Merge tag 'pci-v5.7-changes' of git://git.kernel.org/pub/scm/linux/kernel/git/helgaas/pci
[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/mmdebug.h>
23 #include <linux/sched/signal.h>
24 #include <linux/rmap.h>
25 #include <linux/string_helpers.h>
26 #include <linux/swap.h>
27 #include <linux/swapops.h>
28 #include <linux/jhash.h>
29 #include <linux/numa.h>
30 #include <linux/llist.h>
31
32 #include <asm/page.h>
33 #include <asm/pgtable.h>
34 #include <asm/tlb.h>
35
36 #include <linux/io.h>
37 #include <linux/hugetlb.h>
38 #include <linux/hugetlb_cgroup.h>
39 #include <linux/node.h>
40 #include <linux/userfaultfd_k.h>
41 #include <linux/page_owner.h>
42 #include "internal.h"
43
44 int hugetlb_max_hstate __read_mostly;
45 unsigned int default_hstate_idx;
46 struct hstate hstates[HUGE_MAX_HSTATE];
47 /*
48  * Minimum page order among possible hugepage sizes, set to a proper value
49  * at boot time.
50  */
51 static unsigned int minimum_order __read_mostly = UINT_MAX;
52
53 __initdata LIST_HEAD(huge_boot_pages);
54
55 /* for command line parsing */
56 static struct hstate * __initdata parsed_hstate;
57 static unsigned long __initdata default_hstate_max_huge_pages;
58 static unsigned long __initdata default_hstate_size;
59 static bool __initdata parsed_valid_hugepagesz = true;
60
61 /*
62  * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
63  * free_huge_pages, and surplus_huge_pages.
64  */
65 DEFINE_SPINLOCK(hugetlb_lock);
66
67 /*
68  * Serializes faults on the same logical page.  This is used to
69  * prevent spurious OOMs when the hugepage pool is fully utilized.
70  */
71 static int num_fault_mutexes;
72 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
73
74 /* Forward declaration */
75 static int hugetlb_acct_memory(struct hstate *h, long delta);
76
77 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
78 {
79         bool free = (spool->count == 0) && (spool->used_hpages == 0);
80
81         spin_unlock(&spool->lock);
82
83         /* If no pages are used, and no other handles to the subpool
84          * remain, give up any reservations mased on minimum size and
85          * free the subpool */
86         if (free) {
87                 if (spool->min_hpages != -1)
88                         hugetlb_acct_memory(spool->hstate,
89                                                 -spool->min_hpages);
90                 kfree(spool);
91         }
92 }
93
94 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
95                                                 long min_hpages)
96 {
97         struct hugepage_subpool *spool;
98
99         spool = kzalloc(sizeof(*spool), GFP_KERNEL);
100         if (!spool)
101                 return NULL;
102
103         spin_lock_init(&spool->lock);
104         spool->count = 1;
105         spool->max_hpages = max_hpages;
106         spool->hstate = h;
107         spool->min_hpages = min_hpages;
108
109         if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
110                 kfree(spool);
111                 return NULL;
112         }
113         spool->rsv_hpages = min_hpages;
114
115         return spool;
116 }
117
118 void hugepage_put_subpool(struct hugepage_subpool *spool)
119 {
120         spin_lock(&spool->lock);
121         BUG_ON(!spool->count);
122         spool->count--;
123         unlock_or_release_subpool(spool);
124 }
125
126 /*
127  * Subpool accounting for allocating and reserving pages.
128  * Return -ENOMEM if there are not enough resources to satisfy the
129  * the request.  Otherwise, return the number of pages by which the
130  * global pools must be adjusted (upward).  The returned value may
131  * only be different than the passed value (delta) in the case where
132  * a subpool minimum size must be manitained.
133  */
134 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
135                                       long delta)
136 {
137         long ret = delta;
138
139         if (!spool)
140                 return ret;
141
142         spin_lock(&spool->lock);
143
144         if (spool->max_hpages != -1) {          /* maximum size accounting */
145                 if ((spool->used_hpages + delta) <= spool->max_hpages)
146                         spool->used_hpages += delta;
147                 else {
148                         ret = -ENOMEM;
149                         goto unlock_ret;
150                 }
151         }
152
153         /* minimum size accounting */
154         if (spool->min_hpages != -1 && spool->rsv_hpages) {
155                 if (delta > spool->rsv_hpages) {
156                         /*
157                          * Asking for more reserves than those already taken on
158                          * behalf of subpool.  Return difference.
159                          */
160                         ret = delta - spool->rsv_hpages;
161                         spool->rsv_hpages = 0;
162                 } else {
163                         ret = 0;        /* reserves already accounted for */
164                         spool->rsv_hpages -= delta;
165                 }
166         }
167
168 unlock_ret:
169         spin_unlock(&spool->lock);
170         return ret;
171 }
172
173 /*
174  * Subpool accounting for freeing and unreserving pages.
175  * Return the number of global page reservations that must be dropped.
176  * The return value may only be different than the passed value (delta)
177  * in the case where a subpool minimum size must be maintained.
178  */
179 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
180                                        long delta)
181 {
182         long ret = delta;
183
184         if (!spool)
185                 return delta;
186
187         spin_lock(&spool->lock);
188
189         if (spool->max_hpages != -1)            /* maximum size accounting */
190                 spool->used_hpages -= delta;
191
192          /* minimum size accounting */
193         if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
194                 if (spool->rsv_hpages + delta <= spool->min_hpages)
195                         ret = 0;
196                 else
197                         ret = spool->rsv_hpages + delta - spool->min_hpages;
198
199                 spool->rsv_hpages += delta;
200                 if (spool->rsv_hpages > spool->min_hpages)
201                         spool->rsv_hpages = spool->min_hpages;
202         }
203
204         /*
205          * If hugetlbfs_put_super couldn't free spool due to an outstanding
206          * quota reference, free it now.
207          */
208         unlock_or_release_subpool(spool);
209
210         return ret;
211 }
212
213 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
214 {
215         return HUGETLBFS_SB(inode->i_sb)->spool;
216 }
217
218 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
219 {
220         return subpool_inode(file_inode(vma->vm_file));
221 }
222
223 /* Helper that removes a struct file_region from the resv_map cache and returns
224  * it for use.
225  */
226 static struct file_region *
227 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
228 {
229         struct file_region *nrg = NULL;
230
231         VM_BUG_ON(resv->region_cache_count <= 0);
232
233         resv->region_cache_count--;
234         nrg = list_first_entry(&resv->region_cache, struct file_region, link);
235         VM_BUG_ON(!nrg);
236         list_del(&nrg->link);
237
238         nrg->from = from;
239         nrg->to = to;
240
241         return nrg;
242 }
243
244 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
245                                               struct file_region *rg)
246 {
247 #ifdef CONFIG_CGROUP_HUGETLB
248         nrg->reservation_counter = rg->reservation_counter;
249         nrg->css = rg->css;
250         if (rg->css)
251                 css_get(rg->css);
252 #endif
253 }
254
255 /* Helper that records hugetlb_cgroup uncharge info. */
256 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
257                                                 struct hstate *h,
258                                                 struct resv_map *resv,
259                                                 struct file_region *nrg)
260 {
261 #ifdef CONFIG_CGROUP_HUGETLB
262         if (h_cg) {
263                 nrg->reservation_counter =
264                         &h_cg->rsvd_hugepage[hstate_index(h)];
265                 nrg->css = &h_cg->css;
266                 if (!resv->pages_per_hpage)
267                         resv->pages_per_hpage = pages_per_huge_page(h);
268                 /* pages_per_hpage should be the same for all entries in
269                  * a resv_map.
270                  */
271                 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
272         } else {
273                 nrg->reservation_counter = NULL;
274                 nrg->css = NULL;
275         }
276 #endif
277 }
278
279 static bool has_same_uncharge_info(struct file_region *rg,
280                                    struct file_region *org)
281 {
282 #ifdef CONFIG_CGROUP_HUGETLB
283         return rg && org &&
284                rg->reservation_counter == org->reservation_counter &&
285                rg->css == org->css;
286
287 #else
288         return true;
289 #endif
290 }
291
292 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
293 {
294         struct file_region *nrg = NULL, *prg = NULL;
295
296         prg = list_prev_entry(rg, link);
297         if (&prg->link != &resv->regions && prg->to == rg->from &&
298             has_same_uncharge_info(prg, rg)) {
299                 prg->to = rg->to;
300
301                 list_del(&rg->link);
302                 kfree(rg);
303
304                 coalesce_file_region(resv, prg);
305                 return;
306         }
307
308         nrg = list_next_entry(rg, link);
309         if (&nrg->link != &resv->regions && nrg->from == rg->to &&
310             has_same_uncharge_info(nrg, rg)) {
311                 nrg->from = rg->from;
312
313                 list_del(&rg->link);
314                 kfree(rg);
315
316                 coalesce_file_region(resv, nrg);
317                 return;
318         }
319 }
320
321 /* Must be called with resv->lock held. Calling this with count_only == true
322  * will count the number of pages to be added but will not modify the linked
323  * list. If regions_needed != NULL and count_only == true, then regions_needed
324  * will indicate the number of file_regions needed in the cache to carry out to
325  * add the regions for this range.
326  */
327 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
328                                      struct hugetlb_cgroup *h_cg,
329                                      struct hstate *h, long *regions_needed,
330                                      bool count_only)
331 {
332         long add = 0;
333         struct list_head *head = &resv->regions;
334         long last_accounted_offset = f;
335         struct file_region *rg = NULL, *trg = NULL, *nrg = NULL;
336
337         if (regions_needed)
338                 *regions_needed = 0;
339
340         /* In this loop, we essentially handle an entry for the range
341          * [last_accounted_offset, rg->from), at every iteration, with some
342          * bounds checking.
343          */
344         list_for_each_entry_safe(rg, trg, head, link) {
345                 /* Skip irrelevant regions that start before our range. */
346                 if (rg->from < f) {
347                         /* If this region ends after the last accounted offset,
348                          * then we need to update last_accounted_offset.
349                          */
350                         if (rg->to > last_accounted_offset)
351                                 last_accounted_offset = rg->to;
352                         continue;
353                 }
354
355                 /* When we find a region that starts beyond our range, we've
356                  * finished.
357                  */
358                 if (rg->from > t)
359                         break;
360
361                 /* Add an entry for last_accounted_offset -> rg->from, and
362                  * update last_accounted_offset.
363                  */
364                 if (rg->from > last_accounted_offset) {
365                         add += rg->from - last_accounted_offset;
366                         if (!count_only) {
367                                 nrg = get_file_region_entry_from_cache(
368                                         resv, last_accounted_offset, rg->from);
369                                 record_hugetlb_cgroup_uncharge_info(h_cg, h,
370                                                                     resv, nrg);
371                                 list_add(&nrg->link, rg->link.prev);
372                                 coalesce_file_region(resv, nrg);
373                         } else if (regions_needed)
374                                 *regions_needed += 1;
375                 }
376
377                 last_accounted_offset = rg->to;
378         }
379
380         /* Handle the case where our range extends beyond
381          * last_accounted_offset.
382          */
383         if (last_accounted_offset < t) {
384                 add += t - last_accounted_offset;
385                 if (!count_only) {
386                         nrg = get_file_region_entry_from_cache(
387                                 resv, last_accounted_offset, t);
388                         record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg);
389                         list_add(&nrg->link, rg->link.prev);
390                         coalesce_file_region(resv, nrg);
391                 } else if (regions_needed)
392                         *regions_needed += 1;
393         }
394
395         VM_BUG_ON(add < 0);
396         return add;
397 }
398
399 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
400  */
401 static int allocate_file_region_entries(struct resv_map *resv,
402                                         int regions_needed)
403         __must_hold(&resv->lock)
404 {
405         struct list_head allocated_regions;
406         int to_allocate = 0, i = 0;
407         struct file_region *trg = NULL, *rg = NULL;
408
409         VM_BUG_ON(regions_needed < 0);
410
411         INIT_LIST_HEAD(&allocated_regions);
412
413         /*
414          * Check for sufficient descriptors in the cache to accommodate
415          * the number of in progress add operations plus regions_needed.
416          *
417          * This is a while loop because when we drop the lock, some other call
418          * to region_add or region_del may have consumed some region_entries,
419          * so we keep looping here until we finally have enough entries for
420          * (adds_in_progress + regions_needed).
421          */
422         while (resv->region_cache_count <
423                (resv->adds_in_progress + regions_needed)) {
424                 to_allocate = resv->adds_in_progress + regions_needed -
425                               resv->region_cache_count;
426
427                 /* At this point, we should have enough entries in the cache
428                  * for all the existings adds_in_progress. We should only be
429                  * needing to allocate for regions_needed.
430                  */
431                 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
432
433                 spin_unlock(&resv->lock);
434                 for (i = 0; i < to_allocate; i++) {
435                         trg = kmalloc(sizeof(*trg), GFP_KERNEL);
436                         if (!trg)
437                                 goto out_of_memory;
438                         list_add(&trg->link, &allocated_regions);
439                 }
440
441                 spin_lock(&resv->lock);
442
443                 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
444                         list_del(&rg->link);
445                         list_add(&rg->link, &resv->region_cache);
446                         resv->region_cache_count++;
447                 }
448         }
449
450         return 0;
451
452 out_of_memory:
453         list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
454                 list_del(&rg->link);
455                 kfree(rg);
456         }
457         return -ENOMEM;
458 }
459
460 /*
461  * Add the huge page range represented by [f, t) to the reserve
462  * map.  Regions will be taken from the cache to fill in this range.
463  * Sufficient regions should exist in the cache due to the previous
464  * call to region_chg with the same range, but in some cases the cache will not
465  * have sufficient entries due to races with other code doing region_add or
466  * region_del.  The extra needed entries will be allocated.
467  *
468  * regions_needed is the out value provided by a previous call to region_chg.
469  *
470  * Return the number of new huge pages added to the map.  This number is greater
471  * than or equal to zero.  If file_region entries needed to be allocated for
472  * this operation and we were not able to allocate, it ruturns -ENOMEM.
473  * region_add of regions of length 1 never allocate file_regions and cannot
474  * fail; region_chg will always allocate at least 1 entry and a region_add for
475  * 1 page will only require at most 1 entry.
476  */
477 static long region_add(struct resv_map *resv, long f, long t,
478                        long in_regions_needed, struct hstate *h,
479                        struct hugetlb_cgroup *h_cg)
480 {
481         long add = 0, actual_regions_needed = 0;
482
483         spin_lock(&resv->lock);
484 retry:
485
486         /* Count how many regions are actually needed to execute this add. */
487         add_reservation_in_range(resv, f, t, NULL, NULL, &actual_regions_needed,
488                                  true);
489
490         /*
491          * Check for sufficient descriptors in the cache to accommodate
492          * this add operation. Note that actual_regions_needed may be greater
493          * than in_regions_needed, as the resv_map may have been modified since
494          * the region_chg call. In this case, we need to make sure that we
495          * allocate extra entries, such that we have enough for all the
496          * existing adds_in_progress, plus the excess needed for this
497          * operation.
498          */
499         if (actual_regions_needed > in_regions_needed &&
500             resv->region_cache_count <
501                     resv->adds_in_progress +
502                             (actual_regions_needed - in_regions_needed)) {
503                 /* region_add operation of range 1 should never need to
504                  * allocate file_region entries.
505                  */
506                 VM_BUG_ON(t - f <= 1);
507
508                 if (allocate_file_region_entries(
509                             resv, actual_regions_needed - in_regions_needed)) {
510                         return -ENOMEM;
511                 }
512
513                 goto retry;
514         }
515
516         add = add_reservation_in_range(resv, f, t, h_cg, h, NULL, false);
517
518         resv->adds_in_progress -= in_regions_needed;
519
520         spin_unlock(&resv->lock);
521         VM_BUG_ON(add < 0);
522         return add;
523 }
524
525 /*
526  * Examine the existing reserve map and determine how many
527  * huge pages in the specified range [f, t) are NOT currently
528  * represented.  This routine is called before a subsequent
529  * call to region_add that will actually modify the reserve
530  * map to add the specified range [f, t).  region_chg does
531  * not change the number of huge pages represented by the
532  * map.  A number of new file_region structures is added to the cache as a
533  * placeholder, for the subsequent region_add call to use. At least 1
534  * file_region structure is added.
535  *
536  * out_regions_needed is the number of regions added to the
537  * resv->adds_in_progress.  This value needs to be provided to a follow up call
538  * to region_add or region_abort for proper accounting.
539  *
540  * Returns the number of huge pages that need to be added to the existing
541  * reservation map for the range [f, t).  This number is greater or equal to
542  * zero.  -ENOMEM is returned if a new file_region structure or cache entry
543  * is needed and can not be allocated.
544  */
545 static long region_chg(struct resv_map *resv, long f, long t,
546                        long *out_regions_needed)
547 {
548         long chg = 0;
549
550         spin_lock(&resv->lock);
551
552         /* Count how many hugepages in this range are NOT respresented. */
553         chg = add_reservation_in_range(resv, f, t, NULL, NULL,
554                                        out_regions_needed, true);
555
556         if (*out_regions_needed == 0)
557                 *out_regions_needed = 1;
558
559         if (allocate_file_region_entries(resv, *out_regions_needed))
560                 return -ENOMEM;
561
562         resv->adds_in_progress += *out_regions_needed;
563
564         spin_unlock(&resv->lock);
565         return chg;
566 }
567
568 /*
569  * Abort the in progress add operation.  The adds_in_progress field
570  * of the resv_map keeps track of the operations in progress between
571  * calls to region_chg and region_add.  Operations are sometimes
572  * aborted after the call to region_chg.  In such cases, region_abort
573  * is called to decrement the adds_in_progress counter. regions_needed
574  * is the value returned by the region_chg call, it is used to decrement
575  * the adds_in_progress counter.
576  *
577  * NOTE: The range arguments [f, t) are not needed or used in this
578  * routine.  They are kept to make reading the calling code easier as
579  * arguments will match the associated region_chg call.
580  */
581 static void region_abort(struct resv_map *resv, long f, long t,
582                          long regions_needed)
583 {
584         spin_lock(&resv->lock);
585         VM_BUG_ON(!resv->region_cache_count);
586         resv->adds_in_progress -= regions_needed;
587         spin_unlock(&resv->lock);
588 }
589
590 /*
591  * Delete the specified range [f, t) from the reserve map.  If the
592  * t parameter is LONG_MAX, this indicates that ALL regions after f
593  * should be deleted.  Locate the regions which intersect [f, t)
594  * and either trim, delete or split the existing regions.
595  *
596  * Returns the number of huge pages deleted from the reserve map.
597  * In the normal case, the return value is zero or more.  In the
598  * case where a region must be split, a new region descriptor must
599  * be allocated.  If the allocation fails, -ENOMEM will be returned.
600  * NOTE: If the parameter t == LONG_MAX, then we will never split
601  * a region and possibly return -ENOMEM.  Callers specifying
602  * t == LONG_MAX do not need to check for -ENOMEM error.
603  */
604 static long region_del(struct resv_map *resv, long f, long t)
605 {
606         struct list_head *head = &resv->regions;
607         struct file_region *rg, *trg;
608         struct file_region *nrg = NULL;
609         long del = 0;
610
611 retry:
612         spin_lock(&resv->lock);
613         list_for_each_entry_safe(rg, trg, head, link) {
614                 /*
615                  * Skip regions before the range to be deleted.  file_region
616                  * ranges are normally of the form [from, to).  However, there
617                  * may be a "placeholder" entry in the map which is of the form
618                  * (from, to) with from == to.  Check for placeholder entries
619                  * at the beginning of the range to be deleted.
620                  */
621                 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
622                         continue;
623
624                 if (rg->from >= t)
625                         break;
626
627                 if (f > rg->from && t < rg->to) { /* Must split region */
628                         /*
629                          * Check for an entry in the cache before dropping
630                          * lock and attempting allocation.
631                          */
632                         if (!nrg &&
633                             resv->region_cache_count > resv->adds_in_progress) {
634                                 nrg = list_first_entry(&resv->region_cache,
635                                                         struct file_region,
636                                                         link);
637                                 list_del(&nrg->link);
638                                 resv->region_cache_count--;
639                         }
640
641                         if (!nrg) {
642                                 spin_unlock(&resv->lock);
643                                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
644                                 if (!nrg)
645                                         return -ENOMEM;
646                                 goto retry;
647                         }
648
649                         del += t - f;
650
651                         /* New entry for end of split region */
652                         nrg->from = t;
653                         nrg->to = rg->to;
654
655                         copy_hugetlb_cgroup_uncharge_info(nrg, rg);
656
657                         INIT_LIST_HEAD(&nrg->link);
658
659                         /* Original entry is trimmed */
660                         rg->to = f;
661
662                         hugetlb_cgroup_uncharge_file_region(
663                                 resv, rg, nrg->to - nrg->from);
664
665                         list_add(&nrg->link, &rg->link);
666                         nrg = NULL;
667                         break;
668                 }
669
670                 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
671                         del += rg->to - rg->from;
672                         hugetlb_cgroup_uncharge_file_region(resv, rg,
673                                                             rg->to - rg->from);
674                         list_del(&rg->link);
675                         kfree(rg);
676                         continue;
677                 }
678
679                 if (f <= rg->from) {    /* Trim beginning of region */
680                         del += t - rg->from;
681                         rg->from = t;
682
683                         hugetlb_cgroup_uncharge_file_region(resv, rg,
684                                                             t - rg->from);
685                 } else {                /* Trim end of region */
686                         del += rg->to - f;
687                         rg->to = f;
688
689                         hugetlb_cgroup_uncharge_file_region(resv, rg,
690                                                             rg->to - f);
691                 }
692         }
693
694         spin_unlock(&resv->lock);
695         kfree(nrg);
696         return del;
697 }
698
699 /*
700  * A rare out of memory error was encountered which prevented removal of
701  * the reserve map region for a page.  The huge page itself was free'ed
702  * and removed from the page cache.  This routine will adjust the subpool
703  * usage count, and the global reserve count if needed.  By incrementing
704  * these counts, the reserve map entry which could not be deleted will
705  * appear as a "reserved" entry instead of simply dangling with incorrect
706  * counts.
