cbf32d2824fd473d309c157c514dfff38783be9b
[linux-2.6-microblaze.git] / mm / hugetlb.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Generic hugetlb support.
4  * (C) Nadia Yvette Chambers, April 2004
5  */
6 #include <linux/list.h>
7 #include <linux/init.h>
8 #include <linux/mm.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/sched/mm.h>
23 #include <linux/mmdebug.h>
24 #include <linux/sched/signal.h>
25 #include <linux/rmap.h>
26 #include <linux/string_helpers.h>
27 #include <linux/swap.h>
28 #include <linux/swapops.h>
29 #include <linux/jhash.h>
30 #include <linux/numa.h>
31 #include <linux/llist.h>
32 #include <linux/cma.h>
33
34 #include <asm/page.h>
35 #include <asm/pgalloc.h>
36 #include <asm/tlb.h>
37
38 #include <linux/io.h>
39 #include <linux/hugetlb.h>
40 #include <linux/hugetlb_cgroup.h>
41 #include <linux/node.h>
42 #include <linux/userfaultfd_k.h>
43 #include <linux/page_owner.h>
44 #include "internal.h"
45
46 int hugetlb_max_hstate __read_mostly;
47 unsigned int default_hstate_idx;
48 struct hstate hstates[HUGE_MAX_HSTATE];
49
50 #ifdef CONFIG_CMA
51 static struct cma *hugetlb_cma[MAX_NUMNODES];
52 #endif
53 static unsigned long hugetlb_cma_size __initdata;
54
55 /*
56  * Minimum page order among possible hugepage sizes, set to a proper value
57  * at boot time.
58  */
59 static unsigned int minimum_order __read_mostly = UINT_MAX;
60
61 __initdata LIST_HEAD(huge_boot_pages);
62
63 /* for command line parsing */
64 static struct hstate * __initdata parsed_hstate;
65 static unsigned long __initdata default_hstate_max_huge_pages;
66 static bool __initdata parsed_valid_hugepagesz = true;
67 static bool __initdata parsed_default_hugepagesz;
68
69 /*
70  * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
71  * free_huge_pages, and surplus_huge_pages.
72  */
73 DEFINE_SPINLOCK(hugetlb_lock);
74
75 /*
76  * Serializes faults on the same logical page.  This is used to
77  * prevent spurious OOMs when the hugepage pool is fully utilized.
78  */
79 static int num_fault_mutexes;
80 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
81
82 /* Forward declaration */
83 static int hugetlb_acct_memory(struct hstate *h, long delta);
84
85 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
86 {
87         bool free = (spool->count == 0) && (spool->used_hpages == 0);
88
89         spin_unlock(&spool->lock);
90
91         /* If no pages are used, and no other handles to the subpool
92          * remain, give up any reservations based on minimum size and
93          * free the subpool */
94         if (free) {
95                 if (spool->min_hpages != -1)
96                         hugetlb_acct_memory(spool->hstate,
97                                                 -spool->min_hpages);
98                 kfree(spool);
99         }
100 }
101
102 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
103                                                 long min_hpages)
104 {
105         struct hugepage_subpool *spool;
106
107         spool = kzalloc(sizeof(*spool), GFP_KERNEL);
108         if (!spool)
109                 return NULL;
110
111         spin_lock_init(&spool->lock);
112         spool->count = 1;
113         spool->max_hpages = max_hpages;
114         spool->hstate = h;
115         spool->min_hpages = min_hpages;
116
117         if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
118                 kfree(spool);
119                 return NULL;
120         }
121         spool->rsv_hpages = min_hpages;
122
123         return spool;
124 }
125
126 void hugepage_put_subpool(struct hugepage_subpool *spool)
127 {
128         spin_lock(&spool->lock);
129         BUG_ON(!spool->count);
130         spool->count--;
131         unlock_or_release_subpool(spool);
132 }
133
134 /*
135  * Subpool accounting for allocating and reserving pages.
136  * Return -ENOMEM if there are not enough resources to satisfy the
137  * request.  Otherwise, return the number of pages by which the
138  * global pools must be adjusted (upward).  The returned value may
139  * only be different than the passed value (delta) in the case where
140  * a subpool minimum size must be maintained.
141  */
142 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
143                                       long delta)
144 {
145         long ret = delta;
146
147         if (!spool)
148                 return ret;
149
150         spin_lock(&spool->lock);
151
152         if (spool->max_hpages != -1) {          /* maximum size accounting */
153                 if ((spool->used_hpages + delta) <= spool->max_hpages)
154                         spool->used_hpages += delta;
155                 else {
156                         ret = -ENOMEM;
157                         goto unlock_ret;
158                 }
159         }
160
161         /* minimum size accounting */
162         if (spool->min_hpages != -1 && spool->rsv_hpages) {
163                 if (delta > spool->rsv_hpages) {
164                         /*
165                          * Asking for more reserves than those already taken on
166                          * behalf of subpool.  Return difference.
167                          */
168                         ret = delta - spool->rsv_hpages;
169                         spool->rsv_hpages = 0;
170                 } else {
171                         ret = 0;        /* reserves already accounted for */
172                         spool->rsv_hpages -= delta;
173                 }
174         }
175
176 unlock_ret:
177         spin_unlock(&spool->lock);
178         return ret;
179 }
180
181 /*
182  * Subpool accounting for freeing and unreserving pages.
183  * Return the number of global page reservations that must be dropped.
184  * The return value may only be different than the passed value (delta)
185  * in the case where a subpool minimum size must be maintained.
186  */
187 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
188                                        long delta)
189 {
190         long ret = delta;
191
192         if (!spool)
193                 return delta;
194
195         spin_lock(&spool->lock);
196
197         if (spool->max_hpages != -1)            /* maximum size accounting */
198                 spool->used_hpages -= delta;
199
200          /* minimum size accounting */
201         if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
202                 if (spool->rsv_hpages + delta <= spool->min_hpages)
203                         ret = 0;
204                 else
205                         ret = spool->rsv_hpages + delta - spool->min_hpages;
206
207                 spool->rsv_hpages += delta;
208                 if (spool->rsv_hpages > spool->min_hpages)
209                         spool->rsv_hpages = spool->min_hpages;
210         }
211
212         /*
213          * If hugetlbfs_put_super couldn't free spool due to an outstanding
214          * quota reference, free it now.
215          */
216         unlock_or_release_subpool(spool);
217
218         return ret;
219 }
220
221 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
222 {
223         return HUGETLBFS_SB(inode->i_sb)->spool;
224 }
225
226 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
227 {
228         return subpool_inode(file_inode(vma->vm_file));
229 }
230
231 /* Helper that removes a struct file_region from the resv_map cache and returns
232  * it for use.
233  */
234 static struct file_region *
235 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
236 {
237         struct file_region *nrg = NULL;
238
239         VM_BUG_ON(resv->region_cache_count <= 0);
240
241         resv->region_cache_count--;
242         nrg = list_first_entry(&resv->region_cache, struct file_region, link);
243         list_del(&nrg->link);
244
245         nrg->from = from;
246         nrg->to = to;
247
248         return nrg;
249 }
250
251 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
252                                               struct file_region *rg)
253 {
254 #ifdef CONFIG_CGROUP_HUGETLB
255         nrg->reservation_counter = rg->reservation_counter;
256         nrg->css = rg->css;
257         if (rg->css)
258                 css_get(rg->css);
259 #endif
260 }
261
262 /* Helper that records hugetlb_cgroup uncharge info. */
263 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
264                                                 struct hstate *h,
265                                                 struct resv_map *resv,
266                                                 struct file_region *nrg)
267 {
268 #ifdef CONFIG_CGROUP_HUGETLB
269         if (h_cg) {
270                 nrg->reservation_counter =
271                         &h_cg->rsvd_hugepage[hstate_index(h)];
272                 nrg->css = &h_cg->css;
273                 if (!resv->pages_per_hpage)
274                         resv->pages_per_hpage = pages_per_huge_page(h);
275                 /* pages_per_hpage should be the same for all entries in
276                  * a resv_map.
277                  */
278                 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
279         } else {
280                 nrg->reservation_counter = NULL;
281                 nrg->css = NULL;
282         }
283 #endif
284 }
285
286 static bool has_same_uncharge_info(struct file_region *rg,
287                                    struct file_region *org)
288 {
289 #ifdef CONFIG_CGROUP_HUGETLB
290         return rg && org &&
291                rg->reservation_counter == org->reservation_counter &&
292                rg->css == org->css;
293
294 #else
295         return true;
296 #endif
297 }
298
299 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
300 {
301         struct file_region *nrg = NULL, *prg = NULL;
302
303         prg = list_prev_entry(rg, link);
304         if (&prg->link != &resv->regions && prg->to == rg->from &&
305             has_same_uncharge_info(prg, rg)) {
306                 prg->to = rg->to;
307
308                 list_del(&rg->link);
309                 kfree(rg);
310
311                 rg = prg;
312         }
313
314         nrg = list_next_entry(rg, link);
315         if (&nrg->link != &resv->regions && nrg->from == rg->to &&
316             has_same_uncharge_info(nrg, rg)) {
317                 nrg->from = rg->from;
318
319                 list_del(&rg->link);
320                 kfree(rg);
321         }
322 }
323
324 /*
325  * Must be called with resv->lock held.
326  *
327  * Calling this with regions_needed != NULL will count the number of pages
328  * to be added but will not modify the linked list. And regions_needed will
329  * indicate the number of file_regions needed in the cache to carry out to add
330  * the regions for this range.
331  */
332 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
333                                      struct hugetlb_cgroup *h_cg,
334                                      struct hstate *h, long *regions_needed)
335 {
336         long add = 0;
337         struct list_head *head = &resv->regions;
338         long last_accounted_offset = f;
339         struct file_region *rg = NULL, *trg = NULL, *nrg = NULL;
340
341         if (regions_needed)
342                 *regions_needed = 0;
343
344         /* In this loop, we essentially handle an entry for the range
345          * [last_accounted_offset, rg->from), at every iteration, with some
346          * bounds checking.
347          */
348         list_for_each_entry_safe(rg, trg, head, link) {
349                 /* Skip irrelevant regions that start before our range. */
350                 if (rg->from < f) {
351                         /* If this region ends after the last accounted offset,
352                          * then we need to update last_accounted_offset.
353                          */
354                         if (rg->to > last_accounted_offset)
355                                 last_accounted_offset = rg->to;
356                         continue;
357                 }
358
359                 /* When we find a region that starts beyond our range, we've
360                  * finished.
361                  */
362                 if (rg->from > t)
363                         break;
364
365                 /* Add an entry for last_accounted_offset -> rg->from, and
366                  * update last_accounted_offset.
367                  */
368                 if (rg->from > last_accounted_offset) {
369                         add += rg->from - last_accounted_offset;
370                         if (!regions_needed) {
371                                 nrg = get_file_region_entry_from_cache(
372                                         resv, last_accounted_offset, rg->from);
373                                 record_hugetlb_cgroup_uncharge_info(h_cg, h,
374                                                                     resv, nrg);
375                                 list_add(&nrg->link, rg->link.prev);
376                                 coalesce_file_region(resv, nrg);
377                         } else
378                                 *regions_needed += 1;
379                 }
380
381                 last_accounted_offset = rg->to;
382         }
383
384         /* Handle the case where our range extends beyond
385          * last_accounted_offset.
386          */
387         if (last_accounted_offset < t) {
388                 add += t - last_accounted_offset;
389                 if (!regions_needed) {
390                         nrg = get_file_region_entry_from_cache(
391                                 resv, last_accounted_offset, t);
392                         record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg);
393                         list_add(&nrg->link, rg->link.prev);
394                         coalesce_file_region(resv, nrg);
395                 } else
396                         *regions_needed += 1;
397         }
398
399         VM_BUG_ON(add < 0);
400         return add;
401 }
402
403 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
404  */
405 static int allocate_file_region_entries(struct resv_map *resv,
406                                         int regions_needed)
407         __must_hold(&resv->lock)
408 {
409         struct list_head allocated_regions;
410         int to_allocate = 0, i = 0;
411         struct file_region *trg = NULL, *rg = NULL;
412
413         VM_BUG_ON(regions_needed < 0);
414
415         INIT_LIST_HEAD(&allocated_regions);
416
417         /*
418          * Check for sufficient descriptors in the cache to accommodate
419          * the number of in progress add operations plus regions_needed.
420          *
421          * This is a while loop because when we drop the lock, some other call
422          * to region_add or region_del may have consumed some region_entries,
423          * so we keep looping here until we finally have enough entries for
424          * (adds_in_progress + regions_needed).
425          */
426         while (resv->region_cache_count <
427                (resv->adds_in_progress + regions_needed)) {
428                 to_allocate = resv->adds_in_progress + regions_needed -
429                               resv->region_cache_count;
430
431                 /* At this point, we should have enough entries in the cache
432                  * for all the existings adds_in_progress. We should only be
433                  * needing to allocate for regions_needed.
434                  */
435                 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
436
437                 spin_unlock(&resv->lock);
438                 for (i = 0; i < to_allocate; i++) {
439                         trg = kmalloc(sizeof(*trg), GFP_KERNEL);
440                         if (!trg)
441                                 goto out_of_memory;
442                         list_add(&trg->link, &allocated_regions);
443                 }
444
445                 spin_lock(&resv->lock);
446
447                 list_splice(&allocated_regions, &resv->region_cache);
448                 resv->region_cache_count += to_allocate;
449         }
450
451         return 0;
452
453 out_of_memory:
454         list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
455                 list_del(&rg->link);
456                 kfree(rg);
457         }
458         return -ENOMEM;
459 }
460
461 /*
462  * Add the huge page range represented by [f, t) to the reserve
463  * map.  Regions will be taken from the cache to fill in this range.
464  * Sufficient regions should exist in the cache due to the previous
465  * call to region_chg with the same range, but in some cases the cache will not
466  * have sufficient entries due to races with other code doing region_add or
467  * region_del.  The extra needed entries will be allocated.
468  *
469  * regions_needed is the out value provided by a previous call to region_chg.
470  *
471  * Return the number of new huge pages added to the map.  This number is greater
472  * than or equal to zero.  If file_region entries needed to be allocated for
473  * this operation and we were not able to allocate, it returns -ENOMEM.
474  * region_add of regions of length 1 never allocate file_regions and cannot
475  * fail; region_chg will always allocate at least 1 entry and a region_add for
476  * 1 page will only require at most 1 entry.
477  */
478 static long region_add(struct resv_map *resv, long f, long t,
479                        long in_regions_needed, struct hstate *h,
480                        struct hugetlb_cgroup *h_cg)
481 {
482         long add = 0, actual_regions_needed = 0;
483
484         spin_lock(&resv->lock);
485 retry:
486
487         /* Count how many regions are actually needed to execute this add. */
488         add_reservation_in_range(resv, f, t, NULL, NULL,
489                                  &actual_regions_needed);
490
491         /*
492          * Check for sufficient descriptors in the cache to accommodate
493          * this add operation. Note that actual_regions_needed may be greater
494          * than in_regions_needed, as the resv_map may have been modified since
495          * the region_chg call. In this case, we need to make sure that we
496          * allocate extra entries, such that we have enough for all the
497          * existing adds_in_progress, plus the excess needed for this
498          * operation.
499          */
500         if (actual_regions_needed > in_regions_needed &&
501             resv->region_cache_count <
502                     resv->adds_in_progress +
503                             (actual_regions_needed - in_regions_needed)) {
504                 /* region_add operation of range 1 should never need to
505                  * allocate file_region entries.
