Merge tag 'sound-fix-5.7-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/tiwai...
[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         __must_hold(&hugetlb_lock)
2014 {
2015         struct list_head surplus_list;
2016         struct page *page, *tmp;
2017         int ret, i;
2018         int needed, allocated;
2019         bool alloc_ok = true;
2020
2021         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2022         if (needed <= 0) {
2023                 h->resv_huge_pages += delta;
2024                 return 0;
2025         }
2026
2027         allocated = 0;
2028         INIT_LIST_HEAD(&surplus_list);
2029
2030         ret = -ENOMEM;
2031 retry:
2032         spin_unlock(&hugetlb_lock);
2033         for (i = 0; i < needed; i++) {
2034                 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2035                                 NUMA_NO_NODE, NULL);
2036                 if (!page) {
2037                         alloc_ok = false;
2038                         break;
2039                 }
2040                 list_add(&page->lru, &surplus_list);
2041                 cond_resched();
2042         }
2043         allocated += i;
2044
2045         /*
2046          * After retaking hugetlb_lock, we need to recalculate 'needed'
2047          * because either resv_huge_pages or free_huge_pages may have changed.
2048          */
2049         spin_lock(&hugetlb_lock);
2050         needed = (h->resv_huge_pages + delta) -
2051                         (h->free_huge_pages + allocated);
2052         if (needed > 0) {
2053                 if (alloc_ok)
2054                         goto retry;
2055                 /*
2056                  * We were not able to allocate enough pages to
2057                  * satisfy the entire reservation so we free what
2058                  * we've allocated so far.
2059                  */
2060                 goto free;
2061         }
2062         /*
2063          * The surplus_list now contains _at_least_ the number of extra pages
2064          * needed to accommodate the reservation.  Add the appropriate number
2065          * of pages to the hugetlb pool and free the extras back to the buddy
2066          * allocator.  Commit the entire reservation here to prevent another
2067          * process from stealing the pages as they are added to the pool but
2068          * before they are reserved.
2069          */
2070         needed += allocated;
2071         h->resv_huge_pages += delta;
2072         ret = 0;
2073
2074         /* Free the needed pages to the hugetlb pool */
2075         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2076                 if ((--needed) < 0)
2077                         break;
2078                 /*
2079                  * This page is now managed by the hugetlb allocator and has
2080                  * no users -- drop the buddy allocator's reference.
2081                  */
2082                 put_page_testzero(page);
2083                 VM_BUG_ON_PAGE(page_count(page), page);
2084                 enqueue_huge_page(h, page);
2085         }
2086 free:
2087         spin_unlock(&hugetlb_lock);
2088
2089         /* Free unnecessary surplus pages to the buddy allocator */
2090         list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2091                 put_page(page);
2092         spin_lock(&hugetlb_lock);
2093
2094         return ret;
2095 }
2096
2097 /*
2098  * This routine has two main purposes:
2099  * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2100  *    in unused_resv_pages.  This corresponds to the prior adjustments made
2101  *    to the associated reservation map.
2102  * 2) Free any unused surplus pages that may have been allocated to satisfy
2103  *    the reservation.  As many as unused_resv_pages may be freed.
2104  *
2105  * Called with hugetlb_lock held.  However, the lock could be dropped (and
2106  * reacquired) during calls to cond_resched_lock.  Whenever dropping the lock,
2107  * we must make sure nobody else can claim pages we are in the process of
2108  * freeing.  Do this by ensuring resv_huge_page always is greater than the
2109  * number of huge pages we plan to free when dropping the lock.
2110  */
2111 static void return_unused_surplus_pages(struct hstate *h,
2112                                         unsigned long unused_resv_pages)
2113 {
2114         unsigned long nr_pages;
2115
2116         /* Cannot return gigantic pages currently */
2117         if (hstate_is_gigantic(h))
2118                 goto out;
2119
2120         /*
2121          * Part (or even all) of the reservation could have been backed
2122          * by pre-allocated pages. Only free surplus pages.
2123          */
2124         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2125
2126         /*
2127          * We want to release as many surplus pages as possible, spread
2128          * evenly across all nodes with memory. Iterate across these nodes
2129          * until we can no longer free unreserved surplus pages. This occurs
2130          * when the nodes with surplus pages have no free pages.
2131          * free_pool_huge_page() will balance the the freed pages across the
2132          * on-line nodes with memory and will handle the hstate accounting.
2133          *
2134          * Note that we decrement resv_huge_pages as we free the pages.  If
2135          * we drop the lock, resv_huge_pages will still be sufficiently large
2136          * to cover subsequent pages we may free.
2137          */
2138         while (nr_pages--) {
2139                 h->resv_huge_pages--;
2140                 unused_resv_pages--;
2141                 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
2142                         goto out;
2143                 cond_resched_lock(&hugetlb_lock);
2144         }
2145
2146 out:
2147         /* Fully uncommit the reservation */
2148         h->resv_huge_pages -= unused_resv_pages;
2149 }
2150
2151
2152 /*
2153  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2154  * are used by the huge page allocation routines to manage reservations.
2155  *
2156  * vma_needs_reservation is called to determine if the huge page at addr
2157  * within the vma has an associated reservation.  If a reservation is
2158  * needed, the value 1 is returned.  The caller is then responsible for
2159  * managing the global reservation and subpool usage counts.  After
2160  * the huge page has been allocated, vma_commit_reservation is called
2161  * to add the page to the reservation map.  If the page allocation fails,
2162  * the reservation must be ended instead of committed.  vma_end_reservation
2163  * is called in such cases.
2164  *
2165  * In the normal case, vma_commit_reservation returns the same value
2166  * as the preceding vma_needs_reservation call.  The only time this
2167  * is not the case is if a reserve map was changed between calls.  It
2168  * is the responsibility of the caller to notice the difference and
2169  * take appropriate action.
2170  *
2171  * vma_add_reservation is used in error paths where a reservation must
2172  * be restored when a newly allocated huge page must be freed.  It is
2173  * to be called after calling vma_needs_reservation to determine if a
2174  * reservation exists.
2175  */
2176 enum vma_resv_mode {
2177         VMA_NEEDS_RESV,
2178         VMA_COMMIT_RESV,
2179         VMA_END_RESV,
2180         VMA_ADD_RESV,
2181 };
2182 static long __vma_reservation_common(struct hstate *h,
2183                                 struct vm_area_struct *vma, unsigned long addr,
2184                                 enum vma_resv_mode mode)
2185 {
2186         struct resv_map *resv;
2187         pgoff_t idx;
2188         long ret;
2189         long dummy_out_regions_needed;
2190
2191         resv = vma_resv_map(vma);
2192         if (!resv)
2193                 return 1;
2194
2195         idx = vma_hugecache_offset(h, vma, addr);
2196         switch (mode) {
2197         case VMA_NEEDS_RESV:
2198                 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2199                 /* We assume that vma_reservation_* routines always operate on
2200                  * 1 page, and that adding to resv map a 1 page entry can only
2201                  * ever require 1 region.
2202                  */
2203                 VM_BUG_ON(dummy_out_regions_needed != 1);
2204                 break;
2205         case VMA_COMMIT_RESV:
2206                 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2207                 /* region_add calls of range 1 should never fail. */
2208                 VM_BUG_ON(ret < 0);
2209                 break;
2210         case VMA_END_RESV:
2211                 region_abort(resv, idx, idx + 1, 1);
2212                 ret = 0;
2213                 break;
2214         case VMA_ADD_RESV:
2215                 if (vma->vm_flags & VM_MAYSHARE) {
2216                         ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2217                         /* region_add calls of range 1 should never fail. */
2218                         VM_BUG_ON(ret < 0);
2219                 } else {
2220                         region_abort(resv, idx, idx + 1, 1);
2221                         ret = region_del(resv, idx, idx + 1);
2222                 }
2223                 break;
2224         default:
2225                 BUG();
2226         }
2227
2228         if (vma->vm_flags & VM_MAYSHARE)
2229                 return ret;
2230         else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
2231                 /*
2232                  * In most cases, reserves always exist for private mappings.
2233                  * However, a file associated with mapping could have been
2234                  * hole punched or truncated after reserves were consumed.
2235                  * As subsequent fault on such a range will not use reserves.
2236                  * Subtle - The reserve map for private mappings has the
2237                  * opposite meaning than that of shared mappings.  If NO
2238                  * entry is in the reserve map, it means a reservation exists.
2239                  * If an entry exists in the reserve map, it means the
2240                  * reservation has already been consumed.  As a result, the
2241                  * return value of this routine is the opposite of the
2242                  * value returned from reserve map manipulation routines above.
2243                  */
2244                 if (ret)
2245                         return 0;
2246                 else
2247                         return 1;
2248         }
2249         else
2250                 return ret < 0 ? ret : 0;
2251 }
2252
2253 static long vma_needs_reservation(struct hstate *h,
2254                         struct vm_area_struct *vma, unsigned long addr)
2255 {
2256         return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2257 }
2258
2259 static long vma_commit_reservation(struct hstate *h,
2260                         struct vm_area_struct *vma, unsigned long addr)
2261 {
2262         return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2263 }
2264
2265 static void vma_end_reservation(struct hstate *h,
2266                         struct vm_area_struct *vma, unsigned long addr)
2267 {
2268         (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2269 }
2270
2271 static long vma_add_reservation(struct hstate *h,
2272                         struct vm_area_struct *vma, unsigned long addr)
2273 {
2274         return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2275 }
2276
2277 /*
2278  * This routine is called to restore a reservation on error paths.  In the
2279  * specific error paths, a huge page was allocated (via alloc_huge_page)
2280  * and is about to be freed.  If a reservation for the page existed,
2281  * alloc_huge_page would have consumed the reservation and set PagePrivate
2282  * in the newly allocated page.  When the page is freed via free_huge_page,
2283  * the global reservation count will be incremented if PagePrivate is set.
2284  * However, free_huge_page can not adjust the reserve map.  Adjust the
2285  * reserve map here to be consistent with global reserve count adjustments
2286  * to be made by free_huge_page.
2287  */
2288 static void restore_reserve_on_error(struct hstate *h,
2289                         struct vm_area_struct *vma, unsigned long address,
2290                         struct page *page)
2291 {
2292         if (unlikely(PagePrivate(page))) {
2293                 long rc = vma_needs_reservation(h, vma, address);
2294
2295                 if (unlikely(rc < 0)) {
2296                         /*
2297                          * Rare out of memory condition in reserve map
2298                          * manipulation.  Clear PagePrivate so that
2299                          * global reserve count will not be incremented
2300                          * by free_huge_page.  This will make it appear
2301                          * as though the reservation for this page was
2302                          * consumed.  This may prevent the task from
2303                          * faulting in the page at a later time.  This
2304                          * is better than inconsistent global huge page
2305                          * accounting of reserve counts.
2306                          */
2307                         ClearPagePrivate(page);
2308                 } else if (rc) {
2309                         rc = vma_add_reservation(h, vma, address);
2310                         if (unlikely(rc < 0))
2311                                 /*
2312                                  * See above comment about rare out of
2313                                  * memory condition.
2314                                  */
2315                                 ClearPagePrivate(page);
2316                 } else
2317                         vma_end_reservation(h, vma, address);
2318         }
2319 }
2320
2321 struct page *alloc_huge_page(struct vm_area_struct *vma,
2322                                     unsigned long addr, int avoid_reserve)
2323 {
2324         struct hugepage_subpool *spool = subpool_vma(vma);
2325         struct hstate *h = hstate_vma(vma);
2326         struct page *page;
2327         long map_chg, map_commit;
2328         long gbl_chg;
2329         int ret, idx;
2330         struct hugetlb_cgroup *h_cg;
2331         bool deferred_reserve;
2332
2333         idx = hstate_index(h);
2334         /*
2335          * Examine the region/reserve map to determine if the process
2336          * has a reservation for the page to be allocated.  A return
2337          * code of zero indicates a reservation exists (no change).
2338          */
2339         map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2340         if (map_chg < 0)
2341                 return ERR_PTR(-ENOMEM);
2342
2343         /*
2344          * Processes that did not create the mapping will have no
2345          * reserves as indicated by the region/reserve map. Check
2346          * that the allocation will not exceed the subpool limit.
2347          * Allocations for MAP_NORESERVE mappings also need to be
2348          * checked against any subpool limit.
2349          */
2350         if (map_chg || avoid_reserve) {
2351                 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2352                 if (gbl_chg < 0) {
2353                         vma_end_reservation(h, vma, addr);
2354                         return ERR_PTR(-ENOSPC);
2355                 }
2356
2357                 /*
2358                  * Even though there was no reservation in the region/reserve
2359                  * map, there could be reservations associated with the
2360                  * subpool that can be used.  This would be indicated if the
2361                  * return value of hugepage_subpool_get_pages() is zero.
2362                  * However, if avoid_reserve is specified we still avoid even
2363                  * the subpool reservations.
2364                  */
2365                 if (avoid_reserve)
2366                         gbl_chg = 1;
2367         }
2368
2369         /* If this allocation is not consuming a reservation, charge it now.
2370          */
2371         deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma);
2372         if (deferred_reserve) {
2373                 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2374                         idx, pages_per_huge_page(h), &h_cg);
2375                 if (ret)
2376                         goto out_subpool_put;
2377         }
2378
2379         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2380         if (ret)
2381                 goto out_uncharge_cgroup_reservation;
2382
2383         spin_lock(&hugetlb_lock);
2384         /*
2385          * glb_chg is passed to indicate whether or not a page must be taken
2386          * from the global free pool (global change).  gbl_chg == 0 indicates
2387          * a reservation exists for the allocation.
2388          */
2389         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2390         if (!page) {
2391                 spin_unlock(&hugetlb_lock);
2392                 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2393                 if (!page)
2394                         goto out_uncharge_cgroup;
2395                 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2396                         SetPagePrivate(page);
2397                         h->resv_huge_pages--;
2398                 }
2399                 spin_lock(&hugetlb_lock);
2400                 list_move(&page->lru, &h->hugepage_activelist);
2401                 /* Fall through */
2402         }
2403         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2404         /* If allocation is not consuming a reservation, also store the
2405          * hugetlb_cgroup pointer on the page.
2406          */
2407         if (deferred_reserve) {
2408                 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2409                                                   h_cg, page);
2410         }
2411
2412         spin_unlock(&hugetlb_lock);
2413
2414         set_page_private(page, (unsigned long)spool);
2415
2416         map_commit = vma_commit_reservation(h, vma, addr);
2417         if (unlikely(map_chg > map_commit)) {
2418                 /*
2419                  * The page was added to the reservation map between
2420                  * vma_needs_reservation and vma_commit_reservation.
