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#include <linux/migrate.h>
34
35#include <asm/page.h>
36#include <asm/pgalloc.h>
37#include <asm/tlb.h>
38
39#include <linux/io.h>
40#include <linux/hugetlb.h>
41#include <linux/hugetlb_cgroup.h>
42#include <linux/node.h>
43#include <linux/page_owner.h>
44#include "internal.h"
45#include "hugetlb_vmemmap.h"
46
47int hugetlb_max_hstate __read_mostly;
48unsigned int default_hstate_idx;
49struct hstate hstates[HUGE_MAX_HSTATE];
50
51#ifdef CONFIG_CMA
52static struct cma *hugetlb_cma[MAX_NUMNODES];
53#endif
54static unsigned long hugetlb_cma_size __initdata;
55
56/*
57 * Minimum page order among possible hugepage sizes, set to a proper value
58 * at boot time.
59 */
60static unsigned int minimum_order __read_mostly = UINT_MAX;
61
62__initdata LIST_HEAD(huge_boot_pages);
63
64/* for command line parsing */
65static struct hstate * __initdata parsed_hstate;
66static unsigned long __initdata default_hstate_max_huge_pages;
67static bool __initdata parsed_valid_hugepagesz = true;
68static bool __initdata parsed_default_hugepagesz;
69
70/*
71 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
72 * free_huge_pages, and surplus_huge_pages.
73 */
74DEFINE_SPINLOCK(hugetlb_lock);
75
76/*
77 * Serializes faults on the same logical page. This is used to
78 * prevent spurious OOMs when the hugepage pool is fully utilized.
79 */
80static int num_fault_mutexes;
81struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
82
83/* Forward declaration */
84static int hugetlb_acct_memory(struct hstate *h, long delta);
85
86static inline bool subpool_is_free(struct hugepage_subpool *spool)
87{
88 if (spool->count)
89 return false;
90 if (spool->max_hpages != -1)
91 return spool->used_hpages == 0;
92 if (spool->min_hpages != -1)
93 return spool->rsv_hpages == spool->min_hpages;
94
95 return true;
96}
97
98static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
99 unsigned long irq_flags)
100{
101 spin_unlock_irqrestore(&spool->lock, irq_flags);
102
103 /* If no pages are used, and no other handles to the subpool
104 * remain, give up any reservations based on minimum size and
105 * free the subpool */
106 if (subpool_is_free(spool)) {
107 if (spool->min_hpages != -1)
108 hugetlb_acct_memory(spool->hstate,
109 -spool->min_hpages);
110 kfree(spool);
111 }
112}
113
114struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
115 long min_hpages)
116{
117 struct hugepage_subpool *spool;
118
119 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
120 if (!spool)
121 return NULL;
122
123 spin_lock_init(&spool->lock);
124 spool->count = 1;
125 spool->max_hpages = max_hpages;
126 spool->hstate = h;
127 spool->min_hpages = min_hpages;
128
129 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
130 kfree(spool);
131 return NULL;
132 }
133 spool->rsv_hpages = min_hpages;
134
135 return spool;
136}
137
138void hugepage_put_subpool(struct hugepage_subpool *spool)
139{
140 unsigned long flags;
141
142 spin_lock_irqsave(&spool->lock, flags);
143 BUG_ON(!spool->count);
144 spool->count--;
145 unlock_or_release_subpool(spool, flags);
146}
147
148/*
149 * Subpool accounting for allocating and reserving pages.
150 * Return -ENOMEM if there are not enough resources to satisfy the
151 * request. Otherwise, return the number of pages by which the
152 * global pools must be adjusted (upward). The returned value may
153 * only be different than the passed value (delta) in the case where
154 * a subpool minimum size must be maintained.
155 */
156static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
157 long delta)
158{
159 long ret = delta;
160
161 if (!spool)
162 return ret;
163
164 spin_lock_irq(&spool->lock);
165
166 if (spool->max_hpages != -1) { /* maximum size accounting */
167 if ((spool->used_hpages + delta) <= spool->max_hpages)
168 spool->used_hpages += delta;
169 else {
170 ret = -ENOMEM;
171 goto unlock_ret;
172 }
173 }
174
175 /* minimum size accounting */
176 if (spool->min_hpages != -1 && spool->rsv_hpages) {
177 if (delta > spool->rsv_hpages) {
178 /*
179 * Asking for more reserves than those already taken on
180 * behalf of subpool. Return difference.
181 */
182 ret = delta - spool->rsv_hpages;
183 spool->rsv_hpages = 0;
184 } else {
185 ret = 0; /* reserves already accounted for */
186 spool->rsv_hpages -= delta;
187 }
188 }
189
190unlock_ret:
191 spin_unlock_irq(&spool->lock);
192 return ret;
193}
194
195/*
196 * Subpool accounting for freeing and unreserving pages.
197 * Return the number of global page reservations that must be dropped.
198 * The return value may only be different than the passed value (delta)
199 * in the case where a subpool minimum size must be maintained.
200 */
201static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
202 long delta)
203{
204 long ret = delta;
205 unsigned long flags;
206
207 if (!spool)
208 return delta;
209
210 spin_lock_irqsave(&spool->lock, flags);
211
212 if (spool->max_hpages != -1) /* maximum size accounting */
213 spool->used_hpages -= delta;
214
215 /* minimum size accounting */
216 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
217 if (spool->rsv_hpages + delta <= spool->min_hpages)
218 ret = 0;
219 else
220 ret = spool->rsv_hpages + delta - spool->min_hpages;
221
222 spool->rsv_hpages += delta;
223 if (spool->rsv_hpages > spool->min_hpages)
224 spool->rsv_hpages = spool->min_hpages;
225 }
226
227 /*
228 * If hugetlbfs_put_super couldn't free spool due to an outstanding
229 * quota reference, free it now.
230 */
231 unlock_or_release_subpool(spool, flags);
232
233 return ret;
234}
235
236static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
237{
238 return HUGETLBFS_SB(inode->i_sb)->spool;
239}
240
241static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
242{
243 return subpool_inode(file_inode(vma->vm_file));
244}
245
246/* Helper that removes a struct file_region from the resv_map cache and returns
247 * it for use.
248 */
249static struct file_region *
250get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
251{
252 struct file_region *nrg = NULL;
253
254 VM_BUG_ON(resv->region_cache_count <= 0);
255
256 resv->region_cache_count--;
257 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
258 list_del(&nrg->link);
259
260 nrg->from = from;
261 nrg->to = to;
262
263 return nrg;
264}
265
266static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
267 struct file_region *rg)
268{
269#ifdef CONFIG_CGROUP_HUGETLB
270 nrg->reservation_counter = rg->reservation_counter;
271 nrg->css = rg->css;
272 if (rg->css)
273 css_get(rg->css);
274#endif
275}
276
277/* Helper that records hugetlb_cgroup uncharge info. */
278static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
279 struct hstate *h,
280 struct resv_map *resv,
281 struct file_region *nrg)
282{
283#ifdef CONFIG_CGROUP_HUGETLB
284 if (h_cg) {
285 nrg->reservation_counter =
286 &h_cg->rsvd_hugepage[hstate_index(h)];
287 nrg->css = &h_cg->css;
288 /*
289 * The caller will hold exactly one h_cg->css reference for the
290 * whole contiguous reservation region. But this area might be
291 * scattered when there are already some file_regions reside in
292 * it. As a result, many file_regions may share only one css
293 * reference. In order to ensure that one file_region must hold
294 * exactly one h_cg->css reference, we should do css_get for
295 * each file_region and leave the reference held by caller
296 * untouched.
297 */
298 css_get(&h_cg->css);
299 if (!resv->pages_per_hpage)
300 resv->pages_per_hpage = pages_per_huge_page(h);
301 /* pages_per_hpage should be the same for all entries in
302 * a resv_map.
303 */
304 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
305 } else {
306 nrg->reservation_counter = NULL;
307 nrg->css = NULL;
308 }
309#endif
310}
311
312static void put_uncharge_info(struct file_region *rg)
313{
314#ifdef CONFIG_CGROUP_HUGETLB
315 if (rg->css)
316 css_put(rg->css);
317#endif
318}
319
320static bool has_same_uncharge_info(struct file_region *rg,
321 struct file_region *org)
322{
323#ifdef CONFIG_CGROUP_HUGETLB
324 return rg && org &&
325 rg->reservation_counter == org->reservation_counter &&
326 rg->css == org->css;
327
328#else
329 return true;
330#endif
331}
332
333static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
334{
335 struct file_region *nrg = NULL, *prg = NULL;
336
337 prg = list_prev_entry(rg, link);
338 if (&prg->link != &resv->regions && prg->to == rg->from &&
339 has_same_uncharge_info(prg, rg)) {
340 prg->to = rg->to;
341
342 list_del(&rg->link);
343 put_uncharge_info(rg);
344 kfree(rg);
345
346 rg = prg;
347 }
348
349 nrg = list_next_entry(rg, link);
350 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
351 has_same_uncharge_info(nrg, rg)) {
352 nrg->from = rg->from;
353
354 list_del(&rg->link);
355 put_uncharge_info(rg);
356 kfree(rg);
357 }
358}
359
360static inline long
361hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from,
362 long to, struct hstate *h, struct hugetlb_cgroup *cg,
363 long *regions_needed)
364{
365 struct file_region *nrg;
366
367 if (!regions_needed) {
368 nrg = get_file_region_entry_from_cache(map, from, to);
369 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
370 list_add(&nrg->link, rg->link.prev);
371 coalesce_file_region(map, nrg);
372 } else
373 *regions_needed += 1;
374
375 return to - from;
376}
377
378/*
379 * Must be called with resv->lock held.
380 *
381 * Calling this with regions_needed != NULL will count the number of pages
382 * to be added but will not modify the linked list. And regions_needed will
383 * indicate the number of file_regions needed in the cache to carry out to add
384 * the regions for this range.
385 */
386static long add_reservation_in_range(struct resv_map *resv, long f, long t,
387 struct hugetlb_cgroup *h_cg,
388 struct hstate *h, long *regions_needed)
389{
390 long add = 0;
391 struct list_head *head = &resv->regions;
392 long last_accounted_offset = f;
393 struct file_region *rg = NULL, *trg = NULL;
394
395 if (regions_needed)
396 *regions_needed = 0;
397
398 /* In this loop, we essentially handle an entry for the range
399 * [last_accounted_offset, rg->from), at every iteration, with some
400 * bounds checking.
401 */
402 list_for_each_entry_safe(rg, trg, head, link) {
403 /* Skip irrelevant regions that start before our range. */
404 if (rg->from < f) {
405 /* If this region ends after the last accounted offset,
406 * then we need to update last_accounted_offset.
407 */
408 if (rg->to > last_accounted_offset)
409 last_accounted_offset = rg->to;
410 continue;
411 }
412
413 /* When we find a region that starts beyond our range, we've
414 * finished.
415 */
416 if (rg->from >= t)
417 break;
418
419 /* Add an entry for last_accounted_offset -> rg->from, and
420 * update last_accounted_offset.
421 */
422 if (rg->from > last_accounted_offset)
423 add += hugetlb_resv_map_add(resv, rg,
424 last_accounted_offset,
425 rg->from, h, h_cg,
426 regions_needed);
427
428 last_accounted_offset = rg->to;
429 }
430
431 /* Handle the case where our range extends beyond
432 * last_accounted_offset.
433 */
434 if (last_accounted_offset < t)
435 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
436 t, h, h_cg, regions_needed);
437
438 VM_BUG_ON(add < 0);
439 return add;
440}
441
442/* Must be called with resv->lock acquired. Will drop lock to allocate entries.
443 */
444static int allocate_file_region_entries(struct resv_map *resv,
445 int regions_needed)
446 __must_hold(&resv->lock)
447{
448 struct list_head allocated_regions;
449 int to_allocate = 0, i = 0;
450 struct file_region *trg = NULL, *rg = NULL;
451
452 VM_BUG_ON(regions_needed < 0);
453
454 INIT_LIST_HEAD(&allocated_regions);
455
456 /*
457 * Check for sufficient descriptors in the cache to accommodate
458 * the number of in progress add operations plus regions_needed.
459 *
460 * This is a while loop because when we drop the lock, some other call
461 * to region_add or region_del may have consumed some region_entries,
462 * so we keep looping here until we finally have enough entries for
463 * (adds_in_progress + regions_needed).
464 */
465 while (resv->region_cache_count <
466 (resv->adds_in_progress + regions_needed)) {
467 to_allocate = resv->adds_in_progress + regions_needed -
468 resv->region_cache_count;
469
470 /* At this point, we should have enough entries in the cache
471 * for all the existing adds_in_progress. We should only be
472 * needing to allocate for regions_needed.
473 */
474 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
475
476 spin_unlock(&resv->lock);
477 for (i = 0; i < to_allocate; i++) {
478 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
479 if (!trg)
480 goto out_of_memory;
481 list_add(&trg->link, &allocated_regions);
482 }
483
484 spin_lock(&resv->lock);
485
486 list_splice(&allocated_regions, &resv->region_cache);
487 resv->region_cache_count += to_allocate;
488 }
489
490 return 0;
491
492out_of_memory:
493 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
494 list_del(&rg->link);
495 kfree(rg);
496 }
497 return -ENOMEM;
498}
499
500/*
501 * Add the huge page range represented by [f, t) to the reserve
502 * map. Regions will be taken from the cache to fill in this range.
503 * Sufficient regions should exist in the cache due to the previous
504 * call to region_chg with the same range, but in some cases the cache will not
505 * have sufficient entries due to races with other code doing region_add or
506 * region_del. The extra needed entries will be allocated.
507 *
508 * regions_needed is the out value provided by a previous call to region_chg.
509 *
510 * Return the number of new huge pages added to the map. This number is greater
511 * than or equal to zero. If file_region entries needed to be allocated for
512 * this operation and we were not able to allocate, it returns -ENOMEM.
513 * region_add of regions of length 1 never allocate file_regions and cannot
514 * fail; region_chg will always allocate at least 1 entry and a region_add for
515 * 1 page will only require at most 1 entry.
516 */
517static long region_add(struct resv_map *resv, long f, long t,
518 long in_regions_needed, struct hstate *h,
519 struct hugetlb_cgroup *h_cg)
520{
521 long add = 0, actual_regions_needed = 0;
522
523 spin_lock(&resv->lock);
524retry:
525
526 /* Count how many regions are actually needed to execute this add. */
527 add_reservation_in_range(resv, f, t, NULL, NULL,
528 &actual_regions_needed);
529
530 /*
531 * Check for sufficient descriptors in the cache to accommodate
532 * this add operation. Note that actual_regions_needed may be greater
533 * than in_regions_needed, as the resv_map may have been modified since
534 * the region_chg call. In this case, we need to make sure that we
535 * allocate extra entries, such that we have enough for all the
536 * existing adds_in_progress, plus the excess needed for this
537 * operation.
538 */
539 if (actual_regions_needed > in_regions_needed &&
540 resv->region_cache_count <
541 resv->adds_in_progress +
542 (actual_regions_needed - in_regions_needed)) {
543 /* region_add operation of range 1 should never need to
544 * allocate file_region entries.
545 */
546 VM_BUG_ON(t - f <= 1);
547
548 if (allocate_file_region_entries(
549 resv, actual_regions_needed - in_regions_needed)) {
550 return -ENOMEM;
551 }
552
553 goto retry;
554 }
555
556 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
557
558 resv->adds_in_progress -= in_regions_needed;
559
560 spin_unlock(&resv->lock);
561 return add;
562}
563
564/*
565 * Examine the existing reserve map and determine how many
566 * huge pages in the specified range [f, t) are NOT currently
567 * represented. This routine is called before a subsequent
568 * call to region_add that will actually modify the reserve
569 * map to add the specified range [f, t). region_chg does
570 * not change the number of huge pages represented by the
571 * map. A number of new file_region structures is added to the cache as a
572 * placeholder, for the subsequent region_add call to use. At least 1
573 * file_region structure is added.
574 *
575 * out_regions_needed is the number of regions added to the
576 * resv->adds_in_progress. This value needs to be provided to a follow up call
577 * to region_add or region_abort for proper accounting.
578 *
579 * Returns the number of huge pages that need to be added to the existing
580 * reservation map for the range [f, t). This number is greater or equal to
581 * zero. -ENOMEM is returned if a new file_region structure or cache entry
582 * is needed and can not be allocated.
583 */
584static long region_chg(struct resv_map *resv, long f, long t,
585 long *out_regions_needed)
586{
587 long chg = 0;
588
589 spin_lock(&resv->lock);
590
591 /* Count how many hugepages in this range are NOT represented. */
592 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
593 out_regions_needed);
594
595 if (*out_regions_needed == 0)
596 *out_regions_needed = 1;
597
598 if (allocate_file_region_entries(resv, *out_regions_needed))
599 return -ENOMEM;
600
601 resv->adds_in_progress += *out_regions_needed;
602
603 spin_unlock(&resv->lock);
604 return chg;
605}
606
607/*
608 * Abort the in progress add operation. The adds_in_progress field
609 * of the resv_map keeps track of the operations in progress between
610 * calls to region_chg and region_add. Operations are sometimes
611 * aborted after the call to region_chg. In such cases, region_abort
612 * is called to decrement the adds_in_progress counter. regions_needed
613 * is the value returned by the region_chg call, it is used to decrement
614 * the adds_in_progress counter.
615 *
616 * NOTE: The range arguments [f, t) are not needed or used in this
617 * routine. They are kept to make reading the calling code easier as
618 * arguments will match the associated region_chg call.
619 */
620static void region_abort(struct resv_map *resv, long f, long t,
621 long regions_needed)
622{
623 spin_lock(&resv->lock);
624 VM_BUG_ON(!resv->region_cache_count);
625 resv->adds_in_progress -= regions_needed;
626 spin_unlock(&resv->lock);
627}
628
629/*
630 * Delete the specified range [f, t) from the reserve map. If the
631 * t parameter is LONG_MAX, this indicates that ALL regions after f
632 * should be deleted. Locate the regions which intersect [f, t)
633 * and either trim, delete or split the existing regions.
634 *
635 * Returns the number of huge pages deleted from the reserve map.
636 * In the normal case, the return value is zero or more. In the
637 * case where a region must be split, a new region descriptor must
638 * be allocated. If the allocation fails, -ENOMEM will be returned.
639 * NOTE: If the parameter t == LONG_MAX, then we will never split
640 * a region and possibly return -ENOMEM. Callers specifying
641 * t == LONG_MAX do not need to check for -ENOMEM error.
642 */
643static long region_del(struct resv_map *resv, long f, long t)
644{
645 struct list_head *head = &resv->regions;
646 struct file_region *rg, *trg;
647 struct file_region *nrg = NULL;
648 long del = 0;
649
650retry:
651 spin_lock(&resv->lock);
652 list_for_each_entry_safe(rg, trg, head, link) {
653 /*
654 * Skip regions before the range to be deleted. file_region
655 * ranges are normally of the form [from, to). However, there
656 * may be a "placeholder" entry in the map which is of the form
657 * (from, to) with from == to. Check for placeholder entries
658 * at the beginning of the range to be deleted.
659 */
660 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
661 continue;
662
663 if (rg->from >= t)
664 break;
665
666 if (f > rg->from && t < rg->to) { /* Must split region */
667 /*
668 * Check for an entry in the cache before dropping
669 * lock and attempting allocation.
670 */
671 if (!nrg &&
672 resv->region_cache_count > resv->adds_in_progress) {
673 nrg = list_first_entry(&resv->region_cache,
674 struct file_region,
675 link);
676 list_del(&nrg->link);
677 resv->region_cache_count--;
678 }
679
680 if (!nrg) {
681 spin_unlock(&resv->lock);
682 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
683 if (!nrg)
684 return -ENOMEM;
685 goto retry;
686 }
687
688 del += t - f;
689 hugetlb_cgroup_uncharge_file_region(
690 resv, rg, t - f, false);
691
692 /* New entry for end of split region */
693 nrg->from = t;
694 nrg->to = rg->to;
695
696 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
697
698 INIT_LIST_HEAD(&nrg->link);
699
700 /* Original entry is trimmed */
701 rg->to = f;
702
703 list_add(&nrg->link, &rg->link);
704 nrg = NULL;
705 break;
706 }
707
708 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
709 del += rg->to - rg->from;
710 hugetlb_cgroup_uncharge_file_region(resv, rg,
711 rg->to - rg->from, true);
712 list_del(&rg->link);
713 kfree(rg);
714 continue;
715 }
716
717 if (f <= rg->from) { /* Trim beginning of region */
718 hugetlb_cgroup_uncharge_file_region(resv, rg,
719 t - rg->from, false);
720
721 del += t - rg->from;
722 rg->from = t;
723 } else { /* Trim end of region */
724 hugetlb_cgroup_uncharge_file_region(resv, rg,
725 rg->to - f, false);
726
727 del += rg->to - f;
728 rg->to = f;
729 }
730 }
731
732 spin_unlock(&resv->lock);
733 kfree(nrg);
734 return del;
735}
736
737/*
738 * A rare out of memory error was encountered which prevented removal of
739 * the reserve map region for a page. The huge page itself was free'ed
740 * and removed from the page cache. This routine will adjust the subpool
741 * usage count, and the global reserve count if needed. By incrementing
742 * these counts, the reserve map entry which could not be deleted will
743 * appear as a "reserved" entry instead of simply dangling with incorrect
744 * counts.
745 */
746void hugetlb_fix_reserve_counts(struct inode *inode)
747{
748 struct hugepage_subpool *spool = subpool_inode(inode);
749 long rsv_adjust;
750 bool reserved = false;
751
752 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
753 if (rsv_adjust > 0) {
754 struct hstate *h = hstate_inode(inode);
755
756 if (!hugetlb_acct_memory(h, 1))
757 reserved = true;
758 } else if (!rsv_adjust) {
759 reserved = true;
760 }
761
762 if (!reserved)
763 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
764}
765
766/*
767 * Count and return the number of huge pages in the reserve map
768 * that intersect with the range [f, t).
769 */
770static long region_count(struct resv_map *resv, long f, long t)
771{
772 struct list_head *head = &resv->regions;
773 struct file_region *rg;
774 long chg = 0;
775
776 spin_lock(&resv->lock);
777 /* Locate each segment we overlap with, and count that overlap. */
778 list_for_each_entry(rg, head, link) {
779 long seg_from;
780 long seg_to;
781
782 if (rg->to <= f)
783 continue;
784 if (rg->from >= t)
785 break;
786
787 seg_from = max(rg->from, f);
788 seg_to = min(rg->to, t);
789
790 chg += seg_to - seg_from;
791 }
792 spin_unlock(&resv->lock);
793
794 return chg;
795}
796
797/*
798 * Convert the address within this vma to the page offset within
799 * the mapping, in pagecache page units; huge pages here.
800 */
801static pgoff_t vma_hugecache_offset(struct hstate *h,
802 struct vm_area_struct *vma, unsigned long address)
803{
804 return ((address - vma->vm_start) >> huge_page_shift(h)) +
805 (vma->vm_pgoff >> huge_page_order(h));
806}
807
808pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
809 unsigned long address)
810{
811 return vma_hugecache_offset(hstate_vma(vma), vma, address);
812}
813EXPORT_SYMBOL_GPL(linear_hugepage_index);
814
815/*
816 * Return the size of the pages allocated when backing a VMA. In the majority
817 * cases this will be same size as used by the page table entries.
818 */
819unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
820{
821 if (vma->vm_ops && vma->vm_ops->pagesize)
822 return vma->vm_ops->pagesize(vma);
823 return PAGE_SIZE;
824}
825EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
826
827/*
828 * Return the page size being used by the MMU to back a VMA. In the majority
829 * of cases, the page size used by the kernel matches the MMU size. On
830 * architectures where it differs, an architecture-specific 'strong'
831 * version of this symbol is required.
832 */
833__weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
834{
835 return vma_kernel_pagesize(vma);
836}
837
838/*
839 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
840 * bits of the reservation map pointer, which are always clear due to
841 * alignment.
842 */
843#define HPAGE_RESV_OWNER (1UL << 0)
844#define HPAGE_RESV_UNMAPPED (1UL << 1)
845#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
846
847/*
848 * These helpers are used to track how many pages are reserved for
849 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
850 * is guaranteed to have their future faults succeed.
851 *
852 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
853 * the reserve counters are updated with the hugetlb_lock held. It is safe
854 * to reset the VMA at fork() time as it is not in use yet and there is no
855 * chance of the global counters getting corrupted as a result of the values.
856 *
857 * The private mapping reservation is represented in a subtly different
858 * manner to a shared mapping. A shared mapping has a region map associated
859 * with the underlying file, this region map represents the backing file
860 * pages which have ever had a reservation assigned which this persists even
861 * after the page is instantiated. A private mapping has a region map
862 * associated with the original mmap which is attached to all VMAs which
863 * reference it, this region map represents those offsets which have consumed
864 * reservation ie. where pages have been instantiated.
865 */
866static unsigned long get_vma_private_data(struct vm_area_struct *vma)
867{
868 return (unsigned long)vma->vm_private_data;
869}
870
871static void set_vma_private_data(struct vm_area_struct *vma,
872 unsigned long value)
873{
874 vma->vm_private_data = (void *)value;
875}
876
877static void
878resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
879 struct hugetlb_cgroup *h_cg,
880 struct hstate *h)
881{
882#ifdef CONFIG_CGROUP_HUGETLB
883 if (!h_cg || !h) {
884 resv_map->reservation_counter = NULL;
885 resv_map->pages_per_hpage = 0;
886 resv_map->css = NULL;
887 } else {
888 resv_map->reservation_counter =
889 &h_cg->rsvd_hugepage[hstate_index(h)];
890 resv_map->pages_per_hpage = pages_per_huge_page(h);
891 resv_map->css = &h_cg->css;
892 }
893#endif
894}
895
896struct resv_map *resv_map_alloc(void)
897{
898 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
899 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
900
901 if (!resv_map || !rg) {
902 kfree(resv_map);
903 kfree(rg);
904 return NULL;
905 }
906
907 kref_init(&resv_map->refs);
908 spin_lock_init(&resv_map->lock);
909 INIT_LIST_HEAD(&resv_map->regions);
910
911 resv_map->adds_in_progress = 0;
912 /*
913 * Initialize these to 0. On shared mappings, 0's here indicate these
914 * fields don't do cgroup accounting. On private mappings, these will be
915 * re-initialized to the proper values, to indicate that hugetlb cgroup
916 * reservations are to be un-charged from here.
917 */
918 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
919
920 INIT_LIST_HEAD(&resv_map->region_cache);
921 list_add(&rg->link, &resv_map->region_cache);
922 resv_map->region_cache_count = 1;
923
924 return resv_map;
925}
926
927void resv_map_release(struct kref *ref)
928{
929 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
930 struct list_head *head = &resv_map->region_cache;
931 struct file_region *rg, *trg;
932
933 /* Clear out any active regions before we release the map. */
934 region_del(resv_map, 0, LONG_MAX);
935
936 /* ... and any entries left in the cache */
937 list_for_each_entry_safe(rg, trg, head, link) {
938 list_del(&rg->link);
939 kfree(rg);
940 }
941
942 VM_BUG_ON(resv_map->adds_in_progress);
943
944 kfree(resv_map);
945}
946
947static inline struct resv_map *inode_resv_map(struct inode *inode)
948{
949 /*
950 * At inode evict time, i_mapping may not point to the original
951 * address space within the inode. This original address space
952 * contains the pointer to the resv_map. So, always use the
953 * address space embedded within the inode.
954 * The VERY common case is inode->mapping == &inode->i_data but,
955 * this may not be true for device special inodes.
956 */
957 return (struct resv_map *)(&inode->i_data)->private_data;
958}
959
960static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
961{
962 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
963 if (vma->vm_flags & VM_MAYSHARE) {
964 struct address_space *mapping = vma->vm_file->f_mapping;
965 struct inode *inode = mapping->host;
966
967 return inode_resv_map(inode);
968
969 } else {
970 return (struct resv_map *)(get_vma_private_data(vma) &
971 ~HPAGE_RESV_MASK);
972 }
973}
974
975static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
976{
977 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
978 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
979
980 set_vma_private_data(vma, (get_vma_private_data(vma) &
981 HPAGE_RESV_MASK) | (unsigned long)map);
982}
983
984static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
985{
986 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
987 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
988
989 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
990}
991
992static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
993{
994 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
995
996 return (get_vma_private_data(vma) & flag) != 0;
997}
998
999/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
1000void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
1001{
1002 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1003 if (!(vma->vm_flags & VM_MAYSHARE))
1004 vma->vm_private_data = (void *)0;
1005}
1006
1007/* Returns true if the VMA has associated reserve pages */
1008static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1009{
1010 if (vma->vm_flags & VM_NORESERVE) {
1011 /*
1012 * This address is already reserved by other process(chg == 0),
1013 * so, we should decrement reserved count. Without decrementing,
1014 * reserve count remains after releasing inode, because this
1015 * allocated page will go into page cache and is regarded as
1016 * coming from reserved pool in releasing step. Currently, we
1017 * don't have any other solution to deal with this situation
1018 * properly, so add work-around here.
1019 */
1020 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1021 return true;
1022 else
1023 return false;
1024 }
1025
1026 /* Shared mappings always use reserves */
1027 if (vma->vm_flags & VM_MAYSHARE) {
1028 /*
1029 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1030 * be a region map for all pages. The only situation where
1031 * there is no region map is if a hole was punched via
1032 * fallocate. In this case, there really are no reserves to
1033 * use. This situation is indicated if chg != 0.
1034 */
1035 if (chg)
1036 return false;
1037 else
1038 return true;
1039 }
1040
1041 /*
1042 * Only the process that called mmap() has reserves for
1043 * private mappings.
1044 */
1045 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1046 /*
1047 * Like the shared case above, a hole punch or truncate
1048 * could have been performed on the private mapping.
1049 * Examine the value of chg to determine if reserves
1050 * actually exist or were previously consumed.
1051 * Very Subtle - The value of chg comes from a previous
1052 * call to vma_needs_reserves(). The reserve map for
1053 * private mappings has different (opposite) semantics
1054 * than that of shared mappings. vma_needs_reserves()
1055 * has already taken this difference in semantics into
1056 * account. Therefore, the meaning of chg is the same
1057 * as in the shared case above. Code could easily be
1058 * combined, but keeping it separate draws attention to
1059 * subtle differences.
1060 */
1061 if (chg)
1062 return false;
1063 else
1064 return true;
1065 }
1066
1067 return false;
1068}
1069
1070static void enqueue_huge_page(struct hstate *h, struct page *page)
1071{
1072 int nid = page_to_nid(page);
1073
1074 lockdep_assert_held(&hugetlb_lock);
1075 list_move(&page->lru, &h->hugepage_freelists[nid]);
1076 h->free_huge_pages++;
1077 h->free_huge_pages_node[nid]++;
1078 SetHPageFreed(page);
1079}
1080
1081static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1082{
1083 struct page *page;
1084 bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1085
1086 lockdep_assert_held(&hugetlb_lock);
1087 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1088 if (pin && !is_pinnable_page(page))
1089 continue;
1090
1091 if (PageHWPoison(page))
1092 continue;
1093
1094 list_move(&page->lru, &h->hugepage_activelist);
1095 set_page_refcounted(page);
1096 ClearHPageFreed(page);
1097 h->free_huge_pages--;
1098 h->free_huge_pages_node[nid]--;
1099 return page;
1100 }
1101
1102 return NULL;
1103}
1104
1105static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1106 nodemask_t *nmask)
1107{
1108 unsigned int cpuset_mems_cookie;
1109 struct zonelist *zonelist;
1110 struct zone *zone;
1111 struct zoneref *z;
1112 int node = NUMA_NO_NODE;
1113
1114 zonelist = node_zonelist(nid, gfp_mask);
1115
1116retry_cpuset:
1117 cpuset_mems_cookie = read_mems_allowed_begin();
1118 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1119 struct page *page;
1120
1121 if (!cpuset_zone_allowed(zone, gfp_mask))
1122 continue;
1123 /*
1124 * no need to ask again on the same node. Pool is node rather than
1125 * zone aware
1126 */
1127 if (zone_to_nid(zone) == node)
1128 continue;
1129 node = zone_to_nid(zone);
1130
1131 page = dequeue_huge_page_node_exact(h, node);
1132 if (page)
1133 return page;
1134 }
1135 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1136 goto retry_cpuset;
1137
1138 return NULL;
1139}
1140
1141static struct page *dequeue_huge_page_vma(struct hstate *h,
1142 struct vm_area_struct *vma,
1143 unsigned long address, int avoid_reserve,
1144 long chg)
1145{
1146 struct page *page;
1147 struct mempolicy *mpol;
1148 gfp_t gfp_mask;
1149 nodemask_t *nodemask;
1150 int nid;
1151
1152 /*
1153 * A child process with MAP_PRIVATE mappings created by their parent
1154 * have no page reserves. This check ensures that reservations are
1155 * not "stolen". The child may still get SIGKILLed
1156 */
1157 if (!vma_has_reserves(vma, chg) &&
1158 h->free_huge_pages - h->resv_huge_pages == 0)
1159 goto err;
1160
1161 /* If reserves cannot be used, ensure enough pages are in the pool */
1162 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1163 goto err;
1164
1165 gfp_mask = htlb_alloc_mask(h);
1166 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1167 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1168 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1169 SetHPageRestoreReserve(page);
1170 h->resv_huge_pages--;
1171 }
1172
1173 mpol_cond_put(mpol);
1174 return page;
1175
1176err:
1177 return NULL;
1178}
1179
1180/*
1181 * common helper functions for hstate_next_node_to_{alloc|free}.
1182 * We may have allocated or freed a huge page based on a different
1183 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1184 * be outside of *nodes_allowed. Ensure that we use an allowed
1185 * node for alloc or free.
1186 */
1187static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1188{
1189 nid = next_node_in(nid, *nodes_allowed);
1190 VM_BUG_ON(nid >= MAX_NUMNODES);
1191
1192 return nid;
1193}
1194
1195static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1196{
1197 if (!node_isset(nid, *nodes_allowed))
1198 nid = next_node_allowed(nid, nodes_allowed);
1199 return nid;
1200}
1201
1202/*
1203 * returns the previously saved node ["this node"] from which to
1204 * allocate a persistent huge page for the pool and advance the
1205 * next node from which to allocate, handling wrap at end of node
1206 * mask.
1207 */
1208static int hstate_next_node_to_alloc(struct hstate *h,
1209 nodemask_t *nodes_allowed)
1210{
1211 int nid;
1212
1213 VM_BUG_ON(!nodes_allowed);
1214
1215 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1216 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1217
1218 return nid;
1219}
1220
1221/*
1222 * helper for remove_pool_huge_page() - return the previously saved
1223 * node ["this node"] from which to free a huge page. Advance the
1224 * next node id whether or not we find a free huge page to free so
1225 * that the next attempt to free addresses the next node.
1226 */
1227static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1228{
1229 int nid;
1230
1231 VM_BUG_ON(!nodes_allowed);
1232
1233 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1234 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1235
1236 return nid;
1237}
1238
1239#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1240 for (nr_nodes = nodes_weight(*mask); \
1241 nr_nodes > 0 && \
1242 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1243 nr_nodes--)
1244
1245#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1246 for (nr_nodes = nodes_weight(*mask); \
1247 nr_nodes > 0 && \
1248 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1249 nr_nodes--)
1250
1251#ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1252static void destroy_compound_gigantic_page(struct page *page,
1253 unsigned int order)
1254{
1255 int i;
1256 int nr_pages = 1 << order;
1257 struct page *p = page + 1;
1258
1259 atomic_set(compound_mapcount_ptr(page), 0);
1260 atomic_set(compound_pincount_ptr(page), 0);
1261
1262 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1263 clear_compound_head(p);
1264 set_page_refcounted(p);
1265 }
1266
1267 set_compound_order(page, 0);
1268 page[1].compound_nr = 0;
1269 __ClearPageHead(page);
1270}
1271
1272static void free_gigantic_page(struct page *page, unsigned int order)
1273{
1274 /*
1275 * If the page isn't allocated using the cma allocator,
1276 * cma_release() returns false.