707  */
708 void hugetlb_fix_reserve_counts(struct inode *inode)
709 {
710         struct hugepage_subpool *spool = subpool_inode(inode);
711         long rsv_adjust;
712
713         rsv_adjust = hugepage_subpool_get_pages(spool, 1);
714         if (rsv_adjust) {
715                 struct hstate *h = hstate_inode(inode);
716
717                 hugetlb_acct_memory(h, 1);
718         }
719 }
720
721 /*
722  * Count and return the number of huge pages in the reserve map
723  * that intersect with the range [f, t).
724  */
725 static long region_count(struct resv_map *resv, long f, long t)
726 {
727         struct list_head *head = &resv->regions;
728         struct file_region *rg;
729         long chg = 0;
730
731         spin_lock(&resv->lock);
732         /* Locate each segment we overlap with, and count that overlap. */
733         list_for_each_entry(rg, head, link) {
734                 long seg_from;
735                 long seg_to;
736
737                 if (rg->to <= f)
738                         continue;
739                 if (rg->from >= t)
740                         break;
741
742                 seg_from = max(rg->from, f);
743                 seg_to = min(rg->to, t);
744
745                 chg += seg_to - seg_from;
746         }
747         spin_unlock(&resv->lock);
748
749         return chg;
750 }
751
752 /*
753  * Convert the address within this vma to the page offset within
754  * the mapping, in pagecache page units; huge pages here.
755  */
756 static pgoff_t vma_hugecache_offset(struct hstate *h,
757                         struct vm_area_struct *vma, unsigned long address)
758 {
759         return ((address - vma->vm_start) >> huge_page_shift(h)) +
760                         (vma->vm_pgoff >> huge_page_order(h));
761 }
762
763 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
764                                      unsigned long address)
765 {
766         return vma_hugecache_offset(hstate_vma(vma), vma, address);
767 }
768 EXPORT_SYMBOL_GPL(linear_hugepage_index);
769
770 /*
771  * Return the size of the pages allocated when backing a VMA. In the majority
772  * cases this will be same size as used by the page table entries.
773  */
774 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
775 {
776         if (vma->vm_ops && vma->vm_ops->pagesize)
777                 return vma->vm_ops->pagesize(vma);
778         return PAGE_SIZE;
779 }
780 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
781
782 /*
783  * Return the page size being used by the MMU to back a VMA. In the majority
784  * of cases, the page size used by the kernel matches the MMU size. On
785  * architectures where it differs, an architecture-specific 'strong'
786  * version of this symbol is required.
787  */
788 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
789 {
790         return vma_kernel_pagesize(vma);
791 }
792
793 /*
794  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
795  * bits of the reservation map pointer, which are always clear due to
796  * alignment.
797  */
798 #define HPAGE_RESV_OWNER    (1UL << 0)
799 #define HPAGE_RESV_UNMAPPED (1UL << 1)
800 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
801
802 /*
803  * These helpers are used to track how many pages are reserved for
804  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
805  * is guaranteed to have their future faults succeed.
806  *
807  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
808  * the reserve counters are updated with the hugetlb_lock held. It is safe
809  * to reset the VMA at fork() time as it is not in use yet and there is no
810  * chance of the global counters getting corrupted as a result of the values.
811  *
812  * The private mapping reservation is represented in a subtly different
813  * manner to a shared mapping.  A shared mapping has a region map associated
814  * with the underlying file, this region map represents the backing file
815  * pages which have ever had a reservation assigned which this persists even
816  * after the page is instantiated.  A private mapping has a region map
817  * associated with the original mmap which is attached to all VMAs which
818  * reference it, this region map represents those offsets which have consumed
819  * reservation ie. where pages have been instantiated.
820  */
821 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
822 {
823         return (unsigned long)vma->vm_private_data;
824 }
825
826 static void set_vma_private_data(struct vm_area_struct *vma,
827                                                         unsigned long value)
828 {
829         vma->vm_private_data = (void *)value;
830 }
831
832 static void
833 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
834                                           struct hugetlb_cgroup *h_cg,
835                                           struct hstate *h)
836 {
837 #ifdef CONFIG_CGROUP_HUGETLB
838         if (!h_cg || !h) {
839                 resv_map->reservation_counter = NULL;
840                 resv_map->pages_per_hpage = 0;
841                 resv_map->css = NULL;
842         } else {
843                 resv_map->reservation_counter =
844                         &h_cg->rsvd_hugepage[hstate_index(h)];
845                 resv_map->pages_per_hpage = pages_per_huge_page(h);
846                 resv_map->css = &h_cg->css;
847         }
848 #endif
849 }
850
851 struct resv_map *resv_map_alloc(void)
852 {
853         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
854         struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
855
856         if (!resv_map || !rg) {
857                 kfree(resv_map);
858                 kfree(rg);
859                 return NULL;
860         }
861
862         kref_init(&resv_map->refs);
863         spin_lock_init(&resv_map->lock);
864         INIT_LIST_HEAD(&resv_map->regions);
865
866         resv_map->adds_in_progress = 0;
867         /*
868          * Initialize these to 0. On shared mappings, 0's here indicate these
869          * fields don't do cgroup accounting. On private mappings, these will be
870          * re-initialized to the proper values, to indicate that hugetlb cgroup
871          * reservations are to be un-charged from here.
872          */
873         resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
874
875         INIT_LIST_HEAD(&resv_map->region_cache);
876         list_add(&rg->link, &resv_map->region_cache);
877         resv_map->region_cache_count = 1;
878
879         return resv_map;
880 }
881
882 void resv_map_release(struct kref *ref)
883 {
884         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
885         struct list_head *head = &resv_map->region_cache;
886         struct file_region *rg, *trg;
887
888         /* Clear out any active regions before we release the map. */
889         region_del(resv_map, 0, LONG_MAX);
890
891         /* ... and any entries left in the cache */
892         list_for_each_entry_safe(rg, trg, head, link) {
893                 list_del(&rg->link);
894                 kfree(rg);
895         }
896
897         VM_BUG_ON(resv_map->adds_in_progress);
898
899         kfree(resv_map);
900 }
901
902 static inline struct resv_map *inode_resv_map(struct inode *inode)
903 {
904         /*
905          * At inode evict time, i_mapping may not point to the original
906          * address space within the inode.  This original address space
907          * contains the pointer to the resv_map.  So, always use the
908          * address space embedded within the inode.
909          * The VERY common case is inode->mapping == &inode->i_data but,
910          * this may not be true for device special inodes.
911          */
912         return (struct resv_map *)(&inode->i_data)->private_data;
913 }
914
915 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
916 {
917         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
918         if (vma->vm_flags & VM_MAYSHARE) {
919                 struct address_space *mapping = vma->vm_file->f_mapping;
920                 struct inode *inode = mapping->host;
921
922                 return inode_resv_map(inode);
923
924         } else {
925                 return (struct resv_map *)(get_vma_private_data(vma) &
926                                                         ~HPAGE_RESV_MASK);
927         }
928 }
929
930 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
931 {
932         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
933         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
934
935         set_vma_private_data(vma, (get_vma_private_data(vma) &
936                                 HPAGE_RESV_MASK) | (unsigned long)map);
937 }
938
939 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
940 {
941         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
942         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
943
944         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
945 }
946
947 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
948 {
949         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
950
951         return (get_vma_private_data(vma) & flag) != 0;
952 }
953
954 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
955 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
956 {
957         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
958         if (!(vma->vm_flags & VM_MAYSHARE))
959                 vma->vm_private_data = (void *)0;
960 }
961
962 /* Returns true if the VMA has associated reserve pages */
963 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
964 {
965         if (vma->vm_flags & VM_NORESERVE) {
966                 /*
967                  * This address is already reserved by other process(chg == 0),
968                  * so, we should decrement reserved count. Without decrementing,
969                  * reserve count remains after releasing inode, because this
970                  * allocated page will go into page cache and is regarded as
971                  * coming from reserved pool in releasing step.  Currently, we
972                  * don't have any other solution to deal with this situation
973                  * properly, so add work-around here.
974                  */
975                 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
976                         return true;
977                 else
978                         return false;
979         }
980
981         /* Shared mappings always use reserves */
982         if (vma->vm_flags & VM_MAYSHARE) {
983                 /*
984                  * We know VM_NORESERVE is not set.  Therefore, there SHOULD
985                  * be a region map for all pages.  The only situation where
986                  * there is no region map is if a hole was punched via
987                  * fallocate.  In this case, there really are no reverves to
988                  * use.  This situation is indicated if chg != 0.
989                  */
990                 if (chg)
991                         return false;
992                 else
993                         return true;
994         }
995
996         /*
997          * Only the process that called mmap() has reserves for
998          * private mappings.
999          */
1000         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1001                 /*
1002                  * Like the shared case above, a hole punch or truncate
1003                  * could have been performed on the private mapping.
1004                  * Examine the value of chg to determine if reserves
1005                  * actually exist or were previously consumed.
1006                  * Very Subtle - The value of chg comes from a previous
1007                  * call to vma_needs_reserves().  The reserve map for
1008                  * private mappings has different (opposite) semantics
1009                  * than that of shared mappings.  vma_needs_reserves()
1010                  * has already taken this difference in semantics into
1011                  * account.  Therefore, the meaning of chg is the same
1012                  * as in the shared case above.  Code could easily be
1013                  * combined, but keeping it separate draws attention to
1014                  * subtle differences.
1015                  */
1016                 if (chg)
1017                         return false;
1018                 else
1019                         return true;
1020         }
1021
1022         return false;
1023 }
1024
1025 static void enqueue_huge_page(struct hstate *h, struct page *page)
1026 {
1027         int nid = page_to_nid(page);
1028         list_move(&page->lru, &h->hugepage_freelists[nid]);
1029         h->free_huge_pages++;
1030         h->free_huge_pages_node[nid]++;
1031 }
1032
1033 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1034 {
1035         struct page *page;
1036
1037         list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
1038                 if (!PageHWPoison(page))
1039                         break;
1040         /*
1041          * if 'non-isolated free hugepage' not found on the list,
1042          * the allocation fails.
1043          */
1044         if (&h->hugepage_freelists[nid] == &page->lru)
1045                 return NULL;
1046         list_move(&page->lru, &h->hugepage_activelist);
1047         set_page_refcounted(page);
1048         h->free_huge_pages--;
1049         h->free_huge_pages_node[nid]--;
1050         return page;
1051 }
1052
1053 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1054                 nodemask_t *nmask)
1055 {
1056         unsigned int cpuset_mems_cookie;
1057         struct zonelist *zonelist;
1058         struct zone *zone;
1059         struct zoneref *z;
1060         int node = NUMA_NO_NODE;
1061
1062         zonelist = node_zonelist(nid, gfp_mask);
1063
1064 retry_cpuset:
1065         cpuset_mems_cookie = read_mems_allowed_begin();
1066         for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1067                 struct page *page;
1068
1069                 if (!cpuset_zone_allowed(zone, gfp_mask))
1070                         continue;
1071                 /*
1072                  * no need to ask again on the same node. Pool is node rather than
1073                  * zone aware
1074                  */
1075                 if (zone_to_nid(zone) == node)
1076                         continue;
1077                 node = zone_to_nid(zone);
1078
1079                 page = dequeue_huge_page_node_exact(h, node);
1080                 if (page)
1081                         return page;
1082         }
1083         if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1084                 goto retry_cpuset;
1085
1086         return NULL;
1087 }
1088
1089 /* Movability of hugepages depends on migration support. */
1090 static inline gfp_t htlb_alloc_mask(struct hstate *h)
1091 {
1092         if (hugepage_movable_supported(h))
1093                 return GFP_HIGHUSER_MOVABLE;
1094         else
1095                 return GFP_HIGHUSER;
1096 }
1097
1098 static struct page *dequeue_huge_page_vma(struct hstate *h,
1099                                 struct vm_area_struct *vma,
1100                                 unsigned long address, int avoid_reserve,
1101                                 long chg)
1102 {
1103         struct page *page;
1104         struct mempolicy *mpol;
1105         gfp_t gfp_mask;
1106         nodemask_t *nodemask;
1107         int nid;
1108
1109         /*
1110          * A child process with MAP_PRIVATE mappings created by their parent
1111          * have no page reserves. This check ensures that reservations are
1112          * not "stolen". The child may still get SIGKILLed
1113          */
1114         if (!vma_has_reserves(vma, chg) &&
1115                         h->free_huge_pages - h->resv_huge_pages == 0)
1116                 goto err;
1117
1118         /* If reserves cannot be used, ensure enough pages are in the pool */
1119         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1120                 goto err;
1121
1122         gfp_mask = htlb_alloc_mask(h);
1123         nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1124         page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1125         if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1126                 SetPagePrivate(page);
1127                 h->resv_huge_pages--;
1128         }
1129
1130         mpol_cond_put(mpol);
1131         return page;
1132
1133 err:
1134         return NULL;
1135 }
1136
1137 /*
1138  * common helper functions for hstate_next_node_to_{alloc|free}.
1139  * We may have allocated or freed a huge page based on a different
1140  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1141  * be outside of *nodes_allowed.  Ensure that we use an allowed
1142  * node for alloc or free.
1143  */
1144 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1145 {
1146         nid = next_node_in(nid, *nodes_allowed);
1147         VM_BUG_ON(nid >= MAX_NUMNODES);
1148
1149         return nid;
1150 }
1151
1152 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1153 {
1154         if (!node_isset(nid, *nodes_allowed))
1155                 nid = next_node_allowed(nid, nodes_allowed);
1156         return nid;
1157 }
1158
1159 /*
1160  * returns the previously saved node ["this node"] from which to
1161  * allocate a persistent huge page for the pool and advance the
1162  * next node from which to allocate, handling wrap at end of node
1163  * mask.
1164  */
1165 static int hstate_next_node_to_alloc(struct hstate *h,
1166                                         nodemask_t *nodes_allowed)
1167 {
1168         int nid;
1169
1170         VM_BUG_ON(!nodes_allowed);
1171
1172         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1173         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1174
1175         return nid;
1176 }
1177
1178 /*
1179  * helper for free_pool_huge_page() - return the previously saved
1180  * node ["this node"] from which to free a huge page.  Advance the
1181  * next node id whether or not we find a free huge page to free so
1182  * that the next attempt to free addresses the next node.
1183  */
1184 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1185 {
1186         int nid;
1187
1188         VM_BUG_ON(!nodes_allowed);
1189
1190         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1191         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1192
1193         return nid;
1194 }
1195
1196 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask)           \
1197         for (nr_nodes = nodes_weight(*mask);                            \
1198                 nr_nodes > 0 &&                                         \
1199                 ((node = hstate_next_node_to_alloc(hs, mask)) || 1);    \
1200                 nr_nodes--)
1201
1202 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)            \
1203         for (nr_nodes = nodes_weight(*mask);                            \
1204                 nr_nodes > 0 &&                                         \
1205                 ((node = hstate_next_node_to_free(hs, mask)) || 1);     \
1206                 nr_nodes--)
1207
1208 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1209 static void destroy_compound_gigantic_page(struct page *page,
1210                                         unsigned int order)
1211 {
1212         int i;
1213         int nr_pages = 1 << order;
1214         struct page *p = page + 1;
1215
1216         atomic_set(compound_mapcount_ptr(page), 0);
1217         if (hpage_pincount_available(page))
1218                 atomic_set(compound_pincount_ptr(page), 0);
1219
1220         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1221                 clear_compound_head(p);
1222                 set_page_refcounted(p);
1223         }
1224
1225         set_compound_order(page, 0);
1226         __ClearPageHead(page);
1227 }
1228
1229 static void free_gigantic_page(struct page *page, unsigned int order)
1230 {
1231         free_contig_range(page_to_pfn(page), 1 << order);
1232 }
1233
1234 #ifdef CONFIG_CONTIG_ALLOC
1235 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1236                 int nid, nodemask_t *nodemask)
1237 {
1238         unsigned long nr_pages = 1UL << huge_page_order(h);
1239
1240         return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1241 }
1242
1243 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1244 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1245 #else /* !CONFIG_CONTIG_ALLOC */
1246 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1247                                         int nid, nodemask_t *nodemask)
1248 {
1249         return NULL;
1250 }
1251 #endif /* CONFIG_CONTIG_ALLOC */
1252
1253 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1254 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1255                                         int nid, nodemask_t *nodemask)
1256 {
1257         return NULL;
1258 }
1259 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1260 static inline void destroy_compound_gigantic_page(struct page *page,
1261                                                 unsigned int order) { }
1262 #endif
1263
1264 static void update_and_free_page(struct hstate *h, struct page *page)
1265 {
1266         int i;
1267
1268         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1269                 return;
1270
1271         h->nr_huge_pages--;
1272         h->nr_huge_pages_node[page_to_nid(page)]--;
1273         for (i = 0; i < pages_per_huge_page(h); i++) {
1274                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1275                                 1 << PG_referenced | 1 << PG_dirty |
1276                                 1 << PG_active | 1 << PG_private |
1277                                 1 << PG_writeback);
1278         }
1279         VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1280         VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1281         set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1282         set_page_refcounted(page);
1283         if (hstate_is_gigantic(h)) {
1284                 destroy_compound_gigantic_page(page, huge_page_order(h));
1285                 free_gigantic_page(page, huge_page_order(h));
1286         } else {
1287                 __free_pages(page, huge_page_order(h));
1288         }
1289 }
1290
1291 struct hstate *size_to_hstate(unsigned long size)
1292 {
1293         struct hstate *h;
1294
1295         for_each_hstate(h) {
1296                 if (huge_page_size(h) == size)
1297                         return h;
1298         }
1299         return NULL;
1300 }
1301
1302 /*
1303  * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1304  * to hstate->hugepage_activelist.)
1305  *
1306  * This function can be called for tail pages, but never returns true for them.
1307  */
1308 bool page_huge_active(struct page *page)
1309 {
1310         VM_BUG_ON_PAGE(!PageHuge(page), page);
1311         return PageHead(page) && PagePrivate(&page[1]);
1312 }
1313
1314 /* never called for tail page */
1315 static void set_page_huge_active(struct page *page)
1316 {
1317         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1318         SetPagePrivate(&page[1]);
1319 }
1320
1321 static void clear_page_huge_active(struct page *page)
1322 {
1323         VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1324         ClearPagePrivate(&page[1]);
1325 }
1326
1327 /*
1328  * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1329  * code
1330  */
1331 static inline bool PageHugeTemporary(struct page *page)
1332 {
1333         if (!PageHuge(page))
1334                 return false;
1335
1336         return (unsigned long)page[2].mapping == -1U;
1337 }
1338
1339 static inline void SetPageHugeTemporary(struct page *page)
1340 {
1341         page[2].mapping = (void *)-1U;
1342 }
1343
1344 static inline void ClearPageHugeTemporary(struct page *page)
1345 {
1346         page[2].mapping = NULL;
1347 }
1348
1349 static void __free_huge_page(struct page *page)
1350 {
1351         /*
1352          * Can't pass hstate in here because it is called from the
1353          * compound page destructor.
1354          */
1355         struct hstate *h = page_hstate(page);
1356         int nid = page_to_nid(page);
1357         struct hugepage_subpool *spool =
1358                 (struct hugepage_subpool *)page_private(page);
1359         bool restore_reserve;
1360
1361         VM_BUG_ON_PAGE(page_count(page), page);
1362         VM_BUG_ON_PAGE(page_mapcount(page), page);
1363
1364         set_page_private(page, 0);
1365         page->mapping = NULL;
1366         restore_reserve = PagePrivate(page);
1367         ClearPagePrivate(page);
1368
1369         /*
1370          * If PagePrivate() was set on page, page allocation consumed a
1371          * reservation.  If the page was associated with a subpool, there
1372          * would have been a page reserved in the subpool before allocation
1373          * via hugepage_subpool_get_pages().  Since we are 'restoring' the
1374          * reservtion, do not call hugepage_subpool_put_pages() as this will
1375          * remove the reserved page from the subpool.
1376          */
1377         if (!restore_reserve) {
1378                 /*
1379                  * A return code of zero implies that the subpool will be
1380                  * under its minimum size if the reservation is not restored
1381                  * after page is free.  Therefore, force restore_reserve
1382                  * operation.
1383                  */
1384                 if (hugepage_subpool_put_pages(spool, 1) == 0)
1385                         restore_reserve = true;
1386         }
1387
1388         spin_lock(&hugetlb_lock);
1389         clear_page_huge_active(page);
1390         hugetlb_cgroup_uncharge_page(hstate_index(h),
1391                                      pages_per_huge_page(h), page);
1392         hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1393                                           pages_per_huge_page(h), page);
1394         if (restore_reserve)
1395                 h->resv_huge_pages++;
1396
1397         if (PageHugeTemporary(page)) {
1398                 list_del(&page->lru);
1399                 ClearPageHugeTemporary(page);
1400                 update_and_free_page(h, page);
1401         } else if (h->surplus_huge_pages_node[nid]) {
1402                 /* remove the page from active list */
1403                 list_del(&page->lru);
1404                 update_and_free_page(h, page);
1405                 h->surplus_huge_pages--;
1406                 h->surplus_huge_pages_node[nid]--;
1407         } else {
1408                 arch_clear_hugepage_flags(page);
1409                 enqueue_huge_page(h, page);
1410         }
1411         spin_unlock(&hugetlb_lock);
1412 }
1413
1414 /*
1415  * As free_huge_page() can be called from a non-task context, we have
1416  * to defer the actual freeing in a workqueue to prevent potential
1417  * hugetlb_lock deadlock.
1418  *
1419  * free_hpage_workfn() locklessly retrieves the linked list of pages to
1420  * be freed and frees them one-by-one. As the page->mapping pointer is
1421  * going to be cleared in __free_huge_page() anyway, it is reused as the
1422  * llist_node structure of a lockless linked list of huge pages to be freed.