2421                  * This indicates a race with hugetlb_reserve_pages.
2422                  * Adjust for the subpool count incremented above AND
2423                  * in hugetlb_reserve_pages for the same page.  Also,
2424                  * the reservation count added in hugetlb_reserve_pages
2425                  * no longer applies.
2426                  */
2427                 long rsv_adjust;
2428
2429                 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2430                 hugetlb_acct_memory(h, -rsv_adjust);
2431         }
2432         return page;
2433
2434 out_uncharge_cgroup:
2435         hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2436 out_uncharge_cgroup_reservation:
2437         if (deferred_reserve)
2438                 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2439                                                     h_cg);
2440 out_subpool_put:
2441         if (map_chg || avoid_reserve)
2442                 hugepage_subpool_put_pages(spool, 1);
2443         vma_end_reservation(h, vma, addr);
2444         return ERR_PTR(-ENOSPC);
2445 }
2446
2447 int alloc_bootmem_huge_page(struct hstate *h)
2448         __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2449 int __alloc_bootmem_huge_page(struct hstate *h)
2450 {
2451         struct huge_bootmem_page *m;
2452         int nr_nodes, node;
2453
2454         for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2455                 void *addr;
2456
2457                 addr = memblock_alloc_try_nid_raw(
2458                                 huge_page_size(h), huge_page_size(h),
2459                                 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2460                 if (addr) {
2461                         /*
2462                          * Use the beginning of the huge page to store the
2463                          * huge_bootmem_page struct (until gather_bootmem
2464                          * puts them into the mem_map).
2465                          */
2466                         m = addr;
2467                         goto found;
2468                 }
2469         }
2470         return 0;
2471
2472 found:
2473         BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2474         /* Put them into a private list first because mem_map is not up yet */
2475         INIT_LIST_HEAD(&m->list);
2476         list_add(&m->list, &huge_boot_pages);
2477         m->hstate = h;
2478         return 1;
2479 }
2480
2481 static void __init prep_compound_huge_page(struct page *page,
2482                 unsigned int order)
2483 {
2484         if (unlikely(order > (MAX_ORDER - 1)))
2485                 prep_compound_gigantic_page(page, order);
2486         else
2487                 prep_compound_page(page, order);
2488 }
2489
2490 /* Put bootmem huge pages into the standard lists after mem_map is up */
2491 static void __init gather_bootmem_prealloc(void)
2492 {
2493         struct huge_bootmem_page *m;
2494
2495         list_for_each_entry(m, &huge_boot_pages, list) {
2496                 struct page *page = virt_to_page(m);
2497                 struct hstate *h = m->hstate;
2498
2499                 WARN_ON(page_count(page) != 1);
2500                 prep_compound_huge_page(page, h->order);
2501                 WARN_ON(PageReserved(page));
2502                 prep_new_huge_page(h, page, page_to_nid(page));
2503                 put_page(page); /* free it into the hugepage allocator */
2504
2505                 /*
2506                  * If we had gigantic hugepages allocated at boot time, we need
2507                  * to restore the 'stolen' pages to totalram_pages in order to
2508                  * fix confusing memory reports from free(1) and another
2509                  * side-effects, like CommitLimit going negative.
2510                  */
2511                 if (hstate_is_gigantic(h))
2512                         adjust_managed_page_count(page, 1 << h->order);
2513                 cond_resched();
2514         }
2515 }
2516
2517 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2518 {
2519         unsigned long i;
2520         nodemask_t *node_alloc_noretry;
2521
2522         if (!hstate_is_gigantic(h)) {
2523                 /*
2524                  * Bit mask controlling how hard we retry per-node allocations.
2525                  * Ignore errors as lower level routines can deal with
2526                  * node_alloc_noretry == NULL.  If this kmalloc fails at boot
2527                  * time, we are likely in bigger trouble.
2528                  */
2529                 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2530                                                 GFP_KERNEL);
2531         } else {
2532                 /* allocations done at boot time */
2533                 node_alloc_noretry = NULL;
2534         }
2535
2536         /* bit mask controlling how hard we retry per-node allocations */
2537         if (node_alloc_noretry)
2538                 nodes_clear(*node_alloc_noretry);
2539
2540         for (i = 0; i < h->max_huge_pages; ++i) {
2541                 if (hstate_is_gigantic(h)) {
2542                         if (!alloc_bootmem_huge_page(h))
2543                                 break;
2544                 } else if (!alloc_pool_huge_page(h,
2545                                          &node_states[N_MEMORY],
2546                                          node_alloc_noretry))
2547                         break;
2548                 cond_resched();
2549         }
2550         if (i < h->max_huge_pages) {
2551                 char buf[32];
2552
2553                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2554                 pr_warn("HugeTLB: allocating %lu of page size %s failed.  Only allocated %lu hugepages.\n",
2555                         h->max_huge_pages, buf, i);
2556                 h->max_huge_pages = i;
2557         }
2558
2559         kfree(node_alloc_noretry);
2560 }
2561
2562 static void __init hugetlb_init_hstates(void)
2563 {
2564         struct hstate *h;
2565
2566         for_each_hstate(h) {
2567                 if (minimum_order > huge_page_order(h))
2568                         minimum_order = huge_page_order(h);
2569
2570                 /* oversize hugepages were init'ed in early boot */
2571                 if (!hstate_is_gigantic(h))
2572                         hugetlb_hstate_alloc_pages(h);
2573         }
2574         VM_BUG_ON(minimum_order == UINT_MAX);
2575 }
2576
2577 static void __init report_hugepages(void)
2578 {
2579         struct hstate *h;
2580
2581         for_each_hstate(h) {
2582                 char buf[32];
2583
2584                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2585                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2586                         buf, h->free_huge_pages);
2587         }
2588 }
2589
2590 #ifdef CONFIG_HIGHMEM
2591 static void try_to_free_low(struct hstate *h, unsigned long count,
2592                                                 nodemask_t *nodes_allowed)
2593 {
2594         int i;
2595
2596         if (hstate_is_gigantic(h))
2597                 return;
2598
2599         for_each_node_mask(i, *nodes_allowed) {
2600                 struct page *page, *next;
2601                 struct list_head *freel = &h->hugepage_freelists[i];
2602                 list_for_each_entry_safe(page, next, freel, lru) {
2603                         if (count >= h->nr_huge_pages)
2604                                 return;
2605                         if (PageHighMem(page))
2606                                 continue;
2607                         list_del(&page->lru);
2608                         update_and_free_page(h, page);
2609                         h->free_huge_pages--;
2610                         h->free_huge_pages_node[page_to_nid(page)]--;
2611                 }
2612         }
2613 }
2614 #else
2615 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2616                                                 nodemask_t *nodes_allowed)
2617 {
2618 }
2619 #endif
2620
2621 /*
2622  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
2623  * balanced by operating on them in a round-robin fashion.
2624  * Returns 1 if an adjustment was made.
2625  */
2626 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2627                                 int delta)
2628 {
2629         int nr_nodes, node;
2630
2631         VM_BUG_ON(delta != -1 && delta != 1);
2632
2633         if (delta < 0) {
2634                 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2635                         if (h->surplus_huge_pages_node[node])
2636                                 goto found;
2637                 }
2638         } else {
2639                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2640                         if (h->surplus_huge_pages_node[node] <
2641                                         h->nr_huge_pages_node[node])
2642                                 goto found;
2643                 }
2644         }
2645         return 0;
2646
2647 found:
2648         h->surplus_huge_pages += delta;
2649         h->surplus_huge_pages_node[node] += delta;
2650         return 1;
2651 }
2652
2653 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2654 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2655                               nodemask_t *nodes_allowed)
2656 {
2657         unsigned long min_count, ret;
2658         NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2659
2660         /*
2661          * Bit mask controlling how hard we retry per-node allocations.
2662          * If we can not allocate the bit mask, do not attempt to allocate
2663          * the requested huge pages.
2664          */
2665         if (node_alloc_noretry)
2666                 nodes_clear(*node_alloc_noretry);
2667         else
2668                 return -ENOMEM;
2669
2670         spin_lock(&hugetlb_lock);
2671
2672         /*
2673          * Check for a node specific request.
2674          * Changing node specific huge page count may require a corresponding
2675          * change to the global count.  In any case, the passed node mask
2676          * (nodes_allowed) will restrict alloc/free to the specified node.
2677          */
2678         if (nid != NUMA_NO_NODE) {
2679                 unsigned long old_count = count;
2680
2681                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2682                 /*
2683                  * User may have specified a large count value which caused the
2684                  * above calculation to overflow.  In this case, they wanted
2685                  * to allocate as many huge pages as possible.  Set count to
2686                  * largest possible value to align with their intention.
2687                  */
2688                 if (count < old_count)
2689                         count = ULONG_MAX;
2690         }
2691
2692         /*
2693          * Gigantic pages runtime allocation depend on the capability for large
2694          * page range allocation.
2695          * If the system does not provide this feature, return an error when
2696          * the user tries to allocate gigantic pages but let the user free the
2697          * boottime allocated gigantic pages.
2698          */
2699         if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2700                 if (count > persistent_huge_pages(h)) {
2701                         spin_unlock(&hugetlb_lock);
2702                         NODEMASK_FREE(node_alloc_noretry);
2703                         return -EINVAL;
2704                 }
2705                 /* Fall through to decrease pool */
2706         }
2707
2708         /*
2709          * Increase the pool size
2710          * First take pages out of surplus state.  Then make up the
2711          * remaining difference by allocating fresh huge pages.
2712          *
2713          * We might race with alloc_surplus_huge_page() here and be unable
2714          * to convert a surplus huge page to a normal huge page. That is
2715          * not critical, though, it just means the overall size of the
2716          * pool might be one hugepage larger than it needs to be, but
2717          * within all the constraints specified by the sysctls.
2718          */
2719         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2720                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2721                         break;
2722         }
2723
2724         while (count > persistent_huge_pages(h)) {
2725                 /*
2726                  * If this allocation races such that we no longer need the
2727                  * page, free_huge_page will handle it by freeing the page
2728                  * and reducing the surplus.
2729                  */
2730                 spin_unlock(&hugetlb_lock);
2731
2732                 /* yield cpu to avoid soft lockup */
2733                 cond_resched();
2734
2735                 ret = alloc_pool_huge_page(h, nodes_allowed,
2736                                                 node_alloc_noretry);
2737                 spin_lock(&hugetlb_lock);
2738                 if (!ret)
2739                         goto out;
2740
2741                 /* Bail for signals. Probably ctrl-c from user */
2742                 if (signal_pending(current))
2743                         goto out;
2744         }
2745
2746         /*
2747          * Decrease the pool size
2748          * First return free pages to the buddy allocator (being careful
2749          * to keep enough around to satisfy reservations).  Then place
2750          * pages into surplus state as needed so the pool will shrink
2751          * to the desired size as pages become free.
2752          *
2753          * By placing pages into the surplus state independent of the
2754          * overcommit value, we are allowing the surplus pool size to
2755          * exceed overcommit. There are few sane options here. Since
2756          * alloc_surplus_huge_page() is checking the global counter,
2757          * though, we'll note that we're not allowed to exceed surplus
2758          * and won't grow the pool anywhere else. Not until one of the
2759          * sysctls are changed, or the surplus pages go out of use.
2760          */
2761         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2762         min_count = max(count, min_count);
2763         try_to_free_low(h, min_count, nodes_allowed);
2764         while (min_count < persistent_huge_pages(h)) {
2765                 if (!free_pool_huge_page(h, nodes_allowed, 0))
2766                         break;
2767                 cond_resched_lock(&hugetlb_lock);
2768         }
2769         while (count < persistent_huge_pages(h)) {
2770                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2771                         break;
2772         }
2773 out:
2774         h->max_huge_pages = persistent_huge_pages(h);
2775         spin_unlock(&hugetlb_lock);
2776
2777         NODEMASK_FREE(node_alloc_noretry);
2778
2779         return 0;
2780 }
2781
2782 #define HSTATE_ATTR_RO(_name) \
2783         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2784
2785 #define HSTATE_ATTR(_name) \
2786         static struct kobj_attribute _name##_attr = \
2787                 __ATTR(_name, 0644, _name##_show, _name##_store)
2788
2789 static struct kobject *hugepages_kobj;
2790 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2791
2792 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2793
2794 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2795 {
2796         int i;
2797
2798         for (i = 0; i < HUGE_MAX_HSTATE; i++)
2799                 if (hstate_kobjs[i] == kobj) {
2800                         if (nidp)
2801                                 *nidp = NUMA_NO_NODE;
2802                         return &hstates[i];
2803                 }
2804
2805         return kobj_to_node_hstate(kobj, nidp);
2806 }
2807
2808 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2809                                         struct kobj_attribute *attr, char *buf)
2810 {
2811         struct hstate *h;
2812         unsigned long nr_huge_pages;
2813         int nid;
2814
2815         h = kobj_to_hstate(kobj, &nid);
2816         if (nid == NUMA_NO_NODE)
2817                 nr_huge_pages = h->nr_huge_pages;
2818         else
2819                 nr_huge_pages = h->nr_huge_pages_node[nid];
2820
2821         return sprintf(buf, "%lu\n", nr_huge_pages);
2822 }
2823
2824 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2825                                            struct hstate *h, int nid,
2826                                            unsigned long count, size_t len)
2827 {
2828         int err;
2829         nodemask_t nodes_allowed, *n_mask;
2830
2831         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2832                 return -EINVAL;
2833
2834         if (nid == NUMA_NO_NODE) {
2835                 /*
2836                  * global hstate attribute
2837                  */
2838                 if (!(obey_mempolicy &&
2839                                 init_nodemask_of_mempolicy(&nodes_allowed)))
2840                         n_mask = &node_states[N_MEMORY];
2841                 else
2842                         n_mask = &nodes_allowed;
2843         } else {
2844                 /*
2845                  * Node specific request.  count adjustment happens in
2846                  * set_max_huge_pages() after acquiring hugetlb_lock.