1277 */
1278#ifdef CONFIG_CMA
1279 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1280 return;
1281#endif
1282
1283 free_contig_range(page_to_pfn(page), 1 << order);
1284}
1285
1286#ifdef CONFIG_CONTIG_ALLOC
1287static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1288 int nid, nodemask_t *nodemask)
1289{
1290 unsigned long nr_pages = pages_per_huge_page(h);
1291 if (nid == NUMA_NO_NODE)
1292 nid = numa_mem_id();
1293
1294#ifdef CONFIG_CMA
1295 {
1296 struct page *page;
1297 int node;
1298
1299 if (hugetlb_cma[nid]) {
1300 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1301 huge_page_order(h), true);
1302 if (page)
1303 return page;
1304 }
1305
1306 if (!(gfp_mask & __GFP_THISNODE)) {
1307 for_each_node_mask(node, *nodemask) {
1308 if (node == nid || !hugetlb_cma[node])
1309 continue;
1310
1311 page = cma_alloc(hugetlb_cma[node], nr_pages,
1312 huge_page_order(h), true);
1313 if (page)
1314 return page;
1315 }
1316 }
1317 }
1318#endif
1319
1320 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1321}
1322
1323#else /* !CONFIG_CONTIG_ALLOC */
1324static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1325 int nid, nodemask_t *nodemask)
1326{
1327 return NULL;
1328}
1329#endif /* CONFIG_CONTIG_ALLOC */
1330
1331#else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1332static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1333 int nid, nodemask_t *nodemask)
1334{
1335 return NULL;
1336}
1337static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1338static inline void destroy_compound_gigantic_page(struct page *page,
1339 unsigned int order) { }
1340#endif
1341
1342/*
1343 * Remove hugetlb page from lists, and update dtor so that page appears
1344 * as just a compound page. A reference is held on the page.
1345 *
1346 * Must be called with hugetlb lock held.
1347 */
1348static void remove_hugetlb_page(struct hstate *h, struct page *page,
1349 bool adjust_surplus)
1350{
1351 int nid = page_to_nid(page);
1352
1353 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1354 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1355
1356 lockdep_assert_held(&hugetlb_lock);
1357 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1358 return;
1359
1360 list_del(&page->lru);
1361
1362 if (HPageFreed(page)) {
1363 h->free_huge_pages--;
1364 h->free_huge_pages_node[nid]--;
1365 }
1366 if (adjust_surplus) {
1367 h->surplus_huge_pages--;
1368 h->surplus_huge_pages_node[nid]--;
1369 }
1370
1371 set_page_refcounted(page);
1372 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1373
1374 h->nr_huge_pages--;
1375 h->nr_huge_pages_node[nid]--;
1376}
1377
1378static void add_hugetlb_page(struct hstate *h, struct page *page,
1379 bool adjust_surplus)
1380{
1381 int zeroed;
1382 int nid = page_to_nid(page);
1383
1384 VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page);
1385
1386 lockdep_assert_held(&hugetlb_lock);
1387
1388 INIT_LIST_HEAD(&page->lru);
1389 h->nr_huge_pages++;
1390 h->nr_huge_pages_node[nid]++;
1391
1392 if (adjust_surplus) {
1393 h->surplus_huge_pages++;
1394 h->surplus_huge_pages_node[nid]++;
1395 }
1396
1397 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1398 set_page_private(page, 0);
1399 SetHPageVmemmapOptimized(page);
1400
1401 /*
1402 * This page is now managed by the hugetlb allocator and has
1403 * no users -- drop the last reference.
1404 */
1405 zeroed = put_page_testzero(page);
1406 VM_BUG_ON_PAGE(!zeroed, page);
1407 arch_clear_hugepage_flags(page);
1408 enqueue_huge_page(h, page);
1409}
1410
1411static void __update_and_free_page(struct hstate *h, struct page *page)
1412{
1413 int i;
1414 struct page *subpage = page;
1415
1416 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1417 return;
1418
1419 if (alloc_huge_page_vmemmap(h, page)) {
1420 spin_lock_irq(&hugetlb_lock);
1421 /*
1422 * If we cannot allocate vmemmap pages, just refuse to free the
1423 * page and put the page back on the hugetlb free list and treat
1424 * as a surplus page.
1425 */
1426 add_hugetlb_page(h, page, true);
1427 spin_unlock_irq(&hugetlb_lock);
1428 return;
1429 }
1430
1431 for (i = 0; i < pages_per_huge_page(h);
1432 i++, subpage = mem_map_next(subpage, page, i)) {
1433 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1434 1 << PG_referenced | 1 << PG_dirty |
1435 1 << PG_active | 1 << PG_private |
1436 1 << PG_writeback);
1437 }
1438 if (hstate_is_gigantic(h)) {
1439 destroy_compound_gigantic_page(page, huge_page_order(h));
1440 free_gigantic_page(page, huge_page_order(h));
1441 } else {
1442 __free_pages(page, huge_page_order(h));
1443 }
1444}
1445
1446/*
1447 * As update_and_free_page() can be called under any context, so we cannot
1448 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1449 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1450 * the vmemmap pages.
1451 *
1452 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1453 * freed and frees them one-by-one. As the page->mapping pointer is going
1454 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1455 * structure of a lockless linked list of huge pages to be freed.
1456 */
1457static LLIST_HEAD(hpage_freelist);
1458
1459static void free_hpage_workfn(struct work_struct *work)
1460{
1461 struct llist_node *node;
1462
1463 node = llist_del_all(&hpage_freelist);
1464
1465 while (node) {
1466 struct page *page;
1467 struct hstate *h;
1468
1469 page = container_of((struct address_space **)node,
1470 struct page, mapping);
1471 node = node->next;
1472 page->mapping = NULL;
1473 /*
1474 * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate()
1475 * is going to trigger because a previous call to
1476 * remove_hugetlb_page() will set_compound_page_dtor(page,
1477 * NULL_COMPOUND_DTOR), so do not use page_hstate() directly.
1478 */
1479 h = size_to_hstate(page_size(page));
1480
1481 __update_and_free_page(h, page);
1482
1483 cond_resched();
1484 }
1485}
1486static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1487
1488static inline void flush_free_hpage_work(struct hstate *h)
1489{
1490 if (free_vmemmap_pages_per_hpage(h))
1491 flush_work(&free_hpage_work);
1492}
1493
1494static void update_and_free_page(struct hstate *h, struct page *page,
1495 bool atomic)
1496{
1497 if (!HPageVmemmapOptimized(page) || !atomic) {
1498 __update_and_free_page(h, page);
1499 return;
1500 }
1501
1502 /*
1503 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1504 *
1505 * Only call schedule_work() if hpage_freelist is previously
1506 * empty. Otherwise, schedule_work() had been called but the workfn
1507 * hasn't retrieved the list yet.
1508 */
1509 if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist))
1510 schedule_work(&free_hpage_work);
1511}
1512
1513static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1514{
1515 struct page *page, *t_page;
1516
1517 list_for_each_entry_safe(page, t_page, list, lru) {
1518 update_and_free_page(h, page, false);
1519 cond_resched();
1520 }
1521}
1522
1523struct hstate *size_to_hstate(unsigned long size)
1524{
1525 struct hstate *h;
1526
1527 for_each_hstate(h) {
1528 if (huge_page_size(h) == size)
1529 return h;
1530 }
1531 return NULL;
1532}
1533
1534void free_huge_page(struct page *page)
1535{
1536 /*
1537 * Can't pass hstate in here because it is called from the
1538 * compound page destructor.
1539 */
1540 struct hstate *h = page_hstate(page);
1541 int nid = page_to_nid(page);
1542 struct hugepage_subpool *spool = hugetlb_page_subpool(page);
1543 bool restore_reserve;
1544 unsigned long flags;
1545
1546 VM_BUG_ON_PAGE(page_count(page), page);
1547 VM_BUG_ON_PAGE(page_mapcount(page), page);
1548
1549 hugetlb_set_page_subpool(page, NULL);
1550 page->mapping = NULL;
1551 restore_reserve = HPageRestoreReserve(page);
1552 ClearHPageRestoreReserve(page);
1553
1554 /*
1555 * If HPageRestoreReserve was set on page, page allocation consumed a
1556 * reservation. If the page was associated with a subpool, there
1557 * would have been a page reserved in the subpool before allocation
1558 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1559 * reservation, do not call hugepage_subpool_put_pages() as this will
1560 * remove the reserved page from the subpool.
1561 */
1562 if (!restore_reserve) {
1563 /*
1564 * A return code of zero implies that the subpool will be
1565 * under its minimum size if the reservation is not restored
1566 * after page is free. Therefore, force restore_reserve
1567 * operation.
1568 */
1569 if (hugepage_subpool_put_pages(spool, 1) == 0)
1570 restore_reserve = true;
1571 }
1572
1573 spin_lock_irqsave(&hugetlb_lock, flags);
1574 ClearHPageMigratable(page);
1575 hugetlb_cgroup_uncharge_page(hstate_index(h),
1576 pages_per_huge_page(h), page);
1577 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1578 pages_per_huge_page(h), page);
1579 if (restore_reserve)
1580 h->resv_huge_pages++;
1581
1582 if (HPageTemporary(page)) {
1583 remove_hugetlb_page(h, page, false);
1584 spin_unlock_irqrestore(&hugetlb_lock, flags);
1585 update_and_free_page(h, page, true);
1586 } else if (h->surplus_huge_pages_node[nid]) {
1587 /* remove the page from active list */
1588 remove_hugetlb_page(h, page, true);
1589 spin_unlock_irqrestore(&hugetlb_lock, flags);
1590 update_and_free_page(h, page, true);
1591 } else {
1592 arch_clear_hugepage_flags(page);
1593 enqueue_huge_page(h, page);
1594 spin_unlock_irqrestore(&hugetlb_lock, flags);
1595 }
1596}
1597
1598/*
1599 * Must be called with the hugetlb lock held
1600 */
1601static void __prep_account_new_huge_page(struct hstate *h, int nid)
1602{
1603 lockdep_assert_held(&hugetlb_lock);
1604 h->nr_huge_pages++;
1605 h->nr_huge_pages_node[nid]++;
1606}
1607
1608static void __prep_new_huge_page(struct hstate *h, struct page *page)
1609{
1610 free_huge_page_vmemmap(h, page);
1611 INIT_LIST_HEAD(&page->lru);
1612 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1613 hugetlb_set_page_subpool(page, NULL);
1614 set_hugetlb_cgroup(page, NULL);
1615 set_hugetlb_cgroup_rsvd(page, NULL);
1616}
1617
1618static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1619{
1620 __prep_new_huge_page(h, page);
1621 spin_lock_irq(&hugetlb_lock);
1622 __prep_account_new_huge_page(h, nid);
1623 spin_unlock_irq(&hugetlb_lock);
1624}
1625
1626static bool prep_compound_gigantic_page(struct page *page, unsigned int order)
1627{
1628 int i, j;
1629 int nr_pages = 1 << order;
1630 struct page *p = page + 1;
1631
1632 /* we rely on prep_new_huge_page to set the destructor */
1633 set_compound_order(page, order);
1634 __ClearPageReserved(page);
1635 __SetPageHead(page);
1636 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1637 /*
1638 * For gigantic hugepages allocated through bootmem at
1639 * boot, it's safer to be consistent with the not-gigantic
1640 * hugepages and clear the PG_reserved bit from all tail pages
1641 * too. Otherwise drivers using get_user_pages() to access tail
1642 * pages may get the reference counting wrong if they see
1643 * PG_reserved set on a tail page (despite the head page not
1644 * having PG_reserved set). Enforcing this consistency between
1645 * head and tail pages allows drivers to optimize away a check
1646 * on the head page when they need know if put_page() is needed
1647 * after get_user_pages().
1648 */
1649 __ClearPageReserved(p);
1650 /*
1651 * Subtle and very unlikely
1652 *
1653 * Gigantic 'page allocators' such as memblock or cma will
1654 * return a set of pages with each page ref counted. We need
1655 * to turn this set of pages into a compound page with tail
1656 * page ref counts set to zero. Code such as speculative page
1657 * cache adding could take a ref on a 'to be' tail page.
1658 * We need to respect any increased ref count, and only set
1659 * the ref count to zero if count is currently 1. If count
1660 * is not 1, we call synchronize_rcu in the hope that a rcu
1661 * grace period will cause ref count to drop and then retry.
1662 * If count is still inflated on retry we return an error and
1663 * must discard the pages.
1664 */
1665 if (!page_ref_freeze(p, 1)) {
1666 pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n");
1667 synchronize_rcu();
1668 if (!page_ref_freeze(p, 1))
1669 goto out_error;
1670 }
1671 set_page_count(p, 0);
1672 set_compound_head(p, page);
1673 }
1674 atomic_set(compound_mapcount_ptr(page), -1);
1675 atomic_set(compound_pincount_ptr(page), 0);
1676 return true;
1677
1678out_error:
1679 /* undo tail page modifications made above */
1680 p = page + 1;
1681 for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) {
1682 clear_compound_head(p);
1683 set_page_refcounted(p);
1684 }
1685 /* need to clear PG_reserved on remaining tail pages */
1686 for (; j < nr_pages; j++, p = mem_map_next(p, page, j))
1687 __ClearPageReserved(p);
1688 set_compound_order(page, 0);
1689 page[1].compound_nr = 0;
1690 __ClearPageHead(page);
1691 return false;
1692}
1693
1694/*
1695 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1696 * transparent huge pages. See the PageTransHuge() documentation for more
1697 * details.
1698 */
1699int PageHuge(struct page *page)
1700{
1701 if (!PageCompound(page))
1702 return 0;
1703
1704 page = compound_head(page);
1705 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1706}
1707EXPORT_SYMBOL_GPL(PageHuge);
1708
1709/*
1710 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1711 * normal or transparent huge pages.
1712 */
1713int PageHeadHuge(struct page *page_head)
1714{
1715 if (!PageHead(page_head))
1716 return 0;
1717
1718 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1719}
1720
1721/*
1722 * Find and lock address space (mapping) in write mode.
1723 *
1724 * Upon entry, the page is locked which means that page_mapping() is
1725 * stable. Due to locking order, we can only trylock_write. If we can
1726 * not get the lock, simply return NULL to caller.
1727 */
1728struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1729{
1730 struct address_space *mapping = page_mapping(hpage);
1731
1732 if (!mapping)
1733 return mapping;
1734
1735 if (i_mmap_trylock_write(mapping))
1736 return mapping;
1737
1738 return NULL;
1739}
1740
1741pgoff_t hugetlb_basepage_index(struct page *page)
1742{
1743 struct page *page_head = compound_head(page);
1744 pgoff_t index = page_index(page_head);
1745 unsigned long compound_idx;
1746
1747 if (compound_order(page_head) >= MAX_ORDER)
1748 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1749 else
1750 compound_idx = page - page_head;
1751
1752 return (index << compound_order(page_head)) + compound_idx;
1753}
1754
1755static struct page *alloc_buddy_huge_page(struct hstate *h,
1756 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1757 nodemask_t *node_alloc_noretry)
1758{
1759 int order = huge_page_order(h);
1760 struct page *page;
1761 bool alloc_try_hard = true;
1762
1763 /*
1764 * By default we always try hard to allocate the page with
1765 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1766 * a loop (to adjust global huge page counts) and previous allocation
1767 * failed, do not continue to try hard on the same node. Use the
1768 * node_alloc_noretry bitmap to manage this state information.
1769 */
1770 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1771 alloc_try_hard = false;
1772 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1773 if (alloc_try_hard)
1774 gfp_mask |= __GFP_RETRY_MAYFAIL;
1775 if (nid == NUMA_NO_NODE)
1776 nid = numa_mem_id();
1777 page = __alloc_pages(gfp_mask, order, nid, nmask);
1778 if (page)
1779 __count_vm_event(HTLB_BUDDY_PGALLOC);
1780 else
1781 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1782
1783 /*
1784 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1785 * indicates an overall state change. Clear bit so that we resume
1786 * normal 'try hard' allocations.
1787 */
1788 if (node_alloc_noretry && page && !alloc_try_hard)
1789 node_clear(nid, *node_alloc_noretry);
1790
1791 /*
1792 * If we tried hard to get a page but failed, set bit so that
1793 * subsequent attempts will not try as hard until there is an
1794 * overall state change.
1795 */
1796 if (node_alloc_noretry && !page && alloc_try_hard)
1797 node_set(nid, *node_alloc_noretry);
1798
1799 return page;
1800}
1801
1802/*
1803 * Common helper to allocate a fresh hugetlb page. All specific allocators
1804 * should use this function to get new hugetlb pages
1805 */
1806static struct page *alloc_fresh_huge_page(struct hstate *h,
1807 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1808 nodemask_t *node_alloc_noretry)
1809{
1810 struct page *page;
1811 bool retry = false;
1812
1813retry:
1814 if (hstate_is_gigantic(h))
1815 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1816 else
1817 page = alloc_buddy_huge_page(h, gfp_mask,
1818 nid, nmask, node_alloc_noretry);
1819 if (!page)
1820 return NULL;
1821
1822 if (hstate_is_gigantic(h)) {
1823 if (!prep_compound_gigantic_page(page, huge_page_order(h))) {
1824 /*
1825 * Rare failure to convert pages to compound page.
1826 * Free pages and try again - ONCE!
1827 */
1828 free_gigantic_page(page, huge_page_order(h));
1829 if (!retry) {
1830 retry = true;
1831 goto retry;
1832 }
1833 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
1834 return NULL;
1835 }
1836 }
1837 prep_new_huge_page(h, page, page_to_nid(page));
1838
1839 return page;
1840}
1841
1842/*
1843 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1844 * manner.
1845 */
1846static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1847 nodemask_t *node_alloc_noretry)
1848{
1849 struct page *page;
1850 int nr_nodes, node;
1851 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1852
1853 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1854 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1855 node_alloc_noretry);
1856 if (page)
1857 break;
1858 }
1859
1860 if (!page)
1861 return 0;
1862
1863 put_page(page); /* free it into the hugepage allocator */
1864
1865 return 1;
1866}
1867
1868/*
1869 * Remove huge page from pool from next node to free. Attempt to keep
1870 * persistent huge pages more or less balanced over allowed nodes.
1871 * This routine only 'removes' the hugetlb page. The caller must make
1872 * an additional call to free the page to low level allocators.
1873 * Called with hugetlb_lock locked.
1874 */
1875static struct page *remove_pool_huge_page(struct hstate *h,
1876 nodemask_t *nodes_allowed,
1877 bool acct_surplus)
1878{
1879 int nr_nodes, node;
1880 struct page *page = NULL;
1881
1882 lockdep_assert_held(&hugetlb_lock);
1883 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1884 /*
1885 * If we're returning unused surplus pages, only examine
1886 * nodes with surplus pages.
1887 */
1888 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1889 !list_empty(&h->hugepage_freelists[node])) {
1890 page = list_entry(h->hugepage_freelists[node].next,
1891 struct page, lru);
1892 remove_hugetlb_page(h, page, acct_surplus);
1893 break;
1894 }
1895 }
1896
1897 return page;
1898}
1899
1900/*
1901 * Dissolve a given free hugepage into free buddy pages. This function does
1902 * nothing for in-use hugepages and non-hugepages.
1903 * This function returns values like below:
1904 *
1905 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
1906 * when the system is under memory pressure and the feature of
1907 * freeing unused vmemmap pages associated with each hugetlb page
1908 * is enabled.
1909 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1910 * (allocated or reserved.)
1911 * 0: successfully dissolved free hugepages or the page is not a
1912 * hugepage (considered as already dissolved)
1913 */
1914int dissolve_free_huge_page(struct page *page)
1915{
1916 int rc = -EBUSY;
1917
1918retry:
1919 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1920 if (!PageHuge(page))
1921 return 0;
1922
1923 spin_lock_irq(&hugetlb_lock);
1924 if (!PageHuge(page)) {
1925 rc = 0;
1926 goto out;
1927 }
1928
1929 if (!page_count(page)) {
1930 struct page *head = compound_head(page);
1931 struct hstate *h = page_hstate(head);
1932 if (h->free_huge_pages - h->resv_huge_pages == 0)
1933 goto out;
1934
1935 /*
1936 * We should make sure that the page is already on the free list
1937 * when it is dissolved.
1938 */
1939 if (unlikely(!HPageFreed(head))) {
1940 spin_unlock_irq(&hugetlb_lock);
1941 cond_resched();
1942
1943 /*
1944 * Theoretically, we should return -EBUSY when we
1945 * encounter this race. In fact, we have a chance
1946 * to successfully dissolve the page if we do a
1947 * retry. Because the race window is quite small.
1948 * If we seize this opportunity, it is an optimization
1949 * for increasing the success rate of dissolving page.
1950 */
1951 goto retry;
1952 }
1953
1954 remove_hugetlb_page(h, head, false);
1955 h->max_huge_pages--;
1956 spin_unlock_irq(&hugetlb_lock);
1957
1958 /*
1959 * Normally update_and_free_page will allocate required vmemmmap
1960 * before freeing the page. update_and_free_page will fail to
1961 * free the page if it can not allocate required vmemmap. We
1962 * need to adjust max_huge_pages if the page is not freed.
1963 * Attempt to allocate vmemmmap here so that we can take
1964 * appropriate action on failure.
1965 */
1966 rc = alloc_huge_page_vmemmap(h, head);
1967 if (!rc) {
1968 /*
1969 * Move PageHWPoison flag from head page to the raw
1970 * error page, which makes any subpages rather than
1971 * the error page reusable.
1972 */
1973 if (PageHWPoison(head) && page != head) {
1974 SetPageHWPoison(page);
1975 ClearPageHWPoison(head);
1976 }
1977 update_and_free_page(h, head, false);
1978 } else {
1979 spin_lock_irq(&hugetlb_lock);
1980 add_hugetlb_page(h, head, false);
1981 h->max_huge_pages++;
1982 spin_unlock_irq(&hugetlb_lock);
1983 }
1984
1985 return rc;
1986 }
1987out:
1988 spin_unlock_irq(&hugetlb_lock);
1989 return rc;
1990}
1991
1992/*
1993 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1994 * make specified memory blocks removable from the system.
1995 * Note that this will dissolve a free gigantic hugepage completely, if any
1996 * part of it lies within the given range.
1997 * Also note that if dissolve_free_huge_page() returns with an error, all
1998 * free hugepages that were dissolved before that error are lost.
1999 */
2000int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2001{
2002 unsigned long pfn;
2003 struct page *page;
2004 int rc = 0;
2005
2006 if (!hugepages_supported())
2007 return rc;
2008
2009 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
2010 page = pfn_to_page(pfn);
2011 rc = dissolve_free_huge_page(page);
2012 if (rc)
2013 break;
2014 }
2015
2016 return rc;
2017}
2018
2019/*
2020 * Allocates a fresh surplus page from the page allocator.
2021 */
2022static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
2023 int nid, nodemask_t *nmask)
2024{
2025 struct page *page = NULL;
2026
2027 if (hstate_is_gigantic(h))
2028 return NULL;
2029
2030 spin_lock_irq(&hugetlb_lock);
2031 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2032 goto out_unlock;
2033 spin_unlock_irq(&hugetlb_lock);
2034
2035 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2036 if (!page)
2037 return NULL;
2038
2039 spin_lock_irq(&hugetlb_lock);
2040 /*
2041 * We could have raced with the pool size change.
2042 * Double check that and simply deallocate the new page
2043 * if we would end up overcommiting the surpluses. Abuse
2044 * temporary page to workaround the nasty free_huge_page
2045 * codeflow
2046 */
2047 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2048 SetHPageTemporary(page);
2049 spin_unlock_irq(&hugetlb_lock);
2050 put_page(page);
2051 return NULL;
2052 } else {
2053 h->surplus_huge_pages++;
2054 h->surplus_huge_pages_node[page_to_nid(page)]++;
2055 }
2056
2057out_unlock:
2058 spin_unlock_irq(&hugetlb_lock);
2059
2060 return page;
2061}
2062
2063static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
2064 int nid, nodemask_t *nmask)
2065{
2066 struct page *page;
2067
2068 if (hstate_is_gigantic(h))
2069 return NULL;
2070
2071 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2072 if (!page)
2073 return NULL;
2074
2075 /*
2076 * We do not account these pages as surplus because they are only
2077 * temporary and will be released properly on the last reference
2078 */
2079 SetHPageTemporary(page);
2080
2081 return page;
2082}
2083
2084/*
2085 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2086 */
2087static
2088struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
2089 struct vm_area_struct *vma, unsigned long addr)
2090{
2091 struct page *page;
2092 struct mempolicy *mpol;
2093 gfp_t gfp_mask = htlb_alloc_mask(h);
2094 int nid;
2095 nodemask_t *nodemask;
2096
2097 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2098 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
2099 mpol_cond_put(mpol);
2100
2101 return page;
2102}
2103
2104/* page migration callback function */
2105struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
2106 nodemask_t *nmask, gfp_t gfp_mask)
2107{
2108 spin_lock_irq(&hugetlb_lock);
2109 if (h->free_huge_pages - h->resv_huge_pages > 0) {
2110 struct page *page;
2111
2112 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
2113 if (page) {
2114 spin_unlock_irq(&hugetlb_lock);
2115 return page;
2116 }
2117 }
2118 spin_unlock_irq(&hugetlb_lock);
2119
2120 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2121}
2122
2123/* mempolicy aware migration callback */
2124struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2125 unsigned long address)
2126{
2127 struct mempolicy *mpol;
2128 nodemask_t *nodemask;
2129 struct page *page;
2130 gfp_t gfp_mask;
2131 int node;
2132
2133 gfp_mask = htlb_alloc_mask(h);
2134 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2135 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
2136 mpol_cond_put(mpol);
2137
2138 return page;
2139}
2140
2141/*
2142 * Increase the hugetlb pool such that it can accommodate a reservation
2143 * of size 'delta'.
2144 */
2145static int gather_surplus_pages(struct hstate *h, long delta)
2146 __must_hold(&hugetlb_lock)
2147{
2148 struct list_head surplus_list;
2149 struct page *page, *tmp;
2150 int ret;
2151 long i;
2152 long needed, allocated;
2153 bool alloc_ok = true;
2154
2155 lockdep_assert_held(&hugetlb_lock);
2156 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2157 if (needed <= 0) {
2158 h->resv_huge_pages += delta;
2159 return 0;
2160 }
2161
2162 allocated = 0;
2163 INIT_LIST_HEAD(&surplus_list);
2164
2165 ret = -ENOMEM;
2166retry:
2167 spin_unlock_irq(&hugetlb_lock);
2168 for (i = 0; i < needed; i++) {
2169 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2170 NUMA_NO_NODE, NULL);
2171 if (!page) {
2172 alloc_ok = false;
2173 break;
2174 }
2175 list_add(&page->lru, &surplus_list);
2176 cond_resched();
2177 }
2178 allocated += i;
2179
2180 /*
2181 * After retaking hugetlb_lock, we need to recalculate 'needed'
2182 * because either resv_huge_pages or free_huge_pages may have changed.
2183 */
2184 spin_lock_irq(&hugetlb_lock);
2185 needed = (h->resv_huge_pages + delta) -
2186 (h->free_huge_pages + allocated);
2187 if (needed > 0) {
2188 if (alloc_ok)
2189 goto retry;
2190 /*
2191 * We were not able to allocate enough pages to
2192 * satisfy the entire reservation so we free what
2193 * we've allocated so far.
2194 */
2195 goto free;
2196 }
2197 /*
2198 * The surplus_list now contains _at_least_ the number of extra pages
2199 * needed to accommodate the reservation. Add the appropriate number
2200 * of pages to the hugetlb pool and free the extras back to the buddy
2201 * allocator. Commit the entire reservation here to prevent another
2202 * process from stealing the pages as they are added to the pool but
2203 * before they are reserved.
2204 */
2205 needed += allocated;
2206 h->resv_huge_pages += delta;
2207 ret = 0;
2208
2209 /* Free the needed pages to the hugetlb pool */
2210 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2211 int zeroed;
2212
2213 if ((--needed) < 0)
2214 break;
2215 /*
2216 * This page is now managed by the hugetlb allocator and has
2217 * no users -- drop the buddy allocator's reference.
2218 */
2219 zeroed = put_page_testzero(page);
2220 VM_BUG_ON_PAGE(!zeroed, page);
2221 enqueue_huge_page(h, page);
2222 }
2223free:
2224 spin_unlock_irq(&hugetlb_lock);
2225
2226 /* Free unnecessary surplus pages to the buddy allocator */
2227 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2228 put_page(page);
2229 spin_lock_irq(&hugetlb_lock);
2230
2231 return ret;
2232}
2233
2234/*
2235 * This routine has two main purposes:
2236 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2237 * in unused_resv_pages. This corresponds to the prior adjustments made
2238 * to the associated reservation map.
2239 * 2) Free any unused surplus pages that may have been allocated to satisfy
2240 * the reservation. As many as unused_resv_pages may be freed.
2241 */
2242static void return_unused_surplus_pages(struct hstate *h,
2243 unsigned long unused_resv_pages)
2244{
2245 unsigned long nr_pages;
2246 struct page *page;
2247 LIST_HEAD(page_list);
2248
2249 lockdep_assert_held(&hugetlb_lock);
2250 /* Uncommit the reservation */
2251 h->resv_huge_pages -= unused_resv_pages;
2252
2253 /* Cannot return gigantic pages currently */
2254 if (hstate_is_gigantic(h))
2255 goto out;
2256
2257 /*
2258 * Part (or even all) of the reservation could have been backed
2259 * by pre-allocated pages. Only free surplus pages.
2260 */
2261 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2262
2263 /*
2264 * We want to release as many surplus pages as possible, spread
2265 * evenly across all nodes with memory. Iterate across these nodes
2266 * until we can no longer free unreserved surplus pages. This occurs
2267 * when the nodes with surplus pages have no free pages.
2268 * remove_pool_huge_page() will balance the freed pages across the
2269 * on-line nodes with memory and will handle the hstate accounting.
2270 */
2271 while (nr_pages--) {
2272 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2273 if (!page)
2274 goto out;
2275
2276 list_add(&page->lru, &page_list);
2277 }
2278
2279out:
2280 spin_unlock_irq(&hugetlb_lock);
2281 update_and_free_pages_bulk(h, &page_list);
2282 spin_lock_irq(&hugetlb_lock);
2283}
2284
2285
2286/*
2287 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2288 * are used by the huge page allocation routines to manage reservations.
2289 *
2290 * vma_needs_reservation is called to determine if the huge page at addr
2291 * within the vma has an associated reservation. If a reservation is
2292 * needed, the value 1 is returned. The caller is then responsible for
2293 * managing the global reservation and subpool usage counts. After
2294 * the huge page has been allocated, vma_commit_reservation is called
2295 * to add the page to the reservation map. If the page allocation fails,
2296 * the reservation must be ended instead of committed. vma_end_reservation
2297 * is called in such cases.
2298 *
2299 * In the normal case, vma_commit_reservation returns the same value
2300 * as the preceding vma_needs_reservation call. The only time this
2301 * is not the case is if a reserve map was changed between calls. It
2302 * is the responsibility of the caller to notice the difference and
2303 * take appropriate action.
2304 *
2305 * vma_add_reservation is used in error paths where a reservation must
2306 * be restored when a newly allocated huge page must be freed. It is
2307 * to be called after calling vma_needs_reservation to determine if a
2308 * reservation exists.
2309 *
2310 * vma_del_reservation is used in error paths where an entry in the reserve
2311 * map was created during huge page allocation and must be removed. It is to
2312 * be called after calling vma_needs_reservation to determine if a reservation
2313 * exists.
2314 */
2315enum vma_resv_mode {
2316 VMA_NEEDS_RESV,
2317 VMA_COMMIT_RESV,
2318 VMA_END_RESV,
2319 VMA_ADD_RESV,
2320 VMA_DEL_RESV,
2321};
2322static long __vma_reservation_common(struct hstate *h,
2323 struct vm_area_struct *vma, unsigned long addr,
2324 enum vma_resv_mode mode)
2325{
2326 struct resv_map *resv;
2327 pgoff_t idx;
2328 long ret;
2329 long dummy_out_regions_needed;
2330
2331 resv = vma_resv_map(vma);
2332 if (!resv)
2333 return 1;
2334
2335 idx = vma_hugecache_offset(h, vma, addr);
2336 switch (mode) {
2337 case VMA_NEEDS_RESV:
2338 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2339 /* We assume that vma_reservation_* routines always operate on
2340 * 1 page, and that adding to resv map a 1 page entry can only
2341 * ever require 1 region.
2342 */
2343 VM_BUG_ON(dummy_out_regions_needed != 1);
2344 break;
2345 case VMA_COMMIT_RESV:
2346 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2347 /* region_add calls of range 1 should never fail. */
2348 VM_BUG_ON(ret < 0);
2349 break;
2350 case VMA_END_RESV:
2351 region_abort(resv, idx, idx + 1, 1);
2352 ret = 0;
2353 break;
2354 case VMA_ADD_RESV:
2355 if (vma->vm_flags & VM_MAYSHARE) {
2356 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2357 /* region_add calls of range 1 should never fail. */
2358 VM_BUG_ON(ret < 0);
2359 } else {
2360 region_abort(resv, idx, idx + 1, 1);
2361 ret = region_del(resv, idx, idx + 1);
2362 }
2363 break;
2364 case VMA_DEL_RESV:
2365 if (vma->vm_flags & VM_MAYSHARE) {
2366 region_abort(resv, idx, idx + 1, 1);
2367 ret = region_del(resv, idx, idx + 1);
2368 } else {
2369 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2370 /* region_add calls of range 1 should never fail. */
2371 VM_BUG_ON(ret < 0);
2372 }
2373 break;
2374 default:
2375 BUG();
2376 }
2377
2378 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2379 return ret;
2380 /*
2381 * We know private mapping must have HPAGE_RESV_OWNER set.
2382 *
2383 * In most cases, reserves always exist for private mappings.
2384 * However, a file associated with mapping could have been
2385 * hole punched or truncated after reserves were consumed.
2386 * As subsequent fault on such a range will not use reserves.
2387 * Subtle - The reserve map for private mappings has the
2388 * opposite meaning than that of shared mappings. If NO
2389 * entry is in the reserve map, it means a reservation exists.
2390 * If an entry exists in the reserve map, it means the
2391 * reservation has already been consumed. As a result, the
2392 * return value of this routine is the opposite of the
2393 * value returned from reserve map manipulation routines above.
2394 */
2395 if (ret > 0)
2396 return 0;
2397 if (ret == 0)
2398 return 1;
2399 return ret;
2400}
2401
2402static long vma_needs_reservation(struct hstate *h,
2403 struct vm_area_struct *vma, unsigned long addr)
2404{
2405 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2406}
2407
2408static long vma_commit_reservation(struct hstate *h,
2409 struct vm_area_struct *vma, unsigned long addr)
2410{
2411 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2412}
2413
2414static void vma_end_reservation(struct hstate *h,
2415 struct vm_area_struct *vma, unsigned long addr)
2416{
2417 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2418}
2419
2420static long vma_add_reservation(struct hstate *h,
2421 struct vm_area_struct *vma, unsigned long addr)
2422{
2423 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2424}
2425
2426static long vma_del_reservation(struct hstate *h,
2427 struct vm_area_struct *vma, unsigned long addr)
2428{
2429 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2430}
2431
2432/*
2433 * This routine is called to restore reservation information on error paths.
2434 * It should ONLY be called for pages allocated via alloc_huge_page(), and
2435 * the hugetlb mutex should remain held when calling this routine.
