Merge tag 'audit-pr-20211019' of git://git.kernel.org/pub/scm/linux/kernel/git/pcmoor...
[platform/kernel/linux-starfive.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 #include <linux/migrate.h>
34
35 #include <asm/page.h>
36 #include <asm/pgalloc.h>
37 #include <asm/tlb.h>
38
39 #include <linux/io.h>
40 #include <linux/hugetlb.h>
41 #include <linux/hugetlb_cgroup.h>
42 #include <linux/node.h>
43 #include <linux/page_owner.h>
44 #include "internal.h"
45 #include "hugetlb_vmemmap.h"
46
47 int hugetlb_max_hstate __read_mostly;
48 unsigned int default_hstate_idx;
49 struct hstate hstates[HUGE_MAX_HSTATE];
50
51 #ifdef CONFIG_CMA
52 static struct cma *hugetlb_cma[MAX_NUMNODES];
53 #endif
54 static unsigned long hugetlb_cma_size __initdata;
55
56 /*
57  * Minimum page order among possible hugepage sizes, set to a proper value
58  * at boot time.
59  */
60 static unsigned int minimum_order __read_mostly = UINT_MAX;
61
62 __initdata LIST_HEAD(huge_boot_pages);
63
64 /* for command line parsing */
65 static struct hstate * __initdata parsed_hstate;
66 static unsigned long __initdata default_hstate_max_huge_pages;
67 static bool __initdata parsed_valid_hugepagesz = true;
68 static bool __initdata parsed_default_hugepagesz;
69
70 /*
71  * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
72  * free_huge_pages, and surplus_huge_pages.
73  */
74 DEFINE_SPINLOCK(hugetlb_lock);
75
76 /*
77  * Serializes faults on the same logical page.  This is used to
78  * prevent spurious OOMs when the hugepage pool is fully utilized.
79  */
80 static int num_fault_mutexes;
81 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
82
83 /* Forward declaration */
84 static int hugetlb_acct_memory(struct hstate *h, long delta);
85
86 static inline bool subpool_is_free(struct hugepage_subpool *spool)
87 {
88         if (spool->count)
89                 return false;
90         if (spool->max_hpages != -1)
91                 return spool->used_hpages == 0;
92         if (spool->min_hpages != -1)
93                 return spool->rsv_hpages == spool->min_hpages;
94
95         return true;
96 }
97
98 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
99                                                 unsigned long irq_flags)
100 {
101         spin_unlock_irqrestore(&spool->lock, irq_flags);
102
103         /* If no pages are used, and no other handles to the subpool
104          * remain, give up any reservations based on minimum size and
105          * free the subpool */
106         if (subpool_is_free(spool)) {
107                 if (spool->min_hpages != -1)
108                         hugetlb_acct_memory(spool->hstate,
109                                                 -spool->min_hpages);
110                 kfree(spool);
111         }
112 }
113
114 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
115                                                 long min_hpages)
116 {
117         struct hugepage_subpool *spool;
118
119         spool = kzalloc(sizeof(*spool), GFP_KERNEL);
120         if (!spool)
121                 return NULL;
122
123         spin_lock_init(&spool->lock);
124         spool->count = 1;
125         spool->max_hpages = max_hpages;
126         spool->hstate = h;
127         spool->min_hpages = min_hpages;
128
129         if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
130                 kfree(spool);
131                 return NULL;
132         }
133         spool->rsv_hpages = min_hpages;
134
135         return spool;
136 }
137
138 void hugepage_put_subpool(struct hugepage_subpool *spool)
139 {
140         unsigned long flags;
141
142         spin_lock_irqsave(&spool->lock, flags);
143         BUG_ON(!spool->count);
144         spool->count--;
145         unlock_or_release_subpool(spool, flags);
146 }
147
148 /*
149  * Subpool accounting for allocating and reserving pages.
150  * Return -ENOMEM if there are not enough resources to satisfy the
151  * request.  Otherwise, return the number of pages by which the
152  * global pools must be adjusted (upward).  The returned value may
153  * only be different than the passed value (delta) in the case where
154  * a subpool minimum size must be maintained.
155  */
156 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
157                                       long delta)
158 {
159         long ret = delta;
160
161         if (!spool)
162                 return ret;
163
164         spin_lock_irq(&spool->lock);
165
166         if (spool->max_hpages != -1) {          /* maximum size accounting */
167                 if ((spool->used_hpages + delta) <= spool->max_hpages)
168                         spool->used_hpages += delta;
169                 else {
170                         ret = -ENOMEM;
171                         goto unlock_ret;
172                 }
173         }
174
175         /* minimum size accounting */
176         if (spool->min_hpages != -1 && spool->rsv_hpages) {
177                 if (delta > spool->rsv_hpages) {
178                         /*
179                          * Asking for more reserves than those already taken on
180                          * behalf of subpool.  Return difference.
181                          */
182                         ret = delta - spool->rsv_hpages;
183                         spool->rsv_hpages = 0;
184                 } else {
185                         ret = 0;        /* reserves already accounted for */
186                         spool->rsv_hpages -= delta;
187                 }
188         }
189
190 unlock_ret:
191         spin_unlock_irq(&spool->lock);
192         return ret;
193 }
194
195 /*
196  * Subpool accounting for freeing and unreserving pages.
197  * Return the number of global page reservations that must be dropped.
198  * The return value may only be different than the passed value (delta)
199  * in the case where a subpool minimum size must be maintained.
200  */
201 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
202                                        long delta)
203 {
204         long ret = delta;
205         unsigned long flags;
206
207         if (!spool)
208                 return delta;
209
210         spin_lock_irqsave(&spool->lock, flags);
211
212         if (spool->max_hpages != -1)            /* maximum size accounting */
213                 spool->used_hpages -= delta;
214
215          /* minimum size accounting */
216         if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
217                 if (spool->rsv_hpages + delta <= spool->min_hpages)
218                         ret = 0;
219                 else
220                         ret = spool->rsv_hpages + delta - spool->min_hpages;
221
222                 spool->rsv_hpages += delta;
223                 if (spool->rsv_hpages > spool->min_hpages)
224                         spool->rsv_hpages = spool->min_hpages;
225         }
226
227         /*
228          * If hugetlbfs_put_super couldn't free spool due to an outstanding
229          * quota reference, free it now.
230          */
231         unlock_or_release_subpool(spool, flags);
232
233         return ret;
234 }
235
236 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
237 {
238         return HUGETLBFS_SB(inode->i_sb)->spool;
239 }
240
241 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
242 {
243         return subpool_inode(file_inode(vma->vm_file));
244 }
245
246 /* Helper that removes a struct file_region from the resv_map cache and returns
247  * it for use.
248  */
249 static struct file_region *
250 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
251 {
252         struct file_region *nrg = NULL;
253
254         VM_BUG_ON(resv->region_cache_count <= 0);
255
256         resv->region_cache_count--;
257         nrg = list_first_entry(&resv->region_cache, struct file_region, link);
258         list_del(&nrg->link);
259
260         nrg->from = from;
261         nrg->to = to;
262
263         return nrg;
264 }
265
266 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
267                                               struct file_region *rg)
268 {
269 #ifdef CONFIG_CGROUP_HUGETLB
270         nrg->reservation_counter = rg->reservation_counter;
271         nrg->css = rg->css;
272         if (rg->css)
273                 css_get(rg->css);
274 #endif
275 }
276
277 /* Helper that records hugetlb_cgroup uncharge info. */
278 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
279                                                 struct hstate *h,
280                                                 struct resv_map *resv,
281                                                 struct file_region *nrg)
282 {
283 #ifdef CONFIG_CGROUP_HUGETLB
284         if (h_cg) {
285                 nrg->reservation_counter =
286                         &h_cg->rsvd_hugepage[hstate_index(h)];
287                 nrg->css = &h_cg->css;
288                 /*
289                  * The caller will hold exactly one h_cg->css reference for the
290                  * whole contiguous reservation region. But this area might be
291                  * scattered when there are already some file_regions reside in
292                  * it. As a result, many file_regions may share only one css
293                  * reference. In order to ensure that one file_region must hold
294                  * exactly one h_cg->css reference, we should do css_get for
295                  * each file_region and leave the reference held by caller
296                  * untouched.
297                  */
298                 css_get(&h_cg->css);
299                 if (!resv->pages_per_hpage)
300                         resv->pages_per_hpage = pages_per_huge_page(h);
301                 /* pages_per_hpage should be the same for all entries in
302                  * a resv_map.
303                  */
304                 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
305         } else {
306                 nrg->reservation_counter = NULL;
307                 nrg->css = NULL;
308         }
309 #endif
310 }
311
312 static void put_uncharge_info(struct file_region *rg)
313 {
314 #ifdef CONFIG_CGROUP_HUGETLB
315         if (rg->css)
316                 css_put(rg->css);
317 #endif
318 }
319
320 static bool has_same_uncharge_info(struct file_region *rg,
321                                    struct file_region *org)
322 {
323 #ifdef CONFIG_CGROUP_HUGETLB
324         return rg && org &&
325                rg->reservation_counter == org->reservation_counter &&
326                rg->css == org->css;
327
328 #else
329         return true;
330 #endif
331 }
332
333 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
334 {
335         struct file_region *nrg = NULL, *prg = NULL;
336
337         prg = list_prev_entry(rg, link);
338         if (&prg->link != &resv->regions && prg->to == rg->from &&
339             has_same_uncharge_info(prg, rg)) {
340                 prg->to = rg->to;
341
342                 list_del(&rg->link);
343                 put_uncharge_info(rg);
344                 kfree(rg);
345
346                 rg = prg;
347         }
348
349         nrg = list_next_entry(rg, link);
350         if (&nrg->link != &resv->regions && nrg->from == rg->to &&
351             has_same_uncharge_info(nrg, rg)) {
352                 nrg->from = rg->from;
353
354                 list_del(&rg->link);
355                 put_uncharge_info(rg);
356                 kfree(rg);
357         }
358 }
359
360 static inline long
361 hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from,
362                      long to, struct hstate *h, struct hugetlb_cgroup *cg,
363                      long *regions_needed)
364 {
365         struct file_region *nrg;
366
367         if (!regions_needed) {
368                 nrg = get_file_region_entry_from_cache(map, from, to);
369                 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
370                 list_add(&nrg->link, rg->link.prev);
371                 coalesce_file_region(map, nrg);
372         } else
373                 *regions_needed += 1;
374
375         return to - from;
376 }
377
378 /*
379  * Must be called with resv->lock held.
380  *
381  * Calling this with regions_needed != NULL will count the number of pages
382  * to be added but will not modify the linked list. And regions_needed will
383  * indicate the number of file_regions needed in the cache to carry out to add
384  * the regions for this range.
385  */
386 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
387                                      struct hugetlb_cgroup *h_cg,
388                                      struct hstate *h, long *regions_needed)
389 {
390         long add = 0;
391         struct list_head *head = &resv->regions;
392         long last_accounted_offset = f;
393         struct file_region *rg = NULL, *trg = NULL;
394
395         if (regions_needed)
396                 *regions_needed = 0;
397
398         /* In this loop, we essentially handle an entry for the range
399          * [last_accounted_offset, rg->from), at every iteration, with some
400          * bounds checking.
401          */
402         list_for_each_entry_safe(rg, trg, head, link) {
403                 /* Skip irrelevant regions that start before our range. */
404                 if (rg->from < f) {
405                         /* If this region ends after the last accounted offset,
406                          * then we need to update last_accounted_offset.
407                          */
408                         if (rg->to > last_accounted_offset)
409                                 last_accounted_offset = rg->to;
410                         continue;
411                 }
412
413                 /* When we find a region that starts beyond our range, we've
414                  * finished.
415                  */
416                 if (rg->from >= t)
417                         break;
418
419                 /* Add an entry for last_accounted_offset -> rg->from, and
420                  * update last_accounted_offset.
421                  */
422                 if (rg->from > last_accounted_offset)
423                         add += hugetlb_resv_map_add(resv, rg,
424                                                     last_accounted_offset,
425                                                     rg->from, h, h_cg,
426                                                     regions_needed);
427
428                 last_accounted_offset = rg->to;
429         }
430
431         /* Handle the case where our range extends beyond
432          * last_accounted_offset.
433          */
434         if (last_accounted_offset < t)
435                 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
436                                             t, h, h_cg, regions_needed);
437
438         VM_BUG_ON(add < 0);
439         return add;
440 }
441
442 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
443  */
444 static int allocate_file_region_entries(struct resv_map *resv,
445                                         int regions_needed)
446         __must_hold(&resv->lock)
447 {
448         struct list_head allocated_regions;
449         int to_allocate = 0, i = 0;
450         struct file_region *trg = NULL, *rg = NULL;
451
452         VM_BUG_ON(regions_needed < 0);
453
454         INIT_LIST_HEAD(&allocated_regions);
455
456         /*
457          * Check for sufficient descriptors in the cache to accommodate
458          * the number of in progress add operations plus regions_needed.
459          *
460          * This is a while loop because when we drop the lock, some other call
461          * to region_add or region_del may have consumed some region_entries,
462          * so we keep looping here until we finally have enough entries for
463          * (adds_in_progress + regions_needed).
464          */
465         while (resv->region_cache_count <
466                (resv->adds_in_progress + regions_needed)) {
467                 to_allocate = resv->adds_in_progress + regions_needed -
468                               resv->region_cache_count;
469
470                 /* At this point, we should have enough entries in the cache
471                  * for all the existing adds_in_progress. We should only be
472                  * needing to allocate for regions_needed.
473                  */
474                 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
475
476                 spin_unlock(&resv->lock);
477                 for (i = 0; i < to_allocate; i++) {
478                         trg = kmalloc(sizeof(*trg), GFP_KERNEL);
479                         if (!trg)
480                                 goto out_of_memory;
481                         list_add(&trg->link, &allocated_regions);
482                 }
483
484                 spin_lock(&resv->lock);
485
486                 list_splice(&allocated_regions, &resv->region_cache);
487                 resv->region_cache_count += to_allocate;
488         }
489
490         return 0;
491
492 out_of_memory:
493         list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
494                 list_del(&rg->link);
495                 kfree(rg);
496         }
497         return -ENOMEM;
498 }
499
500 /*
501  * Add the huge page range represented by [f, t) to the reserve
502  * map.  Regions will be taken from the cache to fill in this range.
503  * Sufficient regions should exist in the cache due to the previous
504  * call to region_chg with the same range, but in some cases the cache will not
505  * have sufficient entries due to races with other code doing region_add or
506  * region_del.  The extra needed entries will be allocated.
507  *
508  * regions_needed is the out value provided by a previous call to region_chg.
509  *
510  * Return the number of new huge pages added to the map.  This number is greater
511  * than or equal to zero.  If file_region entries needed to be allocated for
512  * this operation and we were not able to allocate, it returns -ENOMEM.
513  * region_add of regions of length 1 never allocate file_regions and cannot
514  * fail; region_chg will always allocate at least 1 entry and a region_add for
515  * 1 page will only require at most 1 entry.
516  */
517 static long region_add(struct resv_map *resv, long f, long t,
518                        long in_regions_needed, struct hstate *h,
519                        struct hugetlb_cgroup *h_cg)
520 {
521         long add = 0, actual_regions_needed = 0;
522
523         spin_lock(&resv->lock);
524 retry:
525
526         /* Count how many regions are actually needed to execute this add. */
527         add_reservation_in_range(resv, f, t, NULL, NULL,
528                                  &actual_regions_needed);
529
530         /*
531          * Check for sufficient descriptors in the cache to accommodate
532          * this add operation. Note that actual_regions_needed may be greater
533          * than in_regions_needed, as the resv_map may have been modified since
534          * the region_chg call. In this case, we need to make sure that we
535          * allocate extra entries, such that we have enough for all the
536          * existing adds_in_progress, plus the excess needed for this
537          * operation.
538          */
539         if (actual_regions_needed > in_regions_needed &&
540             resv->region_cache_count <
541                     resv->adds_in_progress +
542                             (actual_regions_needed - in_regions_needed)) {
543                 /* region_add operation of range 1 should never need to
544                  * allocate file_region entries.
545                  */
546                 VM_BUG_ON(t - f <= 1);
547
548                 if (allocate_file_region_entries(
549                             resv, actual_regions_needed - in_regions_needed)) {
550                         return -ENOMEM;
551                 }
552
553                 goto retry;
554         }
555
556         add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
557
558         resv->adds_in_progress -= in_regions_needed;
559
560         spin_unlock(&resv->lock);
561         return add;
562 }
563
564 /*
565  * Examine the existing reserve map and determine how many
566  * huge pages in the specified range [f, t) are NOT currently
567  * represented.  This routine is called before a subsequent
568  * call to region_add that will actually modify the reserve
569  * map to add the specified range [f, t).  region_chg does
570  * not change the number of huge pages represented by the
571  * map.  A number of new file_region structures is added to the cache as a
572  * placeholder, for the subsequent region_add call to use. At least 1
573  * file_region structure is added.
574  *
575  * out_regions_needed is the number of regions added to the
576  * resv->adds_in_progress.  This value needs to be provided to a follow up call
577  * to region_add or region_abort for proper accounting.
578  *
579  * Returns the number of huge pages that need to be added to the existing
580  * reservation map for the range [f, t).  This number is greater or equal to
581  * zero.  -ENOMEM is returned if a new file_region structure or cache entry
582  * is needed and can not be allocated.
583  */
584 static long region_chg(struct resv_map *resv, long f, long t,
585                        long *out_regions_needed)
586 {
587         long chg = 0;
588
589         spin_lock(&resv->lock);
590
591         /* Count how many hugepages in this range are NOT represented. */
592         chg = add_reservation_in_range(resv, f, t, NULL, NULL,
593                                        out_regions_needed);
594
595         if (*out_regions_needed == 0)
596                 *out_regions_needed = 1;
597
598         if (allocate_file_region_entries(resv, *out_regions_needed))
599                 return -ENOMEM;
600
601         resv->adds_in_progress += *out_regions_needed;
602
603         spin_unlock(&resv->lock);
604         return chg;
605 }
606
607 /*
608  * Abort the in progress add operation.  The adds_in_progress field
609  * of the resv_map keeps track of the operations in progress between
610  * calls to region_chg and region_add.  Operations are sometimes
611  * aborted after the call to region_chg.  In such cases, region_abort
612  * is called to decrement the adds_in_progress counter. regions_needed
613  * is the value returned by the region_chg call, it is used to decrement
614  * the adds_in_progress counter.
615  *
616  * NOTE: The range arguments [f, t) are not needed or used in this
617  * routine.  They are kept to make reading the calling code easier as
618  * arguments will match the associated region_chg call.
619  */
620 static void region_abort(struct resv_map *resv, long f, long t,
621                          long regions_needed)
622 {
623         spin_lock(&resv->lock);
624         VM_BUG_ON(!resv->region_cache_count);
625         resv->adds_in_progress -= regions_needed;
626         spin_unlock(&resv->lock);
627 }
628
629 /*
630  * Delete the specified range [f, t) from the reserve map.  If the
631  * t parameter is LONG_MAX, this indicates that ALL regions after f
632  * should be deleted.  Locate the regions which intersect [f, t)
633  * and either trim, delete or split the existing regions.
634  *
635  * Returns the number of huge pages deleted from the reserve map.
636  * In the normal case, the return value is zero or more.  In the
637  * case where a region must be split, a new region descriptor must
638  * be allocated.  If the allocation fails, -ENOMEM will be returned.
639  * NOTE: If the parameter t == LONG_MAX, then we will never split
640  * a region and possibly return -ENOMEM.  Callers specifying
641  * t == LONG_MAX do not need to check for -ENOMEM error.
642  */
643 static long region_del(struct resv_map *resv, long f, long t)
644 {
645         struct list_head *head = &resv->regions;
646         struct file_region *rg, *trg;
647         struct file_region *nrg = NULL;
648         long del = 0;
649
650 retry:
651         spin_lock(&resv->lock);
652         list_for_each_entry_safe(rg, trg, head, link) {
653                 /*
654                  * Skip regions before the range to be deleted.  file_region
655                  * ranges are normally of the form [from, to).  However, there
656                  * may be a "placeholder" entry in the map which is of the form
657                  * (from, to) with from == to.  Check for placeholder entries
658                  * at the beginning of the range to be deleted.
659                  */
660                 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
661                         continue;
662
663                 if (rg->from >= t)
664                         break;
665
666                 if (f > rg->from && t < rg->to) { /* Must split region */
667                         /*
668                          * Check for an entry in the cache before dropping
669                          * lock and attempting allocation.
670                          */
671                         if (!nrg &&
672                             resv->region_cache_count > resv->adds_in_progress) {
673                                 nrg = list_first_entry(&resv->region_cache,
674                                                         struct file_region,
675                                                         link);
676                                 list_del(&nrg->link);
677                                 resv->region_cache_count--;
678                         }
679
680                         if (!nrg) {
681                                 spin_unlock(&resv->lock);
682                                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
683                                 if (!nrg)
684                                         return -ENOMEM;
685                                 goto retry;
686                         }
687
688                         del += t - f;
689                         hugetlb_cgroup_uncharge_file_region(
690                                 resv, rg, t - f, false);
691
692                         /* New entry for end of split region */
693                         nrg->from = t;
694                         nrg->to = rg->to;
695
696                         copy_hugetlb_cgroup_uncharge_info(nrg, rg);
697
698                         INIT_LIST_HEAD(&nrg->link);
699
700                         /* Original entry is trimmed */
701                         rg->to = f;
702
703                         list_add(&nrg->link, &rg->link);
704                         nrg = NULL;
705                         break;
706                 }
707
708                 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
709                         del += rg->to - rg->from;
710                         hugetlb_cgroup_uncharge_file_region(resv, rg,
711                                                             rg->to - rg->from, true);
712                         list_del(&rg->link);
713                         kfree(rg);
714                         continue;
715                 }
716
717                 if (f <= rg->from) {    /* Trim beginning of region */
718                         hugetlb_cgroup_uncharge_file_region(resv, rg,
719                                                             t - rg->from, false);
720
721                         del += t - rg->from;
722                         rg->from = t;
723                 } else {                /* Trim end of region */
724                         hugetlb_cgroup_uncharge_file_region(resv, rg,
725                                                             rg->to - f, false);
726
727                         del += rg->to - f;
728                         rg->to = f;
729                 }
730         }
731
732         spin_unlock(&resv->lock);
733         kfree(nrg);
734         return del;
735 }
736
737 /*
738  * A rare out of memory error was encountered which prevented removal of
739  * the reserve map region for a page.  The huge page itself was free'ed
740  * and removed from the page cache.  This routine will adjust the subpool
741  * usage count, and the global reserve count if needed.  By incrementing
742  * these counts, the reserve map entry which could not be deleted will
743  * appear as a "reserved" entry instead of simply dangling with incorrect
744  * counts.
745  */
746 void hugetlb_fix_reserve_counts(struct inode *inode)
747 {
748         struct hugepage_subpool *spool = subpool_inode(inode);
749         long rsv_adjust;
750         bool reserved = false;
751
752         rsv_adjust = hugepage_subpool_get_pages(spool, 1);
753         if (rsv_adjust > 0) {
754                 struct hstate *h = hstate_inode(inode);
755
756                 if (!hugetlb_acct_memory(h, 1))
757                         reserved = true;
758         } else if (!rsv_adjust) {
759                 reserved = true;
760         }
761
762         if (!reserved)
763                 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
764 }
765
766 /*
767  * Count and return the number of huge pages in the reserve map
768  * that intersect with the range [f, t).
769  */
770 static long region_count(struct resv_map *resv, long f, long t)
771 {
772         struct list_head *head = &resv->regions;
773         struct file_region *rg;
774         long chg = 0;
775
776         spin_lock(&resv->lock);
777         /* Locate each segment we overlap with, and count that overlap. */
778         list_for_each_entry(rg, head, link) {
779                 long seg_from;
780                 long seg_to;
781
782                 if (rg->to <= f)
783                         continue;
784                 if (rg->from >= t)
785                         break;
786
787                 seg_from = max(rg->from, f);
788                 seg_to = min(rg->to, t);
789
790                 chg += seg_to - seg_from;
791         }
792         spin_unlock(&resv->lock);
793
794         return chg;
795 }
796
797 /*
798  * Convert the address within this vma to the page offset within
799  * the mapping, in pagecache page units; huge pages here.
800  */
801 static pgoff_t vma_hugecache_offset(struct hstate *h,
802                         struct vm_area_struct *vma, unsigned long address)
803 {
804         return ((address - vma->vm_start) >> huge_page_shift(h)) +
805                         (vma->vm_pgoff >> huge_page_order(h));
806 }
807
808 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
809                                      unsigned long address)
810 {
811         return vma_hugecache_offset(hstate_vma(vma), vma, address);
812 }
813 EXPORT_SYMBOL_GPL(linear_hugepage_index);
814
815 /*
816  * Return the size of the pages allocated when backing a VMA. In the majority
817  * cases this will be same size as used by the page table entries.
818  */
819 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
820 {
821         if (vma->vm_ops && vma->vm_ops->pagesize)
822                 return vma->vm_ops->pagesize(vma);
823         return PAGE_SIZE;
824 }
825 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
826
827 /*
828  * Return the page size being used by the MMU to back a VMA. In the majority
829  * of cases, the page size used by the kernel matches the MMU size. On
830  * architectures where it differs, an architecture-specific 'strong'
831  * version of this symbol is required.
832  */
833 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
834 {
835         return vma_kernel_pagesize(vma);
836 }
837
838 /*
839  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
840  * bits of the reservation map pointer, which are always clear due to
841  * alignment.
842  */
843 #define HPAGE_RESV_OWNER    (1UL << 0)
844 #define HPAGE_RESV_UNMAPPED (1UL << 1)
845 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
846
847 /*
848  * These helpers are used to track how many pages are reserved for
849  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
850  * is guaranteed to have their future faults succeed.
851  *
852  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
853  * the reserve counters are updated with the hugetlb_lock held. It is safe
854  * to reset the VMA at fork() time as it is not in use yet and there is no
855  * chance of the global counters getting corrupted as a result of the values.