2847                  */
2848                 init_nodemask_of_node(&nodes_allowed, nid);
2849                 n_mask = &nodes_allowed;
2850         }
2851
2852         err = set_max_huge_pages(h, count, nid, n_mask);
2853
2854         return err ? err : len;
2855 }
2856
2857 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2858                                          struct kobject *kobj, const char *buf,
2859                                          size_t len)
2860 {
2861         struct hstate *h;
2862         unsigned long count;
2863         int nid;
2864         int err;
2865
2866         err = kstrtoul(buf, 10, &count);
2867         if (err)
2868                 return err;
2869
2870         h = kobj_to_hstate(kobj, &nid);
2871         return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2872 }
2873
2874 static ssize_t nr_hugepages_show(struct kobject *kobj,
2875                                        struct kobj_attribute *attr, char *buf)
2876 {
2877         return nr_hugepages_show_common(kobj, attr, buf);
2878 }
2879
2880 static ssize_t nr_hugepages_store(struct kobject *kobj,
2881                struct kobj_attribute *attr, const char *buf, size_t len)
2882 {
2883         return nr_hugepages_store_common(false, kobj, buf, len);
2884 }
2885 HSTATE_ATTR(nr_hugepages);
2886
2887 #ifdef CONFIG_NUMA
2888
2889 /*
2890  * hstate attribute for optionally mempolicy-based constraint on persistent
2891  * huge page alloc/free.
2892  */
2893 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2894                                        struct kobj_attribute *attr, char *buf)
2895 {
2896         return nr_hugepages_show_common(kobj, attr, buf);
2897 }
2898
2899 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2900                struct kobj_attribute *attr, const char *buf, size_t len)
2901 {
2902         return nr_hugepages_store_common(true, kobj, buf, len);
2903 }
2904 HSTATE_ATTR(nr_hugepages_mempolicy);
2905 #endif
2906
2907
2908 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2909                                         struct kobj_attribute *attr, char *buf)
2910 {
2911         struct hstate *h = kobj_to_hstate(kobj, NULL);
2912         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2913 }
2914
2915 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2916                 struct kobj_attribute *attr, const char *buf, size_t count)
2917 {
2918         int err;
2919         unsigned long input;
2920         struct hstate *h = kobj_to_hstate(kobj, NULL);
2921
2922         if (hstate_is_gigantic(h))
2923                 return -EINVAL;
2924
2925         err = kstrtoul(buf, 10, &input);
2926         if (err)
2927                 return err;
2928
2929         spin_lock(&hugetlb_lock);
2930         h->nr_overcommit_huge_pages = input;
2931         spin_unlock(&hugetlb_lock);
2932
2933         return count;
2934 }
2935 HSTATE_ATTR(nr_overcommit_hugepages);
2936
2937 static ssize_t free_hugepages_show(struct kobject *kobj,
2938                                         struct kobj_attribute *attr, char *buf)
2939 {
2940         struct hstate *h;
2941         unsigned long free_huge_pages;
2942         int nid;
2943
2944         h = kobj_to_hstate(kobj, &nid);
2945         if (nid == NUMA_NO_NODE)
2946                 free_huge_pages = h->free_huge_pages;
2947         else
2948                 free_huge_pages = h->free_huge_pages_node[nid];
2949
2950         return sprintf(buf, "%lu\n", free_huge_pages);
2951 }
2952 HSTATE_ATTR_RO(free_hugepages);
2953
2954 static ssize_t resv_hugepages_show(struct kobject *kobj,
2955                                         struct kobj_attribute *attr, char *buf)
2956 {
2957         struct hstate *h = kobj_to_hstate(kobj, NULL);
2958         return sprintf(buf, "%lu\n", h->resv_huge_pages);
2959 }
2960 HSTATE_ATTR_RO(resv_hugepages);
2961
2962 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2963                                         struct kobj_attribute *attr, char *buf)
2964 {
2965         struct hstate *h;
2966         unsigned long surplus_huge_pages;
2967         int nid;
2968
2969         h = kobj_to_hstate(kobj, &nid);
2970         if (nid == NUMA_NO_NODE)
2971                 surplus_huge_pages = h->surplus_huge_pages;
2972         else
2973                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2974
2975         return sprintf(buf, "%lu\n", surplus_huge_pages);
2976 }
2977 HSTATE_ATTR_RO(surplus_hugepages);
2978
2979 static struct attribute *hstate_attrs[] = {
2980         &nr_hugepages_attr.attr,
2981         &nr_overcommit_hugepages_attr.attr,
2982         &free_hugepages_attr.attr,
2983         &resv_hugepages_attr.attr,
2984         &surplus_hugepages_attr.attr,
2985 #ifdef CONFIG_NUMA
2986         &nr_hugepages_mempolicy_attr.attr,
2987 #endif
2988         NULL,
2989 };
2990
2991 static const struct attribute_group hstate_attr_group = {
2992         .attrs = hstate_attrs,
2993 };
2994
2995 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2996                                     struct kobject **hstate_kobjs,
2997                                     const struct attribute_group *hstate_attr_group)
2998 {
2999         int retval;
3000         int hi = hstate_index(h);
3001
3002         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3003         if (!hstate_kobjs[hi])
3004                 return -ENOMEM;
3005
3006         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3007         if (retval)
3008                 kobject_put(hstate_kobjs[hi]);
3009
3010         return retval;
3011 }
3012
3013 static void __init hugetlb_sysfs_init(void)
3014 {
3015         struct hstate *h;
3016         int err;
3017
3018         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3019         if (!hugepages_kobj)
3020                 return;
3021
3022         for_each_hstate(h) {
3023                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3024                                          hstate_kobjs, &hstate_attr_group);
3025                 if (err)
3026                         pr_err("Hugetlb: Unable to add hstate %s", h->name);
3027         }
3028 }
3029
3030 #ifdef CONFIG_NUMA
3031
3032 /*
3033  * node_hstate/s - associate per node hstate attributes, via their kobjects,
3034  * with node devices in node_devices[] using a parallel array.  The array
3035  * index of a node device or _hstate == node id.
3036  * This is here to avoid any static dependency of the node device driver, in
3037  * the base kernel, on the hugetlb module.
3038  */
3039 struct node_hstate {
3040         struct kobject          *hugepages_kobj;
3041         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
3042 };
3043 static struct node_hstate node_hstates[MAX_NUMNODES];
3044
3045 /*
3046  * A subset of global hstate attributes for node devices
3047  */
3048 static struct attribute *per_node_hstate_attrs[] = {
3049         &nr_hugepages_attr.attr,
3050         &free_hugepages_attr.attr,
3051         &surplus_hugepages_attr.attr,
3052         NULL,
3053 };
3054
3055 static const struct attribute_group per_node_hstate_attr_group = {
3056         .attrs = per_node_hstate_attrs,
3057 };
3058
3059 /*
3060  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3061  * Returns node id via non-NULL nidp.
3062  */
3063 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3064 {
3065         int nid;
3066
3067         for (nid = 0; nid < nr_node_ids; nid++) {
3068                 struct node_hstate *nhs = &node_hstates[nid];
3069                 int i;
3070                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3071                         if (nhs->hstate_kobjs[i] == kobj) {
3072                                 if (nidp)
3073                                         *nidp = nid;
3074                                 return &hstates[i];
3075                         }
3076         }
3077
3078         BUG();
3079         return NULL;
3080 }
3081
3082 /*
3083  * Unregister hstate attributes from a single node device.
3084  * No-op if no hstate attributes attached.
3085  */
3086 static void hugetlb_unregister_node(struct node *node)
3087 {
3088         struct hstate *h;
3089         struct node_hstate *nhs = &node_hstates[node->dev.id];
3090
3091         if (!nhs->hugepages_kobj)
3092                 return;         /* no hstate attributes */
3093
3094         for_each_hstate(h) {
3095                 int idx = hstate_index(h);
3096                 if (nhs->hstate_kobjs[idx]) {
3097                         kobject_put(nhs->hstate_kobjs[idx]);
3098                         nhs->hstate_kobjs[idx] = NULL;
3099                 }
3100         }
3101
3102         kobject_put(nhs->hugepages_kobj);
3103         nhs->hugepages_kobj = NULL;
3104 }
3105
3106
3107 /*
3108  * Register hstate attributes for a single node device.
3109  * No-op if attributes already registered.
3110  */
3111 static void hugetlb_register_node(struct node *node)
3112 {
3113         struct hstate *h;
3114         struct node_hstate *nhs = &node_hstates[node->dev.id];
3115         int err;
3116
3117         if (nhs->hugepages_kobj)
3118                 return;         /* already allocated */
3119
3120         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3121                                                         &node->dev.kobj);
3122         if (!nhs->hugepages_kobj)
3123                 return;
3124
3125         for_each_hstate(h) {
3126                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3127                                                 nhs->hstate_kobjs,
3128                                                 &per_node_hstate_attr_group);
3129                 if (err) {
3130                         pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
3131                                 h->name, node->dev.id);
3132                         hugetlb_unregister_node(node);
3133                         break;
3134                 }
3135         }
3136 }
3137
3138 /*
3139  * hugetlb init time:  register hstate attributes for all registered node
3140  * devices of nodes that have memory.  All on-line nodes should have
3141  * registered their associated device by this time.
3142  */
3143 static void __init hugetlb_register_all_nodes(void)
3144 {
3145         int nid;
3146
3147         for_each_node_state(nid, N_MEMORY) {
3148                 struct node *node = node_devices[nid];
3149                 if (node->dev.id == nid)
3150                         hugetlb_register_node(node);
3151         }
3152
3153         /*
3154          * Let the node device driver know we're here so it can
3155          * [un]register hstate attributes on node hotplug.
3156          */
3157         register_hugetlbfs_with_node(hugetlb_register_node,
3158                                      hugetlb_unregister_node);
3159 }
3160 #else   /* !CONFIG_NUMA */
3161
3162 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3163 {
3164         BUG();
3165         if (nidp)
3166                 *nidp = -1;
3167         return NULL;
3168 }
3169
3170 static void hugetlb_register_all_nodes(void) { }
3171
3172 #endif
3173
3174 static int __init hugetlb_init(void)
3175 {
3176         int i;
3177
3178         if (!hugepages_supported())
3179                 return 0;
3180
3181         if (!size_to_hstate(default_hstate_size)) {
3182                 if (default_hstate_size != 0) {
3183                         pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
3184                                default_hstate_size, HPAGE_SIZE);
3185                 }
3186
3187                 default_hstate_size = HPAGE_SIZE;
3188                 if (!size_to_hstate(default_hstate_size))
3189                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3190         }
3191         default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
3192         if (default_hstate_max_huge_pages) {
3193                 if (!default_hstate.max_huge_pages)
3194                         default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3195         }
3196
3197         hugetlb_init_hstates();
3198         gather_bootmem_prealloc();
3199         report_hugepages();
3200
3201         hugetlb_sysfs_init();
3202         hugetlb_register_all_nodes();
3203         hugetlb_cgroup_file_init();
3204
3205 #ifdef CONFIG_SMP
3206         num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3207 #else
3208         num_fault_mutexes = 1;
3209 #endif
3210         hugetlb_fault_mutex_table =
3211                 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3212                               GFP_KERNEL);
3213         BUG_ON(!hugetlb_fault_mutex_table);
3214
3215         for (i = 0; i < num_fault_mutexes; i++)
3216                 mutex_init(&hugetlb_fault_mutex_table[i]);
3217         return 0;
3218 }
3219 subsys_initcall(hugetlb_init);
3220
3221 /* Should be called on processing a hugepagesz=... option */
3222 void __init hugetlb_bad_size(void)
3223 {
3224         parsed_valid_hugepagesz = false;
3225 }
3226
3227 void __init hugetlb_add_hstate(unsigned int order)
3228 {
3229         struct hstate *h;
3230         unsigned long i;
3231
3232         if (size_to_hstate(PAGE_SIZE << order)) {
3233                 pr_warn("hugepagesz= specified twice, ignoring\n");
3234                 return;
3235         }
3236         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3237         BUG_ON(order == 0);
3238         h = &hstates[hugetlb_max_hstate++];
3239         h->order = order;
3240         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
3241         h->nr_huge_pages = 0;
3242         h->free_huge_pages = 0;
3243         for (i = 0; i < MAX_NUMNODES; ++i)
3244                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3245         INIT_LIST_HEAD(&h->hugepage_activelist);
3246         h->next_nid_to_alloc = first_memory_node;
3247         h->next_nid_to_free = first_memory_node;
3248         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3249                                         huge_page_size(h)/1024);
3250
3251         parsed_hstate = h;
3252 }
3253
3254 static int __init hugetlb_nrpages_setup(char *s)
3255 {
3256         unsigned long *mhp;
3257         static unsigned long *last_mhp;
3258
3259         if (!parsed_valid_hugepagesz) {
3260                 pr_warn("hugepages = %s preceded by "
3261                         "an unsupported hugepagesz, ignoring\n", s);
3262                 parsed_valid_hugepagesz = true;
3263                 return 1;
3264         }
3265         /*
3266          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
3267          * so this hugepages= parameter goes to the "default hstate".
3268          */
3269         else if (!hugetlb_max_hstate)
3270                 mhp = &default_hstate_max_huge_pages;
3271         else
3272                 mhp = &parsed_hstate->max_huge_pages;
3273
3274         if (mhp == last_mhp) {
3275                 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
3276                 return 1;
3277         }
3278
3279         if (sscanf(s, "%lu", mhp) <= 0)
3280                 *mhp = 0;
3281
3282         /*
3283          * Global state is always initialized later in hugetlb_init.
3284          * But we need to allocate >= MAX_ORDER hstates here early to still
3285          * use the bootmem allocator.
