try_hard = true;
1413
1414 /*
1415 * By default we always try hard to allocate the page with
1416 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1417 * a loop (to adjust global huge page counts) and previous allocation
1418 * failed, do not continue to try hard on the same node. Use the
1419 * node_alloc_noretry bitmap to manage this state information.
1420 */
1421 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1422 alloc_try_hard = false;
1423 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1424 if (alloc_try_hard)
1425 gfp_mask |= __GFP_RETRY_MAYFAIL;
1426 if (nid == NUMA_NO_NODE)
1427 nid = numa_mem_id();
1428 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1429 if (page)
1430 __count_vm_event(HTLB_BUDDY_PGALLOC);
1431 else
1432 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1433
1434 /*
1435 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1436 * indicates an overall state change. Clear bit so that we resume
1437 * normal 'try hard' allocations.
1438 */
1439 if (node_alloc_noretry && page && !alloc_try_hard)
1440 node_clear(nid, *node_alloc_noretry);
1441
1442 /*
1443 * If we tried hard to get a page but failed, set bit so that
1444 * subsequent attempts will not try as hard until there is an
1445 * overall state change.
1446 */
1447 if (node_alloc_noretry && !page && alloc_try_hard)
1448 node_set(nid, *node_alloc_noretry);
1449
1450 return page;
1451}
1452
1453/*
1454 * Common helper to allocate a fresh hugetlb page. All specific allocators
1455 * should use this function to get new hugetlb pages
1456 */
1457static struct page *alloc_fresh_huge_page(struct hstate *h,
1458 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1459 nodemask_t *node_alloc_noretry)
1460{
1461 struct page *page;
1462
1463 if (hstate_is_gigantic(h))
1464 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1465 else
1466 page = alloc_buddy_huge_page(h, gfp_mask,
1467 nid, nmask, node_alloc_noretry);
1468 if (!page)
1469 return NULL;
1470
1471 if (hstate_is_gigantic(h))
1472 prep_compound_gigantic_page(page, huge_page_order(h));
1473 prep_new_huge_page(h, page, page_to_nid(page));
1474
1475 return page;
1476}
1477
1478/*
1479 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1480 * manner.
1481 */
1482static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1483 nodemask_t *node_alloc_noretry)
1484{
1485 struct page *page;
1486 int nr_nodes, node;
1487 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1488
1489 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1490 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1491 node_alloc_noretry);
1492 if (page)
1493 break;
1494 }
1495
1496 if (!page)
1497 return 0;
1498
1499 put_page(page); /* free it into the hugepage allocator */
1500
1501 return 1;
1502}
1503
1504/*
1505 * Free huge page from pool from next node to free.
1506 * Attempt to keep persistent huge pages more or less
1507 * balanced over allowed nodes.
1508 * Called with hugetlb_lock locked.
1509 */
1510static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1511 bool acct_surplus)
1512{
1513 int nr_nodes, node;
1514 int ret = 0;
1515
1516 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1517 /*
1518 * If we're returning unused surplus pages, only examine
1519 * nodes with surplus pages.
1520 */
1521 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1522 !list_empty(&h->hugepage_freelists[node])) {
1523 struct page *page =
1524 list_entry(h->hugepage_freelists[node].next,
1525 struct page, lru);
1526 list_del(&page->lru);
1527 h->free_huge_pages--;
1528 h->free_huge_pages_node[node]--;
1529 if (acct_surplus) {
1530 h->surplus_huge_pages--;
1531 h->surplus_huge_pages_node[node]--;
1532 }
1533 update_and_free_page(h, page);
1534 ret = 1;
1535 break;
1536 }
1537 }
1538
1539 return ret;
1540}
1541
1542/*
1543 * Dissolve a given free hugepage into free buddy pages. This function does
1544 * nothing for in-use hugepages and non-hugepages.
1545 * This function returns values like below:
1546 *
1547 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1548 * (allocated or reserved.)
1549 * 0: successfully dissolved free hugepages or the page is not a
1550 * hugepage (considered as already dissolved)
1551 */
1552int dissolve_free_huge_page(struct page *page)
1553{
1554 int rc = -EBUSY;
1555
1556 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1557 if (!PageHuge(page))
1558 return 0;
1559
1560 spin_lock(&hugetlb_lock);
1561 if (!PageHuge(page)) {
1562 rc = 0;
1563 goto out;
1564 }
1565
1566 if (!page_count(page)) {
1567 struct page *head = compound_head(page);
1568 struct hstate *h = page_hstate(head);
1569 int nid = page_to_nid(head);
1570 if (h->free_huge_pages - h->resv_huge_pages == 0)
1571 goto out;
1572 /*
1573 * Move PageHWPoison flag from head page to the raw error page,
1574 * which makes any subpages rather than the error page reusable.
1575 */
1576 if (PageHWPoison(head) && page != head) {
1577 SetPageHWPoison(page);
1578 ClearPageHWPoison(head);
1579 }
1580 list_del(&head->lru);
1581 h->free_huge_pages--;
1582 h->free_huge_pages_node[nid]--;
1583 h->max_huge_pages--;
1584 update_and_free_page(h, head);
1585 rc = 0;
1586 }
1587out:
1588 spin_unlock(&hugetlb_lock);
1589 return rc;
1590}
1591
1592/*
1593 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1594 * make specified memory blocks removable from the system.
1595 * Note that this will dissolve a free gigantic hugepage completely, if any
1596 * part of it lies within the given range.
1597 * Also note that if dissolve_free_huge_page() returns with an error, all
1598 * free hugepages that were dissolved before that error are lost.
1599 */
1600int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1601{
1602 unsigned long pfn;
1603 struct page *page;
1604 int rc = 0;
1605
1606 if (!hugepages_supported())
1607 return rc;
1608
1609 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1610 page = pfn_to_page(pfn);
1611 rc = dissolve_free_huge_page(page);
1612 if (rc)
1613 break;
1614 }
1615
1616 return rc;
1617}
1618
1619/*
1620 * Allocates a fresh surplus page from the page allocator.
1621 */
1622static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1623 int nid, nodemask_t *nmask)
1624{
1625 struct page *page = NULL;
1626
1627 if (hstate_is_gigantic(h))
1628 return NULL;
1629
1630 spin_lock(&hugetlb_lock);
1631 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1632 goto out_unlock;
1633 spin_unlock(&hugetlb_lock);
1634
1635 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1636 if (!page)
1637 return NULL;
1638
1639 spin_lock(&hugetlb_lock);
1640 /*
1641 * We could have raced with the pool size change.
1642 * Double check that and simply deallocate the new page
1643 * if we would end up overcommiting the surpluses. Abuse
1644 * temporary page to workaround the nasty free_huge_page
1645 * codeflow
1646 */
1647 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1648 SetPageHugeTemporary(page);
1649 spin_unlock(&hugetlb_lock);
1650 put_page(page);
1651 return NULL;
1652 } else {
1653 h->surplus_huge_pages++;
1654 h->surplus_huge_pages_node[page_to_nid(page)]++;
1655 }
1656
1657out_unlock:
1658 spin_unlock(&hugetlb_lock);
1659
1660 return page;
1661}
1662
1663struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1664 int nid, nodemask_t *nmask)
1665{
1666 struct page *page;
1667
1668 if (hstate_is_gigantic(h))
1669 return NULL;
1670
1671 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1672 if (!page)
1673 return NULL;
1674
1675 /*
1676 * We do not account these pages as surplus because they are only
1677 * temporary and will be released properly on the last reference
1678 */
1679 SetPageHugeTemporary(page);
1680
1681 return page;
1682}
1683
1684/*
1685 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1686 */
1687static
1688struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1689 struct vm_area_struct *vma, unsigned long addr)
1690{
1691 struct page *page;
1692 struct mempolicy *mpol;
1693 gfp_t gfp_mask = htlb_alloc_mask(h);
1694 int nid;
1695 nodemask_t *nodemask;
1696
1697 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1698 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1699 mpol_cond_put(mpol);
1700
1701 return page;
1702}
1703
1704/* page migration callback function */
1705struct page *alloc_huge_page_node(struct hstate *h, int nid)
1706{
1707 gfp_t gfp_mask = htlb_alloc_mask(h);
1708 struct page *page = NULL;
1709
1710 if (nid != NUMA_NO_NODE)
1711 gfp_mask |= __GFP_THISNODE;
1712
1713 spin_lock(&hugetlb_lock);
1714 if (h->free_huge_pages - h->resv_huge_pages > 0)
1715 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1716 spin_unlock(&hugetlb_lock);
1717
1718 if (!page)
1719 page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1720
1721 return page;
1722}
1723
1724/* page migration callback function */
1725struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1726 nodemask_t *nmask)
1727{
1728 gfp_t gfp_mask = htlb_alloc_mask(h);
1729
1730 spin_lock(&hugetlb_lock);
1731 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1732 struct page *page;
1733
1734 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1735 if (page) {
1736 spin_unlock(&hugetlb_lock);
1737 return page;
1738 }
1739 }
1740 spin_unlock(&hugetlb_lock);
1741
1742 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1743}
1744
1745/* mempolicy aware migration callback */
1746struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1747 unsigned long address)
1748{
1749 struct mempolicy *mpol;
1750 nodemask_t *nodemask;
1751 struct page *page;
1752 gfp_t gfp_mask;
1753 int node;
1754
1755 gfp_mask = htlb_alloc_mask(h);
1756 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1757 page = alloc_huge_page_nodemask(h, node, nodemask);
1758 mpol_cond_put(mpol);
1759
1760 return page;
1761}
1762
1763/*
1764 * Increase the hugetlb pool such that it can accommodate a reservation
1765 * of size 'delta'.
1766 */
1767static int gather_surplus_pages(struct hstate *h, int delta)
1768{
1769 struct list_head surplus_list;
1770 struct page *page, *tmp;
1771 int ret, i;
1772 int needed, allocated;
1773 bool alloc_ok = true;
1774
1775 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1776 if (needed <= 0) {
1777 h->resv_huge_pages += delta;
1778 return 0;
1779 }
1780
1781 allocated = 0;
1782 INIT_LIST_HEAD(&surplus_list);
1783
1784 ret = -ENOMEM;
1785retry:
1786 spin_unlock(&hugetlb_lock);
1787 for (i = 0; i < needed; i++) {
1788 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1789 NUMA_NO_NODE, NULL);
1790 if (!page) {
1791 alloc_ok = false;
1792 break;
1793 }
1794 list_add(&page->lru, &surplus_list);
1795 cond_resched();
1796 }
1797 allocated += i;
1798
1799 /*
1800 * After retaking hugetlb_lock, we need to recalculate 'needed'
1801 * because either resv_huge_pages or free_huge_pages may have changed.
1802 */
1803 spin_lock(&hugetlb_lock);
1804 needed = (h->resv_huge_pages + delta) -
1805 (h->free_huge_pages + allocated);
1806 if (needed > 0) {
1807 if (alloc_ok)
1808 goto retry;
1809 /*
1810 * We were not able to allocate enough pages to
1811 * satisfy the entire reservation so we free what
1812 * we've allocated so far.
1813 */
1814 goto free;
1815 }
1816 /*
1817 * The surplus_list now contains _at_least_ the number of extra pages
1818 * needed to accommodate the reservation. Add the appropriate number
1819 * of pages to the hugetlb pool and free the extras back to the buddy
1820 * allocator. Commit the entire reservation here to prevent another
1821 * process from stealing the pages as they are added to the pool but
1822 * before they are reserved.
1823 */
1824 needed += allocated;
1825 h->resv_huge_pages += delta;
1826 ret = 0;
1827
1828 /* Free the needed pages to the hugetlb pool */
1829 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1830 if ((--needed) < 0)
1831 break;
1832 /*
1833 * This page is now managed by the hugetlb allocator and has
1834 * no users -- drop the buddy allocator's reference.
1835 */
1836 put_page_testzero(page);
1837 VM_BUG_ON_PAGE(page_count(page), page);
1838 enqueue_huge_page(h, page);
1839 }
1840free:
1841 spin_unlock(&hugetlb_lock);
1842
1843 /* Free unnecessary surplus pages to the buddy allocator */
1844 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1845 put_page(page);
1846 spin_lock(&hugetlb_lock);
1847
1848 return ret;
1849}
1850
1851/*
1852 * This routine has two main purposes:
1853 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1854 * in unused_resv_pages. This corresponds to the prior adjustments made
1855 * to the associated reservation map.
1856 * 2) Free any unused surplus pages that may have been allocated to satisfy
1857 * the reservation. As many as unused_resv_pages may be freed.
1858 *
1859 * Called with hugetlb_lock held. However, the lock could be dropped (and
1860 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1861 * we must make sure nobody else can claim pages we are in the process of
1862 * freeing. Do this by ensuring resv_huge_page always is greater than the
1863 * number of huge pages we plan to free when dropping the lock.
1864 */
1865static void return_unused_surplus_pages(struct hstate *h,
1866 unsigned long unused_resv_pages)
1867{
1868 unsigned long nr_pages;
1869
1870 /* Cannot return gigantic pages currently */
1871 if (hstate_is_gigantic(h))
1872 goto out;
1873
1874 /*
1875 * Part (or even all) of the reservation could have been backed
1876 * by pre-allocated pages. Only free surplus pages.
1877 */
1878 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1879
1880 /*
1881 * We want to release as many surplus pages as possible, spread
1882 * evenly across all nodes with memory. Iterate across these nodes
1883 * until we can no longer free unreserved surplus pages. This occurs
1884 * when the nodes with surplus pages have no free pages.
1885 * free_pool_huge_page() will balance the the freed pages across the
1886 * on-line nodes with memory and will handle the hstate accounting.
1887 *
1888 * Note that we decrement resv_huge_pages as we free the pages. If
1889 * we drop the lock, resv_huge_pages will still be sufficiently large
1890 * to cover subsequent pages we may free.
1891 */
1892 while (nr_pages--) {
1893 h->resv_huge_pages--;
1894 unused_resv_pages--;
1895 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1896 goto out;
1897 cond_resched_lock(&hugetlb_lock);
1898 }
1899
1900out:
1901 /* Fully uncommit the reservation */
1902 h->resv_huge_pages -= unused_resv_pages;
1903}
1904
1905
1906/*
1907 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1908 * are used by the huge page allocation routines to manage reservations.