1423  */
1424 static LLIST_HEAD(hpage_freelist);
1425
1426 static void free_hpage_workfn(struct work_struct *work)
1427 {
1428         struct llist_node *node;
1429         struct page *page;
1430
1431         node = llist_del_all(&hpage_freelist);
1432
1433         while (node) {
1434                 page = container_of((struct address_space **)node,
1435                                      struct page, mapping);
1436                 node = node->next;
1437                 __free_huge_page(page);
1438         }
1439 }
1440 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1441
1442 void free_huge_page(struct page *page)
1443 {
1444         /*
1445          * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1446          */
1447         if (!in_task()) {
1448                 /*
1449                  * Only call schedule_work() if hpage_freelist is previously
1450                  * empty. Otherwise, schedule_work() had been called but the
1451                  * workfn hasn't retrieved the list yet.
1452                  */
1453                 if (llist_add((struct llist_node *)&page->mapping,
1454                               &hpage_freelist))
1455                         schedule_work(&free_hpage_work);
1456                 return;
1457         }
1458
1459         __free_huge_page(page);
1460 }
1461
1462 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1463 {
1464         INIT_LIST_HEAD(&page->lru);
1465         set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1466         spin_lock(&hugetlb_lock);
1467         set_hugetlb_cgroup(page, NULL);
1468         set_hugetlb_cgroup_rsvd(page, NULL);
1469         h->nr_huge_pages++;
1470         h->nr_huge_pages_node[nid]++;
1471         spin_unlock(&hugetlb_lock);
1472 }
1473
1474 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1475 {
1476         int i;
1477         int nr_pages = 1 << order;
1478         struct page *p = page + 1;
1479
1480         /* we rely on prep_new_huge_page to set the destructor */
1481         set_compound_order(page, order);
1482         __ClearPageReserved(page);
1483         __SetPageHead(page);
1484         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1485                 /*
1486                  * For gigantic hugepages allocated through bootmem at
1487                  * boot, it's safer to be consistent with the not-gigantic
1488                  * hugepages and clear the PG_reserved bit from all tail pages
1489                  * too.  Otherwse drivers using get_user_pages() to access tail
1490                  * pages may get the reference counting wrong if they see
1491                  * PG_reserved set on a tail page (despite the head page not
1492                  * having PG_reserved set).  Enforcing this consistency between
1493                  * head and tail pages allows drivers to optimize away a check
1494                  * on the head page when they need know if put_page() is needed
1495                  * after get_user_pages().
1496                  */
1497                 __ClearPageReserved(p);
1498                 set_page_count(p, 0);
1499                 set_compound_head(p, page);
1500         }
1501         atomic_set(compound_mapcount_ptr(page), -1);
1502
1503         if (hpage_pincount_available(page))
1504                 atomic_set(compound_pincount_ptr(page), 0);
1505 }
1506
1507 /*
1508  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1509  * transparent huge pages.  See the PageTransHuge() documentation for more
1510  * details.
1511  */
1512 int PageHuge(struct page *page)
1513 {
1514         if (!PageCompound(page))
1515                 return 0;
1516
1517         page = compound_head(page);
1518         return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1519 }
1520 EXPORT_SYMBOL_GPL(PageHuge);
1521
1522 /*
1523  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1524  * normal or transparent huge pages.
1525  */
1526 int PageHeadHuge(struct page *page_head)
1527 {
1528         if (!PageHead(page_head))
1529                 return 0;
1530
1531         return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1532 }
1533
1534 /*
1535  * Find address_space associated with hugetlbfs page.
1536  * Upon entry page is locked and page 'was' mapped although mapped state
1537  * could change.  If necessary, use anon_vma to find vma and associated
1538  * address space.  The returned mapping may be stale, but it can not be
1539  * invalid as page lock (which is held) is required to destroy mapping.
1540  */
1541 static struct address_space *_get_hugetlb_page_mapping(struct page *hpage)
1542 {
1543         struct anon_vma *anon_vma;
1544         pgoff_t pgoff_start, pgoff_end;
1545         struct anon_vma_chain *avc;
1546         struct address_space *mapping = page_mapping(hpage);
1547
1548         /* Simple file based mapping */
1549         if (mapping)
1550                 return mapping;
1551
1552         /*
1553          * Even anonymous hugetlbfs mappings are associated with an
1554          * underlying hugetlbfs file (see hugetlb_file_setup in mmap
1555          * code).  Find a vma associated with the anonymous vma, and
1556          * use the file pointer to get address_space.
1557          */
1558         anon_vma = page_lock_anon_vma_read(hpage);
1559         if (!anon_vma)
1560                 return mapping;  /* NULL */
1561
1562         /* Use first found vma */
1563         pgoff_start = page_to_pgoff(hpage);
1564         pgoff_end = pgoff_start + hpage_nr_pages(hpage) - 1;
1565         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1566                                         pgoff_start, pgoff_end) {
1567                 struct vm_area_struct *vma = avc->vma;
1568
1569                 mapping = vma->vm_file->f_mapping;
1570                 break;
1571         }
1572
1573         anon_vma_unlock_read(anon_vma);
1574         return mapping;
1575 }
1576
1577 /*
1578  * Find and lock address space (mapping) in write mode.
1579  *
1580  * Upon entry, the page is locked which allows us to find the mapping
1581  * even in the case of an anon page.  However, locking order dictates
1582  * the i_mmap_rwsem be acquired BEFORE the page lock.  This is hugetlbfs
1583  * specific.  So, we first try to lock the sema while still holding the
1584  * page lock.  If this works, great!  If not, then we need to drop the
1585  * page lock and then acquire i_mmap_rwsem and reacquire page lock.  Of
1586  * course, need to revalidate state along the way.
1587  */
1588 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1589 {
1590         struct address_space *mapping, *mapping2;
1591
1592         mapping = _get_hugetlb_page_mapping(hpage);
1593 retry:
1594         if (!mapping)
1595                 return mapping;
1596
1597         /*
1598          * If no contention, take lock and return
1599          */
1600         if (i_mmap_trylock_write(mapping))
1601                 return mapping;
1602
1603         /*
1604          * Must drop page lock and wait on mapping sema.
1605          * Note:  Once page lock is dropped, mapping could become invalid.
1606          * As a hack, increase map count until we lock page again.
1607          */
1608         atomic_inc(&hpage->_mapcount);
1609         unlock_page(hpage);
1610         i_mmap_lock_write(mapping);
1611         lock_page(hpage);
1612         atomic_add_negative(-1, &hpage->_mapcount);
1613
1614         /* verify page is still mapped */
1615         if (!page_mapped(hpage)) {
1616                 i_mmap_unlock_write(mapping);
1617                 return NULL;
1618         }
1619
1620         /*
1621          * Get address space again and verify it is the same one
1622          * we locked.  If not, drop lock and retry.
1623          */
1624         mapping2 = _get_hugetlb_page_mapping(hpage);
1625         if (mapping2 != mapping) {
1626                 i_mmap_unlock_write(mapping);
1627                 mapping = mapping2;
1628                 goto retry;
1629         }
1630
1631         return mapping;
1632 }
1633
1634 pgoff_t __basepage_index(struct page *page)
1635 {
1636         struct page *page_head = compound_head(page);
1637         pgoff_t index = page_index(page_head);
1638         unsigned long compound_idx;
1639
1640         if (!PageHuge(page_head))
1641                 return page_index(page);
1642
1643         if (compound_order(page_head) >= MAX_ORDER)
1644                 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1645         else
1646                 compound_idx = page - page_head;
1647
1648         return (index << compound_order(page_head)) + compound_idx;
1649 }
1650
1651 static struct page *alloc_buddy_huge_page(struct hstate *h,
1652                 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1653                 nodemask_t *node_alloc_noretry)
1654 {
1655         int order = huge_page_order(h);
1656         struct page *page;
1657         bool alloc_try_hard = true;
1658
1659         /*
1660          * By default we always try hard to allocate the page with
1661          * __GFP_RETRY_MAYFAIL flag.  However, if we are allocating pages in
1662          * a loop (to adjust global huge page counts) and previous allocation
1663          * failed, do not continue to try hard on the same node.  Use the
1664          * node_alloc_noretry bitmap to manage this state information.
1665          */
1666         if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1667                 alloc_try_hard = false;
1668         gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1669         if (alloc_try_hard)
1670                 gfp_mask |= __GFP_RETRY_MAYFAIL;
1671         if (nid == NUMA_NO_NODE)
1672                 nid = numa_mem_id();
1673         page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1674         if (page)
1675                 __count_vm_event(HTLB_BUDDY_PGALLOC);
1676         else
1677                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1678
1679         /*
1680          * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1681          * indicates an overall state change.  Clear bit so that we resume
1682          * normal 'try hard' allocations.
1683          */
1684         if (node_alloc_noretry && page && !alloc_try_hard)
1685                 node_clear(nid, *node_alloc_noretry);
1686
1687         /*
1688          * If we tried hard to get a page but failed, set bit so that
1689          * subsequent attempts will not try as hard until there is an
1690          * overall state change.
1691          */
1692         if (node_alloc_noretry && !page && alloc_try_hard)
1693                 node_set(nid, *node_alloc_noretry);
1694
1695         return page;
1696 }
1697
1698 /*
1699  * Common helper to allocate a fresh hugetlb page. All specific allocators
1700  * should use this function to get new hugetlb pages
1701  */
1702 static struct page *alloc_fresh_huge_page(struct hstate *h,
1703                 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1704                 nodemask_t *node_alloc_noretry)
1705 {
1706         struct page *page;
1707
1708         if (hstate_is_gigantic(h))
1709                 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1710         else
1711                 page = alloc_buddy_huge_page(h, gfp_mask,
1712                                 nid, nmask, node_alloc_noretry);
1713         if (!page)
1714                 return NULL;
1715
1716         if (hstate_is_gigantic(h))
1717                 prep_compound_gigantic_page(page, huge_page_order(h));
1718         prep_new_huge_page(h, page, page_to_nid(page));
1719
1720         return page;
1721 }
1722
1723 /*
1724  * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1725  * manner.
1726  */
1727 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1728                                 nodemask_t *node_alloc_noretry)
1729 {
1730         struct page *page;
1731         int nr_nodes, node;
1732         gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1733
1734         for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1735                 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1736                                                 node_alloc_noretry);
1737                 if (page)
1738                         break;
1739         }
1740
1741         if (!page)
1742                 return 0;
1743
1744         put_page(page); /* free it into the hugepage allocator */
1745
1746         return 1;
1747 }
1748
1749 /*
1750  * Free huge page from pool from next node to free.
1751  * Attempt to keep persistent huge pages more or less
1752  * balanced over allowed nodes.
1753  * Called with hugetlb_lock locked.
1754  */
1755 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1756                                                          bool acct_surplus)
1757 {
1758         int nr_nodes, node;
1759         int ret = 0;
1760
1761         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1762                 /*
1763                  * If we're returning unused surplus pages, only examine
1764                  * nodes with surplus pages.
1765                  */
1766                 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1767                     !list_empty(&h->hugepage_freelists[node])) {
1768                         struct page *page =
1769                                 list_entry(h->hugepage_freelists[node].next,
1770                                           struct page, lru);
1771                         list_del(&page->lru);
1772                         h->free_huge_pages--;
1773                         h->free_huge_pages_node[node]--;
1774                         if (acct_surplus) {
1775                                 h->surplus_huge_pages--;
1776                                 h->surplus_huge_pages_node[node]--;
1777                         }
1778                         update_and_free_page(h, page);
1779                         ret = 1;
1780                         break;
1781                 }
1782         }
1783
1784         return ret;
1785 }
1786
1787 /*
1788  * Dissolve a given free hugepage into free buddy pages. This function does
1789  * nothing for in-use hugepages and non-hugepages.
1790  * This function returns values like below:
1791  *
1792  *  -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1793  *          (allocated or reserved.)
1794  *       0: successfully dissolved free hugepages or the page is not a
1795  *          hugepage (considered as already dissolved)
1796  */
1797 int dissolve_free_huge_page(struct page *page)
1798 {
1799         int rc = -EBUSY;
1800
1801         /* Not to disrupt normal path by vainly holding hugetlb_lock */
1802         if (!PageHuge(page))
1803                 return 0;
1804
1805         spin_lock(&hugetlb_lock);
1806         if (!PageHuge(page)) {
1807                 rc = 0;
1808                 goto out;
1809         }
1810
1811         if (!page_count(page)) {
1812                 struct page *head = compound_head(page);
1813                 struct hstate *h = page_hstate(head);
1814                 int nid = page_to_nid(head);
1815                 if (h->free_huge_pages - h->resv_huge_pages == 0)
1816                         goto out;
1817                 /*
1818                  * Move PageHWPoison flag from head page to the raw error page,
1819                  * which makes any subpages rather than the error page reusable.
1820                  */
1821                 if (PageHWPoison(head) && page != head) {
1822                         SetPageHWPoison(page);
1823                         ClearPageHWPoison(head);
1824                 }
1825                 list_del(&head->lru);
1826                 h->free_huge_pages--;
1827                 h->free_huge_pages_node[nid]--;
1828                 h->max_huge_pages--;
1829                 update_and_free_page(h, head);
1830                 rc = 0;
1831         }
1832 out:
1833         spin_unlock(&hugetlb_lock);
1834         return rc;
1835 }
1836
1837 /*
1838  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1839  * make specified memory blocks removable from the system.
1840  * Note that this will dissolve a free gigantic hugepage completely, if any
1841  * part of it lies within the given range.
1842  * Also note that if dissolve_free_huge_page() returns with an error, all
1843  * free hugepages that were dissolved before that error are lost.
1844  */
1845 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1846 {
1847         unsigned long pfn;
1848         struct page *page;
1849         int rc = 0;
1850
1851         if (!hugepages_supported())
1852                 return rc;
1853
1854         for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1855                 page = pfn_to_page(pfn);
1856                 rc = dissolve_free_huge_page(page);
1857                 if (rc)
1858                         break;
1859         }
1860
1861         return rc;
1862 }
1863
1864 /*
1865  * Allocates a fresh surplus page from the page allocator.
1866  */
1867 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1868                 int nid, nodemask_t *nmask)
1869 {
1870         struct page *page = NULL;
1871
1872         if (hstate_is_gigantic(h))
1873                 return NULL;
1874
1875         spin_lock(&hugetlb_lock);
1876         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1877                 goto out_unlock;
1878         spin_unlock(&hugetlb_lock);
1879
1880         page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1881         if (!page)
1882                 return NULL;
1883
1884         spin_lock(&hugetlb_lock);
1885         /*
1886          * We could have raced with the pool size change.
1887          * Double check that and simply deallocate the new page
1888          * if we would end up overcommiting the surpluses. Abuse
1889          * temporary page to workaround the nasty free_huge_page
1890          * codeflow
1891          */
1892         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1893                 SetPageHugeTemporary(page);
1894                 spin_unlock(&hugetlb_lock);
1895                 put_page(page);
1896                 return NULL;
1897         } else {
1898                 h->surplus_huge_pages++;
1899                 h->surplus_huge_pages_node[page_to_nid(page)]++;
1900         }
1901
1902 out_unlock:
1903         spin_unlock(&hugetlb_lock);
1904
1905         return page;
1906 }
1907
1908 struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1909                                      int nid, nodemask_t *nmask)
1910 {
1911         struct page *page;
1912
1913         if (hstate_is_gigantic(h))
1914                 return NULL;
1915
1916         page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1917         if (!page)
1918                 return NULL;
1919
1920         /*
1921          * We do not account these pages as surplus because they are only
1922          * temporary and will be released properly on the last reference
1923          */
1924         SetPageHugeTemporary(page);
1925
1926         return page;
1927 }
1928
1929 /*
1930  * Use the VMA's mpolicy to allocate a huge page from the buddy.
1931  */
1932 static
1933 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1934                 struct vm_area_struct *vma, unsigned long addr)
1935 {
1936         struct page *page;
1937         struct mempolicy *mpol;
1938         gfp_t gfp_mask = htlb_alloc_mask(h);
1939         int nid;
1940         nodemask_t *nodemask;
1941
1942         nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1943         page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1944         mpol_cond_put(mpol);
1945
1946         return page;
1947 }
1948
1949 /* page migration callback function */
1950 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1951 {
1952         gfp_t gfp_mask = htlb_alloc_mask(h);
1953         struct page *page = NULL;
1954
1955         if (nid != NUMA_NO_NODE)
1956                 gfp_mask |= __GFP_THISNODE;
1957
1958         spin_lock(&hugetlb_lock);
1959         if (h->free_huge_pages - h->resv_huge_pages > 0)
1960                 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1961         spin_unlock(&hugetlb_lock);
1962
1963         if (!page)
1964                 page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1965
1966         return page;
1967 }
1968
1969 /* page migration callback function */
1970 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1971                 nodemask_t *nmask)
1972 {
1973         gfp_t gfp_mask = htlb_alloc_mask(h);
1974
1975         spin_lock(&hugetlb_lock);
1976         if (h->free_huge_pages - h->resv_huge_pages > 0) {
1977                 struct page *page;
1978
1979                 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1980                 if (page) {
1981                         spin_unlock(&hugetlb_lock);
1982                         return page;
1983                 }
1984         }
1985         spin_unlock(&hugetlb_lock);
1986
1987         return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1988 }
1989
1990 /* mempolicy aware migration callback */
1991 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1992                 unsigned long address)
1993 {
1994         struct mempolicy *mpol;
1995         nodemask_t *nodemask;
1996         struct page *page;
1997         gfp_t gfp_mask;
1998         int node;
1999
2000         gfp_mask = htlb_alloc_mask(h);
2001         node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2002         page = alloc_huge_page_nodemask(h, node, nodemask);
2003         mpol_cond_put(mpol);
2004
2005         return page;
2006 }
2007
2008 /*
2009  * Increase the hugetlb pool such that it can accommodate a reservation
2010  * of size 'delta'.
2011  */
2012 static int gather_surplus_pages(struct hstate *h, int delta)
2013 {
2014         struct list_head surplus_list;
2015         struct page *page, *tmp;
2016         int ret, i;
2017         int needed, allocated;
2018         bool alloc_ok = true;
2019
2020         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2021         if (needed <= 0) {
2022                 h->resv_huge_pages += delta;
2023                 return 0;
2024         }
2025
2026         allocated = 0;
2027         INIT_LIST_HEAD(&surplus_list);
2028
2029         ret = -ENOMEM;
2030 retry:
2031         spin_unlock(&hugetlb_lock);
2032         for (i = 0; i < needed; i++) {
2033                 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2034                                 NUMA_NO_NODE, NULL);
2035                 if (!page) {
2036                         alloc_ok = false;
2037                         break;
2038                 }
2039                 list_add(&page->lru, &surplus_list);
2040                 cond_resched();
2041         }
2042         allocated += i;
2043
2044         /*
2045          * After retaking hugetlb_lock, we need to recalculate 'needed'
2046          * because either resv_huge_pages or free_huge_pages may have changed.
2047          */
2048         spin_lock(&hugetlb_lock);
2049         needed = (h->resv_huge_pages + delta) -
2050                         (h->free_huge_pages + allocated);
2051         if (needed > 0) {
2052                 if (alloc_ok)
2053                         goto retry;
2054                 /*
2055                  * We were not able to allocate enough pages to
2056                  * satisfy the entire reservation so we free what
2057                  * we've allocated so far.
2058                  */
2059                 goto free;
2060         }
2061         /*
2062          * The surplus_list now contains _at_least_ the number of extra pages
2063          * needed to accommodate the reservation.  Add the appropriate number
2064          * of pages to the hugetlb pool and free the extras back to the buddy
2065          * allocator.  Commit the entire reservation here to prevent another
2066          * process from stealing the pages as they are added to the pool but
2067          * before they are reserved.
2068          */
2069         needed += allocated;
2070         h->resv_huge_pages += delta;
2071         ret = 0;
2072
2073         /* Free the needed pages to the hugetlb pool */
2074         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2075                 if ((--needed) < 0)
2076                         break;
2077                 /*
2078                  * This page is now managed by the hugetlb allocator and has
2079                  * no users -- drop the buddy allocator's reference.
2080                  */
2081                 put_page_testzero(page);
2082                 VM_BUG_ON_PAGE(page_count(page), page);
2083                 enqueue_huge_page(h, page);
2084         }
2085 free:
2086         spin_unlock(&hugetlb_lock);
2087
2088         /* Free unnecessary surplus pages to the buddy allocator */
2089         list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2090                 put_page(page);
2091         spin_lock(&hugetlb_lock);
2092
2093         return ret;
2094 }
2095
2096 /*
2097  * This routine has two main purposes:
2098  * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2099  *    in unused_resv_pages.  This corresponds to the prior adjustments made
2100  *    to the associated reservation map.
2101  * 2) Free any unused surplus pages that may have been allocated to satisfy
2102  *    the reservation.  As many as unused_resv_pages may be freed.
2103  *
2104  * Called with hugetlb_lock held.  However, the lock could be dropped (and
2105  * reacquired) during calls to cond_resched_lock.  Whenever dropping the lock,
2106  * we must make sure nobody else can claim pages we are in the process of
2107  * freeing.  Do this by ensuring resv_huge_page always is greater than the
2108  * number of huge pages we plan to free when dropping the lock.
2109  */
2110 static void return_unused_surplus_pages(struct hstate *h,
2111                                         unsigned long unused_resv_pages)
2112 {
2113         unsigned long nr_pages;
2114
2115         /* Cannot return gigantic pages currently */
2116         if (hstate_is_gigantic(h))
2117                 goto out;
2118
2119         /*
2120          * Part (or even all) of the reservation could have been backed
2121          * by pre-allocated pages. Only free surplus pages.
2122          */
2123         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2124
2125         /*
2126          * We want to release as many surplus pages as possible, spread
2127          * evenly across all nodes with memory. Iterate across these nodes
2128          * until we can no longer free unreserved surplus pages. This occurs
2129          * when the nodes with surplus pages have no free pages.
2130          * free_pool_huge_page() will balance the the freed pages across the
2131          * on-line nodes with memory and will handle the hstate accounting.
2132          *
2133          * Note that we decrement resv_huge_pages as we free the pages.  If
2134          * we drop the lock, resv_huge_pages will still be sufficiently large
2135          * to cover subsequent pages we may free.