3286          */
3287         if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
3288                 hugetlb_hstate_alloc_pages(parsed_hstate);
3289
3290         last_mhp = mhp;
3291
3292         return 1;
3293 }
3294 __setup("hugepages=", hugetlb_nrpages_setup);
3295
3296 static int __init hugetlb_default_setup(char *s)
3297 {
3298         default_hstate_size = memparse(s, &s);
3299         return 1;
3300 }
3301 __setup("default_hugepagesz=", hugetlb_default_setup);
3302
3303 static unsigned int cpuset_mems_nr(unsigned int *array)
3304 {
3305         int node;
3306         unsigned int nr = 0;
3307
3308         for_each_node_mask(node, cpuset_current_mems_allowed)
3309                 nr += array[node];
3310
3311         return nr;
3312 }
3313
3314 #ifdef CONFIG_SYSCTL
3315 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3316                          struct ctl_table *table, int write,
3317                          void __user *buffer, size_t *length, loff_t *ppos)
3318 {
3319         struct hstate *h = &default_hstate;
3320         unsigned long tmp = h->max_huge_pages;
3321         int ret;
3322
3323         if (!hugepages_supported())
3324                 return -EOPNOTSUPP;
3325
3326         table->data = &tmp;
3327         table->maxlen = sizeof(unsigned long);
3328         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3329         if (ret)
3330                 goto out;
3331
3332         if (write)
3333                 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3334                                                   NUMA_NO_NODE, tmp, *length);
3335 out:
3336         return ret;
3337 }
3338
3339 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3340                           void __user *buffer, size_t *length, loff_t *ppos)
3341 {
3342
3343         return hugetlb_sysctl_handler_common(false, table, write,
3344                                                         buffer, length, ppos);
3345 }
3346
3347 #ifdef CONFIG_NUMA
3348 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3349                           void __user *buffer, size_t *length, loff_t *ppos)
3350 {
3351         return hugetlb_sysctl_handler_common(true, table, write,
3352                                                         buffer, length, ppos);
3353 }
3354 #endif /* CONFIG_NUMA */
3355
3356 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3357                         void __user *buffer,
3358                         size_t *length, loff_t *ppos)
3359 {
3360         struct hstate *h = &default_hstate;
3361         unsigned long tmp;
3362         int ret;
3363
3364         if (!hugepages_supported())
3365                 return -EOPNOTSUPP;
3366
3367         tmp = h->nr_overcommit_huge_pages;
3368
3369         if (write && hstate_is_gigantic(h))
3370                 return -EINVAL;
3371
3372         table->data = &tmp;
3373         table->maxlen = sizeof(unsigned long);
3374         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3375         if (ret)
3376                 goto out;
3377
3378         if (write) {
3379                 spin_lock(&hugetlb_lock);
3380                 h->nr_overcommit_huge_pages = tmp;
3381                 spin_unlock(&hugetlb_lock);
3382         }
3383 out:
3384         return ret;
3385 }
3386
3387 #endif /* CONFIG_SYSCTL */
3388
3389 void hugetlb_report_meminfo(struct seq_file *m)
3390 {
3391         struct hstate *h;
3392         unsigned long total = 0;
3393
3394         if (!hugepages_supported())
3395                 return;
3396
3397         for_each_hstate(h) {
3398                 unsigned long count = h->nr_huge_pages;
3399
3400                 total += (PAGE_SIZE << huge_page_order(h)) * count;
3401
3402                 if (h == &default_hstate)
3403                         seq_printf(m,
3404                                    "HugePages_Total:   %5lu\n"
3405                                    "HugePages_Free:    %5lu\n"
3406                                    "HugePages_Rsvd:    %5lu\n"
3407                                    "HugePages_Surp:    %5lu\n"
3408                                    "Hugepagesize:   %8lu kB\n",
3409                                    count,
3410                                    h->free_huge_pages,
3411                                    h->resv_huge_pages,
3412                                    h->surplus_huge_pages,
3413                                    (PAGE_SIZE << huge_page_order(h)) / 1024);
3414         }
3415
3416         seq_printf(m, "Hugetlb:        %8lu kB\n", total / 1024);
3417 }
3418
3419 int hugetlb_report_node_meminfo(int nid, char *buf)
3420 {
3421         struct hstate *h = &default_hstate;
3422         if (!hugepages_supported())
3423                 return 0;
3424         return sprintf(buf,
3425                 "Node %d HugePages_Total: %5u\n"
3426                 "Node %d HugePages_Free:  %5u\n"
3427                 "Node %d HugePages_Surp:  %5u\n",
3428                 nid, h->nr_huge_pages_node[nid],
3429                 nid, h->free_huge_pages_node[nid],
3430                 nid, h->surplus_huge_pages_node[nid]);
3431 }
3432
3433 void hugetlb_show_meminfo(void)
3434 {
3435         struct hstate *h;
3436         int nid;
3437
3438         if (!hugepages_supported())
3439                 return;
3440
3441         for_each_node_state(nid, N_MEMORY)
3442                 for_each_hstate(h)
3443                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3444                                 nid,
3445                                 h->nr_huge_pages_node[nid],
3446                                 h->free_huge_pages_node[nid],
3447                                 h->surplus_huge_pages_node[nid],
3448                                 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3449 }
3450
3451 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3452 {
3453         seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3454                    atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3455 }
3456
3457 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3458 unsigned long hugetlb_total_pages(void)
3459 {
3460         struct hstate *h;
3461         unsigned long nr_total_pages = 0;
3462
3463         for_each_hstate(h)
3464                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3465         return nr_total_pages;
3466 }
3467
3468 static int hugetlb_acct_memory(struct hstate *h, long delta)
3469 {
3470         int ret = -ENOMEM;
3471
3472         spin_lock(&hugetlb_lock);
3473         /*
3474          * When cpuset is configured, it breaks the strict hugetlb page
3475          * reservation as the accounting is done on a global variable. Such
3476          * reservation is completely rubbish in the presence of cpuset because
3477          * the reservation is not checked against page availability for the
3478          * current cpuset. Application can still potentially OOM'ed by kernel
3479          * with lack of free htlb page in cpuset that the task is in.
3480          * Attempt to enforce strict accounting with cpuset is almost
3481          * impossible (or too ugly) because cpuset is too fluid that
3482          * task or memory node can be dynamically moved between cpusets.
3483          *
3484          * The change of semantics for shared hugetlb mapping with cpuset is
3485          * undesirable. However, in order to preserve some of the semantics,
3486          * we fall back to check against current free page availability as
3487          * a best attempt and hopefully to minimize the impact of changing
3488          * semantics that cpuset has.
3489          */
3490         if (delta > 0) {
3491                 if (gather_surplus_pages(h, delta) < 0)
3492                         goto out;
3493
3494                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3495                         return_unused_surplus_pages(h, delta);
3496                         goto out;
3497                 }
3498         }
3499
3500         ret = 0;
3501         if (delta < 0)
3502                 return_unused_surplus_pages(h, (unsigned long) -delta);
3503
3504 out:
3505         spin_unlock(&hugetlb_lock);
3506         return ret;
3507 }
3508
3509 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3510 {
3511         struct resv_map *resv = vma_resv_map(vma);
3512
3513         /*
3514          * This new VMA should share its siblings reservation map if present.
3515          * The VMA will only ever have a valid reservation map pointer where
3516          * it is being copied for another still existing VMA.  As that VMA
3517          * has a reference to the reservation map it cannot disappear until
3518          * after this open call completes.  It is therefore safe to take a
3519          * new reference here without additional locking.
3520          */
3521         if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3522                 kref_get(&resv->refs);
3523 }
3524
3525 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3526 {
3527         struct hstate *h = hstate_vma(vma);
3528         struct resv_map *resv = vma_resv_map(vma);
3529         struct hugepage_subpool *spool = subpool_vma(vma);
3530         unsigned long reserve, start, end;
3531         long gbl_reserve;
3532
3533         if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3534                 return;
3535
3536         start = vma_hugecache_offset(h, vma, vma->vm_start);
3537         end = vma_hugecache_offset(h, vma, vma->vm_end);
3538
3539         reserve = (end - start) - region_count(resv, start, end);
3540         hugetlb_cgroup_uncharge_counter(resv, start, end);
3541         if (reserve) {
3542                 /*
3543                  * Decrement reserve counts.  The global reserve count may be
3544                  * adjusted if the subpool has a minimum size.
3545                  */
3546                 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3547                 hugetlb_acct_memory(h, -gbl_reserve);
3548         }
3549
3550         kref_put(&resv->refs, resv_map_release);
3551 }
3552
3553 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3554 {
3555         if (addr & ~(huge_page_mask(hstate_vma(vma))))
3556                 return -EINVAL;
3557         return 0;
3558 }
3559
3560 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3561 {
3562         struct hstate *hstate = hstate_vma(vma);
3563
3564         return 1UL << huge_page_shift(hstate);
3565 }
3566
3567 /*
3568  * We cannot handle pagefaults against hugetlb pages at all.  They cause
3569  * handle_mm_fault() to try to instantiate regular-sized pages in the
3570  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
3571  * this far.
3572  */
3573 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3574 {
3575         BUG();
3576         return 0;
3577 }
3578
3579 /*
3580  * When a new function is introduced to vm_operations_struct and added
3581  * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3582  * This is because under System V memory model, mappings created via
3583  * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3584  * their original vm_ops are overwritten with shm_vm_ops.
3585  */
3586 const struct vm_operations_struct hugetlb_vm_ops = {
3587         .fault = hugetlb_vm_op_fault,
3588         .open = hugetlb_vm_op_open,
3589         .close = hugetlb_vm_op_close,
3590         .split = hugetlb_vm_op_split,
3591         .pagesize = hugetlb_vm_op_pagesize,
3592 };
3593
3594 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3595                                 int writable)
3596 {
3597         pte_t entry;
3598
3599         if (writable) {
3600                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3601                                          vma->vm_page_prot)));
3602         } else {
3603                 entry = huge_pte_wrprotect(mk_huge_pte(page,
3604                                            vma->vm_page_prot));
3605         }
3606         entry = pte_mkyoung(entry);
3607         entry = pte_mkhuge(entry);
3608         entry = arch_make_huge_pte(entry, vma, page, writable);
3609
3610         return entry;
3611 }
3612
3613 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3614                                    unsigned long address, pte_t *ptep)
3615 {
3616         pte_t entry;
3617
3618         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3619         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3620                 update_mmu_cache(vma, address, ptep);
3621 }
3622
3623 bool is_hugetlb_entry_migration(pte_t pte)
3624 {
3625         swp_entry_t swp;
3626
3627         if (huge_pte_none(pte) || pte_present(pte))
3628                 return false;
3629         swp = pte_to_swp_entry(pte);
3630         if (non_swap_entry(swp) && is_migration_entry(swp))
3631                 return true;
3632         else
3633                 return false;
3634 }
3635
3636 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3637 {
3638         swp_entry_t swp;
3639
3640         if (huge_pte_none(pte) || pte_present(pte))
3641                 return 0;
3642         swp = pte_to_swp_entry(pte);
3643         if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3644                 return 1;
3645         else
3646                 return 0;
3647 }
3648
3649 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3650                             struct vm_area_struct *vma)
3651 {
3652         pte_t *src_pte, *dst_pte, entry, dst_entry;
3653         struct page *ptepage;
3654         unsigned long addr;
3655         int cow;
3656         struct hstate *h = hstate_vma(vma);
3657         unsigned long sz = huge_page_size(h);
3658         struct address_space *mapping = vma->vm_file->f_mapping;
3659         struct mmu_notifier_range range;
3660         int ret = 0;
3661
3662         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3663
3664         if (cow) {
3665                 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3666                                         vma->vm_start,
3667                                         vma->vm_end);
3668                 mmu_notifier_invalidate_range_start(&range);
3669         } else {
3670                 /*
3671                  * For shared mappings i_mmap_rwsem must be held to call
3672                  * huge_pte_alloc, otherwise the returned ptep could go
3673                  * away if part of a shared pmd and another thread calls
3674                  * huge_pmd_unshare.
3675                  */
3676                 i_mmap_lock_read(mapping);
3677         }
3678
3679         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3680                 spinlock_t *src_ptl, *dst_ptl;
3681                 src_pte = huge_pte_offset(src, addr, sz);
3682                 if (!src_pte)
3683                         continue;
3684                 dst_pte = huge_pte_alloc(dst, addr, sz);
3685                 if (!dst_pte) {
3686                         ret = -ENOMEM;
3687                         break;
3688                 }
3689
3690                 /*
3691                  * If the pagetables are shared don't copy or take references.
3692                  * dst_pte == src_pte is the common case of src/dest sharing.
3693                  *
3694                  * However, src could have 'unshared' and dst shares with
3695                  * another vma.  If dst_pte !none, this implies sharing.
3696                  * Check here before taking page table lock, and once again
3697                  * after taking the lock below.
3698                  */
3699                 dst_entry = huge_ptep_get(dst_pte);
3700                 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3701                         continue;
3702
3703                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3704                 src_ptl = huge_pte_lockptr(h, src, src_pte);
3705                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3706                 entry = huge_ptep_get(src_pte);
3707                 dst_entry = huge_ptep_get(dst_pte);
3708                 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3709                         /*
3710                          * Skip if src entry none.  Also, skip in the
3711                          * unlikely case dst entry !none as this implies
3712                          * sharing with another vma.
3713                          */
3714                         ;
3715                 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3716                                     is_hugetlb_entry_hwpoisoned(entry))) {
3717                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
3718
3719                         if (is_write_migration_entry(swp_entry) && cow) {
3720                                 /*
3721                                  * COW mappings require pages in both
3722                                  * parent and child to be set to read.
3723                                  */
3724                                 make_migration_entry_read(&swp_entry);
3725                                 entry = swp_entry_to_pte(swp_entry);
3726                                 set_huge_swap_pte_at(src, addr, src_pte,
3727                                                      entry, sz);
3728                         }
3729                         set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3730                 } else {
3731                         if (cow) {
3732                                 /*
3733                                  * No need to notify as we are downgrading page
3734                                  * table protection not changing it to point
3735                                  * to a new page.
3736                                  *
3737                                  * See Documentation/vm/mmu_notifier.rst
3738                                  */
3739                                 huge_ptep_set_wrprotect(src, addr, src_pte);
3740                         }
3741                         entry = huge_ptep_get(src_pte);
3742                         ptepage = pte_page(entry);
3743                         get_page(ptepage);
3744                         page_dup_rmap(ptepage, true);
3745                         set_huge_pte_at(dst, addr, dst_pte, entry);
3746                         hugetlb_count_add(pages_per_huge_page(h), dst);
3747                 }
3748                 spin_unlock(src_ptl);
3749                 spin_unlock(dst_ptl);
3750         }
3751
3752         if (cow)
3753                 mmu_notifier_invalidate_range_end(&range);
3754         else
3755                 i_mmap_unlock_read(mapping);
3756
3757         return ret;
3758 }
3759
3760 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3761                             unsigned long start, unsigned long end,
3762                             struct page *ref_page)
3763 {
3764         struct mm_struct *mm = vma->vm_mm;
3765         unsigned long address;
3766         pte_t *ptep;
3767         pte_t pte;
3768         spinlock_t *ptl;
3769         struct page *page;
3770         struct hstate *h = hstate_vma(vma);
3771         unsigned long sz = huge_page_size(h);
3772         struct mmu_notifier_range range;
3773
3774         WARN_ON(!is_vm_hugetlb_page(vma));
3775         BUG_ON(start & ~huge_page_mask(h));
3776         BUG_ON(end & ~huge_page_mask(h));
3777
3778         /*
3779          * This is a hugetlb vma, all the pte entries should point
3780          * to huge page.