506                  */
507                 VM_BUG_ON(t - f <= 1);
508
509                 if (allocate_file_region_entries(
510                             resv, actual_regions_needed - in_regions_needed)) {
511                         return -ENOMEM;
512                 }
513
514                 goto retry;
515         }
516
517         add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
518
519         resv->adds_in_progress -= in_regions_needed;
520
521         spin_unlock(&resv->lock);
522         VM_BUG_ON(add < 0);
523         return add;
524 }
525
526 /*
527  * Examine the existing reserve map and determine how many
528  * huge pages in the specified range [f, t) are NOT currently
529  * represented.  This routine is called before a subsequent
530  * call to region_add that will actually modify the reserve
531  * map to add the specified range [f, t).  region_chg does
532  * not change the number of huge pages represented by the
533  * map.  A number of new file_region structures is added to the cache as a
534  * placeholder, for the subsequent region_add call to use. At least 1
535  * file_region structure is added.
536  *
537  * out_regions_needed is the number of regions added to the
538  * resv->adds_in_progress.  This value needs to be provided to a follow up call
539  * to region_add or region_abort for proper accounting.
540  *
541  * Returns the number of huge pages that need to be added to the existing
542  * reservation map for the range [f, t).  This number is greater or equal to
543  * zero.  -ENOMEM is returned if a new file_region structure or cache entry
544  * is needed and can not be allocated.
545  */
546 static long region_chg(struct resv_map *resv, long f, long t,
547                        long *out_regions_needed)
548 {
549         long chg = 0;
550
551         spin_lock(&resv->lock);
552
553         /* Count how many hugepages in this range are NOT represented. */
554         chg = add_reservation_in_range(resv, f, t, NULL, NULL,
555                                        out_regions_needed);
556
557         if (*out_regions_needed == 0)
558                 *out_regions_needed = 1;
559
560         if (allocate_file_region_entries(resv, *out_regions_needed))
561                 return -ENOMEM;
562
563         resv->adds_in_progress += *out_regions_needed;
564
565         spin_unlock(&resv->lock);
566         return chg;
567 }
568
569 /*
570  * Abort the in progress add operation.  The adds_in_progress field
571  * of the resv_map keeps track of the operations in progress between
572  * calls to region_chg and region_add.  Operations are sometimes
573  * aborted after the call to region_chg.  In such cases, region_abort
574  * is called to decrement the adds_in_progress counter. regions_needed
575  * is the value returned by the region_chg call, it is used to decrement
576  * the adds_in_progress counter.
577  *
578  * NOTE: The range arguments [f, t) are not needed or used in this
579  * routine.  They are kept to make reading the calling code easier as
580  * arguments will match the associated region_chg call.
581  */
582 static void region_abort(struct resv_map *resv, long f, long t,
583                          long regions_needed)
584 {
585         spin_lock(&resv->lock);
586         VM_BUG_ON(!resv->region_cache_count);
587         resv->adds_in_progress -= regions_needed;
588         spin_unlock(&resv->lock);
589 }
590
591 /*
592  * Delete the specified range [f, t) from the reserve map.  If the
593  * t parameter is LONG_MAX, this indicates that ALL regions after f
594  * should be deleted.  Locate the regions which intersect [f, t)
595  * and either trim, delete or split the existing regions.
596  *
597  * Returns the number of huge pages deleted from the reserve map.
598  * In the normal case, the return value is zero or more.  In the
599  * case where a region must be split, a new region descriptor must
600  * be allocated.  If the allocation fails, -ENOMEM will be returned.
601  * NOTE: If the parameter t == LONG_MAX, then we will never split
602  * a region and possibly return -ENOMEM.  Callers specifying
603  * t == LONG_MAX do not need to check for -ENOMEM error.
604  */
605 static long region_del(struct resv_map *resv, long f, long t)
606 {
607         struct list_head *head = &resv->regions;
608         struct file_region *rg, *trg;
609         struct file_region *nrg = NULL;
610         long del = 0;
611
612 retry:
613         spin_lock(&resv->lock);
614         list_for_each_entry_safe(rg, trg, head, link) {
615                 /*
616                  * Skip regions before the range to be deleted.  file_region
617                  * ranges are normally of the form [from, to).  However, there
618                  * may be a "placeholder" entry in the map which is of the form
619                  * (from, to) with from == to.  Check for placeholder entries
620                  * at the beginning of the range to be deleted.
621                  */
622                 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
623                         continue;
624
625                 if (rg->from >= t)
626                         break;
627
628                 if (f > rg->from && t < rg->to) { /* Must split region */
629                         /*
630                          * Check for an entry in the cache before dropping
631                          * lock and attempting allocation.
632                          */
633                         if (!nrg &&
634                             resv->region_cache_count > resv->adds_in_progress) {
635                                 nrg = list_first_entry(&resv->region_cache,
636                                                         struct file_region,
637                                                         link);
638                                 list_del(&nrg->link);
639                                 resv->region_cache_count--;
640                         }
641
642                         if (!nrg) {
643                                 spin_unlock(&resv->lock);
644                                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
645                                 if (!nrg)
646                                         return -ENOMEM;
647                                 goto retry;
648                         }
649
650                         del += t - f;
651                         hugetlb_cgroup_uncharge_file_region(
652                                 resv, rg, t - f);
653
654                         /* New entry for end of split region */
655                         nrg->from = t;
656                         nrg->to = rg->to;
657
658                         copy_hugetlb_cgroup_uncharge_info(nrg, rg);
659
660                         INIT_LIST_HEAD(&nrg->link);
661
662                         /* Original entry is trimmed */
663                         rg->to = f;
664
665                         list_add(&nrg->link, &rg->link);
666                         nrg = NULL;
667                         break;
668                 }
669
670                 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
671                         del += rg->to - rg->from;
672                         hugetlb_cgroup_uncharge_file_region(resv, rg,
673                                                             rg->to - rg->from);
674                         list_del(&rg->link);
675                         kfree(rg);
676                         continue;
677                 }
678
679                 if (f <= rg->from) {    /* Trim beginning of region */
680                         hugetlb_cgroup_uncharge_file_region(resv, rg,
681                                                             t - rg->from);
682
683                         del += t - rg->from;
684                         rg->from = t;
685                 } else {                /* Trim end of region */
686                         hugetlb_cgroup_uncharge_file_region(resv, rg,
687                                                             rg->to - f);
688
689                         del += rg->to - f;
690                         rg->to = f;
691                 }
692         }
693
694         spin_unlock(&resv->lock);
695         kfree(nrg);
696         return del;
697 }
698
699 /*
700  * A rare out of memory error was encountered which prevented removal of
701  * the reserve map region for a page.  The huge page itself was free'ed
702  * and removed from the page cache.  This routine will adjust the subpool
703  * usage count, and the global reserve count if needed.  By incrementing
704  * these counts, the reserve map entry which could not be deleted will
705  * appear as a "reserved" entry instead of simply dangling with incorrect
706  * counts.
707  */
708 void hugetlb_fix_reserve_counts(struct inode *inode)
709 {
710         struct hugepage_subpool *spool = subpool_inode(inode);
711         long rsv_adjust;
712
713         rsv_adjust = hugepage_subpool_get_pages(spool, 1);
714         if (rsv_adjust) {
715                 struct hstate *h = hstate_inode(inode);
716
717                 hugetlb_acct_memory(h, 1);
718         }
719 }
720
721 /*
722  * Count and return the number of huge pages in the reserve map
723  * that intersect with the range [f, t).
724  */
725 static long region_count(struct resv_map *resv, long f, long t)
726 {
727         struct list_head *head = &resv->regions;
728         struct file_region *rg;
729         long chg = 0;
730
731         spin_lock(&resv->lock);
732         /* Locate each segment we overlap with, and count that overlap. */
733         list_for_each_entry(rg, head, link) {
734                 long seg_from;
735                 long seg_to;
736
737                 if (rg->to <= f)
738                         continue;
739                 if (rg->from >= t)
740                         break;
741
742                 seg_from = max(rg->from, f);
743                 seg_to = min(rg->to, t);
744
745                 chg += seg_to - seg_from;
746         }
747         spin_unlock(&resv->lock);
748
749         return chg;
750 }
751
752 /*
753  * Convert the address within this vma to the page offset within
754  * the mapping, in pagecache page units; huge pages here.
755  */
756 static pgoff_t vma_hugecache_offset(struct hstate *h,
757                         struct vm_area_struct *vma, unsigned long address)
758 {
759         return ((address - vma->vm_start) >> huge_page_shift(h)) +
760                         (vma->vm_pgoff >> huge_page_order(h));
761 }
762
763 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
764                                      unsigned long address)
765 {
766         return vma_hugecache_offset(hstate_vma(vma), vma, address);
767 }
768 EXPORT_SYMBOL_GPL(linear_hugepage_index);
769
770 /*
771  * Return the size of the pages allocated when backing a VMA. In the majority
772  * cases this will be same size as used by the page table entries.
773  */
774 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
775 {
776         if (vma->vm_ops && vma->vm_ops->pagesize)
777                 return vma->vm_ops->pagesize(vma);
778         return PAGE_SIZE;
779 }
780 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
781
782 /*
783  * Return the page size being used by the MMU to back a VMA. In the majority
784  * of cases, the page size used by the kernel matches the MMU size. On
785  * architectures where it differs, an architecture-specific 'strong'
786  * version of this symbol is required.
787  */
788 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
789 {
790         return vma_kernel_pagesize(vma);
791 }
792
793 /*
794  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
795  * bits of the reservation map pointer, which are always clear due to
796  * alignment.
797  */
798 #define HPAGE_RESV_OWNER    (1UL << 0)
799 #define HPAGE_RESV_UNMAPPED (1UL << 1)
800 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
801
802 /*
803  * These helpers are used to track how many pages are reserved for
804  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
805  * is guaranteed to have their future faults succeed.
806  *
807  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
808  * the reserve counters are updated with the hugetlb_lock held. It is safe
809  * to reset the VMA at fork() time as it is not in use yet and there is no
810  * chance of the global counters getting corrupted as a result of the values.
811  *
812  * The private mapping reservation is represented in a subtly different
813  * manner to a shared mapping.  A shared mapping has a region map associated
814  * with the underlying file, this region map represents the backing file
815  * pages which have ever had a reservation assigned which this persists even
816  * after the page is instantiated.  A private mapping has a region map
817  * associated with the original mmap which is attached to all VMAs which
818  * reference it, this region map represents those offsets which have consumed
819  * reservation ie. where pages have been instantiated.
820  */
821 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
822 {
823         return (unsigned long)vma->vm_private_data;
824 }
825
826 static void set_vma_private_data(struct vm_area_struct *vma,
827                                                         unsigned long value)
828 {
829         vma->vm_private_data = (void *)value;
830 }
831
832 static void
833 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
834                                           struct hugetlb_cgroup *h_cg,
835                                           struct hstate *h)
836 {
837 #ifdef CONFIG_CGROUP_HUGETLB
838         if (!h_cg || !h) {
839                 resv_map->reservation_counter = NULL;
840                 resv_map->pages_per_hpage = 0;
841                 resv_map->css = NULL;
842         } else {
843                 resv_map->reservation_counter =
844                         &h_cg->rsvd_hugepage[hstate_index(h)];
845                 resv_map->pages_per_hpage = pages_per_huge_page(h);
846                 resv_map->css = &h_cg->css;
847         }
848 #endif
849 }
850
851 struct resv_map *resv_map_alloc(void)
852 {
853         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
854         struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
855
856         if (!resv_map || !rg) {
857                 kfree(resv_map);
858                 kfree(rg);
859                 return NULL;
860         }
861
862         kref_init(&resv_map->refs);
863         spin_lock_init(&resv_map->lock);
864         INIT_LIST_HEAD(&resv_map->regions);
865
866         resv_map->adds_in_progress = 0;
867         /*
868          * Initialize these to 0. On shared mappings, 0's here indicate these
869          * fields don't do cgroup accounting. On private mappings, these will be
870          * re-initialized to the proper values, to indicate that hugetlb cgroup
871          * reservations are to be un-charged from here.
872          */
873         resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
874
875         INIT_LIST_HEAD(&resv_map->region_cache);
876         list_add(&rg->link, &resv_map->region_cache);
877         resv_map->region_cache_count = 1;
878
879         return resv_map;
880 }
881
882 void resv_map_release(struct kref *ref)
883 {
884         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
885         struct list_head *head = &resv_map->region_cache;
886         struct file_region *rg, *trg;
887
888         /* Clear out any active regions before we release the map. */
889         region_del(resv_map, 0, LONG_MAX);
890
891         /* ... and any entries left in the cache */
892         list_for_each_entry_safe(rg, trg, head, link) {
893                 list_del(&rg->link);
894                 kfree(rg);
895         }
896
897         VM_BUG_ON(resv_map->adds_in_progress);
898
899         kfree(resv_map);
900 }
901
902 static inline struct resv_map *inode_resv_map(struct inode *inode)
903 {
904         /*
905          * At inode evict time, i_mapping may not point to the original
906          * address space within the inode.  This original address space
907          * contains the pointer to the resv_map.  So, always use the
908          * address space embedded within the inode.
909          * The VERY common case is inode->mapping == &inode->i_data but,
910          * this may not be true for device special inodes.
911          */
912         return (struct resv_map *)(&inode->i_data)->private_data;
913 }
914
915 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
916 {
917         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
918         if (vma->vm_flags & VM_MAYSHARE) {
919                 struct address_space *mapping = vma->vm_file->f_mapping;
920                 struct inode *inode = mapping->host;
921
922                 return inode_resv_map(inode);
923
924         } else {
925                 return (struct resv_map *)(get_vma_private_data(vma) &
926                                                         ~HPAGE_RESV_MASK);
927         }
928 }
929
930 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
931 {
932         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
933         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
934
935         set_vma_private_data(vma, (get_vma_private_data(vma) &
936                                 HPAGE_RESV_MASK) | (unsigned long)map);
937 }
938
939 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
940 {
941         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
942         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
943
944         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
945 }
946
947 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
948 {
949         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
950
951         return (get_vma_private_data(vma) & flag) != 0;
952 }
953
954 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
955 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
956 {
957         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
958         if (!(vma->vm_flags & VM_MAYSHARE))
959                 vma->vm_private_data = (void *)0;
960 }
961
962 /* Returns true if the VMA has associated reserve pages */
963 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
964 {
965         if (vma->vm_flags & VM_NORESERVE) {
966                 /*
967                  * This address is already reserved by other process(chg == 0),
968                  * so, we should decrement reserved count. Without decrementing,
969                  * reserve count remains after releasing inode, because this
970                  * allocated page will go into page cache and is regarded as
971                  * coming from reserved pool in releasing step.  Currently, we
972                  * don't have any other solution to deal with this situation
973                  * properly, so add work-around here.
974                  */
975                 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
976                         return true;
977                 else
978                         return false;
979         }
980
981         /* Shared mappings always use reserves */
982         if (vma->vm_flags & VM_MAYSHARE) {
983                 /*
984                  * We know VM_NORESERVE is not set.  Therefore, there SHOULD
985                  * be a region map for all pages.  The only situation where
986                  * there is no region map is if a hole was punched via
987                  * fallocate.  In this case, there really are no reserves to
988                  * use.  This situation is indicated if chg != 0.
989                  */
990                 if (chg)
991                         return false;
992                 else
993                         return true;
994         }
995
996         /*
997          * Only the process that called mmap() has reserves for
998          * private mappings.
999          */
1000         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1001                 /*
1002                  * Like the shared case above, a hole punch or truncate
1003                  * could have been performed on the private mapping.
1004                  * Examine the value of chg to determine if reserves
1005                  * actually exist or were previously consumed.