1909 *
1910 * vma_needs_reservation is called to determine if the huge page at addr
1911 * within the vma has an associated reservation. If a reservation is
1912 * needed, the value 1 is returned. The caller is then responsible for
1913 * managing the global reservation and subpool usage counts. After
1914 * the huge page has been allocated, vma_commit_reservation is called
1915 * to add the page to the reservation map. If the page allocation fails,
1916 * the reservation must be ended instead of committed. vma_end_reservation
1917 * is called in such cases.
1918 *
1919 * In the normal case, vma_commit_reservation returns the same value
1920 * as the preceding vma_needs_reservation call. The only time this
1921 * is not the case is if a reserve map was changed between calls. It
1922 * is the responsibility of the caller to notice the difference and
1923 * take appropriate action.
1924 *
1925 * vma_add_reservation is used in error paths where a reservation must
1926 * be restored when a newly allocated huge page must be freed. It is
1927 * to be called after calling vma_needs_reservation to determine if a
1928 * reservation exists.
1929 */
1930enum vma_resv_mode {
1931 VMA_NEEDS_RESV,
1932 VMA_COMMIT_RESV,
1933 VMA_END_RESV,
1934 VMA_ADD_RESV,
1935};
1936static long __vma_reservation_common(struct hstate *h,
1937 struct vm_area_struct *vma, unsigned long addr,
1938 enum vma_resv_mode mode)
1939{
1940 struct resv_map *resv;
1941 pgoff_t idx;
1942 long ret;
1943
1944 resv = vma_resv_map(vma);
1945 if (!resv)
1946 return 1;
1947
1948 idx = vma_hugecache_offset(h, vma, addr);
1949 switch (mode) {
1950 case VMA_NEEDS_RESV:
1951 ret = region_chg(resv, idx, idx + 1);
1952 break;
1953 case VMA_COMMIT_RESV:
1954 ret = region_add(resv, idx, idx + 1);
1955 break;
1956 case VMA_END_RESV:
1957 region_abort(resv, idx, idx + 1);
1958 ret = 0;
1959 break;
1960 case VMA_ADD_RESV:
1961 if (vma->vm_flags & VM_MAYSHARE)
1962 ret = region_add(resv, idx, idx + 1);
1963 else {
1964 region_abort(resv, idx, idx + 1);
1965 ret = region_del(resv, idx, idx + 1);
1966 }
1967 break;
1968 default:
1969 BUG();
1970 }
1971
1972 if (vma->vm_flags & VM_MAYSHARE)
1973 return ret;
1974 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1975 /*
1976 * In most cases, reserves always exist for private mappings.
1977 * However, a file associated with mapping could have been
1978 * hole punched or truncated after reserves were consumed.
1979 * As subsequent fault on such a range will not use reserves.
1980 * Subtle - The reserve map for private mappings has the
1981 * opposite meaning than that of shared mappings. If NO
1982 * entry is in the reserve map, it means a reservation exists.
1983 * If an entry exists in the reserve map, it means the
1984 * reservation has already been consumed. As a result, the
1985 * return value of this routine is the opposite of the
1986 * value returned from reserve map manipulation routines above.
1987 */
1988 if (ret)
1989 return 0;
1990 else
1991 return 1;
1992 }
1993 else
1994 return ret < 0 ? ret : 0;
1995}
1996
1997static long vma_needs_reservation(struct hstate *h,
1998 struct vm_area_struct *vma, unsigned long addr)
1999{
2000 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2001}
2002
2003static long vma_commit_reservation(struct hstate *h,
2004 struct vm_area_struct *vma, unsigned long addr)
2005{
2006 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2007}
2008
2009static void vma_end_reservation(struct hstate *h,
2010 struct vm_area_struct *vma, unsigned long addr)
2011{
2012 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2013}
2014
2015static long vma_add_reservation(struct hstate *h,
2016 struct vm_area_struct *vma, unsigned long addr)
2017{
2018 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2019}
2020
2021/*
2022 * This routine is called to restore a reservation on error paths. In the
2023 * specific error paths, a huge page was allocated (via alloc_huge_page)
2024 * and is about to be freed. If a reservation for the page existed,
2025 * alloc_huge_page would have consumed the reservation and set PagePrivate
2026 * in the newly allocated page. When the page is freed via free_huge_page,
2027 * the global reservation count will be incremented if PagePrivate is set.
2028 * However, free_huge_page can not adjust the reserve map. Adjust the
2029 * reserve map here to be consistent with global reserve count adjustments
2030 * to be made by free_huge_page.
2031 */
2032static void restore_reserve_on_error(struct hstate *h,
2033 struct vm_area_struct *vma, unsigned long address,
2034 struct page *page)
2035{
2036 if (unlikely(PagePrivate(page))) {
2037 long rc = vma_needs_reservation(h, vma, address);
2038
2039 if (unlikely(rc < 0)) {
2040 /*
2041 * Rare out of memory condition in reserve map
2042 * manipulation. Clear PagePrivate so that
2043 * global reserve count will not be incremented
2044 * by free_huge_page. This will make it appear
2045 * as though the reservation for this page was
2046 * consumed. This may prevent the task from
2047 * faulting in the page at a later time. This
2048 * is better than inconsistent global huge page
2049 * accounting of reserve counts.
2050 */
2051 ClearPagePrivate(page);
2052 } else if (rc) {
2053 rc = vma_add_reservation(h, vma, address);
2054 if (unlikely(rc < 0))
2055 /*
2056 * See above comment about rare out of
2057 * memory condition.
2058 */
2059 ClearPagePrivate(page);
2060 } else
2061 vma_end_reservation(h, vma, address);
2062 }
2063}
2064
2065struct page *alloc_huge_page(struct vm_area_struct *vma,
2066 unsigned long addr, int avoid_reserve)
2067{
2068 struct hugepage_subpool *spool = subpool_vma(vma);
2069 struct hstate *h = hstate_vma(vma);
2070 struct page *page;
2071 long map_chg, map_commit;
2072 long gbl_chg;
2073 int ret, idx;
2074 struct hugetlb_cgroup *h_cg;
2075
2076 idx = hstate_index(h);
2077 /*
2078 * Examine the region/reserve map to determine if the process
2079 * has a reservation for the page to be allocated. A return
2080 * code of zero indicates a reservation exists (no change).
2081 */
2082 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2083 if (map_chg < 0)
2084 return ERR_PTR(-ENOMEM);
2085
2086 /*
2087 * Processes that did not create the mapping will have no
2088 * reserves as indicated by the region/reserve map. Check
2089 * that the allocation will not exceed the subpool limit.
2090 * Allocations for MAP_NORESERVE mappings also need to be
2091 * checked against any subpool limit.
2092 */
2093 if (map_chg || avoid_reserve) {
2094 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2095 if (gbl_chg < 0) {
2096 vma_end_reservation(h, vma, addr);
2097 return ERR_PTR(-ENOSPC);
2098 }
2099
2100 /*
2101 * Even though there was no reservation in the region/reserve
2102 * map, there could be reservations associated with the
2103 * subpool that can be used. This would be indicated if the
2104 * return value of hugepage_subpool_get_pages() is zero.
2105 * However, if avoid_reserve is specified we still avoid even
2106 * the subpool reservations.
2107 */
2108 if (avoid_reserve)
2109 gbl_chg = 1;
2110 }
2111
2112 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2113 if (ret)
2114 goto out_subpool_put;
2115
2116 spin_lock(&hugetlb_lock);
2117 /*
2118 * glb_chg is passed to indicate whether or not a page must be taken
2119 * from the global free pool (global change). gbl_chg == 0 indicates
2120 * a reservation exists for the allocation.
2121 */
2122 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2123 if (!page) {
2124 spin_unlock(&hugetlb_lock);
2125 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2126 if (!page)
2127 goto out_uncharge_cgroup;
2128 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2129 SetPagePrivate(page);
2130 h->resv_huge_pages--;
2131 }
2132 spin_lock(&hugetlb_lock);
2133 list_move(&page->lru, &h->hugepage_activelist);
2134 /* Fall through */
2135 }
2136 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2137 spin_unlock(&hugetlb_lock);
2138
2139 set_page_private(page, (unsigned long)spool);
2140
2141 map_commit = vma_commit_reservation(h, vma, addr);
2142 if (unlikely(map_chg > map_commit)) {
2143 /*
2144 * The page was added to the reservation map between
2145 * vma_needs_reservation and vma_commit_reservation.
2146 * This indicates a race with hugetlb_reserve_pages.
2147 * Adjust for the subpool count incremented above AND
2148 * in hugetlb_reserve_pages for the same page. Also,
2149 * the reservation count added in hugetlb_reserve_pages
2150 * no longer applies.
2151 */
2152 long rsv_adjust;
2153
2154 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2155 hugetlb_acct_memory(h, -rsv_adjust);
2156 }
2157 return page;
2158
2159out_uncharge_cgroup:
2160 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2161out_subpool_put:
2162 if (map_chg || avoid_reserve)
2163 hugepage_subpool_put_pages(spool, 1);
2164 vma_end_reservation(h, vma, addr);
2165 return ERR_PTR(-ENOSPC);
2166}
2167
2168int alloc_bootmem_huge_page(struct hstate *h)
2169 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2170int __alloc_bootmem_huge_page(struct hstate *h)
2171{
2172 struct huge_bootmem_page *m;
2173 int nr_nodes, node;
2174
2175 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2176 void *addr;
2177
2178 addr = memblock_alloc_try_nid_raw(
2179 huge_page_size(h), huge_page_size(h),
2180 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2181 if (addr) {
2182 /*
2183 * Use the beginning of the huge page to store the
2184 * huge_bootmem_page struct (until gather_bootmem
2185 * puts them into the mem_map).
2186 */
2187 m = addr;
2188 goto found;
2189 }
2190 }
2191 return 0;
2192
2193found:
2194 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2195 /* Put them into a private list first because mem_map is not up yet */
2196 INIT_LIST_HEAD(&m->list);
2197 list_add(&m->list, &huge_boot_pages);
2198 m->hstate = h;
2199 return 1;
2200}
2201
2202static void __init prep_compound_huge_page(struct page *page,
2203 unsigned int order)
2204{
2205 if (unlikely(order > (MAX_ORDER - 1)))
2206 prep_compound_gigantic_page(page, order);
2207 else
2208 prep_compound_page(page, order);
2209}
2210
2211/* Put bootmem huge pages into the standard lists after mem_map is up */
2212static void __init gather_bootmem_prealloc(void)
2213{
2214 struct huge_bootmem_page *m;
2215
2216 list_for_each_entry(m, &huge_boot_pages, list) {
2217 struct page *page = virt_to_page(m);
2218 struct hstate *h = m->hstate;
2219
2220 WARN_ON(page_count(page) != 1);
2221 prep_compound_huge_page(page, h->order);
2222 WARN_ON(PageReserved(page));
2223 prep_new_huge_page(h, page, page_to_nid(page));
2224 put_page(page); /* free it into the hugepage allocator */
2225
2226 /*
2227 * If we had gigantic hugepages allocated at boot time, we need
2228 * to restore the 'stolen' pages to totalram_pages in order to
2229 * fix confusing memory reports from free(1) and another
2230 * side-effects, like CommitLimit going negative.
2231 */
2232 if (hstate_is_gigantic(h))
2233 adjust_managed_page_count(page, 1 << h->order);
2234 cond_resched();
2235 }
2236}
2237
2238static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2239{
2240 unsigned long i;
2241 nodemask_t *node_alloc_noretry;
2242
2243 if (!hstate_is_gigantic(h)) {
2244 /*
2245 * Bit mask controlling how hard we retry per-node allocations.
2246 * Ignore errors as lower level routines can deal with
2247 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2248 * time, we are likely in bigger trouble.
2249 */
2250 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2251 GFP_KERNEL);
2252 } else {
2253 /* allocations done at boot time */
2254 node_alloc_noretry = NULL;
2255 }
2256
2257 /* bit mask controlling how hard we retry per-node allocations */
2258 if (node_alloc_noretry)
2259 nodes_clear(*node_alloc_noretry);
2260
2261 for (i = 0; i < h->max_huge_pages; ++i) {
2262 if (hstate_is_gigantic(h)) {
2263 if (!alloc_bootmem_huge_page(h))
2264 break;
2265 } else if (!alloc_pool_huge_page(h,
2266 &node_states[N_MEMORY],
2267 node_alloc_noretry))
2268 break;
2269 cond_resched();
2270 }
2271 if (i < h->max_huge_pages) {
2272 char buf[32];
2273
2274 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2275 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2276 h->max_huge_pages, buf, i);
2277 h->max_huge_pages = i;
2278 }
2279
2280 kfree(node_alloc_noretry);
2281}
2282
2283static void __init hugetlb_init_hstates(void)
2284{
2285 struct hstate *h;
2286
2287 for_each_hstate(h) {
2288 if (minimum_order > huge_page_order(h))
2289 minimum_order = huge_page_order(h);
2290
2291 /* oversize hugepages were init'ed in early boot */
2292 if (!hstate_is_gigantic(h))
2293 hugetlb_hstate_alloc_pages(h);
2294 }
2295 VM_BUG_ON(minimum_order == UINT_MAX);
2296}
2297
2298static void __init report_hugepages(void)
2299{
2300 struct hstate *h;
2301
2302 for_each_hstate(h) {
2303 char buf[32];
2304
2305 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2306 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2307 buf, h->free_huge_pages);
2308 }
2309}
2310
2311#ifdef CONFIG_HIGHMEM
2312static void try_to_free_low(struct hstate *h, unsigned long count,
2313 nodemask_t *nodes_allowed)
2314{
2315 int i;
2316
2317 if (hstate_is_gigantic(h))
2318 return;
2319
2320 for_each_node_mask(i, *nodes_allowed) {
2321 struct page *page, *next;
2322 struct list_head *freel = &h->hugepage_freelists[i];
2323 list_for_each_entry_safe(page, next, freel, lru) {
2324 if (count >= h->nr_huge_pages)
2325 return;
2326 if (PageHighMem(page))
2327 continue;
2328 list_del(&page->lru);
2329 update_and_free_page(h, page);
2330 h->free_huge_pages--;
2331 h->free_huge_pages_node[page_to_nid(page)]--;
2332 }
2333 }
2334}
2335#else
2336static inline void try_to_free_low(struct hstate *h, unsigned long count,
2337 nodemask_t *nodes_allowed)
2338{
2339}
2340#endif
2341
2342/*
2343 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2344 * balanced by operating on them in a round-robin fashion.
