Merge tag 'fuse-update-5.14' of git://git.kernel.org/pub/scm/linux/kernel/git/mszered...
[linux-2.6-microblaze.git] / mm / hugetlb.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Generic hugetlb support.
4  * (C) Nadia Yvette Chambers, April 2004
5  */
6 #include <linux/list.h>
7 #include <linux/init.h>
8 #include <linux/mm.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/sched/mm.h>
23 #include <linux/mmdebug.h>
24 #include <linux/sched/signal.h>
25 #include <linux/rmap.h>
26 #include <linux/string_helpers.h>
27 #include <linux/swap.h>
28 #include <linux/swapops.h>
29 #include <linux/jhash.h>
30 #include <linux/numa.h>
31 #include <linux/llist.h>
32 #include <linux/cma.h>
33 #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         list_move(&page->lru, &h->hugepage_freelists[nid]);
1076         h->free_huge_pages++;
1077         h->free_huge_pages_node[nid]++;
1078         SetHPageFreed(page);
1079 }
1080
1081 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1082 {
1083         struct page *page;
1084         bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1085
1086         lockdep_assert_held(&hugetlb_lock);
1087         list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1088                 if (pin && !is_pinnable_page(page))
1089                         continue;
1090
1091                 if (PageHWPoison(page))
1092                         continue;
1093
1094                 list_move(&page->lru, &h->hugepage_activelist);
1095                 set_page_refcounted(page);
1096                 ClearHPageFreed(page);
1097                 h->free_huge_pages--;
1098                 h->free_huge_pages_node[nid]--;
1099                 return page;
1100         }
1101
1102         return NULL;
1103 }
1104
1105 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1106                 nodemask_t *nmask)
1107 {
1108         unsigned int cpuset_mems_cookie;
1109         struct zonelist *zonelist;
1110         struct zone *zone;
1111         struct zoneref *z;
1112         int node = NUMA_NO_NODE;
1113
1114         zonelist = node_zonelist(nid, gfp_mask);
1115
1116 retry_cpuset:
1117         cpuset_mems_cookie = read_mems_allowed_begin();
1118         for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1119                 struct page *page;
1120
1121                 if (!cpuset_zone_allowed(zone, gfp_mask))
1122                         continue;
1123                 /*
1124                  * no need to ask again on the same node. Pool is node rather than
1125                  * zone aware
1126                  */
1127                 if (zone_to_nid(zone) == node)
1128                         continue;
1129                 node = zone_to_nid(zone);
1130
1131                 page = dequeue_huge_page_node_exact(h, node);
1132                 if (page)
1133                         return page;
1134         }
1135         if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1136                 goto retry_cpuset;
1137
1138         return NULL;
1139 }
1140
1141 static struct page *dequeue_huge_page_vma(struct hstate *h,
1142                                 struct vm_area_struct *vma,
1143                                 unsigned long address, int avoid_reserve,
1144                                 long chg)
1145 {
1146         struct page *page;
1147         struct mempolicy *mpol;
1148         gfp_t gfp_mask;
1149         nodemask_t *nodemask;
1150         int nid;
1151
1152         /*
1153          * A child process with MAP_PRIVATE mappings created by their parent
1154          * have no page reserves. This check ensures that reservations are
1155          * not "stolen". The child may still get SIGKILLed
1156          */
1157         if (!vma_has_reserves(vma, chg) &&
1158                         h->free_huge_pages - h->resv_huge_pages == 0)
1159                 goto err;
1160
1161         /* If reserves cannot be used, ensure enough pages are in the pool */
1162         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1163                 goto err;
1164
1165         gfp_mask = htlb_alloc_mask(h);
1166         nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1167         page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1168         if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1169                 SetHPageRestoreReserve(page);
1170                 h->resv_huge_pages--;
1171         }
1172
1173         mpol_cond_put(mpol);
1174         return page;
1175
1176 err:
1177         return NULL;
1178 }
1179
1180 /*
1181  * common helper functions for hstate_next_node_to_{alloc|free}.
1182  * We may have allocated or freed a huge page based on a different
1183  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1184  * be outside of *nodes_allowed.  Ensure that we use an allowed
1185  * node for alloc or free.
1186  */
1187 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1188 {
1189         nid = next_node_in(nid, *nodes_allowed);
1190         VM_BUG_ON(nid >= MAX_NUMNODES);
1191
1192         return nid;
1193 }
1194
1195 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1196 {
1197         if (!node_isset(nid, *nodes_allowed))
1198                 nid = next_node_allowed(nid, nodes_allowed);
1199         return nid;
1200 }
1201
1202 /*
1203  * returns the previously saved node ["this node"] from which to
1204  * allocate a persistent huge page for the pool and advance the
1205  * next node from which to allocate, handling wrap at end of node
1206  * mask.
1207  */
1208 static int hstate_next_node_to_alloc(struct hstate *h,
1209                                         nodemask_t *nodes_allowed)
1210 {
1211         int nid;
1212
1213         VM_BUG_ON(!nodes_allowed);
1214
1215         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1216         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1217
1218         return nid;
1219 }
1220
1221 /*
1222  * helper for remove_pool_huge_page() - return the previously saved
1223  * node ["this node"] from which to free a huge page.  Advance the
1224  * next node id whether or not we find a free huge page to free so
1225  * that the next attempt to free addresses the next node.
1226  */
1227 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1228 {
1229         int nid;
1230
1231         VM_BUG_ON(!nodes_allowed);
1232
1233         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1234         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1235
1236         return nid;
1237 }
1238
1239 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)           \
1240         for (nr_nodes = nodes_weight(*mask);                            \
1241                 nr_nodes > 0 &&                                         \
1242                 ((node = hstate_next_node_to_alloc(hs, mask)) || 1);    \
1243                 nr_nodes--)
1244
1245 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)            \
1246         for (nr_nodes = nodes_weight(*mask);                            \
1247                 nr_nodes > 0 &&                                         \
1248                 ((node = hstate_next_node_to_free(hs, mask)) || 1);     \
1249                 nr_nodes--)
1250
1251 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1252 static void destroy_compound_gigantic_page(struct page *page,
1253                                         unsigned int order)
1254 {
1255         int i;
1256         int nr_pages = 1 << order;
1257         struct page *p = page + 1;
1258
1259         atomic_set(compound_mapcount_ptr(page), 0);
1260         atomic_set(compound_pincount_ptr(page), 0);
1261
1262         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1263                 clear_compound_head(p);
1264                 set_page_refcounted(p);
1265         }
1266
1267         set_compound_order(page, 0);
1268         page[1].compound_nr = 0;
1269         __ClearPageHead(page);
1270 }
1271
1272 static void free_gigantic_page(struct page *page, unsigned int order)
1273 {
1274         /*
1275          * If the page isn't allocated using the cma allocator,
1276          * cma_release() returns false.
1277          */
1278 #ifdef CONFIG_CMA
1279         if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1280                 return;
1281 #endif
1282
1283         free_contig_range(page_to_pfn(page), 1 << order);
1284 }
1285
1286 #ifdef CONFIG_CONTIG_ALLOC
1287 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1288                 int nid, nodemask_t *nodemask)
1289 {
1290         unsigned long nr_pages = pages_per_huge_page(h);
1291         if (nid == NUMA_NO_NODE)
1292                 nid = numa_mem_id();
1293
1294 #ifdef CONFIG_CMA
1295         {
1296                 struct page *page;
1297                 int node;
1298
1299                 if (hugetlb_cma[nid]) {
1300                         page = cma_alloc(hugetlb_cma[nid], nr_pages,
1301                                         huge_page_order(h), true);
1302                         if (page)
1303                                 return page;
1304                 }
1305
1306                 if (!(gfp_mask & __GFP_THISNODE)) {
1307                         for_each_node_mask(node, *nodemask) {
1308                                 if (node == nid || !hugetlb_cma[node])
1309                                         continue;
1310
1311                                 page = cma_alloc(hugetlb_cma[node], nr_pages,
1312                                                 huge_page_order(h), true);
1313                                 if (page)
1314                                         return page;
1315                         }
1316                 }
1317         }
1318 #endif
1319
1320         return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1321 }
1322
1323 #else /* !CONFIG_CONTIG_ALLOC */
1324 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1325                                         int nid, nodemask_t *nodemask)
1326 {
1327         return NULL;
1328 }
1329 #endif /* CONFIG_CONTIG_ALLOC */
1330
1331 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1332 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1333                                         int nid, nodemask_t *nodemask)
1334 {
1335         return NULL;
1336 }
1337 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1338 static inline void destroy_compound_gigantic_page(struct page *page,
1339                                                 unsigned int order) { }
1340 #endif
1341
1342 /*
1343  * Remove hugetlb page from lists, and update dtor so that page appears
1344  * as just a compound page.  A reference is held on the page.
1345  *
1346  * Must be called with hugetlb lock held.
1347  */
1348 static void remove_hugetlb_page(struct hstate *h, struct page *page,
1349                                                         bool adjust_surplus)
1350 {
1351         int nid = page_to_nid(page);
1352
1353         VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1354         VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1355
1356         lockdep_assert_held(&hugetlb_lock);
1357         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1358                 return;
1359
1360         list_del(&page->lru);
1361
1362         if (HPageFreed(page)) {
1363                 h->free_huge_pages--;
1364                 h->free_huge_pages_node[nid]--;
1365         }
1366         if (adjust_surplus) {
1367                 h->surplus_huge_pages--;
1368                 h->surplus_huge_pages_node[nid]--;
1369         }
1370
1371         set_page_refcounted(page);
1372         set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1373
1374         h->nr_huge_pages--;
1375         h->nr_huge_pages_node[nid]--;
1376 }
1377
1378 static void add_hugetlb_page(struct hstate *h, struct page *page,
1379                              bool adjust_surplus)
1380 {
1381         int zeroed;
1382         int nid = page_to_nid(page);
1383
1384         VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page);
1385
1386         lockdep_assert_held(&hugetlb_lock);
1387
1388         INIT_LIST_HEAD(&page->lru);
1389         h->nr_huge_pages++;
1390         h->nr_huge_pages_node[nid]++;
1391
1392         if (adjust_surplus) {
1393                 h->surplus_huge_pages++;
1394                 h->surplus_huge_pages_node[nid]++;
1395         }
1396
1397         set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1398         set_page_private(page, 0);
1399         SetHPageVmemmapOptimized(page);
1400
1401         /*
1402          * This page is now managed by the hugetlb allocator and has
1403          * no users -- drop the last reference.
1404          */
1405         zeroed = put_page_testzero(page);
1406         VM_BUG_ON_PAGE(!zeroed, page);
1407         arch_clear_hugepage_flags(page);
1408         enqueue_huge_page(h, page);
1409 }
1410
1411 static void __update_and_free_page(struct hstate *h, struct page *page)
1412 {
1413         int i;
1414         struct page *subpage = page;
1415
1416         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1417                 return;
1418
1419         if (alloc_huge_page_vmemmap(h, page)) {
1420                 spin_lock_irq(&hugetlb_lock);
1421                 /*
1422                  * If we cannot allocate vmemmap pages, just refuse to free the
1423                  * page and put the page back on the hugetlb free list and treat
1424                  * as a surplus page.
1425                  */
1426                 add_hugetlb_page(h, page, true);
1427                 spin_unlock_irq(&hugetlb_lock);
1428                 return;
1429         }
1430
1431         for (i = 0; i < pages_per_huge_page(h);
1432              i++, subpage = mem_map_next(subpage, page, i)) {
1433                 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1434                                 1 << PG_referenced | 1 << PG_dirty |
1435                                 1 << PG_active | 1 << PG_private |
1436                                 1 << PG_writeback);
1437         }
1438         if (hstate_is_gigantic(h)) {
1439                 destroy_compound_gigantic_page(page, huge_page_order(h));
1440                 free_gigantic_page(page, huge_page_order(h));
1441         } else {
1442                 __free_pages(page, huge_page_order(h));
1443         }
1444 }
1445
1446 /*
1447  * As update_and_free_page() can be called under any context, so we cannot
1448  * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1449  * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1450  * the vmemmap pages.
1451  *
1452  * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1453  * freed and frees them one-by-one. As the page->mapping pointer is going
1454  * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1455  * structure of a lockless linked list of huge pages to be freed.
1456  */
1457 static LLIST_HEAD(hpage_freelist);
1458
1459 static void free_hpage_workfn(struct work_struct *work)
1460 {
1461         struct llist_node *node;
1462
1463         node = llist_del_all(&hpage_freelist);
1464
1465         while (node) {
1466                 struct page *page;
1467                 struct hstate *h;
1468
1469                 page = container_of((struct address_space **)node,
1470                                      struct page, mapping);
1471                 node = node->next;
1472                 page->mapping = NULL;
1473                 /*
1474                  * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate()
1475                  * is going to trigger because a previous call to
1476                  * remove_hugetlb_page() will set_compound_page_dtor(page,
1477                  * NULL_COMPOUND_DTOR), so do not use page_hstate() directly.
1478                  */
1479                 h = size_to_hstate(page_size(page));
1480
1481                 __update_and_free_page(h, page);
1482
1483                 cond_resched();
1484         }
1485 }
1486 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1487
1488 static inline void flush_free_hpage_work(struct hstate *h)
1489 {
1490         if (free_vmemmap_pages_per_hpage(h))
1491                 flush_work(&free_hpage_work);
1492 }
1493
1494 static void update_and_free_page(struct hstate *h, struct page *page,
1495                                  bool atomic)
1496 {
1497         if (!HPageVmemmapOptimized(page) || !atomic) {
1498                 __update_and_free_page(h, page);
1499                 return;
1500         }
1501
1502         /*
1503          * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1504          *
1505          * Only call schedule_work() if hpage_freelist is previously
1506          * empty. Otherwise, schedule_work() had been called but the workfn
1507          * hasn't retrieved the list yet.
1508          */
1509         if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist))
1510                 schedule_work(&free_hpage_work);
1511 }
1512
1513 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1514 {
1515         struct page *page, *t_page;
1516
1517         list_for_each_entry_safe(page, t_page, list, lru) {
1518                 update_and_free_page(h, page, false);
1519                 cond_resched();
1520         }
1521 }
1522
1523 struct hstate *size_to_hstate(unsigned long size)
1524 {
1525         struct hstate *h;
1526
1527         for_each_hstate(h) {
1528                 if (huge_page_size(h) == size)
1529                         return h;
1530         }
1531         return NULL;
1532 }
1533
1534 void free_huge_page(struct page *page)
1535 {
1536         /*
1537          * Can't pass hstate in here because it is called from the
1538          * compound page destructor.
1539          */
1540         struct hstate *h = page_hstate(page);
1541         int nid = page_to_nid(page);
1542         struct hugepage_subpool *spool = hugetlb_page_subpool(page);
1543         bool restore_reserve;
1544         unsigned long flags;
1545
1546         VM_BUG_ON_PAGE(page_count(page), page);
1547         VM_BUG_ON_PAGE(page_mapcount(page), page);
1548
1549         hugetlb_set_page_subpool(page, NULL);
1550         page->mapping = NULL;
1551         restore_reserve = HPageRestoreReserve(page);
1552         ClearHPageRestoreReserve(page);
1553
1554         /*
1555          * If HPageRestoreReserve was set on page, page allocation consumed a
1556          * reservation.  If the page was associated with a subpool, there
1557          * would have been a page reserved in the subpool before allocation
1558          * via hugepage_subpool_get_pages().  Since we are 'restoring' the
1559          * reservation, do not call hugepage_subpool_put_pages() as this will
1560          * remove the reserved page from the subpool.
1561          */
1562         if (!restore_reserve) {
1563                 /*
1564                  * A return code of zero implies that the subpool will be
1565                  * under its minimum size if the reservation is not restored
1566                  * after page is free.  Therefore, force restore_reserve
1567                  * operation.
1568                  */
1569                 if (hugepage_subpool_put_pages(spool, 1) == 0)
1570                         restore_reserve = true;
1571         }
1572
1573         spin_lock_irqsave(&hugetlb_lock, flags);
1574         ClearHPageMigratable(page);
1575         hugetlb_cgroup_uncharge_page(hstate_index(h),
1576                                      pages_per_huge_page(h), page);
1577         hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1578                                           pages_per_huge_page(h), page);
1579         if (restore_reserve)
1580                 h->resv_huge_pages++;
1581
1582         if (HPageTemporary(page)) {
1583                 remove_hugetlb_page(h, page, false);
1584                 spin_unlock_irqrestore(&hugetlb_lock, flags);
1585                 update_and_free_page(h, page, true);
1586         } else if (h->surplus_huge_pages_node[nid]) {
1587                 /* remove the page from active list */
1588                 remove_hugetlb_page(h, page, true);
1589                 spin_unlock_irqrestore(&hugetlb_lock, flags);
1590                 update_and_free_page(h, page, true);
1591         } else {
1592                 arch_clear_hugepage_flags(page);
1593                 enqueue_huge_page(h, page);
1594                 spin_unlock_irqrestore(&hugetlb_lock, flags);
1595         }
1596 }
1597
1598 /*
1599  * Must be called with the hugetlb lock held
1600  */
1601 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1602 {
1603         lockdep_assert_held(&hugetlb_lock);
1604         h->nr_huge_pages++;
1605         h->nr_huge_pages_node[nid]++;
1606 }
1607
1608 static void __prep_new_huge_page(struct hstate *h, struct page *page)
1609 {
1610         free_huge_page_vmemmap(h, page);
1611         INIT_LIST_HEAD(&page->lru);
1612         set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1613         hugetlb_set_page_subpool(page, NULL);
1614         set_hugetlb_cgroup(page, NULL);
1615         set_hugetlb_cgroup_rsvd(page, NULL);
1616 }
1617
1618 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1619 {
1620         __prep_new_huge_page(h, page);
1621         spin_lock_irq(&hugetlb_lock);
1622         __prep_account_new_huge_page(h, nid);
1623         spin_unlock_irq(&hugetlb_lock);
1624 }
1625
1626 static bool prep_compound_gigantic_page(struct page *page, unsigned int order)
1627 {
1628         int i, j;
1629         int nr_pages = 1 << order;
1630         struct page *p = page + 1;
1631
1632         /* we rely on prep_new_huge_page to set the destructor */
1633         set_compound_order(page, order);
1634         __ClearPageReserved(page);
1635         __SetPageHead(page);
1636         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1637                 /*
1638                  * For gigantic hugepages allocated through bootmem at
1639                  * boot, it's safer to be consistent with the not-gigantic
1640                  * hugepages and clear the PG_reserved bit from all tail pages
1641                  * too.  Otherwise drivers using get_user_pages() to access tail
1642                  * pages may get the reference counting wrong if they see
1643                  * PG_reserved set on a tail page (despite the head page not
1644                  * having PG_reserved set).  Enforcing this consistency between
1645                  * head and tail pages allows drivers to optimize away a check
1646                  * on the head page when they need know if put_page() is needed
1647                  * after get_user_pages().
1648                  */
1649                 __ClearPageReserved(p);
1650                 /*
1651                  * Subtle and very unlikely
1652                  *
1653                  * Gigantic 'page allocators' such as memblock or cma will
1654                  * return a set of pages with each page ref counted.  We need
1655                  * to turn this set of pages into a compound page with tail
1656                  * page ref counts set to zero.  Code such as speculative page
1657                  * cache adding could take a ref on a 'to be' tail page.
1658                  * We need to respect any increased ref count, and only set
1659                  * the ref count to zero if count is currently 1.  If count
1660                  * is not 1, we call synchronize_rcu in the hope that a rcu
1661                  * grace period will cause ref count to drop and then retry.
1662                  * If count is still inflated on retry we return an error and
1663                  * must discard the pages.
1664                  */
1665                 if (!page_ref_freeze(p, 1)) {
1666                         pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n");
1667                         synchronize_rcu();
1668                         if (!page_ref_freeze(p, 1))
1669                                 goto out_error;
1670                 }
1671                 set_page_count(p, 0);
1672                 set_compound_head(p, page);
1673         }
1674         atomic_set(compound_mapcount_ptr(page), -1);
1675         atomic_set(compound_pincount_ptr(page), 0);
1676         return true;
1677
1678 out_error:
1679         /* undo tail page modifications made above */
1680         p = page + 1;
1681         for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) {
1682                 clear_compound_head(p);
1683                 set_page_refcounted(p);
1684         }
1685         /* need to clear PG_reserved on remaining tail pages  */
1686         for (; j < nr_pages; j++, p = mem_map_next(p, page, j))
1687                 __ClearPageReserved(p);
1688         set_compound_order(page, 0);
1689         page[1].compound_nr = 0;
1690         __ClearPageHead(page);
1691         return false;
1692 }
1693
1694 /*
1695  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1696  * transparent huge pages.  See the PageTransHuge() documentation for more
1697  * details.
1698  */
1699 int PageHuge(struct page *page)
1700 {
1701         if (!PageCompound(page))
1702                 return 0;
1703
1704         page = compound_head(page);
1705         return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1706 }
1707 EXPORT_SYMBOL_GPL(PageHuge);
1708
1709 /*
1710  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1711  * normal or transparent huge pages.
1712  */
1713 int PageHeadHuge(struct page *page_head)
1714 {
1715         if (!PageHead(page_head))
1716                 return 0;
1717
1718         return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1719 }
1720
1721 /*
1722  * Find and lock address space (mapping) in write mode.
1723  *
1724  * Upon entry, the page is locked which means that page_mapping() is
1725  * stable.  Due to locking order, we can only trylock_write.  If we can
1726  * not get the lock, simply return NULL to caller.