2436 *
2437 * It handles two specific cases:
2438 * 1) A reservation was in place and the page consumed the reservation.
2439 * HPageRestoreReserve is set in the page.
2440 * 2) No reservation was in place for the page, so HPageRestoreReserve is
2441 * not set. However, alloc_huge_page always updates the reserve map.
2442 *
2443 * In case 1, free_huge_page later in the error path will increment the
2444 * global reserve count. But, free_huge_page does not have enough context
2445 * to adjust the reservation map. This case deals primarily with private
2446 * mappings. Adjust the reserve map here to be consistent with global
2447 * reserve count adjustments to be made by free_huge_page. Make sure the
2448 * reserve map indicates there is a reservation present.
2449 *
2450 * In case 2, simply undo reserve map modifications done by alloc_huge_page.
2451 */
2452void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2453 unsigned long address, struct page *page)
2454{
2455 long rc = vma_needs_reservation(h, vma, address);
2456
2457 if (HPageRestoreReserve(page)) {
2458 if (unlikely(rc < 0))
2459 /*
2460 * Rare out of memory condition in reserve map
2461 * manipulation. Clear HPageRestoreReserve so that
2462 * global reserve count will not be incremented
2463 * by free_huge_page. This will make it appear
2464 * as though the reservation for this page was
2465 * consumed. This may prevent the task from
2466 * faulting in the page at a later time. This
2467 * is better than inconsistent global huge page
2468 * accounting of reserve counts.
2469 */
2470 ClearHPageRestoreReserve(page);
2471 else if (rc)
2472 (void)vma_add_reservation(h, vma, address);
2473 else
2474 vma_end_reservation(h, vma, address);
2475 } else {
2476 if (!rc) {
2477 /*
2478 * This indicates there is an entry in the reserve map
2479 * not added by alloc_huge_page. We know it was added
2480 * before the alloc_huge_page call, otherwise
2481 * HPageRestoreReserve would be set on the page.
2482 * Remove the entry so that a subsequent allocation
2483 * does not consume a reservation.
2484 */
2485 rc = vma_del_reservation(h, vma, address);
2486 if (rc < 0)
2487 /*
2488 * VERY rare out of memory condition. Since
2489 * we can not delete the entry, set
2490 * HPageRestoreReserve so that the reserve
2491 * count will be incremented when the page
2492 * is freed. This reserve will be consumed
2493 * on a subsequent allocation.
2494 */
2495 SetHPageRestoreReserve(page);
2496 } else if (rc < 0) {
2497 /*
2498 * Rare out of memory condition from
2499 * vma_needs_reservation call. Memory allocation is
2500 * only attempted if a new entry is needed. Therefore,
2501 * this implies there is not an entry in the
2502 * reserve map.
2503 *
2504 * For shared mappings, no entry in the map indicates
2505 * no reservation. We are done.
2506 */
2507 if (!(vma->vm_flags & VM_MAYSHARE))
2508 /*
2509 * For private mappings, no entry indicates
2510 * a reservation is present. Since we can
2511 * not add an entry, set SetHPageRestoreReserve
2512 * on the page so reserve count will be
2513 * incremented when freed. This reserve will
2514 * be consumed on a subsequent allocation.
2515 */
2516 SetHPageRestoreReserve(page);
2517 } else
2518 /*
2519 * No reservation present, do nothing
2520 */
2521 vma_end_reservation(h, vma, address);
2522 }
2523}
2524
2525/*
2526 * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one
2527 * @h: struct hstate old page belongs to
2528 * @old_page: Old page to dissolve
2529 * @list: List to isolate the page in case we need to
2530 * Returns 0 on success, otherwise negated error.
2531 */
2532static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page,
2533 struct list_head *list)
2534{
2535 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2536 int nid = page_to_nid(old_page);
2537 struct page *new_page;
2538 int ret = 0;
2539
2540 /*
2541 * Before dissolving the page, we need to allocate a new one for the
2542 * pool to remain stable. Here, we allocate the page and 'prep' it
2543 * by doing everything but actually updating counters and adding to
2544 * the pool. This simplifies and let us do most of the processing
2545 * under the lock.
2546 */
2547 new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL);
2548 if (!new_page)
2549 return -ENOMEM;
2550 __prep_new_huge_page(h, new_page);
2551
2552retry:
2553 spin_lock_irq(&hugetlb_lock);
2554 if (!PageHuge(old_page)) {
2555 /*
2556 * Freed from under us. Drop new_page too.
2557 */
2558 goto free_new;
2559 } else if (page_count(old_page)) {
2560 /*
2561 * Someone has grabbed the page, try to isolate it here.
2562 * Fail with -EBUSY if not possible.
2563 */
2564 spin_unlock_irq(&hugetlb_lock);
2565 if (!isolate_huge_page(old_page, list))
2566 ret = -EBUSY;
2567 spin_lock_irq(&hugetlb_lock);
2568 goto free_new;
2569 } else if (!HPageFreed(old_page)) {
2570 /*
2571 * Page's refcount is 0 but it has not been enqueued in the
2572 * freelist yet. Race window is small, so we can succeed here if
2573 * we retry.
2574 */
2575 spin_unlock_irq(&hugetlb_lock);
2576 cond_resched();
2577 goto retry;
2578 } else {
2579 /*
2580 * Ok, old_page is still a genuine free hugepage. Remove it from
2581 * the freelist and decrease the counters. These will be
2582 * incremented again when calling __prep_account_new_huge_page()
2583 * and enqueue_huge_page() for new_page. The counters will remain
2584 * stable since this happens under the lock.
2585 */
2586 remove_hugetlb_page(h, old_page, false);
2587
2588 /*
2589 * Reference count trick is needed because allocator gives us
2590 * referenced page but the pool requires pages with 0 refcount.
2591 */
2592 __prep_account_new_huge_page(h, nid);
2593 page_ref_dec(new_page);
2594 enqueue_huge_page(h, new_page);
2595
2596 /*
2597 * Pages have been replaced, we can safely free the old one.
2598 */
2599 spin_unlock_irq(&hugetlb_lock);
2600 update_and_free_page(h, old_page, false);
2601 }
2602
2603 return ret;
2604
2605free_new:
2606 spin_unlock_irq(&hugetlb_lock);
2607 update_and_free_page(h, new_page, false);
2608
2609 return ret;
2610}
2611
2612int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
2613{
2614 struct hstate *h;
2615 struct page *head;
2616 int ret = -EBUSY;
2617
2618 /*
2619 * The page might have been dissolved from under our feet, so make sure
2620 * to carefully check the state under the lock.
2621 * Return success when racing as if we dissolved the page ourselves.
2622 */
2623 spin_lock_irq(&hugetlb_lock);
2624 if (PageHuge(page)) {
2625 head = compound_head(page);
2626 h = page_hstate(head);
2627 } else {
2628 spin_unlock_irq(&hugetlb_lock);
2629 return 0;
2630 }
2631 spin_unlock_irq(&hugetlb_lock);
2632
2633 /*
2634 * Fence off gigantic pages as there is a cyclic dependency between
2635 * alloc_contig_range and them. Return -ENOMEM as this has the effect
2636 * of bailing out right away without further retrying.
2637 */
2638 if (hstate_is_gigantic(h))
2639 return -ENOMEM;
2640
2641 if (page_count(head) && isolate_huge_page(head, list))
2642 ret = 0;
2643 else if (!page_count(head))
2644 ret = alloc_and_dissolve_huge_page(h, head, list);
2645
2646 return ret;
2647}
2648
2649struct page *alloc_huge_page(struct vm_area_struct *vma,
2650 unsigned long addr, int avoid_reserve)
2651{
2652 struct hugepage_subpool *spool = subpool_vma(vma);
2653 struct hstate *h = hstate_vma(vma);
2654 struct page *page;
2655 long map_chg, map_commit;
2656 long gbl_chg;
2657 int ret, idx;
2658 struct hugetlb_cgroup *h_cg;
2659 bool deferred_reserve;
2660
2661 idx = hstate_index(h);
2662 /*
2663 * Examine the region/reserve map to determine if the process
2664 * has a reservation for the page to be allocated. A return
2665 * code of zero indicates a reservation exists (no change).
2666 */
2667 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2668 if (map_chg < 0)
2669 return ERR_PTR(-ENOMEM);
2670
2671 /*
2672 * Processes that did not create the mapping will have no
2673 * reserves as indicated by the region/reserve map. Check
2674 * that the allocation will not exceed the subpool limit.
2675 * Allocations for MAP_NORESERVE mappings also need to be
2676 * checked against any subpool limit.
2677 */
2678 if (map_chg || avoid_reserve) {
2679 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2680 if (gbl_chg < 0) {
2681 vma_end_reservation(h, vma, addr);
2682 return ERR_PTR(-ENOSPC);
2683 }
2684
2685 /*
2686 * Even though there was no reservation in the region/reserve
2687 * map, there could be reservations associated with the
2688 * subpool that can be used. This would be indicated if the
2689 * return value of hugepage_subpool_get_pages() is zero.
2690 * However, if avoid_reserve is specified we still avoid even
2691 * the subpool reservations.
2692 */
2693 if (avoid_reserve)
2694 gbl_chg = 1;
2695 }
2696
2697 /* If this allocation is not consuming a reservation, charge it now.
2698 */
2699 deferred_reserve = map_chg || avoid_reserve;
2700 if (deferred_reserve) {
2701 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2702 idx, pages_per_huge_page(h), &h_cg);
2703 if (ret)
2704 goto out_subpool_put;
2705 }
2706
2707 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2708 if (ret)
2709 goto out_uncharge_cgroup_reservation;
2710
2711 spin_lock_irq(&hugetlb_lock);
2712 /*
2713 * glb_chg is passed to indicate whether or not a page must be taken
2714 * from the global free pool (global change). gbl_chg == 0 indicates
2715 * a reservation exists for the allocation.
2716 */
2717 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2718 if (!page) {
2719 spin_unlock_irq(&hugetlb_lock);
2720 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2721 if (!page)
2722 goto out_uncharge_cgroup;
2723 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2724 SetHPageRestoreReserve(page);
2725 h->resv_huge_pages--;
2726 }
2727 spin_lock_irq(&hugetlb_lock);
2728 list_add(&page->lru, &h->hugepage_activelist);
2729 /* Fall through */
2730 }
2731 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2732 /* If allocation is not consuming a reservation, also store the
2733 * hugetlb_cgroup pointer on the page.
2734 */
2735 if (deferred_reserve) {
2736 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2737 h_cg, page);
2738 }
2739
2740 spin_unlock_irq(&hugetlb_lock);
2741
2742 hugetlb_set_page_subpool(page, spool);
2743
2744 map_commit = vma_commit_reservation(h, vma, addr);
2745 if (unlikely(map_chg > map_commit)) {
2746 /*
2747 * The page was added to the reservation map between
2748 * vma_needs_reservation and vma_commit_reservation.
2749 * This indicates a race with hugetlb_reserve_pages.
2750 * Adjust for the subpool count incremented above AND
2751 * in hugetlb_reserve_pages for the same page. Also,
2752 * the reservation count added in hugetlb_reserve_pages
2753 * no longer applies.
2754 */
2755 long rsv_adjust;
2756
2757 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2758 hugetlb_acct_memory(h, -rsv_adjust);
2759 if (deferred_reserve)
2760 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2761 pages_per_huge_page(h), page);
2762 }
2763 return page;
2764
2765out_uncharge_cgroup:
2766 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2767out_uncharge_cgroup_reservation:
2768 if (deferred_reserve)
2769 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2770 h_cg);
2771out_subpool_put:
2772 if (map_chg || avoid_reserve)
2773 hugepage_subpool_put_pages(spool, 1);
2774 vma_end_reservation(h, vma, addr);
2775 return ERR_PTR(-ENOSPC);
2776}
2777
2778int alloc_bootmem_huge_page(struct hstate *h)
2779 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2780int __alloc_bootmem_huge_page(struct hstate *h)
2781{
2782 struct huge_bootmem_page *m;
2783 int nr_nodes, node;
2784
2785 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2786 void *addr;
2787
2788 addr = memblock_alloc_try_nid_raw(
2789 huge_page_size(h), huge_page_size(h),
2790 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2791 if (addr) {
2792 /*
2793 * Use the beginning of the huge page to store the
2794 * huge_bootmem_page struct (until gather_bootmem
2795 * puts them into the mem_map).
2796 */
2797 m = addr;
2798 goto found;
2799 }
2800 }
2801 return 0;
2802
2803found:
2804 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2805 /* Put them into a private list first because mem_map is not up yet */
2806 INIT_LIST_HEAD(&m->list);
2807 list_add(&m->list, &huge_boot_pages);
2808 m->hstate = h;
2809 return 1;
2810}
2811
2812/*
2813 * Put bootmem huge pages into the standard lists after mem_map is up.
2814 * Note: This only applies to gigantic (order > MAX_ORDER) pages.
2815 */
2816static void __init gather_bootmem_prealloc(void)
2817{
2818 struct huge_bootmem_page *m;
2819
2820 list_for_each_entry(m, &huge_boot_pages, list) {
2821 struct page *page = virt_to_page(m);
2822 struct hstate *h = m->hstate;
2823
2824 VM_BUG_ON(!hstate_is_gigantic(h));
2825 WARN_ON(page_count(page) != 1);
2826 if (prep_compound_gigantic_page(page, huge_page_order(h))) {
2827 WARN_ON(PageReserved(page));
2828 prep_new_huge_page(h, page, page_to_nid(page));
2829 put_page(page); /* add to the hugepage allocator */
2830 } else {
2831 free_gigantic_page(page, huge_page_order(h));
2832 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
2833 }
2834
2835 /*
2836 * We need to restore the 'stolen' pages to totalram_pages
2837 * in order to fix confusing memory reports from free(1) and
2838 * other side-effects, like CommitLimit going negative.
2839 */
2840 adjust_managed_page_count(page, pages_per_huge_page(h));
2841 cond_resched();
2842 }
2843}
2844
2845static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2846{
2847 unsigned long i;
2848 nodemask_t *node_alloc_noretry;
2849
2850 if (!hstate_is_gigantic(h)) {
2851 /*
2852 * Bit mask controlling how hard we retry per-node allocations.
2853 * Ignore errors as lower level routines can deal with
2854 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2855 * time, we are likely in bigger trouble.
2856 */
2857 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2858 GFP_KERNEL);
2859 } else {
2860 /* allocations done at boot time */
2861 node_alloc_noretry = NULL;
2862 }
2863
2864 /* bit mask controlling how hard we retry per-node allocations */
2865 if (node_alloc_noretry)
2866 nodes_clear(*node_alloc_noretry);
2867
2868 for (i = 0; i < h->max_huge_pages; ++i) {
2869 if (hstate_is_gigantic(h)) {
2870 if (hugetlb_cma_size) {
2871 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2872 goto free;
2873 }
2874 if (!alloc_bootmem_huge_page(h))
2875 break;
2876 } else if (!alloc_pool_huge_page(h,
2877 &node_states[N_MEMORY],
2878 node_alloc_noretry))
2879 break;
2880 cond_resched();
2881 }
2882 if (i < h->max_huge_pages) {
2883 char buf[32];
2884
2885 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2886 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2887 h->max_huge_pages, buf, i);
2888 h->max_huge_pages = i;
2889 }
2890free:
2891 kfree(node_alloc_noretry);
2892}
2893
2894static void __init hugetlb_init_hstates(void)
2895{
2896 struct hstate *h;
2897
2898 for_each_hstate(h) {
2899 if (minimum_order > huge_page_order(h))
2900 minimum_order = huge_page_order(h);
2901
2902 /* oversize hugepages were init'ed in early boot */
2903 if (!hstate_is_gigantic(h))
2904 hugetlb_hstate_alloc_pages(h);
2905 }
2906 VM_BUG_ON(minimum_order == UINT_MAX);
2907}
2908
2909static void __init report_hugepages(void)
2910{
2911 struct hstate *h;
2912
2913 for_each_hstate(h) {
2914 char buf[32];
2915
2916 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2917 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2918 buf, h->free_huge_pages);
2919 }
2920}
2921
2922#ifdef CONFIG_HIGHMEM
2923static void try_to_free_low(struct hstate *h, unsigned long count,
2924 nodemask_t *nodes_allowed)
2925{
2926 int i;
2927 LIST_HEAD(page_list);
2928
2929 lockdep_assert_held(&hugetlb_lock);
2930 if (hstate_is_gigantic(h))
2931 return;
2932
2933 /*
2934 * Collect pages to be freed on a list, and free after dropping lock
2935 */
2936 for_each_node_mask(i, *nodes_allowed) {
2937 struct page *page, *next;
2938 struct list_head *freel = &h->hugepage_freelists[i];
2939 list_for_each_entry_safe(page, next, freel, lru) {
2940 if (count >= h->nr_huge_pages)
2941 goto out;
2942 if (PageHighMem(page))
2943 continue;
2944 remove_hugetlb_page(h, page, false);
2945 list_add(&page->lru, &page_list);
2946 }
2947 }
2948
2949out:
2950 spin_unlock_irq(&hugetlb_lock);
2951 update_and_free_pages_bulk(h, &page_list);
2952 spin_lock_irq(&hugetlb_lock);
2953}
2954#else
2955static inline void try_to_free_low(struct hstate *h, unsigned long count,
2956 nodemask_t *nodes_allowed)
2957{
2958}
2959#endif
2960
2961/*
2962 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2963 * balanced by operating on them in a round-robin fashion.
2964 * Returns 1 if an adjustment was made.
2965 */
2966static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2967 int delta)
2968{
2969 int nr_nodes, node;
2970
2971 lockdep_assert_held(&hugetlb_lock);
2972 VM_BUG_ON(delta != -1 && delta != 1);
2973
2974 if (delta < 0) {
2975 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2976 if (h->surplus_huge_pages_node[node])
2977 goto found;
2978 }
2979 } else {
2980 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2981 if (h->surplus_huge_pages_node[node] <
2982 h->nr_huge_pages_node[node])
2983 goto found;
2984 }
2985 }
2986 return 0;
2987
2988found:
2989 h->surplus_huge_pages += delta;
2990 h->surplus_huge_pages_node[node] += delta;
2991 return 1;
2992}
2993
2994#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2995static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2996 nodemask_t *nodes_allowed)
2997{
2998 unsigned long min_count, ret;
2999 struct page *page;
3000 LIST_HEAD(page_list);
3001 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3002
3003 /*
3004 * Bit mask controlling how hard we retry per-node allocations.
3005 * If we can not allocate the bit mask, do not attempt to allocate
3006 * the requested huge pages.
3007 */
3008 if (node_alloc_noretry)
3009 nodes_clear(*node_alloc_noretry);
3010 else
3011 return -ENOMEM;
3012
3013 /*
3014 * resize_lock mutex prevents concurrent adjustments to number of
3015 * pages in hstate via the proc/sysfs interfaces.
3016 */
3017 mutex_lock(&h->resize_lock);
3018 flush_free_hpage_work(h);
3019 spin_lock_irq(&hugetlb_lock);
3020
3021 /*
3022 * Check for a node specific request.
3023 * Changing node specific huge page count may require a corresponding
3024 * change to the global count. In any case, the passed node mask
3025 * (nodes_allowed) will restrict alloc/free to the specified node.
3026 */
3027 if (nid != NUMA_NO_NODE) {
3028 unsigned long old_count = count;
3029
3030 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3031 /*
3032 * User may have specified a large count value which caused the
3033 * above calculation to overflow. In this case, they wanted
3034 * to allocate as many huge pages as possible. Set count to
3035 * largest possible value to align with their intention.
3036 */
3037 if (count < old_count)
3038 count = ULONG_MAX;
3039 }
3040
3041 /*
3042 * Gigantic pages runtime allocation depend on the capability for large
3043 * page range allocation.
3044 * If the system does not provide this feature, return an error when
3045 * the user tries to allocate gigantic pages but let the user free the
3046 * boottime allocated gigantic pages.
3047 */
3048 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3049 if (count > persistent_huge_pages(h)) {
3050 spin_unlock_irq(&hugetlb_lock);
3051 mutex_unlock(&h->resize_lock);
3052 NODEMASK_FREE(node_alloc_noretry);
3053 return -EINVAL;
3054 }
3055 /* Fall through to decrease pool */
3056 }
3057
3058 /*
3059 * Increase the pool size
3060 * First take pages out of surplus state. Then make up the
3061 * remaining difference by allocating fresh huge pages.
3062 *
3063 * We might race with alloc_surplus_huge_page() here and be unable
3064 * to convert a surplus huge page to a normal huge page. That is
3065 * not critical, though, it just means the overall size of the
3066 * pool might be one hugepage larger than it needs to be, but
3067 * within all the constraints specified by the sysctls.
3068 */
3069 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3070 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3071 break;
3072 }
3073
3074 while (count > persistent_huge_pages(h)) {
3075 /*
3076 * If this allocation races such that we no longer need the
3077 * page, free_huge_page will handle it by freeing the page
3078 * and reducing the surplus.
3079 */
3080 spin_unlock_irq(&hugetlb_lock);
3081
3082 /* yield cpu to avoid soft lockup */
3083 cond_resched();
3084
3085 ret = alloc_pool_huge_page(h, nodes_allowed,
3086 node_alloc_noretry);
3087 spin_lock_irq(&hugetlb_lock);
3088 if (!ret)
3089 goto out;
3090
3091 /* Bail for signals. Probably ctrl-c from user */
3092 if (signal_pending(current))
3093 goto out;
3094 }
3095
3096 /*
3097 * Decrease the pool size
3098 * First return free pages to the buddy allocator (being careful
3099 * to keep enough around to satisfy reservations). Then place
3100 * pages into surplus state as needed so the pool will shrink
3101 * to the desired size as pages become free.
3102 *
3103 * By placing pages into the surplus state independent of the
3104 * overcommit value, we are allowing the surplus pool size to
3105 * exceed overcommit. There are few sane options here. Since
3106 * alloc_surplus_huge_page() is checking the global counter,
3107 * though, we'll note that we're not allowed to exceed surplus
3108 * and won't grow the pool anywhere else. Not until one of the
3109 * sysctls are changed, or the surplus pages go out of use.
3110 */
3111 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3112 min_count = max(count, min_count);
3113 try_to_free_low(h, min_count, nodes_allowed);
3114
3115 /*
3116 * Collect pages to be removed on list without dropping lock
3117 */
3118 while (min_count < persistent_huge_pages(h)) {
3119 page = remove_pool_huge_page(h, nodes_allowed, 0);
3120 if (!page)
3121 break;
3122
3123 list_add(&page->lru, &page_list);
3124 }
3125 /* free the pages after dropping lock */
3126 spin_unlock_irq(&hugetlb_lock);
3127 update_and_free_pages_bulk(h, &page_list);
3128 flush_free_hpage_work(h);
3129 spin_lock_irq(&hugetlb_lock);
3130
3131 while (count < persistent_huge_pages(h)) {
3132 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3133 break;
3134 }
3135out:
3136 h->max_huge_pages = persistent_huge_pages(h);
3137 spin_unlock_irq(&hugetlb_lock);
3138 mutex_unlock(&h->resize_lock);
3139
3140 NODEMASK_FREE(node_alloc_noretry);
3141
3142 return 0;
3143}
3144
3145#define HSTATE_ATTR_RO(_name) \
3146 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3147
3148#define HSTATE_ATTR(_name) \
3149 static struct kobj_attribute _name##_attr = \
3150 __ATTR(_name, 0644, _name##_show, _name##_store)
3151
3152static struct kobject *hugepages_kobj;
3153static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3154
3155static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3156
3157static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3158{
3159 int i;
3160
3161 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3162 if (hstate_kobjs[i] == kobj) {
3163 if (nidp)
3164 *nidp = NUMA_NO_NODE;
3165 return &hstates[i];
3166 }
3167
3168 return kobj_to_node_hstate(kobj, nidp);
3169}
3170
3171static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3172 struct kobj_attribute *attr, char *buf)
3173{
3174 struct hstate *h;
3175 unsigned long nr_huge_pages;
3176 int nid;
3177
3178 h = kobj_to_hstate(kobj, &nid);
3179 if (nid == NUMA_NO_NODE)
3180 nr_huge_pages = h->nr_huge_pages;
3181 else
3182 nr_huge_pages = h->nr_huge_pages_node[nid];
3183
3184 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3185}
3186
3187static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3188 struct hstate *h, int nid,
3189 unsigned long count, size_t len)
3190{
3191 int err;
3192 nodemask_t nodes_allowed, *n_mask;
3193
3194 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3195 return -EINVAL;
3196
3197 if (nid == NUMA_NO_NODE) {
3198 /*
3199 * global hstate attribute
3200 */
3201 if (!(obey_mempolicy &&
3202 init_nodemask_of_mempolicy(&nodes_allowed)))
3203 n_mask = &node_states[N_MEMORY];
3204 else
3205 n_mask = &nodes_allowed;
3206 } else {
3207 /*
3208 * Node specific request. count adjustment happens in
3209 * set_max_huge_pages() after acquiring hugetlb_lock.
3210 */
3211 init_nodemask_of_node(&nodes_allowed, nid);
3212 n_mask = &nodes_allowed;
3213 }
3214
3215 err = set_max_huge_pages(h, count, nid, n_mask);
3216
3217 return err ? err : len;
3218}
3219
3220static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3221 struct kobject *kobj, const char *buf,
3222 size_t len)
3223{
3224 struct hstate *h;
3225 unsigned long count;
3226 int nid;
3227 int err;
3228
3229 err = kstrtoul(buf, 10, &count);
3230 if (err)
3231 return err;
3232
3233 h = kobj_to_hstate(kobj, &nid);
3234 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3235}
3236
3237static ssize_t nr_hugepages_show(struct kobject *kobj,
3238 struct kobj_attribute *attr, char *buf)
3239{
3240 return nr_hugepages_show_common(kobj, attr, buf);
3241}
3242
3243static ssize_t nr_hugepages_store(struct kobject *kobj,
3244 struct kobj_attribute *attr, const char *buf, size_t len)
3245{
3246 return nr_hugepages_store_common(false, kobj, buf, len);
3247}
3248HSTATE_ATTR(nr_hugepages);
3249
3250#ifdef CONFIG_NUMA
3251
3252/*
3253 * hstate attribute for optionally mempolicy-based constraint on persistent
3254 * huge page alloc/free.
3255 */
3256static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3257 struct kobj_attribute *attr,
3258 char *buf)
3259{
3260 return nr_hugepages_show_common(kobj, attr, buf);
3261}
3262
3263static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3264 struct kobj_attribute *attr, const char *buf, size_t len)
3265{
3266 return nr_hugepages_store_common(true, kobj, buf, len);
3267}
3268HSTATE_ATTR(nr_hugepages_mempolicy);
3269#endif
3270
3271
3272static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3273 struct kobj_attribute *attr, char *buf)
3274{
3275 struct hstate *h = kobj_to_hstate(kobj, NULL);
3276 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3277}
3278
3279static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3280 struct kobj_attribute *attr, const char *buf, size_t count)
3281{
3282 int err;
3283 unsigned long input;
3284 struct hstate *h = kobj_to_hstate(kobj, NULL);
3285
3286 if (hstate_is_gigantic(h))
3287 return -EINVAL;
3288
3289 err = kstrtoul(buf, 10, &input);
3290 if (err)
3291 return err;
3292
3293 spin_lock_irq(&hugetlb_lock);
3294 h->nr_overcommit_huge_pages = input;
3295 spin_unlock_irq(&hugetlb_lock);
3296
3297 return count;
3298}
3299HSTATE_ATTR(nr_overcommit_hugepages);
3300
3301static ssize_t free_hugepages_show(struct kobject *kobj,
3302 struct kobj_attribute *attr, char *buf)
3303{
3304 struct hstate *h;
3305 unsigned long free_huge_pages;
3306 int nid;
3307
3308 h = kobj_to_hstate(kobj, &nid);
3309 if (nid == NUMA_NO_NODE)
3310 free_huge_pages = h->free_huge_pages;
3311 else
3312 free_huge_pages = h->free_huge_pages_node[nid];
3313
3314 return sysfs_emit(buf, "%lu\n", free_huge_pages);
3315}
3316HSTATE_ATTR_RO(free_hugepages);
3317
3318static ssize_t resv_hugepages_show(struct kobject *kobj,
3319 struct kobj_attribute *attr, char *buf)
3320{
3321 struct hstate *h = kobj_to_hstate(kobj, NULL);
3322 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3323}
3324HSTATE_ATTR_RO(resv_hugepages);
3325
3326static ssize_t surplus_hugepages_show(struct kobject *kobj,
3327 struct kobj_attribute *attr, char *buf)
3328{
3329 struct hstate *h;
3330 unsigned long surplus_huge_pages;
3331 int nid;
3332
3333 h = kobj_to_hstate(kobj, &nid);
3334 if (nid == NUMA_NO_NODE)
3335 surplus_huge_pages = h->surplus_huge_pages;
3336 else
3337 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3338
3339 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3340}
3341HSTATE_ATTR_RO(surplus_hugepages);
3342
3343static struct attribute *hstate_attrs[] = {
3344 &nr_hugepages_attr.attr,
3345 &nr_overcommit_hugepages_attr.attr,
3346 &free_hugepages_attr.attr,
3347 &resv_hugepages_attr.attr,
3348 &surplus_hugepages_attr.attr,
3349#ifdef CONFIG_NUMA
3350 &nr_hugepages_mempolicy_attr.attr,
3351#endif
3352 NULL,
3353};
3354
3355static const struct attribute_group hstate_attr_group = {
3356 .attrs = hstate_attrs,
3357};
3358
3359static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3360 struct kobject **hstate_kobjs,
3361 const struct attribute_group *hstate_attr_group)
3362{
3363 int retval;
3364 int hi = hstate_index(h);
3365
3366 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3367 if (!hstate_kobjs[hi])
3368 return -ENOMEM;
3369
3370 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3371 if (retval) {
3372 kobject_put(hstate_kobjs[hi]);
3373 hstate_kobjs[hi] = NULL;
3374 }
3375
3376 return retval;
3377}
3378
3379static void __init hugetlb_sysfs_init(void)
3380{
3381 struct hstate *h;
3382 int err;
3383
3384 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3385 if (!hugepages_kobj)
3386 return;
3387
3388 for_each_hstate(h) {
3389 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3390 hstate_kobjs, &hstate_attr_group);
3391 if (err)
3392 pr_err("HugeTLB: Unable to add hstate %s", h->name);
3393 }
3394}
3395
3396#ifdef CONFIG_NUMA
3397
3398/*
3399 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3400 * with node devices in node_devices[] using a parallel array. The array
3401 * index of a node device or _hstate == node id.
3402 * This is here to avoid any static dependency of the node device driver, in
3403 * the base kernel, on the hugetlb module.
3404 */
3405struct node_hstate {
3406 struct kobject *hugepages_kobj;
3407 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3408};
3409static struct node_hstate node_hstates[MAX_NUMNODES];
3410
3411/*
3412 * A subset of global hstate attributes for node devices
3413 */
3414static struct attribute *per_node_hstate_attrs[] = {
3415 &nr_hugepages_attr.attr,
3416 &free_hugepages_attr.attr,
3417 &surplus_hugepages_attr.attr,
3418 NULL,
3419};
3420
3421static const struct attribute_group per_node_hstate_attr_group = {
3422 .attrs = per_node_hstate_attrs,
3423};
3424
3425/*
3426 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3427 * Returns node id via non-NULL nidp.
3428 */
3429static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3430{
3431 int nid;
3432
3433 for (nid = 0; nid < nr_node_ids; nid++) {
3434 struct node_hstate *nhs = &node_hstates[nid];
3435 int i;
3436 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3437 if (nhs->hstate_kobjs[i] == kobj) {
3438 if (nidp)
3439 *nidp = nid;
3440 return &hstates[i];
3441 }
3442 }
3443
3444 BUG();
3445 return NULL;
3446}
3447
3448/*
3449 * Unregister hstate attributes from a single node device.
3450 * No-op if no hstate attributes attached.
3451 */
3452static void hugetlb_unregister_node(struct node *node)
3453{
3454 struct hstate *h;
3455 struct node_hstate *nhs = &node_hstates[node->dev.id];
3456
3457 if (!nhs->hugepages_kobj)
3458 return; /* no hstate attributes */
3459
3460 for_each_hstate(h) {
3461 int idx = hstate_index(h);
3462 if (nhs->hstate_kobjs[idx]) {
3463 kobject_put(nhs->hstate_kobjs[idx]);
3464 nhs->hstate_kobjs[idx] = NULL;
3465 }
3466 }
3467
3468 kobject_put(nhs->hugepages_kobj);
3469 nhs->hugepages_kobj = NULL;
3470}
3471
3472
3473/*
3474 * Register hstate attributes for a single node device.
3475 * No-op if attributes already registered.
3476 */
3477static void hugetlb_register_node(struct node *node)
3478{
3479 struct hstate *h;
3480 struct node_hstate *nhs = &node_hstates[node->dev.id];
3481 int err;
3482
3483 if (nhs->hugepages_kobj)
3484 return; /* already allocated */
3485
3486 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3487 &node->dev.kobj);
3488 if (!nhs->hugepages_kobj)
3489 return;
3490
3491 for_each_hstate(h) {
3492 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3493 nhs->hstate_kobjs,
3494 &per_node_hstate_attr_group);
3495 if (err) {
3496 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3497 h->name, node->dev.id);
3498 hugetlb_unregister_node(node);
3499 break;
3500 }
3501 }
3502}
3503
3504/*
3505 * hugetlb init time: register hstate attributes for all registered node
3506 * devices of nodes that have memory. All on-line nodes should have
3507 * registered their associated device by this time.
3508 */
3509static void __init hugetlb_register_all_nodes(void)
3510{
3511 int nid;
3512
3513 for_each_node_state(nid, N_MEMORY) {
3514 struct node *node = node_devices[nid];
3515 if (node->dev.id == nid)
3516 hugetlb_register_node(node);
3517 }
3518
3519 /*
3520 * Let the node device driver know we're here so it can
3521 * [un]register hstate attributes on node hotplug.
3522 */
3523 register_hugetlbfs_with_node(hugetlb_register_node,
3524 hugetlb_unregister_node);
3525}
3526#else /* !CONFIG_NUMA */
3527
3528static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3529{
3530 BUG();
3531 if (nidp)
3532 *nidp = -1;
3533 return NULL;
3534}
3535
3536static void hugetlb_register_all_nodes(void) { }
3537
3538#endif
3539
3540static int __init hugetlb_init(void)
3541{
3542 int i;
3543
3544 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
3545 __NR_HPAGEFLAGS);
3546
3547 if (!hugepages_supported()) {
3548 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3549 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3550 return 0;
3551 }
3552
3553 /*
3554 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3555 * architectures depend on setup being done here.
3556 */
3557 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3558 if (!parsed_default_hugepagesz) {
3559 /*
3560 * If we did not parse a default huge page size, set
3561 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3562 * number of huge pages for this default size was implicitly
3563 * specified, set that here as well.
3564 * Note that the implicit setting will overwrite an explicit
3565 * setting. A warning will be printed in this case.