856  *
857  * The private mapping reservation is represented in a subtly different
858  * manner to a shared mapping.  A shared mapping has a region map associated
859  * with the underlying file, this region map represents the backing file
860  * pages which have ever had a reservation assigned which this persists even
861  * after the page is instantiated.  A private mapping has a region map
862  * associated with the original mmap which is attached to all VMAs which
863  * reference it, this region map represents those offsets which have consumed
864  * reservation ie. where pages have been instantiated.
865  */
866 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
867 {
868         return (unsigned long)vma->vm_private_data;
869 }
870
871 static void set_vma_private_data(struct vm_area_struct *vma,
872                                                         unsigned long value)
873 {
874         vma->vm_private_data = (void *)value;
875 }
876
877 static void
878 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
879                                           struct hugetlb_cgroup *h_cg,
880                                           struct hstate *h)
881 {
882 #ifdef CONFIG_CGROUP_HUGETLB
883         if (!h_cg || !h) {
884                 resv_map->reservation_counter = NULL;
885                 resv_map->pages_per_hpage = 0;
886                 resv_map->css = NULL;
887         } else {
888                 resv_map->reservation_counter =
889                         &h_cg->rsvd_hugepage[hstate_index(h)];
890                 resv_map->pages_per_hpage = pages_per_huge_page(h);
891                 resv_map->css = &h_cg->css;
892         }
893 #endif
894 }
895
896 struct resv_map *resv_map_alloc(void)
897 {
898         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
899         struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
900
901         if (!resv_map || !rg) {
902                 kfree(resv_map);
903                 kfree(rg);
904                 return NULL;
905         }
906
907         kref_init(&resv_map->refs);
908         spin_lock_init(&resv_map->lock);
909         INIT_LIST_HEAD(&resv_map->regions);
910
911         resv_map->adds_in_progress = 0;
912         /*
913          * Initialize these to 0. On shared mappings, 0's here indicate these
914          * fields don't do cgroup accounting. On private mappings, these will be
915          * re-initialized to the proper values, to indicate that hugetlb cgroup
916          * reservations are to be un-charged from here.
917          */
918         resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
919
920         INIT_LIST_HEAD(&resv_map->region_cache);
921         list_add(&rg->link, &resv_map->region_cache);
922         resv_map->region_cache_count = 1;
923
924         return resv_map;
925 }
926
927 void resv_map_release(struct kref *ref)
928 {
929         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
930         struct list_head *head = &resv_map->region_cache;
931         struct file_region *rg, *trg;
932
933         /* Clear out any active regions before we release the map. */
934         region_del(resv_map, 0, LONG_MAX);
935
936         /* ... and any entries left in the cache */
937         list_for_each_entry_safe(rg, trg, head, link) {
938                 list_del(&rg->link);
939                 kfree(rg);
940         }
941
942         VM_BUG_ON(resv_map->adds_in_progress);
943
944         kfree(resv_map);
945 }
946
947 static inline struct resv_map *inode_resv_map(struct inode *inode)
948 {
949         /*
950          * At inode evict time, i_mapping may not point to the original
951          * address space within the inode.  This original address space
952          * contains the pointer to the resv_map.  So, always use the
953          * address space embedded within the inode.
954          * The VERY common case is inode->mapping == &inode->i_data but,
955          * this may not be true for device special inodes.
956          */
957         return (struct resv_map *)(&inode->i_data)->private_data;
958 }
959
960 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
961 {
962         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
963         if (vma->vm_flags & VM_MAYSHARE) {
964                 struct address_space *mapping = vma->vm_file->f_mapping;
965                 struct inode *inode = mapping->host;
966
967                 return inode_resv_map(inode);
968
969         } else {
970                 return (struct resv_map *)(get_vma_private_data(vma) &
971                                                         ~HPAGE_RESV_MASK);
972         }
973 }
974
975 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
976 {
977         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
978         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
979
980         set_vma_private_data(vma, (get_vma_private_data(vma) &
981                                 HPAGE_RESV_MASK) | (unsigned long)map);
982 }
983
984 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
985 {
986         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
987         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
988
989         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
990 }
991
992 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
993 {
994         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
995
996         return (get_vma_private_data(vma) & flag) != 0;
997 }
998
999 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
1000 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
1001 {
1002         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1003         if (!(vma->vm_flags & VM_MAYSHARE))
1004                 vma->vm_private_data = (void *)0;
1005 }
1006
1007 /* Returns true if the VMA has associated reserve pages */
1008 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1009 {
1010         if (vma->vm_flags & VM_NORESERVE) {
1011                 /*
1012                  * This address is already reserved by other process(chg == 0),
1013                  * so, we should decrement reserved count. Without decrementing,
1014                  * reserve count remains after releasing inode, because this
1015                  * allocated page will go into page cache and is regarded as
1016                  * coming from reserved pool in releasing step.  Currently, we
1017                  * don't have any other solution to deal with this situation
1018                  * properly, so add work-around here.
1019                  */
1020                 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1021                         return true;
1022                 else
1023                         return false;
1024         }
1025
1026         /* Shared mappings always use reserves */
1027         if (vma->vm_flags & VM_MAYSHARE) {
1028                 /*
1029                  * We know VM_NORESERVE is not set.  Therefore, there SHOULD
1030                  * be a region map for all pages.  The only situation where
1031                  * there is no region map is if a hole was punched via
1032                  * fallocate.  In this case, there really are no reserves to
1033                  * use.  This situation is indicated if chg != 0.
1034                  */
1035                 if (chg)
1036                         return false;
1037                 else
1038                         return true;
1039         }
1040
1041         /*
1042          * Only the process that called mmap() has reserves for
1043          * private mappings.
1044          */
1045         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1046                 /*
1047                  * Like the shared case above, a hole punch or truncate
1048                  * could have been performed on the private mapping.
1049                  * Examine the value of chg to determine if reserves
1050                  * actually exist or were previously consumed.
1051                  * Very Subtle - The value of chg comes from a previous
1052                  * call to vma_needs_reserves().  The reserve map for
1053                  * private mappings has different (opposite) semantics
1054                  * than that of shared mappings.  vma_needs_reserves()
1055                  * has already taken this difference in semantics into
1056                  * account.  Therefore, the meaning of chg is the same
1057                  * as in the shared case above.  Code could easily be
1058                  * combined, but keeping it separate draws attention to
1059                  * subtle differences.
1060                  */
1061                 if (chg)
1062                         return false;
1063                 else
1064                         return true;
1065         }
1066
1067         return false;
1068 }
1069
1070 static void enqueue_huge_page(struct hstate *h, struct page *page)
1071 {
1072         int nid = page_to_nid(page);
1073
1074         lockdep_assert_held(&hugetlb_lock);
1075         VM_BUG_ON_PAGE(page_count(page), page);
1076
1077         list_move(&page->lru, &h->hugepage_freelists[nid]);
1078         h->free_huge_pages++;
1079         h->free_huge_pages_node[nid]++;
1080         SetHPageFreed(page);
1081 }
1082
1083 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1084 {
1085         struct page *page;
1086         bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1087
1088         lockdep_assert_held(&hugetlb_lock);
1089         list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1090                 if (pin && !is_pinnable_page(page))
1091                         continue;
1092
1093                 if (PageHWPoison(page))
1094                         continue;
1095
1096                 list_move(&page->lru, &h->hugepage_activelist);
1097                 set_page_refcounted(page);
1098                 ClearHPageFreed(page);
1099                 h->free_huge_pages--;
1100                 h->free_huge_pages_node[nid]--;
1101                 return page;
1102         }
1103
1104         return NULL;
1105 }
1106
1107 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1108                 nodemask_t *nmask)
1109 {
1110         unsigned int cpuset_mems_cookie;
1111         struct zonelist *zonelist;
1112         struct zone *zone;
1113         struct zoneref *z;
1114         int node = NUMA_NO_NODE;
1115
1116         zonelist = node_zonelist(nid, gfp_mask);
1117
1118 retry_cpuset:
1119         cpuset_mems_cookie = read_mems_allowed_begin();
1120         for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1121                 struct page *page;
1122
1123                 if (!cpuset_zone_allowed(zone, gfp_mask))
1124                         continue;
1125                 /*
1126                  * no need to ask again on the same node. Pool is node rather than
1127                  * zone aware
1128                  */
1129                 if (zone_to_nid(zone) == node)
1130                         continue;
1131                 node = zone_to_nid(zone);
1132
1133                 page = dequeue_huge_page_node_exact(h, node);
1134                 if (page)
1135                         return page;
1136         }
1137         if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1138                 goto retry_cpuset;
1139
1140         return NULL;
1141 }
1142
1143 static struct page *dequeue_huge_page_vma(struct hstate *h,
1144                                 struct vm_area_struct *vma,
1145                                 unsigned long address, int avoid_reserve,
1146                                 long chg)
1147 {
1148         struct page *page = NULL;
1149         struct mempolicy *mpol;
1150         gfp_t gfp_mask;
1151         nodemask_t *nodemask;
1152         int nid;
1153
1154         /*
1155          * A child process with MAP_PRIVATE mappings created by their parent
1156          * have no page reserves. This check ensures that reservations are
1157          * not "stolen". The child may still get SIGKILLed
1158          */
1159         if (!vma_has_reserves(vma, chg) &&
1160                         h->free_huge_pages - h->resv_huge_pages == 0)
1161                 goto err;
1162
1163         /* If reserves cannot be used, ensure enough pages are in the pool */
1164         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1165                 goto err;
1166
1167         gfp_mask = htlb_alloc_mask(h);
1168         nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1169
1170         if (mpol_is_preferred_many(mpol)) {
1171                 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1172
1173                 /* Fallback to all nodes if page==NULL */
1174                 nodemask = NULL;
1175         }
1176
1177         if (!page)
1178                 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1179
1180         if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1181                 SetHPageRestoreReserve(page);
1182                 h->resv_huge_pages--;
1183         }
1184
1185         mpol_cond_put(mpol);
1186         return page;
1187
1188 err:
1189         return NULL;
1190 }
1191
1192 /*
1193  * common helper functions for hstate_next_node_to_{alloc|free}.
1194  * We may have allocated or freed a huge page based on a different
1195  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1196  * be outside of *nodes_allowed.  Ensure that we use an allowed
1197  * node for alloc or free.
1198  */
1199 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1200 {
1201         nid = next_node_in(nid, *nodes_allowed);
1202         VM_BUG_ON(nid >= MAX_NUMNODES);
1203
1204         return nid;
1205 }
1206
1207 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1208 {
1209         if (!node_isset(nid, *nodes_allowed))
1210                 nid = next_node_allowed(nid, nodes_allowed);
1211         return nid;
1212 }
1213
1214 /*
1215  * returns the previously saved node ["this node"] from which to
1216  * allocate a persistent huge page for the pool and advance the
1217  * next node from which to allocate, handling wrap at end of node
1218  * mask.
1219  */
1220 static int hstate_next_node_to_alloc(struct hstate *h,
1221                                         nodemask_t *nodes_allowed)
1222 {
1223         int nid;
1224
1225         VM_BUG_ON(!nodes_allowed);
1226
1227         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1228         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1229
1230         return nid;
1231 }
1232
1233 /*
1234  * helper for remove_pool_huge_page() - return the previously saved
1235  * node ["this node"] from which to free a huge page.  Advance the
1236  * next node id whether or not we find a free huge page to free so
1237  * that the next attempt to free addresses the next node.
1238  */
1239 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1240 {
1241         int nid;
1242
1243         VM_BUG_ON(!nodes_allowed);
1244
1245         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1246         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1247
1248         return nid;
1249 }
1250
1251 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)           \
1252         for (nr_nodes = nodes_weight(*mask);                            \
1253                 nr_nodes > 0 &&                                         \
1254                 ((node = hstate_next_node_to_alloc(hs, mask)) || 1);    \
1255                 nr_nodes--)
1256
1257 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)            \
1258         for (nr_nodes = nodes_weight(*mask);                            \
1259                 nr_nodes > 0 &&                                         \
1260                 ((node = hstate_next_node_to_free(hs, mask)) || 1);     \
1261                 nr_nodes--)
1262
1263 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1264 static void destroy_compound_gigantic_page(struct page *page,
1265                                         unsigned int order)
1266 {
1267         int i;
1268         int nr_pages = 1 << order;
1269         struct page *p = page + 1;
1270
1271         atomic_set(compound_mapcount_ptr(page), 0);
1272         atomic_set(compound_pincount_ptr(page), 0);
1273
1274         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1275                 clear_compound_head(p);
1276                 set_page_refcounted(p);
1277         }
1278
1279         set_compound_order(page, 0);
1280         page[1].compound_nr = 0;
1281         __ClearPageHead(page);
1282 }
1283
1284 static void free_gigantic_page(struct page *page, unsigned int order)
1285 {
1286         /*
1287          * If the page isn't allocated using the cma allocator,
1288          * cma_release() returns false.
1289          */
1290 #ifdef CONFIG_CMA
1291         if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1292                 return;
1293 #endif
1294
1295         free_contig_range(page_to_pfn(page), 1 << order);
1296 }
1297
1298 #ifdef CONFIG_CONTIG_ALLOC
1299 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1300                 int nid, nodemask_t *nodemask)
1301 {
1302         unsigned long nr_pages = pages_per_huge_page(h);
1303         if (nid == NUMA_NO_NODE)
1304                 nid = numa_mem_id();
1305
1306 #ifdef CONFIG_CMA
1307         {
1308                 struct page *page;
1309                 int node;
1310
1311                 if (hugetlb_cma[nid]) {
1312                         page = cma_alloc(hugetlb_cma[nid], nr_pages,
1313                                         huge_page_order(h), true);
1314                         if (page)
1315                                 return page;
1316                 }
1317
1318                 if (!(gfp_mask & __GFP_THISNODE)) {
1319                         for_each_node_mask(node, *nodemask) {
1320                                 if (node == nid || !hugetlb_cma[node])
1321                                         continue;
1322
1323                                 page = cma_alloc(hugetlb_cma[node], nr_pages,
1324                                                 huge_page_order(h), true);
1325                                 if (page)
1326                                         return page;
1327                         }
1328                 }
1329         }
1330 #endif
1331
1332         return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1333 }
1334
1335 #else /* !CONFIG_CONTIG_ALLOC */
1336 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1337                                         int nid, nodemask_t *nodemask)
1338 {
1339         return NULL;
1340 }
1341 #endif /* CONFIG_CONTIG_ALLOC */
1342
1343 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1344 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1345                                         int nid, nodemask_t *nodemask)
1346 {
1347         return NULL;
1348 }
1349 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1350 static inline void destroy_compound_gigantic_page(struct page *page,
1351                                                 unsigned int order) { }
1352 #endif
1353
1354 /*
1355  * Remove hugetlb page from lists, and update dtor so that page appears
1356  * as just a compound page.  A reference is held on the page.
1357  *
1358  * Must be called with hugetlb lock held.
1359  */
1360 static void remove_hugetlb_page(struct hstate *h, struct page *page,
1361                                                         bool adjust_surplus)
1362 {
1363         int nid = page_to_nid(page);
1364
1365         VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1366         VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1367
1368         lockdep_assert_held(&hugetlb_lock);
1369         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1370                 return;
1371
1372         list_del(&page->lru);
1373
1374         if (HPageFreed(page)) {
1375                 h->free_huge_pages--;
1376                 h->free_huge_pages_node[nid]--;
1377         }
1378         if (adjust_surplus) {
1379                 h->surplus_huge_pages--;
1380                 h->surplus_huge_pages_node[nid]--;
1381         }
1382
1383         /*
1384          * Very subtle
1385          *
1386          * For non-gigantic pages set the destructor to the normal compound
1387          * page dtor.  This is needed in case someone takes an additional
1388          * temporary ref to the page, and freeing is delayed until they drop
1389          * their reference.
1390          *
1391          * For gigantic pages set the destructor to the null dtor.  This
1392          * destructor will never be called.  Before freeing the gigantic
1393          * page destroy_compound_gigantic_page will turn the compound page
1394          * into a simple group of pages.  After this the destructor does not
1395          * apply.
1396          *
1397          * This handles the case where more than one ref is held when and
1398          * after update_and_free_page is called.
1399          */
1400         set_page_refcounted(page);
1401         if (hstate_is_gigantic(h))
1402                 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1403         else
1404                 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
1405
1406         h->nr_huge_pages--;
1407         h->nr_huge_pages_node[nid]--;
1408 }
1409
1410 static void add_hugetlb_page(struct hstate *h, struct page *page,
1411                              bool adjust_surplus)
1412 {
1413         int zeroed;
1414         int nid = page_to_nid(page);
1415
1416         VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page);
1417
1418         lockdep_assert_held(&hugetlb_lock);
1419
1420         INIT_LIST_HEAD(&page->lru);
1421         h->nr_huge_pages++;
1422         h->nr_huge_pages_node[nid]++;
1423
1424         if (adjust_surplus) {
1425                 h->surplus_huge_pages++;
1426                 h->surplus_huge_pages_node[nid]++;
1427         }
1428
1429         set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1430         set_page_private(page, 0);
1431         SetHPageVmemmapOptimized(page);
1432
1433         /*
1434          * This page is about to be managed by the hugetlb allocator and
1435          * should have no users.  Drop our reference, and check for others
1436          * just in case.
1437          */
1438         zeroed = put_page_testzero(page);
1439         if (!zeroed)
1440                 /*
1441                  * It is VERY unlikely soneone else has taken a ref on
1442                  * the page.  In this case, we simply return as the
1443                  * hugetlb destructor (free_huge_page) will be called
1444                  * when this other ref is dropped.
1445                  */
1446                 return;
1447
1448         arch_clear_hugepage_flags(page);
1449         enqueue_huge_page(h, page);
1450 }
1451
1452 static void __update_and_free_page(struct hstate *h, struct page *page)
1453 {
1454         int i;
1455         struct page *subpage = page;
1456
1457         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1458                 return;
1459
1460         if (alloc_huge_page_vmemmap(h, page)) {
1461                 spin_lock_irq(&hugetlb_lock);
1462                 /*
1463                  * If we cannot allocate vmemmap pages, just refuse to free the
1464                  * page and put the page back on the hugetlb free list and treat
1465                  * as a surplus page.
1466                  */
1467                 add_hugetlb_page(h, page, true);
1468                 spin_unlock_irq(&hugetlb_lock);
1469                 return;
1470         }
1471
1472         for (i = 0; i < pages_per_huge_page(h);
1473              i++, subpage = mem_map_next(subpage, page, i)) {
1474                 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1475                                 1 << PG_referenced | 1 << PG_dirty |
1476                                 1 << PG_active | 1 << PG_private |
1477                                 1 << PG_writeback);
1478         }
1479         if (hstate_is_gigantic(h)) {
1480                 destroy_compound_gigantic_page(page, huge_page_order(h));
1481                 free_gigantic_page(page, huge_page_order(h));
1482         } else {
1483                 __free_pages(page, huge_page_order(h));
1484         }
1485 }
1486
1487 /*
1488  * As update_and_free_page() can be called under any context, so we cannot
1489  * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1490  * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1491  * the vmemmap pages.
1492  *
1493  * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1494  * freed and frees them one-by-one. As the page->mapping pointer is going
1495  * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1496  * structure of a lockless linked list of huge pages to be freed.
1497  */
1498 static LLIST_HEAD(hpage_freelist);
1499
1500 static void free_hpage_workfn(struct work_struct *work)
1501 {
1502         struct llist_node *node;
1503
1504         node = llist_del_all(&hpage_freelist);
1505
1506         while (node) {
1507                 struct page *page;
1508                 struct hstate *h;
1509
1510                 page = container_of((struct address_space **)node,
1511                                      struct page, mapping);
1512                 node = node->next;
1513                 page->mapping = NULL;
1514                 /*
1515                  * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate()
1516                  * is going to trigger because a previous call to
1517                  * remove_hugetlb_page() will set_compound_page_dtor(page,
1518                  * NULL_COMPOUND_DTOR), so do not use page_hstate() directly.
1519                  */
1520                 h = size_to_hstate(page_size(page));
1521
1522                 __update_and_free_page(h, page);
1523
1524                 cond_resched();
1525         }
1526 }
1527 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1528
1529 static inline void flush_free_hpage_work(struct hstate *h)
1530 {
1531         if (free_vmemmap_pages_per_hpage(h))
1532                 flush_work(&free_hpage_work);
1533 }
1534
1535 static void update_and_free_page(struct hstate *h, struct page *page,
1536                                  bool atomic)
1537 {
1538         if (!HPageVmemmapOptimized(page) || !atomic) {
1539                 __update_and_free_page(h, page);
1540                 return;
1541         }
1542
1543         /*
1544          * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1545          *
1546          * Only call schedule_work() if hpage_freelist is previously
1547          * empty. Otherwise, schedule_work() had been called but the workfn
1548          * hasn't retrieved the list yet.
1549          */
1550         if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist))
1551                 schedule_work(&free_hpage_work);
1552 }
1553
1554 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1555 {
1556         struct page *page, *t_page;
1557
1558         list_for_each_entry_safe(page, t_page, list, lru) {
1559                 update_and_free_page(h, page, false);
1560                 cond_resched();
1561         }
1562 }
1563
1564 struct hstate *size_to_hstate(unsigned long size)
1565 {
1566         struct hstate *h;
1567
1568         for_each_hstate(h) {
1569                 if (huge_page_size(h) == size)
1570                         return h;
1571         }
1572         return NULL;
1573 }
1574
1575 void free_huge_page(struct page *page)
1576 {
1577         /*
1578          * Can't pass hstate in here because it is called from the
1579          * compound page destructor.
1580          */
1581         struct hstate *h = page_hstate(page);
1582         int nid = page_to_nid(page);
1583         struct hugepage_subpool *spool = hugetlb_page_subpool(page);
1584         bool restore_reserve;
1585         unsigned long flags;
1586
1587         VM_BUG_ON_PAGE(page_count(page), page);
1588         VM_BUG_ON_PAGE(page_mapcount(page), page);
1589
1590         hugetlb_set_page_subpool(page, NULL);
1591         page->mapping = NULL;
1592         restore_reserve = HPageRestoreReserve(page);
1593         ClearHPageRestoreReserve(page);
1594
1595         /*
1596          * If HPageRestoreReserve was set on page, page allocation consumed a
1597          * reservation.  If the page was associated with a subpool, there
1598          * would have been a page reserved in the subpool before allocation
1599          * via hugepage_subpool_get_pages().  Since we are 'restoring' the
1600          * reservation, do not call hugepage_subpool_put_pages() as this will
1601          * remove the reserved page from the subpool.
1602          */
1603         if (!restore_reserve) {
1604                 /*
1605                  * A return code of zero implies that the subpool will be
1606                  * under its minimum size if the reservation is not restored
1607                  * after page is free.  Therefore, force restore_reserve
1608                  * operation.
1609                  */
1610                 if (hugepage_subpool_put_pages(spool, 1) == 0)
1611                         restore_reserve = true;
1612         }
1613
1614         spin_lock_irqsave(&hugetlb_lock, flags);
1615         ClearHPageMigratable(page);
1616         hugetlb_cgroup_uncharge_page(hstate_index(h),
1617                                      pages_per_huge_page(h), page);
1618         hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1619                                           pages_per_huge_page(h), page);
1620         if (restore_reserve)
1621                 h->resv_huge_pages++;
1622
1623         if (HPageTemporary(page)) {
1624                 remove_hugetlb_page(h, page, false);
1625                 spin_unlock_irqrestore(&hugetlb_lock, flags);
1626                 update_and_free_page(h, page, true);
1627         } else if (h->surplus_huge_pages_node[nid]) {
1628                 /* remove the page from active list */
1629                 remove_hugetlb_page(h, page, true);
1630                 spin_unlock_irqrestore(&hugetlb_lock, flags);
1631                 update_and_free_page(h, page, true);
1632         } else {
1633                 arch_clear_hugepage_flags(page);
1634                 enqueue_huge_page(h, page);
1635                 spin_unlock_irqrestore(&hugetlb_lock, flags);
1636         }
1637 }
1638
1639 /*
1640  * Must be called with the hugetlb lock held
1641  */
1642 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1643 {
1644         lockdep_assert_held(&hugetlb_lock);
1645         h->nr_huge_pages++;
1646         h->nr_huge_pages_node[nid]++;
1647 }
1648
1649 static void __prep_new_huge_page(struct hstate *h, struct page *page)
1650 {
1651         free_huge_page_vmemmap(h, page);
1652         INIT_LIST_HEAD(&page->lru);
1653         set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1654         hugetlb_set_page_subpool(page, NULL);
1655         set_hugetlb_cgroup(page, NULL);
1656         set_hugetlb_cgroup_rsvd(page, NULL);
1657 }
1658
1659 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1660 {
1661         __prep_new_huge_page(h, page);
1662         spin_lock_irq(&hugetlb_lock);
1663         __prep_account_new_huge_page(h, nid);
1664         spin_unlock_irq(&hugetlb_lock);
1665 }
1666
1667 static bool prep_compound_gigantic_page(struct page *page, unsigned int order)
1668 {
1669         int i, j;
1670         int nr_pages = 1 << order;
1671         struct page *p = page + 1;
1672
1673         /* we rely on prep_new_huge_page to set the destructor */
1674         set_compound_order(page, order);
1675         __ClearPageReserved(page);
1676         __SetPageHead(page);
1677         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1678                 /*
1679                  * For gigantic hugepages allocated through bootmem at
1680                  * boot, it's safer to be consistent with the not-gigantic
1681                  * hugepages and clear the PG_reserved bit from all tail pages
1682                  * too.  Otherwise drivers using get_user_pages() to access tail
1683                  * pages may get the reference counting wrong if they see
1684                  * PG_reserved set on a tail page (despite the head page not
1685                  * having PG_reserved set).  Enforcing this consistency between
1686                  * head and tail pages allows drivers to optimize away a check
1687                  * on the head page when they need know if put_page() is needed
1688                  * after get_user_pages().