1727  */
1728 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1729 {
1730         struct address_space *mapping = page_mapping(hpage);
1731
1732         if (!mapping)
1733                 return mapping;
1734
1735         if (i_mmap_trylock_write(mapping))
1736                 return mapping;
1737
1738         return NULL;
1739 }
1740
1741 pgoff_t hugetlb_basepage_index(struct page *page)
1742 {
1743         struct page *page_head = compound_head(page);
1744         pgoff_t index = page_index(page_head);
1745         unsigned long compound_idx;
1746
1747         if (compound_order(page_head) >= MAX_ORDER)
1748                 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1749         else
1750                 compound_idx = page - page_head;
1751
1752         return (index << compound_order(page_head)) + compound_idx;
1753 }
1754
1755 static struct page *alloc_buddy_huge_page(struct hstate *h,
1756                 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1757                 nodemask_t *node_alloc_noretry)
1758 {
1759         int order = huge_page_order(h);
1760         struct page *page;
1761         bool alloc_try_hard = true;
1762
1763         /*
1764          * By default we always try hard to allocate the page with
1765          * __GFP_RETRY_MAYFAIL flag.  However, if we are allocating pages in
1766          * a loop (to adjust global huge page counts) and previous allocation
1767          * failed, do not continue to try hard on the same node.  Use the
1768          * node_alloc_noretry bitmap to manage this state information.
1769          */
1770         if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1771                 alloc_try_hard = false;
1772         gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1773         if (alloc_try_hard)
1774                 gfp_mask |= __GFP_RETRY_MAYFAIL;
1775         if (nid == NUMA_NO_NODE)
1776                 nid = numa_mem_id();
1777         page = __alloc_pages(gfp_mask, order, nid, nmask);
1778         if (page)
1779                 __count_vm_event(HTLB_BUDDY_PGALLOC);
1780         else
1781                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1782
1783         /*
1784          * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1785          * indicates an overall state change.  Clear bit so that we resume
1786          * normal 'try hard' allocations.
1787          */
1788         if (node_alloc_noretry && page && !alloc_try_hard)
1789                 node_clear(nid, *node_alloc_noretry);
1790
1791         /*
1792          * If we tried hard to get a page but failed, set bit so that
1793          * subsequent attempts will not try as hard until there is an
1794          * overall state change.
1795          */
1796         if (node_alloc_noretry && !page && alloc_try_hard)
1797                 node_set(nid, *node_alloc_noretry);
1798
1799         return page;
1800 }
1801
1802 /*
1803  * Common helper to allocate a fresh hugetlb page. All specific allocators
1804  * should use this function to get new hugetlb pages
1805  */
1806 static struct page *alloc_fresh_huge_page(struct hstate *h,
1807                 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1808                 nodemask_t *node_alloc_noretry)
1809 {
1810         struct page *page;
1811         bool retry = false;
1812
1813 retry:
1814         if (hstate_is_gigantic(h))
1815                 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1816         else
1817                 page = alloc_buddy_huge_page(h, gfp_mask,
1818                                 nid, nmask, node_alloc_noretry);
1819         if (!page)
1820                 return NULL;
1821
1822         if (hstate_is_gigantic(h)) {
1823                 if (!prep_compound_gigantic_page(page, huge_page_order(h))) {
1824                         /*
1825                          * Rare failure to convert pages to compound page.
1826                          * Free pages and try again - ONCE!
1827                          */
1828                         free_gigantic_page(page, huge_page_order(h));
1829                         if (!retry) {
1830                                 retry = true;
1831                                 goto retry;
1832                         }
1833                         pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
1834                         return NULL;
1835                 }
1836         }
1837         prep_new_huge_page(h, page, page_to_nid(page));
1838
1839         return page;
1840 }
1841
1842 /*
1843  * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1844  * manner.
1845  */
1846 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1847                                 nodemask_t *node_alloc_noretry)
1848 {
1849         struct page *page;
1850         int nr_nodes, node;
1851         gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1852
1853         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1854                 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1855                                                 node_alloc_noretry);
1856                 if (page)
1857                         break;
1858         }
1859
1860         if (!page)
1861                 return 0;
1862
1863         put_page(page); /* free it into the hugepage allocator */
1864
1865         return 1;
1866 }
1867
1868 /*
1869  * Remove huge page from pool from next node to free.  Attempt to keep
1870  * persistent huge pages more or less balanced over allowed nodes.
1871  * This routine only 'removes' the hugetlb page.  The caller must make
1872  * an additional call to free the page to low level allocators.
1873  * Called with hugetlb_lock locked.
1874  */
1875 static struct page *remove_pool_huge_page(struct hstate *h,
1876                                                 nodemask_t *nodes_allowed,
1877                                                  bool acct_surplus)
1878 {
1879         int nr_nodes, node;
1880         struct page *page = NULL;
1881
1882         lockdep_assert_held(&hugetlb_lock);
1883         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1884                 /*
1885                  * If we're returning unused surplus pages, only examine
1886                  * nodes with surplus pages.
1887                  */
1888                 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1889                     !list_empty(&h->hugepage_freelists[node])) {
1890                         page = list_entry(h->hugepage_freelists[node].next,
1891                                           struct page, lru);
1892                         remove_hugetlb_page(h, page, acct_surplus);
1893                         break;
1894                 }
1895         }
1896
1897         return page;
1898 }
1899
1900 /*
1901  * Dissolve a given free hugepage into free buddy pages. This function does
1902  * nothing for in-use hugepages and non-hugepages.
1903  * This function returns values like below:
1904  *
1905  *  -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
1906  *           when the system is under memory pressure and the feature of
1907  *           freeing unused vmemmap pages associated with each hugetlb page
1908  *           is enabled.
1909  *  -EBUSY:  failed to dissolved free hugepages or the hugepage is in-use
1910  *           (allocated or reserved.)
1911  *       0:  successfully dissolved free hugepages or the page is not a
1912  *           hugepage (considered as already dissolved)
1913  */
1914 int dissolve_free_huge_page(struct page *page)
1915 {
1916         int rc = -EBUSY;
1917
1918 retry:
1919         /* Not to disrupt normal path by vainly holding hugetlb_lock */
1920         if (!PageHuge(page))
1921                 return 0;
1922
1923         spin_lock_irq(&hugetlb_lock);
1924         if (!PageHuge(page)) {
1925                 rc = 0;
1926                 goto out;
1927         }
1928
1929         if (!page_count(page)) {
1930                 struct page *head = compound_head(page);
1931                 struct hstate *h = page_hstate(head);
1932                 if (h->free_huge_pages - h->resv_huge_pages == 0)
1933                         goto out;
1934
1935                 /*
1936                  * We should make sure that the page is already on the free list
1937                  * when it is dissolved.
1938                  */
1939                 if (unlikely(!HPageFreed(head))) {
1940                         spin_unlock_irq(&hugetlb_lock);
1941                         cond_resched();
1942
1943                         /*
1944                          * Theoretically, we should return -EBUSY when we
1945                          * encounter this race. In fact, we have a chance
1946                          * to successfully dissolve the page if we do a
1947                          * retry. Because the race window is quite small.
1948                          * If we seize this opportunity, it is an optimization
1949                          * for increasing the success rate of dissolving page.
1950                          */
1951                         goto retry;
1952                 }
1953
1954                 remove_hugetlb_page(h, head, false);
1955                 h->max_huge_pages--;
1956                 spin_unlock_irq(&hugetlb_lock);
1957
1958                 /*
1959                  * Normally update_and_free_page will allocate required vmemmmap
1960                  * before freeing the page.  update_and_free_page will fail to
1961                  * free the page if it can not allocate required vmemmap.  We
1962                  * need to adjust max_huge_pages if the page is not freed.
1963                  * Attempt to allocate vmemmmap here so that we can take
1964                  * appropriate action on failure.
1965                  */
1966                 rc = alloc_huge_page_vmemmap(h, head);
1967                 if (!rc) {
1968                         /*
1969                          * Move PageHWPoison flag from head page to the raw
1970                          * error page, which makes any subpages rather than
1971                          * the error page reusable.
1972                          */
1973                         if (PageHWPoison(head) && page != head) {
1974                                 SetPageHWPoison(page);
1975                                 ClearPageHWPoison(head);
1976                         }
1977                         update_and_free_page(h, head, false);
1978                 } else {
1979                         spin_lock_irq(&hugetlb_lock);
1980                         add_hugetlb_page(h, head, false);
1981                         h->max_huge_pages++;
1982                         spin_unlock_irq(&hugetlb_lock);
1983                 }
1984
1985                 return rc;
1986         }
1987 out:
1988         spin_unlock_irq(&hugetlb_lock);
1989         return rc;
1990 }
1991
1992 /*
1993  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1994  * make specified memory blocks removable from the system.
1995  * Note that this will dissolve a free gigantic hugepage completely, if any
1996  * part of it lies within the given range.
1997  * Also note that if dissolve_free_huge_page() returns with an error, all
1998  * free hugepages that were dissolved before that error are lost.
1999  */
2000 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2001 {
2002         unsigned long pfn;
2003         struct page *page;
2004         int rc = 0;
2005
2006         if (!hugepages_supported())
2007                 return rc;
2008
2009         for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
2010                 page = pfn_to_page(pfn);
2011                 rc = dissolve_free_huge_page(page);
2012                 if (rc)
2013                         break;
2014         }
2015
2016         return rc;
2017 }
2018
2019 /*
2020  * Allocates a fresh surplus page from the page allocator.
2021  */
2022 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
2023                 int nid, nodemask_t *nmask)
2024 {
2025         struct page *page = NULL;
2026
2027         if (hstate_is_gigantic(h))
2028                 return NULL;
2029
2030         spin_lock_irq(&hugetlb_lock);
2031         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2032                 goto out_unlock;
2033         spin_unlock_irq(&hugetlb_lock);
2034
2035         page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2036         if (!page)
2037                 return NULL;
2038
2039         spin_lock_irq(&hugetlb_lock);
2040         /*
2041          * We could have raced with the pool size change.
2042          * Double check that and simply deallocate the new page
2043          * if we would end up overcommiting the surpluses. Abuse
2044          * temporary page to workaround the nasty free_huge_page
2045          * codeflow
2046          */
2047         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2048                 SetHPageTemporary(page);
2049                 spin_unlock_irq(&hugetlb_lock);
2050                 put_page(page);
2051                 return NULL;
2052         } else {
2053                 h->surplus_huge_pages++;
2054                 h->surplus_huge_pages_node[page_to_nid(page)]++;
2055         }
2056
2057 out_unlock:
2058         spin_unlock_irq(&hugetlb_lock);
2059
2060         return page;
2061 }
2062
2063 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
2064                                      int nid, nodemask_t *nmask)
2065 {
2066         struct page *page;
2067
2068         if (hstate_is_gigantic(h))
2069                 return NULL;
2070
2071         page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2072         if (!page)
2073                 return NULL;
2074
2075         /*
2076          * We do not account these pages as surplus because they are only
2077          * temporary and will be released properly on the last reference
2078          */
2079         SetHPageTemporary(page);
2080
2081         return page;
2082 }
2083
2084 /*
2085  * Use the VMA's mpolicy to allocate a huge page from the buddy.
2086  */
2087 static
2088 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
2089                 struct vm_area_struct *vma, unsigned long addr)
2090 {
2091         struct page *page;
2092         struct mempolicy *mpol;
2093         gfp_t gfp_mask = htlb_alloc_mask(h);
2094         int nid;
2095         nodemask_t *nodemask;
2096
2097         nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2098         page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
2099         mpol_cond_put(mpol);
2100
2101         return page;
2102 }
2103
2104 /* page migration callback function */
2105 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
2106                 nodemask_t *nmask, gfp_t gfp_mask)
2107 {
2108         spin_lock_irq(&hugetlb_lock);
2109         if (h->free_huge_pages - h->resv_huge_pages > 0) {
2110                 struct page *page;
2111
2112                 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
2113                 if (page) {
2114                         spin_unlock_irq(&hugetlb_lock);
2115                         return page;
2116                 }
2117         }
2118         spin_unlock_irq(&hugetlb_lock);
2119
2120         return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2121 }
2122
2123 /* mempolicy aware migration callback */
2124 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2125                 unsigned long address)
2126 {
2127         struct mempolicy *mpol;
2128         nodemask_t *nodemask;
2129         struct page *page;
2130         gfp_t gfp_mask;
2131         int node;
2132
2133         gfp_mask = htlb_alloc_mask(h);
2134         node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2135         page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
2136         mpol_cond_put(mpol);
2137
2138         return page;
2139 }
2140
2141 /*
2142  * Increase the hugetlb pool such that it can accommodate a reservation
2143  * of size 'delta'.
2144  */
2145 static int gather_surplus_pages(struct hstate *h, long delta)
2146         __must_hold(&hugetlb_lock)
2147 {
2148         struct list_head surplus_list;
2149         struct page *page, *tmp;
2150         int ret;
2151         long i;
2152         long needed, allocated;
2153         bool alloc_ok = true;
2154
2155         lockdep_assert_held(&hugetlb_lock);
2156         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2157         if (needed <= 0) {
2158                 h->resv_huge_pages += delta;
2159                 return 0;
2160         }
2161
2162         allocated = 0;
2163         INIT_LIST_HEAD(&surplus_list);
2164
2165         ret = -ENOMEM;
2166 retry:
2167         spin_unlock_irq(&hugetlb_lock);
2168         for (i = 0; i < needed; i++) {
2169                 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2170                                 NUMA_NO_NODE, NULL);
2171                 if (!page) {
2172                         alloc_ok = false;
2173                         break;
2174                 }
2175                 list_add(&page->lru, &surplus_list);
2176                 cond_resched();
2177         }
2178         allocated += i;
2179
2180         /*
2181          * After retaking hugetlb_lock, we need to recalculate 'needed'
2182          * because either resv_huge_pages or free_huge_pages may have changed.
2183          */
2184         spin_lock_irq(&hugetlb_lock);
2185         needed = (h->resv_huge_pages + delta) -
2186                         (h->free_huge_pages + allocated);
2187         if (needed > 0) {
2188                 if (alloc_ok)
2189                         goto retry;
2190                 /*
2191                  * We were not able to allocate enough pages to
2192                  * satisfy the entire reservation so we free what
2193                  * we've allocated so far.
2194                  */
2195                 goto free;
2196         }
2197         /*
2198          * The surplus_list now contains _at_least_ the number of extra pages
2199          * needed to accommodate the reservation.  Add the appropriate number
2200          * of pages to the hugetlb pool and free the extras back to the buddy
2201          * allocator.  Commit the entire reservation here to prevent another
2202          * process from stealing the pages as they are added to the pool but
2203          * before they are reserved.
2204          */
2205         needed += allocated;
2206         h->resv_huge_pages += delta;
2207         ret = 0;
2208
2209         /* Free the needed pages to the hugetlb pool */
2210         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2211                 int zeroed;
2212
2213                 if ((--needed) < 0)
2214                         break;
2215                 /*
2216                  * This page is now managed by the hugetlb allocator and has
2217                  * no users -- drop the buddy allocator's reference.
2218                  */
2219                 zeroed = put_page_testzero(page);
2220                 VM_BUG_ON_PAGE(!zeroed, page);
2221                 enqueue_huge_page(h, page);
2222         }
2223 free:
2224         spin_unlock_irq(&hugetlb_lock);
2225
2226         /* Free unnecessary surplus pages to the buddy allocator */
2227         list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2228                 put_page(page);
2229         spin_lock_irq(&hugetlb_lock);
2230
2231         return ret;
2232 }
2233
2234 /*
2235  * This routine has two main purposes:
2236  * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2237  *    in unused_resv_pages.  This corresponds to the prior adjustments made
2238  *    to the associated reservation map.
2239  * 2) Free any unused surplus pages that may have been allocated to satisfy
2240  *    the reservation.  As many as unused_resv_pages may be freed.
2241  */
2242 static void return_unused_surplus_pages(struct hstate *h,
2243                                         unsigned long unused_resv_pages)
2244 {
2245         unsigned long nr_pages;
2246         struct page *page;
2247         LIST_HEAD(page_list);
2248
2249         lockdep_assert_held(&hugetlb_lock);
2250         /* Uncommit the reservation */
2251         h->resv_huge_pages -= unused_resv_pages;
2252
2253         /* Cannot return gigantic pages currently */
2254         if (hstate_is_gigantic(h))
2255                 goto out;
2256
2257         /*
2258          * Part (or even all) of the reservation could have been backed
2259          * by pre-allocated pages. Only free surplus pages.
2260          */
2261         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2262
2263         /*
2264          * We want to release as many surplus pages as possible, spread
2265          * evenly across all nodes with memory. Iterate across these nodes
2266          * until we can no longer free unreserved surplus pages. This occurs
2267          * when the nodes with surplus pages have no free pages.
2268          * remove_pool_huge_page() will balance the freed pages across the
2269          * on-line nodes with memory and will handle the hstate accounting.
2270          */
2271         while (nr_pages--) {
2272                 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2273                 if (!page)
2274                         goto out;
2275
2276                 list_add(&page->lru, &page_list);
2277         }
2278
2279 out:
2280         spin_unlock_irq(&hugetlb_lock);
2281         update_and_free_pages_bulk(h, &page_list);
2282         spin_lock_irq(&hugetlb_lock);
2283 }
2284
2285
2286 /*
2287  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2288  * are used by the huge page allocation routines to manage reservations.
2289  *
2290  * vma_needs_reservation is called to determine if the huge page at addr
2291  * within the vma has an associated reservation.  If a reservation is
2292  * needed, the value 1 is returned.  The caller is then responsible for
2293  * managing the global reservation and subpool usage counts.  After
2294  * the huge page has been allocated, vma_commit_reservation is called
2295  * to add the page to the reservation map.  If the page allocation fails,
2296  * the reservation must be ended instead of committed.  vma_end_reservation
2297  * is called in such cases.
2298  *
2299  * In the normal case, vma_commit_reservation returns the same value
2300  * as the preceding vma_needs_reservation call.  The only time this
2301  * is not the case is if a reserve map was changed between calls.  It
2302  * is the responsibility of the caller to notice the difference and
2303  * take appropriate action.
2304  *
2305  * vma_add_reservation is used in error paths where a reservation must
2306  * be restored when a newly allocated huge page must be freed.  It is
2307  * to be called after calling vma_needs_reservation to determine if a
2308  * reservation exists.
2309  *
2310  * vma_del_reservation is used in error paths where an entry in the reserve
2311  * map was created during huge page allocation and must be removed.  It is to
2312  * be called after calling vma_needs_reservation to determine if a reservation
2313  * exists.
2314  */
2315 enum vma_resv_mode {
2316         VMA_NEEDS_RESV,
2317         VMA_COMMIT_RESV,
2318         VMA_END_RESV,
2319         VMA_ADD_RESV,
2320         VMA_DEL_RESV,
2321 };
2322 static long __vma_reservation_common(struct hstate *h,
2323                                 struct vm_area_struct *vma, unsigned long addr,
2324                                 enum vma_resv_mode mode)
2325 {
2326         struct resv_map *resv;
2327         pgoff_t idx;
2328         long ret;
2329         long dummy_out_regions_needed;
2330
2331         resv = vma_resv_map(vma);
2332         if (!resv)
2333                 return 1;
2334
2335         idx = vma_hugecache_offset(h, vma, addr);
2336         switch (mode) {
2337         case VMA_NEEDS_RESV:
2338                 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2339                 /* We assume that vma_reservation_* routines always operate on
2340                  * 1 page, and that adding to resv map a 1 page entry can only
2341                  * ever require 1 region.
2342                  */
2343                 VM_BUG_ON(dummy_out_regions_needed != 1);
2344                 break;
2345         case VMA_COMMIT_RESV:
2346                 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2347                 /* region_add calls of range 1 should never fail. */
2348                 VM_BUG_ON(ret < 0);
2349                 break;
2350         case VMA_END_RESV:
2351                 region_abort(resv, idx, idx + 1, 1);
2352                 ret = 0;
2353                 break;
2354         case VMA_ADD_RESV:
2355                 if (vma->vm_flags & VM_MAYSHARE) {
2356                         ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2357                         /* region_add calls of range 1 should never fail. */
2358                         VM_BUG_ON(ret < 0);
2359                 } else {
2360                         region_abort(resv, idx, idx + 1, 1);
2361                         ret = region_del(resv, idx, idx + 1);
2362                 }
2363                 break;
2364         case VMA_DEL_RESV:
2365                 if (vma->vm_flags & VM_MAYSHARE) {
2366                         region_abort(resv, idx, idx + 1, 1);
2367                         ret = region_del(resv, idx, idx + 1);
2368                 } else {
2369                         ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2370                         /* region_add calls of range 1 should never fail. */
2371                         VM_BUG_ON(ret < 0);
2372                 }
2373                 break;
2374         default:
2375                 BUG();
2376         }
2377
2378         if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2379                 return ret;
2380         /*
2381          * We know private mapping must have HPAGE_RESV_OWNER set.
2382          *
2383          * In most cases, reserves always exist for private mappings.
2384          * However, a file associated with mapping could have been
2385          * hole punched or truncated after reserves were consumed.
2386          * As subsequent fault on such a range will not use reserves.
2387          * Subtle - The reserve map for private mappings has the
2388          * opposite meaning than that of shared mappings.  If NO
2389          * entry is in the reserve map, it means a reservation exists.
2390          * If an entry exists in the reserve map, it means the
2391          * reservation has already been consumed.  As a result, the
2392          * return value of this routine is the opposite of the
2393          * value returned from reserve map manipulation routines above.