3566 */
3567 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3568 if (default_hstate_max_huge_pages) {
3569 if (default_hstate.max_huge_pages) {
3570 char buf[32];
3571
3572 string_get_size(huge_page_size(&default_hstate),
3573 1, STRING_UNITS_2, buf, 32);
3574 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3575 default_hstate.max_huge_pages, buf);
3576 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3577 default_hstate_max_huge_pages);
3578 }
3579 default_hstate.max_huge_pages =
3580 default_hstate_max_huge_pages;
3581 }
3582 }
3583
3584 hugetlb_cma_check();
3585 hugetlb_init_hstates();
3586 gather_bootmem_prealloc();
3587 report_hugepages();
3588
3589 hugetlb_sysfs_init();
3590 hugetlb_register_all_nodes();
3591 hugetlb_cgroup_file_init();
3592
3593#ifdef CONFIG_SMP
3594 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3595#else
3596 num_fault_mutexes = 1;
3597#endif
3598 hugetlb_fault_mutex_table =
3599 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3600 GFP_KERNEL);
3601 BUG_ON(!hugetlb_fault_mutex_table);
3602
3603 for (i = 0; i < num_fault_mutexes; i++)
3604 mutex_init(&hugetlb_fault_mutex_table[i]);
3605 return 0;
3606}
3607subsys_initcall(hugetlb_init);
3608
3609/* Overwritten by architectures with more huge page sizes */
3610bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3611{
3612 return size == HPAGE_SIZE;
3613}
3614
3615void __init hugetlb_add_hstate(unsigned int order)
3616{
3617 struct hstate *h;
3618 unsigned long i;
3619
3620 if (size_to_hstate(PAGE_SIZE << order)) {
3621 return;
3622 }
3623 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3624 BUG_ON(order == 0);
3625 h = &hstates[hugetlb_max_hstate++];
3626 mutex_init(&h->resize_lock);
3627 h->order = order;
3628 h->mask = ~(huge_page_size(h) - 1);
3629 for (i = 0; i < MAX_NUMNODES; ++i)
3630 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3631 INIT_LIST_HEAD(&h->hugepage_activelist);
3632 h->next_nid_to_alloc = first_memory_node;
3633 h->next_nid_to_free = first_memory_node;
3634 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3635 huge_page_size(h)/1024);
3636 hugetlb_vmemmap_init(h);
3637
3638 parsed_hstate = h;
3639}
3640
3641/*
3642 * hugepages command line processing
3643 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3644 * specification. If not, ignore the hugepages value. hugepages can also
3645 * be the first huge page command line option in which case it implicitly
3646 * specifies the number of huge pages for the default size.
3647 */
3648static int __init hugepages_setup(char *s)
3649{
3650 unsigned long *mhp;
3651 static unsigned long *last_mhp;
3652
3653 if (!parsed_valid_hugepagesz) {
3654 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3655 parsed_valid_hugepagesz = true;
3656 return 0;
3657 }
3658
3659 /*
3660 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3661 * yet, so this hugepages= parameter goes to the "default hstate".
3662 * Otherwise, it goes with the previously parsed hugepagesz or
3663 * default_hugepagesz.
3664 */
3665 else if (!hugetlb_max_hstate)
3666 mhp = &default_hstate_max_huge_pages;
3667 else
3668 mhp = &parsed_hstate->max_huge_pages;
3669
3670 if (mhp == last_mhp) {
3671 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3672 return 0;
3673 }
3674
3675 if (sscanf(s, "%lu", mhp) <= 0)
3676 *mhp = 0;
3677
3678 /*
3679 * Global state is always initialized later in hugetlb_init.
3680 * But we need to allocate gigantic hstates here early to still
3681 * use the bootmem allocator.
3682 */
3683 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
3684 hugetlb_hstate_alloc_pages(parsed_hstate);
3685
3686 last_mhp = mhp;
3687
3688 return 1;
3689}
3690__setup("hugepages=", hugepages_setup);
3691
3692/*
3693 * hugepagesz command line processing
3694 * A specific huge page size can only be specified once with hugepagesz.
3695 * hugepagesz is followed by hugepages on the command line. The global
3696 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3697 * hugepagesz argument was valid.
3698 */
3699static int __init hugepagesz_setup(char *s)
3700{
3701 unsigned long size;
3702 struct hstate *h;
3703
3704 parsed_valid_hugepagesz = false;
3705 size = (unsigned long)memparse(s, NULL);
3706
3707 if (!arch_hugetlb_valid_size(size)) {
3708 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3709 return 0;
3710 }
3711
3712 h = size_to_hstate(size);
3713 if (h) {
3714 /*
3715 * hstate for this size already exists. This is normally
3716 * an error, but is allowed if the existing hstate is the
3717 * default hstate. More specifically, it is only allowed if
3718 * the number of huge pages for the default hstate was not
3719 * previously specified.
3720 */
3721 if (!parsed_default_hugepagesz || h != &default_hstate ||
3722 default_hstate.max_huge_pages) {
3723 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3724 return 0;
3725 }
3726
3727 /*
3728 * No need to call hugetlb_add_hstate() as hstate already
3729 * exists. But, do set parsed_hstate so that a following
3730 * hugepages= parameter will be applied to this hstate.
3731 */
3732 parsed_hstate = h;
3733 parsed_valid_hugepagesz = true;
3734 return 1;
3735 }
3736
3737 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3738 parsed_valid_hugepagesz = true;
3739 return 1;
3740}
3741__setup("hugepagesz=", hugepagesz_setup);
3742
3743/*
3744 * default_hugepagesz command line input
3745 * Only one instance of default_hugepagesz allowed on command line.
3746 */
3747static int __init default_hugepagesz_setup(char *s)
3748{
3749 unsigned long size;
3750
3751 parsed_valid_hugepagesz = false;
3752 if (parsed_default_hugepagesz) {
3753 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3754 return 0;
3755 }
3756
3757 size = (unsigned long)memparse(s, NULL);
3758
3759 if (!arch_hugetlb_valid_size(size)) {
3760 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3761 return 0;
3762 }
3763
3764 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3765 parsed_valid_hugepagesz = true;
3766 parsed_default_hugepagesz = true;
3767 default_hstate_idx = hstate_index(size_to_hstate(size));
3768
3769 /*
3770 * The number of default huge pages (for this size) could have been
3771 * specified as the first hugetlb parameter: hugepages=X. If so,
3772 * then default_hstate_max_huge_pages is set. If the default huge
3773 * page size is gigantic (>= MAX_ORDER), then the pages must be
3774 * allocated here from bootmem allocator.
3775 */
3776 if (default_hstate_max_huge_pages) {
3777 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3778 if (hstate_is_gigantic(&default_hstate))
3779 hugetlb_hstate_alloc_pages(&default_hstate);
3780 default_hstate_max_huge_pages = 0;
3781 }
3782
3783 return 1;
3784}
3785__setup("default_hugepagesz=", default_hugepagesz_setup);
3786
3787static unsigned int allowed_mems_nr(struct hstate *h)
3788{
3789 int node;
3790 unsigned int nr = 0;
3791 nodemask_t *mpol_allowed;
3792 unsigned int *array = h->free_huge_pages_node;
3793 gfp_t gfp_mask = htlb_alloc_mask(h);
3794
3795 mpol_allowed = policy_nodemask_current(gfp_mask);
3796
3797 for_each_node_mask(node, cpuset_current_mems_allowed) {
3798 if (!mpol_allowed || node_isset(node, *mpol_allowed))
3799 nr += array[node];
3800 }
3801
3802 return nr;
3803}
3804
3805#ifdef CONFIG_SYSCTL
3806static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3807 void *buffer, size_t *length,
3808 loff_t *ppos, unsigned long *out)
3809{
3810 struct ctl_table dup_table;
3811
3812 /*
3813 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3814 * can duplicate the @table and alter the duplicate of it.
3815 */
3816 dup_table = *table;
3817 dup_table.data = out;
3818
3819 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3820}
3821
3822static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3823 struct ctl_table *table, int write,
3824 void *buffer, size_t *length, loff_t *ppos)
3825{
3826 struct hstate *h = &default_hstate;
3827 unsigned long tmp = h->max_huge_pages;
3828 int ret;
3829
3830 if (!hugepages_supported())
3831 return -EOPNOTSUPP;
3832
3833 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3834 &tmp);
3835 if (ret)
3836 goto out;
3837
3838 if (write)
3839 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3840 NUMA_NO_NODE, tmp, *length);
3841out:
3842 return ret;
3843}
3844
3845int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3846 void *buffer, size_t *length, loff_t *ppos)
3847{
3848
3849 return hugetlb_sysctl_handler_common(false, table, write,
3850 buffer, length, ppos);
3851}
3852
3853#ifdef CONFIG_NUMA
3854int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3855 void *buffer, size_t *length, loff_t *ppos)
3856{
3857 return hugetlb_sysctl_handler_common(true, table, write,
3858 buffer, length, ppos);
3859}
3860#endif /* CONFIG_NUMA */
3861
3862int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3863 void *buffer, size_t *length, loff_t *ppos)
3864{
3865 struct hstate *h = &default_hstate;
3866 unsigned long tmp;
3867 int ret;
3868
3869 if (!hugepages_supported())
3870 return -EOPNOTSUPP;
3871
3872 tmp = h->nr_overcommit_huge_pages;
3873
3874 if (write && hstate_is_gigantic(h))
3875 return -EINVAL;
3876
3877 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3878 &tmp);
3879 if (ret)
3880 goto out;
3881
3882 if (write) {
3883 spin_lock_irq(&hugetlb_lock);
3884 h->nr_overcommit_huge_pages = tmp;
3885 spin_unlock_irq(&hugetlb_lock);
3886 }
3887out:
3888 return ret;
3889}
3890
3891#endif /* CONFIG_SYSCTL */
3892
3893void hugetlb_report_meminfo(struct seq_file *m)
3894{
3895 struct hstate *h;
3896 unsigned long total = 0;
3897
3898 if (!hugepages_supported())
3899 return;
3900
3901 for_each_hstate(h) {
3902 unsigned long count = h->nr_huge_pages;
3903
3904 total += huge_page_size(h) * count;
3905
3906 if (h == &default_hstate)
3907 seq_printf(m,
3908 "HugePages_Total: %5lu\n"
3909 "HugePages_Free: %5lu\n"
3910 "HugePages_Rsvd: %5lu\n"
3911 "HugePages_Surp: %5lu\n"
3912 "Hugepagesize: %8lu kB\n",
3913 count,
3914 h->free_huge_pages,
3915 h->resv_huge_pages,
3916 h->surplus_huge_pages,
3917 huge_page_size(h) / SZ_1K);
3918 }
3919
3920 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
3921}
3922
3923int hugetlb_report_node_meminfo(char *buf, int len, int nid)
3924{
3925 struct hstate *h = &default_hstate;
3926
3927 if (!hugepages_supported())
3928 return 0;
3929
3930 return sysfs_emit_at(buf, len,
3931 "Node %d HugePages_Total: %5u\n"
3932 "Node %d HugePages_Free: %5u\n"
3933 "Node %d HugePages_Surp: %5u\n",
3934 nid, h->nr_huge_pages_node[nid],
3935 nid, h->free_huge_pages_node[nid],
3936 nid, h->surplus_huge_pages_node[nid]);
3937}
3938
3939void hugetlb_show_meminfo(void)
3940{
3941 struct hstate *h;
3942 int nid;
3943
3944 if (!hugepages_supported())
3945 return;
3946
3947 for_each_node_state(nid, N_MEMORY)
3948 for_each_hstate(h)
3949 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3950 nid,
3951 h->nr_huge_pages_node[nid],
3952 h->free_huge_pages_node[nid],
3953 h->surplus_huge_pages_node[nid],
3954 huge_page_size(h) / SZ_1K);
3955}
3956
3957void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3958{
3959 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3960 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3961}
3962
3963/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3964unsigned long hugetlb_total_pages(void)
3965{
3966 struct hstate *h;
3967 unsigned long nr_total_pages = 0;
3968
3969 for_each_hstate(h)
3970 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3971 return nr_total_pages;
3972}
3973
3974static int hugetlb_acct_memory(struct hstate *h, long delta)
3975{
3976 int ret = -ENOMEM;
3977
3978 if (!delta)
3979 return 0;
3980
3981 spin_lock_irq(&hugetlb_lock);
3982 /*
3983 * When cpuset is configured, it breaks the strict hugetlb page
3984 * reservation as the accounting is done on a global variable. Such
3985 * reservation is completely rubbish in the presence of cpuset because
3986 * the reservation is not checked against page availability for the
3987 * current cpuset. Application can still potentially OOM'ed by kernel
3988 * with lack of free htlb page in cpuset that the task is in.
3989 * Attempt to enforce strict accounting with cpuset is almost
3990 * impossible (or too ugly) because cpuset is too fluid that
3991 * task or memory node can be dynamically moved between cpusets.
3992 *
3993 * The change of semantics for shared hugetlb mapping with cpuset is
3994 * undesirable. However, in order to preserve some of the semantics,
3995 * we fall back to check against current free page availability as
3996 * a best attempt and hopefully to minimize the impact of changing
3997 * semantics that cpuset has.
3998 *
3999 * Apart from cpuset, we also have memory policy mechanism that
4000 * also determines from which node the kernel will allocate memory
4001 * in a NUMA system. So similar to cpuset, we also should consider
4002 * the memory policy of the current task. Similar to the description
4003 * above.
4004 */
4005 if (delta > 0) {
4006 if (gather_surplus_pages(h, delta) < 0)
4007 goto out;
4008
4009 if (delta > allowed_mems_nr(h)) {
4010 return_unused_surplus_pages(h, delta);
4011 goto out;
4012 }
4013 }
4014
4015 ret = 0;
4016 if (delta < 0)
4017 return_unused_surplus_pages(h, (unsigned long) -delta);
4018
4019out:
4020 spin_unlock_irq(&hugetlb_lock);
4021 return ret;
4022}
4023
4024static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4025{
4026 struct resv_map *resv = vma_resv_map(vma);
4027
4028 /*
4029 * This new VMA should share its siblings reservation map if present.
4030 * The VMA will only ever have a valid reservation map pointer where
4031 * it is being copied for another still existing VMA. As that VMA
4032 * has a reference to the reservation map it cannot disappear until
4033 * after this open call completes. It is therefore safe to take a
4034 * new reference here without additional locking.
4035 */
4036 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
4037 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
4038 kref_get(&resv->refs);
4039 }
4040}
4041
4042static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4043{
4044 struct hstate *h = hstate_vma(vma);
4045 struct resv_map *resv = vma_resv_map(vma);
4046 struct hugepage_subpool *spool = subpool_vma(vma);
4047 unsigned long reserve, start, end;
4048 long gbl_reserve;
4049
4050 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4051 return;
4052
4053 start = vma_hugecache_offset(h, vma, vma->vm_start);
4054 end = vma_hugecache_offset(h, vma, vma->vm_end);
4055
4056 reserve = (end - start) - region_count(resv, start, end);
4057 hugetlb_cgroup_uncharge_counter(resv, start, end);
4058 if (reserve) {
4059 /*
4060 * Decrement reserve counts. The global reserve count may be
4061 * adjusted if the subpool has a minimum size.
4062 */
4063 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4064 hugetlb_acct_memory(h, -gbl_reserve);
4065 }
4066
4067 kref_put(&resv->refs, resv_map_release);
4068}
4069
4070static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4071{
4072 if (addr & ~(huge_page_mask(hstate_vma(vma))))
4073 return -EINVAL;
4074 return 0;
4075}
4076
4077static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4078{
4079 return huge_page_size(hstate_vma(vma));
4080}
4081
4082/*
4083 * We cannot handle pagefaults against hugetlb pages at all. They cause
4084 * handle_mm_fault() to try to instantiate regular-sized pages in the
4085 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
4086 * this far.
4087 */
4088static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4089{
4090 BUG();
4091 return 0;
4092}
4093
4094/*
4095 * When a new function is introduced to vm_operations_struct and added
4096 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4097 * This is because under System V memory model, mappings created via
4098 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4099 * their original vm_ops are overwritten with shm_vm_ops.
4100 */
4101const struct vm_operations_struct hugetlb_vm_ops = {
4102 .fault = hugetlb_vm_op_fault,
4103 .open = hugetlb_vm_op_open,
4104 .close = hugetlb_vm_op_close,
4105 .may_split = hugetlb_vm_op_split,
4106 .pagesize = hugetlb_vm_op_pagesize,
4107};
4108
4109static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4110 int writable)
4111{
4112 pte_t entry;
4113 unsigned int shift = huge_page_shift(hstate_vma(vma));
4114
4115 if (writable) {
4116 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4117 vma->vm_page_prot)));
4118 } else {
4119 entry = huge_pte_wrprotect(mk_huge_pte(page,
4120 vma->vm_page_prot));
4121 }
4122 entry = pte_mkyoung(entry);
4123 entry = pte_mkhuge(entry);
4124 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4125
4126 return entry;
4127}
4128
4129static void set_huge_ptep_writable(struct vm_area_struct *vma,
4130 unsigned long address, pte_t *ptep)
4131{
4132 pte_t entry;
4133
4134 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4135 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4136 update_mmu_cache(vma, address, ptep);
4137}
4138
4139bool is_hugetlb_entry_migration(pte_t pte)
4140{
4141 swp_entry_t swp;
4142
4143 if (huge_pte_none(pte) || pte_present(pte))
4144 return false;
4145 swp = pte_to_swp_entry(pte);
4146 if (is_migration_entry(swp))
4147 return true;
4148 else
4149 return false;
4150}
4151
4152static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4153{
4154 swp_entry_t swp;
4155
4156 if (huge_pte_none(pte) || pte_present(pte))
4157 return false;
4158 swp = pte_to_swp_entry(pte);
4159 if (is_hwpoison_entry(swp))
4160 return true;
4161 else
4162 return false;
4163}
4164
4165static void
4166hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
4167 struct page *new_page)
4168{
4169 __SetPageUptodate(new_page);
4170 set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
4171 hugepage_add_new_anon_rmap(new_page, vma, addr);
4172 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
4173 ClearHPageRestoreReserve(new_page);
4174 SetHPageMigratable(new_page);
4175}
4176
4177int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
4178 struct vm_area_struct *vma)
4179{
4180 pte_t *src_pte, *dst_pte, entry, dst_entry;
4181 struct page *ptepage;
4182 unsigned long addr;
4183 bool cow = is_cow_mapping(vma->vm_flags);
4184 struct hstate *h = hstate_vma(vma);
4185 unsigned long sz = huge_page_size(h);
4186 unsigned long npages = pages_per_huge_page(h);
4187 struct address_space *mapping = vma->vm_file->f_mapping;
4188 struct mmu_notifier_range range;
4189 int ret = 0;
4190
4191 if (cow) {
4192 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
4193 vma->vm_start,
4194 vma->vm_end);
4195 mmu_notifier_invalidate_range_start(&range);
4196 } else {
4197 /*
4198 * For shared mappings i_mmap_rwsem must be held to call
4199 * huge_pte_alloc, otherwise the returned ptep could go
4200 * away if part of a shared pmd and another thread calls
4201 * huge_pmd_unshare.
4202 */
4203 i_mmap_lock_read(mapping);
4204 }
4205
4206 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
4207 spinlock_t *src_ptl, *dst_ptl;
4208 src_pte = huge_pte_offset(src, addr, sz);
4209 if (!src_pte)
4210 continue;
4211 dst_pte = huge_pte_alloc(dst, vma, addr, sz);
4212 if (!dst_pte) {
4213 ret = -ENOMEM;
4214 break;
4215 }
4216
4217 /*
4218 * If the pagetables are shared don't copy or take references.
4219 * dst_pte == src_pte is the common case of src/dest sharing.
4220 *
4221 * However, src could have 'unshared' and dst shares with
4222 * another vma. If dst_pte !none, this implies sharing.
4223 * Check here before taking page table lock, and once again
4224 * after taking the lock below.
4225 */
4226 dst_entry = huge_ptep_get(dst_pte);
4227 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
4228 continue;
4229
4230 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4231 src_ptl = huge_pte_lockptr(h, src, src_pte);
4232 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4233 entry = huge_ptep_get(src_pte);
4234 dst_entry = huge_ptep_get(dst_pte);
4235again:
4236 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
4237 /*
4238 * Skip if src entry none. Also, skip in the
4239 * unlikely case dst entry !none as this implies
4240 * sharing with another vma.
4241 */
4242 ;
4243 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
4244 is_hugetlb_entry_hwpoisoned(entry))) {
4245 swp_entry_t swp_entry = pte_to_swp_entry(entry);
4246
4247 if (is_writable_migration_entry(swp_entry) && cow) {
4248 /*
4249 * COW mappings require pages in both
4250 * parent and child to be set to read.
4251 */
4252 swp_entry = make_readable_migration_entry(
4253 swp_offset(swp_entry));
4254 entry = swp_entry_to_pte(swp_entry);
4255 set_huge_swap_pte_at(src, addr, src_pte,
4256 entry, sz);
4257 }
4258 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
4259 } else {
4260 entry = huge_ptep_get(src_pte);
4261 ptepage = pte_page(entry);
4262 get_page(ptepage);
4263
4264 /*
4265 * This is a rare case where we see pinned hugetlb
4266 * pages while they're prone to COW. We need to do the
4267 * COW earlier during fork.
4268 *
4269 * When pre-allocating the page or copying data, we
4270 * need to be without the pgtable locks since we could
4271 * sleep during the process.
4272 */
4273 if (unlikely(page_needs_cow_for_dma(vma, ptepage))) {
4274 pte_t src_pte_old = entry;
4275 struct page *new;
4276
4277 spin_unlock(src_ptl);
4278 spin_unlock(dst_ptl);
4279 /* Do not use reserve as it's private owned */
4280 new = alloc_huge_page(vma, addr, 1);
4281 if (IS_ERR(new)) {
4282 put_page(ptepage);
4283 ret = PTR_ERR(new);
4284 break;
4285 }
4286 copy_user_huge_page(new, ptepage, addr, vma,
4287 npages);
4288 put_page(ptepage);
4289
4290 /* Install the new huge page if src pte stable */
4291 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4292 src_ptl = huge_pte_lockptr(h, src, src_pte);
4293 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4294 entry = huge_ptep_get(src_pte);
4295 if (!pte_same(src_pte_old, entry)) {
4296 restore_reserve_on_error(h, vma, addr,
4297 new);
4298 put_page(new);
4299 /* dst_entry won't change as in child */
4300 goto again;
4301 }
4302 hugetlb_install_page(vma, dst_pte, addr, new);
4303 spin_unlock(src_ptl);
4304 spin_unlock(dst_ptl);
4305 continue;
4306 }
4307
4308 if (cow) {
4309 /*
4310 * No need to notify as we are downgrading page
4311 * table protection not changing it to point
4312 * to a new page.
4313 *
4314 * See Documentation/vm/mmu_notifier.rst
4315 */
4316 huge_ptep_set_wrprotect(src, addr, src_pte);
4317 entry = huge_pte_wrprotect(entry);
4318 }
4319
4320 page_dup_rmap(ptepage, true);
4321 set_huge_pte_at(dst, addr, dst_pte, entry);
4322 hugetlb_count_add(npages, dst);
4323 }
4324 spin_unlock(src_ptl);
4325 spin_unlock(dst_ptl);
4326 }
4327
4328 if (cow)
4329 mmu_notifier_invalidate_range_end(&range);
4330 else
4331 i_mmap_unlock_read(mapping);
4332
4333 return ret;
4334}
4335
4336void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
4337 unsigned long start, unsigned long end,
4338 struct page *ref_page)
4339{
4340 struct mm_struct *mm = vma->vm_mm;
4341 unsigned long address;
4342 pte_t *ptep;
4343 pte_t pte;
4344 spinlock_t *ptl;
4345 struct page *page;
4346 struct hstate *h = hstate_vma(vma);
4347 unsigned long sz = huge_page_size(h);
4348 struct mmu_notifier_range range;
4349
4350 WARN_ON(!is_vm_hugetlb_page(vma));
4351 BUG_ON(start & ~huge_page_mask(h));
4352 BUG_ON(end & ~huge_page_mask(h));
4353
4354 /*
4355 * This is a hugetlb vma, all the pte entries should point
4356 * to huge page.
4357 */
4358 tlb_change_page_size(tlb, sz);
4359 tlb_start_vma(tlb, vma);
4360
4361 /*
4362 * If sharing possible, alert mmu notifiers of worst case.
4363 */
4364 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
4365 end);
4366 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4367 mmu_notifier_invalidate_range_start(&range);
4368 address = start;
4369 for (; address < end; address += sz) {
4370 ptep = huge_pte_offset(mm, address, sz);
4371 if (!ptep)
4372 continue;
4373
4374 ptl = huge_pte_lock(h, mm, ptep);
4375 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
4376 spin_unlock(ptl);
4377 /*
4378 * We just unmapped a page of PMDs by clearing a PUD.
4379 * The caller's TLB flush range should cover this area.
4380 */
4381 continue;
4382 }
4383
4384 pte = huge_ptep_get(ptep);
4385 if (huge_pte_none(pte)) {
4386 spin_unlock(ptl);
4387 continue;
4388 }
4389
4390 /*
4391 * Migrating hugepage or HWPoisoned hugepage is already
4392 * unmapped and its refcount is dropped, so just clear pte here.
4393 */
4394 if (unlikely(!pte_present(pte))) {
4395 huge_pte_clear(mm, address, ptep, sz);
4396 spin_unlock(ptl);
4397 continue;
4398 }
4399
4400 page = pte_page(pte);
4401 /*
4402 * If a reference page is supplied, it is because a specific
4403 * page is being unmapped, not a range. Ensure the page we
4404 * are about to unmap is the actual page of interest.
4405 */
4406 if (ref_page) {
4407 if (page != ref_page) {
4408 spin_unlock(ptl);
4409 continue;
4410 }
4411 /*
4412 * Mark the VMA as having unmapped its page so that
4413 * future faults in this VMA will fail rather than
4414 * looking like data was lost
4415 */
4416 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
4417 }
4418
4419 pte = huge_ptep_get_and_clear(mm, address, ptep);
4420 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
4421 if (huge_pte_dirty(pte))
4422 set_page_dirty(page);
4423
4424 hugetlb_count_sub(pages_per_huge_page(h), mm);
4425 page_remove_rmap(page, true);
4426
4427 spin_unlock(ptl);
4428 tlb_remove_page_size(tlb, page, huge_page_size(h));
4429 /*
4430 * Bail out after unmapping reference page if supplied
4431 */
4432 if (ref_page)
4433 break;
4434 }
4435 mmu_notifier_invalidate_range_end(&range);
4436 tlb_end_vma(tlb, vma);
4437}
4438
4439void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4440 struct vm_area_struct *vma, unsigned long start,
4441 unsigned long end, struct page *ref_page)
4442{
4443 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4444
4445 /*
4446 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4447 * test will fail on a vma being torn down, and not grab a page table
4448 * on its way out. We're lucky that the flag has such an appropriate
4449 * name, and can in fact be safely cleared here. We could clear it
4450 * before the __unmap_hugepage_range above, but all that's necessary
4451 * is to clear it before releasing the i_mmap_rwsem. This works
4452 * because in the context this is called, the VMA is about to be
4453 * destroyed and the i_mmap_rwsem is held.
4454 */
4455 vma->vm_flags &= ~VM_MAYSHARE;
4456}
4457
4458void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
4459 unsigned long end, struct page *ref_page)
4460{
4461 struct mmu_gather tlb;
4462
4463 tlb_gather_mmu(&tlb, vma->vm_mm);
4464 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
4465 tlb_finish_mmu(&tlb);
4466}
4467
4468/*
4469 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4470 * mapping it owns the reserve page for. The intention is to unmap the page
4471 * from other VMAs and let the children be SIGKILLed if they are faulting the
4472 * same region.
4473 */
4474static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4475 struct page *page, unsigned long address)
4476{
4477 struct hstate *h = hstate_vma(vma);
4478 struct vm_area_struct *iter_vma;
4479 struct address_space *mapping;
4480 pgoff_t pgoff;
4481
4482 /*
4483 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4484 * from page cache lookup which is in HPAGE_SIZE units.
4485 */
4486 address = address & huge_page_mask(h);
4487 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4488 vma->vm_pgoff;
4489 mapping = vma->vm_file->f_mapping;
4490
4491 /*
4492 * Take the mapping lock for the duration of the table walk. As
4493 * this mapping should be shared between all the VMAs,
4494 * __unmap_hugepage_range() is called as the lock is already held
4495 */
4496 i_mmap_lock_write(mapping);
4497 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
4498 /* Do not unmap the current VMA */
4499 if (iter_vma == vma)
4500 continue;
4501
4502 /*
4503 * Shared VMAs have their own reserves and do not affect
4504 * MAP_PRIVATE accounting but it is possible that a shared
4505 * VMA is using the same page so check and skip such VMAs.
4506 */
4507 if (iter_vma->vm_flags & VM_MAYSHARE)
4508 continue;
4509
4510 /*
4511 * Unmap the page from other VMAs without their own reserves.
4512 * They get marked to be SIGKILLed if they fault in these
4513 * areas. This is because a future no-page fault on this VMA
4514 * could insert a zeroed page instead of the data existing
4515 * from the time of fork. This would look like data corruption
4516 */
4517 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
4518 unmap_hugepage_range(iter_vma, address,
4519 address + huge_page_size(h), page);
4520 }
4521 i_mmap_unlock_write(mapping);
4522}
4523
4524/*
4525 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4526 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4527 * cannot race with other handlers or page migration.
4528 * Keep the pte_same checks anyway to make transition from the mutex easier.
4529 */
4530static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
4531 unsigned long address, pte_t *ptep,
4532 struct page *pagecache_page, spinlock_t *ptl)
4533{
4534 pte_t pte;
4535 struct hstate *h = hstate_vma(vma);
4536 struct page *old_page, *new_page;
4537 int outside_reserve = 0;
4538 vm_fault_t ret = 0;
4539 unsigned long haddr = address & huge_page_mask(h);
4540 struct mmu_notifier_range range;
4541
4542 pte = huge_ptep_get(ptep);
4543 old_page = pte_page(pte);
4544
4545retry_avoidcopy:
4546 /* If no-one else is actually using this page, avoid the copy
4547 * and just make the page writable */
4548 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
4549 page_move_anon_rmap(old_page, vma);
4550 set_huge_ptep_writable(vma, haddr, ptep);
4551 return 0;
4552 }
4553
4554 /*
4555 * If the process that created a MAP_PRIVATE mapping is about to
4556 * perform a COW due to a shared page count, attempt to satisfy
4557 * the allocation without using the existing reserves. The pagecache
4558 * page is used to determine if the reserve at this address was
4559 * consumed or not. If reserves were used, a partial faulted mapping
4560 * at the time of fork() could consume its reserves on COW instead
4561 * of the full address range.
4562 */
4563 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4564 old_page != pagecache_page)
4565 outside_reserve = 1;
4566
4567 get_page(old_page);
4568
4569 /*
4570 * Drop page table lock as buddy allocator may be called. It will
4571 * be acquired again before returning to the caller, as expected.
4572 */
4573 spin_unlock(ptl);
4574 new_page = alloc_huge_page(vma, haddr, outside_reserve);
4575
4576 if (IS_ERR(new_page)) {
4577 /*
4578 * If a process owning a MAP_PRIVATE mapping fails to COW,
4579 * it is due to references held by a child and an insufficient
4580 * huge page pool. To guarantee the original mappers
4581 * reliability, unmap the page from child processes. The child
4582 * may get SIGKILLed if it later faults.
4583 */
4584 if (outside_reserve) {
4585 struct address_space *mapping = vma->vm_file->f_mapping;
4586 pgoff_t idx;
4587 u32 hash;
4588
4589 put_page(old_page);
4590 BUG_ON(huge_pte_none(pte));
4591 /*
4592 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4593 * unmapping. unmapping needs to hold i_mmap_rwsem
4594 * in write mode. Dropping i_mmap_rwsem in read mode
4595 * here is OK as COW mappings do not interact with
4596 * PMD sharing.
4597 *
4598 * Reacquire both after unmap operation.
4599 */
4600 idx = vma_hugecache_offset(h, vma, haddr);
4601 hash = hugetlb_fault_mutex_hash(mapping, idx);
4602 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4603 i_mmap_unlock_read(mapping);
4604
4605 unmap_ref_private(mm, vma, old_page, haddr);
4606
4607 i_mmap_lock_read(mapping);
4608 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4609 spin_lock(ptl);
4610 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4611 if (likely(ptep &&
4612 pte_same(huge_ptep_get(ptep), pte)))
4613 goto retry_avoidcopy;
4614 /*
4615 * race occurs while re-acquiring page table
4616 * lock, and our job is done.
4617 */
4618 return 0;
4619 }
4620
4621 ret = vmf_error(PTR_ERR(new_page));
4622 goto out_release_old;
4623 }
4624
4625 /*
4626 * When the original hugepage is shared one, it does not have
4627 * anon_vma prepared.
4628 */
4629 if (unlikely(anon_vma_prepare(vma))) {
4630 ret = VM_FAULT_OOM;
4631 goto out_release_all;
4632 }
4633
4634 copy_user_huge_page(new_page, old_page, address, vma,
4635 pages_per_huge_page(h));
4636 __SetPageUptodate(new_page);
4637
4638 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4639 haddr + huge_page_size(h));
4640 mmu_notifier_invalidate_range_start(&range);
4641
4642 /*
4643 * Retake the page table lock to check for racing updates
4644 * before the page tables are altered
4645 */
4646 spin_lock(ptl);
4647 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4648 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4649 ClearHPageRestoreReserve(new_page);
4650
4651 /* Break COW */
4652 huge_ptep_clear_flush(vma, haddr, ptep);
4653 mmu_notifier_invalidate_range(mm, range.start, range.end);
4654 set_huge_pte_at(mm, haddr, ptep,
4655 make_huge_pte(vma, new_page, 1));
4656 page_remove_rmap(old_page, true);
4657 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4658 SetHPageMigratable(new_page);
4659 /* Make the old page be freed below */
4660 new_page = old_page;
4661 }
4662 spin_unlock(ptl);
4663 mmu_notifier_invalidate_range_end(&range);
4664out_release_all:
4665 /* No restore in case of successful pagetable update (Break COW) */
4666 if (new_page != old_page)
4667 restore_reserve_on_error(h, vma, haddr, new_page);
4668 put_page(new_page);
4669out_release_old:
4670 put_page(old_page);
4671
4672 spin_lock(ptl); /* Caller expects lock to be held */
4673 return ret;
4674}
4675
4676/* Return the pagecache page at a given address within a VMA */
4677static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4678 struct vm_area_struct *vma, unsigned long address)
4679{
4680 struct address_space *mapping;
4681 pgoff_t idx;
4682
4683 mapping = vma->vm_file->f_mapping;
4684 idx = vma_hugecache_offset(h, vma, address);
4685
4686 return find_lock_page(mapping, idx);
4687}
4688
4689/*
4690 * Return whether there is a pagecache page to back given address within VMA.
4691 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4692 */
4693static bool hugetlbfs_pagecache_present(struct hstate *h,
4694 struct vm_area_struct *vma, unsigned long address)
4695{
4696 struct address_space *mapping;
4697 pgoff_t idx;
4698 struct page *page;
4699
4700 mapping = vma->vm_file->f_mapping;
4701 idx = vma_hugecache_offset(h, vma, address);
4702
4703 page = find_get_page(mapping, idx);
4704 if (page)
4705 put_page(page);
4706 return page != NULL;
4707}
4708
4709int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4710 pgoff_t idx)
4711{
4712 struct inode *inode = mapping->host;
4713 struct hstate *h = hstate_inode(inode);
4714 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4715
4716 if (err)
4717 return err;
4718 ClearHPageRestoreReserve(page);
4719
4720 /*
4721 * set page dirty so that it will not be removed from cache/file
4722 * by non-hugetlbfs specific code paths.
4723 */
4724 set_page_dirty(page);
4725
4726 spin_lock(&inode->i_lock);
4727 inode->i_blocks += blocks_per_huge_page(h);
4728 spin_unlock(&inode->i_lock);
4729 return 0;
4730}
4731
4732static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
4733 struct address_space *mapping,
4734 pgoff_t idx,
4735 unsigned int flags,
4736 unsigned long haddr,
4737 unsigned long reason)
4738{
4739 vm_fault_t ret;
4740 u32 hash;
4741 struct vm_fault vmf = {
4742 .vma = vma,
4743 .address = haddr,
4744 .flags = flags,
4745
4746 /*
4747 * Hard to debug if it ends up being
4748 * used by a callee that assumes
4749 * something about the other
4750 * uninitialized fields... same as in
4751 * memory.c
4752 */
4753 };
4754
4755 /*
4756 * hugetlb_fault_mutex and i_mmap_rwsem must be
4757 * dropped before handling userfault. Reacquire
4758 * after handling fault to make calling code simpler.