2345 * Returns 1 if an adjustment was made.
2346 */
2347static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2348 int delta)
2349{
2350 int nr_nodes, node;
2351
2352 VM_BUG_ON(delta != -1 && delta != 1);
2353
2354 if (delta < 0) {
2355 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2356 if (h->surplus_huge_pages_node[node])
2357 goto found;
2358 }
2359 } else {
2360 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2361 if (h->surplus_huge_pages_node[node] <
2362 h->nr_huge_pages_node[node])
2363 goto found;
2364 }
2365 }
2366 return 0;
2367
2368found:
2369 h->surplus_huge_pages += delta;
2370 h->surplus_huge_pages_node[node] += delta;
2371 return 1;
2372}
2373
2374#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2375static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2376 nodemask_t *nodes_allowed)
2377{
2378 unsigned long min_count, ret;
2379 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2380
2381 /*
2382 * Bit mask controlling how hard we retry per-node allocations.
2383 * If we can not allocate the bit mask, do not attempt to allocate
2384 * the requested huge pages.
2385 */
2386 if (node_alloc_noretry)
2387 nodes_clear(*node_alloc_noretry);
2388 else
2389 return -ENOMEM;
2390
2391 spin_lock(&hugetlb_lock);
2392
2393 /*
2394 * Check for a node specific request.
2395 * Changing node specific huge page count may require a corresponding
2396 * change to the global count. In any case, the passed node mask
2397 * (nodes_allowed) will restrict alloc/free to the specified node.
2398 */
2399 if (nid != NUMA_NO_NODE) {
2400 unsigned long old_count = count;
2401
2402 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2403 /*
2404 * User may have specified a large count value which caused the
2405 * above calculation to overflow. In this case, they wanted
2406 * to allocate as many huge pages as possible. Set count to
2407 * largest possible value to align with their intention.
2408 */
2409 if (count < old_count)
2410 count = ULONG_MAX;
2411 }
2412
2413 /*
2414 * Gigantic pages runtime allocation depend on the capability for large
2415 * page range allocation.
2416 * If the system does not provide this feature, return an error when
2417 * the user tries to allocate gigantic pages but let the user free the
2418 * boottime allocated gigantic pages.
2419 */
2420 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2421 if (count > persistent_huge_pages(h)) {
2422 spin_unlock(&hugetlb_lock);
2423 NODEMASK_FREE(node_alloc_noretry);
2424 return -EINVAL;
2425 }
2426 /* Fall through to decrease pool */
2427 }
2428
2429 /*
2430 * Increase the pool size
2431 * First take pages out of surplus state. Then make up the
2432 * remaining difference by allocating fresh huge pages.
2433 *
2434 * We might race with alloc_surplus_huge_page() here and be unable
2435 * to convert a surplus huge page to a normal huge page. That is
2436 * not critical, though, it just means the overall size of the
2437 * pool might be one hugepage larger than it needs to be, but
2438 * within all the constraints specified by the sysctls.
2439 */
2440 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2441 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2442 break;
2443 }
2444
2445 while (count > persistent_huge_pages(h)) {
2446 /*
2447 * If this allocation races such that we no longer need the
2448 * page, free_huge_page will handle it by freeing the page
2449 * and reducing the surplus.
2450 */
2451 spin_unlock(&hugetlb_lock);
2452
2453 /* yield cpu to avoid soft lockup */
2454 cond_resched();
2455
2456 ret = alloc_pool_huge_page(h, nodes_allowed,
2457 node_alloc_noretry);
2458 spin_lock(&hugetlb_lock);
2459 if (!ret)
2460 goto out;
2461
2462 /* Bail for signals. Probably ctrl-c from user */
2463 if (signal_pending(current))
2464 goto out;
2465 }
2466
2467 /*
2468 * Decrease the pool size
2469 * First return free pages to the buddy allocator (being careful
2470 * to keep enough around to satisfy reservations). Then place
2471 * pages into surplus state as needed so the pool will shrink
2472 * to the desired size as pages become free.
2473 *
2474 * By placing pages into the surplus state independent of the
2475 * overcommit value, we are allowing the surplus pool size to
2476 * exceed overcommit. There are few sane options here. Since
2477 * alloc_surplus_huge_page() is checking the global counter,
2478 * though, we'll note that we're not allowed to exceed surplus
2479 * and won't grow the pool anywhere else. Not until one of the
2480 * sysctls are changed, or the surplus pages go out of use.
2481 */
2482 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2483 min_count = max(count, min_count);
2484 try_to_free_low(h, min_count, nodes_allowed);
2485 while (min_count < persistent_huge_pages(h)) {
2486 if (!free_pool_huge_page(h, nodes_allowed, 0))
2487 break;
2488 cond_resched_lock(&hugetlb_lock);
2489 }
2490 while (count < persistent_huge_pages(h)) {
2491 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2492 break;
2493 }
2494out:
2495 h->max_huge_pages = persistent_huge_pages(h);
2496 spin_unlock(&hugetlb_lock);
2497
2498 NODEMASK_FREE(node_alloc_noretry);
2499
2500 return 0;
2501}
2502
2503#define HSTATE_ATTR_RO(_name) \
2504 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2505
2506#define HSTATE_ATTR(_name) \
2507 static struct kobj_attribute _name##_attr = \
2508 __ATTR(_name, 0644, _name##_show, _name##_store)
2509
2510static struct kobject *hugepages_kobj;
2511static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2512
2513static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2514
2515static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2516{
2517 int i;
2518
2519 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2520 if (hstate_kobjs[i] == kobj) {
2521 if (nidp)
2522 *nidp = NUMA_NO_NODE;
2523 return &hstates[i];
2524 }
2525
2526 return kobj_to_node_hstate(kobj, nidp);
2527}
2528
2529static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2530 struct kobj_attribute *attr, char *buf)
2531{
2532 struct hstate *h;
2533 unsigned long nr_huge_pages;
2534 int nid;
2535
2536 h = kobj_to_hstate(kobj, &nid);
2537 if (nid == NUMA_NO_NODE)
2538 nr_huge_pages = h->nr_huge_pages;
2539 else
2540 nr_huge_pages = h->nr_huge_pages_node[nid];
2541
2542 return sprintf(buf, "%lu\n", nr_huge_pages);
2543}
2544
2545static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2546 struct hstate *h, int nid,
2547 unsigned long count, size_t len)
2548{
2549 int err;
2550 nodemask_t nodes_allowed, *n_mask;
2551
2552 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2553 return -EINVAL;
2554
2555 if (nid == NUMA_NO_NODE) {
2556 /*
2557 * global hstate attribute
2558 */
2559 if (!(obey_mempolicy &&
2560 init_nodemask_of_mempolicy(&nodes_allowed)))
2561 n_mask = &node_states[N_MEMORY];
2562 else
2563 n_mask = &nodes_allowed;
2564 } else {
2565 /*
2566 * Node specific request. count adjustment happens in
2567 * set_max_huge_pages() after acquiring hugetlb_lock.
2568 */
2569 init_nodemask_of_node(&nodes_allowed, nid);
2570 n_mask = &nodes_allowed;
2571 }
2572
2573 err = set_max_huge_pages(h, count, nid, n_mask);
2574
2575 return err ? err : len;
2576}
2577
2578static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2579 struct kobject *kobj, const char *buf,
2580 size_t len)
2581{
2582 struct hstate *h;
2583 unsigned long count;
2584 int nid;
2585 int err;
2586
2587 err = kstrtoul(buf, 10, &count);
2588 if (err)
2589 return err;
2590
2591 h = kobj_to_hstate(kobj, &nid);
2592 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2593}
2594
2595static ssize_t nr_hugepages_show(struct kobject *kobj,
2596 struct kobj_attribute *attr, char *buf)
2597{
2598 return nr_hugepages_show_common(kobj, attr, buf);
2599}
2600
2601static ssize_t nr_hugepages_store(struct kobject *kobj,
2602 struct kobj_attribute *attr, const char *buf, size_t len)
2603{
2604 return nr_hugepages_store_common(false, kobj, buf, len);
2605}
2606HSTATE_ATTR(nr_hugepages);
2607
2608#ifdef CONFIG_NUMA
2609
2610/*
2611 * hstate attribute for optionally mempolicy-based constraint on persistent
2612 * huge page alloc/free.
2613 */
2614static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2615 struct kobj_attribute *attr, char *buf)
2616{
2617 return nr_hugepages_show_common(kobj, attr, buf);
2618}
2619
2620static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2621 struct kobj_attribute *attr, const char *buf, size_t len)
2622{
2623 return nr_hugepages_store_common(true, kobj, buf, len);
2624}
2625HSTATE_ATTR(nr_hugepages_mempolicy);
2626#endif
2627
2628
2629static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2630 struct kobj_attribute *attr, char *buf)
2631{
2632 struct hstate *h = kobj_to_hstate(kobj, NULL);
2633 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2634}
2635
2636static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2637 struct kobj_attribute *attr, const char *buf, size_t count)
2638{
2639 int err;
2640 unsigned long input;
2641 struct hstate *h = kobj_to_hstate(kobj, NULL);
2642
2643 if (hstate_is_gigantic(h))
2644 return -EINVAL;
2645
2646 err = kstrtoul(buf, 10, &input);
2647 if (err)
2648 return err;
2649
2650 spin_lock(&hugetlb_lock);
2651 h->nr_overcommit_huge_pages = input;
2652 spin_unlock(&hugetlb_lock);
2653
2654 return count;
2655}
2656HSTATE_ATTR(nr_overcommit_hugepages);
2657
2658static ssize_t free_hugepages_show(struct kobject *kobj,
2659 struct kobj_attribute *attr, char *buf)
2660{
2661 struct hstate *h;
2662 unsigned long free_huge_pages;
2663 int nid;
2664
2665 h = kobj_to_hstate(kobj, &nid);
2666 if (nid == NUMA_NO_NODE)
2667 free_huge_pages = h->free_huge_pages;
2668 else
2669 free_huge_pages = h->free_huge_pages_node[nid];
2670
2671 return sprintf(buf, "%lu\n", free_huge_pages);
2672}
2673HSTATE_ATTR_RO(free_hugepages);
2674
2675static ssize_t resv_hugepages_show(struct kobject *kobj,
2676 struct kobj_attribute *attr, char *buf)
2677{
2678 struct hstate *h = kobj_to_hstate(kobj, NULL);
2679 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2680}
2681HSTATE_ATTR_RO(resv_hugepages);
2682
2683static ssize_t surplus_hugepages_show(struct kobject *kobj,
2684 struct kobj_attribute *attr, char *buf)
2685{
2686 struct hstate *h;
2687 unsigned long surplus_huge_pages;
2688 int nid;
2689
2690 h = kobj_to_hstate(kobj, &nid);
2691 if (nid == NUMA_NO_NODE)
2692 surplus_huge_pages = h->surplus_huge_pages;
2693 else
2694 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2695
2696 return sprintf(buf, "%lu\n", surplus_huge_pages);
2697}
2698HSTATE_ATTR_RO(surplus_hugepages);
2699
2700static struct attribute *hstate_attrs[] = {
2701 &nr_hugepages_attr.attr,
2702 &nr_overcommit_hugepages_attr.attr,
2703 &free_hugepages_attr.attr,
2704 &resv_hugepages_attr.attr,
2705 &surplus_hugepages_attr.attr,
2706#ifdef CONFIG_NUMA
2707 &nr_hugepages_mempolicy_attr.attr,
2708#endif
2709 NULL,
2710};
2711
2712static const struct attribute_group hstate_attr_group = {
2713 .attrs = hstate_attrs,
2714};
2715
2716static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2717 struct kobject **hstate_kobjs,
2718 const struct attribute_group *hstate_attr_group)
2719{
2720 int retval;
2721 int hi = hstate_index(h);
2722
2723 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2724 if (!hstate_kobjs[hi])
2725 return -ENOMEM;
2726
2727 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2728 if (retval)
2729 kobject_put(hstate_kobjs[hi]);
2730
2731 return retval;
2732}
2733
2734static void __init hugetlb_sysfs_init(void)
2735{
2736 struct hstate *h;
2737 int err;
2738
2739 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2740 if (!hugepages_kobj)
2741 return;
2742
2743 for_each_hstate(h) {
2744 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2745 hstate_kobjs, &hstate_attr_group);
2746 if (err)
2747 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2748 }
2749}
2750
2751#ifdef CONFIG_NUMA
2752
2753/*
2754 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2755 * with node devices in node_devices[] using a parallel array. The array
2756 * index of a node device or _hstate == node id.
2757 * This is here to avoid any static dependency of the node device driver, in
2758 * the base kernel, on the hugetlb module.
2759 */
2760struct node_hstate {
2761 struct kobject *hugepages_kobj;
2762 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2763};
2764static struct node_hstate node_hstates[MAX_NUMNODES];
2765
2766/*
2767 * A subset of global hstate attributes for node devices
2768 */
2769static struct attribute *per_node_hstate_attrs[] = {
2770 &nr_hugepages_attr.attr,
2771 &free_hugepages_attr.attr,
2772 &surplus_hugepages_attr.attr,
2773 NULL,
2774};
2775
2776static const struct attribute_group per_node_hstate_attr_group = {
2777 .attrs = per_node_hstate_attrs,
2778};
2779
2780/*
2781 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2782 * Returns node id via non-NULL nidp.
2783 */
2784static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2785{
2786 int nid;
2787
2788 for (nid = 0; nid < nr_node_ids; nid++) {
2789 struct node_hstate *nhs = &node_hstates[nid];
2790 int i;
2791 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2792 if (nhs->hstate_kobjs[i] == kobj) {
2793 if (nidp)
2794 *nidp = nid;
2795 return &hstates[i];
2796 }
2797 }
2798
2799 BUG();
2800 return NULL;
2801}
2802
2803/*
2804 * Unregister hstate attributes from a single node device.
2805 * No-op if no hstate attributes attached.