2394          */
2395         if (ret > 0)
2396                 return 0;
2397         if (ret == 0)
2398                 return 1;
2399         return ret;
2400 }
2401
2402 static long vma_needs_reservation(struct hstate *h,
2403                         struct vm_area_struct *vma, unsigned long addr)
2404 {
2405         return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2406 }
2407
2408 static long vma_commit_reservation(struct hstate *h,
2409                         struct vm_area_struct *vma, unsigned long addr)
2410 {
2411         return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2412 }
2413
2414 static void vma_end_reservation(struct hstate *h,
2415                         struct vm_area_struct *vma, unsigned long addr)
2416 {
2417         (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2418 }
2419
2420 static long vma_add_reservation(struct hstate *h,
2421                         struct vm_area_struct *vma, unsigned long addr)
2422 {
2423         return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2424 }
2425
2426 static long vma_del_reservation(struct hstate *h,
2427                         struct vm_area_struct *vma, unsigned long addr)
2428 {
2429         return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2430 }
2431
2432 /*
2433  * This routine is called to restore reservation information on error paths.
2434  * It should ONLY be called for pages allocated via alloc_huge_page(), and
2435  * the hugetlb mutex should remain held when calling this routine.
2436  *
2437  * It handles two specific cases:
2438  * 1) A reservation was in place and the page consumed the reservation.
2439  *    HPageRestoreReserve is set in the page.
2440  * 2) No reservation was in place for the page, so HPageRestoreReserve is
2441  *    not set.  However, alloc_huge_page always updates the reserve map.
2442  *
2443  * In case 1, free_huge_page later in the error path will increment the
2444  * global reserve count.  But, free_huge_page does not have enough context
2445  * to adjust the reservation map.  This case deals primarily with private
2446  * mappings.  Adjust the reserve map here to be consistent with global
2447  * reserve count adjustments to be made by free_huge_page.  Make sure the
2448  * reserve map indicates there is a reservation present.
2449  *
2450  * In case 2, simply undo reserve map modifications done by alloc_huge_page.
2451  */
2452 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2453                         unsigned long address, struct page *page)
2454 {
2455         long rc = vma_needs_reservation(h, vma, address);
2456
2457         if (HPageRestoreReserve(page)) {
2458                 if (unlikely(rc < 0))
2459                         /*
2460                          * Rare out of memory condition in reserve map
2461                          * manipulation.  Clear HPageRestoreReserve so that
2462                          * global reserve count will not be incremented
2463                          * by free_huge_page.  This will make it appear
2464                          * as though the reservation for this page was
2465                          * consumed.  This may prevent the task from
2466                          * faulting in the page at a later time.  This
2467                          * is better than inconsistent global huge page
2468                          * accounting of reserve counts.
2469                          */
2470                         ClearHPageRestoreReserve(page);
2471                 else if (rc)
2472                         (void)vma_add_reservation(h, vma, address);
2473                 else
2474                         vma_end_reservation(h, vma, address);
2475         } else {
2476                 if (!rc) {
2477                         /*
2478                          * This indicates there is an entry in the reserve map
2479                          * added by alloc_huge_page.  We know it was added
2480                          * before the alloc_huge_page call, otherwise
2481                          * HPageRestoreReserve would be set on the page.
2482                          * Remove the entry so that a subsequent allocation
2483                          * does not consume a reservation.
2484                          */
2485                         rc = vma_del_reservation(h, vma, address);
2486                         if (rc < 0)
2487                                 /*
2488                                  * VERY rare out of memory condition.  Since
2489                                  * we can not delete the entry, set
2490                                  * HPageRestoreReserve so that the reserve
2491                                  * count will be incremented when the page
2492                                  * is freed.  This reserve will be consumed
2493                                  * on a subsequent allocation.
2494                                  */
2495                                 SetHPageRestoreReserve(page);
2496                 } else if (rc < 0) {
2497                         /*
2498                          * Rare out of memory condition from
2499                          * vma_needs_reservation call.  Memory allocation is
2500                          * only attempted if a new entry is needed.  Therefore,
2501                          * this implies there is not an entry in the
2502                          * reserve map.
2503                          *
2504                          * For shared mappings, no entry in the map indicates
2505                          * no reservation.  We are done.
2506                          */
2507                         if (!(vma->vm_flags & VM_MAYSHARE))
2508                                 /*
2509                                  * For private mappings, no entry indicates
2510                                  * a reservation is present.  Since we can
2511                                  * not add an entry, set SetHPageRestoreReserve
2512                                  * on the page so reserve count will be
2513                                  * incremented when freed.  This reserve will
2514                                  * be consumed on a subsequent allocation.
2515                                  */
2516                                 SetHPageRestoreReserve(page);
2517                 } else
2518                         /*
2519                          * No reservation present, do nothing
2520                          */
2521                          vma_end_reservation(h, vma, address);
2522         }
2523 }
2524
2525 /*
2526  * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one
2527  * @h: struct hstate old page belongs to
2528  * @old_page: Old page to dissolve
2529  * @list: List to isolate the page in case we need to
2530  * Returns 0 on success, otherwise negated error.
2531  */
2532 static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page,
2533                                         struct list_head *list)
2534 {
2535         gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2536         int nid = page_to_nid(old_page);
2537         struct page *new_page;
2538         int ret = 0;
2539
2540         /*
2541          * Before dissolving the page, we need to allocate a new one for the
2542          * pool to remain stable.  Here, we allocate the page and 'prep' it
2543          * by doing everything but actually updating counters and adding to
2544          * the pool.  This simplifies and let us do most of the processing
2545          * under the lock.
2546          */
2547         new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL);
2548         if (!new_page)
2549                 return -ENOMEM;
2550         __prep_new_huge_page(h, new_page);
2551
2552 retry:
2553         spin_lock_irq(&hugetlb_lock);
2554         if (!PageHuge(old_page)) {
2555                 /*
2556                  * Freed from under us. Drop new_page too.
2557                  */
2558                 goto free_new;
2559         } else if (page_count(old_page)) {
2560                 /*
2561                  * Someone has grabbed the page, try to isolate it here.
2562                  * Fail with -EBUSY if not possible.
2563                  */
2564                 spin_unlock_irq(&hugetlb_lock);
2565                 if (!isolate_huge_page(old_page, list))
2566                         ret = -EBUSY;
2567                 spin_lock_irq(&hugetlb_lock);
2568                 goto free_new;
2569         } else if (!HPageFreed(old_page)) {
2570                 /*
2571                  * Page's refcount is 0 but it has not been enqueued in the
2572                  * freelist yet. Race window is small, so we can succeed here if
2573                  * we retry.
2574                  */
2575                 spin_unlock_irq(&hugetlb_lock);
2576                 cond_resched();
2577                 goto retry;
2578         } else {
2579                 /*
2580                  * Ok, old_page is still a genuine free hugepage. Remove it from
2581                  * the freelist and decrease the counters. These will be
2582                  * incremented again when calling __prep_account_new_huge_page()
2583                  * and enqueue_huge_page() for new_page. The counters will remain
2584                  * stable since this happens under the lock.
2585                  */
2586                 remove_hugetlb_page(h, old_page, false);
2587
2588                 /*
2589                  * Reference count trick is needed because allocator gives us
2590                  * referenced page but the pool requires pages with 0 refcount.
2591                  */
2592                 __prep_account_new_huge_page(h, nid);
2593                 page_ref_dec(new_page);
2594                 enqueue_huge_page(h, new_page);
2595
2596                 /*
2597                  * Pages have been replaced, we can safely free the old one.
2598                  */
2599                 spin_unlock_irq(&hugetlb_lock);
2600                 update_and_free_page(h, old_page, false);
2601         }
2602
2603         return ret;
2604
2605 free_new:
2606         spin_unlock_irq(&hugetlb_lock);
2607         update_and_free_page(h, new_page, false);
2608
2609         return ret;
2610 }
2611
2612 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
2613 {
2614         struct hstate *h;
2615         struct page *head;
2616         int ret = -EBUSY;
2617
2618         /*
2619          * The page might have been dissolved from under our feet, so make sure
2620          * to carefully check the state under the lock.
2621          * Return success when racing as if we dissolved the page ourselves.
2622          */
2623         spin_lock_irq(&hugetlb_lock);
2624         if (PageHuge(page)) {
2625                 head = compound_head(page);
2626                 h = page_hstate(head);
2627         } else {
2628                 spin_unlock_irq(&hugetlb_lock);
2629                 return 0;
2630         }
2631         spin_unlock_irq(&hugetlb_lock);
2632
2633         /*
2634          * Fence off gigantic pages as there is a cyclic dependency between
2635          * alloc_contig_range and them. Return -ENOMEM as this has the effect
2636          * of bailing out right away without further retrying.
2637          */
2638         if (hstate_is_gigantic(h))
2639                 return -ENOMEM;
2640
2641         if (page_count(head) && isolate_huge_page(head, list))
2642                 ret = 0;
2643         else if (!page_count(head))
2644                 ret = alloc_and_dissolve_huge_page(h, head, list);
2645
2646         return ret;
2647 }
2648
2649 struct page *alloc_huge_page(struct vm_area_struct *vma,
2650                                     unsigned long addr, int avoid_reserve)
2651 {
2652         struct hugepage_subpool *spool = subpool_vma(vma);
2653         struct hstate *h = hstate_vma(vma);
2654         struct page *page;
2655         long map_chg, map_commit;
2656         long gbl_chg;
2657         int ret, idx;
2658         struct hugetlb_cgroup *h_cg;
2659         bool deferred_reserve;
2660
2661         idx = hstate_index(h);
2662         /*
2663          * Examine the region/reserve map to determine if the process
2664          * has a reservation for the page to be allocated.  A return
2665          * code of zero indicates a reservation exists (no change).
2666          */
2667         map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2668         if (map_chg < 0)
2669                 return ERR_PTR(-ENOMEM);
2670
2671         /*
2672          * Processes that did not create the mapping will have no
2673          * reserves as indicated by the region/reserve map. Check
2674          * that the allocation will not exceed the subpool limit.
2675          * Allocations for MAP_NORESERVE mappings also need to be
2676          * checked against any subpool limit.
2677          */
2678         if (map_chg || avoid_reserve) {
2679                 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2680                 if (gbl_chg < 0) {
2681                         vma_end_reservation(h, vma, addr);
2682                         return ERR_PTR(-ENOSPC);
2683                 }
2684
2685                 /*
2686                  * Even though there was no reservation in the region/reserve
2687                  * map, there could be reservations associated with the
2688                  * subpool that can be used.  This would be indicated if the
2689                  * return value of hugepage_subpool_get_pages() is zero.
2690                  * However, if avoid_reserve is specified we still avoid even
2691                  * the subpool reservations.
2692                  */
2693                 if (avoid_reserve)
2694                         gbl_chg = 1;
2695         }
2696
2697         /* If this allocation is not consuming a reservation, charge it now.
2698          */
2699         deferred_reserve = map_chg || avoid_reserve;
2700         if (deferred_reserve) {
2701                 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2702                         idx, pages_per_huge_page(h), &h_cg);
2703                 if (ret)
2704                         goto out_subpool_put;
2705         }
2706
2707         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2708         if (ret)
2709                 goto out_uncharge_cgroup_reservation;
2710
2711         spin_lock_irq(&hugetlb_lock);
2712         /*
2713          * glb_chg is passed to indicate whether or not a page must be taken
2714          * from the global free pool (global change).  gbl_chg == 0 indicates
2715          * a reservation exists for the allocation.
2716          */
2717         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2718         if (!page) {
2719                 spin_unlock_irq(&hugetlb_lock);
2720                 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2721                 if (!page)
2722                         goto out_uncharge_cgroup;
2723                 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2724                         SetHPageRestoreReserve(page);
2725                         h->resv_huge_pages--;
2726                 }
2727                 spin_lock_irq(&hugetlb_lock);
2728                 list_add(&page->lru, &h->hugepage_activelist);
2729                 /* Fall through */
2730         }
2731         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2732         /* If allocation is not consuming a reservation, also store the
2733          * hugetlb_cgroup pointer on the page.
2734          */
2735         if (deferred_reserve) {
2736                 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2737                                                   h_cg, page);
2738         }
2739
2740         spin_unlock_irq(&hugetlb_lock);
2741
2742         hugetlb_set_page_subpool(page, spool);
2743
2744         map_commit = vma_commit_reservation(h, vma, addr);
2745         if (unlikely(map_chg > map_commit)) {
2746                 /*
2747                  * The page was added to the reservation map between
2748                  * vma_needs_reservation and vma_commit_reservation.
2749                  * This indicates a race with hugetlb_reserve_pages.
2750                  * Adjust for the subpool count incremented above AND
2751                  * in hugetlb_reserve_pages for the same page.  Also,
2752                  * the reservation count added in hugetlb_reserve_pages
2753                  * no longer applies.
2754                  */
2755                 long rsv_adjust;
2756
2757                 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2758                 hugetlb_acct_memory(h, -rsv_adjust);
2759                 if (deferred_reserve)
2760                         hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2761                                         pages_per_huge_page(h), page);
2762         }
2763         return page;
2764
2765 out_uncharge_cgroup:
2766         hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2767 out_uncharge_cgroup_reservation:
2768         if (deferred_reserve)
2769                 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2770                                                     h_cg);
2771 out_subpool_put:
2772         if (map_chg || avoid_reserve)
2773                 hugepage_subpool_put_pages(spool, 1);
2774         vma_end_reservation(h, vma, addr);
2775         return ERR_PTR(-ENOSPC);
2776 }
2777
2778 int alloc_bootmem_huge_page(struct hstate *h)
2779         __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2780 int __alloc_bootmem_huge_page(struct hstate *h)
2781 {
2782         struct huge_bootmem_page *m;
2783         int nr_nodes, node;
2784
2785         for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2786                 void *addr;
2787
2788                 addr = memblock_alloc_try_nid_raw(
2789                                 huge_page_size(h), huge_page_size(h),
2790                                 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2791                 if (addr) {
2792                         /*
2793                          * Use the beginning of the huge page to store the
2794                          * huge_bootmem_page struct (until gather_bootmem
2795                          * puts them into the mem_map).
2796                          */
2797                         m = addr;
2798                         goto found;
2799                 }
2800         }
2801         return 0;
2802
2803 found:
2804         BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2805         /* Put them into a private list first because mem_map is not up yet */
2806         INIT_LIST_HEAD(&m->list);
2807         list_add(&m->list, &huge_boot_pages);
2808         m->hstate = h;
2809         return 1;
2810 }
2811
2812 /*
2813  * Put bootmem huge pages into the standard lists after mem_map is up.
2814  * Note: This only applies to gigantic (order > MAX_ORDER) pages.
2815  */
2816 static void __init gather_bootmem_prealloc(void)
2817 {
2818         struct huge_bootmem_page *m;
2819
2820         list_for_each_entry(m, &huge_boot_pages, list) {
2821                 struct page *page = virt_to_page(m);
2822                 struct hstate *h = m->hstate;
2823
2824                 VM_BUG_ON(!hstate_is_gigantic(h));
2825                 WARN_ON(page_count(page) != 1);
2826                 if (prep_compound_gigantic_page(page, huge_page_order(h))) {
2827                         WARN_ON(PageReserved(page));
2828                         prep_new_huge_page(h, page, page_to_nid(page));
2829                         put_page(page); /* add to the hugepage allocator */
2830                 } else {
2831                         free_gigantic_page(page, huge_page_order(h));
2832                         pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
2833                 }
2834
2835                 /*
2836                  * We need to restore the 'stolen' pages to totalram_pages
2837                  * in order to fix confusing memory reports from free(1) and
2838                  * other side-effects, like CommitLimit going negative.
2839                  */
2840                 adjust_managed_page_count(page, pages_per_huge_page(h));
2841                 cond_resched();
2842         }
2843 }
2844
2845 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2846 {
2847         unsigned long i;
2848         nodemask_t *node_alloc_noretry;
2849
2850         if (!hstate_is_gigantic(h)) {
2851                 /*
2852                  * Bit mask controlling how hard we retry per-node allocations.
2853                  * Ignore errors as lower level routines can deal with
2854                  * node_alloc_noretry == NULL.  If this kmalloc fails at boot
2855                  * time, we are likely in bigger trouble.
2856                  */
2857                 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2858                                                 GFP_KERNEL);
2859         } else {
2860                 /* allocations done at boot time */
2861                 node_alloc_noretry = NULL;
2862         }
2863
2864         /* bit mask controlling how hard we retry per-node allocations */
2865         if (node_alloc_noretry)
2866                 nodes_clear(*node_alloc_noretry);
2867
2868         for (i = 0; i < h->max_huge_pages; ++i) {
2869                 if (hstate_is_gigantic(h)) {
2870                         if (hugetlb_cma_size) {
2871                                 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2872                                 goto free;
2873                         }
2874                         if (!alloc_bootmem_huge_page(h))
2875                                 break;
2876                 } else if (!alloc_pool_huge_page(h,
2877                                          &node_states[N_MEMORY],
2878                                          node_alloc_noretry))
2879                         break;
2880                 cond_resched();
2881         }
2882         if (i < h->max_huge_pages) {
2883                 char buf[32];
2884
2885                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2886                 pr_warn("HugeTLB: allocating %lu of page size %s failed.  Only allocated %lu hugepages.\n",
2887                         h->max_huge_pages, buf, i);
2888                 h->max_huge_pages = i;
2889         }
2890 free:
2891         kfree(node_alloc_noretry);
2892 }
2893
2894 static void __init hugetlb_init_hstates(void)
2895 {
2896         struct hstate *h;
2897
2898         for_each_hstate(h) {
2899                 if (minimum_order > huge_page_order(h))
2900                         minimum_order = huge_page_order(h);
2901
2902                 /* oversize hugepages were init'ed in early boot */
2903                 if (!hstate_is_gigantic(h))
2904                         hugetlb_hstate_alloc_pages(h);
2905         }
2906         VM_BUG_ON(minimum_order == UINT_MAX);
2907 }
2908
2909 static void __init report_hugepages(void)
2910 {
2911         struct hstate *h;
2912
2913         for_each_hstate(h) {
2914                 char buf[32];
2915
2916                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2917                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2918                         buf, h->free_huge_pages);
2919         }
2920 }
2921
2922 #ifdef CONFIG_HIGHMEM
2923 static void try_to_free_low(struct hstate *h, unsigned long count,
2924                                                 nodemask_t *nodes_allowed)
2925 {
2926         int i;
2927         LIST_HEAD(page_list);
2928
2929         lockdep_assert_held(&hugetlb_lock);
2930         if (hstate_is_gigantic(h))
2931                 return;
2932
2933         /*
2934          * Collect pages to be freed on a list, and free after dropping lock
2935          */
2936         for_each_node_mask(i, *nodes_allowed) {
2937                 struct page *page, *next;
2938                 struct list_head *freel = &h->hugepage_freelists[i];
2939                 list_for_each_entry_safe(page, next, freel, lru) {
2940                         if (count >= h->nr_huge_pages)
2941                                 goto out;
2942                         if (PageHighMem(page))
2943                                 continue;
2944                         remove_hugetlb_page(h, page, false);
2945                         list_add(&page->lru, &page_list);
2946                 }
2947         }
2948
2949 out:
2950         spin_unlock_irq(&hugetlb_lock);
2951         update_and_free_pages_bulk(h, &page_list);
2952         spin_lock_irq(&hugetlb_lock);
2953 }
2954 #else
2955 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2956                                                 nodemask_t *nodes_allowed)
2957 {
2958 }
2959 #endif
2960
2961 /*
2962  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
2963  * balanced by operating on them in a round-robin fashion.
2964  * Returns 1 if an adjustment was made.
2965  */
2966 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2967                                 int delta)
2968 {
2969         int nr_nodes, node;
2970
2971         lockdep_assert_held(&hugetlb_lock);
2972         VM_BUG_ON(delta != -1 && delta != 1);
2973
2974         if (delta < 0) {
2975                 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2976                         if (h->surplus_huge_pages_node[node])
2977                                 goto found;
2978                 }
2979         } else {
2980                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2981                         if (h->surplus_huge_pages_node[node] <
2982                                         h->nr_huge_pages_node[node])
2983                                 goto found;
2984                 }
2985         }
2986         return 0;
2987
2988 found:
2989         h->surplus_huge_pages += delta;
2990         h->surplus_huge_pages_node[node] += delta;
2991         return 1;
2992 }
2993
2994 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2995 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2996                               nodemask_t *nodes_allowed)
2997 {
2998         unsigned long min_count, ret;
2999         struct page *page;
3000         LIST_HEAD(page_list);
3001         NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3002
3003         /*
3004          * Bit mask controlling how hard we retry per-node allocations.
3005          * If we can not allocate the bit mask, do not attempt to allocate
3006          * the requested huge pages.
3007          */
3008         if (node_alloc_noretry)
3009                 nodes_clear(*node_alloc_noretry);
3010         else
3011                 return -ENOMEM;
3012
3013         /*
3014          * resize_lock mutex prevents concurrent adjustments to number of
3015          * pages in hstate via the proc/sysfs interfaces.
3016          */
3017         mutex_lock(&h->resize_lock);
3018         flush_free_hpage_work(h);
3019         spin_lock_irq(&hugetlb_lock);
3020
3021         /*
3022          * Check for a node specific request.
3023          * Changing node specific huge page count may require a corresponding
3024          * change to the global count.  In any case, the passed node mask
3025          * (nodes_allowed) will restrict alloc/free to the specified node.
3026          */
3027         if (nid != NUMA_NO_NODE) {
3028                 unsigned long old_count = count;
3029
3030                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3031                 /*
3032                  * User may have specified a large count value which caused the
3033                  * above calculation to overflow.  In this case, they wanted
3034                  * to allocate as many huge pages as possible.  Set count to
3035                  * largest possible value to align with their intention.
3036                  */
3037                 if (count < old_count)
3038                         count = ULONG_MAX;
3039         }
3040
3041         /*
3042          * Gigantic pages runtime allocation depend on the capability for large
3043          * page range allocation.