4759 */
4760 hash = hugetlb_fault_mutex_hash(mapping, idx);
4761 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4762 i_mmap_unlock_read(mapping);
4763 ret = handle_userfault(&vmf, reason);
4764 i_mmap_lock_read(mapping);
4765 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4766
4767 return ret;
4768}
4769
4770static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4771 struct vm_area_struct *vma,
4772 struct address_space *mapping, pgoff_t idx,
4773 unsigned long address, pte_t *ptep, unsigned int flags)
4774{
4775 struct hstate *h = hstate_vma(vma);
4776 vm_fault_t ret = VM_FAULT_SIGBUS;
4777 int anon_rmap = 0;
4778 unsigned long size;
4779 struct page *page;
4780 pte_t new_pte;
4781 spinlock_t *ptl;
4782 unsigned long haddr = address & huge_page_mask(h);
4783 bool new_page, new_pagecache_page = false;
4784
4785 /*
4786 * Currently, we are forced to kill the process in the event the
4787 * original mapper has unmapped pages from the child due to a failed
4788 * COW. Warn that such a situation has occurred as it may not be obvious
4789 */
4790 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4791 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4792 current->pid);
4793 return ret;
4794 }
4795
4796 /*
4797 * We can not race with truncation due to holding i_mmap_rwsem.
4798 * i_size is modified when holding i_mmap_rwsem, so check here
4799 * once for faults beyond end of file.
4800 */
4801 size = i_size_read(mapping->host) >> huge_page_shift(h);
4802 if (idx >= size)
4803 goto out;
4804
4805retry:
4806 new_page = false;
4807 page = find_lock_page(mapping, idx);
4808 if (!page) {
4809 /* Check for page in userfault range */
4810 if (userfaultfd_missing(vma)) {
4811 ret = hugetlb_handle_userfault(vma, mapping, idx,
4812 flags, haddr,
4813 VM_UFFD_MISSING);
4814 goto out;
4815 }
4816
4817 page = alloc_huge_page(vma, haddr, 0);
4818 if (IS_ERR(page)) {
4819 /*
4820 * Returning error will result in faulting task being
4821 * sent SIGBUS. The hugetlb fault mutex prevents two
4822 * tasks from racing to fault in the same page which
4823 * could result in false unable to allocate errors.
4824 * Page migration does not take the fault mutex, but
4825 * does a clear then write of pte's under page table
4826 * lock. Page fault code could race with migration,
4827 * notice the clear pte and try to allocate a page
4828 * here. Before returning error, get ptl and make
4829 * sure there really is no pte entry.
4830 */
4831 ptl = huge_pte_lock(h, mm, ptep);
4832 ret = 0;
4833 if (huge_pte_none(huge_ptep_get(ptep)))
4834 ret = vmf_error(PTR_ERR(page));
4835 spin_unlock(ptl);
4836 goto out;
4837 }
4838 clear_huge_page(page, address, pages_per_huge_page(h));
4839 __SetPageUptodate(page);
4840 new_page = true;
4841
4842 if (vma->vm_flags & VM_MAYSHARE) {
4843 int err = huge_add_to_page_cache(page, mapping, idx);
4844 if (err) {
4845 put_page(page);
4846 if (err == -EEXIST)
4847 goto retry;
4848 goto out;
4849 }
4850 new_pagecache_page = true;
4851 } else {
4852 lock_page(page);
4853 if (unlikely(anon_vma_prepare(vma))) {
4854 ret = VM_FAULT_OOM;
4855 goto backout_unlocked;
4856 }
4857 anon_rmap = 1;
4858 }
4859 } else {
4860 /*
4861 * If memory error occurs between mmap() and fault, some process
4862 * don't have hwpoisoned swap entry for errored virtual address.
4863 * So we need to block hugepage fault by PG_hwpoison bit check.
4864 */
4865 if (unlikely(PageHWPoison(page))) {
4866 ret = VM_FAULT_HWPOISON_LARGE |
4867 VM_FAULT_SET_HINDEX(hstate_index(h));
4868 goto backout_unlocked;
4869 }
4870
4871 /* Check for page in userfault range. */
4872 if (userfaultfd_minor(vma)) {
4873 unlock_page(page);
4874 put_page(page);
4875 ret = hugetlb_handle_userfault(vma, mapping, idx,
4876 flags, haddr,
4877 VM_UFFD_MINOR);
4878 goto out;
4879 }
4880 }
4881
4882 /*
4883 * If we are going to COW a private mapping later, we examine the
4884 * pending reservations for this page now. This will ensure that
4885 * any allocations necessary to record that reservation occur outside
4886 * the spinlock.
4887 */
4888 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4889 if (vma_needs_reservation(h, vma, haddr) < 0) {
4890 ret = VM_FAULT_OOM;
4891 goto backout_unlocked;
4892 }
4893 /* Just decrements count, does not deallocate */
4894 vma_end_reservation(h, vma, haddr);
4895 }
4896
4897 ptl = huge_pte_lock(h, mm, ptep);
4898 ret = 0;
4899 if (!huge_pte_none(huge_ptep_get(ptep)))
4900 goto backout;
4901
4902 if (anon_rmap) {
4903 ClearHPageRestoreReserve(page);
4904 hugepage_add_new_anon_rmap(page, vma, haddr);
4905 } else
4906 page_dup_rmap(page, true);
4907 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4908 && (vma->vm_flags & VM_SHARED)));
4909 set_huge_pte_at(mm, haddr, ptep, new_pte);
4910
4911 hugetlb_count_add(pages_per_huge_page(h), mm);
4912 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4913 /* Optimization, do the COW without a second fault */
4914 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4915 }
4916
4917 spin_unlock(ptl);
4918
4919 /*
4920 * Only set HPageMigratable in newly allocated pages. Existing pages
4921 * found in the pagecache may not have HPageMigratableset if they have
4922 * been isolated for migration.
4923 */
4924 if (new_page)
4925 SetHPageMigratable(page);
4926
4927 unlock_page(page);
4928out:
4929 return ret;
4930
4931backout:
4932 spin_unlock(ptl);
4933backout_unlocked:
4934 unlock_page(page);
4935 /* restore reserve for newly allocated pages not in page cache */
4936 if (new_page && !new_pagecache_page)
4937 restore_reserve_on_error(h, vma, haddr, page);
4938 put_page(page);
4939 goto out;
4940}
4941
4942#ifdef CONFIG_SMP
4943u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4944{
4945 unsigned long key[2];
4946 u32 hash;
4947
4948 key[0] = (unsigned long) mapping;
4949 key[1] = idx;
4950
4951 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4952
4953 return hash & (num_fault_mutexes - 1);
4954}
4955#else
4956/*
4957 * For uniprocessor systems we always use a single mutex, so just
4958 * return 0 and avoid the hashing overhead.
4959 */
4960u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4961{
4962 return 0;
4963}
4964#endif
4965
4966vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4967 unsigned long address, unsigned int flags)
4968{
4969 pte_t *ptep, entry;
4970 spinlock_t *ptl;
4971 vm_fault_t ret;
4972 u32 hash;
4973 pgoff_t idx;
4974 struct page *page = NULL;
4975 struct page *pagecache_page = NULL;
4976 struct hstate *h = hstate_vma(vma);
4977 struct address_space *mapping;
4978 int need_wait_lock = 0;
4979 unsigned long haddr = address & huge_page_mask(h);
4980
4981 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4982 if (ptep) {
4983 /*
4984 * Since we hold no locks, ptep could be stale. That is
4985 * OK as we are only making decisions based on content and
4986 * not actually modifying content here.
4987 */
4988 entry = huge_ptep_get(ptep);
4989 if (unlikely(is_hugetlb_entry_migration(entry))) {
4990 migration_entry_wait_huge(vma, mm, ptep);
4991 return 0;
4992 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4993 return VM_FAULT_HWPOISON_LARGE |
4994 VM_FAULT_SET_HINDEX(hstate_index(h));
4995 }
4996
4997 /*
4998 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4999 * until finished with ptep. This serves two purposes:
5000 * 1) It prevents huge_pmd_unshare from being called elsewhere
5001 * and making the ptep no longer valid.
5002 * 2) It synchronizes us with i_size modifications during truncation.
5003 *
5004 * ptep could have already be assigned via huge_pte_offset. That
5005 * is OK, as huge_pte_alloc will return the same value unless
5006 * something has changed.
5007 */
5008 mapping = vma->vm_file->f_mapping;
5009 i_mmap_lock_read(mapping);
5010 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
5011 if (!ptep) {
5012 i_mmap_unlock_read(mapping);
5013 return VM_FAULT_OOM;
5014 }
5015
5016 /*
5017 * Serialize hugepage allocation and instantiation, so that we don't
5018 * get spurious allocation failures if two CPUs race to instantiate
5019 * the same page in the page cache.
5020 */
5021 idx = vma_hugecache_offset(h, vma, haddr);
5022 hash = hugetlb_fault_mutex_hash(mapping, idx);
5023 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5024
5025 entry = huge_ptep_get(ptep);
5026 if (huge_pte_none(entry)) {
5027 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
5028 goto out_mutex;
5029 }
5030
5031 ret = 0;
5032
5033 /*
5034 * entry could be a migration/hwpoison entry at this point, so this
5035 * check prevents the kernel from going below assuming that we have
5036 * an active hugepage in pagecache. This goto expects the 2nd page
5037 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
5038 * properly handle it.
5039 */
5040 if (!pte_present(entry))
5041 goto out_mutex;
5042
5043 /*
5044 * If we are going to COW the mapping later, we examine the pending
5045 * reservations for this page now. This will ensure that any
5046 * allocations necessary to record that reservation occur outside the
5047 * spinlock. For private mappings, we also lookup the pagecache
5048 * page now as it is used to determine if a reservation has been
5049 * consumed.
5050 */
5051 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
5052 if (vma_needs_reservation(h, vma, haddr) < 0) {
5053 ret = VM_FAULT_OOM;
5054 goto out_mutex;
5055 }
5056 /* Just decrements count, does not deallocate */
5057 vma_end_reservation(h, vma, haddr);
5058
5059 if (!(vma->vm_flags & VM_MAYSHARE))
5060 pagecache_page = hugetlbfs_pagecache_page(h,
5061 vma, haddr);
5062 }
5063
5064 ptl = huge_pte_lock(h, mm, ptep);
5065
5066 /* Check for a racing update before calling hugetlb_cow */
5067 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
5068 goto out_ptl;
5069
5070 /*
5071 * hugetlb_cow() requires page locks of pte_page(entry) and
5072 * pagecache_page, so here we need take the former one
5073 * when page != pagecache_page or !pagecache_page.
5074 */
5075 page = pte_page(entry);
5076 if (page != pagecache_page)
5077 if (!trylock_page(page)) {
5078 need_wait_lock = 1;
5079 goto out_ptl;
5080 }
5081
5082 get_page(page);
5083
5084 if (flags & FAULT_FLAG_WRITE) {
5085 if (!huge_pte_write(entry)) {
5086 ret = hugetlb_cow(mm, vma, address, ptep,
5087 pagecache_page, ptl);
5088 goto out_put_page;
5089 }
5090 entry = huge_pte_mkdirty(entry);
5091 }
5092 entry = pte_mkyoung(entry);
5093 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
5094 flags & FAULT_FLAG_WRITE))
5095 update_mmu_cache(vma, haddr, ptep);
5096out_put_page:
5097 if (page != pagecache_page)
5098 unlock_page(page);
5099 put_page(page);
5100out_ptl:
5101 spin_unlock(ptl);
5102
5103 if (pagecache_page) {
5104 unlock_page(pagecache_page);
5105 put_page(pagecache_page);
5106 }
5107out_mutex:
5108 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5109 i_mmap_unlock_read(mapping);
5110 /*
5111 * Generally it's safe to hold refcount during waiting page lock. But
5112 * here we just wait to defer the next page fault to avoid busy loop and
5113 * the page is not used after unlocked before returning from the current
5114 * page fault. So we are safe from accessing freed page, even if we wait
5115 * here without taking refcount.
5116 */
5117 if (need_wait_lock)
5118 wait_on_page_locked(page);
5119 return ret;
5120}
5121
5122#ifdef CONFIG_USERFAULTFD
5123/*
5124 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
5125 * modifications for huge pages.
5126 */
5127int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
5128 pte_t *dst_pte,
5129 struct vm_area_struct *dst_vma,
5130 unsigned long dst_addr,
5131 unsigned long src_addr,
5132 enum mcopy_atomic_mode mode,
5133 struct page **pagep)
5134{
5135 bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE);
5136 struct hstate *h = hstate_vma(dst_vma);
5137 struct address_space *mapping = dst_vma->vm_file->f_mapping;
5138 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
5139 unsigned long size;
5140 int vm_shared = dst_vma->vm_flags & VM_SHARED;
5141 pte_t _dst_pte;
5142 spinlock_t *ptl;
5143 int ret = -ENOMEM;
5144 struct page *page;
5145 int writable;
5146 bool new_pagecache_page = false;
5147
5148 if (is_continue) {
5149 ret = -EFAULT;
5150 page = find_lock_page(mapping, idx);
5151 if (!page)
5152 goto out;
5153 } else if (!*pagep) {
5154 /* If a page already exists, then it's UFFDIO_COPY for
5155 * a non-missing case. Return -EEXIST.
5156 */
5157 if (vm_shared &&
5158 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5159 ret = -EEXIST;
5160 goto out;
5161 }
5162
5163 page = alloc_huge_page(dst_vma, dst_addr, 0);
5164 if (IS_ERR(page)) {
5165 ret = -ENOMEM;
5166 goto out;
5167 }
5168
5169 ret = copy_huge_page_from_user(page,
5170 (const void __user *) src_addr,
5171 pages_per_huge_page(h), false);
5172
5173 /* fallback to copy_from_user outside mmap_lock */
5174 if (unlikely(ret)) {
5175 ret = -ENOENT;
5176 /* Free the allocated page which may have
5177 * consumed a reservation.
5178 */
5179 restore_reserve_on_error(h, dst_vma, dst_addr, page);
5180 put_page(page);
5181
5182 /* Allocate a temporary page to hold the copied
5183 * contents.
5184 */
5185 page = alloc_huge_page_vma(h, dst_vma, dst_addr);
5186 if (!page) {
5187 ret = -ENOMEM;
5188 goto out;
5189 }
5190 *pagep = page;
5191 /* Set the outparam pagep and return to the caller to
5192 * copy the contents outside the lock. Don't free the
5193 * page.
5194 */
5195 goto out;
5196 }
5197 } else {
5198 if (vm_shared &&
5199 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5200 put_page(*pagep);
5201 ret = -EEXIST;
5202 *pagep = NULL;
5203 goto out;
5204 }
5205
5206 page = alloc_huge_page(dst_vma, dst_addr, 0);
5207 if (IS_ERR(page)) {
5208 ret = -ENOMEM;
5209 *pagep = NULL;
5210 goto out;
5211 }
5212 copy_huge_page(page, *pagep);
5213 put_page(*pagep);
5214 *pagep = NULL;
5215 }
5216
5217 /*
5218 * The memory barrier inside __SetPageUptodate makes sure that
5219 * preceding stores to the page contents become visible before
5220 * the set_pte_at() write.
5221 */
5222 __SetPageUptodate(page);
5223
5224 /* Add shared, newly allocated pages to the page cache. */
5225 if (vm_shared && !is_continue) {
5226 size = i_size_read(mapping->host) >> huge_page_shift(h);
5227 ret = -EFAULT;
5228 if (idx >= size)
5229 goto out_release_nounlock;
5230
5231 /*
5232 * Serialization between remove_inode_hugepages() and
5233 * huge_add_to_page_cache() below happens through the
5234 * hugetlb_fault_mutex_table that here must be hold by
5235 * the caller.
5236 */
5237 ret = huge_add_to_page_cache(page, mapping, idx);
5238 if (ret)
5239 goto out_release_nounlock;
5240 new_pagecache_page = true;
5241 }
5242
5243 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
5244 spin_lock(ptl);
5245
5246 /*
5247 * Recheck the i_size after holding PT lock to make sure not
5248 * to leave any page mapped (as page_mapped()) beyond the end
5249 * of the i_size (remove_inode_hugepages() is strict about
5250 * enforcing that). If we bail out here, we'll also leave a
5251 * page in the radix tree in the vm_shared case beyond the end
5252 * of the i_size, but remove_inode_hugepages() will take care
5253 * of it as soon as we drop the hugetlb_fault_mutex_table.
5254 */
5255 size = i_size_read(mapping->host) >> huge_page_shift(h);
5256 ret = -EFAULT;
5257 if (idx >= size)
5258 goto out_release_unlock;
5259
5260 ret = -EEXIST;
5261 if (!huge_pte_none(huge_ptep_get(dst_pte)))
5262 goto out_release_unlock;
5263
5264 if (vm_shared) {
5265 page_dup_rmap(page, true);
5266 } else {
5267 ClearHPageRestoreReserve(page);
5268 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
5269 }
5270
5271 /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */
5272 if (is_continue && !vm_shared)
5273 writable = 0;
5274 else
5275 writable = dst_vma->vm_flags & VM_WRITE;
5276
5277 _dst_pte = make_huge_pte(dst_vma, page, writable);
5278 if (writable)
5279 _dst_pte = huge_pte_mkdirty(_dst_pte);
5280 _dst_pte = pte_mkyoung(_dst_pte);
5281
5282 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
5283
5284 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
5285 dst_vma->vm_flags & VM_WRITE);
5286 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
5287
5288 /* No need to invalidate - it was non-present before */
5289 update_mmu_cache(dst_vma, dst_addr, dst_pte);
5290
5291 spin_unlock(ptl);
5292 if (!is_continue)
5293 SetHPageMigratable(page);
5294 if (vm_shared || is_continue)
5295 unlock_page(page);
5296 ret = 0;
5297out:
5298 return ret;
5299out_release_unlock:
5300 spin_unlock(ptl);
5301 if (vm_shared || is_continue)
5302 unlock_page(page);
5303out_release_nounlock:
5304 if (!new_pagecache_page)
5305 restore_reserve_on_error(h, dst_vma, dst_addr, page);
5306 put_page(page);
5307 goto out;
5308}
5309#endif /* CONFIG_USERFAULTFD */
5310
5311static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
5312 int refs, struct page **pages,
5313 struct vm_area_struct **vmas)
5314{
5315 int nr;
5316
5317 for (nr = 0; nr < refs; nr++) {
5318 if (likely(pages))
5319 pages[nr] = mem_map_offset(page, nr);
5320 if (vmas)
5321 vmas[nr] = vma;
5322 }
5323}
5324
5325long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
5326 struct page **pages, struct vm_area_struct **vmas,
5327 unsigned long *position, unsigned long *nr_pages,
5328 long i, unsigned int flags, int *locked)
5329{
5330 unsigned long pfn_offset;
5331 unsigned long vaddr = *position;
5332 unsigned long remainder = *nr_pages;
5333 struct hstate *h = hstate_vma(vma);
5334 int err = -EFAULT, refs;
5335
5336 while (vaddr < vma->vm_end && remainder) {
5337 pte_t *pte;
5338 spinlock_t *ptl = NULL;
5339 int absent;
5340 struct page *page;
5341
5342 /*
5343 * If we have a pending SIGKILL, don't keep faulting pages and
5344 * potentially allocating memory.
5345 */
5346 if (fatal_signal_pending(current)) {
5347 remainder = 0;
5348 break;
5349 }
5350
5351 /*
5352 * Some archs (sparc64, sh*) have multiple pte_ts to
5353 * each hugepage. We have to make sure we get the
5354 * first, for the page indexing below to work.
5355 *
5356 * Note that page table lock is not held when pte is null.
5357 */
5358 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
5359 huge_page_size(h));
5360 if (pte)
5361 ptl = huge_pte_lock(h, mm, pte);
5362 absent = !pte || huge_pte_none(huge_ptep_get(pte));
5363
5364 /*
5365 * When coredumping, it suits get_dump_page if we just return
5366 * an error where there's an empty slot with no huge pagecache
5367 * to back it. This way, we avoid allocating a hugepage, and
5368 * the sparse dumpfile avoids allocating disk blocks, but its
5369 * huge holes still show up with zeroes where they need to be.
5370 */
5371 if (absent && (flags & FOLL_DUMP) &&
5372 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
5373 if (pte)
5374 spin_unlock(ptl);
5375 remainder = 0;
5376 break;
5377 }
5378
5379 /*
5380 * We need call hugetlb_fault for both hugepages under migration
5381 * (in which case hugetlb_fault waits for the migration,) and
5382 * hwpoisoned hugepages (in which case we need to prevent the
5383 * caller from accessing to them.) In order to do this, we use
5384 * here is_swap_pte instead of is_hugetlb_entry_migration and
5385 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
5386 * both cases, and because we can't follow correct pages
5387 * directly from any kind of swap entries.
5388 */
5389 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
5390 ((flags & FOLL_WRITE) &&
5391 !huge_pte_write(huge_ptep_get(pte)))) {
5392 vm_fault_t ret;
5393 unsigned int fault_flags = 0;
5394
5395 if (pte)
5396 spin_unlock(ptl);
5397 if (flags & FOLL_WRITE)
5398 fault_flags |= FAULT_FLAG_WRITE;
5399 if (locked)
5400 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
5401 FAULT_FLAG_KILLABLE;
5402 if (flags & FOLL_NOWAIT)
5403 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
5404 FAULT_FLAG_RETRY_NOWAIT;
5405 if (flags & FOLL_TRIED) {
5406 /*
5407 * Note: FAULT_FLAG_ALLOW_RETRY and
5408 * FAULT_FLAG_TRIED can co-exist
5409 */
5410 fault_flags |= FAULT_FLAG_TRIED;
5411 }
5412 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
5413 if (ret & VM_FAULT_ERROR) {
5414 err = vm_fault_to_errno(ret, flags);
5415 remainder = 0;
5416 break;
5417 }
5418 if (ret & VM_FAULT_RETRY) {
5419 if (locked &&
5420 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
5421 *locked = 0;
5422 *nr_pages = 0;
5423 /*
5424 * VM_FAULT_RETRY must not return an
5425 * error, it will return zero
5426 * instead.
5427 *
5428 * No need to update "position" as the
5429 * caller will not check it after
5430 * *nr_pages is set to 0.
5431 */
5432 return i;
5433 }
5434 continue;
5435 }
5436
5437 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
5438 page = pte_page(huge_ptep_get(pte));
5439
5440 /*
5441 * If subpage information not requested, update counters
5442 * and skip the same_page loop below.
5443 */
5444 if (!pages && !vmas && !pfn_offset &&
5445 (vaddr + huge_page_size(h) < vma->vm_end) &&
5446 (remainder >= pages_per_huge_page(h))) {
5447 vaddr += huge_page_size(h);
5448 remainder -= pages_per_huge_page(h);
5449 i += pages_per_huge_page(h);
5450 spin_unlock(ptl);
5451 continue;
5452 }
5453
5454 /* vaddr may not be aligned to PAGE_SIZE */
5455 refs = min3(pages_per_huge_page(h) - pfn_offset, remainder,
5456 (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT);
5457
5458 if (pages || vmas)
5459 record_subpages_vmas(mem_map_offset(page, pfn_offset),
5460 vma, refs,
5461 likely(pages) ? pages + i : NULL,
5462 vmas ? vmas + i : NULL);
5463
5464 if (pages) {
5465 /*
5466 * try_grab_compound_head() should always succeed here,
5467 * because: a) we hold the ptl lock, and b) we've just
5468 * checked that the huge page is present in the page
5469 * tables. If the huge page is present, then the tail
5470 * pages must also be present. The ptl prevents the
5471 * head page and tail pages from being rearranged in
5472 * any way. So this page must be available at this
5473 * point, unless the page refcount overflowed:
5474 */
5475 if (WARN_ON_ONCE(!try_grab_compound_head(pages[i],
5476 refs,
5477 flags))) {
5478 spin_unlock(ptl);
5479 remainder = 0;
5480 err = -ENOMEM;
5481 break;
5482 }
5483 }
5484
5485 vaddr += (refs << PAGE_SHIFT);
5486 remainder -= refs;
5487 i += refs;
5488
5489 spin_unlock(ptl);
5490 }
5491 *nr_pages = remainder;
5492 /*
5493 * setting position is actually required only if remainder is
5494 * not zero but it's faster not to add a "if (remainder)"
5495 * branch.
5496 */
5497 *position = vaddr;
5498
5499 return i ? i : err;
5500}
5501
5502unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
5503 unsigned long address, unsigned long end, pgprot_t newprot)
5504{
5505 struct mm_struct *mm = vma->vm_mm;
5506 unsigned long start = address;
5507 pte_t *ptep;
5508 pte_t pte;
5509 struct hstate *h = hstate_vma(vma);
5510 unsigned long pages = 0;
5511 bool shared_pmd = false;
5512 struct mmu_notifier_range range;
5513
5514 /*
5515 * In the case of shared PMDs, the area to flush could be beyond
5516 * start/end. Set range.start/range.end to cover the maximum possible
5517 * range if PMD sharing is possible.
5518 */
5519 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5520 0, vma, mm, start, end);
5521 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5522
5523 BUG_ON(address >= end);
5524 flush_cache_range(vma, range.start, range.end);
5525
5526 mmu_notifier_invalidate_range_start(&range);
5527 i_mmap_lock_write(vma->vm_file->f_mapping);
5528 for (; address < end; address += huge_page_size(h)) {
5529 spinlock_t *ptl;
5530 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5531 if (!ptep)
5532 continue;
5533 ptl = huge_pte_lock(h, mm, ptep);
5534 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
5535 pages++;
5536 spin_unlock(ptl);
5537 shared_pmd = true;
5538 continue;
5539 }
5540 pte = huge_ptep_get(ptep);
5541 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5542 spin_unlock(ptl);
5543 continue;
5544 }
5545 if (unlikely(is_hugetlb_entry_migration(pte))) {
5546 swp_entry_t entry = pte_to_swp_entry(pte);
5547
5548 if (is_writable_migration_entry(entry)) {
5549 pte_t newpte;
5550
5551 entry = make_readable_migration_entry(
5552 swp_offset(entry));
5553 newpte = swp_entry_to_pte(entry);
5554 set_huge_swap_pte_at(mm, address, ptep,
5555 newpte, huge_page_size(h));
5556 pages++;
5557 }
5558 spin_unlock(ptl);
5559 continue;
5560 }
5561 if (!huge_pte_none(pte)) {
5562 pte_t old_pte;
5563 unsigned int shift = huge_page_shift(hstate_vma(vma));
5564
5565 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5566 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
5567 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
5568 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
5569 pages++;
5570 }
5571 spin_unlock(ptl);
5572 }
5573 /*
5574 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5575 * may have cleared our pud entry and done put_page on the page table:
5576 * once we release i_mmap_rwsem, another task can do the final put_page
5577 * and that page table be reused and filled with junk. If we actually
5578 * did unshare a page of pmds, flush the range corresponding to the pud.
5579 */
5580 if (shared_pmd)
5581 flush_hugetlb_tlb_range(vma, range.start, range.end);
5582 else
5583 flush_hugetlb_tlb_range(vma, start, end);
5584 /*
5585 * No need to call mmu_notifier_invalidate_range() we are downgrading
5586 * page table protection not changing it to point to a new page.
5587 *
5588 * See Documentation/vm/mmu_notifier.rst
5589 */
5590 i_mmap_unlock_write(vma->vm_file->f_mapping);
5591 mmu_notifier_invalidate_range_end(&range);
5592
5593 return pages << h->order;
5594}
5595
5596/* Return true if reservation was successful, false otherwise. */
5597bool hugetlb_reserve_pages(struct inode *inode,
5598 long from, long to,
5599 struct vm_area_struct *vma,
5600 vm_flags_t vm_flags)
5601{
5602 long chg, add = -1;
5603 struct hstate *h = hstate_inode(inode);
5604 struct hugepage_subpool *spool = subpool_inode(inode);
5605 struct resv_map *resv_map;
5606 struct hugetlb_cgroup *h_cg = NULL;
5607 long gbl_reserve, regions_needed = 0;
5608
5609 /* This should never happen */
5610 if (from > to) {
5611 VM_WARN(1, "%s called with a negative range\n", __func__);
5612 return false;
5613 }
5614
5615 /*
5616 * Only apply hugepage reservation if asked. At fault time, an
5617 * attempt will be made for VM_NORESERVE to allocate a page
5618 * without using reserves
5619 */
5620 if (vm_flags & VM_NORESERVE)
5621 return true;
5622
5623 /*
5624 * Shared mappings base their reservation on the number of pages that
5625 * are already allocated on behalf of the file. Private mappings need
5626 * to reserve the full area even if read-only as mprotect() may be
5627 * called to make the mapping read-write. Assume !vma is a shm mapping
5628 */
5629 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5630 /*
5631 * resv_map can not be NULL as hugetlb_reserve_pages is only
5632 * called for inodes for which resv_maps were created (see
5633 * hugetlbfs_get_inode).
5634 */
5635 resv_map = inode_resv_map(inode);
5636
5637 chg = region_chg(resv_map, from, to, ®ions_needed);
5638
5639 } else {
5640 /* Private mapping. */
5641 resv_map = resv_map_alloc();
5642 if (!resv_map)
5643 return false;
5644
5645 chg = to - from;
5646
5647 set_vma_resv_map(vma, resv_map);
5648 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5649 }
5650
5651 if (chg < 0)
5652 goto out_err;
5653
5654 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
5655 chg * pages_per_huge_page(h), &h_cg) < 0)
5656 goto out_err;
5657
5658 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5659 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5660 * of the resv_map.
5661 */
5662 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5663 }
5664
5665 /*
5666 * There must be enough pages in the subpool for the mapping. If
5667 * the subpool has a minimum size, there may be some global
5668 * reservations already in place (gbl_reserve).
5669 */
5670 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5671 if (gbl_reserve < 0)
5672 goto out_uncharge_cgroup;
5673
5674 /*
5675 * Check enough hugepages are available for the reservation.
5676 * Hand the pages back to the subpool if there are not
5677 */
5678 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
5679 goto out_put_pages;
5680
5681 /*
5682 * Account for the reservations made. Shared mappings record regions
5683 * that have reservations as they are shared by multiple VMAs.
5684 * When the last VMA disappears, the region map says how much
5685 * the reservation was and the page cache tells how much of
5686 * the reservation was consumed. Private mappings are per-VMA and
5687 * only the consumed reservations are tracked. When the VMA
5688 * disappears, the original reservation is the VMA size and the
5689 * consumed reservations are stored in the map. Hence, nothing
5690 * else has to be done for private mappings here
5691 */
5692 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5693 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5694
5695 if (unlikely(add < 0)) {
5696 hugetlb_acct_memory(h, -gbl_reserve);
5697 goto out_put_pages;
5698 } else if (unlikely(chg > add)) {
5699 /*
5700 * pages in this range were added to the reserve
5701 * map between region_chg and region_add. This
5702 * indicates a race with alloc_huge_page. Adjust
5703 * the subpool and reserve counts modified above
5704 * based on the difference.
5705 */
5706 long rsv_adjust;
5707
5708 /*
5709 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
5710 * reference to h_cg->css. See comment below for detail.
5711 */
5712 hugetlb_cgroup_uncharge_cgroup_rsvd(
5713 hstate_index(h),
5714 (chg - add) * pages_per_huge_page(h), h_cg);
5715
5716 rsv_adjust = hugepage_subpool_put_pages(spool,
5717 chg - add);
5718 hugetlb_acct_memory(h, -rsv_adjust);
5719 } else if (h_cg) {
5720 /*
5721 * The file_regions will hold their own reference to
5722 * h_cg->css. So we should release the reference held
5723 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
5724 * done.
5725 */
5726 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
5727 }
5728 }
5729 return true;
5730
5731out_put_pages:
5732 /* put back original number of pages, chg */
5733 (void)hugepage_subpool_put_pages(spool, chg);
5734out_uncharge_cgroup:
5735 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5736 chg * pages_per_huge_page(h), h_cg);
5737out_err:
5738 if (!vma || vma->vm_flags & VM_MAYSHARE)
5739 /* Only call region_abort if the region_chg succeeded but the
5740 * region_add failed or didn't run.
5741 */
5742 if (chg >= 0 && add < 0)
5743 region_abort(resv_map, from, to, regions_needed);
5744 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5745 kref_put(&resv_map->refs, resv_map_release);
5746 return false;
5747}
5748
5749long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5750 long freed)
5751{
5752 struct hstate *h = hstate_inode(inode);
5753 struct resv_map *resv_map = inode_resv_map(inode);
5754 long chg = 0;
5755 struct hugepage_subpool *spool = subpool_inode(inode);
5756 long gbl_reserve;
5757
5758 /*
5759 * Since this routine can be called in the evict inode path for all
5760 * hugetlbfs inodes, resv_map could be NULL.
5761 */
5762 if (resv_map) {
5763 chg = region_del(resv_map, start, end);
5764 /*
5765 * region_del() can fail in the rare case where a region
5766 * must be split and another region descriptor can not be
5767 * allocated. If end == LONG_MAX, it will not fail.
5768 */
5769 if (chg < 0)
5770 return chg;
5771 }
5772
5773 spin_lock(&inode->i_lock);
5774 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5775 spin_unlock(&inode->i_lock);
5776
5777 /*
5778 * If the subpool has a minimum size, the number of global
5779 * reservations to be released may be adjusted.
5780 *
5781 * Note that !resv_map implies freed == 0. So (chg - freed)
5782 * won't go negative.
5783 */
5784 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5785 hugetlb_acct_memory(h, -gbl_reserve);
5786
5787 return 0;
5788}
5789
5790#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5791static unsigned long page_table_shareable(struct vm_area_struct *svma,
5792 struct vm_area_struct *vma,
5793 unsigned long addr, pgoff_t idx)
5794{
5795 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5796 svma->vm_start;
5797 unsigned long sbase = saddr & PUD_MASK;
5798 unsigned long s_end = sbase + PUD_SIZE;
5799
5800 /* Allow segments to share if only one is marked locked */
5801 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5802 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5803
5804 /*
5805 * match the virtual addresses, permission and the alignment of the
5806 * page table page.
5807 */
5808 if (pmd_index(addr) != pmd_index(saddr) ||
5809 vm_flags != svm_flags ||
5810 !range_in_vma(svma, sbase, s_end))
5811 return 0;
5812
5813 return saddr;
5814}
5815
5816static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5817{
5818 unsigned long base = addr & PUD_MASK;
5819 unsigned long end = base + PUD_SIZE;
5820
5821 /*
5822 * check on proper vm_flags and page table alignment
5823 */
5824 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5825 return true;
5826 return false;
5827}
5828
5829bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5830{
5831#ifdef CONFIG_USERFAULTFD
5832 if (uffd_disable_huge_pmd_share(vma))
5833 return false;
5834#endif
5835 return vma_shareable(vma, addr);
5836}
5837
5838/*
5839 * Determine if start,end range within vma could be mapped by shared pmd.
5840 * If yes, adjust start and end to cover range associated with possible
5841 * shared pmd mappings.
5842 */
5843void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5844 unsigned long *start, unsigned long *end)
5845{
5846 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
5847 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
5848
5849 /*
5850 * vma needs to span at least one aligned PUD size, and the range
5851 * must be at least partially within in.