1006                  * Very Subtle - The value of chg comes from a previous
1007                  * call to vma_needs_reserves().  The reserve map for
1008                  * private mappings has different (opposite) semantics
1009                  * than that of shared mappings.  vma_needs_reserves()
1010                  * has already taken this difference in semantics into
1011                  * account.  Therefore, the meaning of chg is the same
1012                  * as in the shared case above.  Code could easily be
1013                  * combined, but keeping it separate draws attention to
1014                  * subtle differences.
1015                  */
1016                 if (chg)
1017                         return false;
1018                 else
1019                         return true;
1020         }
1021
1022         return false;
1023 }
1024
1025 static void enqueue_huge_page(struct hstate *h, struct page *page)
1026 {
1027         int nid = page_to_nid(page);
1028         list_move(&page->lru, &h->hugepage_freelists[nid]);
1029         h->free_huge_pages++;
1030         h->free_huge_pages_node[nid]++;
1031 }
1032
1033 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1034 {
1035         struct page *page;
1036         bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA);
1037
1038         list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1039                 if (nocma && is_migrate_cma_page(page))
1040                         continue;
1041
1042                 if (PageHWPoison(page))
1043                         continue;
1044
1045                 list_move(&page->lru, &h->hugepage_activelist);
1046                 set_page_refcounted(page);
1047                 h->free_huge_pages--;
1048                 h->free_huge_pages_node[nid]--;
1049                 return page;
1050         }
1051
1052         return NULL;
1053 }
1054
1055 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1056                 nodemask_t *nmask)
1057 {
1058         unsigned int cpuset_mems_cookie;
1059         struct zonelist *zonelist;
1060         struct zone *zone;
1061         struct zoneref *z;
1062         int node = NUMA_NO_NODE;
1063
1064         zonelist = node_zonelist(nid, gfp_mask);
1065
1066 retry_cpuset:
1067         cpuset_mems_cookie = read_mems_allowed_begin();
1068         for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1069                 struct page *page;
1070
1071                 if (!cpuset_zone_allowed(zone, gfp_mask))
1072                         continue;
1073                 /*
1074                  * no need to ask again on the same node. Pool is node rather than
1075                  * zone aware
1076                  */
1077                 if (zone_to_nid(zone) == node)
1078                         continue;
1079                 node = zone_to_nid(zone);
1080
1081                 page = dequeue_huge_page_node_exact(h, node);
1082                 if (page)
1083                         return page;
1084         }
1085         if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1086                 goto retry_cpuset;
1087
1088         return NULL;
1089 }
1090
1091 static struct page *dequeue_huge_page_vma(struct hstate *h,
1092                                 struct vm_area_struct *vma,
1093                                 unsigned long address, int avoid_reserve,
1094                                 long chg)
1095 {
1096         struct page *page;
1097         struct mempolicy *mpol;
1098         gfp_t gfp_mask;
1099         nodemask_t *nodemask;
1100         int nid;
1101
1102         /*
1103          * A child process with MAP_PRIVATE mappings created by their parent
1104          * have no page reserves. This check ensures that reservations are
1105          * not "stolen". The child may still get SIGKILLed
1106          */
1107         if (!vma_has_reserves(vma, chg) &&
1108                         h->free_huge_pages - h->resv_huge_pages == 0)
1109                 goto err;
1110
1111         /* If reserves cannot be used, ensure enough pages are in the pool */
1112         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1113                 goto err;
1114
1115         gfp_mask = htlb_alloc_mask(h);
1116         nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1117         page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1118         if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1119                 SetPagePrivate(page);
1120                 h->resv_huge_pages--;
1121         }
1122
1123         mpol_cond_put(mpol);
1124         return page;
1125
1126 err:
1127         return NULL;
1128 }
1129
1130 /*
1131  * common helper functions for hstate_next_node_to_{alloc|free}.
1132  * We may have allocated or freed a huge page based on a different
1133  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1134  * be outside of *nodes_allowed.  Ensure that we use an allowed
1135  * node for alloc or free.
1136  */
1137 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1138 {
1139         nid = next_node_in(nid, *nodes_allowed);
1140         VM_BUG_ON(nid >= MAX_NUMNODES);
1141
1142         return nid;
1143 }
1144
1145 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1146 {
1147         if (!node_isset(nid, *nodes_allowed))
1148                 nid = next_node_allowed(nid, nodes_allowed);
1149         return nid;
1150 }
1151
1152 /*
1153  * returns the previously saved node ["this node"] from which to
1154  * allocate a persistent huge page for the pool and advance the
1155  * next node from which to allocate, handling wrap at end of node
1156  * mask.
1157  */
1158 static int hstate_next_node_to_alloc(struct hstate *h,
1159                                         nodemask_t *nodes_allowed)
1160 {
1161         int nid;
1162
1163         VM_BUG_ON(!nodes_allowed);
1164
1165         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1166         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1167
1168         return nid;
1169 }
1170
1171 /*
1172  * helper for free_pool_huge_page() - return the previously saved
1173  * node ["this node"] from which to free a huge page.  Advance the
1174  * next node id whether or not we find a free huge page to free so
1175  * that the next attempt to free addresses the next node.
1176  */
1177 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1178 {
1179         int nid;
1180
1181         VM_BUG_ON(!nodes_allowed);
1182
1183         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1184         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1185
1186         return nid;
1187 }
1188
1189 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)           \
1190         for (nr_nodes = nodes_weight(*mask);                            \
1191                 nr_nodes > 0 &&                                         \
1192                 ((node = hstate_next_node_to_alloc(hs, mask)) || 1);    \
1193                 nr_nodes--)
1194
1195 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)            \
1196         for (nr_nodes = nodes_weight(*mask);                            \
1197                 nr_nodes > 0 &&                                         \
1198                 ((node = hstate_next_node_to_free(hs, mask)) || 1);     \
1199                 nr_nodes--)
1200
1201 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1202 static void destroy_compound_gigantic_page(struct page *page,
1203                                         unsigned int order)
1204 {
1205         int i;
1206         int nr_pages = 1 << order;
1207         struct page *p = page + 1;
1208
1209         atomic_set(compound_mapcount_ptr(page), 0);
1210         if (hpage_pincount_available(page))
1211                 atomic_set(compound_pincount_ptr(page), 0);
1212
1213         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1214                 clear_compound_head(p);
1215                 set_page_refcounted(p);
1216         }
1217
1218         set_compound_order(page, 0);
1219         page[1].compound_nr = 0;
1220         __ClearPageHead(page);
1221 }
1222
1223 static void free_gigantic_page(struct page *page, unsigned int order)
1224 {
1225         /*
1226          * If the page isn't allocated using the cma allocator,
1227          * cma_release() returns false.
1228          */
1229 #ifdef CONFIG_CMA
1230         if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1231                 return;
1232 #endif
1233
1234         free_contig_range(page_to_pfn(page), 1 << order);
1235 }
1236
1237 #ifdef CONFIG_CONTIG_ALLOC
1238 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1239                 int nid, nodemask_t *nodemask)
1240 {
1241         unsigned long nr_pages = 1UL << huge_page_order(h);
1242         if (nid == NUMA_NO_NODE)
1243                 nid = numa_mem_id();
1244
1245 #ifdef CONFIG_CMA
1246         {
1247                 struct page *page;
1248                 int node;
1249
1250                 if (hugetlb_cma[nid]) {
1251                         page = cma_alloc(hugetlb_cma[nid], nr_pages,
1252                                         huge_page_order(h), true);
1253                         if (page)
1254                                 return page;
1255                 }
1256
1257                 if (!(gfp_mask & __GFP_THISNODE)) {
1258                         for_each_node_mask(node, *nodemask) {
1259                                 if (node == nid || !hugetlb_cma[node])
1260                                         continue;
1261
1262                                 page = cma_alloc(hugetlb_cma[node], nr_pages,
1263                                                 huge_page_order(h), true);
1264                                 if (page)
1265                                         return page;
1266                         }
1267                 }
1268         }
1269 #endif
1270
1271         return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1272 }
1273
1274 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1275 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1276 #else /* !CONFIG_CONTIG_ALLOC */
1277 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1278                                         int nid, nodemask_t *nodemask)
1279 {
1280         return NULL;
1281 }
1282 #endif /* CONFIG_CONTIG_ALLOC */
1283
1284 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1285 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1286                                         int nid, nodemask_t *nodemask)
1287 {
1288         return NULL;
1289 }
1290 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1291 static inline void destroy_compound_gigantic_page(struct page *page,
1292                                                 unsigned int order) { }
1293 #endif
1294
1295 static void update_and_free_page(struct hstate *h, struct page *page)
1296 {
1297         int i;
1298
1299         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1300                 return;
1301
1302         h->nr_huge_pages--;
1303         h->nr_huge_pages_node[page_to_nid(page)]--;
1304         for (i = 0; i < pages_per_huge_page(h); i++) {
1305                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1306                                 1 << PG_referenced | 1 << PG_dirty |
1307                                 1 << PG_active | 1 << PG_private |
1308                                 1 << PG_writeback);
1309         }
1310         VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1311         VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1312         set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1313         set_page_refcounted(page);
1314         if (hstate_is_gigantic(h)) {
1315                 /*
1316                  * Temporarily drop the hugetlb_lock, because
1317                  * we might block in free_gigantic_page().
1318                  */
1319                 spin_unlock(&hugetlb_lock);
1320                 destroy_compound_gigantic_page(page, huge_page_order(h));
1321                 free_gigantic_page(page, huge_page_order(h));
1322                 spin_lock(&hugetlb_lock);
1323         } else {
1324                 __free_pages(page, huge_page_order(h));
1325         }
1326 }
1327
1328 struct hstate *size_to_hstate(unsigned long size)
1329 {
1330         struct hstate *h;
1331
1332         for_each_hstate(h) {
1333                 if (huge_page_size(h) == size)
1334                         return h;
1335         }
1336         return NULL;
1337 }
1338
1339 /*
1340  * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1341  * to hstate->hugepage_activelist.)
1342  *
1343  * This function can be called for tail pages, but never returns true for them.
1344  */
1345 bool page_huge_active(struct page *page)
1346 {
1347         VM_BUG_ON_PAGE(!PageHuge(page), page);
1348         return PageHead(page) && PagePrivate(&page[1]);
1349 }
1350
1351 /* never called for tail page */
1352 static void set_page_huge_active(struct page *page)
1353 {
1354         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1355         SetPagePrivate(&page[1]);
1356 }
1357
1358 static void clear_page_huge_active(struct page *page)
1359 {
1360         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1361         ClearPagePrivate(&page[1]);
1362 }
1363
1364 /*
1365  * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1366  * code
1367  */
1368 static inline bool PageHugeTemporary(struct page *page)
1369 {
1370         if (!PageHuge(page))
1371                 return false;
1372
1373         return (unsigned long)page[2].mapping == -1U;
1374 }
1375
1376 static inline void SetPageHugeTemporary(struct page *page)
1377 {
1378         page[2].mapping = (void *)-1U;
1379 }
1380
1381 static inline void ClearPageHugeTemporary(struct page *page)
1382 {
1383         page[2].mapping = NULL;
1384 }
1385
1386 static void __free_huge_page(struct page *page)
1387 {
1388         /*
1389          * Can't pass hstate in here because it is called from the
1390          * compound page destructor.
1391          */
1392         struct hstate *h = page_hstate(page);
1393         int nid = page_to_nid(page);
1394         struct hugepage_subpool *spool =
1395                 (struct hugepage_subpool *)page_private(page);
1396         bool restore_reserve;
1397
1398         VM_BUG_ON_PAGE(page_count(page), page);
1399         VM_BUG_ON_PAGE(page_mapcount(page), page);
1400
1401         set_page_private(page, 0);
1402         page->mapping = NULL;
1403         restore_reserve = PagePrivate(page);
1404         ClearPagePrivate(page);
1405
1406         /*
1407          * If PagePrivate() was set on page, page allocation consumed a
1408          * reservation.  If the page was associated with a subpool, there
1409          * would have been a page reserved in the subpool before allocation
1410          * via hugepage_subpool_get_pages().  Since we are 'restoring' the
1411          * reservtion, do not call hugepage_subpool_put_pages() as this will
1412          * remove the reserved page from the subpool.
1413          */
1414         if (!restore_reserve) {
1415                 /*
1416                  * A return code of zero implies that the subpool will be
1417                  * under its minimum size if the reservation is not restored
1418                  * after page is free.  Therefore, force restore_reserve
1419                  * operation.
1420                  */
1421                 if (hugepage_subpool_put_pages(spool, 1) == 0)
1422                         restore_reserve = true;
1423         }
1424
1425         spin_lock(&hugetlb_lock);
1426         clear_page_huge_active(page);
1427         hugetlb_cgroup_uncharge_page(hstate_index(h),
1428                                      pages_per_huge_page(h), page);
1429         hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1430                                           pages_per_huge_page(h), page);
1431         if (restore_reserve)
1432                 h->resv_huge_pages++;
1433
1434         if (PageHugeTemporary(page)) {
1435                 list_del(&page->lru);
1436                 ClearPageHugeTemporary(page);
1437                 update_and_free_page(h, page);
1438         } else if (h->surplus_huge_pages_node[nid]) {
1439                 /* remove the page from active list */
1440                 list_del(&page->lru);
1441                 update_and_free_page(h, page);
1442                 h->surplus_huge_pages--;
1443                 h->surplus_huge_pages_node[nid]--;
1444         } else {
1445                 arch_clear_hugepage_flags(page);
1446                 enqueue_huge_page(h, page);
1447         }
1448         spin_unlock(&hugetlb_lock);
1449 }
1450
1451 /*
1452  * As free_huge_page() can be called from a non-task context, we have
1453  * to defer the actual freeing in a workqueue to prevent potential
1454  * hugetlb_lock deadlock.
1455  *
1456  * free_hpage_workfn() locklessly retrieves the linked list of pages to
1457  * be freed and frees them one-by-one. As the page->mapping pointer is
1458  * going to be cleared in __free_huge_page() anyway, it is reused as the
1459  * llist_node structure of a lockless linked list of huge pages to be freed.
1460  */
1461 static LLIST_HEAD(hpage_freelist);
1462
1463 static void free_hpage_workfn(struct work_struct *work)
1464 {
1465         struct llist_node *node;
1466         struct page *page;
1467
1468         node = llist_del_all(&hpage_freelist);
1469
1470         while (node) {
1471                 page = container_of((struct address_space **)node,
1472                                      struct page, mapping);
1473                 node = node->next;
1474                 __free_huge_page(page);
1475         }
1476 }
1477 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1478
1479 void free_huge_page(struct page *page)
1480 {
1481         /*
1482          * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1483          */
1484         if (!in_task()) {
1485                 /*
1486                  * Only call schedule_work() if hpage_freelist is previously
1487                  * empty. Otherwise, schedule_work() had been called but the
1488                  * workfn hasn't retrieved the list yet.
1489                  */
1490                 if (llist_add((struct llist_node *)&page->mapping,
1491                               &hpage_freelist))
1492                         schedule_work(&free_hpage_work);
1493                 return;
1494         }
1495
1496         __free_huge_page(page);
1497 }
1498
1499 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1500 {
1501         INIT_LIST_HEAD(&page->lru);
1502         set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1503         set_hugetlb_cgroup(page, NULL);
1504         set_hugetlb_cgroup_rsvd(page, NULL);
1505         spin_lock(&hugetlb_lock);
1506         h->nr_huge_pages++;
1507         h->nr_huge_pages_node[nid]++;
1508         spin_unlock(&hugetlb_lock);
1509 }
1510
1511 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1512 {
1513         int i;
1514         int nr_pages = 1 << order;
1515         struct page *p = page + 1;
1516
1517         /* we rely on prep_new_huge_page to set the destructor */
1518         set_compound_order(page, order);
1519         __ClearPageReserved(page);
1520         __SetPageHead(page);
1521         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1522                 /*
1523                  * For gigantic hugepages allocated through bootmem at
1524                  * boot, it's safer to be consistent with the not-gigantic
1525                  * hugepages and clear the PG_reserved bit from all tail pages
1526                  * too.  Otherwise drivers using get_user_pages() to access tail
1527                  * pages may get the reference counting wrong if they see
1528                  * PG_reserved set on a tail page (despite the head page not
1529                  * having PG_reserved set).  Enforcing this consistency between
1530                  * head and tail pages allows drivers to optimize away a check
1531                  * on the head page when they need know if put_page() is needed
1532                  * after get_user_pages().