3044          * If the system does not provide this feature, return an error when
3045          * the user tries to allocate gigantic pages but let the user free the
3046          * boottime allocated gigantic pages.
3047          */
3048         if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3049                 if (count > persistent_huge_pages(h)) {
3050                         spin_unlock_irq(&hugetlb_lock);
3051                         mutex_unlock(&h->resize_lock);
3052                         NODEMASK_FREE(node_alloc_noretry);
3053                         return -EINVAL;
3054                 }
3055                 /* Fall through to decrease pool */
3056         }
3057
3058         /*
3059          * Increase the pool size
3060          * First take pages out of surplus state.  Then make up the
3061          * remaining difference by allocating fresh huge pages.
3062          *
3063          * We might race with alloc_surplus_huge_page() here and be unable
3064          * to convert a surplus huge page to a normal huge page. That is
3065          * not critical, though, it just means the overall size of the
3066          * pool might be one hugepage larger than it needs to be, but
3067          * within all the constraints specified by the sysctls.
3068          */
3069         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3070                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3071                         break;
3072         }
3073
3074         while (count > persistent_huge_pages(h)) {
3075                 /*
3076                  * If this allocation races such that we no longer need the
3077                  * page, free_huge_page will handle it by freeing the page
3078                  * and reducing the surplus.
3079                  */
3080                 spin_unlock_irq(&hugetlb_lock);
3081
3082                 /* yield cpu to avoid soft lockup */
3083                 cond_resched();
3084
3085                 ret = alloc_pool_huge_page(h, nodes_allowed,
3086                                                 node_alloc_noretry);
3087                 spin_lock_irq(&hugetlb_lock);
3088                 if (!ret)
3089                         goto out;
3090
3091                 /* Bail for signals. Probably ctrl-c from user */
3092                 if (signal_pending(current))
3093                         goto out;
3094         }
3095
3096         /*
3097          * Decrease the pool size
3098          * First return free pages to the buddy allocator (being careful
3099          * to keep enough around to satisfy reservations).  Then place
3100          * pages into surplus state as needed so the pool will shrink
3101          * to the desired size as pages become free.
3102          *
3103          * By placing pages into the surplus state independent of the
3104          * overcommit value, we are allowing the surplus pool size to
3105          * exceed overcommit. There are few sane options here. Since
3106          * alloc_surplus_huge_page() is checking the global counter,
3107          * though, we'll note that we're not allowed to exceed surplus
3108          * and won't grow the pool anywhere else. Not until one of the
3109          * sysctls are changed, or the surplus pages go out of use.
3110          */
3111         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3112         min_count = max(count, min_count);
3113         try_to_free_low(h, min_count, nodes_allowed);
3114
3115         /*
3116          * Collect pages to be removed on list without dropping lock
3117          */
3118         while (min_count < persistent_huge_pages(h)) {
3119                 page = remove_pool_huge_page(h, nodes_allowed, 0);
3120                 if (!page)
3121                         break;
3122
3123                 list_add(&page->lru, &page_list);
3124         }
3125         /* free the pages after dropping lock */
3126         spin_unlock_irq(&hugetlb_lock);
3127         update_and_free_pages_bulk(h, &page_list);
3128         flush_free_hpage_work(h);
3129         spin_lock_irq(&hugetlb_lock);
3130
3131         while (count < persistent_huge_pages(h)) {
3132                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3133                         break;
3134         }
3135 out:
3136         h->max_huge_pages = persistent_huge_pages(h);
3137         spin_unlock_irq(&hugetlb_lock);
3138         mutex_unlock(&h->resize_lock);
3139
3140         NODEMASK_FREE(node_alloc_noretry);
3141
3142         return 0;
3143 }
3144
3145 #define HSTATE_ATTR_RO(_name) \
3146         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3147
3148 #define HSTATE_ATTR(_name) \
3149         static struct kobj_attribute _name##_attr = \
3150                 __ATTR(_name, 0644, _name##_show, _name##_store)
3151
3152 static struct kobject *hugepages_kobj;
3153 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3154
3155 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3156
3157 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3158 {
3159         int i;
3160
3161         for (i = 0; i < HUGE_MAX_HSTATE; i++)
3162                 if (hstate_kobjs[i] == kobj) {
3163                         if (nidp)
3164                                 *nidp = NUMA_NO_NODE;
3165                         return &hstates[i];
3166                 }
3167
3168         return kobj_to_node_hstate(kobj, nidp);
3169 }
3170
3171 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3172                                         struct kobj_attribute *attr, char *buf)
3173 {
3174         struct hstate *h;
3175         unsigned long nr_huge_pages;
3176         int nid;
3177
3178         h = kobj_to_hstate(kobj, &nid);
3179         if (nid == NUMA_NO_NODE)
3180                 nr_huge_pages = h->nr_huge_pages;
3181         else
3182                 nr_huge_pages = h->nr_huge_pages_node[nid];
3183
3184         return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3185 }
3186
3187 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3188                                            struct hstate *h, int nid,
3189                                            unsigned long count, size_t len)
3190 {
3191         int err;
3192         nodemask_t nodes_allowed, *n_mask;
3193
3194         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3195                 return -EINVAL;
3196
3197         if (nid == NUMA_NO_NODE) {
3198                 /*
3199                  * global hstate attribute
3200                  */
3201                 if (!(obey_mempolicy &&
3202                                 init_nodemask_of_mempolicy(&nodes_allowed)))
3203                         n_mask = &node_states[N_MEMORY];
3204                 else
3205                         n_mask = &nodes_allowed;
3206         } else {
3207                 /*
3208                  * Node specific request.  count adjustment happens in
3209                  * set_max_huge_pages() after acquiring hugetlb_lock.
3210                  */
3211                 init_nodemask_of_node(&nodes_allowed, nid);
3212                 n_mask = &nodes_allowed;
3213         }
3214
3215         err = set_max_huge_pages(h, count, nid, n_mask);
3216
3217         return err ? err : len;
3218 }
3219
3220 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3221                                          struct kobject *kobj, const char *buf,
3222                                          size_t len)
3223 {
3224         struct hstate *h;
3225         unsigned long count;
3226         int nid;
3227         int err;
3228
3229         err = kstrtoul(buf, 10, &count);
3230         if (err)
3231                 return err;
3232
3233         h = kobj_to_hstate(kobj, &nid);
3234         return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3235 }
3236
3237 static ssize_t nr_hugepages_show(struct kobject *kobj,
3238                                        struct kobj_attribute *attr, char *buf)
3239 {
3240         return nr_hugepages_show_common(kobj, attr, buf);
3241 }
3242
3243 static ssize_t nr_hugepages_store(struct kobject *kobj,
3244                struct kobj_attribute *attr, const char *buf, size_t len)
3245 {
3246         return nr_hugepages_store_common(false, kobj, buf, len);
3247 }
3248 HSTATE_ATTR(nr_hugepages);
3249
3250 #ifdef CONFIG_NUMA
3251
3252 /*
3253  * hstate attribute for optionally mempolicy-based constraint on persistent
3254  * huge page alloc/free.
3255  */
3256 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3257                                            struct kobj_attribute *attr,
3258                                            char *buf)
3259 {
3260         return nr_hugepages_show_common(kobj, attr, buf);
3261 }
3262
3263 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3264                struct kobj_attribute *attr, const char *buf, size_t len)
3265 {
3266         return nr_hugepages_store_common(true, kobj, buf, len);
3267 }
3268 HSTATE_ATTR(nr_hugepages_mempolicy);
3269 #endif
3270
3271
3272 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3273                                         struct kobj_attribute *attr, char *buf)
3274 {
3275         struct hstate *h = kobj_to_hstate(kobj, NULL);
3276         return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3277 }
3278
3279 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3280                 struct kobj_attribute *attr, const char *buf, size_t count)
3281 {
3282         int err;
3283         unsigned long input;
3284         struct hstate *h = kobj_to_hstate(kobj, NULL);
3285
3286         if (hstate_is_gigantic(h))
3287                 return -EINVAL;
3288
3289         err = kstrtoul(buf, 10, &input);
3290         if (err)
3291                 return err;
3292
3293         spin_lock_irq(&hugetlb_lock);
3294         h->nr_overcommit_huge_pages = input;
3295         spin_unlock_irq(&hugetlb_lock);
3296
3297         return count;
3298 }
3299 HSTATE_ATTR(nr_overcommit_hugepages);
3300
3301 static ssize_t free_hugepages_show(struct kobject *kobj,
3302                                         struct kobj_attribute *attr, char *buf)
3303 {
3304         struct hstate *h;
3305         unsigned long free_huge_pages;
3306         int nid;
3307
3308         h = kobj_to_hstate(kobj, &nid);
3309         if (nid == NUMA_NO_NODE)
3310                 free_huge_pages = h->free_huge_pages;
3311         else
3312                 free_huge_pages = h->free_huge_pages_node[nid];
3313
3314         return sysfs_emit(buf, "%lu\n", free_huge_pages);
3315 }
3316 HSTATE_ATTR_RO(free_hugepages);
3317
3318 static ssize_t resv_hugepages_show(struct kobject *kobj,
3319                                         struct kobj_attribute *attr, char *buf)
3320 {
3321         struct hstate *h = kobj_to_hstate(kobj, NULL);
3322         return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3323 }
3324 HSTATE_ATTR_RO(resv_hugepages);
3325
3326 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3327                                         struct kobj_attribute *attr, char *buf)
3328 {
3329         struct hstate *h;
3330         unsigned long surplus_huge_pages;
3331         int nid;
3332
3333         h = kobj_to_hstate(kobj, &nid);
3334         if (nid == NUMA_NO_NODE)
3335                 surplus_huge_pages = h->surplus_huge_pages;
3336         else
3337                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3338
3339         return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3340 }
3341 HSTATE_ATTR_RO(surplus_hugepages);
3342
3343 static struct attribute *hstate_attrs[] = {
3344         &nr_hugepages_attr.attr,
3345         &nr_overcommit_hugepages_attr.attr,
3346         &free_hugepages_attr.attr,
3347         &resv_hugepages_attr.attr,
3348         &surplus_hugepages_attr.attr,
3349 #ifdef CONFIG_NUMA
3350         &nr_hugepages_mempolicy_attr.attr,
3351 #endif
3352         NULL,
3353 };
3354
3355 static const struct attribute_group hstate_attr_group = {
3356         .attrs = hstate_attrs,
3357 };
3358
3359 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3360                                     struct kobject **hstate_kobjs,
3361                                     const struct attribute_group *hstate_attr_group)
3362 {
3363         int retval;
3364         int hi = hstate_index(h);
3365
3366         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3367         if (!hstate_kobjs[hi])
3368                 return -ENOMEM;
3369
3370         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3371         if (retval) {
3372                 kobject_put(hstate_kobjs[hi]);
3373                 hstate_kobjs[hi] = NULL;
3374         }
3375
3376         return retval;
3377 }
3378
3379 static void __init hugetlb_sysfs_init(void)
3380 {
3381         struct hstate *h;
3382         int err;
3383
3384         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3385         if (!hugepages_kobj)
3386                 return;
3387
3388         for_each_hstate(h) {
3389                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3390                                          hstate_kobjs, &hstate_attr_group);
3391                 if (err)
3392                         pr_err("HugeTLB: Unable to add hstate %s", h->name);
3393         }
3394 }
3395
3396 #ifdef CONFIG_NUMA
3397
3398 /*
3399  * node_hstate/s - associate per node hstate attributes, via their kobjects,
3400  * with node devices in node_devices[] using a parallel array.  The array
3401  * index of a node device or _hstate == node id.
3402  * This is here to avoid any static dependency of the node device driver, in
3403  * the base kernel, on the hugetlb module.
3404  */
3405 struct node_hstate {
3406         struct kobject          *hugepages_kobj;
3407         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
3408 };
3409 static struct node_hstate node_hstates[MAX_NUMNODES];
3410
3411 /*
3412  * A subset of global hstate attributes for node devices
3413  */
3414 static struct attribute *per_node_hstate_attrs[] = {
3415         &nr_hugepages_attr.attr,
3416         &free_hugepages_attr.attr,
3417         &surplus_hugepages_attr.attr,
3418         NULL,
3419 };
3420
3421 static const struct attribute_group per_node_hstate_attr_group = {
3422         .attrs = per_node_hstate_attrs,
3423 };
3424
3425 /*
3426  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3427  * Returns node id via non-NULL nidp.
3428  */
3429 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3430 {
3431         int nid;
3432
3433         for (nid = 0; nid < nr_node_ids; nid++) {
3434                 struct node_hstate *nhs = &node_hstates[nid];
3435                 int i;
3436                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3437                         if (nhs->hstate_kobjs[i] == kobj) {
3438                                 if (nidp)
3439                                         *nidp = nid;
3440                                 return &hstates[i];
3441                         }
3442         }
3443
3444         BUG();
3445         return NULL;
3446 }
3447
3448 /*
3449  * Unregister hstate attributes from a single node device.
3450  * No-op if no hstate attributes attached.
3451  */
3452 static void hugetlb_unregister_node(struct node *node)
3453 {
3454         struct hstate *h;
3455         struct node_hstate *nhs = &node_hstates[node->dev.id];
3456
3457         if (!nhs->hugepages_kobj)
3458                 return;         /* no hstate attributes */
3459
3460         for_each_hstate(h) {
3461                 int idx = hstate_index(h);
3462                 if (nhs->hstate_kobjs[idx]) {
3463                         kobject_put(nhs->hstate_kobjs[idx]);
3464                         nhs->hstate_kobjs[idx] = NULL;
3465                 }
3466         }
3467
3468         kobject_put(nhs->hugepages_kobj);
3469         nhs->hugepages_kobj = NULL;
3470 }
3471
3472
3473 /*
3474  * Register hstate attributes for a single node device.
3475  * No-op if attributes already registered.
3476  */
3477 static void hugetlb_register_node(struct node *node)
3478 {
3479         struct hstate *h;
3480         struct node_hstate *nhs = &node_hstates[node->dev.id];
3481         int err;
3482
3483         if (nhs->hugepages_kobj)
3484                 return;         /* already allocated */
3485
3486         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3487                                                         &node->dev.kobj);
3488         if (!nhs->hugepages_kobj)
3489                 return;
3490
3491         for_each_hstate(h) {
3492                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3493                                                 nhs->hstate_kobjs,
3494                                                 &per_node_hstate_attr_group);
3495                 if (err) {
3496                         pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3497                                 h->name, node->dev.id);
3498                         hugetlb_unregister_node(node);
3499                         break;
3500                 }
3501         }
3502 }
3503
3504 /*
3505  * hugetlb init time:  register hstate attributes for all registered node
3506  * devices of nodes that have memory.  All on-line nodes should have
3507  * registered their associated device by this time.
3508  */
3509 static void __init hugetlb_register_all_nodes(void)
3510 {
3511         int nid;
3512
3513         for_each_node_state(nid, N_MEMORY) {
3514                 struct node *node = node_devices[nid];
3515                 if (node->dev.id == nid)
3516                         hugetlb_register_node(node);
3517         }
3518
3519         /*
3520          * Let the node device driver know we're here so it can
3521          * [un]register hstate attributes on node hotplug.
3522          */
3523         register_hugetlbfs_with_node(hugetlb_register_node,
3524                                      hugetlb_unregister_node);
3525 }
3526 #else   /* !CONFIG_NUMA */
3527
3528 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3529 {
3530         BUG();
3531         if (nidp)
3532                 *nidp = -1;
3533         return NULL;
3534 }
3535
3536 static void hugetlb_register_all_nodes(void) { }
3537
3538 #endif
3539
3540 static int __init hugetlb_init(void)
3541 {
3542         int i;
3543
3544         BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
3545                         __NR_HPAGEFLAGS);
3546
3547         if (!hugepages_supported()) {
3548                 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3549                         pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3550                 return 0;
3551         }
3552
3553         /*
3554          * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists.  Some
3555          * architectures depend on setup being done here.
3556          */
3557         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3558         if (!parsed_default_hugepagesz) {
3559                 /*
3560                  * If we did not parse a default huge page size, set
3561                  * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3562                  * number of huge pages for this default size was implicitly
3563                  * specified, set that here as well.
3564                  * Note that the implicit setting will overwrite an explicit
3565                  * setting.  A warning will be printed in this case.
3566                  */
3567                 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3568                 if (default_hstate_max_huge_pages) {
3569                         if (default_hstate.max_huge_pages) {
3570                                 char buf[32];
3571
3572                                 string_get_size(huge_page_size(&default_hstate),
3573                                         1, STRING_UNITS_2, buf, 32);
3574                                 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3575                                         default_hstate.max_huge_pages, buf);
3576                                 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3577                                         default_hstate_max_huge_pages);
3578                         }
3579                         default_hstate.max_huge_pages =
3580                                 default_hstate_max_huge_pages;
3581                 }
3582         }
3583
3584         hugetlb_cma_check();
3585         hugetlb_init_hstates();
3586         gather_bootmem_prealloc();
3587         report_hugepages();
3588
3589         hugetlb_sysfs_init();
3590         hugetlb_register_all_nodes();
3591         hugetlb_cgroup_file_init();
3592
3593 #ifdef CONFIG_SMP
3594         num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3595 #else
3596         num_fault_mutexes = 1;
3597 #endif
3598         hugetlb_fault_mutex_table =
3599                 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3600                               GFP_KERNEL);
3601         BUG_ON(!hugetlb_fault_mutex_table);
3602
3603         for (i = 0; i < num_fault_mutexes; i++)
3604                 mutex_init(&hugetlb_fault_mutex_table[i]);
3605         return 0;
3606 }
3607 subsys_initcall(hugetlb_init);
3608
3609 /* Overwritten by architectures with more huge page sizes */
3610 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3611 {
3612         return size == HPAGE_SIZE;
3613 }
3614
3615 void __init hugetlb_add_hstate(unsigned int order)
3616 {
3617         struct hstate *h;
3618         unsigned long i;
3619
3620         if (size_to_hstate(PAGE_SIZE << order)) {
3621                 return;
3622         }
3623         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3624         BUG_ON(order == 0);
3625         h = &hstates[hugetlb_max_hstate++];
3626         mutex_init(&h->resize_lock);
3627         h->order = order;
3628         h->mask = ~(huge_page_size(h) - 1);
3629         for (i = 0; i < MAX_NUMNODES; ++i)
3630                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3631         INIT_LIST_HEAD(&h->hugepage_activelist);
3632         h->next_nid_to_alloc = first_memory_node;
3633         h->next_nid_to_free = first_memory_node;
3634         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3635                                         huge_page_size(h)/1024);
3636         hugetlb_vmemmap_init(h);
3637
3638         parsed_hstate = h;
3639 }
3640
3641 /*
3642  * hugepages command line processing
3643  * hugepages normally follows a valid hugepagsz or default_hugepagsz
3644  * specification.  If not, ignore the hugepages value.  hugepages can also
3645  * be the first huge page command line  option in which case it implicitly
3646  * specifies the number of huge pages for the default size.
3647  */
3648 static int __init hugepages_setup(char *s)
3649 {
3650         unsigned long *mhp;
3651         static unsigned long *last_mhp;
3652
3653         if (!parsed_valid_hugepagesz) {
3654                 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3655                 parsed_valid_hugepagesz = true;
3656                 return 0;
3657         }
3658
3659         /*
3660          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3661          * yet, so this hugepages= parameter goes to the "default hstate".
3662          * Otherwise, it goes with the previously parsed hugepagesz or
3663          * default_hugepagesz.
3664          */
3665         else if (!hugetlb_max_hstate)
3666                 mhp = &default_hstate_max_huge_pages;
3667         else
3668                 mhp = &parsed_hstate->max_huge_pages;
3669
3670         if (mhp == last_mhp) {
3671                 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3672                 return 0;
3673         }
3674
3675         if (sscanf(s, "%lu", mhp) <= 0)
3676                 *mhp = 0;
3677
3678         /*
3679          * Global state is always initialized later in hugetlb_init.
3680          * But we need to allocate gigantic hstates here early to still
3681          * use the bootmem allocator.
3682          */
3683         if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
3684                 hugetlb_hstate_alloc_pages(parsed_hstate);
3685
3686         last_mhp = mhp;
3687
3688         return 1;
3689 }
3690 __setup("hugepages=", hugepages_setup);
3691
3692 /*
3693  * hugepagesz command line processing
3694  * A specific huge page size can only be specified once with hugepagesz.
3695  * hugepagesz is followed by hugepages on the command line.  The global
3696  * variable 'parsed_valid_hugepagesz' is used to determine if prior
3697  * hugepagesz argument was valid.
3698  */
3699 static int __init hugepagesz_setup(char *s)
3700 {
3701         unsigned long size;
3702         struct hstate *h;
3703
3704         parsed_valid_hugepagesz = false;
3705         size = (unsigned long)memparse(s, NULL);
3706
3707         if (!arch_hugetlb_valid_size(size)) {
3708                 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3709                 return 0;
3710         }
3711
3712         h = size_to_hstate(size);
3713         if (h) {
3714                 /*
3715                  * hstate for this size already exists.  This is normally
3716                  * an error, but is allowed if the existing hstate is the
3717                  * default hstate.  More specifically, it is only allowed if
3718                  * the number of huge pages for the default hstate was not
3719                  * previously specified.
3720                  */
3721                 if (!parsed_default_hugepagesz ||  h != &default_hstate ||
3722                     default_hstate.max_huge_pages) {
3723                         pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3724                         return 0;
3725                 }
3726
3727                 /*
3728                  * No need to call hugetlb_add_hstate() as hstate already
3729                  * exists.  But, do set parsed_hstate so that a following
3730                  * hugepages= parameter will be applied to this hstate.