3781          */
3782         tlb_change_page_size(tlb, sz);
3783         tlb_start_vma(tlb, vma);
3784
3785         /*
3786          * If sharing possible, alert mmu notifiers of worst case.
3787          */
3788         mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3789                                 end);
3790         adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3791         mmu_notifier_invalidate_range_start(&range);
3792         address = start;
3793         for (; address < end; address += sz) {
3794                 ptep = huge_pte_offset(mm, address, sz);
3795                 if (!ptep)
3796                         continue;
3797
3798                 ptl = huge_pte_lock(h, mm, ptep);
3799                 if (huge_pmd_unshare(mm, &address, ptep)) {
3800                         spin_unlock(ptl);
3801                         /*
3802                          * We just unmapped a page of PMDs by clearing a PUD.
3803                          * The caller's TLB flush range should cover this area.
3804                          */
3805                         continue;
3806                 }
3807
3808                 pte = huge_ptep_get(ptep);
3809                 if (huge_pte_none(pte)) {
3810                         spin_unlock(ptl);
3811                         continue;
3812                 }
3813
3814                 /*
3815                  * Migrating hugepage or HWPoisoned hugepage is already
3816                  * unmapped and its refcount is dropped, so just clear pte here.
3817                  */
3818                 if (unlikely(!pte_present(pte))) {
3819                         huge_pte_clear(mm, address, ptep, sz);
3820                         spin_unlock(ptl);
3821                         continue;
3822                 }
3823
3824                 page = pte_page(pte);
3825                 /*
3826                  * If a reference page is supplied, it is because a specific
3827                  * page is being unmapped, not a range. Ensure the page we
3828                  * are about to unmap is the actual page of interest.
3829                  */
3830                 if (ref_page) {
3831                         if (page != ref_page) {
3832                                 spin_unlock(ptl);
3833                                 continue;
3834                         }
3835                         /*
3836                          * Mark the VMA as having unmapped its page so that
3837                          * future faults in this VMA will fail rather than
3838                          * looking like data was lost
3839                          */
3840                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3841                 }
3842
3843                 pte = huge_ptep_get_and_clear(mm, address, ptep);
3844                 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3845                 if (huge_pte_dirty(pte))
3846                         set_page_dirty(page);
3847
3848                 hugetlb_count_sub(pages_per_huge_page(h), mm);
3849                 page_remove_rmap(page, true);
3850
3851                 spin_unlock(ptl);
3852                 tlb_remove_page_size(tlb, page, huge_page_size(h));
3853                 /*
3854                  * Bail out after unmapping reference page if supplied
3855                  */
3856                 if (ref_page)
3857                         break;
3858         }
3859         mmu_notifier_invalidate_range_end(&range);
3860         tlb_end_vma(tlb, vma);
3861 }
3862
3863 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3864                           struct vm_area_struct *vma, unsigned long start,
3865                           unsigned long end, struct page *ref_page)
3866 {
3867         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3868
3869         /*
3870          * Clear this flag so that x86's huge_pmd_share page_table_shareable
3871          * test will fail on a vma being torn down, and not grab a page table
3872          * on its way out.  We're lucky that the flag has such an appropriate
3873          * name, and can in fact be safely cleared here. We could clear it
3874          * before the __unmap_hugepage_range above, but all that's necessary
3875          * is to clear it before releasing the i_mmap_rwsem. This works
3876          * because in the context this is called, the VMA is about to be
3877          * destroyed and the i_mmap_rwsem is held.
3878          */
3879         vma->vm_flags &= ~VM_MAYSHARE;
3880 }
3881
3882 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3883                           unsigned long end, struct page *ref_page)
3884 {
3885         struct mm_struct *mm;
3886         struct mmu_gather tlb;
3887         unsigned long tlb_start = start;
3888         unsigned long tlb_end = end;
3889
3890         /*
3891          * If shared PMDs were possibly used within this vma range, adjust
3892          * start/end for worst case tlb flushing.
3893          * Note that we can not be sure if PMDs are shared until we try to
3894          * unmap pages.  However, we want to make sure TLB flushing covers
3895          * the largest possible range.
3896          */
3897         adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
3898
3899         mm = vma->vm_mm;
3900
3901         tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
3902         __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3903         tlb_finish_mmu(&tlb, tlb_start, tlb_end);
3904 }
3905
3906 /*
3907  * This is called when the original mapper is failing to COW a MAP_PRIVATE
3908  * mappping it owns the reserve page for. The intention is to unmap the page
3909  * from other VMAs and let the children be SIGKILLed if they are faulting the
3910  * same region.
3911  */
3912 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3913                               struct page *page, unsigned long address)
3914 {
3915         struct hstate *h = hstate_vma(vma);
3916         struct vm_area_struct *iter_vma;
3917         struct address_space *mapping;
3918         pgoff_t pgoff;
3919
3920         /*
3921          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3922          * from page cache lookup which is in HPAGE_SIZE units.
3923          */
3924         address = address & huge_page_mask(h);
3925         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3926                         vma->vm_pgoff;
3927         mapping = vma->vm_file->f_mapping;
3928
3929         /*
3930          * Take the mapping lock for the duration of the table walk. As
3931          * this mapping should be shared between all the VMAs,
3932          * __unmap_hugepage_range() is called as the lock is already held
3933          */
3934         i_mmap_lock_write(mapping);
3935         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3936                 /* Do not unmap the current VMA */
3937                 if (iter_vma == vma)
3938                         continue;
3939
3940                 /*
3941                  * Shared VMAs have their own reserves and do not affect
3942                  * MAP_PRIVATE accounting but it is possible that a shared
3943                  * VMA is using the same page so check and skip such VMAs.
3944                  */
3945                 if (iter_vma->vm_flags & VM_MAYSHARE)
3946                         continue;
3947
3948                 /*
3949                  * Unmap the page from other VMAs without their own reserves.
3950                  * They get marked to be SIGKILLed if they fault in these
3951                  * areas. This is because a future no-page fault on this VMA
3952                  * could insert a zeroed page instead of the data existing
3953                  * from the time of fork. This would look like data corruption
3954                  */
3955                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3956                         unmap_hugepage_range(iter_vma, address,
3957                                              address + huge_page_size(h), page);
3958         }
3959         i_mmap_unlock_write(mapping);
3960 }
3961
3962 /*
3963  * Hugetlb_cow() should be called with page lock of the original hugepage held.
3964  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3965  * cannot race with other handlers or page migration.
3966  * Keep the pte_same checks anyway to make transition from the mutex easier.
3967  */
3968 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3969                        unsigned long address, pte_t *ptep,
3970                        struct page *pagecache_page, spinlock_t *ptl)
3971 {
3972         pte_t pte;
3973         struct hstate *h = hstate_vma(vma);
3974         struct page *old_page, *new_page;
3975         int outside_reserve = 0;
3976         vm_fault_t ret = 0;
3977         unsigned long haddr = address & huge_page_mask(h);
3978         struct mmu_notifier_range range;
3979
3980         pte = huge_ptep_get(ptep);
3981         old_page = pte_page(pte);
3982
3983 retry_avoidcopy:
3984         /* If no-one else is actually using this page, avoid the copy
3985          * and just make the page writable */
3986         if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3987                 page_move_anon_rmap(old_page, vma);
3988                 set_huge_ptep_writable(vma, haddr, ptep);
3989                 return 0;
3990         }
3991
3992         /*
3993          * If the process that created a MAP_PRIVATE mapping is about to
3994          * perform a COW due to a shared page count, attempt to satisfy
3995          * the allocation without using the existing reserves. The pagecache
3996          * page is used to determine if the reserve at this address was
3997          * consumed or not. If reserves were used, a partial faulted mapping
3998          * at the time of fork() could consume its reserves on COW instead
3999          * of the full address range.
4000          */
4001         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4002                         old_page != pagecache_page)
4003                 outside_reserve = 1;
4004
4005         get_page(old_page);
4006
4007         /*
4008          * Drop page table lock as buddy allocator may be called. It will
4009          * be acquired again before returning to the caller, as expected.
4010          */
4011         spin_unlock(ptl);
4012         new_page = alloc_huge_page(vma, haddr, outside_reserve);
4013
4014         if (IS_ERR(new_page)) {
4015                 /*
4016                  * If a process owning a MAP_PRIVATE mapping fails to COW,
4017                  * it is due to references held by a child and an insufficient
4018                  * huge page pool. To guarantee the original mappers
4019                  * reliability, unmap the page from child processes. The child
4020                  * may get SIGKILLed if it later faults.
4021                  */
4022                 if (outside_reserve) {
4023                         put_page(old_page);
4024                         BUG_ON(huge_pte_none(pte));
4025                         unmap_ref_private(mm, vma, old_page, haddr);
4026                         BUG_ON(huge_pte_none(pte));
4027                         spin_lock(ptl);
4028                         ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4029                         if (likely(ptep &&
4030                                    pte_same(huge_ptep_get(ptep), pte)))
4031                                 goto retry_avoidcopy;
4032                         /*
4033                          * race occurs while re-acquiring page table
4034                          * lock, and our job is done.
4035                          */
4036                         return 0;
4037                 }
4038
4039                 ret = vmf_error(PTR_ERR(new_page));
4040                 goto out_release_old;
4041         }
4042
4043         /*
4044          * When the original hugepage is shared one, it does not have
4045          * anon_vma prepared.
4046          */
4047         if (unlikely(anon_vma_prepare(vma))) {
4048                 ret = VM_FAULT_OOM;
4049                 goto out_release_all;
4050         }
4051
4052         copy_user_huge_page(new_page, old_page, address, vma,
4053                             pages_per_huge_page(h));
4054         __SetPageUptodate(new_page);
4055
4056         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4057                                 haddr + huge_page_size(h));
4058         mmu_notifier_invalidate_range_start(&range);
4059
4060         /*
4061          * Retake the page table lock to check for racing updates
4062          * before the page tables are altered
4063          */
4064         spin_lock(ptl);
4065         ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4066         if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4067                 ClearPagePrivate(new_page);
4068
4069                 /* Break COW */
4070                 huge_ptep_clear_flush(vma, haddr, ptep);
4071                 mmu_notifier_invalidate_range(mm, range.start, range.end);
4072                 set_huge_pte_at(mm, haddr, ptep,
4073                                 make_huge_pte(vma, new_page, 1));
4074                 page_remove_rmap(old_page, true);
4075                 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4076                 set_page_huge_active(new_page);
4077                 /* Make the old page be freed below */
4078                 new_page = old_page;
4079         }
4080         spin_unlock(ptl);
4081         mmu_notifier_invalidate_range_end(&range);
4082 out_release_all:
4083         restore_reserve_on_error(h, vma, haddr, new_page);
4084         put_page(new_page);
4085 out_release_old:
4086         put_page(old_page);
4087
4088         spin_lock(ptl); /* Caller expects lock to be held */
4089         return ret;
4090 }
4091
4092 /* Return the pagecache page at a given address within a VMA */
4093 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4094                         struct vm_area_struct *vma, unsigned long address)
4095 {
4096         struct address_space *mapping;
4097         pgoff_t idx;
4098
4099         mapping = vma->vm_file->f_mapping;
4100         idx = vma_hugecache_offset(h, vma, address);
4101
4102         return find_lock_page(mapping, idx);
4103 }
4104
4105 /*
4106  * Return whether there is a pagecache page to back given address within VMA.
4107  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4108  */
4109 static bool hugetlbfs_pagecache_present(struct hstate *h,
4110                         struct vm_area_struct *vma, unsigned long address)
4111 {
4112         struct address_space *mapping;
4113         pgoff_t idx;
4114         struct page *page;
4115
4116         mapping = vma->vm_file->f_mapping;
4117         idx = vma_hugecache_offset(h, vma, address);
4118
4119         page = find_get_page(mapping, idx);
4120         if (page)
4121                 put_page(page);
4122         return page != NULL;
4123 }
4124
4125 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4126                            pgoff_t idx)
4127 {
4128         struct inode *inode = mapping->host;
4129         struct hstate *h = hstate_inode(inode);
4130         int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4131
4132         if (err)
4133                 return err;
4134         ClearPagePrivate(page);
4135
4136         /*
4137          * set page dirty so that it will not be removed from cache/file
4138          * by non-hugetlbfs specific code paths.
4139          */
4140         set_page_dirty(page);
4141
4142         spin_lock(&inode->i_lock);
4143         inode->i_blocks += blocks_per_huge_page(h);
4144         spin_unlock(&inode->i_lock);
4145         return 0;
4146 }
4147
4148 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4149                         struct vm_area_struct *vma,
4150                         struct address_space *mapping, pgoff_t idx,
4151                         unsigned long address, pte_t *ptep, unsigned int flags)
4152 {
4153         struct hstate *h = hstate_vma(vma);
4154         vm_fault_t ret = VM_FAULT_SIGBUS;
4155         int anon_rmap = 0;
4156         unsigned long size;
4157         struct page *page;
4158         pte_t new_pte;
4159         spinlock_t *ptl;
4160         unsigned long haddr = address & huge_page_mask(h);
4161         bool new_page = false;
4162
4163         /*
4164          * Currently, we are forced to kill the process in the event the
4165          * original mapper has unmapped pages from the child due to a failed
4166          * COW. Warn that such a situation has occurred as it may not be obvious
4167          */
4168         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4169                 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4170                            current->pid);
4171                 return ret;
4172         }
4173
4174         /*
4175          * We can not race with truncation due to holding i_mmap_rwsem.
4176          * i_size is modified when holding i_mmap_rwsem, so check here
4177          * once for faults beyond end of file.
4178          */
4179         size = i_size_read(mapping->host) >> huge_page_shift(h);
4180         if (idx >= size)
4181                 goto out;
4182
4183 retry:
4184         page = find_lock_page(mapping, idx);
4185         if (!page) {
4186                 /*
4187                  * Check for page in userfault range
4188                  */
4189                 if (userfaultfd_missing(vma)) {
4190                         u32 hash;
4191                         struct vm_fault vmf = {
4192                                 .vma = vma,
4193                                 .address = haddr,
4194                                 .flags = flags,
4195                                 /*
4196                                  * Hard to debug if it ends up being
4197                                  * used by a callee that assumes
4198                                  * something about the other
4199                                  * uninitialized fields... same as in
4200                                  * memory.c
4201                                  */
4202                         };
4203
4204                         /*
4205                          * hugetlb_fault_mutex and i_mmap_rwsem must be
4206                          * dropped before handling userfault.  Reacquire
4207                          * after handling fault to make calling code simpler.