1689                  */
1690                 __ClearPageReserved(p);
1691                 /*
1692                  * Subtle and very unlikely
1693                  *
1694                  * Gigantic 'page allocators' such as memblock or cma will
1695                  * return a set of pages with each page ref counted.  We need
1696                  * to turn this set of pages into a compound page with tail
1697                  * page ref counts set to zero.  Code such as speculative page
1698                  * cache adding could take a ref on a 'to be' tail page.
1699                  * We need to respect any increased ref count, and only set
1700                  * the ref count to zero if count is currently 1.  If count
1701                  * is not 1, we return an error.  An error return indicates
1702                  * the set of pages can not be converted to a gigantic page.
1703                  * The caller who allocated the pages should then discard the
1704                  * pages using the appropriate free interface.
1705                  */
1706                 if (!page_ref_freeze(p, 1)) {
1707                         pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
1708                         goto out_error;
1709                 }
1710                 set_page_count(p, 0);
1711                 set_compound_head(p, page);
1712         }
1713         atomic_set(compound_mapcount_ptr(page), -1);
1714         atomic_set(compound_pincount_ptr(page), 0);
1715         return true;
1716
1717 out_error:
1718         /* undo tail page modifications made above */
1719         p = page + 1;
1720         for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) {
1721                 clear_compound_head(p);
1722                 set_page_refcounted(p);
1723         }
1724         /* need to clear PG_reserved on remaining tail pages  */
1725         for (; j < nr_pages; j++, p = mem_map_next(p, page, j))
1726                 __ClearPageReserved(p);
1727         set_compound_order(page, 0);
1728         page[1].compound_nr = 0;
1729         __ClearPageHead(page);
1730         return false;
1731 }
1732
1733 /*
1734  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1735  * transparent huge pages.  See the PageTransHuge() documentation for more
1736  * details.
1737  */
1738 int PageHuge(struct page *page)
1739 {
1740         if (!PageCompound(page))
1741                 return 0;
1742
1743         page = compound_head(page);
1744         return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1745 }
1746 EXPORT_SYMBOL_GPL(PageHuge);
1747
1748 /*
1749  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1750  * normal or transparent huge pages.
1751  */
1752 int PageHeadHuge(struct page *page_head)
1753 {
1754         if (!PageHead(page_head))
1755                 return 0;
1756
1757         return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1758 }
1759
1760 /*
1761  * Find and lock address space (mapping) in write mode.
1762  *
1763  * Upon entry, the page is locked which means that page_mapping() is
1764  * stable.  Due to locking order, we can only trylock_write.  If we can
1765  * not get the lock, simply return NULL to caller.
1766  */
1767 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1768 {
1769         struct address_space *mapping = page_mapping(hpage);
1770
1771         if (!mapping)
1772                 return mapping;
1773
1774         if (i_mmap_trylock_write(mapping))
1775                 return mapping;
1776
1777         return NULL;
1778 }
1779
1780 pgoff_t hugetlb_basepage_index(struct page *page)
1781 {
1782         struct page *page_head = compound_head(page);
1783         pgoff_t index = page_index(page_head);
1784         unsigned long compound_idx;
1785
1786         if (compound_order(page_head) >= MAX_ORDER)
1787                 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1788         else
1789                 compound_idx = page - page_head;
1790
1791         return (index << compound_order(page_head)) + compound_idx;
1792 }
1793
1794 static struct page *alloc_buddy_huge_page(struct hstate *h,
1795                 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1796                 nodemask_t *node_alloc_noretry)
1797 {
1798         int order = huge_page_order(h);
1799         struct page *page;
1800         bool alloc_try_hard = true;
1801
1802         /*
1803          * By default we always try hard to allocate the page with
1804          * __GFP_RETRY_MAYFAIL flag.  However, if we are allocating pages in
1805          * a loop (to adjust global huge page counts) and previous allocation
1806          * failed, do not continue to try hard on the same node.  Use the
1807          * node_alloc_noretry bitmap to manage this state information.
1808          */
1809         if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1810                 alloc_try_hard = false;
1811         gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1812         if (alloc_try_hard)
1813                 gfp_mask |= __GFP_RETRY_MAYFAIL;
1814         if (nid == NUMA_NO_NODE)
1815                 nid = numa_mem_id();
1816         page = __alloc_pages(gfp_mask, order, nid, nmask);
1817         if (page)
1818                 __count_vm_event(HTLB_BUDDY_PGALLOC);
1819         else
1820                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1821
1822         /*
1823          * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1824          * indicates an overall state change.  Clear bit so that we resume
1825          * normal 'try hard' allocations.
1826          */
1827         if (node_alloc_noretry && page && !alloc_try_hard)
1828                 node_clear(nid, *node_alloc_noretry);
1829
1830         /*
1831          * If we tried hard to get a page but failed, set bit so that
1832          * subsequent attempts will not try as hard until there is an
1833          * overall state change.
1834          */
1835         if (node_alloc_noretry && !page && alloc_try_hard)
1836                 node_set(nid, *node_alloc_noretry);
1837
1838         return page;
1839 }
1840
1841 /*
1842  * Common helper to allocate a fresh hugetlb page. All specific allocators
1843  * should use this function to get new hugetlb pages
1844  */
1845 static struct page *alloc_fresh_huge_page(struct hstate *h,
1846                 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1847                 nodemask_t *node_alloc_noretry)
1848 {
1849         struct page *page;
1850         bool retry = false;
1851
1852 retry:
1853         if (hstate_is_gigantic(h))
1854                 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1855         else
1856                 page = alloc_buddy_huge_page(h, gfp_mask,
1857                                 nid, nmask, node_alloc_noretry);
1858         if (!page)
1859                 return NULL;
1860
1861         if (hstate_is_gigantic(h)) {
1862                 if (!prep_compound_gigantic_page(page, huge_page_order(h))) {
1863                         /*
1864                          * Rare failure to convert pages to compound page.
1865                          * Free pages and try again - ONCE!
1866                          */
1867                         free_gigantic_page(page, huge_page_order(h));
1868                         if (!retry) {
1869                                 retry = true;
1870                                 goto retry;
1871                         }
1872                         return NULL;
1873                 }
1874         }
1875         prep_new_huge_page(h, page, page_to_nid(page));
1876
1877         return page;
1878 }
1879
1880 /*
1881  * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1882  * manner.
1883  */
1884 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1885                                 nodemask_t *node_alloc_noretry)
1886 {
1887         struct page *page;
1888         int nr_nodes, node;
1889         gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1890
1891         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1892                 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1893                                                 node_alloc_noretry);
1894                 if (page)
1895                         break;
1896         }
1897
1898         if (!page)
1899                 return 0;
1900
1901         put_page(page); /* free it into the hugepage allocator */
1902
1903         return 1;
1904 }
1905
1906 /*
1907  * Remove huge page from pool from next node to free.  Attempt to keep
1908  * persistent huge pages more or less balanced over allowed nodes.
1909  * This routine only 'removes' the hugetlb page.  The caller must make
1910  * an additional call to free the page to low level allocators.
1911  * Called with hugetlb_lock locked.
1912  */
1913 static struct page *remove_pool_huge_page(struct hstate *h,
1914                                                 nodemask_t *nodes_allowed,
1915                                                  bool acct_surplus)
1916 {
1917         int nr_nodes, node;
1918         struct page *page = NULL;
1919
1920         lockdep_assert_held(&hugetlb_lock);
1921         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1922                 /*
1923                  * If we're returning unused surplus pages, only examine
1924                  * nodes with surplus pages.
1925                  */
1926                 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1927                     !list_empty(&h->hugepage_freelists[node])) {
1928                         page = list_entry(h->hugepage_freelists[node].next,
1929                                           struct page, lru);
1930                         remove_hugetlb_page(h, page, acct_surplus);
1931                         break;
1932                 }
1933         }
1934
1935         return page;
1936 }
1937
1938 /*
1939  * Dissolve a given free hugepage into free buddy pages. This function does
1940  * nothing for in-use hugepages and non-hugepages.
1941  * This function returns values like below:
1942  *
1943  *  -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
1944  *           when the system is under memory pressure and the feature of
1945  *           freeing unused vmemmap pages associated with each hugetlb page
1946  *           is enabled.
1947  *  -EBUSY:  failed to dissolved free hugepages or the hugepage is in-use
1948  *           (allocated or reserved.)
1949  *       0:  successfully dissolved free hugepages or the page is not a
1950  *           hugepage (considered as already dissolved)
1951  */
1952 int dissolve_free_huge_page(struct page *page)
1953 {
1954         int rc = -EBUSY;
1955
1956 retry:
1957         /* Not to disrupt normal path by vainly holding hugetlb_lock */
1958         if (!PageHuge(page))
1959                 return 0;
1960
1961         spin_lock_irq(&hugetlb_lock);
1962         if (!PageHuge(page)) {
1963                 rc = 0;
1964                 goto out;
1965         }
1966
1967         if (!page_count(page)) {
1968                 struct page *head = compound_head(page);
1969                 struct hstate *h = page_hstate(head);
1970                 if (h->free_huge_pages - h->resv_huge_pages == 0)
1971                         goto out;
1972
1973                 /*
1974                  * We should make sure that the page is already on the free list
1975                  * when it is dissolved.
1976                  */
1977                 if (unlikely(!HPageFreed(head))) {
1978                         spin_unlock_irq(&hugetlb_lock);
1979                         cond_resched();
1980
1981                         /*
1982                          * Theoretically, we should return -EBUSY when we
1983                          * encounter this race. In fact, we have a chance
1984                          * to successfully dissolve the page if we do a
1985                          * retry. Because the race window is quite small.
1986                          * If we seize this opportunity, it is an optimization
1987                          * for increasing the success rate of dissolving page.
1988                          */
1989                         goto retry;
1990                 }
1991
1992                 remove_hugetlb_page(h, head, false);
1993                 h->max_huge_pages--;
1994                 spin_unlock_irq(&hugetlb_lock);
1995
1996                 /*
1997                  * Normally update_and_free_page will allocate required vmemmmap
1998                  * before freeing the page.  update_and_free_page will fail to
1999                  * free the page if it can not allocate required vmemmap.  We
2000                  * need to adjust max_huge_pages if the page is not freed.
2001                  * Attempt to allocate vmemmmap here so that we can take
2002                  * appropriate action on failure.
2003                  */
2004                 rc = alloc_huge_page_vmemmap(h, head);
2005                 if (!rc) {
2006                         /*
2007                          * Move PageHWPoison flag from head page to the raw
2008                          * error page, which makes any subpages rather than
2009                          * the error page reusable.
2010                          */
2011                         if (PageHWPoison(head) && page != head) {
2012                                 SetPageHWPoison(page);
2013                                 ClearPageHWPoison(head);
2014                         }
2015                         update_and_free_page(h, head, false);
2016                 } else {
2017                         spin_lock_irq(&hugetlb_lock);
2018                         add_hugetlb_page(h, head, false);
2019                         h->max_huge_pages++;
2020                         spin_unlock_irq(&hugetlb_lock);
2021                 }
2022
2023                 return rc;
2024         }
2025 out:
2026         spin_unlock_irq(&hugetlb_lock);
2027         return rc;
2028 }
2029
2030 /*
2031  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2032  * make specified memory blocks removable from the system.
2033  * Note that this will dissolve a free gigantic hugepage completely, if any
2034  * part of it lies within the given range.
2035  * Also note that if dissolve_free_huge_page() returns with an error, all
2036  * free hugepages that were dissolved before that error are lost.
2037  */
2038 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2039 {
2040         unsigned long pfn;
2041         struct page *page;
2042         int rc = 0;
2043
2044         if (!hugepages_supported())
2045                 return rc;
2046
2047         for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
2048                 page = pfn_to_page(pfn);
2049                 rc = dissolve_free_huge_page(page);
2050                 if (rc)
2051                         break;
2052         }
2053
2054         return rc;
2055 }
2056
2057 /*
2058  * Allocates a fresh surplus page from the page allocator.
2059  */
2060 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
2061                 int nid, nodemask_t *nmask, bool zero_ref)
2062 {
2063         struct page *page = NULL;
2064         bool retry = false;
2065
2066         if (hstate_is_gigantic(h))
2067                 return NULL;
2068
2069         spin_lock_irq(&hugetlb_lock);
2070         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2071                 goto out_unlock;
2072         spin_unlock_irq(&hugetlb_lock);
2073
2074 retry:
2075         page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2076         if (!page)
2077                 return NULL;
2078
2079         spin_lock_irq(&hugetlb_lock);
2080         /*
2081          * We could have raced with the pool size change.
2082          * Double check that and simply deallocate the new page
2083          * if we would end up overcommiting the surpluses. Abuse
2084          * temporary page to workaround the nasty free_huge_page
2085          * codeflow
2086          */
2087         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2088                 SetHPageTemporary(page);
2089                 spin_unlock_irq(&hugetlb_lock);
2090                 put_page(page);
2091                 return NULL;
2092         }
2093
2094         if (zero_ref) {
2095                 /*
2096                  * Caller requires a page with zero ref count.
2097                  * We will drop ref count here.  If someone else is holding
2098                  * a ref, the page will be freed when they drop it.  Abuse
2099                  * temporary page flag to accomplish this.
2100                  */
2101                 SetHPageTemporary(page);
2102                 if (!put_page_testzero(page)) {
2103                         /*
2104                          * Unexpected inflated ref count on freshly allocated
2105                          * huge.  Retry once.
2106                          */
2107                         pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n");
2108                         spin_unlock_irq(&hugetlb_lock);
2109                         if (retry)
2110                                 return NULL;
2111
2112                         retry = true;
2113                         goto retry;
2114                 }
2115                 ClearHPageTemporary(page);
2116         }
2117
2118         h->surplus_huge_pages++;
2119         h->surplus_huge_pages_node[page_to_nid(page)]++;
2120
2121 out_unlock:
2122         spin_unlock_irq(&hugetlb_lock);
2123
2124         return page;
2125 }
2126
2127 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
2128                                      int nid, nodemask_t *nmask)
2129 {
2130         struct page *page;
2131
2132         if (hstate_is_gigantic(h))
2133                 return NULL;
2134
2135         page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2136         if (!page)
2137                 return NULL;
2138
2139         /*
2140          * We do not account these pages as surplus because they are only
2141          * temporary and will be released properly on the last reference
2142          */
2143         SetHPageTemporary(page);
2144
2145         return page;
2146 }
2147
2148 /*
2149  * Use the VMA's mpolicy to allocate a huge page from the buddy.
2150  */
2151 static
2152 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
2153                 struct vm_area_struct *vma, unsigned long addr)
2154 {
2155         struct page *page = NULL;
2156         struct mempolicy *mpol;
2157         gfp_t gfp_mask = htlb_alloc_mask(h);
2158         int nid;
2159         nodemask_t *nodemask;
2160
2161         nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2162         if (mpol_is_preferred_many(mpol)) {
2163                 gfp_t gfp = gfp_mask | __GFP_NOWARN;
2164
2165                 gfp &=  ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2166                 page = alloc_surplus_huge_page(h, gfp, nid, nodemask, false);
2167
2168                 /* Fallback to all nodes if page==NULL */
2169                 nodemask = NULL;
2170         }
2171
2172         if (!page)
2173                 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask, false);
2174         mpol_cond_put(mpol);
2175         return page;
2176 }
2177
2178 /* page migration callback function */
2179 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
2180                 nodemask_t *nmask, gfp_t gfp_mask)
2181 {
2182         spin_lock_irq(&hugetlb_lock);
2183         if (h->free_huge_pages - h->resv_huge_pages > 0) {
2184                 struct page *page;
2185
2186                 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
2187                 if (page) {
2188                         spin_unlock_irq(&hugetlb_lock);
2189                         return page;
2190                 }
2191         }
2192         spin_unlock_irq(&hugetlb_lock);
2193
2194         return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2195 }
2196
2197 /* mempolicy aware migration callback */
2198 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2199                 unsigned long address)
2200 {
2201         struct mempolicy *mpol;
2202         nodemask_t *nodemask;
2203         struct page *page;
2204         gfp_t gfp_mask;
2205         int node;
2206
2207         gfp_mask = htlb_alloc_mask(h);
2208         node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2209         page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
2210         mpol_cond_put(mpol);
2211
2212         return page;
2213 }
2214
2215 /*
2216  * Increase the hugetlb pool such that it can accommodate a reservation
2217  * of size 'delta'.
2218  */
2219 static int gather_surplus_pages(struct hstate *h, long delta)
2220         __must_hold(&hugetlb_lock)
2221 {
2222         struct list_head surplus_list;
2223         struct page *page, *tmp;
2224         int ret;
2225         long i;
2226         long needed, allocated;
2227         bool alloc_ok = true;
2228
2229         lockdep_assert_held(&hugetlb_lock);
2230         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2231         if (needed <= 0) {
2232                 h->resv_huge_pages += delta;
2233                 return 0;
2234         }
2235
2236         allocated = 0;
2237         INIT_LIST_HEAD(&surplus_list);
2238
2239         ret = -ENOMEM;
2240 retry:
2241         spin_unlock_irq(&hugetlb_lock);
2242         for (i = 0; i < needed; i++) {
2243                 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2244                                 NUMA_NO_NODE, NULL, true);
2245                 if (!page) {
2246                         alloc_ok = false;
2247                         break;
2248                 }
2249                 list_add(&page->lru, &surplus_list);
2250                 cond_resched();
2251         }
2252         allocated += i;
2253
2254         /*
2255          * After retaking hugetlb_lock, we need to recalculate 'needed'
2256          * because either resv_huge_pages or free_huge_pages may have changed.
2257          */
2258         spin_lock_irq(&hugetlb_lock);
2259         needed = (h->resv_huge_pages + delta) -
2260                         (h->free_huge_pages + allocated);
2261         if (needed > 0) {
2262                 if (alloc_ok)
2263                         goto retry;
2264                 /*
2265                  * We were not able to allocate enough pages to
2266                  * satisfy the entire reservation so we free what
2267                  * we've allocated so far.
2268                  */
2269                 goto free;
2270         }
2271         /*
2272          * The surplus_list now contains _at_least_ the number of extra pages
2273          * needed to accommodate the reservation.  Add the appropriate number
2274          * of pages to the hugetlb pool and free the extras back to the buddy
2275          * allocator.  Commit the entire reservation here to prevent another
2276          * process from stealing the pages as they are added to the pool but
2277          * before they are reserved.
2278          */
2279         needed += allocated;
2280         h->resv_huge_pages += delta;
2281         ret = 0;
2282
2283         /* Free the needed pages to the hugetlb pool */
2284         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2285                 if ((--needed) < 0)
2286                         break;
2287                 /* Add the page to the hugetlb allocator */
2288                 enqueue_huge_page(h, page);
2289         }
2290 free:
2291         spin_unlock_irq(&hugetlb_lock);
2292
2293         /*
2294          * Free unnecessary surplus pages to the buddy allocator.
2295          * Pages have no ref count, call free_huge_page directly.
2296          */
2297         list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2298                 free_huge_page(page);
2299         spin_lock_irq(&hugetlb_lock);
2300
2301         return ret;
2302 }
2303
2304 /*
2305  * This routine has two main purposes:
2306  * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2307  *    in unused_resv_pages.  This corresponds to the prior adjustments made
2308  *    to the associated reservation map.
2309  * 2) Free any unused surplus pages that may have been allocated to satisfy
2310  *    the reservation.  As many as unused_resv_pages may be freed.
2311  */
2312 static void return_unused_surplus_pages(struct hstate *h,
2313                                         unsigned long unused_resv_pages)
2314 {
2315         unsigned long nr_pages;
2316         struct page *page;
2317         LIST_HEAD(page_list);
2318
2319         lockdep_assert_held(&hugetlb_lock);
2320         /* Uncommit the reservation */
2321         h->resv_huge_pages -= unused_resv_pages;
2322
2323         /* Cannot return gigantic pages currently */
2324         if (hstate_is_gigantic(h))
2325                 goto out;
2326
2327         /*
2328          * Part (or even all) of the reservation could have been backed
2329          * by pre-allocated pages. Only free surplus pages.
2330          */
2331         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2332
2333         /*
2334          * We want to release as many surplus pages as possible, spread
2335          * evenly across all nodes with memory. Iterate across these nodes
2336          * until we can no longer free unreserved surplus pages. This occurs
2337          * when the nodes with surplus pages have no free pages.
2338          * remove_pool_huge_page() will balance the freed pages across the
2339          * on-line nodes with memory and will handle the hstate accounting.
2340          */
2341         while (nr_pages--) {
2342                 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2343                 if (!page)
2344                         goto out;
2345
2346                 list_add(&page->lru, &page_list);
2347         }
2348
2349 out:
2350         spin_unlock_irq(&hugetlb_lock);
2351         update_and_free_pages_bulk(h, &page_list);
2352         spin_lock_irq(&hugetlb_lock);
2353 }
2354
2355
2356 /*
2357  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2358  * are used by the huge page allocation routines to manage reservations.
2359  *
2360  * vma_needs_reservation is called to determine if the huge page at addr
2361  * within the vma has an associated reservation.  If a reservation is
2362  * needed, the value 1 is returned.  The caller is then responsible for
2363  * managing the global reservation and subpool usage counts.  After
2364  * the huge page has been allocated, vma_commit_reservation is called
2365  * to add the page to the reservation map.  If the page allocation fails,
2366  * the reservation must be ended instead of committed.  vma_end_reservation
2367  * is called in such cases.
2368  *
2369  * In the normal case, vma_commit_reservation returns the same value
2370  * as the preceding vma_needs_reservation call.  The only time this
2371  * is not the case is if a reserve map was changed between calls.  It
2372  * is the responsibility of the caller to notice the difference and
2373  * take appropriate action.
2374  *
2375  * vma_add_reservation is used in error paths where a reservation must
2376  * be restored when a newly allocated huge page must be freed.  It is
2377  * to be called after calling vma_needs_reservation to determine if a
2378  * reservation exists.
2379  *
2380  * vma_del_reservation is used in error paths where an entry in the reserve
2381  * map was created during huge page allocation and must be removed.  It is to
2382  * be called after calling vma_needs_reservation to determine if a reservation
2383  * exists.
2384  */
2385 enum vma_resv_mode {
2386         VMA_NEEDS_RESV,
2387         VMA_COMMIT_RESV,
2388         VMA_END_RESV,
2389         VMA_ADD_RESV,
2390         VMA_DEL_RESV,
2391 };
2392 static long __vma_reservation_common(struct hstate *h,
2393                                 struct vm_area_struct *vma, unsigned long addr,
2394                                 enum vma_resv_mode mode)
2395 {
2396         struct resv_map *resv;
2397         pgoff_t idx;
2398         long ret;
2399         long dummy_out_regions_needed;
2400
2401         resv = vma_resv_map(vma);
2402         if (!resv)
2403                 return 1;
2404
2405         idx = vma_hugecache_offset(h, vma, addr);
2406         switch (mode) {
2407         case VMA_NEEDS_RESV:
2408                 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2409                 /* We assume that vma_reservation_* routines always operate on
2410                  * 1 page, and that adding to resv map a 1 page entry can only
2411                  * ever require 1 region.
2412                  */
2413                 VM_BUG_ON(dummy_out_regions_needed != 1);
2414                 break;
2415         case VMA_COMMIT_RESV:
2416                 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2417                 /* region_add calls of range 1 should never fail. */
2418                 VM_BUG_ON(ret < 0);
2419                 break;
2420         case VMA_END_RESV:
2421                 region_abort(resv, idx, idx + 1, 1);
2422                 ret = 0;
2423                 break;
2424         case VMA_ADD_RESV:
2425                 if (vma->vm_flags & VM_MAYSHARE) {
2426                         ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2427                         /* region_add calls of range 1 should never fail. */
2428                         VM_BUG_ON(ret < 0);
2429                 } else {
2430                         region_abort(resv, idx, idx + 1, 1);
2431                         ret = region_del(resv, idx, idx + 1);
2432                 }
2433                 break;
2434         case VMA_DEL_RESV:
2435                 if (vma->vm_flags & VM_MAYSHARE) {
2436                         region_abort(resv, idx, idx + 1, 1);
2437                         ret = region_del(resv, idx, idx + 1);
2438                 } else {
2439                         ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2440                         /* region_add calls of range 1 should never fail. */
2441                         VM_BUG_ON(ret < 0);
2442                 }
2443                 break;
2444         default:
2445                 BUG();
2446         }
2447
2448         if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2449                 return ret;
2450         /*
2451          * We know private mapping must have HPAGE_RESV_OWNER set.
2452          *
2453          * In most cases, reserves always exist for private mappings.
2454          * However, a file associated with mapping could have been
2455          * hole punched or truncated after reserves were consumed.
2456          * As subsequent fault on such a range will not use reserves.
2457          * Subtle - The reserve map for private mappings has the
2458          * opposite meaning than that of shared mappings.  If NO
2459          * entry is in the reserve map, it means a reservation exists.
2460          * If an entry exists in the reserve map, it means the
2461          * reservation has already been consumed.  As a result, the
2462          * return value of this routine is the opposite of the
2463          * value returned from reserve map manipulation routines above.
2464          */
2465         if (ret > 0)
2466                 return 0;
2467         if (ret == 0)
2468                 return 1;
2469         return ret;
2470 }
2471
2472 static long vma_needs_reservation(struct hstate *h,
2473                         struct vm_area_struct *vma, unsigned long addr)
2474 {
2475         return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2476 }
2477
2478 static long vma_commit_reservation(struct hstate *h,
2479                         struct vm_area_struct *vma, unsigned long addr)
2480 {
2481         return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2482 }
2483
2484 static void vma_end_reservation(struct hstate *h,
2485                         struct vm_area_struct *vma, unsigned long addr)
2486 {
2487         (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2488 }
2489
2490 static long vma_add_reservation(struct hstate *h,
2491                         struct vm_area_struct *vma, unsigned long addr)
2492 {
2493         return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2494 }
2495
2496 static long vma_del_reservation(struct hstate *h,
2497                         struct vm_area_struct *vma, unsigned long addr)
2498 {
2499         return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2500 }
2501
2502 /*
2503  * This routine is called to restore reservation information on error paths.
2504  * It should ONLY be called for pages allocated via alloc_huge_page(), and
2505  * the hugetlb mutex should remain held when calling this routine.