2136          */
2137         while (nr_pages--) {
2138                 h->resv_huge_pages--;
2139                 unused_resv_pages--;
2140                 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
2141                         goto out;
2142                 cond_resched_lock(&hugetlb_lock);
2143         }
2144
2145 out:
2146         /* Fully uncommit the reservation */
2147         h->resv_huge_pages -= unused_resv_pages;
2148 }
2149
2150
2151 /*
2152  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2153  * are used by the huge page allocation routines to manage reservations.
2154  *
2155  * vma_needs_reservation is called to determine if the huge page at addr
2156  * within the vma has an associated reservation.  If a reservation is
2157  * needed, the value 1 is returned.  The caller is then responsible for
2158  * managing the global reservation and subpool usage counts.  After
2159  * the huge page has been allocated, vma_commit_reservation is called
2160  * to add the page to the reservation map.  If the page allocation fails,
2161  * the reservation must be ended instead of committed.  vma_end_reservation
2162  * is called in such cases.
2163  *
2164  * In the normal case, vma_commit_reservation returns the same value
2165  * as the preceding vma_needs_reservation call.  The only time this
2166  * is not the case is if a reserve map was changed between calls.  It
2167  * is the responsibility of the caller to notice the difference and
2168  * take appropriate action.
2169  *
2170  * vma_add_reservation is used in error paths where a reservation must
2171  * be restored when a newly allocated huge page must be freed.  It is
2172  * to be called after calling vma_needs_reservation to determine if a
2173  * reservation exists.
2174  */
2175 enum vma_resv_mode {
2176         VMA_NEEDS_RESV,
2177         VMA_COMMIT_RESV,
2178         VMA_END_RESV,
2179         VMA_ADD_RESV,
2180 };
2181 static long __vma_reservation_common(struct hstate *h,
2182                                 struct vm_area_struct *vma, unsigned long addr,
2183                                 enum vma_resv_mode mode)
2184 {
2185         struct resv_map *resv;
2186         pgoff_t idx;
2187         long ret;
2188         long dummy_out_regions_needed;
2189
2190         resv = vma_resv_map(vma);
2191         if (!resv)
2192                 return 1;
2193
2194         idx = vma_hugecache_offset(h, vma, addr);
2195         switch (mode) {
2196         case VMA_NEEDS_RESV:
2197                 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2198                 /* We assume that vma_reservation_* routines always operate on
2199                  * 1 page, and that adding to resv map a 1 page entry can only
2200                  * ever require 1 region.
2201                  */
2202                 VM_BUG_ON(dummy_out_regions_needed != 1);
2203                 break;
2204         case VMA_COMMIT_RESV:
2205                 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2206                 /* region_add calls of range 1 should never fail. */
2207                 VM_BUG_ON(ret < 0);
2208                 break;
2209         case VMA_END_RESV:
2210                 region_abort(resv, idx, idx + 1, 1);
2211                 ret = 0;
2212                 break;
2213         case VMA_ADD_RESV:
2214                 if (vma->vm_flags & VM_MAYSHARE) {
2215                         ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2216                         /* region_add calls of range 1 should never fail. */
2217                         VM_BUG_ON(ret < 0);
2218                 } else {
2219                         region_abort(resv, idx, idx + 1, 1);
2220                         ret = region_del(resv, idx, idx + 1);
2221                 }
2222                 break;
2223         default:
2224                 BUG();
2225         }
2226
2227         if (vma->vm_flags & VM_MAYSHARE)
2228                 return ret;
2229         else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
2230                 /*
2231                  * In most cases, reserves always exist for private mappings.
2232                  * However, a file associated with mapping could have been
2233                  * hole punched or truncated after reserves were consumed.
2234                  * As subsequent fault on such a range will not use reserves.
2235                  * Subtle - The reserve map for private mappings has the
2236                  * opposite meaning than that of shared mappings.  If NO
2237                  * entry is in the reserve map, it means a reservation exists.
2238                  * If an entry exists in the reserve map, it means the
2239                  * reservation has already been consumed.  As a result, the
2240                  * return value of this routine is the opposite of the
2241                  * value returned from reserve map manipulation routines above.
2242                  */
2243                 if (ret)
2244                         return 0;
2245                 else
2246                         return 1;
2247         }
2248         else
2249                 return ret < 0 ? ret : 0;
2250 }
2251
2252 static long vma_needs_reservation(struct hstate *h,
2253                         struct vm_area_struct *vma, unsigned long addr)
2254 {
2255         return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2256 }
2257
2258 static long vma_commit_reservation(struct hstate *h,
2259                         struct vm_area_struct *vma, unsigned long addr)
2260 {
2261         return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2262 }
2263
2264 static void vma_end_reservation(struct hstate *h,
2265                         struct vm_area_struct *vma, unsigned long addr)
2266 {
2267         (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2268 }
2269
2270 static long vma_add_reservation(struct hstate *h,
2271                         struct vm_area_struct *vma, unsigned long addr)
2272 {
2273         return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2274 }
2275
2276 /*
2277  * This routine is called to restore a reservation on error paths.  In the
2278  * specific error paths, a huge page was allocated (via alloc_huge_page)
2279  * and is about to be freed.  If a reservation for the page existed,
2280  * alloc_huge_page would have consumed the reservation and set PagePrivate
2281  * in the newly allocated page.  When the page is freed via free_huge_page,
2282  * the global reservation count will be incremented if PagePrivate is set.
2283  * However, free_huge_page can not adjust the reserve map.  Adjust the
2284  * reserve map here to be consistent with global reserve count adjustments
2285  * to be made by free_huge_page.
2286  */
2287 static void restore_reserve_on_error(struct hstate *h,
2288                         struct vm_area_struct *vma, unsigned long address,
2289                         struct page *page)
2290 {
2291         if (unlikely(PagePrivate(page))) {
2292                 long rc = vma_needs_reservation(h, vma, address);
2293
2294                 if (unlikely(rc < 0)) {
2295                         /*
2296                          * Rare out of memory condition in reserve map
2297                          * manipulation.  Clear PagePrivate so that
2298                          * global reserve count will not be incremented
2299                          * by free_huge_page.  This will make it appear
2300                          * as though the reservation for this page was
2301                          * consumed.  This may prevent the task from
2302                          * faulting in the page at a later time.  This
2303                          * is better than inconsistent global huge page
2304                          * accounting of reserve counts.
2305                          */
2306                         ClearPagePrivate(page);
2307                 } else if (rc) {
2308                         rc = vma_add_reservation(h, vma, address);
2309                         if (unlikely(rc < 0))
2310                                 /*
2311                                  * See above comment about rare out of
2312                                  * memory condition.
2313                                  */
2314                                 ClearPagePrivate(page);
2315                 } else
2316                         vma_end_reservation(h, vma, address);
2317         }
2318 }
2319
2320 struct page *alloc_huge_page(struct vm_area_struct *vma,
2321                                     unsigned long addr, int avoid_reserve)
2322 {
2323         struct hugepage_subpool *spool = subpool_vma(vma);
2324         struct hstate *h = hstate_vma(vma);
2325         struct page *page;
2326         long map_chg, map_commit;
2327         long gbl_chg;
2328         int ret, idx;
2329         struct hugetlb_cgroup *h_cg;
2330         bool deferred_reserve;
2331
2332         idx = hstate_index(h);
2333         /*
2334          * Examine the region/reserve map to determine if the process
2335          * has a reservation for the page to be allocated.  A return
2336          * code of zero indicates a reservation exists (no change).
2337          */
2338         map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2339         if (map_chg < 0)
2340                 return ERR_PTR(-ENOMEM);
2341
2342         /*
2343          * Processes that did not create the mapping will have no
2344          * reserves as indicated by the region/reserve map. Check
2345          * that the allocation will not exceed the subpool limit.
2346          * Allocations for MAP_NORESERVE mappings also need to be
2347          * checked against any subpool limit.
2348          */
2349         if (map_chg || avoid_reserve) {
2350                 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2351                 if (gbl_chg < 0) {
2352                         vma_end_reservation(h, vma, addr);
2353                         return ERR_PTR(-ENOSPC);
2354                 }
2355
2356                 /*
2357                  * Even though there was no reservation in the region/reserve
2358                  * map, there could be reservations associated with the
2359                  * subpool that can be used.  This would be indicated if the
2360                  * return value of hugepage_subpool_get_pages() is zero.
2361                  * However, if avoid_reserve is specified we still avoid even
2362                  * the subpool reservations.
2363                  */
2364                 if (avoid_reserve)
2365                         gbl_chg = 1;
2366         }
2367
2368         /* If this allocation is not consuming a reservation, charge it now.
2369          */
2370         deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma);
2371         if (deferred_reserve) {
2372                 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2373                         idx, pages_per_huge_page(h), &h_cg);
2374                 if (ret)
2375                         goto out_subpool_put;
2376         }
2377
2378         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2379         if (ret)
2380                 goto out_uncharge_cgroup_reservation;
2381
2382         spin_lock(&hugetlb_lock);
2383         /*
2384          * glb_chg is passed to indicate whether or not a page must be taken
2385          * from the global free pool (global change).  gbl_chg == 0 indicates
2386          * a reservation exists for the allocation.
2387          */
2388         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2389         if (!page) {
2390                 spin_unlock(&hugetlb_lock);
2391                 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2392                 if (!page)
2393                         goto out_uncharge_cgroup;
2394                 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2395                         SetPagePrivate(page);
2396                         h->resv_huge_pages--;
2397                 }
2398                 spin_lock(&hugetlb_lock);
2399                 list_move(&page->lru, &h->hugepage_activelist);
2400                 /* Fall through */
2401         }
2402         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2403         /* If allocation is not consuming a reservation, also store the
2404          * hugetlb_cgroup pointer on the page.
2405          */
2406         if (deferred_reserve) {
2407                 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2408                                                   h_cg, page);
2409         }
2410
2411         spin_unlock(&hugetlb_lock);
2412
2413         set_page_private(page, (unsigned long)spool);
2414
2415         map_commit = vma_commit_reservation(h, vma, addr);
2416         if (unlikely(map_chg > map_commit)) {
2417                 /*
2418                  * The page was added to the reservation map between
2419                  * vma_needs_reservation and vma_commit_reservation.
2420                  * This indicates a race with hugetlb_reserve_pages.
2421                  * Adjust for the subpool count incremented above AND
2422                  * in hugetlb_reserve_pages for the same page.  Also,
2423                  * the reservation count added in hugetlb_reserve_pages
2424                  * no longer applies.
2425                  */
2426                 long rsv_adjust;
2427
2428                 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2429                 hugetlb_acct_memory(h, -rsv_adjust);
2430         }
2431         return page;
2432
2433 out_uncharge_cgroup:
2434         hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2435 out_uncharge_cgroup_reservation:
2436         if (deferred_reserve)
2437                 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2438                                                     h_cg);
2439 out_subpool_put:
2440         if (map_chg || avoid_reserve)
2441                 hugepage_subpool_put_pages(spool, 1);
2442         vma_end_reservation(h, vma, addr);
2443         return ERR_PTR(-ENOSPC);
2444 }
2445
2446 int alloc_bootmem_huge_page(struct hstate *h)
2447         __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2448 int __alloc_bootmem_huge_page(struct hstate *h)
2449 {
2450         struct huge_bootmem_page *m;
2451         int nr_nodes, node;
2452
2453         for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2454                 void *addr;
2455
2456                 addr = memblock_alloc_try_nid_raw(
2457                                 huge_page_size(h), huge_page_size(h),
2458                                 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2459                 if (addr) {
2460                         /*
2461                          * Use the beginning of the huge page to store the
2462                          * huge_bootmem_page struct (until gather_bootmem
2463                          * puts them into the mem_map).
2464                          */
2465                         m = addr;
2466                         goto found;
2467                 }
2468         }
2469         return 0;
2470
2471 found:
2472         BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2473         /* Put them into a private list first because mem_map is not up yet */
2474         INIT_LIST_HEAD(&m->list);
2475         list_add(&m->list, &huge_boot_pages);
2476         m->hstate = h;
2477         return 1;
2478 }
2479
2480 static void __init prep_compound_huge_page(struct page *page,
2481                 unsigned int order)
2482 {
2483         if (unlikely(order > (MAX_ORDER - 1)))
2484                 prep_compound_gigantic_page(page, order);
2485         else
2486                 prep_compound_page(page, order);
2487 }
2488
2489 /* Put bootmem huge pages into the standard lists after mem_map is up */
2490 static void __init gather_bootmem_prealloc(void)
2491 {
2492         struct huge_bootmem_page *m;
2493
2494         list_for_each_entry(m, &huge_boot_pages, list) {
2495                 struct page *page = virt_to_page(m);
2496                 struct hstate *h = m->hstate;
2497
2498                 WARN_ON(page_count(page) != 1);
2499                 prep_compound_huge_page(page, h->order);
2500                 WARN_ON(PageReserved(page));
2501                 prep_new_huge_page(h, page, page_to_nid(page));
2502                 put_page(page); /* free it into the hugepage allocator */
2503
2504                 /*
2505                  * If we had gigantic hugepages allocated at boot time, we need
2506                  * to restore the 'stolen' pages to totalram_pages in order to
2507                  * fix confusing memory reports from free(1) and another
2508                  * side-effects, like CommitLimit going negative.
2509                  */
2510                 if (hstate_is_gigantic(h))
2511                         adjust_managed_page_count(page, 1 << h->order);
2512                 cond_resched();
2513         }
2514 }
2515
2516 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2517 {
2518         unsigned long i;
2519         nodemask_t *node_alloc_noretry;
2520
2521         if (!hstate_is_gigantic(h)) {
2522                 /*
2523                  * Bit mask controlling how hard we retry per-node allocations.
2524                  * Ignore errors as lower level routines can deal with
2525                  * node_alloc_noretry == NULL.  If this kmalloc fails at boot
2526                  * time, we are likely in bigger trouble.
2527                  */
2528                 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2529                                                 GFP_KERNEL);
2530         } else {
2531                 /* allocations done at boot time */
2532                 node_alloc_noretry = NULL;
2533         }
2534
2535         /* bit mask controlling how hard we retry per-node allocations */
2536         if (node_alloc_noretry)
2537                 nodes_clear(*node_alloc_noretry);
2538
2539         for (i = 0; i < h->max_huge_pages; ++i) {
2540                 if (hstate_is_gigantic(h)) {
2541                         if (!alloc_bootmem_huge_page(h))
2542                                 break;
2543                 } else if (!alloc_pool_huge_page(h,
2544                                          &node_states[N_MEMORY],
2545                                          node_alloc_noretry))
2546                         break;
2547                 cond_resched();
2548         }
2549         if (i < h->max_huge_pages) {
2550                 char buf[32];
2551
2552                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2553                 pr_warn("HugeTLB: allocating %lu of page size %s failed.  Only allocated %lu hugepages.\n",
2554                         h->max_huge_pages, buf, i);
2555                 h->max_huge_pages = i;
2556         }
2557
2558         kfree(node_alloc_noretry);
2559 }
2560
2561 static void __init hugetlb_init_hstates(void)
2562 {
2563         struct hstate *h;
2564
2565         for_each_hstate(h) {
2566                 if (minimum_order > huge_page_order(h))
2567                         minimum_order = huge_page_order(h);
2568
2569                 /* oversize hugepages were init'ed in early boot */
2570                 if (!hstate_is_gigantic(h))
2571                         hugetlb_hstate_alloc_pages(h);
2572         }
2573         VM_BUG_ON(minimum_order == UINT_MAX);
2574 }
2575
2576 static void __init report_hugepages(void)
2577 {
2578         struct hstate *h;
2579
2580         for_each_hstate(h) {
2581                 char buf[32];
2582
2583                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2584                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2585                         buf, h->free_huge_pages);
2586         }
2587 }
2588
2589 #ifdef CONFIG_HIGHMEM
2590 static void try_to_free_low(struct hstate *h, unsigned long count,
2591                                                 nodemask_t *nodes_allowed)
2592 {
2593         int i;
2594
2595         if (hstate_is_gigantic(h))
2596                 return;
2597
2598         for_each_node_mask(i, *nodes_allowed) {
2599                 struct page *page, *next;
2600                 struct list_head *freel = &h->hugepage_freelists[i];
2601                 list_for_each_entry_safe(page, next, freel, lru) {
2602                         if (count >= h->nr_huge_pages)
2603                                 return;
2604                         if (PageHighMem(page))
2605                                 continue;
2606                         list_del(&page->lru);
2607                         update_and_free_page(h, page);
2608                         h->free_huge_pages--;
2609                         h->free_huge_pages_node[page_to_nid(page)]--;
2610                 }
2611         }
2612 }
2613 #else
2614 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2615                                                 nodemask_t *nodes_allowed)
2616 {
2617 }
2618 #endif
2619
2620 /*
2621  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
2622  * balanced by operating on them in a round-robin fashion.
2623  * Returns 1 if an adjustment was made.
2624  */
2625 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2626                                 int delta)
2627 {
2628         int nr_nodes, node;
2629
2630         VM_BUG_ON(delta != -1 && delta != 1);
2631
2632         if (delta < 0) {
2633                 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2634                         if (h->surplus_huge_pages_node[node])
2635                                 goto found;
2636                 }
2637         } else {
2638                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2639                         if (h->surplus_huge_pages_node[node] <
2640                                         h->nr_huge_pages_node[node])
2641                                 goto found;
2642                 }
2643         }
2644         return 0;
2645
2646 found:
2647         h->surplus_huge_pages += delta;
2648         h->surplus_huge_pages_node[node] += delta;
2649         return 1;
2650 }
2651
2652 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2653 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2654                               nodemask_t *nodes_allowed)
2655 {
2656         unsigned long min_count, ret;
2657         NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2658
2659         /*
2660          * Bit mask controlling how hard we retry per-node allocations.
2661          * If we can not allocate the bit mask, do not attempt to allocate
2662          * the requested huge pages.
2663          */
2664         if (node_alloc_noretry)
2665                 nodes_clear(*node_alloc_noretry);
2666         else
2667                 return -ENOMEM;
2668
2669         spin_lock(&hugetlb_lock);
2670
2671         /*
2672          * Check for a node specific request.
2673          * Changing node specific huge page count may require a corresponding
2674          * change to the global count.  In any case, the passed node mask
2675          * (nodes_allowed) will restrict alloc/free to the specified node.
2676          */
2677         if (nid != NUMA_NO_NODE) {
2678                 unsigned long old_count = count;
2679
2680                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2681                 /*
2682                  * User may have specified a large count value which caused the
2683                  * above calculation to overflow.  In this case, they wanted
2684                  * to allocate as many huge pages as possible.  Set count to
2685                  * largest possible value to align with their intention.
2686                  */
2687                 if (count < old_count)
2688                         count = ULONG_MAX;
2689         }
2690
2691         /*
2692          * Gigantic pages runtime allocation depend on the capability for large
2693          * page range allocation.
2694          * If the system does not provide this feature, return an error when
2695          * the user tries to allocate gigantic pages but let the user free the
2696          * boottime allocated gigantic pages.
2697          */
2698         if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2699                 if (count > persistent_huge_pages(h)) {
2700                         spin_unlock(&hugetlb_lock);
2701                         NODEMASK_FREE(node_alloc_noretry);
2702                         return -EINVAL;
2703                 }
2704                 /* Fall through to decrease pool */
2705         }
2706
2707         /*
2708          * Increase the pool size
2709          * First take pages out of surplus state.  Then make up the
2710          * remaining difference by allocating fresh huge pages.
2711          *
2712          * We might race with alloc_surplus_huge_page() here and be unable
2713          * to convert a surplus huge page to a normal huge page. That is
2714          * not critical, though, it just means the overall size of the
2715          * pool might be one hugepage larger than it needs to be, but
2716          * within all the constraints specified by the sysctls.
2717          */
2718         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2719                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2720                         break;
2721         }
2722
2723         while (count > persistent_huge_pages(h)) {
2724                 /*
2725                  * If this allocation races such that we no longer need the
2726                  * page, free_huge_page will handle it by freeing the page
2727                  * and reducing the surplus.
2728                  */
2729                 spin_unlock(&hugetlb_lock);
2730
2731                 /* yield cpu to avoid soft lockup */
2732                 cond_resched();
2733
2734                 ret = alloc_pool_huge_page(h, nodes_allowed,
2735                                                 node_alloc_noretry);
2736                 spin_lock(&hugetlb_lock);
2737                 if (!ret)
2738                         goto out;
2739
2740                 /* Bail for signals. Probably ctrl-c from user */
2741                 if (signal_pending(current))
2742                         goto out;
2743         }
2744
2745         /*
2746          * Decrease the pool size
2747          * First return free pages to the buddy allocator (being careful
2748          * to keep enough around to satisfy reservations).  Then place
2749          * pages into surplus state as needed so the pool will shrink
2750          * to the desired size as pages become free.
2751          *
2752          * By placing pages into the surplus state independent of the
2753          * overcommit value, we are allowing the surplus pool size to
2754          * exceed overcommit. There are few sane options here. Since
2755          * alloc_surplus_huge_page() is checking the global counter,
2756          * though, we'll note that we're not allowed to exceed surplus
2757          * and won't grow the pool anywhere else. Not until one of the
2758          * sysctls are changed, or the surplus pages go out of use.