1533                  */
1534                 __ClearPageReserved(p);
1535                 set_page_count(p, 0);
1536                 set_compound_head(p, page);
1537         }
1538         atomic_set(compound_mapcount_ptr(page), -1);
1539
1540         if (hpage_pincount_available(page))
1541                 atomic_set(compound_pincount_ptr(page), 0);
1542 }
1543
1544 /*
1545  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1546  * transparent huge pages.  See the PageTransHuge() documentation for more
1547  * details.
1548  */
1549 int PageHuge(struct page *page)
1550 {
1551         if (!PageCompound(page))
1552                 return 0;
1553
1554         page = compound_head(page);
1555         return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1556 }
1557 EXPORT_SYMBOL_GPL(PageHuge);
1558
1559 /*
1560  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1561  * normal or transparent huge pages.
1562  */
1563 int PageHeadHuge(struct page *page_head)
1564 {
1565         if (!PageHead(page_head))
1566                 return 0;
1567
1568         return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1569 }
1570
1571 /*
1572  * Find and lock address space (mapping) in write mode.
1573  *
1574  * Upon entry, the page is locked which means that page_mapping() is
1575  * stable.  Due to locking order, we can only trylock_write.  If we can
1576  * not get the lock, simply return NULL to caller.
1577  */
1578 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1579 {
1580         struct address_space *mapping = page_mapping(hpage);
1581
1582         if (!mapping)
1583                 return mapping;
1584
1585         if (i_mmap_trylock_write(mapping))
1586                 return mapping;
1587
1588         return NULL;
1589 }
1590
1591 pgoff_t __basepage_index(struct page *page)
1592 {
1593         struct page *page_head = compound_head(page);
1594         pgoff_t index = page_index(page_head);
1595         unsigned long compound_idx;
1596
1597         if (!PageHuge(page_head))
1598                 return page_index(page);
1599
1600         if (compound_order(page_head) >= MAX_ORDER)
1601                 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1602         else
1603                 compound_idx = page - page_head;
1604
1605         return (index << compound_order(page_head)) + compound_idx;
1606 }
1607
1608 static struct page *alloc_buddy_huge_page(struct hstate *h,
1609                 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1610                 nodemask_t *node_alloc_noretry)
1611 {
1612         int order = huge_page_order(h);
1613         struct page *page;
1614         bool alloc_try_hard = true;
1615
1616         /*
1617          * By default we always try hard to allocate the page with
1618          * __GFP_RETRY_MAYFAIL flag.  However, if we are allocating pages in
1619          * a loop (to adjust global huge page counts) and previous allocation
1620          * failed, do not continue to try hard on the same node.  Use the
1621          * node_alloc_noretry bitmap to manage this state information.
1622          */
1623         if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1624                 alloc_try_hard = false;
1625         gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1626         if (alloc_try_hard)
1627                 gfp_mask |= __GFP_RETRY_MAYFAIL;
1628         if (nid == NUMA_NO_NODE)
1629                 nid = numa_mem_id();
1630         page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1631         if (page)
1632                 __count_vm_event(HTLB_BUDDY_PGALLOC);
1633         else
1634                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1635
1636         /*
1637          * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1638          * indicates an overall state change.  Clear bit so that we resume
1639          * normal 'try hard' allocations.
1640          */
1641         if (node_alloc_noretry && page && !alloc_try_hard)
1642                 node_clear(nid, *node_alloc_noretry);
1643
1644         /*
1645          * If we tried hard to get a page but failed, set bit so that
1646          * subsequent attempts will not try as hard until there is an
1647          * overall state change.
1648          */
1649         if (node_alloc_noretry && !page && alloc_try_hard)
1650                 node_set(nid, *node_alloc_noretry);
1651
1652         return page;
1653 }
1654
1655 /*
1656  * Common helper to allocate a fresh hugetlb page. All specific allocators
1657  * should use this function to get new hugetlb pages
1658  */
1659 static struct page *alloc_fresh_huge_page(struct hstate *h,
1660                 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1661                 nodemask_t *node_alloc_noretry)
1662 {
1663         struct page *page;
1664
1665         if (hstate_is_gigantic(h))
1666                 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1667         else
1668                 page = alloc_buddy_huge_page(h, gfp_mask,
1669                                 nid, nmask, node_alloc_noretry);
1670         if (!page)
1671                 return NULL;
1672
1673         if (hstate_is_gigantic(h))
1674                 prep_compound_gigantic_page(page, huge_page_order(h));
1675         prep_new_huge_page(h, page, page_to_nid(page));
1676
1677         return page;
1678 }
1679
1680 /*
1681  * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1682  * manner.
1683  */
1684 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1685                                 nodemask_t *node_alloc_noretry)
1686 {
1687         struct page *page;
1688         int nr_nodes, node;
1689         gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1690
1691         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1692                 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1693                                                 node_alloc_noretry);
1694                 if (page)
1695                         break;
1696         }
1697
1698         if (!page)
1699                 return 0;
1700
1701         put_page(page); /* free it into the hugepage allocator */
1702
1703         return 1;
1704 }
1705
1706 /*
1707  * Free huge page from pool from next node to free.
1708  * Attempt to keep persistent huge pages more or less
1709  * balanced over allowed nodes.
1710  * Called with hugetlb_lock locked.
1711  */
1712 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1713                                                          bool acct_surplus)
1714 {
1715         int nr_nodes, node;
1716         int ret = 0;
1717
1718         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1719                 /*
1720                  * If we're returning unused surplus pages, only examine
1721                  * nodes with surplus pages.
1722                  */
1723                 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1724                     !list_empty(&h->hugepage_freelists[node])) {
1725                         struct page *page =
1726                                 list_entry(h->hugepage_freelists[node].next,
1727                                           struct page, lru);
1728                         list_del(&page->lru);
1729                         h->free_huge_pages--;
1730                         h->free_huge_pages_node[node]--;
1731                         if (acct_surplus) {
1732                                 h->surplus_huge_pages--;
1733                                 h->surplus_huge_pages_node[node]--;
1734                         }
1735                         update_and_free_page(h, page);
1736                         ret = 1;
1737                         break;
1738                 }
1739         }
1740
1741         return ret;
1742 }
1743
1744 /*
1745  * Dissolve a given free hugepage into free buddy pages. This function does
1746  * nothing for in-use hugepages and non-hugepages.
1747  * This function returns values like below:
1748  *
1749  *  -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1750  *          (allocated or reserved.)
1751  *       0: successfully dissolved free hugepages or the page is not a
1752  *          hugepage (considered as already dissolved)
1753  */
1754 int dissolve_free_huge_page(struct page *page)
1755 {
1756         int rc = -EBUSY;
1757
1758         /* Not to disrupt normal path by vainly holding hugetlb_lock */
1759         if (!PageHuge(page))
1760                 return 0;
1761
1762         spin_lock(&hugetlb_lock);
1763         if (!PageHuge(page)) {
1764                 rc = 0;
1765                 goto out;
1766         }
1767
1768         if (!page_count(page)) {
1769                 struct page *head = compound_head(page);
1770                 struct hstate *h = page_hstate(head);
1771                 int nid = page_to_nid(head);
1772                 if (h->free_huge_pages - h->resv_huge_pages == 0)
1773                         goto out;
1774                 /*
1775                  * Move PageHWPoison flag from head page to the raw error page,
1776                  * which makes any subpages rather than the error page reusable.
1777                  */
1778                 if (PageHWPoison(head) && page != head) {
1779                         SetPageHWPoison(page);
1780                         ClearPageHWPoison(head);
1781                 }
1782                 list_del(&head->lru);
1783                 h->free_huge_pages--;
1784                 h->free_huge_pages_node[nid]--;
1785                 h->max_huge_pages--;
1786                 update_and_free_page(h, head);
1787                 rc = 0;
1788         }
1789 out:
1790         spin_unlock(&hugetlb_lock);
1791         return rc;
1792 }
1793
1794 /*
1795  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1796  * make specified memory blocks removable from the system.
1797  * Note that this will dissolve a free gigantic hugepage completely, if any
1798  * part of it lies within the given range.
1799  * Also note that if dissolve_free_huge_page() returns with an error, all
1800  * free hugepages that were dissolved before that error are lost.
1801  */
1802 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1803 {
1804         unsigned long pfn;
1805         struct page *page;
1806         int rc = 0;
1807
1808         if (!hugepages_supported())
1809                 return rc;
1810
1811         for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1812                 page = pfn_to_page(pfn);
1813                 rc = dissolve_free_huge_page(page);
1814                 if (rc)
1815                         break;
1816         }
1817
1818         return rc;
1819 }
1820
1821 /*
1822  * Allocates a fresh surplus page from the page allocator.
1823  */
1824 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1825                 int nid, nodemask_t *nmask)
1826 {
1827         struct page *page = NULL;
1828
1829         if (hstate_is_gigantic(h))
1830                 return NULL;
1831
1832         spin_lock(&hugetlb_lock);
1833         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1834                 goto out_unlock;
1835         spin_unlock(&hugetlb_lock);
1836
1837         page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1838         if (!page)
1839                 return NULL;
1840
1841         spin_lock(&hugetlb_lock);
1842         /*
1843          * We could have raced with the pool size change.
1844          * Double check that and simply deallocate the new page
1845          * if we would end up overcommiting the surpluses. Abuse
1846          * temporary page to workaround the nasty free_huge_page
1847          * codeflow
1848          */
1849         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1850                 SetPageHugeTemporary(page);
1851                 spin_unlock(&hugetlb_lock);
1852                 put_page(page);
1853                 return NULL;
1854         } else {
1855                 h->surplus_huge_pages++;
1856                 h->surplus_huge_pages_node[page_to_nid(page)]++;
1857         }
1858
1859 out_unlock:
1860         spin_unlock(&hugetlb_lock);
1861
1862         return page;
1863 }
1864
1865 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1866                                      int nid, nodemask_t *nmask)
1867 {
1868         struct page *page;
1869
1870         if (hstate_is_gigantic(h))
1871                 return NULL;
1872
1873         page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1874         if (!page)
1875                 return NULL;
1876
1877         /*
1878          * We do not account these pages as surplus because they are only
1879          * temporary and will be released properly on the last reference
1880          */
1881         SetPageHugeTemporary(page);
1882
1883         return page;
1884 }
1885
1886 /*
1887  * Use the VMA's mpolicy to allocate a huge page from the buddy.
1888  */
1889 static
1890 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1891                 struct vm_area_struct *vma, unsigned long addr)
1892 {
1893         struct page *page;
1894         struct mempolicy *mpol;
1895         gfp_t gfp_mask = htlb_alloc_mask(h);
1896         int nid;
1897         nodemask_t *nodemask;
1898
1899         nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1900         page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1901         mpol_cond_put(mpol);
1902
1903         return page;
1904 }
1905
1906 /* page migration callback function */
1907 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1908                 nodemask_t *nmask, gfp_t gfp_mask)
1909 {
1910         spin_lock(&hugetlb_lock);
1911         if (h->free_huge_pages - h->resv_huge_pages > 0) {
1912                 struct page *page;
1913
1914                 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1915                 if (page) {
1916                         spin_unlock(&hugetlb_lock);
1917                         return page;
1918                 }
1919         }
1920         spin_unlock(&hugetlb_lock);
1921
1922         return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1923 }
1924
1925 /* mempolicy aware migration callback */
1926 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1927                 unsigned long address)
1928 {
1929         struct mempolicy *mpol;
1930         nodemask_t *nodemask;
1931         struct page *page;
1932         gfp_t gfp_mask;
1933         int node;
1934
1935         gfp_mask = htlb_alloc_mask(h);
1936         node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1937         page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
1938         mpol_cond_put(mpol);
1939
1940         return page;
1941 }
1942
1943 /*
1944  * Increase the hugetlb pool such that it can accommodate a reservation
1945  * of size 'delta'.
1946  */
1947 static int gather_surplus_pages(struct hstate *h, long delta)
1948         __must_hold(&hugetlb_lock)
1949 {
1950         struct list_head surplus_list;
1951         struct page *page, *tmp;
1952         int ret;
1953         long i;
1954         long needed, allocated;
1955         bool alloc_ok = true;
1956
1957         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1958         if (needed <= 0) {
1959                 h->resv_huge_pages += delta;
1960                 return 0;
1961         }
1962
1963         allocated = 0;
1964         INIT_LIST_HEAD(&surplus_list);
1965
1966         ret = -ENOMEM;
1967 retry:
1968         spin_unlock(&hugetlb_lock);
1969         for (i = 0; i < needed; i++) {
1970                 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1971                                 NUMA_NO_NODE, NULL);
1972                 if (!page) {
1973                         alloc_ok = false;
1974                         break;
1975                 }
1976                 list_add(&page->lru, &surplus_list);
1977                 cond_resched();
1978         }
1979         allocated += i;
1980
1981         /*
1982          * After retaking hugetlb_lock, we need to recalculate 'needed'
1983          * because either resv_huge_pages or free_huge_pages may have changed.
1984          */
1985         spin_lock(&hugetlb_lock);
1986         needed = (h->resv_huge_pages + delta) -
1987                         (h->free_huge_pages + allocated);
1988         if (needed > 0) {
1989                 if (alloc_ok)
1990                         goto retry;
1991                 /*
1992                  * We were not able to allocate enough pages to
1993                  * satisfy the entire reservation so we free what
1994                  * we've allocated so far.
1995                  */
1996                 goto free;
1997         }
1998         /*
1999          * The surplus_list now contains _at_least_ the number of extra pages
2000          * needed to accommodate the reservation.  Add the appropriate number
2001          * of pages to the hugetlb pool and free the extras back to the buddy
2002          * allocator.  Commit the entire reservation here to prevent another
2003          * process from stealing the pages as they are added to the pool but
2004          * before they are reserved.
2005          */
2006         needed += allocated;
2007         h->resv_huge_pages += delta;
2008         ret = 0;
2009
2010         /* Free the needed pages to the hugetlb pool */
2011         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2012                 if ((--needed) < 0)
2013                         break;
2014                 /*
2015                  * This page is now managed by the hugetlb allocator and has
2016                  * no users -- drop the buddy allocator's reference.
2017                  */
2018                 VM_BUG_ON_PAGE(!put_page_testzero(page), page);
2019                 enqueue_huge_page(h, page);
2020         }
2021 free:
2022         spin_unlock(&hugetlb_lock);
2023
2024         /* Free unnecessary surplus pages to the buddy allocator */
2025         list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2026                 put_page(page);
2027         spin_lock(&hugetlb_lock);
2028
2029         return ret;
2030 }
2031
2032 /*
2033  * This routine has two main purposes:
2034  * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2035  *    in unused_resv_pages.  This corresponds to the prior adjustments made
2036  *    to the associated reservation map.
2037  * 2) Free any unused surplus pages that may have been allocated to satisfy
2038  *    the reservation.  As many as unused_resv_pages may be freed.
2039  *
2040  * Called with hugetlb_lock held.  However, the lock could be dropped (and
2041  * reacquired) during calls to cond_resched_lock.  Whenever dropping the lock,
2042  * we must make sure nobody else can claim pages we are in the process of
2043  * freeing.  Do this by ensuring resv_huge_page always is greater than the
2044  * number of huge pages we plan to free when dropping the lock.