2506  *
2507  * It handles two specific cases:
2508  * 1) A reservation was in place and the page consumed the reservation.
2509  *    HPageRestoreReserve is set in the page.
2510  * 2) No reservation was in place for the page, so HPageRestoreReserve is
2511  *    not set.  However, alloc_huge_page always updates the reserve map.
2512  *
2513  * In case 1, free_huge_page later in the error path will increment the
2514  * global reserve count.  But, free_huge_page does not have enough context
2515  * to adjust the reservation map.  This case deals primarily with private
2516  * mappings.  Adjust the reserve map here to be consistent with global
2517  * reserve count adjustments to be made by free_huge_page.  Make sure the
2518  * reserve map indicates there is a reservation present.
2519  *
2520  * In case 2, simply undo reserve map modifications done by alloc_huge_page.
2521  */
2522 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2523                         unsigned long address, struct page *page)
2524 {
2525         long rc = vma_needs_reservation(h, vma, address);
2526
2527         if (HPageRestoreReserve(page)) {
2528                 if (unlikely(rc < 0))
2529                         /*
2530                          * Rare out of memory condition in reserve map
2531                          * manipulation.  Clear HPageRestoreReserve so that
2532                          * global reserve count will not be incremented
2533                          * by free_huge_page.  This will make it appear
2534                          * as though the reservation for this page was
2535                          * consumed.  This may prevent the task from
2536                          * faulting in the page at a later time.  This
2537                          * is better than inconsistent global huge page
2538                          * accounting of reserve counts.
2539                          */
2540                         ClearHPageRestoreReserve(page);
2541                 else if (rc)
2542                         (void)vma_add_reservation(h, vma, address);
2543                 else
2544                         vma_end_reservation(h, vma, address);
2545         } else {
2546                 if (!rc) {
2547                         /*
2548                          * This indicates there is an entry in the reserve map
2549                          * not added by alloc_huge_page.  We know it was added
2550                          * before the alloc_huge_page call, otherwise
2551                          * HPageRestoreReserve would be set on the page.
2552                          * Remove the entry so that a subsequent allocation
2553                          * does not consume a reservation.
2554                          */
2555                         rc = vma_del_reservation(h, vma, address);
2556                         if (rc < 0)
2557                                 /*
2558                                  * VERY rare out of memory condition.  Since
2559                                  * we can not delete the entry, set
2560                                  * HPageRestoreReserve so that the reserve
2561                                  * count will be incremented when the page
2562                                  * is freed.  This reserve will be consumed
2563                                  * on a subsequent allocation.
2564                                  */
2565                                 SetHPageRestoreReserve(page);
2566                 } else if (rc < 0) {
2567                         /*
2568                          * Rare out of memory condition from
2569                          * vma_needs_reservation call.  Memory allocation is
2570                          * only attempted if a new entry is needed.  Therefore,
2571                          * this implies there is not an entry in the
2572                          * reserve map.
2573                          *
2574                          * For shared mappings, no entry in the map indicates
2575                          * no reservation.  We are done.
2576                          */
2577                         if (!(vma->vm_flags & VM_MAYSHARE))
2578                                 /*
2579                                  * For private mappings, no entry indicates
2580                                  * a reservation is present.  Since we can
2581                                  * not add an entry, set SetHPageRestoreReserve
2582                                  * on the page so reserve count will be
2583                                  * incremented when freed.  This reserve will
2584                                  * be consumed on a subsequent allocation.
2585                                  */
2586                                 SetHPageRestoreReserve(page);
2587                 } else
2588                         /*
2589                          * No reservation present, do nothing
2590                          */
2591                          vma_end_reservation(h, vma, address);
2592         }
2593 }
2594
2595 /*
2596  * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one
2597  * @h: struct hstate old page belongs to
2598  * @old_page: Old page to dissolve
2599  * @list: List to isolate the page in case we need to
2600  * Returns 0 on success, otherwise negated error.
2601  */
2602 static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page,
2603                                         struct list_head *list)
2604 {
2605         gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2606         int nid = page_to_nid(old_page);
2607         bool alloc_retry = false;
2608         struct page *new_page;
2609         int ret = 0;
2610
2611         /*
2612          * Before dissolving the page, we need to allocate a new one for the
2613          * pool to remain stable.  Here, we allocate the page and 'prep' it
2614          * by doing everything but actually updating counters and adding to
2615          * the pool.  This simplifies and let us do most of the processing
2616          * under the lock.
2617          */
2618 alloc_retry:
2619         new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL);
2620         if (!new_page)
2621                 return -ENOMEM;
2622         /*
2623          * If all goes well, this page will be directly added to the free
2624          * list in the pool.  For this the ref count needs to be zero.
2625          * Attempt to drop now, and retry once if needed.  It is VERY
2626          * unlikely there is another ref on the page.
2627          *
2628          * If someone else has a reference to the page, it will be freed
2629          * when they drop their ref.  Abuse temporary page flag to accomplish
2630          * this.  Retry once if there is an inflated ref count.
2631          */
2632         SetHPageTemporary(new_page);
2633         if (!put_page_testzero(new_page)) {
2634                 if (alloc_retry)
2635                         return -EBUSY;
2636
2637                 alloc_retry = true;
2638                 goto alloc_retry;
2639         }
2640         ClearHPageTemporary(new_page);
2641
2642         __prep_new_huge_page(h, new_page);
2643
2644 retry:
2645         spin_lock_irq(&hugetlb_lock);
2646         if (!PageHuge(old_page)) {
2647                 /*
2648                  * Freed from under us. Drop new_page too.
2649                  */
2650                 goto free_new;
2651         } else if (page_count(old_page)) {
2652                 /*
2653                  * Someone has grabbed the page, try to isolate it here.
2654                  * Fail with -EBUSY if not possible.
2655                  */
2656                 spin_unlock_irq(&hugetlb_lock);
2657                 if (!isolate_huge_page(old_page, list))
2658                         ret = -EBUSY;
2659                 spin_lock_irq(&hugetlb_lock);
2660                 goto free_new;
2661         } else if (!HPageFreed(old_page)) {
2662                 /*
2663                  * Page's refcount is 0 but it has not been enqueued in the
2664                  * freelist yet. Race window is small, so we can succeed here if
2665                  * we retry.
2666                  */
2667                 spin_unlock_irq(&hugetlb_lock);
2668                 cond_resched();
2669                 goto retry;
2670         } else {
2671                 /*
2672                  * Ok, old_page is still a genuine free hugepage. Remove it from
2673                  * the freelist and decrease the counters. These will be
2674                  * incremented again when calling __prep_account_new_huge_page()
2675                  * and enqueue_huge_page() for new_page. The counters will remain
2676                  * stable since this happens under the lock.
2677                  */
2678                 remove_hugetlb_page(h, old_page, false);
2679
2680                 /*
2681                  * Ref count on new page is already zero as it was dropped
2682                  * earlier.  It can be directly added to the pool free list.
2683                  */
2684                 __prep_account_new_huge_page(h, nid);
2685                 enqueue_huge_page(h, new_page);
2686
2687                 /*
2688                  * Pages have been replaced, we can safely free the old one.
2689                  */
2690                 spin_unlock_irq(&hugetlb_lock);
2691                 update_and_free_page(h, old_page, false);
2692         }
2693
2694         return ret;
2695
2696 free_new:
2697         spin_unlock_irq(&hugetlb_lock);
2698         /* Page has a zero ref count, but needs a ref to be freed */
2699         set_page_refcounted(new_page);
2700         update_and_free_page(h, new_page, false);
2701
2702         return ret;
2703 }
2704
2705 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
2706 {
2707         struct hstate *h;
2708         struct page *head;
2709         int ret = -EBUSY;
2710
2711         /*
2712          * The page might have been dissolved from under our feet, so make sure
2713          * to carefully check the state under the lock.
2714          * Return success when racing as if we dissolved the page ourselves.
2715          */
2716         spin_lock_irq(&hugetlb_lock);
2717         if (PageHuge(page)) {
2718                 head = compound_head(page);
2719                 h = page_hstate(head);
2720         } else {
2721                 spin_unlock_irq(&hugetlb_lock);
2722                 return 0;
2723         }
2724         spin_unlock_irq(&hugetlb_lock);
2725
2726         /*
2727          * Fence off gigantic pages as there is a cyclic dependency between
2728          * alloc_contig_range and them. Return -ENOMEM as this has the effect
2729          * of bailing out right away without further retrying.
2730          */
2731         if (hstate_is_gigantic(h))
2732                 return -ENOMEM;
2733
2734         if (page_count(head) && isolate_huge_page(head, list))
2735                 ret = 0;
2736         else if (!page_count(head))
2737                 ret = alloc_and_dissolve_huge_page(h, head, list);
2738
2739         return ret;
2740 }
2741
2742 struct page *alloc_huge_page(struct vm_area_struct *vma,
2743                                     unsigned long addr, int avoid_reserve)
2744 {
2745         struct hugepage_subpool *spool = subpool_vma(vma);
2746         struct hstate *h = hstate_vma(vma);
2747         struct page *page;
2748         long map_chg, map_commit;
2749         long gbl_chg;
2750         int ret, idx;
2751         struct hugetlb_cgroup *h_cg;
2752         bool deferred_reserve;
2753
2754         idx = hstate_index(h);
2755         /*
2756          * Examine the region/reserve map to determine if the process
2757          * has a reservation for the page to be allocated.  A return
2758          * code of zero indicates a reservation exists (no change).
2759          */
2760         map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2761         if (map_chg < 0)
2762                 return ERR_PTR(-ENOMEM);
2763
2764         /*
2765          * Processes that did not create the mapping will have no
2766          * reserves as indicated by the region/reserve map. Check
2767          * that the allocation will not exceed the subpool limit.
2768          * Allocations for MAP_NORESERVE mappings also need to be
2769          * checked against any subpool limit.
2770          */
2771         if (map_chg || avoid_reserve) {
2772                 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2773                 if (gbl_chg < 0) {
2774                         vma_end_reservation(h, vma, addr);
2775                         return ERR_PTR(-ENOSPC);
2776                 }
2777
2778                 /*
2779                  * Even though there was no reservation in the region/reserve
2780                  * map, there could be reservations associated with the
2781                  * subpool that can be used.  This would be indicated if the
2782                  * return value of hugepage_subpool_get_pages() is zero.
2783                  * However, if avoid_reserve is specified we still avoid even
2784                  * the subpool reservations.
2785                  */
2786                 if (avoid_reserve)
2787                         gbl_chg = 1;
2788         }
2789
2790         /* If this allocation is not consuming a reservation, charge it now.
2791          */
2792         deferred_reserve = map_chg || avoid_reserve;
2793         if (deferred_reserve) {
2794                 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2795                         idx, pages_per_huge_page(h), &h_cg);
2796                 if (ret)
2797                         goto out_subpool_put;
2798         }
2799
2800         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2801         if (ret)
2802                 goto out_uncharge_cgroup_reservation;
2803
2804         spin_lock_irq(&hugetlb_lock);
2805         /*
2806          * glb_chg is passed to indicate whether or not a page must be taken
2807          * from the global free pool (global change).  gbl_chg == 0 indicates
2808          * a reservation exists for the allocation.
2809          */
2810         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2811         if (!page) {
2812                 spin_unlock_irq(&hugetlb_lock);
2813                 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2814                 if (!page)
2815                         goto out_uncharge_cgroup;
2816                 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2817                         SetHPageRestoreReserve(page);
2818                         h->resv_huge_pages--;
2819                 }
2820                 spin_lock_irq(&hugetlb_lock);
2821                 list_add(&page->lru, &h->hugepage_activelist);
2822                 /* Fall through */
2823         }
2824         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2825         /* If allocation is not consuming a reservation, also store the
2826          * hugetlb_cgroup pointer on the page.
2827          */
2828         if (deferred_reserve) {
2829                 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2830                                                   h_cg, page);
2831         }
2832
2833         spin_unlock_irq(&hugetlb_lock);
2834
2835         hugetlb_set_page_subpool(page, spool);
2836
2837         map_commit = vma_commit_reservation(h, vma, addr);
2838         if (unlikely(map_chg > map_commit)) {
2839                 /*
2840                  * The page was added to the reservation map between
2841                  * vma_needs_reservation and vma_commit_reservation.
2842                  * This indicates a race with hugetlb_reserve_pages.
2843                  * Adjust for the subpool count incremented above AND
2844                  * in hugetlb_reserve_pages for the same page.  Also,
2845                  * the reservation count added in hugetlb_reserve_pages
2846                  * no longer applies.
2847                  */
2848                 long rsv_adjust;
2849
2850                 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2851                 hugetlb_acct_memory(h, -rsv_adjust);
2852                 if (deferred_reserve)
2853                         hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2854                                         pages_per_huge_page(h), page);
2855         }
2856         return page;
2857
2858 out_uncharge_cgroup:
2859         hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2860 out_uncharge_cgroup_reservation:
2861         if (deferred_reserve)
2862                 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2863                                                     h_cg);
2864 out_subpool_put:
2865         if (map_chg || avoid_reserve)
2866                 hugepage_subpool_put_pages(spool, 1);
2867         vma_end_reservation(h, vma, addr);
2868         return ERR_PTR(-ENOSPC);
2869 }
2870
2871 int alloc_bootmem_huge_page(struct hstate *h)
2872         __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2873 int __alloc_bootmem_huge_page(struct hstate *h)
2874 {
2875         struct huge_bootmem_page *m;
2876         int nr_nodes, node;
2877
2878         for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2879                 void *addr;
2880
2881                 addr = memblock_alloc_try_nid_raw(
2882                                 huge_page_size(h), huge_page_size(h),
2883                                 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2884                 if (addr) {
2885                         /*
2886                          * Use the beginning of the huge page to store the
2887                          * huge_bootmem_page struct (until gather_bootmem
2888                          * puts them into the mem_map).
2889                          */
2890                         m = addr;
2891                         goto found;
2892                 }
2893         }
2894         return 0;
2895
2896 found:
2897         BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2898         /* Put them into a private list first because mem_map is not up yet */
2899         INIT_LIST_HEAD(&m->list);
2900         list_add(&m->list, &huge_boot_pages);
2901         m->hstate = h;
2902         return 1;
2903 }
2904
2905 /*
2906  * Put bootmem huge pages into the standard lists after mem_map is up.
2907  * Note: This only applies to gigantic (order > MAX_ORDER) pages.
2908  */
2909 static void __init gather_bootmem_prealloc(void)
2910 {
2911         struct huge_bootmem_page *m;
2912
2913         list_for_each_entry(m, &huge_boot_pages, list) {
2914                 struct page *page = virt_to_page(m);
2915                 struct hstate *h = m->hstate;
2916
2917                 VM_BUG_ON(!hstate_is_gigantic(h));
2918                 WARN_ON(page_count(page) != 1);
2919                 if (prep_compound_gigantic_page(page, huge_page_order(h))) {
2920                         WARN_ON(PageReserved(page));
2921                         prep_new_huge_page(h, page, page_to_nid(page));
2922                         put_page(page); /* add to the hugepage allocator */
2923                 } else {
2924                         /* VERY unlikely inflated ref count on a tail page */
2925                         free_gigantic_page(page, huge_page_order(h));
2926                 }
2927
2928                 /*
2929                  * We need to restore the 'stolen' pages to totalram_pages
2930                  * in order to fix confusing memory reports from free(1) and
2931                  * other side-effects, like CommitLimit going negative.
2932                  */
2933                 adjust_managed_page_count(page, pages_per_huge_page(h));
2934                 cond_resched();
2935         }
2936 }
2937
2938 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2939 {
2940         unsigned long i;
2941         nodemask_t *node_alloc_noretry;
2942
2943         if (!hstate_is_gigantic(h)) {
2944                 /*
2945                  * Bit mask controlling how hard we retry per-node allocations.
2946                  * Ignore errors as lower level routines can deal with
2947                  * node_alloc_noretry == NULL.  If this kmalloc fails at boot
2948                  * time, we are likely in bigger trouble.
2949                  */
2950                 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2951                                                 GFP_KERNEL);
2952         } else {
2953                 /* allocations done at boot time */
2954                 node_alloc_noretry = NULL;
2955         }
2956
2957         /* bit mask controlling how hard we retry per-node allocations */
2958         if (node_alloc_noretry)
2959                 nodes_clear(*node_alloc_noretry);
2960
2961         for (i = 0; i < h->max_huge_pages; ++i) {
2962                 if (hstate_is_gigantic(h)) {
2963                         if (hugetlb_cma_size) {
2964                                 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2965                                 goto free;
2966                         }
2967                         if (!alloc_bootmem_huge_page(h))
2968                                 break;
2969                 } else if (!alloc_pool_huge_page(h,
2970                                          &node_states[N_MEMORY],
2971                                          node_alloc_noretry))
2972                         break;
2973                 cond_resched();
2974         }
2975         if (i < h->max_huge_pages) {
2976                 char buf[32];
2977
2978                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2979                 pr_warn("HugeTLB: allocating %lu of page size %s failed.  Only allocated %lu hugepages.\n",
2980                         h->max_huge_pages, buf, i);
2981                 h->max_huge_pages = i;
2982         }
2983 free:
2984         kfree(node_alloc_noretry);
2985 }
2986
2987 static void __init hugetlb_init_hstates(void)
2988 {
2989         struct hstate *h;
2990
2991         for_each_hstate(h) {
2992                 if (minimum_order > huge_page_order(h))
2993                         minimum_order = huge_page_order(h);
2994
2995                 /* oversize hugepages were init'ed in early boot */
2996                 if (!hstate_is_gigantic(h))
2997                         hugetlb_hstate_alloc_pages(h);
2998         }
2999         VM_BUG_ON(minimum_order == UINT_MAX);
3000 }
3001
3002 static void __init report_hugepages(void)
3003 {
3004         struct hstate *h;
3005
3006         for_each_hstate(h) {
3007                 char buf[32];
3008
3009                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3010                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
3011                         buf, h->free_huge_pages);
3012         }
3013 }
3014
3015 #ifdef CONFIG_HIGHMEM
3016 static void try_to_free_low(struct hstate *h, unsigned long count,
3017                                                 nodemask_t *nodes_allowed)
3018 {
3019         int i;
3020         LIST_HEAD(page_list);
3021
3022         lockdep_assert_held(&hugetlb_lock);
3023         if (hstate_is_gigantic(h))
3024                 return;
3025
3026         /*
3027          * Collect pages to be freed on a list, and free after dropping lock
3028          */
3029         for_each_node_mask(i, *nodes_allowed) {
3030                 struct page *page, *next;
3031                 struct list_head *freel = &h->hugepage_freelists[i];
3032                 list_for_each_entry_safe(page, next, freel, lru) {
3033                         if (count >= h->nr_huge_pages)
3034                                 goto out;
3035                         if (PageHighMem(page))
3036                                 continue;
3037                         remove_hugetlb_page(h, page, false);
3038                         list_add(&page->lru, &page_list);
3039                 }
3040         }
3041
3042 out:
3043         spin_unlock_irq(&hugetlb_lock);
3044         update_and_free_pages_bulk(h, &page_list);
3045         spin_lock_irq(&hugetlb_lock);
3046 }
3047 #else
3048 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3049                                                 nodemask_t *nodes_allowed)
3050 {
3051 }
3052 #endif
3053
3054 /*
3055  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
3056  * balanced by operating on them in a round-robin fashion.
3057  * Returns 1 if an adjustment was made.
3058  */
3059 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3060                                 int delta)
3061 {
3062         int nr_nodes, node;
3063
3064         lockdep_assert_held(&hugetlb_lock);
3065         VM_BUG_ON(delta != -1 && delta != 1);
3066
3067         if (delta < 0) {
3068                 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3069                         if (h->surplus_huge_pages_node[node])
3070                                 goto found;
3071                 }
3072         } else {
3073                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3074                         if (h->surplus_huge_pages_node[node] <
3075                                         h->nr_huge_pages_node[node])
3076                                 goto found;
3077                 }
3078         }
3079         return 0;
3080
3081 found:
3082         h->surplus_huge_pages += delta;
3083         h->surplus_huge_pages_node[node] += delta;
3084         return 1;
3085 }
3086
3087 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3088 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3089                               nodemask_t *nodes_allowed)
3090 {
3091         unsigned long min_count, ret;
3092         struct page *page;
3093         LIST_HEAD(page_list);
3094         NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3095
3096         /*
3097          * Bit mask controlling how hard we retry per-node allocations.
3098          * If we can not allocate the bit mask, do not attempt to allocate
3099          * the requested huge pages.
3100          */
3101         if (node_alloc_noretry)
3102                 nodes_clear(*node_alloc_noretry);
3103         else
3104                 return -ENOMEM;
3105
3106         /*
3107          * resize_lock mutex prevents concurrent adjustments to number of
3108          * pages in hstate via the proc/sysfs interfaces.
3109          */
3110         mutex_lock(&h->resize_lock);
3111         flush_free_hpage_work(h);
3112         spin_lock_irq(&hugetlb_lock);
3113
3114         /*
3115          * Check for a node specific request.
3116          * Changing node specific huge page count may require a corresponding
3117          * change to the global count.  In any case, the passed node mask
3118          * (nodes_allowed) will restrict alloc/free to the specified node.
3119          */
3120         if (nid != NUMA_NO_NODE) {
3121                 unsigned long old_count = count;
3122
3123                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3124                 /*
3125                  * User may have specified a large count value which caused the
3126                  * above calculation to overflow.  In this case, they wanted
3127                  * to allocate as many huge pages as possible.  Set count to
3128                  * largest possible value to align with their intention.
3129                  */
3130                 if (count < old_count)
3131                         count = ULONG_MAX;
3132         }
3133
3134         /*
3135          * Gigantic pages runtime allocation depend on the capability for large
3136          * page range allocation.
3137          * If the system does not provide this feature, return an error when
3138          * the user tries to allocate gigantic pages but let the user free the
3139          * boottime allocated gigantic pages.
3140          */
3141         if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3142                 if (count > persistent_huge_pages(h)) {
3143                         spin_unlock_irq(&hugetlb_lock);
3144                         mutex_unlock(&h->resize_lock);
3145                         NODEMASK_FREE(node_alloc_noretry);
3146                         return -EINVAL;
3147                 }
3148                 /* Fall through to decrease pool */
3149         }
3150
3151         /*
3152          * Increase the pool size
3153          * First take pages out of surplus state.  Then make up the
3154          * remaining difference by allocating fresh huge pages.
3155          *
3156          * We might race with alloc_surplus_huge_page() here and be unable
3157          * to convert a surplus huge page to a normal huge page. That is
3158          * not critical, though, it just means the overall size of the
3159          * pool might be one hugepage larger than it needs to be, but
3160          * within all the constraints specified by the sysctls.
3161          */
3162         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3163                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3164                         break;
3165         }
3166
3167         while (count > persistent_huge_pages(h)) {
3168                 /*
3169                  * If this allocation races such that we no longer need the
3170                  * page, free_huge_page will handle it by freeing the page
3171                  * and reducing the surplus.
3172                  */
3173                 spin_unlock_irq(&hugetlb_lock);
3174
3175                 /* yield cpu to avoid soft lockup */
3176                 cond_resched();
3177
3178                 ret = alloc_pool_huge_page(h, nodes_allowed,
3179                                                 node_alloc_noretry);
3180                 spin_lock_irq(&hugetlb_lock);
3181                 if (!ret)
3182                         goto out;
3183
3184                 /* Bail for signals. Probably ctrl-c from user */
3185                 if (signal_pending(current))
3186                         goto out;
3187         }
3188
3189         /*
3190          * Decrease the pool size
3191          * First return free pages to the buddy allocator (being careful
3192          * to keep enough around to satisfy reservations).  Then place
3193          * pages into surplus state as needed so the pool will shrink
3194          * to the desired size as pages become free.
3195          *
3196          * By placing pages into the surplus state independent of the
3197          * overcommit value, we are allowing the surplus pool size to
3198          * exceed overcommit. There are few sane options here. Since
3199          * alloc_surplus_huge_page() is checking the global counter,
3200          * though, we'll note that we're not allowed to exceed surplus
3201          * and won't grow the pool anywhere else. Not until one of the
3202          * sysctls are changed, or the surplus pages go out of use.
3203          */
3204         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3205         min_count = max(count, min_count);
3206         try_to_free_low(h, min_count, nodes_allowed);
3207
3208         /*
3209          * Collect pages to be removed on list without dropping lock
3210          */
3211         while (min_count < persistent_huge_pages(h)) {
3212                 page = remove_pool_huge_page(h, nodes_allowed, 0);
3213                 if (!page)
3214                         break;
3215
3216                 list_add(&page->lru, &page_list);
3217         }
3218         /* free the pages after dropping lock */
3219         spin_unlock_irq(&hugetlb_lock);
3220         update_and_free_pages_bulk(h, &page_list);
3221         flush_free_hpage_work(h);
3222         spin_lock_irq(&hugetlb_lock);
3223
3224         while (count < persistent_huge_pages(h)) {
3225                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3226                         break;
3227         }
3228 out:
3229         h->max_huge_pages = persistent_huge_pages(h);
3230         spin_unlock_irq(&hugetlb_lock);
3231         mutex_unlock(&h->resize_lock);
3232
3233         NODEMASK_FREE(node_alloc_noretry);
3234
3235         return 0;
3236 }
3237
3238 #define HSTATE_ATTR_RO(_name) \
3239         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3240
3241 #define HSTATE_ATTR(_name) \
3242         static struct kobj_attribute _name##_attr = \
3243                 __ATTR(_name, 0644, _name##_show, _name##_store)
3244
3245 static struct kobject *hugepages_kobj;
3246 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3247
3248 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3249
3250 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3251 {
3252         int i;
3253
3254         for (i = 0; i < HUGE_MAX_HSTATE; i++)
3255                 if (hstate_kobjs[i] == kobj) {
3256                         if (nidp)
3257                                 *nidp = NUMA_NO_NODE;
3258                         return &hstates[i];
3259                 }
3260
3261         return kobj_to_node_hstate(kobj, nidp);
3262 }
3263
3264 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3265                                         struct kobj_attribute *attr, char *buf)
3266 {
3267         struct hstate *h;
3268         unsigned long nr_huge_pages;
3269         int nid;
3270
3271         h = kobj_to_hstate(kobj, &nid);
3272         if (nid == NUMA_NO_NODE)
3273                 nr_huge_pages = h->nr_huge_pages;
3274         else
3275                 nr_huge_pages = h->nr_huge_pages_node[nid];
3276
3277         return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3278 }
3279
3280 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3281                                            struct hstate *h, int nid,
3282                                            unsigned long count, size_t len)
3283 {
3284         int err;
3285         nodemask_t nodes_allowed, *n_mask;
3286
3287         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3288                 return -EINVAL;
3289
3290         if (nid == NUMA_NO_NODE) {
3291                 /*
3292                  * global hstate attribute
3293                  */
3294                 if (!(obey_mempolicy &&
3295                                 init_nodemask_of_mempolicy(&nodes_allowed)))
3296                         n_mask = &node_states[N_MEMORY];
3297                 else
3298                         n_mask = &nodes_allowed;
3299         } else {
3300                 /*
3301                  * Node specific request.  count adjustment happens in
3302                  * set_max_huge_pages() after acquiring hugetlb_lock.