2806 */
2807static void hugetlb_unregister_node(struct node *node)
2808{
2809 struct hstate *h;
2810 struct node_hstate *nhs = &node_hstates[node->dev.id];
2811
2812 if (!nhs->hugepages_kobj)
2813 return; /* no hstate attributes */
2814
2815 for_each_hstate(h) {
2816 int idx = hstate_index(h);
2817 if (nhs->hstate_kobjs[idx]) {
2818 kobject_put(nhs->hstate_kobjs[idx]);
2819 nhs->hstate_kobjs[idx] = NULL;
2820 }
2821 }
2822
2823 kobject_put(nhs->hugepages_kobj);
2824 nhs->hugepages_kobj = NULL;
2825}
2826
2827
2828/*
2829 * Register hstate attributes for a single node device.
2830 * No-op if attributes already registered.
2831 */
2832static void hugetlb_register_node(struct node *node)
2833{
2834 struct hstate *h;
2835 struct node_hstate *nhs = &node_hstates[node->dev.id];
2836 int err;
2837
2838 if (nhs->hugepages_kobj)
2839 return; /* already allocated */
2840
2841 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2842 &node->dev.kobj);
2843 if (!nhs->hugepages_kobj)
2844 return;
2845
2846 for_each_hstate(h) {
2847 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2848 nhs->hstate_kobjs,
2849 &per_node_hstate_attr_group);
2850 if (err) {
2851 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2852 h->name, node->dev.id);
2853 hugetlb_unregister_node(node);
2854 break;
2855 }
2856 }
2857}
2858
2859/*
2860 * hugetlb init time: register hstate attributes for all registered node
2861 * devices of nodes that have memory. All on-line nodes should have
2862 * registered their associated device by this time.
2863 */
2864static void __init hugetlb_register_all_nodes(void)
2865{
2866 int nid;
2867
2868 for_each_node_state(nid, N_MEMORY) {
2869 struct node *node = node_devices[nid];
2870 if (node->dev.id == nid)
2871 hugetlb_register_node(node);
2872 }
2873
2874 /*
2875 * Let the node device driver know we're here so it can
2876 * [un]register hstate attributes on node hotplug.
2877 */
2878 register_hugetlbfs_with_node(hugetlb_register_node,
2879 hugetlb_unregister_node);
2880}
2881#else /* !CONFIG_NUMA */
2882
2883static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2884{
2885 BUG();
2886 if (nidp)
2887 *nidp = -1;
2888 return NULL;
2889}
2890
2891static void hugetlb_register_all_nodes(void) { }
2892
2893#endif
2894
2895static int __init hugetlb_init(void)
2896{
2897 int i;
2898
2899 if (!hugepages_supported())
2900 return 0;
2901
2902 if (!size_to_hstate(default_hstate_size)) {
2903 if (default_hstate_size != 0) {
2904 pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2905 default_hstate_size, HPAGE_SIZE);
2906 }
2907
2908 default_hstate_size = HPAGE_SIZE;
2909 if (!size_to_hstate(default_hstate_size))
2910 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2911 }
2912 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2913 if (default_hstate_max_huge_pages) {
2914 if (!default_hstate.max_huge_pages)
2915 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2916 }
2917
2918 hugetlb_init_hstates();
2919 gather_bootmem_prealloc();
2920 report_hugepages();
2921
2922 hugetlb_sysfs_init();
2923 hugetlb_register_all_nodes();
2924 hugetlb_cgroup_file_init();
2925
2926#ifdef CONFIG_SMP
2927 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2928#else
2929 num_fault_mutexes = 1;
2930#endif
2931 hugetlb_fault_mutex_table =
2932 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
2933 GFP_KERNEL);
2934 BUG_ON(!hugetlb_fault_mutex_table);
2935
2936 for (i = 0; i < num_fault_mutexes; i++)
2937 mutex_init(&hugetlb_fault_mutex_table[i]);
2938 return 0;
2939}
2940subsys_initcall(hugetlb_init);
2941
2942/* Should be called on processing a hugepagesz=... option */
2943void __init hugetlb_bad_size(void)
2944{
2945 parsed_valid_hugepagesz = false;
2946}
2947
2948void __init hugetlb_add_hstate(unsigned int order)
2949{
2950 struct hstate *h;
2951 unsigned long i;
2952
2953 if (size_to_hstate(PAGE_SIZE << order)) {
2954 pr_warn("hugepagesz= specified twice, ignoring\n");
2955 return;
2956 }
2957 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2958 BUG_ON(order == 0);
2959 h = &hstates[hugetlb_max_hstate++];
2960 h->order = order;
2961 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2962 h->nr_huge_pages = 0;
2963 h->free_huge_pages = 0;
2964 for (i = 0; i < MAX_NUMNODES; ++i)
2965 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2966 INIT_LIST_HEAD(&h->hugepage_activelist);
2967 h->next_nid_to_alloc = first_memory_node;
2968 h->next_nid_to_free = first_memory_node;
2969 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2970 huge_page_size(h)/1024);
2971
2972 parsed_hstate = h;
2973}
2974
2975static int __init hugetlb_nrpages_setup(char *s)
2976{
2977 unsigned long *mhp;
2978 static unsigned long *last_mhp;
2979
2980 if (!parsed_valid_hugepagesz) {
2981 pr_warn("hugepages = %s preceded by "
2982 "an unsupported hugepagesz, ignoring\n", s);
2983 parsed_valid_hugepagesz = true;
2984 return 1;
2985 }
2986 /*
2987 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2988 * so this hugepages= parameter goes to the "default hstate".
2989 */
2990 else if (!hugetlb_max_hstate)
2991 mhp = &default_hstate_max_huge_pages;
2992 else
2993 mhp = &parsed_hstate->max_huge_pages;
2994
2995 if (mhp == last_mhp) {
2996 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2997 return 1;
2998 }
2999
3000 if (sscanf(s, "%lu", mhp) <= 0)
3001 *mhp = 0;
3002
3003 /*
3004 * Global state is always initialized later in hugetlb_init.
3005 * But we need to allocate >= MAX_ORDER hstates here early to still
3006 * use the bootmem allocator.
3007 */
3008 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
3009 hugetlb_hstate_alloc_pages(parsed_hstate);
3010
3011 last_mhp = mhp;
3012
3013 return 1;
3014}
3015__setup("hugepages=", hugetlb_nrpages_setup);
3016
3017static int __init hugetlb_default_setup(char *s)
3018{
3019 default_hstate_size = memparse(s, &s);
3020 return 1;
3021}
3022__setup("default_hugepagesz=", hugetlb_default_setup);
3023
3024static unsigned int cpuset_mems_nr(unsigned int *array)
3025{
3026 int node;
3027 unsigned int nr = 0;
3028
3029 for_each_node_mask(node, cpuset_current_mems_allowed)
3030 nr += array[node];
3031
3032 return nr;
3033}
3034
3035#ifdef CONFIG_SYSCTL
3036static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3037 struct ctl_table *table, int write,
3038 void __user *buffer, size_t *length, loff_t *ppos)
3039{
3040 struct hstate *h = &default_hstate;
3041 unsigned long tmp = h->max_huge_pages;
3042 int ret;
3043
3044 if (!hugepages_supported())
3045 return -EOPNOTSUPP;
3046
3047 table->data = &tmp;
3048 table->maxlen = sizeof(unsigned long);
3049 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3050 if (ret)
3051 goto out;
3052
3053 if (write)
3054 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3055 NUMA_NO_NODE, tmp, *length);
3056out:
3057 return ret;
3058}
3059
3060int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3061 void __user *buffer, size_t *length, loff_t *ppos)
3062{
3063
3064 return hugetlb_sysctl_handler_common(false, table, write,
3065 buffer, length, ppos);
3066}
3067
3068#ifdef CONFIG_NUMA
3069int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3070 void __user *buffer, size_t *length, loff_t *ppos)
3071{
3072 return hugetlb_sysctl_handler_common(true, table, write,
3073 buffer, length, ppos);
3074}
3075#endif /* CONFIG_NUMA */
3076
3077int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3078 void __user *buffer,
3079 size_t *length, loff_t *ppos)
3080{
3081 struct hstate *h = &default_hstate;
3082 unsigned long tmp;
3083 int ret;
3084
3085 if (!hugepages_supported())
3086 return -EOPNOTSUPP;
3087
3088 tmp = h->nr_overcommit_huge_pages;
3089
3090 if (write && hstate_is_gigantic(h))
3091 return -EINVAL;
3092
3093 table->data = &tmp;
3094 table->maxlen = sizeof(unsigned long);
3095 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3096 if (ret)
3097 goto out;
3098
3099 if (write) {
3100 spin_lock(&hugetlb_lock);
3101 h->nr_overcommit_huge_pages = tmp;
3102 spin_unlock(&hugetlb_lock);
3103 }
3104out:
3105 return ret;
3106}
3107
3108#endif /* CONFIG_SYSCTL */
3109
3110void hugetlb_report_meminfo(struct seq_file *m)
3111{
3112 struct hstate *h;
3113 unsigned long total = 0;
3114
3115 if (!hugepages_supported())
3116 return;
3117
3118 for_each_hstate(h) {
3119 unsigned long count = h->nr_huge_pages;
3120
3121 total += (PAGE_SIZE << huge_page_order(h)) * count;
3122
3123 if (h == &default_hstate)
3124 seq_printf(m,
3125 "HugePages_Total: %5lu\n"
3126 "HugePages_Free: %5lu\n"
3127 "HugePages_Rsvd: %5lu\n"
3128 "HugePages_Surp: %5lu\n"
3129 "Hugepagesize: %8lu kB\n",
3130 count,
3131 h->free_huge_pages,
3132 h->resv_huge_pages,
3133 h->surplus_huge_pages,
3134 (PAGE_SIZE << huge_page_order(h)) / 1024);
3135 }
3136
3137 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3138}
3139
3140int hugetlb_report_node_meminfo(int nid, char *buf)
3141{
3142 struct hstate *h = &default_hstate;
3143 if (!hugepages_supported())
3144 return 0;
3145 return sprintf(buf,
3146 "Node %d HugePages_Total: %5u\n"
3147 "Node %d HugePages_Free: %5u\n"
3148 "Node %d HugePages_Surp: %5u\n",
3149 nid, h->nr_huge_pages_node[nid],
3150 nid, h->free_huge_pages_node[nid],
3151 nid, h->surplus_huge_pages_node[nid]);
3152}
3153
3154void hugetlb_show_meminfo(void)
3155{
3156 struct hstate *h;
3157 int nid;
3158
3159 if (!hugepages_supported())
3160 return;
3161
3162 for_each_node_state(nid, N_MEMORY)
3163 for_each_hstate(h)
3164 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3165 nid,
3166 h->nr_huge_pages_node[nid],
3167 h->free_huge_pages_node[nid],
3168 h->surplus_huge_pages_node[nid],
3169 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3170}
3171
3172void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3173{
3174 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3175 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3176}
3177
3178/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3179unsigned long hugetlb_total_pages(void)
3180{
3181 struct hstate *h;
3182 unsigned long nr_total_pages = 0;
3183
3184 for_each_hstate(h)
3185 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3186 return nr_total_pages;
3187}
3188
3189static int hugetlb_acct_memory(struct hstate *h, long delta)
3190{
3191 int ret = -ENOMEM;
3192
3193 spin_lock(&hugetlb_lock);
3194 /*
3195 * When cpuset is configured, it breaks the strict hugetlb page
3196 * reservation as the accounting is done on a global variable. Such
3197 * reservation is completely rubbish in the presence of cpuset because
3198 * the reservation is not checked against page availability for the
3199 * current cpuset. Application can still potentially OOM'ed by kernel
3200 * with lack of free htlb page in cpuset that the task is in.
3201 * Attempt to enforce strict accounting with cpuset is almost
3202 * impossible (or too ugly) because cpuset is too fluid that
3203 * task or memory node can be dynamically moved between cpusets.
3204 *
3205 * The change of semantics for shared hugetlb mapping with cpuset is
3206 * undesirable. However, in order to preserve some of the semantics,
3207 * we fall back to check against current free page availability as
3208 * a best attempt and hopefully to minimize the impact of changing
3209 * semantics that cpuset has.
3210 */
3211 if (delta > 0) {
3212 if (gather_surplus_pages(h, delta) < 0)
3213 goto out;
3214
3215 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3216 return_unused_surplus_pages(h, delta);
3217 goto out;
3218 }
3219 }
3220
3221 ret = 0;
3222 if (delta < 0)
3223 return_unused_surplus_pages(h, (unsigned long) -delta);
3224
3225out:
3226 spin_unlock(&hugetlb_lock);
3227 return ret;
3228}
3229
3230static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3231{
3232 struct resv_map *resv = vma_resv_map(vma);
3233
3234 /*
3235 * This new VMA should share its siblings reservation map if present.
3236 * The VMA will only ever have a valid reservation map pointer where
3237 * it is being copied for another still existing VMA. As that VMA
3238 * has a reference to the reservation map it cannot disappear until
3239 * after this open call completes. It is therefore safe to take a
3240 * new reference here without additional locking.
3241 */
3242 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3243 kref_get(&resv->refs);
3244}
3245
3246static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3247{
3248 struct hstate *h = hstate_vma(vma);
3249 struct resv_map *resv = vma_resv_map(vma);
3250 struct hugepage_subpool *spool = subpool_vma(vma);
3251 unsigned long reserve, start, end;
3252 long gbl_reserve;
3253
3254 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3255 return;
3256
3257 start = vma_hugecache_offset(h, vma, vma->vm_start);
3258 end = vma_hugecache_offset(h, vma, vma->vm_end);
3259
3260 reserve = (end - start) - region_count(resv, start, end);
3261
3262 kref_put(&resv->refs, resv_map_release);
3263
3264 if (reserve) {
3265 /*
3266 * Decrement reserve counts. The global reserve count may be
3267 * adjusted if the subpool has a minimum size.
3268 */
3269 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3270 hugetlb_acct_memory(h, -gbl_reserve);
3271 }
3272}
3273
3274static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3275{
3276 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3277 return -EINVAL;
3278 return 0;
3279}
3280
3281static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3282{
3283 struct hstate *hstate = hstate_vma(vma);
3284
3285 return 1UL << huge_page_shift(hstate);
3286}
3287
3288/*
3289 * We cannot handle pagefaults against hugetlb pages at all. They cause
3290 * handle_mm_fault() to try to instantiate regular-sized pages in the
3291 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3292 * this far.
3293 */
3294static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3295{
3296 BUG();
3297 return 0;
3298}
3299
3300/*
3301 * When a new function is introduced to vm_operations_struct and added
3302 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3303 * This is because under System V memory model, mappings created via
3304 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3305 * their original vm_ops are overwritten with shm_vm_ops.