4208                          */
4209                         hash = hugetlb_fault_mutex_hash(mapping, idx);
4210                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4211                         i_mmap_unlock_read(mapping);
4212                         ret = handle_userfault(&vmf, VM_UFFD_MISSING);
4213                         i_mmap_lock_read(mapping);
4214                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
4215                         goto out;
4216                 }
4217
4218                 page = alloc_huge_page(vma, haddr, 0);
4219                 if (IS_ERR(page)) {
4220                         /*
4221                          * Returning error will result in faulting task being
4222                          * sent SIGBUS.  The hugetlb fault mutex prevents two
4223                          * tasks from racing to fault in the same page which
4224                          * could result in false unable to allocate errors.
4225                          * Page migration does not take the fault mutex, but
4226                          * does a clear then write of pte's under page table
4227                          * lock.  Page fault code could race with migration,
4228                          * notice the clear pte and try to allocate a page
4229                          * here.  Before returning error, get ptl and make
4230                          * sure there really is no pte entry.
4231                          */
4232                         ptl = huge_pte_lock(h, mm, ptep);
4233                         if (!huge_pte_none(huge_ptep_get(ptep))) {
4234                                 ret = 0;
4235                                 spin_unlock(ptl);
4236                                 goto out;
4237                         }
4238                         spin_unlock(ptl);
4239                         ret = vmf_error(PTR_ERR(page));
4240                         goto out;
4241                 }
4242                 clear_huge_page(page, address, pages_per_huge_page(h));
4243                 __SetPageUptodate(page);
4244                 new_page = true;
4245
4246                 if (vma->vm_flags & VM_MAYSHARE) {
4247                         int err = huge_add_to_page_cache(page, mapping, idx);
4248                         if (err) {
4249                                 put_page(page);
4250                                 if (err == -EEXIST)
4251                                         goto retry;
4252                                 goto out;
4253                         }
4254                 } else {
4255                         lock_page(page);
4256                         if (unlikely(anon_vma_prepare(vma))) {
4257                                 ret = VM_FAULT_OOM;
4258                                 goto backout_unlocked;
4259                         }
4260                         anon_rmap = 1;
4261                 }
4262         } else {
4263                 /*
4264                  * If memory error occurs between mmap() and fault, some process
4265                  * don't have hwpoisoned swap entry for errored virtual address.
4266                  * So we need to block hugepage fault by PG_hwpoison bit check.
4267                  */
4268                 if (unlikely(PageHWPoison(page))) {
4269                         ret = VM_FAULT_HWPOISON |
4270                                 VM_FAULT_SET_HINDEX(hstate_index(h));
4271                         goto backout_unlocked;
4272                 }
4273         }
4274
4275         /*
4276          * If we are going to COW a private mapping later, we examine the
4277          * pending reservations for this page now. This will ensure that
4278          * any allocations necessary to record that reservation occur outside
4279          * the spinlock.
4280          */
4281         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4282                 if (vma_needs_reservation(h, vma, haddr) < 0) {
4283                         ret = VM_FAULT_OOM;
4284                         goto backout_unlocked;
4285                 }
4286                 /* Just decrements count, does not deallocate */
4287                 vma_end_reservation(h, vma, haddr);
4288         }
4289
4290         ptl = huge_pte_lock(h, mm, ptep);
4291         ret = 0;
4292         if (!huge_pte_none(huge_ptep_get(ptep)))
4293                 goto backout;
4294
4295         if (anon_rmap) {
4296                 ClearPagePrivate(page);
4297                 hugepage_add_new_anon_rmap(page, vma, haddr);
4298         } else
4299                 page_dup_rmap(page, true);
4300         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4301                                 && (vma->vm_flags & VM_SHARED)));
4302         set_huge_pte_at(mm, haddr, ptep, new_pte);
4303
4304         hugetlb_count_add(pages_per_huge_page(h), mm);
4305         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4306                 /* Optimization, do the COW without a second fault */
4307                 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4308         }
4309
4310         spin_unlock(ptl);
4311
4312         /*
4313          * Only make newly allocated pages active.  Existing pages found
4314          * in the pagecache could be !page_huge_active() if they have been
4315          * isolated for migration.
4316          */
4317         if (new_page)
4318                 set_page_huge_active(page);
4319
4320         unlock_page(page);
4321 out:
4322         return ret;
4323
4324 backout:
4325         spin_unlock(ptl);
4326 backout_unlocked:
4327         unlock_page(page);
4328         restore_reserve_on_error(h, vma, haddr, page);
4329         put_page(page);
4330         goto out;
4331 }
4332
4333 #ifdef CONFIG_SMP
4334 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4335 {
4336         unsigned long key[2];
4337         u32 hash;
4338
4339         key[0] = (unsigned long) mapping;
4340         key[1] = idx;
4341
4342         hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4343
4344         return hash & (num_fault_mutexes - 1);
4345 }
4346 #else
4347 /*
4348  * For uniprocesor systems we always use a single mutex, so just
4349  * return 0 and avoid the hashing overhead.
4350  */
4351 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4352 {
4353         return 0;
4354 }
4355 #endif
4356
4357 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4358                         unsigned long address, unsigned int flags)
4359 {
4360         pte_t *ptep, entry;
4361         spinlock_t *ptl;
4362         vm_fault_t ret;
4363         u32 hash;
4364         pgoff_t idx;
4365         struct page *page = NULL;
4366         struct page *pagecache_page = NULL;
4367         struct hstate *h = hstate_vma(vma);
4368         struct address_space *mapping;
4369         int need_wait_lock = 0;
4370         unsigned long haddr = address & huge_page_mask(h);
4371
4372         ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4373         if (ptep) {
4374                 /*
4375                  * Since we hold no locks, ptep could be stale.  That is
4376                  * OK as we are only making decisions based on content and
4377                  * not actually modifying content here.
4378                  */
4379                 entry = huge_ptep_get(ptep);
4380                 if (unlikely(is_hugetlb_entry_migration(entry))) {
4381                         migration_entry_wait_huge(vma, mm, ptep);
4382                         return 0;
4383                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4384                         return VM_FAULT_HWPOISON_LARGE |
4385                                 VM_FAULT_SET_HINDEX(hstate_index(h));
4386         } else {
4387                 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4388                 if (!ptep)
4389                         return VM_FAULT_OOM;
4390         }
4391
4392         /*
4393          * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4394          * until finished with ptep.  This serves two purposes:
4395          * 1) It prevents huge_pmd_unshare from being called elsewhere
4396          *    and making the ptep no longer valid.
4397          * 2) It synchronizes us with i_size modifications during truncation.
4398          *
4399          * ptep could have already be assigned via huge_pte_offset.  That
4400          * is OK, as huge_pte_alloc will return the same value unless
4401          * something has changed.
4402          */
4403         mapping = vma->vm_file->f_mapping;
4404         i_mmap_lock_read(mapping);
4405         ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4406         if (!ptep) {
4407                 i_mmap_unlock_read(mapping);
4408                 return VM_FAULT_OOM;
4409         }
4410
4411         /*
4412          * Serialize hugepage allocation and instantiation, so that we don't
4413          * get spurious allocation failures if two CPUs race to instantiate
4414          * the same page in the page cache.
4415          */
4416         idx = vma_hugecache_offset(h, vma, haddr);
4417         hash = hugetlb_fault_mutex_hash(mapping, idx);
4418         mutex_lock(&hugetlb_fault_mutex_table[hash]);
4419
4420         entry = huge_ptep_get(ptep);
4421         if (huge_pte_none(entry)) {
4422                 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4423                 goto out_mutex;
4424         }
4425
4426         ret = 0;
4427
4428         /*
4429          * entry could be a migration/hwpoison entry at this point, so this
4430          * check prevents the kernel from going below assuming that we have
4431          * a active hugepage in pagecache. This goto expects the 2nd page fault,
4432          * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
4433          * handle it.
4434          */
4435         if (!pte_present(entry))
4436                 goto out_mutex;
4437
4438         /*
4439          * If we are going to COW the mapping later, we examine the pending
4440          * reservations for this page now. This will ensure that any
4441          * allocations necessary to record that reservation occur outside the
4442          * spinlock. For private mappings, we also lookup the pagecache
4443          * page now as it is used to determine if a reservation has been
4444          * consumed.
4445          */
4446         if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4447                 if (vma_needs_reservation(h, vma, haddr) < 0) {
4448                         ret = VM_FAULT_OOM;
4449                         goto out_mutex;
4450                 }
4451                 /* Just decrements count, does not deallocate */
4452                 vma_end_reservation(h, vma, haddr);
4453
4454                 if (!(vma->vm_flags & VM_MAYSHARE))
4455                         pagecache_page = hugetlbfs_pagecache_page(h,
4456                                                                 vma, haddr);
4457         }
4458
4459         ptl = huge_pte_lock(h, mm, ptep);
4460
4461         /* Check for a racing update before calling hugetlb_cow */
4462         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4463                 goto out_ptl;
4464
4465         /*
4466          * hugetlb_cow() requires page locks of pte_page(entry) and
4467          * pagecache_page, so here we need take the former one
4468          * when page != pagecache_page or !pagecache_page.
4469          */
4470         page = pte_page(entry);
4471         if (page != pagecache_page)
4472                 if (!trylock_page(page)) {
4473                         need_wait_lock = 1;
4474                         goto out_ptl;
4475                 }
4476
4477         get_page(page);
4478
4479         if (flags & FAULT_FLAG_WRITE) {
4480                 if (!huge_pte_write(entry)) {
4481                         ret = hugetlb_cow(mm, vma, address, ptep,
4482                                           pagecache_page, ptl);
4483                         goto out_put_page;
4484                 }
4485                 entry = huge_pte_mkdirty(entry);
4486         }
4487         entry = pte_mkyoung(entry);
4488         if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4489                                                 flags & FAULT_FLAG_WRITE))
4490                 update_mmu_cache(vma, haddr, ptep);
4491 out_put_page:
4492         if (page != pagecache_page)
4493                 unlock_page(page);
4494         put_page(page);
4495 out_ptl:
4496         spin_unlock(ptl);
4497
4498         if (pagecache_page) {
4499                 unlock_page(pagecache_page);
4500                 put_page(pagecache_page);
4501         }
4502 out_mutex:
4503         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4504         i_mmap_unlock_read(mapping);
4505         /*
4506          * Generally it's safe to hold refcount during waiting page lock. But
4507          * here we just wait to defer the next page fault to avoid busy loop and
4508          * the page is not used after unlocked before returning from the current
4509          * page fault. So we are safe from accessing freed page, even if we wait
4510          * here without taking refcount.
4511          */
4512         if (need_wait_lock)
4513                 wait_on_page_locked(page);
4514         return ret;
4515 }
4516
4517 /*
4518  * Used by userfaultfd UFFDIO_COPY.  Based on mcopy_atomic_pte with
4519  * modifications for huge pages.
4520  */
4521 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4522                             pte_t *dst_pte,
4523                             struct vm_area_struct *dst_vma,
4524                             unsigned long dst_addr,
4525                             unsigned long src_addr,
4526                             struct page **pagep)
4527 {
4528         struct address_space *mapping;
4529         pgoff_t idx;
4530         unsigned long size;
4531         int vm_shared = dst_vma->vm_flags & VM_SHARED;
4532         struct hstate *h = hstate_vma(dst_vma);
4533         pte_t _dst_pte;
4534         spinlock_t *ptl;
4535         int ret;
4536         struct page *page;
4537
4538         if (!*pagep) {
4539                 ret = -ENOMEM;
4540                 page = alloc_huge_page(dst_vma, dst_addr, 0);
4541                 if (IS_ERR(page))
4542                         goto out;
4543
4544                 ret = copy_huge_page_from_user(page,
4545                                                 (const void __user *) src_addr,
4546                                                 pages_per_huge_page(h), false);
4547
4548                 /* fallback to copy_from_user outside mmap_sem */
4549                 if (unlikely(ret)) {
4550                         ret = -ENOENT;
4551                         *pagep = page;
4552                         /* don't free the page */
4553                         goto out;
4554                 }
4555         } else {
4556                 page = *pagep;
4557                 *pagep = NULL;
4558         }
4559
4560         /*
4561          * The memory barrier inside __SetPageUptodate makes sure that
4562          * preceding stores to the page contents become visible before
4563          * the set_pte_at() write.
4564          */
4565         __SetPageUptodate(page);
4566
4567         mapping = dst_vma->vm_file->f_mapping;
4568         idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4569
4570         /*
4571          * If shared, add to page cache
4572          */
4573         if (vm_shared) {
4574                 size = i_size_read(mapping->host) >> huge_page_shift(h);
4575                 ret = -EFAULT;
4576                 if (idx >= size)
4577                         goto out_release_nounlock;
4578
4579                 /*
4580                  * Serialization between remove_inode_hugepages() and
4581                  * huge_add_to_page_cache() below happens through the
4582                  * hugetlb_fault_mutex_table that here must be hold by
4583                  * the caller.
4584                  */
4585                 ret = huge_add_to_page_cache(page, mapping, idx);
4586                 if (ret)
4587                         goto out_release_nounlock;
4588         }
4589
4590         ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4591         spin_lock(ptl);
4592
4593         /*
4594          * Recheck the i_size after holding PT lock to make sure not
4595          * to leave any page mapped (as page_mapped()) beyond the end
4596          * of the i_size (remove_inode_hugepages() is strict about
4597          * enforcing that). If we bail out here, we'll also leave a
4598          * page in the radix tree in the vm_shared case beyond the end
4599          * of the i_size, but remove_inode_hugepages() will take care
4600          * of it as soon as we drop the hugetlb_fault_mutex_table.