3303                  */
3304                 init_nodemask_of_node(&nodes_allowed, nid);
3305                 n_mask = &nodes_allowed;
3306         }
3307
3308         err = set_max_huge_pages(h, count, nid, n_mask);
3309
3310         return err ? err : len;
3311 }
3312
3313 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3314                                          struct kobject *kobj, const char *buf,
3315                                          size_t len)
3316 {
3317         struct hstate *h;
3318         unsigned long count;
3319         int nid;
3320         int err;
3321
3322         err = kstrtoul(buf, 10, &count);
3323         if (err)
3324                 return err;
3325
3326         h = kobj_to_hstate(kobj, &nid);
3327         return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3328 }
3329
3330 static ssize_t nr_hugepages_show(struct kobject *kobj,
3331                                        struct kobj_attribute *attr, char *buf)
3332 {
3333         return nr_hugepages_show_common(kobj, attr, buf);
3334 }
3335
3336 static ssize_t nr_hugepages_store(struct kobject *kobj,
3337                struct kobj_attribute *attr, const char *buf, size_t len)
3338 {
3339         return nr_hugepages_store_common(false, kobj, buf, len);
3340 }
3341 HSTATE_ATTR(nr_hugepages);
3342
3343 #ifdef CONFIG_NUMA
3344
3345 /*
3346  * hstate attribute for optionally mempolicy-based constraint on persistent
3347  * huge page alloc/free.
3348  */
3349 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3350                                            struct kobj_attribute *attr,
3351                                            char *buf)
3352 {
3353         return nr_hugepages_show_common(kobj, attr, buf);
3354 }
3355
3356 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3357                struct kobj_attribute *attr, const char *buf, size_t len)
3358 {
3359         return nr_hugepages_store_common(true, kobj, buf, len);
3360 }
3361 HSTATE_ATTR(nr_hugepages_mempolicy);
3362 #endif
3363
3364
3365 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3366                                         struct kobj_attribute *attr, char *buf)
3367 {
3368         struct hstate *h = kobj_to_hstate(kobj, NULL);
3369         return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3370 }
3371
3372 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3373                 struct kobj_attribute *attr, const char *buf, size_t count)
3374 {
3375         int err;
3376         unsigned long input;
3377         struct hstate *h = kobj_to_hstate(kobj, NULL);
3378
3379         if (hstate_is_gigantic(h))
3380                 return -EINVAL;
3381
3382         err = kstrtoul(buf, 10, &input);
3383         if (err)
3384                 return err;
3385
3386         spin_lock_irq(&hugetlb_lock);
3387         h->nr_overcommit_huge_pages = input;
3388         spin_unlock_irq(&hugetlb_lock);
3389
3390         return count;
3391 }
3392 HSTATE_ATTR(nr_overcommit_hugepages);
3393
3394 static ssize_t free_hugepages_show(struct kobject *kobj,
3395                                         struct kobj_attribute *attr, char *buf)
3396 {
3397         struct hstate *h;
3398         unsigned long free_huge_pages;
3399         int nid;
3400
3401         h = kobj_to_hstate(kobj, &nid);
3402         if (nid == NUMA_NO_NODE)
3403                 free_huge_pages = h->free_huge_pages;
3404         else
3405                 free_huge_pages = h->free_huge_pages_node[nid];
3406
3407         return sysfs_emit(buf, "%lu\n", free_huge_pages);
3408 }
3409 HSTATE_ATTR_RO(free_hugepages);
3410
3411 static ssize_t resv_hugepages_show(struct kobject *kobj,
3412                                         struct kobj_attribute *attr, char *buf)
3413 {
3414         struct hstate *h = kobj_to_hstate(kobj, NULL);
3415         return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3416 }
3417 HSTATE_ATTR_RO(resv_hugepages);
3418
3419 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3420                                         struct kobj_attribute *attr, char *buf)
3421 {
3422         struct hstate *h;
3423         unsigned long surplus_huge_pages;
3424         int nid;
3425
3426         h = kobj_to_hstate(kobj, &nid);
3427         if (nid == NUMA_NO_NODE)
3428                 surplus_huge_pages = h->surplus_huge_pages;
3429         else
3430                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3431
3432         return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3433 }
3434 HSTATE_ATTR_RO(surplus_hugepages);
3435
3436 static struct attribute *hstate_attrs[] = {
3437         &nr_hugepages_attr.attr,
3438         &nr_overcommit_hugepages_attr.attr,
3439         &free_hugepages_attr.attr,
3440         &resv_hugepages_attr.attr,
3441         &surplus_hugepages_attr.attr,
3442 #ifdef CONFIG_NUMA
3443         &nr_hugepages_mempolicy_attr.attr,
3444 #endif
3445         NULL,
3446 };
3447
3448 static const struct attribute_group hstate_attr_group = {
3449         .attrs = hstate_attrs,
3450 };
3451
3452 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3453                                     struct kobject **hstate_kobjs,
3454                                     const struct attribute_group *hstate_attr_group)
3455 {
3456         int retval;
3457         int hi = hstate_index(h);
3458
3459         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3460         if (!hstate_kobjs[hi])
3461                 return -ENOMEM;
3462
3463         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3464         if (retval) {
3465                 kobject_put(hstate_kobjs[hi]);
3466                 hstate_kobjs[hi] = NULL;
3467         }
3468
3469         return retval;
3470 }
3471
3472 static void __init hugetlb_sysfs_init(void)
3473 {
3474         struct hstate *h;
3475         int err;
3476
3477         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3478         if (!hugepages_kobj)
3479                 return;
3480
3481         for_each_hstate(h) {
3482                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3483                                          hstate_kobjs, &hstate_attr_group);
3484                 if (err)
3485                         pr_err("HugeTLB: Unable to add hstate %s", h->name);
3486         }
3487 }
3488
3489 #ifdef CONFIG_NUMA
3490
3491 /*
3492  * node_hstate/s - associate per node hstate attributes, via their kobjects,
3493  * with node devices in node_devices[] using a parallel array.  The array
3494  * index of a node device or _hstate == node id.
3495  * This is here to avoid any static dependency of the node device driver, in
3496  * the base kernel, on the hugetlb module.
3497  */
3498 struct node_hstate {
3499         struct kobject          *hugepages_kobj;
3500         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
3501 };
3502 static struct node_hstate node_hstates[MAX_NUMNODES];
3503
3504 /*
3505  * A subset of global hstate attributes for node devices
3506  */
3507 static struct attribute *per_node_hstate_attrs[] = {
3508         &nr_hugepages_attr.attr,
3509         &free_hugepages_attr.attr,
3510         &surplus_hugepages_attr.attr,
3511         NULL,
3512 };
3513
3514 static const struct attribute_group per_node_hstate_attr_group = {
3515         .attrs = per_node_hstate_attrs,
3516 };
3517
3518 /*
3519  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3520  * Returns node id via non-NULL nidp.
3521  */
3522 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3523 {
3524         int nid;
3525
3526         for (nid = 0; nid < nr_node_ids; nid++) {
3527                 struct node_hstate *nhs = &node_hstates[nid];
3528                 int i;
3529                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3530                         if (nhs->hstate_kobjs[i] == kobj) {
3531                                 if (nidp)
3532                                         *nidp = nid;
3533                                 return &hstates[i];
3534                         }
3535         }
3536
3537         BUG();
3538         return NULL;
3539 }
3540
3541 /*
3542  * Unregister hstate attributes from a single node device.
3543  * No-op if no hstate attributes attached.
3544  */
3545 static void hugetlb_unregister_node(struct node *node)
3546 {
3547         struct hstate *h;
3548         struct node_hstate *nhs = &node_hstates[node->dev.id];
3549
3550         if (!nhs->hugepages_kobj)
3551                 return;         /* no hstate attributes */
3552
3553         for_each_hstate(h) {
3554                 int idx = hstate_index(h);
3555                 if (nhs->hstate_kobjs[idx]) {
3556                         kobject_put(nhs->hstate_kobjs[idx]);
3557                         nhs->hstate_kobjs[idx] = NULL;
3558                 }
3559         }
3560
3561         kobject_put(nhs->hugepages_kobj);
3562         nhs->hugepages_kobj = NULL;
3563 }
3564
3565
3566 /*
3567  * Register hstate attributes for a single node device.
3568  * No-op if attributes already registered.
3569  */
3570 static void hugetlb_register_node(struct node *node)
3571 {
3572         struct hstate *h;
3573         struct node_hstate *nhs = &node_hstates[node->dev.id];
3574         int err;
3575
3576         if (nhs->hugepages_kobj)
3577                 return;         /* already allocated */
3578
3579         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3580                                                         &node->dev.kobj);
3581         if (!nhs->hugepages_kobj)
3582                 return;
3583
3584         for_each_hstate(h) {
3585                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3586                                                 nhs->hstate_kobjs,
3587                                                 &per_node_hstate_attr_group);
3588                 if (err) {
3589                         pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3590                                 h->name, node->dev.id);
3591                         hugetlb_unregister_node(node);
3592                         break;
3593                 }
3594         }
3595 }
3596
3597 /*
3598  * hugetlb init time:  register hstate attributes for all registered node
3599  * devices of nodes that have memory.  All on-line nodes should have
3600  * registered their associated device by this time.
3601  */
3602 static void __init hugetlb_register_all_nodes(void)
3603 {
3604         int nid;
3605
3606         for_each_node_state(nid, N_MEMORY) {
3607                 struct node *node = node_devices[nid];
3608                 if (node->dev.id == nid)
3609                         hugetlb_register_node(node);
3610         }
3611
3612         /*
3613          * Let the node device driver know we're here so it can
3614          * [un]register hstate attributes on node hotplug.
3615          */
3616         register_hugetlbfs_with_node(hugetlb_register_node,
3617                                      hugetlb_unregister_node);
3618 }
3619 #else   /* !CONFIG_NUMA */
3620
3621 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3622 {
3623         BUG();
3624         if (nidp)
3625                 *nidp = -1;
3626         return NULL;
3627 }
3628
3629 static void hugetlb_register_all_nodes(void) { }
3630
3631 #endif
3632
3633 static int __init hugetlb_init(void)
3634 {
3635         int i;
3636
3637         BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
3638                         __NR_HPAGEFLAGS);
3639
3640         if (!hugepages_supported()) {
3641                 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3642                         pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3643                 return 0;
3644         }
3645
3646         /*
3647          * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists.  Some
3648          * architectures depend on setup being done here.
3649          */
3650         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3651         if (!parsed_default_hugepagesz) {
3652                 /*
3653                  * If we did not parse a default huge page size, set
3654                  * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3655                  * number of huge pages for this default size was implicitly
3656                  * specified, set that here as well.
3657                  * Note that the implicit setting will overwrite an explicit
3658                  * setting.  A warning will be printed in this case.
3659                  */
3660                 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3661                 if (default_hstate_max_huge_pages) {
3662                         if (default_hstate.max_huge_pages) {
3663                                 char buf[32];
3664
3665                                 string_get_size(huge_page_size(&default_hstate),
3666                                         1, STRING_UNITS_2, buf, 32);
3667                                 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3668                                         default_hstate.max_huge_pages, buf);
3669                                 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3670                                         default_hstate_max_huge_pages);
3671                         }
3672                         default_hstate.max_huge_pages =
3673                                 default_hstate_max_huge_pages;
3674                 }
3675         }
3676
3677         hugetlb_cma_check();
3678         hugetlb_init_hstates();
3679         gather_bootmem_prealloc();
3680         report_hugepages();
3681
3682         hugetlb_sysfs_init();
3683         hugetlb_register_all_nodes();
3684         hugetlb_cgroup_file_init();
3685
3686 #ifdef CONFIG_SMP
3687         num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3688 #else
3689         num_fault_mutexes = 1;
3690 #endif
3691         hugetlb_fault_mutex_table =
3692                 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3693                               GFP_KERNEL);
3694         BUG_ON(!hugetlb_fault_mutex_table);
3695
3696         for (i = 0; i < num_fault_mutexes; i++)
3697                 mutex_init(&hugetlb_fault_mutex_table[i]);
3698         return 0;
3699 }
3700 subsys_initcall(hugetlb_init);
3701
3702 /* Overwritten by architectures with more huge page sizes */
3703 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3704 {
3705         return size == HPAGE_SIZE;
3706 }
3707
3708 void __init hugetlb_add_hstate(unsigned int order)
3709 {
3710         struct hstate *h;
3711         unsigned long i;
3712
3713         if (size_to_hstate(PAGE_SIZE << order)) {
3714                 return;
3715         }
3716         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3717         BUG_ON(order == 0);
3718         h = &hstates[hugetlb_max_hstate++];
3719         mutex_init(&h->resize_lock);
3720         h->order = order;
3721         h->mask = ~(huge_page_size(h) - 1);
3722         for (i = 0; i < MAX_NUMNODES; ++i)
3723                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3724         INIT_LIST_HEAD(&h->hugepage_activelist);
3725         h->next_nid_to_alloc = first_memory_node;
3726         h->next_nid_to_free = first_memory_node;
3727         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3728                                         huge_page_size(h)/1024);
3729         hugetlb_vmemmap_init(h);
3730
3731         parsed_hstate = h;
3732 }
3733
3734 /*
3735  * hugepages command line processing
3736  * hugepages normally follows a valid hugepagsz or default_hugepagsz
3737  * specification.  If not, ignore the hugepages value.  hugepages can also
3738  * be the first huge page command line  option in which case it implicitly
3739  * specifies the number of huge pages for the default size.
3740  */
3741 static int __init hugepages_setup(char *s)
3742 {
3743         unsigned long *mhp;
3744         static unsigned long *last_mhp;
3745
3746         if (!parsed_valid_hugepagesz) {
3747                 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3748                 parsed_valid_hugepagesz = true;
3749                 return 0;
3750         }
3751
3752         /*
3753          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3754          * yet, so this hugepages= parameter goes to the "default hstate".
3755          * Otherwise, it goes with the previously parsed hugepagesz or
3756          * default_hugepagesz.
3757          */
3758         else if (!hugetlb_max_hstate)
3759                 mhp = &default_hstate_max_huge_pages;
3760         else
3761                 mhp = &parsed_hstate->max_huge_pages;
3762
3763         if (mhp == last_mhp) {
3764                 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3765                 return 0;
3766         }
3767
3768         if (sscanf(s, "%lu", mhp) <= 0)
3769                 *mhp = 0;
3770
3771         /*
3772          * Global state is always initialized later in hugetlb_init.
3773          * But we need to allocate gigantic hstates here early to still
3774          * use the bootmem allocator.
3775          */
3776         if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
3777                 hugetlb_hstate_alloc_pages(parsed_hstate);
3778
3779         last_mhp = mhp;
3780
3781         return 1;
3782 }
3783 __setup("hugepages=", hugepages_setup);
3784
3785 /*
3786  * hugepagesz command line processing
3787  * A specific huge page size can only be specified once with hugepagesz.
3788  * hugepagesz is followed by hugepages on the command line.  The global
3789  * variable 'parsed_valid_hugepagesz' is used to determine if prior
3790  * hugepagesz argument was valid.
3791  */
3792 static int __init hugepagesz_setup(char *s)
3793 {
3794         unsigned long size;
3795         struct hstate *h;
3796
3797         parsed_valid_hugepagesz = false;
3798         size = (unsigned long)memparse(s, NULL);
3799
3800         if (!arch_hugetlb_valid_size(size)) {
3801                 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3802                 return 0;
3803         }
3804
3805         h = size_to_hstate(size);
3806         if (h) {
3807                 /*
3808                  * hstate for this size already exists.  This is normally
3809                  * an error, but is allowed if the existing hstate is the
3810                  * default hstate.  More specifically, it is only allowed if
3811                  * the number of huge pages for the default hstate was not
3812                  * previously specified.
3813                  */
3814                 if (!parsed_default_hugepagesz ||  h != &default_hstate ||
3815                     default_hstate.max_huge_pages) {
3816                         pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3817                         return 0;
3818                 }
3819
3820                 /*
3821                  * No need to call hugetlb_add_hstate() as hstate already
3822                  * exists.  But, do set parsed_hstate so that a following
3823                  * hugepages= parameter will be applied to this hstate.
3824                  */
3825                 parsed_hstate = h;
3826                 parsed_valid_hugepagesz = true;
3827                 return 1;
3828         }
3829
3830         hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3831         parsed_valid_hugepagesz = true;
3832         return 1;
3833 }
3834 __setup("hugepagesz=", hugepagesz_setup);
3835
3836 /*
3837  * default_hugepagesz command line input
3838  * Only one instance of default_hugepagesz allowed on command line.
3839  */
3840 static int __init default_hugepagesz_setup(char *s)
3841 {
3842         unsigned long size;
3843
3844         parsed_valid_hugepagesz = false;
3845         if (parsed_default_hugepagesz) {
3846                 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3847                 return 0;
3848         }
3849
3850         size = (unsigned long)memparse(s, NULL);
3851
3852         if (!arch_hugetlb_valid_size(size)) {
3853                 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3854                 return 0;
3855         }
3856
3857         hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3858         parsed_valid_hugepagesz = true;
3859         parsed_default_hugepagesz = true;
3860         default_hstate_idx = hstate_index(size_to_hstate(size));
3861
3862         /*
3863          * The number of default huge pages (for this size) could have been
3864          * specified as the first hugetlb parameter: hugepages=X.  If so,
3865          * then default_hstate_max_huge_pages is set.  If the default huge
3866          * page size is gigantic (>= MAX_ORDER), then the pages must be
3867          * allocated here from bootmem allocator.
3868          */
3869         if (default_hstate_max_huge_pages) {
3870                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3871                 if (hstate_is_gigantic(&default_hstate))
3872                         hugetlb_hstate_alloc_pages(&default_hstate);
3873                 default_hstate_max_huge_pages = 0;
3874         }
3875
3876         return 1;
3877 }
3878 __setup("default_hugepagesz=", default_hugepagesz_setup);
3879
3880 static unsigned int allowed_mems_nr(struct hstate *h)
3881 {
3882         int node;
3883         unsigned int nr = 0;
3884         nodemask_t *mpol_allowed;
3885         unsigned int *array = h->free_huge_pages_node;
3886         gfp_t gfp_mask = htlb_alloc_mask(h);
3887
3888         mpol_allowed = policy_nodemask_current(gfp_mask);
3889
3890         for_each_node_mask(node, cpuset_current_mems_allowed) {
3891                 if (!mpol_allowed || node_isset(node, *mpol_allowed))
3892                         nr += array[node];
3893         }
3894
3895         return nr;
3896 }
3897
3898 #ifdef CONFIG_SYSCTL
3899 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3900                                           void *buffer, size_t *length,
3901                                           loff_t *ppos, unsigned long *out)
3902 {
3903         struct ctl_table dup_table;
3904
3905         /*
3906          * In order to avoid races with __do_proc_doulongvec_minmax(), we
3907          * can duplicate the @table and alter the duplicate of it.
3908          */
3909         dup_table = *table;
3910         dup_table.data = out;
3911
3912         return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3913 }
3914
3915 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3916                          struct ctl_table *table, int write,
3917                          void *buffer, size_t *length, loff_t *ppos)
3918 {
3919         struct hstate *h = &default_hstate;
3920         unsigned long tmp = h->max_huge_pages;
3921         int ret;
3922
3923         if (!hugepages_supported())
3924                 return -EOPNOTSUPP;
3925
3926         ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3927                                              &tmp);
3928         if (ret)
3929                 goto out;
3930
3931         if (write)
3932                 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3933                                                   NUMA_NO_NODE, tmp, *length);
3934 out:
3935         return ret;
3936 }
3937
3938 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3939                           void *buffer, size_t *length, loff_t *ppos)
3940 {
3941
3942         return hugetlb_sysctl_handler_common(false, table, write,
3943                                                         buffer, length, ppos);
3944 }
3945
3946 #ifdef CONFIG_NUMA
3947 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3948                           void *buffer, size_t *length, loff_t *ppos)
3949 {
3950         return hugetlb_sysctl_handler_common(true, table, write,
3951                                                         buffer, length, ppos);
3952 }
3953 #endif /* CONFIG_NUMA */
3954
3955 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3956                 void *buffer, size_t *length, loff_t *ppos)
3957 {
3958         struct hstate *h = &default_hstate;
3959         unsigned long tmp;
3960         int ret;
3961
3962         if (!hugepages_supported())
3963                 return -EOPNOTSUPP;
3964
3965         tmp = h->nr_overcommit_huge_pages;
3966
3967         if (write && hstate_is_gigantic(h))
3968                 return -EINVAL;
3969
3970         ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3971                                              &tmp);
3972         if (ret)
3973                 goto out;
3974
3975         if (write) {
3976                 spin_lock_irq(&hugetlb_lock);
3977                 h->nr_overcommit_huge_pages = tmp;
3978                 spin_unlock_irq(&hugetlb_lock);
3979         }
3980 out:
3981         return ret;
3982 }
3983
3984 #endif /* CONFIG_SYSCTL */
3985
3986 void hugetlb_report_meminfo(struct seq_file *m)
3987 {
3988         struct hstate *h;
3989         unsigned long total = 0;
3990
3991         if (!hugepages_supported())
3992                 return;
3993
3994         for_each_hstate(h) {
3995                 unsigned long count = h->nr_huge_pages;
3996
3997                 total += huge_page_size(h) * count;
3998
3999                 if (h == &default_hstate)
4000                         seq_printf(m,
4001                                    "HugePages_Total:   %5lu\n"
4002                                    "HugePages_Free:    %5lu\n"
4003                                    "HugePages_Rsvd:    %5lu\n"
4004                                    "HugePages_Surp:    %5lu\n"
4005                                    "Hugepagesize:   %8lu kB\n",
4006                                    count,
4007                                    h->free_huge_pages,
4008                                    h->resv_huge_pages,
4009                                    h->surplus_huge_pages,
4010                                    huge_page_size(h) / SZ_1K);
4011         }
4012
4013         seq_printf(m, "Hugetlb:        %8lu kB\n", total / SZ_1K);
4014 }
4015
4016 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
4017 {
4018         struct hstate *h = &default_hstate;
4019
4020         if (!hugepages_supported())
4021                 return 0;
4022
4023         return sysfs_emit_at(buf, len,
4024                              "Node %d HugePages_Total: %5u\n"
4025                              "Node %d HugePages_Free:  %5u\n"
4026                              "Node %d HugePages_Surp:  %5u\n",
4027                              nid, h->nr_huge_pages_node[nid],
4028                              nid, h->free_huge_pages_node[nid],
4029                              nid, h->surplus_huge_pages_node[nid]);
4030 }
4031
4032 void hugetlb_show_meminfo(void)
4033 {
4034         struct hstate *h;
4035         int nid;
4036
4037         if (!hugepages_supported())
4038                 return;
4039
4040         for_each_node_state(nid, N_MEMORY)
4041                 for_each_hstate(h)
4042                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
4043                                 nid,
4044                                 h->nr_huge_pages_node[nid],
4045                                 h->free_huge_pages_node[nid],
4046                                 h->surplus_huge_pages_node[nid],
4047                                 huge_page_size(h) / SZ_1K);
4048 }
4049
4050 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
4051 {
4052         seq_printf(m, "HugetlbPages:\t%8lu kB\n",
4053                    atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
4054 }
4055
4056 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
4057 unsigned long hugetlb_total_pages(void)
4058 {
4059         struct hstate *h;
4060         unsigned long nr_total_pages = 0;
4061
4062         for_each_hstate(h)
4063                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
4064         return nr_total_pages;
4065 }
4066
4067 static int hugetlb_acct_memory(struct hstate *h, long delta)
4068 {
4069         int ret = -ENOMEM;
4070
4071         if (!delta)
4072                 return 0;
4073
4074         spin_lock_irq(&hugetlb_lock);
4075         /*
4076          * When cpuset is configured, it breaks the strict hugetlb page
4077          * reservation as the accounting is done on a global variable. Such
4078          * reservation is completely rubbish in the presence of cpuset because
4079          * the reservation is not checked against page availability for the
4080          * current cpuset. Application can still potentially OOM'ed by kernel
4081          * with lack of free htlb page in cpuset that the task is in.
4082          * Attempt to enforce strict accounting with cpuset is almost
4083          * impossible (or too ugly) because cpuset is too fluid that
4084          * task or memory node can be dynamically moved between cpusets.
4085          *
4086          * The change of semantics for shared hugetlb mapping with cpuset is
4087          * undesirable. However, in order to preserve some of the semantics,
4088          * we fall back to check against current free page availability as
4089          * a best attempt and hopefully to minimize the impact of changing
4090          * semantics that cpuset has.
4091          *
4092          * Apart from cpuset, we also have memory policy mechanism that
4093          * also determines from which node the kernel will allocate memory
4094          * in a NUMA system. So similar to cpuset, we also should consider
4095          * the memory policy of the current task. Similar to the description
4096          * above.