2759          */
2760         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2761         min_count = max(count, min_count);
2762         try_to_free_low(h, min_count, nodes_allowed);
2763         while (min_count < persistent_huge_pages(h)) {
2764                 if (!free_pool_huge_page(h, nodes_allowed, 0))
2765                         break;
2766                 cond_resched_lock(&hugetlb_lock);
2767         }
2768         while (count < persistent_huge_pages(h)) {
2769                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2770                         break;
2771         }
2772 out:
2773         h->max_huge_pages = persistent_huge_pages(h);
2774         spin_unlock(&hugetlb_lock);
2775
2776         NODEMASK_FREE(node_alloc_noretry);
2777
2778         return 0;
2779 }
2780
2781 #define HSTATE_ATTR_RO(_name) \
2782         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2783
2784 #define HSTATE_ATTR(_name) \
2785         static struct kobj_attribute _name##_attr = \
2786                 __ATTR(_name, 0644, _name##_show, _name##_store)
2787
2788 static struct kobject *hugepages_kobj;
2789 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2790
2791 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2792
2793 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2794 {
2795         int i;
2796
2797         for (i = 0; i < HUGE_MAX_HSTATE; i++)
2798                 if (hstate_kobjs[i] == kobj) {
2799                         if (nidp)
2800                                 *nidp = NUMA_NO_NODE;
2801                         return &hstates[i];
2802                 }
2803
2804         return kobj_to_node_hstate(kobj, nidp);
2805 }
2806
2807 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2808                                         struct kobj_attribute *attr, char *buf)
2809 {
2810         struct hstate *h;
2811         unsigned long nr_huge_pages;
2812         int nid;
2813
2814         h = kobj_to_hstate(kobj, &nid);
2815         if (nid == NUMA_NO_NODE)
2816                 nr_huge_pages = h->nr_huge_pages;
2817         else
2818                 nr_huge_pages = h->nr_huge_pages_node[nid];
2819
2820         return sprintf(buf, "%lu\n", nr_huge_pages);
2821 }
2822
2823 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2824                                            struct hstate *h, int nid,
2825                                            unsigned long count, size_t len)
2826 {
2827         int err;
2828         nodemask_t nodes_allowed, *n_mask;
2829
2830         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2831                 return -EINVAL;
2832
2833         if (nid == NUMA_NO_NODE) {
2834                 /*
2835                  * global hstate attribute
2836                  */
2837                 if (!(obey_mempolicy &&
2838                                 init_nodemask_of_mempolicy(&nodes_allowed)))
2839                         n_mask = &node_states[N_MEMORY];
2840                 else
2841                         n_mask = &nodes_allowed;
2842         } else {
2843                 /*
2844                  * Node specific request.  count adjustment happens in
2845                  * set_max_huge_pages() after acquiring hugetlb_lock.
2846                  */
2847                 init_nodemask_of_node(&nodes_allowed, nid);
2848                 n_mask = &nodes_allowed;
2849         }
2850
2851         err = set_max_huge_pages(h, count, nid, n_mask);
2852
2853         return err ? err : len;
2854 }
2855
2856 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2857                                          struct kobject *kobj, const char *buf,
2858                                          size_t len)
2859 {
2860         struct hstate *h;
2861         unsigned long count;
2862         int nid;
2863         int err;
2864
2865         err = kstrtoul(buf, 10, &count);
2866         if (err)
2867                 return err;
2868
2869         h = kobj_to_hstate(kobj, &nid);
2870         return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2871 }
2872
2873 static ssize_t nr_hugepages_show(struct kobject *kobj,
2874                                        struct kobj_attribute *attr, char *buf)
2875 {
2876         return nr_hugepages_show_common(kobj, attr, buf);
2877 }
2878
2879 static ssize_t nr_hugepages_store(struct kobject *kobj,
2880                struct kobj_attribute *attr, const char *buf, size_t len)
2881 {
2882         return nr_hugepages_store_common(false, kobj, buf, len);
2883 }
2884 HSTATE_ATTR(nr_hugepages);
2885
2886 #ifdef CONFIG_NUMA
2887
2888 /*
2889  * hstate attribute for optionally mempolicy-based constraint on persistent
2890  * huge page alloc/free.
2891  */
2892 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2893                                        struct kobj_attribute *attr, char *buf)
2894 {
2895         return nr_hugepages_show_common(kobj, attr, buf);
2896 }
2897
2898 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2899                struct kobj_attribute *attr, const char *buf, size_t len)
2900 {
2901         return nr_hugepages_store_common(true, kobj, buf, len);
2902 }
2903 HSTATE_ATTR(nr_hugepages_mempolicy);
2904 #endif
2905
2906
2907 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2908                                         struct kobj_attribute *attr, char *buf)
2909 {
2910         struct hstate *h = kobj_to_hstate(kobj, NULL);
2911         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2912 }
2913
2914 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2915                 struct kobj_attribute *attr, const char *buf, size_t count)
2916 {
2917         int err;
2918         unsigned long input;
2919         struct hstate *h = kobj_to_hstate(kobj, NULL);
2920
2921         if (hstate_is_gigantic(h))
2922                 return -EINVAL;
2923
2924         err = kstrtoul(buf, 10, &input);
2925         if (err)
2926                 return err;
2927
2928         spin_lock(&hugetlb_lock);
2929         h->nr_overcommit_huge_pages = input;
2930         spin_unlock(&hugetlb_lock);
2931
2932         return count;
2933 }
2934 HSTATE_ATTR(nr_overcommit_hugepages);
2935
2936 static ssize_t free_hugepages_show(struct kobject *kobj,
2937                                         struct kobj_attribute *attr, char *buf)
2938 {
2939         struct hstate *h;
2940         unsigned long free_huge_pages;
2941         int nid;
2942
2943         h = kobj_to_hstate(kobj, &nid);
2944         if (nid == NUMA_NO_NODE)
2945                 free_huge_pages = h->free_huge_pages;
2946         else
2947                 free_huge_pages = h->free_huge_pages_node[nid];
2948
2949         return sprintf(buf, "%lu\n", free_huge_pages);
2950 }
2951 HSTATE_ATTR_RO(free_hugepages);
2952
2953 static ssize_t resv_hugepages_show(struct kobject *kobj,
2954                                         struct kobj_attribute *attr, char *buf)
2955 {
2956         struct hstate *h = kobj_to_hstate(kobj, NULL);
2957         return sprintf(buf, "%lu\n", h->resv_huge_pages);
2958 }
2959 HSTATE_ATTR_RO(resv_hugepages);
2960
2961 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2962                                         struct kobj_attribute *attr, char *buf)
2963 {
2964         struct hstate *h;
2965         unsigned long surplus_huge_pages;
2966         int nid;
2967
2968         h = kobj_to_hstate(kobj, &nid);
2969         if (nid == NUMA_NO_NODE)
2970                 surplus_huge_pages = h->surplus_huge_pages;
2971         else
2972                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2973
2974         return sprintf(buf, "%lu\n", surplus_huge_pages);
2975 }
2976 HSTATE_ATTR_RO(surplus_hugepages);
2977
2978 static struct attribute *hstate_attrs[] = {
2979         &nr_hugepages_attr.attr,
2980         &nr_overcommit_hugepages_attr.attr,
2981         &free_hugepages_attr.attr,
2982         &resv_hugepages_attr.attr,
2983         &surplus_hugepages_attr.attr,
2984 #ifdef CONFIG_NUMA
2985         &nr_hugepages_mempolicy_attr.attr,
2986 #endif
2987         NULL,
2988 };
2989
2990 static const struct attribute_group hstate_attr_group = {
2991         .attrs = hstate_attrs,
2992 };
2993
2994 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2995                                     struct kobject **hstate_kobjs,
2996                                     const struct attribute_group *hstate_attr_group)
2997 {
2998         int retval;
2999         int hi = hstate_index(h);
3000
3001         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3002         if (!hstate_kobjs[hi])
3003                 return -ENOMEM;
3004
3005         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3006         if (retval)
3007                 kobject_put(hstate_kobjs[hi]);
3008
3009         return retval;
3010 }
3011
3012 static void __init hugetlb_sysfs_init(void)
3013 {
3014         struct hstate *h;
3015         int err;
3016
3017         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3018         if (!hugepages_kobj)
3019                 return;
3020
3021         for_each_hstate(h) {
3022                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3023                                          hstate_kobjs, &hstate_attr_group);
3024                 if (err)
3025                         pr_err("Hugetlb: Unable to add hstate %s", h->name);
3026         }
3027 }
3028
3029 #ifdef CONFIG_NUMA
3030
3031 /*
3032  * node_hstate/s - associate per node hstate attributes, via their kobjects,
3033  * with node devices in node_devices[] using a parallel array.  The array
3034  * index of a node device or _hstate == node id.
3035  * This is here to avoid any static dependency of the node device driver, in
3036  * the base kernel, on the hugetlb module.
3037  */
3038 struct node_hstate {
3039         struct kobject          *hugepages_kobj;
3040         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
3041 };
3042 static struct node_hstate node_hstates[MAX_NUMNODES];
3043
3044 /*
3045  * A subset of global hstate attributes for node devices
3046  */
3047 static struct attribute *per_node_hstate_attrs[] = {
3048         &nr_hugepages_attr.attr,
3049         &free_hugepages_attr.attr,
3050         &surplus_hugepages_attr.attr,
3051         NULL,
3052 };
3053
3054 static const struct attribute_group per_node_hstate_attr_group = {
3055         .attrs = per_node_hstate_attrs,
3056 };
3057
3058 /*
3059  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3060  * Returns node id via non-NULL nidp.
3061  */
3062 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3063 {
3064         int nid;
3065
3066         for (nid = 0; nid < nr_node_ids; nid++) {
3067                 struct node_hstate *nhs = &node_hstates[nid];
3068                 int i;
3069                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3070                         if (nhs->hstate_kobjs[i] == kobj) {
3071                                 if (nidp)
3072                                         *nidp = nid;
3073                                 return &hstates[i];
3074                         }
3075         }
3076
3077         BUG();
3078         return NULL;
3079 }
3080
3081 /*
3082  * Unregister hstate attributes from a single node device.
3083  * No-op if no hstate attributes attached.
3084  */
3085 static void hugetlb_unregister_node(struct node *node)
3086 {
3087         struct hstate *h;
3088         struct node_hstate *nhs = &node_hstates[node->dev.id];
3089
3090         if (!nhs->hugepages_kobj)
3091                 return;         /* no hstate attributes */
3092
3093         for_each_hstate(h) {
3094                 int idx = hstate_index(h);
3095                 if (nhs->hstate_kobjs[idx]) {
3096                         kobject_put(nhs->hstate_kobjs[idx]);
3097                         nhs->hstate_kobjs[idx] = NULL;
3098                 }
3099         }
3100
3101         kobject_put(nhs->hugepages_kobj);
3102         nhs->hugepages_kobj = NULL;
3103 }
3104
3105
3106 /*
3107  * Register hstate attributes for a single node device.
3108  * No-op if attributes already registered.
3109  */
3110 static void hugetlb_register_node(struct node *node)
3111 {
3112         struct hstate *h;
3113         struct node_hstate *nhs = &node_hstates[node->dev.id];
3114         int err;
3115
3116         if (nhs->hugepages_kobj)
3117                 return;         /* already allocated */
3118
3119         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3120                                                         &node->dev.kobj);
3121         if (!nhs->hugepages_kobj)
3122                 return;
3123
3124         for_each_hstate(h) {
3125                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3126                                                 nhs->hstate_kobjs,
3127                                                 &per_node_hstate_attr_group);
3128                 if (err) {
3129                         pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
3130                                 h->name, node->dev.id);
3131                         hugetlb_unregister_node(node);
3132                         break;
3133                 }
3134         }
3135 }
3136
3137 /*
3138  * hugetlb init time:  register hstate attributes for all registered node
3139  * devices of nodes that have memory.  All on-line nodes should have
3140  * registered their associated device by this time.
3141  */
3142 static void __init hugetlb_register_all_nodes(void)
3143 {
3144         int nid;
3145
3146         for_each_node_state(nid, N_MEMORY) {
3147                 struct node *node = node_devices[nid];
3148                 if (node->dev.id == nid)
3149                         hugetlb_register_node(node);
3150         }
3151
3152         /*
3153          * Let the node device driver know we're here so it can
3154          * [un]register hstate attributes on node hotplug.
3155          */
3156         register_hugetlbfs_with_node(hugetlb_register_node,
3157                                      hugetlb_unregister_node);
3158 }
3159 #else   /* !CONFIG_NUMA */
3160
3161 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3162 {
3163         BUG();
3164         if (nidp)
3165                 *nidp = -1;
3166         return NULL;
3167 }
3168
3169 static void hugetlb_register_all_nodes(void) { }
3170
3171 #endif
3172
3173 static int __init hugetlb_init(void)
3174 {
3175         int i;
3176
3177         if (!hugepages_supported())
3178                 return 0;
3179
3180         if (!size_to_hstate(default_hstate_size)) {
3181                 if (default_hstate_size != 0) {
3182                         pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
3183                                default_hstate_size, HPAGE_SIZE);
3184                 }
3185
3186                 default_hstate_size = HPAGE_SIZE;
3187                 if (!size_to_hstate(default_hstate_size))
3188                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3189         }
3190         default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
3191         if (default_hstate_max_huge_pages) {
3192                 if (!default_hstate.max_huge_pages)
3193                         default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3194         }
3195
3196         hugetlb_init_hstates();
3197         gather_bootmem_prealloc();
3198         report_hugepages();
3199
3200         hugetlb_sysfs_init();
3201         hugetlb_register_all_nodes();
3202         hugetlb_cgroup_file_init();
3203
3204 #ifdef CONFIG_SMP
3205         num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3206 #else
3207         num_fault_mutexes = 1;
3208 #endif
3209         hugetlb_fault_mutex_table =
3210                 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3211                               GFP_KERNEL);
3212         BUG_ON(!hugetlb_fault_mutex_table);
3213
3214         for (i = 0; i < num_fault_mutexes; i++)
3215                 mutex_init(&hugetlb_fault_mutex_table[i]);
3216         return 0;
3217 }
3218 subsys_initcall(hugetlb_init);
3219
3220 /* Should be called on processing a hugepagesz=... option */
3221 void __init hugetlb_bad_size(void)
3222 {
3223         parsed_valid_hugepagesz = false;
3224 }
3225
3226 void __init hugetlb_add_hstate(unsigned int order)
3227 {
3228         struct hstate *h;
3229         unsigned long i;
3230
3231         if (size_to_hstate(PAGE_SIZE << order)) {
3232                 pr_warn("hugepagesz= specified twice, ignoring\n");
3233                 return;
3234         }
3235         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3236         BUG_ON(order == 0);
3237         h = &hstates[hugetlb_max_hstate++];
3238         h->order = order;
3239         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
3240         h->nr_huge_pages = 0;
3241         h->free_huge_pages = 0;
3242         for (i = 0; i < MAX_NUMNODES; ++i)
3243                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3244         INIT_LIST_HEAD(&h->hugepage_activelist);
3245         h->next_nid_to_alloc = first_memory_node;
3246         h->next_nid_to_free = first_memory_node;
3247         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3248                                         huge_page_size(h)/1024);
3249
3250         parsed_hstate = h;
3251 }
3252
3253 static int __init hugetlb_nrpages_setup(char *s)
3254 {
3255         unsigned long *mhp;
3256         static unsigned long *last_mhp;
3257
3258         if (!parsed_valid_hugepagesz) {
3259                 pr_warn("hugepages = %s preceded by "
3260                         "an unsupported hugepagesz, ignoring\n", s);
3261                 parsed_valid_hugepagesz = true;
3262                 return 1;
3263         }
3264         /*
3265          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
3266          * so this hugepages= parameter goes to the "default hstate".
3267          */
3268         else if (!hugetlb_max_hstate)
3269                 mhp = &default_hstate_max_huge_pages;
3270         else
3271                 mhp = &parsed_hstate->max_huge_pages;
3272
3273         if (mhp == last_mhp) {
3274                 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
3275                 return 1;
3276         }
3277
3278         if (sscanf(s, "%lu", mhp) <= 0)
3279                 *mhp = 0;
3280
3281         /*
3282          * Global state is always initialized later in hugetlb_init.
3283          * But we need to allocate >= MAX_ORDER hstates here early to still
3284          * use the bootmem allocator.
3285          */
3286         if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
3287                 hugetlb_hstate_alloc_pages(parsed_hstate);
3288
3289         last_mhp = mhp;
3290
3291         return 1;
3292 }
3293 __setup("hugepages=", hugetlb_nrpages_setup);
3294
3295 static int __init hugetlb_default_setup(char *s)
3296 {
3297         default_hstate_size = memparse(s, &s);
3298         return 1;
3299 }
3300 __setup("default_hugepagesz=", hugetlb_default_setup);
3301
3302 static unsigned int cpuset_mems_nr(unsigned int *array)
3303 {
3304         int node;
3305         unsigned int nr = 0;
3306
3307         for_each_node_mask(node, cpuset_current_mems_allowed)
3308                 nr += array[node];
3309
3310         return nr;
3311 }
3312
3313 #ifdef CONFIG_SYSCTL
3314 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3315                          struct ctl_table *table, int write,
3316                          void __user *buffer, size_t *length, loff_t *ppos)
3317 {
3318         struct hstate *h = &default_hstate;
3319         unsigned long tmp = h->max_huge_pages;
3320         int ret;
3321
3322         if (!hugepages_supported())
3323                 return -EOPNOTSUPP;
3324
3325         table->data = &tmp;
3326         table->maxlen = sizeof(unsigned long);
3327         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3328         if (ret)
3329                 goto out;
3330
3331         if (write)
3332                 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3333                                                   NUMA_NO_NODE, tmp, *length);
3334 out:
3335         return ret;
3336 }
3337
3338 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3339                           void __user *buffer, size_t *length, loff_t *ppos)
3340 {
3341
3342         return hugetlb_sysctl_handler_common(false, table, write,
3343                                                         buffer, length, ppos);
3344 }
3345
3346 #ifdef CONFIG_NUMA
3347 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3348                           void __user *buffer, size_t *length, loff_t *ppos)
3349 {
3350         return hugetlb_sysctl_handler_common(true, table, write,
3351                                                         buffer, length, ppos);
3352 }
3353 #endif /* CONFIG_NUMA */
3354
3355 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3356                         void __user *buffer,
3357                         size_t *length, loff_t *ppos)
3358 {
3359         struct hstate *h = &default_hstate;
3360         unsigned long tmp;
3361         int ret;
3362
3363         if (!hugepages_supported())
3364                 return -EOPNOTSUPP;
3365
3366         tmp = h->nr_overcommit_huge_pages;
3367
3368         if (write && hstate_is_gigantic(h))
3369                 return -EINVAL;
3370
3371         table->data = &tmp;
3372         table->maxlen = sizeof(unsigned long);
3373         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3374         if (ret)
3375                 goto out;
3376
3377         if (write) {
3378                 spin_lock(&hugetlb_lock);
3379                 h->nr_overcommit_huge_pages = tmp;
3380                 spin_unlock(&hugetlb_lock);
3381         }
3382 out:
3383         return ret;
3384 }
3385
3386 #endif /* CONFIG_SYSCTL */
3387
3388 void hugetlb_report_meminfo(struct seq_file *m)
3389 {
3390         struct hstate *h;
3391         unsigned long total = 0;
3392
3393         if (!hugepages_supported())
3394                 return;
3395
3396         for_each_hstate(h) {
3397                 unsigned long count = h->nr_huge_pages;
3398
3399                 total += (PAGE_SIZE << huge_page_order(h)) * count;
3400
3401                 if (h == &default_hstate)
3402                         seq_printf(m,
3403                                    "HugePages_Total:   %5lu\n"
3404                                    "HugePages_Free:    %5lu\n"
3405                                    "HugePages_Rsvd:    %5lu\n"
3406                                    "HugePages_Surp:    %5lu\n"
3407                                    "Hugepagesize:   %8lu kB\n",
3408                                    count,
3409                                    h->free_huge_pages,
3410                                    h->resv_huge_pages,
3411                                    h->surplus_huge_pages,
3412                                    (PAGE_SIZE << huge_page_order(h)) / 1024);
3413         }
3414
3415         seq_printf(m, "Hugetlb:        %8lu kB\n", total / 1024);
3416 }
3417
3418 int hugetlb_report_node_meminfo(int nid, char *buf)
3419 {
3420         struct hstate *h = &default_hstate;
3421         if (!hugepages_supported())
3422                 return 0;
3423         return sprintf(buf,
3424                 "Node %d HugePages_Total: %5u\n"
3425                 "Node %d HugePages_Free:  %5u\n"
3426                 "Node %d HugePages_Surp:  %5u\n",
3427                 nid, h->nr_huge_pages_node[nid],
3428                 nid, h->free_huge_pages_node[nid],
3429                 nid, h->surplus_huge_pages_node[nid]);
3430 }
3431
3432 void hugetlb_show_meminfo(void)
3433 {
3434         struct hstate *h;
3435         int nid;
3436
3437         if (!hugepages_supported())
3438                 return;
3439
3440         for_each_node_state(nid, N_MEMORY)
3441                 for_each_hstate(h)
3442                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3443                                 nid,
3444                                 h->nr_huge_pages_node[nid],
3445                                 h->free_huge_pages_node[nid],
3446                                 h->surplus_huge_pages_node[nid],
3447                                 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3448 }
3449
3450 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3451 {
3452         seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3453                    atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3454 }
3455
3456 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3457 unsigned long hugetlb_total_pages(void)
3458 {
3459         struct hstate *h;
3460         unsigned long nr_total_pages = 0;
3461
3462         for_each_hstate(h)
3463                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3464         return nr_total_pages;
3465 }
3466
3467 static int hugetlb_acct_memory(struct hstate *h, long delta)
3468 {
3469         int ret = -ENOMEM;
3470
3471         spin_lock(&hugetlb_lock);
3472         /*
3473          * When cpuset is configured, it breaks the strict hugetlb page
3474          * reservation as the accounting is done on a global variable. Such
3475          * reservation is completely rubbish in the presence of cpuset because
3476          * the reservation is not checked against page availability for the
3477          * current cpuset. Application can still potentially OOM'ed by kernel
3478          * with lack of free htlb page in cpuset that the task is in.
3479          * Attempt to enforce strict accounting with cpuset is almost
3480          * impossible (or too ugly) because cpuset is too fluid that
3481          * task or memory node can be dynamically moved between cpusets.
3482          *
3483          * The change of semantics for shared hugetlb mapping with cpuset is
3484          * undesirable. However, in order to preserve some of the semantics,
3485          * we fall back to check against current free page availability as
3486          * a best attempt and hopefully to minimize the impact of changing
3487          * semantics that cpuset has.
3488          */
3489         if (delta > 0) {
3490                 if (gather_surplus_pages(h, delta) < 0)
3491                         goto out;
3492
3493                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3494                         return_unused_surplus_pages(h, delta);
3495                         goto out;
3496                 }
3497         }
3498
3499         ret = 0;
3500         if (delta < 0)
3501                 return_unused_surplus_pages(h, (unsigned long) -delta);
3502
3503 out:
3504         spin_unlock(&hugetlb_lock);
3505         return ret;
3506 }
3507
3508 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3509 {
3510         struct resv_map *resv = vma_resv_map(vma);
3511
3512         /*
3513          * This new VMA should share its siblings reservation map if present.