4601          */
4602         size = i_size_read(mapping->host) >> huge_page_shift(h);
4603         ret = -EFAULT;
4604         if (idx >= size)
4605                 goto out_release_unlock;
4606
4607         ret = -EEXIST;
4608         if (!huge_pte_none(huge_ptep_get(dst_pte)))
4609                 goto out_release_unlock;
4610
4611         if (vm_shared) {
4612                 page_dup_rmap(page, true);
4613         } else {
4614                 ClearPagePrivate(page);
4615                 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4616         }
4617
4618         _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4619         if (dst_vma->vm_flags & VM_WRITE)
4620                 _dst_pte = huge_pte_mkdirty(_dst_pte);
4621         _dst_pte = pte_mkyoung(_dst_pte);
4622
4623         set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4624
4625         (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4626                                         dst_vma->vm_flags & VM_WRITE);
4627         hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4628
4629         /* No need to invalidate - it was non-present before */
4630         update_mmu_cache(dst_vma, dst_addr, dst_pte);
4631
4632         spin_unlock(ptl);
4633         set_page_huge_active(page);
4634         if (vm_shared)
4635                 unlock_page(page);
4636         ret = 0;
4637 out:
4638         return ret;
4639 out_release_unlock:
4640         spin_unlock(ptl);
4641         if (vm_shared)
4642                 unlock_page(page);
4643 out_release_nounlock:
4644         put_page(page);
4645         goto out;
4646 }
4647
4648 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4649                          struct page **pages, struct vm_area_struct **vmas,
4650                          unsigned long *position, unsigned long *nr_pages,
4651                          long i, unsigned int flags, int *locked)
4652 {
4653         unsigned long pfn_offset;
4654         unsigned long vaddr = *position;
4655         unsigned long remainder = *nr_pages;
4656         struct hstate *h = hstate_vma(vma);
4657         int err = -EFAULT;
4658
4659         while (vaddr < vma->vm_end && remainder) {
4660                 pte_t *pte;
4661                 spinlock_t *ptl = NULL;
4662                 int absent;
4663                 struct page *page;
4664
4665                 /*
4666                  * If we have a pending SIGKILL, don't keep faulting pages and
4667                  * potentially allocating memory.
4668                  */
4669                 if (fatal_signal_pending(current)) {
4670                         remainder = 0;
4671                         break;
4672                 }
4673
4674                 /*
4675                  * Some archs (sparc64, sh*) have multiple pte_ts to
4676                  * each hugepage.  We have to make sure we get the
4677                  * first, for the page indexing below to work.
4678                  *
4679                  * Note that page table lock is not held when pte is null.
4680                  */
4681                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4682                                       huge_page_size(h));
4683                 if (pte)
4684                         ptl = huge_pte_lock(h, mm, pte);
4685                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4686
4687                 /*
4688                  * When coredumping, it suits get_dump_page if we just return
4689                  * an error where there's an empty slot with no huge pagecache
4690                  * to back it.  This way, we avoid allocating a hugepage, and
4691                  * the sparse dumpfile avoids allocating disk blocks, but its
4692                  * huge holes still show up with zeroes where they need to be.
4693                  */
4694                 if (absent && (flags & FOLL_DUMP) &&
4695                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4696                         if (pte)
4697                                 spin_unlock(ptl);
4698                         remainder = 0;
4699                         break;
4700                 }
4701
4702                 /*
4703                  * We need call hugetlb_fault for both hugepages under migration
4704                  * (in which case hugetlb_fault waits for the migration,) and
4705                  * hwpoisoned hugepages (in which case we need to prevent the
4706                  * caller from accessing to them.) In order to do this, we use
4707                  * here is_swap_pte instead of is_hugetlb_entry_migration and
4708                  * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4709                  * both cases, and because we can't follow correct pages
4710                  * directly from any kind of swap entries.
4711                  */
4712                 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4713                     ((flags & FOLL_WRITE) &&
4714                       !huge_pte_write(huge_ptep_get(pte)))) {
4715                         vm_fault_t ret;
4716                         unsigned int fault_flags = 0;
4717
4718                         if (pte)
4719                                 spin_unlock(ptl);
4720                         if (flags & FOLL_WRITE)
4721                                 fault_flags |= FAULT_FLAG_WRITE;
4722                         if (locked)
4723                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4724                                         FAULT_FLAG_KILLABLE;
4725                         if (flags & FOLL_NOWAIT)
4726                                 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4727                                         FAULT_FLAG_RETRY_NOWAIT;
4728                         if (flags & FOLL_TRIED) {
4729                                 /*
4730                                  * Note: FAULT_FLAG_ALLOW_RETRY and
4731                                  * FAULT_FLAG_TRIED can co-exist
4732                                  */
4733                                 fault_flags |= FAULT_FLAG_TRIED;
4734                         }
4735                         ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4736                         if (ret & VM_FAULT_ERROR) {
4737                                 err = vm_fault_to_errno(ret, flags);
4738                                 remainder = 0;
4739                                 break;
4740                         }
4741                         if (ret & VM_FAULT_RETRY) {
4742                                 if (locked &&
4743                                     !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4744                                         *locked = 0;
4745                                 *nr_pages = 0;
4746                                 /*
4747                                  * VM_FAULT_RETRY must not return an
4748                                  * error, it will return zero
4749                                  * instead.
4750                                  *
4751                                  * No need to update "position" as the
4752                                  * caller will not check it after
4753                                  * *nr_pages is set to 0.
4754                                  */
4755                                 return i;
4756                         }
4757                         continue;
4758                 }
4759
4760                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4761                 page = pte_page(huge_ptep_get(pte));
4762
4763                 /*
4764                  * If subpage information not requested, update counters
4765                  * and skip the same_page loop below.
4766                  */
4767                 if (!pages && !vmas && !pfn_offset &&
4768                     (vaddr + huge_page_size(h) < vma->vm_end) &&
4769                     (remainder >= pages_per_huge_page(h))) {
4770                         vaddr += huge_page_size(h);
4771                         remainder -= pages_per_huge_page(h);
4772                         i += pages_per_huge_page(h);
4773                         spin_unlock(ptl);
4774                         continue;
4775                 }
4776
4777 same_page:
4778                 if (pages) {
4779                         pages[i] = mem_map_offset(page, pfn_offset);
4780                         /*
4781                          * try_grab_page() should always succeed here, because:
4782                          * a) we hold the ptl lock, and b) we've just checked
4783                          * that the huge page is present in the page tables. If
4784                          * the huge page is present, then the tail pages must
4785                          * also be present. The ptl prevents the head page and
4786                          * tail pages from being rearranged in any way. So this
4787                          * page must be available at this point, unless the page
4788                          * refcount overflowed:
4789                          */
4790                         if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) {
4791                                 spin_unlock(ptl);
4792                                 remainder = 0;
4793                                 err = -ENOMEM;
4794                                 break;
4795                         }
4796                 }
4797
4798                 if (vmas)
4799                         vmas[i] = vma;
4800
4801                 vaddr += PAGE_SIZE;
4802                 ++pfn_offset;
4803                 --remainder;
4804                 ++i;
4805                 if (vaddr < vma->vm_end && remainder &&
4806                                 pfn_offset < pages_per_huge_page(h)) {
4807                         /*
4808                          * We use pfn_offset to avoid touching the pageframes
4809                          * of this compound page.
4810                          */
4811                         goto same_page;
4812                 }
4813                 spin_unlock(ptl);
4814         }
4815         *nr_pages = remainder;
4816         /*
4817          * setting position is actually required only if remainder is
4818          * not zero but it's faster not to add a "if (remainder)"
4819          * branch.
4820          */
4821         *position = vaddr;
4822
4823         return i ? i : err;
4824 }
4825
4826 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4827 /*
4828  * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4829  * implement this.
4830  */
4831 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4832 #endif
4833
4834 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4835                 unsigned long address, unsigned long end, pgprot_t newprot)
4836 {
4837         struct mm_struct *mm = vma->vm_mm;
4838         unsigned long start = address;
4839         pte_t *ptep;
4840         pte_t pte;
4841         struct hstate *h = hstate_vma(vma);
4842         unsigned long pages = 0;
4843         bool shared_pmd = false;
4844         struct mmu_notifier_range range;
4845
4846         /*
4847          * In the case of shared PMDs, the area to flush could be beyond
4848          * start/end.  Set range.start/range.end to cover the maximum possible
4849          * range if PMD sharing is possible.
4850          */
4851         mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
4852                                 0, vma, mm, start, end);
4853         adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4854
4855         BUG_ON(address >= end);
4856         flush_cache_range(vma, range.start, range.end);
4857
4858         mmu_notifier_invalidate_range_start(&range);
4859         i_mmap_lock_write(vma->vm_file->f_mapping);
4860         for (; address < end; address += huge_page_size(h)) {
4861                 spinlock_t *ptl;
4862                 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4863                 if (!ptep)
4864                         continue;
4865                 ptl = huge_pte_lock(h, mm, ptep);
4866                 if (huge_pmd_unshare(mm, &address, ptep)) {
4867                         pages++;
4868                         spin_unlock(ptl);
4869                         shared_pmd = true;
4870                         continue;
4871                 }
4872                 pte = huge_ptep_get(ptep);
4873                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4874                         spin_unlock(ptl);
4875                         continue;
4876                 }
4877                 if (unlikely(is_hugetlb_entry_migration(pte))) {
4878                         swp_entry_t entry = pte_to_swp_entry(pte);
4879
4880                         if (is_write_migration_entry(entry)) {
4881                                 pte_t newpte;
4882
4883                                 make_migration_entry_read(&entry);
4884                                 newpte = swp_entry_to_pte(entry);
4885                                 set_huge_swap_pte_at(mm, address, ptep,
4886                                                      newpte, huge_page_size(h));
4887                                 pages++;
4888                         }
4889                         spin_unlock(ptl);
4890                         continue;
4891                 }
4892                 if (!huge_pte_none(pte)) {
4893                         pte_t old_pte;
4894
4895                         old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
4896                         pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
4897                         pte = arch_make_huge_pte(pte, vma, NULL, 0);
4898                         huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
4899                         pages++;
4900                 }
4901                 spin_unlock(ptl);
4902         }
4903         /*
4904          * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4905          * may have cleared our pud entry and done put_page on the page table:
4906          * once we release i_mmap_rwsem, another task can do the final put_page
4907          * and that page table be reused and filled with junk.  If we actually
4908          * did unshare a page of pmds, flush the range corresponding to the pud.
4909          */
4910         if (shared_pmd)
4911                 flush_hugetlb_tlb_range(vma, range.start, range.end);
4912         else
4913                 flush_hugetlb_tlb_range(vma, start, end);
4914         /*
4915          * No need to call mmu_notifier_invalidate_range() we are downgrading
4916          * page table protection not changing it to point to a new page.
4917          *
4918          * See Documentation/vm/mmu_notifier.rst
4919          */
4920         i_mmap_unlock_write(vma->vm_file->f_mapping);
4921         mmu_notifier_invalidate_range_end(&range);
4922
4923         return pages << h->order;
4924 }
4925
4926 int hugetlb_reserve_pages(struct inode *inode,
4927                                         long from, long to,
4928                                         struct vm_area_struct *vma,
4929                                         vm_flags_t vm_flags)
4930 {
4931         long ret, chg, add = -1;
4932         struct hstate *h = hstate_inode(inode);
4933         struct hugepage_subpool *spool = subpool_inode(inode);
4934         struct resv_map *resv_map;
4935         struct hugetlb_cgroup *h_cg = NULL;
4936         long gbl_reserve, regions_needed = 0;
4937
4938         /* This should never happen */
4939         if (from > to) {
4940                 VM_WARN(1, "%s called with a negative range\n", __func__);
4941                 return -EINVAL;
4942         }
4943
4944         /*
4945          * Only apply hugepage reservation if asked. At fault time, an
4946          * attempt will be made for VM_NORESERVE to allocate a page
4947          * without using reserves
4948          */
4949         if (vm_flags & VM_NORESERVE)
4950                 return 0;
4951
4952         /*
4953          * Shared mappings base their reservation on the number of pages that
4954          * are already allocated on behalf of the file. Private mappings need
4955          * to reserve the full area even if read-only as mprotect() may be
4956          * called to make the mapping read-write. Assume !vma is a shm mapping
4957          */
4958         if (!vma || vma->vm_flags & VM_MAYSHARE) {
4959                 /*
4960                  * resv_map can not be NULL as hugetlb_reserve_pages is only
4961                  * called for inodes for which resv_maps were created (see
4962                  * hugetlbfs_get_inode).
4963                  */
4964                 resv_map = inode_resv_map(inode);
4965
4966                 chg = region_chg(resv_map, from, to, &regions_needed);
4967
4968         } else {
4969                 /* Private mapping. */
4970                 resv_map = resv_map_alloc();
4971                 if (!resv_map)
4972                         return -ENOMEM;
4973
4974                 chg = to - from;
4975
4976                 set_vma_resv_map(vma, resv_map);
4977                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4978         }
4979
4980         if (chg < 0) {
4981                 ret = chg;
4982                 goto out_err;
4983         }
4984
4985         ret = hugetlb_cgroup_charge_cgroup_rsvd(
4986                 hstate_index(h), chg * pages_per_huge_page(h), &h_cg);
4987
4988         if (ret < 0) {
4989                 ret = -ENOMEM;
4990                 goto out_err;
4991         }
4992
4993         if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
4994                 /* For private mappings, the hugetlb_cgroup uncharge info hangs
4995                  * of the resv_map.
4996                  */
4997                 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
4998         }
4999
5000         /*
5001          * There must be enough pages in the subpool for the mapping. If
5002          * the subpool has a minimum size, there may be some global
5003          * reservations already in place (gbl_reserve).
5004          */
5005         gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5006         if (gbl_reserve < 0) {
5007                 ret = -ENOSPC;
5008                 goto out_uncharge_cgroup;
5009         }
5010
5011         /*
5012          * Check enough hugepages are available for the reservation.
5013          * Hand the pages back to the subpool if there are not
5014          */
5015         ret = hugetlb_acct_memory(h, gbl_reserve);
5016         if (ret < 0) {
5017                 goto out_put_pages;
5018         }
5019
5020         /*
5021          * Account for the reservations made. Shared mappings record regions
5022          * that have reservations as they are shared by multiple VMAs.
5023          * When the last VMA disappears, the region map says how much
5024          * the reservation was and the page cache tells how much of
5025          * the reservation was consumed. Private mappings are per-VMA and
5026          * only the consumed reservations are tracked. When the VMA
5027          * disappears, the original reservation is the VMA size and the
5028          * consumed reservations are stored in the map. Hence, nothing
5029          * else has to be done for private mappings here
5030          */
5031         if (!vma || vma->vm_flags & VM_MAYSHARE) {
5032                 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5033
5034                 if (unlikely(add < 0)) {
5035                         hugetlb_acct_memory(h, -gbl_reserve);
5036                         goto out_put_pages;
5037                 } else if (unlikely(chg > add)) {
5038                         /*
5039                          * pages in this range were added to the reserve
5040                          * map between region_chg and region_add.  This
5041                          * indicates a race with alloc_huge_page.  Adjust
5042                          * the subpool and reserve counts modified above
5043                          * based on the difference.