4097          */
4098         if (delta > 0) {
4099                 if (gather_surplus_pages(h, delta) < 0)
4100                         goto out;
4101
4102                 if (delta > allowed_mems_nr(h)) {
4103                         return_unused_surplus_pages(h, delta);
4104                         goto out;
4105                 }
4106         }
4107
4108         ret = 0;
4109         if (delta < 0)
4110                 return_unused_surplus_pages(h, (unsigned long) -delta);
4111
4112 out:
4113         spin_unlock_irq(&hugetlb_lock);
4114         return ret;
4115 }
4116
4117 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4118 {
4119         struct resv_map *resv = vma_resv_map(vma);
4120
4121         /*
4122          * This new VMA should share its siblings reservation map if present.
4123          * The VMA will only ever have a valid reservation map pointer where
4124          * it is being copied for another still existing VMA.  As that VMA
4125          * has a reference to the reservation map it cannot disappear until
4126          * after this open call completes.  It is therefore safe to take a
4127          * new reference here without additional locking.
4128          */
4129         if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
4130                 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
4131                 kref_get(&resv->refs);
4132         }
4133 }
4134
4135 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4136 {
4137         struct hstate *h = hstate_vma(vma);
4138         struct resv_map *resv = vma_resv_map(vma);
4139         struct hugepage_subpool *spool = subpool_vma(vma);
4140         unsigned long reserve, start, end;
4141         long gbl_reserve;
4142
4143         if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4144                 return;
4145
4146         start = vma_hugecache_offset(h, vma, vma->vm_start);
4147         end = vma_hugecache_offset(h, vma, vma->vm_end);
4148
4149         reserve = (end - start) - region_count(resv, start, end);
4150         hugetlb_cgroup_uncharge_counter(resv, start, end);
4151         if (reserve) {
4152                 /*
4153                  * Decrement reserve counts.  The global reserve count may be
4154                  * adjusted if the subpool has a minimum size.
4155                  */
4156                 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4157                 hugetlb_acct_memory(h, -gbl_reserve);
4158         }
4159
4160         kref_put(&resv->refs, resv_map_release);
4161 }
4162
4163 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4164 {
4165         if (addr & ~(huge_page_mask(hstate_vma(vma))))
4166                 return -EINVAL;
4167         return 0;
4168 }
4169
4170 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4171 {
4172         return huge_page_size(hstate_vma(vma));
4173 }
4174
4175 /*
4176  * We cannot handle pagefaults against hugetlb pages at all.  They cause
4177  * handle_mm_fault() to try to instantiate regular-sized pages in the
4178  * hugepage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
4179  * this far.
4180  */
4181 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4182 {
4183         BUG();
4184         return 0;
4185 }
4186
4187 /*
4188  * When a new function is introduced to vm_operations_struct and added
4189  * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4190  * This is because under System V memory model, mappings created via
4191  * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4192  * their original vm_ops are overwritten with shm_vm_ops.
4193  */
4194 const struct vm_operations_struct hugetlb_vm_ops = {
4195         .fault = hugetlb_vm_op_fault,
4196         .open = hugetlb_vm_op_open,
4197         .close = hugetlb_vm_op_close,
4198         .may_split = hugetlb_vm_op_split,
4199         .pagesize = hugetlb_vm_op_pagesize,
4200 };
4201
4202 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4203                                 int writable)
4204 {
4205         pte_t entry;
4206         unsigned int shift = huge_page_shift(hstate_vma(vma));
4207
4208         if (writable) {
4209                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4210                                          vma->vm_page_prot)));
4211         } else {
4212                 entry = huge_pte_wrprotect(mk_huge_pte(page,
4213                                            vma->vm_page_prot));
4214         }
4215         entry = pte_mkyoung(entry);
4216         entry = pte_mkhuge(entry);
4217         entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4218
4219         return entry;
4220 }
4221
4222 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4223                                    unsigned long address, pte_t *ptep)
4224 {
4225         pte_t entry;
4226
4227         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4228         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4229                 update_mmu_cache(vma, address, ptep);
4230 }
4231
4232 bool is_hugetlb_entry_migration(pte_t pte)
4233 {
4234         swp_entry_t swp;
4235
4236         if (huge_pte_none(pte) || pte_present(pte))
4237                 return false;
4238         swp = pte_to_swp_entry(pte);
4239         if (is_migration_entry(swp))
4240                 return true;
4241         else
4242                 return false;
4243 }
4244
4245 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4246 {
4247         swp_entry_t swp;
4248
4249         if (huge_pte_none(pte) || pte_present(pte))
4250                 return false;
4251         swp = pte_to_swp_entry(pte);
4252         if (is_hwpoison_entry(swp))
4253                 return true;
4254         else
4255                 return false;
4256 }
4257
4258 static void
4259 hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
4260                      struct page *new_page)
4261 {
4262         __SetPageUptodate(new_page);
4263         set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
4264         hugepage_add_new_anon_rmap(new_page, vma, addr);
4265         hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
4266         ClearHPageRestoreReserve(new_page);
4267         SetHPageMigratable(new_page);
4268 }
4269
4270 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
4271                             struct vm_area_struct *vma)
4272 {
4273         pte_t *src_pte, *dst_pte, entry, dst_entry;
4274         struct page *ptepage;
4275         unsigned long addr;
4276         bool cow = is_cow_mapping(vma->vm_flags);
4277         struct hstate *h = hstate_vma(vma);
4278         unsigned long sz = huge_page_size(h);
4279         unsigned long npages = pages_per_huge_page(h);
4280         struct address_space *mapping = vma->vm_file->f_mapping;
4281         struct mmu_notifier_range range;
4282         int ret = 0;
4283
4284         if (cow) {
4285                 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
4286                                         vma->vm_start,
4287                                         vma->vm_end);
4288                 mmu_notifier_invalidate_range_start(&range);
4289         } else {
4290                 /*
4291                  * For shared mappings i_mmap_rwsem must be held to call
4292                  * huge_pte_alloc, otherwise the returned ptep could go
4293                  * away if part of a shared pmd and another thread calls
4294                  * huge_pmd_unshare.
4295                  */
4296                 i_mmap_lock_read(mapping);
4297         }
4298
4299         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
4300                 spinlock_t *src_ptl, *dst_ptl;
4301                 src_pte = huge_pte_offset(src, addr, sz);
4302                 if (!src_pte)
4303                         continue;
4304                 dst_pte = huge_pte_alloc(dst, vma, addr, sz);
4305                 if (!dst_pte) {
4306                         ret = -ENOMEM;
4307                         break;
4308                 }
4309
4310                 /*
4311                  * If the pagetables are shared don't copy or take references.
4312                  * dst_pte == src_pte is the common case of src/dest sharing.
4313                  *
4314                  * However, src could have 'unshared' and dst shares with
4315                  * another vma.  If dst_pte !none, this implies sharing.
4316                  * Check here before taking page table lock, and once again
4317                  * after taking the lock below.
4318                  */
4319                 dst_entry = huge_ptep_get(dst_pte);
4320                 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
4321                         continue;
4322
4323                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4324                 src_ptl = huge_pte_lockptr(h, src, src_pte);
4325                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4326                 entry = huge_ptep_get(src_pte);
4327                 dst_entry = huge_ptep_get(dst_pte);
4328 again:
4329                 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
4330                         /*
4331                          * Skip if src entry none.  Also, skip in the
4332                          * unlikely case dst entry !none as this implies
4333                          * sharing with another vma.
4334                          */
4335                         ;
4336                 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
4337                                     is_hugetlb_entry_hwpoisoned(entry))) {
4338                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
4339
4340                         if (is_writable_migration_entry(swp_entry) && cow) {
4341                                 /*
4342                                  * COW mappings require pages in both
4343                                  * parent and child to be set to read.
4344                                  */
4345                                 swp_entry = make_readable_migration_entry(
4346                                                         swp_offset(swp_entry));
4347                                 entry = swp_entry_to_pte(swp_entry);
4348                                 set_huge_swap_pte_at(src, addr, src_pte,
4349                                                      entry, sz);
4350                         }
4351                         set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
4352                 } else {
4353                         entry = huge_ptep_get(src_pte);
4354                         ptepage = pte_page(entry);
4355                         get_page(ptepage);
4356
4357                         /*
4358                          * This is a rare case where we see pinned hugetlb
4359                          * pages while they're prone to COW.  We need to do the
4360                          * COW earlier during fork.
4361                          *
4362                          * When pre-allocating the page or copying data, we
4363                          * need to be without the pgtable locks since we could
4364                          * sleep during the process.
4365                          */
4366                         if (unlikely(page_needs_cow_for_dma(vma, ptepage))) {
4367                                 pte_t src_pte_old = entry;
4368                                 struct page *new;
4369
4370                                 spin_unlock(src_ptl);
4371                                 spin_unlock(dst_ptl);
4372                                 /* Do not use reserve as it's private owned */
4373                                 new = alloc_huge_page(vma, addr, 1);
4374                                 if (IS_ERR(new)) {
4375                                         put_page(ptepage);
4376                                         ret = PTR_ERR(new);
4377                                         break;
4378                                 }
4379                                 copy_user_huge_page(new, ptepage, addr, vma,
4380                                                     npages);
4381                                 put_page(ptepage);
4382
4383                                 /* Install the new huge page if src pte stable */
4384                                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4385                                 src_ptl = huge_pte_lockptr(h, src, src_pte);
4386                                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4387                                 entry = huge_ptep_get(src_pte);
4388                                 if (!pte_same(src_pte_old, entry)) {
4389                                         restore_reserve_on_error(h, vma, addr,
4390                                                                 new);
4391                                         put_page(new);
4392                                         /* dst_entry won't change as in child */
4393                                         goto again;
4394                                 }
4395                                 hugetlb_install_page(vma, dst_pte, addr, new);
4396                                 spin_unlock(src_ptl);
4397                                 spin_unlock(dst_ptl);
4398                                 continue;
4399                         }
4400
4401                         if (cow) {
4402                                 /*
4403                                  * No need to notify as we are downgrading page
4404                                  * table protection not changing it to point
4405                                  * to a new page.
4406                                  *
4407                                  * See Documentation/vm/mmu_notifier.rst
4408                                  */
4409                                 huge_ptep_set_wrprotect(src, addr, src_pte);
4410                                 entry = huge_pte_wrprotect(entry);
4411                         }
4412
4413                         page_dup_rmap(ptepage, true);
4414                         set_huge_pte_at(dst, addr, dst_pte, entry);
4415                         hugetlb_count_add(npages, dst);
4416                 }
4417                 spin_unlock(src_ptl);
4418                 spin_unlock(dst_ptl);
4419         }
4420
4421         if (cow)
4422                 mmu_notifier_invalidate_range_end(&range);
4423         else
4424                 i_mmap_unlock_read(mapping);
4425
4426         return ret;
4427 }
4428
4429 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
4430                             unsigned long start, unsigned long end,
4431                             struct page *ref_page)
4432 {
4433         struct mm_struct *mm = vma->vm_mm;
4434         unsigned long address;
4435         pte_t *ptep;
4436         pte_t pte;
4437         spinlock_t *ptl;
4438         struct page *page;
4439         struct hstate *h = hstate_vma(vma);
4440         unsigned long sz = huge_page_size(h);
4441         struct mmu_notifier_range range;
4442
4443         WARN_ON(!is_vm_hugetlb_page(vma));
4444         BUG_ON(start & ~huge_page_mask(h));
4445         BUG_ON(end & ~huge_page_mask(h));
4446
4447         /*
4448          * This is a hugetlb vma, all the pte entries should point
4449          * to huge page.
4450          */
4451         tlb_change_page_size(tlb, sz);
4452         tlb_start_vma(tlb, vma);
4453
4454         /*
4455          * If sharing possible, alert mmu notifiers of worst case.
4456          */
4457         mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
4458                                 end);
4459         adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4460         mmu_notifier_invalidate_range_start(&range);
4461         address = start;
4462         for (; address < end; address += sz) {
4463                 ptep = huge_pte_offset(mm, address, sz);
4464                 if (!ptep)
4465                         continue;
4466
4467                 ptl = huge_pte_lock(h, mm, ptep);
4468                 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
4469                         spin_unlock(ptl);
4470                         /*
4471                          * We just unmapped a page of PMDs by clearing a PUD.
4472                          * The caller's TLB flush range should cover this area.
4473                          */
4474                         continue;
4475                 }
4476
4477                 pte = huge_ptep_get(ptep);
4478                 if (huge_pte_none(pte)) {
4479                         spin_unlock(ptl);
4480                         continue;
4481                 }
4482
4483                 /*
4484                  * Migrating hugepage or HWPoisoned hugepage is already
4485                  * unmapped and its refcount is dropped, so just clear pte here.
4486                  */
4487                 if (unlikely(!pte_present(pte))) {
4488                         huge_pte_clear(mm, address, ptep, sz);
4489                         spin_unlock(ptl);
4490                         continue;
4491                 }
4492
4493                 page = pte_page(pte);
4494                 /*
4495                  * If a reference page is supplied, it is because a specific
4496                  * page is being unmapped, not a range. Ensure the page we
4497                  * are about to unmap is the actual page of interest.
4498                  */
4499                 if (ref_page) {
4500                         if (page != ref_page) {
4501                                 spin_unlock(ptl);
4502                                 continue;
4503                         }
4504                         /*
4505                          * Mark the VMA as having unmapped its page so that
4506                          * future faults in this VMA will fail rather than
4507                          * looking like data was lost
4508                          */
4509                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
4510                 }
4511
4512                 pte = huge_ptep_get_and_clear(mm, address, ptep);
4513                 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
4514                 if (huge_pte_dirty(pte))
4515                         set_page_dirty(page);
4516
4517                 hugetlb_count_sub(pages_per_huge_page(h), mm);
4518                 page_remove_rmap(page, true);
4519
4520                 spin_unlock(ptl);
4521                 tlb_remove_page_size(tlb, page, huge_page_size(h));
4522                 /*
4523                  * Bail out after unmapping reference page if supplied
4524                  */
4525                 if (ref_page)
4526                         break;
4527         }
4528         mmu_notifier_invalidate_range_end(&range);
4529         tlb_end_vma(tlb, vma);
4530 }
4531
4532 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4533                           struct vm_area_struct *vma, unsigned long start,
4534                           unsigned long end, struct page *ref_page)
4535 {
4536         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4537
4538         /*
4539          * Clear this flag so that x86's huge_pmd_share page_table_shareable
4540          * test will fail on a vma being torn down, and not grab a page table
4541          * on its way out.  We're lucky that the flag has such an appropriate
4542          * name, and can in fact be safely cleared here. We could clear it
4543          * before the __unmap_hugepage_range above, but all that's necessary
4544          * is to clear it before releasing the i_mmap_rwsem. This works
4545          * because in the context this is called, the VMA is about to be
4546          * destroyed and the i_mmap_rwsem is held.
4547          */
4548         vma->vm_flags &= ~VM_MAYSHARE;
4549 }
4550
4551 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
4552                           unsigned long end, struct page *ref_page)
4553 {
4554         struct mmu_gather tlb;
4555
4556         tlb_gather_mmu(&tlb, vma->vm_mm);
4557         __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
4558         tlb_finish_mmu(&tlb);
4559 }
4560
4561 /*
4562  * This is called when the original mapper is failing to COW a MAP_PRIVATE
4563  * mapping it owns the reserve page for. The intention is to unmap the page
4564  * from other VMAs and let the children be SIGKILLed if they are faulting the
4565  * same region.
4566  */
4567 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4568                               struct page *page, unsigned long address)
4569 {
4570         struct hstate *h = hstate_vma(vma);
4571         struct vm_area_struct *iter_vma;
4572         struct address_space *mapping;
4573         pgoff_t pgoff;
4574
4575         /*
4576          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4577          * from page cache lookup which is in HPAGE_SIZE units.
4578          */
4579         address = address & huge_page_mask(h);
4580         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4581                         vma->vm_pgoff;
4582         mapping = vma->vm_file->f_mapping;
4583
4584         /*
4585          * Take the mapping lock for the duration of the table walk. As
4586          * this mapping should be shared between all the VMAs,
4587          * __unmap_hugepage_range() is called as the lock is already held
4588          */
4589         i_mmap_lock_write(mapping);
4590         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
4591                 /* Do not unmap the current VMA */
4592                 if (iter_vma == vma)
4593                         continue;
4594
4595                 /*
4596                  * Shared VMAs have their own reserves and do not affect
4597                  * MAP_PRIVATE accounting but it is possible that a shared
4598                  * VMA is using the same page so check and skip such VMAs.
4599                  */
4600                 if (iter_vma->vm_flags & VM_MAYSHARE)
4601                         continue;
4602
4603                 /*
4604                  * Unmap the page from other VMAs without their own reserves.
4605                  * They get marked to be SIGKILLed if they fault in these
4606                  * areas. This is because a future no-page fault on this VMA
4607                  * could insert a zeroed page instead of the data existing
4608                  * from the time of fork. This would look like data corruption
4609                  */
4610                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
4611                         unmap_hugepage_range(iter_vma, address,
4612                                              address + huge_page_size(h), page);
4613         }
4614         i_mmap_unlock_write(mapping);
4615 }
4616
4617 /*
4618  * Hugetlb_cow() should be called with page lock of the original hugepage held.
4619  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4620  * cannot race with other handlers or page migration.
4621  * Keep the pte_same checks anyway to make transition from the mutex easier.
4622  */
4623 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
4624                        unsigned long address, pte_t *ptep,
4625                        struct page *pagecache_page, spinlock_t *ptl)
4626 {
4627         pte_t pte;
4628         struct hstate *h = hstate_vma(vma);
4629         struct page *old_page, *new_page;
4630         int outside_reserve = 0;
4631         vm_fault_t ret = 0;
4632         unsigned long haddr = address & huge_page_mask(h);
4633         struct mmu_notifier_range range;
4634
4635         pte = huge_ptep_get(ptep);
4636         old_page = pte_page(pte);
4637
4638 retry_avoidcopy:
4639         /* If no-one else is actually using this page, avoid the copy
4640          * and just make the page writable */
4641         if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
4642                 page_move_anon_rmap(old_page, vma);
4643                 set_huge_ptep_writable(vma, haddr, ptep);
4644                 return 0;
4645         }
4646
4647         /*
4648          * If the process that created a MAP_PRIVATE mapping is about to
4649          * perform a COW due to a shared page count, attempt to satisfy
4650          * the allocation without using the existing reserves. The pagecache
4651          * page is used to determine if the reserve at this address was
4652          * consumed or not. If reserves were used, a partial faulted mapping
4653          * at the time of fork() could consume its reserves on COW instead
4654          * of the full address range.
4655          */
4656         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4657                         old_page != pagecache_page)
4658                 outside_reserve = 1;
4659
4660         get_page(old_page);
4661
4662         /*
4663          * Drop page table lock as buddy allocator may be called. It will
4664          * be acquired again before returning to the caller, as expected.
4665          */
4666         spin_unlock(ptl);
4667         new_page = alloc_huge_page(vma, haddr, outside_reserve);
4668
4669         if (IS_ERR(new_page)) {
4670                 /*
4671                  * If a process owning a MAP_PRIVATE mapping fails to COW,
4672                  * it is due to references held by a child and an insufficient
4673                  * huge page pool. To guarantee the original mappers
4674                  * reliability, unmap the page from child processes. The child
4675                  * may get SIGKILLed if it later faults.
4676                  */
4677                 if (outside_reserve) {
4678                         struct address_space *mapping = vma->vm_file->f_mapping;
4679                         pgoff_t idx;
4680                         u32 hash;
4681
4682                         put_page(old_page);
4683                         BUG_ON(huge_pte_none(pte));
4684                         /*
4685                          * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4686                          * unmapping.  unmapping needs to hold i_mmap_rwsem
4687                          * in write mode.  Dropping i_mmap_rwsem in read mode
4688                          * here is OK as COW mappings do not interact with
4689                          * PMD sharing.
4690                          *
4691                          * Reacquire both after unmap operation.
4692                          */
4693                         idx = vma_hugecache_offset(h, vma, haddr);
4694                         hash = hugetlb_fault_mutex_hash(mapping, idx);
4695                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4696                         i_mmap_unlock_read(mapping);
4697
4698                         unmap_ref_private(mm, vma, old_page, haddr);
4699
4700                         i_mmap_lock_read(mapping);
4701                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
4702                         spin_lock(ptl);
4703                         ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4704                         if (likely(ptep &&
4705                                    pte_same(huge_ptep_get(ptep), pte)))
4706                                 goto retry_avoidcopy;
4707                         /*
4708                          * race occurs while re-acquiring page table
4709                          * lock, and our job is done.
4710                          */
4711                         return 0;
4712                 }
4713
4714                 ret = vmf_error(PTR_ERR(new_page));
4715                 goto out_release_old;
4716         }
4717
4718         /*
4719          * When the original hugepage is shared one, it does not have
4720          * anon_vma prepared.
4721          */
4722         if (unlikely(anon_vma_prepare(vma))) {
4723                 ret = VM_FAULT_OOM;
4724                 goto out_release_all;
4725         }
4726
4727         copy_user_huge_page(new_page, old_page, address, vma,
4728                             pages_per_huge_page(h));
4729         __SetPageUptodate(new_page);
4730
4731         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4732                                 haddr + huge_page_size(h));
4733         mmu_notifier_invalidate_range_start(&range);
4734
4735         /*
4736          * Retake the page table lock to check for racing updates
4737          * before the page tables are altered
4738          */
4739         spin_lock(ptl);
4740         ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4741         if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4742                 ClearHPageRestoreReserve(new_page);
4743
4744                 /* Break COW */
4745                 huge_ptep_clear_flush(vma, haddr, ptep);
4746                 mmu_notifier_invalidate_range(mm, range.start, range.end);
4747                 set_huge_pte_at(mm, haddr, ptep,
4748                                 make_huge_pte(vma, new_page, 1));
4749                 page_remove_rmap(old_page, true);
4750                 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4751                 SetHPageMigratable(new_page);
4752                 /* Make the old page be freed below */
4753                 new_page = old_page;
4754         }
4755         spin_unlock(ptl);
4756         mmu_notifier_invalidate_range_end(&range);
4757 out_release_all:
4758         /* No restore in case of successful pagetable update (Break COW) */
4759         if (new_page != old_page)
4760                 restore_reserve_on_error(h, vma, haddr, new_page);
4761         put_page(new_page);
4762 out_release_old:
4763         put_page(old_page);
4764
4765         spin_lock(ptl); /* Caller expects lock to be held */
4766         return ret;
4767 }
4768
4769 /* Return the pagecache page at a given address within a VMA */
4770 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4771                         struct vm_area_struct *vma, unsigned long address)
4772 {
4773         struct address_space *mapping;
4774         pgoff_t idx;
4775
4776         mapping = vma->vm_file->f_mapping;
4777         idx = vma_hugecache_offset(h, vma, address);
4778
4779         return find_lock_page(mapping, idx);
4780 }
4781
4782 /*
4783  * Return whether there is a pagecache page to back given address within VMA.
4784  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4785  */
4786 static bool hugetlbfs_pagecache_present(struct hstate *h,
4787                         struct vm_area_struct *vma, unsigned long address)
4788 {
4789         struct address_space *mapping;
4790         pgoff_t idx;
4791         struct page *page;
4792
4793         mapping = vma->vm_file->f_mapping;
4794         idx = vma_hugecache_offset(h, vma, address);
4795
4796         page = find_get_page(mapping, idx);
4797         if (page)
4798                 put_page(page);
4799         return page != NULL;
4800 }
4801
4802 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4803                            pgoff_t idx)
4804 {
4805         struct inode *inode = mapping->host;
4806         struct hstate *h = hstate_inode(inode);
4807         int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4808
4809         if (err)
4810                 return err;
4811         ClearHPageRestoreReserve(page);
4812
4813         /*
4814          * set page dirty so that it will not be removed from cache/file
4815          * by non-hugetlbfs specific code paths.
4816          */
4817         set_page_dirty(page);
4818
4819         spin_lock(&inode->i_lock);
4820         inode->i_blocks += blocks_per_huge_page(h);
4821         spin_unlock(&inode->i_lock);
4822         return 0;
4823 }
4824
4825 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
4826                                                   struct address_space *mapping,
4827                                                   pgoff_t idx,
4828                                                   unsigned int flags,
4829                                                   unsigned long haddr,
4830                                                   unsigned long reason)
4831 {
4832         vm_fault_t ret;
4833         u32 hash;
4834         struct vm_fault vmf = {
4835                 .vma = vma,
4836                 .address = haddr,
4837                 .flags = flags,
4838
4839                 /*
4840                  * Hard to debug if it ends up being
4841                  * used by a callee that assumes
4842                  * something about the other
4843                  * uninitialized fields... same as in
4844                  * memory.c
4845                  */
4846         };
4847
4848         /*
4849          * hugetlb_fault_mutex and i_mmap_rwsem must be
4850          * dropped before handling userfault.  Reacquire
4851          * after handling fault to make calling code simpler.
4852          */
4853         hash = hugetlb_fault_mutex_hash(mapping, idx);
4854         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4855         i_mmap_unlock_read(mapping);
4856         ret = handle_userfault(&vmf, reason);
4857         i_mmap_lock_read(mapping);
4858         mutex_lock(&hugetlb_fault_mutex_table[hash]);
4859
4860         return ret;
4861 }
4862
4863 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4864                         struct vm_area_struct *vma,
4865                         struct address_space *mapping, pgoff_t idx,
4866                         unsigned long address, pte_t *ptep, unsigned int flags)
4867 {
4868         struct hstate *h = hstate_vma(vma);
4869         vm_fault_t ret = VM_FAULT_SIGBUS;
4870         int anon_rmap = 0;
4871         unsigned long size;
4872         struct page *page;
4873         pte_t new_pte;
4874         spinlock_t *ptl;
4875         unsigned long haddr = address & huge_page_mask(h);
4876         bool new_page, new_pagecache_page = false;
4877
4878         /*
4879          * Currently, we are forced to kill the process in the event the
4880          * original mapper has unmapped pages from the child due to a failed
4881          * COW. Warn that such a situation has occurred as it may not be obvious
4882          */
4883         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4884                 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4885                            current->pid);
4886                 return ret;
4887         }
4888
4889         /*
4890          * We can not race with truncation due to holding i_mmap_rwsem.