3514          * The VMA will only ever have a valid reservation map pointer where
3515          * it is being copied for another still existing VMA.  As that VMA
3516          * has a reference to the reservation map it cannot disappear until
3517          * after this open call completes.  It is therefore safe to take a
3518          * new reference here without additional locking.
3519          */
3520         if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3521                 kref_get(&resv->refs);
3522 }
3523
3524 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3525 {
3526         struct hstate *h = hstate_vma(vma);
3527         struct resv_map *resv = vma_resv_map(vma);
3528         struct hugepage_subpool *spool = subpool_vma(vma);
3529         unsigned long reserve, start, end;
3530         long gbl_reserve;
3531
3532         if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3533                 return;
3534
3535         start = vma_hugecache_offset(h, vma, vma->vm_start);
3536         end = vma_hugecache_offset(h, vma, vma->vm_end);
3537
3538         reserve = (end - start) - region_count(resv, start, end);
3539         hugetlb_cgroup_uncharge_counter(resv, start, end);
3540         if (reserve) {
3541                 /*
3542                  * Decrement reserve counts.  The global reserve count may be
3543                  * adjusted if the subpool has a minimum size.
3544                  */
3545                 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3546                 hugetlb_acct_memory(h, -gbl_reserve);
3547         }
3548
3549         kref_put(&resv->refs, resv_map_release);
3550 }
3551
3552 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3553 {
3554         if (addr & ~(huge_page_mask(hstate_vma(vma))))
3555                 return -EINVAL;
3556         return 0;
3557 }
3558
3559 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3560 {
3561         struct hstate *hstate = hstate_vma(vma);
3562
3563         return 1UL << huge_page_shift(hstate);
3564 }
3565
3566 /*
3567  * We cannot handle pagefaults against hugetlb pages at all.  They cause
3568  * handle_mm_fault() to try to instantiate regular-sized pages in the
3569  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
3570  * this far.
3571  */
3572 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3573 {
3574         BUG();
3575         return 0;
3576 }
3577
3578 /*
3579  * When a new function is introduced to vm_operations_struct and added
3580  * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3581  * This is because under System V memory model, mappings created via
3582  * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3583  * their original vm_ops are overwritten with shm_vm_ops.
3584  */
3585 const struct vm_operations_struct hugetlb_vm_ops = {
3586         .fault = hugetlb_vm_op_fault,
3587         .open = hugetlb_vm_op_open,
3588         .close = hugetlb_vm_op_close,
3589         .split = hugetlb_vm_op_split,
3590         .pagesize = hugetlb_vm_op_pagesize,
3591 };
3592
3593 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3594                                 int writable)
3595 {
3596         pte_t entry;
3597
3598         if (writable) {
3599                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3600                                          vma->vm_page_prot)));
3601         } else {
3602                 entry = huge_pte_wrprotect(mk_huge_pte(page,
3603                                            vma->vm_page_prot));
3604         }
3605         entry = pte_mkyoung(entry);
3606         entry = pte_mkhuge(entry);
3607         entry = arch_make_huge_pte(entry, vma, page, writable);
3608
3609         return entry;
3610 }
3611
3612 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3613                                    unsigned long address, pte_t *ptep)
3614 {
3615         pte_t entry;
3616
3617         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3618         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3619                 update_mmu_cache(vma, address, ptep);
3620 }
3621
3622 bool is_hugetlb_entry_migration(pte_t pte)
3623 {
3624         swp_entry_t swp;
3625
3626         if (huge_pte_none(pte) || pte_present(pte))
3627                 return false;
3628         swp = pte_to_swp_entry(pte);
3629         if (non_swap_entry(swp) && is_migration_entry(swp))
3630                 return true;
3631         else
3632                 return false;
3633 }
3634
3635 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3636 {
3637         swp_entry_t swp;
3638
3639         if (huge_pte_none(pte) || pte_present(pte))
3640                 return 0;
3641         swp = pte_to_swp_entry(pte);
3642         if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3643                 return 1;
3644         else
3645                 return 0;
3646 }
3647
3648 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3649                             struct vm_area_struct *vma)
3650 {
3651         pte_t *src_pte, *dst_pte, entry, dst_entry;
3652         struct page *ptepage;
3653         unsigned long addr;
3654         int cow;
3655         struct hstate *h = hstate_vma(vma);
3656         unsigned long sz = huge_page_size(h);
3657         struct address_space *mapping = vma->vm_file->f_mapping;
3658         struct mmu_notifier_range range;
3659         int ret = 0;
3660
3661         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3662
3663         if (cow) {
3664                 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3665                                         vma->vm_start,
3666                                         vma->vm_end);
3667                 mmu_notifier_invalidate_range_start(&range);
3668         } else {
3669                 /*
3670                  * For shared mappings i_mmap_rwsem must be held to call
3671                  * huge_pte_alloc, otherwise the returned ptep could go
3672                  * away if part of a shared pmd and another thread calls
3673                  * huge_pmd_unshare.
3674                  */
3675                 i_mmap_lock_read(mapping);
3676         }
3677
3678         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3679                 spinlock_t *src_ptl, *dst_ptl;
3680                 src_pte = huge_pte_offset(src, addr, sz);
3681                 if (!src_pte)
3682                         continue;
3683                 dst_pte = huge_pte_alloc(dst, addr, sz);
3684                 if (!dst_pte) {
3685                         ret = -ENOMEM;
3686                         break;
3687                 }
3688
3689                 /*
3690                  * If the pagetables are shared don't copy or take references.
3691                  * dst_pte == src_pte is the common case of src/dest sharing.
3692                  *
3693                  * However, src could have 'unshared' and dst shares with
3694                  * another vma.  If dst_pte !none, this implies sharing.
3695                  * Check here before taking page table lock, and once again
3696                  * after taking the lock below.
3697                  */
3698                 dst_entry = huge_ptep_get(dst_pte);
3699                 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3700                         continue;
3701
3702                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3703                 src_ptl = huge_pte_lockptr(h, src, src_pte);
3704                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3705                 entry = huge_ptep_get(src_pte);
3706                 dst_entry = huge_ptep_get(dst_pte);
3707                 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3708                         /*
3709                          * Skip if src entry none.  Also, skip in the
3710                          * unlikely case dst entry !none as this implies
3711                          * sharing with another vma.
3712                          */
3713                         ;
3714                 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3715                                     is_hugetlb_entry_hwpoisoned(entry))) {
3716                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
3717
3718                         if (is_write_migration_entry(swp_entry) && cow) {
3719                                 /*
3720                                  * COW mappings require pages in both
3721                                  * parent and child to be set to read.
3722                                  */
3723                                 make_migration_entry_read(&swp_entry);
3724                                 entry = swp_entry_to_pte(swp_entry);
3725                                 set_huge_swap_pte_at(src, addr, src_pte,
3726                                                      entry, sz);
3727                         }
3728                         set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3729                 } else {
3730                         if (cow) {
3731                                 /*
3732                                  * No need to notify as we are downgrading page
3733                                  * table protection not changing it to point
3734                                  * to a new page.
3735                                  *
3736                                  * See Documentation/vm/mmu_notifier.rst
3737                                  */
3738                                 huge_ptep_set_wrprotect(src, addr, src_pte);
3739                         }
3740                         entry = huge_ptep_get(src_pte);
3741                         ptepage = pte_page(entry);
3742                         get_page(ptepage);
3743                         page_dup_rmap(ptepage, true);
3744                         set_huge_pte_at(dst, addr, dst_pte, entry);
3745                         hugetlb_count_add(pages_per_huge_page(h), dst);
3746                 }
3747                 spin_unlock(src_ptl);
3748                 spin_unlock(dst_ptl);
3749         }
3750
3751         if (cow)
3752                 mmu_notifier_invalidate_range_end(&range);
3753         else
3754                 i_mmap_unlock_read(mapping);
3755
3756         return ret;
3757 }
3758
3759 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3760                             unsigned long start, unsigned long end,
3761                             struct page *ref_page)
3762 {
3763         struct mm_struct *mm = vma->vm_mm;
3764         unsigned long address;
3765         pte_t *ptep;
3766         pte_t pte;
3767         spinlock_t *ptl;
3768         struct page *page;
3769         struct hstate *h = hstate_vma(vma);
3770         unsigned long sz = huge_page_size(h);
3771         struct mmu_notifier_range range;
3772
3773         WARN_ON(!is_vm_hugetlb_page(vma));
3774         BUG_ON(start & ~huge_page_mask(h));
3775         BUG_ON(end & ~huge_page_mask(h));
3776
3777         /*
3778          * This is a hugetlb vma, all the pte entries should point
3779          * to huge page.
3780          */
3781         tlb_change_page_size(tlb, sz);
3782         tlb_start_vma(tlb, vma);
3783
3784         /*
3785          * If sharing possible, alert mmu notifiers of worst case.
3786          */
3787         mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3788                                 end);
3789         adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3790         mmu_notifier_invalidate_range_start(&range);
3791         address = start;
3792         for (; address < end; address += sz) {
3793                 ptep = huge_pte_offset(mm, address, sz);
3794                 if (!ptep)
3795                         continue;
3796
3797                 ptl = huge_pte_lock(h, mm, ptep);
3798                 if (huge_pmd_unshare(mm, &address, ptep)) {
3799                         spin_unlock(ptl);
3800                         /*
3801                          * We just unmapped a page of PMDs by clearing a PUD.
3802                          * The caller's TLB flush range should cover this area.
3803                          */
3804                         continue;
3805                 }
3806
3807                 pte = huge_ptep_get(ptep);
3808                 if (huge_pte_none(pte)) {
3809                         spin_unlock(ptl);
3810                         continue;
3811                 }
3812
3813                 /*
3814                  * Migrating hugepage or HWPoisoned hugepage is already
3815                  * unmapped and its refcount is dropped, so just clear pte here.
3816                  */
3817                 if (unlikely(!pte_present(pte))) {
3818                         huge_pte_clear(mm, address, ptep, sz);
3819                         spin_unlock(ptl);
3820                         continue;
3821                 }
3822
3823                 page = pte_page(pte);
3824                 /*
3825                  * If a reference page is supplied, it is because a specific
3826                  * page is being unmapped, not a range. Ensure the page we
3827                  * are about to unmap is the actual page of interest.
3828                  */
3829                 if (ref_page) {
3830                         if (page != ref_page) {
3831                                 spin_unlock(ptl);
3832                                 continue;
3833                         }
3834                         /*
3835                          * Mark the VMA as having unmapped its page so that
3836                          * future faults in this VMA will fail rather than
3837                          * looking like data was lost
3838                          */
3839                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3840                 }
3841
3842                 pte = huge_ptep_get_and_clear(mm, address, ptep);
3843                 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3844                 if (huge_pte_dirty(pte))
3845                         set_page_dirty(page);
3846
3847                 hugetlb_count_sub(pages_per_huge_page(h), mm);
3848                 page_remove_rmap(page, true);
3849
3850                 spin_unlock(ptl);
3851                 tlb_remove_page_size(tlb, page, huge_page_size(h));
3852                 /*
3853                  * Bail out after unmapping reference page if supplied
3854                  */
3855                 if (ref_page)
3856                         break;
3857         }
3858         mmu_notifier_invalidate_range_end(&range);
3859         tlb_end_vma(tlb, vma);
3860 }
3861
3862 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3863                           struct vm_area_struct *vma, unsigned long start,
3864                           unsigned long end, struct page *ref_page)
3865 {
3866         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3867
3868         /*
3869          * Clear this flag so that x86's huge_pmd_share page_table_shareable
3870          * test will fail on a vma being torn down, and not grab a page table
3871          * on its way out.  We're lucky that the flag has such an appropriate
3872          * name, and can in fact be safely cleared here. We could clear it
3873          * before the __unmap_hugepage_range above, but all that's necessary
3874          * is to clear it before releasing the i_mmap_rwsem. This works
3875          * because in the context this is called, the VMA is about to be
3876          * destroyed and the i_mmap_rwsem is held.
3877          */
3878         vma->vm_flags &= ~VM_MAYSHARE;
3879 }
3880
3881 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3882                           unsigned long end, struct page *ref_page)
3883 {
3884         struct mm_struct *mm;
3885         struct mmu_gather tlb;
3886         unsigned long tlb_start = start;
3887         unsigned long tlb_end = end;
3888
3889         /*
3890          * If shared PMDs were possibly used within this vma range, adjust
3891          * start/end for worst case tlb flushing.
3892          * Note that we can not be sure if PMDs are shared until we try to
3893          * unmap pages.  However, we want to make sure TLB flushing covers
3894          * the largest possible range.
3895          */
3896         adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
3897
3898         mm = vma->vm_mm;
3899
3900         tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
3901         __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3902         tlb_finish_mmu(&tlb, tlb_start, tlb_end);
3903 }
3904
3905 /*
3906  * This is called when the original mapper is failing to COW a MAP_PRIVATE
3907  * mappping it owns the reserve page for. The intention is to unmap the page
3908  * from other VMAs and let the children be SIGKILLed if they are faulting the
3909  * same region.
3910  */
3911 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3912                               struct page *page, unsigned long address)
3913 {
3914         struct hstate *h = hstate_vma(vma);
3915         struct vm_area_struct *iter_vma;
3916         struct address_space *mapping;
3917         pgoff_t pgoff;
3918
3919         /*
3920          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3921          * from page cache lookup which is in HPAGE_SIZE units.
3922          */
3923         address = address & huge_page_mask(h);
3924         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3925                         vma->vm_pgoff;
3926         mapping = vma->vm_file->f_mapping;
3927
3928         /*
3929          * Take the mapping lock for the duration of the table walk. As
3930          * this mapping should be shared between all the VMAs,
3931          * __unmap_hugepage_range() is called as the lock is already held
3932          */
3933         i_mmap_lock_write(mapping);
3934         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3935                 /* Do not unmap the current VMA */
3936                 if (iter_vma == vma)
3937                         continue;
3938
3939                 /*
3940                  * Shared VMAs have their own reserves and do not affect
3941                  * MAP_PRIVATE accounting but it is possible that a shared
3942                  * VMA is using the same page so check and skip such VMAs.
3943                  */
3944                 if (iter_vma->vm_flags & VM_MAYSHARE)
3945                         continue;
3946
3947                 /*
3948                  * Unmap the page from other VMAs without their own reserves.
3949                  * They get marked to be SIGKILLed if they fault in these
3950                  * areas. This is because a future no-page fault on this VMA
3951                  * could insert a zeroed page instead of the data existing
3952                  * from the time of fork. This would look like data corruption
3953                  */
3954                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3955                         unmap_hugepage_range(iter_vma, address,
3956                                              address + huge_page_size(h), page);
3957         }
3958         i_mmap_unlock_write(mapping);
3959 }
3960
3961 /*
3962  * Hugetlb_cow() should be called with page lock of the original hugepage held.
3963  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3964  * cannot race with other handlers or page migration.
3965  * Keep the pte_same checks anyway to make transition from the mutex easier.
3966  */
3967 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3968                        unsigned long address, pte_t *ptep,
3969                        struct page *pagecache_page, spinlock_t *ptl)
3970 {
3971         pte_t pte;
3972         struct hstate *h = hstate_vma(vma);
3973         struct page *old_page, *new_page;
3974         int outside_reserve = 0;
3975         vm_fault_t ret = 0;
3976         unsigned long haddr = address & huge_page_mask(h);
3977         struct mmu_notifier_range range;
3978
3979         pte = huge_ptep_get(ptep);
3980         old_page = pte_page(pte);
3981
3982 retry_avoidcopy:
3983         /* If no-one else is actually using this page, avoid the copy
3984          * and just make the page writable */
3985         if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3986                 page_move_anon_rmap(old_page, vma);
3987                 set_huge_ptep_writable(vma, haddr, ptep);
3988                 return 0;
3989         }
3990
3991         /*
3992          * If the process that created a MAP_PRIVATE mapping is about to
3993          * perform a COW due to a shared page count, attempt to satisfy
3994          * the allocation without using the existing reserves. The pagecache
3995          * page is used to determine if the reserve at this address was
3996          * consumed or not. If reserves were used, a partial faulted mapping
3997          * at the time of fork() could consume its reserves on COW instead
3998          * of the full address range.
3999          */
4000         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4001                         old_page != pagecache_page)
4002                 outside_reserve = 1;
4003
4004         get_page(old_page);
4005
4006         /*
4007          * Drop page table lock as buddy allocator may be called. It will
4008          * be acquired again before returning to the caller, as expected.
4009          */
4010         spin_unlock(ptl);
4011         new_page = alloc_huge_page(vma, haddr, outside_reserve);
4012
4013         if (IS_ERR(new_page)) {
4014                 /*
4015                  * If a process owning a MAP_PRIVATE mapping fails to COW,
4016                  * it is due to references held by a child and an insufficient
4017                  * huge page pool. To guarantee the original mappers
4018                  * reliability, unmap the page from child processes. The child
4019                  * may get SIGKILLed if it later faults.
4020                  */
4021                 if (outside_reserve) {
4022                         put_page(old_page);
4023                         BUG_ON(huge_pte_none(pte));
4024                         unmap_ref_private(mm, vma, old_page, haddr);
4025                         BUG_ON(huge_pte_none(pte));
4026                         spin_lock(ptl);
4027                         ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4028                         if (likely(ptep &&
4029                                    pte_same(huge_ptep_get(ptep), pte)))
4030                                 goto retry_avoidcopy;
4031                         /*
4032                          * race occurs while re-acquiring page table
4033                          * lock, and our job is done.
4034                          */
4035                         return 0;
4036                 }
4037
4038                 ret = vmf_error(PTR_ERR(new_page));
4039                 goto out_release_old;
4040         }
4041
4042         /*
4043          * When the original hugepage is shared one, it does not have
4044          * anon_vma prepared.
4045          */
4046         if (unlikely(anon_vma_prepare(vma))) {
4047                 ret = VM_FAULT_OOM;
4048                 goto out_release_all;
4049         }
4050
4051         copy_user_huge_page(new_page, old_page, address, vma,
4052                             pages_per_huge_page(h));
4053         __SetPageUptodate(new_page);
4054
4055         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4056                                 haddr + huge_page_size(h));
4057         mmu_notifier_invalidate_range_start(&range);
4058
4059         /*
4060          * Retake the page table lock to check for racing updates
4061          * before the page tables are altered
4062          */
4063         spin_lock(ptl);
4064         ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4065         if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4066                 ClearPagePrivate(new_page);
4067
4068                 /* Break COW */
4069                 huge_ptep_clear_flush(vma, haddr, ptep);
4070                 mmu_notifier_invalidate_range(mm, range.start, range.end);
4071                 set_huge_pte_at(mm, haddr, ptep,
4072                                 make_huge_pte(vma, new_page, 1));
4073                 page_remove_rmap(old_page, true);
4074                 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4075                 set_page_huge_active(new_page);
4076                 /* Make the old page be freed below */
4077                 new_page = old_page;
4078         }
4079         spin_unlock(ptl);
4080         mmu_notifier_invalidate_range_end(&range);
4081 out_release_all:
4082         restore_reserve_on_error(h, vma, haddr, new_page);
4083         put_page(new_page);
4084 out_release_old:
4085         put_page(old_page);
4086
4087         spin_lock(ptl); /* Caller expects lock to be held */
4088         return ret;
4089 }
4090
4091 /* Return the pagecache page at a given address within a VMA */
4092 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4093                         struct vm_area_struct *vma, unsigned long address)
4094 {
4095         struct address_space *mapping;
4096         pgoff_t idx;
4097
4098         mapping = vma->vm_file->f_mapping;
4099         idx = vma_hugecache_offset(h, vma, address);
4100
4101         return find_lock_page(mapping, idx);
4102 }
4103
4104 /*
4105  * Return whether there is a pagecache page to back given address within VMA.
4106  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4107  */
4108 static bool hugetlbfs_pagecache_present(struct hstate *h,
4109                         struct vm_area_struct *vma, unsigned long address)
4110 {
4111         struct address_space *mapping;
4112         pgoff_t idx;
4113         struct page *page;
4114
4115         mapping = vma->vm_file->f_mapping;
4116         idx = vma_hugecache_offset(h, vma, address);
4117
4118         page = find_get_page(mapping, idx);
4119         if (page)
4120                 put_page(page);
4121         return page != NULL;
4122 }
4123
4124 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4125                            pgoff_t idx)
4126 {
4127         struct inode *inode = mapping->host;
4128         struct hstate *h = hstate_inode(inode);
4129         int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4130
4131         if (err)
4132                 return err;
4133         ClearPagePrivate(page);
4134
4135         /*
4136          * set page dirty so that it will not be removed from cache/file
4137          * by non-hugetlbfs specific code paths.
4138          */
4139         set_page_dirty(page);
4140
4141         spin_lock(&inode->i_lock);
4142         inode->i_blocks += blocks_per_huge_page(h);
4143         spin_unlock(&inode->i_lock);
4144         return 0;
4145 }
4146
4147 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4148                         struct vm_area_struct *vma,
4149                         struct address_space *mapping, pgoff_t idx,
4150                         unsigned long address, pte_t *ptep, unsigned int flags)
4151 {
4152         struct hstate *h = hstate_vma(vma);
4153         vm_fault_t ret = VM_FAULT_SIGBUS;
4154         int anon_rmap = 0;
4155         unsigned long size;
4156         struct page *page;
4157         pte_t new_pte;
4158         spinlock_t *ptl;
4159         unsigned long haddr = address & huge_page_mask(h);
4160         bool new_page = false;
4161
4162         /*
4163          * Currently, we are forced to kill the process in the event the
4164          * original mapper has unmapped pages from the child due to a failed
4165          * COW. Warn that such a situation has occurred as it may not be obvious
4166          */
4167         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4168                 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4169                            current->pid);
4170                 return ret;
4171         }
4172
4173         /*
4174          * We can not race with truncation due to holding i_mmap_rwsem.
4175          * i_size is modified when holding i_mmap_rwsem, so check here
4176          * once for faults beyond end of file.