5044                          */
5045                         long rsv_adjust;
5046
5047                         hugetlb_cgroup_uncharge_cgroup_rsvd(
5048                                 hstate_index(h),
5049                                 (chg - add) * pages_per_huge_page(h), h_cg);
5050
5051                         rsv_adjust = hugepage_subpool_put_pages(spool,
5052                                                                 chg - add);
5053                         hugetlb_acct_memory(h, -rsv_adjust);
5054                 }
5055         }
5056         return 0;
5057 out_put_pages:
5058         /* put back original number of pages, chg */
5059         (void)hugepage_subpool_put_pages(spool, chg);
5060 out_uncharge_cgroup:
5061         hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5062                                             chg * pages_per_huge_page(h), h_cg);
5063 out_err:
5064         if (!vma || vma->vm_flags & VM_MAYSHARE)
5065                 /* Only call region_abort if the region_chg succeeded but the
5066                  * region_add failed or didn't run.
5067                  */
5068                 if (chg >= 0 && add < 0)
5069                         region_abort(resv_map, from, to, regions_needed);
5070         if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5071                 kref_put(&resv_map->refs, resv_map_release);
5072         return ret;
5073 }
5074
5075 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5076                                                                 long freed)
5077 {
5078         struct hstate *h = hstate_inode(inode);
5079         struct resv_map *resv_map = inode_resv_map(inode);
5080         long chg = 0;
5081         struct hugepage_subpool *spool = subpool_inode(inode);
5082         long gbl_reserve;
5083
5084         /*
5085          * Since this routine can be called in the evict inode path for all
5086          * hugetlbfs inodes, resv_map could be NULL.
5087          */
5088         if (resv_map) {
5089                 chg = region_del(resv_map, start, end);
5090                 /*
5091                  * region_del() can fail in the rare case where a region
5092                  * must be split and another region descriptor can not be
5093                  * allocated.  If end == LONG_MAX, it will not fail.
5094                  */
5095                 if (chg < 0)
5096                         return chg;
5097         }
5098
5099         spin_lock(&inode->i_lock);
5100         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5101         spin_unlock(&inode->i_lock);
5102
5103         /*
5104          * If the subpool has a minimum size, the number of global
5105          * reservations to be released may be adjusted.
5106          */
5107         gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5108         hugetlb_acct_memory(h, -gbl_reserve);
5109
5110         return 0;
5111 }
5112
5113 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5114 static unsigned long page_table_shareable(struct vm_area_struct *svma,
5115                                 struct vm_area_struct *vma,
5116                                 unsigned long addr, pgoff_t idx)
5117 {
5118         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5119                                 svma->vm_start;
5120         unsigned long sbase = saddr & PUD_MASK;
5121         unsigned long s_end = sbase + PUD_SIZE;
5122
5123         /* Allow segments to share if only one is marked locked */
5124         unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5125         unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5126
5127         /*
5128          * match the virtual addresses, permission and the alignment of the
5129          * page table page.
5130          */
5131         if (pmd_index(addr) != pmd_index(saddr) ||
5132             vm_flags != svm_flags ||
5133             sbase < svma->vm_start || svma->vm_end < s_end)
5134                 return 0;
5135
5136         return saddr;
5137 }
5138
5139 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5140 {
5141         unsigned long base = addr & PUD_MASK;
5142         unsigned long end = base + PUD_SIZE;
5143
5144         /*
5145          * check on proper vm_flags and page table alignment
5146          */
5147         if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5148                 return true;
5149         return false;
5150 }
5151
5152 /*
5153  * Determine if start,end range within vma could be mapped by shared pmd.
5154  * If yes, adjust start and end to cover range associated with possible
5155  * shared pmd mappings.
5156  */
5157 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5158                                 unsigned long *start, unsigned long *end)
5159 {
5160         unsigned long check_addr;
5161
5162         if (!(vma->vm_flags & VM_MAYSHARE))
5163                 return;
5164
5165         for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
5166                 unsigned long a_start = check_addr & PUD_MASK;
5167                 unsigned long a_end = a_start + PUD_SIZE;
5168
5169                 /*
5170                  * If sharing is possible, adjust start/end if necessary.
5171                  */
5172                 if (range_in_vma(vma, a_start, a_end)) {
5173                         if (a_start < *start)
5174                                 *start = a_start;
5175                         if (a_end > *end)
5176                                 *end = a_end;
5177                 }
5178         }
5179 }
5180
5181 /*
5182  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5183  * and returns the corresponding pte. While this is not necessary for the
5184  * !shared pmd case because we can allocate the pmd later as well, it makes the
5185  * code much cleaner.
5186  *
5187  * This routine must be called with i_mmap_rwsem held in at least read mode.
5188  * For hugetlbfs, this prevents removal of any page table entries associated
5189  * with the address space.  This is important as we are setting up sharing
5190  * based on existing page table entries (mappings).
5191  */
5192 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5193 {
5194         struct vm_area_struct *vma = find_vma(mm, addr);
5195         struct address_space *mapping = vma->vm_file->f_mapping;
5196         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5197                         vma->vm_pgoff;
5198         struct vm_area_struct *svma;
5199         unsigned long saddr;
5200         pte_t *spte = NULL;
5201         pte_t *pte;
5202         spinlock_t *ptl;
5203
5204         if (!vma_shareable(vma, addr))
5205                 return (pte_t *)pmd_alloc(mm, pud, addr);
5206
5207         vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5208                 if (svma == vma)
5209                         continue;
5210
5211                 saddr = page_table_shareable(svma, vma, addr, idx);
5212                 if (saddr) {
5213                         spte = huge_pte_offset(svma->vm_mm, saddr,
5214                                                vma_mmu_pagesize(svma));
5215                         if (spte) {
5216                                 get_page(virt_to_page(spte));
5217                                 break;
5218                         }
5219                 }
5220         }
5221
5222         if (!spte)
5223                 goto out;
5224
5225         ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
5226         if (pud_none(*pud)) {
5227                 pud_populate(mm, pud,
5228                                 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5229                 mm_inc_nr_pmds(mm);
5230         } else {
5231                 put_page(virt_to_page(spte));
5232         }
5233         spin_unlock(ptl);
5234 out:
5235         pte = (pte_t *)pmd_alloc(mm, pud, addr);
5236         return pte;
5237 }
5238
5239 /*
5240  * unmap huge page backed by shared pte.
5241  *
5242  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
5243  * indicated by page_count > 1, unmap is achieved by clearing pud and
5244  * decrementing the ref count. If count == 1, the pte page is not shared.
5245  *
5246  * Called with page table lock held and i_mmap_rwsem held in write mode.
5247  *
5248  * returns: 1 successfully unmapped a shared pte page
5249  *          0 the underlying pte page is not shared, or it is the last user
5250  */
5251 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
5252 {
5253         pgd_t *pgd = pgd_offset(mm, *addr);
5254         p4d_t *p4d = p4d_offset(pgd, *addr);
5255         pud_t *pud = pud_offset(p4d, *addr);
5256
5257         BUG_ON(page_count(virt_to_page(ptep)) == 0);
5258         if (page_count(virt_to_page(ptep)) == 1)
5259                 return 0;
5260
5261         pud_clear(pud);
5262         put_page(virt_to_page(ptep));
5263         mm_dec_nr_pmds(mm);
5264         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5265         return 1;
5266 }
5267 #define want_pmd_share()        (1)
5268 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5269 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5270 {
5271         return NULL;
5272 }
5273
5274 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
5275 {
5276         return 0;
5277 }
5278
5279 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5280                                 unsigned long *start, unsigned long *end)
5281 {
5282 }
5283 #define want_pmd_share()        (0)
5284 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5285
5286 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5287 pte_t *huge_pte_alloc(struct mm_struct *mm,
5288                         unsigned long addr, unsigned long sz)
5289 {
5290         pgd_t *pgd;
5291         p4d_t *p4d;
5292         pud_t *pud;
5293         pte_t *pte = NULL;
5294
5295         pgd = pgd_offset(mm, addr);
5296         p4d = p4d_alloc(mm, pgd, addr);
5297         if (!p4d)
5298                 return NULL;
5299         pud = pud_alloc(mm, p4d, addr);
5300         if (pud) {
5301                 if (sz == PUD_SIZE) {
5302                         pte = (pte_t *)pud;
5303                 } else {
5304                         BUG_ON(sz != PMD_SIZE);
5305                         if (want_pmd_share() && pud_none(*pud))
5306                                 pte = huge_pmd_share(mm, addr, pud);
5307                         else
5308                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5309                 }
5310         }
5311         BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
5312
5313         return pte;
5314 }
5315
5316 /*
5317  * huge_pte_offset() - Walk the page table to resolve the hugepage
5318  * entry at address @addr
5319  *
5320  * Return: Pointer to page table or swap entry (PUD or PMD) for
5321  * address @addr, or NULL if a p*d_none() entry is encountered and the
5322  * size @sz doesn't match the hugepage size at this level of the page
5323  * table.
5324  */
5325 pte_t *huge_pte_offset(struct mm_struct *mm,
5326                        unsigned long addr, unsigned long sz)
5327 {
5328         pgd_t *pgd;
5329         p4d_t *p4d;
5330         pud_t *pud;
5331         pmd_t *pmd;
5332
5333         pgd = pgd_offset(mm, addr);
5334         if (!pgd_present(*pgd))
5335                 return NULL;
5336         p4d = p4d_offset(pgd, addr);
5337         if (!p4d_present(*p4d))
5338                 return NULL;
5339
5340         pud = pud_offset(p4d, addr);
5341         if (sz != PUD_SIZE && pud_none(*pud))
5342                 return NULL;
5343         /* hugepage or swap? */
5344         if (pud_huge(*pud) || !pud_present(*pud))
5345                 return (pte_t *)pud;
5346
5347         pmd = pmd_offset(pud, addr);
5348         if (sz != PMD_SIZE && pmd_none(*pmd))
5349                 return NULL;
5350         /* hugepage or swap? */
5351         if (pmd_huge(*pmd) || !pmd_present(*pmd))
5352                 return (pte_t *)pmd;
5353
5354         return NULL;
5355 }
5356
5357 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5358
5359 /*
5360  * These functions are overwritable if your architecture needs its own
5361  * behavior.
5362  */
5363 struct page * __weak
5364 follow_huge_addr(struct mm_struct *mm, unsigned long address,
5365                               int write)
5366 {
5367         return ERR_PTR(-EINVAL);
5368 }
5369
5370 struct page * __weak
5371 follow_huge_pd(struct vm_area_struct *vma,
5372                unsigned long address, hugepd_t hpd, int flags, int pdshift)
5373 {
5374         WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5375         return NULL;
5376 }
5377
5378 struct page * __weak
5379 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
5380                 pmd_t *pmd, int flags)
5381 {
5382         struct page *page = NULL;
5383         spinlock_t *ptl;
5384         pte_t pte;
5385
5386         /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5387         if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
5388                          (FOLL_PIN | FOLL_GET)))
5389                 return NULL;
5390
5391 retry:
5392         ptl = pmd_lockptr(mm, pmd);
5393         spin_lock(ptl);
5394         /*
5395          * make sure that the address range covered by this pmd is not
5396          * unmapped from other threads.
5397          */
5398         if (!pmd_huge(*pmd))
5399                 goto out;
5400         pte = huge_ptep_get((pte_t *)pmd);
5401         if (pte_present(pte)) {
5402                 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
5403                 /*
5404                  * try_grab_page() should always succeed here, because: a) we
5405                  * hold the pmd (ptl) lock, and b) we've just checked that the
5406                  * huge pmd (head) page is present in the page tables. The ptl
5407                  * prevents the head page and tail pages from being rearranged
5408                  * in any way. So this page must be available at this point,
5409                  * unless the page refcount overflowed:
5410                  */
5411                 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5412                         page = NULL;
5413                         goto out;
5414                 }
5415         } else {
5416                 if (is_hugetlb_entry_migration(pte)) {
5417                         spin_unlock(ptl);
5418                         __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5419                         goto retry;
5420                 }
5421                 /*
5422                  * hwpoisoned entry is treated as no_page_table in
5423                  * follow_page_mask().
5424                  */
5425         }
5426 out:
5427         spin_unlock(ptl);
5428         return page;
5429 }
5430
5431 struct page * __weak
5432 follow_huge_pud(struct mm_struct *mm, unsigned long address,
5433                 pud_t *pud, int flags)
5434 {
5435         if (flags & (FOLL_GET | FOLL_PIN))
5436                 return NULL;
5437
5438         return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5439 }
5440
5441 struct page * __weak
5442 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5443 {
5444         if (flags & (FOLL_GET | FOLL_PIN))
5445                 return NULL;
5446
5447         return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5448 }
5449
5450 bool isolate_huge_page(struct page *page, struct list_head *list)
5451 {
5452         bool ret = true;
5453
5454         VM_BUG_ON_PAGE(!PageHead(page), page);
5455         spin_lock(&hugetlb_lock);
5456         if (!page_huge_active(page) || !get_page_unless_zero(page)) {
5457                 ret = false;
5458                 goto unlock;
5459         }
5460         clear_page_huge_active(page);
5461         list_move_tail(&page->lru, list);
5462 unlock:
5463         spin_unlock(&hugetlb_lock);
5464         return ret;
5465 }
5466
5467 void putback_active_hugepage(struct page *page)
5468 {
5469         VM_BUG_ON_PAGE(!PageHead(page), page);
5470         spin_lock(&hugetlb_lock);
5471         set_page_huge_active(page);
5472         list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5473         spin_unlock(&hugetlb_lock);
5474         put_page(page);
5475 }
5476
5477 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5478 {
5479         struct hstate *h = page_hstate(oldpage);
5480
5481         hugetlb_cgroup_migrate(oldpage, newpage);
5482         set_page_owner_migrate_reason(newpage, reason);
5483
5484         /*
5485          * transfer temporary state of the new huge page. This is
5486          * reverse to other transitions because the newpage is going to
5487          * be final while the old one will be freed so it takes over
5488          * the temporary status.
5489          *
5490          * Also note that we have to transfer the per-node surplus state
5491          * here as well otherwise the global surplus count will not match
5492          * the per-node's.
5493          */
5494         if (PageHugeTemporary(newpage)) {
5495                 int old_nid = page_to_nid(oldpage);
5496                 int new_nid = page_to_nid(newpage);
5497
5498                 SetPageHugeTemporary(oldpage);
5499                 ClearPageHugeTemporary(newpage);
5500
5501                 spin_lock(&hugetlb_lock);
5502                 if (h->surplus_huge_pages_node[old_nid]) {
5503                         h->surplus_huge_pages_node[old_nid]--;
5504                         h->surplus_huge_pages_node[new_nid]++;
5505                 }
5506                 spin_unlock(&hugetlb_lock);
5507         }
5508 }