4891          * i_size is modified when holding i_mmap_rwsem, so check here
4892          * once for faults beyond end of file.
4893          */
4894         size = i_size_read(mapping->host) >> huge_page_shift(h);
4895         if (idx >= size)
4896                 goto out;
4897
4898 retry:
4899         new_page = false;
4900         page = find_lock_page(mapping, idx);
4901         if (!page) {
4902                 /* Check for page in userfault range */
4903                 if (userfaultfd_missing(vma)) {
4904                         ret = hugetlb_handle_userfault(vma, mapping, idx,
4905                                                        flags, haddr,
4906                                                        VM_UFFD_MISSING);
4907                         goto out;
4908                 }
4909
4910                 page = alloc_huge_page(vma, haddr, 0);
4911                 if (IS_ERR(page)) {
4912                         /*
4913                          * Returning error will result in faulting task being
4914                          * sent SIGBUS.  The hugetlb fault mutex prevents two
4915                          * tasks from racing to fault in the same page which
4916                          * could result in false unable to allocate errors.
4917                          * Page migration does not take the fault mutex, but
4918                          * does a clear then write of pte's under page table
4919                          * lock.  Page fault code could race with migration,
4920                          * notice the clear pte and try to allocate a page
4921                          * here.  Before returning error, get ptl and make
4922                          * sure there really is no pte entry.
4923                          */
4924                         ptl = huge_pte_lock(h, mm, ptep);
4925                         ret = 0;
4926                         if (huge_pte_none(huge_ptep_get(ptep)))
4927                                 ret = vmf_error(PTR_ERR(page));
4928                         spin_unlock(ptl);
4929                         goto out;
4930                 }
4931                 clear_huge_page(page, address, pages_per_huge_page(h));
4932                 __SetPageUptodate(page);
4933                 new_page = true;
4934
4935                 if (vma->vm_flags & VM_MAYSHARE) {
4936                         int err = huge_add_to_page_cache(page, mapping, idx);
4937                         if (err) {
4938                                 put_page(page);
4939                                 if (err == -EEXIST)
4940                                         goto retry;
4941                                 goto out;
4942                         }
4943                         new_pagecache_page = true;
4944                 } else {
4945                         lock_page(page);
4946                         if (unlikely(anon_vma_prepare(vma))) {
4947                                 ret = VM_FAULT_OOM;
4948                                 goto backout_unlocked;
4949                         }
4950                         anon_rmap = 1;
4951                 }
4952         } else {
4953                 /*
4954                  * If memory error occurs between mmap() and fault, some process
4955                  * don't have hwpoisoned swap entry for errored virtual address.
4956                  * So we need to block hugepage fault by PG_hwpoison bit check.
4957                  */
4958                 if (unlikely(PageHWPoison(page))) {
4959                         ret = VM_FAULT_HWPOISON_LARGE |
4960                                 VM_FAULT_SET_HINDEX(hstate_index(h));
4961                         goto backout_unlocked;
4962                 }
4963
4964                 /* Check for page in userfault range. */
4965                 if (userfaultfd_minor(vma)) {
4966                         unlock_page(page);
4967                         put_page(page);
4968                         ret = hugetlb_handle_userfault(vma, mapping, idx,
4969                                                        flags, haddr,
4970                                                        VM_UFFD_MINOR);
4971                         goto out;
4972                 }
4973         }
4974
4975         /*
4976          * If we are going to COW a private mapping later, we examine the
4977          * pending reservations for this page now. This will ensure that
4978          * any allocations necessary to record that reservation occur outside
4979          * the spinlock.
4980          */
4981         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4982                 if (vma_needs_reservation(h, vma, haddr) < 0) {
4983                         ret = VM_FAULT_OOM;
4984                         goto backout_unlocked;
4985                 }
4986                 /* Just decrements count, does not deallocate */
4987                 vma_end_reservation(h, vma, haddr);
4988         }
4989
4990         ptl = huge_pte_lock(h, mm, ptep);
4991         ret = 0;
4992         if (!huge_pte_none(huge_ptep_get(ptep)))
4993                 goto backout;
4994
4995         if (anon_rmap) {
4996                 ClearHPageRestoreReserve(page);
4997                 hugepage_add_new_anon_rmap(page, vma, haddr);
4998         } else
4999                 page_dup_rmap(page, true);
5000         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
5001                                 && (vma->vm_flags & VM_SHARED)));
5002         set_huge_pte_at(mm, haddr, ptep, new_pte);
5003
5004         hugetlb_count_add(pages_per_huge_page(h), mm);
5005         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5006                 /* Optimization, do the COW without a second fault */
5007                 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
5008         }
5009
5010         spin_unlock(ptl);
5011
5012         /*
5013          * Only set HPageMigratable in newly allocated pages.  Existing pages
5014          * found in the pagecache may not have HPageMigratableset if they have
5015          * been isolated for migration.
5016          */
5017         if (new_page)
5018                 SetHPageMigratable(page);
5019
5020         unlock_page(page);
5021 out:
5022         return ret;
5023
5024 backout:
5025         spin_unlock(ptl);
5026 backout_unlocked:
5027         unlock_page(page);
5028         /* restore reserve for newly allocated pages not in page cache */
5029         if (new_page && !new_pagecache_page)
5030                 restore_reserve_on_error(h, vma, haddr, page);
5031         put_page(page);
5032         goto out;
5033 }
5034
5035 #ifdef CONFIG_SMP
5036 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5037 {
5038         unsigned long key[2];
5039         u32 hash;
5040
5041         key[0] = (unsigned long) mapping;
5042         key[1] = idx;
5043
5044         hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
5045
5046         return hash & (num_fault_mutexes - 1);
5047 }
5048 #else
5049 /*
5050  * For uniprocessor systems we always use a single mutex, so just
5051  * return 0 and avoid the hashing overhead.
5052  */
5053 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5054 {
5055         return 0;
5056 }
5057 #endif
5058
5059 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
5060                         unsigned long address, unsigned int flags)
5061 {
5062         pte_t *ptep, entry;
5063         spinlock_t *ptl;
5064         vm_fault_t ret;
5065         u32 hash;
5066         pgoff_t idx;
5067         struct page *page = NULL;
5068         struct page *pagecache_page = NULL;
5069         struct hstate *h = hstate_vma(vma);
5070         struct address_space *mapping;
5071         int need_wait_lock = 0;
5072         unsigned long haddr = address & huge_page_mask(h);
5073
5074         ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5075         if (ptep) {
5076                 /*
5077                  * Since we hold no locks, ptep could be stale.  That is
5078                  * OK as we are only making decisions based on content and
5079                  * not actually modifying content here.
5080                  */
5081                 entry = huge_ptep_get(ptep);
5082                 if (unlikely(is_hugetlb_entry_migration(entry))) {
5083                         migration_entry_wait_huge(vma, mm, ptep);
5084                         return 0;
5085                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
5086                         return VM_FAULT_HWPOISON_LARGE |
5087                                 VM_FAULT_SET_HINDEX(hstate_index(h));
5088         }
5089
5090         /*
5091          * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
5092          * until finished with ptep.  This serves two purposes:
5093          * 1) It prevents huge_pmd_unshare from being called elsewhere
5094          *    and making the ptep no longer valid.
5095          * 2) It synchronizes us with i_size modifications during truncation.
5096          *
5097          * ptep could have already be assigned via huge_pte_offset.  That
5098          * is OK, as huge_pte_alloc will return the same value unless
5099          * something has changed.
5100          */
5101         mapping = vma->vm_file->f_mapping;
5102         i_mmap_lock_read(mapping);
5103         ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
5104         if (!ptep) {
5105                 i_mmap_unlock_read(mapping);
5106                 return VM_FAULT_OOM;
5107         }
5108
5109         /*
5110          * Serialize hugepage allocation and instantiation, so that we don't
5111          * get spurious allocation failures if two CPUs race to instantiate
5112          * the same page in the page cache.
5113          */
5114         idx = vma_hugecache_offset(h, vma, haddr);
5115         hash = hugetlb_fault_mutex_hash(mapping, idx);
5116         mutex_lock(&hugetlb_fault_mutex_table[hash]);
5117
5118         entry = huge_ptep_get(ptep);
5119         if (huge_pte_none(entry)) {
5120                 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
5121                 goto out_mutex;
5122         }
5123
5124         ret = 0;
5125
5126         /*
5127          * entry could be a migration/hwpoison entry at this point, so this
5128          * check prevents the kernel from going below assuming that we have
5129          * an active hugepage in pagecache. This goto expects the 2nd page
5130          * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
5131          * properly handle it.
5132          */
5133         if (!pte_present(entry))
5134                 goto out_mutex;
5135
5136         /*
5137          * If we are going to COW the mapping later, we examine the pending
5138          * reservations for this page now. This will ensure that any
5139          * allocations necessary to record that reservation occur outside the
5140          * spinlock. For private mappings, we also lookup the pagecache
5141          * page now as it is used to determine if a reservation has been
5142          * consumed.
5143          */
5144         if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
5145                 if (vma_needs_reservation(h, vma, haddr) < 0) {
5146                         ret = VM_FAULT_OOM;
5147                         goto out_mutex;
5148                 }
5149                 /* Just decrements count, does not deallocate */
5150                 vma_end_reservation(h, vma, haddr);
5151
5152                 if (!(vma->vm_flags & VM_MAYSHARE))
5153                         pagecache_page = hugetlbfs_pagecache_page(h,
5154                                                                 vma, haddr);
5155         }
5156
5157         ptl = huge_pte_lock(h, mm, ptep);
5158
5159         /* Check for a racing update before calling hugetlb_cow */
5160         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
5161                 goto out_ptl;
5162
5163         /*
5164          * hugetlb_cow() requires page locks of pte_page(entry) and
5165          * pagecache_page, so here we need take the former one
5166          * when page != pagecache_page or !pagecache_page.
5167          */
5168         page = pte_page(entry);
5169         if (page != pagecache_page)
5170                 if (!trylock_page(page)) {
5171                         need_wait_lock = 1;
5172                         goto out_ptl;
5173                 }
5174
5175         get_page(page);
5176
5177         if (flags & FAULT_FLAG_WRITE) {
5178                 if (!huge_pte_write(entry)) {
5179                         ret = hugetlb_cow(mm, vma, address, ptep,
5180                                           pagecache_page, ptl);
5181                         goto out_put_page;
5182                 }
5183                 entry = huge_pte_mkdirty(entry);
5184         }
5185         entry = pte_mkyoung(entry);
5186         if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
5187                                                 flags & FAULT_FLAG_WRITE))
5188                 update_mmu_cache(vma, haddr, ptep);
5189 out_put_page:
5190         if (page != pagecache_page)
5191                 unlock_page(page);
5192         put_page(page);
5193 out_ptl:
5194         spin_unlock(ptl);
5195
5196         if (pagecache_page) {
5197                 unlock_page(pagecache_page);
5198                 put_page(pagecache_page);
5199         }
5200 out_mutex:
5201         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5202         i_mmap_unlock_read(mapping);
5203         /*
5204          * Generally it's safe to hold refcount during waiting page lock. But
5205          * here we just wait to defer the next page fault to avoid busy loop and
5206          * the page is not used after unlocked before returning from the current
5207          * page fault. So we are safe from accessing freed page, even if we wait
5208          * here without taking refcount.
5209          */
5210         if (need_wait_lock)
5211                 wait_on_page_locked(page);
5212         return ret;
5213 }
5214
5215 #ifdef CONFIG_USERFAULTFD
5216 /*
5217  * Used by userfaultfd UFFDIO_COPY.  Based on mcopy_atomic_pte with
5218  * modifications for huge pages.
5219  */
5220 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
5221                             pte_t *dst_pte,
5222                             struct vm_area_struct *dst_vma,
5223                             unsigned long dst_addr,
5224                             unsigned long src_addr,
5225                             enum mcopy_atomic_mode mode,
5226                             struct page **pagep)
5227 {
5228         bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE);
5229         struct hstate *h = hstate_vma(dst_vma);
5230         struct address_space *mapping = dst_vma->vm_file->f_mapping;
5231         pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
5232         unsigned long size;
5233         int vm_shared = dst_vma->vm_flags & VM_SHARED;
5234         pte_t _dst_pte;
5235         spinlock_t *ptl;
5236         int ret = -ENOMEM;
5237         struct page *page;
5238         int writable;
5239         bool new_pagecache_page = false;
5240
5241         if (is_continue) {
5242                 ret = -EFAULT;
5243                 page = find_lock_page(mapping, idx);
5244                 if (!page)
5245                         goto out;
5246         } else if (!*pagep) {
5247                 /* If a page already exists, then it's UFFDIO_COPY for
5248                  * a non-missing case. Return -EEXIST.
5249                  */
5250                 if (vm_shared &&
5251                     hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5252                         ret = -EEXIST;
5253                         goto out;
5254                 }
5255
5256                 page = alloc_huge_page(dst_vma, dst_addr, 0);
5257                 if (IS_ERR(page)) {
5258                         ret = -ENOMEM;
5259                         goto out;
5260                 }
5261
5262                 ret = copy_huge_page_from_user(page,
5263                                                 (const void __user *) src_addr,
5264                                                 pages_per_huge_page(h), false);
5265
5266                 /* fallback to copy_from_user outside mmap_lock */
5267                 if (unlikely(ret)) {
5268                         ret = -ENOENT;
5269                         /* Free the allocated page which may have
5270                          * consumed a reservation.
5271                          */
5272                         restore_reserve_on_error(h, dst_vma, dst_addr, page);
5273                         put_page(page);
5274
5275                         /* Allocate a temporary page to hold the copied
5276                          * contents.
5277                          */
5278                         page = alloc_huge_page_vma(h, dst_vma, dst_addr);
5279                         if (!page) {
5280                                 ret = -ENOMEM;
5281                                 goto out;
5282                         }
5283                         *pagep = page;
5284                         /* Set the outparam pagep and return to the caller to
5285                          * copy the contents outside the lock. Don't free the
5286                          * page.
5287                          */
5288                         goto out;
5289                 }
5290         } else {
5291                 if (vm_shared &&
5292                     hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5293                         put_page(*pagep);
5294                         ret = -EEXIST;
5295                         *pagep = NULL;
5296                         goto out;
5297                 }
5298
5299                 page = alloc_huge_page(dst_vma, dst_addr, 0);
5300                 if (IS_ERR(page)) {
5301                         ret = -ENOMEM;
5302                         *pagep = NULL;
5303                         goto out;
5304                 }
5305                 copy_huge_page(page, *pagep);
5306                 put_page(*pagep);
5307                 *pagep = NULL;
5308         }
5309
5310         /*
5311          * The memory barrier inside __SetPageUptodate makes sure that
5312          * preceding stores to the page contents become visible before
5313          * the set_pte_at() write.
5314          */
5315         __SetPageUptodate(page);
5316
5317         /* Add shared, newly allocated pages to the page cache. */
5318         if (vm_shared && !is_continue) {
5319                 size = i_size_read(mapping->host) >> huge_page_shift(h);
5320                 ret = -EFAULT;
5321                 if (idx >= size)
5322                         goto out_release_nounlock;
5323
5324                 /*
5325                  * Serialization between remove_inode_hugepages() and
5326                  * huge_add_to_page_cache() below happens through the
5327                  * hugetlb_fault_mutex_table that here must be hold by
5328                  * the caller.
5329                  */
5330                 ret = huge_add_to_page_cache(page, mapping, idx);
5331                 if (ret)
5332                         goto out_release_nounlock;
5333                 new_pagecache_page = true;
5334         }
5335
5336         ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
5337         spin_lock(ptl);
5338
5339         /*
5340          * Recheck the i_size after holding PT lock to make sure not
5341          * to leave any page mapped (as page_mapped()) beyond the end
5342          * of the i_size (remove_inode_hugepages() is strict about
5343          * enforcing that). If we bail out here, we'll also leave a
5344          * page in the radix tree in the vm_shared case beyond the end
5345          * of the i_size, but remove_inode_hugepages() will take care
5346          * of it as soon as we drop the hugetlb_fault_mutex_table.
5347          */
5348         size = i_size_read(mapping->host) >> huge_page_shift(h);
5349         ret = -EFAULT;
5350         if (idx >= size)
5351                 goto out_release_unlock;
5352
5353         ret = -EEXIST;
5354         if (!huge_pte_none(huge_ptep_get(dst_pte)))
5355                 goto out_release_unlock;
5356
5357         if (vm_shared) {
5358                 page_dup_rmap(page, true);
5359         } else {
5360                 ClearHPageRestoreReserve(page);
5361                 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
5362         }
5363
5364         /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */
5365         if (is_continue && !vm_shared)
5366                 writable = 0;
5367         else
5368                 writable = dst_vma->vm_flags & VM_WRITE;
5369
5370         _dst_pte = make_huge_pte(dst_vma, page, writable);
5371         if (writable)
5372                 _dst_pte = huge_pte_mkdirty(_dst_pte);
5373         _dst_pte = pte_mkyoung(_dst_pte);
5374
5375         set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
5376
5377         (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
5378                                         dst_vma->vm_flags & VM_WRITE);
5379         hugetlb_count_add(pages_per_huge_page(h), dst_mm);
5380
5381         /* No need to invalidate - it was non-present before */
5382         update_mmu_cache(dst_vma, dst_addr, dst_pte);
5383
5384         spin_unlock(ptl);
5385         if (!is_continue)
5386                 SetHPageMigratable(page);
5387         if (vm_shared || is_continue)
5388                 unlock_page(page);
5389         ret = 0;
5390 out:
5391         return ret;
5392 out_release_unlock:
5393         spin_unlock(ptl);
5394         if (vm_shared || is_continue)
5395                 unlock_page(page);
5396 out_release_nounlock:
5397         if (!new_pagecache_page)
5398                 restore_reserve_on_error(h, dst_vma, dst_addr, page);
5399         put_page(page);
5400         goto out;
5401 }
5402 #endif /* CONFIG_USERFAULTFD */
5403
5404 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
5405                                  int refs, struct page **pages,
5406                                  struct vm_area_struct **vmas)
5407 {
5408         int nr;
5409
5410         for (nr = 0; nr < refs; nr++) {
5411                 if (likely(pages))
5412                         pages[nr] = mem_map_offset(page, nr);
5413                 if (vmas)
5414                         vmas[nr] = vma;
5415         }
5416 }
5417
5418 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
5419                          struct page **pages, struct vm_area_struct **vmas,
5420                          unsigned long *position, unsigned long *nr_pages,
5421                          long i, unsigned int flags, int *locked)
5422 {
5423         unsigned long pfn_offset;
5424         unsigned long vaddr = *position;
5425         unsigned long remainder = *nr_pages;
5426         struct hstate *h = hstate_vma(vma);
5427         int err = -EFAULT, refs;
5428
5429         while (vaddr < vma->vm_end && remainder) {
5430                 pte_t *pte;
5431                 spinlock_t *ptl = NULL;
5432                 int absent;
5433                 struct page *page;
5434
5435                 /*
5436                  * If we have a pending SIGKILL, don't keep faulting pages and
5437                  * potentially allocating memory.
5438                  */
5439                 if (fatal_signal_pending(current)) {
5440                         remainder = 0;
5441                         break;
5442                 }
5443
5444                 /*
5445                  * Some archs (sparc64, sh*) have multiple pte_ts to
5446                  * each hugepage.  We have to make sure we get the
5447                  * first, for the page indexing below to work.
5448                  *
5449                  * Note that page table lock is not held when pte is null.
5450                  */
5451                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
5452                                       huge_page_size(h));
5453                 if (pte)
5454                         ptl = huge_pte_lock(h, mm, pte);
5455                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
5456
5457                 /*
5458                  * When coredumping, it suits get_dump_page if we just return
5459                  * an error where there's an empty slot with no huge pagecache
5460                  * to back it.  This way, we avoid allocating a hugepage, and
5461                  * the sparse dumpfile avoids allocating disk blocks, but its
5462                  * huge holes still show up with zeroes where they need to be.
5463                  */
5464                 if (absent && (flags & FOLL_DUMP) &&
5465                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
5466                         if (pte)
5467                                 spin_unlock(ptl);
5468                         remainder = 0;
5469                         break;
5470                 }
5471
5472                 /*
5473                  * We need call hugetlb_fault for both hugepages under migration
5474                  * (in which case hugetlb_fault waits for the migration,) and
5475                  * hwpoisoned hugepages (in which case we need to prevent the
5476                  * caller from accessing to them.) In order to do this, we use
5477                  * here is_swap_pte instead of is_hugetlb_entry_migration and
5478                  * is_hugetlb_entry_hwpoisoned. This is because it simply covers
5479                  * both cases, and because we can't follow correct pages
5480                  * directly from any kind of swap entries.
5481                  */
5482                 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
5483                     ((flags & FOLL_WRITE) &&
5484                       !huge_pte_write(huge_ptep_get(pte)))) {
5485                         vm_fault_t ret;
5486                         unsigned int fault_flags = 0;
5487
5488                         if (pte)
5489                                 spin_unlock(ptl);
5490                         if (flags & FOLL_WRITE)
5491                                 fault_flags |= FAULT_FLAG_WRITE;
5492                         if (locked)
5493                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
5494                                         FAULT_FLAG_KILLABLE;
5495                         if (flags & FOLL_NOWAIT)
5496                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
5497                                         FAULT_FLAG_RETRY_NOWAIT;
5498                         if (flags & FOLL_TRIED) {
5499                                 /*
5500                                  * Note: FAULT_FLAG_ALLOW_RETRY and
5501                                  * FAULT_FLAG_TRIED can co-exist
5502                                  */
5503                                 fault_flags |= FAULT_FLAG_TRIED;
5504                         }
5505                         ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
5506                         if (ret & VM_FAULT_ERROR) {
5507                                 err = vm_fault_to_errno(ret, flags);
5508                                 remainder = 0;
5509                                 break;
5510                         }
5511                         if (ret & VM_FAULT_RETRY) {
5512                                 if (locked &&
5513                                     !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
5514                                         *locked = 0;
5515                                 *nr_pages = 0;
5516                                 /*
5517                                  * VM_FAULT_RETRY must not return an
5518                                  * error, it will return zero
5519                                  * instead.
5520                                  *
5521                                  * No need to update "position" as the
5522                                  * caller will not check it after
5523                                  * *nr_pages is set to 0.
5524                                  */
5525                                 return i;
5526                         }
5527                         continue;
5528                 }
5529
5530                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
5531                 page = pte_page(huge_ptep_get(pte));
5532
5533                 /*
5534                  * If subpage information not requested, update counters
5535                  * and skip the same_page loop below.
5536                  */
5537                 if (!pages && !vmas && !pfn_offset &&
5538                     (vaddr + huge_page_size(h) < vma->vm_end) &&
5539                     (remainder >= pages_per_huge_page(h))) {
5540                         vaddr += huge_page_size(h);
5541                         remainder -= pages_per_huge_page(h);
5542                         i += pages_per_huge_page(h);
5543                         spin_unlock(ptl);
5544                         continue;
5545                 }
5546
5547                 /* vaddr may not be aligned to PAGE_SIZE */
5548                 refs = min3(pages_per_huge_page(h) - pfn_offset, remainder,
5549                     (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT);
5550
5551                 if (pages || vmas)
5552                         record_subpages_vmas(mem_map_offset(page, pfn_offset),
5553                                              vma, refs,
5554                                              likely(pages) ? pages + i : NULL,
5555                                              vmas ? vmas + i : NULL);
5556
5557                 if (pages) {
5558                         /*
5559                          * try_grab_compound_head() should always succeed here,
5560                          * because: a) we hold the ptl lock, and b) we've just
5561                          * checked that the huge page is present in the page
5562                          * tables. If the huge page is present, then the tail
5563                          * pages must also be present. The ptl prevents the
5564                          * head page and tail pages from being rearranged in
5565                          * any way. So this page must be available at this
5566                          * point, unless the page refcount overflowed:
5567                          */
5568                         if (WARN_ON_ONCE(!try_grab_compound_head(pages[i],
5569                                                                  refs,
5570                                                                  flags))) {
5571                                 spin_unlock(ptl);
5572                                 remainder = 0;
5573                                 err = -ENOMEM;
5574                                 break;
5575                         }
5576                 }
5577
5578                 vaddr += (refs << PAGE_SHIFT);
5579                 remainder -= refs;
5580                 i += refs;
5581
5582                 spin_unlock(ptl);
5583         }
5584         *nr_pages = remainder;
5585         /*
5586          * setting position is actually required only if remainder is
5587          * not zero but it's faster not to add a "if (remainder)"
5588          * branch.
5589          */
5590         *position = vaddr;
5591
5592         return i ? i : err;
5593 }
5594
5595 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
5596                 unsigned long address, unsigned long end, pgprot_t newprot)
5597 {
5598         struct mm_struct *mm = vma->vm_mm;
5599         unsigned long start = address;
5600         pte_t *ptep;
5601         pte_t pte;
5602         struct hstate *h = hstate_vma(vma);
5603         unsigned long pages = 0;
5604         bool shared_pmd = false;
5605         struct mmu_notifier_range range;
5606
5607         /*
5608          * In the case of shared PMDs, the area to flush could be beyond
5609          * start/end.  Set range.start/range.end to cover the maximum possible
5610          * range if PMD sharing is possible.