3306 */
3307const struct vm_operations_struct hugetlb_vm_ops = {
3308 .fault = hugetlb_vm_op_fault,
3309 .open = hugetlb_vm_op_open,
3310 .close = hugetlb_vm_op_close,
3311 .split = hugetlb_vm_op_split,
3312 .pagesize = hugetlb_vm_op_pagesize,
3313};
3314
3315static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3316 int writable)
3317{
3318 pte_t entry;
3319
3320 if (writable) {
3321 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3322 vma->vm_page_prot)));
3323 } else {
3324 entry = huge_pte_wrprotect(mk_huge_pte(page,
3325 vma->vm_page_prot));
3326 }
3327 entry = pte_mkyoung(entry);
3328 entry = pte_mkhuge(entry);
3329 entry = arch_make_huge_pte(entry, vma, page, writable);
3330
3331 return entry;
3332}
3333
3334static void set_huge_ptep_writable(struct vm_area_struct *vma,
3335 unsigned long address, pte_t *ptep)
3336{
3337 pte_t entry;
3338
3339 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3340 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3341 update_mmu_cache(vma, address, ptep);
3342}
3343
3344bool is_hugetlb_entry_migration(pte_t pte)
3345{
3346 swp_entry_t swp;
3347
3348 if (huge_pte_none(pte) || pte_present(pte))
3349 return false;
3350 swp = pte_to_swp_entry(pte);
3351 if (non_swap_entry(swp) && is_migration_entry(swp))
3352 return true;
3353 else
3354 return false;
3355}
3356
3357static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3358{
3359 swp_entry_t swp;
3360
3361 if (huge_pte_none(pte) || pte_present(pte))
3362 return 0;
3363 swp = pte_to_swp_entry(pte);
3364 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3365 return 1;
3366 else
3367 return 0;
3368}
3369
3370int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3371 struct vm_area_struct *vma)
3372{
3373 pte_t *src_pte, *dst_pte, entry, dst_entry;
3374 struct page *ptepage;
3375 unsigned long addr;
3376 int cow;
3377 struct hstate *h = hstate_vma(vma);
3378 unsigned long sz = huge_page_size(h);
3379 struct mmu_notifier_range range;
3380 int ret = 0;
3381
3382 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3383
3384 if (cow) {
3385 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3386 vma->vm_start,
3387 vma->vm_end);
3388 mmu_notifier_invalidate_range_start(&range);
3389 }
3390
3391 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3392 spinlock_t *src_ptl, *dst_ptl;
3393 src_pte = huge_pte_offset(src, addr, sz);
3394 if (!src_pte)
3395 continue;
3396 dst_pte = huge_pte_alloc(dst, addr, sz);
3397 if (!dst_pte) {
3398 ret = -ENOMEM;
3399 break;
3400 }
3401
3402 /*
3403 * If the pagetables are shared don't copy or take references.
3404 * dst_pte == src_pte is the common case of src/dest sharing.
3405 *
3406 * However, src could have 'unshared' and dst shares with
3407 * another vma. If dst_pte !none, this implies sharing.
3408 * Check here before taking page table lock, and once again
3409 * after taking the lock below.
3410 */
3411 dst_entry = huge_ptep_get(dst_pte);
3412 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3413 continue;
3414
3415 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3416 src_ptl = huge_pte_lockptr(h, src, src_pte);
3417 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3418 entry = huge_ptep_get(src_pte);
3419 dst_entry = huge_ptep_get(dst_pte);
3420 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3421 /*
3422 * Skip if src entry none. Also, skip in the
3423 * unlikely case dst entry !none as this implies
3424 * sharing with another vma.
3425 */
3426 ;
3427 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3428 is_hugetlb_entry_hwpoisoned(entry))) {
3429 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3430
3431 if (is_write_migration_entry(swp_entry) && cow) {
3432 /*
3433 * COW mappings require pages in both
3434 * parent and child to be set to read.
3435 */
3436 make_migration_entry_read(&swp_entry);
3437 entry = swp_entry_to_pte(swp_entry);
3438 set_huge_swap_pte_at(src, addr, src_pte,
3439 entry, sz);
3440 }
3441 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3442 } else {
3443 if (cow) {
3444 /*
3445 * No need to notify as we are downgrading page
3446 * table protection not changing it to point
3447 * to a new page.
3448 *
3449 * See Documentation/vm/mmu_notifier.rst
3450 */
3451 huge_ptep_set_wrprotect(src, addr, src_pte);
3452 }
3453 entry = huge_ptep_get(src_pte);
3454 ptepage = pte_page(entry);
3455 get_page(ptepage);
3456 page_dup_rmap(ptepage, true);
3457 set_huge_pte_at(dst, addr, dst_pte, entry);
3458 hugetlb_count_add(pages_per_huge_page(h), dst);
3459 }
3460 spin_unlock(src_ptl);
3461 spin_unlock(dst_ptl);
3462 }
3463
3464 if (cow)
3465 mmu_notifier_invalidate_range_end(&range);
3466
3467 return ret;
3468}
3469
3470void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3471 unsigned long start, unsigned long end,
3472 struct page *ref_page)
3473{
3474 struct mm_struct *mm = vma->vm_mm;
3475 unsigned long address;
3476 pte_t *ptep;
3477 pte_t pte;
3478 spinlock_t *ptl;
3479 struct page *page;
3480 struct hstate *h = hstate_vma(vma);
3481 unsigned long sz = huge_page_size(h);
3482 struct mmu_notifier_range range;
3483
3484 WARN_ON(!is_vm_hugetlb_page(vma));
3485 BUG_ON(start & ~huge_page_mask(h));
3486 BUG_ON(end & ~huge_page_mask(h));
3487
3488 /*
3489 * This is a hugetlb vma, all the pte entries should point
3490 * to huge page.
3491 */
3492 tlb_change_page_size(tlb, sz);
3493 tlb_start_vma(tlb, vma);
3494
3495 /*
3496 * If sharing possible, alert mmu notifiers of worst case.
3497 */
3498 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3499 end);
3500 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3501 mmu_notifier_invalidate_range_start(&range);
3502 address = start;
3503 for (; address < end; address += sz) {
3504 ptep = huge_pte_offset(mm, address, sz);
3505 if (!ptep)
3506 continue;
3507
3508 ptl = huge_pte_lock(h, mm, ptep);
3509 if (huge_pmd_unshare(mm, &address, ptep)) {
3510 spin_unlock(ptl);
3511 /*
3512 * We just unmapped a page of PMDs by clearing a PUD.
3513 * The caller's TLB flush range should cover this area.
3514 */
3515 continue;
3516 }
3517
3518 pte = huge_ptep_get(ptep);
3519 if (huge_pte_none(pte)) {
3520 spin_unlock(ptl);
3521 continue;
3522 }
3523
3524 /*
3525 * Migrating hugepage or HWPoisoned hugepage is already
3526 * unmapped and its refcount is dropped, so just clear pte here.
3527 */
3528 if (unlikely(!pte_present(pte))) {
3529 huge_pte_clear(mm, address, ptep, sz);
3530 spin_unlock(ptl);
3531 continue;
3532 }
3533
3534 page = pte_page(pte);
3535 /*
3536 * If a reference page is supplied, it is because a specific
3537 * page is being unmapped, not a range. Ensure the page we
3538 * are about to unmap is the actual page of interest.
3539 */
3540 if (ref_page) {
3541 if (page != ref_page) {
3542 spin_unlock(ptl);
3543 continue;
3544 }
3545 /*
3546 * Mark the VMA as having unmapped its page so that
3547 * future faults in this VMA will fail rather than
3548 * looking like data was lost
3549 */
3550 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3551 }
3552
3553 pte = huge_ptep_get_and_clear(mm, address, ptep);
3554 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3555 if (huge_pte_dirty(pte))
3556 set_page_dirty(page);
3557
3558 hugetlb_count_sub(pages_per_huge_page(h), mm);
3559 page_remove_rmap(page, true);
3560
3561 spin_unlock(ptl);
3562 tlb_remove_page_size(tlb, page, huge_page_size(h));
3563 /*
3564 * Bail out after unmapping reference page if supplied
3565 */
3566 if (ref_page)
3567 break;
3568 }
3569 mmu_notifier_invalidate_range_end(&range);
3570 tlb_end_vma(tlb, vma);
3571}
3572
3573void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3574 struct vm_area_struct *vma, unsigned long start,
3575 unsigned long end, struct page *ref_page)
3576{
3577 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3578
3579 /*
3580 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3581 * test will fail on a vma being torn down, and not grab a page table
3582 * on its way out. We're lucky that the flag has such an appropriate
3583 * name, and can in fact be safely cleared here. We could clear it
3584 * before the __unmap_hugepage_range above, but all that's necessary
3585 * is to clear it before releasing the i_mmap_rwsem. This works
3586 * because in the context this is called, the VMA is about to be
3587 * destroyed and the i_mmap_rwsem is held.
3588 */
3589 vma->vm_flags &= ~VM_MAYSHARE;
3590}
3591
3592void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3593 unsigned long end, struct page *ref_page)
3594{
3595 struct mm_struct *mm;
3596 struct mmu_gather tlb;
3597 unsigned long tlb_start = start;
3598 unsigned long tlb_end = end;
3599
3600 /*
3601 * If shared PMDs were possibly used within this vma range, adjust
3602 * start/end for worst case tlb flushing.
3603 * Note that we can not be sure if PMDs are shared until we try to
3604 * unmap pages. However, we want to make sure TLB flushing covers
3605 * the largest possible range.
3606 */
3607 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
3608
3609 mm = vma->vm_mm;
3610
3611 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
3612 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3613 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
3614}
3615
3616/*
3617 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3618 * mappping it owns the reserve page for. The intention is to unmap the page
3619 * from other VMAs and let the children be SIGKILLed if they are faulting the
3620 * same region.
3621 */
3622static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3623 struct page *page, unsigned long address)
3624{
3625 struct hstate *h = hstate_vma(vma);
3626 struct vm_area_struct *iter_vma;
3627 struct address_space *mapping;
3628 pgoff_t pgoff;
3629
3630 /*
3631 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3632 * from page cache lookup which is in HPAGE_SIZE units.
3633 */
3634 address = address & huge_page_mask(h);
3635 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3636 vma->vm_pgoff;
3637 mapping = vma->vm_file->f_mapping;
3638
3639 /*
3640 * Take the mapping lock for the duration of the table walk. As
3641 * this mapping should be shared between all the VMAs,
3642 * __unmap_hugepage_range() is called as the lock is already held
3643 */
3644 i_mmap_lock_write(mapping);
3645 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3646 /* Do not unmap the current VMA */
3647 if (iter_vma == vma)
3648 continue;
3649
3650 /*
3651 * Shared VMAs have their own reserves and do not affect
3652 * MAP_PRIVATE accounting but it is possible that a shared
3653 * VMA is using the same page so check and skip such VMAs.
3654 */
3655 if (iter_vma->vm_flags & VM_MAYSHARE)
3656 continue;
3657
3658 /*
3659 * Unmap the page from other VMAs without their own reserves.
3660 * They get marked to be SIGKILLed if they fault in these
3661 * areas. This is because a future no-page fault on this VMA
3662 * could insert a zeroed page instead of the data existing
3663 * from the time of fork. This would look like data corruption
3664 */
3665 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3666 unmap_hugepage_range(iter_vma, address,
3667 address + huge_page_size(h), page);
3668 }
3669 i_mmap_unlock_write(mapping);
3670}
3671
3672/*
3673 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3674 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3675 * cannot race with other handlers or page migration.
3676 * Keep the pte_same checks anyway to make transition from the mutex easier.
3677 */
3678static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3679 unsigned long address, pte_t *ptep,
3680 struct page *pagecache_page, spinlock_t *ptl)
3681{
3682 pte_t pte;
3683 struct hstate *h = hstate_vma(vma);
3684 struct page *old_page, *new_page;
3685 int outside_reserve = 0;
3686 vm_fault_t ret = 0;
3687 unsigned long haddr = address & huge_page_mask(h);
3688 struct mmu_notifier_range range;
3689
3690 pte = huge_ptep_get(ptep);
3691 old_page = pte_page(pte);
3692
3693retry_avoidcopy:
3694 /* If no-one else is actually using this page, avoid the copy
3695 * and just make the page writable */
3696 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3697 page_move_anon_rmap(old_page, vma);
3698 set_huge_ptep_writable(vma, haddr, ptep);
3699 return 0;
3700 }
3701
3702 /*
3703 * If the process that created a MAP_PRIVATE mapping is about to
3704 * perform a COW due to a shared page count, attempt to satisfy
3705 * the allocation without using the existing reserves. The pagecache
3706 * page is used to determine if the reserve at this address was
3707 * consumed or not. If reserves were used, a partial faulted mapping
3708 * at the time of fork() could consume its reserves on COW instead
3709 * of the full address range.
3710 */
3711 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3712 old_page != pagecache_page)
3713 outside_reserve = 1;
3714
3715 get_page(old_page);
3716
3717 /*
3718 * Drop page table lock as buddy allocator may be called. It will
3719 * be acquired again before returning to the caller, as expected.
3720 */
3721 spin_unlock(ptl);
3722 new_page = alloc_huge_page(vma, haddr, outside_reserve);
3723
3724 if (IS_ERR(new_page)) {
3725 /*
3726 * If a process owning a MAP_PRIVATE mapping fails to COW,
3727 * it is due to references held by a child and an insufficient
3728 * huge page pool. To guarantee the original mappers
3729 * reliability, unmap the page from child processes. The child
3730 * may get SIGKILLed if it later faults.
3731 */
3732 if (outside_reserve) {
3733 put_page(old_page);
3734 BUG_ON(huge_pte_none(pte));
3735 unmap_ref_private(mm, vma, old_page, haddr);
3736 BUG_ON(huge_pte_none(pte));
3737 spin_lock(ptl);
3738 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3739 if (likely(ptep &&
3740 pte_same(huge_ptep_get(ptep), pte)))
3741 goto retry_avoidcopy;
3742 /*
3743 * race occurs while re-acquiring page table
3744 * lock, and our job is done.
3745 */
3746 return 0;
3747 }
3748
3749 ret = vmf_error(PTR_ERR(new_page));
3750 goto out_release_old;
3751 }
3752
3753 /*
3754 * When the original hugepage is shared one, it does not have
3755 * anon_vma prepared.