3731                  */
3732                 parsed_hstate = h;
3733                 parsed_valid_hugepagesz = true;
3734                 return 1;
3735         }
3736
3737         hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3738         parsed_valid_hugepagesz = true;
3739         return 1;
3740 }
3741 __setup("hugepagesz=", hugepagesz_setup);
3742
3743 /*
3744  * default_hugepagesz command line input
3745  * Only one instance of default_hugepagesz allowed on command line.
3746  */
3747 static int __init default_hugepagesz_setup(char *s)
3748 {
3749         unsigned long size;
3750
3751         parsed_valid_hugepagesz = false;
3752         if (parsed_default_hugepagesz) {
3753                 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3754                 return 0;
3755         }
3756
3757         size = (unsigned long)memparse(s, NULL);
3758
3759         if (!arch_hugetlb_valid_size(size)) {
3760                 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3761                 return 0;
3762         }
3763
3764         hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3765         parsed_valid_hugepagesz = true;
3766         parsed_default_hugepagesz = true;
3767         default_hstate_idx = hstate_index(size_to_hstate(size));
3768
3769         /*
3770          * The number of default huge pages (for this size) could have been
3771          * specified as the first hugetlb parameter: hugepages=X.  If so,
3772          * then default_hstate_max_huge_pages is set.  If the default huge
3773          * page size is gigantic (>= MAX_ORDER), then the pages must be
3774          * allocated here from bootmem allocator.
3775          */
3776         if (default_hstate_max_huge_pages) {
3777                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3778                 if (hstate_is_gigantic(&default_hstate))
3779                         hugetlb_hstate_alloc_pages(&default_hstate);
3780                 default_hstate_max_huge_pages = 0;
3781         }
3782
3783         return 1;
3784 }
3785 __setup("default_hugepagesz=", default_hugepagesz_setup);
3786
3787 static unsigned int allowed_mems_nr(struct hstate *h)
3788 {
3789         int node;
3790         unsigned int nr = 0;
3791         nodemask_t *mpol_allowed;
3792         unsigned int *array = h->free_huge_pages_node;
3793         gfp_t gfp_mask = htlb_alloc_mask(h);
3794
3795         mpol_allowed = policy_nodemask_current(gfp_mask);
3796
3797         for_each_node_mask(node, cpuset_current_mems_allowed) {
3798                 if (!mpol_allowed || node_isset(node, *mpol_allowed))
3799                         nr += array[node];
3800         }
3801
3802         return nr;
3803 }
3804
3805 #ifdef CONFIG_SYSCTL
3806 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3807                                           void *buffer, size_t *length,
3808                                           loff_t *ppos, unsigned long *out)
3809 {
3810         struct ctl_table dup_table;
3811
3812         /*
3813          * In order to avoid races with __do_proc_doulongvec_minmax(), we
3814          * can duplicate the @table and alter the duplicate of it.
3815          */
3816         dup_table = *table;
3817         dup_table.data = out;
3818
3819         return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3820 }
3821
3822 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3823                          struct ctl_table *table, int write,
3824                          void *buffer, size_t *length, loff_t *ppos)
3825 {
3826         struct hstate *h = &default_hstate;
3827         unsigned long tmp = h->max_huge_pages;
3828         int ret;
3829
3830         if (!hugepages_supported())
3831                 return -EOPNOTSUPP;
3832
3833         ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3834                                              &tmp);
3835         if (ret)
3836                 goto out;
3837
3838         if (write)
3839                 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3840                                                   NUMA_NO_NODE, tmp, *length);
3841 out:
3842         return ret;
3843 }
3844
3845 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3846                           void *buffer, size_t *length, loff_t *ppos)
3847 {
3848
3849         return hugetlb_sysctl_handler_common(false, table, write,
3850                                                         buffer, length, ppos);
3851 }
3852
3853 #ifdef CONFIG_NUMA
3854 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3855                           void *buffer, size_t *length, loff_t *ppos)
3856 {
3857         return hugetlb_sysctl_handler_common(true, table, write,
3858                                                         buffer, length, ppos);
3859 }
3860 #endif /* CONFIG_NUMA */
3861
3862 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3863                 void *buffer, size_t *length, loff_t *ppos)
3864 {
3865         struct hstate *h = &default_hstate;
3866         unsigned long tmp;
3867         int ret;
3868
3869         if (!hugepages_supported())
3870                 return -EOPNOTSUPP;
3871
3872         tmp = h->nr_overcommit_huge_pages;
3873
3874         if (write && hstate_is_gigantic(h))
3875                 return -EINVAL;
3876
3877         ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3878                                              &tmp);
3879         if (ret)
3880                 goto out;
3881
3882         if (write) {
3883                 spin_lock_irq(&hugetlb_lock);
3884                 h->nr_overcommit_huge_pages = tmp;
3885                 spin_unlock_irq(&hugetlb_lock);
3886         }
3887 out:
3888         return ret;
3889 }
3890
3891 #endif /* CONFIG_SYSCTL */
3892
3893 void hugetlb_report_meminfo(struct seq_file *m)
3894 {
3895         struct hstate *h;
3896         unsigned long total = 0;
3897
3898         if (!hugepages_supported())
3899                 return;
3900
3901         for_each_hstate(h) {
3902                 unsigned long count = h->nr_huge_pages;
3903
3904                 total += huge_page_size(h) * count;
3905
3906                 if (h == &default_hstate)
3907                         seq_printf(m,
3908                                    "HugePages_Total:   %5lu\n"
3909                                    "HugePages_Free:    %5lu\n"
3910                                    "HugePages_Rsvd:    %5lu\n"
3911                                    "HugePages_Surp:    %5lu\n"
3912                                    "Hugepagesize:   %8lu kB\n",
3913                                    count,
3914                                    h->free_huge_pages,
3915                                    h->resv_huge_pages,
3916                                    h->surplus_huge_pages,
3917                                    huge_page_size(h) / SZ_1K);
3918         }
3919
3920         seq_printf(m, "Hugetlb:        %8lu kB\n", total / SZ_1K);
3921 }
3922
3923 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
3924 {
3925         struct hstate *h = &default_hstate;
3926
3927         if (!hugepages_supported())
3928                 return 0;
3929
3930         return sysfs_emit_at(buf, len,
3931                              "Node %d HugePages_Total: %5u\n"
3932                              "Node %d HugePages_Free:  %5u\n"
3933                              "Node %d HugePages_Surp:  %5u\n",
3934                              nid, h->nr_huge_pages_node[nid],
3935                              nid, h->free_huge_pages_node[nid],
3936                              nid, h->surplus_huge_pages_node[nid]);
3937 }
3938
3939 void hugetlb_show_meminfo(void)
3940 {
3941         struct hstate *h;
3942         int nid;
3943
3944         if (!hugepages_supported())
3945                 return;
3946
3947         for_each_node_state(nid, N_MEMORY)
3948                 for_each_hstate(h)
3949                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3950                                 nid,
3951                                 h->nr_huge_pages_node[nid],
3952                                 h->free_huge_pages_node[nid],
3953                                 h->surplus_huge_pages_node[nid],
3954                                 huge_page_size(h) / SZ_1K);
3955 }
3956
3957 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3958 {
3959         seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3960                    atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3961 }
3962
3963 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3964 unsigned long hugetlb_total_pages(void)
3965 {
3966         struct hstate *h;
3967         unsigned long nr_total_pages = 0;
3968
3969         for_each_hstate(h)
3970                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3971         return nr_total_pages;
3972 }
3973
3974 static int hugetlb_acct_memory(struct hstate *h, long delta)
3975 {
3976         int ret = -ENOMEM;
3977
3978         if (!delta)
3979                 return 0;
3980
3981         spin_lock_irq(&hugetlb_lock);
3982         /*
3983          * When cpuset is configured, it breaks the strict hugetlb page
3984          * reservation as the accounting is done on a global variable. Such
3985          * reservation is completely rubbish in the presence of cpuset because
3986          * the reservation is not checked against page availability for the
3987          * current cpuset. Application can still potentially OOM'ed by kernel
3988          * with lack of free htlb page in cpuset that the task is in.
3989          * Attempt to enforce strict accounting with cpuset is almost
3990          * impossible (or too ugly) because cpuset is too fluid that
3991          * task or memory node can be dynamically moved between cpusets.
3992          *
3993          * The change of semantics for shared hugetlb mapping with cpuset is
3994          * undesirable. However, in order to preserve some of the semantics,
3995          * we fall back to check against current free page availability as
3996          * a best attempt and hopefully to minimize the impact of changing
3997          * semantics that cpuset has.
3998          *
3999          * Apart from cpuset, we also have memory policy mechanism that
4000          * also determines from which node the kernel will allocate memory
4001          * in a NUMA system. So similar to cpuset, we also should consider
4002          * the memory policy of the current task. Similar to the description
4003          * above.
4004          */
4005         if (delta > 0) {
4006                 if (gather_surplus_pages(h, delta) < 0)
4007                         goto out;
4008
4009                 if (delta > allowed_mems_nr(h)) {
4010                         return_unused_surplus_pages(h, delta);
4011                         goto out;
4012                 }
4013         }
4014
4015         ret = 0;
4016         if (delta < 0)
4017                 return_unused_surplus_pages(h, (unsigned long) -delta);
4018
4019 out:
4020         spin_unlock_irq(&hugetlb_lock);
4021         return ret;
4022 }
4023
4024 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4025 {
4026         struct resv_map *resv = vma_resv_map(vma);
4027
4028         /*
4029          * This new VMA should share its siblings reservation map if present.
4030          * The VMA will only ever have a valid reservation map pointer where
4031          * it is being copied for another still existing VMA.  As that VMA
4032          * has a reference to the reservation map it cannot disappear until
4033          * after this open call completes.  It is therefore safe to take a
4034          * new reference here without additional locking.
4035          */
4036         if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4037                 kref_get(&resv->refs);
4038 }
4039
4040 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4041 {
4042         struct hstate *h = hstate_vma(vma);
4043         struct resv_map *resv = vma_resv_map(vma);
4044         struct hugepage_subpool *spool = subpool_vma(vma);
4045         unsigned long reserve, start, end;
4046         long gbl_reserve;
4047
4048         if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4049                 return;
4050
4051         start = vma_hugecache_offset(h, vma, vma->vm_start);
4052         end = vma_hugecache_offset(h, vma, vma->vm_end);
4053
4054         reserve = (end - start) - region_count(resv, start, end);
4055         hugetlb_cgroup_uncharge_counter(resv, start, end);
4056         if (reserve) {
4057                 /*
4058                  * Decrement reserve counts.  The global reserve count may be
4059                  * adjusted if the subpool has a minimum size.
4060                  */
4061                 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4062                 hugetlb_acct_memory(h, -gbl_reserve);
4063         }
4064
4065         kref_put(&resv->refs, resv_map_release);
4066 }
4067
4068 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4069 {
4070         if (addr & ~(huge_page_mask(hstate_vma(vma))))
4071                 return -EINVAL;
4072         return 0;
4073 }
4074
4075 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4076 {
4077         return huge_page_size(hstate_vma(vma));
4078 }
4079
4080 /*
4081  * We cannot handle pagefaults against hugetlb pages at all.  They cause
4082  * handle_mm_fault() to try to instantiate regular-sized pages in the
4083  * hugepage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
4084  * this far.
4085  */
4086 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4087 {
4088         BUG();
4089         return 0;
4090 }
4091
4092 /*
4093  * When a new function is introduced to vm_operations_struct and added
4094  * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4095  * This is because under System V memory model, mappings created via
4096  * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4097  * their original vm_ops are overwritten with shm_vm_ops.
4098  */
4099 const struct vm_operations_struct hugetlb_vm_ops = {
4100         .fault = hugetlb_vm_op_fault,
4101         .open = hugetlb_vm_op_open,
4102         .close = hugetlb_vm_op_close,
4103         .may_split = hugetlb_vm_op_split,
4104         .pagesize = hugetlb_vm_op_pagesize,
4105 };
4106
4107 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4108                                 int writable)
4109 {
4110         pte_t entry;
4111         unsigned int shift = huge_page_shift(hstate_vma(vma));
4112
4113         if (writable) {
4114                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4115                                          vma->vm_page_prot)));
4116         } else {
4117                 entry = huge_pte_wrprotect(mk_huge_pte(page,
4118                                            vma->vm_page_prot));
4119         }
4120         entry = pte_mkyoung(entry);
4121         entry = pte_mkhuge(entry);
4122         entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4123
4124         return entry;
4125 }
4126
4127 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4128                                    unsigned long address, pte_t *ptep)
4129 {
4130         pte_t entry;
4131
4132         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4133         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4134                 update_mmu_cache(vma, address, ptep);
4135 }
4136
4137 bool is_hugetlb_entry_migration(pte_t pte)
4138 {
4139         swp_entry_t swp;
4140
4141         if (huge_pte_none(pte) || pte_present(pte))
4142                 return false;
4143         swp = pte_to_swp_entry(pte);
4144         if (is_migration_entry(swp))
4145                 return true;
4146         else
4147                 return false;
4148 }
4149
4150 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4151 {
4152         swp_entry_t swp;
4153
4154         if (huge_pte_none(pte) || pte_present(pte))
4155                 return false;
4156         swp = pte_to_swp_entry(pte);
4157         if (is_hwpoison_entry(swp))
4158                 return true;
4159         else
4160                 return false;
4161 }
4162
4163 static void
4164 hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
4165                      struct page *new_page)
4166 {
4167         __SetPageUptodate(new_page);
4168         set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
4169         hugepage_add_new_anon_rmap(new_page, vma, addr);
4170         hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
4171         ClearHPageRestoreReserve(new_page);
4172         SetHPageMigratable(new_page);
4173 }
4174
4175 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
4176                             struct vm_area_struct *vma)
4177 {
4178         pte_t *src_pte, *dst_pte, entry, dst_entry;
4179         struct page *ptepage;
4180         unsigned long addr;
4181         bool cow = is_cow_mapping(vma->vm_flags);
4182         struct hstate *h = hstate_vma(vma);
4183         unsigned long sz = huge_page_size(h);
4184         unsigned long npages = pages_per_huge_page(h);
4185         struct address_space *mapping = vma->vm_file->f_mapping;
4186         struct mmu_notifier_range range;
4187         int ret = 0;
4188
4189         if (cow) {
4190                 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
4191                                         vma->vm_start,
4192                                         vma->vm_end);
4193                 mmu_notifier_invalidate_range_start(&range);
4194         } else {
4195                 /*
4196                  * For shared mappings i_mmap_rwsem must be held to call
4197                  * huge_pte_alloc, otherwise the returned ptep could go
4198                  * away if part of a shared pmd and another thread calls
4199                  * huge_pmd_unshare.
4200                  */
4201                 i_mmap_lock_read(mapping);
4202         }
4203
4204         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
4205                 spinlock_t *src_ptl, *dst_ptl;
4206                 src_pte = huge_pte_offset(src, addr, sz);
4207                 if (!src_pte)
4208                         continue;
4209                 dst_pte = huge_pte_alloc(dst, vma, addr, sz);
4210                 if (!dst_pte) {
4211                         ret = -ENOMEM;
4212                         break;
4213                 }
4214
4215                 /*
4216                  * If the pagetables are shared don't copy or take references.
4217                  * dst_pte == src_pte is the common case of src/dest sharing.
4218                  *
4219                  * However, src could have 'unshared' and dst shares with
4220                  * another vma.  If dst_pte !none, this implies sharing.
4221                  * Check here before taking page table lock, and once again
4222                  * after taking the lock below.
4223                  */
4224                 dst_entry = huge_ptep_get(dst_pte);
4225                 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
4226                         continue;
4227
4228                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4229                 src_ptl = huge_pte_lockptr(h, src, src_pte);
4230                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4231                 entry = huge_ptep_get(src_pte);
4232                 dst_entry = huge_ptep_get(dst_pte);
4233 again:
4234                 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
4235                         /*
4236                          * Skip if src entry none.  Also, skip in the
4237                          * unlikely case dst entry !none as this implies
4238                          * sharing with another vma.
4239                          */
4240                         ;
4241                 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
4242                                     is_hugetlb_entry_hwpoisoned(entry))) {
4243                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
4244
4245                         if (is_writable_migration_entry(swp_entry) && cow) {
4246                                 /*
4247                                  * COW mappings require pages in both
4248                                  * parent and child to be set to read.
4249                                  */
4250                                 swp_entry = make_readable_migration_entry(
4251                                                         swp_offset(swp_entry));
4252                                 entry = swp_entry_to_pte(swp_entry);
4253                                 set_huge_swap_pte_at(src, addr, src_pte,
4254                                                      entry, sz);
4255                         }
4256                         set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
4257                 } else {
4258                         entry = huge_ptep_get(src_pte);
4259                         ptepage = pte_page(entry);
4260                         get_page(ptepage);
4261
4262                         /*
4263                          * This is a rare case where we see pinned hugetlb
4264                          * pages while they're prone to COW.  We need to do the
4265                          * COW earlier during fork.
4266                          *
4267                          * When pre-allocating the page or copying data, we
4268                          * need to be without the pgtable locks since we could
4269                          * sleep during the process.
4270                          */
4271                         if (unlikely(page_needs_cow_for_dma(vma, ptepage))) {
4272                                 pte_t src_pte_old = entry;
4273                                 struct page *new;
4274
4275                                 spin_unlock(src_ptl);
4276                                 spin_unlock(dst_ptl);
4277                                 /* Do not use reserve as it's private owned */
4278                                 new = alloc_huge_page(vma, addr, 1);
4279                                 if (IS_ERR(new)) {
4280                                         put_page(ptepage);
4281                                         ret = PTR_ERR(new);
4282                                         break;
4283                                 }
4284                                 copy_user_huge_page(new, ptepage, addr, vma,
4285                                                     npages);
4286                                 put_page(ptepage);
4287
4288                                 /* Install the new huge page if src pte stable */
4289                                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4290                                 src_ptl = huge_pte_lockptr(h, src, src_pte);
4291                                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4292                                 entry = huge_ptep_get(src_pte);
4293                                 if (!pte_same(src_pte_old, entry)) {
4294                                         restore_reserve_on_error(h, vma, addr,
4295                                                                 new);
4296                                         put_page(new);
4297                                         /* dst_entry won't change as in child */
4298                                         goto again;
4299                                 }
4300                                 hugetlb_install_page(vma, dst_pte, addr, new);
4301                                 spin_unlock(src_ptl);
4302                                 spin_unlock(dst_ptl);
4303                                 continue;
4304                         }
4305
4306                         if (cow) {
4307                                 /*
4308                                  * No need to notify as we are downgrading page
4309                                  * table protection not changing it to point
4310                                  * to a new page.
4311                                  *
4312                                  * See Documentation/vm/mmu_notifier.rst
4313                                  */
4314                                 huge_ptep_set_wrprotect(src, addr, src_pte);
4315                                 entry = huge_pte_wrprotect(entry);
4316                         }
4317
4318                         page_dup_rmap(ptepage, true);
4319                         set_huge_pte_at(dst, addr, dst_pte, entry);
4320                         hugetlb_count_add(npages, dst);
4321                 }
4322                 spin_unlock(src_ptl);
4323                 spin_unlock(dst_ptl);
4324         }
4325
4326         if (cow)
4327                 mmu_notifier_invalidate_range_end(&range);
4328         else
4329                 i_mmap_unlock_read(mapping);
4330
4331         return ret;
4332 }
4333
4334 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
4335                             unsigned long start, unsigned long end,
4336                             struct page *ref_page)
4337 {
4338         struct mm_struct *mm = vma->vm_mm;
4339         unsigned long address;
4340         pte_t *ptep;
4341         pte_t pte;
4342         spinlock_t *ptl;
4343         struct page *page;
4344         struct hstate *h = hstate_vma(vma);
4345         unsigned long sz = huge_page_size(h);
4346         struct mmu_notifier_range range;
4347
4348         WARN_ON(!is_vm_hugetlb_page(vma));
4349         BUG_ON(start & ~huge_page_mask(h));
4350         BUG_ON(end & ~huge_page_mask(h));
4351
4352         /*
4353          * This is a hugetlb vma, all the pte entries should point
4354          * to huge page.
4355          */
4356         tlb_change_page_size(tlb, sz);
4357         tlb_start_vma(tlb, vma);
4358
4359         /*
4360          * If sharing possible, alert mmu notifiers of worst case.
4361          */
4362         mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
4363                                 end);
4364         adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4365         mmu_notifier_invalidate_range_start(&range);
4366         address = start;
4367         for (; address < end; address += sz) {
4368                 ptep = huge_pte_offset(mm, address, sz);
4369                 if (!ptep)
4370                         continue;
4371
4372                 ptl = huge_pte_lock(h, mm, ptep);
4373                 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
4374                         spin_unlock(ptl);
4375                         /*
4376                          * We just unmapped a page of PMDs by clearing a PUD.
4377                          * The caller's TLB flush range should cover this area.
4378                          */
4379                         continue;
4380                 }
4381
4382                 pte = huge_ptep_get(ptep);
4383                 if (huge_pte_none(pte)) {
4384                         spin_unlock(ptl);
4385                         continue;
4386                 }
4387
4388                 /*
4389                  * Migrating hugepage or HWPoisoned hugepage is already
4390                  * unmapped and its refcount is dropped, so just clear pte here.
4391                  */
4392                 if (unlikely(!pte_present(pte))) {
4393                         huge_pte_clear(mm, address, ptep, sz);
4394                         spin_unlock(ptl);
4395                         continue;
4396                 }
4397
4398                 page = pte_page(pte);
4399                 /*
4400                  * If a reference page is supplied, it is because a specific
4401                  * page is being unmapped, not a range. Ensure the page we
4402                  * are about to unmap is the actual page of interest.