5611          */
5612         mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5613                                 0, vma, mm, start, end);
5614         adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5615
5616         BUG_ON(address >= end);
5617         flush_cache_range(vma, range.start, range.end);
5618
5619         mmu_notifier_invalidate_range_start(&range);
5620         i_mmap_lock_write(vma->vm_file->f_mapping);
5621         for (; address < end; address += huge_page_size(h)) {
5622                 spinlock_t *ptl;
5623                 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5624                 if (!ptep)
5625                         continue;
5626                 ptl = huge_pte_lock(h, mm, ptep);
5627                 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
5628                         pages++;
5629                         spin_unlock(ptl);
5630                         shared_pmd = true;
5631                         continue;
5632                 }
5633                 pte = huge_ptep_get(ptep);
5634                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5635                         spin_unlock(ptl);
5636                         continue;
5637                 }
5638                 if (unlikely(is_hugetlb_entry_migration(pte))) {
5639                         swp_entry_t entry = pte_to_swp_entry(pte);
5640
5641                         if (is_writable_migration_entry(entry)) {
5642                                 pte_t newpte;
5643
5644                                 entry = make_readable_migration_entry(
5645                                                         swp_offset(entry));
5646                                 newpte = swp_entry_to_pte(entry);
5647                                 set_huge_swap_pte_at(mm, address, ptep,
5648                                                      newpte, huge_page_size(h));
5649                                 pages++;
5650                         }
5651                         spin_unlock(ptl);
5652                         continue;
5653                 }
5654                 if (!huge_pte_none(pte)) {
5655                         pte_t old_pte;
5656                         unsigned int shift = huge_page_shift(hstate_vma(vma));
5657
5658                         old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5659                         pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
5660                         pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
5661                         huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
5662                         pages++;
5663                 }
5664                 spin_unlock(ptl);
5665         }
5666         /*
5667          * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5668          * may have cleared our pud entry and done put_page on the page table:
5669          * once we release i_mmap_rwsem, another task can do the final put_page
5670          * and that page table be reused and filled with junk.  If we actually
5671          * did unshare a page of pmds, flush the range corresponding to the pud.
5672          */
5673         if (shared_pmd)
5674                 flush_hugetlb_tlb_range(vma, range.start, range.end);
5675         else
5676                 flush_hugetlb_tlb_range(vma, start, end);
5677         /*
5678          * No need to call mmu_notifier_invalidate_range() we are downgrading
5679          * page table protection not changing it to point to a new page.
5680          *
5681          * See Documentation/vm/mmu_notifier.rst
5682          */
5683         i_mmap_unlock_write(vma->vm_file->f_mapping);
5684         mmu_notifier_invalidate_range_end(&range);
5685
5686         return pages << h->order;
5687 }
5688
5689 /* Return true if reservation was successful, false otherwise.  */
5690 bool hugetlb_reserve_pages(struct inode *inode,
5691                                         long from, long to,
5692                                         struct vm_area_struct *vma,
5693                                         vm_flags_t vm_flags)
5694 {
5695         long chg, add = -1;
5696         struct hstate *h = hstate_inode(inode);
5697         struct hugepage_subpool *spool = subpool_inode(inode);
5698         struct resv_map *resv_map;
5699         struct hugetlb_cgroup *h_cg = NULL;
5700         long gbl_reserve, regions_needed = 0;
5701
5702         /* This should never happen */
5703         if (from > to) {
5704                 VM_WARN(1, "%s called with a negative range\n", __func__);
5705                 return false;
5706         }
5707
5708         /*
5709          * Only apply hugepage reservation if asked. At fault time, an
5710          * attempt will be made for VM_NORESERVE to allocate a page
5711          * without using reserves
5712          */
5713         if (vm_flags & VM_NORESERVE)
5714                 return true;
5715
5716         /*
5717          * Shared mappings base their reservation on the number of pages that
5718          * are already allocated on behalf of the file. Private mappings need
5719          * to reserve the full area even if read-only as mprotect() may be
5720          * called to make the mapping read-write. Assume !vma is a shm mapping
5721          */
5722         if (!vma || vma->vm_flags & VM_MAYSHARE) {
5723                 /*
5724                  * resv_map can not be NULL as hugetlb_reserve_pages is only
5725                  * called for inodes for which resv_maps were created (see
5726                  * hugetlbfs_get_inode).
5727                  */
5728                 resv_map = inode_resv_map(inode);
5729
5730                 chg = region_chg(resv_map, from, to, &regions_needed);
5731
5732         } else {
5733                 /* Private mapping. */
5734                 resv_map = resv_map_alloc();
5735                 if (!resv_map)
5736                         return false;
5737
5738                 chg = to - from;
5739
5740                 set_vma_resv_map(vma, resv_map);
5741                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5742         }
5743
5744         if (chg < 0)
5745                 goto out_err;
5746
5747         if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
5748                                 chg * pages_per_huge_page(h), &h_cg) < 0)
5749                 goto out_err;
5750
5751         if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5752                 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5753                  * of the resv_map.
5754                  */
5755                 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5756         }
5757
5758         /*
5759          * There must be enough pages in the subpool for the mapping. If
5760          * the subpool has a minimum size, there may be some global
5761          * reservations already in place (gbl_reserve).
5762          */
5763         gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5764         if (gbl_reserve < 0)
5765                 goto out_uncharge_cgroup;
5766
5767         /*
5768          * Check enough hugepages are available for the reservation.
5769          * Hand the pages back to the subpool if there are not
5770          */
5771         if (hugetlb_acct_memory(h, gbl_reserve) < 0)
5772                 goto out_put_pages;
5773
5774         /*
5775          * Account for the reservations made. Shared mappings record regions
5776          * that have reservations as they are shared by multiple VMAs.
5777          * When the last VMA disappears, the region map says how much
5778          * the reservation was and the page cache tells how much of
5779          * the reservation was consumed. Private mappings are per-VMA and
5780          * only the consumed reservations are tracked. When the VMA
5781          * disappears, the original reservation is the VMA size and the
5782          * consumed reservations are stored in the map. Hence, nothing
5783          * else has to be done for private mappings here
5784          */
5785         if (!vma || vma->vm_flags & VM_MAYSHARE) {
5786                 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5787
5788                 if (unlikely(add < 0)) {
5789                         hugetlb_acct_memory(h, -gbl_reserve);
5790                         goto out_put_pages;
5791                 } else if (unlikely(chg > add)) {
5792                         /*
5793                          * pages in this range were added to the reserve
5794                          * map between region_chg and region_add.  This
5795                          * indicates a race with alloc_huge_page.  Adjust
5796                          * the subpool and reserve counts modified above
5797                          * based on the difference.
5798                          */
5799                         long rsv_adjust;
5800
5801                         /*
5802                          * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
5803                          * reference to h_cg->css. See comment below for detail.
5804                          */
5805                         hugetlb_cgroup_uncharge_cgroup_rsvd(
5806                                 hstate_index(h),
5807                                 (chg - add) * pages_per_huge_page(h), h_cg);
5808
5809                         rsv_adjust = hugepage_subpool_put_pages(spool,
5810                                                                 chg - add);
5811                         hugetlb_acct_memory(h, -rsv_adjust);
5812                 } else if (h_cg) {
5813                         /*
5814                          * The file_regions will hold their own reference to
5815                          * h_cg->css. So we should release the reference held
5816                          * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
5817                          * done.
5818                          */
5819                         hugetlb_cgroup_put_rsvd_cgroup(h_cg);
5820                 }
5821         }
5822         return true;
5823
5824 out_put_pages:
5825         /* put back original number of pages, chg */
5826         (void)hugepage_subpool_put_pages(spool, chg);
5827 out_uncharge_cgroup:
5828         hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5829                                             chg * pages_per_huge_page(h), h_cg);
5830 out_err:
5831         if (!vma || vma->vm_flags & VM_MAYSHARE)
5832                 /* Only call region_abort if the region_chg succeeded but the
5833                  * region_add failed or didn't run.
5834                  */
5835                 if (chg >= 0 && add < 0)
5836                         region_abort(resv_map, from, to, regions_needed);
5837         if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5838                 kref_put(&resv_map->refs, resv_map_release);
5839         return false;
5840 }
5841
5842 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5843                                                                 long freed)
5844 {
5845         struct hstate *h = hstate_inode(inode);
5846         struct resv_map *resv_map = inode_resv_map(inode);
5847         long chg = 0;
5848         struct hugepage_subpool *spool = subpool_inode(inode);
5849         long gbl_reserve;
5850
5851         /*
5852          * Since this routine can be called in the evict inode path for all
5853          * hugetlbfs inodes, resv_map could be NULL.
5854          */
5855         if (resv_map) {
5856                 chg = region_del(resv_map, start, end);
5857                 /*
5858                  * region_del() can fail in the rare case where a region
5859                  * must be split and another region descriptor can not be
5860                  * allocated.  If end == LONG_MAX, it will not fail.
5861                  */
5862                 if (chg < 0)
5863                         return chg;
5864         }
5865
5866         spin_lock(&inode->i_lock);
5867         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5868         spin_unlock(&inode->i_lock);
5869
5870         /*
5871          * If the subpool has a minimum size, the number of global
5872          * reservations to be released may be adjusted.
5873          *
5874          * Note that !resv_map implies freed == 0. So (chg - freed)
5875          * won't go negative.
5876          */
5877         gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5878         hugetlb_acct_memory(h, -gbl_reserve);
5879
5880         return 0;
5881 }
5882
5883 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5884 static unsigned long page_table_shareable(struct vm_area_struct *svma,
5885                                 struct vm_area_struct *vma,
5886                                 unsigned long addr, pgoff_t idx)
5887 {
5888         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5889                                 svma->vm_start;
5890         unsigned long sbase = saddr & PUD_MASK;
5891         unsigned long s_end = sbase + PUD_SIZE;
5892
5893         /* Allow segments to share if only one is marked locked */
5894         unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5895         unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5896
5897         /*
5898          * match the virtual addresses, permission and the alignment of the
5899          * page table page.
5900          */
5901         if (pmd_index(addr) != pmd_index(saddr) ||
5902             vm_flags != svm_flags ||
5903             !range_in_vma(svma, sbase, s_end))
5904                 return 0;
5905
5906         return saddr;
5907 }
5908
5909 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5910 {
5911         unsigned long base = addr & PUD_MASK;
5912         unsigned long end = base + PUD_SIZE;
5913
5914         /*
5915          * check on proper vm_flags and page table alignment
5916          */
5917         if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5918                 return true;
5919         return false;
5920 }
5921
5922 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5923 {
5924 #ifdef CONFIG_USERFAULTFD
5925         if (uffd_disable_huge_pmd_share(vma))
5926                 return false;
5927 #endif
5928         return vma_shareable(vma, addr);
5929 }
5930
5931 /*
5932  * Determine if start,end range within vma could be mapped by shared pmd.
5933  * If yes, adjust start and end to cover range associated with possible
5934  * shared pmd mappings.
5935  */
5936 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5937                                 unsigned long *start, unsigned long *end)
5938 {
5939         unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
5940                 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
5941
5942         /*
5943          * vma needs to span at least one aligned PUD size, and the range
5944          * must be at least partially within in.
5945          */
5946         if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
5947                 (*end <= v_start) || (*start >= v_end))
5948                 return;
5949
5950         /* Extend the range to be PUD aligned for a worst case scenario */
5951         if (*start > v_start)
5952                 *start = ALIGN_DOWN(*start, PUD_SIZE);
5953
5954         if (*end < v_end)
5955                 *end = ALIGN(*end, PUD_SIZE);
5956 }
5957
5958 /*
5959  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5960  * and returns the corresponding pte. While this is not necessary for the
5961  * !shared pmd case because we can allocate the pmd later as well, it makes the
5962  * code much cleaner.
5963  *
5964  * This routine must be called with i_mmap_rwsem held in at least read mode if
5965  * sharing is possible.  For hugetlbfs, this prevents removal of any page
5966  * table entries associated with the address space.  This is important as we
5967  * are setting up sharing based on existing page table entries (mappings).
5968  *
5969  * NOTE: This routine is only called from huge_pte_alloc.  Some callers of
5970  * huge_pte_alloc know that sharing is not possible and do not take
5971  * i_mmap_rwsem as a performance optimization.  This is handled by the
5972  * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5973  * only required for subsequent processing.
5974  */
5975 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5976                       unsigned long addr, pud_t *pud)
5977 {
5978         struct address_space *mapping = vma->vm_file->f_mapping;
5979         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5980                         vma->vm_pgoff;
5981         struct vm_area_struct *svma;
5982         unsigned long saddr;
5983         pte_t *spte = NULL;
5984         pte_t *pte;
5985         spinlock_t *ptl;
5986
5987         i_mmap_assert_locked(mapping);
5988         vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5989                 if (svma == vma)
5990                         continue;
5991
5992                 saddr = page_table_shareable(svma, vma, addr, idx);
5993                 if (saddr) {
5994                         spte = huge_pte_offset(svma->vm_mm, saddr,
5995                                                vma_mmu_pagesize(svma));
5996                         if (spte) {
5997                                 get_page(virt_to_page(spte));
5998                                 break;
5999                         }
6000                 }
6001         }
6002
6003         if (!spte)
6004                 goto out;
6005
6006         ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
6007         if (pud_none(*pud)) {
6008                 pud_populate(mm, pud,
6009                                 (pmd_t *)((unsigned long)spte & PAGE_MASK));
6010                 mm_inc_nr_pmds(mm);
6011         } else {
6012                 put_page(virt_to_page(spte));
6013         }
6014         spin_unlock(ptl);
6015 out:
6016         pte = (pte_t *)pmd_alloc(mm, pud, addr);
6017         return pte;
6018 }
6019
6020 /*
6021  * unmap huge page backed by shared pte.
6022  *
6023  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
6024  * indicated by page_count > 1, unmap is achieved by clearing pud and
6025  * decrementing the ref count. If count == 1, the pte page is not shared.
6026  *
6027  * Called with page table lock held and i_mmap_rwsem held in write mode.
6028  *
6029  * returns: 1 successfully unmapped a shared pte page
6030  *          0 the underlying pte page is not shared, or it is the last user
6031  */
6032 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
6033                                         unsigned long *addr, pte_t *ptep)
6034 {
6035         pgd_t *pgd = pgd_offset(mm, *addr);
6036         p4d_t *p4d = p4d_offset(pgd, *addr);
6037         pud_t *pud = pud_offset(p4d, *addr);
6038
6039         i_mmap_assert_write_locked(vma->vm_file->f_mapping);
6040         BUG_ON(page_count(virt_to_page(ptep)) == 0);
6041         if (page_count(virt_to_page(ptep)) == 1)
6042                 return 0;
6043
6044         pud_clear(pud);
6045         put_page(virt_to_page(ptep));
6046         mm_dec_nr_pmds(mm);
6047         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
6048         return 1;
6049 }
6050
6051 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
6052 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
6053                       unsigned long addr, pud_t *pud)
6054 {
6055         return NULL;
6056 }
6057
6058 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
6059                                 unsigned long *addr, pte_t *ptep)
6060 {
6061         return 0;
6062 }
6063
6064 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
6065                                 unsigned long *start, unsigned long *end)
6066 {
6067 }
6068
6069 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6070 {
6071         return false;
6072 }
6073 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
6074
6075 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
6076 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
6077                         unsigned long addr, unsigned long sz)
6078 {
6079         pgd_t *pgd;
6080         p4d_t *p4d;
6081         pud_t *pud;
6082         pte_t *pte = NULL;
6083
6084         pgd = pgd_offset(mm, addr);
6085         p4d = p4d_alloc(mm, pgd, addr);
6086         if (!p4d)
6087                 return NULL;
6088         pud = pud_alloc(mm, p4d, addr);
6089         if (pud) {
6090                 if (sz == PUD_SIZE) {
6091                         pte = (pte_t *)pud;
6092                 } else {
6093                         BUG_ON(sz != PMD_SIZE);
6094                         if (want_pmd_share(vma, addr) && pud_none(*pud))
6095                                 pte = huge_pmd_share(mm, vma, addr, pud);
6096                         else
6097                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
6098                 }
6099         }
6100         BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
6101
6102         return pte;
6103 }
6104
6105 /*
6106  * huge_pte_offset() - Walk the page table to resolve the hugepage
6107  * entry at address @addr
6108  *
6109  * Return: Pointer to page table entry (PUD or PMD) for
6110  * address @addr, or NULL if a !p*d_present() entry is encountered and the
6111  * size @sz doesn't match the hugepage size at this level of the page
6112  * table.
6113  */
6114 pte_t *huge_pte_offset(struct mm_struct *mm,
6115                        unsigned long addr, unsigned long sz)
6116 {
6117         pgd_t *pgd;
6118         p4d_t *p4d;
6119         pud_t *pud;
6120         pmd_t *pmd;
6121
6122         pgd = pgd_offset(mm, addr);
6123         if (!pgd_present(*pgd))
6124                 return NULL;
6125         p4d = p4d_offset(pgd, addr);
6126         if (!p4d_present(*p4d))
6127                 return NULL;
6128
6129         pud = pud_offset(p4d, addr);
6130         if (sz == PUD_SIZE)
6131                 /* must be pud huge, non-present or none */
6132                 return (pte_t *)pud;
6133         if (!pud_present(*pud))
6134                 return NULL;
6135         /* must have a valid entry and size to go further */
6136
6137         pmd = pmd_offset(pud, addr);
6138         /* must be pmd huge, non-present or none */
6139         return (pte_t *)pmd;
6140 }
6141
6142 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
6143
6144 /*
6145  * These functions are overwritable if your architecture needs its own
6146  * behavior.
6147  */
6148 struct page * __weak
6149 follow_huge_addr(struct mm_struct *mm, unsigned long address,
6150                               int write)
6151 {
6152         return ERR_PTR(-EINVAL);
6153 }
6154
6155 struct page * __weak
6156 follow_huge_pd(struct vm_area_struct *vma,
6157                unsigned long address, hugepd_t hpd, int flags, int pdshift)
6158 {
6159         WARN(1, "hugepd follow called with no support for hugepage directory format\n");
6160         return NULL;
6161 }
6162
6163 struct page * __weak
6164 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
6165                 pmd_t *pmd, int flags)
6166 {
6167         struct page *page = NULL;
6168         spinlock_t *ptl;
6169         pte_t pte;
6170
6171         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
6172         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
6173                          (FOLL_PIN | FOLL_GET)))
6174                 return NULL;
6175
6176 retry:
6177         ptl = pmd_lockptr(mm, pmd);
6178         spin_lock(ptl);
6179         /*
6180          * make sure that the address range covered by this pmd is not
6181          * unmapped from other threads.
6182          */
6183         if (!pmd_huge(*pmd))
6184                 goto out;
6185         pte = huge_ptep_get((pte_t *)pmd);
6186         if (pte_present(pte)) {
6187                 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
6188                 /*
6189                  * try_grab_page() should always succeed here, because: a) we
6190                  * hold the pmd (ptl) lock, and b) we've just checked that the
6191                  * huge pmd (head) page is present in the page tables. The ptl
6192                  * prevents the head page and tail pages from being rearranged
6193                  * in any way. So this page must be available at this point,
6194                  * unless the page refcount overflowed:
6195                  */
6196                 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
6197                         page = NULL;
6198                         goto out;
6199                 }
6200         } else {
6201                 if (is_hugetlb_entry_migration(pte)) {
6202                         spin_unlock(ptl);
6203                         __migration_entry_wait(mm, (pte_t *)pmd, ptl);
6204                         goto retry;
6205                 }
6206                 /*
6207                  * hwpoisoned entry is treated as no_page_table in
6208                  * follow_page_mask().
6209                  */
6210         }
6211 out:
6212         spin_unlock(ptl);
6213         return page;
6214 }
6215
6216 struct page * __weak
6217 follow_huge_pud(struct mm_struct *mm, unsigned long address,
6218                 pud_t *pud, int flags)
6219 {
6220         if (flags & (FOLL_GET | FOLL_PIN))
6221                 return NULL;
6222
6223         return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
6224 }
6225
6226 struct page * __weak
6227 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
6228 {
6229         if (flags & (FOLL_GET | FOLL_PIN))
6230                 return NULL;
6231
6232         return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
6233 }
6234
6235 bool isolate_huge_page(struct page *page, struct list_head *list)
6236 {
6237         bool ret = true;
6238
6239         spin_lock_irq(&hugetlb_lock);
6240         if (!PageHeadHuge(page) ||
6241             !HPageMigratable(page) ||
6242             !get_page_unless_zero(page)) {
6243                 ret = false;
6244                 goto unlock;
6245         }
6246         ClearHPageMigratable(page);
6247         list_move_tail(&page->lru, list);
6248 unlock:
6249         spin_unlock_irq(&hugetlb_lock);
6250         return ret;
6251 }
6252
6253 int get_hwpoison_huge_page(struct page *page, bool *hugetlb)
6254 {
6255         int ret = 0;
6256
6257         *hugetlb = false;
6258         spin_lock_irq(&hugetlb_lock);
6259         if (PageHeadHuge(page)) {
6260                 *hugetlb = true;
6261                 if (HPageFreed(page) || HPageMigratable(page))
6262                         ret = get_page_unless_zero(page);
6263                 else
6264                         ret = -EBUSY;
6265         }
6266         spin_unlock_irq(&hugetlb_lock);
6267         return ret;
6268 }
6269
6270 void putback_active_hugepage(struct page *page)
6271 {
6272         spin_lock_irq(&hugetlb_lock);
6273         SetHPageMigratable(page);
6274         list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
6275         spin_unlock_irq(&hugetlb_lock);
6276         put_page(page);
6277 }
6278
6279 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
6280 {
6281         struct hstate *h = page_hstate(oldpage);
6282
6283         hugetlb_cgroup_migrate(oldpage, newpage);
6284         set_page_owner_migrate_reason(newpage, reason);
6285
6286         /*
6287          * transfer temporary state of the new huge page. This is
6288          * reverse to other transitions because the newpage is going to
6289          * be final while the old one will be freed so it takes over
6290          * the temporary status.
6291          *
6292          * Also note that we have to transfer the per-node surplus state
6293          * here as well otherwise the global surplus count will not match
6294          * the per-node's.
6295          */
6296         if (HPageTemporary(newpage)) {
6297                 int old_nid = page_to_nid(oldpage);
6298                 int new_nid = page_to_nid(newpage);
6299
6300                 SetHPageTemporary(oldpage);
6301                 ClearHPageTemporary(newpage);
6302
6303                 /*
6304                  * There is no need to transfer the per-node surplus state
6305                  * when we do not cross the node.
6306                  */
6307                 if (new_nid == old_nid)
6308                         return;
6309                 spin_lock_irq(&hugetlb_lock);
6310                 if (h->surplus_huge_pages_node[old_nid]) {
6311                         h->surplus_huge_pages_node[old_nid]--;
6312                         h->surplus_huge_pages_node[new_nid]++;
6313                 }
6314                 spin_unlock_irq(&hugetlb_lock);
6315         }
6316 }
6317
6318 /*
6319  * This function will unconditionally remove all the shared pmd pgtable entries
6320  * within the specific vma for a hugetlbfs memory range.
6321  */
6322 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
6323 {
6324         struct hstate *h = hstate_vma(vma);
6325         unsigned long sz = huge_page_size(h);
6326         struct mm_struct *mm = vma->vm_mm;
6327         struct mmu_notifier_range range;
6328         unsigned long address, start, end;
6329         spinlock_t *ptl;
6330         pte_t *ptep;
6331
6332         if (!(vma->vm_flags & VM_MAYSHARE))
6333                 return;
6334
6335         start = ALIGN(vma->vm_start, PUD_SIZE);
6336         end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6337
6338         if (start >= end)
6339                 return;
6340
6341         /*
6342          * No need to call adjust_range_if_pmd_sharing_possible(), because
6343          * we have already done the PUD_SIZE alignment.
6344          */
6345         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
6346                                 start, end);
6347         mmu_notifier_invalidate_range_start(&range);
6348         i_mmap_lock_write(vma->vm_file->f_mapping);
6349         for (address = start; address < end; address += PUD_SIZE) {
6350                 unsigned long tmp = address;
6351
6352                 ptep = huge_pte_offset(mm, address, sz);
6353                 if (!ptep)
6354                         continue;
6355                 ptl = huge_pte_lock(h, mm, ptep);
6356                 /* We don't want 'address' to be changed */
6357                 huge_pmd_unshare(mm, vma, &tmp, ptep);
6358                 spin_unlock(ptl);
6359         }
6360         flush_hugetlb_tlb_range(vma, start, end);
6361         i_mmap_unlock_write(vma->vm_file->f_mapping);
6362         /*
6363          * No need to call mmu_notifier_invalidate_range(), see
6364          * Documentation/vm/mmu_notifier.rst.
6365          */
6366         mmu_notifier_invalidate_range_end(&range);
6367 }
6368
6369 #ifdef CONFIG_CMA
6370 static bool cma_reserve_called __initdata;
6371
6372 static int __init cmdline_parse_hugetlb_cma(char *p)
6373 {
6374         hugetlb_cma_size = memparse(p, &p);
6375         return 0;
6376 }
6377
6378 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
6379
6380 void __init hugetlb_cma_reserve(int order)
6381 {
6382         unsigned long size, reserved, per_node;
6383         int nid;
6384
6385         cma_reserve_called = true;
6386
6387         if (!hugetlb_cma_size)
6388                 return;
6389
6390         if (hugetlb_cma_size < (PAGE_SIZE << order)) {
6391                 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
6392                         (PAGE_SIZE << order) / SZ_1M);
6393                 return;
6394         }
6395
6396         /*
6397          * If 3 GB area is requested on a machine with 4 numa nodes,
6398          * let's allocate 1 GB on first three nodes and ignore the last one.
6399          */
6400         per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
6401         pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
6402                 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
6403
6404         reserved = 0;
6405         for_each_node_state(nid, N_ONLINE) {
6406                 int res;
6407                 char name[CMA_MAX_NAME];
6408
6409                 size = min(per_node, hugetlb_cma_size - reserved);
6410                 size = round_up(size, PAGE_SIZE << order);
6411
6412                 snprintf(name, sizeof(name), "hugetlb%d", nid);
6413                 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
6414                                                  0, false, name,
6415                                                  &hugetlb_cma[nid], nid);
6416                 if (res) {
6417                         pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
6418                                 res, nid);
6419                         continue;
6420                 }
6421
6422                 reserved += size;
6423                 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
6424                         size / SZ_1M, nid);
6425
6426                 if (reserved >= hugetlb_cma_size)
6427                         break;
6428         }
6429 }
6430
6431 void __init hugetlb_cma_check(void)
6432 {
6433         if (!hugetlb_cma_size || cma_reserve_called)
6434                 return;
6435
6436         pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
6437 }
6438
6439 #endif /* CONFIG_CMA */