3756 */
3757 if (unlikely(anon_vma_prepare(vma))) {
3758 ret = VM_FAULT_OOM;
3759 goto out_release_all;
3760 }
3761
3762 copy_user_huge_page(new_page, old_page, address, vma,
3763 pages_per_huge_page(h));
3764 __SetPageUptodate(new_page);
3765
3766 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
3767 haddr + huge_page_size(h));
3768 mmu_notifier_invalidate_range_start(&range);
3769
3770 /*
3771 * Retake the page table lock to check for racing updates
3772 * before the page tables are altered
3773 */
3774 spin_lock(ptl);
3775 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3776 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3777 ClearPagePrivate(new_page);
3778
3779 /* Break COW */
3780 huge_ptep_clear_flush(vma, haddr, ptep);
3781 mmu_notifier_invalidate_range(mm, range.start, range.end);
3782 set_huge_pte_at(mm, haddr, ptep,
3783 make_huge_pte(vma, new_page, 1));
3784 page_remove_rmap(old_page, true);
3785 hugepage_add_new_anon_rmap(new_page, vma, haddr);
3786 set_page_huge_active(new_page);
3787 /* Make the old page be freed below */
3788 new_page = old_page;
3789 }
3790 spin_unlock(ptl);
3791 mmu_notifier_invalidate_range_end(&range);
3792out_release_all:
3793 restore_reserve_on_error(h, vma, haddr, new_page);
3794 put_page(new_page);
3795out_release_old:
3796 put_page(old_page);
3797
3798 spin_lock(ptl); /* Caller expects lock to be held */
3799 return ret;
3800}
3801
3802/* Return the pagecache page at a given address within a VMA */
3803static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3804 struct vm_area_struct *vma, unsigned long address)
3805{
3806 struct address_space *mapping;
3807 pgoff_t idx;
3808
3809 mapping = vma->vm_file->f_mapping;
3810 idx = vma_hugecache_offset(h, vma, address);
3811
3812 return find_lock_page(mapping, idx);
3813}
3814
3815/*
3816 * Return whether there is a pagecache page to back given address within VMA.
3817 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3818 */
3819static bool hugetlbfs_pagecache_present(struct hstate *h,
3820 struct vm_area_struct *vma, unsigned long address)
3821{
3822 struct address_space *mapping;
3823 pgoff_t idx;
3824 struct page *page;
3825
3826 mapping = vma->vm_file->f_mapping;
3827 idx = vma_hugecache_offset(h, vma, address);
3828
3829 page = find_get_page(mapping, idx);
3830 if (page)
3831 put_page(page);
3832 return page != NULL;
3833}
3834
3835int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3836 pgoff_t idx)
3837{
3838 struct inode *inode = mapping->host;
3839 struct hstate *h = hstate_inode(inode);
3840 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3841
3842 if (err)
3843 return err;
3844 ClearPagePrivate(page);
3845
3846 /*
3847 * set page dirty so that it will not be removed from cache/file
3848 * by non-hugetlbfs specific code paths.
3849 */
3850 set_page_dirty(page);
3851
3852 spin_lock(&inode->i_lock);
3853 inode->i_blocks += blocks_per_huge_page(h);
3854 spin_unlock(&inode->i_lock);
3855 return 0;
3856}
3857
3858static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
3859 struct vm_area_struct *vma,
3860 struct address_space *mapping, pgoff_t idx,
3861 unsigned long address, pte_t *ptep, unsigned int flags)
3862{
3863 struct hstate *h = hstate_vma(vma);
3864 vm_fault_t ret = VM_FAULT_SIGBUS;
3865 int anon_rmap = 0;
3866 unsigned long size;
3867 struct page *page;
3868 pte_t new_pte;
3869 spinlock_t *ptl;
3870 unsigned long haddr = address & huge_page_mask(h);
3871 bool new_page = false;
3872
3873 /*
3874 * Currently, we are forced to kill the process in the event the
3875 * original mapper has unmapped pages from the child due to a failed
3876 * COW. Warn that such a situation has occurred as it may not be obvious
3877 */
3878 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3879 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3880 current->pid);
3881 return ret;
3882 }
3883
3884 /*
3885 * Use page lock to guard against racing truncation
3886 * before we get page_table_lock.
3887 */
3888retry:
3889 page = find_lock_page(mapping, idx);
3890 if (!page) {
3891 size = i_size_read(mapping->host) >> huge_page_shift(h);
3892 if (idx >= size)
3893 goto out;
3894
3895 /*
3896 * Check for page in userfault range
3897 */
3898 if (userfaultfd_missing(vma)) {
3899 u32 hash;
3900 struct vm_fault vmf = {
3901 .vma = vma,
3902 .address = haddr,
3903 .flags = flags,
3904 /*
3905 * Hard to debug if it ends up being
3906 * used by a callee that assumes
3907 * something about the other
3908 * uninitialized fields... same as in
3909 * memory.c
3910 */
3911 };
3912
3913 /*
3914 * hugetlb_fault_mutex must be dropped before
3915 * handling userfault. Reacquire after handling
3916 * fault to make calling code simpler.
3917 */
3918 hash = hugetlb_fault_mutex_hash(h, mapping, idx, haddr);
3919 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3920 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3921 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3922 goto out;
3923 }
3924
3925 page = alloc_huge_page(vma, haddr, 0);
3926 if (IS_ERR(page)) {
3927 /*
3928 * Returning error will result in faulting task being
3929 * sent SIGBUS. The hugetlb fault mutex prevents two
3930 * tasks from racing to fault in the same page which
3931 * could result in false unable to allocate errors.
3932 * Page migration does not take the fault mutex, but
3933 * does a clear then write of pte's under page table
3934 * lock. Page fault code could race with migration,
3935 * notice the clear pte and try to allocate a page
3936 * here. Before returning error, get ptl and make
3937 * sure there really is no pte entry.
3938 */
3939 ptl = huge_pte_lock(h, mm, ptep);
3940 if (!huge_pte_none(huge_ptep_get(ptep))) {
3941 ret = 0;
3942 spin_unlock(ptl);
3943 goto out;
3944 }
3945 spin_unlock(ptl);
3946 ret = vmf_error(PTR_ERR(page));
3947 goto out;
3948 }
3949 clear_huge_page(page, address, pages_per_huge_page(h));
3950 __SetPageUptodate(page);
3951 new_page = true;
3952
3953 if (vma->vm_flags & VM_MAYSHARE) {
3954 int err = huge_add_to_page_cache(page, mapping, idx);
3955 if (err) {
3956 put_page(page);
3957 if (err == -EEXIST)
3958 goto retry;
3959 goto out;
3960 }
3961 } else {
3962 lock_page(page);
3963 if (unlikely(anon_vma_prepare(vma))) {
3964 ret = VM_FAULT_OOM;
3965 goto backout_unlocked;
3966 }
3967 anon_rmap = 1;
3968 }
3969 } else {
3970 /*
3971 * If memory error occurs between mmap() and fault, some process
3972 * don't have hwpoisoned swap entry for errored virtual address.
3973 * So we need to block hugepage fault by PG_hwpoison bit check.
3974 */
3975 if (unlikely(PageHWPoison(page))) {
3976 ret = VM_FAULT_HWPOISON |
3977 VM_FAULT_SET_HINDEX(hstate_index(h));
3978 goto backout_unlocked;
3979 }
3980 }
3981
3982 /*
3983 * If we are going to COW a private mapping later, we examine the
3984 * pending reservations for this page now. This will ensure that
3985 * any allocations necessary to record that reservation occur outside
3986 * the spinlock.
3987 */
3988 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3989 if (vma_needs_reservation(h, vma, haddr) < 0) {
3990 ret = VM_FAULT_OOM;
3991 goto backout_unlocked;
3992 }
3993 /* Just decrements count, does not deallocate */
3994 vma_end_reservation(h, vma, haddr);
3995 }
3996
3997 ptl = huge_pte_lock(h, mm, ptep);
3998 size = i_size_read(mapping->host) >> huge_page_shift(h);
3999 if (idx >= size)
4000 goto backout;
4001
4002 ret = 0;
4003 if (!huge_pte_none(huge_ptep_get(ptep)))
4004 goto backout;
4005
4006 if (anon_rmap) {
4007 ClearPagePrivate(page);
4008 hugepage_add_new_anon_rmap(page, vma, haddr);
4009 } else
4010 page_dup_rmap(page, true);
4011 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4012 && (vma->vm_flags & VM_SHARED)));
4013 set_huge_pte_at(mm, haddr, ptep, new_pte);
4014
4015 hugetlb_count_add(pages_per_huge_page(h), mm);
4016 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4017 /* Optimization, do the COW without a second fault */
4018 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4019 }
4020
4021 spin_unlock(ptl);
4022
4023 /*
4024 * Only make newly allocated pages active. Existing pages found
4025 * in the pagecache could be !page_huge_active() if they have been
4026 * isolated for migration.
4027 */
4028 if (new_page)
4029 set_page_huge_active(page);
4030
4031 unlock_page(page);
4032out:
4033 return ret;
4034
4035backout:
4036 spin_unlock(ptl);
4037backout_unlocked:
4038 unlock_page(page);
4039 restore_reserve_on_error(h, vma, haddr, page);
4040 put_page(page);
4041 goto out;
4042}
4043
4044#ifdef CONFIG_SMP
4045u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping,
4046 pgoff_t idx, unsigned long address)
4047{
4048 unsigned long key[2];
4049 u32 hash;
4050
4051 key[0] = (unsigned long) mapping;
4052 key[1] = idx;
4053
4054 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
4055
4056 return hash & (num_fault_mutexes - 1);
4057}
4058#else
4059/*
4060 * For uniprocesor systems we always use a single mutex, so just
4061 * return 0 and avoid the hashing overhead.
4062 */
4063u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping,
4064 pgoff_t idx, unsigned long address)
4065{
4066 return 0;
4067}
4068#endif
4069
4070vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4071 unsigned long address, unsigned int flags)
4072{
4073 pte_t *ptep, entry;
4074 spinlock_t *ptl;
4075 vm_fault_t ret;
4076 u32 hash;
4077 pgoff_t idx;
4078 struct page *page = NULL;
4079 struct page *pagecache_page = NULL;
4080 struct hstate *h = hstate_vma(vma);
4081 struct address_space *mapping;
4082 int need_wait_lock = 0;
4083 unsigned long haddr = address & huge_page_mask(h);
4084
4085 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4086 if (ptep) {
4087 entry = huge_ptep_get(ptep);
4088 if (unlikely(is_hugetlb_entry_migration(entry))) {
4089 migration_entry_wait_huge(vma, mm, ptep);
4090 return 0;
4091 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4092 return VM_FAULT_HWPOISON_LARGE |
4093 VM_FAULT_SET_HINDEX(hstate_index(h));
4094 } else {
4095 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4096 if (!ptep)
4097 return VM_FAULT_OOM;
4098 }
4099
4100 mapping = vma->vm_file->f_mapping;
4101 idx = vma_hugecache_offset(h, vma, haddr);
4102
4103 /*
4104 * Serialize hugepage allocation and instantiation, so that we don't
4105 * get spurious allocation failures if two CPUs race to instantiate
4106 * the same page in the page cache.
4107 */
4108 hash = hugetlb_fault_mutex_hash(h, mapping, idx, haddr);
4109 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4110
4111 entry = huge_ptep_get(ptep);
4112 if (huge_pte_none(entry)) {
4113 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4114 goto out_mutex;
4115 }
4116
4117 ret = 0;
4118
4119 /*
4120 * entry could be a migration/hwpoison entry at this point, so this
4121 * check prevents the kernel from going below assuming that we have
4122 * a active hugepage in pagecache. This goto expects the 2nd page fault,
4123 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
4124 * handle it.
4125 */
4126 if (!pte_present(entry))
4127 goto out_mutex;
4128
4129 /*
4130 * If we are going to COW the mapping later, we examine the pending
4131 * reservations for this page now. This will ensure that any
4132 * allocations necessary to record that reservation occur outside the
4133 * spinlock. For private mappings, we also lookup the pagecache
4134 * page now as it is used to determine if a reservation has been
4135 * consumed.
4136 */
4137 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4138 if (vma_needs_reservation(h, vma, haddr) < 0) {
4139 ret = VM_FAULT_OOM;
4140 goto out_mutex;
4141 }
4142 /* Just decrements count, does not deallocate */
4143 vma_end_reservation(h, vma, haddr);
4144
4145 if (!(vma->vm_flags & VM_MAYSHARE))
4146 pagecache_page = hugetlbfs_pagecache_page(h,
4147 vma, haddr);
4148 }
4149
4150 ptl = huge_pte_lock(h, mm, ptep);
4151
4152 /* Check for a racing update before calling hugetlb_cow */
4153 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4154 goto out_ptl;
4155
4156 /*
4157 * hugetlb_cow() requires page locks of pte_page(entry) and
4158 * pagecache_page, so here we need take the former one
4159 * when page != pagecache_page or !pagecache_page.
4160 */
4161 page = pte_page(entry);
4162 if (page != pagecache_page)
4163 if (!trylock_page(page)) {
4164 need_wait_lock = 1;
4165 goto out_ptl;
4166 }
4167
4168 get_page(page);
4169
4170 if (flags & FAULT_FLAG_WRITE) {
4171 if (!huge_pte_write(entry)) {
4172 ret = hugetlb_cow(mm, vma, address, ptep,
4173 pagecache_page, ptl);
4174 goto out_put_page;
4175 }
4176 entry = huge_pte_mkdirty(entry);
4177 }
4178 entry = pte_mkyoung(entry);
4179 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4180 flags & FAULT_FLAG_WRITE))
4181 update_mmu_cache(vma, haddr, ptep);
4182out_put_page:
4183 if (page != pagecache_page)
4184 unlock_page(page);
4185 put_page(page);
4186out_ptl:
4187 spin_unlock(ptl);
4188
4189 if (pagecache_page) {
4190 unlock_page(pagecache_page);
4191 put_page(pagecache_page);
4192 }
4193out_mutex:
4194 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4195 /*
4196 * Generally it's safe to hold refcount during waiting page lock. But
4197 * here we just wait to defer the next page fault to avoid busy loop and
4198 * the page is not used after unlocked before returning from the current
4199 * page fault. So we are safe from accessing freed page, even if we wait
4200 * here without taking refcount.