4403                  */
4404                 if (ref_page) {
4405                         if (page != ref_page) {
4406                                 spin_unlock(ptl);
4407                                 continue;
4408                         }
4409                         /*
4410                          * Mark the VMA as having unmapped its page so that
4411                          * future faults in this VMA will fail rather than
4412                          * looking like data was lost
4413                          */
4414                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
4415                 }
4416
4417                 pte = huge_ptep_get_and_clear(mm, address, ptep);
4418                 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
4419                 if (huge_pte_dirty(pte))
4420                         set_page_dirty(page);
4421
4422                 hugetlb_count_sub(pages_per_huge_page(h), mm);
4423                 page_remove_rmap(page, true);
4424
4425                 spin_unlock(ptl);
4426                 tlb_remove_page_size(tlb, page, huge_page_size(h));
4427                 /*
4428                  * Bail out after unmapping reference page if supplied
4429                  */
4430                 if (ref_page)
4431                         break;
4432         }
4433         mmu_notifier_invalidate_range_end(&range);
4434         tlb_end_vma(tlb, vma);
4435 }
4436
4437 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4438                           struct vm_area_struct *vma, unsigned long start,
4439                           unsigned long end, struct page *ref_page)
4440 {
4441         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4442
4443         /*
4444          * Clear this flag so that x86's huge_pmd_share page_table_shareable
4445          * test will fail on a vma being torn down, and not grab a page table
4446          * on its way out.  We're lucky that the flag has such an appropriate
4447          * name, and can in fact be safely cleared here. We could clear it
4448          * before the __unmap_hugepage_range above, but all that's necessary
4449          * is to clear it before releasing the i_mmap_rwsem. This works
4450          * because in the context this is called, the VMA is about to be
4451          * destroyed and the i_mmap_rwsem is held.
4452          */
4453         vma->vm_flags &= ~VM_MAYSHARE;
4454 }
4455
4456 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
4457                           unsigned long end, struct page *ref_page)
4458 {
4459         struct mmu_gather tlb;
4460
4461         tlb_gather_mmu(&tlb, vma->vm_mm);
4462         __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
4463         tlb_finish_mmu(&tlb);
4464 }
4465
4466 /*
4467  * This is called when the original mapper is failing to COW a MAP_PRIVATE
4468  * mapping it owns the reserve page for. The intention is to unmap the page
4469  * from other VMAs and let the children be SIGKILLed if they are faulting the
4470  * same region.
4471  */
4472 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4473                               struct page *page, unsigned long address)
4474 {
4475         struct hstate *h = hstate_vma(vma);
4476         struct vm_area_struct *iter_vma;
4477         struct address_space *mapping;
4478         pgoff_t pgoff;
4479
4480         /*
4481          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4482          * from page cache lookup which is in HPAGE_SIZE units.
4483          */
4484         address = address & huge_page_mask(h);
4485         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4486                         vma->vm_pgoff;
4487         mapping = vma->vm_file->f_mapping;
4488
4489         /*
4490          * Take the mapping lock for the duration of the table walk. As
4491          * this mapping should be shared between all the VMAs,
4492          * __unmap_hugepage_range() is called as the lock is already held
4493          */
4494         i_mmap_lock_write(mapping);
4495         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
4496                 /* Do not unmap the current VMA */
4497                 if (iter_vma == vma)
4498                         continue;
4499
4500                 /*
4501                  * Shared VMAs have their own reserves and do not affect
4502                  * MAP_PRIVATE accounting but it is possible that a shared
4503                  * VMA is using the same page so check and skip such VMAs.
4504                  */
4505                 if (iter_vma->vm_flags & VM_MAYSHARE)
4506                         continue;
4507
4508                 /*
4509                  * Unmap the page from other VMAs without their own reserves.
4510                  * They get marked to be SIGKILLed if they fault in these
4511                  * areas. This is because a future no-page fault on this VMA
4512                  * could insert a zeroed page instead of the data existing
4513                  * from the time of fork. This would look like data corruption
4514                  */
4515                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
4516                         unmap_hugepage_range(iter_vma, address,
4517                                              address + huge_page_size(h), page);
4518         }
4519         i_mmap_unlock_write(mapping);
4520 }
4521
4522 /*
4523  * Hugetlb_cow() should be called with page lock of the original hugepage held.
4524  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4525  * cannot race with other handlers or page migration.
4526  * Keep the pte_same checks anyway to make transition from the mutex easier.
4527  */
4528 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
4529                        unsigned long address, pte_t *ptep,
4530                        struct page *pagecache_page, spinlock_t *ptl)
4531 {
4532         pte_t pte;
4533         struct hstate *h = hstate_vma(vma);
4534         struct page *old_page, *new_page;
4535         int outside_reserve = 0;
4536         vm_fault_t ret = 0;
4537         unsigned long haddr = address & huge_page_mask(h);
4538         struct mmu_notifier_range range;
4539
4540         pte = huge_ptep_get(ptep);
4541         old_page = pte_page(pte);
4542
4543 retry_avoidcopy:
4544         /* If no-one else is actually using this page, avoid the copy
4545          * and just make the page writable */
4546         if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
4547                 page_move_anon_rmap(old_page, vma);
4548                 set_huge_ptep_writable(vma, haddr, ptep);
4549                 return 0;
4550         }
4551
4552         /*
4553          * If the process that created a MAP_PRIVATE mapping is about to
4554          * perform a COW due to a shared page count, attempt to satisfy
4555          * the allocation without using the existing reserves. The pagecache
4556          * page is used to determine if the reserve at this address was
4557          * consumed or not. If reserves were used, a partial faulted mapping
4558          * at the time of fork() could consume its reserves on COW instead
4559          * of the full address range.
4560          */
4561         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4562                         old_page != pagecache_page)
4563                 outside_reserve = 1;
4564
4565         get_page(old_page);
4566
4567         /*
4568          * Drop page table lock as buddy allocator may be called. It will
4569          * be acquired again before returning to the caller, as expected.
4570          */
4571         spin_unlock(ptl);
4572         new_page = alloc_huge_page(vma, haddr, outside_reserve);
4573
4574         if (IS_ERR(new_page)) {
4575                 /*
4576                  * If a process owning a MAP_PRIVATE mapping fails to COW,
4577                  * it is due to references held by a child and an insufficient
4578                  * huge page pool. To guarantee the original mappers
4579                  * reliability, unmap the page from child processes. The child
4580                  * may get SIGKILLed if it later faults.
4581                  */
4582                 if (outside_reserve) {
4583                         struct address_space *mapping = vma->vm_file->f_mapping;
4584                         pgoff_t idx;
4585                         u32 hash;
4586
4587                         put_page(old_page);
4588                         BUG_ON(huge_pte_none(pte));
4589                         /*
4590                          * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4591                          * unmapping.  unmapping needs to hold i_mmap_rwsem
4592                          * in write mode.  Dropping i_mmap_rwsem in read mode
4593                          * here is OK as COW mappings do not interact with
4594                          * PMD sharing.
4595                          *
4596                          * Reacquire both after unmap operation.
4597                          */
4598                         idx = vma_hugecache_offset(h, vma, haddr);
4599                         hash = hugetlb_fault_mutex_hash(mapping, idx);
4600                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4601                         i_mmap_unlock_read(mapping);
4602
4603                         unmap_ref_private(mm, vma, old_page, haddr);
4604
4605                         i_mmap_lock_read(mapping);
4606                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
4607                         spin_lock(ptl);
4608                         ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4609                         if (likely(ptep &&
4610                                    pte_same(huge_ptep_get(ptep), pte)))
4611                                 goto retry_avoidcopy;
4612                         /*
4613                          * race occurs while re-acquiring page table
4614                          * lock, and our job is done.
4615                          */
4616                         return 0;
4617                 }
4618
4619                 ret = vmf_error(PTR_ERR(new_page));
4620                 goto out_release_old;
4621         }
4622
4623         /*
4624          * When the original hugepage is shared one, it does not have
4625          * anon_vma prepared.
4626          */
4627         if (unlikely(anon_vma_prepare(vma))) {
4628                 ret = VM_FAULT_OOM;
4629                 goto out_release_all;
4630         }
4631
4632         copy_user_huge_page(new_page, old_page, address, vma,
4633                             pages_per_huge_page(h));
4634         __SetPageUptodate(new_page);
4635
4636         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4637                                 haddr + huge_page_size(h));
4638         mmu_notifier_invalidate_range_start(&range);
4639
4640         /*
4641          * Retake the page table lock to check for racing updates
4642          * before the page tables are altered
4643          */
4644         spin_lock(ptl);
4645         ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4646         if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4647                 ClearHPageRestoreReserve(new_page);
4648
4649                 /* Break COW */
4650                 huge_ptep_clear_flush(vma, haddr, ptep);
4651                 mmu_notifier_invalidate_range(mm, range.start, range.end);
4652                 set_huge_pte_at(mm, haddr, ptep,
4653                                 make_huge_pte(vma, new_page, 1));
4654                 page_remove_rmap(old_page, true);
4655                 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4656                 SetHPageMigratable(new_page);
4657                 /* Make the old page be freed below */
4658                 new_page = old_page;
4659         }
4660         spin_unlock(ptl);
4661         mmu_notifier_invalidate_range_end(&range);
4662 out_release_all:
4663         restore_reserve_on_error(h, vma, haddr, new_page);
4664         put_page(new_page);
4665 out_release_old:
4666         put_page(old_page);
4667
4668         spin_lock(ptl); /* Caller expects lock to be held */
4669         return ret;
4670 }
4671
4672 /* Return the pagecache page at a given address within a VMA */
4673 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4674                         struct vm_area_struct *vma, unsigned long address)
4675 {
4676         struct address_space *mapping;
4677         pgoff_t idx;
4678
4679         mapping = vma->vm_file->f_mapping;
4680         idx = vma_hugecache_offset(h, vma, address);
4681
4682         return find_lock_page(mapping, idx);
4683 }
4684
4685 /*
4686  * Return whether there is a pagecache page to back given address within VMA.
4687  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4688  */
4689 static bool hugetlbfs_pagecache_present(struct hstate *h,
4690                         struct vm_area_struct *vma, unsigned long address)
4691 {
4692         struct address_space *mapping;
4693         pgoff_t idx;
4694         struct page *page;
4695
4696         mapping = vma->vm_file->f_mapping;
4697         idx = vma_hugecache_offset(h, vma, address);
4698
4699         page = find_get_page(mapping, idx);
4700         if (page)
4701                 put_page(page);
4702         return page != NULL;
4703 }
4704
4705 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4706                            pgoff_t idx)
4707 {
4708         struct inode *inode = mapping->host;
4709         struct hstate *h = hstate_inode(inode);
4710         int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4711
4712         if (err)
4713                 return err;
4714         ClearHPageRestoreReserve(page);
4715
4716         /*
4717          * set page dirty so that it will not be removed from cache/file
4718          * by non-hugetlbfs specific code paths.
4719          */
4720         set_page_dirty(page);
4721
4722         spin_lock(&inode->i_lock);
4723         inode->i_blocks += blocks_per_huge_page(h);
4724         spin_unlock(&inode->i_lock);
4725         return 0;
4726 }
4727
4728 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
4729                                                   struct address_space *mapping,
4730                                                   pgoff_t idx,
4731                                                   unsigned int flags,
4732                                                   unsigned long haddr,
4733                                                   unsigned long reason)
4734 {
4735         vm_fault_t ret;
4736         u32 hash;
4737         struct vm_fault vmf = {
4738                 .vma = vma,
4739                 .address = haddr,
4740                 .flags = flags,
4741
4742                 /*
4743                  * Hard to debug if it ends up being
4744                  * used by a callee that assumes
4745                  * something about the other
4746                  * uninitialized fields... same as in
4747                  * memory.c
4748                  */
4749         };
4750
4751         /*
4752          * hugetlb_fault_mutex and i_mmap_rwsem must be
4753          * dropped before handling userfault.  Reacquire
4754          * after handling fault to make calling code simpler.
4755          */
4756         hash = hugetlb_fault_mutex_hash(mapping, idx);
4757         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4758         i_mmap_unlock_read(mapping);
4759         ret = handle_userfault(&vmf, reason);
4760         i_mmap_lock_read(mapping);
4761         mutex_lock(&hugetlb_fault_mutex_table[hash]);
4762
4763         return ret;
4764 }
4765
4766 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4767                         struct vm_area_struct *vma,
4768                         struct address_space *mapping, pgoff_t idx,
4769                         unsigned long address, pte_t *ptep, unsigned int flags)
4770 {
4771         struct hstate *h = hstate_vma(vma);
4772         vm_fault_t ret = VM_FAULT_SIGBUS;
4773         int anon_rmap = 0;
4774         unsigned long size;
4775         struct page *page;
4776         pte_t new_pte;
4777         spinlock_t *ptl;
4778         unsigned long haddr = address & huge_page_mask(h);
4779         bool new_page = false;
4780
4781         /*
4782          * Currently, we are forced to kill the process in the event the
4783          * original mapper has unmapped pages from the child due to a failed
4784          * COW. Warn that such a situation has occurred as it may not be obvious
4785          */
4786         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4787                 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4788                            current->pid);
4789                 return ret;
4790         }
4791
4792         /*
4793          * We can not race with truncation due to holding i_mmap_rwsem.
4794          * i_size is modified when holding i_mmap_rwsem, so check here
4795          * once for faults beyond end of file.
4796          */
4797         size = i_size_read(mapping->host) >> huge_page_shift(h);
4798         if (idx >= size)
4799                 goto out;
4800
4801 retry:
4802         page = find_lock_page(mapping, idx);
4803         if (!page) {
4804                 /* Check for page in userfault range */
4805                 if (userfaultfd_missing(vma)) {
4806                         ret = hugetlb_handle_userfault(vma, mapping, idx,
4807                                                        flags, haddr,
4808                                                        VM_UFFD_MISSING);
4809                         goto out;
4810                 }
4811
4812                 page = alloc_huge_page(vma, haddr, 0);
4813                 if (IS_ERR(page)) {
4814                         /*
4815                          * Returning error will result in faulting task being
4816                          * sent SIGBUS.  The hugetlb fault mutex prevents two
4817                          * tasks from racing to fault in the same page which
4818                          * could result in false unable to allocate errors.
4819                          * Page migration does not take the fault mutex, but
4820                          * does a clear then write of pte's under page table
4821                          * lock.  Page fault code could race with migration,
4822                          * notice the clear pte and try to allocate a page
4823                          * here.  Before returning error, get ptl and make
4824                          * sure there really is no pte entry.
4825                          */
4826                         ptl = huge_pte_lock(h, mm, ptep);
4827                         ret = 0;
4828                         if (huge_pte_none(huge_ptep_get(ptep)))
4829                                 ret = vmf_error(PTR_ERR(page));
4830                         spin_unlock(ptl);
4831                         goto out;
4832                 }
4833                 clear_huge_page(page, address, pages_per_huge_page(h));
4834                 __SetPageUptodate(page);
4835                 new_page = true;
4836
4837                 if (vma->vm_flags & VM_MAYSHARE) {
4838                         int err = huge_add_to_page_cache(page, mapping, idx);
4839                         if (err) {
4840                                 put_page(page);
4841                                 if (err == -EEXIST)
4842                                         goto retry;
4843                                 goto out;
4844                         }
4845                 } else {
4846                         lock_page(page);
4847                         if (unlikely(anon_vma_prepare(vma))) {
4848                                 ret = VM_FAULT_OOM;
4849                                 goto backout_unlocked;
4850                         }
4851                         anon_rmap = 1;
4852                 }
4853         } else {
4854                 /*
4855                  * If memory error occurs between mmap() and fault, some process
4856                  * don't have hwpoisoned swap entry for errored virtual address.
4857                  * So we need to block hugepage fault by PG_hwpoison bit check.
4858                  */
4859                 if (unlikely(PageHWPoison(page))) {
4860                         ret = VM_FAULT_HWPOISON_LARGE |
4861                                 VM_FAULT_SET_HINDEX(hstate_index(h));
4862                         goto backout_unlocked;
4863                 }
4864
4865                 /* Check for page in userfault range. */
4866                 if (userfaultfd_minor(vma)) {
4867                         unlock_page(page);
4868                         put_page(page);
4869                         ret = hugetlb_handle_userfault(vma, mapping, idx,
4870                                                        flags, haddr,
4871                                                        VM_UFFD_MINOR);
4872                         goto out;
4873                 }
4874         }
4875
4876         /*
4877          * If we are going to COW a private mapping later, we examine the
4878          * pending reservations for this page now. This will ensure that
4879          * any allocations necessary to record that reservation occur outside
4880          * the spinlock.
4881          */
4882         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4883                 if (vma_needs_reservation(h, vma, haddr) < 0) {
4884                         ret = VM_FAULT_OOM;
4885                         goto backout_unlocked;
4886                 }
4887                 /* Just decrements count, does not deallocate */
4888                 vma_end_reservation(h, vma, haddr);
4889         }
4890
4891         ptl = huge_pte_lock(h, mm, ptep);
4892         ret = 0;
4893         if (!huge_pte_none(huge_ptep_get(ptep)))
4894                 goto backout;
4895
4896         if (anon_rmap) {
4897                 ClearHPageRestoreReserve(page);
4898                 hugepage_add_new_anon_rmap(page, vma, haddr);
4899         } else
4900                 page_dup_rmap(page, true);
4901         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4902                                 && (vma->vm_flags & VM_SHARED)));
4903         set_huge_pte_at(mm, haddr, ptep, new_pte);
4904
4905         hugetlb_count_add(pages_per_huge_page(h), mm);
4906         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4907                 /* Optimization, do the COW without a second fault */
4908                 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4909         }
4910
4911         spin_unlock(ptl);
4912
4913         /*
4914          * Only set HPageMigratable in newly allocated pages.  Existing pages
4915          * found in the pagecache may not have HPageMigratableset if they have
4916          * been isolated for migration.
4917          */
4918         if (new_page)
4919                 SetHPageMigratable(page);
4920
4921         unlock_page(page);
4922 out:
4923         return ret;
4924
4925 backout:
4926         spin_unlock(ptl);
4927 backout_unlocked:
4928         unlock_page(page);
4929         restore_reserve_on_error(h, vma, haddr, page);
4930         put_page(page);
4931         goto out;
4932 }
4933
4934 #ifdef CONFIG_SMP
4935 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4936 {
4937         unsigned long key[2];
4938         u32 hash;
4939
4940         key[0] = (unsigned long) mapping;
4941         key[1] = idx;
4942
4943         hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4944
4945         return hash & (num_fault_mutexes - 1);
4946 }
4947 #else
4948 /*
4949  * For uniprocessor systems we always use a single mutex, so just
4950  * return 0 and avoid the hashing overhead.
4951  */
4952 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4953 {
4954         return 0;
4955 }
4956 #endif
4957
4958 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4959                         unsigned long address, unsigned int flags)
4960 {
4961         pte_t *ptep, entry;
4962         spinlock_t *ptl;
4963         vm_fault_t ret;
4964         u32 hash;
4965         pgoff_t idx;
4966         struct page *page = NULL;
4967         struct page *pagecache_page = NULL;
4968         struct hstate *h = hstate_vma(vma);
4969         struct address_space *mapping;
4970         int need_wait_lock = 0;
4971         unsigned long haddr = address & huge_page_mask(h);
4972
4973         ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4974         if (ptep) {
4975                 /*
4976                  * Since we hold no locks, ptep could be stale.  That is
4977                  * OK as we are only making decisions based on content and
4978                  * not actually modifying content here.
4979                  */
4980                 entry = huge_ptep_get(ptep);
4981                 if (unlikely(is_hugetlb_entry_migration(entry))) {
4982                         migration_entry_wait_huge(vma, mm, ptep);
4983                         return 0;
4984                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4985                         return VM_FAULT_HWPOISON_LARGE |
4986                                 VM_FAULT_SET_HINDEX(hstate_index(h));
4987         }
4988
4989         /*
4990          * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4991          * until finished with ptep.  This serves two purposes:
4992          * 1) It prevents huge_pmd_unshare from being called elsewhere
4993          *    and making the ptep no longer valid.
4994          * 2) It synchronizes us with i_size modifications during truncation.
4995          *
4996          * ptep could have already be assigned via huge_pte_offset.  That
4997          * is OK, as huge_pte_alloc will return the same value unless
4998          * something has changed.
4999          */
5000         mapping = vma->vm_file->f_mapping;
5001         i_mmap_lock_read(mapping);
5002         ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
5003         if (!ptep) {
5004                 i_mmap_unlock_read(mapping);
5005                 return VM_FAULT_OOM;
5006         }
5007
5008         /*
5009          * Serialize hugepage allocation and instantiation, so that we don't
5010          * get spurious allocation failures if two CPUs race to instantiate
5011          * the same page in the page cache.
5012          */
5013         idx = vma_hugecache_offset(h, vma, haddr);
5014         hash = hugetlb_fault_mutex_hash(mapping, idx);
5015         mutex_lock(&hugetlb_fault_mutex_table[hash]);
5016
5017         entry = huge_ptep_get(ptep);
5018         if (huge_pte_none(entry)) {
5019                 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
5020                 goto out_mutex;
5021         }
5022
5023         ret = 0;
5024
5025         /*
5026          * entry could be a migration/hwpoison entry at this point, so this
5027          * check prevents the kernel from going below assuming that we have
5028          * an active hugepage in pagecache. This goto expects the 2nd page
5029          * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
5030          * properly handle it.