4177          */
4178         size = i_size_read(mapping->host) >> huge_page_shift(h);
4179         if (idx >= size)
4180                 goto out;
4181
4182 retry:
4183         page = find_lock_page(mapping, idx);
4184         if (!page) {
4185                 /*
4186                  * Check for page in userfault range
4187                  */
4188                 if (userfaultfd_missing(vma)) {
4189                         u32 hash;
4190                         struct vm_fault vmf = {
4191                                 .vma = vma,
4192                                 .address = haddr,
4193                                 .flags = flags,
4194                                 /*
4195                                  * Hard to debug if it ends up being
4196                                  * used by a callee that assumes
4197                                  * something about the other
4198                                  * uninitialized fields... same as in
4199                                  * memory.c
4200                                  */
4201                         };
4202
4203                         /*
4204                          * hugetlb_fault_mutex and i_mmap_rwsem must be
4205                          * dropped before handling userfault.  Reacquire
4206                          * after handling fault to make calling code simpler.
4207                          */
4208                         hash = hugetlb_fault_mutex_hash(mapping, idx);
4209                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4210                         i_mmap_unlock_read(mapping);
4211                         ret = handle_userfault(&vmf, VM_UFFD_MISSING);
4212                         i_mmap_lock_read(mapping);
4213                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
4214                         goto out;
4215                 }
4216
4217                 page = alloc_huge_page(vma, haddr, 0);
4218                 if (IS_ERR(page)) {
4219                         /*
4220                          * Returning error will result in faulting task being
4221                          * sent SIGBUS.  The hugetlb fault mutex prevents two
4222                          * tasks from racing to fault in the same page which
4223                          * could result in false unable to allocate errors.
4224                          * Page migration does not take the fault mutex, but
4225                          * does a clear then write of pte's under page table
4226                          * lock.  Page fault code could race with migration,
4227                          * notice the clear pte and try to allocate a page
4228                          * here.  Before returning error, get ptl and make
4229                          * sure there really is no pte entry.
4230                          */
4231                         ptl = huge_pte_lock(h, mm, ptep);
4232                         if (!huge_pte_none(huge_ptep_get(ptep))) {
4233                                 ret = 0;
4234                                 spin_unlock(ptl);
4235                                 goto out;
4236                         }
4237                         spin_unlock(ptl);
4238                         ret = vmf_error(PTR_ERR(page));
4239                         goto out;
4240                 }
4241                 clear_huge_page(page, address, pages_per_huge_page(h));
4242                 __SetPageUptodate(page);
4243                 new_page = true;
4244
4245                 if (vma->vm_flags & VM_MAYSHARE) {
4246                         int err = huge_add_to_page_cache(page, mapping, idx);
4247                         if (err) {
4248                                 put_page(page);
4249                                 if (err == -EEXIST)
4250                                         goto retry;
4251                                 goto out;
4252                         }
4253                 } else {
4254                         lock_page(page);
4255                         if (unlikely(anon_vma_prepare(vma))) {
4256                                 ret = VM_FAULT_OOM;
4257                                 goto backout_unlocked;
4258                         }
4259                         anon_rmap = 1;
4260                 }
4261         } else {
4262                 /*
4263                  * If memory error occurs between mmap() and fault, some process
4264                  * don't have hwpoisoned swap entry for errored virtual address.
4265                  * So we need to block hugepage fault by PG_hwpoison bit check.
4266                  */
4267                 if (unlikely(PageHWPoison(page))) {
4268                         ret = VM_FAULT_HWPOISON |
4269                                 VM_FAULT_SET_HINDEX(hstate_index(h));
4270                         goto backout_unlocked;
4271                 }
4272         }
4273
4274         /*
4275          * If we are going to COW a private mapping later, we examine the
4276          * pending reservations for this page now. This will ensure that
4277          * any allocations necessary to record that reservation occur outside
4278          * the spinlock.
4279          */
4280         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4281                 if (vma_needs_reservation(h, vma, haddr) < 0) {
4282                         ret = VM_FAULT_OOM;
4283                         goto backout_unlocked;
4284                 }
4285                 /* Just decrements count, does not deallocate */
4286                 vma_end_reservation(h, vma, haddr);
4287         }
4288
4289         ptl = huge_pte_lock(h, mm, ptep);
4290         ret = 0;
4291         if (!huge_pte_none(huge_ptep_get(ptep)))
4292                 goto backout;
4293
4294         if (anon_rmap) {
4295                 ClearPagePrivate(page);
4296                 hugepage_add_new_anon_rmap(page, vma, haddr);
4297         } else
4298                 page_dup_rmap(page, true);
4299         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4300                                 && (vma->vm_flags & VM_SHARED)));
4301         set_huge_pte_at(mm, haddr, ptep, new_pte);
4302
4303         hugetlb_count_add(pages_per_huge_page(h), mm);
4304         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4305                 /* Optimization, do the COW without a second fault */
4306                 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4307         }
4308
4309         spin_unlock(ptl);
4310
4311         /*
4312          * Only make newly allocated pages active.  Existing pages found
4313          * in the pagecache could be !page_huge_active() if they have been
4314          * isolated for migration.
4315          */
4316         if (new_page)
4317                 set_page_huge_active(page);
4318
4319         unlock_page(page);
4320 out:
4321         return ret;
4322
4323 backout:
4324         spin_unlock(ptl);
4325 backout_unlocked:
4326         unlock_page(page);
4327         restore_reserve_on_error(h, vma, haddr, page);
4328         put_page(page);
4329         goto out;
4330 }
4331
4332 #ifdef CONFIG_SMP
4333 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4334 {
4335         unsigned long key[2];
4336         u32 hash;
4337
4338         key[0] = (unsigned long) mapping;
4339         key[1] = idx;
4340
4341         hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4342
4343         return hash & (num_fault_mutexes - 1);
4344 }
4345 #else
4346 /*
4347  * For uniprocesor systems we always use a single mutex, so just
4348  * return 0 and avoid the hashing overhead.
4349  */
4350 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4351 {
4352         return 0;
4353 }
4354 #endif
4355
4356 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4357                         unsigned long address, unsigned int flags)
4358 {
4359         pte_t *ptep, entry;
4360         spinlock_t *ptl;
4361         vm_fault_t ret;
4362         u32 hash;
4363         pgoff_t idx;
4364         struct page *page = NULL;
4365         struct page *pagecache_page = NULL;
4366         struct hstate *h = hstate_vma(vma);
4367         struct address_space *mapping;
4368         int need_wait_lock = 0;
4369         unsigned long haddr = address & huge_page_mask(h);
4370
4371         ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4372         if (ptep) {
4373                 /*
4374                  * Since we hold no locks, ptep could be stale.  That is
4375                  * OK as we are only making decisions based on content and
4376                  * not actually modifying content here.
4377                  */
4378                 entry = huge_ptep_get(ptep);
4379                 if (unlikely(is_hugetlb_entry_migration(entry))) {
4380                         migration_entry_wait_huge(vma, mm, ptep);
4381                         return 0;
4382                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4383                         return VM_FAULT_HWPOISON_LARGE |
4384                                 VM_FAULT_SET_HINDEX(hstate_index(h));
4385         } else {
4386                 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4387                 if (!ptep)
4388                         return VM_FAULT_OOM;
4389         }
4390
4391         /*
4392          * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4393          * until finished with ptep.  This serves two purposes:
4394          * 1) It prevents huge_pmd_unshare from being called elsewhere
4395          *    and making the ptep no longer valid.
4396          * 2) It synchronizes us with i_size modifications during truncation.
4397          *
4398          * ptep could have already be assigned via huge_pte_offset.  That
4399          * is OK, as huge_pte_alloc will return the same value unless
4400          * something has changed.
4401          */
4402         mapping = vma->vm_file->f_mapping;
4403         i_mmap_lock_read(mapping);
4404         ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4405         if (!ptep) {
4406                 i_mmap_unlock_read(mapping);
4407                 return VM_FAULT_OOM;
4408         }
4409
4410         /*
4411          * Serialize hugepage allocation and instantiation, so that we don't
4412          * get spurious allocation failures if two CPUs race to instantiate
4413          * the same page in the page cache.
4414          */
4415         idx = vma_hugecache_offset(h, vma, haddr);
4416         hash = hugetlb_fault_mutex_hash(mapping, idx);
4417         mutex_lock(&hugetlb_fault_mutex_table[hash]);
4418
4419         entry = huge_ptep_get(ptep);
4420         if (huge_pte_none(entry)) {
4421                 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4422                 goto out_mutex;
4423         }
4424
4425         ret = 0;
4426
4427         /*
4428          * entry could be a migration/hwpoison entry at this point, so this
4429          * check prevents the kernel from going below assuming that we have
4430          * a active hugepage in pagecache. This goto expects the 2nd page fault,
4431          * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
4432          * handle it.
4433          */
4434         if (!pte_present(entry))
4435                 goto out_mutex;
4436
4437         /*
4438          * If we are going to COW the mapping later, we examine the pending
4439          * reservations for this page now. This will ensure that any
4440          * allocations necessary to record that reservation occur outside the
4441          * spinlock. For private mappings, we also lookup the pagecache
4442          * page now as it is used to determine if a reservation has been
4443          * consumed.
4444          */
4445         if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4446                 if (vma_needs_reservation(h, vma, haddr) < 0) {
4447                         ret = VM_FAULT_OOM;
4448                         goto out_mutex;
4449                 }
4450                 /* Just decrements count, does not deallocate */
4451                 vma_end_reservation(h, vma, haddr);
4452
4453                 if (!(vma->vm_flags & VM_MAYSHARE))
4454                         pagecache_page = hugetlbfs_pagecache_page(h,
4455                                                                 vma, haddr);
4456         }
4457
4458         ptl = huge_pte_lock(h, mm, ptep);
4459
4460         /* Check for a racing update before calling hugetlb_cow */
4461         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4462                 goto out_ptl;
4463
4464         /*
4465          * hugetlb_cow() requires page locks of pte_page(entry) and
4466          * pagecache_page, so here we need take the former one
4467          * when page != pagecache_page or !pagecache_page.
4468          */
4469         page = pte_page(entry);
4470         if (page != pagecache_page)
4471                 if (!trylock_page(page)) {
4472                         need_wait_lock = 1;
4473                         goto out_ptl;
4474                 }
4475
4476         get_page(page);
4477
4478         if (flags & FAULT_FLAG_WRITE) {
4479                 if (!huge_pte_write(entry)) {
4480                         ret = hugetlb_cow(mm, vma, address, ptep,
4481                                           pagecache_page, ptl);
4482                         goto out_put_page;
4483                 }
4484                 entry = huge_pte_mkdirty(entry);
4485         }
4486         entry = pte_mkyoung(entry);
4487         if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4488                                                 flags & FAULT_FLAG_WRITE))
4489                 update_mmu_cache(vma, haddr, ptep);
4490 out_put_page:
4491         if (page != pagecache_page)
4492                 unlock_page(page);
4493         put_page(page);
4494 out_ptl:
4495         spin_unlock(ptl);
4496
4497         if (pagecache_page) {
4498                 unlock_page(pagecache_page);
4499                 put_page(pagecache_page);
4500         }
4501 out_mutex:
4502         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4503         i_mmap_unlock_read(mapping);
4504         /*
4505          * Generally it's safe to hold refcount during waiting page lock. But
4506          * here we just wait to defer the next page fault to avoid busy loop and
4507          * the page is not used after unlocked before returning from the current
4508          * page fault. So we are safe from accessing freed page, even if we wait
4509          * here without taking refcount.
4510          */
4511         if (need_wait_lock)
4512                 wait_on_page_locked(page);
4513         return ret;
4514 }
4515
4516 /*
4517  * Used by userfaultfd UFFDIO_COPY.  Based on mcopy_atomic_pte with
4518  * modifications for huge pages.
4519  */
4520 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4521                             pte_t *dst_pte,
4522                             struct vm_area_struct *dst_vma,
4523                             unsigned long dst_addr,
4524                             unsigned long src_addr,
4525                             struct page **pagep)
4526 {
4527         struct address_space *mapping;
4528         pgoff_t idx;
4529         unsigned long size;
4530         int vm_shared = dst_vma->vm_flags & VM_SHARED;
4531         struct hstate *h = hstate_vma(dst_vma);
4532         pte_t _dst_pte;
4533         spinlock_t *ptl;
4534         int ret;
4535         struct page *page;
4536
4537         if (!*pagep) {
4538                 ret = -ENOMEM;
4539                 page = alloc_huge_page(dst_vma, dst_addr, 0);
4540                 if (IS_ERR(page))
4541                         goto out;
4542
4543                 ret = copy_huge_page_from_user(page,
4544                                                 (const void __user *) src_addr,
4545                                                 pages_per_huge_page(h), false);
4546
4547                 /* fallback to copy_from_user outside mmap_sem */
4548                 if (unlikely(ret)) {
4549                         ret = -ENOENT;
4550                         *pagep = page;
4551                         /* don't free the page */
4552                         goto out;
4553                 }
4554         } else {
4555                 page = *pagep;
4556                 *pagep = NULL;
4557         }
4558
4559         /*
4560          * The memory barrier inside __SetPageUptodate makes sure that
4561          * preceding stores to the page contents become visible before
4562          * the set_pte_at() write.
4563          */
4564         __SetPageUptodate(page);
4565
4566         mapping = dst_vma->vm_file->f_mapping;
4567         idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4568
4569         /*
4570          * If shared, add to page cache
4571          */
4572         if (vm_shared) {
4573                 size = i_size_read(mapping->host) >> huge_page_shift(h);
4574                 ret = -EFAULT;
4575                 if (idx >= size)
4576                         goto out_release_nounlock;
4577
4578                 /*
4579                  * Serialization between remove_inode_hugepages() and
4580                  * huge_add_to_page_cache() below happens through the
4581                  * hugetlb_fault_mutex_table that here must be hold by
4582                  * the caller.
4583                  */
4584                 ret = huge_add_to_page_cache(page, mapping, idx);
4585                 if (ret)
4586                         goto out_release_nounlock;
4587         }
4588
4589         ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4590         spin_lock(ptl);
4591
4592         /*
4593          * Recheck the i_size after holding PT lock to make sure not
4594          * to leave any page mapped (as page_mapped()) beyond the end
4595          * of the i_size (remove_inode_hugepages() is strict about
4596          * enforcing that). If we bail out here, we'll also leave a
4597          * page in the radix tree in the vm_shared case beyond the end
4598          * of the i_size, but remove_inode_hugepages() will take care
4599          * of it as soon as we drop the hugetlb_fault_mutex_table.
4600          */
4601         size = i_size_read(mapping->host) >> huge_page_shift(h);
4602         ret = -EFAULT;
4603         if (idx >= size)
4604                 goto out_release_unlock;
4605
4606         ret = -EEXIST;
4607         if (!huge_pte_none(huge_ptep_get(dst_pte)))
4608                 goto out_release_unlock;
4609
4610         if (vm_shared) {
4611                 page_dup_rmap(page, true);
4612         } else {
4613                 ClearPagePrivate(page);
4614                 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4615         }
4616
4617         _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4618         if (dst_vma->vm_flags & VM_WRITE)
4619                 _dst_pte = huge_pte_mkdirty(_dst_pte);
4620         _dst_pte = pte_mkyoung(_dst_pte);
4621
4622         set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4623
4624         (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4625                                         dst_vma->vm_flags & VM_WRITE);
4626         hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4627
4628         /* No need to invalidate - it was non-present before */
4629         update_mmu_cache(dst_vma, dst_addr, dst_pte);
4630
4631         spin_unlock(ptl);
4632         set_page_huge_active(page);
4633         if (vm_shared)
4634                 unlock_page(page);
4635         ret = 0;
4636 out:
4637         return ret;
4638 out_release_unlock:
4639         spin_unlock(ptl);
4640         if (vm_shared)
4641                 unlock_page(page);
4642 out_release_nounlock:
4643         put_page(page);
4644         goto out;
4645 }
4646
4647 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4648                          struct page **pages, struct vm_area_struct **vmas,
4649                          unsigned long *position, unsigned long *nr_pages,
4650                          long i, unsigned int flags, int *locked)
4651 {
4652         unsigned long pfn_offset;
4653         unsigned long vaddr = *position;
4654         unsigned long remainder = *nr_pages;
4655         struct hstate *h = hstate_vma(vma);
4656         int err = -EFAULT;
4657
4658         while (vaddr < vma->vm_end && remainder) {
4659                 pte_t *pte;
4660                 spinlock_t *ptl = NULL;
4661                 int absent;
4662                 struct page *page;
4663
4664                 /*
4665                  * If we have a pending SIGKILL, don't keep faulting pages and
4666                  * potentially allocating memory.
4667                  */
4668                 if (fatal_signal_pending(current)) {
4669                         remainder = 0;
4670                         break;
4671                 }
4672
4673                 /*
4674                  * Some archs (sparc64, sh*) have multiple pte_ts to
4675                  * each hugepage.  We have to make sure we get the
4676                  * first, for the page indexing below to work.
4677                  *
4678                  * Note that page table lock is not held when pte is null.
4679                  */
4680                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4681                                       huge_page_size(h));
4682                 if (pte)
4683                         ptl = huge_pte_lock(h, mm, pte);
4684                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4685
4686                 /*
4687                  * When coredumping, it suits get_dump_page if we just return
4688                  * an error where there's an empty slot with no huge pagecache
4689                  * to back it.  This way, we avoid allocating a hugepage, and
4690                  * the sparse dumpfile avoids allocating disk blocks, but its
4691                  * huge holes still show up with zeroes where they need to be.
4692                  */
4693                 if (absent && (flags & FOLL_DUMP) &&
4694                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4695                         if (pte)
4696                                 spin_unlock(ptl);
4697                         remainder = 0;
4698                         break;
4699                 }
4700
4701                 /*
4702                  * We need call hugetlb_fault for both hugepages under migration
4703                  * (in which case hugetlb_fault waits for the migration,) and
4704                  * hwpoisoned hugepages (in which case we need to prevent the
4705                  * caller from accessing to them.) In order to do this, we use
4706                  * here is_swap_pte instead of is_hugetlb_entry_migration and
4707                  * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4708                  * both cases, and because we can't follow correct pages
4709                  * directly from any kind of swap entries.
4710                  */
4711                 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4712                     ((flags & FOLL_WRITE) &&
4713                       !huge_pte_write(huge_ptep_get(pte)))) {
4714                         vm_fault_t ret;
4715                         unsigned int fault_flags = 0;
4716
4717                         if (pte)
4718                                 spin_unlock(ptl);
4719                         if (flags & FOLL_WRITE)
4720                                 fault_flags |= FAULT_FLAG_WRITE;
4721                         if (locked)
4722                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4723                                         FAULT_FLAG_KILLABLE;
4724                         if (flags & FOLL_NOWAIT)
4725                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4726                                         FAULT_FLAG_RETRY_NOWAIT;
4727                         if (flags & FOLL_TRIED) {
4728                                 /*
4729                                  * Note: FAULT_FLAG_ALLOW_RETRY and
4730                                  * FAULT_FLAG_TRIED can co-exist
4731                                  */
4732                                 fault_flags |= FAULT_FLAG_TRIED;
4733                         }
4734                         ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4735                         if (ret & VM_FAULT_ERROR) {
4736                                 err = vm_fault_to_errno(ret, flags);
4737                                 remainder = 0;
4738                                 break;
4739                         }
4740                         if (ret & VM_FAULT_RETRY) {
4741                                 if (locked &&
4742                                     !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4743                                         *locked = 0;
4744                                 *nr_pages = 0;
4745                                 /*
4746                                  * VM_FAULT_RETRY must not return an
4747                                  * error, it will return zero
4748                                  * instead.
4749                                  *
4750                                  * No need to update "position" as the
4751                                  * caller will not check it after
4752                                  * *nr_pages is set to 0.
4753                                  */
4754                                 return i;
4755                         }
4756                         continue;
4757                 }
4758
4759                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4760                 page = pte_page(huge_ptep_get(pte));
4761
4762                 /*
4763                  * If subpage information not requested, update counters
4764                  * and skip the same_page loop below.
4765                  */
4766                 if (!pages && !vmas && !pfn_offset &&
4767                     (vaddr + huge_page_size(h) < vma->vm_end) &&
4768                     (remainder >= pages_per_huge_page(h))) {
4769                         vaddr += huge_page_size(h);
4770                         remainder -= pages_per_huge_page(h);
4771                         i += pages_per_huge_page(h);
4772                         spin_unlock(ptl);
4773                         continue;
4774                 }
4775
4776 same_page:
4777                 if (pages) {
4778                         pages[i] = mem_map_offset(page, pfn_offset);
4779                         /*
4780                          * try_grab_page() should always succeed here, because:
4781                          * a) we hold the ptl lock, and b) we've just checked
4782                          * that the huge page is present in the page tables. If
4783                          * the huge page is present, then the tail pages must
4784                          * also be present. The ptl prevents the head page and
4785                          * tail pages from being rearranged in any way. So this
4786                          * page must be available at this point, unless the page
4787                          * refcount overflowed:
4788                          */
4789                         if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) {
4790                                 spin_unlock(ptl);
4791                                 remainder = 0;
4792                                 err = -ENOMEM;
4793                                 break;
4794                         }
4795                 }
4796
4797                 if (vmas)
4798                         vmas[i] = vma;
4799
4800                 vaddr += PAGE_SIZE;
4801                 ++pfn_offset;
4802                 --remainder;
4803                 ++i;
4804                 if (vaddr < vma->vm_end && remainder &&
4805                                 pfn_offset < pages_per_huge_page(h)) {
4806                         /*
4807                          * We use pfn_offset to avoid touching the pageframes
4808                          * of this compound page.
4809                          */
4810                         goto same_page;
4811                 }
4812                 spin_unlock(ptl);
4813         }
4814         *nr_pages = remainder;
4815         /*
4816          * setting position is actually required only if remainder is
4817          * not zero but it's faster not to add a "if (remainder)"
4818          * branch.
4819          */
4820         *position = vaddr;
4821
4822         return i ? i : err;
4823 }
4824
4825 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4826 /*
4827  * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4828  * implement this.