2045  */
2046 static void return_unused_surplus_pages(struct hstate *h,
2047                                         unsigned long unused_resv_pages)
2048 {
2049         unsigned long nr_pages;
2050
2051         /* Cannot return gigantic pages currently */
2052         if (hstate_is_gigantic(h))
2053                 goto out;
2054
2055         /*
2056          * Part (or even all) of the reservation could have been backed
2057          * by pre-allocated pages. Only free surplus pages.
2058          */
2059         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2060
2061         /*
2062          * We want to release as many surplus pages as possible, spread
2063          * evenly across all nodes with memory. Iterate across these nodes
2064          * until we can no longer free unreserved surplus pages. This occurs
2065          * when the nodes with surplus pages have no free pages.
2066          * free_pool_huge_page() will balance the freed pages across the
2067          * on-line nodes with memory and will handle the hstate accounting.
2068          *
2069          * Note that we decrement resv_huge_pages as we free the pages.  If
2070          * we drop the lock, resv_huge_pages will still be sufficiently large
2071          * to cover subsequent pages we may free.
2072          */
2073         while (nr_pages--) {
2074                 h->resv_huge_pages--;
2075                 unused_resv_pages--;
2076                 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
2077                         goto out;
2078                 cond_resched_lock(&hugetlb_lock);
2079         }
2080
2081 out:
2082         /* Fully uncommit the reservation */
2083         h->resv_huge_pages -= unused_resv_pages;
2084 }
2085
2086
2087 /*
2088  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2089  * are used by the huge page allocation routines to manage reservations.
2090  *
2091  * vma_needs_reservation is called to determine if the huge page at addr
2092  * within the vma has an associated reservation.  If a reservation is
2093  * needed, the value 1 is returned.  The caller is then responsible for
2094  * managing the global reservation and subpool usage counts.  After
2095  * the huge page has been allocated, vma_commit_reservation is called
2096  * to add the page to the reservation map.  If the page allocation fails,
2097  * the reservation must be ended instead of committed.  vma_end_reservation
2098  * is called in such cases.
2099  *
2100  * In the normal case, vma_commit_reservation returns the same value
2101  * as the preceding vma_needs_reservation call.  The only time this
2102  * is not the case is if a reserve map was changed between calls.  It
2103  * is the responsibility of the caller to notice the difference and
2104  * take appropriate action.
2105  *
2106  * vma_add_reservation is used in error paths where a reservation must
2107  * be restored when a newly allocated huge page must be freed.  It is
2108  * to be called after calling vma_needs_reservation to determine if a
2109  * reservation exists.
2110  */
2111 enum vma_resv_mode {
2112         VMA_NEEDS_RESV,
2113         VMA_COMMIT_RESV,
2114         VMA_END_RESV,
2115         VMA_ADD_RESV,
2116 };
2117 static long __vma_reservation_common(struct hstate *h,
2118                                 struct vm_area_struct *vma, unsigned long addr,
2119                                 enum vma_resv_mode mode)
2120 {
2121         struct resv_map *resv;
2122         pgoff_t idx;
2123         long ret;
2124         long dummy_out_regions_needed;
2125
2126         resv = vma_resv_map(vma);
2127         if (!resv)
2128                 return 1;
2129
2130         idx = vma_hugecache_offset(h, vma, addr);
2131         switch (mode) {
2132         case VMA_NEEDS_RESV:
2133                 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2134                 /* We assume that vma_reservation_* routines always operate on
2135                  * 1 page, and that adding to resv map a 1 page entry can only
2136                  * ever require 1 region.
2137                  */
2138                 VM_BUG_ON(dummy_out_regions_needed != 1);
2139                 break;
2140         case VMA_COMMIT_RESV:
2141                 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2142                 /* region_add calls of range 1 should never fail. */
2143                 VM_BUG_ON(ret < 0);
2144                 break;
2145         case VMA_END_RESV:
2146                 region_abort(resv, idx, idx + 1, 1);
2147                 ret = 0;
2148                 break;
2149         case VMA_ADD_RESV:
2150                 if (vma->vm_flags & VM_MAYSHARE) {
2151                         ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2152                         /* region_add calls of range 1 should never fail. */
2153                         VM_BUG_ON(ret < 0);
2154                 } else {
2155                         region_abort(resv, idx, idx + 1, 1);
2156                         ret = region_del(resv, idx, idx + 1);
2157                 }
2158                 break;
2159         default:
2160                 BUG();
2161         }
2162
2163         if (vma->vm_flags & VM_MAYSHARE)
2164                 return ret;
2165         else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
2166                 /*
2167                  * In most cases, reserves always exist for private mappings.
2168                  * However, a file associated with mapping could have been
2169                  * hole punched or truncated after reserves were consumed.
2170                  * As subsequent fault on such a range will not use reserves.
2171                  * Subtle - The reserve map for private mappings has the
2172                  * opposite meaning than that of shared mappings.  If NO
2173                  * entry is in the reserve map, it means a reservation exists.
2174                  * If an entry exists in the reserve map, it means the
2175                  * reservation has already been consumed.  As a result, the
2176                  * return value of this routine is the opposite of the
2177                  * value returned from reserve map manipulation routines above.
2178                  */
2179                 if (ret)
2180                         return 0;
2181                 else
2182                         return 1;
2183         }
2184         else
2185                 return ret < 0 ? ret : 0;
2186 }
2187
2188 static long vma_needs_reservation(struct hstate *h,
2189                         struct vm_area_struct *vma, unsigned long addr)
2190 {
2191         return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2192 }
2193
2194 static long vma_commit_reservation(struct hstate *h,
2195                         struct vm_area_struct *vma, unsigned long addr)
2196 {
2197         return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2198 }
2199
2200 static void vma_end_reservation(struct hstate *h,
2201                         struct vm_area_struct *vma, unsigned long addr)
2202 {
2203         (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2204 }
2205
2206 static long vma_add_reservation(struct hstate *h,
2207                         struct vm_area_struct *vma, unsigned long addr)
2208 {
2209         return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2210 }
2211
2212 /*
2213  * This routine is called to restore a reservation on error paths.  In the
2214  * specific error paths, a huge page was allocated (via alloc_huge_page)
2215  * and is about to be freed.  If a reservation for the page existed,
2216  * alloc_huge_page would have consumed the reservation and set PagePrivate
2217  * in the newly allocated page.  When the page is freed via free_huge_page,
2218  * the global reservation count will be incremented if PagePrivate is set.
2219  * However, free_huge_page can not adjust the reserve map.  Adjust the
2220  * reserve map here to be consistent with global reserve count adjustments
2221  * to be made by free_huge_page.
2222  */
2223 static void restore_reserve_on_error(struct hstate *h,
2224                         struct vm_area_struct *vma, unsigned long address,
2225                         struct page *page)
2226 {
2227         if (unlikely(PagePrivate(page))) {
2228                 long rc = vma_needs_reservation(h, vma, address);
2229
2230                 if (unlikely(rc < 0)) {
2231                         /*
2232                          * Rare out of memory condition in reserve map
2233                          * manipulation.  Clear PagePrivate so that
2234                          * global reserve count will not be incremented
2235                          * by free_huge_page.  This will make it appear
2236                          * as though the reservation for this page was
2237                          * consumed.  This may prevent the task from
2238                          * faulting in the page at a later time.  This
2239                          * is better than inconsistent global huge page
2240                          * accounting of reserve counts.
2241                          */
2242                         ClearPagePrivate(page);
2243                 } else if (rc) {
2244                         rc = vma_add_reservation(h, vma, address);
2245                         if (unlikely(rc < 0))
2246                                 /*
2247                                  * See above comment about rare out of
2248                                  * memory condition.
2249                                  */
2250                                 ClearPagePrivate(page);
2251                 } else
2252                         vma_end_reservation(h, vma, address);
2253         }
2254 }
2255
2256 struct page *alloc_huge_page(struct vm_area_struct *vma,
2257                                     unsigned long addr, int avoid_reserve)
2258 {
2259         struct hugepage_subpool *spool = subpool_vma(vma);
2260         struct hstate *h = hstate_vma(vma);
2261         struct page *page;
2262         long map_chg, map_commit;
2263         long gbl_chg;
2264         int ret, idx;
2265         struct hugetlb_cgroup *h_cg;
2266         bool deferred_reserve;
2267
2268         idx = hstate_index(h);
2269         /*
2270          * Examine the region/reserve map to determine if the process
2271          * has a reservation for the page to be allocated.  A return
2272          * code of zero indicates a reservation exists (no change).
2273          */
2274         map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2275         if (map_chg < 0)
2276                 return ERR_PTR(-ENOMEM);
2277
2278         /*
2279          * Processes that did not create the mapping will have no
2280          * reserves as indicated by the region/reserve map. Check
2281          * that the allocation will not exceed the subpool limit.
2282          * Allocations for MAP_NORESERVE mappings also need to be
2283          * checked against any subpool limit.
2284          */
2285         if (map_chg || avoid_reserve) {
2286                 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2287                 if (gbl_chg < 0) {
2288                         vma_end_reservation(h, vma, addr);
2289                         return ERR_PTR(-ENOSPC);
2290                 }
2291
2292                 /*
2293                  * Even though there was no reservation in the region/reserve
2294                  * map, there could be reservations associated with the
2295                  * subpool that can be used.  This would be indicated if the
2296                  * return value of hugepage_subpool_get_pages() is zero.
2297                  * However, if avoid_reserve is specified we still avoid even
2298                  * the subpool reservations.
2299                  */
2300                 if (avoid_reserve)
2301                         gbl_chg = 1;
2302         }
2303
2304         /* If this allocation is not consuming a reservation, charge it now.
2305          */
2306         deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma);
2307         if (deferred_reserve) {
2308                 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2309                         idx, pages_per_huge_page(h), &h_cg);
2310                 if (ret)
2311                         goto out_subpool_put;
2312         }
2313
2314         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2315         if (ret)
2316                 goto out_uncharge_cgroup_reservation;
2317
2318         spin_lock(&hugetlb_lock);
2319         /*
2320          * glb_chg is passed to indicate whether or not a page must be taken
2321          * from the global free pool (global change).  gbl_chg == 0 indicates
2322          * a reservation exists for the allocation.
2323          */
2324         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2325         if (!page) {
2326                 spin_unlock(&hugetlb_lock);
2327                 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2328                 if (!page)
2329                         goto out_uncharge_cgroup;
2330                 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2331                         SetPagePrivate(page);
2332                         h->resv_huge_pages--;
2333                 }
2334                 spin_lock(&hugetlb_lock);
2335                 list_add(&page->lru, &h->hugepage_activelist);
2336                 /* Fall through */
2337         }
2338         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2339         /* If allocation is not consuming a reservation, also store the
2340          * hugetlb_cgroup pointer on the page.
2341          */
2342         if (deferred_reserve) {
2343                 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2344                                                   h_cg, page);
2345         }
2346
2347         spin_unlock(&hugetlb_lock);
2348
2349         set_page_private(page, (unsigned long)spool);
2350
2351         map_commit = vma_commit_reservation(h, vma, addr);
2352         if (unlikely(map_chg > map_commit)) {
2353                 /*
2354                  * The page was added to the reservation map between
2355                  * vma_needs_reservation and vma_commit_reservation.
2356                  * This indicates a race with hugetlb_reserve_pages.
2357                  * Adjust for the subpool count incremented above AND
2358                  * in hugetlb_reserve_pages for the same page.  Also,
2359                  * the reservation count added in hugetlb_reserve_pages
2360                  * no longer applies.
2361                  */
2362                 long rsv_adjust;
2363
2364                 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2365                 hugetlb_acct_memory(h, -rsv_adjust);
2366                 if (deferred_reserve)
2367                         hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2368                                         pages_per_huge_page(h), page);
2369         }
2370         return page;
2371
2372 out_uncharge_cgroup:
2373         hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2374 out_uncharge_cgroup_reservation:
2375         if (deferred_reserve)
2376                 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2377                                                     h_cg);
2378 out_subpool_put:
2379         if (map_chg || avoid_reserve)
2380                 hugepage_subpool_put_pages(spool, 1);
2381         vma_end_reservation(h, vma, addr);
2382         return ERR_PTR(-ENOSPC);
2383 }
2384
2385 int alloc_bootmem_huge_page(struct hstate *h)
2386         __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2387 int __alloc_bootmem_huge_page(struct hstate *h)
2388 {
2389         struct huge_bootmem_page *m;
2390         int nr_nodes, node;
2391
2392         for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2393                 void *addr;
2394
2395                 addr = memblock_alloc_try_nid_raw(
2396                                 huge_page_size(h), huge_page_size(h),
2397                                 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2398                 if (addr) {
2399                         /*
2400                          * Use the beginning of the huge page to store the
2401                          * huge_bootmem_page struct (until gather_bootmem
2402                          * puts them into the mem_map).
2403                          */
2404                         m = addr;
2405                         goto found;
2406                 }
2407         }
2408         return 0;
2409
2410 found:
2411         BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2412         /* Put them into a private list first because mem_map is not up yet */
2413         INIT_LIST_HEAD(&m->list);
2414         list_add(&m->list, &huge_boot_pages);
2415         m->hstate = h;
2416         return 1;
2417 }
2418
2419 static void __init prep_compound_huge_page(struct page *page,
2420                 unsigned int order)
2421 {
2422         if (unlikely(order > (MAX_ORDER - 1)))
2423                 prep_compound_gigantic_page(page, order);
2424         else
2425                 prep_compound_page(page, order);
2426 }
2427
2428 /* Put bootmem huge pages into the standard lists after mem_map is up */
2429 static void __init gather_bootmem_prealloc(void)
2430 {
2431         struct huge_bootmem_page *m;
2432
2433         list_for_each_entry(m, &huge_boot_pages, list) {
2434                 struct page *page = virt_to_page(m);
2435                 struct hstate *h = m->hstate;
2436
2437                 WARN_ON(page_count(page) != 1);
2438                 prep_compound_huge_page(page, h->order);
2439                 WARN_ON(PageReserved(page));
2440                 prep_new_huge_page(h, page, page_to_nid(page));
2441                 put_page(page); /* free it into the hugepage allocator */
2442
2443                 /*
2444                  * If we had gigantic hugepages allocated at boot time, we need
2445                  * to restore the 'stolen' pages to totalram_pages in order to
2446                  * fix confusing memory reports from free(1) and another
2447                  * side-effects, like CommitLimit going negative.
2448                  */
2449                 if (hstate_is_gigantic(h))
2450                         adjust_managed_page_count(page, 1 << h->order);
2451                 cond_resched();
2452         }
2453 }
2454
2455 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2456 {
2457         unsigned long i;
2458         nodemask_t *node_alloc_noretry;
2459
2460         if (!hstate_is_gigantic(h)) {
2461                 /*
2462                  * Bit mask controlling how hard we retry per-node allocations.
2463                  * Ignore errors as lower level routines can deal with
2464                  * node_alloc_noretry == NULL.  If this kmalloc fails at boot
2465                  * time, we are likely in bigger trouble.