5031          */
5032         if (!pte_present(entry))
5033                 goto out_mutex;
5034
5035         /*
5036          * If we are going to COW the mapping later, we examine the pending
5037          * reservations for this page now. This will ensure that any
5038          * allocations necessary to record that reservation occur outside the
5039          * spinlock. For private mappings, we also lookup the pagecache
5040          * page now as it is used to determine if a reservation has been
5041          * consumed.
5042          */
5043         if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
5044                 if (vma_needs_reservation(h, vma, haddr) < 0) {
5045                         ret = VM_FAULT_OOM;
5046                         goto out_mutex;
5047                 }
5048                 /* Just decrements count, does not deallocate */
5049                 vma_end_reservation(h, vma, haddr);
5050
5051                 if (!(vma->vm_flags & VM_MAYSHARE))
5052                         pagecache_page = hugetlbfs_pagecache_page(h,
5053                                                                 vma, haddr);
5054         }
5055
5056         ptl = huge_pte_lock(h, mm, ptep);
5057
5058         /* Check for a racing update before calling hugetlb_cow */
5059         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
5060                 goto out_ptl;
5061
5062         /*
5063          * hugetlb_cow() requires page locks of pte_page(entry) and
5064          * pagecache_page, so here we need take the former one
5065          * when page != pagecache_page or !pagecache_page.
5066          */
5067         page = pte_page(entry);
5068         if (page != pagecache_page)
5069                 if (!trylock_page(page)) {
5070                         need_wait_lock = 1;
5071                         goto out_ptl;
5072                 }
5073
5074         get_page(page);
5075
5076         if (flags & FAULT_FLAG_WRITE) {
5077                 if (!huge_pte_write(entry)) {
5078                         ret = hugetlb_cow(mm, vma, address, ptep,
5079                                           pagecache_page, ptl);
5080                         goto out_put_page;
5081                 }
5082                 entry = huge_pte_mkdirty(entry);
5083         }
5084         entry = pte_mkyoung(entry);
5085         if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
5086                                                 flags & FAULT_FLAG_WRITE))
5087                 update_mmu_cache(vma, haddr, ptep);
5088 out_put_page:
5089         if (page != pagecache_page)
5090                 unlock_page(page);
5091         put_page(page);
5092 out_ptl:
5093         spin_unlock(ptl);
5094
5095         if (pagecache_page) {
5096                 unlock_page(pagecache_page);
5097                 put_page(pagecache_page);
5098         }
5099 out_mutex:
5100         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5101         i_mmap_unlock_read(mapping);
5102         /*
5103          * Generally it's safe to hold refcount during waiting page lock. But
5104          * here we just wait to defer the next page fault to avoid busy loop and
5105          * the page is not used after unlocked before returning from the current
5106          * page fault. So we are safe from accessing freed page, even if we wait
5107          * here without taking refcount.
5108          */
5109         if (need_wait_lock)
5110                 wait_on_page_locked(page);
5111         return ret;
5112 }
5113
5114 #ifdef CONFIG_USERFAULTFD
5115 /*
5116  * Used by userfaultfd UFFDIO_COPY.  Based on mcopy_atomic_pte with
5117  * modifications for huge pages.
5118  */
5119 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
5120                             pte_t *dst_pte,
5121                             struct vm_area_struct *dst_vma,
5122                             unsigned long dst_addr,
5123                             unsigned long src_addr,
5124                             enum mcopy_atomic_mode mode,
5125                             struct page **pagep)
5126 {
5127         bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE);
5128         struct hstate *h = hstate_vma(dst_vma);
5129         struct address_space *mapping = dst_vma->vm_file->f_mapping;
5130         pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
5131         unsigned long size;
5132         int vm_shared = dst_vma->vm_flags & VM_SHARED;
5133         pte_t _dst_pte;
5134         spinlock_t *ptl;
5135         int ret = -ENOMEM;
5136         struct page *page;
5137         int writable;
5138
5139         if (is_continue) {
5140                 ret = -EFAULT;
5141                 page = find_lock_page(mapping, idx);
5142                 if (!page)
5143                         goto out;
5144         } else if (!*pagep) {
5145                 /* If a page already exists, then it's UFFDIO_COPY for
5146                  * a non-missing case. Return -EEXIST.
5147                  */
5148                 if (vm_shared &&
5149                     hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5150                         ret = -EEXIST;
5151                         goto out;
5152                 }
5153
5154                 page = alloc_huge_page(dst_vma, dst_addr, 0);
5155                 if (IS_ERR(page)) {
5156                         ret = -ENOMEM;
5157                         goto out;
5158                 }
5159
5160                 ret = copy_huge_page_from_user(page,
5161                                                 (const void __user *) src_addr,
5162                                                 pages_per_huge_page(h), false);
5163
5164                 /* fallback to copy_from_user outside mmap_lock */
5165                 if (unlikely(ret)) {
5166                         ret = -ENOENT;
5167                         /* Free the allocated page which may have
5168                          * consumed a reservation.
5169                          */
5170                         restore_reserve_on_error(h, dst_vma, dst_addr, page);
5171                         put_page(page);
5172
5173                         /* Allocate a temporary page to hold the copied
5174                          * contents.
5175                          */
5176                         page = alloc_huge_page_vma(h, dst_vma, dst_addr);
5177                         if (!page) {
5178                                 ret = -ENOMEM;
5179                                 goto out;
5180                         }
5181                         *pagep = page;
5182                         /* Set the outparam pagep and return to the caller to
5183                          * copy the contents outside the lock. Don't free the
5184                          * page.
5185                          */
5186                         goto out;
5187                 }
5188         } else {
5189                 if (vm_shared &&
5190                     hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5191                         put_page(*pagep);
5192                         ret = -EEXIST;
5193                         *pagep = NULL;
5194                         goto out;
5195                 }
5196
5197                 page = alloc_huge_page(dst_vma, dst_addr, 0);
5198                 if (IS_ERR(page)) {
5199                         ret = -ENOMEM;
5200                         *pagep = NULL;
5201                         goto out;
5202                 }
5203                 copy_huge_page(page, *pagep);
5204                 put_page(*pagep);
5205                 *pagep = NULL;
5206         }
5207
5208         /*
5209          * The memory barrier inside __SetPageUptodate makes sure that
5210          * preceding stores to the page contents become visible before
5211          * the set_pte_at() write.
5212          */
5213         __SetPageUptodate(page);
5214
5215         /* Add shared, newly allocated pages to the page cache. */
5216         if (vm_shared && !is_continue) {
5217                 size = i_size_read(mapping->host) >> huge_page_shift(h);
5218                 ret = -EFAULT;
5219                 if (idx >= size)
5220                         goto out_release_nounlock;
5221
5222                 /*
5223                  * Serialization between remove_inode_hugepages() and
5224                  * huge_add_to_page_cache() below happens through the
5225                  * hugetlb_fault_mutex_table that here must be hold by
5226                  * the caller.
5227                  */
5228                 ret = huge_add_to_page_cache(page, mapping, idx);
5229                 if (ret)
5230                         goto out_release_nounlock;
5231         }
5232
5233         ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
5234         spin_lock(ptl);
5235
5236         /*
5237          * Recheck the i_size after holding PT lock to make sure not
5238          * to leave any page mapped (as page_mapped()) beyond the end
5239          * of the i_size (remove_inode_hugepages() is strict about
5240          * enforcing that). If we bail out here, we'll also leave a
5241          * page in the radix tree in the vm_shared case beyond the end
5242          * of the i_size, but remove_inode_hugepages() will take care
5243          * of it as soon as we drop the hugetlb_fault_mutex_table.
5244          */
5245         size = i_size_read(mapping->host) >> huge_page_shift(h);
5246         ret = -EFAULT;
5247         if (idx >= size)
5248                 goto out_release_unlock;
5249
5250         ret = -EEXIST;
5251         if (!huge_pte_none(huge_ptep_get(dst_pte)))
5252                 goto out_release_unlock;
5253
5254         if (vm_shared) {
5255                 page_dup_rmap(page, true);
5256         } else {
5257                 ClearHPageRestoreReserve(page);
5258                 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
5259         }
5260
5261         /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */
5262         if (is_continue && !vm_shared)
5263                 writable = 0;
5264         else
5265                 writable = dst_vma->vm_flags & VM_WRITE;
5266
5267         _dst_pte = make_huge_pte(dst_vma, page, writable);
5268         if (writable)
5269                 _dst_pte = huge_pte_mkdirty(_dst_pte);
5270         _dst_pte = pte_mkyoung(_dst_pte);
5271
5272         set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
5273
5274         (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
5275                                         dst_vma->vm_flags & VM_WRITE);
5276         hugetlb_count_add(pages_per_huge_page(h), dst_mm);
5277
5278         /* No need to invalidate - it was non-present before */
5279         update_mmu_cache(dst_vma, dst_addr, dst_pte);
5280
5281         spin_unlock(ptl);
5282         if (!is_continue)
5283                 SetHPageMigratable(page);
5284         if (vm_shared || is_continue)
5285                 unlock_page(page);
5286         ret = 0;
5287 out:
5288         return ret;
5289 out_release_unlock:
5290         spin_unlock(ptl);
5291         if (vm_shared || is_continue)
5292                 unlock_page(page);
5293 out_release_nounlock:
5294         restore_reserve_on_error(h, dst_vma, dst_addr, page);
5295         put_page(page);
5296         goto out;
5297 }
5298 #endif /* CONFIG_USERFAULTFD */
5299
5300 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
5301                                  int refs, struct page **pages,
5302                                  struct vm_area_struct **vmas)
5303 {
5304         int nr;
5305
5306         for (nr = 0; nr < refs; nr++) {
5307                 if (likely(pages))
5308                         pages[nr] = mem_map_offset(page, nr);
5309                 if (vmas)
5310                         vmas[nr] = vma;
5311         }
5312 }
5313
5314 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
5315                          struct page **pages, struct vm_area_struct **vmas,
5316                          unsigned long *position, unsigned long *nr_pages,
5317                          long i, unsigned int flags, int *locked)
5318 {
5319         unsigned long pfn_offset;
5320         unsigned long vaddr = *position;
5321         unsigned long remainder = *nr_pages;
5322         struct hstate *h = hstate_vma(vma);
5323         int err = -EFAULT, refs;
5324
5325         while (vaddr < vma->vm_end && remainder) {
5326                 pte_t *pte;
5327                 spinlock_t *ptl = NULL;
5328                 int absent;
5329                 struct page *page;
5330
5331                 /*
5332                  * If we have a pending SIGKILL, don't keep faulting pages and
5333                  * potentially allocating memory.
5334                  */
5335                 if (fatal_signal_pending(current)) {
5336                         remainder = 0;
5337                         break;
5338                 }
5339
5340                 /*
5341                  * Some archs (sparc64, sh*) have multiple pte_ts to
5342                  * each hugepage.  We have to make sure we get the
5343                  * first, for the page indexing below to work.
5344                  *
5345                  * Note that page table lock is not held when pte is null.
5346                  */
5347                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
5348                                       huge_page_size(h));
5349                 if (pte)
5350                         ptl = huge_pte_lock(h, mm, pte);
5351                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
5352
5353                 /*
5354                  * When coredumping, it suits get_dump_page if we just return
5355                  * an error where there's an empty slot with no huge pagecache
5356                  * to back it.  This way, we avoid allocating a hugepage, and
5357                  * the sparse dumpfile avoids allocating disk blocks, but its
5358                  * huge holes still show up with zeroes where they need to be.
5359                  */
5360                 if (absent && (flags & FOLL_DUMP) &&
5361                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
5362                         if (pte)
5363                                 spin_unlock(ptl);
5364                         remainder = 0;
5365                         break;
5366                 }
5367
5368                 /*
5369                  * We need call hugetlb_fault for both hugepages under migration
5370                  * (in which case hugetlb_fault waits for the migration,) and
5371                  * hwpoisoned hugepages (in which case we need to prevent the
5372                  * caller from accessing to them.) In order to do this, we use
5373                  * here is_swap_pte instead of is_hugetlb_entry_migration and
5374                  * is_hugetlb_entry_hwpoisoned. This is because it simply covers
5375                  * both cases, and because we can't follow correct pages
5376                  * directly from any kind of swap entries.
5377                  */
5378                 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
5379                     ((flags & FOLL_WRITE) &&
5380                       !huge_pte_write(huge_ptep_get(pte)))) {
5381                         vm_fault_t ret;
5382                         unsigned int fault_flags = 0;
5383
5384                         if (pte)
5385                                 spin_unlock(ptl);
5386                         if (flags & FOLL_WRITE)
5387                                 fault_flags |= FAULT_FLAG_WRITE;
5388                         if (locked)
5389                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
5390                                         FAULT_FLAG_KILLABLE;
5391                         if (flags & FOLL_NOWAIT)
5392                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
5393                                         FAULT_FLAG_RETRY_NOWAIT;
5394                         if (flags & FOLL_TRIED) {
5395                                 /*
5396                                  * Note: FAULT_FLAG_ALLOW_RETRY and
5397                                  * FAULT_FLAG_TRIED can co-exist
5398                                  */
5399                                 fault_flags |= FAULT_FLAG_TRIED;
5400                         }
5401                         ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
5402                         if (ret & VM_FAULT_ERROR) {
5403                                 err = vm_fault_to_errno(ret, flags);
5404                                 remainder = 0;
5405                                 break;
5406                         }
5407                         if (ret & VM_FAULT_RETRY) {
5408                                 if (locked &&
5409                                     !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
5410                                         *locked = 0;
5411                                 *nr_pages = 0;
5412                                 /*
5413                                  * VM_FAULT_RETRY must not return an
5414                                  * error, it will return zero
5415                                  * instead.
5416                                  *
5417                                  * No need to update "position" as the
5418                                  * caller will not check it after
5419                                  * *nr_pages is set to 0.
5420                                  */
5421                                 return i;
5422                         }
5423                         continue;
5424                 }
5425
5426                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
5427                 page = pte_page(huge_ptep_get(pte));
5428
5429                 /*
5430                  * If subpage information not requested, update counters
5431                  * and skip the same_page loop below.
5432                  */
5433                 if (!pages && !vmas && !pfn_offset &&
5434                     (vaddr + huge_page_size(h) < vma->vm_end) &&
5435                     (remainder >= pages_per_huge_page(h))) {
5436                         vaddr += huge_page_size(h);
5437                         remainder -= pages_per_huge_page(h);
5438                         i += pages_per_huge_page(h);
5439                         spin_unlock(ptl);
5440                         continue;
5441                 }
5442
5443                 refs = min3(pages_per_huge_page(h) - pfn_offset,
5444                             (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder);
5445
5446                 if (pages || vmas)
5447                         record_subpages_vmas(mem_map_offset(page, pfn_offset),
5448                                              vma, refs,
5449                                              likely(pages) ? pages + i : NULL,
5450                                              vmas ? vmas + i : NULL);
5451
5452                 if (pages) {
5453                         /*
5454                          * try_grab_compound_head() should always succeed here,
5455                          * because: a) we hold the ptl lock, and b) we've just
5456                          * checked that the huge page is present in the page
5457                          * tables. If the huge page is present, then the tail
5458                          * pages must also be present. The ptl prevents the
5459                          * head page and tail pages from being rearranged in
5460                          * any way. So this page must be available at this
5461                          * point, unless the page refcount overflowed:
5462                          */
5463                         if (WARN_ON_ONCE(!try_grab_compound_head(pages[i],
5464                                                                  refs,
5465                                                                  flags))) {
5466                                 spin_unlock(ptl);
5467                                 remainder = 0;
5468                                 err = -ENOMEM;
5469                                 break;
5470                         }
5471                 }
5472
5473                 vaddr += (refs << PAGE_SHIFT);
5474                 remainder -= refs;
5475                 i += refs;
5476
5477                 spin_unlock(ptl);
5478         }
5479         *nr_pages = remainder;
5480         /*
5481          * setting position is actually required only if remainder is
5482          * not zero but it's faster not to add a "if (remainder)"
5483          * branch.
5484          */
5485         *position = vaddr;
5486
5487         return i ? i : err;
5488 }
5489
5490 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
5491                 unsigned long address, unsigned long end, pgprot_t newprot)
5492 {
5493         struct mm_struct *mm = vma->vm_mm;
5494         unsigned long start = address;
5495         pte_t *ptep;
5496         pte_t pte;
5497         struct hstate *h = hstate_vma(vma);
5498         unsigned long pages = 0;
5499         bool shared_pmd = false;
5500         struct mmu_notifier_range range;
5501
5502         /*
5503          * In the case of shared PMDs, the area to flush could be beyond
5504          * start/end.  Set range.start/range.end to cover the maximum possible
5505          * range if PMD sharing is possible.
5506          */
5507         mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5508                                 0, vma, mm, start, end);
5509         adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5510
5511         BUG_ON(address >= end);
5512         flush_cache_range(vma, range.start, range.end);
5513
5514         mmu_notifier_invalidate_range_start(&range);
5515         i_mmap_lock_write(vma->vm_file->f_mapping);
5516         for (; address < end; address += huge_page_size(h)) {
5517                 spinlock_t *ptl;
5518                 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5519                 if (!ptep)
5520                         continue;
5521                 ptl = huge_pte_lock(h, mm, ptep);
5522                 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
5523                         pages++;
5524                         spin_unlock(ptl);
5525                         shared_pmd = true;
5526                         continue;
5527                 }
5528                 pte = huge_ptep_get(ptep);
5529                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5530                         spin_unlock(ptl);
5531                         continue;
5532                 }
5533                 if (unlikely(is_hugetlb_entry_migration(pte))) {
5534                         swp_entry_t entry = pte_to_swp_entry(pte);
5535
5536                         if (is_writable_migration_entry(entry)) {
5537                                 pte_t newpte;
5538
5539                                 entry = make_readable_migration_entry(
5540                                                         swp_offset(entry));
5541                                 newpte = swp_entry_to_pte(entry);
5542                                 set_huge_swap_pte_at(mm, address, ptep,
5543                                                      newpte, huge_page_size(h));
5544                                 pages++;
5545                         }
5546                         spin_unlock(ptl);
5547                         continue;
5548                 }
5549                 if (!huge_pte_none(pte)) {
5550                         pte_t old_pte;
5551                         unsigned int shift = huge_page_shift(hstate_vma(vma));
5552
5553                         old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5554                         pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
5555                         pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
5556                         huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
5557                         pages++;
5558                 }
5559                 spin_unlock(ptl);
5560         }
5561         /*
5562          * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5563          * may have cleared our pud entry and done put_page on the page table:
5564          * once we release i_mmap_rwsem, another task can do the final put_page
5565          * and that page table be reused and filled with junk.  If we actually
5566          * did unshare a page of pmds, flush the range corresponding to the pud.
5567          */
5568         if (shared_pmd)
5569                 flush_hugetlb_tlb_range(vma, range.start, range.end);
5570         else
5571                 flush_hugetlb_tlb_range(vma, start, end);
5572         /*
5573          * No need to call mmu_notifier_invalidate_range() we are downgrading
5574          * page table protection not changing it to point to a new page.
5575          *
5576          * See Documentation/vm/mmu_notifier.rst
5577          */
5578         i_mmap_unlock_write(vma->vm_file->f_mapping);
5579         mmu_notifier_invalidate_range_end(&range);
5580
5581         return pages << h->order;
5582 }
5583
5584 /* Return true if reservation was successful, false otherwise.  */
5585 bool hugetlb_reserve_pages(struct inode *inode,
5586                                         long from, long to,
5587                                         struct vm_area_struct *vma,
5588                                         vm_flags_t vm_flags)
5589 {
5590         long chg, add = -1;
5591         struct hstate *h = hstate_inode(inode);
5592         struct hugepage_subpool *spool = subpool_inode(inode);
5593         struct resv_map *resv_map;
5594         struct hugetlb_cgroup *h_cg = NULL;
5595         long gbl_reserve, regions_needed = 0;
5596
5597         /* This should never happen */
5598         if (from > to) {
5599                 VM_WARN(1, "%s called with a negative range\n", __func__);
5600                 return false;
5601         }
5602
5603         /*
5604          * Only apply hugepage reservation if asked. At fault time, an
5605          * attempt will be made for VM_NORESERVE to allocate a page
5606          * without using reserves
5607          */
5608         if (vm_flags & VM_NORESERVE)
5609                 return true;
5610
5611         /*
5612          * Shared mappings base their reservation on the number of pages that
5613          * are already allocated on behalf of the file. Private mappings need
5614          * to reserve the full area even if read-only as mprotect() may be
5615          * called to make the mapping read-write. Assume !vma is a shm mapping
5616          */
5617         if (!vma || vma->vm_flags & VM_MAYSHARE) {
5618                 /*
5619                  * resv_map can not be NULL as hugetlb_reserve_pages is only
5620                  * called for inodes for which resv_maps were created (see
5621                  * hugetlbfs_get_inode).
5622                  */
5623                 resv_map = inode_resv_map(inode);
5624
5625                 chg = region_chg(resv_map, from, to, &regions_needed);
5626
5627         } else {
5628                 /* Private mapping. */
5629                 resv_map = resv_map_alloc();
5630                 if (!resv_map)
5631                         return false;
5632
5633                 chg = to - from;
5634
5635                 set_vma_resv_map(vma, resv_map);
5636                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5637         }
5638
5639         if (chg < 0)
5640                 goto out_err;
5641
5642         if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
5643                                 chg * pages_per_huge_page(h), &h_cg) < 0)
5644                 goto out_err;
5645
5646         if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5647                 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5648                  * of the resv_map.
5649                  */
5650                 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5651         }
5652
5653         /*
5654          * There must be enough pages in the subpool for the mapping. If
5655          * the subpool has a minimum size, there may be some global
5656          * reservations already in place (gbl_reserve).