4829  */
4830 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4831 #endif
4832
4833 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4834                 unsigned long address, unsigned long end, pgprot_t newprot)
4835 {
4836         struct mm_struct *mm = vma->vm_mm;
4837         unsigned long start = address;
4838         pte_t *ptep;
4839         pte_t pte;
4840         struct hstate *h = hstate_vma(vma);
4841         unsigned long pages = 0;
4842         bool shared_pmd = false;
4843         struct mmu_notifier_range range;
4844
4845         /*
4846          * In the case of shared PMDs, the area to flush could be beyond
4847          * start/end.  Set range.start/range.end to cover the maximum possible
4848          * range if PMD sharing is possible.
4849          */
4850         mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
4851                                 0, vma, mm, start, end);
4852         adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4853
4854         BUG_ON(address >= end);
4855         flush_cache_range(vma, range.start, range.end);
4856
4857         mmu_notifier_invalidate_range_start(&range);
4858         i_mmap_lock_write(vma->vm_file->f_mapping);
4859         for (; address < end; address += huge_page_size(h)) {
4860                 spinlock_t *ptl;
4861                 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4862                 if (!ptep)
4863                         continue;
4864                 ptl = huge_pte_lock(h, mm, ptep);
4865                 if (huge_pmd_unshare(mm, &address, ptep)) {
4866                         pages++;
4867                         spin_unlock(ptl);
4868                         shared_pmd = true;
4869                         continue;
4870                 }
4871                 pte = huge_ptep_get(ptep);
4872                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4873                         spin_unlock(ptl);
4874                         continue;
4875                 }
4876                 if (unlikely(is_hugetlb_entry_migration(pte))) {
4877                         swp_entry_t entry = pte_to_swp_entry(pte);
4878
4879                         if (is_write_migration_entry(entry)) {
4880                                 pte_t newpte;
4881
4882                                 make_migration_entry_read(&entry);
4883                                 newpte = swp_entry_to_pte(entry);
4884                                 set_huge_swap_pte_at(mm, address, ptep,
4885                                                      newpte, huge_page_size(h));
4886                                 pages++;
4887                         }
4888                         spin_unlock(ptl);
4889                         continue;
4890                 }
4891                 if (!huge_pte_none(pte)) {
4892                         pte_t old_pte;
4893
4894                         old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
4895                         pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
4896                         pte = arch_make_huge_pte(pte, vma, NULL, 0);
4897                         huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
4898                         pages++;
4899                 }
4900                 spin_unlock(ptl);
4901         }
4902         /*
4903          * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4904          * may have cleared our pud entry and done put_page on the page table:
4905          * once we release i_mmap_rwsem, another task can do the final put_page
4906          * and that page table be reused and filled with junk.  If we actually
4907          * did unshare a page of pmds, flush the range corresponding to the pud.
4908          */
4909         if (shared_pmd)
4910                 flush_hugetlb_tlb_range(vma, range.start, range.end);
4911         else
4912                 flush_hugetlb_tlb_range(vma, start, end);
4913         /*
4914          * No need to call mmu_notifier_invalidate_range() we are downgrading
4915          * page table protection not changing it to point to a new page.
4916          *
4917          * See Documentation/vm/mmu_notifier.rst
4918          */
4919         i_mmap_unlock_write(vma->vm_file->f_mapping);
4920         mmu_notifier_invalidate_range_end(&range);
4921
4922         return pages << h->order;
4923 }
4924
4925 int hugetlb_reserve_pages(struct inode *inode,
4926                                         long from, long to,
4927                                         struct vm_area_struct *vma,
4928                                         vm_flags_t vm_flags)
4929 {
4930         long ret, chg, add = -1;
4931         struct hstate *h = hstate_inode(inode);
4932         struct hugepage_subpool *spool = subpool_inode(inode);
4933         struct resv_map *resv_map;
4934         struct hugetlb_cgroup *h_cg = NULL;
4935         long gbl_reserve, regions_needed = 0;
4936
4937         /* This should never happen */
4938         if (from > to) {
4939                 VM_WARN(1, "%s called with a negative range\n", __func__);
4940                 return -EINVAL;
4941         }
4942
4943         /*
4944          * Only apply hugepage reservation if asked. At fault time, an
4945          * attempt will be made for VM_NORESERVE to allocate a page
4946          * without using reserves
4947          */
4948         if (vm_flags & VM_NORESERVE)
4949                 return 0;
4950
4951         /*
4952          * Shared mappings base their reservation on the number of pages that
4953          * are already allocated on behalf of the file. Private mappings need
4954          * to reserve the full area even if read-only as mprotect() may be
4955          * called to make the mapping read-write. Assume !vma is a shm mapping
4956          */
4957         if (!vma || vma->vm_flags & VM_MAYSHARE) {
4958                 /*
4959                  * resv_map can not be NULL as hugetlb_reserve_pages is only
4960                  * called for inodes for which resv_maps were created (see
4961                  * hugetlbfs_get_inode).
4962                  */
4963                 resv_map = inode_resv_map(inode);
4964
4965                 chg = region_chg(resv_map, from, to, &regions_needed);
4966
4967         } else {
4968                 /* Private mapping. */
4969                 resv_map = resv_map_alloc();
4970                 if (!resv_map)
4971                         return -ENOMEM;
4972
4973                 chg = to - from;
4974
4975                 set_vma_resv_map(vma, resv_map);
4976                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4977         }
4978
4979         if (chg < 0) {
4980                 ret = chg;
4981                 goto out_err;
4982         }
4983
4984         ret = hugetlb_cgroup_charge_cgroup_rsvd(
4985                 hstate_index(h), chg * pages_per_huge_page(h), &h_cg);
4986
4987         if (ret < 0) {
4988                 ret = -ENOMEM;
4989                 goto out_err;
4990         }
4991
4992         if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
4993                 /* For private mappings, the hugetlb_cgroup uncharge info hangs
4994                  * of the resv_map.
4995                  */
4996                 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
4997         }
4998
4999         /*
5000          * There must be enough pages in the subpool for the mapping. If
5001          * the subpool has a minimum size, there may be some global
5002          * reservations already in place (gbl_reserve).
5003          */
5004         gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5005         if (gbl_reserve < 0) {
5006                 ret = -ENOSPC;
5007                 goto out_uncharge_cgroup;
5008         }
5009
5010         /*
5011          * Check enough hugepages are available for the reservation.
5012          * Hand the pages back to the subpool if there are not
5013          */
5014         ret = hugetlb_acct_memory(h, gbl_reserve);
5015         if (ret < 0) {
5016                 goto out_put_pages;
5017         }
5018
5019         /*
5020          * Account for the reservations made. Shared mappings record regions
5021          * that have reservations as they are shared by multiple VMAs.
5022          * When the last VMA disappears, the region map says how much
5023          * the reservation was and the page cache tells how much of
5024          * the reservation was consumed. Private mappings are per-VMA and
5025          * only the consumed reservations are tracked. When the VMA
5026          * disappears, the original reservation is the VMA size and the
5027          * consumed reservations are stored in the map. Hence, nothing
5028          * else has to be done for private mappings here
5029          */
5030         if (!vma || vma->vm_flags & VM_MAYSHARE) {
5031                 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5032
5033                 if (unlikely(add < 0)) {
5034                         hugetlb_acct_memory(h, -gbl_reserve);
5035                         goto out_put_pages;
5036                 } else if (unlikely(chg > add)) {
5037                         /*
5038                          * pages in this range were added to the reserve
5039                          * map between region_chg and region_add.  This
5040                          * indicates a race with alloc_huge_page.  Adjust
5041                          * the subpool and reserve counts modified above
5042                          * based on the difference.
5043                          */
5044                         long rsv_adjust;
5045
5046                         hugetlb_cgroup_uncharge_cgroup_rsvd(
5047                                 hstate_index(h),
5048                                 (chg - add) * pages_per_huge_page(h), h_cg);
5049
5050                         rsv_adjust = hugepage_subpool_put_pages(spool,
5051                                                                 chg - add);
5052                         hugetlb_acct_memory(h, -rsv_adjust);
5053                 }
5054         }
5055         return 0;
5056 out_put_pages:
5057         /* put back original number of pages, chg */
5058         (void)hugepage_subpool_put_pages(spool, chg);
5059 out_uncharge_cgroup:
5060         hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5061                                             chg * pages_per_huge_page(h), h_cg);
5062 out_err:
5063         if (!vma || vma->vm_flags & VM_MAYSHARE)
5064                 /* Only call region_abort if the region_chg succeeded but the
5065                  * region_add failed or didn't run.
5066                  */
5067                 if (chg >= 0 && add < 0)
5068                         region_abort(resv_map, from, to, regions_needed);
5069         if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5070                 kref_put(&resv_map->refs, resv_map_release);
5071         return ret;
5072 }
5073
5074 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5075                                                                 long freed)
5076 {
5077         struct hstate *h = hstate_inode(inode);
5078         struct resv_map *resv_map = inode_resv_map(inode);
5079         long chg = 0;
5080         struct hugepage_subpool *spool = subpool_inode(inode);
5081         long gbl_reserve;
5082
5083         /*
5084          * Since this routine can be called in the evict inode path for all
5085          * hugetlbfs inodes, resv_map could be NULL.
5086          */
5087         if (resv_map) {
5088                 chg = region_del(resv_map, start, end);
5089                 /*
5090                  * region_del() can fail in the rare case where a region
5091                  * must be split and another region descriptor can not be
5092                  * allocated.  If end == LONG_MAX, it will not fail.
5093                  */
5094                 if (chg < 0)
5095                         return chg;
5096         }
5097
5098         spin_lock(&inode->i_lock);
5099         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5100         spin_unlock(&inode->i_lock);
5101
5102         /*
5103          * If the subpool has a minimum size, the number of global
5104          * reservations to be released may be adjusted.
5105          */
5106         gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5107         hugetlb_acct_memory(h, -gbl_reserve);
5108
5109         return 0;
5110 }
5111
5112 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5113 static unsigned long page_table_shareable(struct vm_area_struct *svma,
5114                                 struct vm_area_struct *vma,
5115                                 unsigned long addr, pgoff_t idx)
5116 {
5117         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5118                                 svma->vm_start;
5119         unsigned long sbase = saddr & PUD_MASK;
5120         unsigned long s_end = sbase + PUD_SIZE;
5121
5122         /* Allow segments to share if only one is marked locked */
5123         unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5124         unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5125
5126         /*
5127          * match the virtual addresses, permission and the alignment of the
5128          * page table page.
5129          */
5130         if (pmd_index(addr) != pmd_index(saddr) ||
5131             vm_flags != svm_flags ||
5132             sbase < svma->vm_start || svma->vm_end < s_end)
5133                 return 0;
5134
5135         return saddr;
5136 }
5137
5138 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5139 {
5140         unsigned long base = addr & PUD_MASK;
5141         unsigned long end = base + PUD_SIZE;
5142
5143         /*
5144          * check on proper vm_flags and page table alignment
5145          */
5146         if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5147                 return true;
5148         return false;
5149 }
5150
5151 /*
5152  * Determine if start,end range within vma could be mapped by shared pmd.
5153  * If yes, adjust start and end to cover range associated with possible
5154  * shared pmd mappings.
5155  */
5156 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5157                                 unsigned long *start, unsigned long *end)
5158 {
5159         unsigned long check_addr;
5160
5161         if (!(vma->vm_flags & VM_MAYSHARE))
5162                 return;
5163
5164         for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
5165                 unsigned long a_start = check_addr & PUD_MASK;
5166                 unsigned long a_end = a_start + PUD_SIZE;
5167
5168                 /*
5169                  * If sharing is possible, adjust start/end if necessary.
5170                  */
5171                 if (range_in_vma(vma, a_start, a_end)) {
5172                         if (a_start < *start)
5173                                 *start = a_start;
5174                         if (a_end > *end)
5175                                 *end = a_end;
5176                 }
5177         }
5178 }
5179
5180 /*
5181  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5182  * and returns the corresponding pte. While this is not necessary for the
5183  * !shared pmd case because we can allocate the pmd later as well, it makes the
5184  * code much cleaner.
5185  *
5186  * This routine must be called with i_mmap_rwsem held in at least read mode.
5187  * For hugetlbfs, this prevents removal of any page table entries associated
5188  * with the address space.  This is important as we are setting up sharing
5189  * based on existing page table entries (mappings).
5190  */
5191 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5192 {
5193         struct vm_area_struct *vma = find_vma(mm, addr);
5194         struct address_space *mapping = vma->vm_file->f_mapping;
5195         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5196                         vma->vm_pgoff;
5197         struct vm_area_struct *svma;
5198         unsigned long saddr;
5199         pte_t *spte = NULL;
5200         pte_t *pte;
5201         spinlock_t *ptl;
5202
5203         if (!vma_shareable(vma, addr))
5204                 return (pte_t *)pmd_alloc(mm, pud, addr);
5205
5206         vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5207                 if (svma == vma)
5208                         continue;
5209
5210                 saddr = page_table_shareable(svma, vma, addr, idx);
5211                 if (saddr) {
5212                         spte = huge_pte_offset(svma->vm_mm, saddr,
5213                                                vma_mmu_pagesize(svma));
5214                         if (spte) {
5215                                 get_page(virt_to_page(spte));
5216                                 break;
5217                         }
5218                 }
5219         }
5220
5221         if (!spte)
5222                 goto out;
5223
5224         ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
5225         if (pud_none(*pud)) {
5226                 pud_populate(mm, pud,
5227                                 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5228                 mm_inc_nr_pmds(mm);
5229         } else {
5230                 put_page(virt_to_page(spte));
5231         }
5232         spin_unlock(ptl);
5233 out:
5234         pte = (pte_t *)pmd_alloc(mm, pud, addr);
5235         return pte;
5236 }
5237
5238 /*
5239  * unmap huge page backed by shared pte.
5240  *
5241  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
5242  * indicated by page_count > 1, unmap is achieved by clearing pud and
5243  * decrementing the ref count. If count == 1, the pte page is not shared.
5244  *
5245  * Called with page table lock held and i_mmap_rwsem held in write mode.
5246  *
5247  * returns: 1 successfully unmapped a shared pte page
5248  *          0 the underlying pte page is not shared, or it is the last user
5249  */
5250 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
5251 {
5252         pgd_t *pgd = pgd_offset(mm, *addr);
5253         p4d_t *p4d = p4d_offset(pgd, *addr);
5254         pud_t *pud = pud_offset(p4d, *addr);
5255
5256         BUG_ON(page_count(virt_to_page(ptep)) == 0);
5257         if (page_count(virt_to_page(ptep)) == 1)
5258                 return 0;
5259
5260         pud_clear(pud);
5261         put_page(virt_to_page(ptep));
5262         mm_dec_nr_pmds(mm);
5263         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5264         return 1;
5265 }
5266 #define want_pmd_share()        (1)
5267 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5268 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5269 {
5270         return NULL;
5271 }
5272
5273 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
5274 {
5275         return 0;
5276 }
5277
5278 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5279                                 unsigned long *start, unsigned long *end)
5280 {
5281 }
5282 #define want_pmd_share()        (0)
5283 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5284
5285 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5286 pte_t *huge_pte_alloc(struct mm_struct *mm,
5287                         unsigned long addr, unsigned long sz)
5288 {
5289         pgd_t *pgd;
5290         p4d_t *p4d;
5291         pud_t *pud;
5292         pte_t *pte = NULL;
5293
5294         pgd = pgd_offset(mm, addr);
5295         p4d = p4d_alloc(mm, pgd, addr);
5296         if (!p4d)
5297                 return NULL;
5298         pud = pud_alloc(mm, p4d, addr);
5299         if (pud) {
5300                 if (sz == PUD_SIZE) {
5301                         pte = (pte_t *)pud;
5302                 } else {
5303                         BUG_ON(sz != PMD_SIZE);
5304                         if (want_pmd_share() && pud_none(*pud))
5305                                 pte = huge_pmd_share(mm, addr, pud);
5306                         else
5307                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5308                 }
5309         }
5310         BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
5311
5312         return pte;
5313 }
5314
5315 /*
5316  * huge_pte_offset() - Walk the page table to resolve the hugepage
5317  * entry at address @addr
5318  *
5319  * Return: Pointer to page table or swap entry (PUD or PMD) for
5320  * address @addr, or NULL if a p*d_none() entry is encountered and the
5321  * size @sz doesn't match the hugepage size at this level of the page
5322  * table.
5323  */
5324 pte_t *huge_pte_offset(struct mm_struct *mm,
5325                        unsigned long addr, unsigned long sz)
5326 {
5327         pgd_t *pgd;
5328         p4d_t *p4d;
5329         pud_t *pud;
5330         pmd_t *pmd;
5331
5332         pgd = pgd_offset(mm, addr);
5333         if (!pgd_present(*pgd))
5334                 return NULL;
5335         p4d = p4d_offset(pgd, addr);
5336         if (!p4d_present(*p4d))
5337                 return NULL;
5338
5339         pud = pud_offset(p4d, addr);
5340         if (sz != PUD_SIZE && pud_none(*pud))
5341                 return NULL;
5342         /* hugepage or swap? */
5343         if (pud_huge(*pud) || !pud_present(*pud))
5344                 return (pte_t *)pud;
5345
5346         pmd = pmd_offset(pud, addr);
5347         if (sz != PMD_SIZE && pmd_none(*pmd))
5348                 return NULL;
5349         /* hugepage or swap? */
5350         if (pmd_huge(*pmd) || !pmd_present(*pmd))
5351                 return (pte_t *)pmd;
5352
5353         return NULL;
5354 }
5355
5356 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5357
5358 /*
5359  * These functions are overwritable if your architecture needs its own
5360  * behavior.
5361  */
5362 struct page * __weak
5363 follow_huge_addr(struct mm_struct *mm, unsigned long address,
5364                               int write)
5365 {
5366         return ERR_PTR(-EINVAL);
5367 }
5368
5369 struct page * __weak
5370 follow_huge_pd(struct vm_area_struct *vma,
5371                unsigned long address, hugepd_t hpd, int flags, int pdshift)
5372 {
5373         WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5374         return NULL;
5375 }
5376
5377 struct page * __weak
5378 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
5379                 pmd_t *pmd, int flags)
5380 {
5381         struct page *page = NULL;
5382         spinlock_t *ptl;
5383         pte_t pte;
5384
5385         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5386         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
5387                          (FOLL_PIN | FOLL_GET)))
5388                 return NULL;
5389
5390 retry:
5391         ptl = pmd_lockptr(mm, pmd);
5392         spin_lock(ptl);
5393         /*
5394          * make sure that the address range covered by this pmd is not
5395          * unmapped from other threads.
5396          */
5397         if (!pmd_huge(*pmd))
5398                 goto out;
5399         pte = huge_ptep_get((pte_t *)pmd);
5400         if (pte_present(pte)) {
5401                 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
5402                 /*
5403                  * try_grab_page() should always succeed here, because: a) we
5404                  * hold the pmd (ptl) lock, and b) we've just checked that the
5405                  * huge pmd (head) page is present in the page tables. The ptl
5406                  * prevents the head page and tail pages from being rearranged
5407                  * in any way. So this page must be available at this point,
5408                  * unless the page refcount overflowed:
5409                  */
5410                 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5411                         page = NULL;
5412                         goto out;
5413                 }
5414         } else {
5415                 if (is_hugetlb_entry_migration(pte)) {
5416                         spin_unlock(ptl);
5417                         __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5418                         goto retry;
5419                 }
5420                 /*
5421                  * hwpoisoned entry is treated as no_page_table in
5422                  * follow_page_mask().
5423                  */
5424         }
5425 out:
5426         spin_unlock(ptl);
5427         return page;
5428 }
5429
5430 struct page * __weak
5431 follow_huge_pud(struct mm_struct *mm, unsigned long address,
5432                 pud_t *pud, int flags)
5433 {
5434         if (flags & (FOLL_GET | FOLL_PIN))
5435                 return NULL;
5436
5437         return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5438 }
5439
5440 struct page * __weak
5441 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5442 {
5443         if (flags & (FOLL_GET | FOLL_PIN))
5444                 return NULL;
5445
5446         return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5447 }
5448
5449 bool isolate_huge_page(struct page *page, struct list_head *list)
5450 {
5451         bool ret = true;
5452
5453         VM_BUG_ON_PAGE(!PageHead(page), page);
5454         spin_lock(&hugetlb_lock);
5455         if (!page_huge_active(page) || !get_page_unless_zero(page)) {
5456                 ret = false;
5457                 goto unlock;
5458         }
5459         clear_page_huge_active(page);
5460         list_move_tail(&page->lru, list);
5461 unlock:
5462         spin_unlock(&hugetlb_lock);
5463         return ret;
5464 }
5465
5466 void putback_active_hugepage(struct page *page)
5467 {
5468         VM_BUG_ON_PAGE(!PageHead(page), page);
5469         spin_lock(&hugetlb_lock);
5470         set_page_huge_active(page);
5471         list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5472         spin_unlock(&hugetlb_lock);
5473         put_page(page);
5474 }
5475
5476 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5477 {
5478         struct hstate *h = page_hstate(oldpage);
5479
5480         hugetlb_cgroup_migrate(oldpage, newpage);
5481         set_page_owner_migrate_reason(newpage, reason);
5482
5483         /*
5484          * transfer temporary state of the new huge page. This is
5485          * reverse to other transitions because the newpage is going to
5486          * be final while the old one will be freed so it takes over
5487          * the temporary status.
5488          *
5489          * Also note that we have to transfer the per-node surplus state
5490          * here as well otherwise the global surplus count will not match
5491          * the per-node's.
5492          */
5493         if (PageHugeTemporary(newpage)) {
5494                 int old_nid = page_to_nid(oldpage);
5495                 int new_nid = page_to_nid(newpage);
5496
5497                 SetPageHugeTemporary(oldpage);
5498                 ClearPageHugeTemporary(newpage);
5499
5500                 spin_lock(&hugetlb_lock);
5501                 if (h->surplus_huge_pages_node[old_nid]) {
5502                         h->surplus_huge_pages_node[old_nid]--;
5503                         h->surplus_huge_pages_node[new_nid]++;
5504                 }
5505                 spin_unlock(&hugetlb_lock);
5506         }
5507 }