4201 */
4202 if (need_wait_lock)
4203 wait_on_page_locked(page);
4204 return ret;
4205}
4206
4207/*
4208 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4209 * modifications for huge pages.
4210 */
4211int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4212 pte_t *dst_pte,
4213 struct vm_area_struct *dst_vma,
4214 unsigned long dst_addr,
4215 unsigned long src_addr,
4216 struct page **pagep)
4217{
4218 struct address_space *mapping;
4219 pgoff_t idx;
4220 unsigned long size;
4221 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4222 struct hstate *h = hstate_vma(dst_vma);
4223 pte_t _dst_pte;
4224 spinlock_t *ptl;
4225 int ret;
4226 struct page *page;
4227
4228 if (!*pagep) {
4229 ret = -ENOMEM;
4230 page = alloc_huge_page(dst_vma, dst_addr, 0);
4231 if (IS_ERR(page))
4232 goto out;
4233
4234 ret = copy_huge_page_from_user(page,
4235 (const void __user *) src_addr,
4236 pages_per_huge_page(h), false);
4237
4238 /* fallback to copy_from_user outside mmap_sem */
4239 if (unlikely(ret)) {
4240 ret = -ENOENT;
4241 *pagep = page;
4242 /* don't free the page */
4243 goto out;
4244 }
4245 } else {
4246 page = *pagep;
4247 *pagep = NULL;
4248 }
4249
4250 /*
4251 * The memory barrier inside __SetPageUptodate makes sure that
4252 * preceding stores to the page contents become visible before
4253 * the set_pte_at() write.
4254 */
4255 __SetPageUptodate(page);
4256
4257 mapping = dst_vma->vm_file->f_mapping;
4258 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4259
4260 /*
4261 * If shared, add to page cache
4262 */
4263 if (vm_shared) {
4264 size = i_size_read(mapping->host) >> huge_page_shift(h);
4265 ret = -EFAULT;
4266 if (idx >= size)
4267 goto out_release_nounlock;
4268
4269 /*
4270 * Serialization between remove_inode_hugepages() and
4271 * huge_add_to_page_cache() below happens through the
4272 * hugetlb_fault_mutex_table that here must be hold by
4273 * the caller.
4274 */
4275 ret = huge_add_to_page_cache(page, mapping, idx);
4276 if (ret)
4277 goto out_release_nounlock;
4278 }
4279
4280 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4281 spin_lock(ptl);
4282
4283 /*
4284 * Recheck the i_size after holding PT lock to make sure not
4285 * to leave any page mapped (as page_mapped()) beyond the end
4286 * of the i_size (remove_inode_hugepages() is strict about
4287 * enforcing that). If we bail out here, we'll also leave a
4288 * page in the radix tree in the vm_shared case beyond the end
4289 * of the i_size, but remove_inode_hugepages() will take care
4290 * of it as soon as we drop the hugetlb_fault_mutex_table.
4291 */
4292 size = i_size_read(mapping->host) >> huge_page_shift(h);
4293 ret = -EFAULT;
4294 if (idx >= size)
4295 goto out_release_unlock;
4296
4297 ret = -EEXIST;
4298 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4299 goto out_release_unlock;
4300
4301 if (vm_shared) {
4302 page_dup_rmap(page, true);
4303 } else {
4304 ClearPagePrivate(page);
4305 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4306 }
4307
4308 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4309 if (dst_vma->vm_flags & VM_WRITE)
4310 _dst_pte = huge_pte_mkdirty(_dst_pte);
4311 _dst_pte = pte_mkyoung(_dst_pte);
4312
4313 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4314
4315 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4316 dst_vma->vm_flags & VM_WRITE);
4317 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4318
4319 /* No need to invalidate - it was non-present before */
4320 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4321
4322 spin_unlock(ptl);
4323 set_page_huge_active(page);
4324 if (vm_shared)
4325 unlock_page(page);
4326 ret = 0;
4327out:
4328 return ret;
4329out_release_unlock:
4330 spin_unlock(ptl);
4331 if (vm_shared)
4332 unlock_page(page);
4333out_release_nounlock:
4334 put_page(page);
4335 goto out;
4336}
4337
4338long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4339 struct page **pages, struct vm_area_struct **vmas,
4340 unsigned long *position, unsigned long *nr_pages,
4341 long i, unsigned int flags, int *nonblocking)
4342{
4343 unsigned long pfn_offset;
4344 unsigned long vaddr = *position;
4345 unsigned long remainder = *nr_pages;
4346 struct hstate *h = hstate_vma(vma);
4347 int err = -EFAULT;
4348
4349 while (vaddr < vma->vm_end && remainder) {
4350 pte_t *pte;
4351 spinlock_t *ptl = NULL;
4352 int absent;
4353 struct page *page;
4354
4355 /*
4356 * If we have a pending SIGKILL, don't keep faulting pages and
4357 * potentially allocating memory.
4358 */
4359 if (fatal_signal_pending(current)) {
4360 remainder = 0;
4361 break;
4362 }
4363
4364 /*
4365 * Some archs (sparc64, sh*) have multiple pte_ts to
4366 * each hugepage. We have to make sure we get the
4367 * first, for the page indexing below to work.
4368 *
4369 * Note that page table lock is not held when pte is null.
4370 */
4371 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4372 huge_page_size(h));
4373 if (pte)
4374 ptl = huge_pte_lock(h, mm, pte);
4375 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4376
4377 /*
4378 * When coredumping, it suits get_dump_page if we just return
4379 * an error where there's an empty slot with no huge pagecache
4380 * to back it. This way, we avoid allocating a hugepage, and
4381 * the sparse dumpfile avoids allocating disk blocks, but its
4382 * huge holes still show up with zeroes where they need to be.
4383 */
4384 if (absent && (flags & FOLL_DUMP) &&
4385 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4386 if (pte)
4387 spin_unlock(ptl);
4388 remainder = 0;
4389 break;
4390 }
4391
4392 /*
4393 * We need call hugetlb_fault for both hugepages under migration
4394 * (in which case hugetlb_fault waits for the migration,) and
4395 * hwpoisoned hugepages (in which case we need to prevent the
4396 * caller from accessing to them.) In order to do this, we use
4397 * here is_swap_pte instead of is_hugetlb_entry_migration and
4398 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4399 * both cases, and because we can't follow correct pages
4400 * directly from any kind of swap entries.
4401 */
4402 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4403 ((flags & FOLL_WRITE) &&
4404 !huge_pte_write(huge_ptep_get(pte)))) {
4405 vm_fault_t ret;
4406 unsigned int fault_flags = 0;
4407
4408 if (pte)
4409 spin_unlock(ptl);
4410 if (flags & FOLL_WRITE)
4411 fault_flags |= FAULT_FLAG_WRITE;
4412 if (nonblocking)
4413 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4414 if (flags & FOLL_NOWAIT)
4415 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4416 FAULT_FLAG_RETRY_NOWAIT;
4417 if (flags & FOLL_TRIED) {
4418 VM_WARN_ON_ONCE(fault_flags &
4419 FAULT_FLAG_ALLOW_RETRY);
4420 fault_flags |= FAULT_FLAG_TRIED;
4421 }
4422 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4423 if (ret & VM_FAULT_ERROR) {
4424 err = vm_fault_to_errno(ret, flags);
4425 remainder = 0;
4426 break;
4427 }
4428 if (ret & VM_FAULT_RETRY) {
4429 if (nonblocking &&
4430 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4431 *nonblocking = 0;
4432 *nr_pages = 0;
4433 /*
4434 * VM_FAULT_RETRY must not return an
4435 * error, it will return zero
4436 * instead.
4437 *
4438 * No need to update "position" as the
4439 * caller will not check it after
4440 * *nr_pages is set to 0.
4441 */
4442 return i;
4443 }
4444 continue;
4445 }
4446
4447 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4448 page = pte_page(huge_ptep_get(pte));
4449
4450 /*
4451 * Instead of doing 'try_get_page()' below in the same_page
4452 * loop, just check the count once here.
4453 */
4454 if (unlikely(page_count(page) <= 0)) {
4455 if (pages) {
4456 spin_unlock(ptl);
4457 remainder = 0;
4458 err = -ENOMEM;
4459 break;
4460 }
4461 }
4462same_page:
4463 if (pages) {
4464 pages[i] = mem_map_offset(page, pfn_offset);
4465 get_page(pages[i]);
4466 }
4467
4468 if (vmas)
4469 vmas[i] = vma;
4470
4471 vaddr += PAGE_SIZE;
4472 ++pfn_offset;
4473 --remainder;
4474 ++i;
4475 if (vaddr < vma->vm_end && remainder &&
4476 pfn_offset < pages_per_huge_page(h)) {
4477 /*
4478 * We use pfn_offset to avoid touching the pageframes
4479 * of this compound page.
4480 */
4481 goto same_page;
4482 }
4483 spin_unlock(ptl);
4484 }
4485 *nr_pages = remainder;
4486 /*
4487 * setting position is actually required only if remainder is
4488 * not zero but it's faster not to add a "if (remainder)"
4489 * branch.
4490 */
4491 *position = vaddr;
4492
4493 return i ? i : err;
4494}
4495
4496#ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4497/*
4498 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4499 * implement this.
4500 */
4501#define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4502#endif
4503
4504unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4505 unsigned long address, unsigned long end, pgprot_t newprot)
4506{
4507 struct mm_struct *mm = vma->vm_mm;
4508 unsigned long start = address;
4509 pte_t *ptep;
4510 pte_t pte;
4511 struct hstate *h = hstate_vma(vma);
4512 unsigned long pages = 0;
4513 bool shared_pmd = false;
4514 struct mmu_notifier_range range;
4515
4516 /*
4517 * In the case of shared PMDs, the area to flush could be beyond
4518 * start/end. Set range.start/range.end to cover the maximum possible
4519 * range if PMD sharing is possible.
4520 */
4521 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
4522 0, vma, mm, start, end);
4523 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4524
4525 BUG_ON(address >= end);
4526 flush_cache_range(vma, range.start, range.end);
4527
4528 mmu_notifier_invalidate_range_start(&range);
4529 i_mmap_lock_write(vma->vm_file->f_mapping);
4530 for (; address < end; address += huge_page_size(h)) {
4531 spinlock_t *ptl;
4532 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4533 if (!ptep)
4534 continue;
4535 ptl = huge_pte_lock(h, mm, ptep);
4536 if (huge_pmd_unshare(mm, &address, ptep)) {
4537 pages++;
4538 spin_unlock(ptl);
4539 shared_pmd = true;
4540 continue;
4541 }
4542 pte = huge_ptep_get(ptep);
4543 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4544 spin_unlock(ptl);
4545 continue;
4546 }
4547 if (unlikely(is_hugetlb_entry_migration(pte))) {
4548 swp_entry_t entry = pte_to_swp_entry(pte);
4549
4550 if (is_write_migration_entry(entry)) {
4551 pte_t newpte;
4552
4553 make_migration_entry_read(&entry);
4554 newpte = swp_entry_to_pte(entry);
4555 set_huge_swap_pte_at(mm, address, ptep,
4556 newpte, huge_page_size(h));
4557 pages++;
4558 }
4559 spin_unlock(ptl);
4560 continue;
4561 }
4562 if (!huge_pte_none(pte)) {
4563 pte_t old_pte;
4564
4565 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
4566 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
4567 pte = arch_make_huge_pte(pte, vma, NULL, 0);
4568 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
4569 pages++;
4570 }
4571 spin_unlock(ptl);
4572 }
4573 /*
4574 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4575 * may have cleared our pud entry and done put_page on the page table:
4576 * once we release i_mmap_rwsem, another task can do the final put_page
4577 * and that page table be reused and filled with junk. If we actually
4578 * did unshare a page of pmds, flush the range corresponding to the pud.
4579 */
4580 if (shared_pmd)
4581 flush_hugetlb_tlb_range(vma, range.start, range.end);
4582 else
4583 flush_hugetlb_tlb_range(vma, start, end);
4584 /*
4585 * No need to call mmu_notifier_invalidate_range() we are downgrading
4586 * page table protection not changing it to point to a new page.
4587 *
4588 * See Documentation/vm/mmu_notifier.rst
4589 */
4590 i_mmap_unlock_write(vma->vm_file->f_mapping);
4591 mmu_notifier_invalidate_range_end(&range);
4592
4593 return pages << h->order;
4594}
4595
4596int hugetlb_reserve_pages(struct inode *inode,
4597 long from, long to,
4598 struct vm_area_struct *vma,
4599 vm_flags_t vm_flags)
4600{
4601 long ret, chg;
4602 struct hstate *h = hstate_inode(inode);
4603 struct hugepage_subpool *spool = subpool_inode(inode);
4604 struct resv_map *resv_map;
4605 long gbl_reserve;
4606
4607 /* This should never happen */
4608 if (from > to) {
4609 VM_WARN(1, "%s called with a negative range\n", __func__);
4610 return -EINVAL;
4611 }
4612
4613 /*
4614 * Only apply hugepage reservation if asked. At fault time, an
4615 * attempt will be made for VM_NORESERVE to allocate a page
4616 * without using reserves
4617 */
4618 if (vm_flags & VM_NORESERVE)
4619 return 0;
4620
4621 /*
4622 * Shared mappings base their reservation on the number of pages that
4623 * are already allocated on behalf of the file. Private mappings need
4624 * to reserve the full area even if read-only as mprotect() may be
4625 * called to make the mapping read-write. Assume !vma is a shm mapping
4626 */
4627 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4628 /*
4629 * resv_map can not be NULL as hugetlb_reserve_pages is only
4630 * called for inodes for which resv_maps were created (see
4631 * hugetlbfs_get_inode).
4632 */
4633 resv_map = inode_resv_map(inode);
4634
4635 chg = region_chg(resv_map, from, to);
4636
4637 } else {
4638 resv_map = resv_map_alloc();
4639 if (!resv_map)
4640 return -ENOMEM;
4641
4642 chg = to - from;
4643
4644 set_vma_resv_map(vma, resv_map);
4645 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4646 }
4647
4648 if (chg < 0) {
4649 ret = chg;
4650 goto out_err;
4651 }
4652
4653 /*
4654 * There must be enough pages in the subpool for the mapping. If
4655 * the subpool has a minimum size, there may be some global
4656 * reservations already in place (gbl_reserve).