2466                  */
2467                 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2468                                                 GFP_KERNEL);
2469         } else {
2470                 /* allocations done at boot time */
2471                 node_alloc_noretry = NULL;
2472         }
2473
2474         /* bit mask controlling how hard we retry per-node allocations */
2475         if (node_alloc_noretry)
2476                 nodes_clear(*node_alloc_noretry);
2477
2478         for (i = 0; i < h->max_huge_pages; ++i) {
2479                 if (hstate_is_gigantic(h)) {
2480                         if (hugetlb_cma_size) {
2481                                 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2482                                 break;
2483                         }
2484                         if (!alloc_bootmem_huge_page(h))
2485                                 break;
2486                 } else if (!alloc_pool_huge_page(h,
2487                                          &node_states[N_MEMORY],
2488                                          node_alloc_noretry))
2489                         break;
2490                 cond_resched();
2491         }
2492         if (i < h->max_huge_pages) {
2493                 char buf[32];
2494
2495                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2496                 pr_warn("HugeTLB: allocating %lu of page size %s failed.  Only allocated %lu hugepages.\n",
2497                         h->max_huge_pages, buf, i);
2498                 h->max_huge_pages = i;
2499         }
2500
2501         kfree(node_alloc_noretry);
2502 }
2503
2504 static void __init hugetlb_init_hstates(void)
2505 {
2506         struct hstate *h;
2507
2508         for_each_hstate(h) {
2509                 if (minimum_order > huge_page_order(h))
2510                         minimum_order = huge_page_order(h);
2511
2512                 /* oversize hugepages were init'ed in early boot */
2513                 if (!hstate_is_gigantic(h))
2514                         hugetlb_hstate_alloc_pages(h);
2515         }
2516         VM_BUG_ON(minimum_order == UINT_MAX);
2517 }
2518
2519 static void __init report_hugepages(void)
2520 {
2521         struct hstate *h;
2522
2523         for_each_hstate(h) {
2524                 char buf[32];
2525
2526                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2527                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2528                         buf, h->free_huge_pages);
2529         }
2530 }
2531
2532 #ifdef CONFIG_HIGHMEM
2533 static void try_to_free_low(struct hstate *h, unsigned long count,
2534                                                 nodemask_t *nodes_allowed)
2535 {
2536         int i;
2537
2538         if (hstate_is_gigantic(h))
2539                 return;
2540
2541         for_each_node_mask(i, *nodes_allowed) {
2542                 struct page *page, *next;
2543                 struct list_head *freel = &h->hugepage_freelists[i];
2544                 list_for_each_entry_safe(page, next, freel, lru) {
2545                         if (count >= h->nr_huge_pages)
2546                                 return;
2547                         if (PageHighMem(page))
2548                                 continue;
2549                         list_del(&page->lru);
2550                         update_and_free_page(h, page);
2551                         h->free_huge_pages--;
2552                         h->free_huge_pages_node[page_to_nid(page)]--;
2553                 }
2554         }
2555 }
2556 #else
2557 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2558                                                 nodemask_t *nodes_allowed)
2559 {
2560 }
2561 #endif
2562
2563 /*
2564  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
2565  * balanced by operating on them in a round-robin fashion.
2566  * Returns 1 if an adjustment was made.
2567  */
2568 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2569                                 int delta)
2570 {
2571         int nr_nodes, node;
2572
2573         VM_BUG_ON(delta != -1 && delta != 1);
2574
2575         if (delta < 0) {
2576                 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2577                         if (h->surplus_huge_pages_node[node])
2578                                 goto found;
2579                 }
2580         } else {
2581                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2582                         if (h->surplus_huge_pages_node[node] <
2583                                         h->nr_huge_pages_node[node])
2584                                 goto found;
2585                 }
2586         }
2587         return 0;
2588
2589 found:
2590         h->surplus_huge_pages += delta;
2591         h->surplus_huge_pages_node[node] += delta;
2592         return 1;
2593 }
2594
2595 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2596 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2597                               nodemask_t *nodes_allowed)
2598 {
2599         unsigned long min_count, ret;
2600         NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2601
2602         /*
2603          * Bit mask controlling how hard we retry per-node allocations.
2604          * If we can not allocate the bit mask, do not attempt to allocate
2605          * the requested huge pages.
2606          */
2607         if (node_alloc_noretry)
2608                 nodes_clear(*node_alloc_noretry);
2609         else
2610                 return -ENOMEM;
2611
2612         spin_lock(&hugetlb_lock);
2613
2614         /*
2615          * Check for a node specific request.
2616          * Changing node specific huge page count may require a corresponding
2617          * change to the global count.  In any case, the passed node mask
2618          * (nodes_allowed) will restrict alloc/free to the specified node.
2619          */
2620         if (nid != NUMA_NO_NODE) {
2621                 unsigned long old_count = count;
2622
2623                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2624                 /*
2625                  * User may have specified a large count value which caused the
2626                  * above calculation to overflow.  In this case, they wanted
2627                  * to allocate as many huge pages as possible.  Set count to
2628                  * largest possible value to align with their intention.
2629                  */
2630                 if (count < old_count)
2631                         count = ULONG_MAX;
2632         }
2633
2634         /*
2635          * Gigantic pages runtime allocation depend on the capability for large
2636          * page range allocation.
2637          * If the system does not provide this feature, return an error when
2638          * the user tries to allocate gigantic pages but let the user free the
2639          * boottime allocated gigantic pages.
2640          */
2641         if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2642                 if (count > persistent_huge_pages(h)) {
2643                         spin_unlock(&hugetlb_lock);
2644                         NODEMASK_FREE(node_alloc_noretry);
2645                         return -EINVAL;
2646                 }
2647                 /* Fall through to decrease pool */
2648         }
2649
2650         /*
2651          * Increase the pool size
2652          * First take pages out of surplus state.  Then make up the
2653          * remaining difference by allocating fresh huge pages.
2654          *
2655          * We might race with alloc_surplus_huge_page() here and be unable
2656          * to convert a surplus huge page to a normal huge page. That is
2657          * not critical, though, it just means the overall size of the
2658          * pool might be one hugepage larger than it needs to be, but
2659          * within all the constraints specified by the sysctls.
2660          */
2661         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2662                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2663                         break;
2664         }
2665
2666         while (count > persistent_huge_pages(h)) {
2667                 /*
2668                  * If this allocation races such that we no longer need the
2669                  * page, free_huge_page will handle it by freeing the page
2670                  * and reducing the surplus.
2671                  */
2672                 spin_unlock(&hugetlb_lock);
2673
2674                 /* yield cpu to avoid soft lockup */
2675                 cond_resched();
2676
2677                 ret = alloc_pool_huge_page(h, nodes_allowed,
2678                                                 node_alloc_noretry);
2679                 spin_lock(&hugetlb_lock);
2680                 if (!ret)
2681                         goto out;
2682
2683                 /* Bail for signals. Probably ctrl-c from user */
2684                 if (signal_pending(current))
2685                         goto out;
2686         }
2687
2688         /*
2689          * Decrease the pool size
2690          * First return free pages to the buddy allocator (being careful
2691          * to keep enough around to satisfy reservations).  Then place
2692          * pages into surplus state as needed so the pool will shrink
2693          * to the desired size as pages become free.
2694          *
2695          * By placing pages into the surplus state independent of the
2696          * overcommit value, we are allowing the surplus pool size to
2697          * exceed overcommit. There are few sane options here. Since
2698          * alloc_surplus_huge_page() is checking the global counter,
2699          * though, we'll note that we're not allowed to exceed surplus
2700          * and won't grow the pool anywhere else. Not until one of the
2701          * sysctls are changed, or the surplus pages go out of use.
2702          */
2703         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2704         min_count = max(count, min_count);
2705         try_to_free_low(h, min_count, nodes_allowed);
2706         while (min_count < persistent_huge_pages(h)) {
2707                 if (!free_pool_huge_page(h, nodes_allowed, 0))
2708                         break;
2709                 cond_resched_lock(&hugetlb_lock);
2710         }
2711         while (count < persistent_huge_pages(h)) {
2712                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2713                         break;
2714         }
2715 out:
2716         h->max_huge_pages = persistent_huge_pages(h);
2717         spin_unlock(&hugetlb_lock);
2718
2719         NODEMASK_FREE(node_alloc_noretry);
2720
2721         return 0;
2722 }
2723
2724 #define HSTATE_ATTR_RO(_name) \
2725         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2726
2727 #define HSTATE_ATTR(_name) \
2728         static struct kobj_attribute _name##_attr = \
2729                 __ATTR(_name, 0644, _name##_show, _name##_store)
2730
2731 static struct kobject *hugepages_kobj;
2732 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2733
2734 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2735
2736 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2737 {
2738         int i;
2739
2740         for (i = 0; i < HUGE_MAX_HSTATE; i++)
2741                 if (hstate_kobjs[i] == kobj) {
2742                         if (nidp)
2743                                 *nidp = NUMA_NO_NODE;
2744                         return &hstates[i];
2745                 }
2746
2747         return kobj_to_node_hstate(kobj, nidp);
2748 }
2749
2750 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2751                                         struct kobj_attribute *attr, char *buf)
2752 {
2753         struct hstate *h;
2754         unsigned long nr_huge_pages;
2755         int nid;
2756
2757         h = kobj_to_hstate(kobj, &nid);
2758         if (nid == NUMA_NO_NODE)
2759                 nr_huge_pages = h->nr_huge_pages;
2760         else
2761                 nr_huge_pages = h->nr_huge_pages_node[nid];
2762
2763         return sysfs_emit(buf, "%lu\n", nr_huge_pages);
2764 }
2765
2766 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2767                                            struct hstate *h, int nid,
2768                                            unsigned long count, size_t len)
2769 {
2770         int err;
2771         nodemask_t nodes_allowed, *n_mask;
2772
2773         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2774                 return -EINVAL;
2775
2776         if (nid == NUMA_NO_NODE) {
2777                 /*
2778                  * global hstate attribute
2779                  */
2780                 if (!(obey_mempolicy &&
2781                                 init_nodemask_of_mempolicy(&nodes_allowed)))
2782                         n_mask = &node_states[N_MEMORY];
2783                 else
2784                         n_mask = &nodes_allowed;
2785         } else {
2786                 /*
2787                  * Node specific request.  count adjustment happens in
2788                  * set_max_huge_pages() after acquiring hugetlb_lock.
2789                  */
2790                 init_nodemask_of_node(&nodes_allowed, nid);
2791                 n_mask = &nodes_allowed;
2792         }
2793
2794         err = set_max_huge_pages(h, count, nid, n_mask);
2795
2796         return err ? err : len;
2797 }
2798
2799 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2800                                          struct kobject *kobj, const char *buf,
2801                                          size_t len)
2802 {
2803         struct hstate *h;
2804         unsigned long count;
2805         int nid;
2806         int err;
2807
2808         err = kstrtoul(buf, 10, &count);
2809         if (err)
2810                 return err;
2811
2812         h = kobj_to_hstate(kobj, &nid);
2813         return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2814 }
2815
2816 static ssize_t nr_hugepages_show(struct kobject *kobj,
2817                                        struct kobj_attribute *attr, char *buf)
2818 {
2819         return nr_hugepages_show_common(kobj, attr, buf);
2820 }
2821
2822 static ssize_t nr_hugepages_store(struct kobject *kobj,
2823                struct kobj_attribute *attr, const char *buf, size_t len)
2824 {
2825         return nr_hugepages_store_common(false, kobj, buf, len);
2826 }
2827 HSTATE_ATTR(nr_hugepages);
2828
2829 #ifdef CONFIG_NUMA
2830
2831 /*
2832  * hstate attribute for optionally mempolicy-based constraint on persistent
2833  * huge page alloc/free.
2834  */
2835 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2836                                            struct kobj_attribute *attr,
2837                                            char *buf)
2838 {
2839         return nr_hugepages_show_common(kobj, attr, buf);
2840 }
2841
2842 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2843                struct kobj_attribute *attr, const char *buf, size_t len)
2844 {
2845         return nr_hugepages_store_common(true, kobj, buf, len);
2846 }
2847 HSTATE_ATTR(nr_hugepages_mempolicy);
2848 #endif
2849
2850
2851 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2852                                         struct kobj_attribute *attr, char *buf)
2853 {
2854         struct hstate *h = kobj_to_hstate(kobj, NULL);
2855         return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
2856 }
2857
2858 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2859                 struct kobj_attribute *attr, const char *buf, size_t count)
2860 {
2861         int err;
2862         unsigned long input;
2863         struct hstate *h = kobj_to_hstate(kobj, NULL);
2864
2865         if (hstate_is_gigantic(h))
2866                 return -EINVAL;
2867
2868         err = kstrtoul(buf, 10, &input);
2869         if (err)
2870                 return err;
2871
2872         spin_lock(&hugetlb_lock);
2873         h->nr_overcommit_huge_pages = input;
2874         spin_unlock(&hugetlb_lock);
2875
2876         return count;
2877 }
2878 HSTATE_ATTR(nr_overcommit_hugepages);
2879
2880 static ssize_t free_hugepages_show(struct kobject *kobj,
2881                                         struct kobj_attribute *attr, char *buf)
2882 {
2883         struct hstate *h;
2884         unsigned long free_huge_pages;
2885         int nid;
2886
2887         h = kobj_to_hstate(kobj, &nid);
2888         if (nid == NUMA_NO_NODE)
2889                 free_huge_pages = h->free_huge_pages;
2890         else
2891                 free_huge_pages = h->free_huge_pages_node[nid];
2892
2893         return sysfs_emit(buf, "%lu\n", free_huge_pages);
2894 }
2895 HSTATE_ATTR_RO(free_hugepages);
2896
2897 static ssize_t resv_hugepages_show(struct kobject *kobj,
2898                                         struct kobj_attribute *attr, char *buf)
2899 {
2900         struct hstate *h = kobj_to_hstate(kobj, NULL);
2901         return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
2902 }
2903 HSTATE_ATTR_RO(resv_hugepages);
2904
2905 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2906                                         struct kobj_attribute *attr, char *buf)
2907 {
2908         struct hstate *h;
2909         unsigned long surplus_huge_pages;
2910         int nid;
2911
2912         h = kobj_to_hstate(kobj, &nid);
2913         if (nid == NUMA_NO_NODE)
2914                 surplus_huge_pages = h->surplus_huge_pages;
2915         else
2916                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2917
2918         return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
2919 }
2920 HSTATE_ATTR_RO(surplus_hugepages);
2921
2922 static struct attribute *hstate_attrs[] = {
2923         &nr_hugepages_attr.attr,
2924         &nr_overcommit_hugepages_attr.attr,
2925         &free_hugepages_attr.attr,
2926         &resv_hugepages_attr.attr,
2927         &surplus_hugepages_attr.attr,
2928 #ifdef CONFIG_NUMA
2929         &nr_hugepages_mempolicy_attr.attr,
2930 #endif
2931         NULL,
2932 };
2933
2934 static const struct attribute_group hstate_attr_group = {
2935         .attrs = hstate_attrs,
2936 };
2937
2938 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2939                                     struct kobject **hstate_kobjs,
2940                                     const struct attribute_group *hstate_attr_group)
2941 {
2942         int retval;
2943         int hi = hstate_index(h);
2944
2945         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2946         if (!hstate_kobjs[hi])
2947                 return -ENOMEM;
2948
2949         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2950         if (retval)
2951                 kobject_put(hstate_kobjs[hi]);
2952
2953         return retval;
2954 }
2955
2956 static void __init hugetlb_sysfs_init(void)
2957 {
2958         struct hstate *h;
2959         int err;
2960
2961         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2962         if (!hugepages_kobj)
2963                 return;
2964
2965         for_each_hstate(h) {
2966                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2967                                          hstate_kobjs, &hstate_attr_group);
2968                 if (err)
2969                         pr_err("HugeTLB: Unable to add hstate %s", h->name);
2970         }
2971 }
2972
2973 #ifdef CONFIG_NUMA
2974
2975 /*
2976  * node_hstate/s - associate per node hstate attributes, via their kobjects,
2977  * with node devices in node_devices[] using a parallel array.  The array
2978  * index of a node device or _hstate == node id.