5852 */
5853 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
5854 (*end <= v_start) || (*start >= v_end))
5855 return;
5856
5857 /* Extend the range to be PUD aligned for a worst case scenario */
5858 if (*start > v_start)
5859 *start = ALIGN_DOWN(*start, PUD_SIZE);
5860
5861 if (*end < v_end)
5862 *end = ALIGN(*end, PUD_SIZE);
5863}
5864
5865/*
5866 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5867 * and returns the corresponding pte. While this is not necessary for the
5868 * !shared pmd case because we can allocate the pmd later as well, it makes the
5869 * code much cleaner.
5870 *
5871 * This routine must be called with i_mmap_rwsem held in at least read mode if
5872 * sharing is possible. For hugetlbfs, this prevents removal of any page
5873 * table entries associated with the address space. This is important as we
5874 * are setting up sharing based on existing page table entries (mappings).
5875 *
5876 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5877 * huge_pte_alloc know that sharing is not possible and do not take
5878 * i_mmap_rwsem as a performance optimization. This is handled by the
5879 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5880 * only required for subsequent processing.
5881 */
5882pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5883 unsigned long addr, pud_t *pud)
5884{
5885 struct address_space *mapping = vma->vm_file->f_mapping;
5886 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5887 vma->vm_pgoff;
5888 struct vm_area_struct *svma;
5889 unsigned long saddr;
5890 pte_t *spte = NULL;
5891 pte_t *pte;
5892 spinlock_t *ptl;
5893
5894 i_mmap_assert_locked(mapping);
5895 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5896 if (svma == vma)
5897 continue;
5898
5899 saddr = page_table_shareable(svma, vma, addr, idx);
5900 if (saddr) {
5901 spte = huge_pte_offset(svma->vm_mm, saddr,
5902 vma_mmu_pagesize(svma));
5903 if (spte) {
5904 get_page(virt_to_page(spte));
5905 break;
5906 }
5907 }
5908 }
5909
5910 if (!spte)
5911 goto out;
5912
5913 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
5914 if (pud_none(*pud)) {
5915 pud_populate(mm, pud,
5916 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5917 mm_inc_nr_pmds(mm);
5918 } else {
5919 put_page(virt_to_page(spte));
5920 }
5921 spin_unlock(ptl);
5922out:
5923 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5924 return pte;
5925}
5926
5927/*
5928 * unmap huge page backed by shared pte.
5929 *
5930 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5931 * indicated by page_count > 1, unmap is achieved by clearing pud and
5932 * decrementing the ref count. If count == 1, the pte page is not shared.
5933 *
5934 * Called with page table lock held and i_mmap_rwsem held in write mode.
5935 *
5936 * returns: 1 successfully unmapped a shared pte page
5937 * 0 the underlying pte page is not shared, or it is the last user
5938 */
5939int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5940 unsigned long *addr, pte_t *ptep)
5941{
5942 pgd_t *pgd = pgd_offset(mm, *addr);
5943 p4d_t *p4d = p4d_offset(pgd, *addr);
5944 pud_t *pud = pud_offset(p4d, *addr);
5945
5946 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
5947 BUG_ON(page_count(virt_to_page(ptep)) == 0);
5948 if (page_count(virt_to_page(ptep)) == 1)
5949 return 0;
5950
5951 pud_clear(pud);
5952 put_page(virt_to_page(ptep));
5953 mm_dec_nr_pmds(mm);
5954 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5955 return 1;
5956}
5957
5958#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5959pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5960 unsigned long addr, pud_t *pud)
5961{
5962 return NULL;
5963}
5964
5965int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5966 unsigned long *addr, pte_t *ptep)
5967{
5968 return 0;
5969}
5970
5971void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5972 unsigned long *start, unsigned long *end)
5973{
5974}
5975
5976bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5977{
5978 return false;
5979}
5980#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5981
5982#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5983pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
5984 unsigned long addr, unsigned long sz)
5985{
5986 pgd_t *pgd;
5987 p4d_t *p4d;
5988 pud_t *pud;
5989 pte_t *pte = NULL;
5990
5991 pgd = pgd_offset(mm, addr);
5992 p4d = p4d_alloc(mm, pgd, addr);
5993 if (!p4d)
5994 return NULL;
5995 pud = pud_alloc(mm, p4d, addr);
5996 if (pud) {
5997 if (sz == PUD_SIZE) {
5998 pte = (pte_t *)pud;
5999 } else {
6000 BUG_ON(sz != PMD_SIZE);
6001 if (want_pmd_share(vma, addr) && pud_none(*pud))
6002 pte = huge_pmd_share(mm, vma, addr, pud);
6003 else
6004 pte = (pte_t *)pmd_alloc(mm, pud, addr);
6005 }
6006 }
6007 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
6008
6009 return pte;
6010}
6011
6012/*
6013 * huge_pte_offset() - Walk the page table to resolve the hugepage
6014 * entry at address @addr
6015 *
6016 * Return: Pointer to page table entry (PUD or PMD) for
6017 * address @addr, or NULL if a !p*d_present() entry is encountered and the
6018 * size @sz doesn't match the hugepage size at this level of the page
6019 * table.
6020 */
6021pte_t *huge_pte_offset(struct mm_struct *mm,
6022 unsigned long addr, unsigned long sz)
6023{
6024 pgd_t *pgd;
6025 p4d_t *p4d;
6026 pud_t *pud;
6027 pmd_t *pmd;
6028
6029 pgd = pgd_offset(mm, addr);
6030 if (!pgd_present(*pgd))
6031 return NULL;
6032 p4d = p4d_offset(pgd, addr);
6033 if (!p4d_present(*p4d))
6034 return NULL;
6035
6036 pud = pud_offset(p4d, addr);
6037 if (sz == PUD_SIZE)
6038 /* must be pud huge, non-present or none */
6039 return (pte_t *)pud;
6040 if (!pud_present(*pud))
6041 return NULL;
6042 /* must have a valid entry and size to go further */
6043
6044 pmd = pmd_offset(pud, addr);
6045 /* must be pmd huge, non-present or none */
6046 return (pte_t *)pmd;
6047}
6048
6049#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
6050
6051/*
6052 * These functions are overwritable if your architecture needs its own
6053 * behavior.
6054 */
6055struct page * __weak
6056follow_huge_addr(struct mm_struct *mm, unsigned long address,
6057 int write)
6058{
6059 return ERR_PTR(-EINVAL);
6060}
6061
6062struct page * __weak
6063follow_huge_pd(struct vm_area_struct *vma,
6064 unsigned long address, hugepd_t hpd, int flags, int pdshift)
6065{
6066 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
6067 return NULL;
6068}
6069
6070struct page * __weak
6071follow_huge_pmd(struct mm_struct *mm, unsigned long address,
6072 pmd_t *pmd, int flags)
6073{
6074 struct page *page = NULL;
6075 spinlock_t *ptl;
6076 pte_t pte;
6077
6078 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
6079 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
6080 (FOLL_PIN | FOLL_GET)))
6081 return NULL;
6082
6083retry:
6084 ptl = pmd_lockptr(mm, pmd);
6085 spin_lock(ptl);
6086 /*
6087 * make sure that the address range covered by this pmd is not
6088 * unmapped from other threads.
6089 */
6090 if (!pmd_huge(*pmd))
6091 goto out;
6092 pte = huge_ptep_get((pte_t *)pmd);
6093 if (pte_present(pte)) {
6094 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
6095 /*
6096 * try_grab_page() should always succeed here, because: a) we
6097 * hold the pmd (ptl) lock, and b) we've just checked that the
6098 * huge pmd (head) page is present in the page tables. The ptl
6099 * prevents the head page and tail pages from being rearranged
6100 * in any way. So this page must be available at this point,
6101 * unless the page refcount overflowed:
6102 */
6103 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
6104 page = NULL;
6105 goto out;
6106 }
6107 } else {
6108 if (is_hugetlb_entry_migration(pte)) {
6109 spin_unlock(ptl);
6110 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
6111 goto retry;
6112 }
6113 /*
6114 * hwpoisoned entry is treated as no_page_table in
6115 * follow_page_mask().
6116 */
6117 }
6118out:
6119 spin_unlock(ptl);
6120 return page;
6121}
6122
6123struct page * __weak
6124follow_huge_pud(struct mm_struct *mm, unsigned long address,
6125 pud_t *pud, int flags)
6126{
6127 if (flags & (FOLL_GET | FOLL_PIN))
6128 return NULL;
6129
6130 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
6131}
6132
6133struct page * __weak
6134follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
6135{
6136 if (flags & (FOLL_GET | FOLL_PIN))
6137 return NULL;
6138
6139 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
6140}
6141
6142bool isolate_huge_page(struct page *page, struct list_head *list)
6143{
6144 bool ret = true;
6145
6146 spin_lock_irq(&hugetlb_lock);
6147 if (!PageHeadHuge(page) ||
6148 !HPageMigratable(page) ||
6149 !get_page_unless_zero(page)) {
6150 ret = false;
6151 goto unlock;
6152 }
6153 ClearHPageMigratable(page);
6154 list_move_tail(&page->lru, list);
6155unlock:
6156 spin_unlock_irq(&hugetlb_lock);
6157 return ret;
6158}
6159
6160int get_hwpoison_huge_page(struct page *page, bool *hugetlb)
6161{
6162 int ret = 0;
6163
6164 *hugetlb = false;
6165 spin_lock_irq(&hugetlb_lock);
6166 if (PageHeadHuge(page)) {
6167 *hugetlb = true;
6168 if (HPageFreed(page) || HPageMigratable(page))
6169 ret = get_page_unless_zero(page);
6170 else
6171 ret = -EBUSY;
6172 }
6173 spin_unlock_irq(&hugetlb_lock);
6174 return ret;
6175}
6176
6177void putback_active_hugepage(struct page *page)
6178{
6179 spin_lock_irq(&hugetlb_lock);
6180 SetHPageMigratable(page);
6181 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
6182 spin_unlock_irq(&hugetlb_lock);
6183 put_page(page);
6184}
6185
6186void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
6187{
6188 struct hstate *h = page_hstate(oldpage);
6189
6190 hugetlb_cgroup_migrate(oldpage, newpage);
6191 set_page_owner_migrate_reason(newpage, reason);
6192
6193 /*
6194 * transfer temporary state of the new huge page. This is
6195 * reverse to other transitions because the newpage is going to
6196 * be final while the old one will be freed so it takes over
6197 * the temporary status.
6198 *
6199 * Also note that we have to transfer the per-node surplus state
6200 * here as well otherwise the global surplus count will not match
6201 * the per-node's.
6202 */
6203 if (HPageTemporary(newpage)) {
6204 int old_nid = page_to_nid(oldpage);
6205 int new_nid = page_to_nid(newpage);
6206
6207 SetHPageTemporary(oldpage);
6208 ClearHPageTemporary(newpage);
6209
6210 /*
6211 * There is no need to transfer the per-node surplus state
6212 * when we do not cross the node.
6213 */
6214 if (new_nid == old_nid)
6215 return;
6216 spin_lock_irq(&hugetlb_lock);
6217 if (h->surplus_huge_pages_node[old_nid]) {
6218 h->surplus_huge_pages_node[old_nid]--;
6219 h->surplus_huge_pages_node[new_nid]++;
6220 }
6221 spin_unlock_irq(&hugetlb_lock);
6222 }
6223}
6224
6225/*
6226 * This function will unconditionally remove all the shared pmd pgtable entries
6227 * within the specific vma for a hugetlbfs memory range.
6228 */
6229void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
6230{
6231 struct hstate *h = hstate_vma(vma);
6232 unsigned long sz = huge_page_size(h);
6233 struct mm_struct *mm = vma->vm_mm;
6234 struct mmu_notifier_range range;
6235 unsigned long address, start, end;
6236 spinlock_t *ptl;
6237 pte_t *ptep;
6238
6239 if (!(vma->vm_flags & VM_MAYSHARE))
6240 return;
6241
6242 start = ALIGN(vma->vm_start, PUD_SIZE);
6243 end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6244
6245 if (start >= end)
6246 return;
6247
6248 /*
6249 * No need to call adjust_range_if_pmd_sharing_possible(), because
6250 * we have already done the PUD_SIZE alignment.
6251 */
6252 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
6253 start, end);
6254 mmu_notifier_invalidate_range_start(&range);
6255 i_mmap_lock_write(vma->vm_file->f_mapping);
6256 for (address = start; address < end; address += PUD_SIZE) {
6257 unsigned long tmp = address;
6258
6259 ptep = huge_pte_offset(mm, address, sz);
6260 if (!ptep)
6261 continue;
6262 ptl = huge_pte_lock(h, mm, ptep);
6263 /* We don't want 'address' to be changed */
6264 huge_pmd_unshare(mm, vma, &tmp, ptep);
6265 spin_unlock(ptl);
6266 }
6267 flush_hugetlb_tlb_range(vma, start, end);
6268 i_mmap_unlock_write(vma->vm_file->f_mapping);
6269 /*
6270 * No need to call mmu_notifier_invalidate_range(), see
6271 * Documentation/vm/mmu_notifier.rst.
6272 */
6273 mmu_notifier_invalidate_range_end(&range);
6274}
6275
6276#ifdef CONFIG_CMA
6277static bool cma_reserve_called __initdata;
6278
6279static int __init cmdline_parse_hugetlb_cma(char *p)
6280{
6281 hugetlb_cma_size = memparse(p, &p);
6282 return 0;
6283}
6284
6285early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
6286
6287void __init hugetlb_cma_reserve(int order)
6288{
6289 unsigned long size, reserved, per_node;
6290 int nid;
6291
6292 cma_reserve_called = true;
6293
6294 if (!hugetlb_cma_size)
6295 return;
6296
6297 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
6298 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
6299 (PAGE_SIZE << order) / SZ_1M);
6300 return;
6301 }
6302
6303 /*
6304 * If 3 GB area is requested on a machine with 4 numa nodes,
6305 * let's allocate 1 GB on first three nodes and ignore the last one.
6306 */
6307 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
6308 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
6309 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
6310
6311 reserved = 0;
6312 for_each_node_state(nid, N_ONLINE) {
6313 int res;
6314 char name[CMA_MAX_NAME];
6315
6316 size = min(per_node, hugetlb_cma_size - reserved);
6317 size = round_up(size, PAGE_SIZE << order);
6318
6319 snprintf(name, sizeof(name), "hugetlb%d", nid);
6320 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
6321 0, false, name,
6322 &hugetlb_cma[nid], nid);
6323 if (res) {
6324 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
6325 res, nid);
6326 continue;
6327 }
6328
6329 reserved += size;
6330 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
6331 size / SZ_1M, nid);
6332
6333 if (reserved >= hugetlb_cma_size)
6334 break;
6335 }
6336}
6337
6338void __init hugetlb_cma_check(void)
6339{
6340 if (!hugetlb_cma_size || cma_reserve_called)
6341 return;
6342
6343 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
6344}
6345
6346#endif /* CONFIG_CMA */
1/*
2 * Generic hugetlb support.
3 * (C) William Irwin, April 2004
4 */
5#include <linux/list.h>
6#include <linux/init.h>
7#include <linux/module.h>
8#include <linux/mm.h>
9#include <linux/seq_file.h>
10#include <linux/sysctl.h>
11#include <linux/highmem.h>
12#include <linux/mmu_notifier.h>
13#include <linux/nodemask.h>
14#include <linux/pagemap.h>
15#include <linux/mempolicy.h>
16#include <linux/cpuset.h>
17#include <linux/mutex.h>
18#include <linux/bootmem.h>
19#include <linux/sysfs.h>
20#include <linux/slab.h>
21#include <linux/rmap.h>
22#include <linux/swap.h>
23#include <linux/swapops.h>
24
25#include <asm/page.h>
26#include <asm/pgtable.h>
27#include <linux/io.h>
28
29#include <linux/hugetlb.h>
30#include <linux/node.h>
31#include "internal.h"
32
33const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
34static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
35unsigned long hugepages_treat_as_movable;
36
37static int max_hstate;
38unsigned int default_hstate_idx;
39struct hstate hstates[HUGE_MAX_HSTATE];
40
41__initdata LIST_HEAD(huge_boot_pages);
42
43/* for command line parsing */
44static struct hstate * __initdata parsed_hstate;
45static unsigned long __initdata default_hstate_max_huge_pages;
46static unsigned long __initdata default_hstate_size;
47
48#define for_each_hstate(h) \
49 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
50
51/*
52 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
53 */
54static DEFINE_SPINLOCK(hugetlb_lock);
55
56static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
57{
58 bool free = (spool->count == 0) && (spool->used_hpages == 0);
59
60 spin_unlock(&spool->lock);
61
62 /* If no pages are used, and no other handles to the subpool
63 * remain, free the subpool the subpool remain */
64 if (free)
65 kfree(spool);
66}
67
68struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
69{
70 struct hugepage_subpool *spool;
71
72 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
73 if (!spool)
74 return NULL;
75
76 spin_lock_init(&spool->lock);
77 spool->count = 1;
78 spool->max_hpages = nr_blocks;
79 spool->used_hpages = 0;
80
81 return spool;
82}
83
84void hugepage_put_subpool(struct hugepage_subpool *spool)
85{
86 spin_lock(&spool->lock);
87 BUG_ON(!spool->count);
88 spool->count--;
89 unlock_or_release_subpool(spool);
90}
91
92static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
93 long delta)
94{
95 int ret = 0;
96
97 if (!spool)
98 return 0;
99
100 spin_lock(&spool->lock);
101 if ((spool->used_hpages + delta) <= spool->max_hpages) {
102 spool->used_hpages += delta;
103 } else {
104 ret = -ENOMEM;
105 }
106 spin_unlock(&spool->lock);
107
108 return ret;
109}
110
111static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
112 long delta)
113{
114 if (!spool)
115 return;
116
117 spin_lock(&spool->lock);
118 spool->used_hpages -= delta;
119 /* If hugetlbfs_put_super couldn't free spool due to
120 * an outstanding quota reference, free it now. */
121 unlock_or_release_subpool(spool);
122}
123
124static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
125{
126 return HUGETLBFS_SB(inode->i_sb)->spool;
127}
128
129static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
130{
131 return subpool_inode(vma->vm_file->f_dentry->d_inode);
132}
133
134/*
135 * Region tracking -- allows tracking of reservations and instantiated pages
136 * across the pages in a mapping.
137 *
138 * The region data structures are protected by a combination of the mmap_sem
139 * and the hugetlb_instantion_mutex. To access or modify a region the caller
140 * must either hold the mmap_sem for write, or the mmap_sem for read and
141 * the hugetlb_instantiation mutex:
142 *
143 * down_write(&mm->mmap_sem);
144 * or
145 * down_read(&mm->mmap_sem);
146 * mutex_lock(&hugetlb_instantiation_mutex);
147 */
148struct file_region {
149 struct list_head link;
150 long from;
151 long to;
152};
153
154static long region_add(struct list_head *head, long f, long t)
155{
156 struct file_region *rg, *nrg, *trg;
157
158 /* Locate the region we are either in or before. */
159 list_for_each_entry(rg, head, link)
160 if (f <= rg->to)
161 break;
162
163 /* Round our left edge to the current segment if it encloses us. */
164 if (f > rg->from)
165 f = rg->from;
166
167 /* Check for and consume any regions we now overlap with. */
168 nrg = rg;
169 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
170 if (&rg->link == head)
171 break;
172 if (rg->from > t)
173 break;
174
175 /* If this area reaches higher then extend our area to
176 * include it completely. If this is not the first area
177 * which we intend to reuse, free it. */
178 if (rg->to > t)
179 t = rg->to;
180 if (rg != nrg) {
181 list_del(&rg->link);
182 kfree(rg);
183 }
184 }
185 nrg->from = f;
186 nrg->to = t;
187 return 0;
188}
189
190static long region_chg(struct list_head *head, long f, long t)
191{
192 struct file_region *rg, *nrg;
193 long chg = 0;
194
195 /* Locate the region we are before or in. */
196 list_for_each_entry(rg, head, link)
197 if (f <= rg->to)
198 break;
199
200 /* If we are below the current region then a new region is required.
201 * Subtle, allocate a new region at the position but make it zero
202 * size such that we can guarantee to record the reservation. */
203 if (&rg->link == head || t < rg->from) {
204 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
205 if (!nrg)
206 return -ENOMEM;
207 nrg->from = f;
208 nrg->to = f;
209 INIT_LIST_HEAD(&nrg->link);
210 list_add(&nrg->link, rg->link.prev);
211
212 return t - f;
213 }
214
215 /* Round our left edge to the current segment if it encloses us. */
216 if (f > rg->from)
217 f = rg->from;
218 chg = t - f;
219
220 /* Check for and consume any regions we now overlap with. */
221 list_for_each_entry(rg, rg->link.prev, link) {
222 if (&rg->link == head)
223 break;
224 if (rg->from > t)
225 return chg;
226
227 /* We overlap with this area, if it extends further than
228 * us then we must extend ourselves. Account for its
229 * existing reservation. */
230 if (rg->to > t) {
231 chg += rg->to - t;
232 t = rg->to;
233 }
234 chg -= rg->to - rg->from;
235 }
236 return chg;
237}
238
239static long region_truncate(struct list_head *head, long end)
240{
241 struct file_region *rg, *trg;
242 long chg = 0;
243
244 /* Locate the region we are either in or before. */
245 list_for_each_entry(rg, head, link)
246 if (end <= rg->to)
247 break;
248 if (&rg->link == head)
249 return 0;
250
251 /* If we are in the middle of a region then adjust it. */
252 if (end > rg->from) {
253 chg = rg->to - end;
254 rg->to = end;
255 rg = list_entry(rg->link.next, typeof(*rg), link);
256 }
257
258 /* Drop any remaining regions. */
259 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
260 if (&rg->link == head)
261 break;
262 chg += rg->to - rg->from;
263 list_del(&rg->link);
264 kfree(rg);
265 }
266 return chg;
267}
268
269static long region_count(struct list_head *head, long f, long t)
270{
271 struct file_region *rg;
272 long chg = 0;
273
274 /* Locate each segment we overlap with, and count that overlap. */
275 list_for_each_entry(rg, head, link) {
276 long seg_from;
277 long seg_to;
278
279 if (rg->to <= f)
280 continue;
281 if (rg->from >= t)
282 break;
283
284 seg_from = max(rg->from, f);
285 seg_to = min(rg->to, t);
286
287 chg += seg_to - seg_from;
288 }
289
290 return chg;
291}
292
293/*
294 * Convert the address within this vma to the page offset within
295 * the mapping, in pagecache page units; huge pages here.
296 */
297static pgoff_t vma_hugecache_offset(struct hstate *h,
298 struct vm_area_struct *vma, unsigned long address)
299{
300 return ((address - vma->vm_start) >> huge_page_shift(h)) +
301 (vma->vm_pgoff >> huge_page_order(h));
302}
303
304pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
305 unsigned long address)
306{
307 return vma_hugecache_offset(hstate_vma(vma), vma, address);
308}
309
310/*
311 * Return the size of the pages allocated when backing a VMA. In the majority
312 * cases this will be same size as used by the page table entries.
313 */
314unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
315{
316 struct hstate *hstate;
317
318 if (!is_vm_hugetlb_page(vma))
319 return PAGE_SIZE;
320
321 hstate = hstate_vma(vma);
322
323 return 1UL << (hstate->order + PAGE_SHIFT);
324}
325EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
326
327/*
328 * Return the page size being used by the MMU to back a VMA. In the majority
329 * of cases, the page size used by the kernel matches the MMU size. On
330 * architectures where it differs, an architecture-specific version of this
331 * function is required.
332 */
333#ifndef vma_mmu_pagesize
334unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
335{
336 return vma_kernel_pagesize(vma);
337}
338#endif
339
340/*
341 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
342 * bits of the reservation map pointer, which are always clear due to
343 * alignment.
344 */
345#define HPAGE_RESV_OWNER (1UL << 0)
346#define HPAGE_RESV_UNMAPPED (1UL << 1)
347#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
348
349/*
350 * These helpers are used to track how many pages are reserved for
351 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
352 * is guaranteed to have their future faults succeed.
353 *
354 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
355 * the reserve counters are updated with the hugetlb_lock held. It is safe
356 * to reset the VMA at fork() time as it is not in use yet and there is no
357 * chance of the global counters getting corrupted as a result of the values.
358 *
359 * The private mapping reservation is represented in a subtly different
360 * manner to a shared mapping. A shared mapping has a region map associated
361 * with the underlying file, this region map represents the backing file
362 * pages which have ever had a reservation assigned which this persists even
363 * after the page is instantiated. A private mapping has a region map
364 * associated with the original mmap which is attached to all VMAs which
365 * reference it, this region map represents those offsets which have consumed
366 * reservation ie. where pages have been instantiated.
367 */
368static unsigned long get_vma_private_data(struct vm_area_struct *vma)
369{
370 return (unsigned long)vma->vm_private_data;
371}
372
373static void set_vma_private_data(struct vm_area_struct *vma,
374 unsigned long value)
375{
376 vma->vm_private_data = (void *)value;
377}
378
379struct resv_map {
380 struct kref refs;
381 struct list_head regions;
382};
383
384static struct resv_map *resv_map_alloc(void)
385{
386 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
387 if (!resv_map)
388 return NULL;
389
390 kref_init(&resv_map->refs);
391 INIT_LIST_HEAD(&resv_map->regions);
392
393 return resv_map;
394}
395
396static void resv_map_release(struct kref *ref)
397{
398 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
399
400 /* Clear out any active regions before we release the map. */
401 region_truncate(&resv_map->regions, 0);
402 kfree(resv_map);
403}
404
405static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
406{
407 VM_BUG_ON(!is_vm_hugetlb_page(vma));
408 if (!(vma->vm_flags & VM_MAYSHARE))
409 return (struct resv_map *)(get_vma_private_data(vma) &
410 ~HPAGE_RESV_MASK);
411 return NULL;
412}
413
414static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
415{
416 VM_BUG_ON(!is_vm_hugetlb_page(vma));
417 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
418
419 set_vma_private_data(vma, (get_vma_private_data(vma) &
420 HPAGE_RESV_MASK) | (unsigned long)map);
421}
422
423static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
424{
425 VM_BUG_ON(!is_vm_hugetlb_page(vma));
426 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
427
428 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
429}
430
431static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
432{
433 VM_BUG_ON(!is_vm_hugetlb_page(vma));
434
435 return (get_vma_private_data(vma) & flag) != 0;
436}
437
438/* Decrement the reserved pages in the hugepage pool by one */
439static void decrement_hugepage_resv_vma(struct hstate *h,
440 struct vm_area_struct *vma)
441{
442 if (vma->vm_flags & VM_NORESERVE)
443 return;
444
445 if (vma->vm_flags & VM_MAYSHARE) {
446 /* Shared mappings always use reserves */
447 h->resv_huge_pages--;
448 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
449 /*
450 * Only the process that called mmap() has reserves for
451 * private mappings.
452 */
453 h->resv_huge_pages--;
454 }
455}
456
457/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
458void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
459{
460 VM_BUG_ON(!is_vm_hugetlb_page(vma));
461 if (!(vma->vm_flags & VM_MAYSHARE))
462 vma->vm_private_data = (void *)0;
463}
464
465/* Returns true if the VMA has associated reserve pages */
466static int vma_has_reserves(struct vm_area_struct *vma)
467{
468 if (vma->vm_flags & VM_MAYSHARE)
469 return 1;
470 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
471 return 1;
472 return 0;
473}
474
475static void copy_gigantic_page(struct page *dst, struct page *src)
476{
477 int i;
478 struct hstate *h = page_hstate(src);
479 struct page *dst_base = dst;
480 struct page *src_base = src;
481
482 for (i = 0; i < pages_per_huge_page(h); ) {
483 cond_resched();
484 copy_highpage(dst, src);
485
486 i++;
487 dst = mem_map_next(dst, dst_base, i);
488 src = mem_map_next(src, src_base, i);
489 }
490}
491
492void copy_huge_page(struct page *dst, struct page *src)
493{
494 int i;
495 struct hstate *h = page_hstate(src);
496
497 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
498 copy_gigantic_page(dst, src);
499 return;
500 }
501
502 might_sleep();
503 for (i = 0; i < pages_per_huge_page(h); i++) {
504 cond_resched();
505 copy_highpage(dst + i, src + i);
506 }
507}
508
509static void enqueue_huge_page(struct hstate *h, struct page *page)
510{
511 int nid = page_to_nid(page);
512 list_add(&page->lru, &h->hugepage_freelists[nid]);
513 h->free_huge_pages++;
514 h->free_huge_pages_node[nid]++;
515}
516
517static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
518{
519 struct page *page;
520
521 if (list_empty(&h->hugepage_freelists[nid]))
522 return NULL;
523 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
524 list_del(&page->lru);
525 set_page_refcounted(page);
526 h->free_huge_pages--;
527 h->free_huge_pages_node[nid]--;
528 return page;
529}
530
531static struct page *dequeue_huge_page_vma(struct hstate *h,
532 struct vm_area_struct *vma,
533 unsigned long address, int avoid_reserve)
534{
535 struct page *page = NULL;
536 struct mempolicy *mpol;
537 nodemask_t *nodemask;
538 struct zonelist *zonelist;
539 struct zone *zone;
540 struct zoneref *z;
541 unsigned int cpuset_mems_cookie;
542
543retry_cpuset:
544 cpuset_mems_cookie = get_mems_allowed();
545 zonelist = huge_zonelist(vma, address,
546 htlb_alloc_mask, &mpol, &nodemask);
547 /*
548 * A child process with MAP_PRIVATE mappings created by their parent
549 * have no page reserves. This check ensures that reservations are
550 * not "stolen". The child may still get SIGKILLed
551 */
552 if (!vma_has_reserves(vma) &&
553 h->free_huge_pages - h->resv_huge_pages == 0)
554 goto err;
555
556 /* If reserves cannot be used, ensure enough pages are in the pool */
557 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
558 goto err;
559
560 for_each_zone_zonelist_nodemask(zone, z, zonelist,
561 MAX_NR_ZONES - 1, nodemask) {
562 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
563 page = dequeue_huge_page_node(h, zone_to_nid(zone));
564 if (page) {
565 if (!avoid_reserve)
566 decrement_hugepage_resv_vma(h, vma);
567 break;
568 }
569 }
570 }
571
572 mpol_cond_put(mpol);
573 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
574 goto retry_cpuset;
575 return page;
576
577err:
578 mpol_cond_put(mpol);
579 return NULL;
580}
581
582static void update_and_free_page(struct hstate *h, struct page *page)
583{
584 int i;
585
586 VM_BUG_ON(h->order >= MAX_ORDER);
587
588 h->nr_huge_pages--;
589 h->nr_huge_pages_node[page_to_nid(page)]--;
590 for (i = 0; i < pages_per_huge_page(h); i++) {
591 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
592 1 << PG_referenced | 1 << PG_dirty |
593 1 << PG_active | 1 << PG_reserved |
594 1 << PG_private | 1 << PG_writeback);
595 }
596 set_compound_page_dtor(page, NULL);
597 set_page_refcounted(page);
598 arch_release_hugepage(page);
599 __free_pages(page, huge_page_order(h));
600}
601
602struct hstate *size_to_hstate(unsigned long size)
603{
604 struct hstate *h;
605
606 for_each_hstate(h) {
607 if (huge_page_size(h) == size)
608 return h;
609 }
610 return NULL;
611}
612
613static void free_huge_page(struct page *page)
614{
615 /*
616 * Can't pass hstate in here because it is called from the
617 * compound page destructor.
618 */
619 struct hstate *h = page_hstate(page);
620 int nid = page_to_nid(page);
621 struct hugepage_subpool *spool =
622 (struct hugepage_subpool *)page_private(page);
623
624 set_page_private(page, 0);
625 page->mapping = NULL;
626 BUG_ON(page_count(page));
627 BUG_ON(page_mapcount(page));
628 INIT_LIST_HEAD(&page->lru);
629
630 spin_lock(&hugetlb_lock);
631 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
632 update_and_free_page(h, page);
633 h->surplus_huge_pages--;
634 h->surplus_huge_pages_node[nid]--;
635 } else {
636 enqueue_huge_page(h, page);
637 }
638 spin_unlock(&hugetlb_lock);
639 hugepage_subpool_put_pages(spool, 1);
640}
641
642static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
643{
644 set_compound_page_dtor(page, free_huge_page);
645 spin_lock(&hugetlb_lock);
646 h->nr_huge_pages++;
647 h->nr_huge_pages_node[nid]++;
648 spin_unlock(&hugetlb_lock);
649 put_page(page); /* free it into the hugepage allocator */
650}
651
652static void prep_compound_gigantic_page(struct page *page, unsigned long order)
653{
654 int i;
655 int nr_pages = 1 << order;
656 struct page *p = page + 1;
657
658 /* we rely on prep_new_huge_page to set the destructor */
659 set_compound_order(page, order);
660 __SetPageHead(page);
661 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
662 __SetPageTail(p);
663 set_page_count(p, 0);
664 p->first_page = page;
665 }
666}
667
668int PageHuge(struct page *page)
669{
670 compound_page_dtor *dtor;
671
672 if (!PageCompound(page))
673 return 0;
674
675 page = compound_head(page);
676 dtor = get_compound_page_dtor(page);
677
678 return dtor == free_huge_page;
679}
680EXPORT_SYMBOL_GPL(PageHuge);
681
682static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
683{
684 struct page *page;
685
686 if (h->order >= MAX_ORDER)
687 return NULL;
688
689 page = alloc_pages_exact_node(nid,
690 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
691 __GFP_REPEAT|__GFP_NOWARN,
692 huge_page_order(h));
693 if (page) {
694 if (arch_prepare_hugepage(page)) {
695 __free_pages(page, huge_page_order(h));
696 return NULL;
697 }
698 prep_new_huge_page(h, page, nid);
699 }
700
701 return page;
702}
703
704/*
705 * common helper functions for hstate_next_node_to_{alloc|free}.
706 * We may have allocated or freed a huge page based on a different
707 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
708 * be outside of *nodes_allowed. Ensure that we use an allowed
709 * node for alloc or free.
710 */
711static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
712{
713 nid = next_node(nid, *nodes_allowed);
714 if (nid == MAX_NUMNODES)
715 nid = first_node(*nodes_allowed);
716 VM_BUG_ON(nid >= MAX_NUMNODES);
717
718 return nid;
719}
720
721static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
722{
723 if (!node_isset(nid, *nodes_allowed))
724 nid = next_node_allowed(nid, nodes_allowed);
725 return nid;
726}
727
728/*
729 * returns the previously saved node ["this node"] from which to
730 * allocate a persistent huge page for the pool and advance the
731 * next node from which to allocate, handling wrap at end of node
732 * mask.
733 */
734static int hstate_next_node_to_alloc(struct hstate *h,
735 nodemask_t *nodes_allowed)
736{
737 int nid;
738
739 VM_BUG_ON(!nodes_allowed);
740
741 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
742 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
743
744 return nid;
745}
746
747static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
748{
749 struct page *page;
750 int start_nid;
751 int next_nid;
752 int ret = 0;
753
754 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
755 next_nid = start_nid;
756
757 do {
758 page = alloc_fresh_huge_page_node(h, next_nid);
759 if (page) {
760 ret = 1;
761 break;
762 }
763 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
764 } while (next_nid != start_nid);
765
766 if (ret)
767 count_vm_event(HTLB_BUDDY_PGALLOC);
768 else
769 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
770
771 return ret;
772}
773
774/*
775 * helper for free_pool_huge_page() - return the previously saved
776 * node ["this node"] from which to free a huge page. Advance the
777 * next node id whether or not we find a free huge page to free so
778 * that the next attempt to free addresses the next node.
779 */
780static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
781{
782 int nid;
783
784 VM_BUG_ON(!nodes_allowed);
785
786 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
787 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
788
789 return nid;
790}
791
792/*
793 * Free huge page from pool from next node to free.
794 * Attempt to keep persistent huge pages more or less
795 * balanced over allowed nodes.