5657          */
5658         gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5659         if (gbl_reserve < 0)
5660                 goto out_uncharge_cgroup;
5661
5662         /*
5663          * Check enough hugepages are available for the reservation.
5664          * Hand the pages back to the subpool if there are not
5665          */
5666         if (hugetlb_acct_memory(h, gbl_reserve) < 0)
5667                 goto out_put_pages;
5668
5669         /*
5670          * Account for the reservations made. Shared mappings record regions
5671          * that have reservations as they are shared by multiple VMAs.
5672          * When the last VMA disappears, the region map says how much
5673          * the reservation was and the page cache tells how much of
5674          * the reservation was consumed. Private mappings are per-VMA and
5675          * only the consumed reservations are tracked. When the VMA
5676          * disappears, the original reservation is the VMA size and the
5677          * consumed reservations are stored in the map. Hence, nothing
5678          * else has to be done for private mappings here
5679          */
5680         if (!vma || vma->vm_flags & VM_MAYSHARE) {
5681                 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5682
5683                 if (unlikely(add < 0)) {
5684                         hugetlb_acct_memory(h, -gbl_reserve);
5685                         goto out_put_pages;
5686                 } else if (unlikely(chg > add)) {
5687                         /*
5688                          * pages in this range were added to the reserve
5689                          * map between region_chg and region_add.  This
5690                          * indicates a race with alloc_huge_page.  Adjust
5691                          * the subpool and reserve counts modified above
5692                          * based on the difference.
5693                          */
5694                         long rsv_adjust;
5695
5696                         /*
5697                          * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
5698                          * reference to h_cg->css. See comment below for detail.
5699                          */
5700                         hugetlb_cgroup_uncharge_cgroup_rsvd(
5701                                 hstate_index(h),
5702                                 (chg - add) * pages_per_huge_page(h), h_cg);
5703
5704                         rsv_adjust = hugepage_subpool_put_pages(spool,
5705                                                                 chg - add);
5706                         hugetlb_acct_memory(h, -rsv_adjust);
5707                 } else if (h_cg) {
5708                         /*
5709                          * The file_regions will hold their own reference to
5710                          * h_cg->css. So we should release the reference held
5711                          * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
5712                          * done.
5713                          */
5714                         hugetlb_cgroup_put_rsvd_cgroup(h_cg);
5715                 }
5716         }
5717         return true;
5718
5719 out_put_pages:
5720         /* put back original number of pages, chg */
5721         (void)hugepage_subpool_put_pages(spool, chg);
5722 out_uncharge_cgroup:
5723         hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5724                                             chg * pages_per_huge_page(h), h_cg);
5725 out_err:
5726         if (!vma || vma->vm_flags & VM_MAYSHARE)
5727                 /* Only call region_abort if the region_chg succeeded but the
5728                  * region_add failed or didn't run.
5729                  */
5730                 if (chg >= 0 && add < 0)
5731                         region_abort(resv_map, from, to, regions_needed);
5732         if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5733                 kref_put(&resv_map->refs, resv_map_release);
5734         return false;
5735 }
5736
5737 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5738                                                                 long freed)
5739 {
5740         struct hstate *h = hstate_inode(inode);
5741         struct resv_map *resv_map = inode_resv_map(inode);
5742         long chg = 0;
5743         struct hugepage_subpool *spool = subpool_inode(inode);
5744         long gbl_reserve;
5745
5746         /*
5747          * Since this routine can be called in the evict inode path for all
5748          * hugetlbfs inodes, resv_map could be NULL.
5749          */
5750         if (resv_map) {
5751                 chg = region_del(resv_map, start, end);
5752                 /*
5753                  * region_del() can fail in the rare case where a region
5754                  * must be split and another region descriptor can not be
5755                  * allocated.  If end == LONG_MAX, it will not fail.
5756                  */
5757                 if (chg < 0)
5758                         return chg;
5759         }
5760
5761         spin_lock(&inode->i_lock);
5762         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5763         spin_unlock(&inode->i_lock);
5764
5765         /*
5766          * If the subpool has a minimum size, the number of global
5767          * reservations to be released may be adjusted.
5768          *
5769          * Note that !resv_map implies freed == 0. So (chg - freed)
5770          * won't go negative.
5771          */
5772         gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5773         hugetlb_acct_memory(h, -gbl_reserve);
5774
5775         return 0;
5776 }
5777
5778 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5779 static unsigned long page_table_shareable(struct vm_area_struct *svma,
5780                                 struct vm_area_struct *vma,
5781                                 unsigned long addr, pgoff_t idx)
5782 {
5783         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5784                                 svma->vm_start;
5785         unsigned long sbase = saddr & PUD_MASK;
5786         unsigned long s_end = sbase + PUD_SIZE;
5787
5788         /* Allow segments to share if only one is marked locked */
5789         unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5790         unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5791
5792         /*
5793          * match the virtual addresses, permission and the alignment of the
5794          * page table page.
5795          */
5796         if (pmd_index(addr) != pmd_index(saddr) ||
5797             vm_flags != svm_flags ||
5798             !range_in_vma(svma, sbase, s_end))
5799                 return 0;
5800
5801         return saddr;
5802 }
5803
5804 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5805 {
5806         unsigned long base = addr & PUD_MASK;
5807         unsigned long end = base + PUD_SIZE;
5808
5809         /*
5810          * check on proper vm_flags and page table alignment
5811          */
5812         if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5813                 return true;
5814         return false;
5815 }
5816
5817 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5818 {
5819 #ifdef CONFIG_USERFAULTFD
5820         if (uffd_disable_huge_pmd_share(vma))
5821                 return false;
5822 #endif
5823         return vma_shareable(vma, addr);
5824 }
5825
5826 /*
5827  * Determine if start,end range within vma could be mapped by shared pmd.
5828  * If yes, adjust start and end to cover range associated with possible
5829  * shared pmd mappings.
5830  */
5831 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5832                                 unsigned long *start, unsigned long *end)
5833 {
5834         unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
5835                 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
5836
5837         /*
5838          * vma needs to span at least one aligned PUD size, and the range
5839          * must be at least partially within in.
5840          */
5841         if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
5842                 (*end <= v_start) || (*start >= v_end))
5843                 return;
5844
5845         /* Extend the range to be PUD aligned for a worst case scenario */
5846         if (*start > v_start)
5847                 *start = ALIGN_DOWN(*start, PUD_SIZE);
5848
5849         if (*end < v_end)
5850                 *end = ALIGN(*end, PUD_SIZE);
5851 }
5852
5853 /*
5854  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5855  * and returns the corresponding pte. While this is not necessary for the
5856  * !shared pmd case because we can allocate the pmd later as well, it makes the
5857  * code much cleaner.
5858  *
5859  * This routine must be called with i_mmap_rwsem held in at least read mode if
5860  * sharing is possible.  For hugetlbfs, this prevents removal of any page
5861  * table entries associated with the address space.  This is important as we
5862  * are setting up sharing based on existing page table entries (mappings).
5863  *
5864  * NOTE: This routine is only called from huge_pte_alloc.  Some callers of
5865  * huge_pte_alloc know that sharing is not possible and do not take
5866  * i_mmap_rwsem as a performance optimization.  This is handled by the
5867  * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5868  * only required for subsequent processing.
5869  */
5870 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5871                       unsigned long addr, pud_t *pud)
5872 {
5873         struct address_space *mapping = vma->vm_file->f_mapping;
5874         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5875                         vma->vm_pgoff;
5876         struct vm_area_struct *svma;
5877         unsigned long saddr;
5878         pte_t *spte = NULL;
5879         pte_t *pte;
5880         spinlock_t *ptl;
5881
5882         i_mmap_assert_locked(mapping);
5883         vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5884                 if (svma == vma)
5885                         continue;
5886
5887                 saddr = page_table_shareable(svma, vma, addr, idx);
5888                 if (saddr) {
5889                         spte = huge_pte_offset(svma->vm_mm, saddr,
5890                                                vma_mmu_pagesize(svma));
5891                         if (spte) {
5892                                 get_page(virt_to_page(spte));
5893                                 break;
5894                         }
5895                 }
5896         }
5897
5898         if (!spte)
5899                 goto out;
5900
5901         ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
5902         if (pud_none(*pud)) {
5903                 pud_populate(mm, pud,
5904                                 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5905                 mm_inc_nr_pmds(mm);
5906         } else {
5907                 put_page(virt_to_page(spte));
5908         }
5909         spin_unlock(ptl);
5910 out:
5911         pte = (pte_t *)pmd_alloc(mm, pud, addr);
5912         return pte;
5913 }
5914
5915 /*
5916  * unmap huge page backed by shared pte.
5917  *
5918  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
5919  * indicated by page_count > 1, unmap is achieved by clearing pud and
5920  * decrementing the ref count. If count == 1, the pte page is not shared.
5921  *
5922  * Called with page table lock held and i_mmap_rwsem held in write mode.
5923  *
5924  * returns: 1 successfully unmapped a shared pte page
5925  *          0 the underlying pte page is not shared, or it is the last user
5926  */
5927 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5928                                         unsigned long *addr, pte_t *ptep)
5929 {
5930         pgd_t *pgd = pgd_offset(mm, *addr);
5931         p4d_t *p4d = p4d_offset(pgd, *addr);
5932         pud_t *pud = pud_offset(p4d, *addr);
5933
5934         i_mmap_assert_write_locked(vma->vm_file->f_mapping);
5935         BUG_ON(page_count(virt_to_page(ptep)) == 0);
5936         if (page_count(virt_to_page(ptep)) == 1)
5937                 return 0;
5938
5939         pud_clear(pud);
5940         put_page(virt_to_page(ptep));
5941         mm_dec_nr_pmds(mm);
5942         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5943         return 1;
5944 }
5945
5946 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5947 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5948                       unsigned long addr, pud_t *pud)
5949 {
5950         return NULL;
5951 }
5952
5953 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5954                                 unsigned long *addr, pte_t *ptep)
5955 {
5956         return 0;
5957 }
5958
5959 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5960                                 unsigned long *start, unsigned long *end)
5961 {
5962 }
5963
5964 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5965 {
5966         return false;
5967 }
5968 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5969
5970 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5971 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
5972                         unsigned long addr, unsigned long sz)
5973 {
5974         pgd_t *pgd;
5975         p4d_t *p4d;
5976         pud_t *pud;
5977         pte_t *pte = NULL;
5978
5979         pgd = pgd_offset(mm, addr);
5980         p4d = p4d_alloc(mm, pgd, addr);
5981         if (!p4d)
5982                 return NULL;
5983         pud = pud_alloc(mm, p4d, addr);
5984         if (pud) {
5985                 if (sz == PUD_SIZE) {
5986                         pte = (pte_t *)pud;
5987                 } else {
5988                         BUG_ON(sz != PMD_SIZE);
5989                         if (want_pmd_share(vma, addr) && pud_none(*pud))
5990                                 pte = huge_pmd_share(mm, vma, addr, pud);
5991                         else
5992                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5993                 }
5994         }
5995         BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
5996
5997         return pte;
5998 }
5999
6000 /*
6001  * huge_pte_offset() - Walk the page table to resolve the hugepage
6002  * entry at address @addr
6003  *
6004  * Return: Pointer to page table entry (PUD or PMD) for
6005  * address @addr, or NULL if a !p*d_present() entry is encountered and the
6006  * size @sz doesn't match the hugepage size at this level of the page
6007  * table.
6008  */
6009 pte_t *huge_pte_offset(struct mm_struct *mm,
6010                        unsigned long addr, unsigned long sz)
6011 {
6012         pgd_t *pgd;
6013         p4d_t *p4d;
6014         pud_t *pud;
6015         pmd_t *pmd;
6016
6017         pgd = pgd_offset(mm, addr);
6018         if (!pgd_present(*pgd))
6019                 return NULL;
6020         p4d = p4d_offset(pgd, addr);
6021         if (!p4d_present(*p4d))
6022                 return NULL;
6023
6024         pud = pud_offset(p4d, addr);
6025         if (sz == PUD_SIZE)
6026                 /* must be pud huge, non-present or none */
6027                 return (pte_t *)pud;
6028         if (!pud_present(*pud))
6029                 return NULL;
6030         /* must have a valid entry and size to go further */
6031
6032         pmd = pmd_offset(pud, addr);
6033         /* must be pmd huge, non-present or none */
6034         return (pte_t *)pmd;
6035 }
6036
6037 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
6038
6039 /*
6040  * These functions are overwritable if your architecture needs its own
6041  * behavior.
6042  */
6043 struct page * __weak
6044 follow_huge_addr(struct mm_struct *mm, unsigned long address,
6045                               int write)
6046 {
6047         return ERR_PTR(-EINVAL);
6048 }
6049
6050 struct page * __weak
6051 follow_huge_pd(struct vm_area_struct *vma,
6052                unsigned long address, hugepd_t hpd, int flags, int pdshift)
6053 {
6054         WARN(1, "hugepd follow called with no support for hugepage directory format\n");
6055         return NULL;
6056 }
6057
6058 struct page * __weak
6059 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
6060                 pmd_t *pmd, int flags)
6061 {
6062         struct page *page = NULL;
6063         spinlock_t *ptl;
6064         pte_t pte;
6065
6066         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
6067         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
6068                          (FOLL_PIN | FOLL_GET)))
6069                 return NULL;
6070
6071 retry:
6072         ptl = pmd_lockptr(mm, pmd);
6073         spin_lock(ptl);
6074         /*
6075          * make sure that the address range covered by this pmd is not
6076          * unmapped from other threads.
6077          */
6078         if (!pmd_huge(*pmd))
6079                 goto out;
6080         pte = huge_ptep_get((pte_t *)pmd);
6081         if (pte_present(pte)) {
6082                 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
6083                 /*
6084                  * try_grab_page() should always succeed here, because: a) we
6085                  * hold the pmd (ptl) lock, and b) we've just checked that the
6086                  * huge pmd (head) page is present in the page tables. The ptl
6087                  * prevents the head page and tail pages from being rearranged
6088                  * in any way. So this page must be available at this point,
6089                  * unless the page refcount overflowed:
6090                  */
6091                 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
6092                         page = NULL;
6093                         goto out;
6094                 }
6095         } else {
6096                 if (is_hugetlb_entry_migration(pte)) {
6097                         spin_unlock(ptl);
6098                         __migration_entry_wait(mm, (pte_t *)pmd, ptl);
6099                         goto retry;
6100                 }
6101                 /*
6102                  * hwpoisoned entry is treated as no_page_table in
6103                  * follow_page_mask().
6104                  */
6105         }
6106 out:
6107         spin_unlock(ptl);
6108         return page;
6109 }
6110
6111 struct page * __weak
6112 follow_huge_pud(struct mm_struct *mm, unsigned long address,
6113                 pud_t *pud, int flags)
6114 {
6115         if (flags & (FOLL_GET | FOLL_PIN))
6116                 return NULL;
6117
6118         return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
6119 }
6120
6121 struct page * __weak
6122 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
6123 {
6124         if (flags & (FOLL_GET | FOLL_PIN))
6125                 return NULL;
6126
6127         return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
6128 }
6129
6130 bool isolate_huge_page(struct page *page, struct list_head *list)
6131 {
6132         bool ret = true;
6133
6134         spin_lock_irq(&hugetlb_lock);
6135         if (!PageHeadHuge(page) ||
6136             !HPageMigratable(page) ||
6137             !get_page_unless_zero(page)) {
6138                 ret = false;
6139                 goto unlock;
6140         }
6141         ClearHPageMigratable(page);
6142         list_move_tail(&page->lru, list);
6143 unlock:
6144         spin_unlock_irq(&hugetlb_lock);
6145         return ret;
6146 }
6147
6148 int get_hwpoison_huge_page(struct page *page, bool *hugetlb)
6149 {
6150         int ret = 0;
6151
6152         *hugetlb = false;
6153         spin_lock_irq(&hugetlb_lock);
6154         if (PageHeadHuge(page)) {
6155                 *hugetlb = true;
6156                 if (HPageFreed(page) || HPageMigratable(page))
6157                         ret = get_page_unless_zero(page);
6158                 else
6159                         ret = -EBUSY;
6160         }
6161         spin_unlock_irq(&hugetlb_lock);
6162         return ret;
6163 }
6164
6165 void putback_active_hugepage(struct page *page)
6166 {
6167         spin_lock_irq(&hugetlb_lock);
6168         SetHPageMigratable(page);
6169         list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
6170         spin_unlock_irq(&hugetlb_lock);
6171         put_page(page);
6172 }
6173
6174 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
6175 {
6176         struct hstate *h = page_hstate(oldpage);
6177
6178         hugetlb_cgroup_migrate(oldpage, newpage);
6179         set_page_owner_migrate_reason(newpage, reason);
6180
6181         /*
6182          * transfer temporary state of the new huge page. This is
6183          * reverse to other transitions because the newpage is going to
6184          * be final while the old one will be freed so it takes over
6185          * the temporary status.
6186          *
6187          * Also note that we have to transfer the per-node surplus state
6188          * here as well otherwise the global surplus count will not match
6189          * the per-node's.
6190          */
6191         if (HPageTemporary(newpage)) {
6192                 int old_nid = page_to_nid(oldpage);
6193                 int new_nid = page_to_nid(newpage);
6194
6195                 SetHPageTemporary(oldpage);
6196                 ClearHPageTemporary(newpage);
6197
6198                 /*
6199                  * There is no need to transfer the per-node surplus state
6200                  * when we do not cross the node.
6201                  */
6202                 if (new_nid == old_nid)
6203                         return;
6204                 spin_lock_irq(&hugetlb_lock);
6205                 if (h->surplus_huge_pages_node[old_nid]) {
6206                         h->surplus_huge_pages_node[old_nid]--;
6207                         h->surplus_huge_pages_node[new_nid]++;
6208                 }
6209                 spin_unlock_irq(&hugetlb_lock);
6210         }
6211 }
6212
6213 /*
6214  * This function will unconditionally remove all the shared pmd pgtable entries
6215  * within the specific vma for a hugetlbfs memory range.
6216  */
6217 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
6218 {
6219         struct hstate *h = hstate_vma(vma);
6220         unsigned long sz = huge_page_size(h);
6221         struct mm_struct *mm = vma->vm_mm;
6222         struct mmu_notifier_range range;
6223         unsigned long address, start, end;
6224         spinlock_t *ptl;
6225         pte_t *ptep;
6226
6227         if (!(vma->vm_flags & VM_MAYSHARE))
6228                 return;
6229
6230         start = ALIGN(vma->vm_start, PUD_SIZE);
6231         end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6232
6233         if (start >= end)
6234                 return;
6235
6236         /*
6237          * No need to call adjust_range_if_pmd_sharing_possible(), because
6238          * we have already done the PUD_SIZE alignment.
6239          */
6240         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
6241                                 start, end);
6242         mmu_notifier_invalidate_range_start(&range);
6243         i_mmap_lock_write(vma->vm_file->f_mapping);
6244         for (address = start; address < end; address += PUD_SIZE) {
6245                 unsigned long tmp = address;
6246
6247                 ptep = huge_pte_offset(mm, address, sz);
6248                 if (!ptep)
6249                         continue;
6250                 ptl = huge_pte_lock(h, mm, ptep);
6251                 /* We don't want 'address' to be changed */
6252                 huge_pmd_unshare(mm, vma, &tmp, ptep);
6253                 spin_unlock(ptl);
6254         }
6255         flush_hugetlb_tlb_range(vma, start, end);
6256         i_mmap_unlock_write(vma->vm_file->f_mapping);
6257         /*
6258          * No need to call mmu_notifier_invalidate_range(), see
6259          * Documentation/vm/mmu_notifier.rst.
6260          */
6261         mmu_notifier_invalidate_range_end(&range);
6262 }
6263
6264 #ifdef CONFIG_CMA
6265 static bool cma_reserve_called __initdata;
6266
6267 static int __init cmdline_parse_hugetlb_cma(char *p)
6268 {
6269         hugetlb_cma_size = memparse(p, &p);
6270         return 0;
6271 }
6272
6273 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
6274
6275 void __init hugetlb_cma_reserve(int order)
6276 {
6277         unsigned long size, reserved, per_node;
6278         int nid;
6279
6280         cma_reserve_called = true;
6281
6282         if (!hugetlb_cma_size)
6283                 return;
6284
6285         if (hugetlb_cma_size < (PAGE_SIZE << order)) {
6286                 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
6287                         (PAGE_SIZE << order) / SZ_1M);
6288                 return;
6289         }
6290
6291         /*
6292          * If 3 GB area is requested on a machine with 4 numa nodes,
6293          * let's allocate 1 GB on first three nodes and ignore the last one.
6294          */
6295         per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
6296         pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
6297                 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
6298
6299         reserved = 0;
6300         for_each_node_state(nid, N_ONLINE) {
6301                 int res;
6302                 char name[CMA_MAX_NAME];
6303
6304                 size = min(per_node, hugetlb_cma_size - reserved);
6305                 size = round_up(size, PAGE_SIZE << order);
6306
6307                 snprintf(name, sizeof(name), "hugetlb%d", nid);
6308                 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
6309                                                  0, false, name,
6310                                                  &hugetlb_cma[nid], nid);
6311                 if (res) {
6312                         pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
6313                                 res, nid);
6314                         continue;
6315                 }
6316
6317                 reserved += size;
6318                 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
6319                         size / SZ_1M, nid);
6320
6321                 if (reserved >= hugetlb_cma_size)
6322                         break;
6323         }
6324 }
6325
6326 void __init hugetlb_cma_check(void)
6327 {
6328         if (!hugetlb_cma_size || cma_reserve_called)
6329                 return;
6330
6331         pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
6332 }
6333
6334 #endif /* CONFIG_CMA */