4657 */
4658 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4659 if (gbl_reserve < 0) {
4660 ret = -ENOSPC;
4661 goto out_err;
4662 }
4663
4664 /*
4665 * Check enough hugepages are available for the reservation.
4666 * Hand the pages back to the subpool if there are not
4667 */
4668 ret = hugetlb_acct_memory(h, gbl_reserve);
4669 if (ret < 0) {
4670 /* put back original number of pages, chg */
4671 (void)hugepage_subpool_put_pages(spool, chg);
4672 goto out_err;
4673 }
4674
4675 /*
4676 * Account for the reservations made. Shared mappings record regions
4677 * that have reservations as they are shared by multiple VMAs.
4678 * When the last VMA disappears, the region map says how much
4679 * the reservation was and the page cache tells how much of
4680 * the reservation was consumed. Private mappings are per-VMA and
4681 * only the consumed reservations are tracked. When the VMA
4682 * disappears, the original reservation is the VMA size and the
4683 * consumed reservations are stored in the map. Hence, nothing
4684 * else has to be done for private mappings here
4685 */
4686 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4687 long add = region_add(resv_map, from, to);
4688
4689 if (unlikely(chg > add)) {
4690 /*
4691 * pages in this range were added to the reserve
4692 * map between region_chg and region_add. This
4693 * indicates a race with alloc_huge_page. Adjust
4694 * the subpool and reserve counts modified above
4695 * based on the difference.
4696 */
4697 long rsv_adjust;
4698
4699 rsv_adjust = hugepage_subpool_put_pages(spool,
4700 chg - add);
4701 hugetlb_acct_memory(h, -rsv_adjust);
4702 }
4703 }
4704 return 0;
4705out_err:
4706 if (!vma || vma->vm_flags & VM_MAYSHARE)
4707 /* Don't call region_abort if region_chg failed */
4708 if (chg >= 0)
4709 region_abort(resv_map, from, to);
4710 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4711 kref_put(&resv_map->refs, resv_map_release);
4712 return ret;
4713}
4714
4715long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4716 long freed)
4717{
4718 struct hstate *h = hstate_inode(inode);
4719 struct resv_map *resv_map = inode_resv_map(inode);
4720 long chg = 0;
4721 struct hugepage_subpool *spool = subpool_inode(inode);
4722 long gbl_reserve;
4723
4724 /*
4725 * Since this routine can be called in the evict inode path for all
4726 * hugetlbfs inodes, resv_map could be NULL.
4727 */
4728 if (resv_map) {
4729 chg = region_del(resv_map, start, end);
4730 /*
4731 * region_del() can fail in the rare case where a region
4732 * must be split and another region descriptor can not be
4733 * allocated. If end == LONG_MAX, it will not fail.
4734 */
4735 if (chg < 0)
4736 return chg;
4737 }
4738
4739 spin_lock(&inode->i_lock);
4740 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4741 spin_unlock(&inode->i_lock);
4742
4743 /*
4744 * If the subpool has a minimum size, the number of global
4745 * reservations to be released may be adjusted.
4746 */
4747 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4748 hugetlb_acct_memory(h, -gbl_reserve);
4749
4750 return 0;
4751}
4752
4753#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4754static unsigned long page_table_shareable(struct vm_area_struct *svma,
4755 struct vm_area_struct *vma,
4756 unsigned long addr, pgoff_t idx)
4757{
4758 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4759 svma->vm_start;
4760 unsigned long sbase = saddr & PUD_MASK;
4761 unsigned long s_end = sbase + PUD_SIZE;
4762
4763 /* Allow segments to share if only one is marked locked */
4764 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4765 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4766
4767 /*
4768 * match the virtual addresses, permission and the alignment of the
4769 * page table page.
4770 */
4771 if (pmd_index(addr) != pmd_index(saddr) ||
4772 vm_flags != svm_flags ||
4773 sbase < svma->vm_start || svma->vm_end < s_end)
4774 return 0;
4775
4776 return saddr;
4777}
4778
4779static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4780{
4781 unsigned long base = addr & PUD_MASK;
4782 unsigned long end = base + PUD_SIZE;
4783
4784 /*
4785 * check on proper vm_flags and page table alignment
4786 */
4787 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
4788 return true;
4789 return false;
4790}
4791
4792/*
4793 * Determine if start,end range within vma could be mapped by shared pmd.
4794 * If yes, adjust start and end to cover range associated with possible
4795 * shared pmd mappings.
4796 */
4797void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4798 unsigned long *start, unsigned long *end)
4799{
4800 unsigned long check_addr = *start;
4801
4802 if (!(vma->vm_flags & VM_MAYSHARE))
4803 return;
4804
4805 for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
4806 unsigned long a_start = check_addr & PUD_MASK;
4807 unsigned long a_end = a_start + PUD_SIZE;
4808
4809 /*
4810 * If sharing is possible, adjust start/end if necessary.
4811 */
4812 if (range_in_vma(vma, a_start, a_end)) {
4813 if (a_start < *start)
4814 *start = a_start;
4815 if (a_end > *end)
4816 *end = a_end;
4817 }
4818 }
4819}
4820
4821/*
4822 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4823 * and returns the corresponding pte. While this is not necessary for the
4824 * !shared pmd case because we can allocate the pmd later as well, it makes the
4825 * code much cleaner. pmd allocation is essential for the shared case because
4826 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4827 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4828 * bad pmd for sharing.
4829 */
4830pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4831{
4832 struct vm_area_struct *vma = find_vma(mm, addr);
4833 struct address_space *mapping = vma->vm_file->f_mapping;
4834 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4835 vma->vm_pgoff;
4836 struct vm_area_struct *svma;
4837 unsigned long saddr;
4838 pte_t *spte = NULL;
4839 pte_t *pte;
4840 spinlock_t *ptl;
4841
4842 if (!vma_shareable(vma, addr))
4843 return (pte_t *)pmd_alloc(mm, pud, addr);
4844
4845 i_mmap_lock_write(mapping);
4846 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4847 if (svma == vma)
4848 continue;
4849
4850 saddr = page_table_shareable(svma, vma, addr, idx);
4851 if (saddr) {
4852 spte = huge_pte_offset(svma->vm_mm, saddr,
4853 vma_mmu_pagesize(svma));
4854 if (spte) {
4855 get_page(virt_to_page(spte));
4856 break;
4857 }
4858 }
4859 }
4860
4861 if (!spte)
4862 goto out;
4863
4864 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4865 if (pud_none(*pud)) {
4866 pud_populate(mm, pud,
4867 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4868 mm_inc_nr_pmds(mm);
4869 } else {
4870 put_page(virt_to_page(spte));
4871 }
4872 spin_unlock(ptl);
4873out:
4874 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4875 i_mmap_unlock_write(mapping);
4876 return pte;
4877}
4878
4879/*
4880 * unmap huge page backed by shared pte.
4881 *
4882 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4883 * indicated by page_count > 1, unmap is achieved by clearing pud and
4884 * decrementing the ref count. If count == 1, the pte page is not shared.
4885 *
4886 * called with page table lock held.
4887 *
4888 * returns: 1 successfully unmapped a shared pte page
4889 * 0 the underlying pte page is not shared, or it is the last user
4890 */
4891int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4892{
4893 pgd_t *pgd = pgd_offset(mm, *addr);
4894 p4d_t *p4d = p4d_offset(pgd, *addr);
4895 pud_t *pud = pud_offset(p4d, *addr);
4896
4897 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4898 if (page_count(virt_to_page(ptep)) == 1)
4899 return 0;
4900
4901 pud_clear(pud);
4902 put_page(virt_to_page(ptep));
4903 mm_dec_nr_pmds(mm);
4904 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4905 return 1;
4906}
4907#define want_pmd_share() (1)
4908#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4909pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4910{
4911 return NULL;
4912}
4913
4914int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4915{
4916 return 0;
4917}
4918
4919void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4920 unsigned long *start, unsigned long *end)
4921{
4922}
4923#define want_pmd_share() (0)
4924#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4925
4926#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4927pte_t *huge_pte_alloc(struct mm_struct *mm,
4928 unsigned long addr, unsigned long sz)
4929{
4930 pgd_t *pgd;
4931 p4d_t *p4d;
4932 pud_t *pud;
4933 pte_t *pte = NULL;
4934
4935 pgd = pgd_offset(mm, addr);
4936 p4d = p4d_alloc(mm, pgd, addr);
4937 if (!p4d)
4938 return NULL;
4939 pud = pud_alloc(mm, p4d, addr);
4940 if (pud) {
4941 if (sz == PUD_SIZE) {
4942 pte = (pte_t *)pud;
4943 } else {
4944 BUG_ON(sz != PMD_SIZE);
4945 if (want_pmd_share() && pud_none(*pud))
4946 pte = huge_pmd_share(mm, addr, pud);
4947 else
4948 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4949 }
4950 }
4951 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4952
4953 return pte;
4954}
4955
4956/*
4957 * huge_pte_offset() - Walk the page table to resolve the hugepage
4958 * entry at address @addr
4959 *
4960 * Return: Pointer to page table or swap entry (PUD or PMD) for
4961 * address @addr, or NULL if a p*d_none() entry is encountered and the
4962 * size @sz doesn't match the hugepage size at this level of the page
4963 * table.
4964 */
4965pte_t *huge_pte_offset(struct mm_struct *mm,
4966 unsigned long addr, unsigned long sz)
4967{
4968 pgd_t *pgd;
4969 p4d_t *p4d;
4970 pud_t *pud;
4971 pmd_t *pmd;
4972
4973 pgd = pgd_offset(mm, addr);
4974 if (!pgd_present(*pgd))
4975 return NULL;
4976 p4d = p4d_offset(pgd, addr);
4977 if (!p4d_present(*p4d))
4978 return NULL;
4979
4980 pud = pud_offset(p4d, addr);
4981 if (sz != PUD_SIZE && pud_none(*pud))
4982 return NULL;
4983 /* hugepage or swap? */
4984 if (pud_huge(*pud) || !pud_present(*pud))
4985 return (pte_t *)pud;
4986
4987 pmd = pmd_offset(pud, addr);
4988 if (sz != PMD_SIZE && pmd_none(*pmd))
4989 return NULL;
4990 /* hugepage or swap? */
4991 if (pmd_huge(*pmd) || !pmd_present(*pmd))
4992 return (pte_t *)pmd;
4993
4994 return NULL;
4995}
4996
4997#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4998
4999/*
5000 * These functions are overwritable if your architecture needs its own
5001 * behavior.
5002 */
5003struct page * __weak
5004follow_huge_addr(struct mm_struct *mm, unsigned long address,
5005 int write)
5006{
5007 return ERR_PTR(-EINVAL);
5008}
5009
5010struct page * __weak
5011follow_huge_pd(struct vm_area_struct *vma,
5012 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5013{
5014 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5015 return NULL;
5016}
5017
5018struct page * __weak
5019follow_huge_pmd(struct mm_struct *mm, unsigned long address,
5020 pmd_t *pmd, int flags)
5021{
5022 struct page *page = NULL;
5023 spinlock_t *ptl;
5024 pte_t pte;
5025retry:
5026 ptl = pmd_lockptr(mm, pmd);
5027 spin_lock(ptl);
5028 /*
5029 * make sure that the address range covered by this pmd is not
5030 * unmapped from other threads.
5031 */
5032 if (!pmd_huge(*pmd))
5033 goto out;
5034 pte = huge_ptep_get((pte_t *)pmd);
5035 if (pte_present(pte)) {
5036 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
5037 if (flags & FOLL_GET)
5038 get_page(page);
5039 } else {
5040 if (is_hugetlb_entry_migration(pte)) {
5041 spin_unlock(ptl);
5042 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5043 goto retry;
5044 }
5045 /*
5046 * hwpoisoned entry is treated as no_page_table in
5047 * follow_page_mask().
5048 */
5049 }
5050out:
5051 spin_unlock(ptl);
5052 return page;
5053}
5054
5055struct page * __weak
5056follow_huge_pud(struct mm_struct *mm, unsigned long address,
5057 pud_t *pud, int flags)
5058{
5059 if (flags & FOLL_GET)
5060 return NULL;
5061
5062 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5063}
5064
5065struct page * __weak
5066follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5067{
5068 if (flags & FOLL_GET)
5069 return NULL;
5070
5071 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5072}
5073
5074bool isolate_huge_page(struct page *page, struct list_head *list)
5075{
5076 bool ret = true;
5077
5078 VM_BUG_ON_PAGE(!PageHead(page), page);
5079 spin_lock(&hugetlb_lock);
5080 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
5081 ret = false;
5082 goto unlock;
5083 }
5084 clear_page_huge_active(page);
5085 list_move_tail(&page->lru, list);
5086unlock:
5087 spin_unlock(&hugetlb_lock);
5088 return ret;
5089}
5090
5091void putback_active_hugepage(struct page *page)
5092{
5093 VM_BUG_ON_PAGE(!PageHead(page), page);
5094 spin_lock(&hugetlb_lock);
5095 set_page_huge_active(page);
5096 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5097 spin_unlock(&hugetlb_lock);
5098 put_page(page);
5099}
5100
5101void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5102{
5103 struct hstate *h = page_hstate(oldpage);
5104
5105 hugetlb_cgroup_migrate(oldpage, newpage);
5106 set_page_owner_migrate_reason(newpage, reason);
5107
5108 /*
5109 * transfer temporary state of the new huge page. This is
5110 * reverse to other transitions because the newpage is going to
5111 * be final while the old one will be freed so it takes over
5112 * the temporary status.
5113 *
5114 * Also note that we have to transfer the per-node surplus state
5115 * here as well otherwise the global surplus count will not match
5116 * the per-node's.
5117 */
5118 if (PageHugeTemporary(newpage)) {
5119 int old_nid = page_to_nid(oldpage);
5120 int new_nid = page_to_nid(newpage);
5121
5122 SetPageHugeTemporary(oldpage);
5123 ClearPageHugeTemporary(newpage);
5124
5125 spin_lock(&hugetlb_lock);
5126 if (h->surplus_huge_pages_node[old_nid]) {
5127 h->surplus_huge_pages_node[old_nid]--;
5128 h->surplus_huge_pages_node[new_nid]++;
5129 }
5130 spin_unlock(&hugetlb_lock);
5131 }
5132}