2979  * This is here to avoid any static dependency of the node device driver, in
2980  * the base kernel, on the hugetlb module.
2981  */
2982 struct node_hstate {
2983         struct kobject          *hugepages_kobj;
2984         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
2985 };
2986 static struct node_hstate node_hstates[MAX_NUMNODES];
2987
2988 /*
2989  * A subset of global hstate attributes for node devices
2990  */
2991 static struct attribute *per_node_hstate_attrs[] = {
2992         &nr_hugepages_attr.attr,
2993         &free_hugepages_attr.attr,
2994         &surplus_hugepages_attr.attr,
2995         NULL,
2996 };
2997
2998 static const struct attribute_group per_node_hstate_attr_group = {
2999         .attrs = per_node_hstate_attrs,
3000 };
3001
3002 /*
3003  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3004  * Returns node id via non-NULL nidp.
3005  */
3006 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3007 {
3008         int nid;
3009
3010         for (nid = 0; nid < nr_node_ids; nid++) {
3011                 struct node_hstate *nhs = &node_hstates[nid];
3012                 int i;
3013                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3014                         if (nhs->hstate_kobjs[i] == kobj) {
3015                                 if (nidp)
3016                                         *nidp = nid;
3017                                 return &hstates[i];
3018                         }
3019         }
3020
3021         BUG();
3022         return NULL;
3023 }
3024
3025 /*
3026  * Unregister hstate attributes from a single node device.
3027  * No-op if no hstate attributes attached.
3028  */
3029 static void hugetlb_unregister_node(struct node *node)
3030 {
3031         struct hstate *h;
3032         struct node_hstate *nhs = &node_hstates[node->dev.id];
3033
3034         if (!nhs->hugepages_kobj)
3035                 return;         /* no hstate attributes */
3036
3037         for_each_hstate(h) {
3038                 int idx = hstate_index(h);
3039                 if (nhs->hstate_kobjs[idx]) {
3040                         kobject_put(nhs->hstate_kobjs[idx]);
3041                         nhs->hstate_kobjs[idx] = NULL;
3042                 }
3043         }
3044
3045         kobject_put(nhs->hugepages_kobj);
3046         nhs->hugepages_kobj = NULL;
3047 }
3048
3049
3050 /*
3051  * Register hstate attributes for a single node device.
3052  * No-op if attributes already registered.
3053  */
3054 static void hugetlb_register_node(struct node *node)
3055 {
3056         struct hstate *h;
3057         struct node_hstate *nhs = &node_hstates[node->dev.id];
3058         int err;
3059
3060         if (nhs->hugepages_kobj)
3061                 return;         /* already allocated */
3062
3063         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3064                                                         &node->dev.kobj);
3065         if (!nhs->hugepages_kobj)
3066                 return;
3067
3068         for_each_hstate(h) {
3069                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3070                                                 nhs->hstate_kobjs,
3071                                                 &per_node_hstate_attr_group);
3072                 if (err) {
3073                         pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3074                                 h->name, node->dev.id);
3075                         hugetlb_unregister_node(node);
3076                         break;
3077                 }
3078         }
3079 }
3080
3081 /*
3082  * hugetlb init time:  register hstate attributes for all registered node
3083  * devices of nodes that have memory.  All on-line nodes should have
3084  * registered their associated device by this time.
3085  */
3086 static void __init hugetlb_register_all_nodes(void)
3087 {
3088         int nid;
3089
3090         for_each_node_state(nid, N_MEMORY) {
3091                 struct node *node = node_devices[nid];
3092                 if (node->dev.id == nid)
3093                         hugetlb_register_node(node);
3094         }
3095
3096         /*
3097          * Let the node device driver know we're here so it can
3098          * [un]register hstate attributes on node hotplug.
3099          */
3100         register_hugetlbfs_with_node(hugetlb_register_node,
3101                                      hugetlb_unregister_node);
3102 }
3103 #else   /* !CONFIG_NUMA */
3104
3105 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3106 {
3107         BUG();
3108         if (nidp)
3109                 *nidp = -1;
3110         return NULL;
3111 }
3112
3113 static void hugetlb_register_all_nodes(void) { }
3114
3115 #endif
3116
3117 static int __init hugetlb_init(void)
3118 {
3119         int i;
3120
3121         if (!hugepages_supported()) {
3122                 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3123                         pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3124                 return 0;
3125         }
3126
3127         /*
3128          * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists.  Some
3129          * architectures depend on setup being done here.
3130          */
3131         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3132         if (!parsed_default_hugepagesz) {
3133                 /*
3134                  * If we did not parse a default huge page size, set
3135                  * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3136                  * number of huge pages for this default size was implicitly
3137                  * specified, set that here as well.
3138                  * Note that the implicit setting will overwrite an explicit
3139                  * setting.  A warning will be printed in this case.
3140                  */
3141                 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3142                 if (default_hstate_max_huge_pages) {
3143                         if (default_hstate.max_huge_pages) {
3144                                 char buf[32];
3145
3146                                 string_get_size(huge_page_size(&default_hstate),
3147                                         1, STRING_UNITS_2, buf, 32);
3148                                 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3149                                         default_hstate.max_huge_pages, buf);
3150                                 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3151                                         default_hstate_max_huge_pages);
3152                         }
3153                         default_hstate.max_huge_pages =
3154                                 default_hstate_max_huge_pages;
3155                 }
3156         }
3157
3158         hugetlb_cma_check();
3159         hugetlb_init_hstates();
3160         gather_bootmem_prealloc();
3161         report_hugepages();
3162
3163         hugetlb_sysfs_init();
3164         hugetlb_register_all_nodes();
3165         hugetlb_cgroup_file_init();
3166
3167 #ifdef CONFIG_SMP
3168         num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3169 #else
3170         num_fault_mutexes = 1;
3171 #endif
3172         hugetlb_fault_mutex_table =
3173                 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3174                               GFP_KERNEL);
3175         BUG_ON(!hugetlb_fault_mutex_table);
3176
3177         for (i = 0; i < num_fault_mutexes; i++)
3178                 mutex_init(&hugetlb_fault_mutex_table[i]);
3179         return 0;
3180 }
3181 subsys_initcall(hugetlb_init);
3182
3183 /* Overwritten by architectures with more huge page sizes */
3184 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3185 {
3186         return size == HPAGE_SIZE;
3187 }
3188
3189 void __init hugetlb_add_hstate(unsigned int order)
3190 {
3191         struct hstate *h;
3192         unsigned long i;
3193
3194         if (size_to_hstate(PAGE_SIZE << order)) {
3195                 return;
3196         }
3197         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3198         BUG_ON(order == 0);
3199         h = &hstates[hugetlb_max_hstate++];
3200         h->order = order;
3201         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
3202         for (i = 0; i < MAX_NUMNODES; ++i)
3203                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3204         INIT_LIST_HEAD(&h->hugepage_activelist);
3205         h->next_nid_to_alloc = first_memory_node;
3206         h->next_nid_to_free = first_memory_node;
3207         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3208                                         huge_page_size(h)/1024);
3209
3210         parsed_hstate = h;
3211 }
3212
3213 /*
3214  * hugepages command line processing
3215  * hugepages normally follows a valid hugepagsz or default_hugepagsz
3216  * specification.  If not, ignore the hugepages value.  hugepages can also
3217  * be the first huge page command line  option in which case it implicitly
3218  * specifies the number of huge pages for the default size.
3219  */
3220 static int __init hugepages_setup(char *s)
3221 {
3222         unsigned long *mhp;
3223         static unsigned long *last_mhp;
3224
3225         if (!parsed_valid_hugepagesz) {
3226                 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3227                 parsed_valid_hugepagesz = true;
3228                 return 0;
3229         }
3230
3231         /*
3232          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3233          * yet, so this hugepages= parameter goes to the "default hstate".
3234          * Otherwise, it goes with the previously parsed hugepagesz or
3235          * default_hugepagesz.
3236          */
3237         else if (!hugetlb_max_hstate)
3238                 mhp = &default_hstate_max_huge_pages;
3239         else
3240                 mhp = &parsed_hstate->max_huge_pages;
3241
3242         if (mhp == last_mhp) {
3243                 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3244                 return 0;
3245         }
3246
3247         if (sscanf(s, "%lu", mhp) <= 0)
3248                 *mhp = 0;
3249
3250         /*
3251          * Global state is always initialized later in hugetlb_init.
3252          * But we need to allocate >= MAX_ORDER hstates here early to still
3253          * use the bootmem allocator.
3254          */
3255         if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
3256                 hugetlb_hstate_alloc_pages(parsed_hstate);
3257
3258         last_mhp = mhp;
3259
3260         return 1;
3261 }
3262 __setup("hugepages=", hugepages_setup);
3263
3264 /*
3265  * hugepagesz command line processing
3266  * A specific huge page size can only be specified once with hugepagesz.
3267  * hugepagesz is followed by hugepages on the command line.  The global
3268  * variable 'parsed_valid_hugepagesz' is used to determine if prior
3269  * hugepagesz argument was valid.
3270  */
3271 static int __init hugepagesz_setup(char *s)
3272 {
3273         unsigned long size;
3274         struct hstate *h;
3275
3276         parsed_valid_hugepagesz = false;
3277         size = (unsigned long)memparse(s, NULL);
3278
3279         if (!arch_hugetlb_valid_size(size)) {
3280                 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3281                 return 0;
3282         }
3283
3284         h = size_to_hstate(size);
3285         if (h) {
3286                 /*
3287                  * hstate for this size already exists.  This is normally
3288                  * an error, but is allowed if the existing hstate is the
3289                  * default hstate.  More specifically, it is only allowed if
3290                  * the number of huge pages for the default hstate was not
3291                  * previously specified.
3292                  */
3293                 if (!parsed_default_hugepagesz ||  h != &default_hstate ||
3294                     default_hstate.max_huge_pages) {
3295                         pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3296                         return 0;
3297                 }
3298
3299                 /*
3300                  * No need to call hugetlb_add_hstate() as hstate already
3301                  * exists.  But, do set parsed_hstate so that a following
3302                  * hugepages= parameter will be applied to this hstate.
3303                  */
3304                 parsed_hstate = h;
3305                 parsed_valid_hugepagesz = true;
3306                 return 1;
3307         }
3308
3309         hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3310         parsed_valid_hugepagesz = true;
3311         return 1;
3312 }
3313 __setup("hugepagesz=", hugepagesz_setup);
3314
3315 /*
3316  * default_hugepagesz command line input
3317  * Only one instance of default_hugepagesz allowed on command line.
3318  */
3319 static int __init default_hugepagesz_setup(char *s)
3320 {
3321         unsigned long size;
3322
3323         parsed_valid_hugepagesz = false;
3324         if (parsed_default_hugepagesz) {
3325                 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3326                 return 0;
3327         }
3328
3329         size = (unsigned long)memparse(s, NULL);
3330
3331         if (!arch_hugetlb_valid_size(size)) {
3332                 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3333                 return 0;
3334         }
3335
3336         hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3337         parsed_valid_hugepagesz = true;
3338         parsed_default_hugepagesz = true;
3339         default_hstate_idx = hstate_index(size_to_hstate(size));
3340
3341         /*
3342          * The number of default huge pages (for this size) could have been
3343          * specified as the first hugetlb parameter: hugepages=X.  If so,
3344          * then default_hstate_max_huge_pages is set.  If the default huge
3345          * page size is gigantic (>= MAX_ORDER), then the pages must be
3346          * allocated here from bootmem allocator.
3347          */
3348         if (default_hstate_max_huge_pages) {
3349                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3350                 if (hstate_is_gigantic(&default_hstate))
3351                         hugetlb_hstate_alloc_pages(&default_hstate);
3352                 default_hstate_max_huge_pages = 0;
3353         }
3354
3355         return 1;
3356 }
3357 __setup("default_hugepagesz=", default_hugepagesz_setup);
3358
3359 static unsigned int allowed_mems_nr(struct hstate *h)
3360 {
3361         int node;
3362         unsigned int nr = 0;
3363         nodemask_t *mpol_allowed;
3364         unsigned int *array = h->free_huge_pages_node;
3365         gfp_t gfp_mask = htlb_alloc_mask(h);
3366
3367         mpol_allowed = policy_nodemask_current(gfp_mask);
3368
3369         for_each_node_mask(node, cpuset_current_mems_allowed) {
3370                 if (!mpol_allowed ||
3371                     (mpol_allowed && node_isset(node, *mpol_allowed)))
3372                         nr += array[node];
3373         }
3374
3375         return nr;
3376 }
3377
3378 #ifdef CONFIG_SYSCTL
3379 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3380                                           void *buffer, size_t *length,
3381                                           loff_t *ppos, unsigned long *out)
3382 {
3383         struct ctl_table dup_table;
3384
3385         /*
3386          * In order to avoid races with __do_proc_doulongvec_minmax(), we
3387          * can duplicate the @table and alter the duplicate of it.
3388          */
3389         dup_table = *table;
3390         dup_table.data = out;
3391
3392         return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3393 }
3394
3395 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3396                          struct ctl_table *table, int write,
3397                          void *buffer, size_t *length, loff_t *ppos)
3398 {
3399         struct hstate *h = &default_hstate;
3400         unsigned long tmp = h->max_huge_pages;
3401         int ret;
3402
3403         if (!hugepages_supported())
3404                 return -EOPNOTSUPP;
3405
3406         ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3407                                              &tmp);
3408         if (ret)
3409                 goto out;
3410
3411         if (write)
3412                 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3413                                                   NUMA_NO_NODE, tmp, *length);
3414 out:
3415         return ret;
3416 }
3417
3418 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3419                           void *buffer, size_t *length, loff_t *ppos)
3420 {
3421
3422         return hugetlb_sysctl_handler_common(false, table, write,
3423                                                         buffer, length, ppos);
3424 }
3425
3426 #ifdef CONFIG_NUMA
3427 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3428                           void *buffer, size_t *length, loff_t *ppos)
3429 {
3430         return hugetlb_sysctl_handler_common(true, table, write,
3431                                                         buffer, length, ppos);
3432 }
3433 #endif /* CONFIG_NUMA */
3434
3435 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3436                 void *buffer, size_t *length, loff_t *ppos)
3437 {
3438         struct hstate *h = &default_hstate;
3439         unsigned long tmp;
3440         int ret;
3441
3442         if (!hugepages_supported())
3443                 return -EOPNOTSUPP;
3444
3445         tmp = h->nr_overcommit_huge_pages;
3446
3447         if (write && hstate_is_gigantic(h))
3448                 return -EINVAL;
3449
3450         ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3451                                              &tmp);
3452         if (ret)
3453                 goto out;
3454
3455         if (write) {
3456                 spin_lock(&hugetlb_lock);
3457                 h->nr_overcommit_huge_pages = tmp;
3458                 spin_unlock(&hugetlb_lock);
3459         }
3460 out:
3461         return ret;
3462 }
3463
3464 #endif /* CONFIG_SYSCTL */
3465
3466 void hugetlb_report_meminfo(struct seq_file *m)
3467 {
3468         struct hstate *h;
3469         unsigned long total = 0;
3470
3471         if (!hugepages_supported())
3472                 return;
3473
3474         for_each_hstate(h) {
3475                 unsigned long count = h->nr_huge_pages;
3476
3477                 total += (PAGE_SIZE << huge_page_order(h)) * count;
3478
3479                 if (h == &default_hstate)
3480                         seq_printf(m,
3481                                    "HugePages_Total:   %5lu\n"
3482                                    "HugePages_Free:    %5lu\n"
3483                                    "HugePages_Rsvd:    %5lu\n"
3484                                    "HugePages_Surp:    %5lu\n"
3485                                    "Hugepagesize:   %8lu kB\n",
3486                                    count,
3487                                    h->free_huge_pages,
3488                                    h->resv_huge_pages,
3489                                    h->surplus_huge_pages,
3490                                    (PAGE_SIZE << huge_page_order(h)) / 1024);
3491         }
3492
3493         seq_printf(m, "Hugetlb:        %8lu kB\n", total / 1024);
3494 }
3495
3496 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
3497 {
3498         struct hstate *h = &default_hstate;
3499
3500         if (!hugepages_supported())
3501                 return 0;
3502
3503         return sysfs_emit_at(buf, len,
3504                              "Node %d HugePages_Total: %5u\n"
3505                              "Node %d HugePages_Free:  %5u\n"
3506                              "Node %d HugePages_Surp:  %5u\n",
3507                              nid, h->nr_huge_pages_node[nid],
3508                              nid, h->free_huge_pages_node[nid],
3509                              nid, h->surplus_huge_pages_node[nid]);
3510 }
3511
3512 void hugetlb_show_meminfo(void)
3513 {
3514         struct hstate *h;
3515         int nid;
3516
3517         if (!hugepages_supported())
3518                 return;
3519
3520         for_each_node_state(nid, N_MEMORY)
3521                 for_each_hstate(h)
3522                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3523                                 nid,
3524                                 h->nr_huge_pages_node[nid],
3525                                 h->free_huge_pages_node[nid],
3526                                 h->surplus_huge_pages_node[nid],
3527                                 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3528 }
3529
3530 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3531 {
3532         seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3533                    atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3534 }
3535
3536 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3537 unsigned long hugetlb_total_pages(void)
3538 {
3539         struct hstate *h;
3540         unsigned long nr_total_pages = 0;
3541
3542         for_each_hstate(h)
3543                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3544         return nr_total_pages;
3545 }
3546
3547 static int hugetlb_acct_memory(struct hstate *h, long delta)
3548 {
3549         int ret = -ENOMEM;
3550
3551         spin_lock(&hugetlb_lock);
3552         /*
3553          * When cpuset is configured, it breaks the strict hugetlb page
3554          * reservation as the accounting is done on a global variable. Such
3555          * reservation is completely rubbish in the presence of cpuset because
3556          * the reservation is not checked against page availability for the
3557          * current cpuset. Application can still potentially OOM'ed by kernel
3558          * with lack of free htlb page in cpuset that the task is in.