796 * Called with hugetlb_lock locked.
797 */
798static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
799 bool acct_surplus)
800{
801 int start_nid;
802 int next_nid;
803 int ret = 0;
804
805 start_nid = hstate_next_node_to_free(h, nodes_allowed);
806 next_nid = start_nid;
807
808 do {
809 /*
810 * If we're returning unused surplus pages, only examine
811 * nodes with surplus pages.
812 */
813 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
814 !list_empty(&h->hugepage_freelists[next_nid])) {
815 struct page *page =
816 list_entry(h->hugepage_freelists[next_nid].next,
817 struct page, lru);
818 list_del(&page->lru);
819 h->free_huge_pages--;
820 h->free_huge_pages_node[next_nid]--;
821 if (acct_surplus) {
822 h->surplus_huge_pages--;
823 h->surplus_huge_pages_node[next_nid]--;
824 }
825 update_and_free_page(h, page);
826 ret = 1;
827 break;
828 }
829 next_nid = hstate_next_node_to_free(h, nodes_allowed);
830 } while (next_nid != start_nid);
831
832 return ret;
833}
834
835static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
836{
837 struct page *page;
838 unsigned int r_nid;
839
840 if (h->order >= MAX_ORDER)
841 return NULL;
842
843 /*
844 * Assume we will successfully allocate the surplus page to
845 * prevent racing processes from causing the surplus to exceed
846 * overcommit
847 *
848 * This however introduces a different race, where a process B
849 * tries to grow the static hugepage pool while alloc_pages() is
850 * called by process A. B will only examine the per-node
851 * counters in determining if surplus huge pages can be
852 * converted to normal huge pages in adjust_pool_surplus(). A
853 * won't be able to increment the per-node counter, until the
854 * lock is dropped by B, but B doesn't drop hugetlb_lock until
855 * no more huge pages can be converted from surplus to normal
856 * state (and doesn't try to convert again). Thus, we have a
857 * case where a surplus huge page exists, the pool is grown, and
858 * the surplus huge page still exists after, even though it
859 * should just have been converted to a normal huge page. This
860 * does not leak memory, though, as the hugepage will be freed
861 * once it is out of use. It also does not allow the counters to
862 * go out of whack in adjust_pool_surplus() as we don't modify
863 * the node values until we've gotten the hugepage and only the
864 * per-node value is checked there.
865 */
866 spin_lock(&hugetlb_lock);
867 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
868 spin_unlock(&hugetlb_lock);
869 return NULL;
870 } else {
871 h->nr_huge_pages++;
872 h->surplus_huge_pages++;
873 }
874 spin_unlock(&hugetlb_lock);
875
876 if (nid == NUMA_NO_NODE)
877 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
878 __GFP_REPEAT|__GFP_NOWARN,
879 huge_page_order(h));
880 else
881 page = alloc_pages_exact_node(nid,
882 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
883 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
884
885 if (page && arch_prepare_hugepage(page)) {
886 __free_pages(page, huge_page_order(h));
887 page = NULL;
888 }
889
890 spin_lock(&hugetlb_lock);
891 if (page) {
892 r_nid = page_to_nid(page);
893 set_compound_page_dtor(page, free_huge_page);
894 /*
895 * We incremented the global counters already
896 */
897 h->nr_huge_pages_node[r_nid]++;
898 h->surplus_huge_pages_node[r_nid]++;
899 __count_vm_event(HTLB_BUDDY_PGALLOC);
900 } else {
901 h->nr_huge_pages--;
902 h->surplus_huge_pages--;
903 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
904 }
905 spin_unlock(&hugetlb_lock);
906
907 return page;
908}
909
910/*
911 * This allocation function is useful in the context where vma is irrelevant.
912 * E.g. soft-offlining uses this function because it only cares physical
913 * address of error page.
914 */
915struct page *alloc_huge_page_node(struct hstate *h, int nid)
916{
917 struct page *page;
918
919 spin_lock(&hugetlb_lock);
920 page = dequeue_huge_page_node(h, nid);
921 spin_unlock(&hugetlb_lock);
922
923 if (!page)
924 page = alloc_buddy_huge_page(h, nid);
925
926 return page;
927}
928
929/*
930 * Increase the hugetlb pool such that it can accommodate a reservation
931 * of size 'delta'.
932 */
933static int gather_surplus_pages(struct hstate *h, int delta)
934{
935 struct list_head surplus_list;
936 struct page *page, *tmp;
937 int ret, i;
938 int needed, allocated;
939 bool alloc_ok = true;
940
941 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
942 if (needed <= 0) {
943 h->resv_huge_pages += delta;
944 return 0;
945 }
946
947 allocated = 0;
948 INIT_LIST_HEAD(&surplus_list);
949
950 ret = -ENOMEM;
951retry:
952 spin_unlock(&hugetlb_lock);
953 for (i = 0; i < needed; i++) {
954 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
955 if (!page) {
956 alloc_ok = false;
957 break;
958 }
959 list_add(&page->lru, &surplus_list);
960 }
961 allocated += i;
962
963 /*
964 * After retaking hugetlb_lock, we need to recalculate 'needed'
965 * because either resv_huge_pages or free_huge_pages may have changed.
966 */
967 spin_lock(&hugetlb_lock);
968 needed = (h->resv_huge_pages + delta) -
969 (h->free_huge_pages + allocated);
970 if (needed > 0) {
971 if (alloc_ok)
972 goto retry;
973 /*
974 * We were not able to allocate enough pages to
975 * satisfy the entire reservation so we free what
976 * we've allocated so far.
977 */
978 goto free;
979 }
980 /*
981 * The surplus_list now contains _at_least_ the number of extra pages
982 * needed to accommodate the reservation. Add the appropriate number
983 * of pages to the hugetlb pool and free the extras back to the buddy
984 * allocator. Commit the entire reservation here to prevent another
985 * process from stealing the pages as they are added to the pool but
986 * before they are reserved.
987 */
988 needed += allocated;
989 h->resv_huge_pages += delta;
990 ret = 0;
991
992 /* Free the needed pages to the hugetlb pool */
993 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
994 if ((--needed) < 0)
995 break;
996 list_del(&page->lru);
997 /*
998 * This page is now managed by the hugetlb allocator and has
999 * no users -- drop the buddy allocator's reference.
1000 */
1001 put_page_testzero(page);
1002 VM_BUG_ON(page_count(page));
1003 enqueue_huge_page(h, page);
1004 }
1005free:
1006 spin_unlock(&hugetlb_lock);
1007
1008 /* Free unnecessary surplus pages to the buddy allocator */
1009 if (!list_empty(&surplus_list)) {
1010 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1011 list_del(&page->lru);
1012 put_page(page);
1013 }
1014 }
1015 spin_lock(&hugetlb_lock);
1016
1017 return ret;
1018}
1019
1020/*
1021 * When releasing a hugetlb pool reservation, any surplus pages that were
1022 * allocated to satisfy the reservation must be explicitly freed if they were
1023 * never used.
1024 * Called with hugetlb_lock held.
1025 */
1026static void return_unused_surplus_pages(struct hstate *h,
1027 unsigned long unused_resv_pages)
1028{
1029 unsigned long nr_pages;
1030
1031 /* Uncommit the reservation */
1032 h->resv_huge_pages -= unused_resv_pages;
1033
1034 /* Cannot return gigantic pages currently */
1035 if (h->order >= MAX_ORDER)
1036 return;
1037
1038 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1039
1040 /*
1041 * We want to release as many surplus pages as possible, spread
1042 * evenly across all nodes with memory. Iterate across these nodes
1043 * until we can no longer free unreserved surplus pages. This occurs
1044 * when the nodes with surplus pages have no free pages.
1045 * free_pool_huge_page() will balance the the freed pages across the
1046 * on-line nodes with memory and will handle the hstate accounting.
1047 */
1048 while (nr_pages--) {
1049 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
1050 break;
1051 }
1052}
1053
1054/*
1055 * Determine if the huge page at addr within the vma has an associated
1056 * reservation. Where it does not we will need to logically increase
1057 * reservation and actually increase subpool usage before an allocation
1058 * can occur. Where any new reservation would be required the
1059 * reservation change is prepared, but not committed. Once the page
1060 * has been allocated from the subpool and instantiated the change should
1061 * be committed via vma_commit_reservation. No action is required on
1062 * failure.
1063 */
1064static long vma_needs_reservation(struct hstate *h,
1065 struct vm_area_struct *vma, unsigned long addr)
1066{
1067 struct address_space *mapping = vma->vm_file->f_mapping;
1068 struct inode *inode = mapping->host;
1069
1070 if (vma->vm_flags & VM_MAYSHARE) {
1071 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1072 return region_chg(&inode->i_mapping->private_list,
1073 idx, idx + 1);
1074
1075 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1076 return 1;
1077
1078 } else {
1079 long err;
1080 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1081 struct resv_map *reservations = vma_resv_map(vma);
1082
1083 err = region_chg(&reservations->regions, idx, idx + 1);
1084 if (err < 0)
1085 return err;
1086 return 0;
1087 }
1088}
1089static void vma_commit_reservation(struct hstate *h,
1090 struct vm_area_struct *vma, unsigned long addr)
1091{
1092 struct address_space *mapping = vma->vm_file->f_mapping;
1093 struct inode *inode = mapping->host;
1094
1095 if (vma->vm_flags & VM_MAYSHARE) {
1096 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1097 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1098
1099 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1100 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1101 struct resv_map *reservations = vma_resv_map(vma);
1102
1103 /* Mark this page used in the map. */
1104 region_add(&reservations->regions, idx, idx + 1);
1105 }
1106}
1107
1108static struct page *alloc_huge_page(struct vm_area_struct *vma,
1109 unsigned long addr, int avoid_reserve)
1110{
1111 struct hugepage_subpool *spool = subpool_vma(vma);
1112 struct hstate *h = hstate_vma(vma);
1113 struct page *page;
1114 long chg;
1115
1116 /*
1117 * Processes that did not create the mapping will have no
1118 * reserves and will not have accounted against subpool
1119 * limit. Check that the subpool limit can be made before
1120 * satisfying the allocation MAP_NORESERVE mappings may also
1121 * need pages and subpool limit allocated allocated if no reserve
1122 * mapping overlaps.
1123 */
1124 chg = vma_needs_reservation(h, vma, addr);
1125 if (chg < 0)
1126 return ERR_PTR(-VM_FAULT_OOM);
1127 if (chg)
1128 if (hugepage_subpool_get_pages(spool, chg))
1129 return ERR_PTR(-VM_FAULT_SIGBUS);
1130
1131 spin_lock(&hugetlb_lock);
1132 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1133 spin_unlock(&hugetlb_lock);
1134
1135 if (!page) {
1136 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1137 if (!page) {
1138 hugepage_subpool_put_pages(spool, chg);
1139 return ERR_PTR(-VM_FAULT_SIGBUS);
1140 }
1141 }
1142
1143 set_page_private(page, (unsigned long)spool);
1144
1145 vma_commit_reservation(h, vma, addr);
1146
1147 return page;
1148}
1149
1150int __weak alloc_bootmem_huge_page(struct hstate *h)
1151{
1152 struct huge_bootmem_page *m;
1153 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1154
1155 while (nr_nodes) {
1156 void *addr;
1157
1158 addr = __alloc_bootmem_node_nopanic(
1159 NODE_DATA(hstate_next_node_to_alloc(h,
1160 &node_states[N_HIGH_MEMORY])),
1161 huge_page_size(h), huge_page_size(h), 0);
1162
1163 if (addr) {
1164 /*
1165 * Use the beginning of the huge page to store the
1166 * huge_bootmem_page struct (until gather_bootmem
1167 * puts them into the mem_map).
1168 */
1169 m = addr;
1170 goto found;
1171 }
1172 nr_nodes--;
1173 }
1174 return 0;
1175
1176found:
1177 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1178 /* Put them into a private list first because mem_map is not up yet */
1179 list_add(&m->list, &huge_boot_pages);
1180 m->hstate = h;
1181 return 1;
1182}
1183
1184static void prep_compound_huge_page(struct page *page, int order)
1185{
1186 if (unlikely(order > (MAX_ORDER - 1)))
1187 prep_compound_gigantic_page(page, order);
1188 else
1189 prep_compound_page(page, order);
1190}
1191
1192/* Put bootmem huge pages into the standard lists after mem_map is up */
1193static void __init gather_bootmem_prealloc(void)
1194{
1195 struct huge_bootmem_page *m;
1196
1197 list_for_each_entry(m, &huge_boot_pages, list) {
1198 struct hstate *h = m->hstate;
1199 struct page *page;
1200
1201#ifdef CONFIG_HIGHMEM
1202 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1203 free_bootmem_late((unsigned long)m,
1204 sizeof(struct huge_bootmem_page));
1205#else
1206 page = virt_to_page(m);
1207#endif
1208 __ClearPageReserved(page);
1209 WARN_ON(page_count(page) != 1);
1210 prep_compound_huge_page(page, h->order);
1211 prep_new_huge_page(h, page, page_to_nid(page));
1212 /*
1213 * If we had gigantic hugepages allocated at boot time, we need
1214 * to restore the 'stolen' pages to totalram_pages in order to
1215 * fix confusing memory reports from free(1) and another
1216 * side-effects, like CommitLimit going negative.
1217 */
1218 if (h->order > (MAX_ORDER - 1))
1219 totalram_pages += 1 << h->order;
1220 }
1221}
1222
1223static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1224{
1225 unsigned long i;
1226
1227 for (i = 0; i < h->max_huge_pages; ++i) {
1228 if (h->order >= MAX_ORDER) {
1229 if (!alloc_bootmem_huge_page(h))
1230 break;
1231 } else if (!alloc_fresh_huge_page(h,
1232 &node_states[N_HIGH_MEMORY]))
1233 break;
1234 }
1235 h->max_huge_pages = i;
1236}
1237
1238static void __init hugetlb_init_hstates(void)
1239{
1240 struct hstate *h;
1241
1242 for_each_hstate(h) {
1243 /* oversize hugepages were init'ed in early boot */
1244 if (h->order < MAX_ORDER)
1245 hugetlb_hstate_alloc_pages(h);
1246 }
1247}
1248
1249static char * __init memfmt(char *buf, unsigned long n)
1250{
1251 if (n >= (1UL << 30))
1252 sprintf(buf, "%lu GB", n >> 30);
1253 else if (n >= (1UL << 20))
1254 sprintf(buf, "%lu MB", n >> 20);
1255 else
1256 sprintf(buf, "%lu KB", n >> 10);
1257 return buf;
1258}
1259
1260static void __init report_hugepages(void)
1261{
1262 struct hstate *h;
1263
1264 for_each_hstate(h) {
1265 char buf[32];
1266 printk(KERN_INFO "HugeTLB registered %s page size, "
1267 "pre-allocated %ld pages\n",
1268 memfmt(buf, huge_page_size(h)),
1269 h->free_huge_pages);
1270 }
1271}
1272
1273#ifdef CONFIG_HIGHMEM
1274static void try_to_free_low(struct hstate *h, unsigned long count,
1275 nodemask_t *nodes_allowed)
1276{
1277 int i;
1278
1279 if (h->order >= MAX_ORDER)
1280 return;
1281
1282 for_each_node_mask(i, *nodes_allowed) {
1283 struct page *page, *next;
1284 struct list_head *freel = &h->hugepage_freelists[i];
1285 list_for_each_entry_safe(page, next, freel, lru) {
1286 if (count >= h->nr_huge_pages)
1287 return;
1288 if (PageHighMem(page))
1289 continue;
1290 list_del(&page->lru);
1291 update_and_free_page(h, page);
1292 h->free_huge_pages--;
1293 h->free_huge_pages_node[page_to_nid(page)]--;
1294 }
1295 }
1296}
1297#else
1298static inline void try_to_free_low(struct hstate *h, unsigned long count,
1299 nodemask_t *nodes_allowed)
1300{
1301}
1302#endif
1303
1304/*
1305 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1306 * balanced by operating on them in a round-robin fashion.
1307 * Returns 1 if an adjustment was made.
1308 */
1309static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1310 int delta)
1311{
1312 int start_nid, next_nid;
1313 int ret = 0;
1314
1315 VM_BUG_ON(delta != -1 && delta != 1);
1316
1317 if (delta < 0)
1318 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1319 else
1320 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1321 next_nid = start_nid;
1322
1323 do {
1324 int nid = next_nid;
1325 if (delta < 0) {
1326 /*
1327 * To shrink on this node, there must be a surplus page
1328 */
1329 if (!h->surplus_huge_pages_node[nid]) {
1330 next_nid = hstate_next_node_to_alloc(h,
1331 nodes_allowed);
1332 continue;
1333 }
1334 }
1335 if (delta > 0) {
1336 /*
1337 * Surplus cannot exceed the total number of pages
1338 */
1339 if (h->surplus_huge_pages_node[nid] >=
1340 h->nr_huge_pages_node[nid]) {
1341 next_nid = hstate_next_node_to_free(h,
1342 nodes_allowed);
1343 continue;
1344 }
1345 }
1346
1347 h->surplus_huge_pages += delta;
1348 h->surplus_huge_pages_node[nid] += delta;
1349 ret = 1;
1350 break;
1351 } while (next_nid != start_nid);
1352
1353 return ret;
1354}
1355
1356#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1357static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1358 nodemask_t *nodes_allowed)
1359{
1360 unsigned long min_count, ret;
1361
1362 if (h->order >= MAX_ORDER)
1363 return h->max_huge_pages;
1364
1365 /*
1366 * Increase the pool size
1367 * First take pages out of surplus state. Then make up the
1368 * remaining difference by allocating fresh huge pages.
1369 *
1370 * We might race with alloc_buddy_huge_page() here and be unable
1371 * to convert a surplus huge page to a normal huge page. That is
1372 * not critical, though, it just means the overall size of the
1373 * pool might be one hugepage larger than it needs to be, but
1374 * within all the constraints specified by the sysctls.
1375 */
1376 spin_lock(&hugetlb_lock);
1377 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1378 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1379 break;
1380 }
1381
1382 while (count > persistent_huge_pages(h)) {
1383 /*
1384 * If this allocation races such that we no longer need the
1385 * page, free_huge_page will handle it by freeing the page
1386 * and reducing the surplus.
1387 */
1388 spin_unlock(&hugetlb_lock);
1389 ret = alloc_fresh_huge_page(h, nodes_allowed);
1390 spin_lock(&hugetlb_lock);
1391 if (!ret)
1392 goto out;
1393
1394 /* Bail for signals. Probably ctrl-c from user */
1395 if (signal_pending(current))
1396 goto out;
1397 }
1398
1399 /*
1400 * Decrease the pool size
1401 * First return free pages to the buddy allocator (being careful
1402 * to keep enough around to satisfy reservations). Then place
1403 * pages into surplus state as needed so the pool will shrink
1404 * to the desired size as pages become free.
1405 *
1406 * By placing pages into the surplus state independent of the
1407 * overcommit value, we are allowing the surplus pool size to
1408 * exceed overcommit. There are few sane options here. Since
1409 * alloc_buddy_huge_page() is checking the global counter,
1410 * though, we'll note that we're not allowed to exceed surplus
1411 * and won't grow the pool anywhere else. Not until one of the
1412 * sysctls are changed, or the surplus pages go out of use.
1413 */
1414 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1415 min_count = max(count, min_count);
1416 try_to_free_low(h, min_count, nodes_allowed);
1417 while (min_count < persistent_huge_pages(h)) {
1418 if (!free_pool_huge_page(h, nodes_allowed, 0))
1419 break;
1420 }
1421 while (count < persistent_huge_pages(h)) {
1422 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1423 break;
1424 }
1425out:
1426 ret = persistent_huge_pages(h);
1427 spin_unlock(&hugetlb_lock);
1428 return ret;
1429}
1430
1431#define HSTATE_ATTR_RO(_name) \
1432 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1433
1434#define HSTATE_ATTR(_name) \
1435 static struct kobj_attribute _name##_attr = \
1436 __ATTR(_name, 0644, _name##_show, _name##_store)
1437
1438static struct kobject *hugepages_kobj;
1439static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1440
1441static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1442
1443static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1444{
1445 int i;
1446
1447 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1448 if (hstate_kobjs[i] == kobj) {
1449 if (nidp)
1450 *nidp = NUMA_NO_NODE;
1451 return &hstates[i];
1452 }
1453
1454 return kobj_to_node_hstate(kobj, nidp);
1455}
1456
1457static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1458 struct kobj_attribute *attr, char *buf)
1459{
1460 struct hstate *h;
1461 unsigned long nr_huge_pages;
1462 int nid;
1463
1464 h = kobj_to_hstate(kobj, &nid);
1465 if (nid == NUMA_NO_NODE)
1466 nr_huge_pages = h->nr_huge_pages;
1467 else
1468 nr_huge_pages = h->nr_huge_pages_node[nid];
1469
1470 return sprintf(buf, "%lu\n", nr_huge_pages);
1471}
1472
1473static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1474 struct kobject *kobj, struct kobj_attribute *attr,
1475 const char *buf, size_t len)
1476{
1477 int err;
1478 int nid;
1479 unsigned long count;
1480 struct hstate *h;
1481 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1482
1483 err = strict_strtoul(buf, 10, &count);
1484 if (err)
1485 goto out;
1486
1487 h = kobj_to_hstate(kobj, &nid);
1488 if (h->order >= MAX_ORDER) {
1489 err = -EINVAL;
1490 goto out;
1491 }
1492
1493 if (nid == NUMA_NO_NODE) {
1494 /*
1495 * global hstate attribute
1496 */
1497 if (!(obey_mempolicy &&
1498 init_nodemask_of_mempolicy(nodes_allowed))) {
1499 NODEMASK_FREE(nodes_allowed);
1500 nodes_allowed = &node_states[N_HIGH_MEMORY];
1501 }
1502 } else if (nodes_allowed) {
1503 /*
1504 * per node hstate attribute: adjust count to global,
1505 * but restrict alloc/free to the specified node.
1506 */
1507 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1508 init_nodemask_of_node(nodes_allowed, nid);
1509 } else
1510 nodes_allowed = &node_states[N_HIGH_MEMORY];
1511
1512 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1513
1514 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1515 NODEMASK_FREE(nodes_allowed);
1516
1517 return len;
1518out:
1519 NODEMASK_FREE(nodes_allowed);
1520 return err;
1521}
1522
1523static ssize_t nr_hugepages_show(struct kobject *kobj,
1524 struct kobj_attribute *attr, char *buf)
1525{
1526 return nr_hugepages_show_common(kobj, attr, buf);
1527}
1528
1529static ssize_t nr_hugepages_store(struct kobject *kobj,
1530 struct kobj_attribute *attr, const char *buf, size_t len)
1531{
1532 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1533}
1534HSTATE_ATTR(nr_hugepages);
1535
1536#ifdef CONFIG_NUMA
1537
1538/*
1539 * hstate attribute for optionally mempolicy-based constraint on persistent
1540 * huge page alloc/free.
1541 */
1542static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1543 struct kobj_attribute *attr, char *buf)
1544{
1545 return nr_hugepages_show_common(kobj, attr, buf);
1546}
1547
1548static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1549 struct kobj_attribute *attr, const char *buf, size_t len)
1550{
1551 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1552}
1553HSTATE_ATTR(nr_hugepages_mempolicy);
1554#endif
1555
1556
1557static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1558 struct kobj_attribute *attr, char *buf)
1559{
1560 struct hstate *h = kobj_to_hstate(kobj, NULL);
1561 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1562}
1563
1564static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1565 struct kobj_attribute *attr, const char *buf, size_t count)
1566{
1567 int err;
1568 unsigned long input;
1569 struct hstate *h = kobj_to_hstate(kobj, NULL);
1570
1571 if (h->order >= MAX_ORDER)
1572 return -EINVAL;
1573
1574 err = strict_strtoul(buf, 10, &input);
1575 if (err)
1576 return err;
1577
1578 spin_lock(&hugetlb_lock);
1579 h->nr_overcommit_huge_pages = input;
1580 spin_unlock(&hugetlb_lock);
1581
1582 return count;
1583}
1584HSTATE_ATTR(nr_overcommit_hugepages);
1585
1586static ssize_t free_hugepages_show(struct kobject *kobj,
1587 struct kobj_attribute *attr, char *buf)
1588{
1589 struct hstate *h;
1590 unsigned long free_huge_pages;
1591 int nid;
1592
1593 h = kobj_to_hstate(kobj, &nid);
1594 if (nid == NUMA_NO_NODE)
1595 free_huge_pages = h->free_huge_pages;
1596 else
1597 free_huge_pages = h->free_huge_pages_node[nid];
1598
1599 return sprintf(buf, "%lu\n", free_huge_pages);
1600}
1601HSTATE_ATTR_RO(free_hugepages);
1602
1603static ssize_t resv_hugepages_show(struct kobject *kobj,
1604 struct kobj_attribute *attr, char *buf)
1605{
1606 struct hstate *h = kobj_to_hstate(kobj, NULL);
1607 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1608}
1609HSTATE_ATTR_RO(resv_hugepages);
1610
1611static ssize_t surplus_hugepages_show(struct kobject *kobj,
1612 struct kobj_attribute *attr, char *buf)
1613{
1614 struct hstate *h;
1615 unsigned long surplus_huge_pages;
1616 int nid;
1617
1618 h = kobj_to_hstate(kobj, &nid);
1619 if (nid == NUMA_NO_NODE)
1620 surplus_huge_pages = h->surplus_huge_pages;
1621 else
1622 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1623
1624 return sprintf(buf, "%lu\n", surplus_huge_pages);
1625}
1626HSTATE_ATTR_RO(surplus_hugepages);
1627
1628static struct attribute *hstate_attrs[] = {
1629 &nr_hugepages_attr.attr,
1630 &nr_overcommit_hugepages_attr.attr,
1631 &free_hugepages_attr.attr,
1632 &resv_hugepages_attr.attr,
1633 &surplus_hugepages_attr.attr,
1634#ifdef CONFIG_NUMA
1635 &nr_hugepages_mempolicy_attr.attr,
1636#endif
1637 NULL,
1638};
1639
1640static struct attribute_group hstate_attr_group = {
1641 .attrs = hstate_attrs,
1642};
1643
1644static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1645 struct kobject **hstate_kobjs,
1646 struct attribute_group *hstate_attr_group)
1647{
1648 int retval;
1649 int hi = h - hstates;
1650
1651 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1652 if (!hstate_kobjs[hi])
1653 return -ENOMEM;
1654
1655 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1656 if (retval)
1657 kobject_put(hstate_kobjs[hi]);
1658
1659 return retval;
1660}
1661
1662static void __init hugetlb_sysfs_init(void)
1663{
1664 struct hstate *h;
1665 int err;
1666
1667 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1668 if (!hugepages_kobj)
1669 return;
1670
1671 for_each_hstate(h) {
1672 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1673 hstate_kobjs, &hstate_attr_group);
1674 if (err)
1675 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1676 h->name);
1677 }
1678}
1679
1680#ifdef CONFIG_NUMA
1681
1682/*
1683 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1684 * with node devices in node_devices[] using a parallel array. The array
1685 * index of a node device or _hstate == node id.
1686 * This is here to avoid any static dependency of the node device driver, in
1687 * the base kernel, on the hugetlb module.
1688 */
1689struct node_hstate {
1690 struct kobject *hugepages_kobj;
1691 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1692};
1693struct node_hstate node_hstates[MAX_NUMNODES];
1694
1695/*
1696 * A subset of global hstate attributes for node devices
1697 */
1698static struct attribute *per_node_hstate_attrs[] = {
1699 &nr_hugepages_attr.attr,
1700 &free_hugepages_attr.attr,
1701 &surplus_hugepages_attr.attr,
1702 NULL,
1703};
1704
1705static struct attribute_group per_node_hstate_attr_group = {
1706 .attrs = per_node_hstate_attrs,
1707};
1708
1709/*
1710 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1711 * Returns node id via non-NULL nidp.
1712 */
1713static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1714{
1715 int nid;
1716
1717 for (nid = 0; nid < nr_node_ids; nid++) {
1718 struct node_hstate *nhs = &node_hstates[nid];
1719 int i;
1720 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1721 if (nhs->hstate_kobjs[i] == kobj) {
1722 if (nidp)
1723 *nidp = nid;
1724 return &hstates[i];
1725 }
1726 }
1727
1728 BUG();
1729 return NULL;
1730}
1731
1732/*
1733 * Unregister hstate attributes from a single node device.
1734 * No-op if no hstate attributes attached.
1735 */
1736void hugetlb_unregister_node(struct node *node)
1737{
1738 struct hstate *h;
1739 struct node_hstate *nhs = &node_hstates[node->dev.id];
1740
1741 if (!nhs->hugepages_kobj)
1742 return; /* no hstate attributes */
1743
1744 for_each_hstate(h)
1745 if (nhs->hstate_kobjs[h - hstates]) {
1746 kobject_put(nhs->hstate_kobjs[h - hstates]);
1747 nhs->hstate_kobjs[h - hstates] = NULL;
1748 }
1749
1750 kobject_put(nhs->hugepages_kobj);
1751 nhs->hugepages_kobj = NULL;
1752}
1753
1754/*
1755 * hugetlb module exit: unregister hstate attributes from node devices
1756 * that have them.
1757 */
1758static void hugetlb_unregister_all_nodes(void)
1759{
1760 int nid;
1761
1762 /*
1763 * disable node device registrations.
1764 */
1765 register_hugetlbfs_with_node(NULL, NULL);
1766
1767 /*
1768 * remove hstate attributes from any nodes that have them.
1769 */
1770 for (nid = 0; nid < nr_node_ids; nid++)
1771 hugetlb_unregister_node(&node_devices[nid]);
1772}
1773
1774/*
1775 * Register hstate attributes for a single node device.
1776 * No-op if attributes already registered.
1777 */
1778void hugetlb_register_node(struct node *node)
1779{
1780 struct hstate *h;
1781 struct node_hstate *nhs = &node_hstates[node->dev.id];
1782 int err;
1783
1784 if (nhs->hugepages_kobj)
1785 return; /* already allocated */
1786
1787 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1788 &node->dev.kobj);
1789 if (!nhs->hugepages_kobj)
1790 return;
1791
1792 for_each_hstate(h) {
1793 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1794 nhs->hstate_kobjs,
1795 &per_node_hstate_attr_group);
1796 if (err) {
1797 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1798 " for node %d\n",
1799 h->name, node->dev.id);
1800 hugetlb_unregister_node(node);
1801 break;
1802 }
1803 }
1804}
1805
1806/*
1807 * hugetlb init time: register hstate attributes for all registered node
1808 * devices of nodes that have memory. All on-line nodes should have
1809 * registered their associated device by this time.
1810 */
1811static void hugetlb_register_all_nodes(void)
1812{
1813 int nid;
1814
1815 for_each_node_state(nid, N_HIGH_MEMORY) {
1816 struct node *node = &node_devices[nid];
1817 if (node->dev.id == nid)
1818 hugetlb_register_node(node);
1819 }
1820
1821 /*
1822 * Let the node device driver know we're here so it can
1823 * [un]register hstate attributes on node hotplug.
1824 */
1825 register_hugetlbfs_with_node(hugetlb_register_node,
1826 hugetlb_unregister_node);
1827}
1828#else /* !CONFIG_NUMA */
1829
1830static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1831{
1832 BUG();
1833 if (nidp)
1834 *nidp = -1;
1835 return NULL;
1836}
1837
1838static void hugetlb_unregister_all_nodes(void) { }
1839
1840static void hugetlb_register_all_nodes(void) { }
1841
1842#endif
1843
1844static void __exit hugetlb_exit(void)
1845{
1846 struct hstate *h;
1847
1848 hugetlb_unregister_all_nodes();
1849
1850 for_each_hstate(h) {
1851 kobject_put(hstate_kobjs[h - hstates]);
1852 }
1853
1854 kobject_put(hugepages_kobj);
1855}
1856module_exit(hugetlb_exit);
1857
1858static int __init hugetlb_init(void)
1859{
1860 /* Some platform decide whether they support huge pages at boot
1861 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1862 * there is no such support
1863 */
1864 if (HPAGE_SHIFT == 0)
1865 return 0;
1866
1867 if (!size_to_hstate(default_hstate_size)) {
1868 default_hstate_size = HPAGE_SIZE;
1869 if (!size_to_hstate(default_hstate_size))
1870 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1871 }
1872 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1873 if (default_hstate_max_huge_pages)
1874 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1875
1876 hugetlb_init_hstates();
1877
1878 gather_bootmem_prealloc();
1879
1880 report_hugepages();
1881
1882 hugetlb_sysfs_init();
1883
1884 hugetlb_register_all_nodes();
1885
1886 return 0;
1887}
1888module_init(hugetlb_init);
1889
1890/* Should be called on processing a hugepagesz=... option */
1891void __init hugetlb_add_hstate(unsigned order)
1892{
1893 struct hstate *h;
1894 unsigned long i;
1895
1896 if (size_to_hstate(PAGE_SIZE << order)) {
1897 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1898 return;
1899 }
1900 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1901 BUG_ON(order == 0);
1902 h = &hstates[max_hstate++];
1903 h->order = order;
1904 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1905 h->nr_huge_pages = 0;
1906 h->free_huge_pages = 0;
1907 for (i = 0; i < MAX_NUMNODES; ++i)
1908 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1909 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1910 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1911 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1912 huge_page_size(h)/1024);
1913
1914 parsed_hstate = h;
1915}
1916
1917static int __init hugetlb_nrpages_setup(char *s)
1918{
1919 unsigned long *mhp;
1920 static unsigned long *last_mhp;
1921
1922 /*
1923 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1924 * so this hugepages= parameter goes to the "default hstate".
1925 */
1926 if (!max_hstate)
1927 mhp = &default_hstate_max_huge_pages;
1928 else
1929 mhp = &parsed_hstate->max_huge_pages;
1930
1931 if (mhp == last_mhp) {
1932 printk(KERN_WARNING "hugepages= specified twice without "
1933 "interleaving hugepagesz=, ignoring\n");
1934 return 1;
1935 }
1936
1937 if (sscanf(s, "%lu", mhp) <= 0)
1938 *mhp = 0;
1939
1940 /*
1941 * Global state is always initialized later in hugetlb_init.
1942 * But we need to allocate >= MAX_ORDER hstates here early to still
1943 * use the bootmem allocator.