3559          * Attempt to enforce strict accounting with cpuset is almost
3560          * impossible (or too ugly) because cpuset is too fluid that
3561          * task or memory node can be dynamically moved between cpusets.
3562          *
3563          * The change of semantics for shared hugetlb mapping with cpuset is
3564          * undesirable. However, in order to preserve some of the semantics,
3565          * we fall back to check against current free page availability as
3566          * a best attempt and hopefully to minimize the impact of changing
3567          * semantics that cpuset has.
3568          *
3569          * Apart from cpuset, we also have memory policy mechanism that
3570          * also determines from which node the kernel will allocate memory
3571          * in a NUMA system. So similar to cpuset, we also should consider
3572          * the memory policy of the current task. Similar to the description
3573          * above.
3574          */
3575         if (delta > 0) {
3576                 if (gather_surplus_pages(h, delta) < 0)
3577                         goto out;
3578
3579                 if (delta > allowed_mems_nr(h)) {
3580                         return_unused_surplus_pages(h, delta);
3581                         goto out;
3582                 }
3583         }
3584
3585         ret = 0;
3586         if (delta < 0)
3587                 return_unused_surplus_pages(h, (unsigned long) -delta);
3588
3589 out:
3590         spin_unlock(&hugetlb_lock);
3591         return ret;
3592 }
3593
3594 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3595 {
3596         struct resv_map *resv = vma_resv_map(vma);
3597
3598         /*
3599          * This new VMA should share its siblings reservation map if present.
3600          * The VMA will only ever have a valid reservation map pointer where
3601          * it is being copied for another still existing VMA.  As that VMA
3602          * has a reference to the reservation map it cannot disappear until
3603          * after this open call completes.  It is therefore safe to take a
3604          * new reference here without additional locking.
3605          */
3606         if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3607                 kref_get(&resv->refs);
3608 }
3609
3610 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3611 {
3612         struct hstate *h = hstate_vma(vma);
3613         struct resv_map *resv = vma_resv_map(vma);
3614         struct hugepage_subpool *spool = subpool_vma(vma);
3615         unsigned long reserve, start, end;
3616         long gbl_reserve;
3617
3618         if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3619                 return;
3620
3621         start = vma_hugecache_offset(h, vma, vma->vm_start);
3622         end = vma_hugecache_offset(h, vma, vma->vm_end);
3623
3624         reserve = (end - start) - region_count(resv, start, end);
3625         hugetlb_cgroup_uncharge_counter(resv, start, end);
3626         if (reserve) {
3627                 /*
3628                  * Decrement reserve counts.  The global reserve count may be
3629                  * adjusted if the subpool has a minimum size.
3630                  */
3631                 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3632                 hugetlb_acct_memory(h, -gbl_reserve);
3633         }
3634
3635         kref_put(&resv->refs, resv_map_release);
3636 }
3637
3638 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3639 {
3640         if (addr & ~(huge_page_mask(hstate_vma(vma))))
3641                 return -EINVAL;
3642         return 0;
3643 }
3644
3645 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3646 {
3647         struct hstate *hstate = hstate_vma(vma);
3648
3649         return 1UL << huge_page_shift(hstate);
3650 }
3651
3652 /*
3653  * We cannot handle pagefaults against hugetlb pages at all.  They cause
3654  * handle_mm_fault() to try to instantiate regular-sized pages in the
3655  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
3656  * this far.
3657  */
3658 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3659 {
3660         BUG();
3661         return 0;
3662 }
3663
3664 /*
3665  * When a new function is introduced to vm_operations_struct and added
3666  * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3667  * This is because under System V memory model, mappings created via
3668  * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3669  * their original vm_ops are overwritten with shm_vm_ops.
3670  */
3671 const struct vm_operations_struct hugetlb_vm_ops = {
3672         .fault = hugetlb_vm_op_fault,
3673         .open = hugetlb_vm_op_open,
3674         .close = hugetlb_vm_op_close,
3675         .may_split = hugetlb_vm_op_split,
3676         .pagesize = hugetlb_vm_op_pagesize,
3677 };
3678
3679 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3680                                 int writable)
3681 {
3682         pte_t entry;
3683
3684         if (writable) {
3685                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3686                                          vma->vm_page_prot)));
3687         } else {
3688                 entry = huge_pte_wrprotect(mk_huge_pte(page,
3689                                            vma->vm_page_prot));
3690         }
3691         entry = pte_mkyoung(entry);
3692         entry = pte_mkhuge(entry);
3693         entry = arch_make_huge_pte(entry, vma, page, writable);
3694
3695         return entry;
3696 }
3697
3698 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3699                                    unsigned long address, pte_t *ptep)
3700 {
3701         pte_t entry;
3702
3703         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3704         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3705                 update_mmu_cache(vma, address, ptep);
3706 }
3707
3708 bool is_hugetlb_entry_migration(pte_t pte)
3709 {
3710         swp_entry_t swp;
3711
3712         if (huge_pte_none(pte) || pte_present(pte))
3713                 return false;
3714         swp = pte_to_swp_entry(pte);
3715         if (is_migration_entry(swp))
3716                 return true;
3717         else
3718                 return false;
3719 }
3720
3721 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
3722 {
3723         swp_entry_t swp;
3724
3725         if (huge_pte_none(pte) || pte_present(pte))
3726                 return false;
3727         swp = pte_to_swp_entry(pte);
3728         if (is_hwpoison_entry(swp))
3729                 return true;
3730         else
3731                 return false;
3732 }
3733
3734 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3735                             struct vm_area_struct *vma)
3736 {
3737         pte_t *src_pte, *dst_pte, entry, dst_entry;
3738         struct page *ptepage;
3739         unsigned long addr;
3740         int cow;
3741         struct hstate *h = hstate_vma(vma);
3742         unsigned long sz = huge_page_size(h);
3743         struct address_space *mapping = vma->vm_file->f_mapping;
3744         struct mmu_notifier_range range;
3745         int ret = 0;
3746
3747         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3748
3749         if (cow) {
3750                 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3751                                         vma->vm_start,
3752                                         vma->vm_end);
3753                 mmu_notifier_invalidate_range_start(&range);
3754         } else {
3755                 /*
3756                  * For shared mappings i_mmap_rwsem must be held to call
3757                  * huge_pte_alloc, otherwise the returned ptep could go
3758                  * away if part of a shared pmd and another thread calls
3759                  * huge_pmd_unshare.
3760                  */
3761                 i_mmap_lock_read(mapping);
3762         }
3763
3764         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3765                 spinlock_t *src_ptl, *dst_ptl;
3766                 src_pte = huge_pte_offset(src, addr, sz);
3767                 if (!src_pte)
3768                         continue;
3769                 dst_pte = huge_pte_alloc(dst, addr, sz);
3770                 if (!dst_pte) {
3771                         ret = -ENOMEM;
3772                         break;
3773                 }
3774
3775                 /*
3776                  * If the pagetables are shared don't copy or take references.
3777                  * dst_pte == src_pte is the common case of src/dest sharing.
3778                  *
3779                  * However, src could have 'unshared' and dst shares with
3780                  * another vma.  If dst_pte !none, this implies sharing.
3781                  * Check here before taking page table lock, and once again
3782                  * after taking the lock below.
3783                  */
3784                 dst_entry = huge_ptep_get(dst_pte);
3785                 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3786                         continue;
3787
3788                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3789                 src_ptl = huge_pte_lockptr(h, src, src_pte);
3790                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3791                 entry = huge_ptep_get(src_pte);
3792                 dst_entry = huge_ptep_get(dst_pte);
3793                 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3794                         /*
3795                          * Skip if src entry none.  Also, skip in the
3796                          * unlikely case dst entry !none as this implies
3797                          * sharing with another vma.
3798                          */
3799                         ;
3800                 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3801                                     is_hugetlb_entry_hwpoisoned(entry))) {
3802                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
3803
3804                         if (is_write_migration_entry(swp_entry) && cow) {
3805                                 /*
3806                                  * COW mappings require pages in both
3807                                  * parent and child to be set to read.
3808                                  */
3809                                 make_migration_entry_read(&swp_entry);
3810                                 entry = swp_entry_to_pte(swp_entry);
3811                                 set_huge_swap_pte_at(src, addr, src_pte,
3812                                                      entry, sz);
3813                         }
3814                         set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3815                 } else {
3816                         if (cow) {
3817                                 /*
3818                                  * No need to notify as we are downgrading page
3819                                  * table protection not changing it to point
3820                                  * to a new page.
3821                                  *
3822                                  * See Documentation/vm/mmu_notifier.rst
3823                                  */
3824                                 huge_ptep_set_wrprotect(src, addr, src_pte);
3825                         }
3826                         entry = huge_ptep_get(src_pte);
3827                         ptepage = pte_page(entry);
3828                         get_page(ptepage);
3829                         page_dup_rmap(ptepage, true);
3830                         set_huge_pte_at(dst, addr, dst_pte, entry);
3831                         hugetlb_count_add(pages_per_huge_page(h), dst);
3832                 }
3833                 spin_unlock(src_ptl);
3834                 spin_unlock(dst_ptl);
3835         }
3836
3837         if (cow)
3838                 mmu_notifier_invalidate_range_end(&range);
3839         else
3840                 i_mmap_unlock_read(mapping);
3841
3842         return ret;
3843 }
3844
3845 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3846                             unsigned long start, unsigned long end,
3847                             struct page *ref_page)
3848 {
3849         struct mm_struct *mm = vma->vm_mm;
3850         unsigned long address;
3851         pte_t *ptep;
3852         pte_t pte;
3853         spinlock_t *ptl;
3854         struct page *page;
3855         struct hstate *h = hstate_vma(vma);
3856         unsigned long sz = huge_page_size(h);
3857         struct mmu_notifier_range range;
3858
3859         WARN_ON(!is_vm_hugetlb_page(vma));
3860         BUG_ON(start & ~huge_page_mask(h));
3861         BUG_ON(end & ~huge_page_mask(h));
3862
3863         /*
3864          * This is a hugetlb vma, all the pte entries should point
3865          * to huge page.
3866          */
3867         tlb_change_page_size(tlb, sz);
3868         tlb_start_vma(tlb, vma);
3869
3870         /*
3871          * If sharing possible, alert mmu notifiers of worst case.
3872          */
3873         mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3874                                 end);
3875         adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3876         mmu_notifier_invalidate_range_start(&range);
3877         address = start;
3878         for (; address < end; address += sz) {
3879                 ptep = huge_pte_offset(mm, address, sz);
3880                 if (!ptep)
3881                         continue;
3882
3883                 ptl = huge_pte_lock(h, mm, ptep);
3884                 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
3885                         spin_unlock(ptl);
3886                         /*
3887                          * We just unmapped a page of PMDs by clearing a PUD.
3888                          * The caller's TLB flush range should cover this area.
3889                          */
3890                         continue;
3891                 }
3892
3893                 pte = huge_ptep_get(ptep);
3894                 if (huge_pte_none(pte)) {
3895                         spin_unlock(ptl);
3896                         continue;
3897                 }
3898
3899                 /*
3900                  * Migrating hugepage or HWPoisoned hugepage is already
3901                  * unmapped and its refcount is dropped, so just clear pte here.
3902                  */
3903                 if (unlikely(!pte_present(pte))) {
3904                         huge_pte_clear(mm, address, ptep, sz);
3905                         spin_unlock(ptl);
3906                         continue;
3907                 }
3908
3909                 page = pte_page(pte);
3910                 /*
3911                  * If a reference page is supplied, it is because a specific
3912                  * page is being unmapped, not a range. Ensure the page we
3913                  * are about to unmap is the actual page of interest.
3914                  */
3915                 if (ref_page) {
3916                         if (page != ref_page) {
3917                                 spin_unlock(ptl);
3918                                 continue;
3919                         }
3920                         /*
3921                          * Mark the VMA as having unmapped its page so that
3922                          * future faults in this VMA will fail rather than
3923                          * looking like data was lost
3924                          */
3925                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3926                 }
3927
3928                 pte = huge_ptep_get_and_clear(mm, address, ptep);
3929                 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3930                 if (huge_pte_dirty(pte))
3931                         set_page_dirty(page);
3932
3933                 hugetlb_count_sub(pages_per_huge_page(h), mm);
3934                 page_remove_rmap(page, true);
3935
3936                 spin_unlock(ptl);
3937                 tlb_remove_page_size(tlb, page, huge_page_size(h));
3938                 /*
3939                  * Bail out after unmapping reference page if supplied
3940                  */
3941                 if (ref_page)
3942                         break;
3943         }
3944         mmu_notifier_invalidate_range_end(&range);
3945         tlb_end_vma(tlb, vma);
3946 }
3947
3948 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3949                           struct vm_area_struct *vma, unsigned long start,
3950                           unsigned long end, struct page *ref_page)
3951 {
3952         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3953
3954         /*
3955          * Clear this flag so that x86's huge_pmd_share page_table_shareable
3956          * test will fail on a vma being torn down, and not grab a page table
3957          * on its way out.  We're lucky that the flag has such an appropriate
3958          * name, and can in fact be safely cleared here. We could clear it
3959          * before the __unmap_hugepage_range above, but all that's necessary
3960          * is to clear it before releasing the i_mmap_rwsem. This works
3961          * because in the context this is called, the VMA is about to be
3962          * destroyed and the i_mmap_rwsem is held.
3963          */
3964         vma->vm_flags &= ~VM_MAYSHARE;
3965 }