1944 */
1945 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1946 hugetlb_hstate_alloc_pages(parsed_hstate);
1947
1948 last_mhp = mhp;
1949
1950 return 1;
1951}
1952__setup("hugepages=", hugetlb_nrpages_setup);
1953
1954static int __init hugetlb_default_setup(char *s)
1955{
1956 default_hstate_size = memparse(s, &s);
1957 return 1;
1958}
1959__setup("default_hugepagesz=", hugetlb_default_setup);
1960
1961static unsigned int cpuset_mems_nr(unsigned int *array)
1962{
1963 int node;
1964 unsigned int nr = 0;
1965
1966 for_each_node_mask(node, cpuset_current_mems_allowed)
1967 nr += array[node];
1968
1969 return nr;
1970}
1971
1972#ifdef CONFIG_SYSCTL
1973static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1974 struct ctl_table *table, int write,
1975 void __user *buffer, size_t *length, loff_t *ppos)
1976{
1977 struct hstate *h = &default_hstate;
1978 unsigned long tmp;
1979 int ret;
1980
1981 tmp = h->max_huge_pages;
1982
1983 if (write && h->order >= MAX_ORDER)
1984 return -EINVAL;
1985
1986 table->data = &tmp;
1987 table->maxlen = sizeof(unsigned long);
1988 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1989 if (ret)
1990 goto out;
1991
1992 if (write) {
1993 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1994 GFP_KERNEL | __GFP_NORETRY);
1995 if (!(obey_mempolicy &&
1996 init_nodemask_of_mempolicy(nodes_allowed))) {
1997 NODEMASK_FREE(nodes_allowed);
1998 nodes_allowed = &node_states[N_HIGH_MEMORY];
1999 }
2000 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2001
2002 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
2003 NODEMASK_FREE(nodes_allowed);
2004 }
2005out:
2006 return ret;
2007}
2008
2009int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2010 void __user *buffer, size_t *length, loff_t *ppos)
2011{
2012
2013 return hugetlb_sysctl_handler_common(false, table, write,
2014 buffer, length, ppos);
2015}
2016
2017#ifdef CONFIG_NUMA
2018int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2019 void __user *buffer, size_t *length, loff_t *ppos)
2020{
2021 return hugetlb_sysctl_handler_common(true, table, write,
2022 buffer, length, ppos);
2023}
2024#endif /* CONFIG_NUMA */
2025
2026int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2027 void __user *buffer,
2028 size_t *length, loff_t *ppos)
2029{
2030 proc_dointvec(table, write, buffer, length, ppos);
2031 if (hugepages_treat_as_movable)
2032 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2033 else
2034 htlb_alloc_mask = GFP_HIGHUSER;
2035 return 0;
2036}
2037
2038int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2039 void __user *buffer,
2040 size_t *length, loff_t *ppos)
2041{
2042 struct hstate *h = &default_hstate;
2043 unsigned long tmp;
2044 int ret;
2045
2046 tmp = h->nr_overcommit_huge_pages;
2047
2048 if (write && h->order >= MAX_ORDER)
2049 return -EINVAL;
2050
2051 table->data = &tmp;
2052 table->maxlen = sizeof(unsigned long);
2053 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2054 if (ret)
2055 goto out;
2056
2057 if (write) {
2058 spin_lock(&hugetlb_lock);
2059 h->nr_overcommit_huge_pages = tmp;
2060 spin_unlock(&hugetlb_lock);
2061 }
2062out:
2063 return ret;
2064}
2065
2066#endif /* CONFIG_SYSCTL */
2067
2068void hugetlb_report_meminfo(struct seq_file *m)
2069{
2070 struct hstate *h = &default_hstate;
2071 seq_printf(m,
2072 "HugePages_Total: %5lu\n"
2073 "HugePages_Free: %5lu\n"
2074 "HugePages_Rsvd: %5lu\n"
2075 "HugePages_Surp: %5lu\n"
2076 "Hugepagesize: %8lu kB\n",
2077 h->nr_huge_pages,
2078 h->free_huge_pages,
2079 h->resv_huge_pages,
2080 h->surplus_huge_pages,
2081 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2082}
2083
2084int hugetlb_report_node_meminfo(int nid, char *buf)
2085{
2086 struct hstate *h = &default_hstate;
2087 return sprintf(buf,
2088 "Node %d HugePages_Total: %5u\n"
2089 "Node %d HugePages_Free: %5u\n"
2090 "Node %d HugePages_Surp: %5u\n",
2091 nid, h->nr_huge_pages_node[nid],
2092 nid, h->free_huge_pages_node[nid],
2093 nid, h->surplus_huge_pages_node[nid]);
2094}
2095
2096/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2097unsigned long hugetlb_total_pages(void)
2098{
2099 struct hstate *h = &default_hstate;
2100 return h->nr_huge_pages * pages_per_huge_page(h);
2101}
2102
2103static int hugetlb_acct_memory(struct hstate *h, long delta)
2104{
2105 int ret = -ENOMEM;
2106
2107 spin_lock(&hugetlb_lock);
2108 /*
2109 * When cpuset is configured, it breaks the strict hugetlb page
2110 * reservation as the accounting is done on a global variable. Such
2111 * reservation is completely rubbish in the presence of cpuset because
2112 * the reservation is not checked against page availability for the
2113 * current cpuset. Application can still potentially OOM'ed by kernel
2114 * with lack of free htlb page in cpuset that the task is in.
2115 * Attempt to enforce strict accounting with cpuset is almost
2116 * impossible (or too ugly) because cpuset is too fluid that
2117 * task or memory node can be dynamically moved between cpusets.
2118 *
2119 * The change of semantics for shared hugetlb mapping with cpuset is
2120 * undesirable. However, in order to preserve some of the semantics,
2121 * we fall back to check against current free page availability as
2122 * a best attempt and hopefully to minimize the impact of changing
2123 * semantics that cpuset has.
2124 */
2125 if (delta > 0) {
2126 if (gather_surplus_pages(h, delta) < 0)
2127 goto out;
2128
2129 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2130 return_unused_surplus_pages(h, delta);
2131 goto out;
2132 }
2133 }
2134
2135 ret = 0;
2136 if (delta < 0)
2137 return_unused_surplus_pages(h, (unsigned long) -delta);
2138
2139out:
2140 spin_unlock(&hugetlb_lock);
2141 return ret;
2142}
2143
2144static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2145{
2146 struct resv_map *reservations = vma_resv_map(vma);
2147
2148 /*
2149 * This new VMA should share its siblings reservation map if present.
2150 * The VMA will only ever have a valid reservation map pointer where
2151 * it is being copied for another still existing VMA. As that VMA
2152 * has a reference to the reservation map it cannot disappear until
2153 * after this open call completes. It is therefore safe to take a
2154 * new reference here without additional locking.
2155 */
2156 if (reservations)
2157 kref_get(&reservations->refs);
2158}
2159
2160static void resv_map_put(struct vm_area_struct *vma)
2161{
2162 struct resv_map *reservations = vma_resv_map(vma);
2163
2164 if (!reservations)
2165 return;
2166 kref_put(&reservations->refs, resv_map_release);
2167}
2168
2169static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2170{
2171 struct hstate *h = hstate_vma(vma);
2172 struct resv_map *reservations = vma_resv_map(vma);
2173 struct hugepage_subpool *spool = subpool_vma(vma);
2174 unsigned long reserve;
2175 unsigned long start;
2176 unsigned long end;
2177
2178 if (reservations) {
2179 start = vma_hugecache_offset(h, vma, vma->vm_start);
2180 end = vma_hugecache_offset(h, vma, vma->vm_end);
2181
2182 reserve = (end - start) -
2183 region_count(&reservations->regions, start, end);
2184
2185 resv_map_put(vma);
2186
2187 if (reserve) {
2188 hugetlb_acct_memory(h, -reserve);
2189 hugepage_subpool_put_pages(spool, reserve);
2190 }
2191 }
2192}
2193
2194/*
2195 * We cannot handle pagefaults against hugetlb pages at all. They cause
2196 * handle_mm_fault() to try to instantiate regular-sized pages in the
2197 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2198 * this far.
2199 */
2200static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2201{
2202 BUG();
2203 return 0;
2204}
2205
2206const struct vm_operations_struct hugetlb_vm_ops = {
2207 .fault = hugetlb_vm_op_fault,
2208 .open = hugetlb_vm_op_open,
2209 .close = hugetlb_vm_op_close,
2210};
2211
2212static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2213 int writable)
2214{
2215 pte_t entry;
2216
2217 if (writable) {
2218 entry =
2219 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2220 } else {
2221 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2222 }
2223 entry = pte_mkyoung(entry);
2224 entry = pte_mkhuge(entry);
2225 entry = arch_make_huge_pte(entry, vma, page, writable);
2226
2227 return entry;
2228}
2229
2230static void set_huge_ptep_writable(struct vm_area_struct *vma,
2231 unsigned long address, pte_t *ptep)
2232{
2233 pte_t entry;
2234
2235 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2236 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2237 update_mmu_cache(vma, address, ptep);
2238}
2239
2240
2241int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2242 struct vm_area_struct *vma)
2243{
2244 pte_t *src_pte, *dst_pte, entry;
2245 struct page *ptepage;
2246 unsigned long addr;
2247 int cow;
2248 struct hstate *h = hstate_vma(vma);
2249 unsigned long sz = huge_page_size(h);
2250
2251 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2252
2253 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2254 src_pte = huge_pte_offset(src, addr);
2255 if (!src_pte)
2256 continue;
2257 dst_pte = huge_pte_alloc(dst, addr, sz);
2258 if (!dst_pte)
2259 goto nomem;
2260
2261 /* If the pagetables are shared don't copy or take references */
2262 if (dst_pte == src_pte)
2263 continue;
2264
2265 spin_lock(&dst->page_table_lock);
2266 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2267 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2268 if (cow)
2269 huge_ptep_set_wrprotect(src, addr, src_pte);
2270 entry = huge_ptep_get(src_pte);
2271 ptepage = pte_page(entry);
2272 get_page(ptepage);
2273 page_dup_rmap(ptepage);
2274 set_huge_pte_at(dst, addr, dst_pte, entry);
2275 }
2276 spin_unlock(&src->page_table_lock);
2277 spin_unlock(&dst->page_table_lock);
2278 }
2279 return 0;
2280
2281nomem:
2282 return -ENOMEM;
2283}
2284
2285static int is_hugetlb_entry_migration(pte_t pte)
2286{
2287 swp_entry_t swp;
2288
2289 if (huge_pte_none(pte) || pte_present(pte))
2290 return 0;
2291 swp = pte_to_swp_entry(pte);
2292 if (non_swap_entry(swp) && is_migration_entry(swp))
2293 return 1;
2294 else
2295 return 0;
2296}
2297
2298static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2299{
2300 swp_entry_t swp;
2301
2302 if (huge_pte_none(pte) || pte_present(pte))
2303 return 0;
2304 swp = pte_to_swp_entry(pte);
2305 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2306 return 1;
2307 else
2308 return 0;
2309}
2310
2311void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2312 unsigned long end, struct page *ref_page)
2313{
2314 struct mm_struct *mm = vma->vm_mm;
2315 unsigned long address;
2316 pte_t *ptep;
2317 pte_t pte;
2318 struct page *page;
2319 struct page *tmp;
2320 struct hstate *h = hstate_vma(vma);
2321 unsigned long sz = huge_page_size(h);
2322
2323 /*
2324 * A page gathering list, protected by per file i_mmap_mutex. The
2325 * lock is used to avoid list corruption from multiple unmapping
2326 * of the same page since we are using page->lru.
2327 */
2328 LIST_HEAD(page_list);
2329
2330 WARN_ON(!is_vm_hugetlb_page(vma));
2331 BUG_ON(start & ~huge_page_mask(h));
2332 BUG_ON(end & ~huge_page_mask(h));
2333
2334 mmu_notifier_invalidate_range_start(mm, start, end);
2335 spin_lock(&mm->page_table_lock);
2336 for (address = start; address < end; address += sz) {
2337 ptep = huge_pte_offset(mm, address);
2338 if (!ptep)
2339 continue;
2340
2341 if (huge_pmd_unshare(mm, &address, ptep))
2342 continue;
2343
2344 pte = huge_ptep_get(ptep);
2345 if (huge_pte_none(pte))
2346 continue;
2347
2348 /*
2349 * HWPoisoned hugepage is already unmapped and dropped reference
2350 */
2351 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2352 continue;
2353
2354 page = pte_page(pte);
2355 /*
2356 * If a reference page is supplied, it is because a specific
2357 * page is being unmapped, not a range. Ensure the page we
2358 * are about to unmap is the actual page of interest.
2359 */
2360 if (ref_page) {
2361 if (page != ref_page)
2362 continue;
2363
2364 /*
2365 * Mark the VMA as having unmapped its page so that
2366 * future faults in this VMA will fail rather than
2367 * looking like data was lost
2368 */
2369 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2370 }
2371
2372 pte = huge_ptep_get_and_clear(mm, address, ptep);
2373 if (pte_dirty(pte))
2374 set_page_dirty(page);
2375 list_add(&page->lru, &page_list);
2376
2377 /* Bail out after unmapping reference page if supplied */
2378 if (ref_page)
2379 break;
2380 }
2381 flush_tlb_range(vma, start, end);
2382 spin_unlock(&mm->page_table_lock);
2383 mmu_notifier_invalidate_range_end(mm, start, end);
2384 list_for_each_entry_safe(page, tmp, &page_list, lru) {
2385 page_remove_rmap(page);
2386 list_del(&page->lru);
2387 put_page(page);
2388 }
2389}
2390
2391void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2392 unsigned long end, struct page *ref_page)
2393{
2394 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2395 __unmap_hugepage_range(vma, start, end, ref_page);
2396 /*
2397 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2398 * test will fail on a vma being torn down, and not grab a page table
2399 * on its way out. We're lucky that the flag has such an appropriate
2400 * name, and can in fact be safely cleared here. We could clear it
2401 * before the __unmap_hugepage_range above, but all that's necessary
2402 * is to clear it before releasing the i_mmap_mutex below.
2403 *
2404 * This works because in the contexts this is called, the VMA is
2405 * going to be destroyed. It is not vunerable to madvise(DONTNEED)
2406 * because madvise is not supported on hugetlbfs. The same applies
2407 * for direct IO. unmap_hugepage_range() is only being called just
2408 * before free_pgtables() so clearing VM_MAYSHARE will not cause
2409 * surprises later.
2410 */
2411 vma->vm_flags &= ~VM_MAYSHARE;
2412 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2413}
2414
2415/*
2416 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2417 * mappping it owns the reserve page for. The intention is to unmap the page
2418 * from other VMAs and let the children be SIGKILLed if they are faulting the
2419 * same region.
2420 */
2421static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2422 struct page *page, unsigned long address)
2423{
2424 struct hstate *h = hstate_vma(vma);
2425 struct vm_area_struct *iter_vma;
2426 struct address_space *mapping;
2427 struct prio_tree_iter iter;
2428 pgoff_t pgoff;
2429
2430 /*
2431 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2432 * from page cache lookup which is in HPAGE_SIZE units.
2433 */
2434 address = address & huge_page_mask(h);
2435 pgoff = vma_hugecache_offset(h, vma, address);
2436 mapping = vma->vm_file->f_dentry->d_inode->i_mapping;
2437
2438 /*
2439 * Take the mapping lock for the duration of the table walk. As
2440 * this mapping should be shared between all the VMAs,
2441 * __unmap_hugepage_range() is called as the lock is already held
2442 */
2443 mutex_lock(&mapping->i_mmap_mutex);
2444 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2445 /* Do not unmap the current VMA */
2446 if (iter_vma == vma)
2447 continue;
2448
2449 /*
2450 * Unmap the page from other VMAs without their own reserves.
2451 * They get marked to be SIGKILLed if they fault in these
2452 * areas. This is because a future no-page fault on this VMA
2453 * could insert a zeroed page instead of the data existing
2454 * from the time of fork. This would look like data corruption
2455 */
2456 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2457 __unmap_hugepage_range(iter_vma,
2458 address, address + huge_page_size(h),
2459 page);
2460 }
2461 mutex_unlock(&mapping->i_mmap_mutex);
2462
2463 return 1;
2464}
2465
2466/*
2467 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2468 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2469 * cannot race with other handlers or page migration.
2470 * Keep the pte_same checks anyway to make transition from the mutex easier.
2471 */
2472static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2473 unsigned long address, pte_t *ptep, pte_t pte,
2474 struct page *pagecache_page)
2475{
2476 struct hstate *h = hstate_vma(vma);
2477 struct page *old_page, *new_page;
2478 int avoidcopy;
2479 int outside_reserve = 0;
2480
2481 old_page = pte_page(pte);
2482
2483retry_avoidcopy:
2484 /* If no-one else is actually using this page, avoid the copy
2485 * and just make the page writable */
2486 avoidcopy = (page_mapcount(old_page) == 1);
2487 if (avoidcopy) {
2488 if (PageAnon(old_page))
2489 page_move_anon_rmap(old_page, vma, address);
2490 set_huge_ptep_writable(vma, address, ptep);
2491 return 0;
2492 }
2493
2494 /*
2495 * If the process that created a MAP_PRIVATE mapping is about to
2496 * perform a COW due to a shared page count, attempt to satisfy
2497 * the allocation without using the existing reserves. The pagecache
2498 * page is used to determine if the reserve at this address was
2499 * consumed or not. If reserves were used, a partial faulted mapping
2500 * at the time of fork() could consume its reserves on COW instead
2501 * of the full address range.
2502 */
2503 if (!(vma->vm_flags & VM_MAYSHARE) &&
2504 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2505 old_page != pagecache_page)
2506 outside_reserve = 1;
2507
2508 page_cache_get(old_page);
2509
2510 /* Drop page_table_lock as buddy allocator may be called */
2511 spin_unlock(&mm->page_table_lock);
2512 new_page = alloc_huge_page(vma, address, outside_reserve);
2513
2514 if (IS_ERR(new_page)) {
2515 page_cache_release(old_page);
2516
2517 /*
2518 * If a process owning a MAP_PRIVATE mapping fails to COW,
2519 * it is due to references held by a child and an insufficient
2520 * huge page pool. To guarantee the original mappers
2521 * reliability, unmap the page from child processes. The child
2522 * may get SIGKILLed if it later faults.
2523 */
2524 if (outside_reserve) {
2525 BUG_ON(huge_pte_none(pte));
2526 if (unmap_ref_private(mm, vma, old_page, address)) {
2527 BUG_ON(huge_pte_none(pte));
2528 spin_lock(&mm->page_table_lock);
2529 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2530 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2531 goto retry_avoidcopy;
2532 /*
2533 * race occurs while re-acquiring page_table_lock, and
2534 * our job is done.
2535 */
2536 return 0;
2537 }
2538 WARN_ON_ONCE(1);
2539 }
2540
2541 /* Caller expects lock to be held */
2542 spin_lock(&mm->page_table_lock);
2543 return -PTR_ERR(new_page);
2544 }
2545
2546 /*
2547 * When the original hugepage is shared one, it does not have
2548 * anon_vma prepared.
2549 */
2550 if (unlikely(anon_vma_prepare(vma))) {
2551 page_cache_release(new_page);
2552 page_cache_release(old_page);
2553 /* Caller expects lock to be held */
2554 spin_lock(&mm->page_table_lock);
2555 return VM_FAULT_OOM;
2556 }
2557
2558 copy_user_huge_page(new_page, old_page, address, vma,
2559 pages_per_huge_page(h));
2560 __SetPageUptodate(new_page);
2561
2562 /*
2563 * Retake the page_table_lock to check for racing updates
2564 * before the page tables are altered
2565 */
2566 spin_lock(&mm->page_table_lock);
2567 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2568 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2569 /* Break COW */
2570 mmu_notifier_invalidate_range_start(mm,
2571 address & huge_page_mask(h),
2572 (address & huge_page_mask(h)) + huge_page_size(h));
2573 huge_ptep_clear_flush(vma, address, ptep);
2574 set_huge_pte_at(mm, address, ptep,
2575 make_huge_pte(vma, new_page, 1));
2576 page_remove_rmap(old_page);
2577 hugepage_add_new_anon_rmap(new_page, vma, address);
2578 /* Make the old page be freed below */
2579 new_page = old_page;
2580 mmu_notifier_invalidate_range_end(mm,
2581 address & huge_page_mask(h),
2582 (address & huge_page_mask(h)) + huge_page_size(h));
2583 }
2584 page_cache_release(new_page);
2585 page_cache_release(old_page);
2586 return 0;
2587}
2588
2589/* Return the pagecache page at a given address within a VMA */
2590static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2591 struct vm_area_struct *vma, unsigned long address)
2592{
2593 struct address_space *mapping;
2594 pgoff_t idx;
2595
2596 mapping = vma->vm_file->f_mapping;
2597 idx = vma_hugecache_offset(h, vma, address);
2598
2599 return find_lock_page(mapping, idx);
2600}
2601
2602/*
2603 * Return whether there is a pagecache page to back given address within VMA.
2604 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2605 */
2606static bool hugetlbfs_pagecache_present(struct hstate *h,
2607 struct vm_area_struct *vma, unsigned long address)
2608{
2609 struct address_space *mapping;
2610 pgoff_t idx;
2611 struct page *page;
2612
2613 mapping = vma->vm_file->f_mapping;
2614 idx = vma_hugecache_offset(h, vma, address);
2615
2616 page = find_get_page(mapping, idx);
2617 if (page)
2618 put_page(page);
2619 return page != NULL;
2620}
2621
2622static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2623 unsigned long address, pte_t *ptep, unsigned int flags)
2624{
2625 struct hstate *h = hstate_vma(vma);
2626 int ret = VM_FAULT_SIGBUS;
2627 int anon_rmap = 0;
2628 pgoff_t idx;
2629 unsigned long size;
2630 struct page *page;
2631 struct address_space *mapping;
2632 pte_t new_pte;
2633
2634 /*
2635 * Currently, we are forced to kill the process in the event the
2636 * original mapper has unmapped pages from the child due to a failed
2637 * COW. Warn that such a situation has occurred as it may not be obvious
2638 */
2639 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2640 printk(KERN_WARNING
2641 "PID %d killed due to inadequate hugepage pool\n",
2642 current->pid);
2643 return ret;
2644 }
2645
2646 mapping = vma->vm_file->f_mapping;
2647 idx = vma_hugecache_offset(h, vma, address);
2648
2649 /*
2650 * Use page lock to guard against racing truncation
2651 * before we get page_table_lock.
2652 */
2653retry:
2654 page = find_lock_page(mapping, idx);
2655 if (!page) {
2656 size = i_size_read(mapping->host) >> huge_page_shift(h);
2657 if (idx >= size)
2658 goto out;
2659 page = alloc_huge_page(vma, address, 0);
2660 if (IS_ERR(page)) {
2661 ret = -PTR_ERR(page);
2662 goto out;
2663 }
2664 clear_huge_page(page, address, pages_per_huge_page(h));
2665 __SetPageUptodate(page);
2666
2667 if (vma->vm_flags & VM_MAYSHARE) {
2668 int err;
2669 struct inode *inode = mapping->host;
2670
2671 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2672 if (err) {
2673 put_page(page);
2674 if (err == -EEXIST)
2675 goto retry;
2676 goto out;
2677 }
2678
2679 spin_lock(&inode->i_lock);
2680 inode->i_blocks += blocks_per_huge_page(h);
2681 spin_unlock(&inode->i_lock);
2682 } else {
2683 lock_page(page);
2684 if (unlikely(anon_vma_prepare(vma))) {
2685 ret = VM_FAULT_OOM;
2686 goto backout_unlocked;
2687 }
2688 anon_rmap = 1;
2689 }
2690 } else {
2691 /*
2692 * If memory error occurs between mmap() and fault, some process
2693 * don't have hwpoisoned swap entry for errored virtual address.
2694 * So we need to block hugepage fault by PG_hwpoison bit check.
2695 */
2696 if (unlikely(PageHWPoison(page))) {
2697 ret = VM_FAULT_HWPOISON |
2698 VM_FAULT_SET_HINDEX(h - hstates);
2699 goto backout_unlocked;
2700 }
2701 }
2702
2703 /*
2704 * If we are going to COW a private mapping later, we examine the
2705 * pending reservations for this page now. This will ensure that
2706 * any allocations necessary to record that reservation occur outside
2707 * the spinlock.
2708 */
2709 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2710 if (vma_needs_reservation(h, vma, address) < 0) {
2711 ret = VM_FAULT_OOM;
2712 goto backout_unlocked;
2713 }
2714
2715 spin_lock(&mm->page_table_lock);
2716 size = i_size_read(mapping->host) >> huge_page_shift(h);
2717 if (idx >= size)
2718 goto backout;
2719
2720 ret = 0;
2721 if (!huge_pte_none(huge_ptep_get(ptep)))
2722 goto backout;
2723
2724 if (anon_rmap)
2725 hugepage_add_new_anon_rmap(page, vma, address);
2726 else
2727 page_dup_rmap(page);
2728 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2729 && (vma->vm_flags & VM_SHARED)));
2730 set_huge_pte_at(mm, address, ptep, new_pte);
2731
2732 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2733 /* Optimization, do the COW without a second fault */
2734 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2735 }
2736
2737 spin_unlock(&mm->page_table_lock);
2738 unlock_page(page);
2739out:
2740 return ret;
2741
2742backout:
2743 spin_unlock(&mm->page_table_lock);
2744backout_unlocked:
2745 unlock_page(page);
2746 put_page(page);
2747 goto out;
2748}
2749
2750int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2751 unsigned long address, unsigned int flags)
2752{
2753 pte_t *ptep;
2754 pte_t entry;
2755 int ret;
2756 struct page *page = NULL;
2757 struct page *pagecache_page = NULL;
2758 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2759 struct hstate *h = hstate_vma(vma);
2760
2761 address &= huge_page_mask(h);
2762
2763 ptep = huge_pte_offset(mm, address);
2764 if (ptep) {
2765 entry = huge_ptep_get(ptep);
2766 if (unlikely(is_hugetlb_entry_migration(entry))) {
2767 migration_entry_wait(mm, (pmd_t *)ptep, address);
2768 return 0;
2769 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2770 return VM_FAULT_HWPOISON_LARGE |
2771 VM_FAULT_SET_HINDEX(h - hstates);
2772 }
2773
2774 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2775 if (!ptep)
2776 return VM_FAULT_OOM;
2777
2778 /*
2779 * Serialize hugepage allocation and instantiation, so that we don't
2780 * get spurious allocation failures if two CPUs race to instantiate
2781 * the same page in the page cache.
2782 */
2783 mutex_lock(&hugetlb_instantiation_mutex);
2784 entry = huge_ptep_get(ptep);
2785 if (huge_pte_none(entry)) {
2786 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2787 goto out_mutex;
2788 }
2789
2790 ret = 0;
2791
2792 /*
2793 * If we are going to COW the mapping later, we examine the pending
2794 * reservations for this page now. This will ensure that any
2795 * allocations necessary to record that reservation occur outside the
2796 * spinlock. For private mappings, we also lookup the pagecache
2797 * page now as it is used to determine if a reservation has been
2798 * consumed.
2799 */
2800 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2801 if (vma_needs_reservation(h, vma, address) < 0) {
2802 ret = VM_FAULT_OOM;
2803 goto out_mutex;
2804 }
2805
2806 if (!(vma->vm_flags & VM_MAYSHARE))
2807 pagecache_page = hugetlbfs_pagecache_page(h,
2808 vma, address);
2809 }
2810
2811 /*
2812 * hugetlb_cow() requires page locks of pte_page(entry) and
2813 * pagecache_page, so here we need take the former one
2814 * when page != pagecache_page or !pagecache_page.
2815 * Note that locking order is always pagecache_page -> page,
2816 * so no worry about deadlock.
2817 */
2818 page = pte_page(entry);
2819 get_page(page);
2820 if (page != pagecache_page)
2821 lock_page(page);
2822
2823 spin_lock(&mm->page_table_lock);
2824 /* Check for a racing update before calling hugetlb_cow */
2825 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2826 goto out_page_table_lock;
2827
2828
2829 if (flags & FAULT_FLAG_WRITE) {
2830 if (!pte_write(entry)) {
2831 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2832 pagecache_page);
2833 goto out_page_table_lock;
2834 }
2835 entry = pte_mkdirty(entry);
2836 }
2837 entry = pte_mkyoung(entry);
2838 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2839 flags & FAULT_FLAG_WRITE))
2840 update_mmu_cache(vma, address, ptep);
2841
2842out_page_table_lock:
2843 spin_unlock(&mm->page_table_lock);
2844
2845 if (pagecache_page) {
2846 unlock_page(pagecache_page);
2847 put_page(pagecache_page);
2848 }
2849 if (page != pagecache_page)
2850 unlock_page(page);
2851 put_page(page);
2852
2853out_mutex:
2854 mutex_unlock(&hugetlb_instantiation_mutex);
2855
2856 return ret;
2857}
2858
2859/* Can be overriden by architectures */
2860__attribute__((weak)) struct page *
2861follow_huge_pud(struct mm_struct *mm, unsigned long address,
2862 pud_t *pud, int write)
2863{
2864 BUG();
2865 return NULL;
2866}
2867
2868int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2869 struct page **pages, struct vm_area_struct **vmas,
2870 unsigned long *position, int *length, int i,
2871 unsigned int flags)
2872{
2873 unsigned long pfn_offset;
2874 unsigned long vaddr = *position;
2875 int remainder = *length;
2876 struct hstate *h = hstate_vma(vma);
2877
2878 spin_lock(&mm->page_table_lock);
2879 while (vaddr < vma->vm_end && remainder) {
2880 pte_t *pte;
2881 int absent;
2882 struct page *page;
2883
2884 /*
2885 * Some archs (sparc64, sh*) have multiple pte_ts to
2886 * each hugepage. We have to make sure we get the
2887 * first, for the page indexing below to work.
2888 */
2889 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2890 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2891
2892 /*
2893 * When coredumping, it suits get_dump_page if we just return
2894 * an error where there's an empty slot with no huge pagecache
2895 * to back it. This way, we avoid allocating a hugepage, and
2896 * the sparse dumpfile avoids allocating disk blocks, but its
2897 * huge holes still show up with zeroes where they need to be.
2898 */
2899 if (absent && (flags & FOLL_DUMP) &&
2900 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2901 remainder = 0;
2902 break;
2903 }
2904
2905 if (absent ||
2906 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2907 int ret;
2908
2909 spin_unlock(&mm->page_table_lock);
2910 ret = hugetlb_fault(mm, vma, vaddr,
2911 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2912 spin_lock(&mm->page_table_lock);
2913 if (!(ret & VM_FAULT_ERROR))
2914 continue;
2915
2916 remainder = 0;
2917 break;
2918 }
2919
2920 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2921 page = pte_page(huge_ptep_get(pte));
2922same_page:
2923 if (pages) {
2924 pages[i] = mem_map_offset(page, pfn_offset);
2925 get_page(pages[i]);
2926 }
2927
2928 if (vmas)
2929 vmas[i] = vma;
2930
2931 vaddr += PAGE_SIZE;
2932 ++pfn_offset;
2933 --remainder;
2934 ++i;
2935 if (vaddr < vma->vm_end && remainder &&
2936 pfn_offset < pages_per_huge_page(h)) {
2937 /*
2938 * We use pfn_offset to avoid touching the pageframes
2939 * of this compound page.
2940 */
2941 goto same_page;
2942 }
2943 }
2944 spin_unlock(&mm->page_table_lock);
2945 *length = remainder;
2946 *position = vaddr;
2947
2948 return i ? i : -EFAULT;
2949}
2950
2951void hugetlb_change_protection(struct vm_area_struct *vma,
2952 unsigned long address, unsigned long end, pgprot_t newprot)
2953{
2954 struct mm_struct *mm = vma->vm_mm;
2955 unsigned long start = address;
2956 pte_t *ptep;
2957 pte_t pte;
2958 struct hstate *h = hstate_vma(vma);
2959
2960 BUG_ON(address >= end);
2961 flush_cache_range(vma, address, end);
2962
2963 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2964 spin_lock(&mm->page_table_lock);
2965 for (; address < end; address += huge_page_size(h)) {
2966 ptep = huge_pte_offset(mm, address);
2967 if (!ptep)
2968 continue;
2969 if (huge_pmd_unshare(mm, &address, ptep))
2970 continue;
2971 if (!huge_pte_none(huge_ptep_get(ptep))) {
2972 pte = huge_ptep_get_and_clear(mm, address, ptep);
2973 pte = pte_mkhuge(pte_modify(pte, newprot));
2974 set_huge_pte_at(mm, address, ptep, pte);
2975 }
2976 }
2977 spin_unlock(&mm->page_table_lock);
2978 /*
2979 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
2980 * may have cleared our pud entry and done put_page on the page table:
2981 * once we release i_mmap_mutex, another task can do the final put_page
2982 * and that page table be reused and filled with junk.
2983 */
2984 flush_tlb_range(vma, start, end);
2985 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2986}
2987
2988int hugetlb_reserve_pages(struct inode *inode,
2989 long from, long to,
2990 struct vm_area_struct *vma,
2991 vm_flags_t vm_flags)
2992{
2993 long ret, chg;
2994 struct hstate *h = hstate_inode(inode);
2995 struct hugepage_subpool *spool = subpool_inode(inode);
2996
2997 /*
2998 * Only apply hugepage reservation if asked. At fault time, an
2999 * attempt will be made for VM_NORESERVE to allocate a page
3000 * without using reserves
3001 */
3002 if (vm_flags & VM_NORESERVE)
3003 return 0;
3004
3005 /*
3006 * Shared mappings base their reservation on the number of pages that
3007 * are already allocated on behalf of the file. Private mappings need
3008 * to reserve the full area even if read-only as mprotect() may be
3009 * called to make the mapping read-write. Assume !vma is a shm mapping
3010 */
3011 if (!vma || vma->vm_flags & VM_MAYSHARE)
3012 chg = region_chg(&inode->i_mapping->private_list, from, to);
3013 else {
3014 struct resv_map *resv_map = resv_map_alloc();
3015 if (!resv_map)
3016 return -ENOMEM;
3017
3018 chg = to - from;
3019
3020 set_vma_resv_map(vma, resv_map);
3021 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3022 }
3023
3024 if (chg < 0) {
3025 ret = chg;
3026 goto out_err;
3027 }
3028
3029 /* There must be enough pages in the subpool for the mapping */
3030 if (hugepage_subpool_get_pages(spool, chg)) {
3031 ret = -ENOSPC;
3032 goto out_err;
3033 }
3034
3035 /*
3036 * Check enough hugepages are available for the reservation.
3037 * Hand the pages back to the subpool if there are not
3038 */
3039 ret = hugetlb_acct_memory(h, chg);
3040 if (ret < 0) {
3041 hugepage_subpool_put_pages(spool, chg);
3042 goto out_err;
3043 }
3044
3045 /*
3046 * Account for the reservations made. Shared mappings record regions
3047 * that have reservations as they are shared by multiple VMAs.
3048 * When the last VMA disappears, the region map says how much
3049 * the reservation was and the page cache tells how much of
3050 * the reservation was consumed. Private mappings are per-VMA and
3051 * only the consumed reservations are tracked. When the VMA
3052 * disappears, the original reservation is the VMA size and the
3053 * consumed reservations are stored in the map. Hence, nothing
3054 * else has to be done for private mappings here
3055 */
3056 if (!vma || vma->vm_flags & VM_MAYSHARE)
3057 region_add(&inode->i_mapping->private_list, from, to);
3058 return 0;
3059out_err:
3060 if (vma)
3061 resv_map_put(vma);
3062 return ret;
3063}
3064
3065void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3066{
3067 struct hstate *h = hstate_inode(inode);
3068 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3069 struct hugepage_subpool *spool = subpool_inode(inode);
3070
3071 spin_lock(&inode->i_lock);
3072 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3073 spin_unlock(&inode->i_lock);
3074
3075 hugepage_subpool_put_pages(spool, (chg - freed));
3076 hugetlb_acct_memory(h, -(chg - freed));
3077}
3078
3079#ifdef CONFIG_MEMORY_FAILURE
3080
3081/* Should be called in hugetlb_lock */
3082static int is_hugepage_on_freelist(struct page *hpage)
3083{
3084 struct page *page;
3085 struct page *tmp;
3086 struct hstate *h = page_hstate(hpage);
3087 int nid = page_to_nid(hpage);
3088
3089 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3090 if (page == hpage)
3091 return 1;
3092 return 0;
3093}
3094
3095/*
3096 * This function is called from memory failure code.
3097 * Assume the caller holds page lock of the head page.
3098 */
3099int dequeue_hwpoisoned_huge_page(struct page *hpage)
3100{
3101 struct hstate *h = page_hstate(hpage);
3102 int nid = page_to_nid(hpage);
3103 int ret = -EBUSY;
3104
3105 spin_lock(&hugetlb_lock);
3106 if (is_hugepage_on_freelist(hpage)) {
3107 list_del(&hpage->lru);
3108 set_page_refcounted(hpage);
3109 h->free_huge_pages--;
3110 h->free_huge_pages_node[nid]--;
3111 ret = 0;
3112 }
3113 spin_unlock(&hugetlb_lock);
3114 return ret;
3115}
3116#endif